Skip to contentSkip to conversation

vfs: serialize updates to file->f_sb_err with f_lock

PATCH RESEND In review Jeff Layton 2 participants 3 messages Jan 4, 2021 VFS

lore
Share & tools
lore
#0 lore raw

PATCH RESEND 1/1 vfs: serialize updates to file->f_sb_err with f_lock

When I added the ability for syncfs to report writeback errors, I neglected to adequately protect file->f_sb_err. While changes to sb->s_wb_err don't require locking, we do need to protect the errseq_t cursor in file->f_sb_err.

We could have racing updates to that value if two tasks are issuing syncfs() on the same fd at the same time, possibly causing an error to be reported twice or not at all.

Fix this by protecting the f_sb_err field with the file->f_lock.

Fixes 735e4ae5ba28 ("vfs: track per-sb writeback errors and report them to syncfs")
---
 fs/sync.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

Al, could you pick up this patch for v5.11 or v5.12? I sent the original version about a month ago, but it didn't get picked up.

In the original posting I marked this for stable, but I'm not sure it really qualifies since it's a pretty unlikely race with an oddball use-case (overlapping syncfs() calls on the same fd).

fs/sync.c +7 -2
index 1373a610dc78..3be26ff72431 100644
@@ -162,7 +162,7 @@ SYSCALL_DEFINE1(syncfs, int, fd)
 {
 	struct fd f = fdget(fd);
 	struct super_block *sb;
-	int ret, ret2;
+	int ret, ret2 = 0;
 
 	if (!f.file)
 		return -EBADF;
@@ -172,7 +172,12 @@ SYSCALL_DEFINE1(syncfs, int, fd)
 	ret = sync_filesystem(sb);
 	up_read(&sb->s_umount);
 
-	ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+	if (errseq_check(&sb->s_wb_err, f.file->f_sb_err)) {
+		/* Something changed, must use slow path */
+		spin_lock(&f.file->f_lock);
+		ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+		spin_unlock(&f.file->f_lock);
+	}
 
 	fdput(f);
 	return ret ? ret : ret2;
Signature · 1 line
2.29.2
maintainer
#1 lore raw

fs/sync.c SYSCALL_DEFINE1(syncfs, int, fd) · line 180

Quoted 11 lines · Jeff Layton
fs/sync.c +6 -1
@@ -172,7 +172,12 @@ SYSCALL_DEFINE1(syncfs, int, fd)
 	ret = sync_filesystem(sb);
 	up_read(&sb->s_umount);
 
-	ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+	if (errseq_check(&sb->s_wb_err, f.file->f_sb_err)) {
+		/* Something changed, must use slow path */
+		spin_lock(&f.file->f_lock);
+		ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+		spin_unlock(&f.file->f_lock);
+	}

Is there any point bothering with the fastpath here?
I mean, look at the up_read() immediately prior to that thing...

#2 lore raw
Quoted 6 lines · Al Viro
Quoted 5 lines · Jeff Layton
fs/sync.c +0 -0
@@ -172,7 +172,12 @@ SYSCALL_DEFINE1(syncfs, int, fd)
  	ret = sync_filesystem(sb);
  	up_read(&sb->s_umount);

Quoted 10 lines

fs/sync.c SYSCALL_DEFINE1(syncfs, int, fd) · line 180

Quoted 7 lines
fs/sync.c +6 -1
-	ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+	if (errseq_check(&sb->s_wb_err, f.file->f_sb_err)) {
+		/* Something changed, must use slow path */
+		spin_lock(&f.file->f_lock);
+		ret2 = errseq_check_and_advance(&sb->s_wb_err, &f.file->f_sb_err);
+		spin_unlock(&f.file->f_lock);
+	}

Is there any point bothering with the fastpath here?
I mean, look at the up_read() immediately prior to that thing...

It is a micro-optimization, but the vastly common case is that we will avoid the spinlock there. That said, I'm fine with dropping the fastpath if you prefer.

Signature · 1 line
Jeff Layton <[email protected]>