Re: [PATCH] 2.4 use __dma_buffer in scsi.h

[Roland Dreier <roland@topspin.com>]

>>>>> "James" == James Bottomley <James.Bottomley@steeleye.com> writes:

    James> This is actually unnecessary.  We've already had this
    James> discussion on the parisc mailing lists (over a year ago, I
    James> think).

Fair enough, the only reason I came up with this patch is that I
looked at all unaligned DMA done by USB, and
drivers/usb/storage/transport.c does DMA into sense_buffer.

    James> The SCSI subsystem is designed with this type of cache
    James> incoherency problem in mind.  The Scsi_Cmnd structure has
    James> an ownership model to forestall cache line bouncing.  The
    James> idea is that when you want to dma to/from components of a
    James> Scsi_Cmnd, you can only do it when the ownership is
    James> SCSI_OWNER_LOW_LEVEL, which guarantees that nothing in the
    James> mid layer will touch the command and trigger an incoherency
    James> (at least until it times out).  The low level driver is
    James> supposed to know about the cache incoherency problem, and
    James> so avoids touching the structure between cache line
    James> invalidation/writeback and DMA completion, thus, with a
    James> correctly implemented driver, there should be no
    James> possibility of DMA corruption.

Out of curiousity, how do you deal with someone writing to Scsi_Cmnd
_before_ the DMA and having that data lost when pci_map_single() with
PCI_DMA_FROMDEVICE invalidates the cache before writeback?

Best,
  Roland