Re: [PATCH 0/2] fs, proc: optimize smaps output formatting
From: Joe Perches
Date: Sat Aug 20 2016 - 03:56:06 EST
On Sat, 2016-08-20 at 09:29 +0200, Michal Hocko wrote:
> On Fri 19-08-16 10:43:15, Joe Perches wrote:
> >
> > On Fri, 2016-08-19 at 12:12 +0200, Michal Hocko wrote:
> > >
> > > Hi,
> > > this is rebased on top of next-20160818. Joe has pointed out that
> > > meminfo is using a similar trick so I have extracted guts of what we
> > > have already and made it more generic to be usable for smaps as well
> > > (patch 1). The second patch then replaces seq_printf with seq_write
> > > and show_val_kb which should have smaller overhead and my measuring (in
> > > kvm) shows quite a nice improvements. I hope kvm is not playing tricks
> > > on me but I didn't get to test on a real HW.
> >
> > Hi Michal.
> >
> > A few comments:
> >
> > For the first patch:
> >
> > I think this isn't worth the expansion in object size (x86-64 defconfig)
> >
> > $ size fs/proc/meminfo.o*
> > text data bss dec hex filename
> > 2698 8 0 2706 a92 fs/proc/meminfo.o.new
> > 2142 8 0 2150 866 fs/proc/meminfo.o.old
> >
> > Creating a new static in task_mmu would be smaller and faster code.
> Hmm, nasty...
> add/remove: 0/0 grow/shrink: 2/1 up/down: 1081/-24 (1057)
> function old new delta
> meminfo_proc_show 1134 1745 +611
> show_smap 560 1030 +470
> show_val_kb 140 116 -24
> Total: Before=91716, After=92773, chg +1.15%
>
> it seems to be calls to seq_write which blown up the size. So I've tried
> to put seq_write back to show_val_kb and did only sizeof() inside those
> macros and that reduced the size but not fully back to the original code
> size. So it seems the value shifts consumed some portion of that as well.
> I've ended up with the following incremental diff which leads to
> text data bss dec hex filename
> 100728 1443 400 102571 190ab fs/proc/built-in.o.next
> 101658 1443 400 103501 1944d fs/proc/built-in.o.patched
> 100951 1443 400 102794 1918a fs/proc/built-in.o.incremental
>
> There is still some increase wrt. the baseline but I guess that can be
> explained by single seq_printf -> many show_name_val_kb calls.
>
> If that looks acceptable I will respin both patches. I would really
> like to prefer to not duplicate show_val_kb into task_mmu as much as
> possible, though.
I think the patch set I'll send you in a few minutes
will speed up /proc/<pid>/smaps a whole lot more.
Please test it using your little test bench.
cheers, Joe