Add a timeout to performing an ssh command. This will let testing if a
machine is alive or not, or if something else may be amiss. A timeout can be
passed to ssh, where ssh will fail if it does not complete within the given
timeout.
Signed-off-by: NSteven Rostedt (VMware) <rostedt@goodmis.org>

The child_exit errno needs to be shifted by 8 bits to compare against the
return values for the bisect variables.

Fixes: c5dacb88 ("ktest: Allow overriding bisect test results")
Cc: stable@vger.kernel.org
Signed-off-by: NSteven Rostedt (VMware) <rostedt@goodmis.org>

The POST_TEST config is to be executed after a test has fully compeleted,
whether the test passed or failed. It currently is executed at the moment
that the test has been decided if it failed or not. As the test does other
clean ups, it isn't truly finished. Move the POST_TEST execution to after
all the test cleanups have been done.
Signed-off-by: NSteven Rostedt (VMware) <rostedt@goodmis.org>

Seems that some of the new console logic causes doprint to possibly
get evaluated. When printing a commit message that contains parenthesis,
it fails with a shell parsing error.

This gets fixed when we add quotes around the $item variable, and prevent
it from being evaluated by any shell commands.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

If dodie() is called with the console open, restore the terminal's
original settings before dying.

Link: http://lkml.kernel.org/r/20150130025453.GB20952@treble.redhat.comSigned-off-by: NJosh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Since both success and failure may shortcut and exit ktest, it is better
to print the status times there too. Once times are printed, the values
for the times are reset, so they will not print more than once.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Allow the user to send input to the console by putting the terminal in
cbreak mode (to allow reading stdin one character at a time) and copying
all stdin data to the console's pty.

Link: http://lkml.kernel.org/r/bb1bbe7d202c95a3ce7894cfffdd8c725875978e.1422473610.git.jpoimboe@redhat.comSigned-off-by: NJosh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Create a pseudoterminal (pty pair) to give the console a dedicated tty
so it doesn't mess with ktest's terminal settings.

Link: http://lkml.kernel.org/r/37b0127f9efad09ff4fc994334db998141e4f6ca.1422473610.git.jpoimboe@redhat.comSigned-off-by: NJosh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The function start_monitor_and_boot is a misnomer. It use to, but
now it starts the monitor and installs. It does not boot. Rename it
before I get confused by it again.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Seeing the times for how long a build, install, reboot and the
test takes is helpful for analyzing the test process. Seeing
how different changes affect the timings.

Show the build, install, boot and test times when at the end of
the test, or between each interval for tests that do those
mulitple times (like bisect and patchcheck).
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

When ktest runs the console program as a child process, the parent and
child share the same tty for stdin and stderr.  This is problematic when
using a libvirt target.  The "virsh console" program makes a lot of
changes to the tty settings, making ktest's output hard to read
(carriage returns don't work).  After ktest exits, the terminal is
unusable (CRs broken, stdin isn't echoed).

I think the best way to fix this issue would be to create a
pseudoterminal (pty pair) so the child process would have a dedicated
tty, and then use pipes to connect the two ttys.  I'm not sure if that's
overkill, but it's far beyond my current Perl abilities.

This patch is a much easier way to (partially) fix this issue.  It saves
the tty settings before opening the console and restores them after
closing it.  There are still a few places where ktest prints mangled
output while the console is open, but the output is much more legible
overall, and the terminal works just fine after ktest exits.

Link: http://lkml.kernel.org/r/1bb89abc0025cf1d6da657c7ba58bbeb4381a515.1422382008.git.jpoimboe@redhat.comSigned-off-by: NJosh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

I find that I usually like to see how long a make or other command takes,
and adding a start and end time and reporting how long each command runs
(in seconds) is helpful.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Commit 52d21580 "ktest: Use make -s kernelrelease" fixed commit
7ff52571 "kbuild: fake the "Entering directory ..." message more simply"
as that commit added output after the make kernelrelease. But there's still
some build scripts that are used by ktest that has output before the make
is executed, and requires that only the last line is printed.

Cc: Michal Marek <mmarek@suse.cz>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Instead of just showing the test type of test in the start of the
test, like this:

  RUNNING TEST 1 of 26 with option build defconfig

Add the name (if it is defined) as well, like this:

  RUNNING TEST 1 of 26 (arm64 aarch64-linux) with option build defconfig
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Tests can set options that override the default ones. But if a test
tries to undefine a default option, it is simply ignored and the
default option stays as is.

For example, if you want to have a test that defines no MIN_CONFIG
then the test should be able to do that with:

   TEST_START
   MIN_CONFIG =

Which should make MIN_CONFIG not defined for that test. But the way
the code currently works, undefined options in tests are dropped.
This is because the NULL options are evaluated during the reading of
the config file and since one can disable default options in the default
section with this method, it is evaluated there (the option turns to a
undef). But undef options in the test section mean to use the default
option.

To fix this, keep the empty string in the option during the reading
of the config file, and then evaluate it when running the test. This
will allow tests to null out default options.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Commit 6071c22e "ktest: Rewrite the config-bisect to actually work"
fixed the config-bisect to work nicely but in doing so it broke
make_min_config by changing the way assign_configs works.

The assign_configs function now adds the config to the hash even if
it is disabled, but changes the hash value to be that of the
line "# CONFIG_FOO is not set". Unfortunately, the make_min_config
test only checks to see if the config is removed. It now needs to
check if the config is in the hash and not set to be disabled.

Cc: stable@vger.kernel.org # 3.17+
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The previous tail -1 broke with commit 7ff52571 ("kbuild: fake the
"Entering directory ..." message more simply")

Link: http://lkml.kernel.org/r/20141022194408.GA20989@pobox.suse.czReported-by: NSteven Rostedt <rostedt@goodmis.org>
Signed-off-by: NMichal Marek <mmarek@suse.cz>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

If git bisect reply is being used in the bisect tests, don't bother
doing the git bisect good or git bisect bad calls. The git bisect
reply will override them anyway, and that's called immediately
after the other two. Going the git bisect (good|bad) is just a
waste of time.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The reboot function when rebooting back to a good kernel has a check
to make sure that a new kernel was indeed booted. But that check
uses a timeout value, which when calling the monitor will still
return success if the timeout is hit (no bug was found). It should
return an error to let the reboot code know that a new kernel was
not reached. Only the reboot code checks the return value of the
monitor.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

When doing a manual bisect, a build can fail or a test can be inconclusive.
In these cases it would be helpful to be able to skip the test entirely.

Link: http://lkml.kernel.org/r/1409164021-2136-1-git-send-email-chris.j.arges@canonical.comReviewed-by: NSatoru Takeuchi <satoru.takeuchi@gmail.com>
Signed-off-by: NChris J Arges <chris.j.arges@canonical.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Add a way to run a patchcheck test on the commits that are in one branch
but not in another. This uses git cherry to find a list of commits to
test each one with.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

After the rewrite of the config bisect, the bisect manual was
removed. Add it back.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

After the rewrite of the config bisect, there were several unused
functions that can be removed.

One of the unused functions printed out the failed config nicer than
what the rewrite did, so I kept that and used it to output the
bad config.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The new rewrite left out the CONFIG_BISECT_CHECK, which allows the
user to test that their "bad" config still is bad and their "good"
config still is good. This is especially important as the configs
are passed through a "make oldconfig" to update them with the lastest
kernel. Things could change that causes a bad config to work, or a
good config to break. The check is done after the configs have run
through the oldconfig processing.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

I never liked the way config-bisect worked. I would assume the bad config
had some config that broke the system. But it would not work if the bad
config just happened to be missing something that the good config had.

I rewrote the config-bisect to do this properly. It does a diff of the two
configs, and sets half of the configs that are in one and not the other.
The way it works is that when it "sets", it really just makes one copy
what the other has. That is, a "set" can be setting a:

  # CONFIG_FOO is not set

Basically, it looks at the differences between the two files and makes
them similar until it comes down to one config that makes it work or
not work depending on if it is set or not.

Note, if more than one config change makes the bad config not work, it
will only find one of them. But this is true with all bisect logic.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Some cleanup for improving readability as follows.

  - Initialize $ktest_config at its definition.
  - Put parentheses around the `config-file' argument in the usage message
    because it's a optional one.
  - Rename get_ktest_config{,s} to more descriptive get_mandatory_config{,s}.

Link: http://lkml.kernel.org/r/87fvmr30kb.wl%satoru.takeuchi@gmail.comSigned-off-by: NSatoru Takeuchi <satoru.takeuchi@gmail.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

If we'd like to set the redirect target file of run_command(),
we should define $redirect before calling this function and should undef it
after calling this function. Since it's user-unfriendly, add 2nd parameter of
run_command() for this purpose.

Link: http://lkml.kernel.org/r/87vbvwokq8.wl%satoru.takeuchi@gmail.comSigned-off-by: NSatoru Takeuchi <satoru.takeuchi@gmail.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

For those cases that it takes several tries to hit a bug, it would be
useful for ktest.pl to try a test multiple times before it considers
the test as a pass. To accomplish this, BISECT_TRIES ktest config
option has been added. It is default to one, as most of the time a
bisect only needs to try a test once. But the user can now up this
to make ktest run a given test multiple times. The first failure
that is detected will set a bisect bad. It only repeats on success.

Note, as with all race bugs, there's no guarantee that if it succeeds,
it is really a good bisect. But it helps in case the bug is somewhat
reliable.

You can set BISECT_TRIES to zero, and all tests will be considered
good, unless you also set BISECT_MANUAL.
Suggested-by: N"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

With the added variable ${KERNEL_VERSION}, it is useful to be
able to use parts of it for other variables.

For example, if you want to create a warnings file for each major
kernel version to test sub versions against you can create
your warnings file with like this:

  WARNINGS_FILE = warnings-file-${KERNEL_VERSION}

But this may add 3.8.12 or something, and we want all 3.8.* to
use the same file, and 3.10.* to use another file, and so on.
With the eval command we can, by adding:

  WARNINGS_FILE =~ s/(-file-\d+\.\d+).*/$1/

Which will chop off the extra characters after the 3.8.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Add a special variable that can be used in other variables called
${KERNEL_VERSION}. This will embed the current kernel version into
the variable. For example:

WARNINGS_FILE = ${OUTPUT_DIR}/warnings-${KERNEL_VERSION}

If the current version is v3.8 then the WARNINGS_FILE will become

  ${OUTPUT_DIR}/warnings-v3.8
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

 Currently ktest sends SIGINT to terminate the console.
 However, there are consoles which do not exit by this signal, for example,
 in my case, "virsh console <guest OS>". In such case, ktest is blocked in
 close_console(). It prevents this automate test.

This patch adds new option CLOSE_CONSOLE_SIGNAL which mean the
signal to terminate the console. Since its default value is "INT",
the original behavior isn't changed.

Link: http://lkml.kernel.org/r/87zjol8pl5.wl%satoru.takeuchi@gmail.comSigned-off-by: NSatoru Takeuchi <satoru.takeuchi@gmail.com>
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Correct spelling typo in various part of drivers
Signed-off-by: NMasanari Iida <standby24x7@gmail.com>
Signed-off-by: NJiri Kosina <jkosina@suse.cz>

Different tests may use a different machine. In such cases, we need to
try to get the current grub menu index. If the same grub menu is used
for two different machines, it may not be at the same index on the
second machine. A search for the index must be performed again.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

To save connecting and searching for a given grub menu for each test,
ktest.pl will cache the grub number it found. The problem is that
different tests might use a different grub menu, but ktest.pl will
ignore it.

Instead, have ktest.pl check if the grub menu it used to cache the
content is the same as when it grabbed the menu. If not, grab it again,
otherwise just return the cached value.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The index of a line where a warning is tested can be returned
differently on different versions of gcc (or same version compiled
differently). That is, a tab + space can give different results. This
causes the warning check to produce a false positive. Removing the
index from the check fixes this issue.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The reboot just wants to get to the next kernel. But if a warning (Call
Trace) appears, the monitor will report an error, and the reboot will
think something went wrong and power cycle the box, even though we
successfully made it to the next kernel.

Ignore warnings during the reboot until we get to the next kernel. It
will still timeout if we never get to the next kernel and then a power
cycle will happen. That's what we want it to do.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Sometimes when a test kernel passed fine, but on reboot it crashed,
ktest could get stuck and not proceed. This would be frustrating if you
let a test run overnight to find out the next morning that it was stuck
on the first test.

To fix this, I made reboot check for the REBOOT_SUCCESS_LINE. If the
line was not detected, then it would power cycle the box.

What it didn't cover was if the REBOOT_SUCCESS_LINE wasn't defined or if
a 'good' kernel did not display the line. Instead have it search for the
Linux banner "Linux version". The reboot just needs to get to the start
of the next kernel, it does not need to test if the next kernel makes it
to a boot prompt.

After we find the next kernel has booted, then we just wait for either
the REBOOT_SUCCESS_LINE to appear or the timeout.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Although the patchcheck test checks for warnings in the files that were
changed, this check does not catch warnings that were caused by header
file changes and the warnings appear in C files not touched by the
commit.

Add a new option called WARNINGS_FILE. If this option is set, then the
file it points to is read before bulid, and the file should contain a
list of known warnings. If a warning appears in the build, this file is
checked, and if the warning does not exist in this file, then it fails
the build showing the new warning.

If the WARNINGS_FILE points to a file that does not exist, this will
cause any warning in the build to fail.

A new test is also added called "make_warnings_file". This test will
create do a build and record any warnings it finds into the
WARNINGS_FILE. This test is something that can be run before other tests
to build a warnings file of "known warnings", ie, warnings that were
there before your changes.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

Options are allowed to use other options, for example:

  LOG_FILE = ${OUTPUT_DIR}/${MACHINE}.log

where the option LOG_FILE used the options OUTPUT_DIR and MACHINE.

But if a test option were to use a default option, it will not get
substituted:

  OUTPUT_DIR = ${THIS_DIR}/${MACHINE}

  TEST_START
  OUTPUT_DIR = ${OUTPUT_DIR}/t1

For the above test, OUTPUT_DIR will stay literally "${OUTPUT_DIR}/t1"
and not be converted to "${THIS_DIR}/${MACHINE}/t1". When the test runs,
it will pass the ${OUTPUT_DIR} to the shell, which would probaly
interpret it as "", and the output directory will end up as "/t1".

Change the code where if a test option has its own option name in
its defined field, and a default option exists, then substitute the
default option in its place.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>

The patchcheck test looks at what files are modified for each patch it
checks and makes sure that those files do not produce any warnings.

Unfortunately, when it read the diffstat, the newlines were added on the
files and this made compares miss warnings, and commits that should not
have passed, ktest let pass.

Fix this by using the perl command "chomp" that strips off whitespace at
the end of lines.
Signed-off-by: NSteven Rostedt <rostedt@goodmis.org>