Re: swiotlb/caamjr regression (Was: [GIT PULL] (swiotlb) stable/for-linus-5.12)

From: 'Dominique MARTINET'
Date: Tue Jun 22 2021 - 03:48:45 EST

Konrad Rzeszutek Wilk wrote on Mon, Jun 21, 2021 at 09:16:43AM -0400:
> The beaty of 'devel' and 'linux-next' is that they can be reshuffled and
> mangled. I pushed them original patch from Bumyong there and will let
> it sit for a day and then create a stable branch and give it to Linus.

Thanks, that should be good.

Do you want me to send a follow-up patch with the two extra checks
(tlb_addr & (IO_TLB_SIZE -1)) > swiotlb_align_offset(dev, orig_addr)
tlb_offset < alloc_size

or are we certain this can't ever happen?
(I didn't see any hit in dmesg when I ran with these, but my opinion is
better safe than sorry...)

> Then I need to expand the test-regression bucket so that this does not
> happen again. Dominique, how easy would it be to purchase one of those
> devices?

My company is making such a device, but it's not on the market yet
(was planned for august, with some delay in approvisionning it'll
probably be a bit late), and would mean buying from Japan so I'm not
sure how convenient that would be...

These are originally NXP devices so I assume Horia would have better
suggestions, if you would?

> I was originally thinking to create a crypto device in QEMU to simulate
> this but that may take longer to write than just getting the real thing.
> Or I could create some fake devices with weird offsets and write a driver
> for it to exercise this.. like this one I had done some time ago that
> needs some brushing off.

Just a fake device with fake offsets as a test is probably good enough,
ideally would need to exerce both failures we've seen (offset in
dma_sync_single_for_device like caam does and in the original mapping (I
assume?) like the NVMe driver does), but that sounds possible :)

Thanks again!