Especially its "next" member: this is what caused the crash in "slabtop:
Reset slab_list if get_slabinfo() fails." (if parse_slabinfo*() fails in
sscanf(), for example, then curr is set to NULL but it is already linked
into the "list" and its "next" member was never initialized).
In proc/slab.c, functions parse_slabinfo20() and parse_slabinfo11(),
sscanf() might overflow curr->name, because "String input conversions
store a terminating null byte ('\0') to mark the end of the input; the
maximum field width does not include this terminator."
Add one byte to name[] for this terminator.
Otherwise this can truncate sizes on 64-bit platforms, and is one of the
reasons the integer overflows in file2strvec() are exploitable at all.
Also: catch potential integer overflow in xstrdup() (should never
happen, but better safe than sorry), and use memcpy() instead of
strcpy() (faster).
Warnings:
- in glibc, realloc(ptr, 0) is equivalent to free(ptr), but not here,
because of the ++size;
- here, xstrdup() can return NULL (if str is NULL), which goes against
the idea of the xalloc wrappers.
We were tempted to call exit() or xerrx() in those cases, but decided
against it, because it might break things in unexpected places; TODO?
This can disclose information from the stack, but is unlikely to have a
security impact in the context of the procps utilities:
user@debian:~$ w 2>&1 | xxd
00000000: a03c 79b7 1420 6661 696c 6564 2074 6f20 .<y.. failed to
00000010: 616c 6c6f 6361 7465 2033 3232 3137 3439 allocate 3221749
00000020: 3738 3020 6279 7465 7320 6f66 206d 656d 780 bytes of mem
00000030: 6f72 79 ory
Do not memleak "copy" in case of an error.
Do not use "sizeof(converted)" in snprintf(), since "converted" is a
"char *" (luckily, 8 >= sizeof(char *)). Also, remove "sizeof(char)"
which is guaranteed to be 1 by the C standard, and replace 8 with 12,
which is enough to hold any stringified int and does not consume more
memory (in both cases, the glibc malloc()ates a minimum-sized chunk).
Right now, if read() returns less than 127 bytes (the most likely case),
the end of the "string" buf will contain garbage from the stack, because
buf is always null-terminated at a fixed offset 127. This is especially
bad because the next operation is a strrchr().
Also, make sure that the whole /proc/PID/stat file is read, otherwise
its parsing may be unsafe (the strrchr() may point into user-controlled
data, comm). This should never happen with the current file format (comm
is very short), but be safe, just in case.
First problem: saved_argc was used to calculate the size of the array,
but saved_argc was never initialized. This triggers an immediate heap-
based buffer overflow:
$ skill -c0 -c0 -c0 -c0
Segmentation fault (core dumped)
Second problem: saved_argc was not the upper bound anyway, because one
argument can ENLIST() several times (for example, in parse_namespaces())
and overflow the array as well.
Third problem: integer overflow of the size of the array.
No need to "pid_count++;" because "ENLIST(pid," does it already. Right
now this can trigger a heap-based buffer overflow.
Also, remove the unneeded "pid_count = 0;" (it is static, and
skillsnice_parse() is called only once; and the other *_count variables
are not initialized explicitly either).
The memmove() itself does not move the NULL-terminator, because nargs is
decremented first. Copy how skill_sig_option() does it: decrement nargs
last, and remove the "if (nargs - i)" (we are in "while (i < nargs)").
man getline: "If *lineptr is set to NULL and *n is set 0 before the
call, then getline() will allocate a buffer for storing the line. This
buffer should be freed by the user program even if getline() failed."
Note: unlike "size" and "omit_size", "path_alloc_size" is not multiplied
by "sizeof(struct el)" but the checks in grow_size() allow for a roughly
100MB path_alloc_size, which should be more than enough for readlink().
Do it explicitly instead of the implicit "longjmp() cannot cause 0 to be
returned. If longjmp() is invoked with a second argument of 0, 1 will be
returned instead."
This is one of the worst issues that we found: if the strlen() of one of
the cmdline arguments is greater than INT_MAX (it is possible), then the
"int bytes" could wrap around completely, back to a very large positive
int, and the next strncat() would be called with a huge number of
destination bytes (a stack-based buffer overflow).
Fortunately, every distribution that we checked compiles its procps
utilities with FORTIFY, and the fortified strncat() detects and aborts
the buffer overflow before it occurs.
This patch also fixes a secondary issue: the old "--bytes;" meant that
cmdline[sizeof (cmdline) - 2] was never written to if the while loop was
never entered; in the example below, "ff" is the uninitialized byte:
((exec -ca `python3 -c 'print("A" * 131000)'` /usr/bin/cat < /dev/zero) | sleep 60) &
pgrep -a -P "$!" 2>/dev/null | hexdump -C
00000000 31 32 34 36 30 20 41 41 41 41 41 41 41 41 41 41 |12460 AAAAAAAAAA|
00000010 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 41 |AAAAAAAAAAAAAAAA|
*
00001000 41 41 41 41 ff 0a 31 32 34 36 32 20 73 6c 65 65 |AAAA..12462 slee|
00001010 70 20 36 30 0a |p 60.|
Otherwise (for example), if the (undocumented) opt_echo is set, but not
opt_long, and not opt_longlong, and not opt_pattern, there is a call to
xstrdup(cmdoutput) but cmdoutput was never initialized:
sleep 60 & echo "$!" > pidfile
env -i LD_DEBUG=`perl -e 'print "A" x 131000'` pkill -e -c -F pidfile | xxd
...
000001c0: 4141 4141 4141 4141 4141 4141 4141 4141 AAAAAAAAAAAAAAAA
000001d0: 4141 4141 4141 4141 fcd4 e6bd e47f 206b AAAAAAAA...... k
000001e0: 696c 6c65 6420 2870 6964 2031 3230 3931 illed (pid 12091
000001f0: 290a 310a ).1.
[1]+ Terminated sleep 60
(the LD_DEBUG is just a trick to fill the initial stack with non-null
bytes, to show that there is uninitialized data from the stack in the
output; here, an address "fcd4 e6bd e47f")
Not exploitable (not under an attacker's control), but still a potential
non-security problem. Copied, fixed, and used the grow_size() macro from
pidof.c.
memset()ing task and subtask inside their loops prevents free_acquired()
(in readproc() and readtask()) from free()ing their contents (especially
cmdline and environ).
Our solution is not perfect, because we still memleak the very last
cmdline/environ, but select_procs() is called only once, so this is not
as bad as it sounds.
It would be better to leave subtask in its block and call
free_acquired() after the loop, but this function is static (not
exported).
The only other solution is to use freeproc(), but this means replacing
the stack task/subtask with xcalloc()s, thus changing a lot of code in
pgrep.c (to pointer accesses).
Hence this imperfect solution for now.
sig.c had this odd logic where on non-Hurd systems it would undefine
SIGLOST. Fine for Hurd or amd64 Linux systems. Bad for a sparc which
has SIGLOST defined *and* is not Hurd.
Just check its defined, its much simpler.
