Messages in this thread | | | Date | Sat, 07 Jun 2003 00:11:40 -0700 (PDT) | Subject | Re: problem with blk_queue_bounce_limit() | From | "David S. Miller" <> |
| |
From: David Mosberger <davidm@napali.hpl.hp.com> Date: Sat, 7 Jun 2003 00:05:06 -0700
But you're creating a new mapping for the old buffer. What if you had a DMA API implementation which consolidates multiple mapping attempts of the same buffer into a single mapping entry (along with a reference count)? That would break the workaround.
I hope nobody is doing this, it would probably break other things we haven't considered yet.
You can't support all the BIO_MERGE_BOUNDARY stuff properly in such a scheme. And you _WANT_ to support that when you have an IOMMU, it shrinks the DMA descriptor addr/len entries a chip has to DMA for each block I/O considerably.
Isn't the proper fix to (a) get a new buffer, (b) create a mapping for the new buffer, (c) destroy the mapping for the old buffer. That should guarantee a different bus address, no matter what the DMA-mapping implementation.
I suppose this would work, fell free to code this up for the tg3 driver for me because I certainly lack the time to do this. - 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/
| |