RDMA/efa: Add rdma write capability to device caps
Subsystems (2)
PATCH 1/1 RDMA/efa: Add rdma write capability to device caps
From: Yonatan Nachum <[email protected]>
Add rdma write capability that is propagated from the device to rdma-core.
Enable MR creation with remote write permissions according to this device capability.
--- .../infiniband/hw/efa/efa_admin_cmds_defs.h | 12 ++++-- drivers/infiniband/hw/efa/efa_io_defs.h | 42 +++++++++++++------ drivers/infiniband/hw/efa/efa_verbs.c | 6 ++- include/uapi/rdma/efa-abi.h | 1 + 4 files changed, 44 insertions(+), 17 deletions(-)
drivers/infiniband/hw/efa/efa_admin_cmds_defs.h +9 -3
@@ -376,7 +376,9 @@ struct efa_admin_reg_mr_cmd {* 0 : local_write_enable - Local write permissions: * must be set for RQ buffers and buffers posted for * RDMA Read requests -* 1 : reserved1 - MBZ +* 1 : remote_write_enable - Remote write +* permissions: must be set to enable RDMA write to +* the region * 2 : remote_read_enable - Remote read permissions: * must be set to enable RDMA read from the region * 7:3 : reserved2 - MBZ
@@ -620,7 +622,9 @@ struct efa_admin_feature_device_attr_desc {* modify QP command * 2 : data_polling_128 - If set, 128 bytes data * polling is supported -* 31:3 : reserved - MBZ +* 3 : rdma_write - If set, RDMA Write is supported +* on TX queues +* 31:4 : reserved - MBZ */ u32 device_caps;
@@ -674,7 +678,7 @@ struct efa_admin_feature_queue_attr_desc {/* The maximum size of LLQ in bytes */ u32 max_llq_size; -/* Maximum number of SGEs for a single RDMA read WQE */ +/* Maximum number of SGEs for a single RDMA read/write WQE */ u16 max_wr_rdma_sges;/*
@@ -979,6 +983,7 @@ struct efa_admin_host_info {#define EFA_ADMIN_REG_MR_CMD_PHYS_PAGE_SIZE_SHIFT_MASK GENMASK(4, 0)#define EFA_ADMIN_REG_MR_CMD_MEM_ADDR_PHY_MODE_EN_MASK BIT(7)#define EFA_ADMIN_REG_MR_CMD_LOCAL_WRITE_ENABLE_MASK BIT(0) +#define EFA_ADMIN_REG_MR_CMD_REMOTE_WRITE_ENABLE_MASK BIT(1)#define EFA_ADMIN_REG_MR_CMD_REMOTE_READ_ENABLE_MASK BIT(2)/* create_cq_cmd */
@@ -994,6 +999,7 @@ struct efa_admin_host_info {#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_READ_MASK BIT(0)#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RNR_RETRY_MASK BIT(1)#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_DATA_POLLING_128_MASK BIT(2) +#define EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_RDMA_WRITE_MASK BIT(3)/* create_eq_cmd */ #define EFA_ADMIN_CREATE_EQ_CMD_ENTRY_SIZE_WORDS_MASK GENMASK(4, 0)
drivers/infiniband/hw/efa/efa_io_defs.h +29 -13
97 lines not on this page · raw message
drivers/infiniband/hw/efa/efa_verbs.c +5 -1
@@ -253,6 +253,9 @@ int efa_query_device(struct ib_device *ibdev,if (EFA_DEV_CAP (dev,DATA_POLLING_128 )) resp.device_caps |=EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 ; +if (EFA_DEV_CAP (dev,RDMA_WRITE )) + resp.device_caps |=EFA_QUERY_DEVICE_CAPS_RDMA_WRITE ; +if (dev->neqs) resp.device_caps |=EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS ;
@@ -1571,7 +1574,8 @@ static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags, supp_access_flags =IB_ACCESS_LOCAL_WRITE | - (EFA_DEV_CAP (dev,RDMA_READ ) ?IB_ACCESS_REMOTE_READ :0 ); + (EFA_DEV_CAP (dev,RDMA_READ ) ?IB_ACCESS_REMOTE_READ :0 ) | + (EFA_DEV_CAP (dev,RDMA_WRITE ) ?IB_ACCESS_REMOTE_WRITE :0 ); access_flags &= ~IB_ACCESS_OPTIONAL ;if (access_flags & ~supp_access_flags) {
include/uapi/rdma/efa-abi.h +1 -0
@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5, }; struct efa_ibv_ex_query_device_resp {
Signature · 1 line
2.39.2
Quoted 16 lines · On Tue, Apr 04, 2023 at 03:43:13PM +0000, [email protected] wrote:
From: Yonatan Nachum <[email protected]>
Add rdma write capability that is propagated from the device to rdma-core.
Enable MR creation with remote write permissions according to this device capability.Reviewed-by Firas Jahjah <[email protected]>Reviewed-by Michael Margolin <[email protected]>Signed-off-by Yonatan Nachum <[email protected]>--- .../infiniband/hw/efa/efa_admin_cmds_defs.h | 12 ++++-- drivers/infiniband/hw/efa/efa_io_defs.h | 42 +++++++++++++------ drivers/infiniband/hw/efa/efa_verbs.c | 6 ++- include/uapi/rdma/efa-abi.h | 1 + 4 files changed, 44 insertions(+), 17 deletions(-)
<...>
Quoted 30 lines
#endif /* _EFA_IO_H_ */
drivers/infiniband/hw/efa/efa_verbs.c +5
-1@@ -253,6 +253,9 @@ int efa_query_device(struct ib_device *ibdev,if (EFA_DEV_CAP (dev,DATA_POLLING_128 )) resp.device_caps |=EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 ; +if (EFA_DEV_CAP (dev,RDMA_WRITE )) + resp.device_caps |=EFA_QUERY_DEVICE_CAPS_RDMA_WRITE ; +if (dev->neqs) resp.device_caps |=EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS ;@@ -1571,7 +1574,8 @@ static struct efa_mr *efa_alloc_mr(struct ib_pd *ibpd, int access_flags, supp_access_flags =IB_ACCESS_LOCAL_WRITE | - (EFA_DEV_CAP (dev,RDMA_READ ) ?IB_ACCESS_REMOTE_READ :0 ); + (EFA_DEV_CAP (dev,RDMA_READ ) ?IB_ACCESS_REMOTE_READ :0 ) | + (EFA_DEV_CAP (dev,RDMA_WRITE ) ?IB_ACCESS_REMOTE_WRITE :0 ); access_flags &= ~IB_ACCESS_OPTIONAL ;if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +0
-0
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 4 lines
include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,
Why do you need special device capability while all rdma-core users set IBV_ACCESS_REMOTE_WRITE anyway without relying on anything from providers?
Thanks
Quoted 6 lines
include/uapi/rdma/efa-abi.h +0
-0};struct efa_ibv_ex_query_device_resp {Signature · 1 line
2.39.2
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?
Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Thanks.
Quoted 18 lines · Yonatan
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?
Quoted 3 lines
Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
Quoted 2 lines
Thanks.
Quoted 18 lines
Quoted 18 lines
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?
Quoted 9 lines
Quoted 3 lines
Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
Thanks
Quoted 22 lines · Yonatan
Quoted 22 lines
Quoted 21 lines
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Quoted 7 lines
Quoted 5 lines
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
AFAIK, RDMA write is different here as fallback is performed in the kernel and not in the rdma-core provider. So why should EFA be different here?
BTW, Please fix your mailer to break lines, so we will be able to reply on specific sentence with less effort.
Thanks
Quoted 2 lines
Thanks
Quoted 30 lines
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Quoted 7 lines
Quoted 5 lines
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
Quoted 7 lines
AFAIK, RDMA write is different here as fallback is performed in the kernel and not in the rdma-core provider. So why should EFA be different here?
BTW, Please fix your mailer to break lines, so we will be able to reply on specific sentence with less effort.
Thanks
Can you please elaborate more on the fallback performed in the kernel? What kind of fallback being performed? Is it in create MR/QP? Does the fallback happens when providing unsupported write capability and to what it fallback to?
Thanks
Quoted 36 lines · Yonatan
Quoted 36 lines
Quoted 22 lines
Quoted 22 lines
Quoted 21 lines
Quoted 18 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 13 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Quoted 7 lines
Quoted 5 lines
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
AFAIK, RDMA write is different here as fallback is performed in the kernel and not in the rdma-core provider. So why should EFA be different here?
BTW, Please fix your mailer to break lines, so we will be able to reply on specific sentence with less effort.
Quoted 6 lines
Quoted 1 line
Thanks
Can you please elaborate more on the fallback performed in the kernel? What kind of fallback being performed? Is it in create MR/QP? Does the fallback happens when providing unsupported write capability and to what it fallback to?
OK, looked again, "Fallback" was in my imagination, sorry about that.
But my main question is continued to be, how other vendors which support RDMA write work without capability?
Thanks
Quoted 2 lines
Thanks
Quoted 42 lines
Quoted 35 lines
Quoted 35 lines
Quoted 21 lines
Quoted 21 lines
Quoted 20 lines
Quoted 17 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 12 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Quoted 7 lines
Quoted 5 lines
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
AFAIK, RDMA write is different here as fallback is performed in the kernel and not in the rdma-core provider. So why should EFA be different here?
BTW, Please fix your mailer to break lines, so we will be able to reply on specific sentence with less effort.
Quoted 6 lines
Quoted 1 line
Thanks
Can you please elaborate more on the fallback performed in the kernel? What kind of fallback being performed? Is it in create MR/QP? Does the fallback happens when providing unsupported write capability and to what it fallback to?
Quoted 6 lines
OK, looked again, "Fallback" was in my imagination, sorry about that.
But my main question is continued to be, how other vendors which support RDMA write work without capability?
Thanks
Vendors that always support RDMA write don’t need a query for this capability. Some EFA devices don’t support write capability so we provide the ability to query the device to know if write is supported.
It is like mlx5 support query capabilities through direct verb.
Thanks
Quoted 47 lines
Quoted 47 lines
Quoted 35 lines
Quoted 35 lines
Quoted 21 lines
Quoted 21 lines
Quoted 20 lines
Quoted 17 lines
include/uapi/rdma/efa-abi.h enum { · line 124
Quoted 12 lines
access_flags &= ~IB_ACCESS_OPTIONAL; if (access_flags & ~supp_access_flags) {include/uapi/rdma/efa-abi.h +1
-0@@ -121,6 +121,7 @@ enum {EFA_QUERY_DEVICE_CAPS_CQ_NOTIFICATIONS =1 <<2 ,EFA_QUERY_DEVICE_CAPS_CQ_WITH_SGID =1 <<3 ,EFA_QUERY_DEVICE_CAPS_DATA_POLLING_128 =1 <<4 , + EFA_QUERY_DEVICE_CAPS_RDMA_WRITE = 1 << 5,Why do you need special device capability while all rdma-core users set
IBV_ACCESS_REMOTE_WRITEanyway without relying on anything from providers?Quoted 1 line
Thanks
We need to query the device because not every device supprort the same RDMA capabilities. Upper layers in the SW stack needs this supported flags to indicate which flows they can use. In addition this is identical to the existing RDMA read support in our code.
Quoted 7 lines
Quoted 5 lines
Nice, but it doesn't answer my question. Please pay attention to the second part of my question "while all rdma-core ....".
Thanks
There are rdma-core users that doesn’t fail on unsupported features but fallback to supported ones. One example is Libfabric EFA provider that emulates RDMA write by read if device write isn’t supported but there are other similar examples. Correct way doing this in user code is by querying rdma-core for device supported capabilities, then selecting a suitable work flow. This is why existing and the additional capability bits are required.
AFAIK, RDMA write is different here as fallback is performed in the kernel and not in the rdma-core provider. So why should EFA be different here?
BTW, Please fix your mailer to break lines, so we will be able to reply on specific sentence with less effort.
Quoted 6 lines
Quoted 1 line
Thanks
Can you please elaborate more on the fallback performed in the kernel? What kind of fallback being performed? Is it in create MR/QP? Does the fallback happens when providing unsupported write capability and to what it fallback to?
OK, looked again, "Fallback" was in my imagination, sorry about that.
But my main question is continued to be, how other vendors which support RDMA write work without capability?
Quoted 8 lines
Quoted 1 line
Thanks
Vendors that always support RDMA write don’t need a query for this capability. Some EFA devices don’t support write capability so we provide the ability to query the device to know if write is supported.
It is like mlx5 support query capabilities through direct verb.Thanks
Hello,
Kind reminder for this patch.
Thanks
Quoted 10 lines · On Tue, Apr 04, 2023 at 03:43:13PM +0000, [email protected] wrote:
From: Yonatan Nachum <[email protected]>
Add rdma write capability that is propagated from the device to rdma-core.
Enable MR creation with remote write permissions according to this device capability.Reviewed-by Firas Jahjah <[email protected]>Reviewed-by Michael Margolin <[email protected]>Signed-off-by Yonatan Nachum <[email protected]>
Quoted 6 lines
--- .../infiniband/hw/efa/efa_admin_cmds_defs.h | 12 ++++-- drivers/infiniband/hw/efa/efa_io_defs.h | 42 +++++++++++++------ drivers/infiniband/hw/efa/efa_verbs.c | 6 ++- include/uapi/rdma/efa-abi.h | 1 + 4 files changed, 44 insertions(+), 17 deletions(-)
Applied to for-next, thanks
Jason