Re: [dpdk-dev] [PATCH v4 6/9] eal: register non-EAL threads as lcores

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

On Wed, Jul 01, 2020 at 09:13:36AM +0200, David Marchand wrote:
> On Tue, Jun 30, 2020 at 12:07 PM Olivier Matz <olivier.matz@6wind.com> wrote:
> >
> > On Fri, Jun 26, 2020 at 04:47:33PM +0200, David Marchand wrote:
> > > DPDK allows calling some part of its API from a non-EAL thread but this
> > > has some limitations.
> > > OVS (and other applications) has its own thread management but still
> > > want to avoid such limitations by hacking RTE_PER_LCORE(_lcore_id) and
> > > faking EAL threads potentially unknown of some DPDK component.
> > >
> > > Introduce a new API to register non-EAL thread and associate them to a
> > > free lcore with a new NON_EAL role.
> > > This role denotes lcores that do not run DPDK mainloop and as such
> > > prevents use of rte_eal_wait_lcore() and consorts.
> > >
> > > Multiprocess is not supported as the need for cohabitation with this new
> > > feature is unclear at the moment.
> > >
> > > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > > Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
> > > ---
> > > Changes since v2:
> > > - refused multiprocess init once rte_thread_register got called, and
> > >   vice versa,
> > > - added warning on multiprocess in rte_thread_register doxygen,
> > >
> > > Changes since v1:
> > > - moved cleanup on lcore role code in patch 5,
> > > - added unit test,
> > > - updated documentation,
> > > - changed naming from "external thread" to "registered non-EAL thread"
> > >
> > > ---
> > >  MAINTAINERS                                   |   1 +
> > >  app/test/Makefile                             |   1 +
> > >  app/test/autotest_data.py                     |   6 +
> > >  app/test/meson.build                          |   2 +
> > >  app/test/test_lcores.c                        | 139 ++++++++++++++++++
> > >  doc/guides/howto/debug_troubleshoot.rst       |   5 +-
> > >  .../prog_guide/env_abstraction_layer.rst      |  22 +--
> > >  doc/guides/prog_guide/mempool_lib.rst         |   2 +-
> > >  lib/librte_eal/common/eal_common_lcore.c      |  50 ++++++-
> > >  lib/librte_eal/common/eal_common_mcfg.c       |  36 +++++
> > >  lib/librte_eal/common/eal_common_thread.c     |  33 +++++
> > >  lib/librte_eal/common/eal_memcfg.h            |  10 ++
> > >  lib/librte_eal/common/eal_private.h           |  18 +++
> > >  lib/librte_eal/freebsd/eal.c                  |   4 +
> > >  lib/librte_eal/include/rte_lcore.h            |  25 +++-
> > >  lib/librte_eal/linux/eal.c                    |   4 +
> > >  lib/librte_eal/rte_eal_version.map            |   2 +
> > >  lib/librte_mempool/rte_mempool.h              |  11 +-
> > >  18 files changed, 349 insertions(+), 22 deletions(-)
> > >  create mode 100644 app/test/test_lcores.c
> > >
> >
> > [...]
> >
> > > diff --git a/app/test/test_lcores.c b/app/test/test_lcores.c
> > > new file mode 100644
> > > index 0000000000..864bcbade7
> > > --- /dev/null
> > > +++ b/app/test/test_lcores.c
> > > @@ -0,0 +1,139 @@
> > > +/* SPDX-License-Identifier: BSD-3-Clause
> > > + * Copyright (c) 2020 Red Hat, Inc.
> > > + */
> > > +
> > > +#include <pthread.h>
> > > +#include <string.h>
> > > +
> > > +#include <rte_lcore.h>
> > > +
> > > +#include "test.h"
> > > +
> > > +struct thread_context {
> > > +     enum { INIT, ERROR, DONE } state;
> > > +     bool lcore_id_any;
> > > +     pthread_t id;
> > > +     unsigned int *registered_count;
> > > +};
> > > +static void *thread_loop(void *arg)
> > > +{
> >
> > missing an empty line here
> >
> > > +     struct thread_context *t = arg;
> > > +     unsigned int lcore_id;
> > > +
> > > +     lcore_id = rte_lcore_id();
> > > +     if (lcore_id != LCORE_ID_ANY) {
> > > +             printf("Incorrect lcore id for new thread %u\n", lcore_id);
> > > +             t->state = ERROR;
> > > +     }
> > > +     rte_thread_register();
> > > +     lcore_id = rte_lcore_id();
> > > +     if ((t->lcore_id_any && lcore_id != LCORE_ID_ANY) ||
> > > +                     (!t->lcore_id_any && lcore_id == LCORE_ID_ANY)) {
> > > +             printf("Could not register new thread, got %u while %sexpecting %u\n",
> > > +                     lcore_id, t->lcore_id_any ? "" : "not ", LCORE_ID_ANY);
> > > +             t->state = ERROR;
> > > +     }
> >
> > To check if rte_thread_register() succedeed, we need to look at
> > lcore_id. I wonder if rte_thread_register() shouldn't return the lcore
> > id on success, and -1 on error (rte_errno could be set to give some
> > info on the error).
> 
> lcore_id are unsigned integers with the special value LCORE_ID_ANY
> mapped to UINT32_MAX (should be UINT_MAX? anyway...).
> 
> rte_thread_register could return an error code as there are no ERROR
> level logs about why a lcore allocation failed.
> We could then distinguish a shortage of lcore (or init callback
> refusal) from an invalid call before rte_eal_init() or when mp is in
> use.
> 
> About returning the lcore_id as part of the return code, this would
> map to -1 for LCORE_ID_ANY.
> This is probably not a problem but still odd.

Yes, it would be a bit odd like this. What about changing the definition of
LCORE_ID_ANY to ((unsigned int)-1) ? I think it does not change the effective
value on any architecture, but would make the above change clearer.


> >
> > The same could be done for rte_thread_init()
> 
> ?
> Not sure where this one could fail.

I was thinking about __rte_trace_mem_per_thread_alloc(), but maybe it's not
needed.

> >
> > [...]
> >
> > > diff --git a/lib/librte_eal/common/eal_common_thread.c b/lib/librte_eal/common/eal_common_thread.c
> > > index a7ae0691bf..1cbddc4b5b 100644
> > > --- a/lib/librte_eal/common/eal_common_thread.c
> > > +++ b/lib/librte_eal/common/eal_common_thread.c
> > > @@ -236,3 +236,36 @@ rte_ctrl_thread_create(pthread_t *thread, const char *name,
> > >       pthread_join(*thread, NULL);
> > >       return -ret;
> > >  }
> > > +
> > > +void
> > > +rte_thread_register(void)
> > > +{
> > > +     unsigned int lcore_id;
> > > +     rte_cpuset_t cpuset;
> > > +
> > > +     /* EAL init flushes all lcores, we can't register before. */
> > > +     assert(internal_config.init_complete == 1);
> > > +     if (pthread_getaffinity_np(pthread_self(), sizeof(cpuset),
> > > +                     &cpuset) != 0)
> > > +             CPU_ZERO(&cpuset);
> > > +     lcore_id = eal_lcore_non_eal_allocate();
> > > +     if (lcore_id >= RTE_MAX_LCORE)
> > > +             lcore_id = LCORE_ID_ANY;
> > > +     rte_thread_init(lcore_id, &cpuset);
> > > +     if (lcore_id != LCORE_ID_ANY)
> > > +             RTE_LOG(DEBUG, EAL, "Registered non-EAL thread as lcore %u.\n",
> > > +                     lcore_id);
> > > +}
> >
> > So, in this case, the affinity of the pthread is kept and saved, in other
> > words there is no link between the lcore id and the affinity. It means we
> > are allowing an application to register lcores for dataplane with conflicting
> > affinities.
> 
> This is not something new, applications using --lcores option already
> live with this.
> We have warnings in the documentation about non-EAL threads and about
> the dangers of conflicting affinities.
> Hopefully, the users of this API know what they are doing since they
> chose not to use EAL threads.
> 
> 
> >
> > I wonder if it could be useful to have an API that automatically sets
> > the affinity according to the lcore_id. Or a function that creates a
> > pthread using the specified lcore id, and setting the correct affinity.
> > I could simplify the work for applications that want to create/destroy
> > dataplane threads dynamically.
> 
> Do you mean EAL threads dynamic creation/suppression?
> 
> 
> >
> > This could be done later however, just an idea.
> 
> For now, I don't see the need.
> 
> 
> >
> > [...]
> > > diff --git a/lib/librte_eal/freebsd/eal.c b/lib/librte_eal/freebsd/eal.c
> > > index 13e5de006f..32a3d999b8 100644
> > > --- a/lib/librte_eal/freebsd/eal.c
> > > +++ b/lib/librte_eal/freebsd/eal.c
> > > @@ -424,6 +424,10 @@ rte_config_init(void)
> > >               }
> > >               if (rte_eal_config_reattach() < 0)
> > >                       return -1;
> > > +             if (!eal_mcfg_enable_multiprocess()) {
> > > +                     RTE_LOG(ERR, EAL, "Primary process refused secondary attachment\n");
> > > +                     return -1;
> > > +             }
> > >               eal_mcfg_update_internal();
> > >               break;
> > >       case RTE_PROC_AUTO:
> > > diff --git a/lib/librte_eal/include/rte_lcore.h b/lib/librte_eal/include/rte_lcore.h
> > > index 3968c40693..43747e88df 100644
> > > --- a/lib/librte_eal/include/rte_lcore.h
> > > +++ b/lib/librte_eal/include/rte_lcore.h
> > > @@ -31,6 +31,7 @@ enum rte_lcore_role_t {
> > >       ROLE_RTE,
> > >       ROLE_OFF,
> > >       ROLE_SERVICE,
> > > +     ROLE_NON_EAL,
> > >  };
> >
> > If find the name ROLE_NON_EAL a bit heavy (this was also my impression
> > when reading the doc part).
> >
> > I understand that there are several types of threads:
> >
> > - eal (pthread created by eal): ROLE_RTE and ROLE_SERVICE
> > - unregistered (pthread not created by eal, and not registered): ROLE_OFF
> >   (note that ROLE_OFF also applies for unexistant threads)
> > - dynamic: pthread not created by eal, but registered
> 
> Last two cases both are non-EAL threads as described in the doc so far.

Yes, but only the last case has the NON_EAL role.
I feel that currently role != thread_type, so may be a bit confusing to
use a pthread_type name in the role enum.

> 
> 
> >
> > What about using ROLE_DYN ? I'm not sure about this name either, it's just
> > to open the discussion :)
> >
> 
> Well, at the moment, all those new lcores are mapped only to non-EAL threads.
> A dynamic role feels like you want to take dynamic EAL threads into
> account from the start.
> I prefer to stick to non-EAL.
> 
> 
> -- 
> David Marchand
>