diff options
author | Babu Moger <babu.moger@amd.com> | 2024-06-12 09:41:51 -0500 |
---|---|---|
committer | Sean Christopherson <seanjc@google.com> | 2024-06-18 08:51:03 -0700 |
commit | c3f3edf73a8f854f8766a69d2734198a58762e33 (patch) | |
tree | 51ebb48bffcf7aecd9d4ea91d35d865ddcb8a2ab /virt/kvm/kvm_main.c | |
parent | 49f683b41f28918df3e51ddc0d928cb2e934ccdb (diff) |
KVM: Stop processing *all* memslots when "null" mmu_notifier handler is found
Bail from outer address space loop, not just the inner memslot loop, when
a "null" handler is encountered by __kvm_handle_hva_range(), which is the
intended behavior. On x86, which has multiple address spaces thanks to
SMM emulation, breaking from just the memslot loop results in undefined
behavior due to assigning the non-existent return value from kvm_null_fn()
to a bool.
In practice, the bug is benign as kvm_mmu_notifier_invalidate_range_end()
is the only caller that passes handler=kvm_null_fn, and it doesn't set
flush_on_ret, i.e. assigning garbage to r.ret is ultimately ignored. And
for most configuration the compiler elides the entire sequence, i.e. there
is no undefined behavior at runtime.
------------[ cut here ]------------
UBSAN: invalid-load in arch/x86/kvm/../../../virt/kvm/kvm_main.c:655:10
load of value 160 is not a valid value for type '_Bool'
CPU: 370 PID: 8246 Comm: CPU 0/KVM Not tainted 6.8.2-amdsos-build58-ubuntu-22.04+ #1
Hardware name: AMD Corporation Sh54p/Sh54p, BIOS WPC4429N 04/25/2024
Call Trace:
<TASK>
dump_stack_lvl+0x48/0x60
ubsan_epilogue+0x5/0x30
__ubsan_handle_load_invalid_value+0x79/0x80
kvm_mmu_notifier_invalidate_range_end.cold+0x18/0x4f [kvm]
__mmu_notifier_invalidate_range_end+0x63/0xe0
__split_huge_pmd+0x367/0xfc0
do_huge_pmd_wp_page+0x1cc/0x380
__handle_mm_fault+0x8ee/0xe50
handle_mm_fault+0xe4/0x4a0
__get_user_pages+0x190/0x840
get_user_pages_unlocked+0xe0/0x590
hva_to_pfn+0x114/0x550 [kvm]
kvm_faultin_pfn+0xed/0x5b0 [kvm]
kvm_tdp_page_fault+0x123/0x170 [kvm]
kvm_mmu_page_fault+0x244/0xaa0 [kvm]
vcpu_enter_guest+0x592/0x1070 [kvm]
kvm_arch_vcpu_ioctl_run+0x145/0x8a0 [kvm]
kvm_vcpu_ioctl+0x288/0x6d0 [kvm]
__x64_sys_ioctl+0x8f/0xd0
do_syscall_64+0x77/0x120
entry_SYSCALL_64_after_hwframe+0x6e/0x76
</TASK>
---[ end trace ]---
Fixes: 071064f14d87 ("KVM: Don't take mmu_lock for range invalidation unless necessary")
Signed-off-by: Babu Moger <babu.moger@amd.com>
Link: https://lore.kernel.org/r/b8723d39903b64c241c50f5513f804390c7b5eec.1718203311.git.babu.moger@amd.com
[sean: massage changelog]
Signed-off-by: Sean Christopherson <seanjc@google.com>
Diffstat (limited to 'virt/kvm/kvm_main.c')
-rw-r--r-- | virt/kvm/kvm_main.c | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c index 843aa68cbcd0..68be4be5d846 100644 --- a/virt/kvm/kvm_main.c +++ b/virt/kvm/kvm_main.c @@ -651,7 +651,7 @@ static __always_inline kvm_mn_ret_t __kvm_handle_hva_range(struct kvm *kvm, range->on_lock(kvm); if (IS_KVM_NULL_FN(range->handler)) - break; + goto mmu_unlock; } r.ret |= range->handler(kvm, &gfn_range); } @@ -660,6 +660,7 @@ static __always_inline kvm_mn_ret_t __kvm_handle_hva_range(struct kvm *kvm, if (range->flush_on_ret && r.ret) kvm_flush_remote_tlbs(kvm); +mmu_unlock: if (r.found_memslot) KVM_MMU_UNLOCK(kvm); |