Skip to contentSkip to conversation

accel/amdxdna: fix double-free on mailbox channel stop

PATCH In review Deniz Aydogan 3 participants 3 messages Aug 28 – Aug 31

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

PATCH 1/1 accel/amdxdna: fix double-free on mailbox channel stop

mailbox_release_msg() frees the message with kfree() but does not remove it from the xarray. The stop function uses two loops to walk pending entries in cyclic order, but since released entries remain in the xarray, overlapping ranges cause the same entry to be freed twice.

In particular, when next_msgid is 0 (the initial value after kzalloc), xa_for_each_start() covers all entries from index 0 onward, and xa_for_each_range() with max=(u32)(0 - 1) = U32_MAX also covers all entries. Every pending message gets double-freed.

Use xa_for_each() to iterate all remaining entries exactly once. At this point the IRQ is already freed and the workqueue is drained, so traversal order does not matter.

Fixes 3ba13f5e7180 ("Merge tag 'devicetree-fixes-for-7.3-1'")
---
 drivers/accel/amdxdna/amdxdna_mailbox.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/accel/amdxdna/amdxdna_mailbox.c +1 -3
index cc8865f4e..271617347 100644
@@ -556,9 +556,7 @@ void xdna_mailbox_stop_channel(struct mailbox_channel *mb_chann)
 	drain_workqueue(mb_chann->work_q);
 
 	/* We can clean up and release resources */
-	xa_for_each_start(&mb_chann->chan_xa, msg_id, mb_msg, mb_chann->next_msgid)
-		mailbox_release_msg(mb_chann, mb_msg);
-	xa_for_each_range(&mb_chann->chan_xa, msg_id, mb_msg, 0, mb_chann->next_msgid - 1)
+	xa_for_each(&mb_chann->chan_xa, msg_id, mb_msg)
 		mailbox_release_msg(mb_chann, mb_msg);
 	xa_destroy(&mb_chann->chan_xa);
 
Signature · 1 line
2.55.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Concurrent execution of xdna_mailbox_send_msg by the DRM scheduler thread and the ioctl thread leads to hardware ring buffer corruption. --

commit f00c302142d5de5c06245223cfbc1a367c1aef38
Author: Deniz Aydogan <[email protected]>

accel/amdxdna: fix double-free on mailbox channel stop

This commit fixes a double-free issue when stopping a mailbox channel. The previous implementation used two xarray iterators that could overlap and release the same message twice. It replaces them with a single iterator to ensure each message is freed exactly once.

Quoted 5 lines
drivers/accel/amdxdna/amdxdna_mailbox.c +0 -0
index cc8865f4e79c9..27161734753d0 100644

drivers/accel/amdxdna/amdxdna_mailbox.c void xdna_mailbox_stop_channel(struct mailbox_channel *mb_chann) · line 561

Quoted 9 lines
drivers/accel/amdxdna/amdxdna_mailbox.c +1 -3
@@ -556,9 +556,7 @@ void xdna_mailbox_stop_channel(struct mailbox_channel *mb_chann)
 	drain_workqueue(mb_chann->work_q);
 
 	/* We can clean up and release resources */
-	xa_for_each_start(&mb_chann->chan_xa, msg_id, mb_msg, mb_chann->next_msgid)
-		mailbox_release_msg(mb_chann, mb_msg);
-	xa_for_each_range(&mb_chann->chan_xa, msg_id, mb_msg, 0, mb_chann->next_msgid - 1)
+	xa_for_each(&mb_chann->chan_xa, msg_id, mb_msg)
 		mailbox_release_msg(mb_chann, mb_msg);
 	xa_destroy(&mb_chann->chan_xa);

[Severity: High]
This is a pre-existing issue, but while reviewing the mailbox channel code in drivers/accel/amdxdna/amdxdna_mailbox.c, I noticed a potential data corruption issue with concurrent messages.

Can concurrent execution of xdna_mailbox_send_msg() lead to hardware ring buffer corruption?

The DRM scheduler thread can submit jobs without holding the device lock, calling:
drivers/accel/amdxdna/aie2_message.c:aie2_execbuf() -> xdna_mailbox_send_msg()

Concurrently, the ioctl thread can configure the same hardware context and channel while holding the device lock, calling:
drivers/accel/amdxdna/aie2_message.c:aie2_config_cu() -> xdna_mailbox_send_msg()

Inside mailbox_send_msg() in drivers/accel/amdxdna/amdxdna_mailbox.c, the tail pointer is read locklessly:

    tail = mb_chann->x2i_tail;

And it is updated locklessly after the message is written:

    mailbox_set_tailptr(mb_chann, tail + mb_msg->pkg_size);

Since there is no common lock held between the DRM scheduler thread and the ioctl thread for the mailbox channel, can they race to read the same tail pointer, overwrite each other's messages in the ring buffer, and write conflicting tail pointers to the hardware?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
maintainer
#2 lore raw
In reply to Deniz Aydogan
Quoted 25 lines · Deniz Aydogan

mailbox_release_msg() frees the message with kfree() but does not remove it from the xarray. The stop function uses two loops to walk pending entries in cyclic order, but since released entries remain in the xarray, overlapping ranges cause the same entry to be freed twice.

In particular, when next_msgid is 0 (the initial value after kzalloc), xa_for_each_start() covers all entries from index 0 onward, and xa_for_each_range() with max=(u32)(0 - 1) = U32_MAX also covers all entries. Every pending message gets double-freed.

Use xa_for_each() to iterate all remaining entries exactly once. At this point the IRQ is already freed and the workqueue is drained, so traversal order does not matter.

Fixes 3ba13f5e7180 ("Merge tag 'devicetree-fixes-for-7.3-1'")
---
  drivers/accel/amdxdna/amdxdna_mailbox.c | 4 +---
  1 file changed, 1 insertion(+), 3 deletions(-)
drivers/accel/amdxdna/amdxdna_mailbox.c +0 -0
index cc8865f4e..271617347 100644

drivers/accel/amdxdna/amdxdna_mailbox.c void xdna_mailbox_stop_channel(struct mailbox_channel *mb_chann) · line 559

Quoted 7 lines
drivers/accel/amdxdna/amdxdna_mailbox.c +1 -3
@@ -556,9 +556,7 @@ void xdna_mailbox_stop_channel(struct mailbox_channel *mb_chann)
  	drain_workqueue(mb_chann->work_q);
 
  	/* We can clean up and release resources */
-	xa_for_each_start(&mb_chann->chan_xa, msg_id, mb_msg, mb_chann->next_msgid)
-		mailbox_release_msg(mb_chann, mb_msg);
-	xa_for_each_range(&mb_chann->chan_xa, msg_id, mb_msg, 0, mb_chann->next_msgid - 1)
+	xa_for_each(&mb_chann->chan_xa, msg_id, mb_msg)

The msg need to be released in sequence. And I would suggest to release msgid in mailbox_release_msg()

@@ -192,6 +192,7 @@ static void mailbox_release_msg(struct mailbox_channel *mb_chann, mb_msg->pkg.header.id, mb_msg->pkg.header.opcode);
        if (mb_msg->notify_cb)
                mb_msg->notify_cb(mb_msg->handle, NULL, 0);
+       mailbox_release_msgid(mb_chann, mb_msg->pkg.header.id); kfree(mb_msg);

Thanks,

Lizhi

Quoted 3 lines
drivers/accel/amdxdna/amdxdna_mailbox.c +0 -0
  		mailbox_release_msg(mb_chann, mb_msg);
  	xa_destroy(&mb_chann->chan_xa);