Re: [RFC 15/20] mm: detect deferred TLB flushes in vma granularity

[Nadav Amit <nadav.amit@gmail.com>]

> On Feb 1, 2021, at 4:14 PM, Andy Lutomirski <luto@amacapital.net> wrote:
> 
> 
>> On Feb 1, 2021, at 2:04 PM, Nadav Amit <nadav.amit@gmail.com> wrote:
>> 
>> Andy’s comments managed to make me realize this code is wrong. We must
>> call inc_mm_tlb_gen(mm) every time.
>> 
>> Otherwise, a CPU that saw the old tlb_gen and updated it in its local
>> cpu_tlbstate on a context-switch. If the process was not running when the
>> TLB flush was issued, no IPI will be sent to the CPU. Therefore, later
>> switch_mm_irqs_off() back to the process will not flush the local TLB.
>> 
>> I need to think if there is a better solution. Multiple calls to
>> inc_mm_tlb_gen() during deferred flushes would trigger a full TLB flush
>> instead of one that is specific to the ranges, once the flush actually takes
>> place. On x86 it’s practically a non-issue, since anyhow any update of more
>> than 33-entries or so would cause a full TLB flush, but this is still ugly.
> 
> What if we had a per-mm ring buffer of flushes?  When starting a flush, we would stick the range in the ring buffer and, when flushing, we would read the ring buffer to catch up.  This would mostly replace the flush_tlb_info struct, and it would let us process multiple partial flushes together.

I wanted to sleep on it, and went back and forth on whether it is the right
direction, hence the late response.

I think that what you say make sense. I think that I even tried to do once
something similar for some reason, but my memory plays tricks on me.

So tell me what you think on this ring-based solution. As you said, you keep
per-mm ring of flush_tlb_info. When you queue an entry, you do something
like:

#define RING_ENTRY_INVALID (0)

  gen = inc_mm_tlb_gen(mm);
  struct flush_tlb_info *info = mm->ring[gen % RING_SIZE];
  spin_lock(&mm->ring_lock);
  WRITE_ONCE(info->new_tlb_gen, RING_ENTRY_INVALID);
  smp_wmb();
  info->start = start;
  info->end = end;
  info->stride_shift = stride_shift;
  info->freed_tables = freed_tables;
  smp_store_release(&info->new_tlb_gen, gen);
  spin_unlock(&mm->ring_lock);
  
When you flush you use the entry generation as a sequence lock. On overflow
of the ring (i.e., sequence number mismatch) you perform a full flush:

  for (gen = mm->tlb_gen_completed; gen < mm->tlb_gen; gen++) {
	struct flush_tlb_info *info = &mm->ring[gen % RING_SIZE];

	// detect overflow and invalid entries
	if (smp_load_acquire(info->new_tlb_gen) != gen)
		goto full_flush;

	start = min(start, info->start);
	end = max(end, info->end);
	stride_shift = min(stride_shift, info->stride_shift);
	freed_tables |= info.freed_tables;
	smp_rmb();

	// seqlock-like check that the information was not updated 
	if (READ_ONCE(info->new_tlb_gen) != gen)
		goto full_flush;
  }

On x86 I suspect that performing a full TLB flush would anyhow be the best
thing to do if there is more than a single entry. I am also not sure that it
makes sense to check the ring from flush_tlb_func_common() (i.e., in each
IPI handler) as it might cause cache thrashing.

Instead it may be better to do so from flush_tlb_mm_range(), when the
flushes are initiated, and use an aggregated flush_tlb_info for the flush.

It may also be better to have the ring arch-independent, so it would
resemble more of mmu_gather (the parts about the TLB flush information,
without the freed pages stuff).

We can detect deferred TLB flushes either by storing “deferred_gen” in the
page-tables/VMA (as I did) or by going over the ring, from tlb_gen_completed
to tlb_gen, and checking for an overlap. I think page-tables would be most
efficient/scalable, but perhaps going over the ring would be easier to
understand logic.

Makes sense? Thoughts?