Re: elevator priorities vs. full request queues

[Werner Almesberger <wa@almesberger.net>]

Jens Axboe wrote:
> Something like this (probably a little half-assed, and definitely very
> untested :-).

Nevertheless, it seems to work well enough :-) The only bug I've
noticed is that calculations related to bi_rw need to be unsigned
long, for 64 bit compatibility, i.e.

+#define bio_set_prio(bio, prio)	do {			\
+	WARN_ON(prio >= (1 << BIO_PRIO_BITS));		\
+	(bio)->bi_rw &= ((1UL << BIO_PRIO_SHIFT) - 1);	\
+	(bio)->bi_rw |= ((unsigned long) (prio) << BIO_PRIO_SHIFT);	\
+} while (0)

I've adapted your per-process IO priority idea, and used it as
follows:

--- linux-2.6.7-orig/include/linux/sched.h	Wed Jun 16 02:18:57 2004
+++ linux-2.6.7/include/linux/sched.h	Sun Jul 11 15:00:31 2004
@@ -505,6 +505,7 @@ struct task_struct {
 	struct backing_dev_info *backing_dev_info;
 
 	struct io_context *io_context;
+	int ioprio;
 
 	unsigned long ptrace_message;
 	siginfo_t *last_siginfo; /* For ptrace use.  */
--- linux-2.6.7-orig/fs/buffer.c	Wed Jun 16 02:19:36 2004
+++ linux-2.6.7/fs/buffer.c	Mon Jul 12 08:25:41 2004
@@ -2789,6 +2789,8 @@ void submit_bh(int rw, struct buffer_hea
 	bio->bi_end_io = end_bio_bh_io_sync;
 	bio->bi_private = bh;
 
+	bio_set_prio(bio, current->ioprio);
+
 	submit_bio(rw, bio);
 }
 
--- linux-2.6.7-orig/drivers/block/ll_rw_blk.c	Sun Jul 11 14:20:07 2004
+++ linux-2.6.7/drivers/block/ll_rw_blk.c	Mon Jul 12 08:20:41 2004
@@ -2320,7 +2320,7 @@ static inline void blk_partition_remap(s
 	if (bdev != bdev->bd_contains) {
 		struct hd_struct *p = bdev->bd_part;
 
-		switch (bio->bi_rw) {
+		switch (bio->bi_rw & BIO_RW) {
 		case READ:
 			p->read_sectors += bio_sectors(bio);
 			p->reads++;
@@ -2451,7 +2451,7 @@ void submit_bio(int rw, struct bio *bio)
 
 	BIO_BUG_ON(!bio->bi_size);
 	BIO_BUG_ON(!bio->bi_io_vec);
-	bio->bi_rw = rw;
+	bio->bi_rw |= rw;
 	if (rw & WRITE)
 		mod_page_state(pgpgout, count);
 	else

Because I'm lazy, I'm using a default priority of zero, so I
don't need any explicit initialization.

I've been playing with this for a few hours, and even a
request-happy load with random accesses through AIO, which
normally basically kills the machine, doesn't impress my
high-priority reader anymore. I haven't looked into fairness
issues, though.

- Werner

-- 
  _________________________________________________________________________
 / Werner Almesberger, Buenos Aires, Argentina         wa@almesberger.net /
/_http://www.almesberger.net/____________________________________________/