Re: [PATCH] Bluetooth: btusb: Add a new quirk to skip HCI_FLT_CLEAR_ALL on fake CSR controllers

[Hans de Goede]

Hi,

On 1/6/22 23:29, Ismael Ferreras Morezuelas wrote:
> Hi again, Marcel.
> 
>>> 		/* Clear the reset quirk since this is not an actual
>>> 		 * early Bluetooth 1.1 device from CSR.
>>> @@ -1942,16 +1943,16 @@ static int btusb_setup_csr(struct hci_dev *hdev)
>>> 		/*
>>> 		 * Special workaround for these BT 4.0 chip clones, and potentially more:
>>> 		 *
>>> -		 * - 0x0134: a Barrot 8041a02                 (HCI rev: 0x1012 sub: 0x0810)
>>> +		 * - 0x0134: a Barrot 8041a02                 (HCI rev: 0x0810 sub: 0x1012)
>>
>> Don’t get this change.
>>
> 
> I swapped both numbers by mistake in my last commit when I moved it from
> a conditional into a comment This is explained in the changeset as:
> 
>  > Also, swapped the HCI subver and LMP subver numbers for the Barrot
>  > in the comment, which I copied wrong the last time around.
> 
> Thought I might as well fix it here, because it may never get corrected otherwise; it's misleading.
> 
> 
>>> 		 * - 0x7558: IC markings FR3191AHAL 749H15143 (HCI rev/sub-version: 0x0709)
>>> 		 *
>>> 		 * These controllers are really messed-up.
>>> 		 *
>>> 		 * 1. Their bulk RX endpoint will never report any data unless
>>> -		 * the device was suspended at least once (yes, really).
>>> +		 *    the device was suspended at least once (yes, really).
>>> 		 * 2. They will not wakeup when autosuspended and receiving data
>>> -		 * on their bulk RX endpoint from e.g. a keyboard or mouse
>>> -		 * (IOW remote-wakeup support is broken for the bulk endpoint).
>>> +		 *    on their bulk RX endpoint from e.g. a keyboard or mouse
>>> +		 *    (IOW remote-wakeup support is broken for the bulk endpoint).
>>
>> Fix the style issues separately.
> 
> Fair enough, I can obviate this part, no worries.
> 
> 
>>
>>> 		 *
>>> 		 * To fix 1. enable runtime-suspend, force-suspend the
>>> 		 * HCI and then wake-it up by disabling runtime-suspend.
>>> @@ -1971,7 +1972,7 @@ static int btusb_setup_csr(struct hci_dev *hdev)
>>> 		if (ret >= 0)
>>> 			msleep(200);
>>> 		else
>>> -			bt_dev_err(hdev, "CSR: Failed to suspend the device for our Barrot 8041a02 receive-issue workaround");
>>> +			bt_dev_warn(hdev, "CSR: Couldn't suspend the device for our Barrot 8041a02 receive-issue workaround");
>>
>> Why change this?
>>
> 
> Because depending on the clone this print may end up showing all the time;
> we *try* doing it, but some clones don't like it. I thought showing it
> as a warning would make sense. Tweaking the text a bit again helps
> when tracking down which version of the workaround the user is
> running, we can narrow it by just grepping the log itself.
> 
> If it doesn't work it doesn't really affect anything, and we can't
> whitelist something half the clones use to get unstuck and the other
> half just ignore and stay peachy. We're trying an unified solution.
> 
> 
>>>
>>> 		pm_runtime_forbid(&data->udev->dev);
>>>
>>> diff --git a/include/net/bluetooth/hci.h b/include/net/bluetooth/hci.h
>>> index 63065bc01..4e5d5979d 100644
>>> --- a/include/net/bluetooth/hci.h
>>> +++ b/include/net/bluetooth/hci.h
>>> @@ -246,6 +246,12 @@ enum {
>>> 	 * HCI after resume.
>>> 	 */
>>> 	HCI_QUIRK_NO_SUSPEND_NOTIFIER,
>>> +
>>> +	/* When this quirk is set, HCI_OP_SET_EVENT_FLT requests with
>>> +	 * HCI_FLT_CLEAR_ALL are ignored. A subset of the CSR controller
>>> +	 * clones struggle with this and instantly lock up.
>>> +	 */
>>> +	HCI_QUIRK_BROKEN_FILTER_CLEAR_ALL,
>>> };
>>>
>>> /* HCI device flags */
>>> diff --git a/net/bluetooth/hci_core.c b/net/bluetooth/hci_core.c
>>> index 8d33aa648..7af649afc 100644
>>> --- a/net/bluetooth/hci_core.c
>>> +++ b/net/bluetooth/hci_core.c
>>> @@ -150,6 +150,7 @@ static void bredr_setup(struct hci_request *req)
>>> {
>>> 	__le16 param;
>>> 	__u8 flt_type;
>>> +	struct hci_dev *hdev = req->hdev;
>>
>> This should always go first in a function.
>>
>>>
>>> 	/* Read Buffer Size (ACL mtu, max pkt, etc.) */
>>> 	hci_req_add(req, HCI_OP_READ_BUFFER_SIZE, 0, NULL);
>>> @@ -169,9 +170,13 @@ static void bredr_setup(struct hci_request *req)
>>> 	/* Read Current IAC LAP */
>>> 	hci_req_add(req, HCI_OP_READ_CURRENT_IAC_LAP, 0, NULL);
>>>
>>> -	/* Clear Event Filters */
>>> -	flt_type = HCI_FLT_CLEAR_ALL;
>>> -	hci_req_add(req, HCI_OP_SET_EVENT_FLT, 1, &flt_type);
>>> +	/* Clear Event Filters; some fake CSR controllers lock up after setting
>>> +	 * this type of filter, so avoid sending the request altogether.
>>> +	 */
>>> +	if (!test_bit(HCI_QUIRK_BROKEN_FILTER_CLEAR_ALL, &hdev->quirks)) {
>>> +		flt_type = HCI_FLT_CLEAR_ALL;
>>> +		hci_req_add(req, HCI_OP_SET_EVENT_FLT, 1, &flt_type);
>>> +	}
>>>
>>> 	/* Connection accept timeout ~20 secs */
>>> 	param = cpu_to_le16(0x7d00);
>>> diff --git a/net/bluetooth/hci_request.c b/net/bluetooth/hci_request.c
>>> index 92611bfc0..cfcf64c0c 100644
>>> --- a/net/bluetooth/hci_request.c
>>> +++ b/net/bluetooth/hci_request.c
>>> @@ -980,11 +980,15 @@ void hci_req_add_le_passive_scan(struct hci_request *req)
>>> static void hci_req_clear_event_filter(struct hci_request *req)
>>> {
>>> 	struct hci_cp_set_event_filter f;
>>> +	struct hci_dev *hdev = req->hdev;
>>> +
>>> +	if (!hci_dev_test_flag(hdev, HCI_BREDR_ENABLED))
>>> +		return;
>>>
>>> -	if (!hci_dev_test_flag(req->hdev, HCI_BREDR_ENABLED))
>>> +	if (test_bit(HCI_QUIRK_BROKEN_FILTER_CLEAR_ALL, &hdev->quirks))
>>> 		return;
>>>
>>> -	if (hci_dev_test_flag(req->hdev, HCI_EVENT_FILTER_CONFIGURED)) {
>>> +	if (hci_dev_test_flag(hdev, HCI_EVENT_FILTER_CONFIGURED)) {
>>> 		memset(&f, 0, sizeof(f));
>>> 		f.flt_type = HCI_FLT_CLEAR_ALL;
>>> 		hci_req_add(req, HCI_OP_SET_EVENT_FLT, 1, &f);
>>
>> This is not enough. If you do not have clear event filter, we need to disable suspend/resume support. These device can for obvious reason not sleep accordingly.
> 
> Would adding HCI_QUIRK_NO_SUSPEND_NOTIFIER be enough here? I'll take another look at the
> codebase, any comments about the best/simplest way to tackle this would help me a lot.
> 
> As you can see I'm still getting my feet wet around the Bluetooth subsystem.
> 
> In theory adding another quirk condition in hci_req_set_event_filter() or anything that leads
> to it (i.e. hci_suspend_dev() / BT_SUSPEND_CONFIGURE_WAKE) may also do the trick.
> 
> The least code I touch the better. Right now I'm thinking of mirroring my hci_req_clear_event_filter()
> nop-out in hci_req_set_event_filter() *and* just setting the NO_SUSPEND_NOTIFIER quirk.

Note we already disable runtime suspend for these broken clones in
the btusb code, I think that in itself might be enough, together with
a comment in the header where the quirk is defined that devices using this
must disable runtime suspend ?

Regards,

Hans