Re: [PATCH v2] mm/mmap: Regression fix for unmapped_area{_topdown}

[Liam R. Howlett]

* Michael Keyes <mgkeyes@xxxxxxxxxxxxxxxxxxxx> [230430 18:41]:
> On 29.04.23 15:32, Tad wrote:
> > This reintroduces the issue described in
> > https://lore.kernel.org/linux-mm/cb8dc31a-fef2-1d09-f133-e9f7b9f9e77a@xxxxxxxx/
> Yes, I also ran into this (even though I'd somehow missed it the
> previous time).

Rick Edgecombe reported something similar [1].

This is probably to do with my stack guard checks I recently added.

> 
> Apparently the issue arises at mm/mmap.c:1582, where low_limit is set to
> vm_end_gap(tmp). Occasionally, this returns a 64-bit address (e.g.
> 0x7fedea581000), which is obviously greater than high_limit for a 32-bit
> mmap, and causes the next call to mas_empty_area() to fail.
> 
> I'm not sure why vm_end_gap(tmp) occasionally returns a 64-bit address,
> or if the best solution is to just check for this and skip the retry if
> it occurs…
> 

Thanks for the debugging.  I will look into it.

I am currently trying to revise how the iterators, prev/next deal with
shifting outside the requested limits.  I suspect it's something to do
with hitting the limit and what someone would assume the next operation
means.

[1] https://lore.kernel.org/linux-mm/32f156ba80010fd97dbaf0a0cdfc84366608624d.camel@xxxxxxxxx/