On Mon, Jun 22, 2009 at 03:44:08PM +0800, Gui Jianfeng wrote: > Preempt the ongoing non-rt ioq if there are rt ioqs waiting for dispatching > in ancestor or sibling groups. It will give other group's rt ioq an chance > to dispatch ASAP. > Hi Gui, Will new preempton logic of traversing up the hiearchy so that both new queue and old queue are at same level to take a preemption decision not take care of above scenario? Please have a look at bfq_find_matching_entity(). At the same time we probably don't want to preempt the non-rt queue with an RT queue in sibling group until and unless sibling group is an RT group. root / \ BEgrpA BEgrpB | | BEioq1 RTioq2 Above we got two BE group A and B and assume ioq in group A is being served and then an RT request in group B comes. Because group B is an BE class group, we should not preempt the queue in group A. Thanks Vivek > Signed-off-by: Gui Jianfeng <guijianfeng@xxxxxxxxxxxxxx> > --- > block/elevator-fq.c | 44 +++++++++++++++++++++++++++++++++++++++----- > block/elevator-fq.h | 1 + > 2 files changed, 40 insertions(+), 5 deletions(-) > > diff --git a/block/elevator-fq.c b/block/elevator-fq.c > index 2ad40eb..80526fd 100644 > --- a/block/elevator-fq.c > +++ b/block/elevator-fq.c > @@ -3245,8 +3245,16 @@ void elv_add_ioq_busy(struct elv_fq_data *efqd, struct io_queue *ioq) > elv_mark_ioq_busy(ioq); > efqd->busy_queues++; > if (elv_ioq_class_rt(ioq)) { > + struct io_entity *entity; > struct io_group *iog = ioq_to_io_group(ioq); > + > iog->busy_rt_queues++; > + entity = iog->entity.parent; > + > + for_each_entity(entity) { > + iog = io_entity_to_iog(entity); > + iog->sub_busy_rt_queues++; > + } > } > > #ifdef CONFIG_DEBUG_GROUP_IOSCHED > @@ -3290,9 +3298,18 @@ void elv_del_ioq_busy(struct elevator_queue *e, struct io_queue *ioq, > elv_clear_ioq_busy(ioq); > BUG_ON(efqd->busy_queues == 0); > efqd->busy_queues--; > + > if (elv_ioq_class_rt(ioq)) { > + struct io_entity *entity; > struct io_group *iog = ioq_to_io_group(ioq); > + > iog->busy_rt_queues--; > + entity = iog->entity.parent; > + > + for_each_entity(entity) { > + iog = io_entity_to_iog(entity); > + iog->sub_busy_rt_queues--; > + } > } > > elv_deactivate_ioq(efqd, ioq, requeue); > @@ -3735,12 +3752,32 @@ int elv_iosched_expire_ioq(struct request_queue *q, int slice_expired, > return ret; > } > > +static int check_rt_queue(struct io_queue *ioq) > +{ > + struct io_group *iog; > + struct io_entity *entity; > + > + iog = ioq_to_io_group(ioq); > + > + if (iog->busy_rt_queues) > + return 1; > + > + entity = iog->entity.parent; > + > + for_each_entity(entity) { > + iog = io_entity_to_iog(entity); > + if (iog->sub_busy_rt_queues) > + return 1; > + } > + > + return 0; > +} > + > /* Common layer function to select the next queue to dispatch from */ > void *elv_fq_select_ioq(struct request_queue *q, int force) > { > struct elv_fq_data *efqd = &q->elevator->efqd; > struct io_queue *new_ioq = NULL, *ioq = elv_active_ioq(q->elevator); > - struct io_group *iog; > int slice_expired = 1; > > if (!elv_nr_busy_ioq(q->elevator)) > @@ -3811,12 +3848,9 @@ void *elv_fq_select_ioq(struct request_queue *q, int force) > /* > * If we have a RT cfqq waiting, then we pre-empt the current non-rt > * cfqq. > - * > - * TODO: This does not seem right across the io groups. Fix it. > */ > - iog = ioq_to_io_group(ioq); > > - if (!elv_ioq_class_rt(ioq) && iog->busy_rt_queues) { > + if (!elv_ioq_class_rt(ioq) && check_rt_queue(ioq)) { > /* > * We simulate this as cfqq timed out so that it gets to bank > * the remaining of its time slice. > diff --git a/block/elevator-fq.h b/block/elevator-fq.h > index b3193f8..be6c1af 100644 > --- a/block/elevator-fq.h > +++ b/block/elevator-fq.h > @@ -248,6 +248,7 @@ struct io_group { > * non-RT cfqq in service when this value is non-zero. > */ > unsigned int busy_rt_queues; > + unsigned int sub_busy_rt_queues; > > int deleting; > unsigned short iocg_id; > -- > 1.5.4.rc3 > -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel