David, On 2019/06/18 8:59, David Sterba wrote: > On Mon, Jun 17, 2019 at 03:16:05AM +0000, Damien Le Moal wrote: >> Josef, >> >> On 2019/06/13 23:15, Josef Bacik wrote: >>> On Fri, Jun 07, 2019 at 10:10:17PM +0900, Naohiro Aota wrote: >>>> Sequential allocation is not enough to maintain sequential delivery of >>>> write IOs to the device. Various features (async compress, async checksum, >>>> ...) of btrfs affect ordering of the IOs. This patch introduces submit >>>> buffer to sort WRITE bios belonging to a block group and sort them out >>>> sequentially in increasing block address to achieve sequential write >>>> sequences with __btrfs_map_bio(). >>>> >>>> Signed-off-by: Naohiro Aota <naohiro.aota@xxxxxxx> >>> >>> I hate everything about this. Can't we just use the plugging infrastructure for >>> this and then make sure it re-orders the bios before submitting them? Also >>> what's to prevent the block layer scheduler from re-arranging these io's? >>> Thanks, >> >> The block I/O scheduler reorders requests in LBA order, but that happens for a >> newly inserted request against pending requests. If there are no pending >> requests because all requests were already issued, no ordering happen, and even >> worse, if the drive queue is not full yet (e.g. there are free tags), then the >> newly inserted request will be dispatched almost immediately, preventing >> reordering with subsequent incoming write requests to happen. > > This would be good to add to the changelog. Sure. No problem. We can add that explanation. >> The other problem is that the mq-deadline scheduler does not track zone WP >> position. Write request issuing is done regardless of the current WP value, >> solely based on LBA ordering. This means that mq-deadline will not prevent >> out-of-order, or rather, unaligned write requests. > > This seems to be the key point. Yes it is. We can also add this to the commit message explanation. >> These will not be detected >> and dispatched whenever possible. The reasons for this are that: >> 1) the disk user (the FS) has to manage zone WP positions anyway. So duplicating >> that management at the block IO scheduler level is inefficient. >> 2) Adding zone WP management at the block IO scheduler level would also need a >> write error processing path to resync the WP value in case of failed writes. But >> the user/FS also needs that anyway. Again duplicated functionalities. >> 3) The block layer will need a timeout to force issue or cancel pending >> unaligned write requests. This is necessary in case the drive user stops issuing >> writes (for whatever reasons) or the scheduler is being switched. This would >> unnecessarily cause write I/O errors or cause deadlocks if the request queue >> quiesce mode is entered at the wrong time (and I do not see a good way to deal >> with that). >> >> blk-mq is already complicated enough. Adding this to the block IO scheduler will >> unnecessarily complicate things further for no real benefits. I would like to >> point out the dm-zoned device mapper and f2fs which are both already dealing >> with write ordering and write error processing directly. Both are fairly >> straightforward but completely different and each optimized for their own structure. > > So the question is where on which layer the decision logic is. The > filesystem(s) or dm-zoned have enough information about the zones and > the writes can be pre-sorted. This is what the patch proposes. Yes, exactly. > From your explanation I get that the io scheduler can throw the wrench > in the sequential ordering, for various reasons depending on state of > internal structures od device queues. This is my simplified > interpretation as I don't understand all the magic below filesystem > layer. Not exactly "throw the wrench". mq-deadline will guarantee per zone write order to be exactly the order in which requests were inserted, that is, issued by the FS. But mq-dealine will not "wait" if the write order is not purely sequential, that is, there are holes/jumps in the LBA sequence for the zone. Order only is guaranteed. The alignment to WP/contiguous sequential write issuing is the responsibility of the issuer (FS or DM or application in the case of raw accesses). > I assume there are some guarantees about the ordering, eg. within one > plug, that apply to all schedulers (maybe not the noop one). Something > like that should be the least common functionality that the filesystem > layer can rely on. The insertion side of the scheduler (upper level from FS to scheduler), which include the per CPU software queues and plug control, will not reorder requests. However, the dispatch side (lower level, from scheduler to HBA driver) can cause reordering. This is what mq-deadline prevents using a per zone write lock to avoid reordering of write requests per zone by allowing only a single write request per zone to be dispatched to the device at any time. Overall order is not guaranteed, nor is read request order. But per zone write requests will not be reordered. But again, this is only ordering. Nothing to do with trying to achieve a purely sequential write stream per zone. This is the responsibility of the issuer to deliver write request per zone without any gap, all requests sequential in LBA within each zone. Overall, the stream of request does not have to be sequential, e.g. if multiple zones are being written at the same time. But per zones, write requests must be sequential. >> Naohiro changes to btrfs IO scheduler have the same intent, that is, efficiently >> integrate and handle write ordering "a la btrfs". Would creating a different >> "hmzoned" btrfs IO scheduler help address your concerns ? > > IMHO this sounds both the same, all we care about is the sequential > ordering, which in some sense is "scheduling", but I would not call it > that way due to the simplicity. OK. And yes, it is only ordering of writes per zone. For all other requests, e.g. reads, order does not matter. And the overall interleaving of write requests to different zones can also be anything. No constraints there. > As implemented, it's a list of bios, but I'd suggest using rb-tree or > xarray, the insertion is fast and submission is start to end traversal. > I'm not sure that the loop in __btrfs_map_bio_zoned after label > send_bios: has reasonable complexity, looks like an O(N^2). OK. We can change that. rbtree is simple enough to use. We can change the list to that. Thank you for your comments. Best regards. -- Damien Le Moal Western Digital Research
