Re: kmsg: lseek errors confuse glibc's dprintf

[Mike Crowe]

On Thursday 21 March 2019 at 11:33:33 +0100, Arnd Bergmann wrote:
> On Thu, Mar 21, 2019 at 10:47 AM Alexander Sverdlin
> <alexander.sverdlin@xxxxxxxxx> wrote:
> >
> > Hello Mike and all,
> >
> > On Thu, 15 Jan 2015 17:31:32 +0000
> > Mike Crowe <mac@xxxxxxxxxx> wrote:
> >
> > > glibc's dprintf implementation does not work correctly with /dev/kmsg file
> > > descriptors because glibc treats receiving EBADF and EINVAL from lseek when
> > > trying to determine the current file position as errors. See
> > > https://sourceware.org/bugzilla/show_bug.cgi?id=17830
> >
> > we need to conclude on this issue. This is a real bug which is ignored
> > for 4 years now. Mike, would you like to re-send a formal patch?
> > I can do it as well, preserving a link to your original patch/report.
> > In case you'd like to post it yourself, I can be a tester and/or
> > provide a reproducer.
> 
> The patch needs to be rebased because of the changed file
> location. I would also suggest adding a "Cc: stable@xxxxxxxxxx"
> tag so it will get backported into stable kernels.
> 
> > > >>From what I can tell prior to Kay Sievers printk record commit
> > > e11fea92e13fb91c50bacca799a6131c81929986, calling lseek(fd, 0, SEEK_CUR)
> > > with such a file descriptor would not return an error.
> > >
> > > Prior to Kay's change, Arnd Bergmann's commit
> > > 6038f373a3dc1f1c26496e60b6c40b164716f07e seemed to go to some lengths to
> > > preserve the successful return code rather than returning (the perhaps more
> > > logical) -ESPIPE.
> > >
> > > glibc is happy with either a successful return or -ESPIPE.
> > >
> > > For maximum compatibility it seems that success should be returned but
> > > given Kay's new seek interface perhaps this isn't helpful.
> > >
> > > This patch ensures that such a seek succeeds:
> > >
> > > diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
> > > index 02d6b6d..b3ff6f0 100644
> > > --- a/kernel/printk/printk.c
> > > +++ b/kernel/printk/printk.c
> > > @@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)
> > >       loff_t ret = 0;
> > >
> > >       if (!user)
> > > -             return -EBADF;
> > > +             return (whence == SEEK_CUR) ? 0 : -EBADF;
> > >       if (offset)
> > >               return -ESPIPE;
> > >
> > > @@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)
> > >               user->idx = log_next_idx;
> > >               user->seq = log_next_seq;
> > >               break;
> > > +     case SEEK_CUR:
> > > +             /* For compatibility with userspace requesting the
> > > +              * current file position. */
> > > +             ret = 0;
> > > +             break;
> > >       default:
> > >               ret = -EINVAL;
> > >       }
> > >
> > > (although it could be argued that the !user case should return -ESPIPE
> > > rather than EBADF since the file descriptor _is_ valid.)
> 
> I don't think the !user case can ever be hit, I would just leave that
> to return -BADF and not touch it.

I originally wrote the patch for linux-3.8(!) and back then a write-only
kmsg instance had no private data. It looks like this changed with
750afe7babd117daabebf4855da18e4418ea845e in v4.8, so user should now always
be valid.

> > > @@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)
> > >               user->idx = log_next_idx;
> > >               user->seq = log_next_seq;
> > >               break;
> > > +     case SEEK_CUR:
> > > +             /* For compatibility with userspace expecting SEEK_CUR
> > > +              * to not yield EINVAL. */
> > > +             ret = -ESPIPE;
> > > +             break;
> > >       default:
> > >               ret = -EINVAL;
> > >       }
> > >
> > > Either makes dprintf work, but is either the right solution?
> 
> I'd vote for -ESPIPE, for consistency with the offset!=0 case, but
> etiher one is fine with me here.

It seems that this is the version we've successfully been running in our
kernel since then, so that's good.

Thanks.

Mike.