Just to be sure that we are following up on the right path, I added the hack
listed below making the JIT memory allocation shared. Then the os.fork() call
works just fine for a simple test program (also included below). Obviously
that's not an actual solution as the parent and the child will need separate
JIT memory allocations, but it pinpoints the issue. I will issue a report to
the Cygwin folks, however a fix may or may not be slow in the coming. How hard
would it be for Pypy to use malloc() instead of mmap() for Cygwin?. -- Uwe
*** pypy/pypy/rlib/rmmap.py Wed May 30 08:11:31 2012
--- pypy/pypy/rlib/rmmap.py Sat Jun 9 16:23:45 2012
***************
*** 14,19 ****
--- 14,20 ----
_MS_WINDOWS = os.name == "nt"
_LINUX = "linux" in sys.platform
_64BIT = "64bit" in platform.architecture()[0]
+ _CYGWIN = "cygwin" == sys.platform
class RValueError(Exception):
def __init__(self, message):
***************
*** 692,698 ****
so the memory has the executable bit set and gets allocated
internally in case of a sandboxed process.
"""
! flags = MAP_PRIVATE | MAP_ANONYMOUS
prot = PROT_EXEC | PROT_READ | PROT_WRITE
hintp = rffi.cast(PTR, hint.pos)
res = c_mmap_safe(hintp, map_size, prot, flags, -1, 0)
--- 693,703 ----
so the memory has the executable bit set and gets allocated
internally in case of a sandboxed process.
"""
! if _CYGWIN:
! # XXX Hack: JIT memory should be private but Cygwin's fork() fails
! flags = MAP_SHARED | MAP_ANONYMOUS
! else:
! flags = MAP_PRIVATE | MAP_ANONYMOUS
prot = PROT_EXEC | PROT_READ | PROT_WRITE
hintp = rffi.cast(PTR, hint.pos)
res = c_mmap_safe(hintp, map_size, prot, flags, -1, 0)
PS Here's a little test program.
import sys, os
from test import pystone
# run some code that uses the JIT
print "top-level process:", pystone.main()
print "top-level process:", pystone.main()
pid = os.fork()
if pid == 0:
print "I'm the child, my PID is", os.getpid()
print "child process:", pystone.main()
print "child process:", pystone.main()
sys.exit()
else:
print "I'm the parent, my PID is", os.getpid(), " the child PID is", pid
print "parent process:", pystone.main()
print "parent process:", pystone.main()
sys.exit()
________________________________
From: Armin Rigo <ar...@tunes.org>
To: Uwe F. Mayer <uwe_f_ma...@yahoo.com>
Cc: Amaury Forgeot d'Arc <amaur...@gmail.com>; Matti Picus
<matti.pi...@gmail.com>; "pypy-dev@python.org" <pypy-dev@python.org>
Sent: Thursday, June 7, 2012 10:27 PM
Subject: Re: [pypy-dev] Patches for Pypy under cygwin
Hi Uwe,
On Thu, Jun 7, 2012 at 9:21 PM, Uwe F. Mayer <uwe_f_ma...@yahoo.com> wrote:
> 3 [main] pypy-c 11788 fixup_mmaps_after_fork: ReadProcessMemory failed
> for MAP_PRIVATE address 0xB0010000, Win32 error 998
Yes, that's one of the mmap addresses used by the JIT, if available.
More to the point, googling for "fork cygwin mmap" shows that there
are open bugs with cygwin's fork() in the presence of MAP_PRIVATE
mmaps. That's obviously why it fails.
I don't really know what we can reasonably do about that, short of
asking people to use some specific versions of cygwin where it seems
to work. I can imagine what we could unreasonably do as work-around,
but that requires a lot of work and gross hacks that don't seem
appropriate to put in the pypy source just to support cygwin... I
would recommend that you try to see it fixed on the cygwin side.
A bientôt,
Armin.
_______________________________________________
pypy-dev mailing list
pypy-dev@python.org
http://mail.python.org/mailman/listinfo/pypy-dev
