dm-crypt: offload writes to thread Submitting write bios directly in the encryption thread caused serious performance degradation. On multiprocessor machine encryption requests finish in a different order than they were submitted in. Consequently, write requests would be submitted in a different order and it could cause severe performance degradation. This patch moves submitting write requests to a separate thread so that the requests can be sorted before submitting. Sorting in implemented in the next patch. Signed-off-by: Mikulas Patocka --- drivers/md/dm-crypt.c | 120 ++++++++++++++++++++++++++++++++++++++++---------- 1 file changed, 97 insertions(+), 23 deletions(-) Index: linux-3.8.6-fast/drivers/md/dm-crypt.c =================================================================== --- linux-3.8.6-fast.orig/drivers/md/dm-crypt.c 2013-04-09 01:50:46.000000000 +0200 +++ linux-3.8.6-fast/drivers/md/dm-crypt.c 2013-04-09 01:53:07.000000000 +0200 @@ -17,6 +17,7 @@ #include #include #include +#include #include #include #include @@ -60,6 +61,8 @@ struct dm_crypt_io { int error; sector_t sector; struct dm_crypt_io *base_io; + + struct list_head list; }; struct dm_crypt_request { @@ -123,6 +126,10 @@ struct crypt_config { struct workqueue_struct *io_queue; struct workqueue_struct *crypt_queue; + struct task_struct *write_thread; + wait_queue_head_t write_thread_wait; + struct list_head write_thread_list; + char *cipher; char *cipher_string; @@ -974,37 +981,89 @@ static int kcryptd_io_read(struct dm_cry return 0; } +static void kcryptd_io_read_work(struct work_struct *work) +{ + struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work); + + crypt_inc_pending(io); + if (kcryptd_io_read(io, GFP_NOIO)) + io->error = -ENOMEM; + crypt_dec_pending(io); +} + +static void kcryptd_queue_read(struct dm_crypt_io *io) +{ + struct crypt_config *cc = io->cc; + + INIT_WORK(&io->work, kcryptd_io_read_work); + queue_work(cc->io_queue, &io->work); +} + static void kcryptd_io_write(struct dm_crypt_io *io) { struct bio *clone = io->ctx.bio_out; + generic_make_request(clone); } -static void kcryptd_io(struct work_struct *work) +static int dmcrypt_write(void *data) { - struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work); + struct crypt_config *cc = data; + while (1) { + struct list_head local_list; + struct blk_plug plug; - if (bio_data_dir(io->base_bio) == READ) { - crypt_inc_pending(io); - if (kcryptd_io_read(io, GFP_NOIO)) - io->error = -ENOMEM; - crypt_dec_pending(io); - } else - kcryptd_io_write(io); -} + DECLARE_WAITQUEUE(wait, current); -static void kcryptd_queue_io(struct dm_crypt_io *io) -{ - struct crypt_config *cc = io->cc; + spin_lock_irq(&cc->write_thread_wait.lock); +continue_locked: - INIT_WORK(&io->work, kcryptd_io); - queue_work(cc->io_queue, &io->work); + if (!list_empty(&cc->write_thread_list)) + goto pop_from_list; + + __set_current_state(TASK_INTERRUPTIBLE); + __add_wait_queue(&cc->write_thread_wait, &wait); + + spin_unlock_irq(&cc->write_thread_wait.lock); + + if (unlikely(kthread_should_stop())) { + set_task_state(current, TASK_RUNNING); + remove_wait_queue(&cc->write_thread_wait, &wait); + break; + } + + schedule(); + + set_task_state(current, TASK_RUNNING); + spin_lock_irq(&cc->write_thread_wait.lock); + __remove_wait_queue(&cc->write_thread_wait, &wait); + goto continue_locked; + +pop_from_list: + local_list = cc->write_thread_list; + local_list.next->prev = &local_list; + local_list.prev->next = &local_list; + INIT_LIST_HEAD(&cc->write_thread_list); + + spin_unlock_irq(&cc->write_thread_wait.lock); + + blk_start_plug(&plug); + do { + struct dm_crypt_io *io = container_of(local_list.next, + struct dm_crypt_io, list); + list_del(&io->list); + kcryptd_io_write(io); + } while (!list_empty(&local_list)); + blk_finish_plug(&plug); + } + return 0; } -static void kcryptd_crypt_write_io_submit(struct dm_crypt_io *io, int async) +static void kcryptd_crypt_write_io_submit(struct dm_crypt_io *io) { struct bio *clone = io->ctx.bio_out; struct crypt_config *cc = io->cc; + unsigned long flags; if (unlikely(io->error < 0)) { crypt_free_buffer_pages(cc, clone); @@ -1018,10 +1077,10 @@ static void kcryptd_crypt_write_io_submi clone->bi_sector = cc->start + io->sector; - if (async) - kcryptd_queue_io(io); - else - generic_make_request(clone); + spin_lock_irqsave(&cc->write_thread_wait.lock, flags); + list_add_tail(&io->list, &cc->write_thread_list); + wake_up_locked(&cc->write_thread_wait); + spin_unlock_irqrestore(&cc->write_thread_wait.lock, flags); } static void kcryptd_crypt_write_convert(struct dm_crypt_io *io) @@ -1068,7 +1127,7 @@ static void kcryptd_crypt_write_convert( /* Encryption was already finished, submit io now */ if (crypt_finished) { - kcryptd_crypt_write_io_submit(io, 0); + kcryptd_crypt_write_io_submit(io); /* * If there was an error, do not try next fragments. @@ -1171,7 +1230,7 @@ static void kcryptd_async_done(struct cr if (bio_data_dir(io->base_bio) == READ) kcryptd_crypt_read_done(io); else - kcryptd_crypt_write_io_submit(io, 1); + kcryptd_crypt_write_io_submit(io); } static void kcryptd_crypt(struct work_struct *work) @@ -1315,6 +1374,9 @@ static void crypt_dtr(struct dm_target * if (!cc) return; + if (cc->write_thread) + kthread_stop(cc->write_thread); + if (cc->io_queue) destroy_workqueue(cc->io_queue); if (cc->crypt_queue) @@ -1629,6 +1691,18 @@ static int crypt_ctr(struct dm_target *t goto bad; } + init_waitqueue_head(&cc->write_thread_wait); + INIT_LIST_HEAD(&cc->write_thread_list); + + cc->write_thread = kthread_create(dmcrypt_write, cc, "dmcrypt_write"); + if (IS_ERR(cc->write_thread)) { + ret = PTR_ERR(cc->write_thread); + cc->write_thread = NULL; + ti->error = "Couldn't spawn write thread"; + goto bad; + } + wake_up_process(cc->write_thread); + ti->num_flush_bios = 1; ti->discard_zeroes_data_unsupported = true; @@ -1660,7 +1734,7 @@ static int crypt_map(struct dm_target *t if (bio_data_dir(io->base_bio) == READ) { if (kcryptd_io_read(io, GFP_NOWAIT)) - kcryptd_queue_io(io); + kcryptd_queue_read(io); } else kcryptd_queue_crypt(io);