Whamcloud - gitweb
LU-9859 libcfs: rename cfs_cpt_table to cfs_cpt_tab
[fs/lustre-release.git] / lustre / ptlrpc / ptlrpcd.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/ptlrpc/ptlrpcd.c
33  */
34
35 /** \defgroup ptlrpcd PortalRPC daemon
36  *
37  * ptlrpcd is a special thread with its own set where other user might add
38  * requests when they don't want to wait for their completion.
39  * PtlRPCD will take care of sending such requests and then processing their
40  * replies and calling completion callbacks as necessary.
41  * The callbacks are called directly from ptlrpcd context.
42  * It is important to never significantly block (esp. on RPCs!) within such
43  * completion handler or a deadlock might occur where ptlrpcd enters some
44  * callback that attempts to send another RPC and wait for it to return,
45  * during which time ptlrpcd is completely blocked, so e.g. if import
46  * fails, recovery cannot progress because connection requests are also
47  * sent by ptlrpcd.
48  *
49  * @{
50  */
51
52 #define DEBUG_SUBSYSTEM S_RPC
53
54 #include <linux/kthread.h>
55 #include <libcfs/libcfs.h>
56 #include <lustre_net.h>
57 #include <lustre_lib.h>
58 #include <lustre_ha.h>
59 #include <obd_class.h>   /* for obd_zombie */
60 #include <obd_support.h> /* for OBD_FAIL_CHECK */
61 #include <cl_object.h> /* cl_env_{get,put}() */
62 #include <lprocfs_status.h>
63
64 #include "ptlrpc_internal.h"
65
66 /* One of these per CPT. */
67 struct ptlrpcd {
68         int                     pd_size;
69         int                     pd_index;
70         int                     pd_cpt;
71         int                     pd_cursor;
72         int                     pd_nthreads;
73         int                     pd_groupsize;
74         struct ptlrpcd_ctl      pd_threads[0];
75 };
76
77 /*
78  * max_ptlrpcds is obsolete, but retained to ensure that the kernel
79  * module will load on a system where it has been tuned.
80  * A value other than 0 implies it was tuned, in which case the value
81  * is used to derive a setting for ptlrpcd_per_cpt_max.
82  */
83 static int max_ptlrpcds;
84 module_param(max_ptlrpcds, int, 0644);
85 MODULE_PARM_DESC(max_ptlrpcds,
86                  "Max ptlrpcd thread count to be started (obsolete).");
87
88 /*
89  * ptlrpcd_bind_policy is obsolete, but retained to ensure that
90  * the kernel module will load on a system where it has been tuned.
91  * A value other than 0 implies it was tuned, in which case the value
92  * is used to derive a setting for ptlrpcd_partner_group_size.
93  */
94 static int ptlrpcd_bind_policy;
95 module_param(ptlrpcd_bind_policy, int, 0644);
96 MODULE_PARM_DESC(ptlrpcd_bind_policy,
97                  "Ptlrpcd threads binding mode (obsolete).");
98
99 /*
100  * ptlrpcd_per_cpt_max: The maximum number of ptlrpcd threads to run
101  * in a CPT.
102  */
103 static int ptlrpcd_per_cpt_max;
104 module_param(ptlrpcd_per_cpt_max, int, 0644);
105 MODULE_PARM_DESC(ptlrpcd_per_cpt_max,
106                  "Max ptlrpcd thread count to be started per CPT.");
107
108 /*
109  * ptlrpcd_partner_group_size: The desired number of threads in each
110  * ptlrpcd partner thread group. Default is 2, corresponding to the
111  * old PDB_POLICY_PAIR. A negative value makes all ptlrpcd threads in
112  * a CPT partners of each other.
113  */
114 static int ptlrpcd_partner_group_size;
115 module_param(ptlrpcd_partner_group_size, int, 0644);
116 MODULE_PARM_DESC(ptlrpcd_partner_group_size,
117                  "Number of ptlrpcd threads in a partner group.");
118
119 /*
120  * ptlrpcd_cpts: A CPT string describing the CPU partitions that
121  * ptlrpcd threads should run on. Used to make ptlrpcd threads run on
122  * a subset of all CPTs.
123  *
124  * ptlrpcd_cpts=2
125  * ptlrpcd_cpts=[2]
126  *   run ptlrpcd threads only on CPT 2.
127  *
128  * ptlrpcd_cpts=0-3
129  * ptlrpcd_cpts=[0-3]
130  *   run ptlrpcd threads on CPTs 0, 1, 2, and 3.
131  *
132  * ptlrpcd_cpts=[0-3,5,7]
133  *   run ptlrpcd threads on CPTS 0, 1, 2, 3, 5, and 7.
134  */
135 static char *ptlrpcd_cpts;
136 module_param(ptlrpcd_cpts, charp, 0644);
137 MODULE_PARM_DESC(ptlrpcd_cpts,
138                  "CPU partitions ptlrpcd threads should run in");
139
140 /* ptlrpcds_cpt_idx maps cpt numbers to an index in the ptlrpcds array. */
141 static int              *ptlrpcds_cpt_idx;
142
143 /* ptlrpcds_num is the number of entries in the ptlrpcds array. */
144 static int              ptlrpcds_num;
145 static struct ptlrpcd   **ptlrpcds;
146
147 /*
148  * In addition to the regular thread pool above, there is a single
149  * global recovery thread. Recovery isn't critical for performance,
150  * and doesn't block, but must always be able to proceed, and it is
151  * possible that all normal ptlrpcd threads are blocked. Hence the
152  * need for a dedicated thread.
153  */
154 static struct ptlrpcd_ctl ptlrpcd_rcv;
155
156 struct mutex ptlrpcd_mutex;
157 static int ptlrpcd_users = 0;
158
159 void ptlrpcd_wake(struct ptlrpc_request *req)
160 {
161         struct ptlrpc_request_set *set = req->rq_set;
162
163         LASSERT(set != NULL);
164         wake_up(&set->set_waitq);
165 }
166 EXPORT_SYMBOL(ptlrpcd_wake);
167
168 static struct ptlrpcd_ctl *
169 ptlrpcd_select_pc(struct ptlrpc_request *req)
170 {
171         struct ptlrpcd  *pd;
172         int             cpt;
173         int             idx;
174
175         if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL)
176                 return &ptlrpcd_rcv;
177
178         cpt = cfs_cpt_current(cfs_cpt_tab, 1);
179         if (ptlrpcds_cpt_idx == NULL)
180                 idx = cpt;
181         else
182                 idx = ptlrpcds_cpt_idx[cpt];
183         pd = ptlrpcds[idx];
184
185         /* We do not care whether it is strict load balance. */
186         idx = pd->pd_cursor;
187         if (++idx == pd->pd_nthreads)
188                 idx = 0;
189         pd->pd_cursor = idx;
190
191         return &pd->pd_threads[idx];
192 }
193
194 /**
195  * Move all request from an existing request set to the ptlrpcd queue.
196  * All requests from the set must be in phase RQ_PHASE_NEW.
197  */
198 void ptlrpcd_add_rqset(struct ptlrpc_request_set *set)
199 {
200         struct list_head *tmp, *pos;
201         struct ptlrpcd_ctl *pc;
202         struct ptlrpc_request_set *new;
203         int count, i;
204
205         pc = ptlrpcd_select_pc(NULL);
206         new = pc->pc_set;
207
208         list_for_each_safe(pos, tmp, &set->set_requests) {
209                 struct ptlrpc_request *req =
210                         list_entry(pos, struct ptlrpc_request,
211                                    rq_set_chain);
212
213                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
214                 req->rq_set = new;
215                 req->rq_queued_time = ktime_get_seconds();
216         }
217
218         spin_lock(&new->set_new_req_lock);
219         list_splice_init(&set->set_requests, &new->set_new_requests);
220         i = atomic_read(&set->set_remaining);
221         count = atomic_add_return(i, &new->set_new_count);
222         atomic_set(&set->set_remaining, 0);
223         spin_unlock(&new->set_new_req_lock);
224         if (count == i) {
225                 wake_up(&new->set_waitq);
226
227                 /*
228                  * XXX: It maybe unnecessary to wakeup all the partners. But to
229                  *      guarantee the async RPC can be processed ASAP, we have
230                  *      no other better choice. It maybe fixed in future.
231                  */
232                 for (i = 0; i < pc->pc_npartners; i++)
233                         wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
234         }
235 }
236
237 /**
238  * Return transferred RPCs count.
239  */
240 static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des,
241                                struct ptlrpc_request_set *src)
242 {
243         struct list_head *tmp, *pos;
244         struct ptlrpc_request *req;
245         int rc = 0;
246
247         spin_lock(&src->set_new_req_lock);
248         if (likely(!list_empty(&src->set_new_requests))) {
249                 list_for_each_safe(pos, tmp, &src->set_new_requests) {
250                         req = list_entry(pos, struct ptlrpc_request,
251                                          rq_set_chain);
252                         req->rq_set = des;
253                 }
254                 list_splice_init(&src->set_new_requests,
255                                  &des->set_requests);
256                 rc = atomic_read(&src->set_new_count);
257                 atomic_add(rc, &des->set_remaining);
258                 atomic_set(&src->set_new_count, 0);
259         }
260         spin_unlock(&src->set_new_req_lock);
261         return rc;
262 }
263
264 /**
265  * Requests that are added to the ptlrpcd queue are sent via
266  * ptlrpcd_check->ptlrpc_check_set().
267  */
268 void ptlrpcd_add_req(struct ptlrpc_request *req)
269 {
270         struct ptlrpcd_ctl *pc;
271
272         if (req->rq_reqmsg)
273                 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
274
275         spin_lock(&req->rq_lock);
276         if (req->rq_invalid_rqset) {
277                 req->rq_invalid_rqset = 0;
278                 spin_unlock(&req->rq_lock);
279                 if (wait_event_idle_timeout(req->rq_set_waitq,
280                                             req->rq_set == NULL,
281                                             cfs_time_seconds(5)) == 0)
282                         l_wait_event_abortable(req->rq_set_waitq,
283                                                req->rq_set == NULL);
284         } else if (req->rq_set) {
285                 /*
286                  * If we have a vaid "rq_set", just reuse it to avoid double
287                  * linked.
288                  */
289                 LASSERT(req->rq_phase == RQ_PHASE_NEW);
290                 LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY);
291
292                 /* ptlrpc_check_set will decrease the count */
293                 atomic_inc(&req->rq_set->set_remaining);
294                 spin_unlock(&req->rq_lock);
295                 wake_up(&req->rq_set->set_waitq);
296                 return;
297         } else {
298                 spin_unlock(&req->rq_lock);
299         }
300
301         pc = ptlrpcd_select_pc(req);
302
303         DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s+%d]",
304                   req, pc->pc_name, pc->pc_index);
305
306         ptlrpc_set_add_new_req(pc, req);
307 }
308 EXPORT_SYMBOL(ptlrpcd_add_req);
309
310 static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set)
311 {
312         atomic_inc(&set->set_refcount);
313 }
314
315 /**
316  * Check if there is more work to do on ptlrpcd set.
317  * Returns 1 if yes.
318  */
319 static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc)
320 {
321         struct list_head *tmp, *pos;
322         struct ptlrpc_request *req;
323         struct ptlrpc_request_set *set = pc->pc_set;
324         int rc = 0;
325         int rc2;
326
327         ENTRY;
328
329         if (atomic_read(&set->set_new_count)) {
330                 spin_lock(&set->set_new_req_lock);
331                 if (likely(!list_empty(&set->set_new_requests))) {
332                         list_splice_init(&set->set_new_requests,
333                                              &set->set_requests);
334                         atomic_add(atomic_read(&set->set_new_count),
335                                    &set->set_remaining);
336                         atomic_set(&set->set_new_count, 0);
337                         /*
338                          * Need to calculate its timeout.
339                          */
340                         rc = 1;
341                 }
342                 spin_unlock(&set->set_new_req_lock);
343         }
344
345         /*
346          * We should call lu_env_refill() before handling new requests to make
347          * sure that env key the requests depending on really exists.
348          */
349         rc2 = lu_env_refill(env);
350         if (rc2 != 0) {
351                 /*
352                  * XXX This is very awkward situation, because
353                  * execution can neither continue (request
354                  * interpreters assume that env is set up), nor repeat
355                  * the loop (as this potentially results in a tight
356                  * loop of -ENOMEM's).
357                  *
358                  * Fortunately, refill only ever does something when
359                  * new modules are loaded, i.e., early during boot up.
360                  */
361                 CERROR("Failure to refill session: %d\n", rc2);
362                 RETURN(rc);
363         }
364
365         if (atomic_read(&set->set_remaining))
366                 rc |= ptlrpc_check_set(env, set);
367
368         /*
369          * NB: ptlrpc_check_set has already moved complted request at the
370          * head of seq::set_requests
371          */
372         list_for_each_safe(pos, tmp, &set->set_requests) {
373                 req = list_entry(pos, struct ptlrpc_request, rq_set_chain);
374                 if (req->rq_phase != RQ_PHASE_COMPLETE)
375                         break;
376
377                 list_del_init(&req->rq_set_chain);
378                 req->rq_set = NULL;
379                 ptlrpc_req_finished(req);
380         }
381
382         if (rc == 0) {
383                 /*
384                  * If new requests have been added, make sure to wake up.
385                  */
386                 rc = atomic_read(&set->set_new_count);
387
388                 /*
389                  * If we have nothing to do, check whether we can take some
390                  * work from our partner threads.
391                  */
392                 if (rc == 0 && pc->pc_npartners > 0) {
393                         struct ptlrpcd_ctl *partner;
394                         struct ptlrpc_request_set *ps;
395                         int first = pc->pc_cursor;
396
397                         do {
398                                 partner = pc->pc_partners[pc->pc_cursor++];
399                                 if (pc->pc_cursor >= pc->pc_npartners)
400                                         pc->pc_cursor = 0;
401                                 if (partner == NULL)
402                                         continue;
403
404                                 spin_lock(&partner->pc_lock);
405                                 ps = partner->pc_set;
406                                 if (ps == NULL) {
407                                         spin_unlock(&partner->pc_lock);
408                                         continue;
409                                 }
410
411                                 ptlrpc_reqset_get(ps);
412                                 spin_unlock(&partner->pc_lock);
413
414                                 if (atomic_read(&ps->set_new_count)) {
415                                         rc = ptlrpcd_steal_rqset(set, ps);
416                                         if (rc > 0)
417                                                 CDEBUG(D_RPCTRACE,
418                                                        "transfer %d async RPCs [%d->%d]\n",
419                                                        rc, partner->pc_index,
420                                                        pc->pc_index);
421                                 }
422                                 ptlrpc_reqset_put(ps);
423                         } while (rc == 0 && pc->pc_cursor != first);
424                 }
425         }
426
427         RETURN(rc || test_bit(LIOD_STOP, &pc->pc_flags));
428 }
429
430 /**
431  * Main ptlrpcd thread.
432  * ptlrpc's code paths like to execute in process context, so we have this
433  * thread which spins on a set which contains the rpcs and sends them.
434  */
435 static int ptlrpcd(void *arg)
436 {
437         struct ptlrpcd_ctl              *pc = arg;
438         struct ptlrpc_request_set       *set;
439         struct lu_context               ses = { 0 };
440         struct lu_env                   env = { .le_ses = &ses };
441         int                             rc = 0;
442         int                             exit = 0;
443
444         ENTRY;
445
446         unshare_fs_struct();
447
448         if (cfs_cpt_bind(cfs_cpt_tab, pc->pc_cpt) != 0)
449                 CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt);
450
451         /*
452          * Allocate the request set after the thread has been bound
453          * above. This is safe because no requests will be queued
454          * until all ptlrpcd threads have confirmed that they have
455          * successfully started.
456          */
457         set = ptlrpc_prep_set();
458         if (set == NULL)
459                 GOTO(failed, rc = -ENOMEM);
460         spin_lock(&pc->pc_lock);
461         pc->pc_set = set;
462         spin_unlock(&pc->pc_lock);
463
464         /* Both client and server (MDT/OST) may use the environment. */
465         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD |
466                                           LCT_DT_THREAD |
467                                           LCT_CL_THREAD |
468                                           LCT_REMEMBER  |
469                                           LCT_NOREF);
470         if (rc != 0)
471                 GOTO(failed, rc);
472         rc = lu_context_init(env.le_ses, LCT_SESSION  |
473                                          LCT_REMEMBER |
474                                          LCT_NOREF);
475         if (rc != 0) {
476                 lu_context_fini(&env.le_ctx);
477                 GOTO(failed, rc);
478         }
479
480         complete(&pc->pc_starting);
481
482         /*
483          * This mainloop strongly resembles ptlrpc_set_wait() except that our
484          * set never completes.  ptlrpcd_check() calls ptlrpc_check_set() when
485          * there are requests in the set. New requests come in on the set's
486          * new_req_list and ptlrpcd_check() moves them into the set.
487          */
488         do {
489                 time64_t timeout;
490
491                 timeout = ptlrpc_set_next_timeout(set);
492
493                 lu_context_enter(&env.le_ctx);
494                 lu_context_enter(env.le_ses);
495                 if (timeout == 0)
496                         wait_event_idle(set->set_waitq,
497                                         ptlrpcd_check(&env, pc));
498                 else if (wait_event_idle_timeout(set->set_waitq,
499                                                  ptlrpcd_check(&env, pc),
500                                                  cfs_time_seconds(timeout))
501                          == 0)
502                         ptlrpc_expired_set(set);
503                 lu_context_exit(&env.le_ctx);
504                 lu_context_exit(env.le_ses);
505
506                 /*
507                  * Abort inflight rpcs for forced stop case.
508                  */
509                 if (test_bit(LIOD_STOP, &pc->pc_flags)) {
510                         if (test_bit(LIOD_FORCE, &pc->pc_flags))
511                                 ptlrpc_abort_set(set);
512                         exit++;
513                 }
514
515                 /*
516                  * Let's make one more loop to make sure that ptlrpcd_check()
517                  * copied all raced new rpcs into the set so we can kill them.
518                  */
519         } while (exit < 2);
520
521         /*
522          * Wait for inflight requests to drain.
523          */
524         if (!list_empty(&set->set_requests))
525                 ptlrpc_set_wait(&env, set);
526         lu_context_fini(&env.le_ctx);
527         lu_context_fini(env.le_ses);
528
529         complete(&pc->pc_finishing);
530
531         return 0;
532
533 failed:
534         pc->pc_error = rc;
535         complete(&pc->pc_starting);
536         RETURN(rc);
537 }
538
539 static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt)
540 {
541         ENTRY;
542
543         pc->pc_index = index;
544         pc->pc_cpt = cpt;
545         init_completion(&pc->pc_starting);
546         init_completion(&pc->pc_finishing);
547         spin_lock_init(&pc->pc_lock);
548
549         if (index < 0) {
550                 /* Recovery thread. */
551                 snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv");
552         } else {
553                 /* Regular thread. */
554                 snprintf(pc->pc_name, sizeof(pc->pc_name),
555                          "ptlrpcd_%02d_%02d", cpt, index);
556         }
557
558         EXIT;
559 }
560
561 /* XXX: We want multiple CPU cores to share the async RPC load. So we
562  *      start many ptlrpcd threads. We also want to reduce the ptlrpcd
563  *      overhead caused by data transfer cross-CPU cores. So we bind
564  *      all ptlrpcd threads to a CPT, in the expectation that CPTs
565  *      will be defined in a way that matches these boundaries. Within
566  *      a CPT a ptlrpcd thread can be scheduled on any available core.
567  *
568  *      Each ptlrpcd thread has its own request queue. This can cause
569  *      response delay if the thread is already busy. To help with
570  *      this we define partner threads: these are other threads bound
571  *      to the same CPT which will check for work in each other's
572  *      request queues if they have no work to do.
573  *
574  *      The desired number of partner threads can be tuned by setting
575  *      ptlrpcd_partner_group_size. The default is to create pairs of
576  *      partner threads.
577  */
578 static int ptlrpcd_partners(struct ptlrpcd *pd, int index)
579 {
580         struct ptlrpcd_ctl      *pc;
581         struct ptlrpcd_ctl      **ppc;
582         int                     first;
583         int                     i;
584         int                     rc = 0;
585
586         ENTRY;
587
588         LASSERT(index >= 0 && index < pd->pd_nthreads);
589         pc = &pd->pd_threads[index];
590         pc->pc_npartners = pd->pd_groupsize - 1;
591
592         if (pc->pc_npartners <= 0)
593                 GOTO(out, rc);
594
595         OBD_CPT_ALLOC(pc->pc_partners, cfs_cpt_tab, pc->pc_cpt,
596                       sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
597         if (pc->pc_partners == NULL) {
598                 pc->pc_npartners = 0;
599                 GOTO(out, rc = -ENOMEM);
600         }
601
602         first = index - index % pd->pd_groupsize;
603         ppc = pc->pc_partners;
604         for (i = first; i < first + pd->pd_groupsize; i++) {
605                 if (i != index)
606                         *ppc++ = &pd->pd_threads[i];
607         }
608 out:
609         RETURN(rc);
610 }
611
612 int ptlrpcd_start(struct ptlrpcd_ctl *pc)
613 {
614         struct task_struct      *task;
615         int                     rc = 0;
616
617         ENTRY;
618
619         /*
620          * Do not allow starting a second thread for one pc.
621          */
622         if (test_and_set_bit(LIOD_START, &pc->pc_flags)) {
623                 CWARN("Starting second thread (%s) for same pc %p\n",
624                       pc->pc_name, pc);
625                 RETURN(0);
626         }
627
628         task = kthread_run(ptlrpcd, pc, pc->pc_name);
629         if (IS_ERR(task))
630                 GOTO(out_set, rc = PTR_ERR(task));
631
632         wait_for_completion(&pc->pc_starting);
633         rc = pc->pc_error;
634         if (rc != 0)
635                 GOTO(out_set, rc);
636
637         RETURN(0);
638
639 out_set:
640         if (pc->pc_set != NULL) {
641                 struct ptlrpc_request_set *set = pc->pc_set;
642
643                 spin_lock(&pc->pc_lock);
644                 pc->pc_set = NULL;
645                 spin_unlock(&pc->pc_lock);
646                 ptlrpc_set_destroy(set);
647         }
648         clear_bit(LIOD_START, &pc->pc_flags);
649         RETURN(rc);
650 }
651
652 void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force)
653 {
654         ENTRY;
655
656         if (!test_bit(LIOD_START, &pc->pc_flags)) {
657                 CWARN("Thread for pc %p was not started\n", pc);
658                 goto out;
659         }
660
661         set_bit(LIOD_STOP, &pc->pc_flags);
662         if (force)
663                 set_bit(LIOD_FORCE, &pc->pc_flags);
664         wake_up(&pc->pc_set->set_waitq);
665
666 out:
667         EXIT;
668 }
669
670 void ptlrpcd_free(struct ptlrpcd_ctl *pc)
671 {
672         struct ptlrpc_request_set *set = pc->pc_set;
673
674         ENTRY;
675
676         if (!test_bit(LIOD_START, &pc->pc_flags)) {
677                 CWARN("Thread for pc %p was not started\n", pc);
678                 goto out;
679         }
680
681         wait_for_completion(&pc->pc_finishing);
682
683         spin_lock(&pc->pc_lock);
684         pc->pc_set = NULL;
685         spin_unlock(&pc->pc_lock);
686         ptlrpc_set_destroy(set);
687
688         clear_bit(LIOD_START, &pc->pc_flags);
689         clear_bit(LIOD_STOP, &pc->pc_flags);
690         clear_bit(LIOD_FORCE, &pc->pc_flags);
691
692 out:
693         if (pc->pc_npartners > 0) {
694                 LASSERT(pc->pc_partners != NULL);
695
696                 OBD_FREE(pc->pc_partners,
697                          sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners);
698                 pc->pc_partners = NULL;
699         }
700         pc->pc_npartners = 0;
701         pc->pc_error = 0;
702         EXIT;
703 }
704
705 static void ptlrpcd_fini(void)
706 {
707         int     i;
708         int     j;
709         int     ncpts;
710
711         ENTRY;
712
713         if (ptlrpcds != NULL) {
714                 for (i = 0; i < ptlrpcds_num; i++) {
715                         if (ptlrpcds[i] == NULL)
716                                 break;
717                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
718                                 ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0);
719                         for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++)
720                                 ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]);
721                         OBD_FREE(ptlrpcds[i], ptlrpcds[i]->pd_size);
722                         ptlrpcds[i] = NULL;
723                 }
724                 OBD_FREE(ptlrpcds, sizeof(ptlrpcds[0]) * ptlrpcds_num);
725         }
726         ptlrpcds_num = 0;
727
728         ptlrpcd_stop(&ptlrpcd_rcv, 0);
729         ptlrpcd_free(&ptlrpcd_rcv);
730
731         if (ptlrpcds_cpt_idx != NULL) {
732                 ncpts = cfs_cpt_number(cfs_cpt_tab);
733                 OBD_FREE(ptlrpcds_cpt_idx, ncpts * sizeof(ptlrpcds_cpt_idx[0]));
734                 ptlrpcds_cpt_idx = NULL;
735         }
736
737         EXIT;
738 }
739
740 static int ptlrpcd_init(void)
741 {
742         int                     nthreads;
743         int                     groupsize;
744         int                     size;
745         int                     i;
746         int                     j;
747         int                     rc = 0;
748         struct cfs_cpt_table    *cptable;
749         __u32                   *cpts = NULL;
750         int                     ncpts;
751         int                     cpt;
752         struct ptlrpcd          *pd;
753
754         ENTRY;
755
756         /*
757          * Determine the CPTs that ptlrpcd threads will run on.
758          */
759         cptable = cfs_cpt_tab;
760         ncpts = cfs_cpt_number(cptable);
761         if (ptlrpcd_cpts != NULL) {
762                 struct cfs_expr_list *el;
763
764                 size = ncpts * sizeof(ptlrpcds_cpt_idx[0]);
765                 OBD_ALLOC(ptlrpcds_cpt_idx, size);
766                 if (ptlrpcds_cpt_idx == NULL)
767                         GOTO(out, rc = -ENOMEM);
768
769                 rc = cfs_expr_list_parse(ptlrpcd_cpts,
770                                          strlen(ptlrpcd_cpts),
771                                          0, ncpts - 1, &el);
772                 if (rc != 0) {
773                         CERROR("%s: invalid CPT pattern string: %s",
774                                "ptlrpcd_cpts", ptlrpcd_cpts);
775                         GOTO(out, rc = -EINVAL);
776                 }
777
778                 rc = cfs_expr_list_values(el, ncpts, &cpts);
779                 cfs_expr_list_free(el);
780                 if (rc <= 0) {
781                         CERROR("%s: failed to parse CPT array %s: %d\n",
782                                "ptlrpcd_cpts", ptlrpcd_cpts, rc);
783                         if (rc == 0)
784                                 rc = -EINVAL;
785                         GOTO(out, rc);
786                 }
787
788                 /*
789                  * Create the cpt-to-index map. When there is no match
790                  * in the cpt table, pick a cpt at random. This could
791                  * be changed to take the topology of the system into
792                  * account.
793                  */
794                 for (cpt = 0; cpt < ncpts; cpt++) {
795                         for (i = 0; i < rc; i++)
796                                 if (cpts[i] == cpt)
797                                         break;
798                         if (i >= rc)
799                                 i = cpt % rc;
800                         ptlrpcds_cpt_idx[cpt] = i;
801                 }
802
803                 cfs_expr_list_values_free(cpts, rc);
804                 ncpts = rc;
805         }
806         ptlrpcds_num = ncpts;
807
808         size = ncpts * sizeof(ptlrpcds[0]);
809         OBD_ALLOC(ptlrpcds, size);
810         if (ptlrpcds == NULL)
811                 GOTO(out, rc = -ENOMEM);
812
813         /*
814          * The max_ptlrpcds parameter is obsolete, but do something
815          * sane if it has been tuned, and complain if
816          * ptlrpcd_per_cpt_max has also been tuned.
817          */
818         if (max_ptlrpcds != 0) {
819                 CWARN("max_ptlrpcds is obsolete.\n");
820                 if (ptlrpcd_per_cpt_max == 0) {
821                         ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts;
822                         /* Round up if there is a remainder. */
823                         if (max_ptlrpcds % ncpts != 0)
824                                 ptlrpcd_per_cpt_max++;
825                         CWARN("Setting ptlrpcd_per_cpt_max = %d\n",
826                               ptlrpcd_per_cpt_max);
827                 } else {
828                         CWARN("ptlrpd_per_cpt_max is also set!\n");
829                 }
830         }
831
832         /*
833          * The ptlrpcd_bind_policy parameter is obsolete, but do
834          * something sane if it has been tuned, and complain if
835          * ptlrpcd_partner_group_size is also tuned.
836          */
837         if (ptlrpcd_bind_policy != 0) {
838                 CWARN("ptlrpcd_bind_policy is obsolete.\n");
839                 if (ptlrpcd_partner_group_size == 0) {
840                         switch (ptlrpcd_bind_policy) {
841                         case 1: /* PDB_POLICY_NONE */
842                         case 2: /* PDB_POLICY_FULL */
843                                 ptlrpcd_partner_group_size = 1;
844                                 break;
845                         case 3: /* PDB_POLICY_PAIR */
846                                 ptlrpcd_partner_group_size = 2;
847                                 break;
848                         case 4: /* PDB_POLICY_NEIGHBOR */
849 #ifdef CONFIG_NUMA
850                                 ptlrpcd_partner_group_size = -1; /* CPT */
851 #else
852                                 ptlrpcd_partner_group_size = 3; /* Triplets */
853 #endif
854                                 break;
855                         default: /* Illegal value, use the default. */
856                                 ptlrpcd_partner_group_size = 2;
857                                 break;
858                         }
859                         CWARN("Setting ptlrpcd_partner_group_size = %d\n",
860                               ptlrpcd_partner_group_size);
861                 } else {
862                         CWARN("ptlrpcd_partner_group_size is also set!\n");
863                 }
864         }
865
866         if (ptlrpcd_partner_group_size == 0)
867                 ptlrpcd_partner_group_size = 2;
868         else if (ptlrpcd_partner_group_size < 0)
869                 ptlrpcd_partner_group_size = -1;
870         else if (ptlrpcd_per_cpt_max > 0 &&
871                  ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max)
872                 ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max;
873
874         /*
875          * Start the recovery thread first.
876          */
877         set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags);
878         ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY);
879         rc = ptlrpcd_start(&ptlrpcd_rcv);
880         if (rc < 0)
881                 GOTO(out, rc);
882
883         for (i = 0; i < ncpts; i++) {
884                 if (cpts == NULL)
885                         cpt = i;
886                 else
887                         cpt = cpts[i];
888
889                 nthreads = cfs_cpt_weight(cptable, cpt);
890                 if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads)
891                         nthreads = ptlrpcd_per_cpt_max;
892                 if (nthreads < 2)
893                         nthreads = 2;
894
895                 if (ptlrpcd_partner_group_size <= 0) {
896                         groupsize = nthreads;
897                 } else if (nthreads <= ptlrpcd_partner_group_size) {
898                         groupsize = nthreads;
899                 } else {
900                         groupsize = ptlrpcd_partner_group_size;
901                         if (nthreads % groupsize != 0)
902                                 nthreads += groupsize - (nthreads % groupsize);
903                 }
904
905                 size = offsetof(struct ptlrpcd, pd_threads[nthreads]);
906                 OBD_CPT_ALLOC(pd, cptable, cpt, size);
907
908                 if (!pd)
909                         GOTO(out, rc = -ENOMEM);
910                 pd->pd_size      = size;
911                 pd->pd_index     = i;
912                 pd->pd_cpt       = cpt;
913                 pd->pd_cursor    = 0;
914                 pd->pd_nthreads  = nthreads;
915                 pd->pd_groupsize = groupsize;
916                 ptlrpcds[i] = pd;
917
918                 /*
919                  * The ptlrpcd threads in a partner group can access
920                  * each other's struct ptlrpcd_ctl, so these must be
921                  * initialized before any thead is started.
922                  */
923                 for (j = 0; j < nthreads; j++) {
924                         ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt);
925                         rc = ptlrpcd_partners(pd, j);
926                         if (rc < 0)
927                                 GOTO(out, rc);
928                 }
929
930                 /* XXX: We start nthreads ptlrpc daemons on this cpt.
931                  *      Each of them can process any non-recovery
932                  *      async RPC to improve overall async RPC
933                  *      efficiency.
934                  *
935                  *      But there are some issues with async I/O RPCs
936                  *      and async non-I/O RPCs processed in the same
937                  *      set under some cases. The ptlrpcd may be
938                  *      blocked by some async I/O RPC(s), then will
939                  *      cause other async non-I/O RPC(s) can not be
940                  *      processed in time.
941                  *
942                  *      Maybe we should distinguish blocked async RPCs
943                  *      from non-blocked async RPCs, and process them
944                  *      in different ptlrpcd sets to avoid unnecessary
945                  *      dependency. But how to distribute async RPCs
946                  *      load among all the ptlrpc daemons becomes
947                  *      another trouble.
948                  */
949                 for (j = 0; j < nthreads; j++) {
950                         rc = ptlrpcd_start(&pd->pd_threads[j]);
951                         if (rc < 0)
952                                 GOTO(out, rc);
953                 }
954         }
955 out:
956         if (rc != 0)
957                 ptlrpcd_fini();
958
959         RETURN(rc);
960 }
961
962 int ptlrpcd_addref(void)
963 {
964         int rc = 0;
965
966         ENTRY;
967
968         mutex_lock(&ptlrpcd_mutex);
969         if (++ptlrpcd_users == 1) {
970                 rc = ptlrpcd_init();
971                 if (rc < 0)
972                         ptlrpcd_users--;
973         }
974         mutex_unlock(&ptlrpcd_mutex);
975         RETURN(rc);
976 }
977 EXPORT_SYMBOL(ptlrpcd_addref);
978
979 void ptlrpcd_decref(void)
980 {
981         mutex_lock(&ptlrpcd_mutex);
982         if (--ptlrpcd_users == 0)
983                 ptlrpcd_fini();
984         mutex_unlock(&ptlrpcd_mutex);
985 }
986 EXPORT_SYMBOL(ptlrpcd_decref);
987 /** @} ptlrpcd */