2021-05-24 13:30:07

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 00/10] Add support for SVM atomics in Nouveau

This is a repost of the previous series to rebase on v5.13-rc2 and to
address comments.

Outside of some code comment updates the primary change was to split the
renaming of migrate_pgmap_owner into a separate patch and to further
simplify the handling of device exclusive entries in copy_pte_range(). This
may result in temporary fork() failures if the process is using a device
whilst forking, but such usage is unlikely to be practical.

This resulted in a new clean-up patch for the series (patch 6) so that
device exclusive entries can be handled inside copy_nonpresent_pte(),
although more extensive clean-ups of copy_pte_range() are planned as
further development work in future.

Introduction
============

Some devices have features such as atomic PTE bits that can be used to
implement atomic access to system memory. To support atomic operations to a
shared virtual memory page such a device needs access to that page which is
exclusive of the CPU. This series introduces a mechanism to temporarily
unmap pages granting exclusive access to a device.

These changes are required to support OpenCL atomic operations in Nouveau
to shared virtual memory (SVM) regions allocated with the
CL_MEM_SVM_ATOMICS clSVMAlloc flag. A more complete description of the
OpenCL SVM feature is available at
https://www.khronos.org/registry/OpenCL/specs/3.0-unified/html/
OpenCL_API.html#_shared_virtual_memory .

Implementation
==============

Exclusive device access is implemented by adding a new swap entry type
(SWAP_DEVICE_EXCLUSIVE) which is similar to a migration entry. The main
difference is that on fault the original entry is immediately restored by
the fault handler instead of waiting.

Restoring the entry triggers calls to MMU notifers which allows a device
driver to revoke the atomic access permission from the GPU prior to the CPU
finalising the entry.

Patches
=======

Patches 1 & 2 refactor existing migration and device private entry
functions.

Patches 3 & 4 rework try_to_unmap_one() by splitting out unrelated
functionality into separate functions - try_to_migrate_one() and
try_to_munlock_one().

Patch 5 renames some existing code but does not introduce functionality.

Patch 6 is a small clean-up to swap entry handling in copy_pte_range().

Patch 7 contains the bulk of the implementation for device exclusive
memory.

Patch 8 contains some additions to the HMM selftests to ensure everything
works as expected.

Patch 9 is a cleanup for the Nouveau SVM implementation.

Patch 10 contains the implementation of atomic access for the Nouveau
driver.

Testing
=======

This has been tested with upstream Mesa 21.1.0 and a simple OpenCL program
which checks that GPU atomic accesses to system memory are atomic. Without
this series the test fails as there is no way of write-protecting the page
mapping which results in the device clobbering CPU writes. For reference
the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/

Further testing has been performed by adding support for testing exclusive
access to the hmm-tests kselftests.

Alistair Popple (10):
mm: Remove special swap entry functions
mm/swapops: Rework swap entry manipulation code
mm/rmap: Split try_to_munlock from try_to_unmap
mm/rmap: Split migration into its own function
mm: Rename migrate_pgmap_owner
mm/memory.c: Allow different return codes for copy_nonpresent_pte()
mm: Device exclusive memory access
mm: Selftests for exclusive device memory
nouveau/svm: Refactor nouveau_range_fault
nouveau/svm: Implement atomic SVM access

Documentation/vm/hmm.rst | 19 +-
Documentation/vm/unevictable-lru.rst | 33 +-
arch/s390/mm/pgtable.c | 2 +-
drivers/gpu/drm/nouveau/include/nvif/if000c.h | 1 +
drivers/gpu/drm/nouveau/nouveau_svm.c | 156 ++++-
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h | 1 +
.../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c | 6 +
fs/proc/task_mmu.c | 23 +-
include/linux/mmu_notifier.h | 26 +-
include/linux/rmap.h | 11 +-
include/linux/swap.h | 11 +-
include/linux/swapops.h | 123 ++--
lib/test_hmm.c | 126 +++-
lib/test_hmm_uapi.h | 2 +
mm/debug_vm_pgtable.c | 12 +-
mm/hmm.c | 12 +-
mm/huge_memory.c | 45 +-
mm/hugetlb.c | 10 +-
mm/memcontrol.c | 2 +-
mm/memory.c | 160 ++++-
mm/migrate.c | 51 +-
mm/mlock.c | 10 +-
mm/mprotect.c | 18 +-
mm/page_vma_mapped.c | 15 +-
mm/rmap.c | 601 +++++++++++++++---
tools/testing/selftests/vm/hmm-tests.c | 158 +++++
26 files changed, 1317 insertions(+), 317 deletions(-)

--
2.20.1


2021-05-24 13:30:14

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 01/10] mm: Remove special swap entry functions

Remove multiple similar inline functions for dealing with different
types of special swap entries.

Both migration and device private swap entries use the swap offset to
store a pfn. Instead of multiple inline functions to obtain a struct
page for each swap entry type use a common function
pfn_swap_entry_to_page(). Also open-code the various entry_to_pfn()
functions as this results is shorter code that is easier to understand.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Ralph Campbell <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>

---

v9:
* Rebased on v5.13-rc2

v8:
* No changes

v7:
* Reworded commit message to include pfn_swap_entry_to_page()
* Added Christoph's Reviewed-by

v6:
* Removed redundant compound_page() call from inside PageLocked()
* Fixed a minor build issue for s390 reported by kernel test bot

v4:
* Added pfn_swap_entry_to_page()
* Reinstated check that migration entries point to locked pages
* Removed #define swapcache_prepare which isn't needed for CONFIG_SWAP=0
builds
---
arch/s390/mm/pgtable.c | 2 +-
fs/proc/task_mmu.c | 23 +++++---------
include/linux/swap.h | 4 +--
include/linux/swapops.h | 69 ++++++++++++++---------------------------
mm/hmm.c | 5 ++-
mm/huge_memory.c | 4 +--
mm/memcontrol.c | 2 +-
mm/memory.c | 10 +++---
mm/migrate.c | 6 ++--
mm/page_vma_mapped.c | 6 ++--
10 files changed, 50 insertions(+), 81 deletions(-)

diff --git a/arch/s390/mm/pgtable.c b/arch/s390/mm/pgtable.c
index 18205f851c24..eec3a9d7176e 100644
--- a/arch/s390/mm/pgtable.c
+++ b/arch/s390/mm/pgtable.c
@@ -691,7 +691,7 @@ static void ptep_zap_swap_entry(struct mm_struct *mm, swp_entry_t entry)
if (!non_swap_entry(entry))
dec_mm_counter(mm, MM_SWAPENTS);
else if (is_migration_entry(entry)) {
- struct page *page = migration_entry_to_page(entry);
+ struct page *page = pfn_swap_entry_to_page(entry);

dec_mm_counter(mm, mm_counter(page));
}
diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
index fc9784544b24..0953732c8ce1 100644
--- a/fs/proc/task_mmu.c
+++ b/fs/proc/task_mmu.c
@@ -514,10 +514,8 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
} else {
mss->swap_pss += (u64)PAGE_SIZE << PSS_SHIFT;
}
- } else if (is_migration_entry(swpent))
- page = migration_entry_to_page(swpent);
- else if (is_device_private_entry(swpent))
- page = device_private_entry_to_page(swpent);
+ } else if (is_pfn_swap_entry(swpent))
+ page = pfn_swap_entry_to_page(swpent);
} else if (unlikely(IS_ENABLED(CONFIG_SHMEM) && mss->check_shmem_swap
&& pte_none(*pte))) {
page = xa_load(&vma->vm_file->f_mapping->i_pages,
@@ -549,7 +547,7 @@ static void smaps_pmd_entry(pmd_t *pmd, unsigned long addr,
swp_entry_t entry = pmd_to_swp_entry(*pmd);

if (is_migration_entry(entry))
- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
}
if (IS_ERR_OR_NULL(page))
return;
@@ -694,10 +692,8 @@ static int smaps_hugetlb_range(pte_t *pte, unsigned long hmask,
} else if (is_swap_pte(*pte)) {
swp_entry_t swpent = pte_to_swp_entry(*pte);

- if (is_migration_entry(swpent))
- page = migration_entry_to_page(swpent);
- else if (is_device_private_entry(swpent))
- page = device_private_entry_to_page(swpent);
+ if (is_pfn_swap_entry(swpent))
+ page = pfn_swap_entry_to_page(swpent);
}
if (page) {
int mapcount = page_mapcount(page);
@@ -1384,11 +1380,8 @@ static pagemap_entry_t pte_to_pagemap_entry(struct pagemapread *pm,
frame = swp_type(entry) |
(swp_offset(entry) << MAX_SWAPFILES_SHIFT);
flags |= PM_SWAP;
- if (is_migration_entry(entry))
- page = migration_entry_to_page(entry);
-
- if (is_device_private_entry(entry))
- page = device_private_entry_to_page(entry);
+ if (is_pfn_swap_entry(entry))
+ page = pfn_swap_entry_to_page(entry);
}

if (page && !PageAnon(page))
@@ -1445,7 +1438,7 @@ static int pagemap_pmd_range(pmd_t *pmdp, unsigned long addr, unsigned long end,
if (pmd_swp_soft_dirty(pmd))
flags |= PM_SOFT_DIRTY;
VM_BUG_ON(!is_pmd_migration_entry(pmd));
- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
}
#endif

diff --git a/include/linux/swap.h b/include/linux/swap.h
index 144727041e78..a6d4505ecf73 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -545,8 +545,8 @@ static inline void show_swap_cache_info(void)
{
}

-#define free_swap_and_cache(e) ({(is_migration_entry(e) || is_device_private_entry(e));})
-#define swapcache_prepare(e) ({(is_migration_entry(e) || is_device_private_entry(e));})
+/* used to sanity check ptes in zap_pte_range when CONFIG_SWAP=0 */
+#define free_swap_and_cache(e) is_pfn_swap_entry(e)

static inline int add_swap_count_continuation(swp_entry_t swp, gfp_t gfp_mask)
{
diff --git a/include/linux/swapops.h b/include/linux/swapops.h
index d9b7c9132c2f..139be8235ad2 100644
--- a/include/linux/swapops.h
+++ b/include/linux/swapops.h
@@ -121,16 +121,6 @@ static inline bool is_write_device_private_entry(swp_entry_t entry)
{
return unlikely(swp_type(entry) == SWP_DEVICE_WRITE);
}
-
-static inline unsigned long device_private_entry_to_pfn(swp_entry_t entry)
-{
- return swp_offset(entry);
-}
-
-static inline struct page *device_private_entry_to_page(swp_entry_t entry)
-{
- return pfn_to_page(swp_offset(entry));
-}
#else /* CONFIG_DEVICE_PRIVATE */
static inline swp_entry_t make_device_private_entry(struct page *page, bool write)
{
@@ -150,16 +140,6 @@ static inline bool is_write_device_private_entry(swp_entry_t entry)
{
return false;
}
-
-static inline unsigned long device_private_entry_to_pfn(swp_entry_t entry)
-{
- return 0;
-}
-
-static inline struct page *device_private_entry_to_page(swp_entry_t entry)
-{
- return NULL;
-}
#endif /* CONFIG_DEVICE_PRIVATE */

#ifdef CONFIG_MIGRATION
@@ -182,22 +162,6 @@ static inline int is_write_migration_entry(swp_entry_t entry)
return unlikely(swp_type(entry) == SWP_MIGRATION_WRITE);
}

-static inline unsigned long migration_entry_to_pfn(swp_entry_t entry)
-{
- return swp_offset(entry);
-}
-
-static inline struct page *migration_entry_to_page(swp_entry_t entry)
-{
- struct page *p = pfn_to_page(swp_offset(entry));
- /*
- * Any use of migration entries may only occur while the
- * corresponding page is locked
- */
- BUG_ON(!PageLocked(compound_head(p)));
- return p;
-}
-
static inline void make_migration_entry_read(swp_entry_t *entry)
{
*entry = swp_entry(SWP_MIGRATION_READ, swp_offset(*entry));
@@ -217,16 +181,6 @@ static inline int is_migration_entry(swp_entry_t swp)
return 0;
}

-static inline unsigned long migration_entry_to_pfn(swp_entry_t entry)
-{
- return 0;
-}
-
-static inline struct page *migration_entry_to_page(swp_entry_t entry)
-{
- return NULL;
-}
-
static inline void make_migration_entry_read(swp_entry_t *entryp) { }
static inline void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
spinlock_t *ptl) { }
@@ -241,6 +195,29 @@ static inline int is_write_migration_entry(swp_entry_t entry)

#endif

+static inline struct page *pfn_swap_entry_to_page(swp_entry_t entry)
+{
+ struct page *p = pfn_to_page(swp_offset(entry));
+
+ /*
+ * Any use of migration entries may only occur while the
+ * corresponding page is locked
+ */
+ BUG_ON(is_migration_entry(entry) && !PageLocked(p));
+
+ return p;
+}
+
+/*
+ * A pfn swap entry is a special type of swap entry that always has a pfn stored
+ * in the swap offset. They are used to represent unaddressable device memory
+ * and to restrict access to a page undergoing migration.
+ */
+static inline bool is_pfn_swap_entry(swp_entry_t entry)
+{
+ return is_migration_entry(entry) || is_device_private_entry(entry);
+}
+
struct page_vma_mapped_walk;

#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
diff --git a/mm/hmm.c b/mm/hmm.c
index 943cb2ba4442..3b2dda71d0ed 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -214,7 +214,7 @@ static inline bool hmm_is_device_private_entry(struct hmm_range *range,
swp_entry_t entry)
{
return is_device_private_entry(entry) &&
- device_private_entry_to_page(entry)->pgmap->owner ==
+ pfn_swap_entry_to_page(entry)->pgmap->owner ==
range->dev_private_owner;
}

@@ -257,8 +257,7 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
cpu_flags = HMM_PFN_VALID;
if (is_write_device_private_entry(entry))
cpu_flags |= HMM_PFN_WRITE;
- *hmm_pfn = device_private_entry_to_pfn(entry) |
- cpu_flags;
+ *hmm_pfn = swp_offset(entry) | cpu_flags;
return 0;
}

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 63ed6b25deaa..7137ab31766a 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1694,7 +1694,7 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,

VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
entry = pmd_to_swp_entry(orig_pmd);
- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
flush_needed = 0;
} else
WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
@@ -2102,7 +2102,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
swp_entry_t entry;

entry = pmd_to_swp_entry(old_pmd);
- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
write = is_write_migration_entry(entry);
young = false;
soft_dirty = pmd_swp_soft_dirty(old_pmd);
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 64ada9e650a5..649758b78d27 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -5349,7 +5349,7 @@ static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
* as special swap entry in the CPU page table.
*/
if (is_device_private_entry(ent)) {
- page = device_private_entry_to_page(ent);
+ page = pfn_swap_entry_to_page(ent);
/*
* MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
* a refcount of 1 when free (unlike normal page)
diff --git a/mm/memory.c b/mm/memory.c
index 730daa00952b..1f5c3f6134fb 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -730,7 +730,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
}
rss[MM_SWAPENTS]++;
} else if (is_migration_entry(entry)) {
- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);

rss[mm_counter(page)]++;

@@ -749,7 +749,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
set_pte_at(src_mm, addr, src_pte, pte);
}
} else if (is_device_private_entry(entry)) {
- page = device_private_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);

/*
* Update rss count even for unaddressable pages, as
@@ -1280,7 +1280,7 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,

entry = pte_to_swp_entry(ptent);
if (is_device_private_entry(entry)) {
- struct page *page = device_private_entry_to_page(entry);
+ struct page *page = pfn_swap_entry_to_page(entry);

if (unlikely(details && details->check_mapping)) {
/*
@@ -1309,7 +1309,7 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
else if (is_migration_entry(entry)) {
struct page *page;

- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
rss[mm_counter(page)]--;
}
if (unlikely(!free_swap_and_cache(entry)))
@@ -3327,7 +3327,7 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
migration_entry_wait(vma->vm_mm, vmf->pmd,
vmf->address);
} else if (is_device_private_entry(entry)) {
- vmf->page = device_private_entry_to_page(entry);
+ vmf->page = pfn_swap_entry_to_page(entry);
ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
} else if (is_hwpoison_entry(entry)) {
ret = VM_FAULT_HWPOISON;
diff --git a/mm/migrate.c b/mm/migrate.c
index b234c3f3acb7..749321ae3026 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -294,7 +294,7 @@ void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
if (!is_migration_entry(entry))
goto out;

- page = migration_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);

/*
* Once page cache replacement of page migration started, page_count
@@ -334,7 +334,7 @@ void pmd_migration_entry_wait(struct mm_struct *mm, pmd_t *pmd)
ptl = pmd_lock(mm, pmd);
if (!is_pmd_migration_entry(*pmd))
goto unlock;
- page = migration_entry_to_page(pmd_to_swp_entry(*pmd));
+ page = pfn_swap_entry_to_page(pmd_to_swp_entry(*pmd));
if (!get_page_unless_zero(page))
goto unlock;
spin_unlock(ptl);
@@ -2399,7 +2399,7 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
if (!is_device_private_entry(entry))
goto next;

- page = device_private_entry_to_page(entry);
+ page = pfn_swap_entry_to_page(entry);
if (!(migrate->flags &
MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
page->pgmap->owner != migrate->pgmap_owner)
diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
index 2cf01d933f13..a6a7febb4d93 100644
--- a/mm/page_vma_mapped.c
+++ b/mm/page_vma_mapped.c
@@ -96,7 +96,7 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
if (!is_migration_entry(entry))
return false;

- pfn = migration_entry_to_pfn(entry);
+ pfn = swp_offset(entry);
} else if (is_swap_pte(*pvmw->pte)) {
swp_entry_t entry;

@@ -105,7 +105,7 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
if (!is_device_private_entry(entry))
return false;

- pfn = device_private_entry_to_pfn(entry);
+ pfn = swp_offset(entry);
} else {
if (!pte_present(*pvmw->pte))
return false;
@@ -200,7 +200,7 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
if (is_migration_entry(pmd_to_swp_entry(*pvmw->pmd))) {
swp_entry_t entry = pmd_to_swp_entry(*pvmw->pmd);

- if (migration_entry_to_page(entry) != page)
+ if (pfn_swap_entry_to_page(entry) != page)
return not_found(pvmw);
return true;
}
--
2.20.1

2021-05-24 13:30:28

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 02/10] mm/swapops: Rework swap entry manipulation code

Both migration and device private pages use special swap entries that
are manipluated by a range of inline functions. The arguments to these
are somewhat inconsitent so rework them to remove flag type arguments
and to make the arguments similar for both read and write entry
creation.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Reviewed-by: Jason Gunthorpe <[email protected]>
Reviewed-by: Ralph Campbell <[email protected]>
---
include/linux/swapops.h | 56 ++++++++++++++++++++++-------------------
mm/debug_vm_pgtable.c | 12 ++++-----
mm/hmm.c | 2 +-
mm/huge_memory.c | 26 +++++++++++++------
mm/hugetlb.c | 10 +++++---
mm/memory.c | 10 +++++---
mm/migrate.c | 26 ++++++++++++++-----
mm/mprotect.c | 10 +++++---
mm/rmap.c | 10 +++++---
9 files changed, 100 insertions(+), 62 deletions(-)

diff --git a/include/linux/swapops.h b/include/linux/swapops.h
index 139be8235ad2..4dfd807ae52a 100644
--- a/include/linux/swapops.h
+++ b/include/linux/swapops.h
@@ -100,35 +100,35 @@ static inline void *swp_to_radix_entry(swp_entry_t entry)
}

#if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
-static inline swp_entry_t make_device_private_entry(struct page *page, bool write)
+static inline swp_entry_t make_readable_device_private_entry(pgoff_t offset)
{
- return swp_entry(write ? SWP_DEVICE_WRITE : SWP_DEVICE_READ,
- page_to_pfn(page));
+ return swp_entry(SWP_DEVICE_READ, offset);
}

-static inline bool is_device_private_entry(swp_entry_t entry)
+static inline swp_entry_t make_writable_device_private_entry(pgoff_t offset)
{
- int type = swp_type(entry);
- return type == SWP_DEVICE_READ || type == SWP_DEVICE_WRITE;
+ return swp_entry(SWP_DEVICE_WRITE, offset);
}

-static inline void make_device_private_entry_read(swp_entry_t *entry)
+static inline bool is_device_private_entry(swp_entry_t entry)
{
- *entry = swp_entry(SWP_DEVICE_READ, swp_offset(*entry));
+ int type = swp_type(entry);
+ return type == SWP_DEVICE_READ || type == SWP_DEVICE_WRITE;
}

-static inline bool is_write_device_private_entry(swp_entry_t entry)
+static inline bool is_writable_device_private_entry(swp_entry_t entry)
{
return unlikely(swp_type(entry) == SWP_DEVICE_WRITE);
}
#else /* CONFIG_DEVICE_PRIVATE */
-static inline swp_entry_t make_device_private_entry(struct page *page, bool write)
+static inline swp_entry_t make_readable_device_private_entry(pgoff_t offset)
{
return swp_entry(0, 0);
}

-static inline void make_device_private_entry_read(swp_entry_t *entry)
+static inline swp_entry_t make_writable_device_private_entry(pgoff_t offset)
{
+ return swp_entry(0, 0);
}

static inline bool is_device_private_entry(swp_entry_t entry)
@@ -136,35 +136,32 @@ static inline bool is_device_private_entry(swp_entry_t entry)
return false;
}

-static inline bool is_write_device_private_entry(swp_entry_t entry)
+static inline bool is_writable_device_private_entry(swp_entry_t entry)
{
return false;
}
#endif /* CONFIG_DEVICE_PRIVATE */

#ifdef CONFIG_MIGRATION
-static inline swp_entry_t make_migration_entry(struct page *page, int write)
-{
- BUG_ON(!PageLocked(compound_head(page)));
-
- return swp_entry(write ? SWP_MIGRATION_WRITE : SWP_MIGRATION_READ,
- page_to_pfn(page));
-}
-
static inline int is_migration_entry(swp_entry_t entry)
{
return unlikely(swp_type(entry) == SWP_MIGRATION_READ ||
swp_type(entry) == SWP_MIGRATION_WRITE);
}

-static inline int is_write_migration_entry(swp_entry_t entry)
+static inline int is_writable_migration_entry(swp_entry_t entry)
{
return unlikely(swp_type(entry) == SWP_MIGRATION_WRITE);
}

-static inline void make_migration_entry_read(swp_entry_t *entry)
+static inline swp_entry_t make_readable_migration_entry(pgoff_t offset)
{
- *entry = swp_entry(SWP_MIGRATION_READ, swp_offset(*entry));
+ return swp_entry(SWP_MIGRATION_READ, offset);
+}
+
+static inline swp_entry_t make_writable_migration_entry(pgoff_t offset)
+{
+ return swp_entry(SWP_MIGRATION_WRITE, offset);
}

extern void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
@@ -174,21 +171,28 @@ extern void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
extern void migration_entry_wait_huge(struct vm_area_struct *vma,
struct mm_struct *mm, pte_t *pte);
#else
+static inline swp_entry_t make_readable_migration_entry(pgoff_t offset)
+{
+ return swp_entry(0, 0);
+}
+
+static inline swp_entry_t make_writable_migration_entry(pgoff_t offset)
+{
+ return swp_entry(0, 0);
+}

-#define make_migration_entry(page, write) swp_entry(0, 0)
static inline int is_migration_entry(swp_entry_t swp)
{
return 0;
}

-static inline void make_migration_entry_read(swp_entry_t *entryp) { }
static inline void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep,
spinlock_t *ptl) { }
static inline void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd,
unsigned long address) { }
static inline void migration_entry_wait_huge(struct vm_area_struct *vma,
struct mm_struct *mm, pte_t *pte) { }
-static inline int is_write_migration_entry(swp_entry_t entry)
+static inline int is_writable_migration_entry(swp_entry_t entry)
{
return 0;
}
diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
index 05efe98a9ac2..1dcc441da377 100644
--- a/mm/debug_vm_pgtable.c
+++ b/mm/debug_vm_pgtable.c
@@ -817,17 +817,17 @@ static void __init swap_migration_tests(void)
* locked, otherwise it stumbles upon a BUG_ON().
*/
__SetPageLocked(page);
- swp = make_migration_entry(page, 1);
+ swp = make_writable_migration_entry(page_to_pfn(page));
WARN_ON(!is_migration_entry(swp));
- WARN_ON(!is_write_migration_entry(swp));
+ WARN_ON(!is_writable_migration_entry(swp));

- make_migration_entry_read(&swp);
+ swp = make_readable_migration_entry(swp_offset(swp));
WARN_ON(!is_migration_entry(swp));
- WARN_ON(is_write_migration_entry(swp));
+ WARN_ON(is_writable_migration_entry(swp));

- swp = make_migration_entry(page, 0);
+ swp = make_readable_migration_entry(page_to_pfn(page));
WARN_ON(!is_migration_entry(swp));
- WARN_ON(is_write_migration_entry(swp));
+ WARN_ON(is_writable_migration_entry(swp));
__ClearPageLocked(page);
__free_page(page);
}
diff --git a/mm/hmm.c b/mm/hmm.c
index 3b2dda71d0ed..11df3ca30b82 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -255,7 +255,7 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
*/
if (hmm_is_device_private_entry(range, entry)) {
cpu_flags = HMM_PFN_VALID;
- if (is_write_device_private_entry(entry))
+ if (is_writable_device_private_entry(entry))
cpu_flags |= HMM_PFN_WRITE;
*hmm_pfn = swp_offset(entry) | cpu_flags;
return 0;
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 7137ab31766a..2ec6dab72217 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1050,8 +1050,9 @@ int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
swp_entry_t entry = pmd_to_swp_entry(pmd);

VM_BUG_ON(!is_pmd_migration_entry(pmd));
- if (is_write_migration_entry(entry)) {
- make_migration_entry_read(&entry);
+ if (is_writable_migration_entry(entry)) {
+ entry = make_readable_migration_entry(
+ swp_offset(entry));
pmd = swp_entry_to_pmd(entry);
if (pmd_swp_soft_dirty(*src_pmd))
pmd = pmd_swp_mksoft_dirty(pmd);
@@ -1819,13 +1820,14 @@ int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
swp_entry_t entry = pmd_to_swp_entry(*pmd);

VM_BUG_ON(!is_pmd_migration_entry(*pmd));
- if (is_write_migration_entry(entry)) {
+ if (is_writable_migration_entry(entry)) {
pmd_t newpmd;
/*
* A protection check is difficult so
* just be safe and disable write
*/
- make_migration_entry_read(&entry);
+ entry = make_readable_migration_entry(
+ swp_offset(entry));
newpmd = swp_entry_to_pmd(entry);
if (pmd_swp_soft_dirty(*pmd))
newpmd = pmd_swp_mksoft_dirty(newpmd);
@@ -2103,7 +2105,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,

entry = pmd_to_swp_entry(old_pmd);
page = pfn_swap_entry_to_page(entry);
- write = is_write_migration_entry(entry);
+ write = is_writable_migration_entry(entry);
young = false;
soft_dirty = pmd_swp_soft_dirty(old_pmd);
uffd_wp = pmd_swp_uffd_wp(old_pmd);
@@ -2135,7 +2137,12 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
*/
if (freeze || pmd_migration) {
swp_entry_t swp_entry;
- swp_entry = make_migration_entry(page + i, write);
+ if (write)
+ swp_entry = make_writable_migration_entry(
+ page_to_pfn(page + i));
+ else
+ swp_entry = make_readable_migration_entry(
+ page_to_pfn(page + i));
entry = swp_entry_to_pte(swp_entry);
if (soft_dirty)
entry = pte_swp_mksoft_dirty(entry);
@@ -3212,7 +3219,10 @@ void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
if (pmd_dirty(pmdval))
set_page_dirty(page);
- entry = make_migration_entry(page, pmd_write(pmdval));
+ if (pmd_write(pmdval))
+ entry = make_writable_migration_entry(page_to_pfn(page));
+ else
+ entry = make_readable_migration_entry(page_to_pfn(page));
pmdswp = swp_entry_to_pmd(entry);
if (pmd_soft_dirty(pmdval))
pmdswp = pmd_swp_mksoft_dirty(pmdswp);
@@ -3238,7 +3248,7 @@ void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
if (pmd_swp_soft_dirty(*pvmw->pmd))
pmde = pmd_mksoft_dirty(pmde);
- if (is_write_migration_entry(entry))
+ if (is_writable_migration_entry(entry))
pmde = maybe_pmd_mkwrite(pmde, vma);

flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 95918f410c0f..5e6ee9c286c0 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3989,12 +3989,13 @@ int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
is_hugetlb_entry_hwpoisoned(entry))) {
swp_entry_t swp_entry = pte_to_swp_entry(entry);

- if (is_write_migration_entry(swp_entry) && cow) {
+ if (is_writable_migration_entry(swp_entry) && cow) {
/*
* COW mappings require pages in both
* parent and child to be set to read.
*/
- make_migration_entry_read(&swp_entry);
+ swp_entry = make_readable_migration_entry(
+ swp_offset(swp_entry));
entry = swp_entry_to_pte(swp_entry);
set_huge_swap_pte_at(src, addr, src_pte,
entry, sz);
@@ -5237,10 +5238,11 @@ unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
if (unlikely(is_hugetlb_entry_migration(pte))) {
swp_entry_t entry = pte_to_swp_entry(pte);

- if (is_write_migration_entry(entry)) {
+ if (is_writable_migration_entry(entry)) {
pte_t newpte;

- make_migration_entry_read(&entry);
+ entry = make_readable_migration_entry(
+ swp_offset(entry));
newpte = swp_entry_to_pte(entry);
set_huge_swap_pte_at(mm, address, ptep,
newpte, huge_page_size(h));
diff --git a/mm/memory.c b/mm/memory.c
index 1f5c3f6134fb..2fb455c365c2 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -734,13 +734,14 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,

rss[mm_counter(page)]++;

- if (is_write_migration_entry(entry) &&
+ if (is_writable_migration_entry(entry) &&
is_cow_mapping(vm_flags)) {
/*
* COW mappings require pages in both
* parent and child to be set to read.
*/
- make_migration_entry_read(&entry);
+ entry = make_readable_migration_entry(
+ swp_offset(entry));
pte = swp_entry_to_pte(entry);
if (pte_swp_soft_dirty(*src_pte))
pte = pte_swp_mksoft_dirty(pte);
@@ -771,9 +772,10 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
* when a device driver is involved (you cannot easily
* save and restore device driver state).
*/
- if (is_write_device_private_entry(entry) &&
+ if (is_writable_device_private_entry(entry) &&
is_cow_mapping(vm_flags)) {
- make_device_private_entry_read(&entry);
+ entry = make_readable_device_private_entry(
+ swp_offset(entry));
pte = swp_entry_to_pte(entry);
if (pte_swp_uffd_wp(*src_pte))
pte = pte_swp_mkuffd_wp(pte);
diff --git a/mm/migrate.c b/mm/migrate.c
index 749321ae3026..930de919b1f2 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -210,13 +210,18 @@ static bool remove_migration_pte(struct page *page, struct vm_area_struct *vma,
* Recheck VMA as permissions can change since migration started
*/
entry = pte_to_swp_entry(*pvmw.pte);
- if (is_write_migration_entry(entry))
+ if (is_writable_migration_entry(entry))
pte = maybe_mkwrite(pte, vma);
else if (pte_swp_uffd_wp(*pvmw.pte))
pte = pte_mkuffd_wp(pte);

if (unlikely(is_device_private_page(new))) {
- entry = make_device_private_entry(new, pte_write(pte));
+ if (pte_write(pte))
+ entry = make_writable_device_private_entry(
+ page_to_pfn(new));
+ else
+ entry = make_readable_device_private_entry(
+ page_to_pfn(new));
pte = swp_entry_to_pte(entry);
if (pte_swp_soft_dirty(*pvmw.pte))
pte = pte_swp_mksoft_dirty(pte);
@@ -2407,7 +2412,7 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,

mpfn = migrate_pfn(page_to_pfn(page)) |
MIGRATE_PFN_MIGRATE;
- if (is_write_device_private_entry(entry))
+ if (is_writable_device_private_entry(entry))
mpfn |= MIGRATE_PFN_WRITE;
} else {
if (!(migrate->flags & MIGRATE_VMA_SELECT_SYSTEM))
@@ -2453,8 +2458,12 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
ptep_get_and_clear(mm, addr, ptep);

/* Setup special migration page table entry */
- entry = make_migration_entry(page, mpfn &
- MIGRATE_PFN_WRITE);
+ if (mpfn & MIGRATE_PFN_WRITE)
+ entry = make_writable_migration_entry(
+ page_to_pfn(page));
+ else
+ entry = make_readable_migration_entry(
+ page_to_pfn(page));
swp_pte = swp_entry_to_pte(entry);
if (pte_present(pte)) {
if (pte_soft_dirty(pte))
@@ -2927,7 +2936,12 @@ static void migrate_vma_insert_page(struct migrate_vma *migrate,
if (is_device_private_page(page)) {
swp_entry_t swp_entry;

- swp_entry = make_device_private_entry(page, vma->vm_flags & VM_WRITE);
+ if (vma->vm_flags & VM_WRITE)
+ swp_entry = make_writable_device_private_entry(
+ page_to_pfn(page));
+ else
+ swp_entry = make_readable_device_private_entry(
+ page_to_pfn(page));
entry = swp_entry_to_pte(swp_entry);
} else {
/*
diff --git a/mm/mprotect.c b/mm/mprotect.c
index e7a443157988..ee5961888e70 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -143,23 +143,25 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
swp_entry_t entry = pte_to_swp_entry(oldpte);
pte_t newpte;

- if (is_write_migration_entry(entry)) {
+ if (is_writable_migration_entry(entry)) {
/*
* A protection check is difficult so
* just be safe and disable write
*/
- make_migration_entry_read(&entry);
+ entry = make_readable_migration_entry(
+ swp_offset(entry));
newpte = swp_entry_to_pte(entry);
if (pte_swp_soft_dirty(oldpte))
newpte = pte_swp_mksoft_dirty(newpte);
if (pte_swp_uffd_wp(oldpte))
newpte = pte_swp_mkuffd_wp(newpte);
- } else if (is_write_device_private_entry(entry)) {
+ } else if (is_writable_device_private_entry(entry)) {
/*
* We do not preserve soft-dirtiness. See
* copy_one_pte() for explanation.
*/
- make_device_private_entry_read(&entry);
+ entry = make_readable_device_private_entry(
+ swp_offset(entry));
newpte = swp_entry_to_pte(entry);
if (pte_swp_uffd_wp(oldpte))
newpte = pte_swp_mkuffd_wp(newpte);
diff --git a/mm/rmap.c b/mm/rmap.c
index 693a610e181d..bc08c4d4b58a 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1526,7 +1526,7 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
* pte. do_swap_page() will wait until the migration
* pte is removed and then restart fault handling.
*/
- entry = make_migration_entry(page, 0);
+ entry = make_readable_migration_entry(page_to_pfn(page));
swp_pte = swp_entry_to_pte(entry);

/*
@@ -1622,8 +1622,12 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
* pte. do_swap_page() will wait until the migration
* pte is removed and then restart fault handling.
*/
- entry = make_migration_entry(subpage,
- pte_write(pteval));
+ if (pte_write(pteval))
+ entry = make_writable_migration_entry(
+ page_to_pfn(subpage));
+ else
+ entry = make_readable_migration_entry(
+ page_to_pfn(subpage));
swp_pte = swp_entry_to_pte(entry);
if (pte_soft_dirty(pteval))
swp_pte = pte_swp_mksoft_dirty(swp_pte);
--
2.20.1

2021-05-24 13:30:33

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

The behaviour of try_to_unmap_one() is difficult to follow because it
performs different operations based on a fairly large set of flags used
in different combinations.

TTU_MUNLOCK is one such flag. However it is exclusively used by
try_to_munlock() which specifies no other flags. Therefore rather than
overload try_to_unmap_one() with unrelated behaviour split this out into
it's own function and remove the flag.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Ralph Campbell <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>

---

v9:
* Improved comments

v8:
* Renamed try_to_munlock to page_mlock to better reflect what the
function actually does.
* Removed the TODO from the documentation that this patch addresses.

v7:
* Added Christoph's Reviewed-by

v4:
* Removed redundant check for VM_LOCKED
---
Documentation/vm/unevictable-lru.rst | 33 ++++++---------
include/linux/rmap.h | 3 +-
mm/mlock.c | 10 ++---
mm/rmap.c | 61 ++++++++++++++++++++--------
4 files changed, 63 insertions(+), 44 deletions(-)

diff --git a/Documentation/vm/unevictable-lru.rst b/Documentation/vm/unevictable-lru.rst
index 0e1490524f53..eae3af17f2d9 100644
--- a/Documentation/vm/unevictable-lru.rst
+++ b/Documentation/vm/unevictable-lru.rst
@@ -389,14 +389,14 @@ mlocked, munlock_vma_page() updates that zone statistics for the number of
mlocked pages. Note, however, that at this point we haven't checked whether
the page is mapped by other VM_LOCKED VMAs.

-We can't call try_to_munlock(), the function that walks the reverse map to
+We can't call page_mlock(), the function that walks the reverse map to
check for other VM_LOCKED VMAs, without first isolating the page from the LRU.
-try_to_munlock() is a variant of try_to_unmap() and thus requires that the page
+page_mlock() is a variant of try_to_unmap() and thus requires that the page
not be on an LRU list [more on these below]. However, the call to
-isolate_lru_page() could fail, in which case we couldn't try_to_munlock(). So,
+isolate_lru_page() could fail, in which case we can't call page_mlock(). So,
we go ahead and clear PG_mlocked up front, as this might be the only chance we
-have. If we can successfully isolate the page, we go ahead and
-try_to_munlock(), which will restore the PG_mlocked flag and update the zone
+have. If we can successfully isolate the page, we go ahead and call
+page_mlock(), which will restore the PG_mlocked flag and update the zone
page statistics if it finds another VMA holding the page mlocked. If we fail
to isolate the page, we'll have left a potentially mlocked page on the LRU.
This is fine, because we'll catch it later if and if vmscan tries to reclaim
@@ -545,31 +545,24 @@ munlock or munmap system calls, mm teardown (munlock_vma_pages_all), reclaim,
holepunching, and truncation of file pages and their anonymous COWed pages.


-try_to_munlock() Reverse Map Scan
+page_mlock() Reverse Map Scan
---------------------------------

-.. warning::
- [!] TODO/FIXME: a better name might be page_mlocked() - analogous to the
- page_referenced() reverse map walker.
-
When munlock_vma_page() [see section :ref:`munlock()/munlockall() System Call
Handling <munlock_munlockall_handling>` above] tries to munlock a
page, it needs to determine whether or not the page is mapped by any
VM_LOCKED VMA without actually attempting to unmap all PTEs from the
page. For this purpose, the unevictable/mlock infrastructure
-introduced a variant of try_to_unmap() called try_to_munlock().
+introduced a variant of try_to_unmap() called page_mlock().

-try_to_munlock() calls the same functions as try_to_unmap() for anonymous and
-mapped file and KSM pages with a flag argument specifying unlock versus unmap
-processing. Again, these functions walk the respective reverse maps looking
-for VM_LOCKED VMAs. When such a VMA is found, as in the try_to_unmap() case,
-the functions mlock the page via mlock_vma_page() and return SWAP_MLOCK. This
-undoes the pre-clearing of the page's PG_mlocked done by munlock_vma_page.
+page_mlock() walks the respective reverse maps looking for VM_LOCKED VMAs. When
+such a VMA is found the page is mlocked via mlock_vma_page(). This undoes the
+pre-clearing of the page's PG_mlocked done by munlock_vma_page.

-Note that try_to_munlock()'s reverse map walk must visit every VMA in a page's
+Note that page_mlock()'s reverse map walk must visit every VMA in a page's
reverse map to determine that a page is NOT mapped into any VM_LOCKED VMA.
However, the scan can terminate when it encounters a VM_LOCKED VMA.
-Although try_to_munlock() might be called a great many times when munlocking a
+Although page_mlock() might be called a great many times when munlocking a
large region or tearing down a large address space that has been mlocked via
mlockall(), overall this is a fairly rare event.

@@ -602,7 +595,7 @@ inactive lists to the appropriate node's unevictable list.
shrink_inactive_list() should only see SHM_LOCK'd pages that became SHM_LOCK'd
after shrink_active_list() had moved them to the inactive list, or pages mapped
into VM_LOCKED VMAs that munlock_vma_page() couldn't isolate from the LRU to
-recheck via try_to_munlock(). shrink_inactive_list() won't notice the latter,
+recheck via page_mlock(). shrink_inactive_list() won't notice the latter,
but will pass on to shrink_page_list().

shrink_page_list() again culls obviously unevictable pages that it could
diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index def5c62c93b3..38a746787c2f 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -87,7 +87,6 @@ struct anon_vma_chain {

enum ttu_flags {
TTU_MIGRATION = 0x1, /* migration mode */
- TTU_MUNLOCK = 0x2, /* munlock mode */

TTU_SPLIT_HUGE_PMD = 0x4, /* split huge PMD if any */
TTU_IGNORE_MLOCK = 0x8, /* ignore mlock */
@@ -239,7 +238,7 @@ int page_mkclean(struct page *);
* called in munlock()/munmap() path to check for other vmas holding
* the page mlocked.
*/
-void try_to_munlock(struct page *);
+void page_mlock(struct page *page);

void remove_migration_ptes(struct page *old, struct page *new, bool locked);

diff --git a/mm/mlock.c b/mm/mlock.c
index df590fda5688..a518d4c48e65 100644
--- a/mm/mlock.c
+++ b/mm/mlock.c
@@ -108,7 +108,7 @@ void mlock_vma_page(struct page *page)
/*
* Finish munlock after successful page isolation
*
- * Page must be locked. This is a wrapper for try_to_munlock()
+ * Page must be locked. This is a wrapper for page_mlock()
* and putback_lru_page() with munlock accounting.
*/
static void __munlock_isolated_page(struct page *page)
@@ -118,7 +118,7 @@ static void __munlock_isolated_page(struct page *page)
* and we don't need to check all the other vmas.
*/
if (page_mapcount(page) > 1)
- try_to_munlock(page);
+ page_mlock(page);

/* Did try_to_unlock() succeed or punt? */
if (!PageMlocked(page))
@@ -158,7 +158,7 @@ static void __munlock_isolation_failed(struct page *page)
* munlock()ed or munmap()ed, we want to check whether other vmas hold the
* page locked so that we can leave it on the unevictable lru list and not
* bother vmscan with it. However, to walk the page's rmap list in
- * try_to_munlock() we must isolate the page from the LRU. If some other
+ * page_mlock() we must isolate the page from the LRU. If some other
* task has removed the page from the LRU, we won't be able to do that.
* So we clear the PageMlocked as we might not get another chance. If we
* can't isolate the page, we leave it for putback_lru_page() and vmscan
@@ -168,7 +168,7 @@ unsigned int munlock_vma_page(struct page *page)
{
int nr_pages;

- /* For try_to_munlock() and to serialize with page migration */
+ /* For page_mlock() and to serialize with page migration */
BUG_ON(!PageLocked(page));
VM_BUG_ON_PAGE(PageTail(page), page);

@@ -205,7 +205,7 @@ static int __mlock_posix_error_return(long retval)
*
* The fast path is available only for evictable pages with single mapping.
* Then we can bypass the per-cpu pvec and get better performance.
- * when mapcount > 1 we need try_to_munlock() which can fail.
+ * when mapcount > 1 we need page_mlock() which can fail.
* when !page_evictable(), we need the full redo logic of putback_lru_page to
* avoid leaving evictable page in unevictable list.
*
diff --git a/mm/rmap.c b/mm/rmap.c
index bc08c4d4b58a..e88966903e1e 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1405,10 +1405,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
struct mmu_notifier_range range;
enum ttu_flags flags = (enum ttu_flags)(long)arg;

- /* munlock has nothing to gain from examining un-locked vmas */
- if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
- return true;
-
if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
is_zone_device_page(page) && !is_device_private_page(page))
return true;
@@ -1469,8 +1465,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
page_vma_mapped_walk_done(&pvmw);
break;
}
- if (flags & TTU_MUNLOCK)
- continue;
}

/* Unexpected PMD-mapped THP? */
@@ -1784,20 +1778,53 @@ bool try_to_unmap(struct page *page, enum ttu_flags flags)
return !page_mapcount(page) ? true : false;
}

+/*
+ * Walks the vma's mapping a page and mlocks the page if any locked vma's are
+ * found. Once one is found the page is locked and the scan can be terminated.
+ */
+static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
+ unsigned long address, void *unused)
+{
+ struct page_vma_mapped_walk pvmw = {
+ .page = page,
+ .vma = vma,
+ .address = address,
+ };
+
+ /* An un-locked vma doesn't have any pages to lock, continue the scan */
+ if (!(vma->vm_flags & VM_LOCKED))
+ return true;
+
+ while (page_vma_mapped_walk(&pvmw)) {
+ /* PTE-mapped THP are never mlocked */
+ if (!PageTransCompound(page))
+ mlock_vma_page(page);
+ page_vma_mapped_walk_done(&pvmw);
+
+ /*
+ * no need to continue scanning other vma's if the page has
+ * been locked.
+ */
+ return false;
+ }
+
+ return true;
+}
+
/**
- * try_to_munlock - try to munlock a page
- * @page: the page to be munlocked
+ * page_mlock - try to mlock a page
+ * @page: the page to be mlocked
*
- * Called from munlock code. Checks all of the VMAs mapping the page
- * to make sure nobody else has this page mlocked. The page will be
- * returned with PG_mlocked cleared if no other vmas have it mlocked.
+ * Called from munlock code. Checks all of the VMAs mapping the page and mlocks
+ * the page if any are found. The page will be returned with PG_mlocked cleared
+ * if it is not mapped by any locked vmas.
+ *
+ * mmap_lock should be held for read or write.
*/
-
-void try_to_munlock(struct page *page)
+void page_mlock(struct page *page)
{
struct rmap_walk_control rwc = {
- .rmap_one = try_to_unmap_one,
- .arg = (void *)TTU_MUNLOCK,
+ .rmap_one = page_mlock_one,
.done = page_not_mapped,
.anon_lock = page_lock_anon_vma_read,

@@ -1849,7 +1876,7 @@ static struct anon_vma *rmap_walk_anon_lock(struct page *page,
* Find all the mappings of a page using the mapping pointer and the vma chains
* contained in the anon_vma struct it points to.
*
- * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
+ * When called from page_mlock(), the mmap_lock of the mm containing the vma
* where the page was found will be held for write. So, we won't recheck
* vm_flags for that VMA. That should be OK, because that vma shouldn't be
* LOCKED.
@@ -1901,7 +1928,7 @@ static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
* Find all the mappings of a page using the mapping pointer and the vma chains
* contained in the address_space struct it points to.
*
- * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
+ * When called from page_mlock(), the mmap_lock of the mm containing the vma
* where the page was found will be held for write. So, we won't recheck
* vm_flags for that VMA. That should be OK, because that vma shouldn't be
* LOCKED.
--
2.20.1

2021-05-24 13:30:46

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 05/10] mm: Rename migrate_pgmap_owner

MMU notifier ranges have a migrate_pgmap_owner field which is used by
drivers to store a pointer. This is subsequently used by the driver
callback to filter MMU_NOTIFY_MIGRATE events. Other notifier event types
can also benefit from this filtering, so rename the
'migrate_pgmap_owner' field to 'owner' and create a new notifier
initialisation function to initialise this field.

Signed-off-by: Alistair Popple <[email protected]>
Suggested-by: Peter Xu <[email protected]>

---

v9:

Previously part of the next patch in the series ('mm: Device exclusive
memory access') but now split out as a separate change as suggested by
Peter Xu.
---
Documentation/vm/hmm.rst | 2 +-
drivers/gpu/drm/nouveau/nouveau_svm.c | 2 +-
include/linux/mmu_notifier.h | 20 ++++++++++----------
lib/test_hmm.c | 2 +-
mm/migrate.c | 10 +++++-----
5 files changed, 18 insertions(+), 18 deletions(-)

diff --git a/Documentation/vm/hmm.rst b/Documentation/vm/hmm.rst
index 09e28507f5b2..3df79307a797 100644
--- a/Documentation/vm/hmm.rst
+++ b/Documentation/vm/hmm.rst
@@ -332,7 +332,7 @@ between device driver specific code and shared common code:
walks to fill in the ``args->src`` array with PFNs to be migrated.
The ``invalidate_range_start()`` callback is passed a
``struct mmu_notifier_range`` with the ``event`` field set to
- ``MMU_NOTIFY_MIGRATE`` and the ``migrate_pgmap_owner`` field set to
+ ``MMU_NOTIFY_MIGRATE`` and the ``owner`` field set to
the ``args->pgmap_owner`` field passed to migrate_vma_setup(). This is
allows the device driver to skip the invalidation callback and only
invalidate device private MMU mappings that are actually migrating.
diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c
index f18bd53da052..94f841026c3b 100644
--- a/drivers/gpu/drm/nouveau/nouveau_svm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_svm.c
@@ -265,7 +265,7 @@ nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
* the invalidation is handled as part of the migration process.
*/
if (update->event == MMU_NOTIFY_MIGRATE &&
- update->migrate_pgmap_owner == svmm->vmm->cli->drm->dev)
+ update->owner == svmm->vmm->cli->drm->dev)
goto out;

if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
diff --git a/include/linux/mmu_notifier.h b/include/linux/mmu_notifier.h
index 1a6a9eb6d3fa..8e428eb813b8 100644
--- a/include/linux/mmu_notifier.h
+++ b/include/linux/mmu_notifier.h
@@ -41,7 +41,7 @@ struct mmu_interval_notifier;
*
* @MMU_NOTIFY_MIGRATE: used during migrate_vma_collect() invalidate to signal
* a device driver to possibly ignore the invalidation if the
- * migrate_pgmap_owner field matches the driver's device private pgmap owner.
+ * owner field matches the driver's device private pgmap owner.
*/
enum mmu_notifier_event {
MMU_NOTIFY_UNMAP = 0,
@@ -269,7 +269,7 @@ struct mmu_notifier_range {
unsigned long end;
unsigned flags;
enum mmu_notifier_event event;
- void *migrate_pgmap_owner;
+ void *owner;
};

static inline int mm_has_notifiers(struct mm_struct *mm)
@@ -521,14 +521,14 @@ static inline void mmu_notifier_range_init(struct mmu_notifier_range *range,
range->flags = flags;
}

-static inline void mmu_notifier_range_init_migrate(
- struct mmu_notifier_range *range, unsigned int flags,
+static inline void mmu_notifier_range_init_owner(
+ struct mmu_notifier_range *range,
+ enum mmu_notifier_event event, unsigned int flags,
struct vm_area_struct *vma, struct mm_struct *mm,
- unsigned long start, unsigned long end, void *pgmap)
+ unsigned long start, unsigned long end, void *owner)
{
- mmu_notifier_range_init(range, MMU_NOTIFY_MIGRATE, flags, vma, mm,
- start, end);
- range->migrate_pgmap_owner = pgmap;
+ mmu_notifier_range_init(range, event, flags, vma, mm, start, end);
+ range->owner = owner;
}

#define ptep_clear_flush_young_notify(__vma, __address, __ptep) \
@@ -655,8 +655,8 @@ static inline void _mmu_notifier_range_init(struct mmu_notifier_range *range,

#define mmu_notifier_range_init(range,event,flags,vma,mm,start,end) \
_mmu_notifier_range_init(range, start, end)
-#define mmu_notifier_range_init_migrate(range, flags, vma, mm, start, end, \
- pgmap) \
+#define mmu_notifier_range_init_owner(range, event, flags, vma, mm, start, \
+ end, owner) \
_mmu_notifier_range_init(range, start, end)

static inline bool
diff --git a/lib/test_hmm.c b/lib/test_hmm.c
index 80a78877bd93..5c9f5a020c1d 100644
--- a/lib/test_hmm.c
+++ b/lib/test_hmm.c
@@ -218,7 +218,7 @@ static bool dmirror_interval_invalidate(struct mmu_interval_notifier *mni,
* the invalidation is handled as part of the migration process.
*/
if (range->event == MMU_NOTIFY_MIGRATE &&
- range->migrate_pgmap_owner == dmirror->mdevice)
+ range->owner == dmirror->mdevice)
return true;

if (mmu_notifier_range_blockable(range))
diff --git a/mm/migrate.c b/mm/migrate.c
index 05740f816bc4..e5429a44e7a3 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -2526,8 +2526,8 @@ static void migrate_vma_collect(struct migrate_vma *migrate)
* that the registered device driver can skip invalidating device
* private page mappings that won't be migrated.
*/
- mmu_notifier_range_init_migrate(&range, 0, migrate->vma,
- migrate->vma->vm_mm, migrate->start, migrate->end,
+ mmu_notifier_range_init_owner(&range, MMU_NOTIFY_MIGRATE, 0,
+ migrate->vma, migrate->vma->vm_mm, migrate->start, migrate->end,
migrate->pgmap_owner);
mmu_notifier_invalidate_range_start(&range);

@@ -3037,9 +3037,9 @@ void migrate_vma_pages(struct migrate_vma *migrate)
if (!notified) {
notified = true;

- mmu_notifier_range_init_migrate(&range, 0,
- migrate->vma, migrate->vma->vm_mm,
- addr, migrate->end,
+ mmu_notifier_range_init_owner(&range,
+ MMU_NOTIFY_MIGRATE, 0, migrate->vma,
+ migrate->vma->vm_mm, addr, migrate->end,
migrate->pgmap_owner);
mmu_notifier_invalidate_range_start(&range);
}
--
2.20.1

2021-05-24 13:31:12

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 07/10] mm: Device exclusive memory access

Some devices require exclusive write access to shared virtual
memory (SVM) ranges to perform atomic operations on that memory. This
requires CPU page tables to be updated to deny access whilst atomic
operations are occurring.

In order to do this introduce a new swap entry
type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for
exclusive access by a device all page table mappings for the particular
range are replaced with device exclusive swap entries. This causes any
CPU access to the page to result in a fault.

Faults are resovled by replacing the faulting entry with the original
mapping. This results in MMU notifiers being called which a driver uses
to update access permissions such as revoking atomic access. After
notifiers have been called the device will no longer have exclusive
access to the region.

Walking of the page tables to find the target pages is handled by
get_user_pages() rather than a direct page table walk. A direct page
table walk similar to what migrate_vma_collect()/unmap() does could also
have been utilised. However this resulted in more code similar in
functionality to what get_user_pages() provides as page faulting is
required to make the PTEs present and to break COW.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>

---

v9:
* Split rename of migrate_pgmap_owner into a separate patch.
* Added comments explaining SWP_DEVICE_EXCLUSIVE_* entries.
* Renamed try_to_protect{_one} to page_make_device_exclusive{_one} based
somewhat on a suggestion from Peter Xu. I was never particularly happy
with try_to_protect() as a name so think this is better.
* Removed unneccesary code and reworded some comments based on feedback
from Peter Xu.
* Removed the VMA walk when restoring PTEs for device-exclusive entries.
* Simplified implementation of copy_pte_range() to fail if the page
cannot be locked. This might lead to occasional fork() failures but at
this stage we don't think that will be an issue.

v8:
* Remove device exclusive entries on fork rather than copy them.

v7:
* Added Christoph's Reviewed-by.
* Minor cosmetic cleanups suggested by Christoph.
* Replace mmu_notifier_range_init_migrate/exclusive with
mmu_notifier_range_init_owner as suggested by Christoph.
* Replaced lock_page() with lock_page_retry() when handling faults.
* Restrict to anonymous pages for now.

v6:
* Fixed a bisectablity issue due to incorrectly applying the rename of
migrate_pgmap_owner to the wrong patches for Nouveau and hmm_test.

v5:
* Renamed range->migrate_pgmap_owner to range->owner.
* Added MMU_NOTIFY_EXCLUSIVE to allow passing of a driver cookie which
allows notifiers called as a result of make_device_exclusive_range() to
be ignored.
* Added a check to try_to_protect_one() to detect if the pages originally
returned from get_user_pages() have been unmapped or not.
* Removed check_device_exclusive_range() as it is no longer required with
the other changes.
* Documentation update.

v4:
* Add function to check that mappings are still valid and exclusive.
* s/long/unsigned long/ in make_device_exclusive_entry().
---
Documentation/vm/hmm.rst | 17 ++++
include/linux/mmu_notifier.h | 6 ++
include/linux/rmap.h | 4 +
include/linux/swap.h | 7 +-
include/linux/swapops.h | 44 ++++++++-
mm/hmm.c | 5 +
mm/memory.c | 128 +++++++++++++++++++++++-
mm/mprotect.c | 8 ++
mm/page_vma_mapped.c | 9 +-
mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
10 files changed, 405 insertions(+), 9 deletions(-)

diff --git a/Documentation/vm/hmm.rst b/Documentation/vm/hmm.rst
index 3df79307a797..a14c2938e7af 100644
--- a/Documentation/vm/hmm.rst
+++ b/Documentation/vm/hmm.rst
@@ -405,6 +405,23 @@ between device driver specific code and shared common code:

The lock can now be released.

+Exclusive access memory
+=======================
+
+Some devices have features such as atomic PTE bits that can be used to implement
+atomic access to system memory. To support atomic operations to a shared virtual
+memory page such a device needs access to that page which is exclusive of any
+userspace access from the CPU. The ``make_device_exclusive_range()`` function
+can be used to make a memory range inaccessible from userspace.
+
+This replaces all mappings for pages in the given range with special swap
+entries. Any attempt to access the swap entry results in a fault which is
+resovled by replacing the entry with the original mapping. A driver gets
+notified that the mapping has been changed by MMU notifiers, after which point
+it will no longer have exclusive access to the page. Exclusive access is
+guranteed to last until the driver drops the page lock and page reference, at
+which point any CPU faults on the page may proceed as described.
+
Memory cgroup (memcg) and rss accounting
========================================

diff --git a/include/linux/mmu_notifier.h b/include/linux/mmu_notifier.h
index 8e428eb813b8..d049e0f6f756 100644
--- a/include/linux/mmu_notifier.h
+++ b/include/linux/mmu_notifier.h
@@ -42,6 +42,11 @@ struct mmu_interval_notifier;
* @MMU_NOTIFY_MIGRATE: used during migrate_vma_collect() invalidate to signal
* a device driver to possibly ignore the invalidation if the
* owner field matches the driver's device private pgmap owner.
+ *
+ * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
+ * longer have exclusive access to the page. May ignore the invalidation that's
+ * part of make_device_exclusive_range() if the owner field
+ * matches the value passed to make_device_exclusive_range().
*/
enum mmu_notifier_event {
MMU_NOTIFY_UNMAP = 0,
@@ -51,6 +56,7 @@ enum mmu_notifier_event {
MMU_NOTIFY_SOFT_DIRTY,
MMU_NOTIFY_RELEASE,
MMU_NOTIFY_MIGRATE,
+ MMU_NOTIFY_EXCLUSIVE,
};

#define MMU_NOTIFIER_RANGE_BLOCKABLE (1 << 0)
diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index 0e25d829f742..3a1ce4ef9276 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -193,6 +193,10 @@ int page_referenced(struct page *, int is_locked,
bool try_to_migrate(struct page *page, enum ttu_flags flags);
bool try_to_unmap(struct page *, enum ttu_flags flags);

+int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
+ unsigned long end, struct page **pages,
+ void *arg);
+
/* Avoid racy checks */
#define PVMW_SYNC (1 << 0)
/* Look for migarion entries rather than present PTEs */
diff --git a/include/linux/swap.h b/include/linux/swap.h
index a6d4505ecf73..306df39d7c67 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -63,11 +63,16 @@ static inline int current_is_kswapd(void)
*
* When a page is migrated from CPU to device, we set the CPU page table entry
* to a special SWP_DEVICE_* entry.
+ *
+ * When a page is mapped by the device for exclusive access we set the CPU page
+ * table entries to special SWP_DEVICE_EXCLUSIVE_* entries.
*/
#ifdef CONFIG_DEVICE_PRIVATE
-#define SWP_DEVICE_NUM 2
+#define SWP_DEVICE_NUM 4
#define SWP_DEVICE_WRITE (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM)
#define SWP_DEVICE_READ (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+1)
+#define SWP_DEVICE_EXCLUSIVE_WRITE (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+2)
+#define SWP_DEVICE_EXCLUSIVE_READ (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+3)
#else
#define SWP_DEVICE_NUM 0
#endif
diff --git a/include/linux/swapops.h b/include/linux/swapops.h
index 4dfd807ae52a..4129bd2ff9d6 100644
--- a/include/linux/swapops.h
+++ b/include/linux/swapops.h
@@ -120,6 +120,27 @@ static inline bool is_writable_device_private_entry(swp_entry_t entry)
{
return unlikely(swp_type(entry) == SWP_DEVICE_WRITE);
}
+
+static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t offset)
+{
+ return swp_entry(SWP_DEVICE_EXCLUSIVE_READ, offset);
+}
+
+static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t offset)
+{
+ return swp_entry(SWP_DEVICE_EXCLUSIVE_WRITE, offset);
+}
+
+static inline bool is_device_exclusive_entry(swp_entry_t entry)
+{
+ return swp_type(entry) == SWP_DEVICE_EXCLUSIVE_READ ||
+ swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE;
+}
+
+static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
+{
+ return unlikely(swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE);
+}
#else /* CONFIG_DEVICE_PRIVATE */
static inline swp_entry_t make_readable_device_private_entry(pgoff_t offset)
{
@@ -140,6 +161,26 @@ static inline bool is_writable_device_private_entry(swp_entry_t entry)
{
return false;
}
+
+static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t offset)
+{
+ return swp_entry(0, 0);
+}
+
+static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t offset)
+{
+ return swp_entry(0, 0);
+}
+
+static inline bool is_device_exclusive_entry(swp_entry_t entry)
+{
+ return false;
+}
+
+static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
+{
+ return false;
+}
#endif /* CONFIG_DEVICE_PRIVATE */

#ifdef CONFIG_MIGRATION
@@ -219,7 +260,8 @@ static inline struct page *pfn_swap_entry_to_page(swp_entry_t entry)
*/
static inline bool is_pfn_swap_entry(swp_entry_t entry)
{
- return is_migration_entry(entry) || is_device_private_entry(entry);
+ return is_migration_entry(entry) || is_device_private_entry(entry) ||
+ is_device_exclusive_entry(entry);
}

struct page_vma_mapped_walk;
diff --git a/mm/hmm.c b/mm/hmm.c
index 11df3ca30b82..fad6be2bf072 100644
--- a/mm/hmm.c
+++ b/mm/hmm.c
@@ -26,6 +26,8 @@
#include <linux/mmu_notifier.h>
#include <linux/memory_hotplug.h>

+#include "internal.h"
+
struct hmm_vma_walk {
struct hmm_range *range;
unsigned long last;
@@ -271,6 +273,9 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
if (!non_swap_entry(entry))
goto fault;

+ if (is_device_exclusive_entry(entry))
+ goto fault;
+
if (is_migration_entry(entry)) {
pte_unmap(ptep);
hmm_vma_walk->last = addr;
diff --git a/mm/memory.c b/mm/memory.c
index e061cfa18c11..c1d2d732f189 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -700,6 +700,68 @@ struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
}
#endif

+static void restore_exclusive_pte(struct vm_area_struct *vma,
+ struct page *page, unsigned long address,
+ pte_t *ptep)
+{
+ pte_t pte;
+ swp_entry_t entry;
+
+ pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
+ if (pte_swp_soft_dirty(*ptep))
+ pte = pte_mksoft_dirty(pte);
+
+ entry = pte_to_swp_entry(*ptep);
+ if (pte_swp_uffd_wp(*ptep))
+ pte = pte_mkuffd_wp(pte);
+ else if (is_writable_device_exclusive_entry(entry))
+ pte = maybe_mkwrite(pte_mkdirty(pte), vma);
+
+ set_pte_at(vma->vm_mm, address, ptep, pte);
+
+ /*
+ * No need to take a page reference as one was already
+ * created when the swap entry was made.
+ */
+ if (PageAnon(page))
+ page_add_anon_rmap(page, vma, address, false);
+ else
+ /*
+ * Currently device exclusive access only supports anonymous
+ * memory so the entry shouldn't point to a filebacked page.
+ */
+ WARN_ON_ONCE(!PageAnon(page));
+
+ if (vma->vm_flags & VM_LOCKED)
+ mlock_vma_page(page);
+
+ /*
+ * No need to invalidate - it was non-present before. However
+ * secondary CPUs may have mappings that need invalidating.
+ */
+ update_mmu_cache(vma, address, ptep);
+}
+
+/*
+ * Tries to restore an exclusive pte if the page lock can be acquired without
+ * sleeping.
+ */
+static unsigned long
+try_restore_exclusive_pte(struct mm_struct *src_mm, pte_t *src_pte,
+ struct vm_area_struct *vma, unsigned long addr)
+{
+ swp_entry_t entry = pte_to_swp_entry(*src_pte);
+ struct page *page = pfn_swap_entry_to_page(entry);
+
+ if (trylock_page(page)) {
+ restore_exclusive_pte(vma, page, addr, src_pte);
+ unlock_page(page);
+ return 0;
+ }
+
+ return -EBUSY;
+}
+
/*
* copy one vm_area from one task to the other. Assumes the page tables
* already present in the new task to be cleared in the whole range
@@ -781,6 +843,17 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
pte = pte_swp_mkuffd_wp(pte);
set_pte_at(src_mm, addr, src_pte, pte);
}
+ } else if (is_device_exclusive_entry(entry)) {
+ /*
+ * Make device exclusive entries present by restoring the
+ * original entry then copying as for a present pte. Device
+ * exclusive entries currently only support private writable
+ * (ie. COW) mappings.
+ */
+ VM_BUG_ON(!is_cow_mapping(vma->vm_flags));
+ if (try_restore_exclusive_pte(src_mm, src_pte, vma, addr))
+ return -EBUSY;
+ return -ENOENT;
}
set_pte_at(dst_mm, addr, dst_pte, pte);
return 0;
@@ -980,9 +1053,18 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
if (ret == -EAGAIN) {
entry = pte_to_swp_entry(*src_pte);
break;
+ } else if (ret == -EBUSY) {
+ break;
+ } else if (!ret) {
+ progress += 8;
+ continue;
}
- progress += 8;
- continue;
+
+ /*
+ * Device exclusive entry restored, continue by copying
+ * the now present pte.
+ */
+ WARN_ON_ONCE(ret != -ENOENT);
}
/* copy_present_pte() will clear `*prealloc' if consumed */
ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
@@ -1019,6 +1101,8 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
goto out;
}
entry.val = 0;
+ } else if (ret == -EBUSY) {
+ return -EBUSY;
} else if (ret) {
WARN_ON_ONCE(ret != -EAGAIN);
prealloc = page_copy_prealloc(src_mm, src_vma, addr);
@@ -1283,7 +1367,8 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
}

entry = pte_to_swp_entry(ptent);
- if (is_device_private_entry(entry)) {
+ if (is_device_private_entry(entry) ||
+ is_device_exclusive_entry(entry)) {
struct page *page = pfn_swap_entry_to_page(entry);

if (unlikely(details && details->check_mapping)) {
@@ -1299,7 +1384,10 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,

pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
rss[mm_counter(page)]--;
- page_remove_rmap(page, false);
+
+ if (is_device_private_entry(entry))
+ page_remove_rmap(page, false);
+
put_page(page);
continue;
}
@@ -3303,6 +3391,35 @@ void unmap_mapping_range(struct address_space *mapping,
}
EXPORT_SYMBOL(unmap_mapping_range);

+/*
+ * Restore a potential device exclusive pte to a working pte entry
+ */
+static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
+{
+ struct page *page = vmf->page;
+ struct vm_area_struct *vma = vmf->vma;
+ vm_fault_t ret = 0;
+ struct mmu_notifier_range range;
+
+ if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
+ return VM_FAULT_RETRY;
+ mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
+ vmf->address & PAGE_MASK,
+ (vmf->address & PAGE_MASK) + PAGE_SIZE);
+ mmu_notifier_invalidate_range_start(&range);
+
+ vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
+ &vmf->ptl);
+ if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
+ restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
+
+ pte_unmap_unlock(vmf->pte, vmf->ptl);
+ unlock_page(page);
+
+ mmu_notifier_invalidate_range_end(&range);
+ return ret;
+}
+
/*
* We enter with non-exclusive mmap_lock (to exclude vma changes,
* but allow concurrent faults), and pte mapped but not yet locked.
@@ -3330,6 +3447,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
if (is_migration_entry(entry)) {
migration_entry_wait(vma->vm_mm, vmf->pmd,
vmf->address);
+ } else if (is_device_exclusive_entry(entry)) {
+ vmf->page = pfn_swap_entry_to_page(entry);
+ ret = remove_device_exclusive_entry(vmf);
} else if (is_device_private_entry(entry)) {
vmf->page = pfn_swap_entry_to_page(entry);
ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
diff --git a/mm/mprotect.c b/mm/mprotect.c
index ee5961888e70..883e2cc85cad 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -165,6 +165,14 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
newpte = swp_entry_to_pte(entry);
if (pte_swp_uffd_wp(oldpte))
newpte = pte_swp_mkuffd_wp(newpte);
+ } else if (is_writable_device_exclusive_entry(entry)) {
+ entry = make_readable_device_exclusive_entry(
+ swp_offset(entry));
+ newpte = swp_entry_to_pte(entry);
+ if (pte_swp_soft_dirty(oldpte))
+ newpte = pte_swp_mksoft_dirty(newpte);
+ if (pte_swp_uffd_wp(oldpte))
+ newpte = pte_swp_mkuffd_wp(newpte);
} else {
newpte = oldpte;
}
diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
index a6a7febb4d93..f535bcb4950c 100644
--- a/mm/page_vma_mapped.c
+++ b/mm/page_vma_mapped.c
@@ -41,7 +41,8 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw)

/* Handle un-addressable ZONE_DEVICE memory */
entry = pte_to_swp_entry(*pvmw->pte);
- if (!is_device_private_entry(entry))
+ if (!is_device_private_entry(entry) &&
+ !is_device_exclusive_entry(entry))
return false;
} else if (!pte_present(*pvmw->pte))
return false;
@@ -93,7 +94,8 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
return false;
entry = pte_to_swp_entry(*pvmw->pte);

- if (!is_migration_entry(entry))
+ if (!is_migration_entry(entry) &&
+ !is_device_exclusive_entry(entry))
return false;

pfn = swp_offset(entry);
@@ -102,7 +104,8 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)

/* Handle un-addressable ZONE_DEVICE memory */
entry = pte_to_swp_entry(*pvmw->pte);
- if (!is_device_private_entry(entry))
+ if (!is_device_private_entry(entry) &&
+ !is_device_exclusive_entry(entry))
return false;

pfn = swp_offset(entry);
diff --git a/mm/rmap.c b/mm/rmap.c
index 8ed1853060cf..fe062f63ef4d 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -2008,6 +2008,192 @@ void page_mlock(struct page *page)
rmap_walk(page, &rwc);
}

+struct make_exclusive_args {
+ struct mm_struct *mm;
+ unsigned long address;
+ void *owner;
+ bool valid;
+};
+
+static bool page_make_device_exclusive_one(struct page *page,
+ struct vm_area_struct *vma, unsigned long address, void *priv)
+{
+ struct mm_struct *mm = vma->vm_mm;
+ struct page_vma_mapped_walk pvmw = {
+ .page = page,
+ .vma = vma,
+ .address = address,
+ };
+ struct make_exclusive_args *args = priv;
+ pte_t pteval;
+ struct page *subpage;
+ bool ret = true;
+ struct mmu_notifier_range range;
+ swp_entry_t entry;
+ pte_t swp_pte;
+
+ mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
+ vma->vm_mm, address, min(vma->vm_end,
+ address + page_size(page)), args->owner);
+ mmu_notifier_invalidate_range_start(&range);
+
+ while (page_vma_mapped_walk(&pvmw)) {
+ /* Unexpected PMD-mapped THP? */
+ VM_BUG_ON_PAGE(!pvmw.pte, page);
+
+ if (!pte_present(*pvmw.pte)) {
+ ret = false;
+ page_vma_mapped_walk_done(&pvmw);
+ break;
+ }
+
+ subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
+ address = pvmw.address;
+
+ /* Nuke the page table entry. */
+ flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
+ pteval = ptep_clear_flush(vma, address, pvmw.pte);
+
+ /* Move the dirty bit to the page. Now the pte is gone. */
+ if (pte_dirty(pteval))
+ set_page_dirty(page);
+
+ if (arch_unmap_one(mm, vma, address, pteval) < 0) {
+ set_pte_at(mm, address, pvmw.pte, pteval);
+ ret = false;
+ page_vma_mapped_walk_done(&pvmw);
+ break;
+ }
+
+ /*
+ * Check that our target page is still mapped at the expected
+ * address.
+ */
+ if (args->mm == mm && args->address == address &&
+ pte_write(pteval))
+ args->valid = true;
+
+ /*
+ * Store the pfn of the page in a special migration
+ * pte. do_swap_page() will wait until the migration
+ * pte is removed and then restart fault handling.
+ */
+ if (pte_write(pteval))
+ entry = make_writable_device_exclusive_entry(
+ page_to_pfn(subpage));
+ else
+ entry = make_readable_device_exclusive_entry(
+ page_to_pfn(subpage));
+ swp_pte = swp_entry_to_pte(entry);
+ if (pte_soft_dirty(pteval))
+ swp_pte = pte_swp_mksoft_dirty(swp_pte);
+ if (pte_uffd_wp(pteval))
+ swp_pte = pte_swp_mkuffd_wp(swp_pte);
+
+ /* Take a reference for the swap entry */
+ get_page(page);
+ set_pte_at(mm, address, pvmw.pte, swp_pte);
+
+ page_remove_rmap(subpage, PageHuge(page));
+ put_page(page);
+ }
+
+ mmu_notifier_invalidate_range_end(&range);
+
+ return ret;
+}
+
+/**
+ * page_make_device_exclusive - replace page table mappings with swap entries
+ * @page: the page to replace page table entries for
+ * @mm: the mm_struct where the page is expected to be mapped
+ * @address: address where the page is expected to be mapped
+ * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
+ *
+ * Tries to remove all the page table entries which are mapping this page and
+ * replace them with special device exclusive swap entries to grant a device
+ * exclusive access to the page. Caller must hold the page lock.
+ *
+ * Returns false if the page is still mapped, or if it could not be unmapped
+ * from the expected address. Otherwise returns true (success).
+ */
+static bool page_make_device_exclusive(struct page *page, struct mm_struct *mm,
+ unsigned long address, void *owner)
+{
+ struct make_exclusive_args args = {
+ .mm = mm,
+ .address = address,
+ .owner = owner,
+ .valid = false,
+ };
+ struct rmap_walk_control rwc = {
+ .rmap_one = page_make_device_exclusive_one,
+ .done = page_not_mapped,
+ .anon_lock = page_lock_anon_vma_read,
+ .arg = &args,
+ };
+
+ /*
+ * Restrict to anonymous pages for now to avoid potential writeback
+ * issues.
+ */
+ if (!PageAnon(page))
+ return false;
+
+ rmap_walk(page, &rwc);
+
+ return args.valid && !page_mapcount(page);
+}
+
+/**
+ * make_device_exclusive_range() - Mark a range for exclusive use by a device
+ * @mm: mm_struct of assoicated target process
+ * @start: start of the region to mark for exclusive device access
+ * @end: end address of region
+ * @pages: returns the pages which were successfully marked for exclusive access
+ * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
+ *
+ * Returns: number of pages found in the range by GUP. A page is marked for
+ * exclusive access only if the page pointer is non-NULL.
+ *
+ * This function finds ptes mapping page(s) to the given address range, locks
+ * them and replaces mappings with special swap entries preventing userspace CPU
+ * access. On fault these entries are replaced with the original mapping after
+ * calling MMU notifiers.
+ *
+ * A driver using this to program access from a device must use a mmu notifier
+ * critical section to hold a device specific lock during programming. Once
+ * programming is complete it should drop the page lock and reference after
+ * which point CPU access to the page will revoke the exclusive access.
+ */
+int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
+ unsigned long end, struct page **pages,
+ void *owner)
+{
+ unsigned long npages = (end - start) >> PAGE_SHIFT;
+ unsigned long i;
+
+ npages = get_user_pages_remote(mm, start, npages,
+ FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
+ pages, NULL, NULL);
+ for (i = 0; i < npages; i++, start += PAGE_SIZE) {
+ if (!trylock_page(pages[i])) {
+ put_page(pages[i]);
+ pages[i] = NULL;
+ continue;
+ }
+
+ if (!page_make_device_exclusive(pages[i], mm, start, owner)) {
+ unlock_page(pages[i]);
+ put_page(pages[i]);
+ pages[i] = NULL;
+ }
+ }
+
+ return npages;
+}
+EXPORT_SYMBOL_GPL(make_device_exclusive_range);
+
void __put_anon_vma(struct anon_vma *anon_vma)
{
struct anon_vma *root = anon_vma->root;
--
2.20.1

2021-05-24 13:32:22

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 09/10] nouveau/svm: Refactor nouveau_range_fault

Call mmu_interval_notifier_insert() as part of nouveau_range_fault().
This doesn't introduce any functional change but makes it easier for a
subsequent patch to alter the behaviour of nouveau_range_fault() to
support GPU atomic operations.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Ben Skeggs <[email protected]>

---

v9:

* Added Ben's Reviewed-By (thanks!)
---
drivers/gpu/drm/nouveau/nouveau_svm.c | 34 ++++++++++++++++-----------
1 file changed, 20 insertions(+), 14 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c
index 94f841026c3b..a195e48c9aee 100644
--- a/drivers/gpu/drm/nouveau/nouveau_svm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_svm.c
@@ -567,18 +567,27 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm,
unsigned long hmm_pfns[1];
struct hmm_range range = {
.notifier = &notifier->notifier,
- .start = notifier->notifier.interval_tree.start,
- .end = notifier->notifier.interval_tree.last + 1,
.default_flags = hmm_flags,
.hmm_pfns = hmm_pfns,
.dev_private_owner = drm->dev,
};
- struct mm_struct *mm = notifier->notifier.mm;
+ struct mm_struct *mm = svmm->notifier.mm;
int ret;

+ ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
+ args->p.addr, args->p.size,
+ &nouveau_svm_mni_ops);
+ if (ret)
+ return ret;
+
+ range.start = notifier->notifier.interval_tree.start;
+ range.end = notifier->notifier.interval_tree.last + 1;
+
while (true) {
- if (time_after(jiffies, timeout))
- return -EBUSY;
+ if (time_after(jiffies, timeout)) {
+ ret = -EBUSY;
+ goto out;
+ }

range.notifier_seq = mmu_interval_read_begin(range.notifier);
mmap_read_lock(mm);
@@ -587,7 +596,7 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm,
if (ret) {
if (ret == -EBUSY)
continue;
- return ret;
+ goto out;
}

mutex_lock(&svmm->mutex);
@@ -606,6 +615,9 @@ static int nouveau_range_fault(struct nouveau_svmm *svmm,
svmm->vmm->vmm.object.client->super = false;
mutex_unlock(&svmm->mutex);

+out:
+ mmu_interval_notifier_remove(&notifier->notifier);
+
return ret;
}

@@ -727,14 +739,8 @@ nouveau_svm_fault(struct nvif_notify *notify)
}

notifier.svmm = svmm;
- ret = mmu_interval_notifier_insert(&notifier.notifier, mm,
- args.i.p.addr, args.i.p.size,
- &nouveau_svm_mni_ops);
- if (!ret) {
- ret = nouveau_range_fault(svmm, svm->drm, &args.i,
- sizeof(args), hmm_flags, &notifier);
- mmu_interval_notifier_remove(&notifier.notifier);
- }
+ ret = nouveau_range_fault(svmm, svm->drm, &args.i,
+ sizeof(args), hmm_flags, &notifier);
mmput(mm);

limit = args.i.p.addr + args.i.p.size;
--
2.20.1

2021-05-24 13:32:40

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 04/10] mm/rmap: Split migration into its own function

Migration is currently implemented as a mode of operation for
try_to_unmap_one() generally specified by passing the TTU_MIGRATION flag
or in the case of splitting a huge anonymous page TTU_SPLIT_FREEZE.

However it does not have much in common with the rest of the unmap
functionality of try_to_unmap_one() and thus splitting it into a
separate function reduces the complexity of try_to_unmap_one() making it
more readable.

Several simplifications can also be made in try_to_migrate_one() based
on the following observations:

- All users of TTU_MIGRATION also set TTU_IGNORE_MLOCK.
- No users of TTU_MIGRATION ever set TTU_IGNORE_HWPOISON.
- No users of TTU_MIGRATION ever set TTU_BATCH_FLUSH.

TTU_SPLIT_FREEZE is a special case of migration used when splitting an
anonymous page. This is most easily dealt with by calling the correct
function from unmap_page() in mm/huge_memory.c - either
try_to_migrate() for PageAnon or try_to_unmap().

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Christoph Hellwig <[email protected]>
Reviewed-by: Ralph Campbell <[email protected]>

---

v5:
* Added comments about how PMD splitting works for migration vs.
unmapping
* Tightened up the flag check in try_to_migrate() to be explicit about
which TTU_XXX flags are supported.
---
include/linux/rmap.h | 4 +-
mm/huge_memory.c | 15 +-
mm/migrate.c | 9 +-
mm/rmap.c | 358 ++++++++++++++++++++++++++++++++-----------
4 files changed, 280 insertions(+), 106 deletions(-)

diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index 38a746787c2f..0e25d829f742 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -86,8 +86,6 @@ struct anon_vma_chain {
};

enum ttu_flags {
- TTU_MIGRATION = 0x1, /* migration mode */
-
TTU_SPLIT_HUGE_PMD = 0x4, /* split huge PMD if any */
TTU_IGNORE_MLOCK = 0x8, /* ignore mlock */
TTU_IGNORE_HWPOISON = 0x20, /* corrupted page is recoverable */
@@ -96,7 +94,6 @@ enum ttu_flags {
* do a final flush if necessary */
TTU_RMAP_LOCKED = 0x80, /* do not grab rmap lock:
* caller holds it */
- TTU_SPLIT_FREEZE = 0x100, /* freeze pte under splitting thp */
};

#ifdef CONFIG_MMU
@@ -193,6 +190,7 @@ static inline void page_dup_rmap(struct page *page, bool compound)
int page_referenced(struct page *, int is_locked,
struct mem_cgroup *memcg, unsigned long *vm_flags);

+bool try_to_migrate(struct page *page, enum ttu_flags flags);
bool try_to_unmap(struct page *, enum ttu_flags flags);

/* Avoid racy checks */
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 2ec6dab72217..6dddc75b89ee 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2345,16 +2345,21 @@ void vma_adjust_trans_huge(struct vm_area_struct *vma,

static void unmap_page(struct page *page)
{
- enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK |
- TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
+ enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
bool unmap_success;

VM_BUG_ON_PAGE(!PageHead(page), page);

if (PageAnon(page))
- ttu_flags |= TTU_SPLIT_FREEZE;
-
- unmap_success = try_to_unmap(page, ttu_flags);
+ unmap_success = try_to_migrate(page, ttu_flags);
+ else
+ /*
+ * Don't install migration entries for file backed pages. This
+ * helps handle cases when i_size is in the middle of the page
+ * as there is no need to unmap pages beyond i_size manually.
+ */
+ unmap_success = try_to_unmap(page, ttu_flags |
+ TTU_IGNORE_MLOCK);
VM_BUG_ON_PAGE(!unmap_success, page);
}

diff --git a/mm/migrate.c b/mm/migrate.c
index 930de919b1f2..05740f816bc4 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1103,7 +1103,7 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
/* Establish migration ptes */
VM_BUG_ON_PAGE(PageAnon(page) && !PageKsm(page) && !anon_vma,
page);
- try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK);
+ try_to_migrate(page, 0);
page_was_mapped = 1;
}

@@ -1305,7 +1305,7 @@ static int unmap_and_move_huge_page(new_page_t get_new_page,

if (page_mapped(hpage)) {
bool mapping_locked = false;
- enum ttu_flags ttu = TTU_MIGRATION|TTU_IGNORE_MLOCK;
+ enum ttu_flags ttu = 0;

if (!PageAnon(hpage)) {
/*
@@ -1322,7 +1322,7 @@ static int unmap_and_move_huge_page(new_page_t get_new_page,
ttu |= TTU_RMAP_LOCKED;
}

- try_to_unmap(hpage, ttu);
+ try_to_migrate(hpage, ttu);
page_was_mapped = 1;

if (mapping_locked)
@@ -2712,7 +2712,6 @@ static void migrate_vma_prepare(struct migrate_vma *migrate)
*/
static void migrate_vma_unmap(struct migrate_vma *migrate)
{
- int flags = TTU_MIGRATION | TTU_IGNORE_MLOCK;
const unsigned long npages = migrate->npages;
const unsigned long start = migrate->start;
unsigned long addr, i, restore = 0;
@@ -2724,7 +2723,7 @@ static void migrate_vma_unmap(struct migrate_vma *migrate)
continue;

if (page_mapped(page)) {
- try_to_unmap(page, flags);
+ try_to_migrate(page, 0);
if (page_mapped(page))
goto restore;
}
diff --git a/mm/rmap.c b/mm/rmap.c
index e88966903e1e..8ed1853060cf 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1405,14 +1405,8 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
struct mmu_notifier_range range;
enum ttu_flags flags = (enum ttu_flags)(long)arg;

- if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
- is_zone_device_page(page) && !is_device_private_page(page))
- return true;
-
- if (flags & TTU_SPLIT_HUGE_PMD) {
- split_huge_pmd_address(vma, address,
- flags & TTU_SPLIT_FREEZE, page);
- }
+ if (flags & TTU_SPLIT_HUGE_PMD)
+ split_huge_pmd_address(vma, address, false, page);

/*
* For THP, we have to assume the worse case ie pmd for invalidation.
@@ -1436,16 +1430,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
mmu_notifier_invalidate_range_start(&range);

while (page_vma_mapped_walk(&pvmw)) {
-#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
- /* PMD-mapped THP migration entry */
- if (!pvmw.pte && (flags & TTU_MIGRATION)) {
- VM_BUG_ON_PAGE(PageHuge(page) || !PageTransCompound(page), page);
-
- set_pmd_migration_entry(&pvmw, page);
- continue;
- }
-#endif
-
/*
* If the page is mlock()d, we cannot swap it out.
* If it's recently referenced (perhaps page_referenced
@@ -1507,46 +1491,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
}
}

- if (IS_ENABLED(CONFIG_MIGRATION) &&
- (flags & TTU_MIGRATION) &&
- is_zone_device_page(page)) {
- swp_entry_t entry;
- pte_t swp_pte;
-
- pteval = ptep_get_and_clear(mm, pvmw.address, pvmw.pte);
-
- /*
- * Store the pfn of the page in a special migration
- * pte. do_swap_page() will wait until the migration
- * pte is removed and then restart fault handling.
- */
- entry = make_readable_migration_entry(page_to_pfn(page));
- swp_pte = swp_entry_to_pte(entry);
-
- /*
- * pteval maps a zone device page and is therefore
- * a swap pte.
- */
- if (pte_swp_soft_dirty(pteval))
- swp_pte = pte_swp_mksoft_dirty(swp_pte);
- if (pte_swp_uffd_wp(pteval))
- swp_pte = pte_swp_mkuffd_wp(swp_pte);
- set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
- /*
- * No need to invalidate here it will synchronize on
- * against the special swap migration pte.
- *
- * The assignment to subpage above was computed from a
- * swap PTE which results in an invalid pointer.
- * Since only PAGE_SIZE pages can currently be
- * migrated, just set it to page. This will need to be
- * changed when hugepage migrations to device private
- * memory are supported.
- */
- subpage = page;
- goto discard;
- }
-
/* Nuke the page table entry. */
flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
if (should_defer_flush(mm, flags)) {
@@ -1599,39 +1543,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
/* We have to invalidate as we cleared the pte */
mmu_notifier_invalidate_range(mm, address,
address + PAGE_SIZE);
- } else if (IS_ENABLED(CONFIG_MIGRATION) &&
- (flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))) {
- swp_entry_t entry;
- pte_t swp_pte;
-
- if (arch_unmap_one(mm, vma, address, pteval) < 0) {
- set_pte_at(mm, address, pvmw.pte, pteval);
- ret = false;
- page_vma_mapped_walk_done(&pvmw);
- break;
- }
-
- /*
- * Store the pfn of the page in a special migration
- * pte. do_swap_page() will wait until the migration
- * pte is removed and then restart fault handling.
- */
- if (pte_write(pteval))
- entry = make_writable_migration_entry(
- page_to_pfn(subpage));
- else
- entry = make_readable_migration_entry(
- page_to_pfn(subpage));
- swp_pte = swp_entry_to_pte(entry);
- if (pte_soft_dirty(pteval))
- swp_pte = pte_swp_mksoft_dirty(swp_pte);
- if (pte_uffd_wp(pteval))
- swp_pte = pte_swp_mkuffd_wp(swp_pte);
- set_pte_at(mm, address, pvmw.pte, swp_pte);
- /*
- * No need to invalidate here it will synchronize on
- * against the special swap migration pte.
- */
} else if (PageAnon(page)) {
swp_entry_t entry = { .val = page_private(subpage) };
pte_t swp_pte;
@@ -1758,6 +1669,268 @@ bool try_to_unmap(struct page *page, enum ttu_flags flags)
.anon_lock = page_lock_anon_vma_read,
};

+ if (flags & TTU_RMAP_LOCKED)
+ rmap_walk_locked(page, &rwc);
+ else
+ rmap_walk(page, &rwc);
+
+ return !page_mapcount(page) ? true : false;
+}
+
+/*
+ * @arg: enum ttu_flags will be passed to this argument.
+ *
+ * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
+ * containing migration entries. This and TTU_RMAP_LOCKED are the only supported
+ * flags.
+ */
+static bool try_to_migrate_one(struct page *page, struct vm_area_struct *vma,
+ unsigned long address, void *arg)
+{
+ struct mm_struct *mm = vma->vm_mm;
+ struct page_vma_mapped_walk pvmw = {
+ .page = page,
+ .vma = vma,
+ .address = address,
+ };
+ pte_t pteval;
+ struct page *subpage;
+ bool ret = true;
+ struct mmu_notifier_range range;
+ enum ttu_flags flags = (enum ttu_flags)(long)arg;
+
+ if (is_zone_device_page(page) && !is_device_private_page(page))
+ return true;
+
+ /*
+ * unmap_page() in mm/huge_memory.c is the only user of migration with
+ * TTU_SPLIT_HUGE_PMD and it wants to freeze.
+ */
+ if (flags & TTU_SPLIT_HUGE_PMD)
+ split_huge_pmd_address(vma, address, true, page);
+
+ /*
+ * For THP, we have to assume the worse case ie pmd for invalidation.
+ * For hugetlb, it could be much worse if we need to do pud
+ * invalidation in the case of pmd sharing.
+ *
+ * Note that the page can not be free in this function as call of
+ * try_to_unmap() must hold a reference on the page.
+ */
+ mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
+ address,
+ min(vma->vm_end, address + page_size(page)));
+ if (PageHuge(page)) {
+ /*
+ * If sharing is possible, start and end will be adjusted
+ * accordingly.
+ */
+ adjust_range_if_pmd_sharing_possible(vma, &range.start,
+ &range.end);
+ }
+ mmu_notifier_invalidate_range_start(&range);
+
+ while (page_vma_mapped_walk(&pvmw)) {
+#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
+ /* PMD-mapped THP migration entry */
+ if (!pvmw.pte) {
+ VM_BUG_ON_PAGE(PageHuge(page) ||
+ !PageTransCompound(page), page);
+
+ set_pmd_migration_entry(&pvmw, page);
+ continue;
+ }
+#endif
+
+ /* Unexpected PMD-mapped THP? */
+ VM_BUG_ON_PAGE(!pvmw.pte, page);
+
+ subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
+ address = pvmw.address;
+
+ if (PageHuge(page) && !PageAnon(page)) {
+ /*
+ * To call huge_pmd_unshare, i_mmap_rwsem must be
+ * held in write mode. Caller needs to explicitly
+ * do this outside rmap routines.
+ */
+ VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
+ if (huge_pmd_unshare(mm, vma, &address, pvmw.pte)) {
+ /*
+ * huge_pmd_unshare unmapped an entire PMD
+ * page. There is no way of knowing exactly
+ * which PMDs may be cached for this mm, so
+ * we must flush them all. start/end were
+ * already adjusted above to cover this range.
+ */
+ flush_cache_range(vma, range.start, range.end);
+ flush_tlb_range(vma, range.start, range.end);
+ mmu_notifier_invalidate_range(mm, range.start,
+ range.end);
+
+ /*
+ * The ref count of the PMD page was dropped
+ * which is part of the way map counting
+ * is done for shared PMDs. Return 'true'
+ * here. When there is no other sharing,
+ * huge_pmd_unshare returns false and we will
+ * unmap the actual page and drop map count
+ * to zero.
+ */
+ page_vma_mapped_walk_done(&pvmw);
+ break;
+ }
+ }
+
+ /* Nuke the page table entry. */
+ flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
+ pteval = ptep_clear_flush(vma, address, pvmw.pte);
+
+ /* Move the dirty bit to the page. Now the pte is gone. */
+ if (pte_dirty(pteval))
+ set_page_dirty(page);
+
+ /* Update high watermark before we lower rss */
+ update_hiwater_rss(mm);
+
+ if (is_zone_device_page(page)) {
+ swp_entry_t entry;
+ pte_t swp_pte;
+
+ /*
+ * Store the pfn of the page in a special migration
+ * pte. do_swap_page() will wait until the migration
+ * pte is removed and then restart fault handling.
+ */
+ entry = make_readable_migration_entry(
+ page_to_pfn(page));
+ swp_pte = swp_entry_to_pte(entry);
+
+ /*
+ * pteval maps a zone device page and is therefore
+ * a swap pte.
+ */
+ if (pte_swp_soft_dirty(pteval))
+ swp_pte = pte_swp_mksoft_dirty(swp_pte);
+ if (pte_swp_uffd_wp(pteval))
+ swp_pte = pte_swp_mkuffd_wp(swp_pte);
+ set_pte_at(mm, pvmw.address, pvmw.pte, swp_pte);
+ /*
+ * No need to invalidate here it will synchronize on
+ * against the special swap migration pte.
+ *
+ * The assignment to subpage above was computed from a
+ * swap PTE which results in an invalid pointer.
+ * Since only PAGE_SIZE pages can currently be
+ * migrated, just set it to page. This will need to be
+ * changed when hugepage migrations to device private
+ * memory are supported.
+ */
+ subpage = page;
+ } else if (PageHWPoison(page)) {
+ pteval = swp_entry_to_pte(make_hwpoison_entry(subpage));
+ if (PageHuge(page)) {
+ hugetlb_count_sub(compound_nr(page), mm);
+ set_huge_swap_pte_at(mm, address,
+ pvmw.pte, pteval,
+ vma_mmu_pagesize(vma));
+ } else {
+ dec_mm_counter(mm, mm_counter(page));
+ set_pte_at(mm, address, pvmw.pte, pteval);
+ }
+
+ } else if (pte_unused(pteval) && !userfaultfd_armed(vma)) {
+ /*
+ * The guest indicated that the page content is of no
+ * interest anymore. Simply discard the pte, vmscan
+ * will take care of the rest.
+ * A future reference will then fault in a new zero
+ * page. When userfaultfd is active, we must not drop
+ * this page though, as its main user (postcopy
+ * migration) will not expect userfaults on already
+ * copied pages.
+ */
+ dec_mm_counter(mm, mm_counter(page));
+ /* We have to invalidate as we cleared the pte */
+ mmu_notifier_invalidate_range(mm, address,
+ address + PAGE_SIZE);
+ } else {
+ swp_entry_t entry;
+ pte_t swp_pte;
+
+ if (arch_unmap_one(mm, vma, address, pteval) < 0) {
+ set_pte_at(mm, address, pvmw.pte, pteval);
+ ret = false;
+ page_vma_mapped_walk_done(&pvmw);
+ break;
+ }
+
+ /*
+ * Store the pfn of the page in a special migration
+ * pte. do_swap_page() will wait until the migration
+ * pte is removed and then restart fault handling.
+ */
+ if (pte_write(pteval))
+ entry = make_writable_migration_entry(
+ page_to_pfn(subpage));
+ else
+ entry = make_readable_migration_entry(
+ page_to_pfn(subpage));
+
+ swp_pte = swp_entry_to_pte(entry);
+ if (pte_soft_dirty(pteval))
+ swp_pte = pte_swp_mksoft_dirty(swp_pte);
+ if (pte_uffd_wp(pteval))
+ swp_pte = pte_swp_mkuffd_wp(swp_pte);
+ set_pte_at(mm, address, pvmw.pte, swp_pte);
+ /*
+ * No need to invalidate here it will synchronize on
+ * against the special swap migration pte.
+ */
+ }
+
+ /*
+ * No need to call mmu_notifier_invalidate_range() it has be
+ * done above for all cases requiring it to happen under page
+ * table lock before mmu_notifier_invalidate_range_end()
+ *
+ * See Documentation/vm/mmu_notifier.rst
+ */
+ page_remove_rmap(subpage, PageHuge(page));
+ put_page(page);
+ }
+
+ mmu_notifier_invalidate_range_end(&range);
+
+ return ret;
+}
+
+/**
+ * try_to_migrate - try to replace all page table mappings with swap entries
+ * @page: the page to replace page table entries for
+ * @flags: action and flags
+ *
+ * Tries to remove all the page table entries which are mapping this page and
+ * replace them with special swap entries. Caller must hold the page lock.
+ *
+ * If is successful, return true. Otherwise, false.
+ */
+bool try_to_migrate(struct page *page, enum ttu_flags flags)
+{
+ struct rmap_walk_control rwc = {
+ .rmap_one = try_to_migrate_one,
+ .arg = (void *)flags,
+ .done = page_not_mapped,
+ .anon_lock = page_lock_anon_vma_read,
+ };
+
+ /*
+ * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
+ * TTU_SPLIT_HUGE_PMD flags.
+ */
+ if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD)))
+ return false;
+
/*
* During exec, a temporary VMA is setup and later moved.
* The VMA is moved under the anon_vma lock but not the
@@ -1766,8 +1939,7 @@ bool try_to_unmap(struct page *page, enum ttu_flags flags)
* locking requirements of exec(), migration skips
* temporary VMAs until after exec() completes.
*/
- if ((flags & (TTU_MIGRATION|TTU_SPLIT_FREEZE))
- && !PageKsm(page) && PageAnon(page))
+ if (!PageKsm(page) && PageAnon(page))
rwc.invalid_vma = invalid_migration_vma;

if (flags & TTU_RMAP_LOCKED)
--
2.20.1

2021-05-24 13:33:02

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 06/10] mm/memory.c: Allow different return codes for copy_nonpresent_pte()

Currently if copy_nonpresent_pte() returns a non-zero value it is
assumed to be a swap entry which requires further processing outside the
loop in copy_pte_range() after dropping locks. This prevents other
values being returned to signal conditions such as failure which a
subsequent change requires.

Instead make copy_nonpresent_pte() return an error code if further
processing is required and read the value for the swap entry in the main
loop under the ptl.

Signed-off-by: Alistair Popple <[email protected]>

---

v9:

New for v9 to allow device exclusive handling to occur in
copy_nonpresent_pte().
---
mm/memory.c | 12 +++++++-----
1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/mm/memory.c b/mm/memory.c
index 2fb455c365c2..e061cfa18c11 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -718,7 +718,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,

if (likely(!non_swap_entry(entry))) {
if (swap_duplicate(entry) < 0)
- return entry.val;
+ return -EAGAIN;

/* make sure dst_mm is on swapoff's mmlist. */
if (unlikely(list_empty(&dst_mm->mmlist))) {
@@ -974,11 +974,13 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
continue;
}
if (unlikely(!pte_present(*src_pte))) {
- entry.val = copy_nonpresent_pte(dst_mm, src_mm,
- dst_pte, src_pte,
- src_vma, addr, rss);
- if (entry.val)
+ ret = copy_nonpresent_pte(dst_mm, src_mm,
+ dst_pte, src_pte,
+ src_vma, addr, rss);
+ if (ret == -EAGAIN) {
+ entry = pte_to_swp_entry(*src_pte);
break;
+ }
progress += 8;
continue;
}
--
2.20.1

2021-05-24 13:33:18

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 08/10] mm: Selftests for exclusive device memory

Adds some selftests for exclusive device memory.

Signed-off-by: Alistair Popple <[email protected]>
Acked-by: Jason Gunthorpe <[email protected]>
Tested-by: Ralph Campbell <[email protected]>
Reviewed-by: Ralph Campbell <[email protected]>
---
lib/test_hmm.c | 124 +++++++++++++++++++
lib/test_hmm_uapi.h | 2 +
tools/testing/selftests/vm/hmm-tests.c | 158 +++++++++++++++++++++++++
3 files changed, 284 insertions(+)

diff --git a/lib/test_hmm.c b/lib/test_hmm.c
index 5c9f5a020c1d..305a9d9e2b4c 100644
--- a/lib/test_hmm.c
+++ b/lib/test_hmm.c
@@ -25,6 +25,7 @@
#include <linux/swapops.h>
#include <linux/sched/mm.h>
#include <linux/platform_device.h>
+#include <linux/rmap.h>

#include "test_hmm_uapi.h"

@@ -46,6 +47,7 @@ struct dmirror_bounce {
unsigned long cpages;
};

+#define DPT_XA_TAG_ATOMIC 1UL
#define DPT_XA_TAG_WRITE 3UL

/*
@@ -619,6 +621,54 @@ static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
}
}

+static int dmirror_check_atomic(struct dmirror *dmirror, unsigned long start,
+ unsigned long end)
+{
+ unsigned long pfn;
+
+ for (pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++) {
+ void *entry;
+ struct page *page;
+
+ entry = xa_load(&dmirror->pt, pfn);
+ page = xa_untag_pointer(entry);
+ if (xa_pointer_tag(entry) == DPT_XA_TAG_ATOMIC)
+ return -EPERM;
+ }
+
+ return 0;
+}
+
+static int dmirror_atomic_map(unsigned long start, unsigned long end,
+ struct page **pages, struct dmirror *dmirror)
+{
+ unsigned long pfn, mapped = 0;
+ int i;
+
+ /* Map the migrated pages into the device's page tables. */
+ mutex_lock(&dmirror->mutex);
+
+ for (i = 0, pfn = start >> PAGE_SHIFT; pfn < (end >> PAGE_SHIFT); pfn++, i++) {
+ void *entry;
+
+ if (!pages[i])
+ continue;
+
+ entry = pages[i];
+ entry = xa_tag_pointer(entry, DPT_XA_TAG_ATOMIC);
+ entry = xa_store(&dmirror->pt, pfn, entry, GFP_ATOMIC);
+ if (xa_is_err(entry)) {
+ mutex_unlock(&dmirror->mutex);
+ return xa_err(entry);
+ }
+
+ mapped++;
+ }
+
+ mutex_unlock(&dmirror->mutex);
+ return mapped;
+}
+
static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
struct dmirror *dmirror)
{
@@ -661,6 +711,71 @@ static int dmirror_migrate_finalize_and_map(struct migrate_vma *args,
return 0;
}

+static int dmirror_exclusive(struct dmirror *dmirror,
+ struct hmm_dmirror_cmd *cmd)
+{
+ unsigned long start, end, addr;
+ unsigned long size = cmd->npages << PAGE_SHIFT;
+ struct mm_struct *mm = dmirror->notifier.mm;
+ struct page *pages[64];
+ struct dmirror_bounce bounce;
+ unsigned long next;
+ int ret;
+
+ start = cmd->addr;
+ end = start + size;
+ if (end < start)
+ return -EINVAL;
+
+ /* Since the mm is for the mirrored process, get a reference first. */
+ if (!mmget_not_zero(mm))
+ return -EINVAL;
+
+ mmap_read_lock(mm);
+ for (addr = start; addr < end; addr = next) {
+ int i, mapped;
+
+ if (end < addr + (ARRAY_SIZE(pages) << PAGE_SHIFT))
+ next = end;
+ else
+ next = addr + (ARRAY_SIZE(pages) << PAGE_SHIFT);
+
+ ret = make_device_exclusive_range(mm, addr, next, pages, NULL);
+ mapped = dmirror_atomic_map(addr, next, pages, dmirror);
+ for (i = 0; i < ret; i++) {
+ if (pages[i]) {
+ unlock_page(pages[i]);
+ put_page(pages[i]);
+ }
+ }
+
+ if (addr + (mapped << PAGE_SHIFT) < next) {
+ mmap_read_unlock(mm);
+ mmput(mm);
+ return -EBUSY;
+ }
+ }
+ mmap_read_unlock(mm);
+ mmput(mm);
+
+ /* Return the migrated data for verification. */
+ ret = dmirror_bounce_init(&bounce, start, size);
+ if (ret)
+ return ret;
+ mutex_lock(&dmirror->mutex);
+ ret = dmirror_do_read(dmirror, start, end, &bounce);
+ mutex_unlock(&dmirror->mutex);
+ if (ret == 0) {
+ if (copy_to_user(u64_to_user_ptr(cmd->ptr), bounce.ptr,
+ bounce.size))
+ ret = -EFAULT;
+ }
+
+ cmd->cpages = bounce.cpages;
+ dmirror_bounce_fini(&bounce);
+ return ret;
+}
+
static int dmirror_migrate(struct dmirror *dmirror,
struct hmm_dmirror_cmd *cmd)
{
@@ -949,6 +1064,15 @@ static long dmirror_fops_unlocked_ioctl(struct file *filp,
ret = dmirror_migrate(dmirror, &cmd);
break;

+ case HMM_DMIRROR_EXCLUSIVE:
+ ret = dmirror_exclusive(dmirror, &cmd);
+ break;
+
+ case HMM_DMIRROR_CHECK_EXCLUSIVE:
+ ret = dmirror_check_atomic(dmirror, cmd.addr,
+ cmd.addr + (cmd.npages << PAGE_SHIFT));
+ break;
+
case HMM_DMIRROR_SNAPSHOT:
ret = dmirror_snapshot(dmirror, &cmd);
break;
diff --git a/lib/test_hmm_uapi.h b/lib/test_hmm_uapi.h
index 670b4ef2a5b6..f14dea5dcd06 100644
--- a/lib/test_hmm_uapi.h
+++ b/lib/test_hmm_uapi.h
@@ -33,6 +33,8 @@ struct hmm_dmirror_cmd {
#define HMM_DMIRROR_WRITE _IOWR('H', 0x01, struct hmm_dmirror_cmd)
#define HMM_DMIRROR_MIGRATE _IOWR('H', 0x02, struct hmm_dmirror_cmd)
#define HMM_DMIRROR_SNAPSHOT _IOWR('H', 0x03, struct hmm_dmirror_cmd)
+#define HMM_DMIRROR_EXCLUSIVE _IOWR('H', 0x04, struct hmm_dmirror_cmd)
+#define HMM_DMIRROR_CHECK_EXCLUSIVE _IOWR('H', 0x05, struct hmm_dmirror_cmd)

/*
* Values returned in hmm_dmirror_cmd.ptr for HMM_DMIRROR_SNAPSHOT.
diff --git a/tools/testing/selftests/vm/hmm-tests.c b/tools/testing/selftests/vm/hmm-tests.c
index 5d1ac691b9f4..864f126ffd78 100644
--- a/tools/testing/selftests/vm/hmm-tests.c
+++ b/tools/testing/selftests/vm/hmm-tests.c
@@ -1485,4 +1485,162 @@ TEST_F(hmm2, double_map)
hmm_buffer_free(buffer);
}

+/*
+ * Basic check of exclusive faulting.
+ */
+TEST_F(hmm, exclusive)
+{
+ struct hmm_buffer *buffer;
+ unsigned long npages;
+ unsigned long size;
+ unsigned long i;
+ int *ptr;
+ int ret;
+
+ npages = ALIGN(HMM_BUFFER_SIZE, self->page_size) >> self->page_shift;
+ ASSERT_NE(npages, 0);
+ size = npages << self->page_shift;
+
+ buffer = malloc(sizeof(*buffer));
+ ASSERT_NE(buffer, NULL);
+
+ buffer->fd = -1;
+ buffer->size = size;
+ buffer->mirror = malloc(size);
+ ASSERT_NE(buffer->mirror, NULL);
+
+ buffer->ptr = mmap(NULL, size,
+ PROT_READ | PROT_WRITE,
+ MAP_PRIVATE | MAP_ANONYMOUS,
+ buffer->fd, 0);
+ ASSERT_NE(buffer->ptr, MAP_FAILED);
+
+ /* Initialize buffer in system memory. */
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ptr[i] = i;
+
+ /* Map memory exclusively for device access. */
+ ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_EXCLUSIVE, buffer, npages);
+ ASSERT_EQ(ret, 0);
+ ASSERT_EQ(buffer->cpages, npages);
+
+ /* Check what the device read. */
+ for (i = 0, ptr = buffer->mirror; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i], i);
+
+ /* Fault pages back to system memory and check them. */
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i]++, i);
+
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i], i+1);
+
+ /* Check atomic access revoked */
+ ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_CHECK_EXCLUSIVE, buffer, npages);
+ ASSERT_EQ(ret, 0);
+
+ hmm_buffer_free(buffer);
+}
+
+TEST_F(hmm, exclusive_mprotect)
+{
+ struct hmm_buffer *buffer;
+ unsigned long npages;
+ unsigned long size;
+ unsigned long i;
+ int *ptr;
+ int ret;
+
+ npages = ALIGN(HMM_BUFFER_SIZE, self->page_size) >> self->page_shift;
+ ASSERT_NE(npages, 0);
+ size = npages << self->page_shift;
+
+ buffer = malloc(sizeof(*buffer));
+ ASSERT_NE(buffer, NULL);
+
+ buffer->fd = -1;
+ buffer->size = size;
+ buffer->mirror = malloc(size);
+ ASSERT_NE(buffer->mirror, NULL);
+
+ buffer->ptr = mmap(NULL, size,
+ PROT_READ | PROT_WRITE,
+ MAP_PRIVATE | MAP_ANONYMOUS,
+ buffer->fd, 0);
+ ASSERT_NE(buffer->ptr, MAP_FAILED);
+
+ /* Initialize buffer in system memory. */
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ptr[i] = i;
+
+ /* Map memory exclusively for device access. */
+ ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_EXCLUSIVE, buffer, npages);
+ ASSERT_EQ(ret, 0);
+ ASSERT_EQ(buffer->cpages, npages);
+
+ /* Check what the device read. */
+ for (i = 0, ptr = buffer->mirror; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i], i);
+
+ ret = mprotect(buffer->ptr, size, PROT_READ);
+ ASSERT_EQ(ret, 0);
+
+ /* Simulate a device writing system memory. */
+ ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_WRITE, buffer, npages);
+ ASSERT_EQ(ret, -EPERM);
+
+ hmm_buffer_free(buffer);
+}
+
+/*
+ * Check copy-on-write works.
+ */
+TEST_F(hmm, exclusive_cow)
+{
+ struct hmm_buffer *buffer;
+ unsigned long npages;
+ unsigned long size;
+ unsigned long i;
+ int *ptr;
+ int ret;
+
+ npages = ALIGN(HMM_BUFFER_SIZE, self->page_size) >> self->page_shift;
+ ASSERT_NE(npages, 0);
+ size = npages << self->page_shift;
+
+ buffer = malloc(sizeof(*buffer));
+ ASSERT_NE(buffer, NULL);
+
+ buffer->fd = -1;
+ buffer->size = size;
+ buffer->mirror = malloc(size);
+ ASSERT_NE(buffer->mirror, NULL);
+
+ buffer->ptr = mmap(NULL, size,
+ PROT_READ | PROT_WRITE,
+ MAP_PRIVATE | MAP_ANONYMOUS,
+ buffer->fd, 0);
+ ASSERT_NE(buffer->ptr, MAP_FAILED);
+
+ /* Initialize buffer in system memory. */
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ptr[i] = i;
+
+ /* Map memory exclusively for device access. */
+ ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_EXCLUSIVE, buffer, npages);
+ ASSERT_EQ(ret, 0);
+ ASSERT_EQ(buffer->cpages, npages);
+
+ fork();
+
+ /* Fault pages back to system memory and check them. */
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i]++, i);
+
+ for (i = 0, ptr = buffer->ptr; i < size / sizeof(*ptr); ++i)
+ ASSERT_EQ(ptr[i], i+1);
+
+ hmm_buffer_free(buffer);
+}
+
TEST_HARNESS_MAIN
--
2.20.1

2021-05-24 13:34:02

by Alistair Popple

[permalink] [raw]
Subject: [PATCH v9 10/10] nouveau/svm: Implement atomic SVM access

Some NVIDIA GPUs do not support direct atomic access to system memory
via PCIe. Instead this must be emulated by granting the GPU exclusive
access to the memory. This is achieved by replacing CPU page table
entries with special swap entries that fault on userspace access.

The driver then grants the GPU permission to update the page undergoing
atomic access via the GPU page tables. When CPU access to the page is
required a CPU fault is raised which calls into the device driver via
MMU notifiers to revoke the atomic access. The original page table
entries are then restored allowing CPU access to proceed.

Signed-off-by: Alistair Popple <[email protected]>
Reviewed-by: Ben Skeggs <[email protected]>

---

v9:
* Added Ben's Reviewed-By

v7:
* Removed magic values for fault access levels
* Improved readability of fault comparison code

v4:
* Check that page table entries haven't changed before mapping on the
device
---
drivers/gpu/drm/nouveau/include/nvif/if000c.h | 1 +
drivers/gpu/drm/nouveau/nouveau_svm.c | 126 ++++++++++++++++--
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h | 1 +
.../drm/nouveau/nvkm/subdev/mmu/vmmgp100.c | 6 +
4 files changed, 123 insertions(+), 11 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/include/nvif/if000c.h b/drivers/gpu/drm/nouveau/include/nvif/if000c.h
index d6dd40f21eed..9c7ff56831c5 100644
--- a/drivers/gpu/drm/nouveau/include/nvif/if000c.h
+++ b/drivers/gpu/drm/nouveau/include/nvif/if000c.h
@@ -77,6 +77,7 @@ struct nvif_vmm_pfnmap_v0 {
#define NVIF_VMM_PFNMAP_V0_APER 0x00000000000000f0ULL
#define NVIF_VMM_PFNMAP_V0_HOST 0x0000000000000000ULL
#define NVIF_VMM_PFNMAP_V0_VRAM 0x0000000000000010ULL
+#define NVIF_VMM_PFNMAP_V0_A 0x0000000000000004ULL
#define NVIF_VMM_PFNMAP_V0_W 0x0000000000000002ULL
#define NVIF_VMM_PFNMAP_V0_V 0x0000000000000001ULL
#define NVIF_VMM_PFNMAP_V0_NONE 0x0000000000000000ULL
diff --git a/drivers/gpu/drm/nouveau/nouveau_svm.c b/drivers/gpu/drm/nouveau/nouveau_svm.c
index a195e48c9aee..81526d65b4e2 100644
--- a/drivers/gpu/drm/nouveau/nouveau_svm.c
+++ b/drivers/gpu/drm/nouveau/nouveau_svm.c
@@ -35,6 +35,7 @@
#include <linux/sched/mm.h>
#include <linux/sort.h>
#include <linux/hmm.h>
+#include <linux/rmap.h>

struct nouveau_svm {
struct nouveau_drm *drm;
@@ -67,6 +68,11 @@ struct nouveau_svm {
} buffer[1];
};

+#define FAULT_ACCESS_READ 0
+#define FAULT_ACCESS_WRITE 1
+#define FAULT_ACCESS_ATOMIC 2
+#define FAULT_ACCESS_PREFETCH 3
+
#define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
#define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)

@@ -411,6 +417,24 @@ nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
fault->client);
}

+static int
+nouveau_svm_fault_priority(u8 fault)
+{
+ switch (fault) {
+ case FAULT_ACCESS_PREFETCH:
+ return 0;
+ case FAULT_ACCESS_READ:
+ return 1;
+ case FAULT_ACCESS_WRITE:
+ return 2;
+ case FAULT_ACCESS_ATOMIC:
+ return 3;
+ default:
+ WARN_ON_ONCE(1);
+ return -1;
+ }
+}
+
static int
nouveau_svm_fault_cmp(const void *a, const void *b)
{
@@ -421,9 +445,8 @@ nouveau_svm_fault_cmp(const void *a, const void *b)
return ret;
if ((ret = (s64)fa->addr - fb->addr))
return ret;
- /*XXX: atomic? */
- return (fa->access == 0 || fa->access == 3) -
- (fb->access == 0 || fb->access == 3);
+ return nouveau_svm_fault_priority(fa->access) -
+ nouveau_svm_fault_priority(fb->access);
}

static void
@@ -487,6 +510,10 @@ static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
struct svm_notifier *sn =
container_of(mni, struct svm_notifier, notifier);

+ if (range->event == MMU_NOTIFY_EXCLUSIVE &&
+ range->owner == sn->svmm->vmm->cli->drm->dev)
+ return true;
+
/*
* serializes the update to mni->invalidate_seq done by caller and
* prevents invalidation of the PTE from progressing while HW is being
@@ -555,6 +582,71 @@ static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm,
args->p.phys[0] |= NVIF_VMM_PFNMAP_V0_W;
}

+static int nouveau_atomic_range_fault(struct nouveau_svmm *svmm,
+ struct nouveau_drm *drm,
+ struct nouveau_pfnmap_args *args, u32 size,
+ struct svm_notifier *notifier)
+{
+ unsigned long timeout =
+ jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
+ struct mm_struct *mm = svmm->notifier.mm;
+ struct page *page;
+ unsigned long start = args->p.addr;
+ unsigned long notifier_seq;
+ int ret = 0;
+
+ ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
+ args->p.addr, args->p.size,
+ &nouveau_svm_mni_ops);
+ if (ret)
+ return ret;
+
+ while (true) {
+ if (time_after(jiffies, timeout)) {
+ ret = -EBUSY;
+ goto out;
+ }
+
+ notifier_seq = mmu_interval_read_begin(&notifier->notifier);
+ mmap_read_lock(mm);
+ make_device_exclusive_range(mm, start, start + PAGE_SIZE,
+ &page, drm->dev);
+ mmap_read_unlock(mm);
+ if (!page) {
+ ret = -EINVAL;
+ goto out;
+ }
+
+ mutex_lock(&svmm->mutex);
+ if (!mmu_interval_read_retry(&notifier->notifier,
+ notifier_seq))
+ break;
+ mutex_unlock(&svmm->mutex);
+ }
+
+ /* Map the page on the GPU. */
+ args->p.page = 12;
+ args->p.size = PAGE_SIZE;
+ args->p.addr = start;
+ args->p.phys[0] = page_to_phys(page) |
+ NVIF_VMM_PFNMAP_V0_V |
+ NVIF_VMM_PFNMAP_V0_W |
+ NVIF_VMM_PFNMAP_V0_A |
+ NVIF_VMM_PFNMAP_V0_HOST;
+
+ svmm->vmm->vmm.object.client->super = true;
+ ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
+ svmm->vmm->vmm.object.client->super = false;
+ mutex_unlock(&svmm->mutex);
+
+ unlock_page(page);
+ put_page(page);
+
+out:
+ mmu_interval_notifier_remove(&notifier->notifier);
+ return ret;
+}
+
static int nouveau_range_fault(struct nouveau_svmm *svmm,
struct nouveau_drm *drm,
struct nouveau_pfnmap_args *args, u32 size,
@@ -637,7 +729,7 @@ nouveau_svm_fault(struct nvif_notify *notify)
unsigned long hmm_flags;
u64 inst, start, limit;
int fi, fn;
- int replay = 0, ret;
+ int replay = 0, atomic = 0, ret;

/* Parse available fault buffer entries into a cache, and update
* the GET pointer so HW can reuse the entries.
@@ -718,12 +810,14 @@ nouveau_svm_fault(struct nvif_notify *notify)
/*
* Determine required permissions based on GPU fault
* access flags.
- * XXX: atomic?
*/
switch (buffer->fault[fi]->access) {
case 0: /* READ. */
hmm_flags = HMM_PFN_REQ_FAULT;
break;
+ case 2: /* ATOMIC. */
+ atomic = true;
+ break;
case 3: /* PREFETCH. */
hmm_flags = 0;
break;
@@ -739,8 +833,14 @@ nouveau_svm_fault(struct nvif_notify *notify)
}

notifier.svmm = svmm;
- ret = nouveau_range_fault(svmm, svm->drm, &args.i,
- sizeof(args), hmm_flags, &notifier);
+ if (atomic)
+ ret = nouveau_atomic_range_fault(svmm, svm->drm,
+ &args.i, sizeof(args),
+ &notifier);
+ else
+ ret = nouveau_range_fault(svmm, svm->drm, &args.i,
+ sizeof(args), hmm_flags,
+ &notifier);
mmput(mm);

limit = args.i.p.addr + args.i.p.size;
@@ -756,11 +856,15 @@ nouveau_svm_fault(struct nvif_notify *notify)
*/
if (buffer->fault[fn]->svmm != svmm ||
buffer->fault[fn]->addr >= limit ||
- (buffer->fault[fi]->access == 0 /* READ. */ &&
+ (buffer->fault[fi]->access == FAULT_ACCESS_READ &&
!(args.phys[0] & NVIF_VMM_PFNMAP_V0_V)) ||
- (buffer->fault[fi]->access != 0 /* READ. */ &&
- buffer->fault[fi]->access != 3 /* PREFETCH. */ &&
- !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W)))
+ (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
+ buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
+ !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W)) ||
+ (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
+ buffer->fault[fi]->access != FAULT_ACCESS_WRITE &&
+ buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
+ !(args.phys[0] & NVIF_VMM_PFNMAP_V0_A)))
break;
}

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
index a2b179568970..f6188aa9171c 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
@@ -178,6 +178,7 @@ void nvkm_vmm_unmap_region(struct nvkm_vmm *, struct nvkm_vma *);
#define NVKM_VMM_PFN_APER 0x00000000000000f0ULL
#define NVKM_VMM_PFN_HOST 0x0000000000000000ULL
#define NVKM_VMM_PFN_VRAM 0x0000000000000010ULL
+#define NVKM_VMM_PFN_A 0x0000000000000004ULL
#define NVKM_VMM_PFN_W 0x0000000000000002ULL
#define NVKM_VMM_PFN_V 0x0000000000000001ULL
#define NVKM_VMM_PFN_NONE 0x0000000000000000ULL
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
index 236db5570771..f02abd9cb4dd 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
@@ -88,6 +88,9 @@ gp100_vmm_pgt_pfn(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
if (!(*map->pfn & NVKM_VMM_PFN_W))
data |= BIT_ULL(6); /* RO. */

+ if (!(*map->pfn & NVKM_VMM_PFN_A))
+ data |= BIT_ULL(7); /* Atomic disable. */
+
if (!(*map->pfn & NVKM_VMM_PFN_VRAM)) {
addr = *map->pfn >> NVKM_VMM_PFN_ADDR_SHIFT;
addr = dma_map_page(dev, pfn_to_page(addr), 0,
@@ -322,6 +325,9 @@ gp100_vmm_pd0_pfn(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
if (!(*map->pfn & NVKM_VMM_PFN_W))
data |= BIT_ULL(6); /* RO. */

+ if (!(*map->pfn & NVKM_VMM_PFN_A))
+ data |= BIT_ULL(7); /* Atomic disable. */
+
if (!(*map->pfn & NVKM_VMM_PFN_VRAM)) {
addr = *map->pfn >> NVKM_VMM_PFN_ADDR_SHIFT;
addr = dma_map_page(dev, pfn_to_page(addr), 0,
--
2.20.1

2021-05-24 22:14:59

by Andrew Morton

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Mon, 24 May 2021 23:27:22 +1000 Alistair Popple <[email protected]> wrote:

> Some devices require exclusive write access to shared virtual
> memory (SVM) ranges to perform atomic operations on that memory. This
> requires CPU page tables to be updated to deny access whilst atomic
> operations are occurring.
>
> In order to do this introduce a new swap entry
> type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for
> exclusive access by a device all page table mappings for the particular
> range are replaced with device exclusive swap entries. This causes any
> CPU access to the page to result in a fault.
>
> Faults are resovled by replacing the faulting entry with the original
> mapping. This results in MMU notifiers being called which a driver uses
> to update access permissions such as revoking atomic access. After
> notifiers have been called the device will no longer have exclusive
> access to the region.
>
> Walking of the page tables to find the target pages is handled by
> get_user_pages() rather than a direct page table walk. A direct page
> table walk similar to what migrate_vma_collect()/unmap() does could also
> have been utilised. However this resulted in more code similar in
> functionality to what get_user_pages() provides as page faulting is
> required to make the PTEs present and to break COW.
>
> ...
>
> Documentation/vm/hmm.rst | 17 ++++
> include/linux/mmu_notifier.h | 6 ++
> include/linux/rmap.h | 4 +
> include/linux/swap.h | 7 +-
> include/linux/swapops.h | 44 ++++++++-
> mm/hmm.c | 5 +
> mm/memory.c | 128 +++++++++++++++++++++++-
> mm/mprotect.c | 8 ++
> mm/page_vma_mapped.c | 9 +-
> mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
> 10 files changed, 405 insertions(+), 9 deletions(-)
>

This is quite a lot of code added to core MM for a single driver.

Is there any expectation that other drivers will use this code?

Is there a way of reducing the impact (code size, at least) for systems
which don't need this code?

How beneficial is this code to nouveau users? I see that it permits a
part of OpenCL to be implemented, but how useful/important is this in
the real world?

Thanks.

2021-05-25 01:40:39

by John Hubbard

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On 5/24/21 3:11 PM, Andrew Morton wrote:
>> ...
>>
>> Documentation/vm/hmm.rst | 17 ++++
>> include/linux/mmu_notifier.h | 6 ++
>> include/linux/rmap.h | 4 +
>> include/linux/swap.h | 7 +-
>> include/linux/swapops.h | 44 ++++++++-
>> mm/hmm.c | 5 +
>> mm/memory.c | 128 +++++++++++++++++++++++-
>> mm/mprotect.c | 8 ++
>> mm/page_vma_mapped.c | 9 +-
>> mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
>> 10 files changed, 405 insertions(+), 9 deletions(-)
>>
>
> This is quite a lot of code added to core MM for a single driver.
>
> Is there any expectation that other drivers will use this code?

Yes! This should work for GPUs (and potentially, other devices) that support
OpenCL SVM atomic accesses on the device. I haven't looked into how amdgpu
works in any detail, but that's certainly at the top of the list of likely
additional callers.

>
> Is there a way of reducing the impact (code size, at least) for systems
> which don't need this code?

I'll leave this question to others for the moment, in order to answer
the "do we need it at all" points.

>
> How beneficial is this code to nouveau users? I see that it permits a
> part of OpenCL to be implemented, but how useful/important is this in
> the real world?
>

So this is interesting. Right now, OpenCL support in Nouveau is rather new
and so probably not a huge impact yet. However, we've built up enough experience
with CUDA and OpenCL to learn that atomic operations, as part of the user
space programming model, are a super big deal. Atomic operations are so
useful and important that I'd expect many OpenCL SVM users to be uninterested in
programming models that lack atomic operations for GPU compute programs.

Again, this doesn't rule out future, non-GPU accelerator devices that may
come along.

Atomic ops are just a really important piece of high-end multi-threaded
programming, it turns out. So this is the beginning of support for an
important building block for general purpose programming on devices that
have GPU-like memory models.


thanks,
--
John Hubbard
NVIDIA

2021-05-25 09:24:15

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Tuesday, 25 May 2021 11:31:17 AM AEST John Hubbard wrote:
> On 5/24/21 3:11 PM, Andrew Morton wrote:
> >> ...
> >>
> >> Documentation/vm/hmm.rst | 17 ++++
> >> include/linux/mmu_notifier.h | 6 ++
> >> include/linux/rmap.h | 4 +
> >> include/linux/swap.h | 7 +-
> >> include/linux/swapops.h | 44 ++++++++-
> >> mm/hmm.c | 5 +
> >> mm/memory.c | 128 +++++++++++++++++++++++-
> >> mm/mprotect.c | 8 ++
> >> mm/page_vma_mapped.c | 9 +-
> >> mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
> >> 10 files changed, 405 insertions(+), 9 deletions(-)
> >
> > This is quite a lot of code added to core MM for a single driver.
> >
> > Is there any expectation that other drivers will use this code?
>
> Yes! This should work for GPUs (and potentially, other devices) that support
> OpenCL SVM atomic accesses on the device. I haven't looked into how amdgpu
> works in any detail, but that's certainly at the top of the list of likely
> additional callers.
>
> > Is there a way of reducing the impact (code size, at least) for systems
> > which don't need this code?

All of the code added to mm/rmap.c is specific to implementing this feature
and not depended on by other core MM code so could be put behind something
like CONFIG_DEVICE_PRIVATE to reduce the code size impact (I realise now it
currently isn't but should be).

The impact on compiled code size in mm/memory.c also ends up being minimised
by the compiler because all of it is of the form:

if (is_device_exclusive_entry(...)) {
[...]
}

Meaning it should get thrown away when the feature is not configured given
is_device_exclusive_entry() is a static inline always returning false in that
case.

> I'll leave this question to others for the moment, in order to answer
> the "do we need it at all" points.
>
> > How beneficial is this code to nouveau users? I see that it permits a
> > part of OpenCL to be implemented, but how useful/important is this in
> > the real world?
>
> So this is interesting. Right now, OpenCL support in Nouveau is rather new
> and so probably not a huge impact yet. However, we've built up enough
> experience with CUDA and OpenCL to learn that atomic operations, as part of
> the user space programming model, are a super big deal. Atomic operations
> are so useful and important that I'd expect many OpenCL SVM users to be
> uninterested in programming models that lack atomic operations for GPU
> compute programs.
>
> Again, this doesn't rule out future, non-GPU accelerator devices that may
> come along.
>
> Atomic ops are just a really important piece of high-end multi-threaded
> programming, it turns out. So this is the beginning of support for an
> important building block for general purpose programming on devices that
> have GPU-like memory models.
>
>
> thanks,




2021-05-25 15:41:56

by Balbir Singh

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Mon, May 24, 2021 at 03:11:57PM -0700, Andrew Morton wrote:
> On Mon, 24 May 2021 23:27:22 +1000 Alistair Popple <[email protected]> wrote:
>
> > Some devices require exclusive write access to shared virtual
> > memory (SVM) ranges to perform atomic operations on that memory. This
> > requires CPU page tables to be updated to deny access whilst atomic
> > operations are occurring.
> >
> > In order to do this introduce a new swap entry
> > type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for
> > exclusive access by a device all page table mappings for the particular
> > range are replaced with device exclusive swap entries. This causes any
> > CPU access to the page to result in a fault.
> >
> > Faults are resovled by replacing the faulting entry with the original
> > mapping. This results in MMU notifiers being called which a driver uses
> > to update access permissions such as revoking atomic access. After
> > notifiers have been called the device will no longer have exclusive
> > access to the region.
> >
> > Walking of the page tables to find the target pages is handled by
> > get_user_pages() rather than a direct page table walk. A direct page
> > table walk similar to what migrate_vma_collect()/unmap() does could also
> > have been utilised. However this resulted in more code similar in
> > functionality to what get_user_pages() provides as page faulting is
> > required to make the PTEs present and to break COW.
> >
> > ...
> >
> > Documentation/vm/hmm.rst | 17 ++++
> > include/linux/mmu_notifier.h | 6 ++
> > include/linux/rmap.h | 4 +
> > include/linux/swap.h | 7 +-
> > include/linux/swapops.h | 44 ++++++++-
> > mm/hmm.c | 5 +
> > mm/memory.c | 128 +++++++++++++++++++++++-
> > mm/mprotect.c | 8 ++
> > mm/page_vma_mapped.c | 9 +-
> > mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
> > 10 files changed, 405 insertions(+), 9 deletions(-)
> >
>
> This is quite a lot of code added to core MM for a single driver.
>
> Is there any expectation that other drivers will use this code?
>
> Is there a way of reducing the impact (code size, at least) for systems
> which don't need this code?
>
> How beneficial is this code to nouveau users? I see that it permits a
> part of OpenCL to be implemented, but how useful/important is this in
> the real world?

That is a very good question! I've not reviewed the code, but a sample
program with the described use case would make things easy to parse.
I suspect that is not easy to build at the moment?

I wonder how we co-ordinate all the work the mm is doing, page migration,
reclaim with device exclusive access? Do we have any numbers for the worst
case page fault latency when something is marked away for exclusive access?
I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE would
only impact the address space of programs using the GPU. Should the exclusively
marked range live in the unreclaimable list and recycled back to active/in-active
to account for the fact that

1. It is not reclaimable and reclaim will only hurt via page faults?
2. It ages the page correctly or at-least allows for that possibility when the
page is used by the GPU.

Balbir Singh.

2021-05-25 21:50:05

by Liam R. Howlett

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

* Alistair Popple <[email protected]> [210524 09:29]:
> The behaviour of try_to_unmap_one() is difficult to follow because it
> performs different operations based on a fairly large set of flags used
> in different combinations.
>
> TTU_MUNLOCK is one such flag. However it is exclusively used by
> try_to_munlock() which specifies no other flags. Therefore rather than
> overload try_to_unmap_one() with unrelated behaviour split this out into
> it's own function and remove the flag.
>
> Signed-off-by: Alistair Popple <[email protected]>
> Reviewed-by: Ralph Campbell <[email protected]>
> Reviewed-by: Christoph Hellwig <[email protected]>
>
> ---
>
> v9:
> * Improved comments
>
> v8:
> * Renamed try_to_munlock to page_mlock to better reflect what the
> function actually does.
> * Removed the TODO from the documentation that this patch addresses.
>
> v7:
> * Added Christoph's Reviewed-by
>
> v4:
> * Removed redundant check for VM_LOCKED
> ---
> Documentation/vm/unevictable-lru.rst | 33 ++++++---------
> include/linux/rmap.h | 3 +-
> mm/mlock.c | 10 ++---
> mm/rmap.c | 61 ++++++++++++++++++++--------
> 4 files changed, 63 insertions(+), 44 deletions(-)
>
> diff --git a/Documentation/vm/unevictable-lru.rst b/Documentation/vm/unevictable-lru.rst
> index 0e1490524f53..eae3af17f2d9 100644
> --- a/Documentation/vm/unevictable-lru.rst
> +++ b/Documentation/vm/unevictable-lru.rst
> @@ -389,14 +389,14 @@ mlocked, munlock_vma_page() updates that zone statistics for the number of
> mlocked pages. Note, however, that at this point we haven't checked whether
> the page is mapped by other VM_LOCKED VMAs.
>
> -We can't call try_to_munlock(), the function that walks the reverse map to
> +We can't call page_mlock(), the function that walks the reverse map to
> check for other VM_LOCKED VMAs, without first isolating the page from the LRU.
> -try_to_munlock() is a variant of try_to_unmap() and thus requires that the page
> +page_mlock() is a variant of try_to_unmap() and thus requires that the page
> not be on an LRU list [more on these below]. However, the call to
> -isolate_lru_page() could fail, in which case we couldn't try_to_munlock(). So,
> +isolate_lru_page() could fail, in which case we can't call page_mlock(). So,
> we go ahead and clear PG_mlocked up front, as this might be the only chance we
> -have. If we can successfully isolate the page, we go ahead and
> -try_to_munlock(), which will restore the PG_mlocked flag and update the zone
> +have. If we can successfully isolate the page, we go ahead and call
> +page_mlock(), which will restore the PG_mlocked flag and update the zone
> page statistics if it finds another VMA holding the page mlocked. If we fail
> to isolate the page, we'll have left a potentially mlocked page on the LRU.
> This is fine, because we'll catch it later if and if vmscan tries to reclaim
> @@ -545,31 +545,24 @@ munlock or munmap system calls, mm teardown (munlock_vma_pages_all), reclaim,
> holepunching, and truncation of file pages and their anonymous COWed pages.
>
>
> -try_to_munlock() Reverse Map Scan
> +page_mlock() Reverse Map Scan
> ---------------------------------
>
> -.. warning::
> - [!] TODO/FIXME: a better name might be page_mlocked() - analogous to the
> - page_referenced() reverse map walker.
> -
> When munlock_vma_page() [see section :ref:`munlock()/munlockall() System Call
> Handling <munlock_munlockall_handling>` above] tries to munlock a
> page, it needs to determine whether or not the page is mapped by any
> VM_LOCKED VMA without actually attempting to unmap all PTEs from the
> page. For this purpose, the unevictable/mlock infrastructure
> -introduced a variant of try_to_unmap() called try_to_munlock().
> +introduced a variant of try_to_unmap() called page_mlock().
>
> -try_to_munlock() calls the same functions as try_to_unmap() for anonymous and
> -mapped file and KSM pages with a flag argument specifying unlock versus unmap
> -processing. Again, these functions walk the respective reverse maps looking
> -for VM_LOCKED VMAs. When such a VMA is found, as in the try_to_unmap() case,
> -the functions mlock the page via mlock_vma_page() and return SWAP_MLOCK. This
> -undoes the pre-clearing of the page's PG_mlocked done by munlock_vma_page.
> +page_mlock() walks the respective reverse maps looking for VM_LOCKED VMAs. When
> +such a VMA is found the page is mlocked via mlock_vma_page(). This undoes the
> +pre-clearing of the page's PG_mlocked done by munlock_vma_page.
>
> -Note that try_to_munlock()'s reverse map walk must visit every VMA in a page's
> +Note that page_mlock()'s reverse map walk must visit every VMA in a page's
> reverse map to determine that a page is NOT mapped into any VM_LOCKED VMA.
> However, the scan can terminate when it encounters a VM_LOCKED VMA.
> -Although try_to_munlock() might be called a great many times when munlocking a
> +Although page_mlock() might be called a great many times when munlocking a
> large region or tearing down a large address space that has been mlocked via
> mlockall(), overall this is a fairly rare event.
>
> @@ -602,7 +595,7 @@ inactive lists to the appropriate node's unevictable list.
> shrink_inactive_list() should only see SHM_LOCK'd pages that became SHM_LOCK'd
> after shrink_active_list() had moved them to the inactive list, or pages mapped
> into VM_LOCKED VMAs that munlock_vma_page() couldn't isolate from the LRU to
> -recheck via try_to_munlock(). shrink_inactive_list() won't notice the latter,
> +recheck via page_mlock(). shrink_inactive_list() won't notice the latter,
> but will pass on to shrink_page_list().
>
> shrink_page_list() again culls obviously unevictable pages that it could
> diff --git a/include/linux/rmap.h b/include/linux/rmap.h
> index def5c62c93b3..38a746787c2f 100644
> --- a/include/linux/rmap.h
> +++ b/include/linux/rmap.h
> @@ -87,7 +87,6 @@ struct anon_vma_chain {
>
> enum ttu_flags {
> TTU_MIGRATION = 0x1, /* migration mode */
> - TTU_MUNLOCK = 0x2, /* munlock mode */
>
> TTU_SPLIT_HUGE_PMD = 0x4, /* split huge PMD if any */
> TTU_IGNORE_MLOCK = 0x8, /* ignore mlock */
> @@ -239,7 +238,7 @@ int page_mkclean(struct page *);
> * called in munlock()/munmap() path to check for other vmas holding
> * the page mlocked.
> */
> -void try_to_munlock(struct page *);
> +void page_mlock(struct page *page);
>
> void remove_migration_ptes(struct page *old, struct page *new, bool locked);
>
> diff --git a/mm/mlock.c b/mm/mlock.c
> index df590fda5688..a518d4c48e65 100644
> --- a/mm/mlock.c
> +++ b/mm/mlock.c
> @@ -108,7 +108,7 @@ void mlock_vma_page(struct page *page)
> /*
> * Finish munlock after successful page isolation
> *
> - * Page must be locked. This is a wrapper for try_to_munlock()
> + * Page must be locked. This is a wrapper for page_mlock()
> * and putback_lru_page() with munlock accounting.
> */
> static void __munlock_isolated_page(struct page *page)
> @@ -118,7 +118,7 @@ static void __munlock_isolated_page(struct page *page)
> * and we don't need to check all the other vmas.
> */
> if (page_mapcount(page) > 1)
> - try_to_munlock(page);
> + page_mlock(page);
>
> /* Did try_to_unlock() succeed or punt? */
> if (!PageMlocked(page))
> @@ -158,7 +158,7 @@ static void __munlock_isolation_failed(struct page *page)
> * munlock()ed or munmap()ed, we want to check whether other vmas hold the
> * page locked so that we can leave it on the unevictable lru list and not
> * bother vmscan with it. However, to walk the page's rmap list in
> - * try_to_munlock() we must isolate the page from the LRU. If some other
> + * page_mlock() we must isolate the page from the LRU. If some other
> * task has removed the page from the LRU, we won't be able to do that.
> * So we clear the PageMlocked as we might not get another chance. If we
> * can't isolate the page, we leave it for putback_lru_page() and vmscan
> @@ -168,7 +168,7 @@ unsigned int munlock_vma_page(struct page *page)
> {
> int nr_pages;
>
> - /* For try_to_munlock() and to serialize with page migration */
> + /* For page_mlock() and to serialize with page migration */
> BUG_ON(!PageLocked(page));
> VM_BUG_ON_PAGE(PageTail(page), page);
>
> @@ -205,7 +205,7 @@ static int __mlock_posix_error_return(long retval)
> *
> * The fast path is available only for evictable pages with single mapping.
> * Then we can bypass the per-cpu pvec and get better performance.
> - * when mapcount > 1 we need try_to_munlock() which can fail.
> + * when mapcount > 1 we need page_mlock() which can fail.
> * when !page_evictable(), we need the full redo logic of putback_lru_page to
> * avoid leaving evictable page in unevictable list.
> *
> diff --git a/mm/rmap.c b/mm/rmap.c
> index bc08c4d4b58a..e88966903e1e 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -1405,10 +1405,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> struct mmu_notifier_range range;
> enum ttu_flags flags = (enum ttu_flags)(long)arg;
>
> - /* munlock has nothing to gain from examining un-locked vmas */
> - if ((flags & TTU_MUNLOCK) && !(vma->vm_flags & VM_LOCKED))
> - return true;
> -
> if (IS_ENABLED(CONFIG_MIGRATION) && (flags & TTU_MIGRATION) &&
> is_zone_device_page(page) && !is_device_private_page(page))
> return true;
> @@ -1469,8 +1465,6 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
> page_vma_mapped_walk_done(&pvmw);
> break;
> }
> - if (flags & TTU_MUNLOCK)
> - continue;
> }
>
> /* Unexpected PMD-mapped THP? */
> @@ -1784,20 +1778,53 @@ bool try_to_unmap(struct page *page, enum ttu_flags flags)
> return !page_mapcount(page) ? true : false;
> }
>
> +/*
> + * Walks the vma's mapping a page and mlocks the page if any locked vma's are
> + * found. Once one is found the page is locked and the scan can be terminated.
> + */

Can you please add that this requires the mmap_sem() lock to the
comments?

> +static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
> + unsigned long address, void *unused)
> +{
> + struct page_vma_mapped_walk pvmw = {
> + .page = page,
> + .vma = vma,
> + .address = address,
> + };
> +
> + /* An un-locked vma doesn't have any pages to lock, continue the scan */
> + if (!(vma->vm_flags & VM_LOCKED))
> + return true;
> +
> + while (page_vma_mapped_walk(&pvmw)) {
> + /* PTE-mapped THP are never mlocked */
> + if (!PageTransCompound(page))
> + mlock_vma_page(page);
> + page_vma_mapped_walk_done(&pvmw);
> +
> + /*
> + * no need to continue scanning other vma's if the page has
> + * been locked.
> + */
> + return false;
> + }
> +
> + return true;
> +}
> +
> /**
> - * try_to_munlock - try to munlock a page
> - * @page: the page to be munlocked
> + * page_mlock - try to mlock a page
> + * @page: the page to be mlocked
> *
> - * Called from munlock code. Checks all of the VMAs mapping the page
> - * to make sure nobody else has this page mlocked. The page will be
> - * returned with PG_mlocked cleared if no other vmas have it mlocked.
> + * Called from munlock code. Checks all of the VMAs mapping the page and mlocks
> + * the page if any are found. The page will be returned with PG_mlocked cleared
> + * if it is not mapped by any locked vmas.
> + *
> + * mmap_lock should be held for read or write.
> */
> -
> -void try_to_munlock(struct page *page)
> +void page_mlock(struct page *page)
> {
> struct rmap_walk_control rwc = {
> - .rmap_one = try_to_unmap_one,
> - .arg = (void *)TTU_MUNLOCK,
> + .rmap_one = page_mlock_one,
> .done = page_not_mapped,
> .anon_lock = page_lock_anon_vma_read,
>
> @@ -1849,7 +1876,7 @@ static struct anon_vma *rmap_walk_anon_lock(struct page *page,
> * Find all the mappings of a page using the mapping pointer and the vma chains
> * contained in the anon_vma struct it points to.
> *
> - * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
> + * When called from page_mlock(), the mmap_lock of the mm containing the vma
> * where the page was found will be held for write. So, we won't recheck
> * vm_flags for that VMA. That should be OK, because that vma shouldn't be
> * LOCKED.
> @@ -1901,7 +1928,7 @@ static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,
> * Find all the mappings of a page using the mapping pointer and the vma chains
> * contained in the address_space struct it points to.
> *
> - * When called from try_to_munlock(), the mmap_lock of the mm containing the vma
> + * When called from page_mlock(), the mmap_lock of the mm containing the vma
> * where the page was found will be held for write. So, we won't recheck
> * vm_flags for that VMA. That should be OK, because that vma shouldn't be
> * LOCKED.
> --
> 2.20.1
>
>

I believe munlock_vma_pages_range() still references the old function
name?

Thanks,
Liam

2021-05-26 03:49:12

by Shakeel Butt

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

On Tue, May 25, 2021 at 11:40 AM Liam Howlett <[email protected]> wrote:
>
[...]
> >
> > +/*
> > + * Walks the vma's mapping a page and mlocks the page if any locked vma's are
> > + * found. Once one is found the page is locked and the scan can be terminated.
> > + */
>
> Can you please add that this requires the mmap_sem() lock to the
> comments?
>

Why does this require mmap_sem() lock? Also mmap_sem() lock of which mm_struct?

> > +static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
> > + unsigned long address, void *unused)
> > +{
> > + struct page_vma_mapped_walk pvmw = {
> > + .page = page,
> > + .vma = vma,
> > + .address = address,
> > + };
> > +
> > + /* An un-locked vma doesn't have any pages to lock, continue the scan */
> > + if (!(vma->vm_flags & VM_LOCKED))
> > + return true;
> > +
> > + while (page_vma_mapped_walk(&pvmw)) {
> > + /* PTE-mapped THP are never mlocked */
> > + if (!PageTransCompound(page))
> > + mlock_vma_page(page);
> > + page_vma_mapped_walk_done(&pvmw);
> > +
> > + /*
> > + * no need to continue scanning other vma's if the page has
> > + * been locked.
> > + */
> > + return false;
> > + }
> > +
> > + return true;
> > +}

2021-05-26 08:03:46

by John Hubbard

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On 5/25/21 4:51 AM, Balbir Singh wrote:
...
>> How beneficial is this code to nouveau users? I see that it permits a
>> part of OpenCL to be implemented, but how useful/important is this in
>> the real world?
>
> That is a very good question! I've not reviewed the code, but a sample
> program with the described use case would make things easy to parse.
> I suspect that is not easy to build at the moment?
>

The cover letter says this:

This has been tested with upstream Mesa 21.1.0 and a simple OpenCL program
which checks that GPU atomic accesses to system memory are atomic. Without
this series the test fails as there is no way of write-protecting the page
mapping which results in the device clobbering CPU writes. For reference
the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/

Further testing has been performed by adding support for testing exclusive
access to the hmm-tests kselftests.

...so that seems to cover the "sample program" request, at least.

> I wonder how we co-ordinate all the work the mm is doing, page migration,
> reclaim with device exclusive access? Do we have any numbers for the worst
> case page fault latency when something is marked away for exclusive access?

CPU page fault latency is approximately "terrible", if a page is resident on
the GPU. We have to spin up a DMA engine on the GPU and have it copy the page
over the PCIe bus, after all.

> I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE would

Yes, for now.

> only impact the address space of programs using the GPU. Should the exclusively
> marked range live in the unreclaimable list and recycled back to active/in-active
> to account for the fact that
>
> 1. It is not reclaimable and reclaim will only hurt via page faults?
> 2. It ages the page correctly or at-least allows for that possibility when the
> page is used by the GPU.

I'm not sure that that is *necessarily* something we can conclude. It depends upon
access patterns of each program. For example, a "reduction" parallel program sends
over lots of data to the GPU, and only a tiny bit of (reduced!) data comes back
to the CPU. In that case, freeing the physical page on the CPU is actually the
best decision for the OS to make (if the OS is sufficiently prescient).

thanks,
--
John Hubbard
NVIDIA

2021-05-26 18:46:48

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Wednesday, 26 May 2021 5:17:18 PM AEST John Hubbard wrote:
> On 5/25/21 4:51 AM, Balbir Singh wrote:
> ...
>
> >> How beneficial is this code to nouveau users? I see that it permits a
> >> part of OpenCL to be implemented, but how useful/important is this in
> >> the real world?
> >
> > That is a very good question! I've not reviewed the code, but a sample
> > program with the described use case would make things easy to parse.
> > I suspect that is not easy to build at the moment?
>
> The cover letter says this:
>
> This has been tested with upstream Mesa 21.1.0 and a simple OpenCL program
> which checks that GPU atomic accesses to system memory are atomic. Without
> this series the test fails as there is no way of write-protecting the page
> mapping which results in the device clobbering CPU writes. For reference
> the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/
>
> Further testing has been performed by adding support for testing exclusive
> access to the hmm-tests kselftests.
>
> ...so that seems to cover the "sample program" request, at least.

It is also sufficiently easy to build, assuming of course you have the
appropriate Mesa/LLVM/OpenCL libraries installed :-)

If you are interested I have some scripts which may help with building Mesa,
etc. Not that that is especially hard either, it's just there are a couple of
different dependencies required.

> > I wonder how we co-ordinate all the work the mm is doing, page migration,
> > reclaim with device exclusive access? Do we have any numbers for the worst
> > case page fault latency when something is marked away for exclusive
> > access?
>
> CPU page fault latency is approximately "terrible", if a page is resident on
> the GPU. We have to spin up a DMA engine on the GPU and have it copy the
> page over the PCIe bus, after all.

Although for clarity that describes latency for CPU faults to device private
pages which are always resident on the GPU. A CPU fault to a page being
exclusively accessed will be slightly less terrible as it only requires the
GPU MMU/TLB mappings to be taken down in much the same as for any other MMU
notifier callback as the page is mapped by the GPU rather than resident there.

> > I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE
> > would
>
> Yes, for now.
>
> > only impact the address space of programs using the GPU. Should the
> > exclusively marked range live in the unreclaimable list and recycled back
> > to active/in-active to account for the fact that
> >
> > 1. It is not reclaimable and reclaim will only hurt via page faults?
> > 2. It ages the page correctly or at-least allows for that possibility when
> > the>
> > page is used by the GPU.
>
> I'm not sure that that is *necessarily* something we can conclude. It
> depends upon access patterns of each program. For example, a "reduction"
> parallel program sends over lots of data to the GPU, and only a tiny bit of
> (reduced!) data comes back to the CPU. In that case, freeing the physical
> page on the CPU is actually the best decision for the OS to make (if the OS
> is sufficiently prescient).
>
> thanks,




2021-05-26 20:42:15

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 05/10] mm: Rename migrate_pgmap_owner

On Mon, May 24, 2021 at 11:27:20PM +1000, Alistair Popple wrote:
> @@ -521,14 +521,14 @@ static inline void mmu_notifier_range_init(struct mmu_notifier_range *range,
> range->flags = flags;
> }
>
> -static inline void mmu_notifier_range_init_migrate(
> - struct mmu_notifier_range *range, unsigned int flags,
> +static inline void mmu_notifier_range_init_owner(
> + struct mmu_notifier_range *range,
> + enum mmu_notifier_event event, unsigned int flags,
> struct vm_area_struct *vma, struct mm_struct *mm,
> - unsigned long start, unsigned long end, void *pgmap)
> + unsigned long start, unsigned long end, void *owner)
> {
> - mmu_notifier_range_init(range, MMU_NOTIFY_MIGRATE, flags, vma, mm,
> - start, end);
> - range->migrate_pgmap_owner = pgmap;
> + mmu_notifier_range_init(range, event, flags, vma, mm, start, end);
> + range->owner = owner;
> }

mmu_notifier_range_init_migrate() can even be kept to just call the new helper,
then existing callers are unaffected. Not a big deal, though:

Reviewed-by: Peter Xu <[email protected]>

Thanks,

--
Peter Xu

2021-05-26 20:47:16

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 06/10] mm/memory.c: Allow different return codes for copy_nonpresent_pte()

On Mon, May 24, 2021 at 11:27:21PM +1000, Alistair Popple wrote:
> Currently if copy_nonpresent_pte() returns a non-zero value it is
> assumed to be a swap entry which requires further processing outside the
> loop in copy_pte_range() after dropping locks. This prevents other
> values being returned to signal conditions such as failure which a
> subsequent change requires.
>
> Instead make copy_nonpresent_pte() return an error code if further
> processing is required and read the value for the swap entry in the main
> loop under the ptl.
>
> Signed-off-by: Alistair Popple <[email protected]>
>
> ---
>
> v9:
>
> New for v9 to allow device exclusive handling to occur in
> copy_nonpresent_pte().
> ---
> mm/memory.c | 12 +++++++-----
> 1 file changed, 7 insertions(+), 5 deletions(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 2fb455c365c2..e061cfa18c11 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -718,7 +718,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
>
> if (likely(!non_swap_entry(entry))) {
> if (swap_duplicate(entry) < 0)
> - return entry.val;
> + return -EAGAIN;
>
> /* make sure dst_mm is on swapoff's mmlist. */
> if (unlikely(list_empty(&dst_mm->mmlist))) {
> @@ -974,11 +974,13 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
> continue;
> }
> if (unlikely(!pte_present(*src_pte))) {
> - entry.val = copy_nonpresent_pte(dst_mm, src_mm,
> - dst_pte, src_pte,
> - src_vma, addr, rss);
> - if (entry.val)
> + ret = copy_nonpresent_pte(dst_mm, src_mm,
> + dst_pte, src_pte,
> + src_vma, addr, rss);
> + if (ret == -EAGAIN) {
> + entry = pte_to_swp_entry(*src_pte);
> break;
> + }
> progress += 8;
> continue;
> }

Note that -EAGAIN was previously used by copy_present_page() for early cow
use. Here later although we check entry.val first:

if (entry.val) {
if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
ret = -ENOMEM;
goto out;
}
entry.val = 0;
} else if (ret) {
WARN_ON_ONCE(ret != -EAGAIN);
prealloc = page_copy_prealloc(src_mm, src_vma, addr);
if (!prealloc)
return -ENOMEM;
/* We've captured and resolved the error. Reset, try again. */
ret = 0;
}

We didn't reset "ret" in entry.val case (maybe we should?). Then in the next
round of "goto again" if "ret" is unluckily untouched, it could reach the 2nd
if check, and I think it could cause an unexpected page_copy_prealloc().

--
Peter Xu

2021-05-27 00:32:43

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Mon, May 24, 2021 at 11:27:22PM +1000, Alistair Popple wrote:
> Some devices require exclusive write access to shared virtual
> memory (SVM) ranges to perform atomic operations on that memory. This
> requires CPU page tables to be updated to deny access whilst atomic
> operations are occurring.
>
> In order to do this introduce a new swap entry
> type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for
> exclusive access by a device all page table mappings for the particular
> range are replaced with device exclusive swap entries. This causes any
> CPU access to the page to result in a fault.
>
> Faults are resovled by replacing the faulting entry with the original
> mapping. This results in MMU notifiers being called which a driver uses
> to update access permissions such as revoking atomic access. After
> notifiers have been called the device will no longer have exclusive
> access to the region.
>
> Walking of the page tables to find the target pages is handled by
> get_user_pages() rather than a direct page table walk. A direct page
> table walk similar to what migrate_vma_collect()/unmap() does could also
> have been utilised. However this resulted in more code similar in
> functionality to what get_user_pages() provides as page faulting is
> required to make the PTEs present and to break COW.
>
> Signed-off-by: Alistair Popple <[email protected]>
> Reviewed-by: Christoph Hellwig <[email protected]>
>
> ---
>
> v9:
> * Split rename of migrate_pgmap_owner into a separate patch.
> * Added comments explaining SWP_DEVICE_EXCLUSIVE_* entries.
> * Renamed try_to_protect{_one} to page_make_device_exclusive{_one} based
> somewhat on a suggestion from Peter Xu. I was never particularly happy
> with try_to_protect() as a name so think this is better.
> * Removed unneccesary code and reworded some comments based on feedback
> from Peter Xu.
> * Removed the VMA walk when restoring PTEs for device-exclusive entries.
> * Simplified implementation of copy_pte_range() to fail if the page
> cannot be locked. This might lead to occasional fork() failures but at
> this stage we don't think that will be an issue.
>
> v8:
> * Remove device exclusive entries on fork rather than copy them.
>
> v7:
> * Added Christoph's Reviewed-by.
> * Minor cosmetic cleanups suggested by Christoph.
> * Replace mmu_notifier_range_init_migrate/exclusive with
> mmu_notifier_range_init_owner as suggested by Christoph.
> * Replaced lock_page() with lock_page_retry() when handling faults.
> * Restrict to anonymous pages for now.
>
> v6:
> * Fixed a bisectablity issue due to incorrectly applying the rename of
> migrate_pgmap_owner to the wrong patches for Nouveau and hmm_test.
>
> v5:
> * Renamed range->migrate_pgmap_owner to range->owner.
> * Added MMU_NOTIFY_EXCLUSIVE to allow passing of a driver cookie which
> allows notifiers called as a result of make_device_exclusive_range() to
> be ignored.
> * Added a check to try_to_protect_one() to detect if the pages originally
> returned from get_user_pages() have been unmapped or not.
> * Removed check_device_exclusive_range() as it is no longer required with
> the other changes.
> * Documentation update.
>
> v4:
> * Add function to check that mappings are still valid and exclusive.
> * s/long/unsigned long/ in make_device_exclusive_entry().
> ---
> Documentation/vm/hmm.rst | 17 ++++
> include/linux/mmu_notifier.h | 6 ++
> include/linux/rmap.h | 4 +
> include/linux/swap.h | 7 +-
> include/linux/swapops.h | 44 ++++++++-
> mm/hmm.c | 5 +
> mm/memory.c | 128 +++++++++++++++++++++++-
> mm/mprotect.c | 8 ++
> mm/page_vma_mapped.c | 9 +-
> mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
> 10 files changed, 405 insertions(+), 9 deletions(-)
>
> diff --git a/Documentation/vm/hmm.rst b/Documentation/vm/hmm.rst
> index 3df79307a797..a14c2938e7af 100644
> --- a/Documentation/vm/hmm.rst
> +++ b/Documentation/vm/hmm.rst
> @@ -405,6 +405,23 @@ between device driver specific code and shared common code:
>
> The lock can now be released.
>
> +Exclusive access memory
> +=======================
> +
> +Some devices have features such as atomic PTE bits that can be used to implement
> +atomic access to system memory. To support atomic operations to a shared virtual
> +memory page such a device needs access to that page which is exclusive of any
> +userspace access from the CPU. The ``make_device_exclusive_range()`` function
> +can be used to make a memory range inaccessible from userspace.
> +
> +This replaces all mappings for pages in the given range with special swap
> +entries. Any attempt to access the swap entry results in a fault which is
> +resovled by replacing the entry with the original mapping. A driver gets
> +notified that the mapping has been changed by MMU notifiers, after which point
> +it will no longer have exclusive access to the page. Exclusive access is
> +guranteed to last until the driver drops the page lock and page reference, at
> +which point any CPU faults on the page may proceed as described.
> +
> Memory cgroup (memcg) and rss accounting
> ========================================
>
> diff --git a/include/linux/mmu_notifier.h b/include/linux/mmu_notifier.h
> index 8e428eb813b8..d049e0f6f756 100644
> --- a/include/linux/mmu_notifier.h
> +++ b/include/linux/mmu_notifier.h
> @@ -42,6 +42,11 @@ struct mmu_interval_notifier;
> * @MMU_NOTIFY_MIGRATE: used during migrate_vma_collect() invalidate to signal
> * a device driver to possibly ignore the invalidation if the
> * owner field matches the driver's device private pgmap owner.
> + *
> + * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
> + * longer have exclusive access to the page. May ignore the invalidation that's
> + * part of make_device_exclusive_range() if the owner field
> + * matches the value passed to make_device_exclusive_range().

Perhaps s/matches/does not match/?

> */
> enum mmu_notifier_event {
> MMU_NOTIFY_UNMAP = 0,
> @@ -51,6 +56,7 @@ enum mmu_notifier_event {
> MMU_NOTIFY_SOFT_DIRTY,
> MMU_NOTIFY_RELEASE,
> MMU_NOTIFY_MIGRATE,
> + MMU_NOTIFY_EXCLUSIVE,
> };
>
> #define MMU_NOTIFIER_RANGE_BLOCKABLE (1 << 0)
> diff --git a/include/linux/rmap.h b/include/linux/rmap.h
> index 0e25d829f742..3a1ce4ef9276 100644
> --- a/include/linux/rmap.h
> +++ b/include/linux/rmap.h
> @@ -193,6 +193,10 @@ int page_referenced(struct page *, int is_locked,
> bool try_to_migrate(struct page *page, enum ttu_flags flags);
> bool try_to_unmap(struct page *, enum ttu_flags flags);
>
> +int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
> + unsigned long end, struct page **pages,
> + void *arg);
> +
> /* Avoid racy checks */
> #define PVMW_SYNC (1 << 0)
> /* Look for migarion entries rather than present PTEs */
> diff --git a/include/linux/swap.h b/include/linux/swap.h
> index a6d4505ecf73..306df39d7c67 100644
> --- a/include/linux/swap.h
> +++ b/include/linux/swap.h
> @@ -63,11 +63,16 @@ static inline int current_is_kswapd(void)
> *
> * When a page is migrated from CPU to device, we set the CPU page table entry
> * to a special SWP_DEVICE_* entry.

s/SWP_DEVICE_*/SWP_DEVICE_{READ|WRITE}/? Since SWP_DEVICE_* covers all four
too.

> + *
> + * When a page is mapped by the device for exclusive access we set the CPU page
> + * table entries to special SWP_DEVICE_EXCLUSIVE_* entries.
> */
> #ifdef CONFIG_DEVICE_PRIVATE
> -#define SWP_DEVICE_NUM 2
> +#define SWP_DEVICE_NUM 4
> #define SWP_DEVICE_WRITE (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM)
> #define SWP_DEVICE_READ (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+1)
> +#define SWP_DEVICE_EXCLUSIVE_WRITE (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+2)
> +#define SWP_DEVICE_EXCLUSIVE_READ (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+3)
> #else
> #define SWP_DEVICE_NUM 0
> #endif
> diff --git a/include/linux/swapops.h b/include/linux/swapops.h
> index 4dfd807ae52a..4129bd2ff9d6 100644
> --- a/include/linux/swapops.h
> +++ b/include/linux/swapops.h
> @@ -120,6 +120,27 @@ static inline bool is_writable_device_private_entry(swp_entry_t entry)
> {
> return unlikely(swp_type(entry) == SWP_DEVICE_WRITE);
> }
> +
> +static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t offset)
> +{
> + return swp_entry(SWP_DEVICE_EXCLUSIVE_READ, offset);
> +}
> +
> +static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t offset)
> +{
> + return swp_entry(SWP_DEVICE_EXCLUSIVE_WRITE, offset);
> +}
> +
> +static inline bool is_device_exclusive_entry(swp_entry_t entry)
> +{
> + return swp_type(entry) == SWP_DEVICE_EXCLUSIVE_READ ||
> + swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE;
> +}
> +
> +static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
> +{
> + return unlikely(swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE);
> +}
> #else /* CONFIG_DEVICE_PRIVATE */
> static inline swp_entry_t make_readable_device_private_entry(pgoff_t offset)
> {
> @@ -140,6 +161,26 @@ static inline bool is_writable_device_private_entry(swp_entry_t entry)
> {
> return false;
> }
> +
> +static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t offset)
> +{
> + return swp_entry(0, 0);
> +}
> +
> +static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t offset)
> +{
> + return swp_entry(0, 0);
> +}
> +
> +static inline bool is_device_exclusive_entry(swp_entry_t entry)
> +{
> + return false;
> +}
> +
> +static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
> +{
> + return false;
> +}
> #endif /* CONFIG_DEVICE_PRIVATE */
>
> #ifdef CONFIG_MIGRATION
> @@ -219,7 +260,8 @@ static inline struct page *pfn_swap_entry_to_page(swp_entry_t entry)
> */
> static inline bool is_pfn_swap_entry(swp_entry_t entry)
> {
> - return is_migration_entry(entry) || is_device_private_entry(entry);
> + return is_migration_entry(entry) || is_device_private_entry(entry) ||
> + is_device_exclusive_entry(entry);
> }
>
> struct page_vma_mapped_walk;
> diff --git a/mm/hmm.c b/mm/hmm.c
> index 11df3ca30b82..fad6be2bf072 100644
> --- a/mm/hmm.c
> +++ b/mm/hmm.c
> @@ -26,6 +26,8 @@
> #include <linux/mmu_notifier.h>
> #include <linux/memory_hotplug.h>
>
> +#include "internal.h"
> +
> struct hmm_vma_walk {
> struct hmm_range *range;
> unsigned long last;
> @@ -271,6 +273,9 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
> if (!non_swap_entry(entry))
> goto fault;
>
> + if (is_device_exclusive_entry(entry))
> + goto fault;
> +
> if (is_migration_entry(entry)) {
> pte_unmap(ptep);
> hmm_vma_walk->last = addr;
> diff --git a/mm/memory.c b/mm/memory.c
> index e061cfa18c11..c1d2d732f189 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -700,6 +700,68 @@ struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
> }
> #endif
>
> +static void restore_exclusive_pte(struct vm_area_struct *vma,
> + struct page *page, unsigned long address,
> + pte_t *ptep)
> +{
> + pte_t pte;
> + swp_entry_t entry;
> +
> + pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
> + if (pte_swp_soft_dirty(*ptep))
> + pte = pte_mksoft_dirty(pte);
> +
> + entry = pte_to_swp_entry(*ptep);
> + if (pte_swp_uffd_wp(*ptep))
> + pte = pte_mkuffd_wp(pte);
> + else if (is_writable_device_exclusive_entry(entry))
> + pte = maybe_mkwrite(pte_mkdirty(pte), vma);
> +
> + set_pte_at(vma->vm_mm, address, ptep, pte);
> +
> + /*
> + * No need to take a page reference as one was already
> + * created when the swap entry was made.
> + */
> + if (PageAnon(page))
> + page_add_anon_rmap(page, vma, address, false);
> + else
> + /*
> + * Currently device exclusive access only supports anonymous
> + * memory so the entry shouldn't point to a filebacked page.
> + */
> + WARN_ON_ONCE(!PageAnon(page));
> +
> + if (vma->vm_flags & VM_LOCKED)
> + mlock_vma_page(page);
> +
> + /*
> + * No need to invalidate - it was non-present before. However
> + * secondary CPUs may have mappings that need invalidating.
> + */
> + update_mmu_cache(vma, address, ptep);
> +}
> +
> +/*
> + * Tries to restore an exclusive pte if the page lock can be acquired without
> + * sleeping.
> + */
> +static unsigned long

Better return a int?

> +try_restore_exclusive_pte(struct mm_struct *src_mm, pte_t *src_pte,
> + struct vm_area_struct *vma, unsigned long addr)

Raised in the other thread too: src_mm can be dropped.

> +{
> + swp_entry_t entry = pte_to_swp_entry(*src_pte);
> + struct page *page = pfn_swap_entry_to_page(entry);
> +
> + if (trylock_page(page)) {
> + restore_exclusive_pte(vma, page, addr, src_pte);
> + unlock_page(page);
> + return 0;
> + }
> +
> + return -EBUSY;
> +}
> +
> /*
> * copy one vm_area from one task to the other. Assumes the page tables
> * already present in the new task to be cleared in the whole range
> @@ -781,6 +843,17 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> pte = pte_swp_mkuffd_wp(pte);
> set_pte_at(src_mm, addr, src_pte, pte);
> }
> + } else if (is_device_exclusive_entry(entry)) {
> + /*
> + * Make device exclusive entries present by restoring the
> + * original entry then copying as for a present pte. Device
> + * exclusive entries currently only support private writable
> + * (ie. COW) mappings.
> + */
> + VM_BUG_ON(!is_cow_mapping(vma->vm_flags));
> + if (try_restore_exclusive_pte(src_mm, src_pte, vma, addr))
> + return -EBUSY;
> + return -ENOENT;
> }
> set_pte_at(dst_mm, addr, dst_pte, pte);
> return 0;
> @@ -980,9 +1053,18 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
> if (ret == -EAGAIN) {
> entry = pte_to_swp_entry(*src_pte);
> break;
> + } else if (ret == -EBUSY) {
> + break;
> + } else if (!ret) {
> + progress += 8;
> + continue;
> }
> - progress += 8;
> - continue;
> +
> + /*
> + * Device exclusive entry restored, continue by copying
> + * the now present pte.
> + */
> + WARN_ON_ONCE(ret != -ENOENT);

The change looks right, thanks. It's just that we should start to consider
document all these err code now in copy_pte_range() some day (perhaps on top of
this patch)..

> }
> /* copy_present_pte() will clear `*prealloc' if consumed */
> ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
> @@ -1019,6 +1101,8 @@ copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
> goto out;
> }
> entry.val = 0;
> + } else if (ret == -EBUSY) {
> + return -EBUSY;
> } else if (ret) {
> WARN_ON_ONCE(ret != -EAGAIN);
> prealloc = page_copy_prealloc(src_mm, src_vma, addr);
> @@ -1283,7 +1367,8 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
> }
>
> entry = pte_to_swp_entry(ptent);
> - if (is_device_private_entry(entry)) {
> + if (is_device_private_entry(entry) ||
> + is_device_exclusive_entry(entry)) {
> struct page *page = pfn_swap_entry_to_page(entry);
>
> if (unlikely(details && details->check_mapping)) {
> @@ -1299,7 +1384,10 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
>
> pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
> rss[mm_counter(page)]--;
> - page_remove_rmap(page, false);
> +
> + if (is_device_private_entry(entry))
> + page_remove_rmap(page, false);
> +
> put_page(page);
> continue;
> }
> @@ -3303,6 +3391,35 @@ void unmap_mapping_range(struct address_space *mapping,
> }
> EXPORT_SYMBOL(unmap_mapping_range);
>
> +/*
> + * Restore a potential device exclusive pte to a working pte entry
> + */
> +static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
> +{
> + struct page *page = vmf->page;
> + struct vm_area_struct *vma = vmf->vma;
> + vm_fault_t ret = 0;
> + struct mmu_notifier_range range;
> +
> + if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
> + return VM_FAULT_RETRY;
> + mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
> + vmf->address & PAGE_MASK,
> + (vmf->address & PAGE_MASK) + PAGE_SIZE);

@MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
longer have exclusive access to the page.

Shouldn't this be the place to use new MMU_NOTIFY_EXCLUSIVE?

> + mmu_notifier_invalidate_range_start(&range);
> +
> + vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
> + &vmf->ptl);
> + if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
> + restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
> +
> + pte_unmap_unlock(vmf->pte, vmf->ptl);
> + unlock_page(page);
> +
> + mmu_notifier_invalidate_range_end(&range);
> + return ret;

We can drop "ret" and return 0 here directly.

> +}
> +
> /*
> * We enter with non-exclusive mmap_lock (to exclude vma changes,
> * but allow concurrent faults), and pte mapped but not yet locked.
> @@ -3330,6 +3447,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
> if (is_migration_entry(entry)) {
> migration_entry_wait(vma->vm_mm, vmf->pmd,
> vmf->address);
> + } else if (is_device_exclusive_entry(entry)) {
> + vmf->page = pfn_swap_entry_to_page(entry);
> + ret = remove_device_exclusive_entry(vmf);
> } else if (is_device_private_entry(entry)) {
> vmf->page = pfn_swap_entry_to_page(entry);
> ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
> diff --git a/mm/mprotect.c b/mm/mprotect.c
> index ee5961888e70..883e2cc85cad 100644
> --- a/mm/mprotect.c
> +++ b/mm/mprotect.c
> @@ -165,6 +165,14 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
> newpte = swp_entry_to_pte(entry);
> if (pte_swp_uffd_wp(oldpte))
> newpte = pte_swp_mkuffd_wp(newpte);
> + } else if (is_writable_device_exclusive_entry(entry)) {
> + entry = make_readable_device_exclusive_entry(
> + swp_offset(entry));
> + newpte = swp_entry_to_pte(entry);
> + if (pte_swp_soft_dirty(oldpte))
> + newpte = pte_swp_mksoft_dirty(newpte);
> + if (pte_swp_uffd_wp(oldpte))
> + newpte = pte_swp_mkuffd_wp(newpte);
> } else {
> newpte = oldpte;
> }
> diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> index a6a7febb4d93..f535bcb4950c 100644
> --- a/mm/page_vma_mapped.c
> +++ b/mm/page_vma_mapped.c
> @@ -41,7 +41,8 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw)
>
> /* Handle un-addressable ZONE_DEVICE memory */
> entry = pte_to_swp_entry(*pvmw->pte);
> - if (!is_device_private_entry(entry))
> + if (!is_device_private_entry(entry) &&
> + !is_device_exclusive_entry(entry))
> return false;
> } else if (!pte_present(*pvmw->pte))
> return false;
> @@ -93,7 +94,8 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
> return false;
> entry = pte_to_swp_entry(*pvmw->pte);
>
> - if (!is_migration_entry(entry))
> + if (!is_migration_entry(entry) &&
> + !is_device_exclusive_entry(entry))
> return false;
>
> pfn = swp_offset(entry);
> @@ -102,7 +104,8 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
>
> /* Handle un-addressable ZONE_DEVICE memory */
> entry = pte_to_swp_entry(*pvmw->pte);
> - if (!is_device_private_entry(entry))
> + if (!is_device_private_entry(entry) &&
> + !is_device_exclusive_entry(entry))
> return false;
>
> pfn = swp_offset(entry);
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 8ed1853060cf..fe062f63ef4d 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -2008,6 +2008,192 @@ void page_mlock(struct page *page)
> rmap_walk(page, &rwc);
> }
>
> +struct make_exclusive_args {
> + struct mm_struct *mm;
> + unsigned long address;
> + void *owner;
> + bool valid;
> +};
> +
> +static bool page_make_device_exclusive_one(struct page *page,
> + struct vm_area_struct *vma, unsigned long address, void *priv)
> +{
> + struct mm_struct *mm = vma->vm_mm;
> + struct page_vma_mapped_walk pvmw = {
> + .page = page,
> + .vma = vma,
> + .address = address,
> + };
> + struct make_exclusive_args *args = priv;
> + pte_t pteval;
> + struct page *subpage;
> + bool ret = true;
> + struct mmu_notifier_range range;
> + swp_entry_t entry;
> + pte_t swp_pte;
> +
> + mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,

Similar question here, EXCLUSIVE comment says it gets notified when the device
does not have exclusive access.

If you prefer to keep using EXCLUSIVE for both mark/restore, then we need to
change the comment above MMU_NOTIFY_EXCLUSIVE?

> + vma->vm_mm, address, min(vma->vm_end,
> + address + page_size(page)), args->owner);
> + mmu_notifier_invalidate_range_start(&range);
> +
> + while (page_vma_mapped_walk(&pvmw)) {
> + /* Unexpected PMD-mapped THP? */
> + VM_BUG_ON_PAGE(!pvmw.pte, page);
> +
> + if (!pte_present(*pvmw.pte)) {
> + ret = false;
> + page_vma_mapped_walk_done(&pvmw);
> + break;
> + }
> +
> + subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);

I see that all pages passed in should be done after FOLL_SPLIT_PMD, so is this
needed? Or say, should subpage==page always be true?

> + address = pvmw.address;
> +
> + /* Nuke the page table entry. */
> + flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
> + pteval = ptep_clear_flush(vma, address, pvmw.pte);
> +
> + /* Move the dirty bit to the page. Now the pte is gone. */
> + if (pte_dirty(pteval))
> + set_page_dirty(page);
> +
> + if (arch_unmap_one(mm, vma, address, pteval) < 0) {
> + set_pte_at(mm, address, pvmw.pte, pteval);
> + ret = false;
> + page_vma_mapped_walk_done(&pvmw);
> + break;
> + }

Didn't notice this previously, but also suggest to drop this.

Two reasons:

1. It's introduced in ca827d55ebaa ("mm, swap: Add infrastructure for saving
page metadata on swap", 2018-03-18) for sparc-only use so far. If we really
want this, we'll also want to call arch_do_swap_page() when restoring the
pte just like what we do in do_swap_page(); NOTE: current code path of
SWP_DEVICE_EXCLUSIVE will skip the arch_do_swap_page() in do_swap_page() so
it's not even paired with the above arch_unmap_one(), so I believe this
won't even work for sparc at all.

2. I highly doubt whether sparc is also on the list of platforms to support for
device atomic ops even in the future. IMHO we'd better not copy-paste code
clips if never used at all, because once merged, removing it would need more
justifications.

> +
> + /*
> + * Check that our target page is still mapped at the expected
> + * address.
> + */
> + if (args->mm == mm && args->address == address &&
> + pte_write(pteval))
> + args->valid = true;
> +
> + /*
> + * Store the pfn of the page in a special migration
> + * pte. do_swap_page() will wait until the migration
> + * pte is removed and then restart fault handling.
> + */
> + if (pte_write(pteval))
> + entry = make_writable_device_exclusive_entry(
> + page_to_pfn(subpage));
> + else
> + entry = make_readable_device_exclusive_entry(
> + page_to_pfn(subpage));
> + swp_pte = swp_entry_to_pte(entry);
> + if (pte_soft_dirty(pteval))
> + swp_pte = pte_swp_mksoft_dirty(swp_pte);
> + if (pte_uffd_wp(pteval))
> + swp_pte = pte_swp_mkuffd_wp(swp_pte);
> +
> + /* Take a reference for the swap entry */
> + get_page(page);
> + set_pte_at(mm, address, pvmw.pte, swp_pte);
> +
> + page_remove_rmap(subpage, PageHuge(page));

Why PageHuge()? Should it be a constant "false"?

> + put_page(page);

Should we drop this put_page() along with get_page() above?

page_count() should be >0 anyway as we've got a mapcount before at least when
dropping the pte. Then IMHO we can simply keep the old page reference.

> + }
> +
> + mmu_notifier_invalidate_range_end(&range);
> +
> + return ret;
> +}
> +
> +/**
> + * page_make_device_exclusive - replace page table mappings with swap entries

"with swap entries" looks a bit blurred to me (although below longer comment
explains much better). How about below (or something similar):

page_make_device_exclusive - Mark the page exclusively owned by the device

?

It'll also match with comment above make_device_exclusive_range().

No strong opinion.

The rest looks good. Thanks,

> + * @page: the page to replace page table entries for
> + * @mm: the mm_struct where the page is expected to be mapped
> + * @address: address where the page is expected to be mapped
> + * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
> + *
> + * Tries to remove all the page table entries which are mapping this page and
> + * replace them with special device exclusive swap entries to grant a device
> + * exclusive access to the page. Caller must hold the page lock.
> + *
> + * Returns false if the page is still mapped, or if it could not be unmapped
> + * from the expected address. Otherwise returns true (success).
> + */
> +static bool page_make_device_exclusive(struct page *page, struct mm_struct *mm,
> + unsigned long address, void *owner)
> +{
> + struct make_exclusive_args args = {
> + .mm = mm,
> + .address = address,
> + .owner = owner,
> + .valid = false,
> + };
> + struct rmap_walk_control rwc = {
> + .rmap_one = page_make_device_exclusive_one,
> + .done = page_not_mapped,
> + .anon_lock = page_lock_anon_vma_read,
> + .arg = &args,
> + };
> +
> + /*
> + * Restrict to anonymous pages for now to avoid potential writeback
> + * issues.
> + */
> + if (!PageAnon(page))
> + return false;
> +
> + rmap_walk(page, &rwc);
> +
> + return args.valid && !page_mapcount(page);
> +}
> +
> +/**
> + * make_device_exclusive_range() - Mark a range for exclusive use by a device
> + * @mm: mm_struct of assoicated target process
> + * @start: start of the region to mark for exclusive device access
> + * @end: end address of region
> + * @pages: returns the pages which were successfully marked for exclusive access
> + * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow filtering
> + *
> + * Returns: number of pages found in the range by GUP. A page is marked for
> + * exclusive access only if the page pointer is non-NULL.
> + *
> + * This function finds ptes mapping page(s) to the given address range, locks
> + * them and replaces mappings with special swap entries preventing userspace CPU
> + * access. On fault these entries are replaced with the original mapping after
> + * calling MMU notifiers.
> + *
> + * A driver using this to program access from a device must use a mmu notifier
> + * critical section to hold a device specific lock during programming. Once
> + * programming is complete it should drop the page lock and reference after
> + * which point CPU access to the page will revoke the exclusive access.
> + */
> +int make_device_exclusive_range(struct mm_struct *mm, unsigned long start,
> + unsigned long end, struct page **pages,
> + void *owner)
> +{
> + unsigned long npages = (end - start) >> PAGE_SHIFT;
> + unsigned long i;
> +
> + npages = get_user_pages_remote(mm, start, npages,
> + FOLL_GET | FOLL_WRITE | FOLL_SPLIT_PMD,
> + pages, NULL, NULL);
> + for (i = 0; i < npages; i++, start += PAGE_SIZE) {
> + if (!trylock_page(pages[i])) {
> + put_page(pages[i]);
> + pages[i] = NULL;
> + continue;
> + }
> +
> + if (!page_make_device_exclusive(pages[i], mm, start, owner)) {
> + unlock_page(pages[i]);
> + put_page(pages[i]);
> + pages[i] = NULL;
> + }
> + }
> +
> + return npages;
> +}
> +EXPORT_SYMBOL_GPL(make_device_exclusive_range);
> +
> void __put_anon_vma(struct anon_vma *anon_vma)
> {
> struct anon_vma *root = anon_vma->root;
> --
> 2.20.1
>

--
Peter Xu

2021-05-27 02:00:30

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 06/10] mm/memory.c: Allow different return codes for copy_nonpresent_pte()

On Thursday, 27 May 2021 5:50:05 AM AEST Peter Xu wrote:
> On Mon, May 24, 2021 at 11:27:21PM +1000, Alistair Popple wrote:
> > Currently if copy_nonpresent_pte() returns a non-zero value it is
> > assumed to be a swap entry which requires further processing outside the
> > loop in copy_pte_range() after dropping locks. This prevents other
> > values being returned to signal conditions such as failure which a
> > subsequent change requires.
> >
> > Instead make copy_nonpresent_pte() return an error code if further
> > processing is required and read the value for the swap entry in the main
> > loop under the ptl.
> >
> > Signed-off-by: Alistair Popple <[email protected]>
> >
> > ---
> >
> > v9:
> >
> > New for v9 to allow device exclusive handling to occur in
> > copy_nonpresent_pte().
> > ---
> >
> > mm/memory.c | 12 +++++++-----
> > 1 file changed, 7 insertions(+), 5 deletions(-)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 2fb455c365c2..e061cfa18c11 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -718,7 +718,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct
> > mm_struct *src_mm,>
> > if (likely(!non_swap_entry(entry))) {
> >
> > if (swap_duplicate(entry) < 0)
> >
> > - return entry.val;
> > + return -EAGAIN;
> >
> > /* make sure dst_mm is on swapoff's mmlist. */
> > if (unlikely(list_empty(&dst_mm->mmlist))) {
> >
> > @@ -974,11 +974,13 @@ copy_pte_range(struct vm_area_struct *dst_vma,
> > struct vm_area_struct *src_vma,>
> > continue;
> >
> > }
> > if (unlikely(!pte_present(*src_pte))) {
> >
> > - entry.val = copy_nonpresent_pte(dst_mm, src_mm,
> > - dst_pte, src_pte,
> > - src_vma, addr, rss);
> > - if (entry.val)
> > + ret = copy_nonpresent_pte(dst_mm, src_mm,
> > + dst_pte, src_pte,
> > + src_vma, addr, rss);
> > + if (ret == -EAGAIN) {
> > + entry = pte_to_swp_entry(*src_pte);
> >
> > break;
> >
> > + }
> >
> > progress += 8;
> > continue;
> >
> > }
>
> Note that -EAGAIN was previously used by copy_present_page() for early cow
> use. Here later although we check entry.val first:
>
> if (entry.val) {
> if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
> ret = -ENOMEM;
> goto out;
> }
> entry.val = 0;
> } else if (ret) {
> WARN_ON_ONCE(ret != -EAGAIN);
> prealloc = page_copy_prealloc(src_mm, src_vma, addr);
> if (!prealloc)
> return -ENOMEM;
> /* We've captured and resolved the error. Reset, try again.
> */ ret = 0;
> }
>
> We didn't reset "ret" in entry.val case (maybe we should?). Then in the next
> round of "goto again" if "ret" is unluckily untouched, it could reach the
> 2nd if check, and I think it could cause an unexpected
> page_copy_prealloc().

Thanks, I had considered that but saw "ret" was always set either by
copy_nonpresent_pte() or copy_present_pte(). However missed the "unlucky" case
at the start of the loop:

if (progress >= 32) {
progress = 0;
if (need_resched() ||
spin_needbreak(src_ptl) || pin_needbreak(dst_ptl))
break;

Looking at this again though checking different variables to figure out what
to do outside the locks and reusing error codes seems error prone. I reused -
EAGAIN for copy_nonpresent_pte() simply because that seemed the most sensible
error code, but I don't think that aids readability and it might be better to
use a unique error code for each case needing extra handling.

So it might be better if I update this patch to:
1) Use unique error codes for each case requiring special handling outside the
lock.
2) Only check "ret" to determine what to do outside locks (ie. not entry.val)
3) Document these.
4) Always reset ret after handling.

Thoughts?

- Alistair

> --
> Peter Xu




2021-05-27 02:15:42

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 06/10] mm/memory.c: Allow different return codes for copy_nonpresent_pte()

On Thu, May 27, 2021 at 11:20:36AM +1000, Alistair Popple wrote:
> On Thursday, 27 May 2021 5:50:05 AM AEST Peter Xu wrote:
> > On Mon, May 24, 2021 at 11:27:21PM +1000, Alistair Popple wrote:
> > > Currently if copy_nonpresent_pte() returns a non-zero value it is
> > > assumed to be a swap entry which requires further processing outside the
> > > loop in copy_pte_range() after dropping locks. This prevents other
> > > values being returned to signal conditions such as failure which a
> > > subsequent change requires.
> > >
> > > Instead make copy_nonpresent_pte() return an error code if further
> > > processing is required and read the value for the swap entry in the main
> > > loop under the ptl.
> > >
> > > Signed-off-by: Alistair Popple <[email protected]>
> > >
> > > ---
> > >
> > > v9:
> > >
> > > New for v9 to allow device exclusive handling to occur in
> > > copy_nonpresent_pte().
> > > ---
> > >
> > > mm/memory.c | 12 +++++++-----
> > > 1 file changed, 7 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index 2fb455c365c2..e061cfa18c11 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -718,7 +718,7 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct
> > > mm_struct *src_mm,>
> > > if (likely(!non_swap_entry(entry))) {
> > >
> > > if (swap_duplicate(entry) < 0)
> > >
> > > - return entry.val;
> > > + return -EAGAIN;
> > >
> > > /* make sure dst_mm is on swapoff's mmlist. */
> > > if (unlikely(list_empty(&dst_mm->mmlist))) {
> > >
> > > @@ -974,11 +974,13 @@ copy_pte_range(struct vm_area_struct *dst_vma,
> > > struct vm_area_struct *src_vma,>
> > > continue;
> > >
> > > }
> > > if (unlikely(!pte_present(*src_pte))) {
> > >
> > > - entry.val = copy_nonpresent_pte(dst_mm, src_mm,
> > > - dst_pte, src_pte,
> > > - src_vma, addr, rss);
> > > - if (entry.val)
> > > + ret = copy_nonpresent_pte(dst_mm, src_mm,
> > > + dst_pte, src_pte,
> > > + src_vma, addr, rss);
> > > + if (ret == -EAGAIN) {
> > > + entry = pte_to_swp_entry(*src_pte);
> > >
> > > break;
> > >
> > > + }
> > >
> > > progress += 8;
> > > continue;
> > >
> > > }
> >
> > Note that -EAGAIN was previously used by copy_present_page() for early cow
> > use. Here later although we check entry.val first:
> >
> > if (entry.val) {
> > if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
> > ret = -ENOMEM;
> > goto out;
> > }
> > entry.val = 0;
> > } else if (ret) {
> > WARN_ON_ONCE(ret != -EAGAIN);
> > prealloc = page_copy_prealloc(src_mm, src_vma, addr);
> > if (!prealloc)
> > return -ENOMEM;
> > /* We've captured and resolved the error. Reset, try again.
> > */ ret = 0;
> > }
> >
> > We didn't reset "ret" in entry.val case (maybe we should?). Then in the next
> > round of "goto again" if "ret" is unluckily untouched, it could reach the
> > 2nd if check, and I think it could cause an unexpected
> > page_copy_prealloc().
>
> Thanks, I had considered that but saw "ret" was always set either by
> copy_nonpresent_pte() or copy_present_pte(). However missed the "unlucky" case
> at the start of the loop:
>
> if (progress >= 32) {
> progress = 0;
> if (need_resched() ||
> spin_needbreak(src_ptl) || pin_needbreak(dst_ptl))
> break;
>
> Looking at this again though checking different variables to figure out what
> to do outside the locks and reusing error codes seems error prone. I reused -
> EAGAIN for copy_nonpresent_pte() simply because that seemed the most sensible
> error code, but I don't think that aids readability and it might be better to
> use a unique error code for each case needing extra handling.
>
> So it might be better if I update this patch to:
> 1) Use unique error codes for each case requiring special handling outside the
> lock.
> 2) Only check "ret" to determine what to do outside locks (ie. not entry.val)
> 3) Document these.
> 4) Always reset ret after handling.
>
> Thoughts?

Looks good to me. Thanks,

--
Peter Xu

2021-05-27 06:14:17

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Thursday, 27 May 2021 5:28:32 AM AEST Peter Xu wrote:
> On Mon, May 24, 2021 at 11:27:22PM +1000, Alistair Popple wrote:
> > Some devices require exclusive write access to shared virtual
> > memory (SVM) ranges to perform atomic operations on that memory. This
> > requires CPU page tables to be updated to deny access whilst atomic
> > operations are occurring.
> >
> > In order to do this introduce a new swap entry
> > type (SWP_DEVICE_EXCLUSIVE). When a SVM range needs to be marked for
> > exclusive access by a device all page table mappings for the particular
> > range are replaced with device exclusive swap entries. This causes any
> > CPU access to the page to result in a fault.
> >
> > Faults are resovled by replacing the faulting entry with the original
> > mapping. This results in MMU notifiers being called which a driver uses
> > to update access permissions such as revoking atomic access. After
> > notifiers have been called the device will no longer have exclusive
> > access to the region.
> >
> > Walking of the page tables to find the target pages is handled by
> > get_user_pages() rather than a direct page table walk. A direct page
> > table walk similar to what migrate_vma_collect()/unmap() does could also
> > have been utilised. However this resulted in more code similar in
> > functionality to what get_user_pages() provides as page faulting is
> > required to make the PTEs present and to break COW.
> >
> > Signed-off-by: Alistair Popple <[email protected]>
> > Reviewed-by: Christoph Hellwig <[email protected]>
> >
> > ---
> >
> > v9:
> > * Split rename of migrate_pgmap_owner into a separate patch.
> > * Added comments explaining SWP_DEVICE_EXCLUSIVE_* entries.
> > * Renamed try_to_protect{_one} to page_make_device_exclusive{_one} based
> >
> > somewhat on a suggestion from Peter Xu. I was never particularly happy
> > with try_to_protect() as a name so think this is better.
> >
> > * Removed unneccesary code and reworded some comments based on feedback
> >
> > from Peter Xu.
> >
> > * Removed the VMA walk when restoring PTEs for device-exclusive entries.
> > * Simplified implementation of copy_pte_range() to fail if the page
> >
> > cannot be locked. This might lead to occasional fork() failures but at
> > this stage we don't think that will be an issue.
> >
> > v8:
> > * Remove device exclusive entries on fork rather than copy them.
> >
> > v7:
> > * Added Christoph's Reviewed-by.
> > * Minor cosmetic cleanups suggested by Christoph.
> > * Replace mmu_notifier_range_init_migrate/exclusive with
> >
> > mmu_notifier_range_init_owner as suggested by Christoph.
> >
> > * Replaced lock_page() with lock_page_retry() when handling faults.
> > * Restrict to anonymous pages for now.
> >
> > v6:
> > * Fixed a bisectablity issue due to incorrectly applying the rename of
> >
> > migrate_pgmap_owner to the wrong patches for Nouveau and hmm_test.
> >
> > v5:
> > * Renamed range->migrate_pgmap_owner to range->owner.
> > * Added MMU_NOTIFY_EXCLUSIVE to allow passing of a driver cookie which
> >
> > allows notifiers called as a result of make_device_exclusive_range() to
> > be ignored.
> >
> > * Added a check to try_to_protect_one() to detect if the pages originally
> >
> > returned from get_user_pages() have been unmapped or not.
> >
> > * Removed check_device_exclusive_range() as it is no longer required with
> >
> > the other changes.
> >
> > * Documentation update.
> >
> > v4:
> > * Add function to check that mappings are still valid and exclusive.
> > * s/long/unsigned long/ in make_device_exclusive_entry().
> > ---
> >
> > Documentation/vm/hmm.rst | 17 ++++
> > include/linux/mmu_notifier.h | 6 ++
> > include/linux/rmap.h | 4 +
> > include/linux/swap.h | 7 +-
> > include/linux/swapops.h | 44 ++++++++-
> > mm/hmm.c | 5 +
> > mm/memory.c | 128 +++++++++++++++++++++++-
> > mm/mprotect.c | 8 ++
> > mm/page_vma_mapped.c | 9 +-
> > mm/rmap.c | 186 +++++++++++++++++++++++++++++++++++
> > 10 files changed, 405 insertions(+), 9 deletions(-)
> >
> > diff --git a/Documentation/vm/hmm.rst b/Documentation/vm/hmm.rst
> > index 3df79307a797..a14c2938e7af 100644
> > --- a/Documentation/vm/hmm.rst
> > +++ b/Documentation/vm/hmm.rst
> >
> > @@ -405,6 +405,23 @@ between device driver specific code and shared common
code:
> > The lock can now be released.
> >
> > +Exclusive access memory
> > +=======================
> > +
> > +Some devices have features such as atomic PTE bits that can be used to
> > implement +atomic access to system memory. To support atomic operations
> > to a shared virtual +memory page such a device needs access to that page
> > which is exclusive of any +userspace access from the CPU. The
> > ``make_device_exclusive_range()`` function +can be used to make a memory
> > range inaccessible from userspace.
> > +
> > +This replaces all mappings for pages in the given range with special swap
> > +entries. Any attempt to access the swap entry results in a fault which is
> > +resovled by replacing the entry with the original mapping. A driver gets
> > +notified that the mapping has been changed by MMU notifiers, after which
> > point +it will no longer have exclusive access to the page. Exclusive
> > access is +guranteed to last until the driver drops the page lock and
> > page reference, at +which point any CPU faults on the page may proceed as
> > described.
> > +
> >
> > Memory cgroup (memcg) and rss accounting
> > ========================================
> >
> > diff --git a/include/linux/mmu_notifier.h b/include/linux/mmu_notifier.h
> > index 8e428eb813b8..d049e0f6f756 100644
> > --- a/include/linux/mmu_notifier.h
> > +++ b/include/linux/mmu_notifier.h
> > @@ -42,6 +42,11 @@ struct mmu_interval_notifier;
> >
> > * @MMU_NOTIFY_MIGRATE: used during migrate_vma_collect() invalidate to
> > signal * a device driver to possibly ignore the invalidation if the
> > * owner field matches the driver's device private pgmap owner.
> >
> > + *
> > + * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will
> > no + * longer have exclusive access to the page. May ignore the
> > invalidation that's + * part of make_device_exclusive_range() if the
> > owner field
> > + * matches the value passed to make_device_exclusive_range().
>
> Perhaps s/matches/does not match/?

No, "matches" is correct. The MMU_NOTIFY_EXCLUSIVE notifier is to notify a
listener that a range is being invalidated for the purpose of making the range
available for some device to have exclusive access to. Which does also mean a
device getting the notification no longer has exclusive access if it already
did.

A unique type is needed because when creating the range a driver needs to form
a mmu critical section (with mmu_interval_read_begin()/
mmu_interval_read_end()) to ensure the entry remains valid long enough to
program the device pte and hasn't been invalidated.

However without a way of filtering any invalidations will result in a retry,
but make_device_exclusive_range() needs to do an invalidation during
installation of the entry. To avoid this causing infinite retries the driver
ignores specific invalidation events that it knows don't apply, ie. the
invalidations that are a result of that driver asking for device exclusive
entries.

Agree the comment could be improved though.

> > */
> >
> > enum mmu_notifier_event {
> >
> > MMU_NOTIFY_UNMAP = 0,
> >
> > @@ -51,6 +56,7 @@ enum mmu_notifier_event {
> >
> > MMU_NOTIFY_SOFT_DIRTY,
> > MMU_NOTIFY_RELEASE,
> > MMU_NOTIFY_MIGRATE,
> >
> > + MMU_NOTIFY_EXCLUSIVE,
> >
> > };
> >
> > #define MMU_NOTIFIER_RANGE_BLOCKABLE (1 << 0)
> >
> > diff --git a/include/linux/rmap.h b/include/linux/rmap.h
> > index 0e25d829f742..3a1ce4ef9276 100644
> > --- a/include/linux/rmap.h
> > +++ b/include/linux/rmap.h
> > @@ -193,6 +193,10 @@ int page_referenced(struct page *, int is_locked,
> >
> > bool try_to_migrate(struct page *page, enum ttu_flags flags);
> > bool try_to_unmap(struct page *, enum ttu_flags flags);
> >
> > +int make_device_exclusive_range(struct mm_struct *mm, unsigned long
> > start,
> > + unsigned long end, struct page **pages,
> > + void *arg);
> > +
> >
> > /* Avoid racy checks */
> > #define PVMW_SYNC (1 << 0)
> > /* Look for migarion entries rather than present PTEs */
> >
> > diff --git a/include/linux/swap.h b/include/linux/swap.h
> > index a6d4505ecf73..306df39d7c67 100644
> > --- a/include/linux/swap.h
> > +++ b/include/linux/swap.h
> > @@ -63,11 +63,16 @@ static inline int current_is_kswapd(void)
> >
> > *
> > * When a page is migrated from CPU to device, we set the CPU page table
> > entry * to a special SWP_DEVICE_* entry.
>
> s/SWP_DEVICE_*/SWP_DEVICE_{READ|WRITE}/? Since SWP_DEVICE_* covers all four
> too.

Sure.

> > + *
> > + * When a page is mapped by the device for exclusive access we set the
> > CPU page + * table entries to special SWP_DEVICE_EXCLUSIVE_* entries.
> >
> > */
> >
> > #ifdef CONFIG_DEVICE_PRIVATE
> >
> > -#define SWP_DEVICE_NUM 2
> > +#define SWP_DEVICE_NUM 4
> >
> > #define SWP_DEVICE_WRITE
> > (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM)
> > #define SWP_DEVICE_READ
> > (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+1)>
> > +#define SWP_DEVICE_EXCLUSIVE_WRITE
> > (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+2) +#define
> > SWP_DEVICE_EXCLUSIVE_READ
> > (MAX_SWAPFILES+SWP_HWPOISON_NUM+SWP_MIGRATION_NUM+3)>
> > #else
> > #define SWP_DEVICE_NUM 0
> > #endif
> >
> > diff --git a/include/linux/swapops.h b/include/linux/swapops.h
> > index 4dfd807ae52a..4129bd2ff9d6 100644
> > --- a/include/linux/swapops.h
> > +++ b/include/linux/swapops.h
> > @@ -120,6 +120,27 @@ static inline bool
> > is_writable_device_private_entry(swp_entry_t entry)>
> > {
> >
> > return unlikely(swp_type(entry) == SWP_DEVICE_WRITE);
> >
> > }
> >
> > +
> > +static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t
> > offset) +{
> > + return swp_entry(SWP_DEVICE_EXCLUSIVE_READ, offset);
> > +}
> > +
> > +static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t
> > offset) +{
> > + return swp_entry(SWP_DEVICE_EXCLUSIVE_WRITE, offset);
> > +}
> > +
> > +static inline bool is_device_exclusive_entry(swp_entry_t entry)
> > +{
> > + return swp_type(entry) == SWP_DEVICE_EXCLUSIVE_READ ||
> > + swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE;
> > +}
> > +
> > +static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
> > +{
> > + return unlikely(swp_type(entry) == SWP_DEVICE_EXCLUSIVE_WRITE);
> > +}
> >
> > #else /* CONFIG_DEVICE_PRIVATE */
> > static inline swp_entry_t make_readable_device_private_entry(pgoff_t
> > offset) {
> >
> > @@ -140,6 +161,26 @@ static inline bool
> > is_writable_device_private_entry(swp_entry_t entry)>
> > {
> >
> > return false;
> >
> > }
> >
> > +
> > +static inline swp_entry_t make_readable_device_exclusive_entry(pgoff_t
> > offset) +{
> > + return swp_entry(0, 0);
> > +}
> > +
> > +static inline swp_entry_t make_writable_device_exclusive_entry(pgoff_t
> > offset) +{
> > + return swp_entry(0, 0);
> > +}
> > +
> > +static inline bool is_device_exclusive_entry(swp_entry_t entry)
> > +{
> > + return false;
> > +}
> > +
> > +static inline bool is_writable_device_exclusive_entry(swp_entry_t entry)
> > +{
> > + return false;
> > +}
> >
> > #endif /* CONFIG_DEVICE_PRIVATE */
> >
> > #ifdef CONFIG_MIGRATION
> >
> > @@ -219,7 +260,8 @@ static inline struct page
> > *pfn_swap_entry_to_page(swp_entry_t entry)>
> > */
> >
> > static inline bool is_pfn_swap_entry(swp_entry_t entry)
> > {
> >
> > - return is_migration_entry(entry) || is_device_private_entry(entry);
> > + return is_migration_entry(entry) || is_device_private_entry(entry)
> > ||
> > + is_device_exclusive_entry(entry);
> >
> > }
> >
> > struct page_vma_mapped_walk;
> >
> > diff --git a/mm/hmm.c b/mm/hmm.c
> > index 11df3ca30b82..fad6be2bf072 100644
> > --- a/mm/hmm.c
> > +++ b/mm/hmm.c
> > @@ -26,6 +26,8 @@
> >
> > #include <linux/mmu_notifier.h>
> > #include <linux/memory_hotplug.h>
> >
> > +#include "internal.h"
> > +
> >
> > struct hmm_vma_walk {
> >
> > struct hmm_range *range;
> > unsigned long last;
> >
> > @@ -271,6 +273,9 @@ static int hmm_vma_handle_pte(struct mm_walk *walk,
> > unsigned long addr,>
> > if (!non_swap_entry(entry))
> >
> > goto fault;
> >
> > + if (is_device_exclusive_entry(entry))
> > + goto fault;
> > +
> >
> > if (is_migration_entry(entry)) {
> >
> > pte_unmap(ptep);
> > hmm_vma_walk->last = addr;
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index e061cfa18c11..c1d2d732f189 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -700,6 +700,68 @@ struct page *vm_normal_page_pmd(struct vm_area_struct
> > *vma, unsigned long addr,>
> > }
> > #endif
> >
> > +static void restore_exclusive_pte(struct vm_area_struct *vma,
> > + struct page *page, unsigned long address,
> > + pte_t *ptep)
> > +{
> > + pte_t pte;
> > + swp_entry_t entry;
> > +
> > + pte = pte_mkold(mk_pte(page, READ_ONCE(vma->vm_page_prot)));
> > + if (pte_swp_soft_dirty(*ptep))
> > + pte = pte_mksoft_dirty(pte);
> > +
> > + entry = pte_to_swp_entry(*ptep);
> > + if (pte_swp_uffd_wp(*ptep))
> > + pte = pte_mkuffd_wp(pte);
> > + else if (is_writable_device_exclusive_entry(entry))
> > + pte = maybe_mkwrite(pte_mkdirty(pte), vma);
> > +
> > + set_pte_at(vma->vm_mm, address, ptep, pte);
> > +
> > + /*
> > + * No need to take a page reference as one was already
> > + * created when the swap entry was made.
> > + */
> > + if (PageAnon(page))
> > + page_add_anon_rmap(page, vma, address, false);
> > + else
> > + /*
> > + * Currently device exclusive access only supports anonymous
> > + * memory so the entry shouldn't point to a filebacked page.
> > + */
> > + WARN_ON_ONCE(!PageAnon(page));
> > +
> > + if (vma->vm_flags & VM_LOCKED)
> > + mlock_vma_page(page);
> > +
> > + /*
> > + * No need to invalidate - it was non-present before. However
> > + * secondary CPUs may have mappings that need invalidating.
> > + */
> > + update_mmu_cache(vma, address, ptep);
> > +}
> > +
> > +/*
> > + * Tries to restore an exclusive pte if the page lock can be acquired
> > without + * sleeping.
> > + */
> > +static unsigned long
>
> Better return a int?

Ok.

> > +try_restore_exclusive_pte(struct mm_struct *src_mm, pte_t *src_pte,
> > + struct vm_area_struct *vma, unsigned long addr)
>
> Raised in the other thread too: src_mm can be dropped.

Ack, sorry I must have missed that.

> > +{
> > + swp_entry_t entry = pte_to_swp_entry(*src_pte);
> > + struct page *page = pfn_swap_entry_to_page(entry);
> > +
> > + if (trylock_page(page)) {
> > + restore_exclusive_pte(vma, page, addr, src_pte);
> > + unlock_page(page);
> > + return 0;
> > + }
> > +
> > + return -EBUSY;
> > +}
> > +
> >
> > /*
> >
> > * copy one vm_area from one task to the other. Assumes the page tables
> > * already present in the new task to be cleared in the whole range
> >
> > @@ -781,6 +843,17 @@ copy_nonpresent_pte(struct mm_struct *dst_mm, struct
> > mm_struct *src_mm,>
> > pte = pte_swp_mkuffd_wp(pte);
> >
> > set_pte_at(src_mm, addr, src_pte, pte);
> >
> > }
> >
> > + } else if (is_device_exclusive_entry(entry)) {
> > + /*
> > + * Make device exclusive entries present by restoring the
> > + * original entry then copying as for a present pte. Device
> > + * exclusive entries currently only support private writable
> > + * (ie. COW) mappings.
> > + */
> > + VM_BUG_ON(!is_cow_mapping(vma->vm_flags));
> > + if (try_restore_exclusive_pte(src_mm, src_pte, vma, addr))
> > + return -EBUSY;
> > + return -ENOENT;
> >
> > }
> > set_pte_at(dst_mm, addr, dst_pte, pte);
> > return 0;
> >
> > @@ -980,9 +1053,18 @@ copy_pte_range(struct vm_area_struct *dst_vma,
> > struct vm_area_struct *src_vma,>
> > if (ret == -EAGAIN) {
> >
> > entry = pte_to_swp_entry(*src_pte);
> > break;
> >
> > + } else if (ret == -EBUSY) {
> > + break;
> > + } else if (!ret) {
> > + progress += 8;
> > + continue;
> >
> > }
> >
> > - progress += 8;
> > - continue;
> > +
> > + /*
> > + * Device exclusive entry restored, continue by
> > copying + * the now present pte.
> > + */
> > + WARN_ON_ONCE(ret != -ENOENT);
>
> The change looks right, thanks. It's just that we should start to consider
> document all these err code now in copy_pte_range() some day (perhaps on top
> of this patch)..

I tried to write the documentation but with the new clean-up patch using a
unique return code for each case the code ends up being rather self
documenting IMHO. It seems reasonably obvious what function returns what due
to the "if (ret == ...) break;" statements after each so the comments ended up
repeating the code (ie. copy_present_pte() returns this for this case, etc.),
but lets see what we think once I've updated.

Of course the whole thing is still a bit clunky, so it's still on my list of
things to look at reworking/cleaning up in future.

> > }
> > /* copy_present_pte() will clear `*prealloc' if consumed */
> > ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
> >
> > @@ -1019,6 +1101,8 @@ copy_pte_range(struct vm_area_struct *dst_vma,
> > struct vm_area_struct *src_vma,>
> > goto out;
> >
> > }
> > entry.val = 0;
> >
> > + } else if (ret == -EBUSY) {
> > + return -EBUSY;
> >
> > } else if (ret) {
> >
> > WARN_ON_ONCE(ret != -EAGAIN);
> > prealloc = page_copy_prealloc(src_mm, src_vma, addr);
> >
> > @@ -1283,7 +1367,8 @@ static unsigned long zap_pte_range(struct mmu_gather
> > *tlb,>
> > }
> >
> > entry = pte_to_swp_entry(ptent);
> >
> > - if (is_device_private_entry(entry)) {
> > + if (is_device_private_entry(entry) ||
> > + is_device_exclusive_entry(entry)) {
> >
> > struct page *page = pfn_swap_entry_to_page(entry);
> >
> > if (unlikely(details && details->check_mapping)) {
> >
> > @@ -1299,7 +1384,10 @@ static unsigned long zap_pte_range(struct
> > mmu_gather *tlb,>
> > pte_clear_not_present_full(mm, addr, pte,
> > tlb->fullmm);
> > rss[mm_counter(page)]--;
> >
> > - page_remove_rmap(page, false);
> > +
> > + if (is_device_private_entry(entry))
> > + page_remove_rmap(page, false);
> > +
> >
> > put_page(page);
> > continue;
> >
> > }
> >
> > @@ -3303,6 +3391,35 @@ void unmap_mapping_range(struct address_space
> > *mapping,>
> > }
> > EXPORT_SYMBOL(unmap_mapping_range);
> >
> > +/*
> > + * Restore a potential device exclusive pte to a working pte entry
> > + */
> > +static vm_fault_t remove_device_exclusive_entry(struct vm_fault *vmf)
> > +{
> > + struct page *page = vmf->page;
> > + struct vm_area_struct *vma = vmf->vma;
> > + vm_fault_t ret = 0;
> > + struct mmu_notifier_range range;
> > +
> > + if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags))
> > + return VM_FAULT_RETRY;
> > + mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma,
> > vma->vm_mm,
> > + vmf->address & PAGE_MASK,
> > + (vmf->address & PAGE_MASK) + PAGE_SIZE);
>
> @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will no
> longer have exclusive access to the page.
>
> Shouldn't this be the place to use new MMU_NOTIFY_EXCLUSIVE?

No. We could introduce another type to notify the range is going away due to
fault but as mentioned in the other thread I didn't think that was necessary
as the only sensible thing a driver can do is invalidate the entry anyway.

MMU_NOTIFY_EXCLUSIVE is to signal the invalidation is occurring because the
range is being marked for exclusive access (hopefully the explanation earlier
makes sense).

> > + mmu_notifier_invalidate_range_start(&range);
> > +
> > + vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
> > + &vmf->ptl);
> > + if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
> > + restore_exclusive_pte(vma, page, vmf->address, vmf->pte);
> > +
> > + pte_unmap_unlock(vmf->pte, vmf->ptl);
> > + unlock_page(page);
> > +
> > + mmu_notifier_invalidate_range_end(&range);
> > + return ret;
>
> We can drop "ret" and return 0 here directly.

Agreed, was left over from cleaning this function up in the last version.

> > +}
> > +
> >
> > /*
> >
> > * We enter with non-exclusive mmap_lock (to exclude vma changes,
> > * but allow concurrent faults), and pte mapped but not yet locked.
> >
> > @@ -3330,6 +3447,9 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
> >
> > if (is_migration_entry(entry)) {
> >
> > migration_entry_wait(vma->vm_mm, vmf->pmd,
> >
> > vmf->address);
> >
> > + } else if (is_device_exclusive_entry(entry)) {
> > + vmf->page = pfn_swap_entry_to_page(entry);
> > + ret = remove_device_exclusive_entry(vmf);
> >
> > } else if (is_device_private_entry(entry)) {
> >
> > vmf->page = pfn_swap_entry_to_page(entry);
> > ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
> >
> > diff --git a/mm/mprotect.c b/mm/mprotect.c
> > index ee5961888e70..883e2cc85cad 100644
> > --- a/mm/mprotect.c
> > +++ b/mm/mprotect.c
> > @@ -165,6 +165,14 @@ static unsigned long change_pte_range(struct
> > vm_area_struct *vma, pmd_t *pmd,>
> > newpte = swp_entry_to_pte(entry);
> > if (pte_swp_uffd_wp(oldpte))
> >
> > newpte = pte_swp_mkuffd_wp(newpte);
> >
> > + } else if
> > (is_writable_device_exclusive_entry(entry)) { +
> > entry = make_readable_device_exclusive_entry( +
> > swp_offset(entry)); +
> > newpte = swp_entry_to_pte(entry);
> > + if (pte_swp_soft_dirty(oldpte))
> > + newpte =
> > pte_swp_mksoft_dirty(newpte); + if
> > (pte_swp_uffd_wp(oldpte))
> > + newpte = pte_swp_mkuffd_wp(newpte);
> >
> > } else {
> >
> > newpte = oldpte;
> >
> > }
> >
> > diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
> > index a6a7febb4d93..f535bcb4950c 100644
> > --- a/mm/page_vma_mapped.c
> > +++ b/mm/page_vma_mapped.c
> > @@ -41,7 +41,8 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw)
> >
> > /* Handle un-addressable ZONE_DEVICE memory
> > */
> > entry = pte_to_swp_entry(*pvmw->pte);
> >
> > - if (!is_device_private_entry(entry))
> > + if (!is_device_private_entry(entry) &&
> > + !is_device_exclusive_entry(entry))
> >
> > return false;
> >
> > } else if (!pte_present(*pvmw->pte))
> >
> > return false;
> >
> > @@ -93,7 +94,8 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
> >
> > return false;
> >
> > entry = pte_to_swp_entry(*pvmw->pte);
> >
> > - if (!is_migration_entry(entry))
> > + if (!is_migration_entry(entry) &&
> > + !is_device_exclusive_entry(entry))
> >
> > return false;
> >
> > pfn = swp_offset(entry);
> >
> > @@ -102,7 +104,8 @@ static bool check_pte(struct page_vma_mapped_walk
> > *pvmw)>
> > /* Handle un-addressable ZONE_DEVICE memory */
> > entry = pte_to_swp_entry(*pvmw->pte);
> >
> > - if (!is_device_private_entry(entry))
> > + if (!is_device_private_entry(entry) &&
> > + !is_device_exclusive_entry(entry))
> >
> > return false;
> >
> > pfn = swp_offset(entry);
> >
> > diff --git a/mm/rmap.c b/mm/rmap.c
> > index 8ed1853060cf..fe062f63ef4d 100644
> > --- a/mm/rmap.c
> > +++ b/mm/rmap.c
> > @@ -2008,6 +2008,192 @@ void page_mlock(struct page *page)
> >
> > rmap_walk(page, &rwc);
> >
> > }
> >
> > +struct make_exclusive_args {
> > + struct mm_struct *mm;
> > + unsigned long address;
> > + void *owner;
> > + bool valid;
> > +};
> > +
> > +static bool page_make_device_exclusive_one(struct page *page,
> > + struct vm_area_struct *vma, unsigned long address, void
> > *priv) +{
> > + struct mm_struct *mm = vma->vm_mm;
> > + struct page_vma_mapped_walk pvmw = {
> > + .page = page,
> > + .vma = vma,
> > + .address = address,
> > + };
> > + struct make_exclusive_args *args = priv;
> > + pte_t pteval;
> > + struct page *subpage;
> > + bool ret = true;
> > + struct mmu_notifier_range range;
> > + swp_entry_t entry;
> > + pte_t swp_pte;
> > +
> > + mmu_notifier_range_init_owner(&range, MMU_NOTIFY_EXCLUSIVE, 0, vma,
>
> Similar question here, EXCLUSIVE comment says it gets notified when the
> device does not have exclusive access.
>
> If you prefer to keep using EXCLUSIVE for both mark/restore, then we need to
> change the comment above MMU_NOTIFY_EXCLUSIVE?

Yeah, sorry for the confusion that comment was stating the somewhat obvious
(any invalidation notifier means a device no longer has exclusive access) but
not enough detail about why a driver might treat this specific invalidation
reason differently.

> > + vma->vm_mm, address, min(vma->vm_end,
> > + address + page_size(page)),
> > args->owner); + mmu_notifier_invalidate_range_start(&range);
> > +
> > + while (page_vma_mapped_walk(&pvmw)) {
> > + /* Unexpected PMD-mapped THP? */
> > + VM_BUG_ON_PAGE(!pvmw.pte, page);
> > +
> > + if (!pte_present(*pvmw.pte)) {
> > + ret = false;
> > + page_vma_mapped_walk_done(&pvmw);
> > + break;
> > + }
> > +
> > + subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
>
> I see that all pages passed in should be done after FOLL_SPLIT_PMD, so is
> this needed? Or say, should subpage==page always be true?

Not always, in the case of a thp there are small ptes which will get device
exclusive entries.

> > + address = pvmw.address;
> > +
> > + /* Nuke the page table entry. */
> > + flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
> > + pteval = ptep_clear_flush(vma, address, pvmw.pte);
> > +
> > + /* Move the dirty bit to the page. Now the pte is gone. */
> > + if (pte_dirty(pteval))
> > + set_page_dirty(page);
> > +
> > + if (arch_unmap_one(mm, vma, address, pteval) < 0) {
> > + set_pte_at(mm, address, pvmw.pte, pteval);
> > + ret = false;
> > + page_vma_mapped_walk_done(&pvmw);
> > + break;
> > + }
>
> Didn't notice this previously, but also suggest to drop this.
>
> Two reasons:
>
> 1. It's introduced in ca827d55ebaa ("mm, swap: Add infrastructure for saving
> page metadata on swap", 2018-03-18) for sparc-only use so far. If we
> really want this, we'll also want to call arch_do_swap_page() when
> restoring the pte just like what we do in do_swap_page(); NOTE: current
> code path of SWP_DEVICE_EXCLUSIVE will skip the arch_do_swap_page() in
> do_swap_page() so it's not even paired with the above arch_unmap_one(), so
> I believe this won't even work for sparc at all.
>
> 2. I highly doubt whether sparc is also on the list of platforms to support
> for device atomic ops even in the future. IMHO we'd better not copy-paste
> code clips if never used at all, because once merged, removing it would
> need more justifications.

That seems reasonable, I am not aware of any need to support this on sparc now
or in the future and we can always add it then. And as you say I had missed
the need to pair it with arch_do_swap_page() anyway.

> > +
> > + /*
> > + * Check that our target page is still mapped at the
> > expected
> > + * address.
> > + */
> > + if (args->mm == mm && args->address == address &&
> > + pte_write(pteval))
> > + args->valid = true;
> > +
> > + /*
> > + * Store the pfn of the page in a special migration
> > + * pte. do_swap_page() will wait until the migration
> > + * pte is removed and then restart fault handling.
> > + */
> > + if (pte_write(pteval))
> > + entry = make_writable_device_exclusive_entry(
> > +
> > page_to_pfn(subpage)); + else
> > + entry = make_readable_device_exclusive_entry(
> > +
> > page_to_pfn(subpage)); + swp_pte = swp_entry_to_pte(entry);
> > + if (pte_soft_dirty(pteval))
> > + swp_pte = pte_swp_mksoft_dirty(swp_pte);
> > + if (pte_uffd_wp(pteval))
> > + swp_pte = pte_swp_mkuffd_wp(swp_pte);
> > +
> > + /* Take a reference for the swap entry */
> > + get_page(page);
> > + set_pte_at(mm, address, pvmw.pte, swp_pte);
> > +
> > + page_remove_rmap(subpage, PageHuge(page));
>
> Why PageHuge()? Should it be a constant "false"?

Yes.

> > + put_page(page);
>
> Should we drop this put_page() along with get_page() above?
>
> page_count() should be >0 anyway as we've got a mapcount before at least
> when dropping the pte. Then IMHO we can simply keep the old page
> reference.

I had debated doing that when I wrote it but left it there to keep things
obvious whilst checking the refcounting. However a comment here works just as
well so have done that.

> > + }
> > +
> > + mmu_notifier_invalidate_range_end(&range);
> > +
> > + return ret;
> > +}
> > +
> > +/**
> > + * page_make_device_exclusive - replace page table mappings with swap
> > entries
> "with swap entries" looks a bit blurred to me (although below longer comment
> explains much better). How about below (or something similar):
>
> page_make_device_exclusive - Mark the page exclusively owned by the device
>
> ?

Seems good, will do.

> It'll also match with comment above make_device_exclusive_range().
>
> No strong opinion.
>
> The rest looks good. Thanks,

Thanks again for looking.

> > + * @page: the page to replace page table entries for
> > + * @mm: the mm_struct where the page is expected to be mapped
> > + * @address: address where the page is expected to be mapped
> > + * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier callbacks
> > + *
> > + * Tries to remove all the page table entries which are mapping this page
> > and + * replace them with special device exclusive swap entries to grant
> > a device + * exclusive access to the page. Caller must hold the page
> > lock.
> > + *
> > + * Returns false if the page is still mapped, or if it could not be
> > unmapped + * from the expected address. Otherwise returns true (success).
> > + */
> > +static bool page_make_device_exclusive(struct page *page, struct
> > mm_struct *mm, + unsigned long address, void
> > *owner)
> > +{
> > + struct make_exclusive_args args = {
> > + .mm = mm,
> > + .address = address,
> > + .owner = owner,
> > + .valid = false,
> > + };
> > + struct rmap_walk_control rwc = {
> > + .rmap_one = page_make_device_exclusive_one,
> > + .done = page_not_mapped,
> > + .anon_lock = page_lock_anon_vma_read,
> > + .arg = &args,
> > + };
> > +
> > + /*
> > + * Restrict to anonymous pages for now to avoid potential writeback
> > + * issues.
> > + */
> > + if (!PageAnon(page))
> > + return false;
> > +
> > + rmap_walk(page, &rwc);
> > +
> > + return args.valid && !page_mapcount(page);
> > +}
> > +
> > +/**
> > + * make_device_exclusive_range() - Mark a range for exclusive use by a
> > device + * @mm: mm_struct of assoicated target process
> > + * @start: start of the region to mark for exclusive device access
> > + * @end: end address of region
> > + * @pages: returns the pages which were successfully marked for exclusive
> > access + * @owner: passed to MMU_NOTIFY_EXCLUSIVE range notifier to allow
> > filtering + *
> > + * Returns: number of pages found in the range by GUP. A page is marked
> > for + * exclusive access only if the page pointer is non-NULL.
> > + *
> > + * This function finds ptes mapping page(s) to the given address range,
> > locks + * them and replaces mappings with special swap entries preventing
> > userspace CPU + * access. On fault these entries are replaced with the
> > original mapping after + * calling MMU notifiers.
> > + *
> > + * A driver using this to program access from a device must use a mmu
> > notifier + * critical section to hold a device specific lock during
> > programming. Once + * programming is complete it should drop the page
> > lock and reference after + * which point CPU access to the page will
> > revoke the exclusive access. + */
> > +int make_device_exclusive_range(struct mm_struct *mm, unsigned long
> > start,
> > + unsigned long end, struct page **pages,
> > + void *owner)
> > +{
> > + unsigned long npages = (end - start) >> PAGE_SHIFT;
> > + unsigned long i;
> > +
> > + npages = get_user_pages_remote(mm, start, npages,
> > + FOLL_GET | FOLL_WRITE |
> > FOLL_SPLIT_PMD, + pages, NULL, NULL);
> > + for (i = 0; i < npages; i++, start += PAGE_SIZE) {
> > + if (!trylock_page(pages[i])) {
> > + put_page(pages[i]);
> > + pages[i] = NULL;
> > + continue;
> > + }
> > +
> > + if (!page_make_device_exclusive(pages[i], mm, start, owner))
> > { + unlock_page(pages[i]);
> > + put_page(pages[i]);
> > + pages[i] = NULL;
> > + }
> > + }
> > +
> > + return npages;
> > +}
> > +EXPORT_SYMBOL_GPL(make_device_exclusive_range);
> > +
> >
> > void __put_anon_vma(struct anon_vma *anon_vma)
> > {
> >
> > struct anon_vma *root = anon_vma->root;
> >
> > --
> > 2.20.1
>
> --
> Peter Xu




2021-05-27 20:50:49

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Thu, May 27, 2021 at 01:35:39PM +1000, Alistair Popple wrote:
> > > + *
> > > + * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device will
> > > no + * longer have exclusive access to the page. May ignore the
> > > invalidation that's + * part of make_device_exclusive_range() if the
> > > owner field
> > > + * matches the value passed to make_device_exclusive_range().
> >
> > Perhaps s/matches/does not match/?
>
> No, "matches" is correct. The MMU_NOTIFY_EXCLUSIVE notifier is to notify a
> listener that a range is being invalidated for the purpose of making the range
> available for some device to have exclusive access to. Which does also mean a
> device getting the notification no longer has exclusive access if it already
> did.
>
> A unique type is needed because when creating the range a driver needs to form
> a mmu critical section (with mmu_interval_read_begin()/
> mmu_interval_read_end()) to ensure the entry remains valid long enough to
> program the device pte and hasn't been invalidated.
>
> However without a way of filtering any invalidations will result in a retry,
> but make_device_exclusive_range() needs to do an invalidation during
> installation of the entry. To avoid this causing infinite retries the driver
> ignores specific invalidation events that it knows don't apply, ie. the
> invalidations that are a result of that driver asking for device exclusive
> entries.

OK I think I get it now.. so the driver checks both EXCLUSIVE and owner, if all
match it skips the notify, otherwise it's treated like all the rest. Thanks.

However then it's still confusing (as I raised it too in previous comment) that
we use CLEAR when re-installing the valid pte. It's merely against what CLEAR
means.

How about sending EXCLUSIVE for both mark/restore? Just that when restore we
notify with owner==NULL telling that no one is owning it anymore so driver
needs to drop the ownership. I assume your driver patch does not need change
too. Would that be much cleaner than CLEAR? I bet it also makes commenting
the new notify easier.

What do you think?

[...]

> > > + vma->vm_mm, address, min(vma->vm_end,
> > > + address + page_size(page)),
> > > args->owner); + mmu_notifier_invalidate_range_start(&range);
> > > +
> > > + while (page_vma_mapped_walk(&pvmw)) {
> > > + /* Unexpected PMD-mapped THP? */
> > > + VM_BUG_ON_PAGE(!pvmw.pte, page);
> > > +
> > > + if (!pte_present(*pvmw.pte)) {
> > > + ret = false;
> > > + page_vma_mapped_walk_done(&pvmw);
> > > + break;
> > > + }
> > > +
> > > + subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
> >
> > I see that all pages passed in should be done after FOLL_SPLIT_PMD, so is
> > this needed? Or say, should subpage==page always be true?
>
> Not always, in the case of a thp there are small ptes which will get device
> exclusive entries.

FOLL_SPLIT_PMD will first split the huge thp into smaller pages, then do
follow_page_pte() on them (in follow_pmd_mask):

if (flags & FOLL_SPLIT_PMD) {
int ret;
page = pmd_page(*pmd);
if (is_huge_zero_page(page)) {
spin_unlock(ptl);
ret = 0;
split_huge_pmd(vma, pmd, address);
if (pmd_trans_unstable(pmd))
ret = -EBUSY;
} else {
spin_unlock(ptl);
split_huge_pmd(vma, pmd, address);
ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
}

return ret ? ERR_PTR(ret) :
follow_page_pte(vma, address, pmd, flags, &ctx->pgmap);
}

So I thought all pages are small pages?

--
Peter Xu

2021-05-28 05:33:41

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Thursday, 27 May 2021 11:04:57 PM AEST Peter Xu wrote:
> On Thu, May 27, 2021 at 01:35:39PM +1000, Alistair Popple wrote:
> > > > + *
> > > > + * @MMU_NOTIFY_EXCLUSIVE: to signal a device driver that the device
> > > > will
> > > > no + * longer have exclusive access to the page. May ignore the
> > > > invalidation that's + * part of make_device_exclusive_range() if the
> > > > owner field
> > > > + * matches the value passed to make_device_exclusive_range().
> > >
> > > Perhaps s/matches/does not match/?
> >
> > No, "matches" is correct. The MMU_NOTIFY_EXCLUSIVE notifier is to notify a
> > listener that a range is being invalidated for the purpose of making the
> > range available for some device to have exclusive access to. Which does
> > also mean a device getting the notification no longer has exclusive
> > access if it already did.
> >
> > A unique type is needed because when creating the range a driver needs to
> > form a mmu critical section (with mmu_interval_read_begin()/
> > mmu_interval_read_end()) to ensure the entry remains valid long enough to
> > program the device pte and hasn't been invalidated.
> >
> > However without a way of filtering any invalidations will result in a
> > retry, but make_device_exclusive_range() needs to do an invalidation
> > during installation of the entry. To avoid this causing infinite retries
> > the driver ignores specific invalidation events that it knows don't
> > apply, ie. the invalidations that are a result of that driver asking for
> > device exclusive entries.
>
> OK I think I get it now.. so the driver checks both EXCLUSIVE and owner, if
> all match it skips the notify, otherwise it's treated like all the rest.
> Thanks.
>
> However then it's still confusing (as I raised it too in previous comment)
> that we use CLEAR when re-installing the valid pte. It's merely against
> what CLEAR means.

Oh, thanks. I understand where you are coming from now - the pte is already
invalid so ordinarily wouldn't need clearing.

> How about sending EXCLUSIVE for both mark/restore? Just that when restore
> we notify with owner==NULL telling that no one is owning it anymore so
> driver needs to drop the ownership. I assume your driver patch does not
> need change too. Would that be much cleaner than CLEAR? I bet it also
> makes commenting the new notify easier.
>
> What do you think?

That seems like a good and avoids adding another type. And as you say they
driver patch shouldn't need changing either (will need to confirm though).

> [...]
>
> > > > + vma->vm_mm, address,
> > > > min(vma->vm_end,
> > > > + address + page_size(page)),
> > > > args->owner); + mmu_notifier_invalidate_range_start(&range);
> > > > +
> > > > + while (page_vma_mapped_walk(&pvmw)) {
> > > > + /* Unexpected PMD-mapped THP? */
> > > > + VM_BUG_ON_PAGE(!pvmw.pte, page);
> > > > +
> > > > + if (!pte_present(*pvmw.pte)) {
> > > > + ret = false;
> > > > + page_vma_mapped_walk_done(&pvmw);
> > > > + break;
> > > > + }
> > > > +
> > > > + subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
> > >
> > > I see that all pages passed in should be done after FOLL_SPLIT_PMD, so
> > > is
> > > this needed? Or say, should subpage==page always be true?
> >
> > Not always, in the case of a thp there are small ptes which will get
> > device
> > exclusive entries.
>
> FOLL_SPLIT_PMD will first split the huge thp into smaller pages, then do
> follow_page_pte() on them (in follow_pmd_mask):
>
> if (flags & FOLL_SPLIT_PMD) {
> int ret;
> page = pmd_page(*pmd);
> if (is_huge_zero_page(page)) {
> spin_unlock(ptl);
> ret = 0;
> split_huge_pmd(vma, pmd, address);
> if (pmd_trans_unstable(pmd))
> ret = -EBUSY;
> } else {
> spin_unlock(ptl);
> split_huge_pmd(vma, pmd, address);
> ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
> }
>
> return ret ? ERR_PTR(ret) :
> follow_page_pte(vma, address, pmd, flags,
> &ctx->pgmap); }
>
> So I thought all pages are small pages?

The page will remain as a transparent huge page though (at least as I
understand things). FOLL_SPLIT_PMD turns it into a pte mapped thp by splitting
the pmd and creating pte's mapping the subpages but doesn't split the page
itself. For comparison FOLL_SPLIT (which has been removed in v5.13 due to lack
of use) is what would be used to split the page in the above GUP code by
calling split_huge_page() rather than split_huge_pmd().

This was done to avoid adding code for handling device exclusive entries at
the pmd level as well which would have made the changes more complicated and
seems unnecessary at least for now.

> --
> Peter Xu




2021-05-28 13:13:10

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Fri, May 28, 2021 at 11:48:40AM +1000, Alistair Popple wrote:

[...]

> > > > > + while (page_vma_mapped_walk(&pvmw)) {
> > > > > + /* Unexpected PMD-mapped THP? */
> > > > > + VM_BUG_ON_PAGE(!pvmw.pte, page);
> > > > > +
> > > > > + if (!pte_present(*pvmw.pte)) {
> > > > > + ret = false;
> > > > > + page_vma_mapped_walk_done(&pvmw);
> > > > > + break;
> > > > > + }
> > > > > +
> > > > > + subpage = page - page_to_pfn(page) + pte_pfn(*pvmw.pte);
> > > >
> > > > I see that all pages passed in should be done after FOLL_SPLIT_PMD, so
> > > > is
> > > > this needed? Or say, should subpage==page always be true?
> > >
> > > Not always, in the case of a thp there are small ptes which will get
> > > device
> > > exclusive entries.
> >
> > FOLL_SPLIT_PMD will first split the huge thp into smaller pages, then do
> > follow_page_pte() on them (in follow_pmd_mask):
> >
> > if (flags & FOLL_SPLIT_PMD) {
> > int ret;
> > page = pmd_page(*pmd);
> > if (is_huge_zero_page(page)) {
> > spin_unlock(ptl);
> > ret = 0;
> > split_huge_pmd(vma, pmd, address);
> > if (pmd_trans_unstable(pmd))
> > ret = -EBUSY;
> > } else {
> > spin_unlock(ptl);
> > split_huge_pmd(vma, pmd, address);
> > ret = pte_alloc(mm, pmd) ? -ENOMEM : 0;
> > }
> >
> > return ret ? ERR_PTR(ret) :
> > follow_page_pte(vma, address, pmd, flags,
> > &ctx->pgmap); }
> >
> > So I thought all pages are small pages?
>
> The page will remain as a transparent huge page though (at least as I
> understand things). FOLL_SPLIT_PMD turns it into a pte mapped thp by splitting
> the pmd and creating pte's mapping the subpages but doesn't split the page
> itself. For comparison FOLL_SPLIT (which has been removed in v5.13 due to lack
> of use) is what would be used to split the page in the above GUP code by
> calling split_huge_page() rather than split_huge_pmd().

But shouldn't FOLL_SPLIT_PMD filled in small pfns for each pte? See
__split_huge_pmd_locked():

for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
...
} else {
entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
...
}
...
set_pte_at(mm, addr, pte, entry);
}

Then iiuc the coming follow_page_pte() will directly fetch the small pages?

--
Peter Xu

2021-06-02 09:50:07

by Balbir Singh

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Wed, May 26, 2021 at 12:17:18AM -0700, John Hubbard wrote:
> On 5/25/21 4:51 AM, Balbir Singh wrote:
> ...
> > > How beneficial is this code to nouveau users? I see that it permits a
> > > part of OpenCL to be implemented, but how useful/important is this in
> > > the real world?
> >
> > That is a very good question! I've not reviewed the code, but a sample
> > program with the described use case would make things easy to parse.
> > I suspect that is not easy to build at the moment?
> >
>
> The cover letter says this:
>
> This has been tested with upstream Mesa 21.1.0 and a simple OpenCL program
> which checks that GPU atomic accesses to system memory are atomic. Without
> this series the test fails as there is no way of write-protecting the page
> mapping which results in the device clobbering CPU writes. For reference
> the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/
>
> Further testing has been performed by adding support for testing exclusive
> access to the hmm-tests kselftests.
>
> ...so that seems to cover the "sample program" request, at least.

Thanks, I'll take a look

>
> > I wonder how we co-ordinate all the work the mm is doing, page migration,
> > reclaim with device exclusive access? Do we have any numbers for the worst
> > case page fault latency when something is marked away for exclusive access?
>
> CPU page fault latency is approximately "terrible", if a page is resident on
> the GPU. We have to spin up a DMA engine on the GPU and have it copy the page
> over the PCIe bus, after all.
>
> > I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE would
>
> Yes, for now.
>
> > only impact the address space of programs using the GPU. Should the exclusively
> > marked range live in the unreclaimable list and recycled back to active/in-active
> > to account for the fact that
> >
> > 1. It is not reclaimable and reclaim will only hurt via page faults?
> > 2. It ages the page correctly or at-least allows for that possibility when the
> > page is used by the GPU.
>
> I'm not sure that that is *necessarily* something we can conclude. It depends upon
> access patterns of each program. For example, a "reduction" parallel program sends
> over lots of data to the GPU, and only a tiny bit of (reduced!) data comes back
> to the CPU. In that case, freeing the physical page on the CPU is actually the
> best decision for the OS to make (if the OS is sufficiently prescient).
>

With a shared device or a device exclusive range, it would be good to get the device
usage pattern and update the mm with that knowledge, so that the LRU can be better
maintained. With your comment you seem to suggest that a page used by the GPU might
be a good candidate for reclaim based on the CPU's understanding of the age of
the page should not account for use by the device
(are GPU workloads - access once and discard?)

Balbir Singh.

2021-06-02 14:39:11

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Wed, Jun 02, 2021 at 06:50:37PM +1000, Balbir Singh wrote:
> On Wed, May 26, 2021 at 12:17:18AM -0700, John Hubbard wrote:
> > On 5/25/21 4:51 AM, Balbir Singh wrote:
> > ...
> > > > How beneficial is this code to nouveau users? I see that it permits a
> > > > part of OpenCL to be implemented, but how useful/important is this in
> > > > the real world?
> > >
> > > That is a very good question! I've not reviewed the code, but a sample
> > > program with the described use case would make things easy to parse.
> > > I suspect that is not easy to build at the moment?
> > >
> >
> > The cover letter says this:
> >
> > This has been tested with upstream Mesa 21.1.0 and a simple OpenCL program
> > which checks that GPU atomic accesses to system memory are atomic. Without
> > this series the test fails as there is no way of write-protecting the page
> > mapping which results in the device clobbering CPU writes. For reference
> > the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/
> >
> > Further testing has been performed by adding support for testing exclusive
> > access to the hmm-tests kselftests.
> >
> > ...so that seems to cover the "sample program" request, at least.
>
> Thanks, I'll take a look
>
> >
> > > I wonder how we co-ordinate all the work the mm is doing, page migration,
> > > reclaim with device exclusive access? Do we have any numbers for the worst
> > > case page fault latency when something is marked away for exclusive access?
> >
> > CPU page fault latency is approximately "terrible", if a page is resident on
> > the GPU. We have to spin up a DMA engine on the GPU and have it copy the page
> > over the PCIe bus, after all.
> >
> > > I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE would
> >
> > Yes, for now.
> >
> > > only impact the address space of programs using the GPU. Should the exclusively
> > > marked range live in the unreclaimable list and recycled back to active/in-active
> > > to account for the fact that
> > >
> > > 1. It is not reclaimable and reclaim will only hurt via page faults?
> > > 2. It ages the page correctly or at-least allows for that possibility when the
> > > page is used by the GPU.
> >
> > I'm not sure that that is *necessarily* something we can conclude. It depends upon
> > access patterns of each program. For example, a "reduction" parallel program sends
> > over lots of data to the GPU, and only a tiny bit of (reduced!) data comes back
> > to the CPU. In that case, freeing the physical page on the CPU is actually the
> > best decision for the OS to make (if the OS is sufficiently prescient).
> >
>
> With a shared device or a device exclusive range, it would be good to get the device
> usage pattern and update the mm with that knowledge, so that the LRU can be better
> maintained. With your comment you seem to suggest that a page used by the GPU might
> be a good candidate for reclaim based on the CPU's understanding of the age of
> the page should not account for use by the device
> (are GPU workloads - access once and discard?)

Hmm, besides the aging info, this reminded me: do we need to isolate the page
from lru too when marking device exclusive access?

Afaict the current patch didn't do that so I think it's reclaimable. If we
still have the rmap then we'll get a mmu notify CLEAR when unmapping that
special pte, so device driver should be able to drop the ownership. However we
dropped the rmap when marking exclusive. Now I don't know whether and how
it'll work if page reclaim runs with the page being exclusively owned if
without isolating the page..

--
Peter Xu

2021-06-03 08:40:58

by John Hubbard

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On 6/2/21 1:50 AM, Balbir Singh wrote:
...
>>> only impact the address space of programs using the GPU. Should the exclusively
>>> marked range live in the unreclaimable list and recycled back to active/in-active
>>> to account for the fact that
>>>
>>> 1. It is not reclaimable and reclaim will only hurt via page faults?
>>> 2. It ages the page correctly or at-least allows for that possibility when the
>>> page is used by the GPU.
>>
>> I'm not sure that that is *necessarily* something we can conclude. It depends upon
>> access patterns of each program. For example, a "reduction" parallel program sends
>> over lots of data to the GPU, and only a tiny bit of (reduced!) data comes back
>> to the CPU. In that case, freeing the physical page on the CPU is actually the
>> best decision for the OS to make (if the OS is sufficiently prescient).
>>
>
> With a shared device or a device exclusive range, it would be good to get the device
> usage pattern and update the mm with that knowledge, so that the LRU can be better

Integrating a GPU (or "device") processor and it's mm behavior with the Linux kernel is
always an interesting concept. Certainly worth exploring, although it's probably
not a small project by any means.

> maintained. With your comment you seem to suggest that a page used by the GPU might
> be a good candidate for reclaim based on the CPU's understanding of the age of
> the page should not account for use by the device
> (are GPU workloads - access once and discard?)
>

Well, that's a little too narrow of an interpretation. The GPU is a fairly general
purpose processor, and so it has all kinds of workloads. I'm trying to discourage
any hopes that one can know, in advance, precisely how the GPU's pages need to be
managed. It's similar to the the CPU, in that regard. My example was just one, out
of a vast pool of possible behaviors.

thanks,
--
John Hubbard
NVIDIA

2021-06-03 11:41:03

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Thursday, 3 June 2021 12:37:30 AM AEST Peter Xu wrote:
> External email: Use caution opening links or attachments
>
> On Wed, Jun 02, 2021 at 06:50:37PM +1000, Balbir Singh wrote:
> > On Wed, May 26, 2021 at 12:17:18AM -0700, John Hubbard wrote:
> > > On 5/25/21 4:51 AM, Balbir Singh wrote:
> > > ...
> > >
> > > > > How beneficial is this code to nouveau users? I see that it permits
> > > > > a
> > > > > part of OpenCL to be implemented, but how useful/important is this
> > > > > in
> > > > > the real world?
> > > >
> > > > That is a very good question! I've not reviewed the code, but a sample
> > > > program with the described use case would make things easy to parse.
> > > > I suspect that is not easy to build at the moment?
> > >
> > > The cover letter says this:
> > >
> > > This has been tested with upstream Mesa 21.1.0 and a simple OpenCL
> > > program
> > > which checks that GPU atomic accesses to system memory are atomic.
> > > Without
> > > this series the test fails as there is no way of write-protecting the
> > > page
> > > mapping which results in the device clobbering CPU writes. For reference
> > > the test is available at https://ozlabs.org/~apopple/opencl_svm_atomics/
> > >
> > > Further testing has been performed by adding support for testing
> > > exclusive
> > > access to the hmm-tests kselftests.
> > >
> > > ...so that seems to cover the "sample program" request, at least.
> >
> > Thanks, I'll take a look
> >
> > > > I wonder how we co-ordinate all the work the mm is doing, page
> > > > migration,
> > > > reclaim with device exclusive access? Do we have any numbers for the
> > > > worst
> > > > case page fault latency when something is marked away for exclusive
> > > > access?
> > >
> > > CPU page fault latency is approximately "terrible", if a page is
> > > resident on the GPU. We have to spin up a DMA engine on the GPU and
> > > have it copy the page over the PCIe bus, after all.
> > >
> > > > I presume for now this is anonymous memory only? SWP_DEVICE_EXCLUSIVE
> > > > would
> > >
> > > Yes, for now.
> > >
> > > > only impact the address space of programs using the GPU. Should the
> > > > exclusively marked range live in the unreclaimable list and recycled
> > > > back to active/in-active to account for the fact that
> > > >
> > > > 1. It is not reclaimable and reclaim will only hurt via page faults?
> > > > 2. It ages the page correctly or at-least allows for that possibility
> > > > when the> > >
> > > > page is used by the GPU.
> > >
> > > I'm not sure that that is *necessarily* something we can conclude. It
> > > depends upon access patterns of each program. For example, a
> > > "reduction" parallel program sends over lots of data to the GPU, and
> > > only a tiny bit of (reduced!) data comes back to the CPU. In that case,
> > > freeing the physical page on the CPU is actually the best decision for
> > > the OS to make (if the OS is sufficiently prescient).>
> > With a shared device or a device exclusive range, it would be good to get
> > the device usage pattern and update the mm with that knowledge, so that
> > the LRU can be better maintained. With your comment you seem to suggest
> > that a page used by the GPU might be a good candidate for reclaim based
> > on the CPU's understanding of the age of the page should not account for
> > use by the device
> > (are GPU workloads - access once and discard?)
>
> Hmm, besides the aging info, this reminded me: do we need to isolate the
> page from lru too when marking device exclusive access?
>
> Afaict the current patch didn't do that so I think it's reclaimable. If we
> still have the rmap then we'll get a mmu notify CLEAR when unmapping that
> special pte, so device driver should be able to drop the ownership. However
> we dropped the rmap when marking exclusive. Now I don't know whether and
> how it'll work if page reclaim runs with the page being exclusively owned
> if without isolating the page..

Reclaim won't run on the page due to the extra references from the special
swap entries.

> --
> Peter Xu




2021-06-03 14:51:39

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Thu, Jun 03, 2021 at 09:39:32PM +1000, Alistair Popple wrote:
> Reclaim won't run on the page due to the extra references from the special
> swap entries.

That sounds reasonable, but I didn't find the point that stops it, probably due
to my limited knowledge on the reclaim code. Could you elaborate?

--
Peter Xu

2021-06-04 01:09:34

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Friday, 4 June 2021 12:47:40 AM AEST Peter Xu wrote:
> External email: Use caution opening links or attachments
>
> On Thu, Jun 03, 2021 at 09:39:32PM +1000, Alistair Popple wrote:
> > Reclaim won't run on the page due to the extra references from the special
> > swap entries.
>
> That sounds reasonable, but I didn't find the point that stops it, probably
> due to my limited knowledge on the reclaim code. Could you elaborate?

Sure, it isn't immediately obvious but it ends up being detected at the start
of is_page_cache_freeable() in the pageout code:


static pageout_t pageout(struct page *page, struct address_space *mapping)
{

[...]

if (!is_page_cache_freeable(page))
return PAGE_KEEP;

- Alistair

> --
> Peter Xu




2021-06-04 15:24:09

by Peter Xu

[permalink] [raw]
Subject: Re: [PATCH v9 07/10] mm: Device exclusive memory access

On Fri, Jun 04, 2021 at 11:07:42AM +1000, Alistair Popple wrote:
> On Friday, 4 June 2021 12:47:40 AM AEST Peter Xu wrote:
> > External email: Use caution opening links or attachments
> >
> > On Thu, Jun 03, 2021 at 09:39:32PM +1000, Alistair Popple wrote:
> > > Reclaim won't run on the page due to the extra references from the special
> > > swap entries.
> >
> > That sounds reasonable, but I didn't find the point that stops it, probably
> > due to my limited knowledge on the reclaim code. Could you elaborate?
>
> Sure, it isn't immediately obvious but it ends up being detected at the start
> of is_page_cache_freeable() in the pageout code:
>
>
> static pageout_t pageout(struct page *page, struct address_space *mapping)
> {
>
> [...]
>
> if (!is_page_cache_freeable(page))
> return PAGE_KEEP;

I did look at pageout() but still missed this small helper indeed (while it's
so important to know..), thanks!

--
Peter Xu

2021-06-04 20:52:13

by Liam R. Howlett

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

* Shakeel Butt <[email protected]> [210525 19:45]:
> On Tue, May 25, 2021 at 11:40 AM Liam Howlett <[email protected]> wrote:
> >
> [...]
> > >
> > > +/*
> > > + * Walks the vma's mapping a page and mlocks the page if any locked vma's are
> > > + * found. Once one is found the page is locked and the scan can be terminated.
> > > + */
> >
> > Can you please add that this requires the mmap_sem() lock to the
> > comments?
> >
>
> Why does this require mmap_sem() lock? Also mmap_sem() lock of which mm_struct?


Doesn't the mlock_vma_page() require the mmap_sem() for reading? The
mm_struct in vma->vm_mm;


From what I can see, at least the following paths have mmap_lock held
for writing:

munlock_vma_pages_range() from __do_munmap()
munlokc_vma_pages_range() from remap_file_pages()

>
> > > +static bool page_mlock_one(struct page *page, struct vm_area_struct *vma,
> > > + unsigned long address, void *unused)
> > > +{
> > > + struct page_vma_mapped_walk pvmw = {
> > > + .page = page,
> > > + .vma = vma,
> > > + .address = address,
> > > + };
> > > +
> > > + /* An un-locked vma doesn't have any pages to lock, continue the scan */
> > > + if (!(vma->vm_flags & VM_LOCKED))
> > > + return true;
> > > +
> > > + while (page_vma_mapped_walk(&pvmw)) {
> > > + /* PTE-mapped THP are never mlocked */
> > > + if (!PageTransCompound(page))
> > > + mlock_vma_page(page);
> > > + page_vma_mapped_walk_done(&pvmw);
> > > +
> > > + /*
> > > + * no need to continue scanning other vma's if the page has
> > > + * been locked.
> > > + */
> > > + return false;
> > > + }
> > > +
> > > + return true;
> > > +}

munlock_vma_pages_range() comments still references try_to_{munlock|unmap}

2021-06-05 00:43:57

by Shakeel Butt

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

On Fri, Jun 4, 2021 at 1:49 PM Liam Howlett <[email protected]> wrote:
>
> * Shakeel Butt <[email protected]> [210525 19:45]:
> > On Tue, May 25, 2021 at 11:40 AM Liam Howlett <[email protected]> wrote:
> > >
> > [...]
> > > >
> > > > +/*
> > > > + * Walks the vma's mapping a page and mlocks the page if any locked vma's are
> > > > + * found. Once one is found the page is locked and the scan can be terminated.
> > > > + */
> > >
> > > Can you please add that this requires the mmap_sem() lock to the
> > > comments?
> > >
> >
> > Why does this require mmap_sem() lock? Also mmap_sem() lock of which mm_struct?
>
>
> Doesn't the mlock_vma_page() require the mmap_sem() for reading? The
> mm_struct in vma->vm_mm;
>

We are traversing all the vmas where this page is mapped of possibly
different mm_structs. I don't think we want to take mmap_sem() of all
those mm_structs. The commit b87537d9e2fe ("mm: rmap use pte lock not
mmap_sem to set PageMlocked") removed exactly that.

>
> From what I can see, at least the following paths have mmap_lock held
> for writing:
>
> munlock_vma_pages_range() from __do_munmap()
> munlokc_vma_pages_range() from remap_file_pages()
>

The following path does not hold mmap_sem:

exit_mmap() -> munlock_vma_pages_all() -> munlock_vma_pages_range().

I would really suggest all to carefully read the commit message of
b87537d9e2fe ("mm: rmap use pte lock not mmap_sem to set
PageMlocked").

Particularly the following paragraph:
...
Vlastimil Babka points out another race which this patch protects against.
try_to_unmap_one() might reach its mlock_vma_page() TestSetPageMlocked a
moment after munlock_vma_pages_all() did its Phase 1 TestClearPageMlocked:
leaving PageMlocked and unevictable when it should be evictable. mmap_sem
is ineffective because exit_mmap() does not hold it; page lock ineffective
because __munlock_pagevec() only takes it afterwards, in Phase 2; pte lock
is effective because __munlock_pagevec_fill() takes it to get the page,
after VM_LOCKED was cleared from vm_flags, so visible to try_to_unmap_one.
...

Alistair, please bring back the VM_LOCKED check with pte lock held and
the comment "Holding pte lock, we do *not* need mmap_lock here".

One positive outcome of this cleanup patch is the removal of
unnecessary invalidation (unmapping for kvm case) of secondary mmus.

2021-06-05 03:43:13

by Liam R. Howlett

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

* Shakeel Butt <[email protected]> [210604 20:41]:
> On Fri, Jun 4, 2021 at 1:49 PM Liam Howlett <[email protected]> wrote:
> >
> > * Shakeel Butt <[email protected]> [210525 19:45]:
> > > On Tue, May 25, 2021 at 11:40 AM Liam Howlett <[email protected]> wrote:
> > > >
> > > [...]
> > > > >
> > > > > +/*
> > > > > + * Walks the vma's mapping a page and mlocks the page if any locked vma's are
> > > > > + * found. Once one is found the page is locked and the scan can be terminated.
> > > > > + */
> > > >
> > > > Can you please add that this requires the mmap_sem() lock to the
> > > > comments?
> > > >
> > >
> > > Why does this require mmap_sem() lock? Also mmap_sem() lock of which mm_struct?
> >
> >
> > Doesn't the mlock_vma_page() require the mmap_sem() for reading? The
> > mm_struct in vma->vm_mm;
> >
>
> We are traversing all the vmas where this page is mapped of possibly
> different mm_structs. I don't think we want to take mmap_sem() of all
> those mm_structs. The commit b87537d9e2fe ("mm: rmap use pte lock not
> mmap_sem to set PageMlocked") removed exactly that.
>
> >
> > From what I can see, at least the following paths have mmap_lock held
> > for writing:
> >
> > munlock_vma_pages_range() from __do_munmap()
> > munlokc_vma_pages_range() from remap_file_pages()
> >
>
> The following path does not hold mmap_sem:
>
> exit_mmap() -> munlock_vma_pages_all() -> munlock_vma_pages_range().

Isn't this the benign race referenced by Hugh in the commit you point to
below?

>
> I would really suggest all to carefully read the commit message of
> b87537d9e2fe ("mm: rmap use pte lock not mmap_sem to set
> PageMlocked").
>
> Particularly the following paragraph:
> ...
> Vlastimil Babka points out another race which this patch protects against.
> try_to_unmap_one() might reach its mlock_vma_page() TestSetPageMlocked a
> moment after munlock_vma_pages_all() did its Phase 1 TestClearPageMlocked:
> leaving PageMlocked and unevictable when it should be evictable. mmap_sem
> is ineffective because exit_mmap() does not hold it; page lock ineffective
> because __munlock_pagevec() only takes it afterwards, in Phase 2; pte lock
> is effective because __munlock_pagevec_fill() takes it to get the page,
> after VM_LOCKED was cleared from vm_flags, so visible to try_to_unmap_one.
> ...

So this is saying the race with exit_mmap() isn't benign after all?

>
> Alistair, please bring back the VM_LOCKED check with pte lock held and
> the comment "Holding pte lock, we do *not* need mmap_lock here".
>
> One positive outcome of this cleanup patch is the removal of
> unnecessary invalidation (unmapping for kvm case) of secondary mmus.

2021-06-05 04:29:36

by Shakeel Butt

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

On Fri, Jun 4, 2021 at 8:39 PM Liam Howlett <[email protected]> wrote:
>
> > Particularly the following paragraph:
> > ...
> > Vlastimil Babka points out another race which this patch protects against.
> > try_to_unmap_one() might reach its mlock_vma_page() TestSetPageMlocked a
> > moment after munlock_vma_pages_all() did its Phase 1 TestClearPageMlocked:
> > leaving PageMlocked and unevictable when it should be evictable. mmap_sem
> > is ineffective because exit_mmap() does not hold it; page lock ineffective
> > because __munlock_pagevec() only takes it afterwards, in Phase 2; pte lock
> > is effective because __munlock_pagevec_fill() takes it to get the page,
> > after VM_LOCKED was cleared from vm_flags, so visible to try_to_unmap_one.
> > ...
>
> So this is saying the race with exit_mmap() isn't benign after all?
>

Yes, not benign at all.

2021-06-07 04:55:11

by Alistair Popple

[permalink] [raw]
Subject: Re: [PATCH v9 03/10] mm/rmap: Split try_to_munlock from try_to_unmap

On Saturday, 5 June 2021 10:41:03 AM AEST Shakeel Butt wrote:
> External email: Use caution opening links or attachments
>
>
> On Fri, Jun 4, 2021 at 1:49 PM Liam Howlett <[email protected]> wrote:
> >
> > * Shakeel Butt <[email protected]> [210525 19:45]:
> > > On Tue, May 25, 2021 at 11:40 AM Liam Howlett <[email protected]>
wrote:
> > > >
> > > [...]
> > > > >
> > > > > +/*
> > > > > + * Walks the vma's mapping a page and mlocks the page if any locked
vma's are
> > > > > + * found. Once one is found the page is locked and the scan can be
terminated.
> > > > > + */
> > > >
> > > > Can you please add that this requires the mmap_sem() lock to the
> > > > comments?
> > > >
> > >
> > > Why does this require mmap_sem() lock? Also mmap_sem() lock of which
mm_struct?
> >
> >
> > Doesn't the mlock_vma_page() require the mmap_sem() for reading? The
> > mm_struct in vma->vm_mm;
> >
>
> We are traversing all the vmas where this page is mapped of possibly
> different mm_structs. I don't think we want to take mmap_sem() of all
> those mm_structs. The commit b87537d9e2fe ("mm: rmap use pte lock not
> mmap_sem to set PageMlocked") removed exactly that.
>
> >
> > From what I can see, at least the following paths have mmap_lock held
> > for writing:
> >
> > munlock_vma_pages_range() from __do_munmap()
> > munlokc_vma_pages_range() from remap_file_pages()
> >
>
> The following path does not hold mmap_sem:
>
> exit_mmap() -> munlock_vma_pages_all() -> munlock_vma_pages_range().
>
> I would really suggest all to carefully read the commit message of
> b87537d9e2fe ("mm: rmap use pte lock not mmap_sem to set
> PageMlocked").
>
> Particularly the following paragraph:
> ...
> Vlastimil Babka points out another race which this patch protects
against.
> try_to_unmap_one() might reach its mlock_vma_page() TestSetPageMlocked
a
> moment after munlock_vma_pages_all() did its Phase 1
TestClearPageMlocked:
> leaving PageMlocked and unevictable when it should be evictable.
mmap_sem
> is ineffective because exit_mmap() does not hold it; page lock
ineffective
> because __munlock_pagevec() only takes it afterwards, in Phase 2; pte
lock
> is effective because __munlock_pagevec_fill() takes it to get the page,
> after VM_LOCKED was cleared from vm_flags, so visible to
try_to_unmap_one.
> ...
>
> Alistair, please bring back the VM_LOCKED check with pte lock held and
> the comment "Holding pte lock, we do *not* need mmap_lock here".

Actually thanks for highlighting that paragraph. I have gone back through the
code again in munlock_vma_pages_range() and think I have a better
understanding of it now. So now I agree - the check of VM_LOCKED under the PTL
is important to ensure mlock_vma_page() does not run after VM_LOCKED has been
cleared and __munlock_pagevec_fill() has run.

Will post v10 to fix this and the try_to_munlock reference pointed out by Liam
which I missed for v9. Thanks Shakeel for taking the time to point this out.

> One positive outcome of this cleanup patch is the removal of
> unnecessary invalidation (unmapping for kvm case) of secondary mmus.