glusterfs-afrv1.git/xlators/performance/write-behind/src, branch release-3.0

performance/write-behind: initialize lock in wb-file before wb-file is set in fd-ctx.

2011-05-08T06:03:47+00:00

- Consider a combination of fuse->quick-read->read-ahead->wb->client.
    quick-read can do open-behind (open is returned as success even
    before it is issued to backend) and hence the fd can already be
    in the list of open fds of the inode. A flush call on some other
    fd opened on the same inode, will result in ra_flush issuing flush
    calls to all the fds opened on the same inode. This can result in
    wb_flush trying to hold a lock on non-initialized lock there by causing
    memory corruption.

Signed-off-by: Raghavendra G 
Signed-off-by: Vijay Bellur 

BUG: 2679 (Crash in GlusterFS 3.0.5 in GSP)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=2679

performance/write-behind: backport write-behind from 3.1

2011-01-27T20:18:17+00:00

Signed-off-by: Raghavendra G 
Signed-off-by: Anand V. Avati 

BUG: 934 (md5sum mismatch when files are transferred using vsftpd)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=934

write-behind: fixes in flush

2010-07-02T16:39:23+00:00

old code was resulting in duplicate flushes and spurious dangling flush frames

Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 

BUG: 1042 (Use correct flock structures in lk fops)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=1042

Revert "performance/write-behind: explicitly enforce ordering of overlapping writes."

2010-05-28T10:41:13+00:00

This reverts commit 6d9b11dba63d86c48450aa956281114962289ef5.

performance/write-behind: explicitly enforce ordering of overlapping writes.

2010-05-26T11:49:00+00:00

- If there are non-contiguous offsets (offsets which do not start where
    previous write ended), wait for completion of previous writes to server,
    before sending new ones.
  - Send flush call to server only when all writes are completed.
  - If a file is opened with O_APPEND, at any point of time a maximum only one
    write call to server should be in transit. This is to avoid reordering of
    writes in the presence of afr which can result in data corruption.
    See bug #934 for more details.

Signed-off-by: Raghavendra G 
Signed-off-by: Anand V. Avati 

BUG: 934 (md5sum mismatch when files are transferred using vsftpd)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=934

write-behind: update default values for better small file performance

2010-05-13T19:12:35+00:00

Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 

BUG: 898 (small file performance enhancements)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=898

performance/write-behind: Resume all the consecutive non-write operations in the request queue in a single go.

2010-03-29T11:29:57+00:00

Signed-off-by: Raghavendra G 
Signed-off-by: Anand V. Avati 

BUG: 740 (read-ahead does not work to its full potential when loaded on top of write-behind)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=740

performance/write-behind: fix data corruption while aggregating the adjacent contiguous iobufs into a single iobuf

2010-02-23T07:39:47+00:00

- while aggregating, we should make sure that the destination has enough
   memory. __wb_collapse_write_bufs assumed that destination vector's iov_base
   was aligned to the start of an iobuf and hence memory of page_size is
   available for aggregation. This assumption is not always true, like in the
   configuration consisting afr->write->io-cache (afr is on top). Refer to the
   bug url for more details.

Signed-off-by: Raghavendra G 
Signed-off-by: Anand V. Avati 

BUG: 542 (write-behind crashes)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=542

performance/write-behind: refactor __wb_collapse_write_bufs

2010-02-23T07:39:40+00:00

- refactor done to remove duplicated code.

Signed-off-by: Raghavendra G 
Signed-off-by: Anand V. Avati 

BUG: 542 (write-behind crashes)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=542

write-behind: fix 'option enable-trickling-writes'

2010-01-26T11:47:32+00:00

fix logic in __wb_mark_winds to properly initialize incomplete_writes
and conditionally check based on enable_trickling_writes flag

Signed-off-by: Anand V. Avati 
Signed-off-by: Anand V. Avati 

BUG: 547 (write-behind enable-trickling-writes working?)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=547