Skip to contentSkip to conversation

x86/tsc: fix 64bit divisor be truncated in calc_hpet_ref

PATCH In review Xiaoming Gao 3 participants 3 messages Apr 13, 2018 – Apr 17, 2018 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/tsc: fix 64bit divisor be truncated in calc_hpet_ref

From ba3d2fb699c4d8ee61b05d7e70be48b9c4e22baf Mon Sep 17 00:00:00 2001
From: Xiaoming Gao <[email protected]>
Date: Fri, 13 Apr 2018 17:05:18 +0800
Subject: [PATCH] x86/tsc: fix 64bit divisor be truncated in calc_hpet_ref

the HPET frequency got larger on intel skylake, thus could cause tmp to exceed 32bits.
do_div will truncate 64bits tmp to 32bits, so the frequency calced via HPET will be wrong, use div64_u64 can fix it.

---
 arch/x86/kernel/tsc.c |    2 +-

 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/tsc.c +0 -0
index 9714a7a..8700269 100644
@@ -160,7 +160,7 @@ static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)

     hpet2 -= hpet1;
     tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD)); do_div(tmp, 1000000);
-    do_div(deltatsc, tmp);
+    deltatsc = div64_u64(deltatsc, tmp);

     return (unsigned long) deltatsc;
 }

Signature · 1 line
1.7.1

From ba3d2fb699c4d8ee61b05d7e70be48b9c4e22baf Mon Sep 17 00:00:00 2001
From: Xiaoming Gao <[email protected]>
Date: Fri, 13 Apr 2018 17:05:18 +0800
Subject: [PATCH] x86/tsc: fix 64bit divisor be truncated in calc_hpet_ref

the HPET frequency got larger on intel skylake, thus could cause tmp to exceed 32bits.
do_div will truncate 64bits tmp to 32bits, so the frequency calced via HPET will be wrong, use div64_u64 can fix it.

---
 arch/x86/kernel/tsc.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/tsc.c +1 -1
index 9714a7a..8700269 100644
@@ -160,7 +160,7 @@ static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 	hpet2 -= hpet1;
 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 	do_div(tmp, 1000000);
-	do_div(deltatsc, tmp);
+	deltatsc = div64_u64(deltatsc, tmp);
 
 	return (unsigned long) deltatsc;
 }
Signature · 1 line
1.7.1
#1 lore raw
Quoted 20 lines · Xiaoming Gao

From ba3d2fb699c4d8ee61b05d7e70be48b9c4e22baf Mon Sep 17 00:00:00 2001
From: Xiaoming Gao <[email protected]>
Date: Fri, 13 Apr 2018 17:05:18 +0800
Subject: [PATCH] x86/tsc: fix 64bit divisor be truncated in calc_hpet_ref

the HPET frequency got larger on intel skylake, thus could cause tmp to exceed 32bits.
do_div will truncate 64bits tmp to 32bits, so the frequency calced via HPET will be wrong, use div64_u64 can fix it.

---
 arch/x86/kernel/tsc.c |    2 +-

 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/tsc.c +0 -0
index 9714a7a..8700269 100644

arch/x86/kernel/tsc.c static unsigned long calc_hpet_ref(u64 deltatsc, u64 · line 166

Quoted 8 lines
arch/x86/kernel/tsc.c +1 -1
@@ -160,7 +160,7 @@ static unsigned long calc_hpet_ref(u64 deltatsc, u64
 hpet1, u64 hpet2)
      hpet2 -= hpet1;
      tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
      do_div(tmp, 1000000);
-    do_div(deltatsc, tmp);
+    deltatsc = div64_u64(deltatsc, tmp);
 
      return (unsigned long) deltatsc;

Please fix your MUA so it does not whitespace damage patches. Documentation/process/email-clients.rst tells you how to teach thunderbird to do the right thing.

Attaching the patch does not help either as it breaks the workflow. I've fixed it up manually this time, but please be more careful next time.

Thanks,

tglx

#2 lore raw

[tip:x86/urgent] x86/tsc: Prevent 32bit truncation in calc_hpet_ref()

In reply to Xiaoming Gao
x86/tsc: Prevent 32bit truncation in calc_hpet_ref()
Commit-ID:  d3878e164dcd3925a237a20e879432400e369172
Gitweb:     https://git.kernel.org/tip/d3878e164dcd3925a237a20e879432400e369172
Author:     Xiaoming Gao <[email protected]>
AuthorDate: Fri, 13 Apr 2018 17:48:08 +0800
Committer:  Thomas Gleixner <[email protected]>
CommitDate: Tue, 17 Apr 2018 11:50:42 +0200

x86/tsc: Prevent 32bit truncation in calc_hpet_ref()

The TSC calibration code uses HPET as reference. The conversion normalizes the delta of two HPET timestamps:

    hpetref = ((tshpet1 - tshpet2) * HPET_PERIOD) / 1e6

and then divides the normalized delta of the corresponding TSC timestamps by the result to calulate the TSC frequency.

    tscfreq = ((tstsc1 - tstsc2 ) * 1e6) / hpetref

This uses do_div() which takes an u32 as the divisor, which worked so far because the HPET frequency was low enough that 'hpetref' never exceeded 32bit.

On Skylake machines the HPET frequency increased so 'hpetref' can exceed 32bit. do_div() truncates the divisor, which causes the calibration to fail.

Use div64_u64() to avoid the problem.

[ tglx: Fixes whitespace mangled patch and rewrote changelog ]

---
 arch/x86/kernel/tsc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kernel/tsc.c +1 -1
index ef32297ff17e..91e6da48cbb6 100644
@@ -317,7 +317,7 @@ static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
 	hpet2 -= hpet1;
 	tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
 	do_div(tmp, 1000000);
-	do_div(deltatsc, tmp);
+	deltatsc = div64_u64(deltatsc, tmp);
 
 	return (unsigned long) deltatsc;
 }