GitLab

The rtdm test relies on successively free file descriptors which is
violated by this test when the run after each other.
Reported-by: Gylstorff Quirin <quirin.gylstorff@siemens.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

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

This is where OpenSUSE puts them, maybe others as well.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Concurrent reads or writes to a buffer causing nested buffer space
reservations could end up in a miscalculation of the number of bytes
remaining to be read or written. Fix this by updating these counts
only once the last reservation is committed.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Since the vfile export and unexport operations are asynchronous,
returning from xnregistry_remove() is no guarantee that the registered
object won't be further accessed, especially by the vfile export
handler.

Plug this race at least for all in-band callers removing objects while
running on root stage like RTIPC protocols by synchronizing with the
workqueue which handles deferred export/unexport requests, before
returning from xnregistry_remove().

This does not cover the issue of removing objects from the head
stage. Fortunately, all users of the vfile export/unexport mechanism
are unregistering objects from the root stage only (typically some
RTDM close() handler).

This issue was reported by KASAN.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Running xnregistry_enter() then xnregistry_remove() in quick
succession for the same object might cause the latter to NULLify the
objaddr field while proc_callback() is busy calling the vfile export
handler for the object, which is going to use that same value,
certainly leading to a crash.

This bug is due to the asynchronous nature of the vfile
export/unexport process which is deferred to a regular workqueue on
the root stage when kicked from the head stage.

Fix this by adding a third logic state describing an aborted vfile
export procedure, due to handling an object removal request while the
workqueue is concurrently running for the same object. In addition,
since an exported object might be stale before it is considered for
export, do not anticipate on the next object to process in
proc_callback() as list_for_each_entry_safe() would do, but always
pick the list head under lock.

At this chance, clarify the naming of the magic vfile values
enumerating those states (i.e. XNOBJECT_EXPORT_xxx).
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

We need this information for debugging purpose when latex stop mode is
enabled (--enable-verbose-latex=yes).
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

WARN_IF_UNDOCUMENTED is switched to "NO" to prevent a truckload of
useless warnings from Doxygen since most ioctl calls are not
documented.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

When we wait for all passed event bits to be set, we need to test for
those bits, not the ones that are currently pending. The current code
fails to detect that the condition is met when both the desired as well
as other events are already pending.

Reported by: Ronny Meeus <ronny.meeus@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

When we wait for all passed event bits to be set, we need to test for
those bits, not the ones that are currently pending. The current code
fails to detect that the condition is met when both the desired as well
as other events are already pending.

Reported by: Ronny Meeus <ronny.meeus@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This fails so far due to a bug in the event wait implementations, both
with cobalt and mercury.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This feature might not be widely used but those debian packages are
supposed to be generic. Having it default on brings more benefits than
drawbacks for generic packages.
Signed-off-by: Henning Schild <henning.schild@siemens.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Updated SPI Makefile and Kconfig in order to contain the option
to compile am335x beaglebone black omap4 spi real-time driver.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
[Jan: enabled in CI build]
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

At least the copied-in files can and should be removed on uninstall, but
we do not make any efforts to remove us again from a pre-existing
udev.rules.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

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

The ioctl_n user parameter must be set to non-zero value
in order to perform ioctl(SPI_RTIOC_TRANSFER_N, TRANSFER_SIZE);
if ioctl_n is 0 or undefined then the test performs ioctl(SPI_RTIOC_TRANSFER).
Default SPI clock speed is set to 40MHz instead of 60MHz
because bbb am335x realtime SPI driver has its maximum set to this value.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

spi-omap2-mcspi-rt.c:
I/O handling is lifted from  omap2/4 mcspi linux driver.
The code follows the model in spi-bcm2835.c.
The theoretical maximum spi frequency of am335x omap4 is 48MHz,
and the driver practical maximum frequency is limited to 40MHz.
The omap4 maximum fifo queue size is 64 bytes,
but when the transfer implies both sending and receiving
it is limited to 32, so, in the driver things are kept simple:
max 32 for tx and max 32 for rx.
When the driver must transfer N bytes
then ((int)(N/32))*32 bytes are transfered with fifo queue size 32
and (N mod 32) bytes are transfered with fifo queue size (N mod 32);
between fifo setup, the slave chip select line is kept active
by using the FORCE bit in the channel configuration register.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

spi-bcm2835.c, spi-sun6i.c:
In order to support SPI_RTIOC_TRANSFER_N ioctl option
add functions bcm2835_transfer_iobufs_n() and sun6i_transfer_iobufs_n()
which are similar to ..._transfer_iobufs(),
but allow user specified len for spi transfer.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

SPI_RTIOC_TRANSFER_N allows the user
to make SPI transfers of arbitrary lengths
by using the same mmapped memory area.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

spi-master.c: spi_master_read_rt(): when master->ops->read()
completes successfully it returns > 0, so, in this case,
copy data to userspace.
Signed-off-by: Laurentiu-Cristian Duca <laurentiu.duca@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Maintenance will continue based on that branch only.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

f116e7eb already tried to enable live updates of a quota group but
got one thing wrong and missed another catch:

First, we must not use group_is_active() to check if the group is
currently consuming time. That has to be the same logic as in
xnsched_quota_pick(), i.e. a check of the currently running thread on
the target sched belongs to the modified group.

The second issue is that, if a group had no budget on refill or consumed
it in this period already, an extension of the budget missed to move
expired threads back to the runqueue.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Add missing forward declaration of timex, avoid incorrect alias
creations by COBALT_IMPL due to wrongly defined fputc and putchar.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Make use of the kernel interface change in c20ed9f1 to fix
truncation on 64-bit systems when passing pointers.
Reported-by: Norbert Lange <nolange79@gmail.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Currently not an issue for the kernels we pull here, but 4.14-ppc would
break e.g. because ls-remote --tags gives

... ipipe-core-4.14.36-ppc32-1.1
... ipipe-core-4.14.36-ppc32-1.1^{}

with the second one representing the dereferentiation of the annotated
tag. We only need the first one, though, while the second breaks git
clone. --refs filters this.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

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

This was missed in cfa2c51d.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Kicking the rescheduling procedure (xnsched_run()) from call sites
running over the root thread with hard irqs enabled may lead to
missing a rescheduling opportunity as a consequence of a CPU
migration, e.g.:

    CPU0					CPU1
    ----                         		----
    t0: ...
            <context switch>
    t1: xnthread_resume(t0)
             xnsched_run()
             <CPU MIGRATION -->
                  |                             |
                  |                             |
                  x                             |
          (missed rescheduling)                 v
                                    ___xnsched_run(CPU1.sched)

To address this issue, make sure every call to xnsched_run() either
runs from the head domain which is not affected by CPU migration, or
moves into the latest atomic section which might have changed the
scheduler state.

Turning on CONFIG_XENO_OPT_DEBUG_COBALT also enables a (oneshot)
assertion which detects invalid calls from the root domain with hard
irqs enabled.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

rtdm_nrtsig_pend() is based on ipipe_post_root_work(), which keeps a
copy of every request descriptor internally until it is consumed by
the secondary mode handler. Triggering rtdm_nrtsig_pend() every time a
packet is relayed from the stack to the tap device may cause such
request buffer to overflow under pressure.

To address this issue, trigger the non-rt signal only when the RX
queue transitions from empty to non-empty state as a result of
enqueuing the next skb.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This feature is useful for applications tuning scheduling parameters
from within RT contexts. Primarily a topic for legacy applications. But
as this doesn't come with noteworthy downsides, make it available in the
generic package configuration.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Now that this path goes via an own syscall, better test it.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

As we didn't migrate to RT, we didn't test properly if lazy parameter
application actually works. Fix that, also taking into account the
related config switch.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This reverts commit b3c70f1a.

Was broken by design and is now no longer needed.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

Switch the implementation of pthread_setschedprio to the new cobalt
syscall in order to avoid scheduler setting corruptions for SCHED_QUOTA
and similar complex policies.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

The approach of using thread_setschedparam_ex with __SCHED_CURRENT
failed for policies with complex parameter sets. Therefore, introduce a
syscall that is fully aware of how to apply only the priority change to
the target thread's scheduling settings.

We need a compat path as well because we are passing down pointers.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

This pattern will be reused by cobalt_thread_setschedprio.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

rtdm_nrtsig_pend() is based on ipipe_post_root_work(), which keeps a
copy of every request descriptor internally until it is consumed by
the secondary mode handler. Triggering rtdm_nrtsig_pend() every time a
packet is relayed from the stack to the tap device may cause such
request buffer to overflow under pressure.

To address this issue, trigger the non-rt signal only when the call
completion queue transitions from empty to non-empty state.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>

rtdm_nrtsig_pend() is based on ipipe_post_root_work(), which keeps a
copy of every request descriptor internally until it is consumed by
the secondary mode handler. Triggering rtdm_nrtsig_pend() every time a
packet is relayed from the stack to the VNIC may cause such request
buffer to overflow under pressure.

To address this issue, trigger the non-rt signal only when the RX
queue transitions from empty to non-empty state as a result of
enqueuing the next skb.
Signed-off-by: Philippe Gerum <rpm@xenomai.org>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>