um: Add initial SMP support

Add initial symmetric multi-processing (SMP) support to UML. With
this support enabled, users can tell UML to start multiple virtual
processors, each represented as a separate host thread.

In UML, kthreads and normal threads (when running in kernel mode)
can be scheduled and executed simultaneously on different virtual
processors. However, the userspace code of normal threads still
runs within their respective single-threaded stubs.

That is, SMP support is currently available both within the kernel
and across different processes, but still remains limited within
threads of the same process in userspace.

Signed-off-by: Tiwei Bie <tiwei.btw@antgroup.com>
Link: https://patch.msgid.link/20251027001815.1666872-6-tiwei.bie@linux.dev
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

9 files changed, 342 insertions, 15 deletions

diff --git a/arch/um/kernel/Makefile b/arch/um/kernel/Makefile
index b8f4e9281599..be60bc451b3f 100644
--- a/arch/um/kernel/Makefile
+++ b/arch/um/kernel/Makefile
@@ -25,6 +25,7 @@ obj-$(CONFIG_GPROF)	+= gprof_syms.o
 obj-$(CONFIG_OF) += dtb.o
 obj-$(CONFIG_EARLY_PRINTK) += early_printk.o
 obj-$(CONFIG_STACKTRACE) += stacktrace.o
+obj-$(CONFIG_SMP) += smp.o
 
 USER_OBJS := config.o
 
diff --git a/arch/um/kernel/irq.c b/arch/um/kernel/irq.c
index e95f6c5a259d..f4b13f15a9c1 100644
--- a/arch/um/kernel/irq.c
+++ b/arch/um/kernel/irq.c
@@ -22,6 +22,9 @@
 #include <irq_kern.h>
 #include <linux/time-internal.h>
 
+DEFINE_PER_CPU_SHARED_ALIGNED(irq_cpustat_t, irq_stat);
+
+#define irq_stats(x)		(&per_cpu(irq_stat, x))
 
 /* When epoll triggers we do not know why it did so
  * we can also have different IRQs for read and write.
@@ -701,3 +704,25 @@ void sigchld_handler(int sig, struct siginfo *unused_si,
 {
 	do_IRQ(SIGCHLD_IRQ, regs);
 }
+
+/*
+ * /proc/interrupts printing for arch specific interrupts
+ */
+int arch_show_interrupts(struct seq_file *p, int prec)
+{
+#if IS_ENABLED(CONFIG_SMP)
+	int cpu;
+
+	seq_printf(p, "%*s: ", prec, "RES");
+	for_each_online_cpu(cpu)
+		seq_printf(p, "%10u ", irq_stats(cpu)->irq_resched_count);
+	seq_puts(p, "  Rescheduling interrupts\n");
+
+	seq_printf(p, "%*s: ", prec, "CAL");
+	for_each_online_cpu(cpu)
+		seq_printf(p, "%10u ", irq_stats(cpu)->irq_call_count);
+	seq_puts(p, "  Function call interrupts\n");
+#endif
+
+	return 0;
+}
diff --git a/arch/um/kernel/process.c b/arch/um/kernel/process.c
index 3b28048f269c..63b38a3f73f7 100644
--- a/arch/um/kernel/process.c
+++ b/arch/um/kernel/process.c
@@ -218,6 +218,11 @@ void arch_cpu_idle(void)
 	um_idle_sleep();
 }
 
+void arch_cpu_idle_prepare(void)
+{
+	os_idle_prepare();
+}
+
 int __uml_cant_sleep(void) {
 	return in_atomic() || irqs_disabled() || in_interrupt();
 	/* Is in_interrupt() really needed? */
diff --git a/arch/um/kernel/skas/mmu.c b/arch/um/kernel/skas/mmu.c
index afe9a2f251ef..00957788591b 100644
--- a/arch/um/kernel/skas/mmu.c
+++ b/arch/um/kernel/skas/mmu.c
@@ -23,17 +23,36 @@ static_assert(sizeof(struct stub_data) == STUB_DATA_PAGES * UM_KERN_PAGE_SIZE);
 static spinlock_t mm_list_lock;
 static struct list_head mm_list;
 
+void enter_turnstile(struct mm_id *mm_id) __acquires(turnstile)
+{
+	struct mm_context *ctx = container_of(mm_id, struct mm_context, id);
+
+	mutex_lock(&ctx->turnstile);
+}
+
+void exit_turnstile(struct mm_id *mm_id) __releases(turnstile)
+{
+	struct mm_context *ctx = container_of(mm_id, struct mm_context, id);
+
+	mutex_unlock(&ctx->turnstile);
+}
+
 int init_new_context(struct task_struct *task, struct mm_struct *mm)
 {
 	struct mm_id *new_id = &mm->context.id;
 	unsigned long stack = 0;
 	int ret = -ENOMEM;
 
+	mutex_init(&mm->context.turnstile);
+	spin_lock_init(&mm->context.sync_tlb_lock);
+
 	stack = __get_free_pages(GFP_KERNEL | __GFP_ZERO, ilog2(STUB_DATA_PAGES));
 	if (stack == 0)
 		goto out;
 
 	new_id->stack = stack;
+	new_id->syscall_data_len = 0;
+	new_id->syscall_fd_num = 0;
 
 	scoped_guard(spinlock_irqsave, &mm_list_lock) {
 		/* Insert into list, used for lookups when the child dies */
@@ -73,6 +92,9 @@ void destroy_context(struct mm_struct *mm)
 		return;
 	}
 
+	scoped_guard(spinlock_irqsave, &mm_list_lock)
+		list_del(&mm->context.list);
+
 	if (mmu->id.pid > 0) {
 		os_kill_ptraced_process(mmu->id.pid, 1);
 		mmu->id.pid = -1;
@@ -82,10 +104,6 @@ void destroy_context(struct mm_struct *mm)
 		os_close_file(mmu->id.sock);
 
 	free_pages(mmu->id.stack, ilog2(STUB_DATA_PAGES));
-
-	guard(spinlock_irqsave)(&mm_list_lock);
-
-	list_del(&mm->context.list);
 }
 
 static irqreturn_t mm_sigchld_irq(int irq, void* dev)
@@ -110,12 +128,11 @@ static irqreturn_t mm_sigchld_irq(int irq, void* dev)
 				/* Marks the MM as dead */
 				mm_context->id.pid = -1;
 
-				/*
-				 * NOTE: If SMP is implemented, a futex_wake
-				 * needs to be added here.
-				 */
 				stub_data = (void *)mm_context->id.stack;
 				stub_data->futex = FUTEX_IN_KERN;
+#if IS_ENABLED(CONFIG_SMP)
+				os_futex_wake(&stub_data->futex);
+#endif
 
 				/*
 				 * NOTE: Currently executing syscalls by
diff --git a/arch/um/kernel/skas/process.c b/arch/um/kernel/skas/process.c
index 5881b17eb987..4a7673b0261a 100644
--- a/arch/um/kernel/skas/process.c
+++ b/arch/um/kernel/skas/process.c
@@ -7,6 +7,7 @@
 #include <linux/sched/mm.h>
 #include <linux/sched/task_stack.h>
 #include <linux/sched/task.h>
+#include <linux/smp-internal.h>
 
 #include <asm/tlbflush.h>
 
@@ -26,12 +27,12 @@ static int __init start_kernel_proc(void *unused)
 	return 0;
 }
 
-static char cpu0_irqstack[THREAD_SIZE] __aligned(THREAD_SIZE);
+char cpu_irqstacks[NR_CPUS][THREAD_SIZE] __aligned(THREAD_SIZE);
 
 int __init start_uml(void)
 {
-	stack_protections((unsigned long) &cpu0_irqstack);
-	set_sigstack(cpu0_irqstack, THREAD_SIZE);
+	stack_protections((unsigned long) &cpu_irqstacks[0]);
+	set_sigstack(cpu_irqstacks[0], THREAD_SIZE);
 
 	init_new_thread_signals();
 
@@ -64,3 +65,15 @@ void current_mm_sync(void)
 
 	um_tlb_sync(current->mm);
 }
+
+static DEFINE_SPINLOCK(initial_jmpbuf_spinlock);
+
+void initial_jmpbuf_lock(void)
+{
+	spin_lock_irq(&initial_jmpbuf_spinlock);
+}
+
+void initial_jmpbuf_unlock(void)
+{
+	spin_unlock_irq(&initial_jmpbuf_spinlock);
+}
diff --git a/arch/um/kernel/smp.c b/arch/um/kernel/smp.c
new file mode 100644
index 000000000000..f1e52b7348fb
--- /dev/null
+++ b/arch/um/kernel/smp.c
@@ -0,0 +1,242 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2025 Ant Group
+ * Author: Tiwei Bie <tiwei.btw@antgroup.com>
+ *
+ * Based on the previous implementation in TT mode
+ * Copyright (C) 2000 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
+ */
+
+#include <linux/sched.h>
+#include <linux/sched/task.h>
+#include <linux/sched/task_stack.h>
+#include <linux/module.h>
+#include <linux/processor.h>
+#include <linux/threads.h>
+#include <linux/cpu.h>
+#include <linux/hardirq.h>
+#include <linux/smp.h>
+#include <linux/smp-internal.h>
+#include <init.h>
+#include <kern.h>
+#include <os.h>
+#include <smp.h>
+
+enum {
+	UML_IPI_RES = 0,
+	UML_IPI_CALL_SINGLE,
+	UML_IPI_CALL,
+	UML_IPI_STOP,
+};
+
+void arch_smp_send_reschedule(int cpu)
+{
+	os_send_ipi(cpu, UML_IPI_RES);
+}
+
+void arch_send_call_function_single_ipi(int cpu)
+{
+	os_send_ipi(cpu, UML_IPI_CALL_SINGLE);
+}
+
+void arch_send_call_function_ipi_mask(const struct cpumask *mask)
+{
+	int cpu;
+
+	for_each_cpu(cpu, mask)
+		os_send_ipi(cpu, UML_IPI_CALL);
+}
+
+void smp_send_stop(void)
+{
+	int cpu, me = smp_processor_id();
+
+	for_each_online_cpu(cpu) {
+		if (cpu == me)
+			continue;
+		os_send_ipi(cpu, UML_IPI_STOP);
+	}
+}
+
+static void ipi_handler(int vector, struct uml_pt_regs *regs)
+{
+	struct pt_regs *old_regs = set_irq_regs((struct pt_regs *)regs);
+	int cpu = raw_smp_processor_id();
+
+	irq_enter();
+
+	if (current->mm)
+		os_alarm_process(current->mm->context.id.pid);
+
+	switch (vector) {
+	case UML_IPI_RES:
+		inc_irq_stat(irq_resched_count);
+		scheduler_ipi();
+		break;
+
+	case UML_IPI_CALL_SINGLE:
+		inc_irq_stat(irq_call_count);
+		generic_smp_call_function_single_interrupt();
+		break;
+
+	case UML_IPI_CALL:
+		inc_irq_stat(irq_call_count);
+		generic_smp_call_function_interrupt();
+		break;
+
+	case UML_IPI_STOP:
+		set_cpu_online(cpu, false);
+		while (1)
+			pause();
+		break;
+
+	default:
+		pr_err("CPU#%d received unknown IPI (vector=%d)!\n", cpu, vector);
+		break;
+	}
+
+	irq_exit();
+	set_irq_regs(old_regs);
+}
+
+void uml_ipi_handler(int vector)
+{
+	struct uml_pt_regs r = { .is_user = 0 };
+
+	preempt_disable();
+	ipi_handler(vector, &r);
+	preempt_enable();
+}
+
+/* AP states used only during CPU startup */
+enum {
+	UML_CPU_PAUSED = 0,
+	UML_CPU_RUNNING,
+};
+
+static int cpu_states[NR_CPUS];
+
+static int start_secondary(void *unused)
+{
+	int err, cpu = raw_smp_processor_id();
+
+	notify_cpu_starting(cpu);
+	set_cpu_online(cpu, true);
+
+	err = um_setup_timer();
+	if (err)
+		panic("CPU#%d failed to setup timer, err = %d", cpu, err);
+
+	local_irq_enable();
+
+	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
+
+	return 0;
+}
+
+void uml_start_secondary(void *opaque)
+{
+	int cpu = raw_smp_processor_id();
+	struct mm_struct *mm = &init_mm;
+	struct task_struct *idle;
+
+	stack_protections((unsigned long) &cpu_irqstacks[cpu]);
+	set_sigstack(&cpu_irqstacks[cpu], THREAD_SIZE);
+
+	set_cpu_present(cpu, true);
+	os_futex_wait(&cpu_states[cpu], UML_CPU_PAUSED);
+
+	smp_rmb(); /* paired with smp_wmb() in __cpu_up() */
+
+	idle = cpu_tasks[cpu];
+	idle->thread_info.cpu = cpu;
+
+	mmgrab(mm);
+	idle->active_mm = mm;
+
+	idle->thread.request.thread.proc = start_secondary;
+	idle->thread.request.thread.arg = NULL;
+
+	new_thread(task_stack_page(idle), &idle->thread.switch_buf,
+		   new_thread_handler);
+	os_start_secondary(opaque, &idle->thread.switch_buf);
+}
+
+void __init smp_prepare_cpus(unsigned int max_cpus)
+{
+	int err, cpu, me = smp_processor_id();
+	unsigned long deadline;
+
+	os_init_smp();
+
+	for_each_possible_cpu(cpu) {
+		if (cpu == me)
+			continue;
+
+		pr_debug("Booting processor %d...\n", cpu);
+		err = os_start_cpu_thread(cpu);
+		if (err) {
+			pr_crit("CPU#%d failed to start cpu thread, err = %d",
+				cpu, err);
+			continue;
+		}
+
+		deadline = jiffies + msecs_to_jiffies(1000);
+		spin_until_cond(cpu_present(cpu) ||
+				time_is_before_jiffies(deadline));
+
+		if (!cpu_present(cpu))
+			pr_crit("CPU#%d failed to boot\n", cpu);
+	}
+}
+
+int __cpu_up(unsigned int cpu, struct task_struct *tidle)
+{
+	cpu_tasks[cpu] = tidle;
+	smp_wmb(); /* paired with smp_rmb() in uml_start_secondary() */
+	cpu_states[cpu] = UML_CPU_RUNNING;
+	os_futex_wake(&cpu_states[cpu]);
+	spin_until_cond(cpu_online(cpu));
+
+	return 0;
+}
+
+void __init smp_cpus_done(unsigned int max_cpus)
+{
+}
+
+/* Set in uml_ncpus_setup */
+int uml_ncpus = 1;
+
+void __init prefill_possible_map(void)
+{
+	int cpu;
+
+	for (cpu = 0; cpu < uml_ncpus; cpu++)
+		set_cpu_possible(cpu, true);
+	for (; cpu < NR_CPUS; cpu++)
+		set_cpu_possible(cpu, false);
+}
+
+static int __init uml_ncpus_setup(char *line, int *add)
+{
+	*add = 0;
+
+	if (kstrtoint(line, 10, &uml_ncpus)) {
+		os_warn("%s: Couldn't parse '%s'\n", __func__, line);
+		return -1;
+	}
+
+	uml_ncpus = clamp(uml_ncpus, 1, NR_CPUS);
+
+	return 0;
+}
+
+__uml_setup("ncpus=", uml_ncpus_setup,
+"ncpus=<# of desired CPUs>\n"
+"    This tells UML how many virtual processors to start. The maximum\n"
+"    number of supported virtual processors can be obtained by querying\n"
+"    the CONFIG_NR_CPUS option using --showconfig.\n\n"
+);
+
+EXPORT_SYMBOL(uml_curr_cpu);
diff --git a/arch/um/kernel/tlb.c b/arch/um/kernel/tlb.c
index cf7e0d4407f2..39608cccf2c6 100644
--- a/arch/um/kernel/tlb.c
+++ b/arch/um/kernel/tlb.c
@@ -162,9 +162,11 @@ int um_tlb_sync(struct mm_struct *mm)
 {
 	pgd_t *pgd;
 	struct vm_ops ops;
-	unsigned long addr = mm->context.sync_tlb_range_from, next;
+	unsigned long addr, next;
 	int ret = 0;
 
+	guard(spinlock_irqsave)(&mm->context.sync_tlb_lock);
+
 	if (mm->context.sync_tlb_range_to == 0)
 		return 0;
 
@@ -177,6 +179,7 @@ int um_tlb_sync(struct mm_struct *mm)
 		ops.unmap = unmap;
 	}
 
+	addr = mm->context.sync_tlb_range_from;
 	pgd = pgd_offset(mm, addr);
 	do {
 		next = pgd_addr_end(addr, mm->context.sync_tlb_range_to);
diff --git a/arch/um/kernel/trap.c b/arch/um/kernel/trap.c
index 5b80a3a89c20..177615820a4c 100644
--- a/arch/um/kernel/trap.c
+++ b/arch/um/kernel/trap.c
@@ -316,7 +316,7 @@ unsigned long segv(struct faultinfo fi, unsigned long ip, int is_user,
 	if (!is_user && regs)
 		current->thread.segv_regs = container_of(regs, struct pt_regs, regs);
 
-	if (!is_user && init_mm.context.sync_tlb_range_to) {
+	if (!is_user && address >= start_vm && address < end_vm) {
 		/*
 		 * Kernel has pending updates from set_ptes that were not
 		 * flushed yet. Syncing them should fix the pagefault (if not
diff --git a/arch/um/kernel/um_arch.c b/arch/um/kernel/um_arch.c
index cf06bb732ed8..e2b24e1ecfa6 100644
--- a/arch/um/kernel/um_arch.c
+++ b/arch/um/kernel/um_arch.c
@@ -19,6 +19,7 @@
 #include <linux/kmsg_dump.h>
 #include <linux/suspend.h>
 #include <linux/random.h>
+#include <linux/smp-internal.h>
 
 #include <asm/processor.h>
 #include <asm/cpufeature.h>
@@ -71,6 +72,12 @@ static int show_cpuinfo(struct seq_file *m, void *v)
 {
 	int i = 0;
 
+#if IS_ENABLED(CONFIG_SMP)
+	i = (uintptr_t) v - 1;
+	if (!cpu_online(i))
+		return 0;
+#endif
+
 	seq_printf(m, "processor\t: %d\n", i);
 	seq_printf(m, "vendor_id\t: User Mode Linux\n");
 	seq_printf(m, "model name\t: UML\n");
@@ -87,13 +94,14 @@ static int show_cpuinfo(struct seq_file *m, void *v)
 		   loops_per_jiffy/(500000/HZ),
 		   (loops_per_jiffy/(5000/HZ)) % 100);
 
-
 	return 0;
 }
 
 static void *c_start(struct seq_file *m, loff_t *pos)
 {
-	return *pos < nr_cpu_ids ? &boot_cpu_data + *pos : NULL;
+	if (*pos < nr_cpu_ids)
+		return (void *)(uintptr_t)(*pos + 1);
+	return NULL;
 }
 
 static void *c_next(struct seq_file *m, void *v, loff_t *pos)
@@ -409,6 +417,7 @@ void __init setup_arch(char **cmdline_p)
 	strscpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
 	*cmdline_p = command_line;
 	setup_hostinfo(host_info, sizeof host_info);
+	prefill_possible_map();
 
 	if (os_getrandom(rng_seed, sizeof(rng_seed), 0) == sizeof(rng_seed)) {
 		add_bootloader_randomness(rng_seed, sizeof(rng_seed));
@@ -443,6 +452,18 @@ void apply_alternatives(struct alt_instr *start, struct alt_instr *end)
 {
 }
 
+#if IS_ENABLED(CONFIG_SMP)
+void alternatives_smp_module_add(struct module *mod, char *name,
+				 void *locks, void *locks_end,
+				 void *text,  void *text_end)
+{
+}
+
+void alternatives_smp_module_del(struct module *mod)
+{
+}
+#endif
+
 void *text_poke(void *addr, const void *opcode, size_t len)
 {
 	/*