Re: IDE device timeout when writing

[David Brownlee <abs%netbsd.org@localhost>]

On Tue, 12 Oct 2021 at 20:32, Carlos Milán Figueredo
<cmilanf%hispamsx.org@localhost> wrote:
>
> I think there has been a little misunderstanding. I didn't had the issues with wsfb kernel[1], I noticed the problem downloading and unpacking it, so it is totally unrelated.

Ah, apologies for my misunderstanding.

So, good news the wsfb kernel looks good, but bad news that there is a
more general issue. (I'll ask for pullups of Jukka's changes to
netbsd-9)

> It looks like the error happens with writes via the built-in IDE of the Amiga 1200. Tests done to produce the errors:
>
> # wget http://sync.absd.org/amiga/netbsd-9-wsfb.gz
> # gunzip -v netbsd-9-wsfb.gz
>
> These commands generate enough writes to produce the system errors. I have tested that with the following kernels:
>
>  1. NetBSD 9.2 stock GENERIC ------> Device timeout error.
>  2. [2] WSCONS --------------------> Device timeout error.
>  3. [3] WSCONS HEAD ---------------> Device timeout error.
>  4. [1] WSFB ----------------------> Device timeout error.
>  5. AmigaOS -----------------------> OK!

OK, so if it's showing up on WSCONS HEAD then the issue is present in
the very latest code.

If it is easy to test (and only if it's easy) it might be nice to
confirm if you can trigger this under earlier versions of NetBSD.
Maybe booing an appropriate miniroot and gzipping/gunzipping. It's
possible that something has changed over time (of course it's also
possible that thing would be the gcc version used to compile the code,
altering timing or something more obscure)

> So it looks like it doesn't matter what kernel do I use to trigger the error. Some additional data from my setup:
>
>  1. CPU is 68060 rev6 50 MHz
>  2. IDE is built-in Amiga 1200 with a CF2IDE+[4]. I can try replacing the CF2IDE+ if that is suspicious, but it is working well on AmigaOS.
>  3. SanDisk 32 GB CF card. NetBSD swap and root partitions are within the < 4 GB boundary.

I think that once the kernel has loaded the 4GB boundary should be
irrelevant - so if your bootloader is 4GB safe then the NetBSD
partitions could be anywhere. If its notm then the root filesystem
only needs to be entirely within the first 4GB. (This isn't relevant
to the issue, just saying :)

>  4. Ethernet PCMCIA D-Link 660+. I wonder if it could have something to do with the PCMCIA.

It's definitely a possibility, if there are interrupt issues. Again,
if it's easy to test it might be nice to see if you can trigger it
with gzip/gunzip without the PCMCIA card.

> I have also checked that even if the kernel is able to recover, if I stress enough the writes, the system hangs and a reboot happens afterwards. Trying some disk benchmarks on AmigaOS didn't show any problems.

It looks pretty likely that the NetBSD code is driving the hardware in
ways that do not work :) (they may or may not be reasonable ways, but
they are ways which do not work on at least your setup, and quite
likely others).

I may also see if someone with more IDE/ATA-fu than me (that's not a
high bar :) might be interested in taking a look at this thread.

Thanks

David