KVM: Fix mmu_notifier vs. pfncache vs. pfncache races
Sean Christopherson
maintainerv2 cover KVM: Fix mmu_notifier vs. pfncache vs. pfncache races
Fix races between mmu_notifier invalidation and pfncache refresh, and within the pfncache itself.
The first two patches are reverts of the patches sitting in kvm/queue, trying to separate and fix the races independently is nigh impossible. I assume/hope they can be ignored and the original patches dropped.
I verified internal races using the attached hack-a-test. Running the test against the current implementation fails due to KVM writing the current GPA into the wrong page.
I don't think the race with the mmu_notifier is technically proven, e.g. I never encountered a use-after-free even running with KASAN.
Ran with PROVE_LOCKING and DEBUG_ATOMIC_SLEEP, so in theory there shouldn't be any lurking locking goofs this time...
v2:
- Map the pfn=>khva outside of gpc->lock. [Maxim]
- Fix a page leak.
- Fix more races.
v1:
https://lore.kernel.org/all/[email protected]
Sean Christopherson (8):
Revert "KVM: Do not speculatively mark pfn cache valid to "fix" race" Revert "KVM: Fix race between mmu_notifier invalidation and pfncache
refresh"
KVM: Drop unused @gpa param from gfn=>pfn cache's __release_gpc()
helper
KVM: Put the extra pfn reference when reusing a pfn in the gpc cache
KVM: Do not incorporate page offset into gfn=>pfn cache user address
KVM: Fix multiple races in gfn=>pfn cache refresh
KVM: Do not pin pages tracked by gfn=>pfn caches
DO NOT MERGE: Hack-a-test to verify gpc invalidation+refresh
arch/x86/kvm/x86.c | 30 ++++ include/linux/kvm_host.h | 2 + include/linux/kvm_types.h | 1 + tools/testing/selftests/kvm/.gitignore | 1 + tools/testing/selftests/kvm/Makefile | 2 + tools/testing/selftests/kvm/gpc_test.c | 217 +++++++++++++++++++++++++ virt/kvm/pfncache.c | 188 +++++++++++++-------- 7 files changed, 372 insertions(+), 69 deletions(-) create mode 100644 tools/testing/selftests/kvm/gpc_test.c
base-commit: 2a39d8b39bffdaf1a4223d0d22f07baee154c8f3
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 1/8 Revert "KVM: Do not speculatively mark pfn cache valid to "fix" race"
This reverts commit 55111927df1cd140aa7b7ea3f33f524b87776381.
--- virt/kvm/pfncache.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
virt/kvm/pfncache.c +6 -3
@@ -81,8 +81,6 @@ bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, {struct kvm_memslots *slots =kvm_memslots (kvm); -lockdep_assert_held_read (&gpc->lock); -if ((gpa & ~PAGE_MASK ) + len >PAGE_SIZE )return false;
@@ -228,6 +226,11 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,if (!old_valid || old_uhva != gpc->uhva) {void *new_khva =NULL ; +/* Placeholders for "hva is valid but not yet mapped" */ + gpc->pfn =KVM_PFN_ERR_FAULT ; + gpc->khva =NULL ; + gpc->valid = true; + new_pfn =hva_to_pfn_retry (kvm, gpc);if (is_error_noslot_pfn (new_pfn)) { ret = -EFAULT ;
@@ -256,7 +259,7 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, gpc->pfn =KVM_PFN_ERR_FAULT ; gpc->khva =NULL ; }else { - gpc->valid = true; +/* At this point, gpc->valid may already have been cleared */ gpc->pfn = new_pfn; gpc->khva = new_khva; }
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 2/8 Revert "KVM: Fix race between mmu_notifier invalidation and pfncache refresh"
In reply to Sean ChristophersonThis reverts commit c496097d2c0bdc229f82d72b4b1e55d64974c316.
--- virt/kvm/kvm_main.c | 9 ------ virt/kvm/pfncache.c | 70 ++++++++++++++------------------------------- 2 files changed, 21 insertions(+), 58 deletions(-)
virt/kvm/kvm_main.c +0 -9
@@ -705,15 +705,6 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, kvm->mn_active_invalidate_count++;spin_unlock (&kvm->mn_invalidate_lock); -/* -* Invalidate pfn caches _before_ invalidating the secondary MMUs, i.e. -* before acquiring mmu_lock, to avoid holding mmu_lock while acquiring -* each cache's lock. There are relatively few caches in existence at -* any given time, and the caches themselves can check for hva overlap, -* i.e. don't need to rely on memslot overlap checks for performance. -* Because this runs without holding mmu_lock, the pfn caches must use -* mn_active_invalidate_count (see above) instead of mmu_notifier_count. -*/ gfn_to_pfn_cache_invalidate_start (kvm, range->start, range->end, hva_range.may_block);
virt/kvm/pfncache.c +21 -49
110 lines not on this page · raw message
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 3/8 KVM: Drop unused @gpa param from gfn=>pfn cache's __release_gpc() helper
In reply to Sean ChristophersonDrop the @pga param from __release_gpc() and rename the helper to make it more obvious that the cache itself is not being released. The helper will be reused by a future commit to release a pfn+khva combination that is _never_ associated with the cache, at which point the current name would go from slightly misleading to blatantly wrong.
No functional change intended.
--- virt/kvm/pfncache.c | 10 +++------- 1 file changed, 3 insertions(+), 7 deletions(-)
virt/kvm/pfncache.c +3 -7
@@ -95,7 +95,7 @@ bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, }EXPORT_SYMBOL_GPL (kvm_gfn_to_pfn_cache_check); -static void __release_gpc (struct kvm *kvm,kvm_pfn_t pfn,void *khva,gpa_t gpa) +static void gpc_release_pfn_and_khva (struct kvm *kvm,kvm_pfn_t pfn,void *khva) {/* Unmap the old page if it was mapped before, and release it */ if (!is_error_noslot_pfn (pfn)) {
@@ -146,7 +146,6 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,unsigned long page_offset = gpa & ~PAGE_MASK ;kvm_pfn_t old_pfn, new_pfn;unsigned long old_uhva; -gpa_t old_gpa;void *old_khva;bool old_valid;int ret =0 ;
@@ -160,7 +159,6 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,write_lock_irq (&gpc->lock); - old_gpa = gpc->gpa; old_pfn = gpc->pfn; old_khva = gpc->khva -offset_in_page (gpc->khva); old_uhva = gpc->uhva;
@@ -244,7 +242,7 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, out:write_unlock_irq (&gpc->lock); -__release_gpc (kvm, old_pfn, old_khva, old_gpa); +gpc_release_pfn_and_khva (kvm, old_pfn, old_khva);return ret; }
@@ -254,14 +252,12 @@ void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc) {void *old_khva;kvm_pfn_t old_pfn; -gpa_t old_gpa;write_lock_irq (&gpc->lock); gpc->valid = false; old_khva = gpc->khva -offset_in_page (gpc->khva); - old_gpa = gpc->gpa; old_pfn = gpc->pfn;/*
@@ -273,7 +269,7 @@ void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)write_unlock_irq (&gpc->lock); -__release_gpc (kvm, old_pfn, old_khva, old_gpa); +gpc_release_pfn_and_khva (kvm, old_pfn, old_khva); }EXPORT_SYMBOL_GPL (kvm_gfn_to_pfn_cache_unmap);
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 4/8 KVM: Put the extra pfn reference when reusing a pfn in the gpc cache
In reply to Sean ChristophersonPut the struct page reference to pfn acquired by hva_to_pfn() when the old and new pfns for a gfn=>pfn cache match. The cache already has a reference via the old/current pfn, and will only put one reference when the cache is done with the pfn.
--- virt/kvm/pfncache.c | 8 ++++++++ 1 file changed, 8 insertions(+)
virt/kvm/pfncache.c +8 -0
@@ -206,6 +206,14 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,if (gpc->usage &KVM_HOST_USES_PFN ) {if (new_pfn == old_pfn) { +/* +* Reuse the existing pfn and khva, but put the +* reference acquired hva_to_pfn_retry(); the +* cache still holds a reference to the pfn +* from the previous refresh. +*/ +gpc_release_pfn_and_khva (kvm, new_pfn,NULL ); + new_khva = old_khva; old_pfn =KVM_PFN_ERR_FAULT ; old_khva =NULL ;
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 5/8 KVM: Do not incorporate page offset into gfn=>pfn cache user address
In reply to Sean ChristophersonDon't adjust the userspace address in the gfn=>pfn cache by the page offset from the gpa. KVM should never use the user address directly, and all KVM operations that translate a user address to something else require the user address to be page aligned. Ignoring the offset will allow the cache to reuse a gfn=>hva translation in the unlikely event that the page offset of the gpa changes, but the gfn does not.
--- virt/kvm/pfncache.c | 2 -- 1 file changed, 2 deletions(-)
virt/kvm/pfncache.c +0 -2
@@ -179,8 +179,6 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, ret = -EFAULT ; goto out; } - - gpc->uhva += page_offset; }/*
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 6/8 KVM: Fix multiple races in gfn=>pfn cache refresh
In reply to Sean ChristophersonRework the gfn=>pfn cache (gpc) refresh logic to address multiple races between the cache itself, and between the cache and mmu_notifier events.
The existing refresh code attempts to guard against races with the mmu_notifier by speculatively marking the cache valid, and then marking it invalid if a mmu_notifier invalidation occurs. That handles the case where an invalidation occurs between dropping and re-acquiring gpc->lock, but it doesn't handle the scenario where the cache is refreshed after the cache was invalidated by the notifier, but before the notifier elevates mmu_notifier_count. The gpc refresh can't use the "retry" helper as its invalidation occurs _before_ mmu_notifier_count is elevated and before mmu_notifier_range_start is set/updated.
CPU0 CPU1
---- ----
gfn_to_pfn_cache_invalidate_start()
|
-> gpc->valid = false;
kvm_gfn_to_pfn_cache_refresh()
|
|-> gpc->valid = true; hva_to_pfn_retry()
|
-> acquire kvm->mmu_lock
kvm->mmu_notifier_count == 0
mmu_seq == kvm->mmu_notifier_seq
drop kvm->mmu_lock
return pfn 'X'
acquire kvm->mmu_lock
kvm_inc_notifier_count()
drop kvm->mmu_lock()
kernel frees pfn 'X'
kvm_gfn_to_pfn_cache_check()
|
|-> gpc->valid == true
caller accesses freed pfn 'X'Key off of mn_active_invalidate_count to detect that a pfncache refresh needs to wait for an in-progress mmu_notifier invalidation. While mn_active_invalidate_count is not guaranteed to be stable, it is guaranteed to be elevated prior to an invalidation acquiring gpc->lock, so either the refresh will see an active invalidation and wait, or the invalidation will run after the refresh completes.
Speculatively marking the cache valid is itself flawed, as a concurrent kvm_gfn_to_pfn_cache_check() would see a valid cache with stale pfn/khva values. The KVM Xen use case explicitly allows/wants multiple users; even though the caches are allocated per vCPU, __kvm_xen_has_interrupt() can read a different vCPU (or vCPUs). Address this race by invalidating the cache prior to dropping gpc->lock (this is made possible by fixing the above mmu_notifier race).
Finally, the refresh logic doesn't protect against concurrent refreshes with different GPAs (which may or may not be a desired use case, but its allowed in the code), nor does it protect against a false negative on the memslot generation. If the first refresh sees a stale memslot generation, it will refresh the hva and generation before moving on to the hva=>pfn translation. If it then drops gpc->lock, a different user can come along, acquire gpc->lock, see that the memslot generation is fresh, and skip the hva=>pfn update due to the userspace address also matching (because it too was updated). Address this race by adding an "in-progress" flag so that the refresh that acquires gpc->lock first runs to completion before other users can start their refresh.
Complicating all of this is the fact that both the hva=>pfn resolution and mapping of the kernel address can sleep, i.e. must be done outside of gpc->lock
Fix the above races in one fell swoop, trying to fix each individual race in a sane manner is impossible, for all intents and purposes.
--- include/linux/kvm_types.h | 1 + virt/kvm/kvm_main.c | 9 ++ virt/kvm/pfncache.c | 209 +++++++++++++++++++++++++------------- 3 files changed, 148 insertions(+), 71 deletions(-)
include/linux/kvm_types.h +1 -0
@@ -74,6 +74,7 @@ struct gfn_to_pfn_cache {enum pfn_cache_usage usage;bool active;bool valid; +bool refresh_in_progress; };#ifdef KVM_ARCH_NR_OBJS_PER_MEMORY_CACHE
virt/kvm/kvm_main.c +9 -0
@@ -705,6 +705,15 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, kvm->mn_active_invalidate_count++;spin_unlock (&kvm->mn_invalidate_lock); +/* +* Invalidate pfn caches _before_ invalidating the secondary MMUs, i.e. +* before acquiring mmu_lock, to avoid holding mmu_lock while acquiring +* each cache's lock. There are relatively few caches in existence at +* any given time, and the caches themselves can check for hva overlap, +* i.e. don't need to rely on memslot overlap checks for performance. +* Because this runs without holding mmu_lock, the pfn caches must use +* mn_active_invalidate_count (see above) instead of mmu_notifier_count. +*/ gfn_to_pfn_cache_invalidate_start (kvm, range->start, range->end, hva_range.may_block);
virt/kvm/pfncache.c +138 -71
262 lines not on this page · raw message
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
Quoted 11 lines · Sean Christopherson
Finally, the refresh logic doesn't protect against concurrent refreshes with different GPAs (which may or may not be a desired use case, but its allowed in the code), nor does it protect against a false negative on the memslot generation. If the first refresh sees a stale memslot generation, it will refresh the hva and generation before moving on to the hva=>pfn translation. If it then drops
gpc->lock, a different user can come along, acquiregpc->lock, see that the memslot generation is fresh, and skip the hva=>pfn update due to the userspace address also matching (because it too was updated). Address this race by adding an "in-progress" flag so that the refresh that acquiresgpc->lockfirst runs to completion before other users can start their refresh.
...
Quoted 7 lines
virt/kvm/pfncache.c +3
-0@@ -159,10 +249,23 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,write_lock_irq (&gpc->lock); +/* +* If another task is refreshing the cache, wait for it to complete. +* There is no guarantee that concurrent refreshes will see the same
virt/kvm/pfncache.c int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, · line 264
Quoted 10 lines
virt/kvm/pfncache.c +10
-0+* gpa, memslots generation, etc..., so they must be fully serialized. +*/ +while (gpc->refresh_in_progress) { +write_unlock_irq (&gpc->lock); + +cond_resched (); + +write_lock_irq (&gpc->lock); + } + gpc->refresh_in_progress = true;
Adding refresh_in_progress can likely go in a separate patch. I'll plan on doing that in a v3 unless it proves to be painful.
Quoted 16 lines
virt/kvm/pfncache.c +12
-0@@ -246,9 +296,26 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, } out: +/* +* Invalidate the cache and purge the pfn/khva if the refresh failed. +* Some/all of the uhva, gpa, and memslot generation info may still be +* valid, leave it as is. +*/ +if (ret) { + gpc->valid = false; + gpc->pfn =KVM_PFN_ERR_FAULT ; + gpc->khva =NULL ; + } + + gpc->refresh_in_progress = false;
Re: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
In reply to Sean ChristophersonReview int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, · line 249
Quoted 2 lines · Sean Christopherson
(quoted diff) +0
-0@@ -159,10 +249,23 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
The following code of refresh_in_progress is somewhat like mutex.
+ mutex_lock(&gpc->refresh_in_progress); // before write_lock_irq(&gpc->lock);
Is it fit for the intention?
Thanks
Lai
Quoted 21 lines
virt/kvm/pfncache.c +14
-1write_lock_irq (&gpc->lock); +/* +* If another task is refreshing the cache, wait for it to complete. +* There is no guarantee that concurrent refreshes will see the same +* gpa, memslots generation, etc..., so they must be fully serialized. +*/ +while (gpc->refresh_in_progress) { +write_unlock_irq (&gpc->lock); + +cond_resched (); + +write_lock_irq (&gpc->lock); + } + gpc->refresh_in_progress = true; + old_pfn = gpc->pfn; old_khva = gpc->khva -offset_in_page (gpc->khva); old_uhva = gpc->uhva; - old_valid = gpc->valid;
Sean Christopherson
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
Quoted 10 lines · Lai Jiangshan
Review int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, · line 249
Quoted 2 lines · Sean Christopherson
(quoted diff) +0
-0@@ -159,10 +249,23 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,The following code of
refresh_in_progressis somewhat like mutex.+
mutex_lock(&gpc->refresh_in_progress); // beforewrite_lock_irq(&gpc->lock);Is it fit for the intention?
Yeah, I mutex should work. Not sure why I shied away from a mutex...
Paolo Bonzini
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
In reply to Sean Christophersonvirt/kvm/pfncache.c static void gpc_release_pfn_and_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva) · line 205
Quoted 13 lines · Sean Christopherson
virt/kvm/pfncache.c +13
-0+* Wait for mn_active_invalidate_count, not mmu_notifier_count, +* to go away, as the invalidation in the mmu_notifier event +* occurs_before_ mmu_notifier_count is elevated. +* +* Note, mn_active_invalidate_count can change at any time as +* it's not protected by gpc->lock. But, it is guaranteed to +* be elevated before the mmu_notifier acquires gpc->lock, and +* isn't dropped until after mmu_notifier_seq is updated. So, +* this task may get a false positive of sorts, i.e. see an +* elevated count and wait even though it's technically safe to +* proceed (becase the mmu_notifier will invalidate the cache +*_after_ it's refreshed here), but the cache will never be +* refreshed with stale data, i.e. won't get false negatives.
I am all for lavish comments, but I think this is even too detailed. What about:
/*
* mn_active_invalidate_count acts for all intents and purposes
* like mmu_notifier_count here; but we cannot use the latter
* because the invalidation in the mmu_notifier event occurs
* _before_ mmu_notifier_count is elevated.
*
* Note, it does not matter that mn_active_invalidate_count
* is not protected by gpc->lock. It is guaranteed to
* be elevated before the mmu_notifier acquires gpc->lock, and
* isn't dropped until after mmu_notifier_seq is updated. */
Paolo
Paolo Bonzini
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
Quoted 15 lines · Paolo Bonzini
Quoted 5 lines · Sean Christopherson
+ * Wait for
mn_active_invalidate_count, notmmu_notifier_count, + * to go away, as the invalidation in themmu_notifierevent + * occurs_before_mmu_notifier_countis elevated. + *
+ * Note,mn_active_invalidate_countcan change at any time asQuoted 8 lines
+ * it's not protected by gpc->lock. But, it is guaranteed to + * be elevated before the
mmu_notifieracquiresgpc->lock, and + * isn't dropped until aftermmu_notifier_seqis updated. So, + * this task may get a false positive of sorts, i.e. see an + * elevated count and wait even though it's technically safe to + * proceed (becase themmu_notifierwill invalidate the cache + *_after_ it's refreshed here), but the cache will never be + * refreshed with stale data, i.e. won't get false negatives.
Quoted 1 line
I am all for lavish comments, but I think this is even too detailed. What about:
And in fact this should be moved to a separate function.
virt/kvm/pfncache.c +32 -27
@@ -112,6 +112,36 @@ static void gpc_release_pfn_and_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva) } } + +static inline bool mmu_notifier_retry_cache (struct kvm *kvm,unsigned long mmu_seq) +{ +/* +* mn_active_invalidate_count acts for all intents and purposes +* like mmu_notifier_count here; but we cannot use the latter +* because the invalidation in the mmu_notifier event occurs +* _before_ mmu_notifier_count is elevated. +* +* Note, it does not matter that mn_active_invalidate_count +* is not protected by gpc->lock. It is guaranteed to +* be elevated before the mmu_notifier acquires gpc->lock, and +* isn't dropped until after mmu_notifier_seq is updated. +*/ +if (kvm->mn_active_invalidate_count) +return true; + +/* +* Ensure mn_active_invalidate_count is read before +* mmu_notifier_seq. This pairs with the smp_wmb() in +* mmu_notifier_invalidate_range_end() to guarantee either the +* old (non-zero) value of mn_active_invalidate_count or the +* new (incremented) value of mmu_notifier_seq is observed. +*/ +smp_rmb (); +if (kvm->mmu_notifier_seq != mmu_seq) +return true; +return false; +} +static kvm_pfn_t hva_to_pfn_retry (struct kvm *kvm,struct gfn_to_pfn_cache *gpc) {/* Note, the new page offset may be different than the old! */
@@ -129,7 +159,7 @@ static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)*/ gpc->valid = false; -for (;;) { + do { mmu_seq = kvm->mmu_notifier_seq; smp_rmb();
@@ -188,32 +218,7 @@ static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)* attempting to refresh. */ WARN_ON_ONCE (gpc->valid); - -/* -* mn_active_invalidate_count acts for all intents and purposes -* like mmu_notifier_count here; but we cannot use the latter -* because the invalidation in the mmu_notifier event occurs -* _before_ mmu_notifier_count is elevated. -* -* Note, it does not matter that mn_active_invalidate_count -* is not protected by gpc->lock. It is guaranteed to -* be elevated before the mmu_notifier acquires gpc->lock, and -* isn't dropped until after mmu_notifier_seq is updated. -*/ -if (kvm->mn_active_invalidate_count) -continue ; - -/* -* Ensure mn_active_invalidate_count is read before -* mmu_notifier_seq. This pairs with the smp_wmb() in -* mmu_notifier_invalidate_range_end() to guarantee either the -* old (non-zero) value of mn_active_invalidate_count or the -* new (incremented) value of mmu_notifier_seq is observed. -*/ -smp_rmb (); -if (kvm->mmu_notifier_seq == mmu_seq) -break ; - } + } while (mmu_notifier_retry_cache(kvm, mmu_seq); gpc->valid = true; gpc->pfn = new_pfn;
Sean Christopherson
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
Quoted 46 lines · Paolo Bonzini
Quoted 15 lines · Paolo Bonzini
Quoted 5 lines · Sean Christopherson
+ * Wait for
mn_active_invalidate_count, notmmu_notifier_count, + * to go away, as the invalidation in themmu_notifierevent + * occurs_before_mmu_notifier_countis elevated. + *
+ * Note,mn_active_invalidate_countcan change at any time asQuoted 8 lines
+ * it's not protected by gpc->lock. But, it is guaranteed to + * be elevated before the
mmu_notifieracquiresgpc->lock, and + * isn't dropped until aftermmu_notifier_seqis updated. So, + * this task may get a false positive of sorts, i.e. see an + * elevated count and wait even though it's technically safe to + * proceed (becase themmu_notifierwill invalidate the cache + *_after_ it's refreshed here), but the cache will never be + * refreshed with stale data, i.e. won't get false negatives.Quoted 2 lines
I am all for lavish comments, but I think this is even too detailed. What about:
And in fact this should be moved to a separate function.
virt/kvm/pfncache.c +18
-0@@ -112,6 +112,36 @@ static void gpc_release_pfn_and_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva) } } + +static inline bool mmu_notifier_retry_cache (struct kvm *kvm,unsigned long mmu_seq) +{ +/* +* mn_active_invalidate_count acts for all intents and purposes +* like mmu_notifier_count here; but we cannot use the latter +* because the invalidation in the mmu_notifier event occurs +* _before_ mmu_notifier_count is elevated. +* +* Note, it does not matter that mn_active_invalidate_count +* is not protected by gpc->lock. It is guaranteed to +* be elevated before the mmu_notifier acquires gpc->lock, and +* isn't dropped until after mmu_notifier_seq is updated. +*/ +if (kvm->mn_active_invalidate_count) +return true; + +/*
virt/kvm/pfncache.c static void gpc_release_pfn_and_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva) · line 141
Quoted 10 lines
virt/kvm/pfncache.c +10
-0+* Ensure mn_active_invalidate_count is read before +* mmu_notifier_seq. This pairs with the smp_wmb() in +* mmu_notifier_invalidate_range_end() to guarantee either the +* old (non-zero) value of mn_active_invalidate_count or the +* new (incremented) value of mmu_notifier_seq is observed. +*/ +smp_rmb (); +if (kvm->mmu_notifier_seq != mmu_seq) +return true; +return false;
This can be
return kvm->mmu_notifier_seq != mmu_seq;
Looks good otherwise. It'll probably yield a smaller diff too.
Sean Christopherson
maintainerRe: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
In reply to Paolo BonziniQuoted 16 lines · Paolo Bonzini
virt/kvm/pfncache.c static void gpc_release_pfn_and_khva(struct kvm *kvm, kvm_pfn_t pfn, void *khva) · line 205
Quoted 13 lines · Sean Christopherson
virt/kvm/pfncache.c +13
-0+* Wait for mn_active_invalidate_count, not mmu_notifier_count, +* to go away, as the invalidation in the mmu_notifier event +* occurs_before_ mmu_notifier_count is elevated. +* +* Note, mn_active_invalidate_count can change at any time as +* it's not protected by gpc->lock. But, it is guaranteed to +* be elevated before the mmu_notifier acquires gpc->lock, and +* isn't dropped until after mmu_notifier_seq is updated. So, +* this task may get a false positive of sorts, i.e. see an +* elevated count and wait even though it's technically safe to +* proceed (becase the mmu_notifier will invalidate the cache +*_after_ it's refreshed here), but the cache will never be +* refreshed with stale data, i.e. won't get false negatives. I am all for lavish comments, but I think this is even too detailed.
Yeah, the false positive/negative stuff is probably overkill.
Quoted 7 lines
What about:
/*
*mn_active_invalidate_countacts for all intents and purposes
* likemmu_notifier_counthere; but we cannot use the latter
* because the invalidation in themmu_notifierevent occurs
* _before_mmu_notifier_countis elevated.
Looks good, though I'd prefer to avoid the "we", and explicitly call out that its the invalidation of the caches.
/*
* mn_active_invalidate_count acts for all intents and purposes
* like mmu_notifier_count here; but the latter cannot be used
* here because the invalidation of caches in the mmu_notifier
* event occurs _before_ mmu_notifier_count is elevated. *
* Note, it does not matter that mn_active_invalidate_count
* is not protected by gpc->lock. It is guaranteed to
* be elevated before the mmu_notifier acquires gpc->lock, and
* isn't dropped until after mmu_notifier_seq is updated. */
Also, you'll definitely want to look at v3 of this series. I'm 99% certain I didn't change the comment though :-)
Re: [PATCH v2 6/8] KVM: Fix multiple races in gfn=>pfn cache refresh
In reply to Sean ChristophersonQuoted 67 lines · Sean Christopherson
Rework the gfn=>pfn cache (gpc) refresh logic to address multiple races between the cache itself, and between the cache and
mmu_notifierevents.The existing refresh code attempts to guard against races with the
mmu_notifierby speculatively marking the cache valid, and then marking it invalid if ammu_notifierinvalidation occurs. That handles the case where an invalidation occurs between dropping and re-acquiringgpc->lock, but it doesn't handle the scenario where the cache is refreshed after the cache was invalidated by the notifier, but before the notifier elevates mmu_notifier_count. The gpc refresh can't use the "retry" helper as its invalidation occurs _before_mmu_notifier_countis elevated and beforemmu_notifier_range_startis set/updated.CPU0 CPU1
---- ----
gfn_to_pfn_cache_invalidate_start()
|
->gpc->valid= false;
kvm_gfn_to_pfn_cache_refresh()|
|->gpc->valid= true;
hva_to_pfn_retry()
|
-> acquirekvm->mmu_lockkvm->mmu_notifier_count== 0mmu_seq==kvm->mmu_notifier_seqdropkvm->mmu_lockreturn pfn 'X' acquirekvm->mmu_lock
kvm_inc_notifier_count()
dropkvm->mmu_lock()
kernel frees pfn 'X'
kvm_gfn_to_pfn_cache_check()|
|->gpc->valid== truecaller accesses freed pfn 'X'
Key off of
mn_active_invalidate_countto detect that a pfncache refresh needs to wait for an in-progressmmu_notifierinvalidation. Whilemn_active_invalidate_countis not guaranteed to be stable, it is guaranteed to be elevated prior to an invalidation acquiringgpc->lock, so either the refresh will see an active invalidation and wait, or the invalidation will run after the refresh completes.Speculatively marking the cache valid is itself flawed, as a concurrent
kvm_gfn_to_pfn_cache_check()would see a valid cache with stale pfn/khva values. The KVM Xen use case explicitly allows/wants multiple users; even though the caches are allocated per vCPU,__kvm_xen_has_interrupt()can read a different vCPU (or vCPUs). Address this race by invalidating the cache prior to droppinggpc->lock(this is made possible by fixing the abovemmu_notifierrace).Finally, the refresh logic doesn't protect against concurrent refreshes with different GPAs (which may or may not be a desired use case, but its allowed in the code), nor does it protect against a false negative on the memslot generation. If the first refresh sees a stale memslot generation, it will refresh the hva and generation before moving on to the hva=>pfn translation. If it then drops
gpc->lock, a different user can come along, acquiregpc->lock, see that the memslot generation is fresh, and skip the hva=>pfn update due to the userspace address also matching (because it too was updated). Address this race by adding an "in-progress" flag so that the refresh that acquiresgpc->lockfirst runs to completion before other users can start their refresh.
Quoted 6 lines
Complicating all of this is the fact that both the hva=>pfn resolution and mapping of the kernel address can sleep, i.e. must be done outside of
gpc->lockFix the above races in one fell swoop, trying to fix each individual race in a sane manner is impossible, for all intents and purposes.
Hm, the problem with this (commit 58cd407ca4c6278) is that it ends up acting like a *really* unfair read/write lock.
If there are a lot of 'writers' invalidating other HVA ranges, then the hva_to_pfn_retry() function as the 'reader' will back off for ever and never make progress. Even while a single other invalidation is active (->mn_active_invalidate_count is non-zero), hva_to_pfn_retry() will just spin in an elaborate busy-wait loop, mapping same page over and over again.
(In the repro case I'm not entirely sure *why* there are userspace threads bashing on MADV_DONTNEED and causing this, but 'Don't Do That Then' is only a partial answer. The kernel's own locking should allow it to make progress regardless.)
If we call the GPC invalidate function from invalidate_range_end() instead of _start, can't we do it concurrently without having to check for active invalidations or mmu_invalidate_seq?
We can then introduce a 'validating' flag, set before the attempt to hva_to_pfn() in the loop, so that the translation can be shot down before it's even made. And use *that* as the trigger for the retry loop so that only has to retry if its *own* uHVA is messed with.
Patch follows...
PATCH 1/1 KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
From: David Woodhouse <[email protected]>
The existing retry loop in hva_to_pfn_retry() is extremely pessimistic. If there is an invalidation running concurrently, it is effectively just a complex busy wait loop because its local mmu_notifier_retry_cache() function will always return true.
It ends up functioning as a very unfair read/write lock. If userspace is acting as a 'writer', performing many unrelated MM changes, then the hva_to_pfn_retry() function acting as the 'reader' just backs off and keep retrying for ever, not making any progress.
Solve this by introducing a separate 'validating' flag to the GPC, so that it can be marked invalid before it's even mapped. This allows the invalidation to be moved to the range_end hook, and the retry loop in hva_to_pfn_retry() can be changed to loop only if its particular uHVA has been affected.
---
I note I'm deleting a big comment in kvm_main.c about doing the invalidation before acquiring mmu_lock. But we don't hold the lock in the range_end callback either, do we?
include/linux/kvm_types.h | 1 + virt/kvm/kvm_main.c | 14 ++------ virt/kvm/kvm_mm.h | 12 +++---- virt/kvm/pfncache.c | 75 +++++++++++++++++++-------------------- 4 files changed, 45 insertions(+), 57 deletions(-)
include/linux/kvm_types.h +1 -0
@@ -69,6 +69,7 @@ struct gfn_to_pfn_cache {void *khva;kvm_pfn_t pfn;bool active; +bool validating;bool valid; };
virt/kvm/kvm_main.c +2 -12
@@ -777,18 +777,6 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, kvm->mn_active_invalidate_count++;spin_unlock (&kvm->mn_invalidate_lock); -/* -* Invalidate pfn caches _before_ invalidating the secondary MMUs, i.e. -* before acquiring mmu_lock, to avoid holding mmu_lock while acquiring -* each cache's lock. There are relatively few caches in existence at -* any given time, and the caches themselves can check for hva overlap, -* i.e. don't need to rely on memslot overlap checks for performance. -* Because this runs without holding mmu_lock, the pfn caches must use -* mn_active_invalidate_count (see above) instead of -* mmu_invalidate_in_progress. -*/ -gfn_to_pfn_cache_invalidate_start (kvm, range->start, range->end); -/* * If one or more memslots were found and thus zapped, notify arch code * that guest memory has been reclaimed. This needs to be done *after*
@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count;spin_unlock (&kvm->mn_invalidate_lock); +gfn_to_pfn_cache_invalidate (kvm, range->start, range->end); +/* * There can only be one waiter, since the wait happens under * slots_lock.
virt/kvm/kvm_mm.h +6 -6
@@ -24,13 +24,13 @@ kvm_pfn_t hva_to_pfn(unsigned long addr, bool atomic, bool interruptible, bool *async, bool write_fault, bool *writable);#ifdef CONFIG_HAVE_KVM_PFNCACHE -void gfn_to_pfn_cache_invalidate_start (struct kvm *kvm, -unsigned long start, -unsigned long end); +void gfn_to_pfn_cache_invalidate (struct kvm *kvm, +unsigned long start, +unsigned long end);#else -static inline void gfn_to_pfn_cache_invalidate_start (struct kvm *kvm, -unsigned long start, -unsigned long end) +static inline void gfn_to_pfn_cache_invalidate (struct kvm *kvm, +unsigned long start, +unsigned long end) { }#endif /* HAVE_KVM_PFNCACHE */
virt/kvm/pfncache.c +36 -39
143 lines not on this page · raw message
Signature · 1 line
2.44.0
Sean Christopherson
maintainerRe: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 12 lines · David Woodhouse
From: David Woodhouse <[email protected]>
The existing retry loop in
hva_to_pfn_retry()is extremely pessimistic. If there is an invalidation running concurrently, it is effectively just a complex busy wait loop because its localmmu_notifier_retry_cache()function will always return true.It ends up functioning as a very unfair read/write lock. If userspace is acting as a 'writer', performing many unrelated MM changes, then the
hva_to_pfn_retry()function acting as the 'reader' just backs off and keep retrying for ever, not making any progress.
Quoted 5 lines
Solve this by introducing a separate 'validating' flag to the GPC, so that it can be marked invalid before it's even mapped. This allows the invalidation to be moved to the
range_endhook, and the retry loop inhva_to_pfn_retry()can be changed to loop only if its particular uHVA has been affected.
I think I'm missing something. How does allowing hva_to_pfn_retry() allow KVM as a whole to make forward progress? Doesn't getting past hva_to_pfn_retry() just move the problem to kvm_gpc_check()?
kvm_gpc_refresh() can't be called with gpc->lock held, and nor does it return with gpc->lock held, so a racing mmu_notifier invalidation can/will acquire gpc->lock and clear gpc->active, no?
Oh, by "unrelated", you mean _completely_ unrelated? As in, KVM happens to do a refresh when userspace is blasting MADV_DONTNEED, and gets stuck retrying for no good reason?
Servicing guest pages faults has the same problem, which is why mmu_invalidate_retry_gfn() was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).
Woefully incomplete, but I think this is the gist of what you want:
virt/kvm/pfncache.c +28 -3
@@ -28,6 +28,26 @@ void gfn_to_pfn_cache_invalidate_start(struct kvm *kvm, unsigned long start,struct gfn_to_pfn_cache *gpc;spin_lock (&kvm->gpc_lock); + + if (likely(kvm_is_error_hva(kvm->mmu_gpc_invalidate_range_start)) { + kvm->mmu_gpc_invalidate_range_start = start; + kvm->mmu_gpc_invalidate_range_end = end; + }else { +/* +* Fully tracking multiple concurrent ranges has diminishing +* returns. Keep things simple and just find the minimal range +* which includes the current and new ranges. As there won't be +* enough information to subtract a range after its invalidate +* completes, any ranges invalidated concurrently will +* accumulate and persist until all outstanding invalidates +* complete. +*/ + kvm->mmu_gpc_invalidate_range_start = +min (kvm->mmu_gpc_invalidate_range_start, start); + kvm->mmu_gpc_invalidate_range_end = +max (kvm->mmu_gpc_invalidate_range_end, end); + } +list_for_each_entry (gpc, &kvm->gpc_list, list) {read_lock_irq (&gpc->lock);
@@ -124,8 +144,11 @@ static void gpc_unmap(kvm_pfn_t pfn, void *khva)#endif } -static inline bool mmu_notifier_retry_cache (struct kvm *kvm,unsigned long mmu_seq) +static inline bool mmu_notifier_retry_cache (struct gfn_to_pfn_cache *gpc, +unsigned long mmu_seq) { +struct kvm *kvm = gpc->kvm; +/* * mn_active_invalidate_count acts for all intents and purposes * like mmu_invalidate_in_progress here; but the latter cannot
@@ -138,7 +161,9 @@ static inline bool mmu_notifier_retry_cache(struct kvm *kvm, unsigned long mmu_s* be elevated before the mmu_notifier acquires gpc->lock, and * isn't dropped until after mmu_invalidate_seq is updated. */ -if (kvm->mn_active_invalidate_count) +if (kvm->mn_active_invalidate_count && + gpc->uhva >= kvm->mmu_gpc_invalidate_range_start && + gpc->uhva < kvm->mmu_gpc_invalidate_range_end)return true;/*
@@ -224,7 +249,7 @@ static kvm_pfn_t hva_to_pfn_retry(struct gfn_to_pfn_cache *gpc)* attempting to refresh. */ WARN_ON_ONCE (gpc->valid); - }while (mmu_notifier_retry_cache (gpc->kvm, mmu_seq)); + }while (mmu_notifier_retry_cache (gpc, mmu_seq)); gpc->valid = true; gpc->pfn = new_pfn;
Quoted 5 lines
Signed-off-by David Woodhouse <[email protected]>---
I note I'm deleting a big comment inkvm_main.cabout doing the invalidation before acquiringmmu_lock. But we don't hold the lock in therange_endcallback either, do we?
Correct, __kvm_handle_hva_range() acquires and releases mmu_lock. However, the intent of the comment was to clarify why GPCs are invalidated in kvm_mmu_notifier_invalidate_range_start(), as opposed to kvm_mmu_invalidate_begin() which _is_ called under mmu_lock and is also called if and only if KVM has a relevant memslot. E.g. that's why the comment also talks about memslot overlap checks.
Quoted 12 lines
include/linux/kvm_types.h | 1 + virt/kvm/kvm_main.c | 14 ++------ virt/kvm/kvm_mm.h | 12 +++---- virt/kvm/pfncache.c | 75 +++++++++++++++++++-------------------- 4 files changed, 45 insertions(+), 57 deletions(-)include/linux/kvm_types.h +0
-0
include/linux/kvm_types.h struct gfn_to_pfn_cache { · line 72
Quoted 4 lines
include/linux/kvm_types.h +1
-0@@ -69,6 +69,7 @@ struct gfn_to_pfn_cache {void *khva;kvm_pfn_t pfn;bool active; +bool validating;
This is a confusing name, partly because KVM usually deals with invalidation events, but also because it's sticky and stays set long after the act of validating the GPC is complete.
Something like "needs_invalidation" is the best I can come up with, but I believe this bikeshed is moot (see above and below).
Quoted 27 lines
include/linux/kvm_types.h +0
-0bool valid; };virt/kvm/kvm_main.c +0
-12@@ -777,18 +777,6 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn, kvm->mn_active_invalidate_count++;spin_unlock (&kvm->mn_invalidate_lock); -/* -* Invalidate pfn caches _before_ invalidating the secondary MMUs, i.e. -* before acquiring mmu_lock, to avoid holding mmu_lock while acquiring -* each cache's lock. There are relatively few caches in existence at -* any given time, and the caches themselves can check for hva overlap, -* i.e. don't need to rely on memslot overlap checks for performance. -* Because this runs without holding mmu_lock, the pfn caches must use -* mn_active_invalidate_count (see above) instead of -* mmu_invalidate_in_progress. -*/ -gfn_to_pfn_cache_invalidate_start (kvm, range->start, range->end); -/* * If one or more memslots were found and thus zapped, notify arch code * that guest memory has been reclaimed. This needs to be done *after*
virt/kvm/kvm_main.c static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, · line 840
Quoted 4 lines
virt/kvm/kvm_main.c +1
-0@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count;spin_unlock (&kvm->mn_invalidate_lock); +gfn_to_pfn_cache_invalidate (kvm, range->start, range->end);
We can't do this. The contract with mmu_notifiers is that secondary MMUs must unmap the hva before returning from invalidate_range_start(), and must not create new mappings until invalidate_range_end().
Re: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 23 lines · Sean Christopherson
Quoted 12 lines · David Woodhouse
From: David Woodhouse <[email protected]>
The existing retry loop in
hva_to_pfn_retry()is extremely pessimistic. If there is an invalidation running concurrently, it is effectively just a complex busy wait loop because its localmmu_notifier_retry_cache()function will always return true.It ends up functioning as a very unfair read/write lock. If userspace is acting as a 'writer', performing many unrelated MM changes, then the
hva_to_pfn_retry()function acting as the 'reader' just backs off and keep retrying for ever, not making any progress.Quoted 5 lines
Solve this by introducing a separate 'validating' flag to the GPC, so that it can be marked invalid before it's even mapped. This allows the invalidation to be moved to the
range_endhook, and the retry loop inhva_to_pfn_retry()can be changed to loop only if its particular uHVA has been affected.I think I'm missing something. How does allowing
hva_to_pfn_retry()allow KVM as a whole to make forward progress? Doesn't getting pasthva_to_pfn_retry()just move the problem tokvm_gpc_check()?
Quoted 7 lines
kvm_gpc_refresh()can't be called withgpc->lockheld, and nor does it return withgpc->lockheld, so a racingmmu_notifierinvalidation can/will acquiregpc->lockand cleargpc->active, no?Oh, by "unrelated", you mean _completely_ unrelated? As in, KVM happens to do a refresh when userspace is blasting
MADV_DONTNEED, and gets stuck retrying for no good reason?
Right. And again, I have no idea why userspace is doing that, and we need to make it stop. But that's not really the point; the kernel should be able to make progress anyway.
Quoted 5 lines
Servicing guest pages faults has the same problem, which is why
mmu_invalidate_retry_gfn()was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).Woefully incomplete, but I think this is the gist of what you want:
Hm, maybe. It does mean that migration occurring all through memory (indeed, just one at top and bottom of guest memory space) would perturb GPCs which remain present.
Quoted 9 lines
Quoted 5 lines
(quoted diff) +0
-0@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count; spin_unlock(&kvm->mn_invalidate_lock);
+ gfn_to_pfn_cache_invalidate(kvm,range->start,range->end);We can't do this. The contract with
mmu_notifiersis that secondary MMUs must unmap the hva before returning frominvalidate_range_start(), and must not create new mappings untilinvalidate_range_end().
But in the context of the GPC, it is only "mapped" when the ->valid bit is set.
Even the invalidation callback just clears the valid bit, and that means nobody is allowed to dereference the ->khva any more. It doesn't matter that the underlying (stale) PFN is still kmapped.
Can we not apply the same logic to the hva_to_pfn_retry() loop? Yes, it might kmap a page that gets removed, but it's not actually created a new mapping if it hasn't set the ->valid bit.
I don't think this version quite meets the constraints, and I might need to hook *both* the start and end notifiers, and might not like it once I get there. But I'll have a go...
Sean Christopherson
maintainerRe: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 7 lines · David Woodhouse
Quoted 6 lines · Sean Christopherson
Quoted 1 line · David Woodhouse
From: David Woodhouse <[email protected]>
Servicing guest pages faults has the same problem, which is why
mmu_invalidate_retry_gfn()was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).
Quoted 5 lines
Quoted 1 line
Woefully incomplete, but I think this is the gist of what you want:
Hm, maybe. It does mean that migration occurring all through memory (indeed, just one at top and bottom of guest memory space) would perturb GPCs which remain present.
If that happens with a real world VMM, and it's not a blatant VMM goof, then we can fix KVM. The stage-2 page fault path hammers the mmu_notifier retry logic far more than GPCs, so if a range-based check is inadequate for some use case, then we definitely need to fix both.
In short, I don't see any reason to invent something different for GPCs.
Quoted 16 lines
Quoted 9 lines
Quoted 5 lines
(quoted diff) +0
-0@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count; spin_unlock(&kvm->mn_invalidate_lock);
+ gfn_to_pfn_cache_invalidate(kvm,range->start,range->end);We can't do this. The contract with
mmu_notifiersis that secondary MMUs must unmap the hva before returning frominvalidate_range_start(), and must not create new mappings untilinvalidate_range_end().But in the context of the GPC, it is only "mapped" when the ->valid bit is set.
Even the invalidation callback just clears the valid bit, and that means nobody is allowed to dereference the ->khva any more. It doesn't matter that the underlying (stale) PFN is still kmapped.
Quoted 7 lines
Can we not apply the same logic to the
hva_to_pfn_retry()loop? Yes, it might kmap a page that gets removed, but it's not actually created a new mapping if it hasn't set the ->valid bit.I don't think this version quite meets the constraints, and I might need to hook *both* the start and end notifiers, and might not like it once I get there. But I'll have a go...
I'm pretty sure you're going to need the range-based retry logic. KVM can't safely set gpc->valid until mn_active_invalidate_count reaches zero, so if a GPC refresh comes along after mn_active_invalidate_count has been elevated, it won't be able to set gpc->valid until the MADV_DONTNEED storm goes away. Without range-based tracking, there's no way to know if a previous invalidation was relevant to the GPC.
Re: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 14 lines · Sean Christopherson
Quoted 7 lines · David Woodhouse
Quoted 6 lines · Sean Christopherson
Quoted 1 line · David Woodhouse
From: David Woodhouse <[email protected]>
Servicing guest pages faults has the same problem, which is why
mmu_invalidate_retry_gfn()was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).Quoted 5 lines
Quoted 1 line
Woefully incomplete, but I think this is the gist of what you want:
Hm, maybe. It does mean that migration occurring all through memory (indeed, just one at top and bottom of guest memory space) would perturb GPCs which remain present.
Quoted 16 lines
If that happens with a real world VMM, and it's not a blatant VMM goof, then we can fix KVM. The stage-2 page fault path hammers the
mmu_notifierretry logic far more than GPCs, so if a range-based check is inadequate for some use case, then we definitely need to fix both.In short, I don't see any reason to invent something different for GPCs.
Quoted 9 lines
Quoted 9 lines
Quoted 5 lines
(quoted diff) +0
-0@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count; spin_unlock(&kvm->mn_invalidate_lock);
+ gfn_to_pfn_cache_invalidate(kvm,range->start,range->end);We can't do this. The contract with
mmu_notifiersis that secondary MMUs must unmap the hva before returning frominvalidate_range_start(), and must not create new mappings untilinvalidate_range_end().
Looking at that assertion harder... where is that rule written? It seems counter-intuitive to me; that isn't how TLBs work. Another CPU can populate a TLB entry right up to the moment the PTE is actually *changed* in the page tables, and then the CPU which is modifying/zapping the PTE needs to perform a remote TLB flush. That remote TLB flush is analogous to the invalidate_range_end() call, surely?
I'm fairly sure that's how it works for PASID support too; nothing prevents the IOMMU+device from populating an IOTLB entry until the PTE is actually changed in the process page tables.
So why can't we do the same for the GPC?
Quoted 14 lines
Quoted 6 lines
But in the context of the GPC, it is only "mapped" when the ->valid bit is set.
Even the invalidation callback just clears the valid bit, and that means nobody is allowed to dereference the ->khva any more. It doesn't matter that the underlying (stale) PFN is still kmapped.
Quoted 7 lines
Can we not apply the same logic to the
hva_to_pfn_retry()loop? Yes, it might kmap a page that gets removed, but it's not actually created a new mapping if it hasn't set the ->valid bit.I don't think this version quite meets the constraints, and I might need to hook *both* the start and end notifiers, and might not like it once I get there. But I'll have a go...
Quoted 6 lines
I'm pretty sure you're going to need the range-based retry logic. KVM can't safely set
gpc->validuntilmn_active_invalidate_countreaches zero, so if a GPC refresh comes along aftermn_active_invalidate_counthas been elevated, it won't be able to setgpc->validuntil theMADV_DONTNEEDstorm goes away. Without range-based tracking, there's no way to know if a previous invalidation was relevant to the GPC.
If it is indeed the case that KVM can't just behave like a normal TLB, so it and can't set gpc->valid until mn_active_invalidate_count reaches zero, it still only needs to *wait* (or spin, maybe). It certainly doesn't need to keep looping and remapping the same PFN over and over again, as it does at the moment.
When mn_active_invalidate_count does reach zero, either the young GPC will have been invalidated by clearing the (to be renamed) ->validating flag, or it won't have been. If it *has* been invalidated, that's when hva_to_pfn_retry() needs to go one more time round its full loop.
So it just needs to wait until any pending (relevant) invalidations have completed, *then* check and potentially loop once more.
And yes, making that *wait* range-based does make some sense, I suppose. It becomes "wait for gpc->uhva not to be within the range of kvm->mmu_gpc_invalidate_range_{start,end}."
Except... that range can never shrink *except* when mn_active_invalidate_count becomes zero, can it? So if we do end up waiting, the wake condition is *still* just that the count has become zero. There's already a wakeup in that case, on kvm-
Quoted 1 line
mn_memslots_update_rcuwait. Can I wait on that?
Sean Christopherson
maintainerRe: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 32 lines · David Woodhouse
Quoted 14 lines · Sean Christopherson
Quoted 7 lines · David Woodhouse
Quoted 6 lines · Sean Christopherson
Quoted 1 line · David Woodhouse
From: David Woodhouse <[email protected]>
Servicing guest pages faults has the same problem, which is why
mmu_invalidate_retry_gfn()was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).Quoted 5 lines
Quoted 1 line
Woefully incomplete, but I think this is the gist of what you want:
Hm, maybe. It does mean that migration occurring all through memory (indeed, just one at top and bottom of guest memory space) would perturb GPCs which remain present.
Quoted 16 lines
If that happens with a real world VMM, and it's not a blatant VMM goof, then we can fix KVM. The stage-2 page fault path hammers the
mmu_notifierretry logic far more than GPCs, so if a range-based check is inadequate for some use case, then we definitely need to fix both.In short, I don't see any reason to invent something different for GPCs.
Quoted 9 lines
Quoted 9 lines
Quoted 5 lines
(quoted diff) +0
-0@@ -849,6 +837,8 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn, wake = !kvm->mn_active_invalidate_count; spin_unlock(&kvm->mn_invalidate_lock);
+ gfn_to_pfn_cache_invalidate(kvm,range->start,range->end);We can't do this. The contract with
mmu_notifiersis that secondary MMUs must unmap the hva before returning frominvalidate_range_start(), and must not create new mappings untilinvalidate_range_end().
Quoted 1 line
Looking at that assertion harder... where is that rule written?
The big comment for invalidate_range_{start,end}() in include/linux/mmu_notifier.h. The relevant snippets are:
* If the subsystem can't guarantee that no additional references are
* taken to the pages in the range, it has to implement the
* invalidate_range() notifier to remove any references taken after
* invalidate_range_start().
* invalidate_range_start() is called when all pages in the
* range are still mapped and have at least a refcount of one. *
* invalidate_range_end() is called when all pages in the
* range have been unmapped and the pages have been freed by
* the VM.
The last one is key: the pages have already been freed when invalidate_range_end() is called, and so unmapping at that time would be too late.
Quoted 5 lines
It seems counter-intuitive to me; that isn't how TLBs work. Another CPU can populate a TLB entry right up to the moment the PTE is actually *changed* in the page tables, and then the CPU which is modifying/zapping the PTE needs to perform a remote TLB flush. That remote TLB flush is analogous to the
invalidate_range_end()call, surely?
KVM's usage isn't about (hardware) TLBs. Ah, and the history is even somewhat evident in the above comment I referenced. invalidate_range() no longer exists, it was morphed into arch_invalidate_secondary_tlbs(). For secondary MMUs that reuse the primary MMU's PTEs, mmu_notifier_arch_invalidate_secondary_tlbs() is indeed called after the PTEs have been modified.
KVM's usage is different. Because KVM has its own (Secondary) PTEs (commit 1af5a8109904 ("mmu_notifiers: rename invalidate_range notifier") calls them "software TLBs", but I find that to be a confusing description), zapping on-demand when the primary PTEs are modified is tricky and ultimately undesirable.
E.g. invoking mmu_notifiers while holding a PTE lock would prevent KVM from blocking, which can be problematic if KVM needs to zap a large number SPTEs.
And doing invalidation on-demand for each primary PTE would be suboptimal for cases where a large VMA range is unmapped/modified, e.g. KVM would get a large number of invalidation events instead of one big, all-encompassing invalidation.
The obvious downside is what you've run into, where the start+end approach forces KVM to wait for all in-flight invalidations to go away. But again, in practice the rudimentary range tracking suffices for all known use cases.
Quoted 38 lines
I'm fairly sure that's how it works for PASID support too; nothing prevents the IOMMU+device from populating an IOTLB entry until the PTE is actually changed in the process page tables.
So why can't we do the same for the GPC?
Quoted 14 lines
Quoted 6 lines
But in the context of the GPC, it is only "mapped" when the ->valid bit is set.
Even the invalidation callback just clears the valid bit, and that means nobody is allowed to dereference the ->khva any more. It doesn't matter that the underlying (stale) PFN is still kmapped.
Quoted 7 lines
Can we not apply the same logic to the
hva_to_pfn_retry()loop? Yes, it might kmap a page that gets removed, but it's not actually created a new mapping if it hasn't set the ->valid bit.I don't think this version quite meets the constraints, and I might need to hook *both* the start and end notifiers, and might not like it once I get there. But I'll have a go...
Quoted 6 lines
I'm pretty sure you're going to need the range-based retry logic. KVM can't safely set
gpc->validuntilmn_active_invalidate_countreaches zero, so if a GPC refresh comes along aftermn_active_invalidate_counthas been elevated, it won't be able to setgpc->validuntil theMADV_DONTNEEDstorm goes away. Without range-based tracking, there's no way to know if a previous invalidation was relevant to the GPC.If it is indeed the case that KVM can't just behave like a normal TLB, so it and can't set
gpc->validuntilmn_active_invalidate_countreaches zero, it still only needs to *wait* (or spin, maybe). It certainly doesn't need to keep looping and remapping the same PFN over and over again, as it does at the moment.When
mn_active_invalidate_countdoes reach zero, either the young GPC will have been invalidated by clearing the (to be renamed) ->validating flag, or it won't have been. If it *has* been invalidated, that's whenhva_to_pfn_retry()needs to go one more time round its full loop.
Quoted 6 lines
So it just needs to wait until any pending (relevant) invalidations have completed, *then* check and potentially loop once more.
And yes, making that *wait* range-based does make some sense, I suppose. It becomes "wait for
gpc->uhvanot to be within the range of kvm->mmu_gpc_invalidate_range_{start,end}."
Yep, exactly. Without range-based tracking, there's no way for KVM to know when a relevant in-flight invalidation has completed.
Quoted 2 lines
Except... that range can never shrink *except* when
mn_active_invalidate_countbecomes zero, can it?
Not without more sophisticated logic, no. E.g. if KVM supported tracking multiple distinct ranges, then individual invalidation ranges could be dropped. But to to avoid memory allocations in invalidate_range_start(), KVM would still need to hardcode the maximum number of in-flight ranges. E.g. even if KVM used a dynamic container, we'd probably want the container entries to be "allocated" out of a cache, and that cache would need a maximum capacity.
With a max limit on the number of ranges, KVM would still be forced to combine ranges if there are too many in-flight invalidations.
So, because tracking a single range has sufficed for all known use cases, and it's significantly simpler than tracking multiple ranges, AFAIK no one has pursued a multi-range tracking implementation.
Quoted 3 lines
So if we do end up waiting, the wake condition is *still* just that the count has become zero. There's already a wakeup in that case, on kvm-
Quoted 1 line
mn_memslots_update_rcuwait. Can I wait on that?
I suspect you're trying to solve a problem that doesn't exist in practice. hva_to_pfn_retry() already has a cond_resched(), so getting stuck for a long duration isn't fatal, just suboptimal. And similar to the range-based tracking, _if_ there's a problem in practice, then it also affects guest page faults. KVM simply resumes the vCPU and keeps re-faulting until the in-flight invalidation(s) has gone away.
Not without reworking mn_memslots_update_rcuwait. KVM assumes there is at most one waiter, as that wait+wake combination is specifically to handle the case where a _relevant_ in-flight mmu_notifier invalidation needs to block a userspace memslot deletion. KVM takes mmu_lock in invalidate_range_{start,end}() if and only if there is an overlapping memslot, and so KVM needs to prevent a memslot from being deleted between start() and end().
Re: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
Quoted 2 lines · Sean Christopherson
The last one is key: the pages have already been freed when
invalidate_range_end()is called, and so unmapping at that time would be too late.
OK, thanks.
Quoted 3 lines
The obvious downside is what you've run into, where the start+end approach forces KVM to wait for all in-flight invalidations to go away. But again, in practice the rudimentary range tracking suffices for all known use cases.
Makes sense.
Quoted 6 lines
I suspect you're trying to solve a problem that doesn't exist in practice.
hva_to_pfn_retry()already has acond_resched(), so getting stuck for a long duration isn't fatal, just suboptimal. And similar to the range-based tracking, _if_ there's a problem in practice, then it also affects guest page faults. KVM simply resumes the vCPU and keeps re-faulting until the in-flight invalidation(s) has gone away.
Yeah. When it's looping on actual page faults it's easy to pretend it isn't a busy-loop :)
Making it wait is fairly simple; it would be easy to just make it cond_resched() instead though. Here's what I have so far (incremental).
https://git.infradead.org/?p=users/dwmw2/linux.git;a=commitdiff;h=pfncache-2
I need to revive the torture test you had at the end of your earlier series.
include/linux/kvm_host.h +3 -0
@@ -770,6 +770,9 @@ struct kvm {/* For management / invalidation of gfn_to_pfn_caches */ spinlock_t gpc_lock;struct list_head gpc_list; +u64 mmu_gpc_invalidate_range_start; +u64 mmu_gpc_invalidate_range_end; +wait_queue_head_t gpc_invalidate_wq;/* * created_vcpus is protected by kvm->lock, and is incremented
virt/kvm/kvm_main.c +30 -4
@@ -775,6 +775,24 @@ static int kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,*/ spin_lock (&kvm->mn_invalidate_lock); kvm->mn_active_invalidate_count++; + if (likely(kvm->mn_active_invalidate_count == 1) { + kvm->mmu_gpc_invalidate_range_start = range->start; + kvm->mmu_gpc_invalidate_range_end = range->end; + }else { +/* +* Fully tracking multiple concurrent ranges has diminishing +* returns. Keep things simple and just find the minimal range +* which includes the current and new ranges. As there won't be +* enough information to subtract a range after its invalidate +* completes, any ranges invalidated concurrently will +* accumulate and persist until all outstanding invalidates +* complete. +*/ + kvm->mmu_gpc_invalidate_range_start = +min (kvm->mmu_gpc_invalidate_range_start, range->start); + kvm->mmu_gpc_invalidate_range_end = +max (kvm->mmu_gpc_invalidate_range_end, range->end); + }spin_unlock (&kvm->mn_invalidate_lock);/*
@@ -830,21 +848,27 @@ static void kvm_mmu_notifier_invalidate_range_end(struct mmu_notifier *mn,__kvm_handle_hva_range (kvm, &hva_range); +gfn_to_pfn_cache_invalidate (kvm, range->start, range->end); +/* Pairs with the increment in range_start(). */ spin_lock (&kvm->mn_invalidate_lock);if (!WARN_ON_ONCE (!kvm->mn_active_invalidate_count)) --kvm->mn_active_invalidate_count; wake = !kvm->mn_active_invalidate_count; +if (wake) { + kvm->mmu_gpc_invalidate_range_start =KVM_HVA_ERR_BAD ; + kvm->mmu_gpc_invalidate_range_end =KVM_HVA_ERR_BAD ; + }spin_unlock (&kvm->mn_invalidate_lock); -gfn_to_pfn_cache_invalidate (kvm, range->start, range->end); -/* * There can only be one waiter, since the wait happens under * slots_lock. */ -if (wake) +if (wake) { +wake_up (&kvm->gpc_invalidate_wq);rcuwait_wake_up (&kvm->mn_memslots_update_rcuwait); + } } static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
@@ -1154,7 +1178,9 @@ static struct kvm *kvm_create_vm(unsigned long type, const char *fdname)INIT_LIST_HEAD (&kvm->gpc_list);spin_lock_init (&kvm->gpc_lock); - +init_waitqueue_head (&kvm->gpc_invalidate_wq); + kvm->mmu_gpc_invalidate_range_start =KVM_HVA_ERR_BAD ; + kvm->mmu_gpc_invalidate_range_end =KVM_HVA_ERR_BAD ;INIT_LIST_HEAD (&kvm->devices); kvm->max_vcpus =KVM_MAX_VCPUS ;
virt/kvm/pfncache.c +42 -0
54 lines not on this page · raw message
Re: [PATCH] KVM: Move gfn_to_pfn_cache invalidation to invalidate_range_end hook
In reply to Sean ChristophersonQuoted 10 lines · Sean Christopherson
Quoted 9 lines · David Woodhouse
Quoted 7 lines · Sean Christopherson
Quoted 1 line · David Woodhouse
From: David Woodhouse <[email protected]>
Servicing guest pages faults has the same problem, which is why
mmu_invalidate_retry_gfn()was added. Supporting hva-only GPCs made our lives a little harder, but not horrifically so (there are ordering differences regardless).Woefully incomplete, but I think this is the gist of what you want:
Quoted 8 lines
Quoted 3 lines
Hm, maybe. It does mean that migration occurring all through memory (indeed, just one at top and bottom of guest memory space) would perturb GPCs which remain present.
If that happens with a real world VMM, and it's not a blatant VMM goof, then we can fix KVM. The stage-2 page fault path hammers the
mmu_notifierretry logic far more than GPCs, so if a range-based check is inadequate for some use case, then we definitely need to fix both.
FWIW this turned out to be because someone used C++ where they shouldn't have done. Opinion is divided on whether to call that a 'blatant VMM goof'.
But if you use std::vector for something that's modified that many times a second, and free() implicitly calls madvise(MADV_DONTNEED) every time, you get what you deserve :)
We should still fix it in the kernel.
Sean Christopherson
maintainerPATCH v2 7/8 KVM: Do not pin pages tracked by gfn=>pfn caches
In reply to Sean ChristophersonPut the reference to any struct page mapped/tracked by a gfn=>pfn cache upon inserting the pfn into its associated cache, as opposed to putting the reference only when the cache is done using the pfn. In other words, don't pin pages while they're in the cache. One of the major roles of the gfn=>pfn cache is to play nicely with invalidation events, i.e. it exists in large part so that KVM doesn't rely on pinning pages.
--- virt/kvm/pfncache.c | 36 ++++++++++++++++++++---------------- 1 file changed, 20 insertions(+), 16 deletions(-)
virt/kvm/pfncache.c +20 -16
@@ -95,20 +95,16 @@ bool kvm_gfn_to_pfn_cache_check(struct kvm *kvm, struct gfn_to_pfn_cache *gpc, }EXPORT_SYMBOL_GPL (kvm_gfn_to_pfn_cache_check); -static void gpc_release_pfn_and_khva (struct kvm *kvm,kvm_pfn_t pfn,void *khva) +static void gpc_unmap_khva (struct kvm *kvm,kvm_pfn_t pfn,void *khva) { -/* Unmap the old page if it was mapped before, and release it */ -if (!is_error_noslot_pfn (pfn)) { -if (khva) { -if (pfn_valid (pfn)) -kunmap (pfn_to_page (pfn)); +/* Unmap the old pfn/page if it was mapped before. */ +if (!is_error_noslot_pfn (pfn) && khva) { +if (pfn_valid (pfn)) +kunmap (pfn_to_page (pfn));#ifdef CONFIG_HAS_IOMEM - else -memunmap (khva ); + else +memunmap (khva );#endif - } - -kvm_release_pfn (pfn, false); } }
@@ -147,10 +143,10 @@ static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)* Keep the mapping if the previous iteration reused * the existing mapping and didn't create a new one. */ -if (new_khva == old_khva) - new_khva =NULL ; +if (new_khva != old_khva) +gpc_unmap_khva (kvm, new_pfn, new_khva); -gpc_release_pfn_and_khva (kvm, new_pfn, new_khva); +kvm_release_pfn_clean (new_pfn);cond_resched (); }
@@ -222,6 +218,14 @@ static kvm_pfn_t hva_to_pfn_retry(struct kvm *kvm, struct gfn_to_pfn_cache *gpc) gpc->valid = true; gpc->pfn = new_pfn; gpc->khva = new_khva + (gpc->gpa & ~PAGE_MASK ); + +/* +* Put the reference to the _new_ pfn. The pfn is now tracked by the +* cache and can be safely migrated, swapped, etc... as the cache will +* invalidate any mappings in response to relevant mmu_notifier events. +*/ +kvm_release_pfn_clean (new_pfn); +return 0 ; out_error:
@@ -315,7 +319,7 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,write_unlock_irq (&gpc->lock);if (old_pfn != new_pfn) -gpc_release_pfn_and_khva (kvm, old_pfn, old_khva); +gpc_unmap_khva (kvm, old_pfn, old_khva);return ret; }
@@ -342,7 +346,7 @@ void kvm_gfn_to_pfn_cache_unmap(struct kvm *kvm, struct gfn_to_pfn_cache *gpc)write_unlock_irq (&gpc->lock); -gpc_release_pfn_and_khva (kvm, old_pfn, old_khva); +gpc_unmap_khva (kvm, old_pfn, old_khva); }EXPORT_SYMBOL_GPL (kvm_gfn_to_pfn_cache_unmap);
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Sean Christopherson
maintainerPATCH v2 8/8 DO NOT MERGE: Hack-a-test to verify gpc invalidation+refresh
In reply to Sean ChristophersonAdd a VM-wide gfn=>pfn cache and a fake MSR to let userspace control the cache. On writes, reflect the value of the MSR into the backing page of a gfn=>pfn cache so that userspace can detect if a value was written to the wrong page, i.e. to a stale mapping.
Spin up 16 vCPUs (arbitrary) to use/refresh the cache, and another thread to trigger mmu_notifier events and memslot updates.
--- arch/x86/kvm/x86.c | 30 ++++ include/linux/kvm_host.h | 2 + tools/testing/selftests/kvm/.gitignore | 1 + tools/testing/selftests/kvm/Makefile | 2 + tools/testing/selftests/kvm/gpc_test.c | 217 +++++++++++++++++++++++++ virt/kvm/pfncache.c | 2 + 6 files changed, 254 insertions(+) create mode 100644 tools/testing/selftests/kvm/gpc_test.c
arch/x86/kvm/x86.c +30 -0
@@ -3473,6 +3473,20 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, struct msr_data *msr_info)return kvm_xen_write_hypercall_page (vcpu, data);switch (msr) { +case 0xdeadbeefu : { +struct gfn_to_pfn_cache *gpc = &vcpu->kvm->test_cache; +unsigned long flags; + +if (kvm_gfn_to_pfn_cache_refresh (vcpu->kvm, gpc, data,8 )) +break ; + +read_lock_irqsave (&gpc->lock, flags); +if (kvm_gfn_to_pfn_cache_check (vcpu->kvm, gpc, data,8 )) + *(u64 *)(gpc->khva) = data; +read_unlock_irqrestore (&gpc->lock, flags); +break ; + } +case MSR_AMD64_NB_CFG :case MSR_IA32_UCODE_WRITE :case MSR_VM_HSAVE_PA :
@@ -3825,6 +3839,19 @@ static int get_msr_mce(struct kvm_vcpu *vcpu, u32 msr, u64 *pdata, bool host)int kvm_get_msr_common (struct kvm_vcpu *vcpu,struct msr_data *msr_info) {switch (msr_info->index) { +case 0xdeadbeefu : { +struct gfn_to_pfn_cache *gpc = &vcpu->kvm->test_cache; +unsigned long flags; + +read_lock_irqsave (&gpc->lock, flags); +if (kvm_gfn_to_pfn_cache_check (vcpu->kvm, gpc, gpc->gpa,8 )) + msr_info->data = gpc->gpa; +else + msr_info->data =0xdeadbeefu ; +read_unlock_irqrestore (&gpc->lock, flags); +return 0 ; + } +case MSR_IA32_PLATFORM_ID :case MSR_IA32_EBL_CR_POWERON :case MSR_IA32_LASTBRANCHFROMIP :
@@ -11794,6 +11821,9 @@ int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)kvm_hv_init_vm (kvm);kvm_xen_init_vm (kvm); +kvm_gfn_to_pfn_cache_init (kvm, &kvm->test_cache,NULL , +KVM_HOST_USES_PFN ,0 ,0 ); +return static_call (kvm_x86_vm_init)(kvm); out_page_track:
include/linux/kvm_host.h +2 -0
@@ -718,6 +718,8 @@ struct kvm {spinlock_t gpc_lock;struct list_head gpc_list; +struct gfn_to_pfn_cache test_cache; +/* * created_vcpus is protected by kvm->lock, and is incremented * at the beginning of KVM_CREATE_VCPU. online_vcpus is only
tools/testing/selftests/kvm/.gitignore +1 -0
@@ -70,3 +70,4 @@ /steal_time /kvm_binary_stats_test /system_counter_offset_test +/gpc_test
tools/testing/selftests/kvm/Makefile +2 -0
@@ -104,6 +104,8 @@ TEST_GEN_PROGS_x86_64 += steal_time TEST_GEN_PROGS_x86_64 += kvm_binary_stats_test TEST_GEN_PROGS_x86_64 += system_counter_offset_test +TEST_GEN_PROGS_x86_64 += gpc_test + TEST_GEN_PROGS_aarch64 += aarch64/arch_timer TEST_GEN_PROGS_aarch64 += aarch64/debug-exceptions TEST_GEN_PROGS_aarch64 += aarch64/get-reg-list
tools/testing/selftests/kvm/gpc_test.c new file +217 -0
217 lines not on this page · raw message
virt/kvm/pfncache.c +2 -0
@@ -372,6 +372,8 @@ int kvm_gfn_to_pfn_cache_init(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,list_add (&gpc->list, &kvm->gpc_list);spin_unlock (&kvm->gpc_lock); } +if (!len) +return -EINVAL ;return kvm_gfn_to_pfn_cache_refresh (kvm, gpc, gpa, len); }EXPORT_SYMBOL_GPL (kvm_gfn_to_pfn_cache_init);
Signature · 1 line
2.36.0.rc2.479.g8af0fa9b8e-goog
Re: [PATCH v2 8/8] DO NOT MERGE: Hack-a-test to verify gpc invalidation+refresh
Quoted 7 lines · Sean Christopherson
Add a VM-wide gfn=>pfn cache and a fake MSR to let userspace control the cache. On writes, reflect the value of the MSR into the backing page of a gfn=>pfn cache so that userspace can detect if a value was written to the wrong page, i.e. to a stale mapping.
Spin up 16 vCPUs (arbitrary) to use/refresh the cache, and another thread to trigger
mmu_notifierevents and memslot updates.
Do you need the MSR hack? Can't you exercise this using Xen interrupt delivery or runstate information and the same kind of thread setup?
Thanks for working on this!
Sean Christopherson
maintainerRe: [PATCH v2 8/8] DO NOT MERGE: Hack-a-test to verify gpc invalidation+refresh
Quoted 11 lines · David Woodhouse
Quoted 7 lines · Sean Christopherson
Add a VM-wide gfn=>pfn cache and a fake MSR to let userspace control the cache. On writes, reflect the value of the MSR into the backing page of a gfn=>pfn cache so that userspace can detect if a value was written to the wrong page, i.e. to a stale mapping.
Spin up 16 vCPUs (arbitrary) to use/refresh the cache, and another thread to trigger
mmu_notifierevents and memslot updates.Do you need the MSR hack? Can't you exercise this using Xen interrupt delivery or runstate information and the same kind of thread setup?
Yeah, I asumme it's possible, and medium/long term I definitely want to have a proper test. I went the hack route to get something that could hammer a cache with minimal chance of a test bug. I only have a rough idea of what the Xen stuff does.