Re: kern/54486: athn driver panics on arm

[Nick Hudson <nick.hudson%gmx.co.uk@localhost>]

The following reply was made to PR kern/54486; it has been noted by GNATS.

From: Nick Hudson <nick.hudson%gmx.co.uk@localhost>
To: Rin Okuyama <rokuyama%meiji.ac.jp@localhost>, matthew green <mrg%eterna.com.au@localhost>,
 Andreas Gustafsson <gson%gson.org@localhost>
Cc: Christos Zoulas <christos%astron.com@localhost>, "maya%NetBSD.org@localhost"
 <maya%NetBSD.org@localhost>, "gnats-bugs%netbsd.org@localhost" <gnats-bugs%netbsd.org@localhost>
Subject: Re: kern/54486: athn driver panics on arm
Date: Sun, 25 Aug 2019 08:54:35 +0100

 On 24/08/2019 13:01, Rin Okuyama wrote:
 > On 2019/08/24 20:11, matthew green wrote:
 >>> I found some related discussion in the port-arm archives, starting at:
 >>>
 >>>     http://mail-index.netbsd.org/port-arm/2019/01/06/msg005357.html
 >>>
 >>> My understanding of that discussion is that there are two alternative
 >>> correct fixes: either apply the patch from the above message to add
 >>> -mno-unaligned-access to CFLAGS, or stop using PMAP_NOCACHE in
 >>> bus_dma.c as mentioned in:
 >>>
 >>>     http://mail-index.netbsd.org/port-arm/2019/01/06/msg005358.html
 >>>
 >>> The problem is that *neither* has been done.  Since the patch is less
 >>> likely to have bad side effects, I think we should apply the patch
 >>> now, and if or when the pmap change is ever made, the patch can be
 >>> reverted.
 >>
 >> apparently i agreed with you back in january.  Rin, can you
 >> commit your change for now.  at least until we fix pmap..
 >
 > IMO, adding -mno-unaligned-access to CFLAGS is not a correct fix,
 > although affects on performance seems negligible as far as I can
 > see (I guess that benefits by ability of unaligned access are
 > canceled by penalty of it).
 >
 > A real fix should be using cacheble memory for DMA buffer. However,
 > it breaks some drivers, at least, awge(4) and vchiq(4). Now, Nick
 > is working on it.
 >
 > Nick, how is the progress? If it takes a long time, we need to
 > consider a workaround of -mno-unaligned-access (and pull it up
 > into netbsd-9).
 
 My focus is elsewhere at the moment, so feel free to apply the workaround.
 
 Perhaps we can compile a handful of drivers with -mno-unaligned-access
 and not the whole kernel?
 
 Thanks,
 Nick