]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - drivers/staging/lustre/lnet/selftest/rpc.c
projects / mirror_ubuntu-artful-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'efi-urgent' of git://git.kernel.org/pub/scm/linux/kernel/git/efi/efi into...
[mirror_ubuntu-artful-kernel.git] / drivers / staging / lustre / lnet / selftest / rpc.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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
25 *
26 * Copyright (c) 2012, 2015, 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 * lnet/selftest/rpc.c
33 *
34 * Author: Isaac Huang <isaac@clusterfs.com>
35 *
36 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
37 * - percpt data for service to improve smp performance
38 * - code cleanup
39 */
40
41 #define DEBUG_SUBSYSTEM S_LNET
42
43 #include "selftest.h"
44
45 enum srpc_state {
46 SRPC_STATE_NONE,
47 SRPC_STATE_NI_INIT,
48 SRPC_STATE_EQ_INIT,
49 SRPC_STATE_RUNNING,
50 SRPC_STATE_STOPPING,
51 };
52
53 static struct smoketest_rpc {
54 spinlock_t rpc_glock; /* global lock */
55 struct srpc_service *rpc_services[SRPC_SERVICE_MAX_ID + 1];
56 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
57 enum srpc_state rpc_state;
58 struct srpc_counters rpc_counters;
59 __u64 rpc_matchbits; /* matchbits counter */
60 } srpc_data;
61
62 static inline int
63 srpc_serv_portal(int svc_id)
64 {
65 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
66 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
67 }
68
69 /* forward ref's */
70 int srpc_handle_rpc(struct swi_workitem *wi);
71
72 void srpc_get_counters(struct srpc_counters *cnt)
73 {
74 spin_lock(&srpc_data.rpc_glock);
75 *cnt = srpc_data.rpc_counters;
76 spin_unlock(&srpc_data.rpc_glock);
77 }
78
79 void srpc_set_counters(const struct srpc_counters *cnt)
80 {
81 spin_lock(&srpc_data.rpc_glock);
82 srpc_data.rpc_counters = *cnt;
83 spin_unlock(&srpc_data.rpc_glock);
84 }
85
86 static int
87 srpc_add_bulk_page(struct srpc_bulk *bk, struct page *pg, int i, int off,
88 int nob)
89 {
90 LASSERT(off < PAGE_SIZE);
91 LASSERT(nob > 0 && nob <= PAGE_SIZE);
92
93 bk->bk_iovs[i].bv_offset = off;
94 bk->bk_iovs[i].bv_page = pg;
95 bk->bk_iovs[i].bv_len = nob;
96 return nob;
97 }
98
99 void
100 srpc_free_bulk(struct srpc_bulk *bk)
101 {
102 int i;
103 struct page *pg;
104
105 LASSERT(bk);
106
107 for (i = 0; i < bk->bk_niov; i++) {
108 pg = bk->bk_iovs[i].bv_page;
109 if (!pg)
110 break;
111
112 __free_page(pg);
113 }
114
115 LIBCFS_FREE(bk, offsetof(struct srpc_bulk, bk_iovs[bk->bk_niov]));
116 }
117
118 struct srpc_bulk *
119 srpc_alloc_bulk(int cpt, unsigned int bulk_off, unsigned int bulk_npg,
120 unsigned int bulk_len, int sink)
121 {
122 struct srpc_bulk *bk;
123 int i;
124
125 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
126
127 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
128 offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
129 if (!bk) {
130 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
131 return NULL;
132 }
133
134 memset(bk, 0, offsetof(struct srpc_bulk, bk_iovs[bulk_npg]));
135 bk->bk_sink = sink;
136 bk->bk_len = bulk_len;
137 bk->bk_niov = bulk_npg;
138
139 for (i = 0; i < bulk_npg; i++) {
140 struct page *pg;
141 int nob;
142
143 pg = alloc_pages_node(cfs_cpt_spread_node(lnet_cpt_table(), cpt),
144 GFP_KERNEL, 0);
145 if (!pg) {
146 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
147 srpc_free_bulk(bk);
148 return NULL;
149 }
150
151 nob = min_t(unsigned int, bulk_off + bulk_len, PAGE_SIZE) -
152 bulk_off;
153 srpc_add_bulk_page(bk, pg, i, bulk_off, nob);
154 bulk_len -= nob;
155 bulk_off = 0;
156 }
157
158 return bk;
159 }
160
161 static inline __u64
162 srpc_next_id(void)
163 {
164 __u64 id;
165
166 spin_lock(&srpc_data.rpc_glock);
167 id = srpc_data.rpc_matchbits++;
168 spin_unlock(&srpc_data.rpc_glock);
169 return id;
170 }
171
172 static void
173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
174 struct srpc_service_cd *scd,
175 struct srpc_buffer *buffer)
176 {
177 memset(rpc, 0, sizeof(*rpc));
178 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
179 srpc_serv_is_framework(scd->scd_svc) ?
180 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
181
182 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
183
184 rpc->srpc_scd = scd;
185 rpc->srpc_reqstbuf = buffer;
186 rpc->srpc_peer = buffer->buf_peer;
187 rpc->srpc_self = buffer->buf_self;
188 LNetInvalidateHandle(&rpc->srpc_replymdh);
189 }
190
191 static void
192 srpc_service_fini(struct srpc_service *svc)
193 {
194 struct srpc_service_cd *scd;
195 struct srpc_server_rpc *rpc;
196 struct srpc_buffer *buf;
197 struct list_head *q;
198 int i;
199
200 if (!svc->sv_cpt_data)
201 return;
202
203 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
204 while (1) {
205 if (!list_empty(&scd->scd_buf_posted))
206 q = &scd->scd_buf_posted;
207 else if (!list_empty(&scd->scd_buf_blocked))
208 q = &scd->scd_buf_blocked;
209 else
210 break;
211
212 while (!list_empty(q)) {
213 buf = list_entry(q->next, struct srpc_buffer,
214 buf_list);
215 list_del(&buf->buf_list);
216 LIBCFS_FREE(buf, sizeof(*buf));
217 }
218 }
219
220 LASSERT(list_empty(&scd->scd_rpc_active));
221
222 while (!list_empty(&scd->scd_rpc_free)) {
223 rpc = list_entry(scd->scd_rpc_free.next,
224 struct srpc_server_rpc,
225 srpc_list);
226 list_del(&rpc->srpc_list);
227 LIBCFS_FREE(rpc, sizeof(*rpc));
228 }
229 }
230
231 cfs_percpt_free(svc->sv_cpt_data);
232 svc->sv_cpt_data = NULL;
233 }
234
235 static int
236 srpc_service_nrpcs(struct srpc_service *svc)
237 {
238 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
239
240 return srpc_serv_is_framework(svc) ?
241 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
242 }
243
244 int srpc_add_buffer(struct swi_workitem *wi);
245
246 static int
247 srpc_service_init(struct srpc_service *svc)
248 {
249 struct srpc_service_cd *scd;
250 struct srpc_server_rpc *rpc;
251 int nrpcs;
252 int i;
253 int j;
254
255 svc->sv_shuttingdown = 0;
256
257 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
258 sizeof(**svc->sv_cpt_data));
259 if (!svc->sv_cpt_data)
260 return -ENOMEM;
261
262 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
263 1 : cfs_cpt_number(lnet_cpt_table());
264 nrpcs = srpc_service_nrpcs(svc);
265
266 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
267 scd->scd_cpt = i;
268 scd->scd_svc = svc;
269 spin_lock_init(&scd->scd_lock);
270 INIT_LIST_HEAD(&scd->scd_rpc_free);
271 INIT_LIST_HEAD(&scd->scd_rpc_active);
272 INIT_LIST_HEAD(&scd->scd_buf_posted);
273 INIT_LIST_HEAD(&scd->scd_buf_blocked);
274
275 scd->scd_ev.ev_data = scd;
276 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
277
278 /*
279 * NB: don't use lst_sched_serial for adding buffer,
280 * see details in srpc_service_add_buffers()
281 */
282 swi_init_workitem(&scd->scd_buf_wi, scd,
283 srpc_add_buffer, lst_sched_test[i]);
284
285 if (i && srpc_serv_is_framework(svc)) {
286 /*
287 * NB: framework service only needs srpc_service_cd for
288 * one partition, but we allocate for all to make
289 * it easier to implement, it will waste a little
290 * memory but nobody should care about this
291 */
292 continue;
293 }
294
295 for (j = 0; j < nrpcs; j++) {
296 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
297 i, sizeof(*rpc));
298 if (!rpc) {
299 srpc_service_fini(svc);
300 return -ENOMEM;
301 }
302 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
303 }
304 }
305
306 return 0;
307 }
308
309 int
310 srpc_add_service(struct srpc_service *sv)
311 {
312 int id = sv->sv_id;
313
314 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
315
316 if (srpc_service_init(sv))
317 return -ENOMEM;
318
319 spin_lock(&srpc_data.rpc_glock);
320
321 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
322
323 if (srpc_data.rpc_services[id]) {
324 spin_unlock(&srpc_data.rpc_glock);
325 goto failed;
326 }
327
328 srpc_data.rpc_services[id] = sv;
329 spin_unlock(&srpc_data.rpc_glock);
330
331 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
332 return 0;
333
334 failed:
335 srpc_service_fini(sv);
336 return -EBUSY;
337 }
338
339 int
340 srpc_remove_service(struct srpc_service *sv)
341 {
342 int id = sv->sv_id;
343
344 spin_lock(&srpc_data.rpc_glock);
345
346 if (srpc_data.rpc_services[id] != sv) {
347 spin_unlock(&srpc_data.rpc_glock);
348 return -ENOENT;
349 }
350
351 srpc_data.rpc_services[id] = NULL;
352 spin_unlock(&srpc_data.rpc_glock);
353 return 0;
354 }
355
356 static int
357 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
358 int len, int options, lnet_process_id_t peer,
359 lnet_handle_md_t *mdh, struct srpc_event *ev)
360 {
361 int rc;
362 lnet_md_t md;
363 lnet_handle_me_t meh;
364
365 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
366 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
367 if (rc) {
368 CERROR("LNetMEAttach failed: %d\n", rc);
369 LASSERT(rc == -ENOMEM);
370 return -ENOMEM;
371 }
372
373 md.threshold = 1;
374 md.user_ptr = ev;
375 md.start = buf;
376 md.length = len;
377 md.options = options;
378 md.eq_handle = srpc_data.rpc_lnet_eq;
379
380 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
381 if (rc) {
382 CERROR("LNetMDAttach failed: %d\n", rc);
383 LASSERT(rc == -ENOMEM);
384
385 rc = LNetMEUnlink(meh);
386 LASSERT(!rc);
387 return -ENOMEM;
388 }
389
390 CDEBUG(D_NET, "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
391 libcfs_id2str(peer), portal, matchbits);
392 return 0;
393 }
394
395 static int
396 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
397 int options, lnet_process_id_t peer, lnet_nid_t self,
398 lnet_handle_md_t *mdh, struct srpc_event *ev)
399 {
400 int rc;
401 lnet_md_t md;
402
403 md.user_ptr = ev;
404 md.start = buf;
405 md.length = len;
406 md.eq_handle = srpc_data.rpc_lnet_eq;
407 md.threshold = options & LNET_MD_OP_GET ? 2 : 1;
408 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
409
410 rc = LNetMDBind(md, LNET_UNLINK, mdh);
411 if (rc) {
412 CERROR("LNetMDBind failed: %d\n", rc);
413 LASSERT(rc == -ENOMEM);
414 return -ENOMEM;
415 }
416
417 /*
418 * this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
419 * they're only meaningful for MDs attached to an ME (i.e. passive
420 * buffers...
421 */
422 if (options & LNET_MD_OP_PUT) {
423 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
424 portal, matchbits, 0, 0);
425 } else {
426 LASSERT(options & LNET_MD_OP_GET);
427
428 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
429 }
430
431 if (rc) {
432 CERROR("LNet%s(%s, %d, %lld) failed: %d\n",
433 options & LNET_MD_OP_PUT ? "Put" : "Get",
434 libcfs_id2str(peer), portal, matchbits, rc);
435
436 /*
437 * The forthcoming unlink event will complete this operation
438 * with failure, so fall through and return success here.
439 */
440 rc = LNetMDUnlink(*mdh);
441 LASSERT(!rc);
442 } else {
443 CDEBUG(D_NET, "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
444 libcfs_id2str(peer), portal, matchbits);
445 }
446 return 0;
447 }
448
449 static int
450 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
451 lnet_handle_md_t *mdh, struct srpc_event *ev)
452 {
453 lnet_process_id_t any = { 0 };
454
455 any.nid = LNET_NID_ANY;
456 any.pid = LNET_PID_ANY;
457
458 return srpc_post_passive_rdma(srpc_serv_portal(service),
459 local, service, buf, len,
460 LNET_MD_OP_PUT, any, mdh, ev);
461 }
462
463 static int
464 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
465 __must_hold(&scd->scd_lock)
466 {
467 struct srpc_service *sv = scd->scd_svc;
468 struct srpc_msg *msg = &buf->buf_msg;
469 int rc;
470
471 LNetInvalidateHandle(&buf->buf_mdh);
472 list_add(&buf->buf_list, &scd->scd_buf_posted);
473 scd->scd_buf_nposted++;
474 spin_unlock(&scd->scd_lock);
475
476 rc = srpc_post_passive_rqtbuf(sv->sv_id,
477 !srpc_serv_is_framework(sv),
478 msg, sizeof(*msg), &buf->buf_mdh,
479 &scd->scd_ev);
480
481 /*
482 * At this point, a RPC (new or delayed) may have arrived in
483 * msg and its event handler has been called. So we must add
484 * buf to scd_buf_posted _before_ dropping scd_lock
485 */
486 spin_lock(&scd->scd_lock);
487
488 if (!rc) {
489 if (!sv->sv_shuttingdown)
490 return 0;
491
492 spin_unlock(&scd->scd_lock);
493 /*
494 * srpc_shutdown_service might have tried to unlink me
495 * when my buf_mdh was still invalid
496 */
497 LNetMDUnlink(buf->buf_mdh);
498 spin_lock(&scd->scd_lock);
499 return 0;
500 }
501
502 scd->scd_buf_nposted--;
503 if (sv->sv_shuttingdown)
504 return rc; /* don't allow to change scd_buf_posted */
505
506 list_del(&buf->buf_list);
507 spin_unlock(&scd->scd_lock);
508
509 LIBCFS_FREE(buf, sizeof(*buf));
510
511 spin_lock(&scd->scd_lock);
512 return rc;
513 }
514
515 int
516 srpc_add_buffer(struct swi_workitem *wi)
517 {
518 struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
519 struct srpc_buffer *buf;
520 int rc = 0;
521
522 /*
523 * it's called by workitem scheduler threads, these threads
524 * should have been set CPT affinity, so buffers will be posted
525 * on CPT local list of Portal
526 */
527 spin_lock(&scd->scd_lock);
528
529 while (scd->scd_buf_adjust > 0 &&
530 !scd->scd_svc->sv_shuttingdown) {
531 scd->scd_buf_adjust--; /* consume it */
532 scd->scd_buf_posting++;
533
534 spin_unlock(&scd->scd_lock);
535
536 LIBCFS_ALLOC(buf, sizeof(*buf));
537 if (!buf) {
538 CERROR("Failed to add new buf to service: %s\n",
539 scd->scd_svc->sv_name);
540 spin_lock(&scd->scd_lock);
541 rc = -ENOMEM;
542 break;
543 }
544
545 spin_lock(&scd->scd_lock);
546 if (scd->scd_svc->sv_shuttingdown) {
547 spin_unlock(&scd->scd_lock);
548 LIBCFS_FREE(buf, sizeof(*buf));
549
550 spin_lock(&scd->scd_lock);
551 rc = -ESHUTDOWN;
552 break;
553 }
554
555 rc = srpc_service_post_buffer(scd, buf);
556 if (rc)
557 break; /* buf has been freed inside */
558
559 LASSERT(scd->scd_buf_posting > 0);
560 scd->scd_buf_posting--;
561 scd->scd_buf_total++;
562 scd->scd_buf_low = max(2, scd->scd_buf_total / 4);
563 }
564
565 if (rc) {
566 scd->scd_buf_err_stamp = ktime_get_real_seconds();
567 scd->scd_buf_err = rc;
568
569 LASSERT(scd->scd_buf_posting > 0);
570 scd->scd_buf_posting--;
571 }
572
573 spin_unlock(&scd->scd_lock);
574 return 0;
575 }
576
577 int
578 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
579 {
580 struct srpc_service_cd *scd;
581 int rc = 0;
582 int i;
583
584 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
585
586 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
587 spin_lock(&scd->scd_lock);
588
589 scd->scd_buf_err = 0;
590 scd->scd_buf_err_stamp = 0;
591 scd->scd_buf_posting = 0;
592 scd->scd_buf_adjust = nbuffer;
593 /* start to post buffers */
594 swi_schedule_workitem(&scd->scd_buf_wi);
595 spin_unlock(&scd->scd_lock);
596
597 /* framework service only post buffer for one partition */
598 if (srpc_serv_is_framework(sv))
599 break;
600 }
601
602 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
603 spin_lock(&scd->scd_lock);
604 /*
605 * NB: srpc_service_add_buffers() can be called inside
606 * thread context of lst_sched_serial, and we don't normally
607 * allow to sleep inside thread context of WI scheduler
608 * because it will block current scheduler thread from doing
609 * anything else, even worse, it could deadlock if it's
610 * waiting on result from another WI of the same scheduler.
611 * However, it's safe at here because scd_buf_wi is scheduled
612 * by thread in a different WI scheduler (lst_sched_test),
613 * so we don't have any risk of deadlock, though this could
614 * block all WIs pending on lst_sched_serial for a moment
615 * which is not good but not fatal.
616 */
617 lst_wait_until(scd->scd_buf_err ||
618 (!scd->scd_buf_adjust &&
619 !scd->scd_buf_posting),
620 scd->scd_lock, "waiting for adding buffer\n");
621
622 if (scd->scd_buf_err && !rc)
623 rc = scd->scd_buf_err;
624
625 spin_unlock(&scd->scd_lock);
626 }
627
628 return rc;
629 }
630
631 void
632 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
633 {
634 struct srpc_service_cd *scd;
635 int num;
636 int i;
637
638 LASSERT(!sv->sv_shuttingdown);
639
640 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
641 spin_lock(&scd->scd_lock);
642
643 num = scd->scd_buf_total + scd->scd_buf_posting;
644 scd->scd_buf_adjust -= min(nbuffer, num);
645
646 spin_unlock(&scd->scd_lock);
647 }
648 }
649
650 /* returns 1 if sv has finished, otherwise 0 */
651 int
652 srpc_finish_service(struct srpc_service *sv)
653 {
654 struct srpc_service_cd *scd;
655 struct srpc_server_rpc *rpc;
656 int i;
657
658 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
659
660 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
661 spin_lock(&scd->scd_lock);
662 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
663 spin_unlock(&scd->scd_lock);
664 return 0;
665 }
666
667 if (scd->scd_buf_nposted > 0) {
668 CDEBUG(D_NET, "waiting for %d posted buffers to unlink\n",
669 scd->scd_buf_nposted);
670 spin_unlock(&scd->scd_lock);
671 return 0;
672 }
673
674 if (list_empty(&scd->scd_rpc_active)) {
675 spin_unlock(&scd->scd_lock);
676 continue;
677 }
678
679 rpc = list_entry(scd->scd_rpc_active.next,
680 struct srpc_server_rpc, srpc_list);
681 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, wi %s scheduled %d running %d, ev fired %d type %d status %d lnet %d\n",
682 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
683 swi_state2str(rpc->srpc_wi.swi_state),
684 rpc->srpc_wi.swi_workitem.wi_scheduled,
685 rpc->srpc_wi.swi_workitem.wi_running,
686 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
687 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
688 spin_unlock(&scd->scd_lock);
689 return 0;
690 }
691
692 /* no lock needed from now on */
693 srpc_service_fini(sv);
694 return 1;
695 }
696
697 /* called with sv->sv_lock held */
698 static void
699 srpc_service_recycle_buffer(struct srpc_service_cd *scd,
700 struct srpc_buffer *buf)
701 __must_hold(&scd->scd_lock)
702 {
703 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
704 if (srpc_service_post_buffer(scd, buf)) {
705 CWARN("Failed to post %s buffer\n",
706 scd->scd_svc->sv_name);
707 }
708 return;
709 }
710
711 /* service is shutting down, or we want to recycle some buffers */
712 scd->scd_buf_total--;
713
714 if (scd->scd_buf_adjust < 0) {
715 scd->scd_buf_adjust++;
716 if (scd->scd_buf_adjust < 0 &&
717 !scd->scd_buf_total && !scd->scd_buf_posting) {
718 CDEBUG(D_INFO,
719 "Try to recycle %d buffers but nothing left\n",
720 scd->scd_buf_adjust);
721 scd->scd_buf_adjust = 0;
722 }
723 }
724
725 spin_unlock(&scd->scd_lock);
726 LIBCFS_FREE(buf, sizeof(*buf));
727 spin_lock(&scd->scd_lock);
728 }
729
730 void
731 srpc_abort_service(struct srpc_service *sv)
732 {
733 struct srpc_service_cd *scd;
734 struct srpc_server_rpc *rpc;
735 int i;
736
737 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
738 sv->sv_id, sv->sv_name);
739
740 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
741 spin_lock(&scd->scd_lock);
742
743 /*
744 * schedule in-flight RPCs to notice the abort, NB:
745 * racing with incoming RPCs; complete fix should make test
746 * RPCs carry session ID in its headers
747 */
748 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
749 rpc->srpc_aborted = 1;
750 swi_schedule_workitem(&rpc->srpc_wi);
751 }
752
753 spin_unlock(&scd->scd_lock);
754 }
755 }
756
757 void
758 srpc_shutdown_service(struct srpc_service *sv)
759 {
760 struct srpc_service_cd *scd;
761 struct srpc_server_rpc *rpc;
762 struct srpc_buffer *buf;
763 int i;
764
765 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
766 sv->sv_id, sv->sv_name);
767
768 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
769 spin_lock(&scd->scd_lock);
770
771 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
772
773 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
774 spin_unlock(&scd->scd_lock);
775
776 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
777 spin_lock(&scd->scd_lock);
778
779 /* schedule in-flight RPCs to notice the shutdown */
780 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
781 swi_schedule_workitem(&rpc->srpc_wi);
782
783 spin_unlock(&scd->scd_lock);
784
785 /*
786 * OK to traverse scd_buf_posted without lock, since no one
787 * touches scd_buf_posted now
788 */
789 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
790 LNetMDUnlink(buf->buf_mdh);
791 }
792 }
793
794 static int
795 srpc_send_request(struct srpc_client_rpc *rpc)
796 {
797 struct srpc_event *ev = &rpc->crpc_reqstev;
798 int rc;
799
800 ev->ev_fired = 0;
801 ev->ev_data = rpc;
802 ev->ev_type = SRPC_REQUEST_SENT;
803
804 rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
805 rpc->crpc_service, &rpc->crpc_reqstmsg,
806 sizeof(struct srpc_msg), LNET_MD_OP_PUT,
807 rpc->crpc_dest, LNET_NID_ANY,
808 &rpc->crpc_reqstmdh, ev);
809 if (rc) {
810 LASSERT(rc == -ENOMEM);
811 ev->ev_fired = 1; /* no more event expected */
812 }
813 return rc;
814 }
815
816 static int
817 srpc_prepare_reply(struct srpc_client_rpc *rpc)
818 {
819 struct srpc_event *ev = &rpc->crpc_replyev;
820 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
821 int rc;
822
823 ev->ev_fired = 0;
824 ev->ev_data = rpc;
825 ev->ev_type = SRPC_REPLY_RCVD;
826
827 *id = srpc_next_id();
828
829 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
830 &rpc->crpc_replymsg,
831 sizeof(struct srpc_msg),
832 LNET_MD_OP_PUT, rpc->crpc_dest,
833 &rpc->crpc_replymdh, ev);
834 if (rc) {
835 LASSERT(rc == -ENOMEM);
836 ev->ev_fired = 1; /* no more event expected */
837 }
838 return rc;
839 }
840
841 static int
842 srpc_prepare_bulk(struct srpc_client_rpc *rpc)
843 {
844 struct srpc_bulk *bk = &rpc->crpc_bulk;
845 struct srpc_event *ev = &rpc->crpc_bulkev;
846 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
847 int rc;
848 int opt;
849
850 LASSERT(bk->bk_niov <= LNET_MAX_IOV);
851
852 if (!bk->bk_niov)
853 return 0; /* nothing to do */
854
855 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
856 opt |= LNET_MD_KIOV;
857
858 ev->ev_fired = 0;
859 ev->ev_data = rpc;
860 ev->ev_type = SRPC_BULK_REQ_RCVD;
861
862 *id = srpc_next_id();
863
864 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
865 &bk->bk_iovs[0], bk->bk_niov, opt,
866 rpc->crpc_dest, &bk->bk_mdh, ev);
867 if (rc) {
868 LASSERT(rc == -ENOMEM);
869 ev->ev_fired = 1; /* no more event expected */
870 }
871 return rc;
872 }
873
874 static int
875 srpc_do_bulk(struct srpc_server_rpc *rpc)
876 {
877 struct srpc_event *ev = &rpc->srpc_ev;
878 struct srpc_bulk *bk = rpc->srpc_bulk;
879 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
880 int rc;
881 int opt;
882
883 LASSERT(bk);
884
885 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
886 opt |= LNET_MD_KIOV;
887
888 ev->ev_fired = 0;
889 ev->ev_data = rpc;
890 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
891
892 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
893 &bk->bk_iovs[0], bk->bk_niov, opt,
894 rpc->srpc_peer, rpc->srpc_self,
895 &bk->bk_mdh, ev);
896 if (rc)
897 ev->ev_fired = 1; /* no more event expected */
898 return rc;
899 }
900
901 /* only called from srpc_handle_rpc */
902 static void
903 srpc_server_rpc_done(struct srpc_server_rpc *rpc, int status)
904 {
905 struct srpc_service_cd *scd = rpc->srpc_scd;
906 struct srpc_service *sv = scd->scd_svc;
907 struct srpc_buffer *buffer;
908
909 LASSERT(status || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
910
911 rpc->srpc_status = status;
912
913 CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
914 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
915 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
916 swi_state2str(rpc->srpc_wi.swi_state), status);
917
918 if (status) {
919 spin_lock(&srpc_data.rpc_glock);
920 srpc_data.rpc_counters.rpcs_dropped++;
921 spin_unlock(&srpc_data.rpc_glock);
922 }
923
924 if (rpc->srpc_done)
925 (*rpc->srpc_done) (rpc);
926 LASSERT(!rpc->srpc_bulk);
927
928 spin_lock(&scd->scd_lock);
929
930 if (rpc->srpc_reqstbuf) {
931 /*
932 * NB might drop sv_lock in srpc_service_recycle_buffer, but
933 * sv won't go away for scd_rpc_active must not be empty
934 */
935 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
936 rpc->srpc_reqstbuf = NULL;
937 }
938
939 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
940
941 /*
942 * No one can schedule me now since:
943 * - I'm not on scd_rpc_active.
944 * - all LNet events have been fired.
945 * Cancel pending schedules and prevent future schedule attempts:
946 */
947 LASSERT(rpc->srpc_ev.ev_fired);
948 swi_exit_workitem(&rpc->srpc_wi);
949
950 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
951 buffer = list_entry(scd->scd_buf_blocked.next,
952 struct srpc_buffer, buf_list);
953 list_del(&buffer->buf_list);
954
955 srpc_init_server_rpc(rpc, scd, buffer);
956 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
957 swi_schedule_workitem(&rpc->srpc_wi);
958 } else {
959 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
960 }
961
962 spin_unlock(&scd->scd_lock);
963 }
964
965 /* handles an incoming RPC */
966 int
967 srpc_handle_rpc(struct swi_workitem *wi)
968 {
969 struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
970 struct srpc_service_cd *scd = rpc->srpc_scd;
971 struct srpc_service *sv = scd->scd_svc;
972 struct srpc_event *ev = &rpc->srpc_ev;
973 int rc = 0;
974
975 LASSERT(wi == &rpc->srpc_wi);
976
977 spin_lock(&scd->scd_lock);
978
979 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
980 spin_unlock(&scd->scd_lock);
981
982 if (rpc->srpc_bulk)
983 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
984 LNetMDUnlink(rpc->srpc_replymdh);
985
986 if (ev->ev_fired) { /* no more event, OK to finish */
987 srpc_server_rpc_done(rpc, -ESHUTDOWN);
988 return 1;
989 }
990 return 0;
991 }
992
993 spin_unlock(&scd->scd_lock);
994
995 switch (wi->swi_state) {
996 default:
997 LBUG();
998 case SWI_STATE_NEWBORN: {
999 struct srpc_msg *msg;
1000 struct srpc_generic_reply *reply;
1001
1002 msg = &rpc->srpc_reqstbuf->buf_msg;
1003 reply = &rpc->srpc_replymsg.msg_body.reply;
1004
1005 if (!msg->msg_magic) {
1006 /* moaned already in srpc_lnet_ev_handler */
1007 srpc_server_rpc_done(rpc, EBADMSG);
1008 return 1;
1009 }
1010
1011 srpc_unpack_msg_hdr(msg);
1012 if (msg->msg_version != SRPC_MSG_VERSION) {
1013 CWARN("Version mismatch: %u, %u expected, from %s\n",
1014 msg->msg_version, SRPC_MSG_VERSION,
1015 libcfs_id2str(rpc->srpc_peer));
1016 reply->status = EPROTO;
1017 /* drop through and send reply */
1018 } else {
1019 reply->status = 0;
1020 rc = (*sv->sv_handler)(rpc);
1021 LASSERT(!reply->status || !rpc->srpc_bulk);
1022 if (rc) {
1023 srpc_server_rpc_done(rpc, rc);
1024 return 1;
1025 }
1026 }
1027
1028 wi->swi_state = SWI_STATE_BULK_STARTED;
1029
1030 if (rpc->srpc_bulk) {
1031 rc = srpc_do_bulk(rpc);
1032 if (!rc)
1033 return 0; /* wait for bulk */
1034
1035 LASSERT(ev->ev_fired);
1036 ev->ev_status = rc;
1037 }
1038 }
1039 case SWI_STATE_BULK_STARTED:
1040 LASSERT(!rpc->srpc_bulk || ev->ev_fired);
1041
1042 if (rpc->srpc_bulk) {
1043 rc = ev->ev_status;
1044
1045 if (sv->sv_bulk_ready)
1046 rc = (*sv->sv_bulk_ready) (rpc, rc);
1047
1048 if (rc) {
1049 srpc_server_rpc_done(rpc, rc);
1050 return 1;
1051 }
1052 }
1053
1054 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1055 rc = srpc_send_reply(rpc);
1056 if (!rc)
1057 return 0; /* wait for reply */
1058 srpc_server_rpc_done(rpc, rc);
1059 return 1;
1060
1061 case SWI_STATE_REPLY_SUBMITTED:
1062 if (!ev->ev_fired) {
1063 CERROR("RPC %p: bulk %p, service %d\n",
1064 rpc, rpc->srpc_bulk, sv->sv_id);
1065 CERROR("Event: status %d, type %d, lnet %d\n",
1066 ev->ev_status, ev->ev_type, ev->ev_lnet);
1067 LASSERT(ev->ev_fired);
1068 }
1069
1070 wi->swi_state = SWI_STATE_DONE;
1071 srpc_server_rpc_done(rpc, ev->ev_status);
1072 return 1;
1073 }
1074
1075 return 0;
1076 }
1077
1078 static void
1079 srpc_client_rpc_expired(void *data)
1080 {
1081 struct srpc_client_rpc *rpc = data;
1082
1083 CWARN("Client RPC expired: service %d, peer %s, timeout %d.\n",
1084 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1085 rpc->crpc_timeout);
1086
1087 spin_lock(&rpc->crpc_lock);
1088
1089 rpc->crpc_timeout = 0;
1090 srpc_abort_rpc(rpc, -ETIMEDOUT);
1091
1092 spin_unlock(&rpc->crpc_lock);
1093
1094 spin_lock(&srpc_data.rpc_glock);
1095 srpc_data.rpc_counters.rpcs_expired++;
1096 spin_unlock(&srpc_data.rpc_glock);
1097 }
1098
1099 static void
1100 srpc_add_client_rpc_timer(struct srpc_client_rpc *rpc)
1101 {
1102 struct stt_timer *timer = &rpc->crpc_timer;
1103
1104 if (!rpc->crpc_timeout)
1105 return;
1106
1107 INIT_LIST_HEAD(&timer->stt_list);
1108 timer->stt_data = rpc;
1109 timer->stt_func = srpc_client_rpc_expired;
1110 timer->stt_expires = ktime_get_real_seconds() + rpc->crpc_timeout;
1111 stt_add_timer(timer);
1112 }
1113
1114 /*
1115 * Called with rpc->crpc_lock held.
1116 *
1117 * Upon exit the RPC expiry timer is not queued and the handler is not
1118 * running on any CPU.
1119 */
1120 static void
1121 srpc_del_client_rpc_timer(struct srpc_client_rpc *rpc)
1122 {
1123 /* timer not planted or already exploded */
1124 if (!rpc->crpc_timeout)
1125 return;
1126
1127 /* timer successfully defused */
1128 if (stt_del_timer(&rpc->crpc_timer))
1129 return;
1130
1131 /* timer detonated, wait for it to explode */
1132 while (rpc->crpc_timeout) {
1133 spin_unlock(&rpc->crpc_lock);
1134
1135 schedule();
1136
1137 spin_lock(&rpc->crpc_lock);
1138 }
1139 }
1140
1141 static void
1142 srpc_client_rpc_done(struct srpc_client_rpc *rpc, int status)
1143 {
1144 struct swi_workitem *wi = &rpc->crpc_wi;
1145
1146 LASSERT(status || wi->swi_state == SWI_STATE_DONE);
1147
1148 spin_lock(&rpc->crpc_lock);
1149
1150 rpc->crpc_closed = 1;
1151 if (!rpc->crpc_status)
1152 rpc->crpc_status = status;
1153
1154 srpc_del_client_rpc_timer(rpc);
1155
1156 CDEBUG_LIMIT(!status ? D_NET : D_NETERROR,
1157 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1158 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1159 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1160
1161 /*
1162 * No one can schedule me now since:
1163 * - RPC timer has been defused.
1164 * - all LNet events have been fired.
1165 * - crpc_closed has been set, preventing srpc_abort_rpc from
1166 * scheduling me.
1167 * Cancel pending schedules and prevent future schedule attempts:
1168 */
1169 LASSERT(!srpc_event_pending(rpc));
1170 swi_exit_workitem(wi);
1171
1172 spin_unlock(&rpc->crpc_lock);
1173
1174 (*rpc->crpc_done)(rpc);
1175 }
1176
1177 /* sends an outgoing RPC */
1178 int
1179 srpc_send_rpc(struct swi_workitem *wi)
1180 {
1181 int rc = 0;
1182 struct srpc_client_rpc *rpc;
1183 struct srpc_msg *reply;
1184 int do_bulk;
1185
1186 LASSERT(wi);
1187
1188 rpc = wi->swi_workitem.wi_data;
1189
1190 LASSERT(rpc);
1191 LASSERT(wi == &rpc->crpc_wi);
1192
1193 reply = &rpc->crpc_replymsg;
1194 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1195
1196 spin_lock(&rpc->crpc_lock);
1197
1198 if (rpc->crpc_aborted) {
1199 spin_unlock(&rpc->crpc_lock);
1200 goto abort;
1201 }
1202
1203 spin_unlock(&rpc->crpc_lock);
1204
1205 switch (wi->swi_state) {
1206 default:
1207 LBUG();
1208 case SWI_STATE_NEWBORN:
1209 LASSERT(!srpc_event_pending(rpc));
1210
1211 rc = srpc_prepare_reply(rpc);
1212 if (rc) {
1213 srpc_client_rpc_done(rpc, rc);
1214 return 1;
1215 }
1216
1217 rc = srpc_prepare_bulk(rpc);
1218 if (rc)
1219 break;
1220
1221 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1222 rc = srpc_send_request(rpc);
1223 break;
1224
1225 case SWI_STATE_REQUEST_SUBMITTED:
1226 /*
1227 * CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1228 * order; however, they're processed in a strict order:
1229 * rqt, rpy, and bulk.
1230 */
1231 if (!rpc->crpc_reqstev.ev_fired)
1232 break;
1233
1234 rc = rpc->crpc_reqstev.ev_status;
1235 if (rc)
1236 break;
1237
1238 wi->swi_state = SWI_STATE_REQUEST_SENT;
1239 /* perhaps more events, fall thru */
1240 case SWI_STATE_REQUEST_SENT: {
1241 enum srpc_msg_type type = srpc_service2reply(rpc->crpc_service);
1242
1243 if (!rpc->crpc_replyev.ev_fired)
1244 break;
1245
1246 rc = rpc->crpc_replyev.ev_status;
1247 if (rc)
1248 break;
1249
1250 srpc_unpack_msg_hdr(reply);
1251 if (reply->msg_type != type ||
1252 (reply->msg_magic != SRPC_MSG_MAGIC &&
1253 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1254 CWARN("Bad message from %s: type %u (%d expected), magic %u (%d expected).\n",
1255 libcfs_id2str(rpc->crpc_dest),
1256 reply->msg_type, type,
1257 reply->msg_magic, SRPC_MSG_MAGIC);
1258 rc = -EBADMSG;
1259 break;
1260 }
1261
1262 if (do_bulk && reply->msg_body.reply.status) {
1263 CWARN("Remote error %d at %s, unlink bulk buffer in case peer didn't initiate bulk transfer\n",
1264 reply->msg_body.reply.status,
1265 libcfs_id2str(rpc->crpc_dest));
1266 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1267 }
1268
1269 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1270 }
1271 case SWI_STATE_REPLY_RECEIVED:
1272 if (do_bulk && !rpc->crpc_bulkev.ev_fired)
1273 break;
1274
1275 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1276
1277 /*
1278 * Bulk buffer was unlinked due to remote error. Clear error
1279 * since reply buffer still contains valid data.
1280 * NB rpc->crpc_done shouldn't look into bulk data in case of
1281 * remote error.
1282 */
1283 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1284 !rpc->crpc_status && reply->msg_body.reply.status)
1285 rc = 0;
1286
1287 wi->swi_state = SWI_STATE_DONE;
1288 srpc_client_rpc_done(rpc, rc);
1289 return 1;
1290 }
1291
1292 if (rc) {
1293 spin_lock(&rpc->crpc_lock);
1294 srpc_abort_rpc(rpc, rc);
1295 spin_unlock(&rpc->crpc_lock);
1296 }
1297
1298 abort:
1299 if (rpc->crpc_aborted) {
1300 LNetMDUnlink(rpc->crpc_reqstmdh);
1301 LNetMDUnlink(rpc->crpc_replymdh);
1302 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1303
1304 if (!srpc_event_pending(rpc)) {
1305 srpc_client_rpc_done(rpc, -EINTR);
1306 return 1;
1307 }
1308 }
1309 return 0;
1310 }
1311
1312 struct srpc_client_rpc *
1313 srpc_create_client_rpc(lnet_process_id_t peer, int service,
1314 int nbulkiov, int bulklen,
1315 void (*rpc_done)(struct srpc_client_rpc *),
1316 void (*rpc_fini)(struct srpc_client_rpc *), void *priv)
1317 {
1318 struct srpc_client_rpc *rpc;
1319
1320 LIBCFS_ALLOC(rpc, offsetof(struct srpc_client_rpc,
1321 crpc_bulk.bk_iovs[nbulkiov]));
1322 if (!rpc)
1323 return NULL;
1324
1325 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1326 bulklen, rpc_done, rpc_fini, priv);
1327 return rpc;
1328 }
1329
1330 /* called with rpc->crpc_lock held */
1331 void
1332 srpc_abort_rpc(struct srpc_client_rpc *rpc, int why)
1333 {
1334 LASSERT(why);
1335
1336 if (rpc->crpc_aborted || /* already aborted */
1337 rpc->crpc_closed) /* callback imminent */
1338 return;
1339
1340 CDEBUG(D_NET, "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1341 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1342 swi_state2str(rpc->crpc_wi.swi_state), why);
1343
1344 rpc->crpc_aborted = 1;
1345 rpc->crpc_status = why;
1346 swi_schedule_workitem(&rpc->crpc_wi);
1347 }
1348
1349 /* called with rpc->crpc_lock held */
1350 void
1351 srpc_post_rpc(struct srpc_client_rpc *rpc)
1352 {
1353 LASSERT(!rpc->crpc_aborted);
1354 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
1355
1356 CDEBUG(D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1357 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1358 rpc->crpc_timeout);
1359
1360 srpc_add_client_rpc_timer(rpc);
1361 swi_schedule_workitem(&rpc->crpc_wi);
1362 }
1363
1364 int
1365 srpc_send_reply(struct srpc_server_rpc *rpc)
1366 {
1367 struct srpc_event *ev = &rpc->srpc_ev;
1368 struct srpc_msg *msg = &rpc->srpc_replymsg;
1369 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1370 struct srpc_service_cd *scd = rpc->srpc_scd;
1371 struct srpc_service *sv = scd->scd_svc;
1372 __u64 rpyid;
1373 int rc;
1374
1375 LASSERT(buffer);
1376 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1377
1378 spin_lock(&scd->scd_lock);
1379
1380 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1381 /*
1382 * Repost buffer before replying since test client
1383 * might send me another RPC once it gets the reply
1384 */
1385 if (srpc_service_post_buffer(scd, buffer))
1386 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1387 rpc->srpc_reqstbuf = NULL;
1388 }
1389
1390 spin_unlock(&scd->scd_lock);
1391
1392 ev->ev_fired = 0;
1393 ev->ev_data = rpc;
1394 ev->ev_type = SRPC_REPLY_SENT;
1395
1396 msg->msg_magic = SRPC_MSG_MAGIC;
1397 msg->msg_version = SRPC_MSG_VERSION;
1398 msg->msg_type = srpc_service2reply(sv->sv_id);
1399
1400 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1401 sizeof(*msg), LNET_MD_OP_PUT,
1402 rpc->srpc_peer, rpc->srpc_self,
1403 &rpc->srpc_replymdh, ev);
1404 if (rc)
1405 ev->ev_fired = 1; /* no more event expected */
1406 return rc;
1407 }
1408
1409 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1410 static void
1411 srpc_lnet_ev_handler(lnet_event_t *ev)
1412 {
1413 struct srpc_service_cd *scd;
1414 struct srpc_event *rpcev = ev->md.user_ptr;
1415 struct srpc_client_rpc *crpc;
1416 struct srpc_server_rpc *srpc;
1417 struct srpc_buffer *buffer;
1418 struct srpc_service *sv;
1419 struct srpc_msg *msg;
1420 enum srpc_msg_type type;
1421
1422 LASSERT(!in_interrupt());
1423
1424 if (ev->status) {
1425 __u32 errors;
1426
1427 spin_lock(&srpc_data.rpc_glock);
1428 if (ev->status != -ECANCELED) /* cancellation is not error */
1429 srpc_data.rpc_counters.errors++;
1430 errors = srpc_data.rpc_counters.errors;
1431 spin_unlock(&srpc_data.rpc_glock);
1432
1433 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1434 ev->status, ev->type, errors);
1435 }
1436
1437 rpcev->ev_lnet = ev->type;
1438
1439 switch (rpcev->ev_type) {
1440 default:
1441 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1442 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1443 LBUG();
1444 case SRPC_REQUEST_SENT:
1445 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1446 spin_lock(&srpc_data.rpc_glock);
1447 srpc_data.rpc_counters.rpcs_sent++;
1448 spin_unlock(&srpc_data.rpc_glock);
1449 }
1450 case SRPC_REPLY_RCVD:
1451 case SRPC_BULK_REQ_RCVD:
1452 crpc = rpcev->ev_data;
1453
1454 if (rpcev != &crpc->crpc_reqstev &&
1455 rpcev != &crpc->crpc_replyev &&
1456 rpcev != &crpc->crpc_bulkev) {
1457 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1458 rpcev, crpc, &crpc->crpc_reqstev,
1459 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1460 CERROR("Bad event: status %d, type %d, lnet %d\n",
1461 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1462 LBUG();
1463 }
1464
1465 spin_lock(&crpc->crpc_lock);
1466
1467 LASSERT(!rpcev->ev_fired);
1468 rpcev->ev_fired = 1;
1469 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1470 -EINTR : ev->status;
1471 swi_schedule_workitem(&crpc->crpc_wi);
1472
1473 spin_unlock(&crpc->crpc_lock);
1474 break;
1475
1476 case SRPC_REQUEST_RCVD:
1477 scd = rpcev->ev_data;
1478 sv = scd->scd_svc;
1479
1480 LASSERT(rpcev == &scd->scd_ev);
1481
1482 spin_lock(&scd->scd_lock);
1483
1484 LASSERT(ev->unlinked);
1485 LASSERT(ev->type == LNET_EVENT_PUT ||
1486 ev->type == LNET_EVENT_UNLINK);
1487 LASSERT(ev->type != LNET_EVENT_UNLINK ||
1488 sv->sv_shuttingdown);
1489
1490 buffer = container_of(ev->md.start, struct srpc_buffer, buf_msg);
1491 buffer->buf_peer = ev->initiator;
1492 buffer->buf_self = ev->target.nid;
1493
1494 LASSERT(scd->scd_buf_nposted > 0);
1495 scd->scd_buf_nposted--;
1496
1497 if (sv->sv_shuttingdown) {
1498 /*
1499 * Leave buffer on scd->scd_buf_nposted since
1500 * srpc_finish_service needs to traverse it.
1501 */
1502 spin_unlock(&scd->scd_lock);
1503 break;
1504 }
1505
1506 if (scd->scd_buf_err_stamp &&
1507 scd->scd_buf_err_stamp < ktime_get_real_seconds()) {
1508 /* re-enable adding buffer */
1509 scd->scd_buf_err_stamp = 0;
1510 scd->scd_buf_err = 0;
1511 }
1512
1513 if (!scd->scd_buf_err && /* adding buffer is enabled */
1514 !scd->scd_buf_adjust &&
1515 scd->scd_buf_nposted < scd->scd_buf_low) {
1516 scd->scd_buf_adjust = max(scd->scd_buf_total / 2,
1517 SFW_TEST_WI_MIN);
1518 swi_schedule_workitem(&scd->scd_buf_wi);
1519 }
1520
1521 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1522 msg = &buffer->buf_msg;
1523 type = srpc_service2request(sv->sv_id);
1524
1525 if (ev->status || ev->mlength != sizeof(*msg) ||
1526 (msg->msg_type != type &&
1527 msg->msg_type != __swab32(type)) ||
1528 (msg->msg_magic != SRPC_MSG_MAGIC &&
1529 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1530 CERROR("Dropping RPC (%s) from %s: status %d mlength %d type %u magic %u.\n",
1531 sv->sv_name, libcfs_id2str(ev->initiator),
1532 ev->status, ev->mlength,
1533 msg->msg_type, msg->msg_magic);
1534
1535 /*
1536 * NB can't call srpc_service_recycle_buffer here since
1537 * it may call LNetM[DE]Attach. The invalid magic tells
1538 * srpc_handle_rpc to drop this RPC
1539 */
1540 msg->msg_magic = 0;
1541 }
1542
1543 if (!list_empty(&scd->scd_rpc_free)) {
1544 srpc = list_entry(scd->scd_rpc_free.next,
1545 struct srpc_server_rpc,
1546 srpc_list);
1547 list_del(&srpc->srpc_list);
1548
1549 srpc_init_server_rpc(srpc, scd, buffer);
1550 list_add_tail(&srpc->srpc_list,
1551 &scd->scd_rpc_active);
1552 swi_schedule_workitem(&srpc->srpc_wi);
1553 } else {
1554 list_add_tail(&buffer->buf_list,
1555 &scd->scd_buf_blocked);
1556 }
1557
1558 spin_unlock(&scd->scd_lock);
1559
1560 spin_lock(&srpc_data.rpc_glock);
1561 srpc_data.rpc_counters.rpcs_rcvd++;
1562 spin_unlock(&srpc_data.rpc_glock);
1563 break;
1564
1565 case SRPC_BULK_GET_RPLD:
1566 LASSERT(ev->type == LNET_EVENT_SEND ||
1567 ev->type == LNET_EVENT_REPLY ||
1568 ev->type == LNET_EVENT_UNLINK);
1569
1570 if (!ev->unlinked)
1571 break; /* wait for final event */
1572
1573 case SRPC_BULK_PUT_SENT:
1574 if (!ev->status && ev->type != LNET_EVENT_UNLINK) {
1575 spin_lock(&srpc_data.rpc_glock);
1576
1577 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1578 srpc_data.rpc_counters.bulk_get += ev->mlength;
1579 else
1580 srpc_data.rpc_counters.bulk_put += ev->mlength;
1581
1582 spin_unlock(&srpc_data.rpc_glock);
1583 }
1584 case SRPC_REPLY_SENT:
1585 srpc = rpcev->ev_data;
1586 scd = srpc->srpc_scd;
1587
1588 LASSERT(rpcev == &srpc->srpc_ev);
1589
1590 spin_lock(&scd->scd_lock);
1591
1592 rpcev->ev_fired = 1;
1593 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1594 -EINTR : ev->status;
1595 swi_schedule_workitem(&srpc->srpc_wi);
1596
1597 spin_unlock(&scd->scd_lock);
1598 break;
1599 }
1600 }
1601
1602 int
1603 srpc_startup(void)
1604 {
1605 int rc;
1606
1607 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1608 spin_lock_init(&srpc_data.rpc_glock);
1609
1610 /* 1 second pause to avoid timestamp reuse */
1611 set_current_state(TASK_UNINTERRUPTIBLE);
1612 schedule_timeout(cfs_time_seconds(1));
1613 srpc_data.rpc_matchbits = ((__u64)ktime_get_real_seconds()) << 48;
1614
1615 srpc_data.rpc_state = SRPC_STATE_NONE;
1616
1617 rc = LNetNIInit(LNET_PID_LUSTRE);
1618 if (rc < 0) {
1619 CERROR("LNetNIInit() has failed: %d\n", rc);
1620 return rc;
1621 }
1622
1623 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1624
1625 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1626 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1627 if (rc) {
1628 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1629 goto bail;
1630 }
1631
1632 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1633 LASSERT(!rc);
1634 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1635 LASSERT(!rc);
1636
1637 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1638
1639 rc = stt_startup();
1640
1641 bail:
1642 if (rc)
1643 srpc_shutdown();
1644 else
1645 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1646
1647 return rc;
1648 }
1649
1650 void
1651 srpc_shutdown(void)
1652 {
1653 int i;
1654 int rc;
1655 int state;
1656
1657 state = srpc_data.rpc_state;
1658 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1659
1660 switch (state) {
1661 default:
1662 LBUG();
1663 case SRPC_STATE_RUNNING:
1664 spin_lock(&srpc_data.rpc_glock);
1665
1666 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1667 struct srpc_service *sv = srpc_data.rpc_services[i];
1668
1669 LASSERTF(!sv, "service not empty: id %d, name %s\n",
1670 i, sv->sv_name);
1671 }
1672
1673 spin_unlock(&srpc_data.rpc_glock);
1674
1675 stt_shutdown();
1676
1677 case SRPC_STATE_EQ_INIT:
1678 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1679 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1680 LASSERT(!rc);
1681 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1682 LASSERT(!rc); /* the EQ should have no user by now */
1683
1684 case SRPC_STATE_NI_INIT:
1685 LNetNIFini();
1686 }
1687 }
Ubuntu Artful kernel mirror