Mathias Nyman <[email protected]> writes:
>> We set CMD_RING_STATE_ABORTED state under locking. I'm not checking what
>> is for taking lock for register though, I guess it should be enough just
>> lock around of read=>write of ->cmd_ring if need lock.
>
> After your patch it should be enough to have the lock only while
> reading and writing the cmd_ring register.
>
> If we want a locking fix that applies more easily to older stable
> releases before your change then the lock needs to cover set
> CMD_RING_STATE_ABORT, read cmd_reg, write cmd_reg and busiloop
> checking CRR bit. Otherwise the stop cmd ring interrupt handler may
> restart the ring just before we start checing CRR. The stop cmd ring
> interrupt will set the CMD_RING_STATE_ABORTED to
> CMD_RING_STATE_RUNNING so ring will really restart in the interrupt
> handler.
Just for record (no chance to make patch I myself for now, sorry), while
checking locking slightly, I noticed unrelated missing locking.
xhci_cleanup_command_queue()
We are calling it without locking, but we need to lock for accessing list.
Thanks.
--
OGAWA Hirofumi <[email protected]>
