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 = 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 /* tgt not connected, this function will be called again later */
117 down_write(&qos->lq_rw_sem);
119 * a bit hacky approach to learn NID of corresponding connection
120 * but there is no official API to access information like this
123 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
124 if (obd_uuid_equals(&svr->lsq_uuid,
125 &exp->exp_connection->c_remote_uuid)) {
129 if (svr->lsq_id > id)
136 GOTO(out, rc = -ENOMEM);
137 memcpy(&svr->lsq_uuid, &exp->exp_connection->c_remote_uuid,
138 sizeof(svr->lsq_uuid));
142 /* Assume we have to move this one */
143 list_del(&svr->lsq_svr_list);
146 svr->lsq_tgt_count++;
147 tgt->ltd_qos.ltq_svr = svr;
149 CDEBUG(D_OTHER, "add tgt %s to server %s (%d targets)\n",
150 obd_uuid2str(&tgt->ltd_uuid), obd_uuid2str(&svr->lsq_uuid),
154 * Add sorted by # of tgts. Find the first entry that we're
157 list_for_each_entry(tempsvr, &qos->lq_svr_list, lsq_svr_list) {
158 if (svr->lsq_tgt_count > tempsvr->lsq_tgt_count)
162 * ...and add before it. If we're the first or smallest, tempsvr
163 * points to the list head, and we add to the end.
165 list_add_tail(&svr->lsq_svr_list, &tempsvr->lsq_svr_list);
168 qos->lq_rr.lqr_dirty = 1;
171 up_write(&qos->lq_rw_sem);
174 EXPORT_SYMBOL(lu_qos_add_tgt);
177 * Remove MDT/OST target from QoS table.
179 * Removes given MDT/OST target from QoS table and releases related
180 * MDS/OSS structure if no target remain on the MDS/OSS.
182 * \param[in] qos lu_qos data
183 * \param[in] ltd target description
185 * \retval 0 on success
186 * \retval -ENOENT if no server was found
188 static int lu_qos_del_tgt(struct lu_qos *qos, struct lu_tgt_desc *ltd)
190 struct lu_svr_qos *svr;
195 down_write(&qos->lq_rw_sem);
196 svr = ltd->ltd_qos.ltq_svr;
198 GOTO(out, rc = -ENOENT);
200 svr->lsq_tgt_count--;
201 if (svr->lsq_tgt_count == 0) {
202 CDEBUG(D_OTHER, "removing server %s\n",
203 obd_uuid2str(&svr->lsq_uuid));
204 list_del(&svr->lsq_svr_list);
205 ltd->ltd_qos.ltq_svr = NULL;
210 qos->lq_rr.lqr_dirty = 1;
212 up_write(&qos->lq_rw_sem);
216 static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
218 struct obd_statfs *statfs = &tgt->ltd_statfs;
220 return statfs->os_bavail * statfs->os_bsize;
223 static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
225 return tgt->ltd_statfs.os_ffree;
229 * Calculate weight for a given tgt.
231 * The final tgt weight is bavail >> 16 * iavail >> 8 minus the tgt and server
232 * penalties. See ltd_qos_penalties_calc() for how penalties are calculated.
234 * \param[in] tgt target descriptor
236 void lu_tgt_qos_weight_calc(struct lu_tgt_desc *tgt)
238 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
241 temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8);
242 temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
246 ltq->ltq_weight = temp - temp2;
248 EXPORT_SYMBOL(lu_tgt_qos_weight_calc);
251 * Allocate and initialize target table.
253 * A helper function to initialize the target table and allocate
254 * a bitmap of the available targets.
256 * \param[in] ltd target's table to initialize
257 * \param[in] is_mdt target table for MDTs
259 * \retval 0 on success
260 * \retval negative negated errno on error
262 int lu_tgt_descs_init(struct lu_tgt_descs *ltd, bool is_mdt)
264 mutex_init(<d->ltd_mutex);
265 init_rwsem(<d->ltd_rw_sem);
268 * the tgt array and bitmap are allocated/grown dynamically as tgts are
269 * added to the LOD/LMV, see lu_tgt_descs_add()
271 ltd->ltd_tgt_bitmap = CFS_ALLOCATE_BITMAP(BITS_PER_LONG);
272 if (!ltd->ltd_tgt_bitmap)
275 ltd->ltd_tgts_size = BITS_PER_LONG;
276 ltd->ltd_death_row = 0;
277 ltd->ltd_refcount = 0;
279 /* Set up allocation policy (QoS and RR) */
280 INIT_LIST_HEAD(<d->ltd_qos.lq_svr_list);
281 init_rwsem(<d->ltd_qos.lq_rw_sem);
282 ltd->ltd_qos.lq_dirty = 1;
283 ltd->ltd_qos.lq_reset = 1;
284 /* Default priority is toward free space balance */
285 ltd->ltd_qos.lq_prio_free = 232;
286 /* Default threshold for rr (roughly 17%) */
287 ltd->ltd_qos.lq_threshold_rr = 43;
288 ltd->ltd_is_mdt = is_mdt;
290 lu_qos_rr_init(<d->ltd_qos.lq_rr);
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 CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
306 for (i = 0; i < TGT_PTRS; 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 struct cfs_bitmap *new_bitmap, *old_bitmap = NULL;
331 /* someone else has already resize the array */
332 if (newsize <= ltd->ltd_tgts_size)
335 new_bitmap = CFS_ALLOCATE_BITMAP(newsize);
339 if (ltd->ltd_tgts_size > 0) {
340 /* the bitmap already exists, copy data from old one */
341 cfs_bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap);
342 old_bitmap = ltd->ltd_tgt_bitmap;
345 ltd->ltd_tgts_size = newsize;
346 ltd->ltd_tgt_bitmap = new_bitmap;
349 CFS_FREE_BITMAP(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 while (newsize < index + 1)
381 newsize = newsize << 1;
383 rc = lu_tgt_descs_resize(ltd, newsize);
386 } else if (cfs_bitmap_check(ltd->ltd_tgt_bitmap, index)) {
390 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
391 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
392 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL)
396 LTD_TGT(ltd, tgt->ltd_index) = tgt;
397 cfs_bitmap_set(ltd->ltd_tgt_bitmap, tgt->ltd_index);
399 ltd->ltd_lov_desc.ld_tgt_count++;
401 ltd->ltd_lov_desc.ld_active_tgt_count++;
405 EXPORT_SYMBOL(ltd_add_tgt);
408 * Delete target from target table
410 void ltd_del_tgt(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt)
412 lu_qos_del_tgt(<d->ltd_qos, tgt);
413 LTD_TGT(ltd, tgt->ltd_index) = NULL;
414 cfs_bitmap_clear(ltd->ltd_tgt_bitmap, tgt->ltd_index);
415 ltd->ltd_lov_desc.ld_tgt_count--;
417 ltd->ltd_lov_desc.ld_active_tgt_count--;
419 EXPORT_SYMBOL(ltd_del_tgt);
422 * Whether QoS data is up-to-date and QoS can be applied.
424 bool ltd_qos_is_usable(struct lu_tgt_descs *ltd)
426 if (!ltd->ltd_qos.lq_dirty && ltd->ltd_qos.lq_same_space)
429 if (ltd->ltd_lov_desc.ld_active_tgt_count < 2)
434 EXPORT_SYMBOL(ltd_qos_is_usable);
437 * Calculate penalties per-tgt and per-server
439 * Re-calculate penalties when the configuration changes, active targets
440 * change and after statfs refresh (all these are reflected by lq_dirty flag).
441 * On every tgt and server: decay the penalty by half for every 8x the update
442 * interval that the device has been idle. That gives lots of time for the
443 * statfs information to be updated (which the penalty is only a proxy for),
444 * and avoids penalizing server/tgt under light load.
445 * See lu_qos_tgt_weight_calc() for how penalties are factored into the weight.
447 * \param[in] ltd lu_tgt_descs
449 * \retval 0 on success
450 * \retval -EAGAIN the number of tgt isn't enough or all tgt spaces are
453 int ltd_qos_penalties_calc(struct lu_tgt_descs *ltd)
455 struct lu_qos *qos = <d->ltd_qos;
456 struct lov_desc *desc = <d->ltd_lov_desc;
457 struct lu_tgt_desc *tgt;
458 struct lu_svr_qos *svr;
459 __u64 ba_max, ba_min, ba;
460 __u64 ia_max, ia_min, ia = 1;
471 num_active = desc->ld_active_tgt_count - 1;
473 GOTO(out, rc = -EAGAIN);
475 /* find bavail on each server */
476 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
478 /* if inode is not counted, set to 1 to ignore */
479 svr->lsq_iavail = ltd->ltd_is_mdt ? 0 : 1;
481 qos->lq_active_svr_count = 0;
484 * How badly user wants to select targets "widely" (not recently chosen
485 * and not on recent MDS's). As opposed to "freely" (free space avail.)
488 prio_wide = 256 - qos->lq_prio_free;
490 ba_min = (__u64)(-1);
492 ia_min = (__u64)(-1);
494 now = ktime_get_real_seconds();
496 /* Calculate server penalty per object */
497 ltd_foreach_tgt(ltd, tgt) {
498 if (!tgt->ltd_active)
501 /* when inode is counted, bavail >> 16 to avoid overflow */
502 ba = tgt_statfs_bavail(tgt);
510 ba_min = min(ba, ba_min);
511 ba_max = max(ba, ba_max);
513 /* Count the number of usable servers */
514 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
515 qos->lq_active_svr_count++;
516 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
518 if (ltd->ltd_is_mdt) {
519 /* iavail >> 8 to avoid overflow */
520 ia = tgt_statfs_iavail(tgt) >> 8;
524 ia_min = min(ia, ia_min);
525 ia_max = max(ia, ia_max);
527 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
532 * prio * bavail * iavail / (num_tgt - 1) / 2
534 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia;
535 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active);
536 tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
538 age = (now - tgt->ltd_qos.ltq_used) >> 3;
539 if (qos->lq_reset || age > 32 * desc->ld_qos_maxage)
540 tgt->ltd_qos.ltq_penalty = 0;
541 else if (age > desc->ld_qos_maxage)
542 /* Decay tgt penalty. */
543 tgt->ltd_qos.ltq_penalty >>= age / desc->ld_qos_maxage;
546 num_active = qos->lq_active_svr_count - 1;
547 if (num_active < 1) {
549 * If there's only 1 server, we can't penalize it, so instead
550 * we have to double the tgt penalty
553 ltd_foreach_tgt(ltd, tgt) {
554 if (!tgt->ltd_active)
557 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
562 * Per-server penalty is
563 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
565 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
566 ba = svr->lsq_bavail;
567 ia = svr->lsq_iavail;
568 svr->lsq_penalty_per_obj = prio_wide * ba * ia;
569 do_div(ba, svr->lsq_tgt_count * num_active);
570 svr->lsq_penalty_per_obj >>= 1;
572 age = (now - svr->lsq_used) >> 3;
573 if (qos->lq_reset || age > 32 * desc->ld_qos_maxage)
574 svr->lsq_penalty = 0;
575 else if (age > desc->ld_qos_maxage)
576 /* Decay server penalty. */
577 svr->lsq_penalty >>= age / desc->ld_qos_maxage;
584 * If each tgt has almost same free space, do rr allocation for better
585 * creation performance
587 qos->lq_same_space = 0;
588 if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min &&
589 (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) {
590 qos->lq_same_space = 1;
591 /* Reset weights for the next time we enter qos mode */
597 if (!rc && qos->lq_same_space)
602 EXPORT_SYMBOL(ltd_qos_penalties_calc);
605 * Re-calculate penalties and weights of all tgts.
607 * The function is called when some target was used for a new object. In
608 * this case we should re-calculate all the weights to keep new allocations
611 * \param[in] ltd lu_tgt_descs
612 * \param[in] tgt recently used tgt
613 * \param[out] total_wt new total weight for the pool
617 int ltd_qos_update(struct lu_tgt_descs *ltd, struct lu_tgt_desc *tgt,
620 struct lu_qos *qos = <d->ltd_qos;
621 struct lu_tgt_qos *ltq;
622 struct lu_svr_qos *svr;
629 /* Don't allocate on this device anymore, until the next alloc_qos */
635 * Decay old penalty by half (we're adding max penalty, and don't
636 * want it to run away.)
638 ltq->ltq_penalty >>= 1;
639 svr->lsq_penalty >>= 1;
641 /* mark the server and tgt as recently used */
642 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
644 /* Set max penalties for this tgt and server */
645 ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
646 ltd->ltd_lov_desc.ld_active_tgt_count;
647 svr->lsq_penalty += svr->lsq_penalty_per_obj *
648 ltd->ltd_lov_desc.ld_active_tgt_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)
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 avail=%llu tgtppo=%llu tgtp=%llu svrppo=%llu svrp=%llu wt=%llu\n",
676 tgt->ltd_index, ltq->ltq_usable,
677 tgt_statfs_bavail(tgt) >> 10,
678 ltq->ltq_penalty_per_obj >> 10,
679 ltq->ltq_penalty >> 10,
680 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
681 ltq->ltq_svr->lsq_penalty >> 10,
682 ltq->ltq_weight >> 10);
687 EXPORT_SYMBOL(ltd_qos_update);