GitLab

Pure cosmetics: We are no longer pulling from a mirror but from the
official source.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Debian 11 is almost ready. It comes with newer gcc-10. So better test
now to avoid surprises.

We just need to hold powerpc at Debian 10 as it does not build with
gcc-10. It's also unclear if the removal of SPE support in recent gcc
would affect us.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

When running in QEMU, the 5 ms holding time of mutex_timed_locker may be
too short when the host is busy. This occasionally leads to failing
tests in our CI:

posix-mutex.c:510, FAILED: get_effective_prio() == prio (=false)

Expand to 50 ms in such setups.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

RTTST_RTIOC_RTDM_DEFER_CLOSE causes a 300 ms delay by the driver. We
only added 1 ms plus the open/close of the second device, and that was
sometimes too short when running in QEMU:

2021-06-18T18:58:41 FAILED run_rtdm:186: open returned -16 instead of 12 - Device or resource busy

Make it 100 ms in that case, to be safe.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

At least on ARM, an illegal instruction event that could also mean
"breakpoint" can become an FPU event after analyzing it in
xnarch_handle_fpu_fault. Avoid registering the process as being stopped
for debugging first - this would be false alarm, and it would not be
unregistered any time soon.

This was exposed by the ptrace helper addition.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

The space allocated for irq stats in xnintr_init() will not be freed
by rtdm_irq_free(), because it just calls xnintr_detach(). Replacing
it with xnintr_destroy() fixes the issue.
Signed-off-by: Wolfgang Grandegger <wg@grandegger.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

The core assumes in some places that CPU 0 is available for real-time.
Therefore fail if the mask is excluding it.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Scheduling a thread on a CPU that is not supported by Xenomai will only
fail and flood the screen with error messages.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

The I-pipe implementation was lucky in its approach to not wake up the
RT thread in case the non-RT part migrated to an unsupported CPU. For
unclear reasons, the scheduled cancellation made it to the thread and
avoided any lockup. On dovetail, this fails.

So switch both implemenations to the safer pattern of actually
completing the migration, using the previous affinity of the thread. It
will still be canceled right after that, now also over dovetail.

As cobalt_affinity_ok's return value becomes irrelevant, refactor this
function to cobalt_adjust_affinity without and return code.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

__this_cpu_read will generate a warning when called by
print_core_clock_status because the context is detected as preemptible.
Avoid this by picking up the name from CPU 0, just like ipipe_timer_name
did.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Rather than duplicating lots of code for the sake of compat userspace
accesses, provide a common version that is given a flag to decide
whether to take the compat or regular code paths.

This also brings restart support to the compat selection version which
was forgotten by 36132cdb

.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

There is no use for this outside of the core that implements select. So
stop exporting these helpers. As there are no users outside, simplify
their return value on success to 0 because their callers do not care
about more.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Commit 36132cdb

 forgot to align this with the native syscall.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Since rtdm_fd_select started to return EADV for unknown fds
(2db562ad

), we were returning that code to applications in case a
selector had been allocated for the caller already. That was because
libcobalt lacked handling of this error.

This fixes it and also establish proper signaling of EBADF in case the
is a real error related to an RTDM fd. For that purpose, detect if
__cobalt_select_bind_all got EADV for the first fd or for a succeeding
one, i.e. on an unsupported mixed fd set.

Also return EADV if the early __cobalt_first_fd_valid_p test fails. That
ensures that we only retry with the regular Linux call when it makes
sense.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

2db562ad

 had a point in avoiding EBADF for file descriptors that are
not under RTDM management. But it unfortunately got a lot of other cases
wrong along this.

This one is about returning the right error code if an RTDM driver does
not implement a particular handler. That must never be EADV because that
error would signal "not an RTDM fd, retry with Linux" which is obviously
wrong. Even worse, mmap clearly defines the right error to return in
case it is not implemented, and that is ENODEV as was returned before.

So this restores ENODEV as invalid default for mmap, uses ENOTTY for
unimplemented IOCTL and switches all others to EBADF which is a match
for select and a reasonable choice for read/write/recvmsg/sendmsg.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

It also returns -EADV since 2db562ad

 - causing troubles in select.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

It worked, but it is not nice to read.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Already at the time of introduction (8ac3db81

), -ENOENT was not
returned by any sane function behind rtdm_fd_select(). Today
rtdm_fd_select() officially only returns -EBADF or -EINVAL. By now it
also returns -EADV, via rtdm_fd_get() or a default select handler. The
only remaining possibility for -ENOENT is a select handler. In-tree,
a4l_rt_select actually has such a case, but for no sane reason this or
any out-of-tree driver case should receive special handling here.

Remove this case, these is also no need for the select_bind_one wrapper
anymore.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This caused random values in those vars, and that led to premature
leaving of __bufp_readbuf and __bufp_writebuf with waiter states not
properly updated, e.g.

Fixes: 0a67a85b

 ("drivers/ipc: bufp: fix rd/rd and wr/wr races")
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Add rtdm_irq_request_affine and rtdm_irq_set_affinity to start IRQs or
specific CPUs or move them around. So far, this was unsupported and made
users bypass layer to configure the desired affinity. Starting with the
right affinity also avoids hitting the wrong CPU at least once before
the IRQ can be live-migrated.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This allows to start up interrupts with a specific affinity already set.
Will be used by a new RTDM service.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This function should respect the supported cpu set and report an error
if the specified set has no intersection with it. It should also return
potential errors for the underlying pipeline code.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Passing a potentially large mask by value was a bad idea of I-pipe.
Dovetail uses the kernel API, and that does it right. So adjust our
internal abstraction - which is unused so far anyway.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This is needed when there is not userspace-readable clocksource
available. This can happen under Dovetail where Linux selects the
suitable clocksource and may see problems with, e.g., the TSC under x86.
When it then switches to a clocksource that require a syscall, we so
far migrate to Linux on every access.

This handler resolves that problem. It's not needed under I-pipe where
clock_gettime is not routed through the vDSO and the clocksource is
separate. Therefore, we simply never match on the related syscall
numbers under I-pipe.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Add missing sigarray initialization to avoid missing signal in
__xnthread_signal.
Signed-off-by: Hongzhan Chen <hongzhan.chen@intel.com>
[Jan: simplify size calculation]
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Same pattern as for previous changes: All that code it common, no need
to duplicate it. The only difference was that I-pipe needs to call
ipipe_disable_user_intret_notifier first while dovetail does this
disabling automatically.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Same pattern as for previous changes: All that code it common, no need
to duplicate it. The only difference between the two versions of
handle_cleanup_event was what is now encapsulated in
pipeline_cleanup_process.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Same pattern as for previous changes: All that code it common, no need
to duplicate it.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Those functions are identical now, with the variations due to some older
I-pipe releases removed.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

All those events have been added long ago to I-pipe versions we want to
support with the next release. Time to drop the tests which will also
allow sharing some code with dovetail.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

I-pipe makes sure that arguments of compat calls are ordered just like
native calls. Dovetail does not. But it is better to use the kernel's
syscall_get_arguments for retrieving the arguments anyway. Introduce
pipeline_get_syscall_args to abstract that difference.

Xenomai/ARM syscalls have an unconventional argument marshalling, with
ARM_ORIG_r0 containing the Xenomai syscall number (ORed with
__COBALT_SYSCALL_BIT), followed by the first actual argument to the
syscall in ARM_r1.

For this reason, using syscall_get_arguments() to collect the syscall
parameters won't yield the expected result, with regs[0] matching
__xn_reg_sys(regs) instead of __xn_reg_arg1(regs).

Allow Dovetail-enabled architectures to override the generic
implementation of pipeline_get_syscall_args(), provide a working one
for ARM.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
[Philippe: add special handling of ARM]
Signed-off-by: Philippe Gerum <rpm@xenomai.org>

At this chance, fix stale and missing copyright notices. Specifically,
Dmitriy Cherkasov <dmitriy@mperpetuo.com> is the original author of
the Xenomai/arm64 port, based on the ARM implementation; add credit
where it is due.

In the same move, introduce SPDX tags as replacement for GPL
boilerplate for the altered files.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Unlike the I-pipe, Dovetail does not need to expose the underlying
timestamp counter to applications. The best equivalent to this counter
would be the common monotonic clock, expressing time as a count of
nanoseconds.

Fork the implementation accordingly between Dovetail and I-pipe
configurations, depending on which pipeline flavour was detected at
init.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Expiration dates of in-band timers are based on the monotonic time
base read by ktime_get(), which is ok for us to use from non-NMI
context.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

When Dovetail is enabled in the kernel, we get timestamps for common
clocks (CLOCK_REALTIME, CLOCK_MONOTONIC) from the vDSO-based gettime()
routine as usual.  Do not expect KUSER_TSC support to be there in such
a case, it is not.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

When Dovetail is enabled in the kernel, we get timestamps for common
clocks (CLOCK_REALTIME, CLOCK_MONOTONIC) from the vDSO-based gettime()
routine as usual.  Do not expect KUSER_TSC support to be there in such
a case, it is not.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>