* [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit
@ 2026-07-01 1:30 Yun Zhou
2026-07-01 1:30 ` [PATCH 1/2] ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive deadlock Yun Zhou
` (2 more replies)
0 siblings, 3 replies; 5+ messages in thread
From: Yun Zhou @ 2026-07-01 1:30 UTC (permalink / raw)
To: almaz.alexandrovich; +Cc: ntfs3, linux-kernel, yun.zhou
Commit d62cf685d12e ("fs/ntfs3: hold ni_lock across readdir metadata
walk") extended ni_lock to cover the entire directory walk in
ntfs_readdir(), including dir_emit() calls that copy data to userspace.
This introduced multiple deadlocks reported by syzkaller:
1. Recursive ni_lock: ntfs_dir_emit() -> ntfs_iget5() -> mi_read()
tries to acquire ni_lock on a child inode with the same lock class
while the parent's ni_lock is already held.
2. Lock-order inversions: dir_emit() may trigger page faults that need
mmap_lock or mapping->invalidate_lock, conflicting with paths that
acquire those locks before ni_lock (ntfs_fallocate, mmap).
This series fixes both issues:
- Patch 1 removes the ntfs_iget5() call that caused recursive locking.
- Patch 2 releases ni_lock before calling dir_emit(), using the XFS
approach of reading data into stable buffers under the lock, then
emitting to userspace after releasing it.
Yun Zhou (2):
ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive
deadlock
ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix
deadlocks
fs/ntfs3/dir.c | 38 ++++++++++++++++++++------------------
1 file changed, 20 insertions(+), 18 deletions(-)
--
2.43.0
^ permalink raw reply [flat|nested] 5+ messages in thread* [PATCH 1/2] ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive deadlock 2026-07-01 1:30 [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Yun Zhou @ 2026-07-01 1:30 ` Yun Zhou 2026-07-01 1:30 ` [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks Yun Zhou 2026-07-19 12:29 ` [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Zhou, Yun 2 siblings, 0 replies; 5+ messages in thread From: Yun Zhou @ 2026-07-01 1:30 UTC (permalink / raw) To: almaz.alexandrovich; +Cc: ntfs3, linux-kernel, yun.zhou ntfs_dir_emit() calls ntfs_iget5() to refine the directory entry type by reading the child inode's MFT record. However, ntfs_dir_emit() is called from ntfs_readdir() under the parent directory's ni_lock, and ntfs_iget5() -> mi_read() tries to acquire ni_lock on the child inode with the same lock class, triggering a recursive locking deadlock. Remove the ntfs_iget5() call. The dt_type derived from the directory entry's FILE_ATTRIBUTE_DIRECTORY flag is sufficient for readdir. The original code's own comment noted this approach was "not good idea". Reported-by: syzbot+17f812893d5906837f33@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=17f812893d5906837f33 Fixes: d62cf685d12e ("fs/ntfs3: hold ni_lock across readdir metadata walk") Signed-off-by: Yun Zhou <yun.zhou@windriver.com> --- fs/ntfs3/dir.c | 16 ---------------- 1 file changed, 16 deletions(-) diff --git a/fs/ntfs3/dir.c b/fs/ntfs3/dir.c index 873d52233003..07de564a2c47 100644 --- a/fs/ntfs3/dir.c +++ b/fs/ntfs3/dir.c @@ -327,22 +327,6 @@ static inline bool ntfs_dir_emit(struct ntfs_sb_info *sbi, */ dt_type = (fname->dup.fa & FILE_ATTRIBUTE_DIRECTORY) ? DT_DIR : DT_REG; - /* - * It is not reliable to detect the type of name using duplicated information - * stored in parent directory. - * The only correct way to get the type of name - read MFT record and find ATTR_STD. - * The code below is not good idea. - * It does additional locks/reads just to get the type of name. - * Should we use additional mount option to enable branch below? - */ - if (fname->dup.extend_data && ino != ni->mi.rno) { - struct inode *inode = ntfs_iget5(sbi->sb, &e->ref, NULL); - if (!IS_ERR_OR_NULL(inode)) { - dt_type = fs_umode_to_dtype(inode->i_mode); - iput(inode); - } - } - return dir_emit(ctx, (s8 *)name, name_len, ino, dt_type); } -- 2.43.0 ^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks 2026-07-01 1:30 [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Yun Zhou 2026-07-01 1:30 ` [PATCH 1/2] ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive deadlock Yun Zhou @ 2026-07-01 1:30 ` Yun Zhou 2026-10-16 14:48 ` Konstantin Komarov 2026-07-19 12:29 ` [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Zhou, Yun 2 siblings, 1 reply; 5+ messages in thread From: Yun Zhou @ 2026-07-01 1:30 UTC (permalink / raw) To: almaz.alexandrovich; +Cc: ntfs3, linux-kernel, yun.zhou ntfs_readdir() holds ni_lock while calling ntfs_read_hdr(), which calls dir_emit() to copy directory entries to userspace. If the user buffer triggers a page fault, the fault handler may need mmap_lock or mapping->invalidate_lock, creating lock-order inversions with paths that acquire those locks before ni_lock (e.g. ntfs_fallocate, mmap). Fix by releasing ni_lock before calling ntfs_read_hdr(): - For non-root index nodes, indx_read_ra() returns data in a kmalloc'd buffer (node->index) that remains valid after ni_unlock. - For the root index header, the data lives in the main MFT record whose lifetime is bound to the inode. i_rwsem shared prevents concurrent directory modifications, so it is safe to access after ni_unlock. Reported-by: syzbot+7736960b837908f3a81d@syzkaller.appspotmail.com Reported-by: syzbot+262a71e9d2faf8747085@syzkaller.appspotmail.com Reported-by: syzbot+cfc6e8106bec1c524f7d@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=7736960b837908f3a81d Closes: https://syzkaller.appspot.com/bug?extid=262a71e9d2faf8747085 Closes: https://syzkaller.appspot.com/bug?extid=cfc6e8106bec1c524f7d Fixes: d62cf685d12e ("fs/ntfs3: hold ni_lock across readdir metadata walk") Signed-off-by: Yun Zhou <yun.zhou@windriver.com> --- fs/ntfs3/dir.c | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-) diff --git a/fs/ntfs3/dir.c b/fs/ntfs3/dir.c index 07de564a2c47..65883409689a 100644 --- a/fs/ntfs3/dir.c +++ b/fs/ntfs3/dir.c @@ -491,12 +491,22 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) if (pos >= sbi->record_size) { bit = (pos - sbi->record_size) >> index_bits; } else { + /* + * Release ni_lock before dir_emit (which may fault into + * user pages). root->ihdr lives in the main MFT record + * whose lifetime is bound to the inode, and i_rwsem shared + * prevents concurrent directory modifications. + */ + ni_unlock(ni); + /* * Add each name from root in 'ctx'. */ err = ntfs_read_hdr(sbi, ni, &root->ihdr, 0, pos, name, ctx); if (err) - goto out_unlock; + goto out; + + ni_lock(ni); bit = 0; } @@ -525,6 +535,12 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) if (err) goto out_unlock; + /* + * node->index is a kmalloc'd buffer, safe after ni_unlock. + * Release lock before dir_emit to avoid lock-order issues. + */ + ni_unlock(ni); + /* * Add each name from index in 'ctx'. */ @@ -532,7 +548,9 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) ((u64)bit << index_bits) + sbi->record_size, pos, name, ctx); if (err) - goto out_unlock; + goto out; + + ni_lock(ni); } out_unlock: -- 2.43.0 ^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks 2026-07-01 1:30 ` [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks Yun Zhou @ 2026-10-16 14:48 ` Konstantin Komarov 0 siblings, 0 replies; 5+ messages in thread From: Konstantin Komarov @ 2026-10-16 14:48 UTC (permalink / raw) To: Yun Zhou; +Cc: ntfs3, linux-kernel On 7/1/26 03:30, Yun Zhou wrote: > ntfs_readdir() holds ni_lock while calling ntfs_read_hdr(), which calls > dir_emit() to copy directory entries to userspace. If the user buffer > triggers a page fault, the fault handler may need mmap_lock or > mapping->invalidate_lock, creating lock-order inversions with paths > that acquire those locks before ni_lock (e.g. ntfs_fallocate, mmap). > > Fix by releasing ni_lock before calling ntfs_read_hdr(): > > - For non-root index nodes, indx_read_ra() returns data in a kmalloc'd > buffer (node->index) that remains valid after ni_unlock. > > - For the root index header, the data lives in the main MFT record > whose lifetime is bound to the inode. i_rwsem shared prevents > concurrent directory modifications, so it is safe to access after > ni_unlock. > > Reported-by: syzbot+7736960b837908f3a81d@syzkaller.appspotmail.com > Reported-by: syzbot+262a71e9d2faf8747085@syzkaller.appspotmail.com > Reported-by: syzbot+cfc6e8106bec1c524f7d@syzkaller.appspotmail.com > Closes: https://syzkaller.appspot.com/bug?extid=7736960b837908f3a81d > Closes: https://syzkaller.appspot.com/bug?extid=262a71e9d2faf8747085 > Closes: https://syzkaller.appspot.com/bug?extid=cfc6e8106bec1c524f7d > Fixes: d62cf685d12e ("fs/ntfs3: hold ni_lock across readdir metadata walk") > Signed-off-by: Yun Zhou <yun.zhou@windriver.com> > --- > fs/ntfs3/dir.c | 22 ++++++++++++++++++++-- > 1 file changed, 20 insertions(+), 2 deletions(-) > > diff --git a/fs/ntfs3/dir.c b/fs/ntfs3/dir.c > index 07de564a2c47..65883409689a 100644 > --- a/fs/ntfs3/dir.c > +++ b/fs/ntfs3/dir.c > @@ -491,12 +491,22 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) > if (pos >= sbi->record_size) { > bit = (pos - sbi->record_size) >> index_bits; > } else { > + /* > + * Release ni_lock before dir_emit (which may fault into > + * user pages). root->ihdr lives in the main MFT record > + * whose lifetime is bound to the inode, and i_rwsem shared > + * prevents concurrent directory modifications. > + */ > + ni_unlock(ni); > + > /* > * Add each name from root in 'ctx'. > */ > err = ntfs_read_hdr(sbi, ni, &root->ihdr, 0, pos, name, ctx); > if (err) > - goto out_unlock; > + goto out; > + > + ni_lock(ni); > bit = 0; > } > > @@ -525,6 +535,12 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) > if (err) > goto out_unlock; > > + /* > + * node->index is a kmalloc'd buffer, safe after ni_unlock. > + * Release lock before dir_emit to avoid lock-order issues. > + */ > + ni_unlock(ni); > + > /* > * Add each name from index in 'ctx'. > */ > @@ -532,7 +548,9 @@ static int ntfs_readdir(struct file *file, struct dir_context *ctx) > ((u64)bit << index_bits) + sbi->record_size, > pos, name, ctx); > if (err) > - goto out_unlock; > + goto out; > + > + ni_lock(ni); > } > > out_unlock: Hello, The patch was applied, thanks. Regards, Konstantin ^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit 2026-07-01 1:30 [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Yun Zhou 2026-07-01 1:30 ` [PATCH 1/2] ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive deadlock Yun Zhou 2026-07-01 1:30 ` [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks Yun Zhou @ 2026-07-19 12:29 ` Zhou, Yun 2 siblings, 0 replies; 5+ messages in thread From: Zhou, Yun @ 2026-07-19 12:29 UTC (permalink / raw) To: almaz.alexandrovich; +Cc: ntfs3, linux-kernel, yun.zhou Friendly ping. On 7/1/2026 9:30 AM, Yun Zhou wrote: > Commit d62cf685d12e ("fs/ntfs3: hold ni_lock across readdir metadata > walk") extended ni_lock to cover the entire directory walk in > ntfs_readdir(), including dir_emit() calls that copy data to userspace. > This introduced multiple deadlocks reported by syzkaller: > > 1. Recursive ni_lock: ntfs_dir_emit() -> ntfs_iget5() -> mi_read() > tries to acquire ni_lock on a child inode with the same lock class > while the parent's ni_lock is already held. > > 2. Lock-order inversions: dir_emit() may trigger page faults that need > mmap_lock or mapping->invalidate_lock, conflicting with paths that > acquire those locks before ni_lock (ntfs_fallocate, mmap). > > This series fixes both issues: > - Patch 1 removes the ntfs_iget5() call that caused recursive locking. > - Patch 2 releases ni_lock before calling dir_emit(), using the XFS > approach of reading data into stable buffers under the lock, then > emitting to userspace after releasing it. > > Yun Zhou (2): > ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive > deadlock > ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix > deadlocks > > fs/ntfs3/dir.c | 38 ++++++++++++++++++++------------------ > 1 file changed, 20 insertions(+), 18 deletions(-) > ^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2026-10-16 14:48 UTC | newest] Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed) -- links below jump to the message on this page -- 2026-07-01 1:30 [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Yun Zhou 2026-07-01 1:30 ` [PATCH 1/2] ntfs3: remove ntfs_iget5 call in ntfs_dir_emit to fix recursive deadlock Yun Zhou 2026-07-01 1:30 ` [PATCH 2/2] ntfs3: release ni_lock before dir_emit in ntfs_readdir to fix deadlocks Yun Zhou 2026-10-16 14:48 ` Konstantin Komarov 2026-07-19 12:29 ` [PATCH 0/2] ntfs3: fix deadlocks in ntfs_readdir caused by ni_lock held across dir_emit Zhou, Yun
This is a public inbox, see mirroring instructions for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®