Re: smp: Start up non-boot CPUs asynchronously

On 02/14/2012 03:18 PM, Srivatsa S. Bhat wrote:

On 02/14/2012 01:47 PM, Srivatsa S. Bhat wrote:

quoted

On 01/31/2012 09:54 PM, Arjan van de Ven wrote:

quoted

From ee65be59057c920042747d46dc174c5a5a56c744 Mon Sep 17 00:00:00 2001
From: Arjan van de Ven <redacted>
Date: Mon, 30 Jan 2012 20:44:51 -0800
Subject: [PATCH] smp: Start up non-boot CPUs asynchronously

The starting of the "not first" CPUs actually takes a lot of boot time
of the kernel... upto "minutes" on some of the bigger SGI boxes.
Right now, this is a fully sequential operation with the rest of the kernel
boot.

This patch turns this bringup of the other cpus into an asynchronous operation.
With some other changes (not in this patch) this can save significant kernel
boot time (upto 40% on my laptop!!).
Basically now CPUs could get brought up in parallel to disk enumeration, graphic
mode bringup etc etc etc.

Note that the implementation in this patch still waits for all CPUs to
be brought up before starting userspace; I would love to remove that
restriction over time (technically that is simple), but that becomes
then a change in behavior... I'd like to see more discussion on that
being a good idea before I write that patch.

Second note on version 2 of the patch:
This patch does currently not save any boot time, due to a situation
where the cpu hotplug lock gets taken for write by the cpu bringup code,
which starves out readers of this lock throughout the kernel.
Ingo specifically requested this behavior to expose this lock problem.

CC: Milton Miller <redacted>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Ingo Molnar <redacted>

Signed-off-by: Arjan van de Ven <redacted>
---
 kernel/smp.c |   21 ++++++++++++++++++++-
 1 files changed, 20 insertions(+), 1 deletions(-)

diff --git a/kernel/smp.c b/kernel/smp.c
index db197d6..ea48418 100644
--- a/kernel/smp.c
+++ b/kernel/smp.c

@@ -12,6 +12,8 @@
 #include <linux/gfp.h>
 #include <linux/smp.h>
 #include <linux/cpu.h>
+#include <linux/async.h>
+#include <linux/delay.h>

 #ifdef CONFIG_USE_GENERIC_SMP_HELPERS
 static struct {

@@ -664,17 +666,34 @@ void __init setup_nr_cpu_ids(void)
 	nr_cpu_ids = find_last_bit(cpumask_bits(cpu_possible_mask),NR_CPUS) + 1;
 }

+void __init async_cpu_up(void *data, async_cookie_t cookie)
+{
+	unsigned long nr = (unsigned long) data;
+	/*
+	 * we can only up one cpu at a time, as enforced by the hotplug
+	 * lock; it's better to wait for all earlier CPUs to be done before
+	 * we bring up ours, so that the bring up order is predictable.
+	 */
+	async_synchronize_cookie(cookie);
+	cpu_up(nr);
+}
+
 /* Called by boot processor to activate the rest. */
 void __init smp_init(void)
 {
 	unsigned int cpu;

 	/* FIXME: This should be done in userspace --RR */
+
+	/*
+	 * But until we do this in userspace, we're going to do this
+	 * in parallel to the rest of the kernel boot up.-- Arjan
+	 */
 	for_each_present_cpu(cpu) {
 		if (num_online_cpus() >= setup_max_cpus)
 			break;
 		if (!cpu_online(cpu))
-			cpu_up(cpu);
+			async_schedule(async_cpu_up, (void *) cpu);
 	}

 	/* Any cleanup work */

If I understand correctly, with this patch, the booting of non-boot CPUs
will happen in parallel with the rest of the kernel boot, but bringing up
of individual CPU is still serialized (due to hotplug lock).

If that is correct, I see several issues with this patch:

1. In smp_init(), after the comment "Any cleanup work" (see above), we print:
printk(KERN_INFO "Brought up %ld CPUs\n", (long)num_online_cpus());
So this can potentially print less than expected number of CPUs and might
surprise users.

2. Just below that we have smp_cpus_done(setup_max_cpus); and this translates
to native_smp_cpus_done() under x86, which calls impress_friends().
And that means, the bogosum calculation and the total activated processor
count which is printed, may get messed up.

3. sched_init_smp() is called immediately after smp_init(). And that calls
init_sched_domains(cpu_active_mask). Of course, it registers a hotplug
notifier callback to handle hot-added cpus.. but with this patch, boot up can
actually become unnecessarily slow at this point - what could have been done
in one go with an appropriately filled up cpu_active_mask, needs to be done
again and again using notifier callbacks. IOW, building sched domains can
potentially become a bottleneck, especially if there are lots and lots of
cpus in the machine.

4. There is an unhandled race condition (tiny window) in sched_init_smp():

get_online_cpus();
...
init_sched_domains(cpu_active_mask);
...
put_online_cpus();
                                           <<<<<<<<<<<<<<<<<<<<<<<< There!

hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);

At the point shown above, some non-boot cpus can get booted up, without
being noticed by the scheduler.

5. And in powerpc, it creates a new race condition, as explained in
https://lkml.org/lkml/2012/2/13/383
(Of course, we can fix it trivially by using get/put_online_cpus().)

Actually, this one is trickier than that, to get it perfectly right.
[see point 8 below].

6. I also observed that in powerpc, a distinction is made implicitly between
a cpu booting for the first time vs a soft CPU online event. That is, for
freshly booted cpus, the following 3 functions are called:
(Refer arch/powerpc/kernel/sysfs.c: topology_init)

	register_cpu(c, cpu);
	device_create_file(&c->dev, &dev_attr_physical_id);
	register_cpu_online(cpu);

However, for a soft CPU Online event, only the last function is called.
(And that looks correct because it matches properly with what is done
upon CPU offline - only unregister_cpu_online() is called).

IOW, with this patch it becomes necessary to carefully examine all code
with such implicit assumptions and modify them to handle the async boot up
properly.

7. And whichever code between smp_init() and async_synchronize_full() didn't
care about CPU hotplug till today but depended on all cpus being online must
suddenly start worrying about CPU Hotplug. They must register a cpu notifier
and handle callbacks etc etc.. Or if they are not worth that complexity, they
should atleast be redesigned or moved around - like the print statements that
tell how many cpus came up, for example.

8. And we should provide a way in which a piece of code can easily "catch" all
CPU_ONLINE/UP_PREPARE events without missing any of them due to race
conditions. Of course register_cpu_notifier() and friends are provided for
that purpose, but they can't be used as it is in this boot up code..
And calling register_cpu_notifier() within get/put_online_cpus() would be a
disaster since that could lead to ABBA deadlock between cpu_add_remove_lock
and cpu_hotplug.lock

9. With this patch, the second statement below in Documentation/cpu-hotplug.txt
won't be true anymore:

Init functions could be of two types:
1. early init (init function called when only the boot processor is online).
2. late init (init function called _after_ all the CPUs are online).

And hence, those parts of the code which depend on this will have to be revisited.

10. Further down, in Documentation/cpu-hotplug.txt, we see:
(referring to early init as first case and late init as second case)

"For the first case, you should add the following to your init function

        register_cpu_notifier(&foobar_cpu_notifier);

For the second case, you should add the following to your init function

        register_hotcpu_notifier(&foobar_cpu_notifier); "

And as of now, hotcpu notifiers are nothing but regular cpu notifiers.
I wonder why do we even have something called hotcpu notifiers, when they do
nothing different.. rather, the distinction between "hotcpu add" vs "just
normal booting" was implicitly handled by choosing when we register our
callbacks:
register at early init => "normal booting" + "hotcpu, including soft online"
register at late init => "hotcpu, including soft online"

So, earlier we had some control over which CPU hotplug events we wanted to be
notified of, by choosing when we register the notifiers. But with this patch,
"careful placement" of our callback registration doesn't make any difference
anymore because late initcalls could run in parallel with smp boot up...

The point I am making is, what was already bad with respect to callback
registration, is made even worse by this patch.
(Btw, this issue is in the light of point 6 above).

quoted

There could be many more things that this patch breaks.. I haven't checked
thoroughly.

 
Regards,
Srivatsa S. Bhat