Florian Westphal <fw@xxxxxxxxx> wrote:
> Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote:
> > Hi Thomas,
> >
> > On Wed, Nov 15, 2023 at 01:36:40PM +0100, Thomas Haller wrote:
> > > On Wed, 2023-11-15 at 13:30 +0100, Pablo Neira Ayuso wrote:
> > [...]
> > > > I see _lots_ of DUMP FAIL with kernel 5.4
> > >
> > > Hi,
> > >
> > > Could you provide more details?
> > >
> > > For example,
> > >
> > > make -j && ./tests/shell/run-tests.sh tests/shell/testcases/include/0007glob_double_0 -x
> > > grep ^ -a -R /tmp/nft-test.latest.*/
> >
> > # cat [...]/ruleset-diff.json
> > --- testcases/include/dumps/0007glob_double_0.json-nft 2023-11-15 13:27:20.272084254 +0100
> > +++ /tmp/nft-test.20231116-170617.584.lrZzMy/test-testcases-include-0007glob_double_0.1/ruleset-after.json 2023-11-16 17:06:18.332535411 +0100
> > @@ -1 +1 @@
> > -{"nftables": [{"metainfo": {"version": "VERSION", "release_name": "RELEASE_NAME", "json_schema_version": 1}}, {"table": {"family": "ip", "name": "x", "handle": 1}}, {"table": {"family": "ip", "name": "y", "handle": 2}}]}
> > +{"nftables": [{"metainfo": {"version": "VERSION", "release_name": "RELEASE_NAME", "json_schema_version": 1}}, {"table": {"family": "ip", "name": "x", "handle": 158}}, {"table": {"family": "ip", "name": "y", "handle": 159}}]}
> >
> > It seems that handles are a problem in this diff.
>
> Are you running tests with -s option?
>
> In that case, modules are removed after each test.
>
> I suspect its because we can then hit -EAGAIN mid-transaction
> because module is missing (again), then replay logic does its
> thing.
>
> But the handle generator isn't transaction aware,
> so it has advanced vs. the aborted partial transaction.
>
> I'm not sure what to do here.
>
> One the one hand those rmmods are plain stupid, but on the other
> hand this adds partial coverage for the rmmod path.
>
> We could make the handle counter transaction aware to
> "fix" this on kernel side; it should not be too much code.
>
> What do you think?
Oh, wait, on older kernels the entire handle counter is global,
so "unshare -n" has no effect on it.
But the rmmod scenario explained above happens as well, this
"breaks" json dumps on centos stream 9, which does have the
pernet handle generator fix backported.
