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 set->set_cookies = NULL;
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");
226 tmp_oa->o_oi = set->set_oi->oi_oa->o_oi;
227 memcpy(set->set_oi->oi_oa, tmp_oa, sizeof(*set->set_oi->oi_oa));
235 int lov_fini_getattr_set(struct lov_request_set *set)
242 LASSERT(set->set_exp);
243 if (atomic_read(&set->set_completes))
244 rc = common_attr_done(set);
251 /* The callback for osc_getattr_async that finilizes a request info when a
252 * response is received. */
253 static int cb_getattr_update(void *cookie, int rc)
255 struct obd_info *oinfo = cookie;
256 struct lov_request *lovreq;
257 lovreq = container_of(oinfo, struct lov_request, rq_oi);
258 return lov_update_common_set(lovreq->rq_rqset, lovreq, rc);
261 int lov_prep_getattr_set(struct obd_export *exp, struct obd_info *oinfo,
262 struct lov_request_set **reqset)
264 struct lov_request_set *set;
265 struct lov_obd *lov = &exp->exp_obd->u.lov;
269 OBD_ALLOC(set, sizeof(*set));
277 for (i = 0; i < oinfo->oi_md->lsm_stripe_count; i++) {
278 struct lov_oinfo *loi;
279 struct lov_request *req;
281 loi = oinfo->oi_md->lsm_oinfo[i];
282 if (lov_oinfo_is_dummy(loi))
285 if (!lov_check_and_wait_active(lov, loi->loi_ost_idx)) {
286 CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
290 OBD_ALLOC(req, sizeof(*req));
292 GOTO(out_set, rc = -ENOMEM);
295 req->rq_idx = loi->loi_ost_idx;
297 OBDO_ALLOC(req->rq_oi.oi_oa);
298 if (req->rq_oi.oi_oa == NULL) {
299 OBD_FREE(req, sizeof(*req));
300 GOTO(out_set, rc = -ENOMEM);
302 memcpy(req->rq_oi.oi_oa, oinfo->oi_oa,
303 sizeof(*req->rq_oi.oi_oa));
304 req->rq_oi.oi_oa->o_oi = loi->loi_oi;
305 req->rq_oi.oi_cb_up = cb_getattr_update;
306 req->rq_oi.oi_capa = oinfo->oi_capa;
308 lov_set_add_req(req, set);
311 GOTO(out_set, rc = -EIO);
315 lov_fini_getattr_set(set);
319 #define LOV_U64_MAX ((__u64)~0ULL)
320 #define LOV_SUM_MAX(tot, add) \
322 if ((tot) + (add) < (tot)) \
323 (tot) = LOV_U64_MAX; \
328 int lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs,int success)
333 __u32 expected_stripes = lov_get_stripecnt(&obd->u.lov,
335 if (osfs->os_files != LOV_U64_MAX)
336 lov_do_div64(osfs->os_files, expected_stripes);
337 if (osfs->os_ffree != LOV_U64_MAX)
338 lov_do_div64(osfs->os_ffree, expected_stripes);
340 spin_lock(&obd->obd_osfs_lock);
341 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
342 obd->obd_osfs_age = cfs_time_current_64();
343 spin_unlock(&obd->obd_osfs_lock);
350 int lov_fini_statfs_set(struct lov_request_set *set)
358 if (atomic_read(&set->set_completes)) {
359 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
360 atomic_read(&set->set_success));
366 void lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
369 int shift = 0, quit = 0;
373 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
375 if (osfs->os_bsize != lov_sfs->os_bsize) {
376 /* assume all block sizes are always powers of 2 */
377 /* get the bits difference */
378 tmp = osfs->os_bsize | lov_sfs->os_bsize;
379 for (shift = 0; shift <= 64; ++shift) {
391 if (osfs->os_bsize < lov_sfs->os_bsize) {
392 osfs->os_bsize = lov_sfs->os_bsize;
394 osfs->os_bfree >>= shift;
395 osfs->os_bavail >>= shift;
396 osfs->os_blocks >>= shift;
397 } else if (shift != 0) {
398 lov_sfs->os_bfree >>= shift;
399 lov_sfs->os_bavail >>= shift;
400 lov_sfs->os_blocks >>= shift;
403 /* Sandia requested that df (and so, statfs) only
404 returned minimal available space on
405 a single OST, so people would be able to
406 write this much data guaranteed. */
407 if (osfs->os_bavail > lov_sfs->os_bavail) {
408 /* Presumably if new bavail is smaller,
409 new bfree is bigger as well */
410 osfs->os_bfree = lov_sfs->os_bfree;
411 osfs->os_bavail = lov_sfs->os_bavail;
414 osfs->os_bfree += lov_sfs->os_bfree;
415 osfs->os_bavail += lov_sfs->os_bavail;
417 osfs->os_blocks += lov_sfs->os_blocks;
418 /* XXX not sure about this one - depends on policy.
419 * - could be minimum if we always stripe on all OBDs
420 * (but that would be wrong for any other policy,
421 * if one of the OBDs has no more objects left)
422 * - could be sum if we stripe whole objects
423 * - could be average, just to give a nice number
425 * To give a "reasonable" (if not wholly accurate)
426 * number, we divide the total number of free objects
427 * by expected stripe count (watch out for overflow).
429 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
430 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
434 /* The callback for osc_statfs_async that finilizes a request info when a
435 * response is received. */
436 static int cb_statfs_update(void *cookie, int rc)
438 struct obd_info *oinfo = cookie;
439 struct lov_request *lovreq;
440 struct lov_request_set *set;
441 struct obd_statfs *osfs, *lov_sfs;
443 struct lov_tgt_desc *tgt;
444 struct obd_device *lovobd, *tgtobd;
448 lovreq = container_of(oinfo, struct lov_request, rq_oi);
449 set = lovreq->rq_rqset;
450 lovobd = set->set_obd;
451 lov = &lovobd->u.lov;
452 osfs = set->set_oi->oi_osfs;
453 lov_sfs = oinfo->oi_osfs;
454 success = atomic_read(&set->set_success);
455 /* XXX: the same is done in lov_update_common_set, however
456 lovset->set_exp is not initialized. */
457 lov_update_set(set, lovreq, rc);
462 tgt = lov->lov_tgts[lovreq->rq_idx];
463 if (!tgt || !tgt->ltd_active)
464 GOTO(out_update, rc);
466 tgtobd = class_exp2obd(tgt->ltd_exp);
467 spin_lock(&tgtobd->obd_osfs_lock);
468 memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
469 if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
470 tgtobd->obd_osfs_age = cfs_time_current_64();
471 spin_unlock(&tgtobd->obd_osfs_lock);
474 lov_update_statfs(osfs, lov_sfs, success);
478 if (set->set_oi->oi_flags & OBD_STATFS_PTLRPCD &&
479 lov_set_finished(set, 0)) {
480 lov_statfs_interpret(NULL, set, set->set_count !=
481 atomic_read(&set->set_success));
487 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
488 struct lov_request_set **reqset)
490 struct lov_request_set *set;
491 struct lov_obd *lov = &obd->u.lov;
495 OBD_ALLOC(set, sizeof(*set));
503 /* We only get block data from the OBD */
504 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
505 struct lov_request *req;
507 if (lov->lov_tgts[i] == NULL ||
508 (oinfo->oi_flags & OBD_STATFS_NODELAY &&
509 !lov->lov_tgts[i]->ltd_active)) {
510 CDEBUG(D_HA, "lov idx %d inactive\n", i);
514 if (!lov->lov_tgts[i]->ltd_active)
515 lov_check_and_wait_active(lov, i);
517 /* skip targets that have been explicitely disabled by the
519 if (!lov->lov_tgts[i]->ltd_exp) {
520 CDEBUG(D_HA, "lov idx %d administratively disabled\n", i);
524 OBD_ALLOC(req, sizeof(*req));
526 GOTO(out_set, rc = -ENOMEM);
528 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
529 if (req->rq_oi.oi_osfs == NULL) {
530 OBD_FREE(req, sizeof(*req));
531 GOTO(out_set, rc = -ENOMEM);
535 req->rq_oi.oi_cb_up = cb_statfs_update;
536 req->rq_oi.oi_flags = oinfo->oi_flags;
538 lov_set_add_req(req, set);
541 GOTO(out_set, rc = -EIO);
545 lov_fini_statfs_set(set);