nvme: merge iod and cmd_info
authorChristoph Hellwig <hch@lst.de>
Sat, 28 Nov 2015 14:43:10 +0000 (15:43 +0100)
committerJens Axboe <axboe@fb.com>
Tue, 22 Dec 2015 16:38:34 +0000 (09:38 -0700)
Merge the two per-request structures in the nvme driver into a single
one.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/nvme/host/pci.c

index 24d695a2f6c4878e71d10df77afed0adc2796b21..b88708affad8ff17111685253ddae2731cd13013 100644 (file)
@@ -164,18 +164,19 @@ struct nvme_queue {
 /*
  * The nvme_iod describes the data in an I/O, including the list of PRP
  * entries.  You can't see it in this data structure because C doesn't let
- * me express that.  Use nvme_alloc_iod to ensure there's enough space
+ * me express that.  Use nvme_init_iod to ensure there's enough space
  * allocated to store the PRP list.
  */
 struct nvme_iod {
-       unsigned long private;  /* For the use of the submitter of the I/O */
+       struct nvme_queue *nvmeq;
+       int aborted;
        int npages;             /* In the PRP list. 0 means small pool in use */
-       int offset;             /* Of PRP list */
        int nents;              /* Used in scatterlist */
        int length;             /* Of data, in bytes */
        dma_addr_t first_dma;
        struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
-       struct scatterlist sg[0];
+       struct scatterlist *sg;
+       struct scatterlist inline_sg[0];
 };
 
 /*
@@ -197,19 +198,11 @@ static inline void _nvme_check_size(void)
        BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
 }
 
-struct nvme_cmd_info {
-       int aborted;
-       struct nvme_queue *nvmeq;
-       struct nvme_iod *iod;
-       struct nvme_iod __iod;
-};
-
 /*
  * Max size of iod being embedded in the request payload
  */
 #define NVME_INT_PAGES         2
 #define NVME_INT_BYTES(dev)    (NVME_INT_PAGES * (dev)->ctrl.page_size)
-#define NVME_INT_MASK          0x01
 
 /*
  * Will slightly overestimate the number of pages needed.  This is OK
@@ -223,15 +216,17 @@ static int nvme_npages(unsigned size, struct nvme_dev *dev)
        return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
 }
 
-static unsigned int nvme_cmd_size(struct nvme_dev *dev)
+static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
+               unsigned int size, unsigned int nseg)
 {
-       unsigned int ret = sizeof(struct nvme_cmd_info);
-
-       ret += sizeof(struct nvme_iod);
-       ret += sizeof(__le64 *) * nvme_npages(NVME_INT_BYTES(dev), dev);
-       ret += sizeof(struct scatterlist) * NVME_INT_PAGES;
+       return sizeof(__le64 *) * nvme_npages(size, dev) +
+                       sizeof(struct scatterlist) * nseg;
+}
 
-       return ret;
+static unsigned int nvme_cmd_size(struct nvme_dev *dev)
+{
+       return sizeof(struct nvme_iod) +
+               nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
 }
 
 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
@@ -261,11 +256,11 @@ static int nvme_admin_init_request(void *data, struct request *req,
                                unsigned int numa_node)
 {
        struct nvme_dev *dev = data;
-       struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct nvme_queue *nvmeq = dev->queues[0];
 
        BUG_ON(!nvmeq);
-       cmd->nvmeq = nvmeq;
+       iod->nvmeq = nvmeq;
        return 0;
 }
 
@@ -288,27 +283,14 @@ static int nvme_init_request(void *data, struct request *req,
                                unsigned int numa_node)
 {
        struct nvme_dev *dev = data;
-       struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
 
        BUG_ON(!nvmeq);
-       cmd->nvmeq = nvmeq;
+       iod->nvmeq = nvmeq;
        return 0;
 }
 
-static void *iod_get_private(struct nvme_iod *iod)
-{
-       return (void *) (iod->private & ~0x1UL);
-}
-
-/*
- * If bit 0 is set, the iod is embedded in the request payload.
- */
-static bool iod_should_kfree(struct nvme_iod *iod)
-{
-       return (iod->private & NVME_INT_MASK) == 0;
-}
-
 static void nvme_complete_async_event(struct nvme_dev *dev,
                struct nvme_completion *cqe)
 {
@@ -352,61 +334,44 @@ static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
        nvmeq->sq_tail = tail;
 }
 
-static __le64 **iod_list(struct nvme_iod *iod)
-{
-       return ((void *)iod) + iod->offset;
-}
-
-static inline void iod_init(struct nvme_iod *iod, unsigned nbytes,
-                           unsigned nseg, unsigned long private)
-{
-       iod->private = private;
-       iod->offset = offsetof(struct nvme_iod, sg[nseg]);
-       iod->npages = -1;
-       iod->length = nbytes;
-       iod->nents = 0;
-}
-
-static struct nvme_iod *
-__nvme_alloc_iod(unsigned nseg, unsigned bytes, struct nvme_dev *dev,
-                unsigned long priv, gfp_t gfp)
+static __le64 **iod_list(struct request *req)
 {
-       struct nvme_iod *iod = kmalloc(sizeof(struct nvme_iod) +
-                               sizeof(__le64 *) * nvme_npages(bytes, dev) +
-                               sizeof(struct scatterlist) * nseg, gfp);
-
-       if (iod)
-               iod_init(iod, bytes, nseg, priv);
-
-       return iod;
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
+       return (__le64 **)(iod->sg + req->nr_phys_segments);
 }
 
-static struct nvme_iod *nvme_alloc_iod(struct request *rq, struct nvme_dev *dev,
-                                      gfp_t gfp)
+static int nvme_init_iod(struct request *rq, struct nvme_dev *dev)
 {
-       unsigned size = !(rq->cmd_flags & REQ_DISCARD) ? blk_rq_bytes(rq) :
-                                                sizeof(struct nvme_dsm_range);
-       struct nvme_iod *iod;
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
+       int nseg = rq->nr_phys_segments;
+       unsigned size;
 
-       if (rq->nr_phys_segments <= NVME_INT_PAGES &&
-           size <= NVME_INT_BYTES(dev)) {
-               struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(rq);
+       if (rq->cmd_flags & REQ_DISCARD)
+               size = sizeof(struct nvme_dsm_range);
+       else
+               size = blk_rq_bytes(rq);
 
-               iod = &cmd->__iod;
-               iod_init(iod, size, rq->nr_phys_segments,
-                               (unsigned long) rq | NVME_INT_MASK);
-               return iod;
+       if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
+               iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
+               if (!iod->sg)
+                       return BLK_MQ_RQ_QUEUE_BUSY;
+       } else {
+               iod->sg = iod->inline_sg;
        }
 
-       return __nvme_alloc_iod(rq->nr_phys_segments, size, dev,
-                               (unsigned long) rq, gfp);
+       iod->aborted = 0;
+       iod->npages = -1;
+       iod->nents = 0;
+       iod->length = size;
+       return 0;
 }
 
-static void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
+static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
 {
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        const int last_prp = dev->ctrl.page_size / 8 - 1;
        int i;
-       __le64 **list = iod_list(iod);
+       __le64 **list = iod_list(req);
        dma_addr_t prp_dma = iod->first_dma;
 
        if (iod->npages == 0)
@@ -418,8 +383,8 @@ static void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
                prp_dma = next_prp_dma;
        }
 
-       if (iod_should_kfree(iod))
-               kfree(iod);
+       if (iod->sg != iod->inline_sg)
+               kfree(iod->sg);
 }
 
 #ifdef CONFIG_BLK_DEV_INTEGRITY
@@ -489,9 +454,10 @@ static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
 }
 #endif
 
-static bool nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
+static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req,
                int total_len)
 {
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct dma_pool *pool;
        int length = total_len;
        struct scatterlist *sg = iod->sg;
@@ -500,7 +466,7 @@ static bool nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
        u32 page_size = dev->ctrl.page_size;
        int offset = dma_addr & (page_size - 1);
        __le64 *prp_list;
-       __le64 **list = iod_list(iod);
+       __le64 **list = iod_list(req);
        dma_addr_t prp_dma;
        int nprps, i;
 
@@ -568,10 +534,10 @@ static bool nvme_setup_prps(struct nvme_dev *dev, struct nvme_iod *iod,
        return true;
 }
 
-static int nvme_map_data(struct nvme_dev *dev, struct nvme_iod *iod,
+static int nvme_map_data(struct nvme_dev *dev, struct request *req,
                struct nvme_command *cmnd)
 {
-       struct request *req = iod_get_private(iod);
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct request_queue *q = req->q;
        enum dma_data_direction dma_dir = rq_data_dir(req) ?
                        DMA_TO_DEVICE : DMA_FROM_DEVICE;
@@ -586,7 +552,7 @@ static int nvme_map_data(struct nvme_dev *dev, struct nvme_iod *iod,
        if (!dma_map_sg(dev->dev, iod->sg, iod->nents, dma_dir))
                goto out;
 
-       if (!nvme_setup_prps(dev, iod, blk_rq_bytes(req)))
+       if (!nvme_setup_prps(dev, req, blk_rq_bytes(req)))
                goto out_unmap;
 
        ret = BLK_MQ_RQ_QUEUE_ERROR;
@@ -617,9 +583,9 @@ out:
        return ret;
 }
 
-static void nvme_unmap_data(struct nvme_dev *dev, struct nvme_iod *iod)
+static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
 {
-       struct request *req = iod_get_private(iod);
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        enum dma_data_direction dma_dir = rq_data_dir(req) ?
                        DMA_TO_DEVICE : DMA_FROM_DEVICE;
 
@@ -632,7 +598,7 @@ static void nvme_unmap_data(struct nvme_dev *dev, struct nvme_iod *iod)
                }
        }
 
-       nvme_free_iod(dev, iod);
+       nvme_free_iod(dev, req);
 }
 
 /*
@@ -641,16 +607,16 @@ static void nvme_unmap_data(struct nvme_dev *dev, struct nvme_iod *iod)
  * the iod.
  */
 static int nvme_setup_discard(struct nvme_queue *nvmeq, struct nvme_ns *ns,
-               struct nvme_iod *iod, struct nvme_command *cmnd)
+               struct request *req, struct nvme_command *cmnd)
 {
-       struct request *req = iod_get_private(iod);
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
        struct nvme_dsm_range *range;
 
        range = dma_pool_alloc(nvmeq->dev->prp_small_pool, GFP_ATOMIC,
                                                &iod->first_dma);
        if (!range)
                return BLK_MQ_RQ_QUEUE_BUSY;
-       iod_list(iod)[0] = (__le64 *)range;
+       iod_list(req)[0] = (__le64 *)range;
        iod->npages = 0;
 
        range->cattr = cpu_to_le32(0);
@@ -676,8 +642,6 @@ static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
        struct nvme_queue *nvmeq = hctx->driver_data;
        struct nvme_dev *dev = nvmeq->dev;
        struct request *req = bd->rq;
-       struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
-       struct nvme_iod *iod;
        struct nvme_command cmnd;
        int ret = BLK_MQ_RQ_QUEUE_OK;
 
@@ -694,12 +658,12 @@ static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
                }
        }
 
-       iod = nvme_alloc_iod(req, dev, GFP_ATOMIC);
-       if (!iod)
-               return BLK_MQ_RQ_QUEUE_BUSY;
+       ret = nvme_init_iod(req, dev);
+       if (ret)
+               return ret;
 
        if (req->cmd_flags & REQ_DISCARD) {
-               ret = nvme_setup_discard(nvmeq, ns, iod, &cmnd);
+               ret = nvme_setup_discard(nvmeq, ns, req, &cmnd);
        } else {
                if (req->cmd_type == REQ_TYPE_DRV_PRIV)
                        memcpy(&cmnd, req->cmd, sizeof(cmnd));
@@ -709,14 +673,12 @@ static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
                        nvme_setup_rw(ns, req, &cmnd);
 
                if (req->nr_phys_segments)
-                       ret = nvme_map_data(dev, iod, &cmnd);
+                       ret = nvme_map_data(dev, req, &cmnd);
        }
 
        if (ret)
                goto out;
 
-       cmd->iod = iod;
-       cmd->aborted = 0;
        cmnd.common.command_id = req->tag;
        blk_mq_start_request(req);
 
@@ -726,17 +688,17 @@ static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
        spin_unlock_irq(&nvmeq->q_lock);
        return BLK_MQ_RQ_QUEUE_OK;
 out:
-       nvme_free_iod(dev, iod);
+       nvme_free_iod(dev, req);
        return ret;
 }
 
 static void nvme_complete_rq(struct request *req)
 {
-       struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
-       struct nvme_dev *dev = cmd->nvmeq->dev;
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
+       struct nvme_dev *dev = iod->nvmeq->dev;
        int error = 0;
 
-       nvme_unmap_data(dev, cmd->iod);
+       nvme_unmap_data(dev, req);
 
        if (unlikely(req->errors)) {
                if (nvme_req_needs_retry(req, req->errors)) {
@@ -750,7 +712,7 @@ static void nvme_complete_rq(struct request *req)
                        error = nvme_error_status(req->errors);
        }
 
-       if (unlikely(cmd->aborted)) {
+       if (unlikely(iod->aborted)) {
                dev_warn(dev->dev,
                        "completing aborted command with status: %04x\n",
                        req->errors);
@@ -955,8 +917,8 @@ static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
 
 static void abort_endio(struct request *req, int error)
 {
-       struct nvme_cmd_info *cmd = blk_mq_rq_to_pdu(req);
-       struct nvme_queue *nvmeq = cmd->nvmeq;
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
+       struct nvme_queue *nvmeq = iod->nvmeq;
        u32 result = (u32)(uintptr_t)req->special;
        u16 status = req->errors;
 
@@ -968,8 +930,8 @@ static void abort_endio(struct request *req, int error)
 
 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
 {
-       struct nvme_cmd_info *cmd_rq = blk_mq_rq_to_pdu(req);
-       struct nvme_queue *nvmeq = cmd_rq->nvmeq;
+       struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
+       struct nvme_queue *nvmeq = iod->nvmeq;
        struct nvme_dev *dev = nvmeq->dev;
        struct request *abort_req;
        struct nvme_command cmd;
@@ -994,7 +956,7 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
         * command was already aborted once before and still hasn't been
         * returned to the driver, or if this is the admin queue.
         */
-       if (!nvmeq->qid || cmd_rq->aborted) {
+       if (!nvmeq->qid || iod->aborted) {
                dev_warn(dev->dev,
                         "I/O %d QID %d timeout, reset controller\n",
                         req->tag, nvmeq->qid);
@@ -1009,7 +971,7 @@ static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
                return BLK_EH_HANDLED;
        }
 
-       cmd_rq->aborted = 1;
+       iod->aborted = 1;
 
        if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
                atomic_inc(&dev->ctrl.abort_limit);