[alepacheco]

Fixes #100.

libppd is used from several threads at once by Printer Applications built on pappl-retrofit, but it is built without its thread support and keeps some state process-wide. Three commits:

1. Build with POSIX threads so the library's locks and per-thread data work. configure checks for <pthread.h> (defining HAVE_PTHREAD_H) and for the library providing pthread_key_create(). The existing mutexes in ppd/string.c and ppd/libcups2.c then lock, and ppdGlobals() (with it ppdLastError() and ppdSetConformance()) is per thread, as in CUPS. Single-threaded programs see no change.
2. Keep the raster error buffer per thread. ppd/raster-error.c keeps each thread's buffer behind a thread key, as ppdGlobals() does (a plain static buffer when built without pthreads), allocated on first use and freed when the thread exits. It is NULL-safe, so ppdRasterExecPS() may report an error before any ppdRasterInterpretPPD() call.
3. Swap the PPD cache status message under a lock. set_error() in ppd/ppd-cache.c builds the new message first, swaps it in under a mutex and frees the previous one afterwards.

The new test ppd/test_ppd_thread.c, run by make check, checks ppdLastError() per thread, runs 8 threads x 2000 loops of the string pool, of ppdRasterInterpretPPD() / ppdRasterExecPS() with their exact error messages, and of failing ppdCacheCreateWithFile() calls, and calls ppdRasterExecPS() before any other raster function. On master it fails (and crashes in a plain GCC -O2 build); with these commits every check passes in 3 of 3 runs under ThreadSanitizer (no warnings), AddressSanitizer with UndefinedBehaviorSanitizer, and a plain GCC -O2 build, and make check passes.

[tillkamppeter]

Thanks a lot for this PR.

When I copied all the PPD-related functions from libcups into libppd, I completely overlooked the thread support and so I did not overtake the appropriate check into configure.ac, leaving libppd always building without thread support. And you even found that the thread support had some bugs and you have fixed them. Great. Could you fix these bugs also in the original libcups (2.4.x and 2.5.x/master in the "cups" repo). Thanks.

[alepacheco]

Thanks for merging!

I checked the CUPS repo on both 2.4.x and master, and none of the three bugs is there:

• Threads compiled out: CUPS's config-scripts/cups-threads.m4 checks for pthread.h, defines HAVE_PTHREAD_H and links the pthread library, so its mutexes (such as the string pool lock in cups/string.c) really lock. Only libppd's configure.ac was missing that check.
• Raster error buffer: in CUPS, _cupsRasterAddError(), _cupsRasterClearError() and _cupsRasterErrorString() (cupsRasterGetErrorString() on master) keep the buffer in the per-thread globals (_cupsGlobals()->raster_error), and that code handles a buffer that was never allocated. The shared buffer is only in libppd's raster-error.c.
• PPD cache status message: CUPS's ppd-cache.c reports errors through _cupsSetError(), which is per thread too. The shared ppd_cache_status_message string and set_error() exist only in libppd.

So CUPS doesn't need these fixes. The bugs came from how libppd adapted the code when it was copied out of libcups.

[tillkamppeter]

OK, so it seems that due to my overlooking of thread-readinessof the code I have filled in missing functions with very simple, non-thread-ready ones.

Thank you for checking CUPS.