Hello! On 29.06.19 10:50, Stanislaw Gruszka wrote: > Hello > > On Wed, Jun 26, 2019 at 03:28:00PM +0200, Soeren Moch wrote: >> Hi Stanislaw, >> >> the good news is: your patch below also solves the issue for me. But >> removing the ENTRY_DATA_STATUS_PENDING check in >> rt2x00usb_kick_rx_entry() alone does not help, while removing this check >> in rt2x00usb_work_rxdone() alone does the trick. >> >> So the real race seems to be that the flags set in the completion >> handler are not yet visible on the cpu core running the workqueue. And >> because the worker is not rescheduled when aborted, the entry can just >> wait forever. >> Do you think this could make sense? > Yes. > >>> I'm somewhat reluctant to change the order, because TX processing >>> might relay on it (we first mark we wait for TX status and >>> then mark entry is no longer owned by hardware). >> OK, maybe it's just good luck that changing the order solves the rx >> problem. Or can memory barriers associated with the spinlock in >> rt2x00lib_dmadone() be responsible for that? >> (I'm testing on a armv7 system, Cortex-A9 quadcore.) > I'm not sure, rt2x00queue_index_inc() also disable/enable interrupts, > so maybe that make race not reproducible. I tested some more, the race is between setting ENTRY_DATA_IO_FAILED (if needed) and enabling workqueue processing. This enabling was done via ENTRY_DATA_STATUS_PENDING in my patch. So setting ENTRY_DATA_STATUS_PENDING behind the spinlock in rt2x00lib_dmadone()/rt2x00queue_index_inc() moved this very close to setting of ENTRY_DATA_IO_FAILED (if needed). While still in the wrong order, this made it very unlikely for the race to show up. > >> While looking at it, why we double-clear ENTRY_OWNER_DEVICE_DATA in >> rt2x00usb_interrupt_rxdone() directly and in rt2x00lib_dmadone() again, > rt2x00lib_dmadone() is called also on other palaces (error paths) > when we have to clear flags. Yes, but also clearing ENTRY_OWNER_DEVICE_DATA in rt2x00usb_interrupt_rxdone() directly is not necessary and can lead to the wrong processing order. >> while not doing the same for tx? > If I remember correctly we have some races on rx (not happened on tx) > that was solved by using test_and_clear_bit(ENTRY_OWNER_DEVICE_DATA). I searched in the history, it actually was the other way around. You changed test_and_clear_bit() to test_bit() in the TX path. I think this is also the right way to go in RX. >> Would it make more sense to possibly >> set ENTRY_DATA_IO_FAILED before clearing ENTRY_OWNER_DEVICE_DATA in >> rt2x00usb_interrupt_rxdone() as for tx? > I don't think so, ENTRY_DATA_IO_FAILED should be only set on error > case. Yes of course. But if the error occurs, it should be signalled before enabling the workqueue processing, see the race described above. After some more testing I'm convinced that this would be the right fix for this problem. I will send a v2 of this patch accordingly. > >>> However on RX >>> side ENTRY_DATA_STATUS_PENDING bit make no sense as we do not >>> wait for status. We should remove ENTRY_DATA_STATUS_PENDING on >>> RX side and perhaps this also will solve issue you observe. >> I agree that removing the unnecessary checks is a good idea in any case. >>> Could you please check below patch, if it fixes the problem as well? >> At least I could not trigger the problem within transferring 10GB of >> data. But maybe the probability for triggering this bug is just lower >> because ENTRY_OWNER_DEVICE_DATA is cleared some time before >> ENTRY_DATA_STATUS_PENDING is set? > Not sure. Anyway, could you post patch removing not needed checks > with proper description/changelog ? > OK, I will do so. Soeren
