Re: [dpdk-dev] [PATCH v6 3/4] bond mode 4: allow external state machine

[Panu Matilainen <pmatilai@redhat.com>]

On 11/20/2015 09:46 PM, Eric Kinzie wrote:
> On Tue Nov 03 11:48:57 +0000 2015, Ferruh Yigit wrote:
>> On Tue, Nov 03, 2015 at 01:31:45PM +0200, Panu Matilainen wrote:
>>> On 11/03/2015 01:02 PM, Ferruh Yigit wrote:
>>>> On Tue, Nov 03, 2015 at 08:48:16AM +0200, Panu Matilainen wrote:
>>>>> On 11/02/2015 06:42 PM, Eric Kinzie wrote:
>>>>>> On Mon Nov 02 12:23:47 +0200 2015, Panu Matilainen wrote:
>>>>>>> On 11/01/2015 08:17 PM, Thomas Monjalon wrote:
>>>>>>>> 2015-10-19 08:36, Eric Kinzie:
>>>>>>>>>      Size of struct rte_eth_bond_8023ad_conf changed.  Increment LIBABIVER
>>>>>>>>>      and version bond_mode_8023ad_setup and bond_mode_8023ad_conf_get
>>>>>>>>>      functions.
>>>>>>>> [...]
>>>>>>>>> +VERSION_SYMBOL(bond_mode_8023ad_setup, _v20, 2.0);
>>>>>>>> [...]
>>>>>>>>> +BIND_DEFAULT_SYMBOL(bond_mode_8023ad_setup, _v22, 2.2);
>>>>>>>>> +MAP_STATIC_SYMBOL(void bond_mode_8023ad_setup(struct rte_eth_dev *dev, struct rte_eth_bond_8023ad_conf *conf), \
>>>>>>>>> +		  bond_mode_8023ad_setup_v22);
>>>>>>>>
>>>>>>>> I'm sorry it doesn't work well when trying to build a combined lib:
>>>>>>>>
>>>>>>>> ld: libdpdk.so: version node not found for symbol bond_mode_8023ad_setup@@DPDK_2.2
>>>>>>>>
>>>>>>>> The symbols are OK in the .o file:
>>>>>>>> 0000000000002340 g     F .text  0000000000000171 bond_mode_8023ad_setup@@DPDK_2.2
>>>>>>>> 0000000000002260 g     F .text  00000000000000da bond_mode_8023ad_setup@DPDK_2.0
>>>>>>>> 0000000000002260 g     F .text  00000000000000da bond_mode_8023ad_setup_v20
>>>>>>>> 0000000000002340 g     F .text  0000000000000171 bond_mode_8023ad_setup_v22
>>>>>>>> 0000000000000000         *UND*  0000000000000000 bond_mode_8023ad_setup
>>>>>>>>
>>>>>>>> I don't understand the problem and I am considering disabling versioning in
>>>>>>>> combined library.
>>>>>>>>
>>>>>>>> Any idea?
>>>>>>>>
>>>>>>>
>>>>>>> The .map additions look incorrect to me:
>>>>>>>
>>>>>>>> diff --git a/drivers/net/bonding/rte_eth_bond_version.map b/drivers/net/bonding/rte_eth_bond_version.map
>>>>>>>> index 22bd920..7f78717 100644
>>>>>>>> --- a/drivers/net/bonding/rte_eth_bond_version.map
>>>>>>>> +++ b/drivers/net/bonding/rte_eth_bond_version.map
>>>>>>>> @@ -17,6 +17,9 @@  DPDK_2.0 {
>>>>>>>> 	rte_eth_bond_slaves_get;
>>>>>>>> 	rte_eth_bond_xmit_policy_get;
>>>>>>>> 	rte_eth_bond_xmit_policy_set;
>>>>>>>> +	rte_eth_bond_8023ad_ext_collect;
>>>>>>>> +	rte_eth_bond_8023ad_ext_distrib;
>>>>>>>> +	rte_eth_bond_8023ad_ext_slowtx;
>>>>>>>
>>>>>>> These symbols didn't exist in DPDK 2.0 but are only being added
>>>>>>> here. So why are they being added to the 2.0 section?
>>>>>>
>>>>>> Yes, I think these should probably be moved.
>>>>>>
>>>>>>
>>>>>>>>
>>>>>>>> 	local: *;
>>>>>>>> };
>>>>>>>> @@ -27,3 +30,10 @@  DPDK_2.1 {
>>>>>>>> 	rte_eth_bond_free;
>>>>>>>>
>>>>>>>> } DPDK_2.0;
>>>>>>>> +
>>>>>>>> +DPDK_2.2 {
>>>>>>>> +	local
>>>>>>>> +
>>>>>>>> +	bond_mode_8023ad_conf_get;
>>>>>>>> +	bond_mode_8023ad_setup;
>>>>>>>> +} DPDK_2.1;
>>>>>>>
>>>>>>> These are marked local, as in, "not exported" which doesn't seem
>>>>>>> right. Also they're lacking the rte_eth_ prefix. AFAICS this is what
>>>>>>> the symbol export map should look like here:
>>>>>>
>>>>>>
>>>>>> These were not exported to begin with.  But after versioning these
>>>>>> functions, they are exported unless explicitly declared to be local here.
>>>>>>
>>>>>
>>>>> And this does not ring any warning bells? :)
>>>>>
>>>>> Sorry, I was not looking at the patch as a whole. You're declaring these
>>>>> symbols as exported with the versioning macros, eg
>>>>>
>>>>> BIND_DEFAULT_SYMBOL(bond_mode_8023ad_setup, _v22, 2.2);
>>>>>
>>>>> ...and then explicitly telling it to not export them by declaring local, and
>>>>> then we wonder why it has trouble finding the symbols.
>>>>> The versioning macros wont invent the librte_ prefix for you, you need to
>>>>> rename the functions accordingly.
>>>>>
>>>>> But all this versioning gymnastics is moot anyway because you declare the
>>>>> ABI incompatible:
>>>>>
>>>>> -LIBABIVER := 1
>>>>> +LIBABIVER := 2
>>>>>
>>>>> This changes the library soname, so no binary compiled against the previous
>>>>> version can possibly use it anymore. As in, by definition there can be no
>>>>> callers of the _v20 variants after this ABI version bump.
>>>>>
>>>> An observation: even soname is different, just renaming .so file itself works.
>>>>
>>>> And this can be useful for the case:
>>>> libx.so.1 provides functions A, B, C
>>>> app1 compiled against libx.so.1, using only function B
>>>>
>>>> libx.so.1 updated only thefunction A and become libx.so.2
>>>> app1 still can run successfully by re-naming lib to libx.so.1 (even soname is libx.so.2)
>>>>
>>>> But for this usage, user needs to know which function updated and is it safe or not to use this library,
>>>> I wonder if there is an automatic way of resolving this dependency.
>>>
>>> Erm, no. The whole point of changing soname and the physical filename is to
>>> tell others it is incompatible with earlier versions. Yes you can rename the
>>> file and get lucky (or not), just like you can play Russian roulette.
>>> Neither are particularly healthy ideas.
>>>
>>> Symbol version exists in part to allow libraries to evolve while maintaining
>>> compatibility, but it requires careful planning and programming. When public
>>> structs change, the structs would have to be versioned too, and from there
>>> on it starts getting more and more complicated.
>>>
>>
>> If we strictly want to prevent using old library, whenever LIBABIVER increased, we should update .map as following, right?
>>
>>   FROM: (dpdk2.1)
>> ================
>> DPDK_2.0 {
>> 	A;
>> 	B;
>> 	C;
>> };
>>
>> DPDK_2.1 {
>> 	D;
>> 	E;
>> };
>>
>> LIBABIVER=1
>> ================
>>
>>   TO: (dpdk2.2)
>> ================
>> DPDK_2.2 {
>> 	A;
>> 	B;
>> 	C;
>> 	D;
>> 	E;
>> };
>>
>> LIBABIVER=2
>> ================
>>
>>
>> So this won't work for anybody without luck factor, I think currently we are not doing this.
>>
>>
>> thanks,
>> ferruh
>>
>>
>
> Panu, Ferruh, is there agreement on an acceptable approach to this?

Define "this" :)

As discussed before in (and around) these:
http://dpdk.org/ml/archives/dev/2015-November/027317.html
http://dpdk.org/ml/archives/dev/2015-November/027321.html

Thomas thinks the struct modification might not even need a LIBABIVER 
bump, but that all depends on the details of how the struct is used, 
I've no clue about that so I wont comment further on it.

Since the added symbols seem to be intended to be private, the 
versioning macro uses and .map changes should be just dropped. Only 
exported symbols are to be versioned.

So ultimately the answer is, "yes, but it depends on the details" :)

	- Panu -

	- Panu -