aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation/virtual/kvm/mmu.txt
diff options
context:
space:
mode:
authorDavid Matlack <dmatlack@google.com>2014-08-18 15:46:06 -0700
committerPaolo Bonzini <pbonzini@redhat.com>2014-09-03 10:03:41 +0200
commitee3d1570b58677885b4552bce8217fda7b226a68 (patch)
treebfc831cf5b242321553fbd6a702085a1916fd04c /Documentation/virtual/kvm/mmu.txt
parentKVM: do not bias the generation number in kvm_current_mmio_generation (diff)
downloadlinux-dev-ee3d1570b58677885b4552bce8217fda7b226a68.tar.xz
linux-dev-ee3d1570b58677885b4552bce8217fda7b226a68.zip
kvm: fix potentially corrupt mmio cache
vcpu exits and memslot mutations can run concurrently as long as the vcpu does not aquire the slots mutex. Thus it is theoretically possible for memslots to change underneath a vcpu that is handling an exit. If we increment the memslot generation number again after synchronize_srcu_expedited(), vcpus can safely cache memslot generation without maintaining a single rcu_dereference through an entire vm exit. And much of the x86/kvm code does not maintain a single rcu_dereference of the current memslots during each exit. We can prevent the following case: vcpu (CPU 0) | thread (CPU 1) --------------------------------------------+-------------------------- 1 vm exit | 2 srcu_read_unlock(&kvm->srcu) | 3 decide to cache something based on | old memslots | 4 | change memslots | (increments generation) 5 | synchronize_srcu(&kvm->srcu); 6 retrieve generation # from new memslots | 7 tag cache with new memslot generation | 8 srcu_read_unlock(&kvm->srcu) | ... | <action based on cache occurs even | though the caching decision was based | on the old memslots> | ... | <action *continues* to occur until next | memslot generation change, which may | be never> | | By incrementing the generation after synchronizing with kvm->srcu readers, we ensure that the generation retrieved in (6) will become invalid soon after (8). Keeping the existing increment is not strictly necessary, but we do keep it and just move it for consistency from update_memslots to install_new_memslots. It invalidates old cached MMIOs immediately, instead of having to wait for the end of synchronize_srcu_expedited, which makes the code more clearly correct in case CPU 1 is preempted right after synchronize_srcu() returns. To avoid halving the generation space in SPTEs, always presume that the low bit of the generation is zero when reconstructing a generation number out of an SPTE. This effectively disables MMIO caching in SPTEs during the call to synchronize_srcu_expedited. Using the low bit this way is somewhat like a seqcount---where the protected thing is a cache, and instead of retrying we can simply punt if we observe the low bit to be 1. Cc: stable@vger.kernel.org Signed-off-by: David Matlack <dmatlack@google.com> Reviewed-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com> Reviewed-by: David Matlack <dmatlack@google.com> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Diffstat (limited to 'Documentation/virtual/kvm/mmu.txt')
-rw-r--r--Documentation/virtual/kvm/mmu.txt14
1 files changed, 14 insertions, 0 deletions
diff --git a/Documentation/virtual/kvm/mmu.txt b/Documentation/virtual/kvm/mmu.txt
index 290894176142..53838d9c6295 100644
--- a/Documentation/virtual/kvm/mmu.txt
+++ b/Documentation/virtual/kvm/mmu.txt
@@ -425,6 +425,20 @@ fault through the slow path.
Since only 19 bits are used to store generation-number on mmio spte, all
pages are zapped when there is an overflow.
+Unfortunately, a single memory access might access kvm_memslots(kvm) multiple
+times, the last one happening when the generation number is retrieved and
+stored into the MMIO spte. Thus, the MMIO spte might be created based on
+out-of-date information, but with an up-to-date generation number.
+
+To avoid this, the generation number is incremented again after synchronize_srcu
+returns; thus, the low bit of kvm_memslots(kvm)->generation is only 1 during a
+memslot update, while some SRCU readers might be using the old copy. We do not
+want to use an MMIO sptes created with an odd generation number, and we can do
+this without losing a bit in the MMIO spte. The low bit of the generation
+is not stored in MMIO spte, and presumed zero when it is extracted out of the
+spte. If KVM is unlucky and creates an MMIO spte while the low bit is 1,
+the next access to the spte will always be a cache miss.
+
Further reading
===============