[PATCHv6] gdb: handle loading shared libraries from /proc/self/fd/

  From: Asaf Fisher via Gdb-patches <gdb-patches@sourceware.org>

Changes since v5:

  - Rebased onto HEAD of master, updated the test script to use
    'require allow_shlib_tests'.

Changes since v4:

  - filter_proc_self_filenames now makes use of gdb_realpath_keepfile to
    (partially) canonicalize the path, this should fix the non-canonical
    path issues,

  - extended the test to include the non-canonical path case.

Changes since v3:

  - Fixed NULL/nullptr issues Lancelot pointed out,

  - Updated the test case to take account of non-probes based shared
    library event handling.

---

Bug PR gdb/29586 describes a situations where a shared library is
created in memory, then written to a memory mapped file.  The memory
mapped file will show up as a file descriptor within /proc/self/fd/,
and this path is then used with dlopen in order to call functions
within the in-memory shared library.

When attempting to debug this GDB hangs.  The problem is that, GDB
stops at the shared-library event, reads the shared library path from
the inferior, which is "/proc/self/fd/<NUM>", and then GDB attempts to
open this file.

Unfortunately, this means GDB tries to open a file within GDB's
/proc/self/fd/ directory, not within the inferior's directory.  In the
case of our hang it turns out that the file descriptor that is opened
is a pipe, and GDB hangs trying to read from the pipe.

However, the behaviour is really just undefined, depending on which
file descriptor the inferior tries to open, GDB will open, or fail to
open, random files within its /proc/self/fd directory.

The solution proposed in this commit is to hook into solib_find_1, and
spot when GDB is looking for any file in /proc/self/, if this is the
case, then the filename is rewritten as /proc/<PID>, where <PID> is
the process-id of the current inferior.

One complexity that we need to consider is how to handle non-canonical
paths, for example, if the user opens /proc/../proc/self/fd/<NUM> then
GDB also needs to be able to spot this.

My solution is to call gdb_realpath_keepfile, this canonicalizes
everything except the filename (the '/proc/../proc/self/fd' part in
the above example).  The problem this then introduces is that the
'self' is replaces with the pid of GDB itself.  I therefore check,
after canonicalization, for a prefix of /proc/<PID OF GDB>, and if
this is the case, I replace this with /proc/<PID OF INFERIOR>.

There is one remaining issue, which I consider very unlikely, and I
don't propose to fix.  What if the inferior actually did try to dlopen
a file from /proc/<PID OF GDB>?  If this happened then GDB will not be
able to distinguish this from the above case and would still replace
the pid of GDB with the pid of the inferior.  However, this seems
really unlikely - why would the inferior be trying to load parts of
GDB?  So I propose to ignore this case until someone can come up with
an actual scenario where that's useful.

The test passes for the unix, native-gdbserver, and
native-extended-gdbserver targets.  When testing with either of the
gdbserver targets, the test is run using the default empty sysroot,
and also using the 'target:' sysroot.

Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29586

Co-authored-by: Andrew Burgess <aburgess@redhat.com>
---
 gdb/solib.c                                |  58 +++++-
 gdb/testsuite/gdb.base/solib-proc-self.cc  |  89 ++++++++++
 gdb/testsuite/gdb.base/solib-proc-self.exp | 197 +++++++++++++++++++++
 3 files changed, 340 insertions(+), 4 deletions(-)
 create mode 100644 gdb/testsuite/gdb.base/solib-proc-self.cc
 create mode 100644 gdb/testsuite/gdb.base/solib-proc-self.exp


base-commit: 7d02a94c8f74613edb0cdde11982b22eaaa9affe
  

On 2023-01-20 12:33 p.m., Andrew Burgess via Gdb-patches wrote:
> From: Asaf Fisher via Gdb-patches <gdb-patches@sourceware.org>
> 
> Changes since v5:
> 
>   - Rebased onto HEAD of master, updated the test script to use
>     'require allow_shlib_tests'.
> 
> Changes since v4:
> 
>   - filter_proc_self_filenames now makes use of gdb_realpath_keepfile to
>     (partially) canonicalize the path, this should fix the non-canonical
>     path issues,
> 
>   - extended the test to include the non-canonical path case.
> 
> Changes since v3:
> 
>   - Fixed NULL/nullptr issues Lancelot pointed out,
> 
>   - Updated the test case to take account of non-probes based shared
>     library event handling.
> 
> ---
> 
> Bug PR gdb/29586 describes a situations where a shared library is

"a situations" -> "a situation"

> created in memory, then written to a memory mapped file.  The memory
> mapped file will show up as a file descriptor within /proc/self/fd/,
> and this path is then used with dlopen in order to call functions
> within the in-memory shared library.
> 
> When attempting to debug this GDB hangs.  The problem is that, GDB
> stops at the shared-library event, reads the shared library path from
> the inferior, which is "/proc/self/fd/<NUM>", and then GDB attempts to
> open this file.
> 
> Unfortunately, this means GDB tries to open a file within GDB's
> /proc/self/fd/ directory, not within the inferior's directory.  In the
> case of our hang it turns out that the file descriptor that is opened
> is a pipe, and GDB hangs trying to read from the pipe.
> 
> However, the behaviour is really just undefined, depending on which
> file descriptor the inferior tries to open, GDB will open, or fail to
> open, random files within its /proc/self/fd directory.
> 
> The solution proposed in this commit is to hook into solib_find_1, and
> spot when GDB is looking for any file in /proc/self/, if this is the
> case, then the filename is rewritten as /proc/<PID>, where <PID> is
> the process-id of the current inferior.
> 
> One complexity that we need to consider is how to handle non-canonical
> paths, for example, if the user opens /proc/../proc/self/fd/<NUM> then
> GDB also needs to be able to spot this.
> 
> My solution is to call gdb_realpath_keepfile, this canonicalizes
> everything except the filename (the '/proc/../proc/self/fd' part in
> the above example).  The problem this then introduces is that the
> 'self' is replaces with the pid of GDB itself.  I therefore check,

"is replaces" -> "is replaced"


> after canonicalization, for a prefix of /proc/<PID OF GDB>, and if
> this is the case, I replace this with /proc/<PID OF INFERIOR>.
> 
> There is one remaining issue, which I consider very unlikely, and I
> don't propose to fix.  What if the inferior actually did try to dlopen
> a file from /proc/<PID OF GDB>?  If this happened then GDB will not be
> able to distinguish this from the above case and would still replace
> the pid of GDB with the pid of the inferior.  However, this seems
> really unlikely - why would the inferior be trying to load parts of
> GDB?  So I propose to ignore this case until someone can come up with
> an actual scenario where that's useful.
> 
> The test passes for the unix, native-gdbserver, and
> native-extended-gdbserver targets.  When testing with either of the
> gdbserver targets, the test is run using the default empty sysroot,
> and also using the 'target:' sysroot.
> 
> Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29586
> 
> Co-authored-by: Andrew Burgess <aburgess@redhat.com>
> ---
>  gdb/solib.c                                |  58 +++++-
>  gdb/testsuite/gdb.base/solib-proc-self.cc  |  89 ++++++++++
>  gdb/testsuite/gdb.base/solib-proc-self.exp | 197 +++++++++++++++++++++
>  3 files changed, 340 insertions(+), 4 deletions(-)
>  create mode 100644 gdb/testsuite/gdb.base/solib-proc-self.cc
>  create mode 100644 gdb/testsuite/gdb.base/solib-proc-self.exp
> 
> diff --git a/gdb/solib.c b/gdb/solib.c
> index 60bdf46cb91..3181b4d5f1f 100644
> --- a/gdb/solib.c
> +++ b/gdb/solib.c
> @@ -50,6 +50,7 @@
>  #include "gdb_bfd.h"
>  #include "gdbsupport/filestuff.h"
>  #include "gdbsupport/scoped_fd.h"
> +#include "gdbsupport/pathstuff.h"
>  #include "debuginfod-support.h"
>  #include "source.h"
>  #include "cli/cli-style.h"
> @@ -79,6 +80,54 @@ show_solib_search_path (struct ui_file *file, int from_tty,
>  #  define DOS_BASED_FILE_SYSTEM 0
>  #endif
>  
> +/* Fix references to files in /proc/self/fd/ when opening a shared library.
> +
> +   SO_NAME is the name of the shared library being loaded.  This function
> +   returns a possibly modified name which should be used as the path to the
> +   shared library.
> +
> +   If SO_NAME starts with /proc/self, then the returned name will be
> +   modified to start with /proc/PID where 'PID' is the pid of the current
> +   inferior.  */
> +
> +static gdb::unique_xmalloc_ptr<char>
> +filter_proc_self_filenames (gdb::unique_xmalloc_ptr<char> so_name)
> +{
> +  /* In order to figure out if SO_NAME points at a file in /proc/self we
> +     need to canonicalize the path.  However, the whole point here is that
> +     if SO_NAME does point at /proc/self then GDB will think this is its
> +     self, while the inferior actually means its /proc/self.  A result of
> +     this is that the file being referenced within the inferior's
> +     /proc/self, might not exist within GDB's /proc/self, a call to
> +     gdb_realpath would then fail.
> +
> +     To avoid this problem we call gdb_realpath_keepfile, this only tries
> +     to canonicalize the basename of SO_NAME, and leaves the final filename
> +     untouched.  */
> +  so_name
> +    = make_unique_xstrdup (gdb_realpath_keepfile (so_name.get ()).c_str ());

gdb_realpath_keepfile resolves symlinks on the local filesystem, so how can
we be using it on target: files?  There's a comment exactly about this in exec.c:

      /* gdb_realpath_keepfile resolves symlinks on the local
	 filesystem and so cannot be used for "target:" files.  */
      gdb_assert (current_program_space->exec_filename == nullptr);
      if (load_via_target)
	current_program_space->exec_filename
	  = (make_unique_xstrdup
	     (bfd_get_filename (current_program_space->exec_bfd ())));
      else
	current_program_space->exec_filename
	  = make_unique_xstrdup (gdb_realpath_keepfile
				   (scratch_pathname).c_str ());


Also, what if the host is not a Linux host, if it is one where /proc/ does
not exist at all, such as e.g., a Windows host debugging a remote Linux target?
Won't gdb_realpath_keepfile fail in that scenario?

Also, do we get here with a relative pathname?  In that case, this will use
the CWD of GDB for path expansion, which just looks wrong.

I wonder whether using the linux namespace support wouldn't be a better approach.

Now how linux_nat_target::fileio_readlink calls linux_mntns_readlink:

 gdb::optional<std::string>
 linux_nat_target::fileio_readlink (struct inferior *inf, const char *filename,
 				   fileio_error *target_errno)
 {
 ...
   len = linux_mntns_readlink (linux_nat_fileio_pid_of (inf),
 			      filename, buf, sizeof (buf));
 ...

That means that the link is resolved in the namespace of the inferior.

GDBserver has similar code.

Also note:

 $ readlink -f /proc/self
 /proc/1957246

That suggests that we would just need to call target_fileio_readlink somewhere
to resolve the path.

That would leave core files to address, as you can't enter the namespace of
a process that is already gone.  But for cores, your solution is also not safe,
as opening "/proc/inferior_pid/" from the live host system is incorrect, that
PID might as well have been assigned to some other process, if it even exists.
Reading from /proc/pid/ files must be handled specially for core files.
So that leaves me to think that the target_fileio_readlink approach could work.

> +
> +  /* A result of the above canonicalization is that /proc/self will have
> +     been replaced with /proc/<pid of gdb>, so that's what we need to check
> +     for.  Of course, it could be the case that the inferior really did try
> +     to dlopen a file within GDB's /proc/<pid> directory, in which case
> +     we're going to do the wrong thing here, but that seems far less likely
> +     than calling dlopen on a file within the inferior's own directory.  */
> +  std::string prefix = string_printf ("/proc/%ld", (long) getpid ());
> +
> +  /* Is the canonical path inside GDB's /proc/<pid> directory?  */
> +  if (!startswith (so_name.get (), prefix.c_str ()))
> +    return so_name;

If GDB's pid is "1234", and so_name is "/proc/12345/something", this will think
the path is GDB's /proc/<pid> directory.  Need to append a slash to
PREFIX to avoid this.  Also need to check for exact "/proc/$GDB_PID" then.

> +
> +  /* Get the part of the path after /proc/<pid>.  For example given
> +     '/proc/123/fd' we find the '/fd' part.  */
> +  const char *tail = so_name.get () + strlen (prefix.c_str ());
> +
> +  /* Build a replacement path within the inferiors directory.  */
> +  int inferior_pid = inferior_ptid.pid ();
> +  return xstrprintf ("/proc/%d%s", inferior_pid, tail);
> +}
> +
>  /* Return the full pathname of a binary file (the main executable or a
>     shared library file), or NULL if not found.  If FD is non-NULL, *FD
>     is set to either -1 or an open file handle for the binary file.
> @@ -168,6 +217,9 @@ solib_find_1 (const char *in_pathname, int *fd, bool is_solib)
>  	}
>      }
>  
> +  temp_pathname.reset (xstrdup (in_pathname));
> +  temp_pathname = filter_proc_self_filenames (std::move (temp_pathname));

FWIW, this could be written as:

  temp_pathname = filter_proc_self_filenames (make_unique_xtrdup (in_pathname));

> +
>    /* Note, we're interested in IS_TARGET_ABSOLUTE_PATH, not
>       IS_ABSOLUTE_PATH.  The latter is for host paths only, while
>       IN_PATHNAME is a target path.  For example, if we're supposed to
> @@ -180,9 +232,7 @@ solib_find_1 (const char *in_pathname, int *fd, bool is_solib)
>         3rd attempt, c:/foo/bar.dll ==> /sysroot/foo/bar.dll
>    */
>  
> -  if (!IS_TARGET_ABSOLUTE_PATH (fskind, in_pathname) || sysroot == NULL)
> -    temp_pathname.reset (xstrdup (in_pathname));
> -  else
> +  if (IS_TARGET_ABSOLUTE_PATH (fskind, in_pathname) && sysroot != nullptr)
>      {
>        bool need_dir_separator;
>  
> @@ -209,7 +259,7 @@ solib_find_1 (const char *in_pathname, int *fd, bool is_solib)
>        /* Cat the prefixed pathname together.  */
>        temp_pathname.reset (concat (sysroot,
>  				   need_dir_separator ? SLASH_STRING : "",
> -				   in_pathname, (char *) NULL));
> +				   temp_pathname.get (), nullptr));
>      }
>  
>    /* Handle files to be accessed via the target.  */
> diff --git a/gdb/testsuite/gdb.base/solib-proc-self.cc b/gdb/testsuite/gdb.base/solib-proc-self.cc
> new file mode 100644
> index 00000000000..aea397a66f9
> --- /dev/null
> +++ b/gdb/testsuite/gdb.base/solib-proc-self.cc
> @@ -0,0 +1,89 @@
> +/* This testcase is part of GDB, the GNU debugger.
> +
> +   Copyright 2022 Free Software Foundation, Inc.

Dates need update to 2022-2023.

> +
> +   This program is free software; you can redistribute it and/or modify
> +   it under the terms of the GNU General Public License as published by
> +   the Free Software Foundation; either version 3 of the License, or
> +   (at your option) any later version.
> +
> +   This program 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 General Public License for more details.
> +
> +   You should have received a copy of the GNU General Public License
> +   along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
> +
> +#include <sys/mman.h>
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <iostream>
> +#include <fstream>
> +#include <sstream>
> +#include <vector>
> +#include <unistd.h>
> +
> +#ifdef __WIN32__
> +#include <windows.h>
> +#define dlopen(name, mode) LoadLibrary (name)
> +#define dlclose(handle) FreeLibrary (handle)
> +#define dlerror() "an error occurred"

/proc/self/ doesn't exit on Windows, so this seems pointless.

Maybe also add some require !istarget windows to the .exp file.


> +#else
> +#include <dlfcn.h>
> +#endif
> +
> +/* Combine PREFIX and NUM into a single string, joined with a '/', and
> +   return the string.  */
> +
> +std::string
> +make_library_path (const char *prefix, int num)
> +{
> +  std::stringstream stream;
> +  stream << prefix << "/" << num;
> +  return stream.str ();
> +}
> +
> +/* Call dlopen on the library pointed to by FILENAME.  */
> +
> +void
> +open_library (const std::string &filename)
> +{
> +  /* Call dlopen on FILENAME.  */
> +  void *handle = dlopen (filename.c_str (), RTLD_LAZY);
> +  if (handle == nullptr)
> +    abort ();
> +
> +  /* It worked.  */
> +  dlclose (handle);
> +}
> +
> +int
> +main ()
> +{
> +  /* Read the shared libraries contents into a buffer.  */
> +  std::ifstream read_so_file = std::ifstream (SHLIB_NAME);
> +  read_so_file.seekg (0, std::ios::end);
> +  std::streamsize size = read_so_file.tellg ();
> +  read_so_file.seekg (0, std::ios::beg);
> +  std::vector<char> buffer (size);
> +  if (!read_so_file.read (buffer.data (), size))
> +    abort ();
> +
> +  /* Create a memory mapped file, then write the shared library to that
> +     new memory mapped file.  */
> +  int mem_fd = memfd_create ("test", 0);
> +  write (mem_fd, buffer.data (), buffer.size ());

Missing error checking here.

> +
> +  /* Generate a canonical /proc/self/fd/[num] path for the memory mapped
> +     file, and call dlopen on it.  */
> +  std::string filename
> +    = make_library_path ("/proc/self/fd", mem_fd);	/* break-here */
> +  open_library (filename);
> +
> +  /* Now generate a new, non-canonical filename, and call dlopen on it.  */
> +  filename = make_library_path ("/proc/../proc/self/fd", mem_fd);
> +  open_library (filename);
> +
> +  return 0;
> +}
> diff --git a/gdb/testsuite/gdb.base/solib-proc-self.exp b/gdb/testsuite/gdb.base/solib-proc-self.exp
> new file mode 100644
> index 00000000000..f026d2544e6
> --- /dev/null
> +++ b/gdb/testsuite/gdb.base/solib-proc-self.exp
> @@ -0,0 +1,197 @@
> +# Copyright 2022 Free Software Foundation, Inc.
> +
> +# This program is free software; you can redistribute it and/or modify
> +# it under the terms of the GNU General Public License as published by
> +# the Free Software Foundation; either version 3 of the License, or
> +# (at your option) any later version.
> +#
> +# This program 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 General Public License for more details.
> +#
> +# You should have received a copy of the GNU General Public License
> +# along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
> +
> +# Test connecting and disconnecting at shared library events.

This comment is stale.

> +
> +require allow_shlib_tests
> +
> +standard_testfile .cc
> +
> +# Reuse an existing library, we don't care about the library contents
> +# for this test.
> +set libfile so-disc-shr
> +set libsrc "${srcdir}/${subdir}/${libfile}.c"
> +set libname "${libfile}.so"
> +set libobj [standard_output_file ${libname}]
> +
> +# Compile the shared library.
> +if { [gdb_compile_shlib $libsrc $libobj {debug}] != ""} {
> +    return -1
> +}
> +
> +# Compile the test executable.
> +if [ build_executable "failed to prepare" $testfile $srcfile \
> +	 [list shlib_load debug c++ additional_flags=-DSHLIB_NAME="${libobj}"]] {
> +    return -1
> +}
> +
> +proc validate_library_load {inferior_pid memfd} {
> +
> +    # Turn on the solib-events so we can see that gdb resolves
> +    # everything correctly.
> +    gdb_test_no_output "set stop-on-solib-events 1"
> +
> +    # Now continue forward until the solib event is detected, and
> +    # check that the loaded library is found through the /proc/PID/fd
> +    # rather than /proc/self/fd.
> +    #
> +    # We need to handle the possibility of the interesting event
> +    # showing the first, or second time we stop, as depending on which
> +    # mechanism GDB is using to handle the shared library events (the
> +    # newer probes based interface, or the old non-probes interface),
> +    # the library will be reported at the first or second stop.
> +    set saw_expected_event false
> +    set saw_no_event_stop false
> +    gdb_test_multiple "continue" "continue to solib evnt" {

"solib evnt" -> "solib event"

> +	-re "^continue\r\n" {
> +	    exp_continue
> +	}
> +
> +	-re "^Continuing\\.\r\n" {
> +	    exp_continue
> +	}
> +
> +	-early -re "Stopped due to shared library event \\(no libraries added or removed\\)\r\n" {
> +	    # This non-interesting event shows up first when using the
> +	    # probes based mechanism for dealing with shared library
> +	    # events.
> +	    #
> +	    # We set a flag here, and, once the prompt has appeared,
> +	    # we send another continue, the next event will contain
> +	    # the information we want.
> +	    set saw_no_event_stop true
> +	    exp_continue
> +	}
> +
> +	-re "Stopped due to shared library event:\r\n  Inferior loaded (?:target:)?/proc/${inferior_pid}/fd/$memfd\r\n" {
> +	    # This event, which includes the information we are
> +	    # looking for, occurs first when using the non-probes
> +	    # based mechanism for handling shared library events, and
> +	    # occurs second when using the probes mechanism.
> +	    #
> +	    # Either way, record here that we say the output we expect.
> +	    set saw_expected_event true
> +	    exp_continue
> +	}
> +
> +	-re "$::gdb_prompt $" {
> +	    if {$saw_no_event_stop} {
> +		set saw_no_event_stop false
> +		send_gdb "continue\n"
> +		exp_continue
> +	    } else {
> +		gdb_assert {$saw_expected_event} $gdb_test_name
> +	    }
> +	}
> +    }
> +
> +    # Turn off solib events.  We're only interested in validating the
> +    # loads for now.
> +    gdb_test_no_output "set stop-on-solib-events 0"
> +}
> +
> +# Start GDB and run to the point where the test program tries to dlopen a file
> +# from within /proc/self/fd/.  Catch the shared library event and check that
> +# we actually try to load a file from /proc/<INFERIOR-PID>/fd/.
> +#
> +# If SYSROOT is not the empty string, then this is set as the value of GDB's
> +# sysroot immediately after starting GDB.  The only value that is (currently)
> +# supported, other than the empty string, is 'target:'.
> +proc do_test { {sysroot ""} } {
> +    clean_restart $::binfile
> +
> +    if {$sysroot != ""} {
> +	gdb_test_no_output "set sysroot ${sysroot}"
> +    }
> +
> +    gdb_load_shlib $::libobj
> +
> +    if ![runto_main] then {
> +	return 0
> +    }
> +
> +    # Get inferior's PID for later.
> +    set inferior_pid [get_inferior_pid]
> +
> +    # Run to the 'break-here' marker.
> +    gdb_breakpoint [gdb_get_line_number "break-here"]
> +    gdb_continue_to_breakpoint "break-here" ".* break-here .*"
> +
> +    set memfd ""
> +    gdb_test_multiple "p mem_fd" "Get file descriptor" {

Lowercase "Get", please.

Pedro Alves

> +	-re -wrap "\\\$$::decimal = (\[^\r\n\]*)" {
> +	    set memfd $expect_out(1,string)
> +	    pass $gdb_test_name
> +	}
> +    }
> +
> +    # The first call to open_library (in the test program) is done
> +    # with a canonical path.
> +    with_test_prefix "canonical path" {
> +	gdb_breakpoint "open_library"
> +	gdb_continue_to_breakpoint "open_library"
> +	validate_library_load $inferior_pid $memfd
> +    }
> +
> +    # The second call to open_library (in the test program) is done
> +    # with a non-canonical path, however, GDB should resolve this to a
> +    # canonical path for display to the user, so the output we see
> +    # should be unchanged.
> +    with_test_prefix "non-canonical path" {
> +	gdb_breakpoint "open_library"
> +	gdb_continue_to_breakpoint "open_library"
> +	validate_library_load $inferior_pid $memfd
> +    }
> +}
> +
> +# First run of the test.
> +do_test
> +
> +# Possible second run of the test.  If we are using a remote target then we
> +# should consider setting the sysroot to 'target:' and re-running the test.
> +if {[target_info exists gdb_protocol]
> +    && ([target_info gdb_protocol] == "remote"
> +	|| [target_info gdb_protocol] == "extended-remote")} {
> +    # GDB will already be running after the first call to do_test, so we can
> +    # take a peek at the current sysroot setting, and decide if we should
> +    # repeat the test with a different setting.
> +
> +    set new_sysroot ""
> +    gdb_test_multiple "show sysroot" "" {
> +	-wrap -re "The current system root is \"\"\\." {
> +	    pass $gdb_test_name
> +
> +	    # Repeat the test with 'target:' sysroot.
> +	    set new_sysroot "target:"
> +	}
> +	-wrap -re "The current system root is \"target:\"\\." {
> +	    pass $gdb_test_name
> +
> +	    # Nothing else to do, we already tested with target: sysroot.
> +	}
> +	-re "$gdb_prompt $" {
> +	    pass $gdb_test_name
> +
> +	    # If already testing with any other sysroot, we probably should
> +	    # not try to adjust things, so don't do any further testing.
> +	}
> +    }
> +
> +    with_test_prefix "sysroot $new_sysroot" {
> +	if { $new_sysroot != "" } {
> +	    do_test $new_sysroot
> +	}
> +    }
> +}
> 
> base-commit: 7d02a94c8f74613edb0cdde11982b22eaaa9affe
>