Remove support for FreeBSD 7 and really old FreeBSD 8. The classifiers

have been in the base for a while, so the gymnastics here aren't
needed. In addition, the bugs in subr_disk.c have been fixed since
2009, so there's no need for an identical copy of it in the tree
anymore. There's really no need to binary patch g_io_request, so let's
get rid of the code (not compiled in anymore) lest others think it is
a good idea.
This commit is contained in:
Warner Losh 2014-12-20 00:04:01 +00:00
parent a5d3a7fba9
commit 0acf08d985
6 changed files with 13 additions and 436 deletions

View File

@ -39,37 +39,17 @@ 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
1. 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
2. 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
3. LOAD THE MODULE, CREATE A GEOM NODE, RUN TESTS
The scheduler's module must be loaded first:

View File

@ -346,17 +346,8 @@ static inline u_long
g_sched_classify(struct bio *bp)
{
#if __FreeBSD_version > 800098
/* we have classifier fields in the struct bio */
#define HAVE_BIO_CLASSIFIER
return ((u_long)bp->bio_classifier1);
#else
#warning old version!!!
while (bp->bio_parent != NULL)
bp = bp->bio_parent;
return ((u_long)bp->bio_caller1);
#endif
}
/* Return the hash chain for the given key. */
@ -705,7 +696,7 @@ g_gsched_global_init(void)
G_SCHED_DEBUG(0, "Initializing global data.");
mtx_init(&me.gs_mtx, "gsched", NULL, MTX_DEF);
LIST_INIT(&me.gs_scheds);
gs_bioq_init(&me.gs_pending);
bioq_init(&me.gs_pending);
me.gs_initialized = 1;
}
}
@ -914,7 +905,7 @@ g_sched_temporary_start(struct bio *bio)
mtx_lock(&me.gs_mtx);
me.gs_npending++;
gs_bioq_disksort(&me.gs_pending, bio);
bioq_disksort(&me.gs_pending, bio);
mtx_unlock(&me.gs_mtx);
}
@ -923,7 +914,7 @@ g_sched_flush_pending(g_start_t *start)
{
struct bio *bp;
while ((bp = gs_bioq_takefirst(&me.gs_pending)))
while ((bp = bioq_takefirst(&me.gs_pending)))
start(bp);
}
@ -1365,162 +1356,8 @@ g_sched_destroy_geom(struct gctl_req *req, struct g_class *mp,
* to the issuer of a request in bp->bio_classifier1 as soon
* as the bio is posted to the geom queue (and not later, because
* requests are managed by the g_down thread afterwards).
*
* On older versions of the system (but this code is not used
* in any existing release), we [ab]use the caller1 field in the
* root element of the bio tree to store the classification info.
* The marking is done at the beginning of g_io_request()
* and only if we find that the field is NULL.
*
* To avoid rebuilding the kernel, this module will patch the
* initial part of g_io_request() so it jumps to some hand-coded
* assembly that does the marking and then executes the original
* body of g_io_request().
*
* fake_ioreq[] is architecture-specific machine code
* that implements the above. CODE_SIZE, STORE_SIZE etc.
* are constants used in the patching routine. Look at the
* code in g_ioreq_patch() for the details.
*/
#ifndef HAVE_BIO_CLASSIFIER
/*
* Support for old FreeBSD versions
*/
#if defined(__i386__)
#define CODE_SIZE 29
#define STORE_SIZE 5
#define EPILOGUE 5
#define SIZE (CODE_SIZE + STORE_SIZE + EPILOGUE)
static u_char fake_ioreq[SIZE] = {
0x8b, 0x44, 0x24, 0x04, /* mov bp, %eax */
/* 1: */
0x89, 0xc2, /* mov %eax, %edx # edx = bp */
0x8b, 0x40, 0x64, /* mov bp->bio_parent, %eax */
0x85, 0xc0, /* test %eax, %eax */
0x75, 0xf7, /* jne 1b */
0x8b, 0x42, 0x30, /* mov bp->bp_caller1, %eax */
0x85, 0xc0, /* test %eax, %eax */
0x75, 0x09, /* jne 2f */
0x64, 0xa1, 0x00, 0x00, /* mov %fs:0, %eax */
0x00, 0x00,
0x89, 0x42, 0x30, /* mov %eax, bp->bio_caller1 */
/* 2: */
0x55, 0x89, 0xe5, 0x57, 0x56,
0xe9, 0x00, 0x00, 0x00, 0x00, /* jmp back... */
};
#elif defined(__amd64)
#define CODE_SIZE 38
#define STORE_SIZE 6
#define EPILOGUE 5
#define SIZE (CODE_SIZE + STORE_SIZE + EPILOGUE)
static u_char fake_ioreq[SIZE] = {
0x48, 0x89, 0xf8, /* mov bp, %rax */
/* 1: */
0x48, 0x89, 0xc2, /* mov %rax, %rdx # rdx = bp */
0x48, 0x8b, 0x82, 0xa8, /* mov bp->bio_parent, %rax */
0x00, 0x00, 0x00,
0x48, 0x85, 0xc0, /* test %rax, %rax */
0x75, 0xf1, /* jne 1b */
0x48, 0x83, 0x7a, 0x58, /* cmp $0, bp->bp_caller1 */
0x00,
0x75, 0x0d, /* jne 2f */
0x65, 0x48, 0x8b, 0x04, /* mov %gs:0, %rax */
0x25, 0x00, 0x00, 0x00,
0x00,
0x48, 0x89, 0x42, 0x58, /* mov %rax, bp->bio_caller1 */
/* 2: */
0x55, 0x48, 0x89, 0xe5, 0x41, 0x56,
0xe9, 0x00, 0x00, 0x00, 0x00, /* jmp back... */
};
#else /* neither x86 nor amd64 */
static void
g_new_io_request(struct bio *bp, struct g_consumer *cp)
{
struct bio *top = bp;
/*
* bio classification: if bio_caller1 is available in the
* root of the 'struct bio' tree, store there the thread id
* of the thread that originated the request.
* More sophisticated classification schemes can be used.
*/
while (top->bio_parent)
top = top->bio_parent;
if (top->bio_caller1 == NULL)
top->bio_caller1 = curthread;
}
#error please add the code above in g_new_io_request() to the beginning of \
/sys/geom/geom_io.c::g_io_request(), and remove this line.
#endif /* end of arch-specific code */
static int
g_ioreq_patch(void)
{
u_char *original;
u_long ofs;
int found;
if (me.gs_patched)
return (-1);
original = (u_char *)g_io_request;
found = !bcmp(original, fake_ioreq + CODE_SIZE, STORE_SIZE);
if (!found)
return (-1);
/* Jump back to the original + STORE_SIZE. */
ofs = (original + STORE_SIZE) - (fake_ioreq + SIZE);
bcopy(&ofs, fake_ioreq + CODE_SIZE + STORE_SIZE + 1, 4);
/* Patch the original address with a jump to the trampoline. */
*original = 0xe9; /* jump opcode */
ofs = fake_ioreq - (original + 5);
bcopy(&ofs, original + 1, 4);
me.gs_patched = 1;
return (0);
}
/*
* Restore the original code, this is easy.
*/
static void
g_ioreq_restore(void)
{
u_char *original;
if (me.gs_patched) {
original = (u_char *)g_io_request;
bcopy(fake_ioreq + CODE_SIZE, original, STORE_SIZE);
me.gs_patched = 0;
}
}
static inline void
g_classifier_ini(void)
{
g_ioreq_patch();
}
static inline void
g_classifier_fini(void)
{
g_ioreq_restore();
}
/*--- end of support code for older FreeBSD versions */
#else /* HAVE_BIO_CLASSIFIER */
/*
* Classifier support for recent FreeBSD versions: we use
* a very simple classifier, only use curthread to tag a request.
@ -1552,7 +1389,6 @@ g_classifier_fini(void)
g_unregister_classifier(&g_sched_classifier);
}
#endif /* HAVE_BIO_CLASSIFIER */
static void
g_sched_init(struct g_class *mp)

View File

@ -120,19 +120,6 @@ struct g_sched_softc {
#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_ */

View File

@ -315,7 +315,7 @@ 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);
bioq_init(&qp->q_bioq);
/*
* Set the initial parameters for the client:
@ -350,7 +350,7 @@ g_rr_fini_class(void *data, void *priv)
{
struct g_rr_queue *qp = priv;
KASSERT(gs_bioq_first(&qp->q_bioq) == NULL,
KASSERT(bioq_first(&qp->q_bioq) == NULL,
("released nonempty queue"));
qp->q_sc->sc_nqueues--;
me.queues--;
@ -438,7 +438,7 @@ g_rr_next(void *data, int force)
qp->q_flags &= ~G_FLAG_COMPLETED;
}
bp = gs_bioq_takefirst(&qp->q_bioq); /* surely not NULL */
bp = bioq_takefirst(&qp->q_bioq); /* surely not NULL */
qp->q_service += bp->bio_length; /* charge the service */
/*
@ -456,7 +456,7 @@ g_rr_next(void *data, int force)
* 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 ? */
next = bioq_first(&qp->q_bioq); /* do we have one more ? */
if (expired) {
sc->sc_active = NULL;
/* Either requeue or release reference. */
@ -538,7 +538,7 @@ g_rr_start(void *data, struct bio *bp)
if (qp == NULL)
return (-1); /* allocation failed, tell upstream */
if (gs_bioq_first(&qp->q_bioq) == NULL) {
if (bioq_first(&qp->q_bioq) == NULL) {
/*
* We are inserting into an empty queue.
* Reset its state if it is sc_active,
@ -560,7 +560,7 @@ g_rr_start(void *data, struct bio *bp)
/* Inherit the reference returned by g_rr_queue_get(). */
bp->bio_caller1 = qp;
gs_bioq_disksort(&qp->q_bioq, bp);
bioq_disksort(&qp->q_bioq, bp);
return (0);
}

View File

@ -1,226 +0,0 @@
/*-
* ----------------------------------------------------------------------------
* "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 as 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 (head->insert_point == NULL) {
if (bp == TAILQ_FIRST(&head->queue))
head->last_offset = bp->bio_offset + bp->bio_length;
} else 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)
{
if (head->insert_point == NULL)
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;
head->last_offset = bp->bio_offset;
}
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;
uoff_t key;
if ((bp->bio_flags & BIO_ORDERED) != 0) {
/*
* Ordered transactions can only be dispatched
* after any currently queued transactions. They
* also have barrier semantics - no transactions
* queued in the future can pass them.
*/
gs_bioq_insert_tail(head, bp);
return;
}
prev = NULL;
key = gs_bioq_bio_key(head, bp);
cur = TAILQ_FIRST(&head->queue);
if (head->insert_point) {
prev = head->insert_point;
cur = TAILQ_NEXT(head->insert_point, bio_queue);
}
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);
}

View File

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