Skip to contentSkip to conversation

powerpc/pseries: IPI doorbell improvements

PATCH Superseded Nicholas Piggin 3 participants 6 messages Jun 30, 2020 – Jul 23, 2020

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Nicholas Piggin

author reviewer
#0 lore raw

v2 cover powerpc/pseries: IPI doorbell improvements

Since v1:
- Fixed SMP compile error.
- Fixed EPAPR / KVM_GUEST breakage.
- Expanded patch 3 changelog a bit.

Thanks,
Nick

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     | 63 ++++++++++++++++++++++++++--
 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, 134 insertions(+), 102 deletions(-)
Signature · 1 line
2.23.0
author reviewer
#1 lore raw

PATCH v2 1/3 powerpc: inline doorbell sending functions

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

---
 arch/powerpc/include/asm/dbell.h | 63 ++++++++++++++++++++++++++++++--
 arch/powerpc/kernel/dbell.c      | 55 ----------------------------
 2 files changed, 60 insertions(+), 58 deletions(-)
arch/powerpc/include/asm/dbell.h +60 -3
index 4ce6808deed3..f19d2282e3f8 100644
@@ -13,6 +13,7 @@
 
 #include <asm/ppc-opcode.h>
 #include <asm/feature-fixups.h>
+#include <asm/kvm_ppc.h>
 
 #define PPC_DBELL_MSG_BRDCAST	(0x04000000)
 #define PPC_DBELL_TYPE(x)	(((x) & 0xf) << (63-36))
@@ -87,9 +88,6 @@ 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 +98,63 @@ 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 */
arch/powerpc/kernel/dbell.c +0 -55

61 lines not on this page · raw message

Signature · 1 line
2.23.0
#2 lore raw

Re: [PATCH v2 1/3] powerpc: inline doorbell sending functions

Quoted 14 lines · Nicholas Piggin

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

---
 arch/powerpc/include/asm/dbell.h | 63 ++++++++++++++++++++++++++++++--
 arch/powerpc/kernel/dbell.c      | 55 ----------------------------
 2 files changed, 60 insertions(+), 58 deletions(-)
arch/powerpc/include/asm/dbell.h +0 -0
index 4ce6808deed3..f19d2282e3f8 100644

arch/powerpc/include/asm/dbell.h line 19

Quoted 7 lines
arch/powerpc/include/asm/dbell.h +1 -0
@@ -13,6 +13,7 @@
 
 #include <asm/ppc-opcode.h>
 #include <asm/feature-fixups.h>
+#include <asm/kvm_ppc.h>
 
 #define PPC_DBELL_MSG_BRDCAST	(0x04000000)
 #define PPC_DBELL_TYPE(x)	(((x) & 0xf) << (63-36))

This somehow breaks ppc40x_defconfig and others:

In file included from /home/michael/linux/arch/powerpc/include/asm/kvm_ppc.h:24,
from /home/michael/linux/arch/powerpc/include/asm/dbell.h:16,
from /home/michael/linux/arch/powerpc/kernel/asm-offsets.c:38:
/home/michael/linux/arch/powerpc/include/asm/kvm_booke.h: In function ‘kvmppc_get_fault_dar’:
/home/michael/linux/arch/powerpc/include/asm/kvm_booke.h:94:19: error: ‘struct kvm_vcpu_arch’ has no member named ‘fault_dear’
94 | return vcpu->arch.fault_dear;
| ^
make[2]: *** [/home/michael/linux/scripts/Makefile.build:114: arch/powerpc/kernel/asm-offsets.s] Error 1 make[1]: *** [/home/michael/linux/Makefile:1175: prepare0] Error 2
make: *** [Makefile:185: __sub-make] Error 2

cheers

author reviewer
#3 lore raw

PATCH v2 2/3 powerpc/pseries: Use doorbells even if XIVE is available

In reply to 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.

---
 arch/powerpc/platforms/pseries/smp.c | 54 ++++++++++++++++++----------
 1 file changed, 36 insertions(+), 18 deletions(-)
arch/powerpc/platforms/pseries/smp.c +36 -18
index 6891710833be..67e6ad5076ce 100644
@@ -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 = {
Signature · 1 line
2.23.0
author reviewer
#4 lore raw

PATCH v2 3/3 powerpc/pseries: Add KVM guest doorbell restrictions

In reply to 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.

---
 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(-)
arch/powerpc/include/asm/firmware.h +6 -0
index 6003c2e533a0..f67efbaba17f 100644
@@ -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
arch/powerpc/include/asm/kvm_para.h +3 -23
index 9c1f6b4b9bbf..744612054c94 100644
@@ -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;
arch/powerpc/kernel/Makefile +2 -3
index 244542ae2a91..852164439dcb 100644
@@ -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
arch/powerpc/kernel/firmware.c +19 -0
index cc4a5e3f51f1..fe48d319d490 100644
@@ -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
arch/powerpc/platforms/pseries/smp.c +23 -15
index 67e6ad5076ce..7af0003b40b6 100644
@@ -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;
 }
Signature · 1 line
2.23.0
Tested-by on the v2 series
#5 lore raw
In reply to Nicholas Piggin
Quoted 7 lines · Nicholas Piggin

Since v1:
- Fixed SMP compile error.
- Fixed EPAPR / KVM_GUEST breakage.
- Expanded patch 3 changelog a bit.

Thanks,
Nick

I gave the patchset a try on KVM guests (P9 and P8) and PowerVM LPARs (P9).

Looks good.

Thanks,

C.

Quoted 15 lines

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     | 63 ++++++++++++++++++++++++++--
 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, 134 insertions(+), 102 deletions(-)