bz253089 Make gfs2_writepage use trylock on the log_flush lock

gfs2 was deadlocking because most code paths acquire the sd_log_flush_lock and then the page locks, while code paths going through gfs2_writepage acquire the page locks and then the log flush lock.

This patch makes gfs2_writepage do a trylock on the sd_log_flush_lock. If it fails, writepage simply redirtys the page and gives up. This should keep the deadlock from happening, however it might cause a performance hit.

Signed-off-by Benjamin E. Marzinski <[email protected]>

Quoted 4 lines · Benjamin Marzinski

gfs2 was deadlocking because most code paths acquire the sd_log_flush_lock and then the page locks, while code paths going through gfs2_writepage acquire the page locks and then the log flush lock.

Quoted 5 lines

This patch makes gfs2_writepage do a trylock on the sd_log_flush_lock. If it fails, writepage simply redirtys the page and gives up. This should keep the deadlock from happening, however it might cause a performance hit.

Signed-off-by Benjamin E. Marzinski <[email protected]>

And here is the actual patch

plain text document attachment (253089_try4.patch)
diff -urpN --exclude-from=linux-2.6.18-44.gfs2abhi.003_test_clean/Documentation/dontdiff linux-2.6.18-44.gfs2abhi.003_test_clean/fs/gfs2/log.c linux-2.6.18-44.gfs2abhi.003_test/fs/gfs2/log.c

fs/gfs2/log.c +19 -4

@@ -288,7 +288,7 @@ static void ail2_empty(struct gfs2_sbd *
  * Returns: errno
  */
 
-int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks)
+int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks, int wait)
 {
 	unsigned int try = 0;
 	unsigned reserved_blks = 6 * (4096 / sdp->sd_vfs->s_blocksize);

@@ -301,6 +301,10 @@ int gfs2_log_reserve(struct gfs2_sbd *sd
 	gfs2_log_lock(sdp);
 	while(sdp->sd_log_blks_free <= (blks + reserved_blks)) {
 		gfs2_log_unlock(sdp);
+		if (!wait) {
+			mutex_unlock(&sdp->sd_log_reserve_mutex);
+			return -EBUSY;
+		}
 		gfs2_ail1_empty(sdp, 0);
 		gfs2_log_flush(sdp, NULL);
 

@@ -310,10 +314,21 @@ int gfs2_log_reserve(struct gfs2_sbd *sd
 	}
 	sdp->sd_log_blks_free -= blks;
 	gfs2_log_unlock(sdp);
-	mutex_unlock(&sdp->sd_log_reserve_mutex);
-
-	down_read(&sdp->sd_log_flush_lock);
 
+	if (wait){
+		mutex_unlock(&sdp->sd_log_reserve_mutex);
+		down_read(&sdp->sd_log_flush_lock);
+	}
+	else {
+		if (!down_read_trylock(&sdp->sd_log_flush_lock)) {
+			gfs2_log_lock(sdp);
+			sdp->sd_log_blks_free += blks;
+			gfs2_log_unlock(sdp);
+			mutex_unlock(&sdp->sd_log_reserve_mutex);
+			return -EBUSY;
+		}
+		mutex_unlock(&sdp->sd_log_reserve_mutex);
+	}
 	return 0;
 }
 

diff -urpN --exclude-from=linux-2.6.18-44.gfs2abhi.003_test_clean/Documentation/dontdiff linux-2.6.18-44.gfs2abhi.003_test_clean/fs/gfs2/log.h linux-2.6.18-44.gfs2abhi.003_test/fs/gfs2/log.h

fs/gfs2/log.h +1 -1

@@ -50,7 +50,7 @@ unsigned int gfs2_struct2blk(struct gfs2
 
 int gfs2_ail1_empty(struct gfs2_sbd *sdp, int flags);
 
-int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks);
+int gfs2_log_reserve(struct gfs2_sbd *sdp, unsigned int blks, int wait);
 void gfs2_log_release(struct gfs2_sbd *sdp, unsigned int blks);
 
 struct buffer_head *gfs2_log_get_buf(struct gfs2_sbd *sdp);

diff -urpN --exclude-from=linux-2.6.18-44.gfs2abhi.003_test_clean/Documentation/dontdiff linux-2.6.18-44.gfs2abhi.003_test_clean/fs/gfs2/ops_address.c linux-2.6.18-44.gfs2abhi.003_test/fs/gfs2/ops_address.c

fs/gfs2/ops_address.c +1 -1

@@ -140,7 +140,7 @@ static int gfs2_writepage(struct page *p
 	if ((sdp->sd_args.ar_data == GFS2_DATA_ORDERED || gfs2_is_jdata(ip)) &&
 	    PageChecked(page)) {
 		ClearPageChecked(page);
-		error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
+		error = gfs2_do_trans_begin(sdp, RES_DINODE + 1, 0, 0);
 		if (error)
 			goto out_ignore;
 		if (!page_has_buffers(page)) {

diff -urpN --exclude-from=linux-2.6.18-44.gfs2abhi.003_test_clean/Documentation/dontdiff linux-2.6.18-44.gfs2abhi.003_test_clean/fs/gfs2/trans.c linux-2.6.18-44.gfs2abhi.003_test/fs/gfs2/trans.c

fs/gfs2/trans.c +3 -3

@@ -25,8 +25,8 @@
 #include "trans.h"
 #include "util.h"
 
-int gfs2_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
-		     unsigned int revokes)
+int gfs2_do_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
+		     unsigned int revokes, int wait)
 {
 	struct gfs2_trans *tr;
 	int error;

@@ -61,7 +61,7 @@ int gfs2_trans_begin(struct gfs2_sbd *sd
 		goto fail_gunlock;
 	}
 
-	error = gfs2_log_reserve(sdp, tr->tr_reserved);
+	error = gfs2_log_reserve(sdp, tr->tr_reserved, wait);
 	if (error)
 		goto fail_gunlock;
 

diff -urpN --exclude-from=linux-2.6.18-44.gfs2abhi.003_test_clean/Documentation/dontdiff linux-2.6.18-44.gfs2abhi.003_test_clean/fs/gfs2/trans.h linux-2.6.18-44.gfs2abhi.003_test/fs/gfs2/trans.h

fs/gfs2/trans.h +9 -2

@@ -25,8 +25,8 @@ struct gfs2_glock;
 #define RES_STATFS	1
 #define RES_QUOTA	2
 
-int gfs2_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
-		     unsigned int revokes);
+int gfs2_do_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
+			unsigned int revokes, int wait);
 
 void gfs2_trans_end(struct gfs2_sbd *sdp);
 

@@ -36,4 +36,11 @@ void gfs2_trans_add_revoke(struct gfs2_s
 void gfs2_trans_add_unrevoke(struct gfs2_sbd *sdp, u64 blkno);
 void gfs2_trans_add_rg(struct gfs2_rgrpd *rgd);
 
+
+static inline int gfs2_trans_begin(struct gfs2_sbd *sdp, unsigned int blocks,
+                     unsigned int revokes)
+{
+        return gfs2_do_trans_begin(sdp, blocks, revokes, 1);
+}
+
 #endif /* __TRANS_DOT_H__ */