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_CLASS
35 #include <linux/module.h>
36 #include <linux/list.h>
37 #include <linux/random.h>
38 #include <libcfs/libcfs.h>
39 #include <libcfs/libcfs_hash.h> /* hash_long() */
40 #include <libcfs/linux/linux-mem.h>
41 #include <obd_class.h>
42 #include <obd_support.h>
43 #include <lustre_disk.h>
44 #include <lustre_fid.h>
45 #include <lu_object.h>
48 * lu_prandom_u64_max - returns a pseudo-random u64 number in interval
51 * \param[in] ep_ro right open interval endpoint
53 * \retval a pseudo-random 64-bit number that is in interval [0, ep_ro).
55 u64 lu_prandom_u64_max(u64 ep_ro)
60 #if BITS_PER_LONG == 32
62 * If ep_ro > 32-bit, first generate the high
63 * 32 bits of the random number, then add in the low
64 * 32 bits (truncated to the upper limit, if needed)
66 if (ep_ro > 0xffffffffULL)
67 rand = (u64)prandom_u32_max((u32)(ep_ro >> 32)) << 32;
69 if (rand == (ep_ro & 0xffffffff00000000ULL))
70 rand |= prandom_u32_max((u32)ep_ro);
72 rand |= prandom_u32();
74 rand = ((u64)prandom_u32() << 32 | prandom_u32()) % ep_ro;
80 EXPORT_SYMBOL(lu_prandom_u64_max);
83 * Add a new target to Quality of Service (QoS) target table.
85 * Add a new MDT/OST target to the structure representing an OSS. Resort the
86 * list of known MDSs/OSSs by the number of MDTs/OSTs attached to each MDS/OSS.
87 * The MDS/OSS list is protected internally and no external locking is required.
89 * \param[in] qos lu_qos data
90 * \param[in] tgt target description
92 * \retval 0 on success
93 * \retval -ENOMEM on error
95 int lu_qos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *tgt)
97 struct lu_svr_qos *svr = NULL;
98 struct lu_svr_qos *tempsvr;
99 struct obd_export *exp = tgt->ltd_exp;
106 down_write(&qos->lq_rw_sem);
108 * a bit hacky approach to learn NID of corresponding connection
109 * but there is no official API to access information like this
112 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
113 if (obd_uuid_equals(&svr->lsq_uuid,
114 &exp->exp_connection->c_remote_uuid)) {
118 if (svr->lsq_id > id)
125 GOTO(out, rc = -ENOMEM);
126 memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid,
127 sizeof(svr->lsq_uuid));
131 /* Assume we have to move this one */
132 list_del(&svr->lsq_svr_list);
135 svr->lsq_tgt_count++;
136 tgt->ltd_qos.ltq_svr = svr;
138 CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n",
139 obd_uuid2str(&tgt->ltd_uuid), obd_uuid2str(&svr->lsq_uuid),
143 * Add sorted by # of tgts. Find the first entry that we're
146 list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) {
147 if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count)
151 * ...and add before it. If we're the first or smallest, tempsvr
152 * points to the list head, and we add to the end.
154 list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list);
156 set_bit(LQ_DIRTY, &qos->lq_flags);
157 #ifdef HAVE_SERVER_SUPPORT
158 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
161 up_write(&qos->lq_rw_sem);
164 EXPORT_SYMBOL(lu_qos_add_tgt);
167 * Remove MDT/OST target from QoS table.
169 * Removes given MDT/OST target from QoS table and releases related
170 * MDS/OSS structure if no target remain on the MDS/OSS.
172 * \param[in] qos lu_qos data
173 * \param[in] ltd target description
175 * \retval 0 on success
176 * \retval -ENOENT if no server was found
178 static int lu_qos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd)
180 struct lu_svr_qos *svr;
185 down_write(&qos->lq_rw_sem);
186 svr = ltd->ltd_qos.ltq_svr;
188 GOTO(out, rc = -ENOENT);
190 svr->lsq_tgt_count--;
191 if (svr->lsq_tgt_count == 0) {
192 CDEBUG(D_OTHER, "removing server %s\n",
193 obd_uuid2str(&svr->lsq_uuid));
194 list_del(&svr->lsq_svr_list);
195 ltd->ltd_qos.ltq_svr = NULL;
199 set_bit(LQ_DIRTY, &qos->lq_flags);
200 #ifdef HAVE_SERVER_SUPPORT
201 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
204 up_write(&qos->lq_rw_sem);
208 static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
210 struct obd_statfs *statfs = &tgt->ltd_statfs;
212 return statfs->os_bavail * statfs->os_bsize;
215 static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
217 return tgt->ltd_statfs.os_ffree;
221 * Calculate weight for a given tgt.
223 * The final tgt weight is bavail >> 16 * iavail >> 8 minus the tgt and server
224 * penalties. See ltd_qos_penalties_calc() for how penalties are calculated.
226 * \param[in] tgt target descriptor
228 void lu_tgt_qos_weight_calc(struct lu_tgt_desc *tgt)
230 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
233 ltq->ltq_avail = (tgt_statfs_bavail(tgt) >> 16) *
234 (tgt_statfs_iavail(tgt) >> 8);
235 penalty = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
236 if (ltq->ltq_avail < penalty)
239 ltq->ltq_weight = ltq->ltq_avail - penalty;
241 EXPORT_SYMBOL(lu_tgt_qos_weight_calc);
244 * Allocate and initialize target table.
246 * A helper function to initialize the target table and allocate
247 * a bitmap of the available targets.
249 * \param[in] ltd target's table to initialize
250 * \param[in] is_mdt target table for MDTs
252 * \retval 0 on success
253 * \retval negative negated errno on error
255 int lu_tgt_descs_init(struct lu_tgt_descs *ltd, bool is_mdt)
257 mutex_init(<d->ltd_mutex);
258 init_rwsem(<d->ltd_rw_sem);
261 * the tgt array and bitmap are allocated/grown dynamically as tgts are
262 * added to the LOD/LMV, see lu_tgt_descs_add()
264 ltd->ltd_tgt_bitmap = bitmap_zalloc(BITS_PER_LONG, GFP_NOFS);
265 if (!ltd->ltd_tgt_bitmap)
268 ltd->ltd_tgts_size = BITS_PER_LONG;
269 ltd->ltd_death_row = 0;
270 ltd->ltd_refcount = 0;
272 /* Set up allocation policy (QoS and RR) */
273 INIT_LIST_HEAD(<d->ltd_qos.lq_svr_list);
274 init_rwsem(<d->ltd_qos.lq_rw_sem);
275 set_bit(LQ_DIRTY, <d->ltd_qos.lq_flags);
276 set_bit(LQ_RESET, <d->ltd_qos.lq_flags);
277 ltd->ltd_is_mdt = is_mdt;
278 /* MDT imbalance threshold is low to balance across MDTs
279 * relatively quickly, because each directory may result
280 * in a large number of files/subdirs created therein.
283 ltd->ltd_lmv_desc.ld_pattern = LMV_HASH_TYPE_DEFAULT;
284 ltd->ltd_qos.lq_prio_free = LMV_QOS_DEF_PRIO_FREE * 256 / 100;
285 ltd->ltd_qos.lq_threshold_rr =
286 LMV_QOS_DEF_THRESHOLD_RR_PCT *
287 QOS_THRESHOLD_MAX / 100;
289 ltd->ltd_qos.lq_prio_free = LOV_QOS_DEF_PRIO_FREE * 256 / 100;
290 ltd->ltd_qos.lq_threshold_rr =
291 LOV_QOS_DEF_THRESHOLD_RR_PCT *
292 QOS_THRESHOLD_MAX / 100;
297 EXPORT_SYMBOL(lu_tgt_descs_init);
300 * Free bitmap and target table pages.
302 * \param[in] ltd target table
304 void lu_tgt_descs_fini(struct lu_tgt_descs *ltd)
308 bitmap_free(ltd->ltd_tgt_bitmap);
309 for (i = 0; i < ARRAY_SIZE(ltd->ltd_tgt_idx); i++) {
310 if (ltd->ltd_tgt_idx[i])
311 OBD_FREE_PTR(ltd->ltd_tgt_idx[i]);
313 ltd->ltd_tgts_size = 0;
315 EXPORT_SYMBOL(lu_tgt_descs_fini);
318 * Expand size of target table.
320 * When the target table is full, we have to extend the table. To do so,
321 * we allocate new memory with some reserve, move data from the old table
322 * to the new one and release memory consumed by the old table.
324 * \param[in] ltd target table
325 * \param[in] newsize new size of the table
327 * \retval 0 on success
328 * \retval -ENOMEM if reallocation failed
330 static int lu_tgt_descs_resize(struct lu_tgt_descs *ltd, __u32 newsize)
332 unsigned long *new_bitmap, *old_bitmap = NULL;
334 /* someone else has already resize the array */
335 if (newsize <= ltd->ltd_tgts_size)
338 new_bitmap = bitmap_zalloc(newsize, GFP_NOFS);
342 if (ltd->ltd_tgts_size > 0) {
343 /* the bitmap already exists, copy data from old one */
344 bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap,
346 old_bitmap = ltd->ltd_tgt_bitmap;
349 ltd->ltd_tgts_size = newsize;
350 ltd->ltd_tgt_bitmap = new_bitmap;
352 bitmap_free(old_bitmap);
354 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
360 * Add new target to target table.
362 * Extend target table if it's full, update target table and bitmap.
363 * Notice we need to take ltd_rw_sem exclusively before entry to ensure
366 * \param[in] ltd target table
367 * \param[in] tgt new target desc
369 * \retval 0 on success
370 * \retval -ENOMEM if reallocation failed
371 * -EEXIST if target existed
373 int ltd_add_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
375 __u32 index = tgt->ltd_index;
380 if (index >= ltd->ltd_tgts_size) {
383 if (index > TGT_PTRS * TGT_PTRS_PER_BLOCK)
386 while (newsize < index + 1)
387 newsize = newsize << 1;
389 rc = lu_tgt_descs_resize(ltd, newsize);
392 } else if (test_bit(index, ltd->ltd_tgt_bitmap)) {
396 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
397 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
398 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
402 LTD_TGT(ltd, tgt->ltd_index) = tgt;
403 set_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
405 ltd->ltd_lov_desc.ld_tgt_count++;
407 ltd->ltd_lov_desc.ld_active_tgt_count++;
411 EXPORT_SYMBOL(ltd_add_tgt);
414 * Delete target from target table
416 void ltd_del_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
418 lu_qos_del_tgt(<d->ltd_qos, tgt);
419 LTD_TGT(ltd, tgt->ltd_index) = NULL;
420 clear_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
421 ltd->ltd_lov_desc.ld_tgt_count--;
423 ltd->ltd_lov_desc.ld_active_tgt_count--;
425 EXPORT_SYMBOL(ltd_del_tgt);
428 * Calculate penalties per-tgt and per-server
430 * Re-calculate penalties when the configuration changes, active targets
431 * change and after statfs refresh (all these are reflected by lq_dirty flag).
432 * On every tgt and server: decay the penalty by half for every 8x the update
433 * interval that the device has been idle. That gives lots of time for the
434 * statfs information to be updated (which the penalty is only a proxy for),
435 * and avoids penalizing server/tgt under light load.
436 * See lu_qos_tgt_weight_calc() for how penalties are factored into the weight.
438 * \param[in] ltd lu_tgt_descs
440 * \retval 0 on success
441 * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
444 int ltd_qos_penalties_calc(struct lu_tgt_descs *ltd)
446 struct lu_qos *qos = <d->ltd_qos;
447 struct lov_desc *desc = <d->ltd_lov_desc;
448 struct lu_tgt_desc *tgt;
449 struct lu_svr_qos *svr;
450 __u64 ba_max, ba_min, ba;
451 __u64 ia_max, ia_min, ia = 1;
459 if (!test_bit(LQ_DIRTY, &qos->lq_flags))
462 num_active = desc->ld_active_tgt_count - 1;
464 GOTO(out, rc = -EAGAIN);
466 /* find bavail on each server */
467 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
469 /* if inode is not counted, set to 1 to ignore */
470 svr->lsq_iavail = ltd->ltd_is_mdt ? 0 : 1;
472 qos->lq_active_svr_count = 0;
475 * How badly user wants to select targets "widely" (not recently chosen
476 * and not on recent MDS's). As opposed to "freely" (free space avail.)
479 prio_wide = 256 - qos->lq_prio_free;
481 ba_min = (__u64)(-1);
483 ia_min = (__u64)(-1);
485 now = ktime_get_real_seconds();
487 /* Calculate server penalty per object */
488 ltd_foreach_tgt(ltd, tgt) {
489 if (!tgt->ltd_active)
492 /* when inode is counted, bavail >> 16 to avoid overflow */
493 ba = tgt_statfs_bavail(tgt);
501 ba_min = min(ba, ba_min);
502 ba_max = max(ba, ba_max);
504 /* Count the number of usable servers */
505 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
506 qos->lq_active_svr_count++;
507 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
509 if (ltd->ltd_is_mdt) {
510 /* iavail >> 8 to avoid overflow */
511 ia = tgt_statfs_iavail(tgt) >> 8;
515 ia_min = min(ia, ia_min);
516 ia_max = max(ia, ia_max);
518 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
523 * prio * bavail * iavail / (num_tgt - 1) / 2
525 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia >> 8;
526 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
527 tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
529 age = (now - tgt->ltd_qos.ltq_used) >> 3;
530 if (test_bit(LQ_RESET, &qos->lq_flags) ||
531 age > 32 * desc->ld_qos_maxage)
532 tgt->ltd_qos.ltq_penalty = 0;
533 else if (age > desc->ld_qos_maxage)
534 /* Decay tgt penalty. */
535 tgt->ltd_qos.ltq_penalty >>= age / desc->ld_qos_maxage;
538 num_active = qos->lq_active_svr_count - 1;
539 if (num_active < 1) {
541 * If there's only 1 server, we can't penalize it, so instead
542 * we have to double the tgt penalty
545 ltd_foreach_tgt(ltd, tgt) {
546 if (!tgt->ltd_active)
549 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
554 * Per-server penalty is
555 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
557 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
558 ba = svr->lsq_bavail;
559 ia = svr->lsq_iavail;
560 svr->lsq_penalty_per_obj = prio_wide * ba * ia >> 8;
561 do_div(svr->lsq_penalty_per_obj,
562 svr->lsq_tgt_count * num_active);
563 svr->lsq_penalty_per_obj >>= 1;
565 age = (now - svr->lsq_used) >> 3;
566 if (test_bit(LQ_RESET, &qos->lq_flags) ||
567 age > 32 * desc->ld_qos_maxage)
568 svr->lsq_penalty = 0;
569 else if (age > desc->ld_qos_maxage)
570 /* Decay server penalty. */
571 svr->lsq_penalty >>= age / desc->ld_qos_maxage;
574 clear_bit(LQ_DIRTY, &qos->lq_flags);
575 clear_bit(LQ_RESET, &qos->lq_flags);
578 * If each tgt has almost same free space, do rr allocation for better
579 * creation performance
581 clear_bit(LQ_SAME_SPACE, &qos->lq_flags);
582 if (((ba_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
583 QOS_THRESHOLD_MAX) < ba_min &&
584 ((ia_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
585 QOS_THRESHOLD_MAX) < ia_min) {
586 set_bit(LQ_SAME_SPACE, &qos->lq_flags);
587 /* Reset weights for the next time we enter qos mode */
588 set_bit(LQ_RESET, &qos->lq_flags);
593 if (!rc && test_bit(LQ_SAME_SPACE, &qos->lq_flags))
598 EXPORT_SYMBOL(ltd_qos_penalties_calc);
601 * Re-calculate penalties and weights of all tgts.
603 * The function is called when some target was used for a new object. In
604 * this case we should re-calculate all the weights to keep new allocations
607 * \param[in] ltd lu_tgt_descs
608 * \param[in] tgt recently used tgt
609 * \param[out] total_wt new total weight for the pool
613 int ltd_qos_update(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt,
616 struct lu_qos *qos = <d->ltd_qos;
617 struct lu_tgt_qos *ltq;
618 struct lu_svr_qos *svr;
625 /* Don't allocate on this device anymore, until the next alloc_qos */
631 * Decay old penalty by half (we're adding max penalty, and don't
632 * want it to run away.)
634 ltq->ltq_penalty >>= 1;
635 svr->lsq_penalty >>= 1;
637 /* mark the server and tgt as recently used */
638 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
640 /* Set max penalties for this tgt and server */
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;
646 /* Decrease all MDS penalties */
647 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
648 if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
649 svr->lsq_penalty = 0;
651 svr->lsq_penalty -= svr->lsq_penalty_per_obj;
655 /* Decrease all tgt penalties */
656 ltd_foreach_tgt(ltd, tgt) {
657 if (!tgt->ltd_active)
661 if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
662 ltq->ltq_penalty = 0;
664 ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
666 lu_tgt_qos_weight_calc(tgt);
668 /* Recalc the total weight of usable osts */
670 *total_wt += ltq->ltq_weight;
672 CDEBUG(D_OTHER, "recalc tgt %d usable=%d bavail=%llu ffree=%llu tgtppo=%llu tgtp=%llu svrppo=%llu svrp=%llu wt=%llu\n",
673 tgt->ltd_index, ltq->ltq_usable,
674 tgt_statfs_bavail(tgt) >> 16,
675 tgt_statfs_iavail(tgt) >> 8,
676 ltq->ltq_penalty_per_obj >> 10,
677 ltq->ltq_penalty >> 10,
678 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
679 ltq->ltq_svr->lsq_penalty >> 10,
680 ltq->ltq_weight >> 10);
685 EXPORT_SYMBOL(ltd_qos_update);