On Thu, Sep 24, 2015 at 10:21 AM, Jiri Kosina <ji...@kernel.org> wrote: > On Thu, 24 Sep 2015, Sedat Dilek wrote: > >> >> > >> [ 24.705767] [<ffffffff8149287d>] dump_stack+0x7d/0xa0 >> >> > >> [ 24.705774] [<ffffffff810cbf7a>] ___might_sleep+0x28a/0x2a0 >> >> > >> [ 24.705779] [<ffffffff810cbc7f>] __might_sleep+0x4f/0xc0 >> >> > >> [ 24.705784] [<ffffffff810ae8ff>] start_flush_work+0x2f/0x290 >> >> > >> [ 24.705789] [<ffffffff810ae8ac>] flush_work+0x5c/0x80 >> >> > >> [ 24.705792] [<ffffffff810ae86a>] ? flush_work+0x1a/0x80 >> >> > >> [ 24.705799] [<ffffffff810eddcd>] ? trace_hardirqs_off+0xd/0x10 >> >> > >> [ 24.705804] [<ffffffff810ad938>] ? try_to_grab_pending+0x48/0x360 >> >> > >> [ 24.705810] [<ffffffff81917e13>] ? >> >> > >> _raw_spin_lock_irqsave+0x73/0x80 >> >> > >> [ 24.705814] [<ffffffff810aecf9>] __cancel_work_timer+0x179/0x260 >> >> >> >> This one is even more strange. It says that flush_work() is being called >> >> from __cancel_work_timer() with IRQs disabled, but flags are explicitly >> >> restored just one statement before that, and usbhid_close() explicitly >> >> calls cancel_work_sync() after unconditionally enabling interrupts. >> >> >> >> So I am not able to make any sense of either of the traces really. >> >> >> >> Are you seeing this with the same .config with GCC-compiled kernel as >> >> well? >> > >> > Actually could you please provide disassembly of your >> > __cancel_work_timer()? >> > >> >> Disassembly of which file - corresponding workqueue or hid file? > > make kernel/workqueue.o > objdump -Dr kernel/workqueue.o > > and copy/paste output for __cancel_work_timer function. > >> > One explanation would be LLVM not considering local_irq_restore() a >> > compiler memory barrier, but I am pretty sure it'll expose much more >> > breakage if that'd be the case. >> >> Can you point me where I can find more informations about "compiler >> memory barrier" or explain in a few words if possible? > > If compiler would not take "memory" clobber (while disabling IRQs) as a > reordering barrier, it wouldn't see any data dependency between > local_irq_restore(flags) and flush_work(data) and could reorder them, > resulting in flush_work() being called with IRQs disabled. >
BTW, is membarrier_test helpful in this case? I asked parallelly on LTP ML is there exists something similiar. - Sedat - [1] http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/tools/testing/selftests/membarrier [2] http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/plain/tools/testing/selftests/membarrier/membarrier_test.c -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
