On Thu, Oct 29 2020 at 23:18, Thomas Gleixner wrote:
>
> There is also a still to be investigated question from Linus on the initial
> posting versus the per cpu / per task mapping stack depth which might need
> to be made larger due to the ability to take page faults within a mapping
> region.
I looked deeper into that and we have a stack depth of 20. That's plenty
and I couldn't find a way to get above 10 nested ones including faults,
interrupts, softirqs. With some stress testing I was not able to get over
a maximum of 6 according to the traceprintk I added.
For some obscure reason when CONFIG_DEBUG_HIGHMEM is enabled the stack
depth is increased from 20 to 41. But the only thing DEBUG_HIGHMEM does
is to enable a few BUG_ON()'s in the mapping code.
That's a leftover from the historical mapping code which had fixed
entries for various purposes. DEBUG_HIGHMEM inserted guard mappings
between the map types. But that got all ditched when kmap_atomic()
switched to a stack based map management. Though the WITH_KM_FENCE magic
survived without being functional. All the thing does today is to
increase the stack depth.
I just made that functional again by keeping the stack depth increase
and utilizing every second slot. That should catch Willy's mapping
problem nicely if he bothers to test on 32bit :)
Thanks,
tglx
