On Tuesday 01 December 2015 15:59:56 Mateusz Guzik wrote:
On Thu, Nov 19, 2015 at 05:38:25PM +0100, Pino Toscano wrote:
When running commands in the mounted guest (using the "command" API, and APIs based on it), provide the /dev/null from the appliance as open fd for stdin. Commands usually assume stdin is open if they didn't close it explicitly, so this should avoid crashes or misbehavings due to that.
This does not look right.
+ /* Provide /dev/null as stdin for the command, since we want + * to make sure processes have an open stdin, and it is not + * possible to rely on the guest to provide it (Linux guests + * get /dev dynamically populated at runtime by udev). + */ + fd = open ("/dev/null", O_RDONLY|O_CLOEXEC); + if (fd == -1) { + reply_with_perror ("/dev/null"); + return NULL; + } +
I disagree with this (see below).
if (bind_mount (&bind_state) == -1) return NULL;
nit: this leaks the fd on error, but it may not matter much.
if (enable_network) { @@ -266,8 +279,10 @@ do_command (char *const *argv) return NULL; }
nit: same.
Both these leaks need to be fixed, thanks for reporting them.
+ flags = COMMAND_FLAG_CHROOT_COPY_FILE_TO_STDIN | fd; + CHROOT_IN; - r = commandv (&out, &err, (const char * const *) argv); + r = commandvf (&out, &err, flags, (const char * const *) argv); CHROOT_OUT;
free_bind_state (&bind_state);
According to the analysis in https://bugzilla.redhat.com/show_bug.cgi?id=1280029 the problem was that the target program was being executed in a chroot which did not have /dev/null populated, hence the open in commandrvf failed and the process was left without fd 0.
commandrvf does the following in the child: close (0); if (flag_copy_stdin) { dup2 (flag_copy_fd, STDIN_FILENO); } else { /* Set stdin to /dev/null (ignore failure) */ ignore_value (open ("/dev/null", O_RDONLY|O_CLOEXEC)); } [..] execvp (argv[0], (void *) argv);
First observation is that regardless of whether this open("/dev/null", ..) succeeds, there is no fd 0 in the process after execvp due to O_CLOEXEC.
Correct, so O_CLOEXEC could be dropped in this case.
So, chances are the chroot did have /dev/null, but the fd simply got closed.
No, the actual issue was that there was nothing at all in the /dev of guest, so open failed.
I would argue that /dev has to be at least partially populated for anything that gets executed in the chroot. At the very least special nodes like null, zero and {u,}random are needed.
We do not assume anything about guests, which could be Linux with a static /dev (rare these days), but also non-Linux guests, including Windows.
CHROOT_IN/OUT around commandvf are definitely problematic. chroot should be done in the child, which also removes the need to chroot out in the parent.
That could be another way to make the command* functions in the daemon safer, which wouldn't solve the issue of this email thread: even if you fork and then chroot, the guest has nothing in /dev, so you cannot open /dev/null at that point. This means you still need to carry on an open /dev/null from the appliance.
Assuming populated /dev is problematic/not feasible, at the very least the open(/dev/null) should be performed in the child just prior to chroot.
See above.
Current patch seems to work around shortcomings of the current API.
Side content: if (flag_copy_stdin) close (flag_copy_fd); waitpid (pid, NULL, 0); return -1;
This is done only in the main "while (quit < 2)" loop, and only on error there, which will return with -1.
but some lines below there is: if (flag_copy_stdin && close (flag_copy_fd) == -1) { perror ("close"); return -1; }
/* Get the exit status of the command. */ if (waitpid (pid, &r, 0) != pid) { perror ("waitpid"); return -1; }
close() does not return an error unless extraterrestial circumstances occur, and even then the fd is no longer in use by the process. As such, I would argue checking for errors here is not necessary. Note that prior sample does not check.
The code for flag_copy_stdin & flag_copy_fd is unrelated to the bit quoted above, see above for the explanation.
In case an error was returned, the code fails to waitpid() for the child.
This needs to be fixed indeed. Thanks, -- Pino Toscano