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

James Bottomley (James.Bottomley@steeleye.com)
Wed, 12 Jun 2002 22:17:59 -0400


> Use __dma_buffer macro to align sense_buffer member of Scsi_Cmnd.
[...]

> - unsigned char sense_buffer[SCSI_SENSE_BUFFERSIZE]; /* obtained by REQUEST SENSE
> + unsigned char sense_buffer[SCSI_SENSE_BUFFERSIZE] __dma_buffer; /* obtained by REQUEST SENSE

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

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

Each of the Scsi_Cmnd blocks is individually allocated with kmalloc, so
they're guaranteed to be non-interfering when it comes to DMA.

Incidentally, if you're really going to insist on padding the structure, some
drivers also use the cmnd element to DMA the command from, so that should be
aligned as well. But I think the best course of action is to fix any low
level drivers that are not following the cache rules rather than expanding
Scsi_Cmnd.

James Bottomley

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/