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 (c) 2003, 2010, Oracle and/or its affiliates. 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>
63 #include "osc_internal.h"
65 /* XXX need AT adjust ? */
66 #define osc_create_timeout (obd_timeout / 2)
68 struct osc_create_async_args {
69 struct osc_creator *rq_oscc;
70 struct lov_stripe_md *rq_lsm;
71 struct obd_info *rq_oinfo;
74 static int oscc_internal_create(struct osc_creator *oscc);
75 static int handle_async_create(struct ptlrpc_request *req, int rc);
77 static int osc_interpret_create(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);
150 /* aka EAGAIN we should not delay create if import failed -
151 * this avoid client stick in create and avoid race with delorphan */
152 /* EINTR means old create request was killed due to mds<>ost
153 * eviction. OSCC_FLAG_RECOVERING can already be 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_RPCTRACE, "prealloc 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;
195 __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
198 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
200 /* Do not check for a degraded OST here - bug21563/bug18539 */
201 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
202 spin_unlock(&oscc->oscc_lock);
206 /* we need check it before OSCC_FLAG_CREATING - because need
207 * see lower number of precreate objects */
208 if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
209 ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
210 (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
211 (oscc->oscc_grow_count / 4 + 1)) {
212 oscc->oscc_flags |= OSCC_FLAG_LOW;
213 oscc->oscc_grow_count *= 2;
216 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
217 spin_unlock(&oscc->oscc_lock);
221 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
222 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
224 oscc->oscc_flags |= OSCC_FLAG_CREATING;
225 spin_unlock(&oscc->oscc_lock);
227 request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
228 LUSTRE_OST_VERSION, OST_CREATE, 2,
230 if (request == NULL) {
231 spin_lock(&oscc->oscc_lock);
232 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
233 spin_unlock(&oscc->oscc_lock);
237 request->rq_request_portal = OST_CREATE_PORTAL;
238 ptlrpc_at_set_req_timeout(request);
239 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
241 spin_lock(&oscc->oscc_lock);
242 body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
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_req_set_repsize(request, 2, size);
255 request->rq_async_args.pointer_arg[0] = oscc;
256 request->rq_interpret_reply = osc_interpret_create;
257 ptlrpcd_add_req(request);
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 objects, may or may not send precreation RPC.
331 1 : the OST has no remaining object, and the sent precreation RPC
332 has not been completed yet.
333 2 : the OST has no remaining object, and will not get any for
334 a potentially very long time
337 int osc_precreate(struct obd_export *exp)
339 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
340 struct obd_import *imp = exp->exp_imp_reverse;
344 LASSERT(oscc != NULL);
345 if (imp != NULL && imp->imp_deactive)
346 GOTO(out_nolock, rc = 1000);
348 /* Handle critical states first */
349 spin_lock(&oscc->oscc_lock);
350 if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
351 oscc->oscc_flags & OSCC_FLAG_RDONLY ||
352 oscc->oscc_flags & OSCC_FLAG_EXITING)
353 GOTO(out, rc = 1000);
355 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING ||
356 oscc->oscc_flags & OSCC_FLAG_DEGRADED)
359 if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2))
362 /* Return 0, if we have at least one object - bug 22884 */
363 rc = oscc_has_objects_nolock(oscc, 1) ? 0 : 1;
365 /* Do not check for OSCC_FLAG_CREATING flag here, let
366 * osc_precreate() call oscc_internal_create() and
367 * adjust oscc_grow_count bug21563 */
368 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS)
371 if (oscc_internal_create(oscc))
372 GOTO(out_nolock, rc = 1000);
376 spin_unlock(&oscc->oscc_lock);
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 /* Handle the critical type errors first.
395 * Should we also test cl_import state as well ? */
396 if (oscc->oscc_flags & OSCC_FLAG_EXITING)
397 GOTO(out_wake, rc = -EIO);
399 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
400 GOTO(out_wake, rc = -ENOSPC);
402 if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
403 GOTO(out_wake, rc = -EROFS);
405 /* should be try wait until recovery finished */
406 if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
407 (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
410 if (oscc_has_objects_nolock(oscc, 1)) {
411 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
412 oa->o_id = oscc->oscc_next_id;
413 lsm->lsm_object_id = oscc->oscc_next_id;
414 oscc->oscc_next_id++;
416 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
418 GOTO(out_wake, rc = 0);
421 /* we don't have objects now - continue wait */
426 rc = oinfo->oi_cb_up(oinfo, rc);
427 ptlrpc_fakereq_finished(req);
432 static int async_create_interpret(struct ptlrpc_request *req, void *data, int rc)
434 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
435 struct osc_creator *oscc = args->rq_oscc;
438 spin_lock(&oscc->oscc_lock);
439 ret = handle_async_create(req, rc);
440 spin_unlock(&oscc->oscc_lock);
445 int osc_create_async(struct obd_export *exp, struct obd_info *oinfo,
446 struct lov_stripe_md **ea, struct obd_trans_info *oti)
449 struct ptlrpc_request *fake_req;
450 struct osc_create_async_args *args;
451 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
452 struct obdo *oa = oinfo->oi_oa;
455 if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0)){
456 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
457 rc = oinfo->oi_cb_up(oinfo, rc);
461 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
462 oa->o_flags == OBD_FL_RECREATE_OBJS) {
463 rc = osc_real_create(exp, oinfo->oi_oa, ea, oti);
464 rc = oinfo->oi_cb_up(oinfo, rc);
468 LASSERT((*ea) != NULL);
470 fake_req = ptlrpc_prep_fakereq(oscc->oscc_obd->u.cli.cl_import,
472 async_create_interpret);
473 if (fake_req == NULL) {
474 rc = oinfo->oi_cb_up(oinfo, -ENOMEM);
478 args = ptlrpc_req_async_args(fake_req);
479 CLASSERT(sizeof(*args) <= sizeof(fake_req->rq_async_args));
481 args->rq_oscc = oscc;
483 args->rq_oinfo = oinfo;
485 spin_lock(&oscc->oscc_lock);
487 rc = handle_async_create(fake_req, 0);
490 /* we not have objects - try wait */
491 is_add = ptlrpcd_add_req(fake_req);
493 list_add(&fake_req->rq_list,
494 &oscc->oscc_wait_create_list);
498 spin_unlock(&oscc->oscc_lock);
501 /* need free request if was error hit or
502 * objects already allocated */
503 ptlrpc_req_finished(fake_req);
505 /* EAGAIN mean - request is delayed */
512 int osc_create(struct obd_export *exp, struct obdo *oa,
513 struct lov_stripe_md **ea, struct obd_trans_info *oti)
515 struct lov_stripe_md *lsm;
516 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
517 int del_orphan = 0, rc = 0;
523 if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0))
524 RETURN(osc_real_create(exp, oa, ea, oti));
526 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
527 oa->o_flags == OBD_FL_RECREATE_OBJS) {
528 RETURN(osc_real_create(exp, oa, ea, oti));
531 /* this is the special case where create removes orphans */
532 if (oa->o_valid & OBD_MD_FLFLAGS &&
533 oa->o_flags == OBD_FL_DELORPHAN) {
534 spin_lock(&oscc->oscc_lock);
535 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
536 spin_unlock(&oscc->oscc_lock);
539 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
540 spin_unlock(&oscc->oscc_lock);
544 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
545 /* seting flag LOW we prevent extra grow precreate size
546 * and enforce use last assigned size */
547 oscc->oscc_flags |= OSCC_FLAG_LOW;
548 spin_unlock(&oscc->oscc_lock);
549 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
550 oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
554 /* delete from next_id on up */
555 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
556 oa->o_id = oscc->oscc_next_id - 1;
559 rc = osc_real_create(exp, oa, ea, NULL);
561 spin_lock(&oscc->oscc_lock);
562 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
563 if (rc == 0 || rc == -ENOSPC) {
564 struct obd_connect_data *ocd;
565 struct obd_import *imp = oscc->oscc_obd->u.cli.cl_import;
568 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
569 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
570 oscc->oscc_last_id = oa->o_id;
572 ocd = &imp->imp_connect_data;
573 if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
575 * The OST reports back in oa->o_id from where
576 * we should restart in order to skip orphan
579 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
580 "objid\n", oscc->oscc_obd->obd_name);
581 oscc->oscc_next_id = oa->o_id + 1;
584 CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
585 LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
586 oscc->oscc_last_id, rc);
588 CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
589 oscc->oscc_obd->obd_name, rc);
592 cfs_waitq_signal(&oscc->oscc_waitq);
593 spin_unlock(&oscc->oscc_lock);
600 rc = obd_alloc_memmd(exp, &lsm);
606 if (oscc_in_sync(oscc))
607 CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
608 oscc->oscc_obd->obd_name);
610 rc = oscc_precreate(oscc);
612 CDEBUG(D_HA,"%s: error create %d\n",
613 oscc->oscc_obd->obd_name, rc);
615 spin_lock(&oscc->oscc_lock);
617 /* wakeup but recovery did not finished */
618 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
619 (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
621 spin_unlock(&oscc->oscc_lock);
625 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
627 spin_unlock(&oscc->oscc_lock);
631 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
633 spin_unlock(&oscc->oscc_lock);
637 // Should we report -EIO error ?
638 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
639 spin_unlock(&oscc->oscc_lock);
643 if (oscc_has_objects_nolock(oscc, 1)) {
644 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
645 oa->o_id = oscc->oscc_next_id;
646 lsm->lsm_object_id = oscc->oscc_next_id;
648 oscc->oscc_next_id++;
649 spin_unlock(&oscc->oscc_lock);
651 CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
652 exp->exp_obd->obd_name, oscc->oscc_next_id);
656 spin_unlock(&oscc->oscc_lock);
660 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
661 obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
664 obd_free_memmd(exp, &lsm);
665 if (del_orphan != 0 && rc != -EIO)
666 /* Ignore non-IO precreate error for clear orphan */
672 void oscc_init(struct obd_device *obd)
674 struct osc_creator *oscc;
679 oscc = &obd->u.cli.cl_oscc;
681 memset(oscc, 0, sizeof(*oscc));
683 cfs_waitq_init(&oscc->oscc_waitq);
684 spin_lock_init(&oscc->oscc_lock);
685 oscc->oscc_obd = obd;
686 oscc->oscc_grow_count = OST_MIN_PRECREATE;
687 oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
689 oscc->oscc_next_id = 2;
690 oscc->oscc_last_id = 1;
691 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
693 CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
695 /* XXX the export handle should give the oscc the last object */
696 /* oed->oed_oscc.oscc_last_id = exph->....; */
699 void oscc_fini(struct obd_device *obd)
701 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
705 spin_lock(&oscc->oscc_lock);
706 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
707 oscc->oscc_flags |= OSCC_FLAG_EXITING;
708 spin_unlock(&oscc->oscc_lock);