BFQ tags some bfq_queues as interactive or soft_rt if it deems that
these bfq_queues contain the I/O of, respectively, interactive or soft
real-time applications. BFQ privileges both these special types of
bfq_queues over normal bfq_queues. To privilege a bfq_queue, BFQ
mainly raises the weight of the bfq_queue. In particular, soft_rt
bfq_queues get a higher weight than interactive bfq_queues.

A bfq_queue may turn from interactive to soft_rt. And this leads to a
tricky issue. Soft real-time applications usually start with an
I/O-bound, interactive phase, in which they load themselves into main
memory. BFQ correctly detects this phase, and keeps the bfq_queues
associated with the application in interactive mode for a
while. Problems arise when the I/O pattern of the application finally
switches to soft real-time. One of the conditions for a bfq_queue to
be deemed as soft_rt is that the bfq_queue does not consume too much
bandwidth. But the bfq_queues associated with a soft real-time
application consume as much bandwidth as they can in the loading phase
of the application. So, after the application becomes truly soft
real-time, a lot of time should pass before the average bandwidth
consumed by its bfq_queues finally drops to a value acceptable for
soft_rt bfq_queues. As a consequence, there might be a time gap during
which the application is not privileged at all, because its bfq_queues
are not interactive any longer, but cannot be deemed as soft_rt yet.

To avoid this problem, BFQ pretends that an interactive bfq_queue
consumes zero bandwidth, and allows an interactive bfq_queue to switch
to soft_rt. Yet, this fake zero-bandwidth consumption easily causes
the bfq_queue to often switch to soft_rt deceptively, during its
loading phase. As in soft_rt mode, the bfq_queue gets its bandwidth
correctly computed, and therefore soon switches back to
interactive. Then it switches again to soft_rt, and so on. These
spurious fluctuations usually cause losses of throughput, because they
deceive BFQ's mechanisms for boosting throughput (injection,
I/O-plugging avoidance, ...).

This commit addresses this issue as follows:
1) It does compute actual bandwidth consumption also for interactive
   bfq_queues. This avoids the above false positives.
2) When a bfq_queue switches from interactive to normal mode, the
   consumed bandwidth is reset (forgotten). This allows the
   bfq_queue to enjoy soft_rt very quickly. In particular, two
   alternatives are possible in this switch:
    - the bfq_queue still has backlog, and therefore there is a budget
      already scheduled to serve the bfq_queue; in this case, the
      scheduling of the current budget of the bfq_queue is not
      hindered, because only the scheduling of the next budget will
      be affected by the weight drop. After that, if the bfq_queue is
      actually in a soft_rt phase, and becomes empty during the
      service of its current budget, which is the natural behavior of
      a soft_rt bfq_queue, then the bfq_queue will be considered as
      soft_rt when its next I/O arrives. If, in contrast, the
      bfq_queue remains constantly non-empty, then its next budget
      will be scheduled with a low weight, which is the natural
      treatment for an I/O-bound (non soft_rt) bfq_queue.
    - the bfq_queue is empty; in this case, the bfq_queue may be
      considered unjustly soft_rt when its new I/O arrives. Yet
      the problem is now much smaller than before, because it is
      unlikely that more than one spurious fluctuation occurs.
Tested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>