Re: [PATCH v2 0/3] vmalloc enhancements

[Roman Gushchin]

On Tue, Feb 12, 2019 at 12:34:09PM -0800, Andrew Morton wrote:
> On Tue, 12 Feb 2019 13:47:24 -0500 Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> 
> > On Tue, Feb 12, 2019 at 09:56:45AM -0800, Roman Gushchin wrote:
> > > The patchset contains few changes to the vmalloc code, which are
> > > leading to some performance gains and code simplification.
> > > 
> > > Also, it exports a number of pages, used by vmalloc(),
> > > in /proc/meminfo.
> > > 
> > > Patch (1) removes some redundancy on __vunmap().
> > > Patch (2) separates memory allocation and data initialization
> > >   in alloc_vmap_area()
> > > Patch (3) adds vmalloc counter to /proc/meminfo.
> > > 
> > > v2->v1:
> > >   - rebased on top of current mm tree
> > >   - switch from atomic to percpu vmalloc page counter
> > 
> > I don't understand what prompted this change to percpu counters.
> > 
> > All writers already write vmap_area_lock and vmap_area_list, so it's
> > not really saving much. The for_each_possible_cpu() for /proc/meminfo
> > on the other hand is troublesome.
> 
> percpu_counters would fit here.  They have probably-unneeded locking
> but I expect that will be acceptable.
> 
> And they address the issues with for_each_possible_cpu() avoidance, CPU
> hotplug and transient negative values.

Using existing vmap_area_lock (as Johannes suggested) is also problematic,
due to different life-cycles of vma_areas and vmalloc pages. A special flag
will be required to decrease the counter during the lazy deletion of
vmap_areas. Allocation path will require passing a bool flag through too many
nested functions. Also it will be semi-accurate, which is probably tolerable.
So, it's doable, but doesn't look nice to me.

So, using a simple per-cpu counter still seems to best option.
Transient negative value is a valid concern, but easily fixable.
Are there any other? What's the problem with for_each_possible_cpu()?
Reading /proc/meminfo is not that hot, no?

Thanks!