[PATCH] Use KM_NOFS for debug trace buffers

[PATCH] Use KM_NOFS for debug trace buffers

[Lachlan McIlroy]

Use KM_NOFS to prevent recursion back into the filesystem which can
cause deadlocks.

In the case of xfs_iread() we hold the lock on the inode cluster buffer
while allocating memory for the trace buffers.  If we recurse back into
XFS to flush data that may require a transaction to allocate extents
which needs log space.  This can deadlock with the xfsaild thread which
can't push the tail of the log because it is trying to get the inode
cluster buffer lock.

--- a/fs/xfs/linux-2.6/xfs_buf.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/linux-2.6/xfs_buf.c	2008-08-05 16:38:43.000000000 +1000
@@ -1795,7 +1795,7 @@ int __init
 xfs_buf_init(void)
 {
 #ifdef XFS_BUF_TRACE
-	xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_SLEEP);
+	xfs_buf_trace_buf = ktrace_alloc(XFS_BUF_TRACE_SIZE, KM_NOFS);
 #endif
 
 	xfs_buf_zone = kmem_zone_init_flags(sizeof(xfs_buf_t), "xfs_buf",
--- a/fs/xfs/quota/xfs_dquot.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/quota/xfs_dquot.c	2008-08-05 13:54:41.000000000 +1000
@@ -105,7 +105,7 @@ xfs_qm_dqinit(
 		sv_init(&dqp->q_pinwait, SV_DEFAULT, "pdq");
 
 #ifdef XFS_DQUOT_TRACE
-		dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_SLEEP);
+		dqp->q_trace = ktrace_alloc(DQUOT_TRACE_SIZE, KM_NOFS);
 		xfs_dqtrace_entry(dqp, "DQINIT");
 #endif
 	} else {
--- a/fs/xfs/xfs_buf_item.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/xfs_buf_item.c	2008-08-05 13:54:42.000000000 +1000
@@ -737,7 +737,7 @@ xfs_buf_item_init(
 	bip->bli_format.blf_len = (ushort)BTOBB(XFS_BUF_COUNT(bp));
 	bip->bli_format.blf_map_size = map_size;
 #ifdef XFS_BLI_TRACE
-	bip->bli_trace = ktrace_alloc(XFS_BLI_TRACE_SIZE, KM_SLEEP);
+	bip->bli_trace = ktrace_alloc(XFS_BLI_TRACE_SIZE, KM_NOFS);
 #endif
 
 #ifdef XFS_TRANS_DEBUG
--- a/fs/xfs/xfs_filestream.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/xfs_filestream.c	2008-08-05 13:54:42.000000000 +1000
@@ -400,7 +400,7 @@ xfs_filestream_init(void)
 	if (!item_zone)
 		return -ENOMEM;
 #ifdef XFS_FILESTREAMS_TRACE
-	xfs_filestreams_trace_buf = ktrace_alloc(XFS_FSTRM_KTRACE_SIZE, KM_SLEEP);
+	xfs_filestreams_trace_buf = ktrace_alloc(XFS_FSTRM_KTRACE_SIZE, KM_NOFS);
 #endif
 	return 0;
 }
--- a/fs/xfs/xfs_inode.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/xfs_inode.c	2008-08-05 13:54:42.000000000 +1000
@@ -835,22 +835,22 @@ xfs_iread(
 	 * Do this before xfs_iformat in case it adds entries.
 	 */
 #ifdef	XFS_INODE_TRACE
-	ip->i_trace = ktrace_alloc(INODE_TRACE_SIZE, KM_SLEEP);
+	ip->i_trace = ktrace_alloc(INODE_TRACE_SIZE, KM_NOFS);
 #endif
 #ifdef XFS_BMAP_TRACE
-	ip->i_xtrace = ktrace_alloc(XFS_BMAP_KTRACE_SIZE, KM_SLEEP);
+	ip->i_xtrace = ktrace_alloc(XFS_BMAP_KTRACE_SIZE, KM_NOFS);
 #endif
 #ifdef XFS_BMBT_TRACE
-	ip->i_btrace = ktrace_alloc(XFS_BMBT_KTRACE_SIZE, KM_SLEEP);
+	ip->i_btrace = ktrace_alloc(XFS_BMBT_KTRACE_SIZE, KM_NOFS);
 #endif
 #ifdef XFS_RW_TRACE
-	ip->i_rwtrace = ktrace_alloc(XFS_RW_KTRACE_SIZE, KM_SLEEP);
+	ip->i_rwtrace = ktrace_alloc(XFS_RW_KTRACE_SIZE, KM_NOFS);
 #endif
 #ifdef XFS_ILOCK_TRACE
-	ip->i_lock_trace = ktrace_alloc(XFS_ILOCK_KTRACE_SIZE, KM_SLEEP);
+	ip->i_lock_trace = ktrace_alloc(XFS_ILOCK_KTRACE_SIZE, KM_NOFS);
 #endif
 #ifdef XFS_DIR2_TRACE
-	ip->i_dir_trace = ktrace_alloc(XFS_DIR2_KTRACE_SIZE, KM_SLEEP);
+	ip->i_dir_trace = ktrace_alloc(XFS_DIR2_KTRACE_SIZE, KM_NOFS);
 #endif
 
 	/*
--- a/fs/xfs/xfs_log.c	2008-08-05 16:38:48.000000000 +1000
+++ b/fs/xfs/xfs_log.c	2008-08-05 13:54:42.000000000 +1000
@@ -160,7 +160,7 @@ void
 xlog_trace_iclog(xlog_in_core_t *iclog, uint state)
 {
 	if (!iclog->ic_trace)
-		iclog->ic_trace = ktrace_alloc(256, KM_SLEEP);
+		iclog->ic_trace = ktrace_alloc(256, KM_NOFS);
 	ktrace_enter(iclog->ic_trace,
 		     (void *)((unsigned long)state),
 		     (void *)((unsigned long)current_pid()),

Re: [PATCH] Use KM_NOFS for debug trace buffers

[Dave Chinner]

On Tue, Aug 05, 2008 at 05:09:59PM +1000, Lachlan McIlroy wrote:
> Use KM_NOFS to prevent recursion back into the filesystem which can
> cause deadlocks.
>
> In the case of xfs_iread() we hold the lock on the inode cluster buffer
> while allocating memory for the trace buffers.  If we recurse back into
> XFS to flush data that may require a transaction to allocate extents
> which needs log space.  This can deadlock with the xfsaild thread which
> can't push the tail of the log because it is trying to get the inode
> cluster buffer lock.

Looks OK.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com