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 * This file is part of Lustre, http://www.lustre.org/
25 * lustre/obdclass/lu_tgt_descs.c
27 * Lustre target descriptions
28 * These are the only exported functions, they provide some generic
29 * infrastructure for target description management used by LOD/LMV
33 #define DEBUG_SUBSYSTEM S_LOV
35 #include <linux/module.h>
36 #include <linux/list.h>
37 #include <linux/random.h>
38 #include <libcfs/libcfs.h>
39 #include <libcfs/linux/linux-mem.h>
40 #include <obd_class.h>
41 #include <obd_support.h>
42 #include <lustre_disk.h>
43 #include <lustre_fid.h>
44 #include <lu_object.h>
47 * lu_prandom_u64_max - returns a pseudo-random u64 number in interval
50 * \param[in] ep_ro right open interval endpoint
52 * \retval a pseudo-random 64-bit number that is in interval [0, ep_ro).
54 u64 lu_prandom_u64_max(u64 ep_ro)
59 #ifdef HAVE_GET_RANDOM_U32_AND_U64
60 rand = get_random_u64() % ep_ro;
61 #elif BITS_PER_LONG == 32
63 * If ep_ro > 32-bit, first generate the high
64 * 32 bits of the random number, then add in the low
65 * 32 bits (truncated to the upper limit, if needed)
67 if (ep_ro > 0xffffffffULL)
68 rand = (u64)get_random_u32_below((u32)(ep_ro >> 32)) << 32;
70 if (rand == (ep_ro & 0xffffffff00000000ULL))
71 rand |= get_random_u32_below((u32)ep_ro);
73 rand |= get_random_u32();
75 rand = ((u64)get_random_u32() << 32 | get_random_u32()) % ep_ro;
81 EXPORT_SYMBOL(lu_prandom_u64_max);
84 * Add a new target to Quality of Service (QoS) target table.
86 * Add a new MDT/OST target to the structure representing an OSS. Resort the
87 * list of known MDSs/OSSs by the number of MDTs/OSTs attached to each MDS/OSS.
88 * The MDS/OSS list is protected internally and no external locking is required.
90 * \param[in] qos lu_qos data
91 * \param[in] tgt target description
93 * \retval 0 on success
94 * \retval -ENOMEM on error
96 int lu_qos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *tgt)
98 struct lu_svr_qos *svr = NULL;
99 struct lu_svr_qos *tempsvr;
100 struct obd_export *exp = tgt->ltd_exp;
107 down_write(&qos->lq_rw_sem);
109 * a bit hacky approach to learn NID of corresponding connection
110 * but there is no official API to access information like this
113 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
114 if (obd_uuid_equals(&svr->lsq_uuid,
115 &exp->exp_connection->c_remote_uuid)) {
119 if (svr->lsq_id > id)
126 GOTO(out, rc = -ENOMEM);
127 memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid,
128 sizeof(svr->lsq_uuid));
132 /* Assume we have to move this one */
133 list_del(&svr->lsq_svr_list);
136 svr->lsq_tgt_count++;
137 tgt->ltd_qos.ltq_svr = svr;
139 CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n",
140 obd_uuid2str(&tgt->ltd_uuid), obd_uuid2str(&svr->lsq_uuid),
144 * Add sorted by # of tgts. Find the first entry that we're
147 list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) {
148 if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count)
152 * ...and add before it. If we're the first or smallest, tempsvr
153 * points to the list head, and we add to the end.
155 list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list);
157 set_bit(LQ_DIRTY, &qos->lq_flags);
158 #ifdef HAVE_SERVER_SUPPORT
159 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
162 up_write(&qos->lq_rw_sem);
165 EXPORT_SYMBOL(lu_qos_add_tgt);
168 * Remove MDT/OST target from QoS table.
170 * Removes given MDT/OST target from QoS table and releases related
171 * MDS/OSS structure if no target remain on the MDS/OSS.
173 * \param[in] qos lu_qos data
174 * \param[in] ltd target description
176 * \retval 0 on success
177 * \retval -ENOENT if no server was found
179 int lu_qos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd)
181 struct lu_svr_qos *svr;
186 down_write(&qos->lq_rw_sem);
187 svr = ltd->ltd_qos.ltq_svr;
189 GOTO(out, rc = -ENOENT);
191 ltd->ltd_qos.ltq_svr = NULL;
192 svr->lsq_tgt_count--;
193 if (svr->lsq_tgt_count == 0) {
194 CDEBUG(D_OTHER, "removing server %s\n",
195 obd_uuid2str(&svr->lsq_uuid));
196 list_del(&svr->lsq_svr_list);
200 set_bit(LQ_DIRTY, &qos->lq_flags);
201 #ifdef HAVE_SERVER_SUPPORT
202 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
205 up_write(&qos->lq_rw_sem);
208 EXPORT_SYMBOL(lu_qos_del_tgt);
211 * Calculate weight for a given tgt.
213 * The final tgt weight uses only free space for OSTs, but combines
214 * both free space and inodes for MDTs, minus tgt and server penalties.
215 * See ltd_qos_penalties_calc() for how penalties are calculated.
217 * \param[in] tgt target descriptor
218 * \param[in] is_mdt target table is for MDT selection (use inodes)
220 void lu_tgt_qos_weight_calc(struct lu_tgt_desc *tgt, bool is_mdt)
222 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
226 ltq->ltq_avail = (tgt_statfs_bavail(tgt) >> 16) *
227 (tgt_statfs_iavail(tgt) >> 8);
229 ltq->ltq_avail = tgt_statfs_bavail(tgt) >> 8;
230 penalty = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
231 CDEBUG(D_OTHER, "ltq_penalty: %llu lsq_penalty: %llu tgt_bavail: %llu\n",
232 ltq->ltq_penalty, ltq->ltq_svr->lsq_penalty, ltq->ltq_avail);
233 if (ltq->ltq_avail < penalty)
236 ltq->ltq_weight = ltq->ltq_avail - penalty;
238 EXPORT_SYMBOL(lu_tgt_qos_weight_calc);
241 * Allocate and initialize target table.
243 * A helper function to initialize the target table and allocate
244 * a bitmap of the available targets.
246 * \param[in] ltd target's table to initialize
247 * \param[in] is_mdt target table for MDTs
249 * \retval 0 on success
250 * \retval negative negated errno on error
252 int lu_tgt_descs_init(struct lu_tgt_descs *ltd, bool is_mdt)
254 mutex_init(<d->ltd_mutex);
255 init_rwsem(<d->ltd_rw_sem);
258 * the tgt array and bitmap are allocated/grown dynamically as tgts are
259 * added to the LOD/LMV, see lu_tgt_descs_add()
261 ltd->ltd_tgt_bitmap = bitmap_zalloc(BITS_PER_LONG, GFP_NOFS);
262 if (!ltd->ltd_tgt_bitmap)
265 ltd->ltd_tgts_size = BITS_PER_LONG;
266 ltd->ltd_death_row = 0;
267 ltd->ltd_refcount = 0;
269 /* Set up allocation policy (QoS and RR) */
270 INIT_LIST_HEAD(<d->ltd_qos.lq_svr_list);
271 init_rwsem(<d->ltd_qos.lq_rw_sem);
272 set_bit(LQ_DIRTY, <d->ltd_qos.lq_flags);
273 set_bit(LQ_RESET, <d->ltd_qos.lq_flags);
274 ltd->ltd_is_mdt = is_mdt;
275 /* MDT imbalance threshold is low to balance across MDTs
276 * relatively quickly, because each directory may result
277 * in a large number of files/subdirs created therein.
280 ltd->ltd_lmv_desc.ld_pattern = LMV_HASH_TYPE_DEFAULT;
281 ltd->ltd_qos.lq_prio_free = LMV_QOS_DEF_PRIO_FREE * 256 / 100;
282 ltd->ltd_qos.lq_threshold_rr =
283 LMV_QOS_DEF_THRESHOLD_RR_PCT *
284 QOS_THRESHOLD_MAX / 100;
286 ltd->ltd_qos.lq_prio_free = LOV_QOS_DEF_PRIO_FREE * 256 / 100;
287 ltd->ltd_qos.lq_threshold_rr =
288 LOV_QOS_DEF_THRESHOLD_RR_PCT *
289 QOS_THRESHOLD_MAX / 100;
294 EXPORT_SYMBOL(lu_tgt_descs_init);
297 * Free bitmap and target table pages.
299 * \param[in] ltd target table
301 void lu_tgt_descs_fini(struct lu_tgt_descs *ltd)
305 bitmap_free(ltd->ltd_tgt_bitmap);
306 for (i = 0; i < ARRAY_SIZE(ltd->ltd_tgt_idx); i++) {
307 if (ltd->ltd_tgt_idx[i])
308 OBD_FREE_PTR(ltd->ltd_tgt_idx[i]);
310 ltd->ltd_tgts_size = 0;
312 EXPORT_SYMBOL(lu_tgt_descs_fini);
315 * Expand size of target table.
317 * When the target table is full, we have to extend the table. To do so,
318 * we allocate new memory with some reserve, move data from the old table
319 * to the new one and release memory consumed by the old table.
321 * \param[in] ltd target table
322 * \param[in] newsize new size of the table
324 * \retval 0 on success
325 * \retval -ENOMEM if reallocation failed
327 static int lu_tgt_descs_resize(struct lu_tgt_descs *ltd, __u32 newsize)
329 unsigned long *new_bitmap, *old_bitmap = NULL;
331 /* someone else has already resize the array */
332 if (newsize <= ltd->ltd_tgts_size)
335 new_bitmap = bitmap_zalloc(newsize, GFP_NOFS);
339 if (ltd->ltd_tgts_size > 0) {
340 /* the bitmap already exists, copy data from old one */
341 bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap,
343 old_bitmap = ltd->ltd_tgt_bitmap;
346 ltd->ltd_tgts_size = newsize;
347 ltd->ltd_tgt_bitmap = new_bitmap;
349 bitmap_free(old_bitmap);
351 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
357 * Add new target to target table.
359 * Extend target table if it's full, update target table and bitmap.
360 * Notice we need to take ltd_rw_sem exclusively before entry to ensure
363 * \param[in] ltd target table
364 * \param[in] tgt new target desc
366 * \retval 0 on success
367 * \retval -ENOMEM if reallocation failed
368 * -EEXIST if target existed
370 int ltd_add_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
372 __u32 index = tgt->ltd_index;
377 if (index >= ltd->ltd_tgts_size) {
380 if (index > TGT_PTRS * TGT_PTRS_PER_BLOCK)
383 while (newsize < index + 1)
384 newsize = newsize << 1;
386 rc = lu_tgt_descs_resize(ltd, newsize);
389 } else if (test_bit(index, ltd->ltd_tgt_bitmap)) {
393 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
394 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
395 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
399 LTD_TGT(ltd, tgt->ltd_index) = tgt;
400 set_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
402 ltd->ltd_lov_desc.ld_tgt_count++;
404 ltd->ltd_lov_desc.ld_active_tgt_count++;
408 EXPORT_SYMBOL(ltd_add_tgt);
411 * Delete target from target table
413 void ltd_del_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
415 lu_qos_del_tgt(<d->ltd_qos, tgt);
416 LTD_TGT(ltd, tgt->ltd_index) = NULL;
417 clear_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
418 ltd->ltd_lov_desc.ld_tgt_count--;
420 ltd->ltd_lov_desc.ld_active_tgt_count--;
422 EXPORT_SYMBOL(ltd_del_tgt);
425 * Calculate penalties per-tgt and per-server
427 * Re-calculate penalties when the configuration changes, active targets
428 * change and after statfs refresh (all these are reflected by LQ_DIRTY flag).
429 * On every tgt and server: decay the penalty by half for every 8x the update
430 * interval that the device has been idle. That gives lots of time for the
431 * statfs information to be updated (which the penalty is only a proxy for),
432 * and avoids penalizing server/tgt under light load.
433 * See lu_qos_tgt_weight_calc() for how penalties are factored into the weight.
435 * \param[in] ltd lu_tgt_descs
437 * \retval 0 on success
438 * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
441 int ltd_qos_penalties_calc(struct lu_tgt_descs *ltd)
443 struct lu_qos *qos = <d->ltd_qos;
444 struct lov_desc *desc = <d->ltd_lov_desc;
445 struct lu_tgt_desc *tgt;
446 struct lu_svr_qos *svr;
447 __u64 ba_max, ba_min, ba;
448 __u64 ia_max, ia_min, ia = 1;
456 if (!test_bit(LQ_DIRTY, &qos->lq_flags))
459 num_active = desc->ld_active_tgt_count - 1;
461 GOTO(out, rc = -EAGAIN);
463 /* find bavail on each server */
464 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
466 /* if inode is not counted, set to 1 to ignore */
467 svr->lsq_iavail = ltd->ltd_is_mdt ? 0 : 1;
469 qos->lq_active_svr_count = 0;
472 * How badly user wants to select targets "widely" (not recently chosen
473 * and not on recent MDS's). As opposed to "freely" (free space avail.)
476 prio_wide = 256 - qos->lq_prio_free;
478 ba_min = (__u64)(-1);
480 ia_min = (__u64)(-1);
482 now = ktime_get_real_seconds();
484 /* Calculate server penalty per object */
485 ltd_foreach_tgt(ltd, tgt) {
486 if (!tgt->ltd_active)
489 /* when inode is counted, bavail >> 16 to avoid overflow */
490 ba = tgt_statfs_bavail(tgt);
498 ba_min = min(ba, ba_min);
499 ba_max = max(ba, ba_max);
501 /* Count the number of usable servers */
502 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
503 qos->lq_active_svr_count++;
504 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
506 if (ltd->ltd_is_mdt) {
507 /* iavail >> 8 to avoid overflow */
508 ia = tgt_statfs_iavail(tgt) >> 8;
512 ia_min = min(ia, ia_min);
513 ia_max = max(ia, ia_max);
515 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
520 * prio * bavail * iavail / (num_tgt - 1) / prio_max / 2
522 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia >> 9;
523 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
525 age = (now - tgt->ltd_qos.ltq_used) >> 3;
526 if (test_bit(LQ_RESET, &qos->lq_flags) ||
527 age > 32 * desc->ld_qos_maxage)
528 tgt->ltd_qos.ltq_penalty = 0;
529 else if (age > desc->ld_qos_maxage)
530 /* Decay tgt penalty. */
531 tgt->ltd_qos.ltq_penalty >>= age / desc->ld_qos_maxage;
534 num_active = qos->lq_active_svr_count - 1;
535 if (num_active < 1) {
537 * If there's only 1 server, we can't penalize it, so instead
538 * we have to double the tgt penalty
541 ltd_foreach_tgt(ltd, tgt) {
542 if (!tgt->ltd_active)
545 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
550 * Per-server penalty is
551 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
553 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
554 ba = svr->lsq_bavail;
555 ia = svr->lsq_iavail;
556 svr->lsq_penalty_per_obj = prio_wide * ba * ia >> 8;
557 do_div(svr->lsq_penalty_per_obj,
558 svr->lsq_tgt_count * num_active);
559 svr->lsq_penalty_per_obj >>= 1;
561 age = (now - svr->lsq_used) >> 3;
562 if (test_bit(LQ_RESET, &qos->lq_flags) ||
563 age > 32 * desc->ld_qos_maxage)
564 svr->lsq_penalty = 0;
565 else if (age > desc->ld_qos_maxage)
566 /* Decay server penalty. */
567 svr->lsq_penalty >>= age / desc->ld_qos_maxage;
572 * If each tgt has almost same free space, do rr allocation for better
573 * creation performance
575 if (((ba_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
576 QOS_THRESHOLD_MAX) < ba_min &&
577 ((ia_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
578 QOS_THRESHOLD_MAX) < ia_min) {
579 set_bit(LQ_SAME_SPACE, &qos->lq_flags);
580 /* Reset weights for the next time we enter qos mode */
581 set_bit(LQ_RESET, &qos->lq_flags);
583 clear_bit(LQ_SAME_SPACE, &qos->lq_flags);
584 clear_bit(LQ_RESET, &qos->lq_flags);
586 clear_bit(LQ_DIRTY, &qos->lq_flags);
590 if (!rc && test_bit(LQ_SAME_SPACE, &qos->lq_flags))
595 EXPORT_SYMBOL(ltd_qos_penalties_calc);
598 * Re-calculate penalties and weights of all tgts.
600 * The function is called when some target was used for a new object. In
601 * this case we should re-calculate all the weights to keep new allocations
604 * \param[in] ltd lu_tgt_descs
605 * \param[in] tgt recently used tgt
606 * \param[out] total_wt new total weight for the pool
610 int ltd_qos_update(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt,
613 struct lu_qos *qos = <d->ltd_qos;
614 struct lu_tgt_qos *ltq;
615 struct lu_svr_qos *svr;
622 /* Don't allocate on this device anymore, until the next alloc_qos */
628 * Decay old penalty by half (we're adding max penalty, and don't
629 * want it to run away.)
631 ltq->ltq_penalty >>= 1;
632 svr->lsq_penalty >>= 1;
634 /* mark the server and tgt as recently used */
635 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
637 /* Set max penalties for this tgt and server */
638 ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
639 ltd->ltd_lov_desc.ld_active_tgt_count;
640 CDEBUG(D_OTHER, "ltq_penalty: %llu per_obj: %llu tgt_count: %d\n",
641 ltq->ltq_penalty, ltq->ltq_penalty_per_obj,
642 ltd->ltd_lov_desc.ld_active_tgt_count);
643 svr->lsq_penalty += svr->lsq_penalty_per_obj *
644 qos->lq_active_svr_count;
645 CDEBUG(D_OTHER, "lsq_penalty: %llu per_obj: %llu srv_count: %d\n",
646 svr->lsq_penalty, svr->lsq_penalty_per_obj,
647 qos->lq_active_svr_count);
650 /* Decrease all MDS penalties */
651 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
652 if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
653 svr->lsq_penalty = 0;
655 svr->lsq_penalty -= svr->lsq_penalty_per_obj;
659 /* Decrease all tgt penalties */
660 ltd_foreach_tgt(ltd, tgt) {
661 if (!tgt->ltd_active)
665 if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
666 ltq->ltq_penalty = 0;
668 ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
670 lu_tgt_qos_weight_calc(tgt, ltd->ltd_is_mdt);
672 /* Recalc the total weight of usable osts */
674 *total_wt += ltq->ltq_weight;
676 CDEBUG(D_OTHER, "recalc tgt %d usable=%d bavail=%llu ffree=%llu tgtppo=%llu tgtp=%llu svrppo=%llu svrp=%llu wt=%llu\n",
677 tgt->ltd_index, ltq->ltq_usable,
678 tgt_statfs_bavail(tgt) >> 16,
679 tgt_statfs_iavail(tgt) >> 8,
680 ltq->ltq_penalty_per_obj >> 10,
681 ltq->ltq_penalty >> 10,
682 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
683 ltq->ltq_svr->lsq_penalty >> 10,
684 ltq->ltq_weight >> 10);
689 EXPORT_SYMBOL(ltd_qos_update);