Re: [PATCH userspace v2] libsepol: cache ebitmap cardinality value

[Ondrej Mosnacek <omosnace@redhat.com>]

On Tue, Feb 25, 2020 at 10:57 PM William Roberts
<bill.c.roberts@gmail.com> wrote:
> On Tue, Feb 25, 2020 at 3:33 PM Nicolas Iooss <nicolas.iooss@m4x.org> wrote:
> >
> > On Tue, Feb 18, 2020 at 5:01 PM Ondrej Mosnacek <omosnace@redhat.com> wrote:
> > >
> > > On Tue, Feb 18, 2020 at 4:40 PM Stephen Smalley <sds@tycho.nsa.gov> wrote:
> > > > On 2/18/20 10:22 AM, Ondrej Mosnacek wrote:
> > > > > On Thu, Feb 13, 2020 at 2:40 PM Ondrej Mosnacek <omosnace@redhat.com> wrote:
> > > > >> According to profiling of semodule -BN, ebitmap_cardinality() is called
> > > > >> quite often and contributes a lot to the total runtime. Cache its result
> > > > >> in the ebitmap struct to reduce this overhead. The cached value is
> > > > >> invalidated on most modifying operations, but ebitmap_cardinality() is
> > > > >> usually called once the ebitmap doesn't change any more.
> > > > >>
> > > > >> After this patch, the time to do 'semodule -BN' on Fedora Rawhide has
> > > > >> decreased from ~14.6s to ~12.4s (2.2s saved).
> > > > >
> > > > > I have no idea why, but I'm now getting completely different times
> > > > > (10.9s vs. 8.9s) with the same builds on the same setup... I can no
> > > > > longer reproduce the slower times anywhere (F31/locally/...) so I have
> > > > > to assume it was some kind of glitch. Since the numbers show a similar
> > > > > magnitude of speed-up (and they depend on a bunch of HW/SW factors
> > > > > anyway), I'm not going to do another respin. The applying person (most
> > > > > likely Stephen) is free to fix the numbers when applying if they wish
> > > > > to do so.
> > > >
> > > > Thanks, applied with fixed times (although I don't really think it
> > > > matters very much).  Maybe you're also picking up the difference from
> > > > the "libsepol/cil: remove unnecessary hash tables" change.
> > >
> > > No, that was actually the reason for the first correction.
> >
> > Hello,
> > About performance issues, the current implementation of
> > ebitmap_cardinality() is quadratic:
> >
> > for (i=ebitmap_startbit(e1); i < ebitmap_length(e1); i++)
> >     if (ebitmap_get_bit(e1, i))
> >         count++;
> >
> > ... because ebitmap_get_bit() browse the bitmap:
> >
> > while (n && (n->startbit <= bit)) {
> >    if ((n->startbit + MAPSIZE) > bit) {
> >       /*... */

Hm... I didn't realize that the function is actually quadratic.

> >
> > A few years ago, I tried modifying this function to make it linear in
> > the bitmap size:
> >
> > unsigned int ebitmap_cardinality(ebitmap_t *e1)
> > {
> >     unsigned int count = 0;
> >     ebitmap_node_t *n;
> >
> >    for (n = e1->node; n; n = n->next) {
> >         count += __builtin_popcountll(n->map);
> >     }
> >     return count;
> > }
> >
> > ... but never actually sent a patch for this, because I wanted to
> > assess how __builtin_popcountll() was supported by several compilers
> > beforehand. Would this be helpful to gain even more performance gain?
>
> Every architecture I've used has an instruction it boils down to:
> x86 - POPCNT
> ARM (neon): vcnt

Note that the compiler will only emit these instructions if you
compile with the right target platform (-mpopcnt or something that
includes it on x86_64). Portable generic builds will usually not use
it. Still, even without the special instruction __builtin_popcountll()
should generate more optimal code than the naive
add-each-bit-one-by-one approach. For example, I came up with this
pure C implementation of 64-bit popcount [1] that both GCC and Clang
can compile down to ~36 instructions. The generic version of
__builtin_popcountll() likely does something similar. (Actually, here
is what Clang seems to use [2], which is pretty close.)

FWIW, I tested the __builtin_popcountll() version with the caching
patch reverted (built without popcnt support) and it actually
performed even better than the old code + caching (it went down to
~0.11% of semodule -B running time). A naive popcount implementation
without caching didn't perform as good (was slower than the old code +
caching).

So... we could just open-code some good generic C implementation
(cleanly written and properly commented, of course) and then we
wouldn't have to rely on the compiler builtin. OTOH, the SELinux
userspace already uses non-standard compiler extensions
(__attribute__(...)), so maybe sticking to pure C is not worth it...
Either way I think we should revert the caching patch along with
switching to an optimized implementation (it would no longer be worth
the added complexity IMO).

[1] https://gcc.godbolt.org/z/39W7qa
[2] https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/popcountdi2.c

>
> For others, (do they even matter at this point) I would imagine GCC
> does something relatively sane.
>
> >
> > Cheers,
> > Nicolas
> >
>
--
Ondrej Mosnacek <omosnace at redhat dot com>
Software Engineer, Security Technologies
Red Hat, Inc.