diff options
author | Mauro Carvalho Chehab <mchehab+huawei@kernel.org> | 2020-02-10 07:02:59 +0100 |
---|---|---|
committer | Paolo Bonzini <pbonzini@redhat.com> | 2020-02-12 20:10:02 +0100 |
commit | 75e7fcdb4a6f394a6644ee1cfe193284945003b5 (patch) | |
tree | a81df8f1ca3e572f0297d5eb615d7eb5671d593d /Documentation | |
parent | 5a0af4806c25aff4b2f8d2e24d635840ec58a87b (diff) | |
download | linux-75e7fcdb4a6f394a6644ee1cfe193284945003b5.tar.bz2 |
docs: kvm: Convert locking.txt to ReST format
- Use document title and chapter markups;
- Add markups for literal blocks;
- use :field: for field descriptions;
- Add blank lines and adjust indentation.
Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Diffstat (limited to 'Documentation')
-rw-r--r-- | Documentation/virt/kvm/index.rst | 1 | ||||
-rw-r--r-- | Documentation/virt/kvm/locking.rst | 243 | ||||
-rw-r--r-- | Documentation/virt/kvm/locking.txt | 215 |
3 files changed, 244 insertions, 215 deletions
diff --git a/Documentation/virt/kvm/index.rst b/Documentation/virt/kvm/index.rst index ac83bc588f7e..9be8f53b729d 100644 --- a/Documentation/virt/kvm/index.rst +++ b/Documentation/virt/kvm/index.rst @@ -12,6 +12,7 @@ KVM cpuid halt-polling hypercalls + locking msr vcpu-requests diff --git a/Documentation/virt/kvm/locking.rst b/Documentation/virt/kvm/locking.rst new file mode 100644 index 000000000000..c02291beac3f --- /dev/null +++ b/Documentation/virt/kvm/locking.rst @@ -0,0 +1,243 @@ +.. SPDX-License-Identifier: GPL-2.0 + +================= +KVM Lock Overview +================= + +1. Acquisition Orders +--------------------- + +The acquisition orders for mutexes are as follows: + +- kvm->lock is taken outside vcpu->mutex + +- kvm->lock is taken outside kvm->slots_lock and kvm->irq_lock + +- kvm->slots_lock is taken outside kvm->irq_lock, though acquiring + them together is quite rare. + +On x86, vcpu->mutex is taken outside kvm->arch.hyperv.hv_lock. + +Everything else is a leaf: no other lock is taken inside the critical +sections. + +2. Exception +------------ + +Fast page fault: + +Fast page fault is the fast path which fixes the guest page fault out of +the mmu-lock on x86. Currently, the page fault can be fast in one of the +following two cases: + +1. Access Tracking: The SPTE is not present, but it is marked for access + tracking i.e. the SPTE_SPECIAL_MASK is set. That means we need to + restore the saved R/X bits. This is described in more detail later below. + +2. Write-Protection: The SPTE is present and the fault is + caused by write-protect. That means we just need to change the W bit of + the spte. + +What we use to avoid all the race is the SPTE_HOST_WRITEABLE bit and +SPTE_MMU_WRITEABLE bit on the spte: + +- SPTE_HOST_WRITEABLE means the gfn is writable on host. +- SPTE_MMU_WRITEABLE means the gfn is writable on mmu. The bit is set when + the gfn is writable on guest mmu and it is not write-protected by shadow + page write-protection. + +On fast page fault path, we will use cmpxchg to atomically set the spte W +bit if spte.SPTE_HOST_WRITEABLE = 1 and spte.SPTE_WRITE_PROTECT = 1, or +restore the saved R/X bits if VMX_EPT_TRACK_ACCESS mask is set, or both. This +is safe because whenever changing these bits can be detected by cmpxchg. + +But we need carefully check these cases: + +1) The mapping from gfn to pfn + +The mapping from gfn to pfn may be changed since we can only ensure the pfn +is not changed during cmpxchg. This is a ABA problem, for example, below case +will happen: + ++------------------------------------------------------------------------+ +| At the beginning:: | +| | +| gpte = gfn1 | +| gfn1 is mapped to pfn1 on host | +| spte is the shadow page table entry corresponding with gpte and | +| spte = pfn1 | ++------------------------------------------------------------------------+ +| On fast page fault path: | ++------------------------------------+-----------------------------------+ +| CPU 0: | CPU 1: | ++------------------------------------+-----------------------------------+ +| :: | | +| | | +| old_spte = *spte; | | ++------------------------------------+-----------------------------------+ +| | pfn1 is swapped out:: | +| | | +| | spte = 0; | +| | | +| | pfn1 is re-alloced for gfn2. | +| | | +| | gpte is changed to point to | +| | gfn2 by the guest:: | +| | | +| | spte = pfn1; | ++------------------------------------+-----------------------------------+ +| :: | +| | +| if (cmpxchg(spte, old_spte, old_spte+W) | +| mark_page_dirty(vcpu->kvm, gfn1) | +| OOPS!!! | ++------------------------------------------------------------------------+ + +We dirty-log for gfn1, that means gfn2 is lost in dirty-bitmap. + +For direct sp, we can easily avoid it since the spte of direct sp is fixed +to gfn. For indirect sp, before we do cmpxchg, we call gfn_to_pfn_atomic() +to pin gfn to pfn, because after gfn_to_pfn_atomic(): + +- We have held the refcount of pfn that means the pfn can not be freed and + be reused for another gfn. +- The pfn is writable that means it can not be shared between different gfns + by KSM. + +Then, we can ensure the dirty bitmaps is correctly set for a gfn. + +Currently, to simplify the whole things, we disable fast page fault for +indirect shadow page. + +2) Dirty bit tracking + +In the origin code, the spte can be fast updated (non-atomically) if the +spte is read-only and the Accessed bit has already been set since the +Accessed bit and Dirty bit can not be lost. + +But it is not true after fast page fault since the spte can be marked +writable between reading spte and updating spte. Like below case: + ++------------------------------------------------------------------------+ +| At the beginning:: | +| | +| spte.W = 0 | +| spte.Accessed = 1 | ++------------------------------------+-----------------------------------+ +| CPU 0: | CPU 1: | ++------------------------------------+-----------------------------------+ +| In mmu_spte_clear_track_bits():: | | +| | | +| old_spte = *spte; | | +| | | +| | | +| /* 'if' condition is satisfied. */| | +| if (old_spte.Accessed == 1 && | | +| old_spte.W == 0) | | +| spte = 0ull; | | ++------------------------------------+-----------------------------------+ +| | on fast page fault path:: | +| | | +| | spte.W = 1 | +| | | +| | memory write on the spte:: | +| | | +| | spte.Dirty = 1 | ++------------------------------------+-----------------------------------+ +| :: | | +| | | +| else | | +| old_spte = xchg(spte, 0ull) | | +| if (old_spte.Accessed == 1) | | +| kvm_set_pfn_accessed(spte.pfn);| | +| if (old_spte.Dirty == 1) | | +| kvm_set_pfn_dirty(spte.pfn); | | +| OOPS!!! | | ++------------------------------------+-----------------------------------+ + +The Dirty bit is lost in this case. + +In order to avoid this kind of issue, we always treat the spte as "volatile" +if it can be updated out of mmu-lock, see spte_has_volatile_bits(), it means, +the spte is always atomically updated in this case. + +3) flush tlbs due to spte updated + +If the spte is updated from writable to readonly, we should flush all TLBs, +otherwise rmap_write_protect will find a read-only spte, even though the +writable spte might be cached on a CPU's TLB. + +As mentioned before, the spte can be updated to writable out of mmu-lock on +fast page fault path, in order to easily audit the path, we see if TLBs need +be flushed caused by this reason in mmu_spte_update() since this is a common +function to update spte (present -> present). + +Since the spte is "volatile" if it can be updated out of mmu-lock, we always +atomically update the spte, the race caused by fast page fault can be avoided, +See the comments in spte_has_volatile_bits() and mmu_spte_update(). + +Lockless Access Tracking: + +This is used for Intel CPUs that are using EPT but do not support the EPT A/D +bits. In this case, when the KVM MMU notifier is called to track accesses to a +page (via kvm_mmu_notifier_clear_flush_young), it marks the PTE as not-present +by clearing the RWX bits in the PTE and storing the original R & X bits in +some unused/ignored bits. In addition, the SPTE_SPECIAL_MASK is also set on the +PTE (using the ignored bit 62). When the VM tries to access the page later on, +a fault is generated and the fast page fault mechanism described above is used +to atomically restore the PTE to a Present state. The W bit is not saved when +the PTE is marked for access tracking and during restoration to the Present +state, the W bit is set depending on whether or not it was a write access. If +it wasn't, then the W bit will remain clear until a write access happens, at +which time it will be set using the Dirty tracking mechanism described above. + +3. Reference +------------ + +:Name: kvm_lock +:Type: mutex +:Arch: any +:Protects: - vm_list + +:Name: kvm_count_lock +:Type: raw_spinlock_t +:Arch: any +:Protects: - hardware virtualization enable/disable +:Comment: 'raw' because hardware enabling/disabling must be atomic /wrt + migration. + +:Name: kvm_arch::tsc_write_lock +:Type: raw_spinlock +:Arch: x86 +:Protects: - kvm_arch::{last_tsc_write,last_tsc_nsec,last_tsc_offset} + - tsc offset in vmcb +:Comment: 'raw' because updating the tsc offsets must not be preempted. + +:Name: kvm->mmu_lock +:Type: spinlock_t +:Arch: any +:Protects: -shadow page/shadow tlb entry +:Comment: it is a spinlock since it is used in mmu notifier. + +:Name: kvm->srcu +:Type: srcu lock +:Arch: any +:Protects: - kvm->memslots + - kvm->buses +:Comment: The srcu read lock must be held while accessing memslots (e.g. + when using gfn_to_* functions) and while accessing in-kernel + MMIO/PIO address->device structure mapping (kvm->buses). + The srcu index can be stored in kvm_vcpu->srcu_idx per vcpu + if it is needed by multiple functions. + +:Name: blocked_vcpu_on_cpu_lock +:Type: spinlock_t +:Arch: x86 +:Protects: blocked_vcpu_on_cpu +:Comment: This is a per-CPU lock and it is used for VT-d posted-interrupts. + When VT-d posted-interrupts is supported and the VM has assigned + devices, we put the blocked vCPU on the list blocked_vcpu_on_cpu + protected by blocked_vcpu_on_cpu_lock, when VT-d hardware issues + wakeup notification event since external interrupts from the + assigned devices happens, we will find the vCPU on the list to + wakeup. diff --git a/Documentation/virt/kvm/locking.txt b/Documentation/virt/kvm/locking.txt deleted file mode 100644 index 635cd6eaf714..000000000000 --- a/Documentation/virt/kvm/locking.txt +++ /dev/null @@ -1,215 +0,0 @@ -KVM Lock Overview -================= - -1. Acquisition Orders ---------------------- - -The acquisition orders for mutexes are as follows: - -- kvm->lock is taken outside vcpu->mutex - -- kvm->lock is taken outside kvm->slots_lock and kvm->irq_lock - -- kvm->slots_lock is taken outside kvm->irq_lock, though acquiring - them together is quite rare. - -On x86, vcpu->mutex is taken outside kvm->arch.hyperv.hv_lock. - -Everything else is a leaf: no other lock is taken inside the critical -sections. - -2: Exception ------------- - -Fast page fault: - -Fast page fault is the fast path which fixes the guest page fault out of -the mmu-lock on x86. Currently, the page fault can be fast in one of the -following two cases: - -1. Access Tracking: The SPTE is not present, but it is marked for access -tracking i.e. the SPTE_SPECIAL_MASK is set. That means we need to -restore the saved R/X bits. This is described in more detail later below. - -2. Write-Protection: The SPTE is present and the fault is -caused by write-protect. That means we just need to change the W bit of the -spte. - -What we use to avoid all the race is the SPTE_HOST_WRITEABLE bit and -SPTE_MMU_WRITEABLE bit on the spte: -- SPTE_HOST_WRITEABLE means the gfn is writable on host. -- SPTE_MMU_WRITEABLE means the gfn is writable on mmu. The bit is set when - the gfn is writable on guest mmu and it is not write-protected by shadow - page write-protection. - -On fast page fault path, we will use cmpxchg to atomically set the spte W -bit if spte.SPTE_HOST_WRITEABLE = 1 and spte.SPTE_WRITE_PROTECT = 1, or -restore the saved R/X bits if VMX_EPT_TRACK_ACCESS mask is set, or both. This -is safe because whenever changing these bits can be detected by cmpxchg. - -But we need carefully check these cases: -1): The mapping from gfn to pfn -The mapping from gfn to pfn may be changed since we can only ensure the pfn -is not changed during cmpxchg. This is a ABA problem, for example, below case -will happen: - -At the beginning: -gpte = gfn1 -gfn1 is mapped to pfn1 on host -spte is the shadow page table entry corresponding with gpte and -spte = pfn1 - - VCPU 0 VCPU0 -on fast page fault path: - - old_spte = *spte; - pfn1 is swapped out: - spte = 0; - - pfn1 is re-alloced for gfn2. - - gpte is changed to point to - gfn2 by the guest: - spte = pfn1; - - if (cmpxchg(spte, old_spte, old_spte+W) - mark_page_dirty(vcpu->kvm, gfn1) - OOPS!!! - -We dirty-log for gfn1, that means gfn2 is lost in dirty-bitmap. - -For direct sp, we can easily avoid it since the spte of direct sp is fixed -to gfn. For indirect sp, before we do cmpxchg, we call gfn_to_pfn_atomic() -to pin gfn to pfn, because after gfn_to_pfn_atomic(): -- We have held the refcount of pfn that means the pfn can not be freed and - be reused for another gfn. -- The pfn is writable that means it can not be shared between different gfns - by KSM. - -Then, we can ensure the dirty bitmaps is correctly set for a gfn. - -Currently, to simplify the whole things, we disable fast page fault for -indirect shadow page. - -2): Dirty bit tracking -In the origin code, the spte can be fast updated (non-atomically) if the -spte is read-only and the Accessed bit has already been set since the -Accessed bit and Dirty bit can not be lost. - -But it is not true after fast page fault since the spte can be marked -writable between reading spte and updating spte. Like below case: - -At the beginning: -spte.W = 0 -spte.Accessed = 1 - - VCPU 0 VCPU0 -In mmu_spte_clear_track_bits(): - - old_spte = *spte; - - /* 'if' condition is satisfied. */ - if (old_spte.Accessed == 1 && - old_spte.W == 0) - spte = 0ull; - on fast page fault path: - spte.W = 1 - memory write on the spte: - spte.Dirty = 1 - - - else - old_spte = xchg(spte, 0ull) - - - if (old_spte.Accessed == 1) - kvm_set_pfn_accessed(spte.pfn); - if (old_spte.Dirty == 1) - kvm_set_pfn_dirty(spte.pfn); - OOPS!!! - -The Dirty bit is lost in this case. - -In order to avoid this kind of issue, we always treat the spte as "volatile" -if it can be updated out of mmu-lock, see spte_has_volatile_bits(), it means, -the spte is always atomically updated in this case. - -3): flush tlbs due to spte updated -If the spte is updated from writable to readonly, we should flush all TLBs, -otherwise rmap_write_protect will find a read-only spte, even though the -writable spte might be cached on a CPU's TLB. - -As mentioned before, the spte can be updated to writable out of mmu-lock on -fast page fault path, in order to easily audit the path, we see if TLBs need -be flushed caused by this reason in mmu_spte_update() since this is a common -function to update spte (present -> present). - -Since the spte is "volatile" if it can be updated out of mmu-lock, we always -atomically update the spte, the race caused by fast page fault can be avoided, -See the comments in spte_has_volatile_bits() and mmu_spte_update(). - -Lockless Access Tracking: - -This is used for Intel CPUs that are using EPT but do not support the EPT A/D -bits. In this case, when the KVM MMU notifier is called to track accesses to a -page (via kvm_mmu_notifier_clear_flush_young), it marks the PTE as not-present -by clearing the RWX bits in the PTE and storing the original R & X bits in -some unused/ignored bits. In addition, the SPTE_SPECIAL_MASK is also set on the -PTE (using the ignored bit 62). When the VM tries to access the page later on, -a fault is generated and the fast page fault mechanism described above is used -to atomically restore the PTE to a Present state. The W bit is not saved when -the PTE is marked for access tracking and during restoration to the Present -state, the W bit is set depending on whether or not it was a write access. If -it wasn't, then the W bit will remain clear until a write access happens, at -which time it will be set using the Dirty tracking mechanism described above. - -3. Reference ------------- - -Name: kvm_lock -Type: mutex -Arch: any -Protects: - vm_list - -Name: kvm_count_lock -Type: raw_spinlock_t -Arch: any -Protects: - hardware virtualization enable/disable -Comment: 'raw' because hardware enabling/disabling must be atomic /wrt - migration. - -Name: kvm_arch::tsc_write_lock -Type: raw_spinlock -Arch: x86 -Protects: - kvm_arch::{last_tsc_write,last_tsc_nsec,last_tsc_offset} - - tsc offset in vmcb -Comment: 'raw' because updating the tsc offsets must not be preempted. - -Name: kvm->mmu_lock -Type: spinlock_t -Arch: any -Protects: -shadow page/shadow tlb entry -Comment: it is a spinlock since it is used in mmu notifier. - -Name: kvm->srcu -Type: srcu lock -Arch: any -Protects: - kvm->memslots - - kvm->buses -Comment: The srcu read lock must be held while accessing memslots (e.g. - when using gfn_to_* functions) and while accessing in-kernel - MMIO/PIO address->device structure mapping (kvm->buses). - The srcu index can be stored in kvm_vcpu->srcu_idx per vcpu - if it is needed by multiple functions. - -Name: blocked_vcpu_on_cpu_lock -Type: spinlock_t -Arch: x86 -Protects: blocked_vcpu_on_cpu -Comment: This is a per-CPU lock and it is used for VT-d posted-interrupts. - When VT-d posted-interrupts is supported and the VM has assigned - devices, we put the blocked vCPU on the list blocked_vcpu_on_cpu - protected by blocked_vcpu_on_cpu_lock, when VT-d hardware issues - wakeup notification event since external interrupts from the - assigned devices happens, we will find the vCPU on the list to - wakeup. |