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

[Andy Lutomirski]

> On Feb 1, 2021, at 2:04 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
> 
> 
>> 
>> On Jan 30, 2021, at 4:11 PM, Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>> 
>> From: Nadav Amit <namit@xxxxxxxxxx>
>> 
>> Currently, deferred TLB flushes are detected in the mm granularity: if
>> there is any deferred TLB flush in the entire address space due to NUMA
>> migration, pte_accessible() in x86 would return true, and
>> ptep_clear_flush() would require a TLB flush. This would happen even if
>> the PTE resides in a completely different vma.
> 
> [ snip ]
> 
>> +static inline void read_defer_tlb_flush_gen(struct mmu_gather *tlb)
>> +{
>> +    struct mm_struct *mm = tlb->mm;
>> +    u64 mm_gen;
>> +
>> +    /*
>> +     * Any change of PTE before calling __track_deferred_tlb_flush() must be
>> +     * performed using RMW atomic operation that provides a memory barriers,
>> +     * such as ptep_modify_prot_start().  The barrier ensure the PTEs are
>> +     * written before the current generation is read, synchronizing
>> +     * (implicitly) with flush_tlb_mm_range().
>> +     */
>> +    smp_mb__after_atomic();
>> +
>> +    mm_gen = atomic64_read(&mm->tlb_gen);
>> +
>> +    /*
>> +     * This condition checks for both first deferred TLB flush and for other
>> +     * TLB pending or executed TLB flushes after the last table that we
>> +     * updated. In the latter case, we are going to skip a generation, which
>> +     * would lead to a full TLB flush. This should therefore not cause
>> +     * correctness issues, and should not induce overheads, since anyhow in
>> +     * TLB storms it is better to perform full TLB flush.
>> +     */
>> +    if (mm_gen != tlb->defer_gen) {
>> +        VM_BUG_ON(mm_gen < tlb->defer_gen);
>> +
>> +        tlb->defer_gen = inc_mm_tlb_gen(mm);
>> +    }
>> +}
> 
> 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.