On Thu, 17 Mar 2005, Jens Axboe wrote:
> See the kmap implementation, mm/highmem.c:map_new_virtual() to be
> precise. If you run out of entries while processing your sglist, you
> will end up waiting on a free kmap entry for an sg entry that will not
> become available before your previous kmaps are released => deadlock.
Right, thanks for the explanation!
> > > > + if (sg_dma_address(sg) <= psge->address &&
> > > > sg_dma_address(sg) + psge->length > psge->address) {
> > >
> > > What is it trying to accomplish?
> >
> > So, I was not sure if the old check was correct, was it? The test is
> > supposed to find the current sg-element.
>
> When this happens, do you know how many bytes of io already completed?
> Most sane drives do, so I would use that to find the current sg entry.
Yeah, it is known at this point. I guess, I'll leave this change out for
now - for this patch at least.
> > > > + sgp[i].address = cpu_to_le32(0xffffffff &
> > > > sg_dma_address(sl +i));
> > > > sgp[i].length = seglen;
> > > > srb->total_xfer_length += seglen;
> > > > }
> > >
> > > I don't understand this change, why the cpu_to_le32?
> >
> > Well, I copied it from tmscsim:
> >
> > pSRB->SGBusAddr = cpu_to_le32(pci_dma_lo32(sg_dma_address(psgl)));
> >
> > I am somewhat confused. In both cases these are bus addresses, right? and
> > sg_dma_address() gives already the bus address, so, both are wrong?
>
> The only reason that would make sense is if ->SGBusAddr is passed to the
> hardware and it requires LE encoding. If it is just driver internal use,
> it doesn't make any sense.
Yes, it is passed to the controller. I think, it is also the case with the
dc395x, although, it is smarter than am53c974 in sg-list processing. Will,
probably, leave it out too.
> > ptr = (u8 *) bus_to_virt( pSRB->SGBusAddr );
...
> > also a case for dma_map_atomic?
>
> It breaks for iommu, for sure.
Ok, expect a patch soon.
So, do we want this kmap_atomic() patch for dc395x or we wait for
helper-functions?
BTW, can one test highmem correctness by booting with, e.g., highmem=xxx
even if no physical highmem is available?
Thanks
Guennadi
---------------------------------
Guennadi Liakhovetski, Ph.D.
DSA Daten- und Systemtechnik GmbH
Pascalstr. 28
D-52076 Aachen
Germany
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
