Re: [PATCH v2] libc_fatal: Get rid of alloca

[Joe Simmons-Talbott <josimmon@redhat.com>]

On Wed, Sep 06, 2023 at 02:45:50PM -0400, Joe Simmons-Talbott wrote:
> On Wed, Sep 06, 2023 at 01:51:12PM -0300, Adhemerval Zanella Netto wrote:
> > 
> > 
> > On 06/09/23 12:43, Joe Simmons-Talbott wrote:
> > > On Fri, Sep 01, 2023 at 11:23:08AM -0300, Adhemerval Zanella Netto wrote:
> > >>
> > >>
> > >> On 31/08/23 17:20, Joe Simmons-Talbott via Libc-alpha wrote:
> > >>> Use fixed size arrays in place of alloca to avoid potential stack overflow.
> > >>> Limit the number of varargs to __libc_message to 10.
> > >>
> > >> I think we enforce the maximum number of arguments internally with some
> > >> macro tricks, so there is no need to bail out to abort without printing
> > >> the message.
> > >>
> > >>> ---
> > >>> Changes to v1:
> > >>>  * Use a fixed size array rather than scratch_buffers since we can only
> > >>>    call async signal safe functions.
> > >>>
> > >>>  sysdeps/posix/libc_fatal.c | 11 +++++++++--
> > >>>  1 file changed, 9 insertions(+), 2 deletions(-)
> > >>>
> > >>> diff --git a/sysdeps/posix/libc_fatal.c b/sysdeps/posix/libc_fatal.c
> > >>> index 70edcc10c1..16929addab 100644
> > >>> --- a/sysdeps/posix/libc_fatal.c
> > >>> +++ b/sysdeps/posix/libc_fatal.c
> > >>> @@ -45,6 +45,9 @@ writev_for_fatal (int fd, const struct iovec *iov, size_t niov, size_t total)
> > >>>  }
> > >>>  #endif
> > >>>  
> > >>> +/* The maximum number of varargs allowed in a __libc_message format string */
> > >>> +#define MAX_NLIST 10
> > >>> +
> > >>>  struct str_list
> > >>>  {
> > >>>    const char *str;
> > >>> @@ -58,6 +61,7 @@ __libc_message (const char *fmt, ...)
> > >>>  {
> > >>>    va_list ap;
> > >>>    int fd = -1;
> > >>> +  struct str_list _newp[MAX_NLIST];
> > >>
> > >> There are no need to track the string list, it is used essentially to
> > >> construct the iovec struct to call writev.  You can construct the
> > >> iovec directly.
> > >>
> > >>>  
> > >>>    va_start (ap, fmt);
> > >>>  
> > >>> @@ -70,6 +74,7 @@ __libc_message (const char *fmt, ...)
> > >>>  
> > >>>    struct str_list *list = NULL;
> > >>>    int nlist = 0;
> > >>> +  struct iovec iov[MAX_NLIST];
> > >>>  
> > >>>    const char *cp = fmt;
> > >>>    while (*cp != '\0')
> > >>> @@ -100,17 +105,18 @@ __libc_message (const char *fmt, ...)
> > >>>  	  cp = next;
> > >>>  	}
> > >>>  
> > >>> -      struct str_list *newp = alloca (sizeof (struct str_list));
> > >>> +      struct str_list *newp = &_newp[nlist];
> > >>>        newp->str = str;
> > >>>        newp->len = len;
> > >>>        newp->next = list;
> > >>>        list = newp;
> > >>>        ++nlist;
> > >>> +      if (nlist > MAX_NLIST)
> > >>> +        goto fail_out;
> > >>>      }
> > >>>  
> > >>>    if (nlist > 0)
> > >>>      {
> > >>> -      struct iovec *iov = alloca (nlist * sizeof (struct iovec));
> > >>>        ssize_t total = 0;
> > >>>  
> > >>>        for (int cnt = nlist - 1; cnt >= 0; --cnt)
> > >>> @@ -146,6 +152,7 @@ __libc_message (const char *fmt, ...)
> > >>>  
> > >>>    va_end (ap);
> > >>>  
> > >>> +fail_out:
> > >>>    /* Kill the application.  */
> > >>>    abort ();
> > >>>  }
> > >>
> > >> Below is a patch on top of your which enforces the maximum number of 
> > >> supported variadic arguments with a similar trick I used for 
> > >> INLINE_SYSCALL_CALL.  One will need to explicit implement a new macro
> > >> for each __libc_message usage with a number of arguments larger than
> > >> LIBC_MESSAGE_MAX_ARGS, but it will fail at build time if you try to 
> > >> use __libc_message with 5 or more arguments.
> > > 
> > > Thanks for the patch.  I applied it and during testing see failures for
> > > stdlib/tst-bz20544 with the following output:
> > > 
> > > <<<
> > > Did not find expected string in error output:
> > >   expected: >>>assertion failed: func != NULL
> > > <<<
> > >   actual:   >>>Fatal glibc error: on_exit.c:31 (__on_exit): assertion failed: ): assertion failed: %s
> > > 
> > > <<<
> > > Did not find expected string in error output:
> > >   expected: >>>assertion failed: func != NULL
> > > <<<
> > >   actual:   >>>Fatal glibc error: cxa_atexit.c:41 (__internal_atexit): assertion failed: ): assertion failed: %s
> > > 
> > > <<<
> > > Did not find expected string in error output:
> > >   expected: >>>assertion failed: func != NULL
> > > <<<
> > >   actual:   >>>Fatal glibc error: cxa_atexit.c:41 (__internal_atexit): assertion failed: ): assertion failed: %s
> > > 
> > > <<<
> > 
> > That is unexpected, I have not see any regression testing here.  Could
> > you check a clean build with a branch I have pushed on sourceware [1]?
> > 
> > [1] https://sourceware.org/git/?p=glibc.git;a=commit;h=742b35228f3efa25d41d14f27c8911f308514b28
> > 
> 
> I saw the same error with your branch on a clean build.  I think the
> issue is that iov needs to have space for the parts of the format
> string that are not varargs too.  I replaced it with:
> 
> struct iovec iov[LIBC_MESSAGE_MAX_ARGS * 2 - 1];
> 
> and that fixed the errors.
> 

I've posted the patch[1]

Thanks,
Joe

[1] https://sourceware.org/pipermail/libc-alpha/2023-September/151403.html