glusterfs.git/xlators/performance, branch v3.3.2qa3

localtime and ctime are not MT-SAFE

2013-02-08T20:04:42+00:00

There are a number of nit-level issues throughout the source with
the use of localtime and ctime. While they apparently aren't causing
too many problems, apart from the one in bz 828058, they ought to be
fixed. Among the "real" problems that are fixed in this patch:
1) general localtime and ctime not MT-SAFE. There's a non-zero chance
   that another thread calling localtime (or ctime) will over-write
   the static data about to be used in another thread
2) localtime(& <64-bit-type>) or ctime(& <64-bit-type>) generally
   not a problem on 64-bit or little-endian 32-bit. But even though
   we probably have zero users on big-ending 32-bit platforms, it's
   still incorrect.
3) multiple nested calls passed as params. Last one wins, i.e. over-
   writes result of prior calls.
4) Inconsistent error handling. Most of these calls are for logging,
   tracing, or dumping. I submit that if an error somehow occurs in
   the call to localtime or ctime, the log/trace/dump still should
   still occur.
5) Appliances should all have their clocks set to UTC, and all log
   entries, traces, and dumps should use GMT.
6) fix strtok(), change to strtok_r()

Other things this patch fixes/changes (that aren't bugs per se):
1) Change "%Y-%m-%d %H:%M:%S" and similar to their equivalent shorthand,
   e.g. "%F %T"
2) change sizeof(timestr) to sizeof timestr. sizeof is an operator,
   not a function. You don't use i +(32), why use sizeof().
   (And yes, you do use parens with sizeof().)
3) change 'char timestr[256]' to 'char timestr[32]' where appropriate.
   Per-thread stack is limited. Time strings are never longer than ~20
   characters, so why waste 220+ bytes on the stack?

Things this patch doesn't fix:
1) hodgepodge of %Y-%m-%d %H:%M:%S versus %Y/%m/%d-%H%M%S and other
   variations. It's not clear to me whether this ever matters, not to
   mention 3rd party log filtering tools may already rely on a
   particular format. Still it would be nice to have a single manifest
   constant and have every call to localtime/strftime consistently use
   the same format.

BUG: 832173

Change-Id: Iee9719db4576eacc6c75694d9107954d0912cba8
Signed-off-by: Kaleb S. KEITHLEY 
Reviewed-on: http://review.gluster.org/3613
Reviewed-by: Anand Avati 
Tested-by: Anand Avati

performance/quick-read: fix race condition in unlink

2013-02-07T22:41:32+00:00

use same lock (inode->lock), while incrementing/decrementing local->open_count.

Change-Id: I08cbab5b5dec09b6057f43324fe3152f1564ce46
BUG: 902174
Signed-off-by: Raghavendra G 
Reviewed-on: http://review.gluster.org/4396
Reviewed-by: Anand Avati 
Tested-by: Anand Avati

performance/io-cache: use pthread_mutex_trylock to hold mutex in statedumps

2012-08-16T14:17:34+00:00

Do not use pthread_mutex_lock and gf_log functions while dumping information
to statedump, to avoid deadlocks.

Change-Id: I6569366856fc2bc0fefb49c8379e2e4337717ce4
BUG: 843787
Signed-off-by: Raghavendra Bhat 
Reviewed-on: http://review.gluster.com/3799
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/write-behind: use pthread_mutex_trylock to hold mutex in statedumps

2012-08-16T05:51:59+00:00

Change-Id: I24c83b1b5e83ef3e38a019043c7fbca13b19ff43
BUG: 841543
Signed-off-by: Raghavendra Bhat 
Reviewed-on: http://review.gluster.com/3815
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/io-threads: Provide option to turn off least-priority

2012-08-12T07:09:06+00:00

RCA:
In cases when self-heal is in progress, self-heal fops are starved
because of least-priority. This affects other fops with conflicting
inode, entry locks with self-heal.

Fix:
This patch provides configuring enable/disable of least-priority.

Additional changes:
Moved RCHECKSUM fop to low instead of least because it will still
affect the performance of other fops if RCHECKSUM is in LEAST
priority.

Tests:
Tested that the enabling/disabling of fops is working fine.
Tested that RCHECKSUM fop priority is assigned LOW when
least-priority is disabled.

BUG: 843704
Signed-off-by: Pranith Kumar K 
Change-Id: I892f99d6d0a3e0ae6c0a280f82e2203af0c346f6
Reviewed-on: http://review.gluster.com/3751
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/read-ahead: use pthread_mutex_trylock to hold mutex in statedumps

2012-08-12T06:58:31+00:00

Change-Id: I4de64915a9c6a46e126ef4a5b987e49de558f827
BUG: 843796
Signed-off-by: Raghavendra Bhat 
Reviewed-on: http://review.gluster.com/3801
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/quick-read: use pthread_mutex_trylock to hold mutex in statedumps

2012-08-12T06:57:02+00:00

Do not use pthread_mutex_lock and gf_log functions while dumping information
to statedump, to avoid deadlocks.

Change-Id: Ic77d96bc52f2a2a32629c0ae20bba797317e0a81
BUG: 843789
Signed-off-by: Raghavendra Bhat 
Reviewed-on: http://review.gluster.com/3800
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

mount/fuse: Use state->lk_lock to print lock information on EAGAIN.

2012-05-25T11:25:20+00:00

Change-Id: I24a4a0b1c8dc0b8e08b380a5bc8efc111ccdb2c3
BUG: 808400
Signed-off-by: Mohammed Junaid 
Reviewed-on: http://review.gluster.com/3438
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/quick-read: fix race-conditions in qr_unlink.

2012-05-23T13:46:07+00:00

The list of fds on which open needs to be done as part of unlink,
was being modified at different places using different locks.
This resulted in a race-condition where open was marked as in-transit,
but fdctx was removed from the list of fds on which open was being
sent even before open was done. Because of this, open_in_transit would
be set forever (as an open was never actually sent, there would be no
open_cbk called and hence we could not reset the variable), blocking
all the future fd based fops on this fd.

Change-Id: Ie84a55bee578869a9a060a094ba28480e7643ae8
BUG: 819490
Signed-off-by: Raghavendra G 
Reviewed-on: http://review.gluster.com/3372
Tested-by: Gluster Build System 
Reviewed-by: Vijay Bellur

performance/md-cache: use mdc_fsetattr_cbk as the callback if setattr is on a fd

2012-05-23T09:04:23+00:00

Change-Id: I114a16055540e0cd3317b83b329600251ffe03c3
BUG: 823886
Signed-off-by: Raghavendra Bhat 
Reviewed-on: http://review.gluster.com/3404
Tested-by: Gluster Build System 
Reviewed-by: Amar Tumballi 
Reviewed-by: Vijay Bellur