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);
82 void lu_qos_rr_init(struct lu_qos_rr *lqr)
84 spin_lock_init(&lqr->lqr_alloc);
87 EXPORT_SYMBOL(lu_qos_rr_init);
90 * Add a new target to Quality of Service (QoS) target table.
92 * Add a new MDT/OST target to the structure representing an OSS. Resort the
93 * list of known MDSs/OSSs by the number of MDTs/OSTs attached to each MDS/OSS.
94 * The MDS/OSS list is protected internally and no external locking is required.
96 * \param[in] qos lu_qos data
97 * \param[in] tgt target description
99 * \retval 0 on success
100 * \retval -ENOMEM on error
102 int lu_qos_add_tgt(struct lu_qos *qos, struct lu_tgt_desc *tgt)
104 struct lu_svr_qos *svr = NULL;
105 struct lu_svr_qos *tempsvr;
106 struct obd_export *exp = tgt->ltd_exp;
113 down_write(&qos->lq_rw_sem);
115 * a bit hacky approach to learn NID of corresponding connection
116 * but there is no official API to access information like this
119 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
120 if (obd_uuid_equals(&svr->lsq_uuid,
121 &exp->exp_connection->c_remote_uuid)) {
125 if (svr->lsq_id > id)
132 GOTO(out, rc = -ENOMEM);
133 memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid,
134 sizeof(svr->lsq_uuid));
138 /* Assume we have to move this one */
139 list_del(&svr->lsq_svr_list);
142 svr->lsq_tgt_count++;
143 tgt->ltd_qos.ltq_svr = svr;
145 CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n",
146 obd_uuid2str(&tgt->ltd_uuid), obd_uuid2str(&svr->lsq_uuid),
150 * Add sorted by # of tgts. Find the first entry that we're
153 list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) {
154 if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count)
158 * ...and add before it. If we're the first or smallest, tempsvr
159 * points to the list head, and we add to the end.
161 list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list);
164 qos->lq_rr.lqr_dirty = 1;
167 up_write(&qos->lq_rw_sem);
170 EXPORT_SYMBOL(lu_qos_add_tgt);
173 * Remove MDT/OST target from QoS table.
175 * Removes given MDT/OST target from QoS table and releases related
176 * MDS/OSS structure if no target remain on the MDS/OSS.
178 * \param[in] qos lu_qos data
179 * \param[in] ltd target description
181 * \retval 0 on success
182 * \retval -ENOENT if no server was found
184 static int lu_qos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd)
186 struct lu_svr_qos *svr;
191 down_write(&qos->lq_rw_sem);
192 svr = ltd->ltd_qos.ltq_svr;
194 GOTO(out, rc = -ENOENT);
196 svr->lsq_tgt_count--;
197 if (svr->lsq_tgt_count == 0) {
198 CDEBUG(D_OTHER, "removing server %s\n",
199 obd_uuid2str(&svr->lsq_uuid));
200 list_del(&svr->lsq_svr_list);
201 ltd->ltd_qos.ltq_svr = NULL;
206 qos->lq_rr.lqr_dirty = 1;
208 up_write(&qos->lq_rw_sem);
212 static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
214 struct obd_statfs *statfs = &tgt->ltd_statfs;
216 return statfs->os_bavail * statfs->os_bsize;
219 static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
221 return tgt->ltd_statfs.os_ffree;
225 * Calculate weight for a given tgt.
227 * The final tgt weight is bavail >> 16 * iavail >> 8 minus the tgt and server
228 * penalties. See ltd_qos_penalties_calc() for how penalties are calculated.
230 * \param[in] tgt target descriptor
232 void lu_tgt_qos_weight_calc(struct lu_tgt_desc *tgt)
234 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
237 temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8);
238 temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
242 ltq->ltq_weight = temp - temp2;
244 EXPORT_SYMBOL(lu_tgt_qos_weight_calc);
247 * Allocate and initialize target table.
249 * A helper function to initialize the target table and allocate
250 * a bitmap of the available targets.
252 * \param[in] ltd target's table to initialize
253 * \param[in] is_mdt target table for MDTs
255 * \retval 0 on success
256 * \retval negative negated errno on error
258 int lu_tgt_descs_init(struct lu_tgt_descs *ltd, bool is_mdt)
260 mutex_init(<d->ltd_mutex);
261 init_rwsem(<d->ltd_rw_sem);
264 * the tgt array and bitmap are allocated/grown dynamically as tgts are
265 * added to the LOD/LMV, see lu_tgt_descs_add()
267 ltd->ltd_tgt_bitmap = bitmap_zalloc(BITS_PER_LONG, GFP_NOFS);
268 if (!ltd->ltd_tgt_bitmap)
271 ltd->ltd_tgts_size = BITS_PER_LONG;
272 ltd->ltd_death_row = 0;
273 ltd->ltd_refcount = 0;
275 /* Set up allocation policy (QoS and RR) */
276 INIT_LIST_HEAD(<d->ltd_qos.lq_svr_list);
277 init_rwsem(<d->ltd_qos.lq_rw_sem);
278 ltd->ltd_qos.lq_dirty = 1;
279 ltd->ltd_qos.lq_reset = 1;
280 /* Default priority is toward free space balance */
281 ltd->ltd_qos.lq_prio_free = 232;
282 /* Default threshold for rr (roughly 17%) */
283 ltd->ltd_qos.lq_threshold_rr = 43;
284 ltd->ltd_is_mdt = is_mdt;
286 ltd->ltd_lmv_desc.ld_pattern = LMV_HASH_TYPE_DEFAULT;
288 lu_qos_rr_init(<d->ltd_qos.lq_rr);
292 EXPORT_SYMBOL(lu_tgt_descs_init);
295 * Free bitmap and target table pages.
297 * \param[in] ltd target table
299 void lu_tgt_descs_fini(struct lu_tgt_descs *ltd)
303 bitmap_free(ltd->ltd_tgt_bitmap);
304 for (i = 0; i < TGT_PTRS; i++) {
305 if (ltd->ltd_tgt_idx[i])
306 OBD_FREE_PTR(ltd->ltd_tgt_idx[i]);
308 ltd->ltd_tgts_size = 0;
310 EXPORT_SYMBOL(lu_tgt_descs_fini);
313 * Expand size of target table.
315 * When the target table is full, we have to extend the table. To do so,
316 * we allocate new memory with some reserve, move data from the old table
317 * to the new one and release memory consumed by the old table.
319 * \param[in] ltd target table
320 * \param[in] newsize new size of the table
322 * \retval 0 on success
323 * \retval -ENOMEM if reallocation failed
325 static int lu_tgt_descs_resize(struct lu_tgt_descs *ltd, __u32 newsize)
327 unsigned long *new_bitmap, *old_bitmap = NULL;
329 /* someone else has already resize the array */
330 if (newsize <= ltd->ltd_tgts_size)
333 new_bitmap = bitmap_zalloc(newsize, GFP_NOFS);
337 if (ltd->ltd_tgts_size > 0) {
338 /* the bitmap already exists, copy data from old one */
339 bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap,
341 old_bitmap = ltd->ltd_tgt_bitmap;
344 ltd->ltd_tgts_size = newsize;
345 ltd->ltd_tgt_bitmap = new_bitmap;
347 bitmap_free(old_bitmap);
349 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
355 * Add new target to target table.
357 * Extend target table if it's full, update target table and bitmap.
358 * Notice we need to take ltd_rw_sem exclusively before entry to ensure
361 * \param[in] ltd target table
362 * \param[in] tgt new target desc
364 * \retval 0 on success
365 * \retval -ENOMEM if reallocation failed
366 * -EEXIST if target existed
368 int ltd_add_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
370 __u32 index = tgt->ltd_index;
375 if (index >= ltd->ltd_tgts_size) {
378 while (newsize < index + 1)
379 newsize = newsize << 1;
381 rc = lu_tgt_descs_resize(ltd, newsize);
384 } else if (test_bit(index, ltd->ltd_tgt_bitmap)) {
388 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
389 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
390 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
394 LTD_TGT(ltd, tgt->ltd_index) = tgt;
395 set_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
397 ltd->ltd_lov_desc.ld_tgt_count++;
399 ltd->ltd_lov_desc.ld_active_tgt_count++;
403 EXPORT_SYMBOL(ltd_add_tgt);
406 * Delete target from target table
408 void ltd_del_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
410 lu_qos_del_tgt(<d->ltd_qos, tgt);
411 LTD_TGT(ltd, tgt->ltd_index) = NULL;
412 clear_bit(tgt->ltd_index, ltd->ltd_tgt_bitmap);
413 ltd->ltd_lov_desc.ld_tgt_count--;
415 ltd->ltd_lov_desc.ld_active_tgt_count--;
417 EXPORT_SYMBOL(ltd_del_tgt);
420 * Whether QoS data is up-to-date and QoS can be applied.
422 bool ltd_qos_is_usable(struct lu_tgt_descs *ltd)
424 if (!ltd->ltd_qos.lq_dirty && ltd->ltd_qos.lq_same_space)
427 if (ltd->ltd_lov_desc.ld_active_tgt_count < 2)
432 EXPORT_SYMBOL(ltd_qos_is_usable);
435 * Calculate penalties per-tgt and per-server
437 * Re-calculate penalties when the configuration changes, active targets
438 * change and after statfs refresh (all these are reflected by lq_dirty flag).
439 * On every tgt and server: decay the penalty by half for every 8x the update
440 * interval that the device has been idle. That gives lots of time for the
441 * statfs information to be updated (which the penalty is only a proxy for),
442 * and avoids penalizing server/tgt under light load.
443 * See lu_qos_tgt_weight_calc() for how penalties are factored into the weight.
445 * \param[in] ltd lu_tgt_descs
447 * \retval 0 on success
448 * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
451 int ltd_qos_penalties_calc(struct lu_tgt_descs *ltd)
453 struct lu_qos *qos = <d->ltd_qos;
454 struct lov_desc *desc = <d->ltd_lov_desc;
455 struct lu_tgt_desc *tgt;
456 struct lu_svr_qos *svr;
457 __u64 ba_max, ba_min, ba;
458 __u64 ia_max, ia_min, ia = 1;
469 num_active = desc->ld_active_tgt_count - 1;
471 GOTO(out, rc = -EAGAIN);
473 /* find bavail on each server */
474 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
476 /* if inode is not counted, set to 1 to ignore */
477 svr->lsq_iavail = ltd->ltd_is_mdt ? 0 : 1;
479 qos->lq_active_svr_count = 0;
482 * How badly user wants to select targets "widely" (not recently chosen
483 * and not on recent MDS's). As opposed to "freely" (free space avail.)
486 prio_wide = 256 - qos->lq_prio_free;
488 ba_min = (__u64)(-1);
490 ia_min = (__u64)(-1);
492 now = ktime_get_real_seconds();
494 /* Calculate server penalty per object */
495 ltd_foreach_tgt(ltd, tgt) {
496 if (!tgt->ltd_active)
499 /* when inode is counted, bavail >> 16 to avoid overflow */
500 ba = tgt_statfs_bavail(tgt);
508 ba_min = min(ba, ba_min);
509 ba_max = max(ba, ba_max);
511 /* Count the number of usable servers */
512 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
513 qos->lq_active_svr_count++;
514 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
516 if (ltd->ltd_is_mdt) {
517 /* iavail >> 8 to avoid overflow */
518 ia = tgt_statfs_iavail(tgt) >> 8;
522 ia_min = min(ia, ia_min);
523 ia_max = max(ia, ia_max);
525 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
530 * prio * bavail * iavail / (num_tgt - 1) / 2
532 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia >> 8;
533 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
534 tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
536 age = (now - tgt->ltd_qos.ltq_used) >> 3;
537 if (qos->lq_reset || age > 32 * desc->ld_qos_maxage)
538 tgt->ltd_qos.ltq_penalty = 0;
539 else if (age > desc->ld_qos_maxage)
540 /* Decay tgt penalty. */
541 tgt->ltd_qos.ltq_penalty >>= age / desc->ld_qos_maxage;
544 num_active = qos->lq_active_svr_count - 1;
545 if (num_active < 1) {
547 * If there's only 1 server, we can't penalize it, so instead
548 * we have to double the tgt penalty
551 ltd_foreach_tgt(ltd, tgt) {
552 if (!tgt->ltd_active)
555 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
560 * Per-server penalty is
561 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
563 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
564 ba = svr->lsq_bavail;
565 ia = svr->lsq_iavail;
566 svr->lsq_penalty_per_obj = prio_wide * ba * ia >> 8;
567 do_div(svr->lsq_penalty_per_obj,
568 svr->lsq_tgt_count * num_active);
569 svr->lsq_penalty_per_obj >>= 1;
571 age = (now - svr->lsq_used) >> 3;
572 if (qos->lq_reset || age > 32 * desc->ld_qos_maxage)
573 svr->lsq_penalty = 0;
574 else if (age > desc->ld_qos_maxage)
575 /* Decay server penalty. */
576 svr->lsq_penalty >>= age / desc->ld_qos_maxage;
583 * If each tgt has almost same free space, do rr allocation for better
584 * creation performance
586 qos->lq_same_space = 0;
587 if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min &&
588 (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) {
589 qos->lq_same_space = 1;
590 /* Reset weights for the next time we enter qos mode */
596 if (!rc && qos->lq_same_space)
601 EXPORT_SYMBOL(ltd_qos_penalties_calc);
604 * Re-calculate penalties and weights of all tgts.
606 * The function is called when some target was used for a new object. In
607 * this case we should re-calculate all the weights to keep new allocations
610 * \param[in] ltd lu_tgt_descs
611 * \param[in] tgt recently used tgt
612 * \param[out] total_wt new total weight for the pool
616 int ltd_qos_update(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt,
619 struct lu_qos *qos = <d->ltd_qos;
620 struct lu_tgt_qos *ltq;
621 struct lu_svr_qos *svr;
628 /* Don't allocate on this device anymore, until the next alloc_qos */
634 * Decay old penalty by half (we're adding max penalty, and don't
635 * want it to run away.)
637 ltq->ltq_penalty >>= 1;
638 svr->lsq_penalty >>= 1;
640 /* mark the server and tgt as recently used */
641 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
643 /* Set max penalties for this tgt and server */
644 ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
645 ltd->ltd_lov_desc.ld_active_tgt_count;
646 svr->lsq_penalty += svr->lsq_penalty_per_obj *
647 ltd->ltd_lov_desc.ld_active_tgt_count;
649 /* Decrease all MDS penalties */
650 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
651 if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
652 svr->lsq_penalty = 0;
654 svr->lsq_penalty -= svr->lsq_penalty_per_obj;
658 /* Decrease all tgt penalties */
659 ltd_foreach_tgt(ltd, tgt) {
660 if (!tgt->ltd_active)
664 if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
665 ltq->ltq_penalty = 0;
667 ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
669 lu_tgt_qos_weight_calc(tgt);
671 /* Recalc the total weight of usable osts */
673 *total_wt += ltq->ltq_weight;
675 CDEBUG(D_OTHER, "recalc tgt %d usable=%d bavail=%llu ffree=%llu tgtppo=%llu tgtp=%llu svrppo=%llu svrp=%llu wt=%llu\n",
676 tgt->ltd_index, ltq->ltq_usable,
677 tgt_statfs_bavail(tgt) >> 16,
678 tgt_statfs_iavail(tgt) >> 8,
679 ltq->ltq_penalty_per_obj >> 10,
680 ltq->ltq_penalty >> 10,
681 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
682 ltq->ltq_svr->lsq_penalty >> 10,
683 ltq->ltq_weight >> 10);
688 EXPORT_SYMBOL(ltd_qos_update);