KVM: Dirty ring fixes and cleanups
Sean Christopherson
reviewerv3 cover KVM: Dirty ring fixes and cleanups
Fix issues with dirty ring harvesting where KVM doesn't bound the processing of entries in any way, which allows userspace to keep KVM in a tight loop indefinitely.
E.g.
struct kvm_dirty_gfn *dirty_gfns = vcpu_map_dirty_ring(vcpu);
if (fork()) {
int r;
for (;;) {
r = kvm_vm_reset_dirty_ring(vcpu->vm);
if (r)
printf("RESET %d dirty ring entries\n", r);
}
} else {
int i;
for (i = 0; i < test_dirty_ring_count; i++) {
dirty_gfns[i].slot = TEST_MEM_SLOT_INDEX;
dirty_gfns[i].offset = (i * 64) % host_num_pages;
}
for (;;) {
for (i = 0; i < test_dirty_ring_count; i++)
WRITE_ONCE(dirty_gfns[i].flags, KVM_DIRTY_GFN_F_RESET);
}
}Patches 1-3 address that class of bugs. Patches 4-6 are cleanups.
v3:
- Fix typos (I apparently can't spell opportunistically to save my life). [Binbin, James]
- Clean up stale comments. [Binbin]
- Collect reviews. [James, Pankaj]
- Add a lockdep assertion on slots_lock, along with a comment. [James]
v2:
- https://lore.kernel.org/all/[email protected]
- Expand on comments in dirty ring harvesting code. [Yan]
v1: https://lore.kernel.org/all/[email protected]
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert that slots_lock is held when resetting per-vCPU dirty
rings
include/linux/kvm_dirty_ring.h | 18 ++---- virt/kvm/dirty_ring.c | 111 +++++++++++++++++++++++---------- virt/kvm/kvm_main.c | 9 ++- 3 files changed, 89 insertions(+), 49 deletions(-)
base-commit: 7ef51a41466bc846ad794d505e2e34ff97157f7f
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Sean Christopherson
reviewerPATCH v3 1/6 KVM: Bound the number of dirty ring entries in a single reset at INT_MAX
Cap the number of ring entries that are reset in a single ioctl to INT_MAX to ensure userspace isn't confused by a wrap into negative space, and so that, in a truly pathological scenario, KVM doesn't miss a TLB flush due to the count wrapping to zero. While the size of the ring is fixed at 0x10000 entries and KVM (currently) supports at most 4096, userspace is allowed to harvest entries from the ring while the reset is in-progress, i.e. it's possible for the ring to always have harvested entries.
Opportunistically return an actual error code from the helper so that a future fix to handle pending signals can gracefully return -EINTR. Drop the function comment now that the return code is a stanard 0/-errno (and because a future commit will add a proper lockdep assertion).
Opportunistically drop a similarly stale comment for kvm_dirty_ring_push().
--- include/linux/kvm_dirty_ring.h | 18 +++++------------- virt/kvm/dirty_ring.c | 10 +++++----- virt/kvm/kvm_main.c | 9 ++++++--- 3 files changed, 16 insertions(+), 21 deletions(-)
include/linux/kvm_dirty_ring.h +5 -13
@@ -49,9 +49,10 @@ static inline int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *r }static inline int kvm_dirty_ring_reset (struct kvm *kvm, -struct kvm_dirty_ring *ring) +struct kvm_dirty_ring *ring, +int *nr_entries_reset) { -return 0 ; +return -ENOENT ; } static inline void kvm_dirty_ring_push(struct kvm_vcpu *vcpu,
@@ -77,17 +78,8 @@ bool kvm_arch_allow_write_without_running_vcpu(struct kvm *kvm);u32 kvm_dirty_ring_get_rsvd_entries (struct kvm *kvm);int kvm_dirty_ring_alloc (struct kvm *kvm,struct kvm_dirty_ring *ring,int index,u32 size); - -/* -* called with kvm->slots_lock held, returns the number of -* processed pages. -*/ -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring); - -/* -* returns =0: successfully pushed -* <0: unable to push, need to wait -*/ +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset);void kvm_dirty_ring_push (struct kvm_vcpu *vcpu,u32 slot,u64 offset);bool kvm_dirty_ring_check_request (struct kvm_vcpu *vcpu);
virt/kvm/dirty_ring.c +5 -5
@@ -105,19 +105,19 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)return smp_load_acquire (&gfn->flags) &KVM_DIRTY_GFN_F_RESET ; } -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring) +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset) {u32 cur_slot, next_slot;u64 cur_offset, next_offset;unsigned long mask; -int count =0 ;struct kvm_dirty_gfn *entry;bool first_round = true;/* This is only needed to make compilers happy */ cur_slot = cur_offset = mask =0 ; -while (true) { +while (likely ((*nr_entries_reset) <INT_MAX )) { entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))
@@ -130,7 +130,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)kvm_dirty_gfn_set_invalid (entry); ring->reset_index++; - count++; + (*nr_entries_reset)++;/* * Try to coalesce the reset operations when the guest is * scanning pages in the same slot.
@@ -167,7 +167,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)trace_kvm_dirty_ring_reset (ring); -return count; +return 0 ; }void kvm_dirty_ring_push (struct kvm_vcpu *vcpu,u32 slot,u64 offset)
virt/kvm/kvm_main.c +6 -3
@@ -4903,15 +4903,18 @@ static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm) {unsigned long i;struct kvm_vcpu *vcpu; -int cleared =0 ; +int cleared =0 , r;if (!kvm->dirty_ring_size)return -EINVAL ;mutex_lock (&kvm->slots_lock); -kvm_for_each_vcpu (i, vcpu, kvm) - cleared +=kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring); +kvm_for_each_vcpu (i, vcpu, kvm) { + r =kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring, &cleared); +if (r) +break ; + }mutex_unlock (&kvm->slots_lock);
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Re: [PATCH v3 1/6] KVM: Bound the number of dirty ring entries in a single reset at INT_MAX
Quoted 11 lines · Sean Christopherson
Cap the number of ring entries that are reset in a single ioctl to
INT_MAXto ensure userspace isn't confused by a wrap into negative space, and so that, in a truly pathological scenario, KVM doesn't miss a TLB flush due to the count wrapping to zero. While the size of the ring is fixed at0x10000entries and KVM (currently) supports at most 4096, userspace is allowed to harvest entries from the ring while the reset is in-progress, i.e. it's possible for the ring to always have harvested entries.Opportunistically return an actual error code from the helper so that a future fix to handle pending signals can gracefully return
-EINTR. Drop the function comment now that the return code is a stanard 0/-errno (and
stanard -> standard
The rest looks good to me.
Quoted 125 lines
because a future commit will add a proper lockdep assertion).
Opportunistically drop a similarly stale comment for
kvm_dirty_ring_push().Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Cc Binbin Wu <[email protected]>Fixes fb04a1eddb1a ("KVM: X86: Implement ring-based dirty memory tracking")Reviewed-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>--- include/linux/kvm_dirty_ring.h | 18 +++++------------- virt/kvm/dirty_ring.c | 10 +++++----- virt/kvm/kvm_main.c | 9 ++++++--- 3 files changed, 16 insertions(+), 21 deletions(-)include/linux/kvm_dirty_ring.h +5
-13@@ -49,9 +49,10 @@ static inline int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *r }static inline int kvm_dirty_ring_reset (struct kvm *kvm, -struct kvm_dirty_ring *ring) +struct kvm_dirty_ring *ring, +int *nr_entries_reset) { -return 0 ; +return -ENOENT ; } static inline void kvm_dirty_ring_push(struct kvm_vcpu *vcpu,@@ -77,17 +78,8 @@ bool kvm_arch_allow_write_without_running_vcpu(struct kvm *kvm);u32 kvm_dirty_ring_get_rsvd_entries (struct kvm *kvm);int kvm_dirty_ring_alloc (struct kvm *kvm,struct kvm_dirty_ring *ring,int index,u32 size); - -/* -* called with kvm->slots_lock held, returns the number of -* processed pages. -*/ -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring); - -/* -* returns =0: successfully pushed -* <0: unable to push, need to wait -*/ +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset);void kvm_dirty_ring_push (struct kvm_vcpu *vcpu,u32 slot,u64 offset);bool kvm_dirty_ring_check_request (struct kvm_vcpu *vcpu);virt/kvm/dirty_ring.c +5
-5@@ -105,19 +105,19 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)return smp_load_acquire (&gfn->flags) &KVM_DIRTY_GFN_F_RESET ; } -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring) +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset) {u32 cur_slot, next_slot;u64 cur_offset, next_offset;unsigned long mask; -int count =0 ;struct kvm_dirty_gfn *entry;bool first_round = true;/* This is only needed to make compilers happy */ cur_slot = cur_offset = mask =0 ; -while (true) { +while (likely ((*nr_entries_reset) <INT_MAX )) { entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))@@ -130,7 +130,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)kvm_dirty_gfn_set_invalid (entry); ring->reset_index++; - count++; + (*nr_entries_reset)++;/* * Try to coalesce the reset operations when the guest is * scanning pages in the same slot. @@ -167,7 +167,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)trace_kvm_dirty_ring_reset (ring); -return count; +return 0 ; }void kvm_dirty_ring_push (struct kvm_vcpu *vcpu,u32 slot,u64 offset)virt/kvm/kvm_main.c +6
-3@@ -4903,15 +4903,18 @@ static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm) {unsigned long i;struct kvm_vcpu *vcpu; -int cleared =0 ; +int cleared =0 , r;if (!kvm->dirty_ring_size)return -EINVAL ;mutex_lock (&kvm->slots_lock); -kvm_for_each_vcpu (i, vcpu, kvm) - cleared +=kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring); +kvm_for_each_vcpu (i, vcpu, kvm) { + r =kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring, &cleared); +if (r) +break ; + }mutex_unlock (&kvm->slots_lock);
Sean Christopherson
reviewerPATCH v3 2/6 KVM: Bail from the dirty ring reset flow if a signal is pending
In reply to Sean ChristophersonAbort a dirty ring reset if the current task has a pending signal, as the hard limit of INT_MAX entries doesn't ensure KVM will respond to a signal in a timely fashion.
--- virt/kvm/dirty_ring.c | 3 +++ 1 file changed, 3 insertions(+)
virt/kvm/dirty_ring.c +3 -0
@@ -118,6 +118,9 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = cur_offset = mask =0 ;while (likely ((*nr_entries_reset) <INT_MAX )) { +if (signal_pending (current)) +return -EINTR ; + entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Re: [PATCH v3 2/6] KVM: Bail from the dirty ring reset flow if a signal is pending
Quoted 7 lines · Sean Christopherson
Abort a dirty ring reset if the current task has a pending signal, as the hard limit of
INT_MAXentries doesn't ensure KVM will respond to a signal in a timely fashion.Fixes fb04a1eddb1a ("KVM: X86: Implement ring-based dirty memory tracking")Reviewed-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
Quoted 18 lines
--- virt/kvm/dirty_ring.c | 3 +++ 1 file changed, 3 insertions(+)virt/kvm/dirty_ring.c +3
-0@@ -118,6 +118,9 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = cur_offset = mask =0 ;while (likely ((*nr_entries_reset) <INT_MAX )) { +if (signal_pending (current)) +return -EINTR ; + entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))
Sean Christopherson
reviewerPATCH v3 3/6 KVM: Conditionally reschedule when resetting the dirty ring
In reply to Sean ChristophersonWhen resetting a dirty ring, conditionally reschedule on each iteration after the first. The recently introduced hard limit mitigates the issue of an endless reset, but isn't sufficient to completely prevent RCU stalls, soft lockups, etc., nor is the hard limit intended to guard against such badness.
Note! Take care to check for reschedule even in the "continue" paths, as a pathological scenario (or malicious userspace) could dirty the same gfn over and over, i.e. always hit the continue path.
rcu: INFO: rcu_sched self-detected stall on CPU rcu: 4-....: (5249 ticks this GP) idle=51e4/1/0x4000000000000000 softirq=309/309 fqs=2563 rcu: (t=5250 jiffies g=-319 q=608 ncpus=24) CPU: 4 UID: 1000 PID: 1067 Comm: dirty_log_test Tainted: G L 6.13.0-rc3-17fa7a24ea1e-HEAD-vm #814 Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x26/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK> Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x17/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK>
--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
virt/kvm/dirty_ring.c +10 -0
@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched (); +/* * Try to coalesce the reset operations when the guest is * scanning pages in the same slot.
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Sean Christopherson
reviewerPATCH v3 4/6 KVM: Check for empty mask of harvested dirty ring entries in caller
In reply to Sean ChristophersonWhen resetting a dirty ring, explicitly check that there is work to be done before calling kvm_reset_dirty_gfn(), e.g. if no harvested entries are found and/or on the loop's first iteration, and delete the extremely misleading comment "This is only needed to make compilers happy". KVM absolutely relies on mask to be zero-initialized, i.e. the comment is an outright lie. Furthermore, the compiler is right to complain that KVM is calling a function with uninitialized data, as there are no guarantees the implementation details of kvm_reset_dirty_gfn() will be visible to kvm_dirty_ring_reset().
While the flaw could be fixed by simply deleting (or rewording) the comment, and duplicating the check is unfortunate, checking mask in the caller will allow for additional cleanups.
Opportunistically drop the zero-initialization of cur_slot and cur_offset. If a bug were introduced where either the slot or offset was consumed before mask is set to a non-zero value, then it is highly desirable for the compiler (or some other sanitizer) to yell.
--- virt/kvm/dirty_ring.c | 44 ++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-)
virt/kvm/dirty_ring.c +35 -9
@@ -55,9 +55,6 @@ static void kvm_reset_dirty_gfn(struct kvm *kvm, u32 slot, u64 offset, u64 mask)struct kvm_memory_slot *memslot;int as_id, id; -if (!mask) -return ; - as_id = slot >>16 ; id = (u16 )slot;
@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For +* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched. +*/ u32 cur_slot, next_slot;u64 cur_offset, next_offset; -unsigned long mask; +unsigned long mask =0 ;struct kvm_dirty_gfn *entry;bool first_round = true; -/* This is only needed to make compilers happy */ - cur_slot = cur_offset = mask =0 ; -while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;
@@ -164,14 +170,34 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,continue ; } } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* Reset the slot for all the harvested entries that have been +* gathered, but not yet fully processed. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* The current slot was reset or this is the first harvested +* entry, (re)initialize the metadata. +*/ cur_slot = next_slot; cur_offset = next_offset; mask =1 ; first_round = false; } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with the "current" entry(s), and that reset processes the +* _current_ entry(s); i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Re: [PATCH v3 4/6] KVM: Check for empty mask of harvested dirty ring entries in caller
Quoted 19 lines · Sean Christopherson
When resetting a dirty ring, explicitly check that there is work to be done before calling
kvm_reset_dirty_gfn(), e.g. if no harvested entries are found and/or on the loop's first iteration, and delete the extremely misleading comment "This is only needed to make compilers happy". KVM absolutely relies on mask to be zero-initialized, i.e. the comment is an outright lie. Furthermore, the compiler is right to complain that KVM is calling a function with uninitialized data, as there are no guarantees the implementation details ofkvm_reset_dirty_gfn()will be visible tokvm_dirty_ring_reset().While the flaw could be fixed by simply deleting (or rewording) the comment, and duplicating the check is unfortunate, checking mask in the caller will allow for additional cleanups.
Opportunistically drop the zero-initialization of
cur_slotandcur_offset. If a bug were introduced where either the slot or offset was consumed before mask is set to a non-zero value, then it is highly desirable for the compiler (or some other sanitizer) to yell.
Quoted 7 lines
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Cc Binbin Wu <[email protected]>Reviewed-by Pankaj Gupta <[email protected]>Reviewed-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
Quoted 84 lines
--- virt/kvm/dirty_ring.c | 44 ++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-)virt/kvm/dirty_ring.c +35
-9@@ -55,9 +55,6 @@ static void kvm_reset_dirty_gfn(struct kvm *kvm, u32 slot, u64 offset, u64 mask)struct kvm_memory_slot *memslot;int as_id, id; -if (!mask) -return ; - as_id = slot >>16 ; id = (u16 )slot;@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For +* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched. +*/ u32 cur_slot, next_slot;u64 cur_offset, next_offset; -unsigned long mask; +unsigned long mask =0 ;struct kvm_dirty_gfn *entry;bool first_round = true; -/* This is only needed to make compilers happy */ - cur_slot = cur_offset = mask =0 ; -while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;@@ -164,14 +170,34 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,continue ; } } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* Reset the slot for all the harvested entries that have been +* gathered, but not yet fully processed. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* The current slot was reset or this is the first harvested +* entry, (re)initialize the metadata. +*/ cur_slot = next_slot; cur_offset = next_offset; mask =1 ; first_round = false; } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with the "current" entry(s), and that reset processes the +* _current_ entry(s); i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared
Re: [PATCH v3 4/6] KVM: Check for empty mask of harvested dirty ring entries in caller
In reply to Sean Christophersonvirt/kvm/dirty_ring.c static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn) · line 111
Quoted 8 lines · Sean Christopherson
virt/kvm/dirty_ring.c +4
-0@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For
Suppose N is 64,
virt/kvm/dirty_ring.c static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn) · line 118
Quoted 7 lines
virt/kvm/dirty_ring.c +7
-0+* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched.
X-N can't be batched, right?
Quoted 54 lines
virt/kvm/dirty_ring.c +24
-6+*/ u32 cur_slot, next_slot;u64 cur_offset, next_offset; -unsigned long mask; +unsigned long mask =0 ;struct kvm_dirty_gfn *entry;bool first_round = true; -/* This is only needed to make compilers happy */ - cur_slot = cur_offset = mask =0 ; -while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;@@ -164,14 +170,34 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,continue ; } } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* Reset the slot for all the harvested entries that have been +* gathered, but not yet fully processed. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* The current slot was reset or this is the first harvested +* entry, (re)initialize the metadata. +*/ cur_slot = next_slot; cur_offset = next_offset; mask =1 ; first_round = false; } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with the "current" entry(s), and that reset processes the +* _current_ entry(s); i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Sean Christopherson
reviewerRe: [PATCH v3 4/6] KVM: Check for empty mask of harvested dirty ring entries in caller
Quoted 11 lines · Yan Zhao
virt/kvm/dirty_ring.c static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn) · line 111
Quoted 8 lines · Sean Christopherson
virt/kvm/dirty_ring.c +4
-0@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For Suppose N is 64,
Quoted 8 lines
virt/kvm/dirty_ring.c static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn) · line 118
Quoted 7 lines
virt/kvm/dirty_ring.c +7
-0+* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched. X-N can't be batched, right?
Hah! Yeah, off-by-one error.
Sean Christopherson
reviewerPATCH v3 5/6 KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
In reply to Sean ChristophersonUse "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to first_round.
Opportunistically combine the adjacent "if (mask)" statements into a single if-statement.
No functional change intended.
--- virt/kvm/dirty_ring.c | 60 +++++++++++++++++++++---------------------- 1 file changed, 29 insertions(+), 31 deletions(-)
virt/kvm/dirty_ring.c +29 -31
@@ -121,7 +121,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; -bool first_round = true;while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))
@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - } -/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + }/* * The current slot was reset or this is the first harvested
@@ -185,7 +184,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = next_slot; cur_offset = next_offset; mask =1 ; - first_round = false; }/*
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Re: [PATCH v3 5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
Quoted 8 lines · Sean Christopherson
Use "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to
first_round.Opportunistically combine the adjacent "if (mask)" statements into a single if-statement.
Quoted 8 lines
No functional change intended.
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Reviewed-by Pankaj Gupta <[email protected]>Reviewed-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
Quoted 96 lines
--- virt/kvm/dirty_ring.c | 60 +++++++++++++++++++++---------------------- 1 file changed, 29 insertions(+), 31 deletions(-)virt/kvm/dirty_ring.c +29
-31@@ -121,7 +121,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; -bool first_round = true;while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - } -/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + }/* * The current slot was reset or this is the first harvested @@ -185,7 +184,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = next_slot; cur_offset = next_offset; mask =1 ; - first_round = false; }/*
Re: [PATCH v3 5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
In reply to Sean ChristophersonQuoted 90 lines · Sean Christopherson
Use "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to
first_round.Opportunistically combine the adjacent "if (mask)" statements into a single if-statement.
No functional change intended.
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Reviewed-by Pankaj Gupta <[email protected]>Reviewed-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>--- virt/kvm/dirty_ring.c | 60 +++++++++++++++++++++---------------------- 1 file changed, 29 insertions(+), 31 deletions(-)virt/kvm/dirty_ring.c +24
-25@@ -121,7 +121,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; -bool first_round = true;while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - }
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 177
Quoted 10 lines
virt/kvm/dirty_ring.c +4
-5-/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);
Nit and feel free to ignore it :)
Would it be better to move the "cond_resched()" to here, i.e., executing it for at most every 64 entries?
Quoted 15 lines
virt/kvm/dirty_ring.c +1
-1+ }/* * The current slot was reset or this is the first harvested @@ -185,7 +184,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = next_slot; cur_offset = next_offset; mask =1 ; - first_round = false; }/* Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Sean Christopherson
reviewerRe: [PATCH v3 5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
Quoted 71 lines · Yan Zhao
Quoted 58 lines · Sean Christopherson
virt/kvm/dirty_ring.c +24
-24@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - }virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 177
Quoted 10 lines
virt/kvm/dirty_ring.c +4
-5-/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);Nit and feel free to ignore it :)
Quoted 2 lines
Would it be better to move the "
cond_resched()" to here, i.e., executing it for at most every 64 entries?
Hmm, yeah, I think that makes sense. The time spent manipulating the ring and mask+offset is quite trivial, so checking on every single entry is unnecessary.
Sean Christopherson
reviewerRe: [PATCH v3 5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
Quoted 72 lines · Sean Christopherson
Quoted 71 lines · Yan Zhao
Quoted 58 lines · Sean Christopherson
virt/kvm/dirty_ring.c +24
-24@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - }virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 177
Quoted 10 lines
virt/kvm/dirty_ring.c +4
-5-/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);Nit and feel free to ignore it :)
Quoted 5 lines
Quoted 2 lines
Would it be better to move the "
cond_resched()" to here, i.e., executing it for at most every 64 entries?Hmm, yeah, I think that makes sense. The time spent manipulating the ring and mask+offset is quite trivial, so checking on every single entry is unnecessary.
Oh, no, scratch that. Thankfully, past me explicitly documented this. From patch 3:
Note! Take care to check for reschedule even in the "continue" paths, as a pathological scenario (or malicious userspace) could dirty the same gfn over and over, i.e. always hit the continue path.
A batch isn't guaranteed to be flushed after processing 64 entries, it's only flushed when an entry more than N gfns away is encountered.
Re: [PATCH v3 5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
Quoted 82 lines · Sean Christopherson
Quoted 72 lines · Sean Christopherson
Quoted 71 lines · Yan Zhao
Quoted 58 lines · Sean Christopherson
virt/kvm/dirty_ring.c +24
-24@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - }virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 177
Quoted 10 lines
virt/kvm/dirty_ring.c +4
-5-/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);Nit and feel free to ignore it :)
Quoted 5 lines
Quoted 2 lines
Would it be better to move the "
cond_resched()" to here, i.e., executing it for at most every 64 entries?Hmm, yeah, I think that makes sense. The time spent manipulating the ring and mask+offset is quite trivial, so checking on every single entry is unnecessary.
Oh, no, scratch that. Thankfully, past me explicitly documented this. From patch 3:
Quoted 6 lines
Note! Take care to check for reschedule even in the "continue" paths, as a pathological scenario (or malicious userspace) could dirty the same gfn over and over, i.e. always hit the continue path.
A batch isn't guaranteed to be flushed after processing 64 entries, it's only flushed when an entry more than N gfns away is encountered.
Oh, I overlooked the "pathological scenario". You are right!
Sean Christopherson
reviewerPATCH v3 6/6 KVM: Assert that slots_lock is held when resetting per-vCPU dirty rings
In reply to Sean ChristophersonAssert that slots_lock is held in kvm_dirty_ring_reset() and add a comment to explain _why_ slots needs to be held for the duration of the reset.
--- virt/kvm/dirty_ring.c | 8 ++++++++ 1 file changed, 8 insertions(+)
virt/kvm/dirty_ring.c +8 -0
@@ -122,6 +122,14 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +/* +* Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, +* e.g. so that KVM fully resets all entries processed by a given call +* before returning to userspace. Holding slots_lock also protects +* the various memslot accesses. +*/ +lockdep_assert_held (&kvm->slots_lock); +while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;
Signature · 1 line
2.49.0.1112.g889b7c5bd8-goog
Re: [PATCH v3 6/6] KVM: Assert that slots_lock is held when resetting per-vCPU dirty rings
Quoted 15 lines · Sean Christopherson
Assert that
slots_lockis held inkvm_dirty_ring_reset()and add a comment to explain _why_ slots needs to be held for the duration of the reset.Suggested-by James Houghton <[email protected]>Signed-off-by Sean Christopherson <[email protected]>--- virt/kvm/dirty_ring.c | 8 ++++++++ 1 file changed, 8 insertions(+)virt/kvm/dirty_ring.c +0
-0
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 127
Quoted 6 lines
virt/kvm/dirty_ring.c +3
-0@@ -122,6 +122,14 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +/* +* Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, +* e.g. so that KVM fully resets all entries processed by a given call
It seems that "e.g." is not needed?
Quoted 8 lines
virt/kvm/dirty_ring.c +5
-0+* before returning to userspace. Holding slots_lock also protects +* the various memslot accesses. +*/ +lockdep_assert_held (&kvm->slots_lock); +while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;
Quoted 10 lines · Sean Christopherson
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert thatslots_lockis held when resetting per-vCPU dirty
rings
For the last one, I'd think it's majorly because of the memslot accesses (or CONFIG_LOCKDEP=y should yell already on resets?). The "serialization of concurrent RESETs" part could be a good side effect. After all, the dirty rings rely a lot on the userspace to do right things.. for example, the userspace better also remember to reset before any slot changes, or it's possible to collect a dirty pfn with a slot index that was already removed and reused with a new one..
Maybe we could switch the sentences there in the comment of last patch, but not a huge deal.
Thanks!
Signature · 1 line
Peter Xu
Sean Christopherson
reviewerQuoted 12 lines · Peter Xu
Quoted 10 lines · Sean Christopherson
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert thatslots_lockis held when resetting per-vCPU dirty
rings
Quoted 2 lines
For the last one, I'd think it's majorly because of the memslot accesses (or
CONFIG_LOCKDEP=yshould yell already on resets?).
No? If KVM only needed to ensure stable memslot accesses, then SRCU would suffice. It sounds like holding slots_lock may have been a somewhat unintentional, but the reason KVM can't switch to SRCU is that doing so would break ordering, not because slots_lock is needed to protect the memslot accesses.
Quoted 16 lines
The "serialization of concurrent RESETs" part could be a good side effect. After all, the dirty rings rely a lot on the userspace to do right things.. for example, the userspace better also remember to reset before any slot changes, or it's possible to collect a dirty pfn with a slot index that was already removed and reused with a new one..
Maybe we could switch the sentences there in the comment of last patch, but not a huge deal.
Reviewed-by Peter Xu <[email protected]>Thanks!
Signature · 1 line
Peter Xu
Quoted 13 lines · Sean Christopherson
Quoted 12 lines · Peter Xu
Quoted 10 lines · Sean Christopherson
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert thatslots_lockis held when resetting per-vCPU dirty
rings
Quoted 7 lines
Quoted 2 lines
For the last one, I'd think it's majorly because of the memslot accesses (or
CONFIG_LOCKDEP=yshould yell already on resets?).No? If KVM only needed to ensure stable memslot accesses, then SRCU would suffice. It sounds like holding
slots_lockmay have been a somewhat unintentional, but the reason KVM can't switch to SRCU is that doing so would break ordering, not becauseslots_lockis needed to protect the memslot accesses.
Hmm.. isn't what you said exactly means a "yes"? :)
I mean, I would still expect lockdep to report this ioctl if without the slots_lock, please correct me if it's not the case. And if using RCU is not trivial (or not necessary either), so far the slots_lock is still required to make sure the memslot accesses are legal?
Signature · 1 line
Peter Xu
Sean Christopherson
reviewerQuoted 22 lines · Peter Xu
Quoted 13 lines · Sean Christopherson
Quoted 12 lines · Peter Xu
Quoted 10 lines · Sean Christopherson
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert thatslots_lockis held when resetting per-vCPU dirty
ringsQuoted 7 lines
Quoted 2 lines
For the last one, I'd think it's majorly because of the memslot accesses (or
CONFIG_LOCKDEP=yshould yell already on resets?).No? If KVM only needed to ensure stable memslot accesses, then SRCU would suffice. It sounds like holding
slots_lockmay have been a somewhat unintentional, but the reason KVM can't switch to SRCU is that doing so would break ordering, not becauseslots_lockis needed to protect the memslot accesses.
Quoted 4 lines
Hmm.. isn't what you said exactly means a "yes"? :)
I mean, I would still expect lockdep to report this ioctl if without the
slots_lock, please correct me if it's not the case.
Yes, one of slots_lock or SRCU needs to be held.
Quoted 2 lines
And if using RCU is not trivial (or not necessary either), so far the
slots_lockis still required to make sure the memslot accesses are legal?
I don't follow this part. The intent of the comment is to document why slots_lock is required, which is exceptional because memslot access for readers are protected by kvm->srcu. The fact that slots_lock also protects memslots is notable only because it makes acquiring kvm->srcu superfluous. But grabbing kvm->srcu is still safe/legal/ok:
virt/kvm/dirty_ring.c +13 -3
@@ -121,18 +121,26 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +int idx;/* * Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, * e.g. so that KVM fully resets all entries processed by a given call -* before returning to userspace. Holding slots_lock also protects -* the various memslot accesses. +* before returning to userspace. */ lockdep_assert_held (&kvm->slots_lock); +/* +* Holding slots_lock also protects the various memslot accesses, but +* acquiring kvm->srcu for read here is still safe, just unnecessary. +*/ + idx =srcu_read_lock (&kvm->srcu); +while (likely ((*nr_entries_reset) <INT_MAX )) { -if (signal_pending (current)) +if (signal_pending (current)) { +srcu_read_unlock (&kvm->srcu, idx);return -EINTR ; + } entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];
@@ -205,6 +213,8 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,if (mask)kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +srcu_read_unlock (&kvm->srcu, idx); +/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared * by the VCPU thread next time when it enters the guest.
--
And unless there are other behaviors that are protected by slots_lock (which is entirely possible), serializing the processing of each ring could be done via a dedicated (for example only, the dedicated mutex could/should be per-vCPU, not global).
This diff in particular shows why I ordered and phrased the comment the way I did. The blurb about protecting memslot accesses is purely a friendly reminder to readers. The sole reason for an assert and comment is to call out the need for ordering.
virt/kvm/dirty_ring.c +12 -4
@@ -102,6 +102,8 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)return smp_load_acquire (&gfn->flags) &KVM_DIRTY_GFN_F_RESET ; } +static DEFINE_MUTEX (per_ring_lock); +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) {
@@ -121,18 +123,22 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +int idx;/* * Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, * e.g. so that KVM fully resets all entries processed by a given call -* before returning to userspace. Holding slots_lock also protects -* the various memslot accesses. +* before returning to userspace. */ -lockdep_assert_held (&kvm->slots_lock); +guard (mutex)(&per_ring_lock); + + idx =srcu_read_lock (&kvm->srcu);while (likely ((*nr_entries_reset) <INT_MAX )) { -if (signal_pending (current)) +if (signal_pending (current)) { +srcu_read_unlock (&kvm->srcu, idx);return -EINTR ; + } entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];
@@ -205,6 +211,8 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,if (mask)kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +srcu_read_unlock (&kvm->srcu, idx); +/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared * by the VCPU thread next time when it enters the guest.
virt/kvm/kvm_main.c +0 -5
@@ -4908,16 +4908,12 @@ static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm)if (!kvm->dirty_ring_size)return -EINVAL ; -mutex_lock (&kvm->slots_lock); -kvm_for_each_vcpu (i, vcpu, kvm) { r =kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring, &cleared);if (r)break ; } -mutex_unlock (&kvm->slots_lock); -if (cleared)kvm_flush_remote_tlbs (kvm); --
Quoted 28 lines · Sean Christopherson
Quoted 22 lines · Peter Xu
Quoted 13 lines · Sean Christopherson
Quoted 12 lines · Peter Xu
Quoted 10 lines · Sean Christopherson
Sean Christopherson (6):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
KVM: Assert thatslots_lockis held when resetting per-vCPU dirty
ringsQuoted 7 lines
Quoted 2 lines
For the last one, I'd think it's majorly because of the memslot accesses (or
CONFIG_LOCKDEP=yshould yell already on resets?).No? If KVM only needed to ensure stable memslot accesses, then SRCU would suffice. It sounds like holding
slots_lockmay have been a somewhat unintentional, but the reason KVM can't switch to SRCU is that doing so would break ordering, not becauseslots_lockis needed to protect the memslot accesses.Quoted 4 lines
Hmm.. isn't what you said exactly means a "yes"? :)
I mean, I would still expect lockdep to report this ioctl if without the
slots_lock, please correct me if it's not the case.
Quoted 8 lines
Yes, one of
slots_lockor SRCU needs to be held.Quoted 2 lines
And if using RCU is not trivial (or not necessary either), so far the
slots_lockis still required to make sure the memslot accesses are legal?I don't follow this part. The intent of the comment is to document why
slots_lockis required, which is exceptional because memslot access for readers are protected bykvm->srcu.
I always think it's fine to take slots_lock for readers too. RCU can definitely be better in most cases..
Quoted 49 lines
The fact that
slots_lockalso protects memslots is notable only because it makes acquiringkvm->srcusuperfluous. But grabbingkvm->srcuis still safe/legal/ok:virt/kvm/dirty_ring.c +13
-4@@ -121,18 +121,26 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +int idx;/* * Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, * e.g. so that KVM fully resets all entries processed by a given call -* before returning to userspace. Holding slots_lock also protects -* the various memslot accesses. +* before returning to userspace. */ lockdep_assert_held (&kvm->slots_lock); +/* +* Holding slots_lock also protects the various memslot accesses, but +* acquiring kvm->srcu for read here is still safe, just unnecessary. +*/ + idx =srcu_read_lock (&kvm->srcu); +while (likely ((*nr_entries_reset) <INT_MAX )) { -if (signal_pending (current)) +if (signal_pending (current)) { +srcu_read_unlock (&kvm->srcu, idx);return -EINTR ; + } entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];@@ -205,6 +213,8 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,if (mask)kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +srcu_read_unlock (&kvm->srcu, idx); +/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared * by the VCPU thread next time when it enters the guest. --
Quoted 2 lines
And unless there are other behaviors that are protected by
slots_lock(which is entirely possible), serializing the processing of each ring could be done via a
Yes, I am not the original author, but when I was working on it I don't remember anything relying on that. However still it's possible it can serialize some operations under the hood (which will be true side effect of using this lock..).
Quoted 69 lines
dedicated (for example only, the dedicated mutex could/should be per-vCPU, not global).
This diff in particular shows why I ordered and phrased the comment the way I did. The blurb about protecting memslot accesses is purely a friendly reminder to readers. The sole reason for an assert and comment is to call out the need for ordering.
virt/kvm/dirty_ring.c +12
-4@@ -102,6 +102,8 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)return smp_load_acquire (&gfn->flags) &KVM_DIRTY_GFN_F_RESET ; } +static DEFINE_MUTEX (per_ring_lock); +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) {@@ -121,18 +123,22 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; +int idx;/* * Ensure concurrent calls to KVM_RESET_DIRTY_RINGS are serialized, * e.g. so that KVM fully resets all entries processed by a given call -* before returning to userspace. Holding slots_lock also protects -* the various memslot accesses. +* before returning to userspace. */ -lockdep_assert_held (&kvm->slots_lock); +guard (mutex)(&per_ring_lock); + + idx =srcu_read_lock (&kvm->srcu);while (likely ((*nr_entries_reset) <INT_MAX )) { -if (signal_pending (current)) +if (signal_pending (current)) { +srcu_read_unlock (&kvm->srcu, idx);return -EINTR ; + } entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];@@ -205,6 +211,8 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,if (mask)kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +srcu_read_unlock (&kvm->srcu, idx); +/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared * by the VCPU thread next time when it enters the guest. virt/kvm/kvm_main.c +0
-2@@ -4908,16 +4908,12 @@ static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm)if (!kvm->dirty_ring_size)return -EINVAL ; -mutex_lock (&kvm->slots_lock); -kvm_for_each_vcpu (i, vcpu, kvm) {
virt/kvm/kvm_main.c static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm) · line 4923
Quoted 10 lines
virt/kvm/kvm_main.c +0
-3r =kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring, &cleared);if (r)break ; } -mutex_unlock (&kvm->slots_lock); -if (cleared)kvm_flush_remote_tlbs (kvm); --
I think we almost agree on each other, and I don't see anything controversial.
It's just that for this path using srcu may have slight risk of breaking what used to be serialized as you said. Said that, I'd be surprised if so.. even if aarch64 is normally even trickier and it now also supports the rings. So it's just that it seems unnecessary yet to switch to srcu, because we don't expect any concurrent writters anyway.
So totally no strong opinion on how the comment should be laid out in the last patch - please feel free to ignore my request. But I hope I stated the fact, that in the current code base the slots_lock is required to access memslots safely when rcu isn't around..
Thanks,
Signature · 1 line
Peter Xu
Aside from the nits,
hi Sean,
Do I need to rebase and repost [1]?
[1] https://lore.kernel.org/all/[email protected]
Thanks
Yan
Quoted 9 lines · Sean Christopherson
Fix issues with dirty ring harvesting where KVM doesn't bound the processing of entries in any way, which allows userspace to keep KVM in a tight loop indefinitely.
E.g.
struct kvm_dirty_gfn *dirty_gfns = vcpu_map_dirty_ring(vcpu);[...]
Applied to kvm-x86 dirty_ring, thanks!
[1/6] KVM: Bound the number of dirty ring entries in a single reset at INT_MAX
https://github.com/kvm-x86/linux/commit/530a8ba71b4c [2/6] KVM: Bail from the dirty ring reset flow if a signal is pending https://github.com/kvm-x86/linux/commit/49005a2a3d2a [3/6] KVM: Conditionally reschedule when resetting the dirty ring https://github.com/kvm-x86/linux/commit/1333c35c4eea [4/6] KVM: Check for empty mask of harvested dirty ring entries in caller https://github.com/kvm-x86/linux/commit/ee188dea1677 [5/6] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets https://github.com/kvm-x86/linux/commit/e46ad851150f [6/6] KVM: Assert that slots_lock is held when resetting per-vCPU dirty rings https://github.com/kvm-x86/linux/commit/614fb9d1479b