Re: [PATCH v3 1/2] mm/ksm: fix flag-dropping behavior in ksm_madvise

[Jakub Acs <acsjakub@amazon.de>]

On Thu, Nov 06, 2025 at 11:39:28AM +0100, Vlastimil Babka wrote:
> On 10/1/25 11:03, Jakub Acs wrote:
> > syzkaller discovered the following crash: (kernel BUG)
> > 
> > [   44.607039] ------------[ cut here ]------------
> > [   44.607422] kernel BUG at mm/userfaultfd.c:2067!
> > [   44.608148] Oops: invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC KASAN NOPTI
> > [   44.608814] CPU: 1 UID: 0 PID: 2475 Comm: reproducer Not tainted 6.16.0-rc6 #1 PREEMPT(none)
> > [   44.609635] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014
> > [   44.610695] RIP: 0010:userfaultfd_release_all+0x3a8/0x460
> > 
> > <snip other registers, drop unreliable trace>
> > 
> > [   44.617726] Call Trace:
> > [   44.617926]  <TASK>
> > [   44.619284]  userfaultfd_release+0xef/0x1b0
> > [   44.620976]  __fput+0x3f9/0xb60
> > [   44.621240]  fput_close_sync+0x110/0x210
> > [   44.622222]  __x64_sys_close+0x8f/0x120
> > [   44.622530]  do_syscall_64+0x5b/0x2f0
> > [   44.622840]  entry_SYSCALL_64_after_hwframe+0x76/0x7e
> > [   44.623244] RIP: 0033:0x7f365bb3f227
> > 
> > Kernel panics because it detects UFFD inconsistency during
> > userfaultfd_release_all(). Specifically, a VMA which has a valid pointer
> > to vma->vm_userfaultfd_ctx, but no UFFD flags in vma->vm_flags.
> > 
> > The inconsistency is caused in ksm_madvise(): when user calls madvise()
> > with MADV_UNMEARGEABLE on a VMA that is registered for UFFD in MINOR
> > mode, it accidentally clears all flags stored in the upper 32 bits of
> > vma->vm_flags.
> > 
> > Assuming x86_64 kernel build, unsigned long is 64-bit and unsigned int
> > and int are 32-bit wide. This setup causes the following mishap during
> > the &= ~VM_MERGEABLE assignment.
> > 
> > VM_MERGEABLE is a 32-bit constant of type unsigned int, 0x8000'0000.
> > After ~ is applied, it becomes 0x7fff'ffff unsigned int, which is then
> > promoted to unsigned long before the & operation. This promotion fills
> > upper 32 bits with leading 0s, as we're doing unsigned conversion (and
> > even for a signed conversion, this wouldn't help as the leading bit is
> > 0). & operation thus ends up AND-ing vm_flags with 0x0000'0000'7fff'ffff
> > instead of intended 0xffff'ffff'7fff'ffff and hence accidentally clears
> > the upper 32-bits of its value.
> > 
> > Fix it by changing `VM_MERGEABLE` constant to unsigned long, using the
> > BIT() macro.
> > 
> > Note: other VM_* flags are not affected:
> > This only happens to the VM_MERGEABLE flag, as the other VM_* flags are
> > all constants of type int and after ~ operation, they end up with
> > leading 1 and are thus converted to unsigned long with leading 1s.
> > 
> > Note 2:
> > After commit 31defc3b01d9 ("userfaultfd: remove (VM_)BUG_ON()s"), this is
> > no longer a kernel BUG, but a WARNING at the same place:
> > 
> > [   45.595973] WARNING: CPU: 1 PID: 2474 at mm/userfaultfd.c:2067
> > 
> > but the root-cause (flag-drop) remains the same.
> > 
> > Fixes: 7677f7fd8be76 ("userfaultfd: add minor fault registration mode")
> 
> Late to the party, but it seems to me the correct Fixes: should be
> f8af4da3b4c1 ("ksm: the mm interface to ksm")
> which introduced the flag and the buggy clearing code, no?
> 
> Commit 7677f7fd8be76 is just one that notices it, right? But there are other
> flags in >32 bit area, including pkeys etc. Sounds rather dangerous if they
> can be cleared using a madvise.
> 
> So we can't amend the Fixes: now but maybe could advise stable to backport
> for even older versions than based on 7677f7fd8be76 ?
> 

Good point. It was a bit tricky to determine the correct "fixes" tag, as
there were more candidates:
- the commit that initially introduced VM_MERGEABLE as a constant with
  different inferred type to other vm_flags constants
- the commit that first started using upper 32 bits of vm_flags and did
  not make sure the constants are defined safely
- f8af4da3b4c1 indeed, as the one that makes the drop actually possible
- 7677f7fd8be76 that shows us a path where the drop manifests

Looking back, I agree f8af4da3b4c1 is the better option, but as you
said, that won't be changed now.

Nevertheless, I'll send the backports after a round of kselftests,
thanks for pointing this out.

Have a good day,
Jakub
 



Amazon Web Services Development Center Germany GmbH
Tamara-Danz-Str. 13
10243 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Christof Hellmis
Eingetragen am Amtsgericht Charlottenburg unter HRB 257764 B
Sitz: Berlin
Ust-ID: DE 365 538 597