mlx5dv_wr_mkey_configure(3) | mlx5dv_wr_mkey_configure(3) |
mlx5dv_wr_mkey_configure - Create a work request to configure an MKEY
mlx5dv_wr_set_mkey_access_flags - Set the memory protection attributes for an MKEY
mlx5dv_wr_set_mkey_layout_list - Set a memory layout for an MKEY based on SGE list
mlx5dv_wr_set_mkey_layout_interleaved - Set an interleaved memory layout for an MKEY
#include <infiniband/mlx5dv.h> static inline void mlx5dv_wr_mkey_configure(struct mlx5dv_qp_ex *mqp,
struct mlx5dv_mkey *mkey,
uint8_t num_setters,
struct mlx5dv_mkey_conf_attr *attr); static inline void mlx5dv_wr_set_mkey_access_flags(struct mlx5dv_qp_ex *mqp,
uint32_t access_flags); static inline void mlx5dv_wr_set_mkey_layout_list(struct mlx5dv_qp_ex *mqp,
uint16_t num_sges,
const struct ibv_sge *sge); static inline void mlx5dv_wr_set_mkey_layout_interleaved(struct mlx5dv_qp_ex *mqp,
uint32_t repeat_count,
uint16_t num_interleaved,
const struct mlx5dv_mr_interleaved *data);
The MLX5DV MKEY configure API and the related setters (mlx5dv_wr_set_mkey*) are an extension of IBV work request API (ibv_wr*) with specific features for MLX5DV MKEY.
MKEYs allow creation of virtually-contiguous address spaces out of non-contiguous chunks of memory regions already registered with the hardware. Additionally it provides access to some advanced hardware offload features, e.g. signature offload.
These APIs are intended to be used to access additional functionality beyond what is provided by mlx5dv_wr_mr_list() and mlx5dv_wr_mr_interleaved(). The MKEY features can be optionally enabled using the mkey configure setters. It allows using different features in the same MKEY.
To use these APIs a QP must be created using mlx5dv_create_qp(3) which allows setting the MLX5DV_QP_EX_WITH_MKEY_CONFIGURE in send_ops_flags.
The MKEY configuration work request is created by calling mlx5dv_wr_mkey_configure(), a WR builder function, followed by required setter functions. num_setters is a number of required setters for the WR. All setters are optional. num_setters can be zero to apply attr only. Each setter can be called only once per the WR builder.
The WR configures mkey and applies attr of the builder function and setter functions’ arguments for it. If mkey is already configured the WR overrides some mkey properties depends on builder and setter functions’ arguments (see details in setters’ description). To clear configuration of mkey, use ibv_post_send() with IBV_WR_LOCAL_INV opcode or ibv_wr_local_inv().
Current implementation requires the IBV_SEND_INLINE option to be set in wr_flags field of ibv_qp_ex structure prior to builder function call. Non-inline payload is currently not supported by this API. Please note that inlining here is done for MKEY configuration data, not for user data referenced by data layouts.
Once MKEY is configured, it may be used in subsequent work requests (SEND, RDMA_READ, RDMA_WRITE, etc). If these work requests are posted on the same QP, there is no need to wait for completion of MKEY configuration work request. They can be posted immediately after the last setter (or builder if no setters). Usually there is no need to even request a completion for MKEY configuration work request.
If completion is requested for MKEY configuration work request it will be delivered with the IBV_WC_DRIVER1 opcode.
MKEY configuration attributes are provided in mlx5dv_mkey_conf_attr structure.
struct mlx5dv_mkey_conf_attr {
uint32_t conf_flags;
uint64_t comp_mask; };
Data layout setters define how data referenced by the MKEY will be scattered/gathered in the memory. In order to use MKEY with RDMA operations it must be configured with a layout.
Not more than one data layout setter may follow builder function. Layout can be updated in the next calls to builder function.
When MKEY is used in RDMA operations, it should be used in a zero-based mode, i.e. the addr field in ibv_sge structure is an offset in the total data.
Default WQE size can fit only 4 SGE entries. To allow more the QP should be created with a larger WQE size that may fit it. This should be done using the max_inline_data attribute of struct ibv_qp_cap upon QP creation.
Default WQE size can fit only 3 interleaved entries. To allow more the QP should be created with a larger WQE size that may fit it. This should be done using the max_inline_data attribute of struct ibv_qp_cap upon QP creation.
As one entry will be consumed for strided header, the MKEY should be created with one more entry than the required num_interleaved.
Interleaved data layout is described by mlx5dv_mr_interleaved structure.
struct mlx5dv_mr_interleaved {
uint64_t addr;
uint32_t bytes_count;
uint32_t bytes_skip;
uint32_t lkey; };
The signature attributes of the MKEY allow adding/modifying/stripping/validating integrity fields when transmitting data from memory to network and when receiving data from network to memory.
Use the signature setters to set/update the signature attributes of the MKEY. To reset the signature attributes without invalidating the MKEY, use the MLX5DV_MKEY_CONF_FLAG_RESET_SIG_ATTR flag.
The crypto attributes of the MKey allow encryption and decryption of transmitted data from memory to network and when receiving data from network to memory.
Use the crypto setter to set/update the crypto attributes of the MKey. When the MKey is created with MLX5DV_MKEY_INIT_ATTR_FLAGS_CRYPTO it must be configured with crypto attributes before the MKey can be used.
Code below creates a QP with MKEY configure operation support and an indirect mkey.
/* Create QP with MKEY configure support */ struct ibv_qp_init_attr_ex attr_ex = {}; attr_ex.comp_mask |= IBV_QP_INIT_ATTR_SEND_OPS_FLAGS; attr_ex.send_ops_flags |= IBV_QP_EX_WITH_RDMA_WRITE; struct mlx5dv_qp_init_attr attr_dv = {}; attr_dv.comp_mask |= MLX5DV_QP_INIT_ATTR_MASK_SEND_OPS_FLAGS; attr_dv.send_ops_flags = MLX5DV_QP_EX_WITH_MKEY_CONFIGURE; ibv_qp *qp = mlx5dv_create_qp(ctx, attr_ex, attr_dv); ibv_qp_ex *qpx = ibv_qp_to_qp_ex(qp); mlx5dv_qp_ex *mqpx = mlx5dv_qp_ex_from_ibv_qp_ex(qpx); mkey_attr.create_flags = MLX5DV_MKEY_INIT_ATTR_FLAGS_INDIRECT; struct mlx5dv_mkey *mkey = mlx5dv_create_mkey(&mkey_attr);
Code below configures an MKEY which allows remote access for read and write and is based on SGE list layout with two entries. When this MKEY is used in RDMA write operation, data will be scattered between two memory regions. The first 64 bytes will go to memory referenced by mr1. The next 4096 bytes will go to memory referenced by mr2.
ibv_wr_start(qpx); qpx->wr_id = my_wr_id_1; qpx->wr_flags = IBV_SEND_INLINE; struct mlx5dv_mkey_conf_attr mkey_attr = {}; mlx5dv_wr_mkey_configure(mqpx, mkey, 2, &mkey_attr); mlx5dv_wr_set_mkey_access_flags(mqpx, IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); struct ibv_sge sgl[2]; sgl[0].addr = mr1->addr; sgl[0].length = 64; sgl[0].lkey = mr1->lkey; sgl[1].addr = mr2->addr; sgl[1].length = 4096; sgl[1].lkey = mr2->lkey; mlx5dv_wr_set_mkey_layout_list(mqpx, 2, sgl); ret = ibv_wr_complete(qpx);
Code below configures an MKEY which allows remote access for read and write and is based on interleaved data layout with two entries and repeat count of two. When this MKEY is used in RDMA write operation, data will be scattered between two memory regions. The first 512 bytes will go to memory referenced by mr1 at offset 0. The next 8 bytes will go to memory referenced by mr2 at offset 0. The next 512 bytes will go to memory referenced by mr1 at offset 516. The next 8 bytes will go to memory referenced by mr2 at offset 8.
ibv_wr_start(qpx); qpx->wr_id = my_wr_id_1; qpx->wr_flags = IBV_SEND_INLINE; struct mlx5dv_mkey_conf_attr mkey_attr = {}; mlx5dv_wr_mkey_configure(mqpx, mkey, 2, &mkey_attr); mlx5dv_wr_set_mkey_access_flags(mqpx, IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE); struct mlx5dv_mr_interleaved data[2]; data[0].addr = mr1->addr; data[0].bytes_count = 512; data[0].bytes_skip = 4; data[0].lkey = mr1->lkey; data[1].addr = mr2->addr; data[1].bytes_count = 8; data[1].bytes_skip = 0; data[1].lkey = mr2->lkey; mlx5dv_wr_set_mkey_layout_interleaved(mqpx, 2, 2, &data); ret = ibv_wr_complete(qpx);
A DEVX context should be opened by using mlx5dv_open_device(3).
mlx5dv_create_mkey(3), mlx5dv_create_qp(3), mlx5dv_wr_set_mkey_sig_block(3), mlx5dv_wr_set_mkey_crypto(3)
Oren Duer <oren@nvidia.com>
Sergey Gorenko <sergeygo@nvidia.com>
Evgenii Kochetov <evgeniik@nvidia.com>