[PATCH] KVM/arm: kernel low level debug support for ARM32 virtual platforms

[Mario Smarduch]

On 11/4/2015 10:51 AM, Ard Biesheuvel wrote:
> On 4 November 2015 at 19:49, Christopher Covington <cov at codeaurora.org> wrote:
>> On 11/04/2015 08:31 AM, Christoffer Dall wrote:
>>> On Tue, Nov 03, 2015 at 01:39:44PM -0600, Rob Herring wrote:
>>>> On Tue, Nov 3, 2015 at 1:17 PM, Mario Smarduch <m.smarduch at samsung.com> wrote:
>>>>> On 11/3/2015 9:55 AM, Will Deacon wrote:
>>>>>> On Tue, Nov 03, 2015 at 09:44:52AM -0800, Mario Smarduch wrote:
>>>>>>> On 11/3/2015 8:33 AM, Christopher Covington wrote:
>>>>>>>> On 11/02/2015 06:51 PM, Mario Smarduch wrote:
>>>>>>>>>    this is a re-post from couple weeks ago, please take time to review this
>>>>>>>>> simple patch which simplifies DEBUG_LL and prevents kernel crash on virtual
>>>>>>>>> platforms.
>>>>>>>>>
>>>>>>>>> Before this patch DEBUG_LL for 'dummy virtual machine':
>>>>>>>>>
>>>>>>>>> ( ) Kernel low-level debugging via EmbeddedICE DCC channel
>>>>>>>>> ( ) Kernel low-level debug output via semihosting I/O
>>>>>>>>> ( ) Kernel low-level debugging via 8250 UART
>>>>>>>>> ( ) Kernel low-level debugging via ARM Ltd PL01x Primecell
>>>>>>>>>
>>>>>>>>> In summary if debug uart is not emulated kernel crashes.
>>>>>>>>> And once you pass that hurdle, uart physical/virtual addresses are unknown.
>>>>>>>>> DEBUG_LL comes in handy on many occasions and should be somewhat
>>>>>>>>> intuitive to use like it is for physical platforms. For virtual platforms
>>>>>>>>> user may start daubting the host and get into a bigger mess.
>>>>>>>>>
>>>>>>>>> After this patch is applied user gets:
>>>>>>>>>
>>>>>>>>> (X) Kernel low-level debugging on QEMU Virtual Platform
>>>>>>>>> ( ) Kernel low-level debugging on Kvmtool Virtual Platform
>>>>>>>>>    ..... above repeated ....
>>>>>>>>>
>>>>>>>>> The virtual addresses selected follow arm reference models, high in vmalloc
>>>>>>>>> section with high mem enabled and guest running with >= 1GB of memory. The
>>>>>>>>> offset is leftover from arm reference models.
>>>>>>>>
>>>>>>>> Which model? It doesn't appear to match the vexpress AEM/RTSM/FVP/whatever
>>>>>>>> which used 0x1c090000 for UART0.
>>>>>>>
>>>>>>> I recall QEMU virt model had it's own physical address map, for sure I saw the
>>>>>>> virtio-mmio regions assigned in some ARM document. Peter would you know?
>>>>>>>
>>>>>>> As far as kvmtool I'm not sure, currently PC1 COM1 port is used? Andre will that
>>>>>>> stay fixed?
>>>>>>
>>>>>> We make absolutely no guarantees about the memory map provided by kvmtool.
>>>>>
>>>>> If that's also the case for qemu, then I guess the best you can do is find a way
>>>>> to dump the device tree. Find the uart, physical address and try figure out the
>>>>> virtual address.
>>>>>
>>>>> Pretty involved, hoped for something more automated since that's a handy feature.
>>>>
>>>> You really only need LL_DEBUG now if you are debugging very early code
>>>> before memory is setup and/or bad memory. Use earlycon instead which
>>>> should already be supported both via the pl011 or semihosting. I used
>>>> it with QEMU semihosting support.
>>>>
>>> Then we should really document how to use that with qemu's virt platform
>>> and kvmtool's platform on both 32-bit and 64-bit so that users can
>>> easily figure out what they're doing wrong when they get no output.
>>>
>>> In practice, the address for the pl011 is quite unlikely to change, I
>>> dare speculate, so that documentation shouldn't need frequent updating.
>>
>> Is it not on by default since the following change?
>>
>> http://git.qemu.org/?p=qemu.git;a=commitdiff;h=f022b8e95379b0433d13509706b66f38fc15dde8
>>
> 
> Yes, but it still requires the plain 'earlycon' argument (i.e, without
> '=pl011,...') to be passed on the kernel command line if you want
> early output.
> 

I spent time debugging 'earlycon' for pl011, ironically using DEBUG_LL, from the
looks of it no mmio uart should work for armv7. It appears earlycon_map()
requires a fixed mapping similar to arm64.

Comparing both options, DEBUG_LL takes you from kernel decompress code to early
FDT parsing. A lot of early_print() calls won't work if DEBUG_LL is not enabled
including dump_machine_table which ends in a endless loop. IMO it's worth
turning this option on for that and other reasons.

'earlycon' is enabled some ways up in setup_arch().

As far as the patch, providing a hint to the user with probable uart addresses
would help and in the worst case "see the latest device tree for the virtual
platform".

- Mario