Re: linux-next: build failure after merge of the akpm-current tree

[J. Bruce Fields]

On Sat, Feb 01, 2020 at 08:54:01AM +0100, Arnd Bergmann wrote:
> On Sat, Feb 1, 2020 at 1:32 AM J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
> >
> > On Fri, Jan 31, 2020 at 02:13:09PM +1100, Stephen Rothwell wrote:
> > > Hi all,
> > >
> > > After merging the akpm-current tree, today's linux-next build (arm
> > > multi_v7_defconfig) failed like this:
> > >
> > > net/sunrpc/auth_gss/svcauth_gss.c: In function 'gss_proxy_save_rsc':
> > > net/sunrpc/auth_gss/svcauth_gss.c:1251:19: error: storage size of 'boot' isn't known
> > >  1251 |   struct timespec boot;
> > >       |                   ^~~~
> > > net/sunrpc/auth_gss/svcauth_gss.c:1273:3: error: implicit declaration of function 'getboottime'; did you mean 'getboottime64'? [-Werror=implicit-function-declaration]
> > >  1273 |   getboottime(&boot);
> > >       |   ^~~~~~~~~~~
> > >       |   getboottime64
> > > net/sunrpc/auth_gss/svcauth_gss.c:1251:19: warning: unused variable 'boot' [-Wunused-variable]
> > >  1251 |   struct timespec boot;
> > >       |                   ^~~~
> > >
> > > Caused by commit
> > >
> > >   a415f20a18c9 ("sunrpc: expiry_time should be seconds not timeval")
> 
> This commit uses the now-removed 'struct timespec' type and 'getboottime()'
> function, so to fix the compilation error, the 64-bit replacements need to e
> used as described in Documentation/core-api/timekeeping.rst
> 
> > > from the nfsd tree interacting with commits
> > >
> > >   de371b6c7b73 ("y2038: remove unused time32 interfaces")
> > >   aa7ff200a719 ("y2038: hide timeval/timespec/itimerval/itimerspec types")
> > >
> > > from the akpm-current tree.
> > >
> > > I have reverted the nfsd commit for today.  A better solution is requested.
> >
> > Unfortunately that expiry time seems to be a signed 32-bit integer in
> > both the kernel<->gss-proxy and the gss-proxy<->krb5 interfaces.
> >
> > I guess we'll have to come to an agreement with the krb5 developers.
> >
> > Simplest might be to agree that the thing's unsigned.  The expiry
> > shouldn't ever need to be decades in the future, so unsigned mod 2^32
> > arithmetic should work forever.
> 
> Can you be more specific which interface you are referring to?
> My change to gss_import_v1_context() is now part of mainline
> as of 294ec5b87a8a ("sunrpc: convert to time64_t for expiry"),
> is anything else needed there?

Oops, thanks, I was looking at a branch without that patch and forgot
that it changed that field to unsigned.

Looking at gss-proxy: it gets this expiration time from the krb5
libraries with a call to gss_inquire_context(..., &lifetime_rec, ...)
where lifetime_rec is uint32_t.

Treating this as a 64-bit time makes everything work till 2106, but, for
what it's worth, keeping it a u32 would make it correct forever (since
mod-2^32 subtraction will still give a reasonable answer).

--b.