RDMA/rxe: Add dma-buf support for Soft-RoCE
Subsystems (2)
cover RDMA/rxe: Add dma-buf support for Soft-RoCE
This patchset introduces dma-buf support for the Soft-RoCE (RXE) driver. By enabling dma-buf, RXE can now participate in zero-copy data transfers with other providers (such as GPUs) that export memory via dma-buf fds.
Traditionally, RXE only supported user-space memory regions (UMEM) based on system RAM. This change extends RXE’s capability to handle peer-to-peer (P2P) like workflows in a software-defined RDMA environment.
This patchset pass the rdma-core tests with the following link for the rdma-core:
https://github.com/linux-rdma/rdma-core/pull/1055
Zhu Yanjun (2):
RDMA/umem: Change for rdma devices has not dma device RDMA/rxe: Add dma-buf support
drivers/infiniband/core/umem_dmabuf.c | 35 ++++++++++- drivers/infiniband/sw/rxe/rxe.c | 2 + drivers/infiniband/sw/rxe/rxe_loc.h | 2 + drivers/infiniband/sw/rxe/rxe_mr.c | 89 ++++++++++++++++++++++++--- drivers/infiniband/sw/rxe/rxe_odp.c | 2 +- drivers/infiniband/sw/rxe/rxe_verbs.c | 40 ++++++++++++ drivers/infiniband/sw/rxe/rxe_verbs.h | 2 +- include/rdma/ib_umem.h | 1 + 8 files changed, 161 insertions(+), 12 deletions(-)
Signature · 1 line
2.53.0
PATCH 1/2 RDMA/umem: Change for rdma devices has not dma device
Current implementation requires a dma device for RDMA driver to use dma-buf memory space as RDMA buffer.
--- drivers/infiniband/core/umem_dmabuf.c | 35 ++++++++++++++++++++++++++- include/rdma/ib_umem.h | 1 + 2 files changed, 35 insertions(+), 1 deletion(-)
drivers/infiniband/core/umem_dmabuf.c +34 -1
@@ -142,6 +142,8 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device, goto out_release_dmabuf; } + umem_dmabuf->dmabuf = dmabuf; + umem = &umem_dmabuf->umem; umem->ibdev = device; umem->length = size;
@@ -152,6 +154,24 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,if (!ib_umem_num_pages (umem)) goto out_free_umem; +/* Software RDMA drivers has not dma device. Just get dmabuf from fd */ +if (!device->dma_device) { +struct sg_table *sgt; + +dma_resv_lock (dmabuf->resv,NULL ); + sgt = dmabuf->ops->map_dma_buf (NULL ,DMA_BIDIRECTIONAL ); +dma_resv_unlock (dmabuf->resv); +if (IS_ERR (sgt)) { + ret =ERR_CAST (sgt); + goto out_free_umem; + } + umem_dmabuf->sgt = sgt; + goto done; + } + +if (unlikely (!ops || !ops->move_notify)) + goto out_free_umem; + umem_dmabuf->attach = dma_buf_dynamic_attach( dmabuf, dma_device,
@@ -161,6 +181,7 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device, ret =ERR_CAST (umem_dmabuf->attach); goto out_free_umem; } +done:return umem_dmabuf; out_free_umem:
@@ -260,11 +281,23 @@ 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; +struct dma_buf *dmabuf = umem_dmabuf->dmabuf; + +if (!umem_dmabuf->attach) { +if (umem_dmabuf->sgt) { +dma_resv_lock (dmabuf->resv,NULL ); + dmabuf->ops->unmap_dma_buf (NULL , umem_dmabuf->sgt, +DMA_BIDIRECTIONAL ); +dma_resv_unlock (dmabuf->resv); + } + goto free_dmabuf; + }ib_umem_dmabuf_revoke (umem_dmabuf);dma_buf_detach (dmabuf, umem_dmabuf->attach); + +free_dmabuf:dma_buf_put (dmabuf);kfree (umem_dmabuf); }
include/rdma/ib_umem.h +1 -0
@@ -31,6 +31,7 @@ struct ib_umem {struct ib_umem_dmabuf {struct ib_umem umem;struct dma_buf_attachment *attach; +struct dma_buf *dmabuf;struct sg_table *sgt;struct scatterlist *first_sg;struct scatterlist *last_sg;
Signature · 1 line
2.53.0
Re: [PATCH 1/2] RDMA/umem: Change for rdma devices has not dma device
Quoted 23 lines · Zhu Yanjun
Current implementation requires a dma device for RDMA driver to use dma-buf memory space as RDMA buffer.
Signed-off-by Zhu Yanjun <[email protected]>--- drivers/infiniband/core/umem_dmabuf.c | 35 ++++++++++++++++++++++++++- include/rdma/ib_umem.h | 1 + 2 files changed, 35 insertions(+), 1 deletion(-)drivers/infiniband/core/umem_dmabuf.c +2
-0@@ -142,6 +142,8 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device, goto out_release_dmabuf; } + umem_dmabuf->dmabuf = dmabuf; + umem = &umem_dmabuf->umem; umem->ibdev = device; umem->length = size;
drivers/infiniband/core/umem_dmabuf.c ib_umem_dmabuf_get_with_dma_device(struct ib_device *device, · line 158
Quoted 5 lines
drivers/infiniband/core/umem_dmabuf.c +2
-0@@ -152,6 +154,24 @@ ib_umem_dmabuf_get_with_dma_device(struct ib_device *device,if (!ib_umem_num_pages (umem)) goto out_free_umem; +/* Software RDMA drivers has not dma device. Just get dmabuf from fd */ +if (!device->dma_device) {
I understand your intent, but RXE cannot support dmabuf, as dmabuf requires a real DMA‑capable device to be attached. In addition, it is unclear how you expect PCI peer‑to‑peer operations to work when one side (RXE) is not a PCI device at all.
Thanks
PATCH 2/2 RDMA/rxe: Add dma-buf support
In reply to Zhu YanjunImplement a ib device operation ‘reg_user_mr_dmabuf’. Generate a rxe_map from the memory space linked the passed dma-buf.
--- drivers/infiniband/sw/rxe/rxe.c | 2 + drivers/infiniband/sw/rxe/rxe_loc.h | 2 + drivers/infiniband/sw/rxe/rxe_mr.c | 89 ++++++++++++++++++++++++--- drivers/infiniband/sw/rxe/rxe_odp.c | 2 +- drivers/infiniband/sw/rxe/rxe_verbs.c | 40 ++++++++++++ drivers/infiniband/sw/rxe/rxe_verbs.h | 2 +- 6 files changed, 126 insertions(+), 11 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +2 -0
@@ -278,3 +278,5 @@ late_initcall(rxe_module_init);module_exit (rxe_module_exit);MODULE_ALIAS_RDMA_LINK ("rxe" ); + +MODULE_IMPORT_NS ("DMA_BUF" );
drivers/infiniband/sw/rxe/rxe_loc.h +2 -0
@@ -66,6 +66,8 @@ int rxe_mr_init_fast(int max_pages, struct rxe_mr *mr);int rxe_flush_pmem_iova (struct rxe_mr *mr,u64 iova,unsigned int length);int rxe_mr_copy (struct rxe_mr *mr,u64 iova,void *addr,unsigned int length,enum rxe_mr_copy_dir dir); +int rxe_mr_dmabuf_init_user (struct rxe_pd *pd,int fd,u64 start,u64 length, +u64 iova,int access,struct rxe_mr *mr);int copy_data (struct rxe_pd *pd,int access,struct rxe_dma_info *dma,void *addr,int length,enum rxe_mr_copy_dir dir); int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
drivers/infiniband/sw/rxe/rxe_mr.c +80 -9
148 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_odp.c +1 -1
@@ -351,7 +351,7 @@ int rxe_odp_flush_pmem_iova(struct rxe_mr *mr, u64 iova, page =hmm_pfn_to_page (umem_odp->map.pfn_list[index]); bytes = min_t(unsigned int, length, -mr_page_size (mr) - page_offset); +rxe_mr_page_size (mr) - page_offset); va =kmap_local_page (page);arch_wb_cache_pmem (va + page_offset, bytes);
drivers/infiniband/sw/rxe/rxe_verbs.c +40 -0
@@ -1358,6 +1358,45 @@ static struct ib_mr *rxe_rereg_user_mr(struct ib_mr *ibmr, int flags,return NULL ; } +static struct ib_mr *rxe_reg_user_mr_dmabuf (struct ib_pd *ibpd,u64 start, +u64 length,u64 iova,int fd, +int access,struct ib_dmah *dmah, +struct uverbs_attr_bundle *udata) +{ +int err; +struct rxe_dev *rxe =to_rdev (ibpd->device); +struct rxe_pd *pd =to_rpd (ibpd); +struct rxe_mr *mr; + + mr =kzalloc_obj (*mr); +if (!mr) { + err = -ENOMEM ; + goto err_out; + } + + err =rxe_add_to_pool (&rxe->mr_pool, mr); +if (err) + goto err_free; + +rxe_get (pd); + + err =rxe_mr_dmabuf_init_user (pd, fd, start, length, iova, access, mr); +if (err) + goto err3; + +return &mr->ibmr; + +err3: +rxe_put (pd); +rxe_put (mr); + +err_free: +kfree (mr); + +err_out: +return ERR_PTR (err); +} +static struct ib_mr *rxe_alloc_mr (struct ib_pd *ibpd,enum ib_mr_type mr_type,u32 max_num_sg) {
@@ -1517,6 +1556,7 @@ static const struct ib_device_ops rxe_dev_ops = { .query_qp = rxe_query_qp, .query_srq = rxe_query_srq, .reg_user_mr = rxe_reg_user_mr, + .reg_user_mr_dmabuf = rxe_reg_user_mr_dmabuf, .req_notify_cq = rxe_req_notify_cq, .rereg_user_mr = rxe_rereg_user_mr, .resize_cq = rxe_resize_cq,
drivers/infiniband/sw/rxe/rxe_verbs.h +1 -1
@@ -364,7 +364,7 @@ struct rxe_mr {struct rxe_mr_page *page_info; }; -static inline unsigned int mr_page_size (struct rxe_mr *mr) +static inline unsigned int rxe_mr_page_size (struct rxe_mr *mr) {return mr ? mr->ibmr.page_size :PAGE_SIZE ; }
Signature · 1 line
2.53.0