goodbye smp again

arch/wasm/Kbuild

@@ -1 +1 @@
-obj-y += kernel/ mm/
+obj-y += kernel/ mm/

arch/wasm/Kconfig

@@ -1,34 +1,25 @@
 config WASM
 	def_bool y
-	depends on !BINFMT_ELF && !MMU && !MODULES && !COREDUMP && !SECCOMP && !UPROBES && !COMPAT
+	depends on !SMP && !BINFMT_ELF && !MMU && !MODULES && !COREDUMP && !SECCOMP && !UPROBES && !COMPAT
 	select ARCH_HAS_BINFMT_WASM
 	select ARCH_HAS_SYSCALL_WRAPPER
 	select ARCH_NO_SWAP
 	select ARCH_USE_QUEUED_RWLOCKS
 	select ARCH_USE_QUEUED_SPINLOCKS
 	select BUG
 	select FLATMEM
-	select FORCE_NR_CPUS
+	select FORCE_NR_CPUS if SMP
 	select GENERIC_ATOMIC64
 	select GENERIC_CSUM
 	select GENERIC_HWEIGHT
 	select OF
 	select OF_EARLY_FLATTREE
 	select PAGE_SIZE_64KB
-	select SMP
+	# select SMP
 	select THREAD_INFO_IN_TASK
 	select UACCESS_MEMCPY if !MMU
 	select USE_PER_CPU_TLS
 
-config ARCH_HAS_BINFMT_WASM
-	bool
-
-menu "Executable file formats"
-	config BINFMT_WASM
-		bool "Kernel support for WASM binaries"
-		depends on ARCH_HAS_BINFMT_WASM
-endmenu
-
 config CMDLINE
 	string "Kernel command string to append"
 	default ""
@@ -40,12 +31,12 @@ config EARLY_PRINTK
 	def_bool y
 
 config SMP
-	def_bool y
+	def_bool n
 
 config NR_CPUS
 	int
 	default 1 if !SMP
-	default 8 if SMP
+	default 16 if SMP
 
 config HZ
 	int

arch/wasm/Makefile

@@ -3,7 +3,7 @@ LD = wasm-ld
 WASM_FEATURES = -matomics -mbulk-memory -mexception-handling -mextended-const -mmultivalue -mmutable-globals -mnontrapping-fptoint -mreference-types -mrelaxed-simd -msign-ext -msimd128
 
 cflags-y += -Wno-incompatible-library-redeclaration '-D__builtin_return_address=wasm_return_address' $(WASM_FEATURES)
-ldflags-y += --import-memory --shared-memory --max-memory=4294967296 --no-gc-sections
+ldflags-y += --import-memory --max-memory=4294967296 --shared-memory
 
 KBUILD_CFLAGS += $(cflags-y)
 KBUILD_AFLAGS += $(aflags-y)

arch/wasm/include/asm/current.h

@@ -1,8 +1,10 @@
 #ifndef _WASM_CURRENT_H
 #define _WASM_CURRENT_H
 
+#include <asm/percpu.h>
+
 struct task_struct;
 
-extern _Thread_local struct task_struct * current;
+DECLARE_PER_CPU(struct task_struct*, current);
 
 #endif

arch/wasm/include/asm/processor.h

@@ -6,10 +6,7 @@
 
 struct task_struct;
 
-static inline void cpu_relax(void)
-{
-	wasm_relax();
-}
+void cpu_relax(void);
 
 #define current_text_addr() wasm_return_address(-1)
 

arch/wasm/include/asm/sigcontext.h

@@ -1,7 +1,11 @@
 #ifndef _WASM_SIGCONTEXT_H
 #define _WASM_SIGCONTEXT_H
 
-struct pt_regs {};
+struct pt_regs {
+	void* current_stack;
+	int (*fn)(void*);
+	void* fn_arg;
+};
 
 struct sigcontext {
 	struct pt_regs regs;

arch/wasm/include/asm/smp.h

@@ -1,10 +1,14 @@
 #ifndef _WASM_SMP_H
 #define _WASM_SMP_H
 
+#ifdef CONFIG_SMP
+
 #define raw_smp_processor_id() (current_thread_info()->cpu)
 
 struct cpumask;
 void arch_send_call_function_ipi_mask(const struct cpumask *mask);
 void arch_send_call_function_single_ipi(int cpu);
 
 #endif
+
+#endif

arch/wasm/include/asm/thread_info.h

@@ -12,10 +12,8 @@ struct thread_info {
 	struct task_struct *task;
 	unsigned long flags;
 	int preempt_count;
-	struct task_struct *from_sched;
-#ifdef CONFIG_SMP
+	// struct task_struct *from_sched;
 	unsigned int cpu;
-#endif
 };
 
 #define INIT_THREAD_INFO(tsk)                                                  \

arch/wasm/include/asm/wasm_imports.h

@@ -7,8 +7,6 @@
 #define import(name) __attribute__((import_module("kernel"), import_name(name)))
 
 import("breakpoint") void wasm_breakpoint(void);
-import("relax") void wasm_relax(void);
-import("idle") void wasm_idle(void);
 import("halt") void wasm_halt(void);
 import("restart") void wasm_restart(void);
 
@@ -27,8 +25,6 @@ import("get_now_nsec") unsigned long long wasm_get_now_nsec(void);
 
 import("get_stacktrace") void wasm_get_stacktrace(char *buf, size_t size);
 
-import("new_worker") void wasm_new_worker(void* arg, char name[16]);
-
 #undef import
 
 #endif

arch/wasm/kernel/Makefile

@@ -7,6 +7,7 @@ obj-y += \
 	process.o \
 	ptrace.o \
 	setup.o \
-	smp.o \
 	stacktrace.o \
 	time.o
+
+obj-$(CONFIG_SMP) += smp.o

arch/wasm/kernel/process.c

@@ -1,16 +1,9 @@
-#include "linux/thread_info.h"
-#include <asm/sysmem.h>
 #include <asm/wasm_imports.h>
 #include <linux/slab.h>
 #include <linux/sched.h>
 #include <linux/sched/task_stack.h>
 #include <linux/sched/task.h>
 
-void arch_cpu_idle(void)
-{
-	wasm_idle();
-}
-
 __asm__(".globaltype __stack_pointer, i32\n");
 
 static void *get_stack_pointer(void)
@@ -31,75 +24,47 @@ static void __always_inline set_stack_pointer(void *ptr)
 struct task_struct *__switch_to(struct task_struct *from,
 				struct task_struct *to)
 {
-	// struct thread_info *from_info = task_thread_info(from);
-	struct thread_info *to_info = task_thread_info(to);
-
-	to_info->from_sched = from;
-
-	pr_info("context switch CPU: %i PID: %u -> %u STACK: %p -> %p\n",
-		smp_processor_id(), from->pid, to->pid, get_stack_pointer(),
-		to->stack);
+	struct pt_regs *from_regs = task_pt_regs(from);
+	struct pt_regs *to_regs = task_pt_regs(to);
 
-	set_stack_pointer(to->stack);
-	current = to;
+	from_regs->current_stack = get_stack_pointer();
 
-	return from;
-}
+	pr_info("context switch PID: %u STACK: %p\n"
+		"->                  %u        %p\n",
+		from->pid, from_regs->current_stack, to->pid,
+		to_regs->current_stack);
 
-struct entry_arg {
-	int cpu;
-	struct thread_info *thread_info;
-	int (*fn)(void *);
-	void *fn_arg;
-};
-void show_stack(struct task_struct *task, unsigned long *sp,
-		const char *loglvl);
+	set_stack_pointer(to_regs->current_stack);
 
-int copy_thread(struct task_struct *p, const struct kernel_clone_args *args)
-{
-	struct entry_arg *entry;
-
-	if (!args->fn) panic("can't copy userspace thread"); // yet
+	current = to;
 
-	entry = kmalloc(sizeof(*entry), GFP_KERNEL);
-	if (!entry)
-		return -ENOMEM;
+	schedule_tail(from);
 
-	entry->fn = args->fn;
-	entry->fn_arg = args->fn_arg;
-	entry->thread_info = task_thread_info(p);
+	pr_info("regs: %p\n", to_regs);
+	BUG_ON(!to_regs->fn);
 
-	wasm_new_worker(entry, p->comm);
+	pr_info("about to call %p %p\n", to_regs->fn, to_regs->fn_arg);
+	to_regs->fn(to_regs->fn_arg);
+	to_regs->fn = NULL;
 
-	return 0;
+	return from;
 }
 
-static atomic_t next_cpu_nr = ATOMIC_INIT(1);
-void noinline_for_stack task_entry_inner(struct entry_arg *entry)
+int copy_thread(struct task_struct *p, const struct kernel_clone_args *args)
 {
-	int (*fn)(void *) = entry->fn;
-	void *fn_arg = entry->fn_arg;
-	struct thread_info *thread_info = entry->thread_info;
-	int cpu;
-	kfree(entry);
+	struct pt_regs *childregs = task_pt_regs(p);
 
-	cpu = atomic_inc_return(&next_cpu_nr);
-	smp_tls_init(cpu);
-	current = thread_info->from_sched;
-	current_thread_info()->cpu = cpu;
-	pr_info("== THREAD ENTRY! CPU: %i ==", cpu);
+	memset(childregs, 0, sizeof(struct pt_regs));
 
-	if (thread_info->from_sched)
-		schedule_tail(thread_info->from_sched);
+	childregs->current_stack = childregs;
 
-	fn(fn_arg);
+	if (!args->fn)
+		panic("can't copy userspace thread"); // yet
 
-	do_exit(0);
-}
+	pr_info("copying thread %i %p\n", p->pid, childregs);
 
-__attribute__((export_name("task_entry"))) void
-task_entry(struct entry_arg *entry)
-{
-	set_stack_pointer(entry->thread_info->task->stack);
-	task_entry_inner(entry);
+	childregs->fn = args->fn;
+	childregs->fn_arg = args->fn_arg;
+
+	return 0;
 }

arch/wasm/kernel/setup.c

@@ -39,7 +39,9 @@ __attribute__((export_name("start"))) void __init _start(void)
 {
 	static char wasm_dt[1024];
 
+#ifdef CONFIG_SMP
 	early_tls_init();
+#endif
 	__wasm_call_ctors();
 
 	setup_early_printk(NULL);

arch/wasm/kernel/smp.c

@@ -116,4 +116,4 @@ void handle_ipi(struct pt_regs *regs)
 
 		mb(); /* Order data access and bit testing. */
 	}
-}
+}

arch/wasm/kernel/time.c

@@ -14,10 +14,14 @@ void calibrate_delay(void)
 
 void __delay(unsigned long cycles)
 {
-	cycles_t start = wasm_get_now_nsec();
+	static int zero = 0;
+	int ret = __builtin_wasm_memory_atomic_wait32(&zero, 0, cycles);
+	BUG_ON(ret != 2); // 2 means timeout
+}
 
-	while ((wasm_get_now_nsec() - start) < cycles)
-		cpu_relax();
+void cpu_relax(void)
+{
+	__delay(100);
 }
 
 void __udelay(unsigned long usecs)
@@ -33,6 +37,10 @@ void __const_udelay(unsigned long xloops)
 	__delay(xloops / 0x10c7ul); /* 2**32 / 1000000 (rounded up) */
 }
 
+unsigned long long sched_clock(void) {
+	return wasm_get_now_nsec();
+}
+
 static u64 clock_read(struct clocksource *cs)
 {
 	return wasm_get_now_nsec();
@@ -50,4 +58,4 @@ void __init time_init(void)
 {
 	if (clocksource_register_khz(&clocksource, 1000000))
 		panic("unable to register clocksource\n");
-}
+}

arch/wasm/mm/Makefile

@@ -1 +1,2 @@
-obj-y := init.o page.o tls.o
+obj-y := init.o page.o
+obj-$(CONFIG_SMP) := tls.o

arch/wasm/mm/tls.c

@@ -48,4 +48,4 @@ void smp_tls_init(int cpu)
 {
 	BUG_ON(__per_cpu_offset[cpu] == (void *)-1);
 	__wasm_init_tls((void *)__per_cpu_offset[cpu]);
-}
+}

flake.nix

@@ -6,8 +6,10 @@
     packages = nixpkgs.lib.genAttrs
       [ "x86_64-linux" "aarch64-linux" "x86_64-darwin" "aarch64-darwin" ]
       (system:
-        let pkgs = nixpkgs.legacyPackages.${system}; in {
-          default = pkgs.clangStdenv.mkDerivation {
+        let pkgs = nixpkgs.legacyPackages.${system}; 
+           llvm = pkgs.llvmPackages_17;
+        in {
+          default = llvm.stdenv.mkDerivation {
             pname = "linux";
             version = "6.1.69-wasm";
             src = ./.;
@@ -35,8 +37,8 @@
               pkg-config
               ncurses
 
-              lld
-              libllvm
+              llvm.lld
+              llvm.libllvm
               wabt
               esbuild
             ];

fs/Kconfig.binfmt

@@ -168,6 +168,13 @@ config BINFMT_MISC
 	  you have use for it; the module is called binfmt_misc. If you
 	  don't know what to answer at this point, say Y.
 
+config ARCH_HAS_BINFMT_WASM
+	bool
+
+config BINFMT_WASM
+	bool "Kernel support for WASM binaries"
+	depends on ARCH_HAS_BINFMT_WASM
+
 config COREDUMP
 	bool "Enable core dump support" if EXPERT
 	default y

include/linux/percpu-defs.h

@@ -220,7 +220,7 @@ do {									\
 	(void)__vpp_verify;						\
 } while (0)
 
-#ifdef CONFIG_SMP
+#if defined(CONFIG_SMP) || defined(CONFIG_USE_PER_CPU_TLS)
 
 /*
  * Add an offset to a pointer but keep the pointer as-is.  Use RELOC_HIDE()