Re: [dpdk-dev] custom align for mempool elements

[Olivier Matz <olivier.matz@6wind.com>]

Hi Gregory,

On Fri, 5 May 2017 20:23:15 +0300, Gregory Etelson <gregory@weka.io> wrote:
> Hello Oliver,
> 
> Our application writes data from incoming MBUFs to storage device.
> For performance considerations we use O_DIRECT storage access
> and work in 'zero copy' data mode.
> To achieve the 'zero copy' we MUST arrange data in all MBUFs to be 512
> bytes aligned
> With pre-calculated custom pool element alignment and right
>  RTE_PKTMBUF_HEADROOM value
> we can set MBUF data alignment to any value. In our case, 512 bytes
> 
> Current implementation sets custom mempool alignment like this:
> 
>     struct rte_mempool *mp = rte_mempool_create_empty("MBUF pool",
> mbufs_count,
>                  elt_size, cache_size, sizeof(struct
> rte_pktmbuf_pool_private), rte_socket_id(), 0);
> 
>     rte_pktmbuf_pool_init(mp, &mbp_priv);
>   *mp->elt_align = align*;
>     rte_mempool_populate_default(mp);

I think we should try to avoid modifying mp->elt_align directly.
A new api could be added for that, maybe something like:

  int rte_mempool_set_obj_align(struct rte_mempool *mp, size_t align)
  size_t rte_mempool_get_obj_align(struct rte_mempool *mp)

The set() function would have to be called before mempool_populate().

We need to take care about conflict with the MEMPOOL_F_NO_CACHE_ALIGN
flag. I think we should keep the compat for this flag user. This flag
could be deprecated and removed later. I think there may also be some
conflicts with MEMPOOL_F_NO_SPREAD.

As I said in my previous mail, if the patch breaks the ABI (the mempool
structure), it has to follow the abi deprecation process (= a notice in
17.05 and the patch for 17.08).

I'm afraid it would be quite late for the notice though.

Regards,
Olivier


> 
> Regards,
> Gregory
> 
> On Fri, May 5, 2017 at 2:26 PM, Olivier Matz <olivier.matz@6wind.com> wrote:
> 
> > Hi Gregory,
> >
> > On Wed, 26 Apr 2017 07:00:49 +0300, Gregory Etelson <gregory@weka.io>
> > wrote:  
> > > Signed-off-by: Gregory Etelson <gregory@weka.io>
> > > ---
> > >  lib/librte_mempool/rte_mempool.c | 27 ++++++++++++++++++++-------
> > >  lib/librte_mempool/rte_mempool.h |  1 +
> > >  2 files changed, 21 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_  
> > mempool.c  
> > > index f65310f..c780df3 100644
> > > --- a/lib/librte_mempool/rte_mempool.c
> > > +++ b/lib/librte_mempool/rte_mempool.c
> > > @@ -382,7 +382,7 @@ rte_mempool_populate_phys(struct rte_mempool *mp,  
> > char *vaddr,  
> > >       if (mp->flags & MEMPOOL_F_NO_CACHE_ALIGN)
> > >               off = RTE_PTR_ALIGN_CEIL(vaddr, 8) - vaddr;
> > >       else
> > > -             off = RTE_PTR_ALIGN_CEIL(vaddr, RTE_CACHE_LINE_SIZE) -  
> > vaddr;  
> > > +             off = RTE_PTR_ALIGN_CEIL(vaddr, mp->elt_align) - vaddr;
> > >
> > >       while (off + total_elt_sz <= len && mp->populated_size < mp->size)  
> > {  
> > >               off += mp->header_size;
> > > @@ -392,6 +392,7 @@ rte_mempool_populate_phys(struct rte_mempool *mp,  
> > char *vaddr,  
> > >               else
> > >                       mempool_add_elem(mp, (char *)vaddr + off, paddr +  
> > off);  
> > >               off += mp->elt_size + mp->trailer_size;
> > > +             off = RTE_ALIGN_CEIL(off, mp->elt_align);
> > >               i++;
> > >       }
> > >
> > > @@ -508,6 +509,20 @@ rte_mempool_populate_virt(struct rte_mempool *mp,  
> > char *addr,  
> > >       return ret;
> > >  }
> > >
> > > +static uint32_t
> > > +mempool_default_elt_aligment(void)
> > > +{
> > > +     uint32_t align;
> > > +     if (rte_xen_dom0_supported()) {
> > > +                align = RTE_PGSIZE_2M;;
> > > +     } else if (rte_eal_has_hugepages()) {
> > > +                align = RTE_CACHE_LINE_SIZE;
> > > +     } else {
> > > +                align = getpagesize();
> > > +     }
> > > +     return align;
> > > +}
> > > +
> > >  /* Default function to populate the mempool: allocate memory in  
> > memzones,  
> > >   * and populate them. Return the number of objects added, or a negative
> > >   * value on error.
> > > @@ -518,7 +533,7 @@ rte_mempool_populate_default(struct rte_mempool *mp)
> > >       int mz_flags = RTE_MEMZONE_1GB|RTE_MEMZONE_SIZE_HINT_ONLY;
> > >       char mz_name[RTE_MEMZONE_NAMESIZE];
> > >       const struct rte_memzone *mz;
> > > -     size_t size, total_elt_sz, align, pg_sz, pg_shift;
> > > +     size_t size, total_elt_sz, pg_sz, pg_shift;
> > >       phys_addr_t paddr;
> > >       unsigned mz_id, n;
> > >       int ret;
> > > @@ -530,15 +545,12 @@ rte_mempool_populate_default(struct rte_mempool  
> > *mp)  
> > >       if (rte_xen_dom0_supported()) {
> > >               pg_sz = RTE_PGSIZE_2M;
> > >               pg_shift = rte_bsf32(pg_sz);
> > > -             align = pg_sz;
> > >       } else if (rte_eal_has_hugepages()) {
> > >               pg_shift = 0; /* not needed, zone is physically contiguous  
> > */  
> > >               pg_sz = 0;
> > > -             align = RTE_CACHE_LINE_SIZE;
> > >       } else {
> > >               pg_sz = getpagesize();
> > >               pg_shift = rte_bsf32(pg_sz);
> > > -             align = pg_sz;
> > >       }
> > >
> > >       total_elt_sz = mp->header_size + mp->elt_size + mp->trailer_size;
> > > @@ -553,11 +565,11 @@ rte_mempool_populate_default(struct rte_mempool  
> > *mp)  
> > >               }
> > >
> > >               mz = rte_memzone_reserve_aligned(mz_name, size,
> > > -                     mp->socket_id, mz_flags, align);
> > > +                     mp->socket_id, mz_flags, mp->elt_align);
> > >               /* not enough memory, retry with the biggest zone we have  
> > */  
> > >               if (mz == NULL)
> > >                       mz = rte_memzone_reserve_aligned(mz_name, 0,
> > > -                             mp->socket_id, mz_flags, align);
> > > +                             mp->socket_id, mz_flags, mp->elt_align);
> > >               if (mz == NULL) {
> > >                       ret = -rte_errno;
> > >                       goto fail;
> > > @@ -827,6 +839,7 @@ rte_mempool_create_empty(const char *name, unsigned  
> > n, unsigned elt_size,  
> > >       /* Size of default caches, zero means disabled. */
> > >       mp->cache_size = cache_size;
> > >       mp->private_data_size = private_data_size;
> > > +     mp->elt_align = mempool_default_elt_aligment();
> > >       STAILQ_INIT(&mp->elt_list);
> > >       STAILQ_INIT(&mp->mem_list);
> > >
> > > diff --git a/lib/librte_mempool/rte_mempool.h b/lib/librte_mempool/rte_  
> > mempool.h  
> > > index 48bc8ea..6631973 100644
> > > --- a/lib/librte_mempool/rte_mempool.h
> > > +++ b/lib/librte_mempool/rte_mempool.h
> > > @@ -245,6 +245,7 @@ struct rte_mempool {
> > >        * this mempool.
> > >        */
> > >       int32_t ops_index;
> > > +     uint32_t elt_align;
> > >
> > >       struct rte_mempool_cache *local_cache; /**< Per-lcore local cache  
> > */  
> > >  
> >
> > It looks the patch will break the ABI (the mempool structure), so it
> > has to follow the abi deprecation process (= a notice in 17.05 and
> > the patch for 17.08).
> >
> > Could you give us some details about why you need such feature, how you
> > use it (since no API is added)?
> >
> > Thanks,
> > Olivier
> >