[PATCH 0/3] LoongArch: CPU parallel bring up

[PATCH 0/3] LoongArch: CPU parallel bring up

[Jiaxun Yang]

Hi all,

This series implemented CPU parallel bring up for LoongArch.

Being the first non-x86 arch enabling that we need to fix some
infra in patch 1 and 2, then implement everything in patch 3.

Please review.
Thanks

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
---
Jiaxun Yang (3):
      cpu/hotplug: Make HOTPLUG_PARALLEL independent of HOTPLUG_SMT
      cpu/hotplug: Weak fallback for arch_cpuhp_init_parallel_bringup
      LoongArch: SMP: Implement parallel CPU bring up

 arch/loongarch/Kconfig              |  1 +
 arch/loongarch/include/asm/smp.h    |  6 ------
 arch/loongarch/kernel/asm-offsets.c | 10 ----------
 arch/loongarch/kernel/head.S        |  7 ++++---
 arch/loongarch/kernel/smp.c         | 35 +++++++----------------------------
 kernel/cpu.c                        |  9 +++++++++
 6 files changed, 21 insertions(+), 47 deletions(-)
---
base-commit: 82e4255305c554b0bb18b7ccf2db86041b4c8b6e
change-id: 20240704-loongarch-hotplug-3f8826b88a43

Best regards,
-- 
Jiaxun Yang <jiaxun.yang@flygoat.com>

[PATCH 1/3] cpu/hotplug: Make HOTPLUG_PARALLEL independent of HOTPLUG_SMT

[Jiaxun Yang]

Use ifdef to guard reference to HOTPLUG_SMT symbols so HOTPLUG_PARALLEL
can be enabled without HOTPLUG_SMT being selected.

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
---
 kernel/cpu.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/kernel/cpu.c b/kernel/cpu.c
index 1209ddaec026..2ab3f68a1bb1 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -1808,6 +1808,7 @@ static int __init parallel_bringup_parse_param(char *arg)
 }
 early_param("cpuhp.parallel", parallel_bringup_parse_param);
 
+#ifdef CONFIG_HOTPLUG_SMT
 static inline bool cpuhp_smt_aware(void)
 {
 	return cpu_smt_max_threads > 1;
@@ -1817,6 +1818,7 @@ static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
 {
 	return cpu_primary_thread_mask;
 }
+#endif
 
 /*
  * On architectures which have enabled parallel bringup this invokes all BP
@@ -1837,6 +1839,7 @@ static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
 	if (!__cpuhp_parallel_bringup)
 		return false;
 
+#ifdef CONFIG_HOTPLUG_SMT
 	if (cpuhp_smt_aware()) {
 		const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
 		static struct cpumask tmp_mask __initdata;
@@ -1857,6 +1860,7 @@ static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
 		cpumask_andnot(&tmp_mask, mask, pmask);
 		mask = &tmp_mask;
 	}
+#endif
 
 	/* Bring the not-yet started CPUs up */
 	cpuhp_bringup_mask(mask, ncpus, CPUHP_BP_KICK_AP);

-- 
2.45.2

[PATCH 2/3] cpu/hotplug: Weak fallback for arch_cpuhp_init_parallel_bringup

[Jiaxun Yang]

It is a general assumption that architectures entitled to parallel
bringup with CONFIG_HOTPLUG_PARALLEL do expect parallel bringup to
be available.

Provide a weak fallback arch_cpuhp_init_parallel_bringup function
to match this assumption.

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
---
 kernel/cpu.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/kernel/cpu.c b/kernel/cpu.c
index 2ab3f68a1bb1..c925d0f4b0c8 100644
--- a/kernel/cpu.c
+++ b/kernel/cpu.c
@@ -1820,6 +1820,11 @@ static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
 }
 #endif
 
+bool __weak arch_cpuhp_init_parallel_bringup(void)
+{
+	return true;
+}
+
 /*
  * On architectures which have enabled parallel bringup this invokes all BP
  * prepare states for each of the to be onlined APs first. The last state

-- 
2.45.2

[PATCH 3/3] LoongArch: SMP: Implement parallel CPU bring up

[Jiaxun Yang]

Implement parallel CPU bring up for LoongArch to reduce
boot time consumption on bring up CPUs.

On my Loongson-3A5000 ~120ms boot time improvement is observed.

tp, sp register values are passed by MBUF now to avoid racing
cpuboot_data global struct.

cpu_running completion is handled by HOTPLUG_CORE_SYNC_FULL.

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
---
 arch/loongarch/Kconfig              |  1 +
 arch/loongarch/include/asm/smp.h    |  6 ------
 arch/loongarch/kernel/asm-offsets.c | 10 ----------
 arch/loongarch/kernel/head.S        |  7 ++++---
 arch/loongarch/kernel/smp.c         | 35 +++++++----------------------------
 5 files changed, 12 insertions(+), 47 deletions(-)

diff --git a/arch/loongarch/Kconfig b/arch/loongarch/Kconfig
index ddc042895d01..656435c1dbd5 100644
--- a/arch/loongarch/Kconfig
+++ b/arch/loongarch/Kconfig
@@ -162,6 +162,7 @@ config LOONGARCH
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_TIF_NOHZ
 	select HAVE_VIRT_CPU_ACCOUNTING_GEN if !SMP
+	select HOTPLUG_PARALLEL if SMP
 	select IRQ_FORCED_THREADING
 	select IRQ_LOONGARCH_CPU
 	select LOCK_MM_AND_FIND_VMA
diff --git a/arch/loongarch/include/asm/smp.h b/arch/loongarch/include/asm/smp.h
index 50db503f44e3..f6953cb16492 100644
--- a/arch/loongarch/include/asm/smp.h
+++ b/arch/loongarch/include/asm/smp.h
@@ -75,12 +75,6 @@ extern int __cpu_logical_map[NR_CPUS];
 #define SMP_CALL_FUNCTION	BIT(ACTION_CALL_FUNCTION)
 #define SMP_IRQ_WORK		BIT(ACTION_IRQ_WORK)
 
-struct secondary_data {
-	unsigned long stack;
-	unsigned long thread_info;
-};
-extern struct secondary_data cpuboot_data;
-
 extern asmlinkage void smpboot_entry(void);
 extern asmlinkage void start_secondary(void);
 
diff --git a/arch/loongarch/kernel/asm-offsets.c b/arch/loongarch/kernel/asm-offsets.c
index bee9f7a3108f..598498f47a4c 100644
--- a/arch/loongarch/kernel/asm-offsets.c
+++ b/arch/loongarch/kernel/asm-offsets.c
@@ -257,16 +257,6 @@ static void __used output_signal_defines(void)
 	BLANK();
 }
 
-#ifdef CONFIG_SMP
-static void __used output_smpboot_defines(void)
-{
-	COMMENT("Linux smp cpu boot offsets.");
-	OFFSET(CPU_BOOT_STACK, secondary_data, stack);
-	OFFSET(CPU_BOOT_TINFO, secondary_data, thread_info);
-	BLANK();
-}
-#endif
-
 #ifdef CONFIG_HIBERNATION
 static void __used output_pbe_defines(void)
 {
diff --git a/arch/loongarch/kernel/head.S b/arch/loongarch/kernel/head.S
index fdb831dc64df..8dd8fb450f46 100644
--- a/arch/loongarch/kernel/head.S
+++ b/arch/loongarch/kernel/head.S
@@ -136,9 +136,10 @@ SYM_CODE_START(smpboot_entry)
 	li.w		t0, 0x00		# FPE=0, SXE=0, ASXE=0, BTE=0
 	csrwr		t0, LOONGARCH_CSR_EUEN
 
-	la.pcrel	t0, cpuboot_data
-	ld.d		sp, t0, CPU_BOOT_STACK
-	ld.d		tp, t0, CPU_BOOT_TINFO
+	li.w		t0, LOONGARCH_IOCSR_MBUF1
+	iocsrrd.d	sp, t0
+	li.w		t0, LOONGARCH_IOCSR_MBUF2
+	iocsrrd.d	tp, t0
 
 	bl		start_secondary
 	ASM_BUG()
diff --git a/arch/loongarch/kernel/smp.c b/arch/loongarch/kernel/smp.c
index ca405ab86aae..ca6a95a0280d 100644
--- a/arch/loongarch/kernel/smp.c
+++ b/arch/loongarch/kernel/smp.c
@@ -48,10 +48,6 @@ EXPORT_SYMBOL(cpu_sibling_map);
 /* Representing the core map of multi-core chips of each logical CPU */
 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
 EXPORT_SYMBOL(cpu_core_map);
-
-static DECLARE_COMPLETION(cpu_starting);
-static DECLARE_COMPLETION(cpu_running);
-
 /*
  * A logcal cpu mask containing only one VPE per core to
  * reduce the number of IPIs on large MT systems.
@@ -65,7 +61,6 @@ static cpumask_t cpu_sibling_setup_map;
 /* representing cpus for which core maps can be computed */
 static cpumask_t cpu_core_setup_map;
 
-struct secondary_data cpuboot_data;
 static DEFINE_PER_CPU(int, cpu_state);
 
 static const char *ipi_types[NR_IPI] __tracepoint_string = {
@@ -340,14 +335,16 @@ void __init loongson_prepare_cpus(unsigned int max_cpus)
  */
 void loongson_boot_secondary(int cpu, struct task_struct *idle)
 {
-	unsigned long entry;
+	unsigned long entry, stack, thread_info;
 
 	pr_info("Booting CPU#%d...\n", cpu);
 
 	entry = __pa_symbol((unsigned long)&smpboot_entry);
-	cpuboot_data.stack = (unsigned long)__KSTK_TOS(idle);
-	cpuboot_data.thread_info = (unsigned long)task_thread_info(idle);
+	stack = (unsigned long)__KSTK_TOS(idle);
+	thread_info = (unsigned long)task_thread_info(idle);
 
+	csr_mail_send(thread_info, cpu_logical_map(cpu), 2);
+	csr_mail_send(stack, cpu_logical_map(cpu), 1);
 	csr_mail_send(entry, cpu_logical_map(cpu), 0);
 
 	loongson_send_ipi_single(cpu, ACTION_BOOT_CPU);
@@ -525,20 +522,10 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 #endif
 }
 
-int __cpu_up(unsigned int cpu, struct task_struct *tidle)
+int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
 {
 	loongson_boot_secondary(cpu, tidle);
 
-	/* Wait for CPU to start and be ready to sync counters */
-	if (!wait_for_completion_timeout(&cpu_starting,
-					 msecs_to_jiffies(5000))) {
-		pr_crit("CPU%u: failed to start\n", cpu);
-		return -EIO;
-	}
-
-	/* Wait for CPU to finish startup & mark itself online before return */
-	wait_for_completion(&cpu_running);
-
 	return 0;
 }
 
@@ -561,22 +548,14 @@ asmlinkage void start_secondary(void)
 	set_cpu_sibling_map(cpu);
 	set_cpu_core_map(cpu);
 
+	cpuhp_ap_sync_alive();
 	notify_cpu_starting(cpu);
 
-	/* Notify boot CPU that we're starting */
-	complete(&cpu_starting);
-
 	/* The CPU is running, now mark it online */
 	set_cpu_online(cpu, true);
 
 	calculate_cpu_foreign_map();
 
-	/*
-	 * Notify boot CPU that we're up & online and it can safely return
-	 * from __cpu_up()
-	 */
-	complete(&cpu_running);
-
 	/*
 	 * irq will be enabled in loongson_smp_finish(), enabling it too
 	 * early is dangerous.

-- 
2.45.2

Re: [PATCH 1/3] cpu/hotplug: Make HOTPLUG_PARALLEL independent of HOTPLUG_SMT

[Thomas Gleixner]

On Fri, Jul 05 2024 at 11:29, Jiaxun Yang wrote:
>  
> +#ifdef CONFIG_HOTPLUG_SMT
>  static inline bool cpuhp_smt_aware(void)
>  {
>  	return cpu_smt_max_threads > 1;
> @@ -1817,6 +1818,7 @@ static inline const struct cpumask *cpuhp_get_primary_thread_mask(void)
>  {
>  	return cpu_primary_thread_mask;
>  }
> +#endif

Please provide stub functions in an #else clause

#else
static inline bool cpuhp_smt_aware(void) { return false; }
static inline cpumask *cpuhp_get_primary_thread_mask(void) { return NULL; }
#endif

or something to that effect.

>  /*
>   * On architectures which have enabled parallel bringup this invokes all BP
> @@ -1837,6 +1839,7 @@ static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
>  	if (!__cpuhp_parallel_bringup)
>  		return false;
>  
> +#ifdef CONFIG_HOTPLUG_SMT

That way you spare the ugly #ifdef here in the middle of the code and
the compiler can optimize it out.

>  	if (cpuhp_smt_aware()) {
>  		const struct cpumask *pmask = cpuhp_get_primary_thread_mask();
>  		static struct cpumask tmp_mask __initdata;
> @@ -1857,6 +1860,7 @@ static bool __init cpuhp_bringup_cpus_parallel(unsigned int ncpus)
>  		cpumask_andnot(&tmp_mask, mask, pmask);
>  		mask = &tmp_mask;
>  	}
> +#endif

Thanks,

        tglx