[PATCH] x86: make delay work in earlier stages

[PATCH] x86: make delay work in earlier stages

[Jiri Slaby]

When a panic happens during bootup, "Rebooting in X seconds.." is
shown, but reboot happens immediatelly. It is because panic uses mdelay
and mdelay calls internally __const_udelay which is broken while booting.

The per_cpu cpu_info.loops_per_jiffy is not initialized yet, so
__const_udelay actually multiplies the number of loops by zero. This
results in __const_udelay to delay the execution only by a nanosecond or
so.

So check whether cpu_info.loops_per_jiffy is zero and use
loops_per_jiffy in that case. mdelay will not be so precise, but it
works relatively good. From the original (mdelay in rest_init):
[    0.170039] delaying 100ms
[    0.170828] done
I get:
[    0.214042] delaying 100ms
[    0.313974] done

I do not think the added check matters given we are about to spin the
processor in the next few hundred cycles.

Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: <x86@kernel.org>
---
 arch/x86/lib/delay.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/arch/x86/lib/delay.c b/arch/x86/lib/delay.c
index 073d1f1a620b..aca0988f0e14 100644
--- a/arch/x86/lib/delay.c
+++ b/arch/x86/lib/delay.c
@@ -156,13 +156,14 @@ EXPORT_SYMBOL(__delay);
 
 inline void __const_udelay(unsigned long xloops)
 {
+	unsigned long lpf = this_cpu_read(cpu_info.loops_per_jiffy) ? :
+		loops_per_jiffy;
 	int d0;
 
 	xloops *= 4;
 	asm("mull %%edx"
 		:"=d" (xloops), "=&a" (d0)
-		:"1" (xloops), "0"
-		(this_cpu_read(cpu_info.loops_per_jiffy) * (HZ/4)));
+		:"1" (xloops), "0" (lpf * (HZ / 4)));
 
 	__delay(++xloops);
 }
-- 
2.11.0

Re: [PATCH] x86: make delay work in earlier stages

[Andy Shevchenko]

> +       unsigned long lpf = this_cpu_read(cpu_info.loops_per_jiffy) ? :
> +               loops_per_jiffy;

Perhaps lpj ?

In any case FWIW:
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>

-- 
With Best Regards,
Andy Shevchenko

[tip:timers/core] x86/timer: Make delay() during early bootup

[tip-bot for Jiri Slaby]

Commit-ID:  bf3304d996fbb993bad6be09cafde39cc2db72bb
Gitweb:     http://git.kernel.org/tip/bf3304d996fbb993bad6be09cafde39cc2db72bb
Author:     Jiri Slaby <jslaby@suse.cz>
AuthorDate: Thu, 19 Jan 2017 12:47:30 +0100
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Fri, 20 Jan 2017 09:45:22 +0100

x86/timer: Make delay() during early bootup

When a panic happens during bootup, "Rebooting in X seconds.." is
shown, but reboot happens immediatelly. It is because panic() uses mdelay()
and mdelay() calls __const_udelay() immediately, which is does not
work while booting.

The per_cpu cpu_info.loops_per_jiffy is not initialized yet, so
__const_udelay() actually multiplies the number of loops by zero. This
results in __const_udelay() to delay the execution only by a nanosecond
or so.

So check whether cpu_info.loops_per_jiffy is zero and use
loops_per_jiffy in that case. mdelay() will not be so precise without
proper calibration, but it works relatively well.

Before:

  [    0.170039] delaying 100ms
  [    0.170828] done

After

  [    0.214042] delaying 100ms
  [    0.313974] done

I do not think the added check matters given we are about to spin the
processor in the next few hundred cycles.

Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170119114730.2670-1-jslaby@suse.cz
[ Minor edits. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 arch/x86/lib/delay.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/lib/delay.c b/arch/x86/lib/delay.c
index 073d1f1..a8e91ae 100644
--- a/arch/x86/lib/delay.c
+++ b/arch/x86/lib/delay.c
@@ -156,13 +156,13 @@ EXPORT_SYMBOL(__delay);
 
 inline void __const_udelay(unsigned long xloops)
 {
+	unsigned long lpj = this_cpu_read(cpu_info.loops_per_jiffy) ? : loops_per_jiffy;
 	int d0;
 
 	xloops *= 4;
 	asm("mull %%edx"
 		:"=d" (xloops), "=&a" (d0)
-		:"1" (xloops), "0"
-		(this_cpu_read(cpu_info.loops_per_jiffy) * (HZ/4)));
+		:"1" (xloops), "0" (lpj * (HZ / 4)));
 
 	__delay(++xloops);
 }

Re: [tip:timers/core] x86/timer: Make delay() during early bootup

[Jiri Slaby]

On 01/20/2017, 03:40 PM, tip-bot for Jiri Slaby wrote:
> Commit-ID:  bf3304d996fbb993bad6be09cafde39cc2db72bb
> Gitweb:     http://git.kernel.org/tip/bf3304d996fbb993bad6be09cafde39cc2db72bb
> Author:     Jiri Slaby <jslaby@suse.cz>
> AuthorDate: Thu, 19 Jan 2017 12:47:30 +0100
> Committer:  Ingo Molnar <mingo@kernel.org>
> CommitDate: Fri, 20 Jan 2017 09:45:22 +0100
> 
> x86/timer: Make delay() during early bootup
> 
> When a panic happens during bootup, "Rebooting in X seconds.." is
> shown, but reboot happens immediatelly. It is because panic() uses mdelay()
> and mdelay() calls __const_udelay() immediately, which is does not
> work while booting.

The 'is' is superfluous after your edit now. But I guess, you cannot
change it given it is in the tree.

> --- a/arch/x86/lib/delay.c
> +++ b/arch/x86/lib/delay.c
> @@ -156,13 +156,13 @@ EXPORT_SYMBOL(__delay);
>  
>  inline void __const_udelay(unsigned long xloops)
>  {
> +	unsigned long lpj = this_cpu_read(cpu_info.loops_per_jiffy) ? : loops_per_jiffy;

And thanks for fixing the 'lpf' brain damage :).

-- 
js
suse labs

[tip:timers/core] x86/timer: Make delay() work during early bootup

[tip-bot for Jiri Slaby]

Commit-ID:  4c45c5167c9563b1a2eee3e2fe954621355e4ca8
Gitweb:     http://git.kernel.org/tip/4c45c5167c9563b1a2eee3e2fe954621355e4ca8
Author:     Jiri Slaby <jslaby@suse.cz>
AuthorDate: Thu, 19 Jan 2017 12:47:30 +0100
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Sun, 22 Jan 2017 10:03:12 +0100

x86/timer: Make delay() work during early bootup

When a panic happens during bootup, "Rebooting in X seconds.." is
shown, but reboot happens immediatelly. It is because panic() uses mdelay()
and mdelay() calls __const_udelay() immediately, which does not
work while booting.

The per_cpu cpu_info.loops_per_jiffy value is not initialized yet, so
__const_udelay() actually multiplies the number of loops by zero. This
results in __const_udelay() to delay the execution only by a nanosecond
or so.

So check whether cpu_info.loops_per_jiffy is zero and use
loops_per_jiffy in that case. mdelay() will not be so precise without
proper calibration, but it works relatively well.

Before:

  [    0.170039] delaying 100ms
  [    0.170828] done

After

  [    0.214042] delaying 100ms
  [    0.313974] done

I do not think the added check matters given we are about to spin the
processor in the next few hundred cycles.

Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170119114730.2670-1-jslaby@suse.cz
[ Minor edits. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 arch/x86/lib/delay.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/x86/lib/delay.c b/arch/x86/lib/delay.c
index 073d1f1..a8e91ae 100644
--- a/arch/x86/lib/delay.c
+++ b/arch/x86/lib/delay.c
@@ -156,13 +156,13 @@ EXPORT_SYMBOL(__delay);
 
 inline void __const_udelay(unsigned long xloops)
 {
+	unsigned long lpj = this_cpu_read(cpu_info.loops_per_jiffy) ? : loops_per_jiffy;
 	int d0;
 
 	xloops *= 4;
 	asm("mull %%edx"
 		:"=d" (xloops), "=&a" (d0)
-		:"1" (xloops), "0"
-		(this_cpu_read(cpu_info.loops_per_jiffy) * (HZ/4)));
+		:"1" (xloops), "0" (lpj * (HZ / 4)));
 
 	__delay(++xloops);
 }