Attachment 920713 Details for Bug 516025 – GFS2: Detect and release delayed bufdata structures

[patch] GFS2: Detect and release delayed bufdata structures

bz516025.addendum2.patch (text/plain), 3.34 KB, created by Robert Peterson on 2014-07-24 19:36:50 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Robert Peterson

Created: 2014-07-24 19:36:50 UTC

Size: 3.34 KB

patch

obsolete

>commit d23534b028f5e273deeca84157ee18ff07c3d0c1
>Author: Bob Peterson <rpeterso@redhat.com>
>Date:   Thu Jul 24 14:13:17 2014 -0500
>
>    GFS2: Detect and release delayed bufdata structures
>    
>    In ordinary cases, writes are buffered on the ordered writes list and
>    inodes are cached for a while. When another node wants to write to the
>    same file, the glock is demoted, the inode's pages and buffers are
>    truncated, and function gfs2_releasepage frees the buffers. However, in
>    some rare circumstances, a timing window can occur when another node
>    requests an inode (and therefore causes an invalidate to occur) while
>    the inode's buffers are actively being written from the ordered writes
>    list. In those cases, the buffer (and its corresponding bd element)
>    cannot be freed because b_count is non-zero until the IO is complete.
>    
>    This timing window does not exist in RHEL6 and above because the code
>    for writing the ordered writes list was reworked so that each inode
>    has its own ordered writes list and uses filemap_fdatawrite to write it.
>    That patch is too big and porting it too risky for RHEL5, so this patch
>    is an attempt to get around the problem and still allow it to free
>    the bd elements properly.
>    
>    When an invalidate operation occurs, this patch marks any buffer_heads
>    that can't be released due to b_count being non-zero during
>    gfs2_releasepage. If these marked buffer_heads are found in
>    gfs2_ordered_wait, we know the buffers are no longer needed, and they
>    are freed at that time.
>    
>    rhbz#516025
>---
> fs/gfs2/incore.h      | 3 +++
> fs/gfs2/log.c         | 6 ++++++
> fs/gfs2/ops_address.c | 8 ++++++--
> 3 files changed, 15 insertions(+), 2 deletions(-)
>
>diff --git a/fs/gfs2/incore.h b/fs/gfs2/incore.h
>index 07cb732..e1eeb6f 100644
>--- a/fs/gfs2/incore.h
>+++ b/fs/gfs2/incore.h
>@@ -96,12 +96,15 @@ struct gfs2_rgrpd {
> enum gfs2_state_bits {
> 	BH_Pinned = BH_PrivateStart,
> 	BH_Escaped = BH_PrivateStart + 1,
>+	BH_Release = BH_PrivateStart + 2, /* Needs its bd released */
> };
> 
> BUFFER_FNS(Pinned, pinned)
> TAS_BUFFER_FNS(Pinned, pinned)
> BUFFER_FNS(Escaped, escaped)
> TAS_BUFFER_FNS(Escaped, escaped)
>+BUFFER_FNS(Release, release)
>+TAS_BUFFER_FNS(Release, release)
> 
> struct gfs2_bufdata {
> 	struct buffer_head *bd_bh;
>diff --git a/fs/gfs2/log.c b/fs/gfs2/log.c
>index b3aeda1..a547106 100644
>--- a/fs/gfs2/log.c
>+++ b/fs/gfs2/log.c
>@@ -844,6 +844,12 @@ static void gfs2_ordered_wait(struct gfs2_sbd *sdp)
> 			get_bh(bh);
> 			gfs2_log_unlock(sdp);
> 			wait_on_buffer(bh);
>+			if (test_clear_buffer_release(bh)) {
>+				bd->bd_bh = NULL;
>+				bh->b_private = NULL;
>+				list_del_init(&bd->bd_le.le_list);
>+				kmem_cache_free(gfs2_bufdata_cachep, bd);
>+			}
> 			brelse(bh);
> 			gfs2_log_lock(sdp);
> 			continue;
>diff --git a/fs/gfs2/ops_address.c b/fs/gfs2/ops_address.c
>index 7ed96ce..f22ee46 100644
>--- a/fs/gfs2/ops_address.c
>+++ b/fs/gfs2/ops_address.c
>@@ -1312,8 +1312,12 @@ int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
> 	gfs2_log_lock(sdp);
> 	head = bh = page_buffers(page);
> 	do {
>-		if (atomic_read(&bh->b_count))
>-			goto cannot_release;
>+		if (atomic_read(&bh->b_count)) {
>+			set_buffer_release(bh); /* release later */
>+			if (atomic_read(&bh->b_count)) /* Sometimes it fixes
>+							  itself. */
>+				goto cannot_release;
>+		}
> 		bd = bh->b_private;
> 		if (bd && bd->bd_ail)
> 			goto cannot_release;

Actions: View | Diff

Attachments on bug 516025: 838395 | 916771 | 920713