From: Thomas Hellström <[email protected]>
Graphics APIs like OpenGL 4.4 and Vulkan require the graphics driver
to provide coherent graphics memory, meaning that the GPU sees any
content written to the coherent memory on the next GPU operation that
touches that memory, and the CPU sees any content written by the GPU
to that memory immediately after any fence object trailing the GPU
operation is signaled.
Paravirtual drivers that otherwise require explicit synchronization
needs to do this by hooking up dirty tracking to pagefault handlers
and buffer object validation.
Provide mm helpers needed for this and that also allow for huge pmd-
and pud entries (patch 1-3), and the associated vmwgfx code (patch 4-7).
The code has been tested and exercised by a tailored version of mesa
where we disable all explicit synchronization and assume graphics memory
is coherent. The performance loss varies of course; a typical number is
around 5%.
I would like to merge this code through the DRM tree, so an ack to include
the new mm helpers in that merge would be greatly appreciated.
Changes since RFC:
- Merge conflict changes moved to the correct patch. Fixes intra-patchset
compile errors.
- Be more aggressive when turning ttm vm code into helpers. This makes sure
we can use a const qualifier on the vmwgfx vm_ops.
- Reinstate a lost comment an fix an error path that was broken when turning
the ttm vm code into helpers.
- Remove explicit type-casts of struct vm_area_struct::vm_private_data
- Clarify the locking inversion that makes us not being able to use the mm
pagewalk code.
Changes since v1:
- Removed the vmwgfx maintainer entry for as_dirty_helpers.c, updated
commit message accordingly
- Removed the TTM patches from the series as they are merged separately
through DRM.
Changes since v2:
- Split out the pagewalk code from as_dirty_helpers.c and document locking.
- Add pre_vma and post_vma callbacks to the pagewalk code.
- Remove huge pmd and -pud asserts that would trip when we protect vmas with
struct address_space::i_mmap_rwsem rather than with
struct vm_area_struct::mmap_sem.
- Do some naming cleanup in as_dirty_helpers.c
Changes since v3:
- Extensive renaming of the dirty helpers including the filename.
- Update walk_page_mapping() doc.
- Update the pagewalk code to not unconditionally split pmds if a pte_entry()
callback is present. Update the dirty helper pmd_entry accordingly.
- Use separate walk ops for the dirty helpers.
- Update the pagewalk code to take the pagetable lock in walk_pte_range.
Changes since v4:
- Fix pte pointer confusion in patch 2/8
- Skip the pagewalk code conditional split patch for now, and update the
mapping_dirty_helper accordingly. That problem will be solved in a cleaner
way in a follow-up patchset.
Changes since v5:
- Fix tlb flushing when we have other pending tlb flushes.
Cc: Andrew Morton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Huang Ying <[email protected]>
Cc: Jérôme Glisse <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
From: Thomas Hellstrom <[email protected]>
The caller needs to make sure that the vma is not torn down during the
lock operation and can also use the i_mmap_rwsem for file-backed vmas.
Remove the BUG_ON. We could, as an alternative, add a test that either
vma->vm_mm->mmap_sem or vma->vm_file->f_mapping->i_mmap_rwsem are held.
Cc: Andrew Morton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Huang Ying <[email protected]>
Cc: Jérôme Glisse <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Signed-off-by: Thomas Hellstrom <[email protected]>
Acked-by: Kirill A. Shutemov <[email protected]>
---
include/linux/huge_mm.h | 2 --
1 file changed, 2 deletions(-)
diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
index 93d5cf0bc716..0b84e13e88e2 100644
--- a/include/linux/huge_mm.h
+++ b/include/linux/huge_mm.h
@@ -216,7 +216,6 @@ static inline int is_swap_pmd(pmd_t pmd)
static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
struct vm_area_struct *vma)
{
- VM_BUG_ON_VMA(!rwsem_is_locked(&vma->vm_mm->mmap_sem), vma);
if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
return __pmd_trans_huge_lock(pmd, vma);
else
@@ -225,7 +224,6 @@ static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
struct vm_area_struct *vma)
{
- VM_BUG_ON_VMA(!rwsem_is_locked(&vma->vm_mm->mmap_sem), vma);
if (pud_trans_huge(*pud) || pud_devmap(*pud))
return __pud_trans_huge_lock(pud, vma);
else
--
2.20.1
From: Thomas Hellstrom <[email protected]>
Add two utilities to 1) write-protect and 2) clean all ptes pointing into
a range of an address space.
The utilities are intended to aid in tracking dirty pages (either
driver-allocated system memory or pci device memory).
The write-protect utility should be used in conjunction with
page_mkwrite() and pfn_mkwrite() to trigger write page-faults on page
accesses. Typically one would want to use this on sparse accesses into
large memory regions. The clean utility should be used to utilize
hardware dirtying functionality and avoid the overhead of page-faults,
typically on large accesses into small memory regions.
Cc: Andrew Morton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Huang Ying <[email protected]>
Cc: Jérôme Glisse <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Signed-off-by: Thomas Hellstrom <[email protected]>
---
include/linux/mm.h | 13 +-
mm/Kconfig | 3 +
mm/Makefile | 1 +
mm/mapping_dirty_helpers.c | 315 +++++++++++++++++++++++++++++++++++++
4 files changed, 331 insertions(+), 1 deletion(-)
create mode 100644 mm/mapping_dirty_helpers.c
diff --git a/include/linux/mm.h b/include/linux/mm.h
index cc292273e6ba..4bc93477375e 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2637,7 +2637,6 @@ typedef int (*pte_fn_t)(pte_t *pte, unsigned long addr, void *data);
extern int apply_to_page_range(struct mm_struct *mm, unsigned long address,
unsigned long size, pte_fn_t fn, void *data);
-
#ifdef CONFIG_PAGE_POISONING
extern bool page_poisoning_enabled(void);
extern void kernel_poison_pages(struct page *page, int numpages, int enable);
@@ -2878,5 +2877,17 @@ static inline int pages_identical(struct page *page1, struct page *page2)
return !memcmp_pages(page1, page2);
}
+#ifdef CONFIG_MAPPING_DIRTY_HELPERS
+unsigned long clean_record_shared_mapping_range(struct address_space *mapping,
+ pgoff_t first_index, pgoff_t nr,
+ pgoff_t bitmap_pgoff,
+ unsigned long *bitmap,
+ pgoff_t *start,
+ pgoff_t *end);
+
+unsigned long wp_shared_mapping_range(struct address_space *mapping,
+ pgoff_t first_index, pgoff_t nr);
+#endif
+
#endif /* __KERNEL__ */
#endif /* _LINUX_MM_H */
diff --git a/mm/Kconfig b/mm/Kconfig
index a5dae9a7eb51..550f7aceb679 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -736,4 +736,7 @@ config ARCH_HAS_PTE_SPECIAL
config ARCH_HAS_HUGEPD
bool
+config MAPPING_DIRTY_HELPERS
+ bool
+
endmenu
diff --git a/mm/Makefile b/mm/Makefile
index d996846697ef..1937cc251883 100644
--- a/mm/Makefile
+++ b/mm/Makefile
@@ -107,3 +107,4 @@ obj-$(CONFIG_PERCPU_STATS) += percpu-stats.o
obj-$(CONFIG_ZONE_DEVICE) += memremap.o
obj-$(CONFIG_HMM_MIRROR) += hmm.o
obj-$(CONFIG_MEMFD_CREATE) += memfd.o
+obj-$(CONFIG_MAPPING_DIRTY_HELPERS) += mapping_dirty_helpers.o
diff --git a/mm/mapping_dirty_helpers.c b/mm/mapping_dirty_helpers.c
new file mode 100644
index 000000000000..71070dda9643
--- /dev/null
+++ b/mm/mapping_dirty_helpers.c
@@ -0,0 +1,315 @@
+// SPDX-License-Identifier: GPL-2.0
+#include <linux/pagewalk.h>
+#include <linux/hugetlb.h>
+#include <linux/bitops.h>
+#include <linux/mmu_notifier.h>
+#include <asm/cacheflush.h>
+#include <asm/tlbflush.h>
+
+/**
+ * struct wp_walk - Private struct for pagetable walk callbacks
+ * @range: Range for mmu notifiers
+ * @tlbflush_start: Address of first modified pte
+ * @tlbflush_end: Address of last modified pte + 1
+ * @total: Total number of modified ptes
+ */
+struct wp_walk {
+ struct mmu_notifier_range range;
+ unsigned long tlbflush_start;
+ unsigned long tlbflush_end;
+ unsigned long total;
+};
+
+/**
+ * wp_pte - Write-protect a pte
+ * @pte: Pointer to the pte
+ * @addr: The virtual page address
+ * @walk: pagetable walk callback argument
+ *
+ * The function write-protects a pte and records the range in
+ * virtual address space of touched ptes for efficient range TLB flushes.
+ */
+static int wp_pte(pte_t *pte, unsigned long addr, unsigned long end,
+ struct mm_walk *walk)
+{
+ struct wp_walk *wpwalk = walk->private;
+ pte_t ptent = *pte;
+
+ if (pte_write(ptent)) {
+ pte_t old_pte = ptep_modify_prot_start(walk->vma, addr, pte);
+
+ ptent = pte_wrprotect(old_pte);
+ ptep_modify_prot_commit(walk->vma, addr, pte, old_pte, ptent);
+ wpwalk->total++;
+ wpwalk->tlbflush_start = min(wpwalk->tlbflush_start, addr);
+ wpwalk->tlbflush_end = max(wpwalk->tlbflush_end,
+ addr + PAGE_SIZE);
+ }
+
+ return 0;
+}
+
+/**
+ * struct clean_walk - Private struct for the clean_record_pte function.
+ * @base: struct wp_walk we derive from
+ * @bitmap_pgoff: Address_space Page offset of the first bit in @bitmap
+ * @bitmap: Bitmap with one bit for each page offset in the address_space range
+ * covered.
+ * @start: Address_space page offset of first modified pte relative
+ * to @bitmap_pgoff
+ * @end: Address_space page offset of last modified pte relative
+ * to @bitmap_pgoff
+ */
+struct clean_walk {
+ struct wp_walk base;
+ pgoff_t bitmap_pgoff;
+ unsigned long *bitmap;
+ pgoff_t start;
+ pgoff_t end;
+};
+
+#define to_clean_walk(_wpwalk) container_of(_wpwalk, struct clean_walk, base)
+
+/**
+ * clean_record_pte - Clean a pte and record its address space offset in a
+ * bitmap
+ * @pte: Pointer to the pte
+ * @addr: The virtual page address
+ * @walk: pagetable walk callback argument
+ *
+ * The function cleans a pte and records the range in
+ * virtual address space of touched ptes for efficient TLB flushes.
+ * It also records dirty ptes in a bitmap representing page offsets
+ * in the address_space, as well as the first and last of the bits
+ * touched.
+ */
+static int clean_record_pte(pte_t *pte, unsigned long addr,
+ unsigned long end, struct mm_walk *walk)
+{
+ struct wp_walk *wpwalk = walk->private;
+ struct clean_walk *cwalk = to_clean_walk(wpwalk);
+ pte_t ptent = *pte;
+
+ if (pte_dirty(ptent)) {
+ pgoff_t pgoff = ((addr - walk->vma->vm_start) >> PAGE_SHIFT) +
+ walk->vma->vm_pgoff - cwalk->bitmap_pgoff;
+ pte_t old_pte = ptep_modify_prot_start(walk->vma, addr, pte);
+
+ ptent = pte_mkclean(old_pte);
+ ptep_modify_prot_commit(walk->vma, addr, pte, old_pte, ptent);
+
+ wpwalk->total++;
+ wpwalk->tlbflush_start = min(wpwalk->tlbflush_start, addr);
+ wpwalk->tlbflush_end = max(wpwalk->tlbflush_end,
+ addr + PAGE_SIZE);
+
+ __set_bit(pgoff, cwalk->bitmap);
+ cwalk->start = min(cwalk->start, pgoff);
+ cwalk->end = max(cwalk->end, pgoff + 1);
+ }
+
+ return 0;
+}
+
+/* wp_clean_pmd_entry - The pagewalk pmd callback. */
+static int wp_clean_pmd_entry(pmd_t *pmd, unsigned long addr, unsigned long end,
+ struct mm_walk *walk)
+{
+ /* Dirty-tracking should be handled on the pte level */
+ pmd_t pmdval = pmd_read_atomic(pmd);
+
+ if (pmd_trans_huge(pmdval) || pmd_devmap(pmdval))
+ WARN_ON(pmd_write(pmdval) || pmd_dirty(pmdval));
+
+ return 0;
+}
+
+/* wp_clean_pud_entry - The pagewalk pud callback. */
+static int wp_clean_pud_entry(pud_t *pud, unsigned long addr, unsigned long end,
+ struct mm_walk *walk)
+{
+ /* Dirty-tracking should be handled on the pte level */
+ pud_t pudval = READ_ONCE(*pud);
+
+ if (pud_trans_huge(pudval) || pud_devmap(pudval))
+ WARN_ON(pud_write(pudval) || pud_dirty(pudval));
+
+ return 0;
+}
+
+/*
+ * wp_clean_pre_vma - The pagewalk pre_vma callback.
+ *
+ * The pre_vma callback performs the cache flush, stages the tlb flush
+ * and calls the necessary mmu notifiers.
+ */
+static int wp_clean_pre_vma(unsigned long start, unsigned long end,
+ struct mm_walk *walk)
+{
+ struct wp_walk *wpwalk = walk->private;
+
+ wpwalk->tlbflush_start = end;
+ wpwalk->tlbflush_end = start;
+
+ mmu_notifier_range_init(&wpwalk->range, MMU_NOTIFY_PROTECTION_PAGE, 0,
+ walk->vma, walk->mm, start, end);
+ mmu_notifier_invalidate_range_start(&wpwalk->range);
+ flush_cache_range(walk->vma, start, end);
+
+ /*
+ * We're not using tlb_gather_mmu() since typically
+ * only a small subrange of PTEs are affected, whereas
+ * tlb_gather_mmu() records the full range.
+ */
+ inc_tlb_flush_pending(walk->mm);
+
+ return 0;
+}
+
+/*
+ * wp_clean_post_vma - The pagewalk post_vma callback.
+ *
+ * The post_vma callback performs the tlb flush and calls necessary mmu
+ * notifiers.
+ */
+static void wp_clean_post_vma(struct mm_walk *walk)
+{
+ struct wp_walk *wpwalk = walk->private;
+
+ if (mm_tlb_flush_nested(walk->mm))
+ flush_tlb_range(walk->vma, wpwalk->range.start,
+ wpwalk->range.end);
+ else if (wpwalk->tlbflush_end > wpwalk->tlbflush_start)
+ flush_tlb_range(walk->vma, wpwalk->tlbflush_start,
+ wpwalk->tlbflush_end);
+
+ mmu_notifier_invalidate_range_end(&wpwalk->range);
+ dec_tlb_flush_pending(walk->mm);
+}
+
+/*
+ * wp_clean_test_walk - The pagewalk test_walk callback.
+ *
+ * Won't perform dirty-tracking on COW, read-only or HUGETLB vmas.
+ */
+static int wp_clean_test_walk(unsigned long start, unsigned long end,
+ struct mm_walk *walk)
+{
+ unsigned long vm_flags = READ_ONCE(walk->vma->vm_flags);
+
+ /* Skip non-applicable VMAs */
+ if ((vm_flags & (VM_SHARED | VM_MAYWRITE | VM_HUGETLB)) !=
+ (VM_SHARED | VM_MAYWRITE))
+ return 1;
+
+ return 0;
+}
+
+static const struct mm_walk_ops clean_walk_ops = {
+ .pte_entry = clean_record_pte,
+ .pmd_entry = wp_clean_pmd_entry,
+ .pud_entry = wp_clean_pud_entry,
+ .test_walk = wp_clean_test_walk,
+ .pre_vma = wp_clean_pre_vma,
+ .post_vma = wp_clean_post_vma
+};
+
+static const struct mm_walk_ops wp_walk_ops = {
+ .pte_entry = wp_pte,
+ .pmd_entry = wp_clean_pmd_entry,
+ .pud_entry = wp_clean_pud_entry,
+ .test_walk = wp_clean_test_walk,
+ .pre_vma = wp_clean_pre_vma,
+ .post_vma = wp_clean_post_vma
+};
+
+/**
+ * wp_shared_mapping_range - Write-protect all ptes in an address space range
+ * @mapping: The address_space we want to write protect
+ * @first_index: The first page offset in the range
+ * @nr: Number of incremental page offsets to cover
+ *
+ * Note: This function currently skips transhuge page-table entries, since
+ * it's intended for dirty-tracking on the PTE level. It will warn on
+ * encountering transhuge write-enabled entries, though, and can easily be
+ * extended to handle them as well.
+ *
+ * Return: The number of ptes actually write-protected. Note that
+ * already write-protected ptes are not counted.
+ */
+unsigned long wp_shared_mapping_range(struct address_space *mapping,
+ pgoff_t first_index, pgoff_t nr)
+{
+ struct wp_walk wpwalk = { .total = 0 };
+
+ i_mmap_lock_read(mapping);
+ WARN_ON(walk_page_mapping(mapping, first_index, nr, &wp_walk_ops,
+ &wpwalk));
+ i_mmap_unlock_read(mapping);
+
+ return wpwalk.total;
+}
+EXPORT_SYMBOL_GPL(wp_shared_mapping_range);
+
+/**
+ * clean_record_shared_mapping_range - Clean and record all ptes in an
+ * address space range
+ * @mapping: The address_space we want to clean
+ * @first_index: The first page offset in the range
+ * @nr: Number of incremental page offsets to cover
+ * @bitmap_pgoff: The page offset of the first bit in @bitmap
+ * @bitmap: Pointer to a bitmap of at least @nr bits. The bitmap needs to
+ * cover the whole range @first_index..@first_index + @nr.
+ * @start: Pointer to number of the first set bit in @bitmap.
+ * is modified as new bits are set by the function.
+ * @end: Pointer to the number of the last set bit in @bitmap.
+ * none set. The value is modified as new bits are set by the function.
+ *
+ * Note: When this function returns there is no guarantee that a CPU has
+ * not already dirtied new ptes. However it will not clean any ptes not
+ * reported in the bitmap. The guarantees are as follows:
+ * a) All ptes dirty when the function starts executing will end up recorded
+ * in the bitmap.
+ * b) All ptes dirtied after that will either remain dirty, be recorded in the
+ * bitmap or both.
+ *
+ * If a caller needs to make sure all dirty ptes are picked up and none
+ * additional are added, it first needs to write-protect the address-space
+ * range and make sure new writers are blocked in page_mkwrite() or
+ * pfn_mkwrite(). And then after a TLB flush following the write-protection
+ * pick up all dirty bits.
+ *
+ * Note: This function currently skips transhuge page-table entries, since
+ * it's intended for dirty-tracking on the PTE level. It will warn on
+ * encountering transhuge dirty entries, though, and can easily be extended
+ * to handle them as well.
+ *
+ * Return: The number of dirty ptes actually cleaned.
+ */
+unsigned long clean_record_shared_mapping_range(struct address_space *mapping,
+ pgoff_t first_index, pgoff_t nr,
+ pgoff_t bitmap_pgoff,
+ unsigned long *bitmap,
+ pgoff_t *start,
+ pgoff_t *end)
+{
+ bool none_set = (*start >= *end);
+ struct clean_walk cwalk = {
+ .base = { .total = 0 },
+ .bitmap_pgoff = bitmap_pgoff,
+ .bitmap = bitmap,
+ .start = none_set ? nr : *start,
+ .end = none_set ? 0 : *end,
+ };
+
+ i_mmap_lock_read(mapping);
+ WARN_ON(walk_page_mapping(mapping, first_index, nr, &clean_walk_ops,
+ &cwalk.base));
+ i_mmap_unlock_read(mapping);
+
+ *start = cwalk.start;
+ *end = cwalk.end;
+
+ return cwalk.base.total;
+}
+EXPORT_SYMBOL_GPL(clean_record_shared_mapping_range);
--
2.20.1
From: Thomas Hellstrom <[email protected]>
Similar to write-coherent resources, make sure that from the user-space
point of view, GPU rendered contents is automatically available for
reading by the CPU.
Cc: Andrew Morton <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Will Deacon <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Rik van Riel <[email protected]>
Cc: Minchan Kim <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Huang Ying <[email protected]>
Cc: Jérôme Glisse <[email protected]>
Cc: Kirill A. Shutemov <[email protected]>
Signed-off-by: Thomas Hellstrom <[email protected]>
Reviewed-by: Deepak Rawat <[email protected]>
---
drivers/gpu/drm/vmwgfx/vmwgfx_drv.h | 7 +-
drivers/gpu/drm/vmwgfx/vmwgfx_page_dirty.c | 77 ++++++++++++-
drivers/gpu/drm/vmwgfx/vmwgfx_resource.c | 103 +++++++++++++++++-
drivers/gpu/drm/vmwgfx/vmwgfx_resource_priv.h | 2 +
drivers/gpu/drm/vmwgfx/vmwgfx_validation.c | 3 +-
5 files changed, 181 insertions(+), 11 deletions(-)
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index 53f8522ae032..729a2e93acf1 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -680,7 +680,8 @@ extern void vmw_resource_unreference(struct vmw_resource **p_res);
extern struct vmw_resource *vmw_resource_reference(struct vmw_resource *res);
extern struct vmw_resource *
vmw_resource_reference_unless_doomed(struct vmw_resource *res);
-extern int vmw_resource_validate(struct vmw_resource *res, bool intr);
+extern int vmw_resource_validate(struct vmw_resource *res, bool intr,
+ bool dirtying);
extern int vmw_resource_reserve(struct vmw_resource *res, bool interruptible,
bool no_backup);
extern bool vmw_resource_needs_backup(const struct vmw_resource *res);
@@ -724,6 +725,8 @@ void vmw_resource_mob_attach(struct vmw_resource *res);
void vmw_resource_mob_detach(struct vmw_resource *res);
void vmw_resource_dirty_update(struct vmw_resource *res, pgoff_t start,
pgoff_t end);
+int vmw_resources_clean(struct vmw_buffer_object *vbo, pgoff_t start,
+ pgoff_t end, pgoff_t *num_prefault);
/**
* vmw_resource_mob_attached - Whether a resource currently has a mob attached
@@ -1417,6 +1420,8 @@ int vmw_bo_dirty_add(struct vmw_buffer_object *vbo);
void vmw_bo_dirty_transfer_to_res(struct vmw_resource *res);
void vmw_bo_dirty_clear_res(struct vmw_resource *res);
void vmw_bo_dirty_release(struct vmw_buffer_object *vbo);
+void vmw_bo_dirty_unmap(struct vmw_buffer_object *vbo,
+ pgoff_t start, pgoff_t end);
vm_fault_t vmw_bo_vm_fault(struct vm_fault *vmf);
vm_fault_t vmw_bo_vm_mkwrite(struct vm_fault *vmf);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_page_dirty.c b/drivers/gpu/drm/vmwgfx/vmwgfx_page_dirty.c
index 060c1e492f25..f07aa857587c 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_page_dirty.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_page_dirty.c
@@ -155,7 +155,6 @@ static void vmw_bo_dirty_scan_mkwrite(struct vmw_buffer_object *vbo)
}
}
-
/**
* vmw_bo_dirty_scan - Scan for dirty pages and add them to the dirty
* tracking structure
@@ -173,6 +172,53 @@ void vmw_bo_dirty_scan(struct vmw_buffer_object *vbo)
vmw_bo_dirty_scan_mkwrite(vbo);
}
+/**
+ * vmw_bo_dirty_pre_unmap - write-protect and pick up dirty pages before
+ * an unmap_mapping_range operation.
+ * @vbo: The buffer object,
+ * @start: First page of the range within the buffer object.
+ * @end: Last page of the range within the buffer object + 1.
+ *
+ * If we're using the _PAGETABLE scan method, we may leak dirty pages
+ * when calling unmap_mapping_range(). This function makes sure we pick
+ * up all dirty pages.
+ */
+static void vmw_bo_dirty_pre_unmap(struct vmw_buffer_object *vbo,
+ pgoff_t start, pgoff_t end)
+{
+ struct vmw_bo_dirty *dirty = vbo->dirty;
+ unsigned long offset = drm_vma_node_start(&vbo->base.base.vma_node);
+ struct address_space *mapping = vbo->base.bdev->dev_mapping;
+
+ if (dirty->method != VMW_BO_DIRTY_PAGETABLE || start >= end)
+ return;
+
+ wp_shared_mapping_range(mapping, start + offset, end - start);
+ clean_record_shared_mapping_range(mapping, start + offset,
+ end - start, offset,
+ &dirty->bitmap[0], &dirty->start,
+ &dirty->end);
+}
+
+/**
+ * vmw_bo_dirty_unmap - Clear all ptes pointing to a range within a bo
+ * @vbo: The buffer object,
+ * @start: First page of the range within the buffer object.
+ * @end: Last page of the range within the buffer object + 1.
+ *
+ * This is similar to ttm_bo_unmap_virtual_locked() except it takes a subrange.
+ */
+void vmw_bo_dirty_unmap(struct vmw_buffer_object *vbo,
+ pgoff_t start, pgoff_t end)
+{
+ unsigned long offset = drm_vma_node_start(&vbo->base.base.vma_node);
+ struct address_space *mapping = vbo->base.bdev->dev_mapping;
+
+ vmw_bo_dirty_pre_unmap(vbo, start, end);
+ unmap_shared_mapping_range(mapping, (offset + start) << PAGE_SHIFT,
+ (loff_t) (end - start) << PAGE_SHIFT);
+}
+
/**
* vmw_bo_dirty_add - Add a dirty-tracking user to a buffer object
* @vbo: The buffer object
@@ -401,21 +447,42 @@ vm_fault_t vmw_bo_vm_fault(struct vm_fault *vmf)
if (ret)
return ret;
+ num_prefault = (vma->vm_flags & VM_RAND_READ) ? 1 :
+ TTM_BO_VM_NUM_PREFAULT;
+
+ if (vbo->dirty) {
+ pgoff_t allowed_prefault;
+ unsigned long page_offset;
+
+ page_offset = vmf->pgoff -
+ drm_vma_node_start(&bo->base.vma_node);
+ if (page_offset >= bo->num_pages ||
+ vmw_resources_clean(vbo, page_offset,
+ page_offset + PAGE_SIZE,
+ &allowed_prefault)) {
+ ret = VM_FAULT_SIGBUS;
+ goto out_unlock;
+ }
+
+ num_prefault = min(num_prefault, allowed_prefault);
+ }
+
/*
- * This will cause mkwrite() to be called for each pte on
- * write-enable vmas.
+ * If we don't track dirty using the MKWRITE method, make sure
+ * sure the page protection is write-enabled so we don't get
+ * a lot of unnecessary write faults.
*/
if (vbo->dirty && vbo->dirty->method == VMW_BO_DIRTY_MKWRITE)
prot = vma->vm_page_prot;
else
prot = vm_get_page_prot(vma->vm_flags);
- num_prefault = (vma->vm_flags & VM_RAND_READ) ? 0 :
- TTM_BO_VM_NUM_PREFAULT;
ret = ttm_bo_vm_fault_reserved(vmf, prot, num_prefault);
if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
return ret;
+out_unlock:
dma_resv_unlock(bo->base.resv);
+
return ret;
}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index 328ad46076ff..c76faf33972e 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -393,7 +393,8 @@ static int vmw_resource_buf_alloc(struct vmw_resource *res,
* should be retried once resources have been freed up.
*/
static int vmw_resource_do_validate(struct vmw_resource *res,
- struct ttm_validate_buffer *val_buf)
+ struct ttm_validate_buffer *val_buf,
+ bool dirtying)
{
int ret = 0;
const struct vmw_res_func *func = res->func;
@@ -435,6 +436,15 @@ static int vmw_resource_do_validate(struct vmw_resource *res,
* the resource.
*/
if (res->dirty) {
+ if (dirtying && !res->res_dirty) {
+ pgoff_t start = res->backup_offset >> PAGE_SHIFT;
+ pgoff_t end = __KERNEL_DIV_ROUND_UP
+ (res->backup_offset + res->backup_size,
+ PAGE_SIZE);
+
+ vmw_bo_dirty_unmap(res->backup, start, end);
+ }
+
vmw_bo_dirty_transfer_to_res(res);
return func->dirty_sync(res);
}
@@ -679,6 +689,7 @@ static int vmw_resource_do_evict(struct ww_acquire_ctx *ticket,
* to the device.
* @res: The resource to make visible to the device.
* @intr: Perform waits interruptible if possible.
+ * @dirtying: Pending GPU operation will dirty the resource
*
* On succesful return, any backup DMA buffer pointed to by @res->backup will
* be reserved and validated.
@@ -688,7 +699,8 @@ static int vmw_resource_do_evict(struct ww_acquire_ctx *ticket,
* Return: Zero on success, -ERESTARTSYS if interrupted, negative error code
* on failure.
*/
-int vmw_resource_validate(struct vmw_resource *res, bool intr)
+int vmw_resource_validate(struct vmw_resource *res, bool intr,
+ bool dirtying)
{
int ret;
struct vmw_resource *evict_res;
@@ -705,7 +717,7 @@ int vmw_resource_validate(struct vmw_resource *res, bool intr)
if (res->backup)
val_buf.bo = &res->backup->base;
do {
- ret = vmw_resource_do_validate(res, &val_buf);
+ ret = vmw_resource_do_validate(res, &val_buf, dirtying);
if (likely(ret != -EBUSY))
break;
@@ -1005,7 +1017,7 @@ int vmw_resource_pin(struct vmw_resource *res, bool interruptible)
/* Do we really need to pin the MOB as well? */
vmw_bo_pin_reserved(vbo, true);
}
- ret = vmw_resource_validate(res, interruptible);
+ ret = vmw_resource_validate(res, interruptible, true);
if (vbo)
ttm_bo_unreserve(&vbo->base);
if (ret)
@@ -1080,3 +1092,86 @@ void vmw_resource_dirty_update(struct vmw_resource *res, pgoff_t start,
res->func->dirty_range_add(res, start << PAGE_SHIFT,
end << PAGE_SHIFT);
}
+
+/**
+ * vmw_resources_clean - Clean resources intersecting a mob range
+ * @vbo: The mob buffer object
+ * @start: The mob page offset starting the range
+ * @end: The mob page offset ending the range
+ * @num_prefault: Returns how many pages including the first have been
+ * cleaned and are ok to prefault
+ */
+int vmw_resources_clean(struct vmw_buffer_object *vbo, pgoff_t start,
+ pgoff_t end, pgoff_t *num_prefault)
+{
+ struct rb_node *cur = vbo->res_tree.rb_node;
+ struct vmw_resource *found = NULL;
+ unsigned long res_start = start << PAGE_SHIFT;
+ unsigned long res_end = end << PAGE_SHIFT;
+ unsigned long last_cleaned = 0;
+
+ /*
+ * Find the resource with lowest backup_offset that intersects the
+ * range.
+ */
+ while (cur) {
+ struct vmw_resource *cur_res =
+ container_of(cur, struct vmw_resource, mob_node);
+
+ if (cur_res->backup_offset >= res_end) {
+ cur = cur->rb_left;
+ } else if (cur_res->backup_offset + cur_res->backup_size <=
+ res_start) {
+ cur = cur->rb_right;
+ } else {
+ found = cur_res;
+ cur = cur->rb_left;
+ /* Continue to look for resources with lower offsets */
+ }
+ }
+
+ /*
+ * In order of increasing backup_offset, clean dirty resorces
+ * intersecting the range.
+ */
+ while (found) {
+ if (found->res_dirty) {
+ int ret;
+
+ if (!found->func->clean)
+ return -EINVAL;
+
+ ret = found->func->clean(found);
+ if (ret)
+ return ret;
+
+ found->res_dirty = false;
+ }
+ last_cleaned = found->backup_offset + found->backup_size;
+ cur = rb_next(&found->mob_node);
+ if (!cur)
+ break;
+
+ found = container_of(cur, struct vmw_resource, mob_node);
+ if (found->backup_offset >= res_end)
+ break;
+ }
+
+ /*
+ * Set number of pages allowed prefaulting and fence the buffer object
+ */
+ *num_prefault = 1;
+ if (last_cleaned > res_start) {
+ struct ttm_buffer_object *bo = &vbo->base;
+
+ *num_prefault = __KERNEL_DIV_ROUND_UP(last_cleaned - res_start,
+ PAGE_SIZE);
+ vmw_bo_fence_single(bo, NULL);
+ if (bo->moving)
+ dma_fence_put(bo->moving);
+ bo->moving = dma_fence_get
+ (dma_resv_get_excl(bo->base.resv));
+ }
+
+ return 0;
+}
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource_priv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_resource_priv.h
index c85144286cfe..3b7438b2d289 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource_priv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource_priv.h
@@ -77,6 +77,7 @@ struct vmw_user_resource_conv {
* @dirty_sync: Upload the dirty mob contents to the resource.
* @dirty_add_range: Add a sequential dirty range to the resource
* dirty tracker.
+ * @clean: Clean the resource.
*/
struct vmw_res_func {
enum vmw_res_type res_type;
@@ -101,6 +102,7 @@ struct vmw_res_func {
int (*dirty_sync)(struct vmw_resource *res);
void (*dirty_range_add)(struct vmw_resource *res, size_t start,
size_t end);
+ int (*clean)(struct vmw_resource *res);
};
/**
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c b/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
index 71349a7bae90..9aaf807ed73c 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_validation.c
@@ -641,7 +641,8 @@ int vmw_validation_res_validate(struct vmw_validation_context *ctx, bool intr)
struct vmw_resource *res = val->res;
struct vmw_buffer_object *backup = res->backup;
- ret = vmw_resource_validate(res, intr);
+ ret = vmw_resource_validate(res, intr, val->dirty_set &&
+ val->dirty);
if (ret) {
if (ret != -ERESTARTSYS)
DRM_ERROR("Failed to validate resource.\n");
--
2.20.1
On 10/14/19 3:22 PM, Thomas Hellstr?m (VMware) wrote:
> From: Thomas Hellstr?m <[email protected]>
>
> Graphics APIs like OpenGL 4.4 and Vulkan require the graphics driver
> to provide coherent graphics memory, meaning that the GPU sees any
> content written to the coherent memory on the next GPU operation that
> touches that memory, and the CPU sees any content written by the GPU
> to that memory immediately after any fence object trailing the GPU
> operation is signaled.
>
> Paravirtual drivers that otherwise require explicit synchronization
> needs to do this by hooking up dirty tracking to pagefault handlers
> and buffer object validation.
>
> Provide mm helpers needed for this and that also allow for huge pmd-
> and pud entries (patch 1-3), and the associated vmwgfx code (patch 4-7).
>
> The code has been tested and exercised by a tailored version of mesa
> where we disable all explicit synchronization and assume graphics memory
> is coherent. The performance loss varies of course; a typical number is
> around 5%.
>
> I would like to merge this code through the DRM tree, so an ack to include
> the new mm helpers in that merge would be greatly appreciated.
>
> Changes since RFC:
> - Merge conflict changes moved to the correct patch. Fixes intra-patchset
> compile errors.
> - Be more aggressive when turning ttm vm code into helpers. This makes sure
> we can use a const qualifier on the vmwgfx vm_ops.
> - Reinstate a lost comment an fix an error path that was broken when turning
> the ttm vm code into helpers.
> - Remove explicit type-casts of struct vm_area_struct::vm_private_data
> - Clarify the locking inversion that makes us not being able to use the mm
> pagewalk code.
>
> Changes since v1:
> - Removed the vmwgfx maintainer entry for as_dirty_helpers.c, updated
> commit message accordingly
> - Removed the TTM patches from the series as they are merged separately
> through DRM.
> Changes since v2:
> - Split out the pagewalk code from as_dirty_helpers.c and document locking.
> - Add pre_vma and post_vma callbacks to the pagewalk code.
> - Remove huge pmd and -pud asserts that would trip when we protect vmas with
> struct address_space::i_mmap_rwsem rather than with
> struct vm_area_struct::mmap_sem.
> - Do some naming cleanup in as_dirty_helpers.c
> Changes since v3:
> - Extensive renaming of the dirty helpers including the filename.
> - Update walk_page_mapping() doc.
> - Update the pagewalk code to not unconditionally split pmds if a pte_entry()
> callback is present. Update the dirty helper pmd_entry accordingly.
> - Use separate walk ops for the dirty helpers.
> - Update the pagewalk code to take the pagetable lock in walk_pte_range.
> Changes since v4:
> - Fix pte pointer confusion in patch 2/8
> - Skip the pagewalk code conditional split patch for now, and update the
> mapping_dirty_helper accordingly. That problem will be solved in a cleaner
> way in a follow-up patchset.
> Changes since v5:
> - Fix tlb flushing when we have other pending tlb flushes.
>
> Cc: Andrew Morton <[email protected]>
> Cc: Matthew Wilcox <[email protected]>
> Cc: Will Deacon <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Rik van Riel <[email protected]>
> Cc: Minchan Kim <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Huang Ying <[email protected]>
> Cc: J?r?me Glisse <[email protected]>
> Cc: Kirill A. Shutemov <[email protected]>
>
Kirill, Linus
I have a formal Ack for two of the four mm patches. Is there a chance I
can get an ack to merge the mm patches of this series through drm with
the vmwgfx patches?
Thanks,
Thomas
Hi, All,
On 10/14/19 3:22 PM, Thomas Hellström (VMware) wrote:
> From: Thomas Hellström <[email protected]>
>
> Graphics APIs like OpenGL 4.4 and Vulkan require the graphics driver
> to provide coherent graphics memory, meaning that the GPU sees any
> content written to the coherent memory on the next GPU operation that
> touches that memory, and the CPU sees any content written by the GPU
> to that memory immediately after any fence object trailing the GPU
> operation is signaled.
>
> Paravirtual drivers that otherwise require explicit synchronization
> needs to do this by hooking up dirty tracking to pagefault handlers
> and buffer object validation.
>
> Provide mm helpers needed for this and that also allow for huge pmd-
> and pud entries (patch 1-3), and the associated vmwgfx code (patch 4-7).
>
> The code has been tested and exercised by a tailored version of mesa
> where we disable all explicit synchronization and assume graphics memory
> is coherent. The performance loss varies of course; a typical number is
> around 5%.
>
> I would like to merge this code through the DRM tree, so an ack to include
> the new mm helpers in that merge would be greatly appreciated.
>
I'm a bit confused as how to get this merged? Is there an -mm maintainer
or who is supposed to ack -mm patches and get them into the kernel?
Any input appreciated,
Thanks,
Thomas
On Mon, Nov 4, 2019 at 1:21 AM Thomas Hellström (VMware)
<[email protected]> wrote:
>
> I'm a bit confused as how to get this merged? Is there an -mm maintainer
> or who is supposed to ack -mm patches and get them into the kernel?
I was assuming they'd go through Andrew Morton.. Andrew?
Linus
On Mon, 14 Oct 2019 15:22:00 +0200 Thomas Hellstr?m (VMware) <[email protected]> wrote:
> Add two utilities to 1) write-protect and 2) clean all ptes pointing into
> a range of an address space.
> The utilities are intended to aid in tracking dirty pages (either
> driver-allocated system memory or pci device memory).
> The write-protect utility should be used in conjunction with
> page_mkwrite() and pfn_mkwrite() to trigger write page-faults on page
> accesses. Typically one would want to use this on sparse accesses into
> large memory regions. The clean utility should be used to utilize
> hardware dirtying functionality and avoid the overhead of page-faults,
> typically on large accesses into small memory regions.
Not fully comfortable reviewing this one.
> --- a/mm/Kconfig
> +++ b/mm/Kconfig
> @@ -736,4 +736,7 @@ config ARCH_HAS_PTE_SPECIAL
> config ARCH_HAS_HUGEPD
> bool
>
> +config MAPPING_DIRTY_HELPERS
> + bool
> +
But given this, it's your problem ;) So
Acked-by: Andrew Morton <[email protected]>
Yes, please proceed with merging [1-4] via a drm tree.