Re: [v3, 3/4] i2c: i2c-qcom-geni: Add bus driver for the Qualcomm GENI I2C controller

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

On Wed, Mar 7, 2018 at 6:42 PM, Sagar Dharia <sdharia@xxxxxxxxxxxxxx> wrote:
> Hi Doug,
> Thank you for reviewing the patch. I will take a stab at a few comments
> below. We will address most of the other comments in next version of I2C
> patch.
>>
>>
>>
>>> +
>>> +#define I2C_AUTO_SUSPEND_DELAY 250
>>
>>
>> Why 250 ms?  That seems like an eternity.  Is it really that expensive
>> to turn resources off and on?  I would sorta just expect clocks and
>> stuff to get turned off right after a transaction finished unless
>> another one was pending right behind it...
>>
> The response from RPMh to turn on/off shared resources also take quite a few
> msecs. The QUP serial bus block sits quite a few shared-NOCs away from the
> memory and runtime-PM is used a bandwidth vote/NOC vote for these NOCs from
> QUP to memory. Also the RPC between apps and RPMh can sometimes take longer
> depending on other tasks running on apps. This 250 msec avoids thrashing of
> these RPCs between apps and RPMh.
> If you plan to keep these NOCs on forever, then your are right: runtime-PM
> will be only used to turn on/off local clocks and we won't even need
> autosuspend. that's not true on products where this driver is currently
> deployed.

OK, fair enough.  I don't know how RPMh works well enough to argue.
It does seem odd that you'd want to design things such that it takes a
few msecs to pull it out of runtime suspend, especially for touch.


>>> +
>>> +static const struct geni_i2c_clk_fld geni_i2c_clk_map[] = {
>>> +       {KHz(100), 7, 10, 11, 26},
>>> +       {KHz(400), 2,  5, 12, 24},
>>> +       {KHz(1000), 1, 3,  9, 18},
>>
>>
>> So I guess this is all relying on an input serial clock of 19.2MHz?
>> Maybe document that?
>>
>> Assuming I'm understanding the math here, is it really OK for your
>> 100kHz and 1MHz mode to be running slightly fast?
>>
>> 19200. / 2 / 24
>>>>>
>>>>> 400.0
>>
>>
>> 19200. / 7 / 26
>>>>>
>>>>> 105.49450549450549
>>
>>
>> 19200. / 1 / 18
>>>>>
>>>>> 1066.6666666666667
>>
>>
>> It seems like you'd want the fastest clock that you can make that's
>> _less than_ the spec.
>>
>>
>> It would also be interesting to know if it's expected that boards
>> might need to tweak the t_high / t_low depending on their electrical
>> characteristics.  In the past I've had lots of requests from board
>> makers to tweak things because they've got a long trace, or a stronger
>> or weaker pull, or ...  If so we might later need to add some dts
>> properties like "i2c-scl-rising-time-ns" and make the math more
>> dynamic here, unless your hardware somehow automatically adjusts for
>> this type of thing...
>> These values are derived by our HW team to comply with the t-high and
>
> t-low specs of I2C. We have confirmed on scope that the frequency of SCL is
> indeed less than/equal to the spec. We have not come across slaves who have
> needed to tweak these things. We are open to adding these properties in dts
> if you have any such slaves not conforming due to board-layout of other
> reasons.

OK, I'm fine with leaving something like this till later if/when it
comes up.  Documenting a little bit more about how these timings work
seems like it would be nice, though, at least mentioning what the
source clock is...


>>>
>>> +       mode = msg->len > 32 ? GENI_SE_DMA : GENI_SE_FIFO;
>>
>>
>> DMA is hard and i2c transfers > 32 bytes are rare.  Do we really gain
>> a lot by transferring i2c commands over DMA compared to a FIFO?
>> Enough to justify the code complexity and the set of bugs that will
>> show up?  I'm sure it will be a controversial assertion given that the
>> code's already written, but personally I'd be supportive of ripping
>> DMA mode out to simplify the driver.  I'd be curious if anyone else
>> agrees.  To me it seems like premature optimization.
>
>
> Yes, we have multiple clients (e.g. touch, NFC) using I2C for data transfers
> bigger than 32 bytes (some transfers are 100s of bytes). The fifo size is
> 32, so we can definitely avoid at least 1 interrupt when DMA mode is used
> with data size > 32.

Does that 1-2 interrupts make any real difference, though?  In theory
it really shouldn't affect the transfer rate.  We should be able to
service the interrupt plenty fast and if we were concerned we would
tweak the watermark code a little bit.  So I guess we're worried about
the extra CPU cycles (and power cost) to service those extra couple
interrupts?

In theory when touch data is coming in or NFC data is coming in then
we're probably not in a super low power state to begin with.  If it's
touch data we likely want to have the CPU boosted a bunch to respond
to the user quickly.  If we've got 8 cores available all of which can
run at 1GHz or more a few interrupts won't kill us.  NFC data is
probably not common enough that we need to optimize power/CPU
utilizatoin for that.


So while i can believe that you do save an interrupt or two, I still
am not convinced that those interrupts are worth a bunch of complex
code (and a whole second code path) to save.


...also note that above you said that coming out of runtime suspend
can take several msec.  That seems like it dwarfs any slight
difference in timing between a FIFO-based operation and DMA.


>>> +       geni_se_select_mode(&gi2c->se, mode);
>>> +       writel_relaxed(msg->len, gi2c->se.base + SE_I2C_RX_TRANS_LEN);
>>> +       geni_se_setup_m_cmd(&gi2c->se, I2C_READ, m_param);
>>> +       if (mode == GENI_SE_DMA) {
>>> +               rx_dma = geni_se_rx_dma_prep(&gi2c->se, msg->buf,
>>> msg->len);
>>
>>
>> Randomly I noticed a flag called "I2C_M_DMA_SAFE".  Do we need to
>> check this flag before using msg->buf for DMA?  ...or use
>> i2c_get_dma_safe_msg_buf()?
>>
>> ...btw: the relative lack of people doing this in the kernel is
>> further evidence of DMA not really being worth it for i2c busses.
>
> I cannot comment about other drivers here using or not using DMA since they
> may not be exercised with slaves like NFC?
>>>
>>> +       ret = pm_runtime_get_sync(gi2c->se.dev);
>>> +       if (ret < 0) {
>>> +               dev_err(gi2c->se.dev, "error turning SE resources:%d\n",
>>> ret);
>>> +               pm_runtime_put_noidle(gi2c->se.dev);
>>> +               /* Set device in suspended since resume failed */
>>> +               pm_runtime_set_suspended(gi2c->se.dev);
>>> +               return ret;
>>
>>
>> Wow, that's a cluster of arcane calls to handle a call that probably
>> will never fail (it just enables clocks and sets pinctrl).  Sigh.
>> ...but as far as I can tell the whole sequence is right.  You
>> definitely need a "put" after a failed get and it looks like
>> pm_runtime_set_suspended() has a special exception where it can be
>> called if you got a runtime error...
>
> We didn't have this in before either. But this condition is somewhat
> frequent if I2C transactions are called on cusp of exiting system suspend.
> (e.g. PMIC slave getting a wakeup-IRQ and trying to read from PMIC through
> I2C to read its status as to what caused that wake-up. At that time,
> get_sync doesn't really enable resources (kernel 4.9) since the runtime-pm
> ref-count isn't decremented. We run the risk of unclocked access if these
> arcane calls aren't present. You can go through runtime-pm documentation
> chapter 6 for more details.

Yeah, I certainly agree that the calls are needed if
pm_runtime_get_sync() and I'm not suggesting removing them.  Hence the
"as far as I can tell the whole sequence is right".

...but I'm actually kinda worried if you're saying that you actually
ran into this case.  Hopefully that got fixed and code no longer tries
to read from the PMIC at a bad time anymore?  That code should be
fixed not to be running so late in suspend.


>>> +       ret = devm_request_irq(&pdev->dev, gi2c->irq, geni_i2c_irq,
>>> +                              IRQF_TRIGGER_HIGH, "i2c_geni", gi2c);
>>> +       if (ret) {
>>> +               dev_err(&pdev->dev, "Request_irq failed:%d: err:%d\n",
>>> +                       gi2c->irq, ret);
>>> +               return ret;
>>> +       }
>>> +       disable_irq(gi2c->irq);
>>
>>
>> Can you explain the goal of the disable_irq() here.  Is it actually
>> needed for something or does it somehow save power?  From drivers I've
>> reviewed in the past this doesn't seem like a common thing to do, so
>> I'm curious what it's supposed to gain for you.  I'd be inclined to
>> just delete the whole disable/enable of the irq from this driver.
>
>
> Qualcomm's power team suggests we enable/disable unused IRQs. Otherwise they
> can block apps from entering some low-power mode (unless the interrupt is in
> some list?) I will confirm again with them and let you know.

Since this is weird (to me anyway), please document w/ a comment.


>>>
>>> +       /* Make sure no transactions are pending */
>>> +       ret = i2c_trylock_bus(&gi2c->adap, I2C_LOCK_SEGMENT);
>>> +       if (!ret) {
>>> +               dev_err(gi2c->se.dev, "late I2C transaction request\n");
>>> +               return -EBUSY;
>>> +       }
>>
>>
>> Does this happen?  How?
>>
>> Nothing about this code looks special for your hardware.  If this is
>> really needed, why is it not part of the i2c core since there's
>> nothing specific about your driver here?
>>
> There have been some clients that don't implement sys-suspend/resume
> callbacks (so i2c adapter has no clue they are done with their transactions)
> and this allows us to be flexible when they call I2C transactions extremely
> late.

Still feels like this belongs in the i2c core, not your driver.  Maybe
you should send a patch for the core and remove it from here?

...and also, it seems like any i2c clients that don't implement the
suspend/resume callbacks and try to do i2c transactions late in the
game need to be fixed.  It should be documented that this isn't a
valid thing for a driver to do and if we end up in this error case
then it's not an i2c issue but it's a bad driver somewhere.

>
>>
>>> +       if (!pm_runtime_status_suspended(device)) {
>>> +               geni_i2c_runtime_suspend(device);
>>> +               pm_runtime_disable(device);
>>> +               pm_runtime_set_suspended(device);
>>> +               pm_runtime_enable(device);
>>> +       }
>>
>>
>> Similar question.  Why do you need this special case code?  Are there
>> cases where we're all the way at suspend_noirq and yet we still
>> haven't runtime suspended?  Can you please document how we get into
>> this state?
>> This is when transaction happens less-than 250 msec of the
>
> system-suspend. PM-runtime has not gotten a chance to auto-suspend us since
> timer hasn't expired before system-suspend is attempted. These calls make
> sure that we truly turn off driver resources and make runtime-PM state
> consistent with the HW state. We can document this.

OK.  PM Runtime always gets me mixed up.  Seems really strange that it
wouldn't autosuspend all devices (regardless of timeout) at system
suspend time.

-Doug



[Index of Archives]     [Linux GPIO]     [Linux SPI]     [Linux Hardward Monitoring]     [LM Sensors]     [Linux USB Devel]     [Linux Media]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux