summaryrefslogtreecommitdiffstats
path: root/drivers/block/loop.h
Commit message (Collapse)AuthorAgeFilesLines
* block: loop: support DIO & AIOMing Lei2015-09-231-0/+2
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | There are at least 3 advantages to use direct I/O and AIO on read/write loop's backing file: 1) double cache can be avoided, then memory usage gets decreased a lot 2) not like user space direct I/O, there isn't cost of pinning pages 3) avoid context switch for obtaining good throughput - in buffered file read, random I/O top throughput is often obtained only if they are submitted concurrently from lots of tasks; but for sequential I/O, most of times they can be hit from page cache, so concurrent submissions often introduce unnecessary context switch and can't improve throughput much. There was such discussion[1] to use non-blocking I/O to improve the problem for application. - with direct I/O and AIO, concurrent submissions can be avoided and random read throughput can't be affected meantime xfstests(-g auto, ext4) is basically passed when running with direct I/O(aio), one exception is generic/232, but it failed in loop buffered I/O(4.2-rc6-next-20150814) too. Follows the fio test result for performance purpose: 4 jobs fio test inside ext4 file system over loop block 1) How to run - KVM: 4 VCPUs, 2G RAM - linux kernel: 4.2-rc6-next-20150814(base) with the patchset - the loop block is over one image on SSD. - linux psync, 4 jobs, size 1500M, ext4 over loop block - test result: IOPS from fio output 2) Throughput(IOPS) becomes a bit better with direct I/O(aio) ------------------------------------------------------------- test cases |randread |read |randwrite |write | ------------------------------------------------------------- base |8015 |113811 |67442 |106978 ------------------------------------------------------------- base+loop aio |8136 |125040 |67811 |111376 ------------------------------------------------------------- - somehow, it should be caused by more page cache avaiable for application or one extra page copy is avoided in case of direct I/O 3) context switch - context switch decreased by ~50% with loop direct I/O(aio) compared with loop buffered I/O(4.2-rc6-next-20150814) 4) memory usage from /proc/meminfo ------------------------------------------------------------- | Buffers | Cached ------------------------------------------------------------- base | > 760MB | ~950MB ------------------------------------------------------------- base+loop direct I/O(aio) | < 5MB | ~1.6GB ------------------------------------------------------------- - so there are much more page caches available for application with direct I/O [1] https://lwn.net/Articles/612483/ Signed-off-by: Ming Lei <ming.lei@canonical.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
* block: loop: prepare for supporing direct IOMing Lei2015-09-231-0/+1
| | | | | | | | | | | | | | | | | | | | This patches provides one interface for enabling direct IO from user space: - userspace(such as losetup) can pass 'file' which is opened/fcntl as O_DIRECT Also __loop_update_dio() is introduced to check if direct I/O can be used on current loop setting. The last big change is to introduce LO_FLAGS_DIRECT_IO flag for userspace to know if direct IO is used to access backing file. Cc: linux-api@vger.kernel.org Signed-off-by: Ming Lei <ming.lei@canonical.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
* block: loop: use kthread_workMing Lei2015-09-231-6/+4
| | | | | | | | | | | | | | | | | | | | | | | The following patch will use dio/aio to submit IO to backing file, then it needn't to schedule IO concurrently from work, so use kthread_work for decreasing context switch cost a lot. For non-AIO case, single thread has been used for long long time, and it was just converted to work in v4.0, which has caused performance regression for fedora live booting already. In discussion[1], even though submitting I/O via work concurrently can improve random read IO throughput, meantime it might hurt sequential read IO performance, so better to restore to single thread behaviour. For the following AIO support, it is better to use multi hw-queue with per-hwq kthread than current work approach suppose there is so high performance requirement for loop. [1] http://marc.info/?t=143082678400002&r=1&w=2 Signed-off-by: Ming Lei <ming.lei@canonical.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@fb.com>
* block: loop: don't hold lo_ctl_mutex in lo_openMing Lei2015-05-201-1/+1
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The lo_ctl_mutex is held for running all ioctl handlers, and in some ioctl handlers, ioctl_by_bdev(BLKRRPART) is called for rereading partitions, which requires bd_mutex. So it is easy to cause failure because trylock(bd_mutex) may fail inside blkdev_reread_part(), and follows the lock context: blkid or other application: ->open() ->mutex_lock(bd_mutex) ->lo_open() ->mutex_lock(lo_ctl_mutex) losetup(set fd ioctl): ->mutex_lock(lo_ctl_mutex) ->ioctl_by_bdev(BLKRRPART) ->trylock(bd_mutex) This patch trys to eliminate the ABBA lock dependency by removing lo_ctl_mutext in lo_open() with the following approach: 1) make lo_refcnt as atomic_t and avoid acquiring lo_ctl_mutex in lo_open(): - for open vs. add/del loop, no any problem because of loop_index_mutex - freeze request queue during clr_fd, so I/O can't come until clearing fd is completed, like the effect of holding lo_ctl_mutex in lo_open - both open() and release() have been serialized by bd_mutex already 2) don't hold lo_ctl_mutex for decreasing/checking lo_refcnt in lo_release(), then lo_ctl_mutex is only required for the last release. Reviewed-by: Christoph Hellwig <hch@lst.de> Tested-by: Jarod Wilson <jarod@redhat.com> Acked-by: Jarod Wilson <jarod@redhat.com> Signed-off-by: Ming Lei <ming.lei@canonical.com> Signed-off-by: Jens Axboe <axboe@fb.com>
* block: loop: convert to per-device workqueueMing Lei2015-05-051-0/+1
| | | | | | | | | | | | | | | | | Documentation/workqueue.txt: If there is dependency among multiple work items used during memory reclaim, they should be queued to separate wq each with WQ_MEM_RECLAIM. Loop devices can be stacked, so we have to convert to per-device workqueue. One example is Fedora live CD. Fixes: b5dd2f6047ca108001328aac0e8588edd15f1778 Cc: stable@vger.kernel.org (v4.0) Cc: Justin M. Forbes <jforbes@fedoraproject.org> Signed-off-by: Ming Lei <ming.lei@canonical.com> Acked-by: Tejun Heo <tj@kernel.org> Signed-off-by: Jens Axboe <axboe@fb.com>
* loop: add blk-mq.h includeJens Axboe2015-01-021-0/+1
| | | | | | | | | | | | | Looks like we pull it in through other ways on x86, but we fail on sparc: In file included from drivers/block/cryptoloop.c:30:0: drivers/block/loop.h:63:24: error: field 'tag_set' has incomplete type struct blk_mq_tag_set tag_set; Add the include to loop.h, kill it from loop.c. Signed-off-by: Jens Axboe <axboe@fb.com>
* block: loop: improve performance via blk-mqMing Lei2015-01-021-6/+11
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | The conversion is a bit straightforward, and use work queue to dispatch requests of loop block, and one big change is that requests is submitted to backend file/device concurrently with work queue, so throughput may get improved much. Given write requests over same file are often run exclusively, so don't handle them concurrently for avoiding extra context switch cost, possible lock contention and work schedule cost. Also with blk-mq, there is opportunity to get loop I/O merged before submitting to backend file/device. In the following test: - base: v3.19-rc2-2041231 - loop over file in ext4 file system on SSD disk - bs: 4k, libaio, io depth: 64, O_DIRECT, num of jobs: 1 - throughput: IOPS ------------------------------------------------------ | | base | base with loop-mq | delta | ------------------------------------------------------ | randread | 1740 | 25318 | +1355%| ------------------------------------------------------ | read | 42196 | 51771 | +22.6%| ----------------------------------------------------- | randwrite | 35709 | 34624 | -3% | ----------------------------------------------------- | write | 39137 | 40326 | +3% | ----------------------------------------------------- So loop-mq can improve throughput for both read and randread, meantime, performance of write and randwrite isn't hurted basically. Another benefit is that loop driver code gets simplified much after blk-mq conversion, and the patch can be thought as cleanup too. Signed-off-by: Ming Lei <ming.lei@canonical.com> Signed-off-by: Jens Axboe <axboe@fb.com>
* move linux/loop.h to drivers/blockAl Viro2013-06-291-0/+85
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>