KVM: x86/xen: Make number of event channels defines less magical
Subsystems (5)
Sean Christopherson
maintainerPATCH 1/1 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.
--- arch/x86/kvm/xen.h | 8 +++++--- include/xen/interface/event_channel.h | 6 ++++-- 2 files changed, 9 insertions(+), 5 deletions(-)
arch/x86/kvm/xen.h +5 -3
@@ -9,6 +9,8 @@#ifndef __ARCH_X86_KVM_XEN_H__#define __ARCH_X86_KVM_XEN_H__ +#include <linux/bits.h> +#ifdef CONFIG_KVM_XEN #include <linux/jump_label_ratelimit.h>
@@ -198,9 +200,9 @@ struct compat_shared_info {struct compat_arch_shared_info arch; }; -#define COMPAT_EVTCHN_2L_NR_CHANNELS (8 * \ - sizeof_field(struct compat_shared_info, \ - evtchn_pending)) +#define COMPAT_EVTCHN_2L_NR_CHANNELS \ + (sizeof_field(struct compat_shared_info, evtchn_pending) * BITS_PER_BYTE) +struct compat_vcpu_runstate_info {int state;uint64_t state_entry_time;
include/xen/interface/event_channel.h +4 -2
@@ -10,6 +10,8 @@#ifndef __XEN_PUBLIC_EVENT_CHANNEL_H__#define __XEN_PUBLIC_EVENT_CHANNEL_H__ +#include <linux/bits.h> +#include <xen/interface/xen.h> typedef uint32_t evtchn_port_t ;
@@ -244,8 +246,8 @@ DEFINE_GUEST_HANDLE_STRUCT(evtchn_op);/* * 2-level ABI */ - -#define EVTCHN_2L_NR_CHANNELS (sizeof(xen_ulong_t) * sizeof(xen_ulong_t) * 64) +#define EVTCHN_2L_NR_CHANNELS \ + (sizeof_field(struct shared_info, evtchn_pending) * BITS_PER_BYTE)/* * FIFO ABI
base-commit: d663b8a285986072428a6a145e5994bc275df994
Signature · 1 line
2.38.1.431.g37b22c650d-goog
David Woodhouse
maintainerQuoted 7 lines · Sean Christopherson
Use
BITS_PER_BYTEandsizeof_field()to compute the number of Xen event channels. The compat version at least usessizeof_field(), but the regular version open codessizeof_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.
Sean Christopherson
maintainerQuoted 11 lines · David Woodhouse
Quoted 7 lines · Sean Christopherson
Use
BITS_PER_BYTEandsizeof_field()to compute the number of Xen event channels. The compat version at least usessizeof_field(), but the regular version open codessizeof_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:
arch/x86/kvm/xen.c +3 -0
@@ -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
David Woodhouse
maintainerQuoted 8 lines · Sean Christopherson
Ugh. I worried that might be the case. An alternative approach to help document things from a KVM perspective would be something like:
arch/x86/kvm/xen.c +0
-0
arch/x86/kvm/xen.c int kvm_xen_hypercall(struct kvm_vcpu *vcpu) · line 1308
Quoted 9 lines
arch/x86/kvm/xen.c +3
-0@@ -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.
Sean Christopherson
maintainerQuoted 45 lines · David Woodhouse
Quoted 8 lines · Sean Christopherson
Ugh. I worried that might be the case. An alternative approach to help document things from a KVM perspective would be something like:
arch/x86/kvm/xen.c +0
-0arch/x86/kvm/xen.c int kvm_xen_hypercall(struct kvm_vcpu *vcpu) · line 1308
Quoted 9 lines
arch/x86/kvm/xen.c +3
-0@@ -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_pendingfield is:
Quoted 8 lines
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.
David Woodhouse
maintainerQuoted 55 lines · Sean Christopherson
Quoted 45 lines · David Woodhouse
Quoted 8 lines · Sean Christopherson
Ugh. I worried that might be the case. An alternative approach to help document things from a KVM perspective would be something like:
arch/x86/kvm/xen.c +0
-0arch/x86/kvm/xen.c int kvm_xen_hypercall(struct kvm_vcpu *vcpu) · line 1308
Quoted 9 lines
arch/x86/kvm/xen.c +3
-0@@ -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_pendingfield is:Quoted 8 lines
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.
Quoted 6 lines
The goal isn't to add a sanity check, it's to document what
EVTCHN_2L_NR_CHANNELSactually represents. My frustration with
sizeof(xen_ulong_t)*sizeof(xen_ulong_t)* 64is that there's nothing there that connects it back to
evtchn_pendingorevtchn_mask.
Heh, welcome to Xen code :)
We could submit patches to Xen which make that clearer.
Quoted 13 lines
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);
Quoted 2 lines
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.
Quoted 3 lines
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 <[email protected]>
arch/x86/kvm/xen.h +12 -5
@@ -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;
Sean Christopherson
maintainerQuoted 8 lines · David Woodhouse
On Tue, 2022-11-22 at 20:31 +0000, Sean Christopherson wrote:
How's something like this? I did start typing that comment in themax_evtchn_port()function inxen.cbut moved it over.Still not utterly convinced, as it's still somewhat circular — we now define
NR_CHANNELSas (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.