Rebased version of #1698 #2902

mihaibudiu · 2021-09-17T01:01:21Z

This is another attempt to fix the problems with recirculate/resubmit/clone3.
The solution is to

use an annotation (called @field_list currently) on the user metadata fields that need to be recirculated
the annotation has a numeric argument
pass the annotation numeric argument as the last argument to the recirculate/etc call
This will cause the runtime to initialize the specified fields before entering the pipeline.
This correctly implements the semantics of P4-14 recirculate/etc.
Here is an excerpt from the resubmit.p4 p4-14 benchmark, as automatically translated by the compiler:

enum bit<8> FieldLists {
    resubmit_FL = 0
}

struct mymeta_t {
    @field_list(FieldLists.resubmit_FL) 
    bit<8> f1;
}

...
    @name("._resubmit") action _resubmit() {
        meta.mymeta.f1 = 8w1;
        resubmit((bit<8>)FieldLists.resubmit_FL);
    }

This changes the signature of the respective v1model functions. We may chose instead to add new functions and deprecate the old ones.

smolkaj · 2021-10-04T20:51:31Z

Awesome, thanks. I agree that we would want to instead add new functions and deprecate the old ones.
Maybe something like clone3_with_field_list?

EDIT: Or clone_preserving_field_list?

mihaibudiu · 2021-11-02T02:21:00Z

@smolkaj can you please test to see whether this variant works for you?

smolkaj · 2021-11-02T17:05:54Z

Thanks, @mbudiu-vmw, I will give it a try!

kheradmandG · 2021-11-09T17:58:47Z

Question, how can one specify not passing any fields to recirculate/resubmit/clone3. E.g., the equivalent of clone3(…, …, {}) ?
I assume we should use a number that is not already used in any @field_list. If true, isn’t that a bit problematic? Say one uses number 0 to specify that clone3_field_list(…, …, 0) and later end up using that number in a @field_list, forgetting that 0 was used as an equivalent of specifying nothing ({}). This can lead to subtle bugs.

mihaibudiu · 2021-11-09T18:19:05Z

Yes, just use a number that has no fields attached. You could use -1, that's less likely to be a common number.
I expect people don't frequently refactor programs in this way.

kheradmandG · 2021-11-09T19:17:58Z

Thanks for the response @mbudiu-vmw. I still feel this can lead to subtle bugs but @smolkaj mentioned that there is the clone primitive that can be used for this purpose, instead.

kheradmandG · 2021-11-10T18:11:12Z

p4include/v1model.p4

- * calling clone3 with the same type and session parameter values,
- * with empty data.
- */
+extern void recirculate_field_list(bit<8> index);
 extern void clone(in CloneType type, in bit<32> session);


There is another declaration of extern void clone(in CloneType type, in bit<32> session) in line 537. This causes compilation errors:

[--Werror=duplicate] error: clone: multiple matching declarations ... v1model.p4(537): [--Werror=duplicate] error: Candidate: clone extern void clone(in CloneType type, in bit<32> session); ^^^^^ v1model.p4(586): [--Werror=duplicate] error: Candidate: clone extern void clone(in CloneType type, in bit<32> session); ^^^^^

how come our tests pass?

Not sure what flags/options we are using for the compilation (maybe @smolkaj can shed some light?). Anyway that declaration is indeed duplicated in v1model.p4 in this pull request (lines 537 and 586).

I guess if you never call the method you don't get the error.
I will remove one of the two declarations, in the meantime can you please do the same if you want to try this?

That makes sense, also maybe somebody should add a test for clone to p4c's test suite to avoid future regressions. Yeah I commented out that line and the error goes away.

kheradmandG · 2021-11-10T19:24:31Z

One more question, I have something like:

enum bit<8> PreservedFieldList {
  CLONE_I2E = 8w1
};

struct metadata_t {
  ...
  @field_list(PreservedFieldList.CLONE_I2E)
  ipv4_addr_t ip;
  ...
}

I get the error:

[--Werror=not-found] error: PreservedFieldList: declaration not found
  @field_list(PreservedFieldList.CLONE_I2E)
              ^^^^^^^^^^^^^^^^^^

However, if I change metadata_t to:

struct metadata_t {
  ...
  @field_list(8w1)
  ipv4_addr_t ip;
  ...
}

It works fine (at least in compilation).

I was wondering whether I'm missing something here.

mihaibudiu · 2021-11-10T19:30:08Z

This is a bit strange, because this exactly the kind of code that is generated by the P4-14 front-end.
It must be a bug related to name resolution.
Can you please post a complete example and I will make sure it compiles and add it as a test?

kheradmandG · 2021-11-13T04:40:01Z

I randomly selected testdata/p4_14_samples_outputs/packet_redirect-frontend.p4 and ran into the same problem. Do you run that test successfully?

mihaibudiu · 2021-11-13T04:41:14Z

If you look at this web page you will see that all tests have passed. I will give it a shot.

castilhorosa · 2021-11-14T19:10:56Z

Sorry if this is not the right place for this kind of question but I'd like to get a sneak peek of this version of p4c, with the extern resubmit_field_list, and I wondering if there is any script that I can use to compile this branch. I've been trying unsuccefully to adapt the script install-p4dev-v2.sh by using the command
git clone -b recirculate --single-branch https://github.com/mbudiu-vmw/p4c-clone.git
I would really appreciate if someone could provide me such script/commands...

mihaibudiu · 2021-11-14T19:13:00Z

Can you already build the main branch of the compiler?
Is yes, then the only thing you need to do is to switch to this branch (git checkout -b recirculate) and then run the same steps.
For me the steps are cmake ..; make

mihaibudiu · 2021-11-16T19:33:33Z

Don't use my clone of the repository, use the official one from https://github.com/p4lang/p4c
My clone is just for development and the main branch may be behind the origin.

castilhorosa · 2021-11-18T20:11:16Z

Can you already build the main branch of the compiler? Is yes, then the only thing you need to do is to switch to this branch (git checkout -b recirculate) and then run the same steps. For me the steps are cmake ..; make

Great. It worked. Thanks @mbudiu-vmw

kheradmandG · 2021-11-20T00:38:10Z

@mbudiu-vmw @smolkaj
I did some internal testing and seems like (beside the 2 issues mentioned above), the pull request works as expected.
A suggestion: might be good idea to name the the action with something that conveys that the field list is being preserved. E.g. clone_preserve_field_list or something like that.

smolkaj · 2021-11-20T00:57:10Z

+1, how about clone_preserving_field_list to make it roll off the tongue a bit easier?

smolkaj · 2021-11-20T00:59:04Z

(As you might be able to tell, Google has trained me to advocate for long, descriptive function names... 😆)

mihaibudiu · 2021-12-01T01:15:50Z

I will submit soon an update which fixes the issues you discussed. There is a new test using an enum, and that passes too, I could not reproduce Ali's error. Make sure that the enum is declared prior to its first use.

mihaibudiu · 2021-12-01T01:53:46Z

Take it back, found and solved the problem with the enums too.
@smolkaj @kheradmandG please take another look.

…g#2971) * Fix bug in file def_use in method setDefinitionsAt class AllDefinitions

kheradmandG · 2021-12-02T02:01:48Z

@mbudiu-vmw Thanks for the changes. One question regarding the naming: what does 3 in clone3 mean? And does it make sense to keep it in clone3_preserving_field_list ?

mihaibudiu · 2021-12-02T02:02:46Z

It comes from the fact that there was a clone with 2 arguments, and one with 3.

* Treat error type as string in p4RuntimeSerializer

kheradmandG · 2021-12-02T02:57:28Z

@mbudiu-vmw Thanks for the info. nit: now that this action is completely distinguishable from the clone with 2 arguments, would it make sense to rename it to clone_preserving_field_list (i.e. drop 3)?

mihaibudiu · 2021-12-03T00:19:02Z

I will make the change to clone3, but we won't be able to merge this until someone approves it.
I asked @jafingerhut for a review. Hopefully he can find the time.

…into recirculate

jfingerh · 2021-12-04T16:03:10Z

control-plane/p4RuntimeSerializer.cpp

+    /* Treat error type as string */
+    if (type->is<IR::Type_Error>())
+        return 0;
+


I'm not reviewing the C++ implementation code parts of this PR very carefully -- simply scanning through them looking for anything odd. Is this change something that was discovered as a small side issue while developing the fixes for preserving user-defined metadata fields for resubmit/recirculate/clone? Perhaps something that was needed for the failing test cases exercising those features to pass, so good to combine them into this PR while you were at it? I don't have any objection, main curious why it is thrown in with this PR.

I think that this is coming from the main branch from a rebase.

jfingerh · 2021-12-04T16:06:02Z

frontends/p4/fromv1.0/programStructure.cpp

@@ -2588,9 +2675,12 @@ void ProgramStructure::populateOutputNames() {
        "mark_to_drop",
        "hash",
        "resubmit",
+        "resubmit_preserving_field_list",


I will double-check this with my own small tests, but is resubmit_preserving_field_list a new primitive internal to p4c only, or do these changes also require new primitives in the BMv2 JSON file that BMv2 reads?

This is compiled to the standard recirculate bmv2 primitive

Sounds good. Thanks.

jfingerh · 2021-12-04T16:09:26Z

p4include/v1model.p4

+ *    @field_list(1, 2)
+ *    bit<32> y;
+ *    bit<32> z;
+ * }


Thanks for adding a short example to this documentation!

I created a separate issue to remind me to either myself, or find some other volunteer to help, update other documentation inside and outside of the p4lang/p4c repository on how these operations work. I think it is best of those changes are made in their own separate PRs after this one is merged: #2977

jfingerh · 2021-12-04T16:24:20Z

testdata/p4_14_samples_outputs/acl1-midend.p4

@@ -134,16 +134,27 @@ header data_t {
 }

 struct metadata {
+    @pa_solitary("ingress", "acl_metadata.if_label") 


How were tests passing before without these annotations? The input program had a pragma or P4_14 annotation-like thing since before this PR, so did something change in this PR to make these annotations appear in the output? Or were the tests somehow passing before even though the files did not match?

This is probably because I rebased on top of other changes. This pr does not know anything about this annotation. I don't know where it's coming from.

OK, I will not worry about it then. There is a pa_solitary pragma in the original P4_14 program, so something in the P4_14 to P4_16 translation must be picking it up.

jfingerh · 2021-12-04T20:15:04Z

testdata/p4_14_samples_outputs/copy_to_cpu-first.p4

@@ -73,7 +77,7 @@ control egress(inout headers hdr, inout metadata meta, inout standard_metadata_t

 control ingress(inout headers hdr, inout metadata meta, inout standard_metadata_t standard_metadata) {
    @name(".do_copy_to_cpu") action do_copy_to_cpu() {
-        clone3<tuple<standard_metadata_t>>(CloneType.I2E, 32w250, { standard_metadata });
+        clone_preserving_field_list(CloneType.I2E, 32w250, (bit<8>)FieldLists.copy_to_cpu_fields);


Does this implementation work to preserve standard metadata fields? If yes, how does that work, if one cannot add annotations to the standard metadata struct definition?

Regardless of the answer, it seems worth adding a sentence or three on this in the v1model.p4 comments that document how this works.

You cannot. The idea is that you should not be able to spoof standard Metadata. It's debatable, but I don't know how to do it anyway.

Seems like a reasonable restriction, and I will put it on a list of details to document elsewhere, including a suggested workaround that of course one is welcome to copy a standard metadata field value into a user-defined metadata field, and preserve the user-defined metadata field.

jfingerh · 2021-12-05T02:59:47Z

testdata/p4_16_samples/issue1653-bmv2.p4

        // Copy another header's data to local variable.
        bh.x = h.bvh0.x;
-        clone3(CloneType.I2E, 0, h);
+        clone_preserving_field_list(CloneType.I2E, 0, 0);


It seems like if we wanted to remain faithful to the original meaning of this test program, there should be @field_list(0) annotations on all fields of struct parsed_packet_t?

The clone only preserves metadata fields,. Adding annotations to parsed_packet_t should have no effect.

jfingerh

Looks good to me. As usual, I have focused on the v1model.p4 include file changes, and the changes to the test programs and their expected outputs. Except for one minor comment where it appears that a test program may have changed its behavior by no longer preserving some fields that the original did, I did not notice any problems.

mihaibudiu · 2021-12-06T22:50:07Z

@jafingerhut if I have your approval please give it officially so we can merge this.
Or suggest alternative reviewers...

jfingerh · 2021-12-07T00:29:35Z

jafingerh is also my user account, and I approved using that one already. I can add approval using my account jafingerhut in a few minutes, in case that makes a difference for tracking purposes.

jafingerhut

Looks good to me. As usual, I have focused on the v1model.p4 include file changes, and the changes to the test programs and their expected outputs. Except for one minor comment where it appears that a test program may have changed its behavior by no longer preserving some fields that the original did, I did not notice any problems.

#1057) * Updates to documentation on metdata preservation after recent p4c changes In particular after this pull request was merged into p4c: p4lang/p4c#2902 * Remove some unnecessary changes from first commit Should make this easier to review. * Fix minor typo * Update the v1model packet_length documentation based on code changes done on 2021-Dec-16.

mihaibudiu mentioned this pull request Sep 17, 2021

A simple way to implement resubmit/recirculate/clone3 in P4-16 on v1model #1698

Closed

mihaibudiu requested a review from jafingerhut September 17, 2021 01:03

mihaibudiu mentioned this pull request Sep 17, 2021

Expanding structs into lists for BMv2 seems to have broken clone3 #2875

Closed

mihaibudiu force-pushed the recirculate branch from db3b0da to 29d8125 Compare November 2, 2021 02:20

mihaibudiu force-pushed the recirculate branch from 29d8125 to 94a3cda Compare November 2, 2021 17:44

mihaibudiu mentioned this pull request Nov 9, 2021

Resubmit with annotations in v1model #2947

Closed

kheradmandG reviewed Nov 10, 2021

View reviewed changes

Mihai Budiu added 4 commits November 30, 2021 16:41

Rebased version of p4lang#1698

c43a616

Rename resubmit to resubmit_field_list, etc.

a65d505

Update tests

e2e5490

reference output

63183a4

Renamed functions as suggested

fe8fa2d

Mihai Budiu and others added 4 commits December 1, 2021 10:30

field_list can take a list of expressions

e19906e

Compiler Bug : overwriting definitions in defuse (p4lang#2900) (p4lan…

f520931

…g#2971) * Fix bug in file def_use in method setDefinitionsAt class AllDefinitions

reference outputs

9943732

Add temporary IR nodes to typeMap (p4lang#2972)

f12dc2e

Add minWidthBits for Type_Error (p4lang#2967)

b958226

* Treat error type as string in p4RuntimeSerializer

Mihai Budiu added 2 commits December 2, 2021 17:00

Merge branch 'recirculate' of https://github.com/mbudiu-vmw/p4c-clone …

5c263dd

…into recirculate

Rename clone3_... to clone_...

e8579cf

jfingerh reviewed Dec 4, 2021

View reviewed changes

jfingerh mentioned this pull request Dec 5, 2021

Update some comments in test programs regarding recirculate, resubmit, and clone operations in v1model #2977

Closed

jfingerh reviewed Dec 5, 2021

View reviewed changes

jfingerh approved these changes Dec 5, 2021

View reviewed changes

jafingerhut approved these changes Dec 7, 2021

View reviewed changes

mihaibudiu merged commit 2be448f into p4lang:main Dec 7, 2021

mihaibudiu deleted the recirculate branch December 7, 2021 00:31

jafingerhut mentioned this pull request Dec 12, 2021

Updates to documentation on metdata preservation after recent p4c cha… p4lang/behavioral-model#1057

Merged

usha1830 mentioned this pull request Dec 21, 2021

Switch statement not correctly translated when default case is not present #2986

Closed

mihaibudiu mentioned this pull request Dec 23, 2021

Is behavior of P4_16+v1model resubmit() primitive action outside of P4_16 language spec? #1514

Closed

MichalKekely mentioned this pull request Jan 5, 2022

CONVERT_PRIMITIVE no longer has second (optional) argument, why? #3019

Closed

Rebased version of #1698 #2902

Rebased version of #1698 #2902

Conversation

mihaibudiu commented Sep 17, 2021

smolkaj commented Oct 4, 2021 • edited Loading

mihaibudiu commented Nov 2, 2021

smolkaj commented Nov 2, 2021

kheradmandG commented Nov 9, 2021 • edited Loading

mihaibudiu commented Nov 9, 2021

kheradmandG commented Nov 9, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kheradmandG commented Nov 10, 2021 • edited Loading

mihaibudiu commented Nov 10, 2021

kheradmandG commented Nov 13, 2021

mihaibudiu commented Nov 13, 2021

castilhorosa commented Nov 14, 2021

mihaibudiu commented Nov 14, 2021

mihaibudiu commented Nov 16, 2021

castilhorosa commented Nov 18, 2021

kheradmandG commented Nov 20, 2021 • edited Loading

smolkaj commented Nov 20, 2021

smolkaj commented Nov 20, 2021

mihaibudiu commented Dec 1, 2021

mihaibudiu commented Dec 1, 2021

kheradmandG commented Dec 2, 2021 • edited Loading

mihaibudiu commented Dec 2, 2021

kheradmandG commented Dec 2, 2021

mihaibudiu commented Dec 3, 2021

Choose a reason for hiding this comment

mihaibudiu Dec 6, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mihaibudiu Dec 4, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jfingerh left a comment

Choose a reason for hiding this comment

mihaibudiu commented Dec 6, 2021

jfingerh commented Dec 7, 2021

jafingerhut left a comment

Choose a reason for hiding this comment

smolkaj commented Oct 4, 2021 •

edited

Loading

kheradmandG commented Nov 9, 2021 •

edited

Loading

kheradmandG commented Nov 9, 2021 •

edited

Loading

kheradmandG commented Nov 10, 2021 •

edited

Loading

kheradmandG commented Nov 20, 2021 •

edited

Loading

kheradmandG commented Dec 2, 2021 •

edited

Loading

mihaibudiu Dec 6, 2021 •

edited

Loading

mihaibudiu Dec 4, 2021 •

edited

Loading