DPDK patches and discussions
 help / color / mirror / Atom feed
From: Honnappa Nagarahalli <Honnappa.Nagarahalli@arm.com>
To: Olivier Matz <olivier.matz@6wind.com>
Cc: "dev@dpdk.org" <dev@dpdk.org>,
	"lucp.at.work@gmail.com" <lucp.at.work@gmail.com>,
	"david.marchand@redhat.com" <david.marchand@redhat.com>,
	"thomas@monjalon.net" <thomas@monjalon.net>,
	Ruifeng Wang <Ruifeng.Wang@arm.com>, nd <nd@arm.com>,
	nd <nd@arm.com>
Subject: Re: [dpdk-dev] [RFC v2] eal: simplify the implementation of rte_ctrl_thread_create
Date: Tue, 24 Aug 2021 20:03:03 +0000
Message-ID: <DBAPR08MB5814DFB414E5D78D57B323AB98C59@DBAPR08MB5814.eurprd08.prod.outlook.com> (raw)
In-Reply-To: <YSN1eveZBseV2t4J@platinum>

<snip>

> 
> Hi Honnappa,
> 
> On Mon, Aug 02, 2021 at 12:16:52AM -0500, Honnappa Nagarahalli wrote:
> > The current described behaviour of rte_ctrl_thread_create is rigid
> > which makes the implementation of the function complex.
> > The behavior is abstracted to allow for simplified implementation.
> 
> I agree that the behavior description should not reference the pthread
> functions, however I don't think the current description prevents from rewriting
> the code as you do.
Ok

> 
> I think it would be better to explain in commit log why the proposed code is
> simpler than the current one.
> 
> > Signed-off-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
> > ---
> > v2: Used compiler's C++11 atomic built-ins to access the shared variable.
> >
> >  lib/eal/common/eal_common_thread.c | 65 +++++++++++++-----------------
> >  lib/eal/include/rte_lcore.h        |  8 ++--
> >  2 files changed, 32 insertions(+), 41 deletions(-)
> >
> > diff --git a/lib/eal/common/eal_common_thread.c
> > b/lib/eal/common/eal_common_thread.c
> > index 1a52f42a2b..e3e0bf4bff 100644
> > --- a/lib/eal/common/eal_common_thread.c
> > +++ b/lib/eal/common/eal_common_thread.c
> > @@ -169,35 +169,35 @@ __rte_thread_uninit(void)  struct
> > rte_thread_ctrl_params {
> >  	void *(*start_routine)(void *);
> >  	void *arg;
> > -	pthread_barrier_t configured;
> > -	unsigned int refcnt;
> > +	int ret;
> > +	/* Synchronization variable between the control thread
> > +	 * and the thread calling rte_ctrl_thread_create function.
> > +	 * 0 - Initialized
> > +	 * 1 - Control thread is running successfully
> > +	 * 2 - Control thread encountered an error. 'ret' has the
> > +	 *     error code.
> > +	 */
> > +	unsigned int sync;
> 
> what do you think about using an enum or defines?
Will use defines

> 
> >  };
> >
> > -static void ctrl_params_free(struct rte_thread_ctrl_params *params)
> > -{
> > -	if (__atomic_sub_fetch(&params->refcnt, 1, __ATOMIC_ACQ_REL) ==
> 0) {
> > -		(void)pthread_barrier_destroy(&params->configured);
> > -		free(params);
> > -	}
> > -}
> > -
> >  static void *ctrl_thread_init(void *arg)  {
> >  	struct internal_config *internal_conf =
> >  		eal_get_internal_configuration();
> >  	rte_cpuset_t *cpuset = &internal_conf->ctrl_cpuset;
> >  	struct rte_thread_ctrl_params *params = arg;
> > -	void *(*start_routine)(void *);
> > +	void *(*start_routine)(void *) = params->start_routine;
> >  	void *routine_arg = params->arg;
> >
> >  	__rte_thread_init(rte_lcore_id(), cpuset);
> > -
> > -	pthread_barrier_wait(&params->configured);
> > -	start_routine = params->start_routine;
> > -	ctrl_params_free(params);
> > -
> > -	if (start_routine == NULL)
> > +	params->ret = pthread_setaffinity_np(pthread_self(),
> > +				sizeof(*cpuset), cpuset);
> > +	if (params->ret != 0) {
> > +		__atomic_store_n(&params->sync, 2, __ATOMIC_RELEASE);
> >  		return NULL;
> > +	}
> 
> Sorry if the question is stupid (I'm still not familiar with C++11 atomic built-ins),
> but do we have the assurance that params->ret is set in memory before
> params->sync is set? See below at [1].
Yes, the '__ATOMIC_RELEASE' store ensures that all prior memory operations are completed before 'sync' is updated.

> 
> > +
> > +	__atomic_store_n(&params->sync, 1, __ATOMIC_RELEASE);
> >
> >  	return start_routine(routine_arg);
> >  }
> > @@ -207,9 +207,6 @@ rte_ctrl_thread_create(pthread_t *thread, const char
> *name,
> >  		const pthread_attr_t *attr,
> >  		void *(*start_routine)(void *), void *arg)  {
> > -	struct internal_config *internal_conf =
> > -		eal_get_internal_configuration();
> > -	rte_cpuset_t *cpuset = &internal_conf->ctrl_cpuset;
> >  	struct rte_thread_ctrl_params *params;
> >  	int ret;
> >
> > @@ -219,15 +216,12 @@ rte_ctrl_thread_create(pthread_t *thread, const
> > char *name,
> >
> >  	params->start_routine = start_routine;
> >  	params->arg = arg;
> > -	params->refcnt = 2;
> > -
> > -	ret = pthread_barrier_init(&params->configured, NULL, 2);
> > -	if (ret != 0)
> > -		goto fail_no_barrier;
> > +	params->ret = 0;
> > +	params->sync = 0;
> >
> >  	ret = pthread_create(thread, attr, ctrl_thread_init, (void *)params);
> >  	if (ret != 0)
> > -		goto fail_with_barrier;
> > +		goto thread_create_failed;
> >
> >  	if (name != NULL) {
> >  		ret = rte_thread_setname(*thread, name); @@ -236,24
> +230,21 @@
> > rte_ctrl_thread_create(pthread_t *thread, const char *name,
> >  				"Cannot set name for ctrl thread\n");
> >  	}
> >
> > -	ret = pthread_setaffinity_np(*thread, sizeof(*cpuset), cpuset);
> > -	if (ret != 0)
> > -		params->start_routine = NULL;
> > +	/* Wait for the control thread to initialize successfully */
> > +	while (!__atomic_load_n(&params->sync, __ATOMIC_ACQUIRE))
> > +		rte_pause();
> 
> One difference between this implementation and the previous one is this busy
> loop. rte_pause() relaxes the cpu, but will not make the calling thread to sleep
> and wait for the sync event. So here we can spin a quite long time until the
> other thread is scheduled by the OS.
Yes, this is a difference. We could add a microsleep to allow for the OS to un-schedule the current thread.

> 
> > +	ret = params->ret;
> 
> [1]
> 
> Here, it is expected that when params->ret is seen as set before
> param->sync.
Yes, the __ATOMIC_ACQUIRE load ensures that the params->ret is loaded only after params->sync is loaded.

> 
> >
> > -	pthread_barrier_wait(&params->configured);
> > -	ctrl_params_free(params);
> > +	free(params);
> >
> > -	if (ret != 0)
> > -		/* start_routine has been set to NULL above; */
> > -		/* ctrl thread will exit immediately */
> > +	if (__atomic_load_n(&params->sync, __ATOMIC_ACQUIRE) != 1)
> 
> it suggest == instead of !=, like this:
> 
>   if (__atomic_load_n(&params->sync, __ATOMIC_ACQUIRE) ==
> CTRL_THREAD_ERR)
Ok

> 
> 
> > +		/* ctrl thread is exiting */
> >  		pthread_join(*thread, NULL);
> >
> >  	return -ret;
> >
> > -fail_with_barrier:
> > -	(void)pthread_barrier_destroy(&params->configured);
> > +thread_create_failed:
> >
> > -fail_no_barrier:
> >  	free(params);
> >
> >  	return -ret;
> > diff --git a/lib/eal/include/rte_lcore.h b/lib/eal/include/rte_lcore.h
> > index 1550b75da0..f1cc5e38dc 100644
> > --- a/lib/eal/include/rte_lcore.h
> > +++ b/lib/eal/include/rte_lcore.h
> > @@ -420,10 +420,10 @@ rte_thread_unregister(void);
> >  /**
> >   * Create a control thread.
> >   *
> > - * Wrapper to pthread_create(), pthread_setname_np() and
> > - * pthread_setaffinity_np(). The affinity of the new thread is based
> > - * on the CPU affinity retrieved at the time rte_eal_init() was
> > called,
> > - * the dataplane and service lcores are then excluded.
> > + * Creates a control thread with the given name and attributes. The
> > + * affinity of the new thread is based on the CPU affinity retrieved
> > + * at the time rte_eal_init() was called, the dataplane and service
> > + * lcores are then excluded.
> 
> The description is indeed better. Maybe it is the opportunity to highlight that if
> the name cannot be set, no error is returned, see commit 368a91d6bdc8 ("eal:
> ignore failure of naming a control thread")
Makes sense, will add.

> 
> >   *
> >   * @param thread
> >   *   Filled with the thread id of the new created thread.
> > --
> > 2.17.1
> >

  reply	other threads:[~2021-08-24 20:03 UTC|newest]

Thread overview: 30+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-07-30 21:37 [dpdk-dev] [RFC] " Honnappa Nagarahalli
2021-08-02  5:16 ` [dpdk-dev] [RFC v2] " Honnappa Nagarahalli
2021-08-23 10:16   ` Olivier Matz
2021-08-24 20:03     ` Honnappa Nagarahalli [this message]
2021-08-24 21:30       ` Stephen Hemminger
2021-08-25  1:26         ` Honnappa Nagarahalli
2021-08-25 15:19         ` Mattias Rönnblom
2021-08-25 15:31           ` Honnappa Nagarahalli
2021-09-01 20:04             ` Mattias Rönnblom
2021-09-01 22:21               ` Honnappa Nagarahalli
2021-08-30 22:30         ` Honnappa Nagarahalli
2021-08-25 15:12   ` Mattias Rönnblom
2021-08-25 15:23     ` Honnappa Nagarahalli
2021-08-25 21:57       ` Stephen Hemminger
2021-08-25 18:48 ` [dpdk-dev] [RFC] " Luc Pelletier
2021-08-25 19:26   ` Honnappa Nagarahalli
2021-08-30 16:34 ` [dpdk-dev] [PATCH v3 1/2] " Honnappa Nagarahalli
2021-08-30 16:34   ` [dpdk-dev] [PATCH v3 2/2] test/eal: add a test for rte_ctrl_thread_create Honnappa Nagarahalli
2021-08-30 16:41     ` Stephen Hemminger
2021-08-30 16:42       ` Honnappa Nagarahalli
2021-08-30 16:44   ` [dpdk-dev] [PATCH v3 1/2] eal: simplify the implementation of rte_ctrl_thread_create Stephen Hemminger
2021-08-30 22:12     ` Honnappa Nagarahalli
2021-10-13 20:18 ` [dpdk-dev] [PATCH v4 " Honnappa Nagarahalli
2021-10-13 20:18   ` [dpdk-dev] [PATCH v4 2/2] test/eal: add a test for rte_ctrl_thread_create Honnappa Nagarahalli
2021-10-21 15:46   ` [dpdk-dev] [PATCH v4 1/2] eal: simplify the implementation of rte_ctrl_thread_create Olivier Matz
2021-10-21 20:49     ` Honnappa Nagarahalli
2021-10-21 21:32 ` [dpdk-dev] [PATCH v5 " Honnappa Nagarahalli
2021-10-21 21:32   ` [dpdk-dev] [PATCH v5 2/2] test/eal: add a test for rte_ctrl_thread_create Honnappa Nagarahalli
2021-10-22  8:35     ` Olivier Matz
2021-10-25 19:46       ` David Marchand

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=DBAPR08MB5814DFB414E5D78D57B323AB98C59@DBAPR08MB5814.eurprd08.prod.outlook.com \
    --to=honnappa.nagarahalli@arm.com \
    --cc=Ruifeng.Wang@arm.com \
    --cc=david.marchand@redhat.com \
    --cc=dev@dpdk.org \
    --cc=lucp.at.work@gmail.com \
    --cc=nd@arm.com \
    --cc=olivier.matz@6wind.com \
    --cc=thomas@monjalon.net \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

DPDK patches and discussions

This inbox may be cloned and mirrored by anyone:

	git clone --mirror http://inbox.dpdk.org/dev/0 dev/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 dev dev/ http://inbox.dpdk.org/dev \
		dev@dpdk.org
	public-inbox-index dev

Example config snippet for mirrors.
Newsgroup available over NNTP:
	nntp://inbox.dpdk.org/inbox.dpdk.dev


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git