pciehp command complete timeout issue

pciehp command complete timeout issue

[Yijing Wang]

When I tried to unbind pciehp driver on a pcie root port(bound pciehp driver),
a lot timeout warning appeared.

The first timeout value is 102387672 msec :(
I debug and found that when pciehp complete pcie_enable_notification(), there was no command complete interrupt
be triggered, so cmd_busy always be set, and once another command post, a very long timeout warning noised.


 +-[0000:40]-+-00.0-[41]--
 |           +-01.0-[42-43]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
 |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
 |           +-03.0-[44-45]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
 |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection

[root@hulk slots]# ls
0  0-1  0-2  0-3  0-4  0-5  1  10  11  12  13  14  15  16  2  3  4  5  6  7  8  9
[root@hulk slots]# cat 6/address
0000:44:00
[root@hulk slots]#

[root@hulk pciehp]# echo 0000:40:03.0:pcie04 > unbind
[root@hulk pciehp]#

...
[102413.749632] pciehp 0000:40:03.0:pcie04: unloading service driver pciehp
[102413.749638] pciehp_remove dev 0000:40:03.0, cmd_busy 1
[102413.754929] pcie_disable_notification: ctrl cmd busy 1
[102413.765903] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x11f1 (issued 102387672 msec ago)
[102413.775171] pcie_do_write_cmd: dev 0000:40:03.0, cmd_busy set to 1
[102415.377950] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x01c0 (issued 1600 msec ago)
...

-- 
Thanks!
Yijing

Re: pciehp command complete timeout issue

[Bjorn Helgaas]

On Thu, Jun 18, 2015 at 08:06:23PM +0800, Yijing Wang wrote:
> When I tried to unbind pciehp driver on a pcie root port(bound pciehp driver),
> a lot timeout warning appeared.
> 
> The first timeout value is 102387672 msec :(
> I debug and found that when pciehp complete pcie_enable_notification(), there was no command complete interrupt
> be triggered, so cmd_busy always be set, and once another command post, a very long timeout warning noised.
> 
> 
>  +-[0000:40]-+-00.0-[41]--
>  |           +-01.0-[42-43]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>  |           +-03.0-[44-45]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
> 
> [root@hulk slots]# ls
> 0  0-1  0-2  0-3  0-4  0-5  1  10  11  12  13  14  15  16  2  3  4  5  6  7  8  9
> [root@hulk slots]# cat 6/address
> 0000:44:00
> [root@hulk slots]#
> 
> [root@hulk pciehp]# echo 0000:40:03.0:pcie04 > unbind
> [root@hulk pciehp]#
> 
> ...
> [102413.749632] pciehp 0000:40:03.0:pcie04: unloading service driver pciehp
> [102413.749638] pciehp_remove dev 0000:40:03.0, cmd_busy 1
> [102413.754929] pcie_disable_notification: ctrl cmd busy 1
> [102413.765903] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x11f1 (issued 102387672 msec ago)

The fact that you got this timeout message means the controller did
not set the "No Command Completed Support" bit, right?  If we had
NO_CMD_CMPL(ctrl), pcie_wait_cmd() becomes a no-op, and we would
never print any timeout message.

Since the "No Command Completed Support" bit is NOT set, we expect
to get an interrupt after every command completes.

This sounds like the Intel CF118 erratum mentioned just above that timeout
message:

         * Controllers with errata like Intel CF118 don't generate
         * completion notifications unless the power/indicator/interlock
         * control bits are changed.  On such controllers, we'll emit this
         * timeout message when we wait for completion of commands that
         * don't change those bits, e.g., commands that merely enable
         * interrupts.

So to me, this sounds like pciehp is working correctly.  What did you
expect to happen instead?

> [102413.775171] pcie_do_write_cmd: dev 0000:40:03.0, cmd_busy set to 1
> [102415.377950] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x01c0 (issued 1600 msec ago)
> ...
> 
> -- 
> Thanks!
> Yijing
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: pciehp command complete timeout issue

[Yijing Wang]

On 2015/6/25 7:05, Bjorn Helgaas wrote:
> On Thu, Jun 18, 2015 at 08:06:23PM +0800, Yijing Wang wrote:
>> When I tried to unbind pciehp driver on a pcie root port(bound pciehp driver),
>> a lot timeout warning appeared.
>>
>> The first timeout value is 102387672 msec :(
>> I debug and found that when pciehp complete pcie_enable_notification(), there was no command complete interrupt
>> be triggered, so cmd_busy always be set, and once another command post, a very long timeout warning noised.
>>
>>
>>  +-[0000:40]-+-00.0-[41]--
>>  |           +-01.0-[42-43]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>  |           +-03.0-[44-45]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>
>> [root@hulk slots]# ls
>> 0  0-1  0-2  0-3  0-4  0-5  1  10  11  12  13  14  15  16  2  3  4  5  6  7  8  9
>> [root@hulk slots]# cat 6/address
>> 0000:44:00
>> [root@hulk slots]#
>>
>> [root@hulk pciehp]# echo 0000:40:03.0:pcie04 > unbind
>> [root@hulk pciehp]#
>>
>> ...
>> [102413.749632] pciehp 0000:40:03.0:pcie04: unloading service driver pciehp
>> [102413.749638] pciehp_remove dev 0000:40:03.0, cmd_busy 1
>> [102413.754929] pcie_disable_notification: ctrl cmd busy 1
>> [102413.765903] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x11f1 (issued 102387672 msec ago)
> 
> The fact that you got this timeout message means the controller did
> not set the "No Command Completed Support" bit, right?  If we had
> NO_CMD_CMPL(ctrl), pcie_wait_cmd() becomes a no-op, and we would
> never print any timeout message.
> 
> Since the "No Command Completed Support" bit is NOT set, we expect
> to get an interrupt after every command completes.
> 
> This sounds like the Intel CF118 erratum mentioned just above that timeout
> message:
> 
>          * Controllers with errata like Intel CF118 don't generate
>          * completion notifications unless the power/indicator/interlock
>          * control bits are changed.  On such controllers, we'll emit this
>          * timeout message when we wait for completion of commands that
>          * don't change those bits, e.g., commands that merely enable
>          * interrupts.
> 
> So to me, this sounds like pciehp is working correctly.  What did you
> expect to happen instead?

I think if we could warn the timeout messages when the timeout is reached, not be detected in
next command write, it would be better.

Something like:

Write A command
trigger a timeout delay work event
interrupt coming (clean the cmd_busy, cancel the timeout delay work event)
timeout delay event work (detect whether the cmd_busy is still set, if yes, warn the timeout message)
..

But this is just my personal 3 seconds idea, it may make code more complex, I am not sure it's worth doing.

Thanks!
Yijing.


> 
>> [102413.775171] pcie_do_write_cmd: dev 0000:40:03.0, cmd_busy set to 1
>> [102415.377950] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x01c0 (issued 1600 msec ago)
>> ...
>>
>> -- 
>> Thanks!
>> Yijing
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-pci" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 
> 


-- 
Thanks!
Yijing

Re: pciehp command complete timeout issue

[Bjorn Helgaas]

On Wed, Jun 24, 2015 at 8:18 PM, Yijing Wang <wangyijing@huawei.com> wrote:
> On 2015/6/25 7:05, Bjorn Helgaas wrote:
>> On Thu, Jun 18, 2015 at 08:06:23PM +0800, Yijing Wang wrote:
>>> When I tried to unbind pciehp driver on a pcie root port(bound pciehp driver),
>>> a lot timeout warning appeared.
>>>
>>> The first timeout value is 102387672 msec :(
>>> I debug and found that when pciehp complete pcie_enable_notification(), there was no command complete interrupt
>>> be triggered, so cmd_busy always be set, and once another command post, a very long timeout warning noised.
>>>
>>>
>>>  +-[0000:40]-+-00.0-[41]--
>>>  |           +-01.0-[42-43]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>>  |           +-03.0-[44-45]--+-00.0  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>>  |           |               \-00.1  Intel Corporation 82599ES 10-Gigabit SFI/SFP+ Network Connection
>>>
>>> [root@hulk slots]# ls
>>> 0  0-1  0-2  0-3  0-4  0-5  1  10  11  12  13  14  15  16  2  3  4  5  6  7  8  9
>>> [root@hulk slots]# cat 6/address
>>> 0000:44:00
>>> [root@hulk slots]#
>>>
>>> [root@hulk pciehp]# echo 0000:40:03.0:pcie04 > unbind
>>> [root@hulk pciehp]#
>>>
>>> ...
>>> [102413.749632] pciehp 0000:40:03.0:pcie04: unloading service driver pciehp
>>> [102413.749638] pciehp_remove dev 0000:40:03.0, cmd_busy 1
>>> [102413.754929] pcie_disable_notification: ctrl cmd busy 1
>>> [102413.765903] pciehp 0000:40:03.0:pcie04: Timeout on hotplug command 0x11f1 (issued 102387672 msec ago)
>>
>> The fact that you got this timeout message means the controller did
>> not set the "No Command Completed Support" bit, right?  If we had
>> NO_CMD_CMPL(ctrl), pcie_wait_cmd() becomes a no-op, and we would
>> never print any timeout message.
>>
>> Since the "No Command Completed Support" bit is NOT set, we expect
>> to get an interrupt after every command completes.
>>
>> This sounds like the Intel CF118 erratum mentioned just above that timeout
>> message:
>>
>>          * Controllers with errata like Intel CF118 don't generate
>>          * completion notifications unless the power/indicator/interlock
>>          * control bits are changed.  On such controllers, we'll emit this
>>          * timeout message when we wait for completion of commands that
>>          * don't change those bits, e.g., commands that merely enable
>>          * interrupts.
>>
>> So to me, this sounds like pciehp is working correctly.  What did you
>> expect to happen instead?
>
> I think if we could warn the timeout messages when the timeout is reached, not be detected in
> next command write, it would be better.
>
> Something like:
>
> Write A command
> trigger a timeout delay work event
> interrupt coming (clean the cmd_busy, cancel the timeout delay work event)
> timeout delay event work (detect whether the cmd_busy is still set, if yes, warn the timeout message)
> ..
>
> But this is just my personal 3 seconds idea, it may make code more complex, I am not sure it's worth doing.

It would make the code more complex, I think.  If you want to code it
up, we could see what it would look like.

I think the real problem is the message itself.  We could emit the
message one second after pciehp claims the device (as you propose), or
the next time pciehp issues a command (as we do today).  EIther way, I
think users will see this as a problem.  You did, and I'm sure I
would, too.

Maybe there's some alternate wording that would be less alarming.

Or maybe we should emit the timeout message only if the previous
command actually changed PCC, PIC, AIC, or EIC, i.e., assume the Intel
CF118 erratum.

Bjorn

Re: pciehp command complete timeout issue

[Yijing Wang]

>>> The fact that you got this timeout message means the controller did
>>> not set the "No Command Completed Support" bit, right?  If we had
>>> NO_CMD_CMPL(ctrl), pcie_wait_cmd() becomes a no-op, and we would
>>> never print any timeout message.
>>>
>>> Since the "No Command Completed Support" bit is NOT set, we expect
>>> to get an interrupt after every command completes.
>>>
>>> This sounds like the Intel CF118 erratum mentioned just above that timeout
>>> message:
>>>
>>>          * Controllers with errata like Intel CF118 don't generate
>>>          * completion notifications unless the power/indicator/interlock
>>>          * control bits are changed.  On such controllers, we'll emit this
>>>          * timeout message when we wait for completion of commands that
>>>          * don't change those bits, e.g., commands that merely enable
>>>          * interrupts.
>>>
>>> So to me, this sounds like pciehp is working correctly.  What did you
>>> expect to happen instead?
>>
>> I think if we could warn the timeout messages when the timeout is reached, not be detected in
>> next command write, it would be better.
>>
>> Something like:
>>
>> Write A command
>> trigger a timeout delay work event
>> interrupt coming (clean the cmd_busy, cancel the timeout delay work event)
>> timeout delay event work (detect whether the cmd_busy is still set, if yes, warn the timeout message)
>> ..
>>
>> But this is just my personal 3 seconds idea, it may make code more complex, I am not sure it's worth doing.
> 
> It would make the code more complex, I think.  If you want to code it
> up, we could see what it would look like.

Make code more complex maybe is not a good idea, after all, this just is a alarming message issue.

> 
> I think the real problem is the message itself.  We could emit the
> message one second after pciehp claims the device (as you propose), or
> the next time pciehp issues a command (as we do today).  EIther way, I
> think users will see this as a problem.  You did, and I'm sure I
> would, too.

Agree.

> 
> Maybe there's some alternate wording that would be less alarming.
> 
> Or maybe we should emit the timeout message only if the previous
> command actually changed PCC, PIC, AIC, or EIC, i.e., assume the Intel
> CF118 erratum.

I like this, because we only touch PCC, PIC, AIC, EIC when we do hotplug,
all other slot control bits ABPE, PFDE, MRLSCE, PDCE, CCIE, HPIE, DLLSCE
only be touched when pciehp probe or reset.

Thanks!
Yijing.

> 
> Bjorn
> 
> .
> 


-- 
Thanks!
Yijing