[dpdk-dev] [RFC v2 01/17] mempool: fix phys contig check if populate default skipped

[Olivier Matz]

On Thu, Feb 01, 2018 at 01:00:12PM +0300, Andrew Rybchenko wrote:
> On 02/01/2018 12:30 PM, santosh wrote:
> > On Thursday 01 February 2018 02:48 PM, Andrew Rybchenko wrote:
> > > On 02/01/2018 12:09 PM, santosh wrote:
> > > > On Thursday 01 February 2018 12:24 PM, Andrew Rybchenko wrote:
> > > > > On 02/01/2018 08:05 AM, santosh wrote:
> > > > > > On Wednesday 31 January 2018 10:15 PM, Olivier Matz wrote:
> > > > > > > On Tue, Jan 23, 2018 at 01:15:56PM +0000, Andrew Rybchenko wrote:
> > > > > > > > There is not specified dependency between rte_mempool_populate_default()
> > > > > > > > and rte_mempool_populate_iova(). So, the second should not rely on the
> > > > > > > > fact that the first adds capability flags to the mempool flags.
> > > > > > > > 
> > > > > > > > Fixes: 65cf769f5e6a ("mempool: detect physical contiguous objects")
> > > > > > > > Cc: stable at dpdk.org
> > > > > > > > 
> > > > > > > > Signed-off-by: Andrew Rybchenko <arybchenko at solarflare.com>
> > > > > > > Looks good to me. I agree it's strange that the mp->flags are
> > > > > > > updated with capabilities only in rte_mempool_populate_default().
> > > > > > > I see that this behavior is removed later in the patchset since the
> > > > > > > get_capa() is removed!
> > > > > > > 
> > > > > > > However maybe this single patch could go in 18.02.
> > > > > > > +Santosh +Jerin since it's mostly about Octeon.
> > > > > > rte_mempool_xmem_size should return correct size if MEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS flag
> > > > > > is set in 'mp->flags'. Thats why _ops_get_capabilities() called in _populate_default() but not
> > > > > > at _populate_iova().
> > > > > > I think, this 'alone' patch may break octeontx mempool.
> > > > > The patch does not touch rte_mempool_populate_default().
> > > > > _ops_get_capabilities() is still called there before
> > > > > rte_mempool_xmem_size(). The theoretical problem which
> > > > > the patch tries to fix is the case when
> > > > > rte_mempool_populate_default() is not called at all. I.e. application
> > > > > calls _ops_get_capabilities() to get flags, then, together with
> > > > > mp->flags, calls rte_mempool_xmem_size() directly, allocates
> > > > > calculated amount of memory and calls _populate_iova().
> > > > > 
> > > > In that case, Application does like below:
> > > > 
> > > >      /* Get mempool capabilities */
> > > >      mp_flags = 0;
> > > >      ret = rte_mempool_ops_get_capabilities(mp, &mp_flags);
> > > >      if ((ret < 0) && (ret != -ENOTSUP))
> > > >          return ret;
> > > > 
> > > >      /* update mempool capabilities */
> > > >      mp->flags |= mp_flags;
> > > Above line is not mandatory. "mp->flags | mp_flags" could be simply
> > > passed to  rte_mempool_xmem_size() below.
> > > 
> > That depends and again upto application requirement, if app further down
> > wants to refer mp->flags for _align/_contig then better update to mp->flags.
> > 
> > But that wasn't the point of discussion, I'm trying to understand that
> > w/o this patch, whats could be the application level problem?
> 
> The problem that it is fragile. If application does not use
> rte_mempool_populate_default() it has to care about addition
> of mempool capability flags into mempool flags. If it is not done,
> rte_mempool_populate_iova/virt/iova_tab() functions will work
> incorrectly since F_CAPA_PHYS_CONTIG and
> F_CAPA_BLK_ALIGNED_OBJECTS are missing.
> 
> The idea of the patch is to make it a bit more robust. I have no
> idea how it can break something. If capability flags are already
> there - no problem. If no, just make sure that we locally have them.

The example given by Santosh will work, but it is *not* the role of the
application to update the mempool flags. And nothing says that it is mandatory
to call rte_mempool_ops_get_capabilities() before the populate functions.

For instance, in testpmd it calls rte_mempool_populate_anon() when using
anonymous memory. The capabilities will never be updated in mp->flags.