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, 2014, 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);
52 INIT_LIST_HEAD(&set->set_list);
53 atomic_set(&set->set_refcount, 1);
54 init_waitqueue_head(&set->set_waitq);
57 void lov_finish_set(struct lov_request_set *set)
59 struct list_head *pos, *n;
60 struct lov_request *req;
64 list_for_each_safe(pos, n, &set->set_list) {
65 req = list_entry(pos, struct lov_request, rq_link);
66 list_del_init(&req->rq_link);
68 if (req->rq_oi.oi_oa != NULL)
69 OBDO_FREE(req->rq_oi.oi_oa);
71 if (req->rq_oi.oi_osfs != NULL)
72 OBD_FREE_PTR(req->rq_oi.oi_osfs);
81 int lov_set_finished(struct lov_request_set *set, int idempotent)
83 int completes = atomic_read(&set->set_completes);
85 CDEBUG(D_INFO, "check set %d/%d\n", completes, set->set_count);
87 if (completes == set->set_count) {
90 if (atomic_inc_return(&set->set_finish_checked) == 1)
96 void lov_update_set(struct lov_request_set *set,
97 struct lov_request *req, int rc)
102 atomic_inc(&set->set_completes);
104 atomic_inc(&set->set_success);
106 wake_up(&set->set_waitq);
109 int lov_update_common_set(struct lov_request_set *set,
110 struct lov_request *req, int rc)
112 struct lov_obd *lov = &set->set_exp->exp_obd->u.lov;
115 lov_update_set(set, req, rc);
117 /* grace error on inactive ost */
118 if (rc && !(lov->lov_tgts[req->rq_idx] &&
119 lov->lov_tgts[req->rq_idx]->ltd_active))
122 /* FIXME in raid1 regime, should return 0 */
126 void lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
128 list_add_tail(&req->rq_link, &set->set_list);
133 static int lov_check_set(struct lov_obd *lov, int idx)
136 mutex_lock(&lov->lov_lock);
138 if (lov->lov_tgts[idx] == NULL ||
139 lov->lov_tgts[idx]->ltd_active ||
140 (lov->lov_tgts[idx]->ltd_exp != NULL &&
141 class_exp2cliimp(lov->lov_tgts[idx]->ltd_exp)->imp_connect_tried))
144 mutex_unlock(&lov->lov_lock);
148 /* Check if the OSC connection exists and is active.
149 * If the OSC has not yet had a chance to connect to the OST the first time,
150 * wait once for it to connect instead of returning an error.
152 int lov_check_and_wait_active(struct lov_obd *lov, int ost_idx)
154 wait_queue_head_t waitq;
155 struct l_wait_info lwi;
156 struct lov_tgt_desc *tgt;
159 mutex_lock(&lov->lov_lock);
161 tgt = lov->lov_tgts[ost_idx];
163 if (unlikely(tgt == NULL))
166 if (likely(tgt->ltd_active))
169 if (tgt->ltd_exp && class_exp2cliimp(tgt->ltd_exp)->imp_connect_tried)
172 mutex_unlock(&lov->lov_lock);
174 init_waitqueue_head(&waitq);
175 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(obd_timeout),
176 cfs_time_seconds(1), NULL, NULL);
178 rc = l_wait_event(waitq, lov_check_set(lov, ost_idx), &lwi);
185 mutex_unlock(&lov->lov_lock);
189 static int common_attr_done(struct lov_request_set *set)
191 struct list_head *pos;
192 struct lov_request *req;
194 int rc = 0, attrset = 0;
197 LASSERT(set->set_oi != NULL);
199 if (set->set_oi->oi_oa == NULL)
202 if (!atomic_read(&set->set_success))
207 GOTO(out, rc = -ENOMEM);
209 list_for_each(pos, &set->set_list) {
210 req = list_entry(pos, struct lov_request, rq_link);
212 if (!req->rq_complete || req->rq_rc)
214 if (req->rq_oi.oi_oa->o_valid == 0) /* inactive stripe */
216 lov_merge_attrs(tmp_oa, req->rq_oi.oi_oa,
217 req->rq_oi.oi_oa->o_valid,
218 set->set_oi->oi_md, req->rq_stripe, &attrset);
221 CERROR("No stripes had valid attrs\n");
225 tmp_oa->o_oi = set->set_oi->oi_oa->o_oi;
226 memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
234 int lov_fini_getattr_set(struct lov_request_set *set)
241 LASSERT(set->set_exp);
242 if (atomic_read(&set->set_completes))
243 rc = common_attr_done(set);
250 /* The callback for osc_getattr_async that finilizes a request info when a
251 * response is received. */
252 static int cb_getattr_update(void *cookie, int rc)
254 struct obd_info *oinfo = cookie;
255 struct lov_request *lovreq;
256 lovreq = container_of(oinfo, struct lov_request, rq_oi);
257 return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
260 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
261 struct lov_request_set **reqset)
263 struct lov_request_set *set;
264 struct lov_obd *lov = &exp->exp_obd->u.lov;
268 OBD_ALLOC(set, sizeof(*set));
276 for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
277 struct lov_oinfo *loi;
278 struct lov_request *req;
280 loi = oinfo->oi_md->lsm_oinfo[i];
281 if (lov_oinfo_is_dummy(loi))
284 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
285 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
289 OBD_ALLOC(req, sizeof(*req));
291 GOTO(out_set, rc = -ENOMEM);
294 req->rq_idx = loi->loi_ost_idx;
296 OBDO_ALLOC(req->rq_oi.oi_oa);
297 if (req->rq_oi.oi_oa == NULL) {
298 OBD_FREE(req, sizeof(*req));
299 GOTO(out_set, rc = -ENOMEM);
301 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
302 sizeof(*req->rq_oi.oi_oa));
303 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
304 req->rq_oi.oi_cb_up = cb_getattr_update;
305 req->rq_oi.oi_capa = oinfo->oi_capa;
307 lov_set_add_req(req, set);
310 GOTO(out_set, rc = -EIO);
314 lov_fini_getattr_set(set);
318 #define LOV_U64_MAX ((__u64)~0ULL)
319 #define LOV_SUM_MAX(tot, add) \
321 if ((tot) + (add) < (tot)) \
322 (tot) = LOV_U64_MAX; \
327 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
332 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
334 if (osfs->os_files != LOV_U64_MAX)
335 lov_do_div64(osfs->os_files, expected_stripes);
336 if (osfs->os_ffree != LOV_U64_MAX)
337 lov_do_div64(osfs->os_ffree, expected_stripes);
339 spin_lock(&obd->obd_osfs_lock);
340 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
341 obd->obd_osfs_age = cfs_time_current_64();
342 spin_unlock(&obd->obd_osfs_lock);
349 int lov_fini_statfs_set(struct lov_request_set *set)
357 if (atomic_read(&set->set_completes)) {
358 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
359 atomic_read(&set->set_success));
365 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
368 int shift = 0, quit = 0;
372 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
374 if (osfs->os_bsize != lov_sfs->os_bsize) {
375 /* assume all block sizes are always powers of 2 */
376 /* get the bits difference */
377 tmp = osfs->os_bsize | lov_sfs->os_bsize;
378 for (shift = 0; shift <= 64; ++shift) {
390 if (osfs->os_bsize < lov_sfs->os_bsize) {
391 osfs->os_bsize = lov_sfs->os_bsize;
393 osfs->os_bfree >>= shift;
394 osfs->os_bavail >>= shift;
395 osfs->os_blocks >>= shift;
396 } else if (shift != 0) {
397 lov_sfs->os_bfree >>= shift;
398 lov_sfs->os_bavail >>= shift;
399 lov_sfs->os_blocks >>= shift;
402 /* Sandia requested that df (and so, statfs) only
403 returned minimal available space on
404 a single OST, so people would be able to
405 write this much data guaranteed. */
406 if (osfs->os_bavail > lov_sfs->os_bavail) {
407 /* Presumably if new bavail is smaller,
408 new bfree is bigger as well */
409 osfs->os_bfree = lov_sfs->os_bfree;
410 osfs->os_bavail = lov_sfs->os_bavail;
413 osfs->os_bfree += lov_sfs->os_bfree;
414 osfs->os_bavail += lov_sfs->os_bavail;
416 osfs->os_blocks += lov_sfs->os_blocks;
417 /* XXX not sure about this one - depends on policy.
418 * - could be minimum if we always stripe on all OBDs
419 * (but that would be wrong for any other policy,
420 * if one of the OBDs has no more objects left)
421 * - could be sum if we stripe whole objects
422 * - could be average, just to give a nice number
424 * To give a "reasonable" (if not wholly accurate)
425 * number, we divide the total number of free objects
426 * by expected stripe count (watch out for overflow).
428 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
429 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
433 /* The callback for osc_statfs_async that finilizes a request info when a
434 * response is received. */
435 static int cb_statfs_update(void *cookie, int rc)
437 struct obd_info *oinfo = cookie;
438 struct lov_request *lovreq;
439 struct lov_request_set *set;
440 struct obd_statfs *osfs, *lov_sfs;
442 struct lov_tgt_desc *tgt;
443 struct obd_device *lovobd, *tgtobd;
447 lovreq = container_of(oinfo, struct lov_request, rq_oi);
448 set = lovreq->rq_rqset;
449 lovobd = set->set_obd;
450 lov = &lovobd->u.lov;
451 osfs = set->set_oi->oi_osfs;
452 lov_sfs = oinfo->oi_osfs;
453 success = atomic_read(&set->set_success);
454 /* XXX: the same is done in lov_update_common_set, however
455 lovset->set_exp is not initialized. */
456 lov_update_set(set, lovreq, rc);
461 tgt = lov->lov_tgts[lovreq->rq_idx];
462 if (!tgt || !tgt->ltd_active)
463 GOTO(out_update, rc);
465 tgtobd = class_exp2obd(tgt->ltd_exp);
466 spin_lock(&tgtobd->obd_osfs_lock);
467 memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
468 if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
469 tgtobd->obd_osfs_age = cfs_time_current_64();
470 spin_unlock(&tgtobd->obd_osfs_lock);
473 lov_update_statfs(osfs, lov_sfs, success);
477 if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
478 lov_set_finished(set, 0)) {
479 lov_statfs_interpret(NULL, set, set->set_count !=
480 atomic_read(&set->set_success));
486 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
487 struct lov_request_set **reqset)
489 struct lov_request_set *set;
490 struct lov_obd *lov = &obd->u.lov;
494 OBD_ALLOC(set, sizeof(*set));
502 /* We only get block data from the OBD */
503 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
504 struct lov_request *req;
506 if (lov->lov_tgts[i] == NULL ||
507 (oinfo->oi_flags & OBD_STATFS_NODELAY &&
508 !lov->lov_tgts[i]->ltd_active)) {
509 CDEBUG(D_HA, "lov idx %d inactive\n", i);
513 if (!lov->lov_tgts[i]->ltd_active)
514 lov_check_and_wait_active(lov, i);
516 /* skip targets that have been explicitely disabled by the
518 if (!lov->lov_tgts[i]->ltd_exp) {
519 CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
523 OBD_ALLOC(req, sizeof(*req));
525 GOTO(out_set, rc = -ENOMEM);
527 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
528 if (req->rq_oi.oi_osfs == NULL) {
529 OBD_FREE(req, sizeof(*req));
530 GOTO(out_set, rc = -ENOMEM);
534 req->rq_oi.oi_cb_up = cb_statfs_update;
535 req->rq_oi.oi_flags = oinfo->oi_flags;
537 lov_set_add_req(req, set);
540 GOTO(out_set, rc = -EIO);
544 lov_fini_statfs_set(set);