Re: [bug report] [rdma] RXE ODP test hangs with new DMA map API

[Leon Romanovsky <leon@kernel.org>]

On Wed, May 21, 2025 at 09:48:27PM +0900, Daisuke Matsuda wrote:
> Hi,
> 
> After these two patches are merged to the for-next tree, RXE ODP test always hangs:
>   RDMA/core: Convert UMEM ODP DMA mapping to caching IOVA and page linkage
>   RDMA/umem: Store ODP access mask information in PFN
> cf. https://lore.kernel.org/linux-rdma/cover.1745831017.git.leon@kernel.org/
> 
> Here is the console log:
> ```
> $ ./build/bin/run_tests.py -v -k odp
> test_odp_dc_traffic (tests.test_mlx5_dc.DCTest.test_odp_dc_traffic) ... skipped 'Can not run the test over non MLX5 device'
> test_devx_rc_qp_odp_traffic (tests.test_mlx5_devx.Mlx5DevxRcTrafficTest.test_devx_rc_qp_odp_traffic) ... skipped 'Can not run the test over non MLX5 device'
> test_odp_mkey_list_new_api (tests.test_mlx5_mkey.Mlx5MkeyTest.test_odp_mkey_list_new_api)
> Create Mkeys above ODP MR, configure it with memory layout using the new API and ... skipped 'Could not open mlx5 context (This is not an MLX5 device)'
> test_odp_async_prefetch_rc_traffic (tests.test_odp.OdpTestCase.test_odp_async_prefetch_rc_traffic) ...
> 
> 
> ```
> 
> It looks that the python process is somehow stuck in uverbs_destroy_ufile_hw():
> ```
> $ sudo cat /proc/1845/task/1845/stack
> [<0>] uverbs_destroy_ufile_hw+0x24/0x100 [ib_uverbs]
> [<0>] ib_uverbs_close+0x1b/0xc0 [ib_uverbs]
> [<0>] __fput+0xea/0x2d0
> [<0>] ____fput+0x15/0x20
> [<0>] task_work_run+0x5d/0xa0
> [<0>] do_exit+0x316/0xa50
> [<0>] make_task_dead+0x81/0x160
> [<0>] rewind_stack_and_make_dead+0x16/0x20
> ```
> 
> I am not sure about the root cause but hope we can fix this before the next merge window.

Can you please try this fix?

diff --git a/drivers/infiniband/sw/rxe/rxe_odp.c b/drivers/infiniband/sw/rxe/rxe_odp.c
index a1416626f61a5..0f67167ddddd1 100644
--- a/drivers/infiniband/sw/rxe/rxe_odp.c
+++ b/drivers/infiniband/sw/rxe/rxe_odp.c
@@ -137,7 +137,7 @@ static inline bool rxe_check_pagefault(struct ib_umem_odp *umem_odp,
        while (addr < iova + length) {
                idx = (addr - ib_umem_start(umem_odp)) >> umem_odp->page_shift;
 
-               if (!(umem_odp->map.pfn_list[idx] & perm)) {
+               if (!(umem_odp->map.pfn_list[idx] & HMM_PFN_VALID)) {
                        need_fault = true;
                        break;
               

> 
> Thanks,
> Daisuke