From ff856bad67cb65cb4dc4ef88b808804fc4265782 Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Mon, 9 Jan 2006 16:02:34 +0100 Subject: [BLOCK] ll_rw_blk: Enable out-of-order request completions through softirq Request completion can be a quite heavy process, since it needs to iterate through the entire request and complete the bio's it holds. This patch adds blk_complete_request() which moves this processing into a dedicated block softirq. Signed-off-by: Jens Axboe --- include/linux/blkdev.h | 21 ++++++++++++++++++--- include/linux/interrupt.h | 1 + 2 files changed, 19 insertions(+), 3 deletions(-) (limited to 'include') diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h index fb098537742..804cc4ec953 100644 --- a/include/linux/blkdev.h +++ b/include/linux/blkdev.h @@ -118,9 +118,9 @@ struct request_list { * try to put the fields that are referenced together in the same cacheline */ struct request { - struct list_head queuelist; /* looking for ->queue? you must _not_ - * access it directly, use - * blkdev_dequeue_request! */ + struct list_head queuelist; + struct list_head donelist; + unsigned long flags; /* see REQ_ bits below */ /* Maintain bio traversal state for part by part I/O submission. @@ -141,6 +141,7 @@ struct request { struct bio *biotail; void *elevator_private; + void *completion_data; unsigned short ioprio; @@ -291,6 +292,7 @@ typedef int (merge_bvec_fn) (request_queue_t *, struct bio *, struct bio_vec *); typedef void (activity_fn) (void *data, int rw); typedef int (issue_flush_fn) (request_queue_t *, struct gendisk *, sector_t *); typedef void (prepare_flush_fn) (request_queue_t *, struct request *); +typedef void (softirq_done_fn)(struct request *); enum blk_queue_state { Queue_down, @@ -332,6 +334,7 @@ struct request_queue activity_fn *activity_fn; issue_flush_fn *issue_flush_fn; prepare_flush_fn *prepare_flush_fn; + softirq_done_fn *softirq_done_fn; /* * Dispatch queue sorting @@ -646,6 +649,17 @@ extern int end_that_request_first(struct request *, int, int); extern int end_that_request_chunk(struct request *, int, int); extern void end_that_request_last(struct request *, int); extern void end_request(struct request *req, int uptodate); +extern void blk_complete_request(struct request *); + +static inline int rq_all_done(struct request *rq, unsigned int nr_bytes) +{ + if (blk_fs_request(rq)) + return (nr_bytes >= (rq->hard_nr_sectors << 9)); + else if (blk_pc_request(rq)) + return nr_bytes >= rq->data_len; + + return 0; +} /* * end_that_request_first/chunk() takes an uptodate argument. we account @@ -694,6 +708,7 @@ extern void blk_queue_segment_boundary(request_queue_t *, unsigned long); extern void blk_queue_prep_rq(request_queue_t *, prep_rq_fn *pfn); extern void blk_queue_merge_bvec(request_queue_t *, merge_bvec_fn *); extern void blk_queue_dma_alignment(request_queue_t *, int); +extern void blk_queue_softirq_done(request_queue_t *, softirq_done_fn *); extern struct backing_dev_info *blk_get_backing_dev_info(struct block_device *bdev); extern int blk_queue_ordered(request_queue_t *, unsigned, prepare_flush_fn *); extern void blk_queue_issue_flush_fn(request_queue_t *, issue_flush_fn *); diff --git a/include/linux/interrupt.h b/include/linux/interrupt.h index e50a95fbeb1..f0220470698 100644 --- a/include/linux/interrupt.h +++ b/include/linux/interrupt.h @@ -112,6 +112,7 @@ enum TIMER_SOFTIRQ, NET_TX_SOFTIRQ, NET_RX_SOFTIRQ, + BLOCK_SOFTIRQ, SCSI_SOFTIRQ, TASKLET_SOFTIRQ }; -- cgit v1.2.3-70-g09d2 From 1aea6434eebd25e532d2e5ddabf2733af4e1ff0b Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Mon, 9 Jan 2006 16:03:03 +0100 Subject: [SCSI] Kill the SCSI softirq handling This patch moves the SCSI softirq handling to the block layer version. There should be no functional changes. Signed-off-by: Jens Axboe --- drivers/scsi/scsi.c | 109 ++++------------------------------------------ drivers/scsi/scsi_lib.c | 36 +++++++++++++++ drivers/scsi/scsi_priv.h | 1 + include/linux/interrupt.h | 1 - 4 files changed, 45 insertions(+), 102 deletions(-) (limited to 'include') diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c index 180676d7115..ee5f4dfdab1 100644 --- a/drivers/scsi/scsi.c +++ b/drivers/scsi/scsi.c @@ -69,7 +69,6 @@ #include "scsi_logging.h" static void scsi_done(struct scsi_cmnd *cmd); -static int scsi_retry_command(struct scsi_cmnd *cmd); /* * Definitions and constants. @@ -752,7 +751,7 @@ static void scsi_done(struct scsi_cmnd *cmd) * isn't running --- used by scsi_times_out */ void __scsi_done(struct scsi_cmnd *cmd) { - unsigned long flags; + struct request *rq = cmd->request; /* * Set the serial numbers back to zero @@ -763,71 +762,14 @@ void __scsi_done(struct scsi_cmnd *cmd) if (cmd->result) atomic_inc(&cmd->device->ioerr_cnt); + BUG_ON(!rq); + /* - * Next, enqueue the command into the done queue. - * It is a per-CPU queue, so we just disable local interrupts - * and need no spinlock. + * The uptodate/nbytes values don't matter, as we allow partial + * completes and thus will check this in the softirq callback */ - local_irq_save(flags); - list_add_tail(&cmd->eh_entry, &__get_cpu_var(scsi_done_q)); - raise_softirq_irqoff(SCSI_SOFTIRQ); - local_irq_restore(flags); -} - -/** - * scsi_softirq - Perform post-interrupt processing of finished SCSI commands. - * - * This is the consumer of the done queue. - * - * This is called with all interrupts enabled. This should reduce - * interrupt latency, stack depth, and reentrancy of the low-level - * drivers. - */ -static void scsi_softirq(struct softirq_action *h) -{ - int disposition; - LIST_HEAD(local_q); - - local_irq_disable(); - list_splice_init(&__get_cpu_var(scsi_done_q), &local_q); - local_irq_enable(); - - while (!list_empty(&local_q)) { - struct scsi_cmnd *cmd = list_entry(local_q.next, - struct scsi_cmnd, eh_entry); - /* The longest time any command should be outstanding is the - * per command timeout multiplied by the number of retries. - * - * For a typical command, this is 2.5 minutes */ - unsigned long wait_for - = cmd->allowed * cmd->timeout_per_command; - list_del_init(&cmd->eh_entry); - - disposition = scsi_decide_disposition(cmd); - if (disposition != SUCCESS && - time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) { - sdev_printk(KERN_ERR, cmd->device, - "timing out command, waited %lus\n", - wait_for/HZ); - disposition = SUCCESS; - } - - scsi_log_completion(cmd, disposition); - switch (disposition) { - case SUCCESS: - scsi_finish_command(cmd); - break; - case NEEDS_RETRY: - scsi_retry_command(cmd); - break; - case ADD_TO_MLQUEUE: - scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY); - break; - default: - if (!scsi_eh_scmd_add(cmd, 0)) - scsi_finish_command(cmd); - } - } + rq->completion_data = cmd; + blk_complete_request(rq); } /* @@ -840,7 +782,7 @@ static void scsi_softirq(struct softirq_action *h) * level drivers should not become re-entrant as a result of * this. */ -static int scsi_retry_command(struct scsi_cmnd *cmd) +int scsi_retry_command(struct scsi_cmnd *cmd) { /* * Restore the SCSI command state. @@ -1273,38 +1215,6 @@ int scsi_device_cancel(struct scsi_device *sdev, int recovery) } EXPORT_SYMBOL(scsi_device_cancel); -#ifdef CONFIG_HOTPLUG_CPU -static int scsi_cpu_notify(struct notifier_block *self, - unsigned long action, void *hcpu) -{ - int cpu = (unsigned long)hcpu; - - switch(action) { - case CPU_DEAD: - /* Drain scsi_done_q. */ - local_irq_disable(); - list_splice_init(&per_cpu(scsi_done_q, cpu), - &__get_cpu_var(scsi_done_q)); - raise_softirq_irqoff(SCSI_SOFTIRQ); - local_irq_enable(); - break; - default: - break; - } - return NOTIFY_OK; -} - -static struct notifier_block __devinitdata scsi_cpu_nb = { - .notifier_call = scsi_cpu_notify, -}; - -#define register_scsi_cpu() register_cpu_notifier(&scsi_cpu_nb) -#define unregister_scsi_cpu() unregister_cpu_notifier(&scsi_cpu_nb) -#else -#define register_scsi_cpu() -#define unregister_scsi_cpu() -#endif /* CONFIG_HOTPLUG_CPU */ - MODULE_DESCRIPTION("SCSI core"); MODULE_LICENSE("GPL"); @@ -1338,8 +1248,6 @@ static int __init init_scsi(void) INIT_LIST_HEAD(&per_cpu(scsi_done_q, i)); devfs_mk_dir("scsi"); - open_softirq(SCSI_SOFTIRQ, scsi_softirq, NULL); - register_scsi_cpu(); printk(KERN_NOTICE "SCSI subsystem initialized\n"); return 0; @@ -1367,7 +1275,6 @@ static void __exit exit_scsi(void) devfs_remove("scsi"); scsi_exit_procfs(); scsi_exit_queue(); - unregister_scsi_cpu(); } subsys_initcall(init_scsi); diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c index ba93d6e66d4..00c9bf383e2 100644 --- a/drivers/scsi/scsi_lib.c +++ b/drivers/scsi/scsi_lib.c @@ -1493,6 +1493,41 @@ static void scsi_kill_request(struct request *req, request_queue_t *q) __scsi_done(cmd); } +static void scsi_softirq_done(struct request *rq) +{ + struct scsi_cmnd *cmd = rq->completion_data; + unsigned long wait_for = cmd->allowed * cmd->timeout_per_command; + int disposition; + + INIT_LIST_HEAD(&cmd->eh_entry); + + disposition = scsi_decide_disposition(cmd); + if (disposition != SUCCESS && + time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) { + sdev_printk(KERN_ERR, cmd->device, + "timing out command, waited %lus\n", + wait_for/HZ); + disposition = SUCCESS; + } + + scsi_log_completion(cmd, disposition); + + switch (disposition) { + case SUCCESS: + scsi_finish_command(cmd); + break; + case NEEDS_RETRY: + scsi_retry_command(cmd); + break; + case ADD_TO_MLQUEUE: + scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY); + break; + default: + if (!scsi_eh_scmd_add(cmd, 0)) + scsi_finish_command(cmd); + } +} + /* * Function: scsi_request_fn() * @@ -1667,6 +1702,7 @@ struct request_queue *scsi_alloc_queue(struct scsi_device *sdev) blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost)); blk_queue_segment_boundary(q, shost->dma_boundary); blk_queue_issue_flush_fn(q, scsi_issue_flush_fn); + blk_queue_softirq_done(q, scsi_softirq_done); if (!shost->use_clustering) clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags); diff --git a/drivers/scsi/scsi_priv.h b/drivers/scsi/scsi_priv.h index f04e7e11f57..14a6198cb8d 100644 --- a/drivers/scsi/scsi_priv.h +++ b/drivers/scsi/scsi_priv.h @@ -44,6 +44,7 @@ extern void scsi_init_cmd_from_req(struct scsi_cmnd *cmd, struct scsi_request *sreq); extern void __scsi_release_request(struct scsi_request *sreq); extern void __scsi_done(struct scsi_cmnd *cmd); +extern int scsi_retry_command(struct scsi_cmnd *cmd); #ifdef CONFIG_SCSI_LOGGING void scsi_log_send(struct scsi_cmnd *cmd); void scsi_log_completion(struct scsi_cmnd *cmd, int disposition); diff --git a/include/linux/interrupt.h b/include/linux/interrupt.h index f0220470698..2c08fdc2bdf 100644 --- a/include/linux/interrupt.h +++ b/include/linux/interrupt.h @@ -113,7 +113,6 @@ enum NET_TX_SOFTIRQ, NET_RX_SOFTIRQ, BLOCK_SOFTIRQ, - SCSI_SOFTIRQ, TASKLET_SOFTIRQ }; -- cgit v1.2.3-70-g09d2 From 8672d57138b34447719cd7749f3d21070e1175a1 Mon Sep 17 00:00:00 2001 From: Jens Axboe Date: Mon, 9 Jan 2006 16:03:35 +0100 Subject: [IDE] Use the block layer deferred softirq request completion This patch makes IDE use the new blk_complete_request() interface. There's still room for improvement, as __ide_end_request() really could drop the lock after getting HWGROUP->rq (why does it need to hold it in the first place? If ->rq access isn't serialized, we are screwed anyways). Signed-off-by: Jens Axboe --- drivers/ide/ide-io.c | 42 +++++++++++++++++++++++++++++++++++------- drivers/ide/ide-probe.c | 2 ++ include/linux/ide.h | 1 + 3 files changed, 38 insertions(+), 7 deletions(-) (limited to 'include') diff --git a/drivers/ide/ide-io.c b/drivers/ide/ide-io.c index b5dc6df8e67..dea2d4dcc69 100644 --- a/drivers/ide/ide-io.c +++ b/drivers/ide/ide-io.c @@ -55,9 +55,22 @@ #include #include +void ide_softirq_done(struct request *rq) +{ + request_queue_t *q = rq->q; + + add_disk_randomness(rq->rq_disk); + end_that_request_chunk(rq, rq->errors, rq->data_len); + + spin_lock_irq(q->queue_lock); + end_that_request_last(rq, rq->errors); + spin_unlock_irq(q->queue_lock); +} + int __ide_end_request(ide_drive_t *drive, struct request *rq, int uptodate, int nr_sectors) { + unsigned int nbytes; int ret = 1; BUG_ON(!(rq->flags & REQ_STARTED)); @@ -81,17 +94,28 @@ int __ide_end_request(ide_drive_t *drive, struct request *rq, int uptodate, HWGROUP(drive)->hwif->ide_dma_on(drive); } - if (!end_that_request_first(rq, uptodate, nr_sectors)) { - add_disk_randomness(rq->rq_disk); - - if (blk_rq_tagged(rq)) - blk_queue_end_tag(drive->queue, rq); - + /* + * For partial completions (or non fs/pc requests), use the regular + * direct completion path. + */ + nbytes = nr_sectors << 9; + if (rq_all_done(rq, nbytes)) { + rq->errors = uptodate; + rq->data_len = nbytes; blkdev_dequeue_request(rq); HWGROUP(drive)->rq = NULL; - end_that_request_last(rq, uptodate); + blk_complete_request(rq); ret = 0; + } else { + if (!end_that_request_first(rq, uptodate, nr_sectors)) { + add_disk_randomness(rq->rq_disk); + blkdev_dequeue_request(rq); + HWGROUP(drive)->rq = NULL; + end_that_request_last(rq, uptodate); + ret = 0; + } } + return ret; } EXPORT_SYMBOL(__ide_end_request); @@ -113,6 +137,10 @@ int ide_end_request (ide_drive_t *drive, int uptodate, int nr_sectors) unsigned long flags; int ret = 1; + /* + * room for locking improvements here, the calls below don't + * need the queue lock held at all + */ spin_lock_irqsave(&ide_lock, flags); rq = HWGROUP(drive)->rq; diff --git a/drivers/ide/ide-probe.c b/drivers/ide/ide-probe.c index 02167a5b751..1ddaa71a8f4 100644 --- a/drivers/ide/ide-probe.c +++ b/drivers/ide/ide-probe.c @@ -1011,6 +1011,8 @@ static int ide_init_queue(ide_drive_t *drive) blk_queue_max_hw_segments(q, max_sg_entries); blk_queue_max_phys_segments(q, max_sg_entries); + blk_queue_softirq_done(q, ide_softirq_done); + /* assign drive queue */ drive->queue = q; diff --git a/include/linux/ide.h b/include/linux/ide.h index 4dd6694963c..ef8d0cbb832 100644 --- a/include/linux/ide.h +++ b/include/linux/ide.h @@ -1001,6 +1001,7 @@ extern int noautodma; extern int ide_end_request (ide_drive_t *drive, int uptodate, int nrsecs); extern int __ide_end_request (ide_drive_t *drive, struct request *rq, int uptodate, int nrsecs); +extern void ide_softirq_done(struct request *rq); /* * This is used on exit from the driver to designate the next irq handler -- cgit v1.2.3-70-g09d2