Re: [PATCH] Chaining sg lists for big IO commands v5

[Jens Axboe <jens.axboe@oracle.com>]

On Tue, May 15 2007, Jens Axboe wrote:
> On Tue, May 15 2007, Badari Pulavarty wrote:
> > On Tue, 2007-05-15 at 19:20 +0200, Jens Axboe wrote:
> > > On Tue, May 15 2007, Badari Pulavarty wrote:
> > > > On Fri, 2007-05-11 at 15:51 +0200, Jens Axboe wrote:
> > > > > Hi,
> > > > > 
> > > > > Updated version of the patch - this time I'll just attach the patch
> > > > > file...
> > > > 
> > > > Missing scatterlist.h inclusions..
> > > > 
> > > > drivers/scsi/sym53c8xx_2/sym_glue.c: In function ???sym_scatter???:
> > > > drivers/scsi/sym53c8xx_2/sym_glue.c:385: warning: implicit declaration
> > > > of function ???for_each_sg???
> > > > drivers/scsi/sym53c8xx_2/sym_glue.c:385: error: expected ???;??? before ???{???
> > > > token
> > > > drivers/scsi/sym53c8xx_2/sym_glue.c:375: warning: unused variable ???tp???
> > > > make[3]: *** [drivers/scsi/sym53c8xx_2/sym_glue.o] Error 1
> > > > 
> > > > 
> > > > drivers/scsi/qla2xxx/qla_iocb.c: In function ???qla24xx_build_scsi_iocbs???:
> > > > drivers/scsi/qla2xxx/qla_iocb.c:678: warning: implicit declaration of
> > > > function ???for_each_sg???
> > > > drivers/scsi/qla2xxx/qla_iocb.c:678: error: expected ???;??? before ???{???
> > > > token
> > > 
> > > Thanks, will fix those. What arch? I tested it here.
> > 
> > I am playing with them on ppc64.
> 
> Ah ok, you need the updated patch series for ppc64 support. Builds fine
> here on ppc64. See the #sglist branch of the block repo:

A little teaser for those that haven't tried it yet. This is the
blktrace stats of doing a dd from a device and piping to md5sum. Just
meant to test that I got the same data back with chained sglists.

Snips from the blktrace data:

  8,16   2     1302    14.958707808     0  C   R 983040 + 10240 [0]
  8,16   2     1303    14.959619712     0  D   R 1005568 + 13312 [swapper]
  8,16   2     1304    15.066257952     0  C   R 993280 + 12288 [0]
  8,16   2     1305    15.067238272     0  D   R 1018880 + 13312 [swapper]

That's completion of a 5120kb request, dispatch of a 6656kb request,
completion of a 6144kb request, and finally dispatch of a 6656kb request
again. Average io size for reading half a gig of data:

 Read Dispatches:      110,  524,288KiB

or ~4766kb average IO size. 6656 are the largest requests in this test,
it's running into the hw sg limit at that point. Just doing a plain
O_DIRECT dd from the device makes basically all requests reach the max,
giving an average io size above 6mb/request.

-- 
Jens Axboe