freebsd-skq/sys/geom/sched/gs_rr.c
luigi 301c433ed1 Correct a subtle bug in the 'gsched_rr' disk scheduler.
The algorithm is supposed to work as follows:
in order to prevent starvation, when a new client starts being served we
record the start time and reset the counter of bytes served.
We then switch to a new client after a certain amount of time or bytes,
even if the current one still has pending requests.
To avoid charging a new client the time of the first seek,
we start counting time when the first request is served.

Unfortunately a bug in the previous version of the code failed
to set the start time in certain cases, resulting in some processes
exceeding their timeslice.

The fix (in this patch) is trivial, though it took a while to find
out and replicate the bug.
Thanks to Tommaso Caprai for investigating and fixing the problem.

Submitted by:	Tommaso Caprai
MFC after:	1 week
2011-02-14 08:09:02 +00:00

699 lines
19 KiB
C

/*-
* Copyright (c) 2009-2010 Fabio Checconi
* Copyright (c) 2009-2010 Luigi Rizzo, Universita` di Pisa
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
/*
* $Id$
* $FreeBSD$
*
* A round-robin (RR) anticipatory scheduler, with per-client queues.
*
* The goal of this implementation is to improve throughput compared
* to the pure elevator algorithm, and insure some fairness among
* clients.
*
* Requests coming from the same client are put in the same queue.
* We use anticipation to help reducing seeks, and each queue
* is never served continuously for more than a given amount of
* time or data. Queues are then served in a round-robin fashion.
*
* Each queue can be in any of the following states:
* READY immediately serve the first pending request;
* BUSY one request is under service, wait for completion;
* IDLING do not serve incoming requests immediately, unless
* they are "eligible" as defined later.
*
* Scheduling is made looking at the status of all queues,
* and the first one in round-robin order is privileged.
*/
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/bio.h>
#include <sys/callout.h>
#include <sys/malloc.h>
#include <sys/module.h>
#include <sys/proc.h>
#include <sys/queue.h>
#include <sys/sysctl.h>
#include "gs_scheduler.h"
/* possible states of the scheduler */
enum g_rr_state {
G_QUEUE_READY = 0, /* Ready to dispatch. */
G_QUEUE_BUSY, /* Waiting for a completion. */
G_QUEUE_IDLING /* Waiting for a new request. */
};
/* possible queue flags */
enum g_rr_flags {
/* G_FLAG_COMPLETED means that the field q_slice_end is valid. */
G_FLAG_COMPLETED = 1, /* Completed a req. in the current budget. */
};
struct g_rr_softc;
/*
* Queue descriptor, containing reference count, scheduling
* state, a queue of pending requests, configuration parameters.
* Queues with pending request(s) and not under service are also
* stored in a Round Robin (RR) list.
*/
struct g_rr_queue {
struct g_rr_softc *q_sc; /* link to the parent */
enum g_rr_state q_status;
unsigned int q_service; /* service received so far */
int q_slice_end; /* actual slice end time, in ticks */
enum g_rr_flags q_flags; /* queue flags */
struct bio_queue_head q_bioq;
/* Scheduling parameters */
unsigned int q_budget; /* slice size in bytes */
unsigned int q_slice_duration; /* slice size in ticks */
unsigned int q_wait_ticks; /* wait time for anticipation */
/* Stats to drive the various heuristics. */
struct g_savg q_thinktime; /* Thinktime average. */
struct g_savg q_seekdist; /* Seek distance average. */
int q_bionum; /* Number of requests. */
off_t q_lastoff; /* Last submitted req. offset. */
int q_lastsub; /* Last submitted req. time. */
/* Expiration deadline for an empty queue. */
int q_expire;
TAILQ_ENTRY(g_rr_queue) q_tailq; /* RR list link field */
};
/* List types. */
TAILQ_HEAD(g_rr_tailq, g_rr_queue);
/* list of scheduler instances */
LIST_HEAD(g_scheds, g_rr_softc);
/* Default quantum for RR between queues. */
#define G_RR_DEFAULT_BUDGET 0x00800000
/*
* Per device descriptor, holding the Round Robin list of queues
* accessing the disk, a reference to the geom, and the timer.
*/
struct g_rr_softc {
struct g_geom *sc_geom;
/*
* sc_active is the queue we are anticipating for.
* It is set only in gs_rr_next(), and possibly cleared
* only in gs_rr_next() or on a timeout.
* The active queue is never in the Round Robin list
* even if it has requests queued.
*/
struct g_rr_queue *sc_active;
struct callout sc_wait; /* timer for sc_active */
struct g_rr_tailq sc_rr_tailq; /* the round-robin list */
int sc_nqueues; /* number of queues */
/* Statistics */
int sc_in_flight; /* requests in the driver */
LIST_ENTRY(g_rr_softc) sc_next;
};
/* Descriptor for bounded values, min and max are constant. */
struct x_bound {
const int x_min;
int x_cur;
const int x_max;
};
/*
* parameters, config and stats
*/
struct g_rr_params {
int queues; /* total number of queues */
int w_anticipate; /* anticipate writes */
int bypass; /* bypass scheduling writes */
int units; /* how many instances */
/* sc_head is used for debugging */
struct g_scheds sc_head; /* first scheduler instance */
struct x_bound queue_depth; /* max parallel requests */
struct x_bound wait_ms; /* wait time, milliseconds */
struct x_bound quantum_ms; /* quantum size, milliseconds */
struct x_bound quantum_kb; /* quantum size, Kb (1024 bytes) */
/* statistics */
int wait_hit; /* success in anticipation */
int wait_miss; /* failure in anticipation */
};
/*
* Default parameters for the scheduler. The quantum sizes target
* a 80MB/s disk; if the hw is faster or slower the minimum of the
* two will have effect: the clients will still be isolated but
* the fairness may be limited. A complete solution would involve
* the on-line measurement of the actual disk throughput to derive
* these parameters. Or we may just choose to ignore service domain
* fairness and accept what can be achieved with time-only budgets.
*/
static struct g_rr_params me = {
.sc_head = LIST_HEAD_INITIALIZER(&me.sc_head),
.w_anticipate = 1,
.queue_depth = { 1, 1, 50 },
.wait_ms = { 1, 10, 30 },
.quantum_ms = { 1, 100, 500 },
.quantum_kb = { 16, 8192, 65536 },
};
struct g_rr_params *gs_rr_me = &me;
SYSCTL_DECL(_kern_geom_sched);
SYSCTL_NODE(_kern_geom_sched, OID_AUTO, rr, CTLFLAG_RW, 0,
"GEOM_SCHED ROUND ROBIN stuff");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, units, CTLFLAG_RD,
&me.units, 0, "Scheduler instances");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, queues, CTLFLAG_RD,
&me.queues, 0, "Total rr queues");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_ms, CTLFLAG_RW,
&me.wait_ms.x_cur, 0, "Wait time milliseconds");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, quantum_ms, CTLFLAG_RW,
&me.quantum_ms.x_cur, 0, "Quantum size milliseconds");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, bypass, CTLFLAG_RW,
&me.bypass, 0, "Bypass scheduler");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, w_anticipate, CTLFLAG_RW,
&me.w_anticipate, 0, "Do anticipation on writes");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, quantum_kb, CTLFLAG_RW,
&me.quantum_kb.x_cur, 0, "Quantum size Kbytes");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, queue_depth, CTLFLAG_RW,
&me.queue_depth.x_cur, 0, "Maximum simultaneous requests");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_hit, CTLFLAG_RW,
&me.wait_hit, 0, "Hits in anticipation");
SYSCTL_INT(_kern_geom_sched_rr, OID_AUTO, wait_miss, CTLFLAG_RW,
&me.wait_miss, 0, "Misses in anticipation");
#ifdef DEBUG_QUEUES
/* print the status of a queue */
static void
gs_rr_dump_q(struct g_rr_queue *qp, int index)
{
int l = 0;
struct bio *bp;
TAILQ_FOREACH(bp, &(qp->q_bioq.queue), bio_queue) {
l++;
}
printf("--- rr queue %d %p status %d len %d ---\n",
index, qp, qp->q_status, l);
}
/*
* Dump the scheduler status when writing to this sysctl variable.
* XXX right now we only dump the status of the last instance created.
* not a severe issue because this is only for debugging
*/
static int
gs_rr_sysctl_status(SYSCTL_HANDLER_ARGS)
{
int error, val = 0;
struct g_rr_softc *sc;
error = sysctl_handle_int(oidp, &val, 0, req);
if (error || !req->newptr )
return (error);
printf("called %s\n", __FUNCTION__);
LIST_FOREACH(sc, &me.sc_head, sc_next) {
int i, tot = 0;
printf("--- sc %p active %p nqueues %d "
"callout %d in_flight %d ---\n",
sc, sc->sc_active, sc->sc_nqueues,
callout_active(&sc->sc_wait),
sc->sc_in_flight);
for (i = 0; i < G_RR_HASH_SIZE; i++) {
struct g_rr_queue *qp;
LIST_FOREACH(qp, &sc->sc_hash[i], q_hash) {
gs_rr_dump_q(qp, tot);
tot++;
}
}
}
return (0);
}
SYSCTL_PROC(_kern_geom_sched_rr, OID_AUTO, status,
CTLTYPE_UINT | CTLFLAG_RW,
0, sizeof(int), gs_rr_sysctl_status, "I", "status");
#endif /* DEBUG_QUEUES */
/*
* Get a bounded value, optionally convert to a min of t_min ticks.
*/
static int
get_bounded(struct x_bound *v, int t_min)
{
int x;
x = v->x_cur;
if (x < v->x_min)
x = v->x_min;
else if (x > v->x_max)
x = v->x_max;
if (t_min) {
x = x * hz / 1000; /* convert to ticks */
if (x < t_min)
x = t_min;
}
return x;
}
/*
* Get a reference to the queue for bp, using the generic
* classification mechanism.
*/
static struct g_rr_queue *
g_rr_queue_get(struct g_rr_softc *sc, struct bio *bp)
{
return (g_sched_get_class(sc->sc_geom, bp));
}
static int
g_rr_init_class(void *data, void *priv)
{
struct g_rr_softc *sc = data;
struct g_rr_queue *qp = priv;
gs_bioq_init(&qp->q_bioq);
/*
* Set the initial parameters for the client:
* slice size in bytes and ticks, and wait ticks.
* Right now these are constant, but we could have
* autoconfiguration code to adjust the values based on
* the actual workload.
*/
qp->q_budget = 1024 * get_bounded(&me.quantum_kb, 0);
qp->q_slice_duration = get_bounded(&me.quantum_ms, 2);
qp->q_wait_ticks = get_bounded(&me.wait_ms, 2);
qp->q_sc = sc; /* link to the parent */
qp->q_sc->sc_nqueues++;
me.queues++;
return (0);
}
/*
* Release a reference to the queue.
*/
static void
g_rr_queue_put(struct g_rr_queue *qp)
{
g_sched_put_class(qp->q_sc->sc_geom, qp);
}
static void
g_rr_fini_class(void *data, void *priv)
{
struct g_rr_queue *qp = priv;
KASSERT(gs_bioq_first(&qp->q_bioq) == NULL,
("released nonempty queue"));
qp->q_sc->sc_nqueues--;
me.queues--;
}
static inline int
g_rr_queue_expired(struct g_rr_queue *qp)
{
if (qp->q_service >= qp->q_budget)
return (1);
if ((qp->q_flags & G_FLAG_COMPLETED) &&
ticks - qp->q_slice_end >= 0)
return (1);
return (0);
}
static inline int
g_rr_should_anticipate(struct g_rr_queue *qp, struct bio *bp)
{
int wait = get_bounded(&me.wait_ms, 2);
if (!me.w_anticipate && (bp->bio_cmd & BIO_WRITE))
return (0);
if (g_savg_valid(&qp->q_thinktime) &&
g_savg_read(&qp->q_thinktime) > wait)
return (0);
if (g_savg_valid(&qp->q_seekdist) &&
g_savg_read(&qp->q_seekdist) > 8192)
return (0);
return (1);
}
/*
* Called on a request arrival, timeout or completion.
* Try to serve a request among those queued.
*/
static struct bio *
g_rr_next(void *data, int force)
{
struct g_rr_softc *sc = data;
struct g_rr_queue *qp;
struct bio *bp, *next;
int expired;
qp = sc->sc_active;
if (me.bypass == 0 && !force) {
if (sc->sc_in_flight >= get_bounded(&me.queue_depth, 0))
return (NULL);
/* Try with the queue under service first. */
if (qp != NULL && qp->q_status != G_QUEUE_READY) {
/*
* Queue is anticipating, ignore request.
* We should check that we are not past
* the timeout, but in that case the timeout
* will fire immediately afterwards so we
* don't bother.
*/
return (NULL);
}
} else if (qp != NULL && qp->q_status != G_QUEUE_READY) {
g_rr_queue_put(qp);
sc->sc_active = qp = NULL;
}
/*
* No queue under service, look for the first in RR order.
* If we find it, select if as sc_active, clear service
* and record the end time of the slice.
*/
if (qp == NULL) {
qp = TAILQ_FIRST(&sc->sc_rr_tailq);
if (qp == NULL)
return (NULL); /* no queues at all, return */
/* otherwise select the new queue for service. */
TAILQ_REMOVE(&sc->sc_rr_tailq, qp, q_tailq);
sc->sc_active = qp;
qp->q_service = 0;
qp->q_flags &= ~G_FLAG_COMPLETED;
}
bp = gs_bioq_takefirst(&qp->q_bioq); /* surely not NULL */
qp->q_service += bp->bio_length; /* charge the service */
/*
* The request at the head of the active queue is always
* dispatched, and gs_rr_next() will be called again
* immediately.
* We need to prepare for what to do next:
*
* 1. have we reached the end of the (time or service) slice ?
* If so, clear sc_active and possibly requeue the previous
* active queue if it has more requests pending;
* 2. do we have more requests in sc_active ?
* If yes, do not anticipate, as gs_rr_next() will run again;
* if no, decide whether or not to anticipate depending
* on read or writes (e.g., anticipate only on reads).
*/
expired = g_rr_queue_expired(qp); /* are we expired ? */
next = gs_bioq_first(&qp->q_bioq); /* do we have one more ? */
if (expired) {
sc->sc_active = NULL;
/* Either requeue or release reference. */
if (next != NULL)
TAILQ_INSERT_TAIL(&sc->sc_rr_tailq, qp, q_tailq);
else
g_rr_queue_put(qp);
} else if (next != NULL) {
qp->q_status = G_QUEUE_READY;
} else {
if (!force && g_rr_should_anticipate(qp, bp)) {
/* anticipate */
qp->q_status = G_QUEUE_BUSY;
} else {
/* do not anticipate, release reference */
g_rr_queue_put(qp);
sc->sc_active = NULL;
}
}
/* If sc_active != NULL, its q_status is always correct. */
sc->sc_in_flight++;
return (bp);
}
static inline void
g_rr_update_thinktime(struct g_rr_queue *qp)
{
int delta = ticks - qp->q_lastsub, wait = get_bounded(&me.wait_ms, 2);
if (qp->q_sc->sc_active != qp)
return;
qp->q_lastsub = ticks;
delta = (delta > 2 * wait) ? 2 * wait : delta;
if (qp->q_bionum > 7)
g_savg_add_sample(&qp->q_thinktime, delta);
}
static inline void
g_rr_update_seekdist(struct g_rr_queue *qp, struct bio *bp)
{
off_t dist;
if (qp->q_lastoff > bp->bio_offset)
dist = qp->q_lastoff - bp->bio_offset;
else
dist = bp->bio_offset - qp->q_lastoff;
if (dist > (8192 * 8))
dist = 8192 * 8;
qp->q_lastoff = bp->bio_offset + bp->bio_length;
if (qp->q_bionum > 7)
g_savg_add_sample(&qp->q_seekdist, dist);
}
/*
* Called when a real request for disk I/O arrives.
* Locate the queue associated with the client.
* If the queue is the one we are anticipating for, reset its timeout;
* if the queue is not in the round robin list, insert it in the list.
* On any error, do not queue the request and return -1, the caller
* will take care of this request.
*/
static int
g_rr_start(void *data, struct bio *bp)
{
struct g_rr_softc *sc = data;
struct g_rr_queue *qp;
if (me.bypass)
return (-1); /* bypass the scheduler */
/* Get the queue for the request. */
qp = g_rr_queue_get(sc, bp);
if (qp == NULL)
return (-1); /* allocation failed, tell upstream */
if (gs_bioq_first(&qp->q_bioq) == NULL) {
/*
* We are inserting into an empty queue.
* Reset its state if it is sc_active,
* otherwise insert it in the RR list.
*/
if (qp == sc->sc_active) {
qp->q_status = G_QUEUE_READY;
callout_stop(&sc->sc_wait);
} else {
g_sched_priv_ref(qp);
TAILQ_INSERT_TAIL(&sc->sc_rr_tailq, qp, q_tailq);
}
}
qp->q_bionum = 1 + qp->q_bionum - (qp->q_bionum >> 3);
g_rr_update_thinktime(qp);
g_rr_update_seekdist(qp, bp);
/* Inherit the reference returned by g_rr_queue_get(). */
bp->bio_caller1 = qp;
gs_bioq_disksort(&qp->q_bioq, bp);
return (0);
}
/*
* Callout executed when a queue times out anticipating a new request.
*/
static void
g_rr_wait_timeout(void *data)
{
struct g_rr_softc *sc = data;
struct g_geom *geom = sc->sc_geom;
g_sched_lock(geom);
/*
* We can race with other events, so check if
* sc_active is still valid.
*/
if (sc->sc_active != NULL) {
/* Release the reference to the queue. */
g_rr_queue_put(sc->sc_active);
sc->sc_active = NULL;
me.wait_hit--;
me.wait_miss++; /* record the miss */
}
g_sched_dispatch(geom);
g_sched_unlock(geom);
}
/*
* Module glue: allocate descriptor, initialize its fields.
*/
static void *
g_rr_init(struct g_geom *geom)
{
struct g_rr_softc *sc;
/* XXX check whether we can sleep */
sc = malloc(sizeof *sc, M_GEOM_SCHED, M_NOWAIT | M_ZERO);
sc->sc_geom = geom;
TAILQ_INIT(&sc->sc_rr_tailq);
callout_init(&sc->sc_wait, CALLOUT_MPSAFE);
LIST_INSERT_HEAD(&me.sc_head, sc, sc_next);
me.units++;
return (sc);
}
/*
* Module glue -- drain the callout structure, destroy the
* hash table and its element, and free the descriptor.
*/
static void
g_rr_fini(void *data)
{
struct g_rr_softc *sc = data;
callout_drain(&sc->sc_wait);
KASSERT(sc->sc_active == NULL, ("still a queue under service"));
KASSERT(TAILQ_EMPTY(&sc->sc_rr_tailq), ("still scheduled queues"));
LIST_REMOVE(sc, sc_next);
me.units--;
free(sc, M_GEOM_SCHED);
}
/*
* Called when the request under service terminates.
* Start the anticipation timer if needed.
*/
static void
g_rr_done(void *data, struct bio *bp)
{
struct g_rr_softc *sc = data;
struct g_rr_queue *qp;
sc->sc_in_flight--;
qp = bp->bio_caller1;
/*
* When the first request for this queue completes, update the
* duration and end of the slice. We do not do it when the
* slice starts to avoid charging to the queue the time for
* the first seek.
*/
if (!(qp->q_flags & G_FLAG_COMPLETED)) {
qp->q_flags |= G_FLAG_COMPLETED;
/*
* recompute the slice duration, in case we want
* to make it adaptive. This is not used right now.
* XXX should we do the same for q_quantum and q_wait_ticks ?
*/
qp->q_slice_duration = get_bounded(&me.quantum_ms, 2);
qp->q_slice_end = ticks + qp->q_slice_duration;
}
if (qp == sc->sc_active && qp->q_status == G_QUEUE_BUSY) {
/* The queue is trying anticipation, start the timer. */
qp->q_status = G_QUEUE_IDLING;
/* may make this adaptive */
qp->q_wait_ticks = get_bounded(&me.wait_ms, 2);
me.wait_hit++;
callout_reset(&sc->sc_wait, qp->q_wait_ticks,
g_rr_wait_timeout, sc);
} else
g_sched_dispatch(sc->sc_geom);
/* Release a reference to the queue. */
g_rr_queue_put(qp);
}
static void
g_rr_dumpconf(struct sbuf *sb, const char *indent, struct g_geom *gp,
struct g_consumer *cp, struct g_provider *pp)
{
if (indent == NULL) { /* plaintext */
sbuf_printf(sb, " units %d queues %d",
me.units, me.queues);
}
}
static struct g_gsched g_rr = {
.gs_name = "rr",
.gs_priv_size = sizeof(struct g_rr_queue),
.gs_init = g_rr_init,
.gs_fini = g_rr_fini,
.gs_start = g_rr_start,
.gs_done = g_rr_done,
.gs_next = g_rr_next,
.gs_dumpconf = g_rr_dumpconf,
.gs_init_class = g_rr_init_class,
.gs_fini_class = g_rr_fini_class,
};
DECLARE_GSCHED_MODULE(rr, &g_rr);