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 #include <linux/irq_work.h> 20 21 #include <asm/paravirt.h> 22 #include <asm/desc.h> 23 #include <asm/pgtable.h> 24 #include <asm/cpu.h> 25 26 #include <xen/interface/xen.h> 27 #include <xen/interface/vcpu.h> 28 29 #include <asm/xen/interface.h> 30 #include <asm/xen/hypercall.h> 31 32 #include <xen/xen.h> 33 #include <xen/page.h> 34 #include <xen/events.h> 35 36 #include <xen/hvc-console.h> 37 #include "xen-ops.h" 38 #include "mmu.h" 39 40 cpumask_var_t xen_cpu_initialized_map; 41 42 static DEFINE_PER_CPU(int, xen_resched_irq); 43 static DEFINE_PER_CPU(int, xen_callfunc_irq); 44 static DEFINE_PER_CPU(int, xen_callfuncsingle_irq); 45 static DEFINE_PER_CPU(int, xen_irq_work); 46 static DEFINE_PER_CPU(int, xen_debug_irq) = -1; 47 48 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id); 49 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id); 50 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id); 51 52 /* 53 * Reschedule call back. 54 */ 55 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id) 56 { 57 inc_irq_stat(irq_resched_count); 58 scheduler_ipi(); 59 60 return IRQ_HANDLED; 61 } 62 63 static void __cpuinit cpu_bringup(void) 64 { 65 int cpu; 66 67 cpu_init(); 68 touch_softlockup_watchdog(); 69 preempt_disable(); 70 71 xen_enable_sysenter(); 72 xen_enable_syscall(); 73 74 cpu = smp_processor_id(); 75 smp_store_cpu_info(cpu); 76 cpu_data(cpu).x86_max_cores = 1; 77 set_cpu_sibling_map(cpu); 78 79 xen_setup_cpu_clockevents(); 80 81 notify_cpu_starting(cpu); 82 83 set_cpu_online(cpu, true); 84 85 this_cpu_write(cpu_state, CPU_ONLINE); 86 87 wmb(); 88 89 /* We can take interrupts now: we're officially "up". */ 90 local_irq_enable(); 91 92 wmb(); /* make sure everything is out */ 93 } 94 95 static void __cpuinit cpu_bringup_and_idle(void) 96 { 97 cpu_bringup(); 98 cpu_idle(); 99 } 100 101 static int xen_smp_intr_init(unsigned int cpu) 102 { 103 int rc; 104 const char *resched_name, *callfunc_name, *debug_name; 105 106 resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu); 107 rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR, 108 cpu, 109 xen_reschedule_interrupt, 110 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 111 resched_name, 112 NULL); 113 if (rc < 0) 114 goto fail; 115 per_cpu(xen_resched_irq, cpu) = rc; 116 117 callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu); 118 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR, 119 cpu, 120 xen_call_function_interrupt, 121 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 122 callfunc_name, 123 NULL); 124 if (rc < 0) 125 goto fail; 126 per_cpu(xen_callfunc_irq, cpu) = rc; 127 128 debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu); 129 rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt, 130 IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING, 131 debug_name, NULL); 132 if (rc < 0) 133 goto fail; 134 per_cpu(xen_debug_irq, cpu) = rc; 135 136 callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu); 137 rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR, 138 cpu, 139 xen_call_function_single_interrupt, 140 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 141 callfunc_name, 142 NULL); 143 if (rc < 0) 144 goto fail; 145 per_cpu(xen_callfuncsingle_irq, cpu) = rc; 146 147 callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu); 148 rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR, 149 cpu, 150 xen_irq_work_interrupt, 151 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING, 152 callfunc_name, 153 NULL); 154 if (rc < 0) 155 goto fail; 156 per_cpu(xen_irq_work, cpu) = rc; 157 158 return 0; 159 160 fail: 161 if (per_cpu(xen_resched_irq, cpu) >= 0) 162 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL); 163 if (per_cpu(xen_callfunc_irq, cpu) >= 0) 164 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL); 165 if (per_cpu(xen_debug_irq, cpu) >= 0) 166 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL); 167 if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0) 168 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), 169 NULL); 170 if (per_cpu(xen_irq_work, cpu) >= 0) 171 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu), NULL); 172 173 return rc; 174 } 175 176 static void __init xen_fill_possible_map(void) 177 { 178 int i, rc; 179 180 if (xen_initial_domain()) 181 return; 182 183 for (i = 0; i < nr_cpu_ids; i++) { 184 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); 185 if (rc >= 0) { 186 num_processors++; 187 set_cpu_possible(i, true); 188 } 189 } 190 } 191 192 static void __init xen_filter_cpu_maps(void) 193 { 194 int i, rc; 195 unsigned int subtract = 0; 196 197 if (!xen_initial_domain()) 198 return; 199 200 num_processors = 0; 201 disabled_cpus = 0; 202 for (i = 0; i < nr_cpu_ids; i++) { 203 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL); 204 if (rc >= 0) { 205 num_processors++; 206 set_cpu_possible(i, true); 207 } else { 208 set_cpu_possible(i, false); 209 set_cpu_present(i, false); 210 subtract++; 211 } 212 } 213 #ifdef CONFIG_HOTPLUG_CPU 214 /* This is akin to using 'nr_cpus' on the Linux command line. 215 * Which is OK as when we use 'dom0_max_vcpus=X' we can only 216 * have up to X, while nr_cpu_ids is greater than X. This 217 * normally is not a problem, except when CPU hotplugging 218 * is involved and then there might be more than X CPUs 219 * in the guest - which will not work as there is no 220 * hypercall to expand the max number of VCPUs an already 221 * running guest has. So cap it up to X. */ 222 if (subtract) 223 nr_cpu_ids = nr_cpu_ids - subtract; 224 #endif 225 226 } 227 228 static void __init xen_smp_prepare_boot_cpu(void) 229 { 230 BUG_ON(smp_processor_id() != 0); 231 native_smp_prepare_boot_cpu(); 232 233 /* We've switched to the "real" per-cpu gdt, so make sure the 234 old memory can be recycled */ 235 make_lowmem_page_readwrite(xen_initial_gdt); 236 237 xen_filter_cpu_maps(); 238 xen_setup_vcpu_info_placement(); 239 } 240 241 static void __init xen_smp_prepare_cpus(unsigned int max_cpus) 242 { 243 unsigned cpu; 244 unsigned int i; 245 246 if (skip_ioapic_setup) { 247 char *m = (max_cpus == 0) ? 248 "The nosmp parameter is incompatible with Xen; " \ 249 "use Xen dom0_max_vcpus=1 parameter" : 250 "The noapic parameter is incompatible with Xen"; 251 252 xen_raw_printk(m); 253 panic(m); 254 } 255 xen_init_lock_cpu(0); 256 257 smp_store_boot_cpu_info(); 258 cpu_data(0).x86_max_cores = 1; 259 260 for_each_possible_cpu(i) { 261 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL); 262 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL); 263 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL); 264 } 265 set_cpu_sibling_map(0); 266 267 if (xen_smp_intr_init(0)) 268 BUG(); 269 270 if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL)) 271 panic("could not allocate xen_cpu_initialized_map\n"); 272 273 cpumask_copy(xen_cpu_initialized_map, cpumask_of(0)); 274 275 /* Restrict the possible_map according to max_cpus. */ 276 while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) { 277 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--) 278 continue; 279 set_cpu_possible(cpu, false); 280 } 281 282 for_each_possible_cpu(cpu) 283 set_cpu_present(cpu, true); 284 } 285 286 static int __cpuinit 287 cpu_initialize_context(unsigned int cpu, struct task_struct *idle) 288 { 289 struct vcpu_guest_context *ctxt; 290 struct desc_struct *gdt; 291 unsigned long gdt_mfn; 292 293 if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map)) 294 return 0; 295 296 ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL); 297 if (ctxt == NULL) 298 return -ENOMEM; 299 300 gdt = get_cpu_gdt_table(cpu); 301 302 ctxt->flags = VGCF_IN_KERNEL; 303 ctxt->user_regs.ds = __USER_DS; 304 ctxt->user_regs.es = __USER_DS; 305 ctxt->user_regs.ss = __KERNEL_DS; 306 #ifdef CONFIG_X86_32 307 ctxt->user_regs.fs = __KERNEL_PERCPU; 308 ctxt->user_regs.gs = __KERNEL_STACK_CANARY; 309 #else 310 ctxt->gs_base_kernel = per_cpu_offset(cpu); 311 #endif 312 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle; 313 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */ 314 315 memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt)); 316 317 xen_copy_trap_info(ctxt->trap_ctxt); 318 319 ctxt->ldt_ents = 0; 320 321 BUG_ON((unsigned long)gdt & ~PAGE_MASK); 322 323 gdt_mfn = arbitrary_virt_to_mfn(gdt); 324 make_lowmem_page_readonly(gdt); 325 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn)); 326 327 ctxt->gdt_frames[0] = gdt_mfn; 328 ctxt->gdt_ents = GDT_ENTRIES; 329 330 ctxt->user_regs.cs = __KERNEL_CS; 331 ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs); 332 333 ctxt->kernel_ss = __KERNEL_DS; 334 ctxt->kernel_sp = idle->thread.sp0; 335 336 #ifdef CONFIG_X86_32 337 ctxt->event_callback_cs = __KERNEL_CS; 338 ctxt->failsafe_callback_cs = __KERNEL_CS; 339 #endif 340 ctxt->event_callback_eip = (unsigned long)xen_hypervisor_callback; 341 ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback; 342 343 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir); 344 ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir)); 345 346 if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt)) 347 BUG(); 348 349 kfree(ctxt); 350 return 0; 351 } 352 353 static int __cpuinit xen_cpu_up(unsigned int cpu, struct task_struct *idle) 354 { 355 int rc; 356 357 per_cpu(current_task, cpu) = idle; 358 #ifdef CONFIG_X86_32 359 irq_ctx_init(cpu); 360 #else 361 clear_tsk_thread_flag(idle, TIF_FORK); 362 per_cpu(kernel_stack, cpu) = 363 (unsigned long)task_stack_page(idle) - 364 KERNEL_STACK_OFFSET + THREAD_SIZE; 365 #endif 366 xen_setup_runstate_info(cpu); 367 xen_setup_timer(cpu); 368 xen_init_lock_cpu(cpu); 369 370 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE; 371 372 /* make sure interrupts start blocked */ 373 per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1; 374 375 rc = cpu_initialize_context(cpu, idle); 376 if (rc) 377 return rc; 378 379 if (num_online_cpus() == 1) 380 /* Just in case we booted with a single CPU. */ 381 alternatives_enable_smp(); 382 383 rc = xen_smp_intr_init(cpu); 384 if (rc) 385 return rc; 386 387 rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL); 388 BUG_ON(rc); 389 390 while(per_cpu(cpu_state, cpu) != CPU_ONLINE) { 391 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 392 barrier(); 393 } 394 395 return 0; 396 } 397 398 static void xen_smp_cpus_done(unsigned int max_cpus) 399 { 400 } 401 402 #ifdef CONFIG_HOTPLUG_CPU 403 static int xen_cpu_disable(void) 404 { 405 unsigned int cpu = smp_processor_id(); 406 if (cpu == 0) 407 return -EBUSY; 408 409 cpu_disable_common(); 410 411 load_cr3(swapper_pg_dir); 412 return 0; 413 } 414 415 static void xen_cpu_die(unsigned int cpu) 416 { 417 while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) { 418 current->state = TASK_UNINTERRUPTIBLE; 419 schedule_timeout(HZ/10); 420 } 421 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL); 422 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL); 423 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL); 424 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL); 425 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu), NULL); 426 xen_uninit_lock_cpu(cpu); 427 xen_teardown_timer(cpu); 428 } 429 430 static void __cpuinit xen_play_dead(void) /* used only with HOTPLUG_CPU */ 431 { 432 play_dead_common(); 433 HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL); 434 cpu_bringup(); 435 } 436 437 #else /* !CONFIG_HOTPLUG_CPU */ 438 static int xen_cpu_disable(void) 439 { 440 return -ENOSYS; 441 } 442 443 static void xen_cpu_die(unsigned int cpu) 444 { 445 BUG(); 446 } 447 448 static void xen_play_dead(void) 449 { 450 BUG(); 451 } 452 453 #endif 454 static void stop_self(void *v) 455 { 456 int cpu = smp_processor_id(); 457 458 /* make sure we're not pinning something down */ 459 load_cr3(swapper_pg_dir); 460 /* should set up a minimal gdt */ 461 462 set_cpu_online(cpu, false); 463 464 HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL); 465 BUG(); 466 } 467 468 static void xen_stop_other_cpus(int wait) 469 { 470 smp_call_function(stop_self, NULL, wait); 471 } 472 473 static void xen_smp_send_reschedule(int cpu) 474 { 475 xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR); 476 } 477 478 static void __xen_send_IPI_mask(const struct cpumask *mask, 479 int vector) 480 { 481 unsigned cpu; 482 483 for_each_cpu_and(cpu, mask, cpu_online_mask) 484 xen_send_IPI_one(cpu, vector); 485 } 486 487 static void xen_smp_send_call_function_ipi(const struct cpumask *mask) 488 { 489 int cpu; 490 491 __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR); 492 493 /* Make sure other vcpus get a chance to run if they need to. */ 494 for_each_cpu(cpu, mask) { 495 if (xen_vcpu_stolen(cpu)) { 496 HYPERVISOR_sched_op(SCHEDOP_yield, NULL); 497 break; 498 } 499 } 500 } 501 502 static void xen_smp_send_call_function_single_ipi(int cpu) 503 { 504 __xen_send_IPI_mask(cpumask_of(cpu), 505 XEN_CALL_FUNCTION_SINGLE_VECTOR); 506 } 507 508 static inline int xen_map_vector(int vector) 509 { 510 int xen_vector; 511 512 switch (vector) { 513 case RESCHEDULE_VECTOR: 514 xen_vector = XEN_RESCHEDULE_VECTOR; 515 break; 516 case CALL_FUNCTION_VECTOR: 517 xen_vector = XEN_CALL_FUNCTION_VECTOR; 518 break; 519 case CALL_FUNCTION_SINGLE_VECTOR: 520 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR; 521 break; 522 case IRQ_WORK_VECTOR: 523 xen_vector = XEN_IRQ_WORK_VECTOR; 524 break; 525 default: 526 xen_vector = -1; 527 printk(KERN_ERR "xen: vector 0x%x is not implemented\n", 528 vector); 529 } 530 531 return xen_vector; 532 } 533 534 void xen_send_IPI_mask(const struct cpumask *mask, 535 int vector) 536 { 537 int xen_vector = xen_map_vector(vector); 538 539 if (xen_vector >= 0) 540 __xen_send_IPI_mask(mask, xen_vector); 541 } 542 543 void xen_send_IPI_all(int vector) 544 { 545 int xen_vector = xen_map_vector(vector); 546 547 if (xen_vector >= 0) 548 __xen_send_IPI_mask(cpu_online_mask, xen_vector); 549 } 550 551 void xen_send_IPI_self(int vector) 552 { 553 int xen_vector = xen_map_vector(vector); 554 555 if (xen_vector >= 0) 556 xen_send_IPI_one(smp_processor_id(), xen_vector); 557 } 558 559 void xen_send_IPI_mask_allbutself(const struct cpumask *mask, 560 int vector) 561 { 562 unsigned cpu; 563 unsigned int this_cpu = smp_processor_id(); 564 565 if (!(num_online_cpus() > 1)) 566 return; 567 568 for_each_cpu_and(cpu, mask, cpu_online_mask) { 569 if (this_cpu == cpu) 570 continue; 571 572 xen_smp_send_call_function_single_ipi(cpu); 573 } 574 } 575 576 void xen_send_IPI_allbutself(int vector) 577 { 578 int xen_vector = xen_map_vector(vector); 579 580 if (xen_vector >= 0) 581 xen_send_IPI_mask_allbutself(cpu_online_mask, xen_vector); 582 } 583 584 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id) 585 { 586 irq_enter(); 587 generic_smp_call_function_interrupt(); 588 inc_irq_stat(irq_call_count); 589 irq_exit(); 590 591 return IRQ_HANDLED; 592 } 593 594 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id) 595 { 596 irq_enter(); 597 generic_smp_call_function_single_interrupt(); 598 inc_irq_stat(irq_call_count); 599 irq_exit(); 600 601 return IRQ_HANDLED; 602 } 603 604 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id) 605 { 606 irq_enter(); 607 irq_work_run(); 608 inc_irq_stat(apic_irq_work_irqs); 609 irq_exit(); 610 611 return IRQ_HANDLED; 612 } 613 614 static const struct smp_ops xen_smp_ops __initconst = { 615 .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu, 616 .smp_prepare_cpus = xen_smp_prepare_cpus, 617 .smp_cpus_done = xen_smp_cpus_done, 618 619 .cpu_up = xen_cpu_up, 620 .cpu_die = xen_cpu_die, 621 .cpu_disable = xen_cpu_disable, 622 .play_dead = xen_play_dead, 623 624 .stop_other_cpus = xen_stop_other_cpus, 625 .smp_send_reschedule = xen_smp_send_reschedule, 626 627 .send_call_func_ipi = xen_smp_send_call_function_ipi, 628 .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi, 629 }; 630 631 void __init xen_smp_init(void) 632 { 633 smp_ops = xen_smp_ops; 634 xen_fill_possible_map(); 635 xen_init_spinlocks(); 636 } 637 638 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus) 639 { 640 native_smp_prepare_cpus(max_cpus); 641 WARN_ON(xen_smp_intr_init(0)); 642 643 xen_init_lock_cpu(0); 644 } 645 646 static int __cpuinit xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle) 647 { 648 int rc; 649 rc = native_cpu_up(cpu, tidle); 650 WARN_ON (xen_smp_intr_init(cpu)); 651 return rc; 652 } 653 654 static void xen_hvm_cpu_die(unsigned int cpu) 655 { 656 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL); 657 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL); 658 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL); 659 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL); 660 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu), NULL); 661 native_cpu_die(cpu); 662 } 663 664 void __init xen_hvm_smp_init(void) 665 { 666 if (!xen_have_vector_callback) 667 return; 668 smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus; 669 smp_ops.smp_send_reschedule = xen_smp_send_reschedule; 670 smp_ops.cpu_up = xen_hvm_cpu_up; 671 smp_ops.cpu_die = xen_hvm_cpu_die; 672 smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi; 673 smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi; 674 } 675