On Sat, May 18, 2019 at 01:02:28AM +0900, Jaewon Kim wrote:
> Hello I don't have enough knowledge on USB core but I've wondered
> why GFP_NOIO has been used in xhci_alloc_dev for
> xhci_alloc_virt_device. I found commit ("a6d940dd759b xhci: Use
> GFP_NOIO during device reset"). But can we just change GFP_NOIO
> to __GFP_RECLAIM | __GFP_FS ?
No. __GFP_FS implies __GFP_IO; you can't set __GFP_FS and clear __GFP_IO.
It seems like the problem you have is using the CMA to do DMA allocation.
Why would you do such a thing?
> Please refer to below case.
>
> I got a report from Lee YongTaek <ytk.lee@xxxxxxxxxxx> that the
> xhci_alloc_virt_device was too slow over 2 seconds only for one page
> allocation.
>
> 1) It was because kernel version was v4.14 and DMA allocation was
> done from CMA(Contiguous Memory Allocator) where CMA region was
> almost filled with file page and CMA passes GFP down to page
> isolation. And the page isolation only allows file page isolation only to
> requests having __GFP_FS.
>
> 2) Historically CMA was changed at v4.19 to use GFP_KERNEL
> regardless of GFP passed to DMA allocation through the
> commit 6518202970c1 "(mm/cma: remove unsupported gfp_mask
> parameter from cma_alloc()".
>
> I think pre v4.19 the xhci_alloc_virt_device could be very slow
> depending on CMA situation but free to USB deadlock issue. But as of
> v4.19, I think, it will be fast but can face the deadlock issue.
> Consequently I think to meet the both cases, I think USB can pass
> __GFP_FS without __GFP_IO.
>
> If __GFP_FS is passed from USB core, of course, the CMA patch also
> need to be changed to pass GFP.
