linux-mm.kvack.org archive mirror
 help / color / mirror / Atom feed
From: Oscar Salvador <osalvador@suse.de>
To: Gavin Guo <gavinguo@igalia.com>
Cc: linux-mm@kvack.org, linux-kernel@vger.kernel.org,
	muchun.song@linux.dev, akpm@linux-foundation.org,
	mike.kravetz@oracle.com, kernel-dev@igalia.com,
	stable@vger.kernel.org, Hugh Dickins <hughd@google.com>,
	Florent Revest <revest@google.com>, Gavin Shan <gshan@redhat.com>
Subject: Re: [PATCH] mm/hugetlb: fix a deadlock with pagecache_folio and hugetlb_fault_mutex_table
Date: Tue, 20 May 2025 21:53:00 +0200	[thread overview]
Message-ID: <aCzdnAmuOylilU1p@localhost.localdomain> (raw)
In-Reply-To: <20250513093448.592150-1-gavinguo@igalia.com>

On Tue, May 13, 2025 at 05:34:48PM +0800, Gavin Guo wrote:
> The patch fixes a deadlock which can be triggered by an internal
> syzkaller [1] reproducer and captured by bpftrace script [2] and its log
> [3] in this scenario:
> 
> Process 1                              Process 2
> ---				       ---
> hugetlb_fault
>   mutex_lock(B) // take B
>   filemap_lock_hugetlb_folio
>     filemap_lock_folio
>       __filemap_get_folio
>         folio_lock(A) // take A
>   hugetlb_wp
>     mutex_unlock(B) // release B
>     ...                                hugetlb_fault
>     ...                                  mutex_lock(B) // take B
>                                          filemap_lock_hugetlb_folio
>                                            filemap_lock_folio
>                                              __filemap_get_folio
>                                                folio_lock(A) // blocked
>     unmap_ref_private
>     ...
>     mutex_lock(B) // retake and blocked
> 
...
> Signed-off-by: Gavin Guo <gavinguo@igalia.com>

I think this is more convoluted that it needs to be?

hugetlb_wp() is called from hugetlb_no_page() and hugetlb_fault().
hugetlb_no_page() locks and unlocks the lock itself, which leaves us
with hugetlb_fault().

hugetlb_fault() always passed the folio locked to hugetlb_wp(), and the
latter only unlocks it when we have a cow from owner happening and we
cannot satisfy the allocation.
So, should not checking whether the folio is still locked after
returning enough?
What speaks against:

 diff --git a/mm/hugetlb.c b/mm/hugetlb.c
 index bd8971388236..23b57c5689a4 100644
 --- a/mm/hugetlb.c
 +++ b/mm/hugetlb.c
 @@ -6228,6 +6228,12 @@ static vm_fault_t hugetlb_wp(struct folio *pagecache_folio,
  			u32 hash;
 
  			folio_put(old_folio);
 +			/*
 +			* The pagecache_folio needs to be unlocked to avoid
 +			* deadlock when the child unmaps the folio.
 +			*/
 +			if (pagecache_folio)
 +				folio_unlock(pagecache_folio);
  			/*
  			 * Drop hugetlb_fault_mutex and vma_lock before
  			 * unmapping.  unmapping needs to hold vma_lock
 @@ -6825,7 +6831,12 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
  	spin_unlock(vmf.ptl);
 
  	if (pagecache_folio) {
 -		folio_unlock(pagecache_folio);
 +		/*
 +		 * hugetlb_wp() might have already unlocked pagecache_folio, so
 +		 * skip it if that is the case.
 +		 */
 +		if (folio_test_locked(pagecache_folio))
 +			folio_unlock(pagecache_folio);
  		folio_put(pagecache_folio);
  	}
  out_mutex:

> ---
>  mm/hugetlb.c | 33 ++++++++++++++++++++++++++++-----
>  1 file changed, 28 insertions(+), 5 deletions(-)
> 
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index e3e6ac991b9c..ad54a74aa563 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -6115,7 +6115,8 @@ static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
>   * Keep the pte_same checks anyway to make transition from the mutex easier.
>   */
>  static vm_fault_t hugetlb_wp(struct folio *pagecache_folio,
> -		       struct vm_fault *vmf)
> +		       struct vm_fault *vmf,
> +		       bool *pagecache_folio_unlocked)
>  {
>  	struct vm_area_struct *vma = vmf->vma;
>  	struct mm_struct *mm = vma->vm_mm;
> @@ -6212,6 +6213,22 @@ static vm_fault_t hugetlb_wp(struct folio *pagecache_folio,
>  			u32 hash;
>  
>  			folio_put(old_folio);
> +			/*
> +			 * The pagecache_folio needs to be unlocked to avoid
> +			 * deadlock and we won't re-lock it in hugetlb_wp(). The
> +			 * pagecache_folio could be truncated after being
> +			 * unlocked. So its state should not be relied
> +			 * subsequently.
> +			 *
> +			 * Setting *pagecache_folio_unlocked to true allows the
> +			 * caller to handle any necessary logic related to the
> +			 * folio's unlocked state.
> +			 */
> +			if (pagecache_folio) {
> +				folio_unlock(pagecache_folio);
> +				if (pagecache_folio_unlocked)
> +					*pagecache_folio_unlocked = true;
> +			}
>  			/*
>  			 * Drop hugetlb_fault_mutex and vma_lock before
>  			 * unmapping.  unmapping needs to hold vma_lock
> @@ -6566,7 +6583,7 @@ static vm_fault_t hugetlb_no_page(struct address_space *mapping,
>  	hugetlb_count_add(pages_per_huge_page(h), mm);
>  	if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
>  		/* Optimization, do the COW without a second fault */
> -		ret = hugetlb_wp(folio, vmf);
> +		ret = hugetlb_wp(folio, vmf, NULL);
>  	}
>  
>  	spin_unlock(vmf->ptl);
> @@ -6638,6 +6655,7 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  	struct hstate *h = hstate_vma(vma);
>  	struct address_space *mapping;
>  	int need_wait_lock = 0;
> +	bool pagecache_folio_unlocked = false;
>  	struct vm_fault vmf = {
>  		.vma = vma,
>  		.address = address & huge_page_mask(h),
> @@ -6792,7 +6810,8 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  
>  	if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
>  		if (!huge_pte_write(vmf.orig_pte)) {
> -			ret = hugetlb_wp(pagecache_folio, &vmf);
> +			ret = hugetlb_wp(pagecache_folio, &vmf,
> +					&pagecache_folio_unlocked);
>  			goto out_put_page;
>  		} else if (likely(flags & FAULT_FLAG_WRITE)) {
>  			vmf.orig_pte = huge_pte_mkdirty(vmf.orig_pte);
> @@ -6809,10 +6828,14 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  out_ptl:
>  	spin_unlock(vmf.ptl);
>  
> -	if (pagecache_folio) {
> +	/*
> +	 * If the pagecache_folio is unlocked in hugetlb_wp(), we skip
> +	 * folio_unlock() here.
> +	 */
> +	if (pagecache_folio && !pagecache_folio_unlocked)
>  		folio_unlock(pagecache_folio);
> +	if (pagecache_folio)
>  		folio_put(pagecache_folio);
> -	}
>  out_mutex:
>  	hugetlb_vma_unlock_read(vma);
>  
> 
> base-commit: d76bb1ebb5587f66b0f8b8099bfbb44722bc08b3
> -- 
> 2.43.0
> 
> 

-- 
Oscar Salvador
SUSE Labs


  parent reply	other threads:[~2025-05-20 19:53 UTC|newest]

Thread overview: 14+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-05-13  9:34 Gavin Guo
2025-05-14  0:56 ` Andrew Morton
2025-05-14  4:33   ` Byungchul Park
2025-05-14  6:47 ` Byungchul Park
2025-05-14  8:10   ` Gavin Guo
2025-05-15  2:22     ` Byungchul Park
2025-05-16  6:03     ` Byungchul Park
2025-05-16  7:32       ` Gavin Guo
2025-05-16  7:43         ` Byungchul Park
2025-05-20 19:53 ` Oscar Salvador [this message]
2025-05-21 11:12   ` Gavin Guo
2025-05-26  4:41 ` Gavin Shan
2025-05-27  9:59   ` Gavin Guo
2025-05-27 10:59     ` Gavin Shan

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=aCzdnAmuOylilU1p@localhost.localdomain \
    --to=osalvador@suse.de \
    --cc=akpm@linux-foundation.org \
    --cc=gavinguo@igalia.com \
    --cc=gshan@redhat.com \
    --cc=hughd@google.com \
    --cc=kernel-dev@igalia.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=mike.kravetz@oracle.com \
    --cc=muchun.song@linux.dev \
    --cc=revest@google.com \
    --cc=stable@vger.kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox