Linux Virtualization - Mar 2017 - [PATCH v3 2/3] x86/hyperv: move TSC reading method to asm/mshyperv.h

Hi,

merge window is about to close so I hope it's OK to make another try here.

Changes since v2:
- Add explicit READ_ONCE() to not rely on 'volatile' [Andy Lutomirski]
- rdtsc() -> rdtsc_ordered() [Andy Lutomirski]
- virt_rmb() -> smp_rmb() [Thomas Gleixner, Andy Lutomirski]

Thomas, Andy, it seems the only blocker for the series was the ambiguity with
TSC page read algorithm. I contacted Microsoft (through K. Y.) and asked what
we should do when we see 'seq=0'. The answer is:

"I have confirmed that the only invalid value is 0 (notwithstanding what
the
spec says). I have asked the Spec to be updated and the current code we have
is correct - it treats 0 as the only invalid value. The invalid value indicates
that the TSC page cannot be used as a time source and a different source is to
be used and this state is going to persist and so looping is not an
option."

I agree it would be wiser to have two invalid values - one for 'try
again' and
another for permanent failure in case it is needed. But this is not the case.

So I keep my algorithm untouched.

I can see two more reasons for us to keep it:
1) It is what we already have in kernel.
2) It is what Windows does (see the disassembly in c35b82ef0294.

Original description:

Hyper-V TSC page clocksource is suitable for vDSO, however, the protocol
defined by the hypervisor is different from VCLOCK_PVCLOCK. Implemented the
required support. Simple sysbench test shows the following results:

Before:
# time sysbench --test=memory --max-requests=500000 run
...
real    1m22.618s
user    0m50.193s
sys     0m32.268s

After:
# time sysbench --test=memory --max-requests=500000 run
...
real	0m47.241s
user	0m47.117s
sys	0m0.008s

Vitaly Kuznetsov (3):
  x86/hyperv: implement hv_get_tsc_page()
  x86/hyperv: move TSC reading method to asm/mshyperv.h
  x86/vdso: Add VCLOCK_HVCLOCK vDSO clock read method

 arch/x86/entry/vdso/vclock_gettime.c  | 24 +++++++++++++++
 arch/x86/entry/vdso/vdso-layout.lds.S |  3 +-
 arch/x86/entry/vdso/vdso2c.c          |  3 ++
 arch/x86/entry/vdso/vma.c             |  7 +++++
 arch/x86/hyperv/hv_init.c             | 48 +++++++++--------------------
 arch/x86/include/asm/clocksource.h    |  3 +-
 arch/x86/include/asm/mshyperv.h       | 58 +++++++++++++++++++++++++++++++++++
 arch/x86/include/asm/vdso.h           |  1 +
 drivers/hv/Kconfig                    |  3 ++
 9 files changed, 114 insertions(+), 36 deletions(-)

-- 
2.9.3

To use Hyper-V TSC page clocksource from vDSO we need to make tsc_pg
available. Implement hv_get_tsc_page() and add CONFIG_HYPERV_TSCPAGE to
make #ifdef-s simple.

Signed-off-by: Vitaly Kuznetsov <vkuznets at redhat.com>
---
 arch/x86/hyperv/hv_init.c       | 9 +++++++--
 arch/x86/include/asm/mshyperv.h | 8 ++++++++
 drivers/hv/Kconfig              | 3 +++
 3 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
index db64baf0..6b64cae 100644
--- a/arch/x86/hyperv/hv_init.c
+++ b/arch/x86/hyperv/hv_init.c
@@ -27,10 +27,15 @@
 #include <linux/clockchips.h>
 
 
-#ifdef CONFIG_X86_64
+#ifdef CONFIG_HYPERV_TSCPAGE
 
 static struct ms_hyperv_tsc_page *tsc_pg;
 
+struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
+{
+	return tsc_pg;
+}
+
 static u64 read_hv_clock_tsc(struct clocksource *arg)
 {
 	u64 current_tick;
@@ -139,7 +144,7 @@ void hyperv_init(void)
 	/*
 	 * Register Hyper-V specific clocksource.
 	 */
-#ifdef CONFIG_X86_64
+#ifdef CONFIG_HYPERV_TSCPAGE
 	if (ms_hyperv.features & HV_X64_MSR_REFERENCE_TSC_AVAILABLE) {
 		union hv_x64_msr_hypercall_contents tsc_msr;
 
diff --git a/arch/x86/include/asm/mshyperv.h b/arch/x86/include/asm/mshyperv.h
index 7c9c895..d324dce 100644
--- a/arch/x86/include/asm/mshyperv.h
+++ b/arch/x86/include/asm/mshyperv.h
@@ -176,4 +176,12 @@ void hyperv_report_panic(struct pt_regs *regs);
 bool hv_is_hypercall_page_setup(void);
 void hyperv_cleanup(void);
 #endif
+#ifdef CONFIG_HYPERV_TSCPAGE
+struct ms_hyperv_tsc_page *hv_get_tsc_page(void);
+#else
+static inline struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
+{
+	return NULL;
+}
+#endif
 #endif
diff --git a/drivers/hv/Kconfig b/drivers/hv/Kconfig
index 0403b51..c29cd53 100644
--- a/drivers/hv/Kconfig
+++ b/drivers/hv/Kconfig
@@ -7,6 +7,9 @@ config HYPERV
 	  Select this option to run Linux as a Hyper-V client operating
 	  system.
 
+config HYPERV_TSCPAGE
+       def_bool HYPERV && X86_64
+
 config HYPERV_UTILS
 	tristate "Microsoft Hyper-V Utilities driver"
 	depends on HYPERV && CONNECTOR && NLS
-- 
2.9.3

As a preparation to making Hyper-V TSC page suitable for vDSO move
the TSC page reading logic to asm/mshyperv.h. While on it, do the
following
- Document the reading algorithm.
- Simplify the code a bit.
- Add explicit READ_ONCE() to not rely on 'volatile'.
- Add explicit barriers to prevent re-ordering (we need to read sequence
  strictly before and after)
- Use mul_u64_u64_shr() instead of assembly, gcc generates a single
'mul'
  instruction on x86_64 anyway.

Signed-off-by: Vitaly Kuznetsov <vkuznets at redhat.com>
---
 arch/x86/hyperv/hv_init.c       | 36 ++++-------------------------
 arch/x86/include/asm/mshyperv.h | 50 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 54 insertions(+), 32 deletions(-)

diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
index 6b64cae..63dd00e 100644
--- a/arch/x86/hyperv/hv_init.c
+++ b/arch/x86/hyperv/hv_init.c
@@ -38,39 +38,11 @@ struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
 
 static u64 read_hv_clock_tsc(struct clocksource *arg)
 {
-	u64 current_tick;
+	u64 current_tick = hv_read_tsc_page(tsc_pg);
+
+	if (current_tick == U64_MAX)
+		rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
 
-	if (tsc_pg->tsc_sequence != 0) {
-		/*
-		 * Use the tsc page to compute the value.
-		 */
-
-		while (1) {
-			u64 tmp;
-			u32 sequence = tsc_pg->tsc_sequence;
-			u64 cur_tsc;
-			u64 scale = tsc_pg->tsc_scale;
-			s64 offset = tsc_pg->tsc_offset;
-
-			rdtscll(cur_tsc);
-			/* current_tick = ((cur_tsc *scale) >> 64) + offset */
-			asm("mulq %3"
-				: "=d" (current_tick), "=a" (tmp)
-				: "a" (cur_tsc), "r" (scale));
-
-			current_tick += offset;
-			if (tsc_pg->tsc_sequence == sequence)
-				return current_tick;
-
-			if (tsc_pg->tsc_sequence != 0)
-				continue;
-			/*
-			 * Fallback using MSR method.
-			 */
-			break;
-		}
-	}
-	rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
 	return current_tick;
 }
 
diff --git a/arch/x86/include/asm/mshyperv.h b/arch/x86/include/asm/mshyperv.h
index d324dce..4ff25436 100644
--- a/arch/x86/include/asm/mshyperv.h
+++ b/arch/x86/include/asm/mshyperv.h
@@ -178,6 +178,56 @@ void hyperv_cleanup(void);
 #endif
 #ifdef CONFIG_HYPERV_TSCPAGE
 struct ms_hyperv_tsc_page *hv_get_tsc_page(void);
+static inline u64 hv_read_tsc_page(const struct ms_hyperv_tsc_page *tsc_pg)
+{
+	u64 scale, offset, current_tick, cur_tsc;
+	u32 sequence;
+
+	/*
+	 * The protocol for reading Hyper-V TSC page is specified in Hypervisor
+	 * Top-Level Functional Specification ver. 3.0 and above. To get the
+	 * reference time we must do the following:
+	 * - READ ReferenceTscSequence
+	 *   A special '0' value indicates the time source is unreliable and
we
+	 *   need to use something else. The currently published specification
+	 *   versions (up to 4.0b) contain a mistake and wrongly claim '-1'
+	 *   instead of '0' as the special value, see commit c35b82ef0294.
+	 * - ReferenceTime +	 *        ((RDTSC() * ReferenceTscScale) >> 64) +
ReferenceTscOffset
+	 * - READ ReferenceTscSequence again. In case its value has changed
+	 *   since our first reading we need to discard ReferenceTime and repeat
+	 *   the whole sequence as the hypervisor was updating the page in
+	 *   between.
+	 */
+	while (1) {
+		sequence = READ_ONCE(tsc_pg->tsc_sequence);
+		if (!sequence)
+			break;
+		/*
+		 * Make sure we read sequence before we read other values from
+		 * TSC page.
+		 */
+		smp_rmb();
+
+		scale = READ_ONCE(tsc_pg->tsc_scale);
+		offset = READ_ONCE(tsc_pg->tsc_offset);
+		cur_tsc = rdtsc_ordered();
+
+		current_tick = mul_u64_u64_shr(cur_tsc, scale, 64) + offset;
+
+		/*
+		 * Make sure we read sequence after we read all other values
+		 * from TSC page.
+		 */
+		smp_rmb();
+
+		if (READ_ONCE(tsc_pg->tsc_sequence) == sequence)
+			return current_tick;
+	}
+
+	return U64_MAX;
+}
+
 #else
 static inline struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
 {
-- 
2.9.3

Hyper-V TSC page clocksource is suitable for vDSO, however, the protocol
defined by the hypervisor is different from VCLOCK_PVCLOCK. Implement the
required support by adding hvclock_page VVAR.

Signed-off-by: Vitaly Kuznetsov <vkuznets at redhat.com>
---
 arch/x86/entry/vdso/vclock_gettime.c  | 24 ++++++++++++++++++++++++
 arch/x86/entry/vdso/vdso-layout.lds.S |  3 ++-
 arch/x86/entry/vdso/vdso2c.c          |  3 +++
 arch/x86/entry/vdso/vma.c             |  7 +++++++
 arch/x86/hyperv/hv_init.c             |  3 +++
 arch/x86/include/asm/clocksource.h    |  3 ++-
 arch/x86/include/asm/vdso.h           |  1 +
 7 files changed, 42 insertions(+), 2 deletions(-)

diff --git a/arch/x86/entry/vdso/vclock_gettime.c
b/arch/x86/entry/vdso/vclock_gettime.c
index 9d4d6e1..fa8dbfc 100644
--- a/arch/x86/entry/vdso/vclock_gettime.c
+++ b/arch/x86/entry/vdso/vclock_gettime.c
@@ -17,6 +17,7 @@
 #include <asm/unistd.h>
 #include <asm/msr.h>
 #include <asm/pvclock.h>
+#include <asm/mshyperv.h>
 #include <linux/math64.h>
 #include <linux/time.h>
 #include <linux/kernel.h>
@@ -32,6 +33,11 @@ extern u8 pvclock_page
 	__attribute__((visibility("hidden")));
 #endif
 
+#ifdef CONFIG_HYPERV_TSCPAGE
+extern u8 hvclock_page
+	__attribute__((visibility("hidden")));
+#endif
+
 #ifndef BUILD_VDSO32
 
 notrace static long vdso_fallback_gettime(long clock, struct timespec *ts)
@@ -141,6 +147,20 @@ static notrace u64 vread_pvclock(int *mode)
 	return last;
 }
 #endif
+#ifdef CONFIG_HYPERV_TSCPAGE
+static notrace u64 vread_hvclock(int *mode)
+{
+	const struct ms_hyperv_tsc_page *tsc_pg +		(const struct ms_hyperv_tsc_page
*)&hvclock_page;
+	u64 current_tick = hv_read_tsc_page(tsc_pg);
+
+	if (current_tick != U64_MAX)
+		return current_tick;
+
+	*mode = VCLOCK_NONE;
+	return 0;
+}
+#endif
 
 notrace static u64 vread_tsc(void)
 {
@@ -173,6 +193,10 @@ notrace static inline u64 vgetsns(int *mode)
 	else if (gtod->vclock_mode == VCLOCK_PVCLOCK)
 		cycles = vread_pvclock(mode);
 #endif
+#ifdef CONFIG_HYPERV_TSCPAGE
+	else if (gtod->vclock_mode == VCLOCK_HVCLOCK)
+		cycles = vread_hvclock(mode);
+#endif
 	else
 		return 0;
 	v = (cycles - gtod->cycle_last) & gtod->mask;
diff --git a/arch/x86/entry/vdso/vdso-layout.lds.S
b/arch/x86/entry/vdso/vdso-layout.lds.S
index a708aa9..8ebb4b6 100644
--- a/arch/x86/entry/vdso/vdso-layout.lds.S
+++ b/arch/x86/entry/vdso/vdso-layout.lds.S
@@ -25,7 +25,7 @@ SECTIONS
 	 * segment.
 	 */
 
-	vvar_start = . - 2 * PAGE_SIZE;
+	vvar_start = . - 3 * PAGE_SIZE;
 	vvar_page = vvar_start;
 
 	/* Place all vvars at the offsets in asm/vvar.h. */
@@ -36,6 +36,7 @@ SECTIONS
 #undef EMIT_VVAR
 
 	pvclock_page = vvar_start + PAGE_SIZE;
+	hvclock_page = vvar_start + 2 * PAGE_SIZE;
 
 	. = SIZEOF_HEADERS;
 
diff --git a/arch/x86/entry/vdso/vdso2c.c b/arch/x86/entry/vdso/vdso2c.c
index 491020b..0780a44 100644
--- a/arch/x86/entry/vdso/vdso2c.c
+++ b/arch/x86/entry/vdso/vdso2c.c
@@ -74,6 +74,7 @@ enum {
 	sym_vvar_page,
 	sym_hpet_page,
 	sym_pvclock_page,
+	sym_hvclock_page,
 	sym_VDSO_FAKE_SECTION_TABLE_START,
 	sym_VDSO_FAKE_SECTION_TABLE_END,
 };
@@ -82,6 +83,7 @@ const int special_pages[] = {
 	sym_vvar_page,
 	sym_hpet_page,
 	sym_pvclock_page,
+	sym_hvclock_page,
 };
 
 struct vdso_sym {
@@ -94,6 +96,7 @@ struct vdso_sym required_syms[] = {
 	[sym_vvar_page] = {"vvar_page", true},
 	[sym_hpet_page] = {"hpet_page", true},
 	[sym_pvclock_page] = {"pvclock_page", true},
+	[sym_hvclock_page] = {"hvclock_page", true},
 	[sym_VDSO_FAKE_SECTION_TABLE_START] = {
 		"VDSO_FAKE_SECTION_TABLE_START", false
 	},
diff --git a/arch/x86/entry/vdso/vma.c b/arch/x86/entry/vdso/vma.c
index 572cee3..71f5d3a 100644
--- a/arch/x86/entry/vdso/vma.c
+++ b/arch/x86/entry/vdso/vma.c
@@ -21,6 +21,7 @@
 #include <asm/page.h>
 #include <asm/desc.h>
 #include <asm/cpufeature.h>
+#include <asm/mshyperv.h>
 
 #if defined(CONFIG_X86_64)
 unsigned int __read_mostly vdso64_enabled = 1;
@@ -120,6 +121,12 @@ static int vvar_fault(const struct vm_special_mapping *sm,
 				vmf->address,
 				__pa(pvti) >> PAGE_SHIFT);
 		}
+	} else if (sym_offset == image->sym_hvclock_page) {
+		struct ms_hyperv_tsc_page *tsc_pg = hv_get_tsc_page();
+
+		if (tsc_pg && vclock_was_used(VCLOCK_HVCLOCK))
+			ret = vm_insert_pfn(vma, vmf->address,
+					    vmalloc_to_pfn(tsc_pg));
 	}
 
 	if (ret == 0 || ret == -EBUSY)
diff --git a/arch/x86/hyperv/hv_init.c b/arch/x86/hyperv/hv_init.c
index 63dd00e..d08ac5e 100644
--- a/arch/x86/hyperv/hv_init.c
+++ b/arch/x86/hyperv/hv_init.c
@@ -132,6 +132,9 @@ void hyperv_init(void)
 		tsc_msr.guest_physical_address = vmalloc_to_pfn(tsc_pg);
 
 		wrmsrl(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
+
+		hyperv_cs_tsc.archdata.vclock_mode = VCLOCK_HVCLOCK;
+
 		clocksource_register_hz(&hyperv_cs_tsc, NSEC_PER_SEC/100);
 		return;
 	}
diff --git a/arch/x86/include/asm/clocksource.h
b/arch/x86/include/asm/clocksource.h
index eae33c7..47bea8c 100644
--- a/arch/x86/include/asm/clocksource.h
+++ b/arch/x86/include/asm/clocksource.h
@@ -6,7 +6,8 @@
 #define VCLOCK_NONE	0	/* No vDSO clock available.		*/
 #define VCLOCK_TSC	1	/* vDSO should use vread_tsc.		*/
 #define VCLOCK_PVCLOCK	2	/* vDSO should use vread_pvclock.	*/
-#define VCLOCK_MAX	2
+#define VCLOCK_HVCLOCK	3	/* vDSO should use vread_hvclock.	*/
+#define VCLOCK_MAX	3
 
 struct arch_clocksource_data {
 	int vclock_mode;
diff --git a/arch/x86/include/asm/vdso.h b/arch/x86/include/asm/vdso.h
index 2444189..bccdf49 100644
--- a/arch/x86/include/asm/vdso.h
+++ b/arch/x86/include/asm/vdso.h
@@ -20,6 +20,7 @@ struct vdso_image {
 	long sym_vvar_page;
 	long sym_hpet_page;
 	long sym_pvclock_page;
+	long sym_hvclock_page;
 	long sym_VDSO32_NOTE_MASK;
 	long sym___kernel_sigreturn;
 	long sym___kernel_rt_sigreturn;
-- 
2.9.3

Minor coding comments
> diff --git a/arch/x86/include/asm/mshyperv.h
b/arch/x86/include/asm/mshyperv.h
> index d324dce..4ff25436 100644
> --- a/arch/x86/include/asm/mshyperv.h
> +++ b/arch/x86/include/asm/mshyperv.h
> @@ -178,6 +178,56 @@ void hyperv_cleanup(void);
>  #endif
>  #ifdef CONFIG_HYPERV_TSCPAGE
>  struct ms_hyperv_tsc_page *hv_get_tsc_page(void);
> +static inline u64 hv_read_tsc_page(const struct ms_hyperv_tsc_page
*tsc_pg)
> +{
> +	u64 scale, offset, current_tick, cur_tsc;
> +	u32 sequence;
> +
> +	/*
> +	 * The protocol for reading Hyper-V TSC page is specified in Hypervisor
> +	 * Top-Level Functional Specification ver. 3.0 and above. To get the
> +	 * reference time we must do the following:
> +	 * - READ ReferenceTscSequence
> +	 *   A special '0' value indicates the time source is unreliable
and we
> +	 *   need to use something else. The currently published specification
> +	 *   versions (up to 4.0b) contain a mistake and wrongly claim
'-1'
> +	 *   instead of '0' as the special value, see commit
c35b82ef0294.
> +	 * - ReferenceTime > +	 *        ((RDTSC() * ReferenceTscScale)
>> 64) + ReferenceTscOffset
> +	 * - READ ReferenceTscSequence again. In case its value has changed
> +	 *   since our first reading we need to discard ReferenceTime and repeat
> +	 *   the whole sequence as the hypervisor was updating the page in
> +	 *   between.
> +	 */
> +	while (1) {
> +		sequence = READ_ONCE(tsc_pg->tsc_sequence);
> +		if (!sequence)
> +			break;
It would be clearer to just return U64_MAX here (and not fall out)
since this is only case here. Also since this failure only occurs if host
clock is not available, probably should be unlikely.
> +		/*
> +		 * Make sure we read sequence before we read other values from
> +		 * TSC page.
> +		 */
> +		smp_rmb();
> +
> +		scale = READ_ONCE(tsc_pg->tsc_scale);
> +		offset = READ_ONCE(tsc_pg->tsc_offset);
> +		cur_tsc = rdtsc_ordered();
Since you already have smp_ barriers and rdtsc_ordered is a barrier,
the compiler barriers (READ_ONCE()) shouldn't be necessary.
> +
> +		current_tick = mul_u64_u64_shr(cur_tsc, scale, 64) + offset;
> +
> +		/*
> +		 * Make sure we read sequence after we read all other values
> +		 * from TSC page.
> +		 */
> +		smp_rmb();
> +
> +		if (READ_ONCE(tsc_pg->tsc_sequence) == sequence)
> +			return current_tick;
> +	}
Why not make do { } while out of this.

	do {
...
	} while (unlikely(READ_ONCE(tsc_pg->tsc_sequence) != sequence);
	return current_tick;

Also don't need to calculate tick value until have good data. As in:

static inline u32 hv_clock_sequence(const struct ms_hyperv_tsc_page *tsc_pg)
{
	u32 sequence 	return sequence;
}

static inline u64 hv_read_tsc_page(const struct ms_hyperv_tsc_page *tsc_pg)
{
	u64 scale, offset, cur_tsc;
	u32 start;

	/*
	 * The protocol for reading Hyper-V TSC page is specified in Hypervisor
	 * Top-Level Functional Specification ver. 3.0 and above. To get the
	 * reference time we must do the following:
	 * - READ ReferenceTscSequence
	 *   A special '0' value indicates the time source is unreliable and we
	 *   need to use something else. The currently published specification
	 *   versions (up to 4.0b) contain a mistake and wrongly claim '-1'
	 *   instead of '0' as the special value, see commit c35b82ef0294.
	 * - ReferenceTime 	 *        ((RDTSC() * ReferenceTscScale) >> 64) +
ReferenceTscOffset
	 * - READ ReferenceTscSequence again. In case its value has changed
	 *   since our first reading we need to discard ReferenceTime and repeat
	 *   the whole sequence as the hypervisor was updating the page in
	 *   between.
	 */
	do {
		start = READ_ONCE(tsc_pg->tsc_sequence);
		smp_rmb();

		if (unlikely(!start))
			return U64_MAX;

		scale = tsc_pg->tsc_scale;
		offset = tsc_pg->tsc_offset;

		/*
		 * Make sure we read sequence after we read all other values
		 * from TSC page.
		 */
		smp_rmb();
	} while (unlikely(READ_ONCE(tsc_pg->tsc_sequence != start)));

	cur_tsc = rdtsc_ordered();
	return mul_u64_u64_shr(cur_tsc, scale, 64) + offset;
}