Automatic regression tests:
---------------------------
The Valgrind gdbserver automatic tests are by default
executed as part of the Valgrind test suite:
make regtest
By default, these tests are running with the gdb found at
Valgrind configure time.
If you want to test with another gdb version, you can do:
make regtest GDB=/path/to/another/gdb
or (to just run the gdbserver tests with another gdb):
cd gdbserver_tests
make check
cd ..
gdbserver_tests/make_local_links /path/to/another/gdb
perl tests/vg_regtest gdbserver_tests
The minimum version to run the tests is gdb >= 6.5.
Previous versions do not have the 'target remote |' command. It
would be possible to use an older version by using the option
--port of vgdb, and using the tcp/ip variant of the 'target remote'
GDB command.
The tests have been run on various platforms using gdb versions >= 7.2
and on some platforms gdb 7.0 and 7.1.
Some gdb tests implies a gdb >= 7.2. (these are automatically disabled
if testing with a lower version).
Test behaviour with gdb < 7.0 is unknown: some might fail,
some might block or loop for a very long time.
Some tests implies to have a vgdb "ptrace invoker" capable.
The prerequisite are established during make regtest (using marker files).
Each test verifies the prerequisite using the prereq: line.
In case of failing tests
------------------------
When executed with a new gdb version and/or depending on the OS version,
gdbserver tests might often fail due to (irrelevant) differences.
Such irrelevant differences have to be filtered by gdbserver_tests/filter_gdb.
You are welcome to fix such bugs by enhancing filter_gdb.
Alternatively, to report such problems, the best is to re-run
the gdbserver tests the following way:
perl tests/vg_regtest --keep-unfiltered gdbserver_tests
Then file a bug in bugzilla, giving the following information:
output of
gdbserver_tests/gdb --version
uname -a
cat /etc/issue
valgrind --version (and/or svn version)
and attach a tar file containing all the *.out and *.diff
files in gdbserver_tests directory
If a gdbserver test fails for other reasons, you can run the test
manually in two windows:
In one window, the valgrind
In another window, you launch gdb yourself, and you copy paste
the command from xxxx.stdinB.gdb. This might help to see what is
wrong.
Another good trick is also to execute the same kind of actions
using a gdb connected to the gdbserver part of gdb.
You can examine what is happening by enabling the trace
of the packets being sent using the gdb command:
set debug remote 1
Note however that the packets might be different
(e.g. the Valgrind gdbserver understands the 'P' packet,
which might not be understood by the gdbserver of gdb).
Naming conventions:
-------------------
The gdbserver tests are done with various Valgrind tools. A gdbserver
test using a tool xxxxxx should have its name starting with the 'short
two letters code' of this tool. For example, the test mcvabits.vgtest
is using Memcheck tool, while the test mssnapshot.vgtest is using
massif tool.
Typically, a gdbserver test implies to launch two programs:
prog: a program running under valgrind
progB: another program (typically, either gdb or vgdb standalone)
The conventions about how to specify the 2nd program in a .vgtest
are explained in the file ../tests/vg_regtest.in
Many tests are using gdb as progB. The input for gdb is named
xxxxxx.stdinB.gdb.
One day, we might imagine to run tests in parallel.
So, each test gets its own '--vgdb-prefix' argument.
This also help to avoid interactions between two successive tests:
if a previous test stayed blocked, the vgdb of the next test
will not report an error due to multiple possible FIFOs.
Rational for test approach
--------------------------
Two approaches have been looked at:
V: use the 'vg_regtest' approach used by the rest of Valgrind tests
G: use the gdb Dejagnu test framework.
Advantages of V: much simpler that G, known by Valgrind developpers,
no additional dependency for the Valgrind build and test.
Disadvantages of V: not well suited to testing of interactive tools,
very unflexible way to test the results (everything is in "template"
files), templates contains often irrelevant data for the test but it
can make the test fail. After writing 13 tests, it looks like the
template diff approach is quite fragile (e.g. changing the gdb version
and/or the OS version influences the output of irrelevant things which
are part of the template).
A more flexible template filtering is needed.
Maybe something like:
The program under test is outputting its instructions to be filtered in
special markers e.g.
#pushf filter_addresses | filter_messages
... some output
#pushf an_additional_filter
... some other output, filtered by both the first and second push
#popf
... here output filtered only by the first pushf
#popf
Advantages of G: much more powerful, well suited to test a gdb with a
gdbserver, tests can verify specifically some output without being
impacted by other output, allow to test Valgrind gdbserver with the
all of the gdb test suite (but a lot of tests will rather test gdb
than Valgrind gdbserver).
Disadvantages: not an easy beast to understand and master, running the
whole gdb testsuite with Valgrind gdbserver looks to be a challenge.
Currently, tests are written with vg_regtest. Approach G will be looked at it
again (e.g. to complement V) once a basic set of tests are available.
Manual tests still to automate:
-------------------------------
Validate monitor commands abbreviation recognition
***************************************************
mo v.info all_errors # to show all errors recorded so far
mo v.i a # the same
mo v # must give an error: v can match v.set v.info
mo v # the same v
mo v. # the same v.
test of gdb detaching or dying
******************************
valgrind --vgdb=yes --vgdb-error=0 --vgdb-poll=500 ./t
in another window
gdb ./t
set remotetimeout 100
target remote|vgdb
detach valgrind continues
target remote|vgdb reattach
detach valgrind continues
target remote|vgdb reattach
monitor v.wait no effect
test of valgrind/gdb output redirection
***************************************
valgrind --vgdb=yes --vgdb-error=1 --vgdb-poll=500 ./t
in another window
**** command to type*** ****** expected behaviour
gdb ./t
set remotetimeout 1000
target remote | vgdb
mo v.set vgdb-error 1000 # so that valgrind does stop only at error 1000 and after
mo v.set gdb_output # to have further valgrind errors output in gdb
c # continue, some errors will appear
C-c # interrupt program
mo v.set log_output # to set back the valgrind output to normal process log output
mo l # leak output to appear in log of process
mo v.set mixed_output
mo l # leak output to appear in gdb
test with a big executable: firefox
***********************************
valgrind --vgdb=yes --vgdb-error=1000 --vgdb-poll=50000 --trace-children=yes firefox 2>&1 | tee f.out
wait for some messages from the "big" firefox executable to appear.
Then:
gdb /usr/lib/firefox-3.5/firefox
target remote | vgdb
... then you can do various next/print/bt/bt full/break/... to see it is working
bulk test with the above
************************
to verify there is no race condition/no reentrance problem
between gdbserver code and valgrind:
start firefox like in the previous test.
In another window, do:
while true
do
vgdb leak
sleep 1
done
NB: this will make firefox run extremely slow, as it will do a leak
search every second.
Test of "jump" functionality
----------------------------
... to be done : put two breaks, jump over one.
... same but when error is encountered
* test with --max-invoke-ms=0
-----------------------------
valgrind --vgdb=yes ./t
... wait till you see "petachounok sleeping 4 of 15
then try to gdb it
!!!! this often causes gdb to report a protocol timeout.
use gdb set remotetimeout <a big time> to avoid that.
The symptoms of a timeout are:
(gdb) tar rem|vgdb --max-invoke-ms=0
Remote debugging using |vgdb --max-invoke-ms=0
relaying data between gdb and process 2930
Ignoring packet error, continuing...
warning: unrecognized item "timeout" in "qSupported" response
* tests of shadow registers
----------------------------
Show/modify shadow registers