lkml.org 
[lkml]   [2020]   [Jan]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v8 3/3] test: Add test for pidfd getfd
On Sun, Jan 05, 2020 at 03:20:23PM +0100, Christian Brauner wrote:
> On Fri, Jan 03, 2020 at 08:29:28AM -0800, Sargun Dhillon wrote:
> > +static int sys_pidfd_getfd(int pidfd, int fd, int flags)
> > +{
> > + return syscall(__NR_pidfd_getfd, pidfd, fd, flags);
> > +}
>
> I think you can move this to the pidfd.h header as:
>
> static inline int sys_pidfd_getfd(int pidfd, int fd, int flags)
> {
> return syscall(__NR_pidfd_getfd, pidfd, fd, flags);
> }
>
> Note, this also needs an
>
> #ifndef __NR_pidfd_getfd
> __NR_pidfd_getfd -1
> #endif
> so that compilation doesn't fail.
>
I'll go ahead and move this into pidfd.h, and follow the pattern there. I
don't think it's worth checking if each time the return code is ENOSYS.

Does it make sense to add something like:
#ifdef __NR_pidfd_getfd
TEST_HARNESS_MAIN
#else
int main(void)
{
fprintf(stderr, "pidfd_getfd syscall not supported\n");
return KSFT_SKIP;
}
#endif

to short-circuit the entire test suite?


> > +
> > +static int sys_memfd_create(const char *name, unsigned int flags)
> > +{
> > + return syscall(__NR_memfd_create, name, flags);
> > +}
> > +
> > +static int __child(int sk, int memfd)
> > +{
> > + int ret;
> > + char buf;
> > +
> > + /*
> > + * Ensure we don't leave around a bunch of orphaned children if our
> > + * tests fail.
> > + */
> > + ret = prctl(PR_SET_PDEATHSIG, SIGKILL);
> > + if (ret) {
> > + fprintf(stderr, "%s: Child could not set DEATHSIG\n",
> > + strerror(errno));
> > + return EXIT_FAILURE;
>
> return -1
>
> > + }
> > +
> > + ret = send(sk, &memfd, sizeof(memfd), 0);
> > + if (ret != sizeof(memfd)) {
> > + fprintf(stderr, "%s: Child failed to send fd number\n",
> > + strerror(errno));
> > + return EXIT_FAILURE;
>
> return -1
>
> > + }
> > +
> > + while ((ret = recv(sk, &buf, sizeof(buf), 0)) > 0) {
> > + if (buf == 'P') {
> > + ret = prctl(PR_SET_DUMPABLE, 0);
> > + if (ret < 0) {
> > + fprintf(stderr,
> > + "%s: Child failed to disable ptrace\n",
> > + strerror(errno));
> > + return EXIT_FAILURE;
>
> return -1
>
> > + }
> > + } else {
> > + fprintf(stderr, "Child received unknown command %c\n",
> > + buf);
> > + return EXIT_FAILURE;
>
> return -1
>
> > + }
> > + ret = send(sk, &buf, sizeof(buf), 0);
> > + if (ret != 1) {
> > + fprintf(stderr, "%s: Child failed to ack\n",
> > + strerror(errno));
> > + return EXIT_FAILURE;
>
> return -1
>
> > + }
> > + }
> > +
> > + if (ret < 0) {
> > + fprintf(stderr, "%s: Child failed to read from socket\n",
> > + strerror(errno));
>
> Is this intentional that this is no failure?
>
My thought here, is the only case where this should happen is if the "ptrace
command" was not properly "transmitted", and the ptrace test itself would fail.

I can add an explicit exit failure here.

> > + }
> > +
> > + return EXIT_SUCCESS;
>
> return 0
>
> > +}
> > +
> > +static int child(int sk)
> > +{
> > + int memfd, ret;
> > +
> > + memfd = sys_memfd_create("test", 0);
> > + if (memfd < 0) {
> > + fprintf(stderr, "%s: Child could not create memfd\n",
> > + strerror(errno));
> > + ret = EXIT_FAILURE;
>
> ret = -1;
>
> > + } else {
> > + ret = __child(sk, memfd);
> > + close(memfd);
> > + }
> > +
> > + close(sk);
> > + return ret;
> > +}
> > +
> > +FIXTURE(child)
> > +{
> > + pid_t pid;
> > + int pidfd, sk, remote_fd;
> > +};
> > +
> > +FIXTURE_SETUP(child)
> > +{
> > + int ret, sk_pair[2];
> > +
> > + ASSERT_EQ(0, socketpair(PF_LOCAL, SOCK_SEQPACKET, 0, sk_pair))
> > + {
> > + TH_LOG("%s: failed to create socketpair", strerror(errno));
> > + }
> > + self->sk = sk_pair[0];
> > +
> > + self->pid = fork();
> > + ASSERT_GE(self->pid, 0);
> > +
> > + if (self->pid == 0) {
> > + close(sk_pair[0]);
> > + exit(child(sk_pair[1]));
>
> if (child(sk_pair[1]))
> _exit(EXIT_FAILURE);
> _exit(EXIT_SUCCESS);
>
> I would like to only use exit macros where one actually calls
> {_}exit()s. It makes the logic easier to follow and ensures that one
> doesn't accidently do an exit(-21345) or something (e.g. when adding new
> code).
>
> > + }
> > +
> > + close(sk_pair[1]);
> > +
> > + self->pidfd = sys_pidfd_open(self->pid, 0);
> > + ASSERT_GE(self->pidfd, 0);
> > +
> > + /*
> > + * Wait for the child to complete setup. It'll send the remote memfd's
> > + * number when ready.
> > + */
> > + ret = recv(sk_pair[0], &self->remote_fd, sizeof(self->remote_fd), 0);
> > + ASSERT_EQ(sizeof(self->remote_fd), ret);
> > +}
> > +
> > +FIXTURE_TEARDOWN(child)
> > +{
> > + int status;
> > +
> > + EXPECT_EQ(0, close(self->pidfd));
> > + EXPECT_EQ(0, close(self->sk));
> > +
> > + EXPECT_EQ(waitpid(self->pid, &status, 0), self->pid);
> > + EXPECT_EQ(true, WIFEXITED(status));
> > + EXPECT_EQ(0, WEXITSTATUS(status));
> > +}
> > +
> > +TEST_F(child, disable_ptrace)
> > +{
> > + int uid, fd;
> > + char c;
> > +
> > + /*
> > + * Turn into nobody if we're root, to avoid CAP_SYS_PTRACE
> > + *
> > + * The tests should run in their own process, so even this test fails,
> > + * it shouldn't result in subsequent tests failing.
> > + */
> > + uid = getuid();
> > + if (uid == 0)
> > + ASSERT_EQ(0, seteuid(USHRT_MAX));
>
> Hm, isn't it safer to do 65535 explicitly? Since USHRT_MAX can
> technically be greater than 65535.
>
I borrowed this from the BPF tests. I can hardcode something like:
#define NOBODY_UID 65535
and setuid to that, if you think it's safer?

> > +
> > + ASSERT_EQ(1, send(self->sk, "P", 1, 0));
> > + ASSERT_EQ(1, recv(self->sk, &c, 1, 0));
> > +
> > + fd = sys_pidfd_getfd(self->pidfd, self->remote_fd, 0);
> > + EXPECT_EQ(-1, fd);
> > + EXPECT_EQ(EPERM, errno);
> > +
> > + if (uid == 0)
> > + ASSERT_EQ(0, seteuid(0));
> > +}
> > +
> > +TEST_F(child, fetch_fd)
> > +{
> > + int fd, ret;
> > +
> > + fd = sys_pidfd_getfd(self->pidfd, self->remote_fd, 0);
> > + ASSERT_GE(fd, 0);
> > +
> > + EXPECT_EQ(0, sys_kcmp(getpid(), self->pid, KCMP_FILE, fd, self->remote_fd));
>
> So most of these tests seem to take place when the child has already
> called exit() - or at least it's very likely that the child has already
> called exit() - and remains a zombie. That's not ideal because
> that's not the common scenario/use-case. Usually the task of which we
> want to get an fd will be alive. Also, if the child has already called
> exit(), by the time it returns to userspace it should have already
> called exit_files() and so I wonder whether this test would fail if it's
> run after the child has exited. Maybe I'm missing something here... Is
> there some ordering enforced by TEST_F()?
Yeah, I think perhaps I was being too clever.
The timeline roughly goes something like this:

# Fixture bringup
[parent] creates socket_pair
[parent] forks, and passes pair down to child
[parent] waits to read sizeof(int) from the sk_pair
[child] creates memfd
[__child] sends local memfd number to parent via sk_pair
[__child] waits to read from sk_pair
[parent] reads remote memfd number from socket
# Test
[parent] performs tests
# Fixture teardown
[parent] closes sk_pair
[__child] reads 0 from recv on sk_pair, implies the other end is closed
[__child] Returns / exits 0
[parent] Reaps child / reads exit code

---
The one case where this is not true, is if the parent sends 'P' to the sk pair,
it triggers setting PR_SET_DUMPABLE to 0, and then resumes waiting for the fd to
close.

Maybe I'm being too clever? Instead, the alternative was to send explicit stop /
start messages across the sk_pair, but that got kind of ugly. Do you have a
better suggestion?

>
> Also, what does self->pid point to? The fd of the already exited child?
It's just the pid of the child. pidfd is the fd of the (unexited) child.

\
 
 \ /
  Last update: 2020-01-05 20:09    [W:0.080 / U:9.776 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site