gfs2: Make flush bios explicitely sync
authorJan Kara <jack@suse.cz>
Tue, 2 May 2017 11:14:13 +0000 (13:14 +0200)
committerJan Kara <jack@suse.cz>
Wed, 24 May 2017 11:35:20 +0000 (13:35 +0200)
Commit b685d3d65ac7 "block: treat REQ_FUA and REQ_PREFLUSH as
synchronous" removed REQ_SYNC flag from WRITE_{FUA|PREFLUSH|...}
definitions.  generic_make_request_checks() however strips REQ_FUA and
REQ_PREFLUSH flags from a bio when the storage doesn't report volatile
write cache and thus write effectively becomes asynchronous which can
lead to performance regressions

Fix the problem by making sure all bios which are synchronous are
properly marked with REQ_SYNC.

Fixes: b685d3d65ac791406e0dfd8779cc9b3707fea5a3
CC: Steven Whitehouse <swhiteho@redhat.com>
CC: cluster-devel@redhat.com
CC: stable@vger.kernel.org
Acked-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Jan Kara <jack@suse.cz>
fs/gfs2/log.c

index f865b96374df2b5c40ecfb13663154499ec09b31..d2955daf17a4fcefa2ded3a412f67732315de12a 100644 (file)
@@ -659,7 +659,7 @@ static void log_write_header(struct gfs2_sbd *sdp, u32 flags)
        struct gfs2_log_header *lh;
        unsigned int tail;
        u32 hash;
        struct gfs2_log_header *lh;
        unsigned int tail;
        u32 hash;
-       int op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META;
+       int op_flags = REQ_PREFLUSH | REQ_FUA | REQ_META | REQ_SYNC;
        struct page *page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
        enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
        lh = page_address(page);
        struct page *page = mempool_alloc(gfs2_page_pool, GFP_NOIO);
        enum gfs2_freeze_state state = atomic_read(&sdp->sd_freeze_state);
        lh = page_address(page);