lkml.org 
[lkml]   [2005]   [Oct]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH linux-2.6-block:master 01/05] blk: implement generic dispatch queue
     Hi, Jens.

    On Thu, Oct 20, 2005 at 12:00:03PM +0200, Jens Axboe wrote:
    > On Wed, Oct 19 2005, Tejun Heo wrote:
    > > @@ -40,6 +40,11 @@
    > > static DEFINE_SPINLOCK(elv_list_lock);
    > > static LIST_HEAD(elv_list);
    > >
    > > +static inline sector_t rq_last_sector(struct request *rq)
    > > +{
    > > + return rq->sector + rq->nr_sectors;
    > > +}
    >
    > Slightly misnamed, since it's really the sector after the last sector
    > :-)
    >
    > I've renamed that to rq_end_sector() instead.

    Maybe rename request_queue->last_sector too?

    >
    > > +/*
    > > + * Insert rq into dispatch queue of q. Queue lock must be held on
    > > + * entry. If sort != 0, rq is sort-inserted; otherwise, rq will be
    > > + * appended to the dispatch queue. To be used by specific elevators.
    > > + */
    > > +void elv_dispatch_insert(request_queue_t *q, struct request *rq, int sort)
    > > +{
    > > + sector_t boundary;
    > > + unsigned max_back;
    > > + struct list_head *entry;
    > > +
    > > + if (!sort) {
    > > + /* Specific elevator is performing sort. Step away. */
    > > + q->last_sector = rq_last_sector(rq);
    > > + q->boundary_rq = rq;
    > > + list_add_tail(&rq->queuelist, &q->queue_head);
    > > + return;
    > > + }
    > > +
    > > + boundary = q->last_sector;
    > > + max_back = q->max_back_kb * 2;
    > > + boundary = boundary > max_back ? boundary - max_back : 0;
    >
    > This looks really strange, what are you doing with boundary here?
    >

    Taking backward seeking into account. I reasonsed that if specific
    elevator chooses the next request with backward seeking,
    elv_dispatch_insert() shouldn't change the order because that may
    result in less efficient seek pattern. At the second thought,
    specific elevators always perform sorting by itself in such cases, so
    this seems unnecessary. I think we can strip this thing out.

    > > + list_for_each_prev(entry, &q->queue_head) {
    > > + struct request *pos = list_entry_rq(entry);
    > > +
    > > + if (pos->flags & (REQ_SOFTBARRIER|REQ_HARDBARRIER|REQ_STARTED))
    > > + break;
    > > + if (rq->sector >= boundary) {
    > > + if (pos->sector < boundary)
    > > + continue;
    > > + } else {
    > > + if (pos->sector >= boundary)
    > > + break;
    > > + }
    > > + if (rq->sector >= pos->sector)
    > > + break;
    > > + }
    > > +
    > > + list_add(&rq->queuelist, entry);
    > > +}
    >
    > I've split this into, I don't like rolled-up functions that really do
    > two seperate things. So elv_dispatch_sort() now does sorting,
    > elv_dispatch_add_tail() does what !sort would have done.

    Fine.

    >
    > > while ((rq = __elv_next_request(q)) != NULL) {
    > > - /*
    > > - * just mark as started even if we don't start it, a request
    > > - * that has been delayed should not be passed by new incoming
    > > - * requests
    > > - */
    > > - rq->flags |= REQ_STARTED;
    > > + if (!(rq->flags & REQ_STARTED)) {
    > > + elevator_t *e = q->elevator;
    > > +
    > > + /*
    > > + * This is the first time the device driver
    > > + * sees this request (possibly after
    > > + * requeueing). Notify IO scheduler.
    > > + */
    > > + if (blk_sorted_rq(rq) &&
    > > + e->ops->elevator_activate_req_fn)
    > > + e->ops->elevator_activate_req_fn(q, rq);
    > > +
    > > + /*
    > > + * just mark as started even if we don't start
    > > + * it, a request that has been delayed should
    > > + * not be passed by new incoming requests
    > > + */
    > > + rq->flags |= REQ_STARTED;
    > > + }
    > >
    > > if (rq == q->last_merge)
    > > q->last_merge = NULL;
    > >
    > > + if (!q->boundary_rq || q->boundary_rq == rq) {
    > > + q->last_sector = rq_last_sector(rq);
    > > + q->boundary_rq = NULL;
    > > + }
    >
    > This seems to be the only place where you clear ->boundary_rq, that
    > can't be right. What about rq-to-rq merging, ->boundary_rq could be
    > freed and you wont notice. Generally I don't really like keeping
    > pointers to rqs around, it's given us problems in the past with the
    > last_merge bits even. For now I've added a clear of this in
    > __blk_put_request() as well.

    Oh, please don't do that. Now, it's guaranteed that there are only
    three paths a request can travel.

    set_req_fn ->

    i. add_req_fn -> (merged_fn ->)* -> dispatch_fn -> activate_req_fn ->
    (deactivate_req_fn -> activate_req_fn ->)* -> completed_req_fn
    ii. add_req_fn -> (merged_fn ->)* -> merge_req_fn
    iii. [none]

    -> put_req_fn

    These three are the only paths a request can travel. Also note that
    dispatched requests don't get merged. So, after dispatched, the only
    way out is via elevator_complete_req_fn and that's why that's the only
    place ->boundary_rq is cleared. I've also documented above in biodoc
    so that we can simplify codes knowing above information.

    boundary_rq is used to keep request sorting sane when some pre-sorted
    requests are present in the dispatch queue. Without it request
    sorting acts wierdly when barrier requests are in the dispatch queue.

    >
    > > int elv_queue_empty(request_queue_t *q)
    > > {
    > > elevator_t *e = q->elevator;
    > >
    > > + if (!list_empty(&q->queue_head))
    > > + return 0;
    > > +
    > > if (e->ops->elevator_queue_empty_fn)
    > > return e->ops->elevator_queue_empty_fn(q);
    > >
    > > - return list_empty(&q->queue_head);
    > > + return 1;
    > > }
    >
    > Agree, this order definitely makes more sense.
    >
    > > @@ -2475,14 +2478,14 @@ static void __blk_put_request(request_qu
    > >
    > > void blk_put_request(struct request *req)
    > > {
    > > + unsigned long flags;
    > > + request_queue_t *q = req->q;
    > > +
    > > /*
    > > - * if req->rl isn't set, this request didnt originate from the
    > > - * block layer, so it's safe to just disregard it
    > > + * Gee, IDE calls in w/ NULL q. Fix IDE and remove the
    > > + * following if (q) test.
    > > */
    > > - if (req->rl) {
    > > - unsigned long flags;
    > > - request_queue_t *q = req->q;
    > > -
    > > + if (q) {
    >
    > The q == NULL is because ide is using requests allocated on the stack,
    > I've wanted for that to die for many years :)

    Somebody, please kill that thing.

    Thanks. :-)

    --
    tejun
    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

    \
     
     \ /
      Last update: 2005-10-20 15:48    [W:2.738 / U:0.756 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site