NVMe: Remove hctx reliance for multi-namespace
authorKeith Busch <keith.busch@intel.com>
Mon, 1 Jun 2015 15:29:54 +0000 (09:29 -0600)
committerJens Axboe <axboe@fb.com>
Mon, 1 Jun 2015 20:36:06 +0000 (14:36 -0600)
The driver needs to track shared tags to support multiple namespaces
that may be dynamically allocated or deleted. Relying on the first
request_queue's hctx's is not appropriate as we cannot clear outstanding
tags for all namespaces using this handle, nor can the driver easily track
all request_queue's hctx as namespaces are attached/detached. Instead,
this patch uses the nvme_dev's tagset to get the shared tag resources
instead of through a request_queue hctx.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/block/nvme-core.c

index 6ed1356e9eb589ad4e6ff08f69f178cb95e84402..513908ff46c4b34129ba6f17f6eaddc1c23ccd06 100644 (file)
@@ -102,6 +102,7 @@ struct nvme_queue {
        spinlock_t q_lock;
        struct nvme_command *sq_cmds;
        volatile struct nvme_completion *cqes;
+       struct blk_mq_tags **tags;
        dma_addr_t sq_dma_addr;
        dma_addr_t cq_dma_addr;
        u32 __iomem *q_db;
@@ -114,7 +115,6 @@ struct nvme_queue {
        u8 cq_phase;
        u8 cqe_seen;
        struct async_cmd_info cmdinfo;
-       struct blk_mq_hw_ctx *hctx;
 };
 
 /*
@@ -182,9 +182,12 @@ static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
        struct nvme_dev *dev = data;
        struct nvme_queue *nvmeq = dev->queues[0];
 
-       WARN_ON(nvmeq->hctx);
-       nvmeq->hctx = hctx;
+       WARN_ON(hctx_idx != 0);
+       WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
+       WARN_ON(nvmeq->tags);
+
        hctx->driver_data = nvmeq;
+       nvmeq->tags = &dev->admin_tagset.tags[0];
        return 0;
 }
 
@@ -201,27 +204,16 @@ static int nvme_admin_init_request(void *data, struct request *req,
        return 0;
 }
 
-static void nvme_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
-{
-       struct nvme_queue *nvmeq = hctx->driver_data;
-
-       nvmeq->hctx = NULL;
-}
-
 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
                          unsigned int hctx_idx)
 {
        struct nvme_dev *dev = data;
-       struct nvme_queue *nvmeq = dev->queues[
-                                       (hctx_idx % dev->queue_count) + 1];
-
-       if (!nvmeq->hctx)
-               nvmeq->hctx = hctx;
+       struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
 
-       /* nvmeq queues are shared between namespaces. We assume here that
-        * blk-mq map the tags so they match up with the nvme queue tags. */
-       WARN_ON(nvmeq->hctx->tags != hctx->tags);
+       if (!nvmeq->tags)
+               nvmeq->tags = &dev->tagset.tags[hctx_idx];
 
+       WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
        hctx->driver_data = nvmeq;
        return 0;
 }
@@ -320,7 +312,7 @@ static void abort_completion(struct nvme_queue *nvmeq, void *ctx,
        u16 status = le16_to_cpup(&cqe->status) >> 1;
        u32 result = le32_to_cpup(&cqe->result);
 
-       blk_mq_free_hctx_request(nvmeq->hctx, req);
+       blk_mq_free_request(req);
 
        dev_warn(nvmeq->q_dmadev, "Abort status:%x result:%x", status, result);
        ++nvmeq->dev->abort_limit;
@@ -333,14 +325,13 @@ static void async_completion(struct nvme_queue *nvmeq, void *ctx,
        cmdinfo->result = le32_to_cpup(&cqe->result);
        cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
        queue_kthread_work(cmdinfo->worker, &cmdinfo->work);
-       blk_mq_free_hctx_request(nvmeq->hctx, cmdinfo->req);
+       blk_mq_free_request(cmdinfo->req);
 }
 
 static inline struct nvme_cmd_info *get_cmd_from_tag(struct nvme_queue *nvmeq,
                                  unsigned int tag)
 {
-       struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
-       struct request *req = blk_mq_tag_to_rq(hctx->tags, tag);
+       struct request *req = blk_mq_tag_to_rq(*nvmeq->tags, tag);
 
        return blk_mq_rq_to_pdu(req);
 }
@@ -1068,7 +1059,7 @@ static int nvme_submit_async_admin_req(struct nvme_dev *dev)
        c.common.opcode = nvme_admin_async_event;
        c.common.command_id = req->tag;
 
-       blk_mq_free_hctx_request(nvmeq->hctx, req);
+       blk_mq_free_request(req);
        return __nvme_submit_cmd(nvmeq, &c);
 }
 
@@ -1310,8 +1301,7 @@ static void nvme_abort_req(struct request *req)
        }
 }
 
-static void nvme_cancel_queue_ios(struct blk_mq_hw_ctx *hctx,
-                               struct request *req, void *data, bool reserved)
+static void nvme_cancel_queue_ios(struct request *req, void *data, bool reserved)
 {
        struct nvme_queue *nvmeq = data;
        void *ctx;
@@ -1408,11 +1398,9 @@ static int nvme_suspend_queue(struct nvme_queue *nvmeq)
 
 static void nvme_clear_queue(struct nvme_queue *nvmeq)
 {
-       struct blk_mq_hw_ctx *hctx = nvmeq->hctx;
-
        spin_lock_irq(&nvmeq->q_lock);
-       if (hctx && hctx->tags)
-               blk_mq_tag_busy_iter(hctx, nvme_cancel_queue_ios, nvmeq);
+       if (nvmeq->tags && *nvmeq->tags)
+               blk_mq_all_tag_busy_iter(*nvmeq->tags, nvme_cancel_queue_ios, nvmeq);
        spin_unlock_irq(&nvmeq->q_lock);
 }
 
@@ -1605,7 +1593,6 @@ static struct blk_mq_ops nvme_mq_admin_ops = {
        .queue_rq       = nvme_queue_rq,
        .map_queue      = blk_mq_map_queue,
        .init_hctx      = nvme_admin_init_hctx,
-       .exit_hctx      = nvme_exit_hctx,
        .init_request   = nvme_admin_init_request,
        .timeout        = nvme_timeout,
 };
@@ -1614,7 +1601,6 @@ static struct blk_mq_ops nvme_mq_ops = {
        .queue_rq       = nvme_queue_rq,
        .map_queue      = blk_mq_map_queue,
        .init_hctx      = nvme_init_hctx,
-       .exit_hctx      = nvme_exit_hctx,
        .init_request   = nvme_init_request,
        .timeout        = nvme_timeout,
 };
@@ -2724,11 +2710,11 @@ static void nvme_set_irq_hints(struct nvme_dev *dev)
        for (i = 0; i < dev->online_queues; i++) {
                nvmeq = dev->queues[i];
 
-               if (!nvmeq->hctx)
+               if (!nvmeq->tags || !(*nvmeq->tags))
                        continue;
 
                irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
-                                                       nvmeq->hctx->cpumask);
+                                       blk_mq_tags_cpumask(*nvmeq->tags));
        }
 }