[dpdk-dev] [PATCH] vfio: fix expanding DMA area in ppc64le

Burakov, Anatoly anatoly.burakov at intel.com
Wed Jun 26 11:43:50 CEST 2019


On 18-Jun-19 3:37 AM, Mo, YufengX wrote:
> From: Takeshi Yoshimura <tyos at jp.ibm.com>
> 
> In ppc64le, expanding DMA areas always fail because we cannot remove
> a DMA window. As a result, we cannot allocate more than one memseg in
> ppc64le. This is because vfio_spapr_dma_mem_map() doesn't unmap all
> the mapped DMA before removing the window. This patch fixes this
> incorrect behavior.
> 
> I added a global variable to track current window size since we do
> not have better ways to get exact size of it than doing so. sPAPR
> IOMMU seems not to provide any ways to get window size with ioctl
> interfaces. rte_memseg_walk*() is currently used to calculate window
> size, but it walks memsegs that are marked as used, not mapped. So,
> we need to determine if a given memseg is mapped or not, otherwise
> the ioctl reports errors due to attempting to unregister memory
> addresses that are not registered. The global variable is excluded
> in non-ppc64le binaries.
> 
> Similar problems happen in user maps. We need to avoid attempting to
> unmap the address that is given as the function's parameter. The
> compaction of user maps prevents us from passing correct length for
> unmapping DMA at the window recreation. So, I removed it in ppc64le.
> 
> I also fixed the order of ioctl for unregister and unmap. The ioctl
> for unregister sometimes report device busy errors due to the
> existence of mapped area.
> 
> Signed-off-by: Takeshi Yoshimura <tyos at jp.ibm.com>
> ---

OK there are three patches, and two v1's with two different authors in 
reply to the same original patch. There's too much going on here, i 
can't review this. Needs splitting.

Also, #ifdef-ing out the map merging seems highly suspect.

With regards to "walking used memsegs, not mapped", unless i'm 
misunderstanding something, these are the same - whenever a segment is 
mapped, it is marked as used, and whenever it is unmapped, it is marked 
as free. Could you please explain what is the difference and why is this 
needed?

Is the point of contention here being the fact that whenever the unmap 
callback arrives, the segments still appear used when iterating over the 
map? If that's the case, then i think it would be OK to mark them as 
unused *before* triggering callbacks, and chances are some of this code 
wouldn't be needed. That would require a deprecation notice though, 
because the API behavior will change (even if this fact is not 
documented properly).

-- 
Thanks,
Anatoly


More information about the dev mailing list