Hide Forgot
Description of problem: While taking snapshot using scheduler one of the brick process crashed. Version-Release number of selected component (if applicable): glusterfs-3.8.4-6.el7rhgs.x86_64 How reproducible: 1/1 Steps to Reproduce: 1. Create 2*2 distributed replicate volume 2. enabled scheduler, 3. scheduled snapshot every one minute Actual results: One of the brick process crashed Expected results: Additional info: bt ======================= #0 0x00007f19a2a12394 in glusterfs_handle_barrier (req=0x7f19a30cffcc) at glusterfsd-mgmt.c:1348 ret = <optimized out> brick_req = {name = 0x7f198c0008e0 "repvol", op = 10, input = {input_len = 1783, input_val = 0x7f198c000900 ""}} brick_rsp = {op_ret = 0, op_errno = 0, output = {output_len = 0, output_val = 0x0}, op_errstr = 0x0} ctx = 0x7f19a3085010 active = 0x0 any = 0x0 xlator = 0x0 old_THIS = 0x0 dict = 0x0 name = '\000' <repeats 1023 times> barrier = _gf_true barrier_err = _gf_false __FUNCTION__ = "glusterfs_handle_barrier" #1 0x00007f19a2550a92 in synctask_wrap (old_task=<optimized out>) at syncop.c:375 task = 0x7f1990002510 #2 0x00007f19a0c0fcf0 in ?? () from /lib64/libc.so.6 No symbol table info available. #3 0x0000000000000000 in ?? () No symbol table info available.
The function from where this core was generated is glusterfs_handle_barrier (). From the core it looks like glusterfsd_ctx (global context) in the brick process didn't have ctx->active initialized which happens during graph initialization. We also saw that when barrier brick op was sent by GlusterD brick process just came up. The hypothesis we have here is as follows: T1. Brick process was in its init. However it still didn't finish doing the graph generation. T2. GlusterD sent a barrier brick op (as a trigger to snapshot initiated by snapshot scheduler) as it understood the brick to be connected (received the rpc connect notify from brick process) The time gap between T1 & T2 is very minimum and currently GlusterD doesn't know whether the brick process has finished all its initialization including the graph generation. One mitigation approach to avoid this crash is to avoid null pointer dereferencing which can be addressed by a simple patch and then even if we hit this race, barrier would fail. But to fix this race entirely we need to come up with a concrete solution which may not be feasible in 3.2.0 time lines.
Many thanks to Rajesh Jospeh for helping in the RCA.
The patch to address null pointer dereferencing is put up for review at upstream master : http://review.gluster.org/#/c/16043
upstream mainline : http://review.gluster.org/#/c/16043 upstream 3.8 : http://review.gluster.org/#/c/16066/ upstream 3.9 : http://review.gluster.org/#/c/16067/ downstream patch : https://code.engineering.redhat.com/gerrit/92447
Created 256 snapshots on volume using scheduler. Not seeing any brick crash. Bug verified on build glusterfs-3.8.4-8.el7rhgs.x86_64
Since the problem described in this bug report should be resolved in a recent advisory, it has been closed with a resolution of ERRATA. For information on the advisory, and where to find the updated files, follow the link below. If the solution does not work for you, open a new bug report. https://rhn.redhat.com/errata/RHSA-2017-0486.html