Skip to contentSkip to conversation

xfs: don't use current->journal_info

PATCH Superseded Dave Chinner 4 participants 6 messages Feb 21, 2024 – Feb 23, 2024 XFS filesystem

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfs: don't use current->journal_info

From: Dave Chinner <[email protected]>

syzbot reported an ext4 panic during a page fault where found a journal handle when it didn't expect to find one. The structure it tripped over had a value of 'TRAN' in the first entry in the structure, and that indicates it tripped over a struct xfs_trans instead of a jbd2 handle.

The reason for this is that the page fault was taken during a copy-out to a user buffer from an xfs bulkstat operation. XFS uses an "empty" transaction context for bulkstat to do automated metadata buffer cleanup, and so the transaction context is valid across the copyout of the bulkstat info into the user buffer.

We are using empty transaction contexts like this in XFS in more places to reduce the risk of failing to release objects we reference during the operation, especially during error handling. Hence we really need to ensure that we can take page faults from these contexts without leaving landmines for the code processing the page fault to trip over.

We really only use current->journal_info for a single warning check in xfs_vm_writepages() to ensure we aren't doing writeback from a transaction context. Writeback might need to do allocation, so it can need to run transactions itself. Hence it's a debug check to warn us that we've done something silly, and largely it is not all that useful.

So let's just remove all the use of current->journal_info in XFS and get rid of all the potential issues from nested contexts where current->journal_info might get misused by another filesytsem context.

---
 fs/xfs/scrub/common.c | 4 +---
 fs/xfs/xfs_aops.c     | 7 -------
 fs/xfs/xfs_icache.c   | 8 +++++---
 fs/xfs/xfs_trans.h    | 9 +--------
 4 files changed, 7 insertions(+), 21 deletions(-)
fs/xfs/scrub/common.c +1 -3
index 81f2b96bb5a7..d853348a48c8 100644
@@ -1000,9 +1000,7 @@ xchk_irele(
 	struct xfs_scrub	*sc,
 	struct xfs_inode	*ip)
 {
-	if (current->journal_info != NULL) {
-		ASSERT(current->journal_info == sc->tp);
-
+	if (sc->tp) {
 		/*
 		 * If we are in a transaction, we /cannot/ drop the inode
 		 * ourselves, because the VFS will trigger writeback, which
fs/xfs/xfs_aops.c +0 -7
index 813f85156b0c..bc3b649d29c4 100644
@@ -502,13 +502,6 @@ xfs_vm_writepages(
 {
 	struct xfs_writepage_ctx wpc = { };
 
-	/*
-	 * Writing back data in a transaction context can result in recursive
-	 * transactions. This is bad, so issue a warning and get out of here.
-	 */
-	if (WARN_ON_ONCE(current->journal_info))
-		return 0;
-
 	xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
 	return iomap_writepages(mapping, wbc, &wpc.ctx, &xfs_writeback_ops);
 }
fs/xfs/xfs_icache.c +5 -3
index 06046827b5fe..9b966af7d55c 100644
@@ -2030,8 +2030,10 @@ xfs_inodegc_want_queue_work(
  *  - Memory shrinkers queued the inactivation worker and it hasn't finished.
  *  - The queue depth exceeds the maximum allowable percpu backlog.
  *
- * Note: If the current thread is running a transaction, we don't ever want to
- * wait for other transactions because that could introduce a deadlock.
+ * Note: If we are in a NOFS context here (e.g. current thread is running a
+ * transaction) the we don't want to block here as inodegc progress may require
+ * filesystem resources we hold to make progress and that could result in a
+ * deadlock. Hence we skip out of here if we are in a scoped NOFS context.
  */
 static inline bool
 xfs_inodegc_want_flush_work(
@@ -2039,7 +2041,7 @@ xfs_inodegc_want_flush_work(
 	unsigned int		items,
 	unsigned int		shrinker_hits)
 {
-	if (current->journal_info)
+	if (current->flags & PF_MEMALLOC_NOFS)
 		return false;
 
 	if (shrinker_hits > 0)
fs/xfs/xfs_trans.h +1 -8
index c6d0795085a3..2bd673715ace 100644
@@ -269,19 +269,14 @@ static inline void
 xfs_trans_set_context(
 	struct xfs_trans	*tp)
 {
-	ASSERT(current->journal_info == NULL);
 	tp->t_pflags = memalloc_nofs_save();
-	current->journal_info = tp;
 }
 
 static inline void
 xfs_trans_clear_context(
 	struct xfs_trans	*tp)
 {
-	if (current->journal_info == tp) {
-		memalloc_nofs_restore(tp->t_pflags);
-		current->journal_info = NULL;
-	}
+	memalloc_nofs_restore(tp->t_pflags);
 }
 
 static inline void
@@ -289,10 +284,8 @@ xfs_trans_switch_context(
 	struct xfs_trans	*old_tp,
 	struct xfs_trans	*new_tp)
 {
-	ASSERT(current->journal_info == old_tp);
 	new_tp->t_pflags = old_tp->t_pflags;
 	old_tp->t_pflags = 0;
-	current->journal_info = new_tp;
 }
 
 #endif	/* __XFS_TRANS_H__ */
Signature · 1 line
2.43.0
#1 lore raw
Quoted 28 lines · Dave Chinner

From: Dave Chinner <[email protected]>

syzbot reported an ext4 panic during a page fault where found a journal handle when it didn't expect to find one. The structure it tripped over had a value of 'TRAN' in the first entry in the structure, and that indicates it tripped over a struct xfs_trans instead of a jbd2 handle.

The reason for this is that the page fault was taken during a copy-out to a user buffer from an xfs bulkstat operation. XFS uses an "empty" transaction context for bulkstat to do automated metadata buffer cleanup, and so the transaction context is valid across the copyout of the bulkstat info into the user buffer.

We are using empty transaction contexts like this in XFS in more places to reduce the risk of failing to release objects we reference during the operation, especially during error handling. Hence we really need to ensure that we can take page faults from these contexts without leaving landmines for the code processing the page fault to trip over.

We really only use current->journal_info for a single warning check in xfs_vm_writepages() to ensure we aren't doing writeback from a transaction context. Writeback might need to do allocation, so it can need to run transactions itself. Hence it's a debug check to warn us that we've done something silly, and largely it is not all that useful.

Quoted 4 lines

So let's just remove all the use of current->journal_info in XFS and get rid of all the potential issues from nested contexts where current->journal_info might get misused by another filesytsem context.

I wonder if this is too much, though?

Conceptually, I'd rather we set current->journal_info to some random number whenever we start a !NO_WRITECOUNT (aka a non-empty) transaction and clear it during transaction commit/cancel. That way, *we* can catch the case where some other filesystem starts a transaction and accidentally bounces into an updating write fault inside XFS.

That might be outweighed by the weird semantics of ext4 where the zeroness of journal_info changes its behavior in ways I don't want to understand. Thoughts?

--D

Quoted 108 lines
---
 fs/xfs/scrub/common.c | 4 +---
 fs/xfs/xfs_aops.c     | 7 -------
 fs/xfs/xfs_icache.c   | 8 +++++---
 fs/xfs/xfs_trans.h    | 9 +--------
 4 files changed, 7 insertions(+), 21 deletions(-)
fs/xfs/scrub/common.c +1 -3
index 81f2b96bb5a7..d853348a48c8 100644
@@ -1000,9 +1000,7 @@ xchk_irele(
 	struct xfs_scrub	*sc,
 	struct xfs_inode	*ip)
 {
-	if (current->journal_info != NULL) {
-		ASSERT(current->journal_info == sc->tp);
-
+	if (sc->tp) {
 		/*
 		 * If we are in a transaction, we /cannot/ drop the inode
 		 * ourselves, because the VFS will trigger writeback, which
fs/xfs/xfs_aops.c +0 -7
index 813f85156b0c..bc3b649d29c4 100644
@@ -502,13 +502,6 @@ xfs_vm_writepages(
 {
 	struct xfs_writepage_ctx wpc = { };
 
-	/*
-	 * Writing back data in a transaction context can result in recursive
-	 * transactions. This is bad, so issue a warning and get out of here.
-	 */
-	if (WARN_ON_ONCE(current->journal_info))
-		return 0;
-
 	xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
 	return iomap_writepages(mapping, wbc, &wpc.ctx, &xfs_writeback_ops);
 }
fs/xfs/xfs_icache.c +5 -3
index 06046827b5fe..9b966af7d55c 100644
@@ -2030,8 +2030,10 @@ xfs_inodegc_want_queue_work(
  *  - Memory shrinkers queued the inactivation worker and it hasn't finished.
  *  - The queue depth exceeds the maximum allowable percpu backlog.
  *
- * Note: If the current thread is running a transaction, we don't ever want to
- * wait for other transactions because that could introduce a deadlock.
+ * Note: If we are in a NOFS context here (e.g. current thread is running a
+ * transaction) the we don't want to block here as inodegc progress may require
+ * filesystem resources we hold to make progress and that could result in a
+ * deadlock. Hence we skip out of here if we are in a scoped NOFS context.
  */
 static inline bool
 xfs_inodegc_want_flush_work(
@@ -2039,7 +2041,7 @@ xfs_inodegc_want_flush_work(
 	unsigned int		items,
 	unsigned int		shrinker_hits)
 {
-	if (current->journal_info)
+	if (current->flags & PF_MEMALLOC_NOFS)
 		return false;
 
 	if (shrinker_hits > 0)
fs/xfs/xfs_trans.h +1 -8
index c6d0795085a3..2bd673715ace 100644
@@ -269,19 +269,14 @@ static inline void
 xfs_trans_set_context(
 	struct xfs_trans	*tp)
 {
-	ASSERT(current->journal_info == NULL);
 	tp->t_pflags = memalloc_nofs_save();
-	current->journal_info = tp;
 }
 
 static inline void
 xfs_trans_clear_context(
 	struct xfs_trans	*tp)
 {
-	if (current->journal_info == tp) {
-		memalloc_nofs_restore(tp->t_pflags);
-		current->journal_info = NULL;
-	}
+	memalloc_nofs_restore(tp->t_pflags);
 }
 
 static inline void
@@ -289,10 +284,8 @@ xfs_trans_switch_context(
 	struct xfs_trans	*old_tp,
 	struct xfs_trans	*new_tp)
 {
-	ASSERT(current->journal_info == old_tp);
 	new_tp->t_pflags = old_tp->t_pflags;
 	old_tp->t_pflags = 0;
-	current->journal_info = new_tp;
 }
 
 #endif	/* __XFS_TRANS_H__ */
Signature · 1 line
2.43.0
#2 lore raw
Quoted 29 lines · Darrick J. Wong
Quoted 28 lines · Dave Chinner

From: Dave Chinner <[email protected]>

syzbot reported an ext4 panic during a page fault where found a journal handle when it didn't expect to find one. The structure it tripped over had a value of 'TRAN' in the first entry in the structure, and that indicates it tripped over a struct xfs_trans instead of a jbd2 handle.

The reason for this is that the page fault was taken during a copy-out to a user buffer from an xfs bulkstat operation. XFS uses an "empty" transaction context for bulkstat to do automated metadata buffer cleanup, and so the transaction context is valid across the copyout of the bulkstat info into the user buffer.

We are using empty transaction contexts like this in XFS in more places to reduce the risk of failing to release objects we reference during the operation, especially during error handling. Hence we really need to ensure that we can take page faults from these contexts without leaving landmines for the code processing the page fault to trip over.

We really only use current->journal_info for a single warning check in xfs_vm_writepages() to ensure we aren't doing writeback from a transaction context. Writeback might need to do allocation, so it can need to run transactions itself. Hence it's a debug check to warn us that we've done something silly, and largely it is not all that useful.

Quoted 6 lines
Quoted 4 lines

So let's just remove all the use of current->journal_info in XFS and get rid of all the potential issues from nested contexts where current->journal_info might get misused by another filesytsem context.

I wonder if this is too much, though?

We ran XFS for 15+ years without setting current->journal_info, so I don't see it as a necessary feature...

Quoted 5 lines

Conceptually, I'd rather we set current->journal_info to some random number whenever we start a !NO_WRITECOUNT (aka a non-empty) transaction and clear it during transaction commit/cancel. That way, *we* can catch the case where some other filesystem starts a transaction and accidentally bounces into an updating write fault inside XFS.

I could just leave the ASSERT(current->journal_info == NULL); in xfs_trans_set_context() and we would catch that case. But, really, having a page fault from some other filesystem bounce into XFS where we then have to run a transaction isn't a bug at all.

The problem is purely that we now have two different contexts that now think they own current->journal_info. IOWs, no filesystem can allow page faults while current->journal_info is set by the filesystem because the page fault processing might use current->journal_info itself.

If we end up with nested XFS transactions from a page fault whilst holding an empty transaction, then it isn't an issue as the outer transaction does not hold a log reservation. The only problem that might occur is a deadlock if the page fault tries to take the same locks the upper context holds, but that's not a problem that setting current->journal_info would solve, anyway...

Hence if XFS doesn't use current->journal_info, then we just don't care what context we are running the transaction in, above or below the fault.

Quoted 3 lines

That might be outweighed by the weird semantics of ext4 where the zeroness of journal_info changes its behavior in ways I don't want to understand. Thoughts?

That's exactly the problem we're trying to avoid. Either we don't allow faults in (empty) transaction context, or we don't use current->journal_info. I prefer the latter as it gives us much more implementation freedom with empty transaction contexts..

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
Reviewed-by Mark Tinguely on v1 1/1
#4 lore raw

Re: [External] : [PATCH] xfs: don't use current->journal_info

In reply to Dave Chinner
Quoted 28 lines · Dave Chinner

From: Dave Chinner <[email protected]>

syzbot reported an ext4 panic during a page fault where found a journal handle when it didn't expect to find one. The structure it tripped over had a value of 'TRAN' in the first entry in the structure, and that indicates it tripped over a struct xfs_trans instead of a jbd2 handle.

The reason for this is that the page fault was taken during a copy-out to a user buffer from an xfs bulkstat operation. XFS uses an "empty" transaction context for bulkstat to do automated metadata buffer cleanup, and so the transaction context is valid across the copyout of the bulkstat info into the user buffer.

We are using empty transaction contexts like this in XFS in more places to reduce the risk of failing to release objects we reference during the operation, especially during error handling. Hence we really need to ensure that we can take page faults from these contexts without leaving landmines for the code processing the page fault to trip over.

We really only use current->journal_info for a single warning check in xfs_vm_writepages() to ensure we aren't doing writeback from a transaction context. Writeback might need to do allocation, so it can need to run transactions itself. Hence it's a debug check to warn us that we've done something silly, and largely it is not all that useful.

Quoted 8 lines

So let's just remove all the use of current->journal_info in XFS and get rid of all the potential issues from nested contexts where current->journal_info might get misused by another filesytsem context.

---

This will also address a problem seen by asyzkaller generated test where  a multithreaded test consisting of XFS_IOC_BULKSTAT and buffered write can unnecessarily trigger the warning in xfs_vm_writepages(). I was thinking of conditionally removing the I_DONTCACHE in xfs_bulkstat_one_int()  but cannot recreate the problem without cheating (forcing the race to happen abnormally).

Reviewed-by
#3 lore raw
In reply to Dave Chinner
Quoted 35 lines · Dave Chinner
Quoted 34 lines · Darrick J. Wong
Quoted 32 lines · Dave Chinner

From: Dave Chinner <[email protected]>

syzbot reported an ext4 panic during a page fault where found a journal handle when it didn't expect to find one. The structure it tripped over had a value of 'TRAN' in the first entry in the structure, and that indicates it tripped over a struct xfs_trans instead of a jbd2 handle.

The reason for this is that the page fault was taken during a copy-out to a user buffer from an xfs bulkstat operation. XFS uses an "empty" transaction context for bulkstat to do automated metadata buffer cleanup, and so the transaction context is valid across the copyout of the bulkstat info into the user buffer.

We are using empty transaction contexts like this in XFS in more places to reduce the risk of failing to release objects we reference during the operation, especially during error handling. Hence we really need to ensure that we can take page faults from these contexts without leaving landmines for the code processing the page fault to trip over.

We really only use current->journal_info for a single warning check in xfs_vm_writepages() to ensure we aren't doing writeback from a transaction context. Writeback might need to do allocation, so it can need to run transactions itself. Hence it's a debug check to warn us that we've done something silly, and largely it is not all that useful.

So let's just remove all the use of current->journal_info in XFS and get rid of all the potential issues from nested contexts where current->journal_info might get misused by another filesytsem context.

Quoted 4 lines
Quoted 1 line

I wonder if this is too much, though?

We ran XFS for 15+ years without setting current->journal_info, so I don't see it as a necessary feature...

Indeed, I don't see it as a necessary feature for XFS, merely a "bs coming in from other parts of the kernel" detector. And boy howdy did I ever find bs.

Quoted 17 lines
Quoted 5 lines

Conceptually, I'd rather we set current->journal_info to some random number whenever we start a !NO_WRITECOUNT (aka a non-empty) transaction and clear it during transaction commit/cancel. That way, *we* can catch the case where some other filesystem starts a transaction and accidentally bounces into an updating write fault inside XFS.

I could just leave the ASSERT(current->journal_info == NULL); in xfs_trans_set_context() and we would catch that case. But, really, having a page fault from some other filesystem bounce into XFS where we then have to run a transaction isn't a bug at all.

The problem is purely that we now have two different contexts that now think they own current->journal_info. IOWs, no filesystem can allow page faults while current->journal_info is set by the filesystem because the page fault processing might use current->journal_info itself.

Quoted 6 lines

If we end up with nested XFS transactions from a page fault whilst holding an empty transaction, then it isn't an issue as the outer transaction does not hold a log reservation. The only problem that might occur is a deadlock if the page fault tries to take the same locks the upper context holds, but that's not a problem that setting current->journal_info would solve, anyway...

This, however, is a much better justification (IMHO) for removing the places where xfs touches journal_info. Would you mind adding that to the commit message?

Quoted 4 lines

Hence if XFS doesn't use current->journal_info, then we just don't care what context we are running the transaction in, above or below the fault.

Quoted 8 lines
Quoted 3 lines

That might be outweighed by the weird semantics of ext4 where the zeroness of journal_info changes its behavior in ways I don't want to understand. Thoughts?

That's exactly the problem we're trying to avoid. Either we don't allow faults in (empty) transaction context, or we don't use current->journal_info. I prefer the latter as it gives us much more implementation freedom with empty transaction contexts..

Heh. I also found much much more bs and shenanigans of exactly the type you'd expect from a global(ish) void pointer.

With the commit message amended, you can add:

--D

Quoted 5 lines

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]