null_blk: multi queue aware block test driver
authorJens Axboe <axboe@kernel.dk>
Fri, 25 Oct 2013 10:52:25 +0000 (11:52 +0100)
committerJens Axboe <axboe@kernel.dk>
Fri, 25 Oct 2013 10:56:00 +0000 (11:56 +0100)
A driver that simply completes IO it receives, it does no
transfers. Written to fascilitate testing of the blk-mq code.
It supports various module options to use either bio queueing,
rq queueing, or mq mode.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
drivers/block/Kconfig
drivers/block/Makefile
drivers/block/null_blk.c [new file with mode: 0644]

index e07a5fd58ad7463f8775f7334fd692c7e51df2ad..4682546c5da730e42ce98ac7e36228bf08b77fdb 100644 (file)
@@ -15,6 +15,9 @@ menuconfig BLK_DEV
 
 if BLK_DEV
 
+config BLK_DEV_NULL_BLK
+       tristate "Null test block driver"
+
 config BLK_DEV_FD
        tristate "Normal floppy disk support"
        depends on ARCH_MAY_HAVE_PC_FDC
index ca07399a8d99efb2d11b80bc266c52c496c902e3..03b3b4a2bd8ae354d7647d3ef4552202d8ea97f0 100644 (file)
@@ -41,6 +41,7 @@ obj-$(CONFIG_BLK_DEV_RBD)     += rbd.o
 obj-$(CONFIG_BLK_DEV_PCIESSD_MTIP32XX) += mtip32xx/
 
 obj-$(CONFIG_BLK_DEV_RSXX) += rsxx/
+obj-$(CONFIG_BLK_DEV_NULL_BLK) += null_blk.o
 
 nvme-y         := nvme-core.o nvme-scsi.o
 swim_mod-y     := swim.o swim_asm.o
diff --git a/drivers/block/null_blk.c b/drivers/block/null_blk.c
new file mode 100644 (file)
index 0000000..b5d8423
--- /dev/null
@@ -0,0 +1,635 @@
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/sched.h>
+#include <linux/fs.h>
+#include <linux/blkdev.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/blk-mq.h>
+#include <linux/hrtimer.h>
+
+struct nullb_cmd {
+       struct list_head list;
+       struct llist_node ll_list;
+       struct call_single_data csd;
+       struct request *rq;
+       struct bio *bio;
+       unsigned int tag;
+       struct nullb_queue *nq;
+};
+
+struct nullb_queue {
+       unsigned long *tag_map;
+       wait_queue_head_t wait;
+       unsigned int queue_depth;
+
+       struct nullb_cmd *cmds;
+};
+
+struct nullb {
+       struct list_head list;
+       unsigned int index;
+       struct request_queue *q;
+       struct gendisk *disk;
+       struct hrtimer timer;
+       unsigned int queue_depth;
+       spinlock_t lock;
+
+       struct nullb_queue *queues;
+       unsigned int nr_queues;
+};
+
+static LIST_HEAD(nullb_list);
+static struct mutex lock;
+static int null_major;
+static int nullb_indexes;
+
+struct completion_queue {
+       struct llist_head list;
+       struct hrtimer timer;
+};
+
+/*
+ * These are per-cpu for now, they will need to be configured by the
+ * complete_queues parameter and appropriately mapped.
+ */
+static DEFINE_PER_CPU(struct completion_queue, completion_queues);
+
+enum {
+       NULL_IRQ_NONE           = 0,
+       NULL_IRQ_SOFTIRQ        = 1,
+       NULL_IRQ_TIMER          = 2,
+
+       NULL_Q_BIO              = 0,
+       NULL_Q_RQ               = 1,
+       NULL_Q_MQ               = 2,
+};
+
+static int submit_queues = 1;
+module_param(submit_queues, int, S_IRUGO);
+MODULE_PARM_DESC(submit_queues, "Number of submission queues");
+
+static int home_node = NUMA_NO_NODE;
+module_param(home_node, int, S_IRUGO);
+MODULE_PARM_DESC(home_node, "Home node for the device");
+
+static int queue_mode = NULL_Q_MQ;
+module_param(queue_mode, int, S_IRUGO);
+MODULE_PARM_DESC(use_mq, "Use blk-mq interface (0=bio,1=rq,2=multiqueue)");
+
+static int gb = 250;
+module_param(gb, int, S_IRUGO);
+MODULE_PARM_DESC(gb, "Size in GB");
+
+static int bs = 512;
+module_param(bs, int, S_IRUGO);
+MODULE_PARM_DESC(bs, "Block size (in bytes)");
+
+static int nr_devices = 2;
+module_param(nr_devices, int, S_IRUGO);
+MODULE_PARM_DESC(nr_devices, "Number of devices to register");
+
+static int irqmode = NULL_IRQ_SOFTIRQ;
+module_param(irqmode, int, S_IRUGO);
+MODULE_PARM_DESC(irqmode, "IRQ completion handler. 0-none, 1-softirq, 2-timer");
+
+static int completion_nsec = 10000;
+module_param(completion_nsec, int, S_IRUGO);
+MODULE_PARM_DESC(completion_nsec, "Time in ns to complete a request in hardware. Default: 10,000ns");
+
+static int hw_queue_depth = 64;
+module_param(hw_queue_depth, int, S_IRUGO);
+MODULE_PARM_DESC(hw_queue_depth, "Queue depth for each hardware queue. Default: 64");
+
+static bool use_per_node_hctx = true;
+module_param(use_per_node_hctx, bool, S_IRUGO);
+MODULE_PARM_DESC(use_per_node_hctx, "Use per-node allocation for hardware context queues. Default: true");
+
+static void put_tag(struct nullb_queue *nq, unsigned int tag)
+{
+       clear_bit_unlock(tag, nq->tag_map);
+
+       if (waitqueue_active(&nq->wait))
+               wake_up(&nq->wait);
+}
+
+static unsigned int get_tag(struct nullb_queue *nq)
+{
+       unsigned int tag;
+
+       do {
+               tag = find_first_zero_bit(nq->tag_map, nq->queue_depth);
+               if (tag >= nq->queue_depth)
+                       return -1U;
+       } while (test_and_set_bit_lock(tag, nq->tag_map));
+
+       return tag;
+}
+
+static void free_cmd(struct nullb_cmd *cmd)
+{
+       put_tag(cmd->nq, cmd->tag);
+}
+
+static struct nullb_cmd *__alloc_cmd(struct nullb_queue *nq)
+{
+       struct nullb_cmd *cmd;
+       unsigned int tag;
+
+       tag = get_tag(nq);
+       if (tag != -1U) {
+               cmd = &nq->cmds[tag];
+               cmd->tag = tag;
+               cmd->nq = nq;
+               return cmd;
+       }
+
+       return NULL;
+}
+
+static struct nullb_cmd *alloc_cmd(struct nullb_queue *nq, int can_wait)
+{
+       struct nullb_cmd *cmd;
+       DEFINE_WAIT(wait);
+
+       cmd = __alloc_cmd(nq);
+       if (cmd || !can_wait)
+               return cmd;
+
+       do {
+               prepare_to_wait(&nq->wait, &wait, TASK_UNINTERRUPTIBLE);
+               cmd = __alloc_cmd(nq);
+               if (cmd)
+                       break;
+
+               io_schedule();
+       } while (1);
+
+       finish_wait(&nq->wait, &wait);
+       return cmd;
+}
+
+static void end_cmd(struct nullb_cmd *cmd)
+{
+       if (cmd->rq) {
+               if (queue_mode == NULL_Q_MQ)
+                       blk_mq_end_io(cmd->rq, 0);
+               else {
+                       INIT_LIST_HEAD(&cmd->rq->queuelist);
+                       blk_end_request_all(cmd->rq, 0);
+               }
+       } else if (cmd->bio)
+               bio_endio(cmd->bio, 0);
+
+       if (queue_mode != NULL_Q_MQ)
+               free_cmd(cmd);
+}
+
+static enum hrtimer_restart null_cmd_timer_expired(struct hrtimer *timer)
+{
+       struct completion_queue *cq;
+       struct llist_node *entry;
+       struct nullb_cmd *cmd;
+
+       cq = &per_cpu(completion_queues, smp_processor_id());
+
+       while ((entry = llist_del_all(&cq->list)) != NULL) {
+               do {
+                       cmd = container_of(entry, struct nullb_cmd, ll_list);
+                       end_cmd(cmd);
+                       entry = entry->next;
+               } while (entry);
+       }
+
+       return HRTIMER_NORESTART;
+}
+
+static void null_cmd_end_timer(struct nullb_cmd *cmd)
+{
+       struct completion_queue *cq = &per_cpu(completion_queues, get_cpu());
+
+       cmd->ll_list.next = NULL;
+       if (llist_add(&cmd->ll_list, &cq->list)) {
+               ktime_t kt = ktime_set(0, completion_nsec);
+
+               hrtimer_start(&cq->timer, kt, HRTIMER_MODE_REL);
+       }
+
+       put_cpu();
+}
+
+static void null_softirq_done_fn(struct request *rq)
+{
+       blk_end_request_all(rq, 0);
+}
+
+#if defined(CONFIG_SMP) && defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+
+static void null_ipi_cmd_end_io(void *data)
+{
+       struct completion_queue *cq;
+       struct llist_node *entry, *next;
+       struct nullb_cmd *cmd;
+
+       cq = &per_cpu(completion_queues, smp_processor_id());
+
+       entry = llist_del_all(&cq->list);
+
+       while (entry) {
+               next = entry->next;
+               cmd = llist_entry(entry, struct nullb_cmd, ll_list);
+               end_cmd(cmd);
+               entry = next;
+       }
+}
+
+static void null_cmd_end_ipi(struct nullb_cmd *cmd)
+{
+       struct call_single_data *data = &cmd->csd;
+       int cpu = get_cpu();
+       struct completion_queue *cq = &per_cpu(completion_queues, cpu);
+
+       cmd->ll_list.next = NULL;
+
+       if (llist_add(&cmd->ll_list, &cq->list)) {
+               data->func = null_ipi_cmd_end_io;
+               data->flags = 0;
+               __smp_call_function_single(cpu, data, 0);
+       }
+
+       put_cpu();
+}
+
+#endif /* CONFIG_SMP && CONFIG_USE_GENERIC_SMP_HELPERS */
+
+static inline void null_handle_cmd(struct nullb_cmd *cmd)
+{
+       /* Complete IO by inline, softirq or timer */
+       switch (irqmode) {
+       case NULL_IRQ_NONE:
+               end_cmd(cmd);
+               break;
+       case NULL_IRQ_SOFTIRQ:
+#if defined(CONFIG_SMP) && defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+               null_cmd_end_ipi(cmd);
+#else
+               end_cmd(cmd);
+#endif
+               break;
+       case NULL_IRQ_TIMER:
+               null_cmd_end_timer(cmd);
+               break;
+       }
+}
+
+static struct nullb_queue *nullb_to_queue(struct nullb *nullb)
+{
+       int index = 0;
+
+       if (nullb->nr_queues != 1)
+               index = raw_smp_processor_id() / ((nr_cpu_ids + nullb->nr_queues - 1) / nullb->nr_queues);
+
+       return &nullb->queues[index];
+}
+
+static void null_queue_bio(struct request_queue *q, struct bio *bio)
+{
+       struct nullb *nullb = q->queuedata;
+       struct nullb_queue *nq = nullb_to_queue(nullb);
+       struct nullb_cmd *cmd;
+
+       cmd = alloc_cmd(nq, 1);
+       cmd->bio = bio;
+
+       null_handle_cmd(cmd);
+}
+
+static int null_rq_prep_fn(struct request_queue *q, struct request *req)
+{
+       struct nullb *nullb = q->queuedata;
+       struct nullb_queue *nq = nullb_to_queue(nullb);
+       struct nullb_cmd *cmd;
+
+       cmd = alloc_cmd(nq, 0);
+       if (cmd) {
+               cmd->rq = req;
+               req->special = cmd;
+               return BLKPREP_OK;
+       }
+
+       return BLKPREP_DEFER;
+}
+
+static void null_request_fn(struct request_queue *q)
+{
+       struct request *rq;
+
+       while ((rq = blk_fetch_request(q)) != NULL) {
+               struct nullb_cmd *cmd = rq->special;
+
+               spin_unlock_irq(q->queue_lock);
+               null_handle_cmd(cmd);
+               spin_lock_irq(q->queue_lock);
+       }
+}
+
+static int null_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *rq)
+{
+       struct nullb_cmd *cmd = rq->special;
+
+       cmd->rq = rq;
+       cmd->nq = hctx->driver_data;
+
+       null_handle_cmd(cmd);
+       return BLK_MQ_RQ_QUEUE_OK;
+}
+
+static struct blk_mq_hw_ctx *null_alloc_hctx(struct blk_mq_reg *reg, unsigned int hctx_index)
+{
+       return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL,
+                               hctx_index);
+}
+
+static void null_free_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_index)
+{
+       kfree(hctx);
+}
+
+static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
+                         unsigned int index)
+{
+       struct nullb *nullb = data;
+       struct nullb_queue *nq = &nullb->queues[index];
+
+       init_waitqueue_head(&nq->wait);
+       nq->queue_depth = nullb->queue_depth;
+       nullb->nr_queues++;
+       hctx->driver_data = nq;
+
+       return 0;
+}
+
+static struct blk_mq_ops null_mq_ops = {
+       .queue_rq       = null_queue_rq,
+       .map_queue      = blk_mq_map_queue,
+       .init_hctx      = null_init_hctx,
+};
+
+static struct blk_mq_reg null_mq_reg = {
+       .ops            = &null_mq_ops,
+       .queue_depth    = 64,
+       .cmd_size       = sizeof(struct nullb_cmd),
+       .flags          = BLK_MQ_F_SHOULD_MERGE,
+};
+
+static void null_del_dev(struct nullb *nullb)
+{
+       list_del_init(&nullb->list);
+
+       del_gendisk(nullb->disk);
+       if (queue_mode == NULL_Q_MQ)
+               blk_mq_free_queue(nullb->q);
+       else
+               blk_cleanup_queue(nullb->q);
+       put_disk(nullb->disk);
+       kfree(nullb);
+}
+
+static int null_open(struct block_device *bdev, fmode_t mode)
+{
+       return 0;
+}
+
+static void null_release(struct gendisk *disk, fmode_t mode)
+{
+}
+
+static const struct block_device_operations null_fops = {
+       .owner =        THIS_MODULE,
+       .open =         null_open,
+       .release =      null_release,
+};
+
+static int setup_commands(struct nullb_queue *nq)
+{
+       struct nullb_cmd *cmd;
+       int i, tag_size;
+
+       nq->cmds = kzalloc(nq->queue_depth * sizeof(*cmd), GFP_KERNEL);
+       if (!nq->cmds)
+               return 1;
+
+       tag_size = ALIGN(nq->queue_depth, BITS_PER_LONG) / BITS_PER_LONG;
+       nq->tag_map = kzalloc(tag_size * sizeof(unsigned long), GFP_KERNEL);
+       if (!nq->tag_map) {
+               kfree(nq->cmds);
+               return 1;
+       }
+
+       for (i = 0; i < nq->queue_depth; i++) {
+               cmd = &nq->cmds[i];
+               INIT_LIST_HEAD(&cmd->list);
+               cmd->ll_list.next = NULL;
+               cmd->tag = -1U;
+       }
+
+       return 0;
+}
+
+static void cleanup_queue(struct nullb_queue *nq)
+{
+       kfree(nq->tag_map);
+       kfree(nq->cmds);
+}
+
+static void cleanup_queues(struct nullb *nullb)
+{
+       int i;
+
+       for (i = 0; i < nullb->nr_queues; i++)
+               cleanup_queue(&nullb->queues[i]);
+
+       kfree(nullb->queues);
+}
+
+static int setup_queues(struct nullb *nullb)
+{
+       struct nullb_queue *nq;
+       int i;
+
+       nullb->queues = kzalloc(submit_queues * sizeof(*nq), GFP_KERNEL);
+       if (!nullb->queues)
+               return 1;
+
+       nullb->nr_queues = 0;
+       nullb->queue_depth = hw_queue_depth;
+
+       if (queue_mode == NULL_Q_MQ)
+               return 0;
+
+       for (i = 0; i < submit_queues; i++) {
+               nq = &nullb->queues[i];
+               init_waitqueue_head(&nq->wait);
+               nq->queue_depth = hw_queue_depth;
+               if (setup_commands(nq))
+                       break;
+               nullb->nr_queues++;
+       }
+
+       if (i == submit_queues)
+               return 0;
+
+       cleanup_queues(nullb);
+       return 1;
+}
+
+static int null_add_dev(void)
+{
+       struct gendisk *disk;
+       struct nullb *nullb;
+       sector_t size;
+
+       nullb = kzalloc_node(sizeof(*nullb), GFP_KERNEL, home_node);
+       if (!nullb)
+               return -ENOMEM;
+
+       spin_lock_init(&nullb->lock);
+
+       if (setup_queues(nullb))
+               goto err;
+
+       if (queue_mode == NULL_Q_MQ) {
+               null_mq_reg.numa_node = home_node;
+               null_mq_reg.queue_depth = hw_queue_depth;
+
+               if (use_per_node_hctx) {
+                       null_mq_reg.ops->alloc_hctx = null_alloc_hctx;
+                       null_mq_reg.ops->free_hctx = null_free_hctx;
+
+                       null_mq_reg.nr_hw_queues = nr_online_nodes;
+               } else {
+                       null_mq_reg.ops->alloc_hctx = blk_mq_alloc_single_hw_queue;
+                       null_mq_reg.ops->free_hctx = blk_mq_free_single_hw_queue;
+
+                       null_mq_reg.nr_hw_queues = submit_queues;
+               }
+
+               nullb->q = blk_mq_init_queue(&null_mq_reg, nullb);
+       } else if (queue_mode == NULL_Q_BIO) {
+               nullb->q = blk_alloc_queue_node(GFP_KERNEL, home_node);
+               blk_queue_make_request(nullb->q, null_queue_bio);
+       } else {
+               nullb->q = blk_init_queue_node(null_request_fn, &nullb->lock, home_node);
+               blk_queue_prep_rq(nullb->q, null_rq_prep_fn);
+               if (nullb->q)
+                       blk_queue_softirq_done(nullb->q, null_softirq_done_fn);
+       }
+
+       if (!nullb->q)
+               goto queue_fail;
+
+       nullb->q->queuedata = nullb;
+       queue_flag_set_unlocked(QUEUE_FLAG_NONROT, nullb->q);
+
+       disk = nullb->disk = alloc_disk_node(1, home_node);
+       if (!disk) {
+queue_fail:
+               if (queue_mode == NULL_Q_MQ)
+                       blk_mq_free_queue(nullb->q);
+               else
+                       blk_cleanup_queue(nullb->q);
+               cleanup_queues(nullb);
+err:
+               kfree(nullb);
+               return -ENOMEM;
+       }
+
+       mutex_lock(&lock);
+       list_add_tail(&nullb->list, &nullb_list);
+       nullb->index = nullb_indexes++;
+       mutex_unlock(&lock);
+
+       blk_queue_logical_block_size(nullb->q, bs);
+       blk_queue_physical_block_size(nullb->q, bs);
+
+       size = gb * 1024 * 1024 * 1024ULL;
+       sector_div(size, bs);
+       set_capacity(disk, size);
+
+       disk->flags |= GENHD_FL_EXT_DEVT;
+       disk->major             = null_major;
+       disk->first_minor       = nullb->index;
+       disk->fops              = &null_fops;
+       disk->private_data      = nullb;
+       disk->queue             = nullb->q;
+       sprintf(disk->disk_name, "nullb%d", nullb->index);
+       add_disk(disk);
+       return 0;
+}
+
+static int __init null_init(void)
+{
+       unsigned int i;
+
+#if !defined(CONFIG_SMP) || !defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+       if (irqmode == NULL_IRQ_SOFTIRQ) {
+               pr_warn("null_blk: softirq completions not available.\n");
+               pr_warn("null_blk: using direct completions.\n");
+               irqmode = NULL_IRQ_NONE;
+       }
+#endif
+
+       if (submit_queues > nr_cpu_ids)
+               submit_queues = nr_cpu_ids;
+       else if (!submit_queues)
+               submit_queues = 1;
+
+       mutex_init(&lock);
+
+       /* Initialize a separate list for each CPU for issuing softirqs */
+       for_each_possible_cpu(i) {
+               struct completion_queue *cq = &per_cpu(completion_queues, i);
+
+               init_llist_head(&cq->list);
+
+               if (irqmode != NULL_IRQ_TIMER)
+                       continue;
+
+               hrtimer_init(&cq->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+               cq->timer.function = null_cmd_timer_expired;
+       }
+
+       null_major = register_blkdev(0, "nullb");
+       if (null_major < 0)
+               return null_major;
+
+       for (i = 0; i < nr_devices; i++) {
+               if (null_add_dev()) {
+                       unregister_blkdev(null_major, "nullb");
+                       return -EINVAL;
+               }
+       }
+
+       pr_info("null: module loaded\n");
+       return 0;
+}
+
+static void __exit null_exit(void)
+{
+       struct nullb *nullb;
+
+       unregister_blkdev(null_major, "nullb");
+
+       mutex_lock(&lock);
+       while (!list_empty(&nullb_list)) {
+               nullb = list_entry(nullb_list.next, struct nullb, list);
+               null_del_dev(nullb);
+       }
+       mutex_unlock(&lock);
+}
+
+module_init(null_init);
+module_exit(null_exit);
+
+MODULE_AUTHOR("Jens Axboe <jaxboe@fusionio.com>");
+MODULE_LICENSE("GPL");