[dpdk-dev] [PATCH v3 9/9] mempool/bucket: handle non-EAL lcores

[David Marchand]

On Tue, Jun 23, 2020 at 7:28 PM Andrew Rybchenko
<arybchenko at solarflare.com> wrote:
>
> On 6/22/20 4:25 PM, David Marchand wrote:
> > Convert to new lcore API to support non-EAL lcores.
> >
> > Signed-off-by: David Marchand <david.marchand at redhat.com>
> > ---
> >  drivers/mempool/bucket/rte_mempool_bucket.c | 131 ++++++++++++--------
> >  1 file changed, 82 insertions(+), 49 deletions(-)
> >
> > diff --git a/drivers/mempool/bucket/rte_mempool_bucket.c b/drivers/mempool/bucket/rte_mempool_bucket.c
> > index 5ce1ef16fb..0b4f42d330 100644
> > --- a/drivers/mempool/bucket/rte_mempool_bucket.c
> > +++ b/drivers/mempool/bucket/rte_mempool_bucket.c
> > @@ -55,6 +55,7 @@ struct bucket_data {
> >       struct rte_ring *shared_orphan_ring;
> >       struct rte_mempool *pool;
> >       unsigned int bucket_mem_size;
> > +     void *lcore_callback_handle;
> >  };
> >
> >  static struct bucket_stack *
> > @@ -345,6 +346,22 @@ bucket_dequeue_contig_blocks(struct rte_mempool *mp, void **first_obj_table,
> >       return 0;
> >  }
> >
> > +struct bucket_per_lcore_ctx {
>
> The structure is not used in per-lcore init and uninit
> functions. So, it is better to add _count to make it
> count specified. I.e. bucket_count_per_lcore_ctx.

Yes, and this aligns with its only user being renamed from
count_per_lcore to bucket_count_per_lcore.

>
> > +     const struct bucket_data *bd;
> > +     unsigned int count;
> > +};
> > +
> > +static int
> > +count_per_lcore(unsigned int lcore_id, void *arg)
> > +{
> > +     struct bucket_per_lcore_ctx *ctx = arg;
> > +
> > +     ctx->count += ctx->bd->obj_per_bucket *
> > +             ctx->bd->buckets[lcore_id]->top;
> > +     ctx->count += rte_ring_count(ctx->bd->adoption_buffer_rings[lcore_id]);
> > +     return 0;
> > +}
> > +
> >  static void
> >  count_underfilled_buckets(struct rte_mempool *mp,
> >                         void *opaque,
> > @@ -373,23 +390,66 @@ count_underfilled_buckets(struct rte_mempool *mp,
> >  static unsigned int
> >  bucket_get_count(const struct rte_mempool *mp)
> >  {
> > -     const struct bucket_data *bd = mp->pool_data;
> > -     unsigned int count =
> > -             bd->obj_per_bucket * rte_ring_count(bd->shared_bucket_ring) +
> > -             rte_ring_count(bd->shared_orphan_ring);
> > -     unsigned int i;
> > +     struct bucket_per_lcore_ctx ctx;
>
> Just a nit, but I think that ctx is too generic.
> (some time ago bucket_data bd was ctx in fact :) )
> May be bplc? Up to you.

Ack.


>
> >
> > -     for (i = 0; i < RTE_MAX_LCORE; i++) {
> > -             if (!rte_lcore_is_enabled(i))
> > -                     continue;
> > -             count += bd->obj_per_bucket * bd->buckets[i]->top +
> > -                     rte_ring_count(bd->adoption_buffer_rings[i]);
> > -     }
> > +     ctx.bd = mp->pool_data;
> > +     ctx.count = ctx.bd->obj_per_bucket *
> > +             rte_ring_count(ctx.bd->shared_bucket_ring);
> > +     ctx.count += rte_ring_count(ctx.bd->shared_orphan_ring);
> >
> > +     rte_lcore_iterate(count_per_lcore, &ctx);
> >       rte_mempool_mem_iter((struct rte_mempool *)(uintptr_t)mp,
> > -                          count_underfilled_buckets, &count);
> > +                          count_underfilled_buckets, &ctx.count);
> > +
> > +     return ctx.count;
> > +}
> > +
> > +static int
> > +bucket_init_per_lcore(unsigned int lcore_id, void *arg)
>
> It should be no bucket_ prefix here, or it should be bucket_
> prefix above in count_per_lcore.

As mentioned before, ack.


>
> > +{
> > +     char rg_name[RTE_RING_NAMESIZE];
> > +     struct bucket_data *bd = arg;
> > +     struct rte_mempool *mp;
> > +     int rg_flags;
> > +     int rc;
> > +
> > +     mp = bd->pool;
> > +     bd->buckets[lcore_id] = bucket_stack_create(mp,
> > +             mp->size / bd->obj_per_bucket);
> > +     if (bd->buckets[lcore_id] == NULL)
> > +             goto error;
> > +
> > +     rc = snprintf(rg_name, sizeof(rg_name), RTE_MEMPOOL_MZ_FORMAT ".a%u",
> > +             mp->name, lcore_id);
> > +     if (rc < 0 || rc >= (int)sizeof(rg_name))
> > +             goto error;
> > +
> > +     rg_flags = RING_F_SC_DEQ;
> > +     if (mp->flags & MEMPOOL_F_SP_PUT)
> > +             rg_flags |= RING_F_SP_ENQ;
> > +     if (mp->flags & MEMPOOL_F_SC_GET)
> > +             rg_flags |= RING_F_SC_DEQ;
>
> There is not point to have two above lines here, since
> RING_F_SC_DEQ is always set.

Ah yes, I did not realise when moving the code.


>
> > +     bd->adoption_buffer_rings[lcore_id] = rte_ring_create(rg_name,
> > +             rte_align32pow2(mp->size + 1), mp->socket_id, rg_flags);
> > +     if (bd->adoption_buffer_rings[lcore_id] == NULL)
> > +             goto error;
> >
> > -     return count;
> > +     return 0;
> > +error:
> > +     rte_free(bd->buckets[lcore_id]);
> > +     bd->buckets[lcore_id] = NULL;
> > +     return -1;
>
> Why does the API collapse all negative errnos into -1?
> (I don't think it is critical, just want to know why).

I collapsed everything as a single error as we have a partial idea of
what went wrong when calling this callback with all lcores at
registration.
We could get a specific error reported by this callback, but then we
would not know on which lcore (programmatically).
And in the end, all errors will summarize as a lack of resources, I do
not expect a need for input validation.

Maybe you have other use cases in mind?


Thanks for the review.

-- 
David Marchand