nvme-rdma: Avoid preallocating big SGL for data

nvme_rdma_alloc_tagset() preallocates a big buffer for the IO SGL based
on SG_CHUNK_SIZE.

Modern DMA engines are often capable of dealing with very big segments so
the SG_CHUNK_SIZE is often too big. SG_CHUNK_SIZE results in a static 4KB
SGL allocation per command.

If a controller has lots of deep queues, preallocation for the sg list can
consume substantial amounts of memory. For nvme-rdma, nr_hw_queues can be
128 and each queue's depth 128. This means the resulting preallocation
for the data SGL is 128*128*4K = 64MB per controller.

Switch to runtime allocation for SGL for lists longer than 2 entries. This
is the approach used by NVMe PCI so it should be reasonable for NVMeOF as
well. Runtime SGL allocation has always been the case for the legacy I/O
path so this is nothing new.

The preallocated small SGL depends on SG_CHAIN so if the ARCH doesn't
support SG_CHAIN, use only runtime allocation for the SGL.

We didn't notice of a performance degradation, since for small IOs we'll
use the inline SG and for the bigger IOs the allocation of a bigger SGL
from slab is fast enough.

Suggested-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Signed-off-by: Keith Busch <kbusch@kernel.org>

drivers/nvme/host/nvme.h

@@ -28,6 +28,12 @@ extern unsigned int admin_timeout;
 #define NVME_DEFAULT_KATO	5
 #define NVME_KATO_GRACE		10
 
+#ifdef CONFIG_ARCH_NO_SG_CHAIN
+#define  NVME_INLINE_SG_CNT  0
+#else
+#define  NVME_INLINE_SG_CNT  2
+#endif
+
 extern struct workqueue_struct *nvme_wq;
 extern struct workqueue_struct *nvme_reset_wq;
 extern struct workqueue_struct *nvme_delete_wq;

drivers/nvme/host/rdma.c

@@ -731,7 +731,7 @@ static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 		set->reserved_tags = 2; /* connect + keep-alive */
 		set->numa_node = nctrl->numa_node;
 		set->cmd_size = sizeof(struct nvme_rdma_request) +
-			SG_CHUNK_SIZE * sizeof(struct scatterlist);
+			NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 		set->driver_data = ctrl;
 		set->nr_hw_queues = 1;
 		set->timeout = ADMIN_TIMEOUT;
@@ -745,7 +745,7 @@ static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
 		set->numa_node = nctrl->numa_node;
 		set->flags = BLK_MQ_F_SHOULD_MERGE;
 		set->cmd_size = sizeof(struct nvme_rdma_request) +
-			SG_CHUNK_SIZE * sizeof(struct scatterlist);
+			NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 		set->driver_data = ctrl;
 		set->nr_hw_queues = nctrl->queue_count - 1;
 		set->timeout = NVME_IO_TIMEOUT;
@@ -1160,7 +1160,7 @@ static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 	}
 
 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq));
-	sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
+	sg_free_table_chained(&req->sg_table, NVME_INLINE_SG_CNT);
 }
 
 static int nvme_rdma_set_sg_null(struct nvme_command *c)
@@ -1276,7 +1276,7 @@ static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 	req->sg_table.sgl = req->first_sgl;
 	ret = sg_alloc_table_chained(&req->sg_table,
 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl,
-			SG_CHUNK_SIZE);
+			NVME_INLINE_SG_CNT);
 	if (ret)
 		return -ENOMEM;
 
@@ -1314,7 +1314,7 @@ out:
 out_unmap_sg:
 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq));
 out_free_table:
-	sg_free_table_chained(&req->sg_table, SG_CHUNK_SIZE);
+	sg_free_table_chained(&req->sg_table, NVME_INLINE_SG_CNT);
 	return ret;
 }