[PATCH] accel/qaic: Add crashdump to Sahara

[Konrad Dybcio]

On 19.09.2024 5:00 PM, Jeffrey Hugo wrote:
> On 9/18/2024 5:41 PM, Konrad Dybcio wrote:
>> On 18.09.2024 5:52 PM, Jeffrey Hugo wrote:
>>> The Sahara protocol has a crashdump functionality. In the hello
>>> exchange, the device can advertise it has a memory dump available for
>>> the host to collect. Instead of the device making requests of the host,
>>> the host requests data from the device which can be later analyzed.
>>>
>>> Implement this functionality and utilize the devcoredump framework for
>>> handing the dump over to userspace.
>>>
>>> Similar to how firmware loading in Sahara involves multiple files,
>>> crashdump can consist of multiple files for different parts of the dump.
>>> Structure these into a single buffer that userspace can parse and
>>> extract the original files from.
>>>
>>> Reviewed-by: Carl Vanderlip <quic_carlv at quicinc.com>
>>> Signed-off-by: Jeffrey Hugo <quic_jhugo at quicinc.com>
>>> ---
>>
>> I gave this a brief read, but.. aren't you dumping however much DRAM the
>> AIC100 has (and then some SRAM) onto the host machine without the user
>> asking for it (i.e. immediately after the AIC crashes)?
> 
> I'm not entirely clear what the concern is.  Too much host RAM usage maybe?

Yes

> In short, I think the direct answer is yes and no.
> 
> We put the dump content in the host RAM and allow the user to decide if they want to save it.  The user has 5 minutes to do something with the dump, then the devcoredump framework automatically frees the content in RAM.  Typically the user would access the sysfs file provided by devcoredump, and save the contents to the file system for offline processing.
> 
> There are a few other GPUs and several other devices that do the same. Panfrost appears to save every BO the user allocated into the dump, which would suggest that the user could create an arbitrarily large dump.

Right, freedreno does something similar. Perhaps a user concerned about this
could simply disable CONFIG_DEVCOREDUMP.

> 
> In the case of AIC100, it is technically possible for the entire device DRAM and SRAM to be offloaded.  That is up to the FW to decide if all of that is relevant.  Current implementation of the FW is heavily aggressive on what it selects for the dump, and current dumps are in the 100-200MB range.

OK this basically invalidates my concerns.. other boards I had ramdump on me
just spat out all the 16 or however much gigabytes of RAM they have..


> It feels like you are implying the user should somehow request the dump by having devcoredump publish something, and then hook into the user's reads of the sysfs to go collect the dump.  I worry that means the driver would then need to determine when there is no user interested in collecting the dump, in order to continue the reboot process.  I expect that would be a 5 minute delay (devcoredump deciding there is no user interest after 5 minutes).  I worry that users would object to such a delay given customer feedback we've had on getting the devices into service quickly.

Yeah no, this wouldn't be good.

Konrad