Whamcloud - gitweb
LU-13920 hsm: process hsm_actions only after mdd setup 28/40028/3
authorSergey Cheremencev <sergey.cheremencev@hpe.com>
Thu, 24 Sep 2020 12:57:47 +0000 (15:57 +0300)
committerOleg Drokin <green@whamcloud.com>
Mon, 12 Oct 2020 05:44:34 +0000 (05:44 +0000)
There is no guarantee that MDD setup is finished
at the moment when coordinator is started by
config params processing. If MDD setup is not finished,
hsm actions llog is not inited(mdd_hsm_actions_llog_init).
Hence hsm_pending_restore will not be called, i.e.
RESTORE requests will be sent to agents without taken
layout locks. I believe it may cause different problems.
I faced at least a kernel panic when llog includes
RESTORE request in ARS_WAITING that hasn't bee sent to
agent before failover. And the 2nd one RESTORE request
to the same fid was resent after recovery. Finally
agent handles to RESTORE to the same FID in parallel
that resulted in a panic with following bt:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000040
IP: [<ffffffffc0b03bec>] thandle_get_sub_by_dt+0x14c/0x420 [ptlrpc]
[<ffffffffc1202732>] lod_sub_get_thandle+0x2f2/0x400 [lod]
[<ffffffffc1205021>] lod_sub_declare_xattr_set+0x61/0x300 [lod]
[<ffffffffc11db0d5>] lod_obj_stripe_replace_parent_fid_cb+0x245/0x450 [lod]
[<ffffffffc11eae0e>] lod_obj_for_each_stripe+0x11e/0x2d0 [lod]
[<ffffffffc11ebfe2>] lod_replace_parent_fid+0x2a2/0x390 [lod]
[<ffffffffc11dae90>] ? lod_attr_get+0x110/0x110 [lod]
[<ffffffffc11f8faf>] lod_declare_xattr_set+0x24f/0xf70 [lod]
[<ffffffffc077b251>] ? lprocfs_counter_sub+0xc1/0x130 [obdclass]
[<ffffffffc1091ff4>] mdo_declare_xattr_set+0x74/0x2b0 [mdd]
[<ffffffffc077b129>] ? lprocfs_counter_add+0xf9/0x160 [obdclass]
[<ffffffffc0fa2f7b>] ? osd_trans_create+0xbb/0x620 [osd_ldiskfs]
[<ffffffffc1094903>] mdd_declare_xattr_set+0x33/0x70 [mdd]
[<ffffffffc1094b4e>] mdd_object_pfid_replace+0x7e/0x1e0 [mdd]
[<ffffffffc109c2c6>] mdd_swap_layouts+0xa76/0x1dc0 [mdd]
[<ffffffffc10a5e1a>] ? mdd_trans_stop+0x3a/0x174 [mdd]
[<ffffffffc114b489>] hsm_cdt_request_completed.isra.14+0xc89/0xf50 [mdt]
[<ffffffffc077b129>] ? lprocfs_counter_add+0xf9/0x160 [obdclass]
[<ffffffffc114d844>] mdt_hsm_update_request_state+0x544/0x7b0 [mdt]
[<ffffffffc0a82277>] ? lustre_msg_buf+0x17/0x60 [ptlrpc]
[<ffffffffc10fec92>] ? ucred_set_audit_enabled.isra.15+0x22/0x60 [mdt]
[<ffffffffc112b98f>] mdt_hsm_progress+0x1ef/0x3f0 [mdt]
[<ffffffffc0aefcfa>] tgt_request_handle+0x96a/0x1640 [ptlrpc]
[<ffffffffc06cca9e>] ? libcfs_nid2str_r+0xfe/0x130 [lnet]
[<ffffffffc0a91466>] ptlrpc_server_handle_request+0x256/0xb10 [ptlrpc]
[<ffffffffc0a95fbc>] ptlrpc_main+0xb3c/0x14d0 [ptlrpc]
[<ffffffffc0a95480>] ? ptlrpc_register_service+0xf90/0xf90 [ptlrpc]
[<ffffffffa04c1c31>] kthread+0xd1/0xe0
[<ffffffffa04c1b60>] ? insert_kthread_work+0x40/0x40
[<ffffffffa0b76c37>] ret_from_fork_nospec_begin+0x21/0x21
[<ffffffffa04c1b60>] ? insert_kthread_work+0x40/0x40
Code: 74 29 4c 3b a0 50 ff ff ff 75 e4 4d 85 ed 74 1b bf 01 00 00 00 e8 c5 b8 ff ff 85 c0 0f 85 98 00 00 00 49 8b 45 00 e9 04 ff ff ff <49> 8b 44 24 40 48 8b 40 08 48 85 c0 0f 84 b3 02 00 00 4c 89 e6
RIP  [<ffffffffc0b03bec>] thandle_get_sub_by_dt+0x14c/0x420 [ptlrpc]

Note, I faced this panic while testing https://review.whamcloud.com/#/c/38867/,
however I believe the same issue may exist even without 38867.

Patch makes mdt_hsm_cdt_start to wait until MDT initialization becomes
finished. Without this fix you should see below error in dmesg
each time after MDS restart if HSM is enbaled.
mdt_hsm_cdt_start()) lustre-MDT0000: cannot take the layout locks needed for registered restore: -2

Test-Parameters: testlist=sanity-hsm
Change-Id: I4c4edaa72a562576ea71d89a4b60618d099ec4f5
Signed-off-by: Sergey Cheremencev <sergey.cheremencev@hpe.com>
Reviewed-on: https://review.whamcloud.com/40028
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Ben Evans <beevans@whamcloud.com>
Reviewed-by: Nathan Rutman <nrutman@gmail.com>
Reviewed-by: Nikitas Angelinas <nikitas.angelinas@hpe.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

index ec4efe8..9f8372a 100644 (file)
@@ -1071,11 +1071,12 @@ int  mdt_hsm_cdt_fini(struct mdt_device *mdt)
 static int mdt_hsm_cdt_start(struct mdt_device *mdt)
-       struct coordinator      *cdt = &mdt->mdt_coordinator;
+       struct coordinator *cdt = &mdt->mdt_coordinator;
        struct mdt_thread_info *cdt_mti;
-       int                      rc;
-       void                    *ptr;
-       struct task_struct      *task;
+       unsigned int i = 0;
+       int rc;
+       void *ptr;
+       struct task_struct *task;
        /* functions defined but not yet used
@@ -1104,6 +1105,14 @@ static int mdt_hsm_cdt_start(struct mdt_device *mdt)
        cdt->cdt_group_request_mask = (1UL << HSMA_RESTORE);
        cdt->cdt_other_request_mask = (1UL << HSMA_RESTORE);
+       /* wait until MDD initialize hsm actions llog */
+       while (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state) && i < obd_timeout) {
+               schedule_timeout_interruptible(cfs_time_seconds(1));
+               i++;
+       }
+       if (!test_bit(MDT_FL_CFGLOG, &mdt->mdt_state))
+               CWARN("%s: trying to init HSM before MDD\n", mdt_obd_name(mdt));
        /* to avoid deadlock when start is made through sysfs
         * sysfs entries are created by the coordinator thread