November 2009 - Darwin-dev - lists.apple.com

Re: Crash on _unload_image(); list of ObjC subclasses appears corrupted
by Greg Parker 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On Nov 19, 2009, at 4:50 AM, Matthew Mott wrote: > We are writing a plugin for a 64-bit Mac host application which is loaded and > unloaded dynamically. The plugin is written in C and C++ and statically linked > against Qt 4.5 on 64-bit Mac OS. We do not use any Objective-C ourselves, > however the 4.5 version of Qt may use them internally in order to support > Cocoa. Objective-C code that is intended to be unloaded must be written very carefully. It's likely that Qt is not written that way. Beware. You can set environment variable OBJC_DEBUG_UNLOAD=YES to get some warnings about poorly-behaved unloadable code. > Although our plugin loads and runs without problems, at application shutdown > the following crash is encountered in the Objective C runtime: > > #0 0x00007fff846d631a in removeSubclass (supercls=0x7fff7070c200, subcls=0x124dfeca0) at runtime/objc-runtime-new.m:870 > #1 0x00007fff846d8067 in _unload_image (hi=0x12223a710) at runtime/objc-runtime-new.m:1722 > #2 0x00007fff846cd418 in _objc_unmap_image (hi=0x12223a710) at runtime/objc-runtime.m:962 > #3 0x00007fff846cd305 in unmap_image (mh=0x12356e000, vmaddr_slide=4887863296) at runtime/objc-runtime.m:920 > #4 0x00007fff5fc02b67 in __dyld__ZN4dyld11removeImageEP11ImageLoader () > #5 0x00007fff5fc02d6f in __dyld__ZN4dyld20garbageCollectImagesEv () > #6 0x00007fff5fc0ac56 in __dyld_dlclose () > #7 0x00007fff80059755 in dlclose () > #8 0x00007fff8413ad09 in _CFBundleDlfcnUnload () > #9 0x00007fff8413cec9 in CFBundleUnloadExecutable () > ... > > The line of code which is crashing is 870 (runtime/objc-runtime-new.m:870) in > the following listing: > > 865 OBJC_CHECK_LOCKED(&runtimeLock); > 866 assert(getSuperclass(subcls) == supercls); > 867 > 868 class_t **cp; > 869 for (cp = &supercls->data->firstSubclass; > 870 *cp && *cp != subcls; > 871 cp = &(*cp)->data->nextSiblingClass) > 872 ; > 873 assert(*cp == subcls); > 874 *cp = subcls->data->nextSiblingClass; > > By tracing through execution we have identified that the problem occurs when > the variable cp points to an invalid location, which is caused by the > nextSiblingClass pointer in 871 being invalid. > > Although it is quite possible that this is a Qt-related problem (the specific > class which is being removed is a QCocoaWindowDelegate), it would be helpful > for us to know more about the construction of this linked list of sibling > classes. Who or what is responsible for populating this list at application > startup (or elsewhere)? Are there any known situations in which the linked > list might be, or become, invalid? firstSubclass and nextSiblingClass together form a tree of classes. Any class that is "used" is added to that tree when realizeClass() calls addSubclass(). I don't know of any reason for that tree to be invalid other than garden-variety memory smashers. -- Greg Parker gparker(a)apple.com Runtime Wrangler _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Jean-Daniel Dupas 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com Le 19 nov. 2009 à 18:31, Damien Sorresso a écrit : > On Nov 19, 2009, at 1:34 AM, Jean-Daniel Dupas wrote: >> Le 19 nov. 2009 à 00:18, Damien Sorresso a écrit : >>> On Nov 18, 2009, at 1:59 PM, Godfrey van der Linden wrote: >>>> Sorry for the repost, just occurred to me. >>>> >>>> Since you are launching a server process directly, have you considered the typical unix solution of using pipes to communicate with the child. See popen(3), which is a single API that that allows a parent to create a child and connect to it. >>> >>> >>> We're trying to discourage projects from spawning their own child processes. >> >> What if you need a child per work ? >> >> A real life example is decoding a movie in a 64 bits application. QuickTime decoders only works on the main thread and in 32 bit process (I know there is some kind of hacks to use them from other threads but they don't work most of the time). So, you have to spawn a child per movie (like QTKit does). >> >> Is there a clean way to accomplish this task using launchd ? > > > launchd doesn't provide support for multiple job instances currently. We are very aware of this need, however, so don't take this to mean that we see that use case as outside of our model. > > For now though, you can sort of achieve this by using the ServiceManagement API (in Snow Leopard) to submit unique-ified jobs to launchd on-the-fly Nice, that's the path I chose (except that I reimplemented it upon launchd API for 10.5 compatibility). Thank you for those answers. -- Jean-Daniel _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Damien Sorresso 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On Nov 19, 2009, at 1:34 AM, Jean-Daniel Dupas wrote: > Le 19 nov. 2009 à 00:18, Damien Sorresso a écrit : >> On Nov 18, 2009, at 1:59 PM, Godfrey van der Linden wrote: >>> Sorry for the repost, just occurred to me. >>> >>> Since you are launching a server process directly, have you considered the typical unix solution of using pipes to communicate with the child. See popen(3), which is a single API that that allows a parent to create a child and connect to it. >> >> >> We're trying to discourage projects from spawning their own child processes. > > What if you need a child per work ? > > A real life example is decoding a movie in a 64 bits application. QuickTime decoders only works on the main thread and in 32 bit process (I know there is some kind of hacks to use them from other threads but they don't work most of the time). So, you have to spawn a child per movie (like QTKit does). > > Is there a clean way to accomplish this task using launchd ? launchd doesn't provide support for multiple job instances currently. We are very aware of this need, however, so don't take this to mean that we see that use case as outside of our model. For now though, you can sort of achieve this by using the ServiceManagement API (in Snow Leopard) to submit unique-ified jobs to launchd on-the-fly. -- Damien Sorresso BSD Engineering Apple Inc. _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Finder tool rewritten in Cocoa
by Jens Alfke 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On Nov 18, 2009, at 11:25 PM, rohan a wrote: Now with 10.6, the Finder tool has been re-written using Cocoa. I am using the Alias manager of Carbon framework in my utility, With 10.6 would this work or I need to port to Cocoa ? —Jens _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… The Finder is just an application; it's not the same as the underlying system frameworks. The Finder could be rewritten in Ruby or FORTRAN and it wouldn't affect the APIs you use :) All of the CarbonCore APIs, including the Alias Manager, are still in 10.6. But there are some new methods on NSURL that provide simpler access to aliases (search for "bookmark" in the NSURL API docs.) This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Damien Sorresso 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On Nov 19, 2009, at 1:15 AM, Alastair Houghton wrote: > On 18 Nov 2009, at 23:18, Damien Sorresso wrote: >> We're trying to discourage projects from spawning their own child processes. > > For the case where the processes that are trying to communicate are parent and child already, so they won't be running unless both of them are running, I can't see an obvious advantage to using launchd (versus just doing a bootstrap_check_in()), and while it isn't significantly more complicated to use it (especially on 10.6), I'm not certain I see the benefit. Maybe I'm missing something? bootstrap_check_in() is designed for on-demand job to obtain their receive rights from launchd. By using launchd, the requestor can simply send a message to a port to get its helper process running, rather than having to go through the fork(2)/exec(3) dance, installing a kevent to monitor for the child's death, respawning if the child dies prematurely, etc. With launchd, you can just send a request and not have to worry about the process' lifecycle. -- Damien Sorresso BSD Engineering Apple Inc. _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Crash on _unload_image(); list of ObjC subclasses appears corrupted
by Matthew Mott 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com User-agent: KMail/1.12.2 (Linux/2.6.31-14-generic; KDE/4.3.2; i686; ; ) Hello, We are writing a plugin for a 64-bit Mac host application which is loaded and unloaded dynamically. The plugin is written in C and C++ and statically linked against Qt 4.5 on 64-bit Mac OS. We do not use any Objective-C ourselves, however the 4.5 version of Qt may use them internally in order to support Cocoa. Although our plugin loads and runs without problems, at application shutdown the following crash is encountered in the Objective C runtime: #0 0x00007fff846d631a in removeSubclass (supercls=0x7fff7070c200, subcls=0x124dfeca0) at runtime/objc-runtime-new.m:870 #1 0x00007fff846d8067 in _unload_image (hi=0x12223a710) at runtime/objc- runtime-new.m:1722 #2 0x00007fff846cd418 in _objc_unmap_image (hi=0x12223a710) at runtime/objc- runtime.m:962 #3 0x00007fff846cd305 in unmap_image (mh=0x12356e000, vmaddr_slide=4887863296) at runtime/objc-runtime.m:920 #4 0x00007fff5fc02b67 in __dyld__ZN4dyld11removeImageEP11ImageLoader () #5 0x00007fff5fc02d6f in __dyld__ZN4dyld20garbageCollectImagesEv () #6 0x00007fff5fc0ac56 in __dyld_dlclose () #7 0x00007fff80059755 in dlclose () #8 0x00007fff8413ad09 in _CFBundleDlfcnUnload () #9 0x00007fff8413cec9 in CFBundleUnloadExecutable () ... The line of code which is crashing is 870 (runtime/objc-runtime-new.m:870) in the following listing: 865 OBJC_CHECK_LOCKED(&runtimeLock); 866 assert(getSuperclass(subcls) == supercls); 867 868 class_t **cp; 869 for (cp = &supercls->data->firstSubclass; 870 *cp && *cp != subcls; 871 cp = &(*cp)->data->nextSiblingClass) 872 ; 873 assert(*cp == subcls); 874 *cp = subcls->data->nextSiblingClass; By tracing through execution we have identified that the problem occurs when the variable cp points to an invalid location, which is caused by the nextSiblingClass pointer in 871 being invalid. Although it is quite possible that this is a Qt-related problem (the specific class which is being removed is a QCocoaWindowDelegate), it would be helpful for us to know more about the construction of this linked list of sibling classes. Who or what is responsible for populating this list at application startup (or elsewhere)? Are there any known situations in which the linked list might be, or become, invalid? Many thanks for any advice, Matthew _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Jean-Daniel Dupas 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com Le 19 nov. 2009 à 00:18, Damien Sorresso a écrit : > On Nov 18, 2009, at 1:59 PM, Godfrey van der Linden wrote: >> Sorry for the repost, just occurred to me. >> >> Since you are launching a server process directly, have you considered the typical unix solution of using pipes to communicate with the child. See popen(3), which is a single API that that allows a parent to create a child and connect to it. > > > We're trying to discourage projects from spawning their own child processes. What if you need a child per work ? A real life example is decoding a movie in a 64 bits application. QuickTime decoders only works on the main thread and in 32 bit process (I know there is some kind of hacks to use them from other threads but they don't work most of the time). So, you have to spawn a child per movie (like QTKit does). Is there a clean way to accomplish this task using launchd ? > I haven't seen any reason the original poster couldn't just make the proposed child process into an on-demand launchd job that registers a MachService. -- Jean-Daniel _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Ed Wynne 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On 18 Nov 2009, at 23:18, Damien Sorresso wrote: Why, pray tell? -Ed _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… We're trying to discourage projects from spawning their own child processes. This email sent to site_archiver(a)lists.apple.com

1 0

Re: Alternative to bootstrap_register?
by Alastair Houghton 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com Dkim-signature: v=1; a=rsa-sha1; c=relaxed; d=alastairs-place.net; h= subject:mime-version:content-type:from:in-reply-to:date:cc: content-transfer-encoding:message-id:references:to; q=dns/txt; s= aug07; bh=IXE/1/+ID+PLpttNoLH+pfmdpbY=; b=L5NKCgkMEbw0aa8OGK2KNA UJYgyURaPj2ZLq36KYnmpxfB/XBd2lT2dWFiFNtUT4UKhx+UH/teX7ok7Mh3FOFb iUvFz7BzJn7N9OQwSHtpO/y4t0WAe/XHbGDfXiYtQhwAi/UICjzjyEyDKWnZGkKX 7VBGWPmvjVBPkNCpDJyFo= On 18 Nov 2009, at 23:18, Damien Sorresso wrote: > We're trying to discourage projects from spawning their own child processes. For the case where the processes that are trying to communicate are parent and child already, so they won't be running unless both of them are running, I can't see an obvious advantage to using launchd (versus just doing a bootstrap_check_in()), and while it isn't significantly more complicated to use it (especially on 10.6), I'm not certain I see the benefit. Maybe I'm missing something? Kind regards, Alastair. -- http://alastairs-place.net _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0

Re: Finder tool rewritten in Cocoa
by Eric Schlegel 19 Nov '09

19 Nov '09

site_archiver(a)lists.apple.com Delivered-To: darwin-dev(a)lists.apple.com On Nov 18, 2009, at 11:25 PM, rohan a wrote: > Hello All, > > Now with 10.6, the Finder tool has been re-written using Cocoa. I am > using the Alias manager of Carbon framework in my utility, With 10.6 > would this work or I need to port to Cocoa ? > I am still waiting for my copy of 10.6 hence this question The Alias Manager is still supported in 10.6. -eric _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev(a)lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.app… This email sent to site_archiver(a)lists.apple.com

1 0