On 7/1/2014 2:43 PM, Albert Chu wrote: > Introduce new sweep state PERFMGR_SWEEP_POST_PROCESSING to fix > race in perfmgr. > > Race occurs as follows: > > Under typical conditions, osm_perfmgr_process() is entered > with sweep_state set to PERFMGR_SWEEP_SLEEP. osm_perfmgr_process() > sets sweep_state to PERFMGR_SWEEP_ACTIVE when it begins to sweep. > > osm_perfmgr_process() will eventually call perfmgr_send_mad() by > way of perfmgr_query_counters() and several other functions. > > Responses to performance counter MADs may initiate the sending > of more MADs via perfmgr_send_mad(), such as through redirection > or the desire to clear counters. > > If too many MADs have been put on the wire, perfmgr_send_mad() > will throttle sending out MADS and temporarily change sweep_state > between PERFMGR_SWEEP_SUSPENDED and PERFMGR_SWEEP_ACTIVE as it > throttles. The sweep_state is set to PERFMGR_SWEEP_ACTIVE > when all performance counter MADs have been sent out by the sweeper. > > osm_perfmgr_process() eventually completes its sweep and puts > sweep_state back into PERFMGR_SWEEP_SLEEP. > > At this point, some MADs may still be on the wire. New MADs may be > put back on the wire if responses necessitate it (redirection or > clearing counters). If enough MADs are put back onto the wire, > perfmgr_send_mad() will throttle as normal, temporarily moving > between PERFMGR_SWEEP_SUSPENDED and PERFMGR_SWEEP_ACTIVE. After > the throttling is complete, sweep_state is put into > PERFMGR_SWEEP_ACTIVE state. > > This is the key problem, the sweep_state is changed from > PERFMGR_SWEEP_SLEEP to PERFMGR_SWEEP_ACTIVE outside of > osm_perfmgr_process(). > > Now that the perfmgr is in ACTIVE state, any future sweep call to > osm_perfmgr_process() will not sweep b/c the sweep_state is set > to PERFMGR_SWEEP_ACTIVE. > > The introduction of a new sweep_state PERFMGR_SWEEP_POST_PROCESSING > fixes this problem. > > If perfmgr_send_mad() throttles mads while in PERFMGR_SWEEP_SLEEP. > sweep_state will be moved into the PERFMGR_SWEEP_POST_PROCESSING > state instead of PERFMGR_SWEEP_SUSPENDED/PERFMGR_SWEEP_ACTIVE. > > When all post-SLEEP state MAD processing is complete, the sweep_state > will move from PERFMGR_SWEEP_POST_PROCESSING back to PERFMGR_SWEEP_SLEEP, > so that future sweeps can operate as normal. > > Signed-off-by: Albert L. Chu <chu11@xxxxxxxx> Thanks. Series applied (with minor cosmetic change). -- Hal -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
