From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753964Ab3EHKXz (ORCPT ); Wed, 8 May 2013 06:23:55 -0400 Received: from 173-166-109-252-newengland.hfc.comcastbusiness.net ([173.166.109.252]:46589 "EHLO bombadil.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753190Ab3EHKXx (ORCPT ); Wed, 8 May 2013 06:23:53 -0400 Date: Wed, 8 May 2013 12:22:20 +0200 From: Peter Zijlstra To: Sasha Levin Cc: torvalds@linux-foundation.org, mingo@kernel.org, linux-kernel@vger.kernel.org Subject: Re: [PATCH 7/9] liblockdep: Support using LD_PRELOAD Message-ID: <20130508102220.GB6131@dyad.programming.kicks-ass.net> References: <1367348080-4680-1-git-send-email-sasha.levin@oracle.com> <1367348080-4680-8-git-send-email-sasha.levin@oracle.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1367348080-4680-8-git-send-email-sasha.levin@oracle.com> User-Agent: Mutt/1.5.21 (2012-12-30) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, Apr 30, 2013 at 02:54:38PM -0400, Sasha Levin wrote: > + > +static struct rb_node **__get_lock_node(void *lock, struct rb_node **parent) > +{ > +} > + > +/** > + * __get_lock - find or create a lock instance > + * @lock: pointer to a pthread lock function > + * > + * Try to find an existing lock in the rbtree using the provided pointer. If > + * one wasn't found - create it. > + */ > +static struct lock_lookup *__get_lock(void *lock) > +{ > +} This needs something like: static void __del_lock(void *lock); Since now you're repeating yourself in pthread_{rwlock,mutex}_destroy() :-) > +int pthread_mutex_lock(pthread_mutex_t *mutex) > +{ > + int r; > + > + try_init_preload(); You seem consistently whitespace challenged on this line. > + > + lock_acquire(&__get_lock(mutex)->dep_map, 0, 0, 0, 2, NULL, > + (unsigned long)_THIS_IP_); > + /* > + * Here's the thing with pthread mutexes: unlike the kernel variant, > + * they can fail. > + * > + * This means that the behaviour here is a bit different from what's > + * going on in the kernel: there we just tell lockdep that we took the > + * lock before actually taking it, but here we must deal with the case > + * that locking failed. mutex_lock_{interruptible,killable}() could be argued to be able to fail too. And if you look at __mutex_lock_common() you'll see that it does the exact same thing -- that is release on fail. > + * To do that we'll "release" the lock if locking failed - this way > + * we'll get lockdep doing the correct checks when we try to take > + * the lock, and if that fails - we'll be back to the correct > + * state by releasing it. > + */ > + r = ll_pthread_mutex_lock(mutex); > + if (r) > + lock_release(&__get_lock(mutex)->dep_map, 0, (unsigned long)_THIS_IP_); > + > + return r; > +} > + > +int pthread_mutex_trylock(pthread_mutex_t *mutex) > +{ > + int r; > + > + try_init_preload(); See.. > + lock_acquire(&__get_lock(mutex)->dep_map, 0, 1, 0, 2, NULL, (unsigned long)_THIS_IP_); > + r = ll_pthread_mutex_trylock(mutex); > + if (r) > + lock_release(&__get_lock(mutex)->dep_map, 0, (unsigned long)_THIS_IP_); > + > + return r; > +} > +__attribute__((constructor)) static void init_preload(void) > +{ > + static bool preload_started; > + > + if (preload_done) > + return; > + > + /* > + * Some programs attempt to initialize and use locks in their > + * allocation path. This means that a call to malloc() would > + * result in locks being initialized and locked. > + * > + * Why is it an issue for us? dlsym() below will try allocating to > + * give us the original function. Since this allocation will result > + * in a locking operations, we have to let pthread deal with it, > + * but we can't! we don't have the pointer to the original API > + * since we're inside dlsym() trying to get it :( > + * > + * We can work around it by telling the program that locking was > + * really okay, and just initialize those locks when we're fully > + * up and running (this is ok because this all happens during > + * initialization phase, when we have just one thread). But > + * this is a big TODO at this point. > + */ Fun.. got any example programs that trigger this? > + if (preload_started) { > + printf( > + "LOCKDEP error: It seems that the program you are trying to " > + "debug is initializing locks in it's allocation path.\n" > + "This means that liblockdep cannot reliably analyze this " > + "program since we need the allocator to work before we can " > + "debug locks.\nSorry!\n"); > + > + exit(1); > + } > + > + preload_started = true; > + > + ll_pthread_mutex_init = dlsym(RTLD_NEXT, "pthread_mutex_init"); > + ll_pthread_mutex_lock = dlsym(RTLD_NEXT, "pthread_mutex_lock"); > + ll_pthread_mutex_trylock = dlsym(RTLD_NEXT, "pthread_mutex_trylock"); > + ll_pthread_mutex_unlock = dlsym(RTLD_NEXT, "pthread_mutex_unlock"); > + ll_pthread_mutex_destroy = dlsym(RTLD_NEXT, "pthread_mutex_destroy"); > + > + ll_pthread_rwlock_init = dlsym(RTLD_NEXT, "pthread_rwlock_init"); > + ll_pthread_rwlock_destroy = dlsym(RTLD_NEXT, "pthread_rwlock_destroy"); > + ll_pthread_rwlock_rdlock = dlsym(RTLD_NEXT, "pthread_rwlock_rdlock"); > + ll_pthread_rwlock_tryrdlock = dlsym(RTLD_NEXT, "pthread_rwlock_tryrdlock"); > + ll_pthread_rwlock_wrlock = dlsym(RTLD_NEXT, "pthread_rwlock_wrlock"); > + ll_pthread_rwlock_trywrlock = dlsym(RTLD_NEXT, "pthread_rwlock_trywrlock"); > + ll_pthread_rwlock_unlock = dlsym(RTLD_NEXT, "pthread_rwlock_unlock"); > + > + lockdep_init(); > + > + preload_done = true; > +}