[ Restoring the recipients after mistakenly pressing reply instead of
reply-all ]
> On May 9, 2019, at 12:11 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, May 09, 2019 at 06:50:00PM +0000, Nadav Amit wrote:
>>> On May 9, 2019, at 11:24 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>>
>>> On Thu, May 09, 2019 at 05:36:29PM +0000, Nadav Amit wrote:
>
>>>> As a simple optimization, I think it is possible to hold multiple nesting
>>>> counters in the mm, similar to tlb_flush_pending, for freed_tables,
>>>> cleared_ptes, etc.
>>>>
>>>> The first time you set tlb->freed_tables, you also atomically increase
>>>> mm->tlb_flush_freed_tables. Then, in tlb_flush_mmu(), you just use
>>>> mm->tlb_flush_freed_tables instead of tlb->freed_tables.
>>>
>>> That sounds fraught with races and expensive; I would much prefer to not
>>> go there for this arguably rare case.
>>>
>>> Consider such fun cases as where CPU-0 sees and clears a PTE, CPU-1
>>> races and doesn't see that PTE. Therefore CPU-0 sets and counts
>>> cleared_ptes. Then if CPU-1 flushes while CPU-0 is still in mmu_gather,
>>> it will see cleared_ptes count increased and flush that granularity,
>>> OTOH if CPU-1 flushes after CPU-0 completes, it will not and potentiall
>>> miss an invalidate it should have had.
>>
>> CPU-0 would send a TLB shootdown request to CPU-1 when it is done, so I
>> don’t see the problem. The TLB shootdown mechanism is independent of the
>> mmu_gather for the matter.
>
> Duh.. I still don't like those unconditional mm wide atomic counters.
>
>>> This whole concurrent mmu_gather stuff is horrible.
>>>
>>> /me ponders more....
>>>
>>> So I think the fundamental race here is this:
>>>
>>> CPU-0 CPU-1
>>>
>>> tlb_gather_mmu(.start=1, tlb_gather_mmu(.start=2,
>>> .end=3); .end=4);
>>>
>>> ptep_get_and_clear_full(2)
>>> tlb_remove_tlb_entry(2);
>>> __tlb_remove_page();
>>> if (pte_present(2)) // nope
>>>
>>> tlb_finish_mmu();
>>>
>>> // continue without TLBI(2)
>>> // whoopsie
>>>
>>> tlb_finish_mmu();
>>> tlb_flush() -> TLBI(2)
>>>
>>>
>>> And we can fix that by having tlb_finish_mmu() sync up. Never let a
>>> concurrent tlb_finish_mmu() complete until all concurrenct mmu_gathers
>>> have completed.
>>>
>>> This should not be too hard to make happen.
>>
>> This synchronization sounds much more expensive than what I proposed. But I
>> agree that cache-lines that move from one CPU to another might become an
>> issue. But I think that the scheme I suggested would minimize this overhead.
>
> Well, it would have a lot more unconditional atomic ops. My scheme only
> waits when there is actual concurrency.
Well, something has to give. I didn’t think that if the same core does the
atomic op it would be too expensive.
> I _think_ something like the below ought to work, but its not even been
> near a compiler. The only problem is the unconditional wakeup; we can
> play games to avoid that if we want to continue with this.
>
> Ideally we'd only do this when there's been actual overlap, but I've not
> found a sensible way to detect that.
>
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 4ef4bbe78a1d..b70e35792d29 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -590,7 +590,12 @@ static inline void dec_tlb_flush_pending(struct mm_struct *mm)
> *
> * Therefore we must rely on tlb_flush_*() to guarantee order.
> */
> - atomic_dec(&mm->tlb_flush_pending);
> + if (atomic_dec_and_test(&mm->tlb_flush_pending)) {
> + wake_up_var(&mm->tlb_flush_pending);
> + } else {
> + wait_event_var(&mm->tlb_flush_pending,
> + !atomic_read_acquire(&mm->tlb_flush_pending));
> + }
> }
It still seems very expensive to me, at least for certain workloads (e.g.,
Apache with multithreaded MPM).
It may be possible to avoid false-positive nesting indications (when the
flushes do not overlap) by creating a new struct mmu_gather_pending, with
something like:
struct mmu_gather_pending {
u64 start;
u64 end;
struct mmu_gather_pending *next;
}
tlb_finish_mmu() would then iterate over the mm->mmu_gather_pending
(pointing to the linked list) and find whether there is any overlap. This
would still require synchronization (acquiring a lock when allocating and
deallocating or something fancier).
