NVMe: Optimise memory usage for I/Os between 4k and 128k
Matthew Wilcox [Thu, 10 Feb 2011 15:30:34 +0000 (10:30 -0500)]
Add a second memory pool for smaller I/Os.  We can pack 16 of these on a
single page instead of using an entire page for each one.

Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>

drivers/block/nvme.c

index 80fe6a7..cd7aeba 100644 (file)
@@ -58,6 +58,7 @@ struct nvme_dev {
        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;
        u32 ctrl_config;
@@ -249,11 +250,6 @@ static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
        return 0;
 }
 
-static __le64 *alloc_prp_list(struct nvme_dev *dev, dma_addr_t *addr)
-{
-       return dma_pool_alloc(dev->prp_page_pool, GFP_ATOMIC, addr);
-}
-
 struct nvme_prps {
        int npages;
        dma_addr_t first_dma;
@@ -271,6 +267,9 @@ static void nvme_free_prps(struct nvme_queue *nvmeq, struct nvme_prps *prps)
                return;
 
        prp_dma = prps->first_dma;
+
+       if (prps->npages == 0)
+               dma_pool_free(dev->prp_small_pool, prps->list[0], prp_dma);
        for (i = 0; i < prps->npages; i++) {
                __le64 *prp_list = prps->list[i];
                dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
@@ -322,6 +321,7 @@ static struct nvme_prps *nvme_setup_prps(struct nvme_queue *nvmeq,
                                        struct scatterlist *sg, int length)
 {
        struct nvme_dev *dev = nvmeq->dev;
+       struct dma_pool *pool;
        int dma_len = sg_dma_len(sg);
        u64 dma_addr = sg_dma_address(sg);
        int offset = offset_in_page(dma_addr);
@@ -352,9 +352,16 @@ static struct nvme_prps *nvme_setup_prps(struct nvme_queue *nvmeq,
        nprps = DIV_ROUND_UP(length, PAGE_SIZE);
        npages = DIV_ROUND_UP(8 * nprps, PAGE_SIZE);
        prps = kmalloc(sizeof(*prps) + sizeof(__le64 *) * npages, GFP_ATOMIC);
-       prps->npages = npages;
        prp_page = 0;
-       prp_list = alloc_prp_list(dev, &prp_dma);
+       if (nprps <= (256 / 8)) {
+               pool = dev->prp_small_pool;
+               prps->npages = 0;
+       } else {
+               pool = dev->prp_page_pool;
+               prps->npages = npages;
+       }
+
+       prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
        prps->list[prp_page++] = prp_list;
        prps->first_dma = prp_dma;
        cmd->prp2 = cpu_to_le64(prp_dma);
@@ -362,7 +369,7 @@ static struct nvme_prps *nvme_setup_prps(struct nvme_queue *nvmeq,
        for (;;) {
                if (i == PAGE_SIZE / 8 - 1) {
                        __le64 *old_prp_list = prp_list;
-                       prp_list = alloc_prp_list(dev, &prp_dma);
+                       prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
                        prps->list[prp_page++] = prp_list;
                        old_prp_list[i] = cpu_to_le64(prp_dma);
                        i = 0;
@@ -1313,12 +1320,20 @@ static int nvme_setup_prp_pools(struct nvme_dev *dev)
        if (!dev->prp_page_pool)
                return -ENOMEM;
 
+       /* Optimisation for I/Os between 4k and 128k */
+       dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
+                                               256, 256, 0);
+       if (!dev->prp_small_pool) {
+               dma_pool_destroy(dev->prp_page_pool);
+               return -ENOMEM;
+       }
        return 0;
 }
 
 static void nvme_release_prp_pools(struct nvme_dev *dev)
 {
        dma_pool_destroy(dev->prp_page_pool);
+       dma_pool_destroy(dev->prp_small_pool);
 }
 
 /* XXX: Use an ida or something to let remove / add work correctly */