Re: [dpdk-dev] [PATCH v4 08/17] eal: apply affinity of EAL thread by assigned cpuset

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

Hi,

On 02/09/2015 02:48 PM, Liang, Cunming wrote:
>> -----Original Message-----
>> From: Olivier MATZ [mailto:olivier.matz@6wind.com]
>> Sent: Monday, February 09, 2015 4:01 AM
>> To: Liang, Cunming; dev@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH v4 08/17] eal: apply affinity of EAL thread by
>> assigned cpuset
>>
>> Hi,
>>
>> On 02/02/2015 03:02 AM, Cunming Liang wrote:
>>> EAL threads use assigned cpuset to set core affinity during startup.
>>> It keeps 1:1 mapping, if no '--lcores' option is used.
>>>
>>> [...]
>>>
>>>  lib/librte_eal/bsdapp/eal/eal.c          | 13 ++++---
>>>  lib/librte_eal/bsdapp/eal/eal_thread.c   | 63 +++++++++---------------------
>>>  lib/librte_eal/linuxapp/eal/eal.c        |  7 +++-
>>>  lib/librte_eal/linuxapp/eal/eal_thread.c | 67 +++++++++++---------------------
>>>  4 files changed, 54 insertions(+), 96 deletions(-)
>>>
>>> diff --git a/lib/librte_eal/bsdapp/eal/eal.c b/lib/librte_eal/bsdapp/eal/eal.c
>>> index 69f3c03..98c5a83 100644
>>> --- a/lib/librte_eal/bsdapp/eal/eal.c
>>> +++ b/lib/librte_eal/bsdapp/eal/eal.c
>>> @@ -432,6 +432,7 @@ rte_eal_init(int argc, char **argv)
>>>  	int i, fctret, ret;
>>>  	pthread_t thread_id;
>>>  	static rte_atomic32_t run_once = RTE_ATOMIC32_INIT(0);
>>> +	char cpuset[CPU_STR_LEN];
>>>
>>>  	if (!rte_atomic32_test_and_set(&run_once))
>>>  		return -1;
>>> @@ -502,13 +503,17 @@ rte_eal_init(int argc, char **argv)
>>>  	if (rte_eal_pci_init() < 0)
>>>  		rte_panic("Cannot init PCI\n");
>>>
>>> -	RTE_LOG(DEBUG, EAL, "Master core %u is ready (tid=%p)\n",
>>> -		rte_config.master_lcore, thread_id);
>>> -
>>>  	eal_check_mem_on_local_socket();
>>>
>>>  	rte_eal_mcfg_complete();
>>>
>>> +	eal_thread_init_master(rte_config.master_lcore);
>>> +
>>> +	eal_thread_dump_affinity(cpuset, CPU_STR_LEN);
>>> +
>>> +	RTE_LOG(DEBUG, EAL, "Master lcore %u is ready (tid=%p;cpuset=[%s])\n",
>>> +		rte_config.master_lcore, thread_id, cpuset);
>>> +
>>>  	if (rte_eal_dev_init() < 0)
>>>  		rte_panic("Cannot init pmd devices\n");
>>>
>>> @@ -532,8 +537,6 @@ rte_eal_init(int argc, char **argv)
>>>  			rte_panic("Cannot create thread\n");
>>>  	}
>>>
>>> -	eal_thread_init_master(rte_config.master_lcore);
>>> -
>>>  	/*
>>>  	 * Launch a dummy function on all slave lcores, so that master lcore
>>>  	 * knows they are all ready when this function returns.
>>
>> I wonder if changing this may have an impact on third-party drivers
>> that already use a management thread. Before the patch, the init()
>> function of the external library was called with default affinities,
>> and now it's called with the affinity from master lcore.
>>
>> I think it should at least be noticed in the commit log.
>>
>> Why are you doing this change? (I don't say it's a bad change, but
>> I don't understand why you are doing it here)
> [LCM] To be honest, the main purpose is I don't found any reason to have linuxapp and freebsdapp in different init sequence.
> I means in linux it init_master before dev_init(), but in freebsd it reverse.


I agree that's something we should fix.


> And as the default value of TLS already changes, if dev_init() first and using those TLS, the result will be not in an EAL thread.
> But actually they're in the EAL master thread. So I prefer to do the change follows linuxapp sequence.

That makes sense. Is it possible to have this reordering in a separate
patch? The title could be
"eal: standardize init sequence between linux and bsd"



>>
>>
>>> diff --git a/lib/librte_eal/bsdapp/eal/eal_thread.c
>> b/lib/librte_eal/bsdapp/eal/eal_thread.c
>>> index d0c077b..5b16302 100644
>>> --- a/lib/librte_eal/bsdapp/eal/eal_thread.c
>>> +++ b/lib/librte_eal/bsdapp/eal/eal_thread.c
>>> @@ -103,55 +103,27 @@ eal_thread_set_affinity(void)
>>>  {
>>>  	int s;
>>>  	pthread_t thread;
>>> -
>>> -/*
>>> - * According to the section VERSIONS of the CPU_ALLOC man page:
>>> - *
>>> - * The CPU_ZERO(), CPU_SET(), CPU_CLR(), and CPU_ISSET() macros were
>> added
>>> - * in glibc 2.3.3.
>>> - *
>>> - * CPU_COUNT() first appeared in glibc 2.6.
>>> - *
>>> - * CPU_AND(),     CPU_OR(),     CPU_XOR(),    CPU_EQUAL(),
>> CPU_ALLOC(),
>>> - * CPU_ALLOC_SIZE(), CPU_FREE(), CPU_ZERO_S(),  CPU_SET_S(),
>> CPU_CLR_S(),
>>> - * CPU_ISSET_S(),  CPU_AND_S(), CPU_OR_S(), CPU_XOR_S(), and
>> CPU_EQUAL_S()
>>> - * first appeared in glibc 2.7.
>>> - */
>>> -#if defined(CPU_ALLOC)
>>> -	size_t size;
>>> -	cpu_set_t *cpusetp;
>>> -
>>> -	cpusetp = CPU_ALLOC(RTE_MAX_LCORE);
>>> -	if (cpusetp == NULL) {
>>> -		RTE_LOG(ERR, EAL, "CPU_ALLOC failed\n");
>>> -		return -1;
>>> -	}
>>> -
>>> -	size = CPU_ALLOC_SIZE(RTE_MAX_LCORE);
>>> -	CPU_ZERO_S(size, cpusetp);
>>> -	CPU_SET_S(rte_lcore_id(), size, cpusetp);
>>> +	unsigned lcore_id = rte_lcore_id();
>>>
>>>  	thread = pthread_self();
>>> -	s = pthread_setaffinity_np(thread, size, cpusetp);
>>> +	s = pthread_setaffinity_np(thread, sizeof(cpuset_t),
>>> +				   &lcore_config[lcore_id].cpuset);
>>>  	if (s != 0) {
>>>  		RTE_LOG(ERR, EAL, "pthread_setaffinity_np failed\n");
>>> -		CPU_FREE(cpusetp);
>>>  		return -1;
>>>  	}
>>>
>>> -	CPU_FREE(cpusetp);
>>> -#else /* CPU_ALLOC */
>>> -	cpuset_t cpuset;
>>> -	CPU_ZERO( &cpuset );
>>> -	CPU_SET( rte_lcore_id(), &cpuset );
>>> +	/* acquire system unique id  */
>>> +	rte_gettid();
>>
>> As suggested in the previous patch, I think having rte_init_tid() would
>> be clearer here.
> [LCM] Sorry, I didn't get your [PATCH v4 07/17] comments, probably the mailbox issue.
> Do you suggest to have a rte_init_tid() but not do syscall on the first time ?
> Any benefit, rte_gettid() looks like more simple and straight forward. 

I think the mail was properly sent, you can see it here:
http://dpdk.org/ml/archives/dev/2015-February/012556.html

Usually, "get" functions return a value and have no side effects.
"init" functions return nothing (or an error code) but have a
side effect which is to initialize an internal state.


>>> +
>>> +	/* store socket_id in TLS for quick access */
>>> +	RTE_PER_LCORE(_socket_id) =
>>> +		eal_cpuset_socket_id(&lcore_config[lcore_id].cpuset);
>>> +
>>> +	CPU_COPY(&lcore_config[lcore_id].cpuset, &RTE_PER_LCORE(_cpuset));
>>> +
>>> +	lcore_config[lcore_id].socket_id = RTE_PER_LCORE(_socket_id);
>>>
>>> -	thread = pthread_self();
>>> -	s = pthread_setaffinity_np(thread, sizeof( cpuset ), &cpuset);
>>> -	if (s != 0) {
>>> -		RTE_LOG(ERR, EAL, "pthread_setaffinity_np failed\n");
>>> -		return -1;
>>> -	}
>>> -#endif
>>
>> You are removing a lot of code that was using CPU_ALLOC().
>> Are we sure that the cpuset_t type is large enough to store all the
>> CPUs?
>>
>> It looks the current value of CPU_SETSIZE is 1024 now, but I wonder
>> if this code was written when this value was lower. Could you check if
>> it can happen today (maybe with an old libc)? A problem can occur if
>> the size of cpuset_t is lower that the size of RTE_MAX_LCORE.
> [LCM] I found actually the MACRO is not just for support CPU_ALLOC(), but for linux or freebsd.
> In freebsdapp, there's no CPU_ALLOC defined, it use fixed width *cpuset_t*.
> In linuxapp, there's CPU_ALLOC defined, it use cpu_set_t* and dynamic CPU_ALLOC(RTE_MAX_LCORE).
> But actually RTE_MAX_LCORE < 1024(sizeof(cpu_set_t)). 
> After using rte_cpuset_t, there's no additional reason to use CPU_ALLOC only for linuxapp and choose a small but dynamic width.

I did a quick search on google, and it seems CPU_SETSIZE is 1024
for a long time. So you are right, there is probably no reason to
keep CPU_ALLOC(). As I said in a previous mail, it could be useful
in the future when the number of CPUs will reach 1024, but we have
some time to handle this.