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/lmv/lmv_qos.c
28 * These are the only exported functions, they provide some generic
29 * infrastructure for object allocation QoS
33 #define DEBUG_SUBSYSTEM S_LMV
35 #include <asm/div64.h>
36 #include <linux/random.h>
38 #include <libcfs/libcfs.h>
39 #include <uapi/linux/lustre/lustre_idl.h>
40 #include <lustre_swab.h>
41 #include <obd_class.h>
43 #include "lmv_internal.h"
45 static inline __u64 tgt_statfs_bavail(struct lu_tgt_desc *tgt)
47 struct obd_statfs *statfs = &tgt->ltd_statfs;
49 return statfs->os_bavail * statfs->os_bsize;
52 static inline __u64 tgt_statfs_iavail(struct lu_tgt_desc *tgt)
54 return tgt->ltd_statfs.os_ffree;
58 * Calculate penalties per-tgt and per-server
60 * Re-calculate penalties when the configuration changes, active targets
61 * change and after statfs refresh (all these are reflected by lq_dirty flag).
62 * On every MDT and MDS: decay the penalty by half for every 8x the update
63 * interval that the device has been idle. That gives lots of time for the
64 * statfs information to be updated (which the penalty is only a proxy for),
65 * and avoids penalizing MDS/MDTs under light load.
66 * See lmv_qos_calc_weight() for how penalties are factored into the weight.
68 * \param[in] lmv LMV device
70 * \retval 0 on success
71 * \retval -EAGAIN the number of MDTs isn't enough or all MDT spaces are
74 static int lmv_qos_calc_ppts(struct lmv_obd *lmv)
76 struct lu_qos *qos = &lmv->lmv_qos;
77 struct lu_tgt_desc *tgt;
78 struct lu_svr_qos *svr;
79 __u64 ba_max, ba_min, ba;
80 __u64 ia_max, ia_min, ia;
85 __u32 maxage = lmv->desc.ld_qos_maxage;
93 num_active = lmv->desc.ld_active_tgt_count;
95 GOTO(out, rc = -EAGAIN);
97 /* find bavail on each server */
98 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
102 qos->lq_active_svr_count = 0;
105 * How badly user wants to select targets "widely" (not recently chosen
106 * and not on recent MDS's). As opposed to "freely" (free space avail.)
109 prio_wide = 256 - qos->lq_prio_free;
111 ba_min = (__u64)(-1);
113 ia_min = (__u64)(-1);
115 now = ktime_get_real_seconds();
117 /* Calculate server penalty per object */
118 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
120 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
123 /* bavail >> 16 to avoid overflow */
124 ba = tgt_statfs_bavail(tgt) >> 16;
128 ba_min = min(ba, ba_min);
129 ba_max = max(ba, ba_max);
131 /* iavail >> 8 to avoid overflow */
132 ia = tgt_statfs_iavail(tgt) >> 8;
136 ia_min = min(ia, ia_min);
137 ia_max = max(ia, ia_max);
139 /* Count the number of usable MDS's */
140 if (tgt->ltd_qos.ltq_svr->lsq_bavail == 0)
141 qos->lq_active_svr_count++;
142 tgt->ltd_qos.ltq_svr->lsq_bavail += ba;
143 tgt->ltd_qos.ltq_svr->lsq_iavail += ia;
147 * prio * bavail * iavail / (num_tgt - 1) / 2
149 tgt->ltd_qos.ltq_penalty_per_obj = prio_wide * ba * ia;
150 do_div(tgt->ltd_qos.ltq_penalty_per_obj, num_active - 1);
151 tgt->ltd_qos.ltq_penalty_per_obj >>= 1;
153 age = (now - tgt->ltd_qos.ltq_used) >> 3;
154 if (qos->lq_reset || age > 32 * maxage)
155 tgt->ltd_qos.ltq_penalty = 0;
156 else if (age > maxage)
157 /* Decay tgt penalty. */
158 tgt->ltd_qos.ltq_penalty >>= (age / maxage);
161 num_active = qos->lq_active_svr_count;
162 if (num_active < 2) {
164 * If there's only 1 MDS, we can't penalize it, so instead
165 * we have to double the MDT penalty
168 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
170 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
173 tgt->ltd_qos.ltq_penalty_per_obj <<= 1;
179 * prio * bavail * iavail / server_tgts / (num_svr - 1) / 2
181 list_for_each_entry(svr, &qos->lq_svr_list, lsq_svr_list) {
182 ba = svr->lsq_bavail;
183 ia = svr->lsq_iavail;
184 svr->lsq_penalty_per_obj = prio_wide * ba * ia;
185 do_div(ba, svr->lsq_tgt_count * (num_active - 1));
186 svr->lsq_penalty_per_obj >>= 1;
188 age = (now - svr->lsq_used) >> 3;
189 if (qos->lq_reset || age > 32 * maxage)
190 svr->lsq_penalty = 0;
191 else if (age > maxage)
192 /* Decay server penalty. */
193 svr->lsq_penalty >>= age / maxage;
200 * If each MDT has almost same free space, do rr allocation for better
201 * creation performance
203 qos->lq_same_space = 0;
204 if ((ba_max * (256 - qos->lq_threshold_rr)) >> 8 < ba_min &&
205 (ia_max * (256 - qos->lq_threshold_rr)) >> 8 < ia_min) {
206 qos->lq_same_space = 1;
207 /* Reset weights for the next time we enter qos mode */
213 if (!rc && qos->lq_same_space)
219 static inline bool lmv_qos_is_usable(struct lmv_obd *lmv)
221 if (!lmv->lmv_qos.lq_dirty && lmv->lmv_qos.lq_same_space)
224 if (lmv->desc.ld_active_tgt_count < 2)
231 * Calculate weight for a given MDT.
233 * The final MDT weight is bavail >> 16 * iavail >> 8 minus the MDT and MDS
234 * penalties. See lmv_qos_calc_ppts() for how penalties are calculated.
236 * \param[in] tgt MDT target descriptor
238 static void lmv_qos_calc_weight(struct lu_tgt_desc *tgt)
240 struct lu_tgt_qos *ltq = &tgt->ltd_qos;
243 temp = (tgt_statfs_bavail(tgt) >> 16) * (tgt_statfs_iavail(tgt) >> 8);
244 temp2 = ltq->ltq_penalty + ltq->ltq_svr->lsq_penalty;
248 ltq->ltq_weight = temp - temp2;
252 * Re-calculate weights.
254 * The function is called when some target was used for a new object. In
255 * this case we should re-calculate all the weights to keep new allocations
258 * \param[in] lmv LMV device
259 * \param[in] tgt target where a new object was placed
260 * \param[out] total_wt new total weight for the pool
264 static int lmv_qos_used(struct lmv_obd *lmv, struct lu_tgt_desc *tgt,
267 struct lu_tgt_qos *ltq;
268 struct lu_svr_qos *svr;
276 /* Don't allocate on this device anymore, until the next alloc_qos */
282 * Decay old penalty by half (we're adding max penalty, and don't
283 * want it to run away.)
285 ltq->ltq_penalty >>= 1;
286 svr->lsq_penalty >>= 1;
288 /* mark the MDS and MDT as recently used */
289 ltq->ltq_used = svr->lsq_used = ktime_get_real_seconds();
291 /* Set max penalties for this MDT and MDS */
292 ltq->ltq_penalty += ltq->ltq_penalty_per_obj *
293 lmv->desc.ld_active_tgt_count;
294 svr->lsq_penalty += svr->lsq_penalty_per_obj *
295 lmv->lmv_qos.lq_active_svr_count;
297 /* Decrease all MDS penalties */
298 list_for_each_entry(svr, &lmv->lmv_qos.lq_svr_list, lsq_svr_list) {
299 if (svr->lsq_penalty < svr->lsq_penalty_per_obj)
300 svr->lsq_penalty = 0;
302 svr->lsq_penalty -= svr->lsq_penalty_per_obj;
306 /* Decrease all MDT penalties */
307 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
308 ltq = &lmv->tgts[i]->ltd_qos;
309 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
312 if (ltq->ltq_penalty < ltq->ltq_penalty_per_obj)
313 ltq->ltq_penalty = 0;
315 ltq->ltq_penalty -= ltq->ltq_penalty_per_obj;
317 lmv_qos_calc_weight(lmv->tgts[i]);
319 /* Recalc the total weight of usable osts */
321 *total_wt += ltq->ltq_weight;
323 CDEBUG(D_OTHER, "recalc tgt %d usable=%d avail=%llu"
324 " tgtppo=%llu tgtp=%llu svrppo=%llu"
325 " svrp=%llu wt=%llu\n",
327 tgt_statfs_bavail(tgt) >> 10,
328 ltq->ltq_penalty_per_obj >> 10,
329 ltq->ltq_penalty >> 10,
330 ltq->ltq_svr->lsq_penalty_per_obj >> 10,
331 ltq->ltq_svr->lsq_penalty >> 10,
332 ltq->ltq_weight >> 10);
338 struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
340 struct lu_tgt_desc *tgt;
341 __u64 total_weight = 0;
342 __u64 cur_weight = 0;
349 if (!lmv_qos_is_usable(lmv))
350 RETURN(ERR_PTR(-EAGAIN));
352 down_write(&lmv->lmv_qos.lq_rw_sem);
354 if (!lmv_qos_is_usable(lmv))
355 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
357 rc = lmv_qos_calc_ppts(lmv);
359 GOTO(unlock, tgt = ERR_PTR(rc));
361 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
366 tgt->ltd_qos.ltq_usable = 0;
367 if (!tgt->ltd_exp || !tgt->ltd_active)
370 tgt->ltd_qos.ltq_usable = 1;
371 lmv_qos_calc_weight(tgt);
372 total_weight += tgt->ltd_qos.ltq_weight;
375 rand = lu_prandom_u64_max(total_weight);
377 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
380 if (!tgt || !tgt->ltd_qos.ltq_usable)
383 cur_weight += tgt->ltd_qos.ltq_weight;
384 if (cur_weight < rand)
387 *mdt = tgt->ltd_index;
388 lmv_qos_used(lmv, tgt, &total_weight);
389 GOTO(unlock, rc = 0);
392 /* no proper target found */
393 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
395 up_write(&lmv->lmv_qos.lq_rw_sem);
400 struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
402 struct lu_tgt_desc *tgt;
407 spin_lock(&lmv->lmv_qos.lq_rr.lqr_alloc);
408 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
409 tgt = lmv->tgts[(i + lmv->lmv_qos_rr_index) %
410 lmv->desc.ld_tgt_count];
411 if (tgt && tgt->ltd_exp && tgt->ltd_active) {
412 *mdt = tgt->ltd_index;
413 lmv->lmv_qos_rr_index =
414 (i + lmv->lmv_qos_rr_index + 1) %
415 lmv->desc.ld_tgt_count;
416 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
421 spin_unlock(&lmv->lmv_qos.lq_rr.lqr_alloc);
423 RETURN(ERR_PTR(-ENODEV));