Commit | Line | Data |
---|---|---|
d7e09d03 PT |
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.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * lustre/ptlrpc/ptlrpcd.c | |
37 | */ | |
38 | ||
39 | /** \defgroup ptlrpcd PortalRPC daemon | |
40 | * | |
41 | * ptlrpcd is a special thread with its own set where other user might add | |
42 | * requests when they don't want to wait for their completion. | |
43 | * PtlRPCD will take care of sending such requests and then processing their | |
44 | * replies and calling completion callbacks as necessary. | |
45 | * The callbacks are called directly from ptlrpcd context. | |
46 | * It is important to never significantly block (esp. on RPCs!) within such | |
47 | * completion handler or a deadlock might occur where ptlrpcd enters some | |
48 | * callback that attempts to send another RPC and wait for it to return, | |
49 | * during which time ptlrpcd is completely blocked, so e.g. if import | |
50 | * fails, recovery cannot progress because connection requests are also | |
51 | * sent by ptlrpcd. | |
52 | * | |
53 | * @{ | |
54 | */ | |
55 | ||
56 | #define DEBUG_SUBSYSTEM S_RPC | |
57 | ||
9fdaf8c0 | 58 | #include "../../include/linux/libcfs/libcfs.h" |
d7e09d03 | 59 | |
e27db149 GKH |
60 | #include "../include/lustre_net.h" |
61 | #include "../include/lustre_lib.h" | |
62 | #include "../include/lustre_ha.h" | |
63 | #include "../include/obd_class.h" /* for obd_zombie */ | |
64 | #include "../include/obd_support.h" /* for OBD_FAIL_CHECK */ | |
65 | #include "../include/cl_object.h" /* cl_env_{get,put}() */ | |
66 | #include "../include/lprocfs_status.h" | |
d7e09d03 PT |
67 | |
68 | #include "ptlrpc_internal.h" | |
69 | ||
c5c4c6fa | 70 | /* One of these per CPT. */ |
d7e09d03 | 71 | struct ptlrpcd { |
d0bfef31 CH |
72 | int pd_size; |
73 | int pd_index; | |
c5c4c6fa OW |
74 | int pd_cpt; |
75 | int pd_cursor; | |
d0bfef31 | 76 | int pd_nthreads; |
c5c4c6fa | 77 | int pd_groupsize; |
d7e09d03 PT |
78 | struct ptlrpcd_ctl pd_threads[0]; |
79 | }; | |
80 | ||
c5c4c6fa OW |
81 | /* |
82 | * max_ptlrpcds is obsolete, but retained to ensure that the kernel | |
83 | * module will load on a system where it has been tuned. | |
84 | * A value other than 0 implies it was tuned, in which case the value | |
85 | * is used to derive a setting for ptlrpcd_per_cpt_max. | |
86 | */ | |
d7e09d03 | 87 | static int max_ptlrpcds; |
8cc7b4b9 PT |
88 | module_param(max_ptlrpcds, int, 0644); |
89 | MODULE_PARM_DESC(max_ptlrpcds, "Max ptlrpcd thread count to be started."); | |
d7e09d03 | 90 | |
c5c4c6fa OW |
91 | /* |
92 | * ptlrpcd_bind_policy is obsolete, but retained to ensure that | |
93 | * the kernel module will load on a system where it has been tuned. | |
94 | * A value other than 0 implies it was tuned, in which case the value | |
95 | * is used to derive a setting for ptlrpcd_partner_group_size. | |
96 | */ | |
97 | static int ptlrpcd_bind_policy; | |
8cc7b4b9 | 98 | module_param(ptlrpcd_bind_policy, int, 0644); |
c5c4c6fa OW |
99 | MODULE_PARM_DESC(ptlrpcd_bind_policy, |
100 | "Ptlrpcd threads binding mode (obsolete)."); | |
101 | ||
102 | /* | |
103 | * ptlrpcd_per_cpt_max: The maximum number of ptlrpcd threads to run | |
104 | * in a CPT. | |
105 | */ | |
106 | static int ptlrpcd_per_cpt_max; | |
107 | module_param(ptlrpcd_per_cpt_max, int, 0644); | |
108 | MODULE_PARM_DESC(ptlrpcd_per_cpt_max, | |
109 | "Max ptlrpcd thread count to be started per cpt."); | |
110 | ||
111 | /* | |
112 | * ptlrpcd_partner_group_size: The desired number of threads in each | |
113 | * ptlrpcd partner thread group. Default is 2, corresponding to the | |
114 | * old PDB_POLICY_PAIR. A negative value makes all ptlrpcd threads in | |
115 | * a CPT partners of each other. | |
116 | */ | |
117 | static int ptlrpcd_partner_group_size; | |
118 | module_param(ptlrpcd_partner_group_size, int, 0644); | |
119 | MODULE_PARM_DESC(ptlrpcd_partner_group_size, | |
120 | "Number of ptlrpcd threads in a partner group."); | |
121 | ||
122 | /* | |
123 | * ptlrpcd_cpts: A CPT string describing the CPU partitions that | |
124 | * ptlrpcd threads should run on. Used to make ptlrpcd threads run on | |
125 | * a subset of all CPTs. | |
126 | * | |
127 | * ptlrpcd_cpts=2 | |
128 | * ptlrpcd_cpts=[2] | |
129 | * run ptlrpcd threads only on CPT 2. | |
130 | * | |
131 | * ptlrpcd_cpts=0-3 | |
132 | * ptlrpcd_cpts=[0-3] | |
133 | * run ptlrpcd threads on CPTs 0, 1, 2, and 3. | |
134 | * | |
135 | * ptlrpcd_cpts=[0-3,5,7] | |
136 | * run ptlrpcd threads on CPTS 0, 1, 2, 3, 5, and 7. | |
137 | */ | |
138 | static char *ptlrpcd_cpts; | |
139 | module_param(ptlrpcd_cpts, charp, 0644); | |
140 | MODULE_PARM_DESC(ptlrpcd_cpts, | |
141 | "CPU partitions ptlrpcd threads should run in"); | |
142 | ||
143 | /* ptlrpcds_cpt_idx maps cpt numbers to an index in the ptlrpcds array. */ | |
144 | static int *ptlrpcds_cpt_idx; | |
145 | ||
146 | /* ptlrpcds_num is the number of entries in the ptlrpcds array. */ | |
147 | static int ptlrpcds_num; | |
148 | static struct ptlrpcd **ptlrpcds; | |
149 | ||
150 | /* | |
151 | * In addition to the regular thread pool above, there is a single | |
152 | * global recovery thread. Recovery isn't critical for performance, | |
153 | * and doesn't block, but must always be able to proceed, and it is | |
154 | * possible that all normal ptlrpcd threads are blocked. Hence the | |
155 | * need for a dedicated thread. | |
156 | */ | |
157 | static struct ptlrpcd_ctl ptlrpcd_rcv; | |
d7e09d03 PT |
158 | |
159 | struct mutex ptlrpcd_mutex; | |
225f597c | 160 | static int ptlrpcd_users; |
d7e09d03 PT |
161 | |
162 | void ptlrpcd_wake(struct ptlrpc_request *req) | |
163 | { | |
164 | struct ptlrpc_request_set *rq_set = req->rq_set; | |
165 | ||
166 | LASSERT(rq_set != NULL); | |
167 | ||
168 | wake_up(&rq_set->set_waitq); | |
169 | } | |
170 | EXPORT_SYMBOL(ptlrpcd_wake); | |
171 | ||
172 | static struct ptlrpcd_ctl * | |
c5c4c6fa | 173 | ptlrpcd_select_pc(struct ptlrpc_request *req) |
d7e09d03 | 174 | { |
c5c4c6fa OW |
175 | struct ptlrpcd *pd; |
176 | int cpt; | |
177 | int idx; | |
d7e09d03 PT |
178 | |
179 | if (req != NULL && req->rq_send_state != LUSTRE_IMP_FULL) | |
c5c4c6fa OW |
180 | return &ptlrpcd_rcv; |
181 | ||
182 | cpt = cfs_cpt_current(cfs_cpt_table, 1); | |
183 | if (!ptlrpcds_cpt_idx) | |
184 | idx = cpt; | |
185 | else | |
186 | idx = ptlrpcds_cpt_idx[cpt]; | |
187 | pd = ptlrpcds[idx]; | |
188 | ||
d7e09d03 | 189 | /* We do not care whether it is strict load balance. */ |
c5c4c6fa OW |
190 | idx = pd->pd_cursor; |
191 | if (++idx == pd->pd_nthreads) | |
192 | idx = 0; | |
193 | pd->pd_cursor = idx; | |
d7e09d03 | 194 | |
c5c4c6fa | 195 | return &pd->pd_threads[idx]; |
d7e09d03 PT |
196 | } |
197 | ||
d7e09d03 PT |
198 | /** |
199 | * Return transferred RPCs count. | |
200 | */ | |
201 | static int ptlrpcd_steal_rqset(struct ptlrpc_request_set *des, | |
202 | struct ptlrpc_request_set *src) | |
203 | { | |
204 | struct list_head *tmp, *pos; | |
205 | struct ptlrpc_request *req; | |
206 | int rc = 0; | |
207 | ||
208 | spin_lock(&src->set_new_req_lock); | |
209 | if (likely(!list_empty(&src->set_new_requests))) { | |
210 | list_for_each_safe(pos, tmp, &src->set_new_requests) { | |
211 | req = list_entry(pos, struct ptlrpc_request, | |
212 | rq_set_chain); | |
213 | req->rq_set = des; | |
214 | } | |
215 | list_splice_init(&src->set_new_requests, | |
216 | &des->set_requests); | |
217 | rc = atomic_read(&src->set_new_count); | |
218 | atomic_add(rc, &des->set_remaining); | |
219 | atomic_set(&src->set_new_count, 0); | |
220 | } | |
221 | spin_unlock(&src->set_new_req_lock); | |
222 | return rc; | |
223 | } | |
224 | ||
225 | /** | |
226 | * Requests that are added to the ptlrpcd queue are sent via | |
227 | * ptlrpcd_check->ptlrpc_check_set(). | |
228 | */ | |
c5c4c6fa | 229 | void ptlrpcd_add_req(struct ptlrpc_request *req) |
d7e09d03 PT |
230 | { |
231 | struct ptlrpcd_ctl *pc; | |
232 | ||
233 | if (req->rq_reqmsg) | |
234 | lustre_msg_set_jobid(req->rq_reqmsg, NULL); | |
235 | ||
236 | spin_lock(&req->rq_lock); | |
237 | if (req->rq_invalid_rqset) { | |
238 | struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(5), | |
239 | back_to_sleep, NULL); | |
240 | ||
241 | req->rq_invalid_rqset = 0; | |
242 | spin_unlock(&req->rq_lock); | |
243 | l_wait_event(req->rq_set_waitq, (req->rq_set == NULL), &lwi); | |
244 | } else if (req->rq_set) { | |
b6da17f3 | 245 | /* If we have a valid "rq_set", just reuse it to avoid double |
d7e09d03 PT |
246 | * linked. */ |
247 | LASSERT(req->rq_phase == RQ_PHASE_NEW); | |
248 | LASSERT(req->rq_send_state == LUSTRE_IMP_REPLAY); | |
249 | ||
250 | /* ptlrpc_check_set will decrease the count */ | |
251 | atomic_inc(&req->rq_set->set_remaining); | |
252 | spin_unlock(&req->rq_lock); | |
253 | wake_up(&req->rq_set->set_waitq); | |
254 | return; | |
255 | } else { | |
256 | spin_unlock(&req->rq_lock); | |
257 | } | |
258 | ||
c5c4c6fa | 259 | pc = ptlrpcd_select_pc(req); |
d7e09d03 PT |
260 | |
261 | DEBUG_REQ(D_INFO, req, "add req [%p] to pc [%s:%d]", | |
262 | req, pc->pc_name, pc->pc_index); | |
263 | ||
264 | ptlrpc_set_add_new_req(pc, req); | |
265 | } | |
266 | EXPORT_SYMBOL(ptlrpcd_add_req); | |
267 | ||
268 | static inline void ptlrpc_reqset_get(struct ptlrpc_request_set *set) | |
269 | { | |
270 | atomic_inc(&set->set_refcount); | |
271 | } | |
272 | ||
273 | /** | |
274 | * Check if there is more work to do on ptlrpcd set. | |
275 | * Returns 1 if yes. | |
276 | */ | |
277 | static int ptlrpcd_check(struct lu_env *env, struct ptlrpcd_ctl *pc) | |
278 | { | |
279 | struct list_head *tmp, *pos; | |
280 | struct ptlrpc_request *req; | |
281 | struct ptlrpc_request_set *set = pc->pc_set; | |
282 | int rc = 0; | |
283 | int rc2; | |
d7e09d03 PT |
284 | |
285 | if (atomic_read(&set->set_new_count)) { | |
286 | spin_lock(&set->set_new_req_lock); | |
287 | if (likely(!list_empty(&set->set_new_requests))) { | |
288 | list_splice_init(&set->set_new_requests, | |
289 | &set->set_requests); | |
290 | atomic_add(atomic_read(&set->set_new_count), | |
291 | &set->set_remaining); | |
292 | atomic_set(&set->set_new_count, 0); | |
293 | /* | |
294 | * Need to calculate its timeout. | |
295 | */ | |
296 | rc = 1; | |
297 | } | |
298 | spin_unlock(&set->set_new_req_lock); | |
299 | } | |
300 | ||
301 | /* We should call lu_env_refill() before handling new requests to make | |
302 | * sure that env key the requests depending on really exists. | |
303 | */ | |
304 | rc2 = lu_env_refill(env); | |
305 | if (rc2 != 0) { | |
306 | /* | |
307 | * XXX This is very awkward situation, because | |
308 | * execution can neither continue (request | |
309 | * interpreters assume that env is set up), nor repeat | |
310 | * the loop (as this potentially results in a tight | |
311 | * loop of -ENOMEM's). | |
312 | * | |
313 | * Fortunately, refill only ever does something when | |
314 | * new modules are loaded, i.e., early during boot up. | |
315 | */ | |
316 | CERROR("Failure to refill session: %d\n", rc2); | |
0a3bdb00 | 317 | return rc; |
d7e09d03 PT |
318 | } |
319 | ||
320 | if (atomic_read(&set->set_remaining)) | |
321 | rc |= ptlrpc_check_set(env, set); | |
322 | ||
fa55c6a4 LZ |
323 | /* NB: ptlrpc_check_set has already moved completed request at the |
324 | * head of seq::set_requests */ | |
325 | list_for_each_safe(pos, tmp, &set->set_requests) { | |
326 | req = list_entry(pos, struct ptlrpc_request, rq_set_chain); | |
327 | if (req->rq_phase != RQ_PHASE_COMPLETE) | |
328 | break; | |
d7e09d03 | 329 | |
fa55c6a4 LZ |
330 | list_del_init(&req->rq_set_chain); |
331 | req->rq_set = NULL; | |
332 | ptlrpc_req_finished(req); | |
d7e09d03 PT |
333 | } |
334 | ||
335 | if (rc == 0) { | |
336 | /* | |
337 | * If new requests have been added, make sure to wake up. | |
338 | */ | |
339 | rc = atomic_read(&set->set_new_count); | |
340 | ||
341 | /* If we have nothing to do, check whether we can take some | |
342 | * work from our partner threads. */ | |
343 | if (rc == 0 && pc->pc_npartners > 0) { | |
344 | struct ptlrpcd_ctl *partner; | |
345 | struct ptlrpc_request_set *ps; | |
346 | int first = pc->pc_cursor; | |
347 | ||
348 | do { | |
349 | partner = pc->pc_partners[pc->pc_cursor++]; | |
350 | if (pc->pc_cursor >= pc->pc_npartners) | |
351 | pc->pc_cursor = 0; | |
352 | if (partner == NULL) | |
353 | continue; | |
354 | ||
355 | spin_lock(&partner->pc_lock); | |
356 | ps = partner->pc_set; | |
357 | if (ps == NULL) { | |
358 | spin_unlock(&partner->pc_lock); | |
359 | continue; | |
360 | } | |
361 | ||
362 | ptlrpc_reqset_get(ps); | |
363 | spin_unlock(&partner->pc_lock); | |
364 | ||
365 | if (atomic_read(&ps->set_new_count)) { | |
366 | rc = ptlrpcd_steal_rqset(set, ps); | |
367 | if (rc > 0) | |
2d00bd17 JP |
368 | CDEBUG(D_RPCTRACE, "transfer %d async RPCs [%d->%d]\n", |
369 | rc, partner->pc_index, | |
370 | pc->pc_index); | |
d7e09d03 PT |
371 | } |
372 | ptlrpc_reqset_put(ps); | |
373 | } while (rc == 0 && pc->pc_cursor != first); | |
374 | } | |
375 | } | |
376 | ||
0a3bdb00 | 377 | return rc; |
d7e09d03 PT |
378 | } |
379 | ||
380 | /** | |
381 | * Main ptlrpcd thread. | |
382 | * ptlrpc's code paths like to execute in process context, so we have this | |
383 | * thread which spins on a set which contains the rpcs and sends them. | |
384 | * | |
385 | */ | |
386 | static int ptlrpcd(void *arg) | |
387 | { | |
388 | struct ptlrpcd_ctl *pc = arg; | |
c5c4c6fa | 389 | struct ptlrpc_request_set *set; |
d7e09d03 | 390 | struct lu_env env = { .le_ses = NULL }; |
c5c4c6fa OW |
391 | int rc = 0; |
392 | int exit = 0; | |
d7e09d03 PT |
393 | |
394 | unshare_fs_struct(); | |
c5c4c6fa OW |
395 | if (cfs_cpt_bind(cfs_cpt_table, pc->pc_cpt) != 0) |
396 | CWARN("Failed to bind %s on CPT %d\n", pc->pc_name, pc->pc_cpt); | |
397 | ||
398 | /* | |
399 | * Allocate the request set after the thread has been bound | |
400 | * above. This is safe because no requests will be queued | |
401 | * until all ptlrpcd threads have confirmed that they have | |
402 | * successfully started. | |
403 | */ | |
404 | set = ptlrpc_prep_set(); | |
405 | if (!set) { | |
406 | rc = -ENOMEM; | |
407 | goto failed; | |
d7e09d03 | 408 | } |
c5c4c6fa OW |
409 | spin_lock(&pc->pc_lock); |
410 | pc->pc_set = set; | |
411 | spin_unlock(&pc->pc_lock); | |
d7e09d03 PT |
412 | /* |
413 | * XXX So far only "client" ptlrpcd uses an environment. In | |
414 | * the future, ptlrpcd thread (or a thread-set) has to given | |
415 | * an argument, describing its "scope". | |
416 | */ | |
417 | rc = lu_context_init(&env.le_ctx, | |
418 | LCT_CL_THREAD|LCT_REMEMBER|LCT_NOREF); | |
d7e09d03 | 419 | if (rc != 0) |
c5c4c6fa OW |
420 | goto failed; |
421 | ||
422 | complete(&pc->pc_starting); | |
d7e09d03 PT |
423 | |
424 | /* | |
9563fe8a | 425 | |
d7e09d03 PT |
426 | * This mainloop strongly resembles ptlrpc_set_wait() except that our |
427 | * set never completes. ptlrpcd_check() calls ptlrpc_check_set() when | |
428 | * there are requests in the set. New requests come in on the set's | |
429 | * new_req_list and ptlrpcd_check() moves them into the set. | |
430 | */ | |
431 | do { | |
432 | struct l_wait_info lwi; | |
433 | int timeout; | |
434 | ||
435 | timeout = ptlrpc_set_next_timeout(set); | |
436 | lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1), | |
437 | ptlrpc_expired_set, set); | |
438 | ||
439 | lu_context_enter(&env.le_ctx); | |
440 | l_wait_event(set->set_waitq, | |
441 | ptlrpcd_check(&env, pc), &lwi); | |
442 | lu_context_exit(&env.le_ctx); | |
443 | ||
444 | /* | |
445 | * Abort inflight rpcs for forced stop case. | |
446 | */ | |
447 | if (test_bit(LIOD_STOP, &pc->pc_flags)) { | |
448 | if (test_bit(LIOD_FORCE, &pc->pc_flags)) | |
449 | ptlrpc_abort_set(set); | |
450 | exit++; | |
451 | } | |
452 | ||
453 | /* | |
454 | * Let's make one more loop to make sure that ptlrpcd_check() | |
455 | * copied all raced new rpcs into the set so we can kill them. | |
456 | */ | |
457 | } while (exit < 2); | |
458 | ||
459 | /* | |
460 | * Wait for inflight requests to drain. | |
461 | */ | |
462 | if (!list_empty(&set->set_requests)) | |
463 | ptlrpc_set_wait(set); | |
464 | lu_context_fini(&env.le_ctx); | |
465 | ||
466 | complete(&pc->pc_finishing); | |
467 | ||
468 | return 0; | |
c5c4c6fa OW |
469 | failed: |
470 | pc->pc_error = rc; | |
471 | complete(&pc->pc_starting); | |
472 | return rc; | |
d7e09d03 PT |
473 | } |
474 | ||
c5c4c6fa OW |
475 | static void ptlrpcd_ctl_init(struct ptlrpcd_ctl *pc, int index, int cpt) |
476 | { | |
477 | pc->pc_index = index; | |
478 | pc->pc_cpt = cpt; | |
479 | init_completion(&pc->pc_starting); | |
480 | init_completion(&pc->pc_finishing); | |
481 | spin_lock_init(&pc->pc_lock); | |
482 | ||
483 | if (index < 0) { | |
484 | /* Recovery thread. */ | |
485 | snprintf(pc->pc_name, sizeof(pc->pc_name), "ptlrpcd_rcv"); | |
486 | } else { | |
487 | /* Regular thread. */ | |
488 | snprintf(pc->pc_name, sizeof(pc->pc_name), | |
489 | "ptlrpcd_%02d_%02d", cpt, index); | |
490 | } | |
491 | } | |
492 | ||
493 | /* XXX: We want multiple CPU cores to share the async RPC load. So we | |
494 | * start many ptlrpcd threads. We also want to reduce the ptlrpcd | |
495 | * overhead caused by data transfer cross-CPU cores. So we bind | |
496 | * all ptlrpcd threads to a CPT, in the expectation that CPTs | |
497 | * will be defined in a way that matches these boundaries. Within | |
498 | * a CPT a ptlrpcd thread can be scheduled on any available core. | |
d7e09d03 | 499 | * |
c5c4c6fa OW |
500 | * Each ptlrpcd thread has its own request queue. This can cause |
501 | * response delay if the thread is already busy. To help with | |
502 | * this we define partner threads: these are other threads bound | |
503 | * to the same CPT which will check for work in each other's | |
504 | * request queues if they have no work to do. | |
d7e09d03 | 505 | * |
c5c4c6fa OW |
506 | * The desired number of partner threads can be tuned by setting |
507 | * ptlrpcd_partner_group_size. The default is to create pairs of | |
508 | * partner threads. | |
d7e09d03 | 509 | */ |
c5c4c6fa | 510 | static int ptlrpcd_partners(struct ptlrpcd *pd, int index) |
d7e09d03 PT |
511 | { |
512 | struct ptlrpcd_ctl *pc; | |
c5c4c6fa OW |
513 | struct ptlrpcd_ctl **ppc; |
514 | int first; | |
515 | int i; | |
d7e09d03 | 516 | int rc = 0; |
c5c4c6fa OW |
517 | int size; |
518 | ||
519 | LASSERT(index >= 0 && index < pd->pd_nthreads); | |
520 | pc = &pd->pd_threads[index]; | |
521 | pc->pc_npartners = pd->pd_groupsize - 1; | |
522 | ||
523 | if (pc->pc_npartners <= 0) | |
524 | goto out; | |
d7e09d03 | 525 | |
c5c4c6fa OW |
526 | size = sizeof(struct ptlrpcd_ctl *) * pc->pc_npartners; |
527 | pc->pc_partners = kzalloc_node(size, GFP_NOFS, | |
528 | cfs_cpt_spread_node(cfs_cpt_table, | |
529 | pc->pc_cpt)); | |
530 | if (!pc->pc_partners) { | |
d7e09d03 | 531 | pc->pc_npartners = 0; |
c5c4c6fa OW |
532 | rc = -ENOMEM; |
533 | goto out; | |
d7e09d03 PT |
534 | } |
535 | ||
c5c4c6fa OW |
536 | first = index - index % pd->pd_groupsize; |
537 | ppc = pc->pc_partners; | |
538 | for (i = first; i < first + pd->pd_groupsize; i++) { | |
539 | if (i != index) | |
540 | *ppc++ = &pd->pd_threads[i]; | |
d7e09d03 | 541 | } |
c5c4c6fa | 542 | out: |
0a3bdb00 | 543 | return rc; |
d7e09d03 PT |
544 | } |
545 | ||
c5c4c6fa | 546 | int ptlrpcd_start(struct ptlrpcd_ctl *pc) |
d7e09d03 | 547 | { |
c5c4c6fa OW |
548 | struct task_struct *task; |
549 | int rc = 0; | |
d7e09d03 PT |
550 | |
551 | /* | |
552 | * Do not allow start second thread for one pc. | |
553 | */ | |
554 | if (test_and_set_bit(LIOD_START, &pc->pc_flags)) { | |
555 | CWARN("Starting second thread (%s) for same pc %p\n", | |
c5c4c6fa | 556 | pc->pc_name, pc); |
0a3bdb00 | 557 | return 0; |
d7e09d03 PT |
558 | } |
559 | ||
d7e09d03 PT |
560 | /* |
561 | * So far only "client" ptlrpcd uses an environment. In the future, | |
562 | * ptlrpcd thread (or a thread-set) has to be given an argument, | |
563 | * describing its "scope". | |
564 | */ | |
565 | rc = lu_context_init(&pc->pc_env.le_ctx, LCT_CL_THREAD|LCT_REMEMBER); | |
566 | if (rc != 0) | |
c5c4c6fa | 567 | goto out; |
d7e09d03 | 568 | |
c5c4c6fa OW |
569 | task = kthread_run(ptlrpcd, pc, "%s", pc->pc_name); |
570 | if (IS_ERR(task)) { | |
571 | rc = PTR_ERR(task); | |
572 | goto out_set; | |
573 | } | |
d7e09d03 | 574 | |
c5c4c6fa OW |
575 | wait_for_completion(&pc->pc_starting); |
576 | rc = pc->pc_error; | |
577 | if (rc != 0) | |
578 | goto out_set; | |
d7e09d03 | 579 | |
87c7d315 DE |
580 | return 0; |
581 | ||
87c7d315 DE |
582 | out_set: |
583 | if (pc->pc_set != NULL) { | |
584 | struct ptlrpc_request_set *set = pc->pc_set; | |
585 | ||
586 | spin_lock(&pc->pc_lock); | |
587 | pc->pc_set = NULL; | |
588 | spin_unlock(&pc->pc_lock); | |
589 | ptlrpc_set_destroy(set); | |
d7e09d03 | 590 | } |
c5c4c6fa | 591 | lu_context_fini(&pc->pc_env.le_ctx); |
87c7d315 DE |
592 | |
593 | out: | |
594 | clear_bit(LIOD_START, &pc->pc_flags); | |
0a3bdb00 | 595 | return rc; |
d7e09d03 PT |
596 | } |
597 | ||
598 | void ptlrpcd_stop(struct ptlrpcd_ctl *pc, int force) | |
599 | { | |
d7e09d03 PT |
600 | if (!test_bit(LIOD_START, &pc->pc_flags)) { |
601 | CWARN("Thread for pc %p was not started\n", pc); | |
23f14e79 | 602 | return; |
d7e09d03 PT |
603 | } |
604 | ||
605 | set_bit(LIOD_STOP, &pc->pc_flags); | |
606 | if (force) | |
607 | set_bit(LIOD_FORCE, &pc->pc_flags); | |
608 | wake_up(&pc->pc_set->set_waitq); | |
d7e09d03 PT |
609 | } |
610 | ||
611 | void ptlrpcd_free(struct ptlrpcd_ctl *pc) | |
612 | { | |
613 | struct ptlrpc_request_set *set = pc->pc_set; | |
d7e09d03 PT |
614 | |
615 | if (!test_bit(LIOD_START, &pc->pc_flags)) { | |
616 | CWARN("Thread for pc %p was not started\n", pc); | |
617 | goto out; | |
618 | } | |
619 | ||
620 | wait_for_completion(&pc->pc_finishing); | |
621 | lu_context_fini(&pc->pc_env.le_ctx); | |
622 | ||
623 | spin_lock(&pc->pc_lock); | |
624 | pc->pc_set = NULL; | |
625 | spin_unlock(&pc->pc_lock); | |
626 | ptlrpc_set_destroy(set); | |
627 | ||
628 | clear_bit(LIOD_START, &pc->pc_flags); | |
629 | clear_bit(LIOD_STOP, &pc->pc_flags); | |
630 | clear_bit(LIOD_FORCE, &pc->pc_flags); | |
d7e09d03 PT |
631 | |
632 | out: | |
633 | if (pc->pc_npartners > 0) { | |
634 | LASSERT(pc->pc_partners != NULL); | |
635 | ||
9ae10597 | 636 | kfree(pc->pc_partners); |
d7e09d03 PT |
637 | pc->pc_partners = NULL; |
638 | } | |
639 | pc->pc_npartners = 0; | |
c5c4c6fa | 640 | pc->pc_error = 0; |
d7e09d03 PT |
641 | } |
642 | ||
643 | static void ptlrpcd_fini(void) | |
644 | { | |
645 | int i; | |
c5c4c6fa | 646 | int j; |
d7e09d03 PT |
647 | |
648 | if (ptlrpcds != NULL) { | |
c5c4c6fa OW |
649 | for (i = 0; i < ptlrpcds_num; i++) { |
650 | if (!ptlrpcds[i]) | |
651 | break; | |
652 | for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++) | |
653 | ptlrpcd_stop(&ptlrpcds[i]->pd_threads[j], 0); | |
654 | for (j = 0; j < ptlrpcds[i]->pd_nthreads; j++) | |
655 | ptlrpcd_free(&ptlrpcds[i]->pd_threads[j]); | |
656 | kfree(ptlrpcds[i]); | |
657 | ptlrpcds[i] = NULL; | |
658 | } | |
9ae10597 | 659 | kfree(ptlrpcds); |
d7e09d03 | 660 | } |
c5c4c6fa OW |
661 | ptlrpcds_num = 0; |
662 | ||
663 | ptlrpcd_stop(&ptlrpcd_rcv, 0); | |
664 | ptlrpcd_free(&ptlrpcd_rcv); | |
665 | ||
666 | kfree(ptlrpcds_cpt_idx); | |
667 | ptlrpcds_cpt_idx = NULL; | |
d7e09d03 PT |
668 | } |
669 | ||
670 | static int ptlrpcd_init(void) | |
671 | { | |
c5c4c6fa OW |
672 | int nthreads; |
673 | int groupsize; | |
674 | int size; | |
675 | int i; | |
676 | int j; | |
677 | int rc = 0; | |
678 | struct cfs_cpt_table *cptable; | |
679 | __u32 *cpts = NULL; | |
680 | int ncpts; | |
681 | int cpt; | |
682 | struct ptlrpcd *pd; | |
683 | ||
684 | /* | |
685 | * Determine the CPTs that ptlrpcd threads will run on. | |
686 | */ | |
687 | cptable = cfs_cpt_table; | |
688 | ncpts = cfs_cpt_number(cptable); | |
689 | if (ptlrpcd_cpts) { | |
690 | struct cfs_expr_list *el; | |
691 | ||
692 | size = ncpts * sizeof(ptlrpcds_cpt_idx[0]); | |
693 | ptlrpcds_cpt_idx = kzalloc(size, GFP_KERNEL); | |
694 | if (!ptlrpcds_cpt_idx) { | |
695 | rc = -ENOMEM; | |
696 | goto out; | |
697 | } | |
698 | ||
699 | rc = cfs_expr_list_parse(ptlrpcd_cpts, | |
700 | strlen(ptlrpcd_cpts), | |
701 | 0, ncpts - 1, &el); | |
702 | ||
703 | if (rc != 0) { | |
704 | CERROR("ptlrpcd_cpts: invalid CPT pattern string: %s", | |
705 | ptlrpcd_cpts); | |
706 | rc = -EINVAL; | |
707 | goto out; | |
708 | } | |
709 | ||
710 | rc = cfs_expr_list_values(el, ncpts, &cpts); | |
711 | cfs_expr_list_free(el); | |
712 | if (rc <= 0) { | |
713 | CERROR("ptlrpcd_cpts: failed to parse CPT array %s: %d\n", | |
714 | ptlrpcd_cpts, rc); | |
715 | if (rc == 0) | |
716 | rc = -EINVAL; | |
717 | goto out; | |
718 | } | |
719 | ||
720 | /* | |
721 | * Create the cpt-to-index map. When there is no match | |
722 | * in the cpt table, pick a cpt at random. This could | |
723 | * be changed to take the topology of the system into | |
724 | * account. | |
725 | */ | |
726 | for (cpt = 0; cpt < ncpts; cpt++) { | |
727 | for (i = 0; i < rc; i++) | |
728 | if (cpts[i] == cpt) | |
729 | break; | |
730 | if (i >= rc) | |
731 | i = cpt % rc; | |
732 | ptlrpcds_cpt_idx[cpt] = i; | |
733 | } | |
734 | ||
735 | cfs_expr_list_values_free(cpts, rc); | |
736 | ncpts = rc; | |
737 | } | |
738 | ptlrpcds_num = ncpts; | |
739 | ||
740 | size = ncpts * sizeof(ptlrpcds[0]); | |
741 | ptlrpcds = kzalloc(size, GFP_KERNEL); | |
597851ac | 742 | if (!ptlrpcds) { |
a9b3e8f3 JL |
743 | rc = -ENOMEM; |
744 | goto out; | |
745 | } | |
d7e09d03 | 746 | |
c5c4c6fa OW |
747 | /* |
748 | * The max_ptlrpcds parameter is obsolete, but do something | |
749 | * sane if it has been tuned, and complain if | |
750 | * ptlrpcd_per_cpt_max has also been tuned. | |
751 | */ | |
752 | if (max_ptlrpcds != 0) { | |
753 | CWARN("max_ptlrpcds is obsolete.\n"); | |
754 | if (ptlrpcd_per_cpt_max == 0) { | |
755 | ptlrpcd_per_cpt_max = max_ptlrpcds / ncpts; | |
756 | /* Round up if there is a remainder. */ | |
757 | if (max_ptlrpcds % ncpts != 0) | |
758 | ptlrpcd_per_cpt_max++; | |
759 | CWARN("Setting ptlrpcd_per_cpt_max = %d\n", | |
760 | ptlrpcd_per_cpt_max); | |
761 | } else { | |
762 | CWARN("ptlrpd_per_cpt_max is also set!\n"); | |
763 | } | |
764 | } | |
765 | ||
766 | /* | |
767 | * The ptlrpcd_bind_policy parameter is obsolete, but do | |
768 | * something sane if it has been tuned, and complain if | |
769 | * ptlrpcd_partner_group_size is also tuned. | |
770 | */ | |
771 | if (ptlrpcd_bind_policy != 0) { | |
772 | CWARN("ptlrpcd_bind_policy is obsolete.\n"); | |
773 | if (ptlrpcd_partner_group_size == 0) { | |
774 | switch (ptlrpcd_bind_policy) { | |
775 | case 1: /* PDB_POLICY_NONE */ | |
776 | case 2: /* PDB_POLICY_FULL */ | |
777 | ptlrpcd_partner_group_size = 1; | |
778 | break; | |
779 | case 3: /* PDB_POLICY_PAIR */ | |
780 | ptlrpcd_partner_group_size = 2; | |
781 | break; | |
782 | case 4: /* PDB_POLICY_NEIGHBOR */ | |
783 | #ifdef CONFIG_NUMA | |
784 | ptlrpcd_partner_group_size = -1; /* CPT */ | |
785 | #else | |
786 | ptlrpcd_partner_group_size = 3; /* Triplets */ | |
787 | #endif | |
788 | break; | |
789 | default: /* Illegal value, use the default. */ | |
790 | ptlrpcd_partner_group_size = 2; | |
791 | break; | |
792 | } | |
793 | CWARN("Setting ptlrpcd_partner_group_size = %d\n", | |
794 | ptlrpcd_partner_group_size); | |
795 | } else { | |
796 | CWARN("ptlrpcd_partner_group_size is also set!\n"); | |
797 | } | |
798 | } | |
799 | ||
800 | if (ptlrpcd_partner_group_size == 0) | |
801 | ptlrpcd_partner_group_size = 2; | |
802 | else if (ptlrpcd_partner_group_size < 0) | |
803 | ptlrpcd_partner_group_size = -1; | |
804 | else if (ptlrpcd_per_cpt_max > 0 && | |
805 | ptlrpcd_partner_group_size > ptlrpcd_per_cpt_max) | |
806 | ptlrpcd_partner_group_size = ptlrpcd_per_cpt_max; | |
807 | ||
808 | /* | |
809 | * Start the recovery thread first. | |
810 | */ | |
811 | set_bit(LIOD_RECOVERY, &ptlrpcd_rcv.pc_flags); | |
812 | ptlrpcd_ctl_init(&ptlrpcd_rcv, -1, CFS_CPT_ANY); | |
813 | rc = ptlrpcd_start(&ptlrpcd_rcv); | |
d7e09d03 | 814 | if (rc < 0) |
a9b3e8f3 | 815 | goto out; |
d7e09d03 | 816 | |
c5c4c6fa OW |
817 | for (i = 0; i < ncpts; i++) { |
818 | if (!cpts) | |
819 | cpt = i; | |
820 | else | |
821 | cpt = cpts[i]; | |
822 | ||
823 | nthreads = cfs_cpt_weight(cptable, cpt); | |
824 | if (ptlrpcd_per_cpt_max > 0 && ptlrpcd_per_cpt_max < nthreads) | |
825 | nthreads = ptlrpcd_per_cpt_max; | |
826 | if (nthreads < 2) | |
827 | nthreads = 2; | |
828 | ||
829 | if (ptlrpcd_partner_group_size <= 0) { | |
830 | groupsize = nthreads; | |
831 | } else if (nthreads <= ptlrpcd_partner_group_size) { | |
832 | groupsize = nthreads; | |
833 | } else { | |
834 | groupsize = ptlrpcd_partner_group_size; | |
835 | if (nthreads % groupsize != 0) | |
836 | nthreads += groupsize - (nthreads % groupsize); | |
837 | } | |
838 | ||
839 | size = offsetof(struct ptlrpcd, pd_threads[nthreads]); | |
840 | pd = kzalloc_node(size, GFP_NOFS, | |
841 | cfs_cpt_spread_node(cfs_cpt_table, cpt)); | |
842 | if (!pd) { | |
843 | rc = -ENOMEM; | |
a9b3e8f3 | 844 | goto out; |
c5c4c6fa OW |
845 | } |
846 | pd->pd_size = size; | |
847 | pd->pd_index = i; | |
848 | pd->pd_cpt = cpt; | |
849 | pd->pd_cursor = 0; | |
850 | pd->pd_nthreads = nthreads; | |
851 | pd->pd_groupsize = groupsize; | |
852 | ptlrpcds[i] = pd; | |
d7e09d03 | 853 | |
c5c4c6fa OW |
854 | /* |
855 | * The ptlrpcd threads in a partner group can access | |
856 | * each other's struct ptlrpcd_ctl, so these must be | |
857 | * initialized before any thread is started. | |
858 | */ | |
859 | for (j = 0; j < nthreads; j++) { | |
860 | ptlrpcd_ctl_init(&pd->pd_threads[j], j, cpt); | |
861 | rc = ptlrpcd_partners(pd, j); | |
862 | if (rc < 0) | |
863 | goto out; | |
864 | } | |
d7e09d03 | 865 | |
c5c4c6fa OW |
866 | /* XXX: We start nthreads ptlrpc daemons. |
867 | * Each of them can process any non-recovery | |
868 | * async RPC to improve overall async RPC | |
869 | * efficiency. | |
870 | * | |
871 | * But there are some issues with async I/O RPCs | |
872 | * and async non-I/O RPCs processed in the same | |
873 | * set under some cases. The ptlrpcd may be | |
874 | * blocked by some async I/O RPC(s), then will | |
875 | * cause other async non-I/O RPC(s) can not be | |
876 | * processed in time. | |
877 | * | |
878 | * Maybe we should distinguish blocked async RPCs | |
879 | * from non-blocked async RPCs, and process them | |
880 | * in different ptlrpcd sets to avoid unnecessary | |
881 | * dependency. But how to distribute async RPCs | |
882 | * load among all the ptlrpc daemons becomes | |
883 | * another trouble. | |
884 | */ | |
885 | for (j = 0; j < nthreads; j++) { | |
886 | rc = ptlrpcd_start(&pd->pd_threads[j]); | |
887 | if (rc < 0) | |
888 | goto out; | |
889 | } | |
d7e09d03 | 890 | } |
c5c4c6fa OW |
891 | out: |
892 | if (rc != 0) | |
893 | ptlrpcd_fini(); | |
d7e09d03 | 894 | |
c5c4c6fa | 895 | return rc; |
d7e09d03 PT |
896 | } |
897 | ||
898 | int ptlrpcd_addref(void) | |
899 | { | |
900 | int rc = 0; | |
d7e09d03 PT |
901 | |
902 | mutex_lock(&ptlrpcd_mutex); | |
903 | if (++ptlrpcd_users == 1) | |
904 | rc = ptlrpcd_init(); | |
905 | mutex_unlock(&ptlrpcd_mutex); | |
0a3bdb00 | 906 | return rc; |
d7e09d03 PT |
907 | } |
908 | EXPORT_SYMBOL(ptlrpcd_addref); | |
909 | ||
910 | void ptlrpcd_decref(void) | |
911 | { | |
912 | mutex_lock(&ptlrpcd_mutex); | |
913 | if (--ptlrpcd_users == 0) | |
914 | ptlrpcd_fini(); | |
915 | mutex_unlock(&ptlrpcd_mutex); | |
916 | } | |
917 | EXPORT_SYMBOL(ptlrpcd_decref); | |
918 | /** @} ptlrpcd */ |