Request-global hashtable initialization

On Thu, Apr 12, 2018 at 9:53 PM, Heinz Wiesinger HMWiesinger@liwjatan.at

wrote:

Hi everyone!

I ventured into writing my first PHP extension and did pretty well so
far
following the existing documentation and example code inside the php
source.

However, I hit an odd bug in my code that I can't quite explain. I'm
obviously
missing something and I hope someone here can point me in the right
direction

:)

I'd like to set up a request global hashtable. For now I'm running
ALLOC_HASHTABLE() and zend_hash_init() inside PHP_GINIT_FUNCTION. I then
have
a user function that's supposed to add entries to that hashtable and
another
user function that reads entries from the hashtable.

I tested with code passed directly on the command line (using php -r)
and
that
works just like I would expect it too. The value gets written to the
hashtable
and I can read it afterwards from the other function. I also verified
it's
the
same value that I first stored, so from what I can tell the basic
functionality is good.

When I take the same testcode and write it into a file, and then run php
on
that file, execution aborts trying to allocate almost 2GB of memory when
trying to store the value in the hashtable (using zend_hash_update()). I
can
see PHP_GINIT_FUNCTION being executed in that case as well so I'd assume
the
hashtable is setup just fine. Yet something is different.

What's the difference between " php -r '' " and "php test.php" in that
regard?

I uploaded my current state of the extension to github in case someone
would
like to take a closer look at the code: https://github.com/pprkut/
autoload-psr

Thanks in advance!

Grs,
Heinz

https://github.com/pprkut/autoload-psr/blob/27cfc53c864116cf0b65b8f7e69464
e1 5b3fdef5/autoload_psr.c#L175 should be using ZVAL_STR_COPY. The
hashtable API does not automatically incref stored values.

I'd suggest running under USE_ZEND_ALLOC=0 valgrind to check for further
memory issues. Differences in the mode of execution are likely just
incidental.

Thanks for the pointers! I tried running it under valgrind, but couldn't
find anything. So I tried without valgrind, but with USE_ZEND_ALLOC=0 and
found that I can't reproduce the problem like that. Looks to me that
something I do doesn't play well with the zend allocator.

Incidentally valgrind also doesn't tell me about the issue with ZVAL_STR
that you pointed out. Maybe I'm using it wrong? I tried

USE_ZEND_ALLOC=0 valgrind --tool=memcheck --leak-check=full php -
dextension=modules/autoload_psr.so test2.php

I moved the hashtable initialization out of PHP_GINIT_FUNCTION and do it on-
demand now in the respective user functions. That seems to have done the
trick, as segfaults and oom messages are gone now.

Grs,
Heinz