Re: [PATCH 6/6] x86/mm: Avoid mmap lock for shadow stack pop fast path

["Edgecombe, Rick P" <rick.p.edgecombe@intel.com>]

On Wed, 2026-04-29 at 11:20 -0700, Dave Hansen wrote:
> +	vma = lock_vma_under_rcu_wait(current->mm, *ssp);
> +	if (!vma)
> +		return -EINVAL;
> +
> +	if (!(vma->vm_flags & VM_SHADOW_STACK)) {
> +		vma_end_read(vma);
> +		return -EINVAL;
> +	}
> +
> +	err = get_shstk_data(&token_addr, (unsigned long __user *)*ssp);

Unfortunately, I think it won't work for the shadow stack case with the user
access. I get this splat from the shadow stack selftests:

======================================================
 WARNING: possible circular locking dependency detected
 7.1.0-rc1+ #2936 Not tainted
 ------------------------------------------------------
 test_shadow_sta/930 is trying to acquire lock:
 ff32a05fbc6a1008 (&mm->mmap_lock){++++}-{4:4}, at: __might_fault+0x3c/0x80
 
               but task is already holding lock:
 ff32a05f4caf3c48 (vm_lock){++++}-{0:0}, at: lock_vma_under_rcu+0xaf/0x2e0
 
               which lock already depends on the new lock.

 
               the existing dependency chain (in reverse order) is:
 
               -> #1 (vm_lock){++++}-{0:0}:
        lock_acquire+0xbd/0x2f0
        __vma_start_exclude_readers+0x8d/0x1e0
        __vma_start_write+0x56/0xe0
        vma_expand+0x7e/0x390
        relocate_vma_down+0x126/0x220
        setup_arg_pages+0x269/0x430
        load_elf_binary+0x3d1/0x1840
        bprm_execve+0x2cf/0x730
        kernel_execve+0xf6/0x160
        kernel_init+0xb9/0x1c0
        ret_from_fork+0x2eb/0x340
        ret_from_fork_asm+0x1a/0x30
 
               -> #0 (&mm->mmap_lock){++++}-{4:4}:
        check_prev_add+0xf1/0xd00
        __lock_acquire+0x14a8/0x1ac0
        lock_acquire+0xbd/0x2f0
        __might_fault+0x5b/0x80
        restore_signal_shadow_stack+0xd6/0x270
        __do_sys_rt_sigreturn+0xdf/0xf0
        do_syscall_64+0x11c/0xf80
        entry_SYSCALL_64_after_hwframe+0x77/0x7f
 
               other info that might help us debug this:

  Possible unsafe locking scenario:

        CPU0                    CPU1
        ----                    ----
   rlock(vm_lock);
                                lock(&mm->mmap_lock);
                                lock(vm_lock);
   rlock(&mm->mmap_lock);
 
                *** DEADLOCK ***

 1 lock held by test_shadow_sta/930:
  #0: ff32a05f4caf3c48 (vm_lock){++++}-{0:0}, at: lock_vma_under_rcu+0xaf/0x2e0
 
               stack backtrace:
 CPU: 18 UID: 0 PID: 930 Comm: test_shadow_sta Not tainted 7.1.0-rc1+ #2936
PREEMPT(full) 
 Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
 Call Trace:
  <TASK>
  dump_stack_lvl+0x68/0xa0
  print_circular_bug+0x2ca/0x400
  check_noncircular+0x12f/0x150
  ? __lock_acquire+0x49c/0x1ac0
  check_prev_add+0xf1/0xd00
  ? reacquire_held_locks+0xe4/0x200
  __lock_acquire+0x14a8/0x1ac0
  lock_acquire+0xbd/0x2f0
  ? __might_fault+0x3c/0x80
  ? lock_is_held_type+0xa0/0x120
  ? __might_fault+0x3c/0x80
  __might_fault+0x5b/0x80
  ? __might_fault+0x3c/0x80
  restore_signal_shadow_stack+0xd6/0x270
  __do_sys_rt_sigreturn+0xdf/0xf0
  do_syscall_64+0x11c/0xf80
  entry_SYSCALL_64_after_hwframe+0x77/0x7f
 RIP: 0033:0x40212f
 Code: 61 00 00 e8 73 f1 ff ff 48 8b 05 4c 61 00 00 31 d2 48 0f 38 f6 10 48 8b
44 24 08 64 48 2b 08
 RSP: 002b:00007ffc286fb208 EFLAGS: 00010202
 RAX: 0000000000000000 RBX: 0000000000000000 RCX: 00007ff628b187b0
 RDX: 0000000000000000 RSI: 00000000066492a0 RDI: 0000000000000000
 RBP: 00007ffc286fb360 R08: 0000000000000000 R09: 0000000000000000
 R10: 0000000000000000 R11: 0000000000000202 R12: 0000000000000000
 R13: 0000000000000001 R14: 00007ff628b6c000 R15: 0000000000406e18


I guess the problem is the lock ordering. Not sure if there is any slow path
avoidance details that could make this splat a false positive. But how about
this simpler munmap() case:

Shadow stack signal                          munmap()
-------------------                          --------
vma_start_read() (VM_SHADOW_STACK check)
                                             mmap_write_lock()
mmap_read_lock() (user fault) <- deadlock
                                             vma_start_write() <-deadlock



> +
> +	vma_end_read(vma);
> +
> +	if (err)
> +		return err;
>  
>  	/* Restore SSP aligned? */
>  	if (unlikely(!IS_ALIGNED(token_addr, 8)))