The nOS-V events are modified in two ways: 1) to create a parallel task
the new VTC event must be used and 2) all task events for both normal
(VTc) and parallel (VTC) tasks require an extra argument in the payload
to indicate the body id. As a consequence, the nOS-V model version is
now increased to 2.0.0.
Additionally, all the channel PRV flags are set to PRV_SKIPDUPNULL, so
duplicates are only emitted if they are not null. It solves the problem
when a task switches to another task with the same body id.
A new Paraver configuration is added for the body id.
The task model is now integrated with the body model. A normal task can
only have one body, while a parallel task can have more.
It inherits the restriction that a task body cannot be nested over
another one unless it is paused (or the relaxed nest model is enabled).
The new task body model (or just body model) allows a task to have
multiple bodies. Generally a body is mapped to the execution of the same
user code of the task with specific input arguments.
The body model can constraint if a given body can be paused or re-run
again (resurrect).
Additionally, the body model can run multiple nested bodies but with the
restriction that the parent body should be paused first. This condition
can be relaxed with the BODY_FLAG_RELAX_NESTING flag.
This flag allows to only skip duplicates if the value is null, otherwise
the value is emitted. Invalid combinations of PRV_* flags are also
checked on prv_register().
Instead of showing the "attached" state with the VH{aA} events, we show
when the call to nosv_attach() and nosv_detach() take place. The old
VH{aA} events are now ignored. Bumps the nOS-V model version to 1.1.0.
Implements a small language parser to define the emulator events. The
event specification is parsed at emulation (when the emulator runs).
The ovnidump output now prints the events with the arguments formatted
as given in the event description.
It also introduces some consistency checks over the event MCVs, which
must begin with the model identifier and cannot be duplicated.
Until now, the value returned by ovni_thread_isready() was still
non-zero when the thread stream was destroyed in ovni_thread_free()
This was making it impossible to detect when a stream was destroyed.
This change makes it return 0 after ovni_thread_free() by using a new
finished flag. The flag keeps track on then the stream is destroyed,
preventing double initialization of a free stream.
This is required on FS like NFS as the cache coherence model states that
only after the file is closed the changes are actually propagated to the
server. It is not recomended to use NFS as OVNI_TMPDIR anyway.
Fixes: https://pm.bsc.es/gitlab/rarias/ovni/-/issues/168
As cmake doesn't provide a way to use the RPATH or the RUNPATH, we
explicitly set the --enable-new-dtags linker flag. If the flag is not
supported by the compiler, the build is aborted.
When running ovniver without LD_LIBRARY_PATH set, the loader couldn't
find any ovni installation and therefore was failing to start the
program. With this change we set the RUNPATH (not RPATH) of ovniver to
the installation libdir, so when no other libovni library is loaded via
LD_LIBRARY_PATH the one provided by the same source as ovniver is used.
The value of LD_LIBRARY_PATH is also printed so we can determine where
it may be loading libovni. Using LD_DEBUG=libs will give more
information of the search process.
Fixes: https://pm.bsc.es/gitlab/rarias/ovni/-/issues/163
The experimental flag -a is used to ease the transition to the usage of
ovni_thread_require(), as it may be posible to have traces in which not
all libraries have requested their model. The flag causes all emulation
models to be enabled. This flag is considered experimental and it may be
removed or renamed in future versions.
The thread metadata file is mandatory for trace version 2, so the
emulator should fail if the file is not there or has the wrong version.
Old traces with version 1 are not supported.
There were leftover temporal directories when using the OVNI_TMPDIR
variable, as we only remove the proccess ones. This change only attempts
to remove them if they are empty, and doesn't care about race
conditions. The removal process is just "best effort".
The metadata was being written to the final directory, which may be too
slow to accept writes during runtime. The thread metadata needs to be
written anyway at ovni_thread_free() to set the finished flag, so we can
simply skip it.
The emulator will now check that all threads are properly finalyzed by
calling ovni_thread_free(), as required by the specification. For now
only a warning is issued, which is enough to determine the cause of
potential emulator panics.
The ovni model is now always enabled.
Allows programs to update to a new libovni library without breaking the
instrumentation. Only until the first call to ovni_thread_require() the
models are enabled on demand.
Traces with metadata version 1 don't use per-thread information, so we
cannot read the 'require' key. All models will be enabled for old traces
(which will cause performance problems).
Until now, emulation models were always being registered via probe(),
which causes the emulator to initialize all the channels. To reduce the
overhead, the channels were not connected or registered in the bay
until the first event of that model was received. This delayed connect
was causing issues in muxes where the newly connected model required
refreshing the touched channels. Which in turn was causing unexpected
PRV events.
By determining which models we need to enable, we can remove the delayed
connect mechanism and just enable those models at initialization time,
and connect the channels.
When running without all the subsystem events enabled in nOS-V, it is
possible to emit two consecutive VTx events, which would push twice the
same value ST_TASK_BODY into the subsystem channel. This change relaxes
the subsystem channel to accept duplicate stacked values. A regression
test is also added.
Reported-By: Raúl Peñacoba Veigas <raul.penacoba@bsc.es>
In nOS-V, when a task was paused via the VTp event, two things were
happening: 1) the task state was set to pause and 2) the subsystem state
"Task: Running" was being popped.
This causes a problem when a task calls nosv_submit() in blocking mode,
as it will call nosv_pause() which will emit a VTp event from a
subsystem different than "Task: Running".
To solve this conflict, we handle the subsystems state and the task
state separately with the VTp and VTr events. The subsystem state "Task:
Running" no longer is connected to the state of the task and only shows
if we entered the body of the task or not. It has now been renamed to
"Task: In body".
The new state "Task: In body" represents that the task body has begun
the execution and is still in the stack, but the task may be paused. The
subsystem is not changed by the VTp (pause) or VTr (resume) events.
Fixes: https://pm.bsc.es/gitlab/rarias/ovni/-/issues/128
The check_clock_gate() was trying to access to the first event in all
streams. However, streams that are empty don't have any event, so they
cause a NULL dereference. Skipping not active streams avoids the
problem.
This commit introduces the OVNI_TRACEDIR environment variable to change
the directory where traces are generated. By default, when the envar is
not defined, the trace is still generated in the ovni directory. The
envar can take a trace directory name, a relative path to the directory,
or its absolute path. In the first case, the directory is created in the
current path $PWD.
Both libovni (rt) and ovnisync read this environment variable.
