How to handle duplicated items in iobref during wb_fulfill_head?

[chen1195585098]

Description of problem:
Does the duplicated items in iobref during wb_fulfill_head meet expectation?

With performance.write-behind on, a writev request will be WOUND in wb_fulfill_head func after iobref_merge.

    list_for_each_entry(req, &head->winds, winds)
    {
        WB_IOV_LOAD(vector, count, req, head);

        if (iobref_merge(head->stub->args.iobref, req->stub->args.iobref))
            goto err;
    }

Normally, such a request will succeed. Howerver, if this request gets a failure from brick, wb_fulfill_err will be called, which eventualy puts the failed request back to wb_inode->todo list by calling __wb_add_request_for_retry without clearing merged iobrefs.

When these retried requests are picked from wb_inode->todo list and ready to WIND, iobref_merge will be called again and same request iobufs from head->winds are repeatly megered into head->stub->args.iobref. This leads to extra memory usage.
For instance,

[2024-12-28 05:01:52.897906] A [MSGID: 0] [mem-pool.c:201:__gf_realloc] : no memory available for size (18446744056529682504) current memory usage in kilobytes 19820560 [call stack follows]
/lib64/libglusterfs.so.0(+0x2a2b4)[0x7f2d2ae2a2b4]
/lib64/libglusterfs.so.0(_gf_msg_nomem+0x292)[0x7f2d2ae2a762]
/lib64/libglusterfs.so.0(__gf_realloc+0x1d8)[0x7f2d2ae52a18]
/lib64/libglusterfs.so.0(+0x57491)[0x7f2d2ae57491]

or

[root@localhost ~]# grep -rnE "cmdlinestr|running\ process" /home/mnt.log |tail -n 2
520590:[2024-12-31 07:49:43.212733 +0000] I [MSGID: 100030] [glusterfsd.c:2949:main] 0-/usr/local/sbin/glusterfs: Started running version [{arg=/usr/local/sbin/glusterfs}, {version=2024.12.31}, {cmdlinestr=/usr/local/sbin/glusterfs --log-file=/home/mnt.log --process-name fuse --volfile-server=localhost --volfile-id=issue /mnt}] 
520591:[2024-12-31 07:49:43.287638 +0000] I [glusterfsd.c:2638:daemonize] 0-glusterfs: Pid of current running process is 2443
[root@localhost ~]# grep -rnw 2443 /var/log/messages|grep oom
25131:Dec 31 17:05:34 localhost kernel: oom-kill:constraint=CONSTRAINT_NONE,nodemask=(null),cpuset=/,mems_allowed=0,global_oom,task_memcg=/user.slice/user-0.slice/session-1.scope,task=glusterfs,pid=2443,uid=0
25132:Dec 31 17:05:34 localhost kernel: Out of memory: Killed process 2443 (glusterfs) total-vm:10298156kB, anon-rss:7152392kB, file-rss:0kB, shmem-rss:0kB, UID:0 pgtables:18480kB oom_score_adj:0

And it seems these additional alloceated iobrefs also reduce the performance of wb request handle, which leads to high response latency of ls <mount_point> or cd <mount_point>.

Can we introduce an uniqueness check in iobref_merge during adding a new iobuf Or taking some measures to clear merged iobrefs in wb_fulfill_err?

The exact command to reproduce the issue:

Additional info:

- The operating system / glusterfs version: Centos, kernel 5.10, latest devel version.

Note: Please hide any confidential data which you don't want to share in public like IP address, file name, hostname or any other configuration