From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:53785) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1g9YaC-0000XC-Bw for qemu-devel@nongnu.org; Mon, 08 Oct 2018 12:40:37 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1g9YaB-0004ms-CX for qemu-devel@nongnu.org; Mon, 08 Oct 2018 12:40:36 -0400 Date: Mon, 8 Oct 2018 18:40:26 +0200 From: Kevin Wolf Message-ID: <20181008164026.GD4604@localhost.localdomain> References: <20181005143802.18412-1-peter.maydell@linaro.org> <20181005161739.GC4606@localhost.localdomain> <20181005180901.GA32677@localhost.localdomain> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: Subject: Re: [Qemu-devel] [PATCH] tests: Disable test-bdrv-drain List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Peter Maydell Cc: Paolo Bonzini , QEMU Developers , Max Reitz , Qemu-block , "patches@linaro.org" Am 08.10.2018 um 17:43 hat Peter Maydell geschrieben: > Looking at the backtraces I'm wondering if this is the result of > an implicit reliance on the order in which per-thread destructors > are called (which is left unspecified by POSIX) -- the destructor > function qemu_thread_atexit_run() is called after some other > destructor, but accesses its memory. > > Specifically, the memory it's trying to read looks like > the __thread local variable pollfds_cleanup_notifier in > util/aio-posix.c. So I think what is happening is: > * util/aio-posix.c calls qemu_thread_atexit_add(), passing > it a pointer to a thread-local variable pollfds_cleanup_notifier > * qemu_thread_atexit_add() works by arranging to run the > notifiers when its 'exit_key' variable's destructor is called > * the destructor for pollfds_cleanup_notifier runs before that > for exit_key, and so the qemu_thread_atexit_run() function > ends up touching freed memory > > I'm pretty confident this analysis of the problem is correct: > unfortunately I have no idea what the right way to fix it is... Yes, I agree with your analysis. If __thread variables can be destructed before pthread_key_create() destructors are called (and in particular if the former are implemented in terms of the latter), this implies at least two rules: 1. The Notfier itself can't be a TLS variable 2. The notifier callback can't access any TLS variables Of course, with these restrictions, qemu_thread_atexit_*() with its existing API is as useless as it could be. The best I can think of at the moment would be to use a separate pthread_key_create() (and therefore a separate destructor) for registering each TLS variable, so that the destructor always gets a valid pointer. Maybe move all __thread variables of a file into a single malloced struct to make it more managable (we could then keep a __thread pointer to it for convenience, but only free the struct with the pointer passed by the pthread_key destructor so that we don't have to access __thread variables in the destructor). By the way, can you reproduce this with virtio-blk/scsi and an iothread in a real QEMU or is it only the test case that fails? In theory, I don't see what would prevent QEMU from hanging at shutdown. Kevin