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/
32 #define DEBUG_SUBSYSTEM S_LOV
34 #include <linux/delay.h>
35 #include <libcfs/libcfs.h>
37 #include <obd_class.h>
38 #include "lov_internal.h"
40 static void lov_init_set(struct lov_request_set *set)
43 atomic_set(&set->set_completes, 0);
44 atomic_set(&set->set_success, 0);
45 INIT_LIST_HEAD(&set->set_list);
48 static void lov_finish_set(struct lov_request_set *set)
50 struct list_head *pos, *n;
51 struct lov_request *req;
56 list_for_each_safe(pos, n, &set->set_list) {
57 req = list_entry(pos, struct lov_request, rq_link);
58 list_del_init(&req->rq_link);
60 if (req->rq_oi.oi_osfs)
61 OBD_FREE_PTR(req->rq_oi.oi_osfs);
71 lov_update_set(struct lov_request_set *set, struct lov_request *req, int rc)
73 atomic_inc(&set->set_completes);
75 atomic_inc(&set->set_success);
79 lov_set_add_req(struct lov_request *req, struct lov_request_set *set)
81 list_add_tail(&req->rq_link, &set->set_list);
86 static int lov_check_set(struct lov_obd *lov, int idx)
90 mutex_lock(&lov->lov_lock);
92 if (!lov->lov_tgts[idx] || lov->lov_tgts[idx]->ltd_active ||
93 (lov->lov_tgts[idx]->ltd_exp &&
94 class_exp2cliimp(lov->lov_tgts[idx]->ltd_exp)->imp_connect_tried))
97 mutex_unlock(&lov->lov_lock);
102 * Check if the OSC connection exists and is active.
103 * If the OSC has not yet had a chance to connect to the OST the first time,
104 * wait once for it to connect instead of returning an error.
106 static int lov_check_and_wait_active(struct lov_obd *lov, int ost_idx)
108 struct lov_tgt_desc *tgt;
109 struct obd_import *imp = NULL;
113 mutex_lock(&lov->lov_lock);
115 tgt = lov->lov_tgts[ost_idx];
120 if (likely(tgt->ltd_active))
124 imp = class_exp2cliimp(tgt->ltd_exp);
125 if (imp && imp->imp_connect_tried)
127 if (imp && imp->imp_state == LUSTRE_IMP_IDLE)
130 mutex_unlock(&lov->lov_lock);
134 !lov_check_set(lov, ost_idx)) {
144 mutex_unlock(&lov->lov_lock);
148 #define LOV_U64_MAX ((__u64)~0ULL)
149 #define LOV_SUM_MAX(tot, add) \
151 if ((tot) + (add) < (tot)) \
152 (tot) = LOV_U64_MAX; \
158 lov_fini_statfs(struct obd_device *obd, struct obd_statfs *osfs, int success)
163 __u32 expected_stripes = lov_get_stripe_count(&obd->u.lov,
165 if (osfs->os_files != LOV_U64_MAX)
166 lov_do_div64(osfs->os_files, expected_stripes);
167 if (osfs->os_ffree != LOV_U64_MAX)
168 lov_do_div64(osfs->os_ffree, expected_stripes);
170 spin_lock(&obd->obd_osfs_lock);
171 memcpy(&obd->obd_osfs, osfs, sizeof(*osfs));
172 obd->obd_osfs_age = ktime_get_seconds();
173 spin_unlock(&obd->obd_osfs_lock);
180 int lov_fini_statfs_set(struct lov_request_set *set)
188 if (atomic_read(&set->set_completes)) {
189 rc = lov_fini_statfs(set->set_obd, set->set_oi->oi_osfs,
190 atomic_read(&set->set_success));
199 lov_update_statfs(struct obd_statfs *osfs, struct obd_statfs *lov_sfs,
202 int shift = 0, quit = 0;
206 memcpy(osfs, lov_sfs, sizeof(*lov_sfs));
208 if (osfs->os_bsize != lov_sfs->os_bsize) {
209 /* assume all block sizes are always powers of 2 */
210 /* get the bits difference */
211 tmp = osfs->os_bsize | lov_sfs->os_bsize;
212 for (shift = 0; shift <= 64; ++shift) {
223 if (osfs->os_bsize < lov_sfs->os_bsize) {
224 osfs->os_bsize = lov_sfs->os_bsize;
226 osfs->os_bfree >>= shift;
227 osfs->os_bavail >>= shift;
228 osfs->os_blocks >>= shift;
229 } else if (shift != 0) {
230 lov_sfs->os_bfree >>= shift;
231 lov_sfs->os_bavail >>= shift;
232 lov_sfs->os_blocks >>= shift;
236 * Sandia requested that df (and so, statfs) only
237 * returned minimal available space on
238 * a single OST, so people would be able to
239 * write this much data guaranteed.
241 if (osfs->os_bavail > lov_sfs->os_bavail) {
243 * Presumably if new bavail is smaller,
244 * new bfree is bigger as well
246 osfs->os_bfree = lov_sfs->os_bfree;
247 osfs->os_bavail = lov_sfs->os_bavail;
250 osfs->os_bfree += lov_sfs->os_bfree;
251 osfs->os_bavail += lov_sfs->os_bavail;
253 osfs->os_blocks += lov_sfs->os_blocks;
255 * XXX not sure about this one - depends on policy.
256 * - could be minimum if we always stripe on all OBDs
257 * (but that would be wrong for any other policy,
258 * if one of the OBDs has no more objects left)
259 * - could be sum if we stripe whole objects
260 * - could be average, just to give a nice number
262 * To give a "reasonable" (if not wholly accurate)
263 * number, we divide the total number of free objects
264 * by expected stripe count (watch out for overflow).
266 LOV_SUM_MAX(osfs->os_files, lov_sfs->os_files);
267 LOV_SUM_MAX(osfs->os_ffree, lov_sfs->os_ffree);
272 * The callback for osc_statfs_async that finilizes a request info when a
273 * response is received.
275 static int cb_statfs_update(void *cookie, int rc)
277 struct obd_info *oinfo = cookie;
278 struct lov_request *lovreq;
279 struct lov_request_set *set;
280 struct obd_statfs *osfs, *lov_sfs;
282 struct lov_tgt_desc *tgt;
283 struct obd_device *lovobd, *tgtobd;
288 lovreq = container_of(oinfo, struct lov_request, rq_oi);
289 set = lovreq->rq_rqset;
290 lovobd = set->set_obd;
291 lov = &lovobd->u.lov;
292 osfs = set->set_oi->oi_osfs;
293 lov_sfs = oinfo->oi_osfs;
294 success = atomic_read(&set->set_success);
296 * XXX: the same is done in lov_update_common_set, however
297 * lovset->set_exp is not initialized.
299 lov_update_set(set, lovreq, rc);
303 lov_tgts_getref(lovobd);
304 tgt = lov->lov_tgts[lovreq->rq_idx];
305 if (!tgt || !tgt->ltd_active)
306 GOTO(out_update, rc);
308 tgtobd = class_exp2obd(tgt->ltd_exp);
309 spin_lock(&tgtobd->obd_osfs_lock);
310 memcpy(&tgtobd->obd_osfs, lov_sfs, sizeof(*lov_sfs));
311 if ((oinfo->oi_flags & OBD_STATFS_FROM_CACHE) == 0)
312 tgtobd->obd_osfs_age = ktime_get_seconds();
313 spin_unlock(&tgtobd->obd_osfs_lock);
316 lov_update_statfs(osfs, lov_sfs, success);
317 lov_tgts_putref(lovobd);
322 int lov_prep_statfs_set(struct obd_device *obd, struct obd_info *oinfo,
323 struct lov_request_set **reqset)
325 struct lov_request_set *set;
326 struct lov_obd *lov = &obd->u.lov;
331 OBD_ALLOC(set, sizeof(*set));
339 /* We only get block data from the OBD */
340 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
341 struct lov_tgt_desc *ltd = lov->lov_tgts[i];
342 struct lov_request *req;
345 CDEBUG(D_HA, "lov idx %d inactive\n", i);
350 * skip targets that have been explicitely disabled by the
354 CDEBUG(D_HA, "lov idx %d administratively disabled\n",
359 if (oinfo->oi_flags & OBD_STATFS_NODELAY &&
360 class_exp2cliimp(ltd->ltd_exp)->imp_state !=
361 LUSTRE_IMP_IDLE && !ltd->ltd_active) {
362 CDEBUG(D_HA, "lov idx %d inactive\n", i);
366 if (!ltd->ltd_active)
367 lov_check_and_wait_active(lov, i);
369 OBD_ALLOC(req, sizeof(*req));
371 GOTO(out_set, rc = -ENOMEM);
373 OBD_ALLOC(req->rq_oi.oi_osfs, sizeof(*req->rq_oi.oi_osfs));
374 if (!req->rq_oi.oi_osfs) {
375 OBD_FREE(req, sizeof(*req));
376 GOTO(out_set, rc = -ENOMEM);
380 req->rq_oi.oi_cb_up = cb_statfs_update;
381 req->rq_oi.oi_flags = oinfo->oi_flags;
383 lov_set_add_req(req, set);
386 GOTO(out_set, rc = -EIO);
390 lov_fini_statfs_set(set);