lkml.org 
[lkml]   [2020]   [Dec]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [GIT pull] locking/urgent for v5.10-rc6
On Tue, Dec 01, 2020 at 08:14:41PM +0100, Peter Zijlstra wrote:
> On Tue, Dec 01, 2020 at 06:57:37PM +0000, Mark Rutland wrote:
> > On Tue, Dec 01, 2020 at 07:15:06PM +0100, Peter Zijlstra wrote:
> > > On Tue, Dec 01, 2020 at 03:55:19PM +0100, Peter Zijlstra wrote:
> > > > On Tue, Dec 01, 2020 at 06:46:44AM -0800, Paul E. McKenney wrote:
> > > >
> > > > > > So after having talked to Sven a bit, the thing that is happening, is
> > > > > > that this is the one place where we take interrupts with RCU being
> > > > > > disabled. Normally RCU is watching and all is well, except during idle.
> > > > >
> > > > > Isn't interrupt entry supposed to invoke rcu_irq_enter() at some point?
> > > > > Or did this fall victim to recent optimizations?
> > > >
> > > > It does, but the problem is that s390 is still using
> > >
> > > I might've been too quick there, I can't actually seem to find where
> > > s390 does rcu_irq_enter()/exit().
> > >
> > > Also, I'm thinking the below might just about solve the current problem.
> > > The next problem would then be it calling TRACE_IRQS_ON after it did
> > > rcu_irq_exit()... :/
> >
> > I gave this patch a go under QEMU TCG atop v5.10-rc6 s390 defconfig with
> > PROVE_LOCKING and DEBUG_ATOMIC_SLEEP. It significantly reduces the
> > number of lockdep splats, but IIUC we need to handle the io_int_handler
> > path in addition to the ext_int_handler path, and there's a remaining
> > lockdep splat (below).
>
> I'm amazed it didn't actually make things worse, given how I failed to
> spot do_IRQ() was arch code etc..
>
> > If this ends up looking like we'll need more point-fixes, I wonder if we
> > should conditionalise the new behaviour of the core idle code under a
> > new CONFIG symbol for now, and opt-in x86 and arm64, then transition the
> > rest once they've had a chance to test. They'll still be broken in the
> > mean time, but no more so than they previously were.
>
> We can do that I suppose... :/

Well, the following small patch works for me (plus an additional call to
trace_hardirqs_on() in our udelay implementation - but that's probably
independent).
Is there a reason why this should be considered broken?

diff --git a/arch/s390/kernel/entry.S b/arch/s390/kernel/entry.S
index 26bb0603c5a1..92beb1444644 100644
--- a/arch/s390/kernel/entry.S
+++ b/arch/s390/kernel/entry.S
@@ -763,12 +763,7 @@ ENTRY(io_int_handler)
xc __PT_FLAGS(8,%r11),__PT_FLAGS(%r11)
TSTMSK __LC_CPU_FLAGS,_CIF_IGNORE_IRQ
jo .Lio_restore
-#if IS_ENABLED(CONFIG_TRACE_IRQFLAGS)
- tmhh %r8,0x300
- jz 1f
TRACE_IRQS_OFF
-1:
-#endif
xc __SF_BACKCHAIN(8,%r15),__SF_BACKCHAIN(%r15)
.Lio_loop:
lgr %r2,%r11 # pass pointer to pt_regs
@@ -791,12 +786,7 @@ ENTRY(io_int_handler)
TSTMSK __LC_CPU_FLAGS,_CIF_WORK
jnz .Lio_work
.Lio_restore:
-#if IS_ENABLED(CONFIG_TRACE_IRQFLAGS)
- tm __PT_PSW(%r11),3
- jno 0f
TRACE_IRQS_ON
-0:
-#endif
mvc __LC_RETURN_PSW(16),__PT_PSW(%r11)
tm __PT_PSW+1(%r11),0x01 # returning to user ?
jno .Lio_exit_kernel
@@ -976,12 +966,7 @@ ENTRY(ext_int_handler)
xc __PT_FLAGS(8,%r11),__PT_FLAGS(%r11)
TSTMSK __LC_CPU_FLAGS,_CIF_IGNORE_IRQ
jo .Lio_restore
-#if IS_ENABLED(CONFIG_TRACE_IRQFLAGS)
- tmhh %r8,0x300
- jz 1f
TRACE_IRQS_OFF
-1:
-#endif
xc __SF_BACKCHAIN(8,%r15),__SF_BACKCHAIN(%r15)
lgr %r2,%r11 # pass pointer to pt_regs
lghi %r3,EXT_INTERRUPT
diff --git a/arch/s390/kernel/idle.c b/arch/s390/kernel/idle.c
index 2b85096964f8..5bd8c1044d09 100644
--- a/arch/s390/kernel/idle.c
+++ b/arch/s390/kernel/idle.c
@@ -123,7 +123,6 @@ void arch_cpu_idle_enter(void)
void arch_cpu_idle(void)
{
enabled_wait();
- raw_local_irq_enable();
}

void arch_cpu_idle_exit(void)
\
 
 \ /
  Last update: 2020-12-01 20:21    [W:0.130 / U:2.260 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site