1 /* 2 * Xen SMP support 3 * 4 * This file implements the Xen versions of smp_ops. SMP under Xen is 5 * very straightforward. Bringing a CPU up is simply a matter of 6 * loading its initial context and setting it running. 7 * 8 * IPIs are handled through the Xen event mechanism. 9 * 10 * Because virtual CPUs can be scheduled onto any real CPU, there's no 11 * useful topology information for the kernel to make use of. As a 12 * result, all CPUs are treated as if they're single-core and 13 * single-threaded. 14 */ 15 #include <linux/sched.h> 16 #include <linux/err.h> 17 #include <linux/slab.h> 18 #include <linux/smp.h> 19 20 #include <asm/paravirt.h> 21 #include <asm/desc.h> 22 #include <asm/pgtable.h> 23 #include <asm/cpu.h> 24 25 #include <xen/interface/xen.h> 26 #include <xen/interface/vcpu.h> 27 28 #include <asm/xen/interface.h> 29 #include <asm/xen/hypercall.h> 30 31 #include <xen/page.h> 32 #include <xen/events.h> 33 34 #include "xen-ops.h" 35 #include "mmu.h" 36 37 cpumask_var_t xen_cpu_initialized_map; 38 39 static DEFINE_PER_CPU(int, xen_resched_irq); 40 static DEFINE_PER_CPU(int, xen_callfunc_irq); 41 static DEFINE_PER_CPU(int, xen_callfuncsingle_irq); 42 static DEFINE_PER_CPU(int, xen_debug_irq) = -1; 43 44 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id); 45 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id); 46 47 /* 48 * Reschedule call back. Nothing to do, 49 * all the work is done automatically when 50 * we return from the interrupt. 51 */ 52 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id) 53 { 54 inc_irq_stat(irq_resched_count); 55 56 return IRQ_HANDLED; 57 } 58 59 static __cpuinit void cpu_bringup(void) 60 { 61 int cpu = smp_processor_id(); 62 63 cpu_init(); 64 touch_softlockup_watchdog(); 65 preempt_disable(); 66 67 xen_enable_sysenter(); 68 xen_enable_syscall(); 69 70 cpu = smp_processor_id(); 71 smp_store_cpu_info(cpu); 72 cpu_data(cpu).x86_max_cores = 1; 73 set_cpu_sibling_map(cpu); 74 75 xen_setup_cpu_clockevents(); 76 77 set_cpu_online(cpu, true); 78 percpu_write(cpu_state, CPU_ONLINE); 79 wmb(); 80 81 /* We can take interrupts now: we're officially "up". */ 82 local_irq_enable(); 83 84 wmb(); /* make sure everything is out */ 85 } 86 87 static __cpuinit void cpu_bringup_and_idle(void) 88 { 89 cpu_bringup(); 90 cpu_idle(); 91 } 92 93 static int xen_smp_intr_init(unsigned int cpu) 94 { 95 int rc; 96 const char *resched_name, *callfunc_name, *debug_name; 97 98 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu); 99 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR, 100 cpu, 101 xen_reschedule_interrupt, 102 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 103 resched_name, 104 NULL); 105 if (rc < 0) 106 goto fail; 107 per_cpu(xen_resched_irq, cpu) = rc; 108 109 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu); 110 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR, 111 cpu, 112 xen_call_function_interrupt, 113 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 114 callfunc_name, 115 NULL); 116 if (rc < 0) 117 goto fail; 118 per_cpu(xen_callfunc_irq, cpu) = rc; 119 120 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu); 121 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt, 122 IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING, 123 debug_name, NULL); 124 if (rc < 0) 125 goto fail; 126 per_cpu(xen_debug_irq, cpu) = rc; 127 128 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu); 129 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR, 130 cpu, 131 xen_call_function_single_interrupt, 132 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 133 callfunc_name, 134 NULL); 135 if (rc < 0) 136 goto fail; 137 per_cpu(xen_callfuncsingle_irq, cpu) = rc; 138 139 return 0; 140 141 fail: 142 if (per_cpu(xen_resched_irq, cpu) >= 0) 143 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL); 144 if (per_cpu(xen_callfunc_irq, cpu) >= 0) 145 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL); 146 if (per_cpu(xen_debug_irq, cpu) >= 0) 147 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL); 148 if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0) 149 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), 150 NULL); 151 152 return rc; 153 } 154 155 static void __init xen_fill_possible_map(void) 156 { 157 int i, rc; 158 159 for (i = 0; i < nr_cpu_ids; i++) { 160 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); 161 if (rc >= 0) { 162 num_processors++; 163 set_cpu_possible(i, true); 164 } 165 } 166 } 167 168 static void __init xen_smp_prepare_boot_cpu(void) 169 { 170 BUG_ON(smp_processor_id() != 0); 171 native_smp_prepare_boot_cpu(); 172 173 /* We've switched to the "real" per-cpu gdt, so make sure the 174 old memory can be recycled */ 175 make_lowmem_page_readwrite(xen_initial_gdt); 176 177 xen_setup_vcpu_info_placement(); 178 } 179 180 static void __init xen_smp_prepare_cpus(unsigned int max_cpus) 181 { 182 unsigned cpu; 183 184 xen_init_lock_cpu(0); 185 186 smp_store_cpu_info(0); 187 cpu_data(0).x86_max_cores = 1; 188 set_cpu_sibling_map(0); 189 190 if (xen_smp_intr_init(0)) 191 BUG(); 192 193 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL)) 194 panic("could not allocate xen_cpu_initialized_map\n"); 195 196 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0)); 197 198 /* Restrict the possible_map according to max_cpus. */ 199 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { 200 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--) 201 continue; 202 set_cpu_possible(cpu, false); 203 } 204 205 for_each_possible_cpu (cpu) { 206 struct task_struct *idle; 207 208 if (cpu == 0) 209 continue; 210 211 idle = fork_idle(cpu); 212 if (IS_ERR(idle)) 213 panic("failed fork for CPU %d", cpu); 214 215 set_cpu_present(cpu, true); 216 } 217 } 218 219 static __cpuinit int 220 cpu_initialize_context(unsigned int cpu, struct task_struct *idle) 221 { 222 struct vcpu_guest_context *ctxt; 223 struct desc_struct *gdt; 224 unsigned long gdt_mfn; 225 226 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map)) 227 return 0; 228 229 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); 230 if (ctxt == NULL) 231 return -ENOMEM; 232 233 gdt = get_cpu_gdt_table(cpu); 234 235 ctxt->flags = VGCF_IN_KERNEL; 236 ctxt->user_regs.ds = __USER_DS; 237 ctxt->user_regs.es = __USER_DS; 238 ctxt->user_regs.ss = __KERNEL_DS; 239 #ifdef CONFIG_X86_32 240 ctxt->user_regs.fs = __KERNEL_PERCPU; 241 ctxt->user_regs.gs = __KERNEL_STACK_CANARY; 242 #else 243 ctxt->gs_base_kernel = per_cpu_offset(cpu); 244 #endif 245 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle; 246 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ 247 248 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt)); 249 250 xen_copy_trap_info(ctxt->trap_ctxt); 251 252 ctxt->ldt_ents = 0; 253 254 BUG_ON((unsigned long)gdt & ~PAGE_MASK); 255 256 gdt_mfn = arbitrary_virt_to_mfn(gdt); 257 make_lowmem_page_readonly(gdt); 258 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn)); 259 260 ctxt->gdt_frames[0] = gdt_mfn; 261 ctxt->gdt_ents = GDT_ENTRIES; 262 263 ctxt->user_regs.cs = __KERNEL_CS; 264 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs); 265 266 ctxt->kernel_ss = __KERNEL_DS; 267 ctxt->kernel_sp = idle->thread.sp0; 268 269 #ifdef CONFIG_X86_32 270 ctxt->event_callback_cs = __KERNEL_CS; 271 ctxt->failsafe_callback_cs = __KERNEL_CS; 272 #endif 273 ctxt->event_callback_eip = (unsigned long)xen_hypervisor_callback; 274 ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback; 275 276 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); 277 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir)); 278 279 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt)) 280 BUG(); 281 282 kfree(ctxt); 283 return 0; 284 } 285 286 static int __cpuinit xen_cpu_up(unsigned int cpu) 287 { 288 struct task_struct *idle = idle_task(cpu); 289 int rc; 290 291 per_cpu(current_task, cpu) = idle; 292 #ifdef CONFIG_X86_32 293 irq_ctx_init(cpu); 294 #else 295 clear_tsk_thread_flag(idle, TIF_FORK); 296 per_cpu(kernel_stack, cpu) = 297 (unsigned long)task_stack_page(idle) - 298 KERNEL_STACK_OFFSET + THREAD_SIZE; 299 #endif 300 xen_setup_runstate_info(cpu); 301 xen_setup_timer(cpu); 302 xen_init_lock_cpu(cpu); 303 304 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE; 305 306 /* make sure interrupts start blocked */ 307 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; 308 309 rc = cpu_initialize_context(cpu, idle); 310 if (rc) 311 return rc; 312 313 if (num_online_cpus() == 1) 314 alternatives_smp_switch(1); 315 316 rc = xen_smp_intr_init(cpu); 317 if (rc) 318 return rc; 319 320 rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL); 321 BUG_ON(rc); 322 323 while(per_cpu(cpu_state, cpu) != CPU_ONLINE) { 324 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 325 barrier(); 326 } 327 328 return 0; 329 } 330 331 static void xen_smp_cpus_done(unsigned int max_cpus) 332 { 333 } 334 335 #ifdef CONFIG_HOTPLUG_CPU 336 static int xen_cpu_disable(void) 337 { 338 unsigned int cpu = smp_processor_id(); 339 if (cpu == 0) 340 return -EBUSY; 341 342 cpu_disable_common(); 343 344 load_cr3(swapper_pg_dir); 345 return 0; 346 } 347 348 static void xen_cpu_die(unsigned int cpu) 349 { 350 while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) { 351 current->state = TASK_UNINTERRUPTIBLE; 352 schedule_timeout(HZ/10); 353 } 354 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL); 355 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL); 356 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL); 357 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL); 358 xen_uninit_lock_cpu(cpu); 359 xen_teardown_timer(cpu); 360 361 if (num_online_cpus() == 1) 362 alternatives_smp_switch(0); 363 } 364 365 static void __cpuinit xen_play_dead(void) /* used only with HOTPLUG_CPU */ 366 { 367 play_dead_common(); 368 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL); 369 cpu_bringup(); 370 } 371 372 #else /* !CONFIG_HOTPLUG_CPU */ 373 static int xen_cpu_disable(void) 374 { 375 return -ENOSYS; 376 } 377 378 static void xen_cpu_die(unsigned int cpu) 379 { 380 BUG(); 381 } 382 383 static void xen_play_dead(void) 384 { 385 BUG(); 386 } 387 388 #endif 389 static void stop_self(void *v) 390 { 391 int cpu = smp_processor_id(); 392 393 /* make sure we're not pinning something down */ 394 load_cr3(swapper_pg_dir); 395 /* should set up a minimal gdt */ 396 397 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL); 398 BUG(); 399 } 400 401 static void xen_smp_send_stop(void) 402 { 403 smp_call_function(stop_self, NULL, 0); 404 } 405 406 static void xen_smp_send_reschedule(int cpu) 407 { 408 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR); 409 } 410 411 static void xen_send_IPI_mask(const struct cpumask *mask, 412 enum ipi_vector vector) 413 { 414 unsigned cpu; 415 416 for_each_cpu_and(cpu, mask, cpu_online_mask) 417 xen_send_IPI_one(cpu, vector); 418 } 419 420 static void xen_smp_send_call_function_ipi(const struct cpumask *mask) 421 { 422 int cpu; 423 424 xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR); 425 426 /* Make sure other vcpus get a chance to run if they need to. */ 427 for_each_cpu(cpu, mask) { 428 if (xen_vcpu_stolen(cpu)) { 429 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 430 break; 431 } 432 } 433 } 434 435 static void xen_smp_send_call_function_single_ipi(int cpu) 436 { 437 xen_send_IPI_mask(cpumask_of(cpu), 438 XEN_CALL_FUNCTION_SINGLE_VECTOR); 439 } 440 441 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id) 442 { 443 irq_enter(); 444 generic_smp_call_function_interrupt(); 445 inc_irq_stat(irq_call_count); 446 irq_exit(); 447 448 return IRQ_HANDLED; 449 } 450 451 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id) 452 { 453 irq_enter(); 454 generic_smp_call_function_single_interrupt(); 455 inc_irq_stat(irq_call_count); 456 irq_exit(); 457 458 return IRQ_HANDLED; 459 } 460 461 static const struct smp_ops xen_smp_ops __initdata = { 462 .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu, 463 .smp_prepare_cpus = xen_smp_prepare_cpus, 464 .smp_cpus_done = xen_smp_cpus_done, 465 466 .cpu_up = xen_cpu_up, 467 .cpu_die = xen_cpu_die, 468 .cpu_disable = xen_cpu_disable, 469 .play_dead = xen_play_dead, 470 471 .smp_send_stop = xen_smp_send_stop, 472 .smp_send_reschedule = xen_smp_send_reschedule, 473 474 .send_call_func_ipi = xen_smp_send_call_function_ipi, 475 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi, 476 }; 477 478 void __init xen_smp_init(void) 479 { 480 smp_ops = xen_smp_ops; 481 xen_fill_possible_map(); 482 xen_init_spinlocks(); 483 } 484