| Ticket UUID: | 1224888 | ||
| Title: | possible to have an empty info nameofexecutable | ||
| Type: | Bug | Created on: | 2005-06-21 14:29:27 |
| Submitter: | davidw | Assigned to: | dgp |
| Subsystem: | 36. Pathname Management | Severity: | Minor |
| Priority: | 5 Medium | Last modified: | 2017-11-17 04:56:48 |
| Status: | Open | Closed by: | nobody |
| Resolution: | Remind | Closed on: | |
| Version: | obsolete: 8.4.19 | ||
| Description: | ||||
How to reproduce:
1) Make a local copy of tclsh called something
*besides* tclsh or anything else that is in your path.
Make sure that '.' is not in your path.
2) Compile this, using the right paths and name of the
executable for your environment:
#include <unistd.h>
int main(int argc, char *argv[], char *envp[]) {
char *myargv[] = {"mytclsh", NULL};
execve("/home/davidw/workshop/protect/mytclsh",
myargv, envp);
}
3) Run it, try 'info name' - it's empty.
Here's why:
5454 execve("/home/davidw/workshop/protect/mytclsh",
["mytclsh"], [/* 29 vars *
/]) = 0
argv0 has no / in it, and mytclsh is not in PATH, so we
hit tclUnixFile.c line 123, and thus have no info
nameofexecutable.
I found this bug while trying to get starpacks working
as CGI's under Apache. The execve program above is a
recreation of how Apache launches CGI's.
Thanks,
Dave
| ||||
| User Comments: | ||||
davidw added on 2005-06-21 21:43:36:
Logged In: YES
user_id=240
Quick patch inline:
This seems to fix the problem, but I don't know if it causes
others.
--- tclUnixFile.c 1 Dec 2004 23:18:55 -0000 1.44
+++ tclUnixFile.c 21 Jun 2005 14:42:29 -0000
@@ -120,8 +120,12 @@
p++;
}
}
- TclSetObjNameOfExecutable(Tcl_NewObj(), NULL);
- goto done;
+ /* It's not in the path, so just use it in the current
working
+ * directory. */
+ Tcl_DStringSetLength(&buffer, 0);
+ name = Tcl_DStringAppend(&buffer, "./", -1);
+ name = Tcl_DStringAppend(&buffer, argv0, -1);
+ goto gotName;
dgp added on 2005-06-21 22:24:12:
Logged In: YES user_id=80530 hmmm.... the proposed patch takes effect only after the PATH has been searched. So whenever a program by the same truncated name *is* on the PATH, Tcl will find the wrong one. Seems that Apache really ought to pass the absolute name in argv0. jcw added on 2005-06-21 22:25:38:
Logged In: YES user_id=1983 Great catch! Possible solutions: on Linux and systems which support /proc, you could get at the exe's own path via "/proc/self/exe" which is a symlink pointing to the right file. Another way out might be http://autopackage.org/docs/binreloc/ As workaround, I s'pose one could use full pathnames in apache? -jcw dgp added on 2005-06-21 22:28:39:
Logged In: YES user_id=80530 and on the other side of things, TclpFindExecutable() comments say [info nameof] will be set to NULL if the executable can't be found. Won't this patch prevent that by always supplying ./foo as a (possibly false) last resort ? gwlester added on 2005-06-21 22:52:56:
Logged In: YES user_id=121110 While this bug appears to be a Unix only, the correct behavior needs to be verified on Windows and, if applicable, MacOsX. davidw added on 2005-06-21 23:35:13:
Logged In: YES user_id=240 The patch was a quick hack and the objections to it are quite valid... I'm sure we can come up with something better. However, even if Apache isn't doing things quite right, I don't think it's egregiously wrong, either, and I think that Tcl ought to be "liberal with what it accepts" as a general principle. How does this sound for an algorithm: 1) If argv0 has slashes in it, we are ok, we just use that. 2) If there are no slashes, and a file that matches pwd+argv0 exists and is executable, use that. 3) Search the PATH. 4) Can't figure anything out, so just make it NULL. dgp added on 2005-06-22 00:23:25:
Logged In: YES user_id=80530 the thing I dislike about that is its implicit assumption that "." is first on the PATH. For executables actually launched from the shell (still the vast majority, right?), that can give false answers. If we have no choice but to give a false answer to one or the other, I'd give it to Apache. patthoyts added on 2005-06-22 04:30:16:
Logged In: YES user_id=202636 To reply to gwlester's point - this isn't really an issue on Windows as on windows there is a reliable way to find the path to the binary using GetModuleFilename() and this is used in win/tclAppInit.c to set argv[0] appropriately. dkf added on 2005-06-22 05:07:19:
Logged In: YES
user_id=79902
I'd be tempted to put the search of the PATH first too.
OTOH, doing a readlink("/proc/self/exe") would be better on
those OSes that support it (alas, neither Solaris or Irix as
far as I can see) though if we do that, it perhaps ought to
be a feature searched for using configure? (We'd need a
fallback anyway in case /proc isn't mounted for some odd
reason.)
dkf added on 2005-06-22 05:45:38:
Logged In: YES user_id=79902 FreeBSD (of uncertain vintage but probably fairly old) vintage supports the following readlink() target: /proc/[pid]/file (no /proc/self, different name inside) das added on 2005-06-23 15:38:17:
Logged In: YES user_id=90580 Mac OS X 10.2 and later has #include <mach-o/dyld.h> int _NSGetExecutablePath(char *buf, unsigned long *bufsize); davidw added on 2005-07-19 22:40:13:
Logged In: YES user_id=240 The apache folks seem to be of the opinion that that system behaves like it does as historical accident. However, they don't really regard it as a bug, and aren't about to change it (who knows who may depend on it at this point), so it doesn't look as if we can shuffle this problem off to them. rkeene added on 2009-03-21 00:59:58:
This is still a problem for me. I would like TclpFindExecutable() to be more robust and look for OS-specific ways to determine the executable name.
Something similar to:
const char *exepaths[] = { "/proc/self/exe", "/proc/%lu/exe", "/proc/%lu/file", "/proc/%lu/object/a.out" };
char pathbuf[PATH_MAX + 1], pathbuf_copy[PATH_MAX + 1];
int pathbuf_len;
int exefileexists, recursedepth, idx;
if (name == NULL || *name == '\0') {
/* Attempt to use "/proc/self/exe" to determine where we are */
for (idx = 0; idx < (sizeof(exepaths) / sizeof(*exepaths)); idx++) {
if (strchr(exepaths[idx], '%') == NULL) {
sprintf(pathbuf_copy, exepaths[idx]);
} else {
sprintf(pathbuf_copy, exepaths[idx], (unsigned long) getpid());
}
exefileexists = access(pathbuf_copy, F_OK);
if (exefileexists == 0) {
for (recursedepth = 0; recursedepth < 100; recursedepth++) {
pathbuf_len = readlink(pathbuf_copy, pathbuf, sizeof(pathbuf));
if (pathbuf_len < 0) {
break;
}
if (pathbuf_len >= sizeof(pathbuf) || pathbuf_len == 0) {
return;
}
memcpy(pathbuf_copy, pathbuf, pathbuf_len);
pathbuf_copy[pathbuf_len] = '\0';
}
name = pathbuf_copy;
goto gotName;
}
}
}
in unix/tclUnixFile.c's TclpFindExecutable if it can't find the executable using the current methods.
stevenaaus added on 2011-10-11 08:39:55:
Hmmm... Scid vs. PC now suffers from this problem with wish8.6b2. We compile our own custom wish - tkscid - and install it into local/bin (or bin - they both bug out), and from it [info nameofexecutable] just returns {}. Never seen this bug before 8.6b2, and not sure what to do.
dkf added on 2011-10-12 16:49:56:
Doing a readlink() on /proc/[pid]/exe is elegant, but won't work on OSX. There, the only reliable method I've found for getting the executable path from the PID is by calling lsof in a subprocess: lsof -a -p [pid] -F n -d txt Then you need the first line that starts with 'n', and you need all of that line except the first character. Ugh. Works though. (Poking in the normal system process information doesn't work; that gives you the process name without the path to it.) stevenaaus added on 2011-10-12 17:43:57:
> Doing a readlink() on /proc/[pid]/exe is elegant This should come in handy. Cheers Just clarifying - i never had this issue with any version of 8.4, 8.5 or 8.6a1, so perhaps it is a regression. rkeene added on 2011-10-12 19:19:43:
In KitCreator I also attempt to use dladdr() if an acceptable implementation is found. Mac OS X appears to support this ( http://developer.apple.com/library/ios/#documentation/system/conceptual/manpages_iphoneos/man3/dladdr.3.html ) http://kitcreator.rkeene.org/fossil/artifact?name=0cfdc14cf7c88b31081e9e0ac689a710a26d6375&ln=294-301 dkf added on 2011-10-13 17:54:23:
I also like the dladdr() solution, though the name it yields is not necessarily normalized. Still, that's a practical solution (and would be easy to check for with autoconf).
My little test code:
#include <dlfcn.h>
#include <stdio.h>
int main(int argc,char**argv) {
Dl_info info;
if (!dladdr(main,&info)) {
printf("not found\n");
return 1;
}
printf("\"%s\"\n",info.dli_fname);
return 0;
}
dkf added on 2011-10-13 17:58:08:
dgp: What's the position on normalization of [info nameofexecutable]? (dladdr() is present on OSX, Solaris and Linux for sure: I've checked.) dkf added on 2011-10-13 19:17:23:
I've committed something on the bug-1224888 branch. It needs more work (especially detection of the validity of the magic!) but it would be interesting to see if it works. https://core.tcl.tk/tcl/timeline?r=bug-1224888 dgp added on 2011-10-14 01:47:45:
The branch produces segfault in test unixFile-1.1 This is on a platform where.. checking for dladdr... no dgp added on 2011-10-14 01:51:55:
Program received signal SIGSEGV, Segmentation fault.
0x0000003716e7288e in free () from /lib64/libc.so.6
(gdb) bt
#0 0x0000003716e7288e in free () from /lib64/libc.so.6
#1 0x00000000004cb7d2 in TclpFree (
cp=0x2f6c74692f3a6e69 <Address 0x2f6c74692f3a6e69 out of bounds>)
at /home/dgp/fossil/tcl8.4/unix/../generic/tclAlloc.c:710
#2 0x0000000000429ea5 in Tcl_Free (
ptr=0x2f6c74692f3a6e69 <Address 0x2f6c74692f3a6e69 out of bounds>)
at /home/dgp/fossil/tcl8.4/unix/../generic/tclCkalloc.c:1167
#3 0x00000000004a86d6 in Tcl_DStringFree (dsPtr=0x7fffffff7f10)
at /home/dgp/fossil/tcl8.4/unix/../generic/tclUtil.c:1698
#4 0x00000000004b6788 in TclpFindExecutable (argv0=0x783ca0 "junk")
at /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:225
#5 0x0000000000454329 in Tcl_FindExecutable (argv0=0x783ca0 "junk")
at /home/dgp/fossil/tcl8.4/unix/../generic/tclEncoding.c:1228
#6 0x0000000000420239 in TestfindexecutableCmd (clientData=0x0,
interp=0x700030, argc=2, argv=0x7fffffff8320)
at /home/dgp/fossil/tcl8.4/unix/tclUnixTest.c:465
dkf added on 2011-10-14 04:04:51:
Thank you for testing; I'd put the code two lines too high. :-) dgp added on 2011-10-18 00:34:17:
==== unixFile-1.1 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) ""
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
/local/tmp/dgp/fossil/tcl8.4/unix/junk
==== unixFile-1.1 FAILED
==== unixFile-1.2 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) "/dummy"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
==== unixFile-1.2 FAILED
==== unixFile-1.3 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) "/dummy:[pwd]"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
/local/tmp/dgp/fossil/tcl8.4/unix/junk
==== unixFile-1.3 FAILED
==== unixFile-1.4 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) "/dummy:"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
/local/tmp/dgp/fossil/tcl8.4/unix/junk
==== unixFile-1.4 FAILED
==== unixFile-1.5 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) "/dummy:/dummy"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
==== unixFile-1.5 FAILED
==== unixFile-1.6 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) "/dummy::/dummy"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
/local/tmp/dgp/fossil/tcl8.4/unix/junk
==== unixFile-1.6 FAILED
==== unixFile-1.7 Tcl_FindExecutable FAILED
==== Contents of test case:
set env(PATH) ":/dummy"
testfindexecutable junk
---- Result was:
/local/tmp/dgp/fossil/tcl8.4/unix/tcltest"j
---- Result should have been (exact matching):
/local/tmp/dgp/fossil/tcl8.4/unix/junk
==== unixFile-1.7 FAILED
dgp added on 2011-11-01 00:14:33:
On another of my systems, this branch fails to build: /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c: In function ‘TclpFindExecutable’: /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:64: error: ‘Dl_info’ undeclared (first use in this function) /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:64: error: (Each undeclared identifier is reported only once /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:64: error: for each function it appears in.) /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:64: error: expected ‘;’ before ‘dlinfoBuffer’ /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:100: error: ‘RTLD_DEFAULT’ undeclared (first use in this function) /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:104: warning: implicit declaration of function ‘dladdr’ /home/dgp/fossil/tcl8.4/unix/tclUnixFile.c:104: error: ‘dlinfoBuffer’ undeclared (first use in this function) nijtmans added on 2011-11-01 02:43:11:
>On another of my systems, this branch fails >to build: Probably because configure.in is changed but configure is not re-generated. done now. Does that help? dgp added on 2011-11-01 03:16:32:
thanks; now back to the unixFile-1.* fails. dgp added on 2015-01-13 14:34:06:
Can't get anywhere with rebasing the branch to something live. Everything I try fails the test suite catastrophically with apparent mem corruption in the name of the tcltest executable. I'll happily test when someone else gets the branch resurrected. jan.nijtmans added on 2015-01-22 20:04:28:
Please try again ;-) The reason probably was that "main" was made MODULE_SCOPE, so dlsym() cannot find it any more ..... dgp added on 2015-01-22 21:02:43:
No improvement.... ... Tests began at Thu Jan 22 16:02:41 EST 2015 aaa_exit.test Test file error: couldn't execute "/local/tmp/dgp/fossil/tcl-only/unix/tcltestQÿ": no such file or directory append.test Test file error: couldn't execute "/local/tmp/dgp/fossil/tcl-only/unix/tcltestQÿ": no such file or directory ... drh added on 2015-02-04 18:12:23:
Adding a comment merely to test Fossil ticket handle. Ignore this comment. dkf added on 2015-03-17 20:32:10:
It's going to be really hard to figure out what is wrong without knowing the path through the tangle of code that is being taken. It's an awful tangle right now. I've added support for _NSGetExecutablePath on OSX to the branch. A bunch of tests in unixFile.test now fail because Tcl ceases to be fool-able as to its executable name. The name isn't (always) a normalized one; will that matter? Anyone digging into the depths of this should start with http://stackoverflow.com/questions/1023306/finding-current-executables-path-without-proc-self-exe which summarizes a lot of the key techniques on different platforms. aspect added on 2017-11-16 12:55:34:
The errors below come from readlink's silly failure to terminate its result. I've pushed a patch that handles it properly and passes the test suite. The implementation in TclpFindExecutable looks pretty good, but it could do with being broken out into individually-testable parts. Looking at the linked SO discussion, it seems all we're missing for "full" coverage is the Solaris getexecname() and FreeBSD sysctl approaches. Both these platforms apparently "often" have /proc/ with one of the paths included in exenames[]. I think this branch (once brought up to date with trunk) should eliminate the need for kitcreator's workaround. I know we have people who can test on FreeBSD (gahr?) but how about Solaris/Illumos? Anyone have access to one of those? aspect added on 2017-11-17 04:56:48:
merged trunk. Correction to previous comment: unixFile-1.x are FAILing, as expected: these tests prove that TclpFindExecutable can be tricked by a bogus argv[0], which is the point of this ticket. Updating these tests is the obvious step, but as mentioned below a test suite which exercises the different paths on various platforms would be a Good Idea. | ||||
