在 2023/1/4 7:11, Jason Gunthorpe 写道:
On Mon, Jan 02, 2023 at 08:34:33PM -0500, Zhu Yanjun wrote:
From: Zhu Yanjun <yanjun.zhu@xxxxxxxxx>
This is a followup to the EFA dmabuf[1]. Irdma driver currently does
not support on-demand-paging(ODP). So it uses habanalabs as the
dmabuf exporter, and irdma as the importer to allow for peer2peer
access through libibverbs.
In this commit, the function ib_umem_dmabuf_get_pinned() is used.
This function is introduced in EFA dmabuf[1] which allows the driver
to get a dmabuf umem which is pinned and does not require move_notify
callback implementation. The returned umem is pinned and DMA mapped
like standard cpu umems, and is released through ib_umem_release().
[1]https://lore.kernel.org/lkml/20211007114018.GD2688930@xxxxxxxx/t/
Signed-off-by: Zhu Yanjun <yanjun.zhu@xxxxxxxxx>
---
drivers/infiniband/hw/irdma/verbs.c | 158 ++++++++++++++++++++++++++++
1 file changed, 158 insertions(+)
diff --git a/drivers/infiniband/hw/irdma/verbs.c b/drivers/infiniband/hw/irdma/verbs.c
index f6973ea55eda..76dc6e65930a 100644
--- a/drivers/infiniband/hw/irdma/verbs.c
+++ b/drivers/infiniband/hw/irdma/verbs.c
@@ -2912,6 +2912,163 @@ static struct ib_mr *irdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 len,
return ERR_PTR(err);
}
+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 irdma_device *iwdev = to_iwdev(pd->device);
+ struct irdma_ucontext *ucontext;
+ struct irdma_pble_alloc *palloc;
+ struct irdma_pbl *iwpbl;
+ struct irdma_mr *iwmr;
+ struct irdma_mem_reg_req req;
+ u32 total, stag = 0;
+ u8 shadow_pgcnt = 1;
+ bool use_pbles = false;
+ unsigned long flags;
+ int err = -EINVAL;
+ struct ib_umem_dmabuf *umem_dmabuf;
+
+ if (len > iwdev->rf->sc_dev.hw_attrs.max_mr_size)
+ return ERR_PTR(-EINVAL);
+
+ if (udata->inlen < IRDMA_MEM_REG_MIN_REQ_LEN)
+ return ERR_PTR(-EINVAL);
+
+ umem_dmabuf = ib_umem_dmabuf_get_pinned(pd->device, start, len, fd,
+ access);
+ if (IS_ERR(umem_dmabuf)) {
+ err = PTR_ERR(umem_dmabuf);
+ ibdev_dbg(&iwdev->ibdev, "Failed to get dmabuf umem[%d]\n", err);
+ return ERR_PTR(err);
+ }
+
+ if (ib_copy_from_udata(&req, udata, min(sizeof(req), udata->inlen))) {
+ ib_umem_release(&umem_dmabuf->umem);
+ return ERR_PTR(-EFAULT);
+ }
+
+ iwmr = kzalloc(sizeof(*iwmr), GFP_KERNEL);
+ if (!iwmr) {
+ ib_umem_release(&umem_dmabuf->umem);
+ return ERR_PTR(-ENOMEM);
+ }
+
+ iwpbl = &iwmr->iwpbl;
+ iwpbl->iwmr = iwmr;
+ iwmr->region = &umem_dmabuf->umem;
+ iwmr->ibmr.pd = pd;
+ iwmr->ibmr.device = pd->device;
+ iwmr->ibmr.iova = virt;
+ iwmr->page_size = PAGE_SIZE;
+
+ if (req.reg_type == IRDMA_MEMREG_TYPE_MEM) {
+ iwmr->page_size = ib_umem_find_best_pgsz(iwmr->region,
+ iwdev->rf->sc_dev.hw_attrs.page_size_cap,
+ virt);
You can't call rdma_umem_for_each_dma_block() without also calling
this function to validate that the page_size passed to
rdma_umem_for_each_dma_block() is correct.
Got it. I will fix it in the latest commit.
This seems to be an existing bug, please fix it.
Also, is there a reason this code is all duplicated from
irdma_reg_user_mr? Please split things up like the other drivers to
obtain the umem then use shared code to process the umem as required.
Got it.
Zhu Yanjun
Jason
