Date: Fri, 13 May 2011 21:52:16 +0300 From: Andriy Gapon <avg@FreeBSD.org> To: Max Laier <max@love2party.net> Cc: freebsd-current@FreeBSD.org Subject: Re: proposed smp_rendezvous change Message-ID: <4DCD7DE0.7070400@FreeBSD.org> In-Reply-To: <201105131150.57548.max@love2party.net> References: <4DCD357D.6000109@FreeBSD.org> <201105131041.59981.max@love2party.net> <4DCD4E21.7020800@FreeBSD.org> <201105131150.57548.max@love2party.net>
next in thread | previous in thread | raw e-mail | index | archive | help
This is a multi-part message in MIME format.
--------------080904030706060507060701
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
on 13/05/2011 18:50 Max Laier said the following:
> On Friday 13 May 2011 11:28:33 Andriy Gapon wrote:
>> on 13/05/2011 17:41 Max Laier said the following:
>>> this ncpus isn't the one you are looking for.
>>
>> Thank you!
>>
>> Here's an updated patch:
>
> Can you attach the patch, so I can apply it locally. This code is really hard
> to read without context. Some more comments inline ...
Attached.
>>
>> Index: sys/kern/subr_smp.c
>> ===================================================================
>> --- sys/kern/subr_smp.c (revision 221835)
>> +++ sys/kern/subr_smp.c (working copy)
>> @@ -316,19 +316,14 @@
>> void (*local_action_func)(void*) = smp_rv_action_func;
>> void (*local_teardown_func)(void*) = smp_rv_teardown_func;
>>
>> - /* Ensure we have up-to-date values. */
>> - atomic_add_acq_int(&smp_rv_waiters[0], 1);
>> - while (smp_rv_waiters[0] < smp_rv_ncpus)
>> - cpu_spinwait();
>> -
>
> You really need this for architectures that need the memory barrier to ensure
> consistency. We also need to move the reads of smp_rv_* below this point to
> provide a consistent view.
I thought that this would be automatically handled by the fact that a master CPU
sets smp_rv_waiters[0] using atomic operation with release semantics.
But I am not very proficient in this matters...
But I fail to see why we need to require that all CPUs should gather at this
point/condition.
That is, my point is that we don't start a new rendezvous until a previous one
is completely finished. Then we set up the new rendezvous, finish the setup
with an operation with release semantics and only then notify the target CPUs.
I can't see how the slave CPUs would see stale values in the rendezvous
pseudo-object, but, OTOH, I am not very familiar with architectures that have
weaker memory consistency rules as compared to x86.
--
Andriy Gapon
--------------080904030706060507060701
Content-Type: text/plain;
name="smp_rv.diff"
Content-Transfer-Encoding: 7bit
Content-Disposition: attachment;
filename="smp_rv.diff"
Index: sys/kern/subr_smp.c
===================================================================
--- sys/kern/subr_smp.c (revision 221835)
+++ sys/kern/subr_smp.c (working copy)
@@ -316,19 +316,14 @@
void (*local_action_func)(void*) = smp_rv_action_func;
void (*local_teardown_func)(void*) = smp_rv_teardown_func;
- /* Ensure we have up-to-date values. */
- atomic_add_acq_int(&smp_rv_waiters[0], 1);
- while (smp_rv_waiters[0] < smp_rv_ncpus)
- cpu_spinwait();
-
/* setup function */
if (local_setup_func != smp_no_rendevous_barrier) {
if (smp_rv_setup_func != NULL)
smp_rv_setup_func(smp_rv_func_arg);
/* spin on entry rendezvous */
- atomic_add_int(&smp_rv_waiters[1], 1);
- while (smp_rv_waiters[1] < smp_rv_ncpus)
+ atomic_add_int(&smp_rv_waiters[0], 1);
+ while (smp_rv_waiters[0] < smp_rv_ncpus)
cpu_spinwait();
}
@@ -337,12 +332,16 @@
local_action_func(local_func_arg);
/* spin on exit rendezvous */
- atomic_add_int(&smp_rv_waiters[2], 1);
- if (local_teardown_func == smp_no_rendevous_barrier)
+ atomic_add_int(&smp_rv_waiters[1], 1);
+ if (local_teardown_func == smp_no_rendevous_barrier) {
+ atomic_add_int(&smp_rv_waiters[2], 1);
return;
- while (smp_rv_waiters[2] < smp_rv_ncpus)
+ }
+ while (smp_rv_waiters[1] < smp_rv_ncpus)
cpu_spinwait();
+ atomic_add_int(&smp_rv_waiters[2], 1);
+
/* teardown function */
if (local_teardown_func != NULL)
local_teardown_func(local_func_arg);
@@ -377,6 +376,10 @@
/* obtain rendezvous lock */
mtx_lock_spin(&smp_ipi_mtx);
+ /* Wait for any previous unwaited rendezvous to finish. */
+ while (atomic_load_acq_int(&smp_rv_waiters[2]) < smp_rv_ncpus)
+ cpu_spinwait();
+
/* set static function pointers */
smp_rv_ncpus = ncpus;
smp_rv_setup_func = setup_func;
@@ -395,7 +398,7 @@
smp_rendezvous_action();
if (teardown_func == smp_no_rendevous_barrier)
- while (atomic_load_acq_int(&smp_rv_waiters[2]) < ncpus)
+ while (atomic_load_acq_int(&smp_rv_waiters[1]) < ncpus)
cpu_spinwait();
/* release lock */
--------------080904030706060507060701--
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4DCD7DE0.7070400>