[PATCH v3 0/3] powerpc/pseries: IPI doorbell improvements

[PATCH v3 0/3] powerpc/pseries: IPI doorbell improvements

[Nicholas Piggin]

Since v2:
- Fixed ppc32 compile error
- Tested-by from Cedric

Nicholas Piggin (3):
  powerpc: inline doorbell sending functions
  powerpc/pseries: Use doorbells even if XIVE is available
  powerpc/pseries: Add KVM guest doorbell restrictions

 arch/powerpc/include/asm/dbell.h     | 67 ++++++++++++++++++++++++++--
 arch/powerpc/include/asm/firmware.h  |  6 +++
 arch/powerpc/include/asm/kvm_para.h  | 26 ++---------
 arch/powerpc/kernel/Makefile         |  5 +--
 arch/powerpc/kernel/dbell.c          | 55 -----------------------
 arch/powerpc/kernel/firmware.c       | 19 ++++++++
 arch/powerpc/platforms/pseries/smp.c | 62 +++++++++++++++++--------
 7 files changed, 138 insertions(+), 102 deletions(-)

-- 
2.23.0

[PATCH v3 1/3] powerpc: inline doorbell sending functions

[Nicholas Piggin]

These are only called in one place for a given platform, so inline them
for performance.

Tested-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/include/asm/dbell.h | 67 ++++++++++++++++++++++++++++++--
 arch/powerpc/kernel/dbell.c      | 55 --------------------------
 2 files changed, 64 insertions(+), 58 deletions(-)

diff --git a/arch/powerpc/include/asm/dbell.h b/arch/powerpc/include/asm/dbell.h
index 4ce6808deed3..1f04f3de96ba 100644
--- a/arch/powerpc/include/asm/dbell.h
+++ b/arch/powerpc/include/asm/dbell.h
@@ -13,6 +13,11 @@
 
 #include <asm/ppc-opcode.h>
 #include <asm/feature-fixups.h>
+#ifdef CONFIG_KVM
+#include <asm/kvm_ppc.h>
+#else
+static inline void kvmppc_set_host_ipi(int cpu) {}
+#endif
 
 #define PPC_DBELL_MSG_BRDCAST	(0x04000000)
 #define PPC_DBELL_TYPE(x)	(((x) & 0xf) << (63-36))
@@ -87,9 +92,6 @@ static inline void ppc_msgsync(void)
 
 #endif /* CONFIG_PPC_BOOK3S */
 
-extern void doorbell_global_ipi(int cpu);
-extern void doorbell_core_ipi(int cpu);
-extern int doorbell_try_core_ipi(int cpu);
 extern void doorbell_exception(struct pt_regs *regs);
 
 static inline void ppc_msgsnd(enum ppc_dbell type, u32 flags, u32 tag)
@@ -100,4 +102,63 @@ static inline void ppc_msgsnd(enum ppc_dbell type, u32 flags, u32 tag)
 	_ppc_msgsnd(msg);
 }
 
+#ifdef CONFIG_SMP
+
+/*
+ * Doorbells must only be used if CPU_FTR_DBELL is available.
+ * msgsnd is used in HV, and msgsndp is used in !HV.
+ *
+ * These should be used by platform code that is aware of restrictions.
+ * Other arch code should use ->cause_ipi.
+ *
+ * doorbell_global_ipi() sends a dbell to any target CPU.
+ * Must be used only by architectures that address msgsnd target
+ * by PIR/get_hard_smp_processor_id.
+ */
+static inline void doorbell_global_ipi(int cpu)
+{
+	u32 tag = get_hard_smp_processor_id(cpu);
+
+	kvmppc_set_host_ipi(cpu);
+	/* Order previous accesses vs. msgsnd, which is treated as a store */
+	ppc_msgsnd_sync();
+	ppc_msgsnd(PPC_DBELL_MSGTYPE, 0, tag);
+}
+
+/*
+ * doorbell_core_ipi() sends a dbell to a target CPU in the same core.
+ * Must be used only by architectures that address msgsnd target
+ * by TIR/cpu_thread_in_core.
+ */
+static inline void doorbell_core_ipi(int cpu)
+{
+	u32 tag = cpu_thread_in_core(cpu);
+
+	kvmppc_set_host_ipi(cpu);
+	/* Order previous accesses vs. msgsnd, which is treated as a store */
+	ppc_msgsnd_sync();
+	ppc_msgsnd(PPC_DBELL_MSGTYPE, 0, tag);
+}
+
+/*
+ * Attempt to cause a core doorbell if destination is on the same core.
+ * Returns 1 on success, 0 on failure.
+ */
+static inline int doorbell_try_core_ipi(int cpu)
+{
+	int this_cpu = get_cpu();
+	int ret = 0;
+
+	if (cpumask_test_cpu(cpu, cpu_sibling_mask(this_cpu))) {
+		doorbell_core_ipi(cpu);
+		ret = 1;
+	}
+
+	put_cpu();
+
+	return ret;
+}
+
+#endif /* CONFIG_SMP */
+
 #endif /* _ASM_POWERPC_DBELL_H */
diff --git a/arch/powerpc/kernel/dbell.c b/arch/powerpc/kernel/dbell.c
index f17ff1200eaa..52680cf07c9d 100644
--- a/arch/powerpc/kernel/dbell.c
+++ b/arch/powerpc/kernel/dbell.c
@@ -18,61 +18,6 @@
 
 #ifdef CONFIG_SMP
 
-/*
- * Doorbells must only be used if CPU_FTR_DBELL is available.
- * msgsnd is used in HV, and msgsndp is used in !HV.
- *
- * These should be used by platform code that is aware of restrictions.
- * Other arch code should use ->cause_ipi.
- *
- * doorbell_global_ipi() sends a dbell to any target CPU.
- * Must be used only by architectures that address msgsnd target
- * by PIR/get_hard_smp_processor_id.
- */
-void doorbell_global_ipi(int cpu)
-{
-	u32 tag = get_hard_smp_processor_id(cpu);
-
-	kvmppc_set_host_ipi(cpu);
-	/* Order previous accesses vs. msgsnd, which is treated as a store */
-	ppc_msgsnd_sync();
-	ppc_msgsnd(PPC_DBELL_MSGTYPE, 0, tag);
-}
-
-/*
- * doorbell_core_ipi() sends a dbell to a target CPU in the same core.
- * Must be used only by architectures that address msgsnd target
- * by TIR/cpu_thread_in_core.
- */
-void doorbell_core_ipi(int cpu)
-{
-	u32 tag = cpu_thread_in_core(cpu);
-
-	kvmppc_set_host_ipi(cpu);
-	/* Order previous accesses vs. msgsnd, which is treated as a store */
-	ppc_msgsnd_sync();
-	ppc_msgsnd(PPC_DBELL_MSGTYPE, 0, tag);
-}
-
-/*
- * Attempt to cause a core doorbell if destination is on the same core.
- * Returns 1 on success, 0 on failure.
- */
-int doorbell_try_core_ipi(int cpu)
-{
-	int this_cpu = get_cpu();
-	int ret = 0;
-
-	if (cpumask_test_cpu(cpu, cpu_sibling_mask(this_cpu))) {
-		doorbell_core_ipi(cpu);
-		ret = 1;
-	}
-
-	put_cpu();
-
-	return ret;
-}
-
 void doorbell_exception(struct pt_regs *regs)
 {
 	struct pt_regs *old_regs = set_irq_regs(regs);
-- 
2.23.0

[PATCH v3 2/3] powerpc/pseries: Use doorbells even if XIVE is available

[Nicholas Piggin]

KVM supports msgsndp in guests by trapping and emulating the
instruction, so it was decided to always use XIVE for IPIs if it is
available. However on PowerVM systems, msgsndp can be used and gives
better performance. On large systems, high XIVE interrupt rates can
have sub-linear scaling, and using msgsndp can reduce the load on
the interrupt controller.

So switch to using core local doorbells even if XIVE is available.
This reduces performance for KVM guests with an SMT topology by
about 50% for ping-pong context switching between SMT vCPUs. An
option vector (or dt-cpu-ftrs) could be defined to disable msgsndp
to get KVM performance back.

Tested-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/platforms/pseries/smp.c | 54 ++++++++++++++++++----------
 1 file changed, 36 insertions(+), 18 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/smp.c b/arch/powerpc/platforms/pseries/smp.c
index 6891710833be..67e6ad5076ce 100644
--- a/arch/powerpc/platforms/pseries/smp.c
+++ b/arch/powerpc/platforms/pseries/smp.c
@@ -188,13 +188,16 @@ static int pseries_smp_prepare_cpu(int cpu)
 	return 0;
 }
 
-static void smp_pseries_cause_ipi(int cpu)
+/* Cause IPI as setup by the interrupt controller (xics or xive) */
+static void (*ic_cause_ipi)(int cpu) __ro_after_init;
+
+/* Use msgsndp doorbells target is a sibling, else use interrupt controller */
+static void dbell_or_ic_cause_ipi(int cpu)
 {
-	/* POWER9 should not use this handler */
 	if (doorbell_try_core_ipi(cpu))
 		return;
 
-	icp_ops->cause_ipi(cpu);
+	ic_cause_ipi(cpu);
 }
 
 static int pseries_cause_nmi_ipi(int cpu)
@@ -218,26 +221,41 @@ static int pseries_cause_nmi_ipi(int cpu)
 	return 0;
 }
 
-static __init void pSeries_smp_probe_xics(void)
-{
-	xics_smp_probe();
-
-	if (cpu_has_feature(CPU_FTR_DBELL) && !is_secure_guest())
-		smp_ops->cause_ipi = smp_pseries_cause_ipi;
-	else
-		smp_ops->cause_ipi = icp_ops->cause_ipi;
-}
-
 static __init void pSeries_smp_probe(void)
 {
 	if (xive_enabled())
-		/*
-		 * Don't use P9 doorbells when XIVE is enabled. IPIs
-		 * using MMIOs should be faster
-		 */
 		xive_smp_probe();
 	else
-		pSeries_smp_probe_xics();
+		xics_smp_probe();
+
+	/* No doorbell facility, must use the interrupt controller for IPIs */
+	if (!cpu_has_feature(CPU_FTR_DBELL))
+		return;
+
+	/* Doorbells can only be used for IPIs between SMT siblings */
+	if (!cpu_has_feature(CPU_FTR_SMT))
+		return;
+
+	/*
+	 * KVM emulates doorbells by disabling FSCR[MSGP] so msgsndp faults
+	 * to the hypervisor which then reads the instruction from guest
+	 * memory. This can't be done if the guest is secure, so don't use
+	 * doorbells in secure guests.
+	 *
+	 * Under PowerVM, FSCR[MSGP] is enabled so doorbells could be used
+	 * by secure guests if we distinguished this from KVM.
+	 */
+	if (is_secure_guest())
+		return;
+
+	/*
+	 * The guest can use doobells for SMT sibling IPIs, which stay in
+	 * the core rather than going to the interrupt controller. This
+	 * tends to be slower under KVM where doorbells are emulated, but
+	 * faster for PowerVM where they're enabled.
+	 */
+	ic_cause_ipi = smp_ops->cause_ipi;
+	smp_ops->cause_ipi = dbell_or_ic_cause_ipi;
 }
 
 static struct smp_ops_t pseries_smp_ops = {
-- 
2.23.0

[PATCH v3 3/3] powerpc/pseries: Add KVM guest doorbell restrictions

[Nicholas Piggin]

KVM guests have certain restrictions and performance quirks when using
doorbells. This patch moves the EPAPR KVM guest test so it can be shared
with PSERIES, and uses that in doorbell setup code to apply the KVM
guest quirks and  improves IPI performance for two cases:

 - PowerVM guests may now use doorbells even if they are secure.

 - KVM guests no longer use doorbells if XIVE is available.

There is a valid complaint that "KVM guest" is not a very reasonable
thing to test for, it's preferable for the hypervisor to advertise
particular behaviours to the guest so they could change if the
hypervisor implementation or configuration changes. However in this case
we were already assuming a KVM guest worst case, so this patch is about
containing those quirks. If KVM later advertises fast doorbells, we
should test for that and override the quirks.

Tested-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 arch/powerpc/include/asm/firmware.h  |  6 +++++
 arch/powerpc/include/asm/kvm_para.h  | 26 +++----------------
 arch/powerpc/kernel/Makefile         |  5 ++--
 arch/powerpc/kernel/firmware.c       | 19 ++++++++++++++
 arch/powerpc/platforms/pseries/smp.c | 38 +++++++++++++++++-----------
 5 files changed, 53 insertions(+), 41 deletions(-)

diff --git a/arch/powerpc/include/asm/firmware.h b/arch/powerpc/include/asm/firmware.h
index 6003c2e533a0..f67efbaba17f 100644
--- a/arch/powerpc/include/asm/firmware.h
+++ b/arch/powerpc/include/asm/firmware.h
@@ -132,6 +132,12 @@ extern int ibm_nmi_interlock_token;
 
 extern unsigned int __start___fw_ftr_fixup, __stop___fw_ftr_fixup;
 
+#if defined(CONFIG_PPC_PSERIES) || defined(CONFIG_KVM_GUEST)
+bool is_kvm_guest(void);
+#else
+static inline bool is_kvm_guest(void) { return false; }
+#endif
+
 #ifdef CONFIG_PPC_PSERIES
 void pseries_probe_fw_features(void);
 #else
diff --git a/arch/powerpc/include/asm/kvm_para.h b/arch/powerpc/include/asm/kvm_para.h
index 9c1f6b4b9bbf..744612054c94 100644
--- a/arch/powerpc/include/asm/kvm_para.h
+++ b/arch/powerpc/include/asm/kvm_para.h
@@ -8,35 +8,15 @@
 #ifndef __POWERPC_KVM_PARA_H__
 #define __POWERPC_KVM_PARA_H__
 
-#include <uapi/asm/kvm_para.h>
-
-#ifdef CONFIG_KVM_GUEST
-
-#include <linux/of.h>
-
-static inline int kvm_para_available(void)
-{
-	struct device_node *hyper_node;
-
-	hyper_node = of_find_node_by_path("/hypervisor");
-	if (!hyper_node)
-		return 0;
+#include <asm/firmware.h>
 
-	if (!of_device_is_compatible(hyper_node, "linux,kvm"))
-		return 0;
-
-	return 1;
-}
-
-#else
+#include <uapi/asm/kvm_para.h>
 
 static inline int kvm_para_available(void)
 {
-	return 0;
+	return IS_ENABLED(CONFIG_KVM_GUEST) && is_kvm_guest();
 }
 
-#endif
-
 static inline unsigned int kvm_arch_para_features(void)
 {
 	unsigned long r;
diff --git a/arch/powerpc/kernel/Makefile b/arch/powerpc/kernel/Makefile
index 244542ae2a91..852164439dcb 100644
--- a/arch/powerpc/kernel/Makefile
+++ b/arch/powerpc/kernel/Makefile
@@ -45,11 +45,10 @@ obj-y				:= cputable.o syscalls.o \
 				   signal.o sysfs.o cacheinfo.o time.o \
 				   prom.o traps.o setup-common.o \
 				   udbg.o misc.o io.o misc_$(BITS).o \
-				   of_platform.o prom_parse.o
+				   of_platform.o prom_parse.o firmware.o
 obj-y				+= ptrace/
 obj-$(CONFIG_PPC64)		+= setup_64.o \
-				   paca.o nvram_64.o firmware.o note.o \
-				   syscall_64.o
+				   paca.o nvram_64.o note.o syscall_64.o
 obj-$(CONFIG_COMPAT)		+= sys_ppc32.o signal_32.o
 obj-$(CONFIG_VDSO32)		+= vdso32/
 obj-$(CONFIG_PPC_WATCHDOG)	+= watchdog.o
diff --git a/arch/powerpc/kernel/firmware.c b/arch/powerpc/kernel/firmware.c
index cc4a5e3f51f1..fe48d319d490 100644
--- a/arch/powerpc/kernel/firmware.c
+++ b/arch/powerpc/kernel/firmware.c
@@ -11,8 +11,27 @@
 
 #include <linux/export.h>
 #include <linux/cache.h>
+#include <linux/of.h>
 
 #include <asm/firmware.h>
 
+#ifdef CONFIG_PPC64
 unsigned long powerpc_firmware_features __read_mostly;
 EXPORT_SYMBOL_GPL(powerpc_firmware_features);
+#endif
+
+#if defined(CONFIG_PPC_PSERIES) || defined(CONFIG_KVM_GUEST)
+bool is_kvm_guest(void)
+{
+	struct device_node *hyper_node;
+
+	hyper_node = of_find_node_by_path("/hypervisor");
+	if (!hyper_node)
+		return 0;
+
+	if (!of_device_is_compatible(hyper_node, "linux,kvm"))
+		return 0;
+
+	return 1;
+}
+#endif
diff --git a/arch/powerpc/platforms/pseries/smp.c b/arch/powerpc/platforms/pseries/smp.c
index 67e6ad5076ce..7af0003b40b6 100644
--- a/arch/powerpc/platforms/pseries/smp.c
+++ b/arch/powerpc/platforms/pseries/smp.c
@@ -236,24 +236,32 @@ static __init void pSeries_smp_probe(void)
 	if (!cpu_has_feature(CPU_FTR_SMT))
 		return;
 
-	/*
-	 * KVM emulates doorbells by disabling FSCR[MSGP] so msgsndp faults
-	 * to the hypervisor which then reads the instruction from guest
-	 * memory. This can't be done if the guest is secure, so don't use
-	 * doorbells in secure guests.
-	 *
-	 * Under PowerVM, FSCR[MSGP] is enabled so doorbells could be used
-	 * by secure guests if we distinguished this from KVM.
-	 */
-	if (is_secure_guest())
-		return;
+	if (is_kvm_guest()) {
+		/*
+		 * KVM emulates doorbells by disabling FSCR[MSGP] so msgsndp
+		 * faults to the hypervisor which then reads the instruction
+		 * from guest memory, which tends to be slower than using XIVE.
+		 */
+		if (xive_enabled())
+			return;
+
+		/*
+		 * XICS hcalls aren't as fast, so we can use msgsndp (which
+		 * also helps exercise KVM emulation), however KVM can't
+		 * emulate secure guests because it can't read the instruction
+		 * out of their memory.
+		 */
+		if (is_secure_guest())
+			return;
+	}
 
 	/*
-	 * The guest can use doobells for SMT sibling IPIs, which stay in
-	 * the core rather than going to the interrupt controller. This
-	 * tends to be slower under KVM where doorbells are emulated, but
-	 * faster for PowerVM where they're enabled.
+	 * Under PowerVM, FSCR[MSGP] is enabled as guest vCPU siblings are
+	 * gang scheduled on the same physical core, so doorbells are always
+	 * faster than the interrupt controller, and they can be used by
+	 * secure guests.
 	 */
+
 	ic_cause_ipi = smp_ops->cause_ipi;
 	smp_ops->cause_ipi = dbell_or_ic_cause_ipi;
 }
-- 
2.23.0

Re: [PATCH v3 1/3] powerpc: inline doorbell sending functions

[Michael Ellerman]

Nicholas Piggin <npiggin@gmail.com> writes:
> diff --git a/arch/powerpc/include/asm/dbell.h b/arch/powerpc/include/asm/dbell.h
> index 4ce6808deed3..1f04f3de96ba 100644
> --- a/arch/powerpc/include/asm/dbell.h
> +++ b/arch/powerpc/include/asm/dbell.h
> @@ -100,4 +102,63 @@ static inline void ppc_msgsnd(enum ppc_dbell type, u32 flags, u32 tag)
>  	_ppc_msgsnd(msg);
>  }
>  
> +#ifdef CONFIG_SMP
> +
> +/*
> + * Doorbells must only be used if CPU_FTR_DBELL is available.
> + * msgsnd is used in HV, and msgsndp is used in !HV.
> + *
> + * These should be used by platform code that is aware of restrictions.
> + * Other arch code should use ->cause_ipi.
> + *
> + * doorbell_global_ipi() sends a dbell to any target CPU.
> + * Must be used only by architectures that address msgsnd target
> + * by PIR/get_hard_smp_processor_id.
> + */
> +static inline void doorbell_global_ipi(int cpu)
> +{
> +	u32 tag = get_hard_smp_processor_id(cpu);
> +
> +	kvmppc_set_host_ipi(cpu);
> +	/* Order previous accesses vs. msgsnd, which is treated as a store */
> +	ppc_msgsnd_sync();
> +	ppc_msgsnd(PPC_DBELL_MSGTYPE, 0, tag);
> +}
> +
> +/*
> + * doorbell_core_ipi() sends a dbell to a target CPU in the same core.
> + * Must be used only by architectures that address msgsnd target
> + * by TIR/cpu_thread_in_core.
> + */
> +static inline void doorbell_core_ipi(int cpu)
> +{
> +	u32 tag = cpu_thread_in_core(cpu);

corenet64_smp_defconfig gives me:

  In file included from /linux/arch/powerpc/kernel/asm-offsets.c:38:
  /linux/arch/powerpc/include/asm/dbell.h: In function 'doorbell_core_ipi':
  /linux/arch/powerpc/include/asm/dbell.h:135:12: error: implicit declaration of function 'cpu_thread_in_core' [-Werror=implicit-function-declaration]
    135 |  u32 tag = cpu_thread_in_core(cpu);
        |            ^~~~~~~~~~~~~~~~~~


Fixed by including cputhreads.h, but then that results in errors due to
your addition of kvmppc_set_host_ipi().

Removing that gets us back to the fault_dear error.

I think I see a way around it, will do some build tests.

cheers

Re: [PATCH v3 0/3] powerpc/pseries: IPI doorbell improvements

[Michael Ellerman]

On Sun, 26 Jul 2020 13:51:52 +1000, Nicholas Piggin wrote:
> Since v2:
> - Fixed ppc32 compile error
> - Tested-by from Cedric
> 
> Nicholas Piggin (3):
>   powerpc: inline doorbell sending functions
>   powerpc/pseries: Use doorbells even if XIVE is available
>   powerpc/pseries: Add KVM guest doorbell restrictions
> 
> [...]

Applied to powerpc/next.

[1/3] powerpc: Inline doorbell sending functions
      https://git.kernel.org/powerpc/c/1f0ce497433f8944045ee1baae218e31a0d295ee
[2/3] powerpc/pseries: Use doorbells even if XIVE is available
      https://git.kernel.org/powerpc/c/5b06d1679f2fe874ef49ea11324cd893ec9e2da8
[3/3] powerpc/pseries: Add KVM guest doorbell restrictions
      https://git.kernel.org/powerpc/c/107c55005fbd5243ee31fb13b6f166cde9e3ade1

cheers