nvme: add separate poll queue map
authorJens Axboe <axboe@kernel.dk>
Mon, 5 Nov 2018 19:44:33 +0000 (12:44 -0700)
committerJens Axboe <axboe@kernel.dk>
Wed, 7 Nov 2018 20:45:00 +0000 (13:45 -0700)
Adds support for defining a variable number of poll queues, currently
configurable with the 'poll_queues' module parameter. Defaults to
a single poll queue.

And now we finally have poll support without triggering interrupts!

Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
drivers/nvme/host/pci.c
include/linux/blk-mq.h

index 1987df13b73e2afe6af621147e2828e435109a87..6aa86dfcb32c126376ee369d68b8e67412e4949a 100644 (file)
@@ -86,6 +86,10 @@ MODULE_PARM_DESC(write_queues,
        "Number of queues to use for writes. If not set, reads and writes "
        "will share a queue set.");
 
+static int poll_queues = 1;
+module_param_cb(poll_queues, &queue_count_ops, &poll_queues, 0644);
+MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
+
 struct nvme_dev;
 struct nvme_queue;
 
@@ -94,6 +98,7 @@ static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
 enum {
        NVMEQ_TYPE_READ,
        NVMEQ_TYPE_WRITE,
+       NVMEQ_TYPE_POLL,
        NVMEQ_TYPE_NR,
 };
 
@@ -202,6 +207,7 @@ struct nvme_queue {
        u16 last_cq_head;
        u16 qid;
        u8 cq_phase;
+       u8 polled;
        u32 *dbbuf_sq_db;
        u32 *dbbuf_cq_db;
        u32 *dbbuf_sq_ei;
@@ -250,7 +256,7 @@ static inline void _nvme_check_size(void)
 
 static unsigned int max_io_queues(void)
 {
-       return num_possible_cpus() + write_queues;
+       return num_possible_cpus() + write_queues + poll_queues;
 }
 
 static unsigned int max_queue_count(void)
@@ -500,8 +506,15 @@ static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
                        offset = queue_irq_offset(dev);
                }
 
+               /*
+                * The poll queue(s) doesn't have an IRQ (and hence IRQ
+                * affinity), so use the regular blk-mq cpu mapping
+                */
                map->queue_offset = qoff;
-               blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
+               if (i != NVMEQ_TYPE_POLL)
+                       blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
+               else
+                       blk_mq_map_queues(map);
                qoff += map->nr_queues;
                offset += map->nr_queues;
        }
@@ -892,7 +905,7 @@ static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
         * We should not need to do this, but we're still using this to
         * ensure we can drain requests on a dying queue.
         */
-       if (unlikely(nvmeq->cq_vector < 0))
+       if (unlikely(nvmeq->cq_vector < 0 && !nvmeq->polled))
                return BLK_STS_IOERR;
 
        ret = nvme_setup_cmd(ns, req, &cmnd);
@@ -921,6 +934,8 @@ out_free_cmd:
 
 static int nvme_rq_flags_to_type(struct request_queue *q, unsigned int flags)
 {
+       if ((flags & REQ_HIPRI) && test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
+               return NVMEQ_TYPE_POLL;
        if ((flags & REQ_OP_MASK) == REQ_OP_READ)
                return NVMEQ_TYPE_READ;
 
@@ -1094,7 +1109,10 @@ static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
                struct nvme_queue *nvmeq, s16 vector)
 {
        struct nvme_command c;
-       int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
+       int flags = NVME_QUEUE_PHYS_CONTIG;
+
+       if (vector != -1)
+               flags |= NVME_CQ_IRQ_ENABLED;
 
        /*
         * Note: we (ab)use the fact that the prp fields survive if no data
@@ -1106,7 +1124,10 @@ static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
        c.create_cq.cqid = cpu_to_le16(qid);
        c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
        c.create_cq.cq_flags = cpu_to_le16(flags);
-       c.create_cq.irq_vector = cpu_to_le16(vector);
+       if (vector != -1)
+               c.create_cq.irq_vector = cpu_to_le16(vector);
+       else
+               c.create_cq.irq_vector = 0;
 
        return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
 }
@@ -1348,13 +1369,14 @@ static int nvme_suspend_queue(struct nvme_queue *nvmeq)
        int vector;
 
        spin_lock_irq(&nvmeq->cq_lock);
-       if (nvmeq->cq_vector == -1) {
+       if (nvmeq->cq_vector == -1 && !nvmeq->polled) {
                spin_unlock_irq(&nvmeq->cq_lock);
                return 1;
        }
        vector = nvmeq->cq_vector;
        nvmeq->dev->online_queues--;
        nvmeq->cq_vector = -1;
+       nvmeq->polled = false;
        spin_unlock_irq(&nvmeq->cq_lock);
 
        /*
@@ -1366,7 +1388,8 @@ static int nvme_suspend_queue(struct nvme_queue *nvmeq)
        if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
                blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
 
-       pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
+       if (vector != -1)
+               pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
 
        return 0;
 }
@@ -1500,7 +1523,7 @@ static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
        spin_unlock_irq(&nvmeq->cq_lock);
 }
 
-static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
+static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
 {
        struct nvme_dev *dev = nvmeq->dev;
        int result;
@@ -1510,7 +1533,11 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
         * A queue's vector matches the queue identifier unless the controller
         * has only one vector available.
         */
-       vector = dev->num_vecs == 1 ? 0 : qid;
+       if (!polled)
+               vector = dev->num_vecs == 1 ? 0 : qid;
+       else
+               vector = -1;
+
        result = adapter_alloc_cq(dev, qid, nvmeq, vector);
        if (result)
                return result;
@@ -1527,15 +1554,20 @@ static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
         * xxx' warning if the create CQ/SQ command times out.
         */
        nvmeq->cq_vector = vector;
+       nvmeq->polled = polled;
        nvme_init_queue(nvmeq, qid);
-       result = queue_request_irq(nvmeq);
-       if (result < 0)
-               goto release_sq;
+
+       if (vector != -1) {
+               result = queue_request_irq(nvmeq);
+               if (result < 0)
+                       goto release_sq;
+       }
 
        return result;
 
 release_sq:
        nvmeq->cq_vector = -1;
+       nvmeq->polled = false;
        dev->online_queues--;
        adapter_delete_sq(dev, qid);
 release_cq:
@@ -1686,7 +1718,7 @@ static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
 
 static int nvme_create_io_queues(struct nvme_dev *dev)
 {
-       unsigned i, max;
+       unsigned i, max, rw_queues;
        int ret = 0;
 
        for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
@@ -1697,8 +1729,17 @@ static int nvme_create_io_queues(struct nvme_dev *dev)
        }
 
        max = min(dev->max_qid, dev->ctrl.queue_count - 1);
+       if (max != 1 && dev->io_queues[NVMEQ_TYPE_POLL]) {
+               rw_queues = dev->io_queues[NVMEQ_TYPE_READ] +
+                               dev->io_queues[NVMEQ_TYPE_WRITE];
+       } else {
+               rw_queues = max;
+       }
+
        for (i = dev->online_queues; i <= max; i++) {
-               ret = nvme_create_queue(&dev->queues[i], i);
+               bool polled = i > rw_queues;
+
+               ret = nvme_create_queue(&dev->queues[i], i, polled);
                if (ret)
                        break;
        }
@@ -1973,6 +2014,7 @@ static int nvme_setup_host_mem(struct nvme_dev *dev)
 static void nvme_calc_io_queues(struct nvme_dev *dev, unsigned int nr_io_queues)
 {
        unsigned int this_w_queues = write_queues;
+       unsigned int this_p_queues = poll_queues;
 
        /*
         * Setup read/write queue split
@@ -1980,9 +2022,28 @@ static void nvme_calc_io_queues(struct nvme_dev *dev, unsigned int nr_io_queues)
        if (nr_io_queues == 1) {
                dev->io_queues[NVMEQ_TYPE_READ] = 1;
                dev->io_queues[NVMEQ_TYPE_WRITE] = 0;
+               dev->io_queues[NVMEQ_TYPE_POLL] = 0;
                return;
        }
 
+       /*
+        * Configure number of poll queues, if set
+        */
+       if (this_p_queues) {
+               /*
+                * We need at least one queue left. With just one queue, we'll
+                * have a single shared read/write set.
+                */
+               if (this_p_queues >= nr_io_queues) {
+                       this_w_queues = 0;
+                       this_p_queues = nr_io_queues - 1;
+               }
+
+               dev->io_queues[NVMEQ_TYPE_POLL] = this_p_queues;
+               nr_io_queues -= this_p_queues;
+       } else
+               dev->io_queues[NVMEQ_TYPE_POLL] = 0;
+
        /*
         * If 'write_queues' is set, ensure it leaves room for at least
         * one read queue
@@ -2099,11 +2160,13 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
                return -EIO;
 
        dev->num_vecs = result;
-       dev->max_qid = max(result - 1, 1);
+       result = max(result - 1, 1);
+       dev->max_qid = result + dev->io_queues[NVMEQ_TYPE_POLL];
 
-       dev_info(dev->ctrl.device, "%d/%d read/write queues\n",
+       dev_info(dev->ctrl.device, "%d/%d/%d read/write/poll queues\n",
                                        dev->io_queues[NVMEQ_TYPE_READ],
-                                       dev->io_queues[NVMEQ_TYPE_WRITE]);
+                                       dev->io_queues[NVMEQ_TYPE_WRITE],
+                                       dev->io_queues[NVMEQ_TYPE_POLL]);
 
        /*
         * Should investigate if there's a performance win from allocating
index 729ce0f00433105426899813ce90e4883270322d..9f5e93f40857c23ae7949f6af93f38eada21e846 100644 (file)
@@ -82,7 +82,7 @@ struct blk_mq_queue_map {
 };
 
 enum {
-       HCTX_MAX_TYPES = 2,
+       HCTX_MAX_TYPES = 3,
 };
 
 struct blk_mq_tag_set {