NVMe: per-cpu io queues
authorKeith Busch <keith.busch@intel.com>
Mon, 24 Mar 2014 16:46:25 +0000 (10:46 -0600)
committerMatthew Wilcox <matthew.r.wilcox@intel.com>
Thu, 10 Apr 2014 21:03:15 +0000 (17:03 -0400)
The device's IO queues are associated with CPUs, so we can use a per-cpu
variable to map the a qid to a cpu. This provides a convienient way
to optimally assign queues to multiple cpus when the device supports
fewer queues than the host has cpus. The previous implementation may
have assigned these poorly in these situations. This patch addresses
this by sharing queues among cpus that are "close" together and should
have a lower lock contention penalty.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
drivers/block/nvme-core.c
include/linux/nvme.h

index e9495f0bfad30d19a8604aa17d687e9ae60a19bd..48d7bd55207a452f43f37ba118d8695e05bf260b 100644 (file)
@@ -20,6 +20,7 @@
 #include <linux/bio.h>
 #include <linux/bitops.h>
 #include <linux/blkdev.h>
+#include <linux/cpu.h>
 #include <linux/delay.h>
 #include <linux/errno.h>
 #include <linux/fs.h>
@@ -35,6 +36,7 @@
 #include <linux/module.h>
 #include <linux/moduleparam.h>
 #include <linux/pci.h>
+#include <linux/percpu.h>
 #include <linux/poison.h>
 #include <linux/ptrace.h>
 #include <linux/sched.h>
@@ -96,6 +98,7 @@ struct nvme_queue {
        u8 cq_phase;
        u8 cqe_seen;
        u8 q_suspended;
+       cpumask_var_t cpu_mask;
        struct async_cmd_info cmdinfo;
        unsigned long cmdid_data[];
 };
@@ -270,14 +273,15 @@ static struct nvme_queue *raw_nvmeq(struct nvme_dev *dev, int qid)
 
 static struct nvme_queue *get_nvmeq(struct nvme_dev *dev) __acquires(RCU)
 {
+       unsigned queue_id = get_cpu_var(*dev->io_queue);
        rcu_read_lock();
-       return rcu_dereference(dev->queues[get_cpu() + 1]);
+       return rcu_dereference(dev->queues[queue_id]);
 }
 
 static void put_nvmeq(struct nvme_queue *nvmeq) __releases(RCU)
 {
-       put_cpu();
        rcu_read_unlock();
+       put_cpu_var(nvmeq->dev->io_queue);
 }
 
 static struct nvme_queue *lock_nvmeq(struct nvme_dev *dev, int q_idx)
@@ -1121,6 +1125,8 @@ static void nvme_free_queue(struct rcu_head *r)
                                (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
        dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
                                        nvmeq->sq_cmds, nvmeq->sq_dma_addr);
+       if (nvmeq->qid)
+               free_cpumask_var(nvmeq->cpu_mask);
        kfree(nvmeq);
 }
 
@@ -1128,8 +1134,6 @@ static void nvme_free_queues(struct nvme_dev *dev, int lowest)
 {
        int i;
 
-       for (i = num_possible_cpus(); i > dev->queue_count - 1; i--)
-               rcu_assign_pointer(dev->queues[i], NULL);
        for (i = dev->queue_count - 1; i >= lowest; i--) {
                struct nvme_queue *nvmeq = raw_nvmeq(dev, i);
                rcu_assign_pointer(dev->queues[i], NULL);
@@ -1154,6 +1158,7 @@ static int nvme_suspend_queue(struct nvme_queue *nvmeq)
                return 1;
        }
        nvmeq->q_suspended = 1;
+       nvmeq->dev->online_queues--;
        spin_unlock_irq(&nvmeq->q_lock);
 
        irq_set_affinity_hint(vector, NULL);
@@ -1208,6 +1213,9 @@ static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
        if (!nvmeq->sq_cmds)
                goto free_cqdma;
 
+       if (qid && !zalloc_cpumask_var(&nvmeq->cpu_mask, GFP_KERNEL))
+               goto free_sqdma;
+
        nvmeq->q_dmadev = dmadev;
        nvmeq->dev = dev;
        snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
@@ -1228,6 +1236,9 @@ static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
 
        return nvmeq;
 
+ free_sqdma:
+       dma_free_coherent(dmadev, SQ_SIZE(depth), (void *)nvmeq->sq_cmds,
+                                                       nvmeq->sq_dma_addr);
  free_cqdma:
        dma_free_coherent(dmadev, CQ_SIZE(depth), (void *)nvmeq->cqes,
                                                        nvmeq->cq_dma_addr);
@@ -1260,6 +1271,7 @@ static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
        memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
        nvme_cancel_ios(nvmeq, false);
        nvmeq->q_suspended = 0;
+       dev->online_queues++;
 }
 
 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
@@ -1835,6 +1847,143 @@ static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid,
        return NULL;
 }
 
+static int nvme_find_closest_node(int node)
+{
+       int n, val, min_val = INT_MAX, best_node = node;
+
+       for_each_online_node(n) {
+               if (n == node)
+                       continue;
+               val = node_distance(node, n);
+               if (val < min_val) {
+                       min_val = val;
+                       best_node = n;
+               }
+       }
+       return best_node;
+}
+
+static void nvme_set_queue_cpus(cpumask_t *qmask, struct nvme_queue *nvmeq,
+                                                               int count)
+{
+       int cpu;
+       for_each_cpu(cpu, qmask) {
+               if (cpumask_weight(nvmeq->cpu_mask) >= count)
+                       break;
+               if (!cpumask_test_and_set_cpu(cpu, nvmeq->cpu_mask))
+                       *per_cpu_ptr(nvmeq->dev->io_queue, cpu) = nvmeq->qid;
+       }
+}
+
+static void nvme_add_cpus(cpumask_t *mask, const cpumask_t *unassigned_cpus,
+       const cpumask_t *new_mask, struct nvme_queue *nvmeq, int cpus_per_queue)
+{
+       int next_cpu;
+       for_each_cpu(next_cpu, new_mask) {
+               cpumask_or(mask, mask, get_cpu_mask(next_cpu));
+               cpumask_or(mask, mask, topology_thread_cpumask(next_cpu));
+               cpumask_and(mask, mask, unassigned_cpus);
+               nvme_set_queue_cpus(mask, nvmeq, cpus_per_queue);
+       }
+}
+
+static void nvme_create_io_queues(struct nvme_dev *dev)
+{
+       unsigned i, max;
+
+       max = min(dev->max_qid, num_online_cpus());
+       for (i = dev->queue_count; i <= max; i++)
+               if (!nvme_alloc_queue(dev, i, dev->q_depth, i - 1))
+                       break;
+
+       max = min(dev->queue_count - 1, num_online_cpus());
+       for (i = dev->online_queues; i <= max; i++)
+               if (nvme_create_queue(raw_nvmeq(dev, i), i))
+                       break;
+}
+
+/*
+ * If there are fewer queues than online cpus, this will try to optimally
+ * assign a queue to multiple cpus by grouping cpus that are "close" together:
+ * thread siblings, core, socket, closest node, then whatever else is
+ * available.
+ */
+static void nvme_assign_io_queues(struct nvme_dev *dev)
+{
+       unsigned cpu, cpus_per_queue, queues, remainder, i;
+       cpumask_var_t unassigned_cpus;
+
+       nvme_create_io_queues(dev);
+
+       queues = min(dev->online_queues - 1, num_online_cpus());
+       if (!queues)
+               return;
+
+       cpus_per_queue = num_online_cpus() / queues;
+       remainder = queues - (num_online_cpus() - queues * cpus_per_queue);
+
+       if (!alloc_cpumask_var(&unassigned_cpus, GFP_KERNEL))
+               return;
+
+       cpumask_copy(unassigned_cpus, cpu_online_mask);
+       cpu = cpumask_first(unassigned_cpus);
+       for (i = 1; i <= queues; i++) {
+               struct nvme_queue *nvmeq = lock_nvmeq(dev, i);
+               cpumask_t mask;
+
+               cpumask_clear(nvmeq->cpu_mask);
+               if (!cpumask_weight(unassigned_cpus)) {
+                       unlock_nvmeq(nvmeq);
+                       break;
+               }
+
+               mask = *get_cpu_mask(cpu);
+               nvme_set_queue_cpus(&mask, nvmeq, cpus_per_queue);
+               if (cpus_weight(mask) < cpus_per_queue)
+                       nvme_add_cpus(&mask, unassigned_cpus,
+                               topology_thread_cpumask(cpu),
+                               nvmeq, cpus_per_queue);
+               if (cpus_weight(mask) < cpus_per_queue)
+                       nvme_add_cpus(&mask, unassigned_cpus,
+                               topology_core_cpumask(cpu),
+                               nvmeq, cpus_per_queue);
+               if (cpus_weight(mask) < cpus_per_queue)
+                       nvme_add_cpus(&mask, unassigned_cpus,
+                               cpumask_of_node(cpu_to_node(cpu)),
+                               nvmeq, cpus_per_queue);
+               if (cpus_weight(mask) < cpus_per_queue)
+                       nvme_add_cpus(&mask, unassigned_cpus,
+                               cpumask_of_node(
+                                       nvme_find_closest_node(
+                                               cpu_to_node(cpu))),
+                               nvmeq, cpus_per_queue);
+               if (cpus_weight(mask) < cpus_per_queue)
+                       nvme_add_cpus(&mask, unassigned_cpus,
+                               unassigned_cpus,
+                               nvmeq, cpus_per_queue);
+
+               WARN(cpumask_weight(nvmeq->cpu_mask) != cpus_per_queue,
+                       "nvme%d qid:%d mis-matched queue-to-cpu assignment\n",
+                       dev->instance, i);
+
+               irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
+                                                       nvmeq->cpu_mask);
+               cpumask_andnot(unassigned_cpus, unassigned_cpus,
+                                               nvmeq->cpu_mask);
+               cpu = cpumask_next(cpu, unassigned_cpus);
+               if (remainder && !--remainder)
+                       cpus_per_queue++;
+               unlock_nvmeq(nvmeq);
+       }
+       WARN(cpumask_weight(unassigned_cpus), "nvme%d unassigned online cpus\n",
+                                                               dev->instance);
+       i = 0;
+       cpumask_andnot(unassigned_cpus, cpu_possible_mask, cpu_online_mask);
+       for_each_cpu(cpu, unassigned_cpus)
+               *per_cpu_ptr(dev->io_queue, cpu) = (i++ % queues) + 1;
+       free_cpumask_var(unassigned_cpus);
+}
+
 static int set_queue_count(struct nvme_dev *dev, int count)
 {
        int status;
@@ -1857,9 +2006,9 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
 {
        struct nvme_queue *adminq = raw_nvmeq(dev, 0);
        struct pci_dev *pdev = dev->pci_dev;
-       int result, cpu, i, vecs, nr_io_queues, size, q_depth;
+       int result, i, vecs, nr_io_queues, size;
 
-       nr_io_queues = num_online_cpus();
+       nr_io_queues = num_possible_cpus();
        result = set_queue_count(dev, nr_io_queues);
        if (result < 0)
                return result;
@@ -1919,6 +2068,7 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
         * number of interrupts.
         */
        nr_io_queues = vecs;
+       dev->max_qid = nr_io_queues;
 
        result = queue_request_irq(dev, adminq, adminq->irqname);
        if (result) {
@@ -1927,36 +2077,8 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
        }
 
        /* Free previously allocated queues that are no longer usable */
-       nvme_free_queues(dev, nr_io_queues);
-
-       cpu = cpumask_first(cpu_online_mask);
-       for (i = 0; i < nr_io_queues; i++) {
-               irq_set_affinity_hint(dev->entry[i].vector, get_cpu_mask(cpu));
-               cpu = cpumask_next(cpu, cpu_online_mask);
-       }
-
-       q_depth = min_t(int, NVME_CAP_MQES(readq(&dev->bar->cap)) + 1,
-                                                               NVME_Q_DEPTH);
-       for (i = dev->queue_count - 1; i < nr_io_queues; i++) {
-               if (!nvme_alloc_queue(dev, i + 1, q_depth, i)) {
-                       result = -ENOMEM;
-                       goto free_queues;
-               }
-       }
-
-       for (; i < num_possible_cpus(); i++) {
-               int target = i % rounddown_pow_of_two(dev->queue_count - 1);
-               rcu_assign_pointer(dev->queues[i + 1], dev->queues[target + 1]);
-       }
-
-       for (i = 1; i < dev->queue_count; i++) {
-               result = nvme_create_queue(raw_nvmeq(dev, i), i);
-               if (result) {
-                       for (--i; i > 0; i--)
-                               nvme_disable_queue(dev, i);
-                       goto free_queues;
-               }
-       }
+       nvme_free_queues(dev, nr_io_queues + 1);
+       nvme_assign_io_queues(dev);
 
        return 0;
 
@@ -2035,6 +2157,7 @@ static int nvme_dev_add(struct nvme_dev *dev)
 
 static int nvme_dev_map(struct nvme_dev *dev)
 {
+       u64 cap;
        int bars, result = -ENOMEM;
        struct pci_dev *pdev = dev->pci_dev;
 
@@ -2058,7 +2181,9 @@ static int nvme_dev_map(struct nvme_dev *dev)
                result = -ENODEV;
                goto unmap;
        }
-       dev->db_stride = 1 << NVME_CAP_STRIDE(readq(&dev->bar->cap));
+       cap = readq(&dev->bar->cap);
+       dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
+       dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
        dev->dbs = ((void __iomem *)dev->bar) + 4096;
 
        return 0;
@@ -2332,6 +2457,7 @@ static void nvme_free_dev(struct kref *kref)
        struct nvme_dev *dev = container_of(kref, struct nvme_dev, kref);
 
        nvme_free_namespaces(dev);
+       free_percpu(dev->io_queue);
        kfree(dev->queues);
        kfree(dev->entry);
        kfree(dev);
@@ -2477,6 +2603,9 @@ static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
                                                                GFP_KERNEL);
        if (!dev->queues)
                goto free;
+       dev->io_queue = alloc_percpu(unsigned short);
+       if (!dev->io_queue)
+               goto free;
 
        INIT_LIST_HEAD(&dev->namespaces);
        INIT_WORK(&dev->reset_work, nvme_reset_failed_dev);
@@ -2526,6 +2655,7 @@ static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
  release:
        nvme_release_instance(dev);
  free:
+       free_percpu(dev->io_queue);
        kfree(dev->queues);
        kfree(dev->entry);
        kfree(dev);
index 7c3f85bc10f19d9c5a3db0e232cf32384a0507b3..f0f95c7196850d2cdd13960c2aa76b034e36bc7e 100644 (file)
@@ -74,12 +74,16 @@ enum {
 struct nvme_dev {
        struct list_head node;
        struct nvme_queue __rcu **queues;
+       unsigned short __percpu *io_queue;
        u32 __iomem *dbs;
        struct pci_dev *pci_dev;
        struct dma_pool *prp_page_pool;
        struct dma_pool *prp_small_pool;
        int instance;
-       int queue_count;
+       unsigned queue_count;
+       unsigned online_queues;
+       unsigned max_qid;
+       int q_depth;
        u32 db_stride;
        u32 ctrl_config;
        struct msix_entry *entry;