difficulty w/ RTE_NEXT_ABI
I was trying to rebase my DPDK onto v2.1.0 and I came across some very confusing code in examples/l3fwd/main.c .
So... this code used the RTE_NEXT_ABI macros on a change which does not appear to affect the API... on a function that is marked always_inline ???
Maybe I missed something but this seems pointless. An always_inline function is going to have to be recompiled in any case.
Now I have no clue what would makes sense for my version of the function either... because RTE_NEXT_ABI is a binary on/off but trying to track a multi-variate quantity of ABI updates.
For now I guess I have to write it like this inside the RTE_NEXT_ABI:
rfc1812_process(struct ipv4_hdr *ipv4_hdr, uint32_t *dp, uint32_t ptype)
This seems unpleasant and kind of painful. What did I miss here?
Matthew.
static inline __attribute__((always_inline)) void
<<<<<<< 8e29af8a2843b6342dbc72db43ac82c9d29695bf
#ifdef RTE_NEXT_ABIrfc1812_process(struct ipv4_hdr *ipv4_hdr, uint16_t *dp, uint32_t ptype) #elserfc1812_process(struct ipv4_hdr *ipv4_hdr, uint16_t *dp, uint32_t flags) #endif
=======rfc1812_process(struct ipv4_hdr *ipv4_hdr, uint32_t *dp, uint32_t flags)
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
examples: update examples to use 24 bit extended next hop
{
2015-11-21 03:49, Matthew Hall:
Quoted 21 lines
I was trying to rebase my DPDK onto v2.1.0 and I came across some very confusing code in
examples/l3fwd/main.c.So... this code used the
RTE_NEXT_ABImacros on a change which does not appear to affect the API... on a function that is markedalways_inline???Maybe I missed something but this seems pointless. An
always_inlinefunction is going to have to be recompiled in any case.Now I have no clue what would makes sense for my version of the function either... because
RTE_NEXT_ABIis a binary on/off but trying to track a multi-variate quantity of ABI updates.For now I guess I have to write it like this inside the
RTE_NEXT_ABI:
rfc1812_process(struct ipv4_hdr *ipv4_hdr, uint32_t *dp, uint32_t ptype)This seems unpleasant and kind of painful. What did I miss here?
Matthew.
Quoted 7 lines
static inline __attribute__((always_inline)) void
<<<<<<<8e29af8a2843b6342dbc72db43ac82c9d29695bf
#ifdefRTE_NEXT_ABIrfc1812_process(struct ipv4_hdr *ipv4_hdr, uint16_t *dp, uint32_t ptype)#elserfc1812_process(struct ipv4_hdr *ipv4_hdr, uint16_t *dp, uint32_t flags)#endif
The new mbuf provides packet type instead of flags.
So the processing in this function is changed and the variable name is different to reflect this.
Quoted 3 lines · Thomas Monjalon
The new mbuf provides packet type instead of flags.
So the processing in this function is changed and the variable name is different to reflect this.
But the data type of the variable is the same, and this is an internal always_inline function.
So again I am confused what advantage we got from RTE_NEXT_ABI here, and how you have multiple copies of RTE_NEXT_ABI on a single symbol when it is a binary variable.
This doesn't really answer the bigger question about the reasoning.
Matthew.
2015-11-21 19:25, Matthew Hall:
Quoted 7 lines · Thomas Monjalon
Quoted 3 lines
The new mbuf provides packet type instead of flags.
So the processing in this function is changed and the variable name is different to reflect this.But the data type of the variable is the same, and this is an internal
always_inlinefunction.
It is an example application using a mbuf feature which changes depending of CONFIG_RTE_NEXT_ABI. The ABI is in the mbuf library not in the app. The header of the app function is changed only for the variable name because the semantic is changed.
Quoted 3 lines
So again I am confused what advantage we got from
RTE_NEXT_ABIhere, and how you have multiple copies ofRTE_NEXT_ABIon a single symbol when it is a binary variable.
I don't understand what is not clear here.
Quoted 1 line
This doesn't really answer the bigger question about the reasoning.
Probably because you don't ask clearly your question.
Please check the code and your question again.
Maybe that this reading may help: doc/guides/contributing/versioning.rst
Quoted 5 lines · Thomas Monjalon
Quoted 3 lines
So again I am confused what advantage we got from
RTE_NEXT_ABIhere, and how you have multiple copies ofRTE_NEXT_ABIon a single symbol when it is a binary variable.I don't understand what is not clear here.
OK. Let me restate it.
I was starting from an assumption that the purpose of RTE_NEXT_ABI was to allow ABI changes.
In most projects I worked on, a renaming of a variable when the data type is unchanged does not count as an ABI change. So it seems like this is different from the usual definition.
Secondly, if one is making an ABI change, like I was, to some code which was already changed once using RTE_NEXT_ABI, which part of the code do you change?
Do you make a third copy different from the first two copies? If you make a third copy, but RTE_NEXT_ABI is binary (i.e. it has two values, on and off) then what labeling do you apply to the third copy?
If you don't make a third copy, I am assuming you edit the copy marked with RTE_NEXT_ABI. But then what happens to a downstream user who wants to have RTE_NEXT_ABI with the first ABI change, and not your second ABI change?
Can you see what I am trying to ask now?
Matthew.
2015-11-22 18:25, Matthew Hall:
Quoted 11 lines · Thomas Monjalon
Quoted 5 lines
Quoted 3 lines
So again I am confused what advantage we got from
RTE_NEXT_ABIhere, and how you have multiple copies ofRTE_NEXT_ABIon a single symbol when it is a binary variable.I don't understand what is not clear here.
OK. Let me restate it.
I was starting from an assumption that the purpose of
RTE_NEXT_ABIwas to allow ABI changes.
Yes it is a preview. The next release will have only the new ABI without #ifdef.
Quoted 3 lines
In most projects I worked on, a renaming of a variable when the data type is unchanged does not count as an ABI change. So it seems like this is different from the usual definition.
No, it is not an ABI change.
As I said previously, the change is in the mbuf ABI. So in the example app, the 2 ABIs are used with #ifdef.
Quoted 2 lines
Secondly, if one is making an ABI change, like I was, to some code which was already changed once using
RTE_NEXT_ABI, which part of the code do you change?
If your change is sent upstream, you must rely on the new ABI because the old one will be removed when your change will be integrated.
If it is a local change, it depends on which ABI you want to use.
Quoted 7 lines
Do you make a third copy different from the first two copies? If you make a third copy, but
RTE_NEXT_ABIis binary (i.e. it has two values, on and off) then what labeling do you apply to the third copy?If you don't make a third copy, I am assuming you edit the copy marked with
RTE_NEXT_ABI. But then what happens to a downstream user who wants to haveRTE_NEXT_ABIwith the first ABI change, and not your second ABI change?
In each release, there is only 2 ABIs: current and next. In 2.2, there is only 1 ABI.
Quoted 1 line
Can you see what I am trying to ask now?
Yes. Hope my detailed answer is enough.
Quoted 3 lines · Thomas Monjalon
If your change is sent upstream, you must rely on the new ABI because the old one will be removed when your change will be integrated.
If it is a local change, it depends on which ABI you want to use.
I submitted separately to Bruce & Co. It is for LPM field expansion.
My question what to do came up in the context of a rebase to master.
Quoted 1 line
Yes. Hope my detailed answer is enough.
Thanks. It is less confusing now.
Matthew.