When obd_get_mod_rpc_slot() waits for an available slot it places the
waiting thread at the HEAD of the queue, so it will be woken before
anything else that is already queued. This is clearly unfair and can
hurt performance.
So change to always add to the tail to ensure a FIFO ordering (except
that CLOSE might sometimes be woken a bit early).
This regression was introduced in a rewrite that was supposed to make
waiting more fair - by avoiding a broadcast wakeup for "close"
requests.
Also fix some stale comments and expose __add_wait_queue_entry_tail
Running mdtest with the patch applied shows about a 3% improvement:
master patched
mdtest-easy-write 350.585906 kIOPS 353.783545 kIOPS
mdtest-easy-stat 1320.329353 kIOPS 1408.320419 kIOPS
mdtest-easy-delete 285.084103 kIOPS 289.625900 kIOPS
[SCORE] 509.115803 kiops 524.516113 kiops
Fixes:
5243630b09d2 ("LU-15947 obdclass: improve precision of wakeups for mod_rpcs")
Signed-off-by: Mr NeilBrown <neilb@suse.de>
Signed-off-by: Shaun Tancheff <shaun.tancheff@hpe.com>
Change-Id: If767c4299bcbab71589b0f3c01e85bf461686ca5
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/52738
Reviewed-by: James Simmons <jsimmons@infradead.org>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
#define wait_queue_entry_t wait_queue_t
#endif
+#ifndef HAVE_PREPARE_TO_WAIT_EVENT
+#define __add_wait_queue_entry_tail __add_wait_queue_tail
+#endif
+
#ifndef HAVE_WAIT_BIT_HEADER_H
struct wait_bit_queue_entry {
struct wait_bit_key key;
#ifndef HAVE_PREPARE_TO_WAIT_EVENT
-#define __add_wait_queue_entry_tail __add_wait_queue_tail
-
long prepare_to_wait_event(wait_queue_head_t *wq_head,
wait_queue_entry_t *wq_entry, int state)
{
wait.wqe.func = claim_mod_rpc_function;
spin_lock_irq(&cli->cl_mod_rpcs_waitq.lock);
- __add_wait_queue(&cli->cl_mod_rpcs_waitq, &wait.wqe);
+ __add_wait_queue_entry_tail(&cli->cl_mod_rpcs_waitq, &wait.wqe);
/* This wakeup will only succeed if the maximums haven't
- * been reached. If that happens, WQ_FLAG_WOKEN will be cleared
+ * been reached. If that happens, wait.woken will be set
* and there will be no need to wait.
*/
wake_up_locked(&cli->cl_mod_rpcs_waitq);
- /* XXX: handle spurious wakeups (from unknown yet source */
while (wait.woken == false) {
spin_unlock_irq(&cli->cl_mod_rpcs_waitq.lock);
wait_woken(&wait.wqe, TASK_UNINTERRUPTIBLE,