[PATCH 0/3] nvme: Avoid preallocating big SGL for data

[PATCH 0/3] nvme: Avoid preallocating big SGL for data

[Israel Rukshin]

When nvme allocates a tagset it 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. 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.

Israel Rukshin (3):
  nvme-rdma: Avoid preallocating big SGL for data
  nvme-fc: Avoid preallocating big SGL for data
  nvmet-loop: Avoid preallocating big SGL for data

 drivers/nvme/host/fc.c     |  8 ++++----
 drivers/nvme/host/nvme.h   |  6 ++++++
 drivers/nvme/host/rdma.c   | 10 +++++-----
 drivers/nvme/target/loop.c |  8 ++++----
 4 files changed, 19 insertions(+), 13 deletions(-)

-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

[PATCH 1/3] nvme-rdma: Avoid preallocating big SGL for data

[Israel Rukshin]

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>
Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
---
 drivers/nvme/host/nvme.h |  6 ++++++
 drivers/nvme/host/rdma.c | 10 +++++-----
 2 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/drivers/nvme/host/nvme.h b/drivers/nvme/host/nvme.h
index 34ac79c..3615145 100644
--- a/drivers/nvme/host/nvme.h
+++ b/drivers/nvme/host/nvme.h
@@ -27,6 +27,12 @@
 #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;
diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index 05f2dfa..9a02fde 100644
--- a/drivers/nvme/host/rdma.c
+++ b/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 @@ static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 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;
 }
 
-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 1/3] nvme-rdma: Avoid preallocating big SGL for data

[Christoph Hellwig]

Looks good, although just like James I'd rather see this split into two
patches.  Otherwise:

Reviewed-by: Christoph Hellwig <hch@lst.de>

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

[PATCH 2/3] nvme-fc: Avoid preallocating big SGL for data

[Israel Rukshin]

nvme_fc_create_io_queues() 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-fc, 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.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
---
 drivers/nvme/host/fc.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/nvme/host/fc.c b/drivers/nvme/host/fc.c
index 679a721..13cb00e 100644
--- a/drivers/nvme/host/fc.c
+++ b/drivers/nvme/host/fc.c
@@ -95,7 +95,7 @@ struct nvme_fc_fcp_op {
 
 struct nvme_fcp_op_w_sgl {
 	struct nvme_fc_fcp_op	op;
-	struct scatterlist	sgl[SG_CHUNK_SIZE];
+	struct scatterlist	sgl[NVME_INLINE_SG_CNT];
 	uint8_t			priv[0];
 };
 
@@ -2141,7 +2141,7 @@ enum {
 	freq->sg_table.sgl = freq->first_sgl;
 	ret = sg_alloc_table_chained(&freq->sg_table,
 			blk_rq_nr_phys_segments(rq), freq->sg_table.sgl,
-			SG_CHUNK_SIZE);
+			NVME_INLINE_SG_CNT);
 	if (ret)
 		return -ENOMEM;
 
@@ -2150,7 +2150,7 @@ enum {
 	freq->sg_cnt = fc_dma_map_sg(ctrl->lport->dev, freq->sg_table.sgl,
 				op->nents, rq_dma_dir(rq));
 	if (unlikely(freq->sg_cnt <= 0)) {
-		sg_free_table_chained(&freq->sg_table, SG_CHUNK_SIZE);
+		sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
 		freq->sg_cnt = 0;
 		return -EFAULT;
 	}
@@ -2173,7 +2173,7 @@ enum {
 	fc_dma_unmap_sg(ctrl->lport->dev, freq->sg_table.sgl, op->nents,
 			rq_dma_dir(rq));
 
-	sg_free_table_chained(&freq->sg_table, SG_CHUNK_SIZE);
+	sg_free_table_chained(&freq->sg_table, NVME_INLINE_SG_CNT);
 
 	freq->sg_cnt = 0;
 }
-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 2/3] nvme-fc: Avoid preallocating big SGL for data

[James Smart]

On 11/24/2019 8:38 AM, Israel Rukshin wrote:
> nvme_fc_create_io_queues() 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-fc, 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.
>
> Signed-off-by: Israel Rukshin <israelr@mellanox.com>
> Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
> ---
>   drivers/nvme/host/fc.c | 8 ++++----
>   1 file changed, 4 insertions(+), 4 deletions(-)
>
>
Look ok to me.

Reviewed-by: James Smart <james.smart@broadcom.com>

Note: would have preferred to see this be 4 patches, with patch 1 be the 
header file addition only, but a minor nit.

-- james


_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

[PATCH 3/3] nvmet-loop: Avoid preallocating big SGL for data

[Israel Rukshin]

nvme_loop_create_io_queues() 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 nvmet-loop, 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.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
---
 drivers/nvme/target/loop.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/nvme/target/loop.c b/drivers/nvme/target/loop.c
index 856eb06..dae31bf 100644
--- a/drivers/nvme/target/loop.c
+++ b/drivers/nvme/target/loop.c
@@ -76,7 +76,7 @@ static void nvme_loop_complete_rq(struct request *req)
 {
 	struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
 
-	sg_free_table_chained(&iod->sg_table, SG_CHUNK_SIZE);
+	sg_free_table_chained(&iod->sg_table, NVME_INLINE_SG_CNT);
 	nvme_complete_rq(req);
 }
 
@@ -156,7 +156,7 @@ static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
 		iod->sg_table.sgl = iod->first_sgl;
 		if (sg_alloc_table_chained(&iod->sg_table,
 				blk_rq_nr_phys_segments(req),
-				iod->sg_table.sgl, SG_CHUNK_SIZE))
+				iod->sg_table.sgl, NVME_INLINE_SG_CNT))
 			return BLK_STS_RESOURCE;
 
 		iod->req.sg = iod->sg_table.sgl;
@@ -340,7 +340,7 @@ static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
 	ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
 	ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
 	ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
-		SG_CHUNK_SIZE * sizeof(struct scatterlist);
+		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 	ctrl->admin_tag_set.driver_data = ctrl;
 	ctrl->admin_tag_set.nr_hw_queues = 1;
 	ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
@@ -514,7 +514,7 @@ static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
 	ctrl->tag_set.numa_node = NUMA_NO_NODE;
 	ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
 	ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
-		SG_CHUNK_SIZE * sizeof(struct scatterlist);
+		NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
 	ctrl->tag_set.driver_data = ctrl;
 	ctrl->tag_set.nr_hw_queues = ctrl->ctrl.queue_count - 1;
 	ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 3/3] nvmet-loop: Avoid preallocating big SGL for data

[Chaitanya Kulkarni]

Looks good, tested with simple loop setup running ranread
4K I/Os with fio.

Tested-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>

On 11/24/2019 08:39 AM, Israel Rukshin wrote:
> nvme_loop_create_io_queues() 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 nvmet-loop, 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.
>
> Signed-off-by: Israel Rukshin<israelr@mellanox.com>
> Reviewed-by: Max Gurtovoy<maxg@mellanox.com>



_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 3/3] nvmet-loop: Avoid preallocating big SGL for data

[Christoph Hellwig]

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 3/3] nvmet-loop: Avoid preallocating big SGL for data

[Keith Busch]

On Sun, Nov 24, 2019 at 06:38:32PM +0200, Israel Rukshin wrote:
> @@ -156,7 +156,7 @@ static blk_status_t nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
>  		iod->sg_table.sgl = iod->first_sgl;
>  		if (sg_alloc_table_chained(&iod->sg_table,
>  				blk_rq_nr_phys_segments(req),
> -				iod->sg_table.sgl, SG_CHUNK_SIZE))
> +				iod->sg_table.sgl, NVME_INLINE_SG_CNT))
>  			return BLK_STS_RESOURCE;

Minor merge conflict here from a resource leak fix from Max, but I fixed
it up.

Series applied to nvme/for-5.5

_______________________________________________
linux-nvme mailing list
linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme