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 #include <linux/irq_work.h>
  20 #include <linux/tick.h>
  21 #include <linux/nmi.h>
  22
  23 #include <asm/paravirt.h>
  24 #include <asm/desc.h>
  25 #include <asm/pgtable.h>
  26 #include <asm/cpu.h>
  27
  28 #include <xen/interface/xen.h>
  29 #include <xen/interface/vcpu.h>
  30 #include <xen/interface/xenpmu.h>
  31
  32 #include <asm/xen/interface.h>
  33 #include <asm/xen/hypercall.h>
  34
  35 #include <xen/xen.h>
  36 #include <xen/page.h>
  37 #include <xen/events.h>
  38
  39 #include <xen/hvc-console.h>
  40 #include "xen-ops.h"
  41 #include "mmu.h"
  42 #include "smp.h"
  43 #include "pmu.h"
  44
  45 cpumask_var_t xen_cpu_initialized_map;
  46
  47 struct xen_common_irq {
  48         int irq;
  49         char *name;
  50 };
  51 static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
  52 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
  53 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
  54 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
  55 static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
  56 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
  57
  58 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  59 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
  60 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
  61
  62 /*
  63  * Reschedule call back.
  64  */
  65 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  66 {
  67         inc_irq_stat(irq_resched_count);
  68         scheduler_ipi();
  69
  70         return IRQ_HANDLED;
  71 }
  72
  73 static void cpu_bringup(void)
  74 {
  75         int cpu;
  76
  77         cpu_init();
  78         touch_softlockup_watchdog();
  79         preempt_disable();
  80
  81         /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
  82         if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
  83                 xen_enable_sysenter();
  84                 xen_enable_syscall();
  85         }
  86         cpu = smp_processor_id();
  87         smp_store_cpu_info(cpu);
  88         cpu_data(cpu).x86_max_cores = 1;
  89         set_cpu_sibling_map(cpu);
  90
  91         xen_setup_cpu_clockevents();
  92
  93         notify_cpu_starting(cpu);
  94
  95         set_cpu_online(cpu, true);
  96
  97         cpu_set_state_online(cpu);  /* Implies full memory barrier. */
  98
  99         /* We can take interrupts now: we're officially "up". */
 100         local_irq_enable();
 101 }
 102
 103 asmlinkage __visible void cpu_bringup_and_idle(void)
 104 {
 105         cpu_bringup();
 106         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 107 }
 108
 109 void xen_smp_intr_free(unsigned int cpu)
 110 {
 111         if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
 112                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
 113                 per_cpu(xen_resched_irq, cpu).irq = -1;
 114                 kfree(per_cpu(xen_resched_irq, cpu).name);
 115                 per_cpu(xen_resched_irq, cpu).name = NULL;
 116         }
 117         if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
 118                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
 119                 per_cpu(xen_callfunc_irq, cpu).irq = -1;
 120                 kfree(per_cpu(xen_callfunc_irq, cpu).name);
 121                 per_cpu(xen_callfunc_irq, cpu).name = NULL;
 122         }
 123         if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
 124                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
 125                 per_cpu(xen_debug_irq, cpu).irq = -1;
 126                 kfree(per_cpu(xen_debug_irq, cpu).name);
 127                 per_cpu(xen_debug_irq, cpu).name = NULL;
 128         }
 129         if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
 130                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
 131                                        NULL);
 132                 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
 133                 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
 134                 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
 135         }
 136 }
 137
 138 void xen_smp_intr_free_pv(unsigned int cpu)
 139 {
 140         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 141                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 142                 per_cpu(xen_irq_work, cpu).irq = -1;
 143                 kfree(per_cpu(xen_irq_work, cpu).name);
 144                 per_cpu(xen_irq_work, cpu).name = NULL;
 145         }
 146
 147         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 148                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 149                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 150                 kfree(per_cpu(xen_pmu_irq, cpu).name);
 151                 per_cpu(xen_pmu_irq, cpu).name = NULL;
 152         }
 153 }
 154
 155 int xen_smp_intr_init(unsigned int cpu)
 156 {
 157         int rc;
 158         char *resched_name, *callfunc_name, *debug_name;
 159
 160         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
 161         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
 162                                     cpu,
 163                                     xen_reschedule_interrupt,
 164                                     IRQF_PERCPU|IRQF_NOBALANCING,
 165                                     resched_name,
 166                                     NULL);
 167         if (rc < 0)
 168                 goto fail;
 169         per_cpu(xen_resched_irq, cpu).irq = rc;
 170         per_cpu(xen_resched_irq, cpu).name = resched_name;
 171
 172         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 173         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 174                                     cpu,
 175                                     xen_call_function_interrupt,
 176                                     IRQF_PERCPU|IRQF_NOBALANCING,
 177                                     callfunc_name,
 178                                     NULL);
 179         if (rc < 0)
 180                 goto fail;
 181         per_cpu(xen_callfunc_irq, cpu).irq = rc;
 182         per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
 183
 184         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 185         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 186                                      IRQF_PERCPU | IRQF_NOBALANCING,
 187                                      debug_name, NULL);
 188         if (rc < 0)
 189                 goto fail;
 190         per_cpu(xen_debug_irq, cpu).irq = rc;
 191         per_cpu(xen_debug_irq, cpu).name = debug_name;
 192
 193         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 194         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 195                                     cpu,
 196                                     xen_call_function_single_interrupt,
 197                                     IRQF_PERCPU|IRQF_NOBALANCING,
 198                                     callfunc_name,
 199                                     NULL);
 200         if (rc < 0)
 201                 goto fail;
 202         per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
 203         per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
 204
 205         return 0;
 206
 207  fail:
 208         xen_smp_intr_free(cpu);
 209         return rc;
 210 }
 211
 212 int xen_smp_intr_init_pv(unsigned int cpu)
 213 {
 214         int rc;
 215         char *callfunc_name, *pmu_name;
 216
 217         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
 218         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
 219                                     cpu,
 220                                     xen_irq_work_interrupt,
 221                                     IRQF_PERCPU|IRQF_NOBALANCING,
 222                                     callfunc_name,
 223                                     NULL);
 224         if (rc < 0)
 225                 goto fail;
 226         per_cpu(xen_irq_work, cpu).irq = rc;
 227         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 228
 229         if (is_xen_pmu(cpu)) {
 230                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 231                 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
 232                                              xen_pmu_irq_handler,
 233                                              IRQF_PERCPU|IRQF_NOBALANCING,
 234                                              pmu_name, NULL);
 235                 if (rc < 0)
 236                         goto fail;
 237                 per_cpu(xen_pmu_irq, cpu).irq = rc;
 238                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 239         }
 240
 241         return 0;
 242
 243  fail:
 244         xen_smp_intr_free_pv(cpu);
 245         return rc;
 246 }
 247
 248 static void __init xen_fill_possible_map(void)
 249 {
 250         int i, rc;
 251
 252         if (xen_initial_domain())
 253                 return;
 254
 255         for (i = 0; i < nr_cpu_ids; i++) {
 256                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 257                 if (rc >= 0) {
 258                         num_processors++;
 259                         set_cpu_possible(i, true);
 260                 }
 261         }
 262 }
 263
 264 static void __init xen_filter_cpu_maps(void)
 265 {
 266         int i, rc;
 267         unsigned int subtract = 0;
 268
 269         if (!xen_initial_domain())
 270                 return;
 271
 272         num_processors = 0;
 273         disabled_cpus = 0;
 274         for (i = 0; i < nr_cpu_ids; i++) {
 275                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 276                 if (rc >= 0) {
 277                         num_processors++;
 278                         set_cpu_possible(i, true);
 279                 } else {
 280                         set_cpu_possible(i, false);
 281                         set_cpu_present(i, false);
 282                         subtract++;
 283                 }
 284         }
 285 #ifdef CONFIG_HOTPLUG_CPU
 286         /* This is akin to using 'nr_cpus' on the Linux command line.
 287          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 288          * have up to X, while nr_cpu_ids is greater than X. This
 289          * normally is not a problem, except when CPU hotplugging
 290          * is involved and then there might be more than X CPUs
 291          * in the guest - which will not work as there is no
 292          * hypercall to expand the max number of VCPUs an already
 293          * running guest has. So cap it up to X. */
 294         if (subtract)
 295                 nr_cpu_ids = nr_cpu_ids - subtract;
 296 #endif
 297
 298 }
 299
 300 static void __init xen_pv_smp_prepare_boot_cpu(void)
 301 {
 302         BUG_ON(smp_processor_id() != 0);
 303         native_smp_prepare_boot_cpu();
 304
 305         if (!xen_feature(XENFEAT_writable_page_tables))
 306                 /* We've switched to the "real" per-cpu gdt, so make
 307                  * sure the old memory can be recycled. */
 308                 make_lowmem_page_readwrite(xen_initial_gdt);
 309
 310 #ifdef CONFIG_X86_32
 311         /*
 312          * Xen starts us with XEN_FLAT_RING1_DS, but linux code
 313          * expects __USER_DS
 314          */
 315         loadsegment(ds, __USER_DS);
 316         loadsegment(es, __USER_DS);
 317 #endif
 318
 319         xen_filter_cpu_maps();
 320         xen_setup_vcpu_info_placement();
 321
 322         /*
 323          * The alternative logic (which patches the unlock/lock) runs before
 324          * the smp bootup up code is activated. Hence we need to set this up
 325          * the core kernel is being patched. Otherwise we will have only
 326          * modules patched but not core code.
 327          */
 328         xen_init_spinlocks();
 329 }
 330
 331 static void __init xen_hvm_smp_prepare_boot_cpu(void)
 332 {
 333         BUG_ON(smp_processor_id() != 0);
 334         native_smp_prepare_boot_cpu();
 335
 336         /*
 337          * Setup vcpu_info for boot CPU.
 338          */
 339         xen_vcpu_setup(0);
 340
 341         /*
 342          * The alternative logic (which patches the unlock/lock) runs before
 343          * the smp bootup up code is activated. Hence we need to set this up
 344          * the core kernel is being patched. Otherwise we will have only
 345          * modules patched but not core code.
 346          */
 347         xen_init_spinlocks();
 348 }
 349
 350 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 351 {
 352         unsigned cpu;
 353         unsigned int i;
 354
 355         if (skip_ioapic_setup) {
 356                 char *m = (max_cpus == 0) ?
 357                         "The nosmp parameter is incompatible with Xen; " \
 358                         "use Xen dom0_max_vcpus=1 parameter" :
 359                         "The noapic parameter is incompatible with Xen";
 360
 361                 xen_raw_printk(m);
 362                 panic(m);
 363         }
 364         xen_init_lock_cpu(0);
 365
 366         smp_store_boot_cpu_info();
 367         cpu_data(0).x86_max_cores = 1;
 368
 369         for_each_possible_cpu(i) {
 370                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 371                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 372                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 373         }
 374         set_cpu_sibling_map(0);
 375
 376         xen_pmu_init(0);
 377
 378         if (xen_smp_intr_init(0))
 379                 BUG();
 380
 381         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 382                 panic("could not allocate xen_cpu_initialized_map\n");
 383
 384         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 385
 386         /* Restrict the possible_map according to max_cpus. */
 387         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 388                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 389                         continue;
 390                 set_cpu_possible(cpu, false);
 391         }
 392
 393         for_each_possible_cpu(cpu)
 394                 set_cpu_present(cpu, true);
 395 }
 396
 397 static int
 398 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 399 {
 400         struct vcpu_guest_context *ctxt;
 401         struct desc_struct *gdt;
 402         unsigned long gdt_mfn;
 403
 404         /* used to tell cpu_init() that it can proceed with initialization */
 405         cpumask_set_cpu(cpu, cpu_callout_mask);
 406         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 407                 return 0;
 408
 409         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 410         if (ctxt == NULL)
 411                 return -ENOMEM;
 412
 413         gdt = get_cpu_gdt_rw(cpu);
 414
 415 #ifdef CONFIG_X86_32
 416         ctxt->user_regs.fs = __KERNEL_PERCPU;
 417         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 418 #endif
 419         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 420
 421         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 422         ctxt->flags = VGCF_IN_KERNEL;
 423         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 424         ctxt->user_regs.ds = __USER_DS;
 425         ctxt->user_regs.es = __USER_DS;
 426         ctxt->user_regs.ss = __KERNEL_DS;
 427
 428         xen_copy_trap_info(ctxt->trap_ctxt);
 429
 430         ctxt->ldt_ents = 0;
 431
 432         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 433
 434         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 435         make_lowmem_page_readonly(gdt);
 436         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 437
 438         ctxt->gdt_frames[0] = gdt_mfn;
 439         ctxt->gdt_ents      = GDT_ENTRIES;
 440
 441         ctxt->kernel_ss = __KERNEL_DS;
 442         ctxt->kernel_sp = idle->thread.sp0;
 443
 444 #ifdef CONFIG_X86_32
 445         ctxt->event_callback_cs     = __KERNEL_CS;
 446         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 447 #else
 448         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 449 #endif
 450         ctxt->event_callback_eip    =
 451                 (unsigned long)xen_hypervisor_callback;
 452         ctxt->failsafe_callback_eip =
 453                 (unsigned long)xen_failsafe_callback;
 454         ctxt->user_regs.cs = __KERNEL_CS;
 455         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 456
 457         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 458         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 459         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 460                 BUG();
 461
 462         kfree(ctxt);
 463         return 0;
 464 }
 465
 466 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
 467 {
 468         int rc;
 469
 470         common_cpu_up(cpu, idle);
 471
 472         xen_setup_runstate_info(cpu);
 473
 474         /*
 475          * PV VCPUs are always successfully taken down (see 'while' loop
 476          * in xen_cpu_die()), so -EBUSY is an error.
 477          */
 478         rc = cpu_check_up_prepare(cpu);
 479         if (rc)
 480                 return rc;
 481
 482         /* make sure interrupts start blocked */
 483         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 484
 485         rc = cpu_initialize_context(cpu, idle);
 486         if (rc)
 487                 return rc;
 488
 489         xen_pmu_init(cpu);
 490
 491         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 492         BUG_ON(rc);
 493
 494         while (cpu_report_state(cpu) != CPU_ONLINE)
 495                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 496
 497         return 0;
 498 }
 499
 500 static void xen_smp_cpus_done(unsigned int max_cpus)
 501 {
 502 }
 503
 504 #ifdef CONFIG_HOTPLUG_CPU
 505 static int xen_cpu_disable(void)
 506 {
 507         unsigned int cpu = smp_processor_id();
 508         if (cpu == 0)
 509                 return -EBUSY;
 510
 511         cpu_disable_common();
 512
 513         load_cr3(swapper_pg_dir);
 514         return 0;
 515 }
 516
 517 static void xen_pv_cpu_die(unsigned int cpu)
 518 {
 519         while (HYPERVISOR_vcpu_op(VCPUOP_is_up,
 520                                   xen_vcpu_nr(cpu), NULL)) {
 521                 __set_current_state(TASK_UNINTERRUPTIBLE);
 522                 schedule_timeout(HZ/10);
 523         }
 524
 525         if (common_cpu_die(cpu) == 0) {
 526                 xen_smp_intr_free(cpu);
 527                 xen_uninit_lock_cpu(cpu);
 528                 xen_teardown_timer(cpu);
 529                 xen_pmu_finish(cpu);
 530         }
 531 }
 532
 533 static void xen_hvm_cpu_die(unsigned int cpu)
 534 {
 535         if (common_cpu_die(cpu) == 0) {
 536                 xen_smp_intr_free(cpu);
 537                 xen_uninit_lock_cpu(cpu);
 538                 xen_teardown_timer(cpu);
 539         }
 540 }
 541
 542 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
 543 {
 544         play_dead_common();
 545         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 546         cpu_bringup();
 547         /*
 548          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 549          * clears certain data that the cpu_idle loop (which called us
 550          * and that we return from) expects. The only way to get that
 551          * data back is to call:
 552          */
 553         tick_nohz_idle_enter();
 554
 555         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 556 }
 557
 558 #else /* !CONFIG_HOTPLUG_CPU */
 559 static int xen_cpu_disable(void)
 560 {
 561         return -ENOSYS;
 562 }
 563
 564 static void xen_pv_cpu_die(unsigned int cpu)
 565 {
 566         BUG();
 567 }
 568
 569 static void xen_hvm_cpu_die(unsigned int cpu)
 570 {
 571         BUG();
 572 }
 573
 574 static void xen_play_dead(void)
 575 {
 576         BUG();
 577 }
 578
 579 #endif
 580 static void stop_self(void *v)
 581 {
 582         int cpu = smp_processor_id();
 583
 584         /* make sure we're not pinning something down */
 585         load_cr3(swapper_pg_dir);
 586         /* should set up a minimal gdt */
 587
 588         set_cpu_online(cpu, false);
 589
 590         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 591         BUG();
 592 }
 593
 594 static void xen_stop_other_cpus(int wait)
 595 {
 596         smp_call_function(stop_self, NULL, wait);
 597 }
 598
 599 static void xen_smp_send_reschedule(int cpu)
 600 {
 601         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 602 }
 603
 604 static void __xen_send_IPI_mask(const struct cpumask *mask,
 605                               int vector)
 606 {
 607         unsigned cpu;
 608
 609         for_each_cpu_and(cpu, mask, cpu_online_mask)
 610                 xen_send_IPI_one(cpu, vector);
 611 }
 612
 613 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 614 {
 615         int cpu;
 616
 617         __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 618
 619         /* Make sure other vcpus get a chance to run if they need to. */
 620         for_each_cpu(cpu, mask) {
 621                 if (xen_vcpu_stolen(cpu)) {
 622                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 623                         break;
 624                 }
 625         }
 626 }
 627
 628 static void xen_smp_send_call_function_single_ipi(int cpu)
 629 {
 630         __xen_send_IPI_mask(cpumask_of(cpu),
 631                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 632 }
 633
 634 static inline int xen_map_vector(int vector)
 635 {
 636         int xen_vector;
 637
 638         switch (vector) {
 639         case RESCHEDULE_VECTOR:
 640                 xen_vector = XEN_RESCHEDULE_VECTOR;
 641                 break;
 642         case CALL_FUNCTION_VECTOR:
 643                 xen_vector = XEN_CALL_FUNCTION_VECTOR;
 644                 break;
 645         case CALL_FUNCTION_SINGLE_VECTOR:
 646                 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
 647                 break;
 648         case IRQ_WORK_VECTOR:
 649                 xen_vector = XEN_IRQ_WORK_VECTOR;
 650                 break;
 651 #ifdef CONFIG_X86_64
 652         case NMI_VECTOR:
 653         case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
 654                 xen_vector = XEN_NMI_VECTOR;
 655                 break;
 656 #endif
 657         default:
 658                 xen_vector = -1;
 659                 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
 660                         vector);
 661         }
 662
 663         return xen_vector;
 664 }
 665
 666 void xen_send_IPI_mask(const struct cpumask *mask,
 667                               int vector)
 668 {
 669         int xen_vector = xen_map_vector(vector);
 670
 671         if (xen_vector >= 0)
 672                 __xen_send_IPI_mask(mask, xen_vector);
 673 }
 674
 675 void xen_send_IPI_all(int vector)
 676 {
 677         int xen_vector = xen_map_vector(vector);
 678
 679         if (xen_vector >= 0)
 680                 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
 681 }
 682
 683 void xen_send_IPI_self(int vector)
 684 {
 685         int xen_vector = xen_map_vector(vector);
 686
 687         if (xen_vector >= 0)
 688                 xen_send_IPI_one(smp_processor_id(), xen_vector);
 689 }
 690
 691 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
 692                                 int vector)
 693 {
 694         unsigned cpu;
 695         unsigned int this_cpu = smp_processor_id();
 696         int xen_vector = xen_map_vector(vector);
 697
 698         if (!(num_online_cpus() > 1) || (xen_vector < 0))
 699                 return;
 700
 701         for_each_cpu_and(cpu, mask, cpu_online_mask) {
 702                 if (this_cpu == cpu)
 703                         continue;
 704
 705                 xen_send_IPI_one(cpu, xen_vector);
 706         }
 707 }
 708
 709 void xen_send_IPI_allbutself(int vector)
 710 {
 711         xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
 712 }
 713
 714 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 715 {
 716         irq_enter();
 717         generic_smp_call_function_interrupt();
 718         inc_irq_stat(irq_call_count);
 719         irq_exit();
 720
 721         return IRQ_HANDLED;
 722 }
 723
 724 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 725 {
 726         irq_enter();
 727         generic_smp_call_function_single_interrupt();
 728         inc_irq_stat(irq_call_count);
 729         irq_exit();
 730
 731         return IRQ_HANDLED;
 732 }
 733
 734 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 735 {
 736         irq_enter();
 737         irq_work_run();
 738         inc_irq_stat(apic_irq_work_irqs);
 739         irq_exit();
 740
 741         return IRQ_HANDLED;
 742 }
 743
 744 static const struct smp_ops xen_smp_ops __initconst = {
 745         .smp_prepare_boot_cpu = xen_pv_smp_prepare_boot_cpu,
 746         .smp_prepare_cpus = xen_smp_prepare_cpus,
 747         .smp_cpus_done = xen_smp_cpus_done,
 748
 749         .cpu_up = xen_cpu_up,
 750         .cpu_die = xen_pv_cpu_die,
 751         .cpu_disable = xen_cpu_disable,
 752         .play_dead = xen_play_dead,
 753
 754         .stop_other_cpus = xen_stop_other_cpus,
 755         .smp_send_reschedule = xen_smp_send_reschedule,
 756
 757         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 758         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 759 };
 760
 761 void __init xen_smp_init(void)
 762 {
 763         smp_ops = xen_smp_ops;
 764         xen_fill_possible_map();
 765 }
 766
 767 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
 768 {
 769         native_smp_prepare_cpus(max_cpus);
 770         WARN_ON(xen_smp_intr_init(0));
 771
 772         xen_init_lock_cpu(0);
 773 }
 774
 775 void __init xen_hvm_smp_init(void)
 776 {
 777         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
 778         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
 779         smp_ops.cpu_die = xen_hvm_cpu_die;
 780         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
 781         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
 782         smp_ops.smp_prepare_boot_cpu = xen_hvm_smp_prepare_boot_cpu;
 783 }