Re: [PATCHv4 00/25] THP-enabled tmpfs/shmem

[Hugh Dickins]

On Fri, 25 Mar 2016, Kirill A. Shutemov wrote:
> On Thu, Mar 24, 2016 at 12:08:55PM -0700, Hugh Dickins wrote:
> > On Thu, 24 Mar 2016, Kirill A. Shutemov wrote:
> > > On Wed, Mar 23, 2016 at 01:09:05PM -0700, Hugh Dickins wrote:
> > > > The small files thing formed my first impression.  My second
> > > > impression was similar, when I tried mmap(NULL, size_of_RAM,
> > > > PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED, -1, 0) and
> > > > cycled around the arena touching all the pages (which of
> > > > course has to push a little into swap): that soon OOMed.
> > > > 
> > > > But there I think you probably just have some minor bug to be fixed:
> > > > I spent a little while trying to debug it, but then decided I'd
> > > > better get back to writing to you.  I didn't really understand what
> > > > I was seeing, but when I hacked some stats into shrink_page_list(),
> > > > converting !is_page_cache_freeable(page) to page_cache_references(page)
> > > > to return the difference instead of the bool, a large proportion of
> > > > huge tmpfs pages seemed to have count 1 too high to be freeable at
> > > > that point (and one huge tmpfs page had a count of 3477).
> > > 
> > > I'll reply to your other points later, but first I wanted to address this
> > > obvious bug.
> > 
> > Thanks.  That works better, but is not yet right: memory isn't freed
> > as it should be, so when I exit then try to run a second time, the
> > mmap() just gets ENOMEM (with /proc/sys/vm/overcommit_memory 0):
> > MemFree is low.  No rush to fix, I've other stuff to do.
> > 
> > I don't get as far as that on the laptop, since the first run is OOM
> > killed while swapping; but I can't vouch for the OOM-kill-correctness
> > of the base tree I'm using, and this laptop has a history of OOMing
> > rather too easily if all's not right.
> 
> Hm. I don't see the issue.
> 
> I tried to reproduce it in my VM with following script:
> 
> #!/bin/sh -efu
> 
> swapon -a
> 
> ram="$(grep MemTotal /proc/meminfo | sed 's,[^0-9\]\+,,; s, kB,k,')"
> 
> usemem -w -f /dev/zero "$ram"
> 
> swapoff -a
> swapon -a
> 
> usemem -w -f /dev/zero "$ram"
> 
> cat /proc/meminfo
> grep thp /proc/vmstat
> 
> -----
> 
> usemem is a tool from this archive:
> 
> http://www.spinics.net/lists/linux-mm/attachments/gtarazbJaHPaAT.gtar
> 
> It works fine even if would double size of mapping.
> 
> Do you have a reproducer?

Yes, my reproducer is simpler (just cycling twice around the arena,
touching each page in order); and I too did not see it running your
script using usemem above.  It looks as if that invocation isn't doing
enough work with swap: if I add a "-r 2" to those usemem lines, then
I get "usemem: mmap failed: Cannot allocate memory" on the second.

I also added a "sleep 2" before the second call to usemem: I'm not sure
of the current state of vmstat, but historically it's slow to gather
back from each cpu to global, and I think it used to leave some cpu
counts stranded indefinitely once upon a time.  In my own testing,
I have a /proc/sys/vm/stat_refresh to touch before checking meminfo
or vmstat - and I think the vm_enough_memory() check in mmap() may
need that same care, since it refers to NR_FREE_PAGES etc.

8GB is my ramsize, if that matters.

Hugh