The block I/O + elevator + I/O scheduler code spend a lot of time trying
to merge I/Os -- rightfully so under "normal" circumstances. However,
if one were to know that the incoming I/O stream was /very/ random in
nature, the cycles are wasted.

This patch adds a per-request_queue tunable that (when set) disables
merge attempts (beyond the simple one-hit cache check), thus freeing up
a non-trivial amount of CPU cycles.
Signed-off-by: NAlan D. Brunelle <alan.brunelle@hp.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch changes rq->cmd from the static array to a pointer to
support large commands.

We rarely handle large commands. So for optimization, a struct request
still has a static array for a command. rq_init sets rq->cmd pointer
to the static array.
Signed-off-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This rename rq_init() blk_rq_init() and export it. Any path that hands
the request to the block layer needs to call it to initialize the
request.

This is a preparation for large command support, which needs to
initialize the request in a proper way (that is, just doing a memset()
will not work).
Signed-off-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

We can save some atomic ops in the IO path, if we clearly define
the rules of how to modify the queue flags.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Only noticed this while hacking something else, no test case.

blk_max_low_pfn is initialized once at bootup by the block layer from
max_low_pfn.  But max_low_pfn is not necessarily constant over the runtime of
the system when you consider memory hotplug.  What could happen if that
someone adds memory later the block layer wouldn't get updated and then start
bouncing memory unnecessarily.

Also on 64bit blk_max_low_pfn actually isn't needed because it just disables
bouncing essentially and there is no highmem.  And nobody can pass pfns >
max_low_pfn to the block layer, because those wouldn't have a struct page and
I suspect block layer wouldn't be very happy without that.

So set BLK_BOUNCE_HIGH to infinity (-1ULL) on 64bit.  That avoids the problem
of having to update it on memory hotadd.

On 32bit I kept the same behaviour because at least on i386
memory hotadd only adds HIGHMEM, never lowmem.

BLK_BOUNCE_ANY is always set to infinity on both 32 and 64bit.
Signed-off-by: NAndi Kleen <ak@suse.de>
Cc: Jens Axboe <jens.axboe@oracle.com>
Acked-by: NYasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

blk_rq_map_user adjusts bi_size of the last bio. It breaks the rule
that req->data_len (the true data length) is equal to sum(bio). It
broke the scsi command completion code.

commit e97a294e was introduced to fix
the above issue. However, the partial completion code doesn't work
with it. The commit is also a layer violation (scsi mid-layer should
not know about the block layer's padding).

This patch moves the padding adjustment to blk_rq_map_sg (suggested by
James). The padding works like the drain buffer. This patch breaks the
rule that req->data_len is equal to sum(sg), however, the drain buffer
already broke it. So this patch just restores the rule that
req->data_len is equal to sub(bio) without breaking anything new.

Now when a low level driver needs padding, blk_rq_map_user and
blk_rq_map_user_iov guarantee there's enough room for padding.
blk_rq_map_sg can safely extend the last entry of a scatter list.

blk_rq_map_sg must extend the last entry of a scatter list only for a
request that got through bio_copy_user_iov. This patches introduces
new REQ_COPY_USER flag.
Signed-off-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Tejun Heo <htejun@gmail.com>
Cc: Mike Christie <michaelc@cs.wisc.edu>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Block layer alignment was used for two different purposes - memory
alignment and padding.  This causes problems in lower layers because
drivers which only require memory alignment ends up with adjusted
rq->data_len.  Separate out padding such that padding occurs iff
driver explicitly requests it.

Tomo: restorethe code to update bio in blk_rq_map_user
      introduced by the commit 40b01b9b
      according to padding alignment.
Signed-off-by: NTejun Heo <htejun@gmail.com>
Signed-off-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

The meaning of rq->data_len was changed to the length of an allocated
buffer from the true data length. It breaks SG_IO friends and
bsg. This patch restores the meaning of rq->data_len to the true data
length and adds rq->extra_len to store an extended length (due to
drain buffer and padding).

This patch also removes the code to update bio in blk_rq_map_user
introduced by the commit 40b01b9b.
The commit adjusts bio according to memory alignment
(queue_dma_alignment). However, memory alignment is NOT padding
alignment. This adjustment also breaks SG_IO friends and bsg. Padding
alignment needs to be fixed in a proper way (by a separate patch).
Signed-off-by: NFUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: NJens Axboe <axboe@carl.home.kernel.dk>

Draining shouldn't be done for commands where overflow may indicate
data integrity issues.  Add dma_drain_needed callback to
request_queue.  Drain buffer is appened iff this function returns
non-zero.
Signed-off-by: NTejun Heo <htejun@gmail.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

With padding and draining moved into it, block layer now may extend
requests as directed by queue parameters, so now a request has two
sizes - the original request size and the extended size which matches
the size of area pointed to by bios and later by sgs.  The latter size
is what lower layers are primarily interested in when allocating,
filling up DMA tables and setting up the controller.

Both padding and draining extend the data area to accomodate
controller characteristics.  As any controller which speaks SCSI can
handle underflows, feeding larger data area is safe.

So, this patch makes the primary data length field, request->data_len,
indicate the size of full data area and add a separate length field,
request->raw_data_len, for the unmodified request size.  The latter is
used to report to higher layer (userland) and where the original
request size should be fed to the controller or device.
Signed-off-by: NTejun Heo <htejun@gmail.com>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Rearrange fields in cache order and initialize some fields that
we didn't previously init. Remove init of ->completion_data, it's
part of a union with ->hash. Luckily clearing the rb node is the same
as setting it to null!
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

It blindly copies everything in the io_context, including the lock.
That doesn't work so well for either lock ordering or lockdep.

There seems zero point in swapping io contexts on a request to request
merge, so the best point of action is to just remove it.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

No point in passing signed integers as the byte count, they can never
be negative.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

struct io_context was not defined, just add an empty forward decl.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

These DMA drain buffer implementations in drivers are pretty horrible
to do in terms of manipulating the scatterlist.  Plus they're being
done at least in drivers/ide and drivers/ata, so we now have code
duplication.

The one use case for this, as I understand it is AHCI controllers doing
PIO mode to mmc devices but translating this to DMA at the controller
level.

So, what about adding a callback to the block layer that permits the
adding of the drain buffer for the problem devices.  The idea is that
you'd do this in slave_configure after you find one of these devices.

The beauty of doing it in the block layer is that it quietly adds the
drain buffer to the end of the sg list, so it automatically gets mapped
(and unmapped) without anything unusual having to be done to the
scatterlist in driver/scsi or drivers/ata and without any alteration to
the transfer length.
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Detach task state from ioc, instead keep track of how many processes
are accessing the ioc.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This is where it belongs and then it doesn't take up space for a
process that doesn't do IO.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch converts 'uptodate' arguments of no longer exported
interfaces, end_that_request_first/last, to 'error', and removes
internal conversions for it in blk_end_request interfaces.

Also, this patch removes no longer needed end_io_error().

Cc: Boaz Harrosh <bharrosh@panasas.com>
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch removes the following functions:
  o end_that_request_first()
  o end_that_request_chunk()
and stops exporting the functions below:
  o end_that_request_last()

Cc: Boaz Harrosh <bharrosh@panasas.com>
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch adds a variant of the interface, blk_end_bidi_request(),
which completes a bidi request.

Bidi request must be completed as a whole, both rq and rq->next_rq
at once.  So the interface has 2 arguments for completion size.

As for ->end_io, only rq->end_io is called (rq->next_rq->end_io is not
called).  So if special completion handling is needed, the handler
must be set to rq->end_io.
And the handler must take care of freeing next_rq too, since
the interface doesn't care of it if rq->end_io is not NULL.

Cc: Boaz Harrosh <bharrosh@panasas.com>
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch adds a variant of the interface, blk_end_request_callback(),
which has driver callback feature.

Drivers may need to do special works between end_that_request_first()
and end_that_request_last().
For such drivers, blk_end_request_callback() allows it to pass
a callback function which is called between end_that_request_first()
and end_that_request_last().

This interface is only for fallback of other blk_end_request interfaces.
Drivers should avoid their tricky behaviors and use other interfaces
as much as possible.

Currently, only one driver, ide-cd, needs this interface.
So this interface should/will be removed, after the driver removes
such tricky behaviors.

o ide-cd (cdrom_newpc_intr())
  In PIO mode, cdrom_newpc_intr() needs to defer end_that_request_last()
  until the device clears DRQ_STAT and raises an interrupt after
  end_that_request_first().
  So end_that_request_first() and end_that_request_last() are called
  separately in cdrom_newpc_intr().

  This means blk_end_request_callback() has to return without
  completing request even if no leftover in the request.
  To satisfy the requirement, callback function has return value
  so that drivers can tell blk_end_request_callback() to return
  without completing request.
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch adds/exports functions to get the size of request in bytes.
They are useful because blk_end_request interfaces take bytes
as a completed I/O size instead of sectors.
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This patch adds 2 new interfaces for request completion:
  o blk_end_request()   : called without queue lock
  o __blk_end_request() : called with queue lock held

blk_end_request takes 'error' as an argument instead of 'uptodate',
which current end_that_request_* take.
The meanings of values are below and the value is used when bio is
completed.
    0 : success
  < 0 : error

Some device drivers call some generic functions below between
end_that_request_{first/chunk} and end_that_request_last().
  o add_disk_randomness()
  o blk_queue_end_tag()
  o blkdev_dequeue_request()
These are called in the blk_end_request interfaces as a part of
generic request completion.
So all device drivers become to call above functions.
To decide whether to call blkdev_dequeue_request(), blk_end_request
uses list_empty(&rq->queuelist) (blk_queued_rq() macro is added for it).
So drivers must re-initialize it using list_init() or so before calling
blk_end_request if drivers use it for its specific purpose.
(Currently, there is no driver which completes request without
 re-initializing the queuelist after used it.  So rq->queuelist
 can be used for the purpose above.)

"Normal" drivers can be converted to use blk_end_request()
in a standard way shown below.

 a) end_that_request_{chunk/first}
    spin_lock_irqsave()
    (add_disk_randomness(), blk_queue_end_tag(), blkdev_dequeue_request())
    end_that_request_last()
    spin_unlock_irqrestore()
    => blk_end_request()

 b) spin_lock_irqsave()
    end_that_request_{chunk/first}
    (add_disk_randomness(), blk_queue_end_tag(), blkdev_dequeue_request())
    end_that_request_last()
    spin_unlock_irqrestore()
    => spin_lock_irqsave()
       __blk_end_request()
       spin_unlock_irqsave()

 c) spin_lock_irqsave()
    (add_disk_randomness(), blk_queue_end_tag(), blkdev_dequeue_request())
    end_that_request_last()
    spin_unlock_irqrestore()
    => blk_end_request()   or   spin_lock_irqsave()
                                __blk_end_request()
                                spin_unlock_irqrestore()
Signed-off-by: NKiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Let queue_dma_alignment return 0 if it was specifically set to 0.
This permits devices with no particular alignment restrictions to
use arbitrary user space buffers without copying.
Signed-off-by: NPete Wyckoff <pw@osc.edu>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Based on the earlier work by Tejun Heo.

All users are gone so we can finally remove it.

Cc: Tejun Heo <htejun@gmail.com>
Signed-off-by: NBartlomiej Zolnierkiewicz <bzolnier@gmail.com>

Based on the earlier work by Tejun Heo.

All users are gone so we can finally remove it.

Cc: Tejun Heo <htejun@gmail.com>
Signed-off-by: NBartlomiej Zolnierkiewicz <bzolnier@gmail.com>

The purpose of this is to allow stacked alignment settings, with the
ultimate queue alignment being set to the largest alignment requirement
in the stack.

The reason for this is so that the SCSI mid-layer can relax the default
alignment requirements (which are basically causing a lot of superfluous
copying to go on in the SG_IO interface) while allowing transports,
devices or HBAs to add stricter limits if they need them.
Acked-by: NJens Axboe <jens.axboe@oracle.com>
Signed-off-by: NJames Bottomley <James.Bottomley@HansenPartnership.com>

Added blk_unplug interface, allowing all invocations of unplugs to result
in a generated blktrace UNPLUG.
Signed-off-by: NAlan D. Brunelle <Alan.Brunelle@hp.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

For the locking to work, only the tag map and tag bit map may be shared
(incidentally, I was just explaining this to Nick yesterday, but I
apparently didn't review the code well enough myself). But we also share
the busy list!  The busy_list must be queue private, or we need a
block_queue_tag covering lock as well.

So we have to move the busy_list to the queue. This'll work fine, and
it'll actually also fix a problem with blk_queue_invalidate_tags() which
will invalidate tags across all shared queues. This is a bit confusing,
the low level driver should call it for each queue seperately since
otherwise you cannot kill tags on just a single queue for eg a hard
drive that stops responding. Since the function has no callers
currently, it's not an issue.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Then we can get rid of ->issue_flush_fn() and all the driver private
implementations of that.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

This implements functionality to pass down or insert a barrier
in a queue, without having data attached to it. The ->prepare_flush_fn()
infrastructure from data barriers are reused to provide this
functionality.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

We can use this helper in the elevator core for BLKPREP_KILL, and it'll
also be useful for the empty barrier patch.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

We need it even if CONFIG_BLOCK is disabled, so move it outside of
the block layer include system.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

The entire function of flush_dry_bio_endio is to undo the effects
of bio_endio (when called on a barrier request).  So remove the
function and the call to bio_endio.

This allows us to remove "bi_size" from "struct request_queue".
Signed-off-by: NNeil Brown <neilb@suse.de>

### Diffstat output
 ./block/ll_rw_blk.c      |   39 ++-------------------------------------
 ./include/linux/blkdev.h |    1 -
 2 files changed, 2 insertions(+), 38 deletions(-)

diff .prev/block/ll_rw_blk.c ./block/ll_rw_blk.c
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Hide everything in blkdev.h with CONFIG_BLOCK isn't set, and fixup
the (few) files that fail to build because they were relying on blkdev.h
pulling in extra includes for them.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

blk_rq_bio_prep is exported for use in exactly
one place.  That place can benefit from using
the new blk_rq_append_bio instead.
So
  - change dm-emc to call blk_rq_append_bio
  - stop exporting blk_rq_bio_prep, and
  - initialise rq_disk in blk_rq_bio_prep,
       as dm-emc needs it.
Signed-off-by: NNeil Brown <neilb@suse.de>

diff .prev/block/ll_rw_blk.c ./block/ll_rw_blk.c
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

ll_back_merge_fn is currently exported to SCSI where is it used,
together with blk_rq_bio_prep, in exactly the same way these
functions are used in __blk_rq_map_user.

So move the common code into a new function (blk_rq_append_bio), and
don't export ll_back_merge_fn any longer.
Signed-off-by: NNeil Brown <neilb@suse.de>

diff .prev/block/ll_rw_blk.c ./block/ll_rw_blk.c
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Every usage of rq_for_each_bio wraps a usage of
bio_for_each_segment, so these can be combined into
rq_for_each_segment.

We define "struct req_iterator" to hold the 'bio' and 'index' that
are needed for the double iteration.
Signed-off-by: NNeil Brown <neilb@suse.de>

Various compile fixes by me...
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

The two comments were transposed.
Signed-off-by: NQi Yong <qiyong@mail.fc-cn.com>
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Andrew thinks I should be nice and allow outside code to at least just
compile, so add the request_queue_t typedef back and mark it deprecated.
It'll warn people that this type is going away soonish.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>