Skip to contentSkip to conversation

habanalabs: fix race when waiting on encaps signal

PATCH New Oded Gabbay 1 participant 1 message Jan 28, 2022 Char and misc drivers

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 habanalabs: fix race when waiting on encaps signal

From: Dani Liberman <[email protected]>

Scenario:
1. CS which is part of encaps signal has been completed and now executing kref_put to its encaps signal handle. The refcount of the handle decremented to 0, and called the encaps signal handle release function - hl_encaps_handle_do_release.

2. At this point the user starts waiting on the signal, and finds the encaps signal handle in the handlers list and increment the habdle refcount to 1.

3. Immediately after, hl_encaps_handle_do_release removed the handle from the list and free its memory.

4. Wait function using the handle although it has been freed.

This scenario caused the slab area which was previously allocated for the handle to be poison overwritten which triggered kernel bug the next time the OS needed to allocate this slab.

Fixed by getting the refcount of the handle only in case it is not zero.

---
v2:
- Use kref_get_unless_zero() instead of kref_get() and then checking if the value is not 0.

 drivers/misc/habanalabs/common/command_submission.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
drivers/misc/habanalabs/common/command_submission.c +8 -5
index 0ea9a73e4aa5..ba5215b77852 100644
@@ -2063,13 +2063,16 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
 			idp = &ctx->sig_mgr.handles;
 			idr_for_each_entry(idp, encaps_sig_hdl, id) {
 				if (encaps_sig_hdl->cs_seq == signal_seq) {
-					handle_found = true;
-					/* get refcount to protect removing
-					 * this handle from idr, needed when
-					 * multiple wait cs are used with offset
+					/* get refcount to protect removing this handle from idr,
+					 * needed when multiple wait cs are used with offset
 					 * to wait on reserved encaps signals.
+					 * Since kref_put of this handle is executed outside the
+					 * current lock, it is possible that the handle refcount
+					 * is 0 but it yet to be removed from the list. In this
+					 * case need to consider the handle as not valid.
 					 */
-					kref_get(&encaps_sig_hdl->refcount);
+					if (kref_get_unless_zero(&encaps_sig_hdl->refcount))
+						handle_found = true;
 					break;
 				}
 			}
Signature · 1 line
2.25.1