Hi, Jens,

If you recall, I posted an RFC patch for this back in July of last year:
http://lkml.org/lkml/2010/7/13/279

The basic problem is that a process can issue a never-ending stream of
async direct I/Os to the same sector on a device, thus starving out
other I/O in the system (due to the way the alias handling works in both
cfq and deadline).  The solution I proposed back then was to start
dispatching from the fifo after a certain number of aliases had been
dispatched.  Vivek asked why we had to treat aliases differently at all,
and I never had a good answer.  So, I put together a simple patch which
allows aliases to be added to the rb tree (it adds them to the right,
though that doesn't matter as the order isn't guaranteed anyway).  I
think this is the preferred solution, as it doesn't break up time slices
in CFQ or batches in deadline.  I've tested it, and it does solve the
starvation issue.  Let me know what you think.

Cheers,
Jeff
Signed-off-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

When struct cfq_data allocation fails, cic_index need to be freed.
Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

The 'group_changed' variable is initialized to 0 and never changed, so
checking the variable is meaningless.

It is a leftover from 0bbfeb83 ("cfq-iosched: Always provide group
iosolation."). Let's get rid of it.
Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Cc: Justin TerAvest <teravest@google.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Reduce the number of bit operations in cfq_choose_req() on average
(and worst) cases.
Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Simplify the calculation in cfq_prio_to_maxrq(), plus replace CFQ_PRIO_LISTS to
IOPRIO_BE_NR since they are the same and IOPRIO_BE_NR looks more reasonable in
this context IMHO.
Signed-off-by: NNamhyung Kim <namhyung@gmail.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

We allocated per cpu stats struct for root group but did not free it.
Fix it.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Currently we take blkg_stat lock for even updating the stats. So even if
a group has no throttling rules (common case for root group), we end
up taking blkg_lock, for updating the stats.

Make dispatch stats per cpu so that these can be updated without taking
blkg lock.

If cpu goes offline, these stats simply disappear. No protection has
been provided for that yet. Do we really need anything for that?
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Currently, all the cfq_group or throtl_group allocations happen while
we are holding ->queue_lock and sleeping is not allowed.

Soon, we will move to per cpu stats and also need to allocate the
per group stats. As one can not call alloc_percpu() from atomic
context as it can sleep, we need to drop ->queue_lock, allocate the
group, retake the lock and continue processing.

In throttling code, I check the queue DEAD flag again to make sure
that driver did not call blk_cleanup_queue() in the mean time.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

blkg->key = cfqd is an rcu protected pointer and hence we used to do
call_rcu(cfqd->rcu_head) to free up cfqd after one rcu grace period.

The problem here is that even though cfqd is around, there are no
gurantees that associated request queue (td->queue) or q->queue_lock
is still around. A driver might have called blk_cleanup_queue() and
release the lock.

It might happen that after freeing up the lock we call
blkg->key->queue->queue_ock and crash. This is possible in following
path.

blkiocg_destroy()
 blkio_unlink_group_fn()
  cfq_unlink_blkio_group()

Hence, wait for an rcu peirod if there are groups which have not
been unlinked from blkcg->blkg_list. That way, if there are any groups
which are taking cfq_unlink_blkio_group() path, can safely take queue
lock.

This is how we have taken care of race in throttling logic also.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Nobody seems to be using cfq_find_alloc_cfqg() function parameter "create".
Get rid of that.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Currentlly we first map the task to cgroup and then cgroup to
blkio_cgroup. There is a more direct way to get to blkio_cgroup
from task using task_subsys_state(). Use that.

The real reason for the fix is that it also avoids a race in generic
cgroup code. During remount/umount rebind_subsystems() is called and
it can do following with and rcu protection.

cgrp->subsys[i] = NULL;

That means if somebody got hold of cgroup under rcu and then it tried
to do cgroup->subsys[] to get to blkio_cgroup, it would get NULL which
is wrong. I was running into this race condition with ltp running on a
upstream derived kernel and that lead to crash.

So ideally we should also fix cgroup generic code to wait for rcu
grace period before setting pointer to NULL. Li Zefan is not very keen
on introducing synchronize_wait() as he thinks it will slow
down moun/remount/umount operations.

So for the time being atleast fix the kernel crash by taking a more
direct route to blkio_cgroup.

One tester had reported a crash while running LTP on a derived kernel
and with this fix crash is no more seen while the test has been
running for over 6 days.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Reviewed-by: NLi Zefan <lizf@cn.fujitsu.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

For some configurations of CONFIG_PREEMPT that is not true. So
get rid of __call_for_each_cic() and always uses the explicitly
rcu_read_lock() protected call_for_each_cic() instead.

This fixes a potential bug related to IO scheduler removal or
online switching.

Thanks to Paul McKenney for clarifying this.
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Instead of overloading __blk_run_queue to force an offload to kblockd
add a new blk_run_queue_async helper to do it explicitly.  I've kept
the blk_queue_stopped check for now, but I suspect it's not needed
as the check we do when the workqueue items runs should be enough.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Fixes generated by 'codespell' and manually reviewed.
Signed-off-by: NLucas De Marchi <lucas.demarchi@profusion.mobi>

Removing think time checking. A high thinktime queue might means the queue
dispatches several requests and then do away. Limitting such queue seems
meaningless. And also this can simplify code. This is suggested by Vivek.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

For v2, I added back lines to cfq_preempt_queue() that were removed
during updates for accounting unaccounted_time. Thanks for pointing out
that I'd missed these, Vivek.

Previous commit "cfq-iosched: Don't set active queue in preempt" wrongly
cleared stats for preempting queues when it shouldn't have, because when
we choose a queue to preempt, it still isn't necessarily scheduled next.

Thanks to Vivek Goyal for figuring this out and understanding how the
preemption code works.
Signed-off-by: NJustin TerAvest <teravest@google.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Commit "Add unaccounted time to timeslice_used" changed the behavior of
cfq_preempt_queue to set cfqq active. Vivek pointed out that other
preemption rules might get involved, so we shouldn't manually set which
queue is active.

This cleans up the code to just clear the queue stats at preemption
time.
Signed-off-by: NJustin TerAvest <teravest@google.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Version 3 is updated to apply to for-2.6.39/core.

For version 2, I took Vivek's advice and made sure we update the group
weight from cfq_group_service_tree_add().

If a weight was updated while a group is on the service tree, the
calculation for the total weight of the service tree can be adjusted
improperly, which either leads to bad service tree weights, or
potentially crashes (if total_weight becomes 0).

This patch defers updates to the weight until a group is off the service
tree.
Signed-off-by: NJustin TerAvest <teravest@google.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

There are two kind of times that tasks are not charged for: the first
seek and the extra time slice used over the allocated timeslice. Both
of these exported as a new unaccounted_time stat.

I think it would be good to have this reported in 'time' as well, but
that is probably a separate discussion.
Signed-off-by: NJustin TerAvest <teravest@google.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Code has been converted over to the new explicit on-stack plugging,
and delay users have been converted to use the new API for that.
So lets kill off the old plugging along with aops->sync_page().
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

The update_vdisktime logic is broken since commit
b54ce60e, st->min_vdisktime never makes
a progress. Fix it.

Thanks Vivek for pointing it out.
Signed-off-by: NGui Jianfeng <guijianfen@cn.fujitsu.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

If there are a sync and an async queue and the sync queue's think time
is small, we can ignore the sync queue's dispatch quantum. Because the
sync queue will always preempt the async queue, we don't need to care
about async's latency.  This can fix a performance regression of
aiostress test, which is introduced by commit f8ae6e3e. The issue
should exist even without the commit, but the commit amplifies the
impact.

The initial post does the same optimization for RT queue too, but since
I have no real workload for it, Vivek suggests to drop it.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Reviewed-by: NGui Jianfeng <guijianfeng@cn.fujitsu.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

We need to hold the queue lock over the reference increment,
it's not atomic anymore.
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

__blk_run_queue() automatically either calls q->request_fn() directly
or schedules kblockd depending on whether the function is recursed.
blk-flush implementation needs to be able to explicitly choose
kblockd.  Add @force_kblockd.

All the current users are converted to specify %false for the
parameter and this patch doesn't introduce any behavior change.

stable: This is prerequisite for fixing ide oops caused by the new
        blk-flush implementation.
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Jan Beulich <JBeulich@novell.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: stable@kernel.org
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Effectively, make group_isolation=1 the default and remove the tunable.
The setting group_isolation=0 was because by default we idle on
sync-noidle tree and on fast devices, this can be very harmful for
throughput.

However, this problem can also be addressed by tuning slice_idle and
possibly group_idle on faster storage devices.

This change simplifies the CFQ code by removing the feature entirely.
Signed-off-by: NJustin TerAvest <teravest@google.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Flush requests are never put on the IO scheduler.  Convert request
structure's elevator_private* into an array and have the flush fields
share a union with it.

Reclaim the space lost in 'struct request' by moving 'completion_data'
back in the union with 'rb_node'.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Acked-by: NTejun Heo <tj@kernel.org>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Commit 7667aa06 added logic to wait for
the last queue of the group to become busy (have at least one request),
so that the group does not lose out for not being continuously
backlogged. The commit did not check for the condition that the last
queue already has some requests. As a result, if the queue already has
requests, wait_busy is set. Later on, cfq_select_queue() checks the
flag, and decides that since the queue has a request now and wait_busy
is set, the queue is expired.  This results in early expiration of the
queue.

This patch fixes the problem by adding a check to see if queue already
has requests. If it does, wait_busy is not set. As a result, time slices
do not expire early.

The queues with more than one request are usually buffered writers.
Testing shows improvement in isolation between buffered writers.

Cc: stable@kernel.org
Signed-off-by: NJustin TerAvest <teravest@google.com>
Reviewed-by: NGui Jianfeng <guijianfeng@cn.fujitsu.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

o Rename a function to give it more approprate name. We are calculating
  cfq queue slice and function name gives the impression as if cfq group
  slice length is being calculated.
Signed-off-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

If a queue is preempted before it gets slice assigned, the queue doesn't get
compensation, which looks unfair. For such queue, we compensate it for a whole
slice.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

I got this:
             fio-874   [007]  2157.724514:   8,32   m   N cfq874 preempt
             fio-874   [007]  2157.724519:   8,32   m   N cfq830 slice expired t=1
             fio-874   [007]  2157.724520:   8,32   m   N cfq830 sl_used=1 disp=0 charge=1 iops=0 sect=0
             fio-874   [007]  2157.724521:   8,32   m   N cfq830 set_active wl_prio:0 wl_type:0
             fio-874   [007]  2157.724522:   8,32   m   N cfq830 Not idling. st->count:1

cfq830 is an async queue, and preempted by a sync queue cfq874. But since we
have cfqg->saved_workload_slice mechanism, the preempt is a nop.
Looks currently our preempt is totally broken if the two queues are not from
the same workload type.
Below patch fixes it. This will might make async queue starvation, but it's
what our old code does before cgroup is added.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

cfq_group->ref is used with queue_lock hold, the only exception is
cfq_set_request, which looks like a bug to me, so ref doesn't need
to be an atomic and atomic operation is slower.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

cfq_queue->ref is used with queue_lock hold, so ref doesn't need to be an atomic
and atomic operation is slower.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

When cfq_choose_cfqg() is called in select_queue(), there must be at least one
backlogged CFQ queue waiting for dispatching, hence there must be at least one
backlogged CFQ group on service tree. So we never call choose_service_tree()
with cfqg == NULL.
Signed-off-by: NGui Jianfeng <guijianfeng@cn.fujitsu.com>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

If priority is changed, continuing to check workload_expires and service tree
count of the previous workload does not make sense. We should always choose
the workload with lowest key of new priority in such case.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Reviewed-by: NJeff Moyer <jmoyer@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

It's able to check whether a CFQ group on a service tree by
checking "cfqg->rb_node". There's no need to maintain an
extra flag here.
Signed-off-by: NGui Jianfeng <guijianfeng@cn.fujitsu.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

When a cfq group is running, it won't be dequeued from service tree, so
there's no need to store the active one in st->active. Just gid rid of it.
Signed-off-by: NGui Jianfeng <guijianfeng@cn.fujitsu.com>
Acked-by: NVivek Goyal <vgoyal@redhat.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Vivek suggests we don't need schedule a dispatch when an idle queue
becomes nonidle. And he is right, cfq_should_preempt already covers
the logic.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

If a deep seek queue slowly deliver requests but disk is much faster, idle
for the queue just wastes disk throughput. If the queue delevers all requests
before half its slice is used, the patch disable idle for it.
In my test, application delivers 32 requests one time, the disk can accept
128 requests at maxium and disk is fast. without the patch, the throughput
is just around 30m/s, while with it, the speed is about 80m/s. The disk is
a SSD, but is detected as a rotational disk. I can configure it as SSD, but
I thought the deep seek queue logic should be fixed too, for example,
considering a fast raid.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

A queue is idle at cfq_dispatch_requests(), but it gets noidle later. Unless
other task explictly does unplug or all requests are drained, we will not
deliever requests to the disk even cfq_arm_slice_timer doesn't make the
queue idle. For example, cfq_should_idle() returns true because of
service_tree->count == 1, and then other queues are added. Note, I didn't
see obvious performance impacts so far with the patch, but just thought
this could be a problem.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>

Some functions should return boolean.
Signed-off-by: NShaohua Li <shaohua.li@intel.com>
Signed-off-by: NJens Axboe <jaxboe@fusionio.com>