[PATCH] powerpc/pseries: explicitly reschedule during drmem_lmb list traversal

[Michael Ellerman]

Nathan Lynch <nathanl at linux.ibm.com> writes:
> Michael Ellerman <mpe at ellerman.id.au> writes:
>> Nathan Lynch <nathanl at linux.ibm.com> writes:
>>> Michael Ellerman <mpe at ellerman.id.au> writes:
>>>> Nathan Lynch <nathanl at linux.ibm.com> writes:
>>>>> Laurent Dufour <ldufour at linux.ibm.com> writes:
>>>>>> Le 28/07/2020 à 19:37, Nathan Lynch a écrit :
>>>>>>> The drmem lmb list can have hundreds of thousands of entries, and
>>>>>>> unfortunately lookups take the form of linear searches. As long as
>>>>>>> this is the case, traversals have the potential to monopolize the CPU
>>>>>>> and provoke lockup reports, workqueue stalls, and the like unless
>>>>>>> they explicitly yield.
>>>>>>> 
>>>>>>> Rather than placing cond_resched() calls within various
>>>>>>> for_each_drmem_lmb() loop blocks in the code, put it in the iteration
>>>>>>> expression of the loop macro itself so users can't omit it.
>>>>>>
>>>>>> Is that not too much to call cond_resched() on every LMB?
>>>>>>
>>>>>> Could that be less frequent, every 10, or 100, I don't really know ?
>>>>>
>>>>> Everything done within for_each_drmem_lmb is relatively heavyweight
>>>>> already. E.g. calling dlpar_remove_lmb()/dlpar_add_lmb() can take dozens
>>>>> of milliseconds. I don't think cond_resched() is an expensive check in
>>>>> this context.
>>>>
>>>> Hmm, mostly.
>>>>
>>>> But there are quite a few cases like drmem_update_dt_v1():
>>>>
>>>> 	for_each_drmem_lmb(lmb) {
>>>> 		dr_cell->base_addr = cpu_to_be64(lmb->base_addr);
>>>> 		dr_cell->drc_index = cpu_to_be32(lmb->drc_index);
>>>> 		dr_cell->aa_index = cpu_to_be32(lmb->aa_index);
>>>> 		dr_cell->flags = cpu_to_be32(drmem_lmb_flags(lmb));
>>>>
>>>> 		dr_cell++;
>>>> 	}
>>>>
>>>> Which will compile to a pretty tight loop at the moment.
>>>>
>>>> Or drmem_update_dt_v2() which has two loops over all lmbs.
>>>>
>>>> And although the actual TIF check is cheap the function call to do it is
>>>> not free.
>>>>
>>>> So I worry this is going to make some of those long loops take even
>>>> longer.
>>>
>>> That's fair, and I was wrong - some of the loop bodies are relatively
>>> simple, not doing allocations or taking locks, etc.
>>>
>>> One way to deal is to keep for_each_drmem_lmb() as-is and add a new
>>> iterator that can reschedule, e.g. for_each_drmem_lmb_slow().
>>
>> If we did that, how many call-sites would need converting?
>> Is it ~2 or ~20 or ~200?
>
> At a glance I would convert 15-20 out of the 24 users in the tree I'm
> looking at. Let me know if I should do a v2 with that approach.

OK, that's a bunch of churn then, if we're planning to rework the code
significantly in the near future.

One thought, which I possibly should not put in writing, is that we
could use the alignment of the pointer as a poor man's substitute for a
counter, eg:

+static inline struct drmem_lmb *drmem_lmb_next(struct drmem_lmb *lmb)
+{
+	if (lmb % PAGE_SIZE == 0)
+		cond_resched();
+
+	return ++lmb;
+}

I think the lmbs are allocated in a block, so I think that will work.
Maybe PAGE_SIZE is not the right size to use, but you get the idea.

Gross I know, but might be OK as short term solution?

cheers