Re: BUG: MAX_STACK_TRACE_ENTRIES too low! (2)

[Bart Van Assche <bvanassche@acm.org>]

On 7/12/19 1:55 AM, Peter Zijlstra wrote:
> On Thu, Jul 11, 2019 at 11:53:12AM -0700, Bart Van Assche wrote:
>> On 7/10/19 3:09 PM, Peter Zijlstra wrote:
>>> One thing I mentioned when Thomas did the unwinder API changes was
>>> trying to move lockdep over to something like stackdepot.
>>>
>>> We can't directly use stackdepot as is, because it uses locks and memory
>>> allocation, but we could maybe add a lower level API to it and use that
>>> under the graph_lock() on static storage or something.
>>>
>>> Otherwise we'll have to (re)implement something like it.
>>>
>>> I've not looked at it in detail.
>>
>> Hi Peter,
>>
>> Is something like the untested patch below perhaps what you had in mind?
> 
> Most excellent, yes! Now I suppose the $64000 question is if it actually
> reduces the amount of storage we use for stack traces..
> 
> Seems to boot just fine.. :-)

Hi Peter,

On my setup after some time the space occupied by stack traces stabilizes
to the following:

# grep stack /proc/lockdep_stats
  stack-trace entries:                169456 [max: 524288]
  number of stack traces:        9073
  number of stack hash chains:   6980

I think these numbers show that there are some but not too many hash
collisions (the code for the hash statistics was added after I e-mailed my
patch).

Bart.