wrong configure check for off_t == long long

[John Spencer]

Dmitry V. Levin wrote:
> On Sat, Nov 02, 2013 at 05:43:50PM +0000, John Spencer wrote:
>> on a 64 bit linux system, with musl libc (off_t is long long), the 
>> following happens:
>>
>> configure:6234: checking for long long off_t
>> conftest.c:42:6: error: size of array 'a' is negative
>>
>> | #include <sys/types.h>
>> | char a[(sizeof (off_t) == sizeof (long long) &&
>> |         sizeof (off_t) > sizeof (long)) - 1];
>> |
>>
>> configure:6254: result: no
>>
>> the problem is the second part of the check, which is completely bogus 
>> on a 64bit system, where sizeof long long == sizeof long.
> 
> This is expected behaviour.  AC_OFF_T_IS_LONG_LONG checks whether long type
> has enough room for off_t, or long long type is required instead.  It has
> nothing to do with the way how off_t is actually defined in the system.
> AC_OFF_T_IS_LONG_LONG was added to strace 12.5 years ago so the name
> of this macro might be slightly misleading nowadays.
> 
> By the way, starting with commit v4.8~83, strace no longer uses
> HAVE_LONG_LONG_OFF_T so the check could be safely removed.

that's good news.

> 
>> btw, while you're at it please consider merging this compatibility patch:
>> http://git.alpinelinux.org/cgit/aports/plain/main/strace/strace-musl.patch
>> (author is Timo Teräs <timo.teras at iki.fi>)
> 
> That patch is not obvious to say the least.  Could you or the author
> clarify the meaning of each change, please?
> 
> 

-    ) && defined(__GLIBC__)
+    ) && defined(__linux__)
  # include <sys/ptrace.h>

apparently without including this header, ptrace defines are missing.

a better check would be if configure checks if the header is available, 
instead of the entire ifdeffery.

-#if _LFS64_LARGEFILE
+#if defined _LFS64_LARGEFILE || defined HAVE_LONG_LONG_OFF_T
  int
  sys_truncate64(struct tcb *tcp)

musl has off_t always as long long (and obviously isnt glibc).
without this patch, we get a link error that sys_truncate64 is missing.
so maybe it would be better if configure checks whether sizeof(off_t) == 
sizeof(int64_t).

-#if _LFS64_LARGEFILE
+#if defined _LFS64_LARGEFILE || defined HAVE_LONG_LONG_OFF_T
  int
  sys_getdents64(struct tcb *tcp)
likewise.

  /* Under Linux these are enums so we can't test for them with ifdef. */
+#if !defined(IPPROTO_MAX)
not under linux, but under glibc. musl defines those as macros.

-			tprintf(", { %d }", p.__sched_priority);
+			tprintf(", { %d }", p.sched_priority);
accessing a glibc internal symbol.

-		execl(_PATH_BSHELL, "sh", "-c", command, NULL);
+		execl(_PATH_BSHELL, "sh", "-c", command, (void*) 0);
NULL is not a valid sentinel for execl.
execl expects char* parameters, not (void*) 0, (char) 0, 0L, 0LL, (enum 
foobar) 0 or whatever the C standard allows it to be.
citation of the C standard is available here: http://ewontfix.com/11/

this is not a build error but a warning fix.
the proper fix would be (char*) 0.
apparently the author just wanted to shut up gcc.

-#if defined(AARCH64)
-# include <asm/ptrace.h>
-#endif
-
-#if defined(XTENSA)
-# include <asm/ptrace.h>
-#endif
+#include <asm/ptrace.h>
seems asm/ptrace.h is required anyway to not get a build error.
and on glibc/uclibc, including it unconditionally doesnt hurt as well.

+#if defined(__GLIBC__)
  			tprintf("{%d}", sev._sigev_un._pad[0]);
+#else
+			tprintf("{%d}", (int) sev.__pad[0]);
+#endif
another glibc implementation specific internal struct member.
since the musl one is __ prefixed as well, i wonder if there's a clean 
fix for this. i suspect though that the __pad macro was added for glibc
compat, so sev.__pad[0] may work on all linux libcs.

thanks,
--JS

P.S. the musl tarball comes with a musl-gcc wrapper which makes it easy 
to test.