sed segfault after commit 9c47c43e0736 with FEATURE_CLEAN_UP

[Denys Vlasenko]

On Mon, Sep 24, 2018 at 4:36 PM Thomas De Schampheleire
<patrickdepinguin at gmail.com> wrote:
> If FEATURE_CLEAN_UP is enabled, following reproduction test case
> causes a segfault:
>
> +testing "sed uses previous regexp test2" \
> +       "sed -n '/foo/ { //p }'" \
> +       "" \
> +       "" \
> +       "q\nw\ne\nr\n"
> +
>
> (note: the test case passes because the output is correct; it seems
> the test suite does not catch segfaults)
>
> The issue is bisected to commit 9c47c43e0736 which adds support for
> '//' to indicate the previous regex. The change in question is:
>
> -               temp = copy_parsing_escapes(pos, next);
> -               *regex = xzalloc(sizeof(regex_t));
> -               xregcomp(*regex, temp, G.regex_type);
> -               free(temp);
> +               if (next != 0) {
> +                       temp = copy_parsing_escapes(pos, next);
> +                       G.previous_regex_ptr = *regex =
> xzalloc(sizeof(regex_t));
> +                       xregcomp(*regex, temp, G.regex_type);
> +                       free(temp);
> +               } else {
> +                       *regex = G.previous_regex_ptr;
> +                       if (!G.previous_regex_ptr)
> +                               bb_error_msg_and_die("no previous regexp");
> +               }
>
>
> In the original code, *regex was always a newly allocated pointer.
> In the new code, there are cases where *regex (mapping to
> sed_cmd.beg_match in the caller) is not a newly allocated pointer, but
> the same as the previously allocated one.
>
> When sed finishes, and sed_free_and_close_stuff() is called, there is
> an iteration over the different sed_cmd structures. Following code:
>
>         if (sed_cmd->beg_match) {
>             regfree(sed_cmd->beg_match);
>             free(sed_cmd->beg_match);
>         }
>
> will work fine in the first iteration that touches the repeated
> pointer, but will fail in 'regfree' on the next iteration, where the
> beg_match pointer points to the same value as was already freed.
>
> Backtrace is:
> Program received signal SIGSEGV, Segmentation fault.
> free_token (node=0x198) at regcomp.c:3808
> 3808    regcomp.c: No such file or directory.
> (gdb) bt
> #0  free_token (node=0x198) at regcomp.c:3808
> #1  0xf6fe474c in free_dfa_content (dfa=0x712220) at regcomp.c:598
> #2  0xf6ff1200 in __regfree (preg=0x712068) at regcomp.c:647
> #3  0x000d53dc in sed_free_and_close_stuff () at editors/sed.c:184
> #4  0xf6f6be94 in __run_exit_handlers (status=0x0, listp=0xf70744b4
> <__exit_funcs>,
>     run_list_atexit=run_list_atexit at entry=0x1) at exit.c:82
> #5  0xf6f6bee8 in __GI_exit (status=<optimized out>) at exit.c:104
> #6  0x0001a194 in xfunc_die () at libbb/xfunc_die.c:20
> #7  0x000196b0 in run_applet_no_and_exit (applet_no=0xa9,
> name=0xfff7ce6a "sed", argv=0xfff7bf14)
>     at libbb/appletlib.c:952
> #8  0x00019718 in run_applet_and_exit (name=0xfff7ce6a "sed",
> argv=0xfff7bf14) at libbb/appletlib.c:968
> #9  0x000197fc in main (argc=0x4, argv=0xfff7bf14) at libbb/appletlib.c:1076
>
>
> I'm not sure how to fix this:
> * either keep a list in sed_free_and_close_stuff of pointers already freed
> * or create a separate list of pointers-to-free when allocating
> pointers rather than assuming that all values in beg_match need to be
> freed
> * or add a flag to indicate this special case in sed_cmd, which would
> need some more communication between get_address and its caller.
> * something else?

How about "don't free them at all"?