1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/osc/osc_create.c
37 * For testing and management it is treated as an obd_device,
38 * although * it does not export a full OBD method table (the
39 * requests are coming * in over the wire, so object target modules
40 * do not have a full * method table.)
42 * Author: Peter Braam <braam@clusterfs.com>
46 # define EXPORT_SYMTAB
48 #define DEBUG_SUBSYSTEM S_OSC
51 # include <libcfs/libcfs.h>
52 #else /* __KERNEL__ */
53 # include <liblustre.h>
60 #include <lustre_dlm.h>
61 #include <obd_class.h>
62 #include "osc_internal.h"
64 /* XXX need AT adjust ? */
65 #define osc_create_timeout (obd_timeout / 2)
67 struct osc_create_async_args {
68 struct osc_creator *rq_oscc;
69 struct lov_stripe_md *rq_lsm;
70 struct obd_info *rq_oinfo;
73 static int oscc_internal_create(struct osc_creator *oscc);
74 static int handle_async_create(struct ptlrpc_request *req, int rc);
76 static int osc_interpret_create(const struct lu_env *env,
77 struct ptlrpc_request *req, void *data, int rc)
79 struct osc_creator *oscc;
80 struct ost_body *body = NULL;
81 struct ptlrpc_request *fake_req, *pos;
85 body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
86 if (body == NULL && rc == 0)
90 oscc = req->rq_async_args.pointer_arg[0];
91 LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
93 spin_lock(&oscc->oscc_lock);
94 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
98 int diff = body->oa.o_id - oscc->oscc_last_id;
100 /* oscc_internal_create() stores the original value of
101 * grow_count in rq_async_args.space[0].
102 * We can't compare against oscc_grow_count directly,
103 * because it may have been increased while the RPC
104 * is in flight, so we would always find ourselves
105 * having created fewer objects and decreasing the
106 * precreate request size. b=18577 */
107 if (diff < (int) req->rq_async_args.space[0]) {
108 /* the OST has not managed to create all the
109 * objects we asked for */
110 oscc->oscc_grow_count = max(diff,
112 /* don't bump grow_count next time */
113 oscc->oscc_flags |= OSCC_FLAG_LOW;
115 /* the OST is able to keep up with the work,
116 * we could consider increasing grow_count
117 * next time if needed */
118 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
120 oscc->oscc_last_id = body->oa.o_id;
122 spin_unlock(&oscc->oscc_lock);
126 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
130 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
131 if (body && rc == -ENOSPC) {
132 oscc->oscc_last_id = body->oa.o_id;
133 oscc->oscc_grow_count = OST_MIN_PRECREATE;
136 spin_unlock(&oscc->oscc_lock);
137 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
140 /* filter always set body->oa.o_id as the last_id
141 * of filter (see filter_handle_precreate for detail)*/
142 if (body && body->oa.o_id > oscc->oscc_last_id)
143 oscc->oscc_last_id = body->oa.o_id;
144 spin_unlock(&oscc->oscc_lock);
149 /* aka EAGAIN we should not delay create if import failed -
150 * this avoid client stick in create and avoid race with
152 /* EINTR say - old create request is killed due mds<>ost
153 * eviction - OSCC_FLAG_RECOVERING can already set due
154 * IMP_DISCONN event */
155 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
156 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
157 spin_unlock(&oscc->oscc_lock);
161 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
162 oscc->oscc_grow_count = OST_MIN_PRECREATE;
163 spin_unlock(&oscc->oscc_lock);
164 DEBUG_REQ(D_ERROR, req,
165 "Unknown rc %d from async create: failing oscc", rc);
166 ptlrpc_fail_import(req->rq_import,
167 lustre_msg_get_conn_cnt(req->rq_reqmsg));
171 CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
172 oscc->oscc_last_id, oscc->oscc_next_id);
174 spin_lock(&oscc->oscc_lock);
175 list_for_each_entry_safe(fake_req, pos,
176 &oscc->oscc_wait_create_list, rq_list) {
177 if (handle_async_create(fake_req, rc) == -EAGAIN) {
178 oscc_internal_create(oscc);
179 /* sending request should be never fail because
180 * osc use preallocated requests pool */
181 GOTO(exit_wakeup, rc);
184 spin_unlock(&oscc->oscc_lock);
187 cfs_waitq_signal(&oscc->oscc_waitq);
191 static int oscc_internal_create(struct osc_creator *oscc)
193 struct ptlrpc_request *request;
194 struct ost_body *body;
197 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
199 if ((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
200 (oscc->oscc_flags & OSCC_FLAG_DEGRADED)) {
201 spin_unlock(&oscc->oscc_lock);
205 /* we need check it before OSCC_FLAG_CREATING - because need
206 * see lower number of precreate objects */
207 if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
208 ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
209 (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
210 (oscc->oscc_grow_count / 4 + 1)) {
211 oscc->oscc_flags |= OSCC_FLAG_LOW;
212 oscc->oscc_grow_count *= 2;
215 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
216 spin_unlock(&oscc->oscc_lock);
220 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
221 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
223 oscc->oscc_flags |= OSCC_FLAG_CREATING;
224 spin_unlock(&oscc->oscc_lock);
226 request = ptlrpc_request_alloc_pack(oscc->oscc_obd->u.cli.cl_import,
228 LUSTRE_OST_VERSION, OST_CREATE);
229 if (request == NULL) {
230 spin_lock(&oscc->oscc_lock);
231 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
232 spin_unlock(&oscc->oscc_lock);
236 request->rq_request_portal = OST_CREATE_PORTAL;
237 ptlrpc_at_set_req_timeout(request);
238 body = req_capsule_client_get(&request->rq_pill, &RMF_OST_BODY);
240 spin_lock(&oscc->oscc_lock);
241 body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
242 body->oa.o_gr = oscc->oscc_oa.o_gr;
243 LASSERT_MDS_GROUP(body->oa.o_gr);
244 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
245 request->rq_async_args.space[0] = oscc->oscc_grow_count;
246 spin_unlock(&oscc->oscc_lock);
247 CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
248 body->oa.o_id, oscc->oscc_last_id);
250 /* we should not resend create request - anyway we will have delorphan
251 * and kill these objects */
252 request->rq_no_delay = request->rq_no_resend = 1;
253 ptlrpc_request_set_replen(request);
255 request->rq_async_args.pointer_arg[0] = oscc;
256 request->rq_interpret_reply = osc_interpret_create;
257 ptlrpcd_add_req(request, PSCOPE_OTHER);
262 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
264 return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
268 static int oscc_has_objects(struct osc_creator *oscc, int count)
272 spin_lock(&oscc->oscc_lock);
273 have_objs = oscc_has_objects_nolock(oscc, count);
274 spin_unlock(&oscc->oscc_lock);
279 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
284 ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
286 spin_lock(&oscc->oscc_lock);
287 ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
288 OSCC_FLAG_EXITING) & oscc->oscc_flags;
289 have_objs = oscc_has_objects_nolock(oscc, count);
291 if (!ost_unusable && !have_objs)
292 /* they release lock himself */
293 have_objs = oscc_internal_create(oscc);
295 spin_unlock(&oscc->oscc_lock);
297 return have_objs || ost_unusable;
300 static int oscc_precreate(struct osc_creator *oscc)
302 struct l_wait_info lwi;
306 if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
309 /* we should be not block forever - because client's create rpc can
310 * stick in mds for long time and forbid client reconnect */
311 lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
314 rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
318 static int oscc_in_sync(struct osc_creator *oscc)
322 spin_lock(&oscc->oscc_lock);
323 sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
324 spin_unlock(&oscc->oscc_lock);
329 /* decide if the OST has remaining object, return value :
330 0 : the OST has remaining object, and don't need to do precreate.
331 1 : the OST has no remaining object, and will send a RPC for precreate.
332 2 : the OST has no remaining object, and will not get any for
333 a potentially very long time
336 int osc_precreate(struct obd_export *exp)
338 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
339 struct obd_import *imp = exp->exp_imp_reverse;
342 LASSERT(oscc != NULL);
343 if (imp != NULL && imp->imp_deactive)
346 /* Handle critical states first */
347 spin_lock(&oscc->oscc_lock);
348 if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
349 oscc->oscc_flags & OSCC_FLAG_RDONLY ||
350 oscc->oscc_flags & OSCC_FLAG_EXITING) {
351 spin_unlock(&oscc->oscc_lock);
355 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING ||
356 oscc->oscc_flags & OSCC_FLAG_DEGRADED) {
357 spin_unlock(&oscc->oscc_lock);
361 if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2)) {
362 spin_unlock(&oscc->oscc_lock);
366 if ((oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) ||
367 (oscc->oscc_flags & OSCC_FLAG_CREATING)) {
368 spin_unlock(&oscc->oscc_lock);
372 if (oscc_internal_create(oscc))
377 static int handle_async_create(struct ptlrpc_request *req, int rc)
379 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
380 struct osc_creator *oscc = args->rq_oscc;
381 struct lov_stripe_md *lsm = args->rq_lsm;
382 struct obd_info *oinfo = args->rq_oinfo;
383 struct obdo *oa = oinfo->oi_oa;
385 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
390 /* Handle the critical type errors first.
391 * Should we also test cl_import state as well ? */
392 if (oscc->oscc_flags & OSCC_FLAG_EXITING)
393 GOTO(out_wake, rc = -EIO);
395 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
396 GOTO(out_wake, rc = -ENOSPC);
398 if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
399 GOTO(out_wake, rc = -EROFS);
401 /* should be try wait until recovery finished */
402 if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
403 (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
406 if (oscc_has_objects_nolock(oscc, 1)) {
407 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
408 oa->o_id = oscc->oscc_next_id;
409 lsm->lsm_object_id = oscc->oscc_next_id;
410 oscc->oscc_next_id++;
412 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
414 GOTO(out_wake, rc = 0);
417 /* we don't have objects now - continue wait */
422 rc = oinfo->oi_cb_up(oinfo, rc);
423 ptlrpc_fakereq_finished(req);
428 static int async_create_interpret(const struct lu_env *env,
429 struct ptlrpc_request *req, void *data, int rc)
431 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
432 struct osc_creator *oscc = args->rq_oscc;
435 spin_lock(&oscc->oscc_lock);
436 ret = handle_async_create(req, rc);
437 spin_unlock(&oscc->oscc_lock);
442 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
443 struct lov_stripe_md **ea, struct obd_trans_info *oti)
446 struct ptlrpc_request *fake_req;
447 struct osc_create_async_args *args;
448 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
449 struct obdo *oa = oinfo->oi_oa;
452 if ((oa->o_valid & OBD_MD_FLGROUP) && !filter_group_is_mds(oa->o_gr)) {
453 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
454 rc = oinfo->oi_cb_up(oinfo, rc);
458 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
459 oa->o_flags == OBD_FL_RECREATE_OBJS) {
460 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
461 rc = oinfo->oi_cb_up(oinfo, rc);
465 LASSERT((*ea) != NULL);
467 fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
469 async_create_interpret);
470 if (fake_req == NULL) {
471 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
475 args = ptlrpc_req_async_args(fake_req);
476 CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
478 args->rq_oscc = oscc;
480 args->rq_oinfo = oinfo;
482 spin_lock(&oscc->oscc_lock);
484 rc = handle_async_create(fake_req, 0);
487 /* we not have objects - try wait */
488 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
490 list_add(&fake_req->rq_list,
491 &oscc->oscc_wait_create_list);
495 spin_unlock(&oscc->oscc_lock);
498 /* need free request if was error hit or
499 * objects already allocated */
500 ptlrpc_req_finished(fake_req);
502 /* EAGAIN mean - request is delayed */
508 int osc_create(struct obd_export *exp, struct obdo *oa,
509 struct lov_stripe_md **ea, struct obd_trans_info *oti)
511 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
512 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
513 struct lov_stripe_md *lsm;
514 int del_orphan = 0, rc = 0;
519 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
521 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
522 oa->o_flags == OBD_FL_RECREATE_OBJS) {
523 RETURN(osc_real_create(exp, oa, ea, oti));
526 if (!filter_group_is_mds(oa->o_gr))
527 RETURN(osc_real_create(exp, oa, ea, oti));
529 /* this is the special case where create removes orphans */
530 if (oa->o_valid & OBD_MD_FLFLAGS &&
531 oa->o_flags == OBD_FL_DELORPHAN) {
532 spin_lock(&oscc->oscc_lock);
533 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
534 spin_unlock(&oscc->oscc_lock);
537 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
538 spin_unlock(&oscc->oscc_lock);
542 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
543 /* seting flag LOW we prevent extra grow precreate size
544 * and enforce use last assigned size */
545 oscc->oscc_flags |= OSCC_FLAG_LOW;
546 spin_unlock(&oscc->oscc_lock);
547 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
548 oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
552 /* delete from next_id on up */
553 oa->o_valid |= OBD_MD_FLID;
554 oa->o_id = oscc->oscc_next_id - 1;
556 rc = osc_real_create(exp, oa, ea, NULL);
558 spin_lock(&oscc->oscc_lock);
559 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
560 if (rc == 0 || rc == -ENOSPC) {
561 struct obd_connect_data *ocd;
564 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
565 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
567 oscc->oscc_last_id = oa->o_id;
568 ocd = &imp->imp_connect_data;
569 if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
570 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
571 "objid\n", oscc->oscc_obd->obd_name);
572 oscc->oscc_next_id = oa->o_id + 1;
575 /* sanity check for next objid. see bug 17025 */
576 LASSERT(oscc->oscc_next_id == oa->o_id + 1);
578 CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
579 LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
580 oscc->oscc_last_id, rc);
582 CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
583 oscc->oscc_obd->obd_name, rc);
586 cfs_waitq_signal(&oscc->oscc_waitq);
587 spin_unlock(&oscc->oscc_lock);
595 rc = obd_alloc_memmd(exp, &lsm);
601 if (oscc_in_sync(oscc))
602 CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
603 oscc->oscc_obd->obd_name);
605 rc = oscc_precreate(oscc);
607 CDEBUG(D_HA,"%s: error create %d\n",
608 oscc->oscc_obd->obd_name, rc);
610 spin_lock(&oscc->oscc_lock);
612 /* wakeup but recovery did not finished */
613 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
614 (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
616 spin_unlock(&oscc->oscc_lock);
620 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
622 spin_unlock(&oscc->oscc_lock);
626 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
628 spin_unlock(&oscc->oscc_lock);
632 // Should we report -EIO error ?
633 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
634 spin_unlock(&oscc->oscc_lock);
638 if (oscc_has_objects_nolock(oscc, 1)) {
639 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
640 oa->o_id = oscc->oscc_next_id;
641 lsm->lsm_object_id = oscc->oscc_next_id;
643 oscc->oscc_next_id++;
644 spin_unlock(&oscc->oscc_lock);
646 CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
647 exp->exp_obd->obd_name, oscc->oscc_next_id);
651 spin_unlock(&oscc->oscc_lock);
655 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
656 obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
659 obd_free_memmd(exp, &lsm);
660 if (del_orphan != 0 && rc != -EIO)
661 /* Ignore non-IO precreate error for clear orphan */
667 void oscc_init(struct obd_device *obd)
669 struct osc_creator *oscc;
674 oscc = &obd->u.cli.cl_oscc;
676 memset(oscc, 0, sizeof(*oscc));
678 cfs_waitq_init(&oscc->oscc_waitq);
679 spin_lock_init(&oscc->oscc_lock);
680 oscc->oscc_obd = obd;
681 oscc->oscc_grow_count = OST_MIN_PRECREATE;
682 oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
684 oscc->oscc_next_id = 2;
685 oscc->oscc_last_id = 1;
686 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
688 CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
690 /* XXX the export handle should give the oscc the last object */
691 /* oed->oed_oscc.oscc_last_id = exph->....; */
694 void oscc_fini(struct obd_device *obd)
696 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
700 spin_lock(&oscc->oscc_lock);
701 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
702 oscc->oscc_flags |= OSCC_FLAG_EXITING;
703 spin_unlock(&oscc->oscc_lock);