From 71f895fb6c0fd6f4146cf0b103be3ed4079d210e Mon Sep 17 00:00:00 2001 From: Mohit Agrawal Date: Tue, 8 Aug 2017 14:36:17 +0530 Subject: [PATCH 600/601] glusterd: Block brick attach request till the brick's ctx is set MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Problem: In multiplexing setup in a container environment we hit a race where before the first brick finishes its handshake with glusterd, the subsequent attach requests went through and they actually failed and glusterd has no mechanism to realize it. This resulted into all the such bricks not to be active resulting into clients not able to connect. Solution: Introduce a new flag port_registered in glusterd_brickinfo to make sure about pmap_signin finish before the subsequent attach bricks can be processed. Test: To reproduce the issue followed below steps 1) Create 100 volumes on 3 nodes(1x3) in CNS environment 2) Enable brick multiplexing 3) Reboot one container 4) Run below command for v in ‛gluster v list‛ do glfsheal $v | grep -i "transport" done After apply the patch command should not fail. Note: A big thanks to Atin for suggest the fix. >Reviewed-on: https://review.gluster.org/17984 >Reviewed-by: Atin Mukherjee >Smoke: Gluster Build System >CentOS-regression: Gluster Build System >Reviewed-by: Jeff Darcy BUG: 1477024 Change-Id: I8e1bd6132122b3a5b0dd49606cea564122f2609b Signed-off-by: Mohit Agrawal Reviewed-on: https://code.engineering.redhat.com/gerrit/114764 Tested-by: Atin Mukherjee Reviewed-by: Atin Mukherjee --- xlators/mgmt/glusterd/src/glusterd-pmap.c | 7 +++ xlators/mgmt/glusterd/src/glusterd-utils.c | 73 ++++++++++++++++++++---------- xlators/mgmt/glusterd/src/glusterd.h | 1 + 3 files changed, 57 insertions(+), 24 deletions(-) diff --git a/xlators/mgmt/glusterd/src/glusterd-pmap.c b/xlators/mgmt/glusterd/src/glusterd-pmap.c index 6fead59..7acee19 100644 --- a/xlators/mgmt/glusterd/src/glusterd-pmap.c +++ b/xlators/mgmt/glusterd/src/glusterd-pmap.c @@ -532,6 +532,9 @@ __gluster_pmap_signin (rpcsvc_request_t *req) GF_PMAP_PORT_BRICKSERVER, req->trans); ret = glusterd_get_brickinfo (THIS, args.brick, args.port, &brickinfo); + /* Update portmap status in brickinfo */ + if (brickinfo) + brickinfo->port_registered = _gf_true; fail: glusterd_submit_reply (req, &rsp, NULL, 0, NULL, @@ -584,6 +587,10 @@ __gluster_pmap_signout (rpcsvc_request_t *req) brick_path, GF_PMAP_PORT_BRICKSERVER, req->trans); } + /* Update portmap status on brickinfo */ + if (brickinfo) + brickinfo->port_registered = _gf_false; + /* Clean up the pidfile for this brick given glusterfsd doesn't clean it * any more. This is required to ensure we don't end up with having * stale pid files in case a brick is killed from the backend diff --git a/xlators/mgmt/glusterd/src/glusterd-utils.c b/xlators/mgmt/glusterd/src/glusterd-utils.c index 3afed1e..407e5b8 100644 --- a/xlators/mgmt/glusterd/src/glusterd-utils.c +++ b/xlators/mgmt/glusterd/src/glusterd-utils.c @@ -2057,6 +2057,7 @@ retry: brickinfo->port = port; brickinfo->rdma_port = rdma_port; brickinfo->status = GF_BRICK_STARTING; + brickinfo->port_registered = _gf_false; if (wait) { synclock_unlock (&priv->big_lock); @@ -5339,13 +5340,6 @@ attach_brick (xlator_t *this, GLUSTERD_REMOVE_SLASH_FROM_PATH (brickinfo->path, unslashed); - ret = pmap_registry_extend (this, other_brick->port, - brickinfo->path); - if (ret != 0) { - gf_log (this->name, GF_LOG_ERROR, - "adding brick to process failed"); - return -1; - } GLUSTERD_GET_BRICK_PIDFILE (pidfile1, other_vol, other_brick, conf); GLUSTERD_GET_BRICK_PIDFILE (pidfile2, volinfo, brickinfo, conf); @@ -5368,6 +5362,14 @@ attach_brick (xlator_t *this, GLUSTERD_BRICK_ATTACH); rpc_clnt_unref (rpc); if (!ret) { + ret = pmap_registry_extend (this, other_brick->port, + brickinfo->path); + if (ret != 0) { + gf_log (this->name, GF_LOG_ERROR, + "adding brick to process failed"); + return ret; + } + /* PID file is copied once brick has attached successfully */ @@ -5376,6 +5378,7 @@ attach_brick (xlator_t *this, brickinfo->status = GF_BRICK_STARTED; brickinfo->rpc = rpc_clnt_ref (other_brick->rpc); + brickinfo->port_registered = _gf_true; ret = glusterd_brick_process_add_brick (brickinfo, volinfo); if (ret) { @@ -5429,6 +5432,7 @@ find_compat_brick_in_vol (glusterd_conf_t *conf, int16_t retries = 15; int mux_limit = -1; int ret = -1; + gf_boolean_t brick_status = _gf_false; /* * If comp_vol is provided, we have to check *volume* compatibility @@ -5503,31 +5507,45 @@ find_compat_brick_in_vol (glusterd_conf_t *conf, "set. Continuing with no limit set for " "brick multiplexing."); } - - GLUSTERD_GET_BRICK_PIDFILE (pidfile2, srch_vol, other_brick, - conf); - - /* It is possible that the pidfile hasn't yet been populated, - * when bricks are started in "no-wait" mode; for example - * when bricks are started by glusterd_restart_bricks(). So - * wait for the pidfile to be populated with a value before - * checking if the service is running */ + /* The first brick process might take some time to finish its + * handshake with glusterd and prepare the graph. We can't + * afford to send attach_req for other bricks till that time. + * brick process sends PMAP_SIGNIN event after processing the + * volfile and hence it's safe to assume that if glusterd has + * received a pmap signin request for the same brick, we are + * good for subsequent attach requests. + */ + retries = 15; while (retries > 0) { - if (sys_access (pidfile2, F_OK) == 0 && - gf_is_service_running (pidfile2, &pid2)) { - break; + if (other_brick->port_registered) { + GLUSTERD_GET_BRICK_PIDFILE (pidfile2, srch_vol, + other_brick, conf); + if (sys_access (pidfile2, F_OK) == 0 && + gf_is_service_running (pidfile2, &pid2)) { + gf_msg_debug (this->name, 0, + "brick %s is running as a pid %d ", + other_brick->path, pid2); + brick_status = _gf_true; + break; + } } - sleep (1); + synclock_unlock (&conf->big_lock); + gf_msg_debug (this->name, 0, "brick %s is still" + " starting, waiting for 2 seconds ", + other_brick->path); + sleep(2); + synclock_lock (&conf->big_lock); retries--; } - if (retries == 0) { + if (!brick_status) { gf_log (this->name, GF_LOG_INFO, - "cleaning up dead brick %s:%s", + "brick has not come up so cleaning up dead brick %s:%s", other_brick->hostname, other_brick->path); other_brick->status = GF_BRICK_STOPPED; - sys_unlink (pidfile2); + if (pidfile2[0]) + sys_unlink (pidfile2); continue; } return other_brick; @@ -5803,6 +5821,12 @@ glusterd_brick_start (glusterd_volinfo_t *volinfo, brickinfo->path); goto out; } + /* We need to set the status back to STARTING so that + * while the other (re)start brick requests come in for + * other bricks, this brick can be considered as + * compatible. + */ + brickinfo->status = GF_BRICK_STARTING; } return 0; } @@ -5846,7 +5870,8 @@ run: * * TBD: pray for GlusterD 2 to be ready soon. */ - + gf_log (this->name, GF_LOG_INFO, "starting a fresh brick process for " + "brick %s", brickinfo->path); ret = glusterd_volume_start_glusterfs (volinfo, brickinfo, wait); if (ret) { gf_msg (this->name, GF_LOG_ERROR, 0, diff --git a/xlators/mgmt/glusterd/src/glusterd.h b/xlators/mgmt/glusterd/src/glusterd.h index c493773..533f9a8 100644 --- a/xlators/mgmt/glusterd/src/glusterd.h +++ b/xlators/mgmt/glusterd/src/glusterd.h @@ -226,6 +226,7 @@ struct glusterd_brickinfo { * a replica 3 volume with arbiter enabled. */ uint16_t group; + gf_boolean_t port_registered; }; typedef struct glusterd_brickinfo glusterd_brickinfo_t; -- 1.8.3.1