Skip to contentSkip to conversation

Optimize SGL registration

PATCH Applied Leon Romanovsky 5 participants 17 messages Oct 7, 2019 – Oct 22, 2019

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

v2 cover Optimize SGL registration

From: Leon Romanovsky <[email protected]>

Changelog
v1->v2: https://lore.kernel.org/linux-rdma/[email protected]
* Used Christoph's comment
* Change patch code flow to have one DMA_FROM_DEVICE check
v0->v1: https://lore.kernel.org/linux-rdma/[email protected]
* Reorganized patches to have IB/core changes separated from mlx5
* Moved SGL check before rdma_rw_force_mr
* Added and rephrased original comment.

-----------------------------------------------------------------------------
Hi,

This series from Yamin implements long standing "TODO" existed in rw.c.

Thanks

Yamin Friedman (3):
net/mlx5: Expose optimal performance scatter entries capability RDMA/rw: Support threshold for registration vs scattering to local
pages
RDMA/mlx5: Add capability for max sge to get optimized performance

 drivers/infiniband/core/rw.c      | 25 +++++++++++++++----------
 drivers/infiniband/hw/mlx5/main.c |  2 ++
 include/linux/mlx5/mlx5_ifc.h     |  2 +-
 include/rdma/ib_verbs.h           |  2 ++
 4 files changed, 20 insertions(+), 11 deletions(-)

--
2.20.1

#1 lore raw

PATCH v2 1/3 net/mlx5: Expose optimal performance scatter entries capability

From: Yamin Friedman <[email protected]>

Expose maximum scatter entries per RDMA READ for optimal performance.

---
 include/linux/mlx5/mlx5_ifc.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/linux/mlx5/mlx5_ifc.h +1 -1
index 138c50d5a353..c0bfb1d90dd2 100644
@@ -1153,7 +1153,7 @@ struct mlx5_ifc_cmd_hca_cap_bits {
 	u8         log_max_srq[0x5];
 	u8         reserved_at_b0[0x10];
 
-	u8         reserved_at_c0[0x8];
+	u8         max_sgl_for_optimized_performance[0x8];
 	u8         log_max_cq_sz[0x8];
 	u8         reserved_at_d0[0xb];
 	u8         log_max_cq[0x5];

--
2.20.1

#5 lore raw

PATCH v2 2/3 RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Leon Romanovsky

From: Yamin Friedman <[email protected]>

If there are more scatter entries than the recommended limit provided by the ib device, UMR registration is used. This will provide optimal performance when performing large RDMA READs over devices that advertise the threshold capability.

With ConnectX-5 running NVMeoF RDMA with FIO single QP 128KB writes:
Without use of cap: 70Gb/sec
With use of cap: 84Gb/sec

---
 drivers/infiniband/core/rw.c | 25 +++++++++++++++----------
 include/rdma/ib_verbs.h      |  2 ++
 2 files changed, 17 insertions(+), 10 deletions(-)
drivers/infiniband/core/rw.c +15 -10
index 5337393d4dfe..c27a543b58ef 100644
@@ -20,14 +20,16 @@ module_param_named(force_mr, rdma_rw_force_mr, bool, 0);
 MODULE_PARM_DESC(force_mr, "Force usage of MRs for RDMA READ/WRITE operations");
 
 /*
- * Check if the device might use memory registration.  This is currently only
- * true for iWarp devices. In the future we can hopefully fine tune this based
- * on HCA driver input.
+ * Check if the device might use memory registration. This is currently
+ * true for iWarp devices and devices that have optimized SGL registration
+ * logic.
  */
 static inline bool rdma_rw_can_use_mr(struct ib_device *dev, u8 port_num)
 {
 	if (rdma_protocol_iwarp(dev, port_num))
 		return true;
+	if (dev->attrs.max_sgl_rd)
+		return true;
 	if (unlikely(rdma_rw_force_mr))
 		return true;
 	return false;
@@ -35,17 +37,20 @@ static inline bool rdma_rw_can_use_mr(struct ib_device *dev, u8 port_num)
 
 /*
  * Check if the device will use memory registration for this RW operation.
- * We currently always use memory registrations for iWarp RDMA READs, and
- * have a debug option to force usage of MRs.
- *
- * XXX: In the future we can hopefully fine tune this based on HCA driver
- * input.
+ * For RDMA READs we must use MRs on iWarp and can optionaly use them as an
+ * optimaztion otherwise.  Additionally we have a debug option to force usage
+ * of MRs to help testing this code path.
  */
+
 static inline bool rdma_rw_io_needs_mr(struct ib_device *dev, u8 port_num,
 		enum dma_data_direction dir, int dma_nents)
 {
-	if (rdma_protocol_iwarp(dev, port_num) && dir == DMA_FROM_DEVICE)
-		return true;
+	if (dir == DMA_FROM_DEVICE) {
+		if (rdma_protocol_iwarp(dev, port_num))
+			return true;
+		if (dev->attrs.max_sgl_rd && dma_nents > dev->attrs.max_sgl_rd)
+			return true;
+	}
 	if (unlikely(rdma_rw_force_mr))
 		return true;
 	return false;
include/rdma/ib_verbs.h +2 -0
index 4f671378dbfc..60fd98a9b7e8 100644
@@ -445,6 +445,8 @@ struct ib_device_attr {
 	struct ib_tm_caps	tm_caps;
 	struct ib_cq_caps       cq_caps;
 	u64			max_dm_size;
+	/* Max entries for sgl for optimized performance per READ */
+	u32			max_sgl_rd;
 };
 
 enum ib_mtu {

--
2.20.1

#14 lore raw

PATCH v2 3/3 RDMA/mlx5: Add capability for max sge to get optimized performance

In reply to Leon Romanovsky

From: Yamin Friedman <[email protected]>

Allows the IB device to provide a value of maximum scatter gather entries per RDMA READ.

In certain cases it may be preferable for a device to perform UMR memory registration rather than have many scatter entries in a single RDMA READ. This provides a significant performance increase in devices capable of using different memory registration schemes based on the number of scatter gather entries. This general capability allows each device vendor to fine tune when it is better to use memory registration.

---
 drivers/infiniband/hw/mlx5/main.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/infiniband/hw/mlx5/main.c +2 -0
index fa23c8e7043b..39d54e285ae9 100644
@@ -1012,6 +1012,8 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		1 << MLX5_CAP_GEN(mdev, log_max_klm_list_size);
 	props->max_pi_fast_reg_page_list_len =
 		props->max_fast_reg_page_list_len / 2;
+	props->max_sgl_rd =
+		MLX5_CAP_GEN(mdev, max_sgl_for_optimized_performance);
 	get_atomic_caps_qp(dev, props);
 	props->masked_atomic_cap   = IB_ATOMIC_NONE;
 	props->max_mcast_grp	   = 1 << MLX5_CAP_GEN(mdev, log_max_mcg);

--
2.20.1

#2 lore raw

Re: [PATCH mlx5-next v2 1/3] net/mlx5: Expose optimal performance scatter entries capability

In reply to Leon Romanovsky

include/linux/mlx5/mlx5_ifc.h struct mlx5_ifc_cmd_hca_cap_bits { · line 1156

Quoted 2 lines · Leon Romanovsky
include/linux/mlx5/mlx5_ifc.h +1 -1
-	u8         reserved_at_c0[0x8];
+	u8         max_sgl_for_optimized_performance[0x8];

Should the name of this member variable perhaps be changed into "max_sgl_for_optimal_performance"?

Thanks,

Bart.

#8 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Leon Romanovsky
Quoted 8 lines · Leon Romanovsky

/*
- * Check if the device might use memory registration. This is currently only
- * true for iWarp devices. In the future we can hopefully fine tune this based
- * on HCA driver input.
+ * Check if the device might use memory registration. This is currently
+ * true for iWarp devices and devices that have optimized SGL registration
+ * logic.
*/

The following sentence in the above comment looks confusing to me: "Check if the device might use memory registration." That sentence suggests that the HCA decides whether or not to use memory registration. Isn't it the RDMA R/W code that decides whether or not to use memory registration?

Quoted 3 lines

+ * For RDMA READs we must use MRs on iWarp and can optionaly use them as an
+ * optimaztion otherwise. Additionally we have a debug option to force usage
+ * of MRs to help testing this code path.

You may want to change "optionaly" into "optionally" and "optimaztion" into "optimization".

Quoted 4 lines

static inline bool rdma_rw_io_needs_mr(struct ib_device *dev, u8 port_num,
enum dma_data_direction dir, int dma_nents)
{

drivers/infiniband/core/rw.c +0 -1
-	if (rdma_protocol_iwarp(dev, port_num) && dir == DMA_FROM_DEVICE)

drivers/infiniband/core/rw.c static inline bool rdma_rw_can_use_mr(struct ib_device *dev, u8 port_num) · line 56

Quoted 10 lines
drivers/infiniband/core/rw.c +6 -1
-		return true;
+	if (dir == DMA_FROM_DEVICE) {
+		if (rdma_protocol_iwarp(dev, port_num))
+			return true;
+		if (dev->attrs.max_sgl_rd && dma_nents > dev->attrs.max_sgl_rd)
+			return true;
+	}
  	if (unlikely(rdma_rw_force_mr))
  		return true;
  	return false;

Should this function be renamed? The function name suggests if this function returns 'true' that using memory registration is mandatory. My understanding is if this function returns true for the mlx5 HCA that using memory registration improves performance but is not mandatory.

Thanks,

Bart.

#12 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Leon Romanovsky

include/rdma/ib_verbs.h struct ib_device_attr { · line 448

Quoted 9 lines · Leon Romanovsky
include/rdma/ib_verbs.h +1 -0
index 4f671378dbfc..60fd98a9b7e8 100644
@@ -445,6 +445,8 @@ struct ib_device_attr {
  	struct ib_tm_caps	tm_caps;
  	struct ib_cq_caps       cq_caps;
  	u64			max_dm_size;
+	/* Max entries for sgl for optimized performance per READ */
                                   ^^^^^^^^^
                                   optimal?

Should it be mentioned that zero means that the HCA has not set this parameter?

include/rdma/ib_verbs.h struct ib_device_attr { · line 449

Quoted 1 line
include/rdma/ib_verbs.h +1 -0
+	u32			max_sgl_rd;

Thanks,

Bart.

#3 lore raw

Re: [PATCH mlx5-next v2 1/3] net/mlx5: Expose optimal performance scatter entries capability

In reply to Bart Van Assche
Quoted 6 lines · Bart Van Assche

include/linux/mlx5/mlx5_ifc.h struct mlx5_ifc_cmd_hca_cap_bits { · line 1156

Quoted 2 lines · Leon Romanovsky
include/linux/mlx5/mlx5_ifc.h +1 -1
-	u8         reserved_at_c0[0x8];
+	u8         max_sgl_for_optimized_performance[0x8];

Should the name of this member variable perhaps be changed into "max_sgl_for_optimal_performance"?

We don't want to force our internal HW/FW names on all uverbs users and drivers. So, the answer is no, it is used for optimized performance, but it is not the proper name for uverbs.

Thanks

Quoted 4 lines

Thanks,

Bart.

#7 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Christoph Hellwig
Quoted 8 lines · Christoph Hellwig
Quoted 3 lines

*/
+
static inline bool rdma_rw_io_needs_mr(struct ib_device *dev, u8 port_num,

Same like an empty line sneaked in here. Except for that the whole series looks fine:

Thanks

#13 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Bart Van Assche
Quoted 13 lines · Bart Van Assche

include/rdma/ib_verbs.h struct ib_device_attr { · line 448

Quoted 9 lines · Leon Romanovsky
include/rdma/ib_verbs.h +1 -0
index 4f671378dbfc..60fd98a9b7e8 100644
@@ -445,6 +445,8 @@ struct ib_device_attr {
  	struct ib_tm_caps	tm_caps;
  	struct ib_cq_caps       cq_caps;
  	u64			max_dm_size;
+	/* Max entries for sgl for optimized performance per READ */
                                   ^^^^^^^^^
                                   optimal?
Quoted 2 lines

Should it be mentioned that zero means that the HCA has not set this parameter?

It is always the case for other max_* values.

Quoted 6 lines

include/rdma/ib_verbs.h struct ib_device_attr { · line 449

Quoted 2 lines
include/rdma/ib_verbs.h +1 -0
+	u32			max_sgl_rd;

Thanks,

Bart.

#9 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

In reply to Bart Van Assche
Quoted 10 lines · Bart Van Assche
Quoted 8 lines · Leon Romanovsky

/*
- * Check if the device might use memory registration. This is currently only
- * true for iWarp devices. In the future we can hopefully fine tune this based
- * on HCA driver input.
+ * Check if the device might use memory registration. This is currently
+ * true for iWarp devices and devices that have optimized SGL registration
+ * logic.
*/

Quoted 4 lines

The following sentence in the above comment looks confusing to me: "Check if the device might use memory registration." That sentence suggests that the HCA decides whether or not to use memory registration. Isn't it the RDMA R/W code that decides whether or not to use memory registration?

I'm open for any reasonable text, what do you expect to be written there?

Quoted 7 lines
Quoted 4 lines

+ * For RDMA READs we must use MRs on iWarp and can optionaly use them as an
+ * optimaztion otherwise. Additionally we have a debug option to force usage
+ * of MRs to help testing this code path.

You may want to change "optionaly" into "optionally" and "optimaztion" into "optimization".

Thanks

Quoted 12 lines
Quoted 12 lines

static inline bool rdma_rw_io_needs_mr(struct ib_device *dev, u8 port_num,
enum dma_data_direction dir, int dma_nents)
{

drivers/infiniband/core/rw.c +6 -2
-	if (rdma_protocol_iwarp(dev, port_num) && dir == DMA_FROM_DEVICE)
-		return true;
+	if (dir == DMA_FROM_DEVICE) {
+		if (rdma_protocol_iwarp(dev, port_num))
+			return true;
+		if (dev->attrs.max_sgl_rd && dma_nents > dev->attrs.max_sgl_rd)
+			return true;
+	}
Quoted 8 lines
Quoted 3 lines

if (unlikely(rdma_rw_force_mr))
return true;
return false;

Should this function be renamed? The function name suggests if this function returns 'true' that using memory registration is mandatory. My understanding is if this function returns true for the mlx5 HCA that using memory registration improves performance but is not mandatory.

The end result the same, better to work with MR while working with mlx5 for "dma_nents > dev->attrs.max_sgl_rd",

Thanks

Quoted 4 lines

Thanks,

Bart.

#10 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

Quoted 11 lines · Leon Romanovsky
Quoted 10 lines · Bart Van Assche
Quoted 8 lines · Leon Romanovsky
   /*

- * Check if the device might use memory registration. This is currently only
- * true for iWarp devices. In the future we can hopefully fine tune this based
- * on HCA driver input.
+ * Check if the device might use memory registration. This is currently
+ * true for iWarp devices and devices that have optimized SGL registration
+ * logic.

    */
Quoted 6 lines
Quoted 4 lines

The following sentence in the above comment looks confusing to me: "Check if the device might use memory registration." That sentence suggests that the HCA decides whether or not to use memory registration. Isn't it the RDMA R/W code that decides whether or not to use memory registration?

I'm open for any reasonable text, what do you expect to be written there?

Hi Leon,

How about the following (not sure whether this is correct)?

/*
* Report whether memory registration should be used. Memory
* registration must be used for iWarp devices because of
* iWARP-specific limitations. Memory registration is also enabled if
* registering memory will yield better performance than using multiple
* SGE entries.
*/

Thanks,

Bart.

#11 lore raw

Re: [PATCH rdma-next v2 2/3] RDMA/rw: Support threshold for registration vs scattering to local pages

Quoted 23 lines · Bart Van Assche
Quoted 11 lines · Leon Romanovsky
Quoted 10 lines · Bart Van Assche
Quoted 8 lines · Leon Romanovsky
   /*

- * Check if the device might use memory registration. This is currently only
- * true for iWarp devices. In the future we can hopefully fine tune this based
- * on HCA driver input.
+ * Check if the device might use memory registration. This is currently
+ * true for iWarp devices and devices that have optimized SGL registration
+ * logic.

    */
Quoted 6 lines
Quoted 4 lines

The following sentence in the above comment looks confusing to me: "Check if the device might use memory registration." That sentence suggests that the HCA decides whether or not to use memory registration. Isn't it the RDMA R/W code that decides whether or not to use memory registration?

I'm open for any reasonable text, what do you expect to be written there?

Hi Leon,

How about the following (not sure whether this is correct)?

Quoted 7 lines

/*
* Report whether memory registration should be used. Memory
* registration must be used for iWarp devices because of
* iWARP-specific limitations. Memory registration is also enabled if
* registering memory will yield better performance than using multiple
* SGE entries.
*/

"Better performance" is relevant for mlx5 only, maybe others will use this max_.. field to overcome their HW limitations.

Thanks

Quoted 4 lines

Thanks,

Bart.

#4 lore raw

Re: [PATCH mlx5-next v2 1/3] net/mlx5: Expose optimal performance scatter entries capability

In reply to Leon Romanovsky
Quoted 11 lines · Leon Romanovsky

From: Yamin Friedman <[email protected]>

Expose maximum scatter entries per RDMA READ for optimal performance.

---
 include/linux/mlx5/mlx5_ifc.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

I'm taking this patch to mlx5-next, since it is not controversial.

Thanks

#15 lore raw
In reply to Leon Romanovsky
Quoted 11 lines · Leon Romanovsky

From: Leon Romanovsky <[email protected]>

Changelog
v1->v2: https://lore.kernel.org/linux-rdma/[email protected]
* Used Christoph's comment
* Change patch code flow to have one DMA_FROM_DEVICE check
v0->v1: https://lore.kernel.org/linux-rdma/[email protected]
* Reorganized patches to have IB/core changes separated from mlx5
* Moved SGL check before rdma_rw_force_mr
* Added and rephrased original comment.

Quoted 6 lines

-----------------------------------------------------------------------------
Hi,

This series from Yamin implements long standing "TODO" existed in rw.c.

Thanks

Jason, Doug

Do you expect anything from me in regards to this series?

Thanks

Quoted 16 lines

Yamin Friedman (3):
net/mlx5: Expose optimal performance scatter entries capability RDMA/rw: Support threshold for registration vs scattering to local
pages
RDMA/mlx5: Add capability for max sge to get optimized performance

 drivers/infiniband/core/rw.c      | 25 +++++++++++++++----------
 drivers/infiniband/hw/mlx5/main.c |  2 ++
 include/linux/mlx5/mlx5_ifc.h     |  2 +-
 include/rdma/ib_verbs.h           |  2 ++
 4 files changed, 20 insertions(+), 11 deletions(-)

--
2.20.1

#16 lore raw
In reply to Leon Romanovsky
Quoted 16 lines · Leon Romanovsky

From: Leon Romanovsky <[email protected]>

Changelog
v1->v2: https://lore.kernel.org/linux-rdma/[email protected]
* Used Christoph's comment
* Change patch code flow to have one DMA_FROM_DEVICE check
v0->v1: https://lore.kernel.org/linux-rdma/[email protected]
* Reorganized patches to have IB/core changes separated from mlx5
* Moved SGL check before rdma_rw_force_mr
* Added and rephrased original comment.

-----------------------------------------------------------------------------
Hi,

This series from Yamin implements long standing "TODO" existed in rw.c.

Quoted 7 lines

Thanks

Yamin Friedman (3):
net/mlx5: Expose optimal performance scatter entries capability RDMA/rw: Support threshold for registration vs scattering to local
pages
RDMA/mlx5: Add capability for max sge to get optimized performance

Applied to for-next with various fixes from Christoph and Bart

Thanks,
Jason