[PATCH rdma-rc] RDMA/rxe: Fix the copy offsets in __rxe_odp_mr_copy()
From: Andrew Stellman
Date: Sat Nov 21 2026 - 19:37:23 EST
__rxe_odp_mr_copy() maps each page it copies and copies to or from the
start of that page. It uses the offset of iova in the first page only
to size the first copy, not to find where in the page to copy, and it
does not advance addr after each page, so each page is copied between
the start of that page and the start of addr.
Tested as root on rxe loopback (path MTU 1024), three runs, with an
ODP MR registered 100 bytes into a page:
- In a 4096-byte RDMA READ or SEND from the start of the MR, the first
16 bytes of each of the first three 1024-byte packets were the same
as the first 16 bytes of the page, not the data at their offset in
the MR.
- A 32-byte RDMA WRITE to the start of the MR changed 32 of the 100
bytes between the start of the page and the MR, which are outside
the MR, and none of the 32 bytes at the start of the MR got the
written value.
- In a 1024-byte RDMA READ that starts 512 bytes before a page
boundary, the first 512 bytes matched the start of the next page.
In each run, over 50 rounds of a 1024-byte RDMA WRITE between two
other MRs followed by such a READ, some of the last 512 bytes of the
READs were neither zero nor the expected data.
Each of these also occurred in three runs as an unprivileged user (uid
65534, no effective capabilities) on an rxe device that root had
created. With this patch (one run, as root), the READ, SEND and WRITE
tests above returned or wrote the expected bytes, and the 50 rounds
returned no such bytes.
Use the offset of iova in the first page, and advance addr after each
page, as rxe_mr_copy_xarray() does.
The issue was found by Claude Opus 5.5 and Claude Fable running
Quality Playbook, an LLM-driven code review tool:
https://github.com/andrewstellman/quality-playbook
The fix and this changelog were drafted with Claude Opus 5.5 and
reviewed by me.
Fixes: 2fae67ab63db ("RDMA/rxe: Add support for Send/Recv/Write/Read with ODP")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Andrew Stellman <astellman@xxxxxxxxxxxxxxxxxxx>
---
Tested in QEMU on v7.4-rc3. rxe_odp.c is the same at rdma for-rc
commit e7732ec135b5. The reproducer is a libibverbs program written
with Claude Opus 5.5; I can send it off-list.
drivers/infiniband/sw/rxe/rxe_odp.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
index 9d27244f89b..73ed4c45fea 100644
--- a/drivers/infiniband/sw/rxe/rxe_odp.c
+++ b/drivers/infiniband/sw/rxe/rxe_odp.c
@@ -211,8 +211,8 @@ static int __rxe_odp_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
page = hmm_pfn_to_page(umem_odp->map.pfn_list[idx]);
user_va = kmap_local_page(page);
- src = (dir == RXE_TO_MR_OBJ) ? addr : user_va;
- dest = (dir == RXE_TO_MR_OBJ) ? user_va : addr;
+ src = (dir == RXE_TO_MR_OBJ) ? addr : user_va + offset;
+ dest = (dir == RXE_TO_MR_OBJ) ? user_va + offset : addr;
bytes = BIT(umem_odp->page_shift) - offset;
if (bytes > length)
@@ -222,6 +222,7 @@ static int __rxe_odp_mr_copy(struct rxe_mr *mr, u64 iova, void *addr,
kunmap_local(user_va);
length -= bytes;
+ addr += bytes;
idx++;
offset = 0;
}
base-commit: e7732ec135b501e87ab060d992b3b5f2ec4396c9
--
2.43.0