Re: [PATCH 13/26] ppc: Convert mmu context allocation to new IDA API

[Matthew Wilcox]

On Fri, Jun 22, 2018 at 12:15:11PM +1000, Nicholas Piggin wrote:
> On Thu, 21 Jun 2018 14:28:22 -0700
> Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >  static int alloc_context_id(int min_id, int max_id)
...
> > -	spin_lock(&mmu_context_lock);
> > -	err = ida_get_new_above(&mmu_context_ida, min_id, &index);
> > -	spin_unlock(&mmu_context_lock);
...
> > @@ -182,13 +148,11 @@ static void destroy_contexts(mm_context_t *ctx)
> >  {
> >  	int index, context_id;
> >  
> > -	spin_lock(&mmu_context_lock);
> >  	for (index = 0; index < ARRAY_SIZE(ctx->extended_id); index++) {
> >  		context_id = ctx->extended_id[index];
> >  		if (context_id)
> > -			ida_remove(&mmu_context_ida, context_id);
> > +			ida_free(&mmu_context_ida, context_id);
> >  	}
> > -	spin_unlock(&mmu_context_lock);
> >  }
> >  
> >  static void pte_frag_destroy(void *pte_frag)
> 
> This hunk should be okay because the mmu_context_lock does not protect
> the extended_id array, right Aneesh?

That's my understanding.  The code today does this:

static inline int alloc_extended_context(struct mm_struct *mm,
                                         unsigned long ea)
{
        int context_id;

        int index = ea >> MAX_EA_BITS_PER_CONTEXT;

        context_id = hash__alloc_context_id();
        if (context_id < 0)
                return context_id;

        VM_WARN_ON(mm->context.extended_id[index]);
        mm->context.extended_id[index] = context_id;

so it's not currently protected by this lock.  I suppose we are currently
protected from destroy_contexts() being called twice simultaneously, but
you'll notice that we don't zero the array elements in destroy_contexts(),
so if we somehow had a code path which could call it concurrently, we'd
be seeing warnings when the second caller tried to remove the context
IDs from the IDA.  I deduced that something else must be preventing
this situation from occurring (like, oh i don't know, this function only
being called on process exit, so implicitly only called once per context).

> Reviewed-by: Nicholas Piggin <npiggin@xxxxxxxxx>

Thanks.