Heavily revised, comprising revisions:
46640 - original trunk revision (backed out in r46655)
46647 - markup fix (backed out in r46655)
46692:46918 merged from branch aimacintyre-sf1454481
branch tested on buildbots (Windows buildbots had problems
not related to these changes).
46640 Patch #1454481: Make thread stack size runtime tunable.
46647 Markup fix
The first is causing many buildbots to fail test runs, and there
are multiple causes with seemingly no immediate prospects for
repairing them. See python-dev discussion.
Note that a branch can (and should) be created for resolving these
problems, like
svn copy svn+ssh://svn.python.org/python/trunk -r46640 svn+ssh://svn.python.org/python/branches/NEW_BRANCH
followed by merging rev 46647 to the new branch.
behavior, creating many threads very quickly. A long debugging session
revealed that the Windows implementation of PyThread_start_new_thread()
was choked with "laziness" errors:
1. It checked MS _beginthread() for a failure return, but when that
happened it returned heap trash as the function result, instead of
an id of -1 (the proper error-return value).
2. It didn't consider that the Win32 CreateSemaphore() can fail.
3. When creating a great many threads very quickly, it's quite possible
that any particular bootstrap call can take virtually any amount of
time to return. But the code waited for a maximum of 5 seconds, and
didn't check to see whether the semaphore it was waiting for got
signaled. If it in fact timed out, the function could again return
heap trash as the function result. This is actually what confused
the test program, as the heap trash usually turned out to be 0, and
then multiple threads all got id 0 simultaneously, confusing the
hell out of threading.py's _active dict (mapping id to thread
object). A variety of baffling behaviors followed from that.
WRT #1 and #2, error returns are checked now, and "thread.error: can't
start new thread" gets raised now if a new thread (or new semaphore)
can't be created. WRT #3, we now wait for the semaphore without a
timeout.
Also removed useless local vrbls, folded long lines, and changed callobj
to a stack auto (it was going thru malloc/free instead, for no discernible
reason).
Bugfix candidate.
A file-static "threads" dict mapped thread IDs to Windows handles, but
was never referenced, and entries never got removed. This gets rid of
the YAGNI-dict entirely.
Bugfix candidate.
This changes Pythread_start_thread() to return the thread ID, or -1
for an error. (It's technically an incompatible API change, but I
doubt anyone calls it.)
pyport.h: typedef a new Py_intptr_t type.
DELICATE ASSUMPTION: That HAVE_UINTPTR_T implies intptr_t is
available as well as uintptr_t. If that turns out not to be
true, things must get uglier (C99 wants both, so I think it's
an assumption we're *likely* to get away with).
thread_nt.h, PyThread_start_new_thread: MS _beginthread is documented
as returning unsigned long; no idea why uintptr_t was being used.
Others: Always use Py_[u]intptr_t, never [u]intptr_t directly.
The common technique for printing out a pointer has been to cast to a long
and use the "%lx" printf modifier. This is incorrect on Win64 where casting
to a long truncates the pointer. The "%p" formatter should be used instead.
The problem as stated by Tim:
> Unfortunately, the C committee refused to define what %p conversion "looks
> like" -- they explicitly allowed it to be implementation-defined. Older
> versions of Microsoft C even stuck a colon in the middle of the address (in
> the days of segment+offset addressing)!
The result is that the hex value of a pointer will maybe/maybe not have a 0x
prepended to it.
Notes on the patch:
There are two main classes of changes:
- in the various repr() functions that print out pointers
- debugging printf's in the various thread_*.h files (these are why the
patch is large)
Closes SourceForge patch #100505.
who wrote:
Here's the new version of thread_nt.h. More particular, there is a
new version of thread lock that uses kernel object (e.g. semaphore)
only in case of contention; in other case it simply uses interlocked
functions, which are faster by the order of magnitude. It doesn't
make much difference without threads present, but as soon as thread
machinery initialised and (mostly) the interpreter global lock is on,
difference becomes tremendous. I've included a small script, which
initialises threads and launches pystone. With original thread_nt.h,
Pystone results with initialised threads are twofold worse then w/o
threads. With the new version, only 10% worse. I have used this
patch for about 6 months (with threaded and non-threaded
applications). It works remarkably well (though I'd desperately
prefer Python was free-threaded; I hope, it will soon).
ExitThread(). As discussed in c.l.p, this takes care of
initialization and finalization of thread-local storage allocated by
the C runtime system. Not sure whether non-MS compilers grok this
though (but who cares :-).
Kristján Valur Jónsson
Andrew MacIntyre
Martin v. Löwis
Tim Peters
Georg Brandl
Fred Drake
Guido van Rossum
Fredrik Lundh
Thomas Wouters