Re: [PATCH v5 02/13] rcu-tasks: Add a Tasks RCU implementation for reader-marked trampolines

[Frederic Weisbecker <frederic@kernel.org>]

Le Tue, Sep 22, 2026 at 02:23:21AM +0000, Josef Bacik a écrit :
> Tasks RCU waits for every task to pass through a voluntary context
> switch, usermode or idle, because a preempted task might be sitting in a
> trampoline that is about to be freed and nothing marks it as such.  With
> PREEMPT_LAZY that is a poor fit for servers: cond_resched() is a no-op,
> so a CPU-bound kthread only ever leaves the CPU by preemption, and one
> such kthread holds every synchronize_rcu_tasks() caller -- ftrace and
> BPF trampoline teardown under their mutexes, the kprobe jump optimizer
> under text_mutex and cpus_read_lock() -- hostage for as long as it runs.
> 
> Following the discussion on v2, take the other road: let the
> architecture make its trampolines Tasks Trace RCU readers.  When an
> architecture selects HAVE_RCU_TRAMPOLINE_READERS it promises that every
> trampoline whose lifetime Tasks RCU guards enters rcu_read_lock_trace()
> (or its assembly equivalent) before calling out and leaves it before
> returning, so a task anywhere inside such a call-out, preempted or not,
> is an ordinary Tasks Trace reader.
> 
> That leaves the few instructions of trampoline text before the reader is
> entered and after it is left (plus, in a later patch, the bytes a kprobe
> jump optimization is about to overwrite). A task can only linger there
> by being interrupted there, and such text never calls anything that
> schedules, so instead of tracking tasks we track CPUs: every pass
> through __schedule() is a per-CPU quiescent event, except that the one
> context switch that can catch a task at an arbitrary instruction -- a
> preemption from irq exit -- first records the interrupted IP in the task
> and parks it on a per-CPU list for the duration (reusing the fields and
> lists the classic flavor keeps for its exit-path bookkeeping), and, if
> the IP is inside such "unmarked" text, puts the task on a short holdout
> list; the task takes itself off at its next context switch outside such
> a preemption or irq-exit check that finds it elsewhere. Usermode (the
> existing tick hook, or a nohz_full CPU in an RCU extended quiescent
> state) and idle count as well. rcu_tasks_trampoline_text() does the
> classification: anything outside core and module text, plus an arch hook
> for things like static ftrace stubs and return thunks.
> 
> The grace period, run by the existing rcu_tasks kthread so that
> call_rcu_tasks(), synchronize_rcu_tasks() and rcu_barrier_tasks() keep
> their names and callers, is: wait for every online CPU to context switch
> or be seen in an RCU extended quiescent state (nudging stragglers with
> resched_cpu() after a jiffy), drain the holdout list as it stood,
> synchronize_rcu_tasks_trace() for everything inside the readers, then
> one more CPU pass and drain for tasks that have since left the reader
> into the trailing instructions. That is bounded by a few jiffies,
> preempt-off latency and an SRCU grace period rather than by the longest
> stretch any task runs without sleeping, needs no per-task scan, and
> makes cond_resched_tasks_rcu_qs() unnecessary on such architectures.
> Unlike the classic flavor it also waits for an idle task caught in a
> trampoline, since an idle CPU only counts while RCU is not watching it.
> rcu_tasks_wait_irq_preempted() walks the parked lists for the one caller
> (the kprobe jump optimizer, later in the series) that makes ordinary
> text unsafe to be parked in and so has to wait out tasks that were
> preempted there before it said so.
> 
> The classic implementation is untouched and remains the default; the
> new one is built only as CONFIG_TASKS_RCU_TRAMPOLINE_READERS when the
> architecture opts in and uses the generic irq entry code, whose
> reschedule check gains the rcu_tasks_irq_resched() call.  Nothing
> selects it yet.
> 
> Suggested-by: Paul E. McKenney <paulmck@kernel.org>
> Suggested-by: Alexei Starovoitov <ast@kernel.org>
> Assisted-by: LLM
> Signed-off-by: Josef Bacik <josef@toxicpanda.com>

One review might have fell into the cracks: https://lore.kernel.org/lkml/aqxLgT41UyA-bV5J@pavilion.home/

-- 
Frederic Weisbecker
SUSE Labs