Re: [PATCH v15 19/26] x86/tsc: calibrate tsc only once

Re: [PATCH v15 19/26] x86/tsc: calibrate tsc only once

[Chuan Hua, Lei]

> static unsigned long __init get_loops_per_jiffy(void)
> {
> 	unsigned long lpj = tsc_khz * KHZ;
> 
> 	do_div(lpj, HZ);
> 	return lpj;
> }
Just tried this with 4.19-rc2 on x86(32bit). lpj return as zero which is not expected
After disassembling the code,
	0xc1239a9e <+199>:   imul   $0x3e8,0xc12296e4,%edx
	0xc1239aa8 <+209>:   xor    %ecx,%ecx
	0xc1239aaa <+211>:   test   %edx,%edx
	0xc1239aac <+213>:   mov    %eax,%ebx
	0xc1239aae <+215>:   je     0xc1239abd <tsc_init+230>
	0xc1239ab0 <+217>:   mov    $0x64,%ecx
	0xc1239ab5 <+222>:   mov    %edx,%eax
	0xc1239ab7 <+224>:   xor    %edx,%edx
	0xc1239ab9 <+226>:   div    %ecx
	0xc1239abb <+228>:   mov    %eax,%ecx
	0xc1239abd <+230>:   mov    %ebx,%eax
	0xc1239abf <+232>:   mov    $0x64,%ebx
	0xc1239ac4 <+237>:   div    %ebx
	0xc1239ac6 <+239>:   mov    %ecx,%edx
imul will load the result into %edx, %edx supposed to be high 32 bit which is not zero,
It should be zero in this case. both lpj and tsc_khz should be u64 to work properly.

Re: [PATCH v15 19/26] x86/tsc: calibrate tsc only once

[Thomas Gleixner]

On Wed, 5 Sep 2018, Chuan Hua, Lei wrote:

> > static unsigned long __init get_loops_per_jiffy(void)
> > {
> > 	unsigned long lpj = tsc_khz * KHZ;
> > 
> > 	do_div(lpj, HZ);
> > 	return lpj;
> > }
> Just tried this with 4.19-rc2 on x86(32bit). lpj return as zero which is not
> expected
> After disassembling the code,
> 	0xc1239a9e <+199>:   imul   $0x3e8,0xc12296e4,%edx
> 	0xc1239aa8 <+209>:   xor    %ecx,%ecx
> 	0xc1239aaa <+211>:   test   %edx,%edx
> 	0xc1239aac <+213>:   mov    %eax,%ebx
> 	0xc1239aae <+215>:   je     0xc1239abd <tsc_init+230>
> 	0xc1239ab0 <+217>:   mov    $0x64,%ecx
> 	0xc1239ab5 <+222>:   mov    %edx,%eax
> 	0xc1239ab7 <+224>:   xor    %edx,%edx
> 	0xc1239ab9 <+226>:   div    %ecx
> 	0xc1239abb <+228>:   mov    %eax,%ecx
> 	0xc1239abd <+230>:   mov    %ebx,%eax
> 	0xc1239abf <+232>:   mov    $0x64,%ebx
> 	0xc1239ac4 <+237>:   div    %ebx
> 	0xc1239ac6 <+239>:   mov    %ecx,%edx
> imul will load the result into %edx, %edx supposed to be high 32 bit which is
> not zero,
> It should be zero in this case. both lpj and tsc_khz should be u64 to work
> properly.

Good catch! Care to send a patch?

Thanks,

	tglx

[PATCH v15 00/26] Early boot time stamps

[Pavel Tatashin]

changelog
---------
v15 - v14
	Repo: https://github.com/soleen/time_15.git
	- dropped "x86/kvmclock: Avoid TSC recalibration" as Paolo Bonzini
	  suggested
	- Fixed in "sched: early boot clock" whenched_clock_running is set,
	   and moved __sched_clock_gtod_offset inside IRQ as Peter noticed.
	- Addressed comments from Dou Liyang: added missing __inits, and
	  X86_FEATURE_TSC_DEADLINE_TIMER, spelling.
	- Fixed xen_sched_clock_offset on xen hvm (noticed by Boris
	  Ostrovsky).
	- Added two patches to address Peter Zijlstra's request to split
	  native cpu calibration into early and late parts. The patches
	  are:
	  x86/tsc:  split native_calibrate_cpu() into early and late parts
	  x86/tsc:  use tsc_calibrate_cpu_early and
	    pit_hpet_ptimer_calibrate_cpu

v14 - v13
	- Included Thomas' KVM clock series, addressed comments from
	  reviewers.
	http://lkml.kernel.org/r/20180706161307.733337643@linutronix.de
	- Fixed xen hvm panic reported by Boris
	- Fixed build issue on microblaze

v13 - v12
	- Addressed comments from Thomas Gleixner.
	- Addressed comments from Peter Zijlstra.
	- Added a patch from Borislav Petkov
	- Added a new patch: sched: use static key for sched_clock_running
	- Added xen pv fixes, so clock is initialized when other
	  hypervisors initialize their clocks.
	Note: I am including kvm/x86: remove kvm memblock dependency, which
	is part of this series:
	http://lkml.kernel.org/r/20180706161307.733337643@linutronix.de
	Because without this patch it is not possible to test this series on
	KVM.

v12 - v11
	- split time: replace read_boot_clock64() with
	  read_persistent_wall_and_boot_offset() into four patches
	- Added two patches one fixes an existing bug with text_poke()
	  another one enables static branches early. Note, because I found
	  and fixed the text_poke() bug, enabling static branching became
	  super easy, as no changes to jump_label* is needed.
	- Modified x86/tsc: use tsc early to use static branches early, and
	  thus native_sched_clock() is not changed at all.
v11 - v10
	- Addressed all the comments from Thomas Gleixner.
	- I added one more patch:
	  "x86/tsc: prepare for early sched_clock" which fixes a problem
	  that I discovered while testing. I am not particularly happy with
	  the fix, as it adds a new argument that is used only in one
	  place, but if you have a suggestion for a different approach on
	  how to address this problem please let me know.

v10 - v9
	- Added another patch to this series that removes dependency
	  between KVM clock, and memblock allocator. The benefit is that
	  all clocks can now be initialized even earlier.
v9 - v8
	- Addressed more comments from Dou Liyang

v8 - v7
	- Addressed comments from Dou Liyang:
	- Moved tsc_early_init() and tsc_early_fini() to be all inside
	  tsc.c, and changed them to be static.
	- Removed warning when notsc parameter is used.
	- Merged with:
	  https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git

v7 - v6
	- Removed tsc_disabled flag, now notsc is equivalent of
	  tsc=unstable
	- Simplified changes to sched/clock.c, by removing the
	  sched_clock_early() and friends as requested by Peter Zijlstra.
	  We know always use sched_clock()
	- Modified x86 sched_clock() to return either early boot time or
	  regular.
	- Added another example why ealry boot time is important

v5 - v6
	- Added a new patch:
		time: sync read_boot_clock64() with persistent clock
	  Which fixes missing __init macro, and enabled time discrepancy
	  fix that was noted by Thomas Gleixner
	- Split "x86/time: read_boot_clock64() implementation" into a
	  separate patch

v4 - v5
	- Fix compiler warnings on systems with stable clocks.

v3 - v4
	- Fixed tsc_early_fini() call to be in the 2nd patch as reported
	  by Dou Liyang
	- Improved comment before __use_sched_clock_early to explain why
	  we need both booleans.
	- Simplified valid_clock logic in read_boot_clock64().

v2 - v3
	- Addressed comment from Thomas Gleixner
	- Timestamps are available a little later in boot but still much
	  earlier than in mainline. This significantly simplified this
	  work.

v1 - v2
	In patch "x86/tsc: tsc early":
	- added tsc_adjusted_early()
	- fixed 32-bit compile error use do_div()

The early boot time stamps were discussed recently in these threads:
http://lkml.kernel.org/r/1527672059-6225-1-git-send-email-feng.tang@intel.com
http://lkml.kernel.org/r/1527672059-6225-2-git-send-email-feng.tang@intel.com

I updated my series to the latest mainline and sending it again.

Peter mentioned he did not like patch 6,7, and we can discuss for a better
way to do that, but I think patches 1-5 can be accepted separetly, since
they already enable early timestamps on platforms where sched_clock() is
available early. Such as KVM.

Adding early boot time stamps support for x86 machines.
SPARC patches for early boot time stamps are already integrated into
mainline linux.

Sample output
-------------
Before:
https://paste.ubuntu.com/26133428/

After:
https://paste.ubuntu.com/26133523/

For exaples how early time stamps are used, see this work:
Example 1:
https://lwn.net/Articles/734374/
- Without early boot time stamps we would not know about the extra time
  that is spent zeroing struct pages early in boot even when deferred
  page initialization.

Example 2:
https://patchwork.kernel.org/patch/10021247/
- If early boot timestamps were available, the engineer who introduced
  this bug would have noticed the extra time that is spent early in boot.
Pavel Tatashin (7):
  x86/tsc: remove tsc_disabled flag
  time: sync read_boot_clock64() with persistent clock
  x86/time: read_boot_clock64() implementation
  sched: early boot clock
  kvm/x86: remove kvm memblock dependency
  x86/paravirt: add active_sched_clock to pv_time_ops
  x86/tsc: use tsc early

Example 3:
http://lkml.kernel.org/r/20180615155733.1175-1-pasha.tatashin@oracle.com
- Needed early time stamps to show improvement
Borislav Petkov (1):
  x86/CPU: Call detect_nopl() only on the BSP

Pavel Tatashin (19):
  x86/kvmclock: Remove memblock dependency
  x86: text_poke() may access uninitialized struct pages
  x86: initialize static branching early
  x86/tsc: redefine notsc to behave as tsc=unstable
  x86/xen/time: initialize pv xen time in init_hypervisor_platform
  x86/xen/time: output xen sched_clock time from 0
  s390/time: add read_persistent_wall_and_boot_offset()
  time: replace read_boot_clock64() with
    read_persistent_wall_and_boot_offset()
  time: default boot time offset to local_clock()
  s390/time: remove read_boot_clock64()
  ARM/time: remove read_boot_clock64()
  x86/tsc: calibrate tsc only once
  x86/tsc: initialize cyc2ns when tsc freq. is determined
  x86/tsc: use tsc early
  sched: move sched clock initialization and merge with generic clock
  sched: early boot clock
  sched: use static key for sched_clock_running
  x86/tsc:  split native_calibrate_cpu() into early and late parts
  x86/tsc:  use tsc_calibrate_cpu_early and
    pit_hpet_ptimer_calibrate_cpu

Thomas Gleixner (6):
  x86/kvmclock: Remove page size requirement from wall_clock
  x86/kvmclock: Decrapify kvm_register_clock()
  x86/kvmclock: Cleanup the code
  x86/kvmclock: Mark variables __initdata and __ro_after_init
  x86/kvmclock: Move kvmclock vsyscall param and init to kvmclock
  x86/kvmclock: Switch kvmclock data to a PER_CPU variable

 .../admin-guide/kernel-parameters.txt         |   2 -
 Documentation/x86/x86_64/boot-options.txt     |   4 +-
 arch/arm/include/asm/mach/time.h              |   3 +-
 arch/arm/kernel/time.c                        |  15 +-
 arch/arm/plat-omap/counter_32k.c              |   2 +-
 arch/s390/kernel/time.c                       |  15 +-
 arch/x86/include/asm/kvm_guest.h              |   7 -
 arch/x86/include/asm/kvm_para.h               |   1 -
 arch/x86/include/asm/text-patching.h          |   1 +
 arch/x86/include/asm/tsc.h                    |   4 +-
 arch/x86/kernel/alternative.c                 |   7 +
 arch/x86/kernel/cpu/amd.c                     |  13 +-
 arch/x86/kernel/cpu/common.c                  |  40 +--
 arch/x86/kernel/jump_label.c                  |  11 +-
 arch/x86/kernel/kvm.c                         |  14 +-
 arch/x86/kernel/kvmclock.c                    | 256 +++++++-----------
 arch/x86/kernel/setup.c                       |  10 +-
 arch/x86/kernel/tsc.c                         | 253 +++++++++--------
 arch/x86/kernel/x86_init.c                    |   2 +-
 arch/x86/xen/enlighten_pv.c                   |  51 ++--
 arch/x86/xen/mmu_pv.c                         |   6 +-
 arch/x86/xen/suspend_pv.c                     |   5 +-
 arch/x86/xen/time.c                           |  18 +-
 arch/x86/xen/xen-ops.h                        |   6 +-
 drivers/clocksource/tegra20_timer.c           |   2 +-
 include/linux/sched_clock.h                   |   5 +-
 include/linux/timekeeping.h                   |   3 +-
 init/main.c                                   |   4 +-
 kernel/sched/clock.c                          |  59 ++--
 kernel/sched/core.c                           |   1 -
 kernel/sched/debug.c                          |   2 -
 kernel/time/sched_clock.c                     |   2 +-
 kernel/time/timekeeping.c                     |  62 +++--
 33 files changed, 439 insertions(+), 447 deletions(-)
 delete mode 100644 arch/x86/include/asm/kvm_guest.h

-- 
2.18.0

[PATCH v15 19/26] x86/tsc: calibrate tsc only once

[Pavel Tatashin]

During boot tsc is calibrated twice: once in tsc_early_delay_calibrate(),
and the second time in tsc_init().

Rename tsc_early_delay_calibrate() to tsc_early_init(), and rework it so
the calibration is done only early, and make tsc_init() to use the values
already determined in tsc_early_init().

Sometimes it is not possible to determine tsc early, as the subsystem that
is required is not yet initialized, in such case try again later in
tsc_init().

Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Pavel Tatashin <pasha.tatashin@oracle.com>
---
 arch/x86/include/asm/tsc.h |  2 +-
 arch/x86/kernel/setup.c    |  2 +-
 arch/x86/kernel/tsc.c      | 87 ++++++++++++++++++++------------------
 3 files changed, 49 insertions(+), 42 deletions(-)

diff --git a/arch/x86/include/asm/tsc.h b/arch/x86/include/asm/tsc.h
index 2701d221583a..c4368ff73652 100644
--- a/arch/x86/include/asm/tsc.h
+++ b/arch/x86/include/asm/tsc.h
@@ -33,7 +33,7 @@ static inline cycles_t get_cycles(void)
 extern struct system_counterval_t convert_art_to_tsc(u64 art);
 extern struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns);
 
-extern void tsc_early_delay_calibrate(void);
+extern void tsc_early_init(void);
 extern void tsc_init(void);
 extern void mark_tsc_unstable(char *reason);
 extern int unsynchronized_tsc(void);
diff --git a/arch/x86/kernel/setup.c b/arch/x86/kernel/setup.c
index 7490de925a81..5d32c55aeb8b 100644
--- a/arch/x86/kernel/setup.c
+++ b/arch/x86/kernel/setup.c
@@ -1014,6 +1014,7 @@ void __init setup_arch(char **cmdline_p)
 	 */
 	init_hypervisor_platform();
 
+	tsc_early_init();
 	x86_init.resources.probe_roms();
 
 	/* after parse_early_param, so could debug it */
@@ -1199,7 +1200,6 @@ void __init setup_arch(char **cmdline_p)
 
 	memblock_find_dma_reserve();
 
-	tsc_early_delay_calibrate();
 	if (!early_xdbc_setup_hardware())
 		early_xdbc_register_console();
 
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
index 186395041725..4cab2236169e 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -33,6 +33,8 @@ EXPORT_SYMBOL(cpu_khz);
 unsigned int __read_mostly tsc_khz;
 EXPORT_SYMBOL(tsc_khz);
 
+#define KHZ	1000
+
 /*
  * TSC can be unstable due to cpufreq or due to unsynced TSCs
  */
@@ -1335,34 +1337,10 @@ static int __init init_tsc_clocksource(void)
  */
 device_initcall(init_tsc_clocksource);
 
-void __init tsc_early_delay_calibrate(void)
-{
-	unsigned long lpj;
-
-	if (!boot_cpu_has(X86_FEATURE_TSC))
-		return;
-
-	cpu_khz = x86_platform.calibrate_cpu();
-	tsc_khz = x86_platform.calibrate_tsc();
-
-	tsc_khz = tsc_khz ? : cpu_khz;
-	if (!tsc_khz)
-		return;
-
-	lpj = tsc_khz * 1000;
-	do_div(lpj, HZ);
-	loops_per_jiffy = lpj;
-}
-
-void __init tsc_init(void)
+static bool __init determine_cpu_tsc_frequencies(void)
 {
-	u64 lpj, cyc;
-	int cpu;
-
-	if (!boot_cpu_has(X86_FEATURE_TSC)) {
-		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
-		return;
-	}
+	/* Make sure that cpu and tsc are not already calibrated */
+	WARN_ON(cpu_khz || tsc_khz);
 
 	cpu_khz = x86_platform.calibrate_cpu();
 	tsc_khz = x86_platform.calibrate_tsc();
@@ -1377,20 +1355,52 @@ void __init tsc_init(void)
 	else if (abs(cpu_khz - tsc_khz) * 10 > tsc_khz)
 		cpu_khz = tsc_khz;
 
-	if (!tsc_khz) {
-		mark_tsc_unstable("could not calculate TSC khz");
-		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
-		return;
-	}
+	if (tsc_khz == 0)
+		return false;
 
 	pr_info("Detected %lu.%03lu MHz processor\n",
-		(unsigned long)cpu_khz / 1000,
-		(unsigned long)cpu_khz % 1000);
+		(unsigned long)cpu_khz / KHZ,
+		(unsigned long)cpu_khz % KHZ);
 
 	if (cpu_khz != tsc_khz) {
 		pr_info("Detected %lu.%03lu MHz TSC",
-			(unsigned long)tsc_khz / 1000,
-			(unsigned long)tsc_khz % 1000);
+			(unsigned long)tsc_khz / KHZ,
+			(unsigned long)tsc_khz % KHZ);
+	}
+	return true;
+}
+
+static unsigned long __init get_loops_per_jiffy(void)
+{
+	unsigned long lpj = tsc_khz * KHZ;
+
+	do_div(lpj, HZ);
+	return lpj;
+}
+
+void __init tsc_early_init(void)
+{
+	if (!boot_cpu_has(X86_FEATURE_TSC))
+		return;
+	if (!determine_cpu_tsc_frequencies())
+		return;
+	loops_per_jiffy = get_loops_per_jiffy();
+}
+
+void __init tsc_init(void)
+{
+	if (!boot_cpu_has(X86_FEATURE_TSC)) {
+		setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
+		return;
+	}
+
+	if (!tsc_khz) {
+		/* We failed to determine frequencies earlier, try again */
+		if (!determine_cpu_tsc_frequencies()) {
+			mark_tsc_unstable("could not calculate TSC khz");
+			setup_clear_cpu_cap(X86_FEATURE_TSC_DEADLINE_TIMER);
+			return;
+		}
 	}
 
 	/* Sanitize TSC ADJUST before cyc2ns gets initialized */
@@ -1413,10 +1423,7 @@ void __init tsc_init(void)
 	if (!no_sched_irq_time)
 		enable_sched_clock_irqtime();
 
-	lpj = ((u64)tsc_khz * 1000);
-	do_div(lpj, HZ);
-	lpj_fine = lpj;
-
+	lpj_fine = get_loops_per_jiffy();
 	use_tsc_delay();
 
 	check_system_tsc_reliable();
-- 
2.18.0