Re: [bug report] dead loop in generic_perform_write() //Re: [PATCH v7 07/12] iov_iter: Convert iterate*() to inline funcs

[Tong Tiangen]

在 2024/3/2 10:59, Linus Torvalds 写道:
> On Thu, 29 Feb 2024 at 09:32, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > One option might be to make a failed memcpy_from_iter_mc() set another
> > flag in the iter, and then make fault_in_iov_iter_readable() test that
> > flag and return 'len' if that flag is set.
> >
> > Something like that (wild handwaving) should get the right error handling.
> >
> > The simpler alternative is maybe something like the attached.
> > COMPLETELY UNTESTED. Maybe I've confused myself with all the different
> > indiraction mazes in the iov_iter code.
>
> Actually, I think the right model is to get rid of that horrendous
> .copy_mc field entirely.
>
> We only have one single place that uses it - that nasty core dumping
> code. And that code is *not* performance critical.
>
> And not only isn't it performance-critical, it already does all the
> core dumping one page at a time because it doesn't want to write pages
> that were never mapped into user space.
>
> So what we can do is
>
>   (a) make the core dumping code *copy* the page to a good location
> with copy_mc_to_kernel() first
>
>   (b) remove this horrendous .copy_mc crap entirely from iov_iter

I think this solution has two impacts:
1. Although it is not a performance-critical path, the CPU usage may be
affected by one more memory copy in some large-memory applications.
2. If a hardware memory error occurs in "good location" and the
".copy_mc" is removed, the kernel will panic.

I would prefer to use the previous solution(modify the implementation of
memcpy_from_iter_mc()).

Thanks,
Tong.

> This is slightly complicated by the fact that copy_mc_to_kernel() may
> not even exist, and architectures that don't have it don't want the
> silly extra copy. So we need to abstract the "copy to temporary page"
> code a bit. But that's probably a good thing anyway in that it forces
> us to have nice interfaces.
>
> End result: something like the attached.
>
> AGAIN: THIS IS ENTIRELY UNTESTED.
>
> But hey, so was clearly all the .copy_mc code too that this removes, so...
>
>                 Linus