Re[2]: [Gc] boehm pohttp://win.mail.ru/cgi-bin/sentmsg?drafts&id=12950891700000000630&folder=500001rt to Native Client

Ivan Maidanski ivmai at mail.ru
Sat Jan 15 04:36:47 PST 2011


Hello Elijah,

Fri, 14 Jan 2011 15:36:34 -0800 письмо от Elijah Taylor < elijahtaylor at google.com (sentmsg?compose&amp;To=elijahtaylor at google.com) >:

> Hi Ivan,
> 
> Thanks for taking a look, I'm pleasantly surprised by the level of detail
> here.  Specific replies inline:
> 
> 
> On Fri, Jan 14, 2011 at 2:50 PM, Ivan Maidanski < ivmai at mail.ru (sentmsg?compose&amp;To=ivmai at mail.ru) > wrote:
> 
> - the patch in  naclports repository contains a typo in a macro definition
> > (MAC_TYPE -> MACH_TYPE);
> 
> Oops, will fix.

Why not to leave MACH_TYPE as-is (e.g. "I386", etc.)? NaCl is a kind of OS not a kind of machine hardware.
Is eg. I386 defined for x86?

Also, please add a mapping comment in gcconfig.h (around "Feel free to add more clauses here").

> 
> > - the patch in  naclports repository looks more suitable for gc v72 than
> > that for mono/libgc;
> >
> 
> This patch is meant to be applied to vanilla gc6.8.  The mono/libgc port is
> already patched directly into mono's code repository.

Yes, I only meant the vanilla gc6.8 patch contains some more code (eg, for PARALLEL_MARK) not present in mono/libgc.

> > - gc_pthread_redirects.h (which is a public one) should not test NACL macro
> > (or, at least, while less desirable, it should be prefixed with GC_);
> >
> 
> Ok, makes sense, I think I didn't realize this was a public header.  Though
> that explains why I had to add a test for __native_client__  (which is
> defined in our toolchain).  I'll fix this.
> 
> > - it's not clear why you need to explicitly undef STACK_GRAN, USE_M[UN]MAP,
> > etc. in gcconfig.h;
> >
> 
> I'll do some investigation, but IIRC these were needed at one point for me.
> There's a good chance these may be unnecessary and vestigial.

There should none "undef" (no other target undefining them).

If mmap is supported by NaCl then it might be possible to support USE_M[UN]MAP.

> > - is MPROTECT_VDB supported or not?;
> 
> Is MPROTECT_VDB equivalent to catching protection violations in the GC code?
> If so, then no, we don't support anything like that right now.  Protection
> violation in NaCl == instant death.

Ok, so MPROTECT_VDB (i.e., incremental/generation collection) is not supported.

> > - if you you want to port gc72 please use the recent CVS snapshot (it would
> > be easier to me to review and commit it);
> 
> I've been grabbing source from
>  http://www.hpl.hp.com/personal/Hans_Boehm/gc/gc_source/ (http://www.hpl.hp.com/personal/Hans_Boehm/gc/gc_source/)  ... can you point me
> to where I should be getting the latest?  It's not immediately obvious to
> me.

http://bdwgc.cvs.sourceforge.net/viewvc/bdwgc/bdwgc/
(For convenience, I have a recent snapshot as a tarball which I use for my project - http://www.ivmaisoft.com/_mirror/hpl/bdwgc-7_2alpha5-20110107.tar.bz2)

> - not sure that HEURISTIC1 really works reliably there (in short, HEURISTIC1
> > means you treat stack pointer at GC_init call as stack bottom - is it
> > guaranteed that GC_init call is always done at higher stack addresses than
> > any other GC call);
> >
> 
> HEURISTIC2 will definitely not work for us as it wants to use a segfault to
> detect running over the stack.  I've set STACK_GRAN to 64K, so as long as
> the stack doesn't grow beyond that size before GC_init, we should be ok,as
> right?  The stack for the main thread right now in NaCl lives at a fixed
> address usually, but that isn't guaranteed for all future time, so I'd
> prefer not to hard code magic numbers here.

No, HEURISTIC2 won't work without signals, but there are other alternatives:
- if threads-support is on then is it possible to use USE_GET_STACKBASE_FOR_MAIN?;
- is it possible to LINUX_STACKBOTTOM (if we are on real Linux)?
- for NaCl on Cygwin, it might be possible to use GC_get_[main_]stack_base based on __asm__ ("%fs:4").

> > - is the GC port compilable (and working) on other (non-Linux) platforms
> > (eg., Cygwin);
> >
> 
> Native Client is meant to be portable, so it should run on any x86 or x86-64
> machine once it's built.  In terms of building, I haven't built this gc port
> personally on Mac or Windows, but I just checked our build bot logs and they
> seem to be building ok on Mac and in Cygwin.

So, eg. DARWIN, GC_DARWIN_THREADS,  WIN32, CYGWIN, GC_WIN32_THREADS won't ever be defined when building NaCl, right?
Is GC_LINUX_THREADS defined when building NaCl with multi-threaded support?

I have very little knowledge of NaCl - could you briefly explain what does stand for NaCl portability - is it possible to call Win32 API if I'm compiling on Cygwin or should I use the NaCl API (and, thus, the compiled binary code will run on any x86 target)?

If NaCl is some kind of OS then LINUX, DARWIN, WIN32, etc shouldn't be defined (even if __linux__ defined) if NACL.
Same for GC_xxx_THREADS - I think GC_NACL_THREADS could be defined instead of GC_LINUX_THREADS, etc.

I also think that I386 and X86_64 should stay defined for respectively the corresponding CPU type (I guess it is already for NaCl but i haven't checked yet)
I think there should be 2 ifdef NACL define OS_TYPE "NACL" ... sections (one for every supported CPU).

> > - for non-static GC-internal symbols use GC_ prefix (eg. for
> > nacl_thread_parked);
> > - define SIG_SUSPEND to -1 (instead of 0) as it is returned by
> > GC_get_suspend_signal;
> > - GC functions called from NaCl it self (eg, nacl_pre_syscall_hook) shoud
> > be tagged with some attribute (like public GC functions are) both for code
> > readability and to prevent that symbols stripping when compiled as a shared
> > lib with -DGC_DLL);
> >
> 
> I'll address these issues.  (note that NaCl currently doesn't support shared
> libs yet so your dll example won't happen, but I agree that these should be
> treated like other public GC functions)

Ok. But what is eg. nacl_pre_syscall_hook() - a callback from the NaCl subsystem? (I guess this should be treated as GC public API)

Of course, use STATIC or static where possible (all STATIC symbols start with GC_, while static typically not).
More tips: use GC_INNER and GC_EXTERN for internal global variables; use GC_INNER for internal functions.

> > - libatomic_ops does not use signals API (except for CAS emulation which is
> > not used for x86/x64).
> 
> I think I saw sigprocmask and related functions and assumed the worst, but I
> see now that's windows code.  Looking at the x86 variants it looks like a
> NaCl port of libatomic_ops is probably not going to be too bad.  I'll look
> into this eventually.

Most probably, it work w/o any porting afforts but it would be good to port atomic_ops.c (similar to what I did for Win32-pthreads targets - see AO_USE_WIN32_PTHREADS, I guess you should add AO_USE_NACL macro testing in that file (looks easy to add). I think it's worth doing first (and submit me a separate patch for libatomics_op when done).

What's about GC_HAVE_BUILTIN_BACKTRACE and GC_CAN_SAVE_CALL_STACKS? At least, gc.h should be consistent with the GC implementation (I mean eg. if GC_HAVE_BUILTIN_BACKTRACE not supported then it shouldn't be defined in gc.h regardless of __linux__, _MSC_VER, etc. provided  __native_client__). Same for GC_ADD_CALLER, GC_RETURN_ADDR.

Regards.

> > PS. Let me not do the benefits analysis (probably someone else can do
> > this).
> >
> Well, if the gc7.2 port is as easy as it's looking now, I think it's
> probably worth doing it.  I would still love to hear anyone chime in on the
> benefits of gc7.2 vs 6.8 though  
> 
> > Regards.
> >
> > Thu, 13 Jan 2011 10:21:03 -0800 Elijah Taylor < elijahtaylor at google.com (sentmsg?compose&amp;To=elijahtaylor at google.com) >:
> >
> > > Hi GC folks,
> >
> > > I saw a little chatter in the archives related to porting libgc to Native
> > Client, so I joined this list to share some details. I'm the engineer at
> > Google who ported of libgc to Native Client for Mono. I've also included a
> > patch for vanilla gc6.8 in our naclports repository:
> >  http://code.google.com/p/naclports/ (http://code.google.com/p/naclports/) . This version will be available to
> > users that want to use libgc as part of their Native Client projects.
> >
> > > Before porting gc6.8 I had attempted to port one of the newer versions,
> > gc7.2alpha4, but ran into snags. The largest snag right now I think is that
> > gc 7+ includes libatomic_ops which will require some non-trivial effort in
> > order to work under Native Client. Most notably we don't support signals;
> > that was the biggest effort in porting libgc in the first place for NaCl,
> > and I assume that will require the most work in porting libatomic_ops too.
> >
> > > Can someone give me the high level details of what kind of things we
> > might be missing if we only support gc6.8 instead of the latest version?
> > Because of our thread stopping implementation, we may not even benefit from
> > some of the newer features. I just wanted to get a sense of what the
> > benefits are of getting a newer version available for users.
> >
> > > -Elijah


-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://napali.hpl.hp.com/pipermail/gc/attachments/20110115/1a695882/attachment-0001.htm


More information about the Gc mailing list