From c772a2c690182410642ead740f7a84b3a7544b2b Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:10 +0300 Subject: net/mlx5: Add IFC related stuff for data direct Add IFC related stuff for data direct. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/82da7f578a567909bb5858a64ba844fe4cc298fa.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- include/linux/mlx5/mlx5_ifc.h | 51 ++++++++++++++++++++++++++++++++++++++----- 1 file changed, 46 insertions(+), 5 deletions(-) (limited to 'include') diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h index cab228cf51c6..970c9d8473ef 100644 --- a/include/linux/mlx5/mlx5_ifc.h +++ b/include/linux/mlx5/mlx5_ifc.h @@ -313,6 +313,7 @@ enum { MLX5_CMD_OP_MODIFY_VHCA_STATE = 0xb0e, MLX5_CMD_OP_SYNC_CRYPTO = 0xb12, MLX5_CMD_OP_ALLOW_OTHER_VHCA_ACCESS = 0xb16, + MLX5_CMD_OPCODE_QUERY_VUID = 0xb22, MLX5_CMD_OP_MAX }; @@ -1885,7 +1886,8 @@ struct mlx5_ifc_cmd_hca_cap_bits { u8 reserved_at_5a0[0x10]; u8 enhanced_cqe_compression[0x1]; - u8 reserved_at_5b1[0x2]; + u8 reserved_at_5b1[0x1]; + u8 crossing_vhca_mkey[0x1]; u8 log_max_dek[0x5]; u8 reserved_at_5b8[0x4]; u8 mini_cqe_resp_stride_index[0x1]; @@ -1954,7 +1956,9 @@ struct mlx5_ifc_cmd_hca_cap_bits { u8 dynamic_msix_table_size[0xc]; u8 reserved_at_740[0xc]; u8 min_dynamic_vf_msix_table_size[0x4]; - u8 reserved_at_750[0x4]; + u8 reserved_at_750[0x2]; + u8 data_direct[0x1]; + u8 reserved_at_753[0x1]; u8 max_dynamic_vf_msix_table_size[0xc]; u8 reserved_at_760[0x3]; @@ -1982,7 +1986,9 @@ struct mlx5_ifc_cmd_hca_cap_2_bits { u8 reserved_at_0[0x80]; u8 migratable[0x1]; - u8 reserved_at_81[0x1f]; + u8 reserved_at_81[0x11]; + u8 query_vuid[0x1]; + u8 reserved_at_93[0xd]; u8 max_reformat_insert_size[0x8]; u8 max_reformat_insert_offset[0x8]; @@ -4154,6 +4160,7 @@ enum { MLX5_MKC_ACCESS_MODE_KSM = 0x3, MLX5_MKC_ACCESS_MODE_SW_ICM = 0x4, MLX5_MKC_ACCESS_MODE_MEMIC = 0x5, + MLX5_MKC_ACCESS_MODE_CROSSING = 0x6, }; struct mlx5_ifc_mkc_bits { @@ -4196,7 +4203,10 @@ struct mlx5_ifc_mkc_bits { u8 bsf_octword_size[0x20]; - u8 reserved_at_120[0x80]; + u8 reserved_at_120[0x60]; + + u8 crossing_target_vhca_id[0x10]; + u8 reserved_at_190[0x10]; u8 translations_octword_size[0x20]; @@ -5124,6 +5134,36 @@ struct mlx5_ifc_query_vport_state_out_bits { u8 state[0x4]; }; +struct mlx5_ifc_array1024_auto_bits { + u8 array1024_auto[32][0x20]; +}; + +struct mlx5_ifc_query_vuid_in_bits { + u8 opcode[0x10]; + u8 uid[0x10]; + + u8 reserved_at_20[0x40]; + + u8 query_vfs_vuid[0x1]; + u8 data_direct[0x1]; + u8 reserved_at_62[0xe]; + u8 vhca_id[0x10]; +}; + +struct mlx5_ifc_query_vuid_out_bits { + u8 status[0x8]; + u8 reserved_at_8[0x18]; + + u8 syndrome[0x20]; + + u8 reserved_at_40[0x1a0]; + + u8 reserved_at_1e0[0x10]; + u8 num_of_entries[0x10]; + + struct mlx5_ifc_array1024_auto_bits vuid[]; +}; + enum { MLX5_VPORT_STATE_OP_MOD_VNIC_VPORT = 0x0, MLX5_VPORT_STATE_OP_MOD_ESW_VPORT = 0x1, @@ -8989,7 +9029,8 @@ struct mlx5_ifc_create_mkey_in_bits { u8 pg_access[0x1]; u8 mkey_umem_valid[0x1]; - u8 reserved_at_62[0x1e]; + u8 data_direct[0x1]; + u8 reserved_at_63[0x1d]; struct mlx5_ifc_mkc_bits memory_key_mkey_entry; -- cgit From 682358fd35dece838e6ae2d9d6a69fc0b9a9d411 Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:13 +0300 Subject: RDMA/umem: Add support for creating pinned DMABUF umem with a given dma device Add support for creating pinned DMABUF umem with a specified DMA device instead of the DMA device of the given IB device. This API will be utilized in the upcoming patches of the series when multiple path DMAs are implemented. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/038aad36a43797e5591b20ba81051fc5758124f9.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/umem_dmabuf.c | 45 ++++++++++++++++++++++++++--------- include/rdma/ib_umem.h | 15 ++++++++++++ 2 files changed, 49 insertions(+), 11 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/core/umem_dmabuf.c b/drivers/infiniband/core/umem_dmabuf.c index 39357dc2d229..726a09786547 100644 --- a/drivers/infiniband/core/umem_dmabuf.c +++ b/drivers/infiniband/core/umem_dmabuf.c @@ -110,10 +110,12 @@ void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf) } EXPORT_SYMBOL(ib_umem_dmabuf_unmap_pages); -struct ib_umem_dmabuf *ib_umem_dmabuf_get(struct ib_device *device, - unsigned long offset, size_t size, - int fd, int access, - const struct dma_buf_attach_ops *ops) +static struct ib_umem_dmabuf * +ib_umem_dmabuf_get_with_dma_device(struct ib_device *device, + struct device *dma_device, + unsigned long offset, size_t size, + int fd, int access, + const struct dma_buf_attach_ops *ops) { struct dma_buf *dmabuf; struct ib_umem_dmabuf *umem_dmabuf; @@ -152,7 +154,7 @@ struct ib_umem_dmabuf *ib_umem_dmabuf_get(struct ib_device *device, umem_dmabuf->attach = dma_buf_dynamic_attach( dmabuf, - device->dma_device, + dma_device, ops, umem_dmabuf); if (IS_ERR(umem_dmabuf->attach)) { @@ -168,6 +170,15 @@ out_release_dmabuf: dma_buf_put(dmabuf); return ret; } + +struct ib_umem_dmabuf *ib_umem_dmabuf_get(struct ib_device *device, + unsigned long offset, size_t size, + int fd, int access, + const struct dma_buf_attach_ops *ops) +{ + return ib_umem_dmabuf_get_with_dma_device(device, device->dma_device, + offset, size, fd, access, ops); +} EXPORT_SYMBOL(ib_umem_dmabuf_get); static void @@ -184,16 +195,18 @@ static struct dma_buf_attach_ops ib_umem_dmabuf_attach_pinned_ops = { .move_notify = ib_umem_dmabuf_unsupported_move_notify, }; -struct ib_umem_dmabuf *ib_umem_dmabuf_get_pinned(struct ib_device *device, - unsigned long offset, - size_t size, int fd, - int access) +struct ib_umem_dmabuf * +ib_umem_dmabuf_get_pinned_with_dma_device(struct ib_device *device, + struct device *dma_device, + unsigned long offset, size_t size, + int fd, int access) { struct ib_umem_dmabuf *umem_dmabuf; int err; - umem_dmabuf = ib_umem_dmabuf_get(device, offset, size, fd, access, - &ib_umem_dmabuf_attach_pinned_ops); + umem_dmabuf = ib_umem_dmabuf_get_with_dma_device(device, dma_device, offset, + size, fd, access, + &ib_umem_dmabuf_attach_pinned_ops); if (IS_ERR(umem_dmabuf)) return umem_dmabuf; @@ -217,6 +230,16 @@ err_release: ib_umem_release(&umem_dmabuf->umem); return ERR_PTR(err); } +EXPORT_SYMBOL(ib_umem_dmabuf_get_pinned_with_dma_device); + +struct ib_umem_dmabuf *ib_umem_dmabuf_get_pinned(struct ib_device *device, + unsigned long offset, + size_t size, int fd, + int access) +{ + return ib_umem_dmabuf_get_pinned_with_dma_device(device, device->dma_device, + offset, size, fd, access); +} EXPORT_SYMBOL(ib_umem_dmabuf_get_pinned); void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf) diff --git a/include/rdma/ib_umem.h b/include/rdma/ib_umem.h index 565a85044541..de05268ed632 100644 --- a/include/rdma/ib_umem.h +++ b/include/rdma/ib_umem.h @@ -150,6 +150,11 @@ struct ib_umem_dmabuf *ib_umem_dmabuf_get_pinned(struct ib_device *device, unsigned long offset, size_t size, int fd, int access); +struct ib_umem_dmabuf * +ib_umem_dmabuf_get_pinned_with_dma_device(struct ib_device *device, + struct device *dma_device, + unsigned long offset, size_t size, + int fd, int access); int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf); void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf); void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf); @@ -196,6 +201,16 @@ ib_umem_dmabuf_get_pinned(struct ib_device *device, unsigned long offset, { return ERR_PTR(-EOPNOTSUPP); } + +static inline struct ib_umem_dmabuf * +ib_umem_dmabuf_get_pinned_with_dma_device(struct ib_device *device, + struct device *dma_device, + unsigned long offset, size_t size, + int fd, int access) +{ + return ERR_PTR(-EOPNOTSUPP); +} + static inline int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf) { return -EOPNOTSUPP; -- cgit From 253c61dc256b3e6be65657f78b4a8452163ce00f Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:14 +0300 Subject: RDMA/umem: Introduce an option to revoke DMABUF umem Introduce an option to revoke DMABUF umem. This option will retain the umem allocation while revoking its DMA mapping. Furthermore, any subsequent attempts to map the pages should fail once the umem has been revoked. This functionality will be utilized in the upcoming patches in the series, where we aim to delay umem deallocation until the mkey deregistration. However, we must unmap its pages immediately. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/a38270f2fe4a194868ca2312f4c1c760e51bcbff.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/umem_dmabuf.c | 21 +++++++++++++++++++-- include/rdma/ib_umem.h | 3 +++ 2 files changed, 22 insertions(+), 2 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/core/umem_dmabuf.c b/drivers/infiniband/core/umem_dmabuf.c index 726a09786547..9fcd37761264 100644 --- a/drivers/infiniband/core/umem_dmabuf.c +++ b/drivers/infiniband/core/umem_dmabuf.c @@ -23,6 +23,9 @@ int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf) dma_resv_assert_held(umem_dmabuf->attach->dmabuf->resv); + if (umem_dmabuf->revoked) + return -EINVAL; + if (umem_dmabuf->sgt) goto wait_fence; @@ -242,15 +245,29 @@ struct ib_umem_dmabuf *ib_umem_dmabuf_get_pinned(struct ib_device *device, } EXPORT_SYMBOL(ib_umem_dmabuf_get_pinned); -void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf) +void ib_umem_dmabuf_revoke(struct ib_umem_dmabuf *umem_dmabuf) { struct dma_buf *dmabuf = umem_dmabuf->attach->dmabuf; dma_resv_lock(dmabuf->resv, NULL); + if (umem_dmabuf->revoked) + goto end; ib_umem_dmabuf_unmap_pages(umem_dmabuf); - if (umem_dmabuf->pinned) + if (umem_dmabuf->pinned) { dma_buf_unpin(umem_dmabuf->attach); + umem_dmabuf->pinned = 0; + } + umem_dmabuf->revoked = 1; +end: dma_resv_unlock(dmabuf->resv); +} +EXPORT_SYMBOL(ib_umem_dmabuf_revoke); + +void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf) +{ + struct dma_buf *dmabuf = umem_dmabuf->attach->dmabuf; + + ib_umem_dmabuf_revoke(umem_dmabuf); dma_buf_detach(dmabuf, umem_dmabuf->attach); dma_buf_put(dmabuf); diff --git a/include/rdma/ib_umem.h b/include/rdma/ib_umem.h index de05268ed632..7dc7b1cc71b5 100644 --- a/include/rdma/ib_umem.h +++ b/include/rdma/ib_umem.h @@ -38,6 +38,7 @@ struct ib_umem_dmabuf { unsigned long last_sg_trim; void *private; u8 pinned : 1; + u8 revoked : 1; }; static inline struct ib_umem_dmabuf *to_ib_umem_dmabuf(struct ib_umem *umem) @@ -158,6 +159,7 @@ ib_umem_dmabuf_get_pinned_with_dma_device(struct ib_device *device, int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf); void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf); void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf); +void ib_umem_dmabuf_revoke(struct ib_umem_dmabuf *umem_dmabuf); #else /* CONFIG_INFINIBAND_USER_MEM */ @@ -217,6 +219,7 @@ static inline int ib_umem_dmabuf_map_pages(struct ib_umem_dmabuf *umem_dmabuf) } static inline void ib_umem_dmabuf_unmap_pages(struct ib_umem_dmabuf *umem_dmabuf) { } static inline void ib_umem_dmabuf_release(struct ib_umem_dmabuf *umem_dmabuf) { } +static inline void ib_umem_dmabuf_revoke(struct ib_umem_dmabuf *umem_dmabuf) {} #endif /* CONFIG_INFINIBAND_USER_MEM */ #endif /* IB_UMEM_H */ -- cgit From 3aa73c6b795b9aaaf933f3c95495d85fc0de39e3 Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:15 +0300 Subject: RDMA: Pass uverbs_attr_bundle as part of '.reg_user_mr_dmabuf' API Pass uverbs_attr_bundle as part of '.reg_user_mr_dmabuf' API instead of udata. This enables passing some new ioctl attributes to the drivers, as will be introduced in the next patches for mlx5 driver. Change the involved drivers accordingly. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/9a25b2fc02443f7c36c2d93499ae25252b6afd40.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/uverbs_std_types_mr.c | 2 +- drivers/infiniband/hw/bnxt_re/ib_verbs.c | 3 ++- drivers/infiniband/hw/bnxt_re/ib_verbs.h | 2 +- drivers/infiniband/hw/efa/efa.h | 2 +- drivers/infiniband/hw/efa/efa_verbs.c | 4 ++-- drivers/infiniband/hw/irdma/verbs.c | 2 +- drivers/infiniband/hw/mlx5/mlx5_ib.h | 2 +- drivers/infiniband/hw/mlx5/mr.c | 2 +- include/rdma/ib_verbs.h | 2 +- 9 files changed, 11 insertions(+), 10 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/core/uverbs_std_types_mr.c b/drivers/infiniband/core/uverbs_std_types_mr.c index 03e1db5d1e8c..7ebc7bd3caae 100644 --- a/drivers/infiniband/core/uverbs_std_types_mr.c +++ b/drivers/infiniband/core/uverbs_std_types_mr.c @@ -239,7 +239,7 @@ static int UVERBS_HANDLER(UVERBS_METHOD_REG_DMABUF_MR)( mr = pd->device->ops.reg_user_mr_dmabuf(pd, offset, length, iova, fd, access_flags, - &attrs->driver_udata); + attrs); if (IS_ERR(mr)) return PTR_ERR(mr); diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c index 7c757351a016..43a68e7de02a 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c @@ -4122,7 +4122,8 @@ struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length, struct ib_mr *bnxt_re_reg_user_mr_dmabuf(struct ib_pd *ib_pd, u64 start, u64 length, u64 virt_addr, int fd, - int mr_access_flags, struct ib_udata *udata) + int mr_access_flags, + struct uverbs_attr_bundle *attrs) { struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); struct bnxt_re_dev *rdev = pd->rdev; diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h index e98cb1717338..3ddeda312376 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h @@ -242,7 +242,7 @@ struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, struct ib_mr *bnxt_re_reg_user_mr_dmabuf(struct ib_pd *ib_pd, u64 start, u64 length, u64 virt_addr, int fd, int mr_access_flags, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata); void bnxt_re_dealloc_ucontext(struct ib_ucontext *context); int bnxt_re_mmap(struct ib_ucontext *context, struct vm_area_struct *vma); diff --git a/drivers/infiniband/hw/efa/efa.h b/drivers/infiniband/hw/efa/efa.h index e580e087e9da..d7fc9d5eeefd 100644 --- a/drivers/infiniband/hw/efa/efa.h +++ b/drivers/infiniband/hw/efa/efa.h @@ -168,7 +168,7 @@ struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length, struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata); int efa_get_port_immutable(struct ib_device *ibdev, u32 port_num, struct ib_port_immutable *immutable); diff --git a/drivers/infiniband/hw/efa/efa_verbs.c b/drivers/infiniband/hw/efa/efa_verbs.c index b1e0a1b7c59d..cc13415ff7e7 100644 --- a/drivers/infiniband/hw/efa/efa_verbs.c +++ b/drivers/infiniband/hw/efa/efa_verbs.c @@ -1684,14 +1684,14 @@ static int efa_register_mr(struct ib_pd *ibpd, struct efa_mr *mr, u64 start, struct ib_mr *efa_reg_user_mr_dmabuf(struct ib_pd *ibpd, u64 start, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata) + struct uverbs_attr_bundle *attrs) { struct efa_dev *dev = to_edev(ibpd->device); struct ib_umem_dmabuf *umem_dmabuf; struct efa_mr *mr; int err; - mr = efa_alloc_mr(ibpd, access_flags, udata); + mr = efa_alloc_mr(ibpd, access_flags, &attrs->driver_udata); if (IS_ERR(mr)) { err = PTR_ERR(mr); goto err_out; diff --git a/drivers/infiniband/hw/irdma/verbs.c b/drivers/infiniband/hw/irdma/verbs.c index fc0ce35da14e..6a107decb704 100644 --- a/drivers/infiniband/hw/irdma/verbs.c +++ b/drivers/infiniband/hw/irdma/verbs.c @@ -3085,7 +3085,7 @@ error: static struct ib_mr *irdma_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start, u64 len, u64 virt, int fd, int access, - struct ib_udata *udata) + struct uverbs_attr_bundle *attrs) { struct irdma_device *iwdev = to_iwdev(pd->device); struct ib_umem_dmabuf *umem_dmabuf; diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index b2ebea173547..e915a62da49c 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -1354,7 +1354,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 start, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); int mlx5_ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice, u32 flags, diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 98bd8eaa393e..1dfd9124bdd1 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -1513,7 +1513,7 @@ static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = { struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, u64 length, u64 virt_addr, int fd, int access_flags, - struct ib_udata *udata) + struct uverbs_attr_bundle *attrs) { struct mlx5_ib_dev *dev = to_mdev(pd->device); struct mlx5_ib_mr *mr = NULL; diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 6c5712ae559d..a1dcf812d787 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -2476,7 +2476,7 @@ struct ib_device_ops { struct ib_mr *(*reg_user_mr_dmabuf)(struct ib_pd *pd, u64 offset, u64 length, u64 virt_addr, int fd, int mr_access_flags, - struct ib_udata *udata); + struct uverbs_attr_bundle *attrs); struct ib_mr *(*rereg_user_mr)(struct ib_mr *mr, int flags, u64 start, u64 length, u64 virt_addr, int mr_access_flags, struct ib_pd *pd, -- cgit From de8f847a5114ff7cfcdfc114af8485c431dec703 Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:16 +0300 Subject: RDMA/mlx5: Add support for DMABUF MR registrations with Data-direct Add support for DMABUF MR registrations with Data-direct device. Upon userspace calling to register a DMABUF MR with the data direct bit set, the below algorithm will be followed. 1) Obtain a pinned DMABUF umem from the IB core using the user input parameters (FD, offset, length) and the DMA PF device. The DMA PF device is needed to allow the IOMMU to enable the DMA PF to access the user buffer over PCI. 2) Create a KSM MKEY by setting its entries according to the user buffer VA to IOVA mapping, with the MKEY being the data direct device-crossed MKEY. This KSM MKEY is umrable and will be used as part of the MR cache. The PD for creating it is the internal device 'data direct' kernel one. 3) Create a crossing MKEY that points to the KSM MKEY using the crossing access mode. 4) Manage the KSM MKEY by adding it to a list of 'data direct' MKEYs managed on the mlx5_ib device. 5) Return the crossing MKEY to the user, created with its supplied PD. Upon DMA PF unbind flow, the driver will revoke the KSM entries. The final deregistration will occur under the hood once the application deregisters its MKEY. Notes: - This version supports only the PINNED UMEM mode, so there is no dependency on ODP. - The IOVA supplied by the application must be system page aligned due to HW translations of KSM. - The crossing MKEY will not be umrable or part of the MR cache, as we cannot change its crossed (i.e. KSM) MKEY over UMR. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/1f99d8020ed540d9702b9e2252a145a439609ba6.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/mlx5/main.c | 11 ++ drivers/infiniband/hw/mlx5/mlx5_ib.h | 8 + drivers/infiniband/hw/mlx5/mr.c | 304 ++++++++++++++++++++++++++---- drivers/infiniband/hw/mlx5/odp.c | 5 +- drivers/infiniband/hw/mlx5/umr.c | 93 ++++++--- drivers/infiniband/hw/mlx5/umr.h | 1 + include/uapi/rdma/mlx5_user_ioctl_cmds.h | 4 + include/uapi/rdma/mlx5_user_ioctl_verbs.h | 4 + 8 files changed, 358 insertions(+), 72 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index fc0562f07249..b85ad3c0bfa1 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -3490,6 +3490,7 @@ static int mlx5_ib_data_direct_init(struct mlx5_ib_dev *dev) if (ret) return ret; + INIT_LIST_HEAD(&dev->data_direct_mr_list); ret = mlx5_data_direct_ib_reg(dev, vuid); if (ret) mlx5_ib_free_data_direct_resources(dev); @@ -3882,6 +3883,14 @@ ADD_UVERBS_ATTRIBUTES_SIMPLE( dump_fill_mkey), UA_MANDATORY)); +ADD_UVERBS_ATTRIBUTES_SIMPLE( + mlx5_ib_reg_dmabuf_mr, + UVERBS_OBJECT_MR, + UVERBS_METHOD_REG_DMABUF_MR, + UVERBS_ATTR_FLAGS_IN(MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS, + enum mlx5_ib_uapi_reg_dmabuf_flags, + UA_OPTIONAL)); + static const struct uapi_definition mlx5_ib_defs[] = { UAPI_DEF_CHAIN(mlx5_ib_devx_defs), UAPI_DEF_CHAIN(mlx5_ib_flow_defs), @@ -3891,6 +3900,7 @@ static const struct uapi_definition mlx5_ib_defs[] = { UAPI_DEF_CHAIN(mlx5_ib_create_cq_defs), UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_DEVICE, &mlx5_ib_query_context), + UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_MR, &mlx5_ib_reg_dmabuf_mr), UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_VAR, UAPI_DEF_IS_OBJ_SUPPORTED(var_is_supported)), UAPI_DEF_CHAIN_OBJ_TREE_NAMED(MLX5_IB_OBJECT_UAR), @@ -4396,6 +4406,7 @@ void mlx5_ib_data_direct_bind(struct mlx5_ib_dev *ibdev, void mlx5_ib_data_direct_unbind(struct mlx5_ib_dev *ibdev) { mutex_lock(&ibdev->data_direct_lock); + mlx5_ib_revoke_data_direct_mrs(ibdev); ibdev->data_direct_dev = NULL; mutex_unlock(&ibdev->data_direct_lock); } diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index e915a62da49c..be83a4d91a34 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -682,6 +682,8 @@ struct mlx5_ib_mr { struct mlx5_ib_mkey mmkey; struct ib_umem *umem; + /* The mr is data direct related */ + u8 data_direct :1; union { /* Used only by kernel MRs (umem == NULL) */ @@ -719,6 +721,10 @@ struct mlx5_ib_mr { } odp_destroy; struct ib_odp_counters odp_stats; bool is_odp_implicit; + /* The affilated data direct crossed mr */ + struct mlx5_ib_mr *dd_crossed_mr; + struct list_head dd_node; + u8 revoked :1; }; }; }; @@ -1169,6 +1175,7 @@ struct mlx5_ib_dev { /* protect resources needed as part of reset flow */ spinlock_t reset_flow_resource_lock; struct list_head qp_list; + struct list_head data_direct_mr_list; /* Array with num_ports elements */ struct mlx5_ib_port *port; struct mlx5_sq_bfreg bfreg; @@ -1437,6 +1444,7 @@ struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm, void mlx5_ib_data_direct_bind(struct mlx5_ib_dev *ibdev, struct mlx5_data_direct_dev *dev); void mlx5_ib_data_direct_unbind(struct mlx5_ib_dev *ibdev); +void mlx5_ib_revoke_data_direct_mrs(struct mlx5_ib_dev *dev); #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev); diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 1dfd9124bdd1..6829e3688b60 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -43,6 +43,7 @@ #include "dm.h" #include "mlx5_ib.h" #include "umr.h" +#include "data_direct.h" enum { MAX_PENDING_REG_MR = 8, @@ -54,7 +55,9 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context); static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, u64 iova, int access_flags, - unsigned int page_size, bool populate); + unsigned int page_size, bool populate, + int access_mode); +static int __mlx5_ib_dereg_mr(struct ib_mr *ibmr); static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr, struct ib_pd *pd) @@ -1126,12 +1129,10 @@ static unsigned int mlx5_umem_dmabuf_default_pgsz(struct ib_umem *umem, static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, struct ib_umem *umem, u64 iova, - int access_flags) + int access_flags, int access_mode) { - struct mlx5r_cache_rb_key rb_key = { - .access_mode = MLX5_MKC_ACCESS_MODE_MTT, - }; struct mlx5_ib_dev *dev = to_mdev(pd->device); + struct mlx5r_cache_rb_key rb_key = {}; struct mlx5_cache_ent *ent; struct mlx5_ib_mr *mr; unsigned int page_size; @@ -1144,6 +1145,7 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, if (WARN_ON(!page_size)) return ERR_PTR(-EINVAL); + rb_key.access_mode = access_mode; rb_key.ndescs = ib_umem_num_dma_blocks(umem, page_size); rb_key.ats = mlx5_umem_needs_ats(dev, umem, access_flags); rb_key.access_flags = get_unchangeable_access_flags(dev, access_flags); @@ -1154,7 +1156,7 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, */ if (!ent) { mutex_lock(&dev->slow_path_mutex); - mr = reg_create(pd, umem, iova, access_flags, page_size, false); + mr = reg_create(pd, umem, iova, access_flags, page_size, false, access_mode); mutex_unlock(&dev->slow_path_mutex); if (IS_ERR(mr)) return mr; @@ -1175,13 +1177,71 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, return mr; } +static struct ib_mr * +reg_create_crossing_vhca_mr(struct ib_pd *pd, u64 iova, u64 length, int access_flags, + u32 crossed_lkey) +{ + struct mlx5_ib_dev *dev = to_mdev(pd->device); + int access_mode = MLX5_MKC_ACCESS_MODE_CROSSING; + struct mlx5_ib_mr *mr; + void *mkc; + int inlen; + u32 *in; + int err; + + if (!MLX5_CAP_GEN(dev->mdev, crossing_vhca_mkey)) + return ERR_PTR(-EOPNOTSUPP); + + mr = kzalloc(sizeof(*mr), GFP_KERNEL); + if (!mr) + return ERR_PTR(-ENOMEM); + + inlen = MLX5_ST_SZ_BYTES(create_mkey_in); + in = kvzalloc(inlen, GFP_KERNEL); + if (!in) { + err = -ENOMEM; + goto err_1; + } + + mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); + MLX5_SET(mkc, mkc, crossing_target_vhca_id, + MLX5_CAP_GEN(dev->mdev, vhca_id)); + MLX5_SET(mkc, mkc, translations_octword_size, crossed_lkey); + MLX5_SET(mkc, mkc, access_mode_1_0, access_mode & 0x3); + MLX5_SET(mkc, mkc, access_mode_4_2, (access_mode >> 2) & 0x7); + + /* for this crossing mkey IOVA should be 0 and len should be IOVA + len */ + set_mkc_access_pd_addr_fields(mkc, access_flags, 0, pd); + MLX5_SET64(mkc, mkc, len, iova + length); + + MLX5_SET(mkc, mkc, free, 0); + MLX5_SET(mkc, mkc, umr_en, 0); + err = mlx5_ib_create_mkey(dev, &mr->mmkey, in, inlen); + if (err) + goto err_2; + + mr->mmkey.type = MLX5_MKEY_MR; + set_mr_fields(dev, mr, length, access_flags, iova); + mr->ibmr.pd = pd; + kvfree(in); + mlx5_ib_dbg(dev, "crossing mkey = 0x%x\n", mr->mmkey.key); + + return &mr->ibmr; +err_2: + kvfree(in); +err_1: + kfree(mr); + return ERR_PTR(err); +} + /* * If ibmr is NULL it will be allocated by reg_create. * Else, the given ibmr will be used. */ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, u64 iova, int access_flags, - unsigned int page_size, bool populate) + unsigned int page_size, bool populate, + int access_mode) { struct mlx5_ib_dev *dev = to_mdev(pd->device); struct mlx5_ib_mr *mr; @@ -1190,7 +1250,9 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, int inlen; u32 *in; int err; - bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg)); + bool pg_cap = !!(MLX5_CAP_GEN(dev->mdev, pg)) && + (access_mode == MLX5_MKC_ACCESS_MODE_MTT); + bool ksm_mode = (access_mode == MLX5_MKC_ACCESS_MODE_KSM); if (!page_size) return ERR_PTR(-EINVAL); @@ -1213,7 +1275,7 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, } pas = (__be64 *)MLX5_ADDR_OF(create_mkey_in, in, klm_pas_mtt); if (populate) { - if (WARN_ON(access_flags & IB_ACCESS_ON_DEMAND)) { + if (WARN_ON(access_flags & IB_ACCESS_ON_DEMAND || ksm_mode)) { err = -EINVAL; goto err_2; } @@ -1229,14 +1291,22 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry); set_mkc_access_pd_addr_fields(mkc, access_flags, iova, populate ? pd : dev->umrc.pd); + /* In case a data direct flow, overwrite the pdn field by its internal kernel PD */ + if (umem->is_dmabuf && ksm_mode) + MLX5_SET(mkc, mkc, pd, dev->ddr.pdn); + MLX5_SET(mkc, mkc, free, !populate); - MLX5_SET(mkc, mkc, access_mode_1_0, MLX5_MKC_ACCESS_MODE_MTT); + MLX5_SET(mkc, mkc, access_mode_1_0, access_mode); MLX5_SET(mkc, mkc, umr_en, 1); MLX5_SET64(mkc, mkc, len, umem->length); MLX5_SET(mkc, mkc, bsf_octword_size, 0); - MLX5_SET(mkc, mkc, translations_octword_size, - get_octo_len(iova, umem->length, mr->page_shift)); + if (ksm_mode) + MLX5_SET(mkc, mkc, translations_octword_size, + get_octo_len(iova, umem->length, mr->page_shift) * 2); + else + MLX5_SET(mkc, mkc, translations_octword_size, + get_octo_len(iova, umem->length, mr->page_shift)); MLX5_SET(mkc, mkc, log_page_size, mr->page_shift); if (mlx5_umem_needs_ats(dev, umem, access_flags)) MLX5_SET(mkc, mkc, ma_translation_mode, 1); @@ -1373,13 +1443,15 @@ static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem, xlt_with_umr = mlx5r_umr_can_load_pas(dev, umem->length); if (xlt_with_umr) { - mr = alloc_cacheable_mr(pd, umem, iova, access_flags); + mr = alloc_cacheable_mr(pd, umem, iova, access_flags, + MLX5_MKC_ACCESS_MODE_MTT); } else { unsigned int page_size = mlx5_umem_find_best_pgsz( umem, mkc, log_page_size, 0, iova); mutex_lock(&dev->slow_path_mutex); - mr = reg_create(pd, umem, iova, access_flags, page_size, true); + mr = reg_create(pd, umem, iova, access_flags, page_size, + true, MLX5_MKC_ACCESS_MODE_MTT); mutex_unlock(&dev->slow_path_mutex); } if (IS_ERR(mr)) { @@ -1442,7 +1514,8 @@ static struct ib_mr *create_user_odp_mr(struct ib_pd *pd, u64 start, u64 length, if (IS_ERR(odp)) return ERR_CAST(odp); - mr = alloc_cacheable_mr(pd, &odp->umem, iova, access_flags); + mr = alloc_cacheable_mr(pd, &odp->umem, iova, access_flags, + MLX5_MKC_ACCESS_MODE_MTT); if (IS_ERR(mr)) { ib_umem_release(&odp->umem); return ERR_CAST(mr); @@ -1510,35 +1583,31 @@ static struct dma_buf_attach_ops mlx5_ib_dmabuf_attach_ops = { .move_notify = mlx5_ib_dmabuf_invalidate_cb, }; -struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, - u64 length, u64 virt_addr, - int fd, int access_flags, - struct uverbs_attr_bundle *attrs) +static struct ib_mr * +reg_user_mr_dmabuf(struct ib_pd *pd, struct device *dma_device, + u64 offset, u64 length, u64 virt_addr, + int fd, int access_flags, int access_mode) { + bool pinned_mode = (access_mode == MLX5_MKC_ACCESS_MODE_KSM); struct mlx5_ib_dev *dev = to_mdev(pd->device); struct mlx5_ib_mr *mr = NULL; struct ib_umem_dmabuf *umem_dmabuf; int err; - if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) || - !IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) - return ERR_PTR(-EOPNOTSUPP); - - mlx5_ib_dbg(dev, - "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x\n", - offset, virt_addr, length, fd, access_flags); - err = mlx5r_umr_resource_init(dev); if (err) return ERR_PTR(err); - /* dmabuf requires xlt update via umr to work. */ - if (!mlx5r_umr_can_load_pas(dev, length)) - return ERR_PTR(-EINVAL); + if (!pinned_mode) + umem_dmabuf = ib_umem_dmabuf_get(&dev->ib_dev, + offset, length, fd, + access_flags, + &mlx5_ib_dmabuf_attach_ops); + else + umem_dmabuf = ib_umem_dmabuf_get_pinned_with_dma_device(&dev->ib_dev, + dma_device, offset, length, + fd, access_flags); - umem_dmabuf = ib_umem_dmabuf_get(&dev->ib_dev, offset, length, fd, - access_flags, - &mlx5_ib_dmabuf_attach_ops); if (IS_ERR(umem_dmabuf)) { mlx5_ib_dbg(dev, "umem_dmabuf get failed (%ld)\n", PTR_ERR(umem_dmabuf)); @@ -1546,7 +1615,7 @@ struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, } mr = alloc_cacheable_mr(pd, &umem_dmabuf->umem, virt_addr, - access_flags); + access_flags, access_mode); if (IS_ERR(mr)) { ib_umem_release(&umem_dmabuf->umem); return ERR_CAST(mr); @@ -1556,9 +1625,13 @@ struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, atomic_add(ib_umem_num_pages(mr->umem), &dev->mdev->priv.reg_pages); umem_dmabuf->private = mr; - err = mlx5r_store_odp_mkey(dev, &mr->mmkey); - if (err) - goto err_dereg_mr; + if (!pinned_mode) { + err = mlx5r_store_odp_mkey(dev, &mr->mmkey); + if (err) + goto err_dereg_mr; + } else { + mr->data_direct = true; + } err = mlx5_ib_init_dmabuf_mr(mr); if (err) @@ -1566,10 +1639,101 @@ struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, return &mr->ibmr; err_dereg_mr: - mlx5_ib_dereg_mr(&mr->ibmr, NULL); + __mlx5_ib_dereg_mr(&mr->ibmr); return ERR_PTR(err); } +static struct ib_mr * +reg_user_mr_dmabuf_by_data_direct(struct ib_pd *pd, u64 offset, + u64 length, u64 virt_addr, + int fd, int access_flags) +{ + struct mlx5_ib_dev *dev = to_mdev(pd->device); + struct mlx5_data_direct_dev *data_direct_dev; + struct ib_mr *crossing_mr; + struct ib_mr *crossed_mr; + int ret = 0; + + /* As of HW behaviour the IOVA must be page aligned in KSM mode */ + if (!PAGE_ALIGNED(virt_addr) || (access_flags & IB_ACCESS_ON_DEMAND)) + return ERR_PTR(-EOPNOTSUPP); + + mutex_lock(&dev->data_direct_lock); + data_direct_dev = dev->data_direct_dev; + if (!data_direct_dev) { + ret = -EINVAL; + goto end; + } + + /* The device's 'data direct mkey' was created without RO flags to + * simplify things and allow for a single mkey per device. + * Since RO is not a must, mask it out accordingly. + */ + access_flags &= ~IB_ACCESS_RELAXED_ORDERING; + crossed_mr = reg_user_mr_dmabuf(pd, &data_direct_dev->pdev->dev, + offset, length, virt_addr, fd, + access_flags, MLX5_MKC_ACCESS_MODE_KSM); + if (IS_ERR(crossed_mr)) { + ret = PTR_ERR(crossed_mr); + goto end; + } + + mutex_lock(&dev->slow_path_mutex); + crossing_mr = reg_create_crossing_vhca_mr(pd, virt_addr, length, access_flags, + crossed_mr->lkey); + mutex_unlock(&dev->slow_path_mutex); + if (IS_ERR(crossing_mr)) { + __mlx5_ib_dereg_mr(crossed_mr); + ret = PTR_ERR(crossing_mr); + goto end; + } + + list_add_tail(&to_mmr(crossed_mr)->dd_node, &dev->data_direct_mr_list); + to_mmr(crossing_mr)->dd_crossed_mr = to_mmr(crossed_mr); + to_mmr(crossing_mr)->data_direct = true; +end: + mutex_unlock(&dev->data_direct_lock); + return ret ? ERR_PTR(ret) : crossing_mr; +} + +struct ib_mr *mlx5_ib_reg_user_mr_dmabuf(struct ib_pd *pd, u64 offset, + u64 length, u64 virt_addr, + int fd, int access_flags, + struct uverbs_attr_bundle *attrs) +{ + struct mlx5_ib_dev *dev = to_mdev(pd->device); + int mlx5_access_flags = 0; + int err; + + if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) || + !IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING)) + return ERR_PTR(-EOPNOTSUPP); + + if (uverbs_attr_is_valid(attrs, MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS)) { + err = uverbs_get_flags32(&mlx5_access_flags, attrs, + MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS, + MLX5_IB_UAPI_REG_DMABUF_ACCESS_DATA_DIRECT); + if (err) + return ERR_PTR(err); + } + + mlx5_ib_dbg(dev, + "offset 0x%llx, virt_addr 0x%llx, length 0x%llx, fd %d, access_flags 0x%x, mlx5_access_flags 0x%x\n", + offset, virt_addr, length, fd, access_flags, mlx5_access_flags); + + /* dmabuf requires xlt update via umr to work. */ + if (!mlx5r_umr_can_load_pas(dev, length)) + return ERR_PTR(-EINVAL); + + if (mlx5_access_flags & MLX5_IB_UAPI_REG_DMABUF_ACCESS_DATA_DIRECT) + return reg_user_mr_dmabuf_by_data_direct(pd, offset, length, virt_addr, + fd, access_flags); + + return reg_user_mr_dmabuf(pd, pd->device->dma_device, + offset, length, virt_addr, + fd, access_flags, MLX5_MKC_ACCESS_MODE_MTT); +} + /* * True if the change in access flags can be done via UMR, only some access * flags can be updated. @@ -1665,7 +1829,7 @@ struct ib_mr *mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start, struct mlx5_ib_mr *mr = to_mmr(ib_mr); int err; - if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM)) + if (!IS_ENABLED(CONFIG_INFINIBAND_USER_MEM) || mr->data_direct) return ERR_PTR(-EOPNOTSUPP); mlx5_ib_dbg( @@ -1793,7 +1957,7 @@ err: static void mlx5_free_priv_descs(struct mlx5_ib_mr *mr) { - if (!mr->umem && mr->descs) { + if (!mr->umem && !mr->data_direct && mr->descs) { struct ib_device *device = mr->ibmr.device; int size = mr->max_descs * mr->desc_size; struct mlx5_ib_dev *dev = to_mdev(device); @@ -1847,6 +2011,34 @@ end: return ret; } +static int mlx5_ib_revoke_data_direct_mr(struct mlx5_ib_mr *mr) +{ + struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device); + struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem); + int err; + + lockdep_assert_held(&dev->data_direct_lock); + mr->revoked = true; + err = mlx5r_umr_revoke_mr(mr); + if (WARN_ON(err)) + return err; + + ib_umem_dmabuf_revoke(umem_dmabuf); + return 0; +} + +void mlx5_ib_revoke_data_direct_mrs(struct mlx5_ib_dev *dev) +{ + struct mlx5_ib_mr *mr, *next; + + lockdep_assert_held(&dev->data_direct_lock); + + list_for_each_entry_safe(mr, next, &dev->data_direct_mr_list, dd_node) { + list_del(&mr->dd_node); + mlx5_ib_revoke_data_direct_mr(mr); + } +} + static int mlx5_revoke_mr(struct mlx5_ib_mr *mr) { struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device); @@ -1864,7 +2056,7 @@ static int mlx5_revoke_mr(struct mlx5_ib_mr *mr) return destroy_mkey(dev, mr); } -int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) +static int __mlx5_ib_dereg_mr(struct ib_mr *ibmr) { struct mlx5_ib_mr *mr = to_mmr(ibmr); struct mlx5_ib_dev *dev = to_mdev(ibmr->device); @@ -1931,6 +2123,36 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) return 0; } +static int dereg_crossing_data_direct_mr(struct mlx5_ib_dev *dev, + struct mlx5_ib_mr *mr) +{ + struct mlx5_ib_mr *dd_crossed_mr = mr->dd_crossed_mr; + int ret; + + ret = __mlx5_ib_dereg_mr(&mr->ibmr); + if (ret) + return ret; + + mutex_lock(&dev->data_direct_lock); + if (!dd_crossed_mr->revoked) + list_del(&dd_crossed_mr->dd_node); + + ret = __mlx5_ib_dereg_mr(&dd_crossed_mr->ibmr); + mutex_unlock(&dev->data_direct_lock); + return ret; +} + +int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) +{ + struct mlx5_ib_mr *mr = to_mmr(ibmr); + struct mlx5_ib_dev *dev = to_mdev(ibmr->device); + + if (mr->data_direct) + return dereg_crossing_data_direct_mr(dev, mr); + + return __mlx5_ib_dereg_mr(ibmr); +} + static void mlx5_set_umr_free_mkey(struct ib_pd *pd, u32 *in, int ndescs, int access_mode, int page_shift) { diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index a524181f34df..44a3428ea342 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -710,7 +710,10 @@ static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt, ib_umem_dmabuf_unmap_pages(umem_dmabuf); err = -EINVAL; } else { - err = mlx5r_umr_update_mr_pas(mr, xlt_flags); + if (mr->data_direct) + err = mlx5r_umr_update_data_direct_ksm_pas(mr, xlt_flags); + else + err = mlx5r_umr_update_mr_pas(mr, xlt_flags); } dma_resv_unlock(umem_dmabuf->attach->dmabuf->resv); diff --git a/drivers/infiniband/hw/mlx5/umr.c b/drivers/infiniband/hw/mlx5/umr.c index ffc31b01f690..eb74c163fd83 100644 --- a/drivers/infiniband/hw/mlx5/umr.c +++ b/drivers/infiniband/hw/mlx5/umr.c @@ -632,44 +632,47 @@ static void mlx5r_umr_final_update_xlt(struct mlx5_ib_dev *dev, wqe->data_seg.byte_count = cpu_to_be32(sg->length); } -/* - * Send the DMA list to the HW for a normal MR using UMR. - * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP - * flag may be used. - */ -int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags) +static int +_mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags, bool dd) { + size_t ent_size = dd ? sizeof(struct mlx5_ksm) : sizeof(struct mlx5_mtt); struct mlx5_ib_dev *dev = mr_to_mdev(mr); struct device *ddev = &dev->mdev->pdev->dev; struct mlx5r_umr_wqe wqe = {}; struct ib_block_iter biter; + struct mlx5_ksm *cur_ksm; struct mlx5_mtt *cur_mtt; size_t orig_sg_length; - struct mlx5_mtt *mtt; size_t final_size; + void *curr_entry; struct ib_sge sg; + void *entry; u64 offset = 0; int err = 0; - if (WARN_ON(mr->umem->is_odp)) - return -EINVAL; - - mtt = mlx5r_umr_create_xlt( - dev, &sg, ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift), - sizeof(*mtt), flags); - if (!mtt) + entry = mlx5r_umr_create_xlt(dev, &sg, + ib_umem_num_dma_blocks(mr->umem, 1 << mr->page_shift), + ent_size, flags); + if (!entry) return -ENOMEM; orig_sg_length = sg.length; - mlx5r_umr_set_update_xlt_ctrl_seg(&wqe.ctrl_seg, flags, &sg); mlx5r_umr_set_update_xlt_mkey_seg(dev, &wqe.mkey_seg, mr, mr->page_shift); + if (dd) { + /* Use the data direct internal kernel PD */ + MLX5_SET(mkc, &wqe.mkey_seg, pd, dev->ddr.pdn); + cur_ksm = entry; + } else { + cur_mtt = entry; + } + mlx5r_umr_set_update_xlt_data_seg(&wqe.data_seg, &sg); - cur_mtt = mtt; + curr_entry = entry; rdma_umem_for_each_dma_block(mr->umem, &biter, BIT(mr->page_shift)) { - if (cur_mtt == (void *)mtt + sg.length) { + if (curr_entry == entry + sg.length) { dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE); @@ -681,23 +684,31 @@ int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags) DMA_TO_DEVICE); offset += sg.length; mlx5r_umr_update_offset(&wqe.ctrl_seg, offset); - - cur_mtt = mtt; + if (dd) + cur_ksm = entry; + else + cur_mtt = entry; } - cur_mtt->ptag = - cpu_to_be64(rdma_block_iter_dma_address(&biter) | - MLX5_IB_MTT_PRESENT); - - if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP)) - cur_mtt->ptag = 0; - - cur_mtt++; + if (dd) { + cur_ksm->va = cpu_to_be64(rdma_block_iter_dma_address(&biter)); + cur_ksm->key = cpu_to_be32(dev->ddr.mkey); + cur_ksm++; + curr_entry = cur_ksm; + } else { + cur_mtt->ptag = + cpu_to_be64(rdma_block_iter_dma_address(&biter) | + MLX5_IB_MTT_PRESENT); + if (mr->umem->is_dmabuf && (flags & MLX5_IB_UPD_XLT_ZAP)) + cur_mtt->ptag = 0; + cur_mtt++; + curr_entry = cur_mtt; + } } - final_size = (void *)cur_mtt - (void *)mtt; + final_size = curr_entry - entry; sg.length = ALIGN(final_size, MLX5_UMR_FLEX_ALIGNMENT); - memset(cur_mtt, 0, sg.length - final_size); + memset(curr_entry, 0, sg.length - final_size); mlx5r_umr_final_update_xlt(dev, &wqe, mr, &sg, flags); dma_sync_single_for_device(ddev, sg.addr, sg.length, DMA_TO_DEVICE); @@ -705,10 +716,32 @@ int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags) err: sg.length = orig_sg_length; - mlx5r_umr_unmap_free_xlt(dev, mtt, &sg); + mlx5r_umr_unmap_free_xlt(dev, entry, &sg); return err; } +int mlx5r_umr_update_data_direct_ksm_pas(struct mlx5_ib_mr *mr, unsigned int flags) +{ + /* No invalidation flow is expected */ + if (WARN_ON(!mr->umem->is_dmabuf) || (flags & MLX5_IB_UPD_XLT_ZAP)) + return -EINVAL; + + return _mlx5r_umr_update_mr_pas(mr, flags, true); +} + +/* + * Send the DMA list to the HW for a normal MR using UMR. + * Dmabuf MR is handled in a similar way, except that the MLX5_IB_UPD_XLT_ZAP + * flag may be used. + */ +int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags) +{ + if (WARN_ON(mr->umem->is_odp)) + return -EINVAL; + + return _mlx5r_umr_update_mr_pas(mr, flags, false); +} + static bool umr_can_use_indirect_mkey(struct mlx5_ib_dev *dev) { return !MLX5_CAP_GEN(dev->mdev, umr_indirect_mkey_disabled); diff --git a/drivers/infiniband/hw/mlx5/umr.h b/drivers/infiniband/hw/mlx5/umr.h index 5f734dc72bef..4a02c9b5aad8 100644 --- a/drivers/infiniband/hw/mlx5/umr.h +++ b/drivers/infiniband/hw/mlx5/umr.h @@ -95,6 +95,7 @@ int mlx5r_umr_revoke_mr(struct mlx5_ib_mr *mr); int mlx5r_umr_rereg_pd_access(struct mlx5_ib_mr *mr, struct ib_pd *pd, int access_flags); int mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags); +int mlx5r_umr_update_data_direct_ksm_pas(struct mlx5_ib_mr *mr, unsigned int flags); int mlx5r_umr_update_xlt(struct mlx5_ib_mr *mr, u64 idx, int npages, int page_shift, int flags); diff --git a/include/uapi/rdma/mlx5_user_ioctl_cmds.h b/include/uapi/rdma/mlx5_user_ioctl_cmds.h index 5b74d6534899..106276a4cce7 100644 --- a/include/uapi/rdma/mlx5_user_ioctl_cmds.h +++ b/include/uapi/rdma/mlx5_user_ioctl_cmds.h @@ -274,6 +274,10 @@ enum mlx5_ib_create_cq_attrs { MLX5_IB_ATTR_CREATE_CQ_UAR_INDEX = UVERBS_ID_DRIVER_NS_WITH_UHW, }; +enum mlx5_ib_reg_dmabuf_mr_attrs { + MLX5_IB_ATTR_REG_DMABUF_MR_ACCESS_FLAGS = (1U << UVERBS_ID_NS_SHIFT), +}; + #define MLX5_IB_DW_MATCH_PARAM 0xA0 struct mlx5_ib_match_params { diff --git a/include/uapi/rdma/mlx5_user_ioctl_verbs.h b/include/uapi/rdma/mlx5_user_ioctl_verbs.h index 3189c7f08d17..7c233df475e7 100644 --- a/include/uapi/rdma/mlx5_user_ioctl_verbs.h +++ b/include/uapi/rdma/mlx5_user_ioctl_verbs.h @@ -54,6 +54,10 @@ enum mlx5_ib_uapi_flow_action_packet_reformat_type { MLX5_IB_UAPI_FLOW_ACTION_PACKET_REFORMAT_TYPE_L2_TO_L3_TUNNEL = 0x3, }; +enum mlx5_ib_uapi_reg_dmabuf_flags { + MLX5_IB_UAPI_REG_DMABUF_ACCESS_DATA_DIRECT = 1 << 0, +}; + struct mlx5_ib_uapi_devx_async_cmd_hdr { __aligned_u64 wr_id; __u8 out_data[]; -- cgit From ec7ad6530909983c8736c80af46e3529ce7bab55 Mon Sep 17 00:00:00 2001 From: Yishai Hadas Date: Thu, 1 Aug 2024 15:05:17 +0300 Subject: RDMA/mlx5: Introduce GET_DATA_DIRECT_SYSFS_PATH ioctl Introduce the 'GET_DATA_DIRECT_SYSFS_PATH' ioctl to return the sysfs path of the affiliated 'data direct' device for a given device. Signed-off-by: Yishai Hadas Link: https://patch.msgid.link/403745463e0ef52adbef681ff09aa6a29a756352.1722512548.git.leon@kernel.org Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/mlx5/std_types.c | 55 +++++++++++++++++++++++++++++++- include/uapi/rdma/mlx5_user_ioctl_cmds.h | 5 +++ 2 files changed, 59 insertions(+), 1 deletion(-) (limited to 'include') diff --git a/drivers/infiniband/hw/mlx5/std_types.c b/drivers/infiniband/hw/mlx5/std_types.c index 5c83765a85e0..bdb568411091 100644 --- a/drivers/infiniband/hw/mlx5/std_types.c +++ b/drivers/infiniband/hw/mlx5/std_types.c @@ -10,6 +10,7 @@ #include #include #include "mlx5_ib.h" +#include "data_direct.h" #define UVERBS_MODULE_NAME mlx5_ib #include @@ -204,6 +205,50 @@ static int UVERBS_HANDLER(MLX5_IB_METHOD_QUERY_PORT)( sizeof(info)); } +static int UVERBS_HANDLER(MLX5_IB_METHOD_GET_DATA_DIRECT_SYSFS_PATH)( + struct uverbs_attr_bundle *attrs) +{ + struct mlx5_data_direct_dev *data_direct_dev; + struct mlx5_ib_ucontext *c; + struct mlx5_ib_dev *dev; + int out_len = uverbs_attr_get_len(attrs, + MLX5_IB_ATTR_GET_DATA_DIRECT_SYSFS_PATH); + u32 dev_path_len; + char *dev_path; + int ret; + + c = to_mucontext(ib_uverbs_get_ucontext(attrs)); + if (IS_ERR(c)) + return PTR_ERR(c); + dev = to_mdev(c->ibucontext.device); + mutex_lock(&dev->data_direct_lock); + data_direct_dev = dev->data_direct_dev; + if (!data_direct_dev) { + ret = -ENODEV; + goto end; + } + + dev_path = kobject_get_path(&data_direct_dev->device->kobj, GFP_KERNEL); + if (!dev_path) { + ret = -ENOMEM; + goto end; + } + + dev_path_len = strlen(dev_path) + 1; + if (dev_path_len > out_len) { + ret = -ENOSPC; + goto end; + } + + ret = uverbs_copy_to(attrs, MLX5_IB_ATTR_GET_DATA_DIRECT_SYSFS_PATH, dev_path, + dev_path_len); + kfree(dev_path); + +end: + mutex_unlock(&dev->data_direct_lock); + return ret; +} + DECLARE_UVERBS_NAMED_METHOD( MLX5_IB_METHOD_QUERY_PORT, UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_QUERY_PORT_PORT_NUM, @@ -214,9 +259,17 @@ DECLARE_UVERBS_NAMED_METHOD( reg_c0), UA_MANDATORY)); +DECLARE_UVERBS_NAMED_METHOD( + MLX5_IB_METHOD_GET_DATA_DIRECT_SYSFS_PATH, + UVERBS_ATTR_PTR_OUT( + MLX5_IB_ATTR_GET_DATA_DIRECT_SYSFS_PATH, + UVERBS_ATTR_MIN_SIZE(0), + UA_MANDATORY)); + ADD_UVERBS_METHODS(mlx5_ib_device, UVERBS_OBJECT_DEVICE, - &UVERBS_METHOD(MLX5_IB_METHOD_QUERY_PORT)); + &UVERBS_METHOD(MLX5_IB_METHOD_QUERY_PORT), + &UVERBS_METHOD(MLX5_IB_METHOD_GET_DATA_DIRECT_SYSFS_PATH)); DECLARE_UVERBS_NAMED_METHOD( MLX5_IB_METHOD_PD_QUERY, diff --git a/include/uapi/rdma/mlx5_user_ioctl_cmds.h b/include/uapi/rdma/mlx5_user_ioctl_cmds.h index 106276a4cce7..fd2e4a3a56b3 100644 --- a/include/uapi/rdma/mlx5_user_ioctl_cmds.h +++ b/include/uapi/rdma/mlx5_user_ioctl_cmds.h @@ -348,6 +348,7 @@ enum mlx5_ib_pd_methods { enum mlx5_ib_device_methods { MLX5_IB_METHOD_QUERY_PORT = (1U << UVERBS_ID_NS_SHIFT), + MLX5_IB_METHOD_GET_DATA_DIRECT_SYSFS_PATH, }; enum mlx5_ib_query_port_attrs { @@ -355,4 +356,8 @@ enum mlx5_ib_query_port_attrs { MLX5_IB_ATTR_QUERY_PORT, }; +enum mlx5_ib_get_data_direct_sysfs_path_attrs { + MLX5_IB_ATTR_GET_DATA_DIRECT_SYSFS_PATH = (1U << UVERBS_ID_NS_SHIFT), +}; + #endif -- cgit From d8ea645d6984c84a87032063a0941f15a323831f Mon Sep 17 00:00:00 2001 From: Selvin Xavier Date: Sun, 18 Aug 2024 21:47:26 -0700 Subject: RDMA/bnxt_re: Handle variable WQE support for user applications User library calculates the number of slots required for user applications and it can pass that information to the driver. Driver can use this value and update the HW directly. This mechanism is currently used only for the newly introduced variable size WQEs. Extend the bnxt_re_qp_req structure to pass the Send Queue slot count. Reorganize the code to get the sq_slots before initializing the Send Queue attributes. Link: https://patch.msgid.link/r/1724042847-1481-5-git-send-email-selvin.xavier@broadcom.com Signed-off-by: Hongguang Gao Signed-off-by: Selvin Xavier Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/bnxt_re/ib_verbs.c | 106 ++++++++++++++++++------------- drivers/infiniband/hw/bnxt_re/ib_verbs.h | 16 ++++- include/uapi/rdma/bnxt_re-abi.h | 6 ++ 3 files changed, 82 insertions(+), 46 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c index 4dd137b7a5ce..2932db129958 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c @@ -1017,20 +1017,15 @@ static int bnxt_re_setup_swqe_size(struct bnxt_re_qp *qp, } static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd, - struct bnxt_re_qp *qp, struct ib_udata *udata) + struct bnxt_re_qp *qp, struct bnxt_re_ucontext *cntx, + struct bnxt_re_qp_req *ureq) { struct bnxt_qplib_qp *qplib_qp; - struct bnxt_re_ucontext *cntx; - struct bnxt_re_qp_req ureq; int bytes = 0, psn_sz; struct ib_umem *umem; int psn_nume; qplib_qp = &qp->qplib_qp; - cntx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, - ib_uctx); - if (ib_copy_from_udata(&ureq, udata, sizeof(ureq))) - return -EFAULT; bytes = (qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size); /* Consider mapping PSN search memory only for RC QPs. */ @@ -1038,17 +1033,20 @@ static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd, psn_sz = bnxt_qplib_is_chip_gen_p5_p7(rdev->chip_ctx) ? sizeof(struct sq_psn_search_ext) : sizeof(struct sq_psn_search); - psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ? - qplib_qp->sq.max_wqe : - ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) / - sizeof(struct bnxt_qplib_sge)); + if (cntx && bnxt_re_is_var_size_supported(rdev, cntx)) { + psn_nume = ureq->sq_slots; + } else { + psn_nume = (qplib_qp->wqe_mode == BNXT_QPLIB_WQE_MODE_STATIC) ? + qplib_qp->sq.max_wqe : ((qplib_qp->sq.max_wqe * qplib_qp->sq.wqe_size) / + sizeof(struct bnxt_qplib_sge)); + } if (_is_host_msn_table(rdev->qplib_res.dattr->dev_cap_flags2)) psn_nume = roundup_pow_of_two(psn_nume); bytes += (psn_nume * psn_sz); } bytes = PAGE_ALIGN(bytes); - umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes, + umem = ib_umem_get(&rdev->ibdev, ureq->qpsva, bytes, IB_ACCESS_LOCAL_WRITE); if (IS_ERR(umem)) return PTR_ERR(umem); @@ -1057,12 +1055,12 @@ static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd, qplib_qp->sq.sg_info.umem = umem; qplib_qp->sq.sg_info.pgsize = PAGE_SIZE; qplib_qp->sq.sg_info.pgshft = PAGE_SHIFT; - qplib_qp->qp_handle = ureq.qp_handle; + qplib_qp->qp_handle = ureq->qp_handle; if (!qp->qplib_qp.srq) { bytes = (qplib_qp->rq.max_wqe * qplib_qp->rq.wqe_size); bytes = PAGE_ALIGN(bytes); - umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes, + umem = ib_umem_get(&rdev->ibdev, ureq->qprva, bytes, IB_ACCESS_LOCAL_WRITE); if (IS_ERR(umem)) goto rqfail; @@ -1261,14 +1259,15 @@ static void bnxt_re_adjust_gsi_rq_attr(struct bnxt_re_qp *qp) static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp, struct ib_qp_init_attr *init_attr, - struct bnxt_re_ucontext *uctx) + struct bnxt_re_ucontext *uctx, + struct bnxt_re_qp_req *ureq) { struct bnxt_qplib_dev_attr *dev_attr; struct bnxt_qplib_qp *qplqp; struct bnxt_re_dev *rdev; struct bnxt_qplib_q *sq; + int diff = 0; int entries; - int diff; int rc; rdev = qp->rdev; @@ -1277,22 +1276,28 @@ static int bnxt_re_init_sq_attr(struct bnxt_re_qp *qp, dev_attr = &rdev->dev_attr; sq->max_sge = init_attr->cap.max_send_sge; - if (sq->max_sge > dev_attr->max_qp_sges) { - sq->max_sge = dev_attr->max_qp_sges; - init_attr->cap.max_send_sge = sq->max_sge; - } + entries = init_attr->cap.max_send_wr; + if (uctx && qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) { + sq->max_wqe = ureq->sq_slots; + sq->max_sw_wqe = ureq->sq_slots; + sq->wqe_size = sizeof(struct sq_sge); + } else { + if (sq->max_sge > dev_attr->max_qp_sges) { + sq->max_sge = dev_attr->max_qp_sges; + init_attr->cap.max_send_sge = sq->max_sge; + } - rc = bnxt_re_setup_swqe_size(qp, init_attr); - if (rc) - return rc; + rc = bnxt_re_setup_swqe_size(qp, init_attr); + if (rc) + return rc; - entries = init_attr->cap.max_send_wr; - /* Allocate 128 + 1 more than what's provided */ - diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ? - 0 : BNXT_QPLIB_RESERVED_QP_WRS; - entries = bnxt_re_init_depth(entries + diff + 1, uctx); - sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1); - sq->max_sw_wqe = bnxt_qplib_get_depth(sq, qplqp->wqe_mode, true); + /* Allocate 128 + 1 more than what's provided */ + diff = (qplqp->wqe_mode == BNXT_QPLIB_WQE_MODE_VARIABLE) ? + 0 : BNXT_QPLIB_RESERVED_QP_WRS; + entries = bnxt_re_init_depth(entries + diff + 1, uctx); + sq->max_wqe = min_t(u32, entries, dev_attr->max_qp_wqes + diff + 1); + sq->max_sw_wqe = bnxt_qplib_get_depth(sq, qplqp->wqe_mode, true); + } sq->q_full_delta = diff + 1; /* * Reserving one slot for Phantom WQE. Application can @@ -1355,10 +1360,10 @@ out: static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, struct ib_qp_init_attr *init_attr, - struct ib_udata *udata) + struct bnxt_re_ucontext *uctx, + struct bnxt_re_qp_req *ureq) { struct bnxt_qplib_dev_attr *dev_attr; - struct bnxt_re_ucontext *uctx; struct bnxt_qplib_qp *qplqp; struct bnxt_re_dev *rdev; struct bnxt_re_cq *cq; @@ -1368,7 +1373,6 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, qplqp = &qp->qplib_qp; dev_attr = &rdev->dev_attr; - uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); /* Setup misc params */ ether_addr_copy(qplqp->smac, rdev->netdev->dev_addr); qplqp->pd = &pd->qplib_pd; @@ -1381,8 +1385,7 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, goto out; } qplqp->type = (u8)qptype; - qplqp->wqe_mode = rdev->chip_ctx->modes.wqe_mode; - + qplqp->wqe_mode = bnxt_re_is_var_size_supported(rdev, uctx); if (init_attr->qp_type == IB_QPT_RC) { qplqp->max_rd_atomic = dev_attr->max_qp_rd_atom; qplqp->max_dest_rd_atomic = dev_attr->max_qp_init_rd_atom; @@ -1417,14 +1420,14 @@ static int bnxt_re_init_qp_attr(struct bnxt_re_qp *qp, struct bnxt_re_pd *pd, bnxt_re_adjust_gsi_rq_attr(qp); /* Setup SQ */ - rc = bnxt_re_init_sq_attr(qp, init_attr, uctx); + rc = bnxt_re_init_sq_attr(qp, init_attr, uctx, ureq); if (rc) goto out; if (init_attr->qp_type == IB_QPT_GSI) bnxt_re_adjust_gsi_sq_attr(qp, init_attr, uctx); - if (udata) /* This will update DPI and qp_handle */ - rc = bnxt_re_init_user_qp(rdev, pd, qp, udata); + if (uctx) /* This will update DPI and qp_handle */ + rc = bnxt_re_init_user_qp(rdev, pd, qp, uctx, ureq); out: return rc; } @@ -1525,14 +1528,27 @@ static bool bnxt_re_test_qp_limits(struct bnxt_re_dev *rdev, int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr, struct ib_udata *udata) { - struct ib_pd *ib_pd = ib_qp->pd; - struct bnxt_re_pd *pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); - struct bnxt_re_dev *rdev = pd->rdev; - struct bnxt_qplib_dev_attr *dev_attr = &rdev->dev_attr; - struct bnxt_re_qp *qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); + struct bnxt_qplib_dev_attr *dev_attr; + struct bnxt_re_ucontext *uctx; + struct bnxt_re_qp_req ureq; + struct bnxt_re_dev *rdev; + struct bnxt_re_pd *pd; + struct bnxt_re_qp *qp; + struct ib_pd *ib_pd; u32 active_qps; int rc; + ib_pd = ib_qp->pd; + pd = container_of(ib_pd, struct bnxt_re_pd, ib_pd); + rdev = pd->rdev; + dev_attr = &rdev->dev_attr; + qp = container_of(ib_qp, struct bnxt_re_qp, ib_qp); + + uctx = rdma_udata_to_drv_context(udata, struct bnxt_re_ucontext, ib_uctx); + if (udata) + if (ib_copy_from_udata(&ureq, udata, min(udata->inlen, sizeof(ureq)))) + return -EFAULT; + rc = bnxt_re_test_qp_limits(rdev, qp_init_attr, dev_attr); if (!rc) { rc = -EINVAL; @@ -1540,7 +1556,7 @@ int bnxt_re_create_qp(struct ib_qp *ib_qp, struct ib_qp_init_attr *qp_init_attr, } qp->rdev = rdev; - rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, udata); + rc = bnxt_re_init_qp_attr(qp, pd, qp_init_attr, uctx, &ureq); if (rc) goto fail; @@ -4215,7 +4231,7 @@ int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata) goto cfail; if (ureq.comp_mask & BNXT_RE_COMP_MASK_REQ_UCNTX_POW2_SUPPORT) { resp.comp_mask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED; - uctx->cmask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED; + uctx->cmask |= BNXT_RE_UCNTX_CAP_POW2_DISABLED; } } diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h index e98cb1717338..7c8350fb8aad 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h @@ -171,12 +171,26 @@ static inline u16 bnxt_re_get_rwqe_size(int nsge) return sizeof(struct rq_wqe_hdr) + (nsge * sizeof(struct sq_sge)); } +enum { + BNXT_RE_UCNTX_CAP_POW2_DISABLED = 0x1ULL, + BNXT_RE_UCNTX_CAP_VAR_WQE_ENABLED = 0x2ULL, +}; + static inline u32 bnxt_re_init_depth(u32 ent, struct bnxt_re_ucontext *uctx) { - return uctx ? (uctx->cmask & BNXT_RE_UCNTX_CMASK_POW2_DISABLED) ? + return uctx ? (uctx->cmask & BNXT_RE_UCNTX_CAP_POW2_DISABLED) ? ent : roundup_pow_of_two(ent) : ent; } +static inline bool bnxt_re_is_var_size_supported(struct bnxt_re_dev *rdev, + struct bnxt_re_ucontext *uctx) +{ + if (uctx) + return uctx->cmask & BNXT_RE_UCNTX_CAP_VAR_WQE_ENABLED; + else + return rdev->chip_ctx->modes.wqe_mode; +} + int bnxt_re_query_device(struct ib_device *ibdev, struct ib_device_attr *ib_attr, struct ib_udata *udata); diff --git a/include/uapi/rdma/bnxt_re-abi.h b/include/uapi/rdma/bnxt_re-abi.h index e61104f35d73..71140618700a 100644 --- a/include/uapi/rdma/bnxt_re-abi.h +++ b/include/uapi/rdma/bnxt_re-abi.h @@ -118,10 +118,16 @@ struct bnxt_re_resize_cq_req { __aligned_u64 cq_va; }; +enum bnxt_re_qp_mask { + BNXT_RE_QP_REQ_MASK_VAR_WQE_SQ_SLOTS = 0x1, +}; + struct bnxt_re_qp_req { __aligned_u64 qpsva; __aligned_u64 qprva; __aligned_u64 qp_handle; + __aligned_u64 comp_mask; + __u32 sq_slots; }; struct bnxt_re_qp_resp { -- cgit From 10a104c0debbb19a1e45193d5670510216e339ff Mon Sep 17 00:00:00 2001 From: Selvin Xavier Date: Sun, 18 Aug 2024 21:47:27 -0700 Subject: RDMA/bnxt_re: Enable variable size WQEs for user space applications Add backward compatibility code to enable variable size WQEs only if the user lib supports it. Link: https://patch.msgid.link/r/1724042847-1481-6-git-send-email-selvin.xavier@broadcom.com Signed-off-by: Hongguang Gao Signed-off-by: Selvin Xavier Signed-off-by: Jason Gunthorpe --- drivers/infiniband/hw/bnxt_re/ib_verbs.c | 5 +++++ include/uapi/rdma/bnxt_re-abi.h | 1 + 2 files changed, 6 insertions(+) (limited to 'include') diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c index 2932db129958..82444fd748f1 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c @@ -4233,6 +4233,11 @@ int bnxt_re_alloc_ucontext(struct ib_ucontext *ctx, struct ib_udata *udata) resp.comp_mask |= BNXT_RE_UCNTX_CMASK_POW2_DISABLED; uctx->cmask |= BNXT_RE_UCNTX_CAP_POW2_DISABLED; } + if (ureq.comp_mask & BNXT_RE_COMP_MASK_REQ_UCNTX_VAR_WQE_SUPPORT) { + resp.comp_mask |= BNXT_RE_UCNTX_CMASK_HAVE_MODE; + resp.mode = rdev->chip_ctx->modes.wqe_mode; + uctx->cmask |= BNXT_RE_UCNTX_CAP_VAR_WQE_ENABLED; + } } rc = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp))); diff --git a/include/uapi/rdma/bnxt_re-abi.h b/include/uapi/rdma/bnxt_re-abi.h index 71140618700a..6821002931c8 100644 --- a/include/uapi/rdma/bnxt_re-abi.h +++ b/include/uapi/rdma/bnxt_re-abi.h @@ -66,6 +66,7 @@ enum bnxt_re_wqe_mode { enum { BNXT_RE_COMP_MASK_REQ_UCNTX_POW2_SUPPORT = 0x01, + BNXT_RE_COMP_MASK_REQ_UCNTX_VAR_WQE_SUPPORT = 0x02, }; struct bnxt_re_uctx_req { -- cgit From 181028a0d84cdcc7ac86d05cc49eaa416ce85c8b Mon Sep 17 00:00:00 2001 From: Chandramohan Akula Date: Thu, 29 Aug 2024 08:34:05 -0700 Subject: RDMA/bnxt_re: Share a page to expose per SRQ info with userspace Gen P7 adapters needs to share a toggle bits information received in kernel driver with the user space. User space needs this info to arm the SRQ. User space application can get this page using the UAPI routines. Library will mmap this page and get the toggle bits to be used in the next ARM Doorbell. Uses a hash list to map the SRQ structure from the SRQ ID. SRQ structure is retrieved from the hash list while the library calls the UAPI routine to get the toggle page mapping. Currently the full page is mapped per SRQ. This can be optimized to enable multiple SRQs from the same application share the same page and different offsets in the page Signed-off-by: Chandramohan Akula Signed-off-by: Selvin Xavier Link: https://patch.msgid.link/1724945645-14989-4-git-send-email-selvin.xavier@broadcom.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/bnxt_re/bnxt_re.h | 2 ++ drivers/infiniband/hw/bnxt_re/ib_verbs.c | 34 +++++++++++++++++++++++++++++++- drivers/infiniband/hw/bnxt_re/ib_verbs.h | 1 + drivers/infiniband/hw/bnxt_re/main.c | 6 +++++- include/uapi/rdma/bnxt_re-abi.h | 6 ++++++ 5 files changed, 47 insertions(+), 2 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/bnxt_re/bnxt_re.h b/drivers/infiniband/hw/bnxt_re/bnxt_re.h index 0912d2fa9634..2be9a62d230f 100644 --- a/drivers/infiniband/hw/bnxt_re/bnxt_re.h +++ b/drivers/infiniband/hw/bnxt_re/bnxt_re.h @@ -141,6 +141,7 @@ struct bnxt_re_pacing { #define BNXT_RE_GRC_FIFO_REG_BASE 0x2000 #define MAX_CQ_HASH_BITS (16) +#define MAX_SRQ_HASH_BITS (16) struct bnxt_re_dev { struct ib_device ibdev; struct list_head list; @@ -196,6 +197,7 @@ struct bnxt_re_dev { struct work_struct dbq_fifo_check_work; struct delayed_work dbq_pacing_work; DECLARE_HASHTABLE(cq_hash, MAX_CQ_HASH_BITS); + DECLARE_HASHTABLE(srq_hash, MAX_SRQ_HASH_BITS); }; #define to_bnxt_re_dev(ptr, member) \ diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.c b/drivers/infiniband/hw/bnxt_re/ib_verbs.c index eafbee4af067..f9f944a6094a 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.c +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.c @@ -1707,6 +1707,10 @@ int bnxt_re_destroy_srq(struct ib_srq *ib_srq, struct ib_udata *udata) if (qplib_srq->cq) nq = qplib_srq->cq->nq; + if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT) { + free_page((unsigned long)srq->uctx_srq_page); + hash_del(&srq->hash_entry); + } bnxt_qplib_destroy_srq(&rdev->qplib_res, qplib_srq); ib_umem_release(srq->umem); atomic_dec(&rdev->stats.res.srq_count); @@ -1811,9 +1815,18 @@ int bnxt_re_create_srq(struct ib_srq *ib_srq, } if (udata) { - struct bnxt_re_srq_resp resp; + struct bnxt_re_srq_resp resp = {}; resp.srqid = srq->qplib_srq.id; + if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT) { + hash_add(rdev->srq_hash, &srq->hash_entry, srq->qplib_srq.id); + srq->uctx_srq_page = (void *)get_zeroed_page(GFP_KERNEL); + if (!srq->uctx_srq_page) { + rc = -ENOMEM; + goto fail; + } + resp.comp_mask |= BNXT_RE_SRQ_TOGGLE_PAGE_SUPPORT; + } rc = ib_copy_to_udata(udata, &resp, sizeof(resp)); if (rc) { ibdev_err(&rdev->ibdev, "SRQ copy to udata failed!"); @@ -4291,6 +4304,19 @@ static struct bnxt_re_cq *bnxt_re_search_for_cq(struct bnxt_re_dev *rdev, u32 cq return cq; } +static struct bnxt_re_srq *bnxt_re_search_for_srq(struct bnxt_re_dev *rdev, u32 srq_id) +{ + struct bnxt_re_srq *srq = NULL, *tmp_srq; + + hash_for_each_possible(rdev->srq_hash, tmp_srq, hash_entry, srq_id) { + if (tmp_srq->qplib_srq.id == srq_id) { + srq = tmp_srq; + break; + } + } + return srq; +} + /* Helper function to mmap the virtual memory from user app */ int bnxt_re_mmap(struct ib_ucontext *ib_uctx, struct vm_area_struct *vma) { @@ -4519,6 +4545,7 @@ static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bund struct bnxt_re_ucontext *uctx; struct ib_ucontext *ib_uctx; struct bnxt_re_dev *rdev; + struct bnxt_re_srq *srq; u32 length = PAGE_SIZE; struct bnxt_re_cq *cq; u64 mem_offset; @@ -4550,6 +4577,11 @@ static int UVERBS_HANDLER(BNXT_RE_METHOD_GET_TOGGLE_MEM)(struct uverbs_attr_bund addr = (u64)cq->uctx_cq_page; break; case BNXT_RE_SRQ_TOGGLE_MEM: + srq = bnxt_re_search_for_srq(rdev, res_id); + if (!srq) + return -EINVAL; + + addr = (u64)srq->uctx_srq_page; break; default: diff --git a/drivers/infiniband/hw/bnxt_re/ib_verbs.h b/drivers/infiniband/hw/bnxt_re/ib_verbs.h index 060bf62eceba..b789e47ec97a 100644 --- a/drivers/infiniband/hw/bnxt_re/ib_verbs.h +++ b/drivers/infiniband/hw/bnxt_re/ib_verbs.h @@ -78,6 +78,7 @@ struct bnxt_re_srq { struct ib_umem *umem; spinlock_t lock; /* protect srq */ void *uctx_srq_page; + struct hlist_node hash_entry; }; struct bnxt_re_qp { diff --git a/drivers/infiniband/hw/bnxt_re/main.c b/drivers/infiniband/hw/bnxt_re/main.c index 31ba89cffe9d..16a84ca1ce48 100644 --- a/drivers/infiniband/hw/bnxt_re/main.c +++ b/drivers/infiniband/hw/bnxt_re/main.c @@ -139,8 +139,10 @@ static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev) if (bnxt_re_hwrm_qcaps(rdev)) dev_err(rdev_to_dev(rdev), "Failed to query hwrm qcaps\n"); - if (bnxt_qplib_is_chip_gen_p7(rdev->chip_ctx)) + if (bnxt_qplib_is_chip_gen_p7(rdev->chip_ctx)) { cctx->modes.toggle_bits |= BNXT_QPLIB_CQ_TOGGLE_BIT; + cctx->modes.toggle_bits |= BNXT_QPLIB_SRQ_TOGGLE_BIT; + } } static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev) @@ -1771,6 +1773,8 @@ static int bnxt_re_dev_init(struct bnxt_re_dev *rdev) bnxt_re_vf_res_config(rdev); } hash_init(rdev->cq_hash); + if (rdev->chip_ctx->modes.toggle_bits & BNXT_QPLIB_SRQ_TOGGLE_BIT) + hash_init(rdev->srq_hash); return 0; free_sctx: diff --git a/include/uapi/rdma/bnxt_re-abi.h b/include/uapi/rdma/bnxt_re-abi.h index 6821002931c8..faa9d62b3b30 100644 --- a/include/uapi/rdma/bnxt_re-abi.h +++ b/include/uapi/rdma/bnxt_re-abi.h @@ -141,8 +141,14 @@ struct bnxt_re_srq_req { __aligned_u64 srq_handle; }; +enum bnxt_re_srq_mask { + BNXT_RE_SRQ_TOGGLE_PAGE_SUPPORT = 0x1, +}; + struct bnxt_re_srq_resp { __u32 srqid; + __u32 rsvd; /* padding */ + __aligned_u64 comp_mask; }; enum bnxt_re_shpg_offt { -- cgit From cef7dde8836ab09a3bfe96ada4f18ef2496eacc9 Mon Sep 17 00:00:00 2001 From: Michael Guralnik Date: Mon, 9 Sep 2024 13:04:57 +0300 Subject: net/mlx5: Expand mkey page size to support 6 bits Protect the usage of the 6th bit with the relevant capability to ensure we are using the new page sizes with FW that supports the bit extension. Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909100504.29797-2-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/mlx5/mlx5_ib.h | 27 ++++++++++++++++----------- drivers/infiniband/hw/mlx5/mr.c | 10 ++++------ drivers/infiniband/hw/mlx5/odp.c | 2 +- include/linux/mlx5/mlx5_ifc.h | 7 ++++--- 4 files changed, 25 insertions(+), 21 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index 5505eb70939b..1c96f209cda6 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -63,17 +63,6 @@ __mlx5_log_page_size_to_bitmap(unsigned int log_pgsz_bits, return GENMASK(largest_pg_shift, pgsz_shift); } -/* - * For mkc users, instead of a page_offset the command has a start_iova which - * specifies both the page_offset and the on-the-wire IOVA - */ -#define mlx5_umem_find_best_pgsz(umem, typ, log_pgsz_fld, pgsz_shift, iova) \ - ib_umem_find_best_pgsz(umem, \ - __mlx5_log_page_size_to_bitmap( \ - __mlx5_bit_sz(typ, log_pgsz_fld), \ - pgsz_shift), \ - iova) - static __always_inline unsigned long __mlx5_page_offset_to_bitmask(unsigned int page_offset_bits, unsigned int offset_shift) @@ -1724,4 +1713,20 @@ static inline u32 smi_to_native_portnum(struct mlx5_ib_dev *dev, u32 port) return (port - 1) / dev->num_ports + 1; } +/* + * For mkc users, instead of a page_offset the command has a start_iova which + * specifies both the page_offset and the on-the-wire IOVA + */ +static __always_inline unsigned long +mlx5_umem_mkc_find_best_pgsz(struct mlx5_ib_dev *dev, struct ib_umem *umem, + u64 iova) +{ + int page_size_bits = + MLX5_CAP_GEN_2(dev->mdev, umr_log_entity_size_5) ? 6 : 5; + unsigned long bitmap = + __mlx5_log_page_size_to_bitmap(page_size_bits, 0); + + return ib_umem_find_best_pgsz(umem, bitmap, iova); +} + #endif /* MLX5_IB_H */ diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index d5b5cd73e20c..45d9dc9c6c8f 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -1120,8 +1120,7 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, if (umem->is_dmabuf) page_size = mlx5_umem_dmabuf_default_pgsz(umem, iova); else - page_size = mlx5_umem_find_best_pgsz(umem, mkc, log_page_size, - 0, iova); + page_size = mlx5_umem_mkc_find_best_pgsz(dev, umem, iova); if (WARN_ON(!page_size)) return ERR_PTR(-EINVAL); @@ -1426,8 +1425,8 @@ static struct ib_mr *create_real_mr(struct ib_pd *pd, struct ib_umem *umem, mr = alloc_cacheable_mr(pd, umem, iova, access_flags, MLX5_MKC_ACCESS_MODE_MTT); } else { - unsigned int page_size = mlx5_umem_find_best_pgsz( - umem, mkc, log_page_size, 0, iova); + unsigned int page_size = + mlx5_umem_mkc_find_best_pgsz(dev, umem, iova); mutex_lock(&dev->slow_path_mutex); mr = reg_create(pd, umem, iova, access_flags, page_size, @@ -1745,8 +1744,7 @@ static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr, if (!mlx5r_umr_can_load_pas(dev, new_umem->length)) return false; - *page_size = - mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova); + *page_size = mlx5_umem_mkc_find_best_pgsz(dev, new_umem, iova); if (WARN_ON(!*page_size)) return false; return (mr->mmkey.cache_ent->rb_key.ndescs) >= diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 44a3428ea342..221820874e7a 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -693,7 +693,7 @@ static int pagefault_dmabuf_mr(struct mlx5_ib_mr *mr, size_t bcnt, struct ib_umem_dmabuf *umem_dmabuf = to_ib_umem_dmabuf(mr->umem); u32 xlt_flags = 0; int err; - unsigned int page_size; + unsigned long page_size; if (flags & MLX5_PF_FLAGS_ENABLE) xlt_flags |= MLX5_IB_UPD_XLT_ENABLE; diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h index 970c9d8473ef..ec1117d4e441 100644 --- a/include/linux/mlx5/mlx5_ifc.h +++ b/include/linux/mlx5/mlx5_ifc.h @@ -1988,7 +1988,9 @@ struct mlx5_ifc_cmd_hca_cap_2_bits { u8 migratable[0x1]; u8 reserved_at_81[0x11]; u8 query_vuid[0x1]; - u8 reserved_at_93[0xd]; + u8 reserved_at_93[0x5]; + u8 umr_log_entity_size_5[0x1]; + u8 reserved_at_99[0x7]; u8 max_reformat_insert_size[0x8]; u8 max_reformat_insert_offset[0x8]; @@ -4212,8 +4214,7 @@ struct mlx5_ifc_mkc_bits { u8 reserved_at_1c0[0x19]; u8 relaxed_ordering_read[0x1]; - u8 reserved_at_1d9[0x1]; - u8 log_page_size[0x5]; + u8 log_page_size[0x6]; u8 reserved_at_1e0[0x20]; }; -- cgit From 6cd9171d04cff79abe78c166927ab8563bf95fe5 Mon Sep 17 00:00:00 2001 From: Michael Guralnik Date: Mon, 9 Sep 2024 13:04:58 +0300 Subject: net/mlx5: Expose HW bits for Memory scheme ODP Expose IFC bits to support the new memory scheme on demand paging. Change the macro reading odp capabilities to be able to read from the new IFC layout and align the code in upper layers to be compiled. Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909100504.29797-3-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/mlx5/odp.c | 40 ++++++++++-------- drivers/net/ethernet/mellanox/mlx5/core/main.c | 28 ++++++------- include/linux/mlx5/device.h | 4 ++ include/linux/mlx5/mlx5_ifc.h | 57 +++++++++++++++++++++----- 4 files changed, 86 insertions(+), 43 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 221820874e7a..300504bf79d7 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -332,46 +332,46 @@ static void internal_fill_odp_caps(struct mlx5_ib_dev *dev) else dev->odp_max_size = BIT_ULL(MLX5_MAX_UMR_SHIFT + PAGE_SHIFT); - if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.send)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, ud_odp_caps.send)) caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SEND; - if (MLX5_CAP_ODP(dev->mdev, ud_odp_caps.srq_receive)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, ud_odp_caps.srq_receive)) caps->per_transport_caps.ud_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.send)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.send)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SEND; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.receive)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.receive)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_RECV; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.write)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.write)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_WRITE; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.read)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.read)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_READ; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.atomic)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.atomic)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; - if (MLX5_CAP_ODP(dev->mdev, rc_odp_caps.srq_receive)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, rc_odp_caps.srq_receive)) caps->per_transport_caps.rc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.send)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.send)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SEND; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.receive)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.receive)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_RECV; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.write)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.write)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_WRITE; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.read)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.read)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_READ; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.atomic)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.atomic)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_ATOMIC; - if (MLX5_CAP_ODP(dev->mdev, xrc_odp_caps.srq_receive)) + if (MLX5_CAP_ODP_SCHEME(dev->mdev, xrc_odp_caps.srq_receive)) caps->per_transport_caps.xrc_odp_caps |= IB_ODP_SUPPORT_SRQ_RECV; if (MLX5_CAP_GEN(dev->mdev, fixed_buffer_size) && @@ -388,13 +388,17 @@ static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev, int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ? pfault->wqe.wq_num : pfault->token; u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)] = {}; + void *info; int err; MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME); - MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type); - MLX5_SET(page_fault_resume_in, in, token, pfault->token); - MLX5_SET(page_fault_resume_in, in, wq_number, wq_num); - MLX5_SET(page_fault_resume_in, in, error, !!error); + + info = MLX5_ADDR_OF(page_fault_resume_in, in, + page_fault_info.trans_page_fault_info); + MLX5_SET(trans_page_fault_info, info, page_fault_type, pfault->type); + MLX5_SET(trans_page_fault_info, info, fault_token, pfault->token); + MLX5_SET(trans_page_fault_info, info, wq_number, wq_num); + MLX5_SET(trans_page_fault_info, info, error, !!error); err = mlx5_cmd_exec_in(dev->mdev, page_fault_resume, in); if (err) diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/net/ethernet/mellanox/mlx5/core/main.c index 5b7e6f4b5c7e..cc2aa46cff04 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/main.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c @@ -479,20 +479,20 @@ static int handle_hca_cap_odp(struct mlx5_core_dev *dev, void *set_ctx) } \ } while (0) - ODP_CAP_SET_MAX(dev, ud_odp_caps.srq_receive); - ODP_CAP_SET_MAX(dev, rc_odp_caps.srq_receive); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.srq_receive); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.send); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.receive); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.write); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.read); - ODP_CAP_SET_MAX(dev, xrc_odp_caps.atomic); - ODP_CAP_SET_MAX(dev, dc_odp_caps.srq_receive); - ODP_CAP_SET_MAX(dev, dc_odp_caps.send); - ODP_CAP_SET_MAX(dev, dc_odp_caps.receive); - ODP_CAP_SET_MAX(dev, dc_odp_caps.write); - ODP_CAP_SET_MAX(dev, dc_odp_caps.read); - ODP_CAP_SET_MAX(dev, dc_odp_caps.atomic); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.ud_odp_caps.srq_receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.rc_odp_caps.srq_receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.srq_receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.send); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.write); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.read); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.xrc_odp_caps.atomic); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.srq_receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.send); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.receive); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.write); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.read); + ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.atomic); if (!do_set) return 0; diff --git a/include/linux/mlx5/device.h b/include/linux/mlx5/device.h index ba875a619b97..bd081f276654 100644 --- a/include/linux/mlx5/device.h +++ b/include/linux/mlx5/device.h @@ -1369,6 +1369,10 @@ enum mlx5_qcam_feature_groups { #define MLX5_CAP_ODP(mdev, cap)\ MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, cap) +#define MLX5_CAP_ODP_SCHEME(mdev, cap) \ + MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, \ + transport_page_fault_scheme_cap.cap) + #define MLX5_CAP_ODP_MAX(mdev, cap)\ MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->max, cap) diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h index ec1117d4e441..3e3336bb9191 100644 --- a/include/linux/mlx5/mlx5_ifc.h +++ b/include/linux/mlx5/mlx5_ifc.h @@ -1326,11 +1326,13 @@ struct mlx5_ifc_atomic_caps_bits { u8 reserved_at_e0[0x720]; }; -struct mlx5_ifc_odp_cap_bits { +struct mlx5_ifc_odp_scheme_cap_bits { u8 reserved_at_0[0x40]; u8 sig[0x1]; - u8 reserved_at_41[0x1f]; + u8 reserved_at_41[0x4]; + u8 page_prefetch[0x1]; + u8 reserved_at_46[0x1a]; u8 reserved_at_60[0x20]; @@ -1344,7 +1346,20 @@ struct mlx5_ifc_odp_cap_bits { struct mlx5_ifc_odp_per_transport_service_cap_bits dc_odp_caps; - u8 reserved_at_120[0x6E0]; + u8 reserved_at_120[0xe0]; +}; + +struct mlx5_ifc_odp_cap_bits { + struct mlx5_ifc_odp_scheme_cap_bits transport_page_fault_scheme_cap; + + struct mlx5_ifc_odp_scheme_cap_bits memory_page_fault_scheme_cap; + + u8 reserved_at_400[0x200]; + + u8 mem_page_fault[0x1]; + u8 reserved_at_601[0x1f]; + + u8 reserved_at_620[0x1e0]; }; struct mlx5_ifc_tls_cap_bits { @@ -2034,7 +2049,8 @@ struct mlx5_ifc_cmd_hca_cap_2_bits { u8 min_mkey_log_entity_size_fixed_buffer[0x5]; u8 ec_vf_vport_base[0x10]; - u8 reserved_at_3a0[0x10]; + u8 reserved_at_3a0[0xa]; + u8 max_mkey_log_entity_size_mtt[0x6]; u8 max_rqt_vhca_id[0x10]; u8 reserved_at_3c0[0x20]; @@ -7258,6 +7274,30 @@ struct mlx5_ifc_qp_2err_in_bits { u8 reserved_at_60[0x20]; }; +struct mlx5_ifc_trans_page_fault_info_bits { + u8 error[0x1]; + u8 reserved_at_1[0x4]; + u8 page_fault_type[0x3]; + u8 wq_number[0x18]; + + u8 reserved_at_20[0x8]; + u8 fault_token[0x18]; +}; + +struct mlx5_ifc_mem_page_fault_info_bits { + u8 error[0x1]; + u8 reserved_at_1[0xf]; + u8 fault_token_47_32[0x10]; + + u8 fault_token_31_0[0x20]; +}; + +union mlx5_ifc_page_fault_resume_in_page_fault_info_auto_bits { + struct mlx5_ifc_trans_page_fault_info_bits trans_page_fault_info; + struct mlx5_ifc_mem_page_fault_info_bits mem_page_fault_info; + u8 reserved_at_0[0x40]; +}; + struct mlx5_ifc_page_fault_resume_out_bits { u8 status[0x8]; u8 reserved_at_8[0x18]; @@ -7274,13 +7314,8 @@ struct mlx5_ifc_page_fault_resume_in_bits { u8 reserved_at_20[0x10]; u8 op_mod[0x10]; - u8 error[0x1]; - u8 reserved_at_41[0x4]; - u8 page_fault_type[0x3]; - u8 wq_number[0x18]; - - u8 reserved_at_60[0x8]; - u8 token[0x18]; + union mlx5_ifc_page_fault_resume_in_page_fault_info_auto_bits + page_fault_info; }; struct mlx5_ifc_nop_out_bits { -- cgit From 64c68385a39bb676c76da36164ab696e8da78842 Mon Sep 17 00:00:00 2001 From: Michael Guralnik Date: Mon, 9 Sep 2024 13:04:59 +0300 Subject: RDMA/mlx5: Add new ODP memory scheme eqe format Add new fields to support the new memory scheme page fault and extend the token field to u64 as in the new scheme the token is 48 bit. Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909100504.29797-4-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/hw/mlx5/odp.c | 48 ++++++++++++++++++++++++---------------- include/linux/mlx5/device.h | 22 +++++++++++++++++- 2 files changed, 50 insertions(+), 20 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c index 300504bf79d7..f01026d507a3 100644 --- a/drivers/infiniband/hw/mlx5/odp.c +++ b/drivers/infiniband/hw/mlx5/odp.c @@ -45,7 +45,7 @@ /* Contains the details of a pagefault. */ struct mlx5_pagefault { u32 bytes_committed; - u32 token; + u64 token; u8 event_subtype; u8 type; union { @@ -74,6 +74,14 @@ struct mlx5_pagefault { u32 rdma_op_len; u64 rdma_va; } rdma; + struct { + u64 va; + u32 mkey; + u32 fault_byte_count; + u32 prefetch_before_byte_count; + u32 prefetch_after_byte_count; + u8 flags; + } memory; }; struct mlx5_ib_pf_eq *eq; @@ -1273,7 +1281,7 @@ read_user: if (ret) mlx5_ib_err( dev, - "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %x\n", + "Failed reading a WQE following page fault, error %d, wqe_index %x, qpn %llx\n", ret, wqe_index, pfault->token); resolve_page_fault: @@ -1332,13 +1340,13 @@ static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev, } else if (ret < 0 || pages_in_range(address, length) > ret) { mlx5_ib_page_fault_resume(dev, pfault, 1); if (ret != -ENOENT) - mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n", + mlx5_ib_dbg(dev, "PAGE FAULT error %d. QP 0x%llx, type: 0x%x\n", ret, pfault->token, pfault->type); return; } mlx5_ib_page_fault_resume(dev, pfault, 0); - mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n", + mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%llx, type: 0x%x, prefetch_activated: %d\n", pfault->token, pfault->type, prefetch_activated); @@ -1354,7 +1362,7 @@ static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev, prefetch_len, &bytes_committed, NULL); if (ret < 0 && ret != -EAGAIN) { - mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n", + mlx5_ib_dbg(dev, "Prefetch failed. ret: %d, QP 0x%llx, address: 0x%.16llx, length = 0x%.16x\n", ret, pfault->token, address, prefetch_len); } } @@ -1405,15 +1413,12 @@ static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) pf_eqe = &eqe->data.page_fault; pfault->event_subtype = eqe->sub_type; - pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed); - - mlx5_ib_dbg(eq->dev, - "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n", - eqe->sub_type, pfault->bytes_committed); switch (eqe->sub_type) { case MLX5_PFAULT_SUBTYPE_RDMA: /* RDMA based event */ + pfault->bytes_committed = + be32_to_cpu(pf_eqe->rdma.bytes_committed); pfault->type = be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24; pfault->token = @@ -1427,10 +1432,12 @@ static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) be32_to_cpu(pf_eqe->rdma.rdma_op_len); pfault->rdma.rdma_va = be64_to_cpu(pf_eqe->rdma.rdma_va); - mlx5_ib_dbg(eq->dev, - "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n", - pfault->type, pfault->token, - pfault->rdma.r_key); + mlx5_ib_dbg( + eq->dev, + "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x, type:0x%x, token: 0x%06llx, r_key: 0x%08x\n", + eqe->sub_type, pfault->bytes_committed, + pfault->type, pfault->token, + pfault->rdma.r_key); mlx5_ib_dbg(eq->dev, "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n", pfault->rdma.rdma_op_len, @@ -1439,6 +1446,8 @@ static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) case MLX5_PFAULT_SUBTYPE_WQE: /* WQE based event */ + pfault->bytes_committed = + be32_to_cpu(pf_eqe->wqe.bytes_committed); pfault->type = (be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7; pfault->token = @@ -1450,11 +1459,12 @@ static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq) be16_to_cpu(pf_eqe->wqe.wqe_index); pfault->wqe.packet_size = be16_to_cpu(pf_eqe->wqe.packet_length); - mlx5_ib_dbg(eq->dev, - "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n", - pfault->type, pfault->token, - pfault->wqe.wq_num, - pfault->wqe.wqe_index); + mlx5_ib_dbg( + eq->dev, + "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x, type:0x%x, token: 0x%06llx, wq_num: 0x%06x, wqe_index: 0x%04x\n", + eqe->sub_type, pfault->bytes_committed, + pfault->type, pfault->token, pfault->wqe.wq_num, + pfault->wqe.wqe_index); break; default: diff --git a/include/linux/mlx5/device.h b/include/linux/mlx5/device.h index bd081f276654..154095256d0d 100644 --- a/include/linux/mlx5/device.h +++ b/include/linux/mlx5/device.h @@ -211,6 +211,7 @@ enum { enum { MLX5_PFAULT_SUBTYPE_WQE = 0, MLX5_PFAULT_SUBTYPE_RDMA = 1, + MLX5_PFAULT_SUBTYPE_MEMORY = 2, }; enum wqe_page_fault_type { @@ -646,10 +647,11 @@ struct mlx5_eqe_page_req { __be32 rsvd1[5]; }; +#define MEMORY_SCHEME_PAGE_FAULT_GRANULARITY 4096 struct mlx5_eqe_page_fault { - __be32 bytes_committed; union { struct { + __be32 bytes_committed; u16 reserved1; __be16 wqe_index; u16 reserved2; @@ -659,6 +661,7 @@ struct mlx5_eqe_page_fault { __be32 pftype_wq; } __packed wqe; struct { + __be32 bytes_committed; __be32 r_key; u16 reserved1; __be16 packet_length; @@ -666,6 +669,23 @@ struct mlx5_eqe_page_fault { __be64 rdma_va; __be32 pftype_token; } __packed rdma; + struct { + u8 flags; + u8 reserved1; + __be16 post_demand_fault_pages; + __be16 pre_demand_fault_pages; + __be16 token47_32; + __be32 token31_0; + /* + * FW changed from specifying the fault size in byte + * count to 4k pages granularity. The size specified + * in pages uses bits 31:12, to keep backward + * compatibility. + */ + __be32 demand_fault_pages; + __be32 mkey; + __be64 va; + } __packed memory; } __packed; } __packed; -- cgit From 907936b6f4e630718cc31ddea79cc76a3e32080a Mon Sep 17 00:00:00 2001 From: Michael Guralnik Date: Mon, 9 Sep 2024 13:05:04 +0300 Subject: net/mlx5: Handle memory scheme ODP capabilities When running over new FW that supports the new memory scheme ODP, set the cap in the FW to signal the FW we are working in the new scheme. In the memory scheme ODP the per_transport_service capabilities are RO for the driver so we skip their setting. Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909100504.29797-9-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/net/ethernet/mellanox/mlx5/core/main.c | 22 ++++++++++++++++++---- include/linux/mlx5/device.h | 10 +++++++--- 2 files changed, 25 insertions(+), 7 deletions(-) (limited to 'include') diff --git a/drivers/net/ethernet/mellanox/mlx5/core/main.c b/drivers/net/ethernet/mellanox/mlx5/core/main.c index cc2aa46cff04..4ec6507d094a 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/main.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/main.c @@ -454,8 +454,8 @@ static int handle_hca_cap_atomic(struct mlx5_core_dev *dev, void *set_ctx) static int handle_hca_cap_odp(struct mlx5_core_dev *dev, void *set_ctx) { + bool do_set = false, mem_page_fault = false; void *set_hca_cap; - bool do_set = false; int err; if (!IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) || @@ -470,6 +470,17 @@ static int handle_hca_cap_odp(struct mlx5_core_dev *dev, void *set_ctx) memcpy(set_hca_cap, dev->caps.hca[MLX5_CAP_ODP]->cur, MLX5_ST_SZ_BYTES(odp_cap)); + /* For best performance, enable memory scheme ODP only when + * it has page prefetch enabled. + */ + if (MLX5_CAP_ODP_MAX(dev, mem_page_fault) && + MLX5_CAP_ODP_MAX(dev, memory_page_fault_scheme_cap.page_prefetch)) { + mem_page_fault = true; + do_set = true; + MLX5_SET(odp_cap, set_hca_cap, mem_page_fault, mem_page_fault); + goto set; + } + #define ODP_CAP_SET_MAX(dev, field) \ do { \ u32 _res = MLX5_CAP_ODP_MAX(dev, field); \ @@ -494,10 +505,13 @@ static int handle_hca_cap_odp(struct mlx5_core_dev *dev, void *set_ctx) ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.read); ODP_CAP_SET_MAX(dev, transport_page_fault_scheme_cap.dc_odp_caps.atomic); - if (!do_set) - return 0; +set: + if (do_set) + err = set_caps(dev, set_ctx, MLX5_SET_HCA_CAP_OP_MOD_ODP); - return set_caps(dev, set_ctx, MLX5_SET_HCA_CAP_OP_MOD_ODP); + mlx5_core_dbg(dev, "Using ODP %s scheme\n", + mem_page_fault ? "memory" : "transport"); + return err; } static int max_uc_list_get_devlink_param(struct mlx5_core_dev *dev) diff --git a/include/linux/mlx5/device.h b/include/linux/mlx5/device.h index 154095256d0d..57c9b18c3adb 100644 --- a/include/linux/mlx5/device.h +++ b/include/linux/mlx5/device.h @@ -1389,9 +1389,13 @@ enum mlx5_qcam_feature_groups { #define MLX5_CAP_ODP(mdev, cap)\ MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, cap) -#define MLX5_CAP_ODP_SCHEME(mdev, cap) \ - MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, \ - transport_page_fault_scheme_cap.cap) +#define MLX5_CAP_ODP_SCHEME(mdev, cap) \ + (MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, \ + mem_page_fault) ? \ + MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, \ + memory_page_fault_scheme_cap.cap) : \ + MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->cur, \ + transport_page_fault_scheme_cap.cap)) #define MLX5_CAP_ODP_MAX(mdev, cap)\ MLX5_GET(odp_cap, mdev->caps.hca[MLX5_CAP_ODP]->max, cap) -- cgit From 8d159eb2117b2e3697a31785662b653938f007cb Mon Sep 17 00:00:00 2001 From: Chiara Meiohas Date: Mon, 9 Sep 2024 20:30:23 +0300 Subject: RDMA/mlx5: Use IB set_netdev and get_netdev functions The IB layer provides a common interface to store and get net devices associated to an IB device port (ib_device_set_netdev() and ib_device_get_netdev()). Previously, mlx5_ib stored and managed the associated net devices internally. Replace internal net device management in mlx5_ib with ib_device_set_netdev() when attaching/detaching a net device and ib_device_get_netdev() when retrieving the net device. Export ib_device_get_netdev(). For mlx5 representors/PFs/VFs and lag creation we replace the netdev assignments with the IB set/get netdev functions. In active-backup mode lag the active slave net device is stored in the lag itself. To assure the net device stored in a lag bond IB device is the active slave we implement the following: - mlx5_core: when modifying the slave of a bond we send the internal driver event MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE. - mlx5_ib: when catching the event call ib_device_set_netdev() This patch also ensures the correct IB events are sent in switchdev lag. While at it, when in multiport eswitch mode, only a single IB device is created for all ports. The said IB device will receive all netdev events of its VFs once loaded, thus to avoid overwriting the mapping of PF IB device to PF netdev, ignore NETDEV_REGISTER events if the ib device has already been mapped to a netdev. Signed-off-by: Chiara Meiohas Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909173025.30422-6-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/device.c | 4 + drivers/infiniband/hw/mlx5/ib_rep.c | 23 +-- drivers/infiniband/hw/mlx5/main.c | 183 +++++++++++++++------- drivers/infiniband/hw/mlx5/mlx5_ib.h | 3 +- drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c | 76 ++++----- include/linux/mlx5/device.h | 1 + include/linux/mlx5/driver.h | 2 +- include/rdma/ib_verbs.h | 2 + 8 files changed, 191 insertions(+), 103 deletions(-) (limited to 'include') diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c index b1377503cb9d..9e765c79a892 100644 --- a/drivers/infiniband/core/device.c +++ b/drivers/infiniband/core/device.c @@ -2236,6 +2236,9 @@ struct net_device *ib_device_get_netdev(struct ib_device *ib_dev, if (!rdma_is_port_valid(ib_dev, port)) return NULL; + if (!ib_dev->port_data) + return NULL; + pdata = &ib_dev->port_data[port]; /* @@ -2254,6 +2257,7 @@ struct net_device *ib_device_get_netdev(struct ib_device *ib_dev, return res; } +EXPORT_SYMBOL(ib_device_get_netdev); /** * ib_device_get_by_netdev - Find an IB device associated with a netdev diff --git a/drivers/infiniband/hw/mlx5/ib_rep.c b/drivers/infiniband/hw/mlx5/ib_rep.c index 1ad934685d80..49af1cfbe6d1 100644 --- a/drivers/infiniband/hw/mlx5/ib_rep.c +++ b/drivers/infiniband/hw/mlx5/ib_rep.c @@ -13,6 +13,7 @@ mlx5_ib_set_vport_rep(struct mlx5_core_dev *dev, int vport_index) { struct mlx5_ib_dev *ibdev; + struct net_device *ndev; ibdev = mlx5_eswitch_uplink_get_proto_dev(dev->priv.eswitch, REP_IB); if (!ibdev) @@ -20,12 +21,9 @@ mlx5_ib_set_vport_rep(struct mlx5_core_dev *dev, ibdev->port[vport_index].rep = rep; rep->rep_data[REP_IB].priv = ibdev; - write_lock(&ibdev->port[vport_index].roce.netdev_lock); - ibdev->port[vport_index].roce.netdev = - mlx5_ib_get_rep_netdev(rep->esw, rep->vport); - write_unlock(&ibdev->port[vport_index].roce.netdev_lock); + ndev = mlx5_ib_get_rep_netdev(rep->esw, rep->vport); - return 0; + return ib_device_set_netdev(&ibdev->ib_dev, ndev, vport_index + 1); } static void mlx5_ib_register_peer_vport_reps(struct mlx5_core_dev *mdev); @@ -104,11 +102,15 @@ mlx5_ib_vport_rep_load(struct mlx5_core_dev *dev, struct mlx5_eswitch_rep *rep) ibdev->is_rep = true; vport_index = rep->vport_index; ibdev->port[vport_index].rep = rep; - ibdev->ib_dev.phys_port_cnt = num_ports; - ibdev->port[vport_index].roce.netdev = - mlx5_ib_get_rep_netdev(lag_master->priv.eswitch, rep->vport); ibdev->mdev = lag_master; ibdev->num_ports = num_ports; + ibdev->ib_dev.phys_port_cnt = num_ports; + ret = ib_device_set_netdev(&ibdev->ib_dev, + mlx5_ib_get_rep_netdev(lag_master->priv.eswitch, + rep->vport), + vport_index + 1); + if (ret) + goto fail_add; ret = __mlx5_ib_add(ibdev, profile); if (ret) @@ -161,9 +163,8 @@ mlx5_ib_vport_rep_unload(struct mlx5_eswitch_rep *rep) } port = &dev->port[vport_index]; - write_lock(&port->roce.netdev_lock); - port->roce.netdev = NULL; - write_unlock(&port->roce.netdev_lock); + + ib_device_set_netdev(&dev->ib_dev, NULL, vport_index + 1); rep->rep_data[REP_IB].priv = NULL; port->rep = NULL; diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c index ad8a2b5517bf..4999239c8f41 100644 --- a/drivers/infiniband/hw/mlx5/main.c +++ b/drivers/infiniband/hw/mlx5/main.c @@ -147,16 +147,52 @@ static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev, if (upper && port->rep->vport == MLX5_VPORT_UPLINK) continue; - - read_lock(&port->roce.netdev_lock); - rep_ndev = mlx5_ib_get_rep_netdev(port->rep->esw, - port->rep->vport); - if (rep_ndev == ndev) { - read_unlock(&port->roce.netdev_lock); + rep_ndev = ib_device_get_netdev(&dev->ib_dev, i + 1); + if (rep_ndev && rep_ndev == ndev) { + dev_put(rep_ndev); *port_num = i + 1; return &port->roce; } - read_unlock(&port->roce.netdev_lock); + + dev_put(rep_ndev); + } + + return NULL; +} + +static bool mlx5_netdev_send_event(struct mlx5_ib_dev *dev, + struct net_device *ndev, + struct net_device *upper, + struct net_device *ib_ndev) +{ + if (!dev->ib_active) + return false; + + /* Event is about our upper device */ + if (upper == ndev) + return true; + + /* RDMA device is not in lag and not in switchdev */ + if (!dev->is_rep && !upper && ndev == ib_ndev) + return true; + + /* RDMA devie is in switchdev */ + if (dev->is_rep && ndev == ib_ndev) + return true; + + return false; +} + +static struct net_device *mlx5_ib_get_rep_uplink_netdev(struct mlx5_ib_dev *ibdev) +{ + struct mlx5_ib_port *port; + int i; + + for (i = 0; i < ibdev->num_ports; i++) { + port = &ibdev->port[i]; + if (port->rep && port->rep->vport == MLX5_VPORT_UPLINK) { + return ib_device_get_netdev(&ibdev->ib_dev, i + 1); + } } return NULL; @@ -168,6 +204,7 @@ static int mlx5_netdev_event(struct notifier_block *this, struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb); struct net_device *ndev = netdev_notifier_info_to_dev(ptr); u32 port_num = roce->native_port_num; + struct net_device *ib_ndev = NULL; struct mlx5_core_dev *mdev; struct mlx5_ib_dev *ibdev; @@ -181,29 +218,38 @@ static int mlx5_netdev_event(struct notifier_block *this, /* Should already be registered during the load */ if (ibdev->is_rep) break; - write_lock(&roce->netdev_lock); + + ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); + /* Exit if already registered */ + if (ib_ndev) + goto put_ndev; + if (ndev->dev.parent == mdev->device) - roce->netdev = ndev; - write_unlock(&roce->netdev_lock); + ib_device_set_netdev(&ibdev->ib_dev, ndev, port_num); break; case NETDEV_UNREGISTER: /* In case of reps, ib device goes away before the netdevs */ - write_lock(&roce->netdev_lock); - if (roce->netdev == ndev) - roce->netdev = NULL; - write_unlock(&roce->netdev_lock); - break; + if (ibdev->is_rep) + break; + ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); + if (ib_ndev == ndev) + ib_device_set_netdev(&ibdev->ib_dev, NULL, port_num); + goto put_ndev; case NETDEV_CHANGE: case NETDEV_UP: case NETDEV_DOWN: { struct net_device *upper = NULL; - if (mlx5_lag_is_roce(mdev)) { + if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) { struct net_device *lag_ndev; - lag_ndev = mlx5_lag_get_roce_netdev(mdev); + if(mlx5_lag_is_roce(mdev)) + lag_ndev = ib_device_get_netdev(&ibdev->ib_dev, 1); + else /* sriov lag */ + lag_ndev = mlx5_ib_get_rep_uplink_netdev(ibdev); + if (lag_ndev) { upper = netdev_master_upper_dev_get(lag_ndev); dev_put(lag_ndev); @@ -216,18 +262,19 @@ static int mlx5_netdev_event(struct notifier_block *this, roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num); if (!roce) return NOTIFY_DONE; - if ((upper == ndev || - ((!upper || ibdev->is_rep) && ndev == roce->netdev)) && - ibdev->ib_active) { + + ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num); + + if (mlx5_netdev_send_event(ibdev, ndev, upper, ib_ndev)) { struct ib_event ibev = { }; enum ib_port_state port_state; if (get_port_state(&ibdev->ib_dev, port_num, &port_state)) - goto done; + goto put_ndev; if (roce->last_port_state == port_state) - goto done; + goto put_ndev; roce->last_port_state = port_state; ibev.device = &ibdev->ib_dev; @@ -236,7 +283,7 @@ static int mlx5_netdev_event(struct notifier_block *this, else if (port_state == IB_PORT_ACTIVE) ibev.event = IB_EVENT_PORT_ACTIVE; else - goto done; + goto put_ndev; ibev.element.port_num = port_num; ib_dispatch_event(&ibev); @@ -247,39 +294,13 @@ static int mlx5_netdev_event(struct notifier_block *this, default: break; } +put_ndev: + dev_put(ib_ndev); done: mlx5_ib_put_native_port_mdev(ibdev, port_num); return NOTIFY_DONE; } -static struct net_device *mlx5_ib_get_netdev(struct ib_device *device, - u32 port_num) -{ - struct mlx5_ib_dev *ibdev = to_mdev(device); - struct net_device *ndev; - struct mlx5_core_dev *mdev; - - mdev = mlx5_ib_get_native_port_mdev(ibdev, port_num, NULL); - if (!mdev) - return NULL; - - if (mlx5_lag_is_roce(mdev)) { - ndev = mlx5_lag_get_roce_netdev(mdev); - goto out; - } - - /* Ensure ndev does not disappear before we invoke dev_hold() - */ - read_lock(&ibdev->port[port_num - 1].roce.netdev_lock); - ndev = ibdev->port[port_num - 1].roce.netdev; - dev_hold(ndev); - read_unlock(&ibdev->port[port_num - 1].roce.netdev_lock); - -out: - mlx5_ib_put_native_port_mdev(ibdev, port_num); - return ndev; -} - struct mlx5_core_dev *mlx5_ib_get_native_port_mdev(struct mlx5_ib_dev *ibdev, u32 ib_port_num, u32 *native_port_num) @@ -554,7 +575,7 @@ static int mlx5_query_port_roce(struct ib_device *device, u32 port_num, if (!put_mdev) goto out; - ndev = mlx5_ib_get_netdev(device, port_num); + ndev = ib_device_get_netdev(device, port_num); if (!ndev) goto out; @@ -3185,6 +3206,60 @@ static void get_dev_fw_str(struct ib_device *ibdev, char *str) fw_rev_sub(dev->mdev)); } +static int lag_event(struct notifier_block *nb, unsigned long event, void *data) +{ + struct mlx5_ib_dev *dev = container_of(nb, struct mlx5_ib_dev, + lag_events); + struct mlx5_core_dev *mdev = dev->mdev; + struct mlx5_ib_port *port; + struct net_device *ndev; + int i, err; + int portnum; + + portnum = 0; + switch (event) { + case MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE: + ndev = data; + if (ndev) { + if (!mlx5_lag_is_roce(mdev)) { + // sriov lag + for (i = 0; i < dev->num_ports; i++) { + port = &dev->port[i]; + if (port->rep && port->rep->vport == + MLX5_VPORT_UPLINK) { + portnum = i; + break; + } + } + } + err = ib_device_set_netdev(&dev->ib_dev, ndev, + portnum + 1); + dev_put(ndev); + if (err) + return err; + /* Rescan gids after new netdev assignment */ + rdma_roce_rescan_device(&dev->ib_dev); + } + break; + default: + return NOTIFY_DONE; + } + return NOTIFY_OK; +} + +static void mlx5e_lag_event_register(struct mlx5_ib_dev *dev) +{ + dev->lag_events.notifier_call = lag_event; + blocking_notifier_chain_register(&dev->mdev->priv.lag_nh, + &dev->lag_events); +} + +static void mlx5e_lag_event_unregister(struct mlx5_ib_dev *dev) +{ + blocking_notifier_chain_unregister(&dev->mdev->priv.lag_nh, + &dev->lag_events); +} + static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev) { struct mlx5_core_dev *mdev = dev->mdev; @@ -3206,6 +3281,7 @@ static int mlx5_eth_lag_init(struct mlx5_ib_dev *dev) goto err_destroy_vport_lag; } + mlx5e_lag_event_register(dev); dev->flow_db->lag_demux_ft = ft; dev->lag_ports = mlx5_lag_get_num_ports(mdev); dev->lag_active = true; @@ -3223,6 +3299,7 @@ static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev) if (dev->lag_active) { dev->lag_active = false; + mlx5e_lag_event_unregister(dev); mlx5_destroy_flow_table(dev->flow_db->lag_demux_ft); dev->flow_db->lag_demux_ft = NULL; @@ -3939,7 +4016,6 @@ static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev) for (i = 0; i < dev->num_ports; i++) { spin_lock_init(&dev->port[i].mp.mpi_lock); - rwlock_init(&dev->port[i].roce.netdev_lock); dev->port[i].roce.dev = dev; dev->port[i].roce.native_port_num = i + 1; dev->port[i].roce.last_port_state = IB_PORT_DOWN; @@ -4204,7 +4280,6 @@ static const struct ib_device_ops mlx5_ib_dev_common_roce_ops = { .create_wq = mlx5_ib_create_wq, .destroy_rwq_ind_table = mlx5_ib_destroy_rwq_ind_table, .destroy_wq = mlx5_ib_destroy_wq, - .get_netdev = mlx5_ib_get_netdev, .modify_wq = mlx5_ib_modify_wq, INIT_RDMA_OBJ_SIZE(ib_rwq_ind_table, mlx5_ib_rwq_ind_table, diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index 59ce407ce505..23fd72f7f63d 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -888,8 +888,6 @@ struct mlx5_roce { /* Protect mlx5_ib_get_netdev from invoking dev_hold() with a NULL * netdev pointer */ - rwlock_t netdev_lock; - struct net_device *netdev; struct notifier_block nb; struct netdev_net_notifier nn; struct notifier_block mdev_nb; @@ -1138,6 +1136,7 @@ struct mlx5_ib_dev { /* protect accessing data_direct_dev */ struct mutex data_direct_lock; struct notifier_block mdev_events; + struct notifier_block lag_events; int num_ports; /* serialize update of capability mask */ diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c index cf8045b92689..8577db3308cc 100644 --- a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c +++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c @@ -445,6 +445,34 @@ static int _mlx5_modify_lag(struct mlx5_lag *ldev, u8 *ports) return mlx5_cmd_modify_lag(dev0, ldev->ports, ports); } +static struct net_device *mlx5_lag_active_backup_get_netdev(struct mlx5_core_dev *dev) +{ + struct net_device *ndev = NULL; + struct mlx5_lag *ldev; + unsigned long flags; + int i; + + spin_lock_irqsave(&lag_lock, flags); + ldev = mlx5_lag_dev(dev); + + if (!ldev) + goto unlock; + + for (i = 0; i < ldev->ports; i++) + if (ldev->tracker.netdev_state[i].tx_enabled) + ndev = ldev->pf[i].netdev; + if (!ndev) + ndev = ldev->pf[ldev->ports - 1].netdev; + + if (ndev) + dev_hold(ndev); + +unlock: + spin_unlock_irqrestore(&lag_lock, flags); + + return ndev; +} + void mlx5_modify_lag(struct mlx5_lag *ldev, struct lag_tracker *tracker) { @@ -477,9 +505,18 @@ void mlx5_modify_lag(struct mlx5_lag *ldev, } } - if (tracker->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP && - !(ldev->mode == MLX5_LAG_MODE_ROCE)) - mlx5_lag_drop_rule_setup(ldev, tracker); + if (tracker->tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) { + struct net_device *ndev = mlx5_lag_active_backup_get_netdev(dev0); + + if(!(ldev->mode == MLX5_LAG_MODE_ROCE)) + mlx5_lag_drop_rule_setup(ldev, tracker); + /** Only sriov and roce lag should have tracker->tx_type set so + * no need to check the mode + */ + blocking_notifier_call_chain(&dev0->priv.lag_nh, + MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE, + ndev); + } } static int mlx5_lag_set_port_sel_mode_roce(struct mlx5_lag *ldev, @@ -613,6 +650,7 @@ static int mlx5_create_lag(struct mlx5_lag *ldev, mlx5_core_err(dev0, "Failed to deactivate RoCE LAG; driver restart required\n"); } + BLOCKING_INIT_NOTIFIER_HEAD(&dev0->priv.lag_nh); return err; } @@ -1492,38 +1530,6 @@ void mlx5_lag_enable_change(struct mlx5_core_dev *dev) mlx5_queue_bond_work(ldev, 0); } -struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev) -{ - struct net_device *ndev = NULL; - struct mlx5_lag *ldev; - unsigned long flags; - int i; - - spin_lock_irqsave(&lag_lock, flags); - ldev = mlx5_lag_dev(dev); - - if (!(ldev && __mlx5_lag_is_roce(ldev))) - goto unlock; - - if (ldev->tracker.tx_type == NETDEV_LAG_TX_TYPE_ACTIVEBACKUP) { - for (i = 0; i < ldev->ports; i++) - if (ldev->tracker.netdev_state[i].tx_enabled) - ndev = ldev->pf[i].netdev; - if (!ndev) - ndev = ldev->pf[ldev->ports - 1].netdev; - } else { - ndev = ldev->pf[MLX5_LAG_P1].netdev; - } - if (ndev) - dev_hold(ndev); - -unlock: - spin_unlock_irqrestore(&lag_lock, flags); - - return ndev; -} -EXPORT_SYMBOL(mlx5_lag_get_roce_netdev); - u8 mlx5_lag_get_slave_port(struct mlx5_core_dev *dev, struct net_device *slave) { diff --git a/include/linux/mlx5/device.h b/include/linux/mlx5/device.h index 57c9b18c3adb..dac33cfe9c0c 100644 --- a/include/linux/mlx5/device.h +++ b/include/linux/mlx5/device.h @@ -371,6 +371,7 @@ enum mlx5_driver_event { MLX5_DRIVER_EVENT_SF_PEER_DEVLINK, MLX5_DRIVER_EVENT_AFFILIATION_DONE, MLX5_DRIVER_EVENT_AFFILIATION_REMOVED, + MLX5_DRIVER_EVENT_ACTIVE_BACKUP_LAG_CHANGE_LOWERSTATE, }; enum { diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h index a96438ded15f..46a7a3d11048 100644 --- a/include/linux/mlx5/driver.h +++ b/include/linux/mlx5/driver.h @@ -643,6 +643,7 @@ struct mlx5_priv { struct mlx5_sf_hw_table *sf_hw_table; struct mlx5_sf_table *sf_table; #endif + struct blocking_notifier_head lag_nh; }; enum mlx5_device_state { @@ -1181,7 +1182,6 @@ bool mlx5_lag_mode_is_hash(struct mlx5_core_dev *dev); bool mlx5_lag_is_master(struct mlx5_core_dev *dev); bool mlx5_lag_is_shared_fdb(struct mlx5_core_dev *dev); bool mlx5_lag_is_mpesw(struct mlx5_core_dev *dev); -struct net_device *mlx5_lag_get_roce_netdev(struct mlx5_core_dev *dev); u8 mlx5_lag_get_slave_port(struct mlx5_core_dev *dev, struct net_device *slave); int mlx5_lag_query_cong_counters(struct mlx5_core_dev *dev, diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index a1dcf812d787..aa8ede439905 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -4453,6 +4453,8 @@ struct net_device *ib_get_net_dev_by_params(struct ib_device *dev, u32 port, const struct sockaddr *addr); int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev, unsigned int port); +struct net_device *ib_device_get_netdev(struct ib_device *ib_dev, + u32 port); struct ib_wq *ib_create_wq(struct ib_pd *pd, struct ib_wq_init_attr *init_attr); int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata); -- cgit From 9cbed5aab5aeea420d0aa945733bf608449d44fb Mon Sep 17 00:00:00 2001 From: Chiara Meiohas Date: Mon, 9 Sep 2024 20:30:24 +0300 Subject: RDMA/nldev: Add support for RDMA monitoring Introduce a new netlink command to allow rdma event monitoring. The rdma events supported now are IB device registration/unregistration and net device attachment/detachment. Example output of rdma monitor and the commands which trigger the events: $ rdma monitor $ rmmod mlx5_ib [UNREGISTER] dev 1 rocep8s0f1 [UNREGISTER] dev 0 rocep8s0f0 $ modprobe mlx5_ib [REGISTER] dev 2 mlx5_0 [NETDEV_ATTACH] dev 2 mlx5_0 port 1 netdev 4 eth2 [REGISTER] dev 3 mlx5_1 [NETDEV_ATTACH] dev 3 mlx5_1 port 1 netdev 5 eth3 $ devlink dev eswitch set pci/0000:08:00.0 mode switchdev [UNREGISTER] dev 2 rocep8s0f0 [REGISTER] dev 4 mlx5_0 [NETDEV_ATTACH] dev 4 mlx5_0 port 30 netdev 4 eth2 $ echo 4 > /sys/class/net/eth2/device/sriov_numvfs [NETDEV_ATTACH] dev 4 rdmap8s0f0 port 2 netdev 7 eth4 [NETDEV_ATTACH] dev 4 rdmap8s0f0 port 3 netdev 8 eth5 [NETDEV_ATTACH] dev 4 rdmap8s0f0 port 4 netdev 9 eth6 [NETDEV_ATTACH] dev 4 rdmap8s0f0 port 5 netdev 10 eth7 [REGISTER] dev 5 mlx5_0 [NETDEV_ATTACH] dev 5 mlx5_0 port 1 netdev 11 eth8 [REGISTER] dev 6 mlx5_0 [NETDEV_ATTACH] dev 6 mlx5_0 port 1 netdev 12 eth9 [REGISTER] dev 7 mlx5_0 [NETDEV_ATTACH] dev 7 mlx5_0 port 1 netdev 13 eth10 [REGISTER] dev 8 mlx5_0 [NETDEV_ATTACH] dev 8 mlx5_0 port 1 netdev 14 eth11 $ echo 0 > /sys/class/net/eth2/device/sriov_numvfs [UNREGISTER] dev 5 rocep8s0f0v0 [UNREGISTER] dev 6 rocep8s0f0v1 [UNREGISTER] dev 7 rocep8s0f0v2 [UNREGISTER] dev 8 rocep8s0f0v3 [NETDEV_DETACH] dev 4 rdmap8s0f0 port 2 [NETDEV_DETACH] dev 4 rdmap8s0f0 port 3 [NETDEV_DETACH] dev 4 rdmap8s0f0 port 4 [NETDEV_DETACH] dev 4 rdmap8s0f0 port 5 Signed-off-by: Chiara Meiohas Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909173025.30422-7-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/device.c | 35 +++++++++++ drivers/infiniband/core/netlink.c | 1 + drivers/infiniband/core/nldev.c | 124 ++++++++++++++++++++++++++++++++++++++ include/rdma/rdma_netlink.h | 12 ++++ include/uapi/rdma/rdma_netlink.h | 15 +++++ 5 files changed, 187 insertions(+) (limited to 'include') diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c index 9e765c79a892..e029401b5680 100644 --- a/drivers/infiniband/core/device.c +++ b/drivers/infiniband/core/device.c @@ -1351,6 +1351,29 @@ static void prevent_dealloc_device(struct ib_device *ib_dev) { } +static void ib_device_notify_register(struct ib_device *device) +{ + struct net_device *netdev; + u32 port; + int ret; + + ret = rdma_nl_notify_event(device, 0, RDMA_REGISTER_EVENT); + if (ret) + return; + + rdma_for_each_port(device, port) { + netdev = ib_device_get_netdev(device, port); + if (!netdev) + continue; + + ret = rdma_nl_notify_event(device, port, + RDMA_NETDEV_ATTACH_EVENT); + dev_put(netdev); + if (ret) + return; + } +} + /** * ib_register_device - Register an IB device with IB core * @device: Device to register @@ -1449,6 +1472,8 @@ int ib_register_device(struct ib_device *device, const char *name, dev_set_uevent_suppress(&device->dev, false); /* Mark for userspace that device is ready */ kobject_uevent(&device->dev.kobj, KOBJ_ADD); + + ib_device_notify_register(device); ib_device_put(device); return 0; @@ -1491,6 +1516,7 @@ static void __ib_unregister_device(struct ib_device *ib_dev) goto out; disable_device(ib_dev); + rdma_nl_notify_event(ib_dev, 0, RDMA_UNREGISTER_EVENT); /* Expedite removing unregistered pointers from the hash table */ free_netdevs(ib_dev); @@ -2159,6 +2185,7 @@ static void add_ndev_hash(struct ib_port_data *pdata) int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev, u32 port) { + enum rdma_nl_notify_event_type etype; struct net_device *old_ndev; struct ib_port_data *pdata; unsigned long flags; @@ -2190,6 +2217,14 @@ int ib_device_set_netdev(struct ib_device *ib_dev, struct net_device *ndev, spin_unlock_irqrestore(&pdata->netdev_lock, flags); add_ndev_hash(pdata); + + /* Make sure that the device is registered before we send events */ + if (xa_load(&devices, ib_dev->index) != ib_dev) + return 0; + + etype = ndev ? RDMA_NETDEV_ATTACH_EVENT : RDMA_NETDEV_DETACH_EVENT; + rdma_nl_notify_event(ib_dev, port, etype); + return 0; } EXPORT_SYMBOL(ib_device_set_netdev); diff --git a/drivers/infiniband/core/netlink.c b/drivers/infiniband/core/netlink.c index ae2db0c70788..def14c54b648 100644 --- a/drivers/infiniband/core/netlink.c +++ b/drivers/infiniband/core/netlink.c @@ -311,6 +311,7 @@ int rdma_nl_net_init(struct rdma_dev_net *rnet) struct net *net = read_pnet(&rnet->net); struct netlink_kernel_cfg cfg = { .input = rdma_nl_rcv, + .flags = NL_CFG_F_NONROOT_RECV, }; struct sock *nls; diff --git a/drivers/infiniband/core/nldev.c b/drivers/infiniband/core/nldev.c index 4d4a1f90e484..70b3fa0469f2 100644 --- a/drivers/infiniband/core/nldev.c +++ b/drivers/infiniband/core/nldev.c @@ -170,6 +170,7 @@ static const struct nla_policy nldev_policy[RDMA_NLDEV_ATTR_MAX] = { [RDMA_NLDEV_ATTR_DEV_TYPE] = { .type = NLA_U8 }, [RDMA_NLDEV_ATTR_PARENT_NAME] = { .type = NLA_NUL_STRING }, [RDMA_NLDEV_ATTR_NAME_ASSIGN_TYPE] = { .type = NLA_U8 }, + [RDMA_NLDEV_ATTR_EVENT_TYPE] = { .type = NLA_U8 }, }; static int put_driver_name_print_type(struct sk_buff *msg, const char *name, @@ -2722,6 +2723,129 @@ static const struct rdma_nl_cbs nldev_cb_table[RDMA_NLDEV_NUM_OPS] = { }, }; +static int fill_mon_netdev_association(struct sk_buff *msg, + struct ib_device *device, u32 port, + const struct net *net) +{ + struct net_device *netdev = ib_device_get_netdev(device, port); + int ret = 0; + + if (netdev && !net_eq(dev_net(netdev), net)) + goto out; + + ret = nla_put_u32(msg, RDMA_NLDEV_ATTR_DEV_INDEX, device->index); + if (ret) + goto out; + + ret = nla_put_string(msg, RDMA_NLDEV_ATTR_DEV_NAME, + dev_name(&device->dev)); + if (ret) + goto out; + + ret = nla_put_u32(msg, RDMA_NLDEV_ATTR_PORT_INDEX, port); + if (ret) + goto out; + + if (netdev) { + ret = nla_put_u32(msg, + RDMA_NLDEV_ATTR_NDEV_INDEX, netdev->ifindex); + if (ret) + goto out; + + ret = nla_put_string(msg, + RDMA_NLDEV_ATTR_NDEV_NAME, netdev->name); + } + +out: + dev_put(netdev); + return ret; +} + +static void rdma_nl_notify_err_msg(struct ib_device *device, u32 port_num, + enum rdma_nl_notify_event_type type) +{ + struct net_device *netdev; + + switch (type) { + case RDMA_REGISTER_EVENT: + dev_warn_ratelimited(&device->dev, + "Failed to send RDMA monitor register device event\n"); + break; + case RDMA_UNREGISTER_EVENT: + dev_warn_ratelimited(&device->dev, + "Failed to send RDMA monitor unregister device event\n"); + break; + case RDMA_NETDEV_ATTACH_EVENT: + netdev = ib_device_get_netdev(device, port_num); + dev_warn_ratelimited(&device->dev, + "Failed to send RDMA monitor netdev attach event: port %d netdev %d\n", + port_num, netdev->ifindex); + dev_put(netdev); + break; + case RDMA_NETDEV_DETACH_EVENT: + dev_warn_ratelimited(&device->dev, + "Failed to send RDMA monitor netdev detach event: port %d\n", + port_num); + default: + break; + } +} + +int rdma_nl_notify_event(struct ib_device *device, u32 port_num, + enum rdma_nl_notify_event_type type) +{ + struct sk_buff *skb; + struct net *net; + int ret = 0; + void *nlh; + + net = read_pnet(&device->coredev.rdma_net); + if (!net) + return -EINVAL; + + skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); + if (!skb) + return -ENOMEM; + nlh = nlmsg_put(skb, 0, 0, + RDMA_NL_GET_TYPE(RDMA_NL_NLDEV, RDMA_NLDEV_CMD_MONITOR), + 0, 0); + + switch (type) { + case RDMA_REGISTER_EVENT: + case RDMA_UNREGISTER_EVENT: + ret = fill_nldev_handle(skb, device); + if (ret) + goto err_free; + break; + case RDMA_NETDEV_ATTACH_EVENT: + case RDMA_NETDEV_DETACH_EVENT: + ret = fill_mon_netdev_association(skb, device, + port_num, net); + if (ret) + goto err_free; + break; + default: + break; + } + + ret = nla_put_u8(skb, RDMA_NLDEV_ATTR_EVENT_TYPE, type); + if (ret) + goto err_free; + + nlmsg_end(skb, nlh); + ret = rdma_nl_multicast(net, skb, RDMA_NL_GROUP_NOTIFY, GFP_KERNEL); + if (ret && ret != -ESRCH) { + skb = NULL; /* skb is freed in the netlink send-op handling */ + goto err_free; + } + return 0; + +err_free: + rdma_nl_notify_err_msg(device, port_num, type); + nlmsg_free(skb); + return ret; +} + void __init nldev_init(void) { rdma_nl_register(RDMA_NL_NLDEV, nldev_cb_table); diff --git a/include/rdma/rdma_netlink.h b/include/rdma/rdma_netlink.h index c2a79aeee113..326deaf56d5d 100644 --- a/include/rdma/rdma_netlink.h +++ b/include/rdma/rdma_netlink.h @@ -6,6 +6,8 @@ #include #include +struct ib_device; + enum { RDMA_NLDEV_ATTR_EMPTY_STRING = 1, RDMA_NLDEV_ATTR_ENTRY_STRLEN = 16, @@ -110,6 +112,16 @@ int rdma_nl_multicast(struct net *net, struct sk_buff *skb, */ bool rdma_nl_chk_listeners(unsigned int group); +/** + * Prepare and send an event message + * @ib: the IB device which triggered the event + * @port_num: the port number which triggered the event - 0 if unused + * @type: the event type + * Returns 0 on success or a negative error code + */ +int rdma_nl_notify_event(struct ib_device *ib, u32 port_num, + enum rdma_nl_notify_event_type type); + struct rdma_link_ops { struct list_head list; const char *type; diff --git a/include/uapi/rdma/rdma_netlink.h b/include/uapi/rdma/rdma_netlink.h index 2f37568f5556..5f9636d26050 100644 --- a/include/uapi/rdma/rdma_netlink.h +++ b/include/uapi/rdma/rdma_netlink.h @@ -15,6 +15,7 @@ enum { enum { RDMA_NL_GROUP_IWPM = 2, RDMA_NL_GROUP_LS, + RDMA_NL_GROUP_NOTIFY, RDMA_NL_NUM_GROUPS }; @@ -305,6 +306,8 @@ enum rdma_nldev_command { RDMA_NLDEV_CMD_DELDEV, + RDMA_NLDEV_CMD_MONITOR, + RDMA_NLDEV_NUM_OPS }; @@ -574,6 +577,8 @@ enum rdma_nldev_attr { RDMA_NLDEV_ATTR_NAME_ASSIGN_TYPE, /* u8 */ + RDMA_NLDEV_ATTR_EVENT_TYPE, /* u8 */ + /* * Always the end */ @@ -624,4 +629,14 @@ enum rdma_nl_name_assign_type { RDMA_NAME_ASSIGN_TYPE_USER = 1, /* Provided by user-space */ }; +/* + * Supported rdma monitoring event types. + */ +enum rdma_nl_notify_event_type { + RDMA_REGISTER_EVENT, + RDMA_UNREGISTER_EVENT, + RDMA_NETDEV_ATTACH_EVENT, + RDMA_NETDEV_DETACH_EVENT, +}; + #endif /* _UAPI_RDMA_NETLINK_H */ -- cgit From 12fb1153c53bf9b53e299c9775b84fa7838640f7 Mon Sep 17 00:00:00 2001 From: Chiara Meiohas Date: Mon, 9 Sep 2024 20:30:25 +0300 Subject: RDMA/nldev: Expose whether RDMA monitoring is supported Extend the "rdma sys" command to display whether RDMA monitoring is supported. RDMA monitoring is not supported in mlx4 because it does not use the ib_device_set_netdev() API, which sends the RDMA events. Example output for kernel where monitoring is supported: $ rdma sys show netns shared privileged-qkey off monitor on copy-on-fork on Example output for kernel where monitoring is not supported: $ rdma sys show netns shared privileged-qkey off monitor off copy-on-fork on Signed-off-by: Chiara Meiohas Signed-off-by: Michael Guralnik Link: https://patch.msgid.link/20240909173025.30422-8-michaelgur@nvidia.com Signed-off-by: Leon Romanovsky --- drivers/infiniband/core/nldev.c | 6 ++++++ include/uapi/rdma/rdma_netlink.h | 1 + 2 files changed, 7 insertions(+) (limited to 'include') diff --git a/drivers/infiniband/core/nldev.c b/drivers/infiniband/core/nldev.c index 70b3fa0469f2..10b1411ac53d 100644 --- a/drivers/infiniband/core/nldev.c +++ b/drivers/infiniband/core/nldev.c @@ -1952,6 +1952,12 @@ static int nldev_sys_get_doit(struct sk_buff *skb, struct nlmsghdr *nlh, nlmsg_free(msg); return err; } + + err = nla_put_u8(msg, RDMA_NLDEV_SYS_ATTR_MONITOR_MODE, 1); + if (err) { + nlmsg_free(msg); + return err; + } /* * Copy-on-fork is supported. * See commits: diff --git a/include/uapi/rdma/rdma_netlink.h b/include/uapi/rdma/rdma_netlink.h index 5f9636d26050..39be09c0ffbb 100644 --- a/include/uapi/rdma/rdma_netlink.h +++ b/include/uapi/rdma/rdma_netlink.h @@ -579,6 +579,7 @@ enum rdma_nldev_attr { RDMA_NLDEV_ATTR_EVENT_TYPE, /* u8 */ + RDMA_NLDEV_SYS_ATTR_MONITOR_MODE, /* u8 */ /* * Always the end */ -- cgit