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/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 #if BITS_PER_LONG == 32
61 * If ep_ro > 32-bit, first generate the high
62 * 32 bits of the random number, then add in the low
63 * 32 bits (truncated to the upper limit, if needed)
65 if (ep_ro > 0xffffffffULL)
66 rand = (u64)prandom_u32_max((u32)(ep_ro >> 32)) << 32;
68 if (rand == (ep_ro & 0xffffffff00000000ULL))
69 rand |= prandom_u32_max((u32)ep_ro);
71 rand |= prandom_u32();
73 rand = ((u64)prandom_u32() << 32 | prandom_u32()) % ep_ro;
79 EXPORT_SYMBOL(lu_prandom_u64_max);
82 * Add a new target to Quality of Service (QoS) target table.
84 * Add a new MDT/OST target to the structure representing an OSS. Resort the
85 * list of known MDSs/OSSs by the number of MDTs/OSTs attached to each MDS/OSS.
86 * The MDS/OSS list is protected internally and no external locking is required.
88 * \param[in] qos lu_qos data
89 * \param[in] tgt target description
91 * \retval 0 on success
92 * \retval -ENOMEM on error
94 int lu_qos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *tgt)
96 struct lu_svr_qos *svr = NULL;
97 struct lu_svr_qos *tempsvr;
98 struct obd_export *exp = tgt->ltd_exp;
105 down_write(&qos->lq_rw_sem);
107 * a bit hacky approach to learn NID of corresponding connection
108 * but there is no official API to access information like this
111 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
112 if (obd_uuid_equals(&svr->lsq_uuid,
113 &exp->exp_connection->c_remote_uuid)) {
117 if (svr->lsq_id > id)
124 GOTO(out, rc = -ENOMEM);
125 memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid,
126 sizeof(svr->lsq_uuid));
130 /* Assume we have to move this one */
131 list_del(&svr->lsq_svr_list);
134 svr->lsq_tgt_count++;
135 tgt->ltd_qos.ltq_svr = svr;
137 CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n",
138 obd_uuid2str(&tgt->ltd_uuid), obd_uuid2str(&svr->lsq_uuid),
142 * Add sorted by # of tgts. Find the first entry that we're
145 list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) {
146 if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count)
150 * ...and add before it. If we're the first or smallest, tempsvr
151 * points to the list head, and we add to the end.
153 list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list);
155 set_bit(LQ_DIRTY, &qos->lq_flags);
156 #ifdef HAVE_SERVER_SUPPORT
157 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
160 up_write(&qos->lq_rw_sem);
163 EXPORT_SYMBOL(lu_qos_add_tgt);
166 * Remove MDT/OST target from QoS table.
168 * Removes given MDT/OST target from QoS table and releases related
169 * MDS/OSS structure if no target remain on the MDS/OSS.
171 * \param[in] qos lu_qos data
172 * \param[in] ltd target description
174 * \retval 0 on success
175 * \retval -ENOENT if no server was found
177 int lu_qos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd)
179 struct lu_svr_qos *svr;
184 down_write(&qos->lq_rw_sem);
185 svr = ltd->ltd_qos.ltq_svr;
187 GOTO(out, rc = -ENOENT);
189 ltd->ltd_qos.ltq_svr = NULL;
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);
198 set_bit(LQ_DIRTY, &qos->lq_flags);
199 #ifdef HAVE_SERVER_SUPPORT
200 set_bit(LQ_DIRTY, &qos->lq_rr.lqr_flags);
203 up_write(&qos->lq_rw_sem);
206 EXPORT_SYMBOL(lu_qos_del_tgt);
209 * Calculate weight for a given tgt.
211 * The final tgt weight uses only free space for OSTs, but combines
212 * both free space and inodes for MDTs, minus tgt and server penalties.
213 * See ltd_qos_penalties_calc() for how penalties are calculated.
215 * \param[in] tgt target descriptor
216 * \param[in] is_mdt target table is for MDT selection (use inodes)
218 void lu_tgt_qos_weight_calc(struct lu_tgt_desc *tgt, bool is_mdt)
220 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
224 ltq->ltq_avail = (tgt_statfs_bavail(tgt) >> 16) *
225 (tgt_statfs_iavail(tgt) >> 8);
227 ltq->ltq_avail = tgt_statfs_bavail(tgt) >> 8;
228 penalty = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
229 if (ltq->ltq_avail < penalty)
232 ltq->ltq_weight = ltq->ltq_avail - penalty;
234 EXPORT_SYMBOL(lu_tgt_qos_weight_calc);
237 * Allocate and initialize target table.
239 * A helper function to initialize the target table and allocate
240 * a bitmap of the available targets.
242 * \param[in] ltd target's table to initialize
243 * \param[in] is_mdt target table for MDTs
245 * \retval 0 on success
246 * \retval negative negated errno on error
248 int lu_tgt_descs_init(struct lu_tgt_descs *ltd, bool is_mdt)
250 mutex_init(<d->ltd_mutex);
251 init_rwsem(<d->ltd_rw_sem);
254 * the tgt array and bitmap are allocated/grown dynamically as tgts are
255 * added to the LOD/LMV, see lu_tgt_descs_add()
257 ltd->ltd_tgt_bitmap = bitmap_zalloc(BITS_PER_LONG, GFP_NOFS);
258 if (!ltd->ltd_tgt_bitmap)
261 ltd->ltd_tgts_size = BITS_PER_LONG;
262 ltd->ltd_death_row = 0;
263 ltd->ltd_refcount = 0;
265 /* Set up allocation policy (QoS and RR) */
266 INIT_LIST_HEAD(<d->ltd_qos.lq_svr_list);
267 init_rwsem(<d->ltd_qos.lq_rw_sem);
268 set_bit(LQ_DIRTY, <d->ltd_qos.lq_flags);
269 set_bit(LQ_RESET, <d->ltd_qos.lq_flags);
270 ltd->ltd_is_mdt = is_mdt;
271 /* MDT imbalance threshold is low to balance across MDTs
272 * relatively quickly, because each directory may result
273 * in a large number of files/subdirs created therein.
276 ltd->ltd_lmv_desc.ld_pattern = LMV_HASH_TYPE_DEFAULT;
277 ltd->ltd_qos.lq_prio_free = LMV_QOS_DEF_PRIO_FREE * 256 / 100;
278 ltd->ltd_qos.lq_threshold_rr =
279 LMV_QOS_DEF_THRESHOLD_RR_PCT *
280 QOS_THRESHOLD_MAX / 100;
282 ltd->ltd_qos.lq_prio_free = LOV_QOS_DEF_PRIO_FREE * 256 / 100;
283 ltd->ltd_qos.lq_threshold_rr =
284 LOV_QOS_DEF_THRESHOLD_RR_PCT *
285 QOS_THRESHOLD_MAX / 100;
290 EXPORT_SYMBOL(lu_tgt_descs_init);
293 * Free bitmap and target table pages.
295 * \param[in] ltd target table
297 void lu_tgt_descs_fini(struct lu_tgt_descs *ltd)
301 bitmap_free(ltd->ltd_tgt_bitmap);
302 for (i = 0; i < ARRAY_SIZE(ltd->ltd_tgt_idx); i++) {
303 if (ltd->ltd_tgt_idx[i])
304 OBD_FREE_PTR(ltd->ltd_tgt_idx[i]);
306 ltd->ltd_tgts_size = 0;
308 EXPORT_SYMBOL(lu_tgt_descs_fini);
311 * Expand size of target table.
313 * When the target table is full, we have to extend the table. To do so,
314 * we allocate new memory with some reserve, move data from the old table
315 * to the new one and release memory consumed by the old table.
317 * \param[in] ltd target table
318 * \param[in] newsize new size of the table
320 * \retval 0 on success
321 * \retval -ENOMEM if reallocation failed
323 static int lu_tgt_descs_resize(struct lu_tgt_descs *ltd, __u32 newsize)
325 unsigned long *new_bitmap, *old_bitmap = NULL;
327 /* someone else has already resize the array */
328 if (newsize <= ltd->ltd_tgts_size)
331 new_bitmap = bitmap_zalloc(newsize, GFP_NOFS);
335 if (ltd->ltd_tgts_size > 0) {
336 /* the bitmap already exists, copy data from old one */
337 bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap,
339 old_bitmap = ltd->ltd_tgt_bitmap;
342 ltd->ltd_tgts_size = newsize;
343 ltd->ltd_tgt_bitmap = new_bitmap;
345 bitmap_free(old_bitmap);
347 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
353 * Add new target to target table.
355 * Extend target table if it's full, update target table and bitmap.
356 * Notice we need to take ltd_rw_sem exclusively before entry to ensure
359 * \param[in] ltd target table
360 * \param[in] tgt new target desc
362 * \retval 0 on success
363 * \retval -ENOMEM if reallocation failed
364 * -EEXIST if target existed
366 int ltd_add_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
368 __u32 index = tgt->ltd_index;
373 if (index >= ltd->ltd_tgts_size) {
376 if (index > TGT_PTRS * TGT_PTRS_PER_BLOCK)
379 while (newsize < index + 1)
380 newsize = newsize << 1;
382 rc = lu_tgt_descs_resize(ltd, newsize);
385 } else if (test_bit(index, ltd->ltd_tgt_bitmap)) {
389 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
390 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
391 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
395 LTD_TGT(ltd, tgt->ltd_index) = tgt;
396 set_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
398 ltd->ltd_lov_desc.ld_tgt_count++;
400 ltd->ltd_lov_desc.ld_active_tgt_count++;
404 EXPORT_SYMBOL(ltd_add_tgt);
407 * Delete target from target table
409 void ltd_del_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
411 lu_qos_del_tgt(<d->ltd_qos, tgt);
412 LTD_TGT(ltd, tgt->ltd_index) = NULL;
413 clear_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
414 ltd->ltd_lov_desc.ld_tgt_count--;
416 ltd->ltd_lov_desc.ld_active_tgt_count--;
418 EXPORT_SYMBOL(ltd_del_tgt);
421 * Calculate penalties per-tgt and per-server
423 * Re-calculate penalties when the configuration changes, active targets
424 * change and after statfs refresh (all these are reflected by lq_dirty flag).
425 * On every tgt and server: decay the penalty by half for every 8x the update
426 * interval that the device has been idle. That gives lots of time for the
427 * statfs information to be updated (which the penalty is only a proxy for),
428 * and avoids penalizing server/tgt under light load.
429 * See lu_qos_tgt_weight_calc() for how penalties are factored into the weight.
431 * \param[in] ltd lu_tgt_descs
433 * \retval 0 on success
434 * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
437 int ltd_qos_penalties_calc(struct lu_tgt_descs *ltd)
439 struct lu_qos *qos = <d->ltd_qos;
440 struct lov_desc *desc = <d->ltd_lov_desc;
441 struct lu_tgt_desc *tgt;
442 struct lu_svr_qos *svr;
443 __u64 ba_max, ba_min, ba;
444 __u64 ia_max, ia_min, ia = 1;
452 if (!test_bit(LQ_DIRTY, &qos->lq_flags))
455 num_active = desc->ld_active_tgt_count - 1;
457 GOTO(out, rc = -EAGAIN);
459 /* find bavail on each server */
460 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
462 /* if inode is not counted, set to 1 to ignore */
463 svr->lsq_iavail = ltd->ltd_is_mdt ? 0 : 1;
465 qos->lq_active_svr_count = 0;
468 * How badly user wants to select targets "widely" (not recently chosen
469 * and not on recent MDS's). As opposed to "freely" (free space avail.)
472 prio_wide = 256 - qos->lq_prio_free;
474 ba_min = (__u64)(-1);
476 ia_min = (__u64)(-1);
478 now = ktime_get_real_seconds();
480 /* Calculate server penalty per object */
481 ltd_foreach_tgt(ltd, tgt) {
482 if (!tgt->ltd_active)
485 /* when inode is counted, bavail >> 16 to avoid overflow */
486 ba = tgt_statfs_bavail(tgt);
494 ba_min = min(ba, ba_min);
495 ba_max = max(ba, ba_max);
497 /* Count the number of usable servers */
498 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
499 qos->lq_active_svr_count++;
500 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
502 if (ltd->ltd_is_mdt) {
503 /* iavail >> 8 to avoid overflow */
504 ia = tgt_statfs_iavail(tgt) >> 8;
508 ia_min = min(ia, ia_min);
509 ia_max = max(ia, ia_max);
511 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
516 * prio * bavail * iavail / (num_tgt - 1) / prio_max / 2
518 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia >> 9;
519 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
521 age = (now - tgt->ltd_qos.ltq_used) >> 3;
522 if (test_bit(LQ_RESET, &qos->lq_flags) ||
523 age > 32 * desc->ld_qos_maxage)
524 tgt->ltd_qos.ltq_penalty = 0;
525 else if (age > desc->ld_qos_maxage)
526 /* Decay tgt penalty. */
527 tgt->ltd_qos.ltq_penalty >>= age / desc->ld_qos_maxage;
530 num_active = qos->lq_active_svr_count - 1;
531 if (num_active < 1) {
533 * If there's only 1 server, we can't penalize it, so instead
534 * we have to double the tgt penalty
537 ltd_foreach_tgt(ltd, tgt) {
538 if (!tgt->ltd_active)
541 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
546 * Per-server penalty is
547 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
549 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
550 ba = svr->lsq_bavail;
551 ia = svr->lsq_iavail;
552 svr->lsq_penalty_per_obj = prio_wide * ba * ia >> 8;
553 do_div(svr->lsq_penalty_per_obj,
554 svr->lsq_tgt_count * num_active);
555 svr->lsq_penalty_per_obj >>= 1;
557 age = (now - svr->lsq_used) >> 3;
558 if (test_bit(LQ_RESET, &qos->lq_flags) ||
559 age > 32 * desc->ld_qos_maxage)
560 svr->lsq_penalty = 0;
561 else if (age > desc->ld_qos_maxage)
562 /* Decay server penalty. */
563 svr->lsq_penalty >>= age / desc->ld_qos_maxage;
568 * If each tgt has almost same free space, do rr allocation for better
569 * creation performance
571 if (((ba_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
572 QOS_THRESHOLD_MAX) < ba_min &&
573 ((ia_max * (QOS_THRESHOLD_MAX - qos->lq_threshold_rr)) /
574 QOS_THRESHOLD_MAX) < ia_min) {
575 set_bit(LQ_SAME_SPACE, &qos->lq_flags);
576 /* Reset weights for the next time we enter qos mode */
577 set_bit(LQ_RESET, &qos->lq_flags);
579 clear_bit(LQ_SAME_SPACE, &qos->lq_flags);
580 clear_bit(LQ_RESET, &qos->lq_flags);
582 clear_bit(LQ_DIRTY, &qos->lq_flags);
586 if (!rc && test_bit(LQ_SAME_SPACE, &qos->lq_flags))
591 EXPORT_SYMBOL(ltd_qos_penalties_calc);
594 * Re-calculate penalties and weights of all tgts.
596 * The function is called when some target was used for a new object. In
597 * this case we should re-calculate all the weights to keep new allocations
600 * \param[in] ltd lu_tgt_descs
601 * \param[in] tgt recently used tgt
602 * \param[out] total_wt new total weight for the pool
606 int ltd_qos_update(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt,
609 struct lu_qos *qos = <d->ltd_qos;
610 struct lu_tgt_qos *ltq;
611 struct lu_svr_qos *svr;
618 /* Don't allocate on this device anymore, until the next alloc_qos */
624 * Decay old penalty by half (we're adding max penalty, and don't
625 * want it to run away.)
627 ltq->ltq_penalty >>= 1;
628 svr->lsq_penalty >>= 1;
630 /* mark the server and tgt as recently used */
631 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
633 /* Set max penalties for this tgt and server */
634 ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
635 ltd->ltd_lov_desc.ld_active_tgt_count;
636 svr->lsq_penalty += svr->lsq_penalty_per_obj *
637 qos->lq_active_svr_count;
639 /* Decrease all MDS penalties */
640 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
641 if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
642 svr->lsq_penalty = 0;
644 svr->lsq_penalty -= svr->lsq_penalty_per_obj;
648 /* Decrease all tgt penalties */
649 ltd_foreach_tgt(ltd, tgt) {
650 if (!tgt->ltd_active)
654 if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
655 ltq->ltq_penalty = 0;
657 ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
659 lu_tgt_qos_weight_calc(tgt, ltd->ltd_is_mdt);
661 /* Recalc the total weight of usable osts */
663 *total_wt += ltq->ltq_weight;
665 CDEBUG(D_OTHER, "recalc tgt %d usable=%d bavail=%llu ffree=%llu tgtppo=%llu tgtp=%llu svrppo=%llu svrp=%llu wt=%llu\n",
666 tgt->ltd_index, ltq->ltq_usable,
667 tgt_statfs_bavail(tgt) >> 16,
668 tgt_statfs_iavail(tgt) >> 8,
669 ltq->ltq_penalty_per_obj >> 10,
670 ltq->ltq_penalty >> 10,
671 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
672 ltq->ltq_svr->lsq_penalty >> 10,
673 ltq->ltq_weight >> 10);
678 EXPORT_SYMBOL(ltd_qos_update);