Introduce new mlx5 CQE format
Subsystems (5)
cover Introduce new mlx5 CQE format
From: Leon Romanovsky <[email protected]>
Introduce new internal to mlx5 CQE format - mini-CQE. It is a CQE in compressed form that holds data needed to extra a single full CQE.
It stride index, byte count and packet checksum.
Thanks
Yonatan Cohen (3):
net/mlx5: Exposing a new mini-CQE format
IB/mlx5: Refactor CQE compression response
IB/mlx5: Introduce a new mini-CQE format
drivers/infiniband/hw/mlx5/cq.c | 42 +++++++++++++++++++++++++++++---------- drivers/infiniband/hw/mlx5/main.c | 20 +++++++++++++------ include/linux/mlx5/mlx5_ifc.h | 3 ++- include/uapi/rdma/mlx5-abi.h | 2 +- 4 files changed, 49 insertions(+), 18 deletions(-)
PATCH 1/3 net/mlx5: Exposing a new mini-CQE format
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes byte-count, checksum and stride index.
--- include/linux/mlx5/mlx5_ifc.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
include/linux/mlx5/mlx5_ifc.h +2 -1
@@ -1160,7 +1160,8 @@ struct mlx5_ifc_cmd_hca_cap_bits {u8 flex_parser_protocols[0x20 ];u8 reserved_at_560[0x20 ]; -u8 reserved_at_580[0x3d ]; +u8 reserved_at_580[0x3c ]; +u8 mini_cqe_resp_stride_index[0x1 ];u8 cqe_128_always[0x1 ];u8 cqe_compression_128[0x1 ];u8 cqe_compression[0x1 ];
Signature · 1 line
2.14.3
Saeed Mahameed
AppliedRe: [PATCH mlx5-next 1/3] net/mlx5: Exposing a new mini-CQE format
Quoted 9 lines · Leon Romanovsky
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes byte-count, checksum and stride index.
Reviewed-by Yishai Hadas <[email protected]>Reviewed-by Guy Levi <[email protected]>Signed-off-by Yonatan Cohen <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>
Applied to mlx5-next.
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux.git mlx5- next
commit-id: ab741b2eed3e456cebd2240d4c9c6be003d5ae72
Thanks!
Re: [PATCH mlx5-next 1/3] net/mlx5: Exposing a new mini-CQE format
Quoted 12 lines · Saeed Mahameed
Quoted 9 lines · Leon Romanovsky
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes byte-count, checksum and stride index.
Reviewed-by Yishai Hadas <[email protected]>Reviewed-by Guy Levi <[email protected]>Signed-off-by Yonatan Cohen <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>
Quoted 6 lines
Applied to mlx5-next.
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux.gitmlx5- nextcommit-id:
ab741b2eed3e456cebd2240d4c9c6be003d5ae72
Thanks, everything is now merged as:
Jason
PATCH 2/3 IB/mlx5: Refactor CQE compression response
In reply to Leon RomanovskyFrom: Yonatan Cohen <[email protected]>
Refactor CQE compression response to be fully set only when it`s really supported. There is no change from user perspective because anyway resp.cqe_comp_caps.max_num was set to zero.
--- drivers/infiniband/hw/mlx5/main.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)
drivers/infiniband/hw/mlx5/main.c +10 -6
@@ -982,13 +982,17 @@ static int mlx5_ib_query_device(struct ib_device *ibdev, }if (field_avail (typeof (resp), cqe_comp_caps, uhw->outlen)) { - resp.cqe_comp_caps.max_num = -MLX5_CAP_GEN (dev->mdev, cqe_compression) ? -MLX5_CAP_GEN (dev->mdev, cqe_compression_max_num) :0 ; - resp.cqe_comp_caps.supported_format = -MLX5_IB_CQE_RES_FORMAT_HASH | -MLX5_IB_CQE_RES_FORMAT_CSUM ; resp.response_length +=sizeof (resp.cqe_comp_caps); + +if (MLX5_CAP_GEN (dev->mdev, cqe_compression)) { + resp.cqe_comp_caps.max_num = +MLX5_CAP_GEN (dev->mdev, + cqe_compression_max_num); + + resp.cqe_comp_caps.supported_format = +MLX5_IB_CQE_RES_FORMAT_HASH | +MLX5_IB_CQE_RES_FORMAT_CSUM ; + } } if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
Signature · 1 line
2.14.3
PATCH 3/3 IB/mlx5: Introduce a new mini-CQE format
In reply to Leon RomanovskyFrom: Yonatan Cohen <[email protected]>
The new mini-CQE format includes the stride index, byte count and packet checksum.
Stride index is needed for striding WQ feature.
This patch exposes this capability and enables its setting via mlx5 UHW data as part of query device and cq creation.
--- drivers/infiniband/hw/mlx5/cq.c | 42 +++++++++++++++++++++++++++++---------- drivers/infiniband/hw/mlx5/main.c | 4 ++++ include/uapi/rdma/mlx5-abi.h | 2 +- 3 files changed, 37 insertions(+), 11 deletions(-)
drivers/infiniband/hw/mlx5/cq.c +32 -10
@@ -751,6 +751,28 @@ static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,return 0 ; } +enum { +MLX5_CQE_RES_FORMAT_HASH =0 , +MLX5_CQE_RES_FORMAT_CSUM =1 , +MLX5_CQE_RES_FORMAT_CSUM_STRIDX =3 , +}; + +static int mini_cqe_res_format_to_hw (struct mlx5_ib_dev *dev,u8 format) +{ +switch (format) { +case MLX5_IB_CQE_RES_FORMAT_HASH : +return MLX5_CQE_RES_FORMAT_HASH ; +case MLX5_IB_CQE_RES_FORMAT_CSUM : +return MLX5_CQE_RES_FORMAT_CSUM ; +case MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX : +if (MLX5_CAP_GEN (dev->mdev, mini_cqe_resp_stride_index)) +return MLX5_CQE_RES_FORMAT_CSUM_STRIDX ; +return -EOPNOTSUPP ; +default : +return -EINVAL ; + } +} + static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, struct ib_ucontext *context, struct mlx5_ib_cq *cq, int entries, u32 **cqb,
@@ -816,6 +838,8 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, *index =to_mucontext (context)->bfregi.sys_pages[0 ];if (ucmd.cqe_comp_en ==1 ) { +int mini_cqe_format; + if (!((*cqe_size == 128 && MLX5_CAP_GEN(dev->mdev, cqe_compression_128)) || (*cqe_size == 64 &&
@@ -826,20 +850,18 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, goto err_cqb; } -if (unlikely (!ucmd.cqe_comp_res_format || - !(ucmd.cqe_comp_res_format < -MLX5_IB_CQE_RES_RESERVED ) || - (ucmd.cqe_comp_res_format & - (ucmd.cqe_comp_res_format -1 )))) { - err = -EOPNOTSUPP ; -mlx5_ib_warn (dev,"CQE compression res format %d is not supported!\n" , - ucmd.cqe_comp_res_format); + mini_cqe_format = +mini_cqe_res_format_to_hw (dev, + ucmd.cqe_comp_res_format); +if (mini_cqe_format <0 ) { + err = mini_cqe_format; +mlx5_ib_dbg (dev,"CQE compression res format %d error: %d\n" , + ucmd.cqe_comp_res_format, err); goto err_cqb; }MLX5_SET (cqc, cqc, cqe_comp_en,1 ); -MLX5_SET (cqc, cqc, mini_cqe_res_format, -ilog2 (ucmd.cqe_comp_res_format)); +MLX5_SET (cqc, cqc, mini_cqe_res_format, mini_cqe_format); }if (ucmd.flags &MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD ) {
drivers/infiniband/hw/mlx5/main.c +4 -0
@@ -992,6 +992,10 @@ static int mlx5_ib_query_device(struct ib_device *ibdev, resp.cqe_comp_caps.supported_format =MLX5_IB_CQE_RES_FORMAT_HASH |MLX5_IB_CQE_RES_FORMAT_CSUM ; + +if (MLX5_CAP_GEN (dev->mdev, mini_cqe_resp_stride_index)) + resp.cqe_comp_caps.supported_format |= +MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX ; } }
include/uapi/rdma/mlx5-abi.h +1 -1
@@ -166,7 +166,7 @@ struct mlx5_ib_rss_caps {enum mlx5_ib_cqe_comp_res_format {MLX5_IB_CQE_RES_FORMAT_HASH =1 <<0 ,MLX5_IB_CQE_RES_FORMAT_CSUM =1 <<1 , -MLX5_IB_CQE_RES_RESERVED =1 <<2 , +MLX5_IB_CQE_RES_FORMAT_CSUM_STRIDX =1 <<2 , };struct mlx5_ib_cqe_comp_caps {
Signature · 1 line
2.14.3
Re: [PATCH rdma-next 3/3] IB/mlx5: Introduce a new mini-CQE format
Quoted 21 lines · Leon Romanovsky
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes the stride index, byte count and packet checksum.
Stride index is needed for striding WQ feature.
This patch exposes this capability and enables its setting via mlx5 UHW data as part of query device and cq creation.Reviewed-by Yishai Hadas <[email protected]>Reviewed-by Guy Levi <[email protected]>Signed-off-by Yonatan Cohen <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>drivers/infiniband/hw/mlx5/cq.c | 42 +++++++++++++++++++++++++++++---------- drivers/infiniband/hw/mlx5/main.c | 4 ++++ include/uapi/rdma/mlx5-abi.h | 2 +- 3 files changed, 37 insertions(+), 11 deletions(-)drivers/infiniband/hw/mlx5/cq.c +1
-0+++ b/drivers/infiniband/hw/mlx5/cq.c
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 758
Quoted 8 lines
drivers/infiniband/hw/mlx5/cq.c +5
-0@@ -751,6 +751,28 @@ static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,return 0 ; } +enum { +MLX5_CQE_RES_FORMAT_HASH =0 , +MLX5_CQE_RES_FORMAT_CSUM =1 , +MLX5_CQE_RES_FORMAT_CSUM_STRIDX =3 , +};
What is this??
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 764
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c +5
-0+static int mini_cqe_res_format_to_hw (struct mlx5_ib_dev *dev,u8 format) +{ +switch (format) { +case MLX5_IB_CQE_RES_FORMAT_HASH : +return MLX5_CQE_RES_FORMAT_HASH ;
Used here..
drivers/infiniband/hw/mlx5/cq.c static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, · line 855
Quoted 3 lines
drivers/infiniband/hw/mlx5/cq.c +3
-0+ mini_cqe_format = +mini_cqe_res_format_to_hw (dev, + ucmd.cqe_comp_res_format);
And format comes from a ucmd, so that enum is upai.
Put it in the right place and put the right comment beside struct mlx5_ib_create_cq's cqe_comp_res_format..
And what is wrong with the user space patches? Where is the update to enum mlx5dv_cqe_comp_res_format ? And why is this wrong?
struct mlx5dv_cq_init_attr {
uint64_t comp_mask; /* Use enum mlx5dv_cq_init_attr_mask */ uint8_t cqe_comp_res_format; /* Use enum mlx5dv_cqe_comp_res_format */
^^^^^^^^^^^^^^^^^^^^^^^^^^
No, it isn't, and there isn't even an enum for it. Are you sure this is designed right? Looks pretty wrong to me.
Fix it all please, and you need to arrange things to share the uapi header with dv just like verbs is doing.
No more of this lax attitude toward uapi!
Jason
Re: [PATCH rdma-next 3/3] IB/mlx5: Introduce a new mini-CQE format
Quoted 25 lines · Jason Gunthorpe
Quoted 24 lines · Leon Romanovsky
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes the stride index, byte count and packet checksum.
Stride index is needed for striding WQ feature.
This patch exposes this capability and enables its setting via mlx5 UHW data as part of query device and cq creation.Reviewed-by Yishai Hadas <[email protected]>Reviewed-by Guy Levi <[email protected]>Signed-off-by Yonatan Cohen <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>drivers/infiniband/hw/mlx5/cq.c | 42 +++++++++++++++++++++++++++++---------- drivers/infiniband/hw/mlx5/main.c | 4 ++++ include/uapi/rdma/mlx5-abi.h | 2 +- 3 files changed, 37 insertions(+), 11 deletions(-)drivers/infiniband/hw/mlx5/cq.c +1
-0+++ b/drivers/infiniband/hw/mlx5/cq.c@@ -751,6 +751,28 @@ static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,return 0 ; }
Quoted 7 lines
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 758
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c +5
-0+enum { +MLX5_CQE_RES_FORMAT_HASH =0 , +MLX5_CQE_RES_FORMAT_CSUM =1 , +MLX5_CQE_RES_FORMAT_CSUM_STRIDX =3 , +};What is this??
Those are mlx5 device values not uapi.
Quoted 7 lines
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 764
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c +5
-0+static int mini_cqe_res_format_to_hw (struct mlx5_ib_dev *dev,u8 format) +{ +switch (format) { +case MLX5_IB_CQE_RES_FORMAT_HASH : +return MLX5_CQE_RES_FORMAT_HASH ;Used here..
This is some conversion between the uapi to the device values.
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, · line 855
Quoted 3 lines
drivers/infiniband/hw/mlx5/cq.c +3
-0+ mini_cqe_format = +mini_cqe_res_format_to_hw (dev, + ucmd.cqe_comp_res_format);And format comes from a ucmd, so that enum is upai.
Correct, see mlx5-abi.h as part of this patch.
Quoted 6 lines
Put it in the right place and put the right comment beside struct mlx5_ib_create_cq's
cqe_comp_res_format..And what is wrong with the user space patches? Where is the update to enum
mlx5dv_cqe_comp_res_format? And why is this wrong?
See the first patch from below PR [1], it brings the new enum value to the user area as part of kernel-headers/rdma/mlx5-abi.h.
[1] https://github.com/linux-rdma/rdma-core/pull/337
Quoted 5 lines
struct
mlx5dv_cq_init_attr{
uint64_tcomp_mask; /* Use enummlx5dv_cq_init_attr_mask*/uint8_tcqe_comp_res_format; /* Use enummlx5dv_cqe_comp_res_format*/
^^^^^^^^^^^^^^^^^^^^^^^^^^
The user space uses the DV prefix (e.g. MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX), no change from previous flags around enum mlx5dv_cqe_comp_res_format.
Re: [PATCH rdma-next 3/3] IB/mlx5: Introduce a new mini-CQE format
Quoted 65 lines · Yishai Hadas
Quoted 18 lines · Jason Gunthorpe
Quoted 17 lines · Leon Romanovsky
From: Yonatan Cohen <[email protected]>
The new mini-CQE format includes the stride index, byte count and packet checksum.
Stride index is needed for striding WQ feature.
This patch exposes this capability and enables its setting via mlx5 UHW data as part of query device and cq creation.Reviewed-by Yishai Hadas <[email protected]>Reviewed-by Guy Levi <[email protected]>Signed-off-by Yonatan Cohen <[email protected]>Signed-off-by Leon Romanovsky <[email protected]>drivers/infiniband/hw/mlx5/cq.c | 42 +++++++++++++++++++++++++++++---------- drivers/infiniband/hw/mlx5/main.c | 4 ++++ include/uapi/rdma/mlx5-abi.h | 2 +- 3 files changed, 37 insertions(+), 11 deletions(-)Quoted 13 lines
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 757
Quoted 11 lines
drivers/infiniband/hw/mlx5/cq.c +6
-0+++ b/drivers/infiniband/hw/mlx5/cq.c@@ -751,6 +751,28 @@ static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev,return 0 ; } +enum { +MLX5_CQE_RES_FORMAT_HASH =0 , +MLX5_CQE_RES_FORMAT_CSUM =1 , +MLX5_CQE_RES_FORMAT_CSUM_STRIDX =3 , +};What is this??
Those are mlx5 device values not uapi.
Quoted 7 lines
drivers/infiniband/hw/mlx5/cq.c static int alloc_cq_frag_buf(struct mlx5_ib_dev *dev, · line 764
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c +5
-0+static int mini_cqe_res_format_to_hw (struct mlx5_ib_dev *dev,u8 format) +{ +switch (format) { +case MLX5_IB_CQE_RES_FORMAT_HASH : +return MLX5_CQE_RES_FORMAT_HASH ;Used here..
This is some conversion between the uapi to the device values.
Quoted 5 lines
drivers/infiniband/hw/mlx5/cq.c static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata, · line 855
Quoted 3 lines
drivers/infiniband/hw/mlx5/cq.c +3
-0+ mini_cqe_format = +mini_cqe_res_format_to_hw (dev, + ucmd.cqe_comp_res_format);And format comes from a ucmd, so that enum is upai.
Correct, see
mlx5-abi.has part of this patch.Quoted 6 lines
Put it in the right place and put the right comment beside struct mlx5_ib_create_cq's
cqe_comp_res_format..And what is wrong with the user space patches? Where is the update to enum
mlx5dv_cqe_comp_res_format? And why is this wrong?See the first patch from below PR [1], it brings the new enum value to the user area as part of
kernel-headers/rdma/mlx5-abi.h.
Quoted 8 lines
Quoted 5 lines
struct
mlx5dv_cq_init_attr{
uint64_tcomp_mask; /* Use enummlx5dv_cq_init_attr_mask*/uint8_tcqe_comp_res_format; /* Use enummlx5dv_cqe_comp_res_format*/
^^^^^^^^^^^^^^^^^^^^^^^^^^The user space uses the DV prefix (e.g.
MLX5DV_CQE_RES_FORMAT_CSUM_STRIDX), no change from previous flags around enummlx5dv_cqe_comp_res_format.
This still needs eventual cleaning up in the verbs_abi.h way.
But OK, this is not what I thought.
Jason
Jason Gunthorpe
AppliedQuoted 5 lines · Leon Romanovsky
From: Leon Romanovsky <[email protected]>
Introduce new internal to mlx5 CQE format - mini-CQE. It is a CQE in compressed form that holds data needed to extra a single full CQE.
Quoted 8 lines
It stride index, byte count and packet checksum.
Thanks
Yonatan Cohen (3):
net/mlx5: Exposing a new mini-CQE format
IB/mlx5: Refactor CQE compression response
IB/mlx5: Introduce a new mini-CQE format
Applied to for-next.
Generally taking new uapi patches that are first the list should have a few weeks of comment period, but since this is just adding a new bit to an existing driver private api it seems OK to go this merge window.
Thanks,
Jason