Hi Stephen,
On 6 February 2016 at 13:34, Stephen Warren swarren@wwwdotorg.org wrote:
On 02/06/2016 01:30 PM, Simon Glass wrote:
On 4 February 2016 at 16:11, Stephen Warren swarren@wwwdotorg.org wrote:
Implement command--line option --gdbserver COMM, which does two things:
a) Run the sandbox process under gdbserver, using COMM as gdbserver's communication channel.
b) Disables all timeouts, so that if U-Boot is halted under the debugger, tests don't fail. If the user gives up in the middle of a debugging session, they can simply CTRL-C the test script to abort it.
This allows easy debugging of test failures without having to manually re-create the failure conditions. Usage is:
Window 1: ./test/py/test.py --bd sandbox --gdbserver localhost:1234
Window 2: gdb ./build-sandbox/u-boot -ex 'target remote localhost:1234'
When using this option, it likely makes sense to use pytest's -k option to limit the set of tests that are executed.
Simply running U-Boot directly under gdb (rather than gdbserver) was also considered. However, this was rejected because:
a) gdb's output would then be processed by the test script, and likely confuse it causing false failures.
b) pytest by default hides stdout from tests, which would prevent the user from interacting with gdb.
While gdb can be told to redirect the debugee's stdio to a separate PTY, this would appear to leave gdb's stdio directed at the test scripts and the debugee's stdio directed elsewhere, which is the opposite of the desired effect. Perhaps some complicated PTY muxing and process hierarchy could invert this. However, the current scheme is simple to implement and use, so it doesn't seem worth complicating matters.
c) Using gdbserver allows arbitrary debuggers to be used, even those with a GUI. If the test scripts invoked the debugger themselves, they'd have to know how to execute arbitary applications. While the user could hide this all in a wrapper script, this feels like extra complication.
An interesting future idea might be a --gdb-screen option, which could spawn both U-Boot and gdb separately, and spawn the screen into a newly created window under screen. Similar options could be envisaged for creating a new xterm/... too.
--gdbserver currently only supports sandbox, and not real hardware. That's primarily because the test hooks are responsible for all aspects of hardware control, so there's nothing for the test scripts themselves can do to enable gdbserver on real hardware. We might consider introducing a separate --disable-timeouts option to support use of debuggers on real hardware, and having --gdbserver imply that option.
Signed-off-by: Stephen Warren swarren@nvidia.com
test/py/conftest.py | 8 ++++++++ test/py/tests/test_sleep.py | 7 ++++--- test/py/u_boot_console_base.py | 3 ++- test/py/u_boot_console_sandbox.py | 5 ++++- test/py/u_boot_spawn.py | 12 ++++++++---- 5 files changed, 26 insertions(+), 9 deletions(-)
Can you please add info about this to the docs?
Also for me this worked up to the point where it ran the test_sandbox_exit.py test. Then the gdb process said that U-Boot exited normally. Is that test not compatible with this feature?
The sandbox_exit test deliberately causes the sandbox process to exit, to make sure that the "reset" command and "typing" Ctrl-C work. To continue the test, simply re-run gdb to re-attach to the new gdbserver and U-Boot process.
That's a bit annoying. Perhaps we should have a flag that disabled such tests? Or perhaps a way to specify what tests are run?
Anyway, can you add this to the docs too?
Regards, Simon