2012-06-20 19:28:13

by Jean-François Dagenais

[permalink] [raw]
Subject: Re: gpio: pca953x: interrupt feature unreliable

(putting Linus in CC 'cause I hear he enjoys interaction with hardware. As do I,
and this is a funny hard/soft timing issue, insignificant maybe in the large
scale of the kernel, but an interesting puzzle. Sorry if doing so is out-of-line.

It seems a few answers missed the lists, the interesting bits are quoted, and
should give a flavour of the discussion.)

On 2012-06-20, at 11:01 AM, David Jander wrote:
>
> First of all: As a hardware designer, if you intended to connect this chip's
> nINT output to a PCA953x input, I assume you know what you are doing, and not
> just trust the features a particular driver advertises.

Of course, I found out about this while "proof of concept"ing chaining the
AD714X INT to a PCA9555. The product this goes in is not in production yet.

> Second: What I try to explain as a possible workaround for your case, is to
> move the complete ISR callback functionality into a thread. But since I don't
> know the details of your driver, I don't know if that is possible.
I trie not to modify the kernel wherever I can. For the "purist" part of me,
hacking a chip driver here to compensate the flaws of another there is not the
best way to do it. I would prefer fixing a chip's shortcomings in it's own
driver.


>
>>> It still is an enormous advantage to have this functionality, because polling
>>> an I2C peripheral just to wait for input state can be a terrible waste of
>>> CPU-time ;-)
>>
>> I totally respect this. However, it may be OK for a student project to have a
>> flaky interrupt controller, but for industrial or medical applications, the
>> interrupt controller must be reliable.
>
> We use it in industrial applications, and we have no problems at all ;-)

If you examine the ASCII art of the timing I showed in the original message, you
will see that my particular timing is not the only problematic one... Can you
tell me how your 953x recovers from the state established at step 5? i.e. when
you clear the slave's INT through whatever mechanism, the 953x re-assert's it's
INT signal.

If I can attempt an answer: since the 953x has requested it's own ISR using
level-low, when the 953x's ISR thread is done, since the INT is still low, it
re-enters it's threaded ISR, and read it's input register again, which clears
953x's INT. (And this is key to the problem here, I hope you can see it, or I'm
mistaken here, but the ISR thread of the 953x runs a SECOND TIME) On this second
run, 953x's ISR examines the bits that changed since the last read, in this
case, two things can happen: (assuming your slave is trigger_falling) 1. Things
are quiet, the last read had your slave chip's INT low (and client's the nested
ISR was called (falling edge)), but this current read now says high. This is
considered like a rising and doesn't trigger calling the nested ISR again. All
is well then. 2. if for whatever reason (fast slave OR some delay scheduling
the 953x's ISR thread a second time) the second ISR run comes after the client's
chip re-asserting it's INT. This event makes the 953x's input the same as when
the input register was read on the first ISR thread run, so by definition, this
de-asserts the 953x's INT. When 953x finally reads the input register moments
later, there is no difference between old_stat and cur_stat and you are
essentially locked for this slave until you reset the whole thing.

Since the timings of the slave and the OS are factors here, it's quite hard to
reproduce. Even with the best intention and diligence at design, a race is
possible.


> Really, you must know what you are doing if you want to use an I2C I/O
> expander as interrupt chain controller!

After a good read of the specs and code, I had the idea to use our existing 9555
to separate the interrupts of two i2c slaves. I put this together on a bread
board and put the idea to the test. So I am in agreement that you must know what
you are doing when doing this. I had to know for sure before telling the
hardware guy to put this on real hardware.

>> For example, we are most likely going to
>> replace the 9555 in our next hardware rev because of this problem.
>
> Then most probably you started your design committing a mistake. You should
> just assume your hardware design is flawed and not try to blame some driver
> for it.

Again, bread board. Plus, I hope you see that since we will most likely not be
using this combination of i2c devices, I no longer have any personal interest of
my own in this, other than my wish to improve the kernel in general.


>
>> Remember though that I was proposing a fix instead of a complete removal (I
>> admit that suggesting removing the feature was a way to attract attention to the
>> problem ;) The only problem is that since we are most likely not keeping this
>> part in our design, putting effort into a fix of the pca953x.c driver is much
>> lower now in my priority list.
>
> Ok, then what about just leaving it as it is now?

If you are comfortable with the scenario I demonstrated originally, plus the one
I mention in this message, or you can explain to me (and others) why I am wrong,
then by all means keep the driver as it is! I just thought, first the world should
know about this, and second, get feedback before attempting modifications, if I
ever do get there.


Cheers!-


2012-06-21 07:36:12

by David Jander

[permalink] [raw]
Subject: Re: gpio: pca953x: interrupt feature unreliable


Dear Jean-François,

On Wed, 20 Jun 2012 15:28:05 -0400
Jean-François Dagenais <[email protected]> wrote:

> (putting Linus in CC 'cause I hear he enjoys interaction with hardware. As do I,
> and this is a funny hard/soft timing issue, insignificant maybe in the large
> scale of the kernel, but an interesting puzzle. Sorry if doing so is out-of-line.

Oh dear. Well, if you really need to do that.... ;-)
I hope we are not boring the hell out of a lot of people with this silly
discussion. My excuses to the rest of the CC if we are...

> It seems a few answers missed the lists, the interesting bits are quoted, and
> should give a flavour of the discussion.)
>
> On 2012-06-20, at 11:01 AM, David Jander wrote:
> >
> > First of all: As a hardware designer, if you intended to connect this chip's
> > nINT output to a PCA953x input, I assume you know what you are doing, and not
> > just trust the features a particular driver advertises.
>
> Of course, I found out about this while "proof of concept"ing chaining the
> AD714X INT to a PCA9555. The product this goes in is not in production yet.

Phew! Then what's the problem? You saved yourself right on time ;-)

> > Second: What I try to explain as a possible workaround for your case, is to
> > move the complete ISR callback functionality into a thread. But since I don't
> > know the details of your driver, I don't know if that is possible.
> I trie not to modify the kernel wherever I can. For the "purist" part of me,
> hacking a chip driver here to compensate the flaws of another there is not the
> best way to do it. I would prefer fixing a chip's shortcomings in it's own
> driver.

I agree, but I thought you had a major problem so I made a suggestion about how
you could work around it.... not to suggest how to write an acceptable driver.

> >>> It still is an enormous advantage to have this functionality, because polling
> >>> an I2C peripheral just to wait for input state can be a terrible waste of
> >>> CPU-time ;-)
> >>
> >> I totally respect this. However, it may be OK for a student project to have a
> >> flaky interrupt controller, but for industrial or medical applications, the
> >> interrupt controller must be reliable.
> >
> > We use it in industrial applications, and we have no problems at all ;-)
>
> If you examine the ASCII art of the timing I showed in the original message, you
> will see that my particular timing is not the only problematic one... Can you
> tell me how your 953x recovers from the state established at step 5? i.e. when
> you clear the slave's INT through whatever mechanism, the 953x re-assert's it's
> INT signal.

Easy: Just release the button and push it again ;-)
(I mentioned our application was a gpio_keys keyboard in my previous e-mail.)
That's the kind of applications you can use the PCA953x as interrupt
controller for. My patches incorporated support for device-tree bindings and a
few fixes for both gpio-pca953x.c and gpio_keys.c, so it became possible to do
just this in a device-tree:

i2c@1700 {
[...]
interrupts = <0x9 0x8>;
[...]
gpio0: gpio@74 {
compatible = "nxp,pca9539";
reg = <0x74>;
interrupts = <0x11 0x8>;
#gpio-cells = <2>;
gpio-controller;
}
}
[...]
gpio_keys {
compatible = "gpio-keys";
#address-cells = <1>;
#size-cells = <0>;
autorepeat;
button@20 {
label = "GPIO Key ESC";
linux,code = <1>;
gpios = <&gpio0 0 1>;
};
button@21 {
label = "GPIO Key UP";
linux,code = <103>;
gpios = <&gpio0 1 1>;
};
...

Get the idea? There is no other code involved, and it "just works"! I am
absolutely amazed by this marvel ;-)

> If I can attempt an answer: since the 953x has requested it's own ISR using
> level-low, when the 953x's ISR thread is done, since the INT is still low, it
> re-enters it's threaded ISR, and read it's input register again, which clears
> 953x's INT. (And this is key to the problem here, I hope you can see it, or I'm
> mistaken here, but the ISR thread of the 953x runs a SECOND TIME) On this second
> run, 953x's ISR examines the bits that changed since the last read, in this
> case, two things can happen: (assuming your slave is trigger_falling) 1. Things

I perfectly understand your reasoning. But it is not the case in our
application, since the user will not keep his finger on the key forever if the
key doesn't react ;-)
Also, in our case, pushing and releasing a button is such a slow process, the
the likelihood this will ever occur is negligible.

> are quiet, the last read had your slave chip's INT low (and client's the nested
> ISR was called (falling edge)), but this current read now says high. This is
> considered like a rising and doesn't trigger calling the nested ISR again. All
> is well then. 2. if for whatever reason (fast slave OR some delay scheduling
> the 953x's ISR thread a second time) the second ISR run comes after the client's
> chip re-asserting it's INT. This event makes the 953x's input the same as when
> the input register was read on the first ISR thread run, so by definition, this
> de-asserts the 953x's INT. When 953x finally reads the input register moments
> later, there is no difference between old_stat and cur_stat and you are
> essentially locked for this slave until you reset the whole thing.
>
> Since the timings of the slave and the OS are factors here, it's quite hard to
> reproduce. Even with the best intention and diligence at design, a race is
> possible.

Yes, of course. But that is a shortcoming of the extremely simple design of
PCA953x gpio expanders. There is not much we can do about this in the driver.
Maybe we could place a warning in the documentation about the shortcomings of
these chips...?

> > Really, you must know what you are doing if you want to use an I2C I/O
> > expander as interrupt chain controller!
>
> After a good read of the specs and code, I had the idea to use our existing 9555
> to separate the interrupts of two i2c slaves. I put this together on a bread
> board and put the idea to the test. So I am in agreement that you must know what
> you are doing when doing this. I had to know for sure before telling the
> hardware guy to put this on real hardware.

Never mind. I am a hardware designer myself and we all make such mistakes.
Just try not to blame someone or something else for it publicly ;-)

> >> For example, we are most likely going to
> >> replace the 9555 in our next hardware rev because of this problem.
> >
> > Then most probably you started your design committing a mistake. You should
> > just assume your hardware design is flawed and not try to blame some driver
> > for it.
>
> Again, bread board. Plus, I hope you see that since we will most likely not be
> using this combination of i2c devices, I no longer have any personal interest of
> my own in this, other than my wish to improve the kernel in general.

Feel free to improve the driver... but please do not remove features that work
for others.

> >> Remember though that I was proposing a fix instead of a complete removal (I
> >> admit that suggesting removing the feature was a way to attract attention to the
> >> problem ;) The only problem is that since we are most likely not keeping this
> >> part in our design, putting effort into a fix of the pca953x.c driver is much
> >> lower now in my priority list.
> >
> > Ok, then what about just leaving it as it is now?
>
> If you are comfortable with the scenario I demonstrated originally, plus the one
> I mention in this message, or you can explain to me (and others) why I am wrong,
> then by all means keep the driver as it is! I just thought, first the world should
> know about this, and second, get feedback before attempting modifications, if I
> ever do get there.

I hope my explanation was clear enough this time. As I said, we may put some
warnings in the documentation of the driver.

Best regards,

--
David Jander

2012-06-21 19:34:12

by Jean-François Dagenais

[permalink] [raw]
Subject: Re: gpio: pca953x: interrupt feature unreliable

Dear David!

On 2012-06-21, at 3:10 AM, David Jander wrote:

> On Wed, 20 Jun 2012 15:28:05 -0400
> Jean-Fran?ois Dagenais <[email protected]> wrote:
>
>> (putting Linus in CC 'cause I hear he enjoys interaction with hardware. As do I,
>> and this is a funny hard/soft timing issue, insignificant maybe in the large
>> scale of the kernel, but an interesting puzzle. Sorry if doing so is out-of-line.
>
> Oh dear. Well, if you really need to do that.... ;-)
> I hope we are not boring the hell out of a lot of people with this silly
> discussion. My excuses to the rest of the CC if we are...

I personally think the "silliness" of the discussion is still in debate. It may be silly in the grand context of the kernel, I agree, but definitely core for pca953x users. I merely involved torvalds for the kick of tricky timing issues and register specs, not to attract a parade on this. Removed on this reply, he can follow if he's interested on the list. Sorry if any etiquette was violated.

>
>> Of course, I found out about this while "proof of concept"ing chaining the
>> AD714X INT to a PCA9555. The product this goes in is not in production yet.
>
> Phew! Then what's the problem?

Uuuh, I'm good myself thanks ;) but the problem is this is not a reliable interrupt controller (in the general sense), especially the way the current driver works with it.

> ...
> Easy: Just release the button and push it again ;-)
> (I mentioned our application was a gpio_keys keyboard in my previous e-mail.)

Yeah I had missed that you were doing one key per GPIO... But that's indeed why you are fine with it. You are using it more like a keyboard driver with an interrupt controller. If you use the 953x as an actual interrupt "controller" for a client that has a hardware-logic controlled INT signal, it won't be reliable for such an application (with the current driver state), contrary to what the driver initially suggests.

> That's the kind of applications you can use the PCA953x as interrupt
> controller for. My patches incorporated support for device-tree bindings and a
> few fixes for both gpio-pca953x.c and gpio_keys.c, so it became possible to do
> just this in a device-tree:
> ...
>
> Get the idea? There is no other code involved, and it "just works"! I am
> absolutely amazed by this marvel ;-)

I wasn't familiar with device-tree, thanks for this intro, I will look into it more. Not being familiar, I couldn't make this part out properly: are each of your keys using their own nested IRQ?

> ...
>> Since the timings of the slave and the OS are factors here, it's quite hard to
>> reproduce. Even with the best intention and diligence at design, a race is
>> possible.
>
> Yes, of course. But that is a shortcoming of the extremely simple design of
> PCA953x gpio expanders. There is not much we can do about this in the driver.

Well, what about my original suggestion? I know you said you agree with serving on only LEVEL, but it had a few other specifics... I guess at this point, my idea would have to be backed by tests and a patch. (Hearing you tell me: Yeah JF, why don't you put your money where your mouth is!? ;)

> Maybe we could place a warning in the documentation about the shortcomings of
> these chips...?

Ok, now we are starting to agree! ;-) Short of an actual mod patch to the driver with results, I think this is a minimum. It could save a lot of people a lot of time, and perhaps save some aggravation too.

> Never mind. I am a hardware designer myself and we all make such mistakes.
> Just try not to blame someone or something else for it publicly ;-)

I'm sorry if you felt my interventions here had anything to do with blame, to you or anyone. I was merely acting on the behalf of helping others avoid encountering issues with this, and saving other engineers some time from my own experiences. And in the spirit of giving back since I take so much from the open source community. I know it sounds corny, but it's still true.

>
> Feel free to improve the driver... but please do not remove features that work
> for others.

Indeed. Hence why I try to get feedback about my improvement idea. It has effects on current users since the interrupt clients would have to adjust since instead of providing EDGE based, it would provide only LEVEL, plus, in existing designs, the pca953x would now ask that it's own interrupt be EDGE(RISING|FALLING) instead of LEVEL_LOW. These changes are important for my theory to work correctly. The only remaining trap at this point would be in the chip, which says it cannot catch edges that happen too quickly, but the minimum length of the edges it detects is sufficiently small not to be of anyone's concern.

> I hope my explanation was clear enough this time. As I said, we may put some
> warnings in the documentation of the driver.
>
> Best regards,
>
> --
> David Jander

Thanks for your participation and work on this driver! ;)
/jfd

2012-06-22 07:53:03

by David Jander

[permalink] [raw]
Subject: Re: gpio: pca953x: interrupt feature unreliable


Dear Jean-François,

On Thu, 21 Jun 2012 15:34:07 -0400
Jean-François Dagenais <[email protected]> wrote:
> Uuuh, I'm good myself thanks ;) but the problem is this is not a reliable
> interrupt controller (in the general sense), especially the way the current
> driver works with it.

I still think that the real problem is with the chip itself. The driver should
just do it's best, and if the problem can't be worked around in the driver,
the driver should stay neutral IMHO.
In theory the PCA953x is still an edge triggered interrupt controller, but its
own interrupt output is level.

> > ...
> > Easy: Just release the button and push it again ;-)
> > (I mentioned our application was a gpio_keys keyboard in my previous e-mail.)
>
> Yeah I had missed that you were doing one key per GPIO... But that's indeed why
> you are fine with it. You are using it more like a keyboard driver with an
> interrupt controller. If you use the 953x as an actual interrupt "controller"
> for a client that has a hardware-logic controlled INT signal, it won't be
> reliable for such an application (with the current driver state), contrary to
> what the driver initially suggests.

I insist. The driver can't really do much about it.

> > That's the kind of applications you can use the PCA953x as interrupt
> > controller for. My patches incorporated support for device-tree bindings and a
> > few fixes for both gpio-pca953x.c and gpio_keys.c, so it became possible to do
> > just this in a device-tree:
> > ...
> >
> > Get the idea? There is no other code involved, and it "just works"! I am
> > absolutely amazed by this marvel ;-)
>
> I wasn't familiar with device-tree, thanks for this intro, I will look into

Be prepared, a device-tree is coming to a SoC near you any day... unless you
are stuck in the x86 world, where such marvel is not likely to appear anytime
soon ;-)

> it more. Not being familiar, I couldn't make this part out properly: are each
> of your keys using their own nested IRQ?

Yes!
Unfortunately gpio_keys.c requests edge-based IRQ's, so if you eliminate edge
functionality from the pca953x driver, my setup will stop working.
And the pca953x really _is_ an edge based interrupt controller (if one really
uses it as such), although a fairly limited (broken if you must) one.

> >> Since the timings of the slave and the OS are factors here, it's quite hard to
> >> reproduce. Even with the best intention and diligence at design, a race is
> >> possible.
> >
> > Yes, of course. But that is a shortcoming of the extremely simple design of
> > PCA953x gpio expanders. There is not much we can do about this in the driver.
>
> Well, what about my original suggestion? I know you said you agree with serving
> on only LEVEL, but it had a few other specifics... I guess at this point, my

Sorry. I was wrong. It is not a LEVEL interrupt controller. It doesn't work as
such. Besides, as I said above, this would break my setup, although the
decice-tree where we use this functionality is not mainlined, so I am probably
not allowed to complain about that :-(

> idea would have to be backed by tests and a patch. (Hearing you tell me: Yeah
> JF, why don't you put your money where your mouth is!? ;)

I wasn't about to say that.

> > Maybe we could place a warning in the documentation about the shortcomings of
> > these chips...?
>
> Ok, now we are starting to agree! ;-) Short of an actual mod patch to the
> driver with results, I think this is a minimum. It could save a lot of people
> a lot of time, and perhaps save some aggravation too.
>
> > Never mind. I am a hardware designer myself and we all make such mistakes.
> > Just try not to blame someone or something else for it publicly ;-)
>
> I'm sorry if you felt my interventions here had anything to do with blame,
> to you or anyone. I was merely acting on the behalf of helping others avoid

Well, to me it looked like you wanted to blame the pca953x driver for your
problem with the chip....

> encountering issues with this, and saving other engineers some time from my
> own experiences. And in the spirit of giving back since I take so much from
> the open source community. I know it sounds corny, but it's still true.
>
> >
> > Feel free to improve the driver... but please do not remove features that work
> > for others.
>
> Indeed. Hence why I try to get feedback about my improvement idea. It has
> effects on current users since the interrupt clients would have to adjust
> since instead of providing EDGE based, it would provide only LEVEL, plus, in
> existing designs, the pca953x would now ask that it's own interrupt be
> EDGE(RISING|FALLING) instead of LEVEL_LOW. These changes are important for my
> theory to work correctly. The only remaining trap at this point would be in
> the chip, which says it cannot catch edges that happen too quickly, but the
> minimum length of the edges it detects is sufficiently small not to be of
> anyone's concern.

You are proposing to change the PCA953x's input sensitivity artificially to
LEVEL, and its output sensitivity to EDGE(RISING|FALLING)?
Weird. Quickly sketching a scenario where you have just one level-based
acknowledged interrupt on a PCA953x input, it would generate no less than 4
interrupts to the primary controller (2 for each edge of the input signal)!!
I don't think this will be accepted.... besides it will break my setup and be
pretty much counter-intuitive to anyone else trying to use this chip as
advertised (talking about pitfalls).
And it will generate quite a significant amount of downstream interrupts,
triggering a lot if I2C traffic.

> > I hope my explanation was clear enough this time. As I said, we may put some
> > warnings in the documentation of the driver.

I still think that would be the best option... specially since you are not
going to use your proposed setup anyway.

> Thanks for your participation and work on this driver! ;)

You are welcome. Thanks for the discussion.

Best regards,

--
David Jander
Protonic Holland.