[v5] Allow for unpriviledged nested containers

[resent as v5]

[Both versions were tested, with identical test results.  Rather than
delete my work, I leave it as an option in the source, in case it's
needed later.  I want to have my cake and eat it too ;-)]

If the build itself is run in a container, we may not be able to
fully set up a nested container for test-container testing.
Notably is the mounting of /proc, since it's critical that it
be mounted from within the same PID namespace as its users, and
thus cannot be bind mounted from outside the container like other
mounts.

This patch chooses to use the parent's PID namespace instead of
creating a new one, as this is more likely to be allowed.

Code for using a separate PID namespace is included as well, in case
in the future a test requires it.  In this configuration,
test-container may not be able to mount /proc but will run the test
anyway, since most containerized tests do not require /proc.  The few
that do may predicate that, and support for such is also added.

On 3/10/22 22:38, DJ Delorie via Libc-alpha wrote:
> 
> [resent as v5]
> 
> [Both versions were tested, with identical test results.  Rather than
> delete my work, I leave it as an option in the source, in case it's
> needed later.  I want to have my cake and eat it too ;-)]
> 
> If the build itself is run in a container, we may not be able to
> fully set up a nested container for test-container testing.
> Notably is the mounting of /proc, since it's critical that it
> be mounted from within the same PID namespace as its users, and
> thus cannot be bind mounted from outside the container like other
> mounts.
> 
> This patch chooses to use the parent's PID namespace instead of
> creating a new one, as this is more likely to be allowed.
> 
> Code for using a separate PID namespace is included as well, in case
> in the future a test requires it.  In this configuration,
> test-container may not be able to mount /proc but will run the test
> anyway, since most containerized tests do not require /proc.  The few
> that do may predicate that, and support for such is also added.

I dislike what we're doing here, but I don't see better solution.

Let me lay out my thinking here:

(a) The implementation has runtime requirements on the host OS and we don't
    express those requirements anywhere on a per-interface definition e.g.
    the need for proc to compute an answer for pthread_getattr_np for
    ATTR_FLAG_STACKADDR. When creating tests we look to specifically test
    such paths, and so in the test we now encode, via support_need_proc()
    to say we expect proc mounted to satisfy the implementation requirement.
    If the impl requirements change then the test can be out of date and
    need updating.

(b) If we don't express the requirements at the test level then the test will
    just fail when proc is not mounted and developers can no longer reason well
    about what is expected to fail and what is not expected to fail. This is not
    a good solution to this problem and it is where we are today.

Your solution is (a), and the only better solutions I can imagine (which doesn't
require root) are as follows:

(1) On a per-impl level provide a specification of the host requirements.
    Have tests gather those requirements in a general way and use them to
    derive UNSUPPORTED behaviour. This is way too much work.

(2) Provide a synthetic /proc that does what we want based on the impl.
    This is also too much work for what we want to achieve.

In the end I think your solution is an acceptable middle ground:

* Markup the tests.
* Attempt to mount /proc in the container, and if we fail then using
  the test markup (API calls) we can mark tests as UNSUPPORTED.

Please review comments below.

Looking forward to a v6.
 
> diff --git a/elf/tst-pldd.c b/elf/tst-pldd.c
> index f31f9956faa..e9e99d11e01 100644
> --- a/elf/tst-pldd.c
> +++ b/elf/tst-pldd.c
> @@ -85,6 +85,9 @@ in_str_list (const char *libname, const char *const strlist[])
>  static int
>  do_test (void)
>  {
> +  /* needs /proc/sys/kernel/yama/ptrace_scope and /proc/$child */
> +  support_need_proc ();

OK.

> +
>    /* Check if our subprocess can be debugged with ptrace.  */
>    {
>      int ptrace_scope = support_ptrace_scope ();
> diff --git a/nptl/tst-pthread-getattr.c b/nptl/tst-pthread-getattr.c
> index d2ebf308ae7..4a8bf1b4846 100644
> --- a/nptl/tst-pthread-getattr.c
> +++ b/nptl/tst-pthread-getattr.c
> @@ -28,6 +28,8 @@
>  #include <unistd.h>
>  #include <inttypes.h>
>  
> +#include <support/support.h>

OK.

> +
>  /* There is an obscure bug in the kernel due to which RLIMIT_STACK is sometimes
>     returned as unlimited when it is not, which may cause this test to fail.
>     There is also the other case where RLIMIT_STACK is intentionally set as
> @@ -153,6 +155,9 @@ check_stack_top (void)
>  static int
>  do_test (void)
>  {
> +  /* Reads /proc/self/maps to get stack size.  */
> +  support_need_proc ();

OK.

> +
>    pagesize = sysconf (_SC_PAGESIZE);
>    return check_stack_top ();
>  }
> diff --git a/nss/tst-reload2.c b/nss/tst-reload2.c
> index fb3b94a1fab..94e2029fd35 100644
> --- a/nss/tst-reload2.c
> +++ b/nss/tst-reload2.c
> @@ -95,6 +95,9 @@ do_test (void)
>    char buf1[PATH_MAX];
>    char buf2[PATH_MAX];
>  
> +  /* The xmkdirp below fails if we can't map our uid, which requires /proc.  */
> +  support_need_proc ();

OK.

> +
>    sprintf (buf1, "/subdir%s", support_slibdir_prefix);
>    xmkdirp (buf1, 0777);
>  
> diff --git a/support/Makefile b/support/Makefile
> index 5ddcb8d1581..f036a813048 100644
> --- a/support/Makefile
> +++ b/support/Makefile
> @@ -64,6 +64,7 @@ libsupport-routines = \
>    support_format_netent \
>    support_isolate_in_subprocess \
>    support_mutex_pi_monotonic \
> +  support_need_proc \

OK.

>    support_path_support_time64 \
>    support_process_state \
>    support_ptrace \
> diff --git a/support/support.h b/support/support.h
> index 73b9fc48f01..bcf7bc43723 100644
> --- a/support/support.h
> +++ b/support/support.h
> @@ -91,6 +91,10 @@ char *support_quote_string (const char *);
>     regular file open for writing, and initially empty.  */
>  int support_descriptor_supports_holes (int fd);
>  
> +/* Predicates that a test requires a working /proc filesystem.  This
> +   call will exit with UNSUPPORTED if /proc is not available.  */
> +void support_need_proc (void);

OK.

> +
>  /* Error-checking wrapper functions which terminate the process on
>     error.  */
>  
> diff --git a/support/support_need_proc.c b/support/support_need_proc.c
> new file mode 100644
> index 00000000000..5d94b25ba8b
> --- /dev/null
> +++ b/support/support_need_proc.c
> @@ -0,0 +1,33 @@
> +/* Indicate that a test requires a working /proc.
> +   Copyright (C) 2022 Free Software Foundation, Inc.
> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#include <unistd.h>
> +#include <support/check.h>
> +#include <support/support.h>
> +
> +/* We test for /proc/self/maps since that's one of the files that one
> +   of our tests actually uses, but the general idea is if Linux's
> +   /proc/ (procfs) filesystem is mounted.  If not, the process exits
> +   with an UNSUPPORTED result code.  */
> +
> +void
> +support_need_proc (void)
> +{
> +  if (access ("/proc/self/maps", R_OK))
> +    FAIL_UNSUPPORTED ("/proc is not available");

This needs wrapping in #ifdef __linux__ for Hurd support.

Please see the various uses of this in support/*.

So on Hurd it should do nothing, but on Linux it should fail unsupported.

> +}
> diff --git a/support/test-container.c b/support/test-container.c
> index 25e7f142193..d574ee6cd89 100644
> --- a/support/test-container.c
> +++ b/support/test-container.c
> @@ -49,6 +49,20 @@
>  #include "check.h"
>  #include "test-driver.h"
>  
> +/* If set to 0, we bind mount the parent's /proc and re-use the
> +   paren't pids (i.e. test runs as the user).  If set to 1, we enter a
> +   new pid namespace and mount our own /proc (i.e. test runs as root).
> +   I was tired of swapping the code back and forth during testing, so
> +   you get both ;-) */
> +#define SEPARATE_PID_NS 1
> +#define SHARED_PID_NS (!SEPARATE_PID_NS)

This kind of conditional leads to bitrot on the side that is not enabled.

Please remove this and use a variable to ensure both paths are always compiled
and therefore we see warnings and error.

> +
> +#if SEPARATE_PID_NS
> +#define MAYBE_CLONE_NEWPID CLONE_NEWPID
> +#else
> +#define MAYBE_CLONE_NEWPID 0
> +#endif

Should be a variable also.

> +
>  #ifndef __linux__
>  #define mount(s,t,fs,f,d) no_mount()
>  int no_mount (void)
> @@ -231,7 +245,7 @@ concat (const char *str, ...)
>  static void
>  trymount (const char *src, const char *dest)
>  {
> -  if (mount (src, dest, "", MS_BIND, NULL) < 0)
> +  if (mount (src, dest, "", MS_BIND | MS_REC, NULL) < 0)
>      FAIL_EXIT1 ("can't mount %s onto %s\n", src, dest);
>  }
>  
> @@ -1068,7 +1082,7 @@ main (int argc, char **argv)
>  
>  #ifdef CLONE_NEWNS
>    /* The unshare here gives us our own spaces and capabilities.  */
> -  if (unshare (CLONE_NEWUSER | CLONE_NEWPID | CLONE_NEWNS) < 0)
> +  if (unshare (CLONE_NEWUSER | MAYBE_CLONE_NEWPID | CLONE_NEWNS) < 0)
>      {
>        /* Older kernels may not support all the options, or security
>  	 policy may block this call.  */
> @@ -1094,6 +1108,15 @@ main (int argc, char **argv)
>    trymount (support_srcdir_root, new_srcdir_path);
>    trymount (support_objdir_root, new_objdir_path);
>  
> +#if SHARED_PID_NS
> +  /* It may not be possible to mount /proc directly.  */
> +  {
> +    char *new_proc = concat (new_root_path, "/proc", NULL);
> +    xmkdirp (new_proc, 0755);
> +    trymount ("/proc", new_proc);
> +  }
> +#endif

Conditional on variable indicating shared pid ns.

> +
>    xmkdirp (concat (new_root_path, "/dev", NULL), 0755);
>    devmount (new_root_path, "null");
>    devmount (new_root_path, "zero");
> @@ -1163,42 +1186,57 @@ main (int argc, char **argv)
>  
>    maybe_xmkdir ("/tmp", 0755);
>  
> +#if SEPARATE_PID_NS

Likewise for separate pid ns.

>    /* Now that we're pid 1 (effectively "root") we can mount /proc  */
>    maybe_xmkdir ("/proc", 0777);
> -  if (mount ("proc", "/proc", "proc", 0, NULL) < 0)
> -    FAIL_EXIT1 ("Unable to mount /proc: ");
> -
> -  /* We map our original UID to the same UID in the container so we
> -     can own our own files normally.  */
> -  UMAP = open ("/proc/self/uid_map", O_WRONLY);
> -  if (UMAP < 0)
> -    FAIL_EXIT1 ("can't write to /proc/self/uid_map\n");
> -
> -  sprintf (tmp, "%lld %lld 1\n",
> -	   (long long) (be_su ? 0 : original_uid), (long long) original_uid);
> -  write (UMAP, tmp, strlen (tmp));
> -  xclose (UMAP);
> -
> -  /* We must disable setgroups () before we can map our groups, else we
> -     get EPERM.  */
> -  GMAP = open ("/proc/self/setgroups", O_WRONLY);
> -  if (GMAP >= 0)
> +  if (mount ("proc", "/proc", "proc", 0, NULL) != 0)
>      {
> -      /* We support kernels old enough to not have this.  */
> -      write (GMAP, "deny\n", 5);
> -      xclose (GMAP);
> +      // This happens if we're trying to create a nested container,
> +      // like if the build is running under podman, and we lack
> +      // priviledges.
> +
> +      // Ideally we would WARN here, but that would just add noise to
> +      // *every* test-container test, and the ones that care should
> +      // have their own relevent diagnostics.
> +
> +      // FAIL_EXIT1 ("Unable to mount /proc: ");

Please use C-style comments.

>      }
> +  else
> +    /* The shared pid namespace case will always run the following block... */
> +#endif
> +    {
> +      /* We map our original UID to the same UID in the container so we
> +	 can own our own files normally.  */
> +      UMAP = open ("/proc/self/uid_map", O_WRONLY);
> +      if (UMAP < 0)
> +	FAIL_EXIT1 ("can't write to /proc/self/uid_map\n");
> +
> +      sprintf (tmp, "%lld %lld 1\n",
> +	       (long long) (be_su ? 0 : original_uid), (long long) original_uid);
> +      write (UMAP, tmp, strlen (tmp));
> +      xclose (UMAP);
> +
> +      /* We must disable setgroups () before we can map our groups, else we
> +	 get EPERM.  */
> +      GMAP = open ("/proc/self/setgroups", O_WRONLY);
> +      if (GMAP >= 0)
> +	{
> +	  /* We support kernels old enough to not have this.  */
> +	  write (GMAP, "deny\n", 5);
> +	  xclose (GMAP);
> +	}
>  
> -  /* We map our original GID to the same GID in the container so we
> -     can own our own files normally.  */
> -  GMAP = open ("/proc/self/gid_map", O_WRONLY);
> -  if (GMAP < 0)
> -    FAIL_EXIT1 ("can't write to /proc/self/gid_map\n");
> +      /* We map our original GID to the same GID in the container so we
> +	 can own our own files normally.  */
> +      GMAP = open ("/proc/self/gid_map", O_WRONLY);
> +      if (GMAP < 0)
> +	FAIL_EXIT1 ("can't write to /proc/self/gid_map\n");
>  
> -  sprintf (tmp, "%lld %lld 1\n",
> -	   (long long) (be_su ? 0 : original_gid), (long long) original_gid);
> -  write (GMAP, tmp, strlen (tmp));
> -  xclose (GMAP);
> +      sprintf (tmp, "%lld %lld 1\n",
> +	       (long long) (be_su ? 0 : original_gid), (long long) original_gid);
> +      write (GMAP, tmp, strlen (tmp));
> +      xclose (GMAP);
> +    }
>  
>    if (change_cwd)
>      {
>