]> git.baikalelectronics.ru Git - kernel.git/commitdiff
kvm: arm/arm64: Fix race in resetting stage2 PGD
authorSuzuki K Poulose <suzuki.poulose@arm.com>
Wed, 3 May 2017 14:17:51 +0000 (15:17 +0100)
committerChristoffer Dall <cdall@linaro.org>
Mon, 15 May 2017 10:05:25 +0000 (12:05 +0200)
In kvm_free_stage2_pgd() we check the stage2 PGD before holding
the lock and proceed to take the lock if it is valid. And we unmap
the page tables, followed by releasing the lock. We reset the PGD
only after dropping this lock, which could cause a race condition
where another thread waiting on or even holding the lock, could
potentially see that the PGD is still valid and proceed to perform
a stage2 operation and later encounter a NULL PGD.

[223090.242280] Unable to handle kernel NULL pointer dereference at
virtual address 00000040
[223090.262330] PC is at unmap_stage2_range+0x8c/0x428
[223090.262332] LR is at kvm_unmap_hva_handler+0x2c/0x3c
[223090.262531] Call trace:
[223090.262533] [<ffff0000080adb78>] unmap_stage2_range+0x8c/0x428
[223090.262535] [<ffff0000080adf40>] kvm_unmap_hva_handler+0x2c/0x3c
[223090.262537] [<ffff0000080ace2c>] handle_hva_to_gpa+0xb0/0x104
[223090.262539] [<ffff0000080af988>] kvm_unmap_hva+0x5c/0xbc
[223090.262543] [<ffff0000080a2478>]
kvm_mmu_notifier_invalidate_page+0x50/0x8c
[223090.262547] [<ffff0000082274f8>]
__mmu_notifier_invalidate_page+0x5c/0x84
[223090.262551] [<ffff00000820b700>] try_to_unmap_one+0x1d0/0x4a0
[223090.262553] [<ffff00000820c5c8>] rmap_walk+0x1cc/0x2e0
[223090.262555] [<ffff00000820c90c>] try_to_unmap+0x74/0xa4
[223090.262557] [<ffff000008230ce4>] migrate_pages+0x31c/0x5ac
[223090.262561] [<ffff0000081f869c>] compact_zone+0x3fc/0x7ac
[223090.262563] [<ffff0000081f8ae0>] compact_zone_order+0x94/0xb0
[223090.262564] [<ffff0000081f91c0>] try_to_compact_pages+0x108/0x290
[223090.262569] [<ffff0000081d5108>] __alloc_pages_direct_compact+0x70/0x1ac
[223090.262571] [<ffff0000081d64a0>] __alloc_pages_nodemask+0x434/0x9f4
[223090.262572] [<ffff0000082256f0>] alloc_pages_vma+0x230/0x254
[223090.262574] [<ffff000008235e5c>] do_huge_pmd_anonymous_page+0x114/0x538
[223090.262576] [<ffff000008201bec>] handle_mm_fault+0xd40/0x17a4
[223090.262577] [<ffff0000081fb324>] __get_user_pages+0x12c/0x36c
[223090.262578] [<ffff0000081fb804>] get_user_pages_unlocked+0xa4/0x1b8
[223090.262579] [<ffff0000080a3ce8>] __gfn_to_pfn_memslot+0x280/0x31c
[223090.262580] [<ffff0000080a3dd0>] gfn_to_pfn_prot+0x4c/0x5c
[223090.262582] [<ffff0000080af3f8>] kvm_handle_guest_abort+0x240/0x774
[223090.262584] [<ffff0000080b2bac>] handle_exit+0x11c/0x1ac
[223090.262586] [<ffff0000080ab99c>] kvm_arch_vcpu_ioctl_run+0x31c/0x648
[223090.262587] [<ffff0000080a1d78>] kvm_vcpu_ioctl+0x378/0x768
[223090.262590] [<ffff00000825df5c>] do_vfs_ioctl+0x324/0x5a4
[223090.262591] [<ffff00000825e26c>] SyS_ioctl+0x90/0xa4
[223090.262595] [<ffff000008085d84>] el0_svc_naked+0x38/0x3c

This patch moves the stage2 PGD manipulation under the lock.

Reported-by: Alexander Graf <agraf@suse.de>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krčmář <rkrcmar@redhat.com>
Reviewed-by: Christoffer Dall <cdall@linaro.org>
Reviewed-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Christoffer Dall <cdall@linaro.org>
virt/kvm/arm/mmu.c

index 313ee646480f4d88264664d7393fcb39969609ba..909a1a793b3103d161f7e9a8f839933e98c41f9a 100644 (file)
@@ -829,22 +829,22 @@ void stage2_unmap_vm(struct kvm *kvm)
  * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
  * underlying level-2 and level-3 tables before freeing the actual level-1 table
  * and setting the struct pointer to NULL.
- *
- * Note we don't need locking here as this is only called when the VM is
- * destroyed, which can only be done once.
  */
 void kvm_free_stage2_pgd(struct kvm *kvm)
 {
-       if (kvm->arch.pgd == NULL)
-               return;
+       void *pgd = NULL;
 
        spin_lock(&kvm->mmu_lock);
-       unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
+       if (kvm->arch.pgd) {
+               unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
+               pgd = kvm->arch.pgd;
+               kvm->arch.pgd = NULL;
+       }
        spin_unlock(&kvm->mmu_lock);
 
        /* Free the HW pgd, one page at a time */
-       free_pages_exact(kvm->arch.pgd, S2_PGD_SIZE);
-       kvm->arch.pgd = NULL;
+       if (pgd)
+               free_pages_exact(pgd, S2_PGD_SIZE);
 }
 
 static pud_t *stage2_get_pud(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,