Re: New 2.6.24.2 SG_IO SCSI problems

[Mike Christie]

Mark Hounschell wrote:
> Mark Hounschell wrote:
> > Mike Christie wrote:
> > > Mike Christie wrote:
> > > > Mark Hounschell wrote:
> > > > > I seem to have run into some sort of regression in the SG_IO
> > > > > interface of 2.6.24.2. I have an application that up until 2.6.24
> > > > > worked fine. The 2.6.23.16 kernel works fine.
> > > > >
> > > > > During reads I get these kernel messages. Writes and other functions
> > > > > _seem_ OK. Actually basic
> > > > > reads  are working. Its with large BC reads using an io_vec list that
> > > > > the problem shows up.
> > > > Are you doing SG_IO to the sg device (/dev/sg*) or to the block device
> > > > (/dev/sdX)?
> > > If you are doing SG_IO to the sg device, then I know of one regression
> > > (well not regression exactly, but I fixed a bug but the patch got
> > > partially overwritten by another patch and that caused a new bug). Both
> > > bugs are fixed in 2.6.25-rc2. Could you try that out if you are doing
> > > SG_IO to the sg device.
> > Yes, I'm using /dev/sg*. And yes again I'll checkout 2.6.25-rc2 ASIC.
> >
> > Thanks
> > Mark
> > -
>  
> 2.6.25-rc2 does fix the problem I'm having. I don't suppose there is a patch
> lying around for 2.6.24.2??

I attached a backport of the patch from Tony (added as cc) that is in 
2.6.25-rc2. Could you try it out against 2.6.24.2 just to make sure it 
was this patch, then we can send it to stable.
Backport
76d78300a6eb8b7f08e47703b7e68a659ffc2053
to 2.6.24

>From Tony Battersby:

When sending a SCSI command to a tape drive via the SCSI Generic (sg)
driver, if the command has a data transfer length more than
scatter_elem_sz (32 KB default) and not a multiple of 512, then I either
hit BUG_ON(!valid_dma_direction(direction)) in dma_unmap_sg() or else
the command never completes (depending on the LLDD).

When constructing scatterlists, the sg driver rounds up the scatterlist
element sizes to be a multiple of 512.  This can result in
sum(scatterlist lengths) > bufflen.  In this case, scsi_req_map_sg()
incorrectly sets bio->bi_size to sum(scatterlist lengths) rather than to
bufflen.  When the command completes, req_bio_endio() detects that
bio->bi_size != 0, and so it doesn't call bio_endio().  This causes the
command to be resubmitted, resulting in BUG_ON or the command never
completing.

This patch makes scsi_req_map_sg() set bio->bi_size to bufflen rather
than to sum(scatterlist lengths), which fixes the problem.

Signed-off-by: Mike Christie <michaelc@xxxxxxxxxxx>

--- linux-2.6.24.2/drivers/scsi/scsi_lib.c	2008-02-10 23:51:11.000000000 -0600
+++ linux-2.6.24.2.work/drivers/scsi/scsi_lib.c	2008-02-22 16:20:09.000000000 -0600
@@ -298,7 +298,6 @@ static int scsi_req_map_sg(struct reques
 		page = sg_page(sg);
 		off = sg->offset;
 		len = sg->length;
- 		data_len += len;
 
 		while (len > 0 && data_len > 0) {
 			/*