On 4/1/22 12:46 PM, Oliver Upton wrote:
It is possible to take a stage-2 permission fault on a page larger than
PAGE_SIZE. For example, when running a guest backed by 2M HugeTLB, KVM
eagerly maps at the largest possible block size. When dirty logging is
enabled on a memslot, KVM does *not* eagerly split these 2M stage-2
mappings and instead clears the write bit on the pte.
Since dirty logging is always performed at PAGE_SIZE granularity, KVM
lazily splits these 2M block mappings down to PAGE_SIZE in the stage-2
fault handler. This operation must be done under the write lock. Since
commit f783ef1c0e82 ("KVM: arm64: Add fast path to handle permission
relaxation during dirty logging"), the stage-2 fault handler
conditionally takes the read lock on permission faults with dirty
logging enabled. To that end, it is possible to split a 2M block mapping
while only holding the read lock.
The problem is demonstrated by running kvm_page_table_test with 2M
anonymous HugeTLB, which splats like so:
WARNING: CPU: 5 PID: 15276 at arch/arm64/kvm/hyp/pgtable.c:153 stage2_map_walk_leaf+0x124/0x158
[...]
Call trace:
stage2_map_walk_leaf+0x124/0x158
stage2_map_walker+0x5c/0xf0
__kvm_pgtable_walk+0x100/0x1d4
__kvm_pgtable_walk+0x140/0x1d4
__kvm_pgtable_walk+0x140/0x1d4
kvm_pgtable_walk+0xa0/0xf8
kvm_pgtable_stage2_map+0x15c/0x198
user_mem_abort+0x56c/0x838
kvm_handle_guest_abort+0x1fc/0x2a4
handle_exit+0xa4/0x120
kvm_arch_vcpu_ioctl_run+0x200/0x448
kvm_vcpu_ioctl+0x588/0x664
__arm64_sys_ioctl+0x9c/0xd4
invoke_syscall+0x4c/0x144
el0_svc_common+0xc4/0x190
do_el0_svc+0x30/0x8c
el0_svc+0x28/0xcc
el0t_64_sync_handler+0x84/0xe4
el0t_64_sync+0x1a4/0x1a8
Fix the issue by only acquiring the read lock if the guest faulted on a
PAGE_SIZE granule w/ dirty logging enabled. Add a WARN to catch locking
bugs in future changes.
Fixes: f783ef1c0e82 ("KVM: arm64: Add fast path to handle permission relaxation during dirty logging")
Cc: Jing Zhang <jingzhangos@xxxxxxxxxx>
Signed-off-by: Oliver Upton <oupton@xxxxxxxxxx>
Reviewed-by: Reiji Watanabe <reijiw@xxxxxxxxxx>
