d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

MFC geom_sched code, a geom-based disk scheduling framework.

Browse Source
This commit is contained in:
luigi 2010-04-20 15:23:12 +00:00
parent 3849eb12d4
commit be47c154d1
15 changed files with 3670 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
sbin/geom/class/Makefile
View File

@ -14,6 +14,7 @@ SUBDIR+=multipath
SUBDIR+=nop
SUBDIR+=part
SUBDIR+=raid3
SUBDIR+=sched
SUBDIR+=shsec
SUBDIR+=stripe
SUBDIR+=virstor

18
sbin/geom/class/sched/Makefile Normal file
View File

@ -0,0 +1,18 @@
# GEOM_LIBRARY_PATH
# $FreeBSD$
.PATH: ${.CURDIR}/../../misc
#CFLAGS += -I/usr/src/sbin/geom
CLASS=sched
WARNS?= 6
CLASS_DIR?=/lib/geom
SHLIBDIR?=${CLASS_DIR}
SHLIB_NAME?=geom_${CLASS}.so
LINKS= ${BINDIR}/geom ${BINDIR}/g${CLASS}
MAN= g${CLASS}.8
SRCS+= geom_${CLASS}.c subr.c
.include <bsd.lib.mk>

124
sbin/geom/class/sched/geom_sched.c Normal file
View File

@ -0,0 +1,124 @@
/*-
* Copyright (c) 2009 Fabio Checconi
* Copyright (c) 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 AUTHORS 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 AUTHORS 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.
*/
/*
* $Id$
* $FreeBSD$
*
* This file implements the userspace library used by the 'geom'
* command to load and manipulate disk schedulers.
*/
#include <sys/cdefs.h>
#include <sys/param.h>
#include <sys/linker.h>
#include <sys/module.h>
#include <stdio.h>
#include <stdint.h>
#include <libgeom.h>
#include "core/geom.h"
#include "misc/subr.h"
#define G_SCHED_VERSION 0
uint32_t lib_version = G_LIB_VERSION;
uint32_t version = G_SCHED_VERSION;
/*
* storage for parameters used by this geom class.
* Right now only the scheduler name is used.
*/
static char algo[] = "rr"; /* default scheduler */
/*
* Adapt to differences in geom library.
* in V1 struct g_command misses gc_argname, eld, and G_BOOL is undefined
*/
#if G_LIB_VERSION == 1
#define G_ARGNAME
#define G_TYPE_BOOL G_TYPE_NUMBER
#else
#define G_ARGNAME NULL,
#endif
static void
gcmd_createinsert(struct gctl_req *req, unsigned flags __unused)
{
const char *reqalgo;
char name[64];
if (gctl_has_param(req, "algo"))
reqalgo = gctl_get_ascii(req, "algo");
else
reqalgo = algo;
snprintf(name, sizeof(name), "gsched_%s", reqalgo);
/*
* Do not complain about errors here, gctl_issue()
* will fail anyway.
*/
if (modfind(name) < 0)
kldload(name);
gctl_issue(req);
}
struct g_command class_commands[] = {
{ "create", G_FLAG_VERBOSE | G_FLAG_LOADKLD, gcmd_createinsert,
{
{ 'a', "algo", algo, G_TYPE_STRING },
G_OPT_SENTINEL
},
G_ARGNAME "[-v] [-a algorithm_name] dev ..."
},
{ "insert", G_FLAG_VERBOSE | G_FLAG_LOADKLD, gcmd_createinsert,
{
{ 'a', "algo", algo, G_TYPE_STRING },
G_OPT_SENTINEL
},
G_ARGNAME "[-v] [-a algorithm_name] dev ..."
},
{ "configure", G_FLAG_VERBOSE, NULL,
{
{ 'a', "algo", algo, G_TYPE_STRING },
G_OPT_SENTINEL
},
G_ARGNAME "[-v] [-a algorithm_name] prov ..."
},
{ "destroy", G_FLAG_VERBOSE, NULL,
{
{ 'f', "force", NULL, G_TYPE_BOOL },
G_OPT_SENTINEL
},
G_ARGNAME "[-fv] prov ..."
},
{ "reset", G_FLAG_VERBOSE, NULL, G_NULL_OPTS,
G_ARGNAME "[-v] prov ..."
},
G_CMD_SENTINEL
};

163
sbin/geom/class/sched/gsched.8 Normal file
View File

@ -0,0 +1,163 @@
.\" 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 AUTHORS 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 AUTHORS 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.
.\"
.\" $FreeBSD$
.\"
.Dd April 12, 2010
.Dt GSCHED 8
.Os
.Sh NAME
.Nm gsched
.Nd "control utility for disk scheduler GEOM class"
.Sh SYNOPSIS
.Nm
.Cm create
.Op Fl v
.Op Fl a Ar algorithm
.Ar provider ...
.Nm
.Cm insert
.Op Fl v
.Op Fl a Ar algorithm
.Ar provider ...
.Nm
.Cm configure
.Op Fl v
.Op Fl a Ar algorithm
.Ar node ...
.Nm
.Cm destroy
.Op Fl fv
.Ar node ...
.Nm
.Cm reset
.Op Fl v
.Ar node ...
.Nm
.Cm { list | status | load | unload }
.Sh DESCRIPTION
The
.Nm
utility (also callable as
.Nm geom sched ... )
changes the scheduling policy of the requests going to a provider.
.Pp
The first argument to
.Nm
indicates an action to be performed:
.Bl -tag -width ".Cm configure"
.It Cm create
Create a new provider and geom node using the specified scheduling algorithm.
.Ar algorithm
is the name of the scheduling algorithm used for the provider.
Available algorithms include:
.Ar rr ,
which implements anticipatory scheduling with round robin service
among clients;
.Ar as ,
which implements a simple form of anticipatory scheduling with
no per-client queue.
.Pp
If the operation succeeds, the new provider should appear with name
.Pa /dev/ Ns Ao Ar dev Ac Ns Pa .sched. .
The kernel module
.Pa geom_sched.ko
will be loaded if it is not loaded already.
.It Cm insert
Operates as "create", but the insertion is "transparent",
i.e. the existing provider is rerouted to the newly created geom,
which in turn forwards requests to the existing geom.
This operation allows one to start/stop a scheduling service
on an already existing provider.
.Pp
A subsequent 'destroy' will remove the newly created geom and
hook the provider back to the original geom.
.Ar algorithm
.It Cm configure
Configure existing scheduling provider. It supports the same options
as the
.Nm create
command.
.It Cm destroy
Destroy the geom specified in the parameter.
.It Cm reset
Do nothing.
.It Cm list | status | load | unload
See
.Xr geom 8 .
.El
.Pp
Additional options:
.Bl -tag -width ".Fl f"
.It Fl f
Force the removal of the specified provider.
.It Fl v
Be more verbose.
.El
.Sh SYSCTL VARIABLES
The following
.Xr sysctl 8
variables can be used to control the behavior of the
.Nm SCHED
GEOM class.
The default value is shown next to each variable.
.Bl -tag -width indent
.It Va kern.geom.sched.debug : No 0
Debug level of the
.Nm SCHED
GEOM class.
This can be set to a number between 0 and 2 inclusive.
If set to 0 minimal debug information is printed, and if set to 2 the
maximum amount of debug information is printed.
.El
.Sh EXIT STATUS
Exit status is 0 on success, and 1 if the command fails.
.Sh EXAMPLES
The following example shows how to create a scheduling provider for disk
.Pa /dev/da0
, and how to destroy it.
.Bd -literal -offset indent
# Load the geom_sched module:
kldload geom_sched
# Load some scheduler classes used by geom_sched:
kldload gsched_rr gsched_as
# Configure device ad0 to use scheduler 'rr':
geom sched insert -s rr ad0
# Now provider ad0 uses the 'rr' algorithm;
# the new geom is ad0.sched.
# Remove the scheduler on the device:
geom sched destroy -v ad0.sched.
.Ed
.Pp
.Sh SEE ALSO
.Xr geom 4 ,
.Xr geom 8
.Sh HISTORY
The
.Nm
utility appeared in April 2010.
.Sh AUTHORS
.An Fabio Checconi Aq fabio@FreeBSD.org
.An Luigi Rizzo Aq luigi@FreeBSD.org

162
sys/geom/sched/README Normal file
View File

@ -0,0 +1,162 @@
--- GEOM BASED DISK SCHEDULERS FOR FREEBSD ---
This code contains a framework for GEOM-based disk schedulers and a
couple of sample scheduling algorithms that use the framework and
implement two forms of "anticipatory scheduling" (see below for more
details).
As a quick example of what this code can give you, try to run "dd",
"tar", or some other program with highly SEQUENTIAL access patterns,
together with "cvs", "cvsup", "svn" or other highly RANDOM access patterns
(this is not a made-up example: it is pretty common for developers
to have one or more apps doing random accesses, and others that do
sequential accesses e.g., loading large binaries from disk, checking
the integrity of tarballs, watching media streams and so on).
These are the results we get on a local machine (AMD BE2400 dual
core CPU, SATA 250GB disk):
/mnt is a partition mounted on /dev/ad0s1f
cvs: cvs -d /mnt/home/ncvs-local update -Pd /mnt/ports
dd-read: dd bs=128k of=/dev/null if=/dev/ad0 (or ad0-sched-)
dd-writew dd bs=128k if=/dev/zero of=/mnt/largefile
NO SCHEDULER RR SCHEDULER
dd cvs dd cvs
dd-read only 72 MB/s ---- 72 MB/s ---
dd-write only 55 MB/s --- 55 MB/s ---
dd-read+cvs 6 MB/s ok 30 MB/s ok
dd-write+cvs 55 MB/s slooow 14 MB/s ok
As you can see, when a cvs is running concurrently with dd, the
performance drops dramatically, and depending on read or write mode,
one of the two is severely penalized. The use of the RR scheduler
in this example makes the dd-reader go much faster when competing
with cvs, and lets cvs progress when competing with a writer.
To try it out:
1. USERS OF FREEBSD 7, PLEASE READ CAREFULLY THE FOLLOWING:
On loading, this module patches one kernel function (g_io_request())
so that I/O requests ("bio's") carry a classification tag, useful
for scheduling purposes.
ON FREEBSD 7, the tag is stored in an existing (though rarely used)
field of the "struct bio", a solution which makes this module
incompatible with other modules using it, such as ZFS and gjournal.
Additionally, g_io_request() is patched in-memory to add a call
to the function that initializes this field (i386/amd64 only;
for other architectures you need to manually patch sys/geom/geom_io.c).
See details in the file g_sched.c.
On FreeBSD 8.0 and above, the above trick is not necessary,
as the struct bio contains dedicated fields for the classifier,
and hooks for request classifiers.
If you don't like the above, don't run this code.
2. PLEASE MAKE SURE THAT THE DISK THAT YOU WILL BE USING FOR TESTS
DOES NOT CONTAIN PRECIOUS DATA.
This is experimental code, so we make no guarantees, though
I am routinely using it on my desktop and laptop.
3. EXTRACT AND BUILD THE PROGRAMS
A 'make install' in the directory should work (with root privs),
or you can even try the binary modules.
If you want to build the modules yourself, look at the Makefile.
4. LOAD THE MODULE, CREATE A GEOM NODE, RUN TESTS
The scheduler's module must be loaded first:
# kldload gsched_rr
substitute with gsched_as to test AS. Then, supposing that you are
using /dev/ad0 for testing, a scheduler can be attached to it with:
# geom sched insert ad0
The scheduler is inserted transparently in the geom chain, so
mounted partitions and filesystems will keep working, but
now requests will go through the scheduler.
To change scheduler on-the-fly, you can reconfigure the geom:
# geom sched configure -a as ad0.sched.
assuming that gsched_as was loaded previously.
5. SCHEDULER REMOVAL
In principle it is possible to remove the scheduler module
even on an active chain by doing
# geom sched destroy ad0.sched.
However, there is some race in the geom subsystem which makes
the removal unsafe if there are active requests on a chain.
So, in order to reduce the risk of data losses, make sure
you don't remove a scheduler from a chain with ongoing transactions.
--- NOTES ON THE SCHEDULERS ---
The important contribution of this code is the framework to experiment
with different scheduling algorithms. 'Anticipatory scheduling'
is a very powerful technique based on the following reasoning:
The disk throughput is much better if it serves sequential requests.
If we have a mix of sequential and random requests, and we see a
non-sequential request, do not serve it immediately but instead wait
a little bit (2..5ms) to see if there is another one coming that
the disk can serve more efficiently.
There are many details that should be added to make sure that the
mechanism is effective with different workloads and systems, to
gain a few extra percent in performance, to improve fairness,
insulation among processes etc. A discussion of the vast literature
on the subject is beyond the purpose of this short note.
--------------------------------------------------------------------------
TRANSPARENT INSERT/DELETE
geom_sched is an ordinary geom module, however it is convenient
to plug it transparently into the geom graph, so that one can
enable or disable scheduling on a mounted filesystem, and the
names in /etc/fstab do not depend on the presence of the scheduler.
To understand how this works in practice, remember that in GEOM
we have "providers" and "geom" objects.
Say that we want to hook a scheduler on provider "ad0",
accessible through pointer 'pp'. Originally, pp is attached to
geom "ad0" (same name, different object) accessible through pointer old_gp
BEFORE ---> [ pp --> old_gp ...]
A normal "geom sched create ad0" call would create a new geom node
on top of provider ad0/pp, and export a newly created provider
("ad0.sched." accessible through pointer newpp).
AFTER create ---> [ newpp --> gp --> cp ] ---> [ pp --> old_gp ... ]
On top of newpp, a whole tree will be created automatically, and we
can e.g. mount partitions on /dev/ad0.sched.s1d, and those requests
will go through the scheduler, whereas any partition mounted on
the pre-existing device entries will not go through the scheduler.
With the transparent insert mechanism, the original provider "ad0"/pp
is hooked to the newly created geom, as follows:
AFTER insert ---> [ pp --> gp --> cp ] ---> [ newpp --> old_gp ... ]
so anything that was previously using provider pp will now have
the requests routed through the scheduler node.
A removal ("geom sched destroy ad0.sched.") will restore the original
configuration.
# $FreeBSD$

1902
sys/geom/sched/g_sched.c Normal file
View File

File diff suppressed because it is too large Load Diff

138
sys/geom/sched/g_sched.h Normal file
View File

@ -0,0 +1,138 @@
/*-
* 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 AUTHORS 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 AUTHORS 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.
*/
#ifndef _G_SCHED_H_
#define _G_SCHED_H_
/*
* $Id$
* $FreeBSD$
*
* Header for the geom_sched class (userland library and kernel part).
* See g_sched.c for documentation.
* The userland code only needs the three G_SCHED_* values below.
*/
#define G_SCHED_CLASS_NAME "SCHED"
#define G_SCHED_VERSION 0
#define G_SCHED_SUFFIX ".sched."
#ifdef _KERNEL
#define G_SCHED_DEBUG(lvl, ...) do { \
if (me.gs_debug >= (lvl)) { \
printf("GEOM_SCHED"); \
if (me.gs_debug > 0) \
printf("[%u]", lvl); \
printf(": "); \
printf(__VA_ARGS__); \
printf("\n"); \
} \
} while (0)
#define G_SCHED_LOGREQ(bp, ...) do { \
if (me.gs_debug >= 2) { \
printf("GEOM_SCHED[2]: "); \
printf(__VA_ARGS__); \
printf(" "); \
g_print_bio(bp); \
printf("\n"); \
} \
} while (0)
LIST_HEAD(g_hash, g_sched_class);
/*
* Descriptor of a scheduler.
* In addition to the obvious fields, sc_flushing and sc_pending
* support dynamic switching of scheduling algorithm.
* Normally, sc_flushing is 0, and requests that are scheduled are
* also added to the sc_pending queue, and removed when we receive
* the 'done' event.
*
* When we are transparently inserted on an existing provider,
* sc_proxying is set. The detach procedure is slightly different.
*
* When switching schedulers, sc_flushing is set so requests bypass us,
* and at the same time we update the pointer in the pending bios
* to ignore us when they return up.
* XXX it would be more efficient to implement sc_pending with
* a generation number: the softc generation is increased when
* we change scheduling algorithm, we store the current generation
* number in the pending bios, and when they come back we ignore
* the done() call if the generation number do not match.
*/
struct g_sched_softc {
/*
* Generic fields used by any scheduling algorithm:
* a mutex, the class descriptor, flags, list of pending
* requests (used when flushing the module) and support
* for hash tables where we store per-flow queues.
*/
struct mtx sc_mtx;
struct g_gsched *sc_gsched; /* Scheduler descriptor. */
int sc_pending; /* Pending requests. */
int sc_flags; /* Various flags. */
/*
* Hash tables to store per-flow queues are generally useful
* so we handle them in the common code.
* sc_hash and sc_mask are parameters of the hash table,
* the last two fields are used to periodically remove
* expired items from the hash table.
*/
struct g_hash *sc_hash;
u_long sc_mask;
int sc_flush_ticks; /* Next tick for a flush. */
int sc_flush_bucket; /* Next bucket to flush. */
/*
* Pointer to the algorithm's private data, which is the value
* returned by sc_gsched->gs_init() . A NULL here means failure.
* XXX intptr_t might be more appropriate.
*/
void *sc_data;
};
#define G_SCHED_PROXYING 1
#define G_SCHED_FLUSHING 2
/*
* Temporary- our own version of the disksort, because the
* version in 7.x and 8.x before march 2009 is buggy.
*/
void gs_bioq_init(struct bio_queue_head *);
void gs_bioq_remove(struct bio_queue_head *, struct bio *);
void gs_bioq_flush(struct bio_queue_head *, struct devstat *, int);
void gs_bioq_insert_head(struct bio_queue_head *, struct bio *);
void gs_bioq_insert_tail(struct bio_queue_head *, struct bio *);
struct bio *gs_bioq_first(struct bio_queue_head *);
struct bio *gs_bioq_takefirst(struct bio_queue_head *);
void gs_bioq_disksort(struct bio_queue_head *, struct bio *);
#endif /* _KERNEL */
#endif /* _G_SCHED_H_ */

686
sys/geom/sched/gs_rr.c Normal file
View File

@ -0,0 +1,686 @@
/*-
* 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 = 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 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_UINT(_kern_geom_sched_rr, OID_AUTO, units, CTLFLAG_RD,
&me.units, 0, "Scheduler instances");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, queues, CTLFLAG_RD,
&me.queues, 0, "Total rr queues");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, wait_ms, CTLFLAG_RW,
&me.wait_ms.x_cur, 0, "Wait time milliseconds");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, quantum_ms, CTLFLAG_RW,
&me.quantum_ms.x_cur, 0, "Quantum size milliseconds");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, bypass, CTLFLAG_RW,
&me.bypass, 0, "Bypass scheduler");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, w_anticipate, CTLFLAG_RW,
&me.w_anticipate, 0, "Do anticipation on writes");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, quantum_kb, CTLFLAG_RW,
&me.quantum_kb.x_cur, 0, "Quantum size Kbytes");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, queue_depth, CTLFLAG_RW,
&me.queue_depth.x_cur, 0, "Maximum simultaneous requests");
SYSCTL_UINT(_kern_geom_sched_rr, OID_AUTO, wait_hit, CTLFLAG_RW,
&me.wait_hit, 0, "Hits in anticipation");
SYSCTL_UINT(_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;
if (qp == sc->sc_active && qp->q_status == G_QUEUE_BUSY) {
if (!(qp->q_flags & G_FLAG_COMPLETED)) {
qp->q_flags |= G_FLAG_COMPLETED;
/* in case we want to make the slice adaptive */
qp->q_slice_duration = get_bounded(&me.quantum_ms, 2);
qp->q_slice_end = ticks + qp->q_slice_duration;
}
/* 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);

237
sys/geom/sched/gs_scheduler.h Normal file
View File

@ -0,0 +1,237 @@
/*-
* 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 AUTHORS 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 AUTHORS 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.
*/
/*
* $Id$
* $FreeBSD$
*
* Prototypes for GEOM-based disk scheduling algorithms.
* See g_sched.c for generic documentation.
*
* This file is used by the kernel modules implementing the various
* scheduling algorithms. They should provide all the methods
* defined in struct g_gsched, and also invoke the macro
* DECLARE_GSCHED_MODULE
* which registers the scheduling algorithm with the geom_sched module.
*
* The various scheduling algorithms do not need to know anything
* about geom, they only need to handle the 'bio' requests they
* receive, pass them down when needed, and use the locking interface
* defined below.
*/
#ifndef _G_GSCHED_H_
#define _G_GSCHED_H_
#ifdef _KERNEL
#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/ktr.h>
#include <sys/module.h>
#include <sys/queue.h>
#include <geom/geom.h>
#include "g_sched.h"
/*
* This is the interface exported to scheduling modules.
*
* gs_init() is called when our scheduling algorithm
* starts being used by a geom 'sched'
*
* gs_fini() is called when the algorithm is released.
*
* gs_start() is called when a new request comes in. It should
* enqueue the request and return 0 if success, or return non-zero
* in case of failure (meaning the request is passed down).
* The scheduler can use bio->bio_caller1 to store a non-null
* pointer meaning the request is under its control.
*
* gs_next() is called in a loop by g_sched_dispatch(), right after
* gs_start(), or on timeouts or 'done' events. It should return
* immediately, either a pointer to the bio to be served or NULL
* if no bio should be served now. If force is specified, a
* work-conserving behavior is expected.
*
* gs_done() is called when a request under service completes.
* In turn the scheduler may decide to call the dispatch loop
* to serve other pending requests (or make sure there is a pending
* timeout to avoid stalls).
*
* gs_init_class() is called when a new client (as determined by
* the classifier) starts being used.
*
* gs_hash_unref() is called right before the class hashtable is
* destroyed; after this call, the scheduler is supposed to hold no
* more references to the elements in the table.
*/
/* Forward declarations for prototypes. */
struct g_geom;
struct g_sched_class;
typedef void *gs_init_t (struct g_geom *geom);
typedef void gs_fini_t (void *data);
typedef int gs_start_t (void *data, struct bio *bio);
typedef void gs_done_t (void *data, struct bio *bio);
typedef struct bio *gs_next_t (void *data, int force);
typedef int gs_init_class_t (void *data, void *priv);
typedef void gs_fini_class_t (void *data, void *priv);
typedef void gs_hash_unref_t (void *data);
struct g_gsched {
const char *gs_name;
int gs_refs;
int gs_priv_size;
gs_init_t *gs_init;
gs_fini_t *gs_fini;
gs_start_t *gs_start;
gs_done_t *gs_done;
gs_next_t *gs_next;
g_dumpconf_t *gs_dumpconf;
gs_init_class_t *gs_init_class;
gs_fini_class_t *gs_fini_class;
gs_hash_unref_t *gs_hash_unref;
LIST_ENTRY(g_gsched) glist;
};
#define KTR_GSCHED KTR_SPARE4
MALLOC_DECLARE(M_GEOM_SCHED);
/*
* Basic classification mechanism. Each request is associated to
* a g_sched_class, and each scheduler has the opportunity to set
* its own private data for the given (class, geom) pair. The
* private data have a base type of g_sched_private, and are
* extended at the end with the actual private fields of each
* scheduler.
*/
struct g_sched_class {
int gsc_refs;
int gsc_expire;
u_long gsc_key;
LIST_ENTRY(g_sched_class) gsc_clist;
void *gsc_priv[0];
};
/*
* Manipulate the classifier's data. g_sched_get_class() gets a reference
* to the the class corresponding to bp in gp, allocating and initializing
* it if necessary. g_sched_put_class() releases the reference.
* The returned value points to the private data for the class.
*/
void *g_sched_get_class(struct g_geom *gp, struct bio *bp);
void g_sched_put_class(struct g_geom *gp, void *priv);
static inline struct g_sched_class *
g_sched_priv2class(void *priv)
{
return ((struct g_sched_class *)((u_long)priv -
offsetof(struct g_sched_class, gsc_priv)));
}
static inline void
g_sched_priv_ref(void *priv)
{
struct g_sched_class *gsc;
gsc = g_sched_priv2class(priv);
gsc->gsc_refs++;
}
/*
* Locking interface. When each operation registered with the
* scheduler is invoked, a per-instance lock is taken to protect
* the data associated with it. If the scheduler needs something
* else to access the same data (e.g., a callout) it must use
* these functions.
*/
void g_sched_lock(struct g_geom *gp);
void g_sched_unlock(struct g_geom *gp);
/*
* Restart request dispatching. Must be called with the per-instance
* mutex held.
*/
void g_sched_dispatch(struct g_geom *geom);
/*
* Simple gathering of statistical data, used by schedulers to collect
* info on process history. Just keep an exponential average of the
* samples, with some extra bits of precision.
*/
struct g_savg {
uint64_t gs_avg;
unsigned int gs_smpl;
};
static inline void
g_savg_add_sample(struct g_savg *ss, uint64_t sample)
{
/* EMA with alpha = 0.125, fixed point, 3 bits of precision. */
ss->gs_avg = sample + ss->gs_avg - (ss->gs_avg >> 3);
ss->gs_smpl = 1 + ss->gs_smpl - (ss->gs_smpl >> 3);
}
static inline int
g_savg_valid(struct g_savg *ss)
{
/* We want at least 8 samples to deem an average as valid. */
return (ss->gs_smpl > 7);
}
static inline uint64_t
g_savg_read(struct g_savg *ss)
{
return (ss->gs_avg / ss->gs_smpl);
}
/*
* Declaration of a scheduler module.
*/
int g_gsched_modevent(module_t mod, int cmd, void *arg);
#define DECLARE_GSCHED_MODULE(name, gsched) \
static moduledata_t name##_mod = { \
#name, \
g_gsched_modevent, \
gsched, \
}; \
DECLARE_MODULE(name, name##_mod, SI_SUB_DRIVERS, SI_ORDER_MIDDLE); \
MODULE_DEPEND(name, geom_sched, 0, 0, 0);
#endif /* _KERNEL */
#endif /* _G_GSCHED_H_ */

209
sys/geom/sched/subr_disk.c Normal file
View File

@ -0,0 +1,209 @@
/*-
* ----------------------------------------------------------------------------
* "THE BEER-WARE LICENSE" (Revision 42):
* <phk@FreeBSD.ORG> wrote this file. As long as you retain this notice you
* can do whatever you want with this stuff. If we meet some day, and you think
* this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp
* ----------------------------------------------------------------------------
*
* The bioq_disksort() (and the specification of the bioq API)
* have been written by Luigi Rizzo and Fabio Checconi under the same
* license as above.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
//#include "opt_geom.h"
#include <sys/param.h>
#include <sys/systm.h>
#include <sys/bio.h>
#include <sys/conf.h>
#include <sys/disk.h>
#include <geom/geom_disk.h>
#include "g_sched.h"
/*
* BIO queue implementation
*
* Please read carefully the description below before making any change
* to the code, or you might change the behaviour of the data structure
* in undesirable ways.
*
* A bioq stores disk I/O request (bio), normally sorted according to
* the distance of the requested position (bio->bio_offset) from the
* current head position (bioq->last_offset) in the scan direction, i.e.
*
* (uoff_t)(bio_offset - last_offset)
*
* Note that the cast to unsigned (uoff_t) is fundamental to insure
* that the distance is computed in the scan direction.
*
* The main methods for manipulating the bioq are:
*
* bioq_disksort() performs an ordered insertion;
*
* bioq_first() return the head of the queue, without removing;
*
* bioq_takefirst() return and remove the head of the queue,
* updating the 'current head position' as
* bioq->last_offset = bio->bio_offset + bio->bio_length;
*
* When updating the 'current head position', we assume that the result of
* bioq_takefirst() is dispatched to the device, so bioq->last_offset
* represents the head position once the request is complete.
*
* If the bioq is manipulated using only the above calls, it starts
* with a sorted sequence of requests with bio_offset >= last_offset,
* possibly followed by another sorted sequence of requests with
* 0 <= bio_offset < bioq->last_offset
*
* NOTE: historical behaviour was to ignore bio->bio_length in the
* update, but its use tracks the head position in a better way.
* Historical behaviour was also to update the head position when
* the request under service is complete, rather than when the
* request is extracted from the queue. However, the current API
* has no method to update the head position; secondly, once
* a request has been submitted to the disk, we have no idea of
* the actual head position, so the final one is our best guess.
*
* --- Direct queue manipulation ---
*
* A bioq uses an underlying TAILQ to store requests, so we also
* export methods to manipulate the TAILQ, in particular:
*
* bioq_insert_tail() insert an entry at the end.
* It also creates a 'barrier' so all subsequent
* insertions through bioq_disksort() will end up
* after this entry;
*
* bioq_insert_head() insert an entry at the head, update
* bioq->last_offset = bio->bio_offset so that
* all subsequent insertions through bioq_disksort()
* will end up after this entry;
*
* bioq_remove() remove a generic element from the queue, act as
* bioq_takefirst() if invoked on the head of the queue.
*
* The semantic of these methods is the same of the operations
* on the underlying TAILQ, but with additional guarantees on
* subsequent bioq_disksort() calls. E.g. bioq_insert_tail()
* can be useful for making sure that all previous ops are flushed
* to disk before continuing.
*
* Updating bioq->last_offset on a bioq_insert_head() guarantees
* that the bio inserted with the last bioq_insert_head() will stay
* at the head of the queue even after subsequent bioq_disksort().
*
* Note that when the direct queue manipulation functions are used,
* the queue may contain multiple inversion points (i.e. more than
* two sorted sequences of requests).
*
*/
void
gs_bioq_init(struct bio_queue_head *head)
{
TAILQ_INIT(&head->queue);
head->last_offset = 0;
head->insert_point = NULL;
}
void
gs_bioq_remove(struct bio_queue_head *head, struct bio *bp)
{
if (bp == TAILQ_FIRST(&head->queue))
head->last_offset = bp->bio_offset + bp->bio_length;
if (bp == head->insert_point)
head->insert_point = NULL;
TAILQ_REMOVE(&head->queue, bp, bio_queue);
}
void
gs_bioq_flush(struct bio_queue_head *head, struct devstat *stp, int error)
{
struct bio *bp;
while ((bp = gs_bioq_takefirst(head)) != NULL)
biofinish(bp, stp, error);
}
void
gs_bioq_insert_head(struct bio_queue_head *head, struct bio *bp)
{
head->last_offset = bp->bio_offset;
TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue);
}
void
gs_bioq_insert_tail(struct bio_queue_head *head, struct bio *bp)
{
TAILQ_INSERT_TAIL(&head->queue, bp, bio_queue);
head->insert_point = bp;
}
struct bio *
gs_bioq_first(struct bio_queue_head *head)
{
return (TAILQ_FIRST(&head->queue));
}
struct bio *
gs_bioq_takefirst(struct bio_queue_head *head)
{
struct bio *bp;
bp = TAILQ_FIRST(&head->queue);
if (bp != NULL)
gs_bioq_remove(head, bp);
return (bp);
}
/*
* Compute the sorting key. The cast to unsigned is
* fundamental for correctness, see the description
* near the beginning of the file.
*/
static inline uoff_t
gs_bioq_bio_key(struct bio_queue_head *head, struct bio *bp)
{
return ((uoff_t)(bp->bio_offset - head->last_offset));
}
/*
* Seek sort for disks.
*
* Sort all requests in a single queue while keeping
* track of the current position of the disk with last_offset.
* See above for details.
*/
void
gs_bioq_disksort(struct bio_queue_head *head, struct bio *bp)
{
struct bio *cur, *prev = NULL;
uoff_t key = gs_bioq_bio_key(head, bp);
cur = TAILQ_FIRST(&head->queue);
if (head->insert_point)
cur = head->insert_point;
while (cur != NULL && key >= gs_bioq_bio_key(head, cur)) {
prev = cur;
cur = TAILQ_NEXT(cur, bio_queue);
}
if (prev == NULL)
TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue);
else
TAILQ_INSERT_AFTER(&head->queue, prev, bp, bio_queue);
}

1
sys/modules/geom/Makefile
View File

@ -18,6 +18,7 @@ SUBDIR= geom_bde \
geom_part \
geom_pc98 \
geom_raid3 \
geom_sched \
geom_shsec \
geom_stripe \
geom_sunlabel \

5
sys/modules/geom/geom_sched/Makefile Normal file
View File

@ -0,0 +1,5 @@
# $FreeBSD$
SUBDIR= gs_sched gsched_rr
.include <bsd.subdir.mk>

9
sys/modules/geom/geom_sched/Makefile.inc Normal file
View File

@ -0,0 +1,9 @@
# $FreeBSD$
# included by geom_sched children
.PATH: ${.CURDIR}/../../../../geom/sched
# 6.x needs this path
#CFLAGS += -I${.CURDIR}/../../../../geom/sched
# .include <bsd.kmod.mk>

6
sys/modules/geom/geom_sched/gs_sched/Makefile Normal file
View File

@ -0,0 +1,6 @@
# $FreeBSD$
KMOD= geom_sched
SRCS= g_sched.c subr_disk.c
# ../Makefile.inc automatically included
.include <bsd.kmod.mk>

9
sys/modules/geom/geom_sched/gsched_rr/Makefile Normal file
View File

@ -0,0 +1,9 @@
# $FreeBSD$
KMOD= gsched_rr
SRCS= gs_rr.c
# hash.h on 6.x has a (char *) cast on a const pointer
#CWARNFLAGS=
# ../Makefile.inc automatically included
.include <bsd.kmod.mk>