diff mbox series

i386/kvm: Fix kvm_enable_x2apic link error in non-KVM builds

Message ID 20241113144923.41225-1-phil@philjordan.eu (mailing list archive)
State New
Headers show
Series i386/kvm: Fix kvm_enable_x2apic link error in non-KVM builds | expand

Commit Message

Phil Dennis-Jordan Nov. 13, 2024, 2:49 p.m. UTC
It appears that existing call sites for the kvm_enable_x2apic()
function rely on the compiler eliding the calls during optimisation
when building with KVM disabled, or on platforms other than Linux,
where that function is declared but not defined.

This fragile reliance recently broke down when commit b12cb38 added
a new call site which apparently failed to be optimised away when
building QEMU on macOS with clang, resulting in a link error.

This change moves the function declaration into the existing
#if CONFIG_KVM
block in the same header file, while the corresponding
#else
block now #defines the symbol as 0, same as for various other
KVM-specific query functions.

Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>
---
 target/i386/kvm/kvm_i386.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

Comments

Paolo Bonzini Nov. 13, 2024, 6:06 p.m. UTC | #1
On 11/13/24 15:49, Phil Dennis-Jordan wrote:
> It appears that existing call sites for the kvm_enable_x2apic()
> function rely on the compiler eliding the calls during optimisation
> when building with KVM disabled, or on platforms other than Linux,
> where that function is declared but not defined.
> 
> This fragile reliance recently broke down when commit b12cb38 added
> a new call site which apparently failed to be optimised away when
> building QEMU on macOS with clang, resulting in a link error.

That's weird, can you check the preprocessor output?  The definition
of kvm_irqchip_in_kernel() should be just "false" on macOS, in fact
even the area you're changing should be simplified like

diff --git a/target/i386/kvm/kvm_i386.h b/target/i386/kvm/kvm_i386.h
index 9de9c0d3038..7edb154a16e 100644
--- a/target/i386/kvm/kvm_i386.h
+++ b/target/i386/kvm/kvm_i386.h
@@ -13,8 +13,7 @@

  #include "sysemu/kvm.h"

-#ifdef CONFIG_KVM
-
+/* always false if !CONFIG_KVM */
  #define kvm_pit_in_kernel() \
      (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())
  #define kvm_pic_in_kernel()  \
@@ -22,14 +21,6 @@
  #define kvm_ioapic_in_kernel() \
      (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())

-#else
-
-#define kvm_pit_in_kernel()      0
-#define kvm_pic_in_kernel()      0
-#define kvm_ioapic_in_kernel()   0
-
-#endif  /* CONFIG_KVM */
-
  bool kvm_has_smm(void);
  bool kvm_enable_x2apic(void);

Paolo

> 
> Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>
> ---
>   target/i386/kvm/kvm_i386.h | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/target/i386/kvm/kvm_i386.h b/target/i386/kvm/kvm_i386.h
> index 9de9c0d3038..7ce47388d90 100644
> --- a/target/i386/kvm/kvm_i386.h
> +++ b/target/i386/kvm/kvm_i386.h
> @@ -21,17 +21,18 @@
>       (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())
>   #define kvm_ioapic_in_kernel() \
>       (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())
> +bool kvm_enable_x2apic(void);
>   
>   #else
>   
>   #define kvm_pit_in_kernel()      0
>   #define kvm_pic_in_kernel()      0
>   #define kvm_ioapic_in_kernel()   0
> +#define kvm_enable_x2apic()      0
>   
>   #endif  /* CONFIG_KVM */
>   
>   bool kvm_has_smm(void);
> -bool kvm_enable_x2apic(void);
>   bool kvm_hv_vpindex_settable(void);
>   bool kvm_enable_hypercall(uint64_t enable_mask);
>
Paolo Bonzini Nov. 13, 2024, 6:11 p.m. UTC | #2
On 11/13/24 15:49, Phil Dennis-Jordan wrote:
> It appears that existing call sites for the kvm_enable_x2apic()
> function rely on the compiler eliding the calls during optimisation
> when building with KVM disabled, or on platforms other than Linux,
> where that function is declared but not defined.
> 
> This fragile reliance recently broke down when commit b12cb38 added
> a new call site which apparently failed to be optimised away when
> building QEMU on macOS with clang, resulting in a link error.
> 
> This change moves the function declaration into the existing
> #if CONFIG_KVM
> block in the same header file, while the corresponding
> #else
> block now #defines the symbol as 0, same as for various other
> KVM-specific query functions.
> 
> Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>

Nevermind, this actually rung a bell and seems to be the same as
this commit from last year:

commit c04cfb4596ad5032a9869a8f77fe9114ca8af9e0
Author: Daniel Hoffman <dhoff749@gmail.com>
Date:   Sun Nov 19 12:31:16 2023 -0800

     hw/i386: fix short-circuit logic with non-optimizing builds
     
     `kvm_enabled()` is compiled down to `0` and short-circuit logic is
     used to remove references to undefined symbols at the compile stage.
     Some build configurations with some compilers don't attempt to
     simplify this logic down in some cases (the pattern appears to be
     that the literal false must be the first term) and this was causing
     some builds to emit references to undefined symbols.
     
     An example of such a configuration is clang 16.0.6 with the following
     configure: ./configure --enable-debug --without-default-features
     --target-list=x86_64-softmmu --enable-tcg-interpreter
     
     Signed-off-by: Daniel Hoffman <dhoff749@gmail.com>
     Message-Id: <20231119203116.3027230-1-dhoff749@gmail.com>
     Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
     Signed-off-by: Michael S. Tsirkin <mst@redhat.com>

So, this should work:

diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
index 13af7211e11..af0f4da1f69 100644
--- a/hw/i386/amd_iommu.c
+++ b/hw/i386/amd_iommu.c
@@ -1657,9 +1657,11 @@ static void amdvi_sysbus_realize(DeviceState *dev, Error **errp)
          error_report("AMD IOMMU with x2APIC confguration requires xtsup=on");
          exit(EXIT_FAILURE);
      }
-    if (s->xtsup && kvm_irqchip_is_split() && !kvm_enable_x2apic()) {
-        error_report("AMD IOMMU xtsup=on requires support on the KVM side");
-        exit(EXIT_FAILURE);
+    if (s->xtsup) {
+        if (kvm_irqchip_is_split() && !kvm_enable_x2apic()) {
+            error_report("AMD IOMMU xtsup=on requires support on the KVM side");
+            exit(EXIT_FAILURE);
+        }
      }
  
      pci_setup_iommu(bus, &amdvi_iommu_ops, s);


It's admittedly a bit brittle, but it's already done in the neighboring
hw/i386/intel_iommu.c so I guess it's okay.

Paolo
Shukla, Santosh Nov. 13, 2024, 6:25 p.m. UTC | #3
On 11/13/2024 11:41 PM, Paolo Bonzini wrote:
> On 11/13/24 15:49, Phil Dennis-Jordan wrote:
>> It appears that existing call sites for the kvm_enable_x2apic()
>> function rely on the compiler eliding the calls during optimisation
>> when building with KVM disabled, or on platforms other than Linux,
>> where that function is declared but not defined.
>>
>> This fragile reliance recently broke down when commit b12cb38 added
>> a new call site which apparently failed to be optimised away when
>> building QEMU on macOS with clang, resulting in a link error.
>>
>> This change moves the function declaration into the existing
>> #if CONFIG_KVM
>> block in the same header file, while the corresponding
>> #else
>> block now #defines the symbol as 0, same as for various other
>> KVM-specific query functions.
>>
>> Signed-off-by: Phil Dennis-Jordan <phil@philjordan.eu>
> 
> Nevermind, this actually rung a bell and seems to be the same as
> this commit from last year:
> 
> commit c04cfb4596ad5032a9869a8f77fe9114ca8af9e0
> Author: Daniel Hoffman <dhoff749@gmail.com>
> Date:   Sun Nov 19 12:31:16 2023 -0800
> 
>     hw/i386: fix short-circuit logic with non-optimizing builds
>         `kvm_enabled()` is compiled down to `0` and short-circuit logic is
>     used to remove references to undefined symbols at the compile stage.
>     Some build configurations with some compilers don't attempt to
>     simplify this logic down in some cases (the pattern appears to be
>     that the literal false must be the first term) and this was causing
>     some builds to emit references to undefined symbols.
>         An example of such a configuration is clang 16.0.6 with the following
>     configure: ./configure --enable-debug --without-default-features
>     --target-list=x86_64-softmmu --enable-tcg-interpreter
>         Signed-off-by: Daniel Hoffman <dhoff749@gmail.com>
>     Message-Id: <20231119203116.3027230-1-dhoff749@gmail.com>
>     Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
>     Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> 
> So, this should work:
> 
> diff --git a/hw/i386/amd_iommu.c b/hw/i386/amd_iommu.c
> index 13af7211e11..af0f4da1f69 100644
> --- a/hw/i386/amd_iommu.c
> +++ b/hw/i386/amd_iommu.c
> @@ -1657,9 +1657,11 @@ static void amdvi_sysbus_realize(DeviceState *dev, Error **errp)
>          error_report("AMD IOMMU with x2APIC confguration requires xtsup=on");
>          exit(EXIT_FAILURE);
>      }
> -    if (s->xtsup && kvm_irqchip_is_split() && !kvm_enable_x2apic()) {
> -        error_report("AMD IOMMU xtsup=on requires support on the KVM side");
> -        exit(EXIT_FAILURE);
> +    if (s->xtsup) {
> +        if (kvm_irqchip_is_split() && !kvm_enable_x2apic()) {
> +            error_report("AMD IOMMU xtsup=on requires support on the KVM side");
> +            exit(EXIT_FAILURE);
> +        }
>      }
>  
>      pci_setup_iommu(bus, &amdvi_iommu_ops, s);
> 
> 
> It's admittedly a bit brittle, but it's already done in the neighboring
> hw/i386/intel_iommu.c so I guess it's okay.
> 

Same proposed at https://lore.kernel.org/qemu-devel/cebca38a-5896-e2a5-8a68-5edad5dc9d8c@amd.com/
and I think Phil confirmed that it works.

Thanks,
Santosh

> Paolo
>
Paolo Bonzini Nov. 13, 2024, 6:39 p.m. UTC | #4
On Wed, Nov 13, 2024 at 7:25 PM Shukla, Santosh <santosh.shukla@amd.com> wrote:
> Same proposed at https://lore.kernel.org/qemu-devel/cebca38a-5896-e2a5-8a68-5edad5dc9d8c@amd.com/
> and I think Phil confirmed that it works.

Thanks Santosh, can you post it with commit message and everything?

Paolo
Shukla, Santosh Nov. 13, 2024, 6:40 p.m. UTC | #5
On 11/14/2024 12:09 AM, Paolo Bonzini wrote:
> On Wed, Nov 13, 2024 at 7:25 PM Shukla, Santosh <santosh.shukla@amd.com> wrote:
>> Same proposed at https://lore.kernel.org/qemu-devel/cebca38a-5896-e2a5-8a68-5edad5dc9d8c@amd.com/
>> and I think Phil confirmed that it works.
> 
> Thanks Santosh, can you post it with commit message and everything?
> 

Sure thing, will send by tomorrow.

Regards,
Santosh

> Paolo
>
diff mbox series

Patch

diff --git a/target/i386/kvm/kvm_i386.h b/target/i386/kvm/kvm_i386.h
index 9de9c0d3038..7ce47388d90 100644
--- a/target/i386/kvm/kvm_i386.h
+++ b/target/i386/kvm/kvm_i386.h
@@ -21,17 +21,18 @@ 
     (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())
 #define kvm_ioapic_in_kernel() \
     (kvm_irqchip_in_kernel() && !kvm_irqchip_is_split())
+bool kvm_enable_x2apic(void);
 
 #else
 
 #define kvm_pit_in_kernel()      0
 #define kvm_pic_in_kernel()      0
 #define kvm_ioapic_in_kernel()   0
+#define kvm_enable_x2apic()      0
 
 #endif  /* CONFIG_KVM */
 
 bool kvm_has_smm(void);
-bool kvm_enable_x2apic(void);
 bool kvm_hv_vpindex_settable(void);
 bool kvm_enable_hypercall(uint64_t enable_mask);