Skip to contentSkip to conversation

xfs: avoid transaction reservation recursion

PATCH Superseded Yafang Shao 3 participants 11 messages Dec 9, 2020 – Dec 15, 2020

Subsystems (6)

v12 of 15 Superseded Latest is v15 →

lore
Share & tools
lore
#0 lore raw

v12 cover xfs: avoid transaction reservation recursion

PF_FSTRANS which is used to avoid transaction reservation recursion, is dropped since commit 9070733b4efa ("xfs: abstract PF_FSTRANS to PF_MEMALLOC_NOFS") and commit 7dea19f9ee63 ("mm: introduce memalloc_nofs_{save,restore} API"), and replaced by PF_MEMALLOC_NOFS which means to avoid filesystem reclaim recursion.

As these two flags have different meanings, we'd better reintroduce PF_FSTRANS back. To avoid wasting the space of PF_* flags in task_struct, we can reuse the current->journal_info to do that, per Willy. As the check of transaction reservation recursion is used by XFS only, we can move the check into xfs_vm_writepage(s), per Dave.

Patch #1 and #2 are to use the memalloc_nofs_{save,restore} API Patch #1 is picked form Willy's patchset "Overhaul memalloc_no*"[1]

Patch #3 is the refactor of xfs_trans context, which is activated when xfs_trans is allocated and deactivated when xfs_trans is freed.

Patch #4 is the implementation of reussing current->journal_info to avoid transaction reservation recursion.

No obvious error occurred after running xfstests.

[1]. https://lore.kernel.org/linux-mm/[email protected]

v12:
Per Darrick's suggestion,
- add the check before calling xfs_trans_context_clear() in xfs_trans_context_free().
- move t_pflags into xfs_trans_context_swap()

v11:
- add the warning at the callsite of xfs_trans_context_active()
- improve the commit log of patch #2

v10:
- refactor the code, per Dave.

v9:
- rebase it on xfs tree.
- Darrick fixed an error occurred in xfs/141
- run xfstests, and no obvious error occurred.

v8:
- check xfs_trans_context_active() in xfs_vm_writepage(s), per Dave.

v7:
- check fstrans recursion for XFS only, by introducing a new member in struct writeback_control.

v6:
- add Michal's ack and comment in patch #1.

v5:
- pick one of Willy's patch
- introduce four new helpers, per Dave

v4:
- retitle from "xfs: introduce task->in_fstrans for transaction reservation recursion protection"
- reuse current->journal_info, per Willy

Matthew Wilcox (Oracle) (1):
mm: Add become_kswapd and restore_kswapd

Yafang Shao (3):
xfs: use memalloc_nofs_{save,restore} in xfs transaction
xfs: refactor the usage around xfs_trans_context_{set,clear}
xfs: use current->journal_info to avoid transaction reservation
recursion

 fs/iomap/buffered-io.c    |  7 -------
 fs/xfs/libxfs/xfs_btree.c | 14 ++++++++------
 fs/xfs/xfs_aops.c         | 21 +++++++++++++++++++--
 fs/xfs/xfs_linux.h        |  4 ----
 fs/xfs/xfs_trans.c        | 33 +++++++++++++++++++--------------
 fs/xfs/xfs_trans.h        | 35 +++++++++++++++++++++++++++++++++++
 include/linux/sched/mm.h  | 23 +++++++++++++++++++++++
 mm/vmscan.c               | 16 +---------------
 8 files changed, 105 insertions(+), 48 deletions(-)
Signature · 1 line
2.18.4
#1 lore raw

PATCH v12 1/4 mm: Add become_kswapd and restore_kswapd

From: "Matthew Wilcox (Oracle)" <[email protected]>

Since XFS needs to pretend to be kswapd in some of its worker threads, create methods to save & restore kswapd state. Don't bother restoring kswapd state in kswapd -- the only time we reach this code is when we're exiting and the task_struct is about to be destroyed anyway.

---
 fs/xfs/libxfs/xfs_btree.c | 14 ++++++++------
 include/linux/sched/mm.h  | 23 +++++++++++++++++++++++
 mm/vmscan.c               | 16 +---------------
 3 files changed, 32 insertions(+), 21 deletions(-)
fs/xfs/libxfs/xfs_btree.c +8 -6
index 2d25bab68764..a04a44238aab 100644
@@ -2813,8 +2813,9 @@ xfs_btree_split_worker(
 {
 	struct xfs_btree_split_args	*args = container_of(work,
 						struct xfs_btree_split_args, work);
+	bool			is_kswapd = args->kswapd;
 	unsigned long		pflags;
-	unsigned long		new_pflags = PF_MEMALLOC_NOFS;
+	int			memalloc_nofs;
 
 	/*
 	 * we are in a transaction context here, but may also be doing work
@@ -2822,16 +2823,17 @@ xfs_btree_split_worker(
 	 * temporarily to ensure that we don't block waiting for memory reclaim
 	 * in any way.
 	 */
-	if (args->kswapd)
-		new_pflags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
-
-	current_set_flags_nested(&pflags, new_pflags);
+	if (is_kswapd)
+		pflags = become_kswapd();
+	memalloc_nofs = memalloc_nofs_save();
 
 	args->result = __xfs_btree_split(args->cur, args->level, args->ptrp,
 					 args->key, args->curp, args->stat);
 	complete(args->done);
 
-	current_restore_flags_nested(&pflags, new_pflags);
+	memalloc_nofs_restore(memalloc_nofs);
+	if (is_kswapd)
+		restore_kswapd(pflags);
 }
 
 /*
include/linux/sched/mm.h +23 -0
index d5ece7a9a403..2faf03e79a1e 100644
@@ -278,6 +278,29 @@ static inline void memalloc_nocma_restore(unsigned int flags)
 }
 #endif
 
+/*
+ * Tell the memory management code that this thread is working on behalf
+ * of background memory reclaim (like kswapd).  That means that it will
+ * get access to memory reserves should it need to allocate memory in
+ * order to make forward progress.  With this great power comes great
+ * responsibility to not exhaust those reserves.
+ */
+#define KSWAPD_PF_FLAGS		(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD)
+
+static inline unsigned long become_kswapd(void)
+{
+	unsigned long flags = current->flags & KSWAPD_PF_FLAGS;
+
+	current->flags |= KSWAPD_PF_FLAGS;
+
+	return flags;
+}
+
+static inline void restore_kswapd(unsigned long flags)
+{
+	current->flags &= ~(flags ^ KSWAPD_PF_FLAGS);
+}
+
 #ifdef CONFIG_MEMCG
 DECLARE_PER_CPU(struct mem_cgroup *, int_active_memcg);
 /**
mm/vmscan.c +1 -15
index 1b8f0e059767..77bc1dda75bf 100644
@@ -3869,19 +3869,7 @@ static int kswapd(void *p)
 	if (!cpumask_empty(cpumask))
 		set_cpus_allowed_ptr(tsk, cpumask);
 
-	/*
-	 * Tell the memory management that we're a "memory allocator",
-	 * and that if we need more memory we should get access to it
-	 * regardless (see "__alloc_pages()"). "kswapd" should
-	 * never get caught in the normal page freeing logic.
-	 *
-	 * (Kswapd normally doesn't need memory anyway, but sometimes
-	 * you need a small amount of memory in order to be able to
-	 * page out something else, and this flag essentially protects
-	 * us from recursively trying to free more memory as we're
-	 * trying to free the first piece of memory in the first place).
-	 */
-	tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
+	become_kswapd();
 	set_freezable();
 
 	WRITE_ONCE(pgdat->kswapd_order, 0);
@@ -3931,8 +3919,6 @@ static int kswapd(void *p)
 			goto kswapd_try_sleep;
 	}
 
-	tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
-
 	return 0;
 }
 
Signature · 1 line
2.18.4
#2 lore raw

PATCH v12 2/4 xfs: use memalloc_nofs_{save,restore} in xfs transaction

In reply to Yafang Shao

Introduce a new API to mark the start and end of XFS transactions. For now, just save and restore the memalloc_nofs flags.

The new helpers as follows,
- xfs_trans_context_set
Mark the start of XFS transactions
- xfs_trans_context_clear
Mark the end of XFS transactions

Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
Cc Michal Hocko <[email protected]>
---
 fs/xfs/xfs_aops.c  |  4 ++--
 fs/xfs/xfs_linux.h |  4 ----
 fs/xfs/xfs_trans.c | 13 +++++++------
 fs/xfs/xfs_trans.h | 12 ++++++++++++
 4 files changed, 21 insertions(+), 12 deletions(-)
fs/xfs/xfs_aops.c +2 -2
index 4304c6416fbb..2371187b7615 100644
@@ -62,7 +62,7 @@ xfs_setfilesize_trans_alloc(
 	 * We hand off the transaction to the completion thread now, so
 	 * clear the flag here.
 	 */
-	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_clear(tp);
 	return 0;
 }
 
@@ -125,7 +125,7 @@ xfs_setfilesize_ioend(
 	 * thus we need to mark ourselves as being in a transaction manually.
 	 * Similarly for freeze protection.
 	 */
-	current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_set(tp);
 	__sb_writers_acquired(VFS_I(ip)->i_sb, SB_FREEZE_FS);
 
 	/* we abort the update if there was an IO error */
fs/xfs/xfs_linux.h +0 -4
index 5b7a1e201559..6ab0f8043c73 100644
@@ -102,10 +102,6 @@ typedef __u32			xfs_nlink_t;
 #define xfs_cowb_secs		xfs_params.cowb_timer.val
 
 #define current_cpu()		(raw_smp_processor_id())
-#define current_set_flags_nested(sp, f)		\
-		(*(sp) = current->flags, current->flags |= (f))
-#define current_restore_flags_nested(sp, f)	\
-		(current->flags = ((current->flags & ~(f)) | (*(sp) & (f))))
 
 #define NBBY		8		/* number of bits per byte */
 
fs/xfs/xfs_trans.c +7 -6
index c94e71f741b6..11d390f0d3f2 100644
@@ -154,7 +154,7 @@ xfs_trans_reserve(
 	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
 	/* Mark this thread as being in a transaction */
-	current_set_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_set(tp);
 
 	/*
 	 * Attempt to reserve the needed disk blocks by decrementing
@@ -164,7 +164,7 @@ xfs_trans_reserve(
 	if (blocks > 0) {
 		error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
 		if (error != 0) {
-			current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+			xfs_trans_context_clear(tp);
 			return -ENOSPC;
 		}
 		tp->t_blk_res += blocks;
@@ -241,7 +241,7 @@ xfs_trans_reserve(
 		tp->t_blk_res = 0;
 	}
 
-	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_clear(tp);
 
 	return error;
 }
@@ -878,7 +878,7 @@ __xfs_trans_commit(
 
 	xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
 
-	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_clear(tp);
 	xfs_trans_free(tp);
 
 	/*
@@ -910,7 +910,8 @@ __xfs_trans_commit(
 			xfs_log_ticket_ungrant(mp->m_log, tp->t_ticket);
 		tp->t_ticket = NULL;
 	}
-	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+
+	xfs_trans_context_clear(tp);
 	xfs_trans_free_items(tp, !!error);
 	xfs_trans_free(tp);
 
@@ -971,7 +972,7 @@ xfs_trans_cancel(
 	}
 
 	/* mark this thread as no longer being in a transaction */
-	current_restore_flags_nested(&tp->t_pflags, PF_MEMALLOC_NOFS);
+	xfs_trans_context_clear(tp);
 
 	xfs_trans_free_items(tp, dirty);
 	xfs_trans_free(tp);
fs/xfs/xfs_trans.h +12 -0
index 084658946cc8..44b11c64a15e 100644
@@ -268,4 +268,16 @@ xfs_trans_item_relog(
 	return lip->li_ops->iop_relog(lip, tp);
 }
 
+static inline void
+xfs_trans_context_set(struct xfs_trans *tp)
+{
+	tp->t_pflags = memalloc_nofs_save();
+}
+
+static inline void
+xfs_trans_context_clear(struct xfs_trans *tp)
+{
+	memalloc_nofs_restore(tp->t_pflags);
+}
+
 #endif	/* __XFS_TRANS_H__ */
Signature · 1 line
2.18.4
#3 lore raw

PATCH v12 3/4 xfs: refactor the usage around xfs_trans_context_{set,clear}

In reply to Yafang Shao

The xfs_trans context should be active after it is allocated, and deactive when it is freed.

So these two helpers are refactored as,
- xfs_trans_context_set()
Used in xfs_trans_alloc()
- xfs_trans_context_clear()
Used in xfs_trans_free()

This patch is based on Darrick's work to fix the issue in xfs/141 in the earlier version. [1]

1. https://lore.kernel.org/linux-xfs/20201104001649.GN7123@magnolia

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
---
 fs/xfs/xfs_trans.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
fs/xfs/xfs_trans.c +15 -13
index 11d390f0d3f2..4f4645329bb2 100644
@@ -67,6 +67,17 @@ xfs_trans_free(
 	xfs_extent_busy_sort(&tp->t_busy);
 	xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);
 
+
+	/* Detach the transaction from this thread. */
+	ASSERT(current->journal_info != NULL);
+	/*
+	 * The PF_MEMALLOC_NOFS is bound to the transaction itself instead
+	 * of the reservation, so we need to check if tp is still the
+	 * current transaction before clearing the flag.
+	 */
+	if (current->journal_info == tp)
+		xfs_trans_context_clear(tp);
+
 	trace_xfs_trans_free(tp, _RET_IP_);
 	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
 		sb_end_intwrite(tp->t_mountp->m_super);
@@ -153,9 +164,6 @@ xfs_trans_reserve(
 	int			error = 0;
 	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
-	/* Mark this thread as being in a transaction */
-	xfs_trans_context_set(tp);
-
 	/*
 	 * Attempt to reserve the needed disk blocks by decrementing
 	 * the number needed from the number available.  This will
@@ -163,10 +171,9 @@ xfs_trans_reserve(
 	 */
 	if (blocks > 0) {
 		error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
-		if (error != 0) {
-			xfs_trans_context_clear(tp);
+		if (error != 0)
 			return -ENOSPC;
-		}
+
 		tp->t_blk_res += blocks;
 	}
 
@@ -241,8 +248,6 @@ xfs_trans_reserve(
 		tp->t_blk_res = 0;
 	}
 
-	xfs_trans_context_clear(tp);
-
 	return error;
 }
 
@@ -284,6 +289,8 @@ xfs_trans_alloc(
 	INIT_LIST_HEAD(&tp->t_dfops);
 	tp->t_firstblock = NULLFSBLOCK;
 
+	/* Mark this thread as being in a transaction */
+	xfs_trans_context_set(tp);
 	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
 	if (error) {
 		xfs_trans_cancel(tp);
@@ -878,7 +885,6 @@ __xfs_trans_commit(
 
 	xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
 
-	xfs_trans_context_clear(tp);
 	xfs_trans_free(tp);
 
 	/*
@@ -911,7 +917,6 @@ __xfs_trans_commit(
 		tp->t_ticket = NULL;
 	}
 
-	xfs_trans_context_clear(tp);
 	xfs_trans_free_items(tp, !!error);
 	xfs_trans_free(tp);
 
@@ -971,9 +976,6 @@ xfs_trans_cancel(
 		tp->t_ticket = NULL;
 	}
 
-	/* mark this thread as no longer being in a transaction */
-	xfs_trans_context_clear(tp);
-
 	xfs_trans_free_items(tp, dirty);
 	xfs_trans_free(tp);
 }
Signature · 1 line
2.18.4
#4 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 30 lines · Yafang Shao

The xfs_trans context should be active after it is allocated, and deactive when it is freed.

So these two helpers are refactored as,
- xfs_trans_context_set()
Used in xfs_trans_alloc()
- xfs_trans_context_clear()
Used in xfs_trans_free()

This patch is based on Darrick's work to fix the issue in xfs/141 in the earlier version. [1]

1. https://lore.kernel.org/linux-xfs/20201104001649.GN7123@magnolia

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
---
 fs/xfs/xfs_trans.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
fs/xfs/xfs_trans.c +0 -0
index 11d390f0d3f2..4f4645329bb2 100644
@@ -67,6 +67,17 @@ xfs_trans_free(
 	xfs_extent_busy_sort(&tp->t_busy);
 	xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);

fs/xfs/xfs_trans.c xfs_trans_free( · line 78

Quoted 10 lines
fs/xfs/xfs_trans.c +9 -0
 
+
+	/* Detach the transaction from this thread. */
+	ASSERT(current->journal_info != NULL);
+	/*
+	 * The PF_MEMALLOC_NOFS is bound to the transaction itself instead
+	 * of the reservation, so we need to check if tp is still the
+	 * current transaction before clearing the flag.
+	 */
+	if (current->journal_info == tp)

Um, you don't start setting journal_info until the next patch, so this means that someone who lands on this commit with git bisect will have a xfs with broken logic.

Because this is the patch that changes where we set and restore NOFS context, I think you have to introduce xfs_trans_context_swap here, and not in the next patch.

I also think the _swap routine has to move the old NOFS state to the new transaction's t_pflags, and then set NOFS in the old transaction's t_pflags so that when we clear the context on the old transaction we don't actually change the thread's NOFS state.

--D

Quoted 75 lines
fs/xfs/xfs_trans.c +6 -13
+		xfs_trans_context_clear(tp);
+
 	trace_xfs_trans_free(tp, _RET_IP_);
 	if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
 		sb_end_intwrite(tp->t_mountp->m_super);
@@ -153,9 +164,6 @@ xfs_trans_reserve(
 	int			error = 0;
 	bool			rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
-	/* Mark this thread as being in a transaction */
-	xfs_trans_context_set(tp);
-
 	/*
 	 * Attempt to reserve the needed disk blocks by decrementing
 	 * the number needed from the number available.  This will
@@ -163,10 +171,9 @@ xfs_trans_reserve(
 	 */
 	if (blocks > 0) {
 		error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
-		if (error != 0) {
-			xfs_trans_context_clear(tp);
+		if (error != 0)
 			return -ENOSPC;
-		}
+
 		tp->t_blk_res += blocks;
 	}
 
@@ -241,8 +248,6 @@ xfs_trans_reserve(
 		tp->t_blk_res = 0;
 	}
 
-	xfs_trans_context_clear(tp);
-
 	return error;
 }
 
@@ -284,6 +289,8 @@ xfs_trans_alloc(
 	INIT_LIST_HEAD(&tp->t_dfops);
 	tp->t_firstblock = NULLFSBLOCK;
 
+	/* Mark this thread as being in a transaction */
+	xfs_trans_context_set(tp);
 	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
 	if (error) {
 		xfs_trans_cancel(tp);
@@ -878,7 +885,6 @@ __xfs_trans_commit(
 
 	xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
 
-	xfs_trans_context_clear(tp);
 	xfs_trans_free(tp);
 
 	/*
@@ -911,7 +917,6 @@ __xfs_trans_commit(
 		tp->t_ticket = NULL;
 	}
 
-	xfs_trans_context_clear(tp);
 	xfs_trans_free_items(tp, !!error);
 	xfs_trans_free(tp);
 
@@ -971,9 +976,6 @@ xfs_trans_cancel(
 		tp->t_ticket = NULL;
 	}
 
-	/* mark this thread as no longer being in a transaction */
-	xfs_trans_context_clear(tp);
-
 	xfs_trans_free_items(tp, dirty);
 	xfs_trans_free(tp);
 }
Signature · 1 line
2.18.4
#5 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 47 lines · Darrick J. Wong
Quoted 30 lines · Yafang Shao

The xfs_trans context should be active after it is allocated, and deactive when it is freed.

So these two helpers are refactored as,
- xfs_trans_context_set()
Used in xfs_trans_alloc()
- xfs_trans_context_clear()
Used in xfs_trans_free()

This patch is based on Darrick's work to fix the issue in xfs/141 in the earlier version. [1]

1. https://lore.kernel.org/linux-xfs/20201104001649.GN7123@magnolia

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
---
 fs/xfs/xfs_trans.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
fs/xfs/xfs_trans.c +0 -0
index 11d390f0d3f2..4f4645329bb2 100644
@@ -67,6 +67,17 @@ xfs_trans_free(
      xfs_extent_busy_sort(&tp->t_busy);
      xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);

fs/xfs/xfs_trans.c xfs_trans_free( · line 78

Quoted 10 lines
fs/xfs/xfs_trans.c +9 -0
 
+
+     /* Detach the transaction from this thread. */
+     ASSERT(current->journal_info != NULL);
+     /*
+      * The PF_MEMALLOC_NOFS is bound to the transaction itself instead
+      * of the reservation, so we need to check if tp is still the
+      * current transaction before clearing the flag.
+      */
+     if (current->journal_info == tp)

Um, you don't start setting journal_info until the next patch, so this means that someone who lands on this commit with git bisect will have a xfs with broken logic.

Quoted 4 lines

Because this is the patch that changes where we set and restore NOFS context, I think you have to introduce xfs_trans_context_swap here, and not in the next patch.

Thanks for the review. I will change it in the next version.

Quoted 2 lines

I also think the _swap routine has to move the old NOFS state to the new transaction's t_pflags,

Sure

Quoted 4 lines

and then set NOFS in the old transaction's
t_pflags so that when we clear the context on the old transaction we don't actually change the thread's NOFS state.

Both thread's NOFS state and thead's journal_info state can't be changed in that case, right ?
So should it better be,

__xfs_trans_commit(tp, regrant)
xfs_trans_free(tp, regrant)
if (!regrant). // don't clear the xfs_trans_context if
regrant is true.
xfs_trans_context_clear()

Quoted 77 lines

--D

Quoted 75 lines
fs/xfs/xfs_trans.c +6 -14
+             xfs_trans_context_clear(tp);
+
      trace_xfs_trans_free(tp, _RET_IP_);
      if (!(tp->t_flags & XFS_TRANS_NO_WRITECOUNT))
              sb_end_intwrite(tp->t_mountp->m_super);
@@ -153,9 +164,6 @@ xfs_trans_reserve(
      int                     error = 0;
      bool                    rsvd = (tp->t_flags & XFS_TRANS_RESERVE) != 0;
 
-     /* Mark this thread as being in a transaction */
-     xfs_trans_context_set(tp);
-
      /*
       * Attempt to reserve the needed disk blocks by decrementing
       * the number needed from the number available.  This will
@@ -163,10 +171,9 @@ xfs_trans_reserve(
       */
      if (blocks > 0) {
              error = xfs_mod_fdblocks(mp, -((int64_t)blocks), rsvd);
-             if (error != 0) {
-                     xfs_trans_context_clear(tp);
+             if (error != 0)
                      return -ENOSPC;
-             }
+
              tp->t_blk_res += blocks;
      }
 
@@ -241,8 +248,6 @@ xfs_trans_reserve(
              tp->t_blk_res = 0;
      }
 
-     xfs_trans_context_clear(tp);
-
      return error;
 }
 
@@ -284,6 +289,8 @@ xfs_trans_alloc(
      INIT_LIST_HEAD(&tp->t_dfops);
      tp->t_firstblock = NULLFSBLOCK;
 
+     /* Mark this thread as being in a transaction */
+     xfs_trans_context_set(tp);
      error = xfs_trans_reserve(tp, resp, blocks, rtextents);
      if (error) {
              xfs_trans_cancel(tp);
@@ -878,7 +885,6 @@ __xfs_trans_commit(
 
      xfs_log_commit_cil(mp, tp, &commit_lsn, regrant);
 
-     xfs_trans_context_clear(tp);
      xfs_trans_free(tp);
 
      /*
@@ -911,7 +917,6 @@ __xfs_trans_commit(
              tp->t_ticket = NULL;
      }
 
-     xfs_trans_context_clear(tp);
      xfs_trans_free_items(tp, !!error);
      xfs_trans_free(tp);
 
@@ -971,9 +976,6 @@ xfs_trans_cancel(
              tp->t_ticket = NULL;
      }
 
-     /* mark this thread as no longer being in a transaction */
-     xfs_trans_context_clear(tp);
-
      xfs_trans_free_items(tp, dirty);
      xfs_trans_free(tp);
 }
--

2.18.4

Signature · 2 lines
Thanks
Yafang
#6 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 69 lines · Yafang Shao
Quoted 47 lines · Darrick J. Wong
Quoted 30 lines · Yafang Shao

The xfs_trans context should be active after it is allocated, and deactive when it is freed.

So these two helpers are refactored as,
- xfs_trans_context_set()
Used in xfs_trans_alloc()
- xfs_trans_context_clear()
Used in xfs_trans_free()

This patch is based on Darrick's work to fix the issue in xfs/141 in the earlier version. [1]

1. https://lore.kernel.org/linux-xfs/20201104001649.GN7123@magnolia

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
---
 fs/xfs/xfs_trans.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
fs/xfs/xfs_trans.c +0 -0
index 11d390f0d3f2..4f4645329bb2 100644
@@ -67,6 +67,17 @@ xfs_trans_free(
      xfs_extent_busy_sort(&tp->t_busy);
      xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);

fs/xfs/xfs_trans.c xfs_trans_free( · line 78

Quoted 10 lines
fs/xfs/xfs_trans.c +9 -0
 
+
+     /* Detach the transaction from this thread. */
+     ASSERT(current->journal_info != NULL);
+     /*
+      * The PF_MEMALLOC_NOFS is bound to the transaction itself instead
+      * of the reservation, so we need to check if tp is still the
+      * current transaction before clearing the flag.
+      */
+     if (current->journal_info == tp)

Um, you don't start setting journal_info until the next patch, so this means that someone who lands on this commit with git bisect will have a xfs with broken logic.

Quoted 4 lines

Because this is the patch that changes where we set and restore NOFS context, I think you have to introduce xfs_trans_context_swap here, and not in the next patch.

Thanks for the review. I will change it in the next version.

Quoted 2 lines

I also think the _swap routine has to move the old NOFS state to the new transaction's t_pflags,

Sure

Quoted 4 lines

and then set NOFS in the old transaction's
t_pflags so that when we clear the context on the old transaction we don't actually change the thread's NOFS state.

Both thread's NOFS state and thead's journal_info state can't be changed in that case, right ?
So should it better be,

Quoted 5 lines

__xfs_trans_commit(tp, regrant)
xfs_trans_free(tp, regrant)
if (!regrant). // don't clear the xfs_trans_context if
regrant is true.
xfs_trans_context_clear()

No. You are trying to make this way more complex than it needs to be. The logic in the core XFS code is *already correct* and all we need to do is move that logic to wrapper functions, then slightly modify the implementation inside the wrapper functions.

That is, xfs_trans_context_clear() should end up like this:

static inline void
xfs_trans_context_clear(struct xfs_trans *tp)

{
	/*
	 * If xfs_trans_context_swap() handed the NOFS context to a
	 * new transaction we do not clear the context here.
	 */
	if (current->journal_info != tp)
		return;
	current->journal_info = NULL;
	memalloc_nofs_restore(tp->t_pflags);
}

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#7 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 77 lines · Dave Chinner
Quoted 69 lines · Yafang Shao
Quoted 47 lines · Darrick J. Wong
Quoted 30 lines · Yafang Shao

The xfs_trans context should be active after it is allocated, and deactive when it is freed.

So these two helpers are refactored as,
- xfs_trans_context_set()
Used in xfs_trans_alloc()
- xfs_trans_context_clear()
Used in xfs_trans_free()

This patch is based on Darrick's work to fix the issue in xfs/141 in the earlier version. [1]

1. https://lore.kernel.org/linux-xfs/20201104001649.GN7123@magnolia

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
---
 fs/xfs/xfs_trans.c | 28 +++++++++++++++-------------
 1 file changed, 15 insertions(+), 13 deletions(-)
fs/xfs/xfs_trans.c +0 -0
index 11d390f0d3f2..4f4645329bb2 100644
@@ -67,6 +67,17 @@ xfs_trans_free(
      xfs_extent_busy_sort(&tp->t_busy);
      xfs_extent_busy_clear(tp->t_mountp, &tp->t_busy, false);

fs/xfs/xfs_trans.c xfs_trans_free( · line 78

Quoted 10 lines
fs/xfs/xfs_trans.c +9 -0
 
+
+     /* Detach the transaction from this thread. */
+     ASSERT(current->journal_info != NULL);
+     /*
+      * The PF_MEMALLOC_NOFS is bound to the transaction itself instead
+      * of the reservation, so we need to check if tp is still the
+      * current transaction before clearing the flag.
+      */
+     if (current->journal_info == tp)

Um, you don't start setting journal_info until the next patch, so this means that someone who lands on this commit with git bisect will have a xfs with broken logic.

Quoted 4 lines

Because this is the patch that changes where we set and restore NOFS context, I think you have to introduce xfs_trans_context_swap here, and not in the next patch.

Thanks for the review. I will change it in the next version.

Quoted 2 lines

I also think the _swap routine has to move the old NOFS state to the new transaction's t_pflags,

Sure

Quoted 4 lines

and then set NOFS in the old transaction's
t_pflags so that when we clear the context on the old transaction we don't actually change the thread's NOFS state.

Both thread's NOFS state and thead's journal_info state can't be changed in that case, right ?
So should it better be,

Quoted 5 lines

__xfs_trans_commit(tp, regrant)
xfs_trans_free(tp, regrant)
if (!regrant). // don't clear the xfs_trans_context if
regrant is true.
xfs_trans_context_clear()

Quoted 5 lines

No. You are trying to make this way more complex than it needs to be. The logic in the core XFS code is *already correct* and all we need to do is move that logic to wrapper functions, then slightly modify the implementation inside the wrapper functions.

Thanks for the explanation.

Quoted 3 lines

That is, xfs_trans_context_clear() should end up like this:

Agreed.

Quoted 8 lines

static inline void
xfs_trans_context_clear(struct xfs_trans *tp)

{
        /*
         * If xfs_trans_context_swap() handed the NOFS context to a
         * new transaction we do not clear the context here.
         */
        if (current->journal_info != tp)

current->journal_info hasn't been used in patch #3, that will make patch #3 a little more complex.
We have to do some workaround in patch #3. I will think about it.

Quoted 9 lines
                return;
        current->journal_info = NULL;
        memalloc_nofs_restore(tp->t_pflags);
}

-Dave.
--
Dave Chinner
[email protected]

Signature · 2 lines
Thanks
Yafang
#8 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 13 lines · Yafang Shao
Quoted 11 lines · Dave Chinner
Quoted 2 lines · Yafang Shao
Quoted 1 line · Darrick J. Wong

On Wed, Dec 09, 2020 at 09:11:45PM +0800, Yafang Shao wrote:

static inline void
xfs_trans_context_clear(struct xfs_trans *tp)

{
        /*
         * If xfs_trans_context_swap() handed the NOFS context to a
         * new transaction we do not clear the context here.
         */
        if (current->journal_info != tp)
Quoted 3 lines

current->journal_info hasn't been used in patch #3, that will make patch #3 a little more complex.
We have to do some workaround in patch #3. I will think about it.

What I wrote is how the function should look at the end of the patch series. Do not add the current->journal_info parts of it until the patch that introduces the current->journal_info tracking.

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#9 lore raw

Re: [PATCH v12 3/4] xfs: refactor the usage around xfs_trans_context_{set,clear}

Quoted 15 lines · Dave Chinner
Quoted 13 lines · Yafang Shao
Quoted 11 lines · Dave Chinner
Quoted 2 lines · Yafang Shao
Quoted 1 line · Darrick J. Wong

On Wed, Dec 09, 2020 at 09:11:45PM +0800, Yafang Shao wrote:

static inline void
xfs_trans_context_clear(struct xfs_trans *tp)

{
        /*
         * If xfs_trans_context_swap() handed the NOFS context to a
         * new transaction we do not clear the context here.
         */
        if (current->journal_info != tp)
Quoted 8 lines
Quoted 3 lines

current->journal_info hasn't been used in patch #3, that will make patch #3 a little more complex.
We have to do some workaround in patch #3. I will think about it.

What I wrote is how the function should look at the end of the patch series. Do not add the current->journal_info parts of it until the patch that introduces the current->journal_info tracking.

I know what you meant.
While I mean we have to do some hack, as suggested by Darrrick that "set NOFS in the old transaction's
t_pflags so that when we clear the context on the old transaction we don't actually change the thread's NOFS state." in patch #3 and then remove it in patch #4.

Signature · 2 lines
Thanks
Yafang
#10 lore raw

PATCH v12 4/4 xfs: use current->journal_info to avoid transaction reservation recursion

In reply to Yafang Shao

PF_FSTRANS which is used to avoid transaction reservation recursion, is dropped since commit 9070733b4efa ("xfs: abstract PF_FSTRANS to PF_MEMALLOC_NOFS") and commit 7dea19f9ee63 ("mm: introduce memalloc_nofs_{save,restore} API") and replaced by PF_MEMALLOC_NOFS which means to avoid filesystem reclaim recursion.

As these two flags have different meanings, we'd better reintroduce PF_FSTRANS back. To avoid wasting the space of PF_* flags in task_struct, we can reuse the current->journal_info to do that, per Willy. As the check of transaction reservation recursion is used by XFS only, we can move the check into xfs_vm_writepage(s), per Dave.

To better abstract that behavoir, two new helpers are introduced, as follows,
- xfs_trans_context_active
To check whehter current is in fs transcation or not
- xfs_trans_context_swap
Transfer the transaction context when rolling a permanent transaction

These two new helpers are instroduced in xfs_trans.h.

Cc Darrick J. Wong <[email protected]>
Cc Matthew Wilcox (Oracle) <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Dave Chinner <[email protected]>
Cc Michal Hocko <[email protected]>
Cc David Howells <[email protected]>
Cc Jeff Layton <[email protected]>
---
 fs/iomap/buffered-io.c |  7 -------
 fs/xfs/xfs_aops.c      | 17 +++++++++++++++++
 fs/xfs/xfs_trans.c     |  4 +++-
 fs/xfs/xfs_trans.h     | 23 +++++++++++++++++++++++
 4 files changed, 43 insertions(+), 8 deletions(-)
fs/iomap/buffered-io.c +0 -7
index 10cc7979ce38..3c53fa6ce64d 100644
@@ -1458,13 +1458,6 @@ iomap_do_writepage(struct page *page, struct writeback_control *wbc, void *data)
 			PF_MEMALLOC))
 		goto redirty;
 
-	/*
-	 * Given that we do not allow direct reclaim to call us, we should
-	 * never be called in a recursive filesystem reclaim context.
-	 */
-	if (WARN_ON_ONCE(current->flags & PF_MEMALLOC_NOFS))
-		goto redirty;
-
 	/*
 	 * Is this page beyond the end of the file?
 	 *
fs/xfs/xfs_aops.c +17 -0
index 2371187b7615..0da0242d42c3 100644
@@ -568,6 +568,16 @@ xfs_vm_writepage(
 {
 	struct xfs_writepage_ctx wpc = { };
 
+	/*
+	 * Given that we do not allow direct reclaim to call us, we should
+	 * never be called while in a filesystem transaction.
+	 */
+	if (WARN_ON_ONCE(xfs_trans_context_active())) {
+		redirty_page_for_writepage(wbc, page);
+		unlock_page(page);
+		return 0;
+	}
+
 	return iomap_writepage(page, wbc, &wpc.ctx, &xfs_writeback_ops);
 }
 
@@ -579,6 +589,13 @@ xfs_vm_writepages(
 	struct xfs_writepage_ctx wpc = { };
 
 	xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
+	/*
+	 * Given that we do not allow direct reclaim to call us, we should
+	 * never be called while in a filesystem transaction.
+	 */
+	if (WARN_ON_ONCE(xfs_trans_context_active()))
+		return 0;
+
 	return iomap_writepages(mapping, wbc, &wpc.ctx, &xfs_writeback_ops);
 }
 
fs/xfs/xfs_trans.c +3 -1
index 4f4645329bb2..7793391fe536 100644
@@ -130,7 +130,9 @@ xfs_trans_dup(
 
 	ntp->t_rtx_res = tp->t_rtx_res - tp->t_rtx_res_used;
 	tp->t_rtx_res = tp->t_rtx_res_used;
-	ntp->t_pflags = tp->t_pflags;
+
+	/* Associate the new transaction with this thread. */
+	xfs_trans_context_swap(tp, ntp);
 
 	/* move deferred ops over to the new tp */
 	xfs_defer_move(ntp, tp);
fs/xfs/xfs_trans.h +23 -0
index 44b11c64a15e..e994b01eeb57 100644
@@ -268,16 +268,39 @@ xfs_trans_item_relog(
 	return lip->li_ops->iop_relog(lip, tp);
 }
 
+static inline bool
+xfs_trans_context_active(void)
+{
+	/* Use journal_info to indicate current is in a transaction */
+	return current->journal_info != NULL;
+}
+
 static inline void
 xfs_trans_context_set(struct xfs_trans *tp)
 {
+	ASSERT(!current->journal_info);
+	current->journal_info = tp;
 	tp->t_pflags = memalloc_nofs_save();
 }
 
 static inline void
 xfs_trans_context_clear(struct xfs_trans *tp)
 {
+	ASSERT(current->journal_info == tp);
+	current->journal_info = NULL;
 	memalloc_nofs_restore(tp->t_pflags);
 }
 
+/*
+ * Transfer the transaction context when rolling a permanent
+ * transaction.
+ */
+static inline void
+xfs_trans_context_swap(struct xfs_trans *tp, struct xfs_trans *ntp)
+{
+	ASSERT(current->journal_info == tp);
+	ntp->t_pflags = tp->t_pflags;
+	current->journal_info = ntp;
+}
+
 #endif	/* __XFS_TRANS_H__ */
Signature · 1 line
2.18.4