[PATCH 3/8] nptl: Fix testcases for new pthread cancellation, mechanism

[Adhemerval Zanella]

On 08-10-2014 20:12, Joseph S. Myers wrote:
> On Wed, 8 Oct 2014, Adhemerval Zanella wrote:
>
>> For testcases that are not tested in blocked case (pread64 for instance)
>> tst-cancel4/5 tests for its cancellation point as 'early exit'.  Although
>> this is not ideal, 'tst-cancel-wrappers.sh' also does not offer more
>> coverage. Ideally and as pointed out in testcases itself, the best option
>> would be to create a way to actually block some cancellable entrypoint.
>> However, for some syscalls calls this is very trick (pread64 for instance
>> does not work on pipe-type descriptors).
> I still don't understand exactly what the effects are on test coverage.
>
> Could you provide a list of all the functions tested in 
> tst-cancel-wrappers.sh, with, for each function, a statement either of the 
> other testcase that tests cancellation of that function, or of why it is 
> not possible to test cancellation of that function in a useful way?
>
I understand your hesitation about remove tst-cancel-wrappers.sh testcase,
however currently it does not offer any more coverage than current other tests.
Below is a chart I compiled with the function tst-cancel-wrappers.sh is intended
to check:

syscall         | testcase       | blocked case | early exit | Notes
----------------|----------------|--------------|------------|------
accept          | tst-cancel4.c  |          yes |        yes |
close           | tst-cancel4.c  |           no |        yes |
connect         | tst-cancel4.c  |           no |        yes |
creat           | tst-cancel4.c  |           no |        yes |
fcntl           | tst-cancel16.c |           no |        yes | *
fdatasync       | tst-cancel4.c  |           no |        yes |
fsync           | tst-cancel4.c  |           no |        yes |
msgrcv          | tst-cancel4.c  |          yes |        yes |
msgsnd          | tst-cancel4.c  |           no |        yes |
msync           | tst-cancel4.c  |           no |        yes |
nanosleep       | tst-cancel4.c  |          yes |        yes |
open/open64     | tst-cancel4.c  |           no |        yes |
pause           | tst-cancel4.c  |          yes |        yes |
poll            | tst-cancel4.c  |          yes |        yes |
pread/pread64   | tst-cancel4.c  |           no |        yes |
pselect         | tst-cancel4.c  |          yes |        yes |
pwrite/pwrite64 | tst-cancel4.c  |           no |        yes |
read            | tst-cancel4.c  |          yes |        yes |
readv           | tst-cancel4.c  |          yes |        yes |
recv            | tst-cancel4.c  |          yes |        yes |
recvfrom        | tst-cancel4.c  |          yes |        yes |
recvmsg         | tst-cancel4.c  |          yes |        yes |
select          | tst-cancel4.c  |          yes |        yes |
send            | tst-cancel4.c  |          yes |        yes |
sendmsg         | tst-cancel4.c  |           no |        yes |
sendto          | tst-cancel4.c  |           no |        yes |
sigpause        | tst-cancel4.c  |          yes |        yes |
sigsuspend      | tst-cancel4.c  |          yes |        yes |
sigwait         | tst-cancel4.c  |          yes |        yes |
sigwaitinfo     | tst-cancel4.c  |          yes |        yes |
tcdrain         | tst-cancel4.c  |           no |        yes |
wait            | tst-cancel4.c  |          yes |        yes |
waitid          | tst-cancel4.c  |          yes |        yes |
waitpid         | tst-cancel4.c  |          yes |        yes |
write           | tst-cancel4.c  |          yes |        yes |
writev          | tst-cancel4.c  |          yes |        yes |

* tested through lockf (F_LOCK) which calls fctnl (F_SETLK)

The 'blocked case' case refers to a cancellation signal send while the syscall
is blocked and the test checks cancellation is done properly. As the 'tst-cancel4.c'
states, some syscalls are hard to put in a blocked case (pread/close for instance)
to exercise all code patch. If you have suggestion of how to add blocked cases
I can add them in a following patch.

The 'early exit' checks if the cancellation entrypoint will check for pending
cancellation signal and act accordingly. As you can for this case 'tst-cancel4.c'
already cover alls the syscalls.