Hi all,
This is a v2 of the patch series previously posted here:
https://lore.kernel.org/r/[email protected]/
Please refer to the cover letter of v1 for the context and motivation
behind the series.
Changes since v1:
- Added comment/doc in page_alloc.c explaining the locking expectations;
- Removed unnecessary casts in the 'page-to-list-node' helpers;
- A few other cosmetic changes in the allocator.
Thanks,
Quentin
Quentin Perret (7):
KVM: arm64: Move hyp_pool locking out of refcount helpers
KVM: arm64: Use refcount at hyp to check page availability
KVM: arm64: Remove list_head from hyp_page
KVM: arm64: Unify MMIO and mem host stage-2 pools
KVM: arm64: Remove hyp_pool pointer from struct hyp_page
KVM: arm64: Use less bits for hyp_page order
KVM: arm64: Use less bits for hyp_page refcount
arch/arm64/kvm/hyp/include/nvhe/gfp.h | 45 +------
arch/arm64/kvm/hyp/include/nvhe/mem_protect.h | 2 +-
arch/arm64/kvm/hyp/include/nvhe/memory.h | 7 +-
arch/arm64/kvm/hyp/include/nvhe/mm.h | 13 +-
arch/arm64/kvm/hyp/nvhe/mem_protect.c | 60 +++++-----
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 112 +++++++++++++-----
arch/arm64/kvm/hyp/nvhe/setup.c | 30 +++--
arch/arm64/kvm/hyp/reserved_mem.c | 3 +-
8 files changed, 145 insertions(+), 127 deletions(-)
--
2.32.0.rc0.204.g9fa02ecfa5-goog
The hyp buddy allocator currently checks the struct hyp_page list node
to see if a page is available for allocation or not when trying to
coalesce memory. Now that decrementing the refcount and attaching to
the buddy tree is done in the same critical section, we can rely on the
refcount of the buddy page to be in sync, which allows to replace the
list node check by a refcount check. This will ease removing the list
node from struct hyp_page later on.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 16 +++++++++++-----
1 file changed, 11 insertions(+), 5 deletions(-)
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index d666f4789e31..2602577daa00 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -55,7 +55,7 @@ static struct hyp_page *__find_buddy_avail(struct hyp_pool *pool,
{
struct hyp_page *buddy = __find_buddy_nocheck(pool, p, order);
- if (!buddy || buddy->order != order || list_empty(&buddy->node))
+ if (!buddy || buddy->order != order || buddy->refcount)
return NULL;
return buddy;
@@ -133,6 +133,12 @@ static inline void hyp_set_page_refcounted(struct hyp_page *p)
p->refcount = 1;
}
+static void __hyp_put_page(struct hyp_pool *pool, struct hyp_page *p)
+{
+ if (hyp_page_ref_dec_and_test(p))
+ __hyp_attach_page(pool, p);
+}
+
/*
* Changes to the buddy tree and page refcounts must be done with the hyp_pool
* lock held. If a refcount change requires an update to the buddy tree (e.g.
@@ -146,8 +152,7 @@ void hyp_put_page(void *addr)
struct hyp_pool *pool = hyp_page_to_pool(p);
hyp_spin_lock(&pool->lock);
- if (hyp_page_ref_dec_and_test(p))
- __hyp_attach_page(pool, p);
+ __hyp_put_page(pool, p);
hyp_spin_unlock(&pool->lock);
}
@@ -202,15 +207,16 @@ int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages,
/* Init the vmemmap portion */
p = hyp_phys_to_page(phys);
- memset(p, 0, sizeof(*p) * nr_pages);
for (i = 0; i < nr_pages; i++) {
p[i].pool = pool;
+ p[i].order = 0;
INIT_LIST_HEAD(&p[i].node);
+ hyp_set_page_refcounted(&p[i]);
}
/* Attach the unused pages to the buddy tree */
for (i = reserved_pages; i < nr_pages; i++)
- __hyp_attach_page(pool, &p[i]);
+ __hyp_put_page(pool, &p[i]);
return 0;
}
--
2.32.0.rc0.204.g9fa02ecfa5-goog
The hyp_page order is currently encoded on 4 bytes even though it is
guaranteed to be smaller than this. Make it 2 bytes to reduce the hyp
vmemmap overhead.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/gfp.h | 6 +++---
arch/arm64/kvm/hyp/include/nvhe/memory.h | 2 +-
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 12 ++++++------
3 files changed, 10 insertions(+), 10 deletions(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/gfp.h b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
index 3ea7bfb6c380..fb0f523d1492 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/gfp.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
@@ -7,7 +7,7 @@
#include <nvhe/memory.h>
#include <nvhe/spinlock.h>
-#define HYP_NO_ORDER UINT_MAX
+#define HYP_NO_ORDER USHRT_MAX
struct hyp_pool {
/*
@@ -19,11 +19,11 @@ struct hyp_pool {
struct list_head free_area[MAX_ORDER];
phys_addr_t range_start;
phys_addr_t range_end;
- unsigned int max_order;
+ unsigned short max_order;
};
/* Allocation */
-void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order);
+void *hyp_alloc_pages(struct hyp_pool *pool, unsigned short order);
void hyp_get_page(struct hyp_pool *pool, void *addr);
void hyp_put_page(struct hyp_pool *pool, void *addr);
diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h
index 991636be2f46..3fe34fa30ea4 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/memory.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h
@@ -9,7 +9,7 @@
struct hyp_page {
unsigned int refcount;
- unsigned int order;
+ unsigned short order;
};
extern u64 __hyp_vmemmap;
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index e3689def7033..be07055bbc10 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -32,7 +32,7 @@ u64 __hyp_vmemmap;
*/
static struct hyp_page *__find_buddy_nocheck(struct hyp_pool *pool,
struct hyp_page *p,
- unsigned int order)
+ unsigned short order)
{
phys_addr_t addr = hyp_page_to_phys(p);
@@ -51,7 +51,7 @@ static struct hyp_page *__find_buddy_nocheck(struct hyp_pool *pool,
/* Find a buddy page currently available for allocation */
static struct hyp_page *__find_buddy_avail(struct hyp_pool *pool,
struct hyp_page *p,
- unsigned int order)
+ unsigned short order)
{
struct hyp_page *buddy = __find_buddy_nocheck(pool, p, order);
@@ -93,7 +93,7 @@ static inline struct hyp_page *node_to_page(struct list_head *node)
static void __hyp_attach_page(struct hyp_pool *pool,
struct hyp_page *p)
{
- unsigned int order = p->order;
+ unsigned short order = p->order;
struct hyp_page *buddy;
memset(hyp_page_to_virt(p), 0, PAGE_SIZE << p->order);
@@ -123,7 +123,7 @@ static void __hyp_attach_page(struct hyp_pool *pool,
static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
struct hyp_page *p,
- unsigned int order)
+ unsigned short order)
{
struct hyp_page *buddy;
@@ -192,9 +192,9 @@ void hyp_get_page(struct hyp_pool *pool, void *addr)
hyp_spin_unlock(&pool->lock);
}
-void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order)
+void *hyp_alloc_pages(struct hyp_pool *pool, unsigned short order)
{
- unsigned int i = order;
+ unsigned short i = order;
struct hyp_page *p;
hyp_spin_lock(&pool->lock);
--
2.32.0.rc0.204.g9fa02ecfa5-goog
The hyp_page refcount helpers currently rely on the hyp_pool lock for
serialization. However, this means the refcounts can't be changed from
the buddy allocator core as it already holds the lock, which means pages
have to go through odd transient states.
For example, when a page is freed, its refcount is set to 0, and the
lock is transiently released before the page can be attached to a free
list in the buddy tree. This is currently harmless as the allocator
checks the list node of each page to see if it is available for
allocation or not, but it means the page refcount can't be trusted to
represent the state of the page even if the pool lock is held.
In order to fix this, remove the pool locking from the refcount helpers,
and move all the logic to the buddy allocator. This will simplify the
removal of the list node from struct hyp_page in a later patch.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/gfp.h | 35 ----------------------
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 43 ++++++++++++++++++++-------
2 files changed, 32 insertions(+), 46 deletions(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/gfp.h b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
index 18a4494337bd..f2c84e4fa40f 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/gfp.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
@@ -22,41 +22,6 @@ struct hyp_pool {
unsigned int max_order;
};
-static inline void hyp_page_ref_inc(struct hyp_page *p)
-{
- struct hyp_pool *pool = hyp_page_to_pool(p);
-
- hyp_spin_lock(&pool->lock);
- p->refcount++;
- hyp_spin_unlock(&pool->lock);
-}
-
-static inline int hyp_page_ref_dec_and_test(struct hyp_page *p)
-{
- struct hyp_pool *pool = hyp_page_to_pool(p);
- int ret;
-
- hyp_spin_lock(&pool->lock);
- p->refcount--;
- ret = (p->refcount == 0);
- hyp_spin_unlock(&pool->lock);
-
- return ret;
-}
-
-static inline void hyp_set_page_refcounted(struct hyp_page *p)
-{
- struct hyp_pool *pool = hyp_page_to_pool(p);
-
- hyp_spin_lock(&pool->lock);
- if (p->refcount) {
- hyp_spin_unlock(&pool->lock);
- BUG();
- }
- p->refcount = 1;
- hyp_spin_unlock(&pool->lock);
-}
-
/* Allocation */
void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order);
void hyp_get_page(void *addr);
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index 237e03bf0cb1..d666f4789e31 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -93,15 +93,6 @@ static void __hyp_attach_page(struct hyp_pool *pool,
list_add_tail(&p->node, &pool->free_area[order]);
}
-static void hyp_attach_page(struct hyp_page *p)
-{
- struct hyp_pool *pool = hyp_page_to_pool(p);
-
- hyp_spin_lock(&pool->lock);
- __hyp_attach_page(pool, p);
- hyp_spin_unlock(&pool->lock);
-}
-
static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
struct hyp_page *p,
unsigned int order)
@@ -125,19 +116,49 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
return p;
}
+static inline void hyp_page_ref_inc(struct hyp_page *p)
+{
+ p->refcount++;
+}
+
+static inline int hyp_page_ref_dec_and_test(struct hyp_page *p)
+{
+ p->refcount--;
+ return (p->refcount == 0);
+}
+
+static inline void hyp_set_page_refcounted(struct hyp_page *p)
+{
+ BUG_ON(p->refcount);
+ p->refcount = 1;
+}
+
+/*
+ * Changes to the buddy tree and page refcounts must be done with the hyp_pool
+ * lock held. If a refcount change requires an update to the buddy tree (e.g.
+ * hyp_put_page()), both operations must be done within the same critical
+ * section to guarantee transient states (e.g. a page with null refcount but
+ * not yet attached to a free list) can't be observed by well-behaved readers.
+ */
void hyp_put_page(void *addr)
{
struct hyp_page *p = hyp_virt_to_page(addr);
+ struct hyp_pool *pool = hyp_page_to_pool(p);
+ hyp_spin_lock(&pool->lock);
if (hyp_page_ref_dec_and_test(p))
- hyp_attach_page(p);
+ __hyp_attach_page(pool, p);
+ hyp_spin_unlock(&pool->lock);
}
void hyp_get_page(void *addr)
{
struct hyp_page *p = hyp_virt_to_page(addr);
+ struct hyp_pool *pool = hyp_page_to_pool(p);
+ hyp_spin_lock(&pool->lock);
hyp_page_ref_inc(p);
+ hyp_spin_unlock(&pool->lock);
}
void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order)
@@ -159,8 +180,8 @@ void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order)
p = list_first_entry(&pool->free_area[i], struct hyp_page, node);
p = __hyp_extract_page(pool, p, order);
- hyp_spin_unlock(&pool->lock);
hyp_set_page_refcounted(p);
+ hyp_spin_unlock(&pool->lock);
return hyp_page_to_virt(p);
}
--
2.32.0.rc0.204.g9fa02ecfa5-goog
The list_head member of struct hyp_page is only needed when the page is
attached to a free-list, which by definition implies the page is free.
As such, nothing prevents us from using the page itself to store the
list_head, hence reducing the size of the vmemmap.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/memory.h | 1 -
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 39 ++++++++++++++++++++----
2 files changed, 33 insertions(+), 7 deletions(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h
index fd78bde939ee..7691ab495eb4 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/memory.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h
@@ -12,7 +12,6 @@ struct hyp_page {
unsigned int refcount;
unsigned int order;
struct hyp_pool *pool;
- struct list_head node;
};
extern u64 __hyp_vmemmap;
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index 2602577daa00..34f0eb026dd2 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -62,6 +62,34 @@ static struct hyp_page *__find_buddy_avail(struct hyp_pool *pool,
}
+/*
+ * Pages that are available for allocation are tracked in free-lists, so we use
+ * the pages themselves to store the list nodes to avoid wasting space. As the
+ * allocator always returns zeroed pages (which are zeroed on the hyp_put_page()
+ * path to optimize allocation speed), we also need to clean-up the list node in
+ * each page when we take it out of the list.
+ */
+static inline void page_remove_from_list(struct hyp_page *p)
+{
+ struct list_head *node = hyp_page_to_virt(p);
+
+ __list_del_entry(node);
+ memset(node, 0, sizeof(*node));
+}
+
+static inline void page_add_to_list(struct hyp_page *p, struct list_head *head)
+{
+ struct list_head *node = hyp_page_to_virt(p);
+
+ INIT_LIST_HEAD(node);
+ list_add_tail(node, head);
+}
+
+static inline struct hyp_page *node_to_page(struct list_head *node)
+{
+ return hyp_virt_to_page(node);
+}
+
static void __hyp_attach_page(struct hyp_pool *pool,
struct hyp_page *p)
{
@@ -83,14 +111,14 @@ static void __hyp_attach_page(struct hyp_pool *pool,
break;
/* Take the buddy out of its list, and coallesce with @p */
- list_del_init(&buddy->node);
+ page_remove_from_list(buddy);
buddy->order = HYP_NO_ORDER;
p = min(p, buddy);
}
/* Mark the new head, and insert it */
p->order = order;
- list_add_tail(&p->node, &pool->free_area[order]);
+ page_add_to_list(p, &pool->free_area[order]);
}
static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
@@ -99,7 +127,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
{
struct hyp_page *buddy;
- list_del_init(&p->node);
+ page_remove_from_list(p);
while (p->order > order) {
/*
* The buddy of order n - 1 currently has HYP_NO_ORDER as it
@@ -110,7 +138,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
p->order--;
buddy = __find_buddy_nocheck(pool, p, p->order);
buddy->order = p->order;
- list_add_tail(&buddy->node, &pool->free_area[buddy->order]);
+ page_add_to_list(buddy, &pool->free_area[buddy->order]);
}
return p;
@@ -182,7 +210,7 @@ void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order)
}
/* Extract it from the tree at the right order */
- p = list_first_entry(&pool->free_area[i], struct hyp_page, node);
+ p = node_to_page(pool->free_area[i].next);
p = __hyp_extract_page(pool, p, order);
hyp_set_page_refcounted(p);
@@ -210,7 +238,6 @@ int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages,
for (i = 0; i < nr_pages; i++) {
p[i].pool = pool;
p[i].order = 0;
- INIT_LIST_HEAD(&p[i].node);
hyp_set_page_refcounted(&p[i]);
}
--
2.32.0.rc0.204.g9fa02ecfa5-goog
Each struct hyp_page currently contains a pointer to a hyp_pool struct
where the page should be freed if its refcount reaches 0. However, this
information can always be inferred from the context in the EL2 code, so
drop the pointer to save a few bytes in the vmemmap.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/gfp.h | 4 ++--
arch/arm64/kvm/hyp/include/nvhe/memory.h | 2 --
arch/arm64/kvm/hyp/nvhe/mem_protect.c | 14 ++++++++++++--
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 7 ++-----
arch/arm64/kvm/hyp/nvhe/setup.c | 14 ++++++++++++--
5 files changed, 28 insertions(+), 13 deletions(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/gfp.h b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
index f2c84e4fa40f..3ea7bfb6c380 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/gfp.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/gfp.h
@@ -24,8 +24,8 @@ struct hyp_pool {
/* Allocation */
void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order);
-void hyp_get_page(void *addr);
-void hyp_put_page(void *addr);
+void hyp_get_page(struct hyp_pool *pool, void *addr);
+void hyp_put_page(struct hyp_pool *pool, void *addr);
/* Used pages cannot be freed */
int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages,
diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h
index 7691ab495eb4..991636be2f46 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/memory.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h
@@ -7,11 +7,9 @@
#include <linux/types.h>
-struct hyp_pool;
struct hyp_page {
unsigned int refcount;
unsigned int order;
- struct hyp_pool *pool;
};
extern u64 __hyp_vmemmap;
diff --git a/arch/arm64/kvm/hyp/nvhe/mem_protect.c b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
index fdd5b5702e8a..7727252890d8 100644
--- a/arch/arm64/kvm/hyp/nvhe/mem_protect.c
+++ b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
@@ -43,6 +43,16 @@ static void *host_s2_zalloc_page(void *pool)
return hyp_alloc_pages(pool, 0);
}
+static void host_s2_get_page(void *addr)
+{
+ hyp_get_page(&host_s2_pool, addr);
+}
+
+static void host_s2_put_page(void *addr)
+{
+ hyp_put_page(&host_s2_pool, addr);
+}
+
static int prepare_s2_pool(void *pgt_pool_base)
{
unsigned long nr_pages, pfn;
@@ -60,8 +70,8 @@ static int prepare_s2_pool(void *pgt_pool_base)
.phys_to_virt = hyp_phys_to_virt,
.virt_to_phys = hyp_virt_to_phys,
.page_count = hyp_page_count,
- .get_page = hyp_get_page,
- .put_page = hyp_put_page,
+ .get_page = host_s2_get_page,
+ .put_page = host_s2_put_page,
};
return 0;
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index 34f0eb026dd2..e3689def7033 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -174,20 +174,18 @@ static void __hyp_put_page(struct hyp_pool *pool, struct hyp_page *p)
* section to guarantee transient states (e.g. a page with null refcount but
* not yet attached to a free list) can't be observed by well-behaved readers.
*/
-void hyp_put_page(void *addr)
+void hyp_put_page(struct hyp_pool *pool, void *addr)
{
struct hyp_page *p = hyp_virt_to_page(addr);
- struct hyp_pool *pool = hyp_page_to_pool(p);
hyp_spin_lock(&pool->lock);
__hyp_put_page(pool, p);
hyp_spin_unlock(&pool->lock);
}
-void hyp_get_page(void *addr)
+void hyp_get_page(struct hyp_pool *pool, void *addr)
{
struct hyp_page *p = hyp_virt_to_page(addr);
- struct hyp_pool *pool = hyp_page_to_pool(p);
hyp_spin_lock(&pool->lock);
hyp_page_ref_inc(p);
@@ -236,7 +234,6 @@ int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages,
/* Init the vmemmap portion */
p = hyp_phys_to_page(phys);
for (i = 0; i < nr_pages; i++) {
- p[i].pool = pool;
p[i].order = 0;
hyp_set_page_refcounted(&p[i]);
}
diff --git a/arch/arm64/kvm/hyp/nvhe/setup.c b/arch/arm64/kvm/hyp/nvhe/setup.c
index 709cb3d19eb7..dee099871865 100644
--- a/arch/arm64/kvm/hyp/nvhe/setup.c
+++ b/arch/arm64/kvm/hyp/nvhe/setup.c
@@ -137,6 +137,16 @@ static void *hyp_zalloc_hyp_page(void *arg)
return hyp_alloc_pages(&hpool, 0);
}
+static void hpool_get_page(void *addr)
+{
+ hyp_get_page(&hpool, addr);
+}
+
+static void hpool_put_page(void *addr)
+{
+ hyp_put_page(&hpool, addr);
+}
+
void __noreturn __pkvm_init_finalise(void)
{
struct kvm_host_data *host_data = this_cpu_ptr(&kvm_host_data);
@@ -160,8 +170,8 @@ void __noreturn __pkvm_init_finalise(void)
.zalloc_page = hyp_zalloc_hyp_page,
.phys_to_virt = hyp_phys_to_virt,
.virt_to_phys = hyp_virt_to_phys,
- .get_page = hyp_get_page,
- .put_page = hyp_put_page,
+ .get_page = hpool_get_page,
+ .put_page = hpool_put_page,
};
pkvm_pgtable.mm_ops = &pkvm_pgtable_mm_ops;
--
2.32.0.rc0.204.g9fa02ecfa5-goog
We currently maintain two separate memory pools for the host stage-2,
one for pages used in the page-table when mapping memory regions, and
the other to map MMIO regions. The former is large enough to map all of
memory with page granularity and the latter can cover an arbitrary
portion of IPA space, but allows to 'recycle' pages.
However, this split makes accounting difficult to manage as pages at
intermediate levels of the page-table may be used to map both memory and
MMIO regions. Simplify the scheme by merging both pools into one. This
means we can now hit the -ENOMEM case in the memory abort path, but
we're still guaranteed forward-progress in the worst case by unmapping
MMIO regions. On the plus side this also means we can usually map a lot
more MMIO space at once if memory ranges happen to be mapped with block
mappings.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/mem_protect.h | 2 +-
arch/arm64/kvm/hyp/include/nvhe/mm.h | 13 +++---
arch/arm64/kvm/hyp/nvhe/mem_protect.c | 46 ++++++++-----------
arch/arm64/kvm/hyp/nvhe/setup.c | 16 ++-----
arch/arm64/kvm/hyp/reserved_mem.c | 3 +-
5 files changed, 32 insertions(+), 48 deletions(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h b/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
index 42d81ec739fa..9c227d87c36d 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/mem_protect.h
@@ -23,7 +23,7 @@ extern struct host_kvm host_kvm;
int __pkvm_prot_finalize(void);
int __pkvm_mark_hyp(phys_addr_t start, phys_addr_t end);
-int kvm_host_prepare_stage2(void *mem_pgt_pool, void *dev_pgt_pool);
+int kvm_host_prepare_stage2(void *pgt_pool_base);
void handle_host_mem_abort(struct kvm_cpu_context *host_ctxt);
static __always_inline void __load_host_stage2(void)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/mm.h b/arch/arm64/kvm/hyp/include/nvhe/mm.h
index 0095f6289742..8ec3a5a7744b 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/mm.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/mm.h
@@ -78,19 +78,20 @@ static inline unsigned long hyp_s1_pgtable_pages(void)
return res;
}
-static inline unsigned long host_s2_mem_pgtable_pages(void)
+static inline unsigned long host_s2_pgtable_pages(void)
{
+ unsigned long res;
+
/*
* Include an extra 16 pages to safely upper-bound the worst case of
* concatenated pgds.
*/
- return __hyp_pgtable_total_pages() + 16;
-}
+ res = __hyp_pgtable_total_pages() + 16;
-static inline unsigned long host_s2_dev_pgtable_pages(void)
-{
/* Allow 1 GiB for MMIO mappings */
- return __hyp_pgtable_max_pages(SZ_1G >> PAGE_SHIFT);
+ res += __hyp_pgtable_max_pages(SZ_1G >> PAGE_SHIFT);
+
+ return res;
}
#endif /* __KVM_HYP_MM_H */
diff --git a/arch/arm64/kvm/hyp/nvhe/mem_protect.c b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
index e342f7f4f4fb..fdd5b5702e8a 100644
--- a/arch/arm64/kvm/hyp/nvhe/mem_protect.c
+++ b/arch/arm64/kvm/hyp/nvhe/mem_protect.c
@@ -23,8 +23,7 @@
extern unsigned long hyp_nr_cpus;
struct host_kvm host_kvm;
-struct hyp_pool host_s2_mem;
-struct hyp_pool host_s2_dev;
+struct hyp_pool host_s2_pool;
/*
* Copies of the host's CPU features registers holding sanitized values.
@@ -36,7 +35,7 @@ static const u8 pkvm_hyp_id = 1;
static void *host_s2_zalloc_pages_exact(size_t size)
{
- return hyp_alloc_pages(&host_s2_mem, get_order(size));
+ return hyp_alloc_pages(&host_s2_pool, get_order(size));
}
static void *host_s2_zalloc_page(void *pool)
@@ -44,20 +43,14 @@ static void *host_s2_zalloc_page(void *pool)
return hyp_alloc_pages(pool, 0);
}
-static int prepare_s2_pools(void *mem_pgt_pool, void *dev_pgt_pool)
+static int prepare_s2_pool(void *pgt_pool_base)
{
unsigned long nr_pages, pfn;
int ret;
- pfn = hyp_virt_to_pfn(mem_pgt_pool);
- nr_pages = host_s2_mem_pgtable_pages();
- ret = hyp_pool_init(&host_s2_mem, pfn, nr_pages, 0);
- if (ret)
- return ret;
-
- pfn = hyp_virt_to_pfn(dev_pgt_pool);
- nr_pages = host_s2_dev_pgtable_pages();
- ret = hyp_pool_init(&host_s2_dev, pfn, nr_pages, 0);
+ pfn = hyp_virt_to_pfn(pgt_pool_base);
+ nr_pages = host_s2_pgtable_pages();
+ ret = hyp_pool_init(&host_s2_pool, pfn, nr_pages, 0);
if (ret)
return ret;
@@ -86,7 +79,7 @@ static void prepare_host_vtcr(void)
id_aa64mmfr1_el1_sys_val, phys_shift);
}
-int kvm_host_prepare_stage2(void *mem_pgt_pool, void *dev_pgt_pool)
+int kvm_host_prepare_stage2(void *pgt_pool_base)
{
struct kvm_s2_mmu *mmu = &host_kvm.arch.mmu;
int ret;
@@ -94,7 +87,7 @@ int kvm_host_prepare_stage2(void *mem_pgt_pool, void *dev_pgt_pool)
prepare_host_vtcr();
hyp_spin_lock_init(&host_kvm.lock);
- ret = prepare_s2_pools(mem_pgt_pool, dev_pgt_pool);
+ ret = prepare_s2_pool(pgt_pool_base);
if (ret)
return ret;
@@ -199,11 +192,10 @@ static bool range_is_memory(u64 start, u64 end)
}
static inline int __host_stage2_idmap(u64 start, u64 end,
- enum kvm_pgtable_prot prot,
- struct hyp_pool *pool)
+ enum kvm_pgtable_prot prot)
{
return kvm_pgtable_stage2_map(&host_kvm.pgt, start, end - start, start,
- prot, pool);
+ prot, &host_s2_pool);
}
static int host_stage2_idmap(u64 addr)
@@ -211,7 +203,6 @@ static int host_stage2_idmap(u64 addr)
enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R | KVM_PGTABLE_PROT_W;
struct kvm_mem_range range;
bool is_memory = find_mem_range(addr, &range);
- struct hyp_pool *pool = is_memory ? &host_s2_mem : &host_s2_dev;
int ret;
if (is_memory)
@@ -222,22 +213,21 @@ static int host_stage2_idmap(u64 addr)
if (ret)
goto unlock;
- ret = __host_stage2_idmap(range.start, range.end, prot, pool);
- if (is_memory || ret != -ENOMEM)
+ ret = __host_stage2_idmap(range.start, range.end, prot);
+ if (ret != -ENOMEM)
goto unlock;
/*
- * host_s2_mem has been provided with enough pages to cover all of
- * memory with page granularity, so we should never hit the ENOMEM case.
- * However, it is difficult to know how much of the MMIO range we will
- * need to cover upfront, so we may need to 'recycle' the pages if we
- * run out.
+ * The pool has been provided with enough pages to cover all of memory
+ * with page granularity, but it is difficult to know how much of the
+ * MMIO range we will need to cover upfront, so we may need to 'recycle'
+ * the pages if we run out.
*/
ret = host_stage2_unmap_dev_all();
if (ret)
goto unlock;
- ret = __host_stage2_idmap(range.start, range.end, prot, pool);
+ ret = __host_stage2_idmap(range.start, range.end, prot);
unlock:
hyp_spin_unlock(&host_kvm.lock);
@@ -258,7 +248,7 @@ int __pkvm_mark_hyp(phys_addr_t start, phys_addr_t end)
hyp_spin_lock(&host_kvm.lock);
ret = kvm_pgtable_stage2_set_owner(&host_kvm.pgt, start, end - start,
- &host_s2_mem, pkvm_hyp_id);
+ &host_s2_pool, pkvm_hyp_id);
hyp_spin_unlock(&host_kvm.lock);
return ret != -EAGAIN ? ret : 0;
diff --git a/arch/arm64/kvm/hyp/nvhe/setup.c b/arch/arm64/kvm/hyp/nvhe/setup.c
index 7488f53b0aa2..709cb3d19eb7 100644
--- a/arch/arm64/kvm/hyp/nvhe/setup.c
+++ b/arch/arm64/kvm/hyp/nvhe/setup.c
@@ -25,8 +25,7 @@ unsigned long hyp_nr_cpus;
static void *vmemmap_base;
static void *hyp_pgt_base;
-static void *host_s2_mem_pgt_base;
-static void *host_s2_dev_pgt_base;
+static void *host_s2_pgt_base;
static int divide_memory_pool(void *virt, unsigned long size)
{
@@ -45,14 +44,9 @@ static int divide_memory_pool(void *virt, unsigned long size)
if (!hyp_pgt_base)
return -ENOMEM;
- nr_pages = host_s2_mem_pgtable_pages();
- host_s2_mem_pgt_base = hyp_early_alloc_contig(nr_pages);
- if (!host_s2_mem_pgt_base)
- return -ENOMEM;
-
- nr_pages = host_s2_dev_pgtable_pages();
- host_s2_dev_pgt_base = hyp_early_alloc_contig(nr_pages);
- if (!host_s2_dev_pgt_base)
+ nr_pages = host_s2_pgtable_pages();
+ host_s2_pgt_base = hyp_early_alloc_contig(nr_pages);
+ if (!host_s2_pgt_base)
return -ENOMEM;
return 0;
@@ -158,7 +152,7 @@ void __noreturn __pkvm_init_finalise(void)
if (ret)
goto out;
- ret = kvm_host_prepare_stage2(host_s2_mem_pgt_base, host_s2_dev_pgt_base);
+ ret = kvm_host_prepare_stage2(host_s2_pgt_base);
if (ret)
goto out;
diff --git a/arch/arm64/kvm/hyp/reserved_mem.c b/arch/arm64/kvm/hyp/reserved_mem.c
index 83ca23ac259b..d654921dd09b 100644
--- a/arch/arm64/kvm/hyp/reserved_mem.c
+++ b/arch/arm64/kvm/hyp/reserved_mem.c
@@ -71,8 +71,7 @@ void __init kvm_hyp_reserve(void)
}
hyp_mem_pages += hyp_s1_pgtable_pages();
- hyp_mem_pages += host_s2_mem_pgtable_pages();
- hyp_mem_pages += host_s2_dev_pgtable_pages();
+ hyp_mem_pages += host_s2_pgtable_pages();
/*
* The hyp_vmemmap needs to be backed by pages, but these pages
--
2.32.0.rc0.204.g9fa02ecfa5-goog
The hyp_page refcount is currently encoded on 4 bytes even though we
never need to count that many objects in a page. Make it 2 bytes to save
some space in the vmemmap.
As overflows are more likely to happen as well, make sure to catch those
with a BUG in the increment function.
Signed-off-by: Quentin Perret <[email protected]>
---
arch/arm64/kvm/hyp/include/nvhe/memory.h | 2 +-
arch/arm64/kvm/hyp/nvhe/page_alloc.c | 1 +
2 files changed, 2 insertions(+), 1 deletion(-)
diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h
index 3fe34fa30ea4..592b7edb3edb 100644
--- a/arch/arm64/kvm/hyp/include/nvhe/memory.h
+++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h
@@ -8,7 +8,7 @@
#include <linux/types.h>
struct hyp_page {
- unsigned int refcount;
+ unsigned short refcount;
unsigned short order;
};
diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
index be07055bbc10..41fc25bdfb34 100644
--- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c
+++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c
@@ -146,6 +146,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool,
static inline void hyp_page_ref_inc(struct hyp_page *p)
{
+ BUG_ON(p->refcount == USHRT_MAX);
p->refcount++;
}
--
2.32.0.rc0.204.g9fa02ecfa5-goog
On 2021-06-02 10:43, Quentin Perret wrote:
> We currently maintain two separate memory pools for the host stage-2,
> one for pages used in the page-table when mapping memory regions, and
> the other to map MMIO regions. The former is large enough to map all of
> memory with page granularity and the latter can cover an arbitrary
> portion of IPA space, but allows to 'recycle' pages.
>
> However, this split makes accounting difficult to manage as pages at
> intermediate levels of the page-table may be used to map both memory
> and
> MMIO regions. Simplify the scheme by merging both pools into one. This
> means we can now hit the -ENOMEM case in the memory abort path, but
> we're still guaranteed forward-progress in the worst case by unmapping
> MMIO regions. On the plus side this also means we can usually map a lot
> more MMIO space at once if memory ranges happen to be mapped with block
> mappings.
>
> Signed-off-by: Quentin Perret <[email protected]>
This patch fails to apply on top of -rc4. Did you use some more exotic
base or, as I suspect, a development tree?
Please check and respin the series if necessary.
M.
--
Jazz is not dead. It just smells funny...
On Sunday 06 Jun 2021 at 11:31:20 (+0100), Marc Zyngier wrote:
> On 2021-06-02 10:43, Quentin Perret wrote:
> > We currently maintain two separate memory pools for the host stage-2,
> > one for pages used in the page-table when mapping memory regions, and
> > the other to map MMIO regions. The former is large enough to map all of
> > memory with page granularity and the latter can cover an arbitrary
> > portion of IPA space, but allows to 'recycle' pages.
> >
> > However, this split makes accounting difficult to manage as pages at
> > intermediate levels of the page-table may be used to map both memory and
> > MMIO regions. Simplify the scheme by merging both pools into one. This
> > means we can now hit the -ENOMEM case in the memory abort path, but
> > we're still guaranteed forward-progress in the worst case by unmapping
> > MMIO regions. On the plus side this also means we can usually map a lot
> > more MMIO space at once if memory ranges happen to be mapped with block
> > mappings.
> >
> > Signed-off-by: Quentin Perret <[email protected]>
>
> This patch fails to apply on top of -rc4. Did you use some more exotic
> base or, as I suspect, a development tree?
>
> Please check and respin the series if necessary.
I used kvmarm/next, but clearly an out-of-date one. I'll respin -- sorry
about that.
Thanks,
Quentin