[PATCH v3] Bluetooth: Don't initialize msft/aosp when using user channel

A race condition is triggered when usermode control is given to
userspace before the kernel's MSFT query responds, resulting in an
unexpected response to userspace's reset command.

Issue can be observed in btmon:
< HCI Command: Vendor (0x3f|0x001e) plen 2                    #3 [hci0]
        05 01                                            ..
@ USER Open: bt_stack_manage (privileged) version 2.22  {0x0002} [hci0]
< HCI Command: Reset (0x03|0x0003) plen 0                     #4 [hci0]

HCI Event: Command Complete (0x0e) plen 5                   #5 [hci0]

      Vendor (0x3f|0x001e) ncmd 1
	Status: Command Disallowed (0x0c)
	05                                               .

HCI Event: Command Complete (0x0e) plen 4                   #6 [hci0]

      Reset (0x03|0x0003) ncmd 2
	Status: Success (0x00)
Reviewed-by: Abhishek Pandit-Subedi <abhishekpandit@chromium.org>
Reviewed-by: Sonny Sasaka <redacted>
Signed-off-by: Jesse Melhuish <redacted>
---

Changes in v3:
- Added guard around *_close calls as well.

Changes in v2:
- Moved guard to the new home for this code.

 net/bluetooth/hci_sync.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/net/bluetooth/hci_sync.c b/net/bluetooth/hci_sync.c
index b794605dc882..06d08dc6ad22 100644
--- a/net/bluetooth/hci_sync.c
+++ b/net/bluetooth/hci_sync.c

@@ -3887,8 +3887,10 @@ int hci_dev_open_sync(struct hci_dev *hdev)
 	    hci_dev_test_flag(hdev, HCI_VENDOR_DIAG) && hdev->set_diag)
 		ret = hdev->set_diag(hdev, true);
 
-	msft_do_open(hdev);
-	aosp_do_open(hdev);
+	if (!hci_dev_test_flag(hdev, HCI_USER_CHANNEL)) {
+		msft_do_open(hdev);
+		aosp_do_open(hdev);
+	}
 
 	clear_bit(HCI_INIT, &hdev->flags);
 

@@ -4031,8 +4033,10 @@ int hci_dev_close_sync(struct hci_dev *hdev)
 
 	hci_sock_dev_event(hdev, HCI_DEV_DOWN);
 
-	aosp_do_close(hdev);
-	msft_do_close(hdev);
+	if (!hci_dev_test_flag(hdev, HCI_USER_CHANNEL)) {
+		aosp_do_close(hdev);
+		msft_do_close(hdev);
+	}
 
 	if (hdev->flush)
 		hdev->flush(hdev);

-- 
2.31.0

Hi Jesse,

A race condition is triggered when usermode control is given to
userspace before the kernel's MSFT query responds, resulting in an
unexpected response to userspace's reset command.

Issue can be observed in btmon:
< HCI Command: Vendor (0x3f|0x001e) plen 2                    #3 [hci0]
       05 01                                            ..
@ USER Open: bt_stack_manage (privileged) version 2.22  {0x0002} [hci0]
< HCI Command: Reset (0x03|0x0003) plen 0                     #4 [hci0]

quoted

HCI Event: Command Complete (0x0e) plen 5                   #5 [hci0]

     Vendor (0x3f|0x001e) ncmd 1
	Status: Command Disallowed (0x0c)
	05                                               .

quoted

HCI Event: Command Complete (0x0e) plen 4                   #6 [hci0]

     Reset (0x03|0x0003) ncmd 2
	Status: Success (0x00)
Reviewed-by: Abhishek Pandit-Subedi <abhishekpandit@chromium.org>
Reviewed-by: Sonny Sasaka <redacted>
Signed-off-by: Jesse Melhuish <redacted>
---

Changes in v3:
- Added guard around *_close calls as well.

Changes in v2:
- Moved guard to the new home for this code.

net/bluetooth/hci_sync.c | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)

patch has been applied to bluetooth-next tree.

Regards

Marcel