Skip to contentSkip to conversation

RDMA/ucma: reject a second join of the same multicast address

PATCH New Andrey Misiurov 2 participants 2 messages Nov 19

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

PATCH 1/1 RDMA/ucma: reject a second join of the same multicast address

ucma_process_join() lets one cm_id join the same multicast address twice, creating two ucma_multicast objects (A and B) and, via rdma_join_multicast(), two cma_multicast entries that both point back at the ucma object.

rdma_leave_multicast() matches a leave only by sockaddr and destroys the first entry it finds. So LEAVE_MCAST on A destroys B's cma_multicast and frees ucma object A, while A's cma_multicast stays on the list with a context pointing at freed memory. The next event for it is a use-after-free:

  BUG: KASAN: slab-use-after-free in ucma_create_uevent+0xb11/0xb70 [rdma_ucm]
  Workqueue: rdma_cm cma_iboe_join_work_handler [rdma_cm]
  Call Trace:
   ucma_create_uevent+0xb11/0xb70 [rdma_ucm]
   ucma_event_handler+0xd5/0xa10 [rdma_ucm]
   cma_cm_event_handler+0x98/0x310 [rdma_cm]
   cma_iboe_join_work_handler+0xac/0x170 [rdma_cm]
   process_scheduled_works+0xb59/0x1890
   worker_thread+0x470/0xe40
   kthread+0x33a/0x450
   ret_from_fork+0x59/0x90
   ret_from_fork_asm+0x1a/0x30

Any user with access to /dev/infiniband/rdma_cm can trigger it.

Reject a second join of an address already joined on the same cm_id. The incomplete fix 662ade4de9ff ("RDMA/ucma: Serialize join and leave on copy_to_user failure") left this path open.

Fixes 662ade4de9ff ("RDMA/ucma: Serialize join and leave on copy_to_user failure")

---
Found during a backport review.

Reproduced on debian13 6.12.88 debug (KASAN=y) over Soft-RoCE: two joins of one address on a cm_id, then LEAVE -> the KASAN use-after-free above within one run. With the fix the second join returns -EADDRINUSE and the run is clean.

 drivers/infiniband/core/ucma.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
drivers/infiniband/core/ucma.c +8 -1
@@ -1509,7 +1509,7 @@
 {
 	struct rdma_ucm_create_id_resp resp;
 	struct ucma_context *ctx;
-	struct ucma_multicast *mc;
+	struct ucma_multicast *mc, *tmp;
 	struct sockaddr *addr;
 	int ret;
 	u8 join_state;
@@ -1544,6 +1544,13 @@
 	memcpy(&mc->addr, addr, cmd->addr_size);
 
 	xa_lock(&multicast_table);
+	/* leave matches by address only; refuse a duplicate */
+	list_for_each_entry(tmp, &ctx->mc_list, list)
+		if (!memcmp(&tmp->addr, addr, cmd->addr_size)) {
+			ret = -EADDRINUSE;
+			goto err_free_mc;
+		}
+
 	if (__xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
 		     GFP_KERNEL)) {
 		ret = -ENOMEM;
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [Critical] RDMA/ucma: UAF race between ucma_leave_multicast() and ucma_process_join() - [High] RDMA/ucma…

Thank you for your contribution! Sashiko AI review found 2 potential issues to consider:
- [Critical] RDMA/ucma: UAF race between ucma_leave_multicast() and ucma_process_join()
- [High] RDMA/ucma: TOCTOU race in ucma_process_join() across __xa_alloc() lock drop
--

commit 92134efcefe26ef534f470da9d6476bb2827da65
Author: Andrey Misiurov <[email protected]>

RDMA/ucma: reject a second join of the same multicast address

This commit rejects duplicate multicast joins for the same address on a ucma_context by checking ctx->mc_list under multicast_table's XArray lock before allocating an ID.

Quoted 14 lines
drivers/infiniband/core/ucma.c +1 -1
index a15182f7a7c5b..3f39de52bc9a6 100644
@@ -1509,7 +1509,7 @@ static ssize_t ucma_process_join(struct ucma_file *file,
 {
 	struct rdma_ucm_create_id_resp resp;
 	struct ucma_context *ctx;
-	struct ucma_multicast *mc;
+	struct ucma_multicast *mc, *tmp;
 	struct sockaddr *addr;
 	int ret;
 	u8 join_state;

drivers/infiniband/core/ucma.c static ssize_t ucma_process_join(struct ucma_file *file, · line 1552

Quoted 9 lines
drivers/infiniband/core/ucma.c +6 -0
@@ -1544,6 +1544,13 @@ static ssize_t ucma_process_join(struct ucma_file *file,
 	memcpy(&mc->addr, addr, cmd->addr_size);
 
 	xa_lock(&multicast_table);
+	/* leave matches by address only; refuse a duplicate */
+	list_for_each_entry(tmp, &ctx->mc_list, list)
+		if (!memcmp(&tmp->addr, addr, cmd->addr_size)) {
+			ret = -EADDRINUSE;
+			goto err_free_mc;
+		}

[Severity: Critical]
Can a concurrent ucma_leave_multicast() and ucma_process_join() for the same address bypass this check and cause a use-after-free of the leaving ucma_multicast?

ucma_leave_multicast() removes the old entry from ctx->mc_list and drops multicast_table's xa_lock before acquiring ctx->mutex to call rdma_leave_multicast(). A join entering in that window will not see the address on ctx->mc_list and can take ctx->mutex first:

CPU0 (leave mc_A) CPU1 (re-join same addr, mc_B)
ucma_leave_multicast()
xa_lock(&multicast_table)
list_del(&mc_A->list)
xa_unlock(&multicast_table)

                                  ucma_process_join()
                                    xa_lock(&multicast_table)
                                    /* mc_A already unlinked */
                                    list_add_tail(&mc_B->list,
                                                  &ctx->mc_list)
                                    xa_unlock(&multicast_table)
                                    mutex_lock(&ctx->mutex)
                                    rdma_join_multicast(..., mc_B)
                                      /* prepends cma_mc_B to
                                         id_priv->mc_list */
                                    mutex_unlock(&ctx->mutex)
  mutex_lock(&ctx->mutex)
  rdma_leave_multicast()
    /* matches by addr from head,
       destroys cma_mc_B instead
       of cma_mc_A */
  mutex_unlock(&ctx->mutex)
  kfree(mc_A)
                                  cma_iboe_join_work_handler()
                                    ...
                                      ucma_create_uevent()
                                        /* dereferences mc_A */
                                        <-- use-after-free

Because rdma_join_multicast() inserts at the head of id_priv->mc_list with list_add(), CPU0's rdma_leave_multicast() destroys CPU1's newly joined cma_multicast entry while leaving cma_mc_A active and pointing to freed mc_A. Checking and updating ctx->mc_list (or serializing the join and leave operations) under ctx->mutex would close this window.

drivers/infiniband/core/ucma.c line 1556

Quoted 4 lines
drivers/infiniband/core/ucma.c +1 -0
+
 	if (__xa_alloc(&multicast_table, &mc->id, NULL, xa_limit_32b,
 		     GFP_KERNEL)) {
 		ret = -ENOMEM;

[Severity: High]
Can two concurrent joins of the same address both pass the duplicate check when __xa_alloc() drops multicast_table's xa_lock to allocate an XArray node?

Because __xa_alloc() is called with GFP_KERNEL before mc is added to ctx->mc_list, __xas_nomem() will drop and reacquire multicast_table's xa_lock if it needs to allocate a radix_tree_node:

CPU0 CPU1
ucma_process_join()
xa_lock(&multicast_table)
/* checks ctx->mc_list: empty */
__xa_alloc(..., GFP_KERNEL)

    __xas_nomem()
      xas_unlock_type()
                                  ucma_process_join()
                                    xa_lock(&multicast_table)
                                    /* checks ctx->mc_list: still empty */
                                    __xa_alloc(...)
                                    list_add_tail(&mc1->list,
                                                  &ctx->mc_list)
                                    xa_unlock(&multicast_table)
      xas_lock_type()
  list_add_tail(&mc0->list,
                &ctx->mc_list)

xa_unlock(&multicast_table)

Both joins then succeed for the same address on the same cm_id, recreating the duplicate cma_multicast state and subsequent use-after-free on leave. Inserting mc into ctx->mc_list before __xa_alloc() (and unlinking it if __xa_alloc() fails), or serializing under ctx->mutex, avoids this race.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1