Updates #353 Change-Id: I755b9208690be76935d763688fa414521eba3a40 Signed-off-by: Poornima G <pgurusid@redhat.com>
Updates: #242 Change-Id: I767e574a26e922760a7130bd209c178d74e8cf69 Signed-off-by: Poornima G <pgurusid@redhat.com>
Change-Id: I27f5e1e34fe3eac96c7dd88e90753fb5d3d14550 BUG: 1272030 Signed-off-by: Anoop C S <anoopcs@redhat.com>
Current use of a per-client mutex to protect fdctx introduces lock contentions when there are dozens of file operations active. Use finer grain spinlock to reduce contention, and put retrieving fdctx out of lock. Change-Id: Iea3e2eb481e76a5d73a582ba81529180c5b88248 BUG: 1519598 Signed-off-by: Zhang Huan <zhanghuan@open-fs.com>
Problem: In glusterfs code base we call mutex_lock/unlock to take
reference/dereference for a object.Sometime it could be
reason for lock contention also.
Solution: There is no need to use mutex to increase/decrease ref
counter, instead of using mutex use gcc builtin ATOMIC
operation.
Test: I have not observed yet how much performance gain after apply
this patch specific to glusterfs but i have tested same
with below small program(mutex and atomic both) and
get good difference.
static int numOuterLoops;
static void *
threadFunc(void *arg)
{
int j;
for (j = 0; j < numOuterLoops; j++) {
__atomic_add_fetch (&glob, 1,__ATOMIC_ACQ_REL);
}
return NULL;
}
int
main(int argc, char *argv[])
{
int opt, s, j;
int numThreads;
pthread_t *thread;
int verbose;
int64_t n = 0;
if (argc < 2 ) {
printf(" Please provide 2 args Num of threads && Outer Loop\n");
exit (-1);
}
numThreads = atoi(argv[1]);
numOuterLoops = atoi (argv[2]);
if (1) {
printf("\tthreads: %d; outer loops: %d;\n",
numThreads, numOuterLoops);
}
thread = calloc(numThreads, sizeof(pthread_t));
if (thread == NULL) {
printf ("calloc error so exit\n");
exit (-1);
}
__atomic_store (&glob, &n, __ATOMIC_RELEASE);
for (j = 0; j < numThreads; j++) {
s = pthread_create(&thread[j], NULL, threadFunc, NULL);
if (s != 0) {
printf ("pthread_create failed so exit\n");
exit (-1);
}
}
for (j = 0; j < numThreads; j++) {
s = pthread_join(thread[j], NULL);
if (s != 0) {
printf ("pthread_join failed so exit\n");
exit (-1);
}
}
printf("glob value is %ld\n",__atomic_load_n (&glob,__ATOMIC_RELAXED));
exit(0);
}
time ./thr_count 800 800000
threads: 800; outer loops: 800000;
glob value is 640000000
real 1m10.288s
user 0m57.269s
sys 3m31.565s
time ./thr_count_atomic 800 800000
threads: 800; outer loops: 800000;
glob value is 640000000
real 0m20.313s
user 1m20.558s
sys 0m0.028
Change-Id: Ie5030a52ea264875e002e108dd4b207b15ab7cc7
Signed-off-by: Mohit Agrawal <moagrawa@redhat.com>
icreate creates inode, while namelink links the basename to it's parent gfid. For now mkdir is the primary user of these fops. Better distribution is acheived by creating the inode on ,(say) mds1 and linking the basename to it's parent gfid on mds2. The inode serves readdirp, stat etc. More details about the fops are present at: https://review.gluster.org/#/c/13395/3/design/DHT2/DHT2_Icreate_Namelink_Notes.md This backport of three patches from experimental branch. 1- https://review.gluster.org/#/c/18085/ 2- https://review.gluster.org/#/c/18086/ 3- https://review.gluster.org/#/c/18094/ Updates gluster/glusterfs#243 Change-Id: I1bd3d5a441a3cfab1acfeb52f15c6c867d362592 Signed-off-by: Susant Palai <spalai@redhat.com>
Updates #302 Change-Id: Ia78e5d8f7b9ee6410965296808ad316c3cfb1d61 Signed-off-by: Amar Tumballi <amarts@redhat.com>
Coverity ID: 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419, 423, 424, 425, 426, 427, 428, 429, 436, 437, 438, 439, 440, 441, 442, 443 Issue: Event include_recursion Removed redundant, recursive includes from the files. Change-Id: I920776b1fa089a2d4917ca722d0075a9239911a7 BUG: 789278 Signed-off-by: Girjesh Rajoria <grajoria@redhat.com>
Summary: Halo Geo-replication is a feature which allows Gluster or NFS clients to write locally to their region (as defined by a latency "halo" or threshold if you like), and have their writes asynchronously propagate from their origin to the rest of the cluster. Clients can also write synchronously to the cluster simply by specifying a halo-latency which is very large (e.g. 10seconds) which will include all bricks. In other words, it allows clients to decide at mount time if they desire synchronous or asynchronous IO into a cluster and the cluster can support both of these modes to any number of clients simultaneously. There are a few new volume options due to this feature: halo-shd-latency: The threshold below which self-heal daemons will consider children (bricks) connected. halo-nfsd-latency: The threshold below which NFS daemons will consider children (bricks) connected. halo-latency: The threshold below which all other clients will consider children (bricks) connected. halo-min-replicas: The minimum number of replicas which are to be enforced regardless of latency specified in the above 3 options. If the number of children falls below this threshold the next best (chosen by latency) shall be swapped in. New FUSE mount options: halo-latency & halo-min-replicas: As descripted above. This feature combined with multi-threaded SHD support (D1271745) results in some pretty cool geo-replication possibilities. Operational Notes: - Global consistency is gaurenteed for synchronous clients, this is provided by the existing entry-locking mechanism. - Asynchronous clients on the other hand and merely consistent to their region. Writes & deletes will be protected via entry-locks as usual preventing concurrent writes into files which are undergoing replication. Read operations on the other hand should never block. - Writes are allowed from _any_ region and propagated from the origin to all other regions. The take away from this is care should be taken to ensure multiple writers do not write the same files resulting in a gfid split-brain which will require resolution via split-brain policies (majority, mtime & size). Recommended method for preventing this is using the nfs-auth feature to define which region for each share has RW permissions, tiers not in the origin region should have RO perms. TODO: - Synchronous clients (including the SHD) should choose clients from their own region as preferred sources for reads. Most of the plumbing is in place for this via the child_latency array. - Better GFID split brain handling & better dent type split brain handling (i.e. create a trash can and move the offending files into it). - Tagging in addition to latency as a means of defining which children you wish to synchronously write to Test Plan: - The usual suspects, clang, gcc w/ address sanitizer & valgrind - Prove tests Reviewers: jackl, dph, cjh, meyering Reviewed By: meyering Subscribers: ethanr Differential Revision: https://phabricator.fb.com/D1272053 Tasks: 4117827 Change-Id: I694a9ab429722da538da171ec528406e77b5e6d1 BUG: 1428061 Signed-off-by: Kevin Vigor <kvigor@fb.com> Reviewed-on: http://review.gluster.org/16099 Reviewed-on: https://review.gluster.org/16177 Tested-by: Pranith Kumar Karampuri <pkarampu@redhat.com> Smoke: Gluster Build System <jenkins@build.gluster.org> NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org> CentOS-regression: Gluster Build System <jenkins@build.gluster.org> Reviewed-by: Pranith Kumar Karampuri <pkarampu@redhat.com>
Locking during notify was introduced as part of commit aa22f24f5db7659387704998ae01520708869873 [1]. The fix was introduced to fix out-of-order CONNECT/DISCONNECT events from rpc-clnt to parent xlators [2]. However as part of handling DISCONNECT protocol/client does unwind saved frames (with failure) waiting for responses. This saved_frames_unwind can be a costly operation and hence ideally shouldn't be included in the critical section of notifylock, as it unnecessarily delays the reconnection to same brick. Also, its not a good practise to pass control to other xlators holding a lock as it can lead to deadlocks. So, this patch removes locking in rpc-clnt while notifying parent xlators. To fix [2], two changes are present in this patch: * notify DISCONNECT before cleaning up rpc connection (same as commit a6b63e11b7758cf1bfcb6798, patch [3]). * protocol/client uses rpc_clnt_cleanup_and_start, which cleans up rpc connection and does a start while handling a DISCONNECT event from rpc. Note that patch [3] was reverted as rpc_clnt_start called in quick_reconnect path of protocol/client didn't invoke connect on transport as the connection was not cleaned up _yet_ (as cleanup was moved post notification in rpc-clnt). This resulted in clients never attempting connect to bricks. Note that one of the neater ways to fix [2] (without using locks) is to introduce generation numbers to map CONNECT and DISCONNECTS across epochs and ignore DISCONNECT events if they don't belong to current epoch. However, this approach is a bit complex to implement and requires time. So, current patch is a hacky stop-gap fix till we come up with a more cleaner solution. [1] http://review.gluster.org/15916 [2] https://bugzilla.redhat.com/show_bug.cgi?id=1386626 [3] http://review.gluster.org/15681 Change-Id: I62daeee8bb1430004e28558f6eb133efd4ccf418 Signed-off-by: Raghavendra G <rgowdapp@redhat.com> BUG: 1427012 Reviewed-on: https://review.gluster.org/16784 Smoke: Gluster Build System <jenkins@build.gluster.org> Reviewed-by: Milind Changire <mchangir@redhat.com> NetBSD-regression: NetBSD Build System <jenkins@build.gluster.org> CentOS-regression: Gluster Build System <jenkins@build.gluster.org>