2 * Copyright (c) 2009-2010 Fabio Checconi
3 * Copyright (c) 2009-2010 Luigi Rizzo, Universita` di Pisa
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * A round-robin (RR) anticipatory scheduler, with per-client queues.
34 * The goal of this implementation is to improve throughput compared
35 * to the pure elevator algorithm, and insure some fairness among
38 * Requests coming from the same client are put in the same queue.
39 * We use anticipation to help reducing seeks, and each queue
40 * is never served continuously for more than a given amount of
41 * time or data. Queues are then served in a round-robin fashion.
43 * Each queue can be in any of the following states:
44 * READY immediately serve the first pending request;
45 * BUSY one request is under service, wait for completion;
46 * IDLING do not serve incoming requests immediately, unless
47 * they are "eligible" as defined later.
49 * Scheduling is made looking at the status of all queues,
50 * and the first one in round-robin order is privileged.
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
57 #include <sys/callout.h>
58 #include <sys/malloc.h>
59 #include <sys/module.h>
61 #include <sys/queue.h>
63 #include <sys/sysctl.h>
64 #include "gs_scheduler.h"
66 /* possible states of the scheduler */
68 G_QUEUE_READY = 0, /* Ready to dispatch. */
69 G_QUEUE_BUSY, /* Waiting for a completion. */
70 G_QUEUE_IDLING /* Waiting for a new request. */
73 /* possible queue flags */
75 /* G_FLAG_COMPLETED means that the field q_slice_end is valid. */
76 G_FLAG_COMPLETED = 1, /* Completed a req. in the current budget. */
82 * Queue descriptor, containing reference count, scheduling
83 * state, a queue of pending requests, configuration parameters.
84 * Queues with pending request(s) and not under service are also
85 * stored in a Round Robin (RR) list.
88 struct g_rr_softc *q_sc; /* link to the parent */
90 enum g_rr_state q_status;
91 unsigned int q_service; /* service received so far */
92 int q_slice_end; /* actual slice end time, in ticks */
93 enum g_rr_flags q_flags; /* queue flags */
94 struct bio_queue_head q_bioq;
96 /* Scheduling parameters */
97 unsigned int q_budget; /* slice size in bytes */
98 unsigned int q_slice_duration; /* slice size in ticks */
99 unsigned int q_wait_ticks; /* wait time for anticipation */
101 /* Stats to drive the various heuristics. */
102 struct g_savg q_thinktime; /* Thinktime average. */
103 struct g_savg q_seekdist; /* Seek distance average. */
105 int q_bionum; /* Number of requests. */
107 off_t q_lastoff; /* Last submitted req. offset. */
108 int q_lastsub; /* Last submitted req. time. */
110 /* Expiration deadline for an empty queue. */
113 TAILQ_ENTRY(g_rr_queue) q_tailq; /* RR list link field */
117 TAILQ_HEAD(g_rr_tailq, g_rr_queue);
119 /* list of scheduler instances */
120 LIST_HEAD(g_scheds, g_rr_softc);
122 /* Default quantum for RR between queues. */
123 #define G_RR_DEFAULT_BUDGET 0x00800000
126 * Per device descriptor, holding the Round Robin list of queues
127 * accessing the disk, a reference to the geom, and the timer.
130 struct g_geom *sc_geom;
133 * sc_active is the queue we are anticipating for.
134 * It is set only in gs_rr_next(), and possibly cleared
135 * only in gs_rr_next() or on a timeout.
136 * The active queue is never in the Round Robin list
137 * even if it has requests queued.
139 struct g_rr_queue *sc_active;
140 struct callout sc_wait; /* timer for sc_active */
142 struct g_rr_tailq sc_rr_tailq; /* the round-robin list */
143 int sc_nqueues; /* number of queues */
146 int sc_in_flight; /* requests in the driver */
148 LIST_ENTRY(g_rr_softc) sc_next;
151 /* Descriptor for bounded values, min and max are constant. */
159 * parameters, config and stats
162 int queues; /* total number of queues */
163 int w_anticipate; /* anticipate writes */
164 int bypass; /* bypass scheduling writes */
166 int units; /* how many instances */
167 /* sc_head is used for debugging */
168 struct g_scheds sc_head; /* first scheduler instance */
170 struct x_bound queue_depth; /* max parallel requests */
171 struct x_bound wait_ms; /* wait time, milliseconds */
172 struct x_bound quantum_ms; /* quantum size, milliseconds */
173 struct x_bound quantum_kb; /* quantum size, Kb (1024 bytes) */
176 int wait_hit; /* success in anticipation */
177 int wait_miss; /* failure in anticipation */
181 * Default parameters for the scheduler. The quantum sizes target
182 * a 80MB/s disk; if the hw is faster or slower the minimum of the
183 * two will have effect: the clients will still be isolated but
184 * the fairness may be limited. A complete solution would involve
185 * the on-line measurement of the actual disk throughput to derive
186 * these parameters. Or we may just choose to ignore service domain
187 * fairness and accept what can be achieved with time-only budgets.
189 static struct g_rr_params me = {
190 .sc_head = LIST_HEAD_INITIALIZER(&me.sc_head),
192 .queue_depth = { 1, 1, 50 },
193 .wait_ms = { 1, 10, 30 },
194 .quantum_ms = { 1, 100, 500 },
195 .quantum_kb = { 16, 8192, 65536 },
198 struct g_rr_params *gs_rr_me = &me;
200 SYSCTL_DECL(_kern_geom_sched);
201 static SYSCTL_NODE(_kern_geom_sched, OID_AUTO, rr, CTLFLAG_RW, 0,
202 "GEOM_SCHED ROUND ROBIN stuff");
203 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, units, CTLFLAG_RD,
204 &me.units, 0, "Scheduler instances");
205 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, queues, CTLFLAG_RD,
206 &me.queues, 0, "Total rr queues");
207 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_ms, CTLFLAG_RW,
208 &me.wait_ms.x_cur, 0, "Wait time milliseconds");
209 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, quantum_ms, CTLFLAG_RW,
210 &me.quantum_ms.x_cur, 0, "Quantum size milliseconds");
211 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, bypass, CTLFLAG_RW,
212 &me.bypass, 0, "Bypass scheduler");
213 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, w_anticipate, CTLFLAG_RW,
214 &me.w_anticipate, 0, "Do anticipation on writes");
215 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, quantum_kb, CTLFLAG_RW,
216 &me.quantum_kb.x_cur, 0, "Quantum size Kbytes");
217 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, queue_depth, CTLFLAG_RW,
218 &me.queue_depth.x_cur, 0, "Maximum simultaneous requests");
219 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_hit, CTLFLAG_RW,
220 &me.wait_hit, 0, "Hits in anticipation");
221 SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_miss, CTLFLAG_RW,
222 &me.wait_miss, 0, "Misses in anticipation");
225 /* print the status of a queue */
227 gs_rr_dump_q(struct g_rr_queue *qp, int index)
232 TAILQ_FOREACH(bp, &(qp->q_bioq.queue), bio_queue) {
235 printf("--- rr queue %d %p status %d len %d ---\n",
236 index, qp, qp->q_status, l);
240 * Dump the scheduler status when writing to this sysctl variable.
241 * XXX right now we only dump the status of the last instance created.
242 * not a severe issue because this is only for debugging
245 gs_rr_sysctl_status(SYSCTL_HANDLER_ARGS)
248 struct g_rr_softc *sc;
250 error = sysctl_handle_int(oidp, &val, 0, req);
251 if (error || !req->newptr )
254 printf("called %s\n", __FUNCTION__);
256 LIST_FOREACH(sc, &me.sc_head, sc_next) {
258 printf("--- sc %p active %p nqueues %d "
259 "callout %d in_flight %d ---\n",
260 sc, sc->sc_active, sc->sc_nqueues,
261 callout_active(&sc->sc_wait),
263 for (i = 0; i < G_RR_HASH_SIZE; i++) {
264 struct g_rr_queue *qp;
265 LIST_FOREACH(qp, &sc->sc_hash[i], q_hash) {
266 gs_rr_dump_q(qp, tot);
274 SYSCTL_PROC(_kern_geom_sched_rr, OID_AUTO, status,
275 CTLTYPE_UINT | CTLFLAG_RW,
276 0, sizeof(int), gs_rr_sysctl_status, "I", "status");
278 #endif /* DEBUG_QUEUES */
281 * Get a bounded value, optionally convert to a min of t_min ticks.
284 get_bounded(struct x_bound *v, int t_min)
291 else if (x > v->x_max)
294 x = x * hz / 1000; /* convert to ticks */
302 * Get a reference to the queue for bp, using the generic
303 * classification mechanism.
305 static struct g_rr_queue *
306 g_rr_queue_get(struct g_rr_softc *sc, struct bio *bp)
309 return (g_sched_get_class(sc->sc_geom, bp));
313 g_rr_init_class(void *data, void *priv)
315 struct g_rr_softc *sc = data;
316 struct g_rr_queue *qp = priv;
318 gs_bioq_init(&qp->q_bioq);
321 * Set the initial parameters for the client:
322 * slice size in bytes and ticks, and wait ticks.
323 * Right now these are constant, but we could have
324 * autoconfiguration code to adjust the values based on
325 * the actual workload.
327 qp->q_budget = 1024 * get_bounded(&me.quantum_kb, 0);
328 qp->q_slice_duration = get_bounded(&me.quantum_ms, 2);
329 qp->q_wait_ticks = get_bounded(&me.wait_ms, 2);
331 qp->q_sc = sc; /* link to the parent */
332 qp->q_sc->sc_nqueues++;
339 * Release a reference to the queue.
342 g_rr_queue_put(struct g_rr_queue *qp)
345 g_sched_put_class(qp->q_sc->sc_geom, qp);
349 g_rr_fini_class(void *data, void *priv)
351 struct g_rr_queue *qp = priv;
353 KASSERT(gs_bioq_first(&qp->q_bioq) == NULL,
354 ("released nonempty queue"));
355 qp->q_sc->sc_nqueues--;
360 g_rr_queue_expired(struct g_rr_queue *qp)
363 if (qp->q_service >= qp->q_budget)
366 if ((qp->q_flags & G_FLAG_COMPLETED) &&
367 ticks - qp->q_slice_end >= 0)
374 g_rr_should_anticipate(struct g_rr_queue *qp, struct bio *bp)
376 int wait = get_bounded(&me.wait_ms, 2);
378 if (!me.w_anticipate && (bp->bio_cmd & BIO_WRITE))
381 if (g_savg_valid(&qp->q_thinktime) &&
382 g_savg_read(&qp->q_thinktime) > wait)
385 if (g_savg_valid(&qp->q_seekdist) &&
386 g_savg_read(&qp->q_seekdist) > 8192)
393 * Called on a request arrival, timeout or completion.
394 * Try to serve a request among those queued.
397 g_rr_next(void *data, int force)
399 struct g_rr_softc *sc = data;
400 struct g_rr_queue *qp;
401 struct bio *bp, *next;
405 if (me.bypass == 0 && !force) {
406 if (sc->sc_in_flight >= get_bounded(&me.queue_depth, 0))
409 /* Try with the queue under service first. */
410 if (qp != NULL && qp->q_status != G_QUEUE_READY) {
412 * Queue is anticipating, ignore request.
413 * We should check that we are not past
414 * the timeout, but in that case the timeout
415 * will fire immediately afterwards so we
420 } else if (qp != NULL && qp->q_status != G_QUEUE_READY) {
422 sc->sc_active = qp = NULL;
426 * No queue under service, look for the first in RR order.
427 * If we find it, select if as sc_active, clear service
428 * and record the end time of the slice.
431 qp = TAILQ_FIRST(&sc->sc_rr_tailq);
433 return (NULL); /* no queues at all, return */
434 /* otherwise select the new queue for service. */
435 TAILQ_REMOVE(&sc->sc_rr_tailq, qp, q_tailq);
438 qp->q_flags &= ~G_FLAG_COMPLETED;
441 bp = gs_bioq_takefirst(&qp->q_bioq); /* surely not NULL */
442 qp->q_service += bp->bio_length; /* charge the service */
445 * The request at the head of the active queue is always
446 * dispatched, and gs_rr_next() will be called again
448 * We need to prepare for what to do next:
450 * 1. have we reached the end of the (time or service) slice ?
451 * If so, clear sc_active and possibly requeue the previous
452 * active queue if it has more requests pending;
453 * 2. do we have more requests in sc_active ?
454 * If yes, do not anticipate, as gs_rr_next() will run again;
455 * if no, decide whether or not to anticipate depending
456 * on read or writes (e.g., anticipate only on reads).
458 expired = g_rr_queue_expired(qp); /* are we expired ? */
459 next = gs_bioq_first(&qp->q_bioq); /* do we have one more ? */
461 sc->sc_active = NULL;
462 /* Either requeue or release reference. */
464 TAILQ_INSERT_TAIL(&sc->sc_rr_tailq, qp, q_tailq);
467 } else if (next != NULL) {
468 qp->q_status = G_QUEUE_READY;
470 if (!force && g_rr_should_anticipate(qp, bp)) {
472 qp->q_status = G_QUEUE_BUSY;
474 /* do not anticipate, release reference */
476 sc->sc_active = NULL;
479 /* If sc_active != NULL, its q_status is always correct. */
487 g_rr_update_thinktime(struct g_rr_queue *qp)
489 int delta = ticks - qp->q_lastsub, wait = get_bounded(&me.wait_ms, 2);
491 if (qp->q_sc->sc_active != qp)
494 qp->q_lastsub = ticks;
495 delta = (delta > 2 * wait) ? 2 * wait : delta;
496 if (qp->q_bionum > 7)
497 g_savg_add_sample(&qp->q_thinktime, delta);
501 g_rr_update_seekdist(struct g_rr_queue *qp, struct bio *bp)
505 if (qp->q_lastoff > bp->bio_offset)
506 dist = qp->q_lastoff - bp->bio_offset;
508 dist = bp->bio_offset - qp->q_lastoff;
510 if (dist > (8192 * 8))
513 qp->q_lastoff = bp->bio_offset + bp->bio_length;
515 if (qp->q_bionum > 7)
516 g_savg_add_sample(&qp->q_seekdist, dist);
520 * Called when a real request for disk I/O arrives.
521 * Locate the queue associated with the client.
522 * If the queue is the one we are anticipating for, reset its timeout;
523 * if the queue is not in the round robin list, insert it in the list.
524 * On any error, do not queue the request and return -1, the caller
525 * will take care of this request.
528 g_rr_start(void *data, struct bio *bp)
530 struct g_rr_softc *sc = data;
531 struct g_rr_queue *qp;
534 return (-1); /* bypass the scheduler */
536 /* Get the queue for the request. */
537 qp = g_rr_queue_get(sc, bp);
539 return (-1); /* allocation failed, tell upstream */
541 if (gs_bioq_first(&qp->q_bioq) == NULL) {
543 * We are inserting into an empty queue.
544 * Reset its state if it is sc_active,
545 * otherwise insert it in the RR list.
547 if (qp == sc->sc_active) {
548 qp->q_status = G_QUEUE_READY;
549 callout_stop(&sc->sc_wait);
551 g_sched_priv_ref(qp);
552 TAILQ_INSERT_TAIL(&sc->sc_rr_tailq, qp, q_tailq);
556 qp->q_bionum = 1 + qp->q_bionum - (qp->q_bionum >> 3);
558 g_rr_update_thinktime(qp);
559 g_rr_update_seekdist(qp, bp);
561 /* Inherit the reference returned by g_rr_queue_get(). */
562 bp->bio_caller1 = qp;
563 gs_bioq_disksort(&qp->q_bioq, bp);
569 * Callout executed when a queue times out anticipating a new request.
572 g_rr_wait_timeout(void *data)
574 struct g_rr_softc *sc = data;
575 struct g_geom *geom = sc->sc_geom;
579 * We can race with other events, so check if
580 * sc_active is still valid.
582 if (sc->sc_active != NULL) {
583 /* Release the reference to the queue. */
584 g_rr_queue_put(sc->sc_active);
585 sc->sc_active = NULL;
587 me.wait_miss++; /* record the miss */
589 g_sched_dispatch(geom);
590 g_sched_unlock(geom);
594 * Module glue: allocate descriptor, initialize its fields.
597 g_rr_init(struct g_geom *geom)
599 struct g_rr_softc *sc;
601 /* XXX check whether we can sleep */
602 sc = malloc(sizeof *sc, M_GEOM_SCHED, M_NOWAIT | M_ZERO);
604 TAILQ_INIT(&sc->sc_rr_tailq);
605 callout_init(&sc->sc_wait, CALLOUT_MPSAFE);
606 LIST_INSERT_HEAD(&me.sc_head, sc, sc_next);
613 * Module glue -- drain the callout structure, destroy the
614 * hash table and its element, and free the descriptor.
617 g_rr_fini(void *data)
619 struct g_rr_softc *sc = data;
621 callout_drain(&sc->sc_wait);
622 KASSERT(sc->sc_active == NULL, ("still a queue under service"));
623 KASSERT(TAILQ_EMPTY(&sc->sc_rr_tailq), ("still scheduled queues"));
625 LIST_REMOVE(sc, sc_next);
627 free(sc, M_GEOM_SCHED);
631 * Called when the request under service terminates.
632 * Start the anticipation timer if needed.
635 g_rr_done(void *data, struct bio *bp)
637 struct g_rr_softc *sc = data;
638 struct g_rr_queue *qp;
642 qp = bp->bio_caller1;
645 * When the first request for this queue completes, update the
646 * duration and end of the slice. We do not do it when the
647 * slice starts to avoid charging to the queue the time for
650 if (!(qp->q_flags & G_FLAG_COMPLETED)) {
651 qp->q_flags |= G_FLAG_COMPLETED;
653 * recompute the slice duration, in case we want
654 * to make it adaptive. This is not used right now.
655 * XXX should we do the same for q_quantum and q_wait_ticks ?
657 qp->q_slice_duration = get_bounded(&me.quantum_ms, 2);
658 qp->q_slice_end = ticks + qp->q_slice_duration;
661 if (qp == sc->sc_active && qp->q_status == G_QUEUE_BUSY) {
662 /* The queue is trying anticipation, start the timer. */
663 qp->q_status = G_QUEUE_IDLING;
664 /* may make this adaptive */
665 qp->q_wait_ticks = get_bounded(&me.wait_ms, 2);
667 callout_reset(&sc->sc_wait, qp->q_wait_ticks,
668 g_rr_wait_timeout, sc);
670 g_sched_dispatch(sc->sc_geom);
672 /* Release a reference to the queue. */
677 g_rr_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
678 struct g_consumer *cp, struct g_provider *pp)
680 if (indent == NULL) { /* plaintext */
681 sbuf_printf(sb, " units %d queues %d",
682 me.units, me.queues);
686 static struct g_gsched g_rr = {
688 .gs_priv_size = sizeof(struct g_rr_queue),
689 .gs_init = g_rr_init,
690 .gs_fini = g_rr_fini,
691 .gs_start = g_rr_start,
692 .gs_done = g_rr_done,
693 .gs_next = g_rr_next,
694 .gs_dumpconf = g_rr_dumpconf,
695 .gs_init_class = g_rr_init_class,
696 .gs_fini_class = g_rr_fini_class,
699 DECLARE_GSCHED_MODULE(rr, &g_rr);