Re: [PATCH v3 0/1] mm: introduce put_user_page*(), placeholder versions

[John Hubbard]

On 3/14/19 1:25 PM, William Kucharski wrote:
> > On Mar 14, 2019, at 7:30 AM, Jan Kara <jack@xxxxxxx> wrote:
> >
> > Well I have some crash reports couple years old and they are not from QA
> > departments. So I'm pretty confident there are real users that use this in
> > production... and just reboot their machine in case it crashes.
>
> Do you know what the use case in those crashes actually was?
>
> I'm curious to know they were actually cases of say DMA from a video
> capture card or if the uses posited to date are simply theoretical.


It's not merely theoretical. In addition to Jan's bug reports, I've
personally investigated a bug that involved an GPU (acting basically as
an AI accelerator in this case) that was doing DMA to memory that turned
out to be file backed.

The backtrace for that is in the commit description.

As others have mentioned, this works well enough to lure people into
using it, but then fails when you load down a powerful system (and put
it under memory pressure).

I think that as systems get larger, and more highly threaded, we might
see more such failures--maybe even in the Direct IO case someday,
although so far that race window is so small that that one truly is
still theoretical (or, we just haven't been in communication with
anyone who hit it).

thanks,
--
John Hubbard
NVIDIA

> It's always good to know who might be doing this and why if for no other
> reason than as something to keep in mind when designing future interfaces.