reducing max segments expected to work?

reducing max segments expected to work?

[Benjamin LaHaise]

Hello Jens et al,

Is reducing the max number of segments in the block layer supposed to 
work (as done in the patch below), or should i be sticking to mucking 
with MAX_PHYS_SEGMENTS?  I seem to get a kernel thatt cannot boot with 
the below patch applied, and was wondering if you're aware of any 
problems in this area.  I'll probably post something more detailed 
tomorrow after trying a few things.

		-ben
-- 
"Time is what keeps everything from happening all at once." -- John Wheeler


diff -purN v2.6.12-rc4/include/linux/blkdev.h test-rc4/include/linux/blkdev.h
--- v2.6.12-rc4/include/linux/blkdev.h	2005-04-28 11:02:01.000000000 -0400
+++ test-rc4/include/linux/blkdev.h	2005-05-11 17:06:10.000000000 -0400
@@ -667,8 +667,8 @@ extern long blk_congestion_wait(int rw, 
 extern void blk_rq_bio_prep(request_queue_t *, struct request *, struct bio *);
 extern int blkdev_issue_flush(struct block_device *, sector_t *);
 
-#define MAX_PHYS_SEGMENTS 128
-#define MAX_HW_SEGMENTS 128
+#define MAX_PHYS_SEGMENTS 32
+#define MAX_HW_SEGMENTS 32
 #define MAX_SECTORS 255
 
 #define MAX_SEGMENT_SIZE	65536

Re: reducing max segments expected to work?

[Jens Axboe]

On Wed, May 11 2005, Benjamin LaHaise wrote:
> Hello Jens et al,
> 
> Is reducing the max number of segments in the block layer supposed to 
> work (as done in the patch below), or should i be sticking to mucking 
> with MAX_PHYS_SEGMENTS?  I seem to get a kernel thatt cannot boot with 
> the below patch applied, and was wondering if you're aware of any 
> problems in this area.  I'll probably post something more detailed 
> tomorrow after trying a few things.
> 
> 		-ben
> -- 
> "Time is what keeps everything from happening all at once." -- John Wheeler
> 
> 
> diff -purN v2.6.12-rc4/include/linux/blkdev.h test-rc4/include/linux/blkdev.h
> --- v2.6.12-rc4/include/linux/blkdev.h	2005-04-28 11:02:01.000000000 -0400
> +++ test-rc4/include/linux/blkdev.h	2005-05-11 17:06:10.000000000 -0400
> @@ -667,8 +667,8 @@ extern long blk_congestion_wait(int rw, 
>  extern void blk_rq_bio_prep(request_queue_t *, struct request *, struct bio *);
>  extern int blkdev_issue_flush(struct block_device *, sector_t *);
>  
> -#define MAX_PHYS_SEGMENTS 128
> -#define MAX_HW_SEGMENTS 128
> +#define MAX_PHYS_SEGMENTS 32
> +#define MAX_HW_SEGMENTS 32
>  #define MAX_SECTORS 255

This doesn't really do what you would think it does - the defines should
be called DEFAULT_PHYS_SEGMENTS etc, since they are just default values
and do not denote any max-allowed-by-driver value.

But it is strange why your system wont boot after applying the above.
What happens (and what kind of storage)?

-- 
Jens Axboe

Re: reducing max segments expected to work?

[Benjamin LaHaise]

On Thu, May 12, 2005 at 08:37:57AM +0200, Jens Axboe wrote:
> This doesn't really do what you would think it does - the defines should
> be called DEFAULT_PHYS_SEGMENTS etc, since they are just default values
> and do not denote any max-allowed-by-driver value.

They do place a limit on athe sgpool entries in scsi_lib.c.  I'm curious 
about the overhead from these data structures, hence the experimentation.

> But it is strange why your system wont boot after applying the above.
> What happens (and what kind of storage)?

The system is a pretty standard P4 with SATA on ICH6.  I tried booting 
with MAX_SECTORS = 31 (with *_SEGMENTS = 32) to no avail.  The system 
usually gets to some point in early userland init with whatever program 
(init) being stuck in D state waiting for io to complete.  I'm curious 
if there is some unwritten dependancy on MAX_SEGMENT_SIZE or some other 
piece of code being hit here...

		-ben
-- 
"Time is what keeps everything from happening all at once." -- John Wheeler