On Tue, Jan 30, 2018 at 12:49:21PM -0800, Eric Biggers wrote:
> On Fri, Jan 19, 2018 at 10:57:16AM -0800, Eric Biggers wrote:
> > +Cc alex.williamson@xxxxxxxxxx
> >
> > On Fri, Jan 19, 2018 at 05:03:47PM +0800, Wanpeng Li wrote:
> > > 2018-01-19 17:01 GMT+08:00 Wanpeng Li <kernellwp@xxxxxxxxx>:
> > > > 2018-01-19 16:18 GMT+08:00 Eric Biggers <ebiggers3@xxxxxxxxx>:
> > > >> From: Eric Biggers <ebiggers@xxxxxxxxxx>
> > > >>
> > > >> Memslots must not overlap in guest physical memory, since otherwise some
> > > >> guest physical addresses will not uniquely map to a memslot. Yet, the
> > > >> overlap check in __kvm_set_memory_region() allows a memslot that
> > > >> overlaps one of the "private" memslots, e.g. the memslot reserved for
> > > >> the TSS on x86.
> > > >>
> > > >> This seems to be a very old bug that was introduced years ago when
> > > >> private memory slots were first added. It seems that later refactoring
> > > >> incorrectly assumed this bug was intentional and preserved it.
> > > >>
> > > >> Fix it by removing the loophole for private memslots, so we just check
> > > >> for overlap against all memslots.
> > > >>
> > > >> This bug was found by syzkaller, which used a memslot overlap to make
> > > >> pte_list_remove() be called for the wrong memslot, hitting a BUG():
> > > >>
> > > >> pte_list_remove: 000000007185ed42 0->BUG
> > > >> kernel BUG at arch/x86/kvm/mmu.c:1209!
> > > >> [...]
> > > >> RIP: 0010:pte_list_remove+0x107/0x110 arch/x86/kvm/mmu.c:1208
> > > >> [...]
> > > >> Call Trace:
> > > >> mmu_page_zap_pte+0x7e/0xd0 arch/x86/kvm/mmu.c:2499
> > > >> kvm_mmu_page_unlink_children arch/x86/kvm/mmu.c:2521 [inline]
> > > >> kvm_mmu_prepare_zap_page+0x4f/0x340 arch/x86/kvm/mmu.c:2565
> > > >> kvm_zap_obsolete_pages arch/x86/kvm/mmu.c:5348 [inline]
> > > >> kvm_mmu_invalidate_zap_all_pages+0xa6/0x100 arch/x86/kvm/mmu.c:5389
> > > >> kvm_mmu_notifier_release+0x4f/0x80 arch/x86/kvm/../../../virt/kvm/kvm_main.c:468
> > > >> __mmu_notifier_release+0x63/0x100 mm/mmu_notifier.c:75
> > > >> mmu_notifier_release include/linux/mmu_notifier.h:244 [inline]
> > > >> exit_mmap+0x160/0x170 mm/mmap.c:3009
> > > >> __mmput kernel/fork.c:966 [inline]
> > > >> mmput+0x44/0xd0 kernel/fork.c:987
> > > >> exit_mm kernel/exit.c:544 [inline]
> > > >> do_exit+0x24a/0xb50 kernel/exit.c:856
> > > >> do_group_exit+0x34/0xb0 kernel/exit.c:972
> > > >> SYSC_exit_group kernel/exit.c:983 [inline]
> > > >> SyS_exit_group+0xb/0x10 kernel/exit.c:981
> > > >> entry_SYSCALL_64_fastpath+0x1e/0x8b
> > > >>
> > > >> Reproducer:
> > > >>
> > > >> #include <fcntl.h>
> > > >> #include <linux/kvm.h>
> > > >> #include <sys/ioctl.h>
> > > >>
> > > >> int main()
> > > >> {
> > > >> static char buf[4096*3] __attribute__((aligned(4096)));
> > > >> int kvm, vm, cpu;
> > > >> struct kvm_mp_state mp_state = { KVM_MP_STATE_SIPI_RECEIVED };
> > > >> struct kvm_userspace_memory_region memreg = {
> > > >> .memory_size = sizeof(buf),
> > > >> .userspace_addr = (__u64)buf,
> > > >> };
> > > >>
> > > >> kvm = open("/dev/kvm", O_RDWR);
> > > >> vm = ioctl(kvm, KVM_CREATE_VM, 0);
> > > >> ioctl(vm, KVM_CREATE_IRQCHIP);
> > > >> cpu = ioctl(vm, KVM_CREATE_VCPU, 0);
> > > >> ioctl(cpu, KVM_SET_MP_STATE, &mp_state);
> > > >> ioctl(vm, KVM_SET_TSS_ADDR, 0);
> > > >> ioctl(cpu, KVM_RUN, 0);
> > > >> ioctl(vm, KVM_SET_USER_MEMORY_REGION, &memreg);
> > > >> }
> > > >>
> > > >> Reported-by: syzbot <syzkaller@xxxxxxxxxxxxxxxx>
> > > >> Fixes: e0d62c7f4860 ("KVM: Add kernel-internal memory slots")
> > > >> Cc: <stable@xxxxxxxxxxxxxxx> # v2.6.25+
> > > >> Signed-off-by: Eric Biggers <ebiggers@xxxxxxxxxx>
> > > >
> > > > Please refer to this one. https://patchwork.kernel.org/patch/9645377/
> > >
> > > https://lkml.org/lkml/2017/3/27/57
> > >
> >
> > Ah, so this was reported before, and you sent the same fix. Well, it was never
> > applied, so the bug is still there, and anyone who can use /dev/kvm can trigger
> > it. So one of these patches needs to be applied, unless there is a better fix.
> >
> > I don't agree with the "Fixes:" line in your version of the patch. The bug was
> > actually there prior to 5419369ed, which might explain why that commit seemed to
> > preserve the behavior intentionally. (Note that KVM_MEMORY_SLOTS did not
> > include the private memory slots; it was later renamed to KVM_USER_MEM_SLOTS.)
> >
> > Eric
>
> Ping. Paolo or Radim, can you please consider applying one of these patches?
>
> Eric
Ping.
