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 = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
86 lustre_swab_ost_body);
87 if (body == NULL && rc == 0)
91 oscc = req->rq_async_args.pointer_arg[0];
92 LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
94 spin_lock(&oscc->oscc_lock);
95 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
99 int diff = body->oa.o_id - oscc->oscc_last_id;
101 /* oscc_internal_create() stores the original value of
102 * grow_count in rq_async_args.space[0].
103 * We can't compare against oscc_grow_count directly,
104 * because it may have been increased while the RPC
105 * is in flight, so we would always find ourselves
106 * having created fewer objects and decreasing the
107 * precreate request size. b=18577 */
108 if (diff < (int) req->rq_async_args.space[0]) {
109 /* the OST has not managed to create all the
110 * objects we asked for */
111 oscc->oscc_grow_count = max(diff,
113 /* don't bump grow_count next time */
114 oscc->oscc_flags |= OSCC_FLAG_LOW;
116 /* the OST is able to keep up with the work,
117 * we could consider increasing grow_count
118 * next time if needed */
119 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
121 oscc->oscc_last_id = body->oa.o_id;
123 spin_unlock(&oscc->oscc_lock);
129 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
130 if (body && rc == -ENOSPC) {
131 oscc->oscc_grow_count = OST_MIN_PRECREATE;
132 oscc->oscc_last_id = body->oa.o_id;
134 spin_unlock(&oscc->oscc_lock);
135 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
139 /* filter always set body->oa.o_id as the last_id
140 * of filter (see filter_handle_precreate for detail)*/
141 if (body && body->oa.o_id > oscc->oscc_last_id)
142 oscc->oscc_last_id = body->oa.o_id;
143 spin_unlock(&oscc->oscc_lock);
147 /* aka EAGAIN we should not delay create if import failed -
148 * this avoid client stick in create and avoid race with
150 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
151 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
152 spin_unlock(&oscc->oscc_lock);
156 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
157 oscc->oscc_grow_count = OST_MIN_PRECREATE;
158 spin_unlock(&oscc->oscc_lock);
159 DEBUG_REQ(D_ERROR, req,
160 "Unknown rc %d from async create: failing oscc", rc);
161 ptlrpc_fail_import(req->rq_import,
162 lustre_msg_get_conn_cnt(req->rq_reqmsg));
166 CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
167 oscc->oscc_last_id, oscc->oscc_next_id);
169 spin_lock(&oscc->oscc_lock);
170 list_for_each_entry_safe(fake_req, pos,
171 &oscc->oscc_wait_create_list, rq_list) {
172 if (handle_async_create(fake_req, rc) == -EAGAIN) {
173 oscc_internal_create(oscc);
174 /* sending request should be never fail because
175 * osc use preallocated requests pool */
176 GOTO(exit_wakeup, rc);
179 spin_unlock(&oscc->oscc_lock);
182 cfs_waitq_signal(&oscc->oscc_waitq);
186 static int oscc_internal_create(struct osc_creator *oscc)
188 struct ptlrpc_request *request;
189 struct ost_body *body;
190 __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
193 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
195 if(oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
196 spin_unlock(&oscc->oscc_lock);
200 /* we need check it before OSCC_FLAG_CREATING - because need
201 * see lower number of precreate objects */
202 if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
203 ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
204 (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
205 (oscc->oscc_grow_count / 4 + 1)) {
206 oscc->oscc_flags |= OSCC_FLAG_LOW;
207 oscc->oscc_grow_count *= 2;
210 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
211 spin_unlock(&oscc->oscc_lock);
215 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
216 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
218 oscc->oscc_flags |= OSCC_FLAG_CREATING;
219 spin_unlock(&oscc->oscc_lock);
221 request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
222 LUSTRE_OST_VERSION, OST_CREATE, 2,
224 if (request == NULL) {
225 spin_lock(&oscc->oscc_lock);
226 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
227 spin_unlock(&oscc->oscc_lock);
231 request->rq_request_portal = OST_CREATE_PORTAL;
232 ptlrpc_at_set_req_timeout(request);
233 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
235 spin_lock(&oscc->oscc_lock);
236 body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
237 body->oa.o_gr = oscc->oscc_oa.o_gr;
238 LASSERT_MDS_GROUP(body->oa.o_gr);
239 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
240 request->rq_async_args.space[0] = oscc->oscc_grow_count;
241 spin_unlock(&oscc->oscc_lock);
242 CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
243 body->oa.o_id, oscc->oscc_last_id);
245 /* we should not resend create request - anyway we will have delorphan
246 * and kill these objects */
247 request->rq_no_delay = request->rq_no_resend = 1;
248 ptlrpc_req_set_repsize(request, 2, size);
250 request->rq_async_args.pointer_arg[0] = oscc;
251 request->rq_interpret_reply = osc_interpret_create;
252 ptlrpcd_add_req(request, PSCOPE_OTHER);
257 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
259 return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
263 static int oscc_has_objects(struct osc_creator *oscc, int count)
267 spin_lock(&oscc->oscc_lock);
268 have_objs = oscc_has_objects_nolock(oscc, count);
269 spin_unlock(&oscc->oscc_lock);
274 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
280 osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
282 spin_lock(&oscc->oscc_lock);
283 ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
284 have_objs = oscc_has_objects_nolock(oscc, count);
285 osc_invalid |= oscc->oscc_flags & OSCC_FLAG_EXITING;
287 if (!ost_full && !osc_invalid)
288 /* they release lock himself */
289 oscc_internal_create(oscc);
291 spin_unlock(&oscc->oscc_lock);
293 return have_objs || ost_full || osc_invalid;
296 static int oscc_precreate(struct osc_creator *oscc)
298 struct l_wait_info lwi;
302 if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
305 /* we should be not block forever - because client's create rpc can
306 * stick in mds for long time and forbid client reconnect */
307 lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
310 rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
312 if (!oscc_has_objects(oscc, 1) || (oscc->oscc_flags & OSCC_FLAG_NOSPC))
315 if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
321 static int oscc_recovering(struct osc_creator *oscc)
325 spin_lock(&oscc->oscc_lock);
326 recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
327 spin_unlock(&oscc->oscc_lock);
332 static int oscc_in_sync(struct osc_creator *oscc)
336 spin_lock(&oscc->oscc_lock);
337 sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
338 spin_unlock(&oscc->oscc_lock);
343 /* decide if the OST has remaining object, return value :
344 0 : the OST has remaining object, and don't need to do precreate.
345 1 : the OST has no remaining object, and will send a RPC for precreate.
346 2 : the OST has no remaining object, and will not get any for
347 a potentially very long time
350 int osc_precreate(struct obd_export *exp)
352 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
353 struct obd_import *imp = exp->exp_imp_reverse;
356 LASSERT(oscc != NULL);
357 if (imp != NULL && imp->imp_deactive)
360 /* until oscc in recovery - other flags is wrong */
361 if (oscc_recovering(oscc))
364 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
367 if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
370 spin_lock(&oscc->oscc_lock);
371 if ((oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) ||
372 (oscc->oscc_flags & OSCC_FLAG_CREATING)) {
373 spin_unlock(&oscc->oscc_lock);
377 oscc_internal_create(oscc);
381 static int handle_async_create(struct ptlrpc_request *req, int rc)
383 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
384 struct osc_creator *oscc = args->rq_oscc;
385 struct lov_stripe_md *lsm = args->rq_lsm;
386 struct obd_info *oinfo = args->rq_oinfo;
387 struct obdo *oa = oinfo->oi_oa;
389 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
394 if ((oscc->oscc_flags & OSCC_FLAG_EXITING))
395 GOTO(out_wake, rc = -EIO);
397 if (oscc_has_objects_nolock(oscc, 1)) {
398 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
399 oa->o_id = oscc->oscc_next_id;
400 lsm->lsm_object_id = oscc->oscc_next_id;
401 oscc->oscc_next_id++;
403 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
405 GOTO(out_wake, rc = 0);
408 /* should be try wait until recovery finished */
409 if(oscc->oscc_flags & OSCC_FLAG_RECOVERING)
412 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
413 GOTO(out_wake, rc = -ENOSPC);
415 /* we not have objects now - continue wait */
420 rc = oinfo->oi_cb_up(oinfo, rc);
421 ptlrpc_fakereq_finished(req);
426 static int async_create_interpret(const struct lu_env *env,
427 struct ptlrpc_request *req, void *data, int rc)
429 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
430 struct osc_creator *oscc = args->rq_oscc;
433 spin_lock(&oscc->oscc_lock);
434 ret = handle_async_create(req, rc);
435 spin_unlock(&oscc->oscc_lock);
440 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
441 struct lov_stripe_md **ea, struct obd_trans_info *oti)
444 struct ptlrpc_request *fake_req;
445 struct osc_create_async_args *args;
446 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
447 struct obdo *oa = oinfo->oi_oa;
450 if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0)){
451 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
452 rc = oinfo->oi_cb_up(oinfo, rc);
456 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
457 oa->o_flags == OBD_FL_RECREATE_OBJS) {
458 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
459 rc = oinfo->oi_cb_up(oinfo, rc);
463 LASSERT((*ea) != NULL);
465 fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
467 async_create_interpret);
468 if (fake_req == NULL) {
469 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
473 args = ptlrpc_req_async_args(fake_req);
474 CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
476 args->rq_oscc = oscc;
478 args->rq_oinfo = oinfo;
480 spin_lock(&oscc->oscc_lock);
482 rc = handle_async_create(fake_req, 0);
485 /* we not have objects - try wait */
486 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
488 list_add(&fake_req->rq_list,
489 &oscc->oscc_wait_create_list);
493 spin_unlock(&oscc->oscc_lock);
496 /* need free request if was error hit or
497 * objects already allocated */
498 ptlrpc_req_finished(fake_req);
500 /* EAGAIN mean - request is delayed */
506 int osc_create(struct obd_export *exp, struct obdo *oa,
507 struct lov_stripe_md **ea, struct obd_trans_info *oti)
509 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
510 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
511 struct lov_stripe_md *lsm;
517 LASSERT_MDS_GROUP(oa->o_gr);
518 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
520 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
521 oa->o_flags == OBD_FL_RECREATE_OBJS) {
522 RETURN(osc_real_create(exp, oa, ea, oti));
525 if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
526 RETURN(osc_real_create(exp, oa, ea, oti));
528 /* this is the special case where create removes orphans */
529 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
530 oa->o_flags == OBD_FL_DELORPHAN) {
531 spin_lock(&oscc->oscc_lock);
532 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
533 spin_unlock(&oscc->oscc_lock);
536 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
537 spin_unlock(&oscc->oscc_lock);
541 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
542 /* seting flag LOW we prevent extra grow precreate size
543 * and enforce use last assigned size */
544 oscc->oscc_flags |= OSCC_FLAG_LOW;
545 spin_unlock(&oscc->oscc_lock);
546 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
547 oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
549 /* delete from next_id on up */
550 oa->o_valid |= OBD_MD_FLID;
551 oa->o_id = oscc->oscc_next_id - 1;
553 rc = osc_real_create(exp, oa, ea, NULL);
555 spin_lock(&oscc->oscc_lock);
556 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
557 if (rc == 0 || rc == -ENOSPC) {
558 struct obd_connect_data *ocd;
561 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
562 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
564 oscc->oscc_last_id = oa->o_id;
565 ocd = &imp->imp_connect_data;
566 if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
567 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
568 "objid\n", oscc->oscc_obd->obd_name);
569 oscc->oscc_next_id = oa->o_id + 1;
572 /* sanity check for next objid. see bug 17025 */
573 LASSERT(oscc->oscc_next_id == oa->o_id + 1);
575 CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
576 LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
577 oscc->oscc_last_id, rc);
579 CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
580 oscc->oscc_obd->obd_name, rc);
583 cfs_waitq_signal(&oscc->oscc_waitq);
584 spin_unlock(&oscc->oscc_lock);
592 rc = obd_alloc_memmd(exp, &lsm);
598 if (oscc_in_sync(oscc))
599 CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
600 oscc->oscc_obd->obd_name);
602 rc = oscc_precreate(oscc);
604 CDEBUG(D_HA,"%s: error create %d\n",
605 oscc->oscc_obd->obd_name, rc);
609 spin_lock(&oscc->oscc_lock);
610 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
611 spin_unlock(&oscc->oscc_lock);
614 /* wakeup but recovery not finished */
615 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
617 spin_unlock(&oscc->oscc_lock);
621 if (oscc_has_objects_nolock(oscc, 1)) {
622 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
623 oa->o_id = oscc->oscc_next_id;
624 lsm->lsm_object_id = oscc->oscc_next_id;
626 oscc->oscc_next_id++;
627 spin_unlock(&oscc->oscc_lock);
629 CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
630 exp->exp_obd->obd_name, oscc->oscc_next_id);
632 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
634 spin_unlock(&oscc->oscc_lock);
638 spin_unlock(&oscc->oscc_lock);
642 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
643 obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
644 else if (*ea == NULL)
645 obd_free_memmd(exp, &lsm);
649 void oscc_init(struct obd_device *obd)
651 struct osc_creator *oscc;
656 oscc = &obd->u.cli.cl_oscc;
658 memset(oscc, 0, sizeof(*oscc));
660 cfs_waitq_init(&oscc->oscc_waitq);
661 spin_lock_init(&oscc->oscc_lock);
662 oscc->oscc_obd = obd;
663 oscc->oscc_grow_count = OST_MIN_PRECREATE;
664 oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
666 oscc->oscc_next_id = 2;
667 oscc->oscc_last_id = 1;
668 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
670 CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
672 /* XXX the export handle should give the oscc the last object */
673 /* oed->oed_oscc.oscc_last_id = exph->....; */
676 void oscc_fini(struct obd_device *obd)
678 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
682 spin_lock(&oscc->oscc_lock);
683 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
684 oscc->oscc_flags |= OSCC_FLAG_EXITING;
685 spin_unlock(&oscc->oscc_lock);