On Tue, Oct 25, 2011 at 1:17 PM, Roland Dreier <[email protected]> wrote: > On Mon, Oct 24, 2011 at 11:07 PM, Bart Van Assche <[email protected]> wrote: > >> As far as I know every HCA supported by Linux does implement this > >> correctly. Which class did you have in mind as not doing that? > > > > At least QDR ConnectX 2 HCAs with fairly recent firmware. This > > behavior can be reproduced easily with the SCST version of ib_srpt as > > follows: > > 1. Log in from an SRP initiator to ib_srpt. > > 2. Start a direct I/O read test, e.g. with fio. > > 3. Issue the command "rmmod ib_srpt" on the target during I/O. > > OK, this is a pretty serious bug in mlx4_ib if true. Are > you sure that you really are seeing some pending work > requests not generating a flush error when the QP transitions > to the error state?
It's a little more complex than that. The original version of ib_srpt stops polling for completions as soon as the last WQE event has been received and after that the queue has been drained. So I don't know whether these flush errors were not delivered or whether these were delivered too late. Bart. -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
