On Tue, 16 Feb 2021 09:30:03 -0800
"Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:
> On Mon, Feb 15, 2021 at 12:28:26PM +0100, Sebastian Andrzej Siewior wrote:
> > On 2021-02-13 08:45:54 [-0800], Paul E. McKenney wrote:
> > > Glad you like it! But let's see which (if any) of these patches solves
> > > the problem for Sebastian.
> >
> > Looking at that, is there any reason for doing this that can not be
> > solved by moving the self-test a little later? Maybe once we reached at
> > least SYSTEM_SCHEDULING?
>
> One problem is that ksoftirqd and the kprobes use are early_initcall(),
> so we cannot count on ksoftirqd being spawned when kprobes first uses
> synchronize_rcu_tasks(). Moving the selftest later won't fix this
> problem, but rather just paper it over.
>
> > This happens now even before lockdep is up or the console is registered.
> > So if something bad happens, you end up with a blank terminal.
>
> I was getting a splat, but I could easily believe that there are
> configurations where the hang is totally silent. In other words, I do
> agree that this needs a proper fix. All we need do is work out an
> agreeable value of "proper". ;-)
>
> > There is nothing else that early in the boot process that requires
> > working softirq. The only exception to this is wait_task_inactive()
> > which is used while starting a new thread (including the ksoftirqd)
> > which is why it was moved to schedule_hrtimeout().
>
> Moving kprobes initialization to early_initcall() [1] means that there
> can be a call to synchronize_rcu_tasks() before the current spawning of
> ksoftirqd. Because synchronize_rcu_tasks() needs timers to work, it needs
> softirq to work. I know two straightforward ways to make that happen:
>
> 1. Spawn ksoftirqd earlier.
>
> 2. Suppress attempts to awaken ksoftirqd before it exists,
> forcing all ksoftirq execution on the back of interrupts.
>
> Uladzislau and I each produced patches for #1, and I produced a patch
> for #2.
>
> The only other option I know of is to push the call to init_kprobes()
> later in the boot sequence, perhaps to its original subsys_initcall(),
> or maybe only as late as core_initcall(). I added Masami and Steve on
> CC for their thoughts on this.
>
> Is there some other proper fix that I am missing?
Oh, I missed that the synchronize_rcu_tasks() will be involved the kprobes
in early stage. Does the problem only exist in the synchronize_rcu_tasks()
instead of synchronize_rcu()? If so I can just stop optimizer in early stage
because I just want to enable kprobes in early stage, but not optprobes.
Does the following patch help?
>From e5fafcda3ff918cd52619f795a3f22fb95c72b11 Mon Sep 17 00:00:00 2001
From: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
Date: Wed, 17 Feb 2021 23:35:20 +0900
Subject: [PATCH] kprobes: Fix to delay the kprobes jump optimization
Since the kprobes jump optimization involves synchronize_rcu_tasks()
which depends on the ksoftirqd, that can not be enabled at the
early_initcall() boot stage. So this makes the kprobe optimization
disabled in the early_initcall() and enables it in subsys_initcall().
Note that non-optimized kprobes is still available after
early_initcall(). Only jump optimization is delayed.
Fixes: 36dadef23fcc ("kprobes: Init kprobes in early_initcall")
Signed-off-by: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
---
kernel/kprobes.c | 31 +++++++++++++++++++++----------
1 file changed, 21 insertions(+), 10 deletions(-)
diff --git a/kernel/kprobes.c b/kernel/kprobes.c
index d5a3eb74a657..779d8322e307 100644
--- a/kernel/kprobes.c
+++ b/kernel/kprobes.c
@@ -861,7 +861,6 @@ static void try_to_optimize_kprobe(struct kprobe *p)
cpus_read_unlock();
}
-#ifdef CONFIG_SYSCTL
static void optimize_all_kprobes(void)
{
struct hlist_head *head;
@@ -887,6 +886,7 @@ static void optimize_all_kprobes(void)
mutex_unlock(&kprobe_mutex);
}
+#ifdef CONFIG_SYSCTL
static void unoptimize_all_kprobes(void)
{
struct hlist_head *head;
@@ -2497,18 +2497,14 @@ static int __init init_kprobes(void)
}
}
-#if defined(CONFIG_OPTPROBES)
-#if defined(__ARCH_WANT_KPROBES_INSN_SLOT)
- /* Init kprobe_optinsn_slots */
- kprobe_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
-#endif
- /* By default, kprobes can be optimized */
- kprobes_allow_optimization = true;
-#endif
-
/* By default, kprobes are armed */
kprobes_all_disarmed = false;
+#if defined(CONFIG_OPTPROBES) && defined(__ARCH_WANT_KPROBES_INSN_SLOT)
+ /* Init kprobe_optinsn_slots for allocation */
+ kprobe_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
+#endif
+
err = arch_init_kprobes();
if (!err)
err = register_die_notifier(&kprobe_exceptions_nb);
@@ -2523,6 +2519,21 @@ static int __init init_kprobes(void)
}
early_initcall(init_kprobes);
+#if defined(CONFIG_OPTPROBES)
+static int __init init_optprobes(void)
+{
+ /*
+ * Enable kprobe optimization - this kicks the optimizer which
+ * depends on synchronize_rcu_tasks() and ksoftirqd, that is
+ * not spawned in early initcall. So delay the optimization.
+ */
+ optimize_all_kprobes();
+
+ return 0;
+}
+subsys_initcall(init_optprobes);
+#endif
+
#ifdef CONFIG_DEBUG_FS
static void report_probe(struct seq_file *pi, struct kprobe *p,
const char *sym, int offset, char *modname, struct kprobe *pp)
--
2.25.1
--
Masami Hiramatsu <mhiramat@xxxxxxxxxx>
