Re: OSX + mmap + mansearch.c + documentation

[Ingo Schwarze <schwarze@usta.de>]

Hi Kristaps,

Kristaps Dzonsons wrote on Sat, Aug 09, 2014 at 01:51:36PM +0200:

> Small thing noticed in OSX:
> 
> % /usr/local/bin/apropos find
> mmap: Invalid argument
> 
> According to the Mac OSX manpage for mmap(2), this is because:
> 
> [EINVAL]   flags does not include either MAP_PRIVATE or MAP_SHARED.

Fixed in rc3.

That release candidate contains another fix for an issue
reported by Thomas Klausner:  config.h was forgotten in
mansearch_const.c.

> It's easy enough to add a mmap_flags variable as MAP_ANON | blah and
> ifdef it to do one or the other.  (I hesitate to say which, but am
> guessing MAP_PRIVATE since the page cache isn't (???) shared across
> processes.)

Well, even on OpenBSD the manual says that either MAP_PRIVATE or
MAP_SHARED is required, it just doesn't seem to be enforced by the
implementation, so i simply added MAP_SHARED like in all
three other places where we use mmap(2), for minimum disruption
for the release.

Now i do wonder why we use MAP_SHARED and not MAP_PRIVATE.
But that's a question for after release, i have taken a note
in the TODO file.

> In general, the section scares me--I vaguely guess at what's
> happening, but have no idea why.

Unfortunately, at one point, i accidentally deleted my gprof(1)
logs, but off the top of my head, i'm quite sure the point was
that without SQLITE_CONFIG_PAGECACHE, apropos(1) spends
considerable extra time, that is, about 10-25% of the total
time it needs, in malloc(3) and free(3) called from within
sqlite3_*() functions.

> So more pertinently than the bit-field would be to document
> this section!  Ingo mentions the speed reduction in his talk
> (pg. 20, http://www.openbsd.org/papers/bsdcan14-mandoc.pdf),
> but not /why/ there's a speed reduction.

It scares you - you mean tedu@ might suspect exploit mitigation
countermeasures?  That might be worth investigating, too.
While espie@ holds the chief SQLite developer in high esteem,
deraadt@ hates the code quality.  Myself, i didn't form an
opinion on the SQLite code, yet.

But for this release, i'll keep the lid on that can of worms.

Yours,
  Ingo
--
 To unsubscribe send an email to tech+unsubscribe@mdocml.bsd.lv