Please add 17839856fd58 to the stable queue

Please add 17839856fd58 to the stable queue

[Linus Torvalds]

Commit 17839856fd58 ("gup: document and work around "COW can break
either way" issue") is a real fix, but wasn't marked for stable
because I wanted it to get more coverage testing in mainline first.
Not because the patch is all that complex or scary, but because I was
worried we'd find some odd case where it would make things slower by
triggering the GUP slowpath much more often due to people doing odd
things.

It turns out my worry seems to have been misplaced. The kernel test
robot did indeed trigger a case where this made a big difference, but
rather than being bad, it improved the odd corner-case test-case
performance by a factor of 20x by breaking the COW and triggering the
fast-case code that way, rather than the other way around.

See

  https://lore.kernel.org/lkml/20200611040453.GK12456@shao2-debian/

for details.

So that commit fixes a bug, isn't expected to really make any
difference on any sane workload, and can apparently help the crazy
cases by a huge amount. Let's just push it to stable..

            Linus

Re: Please add 17839856fd58 to the stable queue

[Greg Kroah-Hartman]

On Mon, Jun 15, 2020 at 12:18:48PM -0700, Linus Torvalds wrote:
> Commit 17839856fd58 ("gup: document and work around "COW can break
> either way" issue") is a real fix, but wasn't marked for stable
> because I wanted it to get more coverage testing in mainline first.
> Not because the patch is all that complex or scary, but because I was
> worried we'd find some odd case where it would make things slower by
> triggering the GUP slowpath much more often due to people doing odd
> things.
> 
> It turns out my worry seems to have been misplaced. The kernel test
> robot did indeed trigger a case where this made a big difference, but
> rather than being bad, it improved the odd corner-case test-case
> performance by a factor of 20x by breaking the COW and triggering the
> fast-case code that way, rather than the other way around.
> 
> See
> 
>   https://lore.kernel.org/lkml/20200611040453.GK12456@shao2-debian/
> 
> for details.
> 
> So that commit fixes a bug, isn't expected to really make any
> difference on any sane workload, and can apparently help the crazy
> cases by a huge amount. Let's just push it to stable..

Will go do that now, was waiting to see if this caused any problems,
glad to see it hasn't.

greg k-h