KVM: x86/xen: Make number of event channels defines less magical

  Use BITS_PER_BYTE and sizeof_field() to compute the number of Xen event
channels.  The compat version at least uses sizeof_field(), but the
regular version open codes sizeof_field(), BITS_PER_BYTE, and combines
literals in the process, which makes it far too difficult to understand
relatively straightforward code.

No functional change intended.

Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Paul Durrant <paul@xen.org>
Signed-off-by: Sean Christopherson <seanjc@google.com>
---
 arch/x86/kvm/xen.h                    | 8 +++++---
 include/xen/interface/event_channel.h | 6 ++++--
 2 files changed, 9 insertions(+), 5 deletions(-)


base-commit: d663b8a285986072428a6a145e5994bc275df994
  

On Mon, 2022-11-14 at 18:16 +0000, Sean Christopherson wrote:
> Use BITS_PER_BYTE and sizeof_field() to compute the number of Xen event
> channels.  The compat version at least uses sizeof_field(), but the
> regular version open codes sizeof_field(), BITS_PER_BYTE, and combines
> literals in the process, which makes it far too difficult to understand
> relatively straightforward code.
> 
> No functional change intended.

Slightly dubious about changing the regular one, since that's just
imported directly from Xen public header files.
  

On Mon, Nov 14, 2022, David Woodhouse wrote:
> On Mon, 2022-11-14 at 18:16 +0000, Sean Christopherson wrote:
> > Use BITS_PER_BYTE and sizeof_field() to compute the number of Xen event
> > channels.  The compat version at least uses sizeof_field(), but the
> > regular version open codes sizeof_field(), BITS_PER_BYTE, and combines
> > literals in the process, which makes it far too difficult to understand
> > relatively straightforward code.
> > 
> > No functional change intended.
> 
> Slightly dubious about changing the regular one, since that's just
> imported directly from Xen public header files.

Ugh.  I worried that might be the case.  An alternative approach to help document
things from a KVM perspective would be something like:

diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index 93c628d3e3a9..7769f3b98af0 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -1300,6 +1300,9 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
 
 static inline int max_evtchn_port(struct kvm *kvm)
 {
+       BUILD_BUG_ON(EVTCHN_2L_NR_CHANNELS !=
+                    (sizeof_field(struct shared_info, evtchn_pending) * BITS_PER_BYTE));
+
        if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
                return EVTCHN_2L_NR_CHANNELS;
        else
  

On Mon, 2022-11-14 at 19:39 +0000, Sean Christopherson wrote:
> Ugh.  I worried that might be the case.  An alternative approach to help document
> things from a KVM perspective would be something like:
> 
> diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> index 93c628d3e3a9..7769f3b98af0 100644
> --- a/arch/x86/kvm/xen.c
> +++ b/arch/x86/kvm/xen.c
> @@ -1300,6 +1300,9 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
>  
>  static inline int max_evtchn_port(struct kvm *kvm)
>  {
> +       BUILD_BUG_ON(EVTCHN_2L_NR_CHANNELS !=
> +                    (sizeof_field(struct shared_info, evtchn_pending) * BITS_PER_BYTE));
> +
>         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
>                 return EVTCHN_2L_NR_CHANNELS;
>         else

Not really sure I see the point of that.

There are two main reasons for that kind of BUILD_BUG_ON(). I've added
a few of them asserting that the size of the structure and its compat
variant are identical, and thus documenting *why* the code lacks compat
handling. For example...

	/*
	 * Next, write the new runstate. This is in the *same* place
	 * for 32-bit and 64-bit guests, asserted here for paranoia.
	 */
	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
		     offsetof(struct compat_vcpu_runstate_info, state));

The second reason is to prevent accidental screwups where our local
definition of a structure varies from the official ABI. Like these:

	/* Paranoia checks on the 32-bit struct layout */
	BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
	BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
	BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);

I don't really see the above fulfilling either of those use cases.
Given that the definition of the evtchn_pending field is:

        xen_ulong_t evtchn_pending[sizeof(xen_ulong_t) * 8];

It's fairly tautological that the number of event channels supported is
BITS_PER_ULONG * BITS_PER_ULONG. Which is sizeof(xen_ulong_t)² * 64 as
defined in the official Xen headers.

I don't know that we really need to add our own sanity check on the
headers we imported from Xen. It doesn't seem to add much.

No objection to cleaning up the COMPAT one though, as in your original
patch.
  

EVTCHN_2L_NR_CHANNELSOn Mon, Nov 14, 2022, David Woodhouse wrote:
> On Mon, 2022-11-14 at 19:39 +0000, Sean Christopherson wrote:
> > Ugh.  I worried that might be the case.  An alternative approach to help document
> > things from a KVM perspective would be something like:
> > 
> > diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> > index 93c628d3e3a9..7769f3b98af0 100644
> > --- a/arch/x86/kvm/xen.c
> > +++ b/arch/x86/kvm/xen.c
> > @@ -1300,6 +1300,9 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
> >  
> >  static inline int max_evtchn_port(struct kvm *kvm)
> >  {
> > +       BUILD_BUG_ON(EVTCHN_2L_NR_CHANNELS !=
> > +                    (sizeof_field(struct shared_info, evtchn_pending) * BITS_PER_BYTE));
> > +
> >         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
> >                 return EVTCHN_2L_NR_CHANNELS;
> >         else
> 
> Not really sure I see the point of that.
> 
> There are two main reasons for that kind of BUILD_BUG_ON(). I've added
> a few of them asserting that the size of the structure and its compat
> variant are identical, and thus documenting *why* the code lacks compat
> handling. For example...
> 
> 	/*
> 	 * Next, write the new runstate. This is in the *same* place
> 	 * for 32-bit and 64-bit guests, asserted here for paranoia.
> 	 */
> 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
> 		     offsetof(struct compat_vcpu_runstate_info, state));
> 
> The second reason is to prevent accidental screwups where our local
> definition of a structure varies from the official ABI. Like these:
> 
> 	/* Paranoia checks on the 32-bit struct layout */
> 	BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
> 	BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
> 	BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
> 
> I don't really see the above fulfilling either of those use cases.
>
> Given that the definition of the evtchn_pending field is:
> 
>         xen_ulong_t evtchn_pending[sizeof(xen_ulong_t) * 8];
> 
> It's fairly tautological that the number of event channels supported is
> BITS_PER_ULONG * BITS_PER_ULONG. Which is sizeof(xen_ulong_t)² * 64 as
> defined in the official Xen headers.
> 
> I don't know that we really need to add our own sanity check on the
> headers we imported from Xen. It doesn't seem to add much.

The goal isn't to add a sanity check, it's to document what EVTCHN_2L_NR_CHANNELS
actually represents.  My frustration with 

  sizeof(xen_ulong_t) * sizeof(xen_ulong_t) * 64

is that there's nothing there that connects it back to evtchn_pending or evtchn_mask.

E.g. ideally the code would be something like

  #define COMPAT_EVTCHN_2L_NR_CHANNELS	256

  #ifdef CONFIG_X86_64
  #define EVTCHN_2L_NR_CHANNELS		512
  #else
  #define EVTCHN_2L_NR_CHANNELS		COMPAT_EVTCHN_2L_NR_CHANNELS


  DECLARE_BITMAP(evtchn_pending, EVTCHN_2L_NR_CHANNELS);
  DECLARE_BITMAP(evtchn_mask, EVTCHN_2L_NR_CHANNELS);

which is much more self-documenting and doesn't require the reader to do math to
grok the basics.

Anyways, we can drop this patch, it was written mostly out of frustration with
how long it took me to understand what is actually a very simple concept that's
written in an unnecessarily obscure way.
  

On Tue, 2022-11-22 at 20:31 +0000, Sean Christopherson wrote:
> EVTCHN_2L_NR_CHANNELSOn Mon, Nov 14, 2022, David Woodhouse wrote:
> > On Mon, 2022-11-14 at 19:39 +0000, Sean Christopherson wrote:
> > > Ugh.  I worried that might be the case.  An alternative approach to help document
> > > things from a KVM perspective would be something like:
> > > 
> > > diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
> > > index 93c628d3e3a9..7769f3b98af0 100644
> > > --- a/arch/x86/kvm/xen.c
> > > +++ b/arch/x86/kvm/xen.c
> > > @@ -1300,6 +1300,9 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
> > >  
> > >  static inline int max_evtchn_port(struct kvm *kvm)
> > >  {
> > > +       BUILD_BUG_ON(EVTCHN_2L_NR_CHANNELS !=
> > > +                    (sizeof_field(struct shared_info, evtchn_pending) * BITS_PER_BYTE));
> > > +
> > >         if (IS_ENABLED(CONFIG_64BIT) && kvm->arch.xen.long_mode)
> > >                 return EVTCHN_2L_NR_CHANNELS;
> > >         else
> > 
> > Not really sure I see the point of that.
> > 
> > There are two main reasons for that kind of BUILD_BUG_ON(). I've added
> > a few of them asserting that the size of the structure and its compat
> > variant are identical, and thus documenting *why* the code lacks compat
> > handling. For example...
> > 
> > 	/*
> > 	 * Next, write the new runstate. This is in the *same* place
> > 	 * for 32-bit and 64-bit guests, asserted here for paranoia.
> > 	 */
> > 	BUILD_BUG_ON(offsetof(struct vcpu_runstate_info, state) !=
> > 		     offsetof(struct compat_vcpu_runstate_info, state));
> > 
> > The second reason is to prevent accidental screwups where our local
> > definition of a structure varies from the official ABI. Like these:
> > 
> > 	/* Paranoia checks on the 32-bit struct layout */
> > 	BUILD_BUG_ON(offsetof(struct compat_shared_info, wc) != 0x900);
> > 	BUILD_BUG_ON(offsetof(struct compat_shared_info, arch.wc_sec_hi) != 0x924);
> > 	BUILD_BUG_ON(offsetof(struct pvclock_vcpu_time_info, version) != 0);
> > 
> > I don't really see the above fulfilling either of those use cases.
> > 
> > Given that the definition of the evtchn_pending field is:
> > 
> >         xen_ulong_t evtchn_pending[sizeof(xen_ulong_t) * 8];
> > 
> > It's fairly tautological that the number of event channels supported is
> > BITS_PER_ULONG * BITS_PER_ULONG. Which is sizeof(xen_ulong_t)² * 64 as
> > defined in the official Xen headers.
> > 
> > I don't know that we really need to add our own sanity check on the
> > headers we imported from Xen. It doesn't seem to add much.
> 
> The goal isn't to add a sanity check, it's to document what EVTCHN_2L_NR_CHANNELS
> actually represents.  My frustration with 
> 
>   sizeof(xen_ulong_t) * sizeof(xen_ulong_t) * 64
> 
> is that there's nothing there that connects it back to evtchn_pending or evtchn_mask.

Heh, welcome to Xen code :)

We could submit patches to Xen which make that clearer. 

> E.g. ideally the code would be something like
> 
>   #define COMPAT_EVTCHN_2L_NR_CHANNELS	256
> 
>   #ifdef CONFIG_X86_64
>   #define EVTCHN_2L_NR_CHANNELS		512
>   #else
>   #define EVTCHN_2L_NR_CHANNELS		COMPAT_EVTCHN_2L_NR_CHANNELS
> 
> 
>   DECLARE_BITMAP(evtchn_pending, EVTCHN_2L_NR_CHANNELS);
>   DECLARE_BITMAP(evtchn_mask, EVTCHN_2L_NR_CHANNELS);
> 
> which is much more self-documenting and doesn't require the reader to do math to
> grok the basics.

For the *compat* case that's entirely within arch/x86/kvm/xen.h so we
really *can* flip it to '#define COMPAT_EVTCHN_2L_NR_CHANNELS 1024' and
then use DECLARE_BITMAP in the struct itself. And I believe that I have
enough BUILD_BUG_ON() sanity checks that if you try that with 256
channels it would kick you in the teeth for it.

The fact that the official definition is an array of uint32_t is
irrelevant in our code, as we always cast to (unsigned long *) when
accessing the bitmaps anyway. DECLARE_BITMAP() is fine.

> Anyways, we can drop this patch, it was written mostly out of frustration with
> how long it took me to understand what is actually a very simple concept that's
> written in an unnecessarily obscure way.

Your experience is valid. This should be as understandable as possible
for people who aren't intimately familiar with the Xen ABIs, and I
certainly can't have an opinion on that.

How's something like this? I did start typing that comment in the
max_evtchn_port() function in xen.c but moved it over.

Still not utterly convinced, as it's still somewhat circular — we now
define NR_CHANNELS as (32*32) with a big comment explaining *why* that
is, and the reason is basically "because that's the number of bits in
an array of uint32_t[32]".

Which might perhaps have been expressed in C instead of prose, as

#define COMPAT_EVTCHN_2L_NR_CHANNELS (BITS_PER_BYTE * \
           sizeof_field(struct compat_shared_info,evtchn_pending))


Signed-off-by-if-you-want-it: David Woodhouse <dwmw@amazon.co.uk>

diff --git a/arch/x86/kvm/xen.h b/arch/x86/kvm/xen.h
index 8503d2c6891e..96735c8ad03f 100644
--- a/arch/x86/kvm/xen.h
+++ b/arch/x86/kvm/xen.h
@@ -190,17 +190,24 @@ struct compat_arch_shared_info {
 	uint32_t wc_sec_hi;
 };
 
+/*
+ * EVTCHN_2L_NR_CHANNELS is the number of event channels supported,
+ * corresponding to bits in the shinfo->evtchn_pending (and _mask)
+ * bitmaps. Those bitmaps are an array of
+ *     unsigned long evtchn_pending[BITS_PER_LONG];
+ * giving 1024 (32*32) or 4096 (64*64) event channels for 32-bit and
+ * 64-bit guests respectively.
+ */
+#define COMPAT_EVTCHN_2L_NR_CHANNELS (32 * 32)
+
 struct compat_shared_info {
 	struct compat_vcpu_info vcpu_info[MAX_VIRT_CPUS];
-	uint32_t evtchn_pending[32];
-	uint32_t evtchn_mask[32];
+	DECLARE_BITMAP(evtchn_pending, COMPAT_EVTCHN_2L_NR_CHANNELS);
+	DECLARE_BITMAP(evtchn_mask, COMPAT_EVTCHN_2L_NR_CHANNELS);
 	struct pvclock_wall_clock wc;
 	struct compat_arch_shared_info arch;
 };
 
-#define COMPAT_EVTCHN_2L_NR_CHANNELS (8 *				\
-				      sizeof_field(struct compat_shared_info, \
-						   evtchn_pending))
 struct compat_vcpu_runstate_info {
     int state;
     uint64_t state_entry_time;
  

On Tue, Nov 22, 2022, David Woodhouse wrote:
> On Tue, 2022-11-22 at 20:31 +0000, Sean Christopherson wrote:
> How's something like this? I did start typing that comment in the
> max_evtchn_port() function in xen.c but moved it over.
> 
> Still not utterly convinced, as it's still somewhat circular — we now
> define NR_CHANNELS as (32*32) with a big comment explaining *why* that
> is, and the reason is basically "because that's the number of bits in
> an array of uint32_t[32]".

Agreed, probably not an improvement across the board.  Consistency with how the
non-compat code declares the fields is also valuable, so unless someone changes
upstream Xen code, let's just leave things as-is.