[CALL FOR TESTERS] New system call: abort2()

[John Baldwin]

On Friday 16 December 2005 05:19 pm, Wojciech A. Koszek wrote:
> On Fri, Dec 16, 2005 at 11:14:12AM -0500, John Baldwin wrote:
> > On Friday 16 December 2005 04:10 am, Peter Jeremy wrote:
> > > On Thu, 2005-Dec-15 22:37:45 +0000, Wojciech A. Koszek wrote:
> > > >	abort2(const char *why, int nargs, void **args);
> > > >
> > > >"why" is reason of program abort, "nargs" is number of arguments
> > > >passed in "args". Both "why" and "args" (with "%p" format) will be
> > > >printed via log(9). Sample output:
> > > >[..]
> > > >pid <3004> <abort2> abort2: ABORT2 <arg0:0x62612f2e>
> > > >pid <3019> <abort2> abort2: invalid argument
> > > >[..]
> > >
> > > I don't believe the following code is correct.  uap->args is a
> > > userspace pointer so uap->args[i] is dereferencing a userspace
> > > argument in kernelspace.
> > > +                       arg = uargs[i] = (void *) fuword(uap->args[i]);
> > > I think it should be fuword(uap->args + i);
> > >
> > > I don't see the point of the following test.  "arg" is printed using
> > > %p and never de-referenced so there's no reason it can't be NULL.  I
> > > would see that a legitimate use of abort2() is when the application
> > > detects that a pointer is unexpectedly NULL.  Aborting on -1 is less
> > > clear - if fuword() fails, it will return -1 but, equally, a faulty
> > > user application may have left -1 in a pointer.  (Note that mmap(2)
> > > returns -1 on error so it's not inconceivable that a pointer could
> > > contain -1).
> > >
> > > +                       /* Prevent from faults in user-space */
> > > +                       if (arg == NULL || arg == (void *)-1) {
> > > +                               error = EINVAL;
> > > +                               break;
> > > +                       }
> > >
> > > Taking the above into account, I believe the code should be:
> > > +               if (uap->args == NULL)
> > > +                       break;
> > > +               error = copyin(uap->args, uargs, uap->nargs * sizeof
> > > (void *)); +               if (error != 0)
> > > +                       break;
> >
> > Agreed.  Also, copyinstr() can provide a better interface for copying the
> > why string in.  Also, the PROC LOCK isn't needed for reading the static
> > p_pid and p_comm fields of struct proc.  Also, I second the other
> > comments of do { } while(0) vs goto.  Many existing syscalls use 'goto
> > out;' for error handling, and I think that is one of the very few cases
> > when goto is useful and not harmful.
>
> Thanks for the suggestions and comments!
>
> My question is related with copying string from user-space: the only
> difference I can see between those functions (other than operating of
> strings/sbufs) is that sbuf_copyin() looses 'done' [1]. Since current
> abort2() makes use of sbuf(9), I'll have to have additional buffer just
> for string copying and than copy it to sbuf. Would you prefer this
> solution or complete migration from sbufs to strl..()?
>
> [1] Couldn't sbuf_copyin() simply return 'done' from copyinstr()
> embedded in it, since it already returns -1 on failure? This function
> is used in two places, which make no use of return value.

That sounds good to me (fixing sbuf_copyin()).

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org