James Bottomley wrote:
> On Sat, 2004-10-16 at 01:51, Douglas Gilbert wrote:
> 
>>So this problem seems related to highmem.
>>
>>The attachment is against the current scsi_debug driver
>>(at least in lk 2.6.9-rc4). It uses dma_map_sg() and friends
>>together with phys_to_virt() which comes highly recommended:
>>"in almost all conceivable cases a device driver should not be
>>using this function" :-) Could you try the patch.
>>
>>Perhaps others my be able to answer this: is setting
>>scsi_host_template::dma_boundary (to some figure other
>>than 0xffffffff) going to help in this case?
> 
> 
> No, the problem is with highmem as you correctly point out, but the
> issue is that your driver cannot see beyond it.  On x86, highmem begins
> at about 900Mb.  The kernel has no page tables for any memory beyond
> this.  In order to see the memory you need to kmap it (i.e. ask the
> kernel to create a temporary page table for it); so the virtual address
> sg uses should be got by kmapping the pages in the sg list:
> 
> kaddr = kmap(sg[i].page) + sg[i].offset
> 
> If you make sure clustering is disabled, you should never get multiple
> pages in the initial sg list.
done

> When you've finished you need to release the mapping with kunmap().
ouch, that requires a re-org

> There are also atomic versions of these depending on where you are.
hopefully not required


The above required a fair few changes to scsi_debug.
Attached is a patch that rolls "kmap" changes with
patches I have sent recently for scsi_debug.
Attachment is gzipped (due to size) and applies against
lk 2.6.8.1 -> lk 2.6.9-rc4 .

Nishanth, could you test this with highmem?

Changelog:
    - use kmap/kunmap to handle normal and highmem cases
        - disable clustering
        - previously could only handle 'use_sg > 1' for
          read and write commands
    - VERIFY (SBC) and REWIND (SSC) command support (dummies)
    - 'dsense' parameter (sysfs + load time) for descriptor
      sense format
    - allow negative 'every_nth' to fail continually after
      |every_nth| commands (or until sysfs intervention)
    - clean up debug messages sent to the log (when opts=1)
        - correct ordering of log messages
    - set scsi_cmnd::resid when underflow on DMA_FROM_DEVICE
      operations
    - clean up std inquiry response; add version descriptors
    - reject mode sense subpage commands since no subpages
      currently supported (i.e. better command filtering)


Doug Gilbert