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);
127 oscc->oscc_flags |= OSCC_FLAG_RDONLY;
131 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
132 if (body && rc == -ENOSPC) {
133 oscc->oscc_last_id = body->oa.o_id;
134 oscc->oscc_grow_count = OST_MIN_PRECREATE;
137 spin_unlock(&oscc->oscc_lock);
138 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
141 /* filter always set body->oa.o_id as the last_id
142 * of filter (see filter_handle_precreate for detail)*/
143 if (body && body->oa.o_id > oscc->oscc_last_id)
144 oscc->oscc_last_id = body->oa.o_id;
145 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 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
153 /* oscc->oscc_grow_count = OST_MIN_PRECREATE; */
154 spin_unlock(&oscc->oscc_lock);
158 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
159 oscc->oscc_grow_count = OST_MIN_PRECREATE;
160 spin_unlock(&oscc->oscc_lock);
161 DEBUG_REQ(D_ERROR, req,
162 "Unknown rc %d from async create: failing oscc", rc);
163 ptlrpc_fail_import(req->rq_import,
164 lustre_msg_get_conn_cnt(req->rq_reqmsg));
168 CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
169 oscc->oscc_last_id, oscc->oscc_next_id);
171 spin_lock(&oscc->oscc_lock);
172 list_for_each_entry_safe(fake_req, pos,
173 &oscc->oscc_wait_create_list, rq_list) {
174 if (handle_async_create(fake_req, rc) == -EAGAIN) {
175 oscc_internal_create(oscc);
176 /* sending request should be never fail because
177 * osc use preallocated requests pool */
178 GOTO(exit_wakeup, rc);
181 spin_unlock(&oscc->oscc_lock);
184 cfs_waitq_signal(&oscc->oscc_waitq);
188 static int oscc_internal_create(struct osc_creator *oscc)
190 struct ptlrpc_request *request;
191 struct ost_body *body;
192 __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
195 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
197 if(oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
198 spin_unlock(&oscc->oscc_lock);
202 /* we need check it before OSCC_FLAG_CREATING - because need
203 * see lower number of precreate objects */
204 if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
205 ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
206 (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
207 (oscc->oscc_grow_count / 4 + 1)) {
208 oscc->oscc_flags |= OSCC_FLAG_LOW;
209 oscc->oscc_grow_count *= 2;
212 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
213 spin_unlock(&oscc->oscc_lock);
217 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
218 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
220 oscc->oscc_flags |= OSCC_FLAG_CREATING;
221 spin_unlock(&oscc->oscc_lock);
223 request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
224 LUSTRE_OST_VERSION, OST_CREATE, 2,
226 if (request == NULL) {
227 spin_lock(&oscc->oscc_lock);
228 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
229 spin_unlock(&oscc->oscc_lock);
233 request->rq_request_portal = OST_CREATE_PORTAL;
234 ptlrpc_at_set_req_timeout(request);
235 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
237 spin_lock(&oscc->oscc_lock);
238 body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
239 body->oa.o_gr = oscc->oscc_oa.o_gr;
240 LASSERT_MDS_GROUP(body->oa.o_gr);
241 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
242 request->rq_async_args.space[0] = oscc->oscc_grow_count;
243 spin_unlock(&oscc->oscc_lock);
244 CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
245 body->oa.o_id, oscc->oscc_last_id);
247 /* we should not resend create request - anyway we will have delorphan
248 * and kill these objects */
249 request->rq_no_delay = request->rq_no_resend = 1;
250 ptlrpc_req_set_repsize(request, 2, size);
252 request->rq_async_args.pointer_arg[0] = oscc;
253 request->rq_interpret_reply = osc_interpret_create;
254 ptlrpcd_add_req(request, PSCOPE_OTHER);
259 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
261 return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
265 static int oscc_has_objects(struct osc_creator *oscc, int count)
269 spin_lock(&oscc->oscc_lock);
270 have_objs = oscc_has_objects_nolock(oscc, count);
271 spin_unlock(&oscc->oscc_lock);
276 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
281 ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
283 spin_lock(&oscc->oscc_lock);
284 ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
285 OSCC_FLAG_EXITING) & oscc->oscc_flags;
286 have_objs = oscc_has_objects_nolock(oscc, count);
289 /* they release lock himself */
290 oscc_internal_create(oscc);
292 spin_unlock(&oscc->oscc_lock);
294 return have_objs || ost_unusable;
297 static int oscc_precreate(struct osc_creator *oscc)
299 struct l_wait_info lwi;
303 if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
306 /* we should be not block forever - because client's create rpc can
307 * stick in mds for long time and forbid client reconnect */
308 lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
311 rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
313 if (!oscc_has_objects(oscc, 1) || (oscc->oscc_flags & OSCC_FLAG_NOSPC))
316 if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
319 if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
325 static int oscc_in_sync(struct osc_creator *oscc)
329 spin_lock(&oscc->oscc_lock);
330 sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
331 spin_unlock(&oscc->oscc_lock);
336 /* decide if the OST has remaining object, return value :
337 0 : the OST has remaining object, and don't need to do precreate.
338 1 : the OST has no remaining object, and will send a RPC for precreate.
339 2 : the OST has no remaining object, and will not get any for
340 a potentially very long time
343 int osc_precreate(struct obd_export *exp)
345 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
346 struct obd_import *imp = exp->exp_imp_reverse;
349 LASSERT(oscc != NULL);
350 if (imp != NULL && imp->imp_deactive)
353 /* until oscc in recovery - other flags is wrong */
354 spin_lock(&oscc->oscc_lock);
355 if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
356 oscc->oscc_flags & OSCC_FLAG_RDONLY) {
357 spin_unlock(&oscc->oscc_lock);
361 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING ||
362 oscc->oscc_flags & OSCC_FLAG_DEGRADED) {
363 spin_unlock(&oscc->oscc_lock);
367 if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2)) {
368 spin_unlock(&oscc->oscc_lock);
372 if ((oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) ||
373 (oscc->oscc_flags & OSCC_FLAG_CREATING)) {
374 spin_unlock(&oscc->oscc_lock);
378 oscc_internal_create(oscc);
382 static int handle_async_create(struct ptlrpc_request *req, int rc)
384 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
385 struct osc_creator *oscc = args->rq_oscc;
386 struct lov_stripe_md *lsm = args->rq_lsm;
387 struct obd_info *oinfo = args->rq_oinfo;
388 struct obdo *oa = oinfo->oi_oa;
390 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
395 if ((oscc->oscc_flags & OSCC_FLAG_EXITING))
396 GOTO(out_wake, rc = -EIO);
398 if (oscc_has_objects_nolock(oscc, 1)) {
399 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
400 oa->o_id = oscc->oscc_next_id;
401 lsm->lsm_object_id = oscc->oscc_next_id;
402 oscc->oscc_next_id++;
404 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
406 GOTO(out_wake, rc = 0);
409 /* should be try wait until recovery finished */
410 if(oscc->oscc_flags & OSCC_FLAG_RECOVERING)
413 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
414 GOTO(out_wake, rc = -ENOSPC);
416 if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
417 GOTO(out_wake, rc = -EROFS);
419 /* we not have objects now - continue wait */
424 rc = oinfo->oi_cb_up(oinfo, rc);
425 ptlrpc_fakereq_finished(req);
430 static int async_create_interpret(const struct lu_env *env,
431 struct ptlrpc_request *req, void *data, int rc)
433 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
434 struct osc_creator *oscc = args->rq_oscc;
437 spin_lock(&oscc->oscc_lock);
438 ret = handle_async_create(req, rc);
439 spin_unlock(&oscc->oscc_lock);
444 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
445 struct lov_stripe_md **ea, struct obd_trans_info *oti)
448 struct ptlrpc_request *fake_req;
449 struct osc_create_async_args *args;
450 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
451 struct obdo *oa = oinfo->oi_oa;
454 if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0)){
455 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
456 rc = oinfo->oi_cb_up(oinfo, rc);
460 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
461 oa->o_flags == OBD_FL_RECREATE_OBJS) {
462 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
463 rc = oinfo->oi_cb_up(oinfo, rc);
467 LASSERT((*ea) != NULL);
469 fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
471 async_create_interpret);
472 if (fake_req == NULL) {
473 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
477 args = ptlrpc_req_async_args(fake_req);
478 CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
480 args->rq_oscc = oscc;
482 args->rq_oinfo = oinfo;
484 spin_lock(&oscc->oscc_lock);
486 rc = handle_async_create(fake_req, 0);
489 /* we not have objects - try wait */
490 is_add = ptlrpcd_add_req(fake_req, PSCOPE_OTHER);
492 list_add(&fake_req->rq_list,
493 &oscc->oscc_wait_create_list);
497 spin_unlock(&oscc->oscc_lock);
500 /* need free request if was error hit or
501 * objects already allocated */
502 ptlrpc_req_finished(fake_req);
504 /* EAGAIN mean - request is delayed */
510 int osc_create(struct obd_export *exp, struct obdo *oa,
511 struct lov_stripe_md **ea, struct obd_trans_info *oti)
513 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
514 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
515 struct lov_stripe_md *lsm;
521 LASSERT_MDS_GROUP(oa->o_gr);
522 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
524 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
525 oa->o_flags == OBD_FL_RECREATE_OBJS) {
526 RETURN(osc_real_create(exp, oa, ea, oti));
529 if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
530 RETURN(osc_real_create(exp, oa, ea, oti));
532 /* this is the special case where create removes orphans */
533 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
534 oa->o_flags == OBD_FL_DELORPHAN) {
535 spin_lock(&oscc->oscc_lock);
536 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
537 spin_unlock(&oscc->oscc_lock);
540 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
541 spin_unlock(&oscc->oscc_lock);
545 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
546 /* seting flag LOW we prevent extra grow precreate size
547 * and enforce use last assigned size */
548 oscc->oscc_flags |= OSCC_FLAG_LOW;
549 spin_unlock(&oscc->oscc_lock);
550 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
551 oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
553 /* delete from next_id on up */
554 oa->o_valid |= OBD_MD_FLID;
555 oa->o_id = oscc->oscc_next_id - 1;
557 rc = osc_real_create(exp, oa, ea, NULL);
559 spin_lock(&oscc->oscc_lock);
560 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
561 if (rc == 0 || rc == -ENOSPC) {
562 struct obd_connect_data *ocd;
565 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
566 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
568 oscc->oscc_last_id = oa->o_id;
569 ocd = &imp->imp_connect_data;
570 if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
571 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
572 "objid\n", oscc->oscc_obd->obd_name);
573 oscc->oscc_next_id = oa->o_id + 1;
576 /* sanity check for next objid. see bug 17025 */
577 LASSERT(oscc->oscc_next_id == oa->o_id + 1);
579 CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
580 LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
581 oscc->oscc_last_id, rc);
583 CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
584 oscc->oscc_obd->obd_name, rc);
587 cfs_waitq_signal(&oscc->oscc_waitq);
588 spin_unlock(&oscc->oscc_lock);
596 rc = obd_alloc_memmd(exp, &lsm);
602 if (oscc_in_sync(oscc))
603 CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
604 oscc->oscc_obd->obd_name);
606 rc = oscc_precreate(oscc);
608 CDEBUG(D_HA,"%s: error create %d\n",
609 oscc->oscc_obd->obd_name, rc);
613 spin_lock(&oscc->oscc_lock);
614 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
615 spin_unlock(&oscc->oscc_lock);
618 /* wakeup but recovery not finished */
619 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
621 spin_unlock(&oscc->oscc_lock);
625 if (oscc_has_objects_nolock(oscc, 1)) {
626 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
627 oa->o_id = oscc->oscc_next_id;
628 lsm->lsm_object_id = oscc->oscc_next_id;
630 oscc->oscc_next_id++;
631 spin_unlock(&oscc->oscc_lock);
633 CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
634 exp->exp_obd->obd_name, oscc->oscc_next_id);
636 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
638 spin_unlock(&oscc->oscc_lock);
640 } else if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
642 spin_unlock(&oscc->oscc_lock);
646 spin_unlock(&oscc->oscc_lock);
650 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
651 obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
652 else if (*ea == NULL)
653 obd_free_memmd(exp, &lsm);
657 void oscc_init(struct obd_device *obd)
659 struct osc_creator *oscc;
664 oscc = &obd->u.cli.cl_oscc;
666 memset(oscc, 0, sizeof(*oscc));
668 cfs_waitq_init(&oscc->oscc_waitq);
669 spin_lock_init(&oscc->oscc_lock);
670 oscc->oscc_obd = obd;
671 oscc->oscc_grow_count = OST_MIN_PRECREATE;
672 oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
674 oscc->oscc_next_id = 2;
675 oscc->oscc_last_id = 1;
676 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
678 CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
680 /* XXX the export handle should give the oscc the last object */
681 /* oed->oed_oscc.oscc_last_id = exph->....; */
684 void oscc_fini(struct obd_device *obd)
686 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
690 spin_lock(&oscc->oscc_lock);
691 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
692 oscc->oscc_flags |= OSCC_FLAG_EXITING;
693 spin_unlock(&oscc->oscc_lock);