Re: Linux 4.9-rc6

[Al Viro]

On Sun, Nov 20, 2016 at 03:27:07PM -0800, Linus Torvalds wrote:
> On Sun, Nov 20, 2016 at 2:27 PM, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> >
> > Hosts with ~100,000 threads have an issue with /prov/vmallocinfo
> >
> > It can take about 800 usec to skip over ~100,000 struct vmap_area
> > in s_start(), while holding vmap_area_lock spinlock, and therefore
> > blocking fork()/pthread_create().
> >
> > I presume we can not switch to the rbtree (vmap_area_root)
> > for /proc/vmallocinfo, because this file is seek-able, right ?
> 
> Well, the good news is that the file is root-only anyway, which means
> that at least it won't have the issue that a lot of other /proc files
> have had - namely being opened by random user programs or libraries.
> 
> Which means that the users of it are likely fairly limited.
> 
> Which in turn means that we can probably afford to play more games
> with it. Including, for example, possibly marking it non-seekable.
> 
> Or even just limit the maximum entries we are willing to walk.
> 
> Or we could decide that that file shouldn't be a seq_file at all, use
> the old "one page buffer" approach that was so common for /proc files,
> and make the position encode the vmalloc address in it (make the lower
> PAGE_MASK bits be the offset in the line), and then we *could* just
> look things up using the btree method.
> 
> Al, do you have any clever ideas?

Umm...  One possibility would be something like fs/namespace.c:m_start() -
if nothing has changed since the last time, just use a cached pointer.
That has sped the damn thing (/proc/mounts et.al.) big way, but it's
dependent upon having an event count updated whenever we change the
mount tree - doing the same for vma_area list might or might not be
a good idea.  /proc/mounts and friends get ->poll() on that as well;
that probably would _not_ be a good idea in this case.