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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_LOV
35 #include <linux/delay.h>
36 #include <libcfs/libcfs.h>
38 #include <obd_class.h>
39 #include "lov_internal.h"
41 static void lov_init_set(struct lov_request_set *set)
44 atomic_set(&set->set_completes, 0);
45 atomic_set(&set->set_success, 0);
46 INIT_LIST_HEAD(&set->set_list);
49 static void lov_finish_set(struct lov_request_set *set)
51 struct list_head *pos, *n;
52 struct lov_request *req;
57 list_for_each_safe(pos, n, &set->set_list) {
58 req = list_entry(pos, struct lov_request, rq_link);
59 list_del_init(&req->rq_link);
61 if (req->rq_oi.oi_osfs)
62 OBD_FREE_PTR(req->rq_oi.oi_osfs);
72 lov_update_set(struct lov_request_set *set, struct lov_request *req, int rc)
74 atomic_inc(&set->set_completes);
76 atomic_inc(&set->set_success);
80 lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
82 list_add_tail(&req->rq_link, &set->set_list);
87 static int lov_check_set(struct lov_obd *lov, int idx)
91 mutex_lock(&lov->lov_lock);
93 if (!lov->lov_tgts[idx] || lov->lov_tgts[idx]->ltd_active ||
94 (lov->lov_tgts[idx]->ltd_exp &&
95 class_exp2cliimp(lov->lov_tgts[idx]->ltd_exp)->imp_connect_tried))
98 mutex_unlock(&lov->lov_lock);
103 * Check if the OSC connection exists and is active.
104 * If the OSC has not yet had a chance to connect to the OST the first time,
105 * wait once for it to connect instead of returning an error.
107 static int lov_check_and_wait_active(struct lov_obd *lov, int ost_idx)
109 struct lov_tgt_desc *tgt;
110 struct obd_import *imp = NULL;
114 mutex_lock(&lov->lov_lock);
116 tgt = lov->lov_tgts[ost_idx];
121 if (likely(tgt->ltd_active))
125 imp = class_exp2cliimp(tgt->ltd_exp);
126 if (imp && imp->imp_connect_tried)
128 if (imp && imp->imp_state == LUSTRE_IMP_IDLE)
131 mutex_unlock(&lov->lov_lock);
135 !lov_check_set(lov, ost_idx)) {
145 mutex_unlock(&lov->lov_lock);
149 #define LOV_U64_MAX ((__u64)~0ULL)
150 #define LOV_SUM_MAX(tot, add) \
152 if ((tot) + (add) < (tot)) \
153 (tot) = LOV_U64_MAX; \
159 lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs, int success)
164 __u32 expected_stripes = lov_get_stripe_count(&obd->u.lov,
166 if (osfs->os_files != LOV_U64_MAX)
167 lov_do_div64(osfs->os_files, expected_stripes);
168 if (osfs->os_ffree != LOV_U64_MAX)
169 lov_do_div64(osfs->os_ffree, expected_stripes);
171 spin_lock(&obd->obd_osfs_lock);
172 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
173 obd->obd_osfs_age = ktime_get_seconds();
174 spin_unlock(&obd->obd_osfs_lock);
181 int lov_fini_statfs_set(struct lov_request_set *set)
189 if (atomic_read(&set->set_completes)) {
190 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
191 atomic_read(&set->set_success));
200 lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
203 int shift = 0, quit = 0;
207 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
209 if (osfs->os_bsize != lov_sfs->os_bsize) {
210 /* assume all block sizes are always powers of 2 */
211 /* get the bits difference */
212 tmp = osfs->os_bsize | lov_sfs->os_bsize;
213 for (shift = 0; shift <= 64; ++shift) {
224 if (osfs->os_bsize < lov_sfs->os_bsize) {
225 osfs->os_bsize = lov_sfs->os_bsize;
227 osfs->os_bfree >>= shift;
228 osfs->os_bavail >>= shift;
229 osfs->os_blocks >>= shift;
230 } else if (shift != 0) {
231 lov_sfs->os_bfree >>= shift;
232 lov_sfs->os_bavail >>= shift;
233 lov_sfs->os_blocks >>= shift;
237 * Sandia requested that df (and so, statfs) only
238 * returned minimal available space on
239 * a single OST, so people would be able to
240 * write this much data guaranteed.
242 if (osfs->os_bavail > lov_sfs->os_bavail) {
244 * Presumably if new bavail is smaller,
245 * new bfree is bigger as well
247 osfs->os_bfree = lov_sfs->os_bfree;
248 osfs->os_bavail = lov_sfs->os_bavail;
251 osfs->os_bfree += lov_sfs->os_bfree;
252 osfs->os_bavail += lov_sfs->os_bavail;
254 osfs->os_blocks += lov_sfs->os_blocks;
256 * XXX not sure about this one - depends on policy.
257 * - could be minimum if we always stripe on all OBDs
258 * (but that would be wrong for any other policy,
259 * if one of the OBDs has no more objects left)
260 * - could be sum if we stripe whole objects
261 * - could be average, just to give a nice number
263 * To give a "reasonable" (if not wholly accurate)
264 * number, we divide the total number of free objects
265 * by expected stripe count (watch out for overflow).
267 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
268 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
273 * The callback for osc_statfs_async that finilizes a request info when a
274 * response is received.
276 static int cb_statfs_update(void *cookie, int rc)
278 struct obd_info *oinfo = cookie;
279 struct lov_request *lovreq;
280 struct lov_request_set *set;
281 struct obd_statfs *osfs, *lov_sfs;
283 struct lov_tgt_desc *tgt;
284 struct obd_device *lovobd, *tgtobd;
289 lovreq = container_of(oinfo, struct lov_request, rq_oi);
290 set = lovreq->rq_rqset;
291 lovobd = set->set_obd;
292 lov = &lovobd->u.lov;
293 osfs = set->set_oi->oi_osfs;
294 lov_sfs = oinfo->oi_osfs;
295 success = atomic_read(&set->set_success);
297 * XXX: the same is done in lov_update_common_set, however
298 * lovset->set_exp is not initialized.
300 lov_update_set(set, lovreq, rc);
304 lov_tgts_getref(lovobd);
305 tgt = lov->lov_tgts[lovreq->rq_idx];
306 if (!tgt || !tgt->ltd_active)
307 GOTO(out_update, rc);
309 tgtobd = class_exp2obd(tgt->ltd_exp);
310 spin_lock(&tgtobd->obd_osfs_lock);
311 memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
312 if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
313 tgtobd->obd_osfs_age = ktime_get_seconds();
314 spin_unlock(&tgtobd->obd_osfs_lock);
317 lov_update_statfs(osfs, lov_sfs, success);
318 lov_tgts_putref(lovobd);
323 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
324 struct lov_request_set **reqset)
326 struct lov_request_set *set;
327 struct lov_obd *lov = &obd->u.lov;
332 OBD_ALLOC(set, sizeof(*set));
340 /* We only get block data from the OBD */
341 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
342 struct lov_tgt_desc *ltd = lov->lov_tgts[i];
343 struct lov_request *req;
346 CDEBUG(D_HA, "lov idx %d inactive\n", i);
351 * skip targets that have been explicitely disabled by the
355 CDEBUG(D_HA, "lov idx %d administratively disabled\n",
360 if (oinfo->oi_flags & OBD_STATFS_NODELAY &&
361 class_exp2cliimp(ltd->ltd_exp)->imp_state !=
362 LUSTRE_IMP_IDLE && !ltd->ltd_active) {
363 CDEBUG(D_HA, "lov idx %d inactive\n", i);
367 if (!ltd->ltd_active)
368 lov_check_and_wait_active(lov, i);
370 OBD_ALLOC(req, sizeof(*req));
372 GOTO(out_set, rc = -ENOMEM);
374 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
375 if (!req->rq_oi.oi_osfs) {
376 OBD_FREE(req, sizeof(*req));
377 GOTO(out_set, rc = -ENOMEM);
381 req->rq_oi.oi_cb_up = cb_statfs_update;
382 req->rq_oi.oi_flags = oinfo->oi_flags;
384 lov_set_add_req(req, set);
387 GOTO(out_set, rc = -EIO);
391 lov_fini_statfs_set(set);