Boot hang regression 3.10.0-rc4 -> 3.10.0

[Rajendra Nayak]

On Monday 08 July 2013 06:50 PM, Rajendra Nayak wrote:
> On Monday 08 July 2013 06:40 PM, Tony Lindgren wrote:
>> * Rajendra Nayak <rnayak at ti.com> [130708 05:48]:
>>> On Monday 08 July 2013 04:55 PM, Tony Lindgren wrote:
>>>> * Bedia, Vaibhav <vaibhav.bedia at ti.com> [130705 06:37]:
>>>>> On Fri, Jul 05, 2013 at 18:50:10, Bedia, Vaibhav wrote:
>>>>>> Hi Tony,
>>>>>>
>>>>>> On Fri, Jul 05, 2013 at 17:29:59, Tony Lindgren wrote:
>>>>>>> * Bedia, Vaibhav <vaibhav.bedia at ti.com> [130705 01:17]:
>>>>>>>>
>>>>>>>> I just checked the behavior on my AM335x-EVM. Current mainline boots fine
>>>>>>>> provided I don't use earlyprintk.  The offending patch [1] in this case is the one
>>>>>>>> that tries to get rid of omap_serial_early_init() for DT boot. This change inadvertently
>>>>>>>> also results in the console UART getting reset and idled during bootup and that's where
>>>>>>>> the boot stops for you. I think if you skip earlyprintk from the bootargs you should see
>>>>>>>> the system booting fine. 
>>>>>>>>
>>>>>>>> I guess we need to retain the NO_IDLE and NO_RESET aspect for the console UART in
>>>>>>>> omap_serial_early_init() to get earlyprintk working again.
>>>>>>>
>>>>>>> Hmm nothing should get idled while earlyprintk is running, and then when the
>>>>>>> serial driver kicks in it should not idle anything by default. And for DT based
>>>>>>> booting we should not have mach-omap2/serial.c initialize anything.
>>>>>>>
>>>>>>
>>>>>> If I add in the HWMOD flags without any reverts I get to the point where the serial driver
>>>>>> comes up but the boot eventually stops [1]. Without the flags the boot stops much earlier [2]
>>>>>> just like Mark reported.
>>>>>>
>>>>>
>>>>> Err.. the log with HWMOD flags added is [2] and without flags is [1]. Sorry for the confusion.
>>>>
>>>> It sounds like something needs to be fixed for am33xx as omap3 and omap4
>>>> won't hang with earlyprintk. Almost certainly mach-omap2/serial.c should not
>>>> be needed at all for am33xx, and the bug is somewhere else.
>>>
>>> Tony, I spent some time on this today and there seem to 2 issues.
>>>
>>> Issue 1: Causing boot to stop much earlier as reported (this is during hmwod setup)
>>>
>>> The commit 'e97f03cb36e9ec8a2ccaa3e4bee5297fe48156fd' 
>>> "ARM: OMAP2+: Fix serial init for device tree based booting" stubbed out omap_serial_early_init()
>>> for DT case thinking its doing the port inits. But that does not seem to be true, the port inits happen
>>> as part of omap_serial_init_port(). What omap_serial_early_init() was doing instead was adding the
>>> HWMOD_INIT_NO_IDLE and HWMOD_INIT_NO_RESET flags which would tell hmwod not to reset and then idle the
>>> console UART. With this not happening now for the DT case, it causes an issue.
>>>
>>> The issue was seen on am33xx and not on some other platforms because some platforms still have these
>>> statically defined in the hwmod data files. I could see these set for uart3 in case of omap4 and omap5.
>>> So I feel the above commit should be reverted and these static flags should be removed from the data
>>> files.
>>
>> Oh OK. That's starting to make a bit more sense then. 
>>   
>>>>>>> I wonder if this is because the timeouts get now initialized to 0 instead
>>>>>>> of -1 for the serial driver?
>>>>>>>
>>>>>>
>>>>>> You meant initialized to -1, right? There's an additional check for timeout being 0. Unless i
>>>>>> am missing something DT-boot will start off with timeout set to 0 and then get forced to -1.
>>>>
>>>> OK
>>>
>>> Issue 2: Causing boot to stop when serial driver is initialized. (After Issue 1 is fixed)
>>>
>>> I could narrow this down to the change done to return -EINVAL instead of 0 in serial_omap_get_context_loss_count()
>>> as part of commit 'a630fbfbb1beeffc5bbe542a7986bf2068874633' "serial: omap: Fix device tree based PM runtime"
>>>
>>> What this change in turn seems to do is cause a serial_omap_restore_context() to get called as part of
>>> serial_omap_runtime_resume() which was not the case when serial_omap_get_context_loss_count() returned 0
>>>
>>> from serial_omap_runtime_resume():
>>> -----
>>>         int loss_cnt = serial_omap_get_context_loss_count(up);
>>>
>>>         if (loss_cnt < 0) {
>>>                 dev_dbg(dev, "serial_omap_get_context_loss_count failed : %d\n",
>>>                         loss_cnt);
>>>                 serial_omap_restore_context(up);
>>>         } else if (up->context_loss_cnt != loss_cnt) {
>>>                 serial_omap_restore_context(up);
>>>         }
>>> -----
>>>
>>> I am still working on why a serial_omap_restore_context() could have caused console to die. I will work with
>>> Sourav on this and post the fixes for both issue 1 and issue2 once its clear on whats really causing issue 2.
>>
>> That's because we don't have the omap specific pdata callbacks for
>> context loss any longer. We may be able to detect when the context
>> was really lost in the serial driver, and only then call the
>> serial_omap_restore_context().
> 
> Right, but calling serial_omap_restore_context() even when the context is not lost, should not
> ideally cause an issue.

So looking some more into this, it turns out a lot of what serial_omap_restore_context() restores
gets set as part of serial_omap_set_termios() which infact gets called much later and any call to
serial_omap_restore_context before a call to serial_omap_set_termios() ends up restoring junk?

> 
>>  
>>> Let me know if the fix I listed for Issue 1: makes sense.
>>
>> Yes makes sense as a fix, but IMHO we should not need any workarounds
>> like that. Is the hwmod code idling the the uarts early? If so, then
>> it should only do that in a late_initcall if no drivers are registered.
> 
> hwmod as part of its setup (early) enables/resets and idles all modules.
> These flags are used to tell hwmod to avoid a reset and idle and leave the
> module enabled (in this case console uart)
> 
> regards
> Rajendra
> 
>>
>> Regards,
>>
>> Tony
>>
>