I think undefined symbols should be skipped while finding symbols from dynamic symbol tables in _dl_addr() in elf/dl-addr.c. Otherwise, it returns wrong symbol names in rare cases. Here is the code for reproducing the problem. I tested it on Debian GNU/Linux sarge on a x86_32 machine (glibc 2.3.2) but looked like the latest glibc in CVS had the same problem. % cat main.c #include <stdio.h> #define __USE_GNU 1 #include <dlfcn.h> extern void foo(); static void dummy_func() {} int main() { Dl_info info; void *p = &foo; dladdr(dummy_func, &info); if (info.dli_sname) { printf("%s\n", info.dli_sname); } return 0; } % cat foo.c void foo() { volatile int a = 0; // Make this function big. ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; ++a; } % gcc -shared -o foo.so foo.c % gcc -rdynamic main.c ./foo.so -ldl % ./a.out foo Here, "foo" shouldn't be printed. % readelf --symbols a.out | egrep 'foo|dummy_func' 13: 080484d8 505 FUNC GLOBAL DEFAULT UND foo 83: 08048594 5 FUNC LOCAL DEFAULT 12 dummy_func 104: 080484d8 505 FUNC GLOBAL DEFAULT UND foo "foo" was picked since the following condition satisfied. 0x080484d8 (foo) <= 0x08048594 (dummmy_func) < 0x80486d1 (0x080484d8 + 505) I haven't tested but I guess the following patch would solve the problem. % diff -u elf/dl-addr.c.orig elf/dl-addr.c --- elf/dl-addr.c.orig 2006-05-22 23:16:21.000000000 +0900 +++ elf/dl-addr.c 2006-05-22 23:16:42.000000000 +0900 @@ -92,6 +92,7 @@ #if defined USE_TLS && ELFW(ST_TYPE) (symtab->st_info) != STT_TLS #endif + && symtab->st_shndx != SHN_UNDEF && DL_ADDR_SYM_MATCH (match, symtab, matchsym, addr) && symtab->st_name < strtabsize) matchsym = (ElfW(Sym) *) symtab;
Correct patch, I applied it to CVS.
(In reply to comment #1) > Correct patch, I applied it to CVS. Thank you for fixing the bug. I just realized that completly skipping undefined references could be a problem. Consider the following code. #include <stdio.h> int main() { printf("%p\n", &printf); return 0; } If I compile it without -fPIC on my machine (Debian GNU/Linux sarge on a x86_32 machine), &printf points to a memory address in PLT. % gcc test.c % ./a.out 0x8048288 % readelf -S a.out |grep " \.plt" [11] .plt PROGBITS 08048270 000270 000030 04 AX 0 0 4 The binary contains the undefined symbols for printf both in .dysym and .symtab sections and the values of the undefined symbols correspond to the memory address in PLT (0x8048288). % readelf --symbols a.out |grep printf 2: 08048288 57 FUNC GLOBAL DEFAULT UND printf@GLIBC_2.0 (2) 99: 08048288 57 FUNC GLOBAL DEFAULT UND printf@@GLIBC_2.0 Hence, to support dladdr(&printf, info) for non-pic binaries, probably we need to allow the condition something like this as well. (ELFW(ST_TYPE) (symtab->st_info) == STT_FUNC && symtab->st_shndx == SHN_UNDEF && match->l_addr + symtab->st_value == addr) On the other hand, if I compile the same code with -fPIC, &printf points to the printf definition in the map of glibc. There is no problem in this case for doing dladdr(&printf, info). % gcc -fPIC test.c % ./a.out 0x4006e8e0 % ldd a.out libc.so.6 => /lib/libc.so.6 (0x4001f000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000) % readelf --symbols /lib/libc.so.6|grep " printf" 157: 0004f8e0 57 FUNC GLOBAL DEFAULT 11 printf@@GLIBC_2.0 % python >>> "%x" % (0x4001f000 + 0x0004f8e0) '4006e8e0'
I adjusted the patch.