Re: [PATCH v7 4/4] Bluetooth: btqca: inject command complete event during fw download

From: Balakrishna Godavarthi
Date: Fri Jan 11 2019 - 09:23:51 EST

Hi Matthias,

On 2019-01-11 02:13, Matthias Kaehlcke wrote:
Hi Balakrishna,

On Thu, Jan 10, 2019 at 08:30:43PM +0530, Balakrishna Godavarthi wrote:
Hi Matthias,

On 2019-01-03 03:45, Matthias Kaehlcke wrote:
> On Mon, Dec 31, 2018 at 11:34:46AM +0530, Balakrishna Godavarthi wrote:
> > Hi Marcel,
> >
> > On 2018-12-30 13:40, Marcel Holtmann wrote:
> > > Hi Balakrishna,
> > >
> > > > > > Latest qualcomm chips are not sending an command complete event for
> > > > > > every firmware packet sent to chip. They only respond with a vendor
> > > > > > specific event for the last firmware packet. This optimization will
> > > > > > decrease the BT ON time. Due to this we are seeing a timeout error
> > > > > > message logs on the console during firmware download. Now we are
> > > > > > injecting a command complete event once we receive an vendor
> > > > > > specific
> > > > > > event for the last RAM firmware packet.
> > > > > > Signed-off-by: Balakrishna Godavarthi <bgodavar@xxxxxxxxxxxxxx>
> > > > > > ---
> > > > > > drivers/bluetooth/btqca.c | 39
> > > > > > ++++++++++++++++++++++++++++++++++++++-
> > > > > > drivers/bluetooth/btqca.h | 3 +++
> > > > > > 2 files changed, 41 insertions(+), 1 deletion(-)
> > > > > > diff --git a/drivers/bluetooth/btqca.c b/drivers/bluetooth/btqca.c
> > > > > > index ec9e03a6b778..0b533f65f652 100644
> > > > > > --- a/drivers/bluetooth/btqca.c
> > > > > > +++ b/drivers/bluetooth/btqca.c
> > > > > > @@ -144,6 +144,7 @@ static void qca_tlv_check_data(struct
> > > > > > rome_config *config,
> > > > > > * In case VSE is skipped, only the last segment is acked.
> > > > > > */
> > > > > > config->dnld_mode = tlv_patch->download_mode;
> > > > > > + config->dnld_type = config->dnld_mode;
> > > > > > BT_DBG("Total Length : %d bytes",
> > > > > > le32_to_cpu(tlv_patch->total_size));
> > > > > > @@ -264,6 +265,31 @@ static int qca_tlv_send_segment(struct
> > > > > > hci_dev *hdev, int seg_size,
> > > > > > return err;
> > > > > > }
> > > > > > +static int qca_inject_cmd_complete_event(struct hci_dev *hdev)
> > > > > > +{
> > > > > > + struct hci_event_hdr *hdr;
> > > > > > + struct hci_ev_cmd_complete *evt;
> > > > > > + struct sk_buff *skb;
> > > > > > +
> > > > > > + skb = bt_skb_alloc(sizeof(*hdr) + sizeof(*evt) + 1, GFP_KERNEL);
> > > > > > + if (!skb)
> > > > > > + return -ENOMEM;
> > > > > > +
> > > > > > + hdr = skb_put(skb, sizeof(*hdr));
> > > > > > + hdr->evt = HCI_EV_CMD_COMPLETE;
> > > > > > + hdr->plen = sizeof(*evt) + 1;
> > > > > > +
> > > > > > + evt = skb_put(skb, sizeof(*evt));
> > > > > > + evt->ncmd = 1;
> > > > > > + evt->opcode = HCI_OP_NOP;
>
> After looking a bit more at it I realize HCI_OP_NOP is not a good
> value in this case:
>
> static void hci_cmd_complete_evt(...)
> {
> ...
>
> if (*opcode != HCI_OP_NOP)
> cancel_delayed_work(&hdev->cmd_timer);
>
> ...
> }
>
> https://elixir.bootlin.com/linux/v4.19/source/net/bluetooth/hci_event.c#L3351
>
> Cancelling the command timeout is precisely what we want. Not sure why
> the patch with HCI_OP_NOP makes the timeouts go away in most cases
> (but not e.g. when inserting an msleep(1000) after downloading the
> NVM.
>
> I suggest to pass the opcode of the command to be completed.
>
> > > > > > +
> > > > > > + skb_put_u8(skb, QCA_HCI_CC_SUCCESS);
> > > > > > +
> > > > > > + hci_skb_pkt_type(skb) = HCI_EVENT_PKT;
> > > > > > +
> > > > > > + return hci_recv_frame(hdev, skb);
> > > > > > +}
> > > > > > +
> > > > > > static int qca_download_firmware(struct hci_dev *hdev,
> > > > > > struct rome_config *config)
> > > > > > {
> > > > > > @@ -297,11 +323,22 @@ static int
> > > > > > qca_download_firmware(struct hci_dev *hdev,
> > > > > > ret = qca_tlv_send_segment(hdev, segsize, segment,
> > > > > > config->dnld_mode);
> > > > > > if (ret)
> > > > > > - break;
> > > > > > + goto out;
> > > > > > segment += segsize;
> > > > > > }
> > > > > > + /* Latest qualcomm chipsets are not sending a command
> > > > > > complete event
> > > > > > + * for every fw packet sent. They only respond with a
> > > > > > vendor specific
> > > > > > + * event for the last packet. This optimization in the chip will
> > > > > > + * decrease the BT in initialization time. Here we will
> > > > > > inject a command
> > > > > > + * complete event to avoid a command timeout error message.
> > > > > > + */
> > > > > > + if ((config->dnld_type == ROME_SKIP_EVT_VSE_CC ||
> > > > > > + config->dnld_type == ROME_SKIP_EVT_VSE))
> > > > > > + return qca_inject_cmd_complete_event(hdev);
> > > > > > +
> > > > > have you actually considered using __hci_cmd_send in that case. It is
> > > > > allowed for vendor OGF to use that command. I see you actually do use
> > > > > it and now I am failing to understand what this is for.
> > > > [Bala]: thanks for reviewing the change.
> > > >
> > > > __hci_cmd_send() can be used only to send the command to the chip.
> > > > it will not wait for the response for the command sent.
> > > >
> > > > as you know that every vendor command sent to chip will respond with
> > > > vendor specific event and command complete event.
> > > > but in our case chip will only respond with vendor specific event
> > > > only. so we are injecting command complete event.
> > >
> > > and __hci_cmd_sync_ev is also not working for you? However since you
> > > are not waiting for the vendor event anyway and just injecting
> > > cmd_complete, I wonder whatâs the difference in just using
> > > __hci_cmd_send and not bothering to wait or inject at all. I am
> > > failing to see where this injection makes a difference.
> > >
> > > For me it is a big difference if we are injecting one event like in
> > > the case of Intel compared to injecting one for every command. It will
> > > show a wrong picture in btmon and that is a bad idea.
> > >
> > > Regards
> > >
> > > Marcel
> >
> > [Bala]: here is the use case, when ever we download the fw packets
> > i.e. RAM
> > image, for every command sent(i.e. fw packet) from
> > the host chip will respond with an vendor specific event and command
> > complete event.
> >
> > the above is taking more time to setup the BT device. then we came
> > up with
> > solution where we enable flags in fw file (i.e. RAM image header)
> > whether to wait for event to be received or sent the total packets
> > and wait
> > for the events for the last packet.
> >
> > So currently we are handling both the cases in the code. i.e wait
> > for event
> > for all packet or wait for an event for the last packet.
> >
> > but in the second case i.e. wait for event for the last packet sent,
> > we are
> > only receiving an vendor specific event from chip which holds the
> > status of
> > fw download.
> >
> > so as __hci_cmd_sync_ev() requires an command complete event. so we
> > are
> > injecting it after the vendor specific event received for the last
> > packet.
> >
> > This helps to overcome 0xfc00 timeout error logging on console.
>
> Some more details:
>
> The timeout error is actually from reading the 'SoC version', which
> uses the same command code as the firmware download
> (EDL_PATCH_CMD_OPCODE). Without reading the 'SoC version' it would be
> from the command to write the first firmware segment.
>
> If the download of a firmware binary takes >= 2s (HCI_CMD_TIMEOUT) the
> timeout would still occur. If necessary this could be mitigated by
> injecting some command complete events during the firmware download,
> though I expect Marcel wouldn't be overly happy with that, since it
> would affect btmon even more.
>
> Regards
>
> Matthias

[Bala]: Basically every vendor specific command we sent to chip,
chip should respond with an vendor specific event followed by an command
complete event
or some times it will only respond with an command complete event.
but in any case command complete event is mandatory to all the command we
sent to the chip.

Is this ("command complete event is mandatory to all the command we
sent to the chip") a description of what the chip actually does, or
what it should be doing according to the spec?

As mentioned earlier, the timeout we see originates from reading the
SoC version:

diff --git a/drivers/bluetooth/btqca.c b/drivers/bluetooth/btqca.c
index 0b533f65f652fc..1e484e61799571 100644
--- a/drivers/bluetooth/btqca.c
+++ b/drivers/bluetooth/btqca.c
@@ -400,6 +400,10 @@ int qca_uart_setup(struct hci_dev *hdev, uint8_t baudrate,
return err;
}

+ printk("DBG: ZZZzzzzzzz\n");
+ msleep(2500);
+ printk("DBG: good morning!\n");
+
/* Download NVM configuration */
config.type = TLV_TYPE_NVM;
if (soc_type == QCA_WCN3990)

When you boot with this patch you'll see something like this:

[ 15.531365] DBG: reading SoC version
[ 15.544963] DBG: ZZZzzzzzzz
[ 17.590282] Bluetooth: hci0: command 0xfc00 tx timeout
[ 18.099110] DBG: good morning!


[Bala]: my previous analysis were wrong. thanks for pointing me to the correct issue.
i am able to see timeout after version command.(used some sleep)
here is the reason for it.

01 00 fc 01 19 (we send the command to chip to request version)
04 ff 0e 00 02 0a 00 00 00 01 00 01 02 14 02 01 40(chip will respond with the vendor specific event payload will be chip version)
04 0e 04 01 00 00 00(command complete event)

issue is with command complete event. ideally command complete event payload holds the command for which chip sends command complete event.

breaking the command

04 : event packet (fixed)
0e : command complete event(fixed)
04 : size of the payload,

01 : no of commands in the payload
00 : OCF (opocde lsb)
00 : OGF (opcode msb)
00 : status of command executed,

payload is form byte 4 to byte 7. which hold the command i.e opcode
form the above response opcode is 0x0000. ideally this should be 0xfc00
because the command complete response from the chip is for the command 0xfc00.

Checked with Chip firmware dev team for the reason, it was an limitation or an bug in the chip ROM firmware
in the current chipset, which is fixed in the coming chipset.

so we expect the command 0xfc00 but recevuies 0x00. so that is the reason we see an timeouts.
for all the vendor commands we receive 0x0000 in the command complete event.

if comes a new question how are we not seeing an 0xfc00 when we inject an command complete event.
i experimented this patch, when i inject the command complete event from the soc version still i can see the command timeouts.

The reason for no timeouts error, we are injecting an command complete after last packet sent out, so this injection helps
.bin file to speed out it download part, so as the bin is dumped fastly we send an HCI RESET Comamnd where we recevie an command complete event
with non zero opcode which is cancelling the command timeout timer.

when we inject command complete event, in short we are completing setup process in less than 2 seconds(command timeout timer value),
so before 2 seconds we are sending an HCI RESET command whose command complete event opcode value is an non zero (which cancels the command timeout timer)


In our case, we have an two fw files i.e. *.tlv and *.bin.
tlv is an RAM image of the chip where as bin is an nvm image of the chip. so
tlv will be of
more size which require an lot more time to dump the file in to chip,
while dumping the tlv, we divide tlv as packet of size 245 bytes and send
them as an command packet to the chip. chip should respond with an command
complete event.
then only we will send the next packet. but size of the tlv is large, to
optimize this we will
not wait for the either an vendor specific event or an command complete
event.

Let's make sure we have an accurate picture, which of the following is
correct:

1. the chip sends a command complete event after each packet, as an
optimization the BT driver doesn't wait for it

2. as an optimization the chip does not send a command complete event
and the driver has to deal with it

My understanding is that it's 2), but the wording above seems to
describe 1)

[Bala]: point 2 is true.

But as we need to be on the sync, i.e. whether are we sending an correct
packets or not,
for the last fw packet we sent to the chip.. chip will to do an CRC check
for the total no of packets received and respond with an vendor specific
event.

We decode the vendor specific event and decide whether the fw download is
success or not. here we send an fw packet as command so stack expects an
command complete event.
where this is missing from the chip. this is expected behavior from chip.

So currently i am inject an command complete event after receiving an vendor
event for the last packet of the tlv.

And the same for the .bin if I'm not mistaken.

This we inject only once for the last command packet sent to the chip.
i don't think this will effect the btmon.

I don't know enough about btmon to comment on that, in any case Marcel
raised concerns.

And I think my comment that triggered this disucssion remains true:

If the download of a firmware binary takes >= 2s (HCI_CMD_TIMEOUT) the
timeout would still occur. If necessary this could be mitigated by
injecting some command complete events during the firmware download.

Not sure it's a likely case, it might be an issue with larger firmware
files and/or slower UART speeds.

Thanks

Matthias

[Bala]: i don't think it is a good idea to handle this kind of limitations in the HOST driver.

Thanks for pointing me the actual issue.

--
Regards
Balakrishna.