Re: [PATCH] erofs: fix PSI memstall accounting

[Gao Xiang <hsiangkao@linux.alibaba.com>]

On 2024/12/3 16:02, Max Kellermann via Linux-erofs wrote:
> On Tue, Dec 3, 2024 at 2:42 AM Gao Xiang <hsiangkao@linux.alibaba.com> wrote:
>> I think at least this patch resolves a recent regression,
>> I guess we'd better to address it first.
> 
> Certainly your patch is an improvement, but maybe this just wasn't the
> bug I was experiencing.
> 
>> As for your reports, I think we might need more information
>> about this, since I don't find more clues from the EROFS
>> codebase itself.
> 
> What kind of information do you need? I posted some high-level
> information about my setup here:
> https://lore.kernel.org/lkml/CAKPOu+-_X9cc723v_f_BW4CwfHJe_mi=+cbUBP2tZO-kEcyoMA@mail.gmail.com/
> What else do you want to know?
> This triggers pretty often now - can we improve Suren's patch to
> record and dump more information when this happens?
> 
> Right now, we don't even know if this is an erofs problem (and if this
> ever was). It is just one of several candidates.

I have no idea since I'm not working on PSI stuffs.

But I guess you could record more frame addresses to get the caller
of readahead_expand()? e.g. __builtin_return_address(1)?

Also btw, is _RET_IP_ stable among all cases as readahead_expand()?

Thanks,
Gao Xiang

> 
> Max