We can sometimes hit scenarios in the scx_userland scheduler where there
is work to be done in user space, but we incorrectly fail to run the
user space scheduler. In order to avoid this, we can use global
variables that are set from both BPF and user space. The BPF-side
variable reflects when one or more tasks have been enqueued, and the
user space-side variable reflects when user space has received tasks but
has not yet dispatched them.
In the ops.update_idle() callback, we can check these variables and send
a resched IPI to a core to ensure that the user-space scheduler is
always scheduled when there's work to be done.
Signed-off-by: David Vernet <void@manifault.com>
Improve build portability by including asm-generic/errno.h, instead of
linux/errno.h.
The difference between these two headers can be summarized as following:
- asm-generic/errno.h contains generic error code definitions that are
intended to be common across different architectures,
- linux/errno.h includes architecture-specific error codes and
provides additional (or overrides) error code definitions based on
the specific architecture where the code is compiled.
Considering the architecture-independent nature of scx, the advantages
of being able to use architecture-specific error codes are marginal or
negligible (and we should probably discourage using them).
Moving towards asm-generic/errno.h, however, allows the removal of
cross-compilation dependencies (such as the gcc-multilib package in
Debian/Ubuntu) and improves the code portability across various
architectures and distributions.
This also allows to remove a symlink hack from the github workflow.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Instead of implementing a garbage collector to periodically free up
exiting tasks' resources, implement a proper synchronous mechanism to
notify the user-space scheduler about the exiting tasks from the BPF
component, using the .disable() callback.
When the user-space scheduler receives a queued task with a negative CPU
number, it can then release all the resources associated with that task
(which currently includes only the entry in the TaskInfoMap for now).
This allows to get rid of the TaskInfoMap periodic garbage collector
routine, save a lot of syscalls in procfs (used to check if the pids
were still alive), and improve the overall scheduler performance.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
No functional change, make the user-space scheduler code a bit more
readable and more Rust idiomatic.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The forumla used to evaluate the weighted time delta is not correct,
it's not considering the weight as a percentage. Fix this by using the
proper formula.
Moreover, take into account also the task weight when evaluating the
maximum time delta to account in vruntime and make sure that we never
charge a task more than slice_ns.
This helps to prevent starvation of low priority tasks.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Prevent newly created short-lived tasks from starving the other tasks
sitting in the user-space scheduler.
This can be done setting an initial vruntime of (min_vruntime + 1) to
newly scheduled tasks, instead of min_vruntime: this ensures a
progressing global vruntime durig each scheduler run, providing a
priority boost to newer tasks (that is still beneficial for potential
short-lived tasks) while also preventing excessive starvation of the
other tasks sitting in the user-space scheduler, waiting to be
dispatched.
Without this change it is really easy to create a stall condition simply
by forking a bunch of short-lived tasks in a busy loop, with this change
applied the scheduler can handle properly the consistent flow of newly
created short-lived tasks, without introducing any stall.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Never dispatch the user-space scheduler to the global DSQ, while all
the other tasks are dispatched to the local per-CPU DSQ.
Since tasks are consumed from the local DSQ first and then from the
global DSQ, we may end up starving the scheduler if we dispatch only
this one on the global DSQ.
In fact it is really easy to trigger a stall with a workload that
triggers many context switches in the system, for example (on a 8 cores
system):
$ stress-ng --cpu 32 --iomix 4 --vm 2 --vm-bytes 128M --fork 4 --timeout 30s
...
09:28:11 [WARN] EXIT: scx_rustland[1455943] failed to run for 5.275s
09:28:11 [INFO] Unregister RustLand scheduler
To prevent this from happening also dispatch the user-space scheduler on
the local DSQ, using the current CPU where .dispatch() is called, if
possible, or the previously used CPU otherwise.
Apply the same logic when the scheduler is congested: dispatch on the
previously used CPU using the local DSQ.
In this way all tasks will always get the same "dispatch priority" and
we can prevent the scheduler starvation issue.
Note that with this change in place dispatch_global() is never used and
we can get rid of it.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
With commit 49f2e7c ("scx_rustland: enable SCX_OPS_ENQ_LAST") we have
enabled SCX_OPS_ENQ_LAST that seems to save some unnecessary user-space
scheduler activations when the system is mostly idle.
We are also checking for the SCX_ENQ_LAST in the enqueue flags, that
apparently it is not needed and we can achieve the same behavior
dropping this check.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
In any case make sure that we never account more than the maximum
slice_ns to a task's vruntime.
This helps to prevent starving a task for too long in the user-space
scheduler.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The user-space scheduler maintains an internal hash map of tasks
information (indexed by their pid). Tasks are only added to this hash
map and never removed. After running the scheduler for a while we may
experience a performance degration, because the hash map keeps growing.
Therefore implement a mechanism of garbage collector to remove the old
entries from the task map (periodically removing pids that don't exist
anymore).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
In the scheduler statistics reported periodically to stdout, instead of
showing "pid=0" for the CPU where the scheduler is running (like an idle
CPU), show "[self]".
This helps to identify exactly where the user-space scheduler is running
(when and where it migrates, etc.).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Bypass the user-space scheduler for kthreads that still have more than
half of their runtime budget.
As they are likely to release the CPU soon, granting them a substantial
priority boost can enhance the overall system performance.
In the event that one of these kthreads turns into a CPU hog, it will
deplete its runtime budget and therefore it will be scheduled like
any other normal task through the user-space scheduler.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Use dispatch_on_cpu() when possible, so that all tasks dispatched by the
user-space scheduler gets the same priority, instead of having some of
them dispatched to the global DSQ and others dispatched to the per-CPU
DSQ.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Make sure the scheduler is not activated if we are deadling with the
last task running.
This allows to consistency reduce scx_rustland CPU usage in systems that
are mostly idle (and avoid unnecessary power consumption).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
When a task is dispatched we always try to pick the previously used CPU
(if idle) to minimize the migration overhead. Alternatively, if such CPU
is not available, we pick any other idle CPU in the system.
However, we don't update the list of idle CPUs as we dispatch tasks,
therefore we may end up sending multiple tasks to the same idle CPU (if
their previously used CPU is the same) and we may even skip some idle
CPUs completely.
Change this logic to make sure that we never dispatch multiple tasks to
the same idle CPU, by updating the list of idle CPUs as we send tasks to
the BPF dispatcher.
This also avoids dispatching tasks with a closely matched vruntime to
the same CPU, thereby negating the advantages of the vruntime ordering.
With this change in place, we ensure that tasks with a similar vruntime
are dispatched to different CPUs, leading to significant improvements in
latency performance.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
When the scheduler decides to assign a different CPU to the task always
make sure the assignment is valid according to the task cpumask. If it's
not valid simply dispatch the task to the global DSQ.
This prevents the scheduler from exiting with errors like this:
09:11:02 [WARN] EXIT: SCX_DSQ_LOCAL[_ON] verdict target cpu 7 not allowed for gcc[440718]
In the future we may want move this check directly into the user-space
scheduler, but for now let's keep this check in the BPF dispatcher as a
quick fix.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The current CPU selection logic in the scheduler presents some
inefficiencies.
When a task is drained from the BPF queue, the scheduler immediately
checks whether the CPU previously assigned to the task is still idle,
assigning it if it is. Otherwise, it iterates through available CPUs,
always starting from CPU #0, and selects the first idle one without
updating its state. This approach is consistently applied to the entire
batch of tasks drained from the BPF queue, resulting in all of them
being assigned to the same idle CPU (also with a higher likelihood of
allocation to lower CPU ids rather than higher ones).
While dispatching a batch of tasks to the same idle CPU is not
necessarily problematic, a fairer distribution among the list of idle
CPUs would be preferable.
Therefore change the CPU selection logic to distribute tasks equally
among the idle CPUs, still maintaining the preference for the previously
used one. Additionally, apply the CPU selection logic just before tasks
are dispatched, rather than assigning a CPU when tasks are drained from
the BPF queue. This adjustment is important, because tasks may linger in
the scheduler's internal structures for a bit and the idle state of the
CPUs in the system may change during that period.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
We want to activate the user-space scheduler only when there are pending
tasks that require scheduling actions.
To do so we keep track of the queued tasks via nr_queued, that is
incremented in .enqueue() when a task is sent to the user-space
scheduler and decremented in .dispatch() when a task is dispatched.
However, we may trigger an unbalance if the same pid is sent to the
scheduler multiple times (because the scheduler store all the tasks by
their unique pid).
When this happens nr_queued is never decremented back to 0, leading the
user-space scheduler to constantly spin, even if there's no activity to
do.
To prevent this from happening split nr_queued into nr_queued and
nr_scheduled. The former will be updated by the BPF component every time
that a task is sent to the scheduler and it's up to the user-space
scheduler to reset the counter when the queue is fully dreained. The
latter is maintained by the user-space scheduler and represents the
amount of tasks that are still processed by the scheduler and are
waiting to be dispatched.
The sum of nr_queued + nr_scheduled will be called nr_waiting and we can
rely on this metric to determine if the user-space scheduler has some
pending work to do or not.
This change makes rust_rustland more reliable and it strongly reduces
the CPU usage of the user-space scheduler by eliminating a lot of
unnecessary activations.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Considering the CPU where the user-space scheduler is running as busy
doesn't really provide any benefit, since the user-space scheduler is
constantly dispatching an amount of tasks equal to the amount of idle
CPUs and then yields (therefore its own CPU should be considered idle).
Considering the CPU where the user-space scheduler is running as busy
doesn't provide any benefit, as the scheduler consistently dispatches
tasks equal to the number of idle CPUs and then yields (therefore its
own CPU should be considered idle).
This also allows to reduce the overall user-space scheduler CPU
utilization, especially when the system is mostly idle, without
introducing any measurable performance regression.
Measuring the average CPU utilization of a (mostly) idle system over a
time period of 60 sec:
- wihout this patch: 5.41% avg cpu util
- with this patch: 2.26% avg cpu util
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Use virtme-ng to run the schedulers after they're built; virtme-ng
allows to pick an arbitrary sched-ext enabled kernel and run it
virtualizing the entire user-space root filesystem, so we can basically
exceute the recompiled schedulers inside such kernel.
This should allow to catch potential run-time issue in advance (both in
the kernel and the schedulers).
The sched-ext kernel is taken from the Ubuntu ppa (ppa:arighi/sched-ext)
at the moment, since it is the easiest / fastest way to get a
precompiled sched-ext kernel to run inside the Ubuntu 22.04 testing
environment.
The schedulers are tested using the new meson target "test_sched", the
specific actions are defined in meson-scripts/test_sched.
By default each test has a timeout of 30 sec, after the virtme-ng
completes the boot (that should be enough to initialize the scheduler
and run the scheduler for some seconds), while the total lifetime of the
virtme-ng guest is set to 60 sec, after this time the guest will be
killed (this allows to catch potential kernel crashes / hangs).
If a single scheduler fails the test, the entire "test_sched" action
will be interrupted and the overall test result will be considered a
failure.
At the moment scx_layered is excluded from the tests, because it
requires a special configuration (we should probably pre-generate a
default config in the workflow actions and change the scheduler to use
the default config if it's executed without any argument).
Moreover, scx_flatcg is also temporarily excluded from the tests,
because of these known issues:
- https://github.com/sched-ext/scx/issues/49
- https://github.com/sched-ext/sched_ext/pull/101
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Move the logic to activate the userspace scheduler to an update_idle()
callback, which is called when the CPU is about to go idle.
This disables the built-in idle tracking mechanism, so it allows to rely
completely on the internal CPU ownership logic (via get_cpu_owner() and
set_cpu_owner()) and it also allows to share the idle state with the
user-space scheduler via the BPF_MAP_TYPE_ARRAY cpu_map.
Moreover, when the user-space scheduler is activated, kick the idle cpu
to trigger immediate dispatch and avoid bubbles in the scheduling
pipeline.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
This reverts commit 9237e1d ("scx_rustland: always dispatch kthreads on
the local CPU").
Do not always prioritize all kthreads, we may have unbound workqueue
workers that can consume a lot of CPU cycles (e.g., encryption workers),
so we definitely want to apply the scheduling for those.
Therefore, restore the old behavior to prioritize only per-CPU kthreads.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Adding extra overhead to any kthread can potentially slow down the
entire system, so make sure this never happens by dispatching all
kthreads directly on the same local CPU (not just the per-CPU kthreads),
bypassing the user-space scheduler.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Trigger the user-space scheduler only upon a task's CPU release event
(avoiding its activation during each enqueue event) and only if there
are tasks waiting to be processed by the user-space scheduler.
This should save unnecessary calls to the user-space scheduler, reducing
the overall overhead of the scheduler.
Moreover, rename nr_enqueues to nr_queued and store the amount of tasks
currently queued to the user-space scheduler (that are waiting to be
dispatched).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Provide the following primitives to get and set CPU ownership in the BPF
part. This improves code readability and these primitives can be used by
the BPF part as a baseline to implement a better CPU idle tracking in
the future.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
BPF doesn't have full memory model yet, and while strict atomicity might
not be necessary in this context, it is advisable to enhance clarity in
the interlocking model.
To achieve this, provide the following primitives to operate on
usersched_needed:
static void set_usersched_needed(void)
static bool test_and_clear_usersched_needed(void)
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Dispatch tasks in a batch equal to the amount of idle CPUs in the
system.
This allows to reduce the pressure on the dispatcher queues, improving
the effectiveness of the scheduler (by having more tasks sitting in the
scheduler task pool) and mitigating potential priority inversion issues.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Provide an interface for the BPF dispatcher and user-space scheduler to
share CPU information. This information can empower the user-space
scheduler to make more informed decisions and enable the implementation
of a broader range of scheduling policies.
With this change the BPF dispatcher provides a CPU map (one entry per
CPU) that stores the pid that is running on each CPU (0 if the CPU is
idle). The CPU map is updated by the BPF dispatcher in the .running()
and .stopping() callbacks.
The dispatcher then sends to the user-space scheduler a suggestion of
the candidate CPU for each task that needs to run (that is always the
previously used CPU), along with all the task's information.
The user-space scheduler can decide to confirm the selected CPU or to
choose a different one, using all the shared CPU information.
Lastly, the selected CPU is communicated back to the dispatcher along
with all the task's information and the BPF dispatcher takes care of
executing the task on the selected CPU, eventually triggering a
migration.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
