Commit ae1b1539622fb46e51b4d13b3f9e5f4c713f86ae, block: reimplement
FLUSH/FUA to support merge, introduced a performance regression when
running any sort of fsyncing workload using dm-multipath and certain
storage (in our case, an HP EVA).  The test I ran was fs_mark, and it
dropped from ~800 files/sec on ext4 to ~100 files/sec.  It turns out
that dm-multipath always advertised flush+fua support, and passed
commands on down the stack, where those flags used to get stripped off.
The above commit changed that behavior:

static inline struct request *__elv_next_request(struct request_queue *q)
{
        struct request *rq;

        while (1) {
-               while (!list_empty(&q->queue_head)) {
+               if (!list_empty(&q->queue_head)) {
                        rq = list_entry_rq(q->queue_head.next);
-                       if (!(rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) ||
-                           (rq->cmd_flags & REQ_FLUSH_SEQ))
-                               return rq;
-                       rq = blk_do_flush(q, rq);
-                       if (rq)
-                               return rq;
+                       return rq;
                }

Note that previously, a command would come in here, have
REQ_FLUSH|REQ_FUA set, and then get handed off to blk_do_flush:

struct request *blk_do_flush(struct request_queue *q, struct request *rq)
{
        unsigned int fflags = q->flush_flags; /* may change, cache it */
        bool has_flush = fflags & REQ_FLUSH, has_fua = fflags & REQ_FUA;
        bool do_preflush = has_flush && (rq->cmd_flags & REQ_FLUSH);
        bool do_postflush = has_flush && !has_fua && (rq->cmd_flags &
        REQ_FUA);
        unsigned skip = 0;
...
        if (blk_rq_sectors(rq) && !do_preflush && !do_postflush) {
                rq->cmd_flags &= ~REQ_FLUSH;
		if (!has_fua)
			rq->cmd_flags &= ~REQ_FUA;
	        return rq;
	}

So, the flush machinery was bypassed in such cases (q->flush_flags == 0
&& rq->cmd_flags & (REQ_FLUSH|REQ_FUA)).

Now, however, we don't get into the flush machinery at all.  Instead,
__elv_next_request just hands a request with flush and fua bits set to
the scsi_request_fn, even if the underlying request_queue does not
support flush or fua.

The agreed upon approach is to fix the flush machinery to allow
stacking.  While this isn't used in practice (since there is only one
request-based dm target, and that target will now reflect the flush
flags of the underlying device), it does future-proof the solution, and
make it function as designed.

In order to make this work, I had to add a field to the struct request,
inside the flush structure (to store the original req->end_io).  Shaohua
had suggested overloading the union with rb_node and completion_data,
but the completion data is used by device mapper and can also be used by
other drivers.  So, I didn't see a way around the additional field.

I tested this patch on an HP EVA with both ext4 and xfs, and it recovers
the lost performance.  Comments and other testers, as always, are
appreciated.

Cheers,
Jeff

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

This patch merges in a fix that missed 2.6.39 final.

Conflicts:
	block/blk.h

Since for-2.6.40/core was forked off the 2.6.39 devel tree, we've
had churn in the core area that makes it difficult to handle
patches for eg cfq or blk-throttle. Instead of requiring that they
be based in older versions with bugs that have been fixed later
in the rc cycle, merge in 2.6.39 final.

Also fixes up conflicts in the below files.

Conflicts:
	drivers/block/paride/pcd.c
	drivers/cdrom/viocd.c
	drivers/ide/ide-cd.c

Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

blk_cleanup_queue() calls elevator_exit() and after this, we can't
touch the elevator without oopsing.  __elv_next_request() must check
for this state because in the refcounted queue model, we can still
call it after blk_cleanup_queue() has been called.

This was reported as causing an oops attributable to scsi.

Signed-off-by: James Bottomley <James.Bottomley@suse.de>
Cc: stable@kernel.org
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

In some drives, flush requests are non-queueable. When flush request is
running, normal read/write requests can't run. If block layer dispatches
such request, driver can't handle it and requeue it.  Tejun suggested we
can hold the queue when flush is running. This can avoid unnecessary
requeue.  Also this can improve performance. For example, we have
request flush1, write1, flush 2. flush1 is dispatched, then queue is
hold, write1 isn't inserted to queue. After flush1 is finished, flush2
will be dispatched. Since disk cache is already clean, flush2 will be
finished very soon, so looks like flush2 is folded to flush1.

In my test, the queue holding completely solves a regression introduced by
commit 53d63e6b0dfb95882ec0219ba6bbd50cde423794:

    block: make the flush insertion use the tail of the dispatch list

    It's not a preempt type request, in fact we have to insert it
    behind requests that do specify INSERT_FRONT.

which causes about 20% regression running a sysbench fileio
workload.

Stable: 2.6.39 only

Cc: stable@kernel.org
Signed-off-by: Shaohua Li <shaohua.li@intel.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

We are currently using this flag to check whether it's safe
to call into ->request_fn(). If it is set, we punt to kblockd.
But we get a lot of false positives and excessive punts to
kblockd, which hurts performance.

The only real abuser of this infrastructure is SCSI. So export
the async queue run and convert SCSI over to use that. There's
room for improvement in that SCSI need not always use the async
call, but this fixes our performance issue and they can fix that
up in due time.

Signed-off-by: Jens 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: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

Fixes generated by 'codespell' and manually reviewed.

Signed-off-by: Lucas De Marchi <lucas.demarchi@profusion.mobi>

One of the disadvantages of on-stack plugging is that we potentially
lose out on merging since all pending IO isn't always visible to
everybody. When we flush the on-stack plugs, right now we don't do
any checks to see if potential merge candidates could be utilized.

Correct this by adding a new insert variant, ELEVATOR_INSERT_SORT_MERGE.
It works just ELEVATOR_INSERT_SORT, but first checks whether we can
merge with an existing request before doing the insertion (if we fail
merging).

This fixes a regression with multiple processes issuing IO that
can be merged.

Thanks to Shaohua Li <shaohua.li@intel.com> for testing and fixing
an accounting bug.

Signed-off-by: Jens 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: Jens Axboe <jaxboe@fusionio.com>