diff options
author | Tejun Heo <htejun@gmail.com> | 2006-01-16 09:45:58 +0100 |
---|---|---|
committer | Jens Axboe <axboe@suse.de> | 2006-01-16 09:45:58 +0100 |
commit | 5a225791374dca35bba871a007c78a3d0b14b559 (patch) | |
tree | e9dd684069c945163af2a0cabfeaa481942d50f5 /Documentation | |
parent | 4a8e4a270b89030bdeb09d2f8cef7cfe9a50e54d (diff) | |
download | blackbird-op-linux-5a225791374dca35bba871a007c78a3d0b14b559.tar.gz blackbird-op-linux-5a225791374dca35bba871a007c78a3d0b14b559.zip |
barrier.txt got lost while the new barrier patchset was climbing up
the ladder to the mainline. Add it back.
Signed-off-by: Tejun Heo <htejun@gmail.com>
Signed-off-by: Jens Axboe <axboe@suse.de>
Diffstat (limited to 'Documentation')
-rw-r--r-- | Documentation/block/barrier.txt | 271 |
1 files changed, 271 insertions, 0 deletions
diff --git a/Documentation/block/barrier.txt b/Documentation/block/barrier.txt new file mode 100644 index 000000000000..03971518b222 --- /dev/null +++ b/Documentation/block/barrier.txt @@ -0,0 +1,271 @@ +I/O Barriers +============ +Tejun Heo <htejun@gmail.com>, 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 contraint 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 +category. + +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 physcial 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 +gurantee 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 +request. + +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)(request_queue_t *q, struct request *rq); + +int blk_queue_ordered(request_queue_t *q, unsigned ordered, + prepare_flush_fn *prepare_flush_fn, + unsigned gfp_mask); + +int blk_queue_ordered_locked(request_queue_t *q, unsigned ordered, + prepare_flush_fn *prepare_flush_fn, + unsigned gfp_mask); + +The only difference between the two functions is whether or not the +caller is holding q->queue_lock on entry. The latter expects the +caller is holding the lock. + +@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 +@gfp_mask : gfp_mask used when allocating data structures + for ordered processing + +For example, SCSI disk driver's prepare_flush_fn looks like the +following. + +static void sd_prepare_flush(request_queue_t *q, struct request *rq) +{ + memset(rq->cmd, 0, sizeof(rq->cmd)); + rq->flags |= REQ_BLOCK_PC; + rq->timeout = SD_TIMEOUT; + rq->cmd[0] = SYNCHRONIZE_CACHE; +} + +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 + + +QUEUE_ORDERED_NONE + I/O barriers are not needed and/or supported. + + Sequence: N/A + +QUEUE_ORDERED_DRAIN + Requests are ordered by draining the request queue and cache + flushing isn't needed. + + Sequence: drain => barrier + +QUEUE_ORDERED_DRAIN_FLUSH + Requests are ordered by draining the request queue and both + pre-barrier and post-barrier cache flushings are needed. + + Sequence: drain => preflush => barrier => postflush + +QUEUE_ORDERED_DRAIN_FUA + 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 + +QUEUE_ORDERED_TAG + Requests are ordered by ordered tag and cache flushing isn't + needed. + + Sequence: barrier + +QUEUE_ORDERED_TAG_FLUSH + Requests are ordered by ordered tag and both pre-barrier and + post-barrier cache flushings are needed. + + Sequence: preflush -> barrier -> postflush + +QUEUE_ORDERED_TAG_FUA + 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. +QUEUE_ORDERED_TAG_FLUSH is in use. + + [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. |