Re: [PATCH v5 01/19] KVM: x86: Allocate new rmap and large page tracking when moving memslot

[Peter Xu <peterx@redhat.com>]

On Wed, Feb 05, 2020 at 03:55:33PM -0800, Sean Christopherson wrote:
> On Wed, Feb 05, 2020 at 04:49:52PM -0500, Peter Xu wrote:
> > On Tue, Jan 21, 2020 at 02:31:39PM -0800, Sean Christopherson wrote:
> > > Reallocate a rmap array and recalcuate large page compatibility when
> > > moving an existing memslot to correctly handle the alignment properties
> > > of the new memslot.  The number of rmap entries required at each level
> > > is dependent on the alignment of the memslot's base gfn with respect to
> > > that level, e.g. moving a large-page aligned memslot so that it becomes
> > > unaligned will increase the number of rmap entries needed at the now
> > > unaligned level.
> > > 
> > > Not updating the rmap array is the most obvious bug, as KVM accesses
> > > garbage data beyond the end of the rmap.  KVM interprets the bad data as
> > > pointers, leading to non-canonical #GPs, unexpected #PFs, etc...
> > > 
> > >   general protection fault: 0000 [#1] SMP
> > >   CPU: 0 PID: 1909 Comm: move_memory_reg Not tainted 5.4.0-rc7+ #139
> > >   Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
> > >   RIP: 0010:rmap_get_first+0x37/0x50 [kvm]
> > >   Code: <48> 8b 3b 48 85 ff 74 ec e8 6c f4 ff ff 85 c0 74 e3 48 89 d8 5b c3
> > >   RSP: 0018:ffffc9000021bbc8 EFLAGS: 00010246
> > >   RAX: ffff00617461642e RBX: ffff00617461642e RCX: 0000000000000012
> > >   RDX: ffff88827400f568 RSI: ffffc9000021bbe0 RDI: ffff88827400f570
> > >   RBP: 0010000000000000 R08: ffffc9000021bd00 R09: ffffc9000021bda8
> > >   R10: ffffc9000021bc48 R11: 0000000000000000 R12: 0030000000000000
> > >   R13: 0000000000000000 R14: ffff88827427d700 R15: ffffc9000021bce8
> > >   FS:  00007f7eda014700(0000) GS:ffff888277a00000(0000) knlGS:0000000000000000
> > >   CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > >   CR2: 00007f7ed9216ff8 CR3: 0000000274391003 CR4: 0000000000162eb0
> > >   Call Trace:
> > >    kvm_mmu_slot_set_dirty+0xa1/0x150 [kvm]
> > >    __kvm_set_memory_region.part.64+0x559/0x960 [kvm]
> > >    kvm_set_memory_region+0x45/0x60 [kvm]
> > >    kvm_vm_ioctl+0x30f/0x920 [kvm]
> > >    do_vfs_ioctl+0xa1/0x620
> > >    ksys_ioctl+0x66/0x70
> > >    __x64_sys_ioctl+0x16/0x20
> > >    do_syscall_64+0x4c/0x170
> > >    entry_SYSCALL_64_after_hwframe+0x44/0xa9
> > >   RIP: 0033:0x7f7ed9911f47
> > >   Code: <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 21 6f 2c 00 f7 d8 64 89 01 48
> > >   RSP: 002b:00007ffc00937498 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
> > >   RAX: ffffffffffffffda RBX: 0000000001ab0010 RCX: 00007f7ed9911f47
> > >   RDX: 0000000001ab1350 RSI: 000000004020ae46 RDI: 0000000000000004
> > >   RBP: 000000000000000a R08: 0000000000000000 R09: 00007f7ed9214700
> > >   R10: 00007f7ed92149d0 R11: 0000000000000246 R12: 00000000bffff000
> > >   R13: 0000000000000003 R14: 00007f7ed9215000 R15: 0000000000000000
> > >   Modules linked in: kvm_intel kvm irqbypass
> > >   ---[ end trace 0c5f570b3358ca89 ]---
> > > 
> > > The disallow_lpage tracking is more subtle.  Failure to update results
> > > in KVM creating large pages when it shouldn't, either due to stale data
> > > or again due to indexing beyond the end of the metadata arrays, which
> > > can lead to memory corruption and/or leaking data to guest/userspace.
> > > 
> > > Note, the arrays for the old memslot are freed by the unconditional call
> > > to kvm_free_memslot() in __kvm_set_memory_region().
> > 
> > If __kvm_set_memory_region() failed, I think the old memslot will be
> > kept and the new memslot will be freed instead?
> 
> This is referring to a successful MOVE operation to note that zeroing @arch
> in kvm_arch_create_memslot() won't leak memory.
> 
> > > 
> > > Fixes: 05da45583de9b ("KVM: MMU: large page support")
> > > Cc: stable@vger.kernel.org
> > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
> > > ---
> > >  arch/x86/kvm/x86.c | 11 +++++++++++
> > >  1 file changed, 11 insertions(+)
> > > 
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 4c30ebe74e5d..1953c71c52f2 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -9793,6 +9793,13 @@ int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
> > >  {
> > >  	int i;
> > >  
> > > +	/*
> > > +	 * Clear out the previous array pointers for the KVM_MR_MOVE case.  The
> > > +	 * old arrays will be freed by __kvm_set_memory_region() if installing
> > > +	 * the new memslot is successful.
> > > +	 */
> > > +	memset(&slot->arch, 0, sizeof(slot->arch));
> > 
> > I actually gave r-b on this patch but it was lost... And then when I
> > read it again I start to confuse on why we need to set these to zeros.
> > Even if they're not zeros, iiuc kvm_free_memslot() will compare each
> > of the array pointer and it will only free the changed pointers, then
> > it looks fine even without zeroing?
> 
> It's for the failure path, the out_free label, which blindy calls kvfree()
> and relies on un-allocated pointers being NULL.  If @arch isn't zeroed, the
> failure path will free metadata from the previous memslot.

IMHO it won't, because kvm_free_memslot() will only free metadata if
the pointer changed.  So:

  - For succeeded kvcalloc(), the pointer will change in the new slot,
    so kvm_free_memslot() will free it,

  - For failed kvcalloc(), the pointer will be NULL, so
    kvm_free_memslot() will skip it,

  - For untouched pointer, it'll be the same as the old, so
    kvm_free_memslot() will skip it as well.

> 
> > > +
> > >  	for (i = 0; i < KVM_NR_PAGE_SIZES; ++i) {
> > >  		struct kvm_lpage_info *linfo;
> > >  		unsigned long ugfn;
> > > @@ -9867,6 +9874,10 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
> > >  				const struct kvm_userspace_memory_region *mem,
> > >  				enum kvm_mr_change change)
> > >  {
> > > +	if (change == KVM_MR_MOVE)
> > > +		return kvm_arch_create_memslot(kvm, memslot,
> > > +					       mem->memory_size >> PAGE_SHIFT);
> > > +
> > 
> > Instead of calling kvm_arch_create_memslot() explicitly again here,
> > can it be replaced by below?
> > 
> > diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> > index 72b45f491692..85a7b02fd752 100644
> > --- a/virt/kvm/kvm_main.c
> > +++ b/virt/kvm/kvm_main.c
> > @@ -1144,7 +1144,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
> >                 new.dirty_bitmap = NULL;
> >  
> >         r = -ENOMEM;
> > -       if (change == KVM_MR_CREATE) {
> > +       if (change == KVM_MR_CREATE || change == KVM_MR_MOVE) {
> >                 new.userspace_addr = mem->userspace_addr;
> >  
> >                 if (kvm_arch_create_memslot(kvm, &new, npages))
> 
> No, because other architectures don't need to re-allocate new metadata on
> MOVE and rely on __kvm_set_memory_region() to copy @arch from old to new,
> e.g. see kvmppc_core_create_memslot_hv().

Yes it's only required in x86, but iiuc it also will still work for
ppc?  Say, in that case ppc won't copy @arch from old to new, and
kvmppc_core_free_memslot_hv() will free the old, however it should
still work.

> 
> That being said, that's effectively what the x86 code looks like once
> kvm_arch_create_memslot() gets merged into kvm_arch_prepare_memory_region().

Right.  I don't have strong opinion on this, but if my above analysis
is correct, it's still slightly cleaner, imho, to have this patch as a
oneliner as I provided, then in the other patch move the whole
CREATE|MOVE into prepare_memory_region().  The final code should be
the same.

-- 
Peter Xu