diff mbox series

[v10,2/5] KVM: stats: Add fd-based API to read binary stats data

Message ID 20210617044146.2667540-3-jingzhangos@google.com
State New
Headers show
Series KVM statistics data fd-based binary interface | expand

Commit Message

Jing Zhang June 17, 2021, 4:41 a.m. UTC 
Provides a file descriptor per VM to read VM stats info/data.
Provides a file descriptor per vCPU to read vCPU stats info/data.

The KVM stats now is only accessible by debugfs, which has some
shortcomings this change are supposed to fix:
1. Debugfs is not a stable interface for production and it is
   disabled when kernel Lockdown mode is enabled.
2. Debugfs is organized as "one value per file", it is good for
   debugging, but not supposed to be used for production.
3. Debugfs read/clear in KVM are protected by the global kvm_lock.

Besides that, there are some other benefits with this change:
1. All KVM VM/VCPU stats can be read out in a bulk by one copy
   to userspace.
2. A schema is used to describe KVM statistics. From userspace's
   perspective, the KVM statistics are self-describing.
3. Fd-based solution provides the possibility that a telemetry can
   read KVM stats in a less privileged situation.
4. After the initial setup by reading in stats descriptors, a
   telemetry only needs to read the stats data itself, no more
   parsing or setup is needed.

Reviewed-by: David Matlack <dmatlack@google.com>
Reviewed-by: Ricardo Koller <ricarkol@google.com>
Reviewed-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Reviewed-by: Fuad Tabba <tabba@google.com>
Tested-by: Fuad Tabba <tabba@google.com> #arm64
Signed-off-by: Jing Zhang <jingzhangos@google.com>
---
 arch/arm64/kvm/guest.c    |  38 +++++++++
 arch/mips/kvm/mips.c      |  61 ++++++++++++++
 arch/powerpc/kvm/book3s.c |  64 +++++++++++++++
 arch/powerpc/kvm/booke.c  |  57 +++++++++++++
 arch/s390/kvm/kvm-s390.c  | 130 ++++++++++++++++++++++++++++++
 arch/x86/kvm/x86.c        |  68 ++++++++++++++++
 include/linux/kvm_host.h  | 164 +++++++++++++++++++++++++++++++++++++-
 include/uapi/linux/kvm.h  |  46 +++++++++++
 virt/kvm/kvm_main.c       | 152 ++++++++++++++++++++++++++++++++++-
 9 files changed, 775 insertions(+), 5 deletions(-)

Comments

Greg KH June 17, 2021, 7:03 a.m. UTC | #1 
On Thu, Jun 17, 2021 at 04:41:43AM +0000, Jing Zhang wrote:
> Provides a file descriptor per VM to read VM stats info/data.
> Provides a file descriptor per vCPU to read vCPU stats info/data.
> 
> The KVM stats now is only accessible by debugfs, which has some
> shortcomings this change are supposed to fix:
> 1. Debugfs is not a stable interface for production and it is
>    disabled when kernel Lockdown mode is enabled.

debugfs _could_ be a stable interface if you want it to be and make that
rule for your subsystem.  Disabling it for lockdown mode is a different
issue, and that is a system-wide-policy-decision, not a debugfs-specific
thing.

> 2. Debugfs is organized as "one value per file", it is good for
>    debugging, but not supposed to be used for production.

debugfs IS NOT one-value-per-file, you can do whatever you want in
there.  sysfs IS one-value-per-file, do not get the two confused there.

> 3. Debugfs read/clear in KVM are protected by the global kvm_lock.

That's your implementation issue, not a debugfs issue.

The only "rule" in debugfs is:
	There are no rules.

So while your subsystem might have issues with using debugfs for
statistics like this, that's not debugfs's fault, that's how you want to
use the debugfs files for your subsystem.

> Besides that, there are some other benefits with this change:
> 1. All KVM VM/VCPU stats can be read out in a bulk by one copy
>    to userspace.
> 2. A schema is used to describe KVM statistics. From userspace's
>    perspective, the KVM statistics are self-describing.
> 3. Fd-based solution provides the possibility that a telemetry can
>    read KVM stats in a less privileged situation.

"possiblity"?  Does this work or not?  Have you tested it?

> +static ssize_t kvm_vm_stats_read(struct file *file, char __user *user_buffer,
> +			      size_t size, loff_t *offset)
> +{
> +	struct kvm *kvm = file->private_data;
> +
> +	snprintf(&kvm_vm_stats_header.id[0], sizeof(kvm_vm_stats_header.id),
> +			"kvm-%d", task_pid_nr(current));

Why do you write to this static variable for EVERY read?  Shouldn't you
just do it once at open?  How can it change?

Wait, it's a single shared variable, what happens when multiple tasks
open this thing and read from it?  You race between writing to this
variable here and then:

> +	return kvm_stats_read(&kvm_vm_stats_header, &kvm_vm_stats_desc[0],
> +		&kvm->stat, sizeof(kvm->stat), user_buffer, size, offset);

Accessing it here.

So how is this really working?

thanks,

greg k-h
Greg KH June 17, 2021, 7:24 a.m. UTC | #2 
On Thu, Jun 17, 2021 at 04:41:43AM +0000, Jing Zhang wrote:
> Provides a file descriptor per VM to read VM stats info/data.
> Provides a file descriptor per vCPU to read vCPU stats info/data.

Shouldn't this be two separate patches, one for each thing as these are
two different features being added?

Anyway, an implementation question for both of these:

> +static ssize_t kvm_stats_read(struct _kvm_stats_header *header,
> +		struct _kvm_stats_desc *desc, void *stats, size_t size_stats,
> +		char __user *user_buffer, size_t size, loff_t *offset)
> +{
> +	ssize_t copylen, len, remain = size;

You are "burying" the fact that remain is initialized here, not nice, I
totally missed it when reading this the first time.

This should be:
	ssize_t copylen;
	ssize_t len;
	ssize_t remain = size;
to be obvious.

Remember you will be looking at this code for the next 20 years, make it
easy to read.

> +	size_t size_header, size_desc;
> +	loff_t pos = *offset;
> +	char __user *dest = user_buffer;
> +	void *src;
> +
> +	size_header = sizeof(*header);
> +	size_desc = header->header.count * sizeof(*desc);
> +
> +	len = size_header + size_desc + size_stats - pos;
> +	len = min(len, remain);
> +	if (len <= 0)
> +		return 0;
> +	remain = len;
> +
> +	/* Copy kvm stats header */
> +	copylen = size_header - pos;
> +	copylen = min(copylen, remain);
> +	if (copylen > 0) {
> +		src = (void *)header + pos;
> +		if (copy_to_user(dest, src, copylen))
> +			return -EFAULT;
> +		remain -= copylen;
> +		pos += copylen;
> +		dest += copylen;
> +	}

I thought you said that you would not provide the header for each read,
if you keep reading from the fd.  It looks like you are adding it here
to each read, or is there some "magic" with pos happening here that I do
not understand?

And if there is "magic" with pos, you should document it as it's not
very obvious :)

> +	/* Copy kvm stats descriptors */
> +	copylen = header->header.desc_offset + size_desc - pos;
> +	copylen = min(copylen, remain);
> +	if (copylen > 0) {
> +		src = (void *)desc + pos - header->header.desc_offset;
> +		if (copy_to_user(dest, src, copylen))
> +			return -EFAULT;
> +		remain -= copylen;
> +		pos += copylen;
> +		dest += copylen;
> +	}
> +	/* Copy kvm stats values */

New lines between code blocks of doing things?

And again, why copy the decriptor again?  or is it being skipped
somehow?  Ah, I think I see how it's being skipped, if I look really
closely.  But again, it's not obvious, and I could be wrong.  Please
document this REALLY well.

Write code for the developer first, compiler second.  Again, you are
going to be maintaining it for 20+ years, think of your future self...


> +	copylen = header->header.data_offset + size_stats - pos;
> +	copylen = min(copylen, remain);
> +	if (copylen > 0) {
> +		src = stats + pos - header->header.data_offset;
> +		if (copy_to_user(dest, src, copylen))
> +			return -EFAULT;
> +		remain -= copylen;
> +		pos += copylen;
> +		dest += copylen;
> +	}
> +
> +	*offset = pos;
> +	return len;
> +}

thanks,

greg k-h
Paolo Bonzini June 17, 2021, 11:23 a.m. UTC | #3 
On 17/06/21 09:03, Greg KH wrote:
>> 3. Fd-based solution provides the possibility that a telemetry can
>>     read KVM stats in a less privileged situation.
> "possiblity"?  Does this work or not?  Have you tested it?
> 

I think this is essentially s/that/for/.  But more precisely:

3. Compared for example to a ioctl, a separate file descriptor makes it 
possible for an external program to read statistics, while maintaining 
privilege separation between VMM and telemetry code.

>>
>> +	snprintf(&kvm_vm_stats_header.id[0], sizeof(kvm_vm_stats_header.id),
>> +			"kvm-%d", task_pid_nr(current));
> 
> Why do you write to this static variable for EVERY read?  Shouldn't you
> just do it once at open?  How can it change?
> 
> Wait, it's a single shared variable, what happens when multiple tasks
> open this thing and read from it?  You race between writing to this
> variable here and then:
> 
>> +	return kvm_stats_read(&kvm_vm_stats_header, &kvm_vm_stats_desc[0],
>> +		&kvm->stat, sizeof(kvm->stat), user_buffer, size, offset);
> 
> Accessing it here.
> 
> So how is this really working?

It's not - Jing, kvm_vm_stats_header is small enough that you can store 
a copy in struct kvm.

Paolo
Paolo Bonzini June 17, 2021, 11:27 a.m. UTC | #4 
On 17/06/21 09:24, Greg KH wrote:
>> Provides a file descriptor per VM to read VM stats info/data.
>> Provides a file descriptor per vCPU to read vCPU stats info/data.
> Shouldn't this be two separate patches, one for each thing as these are
> two different features being added?

They share a lot of code.  We could have three patches though:

- add common code for binary statistics file descriptor

- add VM ioctl to retrieve a statistics file descriptor [including the 
definition of VM stats descriptors]

- add VCPU ioctl to retrieve a statistics file descriptor [again 
including the definitions of VCPU stats].

Paolo
Jing Zhang June 17, 2021, 2:48 p.m. UTC | #5 
Hi Greg,

On Thu, Jun 17, 2021 at 2:03 AM Greg KH <greg@kroah.com> wrote:
>
> On Thu, Jun 17, 2021 at 04:41:43AM +0000, Jing Zhang wrote:
> > Provides a file descriptor per VM to read VM stats info/data.
> > Provides a file descriptor per vCPU to read vCPU stats info/data.
> >
> > The KVM stats now is only accessible by debugfs, which has some
> > shortcomings this change are supposed to fix:
> > 1. Debugfs is not a stable interface for production and it is
> >    disabled when kernel Lockdown mode is enabled.
>
> debugfs _could_ be a stable interface if you want it to be and make that
> rule for your subsystem.  Disabling it for lockdown mode is a different
> issue, and that is a system-wide-policy-decision, not a debugfs-specific
> thing.
>
> > 2. Debugfs is organized as "one value per file", it is good for
> >    debugging, but not supposed to be used for production.
>
> debugfs IS NOT one-value-per-file, you can do whatever you want in
> there.  sysfs IS one-value-per-file, do not get the two confused there.
>
> > 3. Debugfs read/clear in KVM are protected by the global kvm_lock.
>
> That's your implementation issue, not a debugfs issue.
>
> The only "rule" in debugfs is:
>         There are no rules.
>
> So while your subsystem might have issues with using debugfs for
> statistics like this, that's not debugfs's fault, that's how you want to
> use the debugfs files for your subsystem.
>
You are right. The issues are from how the debugfs is used in KVM stats.
Will fix the text accordingly.
> > Besides that, there are some other benefits with this change:
> > 1. All KVM VM/VCPU stats can be read out in a bulk by one copy
> >    to userspace.
> > 2. A schema is used to describe KVM statistics. From userspace's
> >    perspective, the KVM statistics are self-describing.
> > 3. Fd-based solution provides the possibility that a telemetry can
> >    read KVM stats in a less privileged situation.
>
> "possiblity"?  Does this work or not?  Have you tested it?
>
I should've said "We are able to read KVM stats in a less privileged process".
> > +static ssize_t kvm_vm_stats_read(struct file *file, char __user *user_buffer,
> > +                           size_t size, loff_t *offset)
> > +{
> > +     struct kvm *kvm = file->private_data;
> > +
> > +     snprintf(&kvm_vm_stats_header.id[0], sizeof(kvm_vm_stats_header.id),
> > +                     "kvm-%d", task_pid_nr(current));
>
> Why do you write to this static variable for EVERY read?  Shouldn't you
> just do it once at open?  How can it change?
>
> Wait, it's a single shared variable, what happens when multiple tasks
> open this thing and read from it?  You race between writing to this
> variable here and then:
>
> > +     return kvm_stats_read(&kvm_vm_stats_header, &kvm_vm_stats_desc[0],
> > +             &kvm->stat, sizeof(kvm->stat), user_buffer, size, offset);
>
> Accessing it here.
>
> So how is this really working?
>
You are right. We only need to do it once at the open. Will fix it according to
Paolo's suggestion.
> thanks,
>
> greg k-h

Thanks,
Jing
Jing Zhang June 17, 2021, 2:56 p.m. UTC | #6 
Hi Greg,

On Thu, Jun 17, 2021 at 2:24 AM Greg KH <gregkh@linuxfoundation.org> wrote:
>
> On Thu, Jun 17, 2021 at 04:41:43AM +0000, Jing Zhang wrote:
> > Provides a file descriptor per VM to read VM stats info/data.
> > Provides a file descriptor per vCPU to read vCPU stats info/data.
>
> Shouldn't this be two separate patches, one for each thing as these are
> two different features being added?
>
Actually, it is really not easy to separate this change into two patches even by
following Paolo's suggestion. And it would be a surprise to userspace to see
only VM stats, no VCPU stats.
I guess it is the text that caused the confusion, I'll change the commit message
for this.
> Anyway, an implementation question for both of these:
>
> > +static ssize_t kvm_stats_read(struct _kvm_stats_header *header,
> > +             struct _kvm_stats_desc *desc, void *stats, size_t size_stats,
> > +             char __user *user_buffer, size_t size, loff_t *offset)
> > +{
> > +     ssize_t copylen, len, remain = size;
>
> You are "burying" the fact that remain is initialized here, not nice, I
> totally missed it when reading this the first time.
>
> This should be:
>         ssize_t copylen;
>         ssize_t len;
>         ssize_t remain = size;
> to be obvious.
>
> Remember you will be looking at this code for the next 20 years, make it
> easy to read.
>
Nice! Will do.
> > +     size_t size_header, size_desc;
> > +     loff_t pos = *offset;
> > +     char __user *dest = user_buffer;
> > +     void *src;
> > +
> > +     size_header = sizeof(*header);
> > +     size_desc = header->header.count * sizeof(*desc);
> > +
> > +     len = size_header + size_desc + size_stats - pos;
> > +     len = min(len, remain);
> > +     if (len <= 0)
> > +             return 0;
> > +     remain = len;
> > +
> > +     /* Copy kvm stats header */
> > +     copylen = size_header - pos;
> > +     copylen = min(copylen, remain);
> > +     if (copylen > 0) {
> > +             src = (void *)header + pos;
> > +             if (copy_to_user(dest, src, copylen))
> > +                     return -EFAULT;
> > +             remain -= copylen;
> > +             pos += copylen;
> > +             dest += copylen;
> > +     }
>
> I thought you said that you would not provide the header for each read,
> if you keep reading from the fd.  It looks like you are adding it here
> to each read, or is there some "magic" with pos happening here that I do
> not understand?
>
> And if there is "magic" with pos, you should document it as it's not
> very obvious :)
>
Will do.
> > +     /* Copy kvm stats descriptors */
> > +     copylen = header->header.desc_offset + size_desc - pos;
> > +     copylen = min(copylen, remain);
> > +     if (copylen > 0) {
> > +             src = (void *)desc + pos - header->header.desc_offset;
> > +             if (copy_to_user(dest, src, copylen))
> > +                     return -EFAULT;
> > +             remain -= copylen;
> > +             pos += copylen;
> > +             dest += copylen;
> > +     }
> > +     /* Copy kvm stats values */
>
> New lines between code blocks of doing things?
>
Will add lines between code blocks.
> And again, why copy the decriptor again?  or is it being skipped
> somehow?  Ah, I think I see how it's being skipped, if I look really
> closely.  But again, it's not obvious, and I could be wrong.  Please
> document this REALLY well.
>
> Write code for the developer first, compiler second.  Again, you are
> going to be maintaining it for 20+ years, think of your future self...
>
>
Sure, will document it.
> > +     copylen = header->header.data_offset + size_stats - pos;
> > +     copylen = min(copylen, remain);
> > +     if (copylen > 0) {
> > +             src = stats + pos - header->header.data_offset;
> > +             if (copy_to_user(dest, src, copylen))
> > +                     return -EFAULT;
> > +             remain -= copylen;
> > +             pos += copylen;
> > +             dest += copylen;
> > +     }
> > +
> > +     *offset = pos;
> > +     return len;
> > +}
>
> thanks,
>
> greg k-h
Thanks,
Jing
Paolo Bonzini June 17, 2021, 3:29 p.m. UTC | #7 
On 17/06/21 16:56, Jing Zhang wrote:
> Actually, it is really not easy to separate this change into two patches even by
> following Paolo's suggestion. And it would be a surprise to userspace to see
> only VM stats, no VCPU stats.

That does not matter.  Having two or three patches is useful because it 
makes review easier; they will never appear separately in a release.

Paolo
Jing Zhang June 17, 2021, 3:42 p.m. UTC | #8 
On Thu, Jun 17, 2021 at 10:29 AM Paolo Bonzini <pbonzini@redhat.com> wrote:
>
> On 17/06/21 16:56, Jing Zhang wrote:
> > Actually, it is really not easy to separate this change into two patches even by
> > following Paolo's suggestion. And it would be a surprise to userspace to see
> > only VM stats, no VCPU stats.
>
> That does not matter.  Having two or three patches is useful because it
> makes review easier; they will never appear separately in a release.
>
> Paolo
>
Cool, then I'll separate them.

Thanks,
Jing
diff mbox series

Patch

diff --git a/arch/arm64/kvm/guest.c b/arch/arm64/kvm/guest.c
index 4962331d01e6..99842e29c2de 100644
--- a/arch/arm64/kvm/guest.c
+++ b/arch/arm64/kvm/guest.c
@@ -28,6 +28,44 @@ 
 
 #include "trace.h"
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS()
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, hvc_exit_stat),
+	STATS_DESC_COUNTER(VCPU, wfe_exit_stat),
+	STATS_DESC_COUNTER(VCPU, wfi_exit_stat),
+	STATS_DESC_COUNTER(VCPU, mmio_exit_user),
+	STATS_DESC_COUNTER(VCPU, mmio_exit_kernel),
+	STATS_DESC_COUNTER(VCPU, exits)
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT_GENERIC("halt_successful_poll", halt_successful_poll),
 	VCPU_STAT_GENERIC("halt_attempted_poll", halt_attempted_poll),
diff --git a/arch/mips/kvm/mips.c b/arch/mips/kvm/mips.c
index ff205b35719b..67404f0947aa 100644
--- a/arch/mips/kvm/mips.c
+++ b/arch/mips/kvm/mips.c
@@ -38,6 +38,67 @@ 
 #define VECTORSPACING 0x100	/* for EI/VI mode */
 #endif
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS()
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, wait_exits),
+	STATS_DESC_COUNTER(VCPU, cache_exits),
+	STATS_DESC_COUNTER(VCPU, signal_exits),
+	STATS_DESC_COUNTER(VCPU, int_exits),
+	STATS_DESC_COUNTER(VCPU, cop_unusable_exits),
+	STATS_DESC_COUNTER(VCPU, tlbmod_exits),
+	STATS_DESC_COUNTER(VCPU, tlbmiss_ld_exits),
+	STATS_DESC_COUNTER(VCPU, tlbmiss_st_exits),
+	STATS_DESC_COUNTER(VCPU, addrerr_st_exits),
+	STATS_DESC_COUNTER(VCPU, addrerr_ld_exits),
+	STATS_DESC_COUNTER(VCPU, syscall_exits),
+	STATS_DESC_COUNTER(VCPU, resvd_inst_exits),
+	STATS_DESC_COUNTER(VCPU, break_inst_exits),
+	STATS_DESC_COUNTER(VCPU, trap_inst_exits),
+	STATS_DESC_COUNTER(VCPU, msa_fpe_exits),
+	STATS_DESC_COUNTER(VCPU, fpe_exits),
+	STATS_DESC_COUNTER(VCPU, msa_disabled_exits),
+	STATS_DESC_COUNTER(VCPU, flush_dcache_exits),
+	STATS_DESC_COUNTER(VCPU, vz_gpsi_exits),
+	STATS_DESC_COUNTER(VCPU, vz_gsfc_exits),
+	STATS_DESC_COUNTER(VCPU, vz_hc_exits),
+	STATS_DESC_COUNTER(VCPU, vz_grr_exits),
+	STATS_DESC_COUNTER(VCPU, vz_gva_exits),
+	STATS_DESC_COUNTER(VCPU, vz_ghfc_exits),
+	STATS_DESC_COUNTER(VCPU, vz_gpa_exits),
+	STATS_DESC_COUNTER(VCPU, vz_resvd_exits),
+#ifdef CONFIG_CPU_LOONGSON64
+	STATS_DESC_COUNTER(VCPU, vz_cpucfg_exits),
+#endif
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT("wait", wait_exits),
 	VCPU_STAT("cache", cache_exits),
diff --git a/arch/powerpc/kvm/book3s.c b/arch/powerpc/kvm/book3s.c
index 92cdb4175945..15436484e521 100644
--- a/arch/powerpc/kvm/book3s.c
+++ b/arch/powerpc/kvm/book3s.c
@@ -38,6 +38,70 @@ 
 
 /* #define EXIT_DEBUG */
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS(),
+	STATS_DESC_ICOUNTER(VM, num_2M_pages),
+	STATS_DESC_ICOUNTER(VM, num_1G_pages)
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, sum_exits),
+	STATS_DESC_COUNTER(VCPU, mmio_exits),
+	STATS_DESC_COUNTER(VCPU, signal_exits),
+	STATS_DESC_COUNTER(VCPU, light_exits),
+	STATS_DESC_COUNTER(VCPU, itlb_real_miss_exits),
+	STATS_DESC_COUNTER(VCPU, itlb_virt_miss_exits),
+	STATS_DESC_COUNTER(VCPU, dtlb_real_miss_exits),
+	STATS_DESC_COUNTER(VCPU, dtlb_virt_miss_exits),
+	STATS_DESC_COUNTER(VCPU, syscall_exits),
+	STATS_DESC_COUNTER(VCPU, isi_exits),
+	STATS_DESC_COUNTER(VCPU, dsi_exits),
+	STATS_DESC_COUNTER(VCPU, emulated_inst_exits),
+	STATS_DESC_COUNTER(VCPU, dec_exits),
+	STATS_DESC_COUNTER(VCPU, ext_intr_exits),
+	STATS_DESC_TIME_NSEC(VCPU, halt_wait_ns),
+	STATS_DESC_COUNTER(VCPU, halt_successful_wait),
+	STATS_DESC_COUNTER(VCPU, dbell_exits),
+	STATS_DESC_COUNTER(VCPU, gdbell_exits),
+	STATS_DESC_COUNTER(VCPU, ld),
+	STATS_DESC_COUNTER(VCPU, st),
+	STATS_DESC_COUNTER(VCPU, pf_storage),
+	STATS_DESC_COUNTER(VCPU, pf_instruc),
+	STATS_DESC_COUNTER(VCPU, sp_storage),
+	STATS_DESC_COUNTER(VCPU, sp_instruc),
+	STATS_DESC_COUNTER(VCPU, queue_intr),
+	STATS_DESC_COUNTER(VCPU, ld_slow),
+	STATS_DESC_COUNTER(VCPU, st_slow),
+	STATS_DESC_COUNTER(VCPU, pthru_all),
+	STATS_DESC_COUNTER(VCPU, pthru_host),
+	STATS_DESC_COUNTER(VCPU, pthru_bad_aff)
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT("exits", sum_exits),
 	VCPU_STAT("mmio", mmio_exits),
diff --git a/arch/powerpc/kvm/booke.c b/arch/powerpc/kvm/booke.c
index 80d3b39aa7ac..fbc9b7b6af57 100644
--- a/arch/powerpc/kvm/booke.c
+++ b/arch/powerpc/kvm/booke.c
@@ -36,6 +36,63 @@ 
 
 unsigned long kvmppc_booke_handlers;
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS(),
+	STATS_DESC_ICOUNTER(VM, num_2M_pages),
+	STATS_DESC_ICOUNTER(VM, num_1G_pages)
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, sum_exits),
+	STATS_DESC_COUNTER(VCPU, mmio_exits),
+	STATS_DESC_COUNTER(VCPU, signal_exits),
+	STATS_DESC_COUNTER(VCPU, light_exits),
+	STATS_DESC_COUNTER(VCPU, itlb_real_miss_exits),
+	STATS_DESC_COUNTER(VCPU, itlb_virt_miss_exits),
+	STATS_DESC_COUNTER(VCPU, dtlb_real_miss_exits),
+	STATS_DESC_COUNTER(VCPU, dtlb_virt_miss_exits),
+	STATS_DESC_COUNTER(VCPU, syscall_exits),
+	STATS_DESC_COUNTER(VCPU, isi_exits),
+	STATS_DESC_COUNTER(VCPU, dsi_exits),
+	STATS_DESC_COUNTER(VCPU, emulated_inst_exits),
+	STATS_DESC_COUNTER(VCPU, dec_exits),
+	STATS_DESC_COUNTER(VCPU, ext_intr_exits),
+	STATS_DESC_TIME_NSEC(VCPU, halt_wait_ns),
+	STATS_DESC_COUNTER(VCPU, halt_successful_wait),
+	STATS_DESC_COUNTER(VCPU, dbell_exits),
+	STATS_DESC_COUNTER(VCPU, gdbell_exits),
+	STATS_DESC_COUNTER(VCPU, ld),
+	STATS_DESC_COUNTER(VCPU, st),
+	STATS_DESC_COUNTER(VCPU, pthru_all),
+	STATS_DESC_COUNTER(VCPU, pthru_host),
+	STATS_DESC_COUNTER(VCPU, pthru_bad_aff)
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT("mmio", mmio_exits),
 	VCPU_STAT("sig", signal_exits),
diff --git a/arch/s390/kvm/kvm-s390.c b/arch/s390/kvm/kvm-s390.c
index e8bc7cd06794..ba50dd775fa8 100644
--- a/arch/s390/kvm/kvm-s390.c
+++ b/arch/s390/kvm/kvm-s390.c
@@ -58,6 +58,136 @@ 
 #define VCPU_IRQS_MAX_BUF (sizeof(struct kvm_s390_irq) * \
 			   (KVM_MAX_VCPUS + LOCAL_IRQS))
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS(),
+	STATS_DESC_COUNTER(VM, inject_io),
+	STATS_DESC_COUNTER(VM, inject_float_mchk),
+	STATS_DESC_COUNTER(VM, inject_pfault_done),
+	STATS_DESC_COUNTER(VM, inject_service_signal),
+	STATS_DESC_COUNTER(VM, inject_virtio)
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, exit_userspace),
+	STATS_DESC_COUNTER(VCPU, exit_null),
+	STATS_DESC_COUNTER(VCPU, exit_external_request),
+	STATS_DESC_COUNTER(VCPU, exit_io_request),
+	STATS_DESC_COUNTER(VCPU, exit_external_interrupt),
+	STATS_DESC_COUNTER(VCPU, exit_stop_request),
+	STATS_DESC_COUNTER(VCPU, exit_validity),
+	STATS_DESC_COUNTER(VCPU, exit_instruction),
+	STATS_DESC_COUNTER(VCPU, exit_pei),
+	STATS_DESC_COUNTER(VCPU, halt_no_poll_steal),
+	STATS_DESC_COUNTER(VCPU, instruction_lctl),
+	STATS_DESC_COUNTER(VCPU, instruction_lctlg),
+	STATS_DESC_COUNTER(VCPU, instruction_stctl),
+	STATS_DESC_COUNTER(VCPU, instruction_stctg),
+	STATS_DESC_COUNTER(VCPU, exit_program_interruption),
+	STATS_DESC_COUNTER(VCPU, exit_instr_and_program),
+	STATS_DESC_COUNTER(VCPU, exit_operation_exception),
+	STATS_DESC_COUNTER(VCPU, deliver_ckc),
+	STATS_DESC_COUNTER(VCPU, deliver_cputm),
+	STATS_DESC_COUNTER(VCPU, deliver_external_call),
+	STATS_DESC_COUNTER(VCPU, deliver_emergency_signal),
+	STATS_DESC_COUNTER(VCPU, deliver_service_signal),
+	STATS_DESC_COUNTER(VCPU, deliver_virtio),
+	STATS_DESC_COUNTER(VCPU, deliver_stop_signal),
+	STATS_DESC_COUNTER(VCPU, deliver_prefix_signal),
+	STATS_DESC_COUNTER(VCPU, deliver_restart_signal),
+	STATS_DESC_COUNTER(VCPU, deliver_program),
+	STATS_DESC_COUNTER(VCPU, deliver_io),
+	STATS_DESC_COUNTER(VCPU, deliver_machine_check),
+	STATS_DESC_COUNTER(VCPU, exit_wait_state),
+	STATS_DESC_COUNTER(VCPU, inject_ckc),
+	STATS_DESC_COUNTER(VCPU, inject_cputm),
+	STATS_DESC_COUNTER(VCPU, inject_external_call),
+	STATS_DESC_COUNTER(VCPU, inject_emergency_signal),
+	STATS_DESC_COUNTER(VCPU, inject_mchk),
+	STATS_DESC_COUNTER(VCPU, inject_pfault_init),
+	STATS_DESC_COUNTER(VCPU, inject_program),
+	STATS_DESC_COUNTER(VCPU, inject_restart),
+	STATS_DESC_COUNTER(VCPU, inject_set_prefix),
+	STATS_DESC_COUNTER(VCPU, inject_stop_signal),
+	STATS_DESC_COUNTER(VCPU, instruction_epsw),
+	STATS_DESC_COUNTER(VCPU, instruction_gs),
+	STATS_DESC_COUNTER(VCPU, instruction_io_other),
+	STATS_DESC_COUNTER(VCPU, instruction_lpsw),
+	STATS_DESC_COUNTER(VCPU, instruction_lpswe),
+	STATS_DESC_COUNTER(VCPU, instruction_pfmf),
+	STATS_DESC_COUNTER(VCPU, instruction_ptff),
+	STATS_DESC_COUNTER(VCPU, instruction_sck),
+	STATS_DESC_COUNTER(VCPU, instruction_sckpf),
+	STATS_DESC_COUNTER(VCPU, instruction_stidp),
+	STATS_DESC_COUNTER(VCPU, instruction_spx),
+	STATS_DESC_COUNTER(VCPU, instruction_stpx),
+	STATS_DESC_COUNTER(VCPU, instruction_stap),
+	STATS_DESC_COUNTER(VCPU, instruction_iske),
+	STATS_DESC_COUNTER(VCPU, instruction_ri),
+	STATS_DESC_COUNTER(VCPU, instruction_rrbe),
+	STATS_DESC_COUNTER(VCPU, instruction_sske),
+	STATS_DESC_COUNTER(VCPU, instruction_ipte_interlock),
+	STATS_DESC_COUNTER(VCPU, instruction_stsi),
+	STATS_DESC_COUNTER(VCPU, instruction_stfl),
+	STATS_DESC_COUNTER(VCPU, instruction_tb),
+	STATS_DESC_COUNTER(VCPU, instruction_tpi),
+	STATS_DESC_COUNTER(VCPU, instruction_tprot),
+	STATS_DESC_COUNTER(VCPU, instruction_tsch),
+	STATS_DESC_COUNTER(VCPU, instruction_sie),
+	STATS_DESC_COUNTER(VCPU, instruction_essa),
+	STATS_DESC_COUNTER(VCPU, instruction_sthyi),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_sense),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_sense_running),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_external_call),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_emergency),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_cond_emergency),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_start),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_stop),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_stop_store_status),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_store_status),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_store_adtl_status),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_arch),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_prefix),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_restart),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_init_cpu_reset),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_cpu_reset),
+	STATS_DESC_COUNTER(VCPU, instruction_sigp_unknown),
+	STATS_DESC_COUNTER(VCPU, diagnose_10),
+	STATS_DESC_COUNTER(VCPU, diagnose_44),
+	STATS_DESC_COUNTER(VCPU, diagnose_9c),
+	STATS_DESC_COUNTER(VCPU, diagnose_9c_ignored),
+	STATS_DESC_COUNTER(VCPU, diagnose_9c_forward),
+	STATS_DESC_COUNTER(VCPU, diagnose_258),
+	STATS_DESC_COUNTER(VCPU, diagnose_308),
+	STATS_DESC_COUNTER(VCPU, diagnose_500),
+	STATS_DESC_COUNTER(VCPU, diagnose_other),
+	STATS_DESC_COUNTER(VCPU, pfault_sync)
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT("userspace_handled", exit_userspace),
 	VCPU_STAT("exit_null", exit_null),
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 157212157aee..84438573b529 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -223,6 +223,74 @@  EXPORT_SYMBOL_GPL(host_xss);
 u64 __read_mostly supported_xss;
 EXPORT_SYMBOL_GPL(supported_xss);
 
+struct _kvm_stats_desc kvm_vm_stats_desc[] = {
+	KVM_GENERIC_VM_STATS(),
+	STATS_DESC_COUNTER(VM, mmu_shadow_zapped),
+	STATS_DESC_COUNTER(VM, mmu_pte_write),
+	STATS_DESC_COUNTER(VM, mmu_pde_zapped),
+	STATS_DESC_COUNTER(VM, mmu_flooded),
+	STATS_DESC_COUNTER(VM, mmu_recycled),
+	STATS_DESC_COUNTER(VM, mmu_cache_miss),
+	STATS_DESC_ICOUNTER(VM, mmu_unsync),
+	STATS_DESC_ICOUNTER(VM, lpages),
+	STATS_DESC_ICOUNTER(VM, nx_lpage_splits),
+	STATS_DESC_ICOUNTER(VM, max_mmu_page_hash_collisions)
+};
+static_assert(ARRAY_SIZE(kvm_vm_stats_desc) ==
+		sizeof(struct kvm_vm_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vm_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vm_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vm_stats_desc),
+	}
+};
+
+struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
+	KVM_GENERIC_VCPU_STATS(),
+	STATS_DESC_COUNTER(VCPU, pf_fixed),
+	STATS_DESC_COUNTER(VCPU, pf_guest),
+	STATS_DESC_COUNTER(VCPU, tlb_flush),
+	STATS_DESC_COUNTER(VCPU, invlpg),
+	STATS_DESC_COUNTER(VCPU, exits),
+	STATS_DESC_COUNTER(VCPU, io_exits),
+	STATS_DESC_COUNTER(VCPU, mmio_exits),
+	STATS_DESC_COUNTER(VCPU, signal_exits),
+	STATS_DESC_COUNTER(VCPU, irq_window_exits),
+	STATS_DESC_COUNTER(VCPU, nmi_window_exits),
+	STATS_DESC_COUNTER(VCPU, l1d_flush),
+	STATS_DESC_COUNTER(VCPU, halt_exits),
+	STATS_DESC_COUNTER(VCPU, request_irq_exits),
+	STATS_DESC_COUNTER(VCPU, irq_exits),
+	STATS_DESC_COUNTER(VCPU, host_state_reload),
+	STATS_DESC_COUNTER(VCPU, fpu_reload),
+	STATS_DESC_COUNTER(VCPU, insn_emulation),
+	STATS_DESC_COUNTER(VCPU, insn_emulation_fail),
+	STATS_DESC_COUNTER(VCPU, hypercalls),
+	STATS_DESC_COUNTER(VCPU, irq_injections),
+	STATS_DESC_COUNTER(VCPU, nmi_injections),
+	STATS_DESC_COUNTER(VCPU, req_event),
+	STATS_DESC_COUNTER(VCPU, nested_run),
+	STATS_DESC_COUNTER(VCPU, directed_yield_attempted),
+	STATS_DESC_COUNTER(VCPU, directed_yield_successful),
+	STATS_DESC_ICOUNTER(VCPU, guest_mode)
+};
+static_assert(ARRAY_SIZE(kvm_vcpu_stats_desc) ==
+		sizeof(struct kvm_vcpu_stat) / sizeof(u64));
+
+struct _kvm_stats_header kvm_vcpu_stats_header = {
+	{
+		.name_size = KVM_STATS_NAME_LEN,
+		.count = ARRAY_SIZE(kvm_vcpu_stats_desc),
+		.desc_offset = sizeof(struct _kvm_stats_header),
+		.data_offset = sizeof(struct _kvm_stats_header) +
+			sizeof(kvm_vcpu_stats_desc),
+	}
+};
+
 struct kvm_stats_debugfs_item debugfs_entries[] = {
 	VCPU_STAT("pf_fixed", pf_fixed),
 	VCPU_STAT("pf_guest", pf_guest),
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 5a31e0696360..a5a32825c83f 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -305,7 +305,6 @@  struct kvm_vcpu {
 	struct pid __rcu *pid;
 	int sigset_active;
 	sigset_t sigset;
-	struct kvm_vcpu_stat stat;
 	unsigned int halt_poll_ns;
 	bool valid_wakeup;
 
@@ -342,6 +341,7 @@  struct kvm_vcpu {
 	bool preempted;
 	bool ready;
 	struct kvm_vcpu_arch arch;
+	struct kvm_vcpu_stat stat;
 	struct kvm_dirty_ring dirty_ring;
 };
 
@@ -1272,6 +1272,17 @@  struct kvm_stats_debugfs_item {
 	int mode;
 };
 
+struct _kvm_stats_header {
+	struct kvm_stats_header header;
+	char id[KVM_STATS_ID_MAXLEN];
+};
+
+#define KVM_STATS_NAME_LEN	48
+struct _kvm_stats_desc {
+	struct kvm_stats_desc desc;
+	char name[KVM_STATS_NAME_LEN];
+};
+
 #define KVM_DBGFS_GET_MODE(dbgfs_item)                                         \
 	((dbgfs_item)->mode ? (dbgfs_item)->mode : 0644)
 
@@ -1285,8 +1296,159 @@  struct kvm_stats_debugfs_item {
 	{ n, offsetof(struct kvm_vcpu, stat.generic.x),			       \
 	  KVM_STAT_VCPU, ## __VA_ARGS__ }
 
+#define STATS_DESC_COMMON(type, unit, base, exp)			       \
+	.flags = type | unit | base |					       \
+	    BUILD_BUG_ON_ZERO(type & ~KVM_STATS_TYPE_MASK) |		       \
+	    BUILD_BUG_ON_ZERO(unit & ~KVM_STATS_UNIT_MASK) |		       \
+	    BUILD_BUG_ON_ZERO(base & ~KVM_STATS_BASE_MASK),		       \
+	.exponent = exp,						       \
+	.size = 1
+
+#define VM_GENERIC_STATS_DESC(stat, type, unit, base, exp)		       \
+	{								       \
+		{							       \
+			STATS_DESC_COMMON(type, unit, base, exp),	       \
+			.offset = offsetof(struct kvm_vm_stat, generic.stat)   \
+		},							       \
+		.name = #stat,						       \
+	}
+#define VCPU_GENERIC_STATS_DESC(stat, type, unit, base, exp)		       \
+	{								       \
+		{							       \
+			STATS_DESC_COMMON(type, unit, base, exp),	       \
+			.offset = offsetof(struct kvm_vcpu_stat, generic.stat) \
+		},							       \
+		.name = #stat,						       \
+	}
+#define VM_STATS_DESC(stat, type, unit, base, exp)			       \
+	{								       \
+		{							       \
+			STATS_DESC_COMMON(type, unit, base, exp),	       \
+			.offset = offsetof(struct kvm_vm_stat, stat)	       \
+		},							       \
+		.name = #stat,						       \
+	}
+#define VCPU_STATS_DESC(stat, type, unit, base, exp)			       \
+	{								       \
+		{							       \
+			STATS_DESC_COMMON(type, unit, base, exp),	       \
+			.offset = offsetof(struct kvm_vcpu_stat, stat)	       \
+		},							       \
+		.name = #stat,						       \
+	}
+/* SCOPE: VM, VM_GENERIC, VCPU, VCPU_GENERIC */
+#define STATS_DESC(SCOPE, stat, type, unit, base, exp)			       \
+	SCOPE##_STATS_DESC(stat, type, unit, base, exp)
+
+#define STATS_DESC_CUMULATIVE(SCOPE, name, unit, base, exponent)	       \
+	STATS_DESC(SCOPE, name, KVM_STATS_TYPE_CUMULATIVE,		       \
+		      unit, base, exponent)
+#define STATS_DESC_INSTANT(SCOPE, name, unit, base, exponent)		       \
+	STATS_DESC(SCOPE, name, KVM_STATS_TYPE_INSTANT, unit, base, exponent)  \
+
+/* Cumulative counter */
+#define STATS_DESC_COUNTER(SCOPE, name)					       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_NONE,		       \
+		KVM_STATS_BASE_POW10, 0)
+/* Instantaneous counter */
+#define STATS_DESC_ICOUNTER(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_NONE,		       \
+		KVM_STATS_BASE_POW10, 0)
+
+/* Cumulative clock cycles */
+#define STATS_DESC_CYCLE(SCOPE, name)					       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_CYCLES,	       \
+		KVM_STATS_BASE_POW10, 0)
+/* Instantaneous clock cycles */
+#define STATS_DESC_ICYCLE(SCOPE, name)					       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_CYCLES,		       \
+		KVM_STATS_BASE_POW10, 0)
+
+/* Cumulative memory size in Byte */
+#define STATS_DESC_SIZE_BYTE(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_BYTES,	       \
+		KVM_STATS_BASE_POW2, 0)
+/* Cumulative memory size in KiByte */
+#define STATS_DESC_SIZE_KBYTE(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_BYTES,	       \
+		KVM_STATS_BASE_POW2, 10)
+/* Cumulative memory size in MiByte */
+#define STATS_DESC_SIZE_MBYTE(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_BYTES,	       \
+		KVM_STATS_BASE_POW2, 20)
+/* Cumulative memory size in GiByte */
+#define STATS_DESC_SIZE_GBYTE(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_BYTES,	       \
+		KVM_STATS_BASE_POW2, 30)
+
+/* Instantaneous memory size in Byte */
+#define STATS_DESC_ISIZE_BYTE(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_BYTES,		       \
+		KVM_STATS_BASE_POW2, 0)
+/* Instantaneous memory size in KiByte */
+#define STATS_DESC_ISIZE_KBYTE(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_BYTES,		       \
+		KVM_STATS_BASE_POW2, 10)
+/* Instantaneous memory size in MiByte */
+#define STATS_DESC_ISIZE_MBYTE(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_BYTES,		       \
+		KVM_STATS_BASE_POW2, 20)
+/* Instantaneous memory size in GiByte */
+#define STATS_DESC_ISIZE_GBYTE(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_BYTES,		       \
+		KVM_STATS_BASE_POW2, 30)
+
+/* Cumulative time in second */
+#define STATS_DESC_TIME_SEC(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_SECONDS,	       \
+		KVM_STATS_BASE_POW10, 0)
+/* Cumulative time in millisecond */
+#define STATS_DESC_TIME_MSEC(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_SECONDS,	       \
+		KVM_STATS_BASE_POW10, -3)
+/* Cumulative time in microsecond */
+#define STATS_DESC_TIME_USEC(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_SECONDS,	       \
+		KVM_STATS_BASE_POW10, -6)
+/* Cumulative time in nanosecond */
+#define STATS_DESC_TIME_NSEC(SCOPE, name)				       \
+	STATS_DESC_CUMULATIVE(SCOPE, name, KVM_STATS_UNIT_SECONDS,	       \
+		KVM_STATS_BASE_POW10, -9)
+
+/* Instantaneous time in second */
+#define STATS_DESC_ITIME_SEC(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_SECONDS,		       \
+		KVM_STATS_BASE_POW10, 0)
+/* Instantaneous time in millisecond */
+#define STATS_DESC_ITIME_MSEC(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_SECONDS,		       \
+		KVM_STATS_BASE_POW10, -3)
+/* Instantaneous time in microsecond */
+#define STATS_DESC_ITIME_USEC(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_SECONDS,		       \
+		KVM_STATS_BASE_POW10, -6)
+/* Instantaneous time in nanosecond */
+#define STATS_DESC_ITIME_NSEC(SCOPE, name)				       \
+	STATS_DESC_INSTANT(SCOPE, name, KVM_STATS_UNIT_SECONDS,		       \
+		KVM_STATS_BASE_POW10, -9)
+
+#define KVM_GENERIC_VM_STATS()						       \
+	STATS_DESC_COUNTER(VM_GENERIC, remote_tlb_flush)
+
+#define KVM_GENERIC_VCPU_STATS()					       \
+	STATS_DESC_COUNTER(VCPU_GENERIC, halt_successful_poll),		       \
+	STATS_DESC_COUNTER(VCPU_GENERIC, halt_attempted_poll),		       \
+	STATS_DESC_COUNTER(VCPU_GENERIC, halt_poll_invalid),		       \
+	STATS_DESC_COUNTER(VCPU_GENERIC, halt_wakeup),			       \
+	STATS_DESC_TIME_NSEC(VCPU_GENERIC, halt_poll_success_ns),	       \
+	STATS_DESC_TIME_NSEC(VCPU_GENERIC, halt_poll_fail_ns)
+
 extern struct kvm_stats_debugfs_item debugfs_entries[];
 extern struct dentry *kvm_debugfs_dir;
+extern struct _kvm_stats_header kvm_vm_stats_header;
+extern struct _kvm_stats_header kvm_vcpu_stats_header;
+extern struct _kvm_stats_desc kvm_vm_stats_desc[];
+extern struct _kvm_stats_desc kvm_vcpu_stats_desc[];
 
 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
 static inline int mmu_notifier_retry(struct kvm *kvm, unsigned long mmu_seq)
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index 9febe1412f7a..d6e97b577d01 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -1086,6 +1086,7 @@  struct kvm_ppc_resize_hpt {
 #define KVM_CAP_HYPERV_ENFORCE_CPUID 199
 #define KVM_CAP_SREGS2 200
 #define KVM_CAP_EXIT_HYPERCALL 201
+#define KVM_CAP_BINARY_STATS_FD 202
 
 #ifdef KVM_CAP_IRQ_ROUTING
 
@@ -1905,4 +1906,49 @@  struct kvm_dirty_gfn {
 #define KVM_BUS_LOCK_DETECTION_OFF             (1 << 0)
 #define KVM_BUS_LOCK_DETECTION_EXIT            (1 << 1)
 
+#define KVM_STATS_ID_MAXLEN		64
+
+struct kvm_stats_header {
+	__u32 name_size;
+	__u32 count;
+	__u32 desc_offset;
+	__u32 data_offset;
+	char id[0];
+};
+
+#define KVM_STATS_TYPE_SHIFT		0
+#define KVM_STATS_TYPE_MASK		(0xF << KVM_STATS_TYPE_SHIFT)
+#define KVM_STATS_TYPE_CUMULATIVE	(0x0 << KVM_STATS_TYPE_SHIFT)
+#define KVM_STATS_TYPE_INSTANT		(0x1 << KVM_STATS_TYPE_SHIFT)
+#define KVM_STATS_TYPE_MAX		KVM_STATS_TYPE_INSTANT
+
+#define KVM_STATS_UNIT_SHIFT		4
+#define KVM_STATS_UNIT_MASK		(0xF << KVM_STATS_UNIT_SHIFT)
+#define KVM_STATS_UNIT_NONE		(0x0 << KVM_STATS_UNIT_SHIFT)
+#define KVM_STATS_UNIT_BYTES		(0x1 << KVM_STATS_UNIT_SHIFT)
+#define KVM_STATS_UNIT_SECONDS		(0x2 << KVM_STATS_UNIT_SHIFT)
+#define KVM_STATS_UNIT_CYCLES		(0x3 << KVM_STATS_UNIT_SHIFT)
+#define KVM_STATS_UNIT_MAX		KVM_STATS_UNIT_CYCLES
+
+#define KVM_STATS_BASE_SHIFT		8
+#define KVM_STATS_BASE_MASK		(0xF << KVM_STATS_BASE_SHIFT)
+#define KVM_STATS_BASE_POW10		(0x0 << KVM_STATS_BASE_SHIFT)
+#define KVM_STATS_BASE_POW2		(0x1 << KVM_STATS_BASE_SHIFT)
+#define KVM_STATS_BASE_MAX		KVM_STATS_BASE_POW2
+
+struct kvm_stats_desc {
+	__u32 flags;
+	__s16 exponent;
+	__u16 size;
+	__u32 offset;
+	__u32 unused;
+	char name[0];
+};
+
+struct kvm_stats_data {
+	__u64 value[0];
+};
+
+#define KVM_GET_STATS_FD  _IO(KVMIO,  0xce)
+
 #endif /* __LINUX_KVM_H */
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index cec986487b30..0ad37c350198 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -3497,6 +3497,103 @@  static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
 	return 0;
 }
 
+static ssize_t kvm_stats_read(struct _kvm_stats_header *header,
+		struct _kvm_stats_desc *desc, void *stats, size_t size_stats,
+		char __user *user_buffer, size_t size, loff_t *offset)
+{
+	ssize_t copylen, len, remain = size;
+	size_t size_header, size_desc;
+	loff_t pos = *offset;
+	char __user *dest = user_buffer;
+	void *src;
+
+	size_header = sizeof(*header);
+	size_desc = header->header.count * sizeof(*desc);
+
+	len = size_header + size_desc + size_stats - pos;
+	len = min(len, remain);
+	if (len <= 0)
+		return 0;
+	remain = len;
+
+	/* Copy kvm stats header */
+	copylen = size_header - pos;
+	copylen = min(copylen, remain);
+	if (copylen > 0) {
+		src = (void *)header + pos;
+		if (copy_to_user(dest, src, copylen))
+			return -EFAULT;
+		remain -= copylen;
+		pos += copylen;
+		dest += copylen;
+	}
+	/* Copy kvm stats descriptors */
+	copylen = header->header.desc_offset + size_desc - pos;
+	copylen = min(copylen, remain);
+	if (copylen > 0) {
+		src = (void *)desc + pos - header->header.desc_offset;
+		if (copy_to_user(dest, src, copylen))
+			return -EFAULT;
+		remain -= copylen;
+		pos += copylen;
+		dest += copylen;
+	}
+	/* Copy kvm stats values */
+	copylen = header->header.data_offset + size_stats - pos;
+	copylen = min(copylen, remain);
+	if (copylen > 0) {
+		src = stats + pos - header->header.data_offset;
+		if (copy_to_user(dest, src, copylen))
+			return -EFAULT;
+		remain -= copylen;
+		pos += copylen;
+		dest += copylen;
+	}
+
+	*offset = pos;
+	return len;
+}
+
+static ssize_t kvm_vcpu_stats_read(struct file *file, char __user *user_buffer,
+			      size_t size, loff_t *offset)
+{
+	struct kvm_vcpu *vcpu = file->private_data;
+
+	snprintf(&kvm_vcpu_stats_header.id[0], sizeof(kvm_vcpu_stats_header.id),
+			"kvm-%d/vcpu-%d", task_pid_nr(current), vcpu->vcpu_id);
+	return kvm_stats_read(&kvm_vcpu_stats_header,
+			&kvm_vcpu_stats_desc[0], &vcpu->stat,
+			sizeof(vcpu->stat), user_buffer, size, offset);
+}
+
+static const struct file_operations kvm_vcpu_stats_fops = {
+	.read = kvm_vcpu_stats_read,
+	.llseek = noop_llseek,
+};
+
+static int kvm_vcpu_ioctl_get_stats_fd(struct kvm_vcpu *vcpu)
+{
+	int fd;
+	struct file *file;
+	char name[15 + ITOA_MAX_LEN + 1];
+
+	snprintf(name, sizeof(name), "kvm-vcpu-stats:%d", vcpu->vcpu_id);
+
+	fd = get_unused_fd_flags(O_CLOEXEC);
+	if (fd < 0)
+		return fd;
+
+	file = anon_inode_getfile(name, &kvm_vcpu_stats_fops, vcpu, O_RDONLY);
+	if (IS_ERR(file)) {
+		put_unused_fd(fd);
+		return PTR_ERR(file);
+	}
+	file->f_mode |= FMODE_PREAD;
+	fd_install(fd, file);
+
+	return fd;
+}
+
 static long kvm_vcpu_ioctl(struct file *filp,
 			   unsigned int ioctl, unsigned long arg)
 {
@@ -3694,6 +3791,10 @@  static long kvm_vcpu_ioctl(struct file *filp,
 		r = kvm_arch_vcpu_ioctl_set_fpu(vcpu, fpu);
 		break;
 	}
+	case KVM_GET_STATS_FD: {
+		r = kvm_vcpu_ioctl_get_stats_fd(vcpu);
+		break;
+	}
 	default:
 		r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
 	}
@@ -3952,6 +4053,8 @@  static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
 #else
 		return 0;
 #endif
+	case KVM_CAP_BINARY_STATS_FD:
+		return 1;
 	default:
 		break;
 	}
@@ -4055,6 +4158,43 @@  static int kvm_vm_ioctl_enable_cap_generic(struct kvm *kvm,
 	}
 }
 
+static ssize_t kvm_vm_stats_read(struct file *file, char __user *user_buffer,
+			      size_t size, loff_t *offset)
+{
+	struct kvm *kvm = file->private_data;
+
+	snprintf(&kvm_vm_stats_header.id[0], sizeof(kvm_vm_stats_header.id),
+			"kvm-%d", task_pid_nr(current));
+	return kvm_stats_read(&kvm_vm_stats_header, &kvm_vm_stats_desc[0],
+		&kvm->stat, sizeof(kvm->stat), user_buffer, size, offset);
+}
+
+static const struct file_operations kvm_vm_stats_fops = {
+	.read = kvm_vm_stats_read,
+	.llseek = noop_llseek,
+};
+
+static int kvm_vm_ioctl_get_stats_fd(struct kvm *kvm)
+{
+	int fd;
+	struct file *file;
+
+	fd = get_unused_fd_flags(O_CLOEXEC);
+	if (fd < 0)
+		return fd;
+
+	file = anon_inode_getfile("kvm-vm-stats",
+			&kvm_vm_stats_fops, kvm, O_RDONLY);
+	if (IS_ERR(file)) {
+		put_unused_fd(fd);
+		return PTR_ERR(file);
+	}
+	file->f_mode |= FMODE_PREAD;
+	fd_install(fd, file);
+
+	return fd;
+}
+
 static long kvm_vm_ioctl(struct file *filp,
 			   unsigned int ioctl, unsigned long arg)
 {
@@ -4237,6 +4377,9 @@  static long kvm_vm_ioctl(struct file *filp,
 	case KVM_RESET_DIRTY_RINGS:
 		r = kvm_vm_ioctl_reset_dirty_pages(kvm);
 		break;
+	case KVM_GET_STATS_FD:
+		r = kvm_vm_ioctl_get_stats_fd(kvm);
+		break;
 	default:
 		r = kvm_arch_vm_ioctl(filp, ioctl, arg);
 	}
@@ -5210,10 +5353,11 @@  int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
 		vcpu_align = __alignof__(struct kvm_vcpu);
 	kvm_vcpu_cache =
 		kmem_cache_create_usercopy("kvm_vcpu", vcpu_size, vcpu_align,
-					   SLAB_ACCOUNT,
-					   offsetof(struct kvm_vcpu, arch),
-					   sizeof_field(struct kvm_vcpu, arch),
-					   NULL);
+				   SLAB_ACCOUNT,
+				   offsetof(struct kvm_vcpu, arch),
+				   sizeof_field(struct kvm_vcpu, arch)
+				   + sizeof_field(struct kvm_vcpu, stat),
+				   NULL);
 	if (!kvm_vcpu_cache) {
 		r = -ENOMEM;
 		goto out_free_3;