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 (oscc->oscc_flags & OSCC_FLAG_DEGRADED)) {
199 spin_unlock(&oscc->oscc_lock);
203 /* we need check it before OSCC_FLAG_CREATING - because need
204 * see lower number of precreate objects */
205 if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
206 ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
207 (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
208 (oscc->oscc_grow_count / 4 + 1)) {
209 oscc->oscc_flags |= OSCC_FLAG_LOW;
210 oscc->oscc_grow_count *= 2;
213 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
214 spin_unlock(&oscc->oscc_lock);
218 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
219 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
221 oscc->oscc_flags |= OSCC_FLAG_CREATING;
222 spin_unlock(&oscc->oscc_lock);
224 request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
225 LUSTRE_OST_VERSION, OST_CREATE, 2,
227 if (request == NULL) {
228 spin_lock(&oscc->oscc_lock);
229 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
230 spin_unlock(&oscc->oscc_lock);
234 request->rq_request_portal = OST_CREATE_PORTAL;
235 ptlrpc_at_set_req_timeout(request);
236 body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
238 spin_lock(&oscc->oscc_lock);
239 body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
240 body->oa.o_gr = oscc->oscc_oa.o_gr;
241 LASSERT_MDS_GROUP(body->oa.o_gr);
242 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
243 request->rq_async_args.space[0] = oscc->oscc_grow_count;
244 spin_unlock(&oscc->oscc_lock);
245 CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
246 body->oa.o_id, oscc->oscc_last_id);
248 /* we should not resend create request - anyway we will have delorphan
249 * and kill these objects */
250 request->rq_no_delay = request->rq_no_resend = 1;
251 ptlrpc_req_set_repsize(request, 2, size);
253 request->rq_async_args.pointer_arg[0] = oscc;
254 request->rq_interpret_reply = osc_interpret_create;
255 ptlrpcd_add_req(request, PSCOPE_OTHER);
260 static int oscc_has_objects_nolock(struct osc_creator *oscc, int count)
262 return ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
266 static int oscc_has_objects(struct osc_creator *oscc, int count)
270 spin_lock(&oscc->oscc_lock);
271 have_objs = oscc_has_objects_nolock(oscc, count);
272 spin_unlock(&oscc->oscc_lock);
277 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
282 ost_unusable = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
284 spin_lock(&oscc->oscc_lock);
285 ost_unusable |= (OSCC_FLAG_NOSPC | OSCC_FLAG_RDONLY |
286 OSCC_FLAG_EXITING) & oscc->oscc_flags;
287 have_objs = oscc_has_objects_nolock(oscc, count);
289 if (!ost_unusable && !have_objs)
290 /* they release lock himself */
291 have_objs = oscc_internal_create(oscc);
293 spin_unlock(&oscc->oscc_lock);
295 return have_objs || ost_unusable;
298 static int oscc_precreate(struct osc_creator *oscc)
300 struct l_wait_info lwi;
304 if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
307 /* we should be not block forever - because client's create rpc can
308 * stick in mds for long time and forbid client reconnect */
309 lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(osc_create_timeout)),
312 rc = l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
316 static int oscc_in_sync(struct osc_creator *oscc)
320 spin_lock(&oscc->oscc_lock);
321 sync = oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS;
322 spin_unlock(&oscc->oscc_lock);
327 /* decide if the OST has remaining object, return value :
328 0 : the OST has remaining object, and don't need to do precreate.
329 1 : the OST has no remaining object, and will send a RPC for precreate.
330 2 : the OST has no remaining object, and will not get any for
331 a potentially very long time
334 int osc_precreate(struct obd_export *exp)
336 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
337 struct obd_import *imp = exp->exp_imp_reverse;
340 LASSERT(oscc != NULL);
341 if (imp != NULL && imp->imp_deactive)
344 /* Handle critical states first */
345 spin_lock(&oscc->oscc_lock);
346 if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
347 oscc->oscc_flags & OSCC_FLAG_RDONLY ||
348 oscc->oscc_flags & OSCC_FLAG_EXITING) {
349 spin_unlock(&oscc->oscc_lock);
353 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING ||
354 oscc->oscc_flags & OSCC_FLAG_DEGRADED) {
355 spin_unlock(&oscc->oscc_lock);
359 if (oscc_has_objects_nolock(oscc, oscc->oscc_grow_count / 2)) {
360 spin_unlock(&oscc->oscc_lock);
364 if ((oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) ||
365 (oscc->oscc_flags & OSCC_FLAG_CREATING)) {
366 spin_unlock(&oscc->oscc_lock);
370 if (oscc_internal_create(oscc))
375 static int handle_async_create(struct ptlrpc_request *req, int rc)
377 struct osc_create_async_args *args = ptlrpc_req_async_args(req);
378 struct osc_creator *oscc = args->rq_oscc;
379 struct lov_stripe_md *lsm = args->rq_lsm;
380 struct obd_info *oinfo = args->rq_oinfo;
381 struct obdo *oa = oinfo->oi_oa;
383 LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
388 /* Handle the critical type errors first.
389 * Should we also test cl_import state as well ? */
390 if (oscc->oscc_flags & OSCC_FLAG_EXITING)
391 GOTO(out_wake, rc = -EIO);
393 if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
394 GOTO(out_wake, rc = -ENOSPC);
396 if (oscc->oscc_flags & OSCC_FLAG_RDONLY)
397 GOTO(out_wake, rc = -EROFS);
399 /* should be try wait until recovery finished */
400 if((oscc->oscc_flags & OSCC_FLAG_RECOVERING) ||
401 (oscc->oscc_flags & OSCC_FLAG_DEGRADED))
404 if (oscc_has_objects_nolock(oscc, 1)) {
405 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
406 oa->o_id = oscc->oscc_next_id;
407 lsm->lsm_object_id = oscc->oscc_next_id;
408 oscc->oscc_next_id++;
410 CDEBUG(D_RPCTRACE, " set oscc_next_id = "LPU64"\n",
412 GOTO(out_wake, rc = 0);
415 /* we don't 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) && !filter_group_is_mds(oa->o_gr)) {
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;
512 int del_orphan = 0, rc = 0;
517 LASSERT(oa->o_valid & OBD_MD_FLGROUP);
519 if ((oa->o_valid & OBD_MD_FLFLAGS) &&
520 oa->o_flags == OBD_FL_RECREATE_OBJS) {
521 RETURN(osc_real_create(exp, oa, ea, oti));
524 if (!filter_group_is_mds(oa->o_gr))
525 RETURN(osc_real_create(exp, oa, ea, oti));
527 /* this is the special case where create removes orphans */
528 if (oa->o_valid & OBD_MD_FLFLAGS &&
529 oa->o_flags == OBD_FL_DELORPHAN) {
530 spin_lock(&oscc->oscc_lock);
531 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
532 spin_unlock(&oscc->oscc_lock);
535 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
536 spin_unlock(&oscc->oscc_lock);
540 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
541 /* seting flag LOW we prevent extra grow precreate size
542 * and enforce use last assigned size */
543 oscc->oscc_flags |= OSCC_FLAG_LOW;
544 spin_unlock(&oscc->oscc_lock);
545 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
546 oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
550 /* delete from next_id on up */
551 oa->o_valid |= OBD_MD_FLID;
552 oa->o_id = oscc->oscc_next_id - 1;
554 rc = osc_real_create(exp, oa, ea, NULL);
556 spin_lock(&oscc->oscc_lock);
557 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
558 if (rc == 0 || rc == -ENOSPC) {
559 struct obd_connect_data *ocd;
562 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
563 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
565 oscc->oscc_last_id = oa->o_id;
566 ocd = &imp->imp_connect_data;
567 if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
568 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
569 "objid\n", oscc->oscc_obd->obd_name);
570 oscc->oscc_next_id = oa->o_id + 1;
573 /* sanity check for next objid. see bug 17025 */
574 LASSERT(oscc->oscc_next_id == oa->o_id + 1);
576 CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
577 LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
578 oscc->oscc_last_id, rc);
580 CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
581 oscc->oscc_obd->obd_name, rc);
584 cfs_waitq_signal(&oscc->oscc_waitq);
585 spin_unlock(&oscc->oscc_lock);
593 rc = obd_alloc_memmd(exp, &lsm);
599 if (oscc_in_sync(oscc))
600 CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
601 oscc->oscc_obd->obd_name);
603 rc = oscc_precreate(oscc);
605 CDEBUG(D_HA,"%s: error create %d\n",
606 oscc->oscc_obd->obd_name, rc);
608 spin_lock(&oscc->oscc_lock);
610 /* wakeup but recovery did not finished */
611 if ((oscc->oscc_obd->u.cli.cl_import->imp_invalid) ||
612 (oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
614 spin_unlock(&oscc->oscc_lock);
618 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
620 spin_unlock(&oscc->oscc_lock);
624 if (oscc->oscc_flags & OSCC_FLAG_RDONLY) {
626 spin_unlock(&oscc->oscc_lock);
630 // Should we report -EIO error ?
631 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
632 spin_unlock(&oscc->oscc_lock);
636 if (oscc_has_objects_nolock(oscc, 1)) {
637 memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
638 oa->o_id = oscc->oscc_next_id;
639 lsm->lsm_object_id = oscc->oscc_next_id;
641 oscc->oscc_next_id++;
642 spin_unlock(&oscc->oscc_lock);
644 CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
645 exp->exp_obd->obd_name, oscc->oscc_next_id);
649 spin_unlock(&oscc->oscc_lock);
653 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
654 obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
657 obd_free_memmd(exp, &lsm);
658 if (del_orphan != 0 && rc != -EIO)
659 /* Ignore non-IO precreate error for clear orphan */
665 void oscc_init(struct obd_device *obd)
667 struct osc_creator *oscc;
672 oscc = &obd->u.cli.cl_oscc;
674 memset(oscc, 0, sizeof(*oscc));
676 cfs_waitq_init(&oscc->oscc_waitq);
677 spin_lock_init(&oscc->oscc_lock);
678 oscc->oscc_obd = obd;
679 oscc->oscc_grow_count = OST_MIN_PRECREATE;
680 oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
682 oscc->oscc_next_id = 2;
683 oscc->oscc_last_id = 1;
684 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
686 CFS_INIT_LIST_HEAD(&oscc->oscc_wait_create_list);
688 /* XXX the export handle should give the oscc the last object */
689 /* oed->oed_oscc.oscc_last_id = exph->....; */
692 void oscc_fini(struct obd_device *obd)
694 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
698 spin_lock(&oscc->oscc_lock);
699 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
700 oscc->oscc_flags |= OSCC_FLAG_EXITING;
701 spin_unlock(&oscc->oscc_lock);