DPDK patches and discussions
 help / color / mirror / Atom feed
From: Matteo Lanzuisi <m.lanzuisi@resi.it>
To: "Wiles, Keith" <keith.wiles@intel.com>
Cc: Olivier Matz <olivier.matz@6wind.com>, "dev@dpdk.org" <dev@dpdk.org>
Subject: Re: [dpdk-dev] Multi-thread mempool usage
Date: Fri, 24 Aug 2018 16:44:34 +0200	[thread overview]
Message-ID: <2d1b12b7-62c9-bfd8-d899-db9990f4f44d@resi.it> (raw)
In-Reply-To: <9970633e-a992-0bb0-31a7-72d78325de0e@resi.it>

Hi,

I used valgrind again for a very long time, and it told me nothing 
strange is happening on my code.
After it, I changed my code this way

  unsigned            lcore_id_start = rte_lcore_id();
RTE_LCORE_FOREACH(lcore_id)
{
         if (lcore_id_start != lcore_id) // <--------- before this 
change, every lcore could use it own mempool and enqueue to its own ring
         {
                 new_work = NULL;
                 result = 
rte_mempool_get(cea_main_lcore_conf[lcore_id].de_conf.cmd_pool, (VOID_P 
*) &new_work);        // mempools are created one for each logical core
                 if (result == 0)
                {
                     if (((uint64_t)(new_work)) < 0x7f0000000000)
                         printf("Result %d, lcore di partenza %u, lcore 
di ricezione %u, pointer %p\n", result, rte_lcore_id(), lcore_id, 
new_work);    // debug print, on my server it should never happen but 
with multi-thread happens always on the last logical core!!!!
                     new_work->command = command; // usage of the memory 
gotten from the mempool... <<<<<- here is where the application crashes!!!!
                     result = 
rte_ring_enqueue(cea_main_lcore_conf[lcore_id].de_conf.cmd_ring, 
(VOID_P) new_work);    // enqueues the gotten buffer on the rings of all 
lcores
                     // check on result value ...
                 }
                 else
                 {
                     // do something if result != 0 ...
                 }
         }
         else
         {
               // don't use mempool but call a function instead ....
         }
}

and now it all goes well.
It is possibile that sending to itself could generate this issue?

Regards,
Matteo

Il 21/08/2018 16:46, Matteo Lanzuisi ha scritto:
> Il 21/08/2018 14:51, Wiles, Keith ha scritto:
>>
>>> On Aug 21, 2018, at 7:44 AM, Matteo Lanzuisi <m.lanzuisi@resi.it> 
>>> wrote:
>>>
>>> Il 21/08/2018 14:17, Wiles, Keith ha scritto:
>>>>> On Aug 21, 2018, at 7:01 AM, Matteo Lanzuisi <m.lanzuisi@resi.it> 
>>>>> wrote:
>>>>>
>>>>> Hi
>>>>>
>>>>> Il 20/08/2018 18:03, Wiles, Keith ha scritto:
>>>>>>> On Aug 20, 2018, at 9:47 AM, Matteo Lanzuisi <m.lanzuisi@resi.it>
>>>>>>>   wrote:
>>>>>>>
>>>>>>> Hello Olivier,
>>>>>>>
>>>>>>> Il 13/08/2018 23:54, Olivier Matz ha scritto:
>>>>>>>
>>>>>>>> Hello Matteo,
>>>>>>>>
>>>>>>>> On Mon, Aug 13, 2018 at 03:20:44PM +0200, Matteo Lanzuisi wrote:
>>>>>>>>
>>>>>>>>> Any suggestion? any idea about this behaviour?
>>>>>>>>>
>>>>>>>>> Il 08/08/2018 11:56, Matteo Lanzuisi ha scritto:
>>>>>>>>>
>>>>>>>>>> Hi all,
>>>>>>>>>>
>>>>>>>>>> recently I began using "dpdk-17.11-11.el7.x86_64" rpm (RedHat 
>>>>>>>>>> rpm) on
>>>>>>>>>> RedHat 7.5 kernel 3.10.0-862.6.3.el7.x86_64 as a porting of an
>>>>>>>>>> application from RH6 to RH7. On RH6 I used dpdk-2.2.0.
>>>>>>>>>>
>>>>>>>>>> This application is made up by one or more threads (each one 
>>>>>>>>>> on a
>>>>>>>>>> different logical core) reading packets from i40e interfaces.
>>>>>>>>>>
>>>>>>>>>> Each thread can call the following code lines when receiving 
>>>>>>>>>> a specific
>>>>>>>>>> packet:
>>>>>>>>>>
>>>>>>>>>> RTE_LCORE_FOREACH(lcore_id)
>>>>>>>>>> {
>>>>>>>>>>          result =
>>>>>>>>>> rte_mempool_get(cea_main_lcore_conf[lcore_id].de_conf.cmd_pool, 
>>>>>>>>>> (VOID_P
>>>>>>>>>> *) &new_work);        // mempools are created one for each 
>>>>>>>>>> logical core
>>>>>>>>>>          if (((uint64_t)(new_work)) < 0x7f0000000000)
>>>>>>>>>>              printf("Result %d, lcore di partenza %u, lcore 
>>>>>>>>>> di ricezione
>>>>>>>>>> %u, pointer %p\n", result, rte_lcore_id(), lcore_id, 
>>>>>>>>>> new_work);    //
>>>>>>>>>> debug print, on my server it should never happen but with 
>>>>>>>>>> multi-thread
>>>>>>>>>> happens always on the last logical core!!!!
>>>>>>>>>>
>>>>>>>> Here, checking the value of new_work looks wrong to me, before
>>>>>>>> ensuring that result == 0. At least, new_work should be set to
>>>>>>>> NULL before calling rte_mempool_get().
>>>>>>>>
>>>>>>> I put the check after result == 0, and just before the 
>>>>>>> rte_mempool_get() I set new_work to NULL, but nothing changed.
>>>>>>> The first time something goes wrong the print is
>>>>>>>
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 635, pointer 0x880002
>>>>>>>
>>>>>>> Sorry for the italian language print :) it means that 
>>>>>>> application is sending a message from the logical core 1 to the 
>>>>>>> logical core 2, it's the 635th time, the result is 0 and the 
>>>>>>> pointer is 0x880002 while all pointers before were 0x7ffxxxxxx.
>>>>>>> One strange thing is that this behaviour happens always from the 
>>>>>>> logical core 1 to the logical core 2 when the counter is 635!!! 
>>>>>>> (Sending messages from 2 to 1 or 1 to 1 or 2 to 2 is all ok)
>>>>>>> Another strange thing is that pointers from counter 636 to 640 
>>>>>>> are NULL, and from 641 begin again to be good... as you can see 
>>>>>>> here following (I attached the result of a test without the "if" 
>>>>>>> of the check on the value of new_work, and only for messages 
>>>>>>> from the lcore 1 to lcore 2)
>>>>>>>
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 627, pointer 0x7ffe8a261880
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 628, pointer 0x7ffe8a261900
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 629, pointer 0x7ffe8a261980
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 630, pointer 0x7ffe8a261a00
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 631, pointer 0x7ffe8a261a80
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 632, pointer 0x7ffe8a261b00
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 633, pointer 0x7ffe8a261b80
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 634, pointer 0x7ffe8a261c00
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 635, pointer 0x880002
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 636, pointer (nil)
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 637, pointer (nil)
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 638, pointer (nil)
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 639, pointer (nil)
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 640, pointer (nil)
>>>>>>>
>>>>>> This sure does seem like a memory over write problem, with maybe 
>>>>>> a memset(0) in the mix as well. Have you tried using hardware 
>>>>>> break points with the 0x880002 or 0x00 being written into this 
>>>>>> range?
>>>>> I put some breakpoints and found this:
>>>>>
>>>>> 1 - using pointer 0x880002, the output is (the pointer comes in 
>>>>> the middle of two rwlock):
>>>>>
>>>>> (gdb) awatch *0x880002
>>>>> Hardware access (read/write) watchpoint 1: *0x880002
>>>>> (gdb) c
>>>>> Continuing.
>>>>> [New Thread 0x7fffeded5700 (LWP 19969)]
>>>>> [New Thread 0x7fffed6d4700 (LWP 19970)]
>>>>> [New Thread 0x7fffeced3700 (LWP 19971)]
>>>>> [New Thread 0x7fffec6d2700 (LWP 19972)]
>>>>> [New Thread 0x7fffebed1700 (LWP 19973)]
>>>>> [New Thread 0x7fffeb6d0700 (LWP 19974)]
>>>>> Hardware access (read/write) watchpoint 1: *0x880002
>>>>>
>>>>> Value = 0
>>>>> rte_rwlock_init (rwl=0x880000 <ikco_sdkif_actlist_lock+677024>)
>>>>>      at 
>>>>> /usr/share/dpdk/x86_64-default-linuxapp-gcc/include/generic/rte_rwlock.h:81
>>>>> 81    }
>>>>> (gdb) c
>>>>> Continuing.
>>>>> Hardware access (read/write) watchpoint 1: *0x880002
>>>> These are most likely false positive hits and not the issue.
>>>>> Value = 0
>>>>> rte_rwlock_init (rwl=0x880004 <ikco_sdkif_actlist_lock+677028>)
>>>>>      at 
>>>>> /usr/share/dpdk/x86_64-default-linuxapp-gcc/include/generic/rte_rwlock.h:81
>>>>> 81    }
>>>>> (gdb) c
>>>>> Continuing.
>>>>>
>>>>> 2 - when using pointers minor or equal than 0x7ffe8a261d64 (in the 
>>>>> range of the mempool), gdb tells nothing about them (I don't use 
>>>>> them, I just get them from the pool and the put them in the pool 
>>>>> again);
>>>>>
>>>>> 3 - when using pointer 0x7ffe8a261d65 or greater, this is the 
>>>>> output of gdb:
>>>>>
>>>>> (gdb) awatch *(int *)0x7ffe8a261d65
>>>>> Hardware access (read/write) watchpoint 1: *(int *)0x7ffe8a261d65
>>>>> (gdb) c
>>>>> Continuing.
>>>>> [New Thread 0x7fffeded5700 (LWP 17689)]
>>>>> [New Thread 0x7fffed6d4700 (LWP 17690)]
>>>>> [New Thread 0x7fffeced3700 (LWP 17691)]
>>>>> [New Thread 0x7fffec6d2700 (LWP 17692)]
>>>>> [New Thread 0x7fffebed1700 (LWP 17693)]
>>>>> [New Thread 0x7fffeb6d0700 (LWP 17694)]
>>>>> Hardware access (read/write) watchpoint 1: *(int *)0x7ffe8a261d65
>>>>>
>>>>> Value = 0
>>>>> 0x00007ffff3798c21 in mempool_add_elem 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0, obj=obj@entry=0x7ffe8a261d80,
>>>>>      iova=iova@entry=4465237376) at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:140
>>>>> 140        STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
>>>>> (gdb) where
>>>>> #0  0x00007ffff3798c21 in mempool_add_elem 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0, obj=obj@entry=0x7ffe8a261d80,
>>>>>      iova=iova@entry=4465237376) at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:140
>>>>> #1  0x00007ffff37990f0 in rte_mempool_populate_iova 
>>>>> (mp=0x7ffebfd8d6c0, vaddr=0x7ffe8a23d540 "",
>>>>>      iova=4465087808, len=8388480, free_cb=<optimized out>, 
>>>>> opaque=<optimized out>)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:424
>>>>> #2  0x00007ffff379967d in rte_mempool_populate_default 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:624
>>>>> #3  0x00007ffff3799e89 in rte_mempool_create (name=<optimized 
>>>>> out>, n=<optimized out>,
>>>>>      elt_size=<optimized out>, cache_size=<optimized out>, 
>>>>> private_data_size=<optimized out>,
>>>>>      mp_init=0x7ffff444e410 <rte_pktmbuf_pool_init>, mp_init_arg=0x0,
>>>>>      obj_init=0x7ffff444e330 <rte_pktmbuf_init>, obj_init_arg=0x0, 
>>>>> socket_id=0, flags=0)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:952
>>>>> #4  0x0000000000548a52 in main (argc=16, argv=0x7fffffffe3c8)
>>>>>      at 
>>>>> /root/gemini-cea-4.6.0/msrc/sys/com/linux-dpdk/cea-app/../../../../sys/com/linux-dpdk/cea-app/main.c:2360
>>>>> (gdb) c
>>>>> Continuing.
>>>>> Hardware access (read/write) watchpoint 1: *(int *)0x7ffe8a261d65
>>>> This seems to be just creating a pktmbuf pool. The 
>>>> STAILQ_INSERT_TAILQ is just putting the mempool on the main tailq 
>>>> list for mempools in DPDK.
>>>>
>>>>> Old value = 0
>>>>> New value = -402653184
>>>>> 0x00007ffff3798c24 in mempool_add_elem 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0, obj=obj@entry=0x7ffe8a261e00,
>>>>>      iova=iova@entry=4465237504) at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:140
>>>>> 140        STAILQ_INSERT_TAIL(&mp->elt_list, hdr, next);
>>>>> (gdb) where
>>>>> #0  0x00007ffff3798c24 in mempool_add_elem 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0, obj=obj@entry=0x7ffe8a261e00,
>>>>>      iova=iova@entry=4465237504) at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:140
>>>>> #1  0x00007ffff37990f0 in rte_mempool_populate_iova 
>>>>> (mp=0x7ffebfd8d6c0, vaddr=0x7ffe8a23d540 "",
>>>>>      iova=4465087808, len=8388480, free_cb=<optimized out>, 
>>>>> opaque=<optimized out>)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:424
>>>>> #2  0x00007ffff379967d in rte_mempool_populate_default 
>>>>> (mp=mp@entry=0x7ffebfd8d6c0)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:624
>>>>> #3  0x00007ffff3799e89 in rte_mempool_create (name=<optimized 
>>>>> out>, n=<optimized out>,
>>>>>      elt_size=<optimized out>, cache_size=<optimized out>, 
>>>>> private_data_size=<optimized out>,
>>>>>      mp_init=0x7ffff444e410 <rte_pktmbuf_pool_init>, mp_init_arg=0x0,
>>>>>      obj_init=0x7ffff444e330 <rte_pktmbuf_init>, obj_init_arg=0x0, 
>>>>> socket_id=0, flags=0)
>>>>>      at 
>>>>> /usr/src/debug/dpdk-17.11/lib/librte_mempool/rte_mempool.c:952
>>>>> #4  0x0000000000548a52 in main (argc=16, argv=0x7fffffffe3c8)
>>>>>      at 
>>>>> /root/gemini-cea-4.6.0/msrc/sys/com/linux-dpdk/cea-app/../../../../sys/com/linux-dpdk/cea-app/main.c:2360
>>>>> (gdb) c
>>>>> Continuing.
>>>>>
>>>>> What do you think? It is normal that the mempool_add_elem is 
>>>>> called only on certain pointers of the mempool?
>>>>> I attached the initialization of the mempool. Can this be wrong?
>>>> All mempools with a cache size will have two queue to put memory 
>>>> on, one is the per lcore list and that one is used as a fast access 
>>>> queue. When the cache becomes empty or has more entries then the 
>>>> cache was created with then it pushed the extra entries to the main 
>>>> list of mempool data.
>>> Why do you say "mempools with a cache size" ? In my initialization 
>>> this mempool has cache_size = 0
>> If you give a cache size then you will have a cache list per lcore, 
>> in your case you do not have a cache. BTW not having a cache will 
>> effect performance a great deal.
>>
>>>> The only time that rwlock is touched is to get/put items on the 
>>>> main mempool.
>>>>
>>>> Just as a data point have you tried this app on 18.08 yet? I do not 
>>>> see the problem yet, sorry.
>>> I'll try 18.08 and let you know
>
> Hi ,
>
> I tried 18.08 but nothing changed about the described behaviour. I'm 
> thinking about some overflow in my code lines but using valgrind on my 
> application tells me nothing more and it seems strange to me.
> Is there any particular way to debug memory issues on dpdk application 
> apart from valgrind?
>
> Regards,
> Matteo
>
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 641, pointer 0x7ffe8a262b00
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 642, pointer 0x7ffe8a262b80
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 643, pointer 0x7ffe8a262d00
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 644, pointer 0x7ffe8a262d80
>>>>>>> Result 0, lcore di partenza 1, lcore di ricezione 2, counter 
>>>>>>> 645, pointer 0x7ffe8a262e00
>>>>>>>
>>>>>>>
>>>>>>>>>>          if (result == 0)
>>>>>>>>>>          {
>>>>>>>>>>              new_work->command = command; // usage of the 
>>>>>>>>>> memory gotten
>>>>>>>>>> from the mempool... <<<<<- here is where the application 
>>>>>>>>>> crashes!!!!
>>>>>>>>>>
>>>>>>>> Do you know why it crashes? Is it that new_work is NULL?
>>>>>>>>
>>>>>>> The pointer is not NULL but is not sequential to the others 
>>>>>>> (0x880002 as written before in this email). It seems to be in a 
>>>>>>> memory zone not in DPDK hugepages or something similar.
>>>>>>> If I use this pointer the application crashes.
>>>>>>>
>>>>>>>> Can you check how the mempool is initialized? It should be in 
>>>>>>>> multi
>>>>>>>> consumer and depending on your use case, single or multi producer.
>>>>>>>>
>>>>>>> Here is the initialization of this mempool
>>>>>>>
>>>>>>> cea_main_cmd_pool[i] = rte_mempool_create(pool_name,
>>>>>>>              (unsigned int) (ikco_cmd_buffers - 1), // 65536 - 1 
>>>>>>> in this case
>>>>>>>              sizeof (CEA_DECODE_CMD_T), // 24 bytes
>>>>>>>              0, 0,
>>>>>>>              rte_pktmbuf_pool_init, NULL,
>>>>>>>              rte_pktmbuf_init, NULL,
>>>>>>>              rte_socket_id(), 0);
>>>>>>>
>>>>>>>> Another thing that could be checked: at all the places where you
>>>>>>>> return your work object to the mempool, you should add a check
>>>>>>>> that it is not NULL. Or just enabling RTE_LIBRTE_MEMPOOL_DEBUG
>>>>>>>> could do the trick: it adds some additional checks when doing
>>>>>>>> mempool operations.
>>>>>>>>
>>>>>>> I think I have already answered this point with the prints up in 
>>>>>>> the email.
>>>>>>>
>>>>>>> What do you think about this behaviour?
>>>>>>>
>>>>>>> Regards,
>>>>>>> Matteo
>>>>>>>
>>>>>>>>>>              result =
>>>>>>>>>> rte_ring_enqueue(cea_main_lcore_conf[lcore_id].de_conf.cmd_ring,
>>>>>>>>>> (VOID_P) new_work);    // enqueues the gotten buffer on the 
>>>>>>>>>> rings of all
>>>>>>>>>> lcores
>>>>>>>>>>              // check on result value ...
>>>>>>>>>>          }
>>>>>>>>>>          else
>>>>>>>>>>          {
>>>>>>>>>>              // do something if result != 0 ...
>>>>>>>>>>          }
>>>>>>>>>> }
>>>>>>>>>>
>>>>>>>>>> This code worked perfectly (never had an issue) on 
>>>>>>>>>> dpdk-2.2.0, while if
>>>>>>>>>> I use more than 1 thread doing these operations on dpdk-17.11 
>>>>>>>>>> it happens
>>>>>>>>>> that after some times the "new_work" pointer is not a good 
>>>>>>>>>> one, and the
>>>>>>>>>> application crashes when using that pointer.
>>>>>>>>>>
>>>>>>>>>> It seems that these lines cannot be used by more than one thread
>>>>>>>>>> simultaneously. I also used many 2017 and 2018 dpdk versions 
>>>>>>>>>> without
>>>>>>>>>> success.
>>>>>>>>>>
>>>>>>>>>> Is this code possible on the new dpdk versions? Or have I to 
>>>>>>>>>> change my
>>>>>>>>>> application so that this code is called just by one lcore at 
>>>>>>>>>> a time?
>>>>>>>>>>
>>>>>>>> Assuming the mempool is properly initialized, I don't see any 
>>>>>>>> reason
>>>>>>>> why it would not work. There has been a lot of changes in 
>>>>>>>> mempool between
>>>>>>>> dpdk-2.2.0 and dpdk-17.11, but this behavior should remain the 
>>>>>>>> same.
>>>>>>>>
>>>>>>>> If the comments above do not help to solve the issue, it could 
>>>>>>>> be helpful
>>>>>>>> to try to reproduce the issue in a minimal program, so we can 
>>>>>>>> help to
>>>>>>>> review it.
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> Olivier
>>>>>>>>
>>>>>> Regards,
>>>>>> Keith
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> Regards,
>>>>>
>>>>> Matteo
>>>>>
>>>> Regards,
>>>> Keith
>>>>
>>>>
>>>>
>>> Regards,
>>> Matteo
>> Regards,
>> Keith
>>
>>
>

  reply	other threads:[~2018-08-24 14:44 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2018-08-08  9:56 Matteo Lanzuisi
2018-08-13 13:20 ` Matteo Lanzuisi
2018-08-13 21:54   ` Olivier Matz
2018-08-20 14:47     ` Matteo Lanzuisi
2018-08-20 16:03       ` Wiles, Keith
2018-08-21 12:01         ` Matteo Lanzuisi
2018-08-21 12:17           ` Wiles, Keith
2018-08-21 12:44             ` Matteo Lanzuisi
2018-08-21 12:51               ` Wiles, Keith
2018-08-21 14:46                 ` Matteo Lanzuisi
2018-08-24 14:44                   ` Matteo Lanzuisi [this message]
2018-08-24 16:47                     ` Wiles, Keith
2018-08-27 15:34                       ` Matteo Lanzuisi

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=2d1b12b7-62c9-bfd8-d899-db9990f4f44d@resi.it \
    --to=m.lanzuisi@resi.it \
    --cc=dev@dpdk.org \
    --cc=keith.wiles@intel.com \
    --cc=olivier.matz@6wind.com \
    /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
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).