The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
even if UART isn't console") worked pretty well and I've been doing a
lot of debugging with it. However, recently I typed "dmesg" in kdb
and then held the space key down to scroll through the pagination. My
device hung. This was repeatable and I found that it was introduced
with the aforementioned commit.
It turns out that there are some strange boundary cases in geni where
in some weird situations it will signal RX_LAST but then will put 0 in
RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
This weird corner case is handled in qcom_geni_serial_handle_rx()
where you can see that we only honor RX_LAST_BYTE if RX_LAST is set
_and_ RX_LAST_BYTE is non-zero. If either of these is not true we use
BYTES_PER_FIFO_WORD (4) for the size of the last FIFO word.
Let's fix kgdb. While at it, also use the proper #define for 4.
Fixes: e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work even if UART isn't console")
Signed-off-by: Douglas Anderson <[email protected]>
---
drivers/tty/serial/qcom_geni_serial.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/drivers/tty/serial/qcom_geni_serial.c b/drivers/tty/serial/qcom_geni_serial.c
index 07b7b6b05b8b..e27077656939 100644
--- a/drivers/tty/serial/qcom_geni_serial.c
+++ b/drivers/tty/serial/qcom_geni_serial.c
@@ -361,11 +361,16 @@ static int qcom_geni_serial_get_char(struct uart_port *uport)
return NO_POLL_CHAR;
if (word_cnt == 1 && (status & RX_LAST))
+ /*
+ * NOTE: If RX_LAST_BYTE_VALID is 0 it needs to be
+ * treated as if it was BYTES_PER_FIFO_WORD.
+ */
private_data->poll_cached_bytes_cnt =
(status & RX_LAST_BYTE_VALID_MSK) >>
RX_LAST_BYTE_VALID_SHFT;
- else
- private_data->poll_cached_bytes_cnt = 4;
+
+ if (private_data->poll_cached_bytes_cnt == 0)
+ private_data->poll_cached_bytes_cnt = BYTES_PER_FIFO_WORD;
private_data->poll_cached_bytes =
readl(uport->membase + SE_GENI_RX_FIFOn);
--
2.28.0.236.gb10cc79966-goog
Hi Doug,
On 8/7/2020 10:49 AM, Douglas Anderson wrote:
> The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
> even if UART isn't console") worked pretty well and I've been doing a
> lot of debugging with it. However, recently I typed "dmesg" in kdb
> and then held the space key down to scroll through the pagination. My
> device hung. This was repeatable and I found that it was introduced
> with the aforementioned commit.
>
> It turns out that there are some strange boundary cases in geni where
> in some weird situations it will signal RX_LAST but then will put 0 in
> RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
IMO that means we received a word in RX_FIFO and it is the last word
hence RX_LAST bit is set.
RX_LAST_BYTE is 0 means none of the bytes are valid in the last word.
In such scenario we should just read RX_FIFO buffer (to empty it),
discard the word and return NO_POLL_CHAR. Something like below.
---------------------------------------------------------------------------------------------------------------------------------------------------------
else
private_data->poll_cached_bytes_cnt = 4;
private_data->poll_cached_bytes =
readl(uport->membase + SE_GENI_RX_FIFOn);
}
+ if (!private_data->poll_cached_bytes_cnt)
+ return NO_POLL_CHAR;
private_data->poll_cached_bytes_cnt--;
ret = private_data->poll_cached_bytes & 0xff;
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Please let me know whether above code helps.
I am not sure about what all scenario can leads to this behavior from
hardware, I will try to get an answer from hardware team.
Any error bit was set for SE_GENI_S_IRQ_STATUS & SE_GENI_M_IRQ_STATUS
registers?
I guess the hang was seen because *poll_cached_bytes_cnt* is unsigned
int and it's value was 0, when it's decremented by 1 it's value become
'4294967295' (very large) and dummy RX (0x00) would happen that
many times before reading any actual RX transfers/bytes.
Regards,
Akash
> This weird corner case is handled in qcom_geni_serial_handle_rx()
> where you can see that we only honor RX_LAST_BYTE if RX_LAST is set
> _and_ RX_LAST_BYTE is non-zero. If either of these is not true we use
> BYTES_PER_FIFO_WORD (4) for the size of the last FIFO word.
>
> Let's fix kgdb. While at it, also use the proper #define for 4.
>
> Fixes: e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work even if UART isn't console")
> Signed-off-by: Douglas Anderson <[email protected]>
> ---
>
> drivers/tty/serial/qcom_geni_serial.c | 9 +++++++--
> 1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/tty/serial/qcom_geni_serial.c b/drivers/tty/serial/qcom_geni_serial.c
> index 07b7b6b05b8b..e27077656939 100644
> --- a/drivers/tty/serial/qcom_geni_serial.c
> +++ b/drivers/tty/serial/qcom_geni_serial.c
> @@ -361,11 +361,16 @@ static int qcom_geni_serial_get_char(struct uart_port *uport)
> return NO_POLL_CHAR;
>
> if (word_cnt == 1 && (status & RX_LAST))
> + /*
> + * NOTE: If RX_LAST_BYTE_VALID is 0 it needs to be
> + * treated as if it was BYTES_PER_FIFO_WORD.
> + */
> private_data->poll_cached_bytes_cnt =
> (status & RX_LAST_BYTE_VALID_MSK) >>
> RX_LAST_BYTE_VALID_SHFT;
> - else
> - private_data->poll_cached_bytes_cnt = 4;
> +
> + if (private_data->poll_cached_bytes_cnt == 0)
> + private_data->poll_cached_bytes_cnt = BYTES_PER_FIFO_WORD;
>
> private_data->poll_cached_bytes =
> readl(uport->membase + SE_GENI_RX_FIFOn);
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project
Hi,
On Mon, Aug 10, 2020 at 5:32 AM Akash Asthana <[email protected]> wrote:
>
> Hi Doug,
>
> On 8/7/2020 10:49 AM, Douglas Anderson wrote:
> > The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
> > even if UART isn't console") worked pretty well and I've been doing a
> > lot of debugging with it. However, recently I typed "dmesg" in kdb
> > and then held the space key down to scroll through the pagination. My
> > device hung. This was repeatable and I found that it was introduced
> > with the aforementioned commit.
> >
> > It turns out that there are some strange boundary cases in geni where
> > in some weird situations it will signal RX_LAST but then will put 0 in
> > RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
>
> IMO that means we received a word in RX_FIFO and it is the last word
> hence RX_LAST bit is set.
What you say would make logical sense, but it's not how I have
observed geni to work. See below.
> RX_LAST_BYTE is 0 means none of the bytes are valid in the last word.
This would imply that qcom_geni_serial_handle_rx() is also broken
though, wouldn't it? Specifically imagine that WORD_CNT is 1 and
RX_LAST is set and RX_LAST_BYTE_VALID is true. Here's the logic from
that function:
total_bytes = BYTES_PER_FIFO_WORD * (word_cnt - 1);
if (last_word_partial && last_word_byte_cnt)
total_bytes += last_word_byte_cnt;
else
total_bytes += BYTES_PER_FIFO_WORD;
port->handle_rx(uport, total_bytes, drop);
As you can see that logic will set "total_bytes" to 4 in the case I'm
talking about.
> In such scenario we should just read RX_FIFO buffer (to empty it),
> discard the word and return NO_POLL_CHAR. Something like below.
>
> ---------------------------------------------------------------------------------------------------------------------------------------------------------
>
> else
> private_data->poll_cached_bytes_cnt = 4;
>
> private_data->poll_cached_bytes =
> readl(uport->membase + SE_GENI_RX_FIFOn);
> }
>
> + if (!private_data->poll_cached_bytes_cnt)
> + return NO_POLL_CHAR;
> private_data->poll_cached_bytes_cnt--;
> ret = private_data->poll_cached_bytes & 0xff;
> -------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Please let me know whether above code helps.
Your code will avoid the hang. Yes. ...but it will drop bytes. I
devised a quick-n-dirty test. Here's a test of your code:
https://crrev.com/c/2346886
...and here's a test of my code:
https://crrev.com/c/2346884
I had to keep a buffer around since it's hard to debug the serial
driver. In both cases I put "DOUG" into the buffer when I detect this
case. If my theory about how geni worked was wrong then we should
expect to see some garbage in the buffer right after the DOUG, right?
...but my code gets the alphabet in nice sequence. Your code drops 4
bytes.
NOTE: while poking around with the above two test patches I found it
was pretty easy to get geni to drop bytes / hit overflow cases and
also to insert bogus 0 bytes in the stream (I believe these are
related). I was able to reproduce this:
* With ${SUBJECT} patch in place.
* With your proposed patch.
* With the recent "geni" patches reverted (in other words back to 1
byte per FIFO entry).
It's not terribly surprising that we're overflowing since I believe
kgdb isn't too keen to read characters at the same time it's writing.
That doesn't explain the weird 0-bytes that geni seemed to be
inserting, but at least it would explain the overflows. However, even
after I fixed this I _still_ was getting problems. Specifically geni
seemed to be hiding bytes from me until it was too late. I put
logging in and would see this:
1 word in FIFO - wxyz
1 word in FIFO (last set, last FIFO has 1 byte) - \n
Check again, still 0 bytes in FIFO
Suddenly 16 bytes are in FIFO and S_RX_FIFO_WR_ERR_EN is set.
I spent a whole bunch of time poking at this and couldn't find any
sort of workaround. Presumably geni is taking some time between me
reading the last word out of the FIFO from the "previous" packet and
then transitioning to the new packet. I found a lot of references to
this process in the hardware register description (see GENI_CFG_REG69,
for instance), but I couldn't manage to make the kick to happen any
faster. Presumably this isn't a problem for things like Bluetooth
since flow control saves them. ...and I guess this isn't a problem in
practice because we usually _send_ a lot of data to the host for
console/kgdb and it's only the host => DUT path that has problems.
> I am not sure about what all scenario can leads to this behavior from
> hardware, I will try to get an answer from hardware team.
>
> Any error bit was set for SE_GENI_S_IRQ_STATUS & SE_GENI_M_IRQ_STATUS
> registers?
As per above I can see overflows in my test case and geni seems to be
behaving pretty badly. If you have ideas on how to fix this I'd love
it. However, it still seems like my patch is right because (at least
in the cases I tested) it avoids dropping bytes in some cases. It
also matches how qcom_geni_serial_handle_rx() works and if that was
broken we'd have noticed by now.
> I guess the hang was seen because *poll_cached_bytes_cnt* is unsigned
> int and it's value was 0, when it's decremented by 1 it's value become
> '4294967295' (very large) and dummy RX (0x00) would happen that
>
> many times before reading any actual RX transfers/bytes.
Right. That would be why it was hanging.
-Doug
On 8/11/2020 2:56 AM, Doug Anderson wrote:
> Hi,
>
> On Mon, Aug 10, 2020 at 5:32 AM Akash Asthana <[email protected]> wrote:
>> Hi Doug,
>>
>> On 8/7/2020 10:49 AM, Douglas Anderson wrote:
>>> The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
>>> even if UART isn't console") worked pretty well and I've been doing a
>>> lot of debugging with it. However, recently I typed "dmesg" in kdb
>>> and then held the space key down to scroll through the pagination. My
>>> device hung. This was repeatable and I found that it was introduced
>>> with the aforementioned commit.
>>>
>>> It turns out that there are some strange boundary cases in geni where
>>> in some weird situations it will signal RX_LAST but then will put 0 in
>>> RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
>> IMO that means we received a word in RX_FIFO and it is the last word
>> hence RX_LAST bit is set.
> What you say would make logical sense, but it's not how I have
> observed geni to work. See below.
>
>
>> RX_LAST_BYTE is 0 means none of the bytes are valid in the last word.
> This would imply that qcom_geni_serial_handle_rx() is also broken
> though, wouldn't it? Specifically imagine that WORD_CNT is 1 and
> RX_LAST is set and RX_LAST_BYTE_VALID is true. Here's the logic from
> that function:
>
> total_bytes = BYTES_PER_FIFO_WORD * (word_cnt - 1);
> if (last_word_partial && last_word_byte_cnt)
> total_bytes += last_word_byte_cnt;
> else
> total_bytes += BYTES_PER_FIFO_WORD;
> port->handle_rx(uport, total_bytes, drop);
>
> As you can see that logic will set "total_bytes" to 4 in the case I'm
> talking about.
Yeah IMO as per theory this should also be corrected but since you have
already pulled out few experiment to prove garbage data issue(which I
was suspecting) is not seen.
It's already consistent with existing logic and it behaves well
practically . So the changes could be merge. Meanwhile I am checking
with HW team to get clarity.
>
>
>> In such scenario we should just read RX_FIFO buffer (to empty it),
>> discard the word and return NO_POLL_CHAR. Something like below.
>>
>> ---------------------------------------------------------------------------------------------------------------------------------------------------------
>>
>> else
>> private_data->poll_cached_bytes_cnt = 4;
>>
>> private_data->poll_cached_bytes =
>> readl(uport->membase + SE_GENI_RX_FIFOn);
>> }
>>
>> + if (!private_data->poll_cached_bytes_cnt)
>> + return NO_POLL_CHAR;
>> private_data->poll_cached_bytes_cnt--;
>> ret = private_data->poll_cached_bytes & 0xff;
>> -------------------------------------------------------------------------------------------------------------------------------------------------------------
>>
>> Please let me know whether above code helps.
> Your code will avoid the hang. Yes. ...but it will drop bytes. I
> devised a quick-n-dirty test. Here's a test of your code:
I assumed those as invalid bytes and don't wanted to read them so yeah
dropping of bytes was expected.
>
> https://crrev.com/c/2346886
>
> ...and here's a test of my code:
>
> https://crrev.com/c/2346884
>
> I had to keep a buffer around since it's hard to debug the serial
> driver. In both cases I put "DOUG" into the buffer when I detect this
> case. If my theory about how geni worked was wrong then we should
> expect to see some garbage in the buffer right after the DOUG, right?
> ...but my code gets the alphabet in nice sequence. Your code drops 4
> bytes.
Yeah I was expecting garbage data.
>
>
> NOTE: while poking around with the above two test patches I found it
> was pretty easy to get geni to drop bytes / hit overflow cases and
> also to insert bogus 0 bytes in the stream (I believe these are
> related). I was able to reproduce this:
> * With ${SUBJECT} patch in place.
> * With your proposed patch.
> * With the recent "geni" patches reverted (in other words back to 1
> byte per FIFO entry).
>
> It's not terribly surprising that we're overflowing since I believe
> kgdb isn't too keen to read characters at the same time it's writing.
> That doesn't explain the weird 0-bytes that geni seemed to be
> inserting, but at least it would explain the overflows. However, even
> after I fixed this I _still_ was getting problems. Specifically geni
> seemed to be hiding bytes from me until it was too late. I put
> logging in and would see this:
>
> 1 word in FIFO - wxyz
> 1 word in FIFO (last set, last FIFO has 1 byte) - \n
> Check again, still 0 bytes in FIFO
> Suddenly 16 bytes are in FIFO and S_RX_FIFO_WR_ERR_EN is set.
RX data first stored in RX_ASYNC_FIFO then it's transfered to RX_FIFO
When get_char is called and we observe 0 bytes in RX_FIFO, most probably
data is not transfered from RX_ASYNC_FIFO to RX_FIFO.
BITS 27:25 of SE_GENI_RX_FIFO_STATUS register shows RX_ASYNC_FIFO word
count.
>
> I spent a whole bunch of time poking at this and couldn't find any
> sort of workaround. Presumably geni is taking some time between me
> reading the last word out of the FIFO from the "previous" packet and
> then transitioning to the new packet. I found a lot of references to
> this process in the hardware register description (see GENI_CFG_REG69,
> for instance), but I couldn't manage to make the kick to happen any
> faster. Presumably this isn't a problem for things like Bluetooth
> since flow control saves them. ...and I guess this isn't a problem in
> practice because we usually _send_ a lot of data to the host for
> console/kgdb and it's only the host => DUT path that has problems.
>
>
>> I am not sure about what all scenario can leads to this behavior from
>> hardware, I will try to get an answer from hardware team.
>>
>> Any error bit was set for SE_GENI_S_IRQ_STATUS & SE_GENI_M_IRQ_STATUS
>> registers?
> As per above I can see overflows in my test case and geni seems to be
> behaving pretty badly. If you have ideas on how to fix this I'd love
> it. However, it still seems like my patch is right because (at least
> in the cases I tested) it avoids dropping bytes in some cases. It
> also matches how qcom_geni_serial_handle_rx() works and if that was
> broken we'd have noticed by now.
Reviewed-by: Akash Asthana <[email protected]>
>
>> I guess the hang was seen because *poll_cached_bytes_cnt* is unsigned
>> int and it's value was 0, when it's decremented by 1 it's value become
>> '4294967295' (very large) and dummy RX (0x00) would happen that
>>
>> many times before reading any actual RX transfers/bytes.
> Right. That would be why it was hanging.
>
>
> -Doug
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,\na Linux Foundation Collaborative Project
Hi,
On Tue, Aug 11, 2020 at 4:54 AM Akash Asthana <[email protected]> wrote:
>
>
> On 8/11/2020 2:56 AM, Doug Anderson wrote:
> > Hi,
> >
> > On Mon, Aug 10, 2020 at 5:32 AM Akash Asthana <[email protected]> wrote:
> >> Hi Doug,
> >>
> >> On 8/7/2020 10:49 AM, Douglas Anderson wrote:
> >>> The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
> >>> even if UART isn't console") worked pretty well and I've been doing a
> >>> lot of debugging with it. However, recently I typed "dmesg" in kdb
> >>> and then held the space key down to scroll through the pagination. My
> >>> device hung. This was repeatable and I found that it was introduced
> >>> with the aforementioned commit.
> >>>
> >>> It turns out that there are some strange boundary cases in geni where
> >>> in some weird situations it will signal RX_LAST but then will put 0 in
> >>> RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
> >> IMO that means we received a word in RX_FIFO and it is the last word
> >> hence RX_LAST bit is set.
> > What you say would make logical sense, but it's not how I have
> > observed geni to work. See below.
> >
> >
> >> RX_LAST_BYTE is 0 means none of the bytes are valid in the last word.
> > This would imply that qcom_geni_serial_handle_rx() is also broken
> > though, wouldn't it? Specifically imagine that WORD_CNT is 1 and
> > RX_LAST is set and RX_LAST_BYTE_VALID is true. Here's the logic from
> > that function:
> >
> > total_bytes = BYTES_PER_FIFO_WORD * (word_cnt - 1);
> > if (last_word_partial && last_word_byte_cnt)
> > total_bytes += last_word_byte_cnt;
> > else
> > total_bytes += BYTES_PER_FIFO_WORD;
> > port->handle_rx(uport, total_bytes, drop);
> >
> > As you can see that logic will set "total_bytes" to 4 in the case I'm
> > talking about.
>
> Yeah IMO as per theory this should also be corrected but since you have
> already pulled out few experiment to prove garbage data issue(which I
> was suspecting) is not seen.
>
> It's already consistent with existing logic and it behaves well
> practically . So the changes could be merge. Meanwhile I am checking
> with HW team to get clarity.
>
> >
> >
> >> In such scenario we should just read RX_FIFO buffer (to empty it),
> >> discard the word and return NO_POLL_CHAR. Something like below.
> >>
> >> ---------------------------------------------------------------------------------------------------------------------------------------------------------
> >>
> >> else
> >> private_data->poll_cached_bytes_cnt = 4;
> >>
> >> private_data->poll_cached_bytes =
> >> readl(uport->membase + SE_GENI_RX_FIFOn);
> >> }
> >>
> >> + if (!private_data->poll_cached_bytes_cnt)
> >> + return NO_POLL_CHAR;
> >> private_data->poll_cached_bytes_cnt--;
> >> ret = private_data->poll_cached_bytes & 0xff;
> >> -------------------------------------------------------------------------------------------------------------------------------------------------------------
> >>
> >> Please let me know whether above code helps.
> > Your code will avoid the hang. Yes. ...but it will drop bytes. I
> > devised a quick-n-dirty test. Here's a test of your code:
> I assumed those as invalid bytes and don't wanted to read them so yeah
> dropping of bytes was expected.
> >
> > https://crrev.com/c/2346886
> >
> > ...and here's a test of my code:
> >
> > https://crrev.com/c/2346884
> >
> > I had to keep a buffer around since it's hard to debug the serial
> > driver. In both cases I put "DOUG" into the buffer when I detect this
> > case. If my theory about how geni worked was wrong then we should
> > expect to see some garbage in the buffer right after the DOUG, right?
> > ...but my code gets the alphabet in nice sequence. Your code drops 4
> > bytes.
> Yeah I was expecting garbage data.
> >
> >
> > NOTE: while poking around with the above two test patches I found it
> > was pretty easy to get geni to drop bytes / hit overflow cases and
> > also to insert bogus 0 bytes in the stream (I believe these are
> > related). I was able to reproduce this:
> > * With ${SUBJECT} patch in place.
> > * With your proposed patch.
> > * With the recent "geni" patches reverted (in other words back to 1
> > byte per FIFO entry).
> >
> > It's not terribly surprising that we're overflowing since I believe
> > kgdb isn't too keen to read characters at the same time it's writing.
> > That doesn't explain the weird 0-bytes that geni seemed to be
> > inserting, but at least it would explain the overflows. However, even
> > after I fixed this I _still_ was getting problems. Specifically geni
> > seemed to be hiding bytes from me until it was too late. I put
> > logging in and would see this:
> >
> > 1 word in FIFO - wxyz
> > 1 word in FIFO (last set, last FIFO has 1 byte) - \n
> > Check again, still 0 bytes in FIFO
> > Suddenly 16 bytes are in FIFO and S_RX_FIFO_WR_ERR_EN is set.
>
> RX data first stored in RX_ASYNC_FIFO then it's transfered to RX_FIFO
>
> When get_char is called and we observe 0 bytes in RX_FIFO, most probably
> data is not transfered from RX_ASYNC_FIFO to RX_FIFO.
>
> BITS 27:25 of SE_GENI_RX_FIFO_STATUS register shows RX_ASYNC_FIFO word
> count.
OK, I did a tad bit more debugging and I think any problems left after
my patch are actually kdb's fault, though geni doesn't behave terribly
well with overflows (it would be better if it just dropped characters
rather than enqueuing garbage). Specifically in the tests I was
running recently kdb would sometimes still take over 30 ms between
polls so it's not surprising that we'd get overflows if we're
spamming. All my test code and results are at:
https://crrev.com/c/2348284
Earlier when I claimed that bytes were showing up in geni too quickly
I was wrong. My timing code was broken at the time.
> > I spent a whole bunch of time poking at this and couldn't find any
> > sort of workaround. Presumably geni is taking some time between me
> > reading the last word out of the FIFO from the "previous" packet and
> > then transitioning to the new packet. I found a lot of references to
> > this process in the hardware register description (see GENI_CFG_REG69,
> > for instance), but I couldn't manage to make the kick to happen any
> > faster. Presumably this isn't a problem for things like Bluetooth
> > since flow control saves them. ...and I guess this isn't a problem in
> > practice because we usually _send_ a lot of data to the host for
> > console/kgdb and it's only the host => DUT path that has problems.
> >
> >
> >> I am not sure about what all scenario can leads to this behavior from
> >> hardware, I will try to get an answer from hardware team.
> >>
> >> Any error bit was set for SE_GENI_S_IRQ_STATUS & SE_GENI_M_IRQ_STATUS
> >> registers?
> > As per above I can see overflows in my test case and geni seems to be
> > behaving pretty badly. If you have ideas on how to fix this I'd love
> > it. However, it still seems like my patch is right because (at least
> > in the cases I tested) it avoids dropping bytes in some cases. It
> > also matches how qcom_geni_serial_handle_rx() works and if that was
> > broken we'd have noticed by now.
>
> Reviewed-by: Akash Asthana <[email protected]>
Thanks! After my extra digging I now think that the weird case is
triggered by the overflows and the FIFO certainly does get garbage in
it when there are overflows, but I think my patch is still correct and
gets correct bytes. The garbage that is queued up is queued up later
and looks like normal transfers. ...so I think we're good to go with
it.
So summary:
1. We should land my patch.
2. I filed <https://crbug.com/1115125> to track trying to make kdb's
polling nicer.
3. I also filed an internal bug to track seeing if we could make geni
not queue up garbage in this case.
-Doug
On Tue, Aug 11, 2020 at 09:21:22AM -0700, Doug Anderson wrote:
> Hi,
>
> On Tue, Aug 11, 2020 at 4:54 AM Akash Asthana <[email protected]> wrote:
> >
> >
> > On 8/11/2020 2:56 AM, Doug Anderson wrote:
> > > Hi,
> > >
> > > On Mon, Aug 10, 2020 at 5:32 AM Akash Asthana <[email protected]> wrote:
> > >> Hi Doug,
> > >>
> > >> On 8/7/2020 10:49 AM, Douglas Anderson wrote:
> > >>> The commit e42d6c3ec0c7 ("serial: qcom_geni_serial: Make kgdb work
> > >>> even if UART isn't console") worked pretty well and I've been doing a
> > >>> lot of debugging with it. However, recently I typed "dmesg" in kdb
> > >>> and then held the space key down to scroll through the pagination. My
> > >>> device hung. This was repeatable and I found that it was introduced
> > >>> with the aforementioned commit.
> > >>>
> > >>> It turns out that there are some strange boundary cases in geni where
> > >>> in some weird situations it will signal RX_LAST but then will put 0 in
> > >>> RX_LAST_BYTE. This means that the entire last FIFO entry is valid.
> > >> IMO that means we received a word in RX_FIFO and it is the last word
> > >> hence RX_LAST bit is set.
> > > What you say would make logical sense, but it's not how I have
> > > observed geni to work. See below.
> > >
> > >
> > >> RX_LAST_BYTE is 0 means none of the bytes are valid in the last word.
> > > This would imply that qcom_geni_serial_handle_rx() is also broken
> > > though, wouldn't it? Specifically imagine that WORD_CNT is 1 and
> > > RX_LAST is set and RX_LAST_BYTE_VALID is true. Here's the logic from
> > > that function:
> > >
> > > total_bytes = BYTES_PER_FIFO_WORD * (word_cnt - 1);
> > > if (last_word_partial && last_word_byte_cnt)
> > > total_bytes += last_word_byte_cnt;
> > > else
> > > total_bytes += BYTES_PER_FIFO_WORD;
> > > port->handle_rx(uport, total_bytes, drop);
> > >
> > > As you can see that logic will set "total_bytes" to 4 in the case I'm
> > > talking about.
> >
> > Yeah IMO as per theory this should also be corrected but since you have
> > already pulled out few experiment to prove garbage data issue(which I
> > was suspecting) is not seen.
> >
> > It's already consistent with existing logic and it behaves well
> > practically . So the changes could be merge. Meanwhile I am checking
> > with HW team to get clarity.
> >
> > >
> > >
> > >> In such scenario we should just read RX_FIFO buffer (to empty it),
> > >> discard the word and return NO_POLL_CHAR. Something like below.
> > >>
> > >> ---------------------------------------------------------------------------------------------------------------------------------------------------------
> > >>
> > >> else
> > >> private_data->poll_cached_bytes_cnt = 4;
> > >>
> > >> private_data->poll_cached_bytes =
> > >> readl(uport->membase + SE_GENI_RX_FIFOn);
> > >> }
> > >>
> > >> + if (!private_data->poll_cached_bytes_cnt)
> > >> + return NO_POLL_CHAR;
> > >> private_data->poll_cached_bytes_cnt--;
> > >> ret = private_data->poll_cached_bytes & 0xff;
> > >> -------------------------------------------------------------------------------------------------------------------------------------------------------------
> > >>
> > >> Please let me know whether above code helps.
> > > Your code will avoid the hang. Yes. ...but it will drop bytes. I
> > > devised a quick-n-dirty test. Here's a test of your code:
> > I assumed those as invalid bytes and don't wanted to read them so yeah
> > dropping of bytes was expected.
> > >
> > > https://crrev.com/c/2346886
> > >
> > > ...and here's a test of my code:
> > >
> > > https://crrev.com/c/2346884
> > >
> > > I had to keep a buffer around since it's hard to debug the serial
> > > driver. In both cases I put "DOUG" into the buffer when I detect this
> > > case. If my theory about how geni worked was wrong then we should
> > > expect to see some garbage in the buffer right after the DOUG, right?
> > > ...but my code gets the alphabet in nice sequence. Your code drops 4
> > > bytes.
> > Yeah I was expecting garbage data.
> > >
> > >
> > > NOTE: while poking around with the above two test patches I found it
> > > was pretty easy to get geni to drop bytes / hit overflow cases and
> > > also to insert bogus 0 bytes in the stream (I believe these are
> > > related). I was able to reproduce this:
> > > * With ${SUBJECT} patch in place.
> > > * With your proposed patch.
> > > * With the recent "geni" patches reverted (in other words back to 1
> > > byte per FIFO entry).
> > >
> > > It's not terribly surprising that we're overflowing since I believe
> > > kgdb isn't too keen to read characters at the same time it's writing.
> > > That doesn't explain the weird 0-bytes that geni seemed to be
> > > inserting, but at least it would explain the overflows. However, even
> > > after I fixed this I _still_ was getting problems. Specifically geni
> > > seemed to be hiding bytes from me until it was too late. I put
> > > logging in and would see this:
> > >
> > > 1 word in FIFO - wxyz
> > > 1 word in FIFO (last set, last FIFO has 1 byte) - \n
> > > Check again, still 0 bytes in FIFO
> > > Suddenly 16 bytes are in FIFO and S_RX_FIFO_WR_ERR_EN is set.
> >
> > RX data first stored in RX_ASYNC_FIFO then it's transfered to RX_FIFO
> >
> > When get_char is called and we observe 0 bytes in RX_FIFO, most probably
> > data is not transfered from RX_ASYNC_FIFO to RX_FIFO.
> >
> > BITS 27:25 of SE_GENI_RX_FIFO_STATUS register shows RX_ASYNC_FIFO word
> > count.
>
> OK, I did a tad bit more debugging and I think any problems left after
> my patch are actually kdb's fault, though geni doesn't behave terribly
> well with overflows (it would be better if it just dropped characters
> rather than enqueuing garbage). Specifically in the tests I was
> running recently kdb would sometimes still take over 30 ms between
> polls so it's not surprising that we'd get overflows if we're
> spamming. All my test code and results are at:
>
> https://crrev.com/c/2348284
>
> Earlier when I claimed that bytes were showing up in geni too quickly
> I was wrong. My timing code was broken at the time.
>
>
> > > I spent a whole bunch of time poking at this and couldn't find any
> > > sort of workaround. Presumably geni is taking some time between me
> > > reading the last word out of the FIFO from the "previous" packet and
> > > then transitioning to the new packet. I found a lot of references to
> > > this process in the hardware register description (see GENI_CFG_REG69,
> > > for instance), but I couldn't manage to make the kick to happen any
> > > faster. Presumably this isn't a problem for things like Bluetooth
> > > since flow control saves them. ...and I guess this isn't a problem in
> > > practice because we usually _send_ a lot of data to the host for
> > > console/kgdb and it's only the host => DUT path that has problems.
> > >
> > >
> > >> I am not sure about what all scenario can leads to this behavior from
> > >> hardware, I will try to get an answer from hardware team.
> > >>
> > >> Any error bit was set for SE_GENI_S_IRQ_STATUS & SE_GENI_M_IRQ_STATUS
> > >> registers?
> > > As per above I can see overflows in my test case and geni seems to be
> > > behaving pretty badly. If you have ideas on how to fix this I'd love
> > > it. However, it still seems like my patch is right because (at least
> > > in the cases I tested) it avoids dropping bytes in some cases. It
> > > also matches how qcom_geni_serial_handle_rx() works and if that was
> > > broken we'd have noticed by now.
> >
> > Reviewed-by: Akash Asthana <[email protected]>
>
> Thanks! After my extra digging I now think that the weird case is
> triggered by the overflows and the FIFO certainly does get garbage in
> it when there are overflows, but I think my patch is still correct and
> gets correct bytes. The garbage that is queued up is queued up later
> and looks like normal transfers. ...so I think we're good to go with
> it.
>
> So summary:
> 1. We should land my patch.
> 2. I filed <https://crbug.com/1115125> to track trying to make kdb's
> polling nicer.
This sort of thing could improve multi-line copy 'n paste (e.g. any case
where we issue more output than we receive in input) but be aware the pager
deliberately discards backlogged characters to improve latency so
in the use case that provoked this thread it makes no difference
whether the UART overflow machinery or the kdb core discards the backlog!
Daniel.