Re: [RFC PATCH v3 2/3] grep: make PCRE2 aware of custom allocator

From: Carlo Arenas <carenas@gmail.com>
To: "René Scharfe" <l.s.r@web.de>
Cc: git@vger.kernel.org, gitster@pobox.com,
	johannes.schindelin@gmx.de, avarab@gmail.com,
	michal.kiedrowicz@gmail.com
Subject: Re: [RFC PATCH v3 2/3] grep: make PCRE2 aware of custom allocator
Date: Wed, 7 Aug 2019 02:49:56 -0700	[thread overview]
Message-ID: <CAPUEspip98Mq8FrKTOkEikZhaLPprZXf=E2x3d0b7=c7e5+Gyw@mail.gmail.com> (raw)
In-Reply-To: <ab8a378c-0a60-9554-b2dd-ecb3d05229cb@web.de>

On Tue, Aug 6, 2019 at 10:38 PM René Scharfe <l.s.r@web.de> wrote:
>
> Am 06.08.19 um 18:36 schrieb Carlo Marcelo Arenas Belón:
> > Move some of the logic that was before done per thread (in the workers)
> > into an earlier phase to avoid degrading performance
>
> Which logic is moved?  In the patch I basically only see additions.

the decision of if we want to create a global context or not, which is now being
done only once per pattern.  "moving" there was to imply the logic now
was no longer going to always use state in the struct_pat that then
will be evaluated at least once per thread, but could use state
globally and reuse some of the work (as I did for the generation of
the chartables in a later patch, that I had yet to send)

> >, but as the use
> > of PCRE2 with NED is better understood it is expected more of its
> > functions will be instructed to use the custom allocator as well as
> > was done in the original code[1] this work was based on.
> >
> > [1] https://public-inbox.org/git/3397e6797f872aedd18c6d795f4976e1c579514b.1565005867.git.gitgitgadget@gmail.com/
>
> I'm not sure I understand that part.  Do you mean there are gaps of
> knowledge about nedmalloc and/or PCRE2 and/or their interaction?  And
> someone is working on closing those gaps, and is going to submit more
> patches in the process?
>
> Your patch uses a custom global context only if USE_NED_ALLOCATOR is
> defined, while [1] does it unconditionally.  The latter is easier to
> debug and requires less preprocessor directives.  What's the upside
> of your approach?

was hoping will perform better but it seems that testing can be done
only in windows

> > +#ifdef USE_NED_ALLOCATOR
> > +static void *pcre2_malloc(PCRE2_SIZE size, MAYBE_UNUSED void *memory_data)
> > +{
> > +     return malloc(size);
>
> Should this be xmalloc() to get consistent out-of-memory handling?

good point, note though that since it is inside a USE_NED_ALLOCATOR
ifdef it is really
nedalloc in disguise
>
> > +}
> > +
> > +static void pcre2_free(void *pointer, MAYBE_UNUSED void *memory_data)
> > +{
> > +     return free(pointer);
> > +}
> > +#endif
> > +#endif
> > +
> >  static const char *color_grep_slots[] = {
> >       [GREP_COLOR_CONTEXT]        = "context",
> >       [GREP_COLOR_FILENAME]       = "filename",
> > @@ -153,6 +169,7 @@ int grep_config(const char *var, const char *value, void *cb)
> >   *
> >   * If using PCRE make sure that the library is configured
> >   * to use the right allocator (ex: NED)
> > + * if any object is created it should be cleaned up in grep_destroy()
> >   */
> >  void grep_init(struct grep_opt *opt, struct repository *repo, const char *prefix)
> >  {
> > @@ -188,6 +205,13 @@ void grep_init(struct grep_opt *opt, struct repository *repo, const char *prefix
> >               color_set(opt->colors[i], def->colors[i]);
> >  }
> >
> > +void grep_destroy(void)
> > +{
> > +#ifdef USE_LIBPCRE2
> > +     pcre2_general_context_free(pcre2_global_context);
> > +#endif
> > +}
> > +
> >  static void grep_set_pattern_type_option(enum grep_pattern_type pattern_type, struct grep_opt *opt)
> >  {
> >       /*
> > @@ -319,6 +343,11 @@ void append_header_grep_pattern(struct grep_opt *opt,
> >  void append_grep_pattern(struct grep_opt *opt, const char *pat,
> >                        const char *origin, int no, enum grep_pat_token t)
> >  {
> > +#if defined(USE_LIBPCRE2) && defined(USE_NED_ALLOCATOR)
> > +     if (!pcre2_global_context && opt->ignore_case && has_non_ascii(pat))
> > +             pcre2_global_context = pcre2_general_context_create(
> > +                                     pcre2_malloc, pcre2_free, NULL);
> > +#endif
> >       append_grep_pat(opt, pat, strlen(pat), origin, no, t);
> >  }
> >
> > @@ -507,9 +536,14 @@ static void compile_pcre2_pattern(struct grep_pat *p, const struct grep_opt *opt
> >
> >       p->pcre2_compile_context = NULL;
> >
> > +     /* pcre2_global_context is initialized in append_grep_pattern */
> >       if (opt->ignore_case) {
> >               if (has_non_ascii(p->pattern)) {
> > -                     character_tables = pcre2_maketables(NULL);
> > +#ifdef USE_NED_ALLOCATOR
> > +                     if (!pcre2_global_context)
> > +                             BUG("pcre2_global_context uninitialized");
>
> [1] initializes on demand.  Why not do that?  To avoid race conditions

this was just for help migrating the code, could go away even now, but
though it was nice to keep for safety (as someone mentioned recently)

> that would lead to occasional double-allocation of the global context?

and a nicer API that allows for cleaning up any global objects (with
the chartables being an easy one to tackle next) and might
help in the future to make the worker threads less messy (ex: compile vs match)

> > +#endif
> > +                     character_tables = pcre2_maketables(pcre2_global_context);
> >                       p->pcre2_compile_context = pcre2_compile_context_create(NULL);
>
> Don't you want to pass pcre2_global_context here as well?  And [1] even
> uses it in some more places.
>
> Oh, that's the "expected more" when "better understood" part from the
> commit message, right?

correct, was trying to be conservative and minimal (since this code
will conflict as well
with other in the fly changes), but considering that the last Azure
build from pu still failed
and I have no windows box to debug, might need to do it anyway.

> Basically I'd expect the custom global context to be used for all PCRE2
> allocations; I can't think of a reason for mixing allocators (e.g.
> system malloc for PCRE2 regexes and nedmalloc for everything else).

ok

Carlo