Re: [PATCH v3 00/37] IB: Optimize DMA mapping

[Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 2499 bytes --]

On Fri, 2017-01-20 at 13:04 -0800, Bart Van Assche wrote:
> Hello Doug,
> 
> As you know there are two sets of DMA mapping operations in the Linux
> kernel:
> - One set of DMA mapping operations that is used by most drivers.
> - Another set of DMA mapping operations that is only used by the RDMA
>   drivers.
> Having two types of DMA mapping operations is not only a source of
> confusion but also a source of unnecessary overhead. The DMA mapping
> operations are in the hot path so it is important that the overhead
> of these operations is as low as possible. Hence this patch series
> that converts the RDMA code to the standard DMA mapping operations
> and thereby eliminates the if (dev->dma_ops) test from the hot path.
> 
> The changes compared to version 2 of this patch series are:
> - Left out the inlining of the ib_dma_map_*() functions to keep the
>   number of patches in this patch series reasonable.
> - Reworked this patch series such that (hopefully) the hfi1 and qib
>   drivers keep working.
> 
> The changes compared to version 1 of this patch series are:
> - Patch "Move dma_ops from archdata into struct device" has been
>   split into three patches.
> - Patch "treewide: Inline ib_dma_map_*() functions" has been split
>   into 15 patches (one per driver).
> - A patch has been added that builds dma_noop_ops only for the
>   architectures that need it.
> - The new dma_virt_ops is only built if it is used by a driver.
> - In these last 15 patches indentation has been adjusted to keep
>   the arguments aligned with the opening parenthesis.

Hi Bart,

I think this has enough Acks to move foward (especially from the larger
kernel community on the core kernel bits, such as GKH's ack on the core
bits).  I've pulled it into a branch (there were a few fixups I had to
do, compile testing that now).  However, I'm going to wait until the
very end to merge this branch right before my pull request.  That way
if Linus has any objections, he can just pop the top 37 patches off of
my pull request and effectively remove this patchset, at which point we
can fix up whatever he objected to and resubmit.  And I'm only doing
that because of the number of patches that are either outside of
drivers/infiniband or treewide in this series.

-- 
Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
    GPG KeyID: B826A3330E572FDD
   
Key fingerprint = AE6B 1BDA 122B 23B4 265B  1274 B826 A333 0E57 2FDD

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 819 bytes --]