Re: [dpdk-dev] [RFC PATCH 0/7] support multi-phtread per lcore

["Liang, Cunming" <cunming.liang@intel.com>]

> -----Original Message-----
> From: Ananyev, Konstantin
> Sent: Friday, January 09, 2015 1:06 AM
> To: Liang, Cunming; Stephen Hemminger; Richardson, Bruce
> Cc: dev@dpdk.org
> Subject: RE: [dpdk-dev] [RFC PATCH 0/7] support multi-phtread per lcore
> 
> 
> Hi Steve,
> 
> > -----Original Message-----
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Liang, Cunming
> > Sent: Tuesday, December 23, 2014 9:52 AM
> > To: Stephen Hemminger; Richardson, Bruce
> > Cc: dev@dpdk.org
> > Subject: Re: [dpdk-dev] [RFC PATCH 0/7] support multi-phtread per lcore
> >
> >
> >
> > > -----Original Message-----
> > > From: Stephen Hemminger [mailto:stephen@networkplumber.org]
> > > Sent: Tuesday, December 23, 2014 2:29 AM
> > > To: Richardson, Bruce
> > > Cc: Liang, Cunming; dev@dpdk.org
> > > Subject: Re: [dpdk-dev] [RFC PATCH 0/7] support multi-phtread per lcore
> > >
> > > On Mon, 22 Dec 2014 09:46:03 +0000
> > > Bruce Richardson <bruce.richardson@intel.com> wrote:
> > >
> > > > On Mon, Dec 22, 2014 at 01:51:27AM +0000, Liang, Cunming wrote:
> > > > > ...
> > > > > > I'm conflicted on this one. However, I think far more applications would
> be
> > > > > > broken
> > > > > > to start having to use thread_id in place of an lcore_id than would be
> > > broken
> > > > > > by having the lcore_id no longer actually correspond to a core.
> > > > > > I'm actually struggling to come up with a large number of scenarios
> where
> > > it's
> > > > > > important to an app to determine the cpu it's running on, compared to
> the
> > > large
> > > > > > number of cases where you need to have a data-structure per thread.
> In
> > > DPDK
> > > > > > libs
> > > > > > alone, you see this assumption that lcore_id == thread_id a large
> number
> > > of
> > > > > > times.
> > > > > >
> > > > > > Despite the slight logical inconsistency, I think it's better to avoid
> > > introducing
> > > > > > a thread-id and continue having lcore_id representing a unique thread.
> > > > > >
> > > > > > /Bruce
> > > > >
> > > > > Ok, I understand it.
> > > > > I list the implicit meaning if using lcore_id representing the unique thread.
> > > > > 1). When lcore_id less than RTE_MAX_LCORE, it still represents the logical
> > > core id.
> > > > > 2). When lcore_id large equal than RTE_MAX_LCORE, it represents an
> unique
> > > id for thread.
> > > > > 3). Most of APIs(except rte_lcore_id()) in rte_lcore.h suggest to be used
> only
> > > in CASE 1)
> > > > > 4). rte_lcore_id() can be used in CASE 2), but the return value no matter
> > > represent a logical core id.
> > > > >
> > > > > If most of us feel it's acceptable, I'll prepare for the RFC v2 base on this
> > > conclusion.
> > > > >
> > > > > /Cunming
> > > >
> > > > Sorry, I don't like that suggestion either, as having lcore_id values greater
> > > > than RTE_MAX_LCORE is terrible, as how will people know how to
> dimension
> > > arrays
> > > > to be indexes by lcore id? Given the choice, if we are not going to just use
> > > > lcore_id as a generic thread id, which is always between 0 and
> > > RTE_MAX_LCORE
> > > > we can look to define a new thread_id variable to hold that. However, it
> should
> > > > have a bounded range.
> > > > From an ease-of-porting perspective, I still think that the simplest option is
> to
> > > > use the existing lcore_id and accept the fact that it's now a thread id rather
> > > > than an actual physical lcore. Question is, is would that cause us lots of
> issues
> > > > in the future?
> > > >
> > > > /Bruce
> > >
> > > The current rte_lcore_id() has different meaning the thread. Your proposal
> will
> > > break code that uses lcore_id to do per-cpu statistics and the lcore_config
> > > code in the samples.
> > > q
> > [Liang, Cunming] +1.
> 
> Few more thoughts on that subject:
> 
> Actually one more place in the lib, where lcore_id is used (and it should be
> unique):
> rte_spinlock_recursive_lock() / rte_spinlock_recursive_trylock().
> So if we going to replace lcore_id with thread_id as uniques thread index, then
> these functions
> have to be updated too.
[Liang, Cunming] You're right, if deciding to use thread_id, we have to check and replace 
rte_lcore_id()/RTE_PER_LCORE(_lcore_id) on all the impact place.
Now I'm buying the proposal to keep using rte_lcore_id() to return the 
unique id. Meanwhile I think it's necessary to have real cpu id.
It's helpful in NUMA socket checking. 
I will provide new API rte_curr_cpu() to return the runtime cpu no matter 
the thread running in coremasked or non-coremasked cpu.
So the socket info stored in lcore_config still useful  to choose the local socket.
> 
> About maintaining our own unique thread_id inside shared memory
> (_get_linear_tid()/_put_linear_tid()).
> There is one thing that worries me with that approach:
> In case of abnormal process termination, TIDs used by that process will remain
> 'reserved'
> and there is no way to know which TIDs were used by terminated process.
> So there could be a situation with DPDK multi-process model,
> when after secondary process abnormal termination, It wouldn't be possible to
> restart it -
> we just run out of 'free' TIDs.
[Liang, Cunming] That's a good point I think. I think it's not only for thread id but 
for all the dynamic allocated resource (e.g. memzone, mempool).
we haven't a garbage collection or heartbeat to process the secondary abnormal exit.

> 
> Which makes me think probably there is no need to introduce new globally
> unique 'thread_id'?
> Might be just lcore_id is enough?
> As Mirek and Bruce suggested we can treat it a sort of 'unique thread id' inside
> EAL.
[Liang, Cunming] I think we'd better have two, one for 'unique thread id', one for real cpu id.
No matter which of them are named lcore_id/thread_id/cpu_id and etc.
For cpu id, we need to check/get the NUMA info.
Pthread may migrate from one core to another, the thread 'socket id' may change, 
The per cpu socket info we have them in lcore_config.

> Or as 'virtual' core id that can run on set of physical cpus, and these subsets for
> different 'virtual' cores can intersect.
> Then basically we can keep legacy behaviour with '-c <lcores_mask>,' where each
> lcore_id matches one to one  with physical cpu, and introduce new one,
> something like:
> --
> lcores='(<lcore_set1>)=(<phys_cpu_set1>),..(<lcore_setN)=(<phys_cpu_setN>)'.
> So let say: --lcores=(0-7)=(0,2-4),(10)=(7),(8)=(all)' would mean:
> Create 10 EAL threads, bind threads with clore_id=[0-7] to cpuset: <0,2,3,4>,
> thread  with lcore_id=10 is binded to  cpu 7, and allow to run lcore_id=8 on any
> cpu in the system.
> Of course '-c' and '-lcores' would be mutually exclusive, and we will need to
> update  rte_lcore_to_socket_id()
> and introduce: rte_lcore_(set|get)_affinity().
> 
> Does it make sense to you?
[Liang, Cunming] If assign lcore_id during the command line, user have to handle 
the conflict for '-c' and '--lcores'. 
In this cases, if lcore_id 0~10 is occupied, the coremasked thread start from 11 ?
In case, application create a new pthread during the runtime.
As there's no lcore id belongs to the new thread mentioned in the command line, it then still back to dynamic allocate.
I means on the startup, user may have no idea of how much pthread they will run.

'rte_pthread_assign_lcore' do the things as 'rte_lcore_(set|get)_affinity()'
If we keeping using lcore_id, I like the name you proposed.

I'll send my code update on next Monday.

> 
> BTW, one more thing: while we are on it  - it is probably a good time to do
> something with our interrupt thread?
> It is a bit strange that we can't use rte_pktmbuf_free() or
> rte_spinlock_recursive_lock() from our own interrupt/alarm handlers
> 
> Konstantin