Message-ID: <197389ce-9f42-4d6a-91c4-fce116e988b4@suse.cz>
Date: Tue, 29 Jul 2025 00:04:01 +0200
From: Vlastimil Babka <vbabka@...e.cz>
To: Suren Baghdasaryan <surenb@...gle.com>, akpm@...ux-foundation.org
Cc: jannh@...gle.com, Liam.Howlett@...cle.com, lorenzo.stoakes@...cle.com,
 pfalcato@...e.de, linux-mm@...ck.org, linux-kernel@...r.kernel.org,
 stable@...r.kernel.org
Subject: Re: [PATCH v2 1/1] mm: fix a UAF when vma->mm is freed after
 vma->vm_refcnt got dropped

On 7/28/25 19:53, Suren Baghdasaryan wrote:
> By inducing delays in the right places, Jann Horn created a reproducer
> for a hard to hit UAF issue that became possible after VMAs were allowed
> to be recycled by adding SLAB_TYPESAFE_BY_RCU to their cache.
> 
> Race description is borrowed from Jann's discovery report:
> lock_vma_under_rcu() looks up a VMA locklessly with mas_walk() under
> rcu_read_lock(). At that point, the VMA may be concurrently freed, and
> it can be recycled by another process. vma_start_read() then
> increments the vma->vm_refcnt (if it is in an acceptable range), and
> if this succeeds, vma_start_read() can return a recycled VMA.
> 
> In this scenario where the VMA has been recycled, lock_vma_under_rcu()
> will then detect the mismatching ->vm_mm pointer and drop the VMA
> through vma_end_read(), which calls vma_refcount_put().
> vma_refcount_put() drops the refcount and then calls rcuwait_wake_up()
> using a copy of vma->vm_mm. This is wrong: It implicitly assumes that
> the caller is keeping the VMA's mm alive, but in this scenario the caller
> has no relation to the VMA's mm, so the rcuwait_wake_up() can cause UAF.
> 
> The diagram depicting the race:
> T1         T2         T3
> ==         ==         ==
> lock_vma_under_rcu
>   mas_walk
>           <VMA gets removed from mm>
>                       mmap
>                         <the same VMA is reallocated>
>   vma_start_read
>     __refcount_inc_not_zero_limited_acquire
>                       munmap
>                         __vma_enter_locked
>                           refcount_add_not_zero
>   vma_end_read
>     vma_refcount_put
>       __refcount_dec_and_test
>                           rcuwait_wait_event
>                             <finish operation>
>       rcuwait_wake_up [UAF]
> 
> Note that rcuwait_wait_event() in T3 does not block because refcount
> was already dropped by T1. At this point T3 can exit and free the mm
> causing UAF in T1.
> To avoid this we move vma->vm_mm verification into vma_start_read() and
> grab vma->vm_mm to stabilize it before vma_refcount_put() operation.
> 
> Fixes: 3104138517fc ("mm: make vma cache SLAB_TYPESAFE_BY_RCU")
> Reported-by: Jann Horn <jannh@...gle.com>
> Closes: https://lore.kernel.org/all/CAG48ez0-deFbVH=E3jbkWx=X3uVbd8nWeo6kbJPQ0KoUD+m2tA@mail.gmail.com/
> Signed-off-by: Suren Baghdasaryan <surenb@...gle.com>
> Cc: <stable@...r.kernel.org>

As for further steps, considered replying to [1] but maybe it's better here.

As for a KISS fix including stable, great. Seems a nice improvement to
actually handle "vma->vm_mm != mm" in vma_start_read() like this - good idea!

Less great is that there's now a subtle assumption that some (but not all!)
cases where vma_start_read() returns NULL imply that we dropped the rcu
lock. And it doesn't matter as the callers abort or fallback to mmap sem
anyway in that case. Hopefully we can improve that a bit.

The idea of moving rcu lock and mas walk inside vma_start_read() is indeed
busted with lock_next_vma(). The iterator difference could be perhaps solved
by having lock_vma_under_rcu() set up its own one (instead of MA_STATE) in a
way that vma_next() would do the right thing for it. However there would
still be the difference that lock_next_vma() expects we are already under
rcu lock, and lock_vma_under_rcu() doesn't.

So what we can perhaps do instead is move vma_start_read() to mm/mmap_lock.c
(no other users so why expose it in a header for potential misuse). And then
indeed just make it drop rcu lock completely (and not reacquire) any time
it's returning NULL, document that and adjust callers to that. I think it's
less subtle than dropping and reacquring, and should simplify the error
handling in the callers a bit.

[1]
https://lore.kernel.org/all/CAJuCfpEMhGe_eZuFm__4CDstM9%3DOG2kTUTziNL-f%3DM3BYQor2A@mail.gmail.com/

> ---
> Changes since v1 [1]
> - Made a copy of vma->mm before using it in vma_start_read(),
> per Vlastimil Babka
> 
> Notes:
> - Applies cleanly over mm-unstable.
> - Should be applied to 6.15 and 6.16 but these branches do not
> have lock_next_vma() function, so the change in lock_next_vma() should be
> skipped when applying to those branches.
> 
> [1] https://lore.kernel.org/all/20250728170950.2216966-1-surenb@google.com/
> 
>  include/linux/mmap_lock.h | 23 +++++++++++++++++++++++
>  mm/mmap_lock.c            | 10 +++-------
>  2 files changed, 26 insertions(+), 7 deletions(-)
> 
> diff --git a/include/linux/mmap_lock.h b/include/linux/mmap_lock.h
> index 1f4f44951abe..da34afa2f8ef 100644
> --- a/include/linux/mmap_lock.h
> +++ b/include/linux/mmap_lock.h
> @@ -12,6 +12,7 @@ extern int rcuwait_wake_up(struct rcuwait *w);
>  #include <linux/tracepoint-defs.h>
>  #include <linux/types.h>
>  #include <linux/cleanup.h>
> +#include <linux/sched/mm.h>
>  
>  #define MMAP_LOCK_INITIALIZER(name) \
>  	.mmap_lock = __RWSEM_INITIALIZER((name).mmap_lock),
> @@ -183,6 +184,28 @@ static inline struct vm_area_struct *vma_start_read(struct mm_struct *mm,
>  	}
>  
>  	rwsem_acquire_read(&vma->vmlock_dep_map, 0, 1, _RET_IP_);
> +
> +	/*
> +	 * If vma got attached to another mm from under us, that mm is not
> +	 * stable and can be freed in the narrow window after vma->vm_refcnt
> +	 * is dropped and before rcuwait_wake_up(mm) is called. Grab it before
> +	 * releasing vma->vm_refcnt.
> +	 */
> +	if (unlikely(vma->vm_mm != mm)) {
> +		/* Use a copy of vm_mm in case vma is freed after we drop vm_refcnt */
> +		struct mm_struct *other_mm = vma->vm_mm;
> +		/*
> +		 * __mmdrop() is a heavy operation and we don't need RCU
> +		 * protection here. Release RCU lock during these operations.
> +		 */
> +		rcu_read_unlock();
> +		mmgrab(other_mm);
> +		vma_refcount_put(vma);
> +		mmdrop(other_mm);
> +		rcu_read_lock();
> +		return NULL;
> +	}
> +
>  	/*
>  	 * Overflow of vm_lock_seq/mm_lock_seq might produce false locked result.
>  	 * False unlocked result is impossible because we modify and check
> diff --git a/mm/mmap_lock.c b/mm/mmap_lock.c
> index 729fb7d0dd59..aa3bc42ecde0 100644
> --- a/mm/mmap_lock.c
> +++ b/mm/mmap_lock.c
> @@ -164,8 +164,7 @@ struct vm_area_struct *lock_vma_under_rcu(struct mm_struct *mm,
>  	 */
>  
>  	/* Check if the vma we locked is the right one. */
> -	if (unlikely(vma->vm_mm != mm ||
> -		     address < vma->vm_start || address >= vma->vm_end))
> +	if (unlikely(address < vma->vm_start || address >= vma->vm_end))
>  		goto inval_end_read;
>  
>  	rcu_read_unlock();
> @@ -236,11 +235,8 @@ struct vm_area_struct *lock_next_vma(struct mm_struct *mm,
>  		goto fallback;
>  	}
>  
> -	/*
> -	 * Verify the vma we locked belongs to the same address space and it's
> -	 * not behind of the last search position.
> -	 */
> -	if (unlikely(vma->vm_mm != mm || from_addr >= vma->vm_end))
> +	/* Verify the vma is not behind of the last search position. */
> +	if (unlikely(from_addr >= vma->vm_end))
>  		goto fallback_unlock;
>  
>  	/*
> 
> base-commit: c617a4dd7102e691fa0fb2bc4f6b369e37d7f509

Powered by blists - more mailing lists