I have read Russel Coker’s nice article on identifying use of thread unsafe functions. This reminded me of a script I wrote a long time ago that is part of GNU SASL‘s regression suite: threadsafety.
As you can see, my script looks for functions mentioned in the latest POSIX specification as being thread unsafe. In the last POSIX release, they actually removed some older interfaces (e.g., gethostbyname) so the script also checks for thread-unsafe functions mentioned in one older POSIX specification.
Russel’s approach is to look for man pages of functions ending with _r and labeling the non-_r-function as a thread unsafe function. Russel’s and my approach are quite different, so I wanted to compare the results. There is potential for me to add more functions to search for. I still want to preserve my approach of explicitly listing known thread unsafe functions, though.
Running Russel’s command, I get a list of functions that my script catches that Russel’s doesn’t, and vice versa. For reference, the functions that my script catches that Russel’s doesn’t are:
basename catgets dbm_clearerr dbm_close dbm_delete dbm_error dbm_fetch dbm_firstkey dbm_nextkey dbm_open dbm_store dirname dlerror endgrent endpwent endutxent ftw gcvt getc_unlocked getchar_unlocked getenv getopt getutxent getutxid getutxline inet_ntoa l64a lgamma lgammaf lgammal localeconv nftw nl_langinfo putc_unlocked putchar_unlocked putenv pututxline setenv setgrent setpwent setutxent strsignal system unsetenv wcstombs wctomb
The list contains lgamma, lgammaf, and lgammal which are all excluded by Russel’s command. I don’t understand why — according to the man page, the functions uses a global variable for sign, which doesn’t seem thread safe. So it seems right to include them?
What’s more interesting (for me) is the list of functions that Russel’s script catches that my script currently doesn’t. Here is the list:
erand48 ether_aton ether_ntoa fgetgrent fgetpwent fgetspent getaliasbyname getaliasent gethostbyname2 getmntent getnetgrent getrpcbyname getrpcbynumber getrpcent getspent getspnam getutent getutid getutline initstate jrand48 lcong48 nrand48 qecvt qfcvt random seed48 setstate sgetspent srand48 srandom tmpnam
I started looking into each function. For erand48 there is a erand48_r function in glibc, and the former does indeed seem to use a global variable. However, as far as I can tell from the POSIX specification, erand48 should be thread safe. So I filed a glibc bug about it. The same concern may hold for jrand48, lcong48, nrand48, seed48, and srand48.
I noticed that initstate, random, setstate, and srandom are defined by latest POSIX, but not mentioned as a thread-unsafe functions. Possibly a bug in the POSIX specification?
I also noticed that I had missed to include tmpnam even though it is mentioned separately in the POSIX link.
The rest of the functions are not documented by POSIX, and presumably thread unsafe (although I didn’t read the man page or source code for each of them).
In the end, I ended up adding several new functions to check for. The latest script is always available from:
http://git.savannah.gnu.org/cgit/gsasl.git/tree/tests/threadsafety
So, finally, did the updated script catch any use of thread-unsafe functions in GNU SASL? Nope.