[lkml]   [2019]   [Sep]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v5 2/2] mm,thp: Add experimental config option RO_EXEC_FILEMAP_HUGE_FAULT_THP
On Tue, Sep 03, 2019 at 02:51:50PM +0200, Michal Hocko wrote:
> On Tue 03-09-19 05:22:08, Matthew Wilcox wrote:
> > On Tue, Sep 03, 2019 at 02:14:24PM +0200, Michal Hocko wrote:
> > > On Mon 02-09-19 03:23:41, William Kucharski wrote:
> > > > Add filemap_huge_fault() to attempt to satisfy page
> > > > faults on memory-mapped read-only text pages using THP when possible.
> > >
> > > This deserves much more description of how the thing is implemented and
> > > expected to work. For one thing it is not really clear to me why you
> > > need CONFIG_RO_EXEC_FILEMAP_HUGE_FAULT_THP at all. You need a support
> > > from the filesystem anyway. So who is going to enable/disable this
> > > config?
> >
> > There are definitely situations in which enabling this code will crash
> > the kernel. But we want to get filesystems to a point where they can
> > start working on their support for large pages. So our workaround is
> > to try to get the core pieces merged under a CONFIG_I_KNOW_WHAT_IM_DOING
> > flag and let people play with it. Then continue to work on the core
> > to eliminate those places that are broken.
> I am not sure I understand. Each fs has to opt in to the feature
> anyway. If it doesn't then there should be no risk of regression, right?
> I do not expect any fs would rush an implementation in while not being
> sure about the correctness. So how exactly does a config option help
> here.

Filesystems won't see large pages unless they've opted into them.
But there's a huge amount of page-cache work that needs to get done
before this can be enabled by default. For example, truncate() won't
work properly.

Rather than try to do all the page cache work upfront, then wait for the
filesystems to catch up, we want to get some basics merged. Since we've
been talking about this for so long without any movement in the kernel
towards actual support, this felt like a good way to go.

We could, of course, develop the entire thing out of tree, but that's
likely to lead to pain and anguish.

 \ /
  Last update: 2019-09-03 17:10    [W:0.056 / U:5.392 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site