arch/x86/xen/smp.c

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