Merge branch 'x86-hyperv-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip

Pull x86/hyper-v update from Thomas Gleixner:
 "Add fast hypercall support for guest running on the Microsoft HyperV(isor)"

* 'x86-hyperv-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86/hyper-v: Fix wrong merge conflict resolution
  x86/hyper-v: Check for VP_INVAL in hyperv_flush_tlb_others()
  x86/hyper-v: Check cpumask_to_vpset() return value in hyperv_flush_tlb_others_ex()
  x86/hyper-v: Trace PV IPI send
  x86/hyper-v: Use cheaper HVCALL_SEND_IPI hypercall when possible
  x86/hyper-v: Use 'fast' hypercall for HVCALL_SEND_IPI
  x86/hyper-v: Implement hv_do_fast_hypercall16
  x86/hyper-v: Use cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE} hypercalls when possible

arch/x86/hyperv/hv_apic.c

@@ -31,6 +31,8 @@
 #include <asm/mshyperv.h>
 #include <asm/apic.h>
 
+#include <asm/trace/hyperv.h>
+
 static struct apic orig_apic;
 
 static u64 hv_apic_icr_read(void)
@@ -99,6 +101,9 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector)
 	int nr_bank = 0;
 	int ret = 1;
 
+	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
+		return false;
+
 	local_irq_save(flags);
 	arg = (struct ipi_arg_ex **)this_cpu_ptr(hyperv_pcpu_input_arg);
 
@@ -130,10 +135,10 @@ static bool __send_ipi_mask_ex(const struct cpumask *mask, int vector)
 static bool __send_ipi_mask(const struct cpumask *mask, int vector)
 {
 	int cur_cpu, vcpu;
-	struct ipi_arg_non_ex **arg;
-	struct ipi_arg_non_ex *ipi_arg;
+	struct ipi_arg_non_ex ipi_arg;
 	int ret = 1;
-	unsigned long flags;
+
+	trace_hyperv_send_ipi_mask(mask, vector);
 
 	if (cpumask_empty(mask))
 		return true;
@@ -144,40 +149,43 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector)
 	if ((vector < HV_IPI_LOW_VECTOR) || (vector > HV_IPI_HIGH_VECTOR))
 		return false;
 
-	if ((ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
-		return __send_ipi_mask_ex(mask, vector);
-
-	local_irq_save(flags);
-	arg = (struct ipi_arg_non_ex **)this_cpu_ptr(hyperv_pcpu_input_arg);
-
-	ipi_arg = *arg;
-	if (unlikely(!ipi_arg))
-		goto ipi_mask_done;
-
-	ipi_arg->vector = vector;
-	ipi_arg->reserved = 0;
-	ipi_arg->cpu_mask = 0;
+	/*
+	 * From the supplied CPU set we need to figure out if we can get away
+	 * with cheaper HVCALL_SEND_IPI hypercall. This is possible when the
+	 * highest VP number in the set is < 64. As VP numbers are usually in
+	 * ascending order and match Linux CPU ids, here is an optimization:
+	 * we check the VP number for the highest bit in the supplied set first
+	 * so we can quickly find out if using HVCALL_SEND_IPI_EX hypercall is
+	 * a must. We will also check all VP numbers when walking the supplied
+	 * CPU set to remain correct in all cases.
+	 */
+	if (hv_cpu_number_to_vp_number(cpumask_last(mask)) >= 64)
+		goto do_ex_hypercall;
+
+	ipi_arg.vector = vector;
+	ipi_arg.cpu_mask = 0;
 
 	for_each_cpu(cur_cpu, mask) {
 		vcpu = hv_cpu_number_to_vp_number(cur_cpu);
 		if (vcpu == VP_INVAL)
-			goto ipi_mask_done;
+			return false;
 
 		/*
 		 * This particular version of the IPI hypercall can
 		 * only target upto 64 CPUs.
 		 */
 		if (vcpu >= 64)
-			goto ipi_mask_done;
+			goto do_ex_hypercall;
 
-		__set_bit(vcpu, (unsigned long *)&ipi_arg->cpu_mask);
+		__set_bit(vcpu, (unsigned long *)&ipi_arg.cpu_mask);
 	}
 
-	ret = hv_do_hypercall(HVCALL_SEND_IPI, ipi_arg, NULL);
-
-ipi_mask_done:
-	local_irq_restore(flags);
+	ret = hv_do_fast_hypercall16(HVCALL_SEND_IPI, ipi_arg.vector,
+				     ipi_arg.cpu_mask);
 	return ((ret == 0) ? true : false);
+
+do_ex_hypercall:
+	return __send_ipi_mask_ex(mask, vector);
 }
 
 static bool __send_ipi_one(int cpu, int vector)
@@ -233,10 +241,7 @@ static void hv_send_ipi_self(int vector)
 void __init hv_apic_init(void)
 {
 	if (ms_hyperv.hints & HV_X64_CLUSTER_IPI_RECOMMENDED) {
-		if ((ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
-			pr_info("Hyper-V: Using ext hypercalls for IPI\n");
-		else
-			pr_info("Hyper-V: Using IPI hypercalls\n");
+		pr_info("Hyper-V: Using IPI hypercalls\n");
 		/*
 		 * Set the IPI entry points.
 		 */

arch/x86/hyperv/mmu.c

@@ -16,6 +16,8 @@
 /* Each gva in gva_list encodes up to 4096 pages to flush */
 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
 
+static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
+				      const struct flush_tlb_info *info);
 
 /*
  * Fills in gva_list starting from offset. Returns the number of items added.
@@ -93,10 +95,29 @@ static void hyperv_flush_tlb_others(const struct cpumask *cpus,
 	if (cpumask_equal(cpus, cpu_present_mask)) {
 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
 	} else {
+		/*
+		 * From the supplied CPU set we need to figure out if we can get
+		 * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
+		 * hypercalls. This is possible when the highest VP number in
+		 * the set is < 64. As VP numbers are usually in ascending order
+		 * and match Linux CPU ids, here is an optimization: we check
+		 * the VP number for the highest bit in the supplied set first
+		 * so we can quickly find out if using *_EX hypercalls is a
+		 * must. We will also check all VP numbers when walking the
+		 * supplied CPU set to remain correct in all cases.
+		 */
+		if (hv_cpu_number_to_vp_number(cpumask_last(cpus)) >= 64)
+			goto do_ex_hypercall;
+
 		for_each_cpu(cpu, cpus) {
 			vcpu = hv_cpu_number_to_vp_number(cpu);
-			if (vcpu >= 64)
+			if (vcpu == VP_INVAL) {
+				local_irq_restore(flags);
 				goto do_native;
+			}
+
+			if (vcpu >= 64)
+				goto do_ex_hypercall;
 
 			__set_bit(vcpu, (unsigned long *)
 				  &flush->processor_mask);
@@ -123,7 +144,12 @@ static void hyperv_flush_tlb_others(const struct cpumask *cpus,
 		status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
 					     gva_n, 0, flush, NULL);
 	}
+	goto check_status;
+
+do_ex_hypercall:
+	status = hyperv_flush_tlb_others_ex(cpus, info);
 
+check_status:
 	local_irq_restore(flags);
 
 	if (!(status & HV_HYPERCALL_RESULT_MASK))
@@ -132,35 +158,22 @@ static void hyperv_flush_tlb_others(const struct cpumask *cpus,
 	native_flush_tlb_others(cpus, info);
 }
 
-static void hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
-				       const struct flush_tlb_info *info)
+static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
+				      const struct flush_tlb_info *info)
 {
 	int nr_bank = 0, max_gvas, gva_n;
 	struct hv_tlb_flush_ex **flush_pcpu;
 	struct hv_tlb_flush_ex *flush;
-	u64 status = U64_MAX;
-	unsigned long flags;
+	u64 status;
 
-	trace_hyperv_mmu_flush_tlb_others(cpus, info);
-
-	if (!hv_hypercall_pg)
-		goto do_native;
-
-	if (cpumask_empty(cpus))
-		return;
-
-	local_irq_save(flags);
+	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
+		return U64_MAX;
 
 	flush_pcpu = (struct hv_tlb_flush_ex **)
 		     this_cpu_ptr(hyperv_pcpu_input_arg);
 
 	flush = *flush_pcpu;
 
-	if (unlikely(!flush)) {
-		local_irq_restore(flags);
-		goto do_native;
-	}
-
 	if (info->mm) {
 		/*
 		 * AddressSpace argument must match the CR3 with PCID bits
@@ -176,15 +189,10 @@ static void hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
 
 	flush->hv_vp_set.valid_bank_mask = 0;
 
-	if (!cpumask_equal(cpus, cpu_present_mask)) {
-		flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
-		nr_bank = cpumask_to_vpset(&(flush->hv_vp_set), cpus);
-	}
-
-	if (!nr_bank) {
-		flush->hv_vp_set.format = HV_GENERIC_SET_ALL;
-		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
-	}
+	flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
+	nr_bank = cpumask_to_vpset(&(flush->hv_vp_set), cpus);
+	if (nr_bank < 0)
+		return U64_MAX;
 
 	/*
 	 * We can flush not more than max_gvas with one hypercall. Flush the
@@ -213,12 +221,7 @@ static void hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
 			gva_n, nr_bank, flush, NULL);
 	}
 
-	local_irq_restore(flags);
-
-	if (!(status & HV_HYPERCALL_RESULT_MASK))
-		return;
-do_native:
-	native_flush_tlb_others(cpus, info);
+	return status;
 }
 
 void hyperv_setup_mmu_ops(void)
@@ -226,11 +229,6 @@ void hyperv_setup_mmu_ops(void)
 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
 		return;
 
-	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED)) {
-		pr_info("Using hypercall for remote TLB flush\n");
-		pv_mmu_ops.flush_tlb_others = hyperv_flush_tlb_others;
-	} else {
-		pr_info("Using ext hypercall for remote TLB flush\n");
-		pv_mmu_ops.flush_tlb_others = hyperv_flush_tlb_others_ex;
-	}
+	pr_info("Using hypercall for remote TLB flush\n");
+	pv_mmu_ops.flush_tlb_others = hyperv_flush_tlb_others;
 }

arch/x86/include/asm/mshyperv.h

@@ -194,6 +194,40 @@ static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1)
 		return hv_status;
 }
 
+/* Fast hypercall with 16 bytes of input */
+static inline u64 hv_do_fast_hypercall16(u16 code, u64 input1, u64 input2)
+{
+	u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
+
+#ifdef CONFIG_X86_64
+	{
+		__asm__ __volatile__("mov %4, %%r8\n"
+				     CALL_NOSPEC
+				     : "=a" (hv_status), ASM_CALL_CONSTRAINT,
+				       "+c" (control), "+d" (input1)
+				     : "r" (input2),
+				       THUNK_TARGET(hv_hypercall_pg)
+				     : "cc", "r8", "r9", "r10", "r11");
+	}
+#else
+	{
+		u32 input1_hi = upper_32_bits(input1);
+		u32 input1_lo = lower_32_bits(input1);
+		u32 input2_hi = upper_32_bits(input2);
+		u32 input2_lo = lower_32_bits(input2);
+
+		__asm__ __volatile__ (CALL_NOSPEC
+				      : "=A"(hv_status),
+					"+c"(input1_lo), ASM_CALL_CONSTRAINT
+				      :	"A" (control), "b" (input1_hi),
+					"D"(input2_hi), "S"(input2_lo),
+					THUNK_TARGET(hv_hypercall_pg)
+				      : "cc");
+	}
+#endif
+		return hv_status;
+}
+
 /*
  * Rep hypercalls. Callers of this functions are supposed to ensure that
  * rep_count and varhead_size comply with Hyper-V hypercall definition.

arch/x86/include/asm/trace/hyperv.h

@@ -28,6 +28,21 @@ TRACE_EVENT(hyperv_mmu_flush_tlb_others,
 		      __entry->addr, __entry->end)
 	);
 
+TRACE_EVENT(hyperv_send_ipi_mask,
+	    TP_PROTO(const struct cpumask *cpus,
+		     int vector),
+	    TP_ARGS(cpus, vector),
+	    TP_STRUCT__entry(
+		    __field(unsigned int, ncpus)
+		    __field(int, vector)
+		    ),
+	    TP_fast_assign(__entry->ncpus = cpumask_weight(cpus);
+			   __entry->vector = vector;
+		    ),
+	    TP_printk("ncpus %d vector %x",
+		      __entry->ncpus, __entry->vector)
+	);
+
 #endif /* CONFIG_HYPERV */
 
 #undef TRACE_INCLUDE_PATH