When clearing all of the ldlm LRUs (as Cray does at the end of
a job), a ldlm_bl_work_item is generated for each namespace
and then they are placed on a list for the ldlm_bl threads to
iterate over.
If the number of namespaces greatly exceeds the number of
ldlm_bl threads, a given thread will iterate over many
namespaces without sleeping looking for work. This can go
on for an extremely long time and result in an RCU stall.
This patch adds a cond_resched() between completing one
work item and looking for the next. This is a fairly cheap
operation, as it will only schedule if there is an
interrupt waiting, and it will not be called too much -
Even the largest file systems have < 100 namespaces per
ldlm_bl_thread currently.
Signed-off-by: Patrick Farrell <paf@cray.com>
Change-Id: Ic8022faf641ad6ab02462ab376a4bfd510dca14c
Reviewed-on: https://review.whamcloud.com/20888
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Ned Bass <bass6@llnl.gov>
Reviewed-by: Ann Koehler <amk@cray.com>
Reviewed-by: Ben Evans <bevans@cray.com>
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
if (rc == LDLM_ITER_STOP)
break;
+
+ /* If there are many namespaces, we will not sleep waiting for
+ * work, and must do a cond_resched to avoid holding the CPU
+ * for too long */
+ cond_resched();
}
atomic_dec(&blp->blp_num_threads);