Skip to contentSkip to conversation

RDMA/irdma: Initialize iwmr->access during MR registration

PATCH Applied Jacob Moroni 2 participants 2 messages Jun 4 – Jun 5

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 RDMA/irdma: Initialize iwmr->access during MR registration

Initialize iwmr->access during initial user mem registration so that it contains a valid value during a subsequent rereg_mr.

Otherwise, a rereg_mr that doesn't set IB_MR_REREG_ACCESS (for example, one that only changes the PD) ends up clearing the access flags in HW since iwmr->access is zero-initialized, which is not intended.

Fixes 5ac388db27c4 ("RDMA/irdma: Add support to re-register a memory region")
---
 drivers/infiniband/hw/irdma/verbs.c | 1 +
 1 file changed, 1 insertion(+)
drivers/infiniband/hw/irdma/verbs.c +1 -0
index baee48df5fd..f2f487e9540 100644
@@ -3309,6 +3309,7 @@ static int irdma_reg_user_mr_type_mem(struct irdma_mr *iwmr, int access,
 	int err;
 
 	lvl = iwmr->page_cnt != 1 ? PBLE_LEVEL_1 | PBLE_LEVEL_2 : PBLE_LEVEL_0;
+	iwmr->access = access;
 
 	err = irdma_setup_pbles(iwdev->rf, iwmr, lvl);
 	if (err)
Signature · 1 line
2.54.0.1032.g2f8565e1d1-goog
maintainer Applied
#1 lore raw
Quoted 8 lines · Jacob Moroni

Initialize iwmr->access during initial user mem registration so that it contains a valid value during a subsequent rereg_mr.

Otherwise, a rereg_mr that doesn't set IB_MR_REREG_ACCESS (for example, one that only changes the PD) ends up clearing the access flags in HW since iwmr->access is zero-initialized, which is not intended.

Quoted 5 lines
Fixes 5ac388db27c4 ("RDMA/irdma: Add support to re-register a memory region")
---
 drivers/infiniband/hw/irdma/verbs.c | 1 +
 1 file changed, 1 insertion(+)

Applied to for-next

Still more preexisting Sashiko issues:

https://sashiko.dev/#/patchset/20260604154104.4035581-1-jmoroni%40google.com

This one sounds bad:

This is a pre-existing issue, but could this sequence allow arbitrary physical memory access when an MR re-registration fails?

Sigh, rereg_mr is a trainwreck in every driver.

I fixed mlx5 a long time ago, there is a new flow now for rereg where the driver can create a new mr from scratch and return that. irdma should do that for any destructive change like replacing the umem. Otherwise it is impossible to get the error semantics correct.

Jason