free electrons

Embedded Linux Experts

I/O Barriers
Tejun Heo <>, July 22 2005

I/O barrier requests are used to guarantee ordering around the barrier
requests.  Unless you're crazy enough to use disk drives for
implementing synchronization constructs (wow, sounds interesting...),
the ordering is meaningful only for write requests for things like
journal checkpoints.  All requests queued before a barrier request
must be finished (made it to the physical medium) before the barrier
request is started, and all requests queued after the barrier request
must be started only after the barrier request is finished (again,
made it to the physical medium).

In other words, I/O barrier requests have the following two properties.

1. Request ordering

Requests cannot pass the barrier request.  Preceding requests are
processed before the barrier and following requests after.

Depending on what features a drive supports, this can be done in one
of the following three ways.

i. For devices which have queue depth greater than 1 (TCQ devices) and
support ordered tags, block layer can just issue the barrier as an
ordered request and the lower level driver, controller and drive
itself are responsible for making sure that the ordering constraint is
met.  Most modern SCSI controllers/drives should support this.

NOTE: SCSI ordered tag isn't currently used due to limitation in the
      SCSI midlayer, see the following random notes section.

ii. For devices which have queue depth greater than 1 but don't
support ordered tags, block layer ensures that the requests preceding
a barrier request finishes before issuing the barrier request.  Also,
it defers requests following the barrier until the barrier request is
finished.  Older SCSI controllers/drives and SATA drives fall in this

iii. Devices which have queue depth of 1.  This is a degenerate case
of ii.  Just keeping issue order suffices.  Ancient SCSI
controllers/drives and IDE drives are in this category.

2. Forced flushing to physical medium

Again, if you're not gonna do synchronization with disk drives (dang,
it sounds even more appealing now!), the reason you use I/O barriers
is mainly to protect filesystem integrity when power failure or some
other events abruptly stop the drive from operating and possibly make
the drive lose data in its cache.  So, I/O barriers need to guarantee
that requests actually get written to non-volatile medium in order.

There are four cases,

i. No write-back cache.  Keeping requests ordered is enough.

ii. Write-back cache but no flush operation.  There's no way to
guarantee physical-medium commit order.  This kind of devices can't to
I/O barriers.

iii. Write-back cache and flush operation but no FUA (forced unit
access).  We need two cache flushes - before and after the barrier

iv. Write-back cache, flush operation and FUA.  We still need one
flush to make sure requests preceding a barrier are written to medium,
but post-barrier flush can be avoided by using FUA write on the
barrier itself.

How to support barrier requests in drivers

All barrier handling is done inside block layer proper.  All low level
drivers have to are implementing its prepare_flush_fn and using one
the following two functions to indicate what barrier type it supports
and how to prepare flush requests.  Note that the term 'ordered' is
used to indicate the whole sequence of performing barrier requests
including draining and flushing.

typedef void (prepare_flush_fn)(struct request_queue *q, struct request *rq);

int blk_queue_ordered(struct request_queue *q, unsigned ordered,
		      prepare_flush_fn *prepare_flush_fn);

@q			: the queue in question
@ordered		: the ordered mode the driver/device supports
@prepare_flush_fn	: this function should prepare @rq such that it
			  flushes cache to physical medium when executed

For example, SCSI disk driver's prepare_flush_fn looks like the

static void sd_prepare_flush(struct request_queue *q, struct request *rq)
	memset(rq->cmd, 0, sizeof(rq->cmd));
	rq->cmd_type = REQ_TYPE_BLOCK_PC;
	rq->timeout = SD_TIMEOUT;
	rq->cmd[0] = SYNCHRONIZE_CACHE;
	rq->cmd_len = 10;

The following seven ordered modes are supported.  The following table
shows which mode should be used depending on what features a
device/driver supports.  In the leftmost column of table,
QUEUE_ORDERED_ prefix is omitted from the mode names to save space.

The table is followed by description of each mode.  Note that in the
descriptions of QUEUE_ORDERED_DRAIN*, '=>' is used whereas '->' is
used for QUEUE_ORDERED_TAG* descriptions.  '=>' indicates that the
preceding step must be complete before proceeding to the next step.
'->' indicates that the next step can start as soon as the previous
step is issued.

	    write-back cache	ordered tag	flush		FUA
NONE		yes/no		N/A		no		N/A
DRAIN		no		no		N/A		N/A
DRAIN_FLUSH	yes		no		yes		no
DRAIN_FUA	yes		no		yes		yes
TAG		no		yes		N/A		N/A
TAG_FLUSH	yes		yes		yes		no
TAG_FUA		yes		yes		yes		yes

	I/O barriers are not needed and/or supported.

	Sequence: N/A

	Requests are ordered by draining the request queue and cache
	flushing isn't needed.

	Sequence: drain => barrier

	Requests are ordered by draining the request queue and both
	pre-barrier and post-barrier cache flushings are needed.

	Sequence: drain => preflush => barrier => postflush

	Requests are ordered by draining the request queue and
	pre-barrier cache flushing is needed.  By using FUA on barrier
	request, post-barrier flushing can be skipped.

	Sequence: drain => preflush => barrier

	Requests are ordered by ordered tag and cache flushing isn't

	Sequence: barrier

	Requests are ordered by ordered tag and both pre-barrier and
	post-barrier cache flushings are needed.

	Sequence: preflush -> barrier -> postflush

	Requests are ordered by ordered tag and pre-barrier cache
	flushing is needed.  By using FUA on barrier request,
	post-barrier flushing can be skipped.

	Sequence: preflush -> barrier

Random notes/caveats

* SCSI layer currently can't use TAG ordering even if the drive,
controller and driver support it.  The problem is that SCSI midlayer
request dispatch function is not atomic.  It releases queue lock and
switch to SCSI host lock during issue and it's possible and likely to
happen in time that requests change their relative positions.  Once
this problem is solved, TAG ordering can be enabled.

* Currently, no matter which ordered mode is used, there can be only
one barrier request in progress.  All I/O barriers are held off by
block layer until the previous I/O barrier is complete.  This doesn't
make any difference for DRAIN ordered devices, but, for TAG ordered
devices with very high command latency, passing multiple I/O barriers
to low level *might* be helpful if they are very frequent.  Well, this
certainly is a non-issue.  I'm writing this just to make clear that no
two I/O barrier is ever passed to low-level driver.

* Completion order.  Requests in ordered sequence are issued in order
but not required to finish in order.  Barrier implementation can
handle out-of-order completion of ordered sequence.  IOW, the requests
MUST be processed in order but the hardware/software completion paths
are allowed to reorder completion notifications - eg. current SCSI
midlayer doesn't preserve completion order during error handling.

* Requeueing order.  Low-level drivers are free to requeue any request
after they removed it from the request queue with
blkdev_dequeue_request().  As barrier sequence should be kept in order
when requeued, generic elevator code takes care of putting requests in
order around barrier.  See blk_ordered_req_seq() and
ELEVATOR_INSERT_REQUEUE handling in __elv_add_request() for details.

Note that block drivers must not requeue preceding requests while
completing latter requests in an ordered sequence.  Currently, no
error checking is done against this.

* Error handling.  Currently, block layer will report error to upper
layer if any of requests in an ordered sequence fails.  Unfortunately,
this doesn't seem to be enough.  Look at the following request flow.

 [0] [1] [2] [3] [pre] [barrier] [post] < [4] [5] [6] ... >
					  still in elevator

Let's say request [2], [3] are write requests to update file system
metadata (journal or whatever) and [barrier] is used to mark that
those updates are valid.  Consider the following sequence.

 i.	Requests [0] ~ [post] leaves the request queue and enters
	low-level driver.
 ii.	After a while, unfortunately, something goes wrong and the
	drive fails [2].  Note that any of [0], [1] and [3] could have
	completed by this time, but [pre] couldn't have been finished
	as the drive must process it in order and it failed before
	processing that command.
 iii.	Error handling kicks in and determines that the error is
	unrecoverable and fails [2], and resumes operation.
 iv.	[pre] [barrier] [post] gets processed.
 v.	*BOOM* power fails

The problem here is that the barrier request is *supposed* to indicate
that filesystem update requests [2] and [3] made it safely to the
physical medium and, if the machine crashes after the barrier is
written, filesystem recovery code can depend on that.  Sadly, that
isn't true in this case anymore.  IOW, the success of a I/O barrier
should also be dependent on success of some of the preceding requests,
where only upper layer (filesystem) knows what 'some' is.

This can be solved by implementing a way to tell the block layer which
requests affect the success of the following barrier request and
making lower lever drivers to resume operation on error only after
block layer tells it to do so.

As the probability of this happening is very low and the drive should
be faulty, implementing the fix is probably an overkill.  But, still,
it's there.

* In previous drafts of barrier implementation, there was fallback
mechanism such that, if FUA or ordered TAG fails, less fancy ordered
mode can be selected and the failed barrier request is retried
automatically.  The rationale for this feature was that as FUA is
pretty new in ATA world and ordered tag was never used widely, there
could be devices which report to support those features but choke when
actually given such requests.

 This was removed for two reasons 1. it's an overkill 2. it's
impossible to implement properly when TAG ordering is used as low
level drivers resume after an error automatically.  If it's ever
needed adding it back and modifying low level drivers accordingly
shouldn't be difficult.