feat: fireworks!

src/platform/pc/apic.zig

@@ -142,10 +142,11 @@ pub fn send_sipi(apic_id: u32, startup_page: u8) void {
 }
 
 pub fn send_ipi(apic_id: u32, vector: u8, delivery_mode: u8) void {
-    const icr_value = (@as(u64, apic_id) << 32) | (@as(u64, delivery_mode) << 8) | vector;
     if (in_x2apic_mode.local().*) {
+        const icr_value = (@as(u64, apic_id) << 32) | (@as(u64, delivery_mode) << 8) | vector;
         x2apic_write(.Icr, icr_value);
     } else {
+        const icr_value = (@as(u32, delivery_mode) << 8) | vector;
         // Write to ICR1 first, because writing to the low word causes
         // the IPI to be sent.
         xapic_write(.Icr1, @as(u32, apic_id) << 24);

src/platform/pc/impl.zig

@@ -5,7 +5,7 @@ const std = @import("std");
 const tsc = @import("tsc.zig");
 const pvclock = @import("pvclock.zig");
 pub const apic = @import("apic.zig");
-const smp = @import("smp.zig");
+pub const smp = @import("smp.zig");
 
 const ki = r.ke.private;
 
@@ -105,3 +105,7 @@ pub fn debug_write(c: u8) void {
 pub fn arm_timer(deadline: r.Nanoseconds) void {
     apic.arm_timer(deadline);
 }
+
+pub fn send_resched_ipi(target_cpu: u32) void {
+    apic.send_ipi(smp.cpu_id_to_apic_id[target_cpu], 240, 0);
+}

src/platform/pc/smp.zig

@@ -17,10 +17,12 @@ extern var AP_TRAMPOLINE_DATA: u8;
 extern var __percpu_start: u8;
 extern var __percpu_end: u8;
 
-export var cpu_id_to_apic_id: [config.CONFIG_NCPUS]u32 = undefined;
+pub export var cpu_id_to_apic_id: [config.CONFIG_NCPUS]u32 = undefined;
 
 const start_stack = ke.ExportedCpuLocal(usize, 0, "ap_start_stack");
 
+pub const start_thread = ke.CpuLocal(*ke.Thread, undefined);
+
 var aps_booted: std.atomic.Value(usize) = .init(0);
 
 const ApData = extern struct {
@@ -35,6 +37,16 @@ fn ap_entry(cpu_id: u32) callconv(.c) noreturn {
     ki.impl.init.ap_entry(cpu_id, &aps_booted);
 }
 
+fn make_thread(entrypoint: *const fn (?*anyopaque) void, stack: usize) *ke.Thread {
+    const td = mm.zone.gpa.create(ke.Thread) catch @panic("Failed to allocate thread for AP");
+
+    td.init(stack, r.kib(32), entrypoint, null);
+
+    td.priority = 0;
+    td.priority_class = .Idle;
+    return td;
+}
+
 pub fn init() linksection(r.init) void {
     const trampoline_start = @intFromPtr(&AP_TRAMPOLINE_START);
     const trampoline_size = @intFromPtr(&AP_TRAMPOLINE_END) - trampoline_start;
@@ -94,10 +106,12 @@ pub fn init() linksection(r.init) void {
         // copy the offset into the AP's self_offset variable
         @as(*usize, @ptrCast(@alignCast(&cpu_data[self_offset_offset]))).* = ki.impl.cpu_offsets[cpu_id];
 
-        const stack_top = @intFromPtr(mm.heap.alloc(r.kib(16)) catch @panic("Failed to allocate AP stack")) + r.kib(16);
+        const stack_top = @intFromPtr(mm.heap.alloc(r.kib(32)) catch @panic("Failed to allocate AP stack")) + r.kib(32);
 
         start_stack.remote(@intCast(cpu_id)).* = stack_top & ~@as(usize, 15);
 
+        start_thread.remote(@intCast(cpu_id)).* = make_thread(@ptrCast(&ap_entry), stack_top - r.kib(32));
+
         rtl.barrier.wmb();
 
         // Send the INIT-SIPI-SIPI sequence to start the AP.
@@ -109,6 +123,8 @@ pub fn init() linksection(r.init) void {
         apic.send_sipi(apic_id, 0x08);
     }
 
+    log.info("Waiting for APs...", .{});
+
     while (aps_booted.load(.acquire) < apic.apics.items.len) {
         std.atomic.spinLoopHint();
     }

src/platform/um/impl.zig

@@ -38,6 +38,7 @@ pub const GlobalState = struct {
 
 pub const CpuState = struct {
     pthread: c.pthread_t,
+    td: *ke.Thread,
 };
 
 pub var global_state: GlobalState = undefined;
@@ -122,6 +123,8 @@ fn other_cpu_entry(_: ?*anyopaque) callconv(.c) ?*anyopaque {
 
     ki.cpu.init_cpu(@intCast(my_cpu_id));
 
+    ki.sched.percpu.local().current_thread = percpu.local().td;
+
     timer.init_cpu();
 
     while (true) {}
@@ -174,6 +177,10 @@ pub fn late_init(_: *pl.BootInfo) linksection(r.init) void {
 
         @memcpy(block, @as([*]u8, @ptrCast(&__percpu_start))[0..size]);
 
+        const td = allocator.create(ke.Thread) catch @panic("oom");
+        make_thread(idle, td);
+        percpu.remote(@intCast(i)).td = td;
+
         var attr: c.pthread_attr_t = undefined;
 
         _ = c.pthread_attr_init(&attr);