4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LOV
39 #include <libcfs/libcfs.h>
41 #include <obd_class.h>
42 #include <lustre/lustre_idl.h>
44 #include "lov_internal.h"
46 static void lov_init_set(struct lov_request_set *set)
49 atomic_set(&set->set_completes, 0);
50 atomic_set(&set->set_success, 0);
51 atomic_set(&set->set_finish_checked, 0);
53 INIT_LIST_HEAD(&set->set_list);
54 atomic_set(&set->set_refcount, 1);
55 init_waitqueue_head(&set->set_waitq);
58 void lov_finish_set(struct lov_request_set *set)
60 struct list_head *pos, *n;
61 struct lov_request *req;
65 list_for_each_safe(pos, n, &set->set_list) {
66 req = list_entry(pos, struct lov_request, rq_link);
67 list_del_init(&req->rq_link);
69 if (req->rq_oi.oi_oa != NULL)
70 OBDO_FREE(req->rq_oi.oi_oa);
72 if (req->rq_oi.oi_osfs != NULL)
73 OBD_FREE_PTR(req->rq_oi.oi_osfs);
82 int lov_set_finished(struct lov_request_set *set, int idempotent)
84 int completes = atomic_read(&set->set_completes);
86 CDEBUG(D_INFO, "check set %d/%d\n", completes, set->set_count);
88 if (completes == set->set_count) {
91 if (atomic_inc_return(&set->set_finish_checked) == 1)
97 void lov_update_set(struct lov_request_set *set,
98 struct lov_request *req, int rc)
100 req->rq_complete = 1;
103 atomic_inc(&set->set_completes);
105 atomic_inc(&set->set_success);
107 wake_up(&set->set_waitq);
110 int lov_update_common_set(struct lov_request_set *set,
111 struct lov_request *req, int rc)
113 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
116 lov_update_set(set, req, rc);
118 /* grace error on inactive ost */
119 if (rc && !(lov->lov_tgts[req->rq_idx] &&
120 lov->lov_tgts[req->rq_idx]->ltd_active))
123 /* FIXME in raid1 regime, should return 0 */
127 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
129 list_add_tail(&req->rq_link, &set->set_list);
134 static int lov_check_set(struct lov_obd *lov, int idx)
137 mutex_lock(&lov->lov_lock);
139 if (lov->lov_tgts[idx] == NULL ||
140 lov->lov_tgts[idx]->ltd_active ||
141 (lov->lov_tgts[idx]->ltd_exp != NULL &&
142 class_exp2cliimp(lov->lov_tgts[idx]->ltd_exp)->imp_connect_tried))
145 mutex_unlock(&lov->lov_lock);
149 /* Check if the OSC connection exists and is active.
150 * If the OSC has not yet had a chance to connect to the OST the first time,
151 * wait once for it to connect instead of returning an error.
153 int lov_check_and_wait_active(struct lov_obd *lov, int ost_idx)
155 wait_queue_head_t waitq;
156 struct l_wait_info lwi;
157 struct lov_tgt_desc *tgt;
160 mutex_lock(&lov->lov_lock);
162 tgt = lov->lov_tgts[ost_idx];
164 if (unlikely(tgt == NULL))
167 if (likely(tgt->ltd_active))
170 if (tgt->ltd_exp && class_exp2cliimp(tgt->ltd_exp)->imp_connect_tried)
173 mutex_unlock(&lov->lov_lock);
175 init_waitqueue_head(&waitq);
176 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(obd_timeout),
177 cfs_time_seconds(1), NULL, NULL);
179 rc = l_wait_event(waitq, lov_check_set(lov, ost_idx), &lwi);
186 mutex_unlock(&lov->lov_lock);
190 static int common_attr_done(struct lov_request_set *set)
192 struct list_head *pos;
193 struct lov_request *req;
195 int rc = 0, attrset = 0;
198 LASSERT(set->set_oi != NULL);
200 if (set->set_oi->oi_oa == NULL)
203 if (!atomic_read(&set->set_success))
208 GOTO(out, rc = -ENOMEM);
210 list_for_each(pos, &set->set_list) {
211 req = list_entry(pos, struct lov_request, rq_link);
213 if (!req->rq_complete || req->rq_rc)
215 if (req->rq_oi.oi_oa->o_valid == 0) /* inactive stripe */
217 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
218 req->rq_oi.oi_oa->o_valid,
219 set->set_oi->oi_md, req->rq_stripe, &attrset);
222 CERROR("No stripes had valid attrs\n");
225 if ((set->set_oi->oi_oa->o_valid & OBD_MD_FLEPOCH) &&
226 (set->set_oi->oi_md->lsm_stripe_count != attrset)) {
227 /* When we take attributes of some epoch, we require all the
228 * ost to be active. */
229 CERROR("Not all the stripes had valid attrs\n");
230 GOTO(out, rc = -EIO);
233 tmp_oa->o_oi = set->set_oi->oi_oa->o_oi;
234 memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
242 int lov_fini_getattr_set(struct lov_request_set *set)
249 LASSERT(set->set_exp);
250 if (atomic_read(&set->set_completes))
251 rc = common_attr_done(set);
258 /* The callback for osc_getattr_async that finilizes a request info when a
259 * response is received. */
260 static int cb_getattr_update(void *cookie, int rc)
262 struct obd_info *oinfo = cookie;
263 struct lov_request *lovreq;
264 lovreq = container_of(oinfo, struct lov_request, rq_oi);
265 return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
268 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
269 struct lov_request_set **reqset)
271 struct lov_request_set *set;
272 struct lov_obd *lov = &exp->exp_obd->u.lov;
276 OBD_ALLOC(set, sizeof(*set));
284 for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
285 struct lov_oinfo *loi;
286 struct lov_request *req;
288 loi = oinfo->oi_md->lsm_oinfo[i];
289 if (lov_oinfo_is_dummy(loi))
292 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
293 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
294 if (oinfo->oi_oa->o_valid & OBD_MD_FLEPOCH)
295 /* SOM requires all the OSTs to be active. */
296 GOTO(out_set, rc = -EIO);
300 OBD_ALLOC(req, sizeof(*req));
302 GOTO(out_set, rc = -ENOMEM);
305 req->rq_idx = loi->loi_ost_idx;
307 OBDO_ALLOC(req->rq_oi.oi_oa);
308 if (req->rq_oi.oi_oa == NULL) {
309 OBD_FREE(req, sizeof(*req));
310 GOTO(out_set, rc = -ENOMEM);
312 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
313 sizeof(*req->rq_oi.oi_oa));
314 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
315 req->rq_oi.oi_cb_up = cb_getattr_update;
316 req->rq_oi.oi_capa = oinfo->oi_capa;
318 lov_set_add_req(req, set);
321 GOTO(out_set, rc = -EIO);
325 lov_fini_getattr_set(set);
329 int lov_fini_destroy_set(struct lov_request_set *set)
335 LASSERT(set->set_exp);
336 if (atomic_read(&set->set_completes)) {
337 /* FIXME update qos data here */
345 int lov_prep_destroy_set(struct obd_export *exp, struct obd_info *oinfo,
346 struct obdo *src_oa, struct lov_stripe_md *lsm,
347 struct obd_trans_info *oti,
348 struct lov_request_set **reqset)
350 struct lov_request_set *set;
351 struct lov_obd *lov = &exp->exp_obd->u.lov;
355 OBD_ALLOC(set, sizeof(*set));
362 set->set_oi->oi_md = lsm;
363 set->set_oi->oi_oa = src_oa;
364 if (oti != NULL && src_oa->o_valid & OBD_MD_FLCOOKIE)
365 set->set_cookies = oti->oti_logcookies;
367 for (i = 0; i < lsm->lsm_stripe_count; i++) {
368 struct lov_oinfo *loi;
369 struct lov_request *req;
371 loi = lsm->lsm_oinfo[i];
372 if (lov_oinfo_is_dummy(loi))
375 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
376 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
380 OBD_ALLOC(req, sizeof(*req));
382 GOTO(out_set, rc = -ENOMEM);
385 req->rq_idx = loi->loi_ost_idx;
387 OBDO_ALLOC(req->rq_oi.oi_oa);
388 if (req->rq_oi.oi_oa == NULL) {
389 OBD_FREE(req, sizeof(*req));
390 GOTO(out_set, rc = -ENOMEM);
392 memcpy(req->rq_oi.oi_oa, src_oa, sizeof(*req->rq_oi.oi_oa));
393 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
394 lov_set_add_req(req, set);
397 GOTO(out_set, rc = -EIO);
401 lov_fini_destroy_set(set);
405 int lov_fini_setattr_set(struct lov_request_set *set)
412 LASSERT(set->set_exp);
413 if (atomic_read(&set->set_completes)) {
414 rc = common_attr_done(set);
415 /* FIXME update qos data here */
422 int lov_update_setattr_set(struct lov_request_set *set,
423 struct lov_request *req, int rc)
425 struct lov_obd *lov = &req->rq_rqset->set_exp->exp_obd->u.lov;
426 struct lov_stripe_md *lsm = req->rq_rqset->set_oi->oi_md;
429 lov_update_set(set, req, rc);
431 /* grace error on inactive ost */
432 if (rc && !(lov->lov_tgts[req->rq_idx] &&
433 lov->lov_tgts[req->rq_idx]->ltd_active))
437 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLCTIME)
438 lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_ctime =
439 req->rq_oi.oi_oa->o_ctime;
440 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLMTIME)
441 lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_mtime =
442 req->rq_oi.oi_oa->o_mtime;
443 if (req->rq_oi.oi_oa->o_valid & OBD_MD_FLATIME)
444 lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_atime =
445 req->rq_oi.oi_oa->o_atime;
451 /* The callback for osc_setattr_async that finilizes a request info when a
452 * response is received. */
453 static int cb_setattr_update(void *cookie, int rc)
455 struct obd_info *oinfo = cookie;
456 struct lov_request *lovreq;
457 lovreq = container_of(oinfo, struct lov_request, rq_oi);
458 return lov_update_setattr_set(lovreq->rq_rqset, lovreq, rc);
461 int lov_prep_setattr_set(struct obd_export *exp, struct obd_info *oinfo,
462 struct obd_trans_info *oti,
463 struct lov_request_set **reqset)
465 struct lov_request_set *set;
466 struct lov_obd *lov = &exp->exp_obd->u.lov;
470 OBD_ALLOC(set, sizeof(*set));
477 if (oti != NULL && oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
478 set->set_cookies = oti->oti_logcookies;
480 for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
481 struct lov_oinfo *loi = oinfo->oi_md->lsm_oinfo[i];
482 struct lov_request *req;
484 if (lov_oinfo_is_dummy(loi))
487 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
488 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
492 OBD_ALLOC(req, sizeof(*req));
494 GOTO(out_set, rc = -ENOMEM);
496 req->rq_idx = loi->loi_ost_idx;
498 OBDO_ALLOC(req->rq_oi.oi_oa);
499 if (req->rq_oi.oi_oa == NULL) {
500 OBD_FREE(req, sizeof(*req));
501 GOTO(out_set, rc = -ENOMEM);
503 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
504 sizeof(*req->rq_oi.oi_oa));
505 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
506 req->rq_oi.oi_oa->o_stripe_idx = i;
507 req->rq_oi.oi_cb_up = cb_setattr_update;
508 req->rq_oi.oi_capa = oinfo->oi_capa;
510 if (oinfo->oi_oa->o_valid & OBD_MD_FLSIZE) {
511 int off = lov_stripe_offset(oinfo->oi_md,
512 oinfo->oi_oa->o_size, i,
513 &req->rq_oi.oi_oa->o_size);
515 if (off < 0 && req->rq_oi.oi_oa->o_size)
516 req->rq_oi.oi_oa->o_size--;
518 CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n",
519 i, req->rq_oi.oi_oa->o_size,
520 oinfo->oi_oa->o_size);
522 lov_set_add_req(req, set);
525 GOTO(out_set, rc = -EIO);
529 lov_fini_setattr_set(set);
533 #define LOV_U64_MAX ((__u64)~0ULL)
534 #define LOV_SUM_MAX(tot, add) \
536 if ((tot) + (add) < (tot)) \
537 (tot) = LOV_U64_MAX; \
542 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
547 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
549 if (osfs->os_files != LOV_U64_MAX)
550 lov_do_div64(osfs->os_files, expected_stripes);
551 if (osfs->os_ffree != LOV_U64_MAX)
552 lov_do_div64(osfs->os_ffree, expected_stripes);
554 spin_lock(&obd->obd_osfs_lock);
555 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
556 obd->obd_osfs_age = cfs_time_current_64();
557 spin_unlock(&obd->obd_osfs_lock);
564 int lov_fini_statfs_set(struct lov_request_set *set)
572 if (atomic_read(&set->set_completes)) {
573 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
574 atomic_read(&set->set_success));
580 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
583 int shift = 0, quit = 0;
587 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
589 if (osfs->os_bsize != lov_sfs->os_bsize) {
590 /* assume all block sizes are always powers of 2 */
591 /* get the bits difference */
592 tmp = osfs->os_bsize | lov_sfs->os_bsize;
593 for (shift = 0; shift <= 64; ++shift) {
605 if (osfs->os_bsize < lov_sfs->os_bsize) {
606 osfs->os_bsize = lov_sfs->os_bsize;
608 osfs->os_bfree >>= shift;
609 osfs->os_bavail >>= shift;
610 osfs->os_blocks >>= shift;
611 } else if (shift != 0) {
612 lov_sfs->os_bfree >>= shift;
613 lov_sfs->os_bavail >>= shift;
614 lov_sfs->os_blocks >>= shift;
617 /* Sandia requested that df (and so, statfs) only
618 returned minimal available space on
619 a single OST, so people would be able to
620 write this much data guaranteed. */
621 if (osfs->os_bavail > lov_sfs->os_bavail) {
622 /* Presumably if new bavail is smaller,
623 new bfree is bigger as well */
624 osfs->os_bfree = lov_sfs->os_bfree;
625 osfs->os_bavail = lov_sfs->os_bavail;
628 osfs->os_bfree += lov_sfs->os_bfree;
629 osfs->os_bavail += lov_sfs->os_bavail;
631 osfs->os_blocks += lov_sfs->os_blocks;
632 /* XXX not sure about this one - depends on policy.
633 * - could be minimum if we always stripe on all OBDs
634 * (but that would be wrong for any other policy,
635 * if one of the OBDs has no more objects left)
636 * - could be sum if we stripe whole objects
637 * - could be average, just to give a nice number
639 * To give a "reasonable" (if not wholly accurate)
640 * number, we divide the total number of free objects
641 * by expected stripe count (watch out for overflow).
643 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
644 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
648 /* The callback for osc_statfs_async that finilizes a request info when a
649 * response is received. */
650 static int cb_statfs_update(void *cookie, int rc)
652 struct obd_info *oinfo = cookie;
653 struct lov_request *lovreq;
654 struct lov_request_set *set;
655 struct obd_statfs *osfs, *lov_sfs;
657 struct lov_tgt_desc *tgt;
658 struct obd_device *lovobd, *tgtobd;
662 lovreq = container_of(oinfo, struct lov_request, rq_oi);
663 set = lovreq->rq_rqset;
664 lovobd = set->set_obd;
665 lov = &lovobd->u.lov;
666 osfs = set->set_oi->oi_osfs;
667 lov_sfs = oinfo->oi_osfs;
668 success = atomic_read(&set->set_success);
669 /* XXX: the same is done in lov_update_common_set, however
670 lovset->set_exp is not initialized. */
671 lov_update_set(set, lovreq, rc);
676 tgt = lov->lov_tgts[lovreq->rq_idx];
677 if (!tgt || !tgt->ltd_active)
678 GOTO(out_update, rc);
680 tgtobd = class_exp2obd(tgt->ltd_exp);
681 spin_lock(&tgtobd->obd_osfs_lock);
682 memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
683 if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
684 tgtobd->obd_osfs_age = cfs_time_current_64();
685 spin_unlock(&tgtobd->obd_osfs_lock);
688 lov_update_statfs(osfs, lov_sfs, success);
692 if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
693 lov_set_finished(set, 0)) {
694 lov_statfs_interpret(NULL, set, set->set_count !=
695 atomic_read(&set->set_success));
701 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
702 struct lov_request_set **reqset)
704 struct lov_request_set *set;
705 struct lov_obd *lov = &obd->u.lov;
709 OBD_ALLOC(set, sizeof(*set));
717 /* We only get block data from the OBD */
718 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
719 struct lov_request *req;
721 if (lov->lov_tgts[i] == NULL ||
722 (oinfo->oi_flags & OBD_STATFS_NODELAY &&
723 !lov->lov_tgts[i]->ltd_active)) {
724 CDEBUG(D_HA, "lov idx %d inactive\n", i);
728 if (!lov->lov_tgts[i]->ltd_active)
729 lov_check_and_wait_active(lov, i);
731 /* skip targets that have been explicitely disabled by the
733 if (!lov->lov_tgts[i]->ltd_exp) {
734 CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
738 OBD_ALLOC(req, sizeof(*req));
740 GOTO(out_set, rc = -ENOMEM);
742 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
743 if (req->rq_oi.oi_osfs == NULL) {
744 OBD_FREE(req, sizeof(*req));
745 GOTO(out_set, rc = -ENOMEM);
749 req->rq_oi.oi_cb_up = cb_statfs_update;
750 req->rq_oi.oi_flags = oinfo->oi_flags;
752 lov_set_add_req(req, set);
755 GOTO(out_set, rc = -EIO);
759 lov_fini_statfs_set(set);