ucx
ucx copied to clipboard
openucx
Broken build on OS X
Error:
checking for ANSI C header files... (cached) yes
checking for pkg-config... /opt/local/bin/pkg-config
checking pkg-config is at least version 0.9.0... yes
checking for shm_open in -lrt... no
configure: error: librt not found
librt is not supported on OS X. Do we really need/use it ?
yes, it's needed for SystemV shared memory we can change this to conditionally compile out systemV support instead of failing configure.
@uccs/ucx-dev
Get the code running on OS X involves much more work that I anticipated. There are plenty of linux specific/non-posix codec. I made some progress here https://github.com/shamisp/ucx/tree/osx_fixes I wanna get more OS X users/experts involved. Right now I'm trying to figure out what to do with debug.[ch] and log.[ch] that relays on pthreads and ELF specific code. I think there is no particular reason to make it linux specific. The lock calls we have to wrap with ucs_* calls. I'm not sure what to do with debug.[ch].
On Oct 30, 2014, at 3:35 PM, Pasha <[email protected]mailto:[email protected]> wrote:
@uccs/ucx-devhttps://github.com/orgs/uccs/teams/ucx-dev Get the code running on OS X involves much more work that I anticipated. There are plenty of linux specific/non-posix codec. I made some progress here https://github.com/shamisp/ucx/tree/osx_fixes I wanna get more OS X users/experts involved. Right now I'm trying to figure out what to do with debug.[ch] and log.[ch] that relays on pthreads and ELF specific code. I think there is no particular reason to make it linux specific. The lock calls we have to wrap with ucs_* calls. I'm not sure what to do with debug.[ch].
So is the OS X pthread issue the lack of spin_[un]lock calls? At least on my Mac (Yosemite, Xcode 6.1). Could we use a mutex instead?
For the ELF, same issue in the openshmem code. Currently hard-coded for ELF, but we could dream up a small API to present the required info about the executable’s memory layout etc.
tony
On Oct 30, 2014, at 3:35 PM, Pasha <[email protected]mailto:[email protected]> wrote:
@uccs/ucx-devhttps://github.com/orgs/uccs/teams/ucx-dev Get the code running on OS X involves much more work that I anticipated. There are plenty of linux specific/non-posix codec. I made some progress here https://github.com/shamisp/ucx/tree/osx_fixes I wanna get more OS X users/experts involved. Right now I'm trying to figure out what to do with debug.[ch] and log.[ch] that relays on pthreads and ELF specific code. I think there is no particular reason to make it linux specific. The lock calls we have to wrap with ucs_* calls. I'm not sure what to do with debug.[ch].
Some code to get the executable name for OS X…
#include <libproc.h> #include <unistd.h>
const char * ucs_get_exe (void) { #ifdef APPLE int ret; static char pathbuf[PROC_PIDPATHINFO_MAXSIZE]; pid_t pid = getpid();
ret = proc_pidpath (pid, pathbuf, sizeof(pathbuf));
return (ret > 0) ? pathbuf : NULL; #endif /* APPLE */ }
On Oct 30, 2014, at 5:06 PM, Curtis, Tony <[email protected]mailto:[email protected]> wrote:
On Oct 30, 2014, at 3:35 PM, Pasha <[email protected]mailto:[email protected]> wrote:
@uccs/ucx-devhttps://github.com/orgs/uccs/teams/ucx-dev Get the code running on OS X involves much more work that I anticipated. There are plenty of linux specific/non-posix codec. I made some progress here https://github.com/shamisp/ucx/tree/osx_fixes I wanna get more OS X users/experts involved. Right now I'm trying to figure out what to do with debug.[ch] and log.[ch] that relays on pthreads and ELF specific code. I think there is no particular reason to make it linux specific. The lock calls we have to wrap with ucs_* calls. I'm not sure what to do with debug.[ch].
Some code to get the executable name for OS X… ...
Implicit was an offer to fold it into sys.c with suitable #ifdef guards and do a PReq.
tony
Pthreads - this part of code we should be able to change to mutex, since it is not critical path (i hope so). The debug part that lookups addresses in libraries is a one that concerns me.
On Oct 30, 2014, at 8:46 PM, Pasha <[email protected]mailto:[email protected]> wrote:
Pthreads - this part of code we should be able to change to mutex, since it is not critical path (i hope so). The debug part that lookups addresses in libraries is a one that concerns me.
the wikipedia article for Mach-O says
"Some versions of NetBSDhttp://en.wikipedia.org/wiki/NetBSD have had Mach-O support added as part of an implementation of binary compatibility”
I’ll go and browse that, it may give some hints about the things we need to do…
tony
Actually the problem is not in elf but all these dl_* (example: dl_match_address) functions from link.h OS X does not have this API at all.
dladdr seems to provide a somewhat similar functionality as indicated @ https://developer.apple.com/library/mac/Documentation/DeveloperTools/Reference/MachOReference/index.html#//apple_ref/doc/uid/TP40001398-CH1g-315644 .
George.
On Thu, Oct 30, 2014 at 11:00 PM, Pasha [email protected] wrote:
Actually the problem is not in elf but all these dl_* (example: dl_match_address) functions from link.h OS X does not have this API at all.
— Reply to this email directly or view it on GitHub https://github.com/uccs/ucx/issues/9#issuecomment-61209142.
Seems like pthread_spin_lock I would have to wrap with https://developer.apple.com/library/mac/documentation/Darwin/Reference/ManPages/man3/spinlock.3.html
@yosefe there are a lot of pthread_spin calls. Shell we wrap it with ucs_thead_lock/unlock/init to make it more convenient ?
On Oct 30, 2014, at 10:31 PM, Pasha <[email protected]mailto:[email protected]> wrote:
Seems like pthread_spin_lock I would have to wrap with https://developer.apple.com/library/mac/documentation/Darwin/Reference/ManPages/man3/spinlock.3.html
@yosefehttps://github.com/yosefe there are a lot of pthread_spin calls. Shell we wrap it with ucs_thead_lock/unlock/init to make it more convenient ?
Well, it’s possible that e.g. qthreads or massivethreads could be a drop-in replacement so an abstraction might be useful. Hide the threading implementation in a ucs/ subdirectory.
tony
@yosefe There is already some wrapper for debug purpose under: src/ucs/type/spinlock.h.
I think we just may extend the wrapper to support OS X.
@shamisp We could definitely wrap the spinlock as you suggest. There are places we can't settle for a mutex. The wrappers in spinlock.h are for "recursive spinlock" and we can extend that.
Regarding debug, we could disable the whole "detailed backtrace" thing for MAC, see HAVE_DETAILED_BACKTRACE. it will fall back to printing plain addresses.
spinlock - we will wrap it around with ucs_*
HAVE_DETAILED_BACKTRACE - in my build it is disabled. There are still debug functions (I guess for attaching debugger?) that relay on sone functionality in the file: dl_match_address, dl_lookup_address, ucs_debugger_attach,ucs_handle_error and plenty of others. Some of these functions are linux specific. George proposed a solution for one of the problems, but I'm sure there are more issues to fix.
@devreal currently it doesn't seem there is enough interest to push this. of course, if anyone is willing to invest in this, the contribution will be welcome :)