kernel/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
  22 #include <asm/paravirt.h>
  23 #include <asm/desc.h>
  24 #include <asm/pgtable.h>
  25 #include <asm/cpu.h>
  26
  27 #include <xen/interface/xen.h>
  28 #include <xen/interface/vcpu.h>
  29 #include <xen/interface/xenpmu.h>
  30
  31 #include <asm/xen/interface.h>
  32 #include <asm/xen/hypercall.h>
  33
  34 #include <xen/xen.h>
  35 #include <xen/page.h>
  36 #include <xen/events.h>
  37
  38 #include <xen/hvc-console.h>
  39 #include "xen-ops.h"
  40 #include "mmu.h"
  41 #include "smp.h"
  42 #include "pmu.h"
  43
  44 cpumask_var_t xen_cpu_initialized_map;
  45
  46 struct xen_common_irq {
  47         int irq;
  48         char *name;
  49 };
  50 static DEFINE_PER_CPU(struct xen_common_irq, xen_resched_irq) = { .irq = -1 };
  51 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfunc_irq) = { .irq = -1 };
  52 static DEFINE_PER_CPU(struct xen_common_irq, xen_callfuncsingle_irq) = { .irq = -1 };
  53 static DEFINE_PER_CPU(struct xen_common_irq, xen_irq_work) = { .irq = -1 };
  54 static DEFINE_PER_CPU(struct xen_common_irq, xen_debug_irq) = { .irq = -1 };
  55 static DEFINE_PER_CPU(struct xen_common_irq, xen_pmu_irq) = { .irq = -1 };
  56
  57 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  58 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
  59 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id);
  60
  61 /*
  62  * Reschedule call back.
  63  */
  64 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  65 {
  66         inc_irq_stat(irq_resched_count);
  67         scheduler_ipi();
  68
  69         return IRQ_HANDLED;
  70 }
  71
  72 static void cpu_bringup(void)
  73 {
  74         int cpu;
  75
  76         cpu_init();
  77         touch_softlockup_watchdog();
  78         preempt_disable();
  79
  80         /* PVH runs in ring 0 and allows us to do native syscalls. Yay! */
  81         if (!xen_feature(XENFEAT_supervisor_mode_kernel)) {
  82                 xen_enable_sysenter();
  83                 xen_enable_syscall();
  84         }
  85         cpu = smp_processor_id();
  86         smp_store_cpu_info(cpu);
  87         cpu_data(cpu).x86_max_cores = 1;
  88         set_cpu_sibling_map(cpu);
  89
  90         xen_setup_cpu_clockevents();
  91
  92         notify_cpu_starting(cpu);
  93
  94         set_cpu_online(cpu, true);
  95
  96         cpu_set_state_online(cpu);  /* Implies full memory barrier. */
  97
  98         /* We can take interrupts now: we're officially "up". */
  99         local_irq_enable();
 100 }
 101
 102 /*
 103  * Note: cpu parameter is only relevant for PVH. The reason for passing it
 104  * is we can't do smp_processor_id until the percpu segments are loaded, for
 105  * which we need the cpu number! So we pass it in rdi as first parameter.
 106  */
 107 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
 108 {
 109 #ifdef CONFIG_XEN_PVH
 110         if (xen_feature(XENFEAT_auto_translated_physmap) &&
 111             xen_feature(XENFEAT_supervisor_mode_kernel))
 112                 xen_pvh_secondary_vcpu_init(cpu);
 113 #endif
 114         cpu_bringup();
 115         cpu_startup_entry(CPUHP_ONLINE);
 116 }
 117
 118 static void xen_smp_intr_free(unsigned int cpu)
 119 {
 120         if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
 121                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
 122                 per_cpu(xen_resched_irq, cpu).irq = -1;
 123                 kfree(per_cpu(xen_resched_irq, cpu).name);
 124                 per_cpu(xen_resched_irq, cpu).name = NULL;
 125         }
 126         if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
 127                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
 128                 per_cpu(xen_callfunc_irq, cpu).irq = -1;
 129                 kfree(per_cpu(xen_callfunc_irq, cpu).name);
 130                 per_cpu(xen_callfunc_irq, cpu).name = NULL;
 131         }
 132         if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
 133                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
 134                 per_cpu(xen_debug_irq, cpu).irq = -1;
 135                 kfree(per_cpu(xen_debug_irq, cpu).name);
 136                 per_cpu(xen_debug_irq, cpu).name = NULL;
 137         }
 138         if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
 139                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
 140                                        NULL);
 141                 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
 142                 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
 143                 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
 144         }
 145         if (xen_hvm_domain())
 146                 return;
 147
 148         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 149                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 150                 per_cpu(xen_irq_work, cpu).irq = -1;
 151                 kfree(per_cpu(xen_irq_work, cpu).name);
 152                 per_cpu(xen_irq_work, cpu).name = NULL;
 153         }
 154
 155         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 156                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 157                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 158                 kfree(per_cpu(xen_pmu_irq, cpu).name);
 159                 per_cpu(xen_pmu_irq, cpu).name = NULL;
 160         }
 161 };
 162 static int xen_smp_intr_init(unsigned int cpu)
 163 {
 164         int rc;
 165         char *resched_name, *callfunc_name, *debug_name, *pmu_name;
 166
 167         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
 168         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
 169                                     cpu,
 170                                     xen_reschedule_interrupt,
 171                                     IRQF_PERCPU|IRQF_NOBALANCING,
 172                                     resched_name,
 173                                     NULL);
 174         if (rc < 0)
 175                 goto fail;
 176         per_cpu(xen_resched_irq, cpu).irq = rc;
 177         per_cpu(xen_resched_irq, cpu).name = resched_name;
 178
 179         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 180         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 181                                     cpu,
 182                                     xen_call_function_interrupt,
 183                                     IRQF_PERCPU|IRQF_NOBALANCING,
 184                                     callfunc_name,
 185                                     NULL);
 186         if (rc < 0)
 187                 goto fail;
 188         per_cpu(xen_callfunc_irq, cpu).irq = rc;
 189         per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
 190
 191         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 192         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 193                                      IRQF_PERCPU | IRQF_NOBALANCING,
 194                                      debug_name, NULL);
 195         if (rc < 0)
 196                 goto fail;
 197         per_cpu(xen_debug_irq, cpu).irq = rc;
 198         per_cpu(xen_debug_irq, cpu).name = debug_name;
 199
 200         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 201         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 202                                     cpu,
 203                                     xen_call_function_single_interrupt,
 204                                     IRQF_PERCPU|IRQF_NOBALANCING,
 205                                     callfunc_name,
 206                                     NULL);
 207         if (rc < 0)
 208                 goto fail;
 209         per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
 210         per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
 211
 212         /*
 213          * The IRQ worker on PVHVM goes through the native path and uses the
 214          * IPI mechanism.
 215          */
 216         if (xen_hvm_domain())
 217                 return 0;
 218
 219         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
 220         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
 221                                     cpu,
 222                                     xen_irq_work_interrupt,
 223                                     IRQF_PERCPU|IRQF_NOBALANCING,
 224                                     callfunc_name,
 225                                     NULL);
 226         if (rc < 0)
 227                 goto fail;
 228         per_cpu(xen_irq_work, cpu).irq = rc;
 229         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 230
 231         if (is_xen_pmu(cpu)) {
 232                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 233                 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
 234                                              xen_pmu_irq_handler,
 235                                              IRQF_PERCPU|IRQF_NOBALANCING,
 236                                              pmu_name, NULL);
 237                 if (rc < 0)
 238                         goto fail;
 239                 per_cpu(xen_pmu_irq, cpu).irq = rc;
 240                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 241         }
 242
 243         return 0;
 244
 245  fail:
 246         xen_smp_intr_free(cpu);
 247         return rc;
 248 }
 249
 250 static void __init xen_fill_possible_map(void)
 251 {
 252         int i, rc;
 253
 254         if (xen_initial_domain())
 255                 return;
 256
 257         for (i = 0; i < nr_cpu_ids; i++) {
 258                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 259                 if (rc >= 0) {
 260                         num_processors++;
 261                         set_cpu_possible(i, true);
 262                 }
 263         }
 264 }
 265
 266 static void __init xen_filter_cpu_maps(void)
 267 {
 268         int i, rc;
 269         unsigned int subtract = 0;
 270
 271         if (!xen_initial_domain())
 272                 return;
 273
 274         num_processors = 0;
 275         disabled_cpus = 0;
 276         for (i = 0; i < nr_cpu_ids; i++) {
 277                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 278                 if (rc >= 0) {
 279                         num_processors++;
 280                         set_cpu_possible(i, true);
 281                 } else {
 282                         set_cpu_possible(i, false);
 283                         set_cpu_present(i, false);
 284                         subtract++;
 285                 }
 286         }
 287 #ifdef CONFIG_HOTPLUG_CPU
 288         /* This is akin to using 'nr_cpus' on the Linux command line.
 289          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 290          * have up to X, while nr_cpu_ids is greater than X. This
 291          * normally is not a problem, except when CPU hotplugging
 292          * is involved and then there might be more than X CPUs
 293          * in the guest - which will not work as there is no
 294          * hypercall to expand the max number of VCPUs an already
 295          * running guest has. So cap it up to X. */
 296         if (subtract)
 297                 nr_cpu_ids = nr_cpu_ids - subtract;
 298 #endif
 299
 300 }
 301
 302 static void __init xen_smp_prepare_boot_cpu(void)
 303 {
 304         BUG_ON(smp_processor_id() != 0);
 305         native_smp_prepare_boot_cpu();
 306
 307         if (xen_pv_domain()) {
 308                 if (!xen_feature(XENFEAT_writable_page_tables))
 309                         /* We've switched to the "real" per-cpu gdt, so make
 310                          * sure the old memory can be recycled. */
 311                         make_lowmem_page_readwrite(xen_initial_gdt);
 312
 313 #ifdef CONFIG_X86_32
 314                 /*
 315                  * Xen starts us with XEN_FLAT_RING1_DS, but linux code
 316                  * expects __USER_DS
 317                  */
 318                 loadsegment(ds, __USER_DS);
 319                 loadsegment(es, __USER_DS);
 320 #endif
 321
 322                 xen_filter_cpu_maps();
 323                 xen_setup_vcpu_info_placement();
 324         }
 325         /*
 326          * The alternative logic (which patches the unlock/lock) runs before
 327          * the smp bootup up code is activated. Hence we need to set this up
 328          * the core kernel is being patched. Otherwise we will have only
 329          * modules patched but not core code.
 330          */
 331         xen_init_spinlocks();
 332 }
 333
 334 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 335 {
 336         unsigned cpu;
 337         unsigned int i;
 338
 339         if (skip_ioapic_setup) {
 340                 char *m = (max_cpus == 0) ?
 341                         "The nosmp parameter is incompatible with Xen; " \
 342                         "use Xen dom0_max_vcpus=1 parameter" :
 343                         "The noapic parameter is incompatible with Xen";
 344
 345                 xen_raw_printk(m);
 346                 panic(m);
 347         }
 348         xen_init_lock_cpu(0);
 349
 350         smp_store_boot_cpu_info();
 351         cpu_data(0).x86_max_cores = 1;
 352
 353         for_each_possible_cpu(i) {
 354                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 355                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 356                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 357         }
 358         set_cpu_sibling_map(0);
 359
 360         xen_pmu_init(0);
 361
 362         if (xen_smp_intr_init(0))
 363                 BUG();
 364
 365         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 366                 panic("could not allocate xen_cpu_initialized_map\n");
 367
 368         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 369
 370         /* Restrict the possible_map according to max_cpus. */
 371         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 372                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 373                         continue;
 374                 set_cpu_possible(cpu, false);
 375         }
 376
 377         for_each_possible_cpu(cpu)
 378                 set_cpu_present(cpu, true);
 379 }
 380
 381 static int
 382 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 383 {
 384         struct vcpu_guest_context *ctxt;
 385         struct desc_struct *gdt;
 386         unsigned long gdt_mfn;
 387
 388         /* used to tell cpu_init() that it can proceed with initialization */
 389         cpumask_set_cpu(cpu, cpu_callout_mask);
 390         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 391                 return 0;
 392
 393         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 394         if (ctxt == NULL)
 395                 return -ENOMEM;
 396
 397         gdt = get_cpu_gdt_table(cpu);
 398
 399 #ifdef CONFIG_X86_32
 400         /* Note: PVH is not yet supported on x86_32. */
 401         ctxt->user_regs.fs = __KERNEL_PERCPU;
 402         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 403 #endif
 404         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 405
 406         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 407                 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 408                 ctxt->flags = VGCF_IN_KERNEL;
 409                 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 410                 ctxt->user_regs.ds = __USER_DS;
 411                 ctxt->user_regs.es = __USER_DS;
 412                 ctxt->user_regs.ss = __KERNEL_DS;
 413
 414                 xen_copy_trap_info(ctxt->trap_ctxt);
 415
 416                 ctxt->ldt_ents = 0;
 417
 418                 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 419
 420                 gdt_mfn = arbitrary_virt_to_mfn(gdt);
 421                 make_lowmem_page_readonly(gdt);
 422                 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 423
 424                 ctxt->gdt_frames[0] = gdt_mfn;
 425                 ctxt->gdt_ents      = GDT_ENTRIES;
 426
 427                 ctxt->kernel_ss = __KERNEL_DS;
 428                 ctxt->kernel_sp = idle->thread.sp0;
 429
 430 #ifdef CONFIG_X86_32
 431                 ctxt->event_callback_cs     = __KERNEL_CS;
 432                 ctxt->failsafe_callback_cs  = __KERNEL_CS;
 433 #else
 434                 ctxt->gs_base_kernel = per_cpu_offset(cpu);
 435 #endif
 436                 ctxt->event_callback_eip    =
 437                                         (unsigned long)xen_hypervisor_callback;
 438                 ctxt->failsafe_callback_eip =
 439                                         (unsigned long)xen_failsafe_callback;
 440                 ctxt->user_regs.cs = __KERNEL_CS;
 441                 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 442         }
 443 #ifdef CONFIG_XEN_PVH
 444         else {
 445                 /*
 446                  * The vcpu comes on kernel page tables which have the NX pte
 447                  * bit set. This means before DS/SS is touched, NX in
 448                  * EFER must be set. Hence the following assembly glue code.
 449                  */
 450                 ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
 451                 ctxt->user_regs.rdi = cpu;
 452                 ctxt->user_regs.rsi = true;  /* entry == true */
 453         }
 454 #endif
 455         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 456         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 457         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 458                 BUG();
 459
 460         kfree(ctxt);
 461         return 0;
 462 }
 463
 464 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
 465 {
 466         int rc;
 467
 468         common_cpu_up(cpu, idle);
 469
 470         xen_setup_runstate_info(cpu);
 471         xen_setup_timer(cpu);
 472         xen_init_lock_cpu(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 = xen_smp_intr_init(cpu);
 492         if (rc)
 493                 return rc;
 494
 495         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 496         BUG_ON(rc);
 497
 498         while (cpu_report_state(cpu) != CPU_ONLINE)
 499                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 500
 501         return 0;
 502 }
 503
 504 static void xen_smp_cpus_done(unsigned int max_cpus)
 505 {
 506 }
 507
 508 #ifdef CONFIG_HOTPLUG_CPU
 509 static int xen_cpu_disable(void)
 510 {
 511         unsigned int cpu = smp_processor_id();
 512         if (cpu == 0)
 513                 return -EBUSY;
 514
 515         cpu_disable_common();
 516
 517         load_cr3(swapper_pg_dir);
 518         return 0;
 519 }
 520
 521 static void xen_cpu_die(unsigned int cpu)
 522 {
 523         while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
 524                 __set_current_state(TASK_UNINTERRUPTIBLE);
 525                 schedule_timeout(HZ/10);
 526         }
 527
 528         if (common_cpu_die(cpu) == 0) {
 529                 xen_smp_intr_free(cpu);
 530                 xen_uninit_lock_cpu(cpu);
 531                 xen_teardown_timer(cpu);
 532                 xen_pmu_finish(cpu);
 533         }
 534 }
 535
 536 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
 537 {
 538         play_dead_common();
 539         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 540         cpu_bringup();
 541         /*
 542          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 543          * clears certain data that the cpu_idle loop (which called us
 544          * and that we return from) expects. The only way to get that
 545          * data back is to call:
 546          */
 547         tick_nohz_idle_enter();
 548 }
 549
 550 #else /* !CONFIG_HOTPLUG_CPU */
 551 static int xen_cpu_disable(void)
 552 {
 553         return -ENOSYS;
 554 }
 555
 556 static void xen_cpu_die(unsigned int cpu)
 557 {
 558         BUG();
 559 }
 560
 561 static void xen_play_dead(void)
 562 {
 563         BUG();
 564 }
 565
 566 #endif
 567 static void stop_self(void *v)
 568 {
 569         int cpu = smp_processor_id();
 570
 571         /* make sure we're not pinning something down */
 572         load_cr3(swapper_pg_dir);
 573         /* should set up a minimal gdt */
 574
 575         set_cpu_online(cpu, false);
 576
 577         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 578         BUG();
 579 }
 580
 581 static void xen_stop_other_cpus(int wait)
 582 {
 583         smp_call_function(stop_self, NULL, wait);
 584 }
 585
 586 static void xen_smp_send_reschedule(int cpu)
 587 {
 588         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 589 }
 590
 591 static void __xen_send_IPI_mask(const struct cpumask *mask,
 592                               int vector)
 593 {
 594         unsigned cpu;
 595
 596         for_each_cpu_and(cpu, mask, cpu_online_mask)
 597                 xen_send_IPI_one(cpu, vector);
 598 }
 599
 600 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 601 {
 602         int cpu;
 603
 604         __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 605
 606         /* Make sure other vcpus get a chance to run if they need to. */
 607         for_each_cpu(cpu, mask) {
 608                 if (xen_vcpu_stolen(cpu)) {
 609                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 610                         break;
 611                 }
 612         }
 613 }
 614
 615 static void xen_smp_send_call_function_single_ipi(int cpu)
 616 {
 617         __xen_send_IPI_mask(cpumask_of(cpu),
 618                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 619 }
 620
 621 static inline int xen_map_vector(int vector)
 622 {
 623         int xen_vector;
 624
 625         switch (vector) {
 626         case RESCHEDULE_VECTOR:
 627                 xen_vector = XEN_RESCHEDULE_VECTOR;
 628                 break;
 629         case CALL_FUNCTION_VECTOR:
 630                 xen_vector = XEN_CALL_FUNCTION_VECTOR;
 631                 break;
 632         case CALL_FUNCTION_SINGLE_VECTOR:
 633                 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
 634                 break;
 635         case IRQ_WORK_VECTOR:
 636                 xen_vector = XEN_IRQ_WORK_VECTOR;
 637                 break;
 638 #ifdef CONFIG_X86_64
 639         case NMI_VECTOR:
 640         case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
 641                 xen_vector = XEN_NMI_VECTOR;
 642                 break;
 643 #endif
 644         default:
 645                 xen_vector = -1;
 646                 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
 647                         vector);
 648         }
 649
 650         return xen_vector;
 651 }
 652
 653 void xen_send_IPI_mask(const struct cpumask *mask,
 654                               int vector)
 655 {
 656         int xen_vector = xen_map_vector(vector);
 657
 658         if (xen_vector >= 0)
 659                 __xen_send_IPI_mask(mask, xen_vector);
 660 }
 661
 662 void xen_send_IPI_all(int vector)
 663 {
 664         int xen_vector = xen_map_vector(vector);
 665
 666         if (xen_vector >= 0)
 667                 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
 668 }
 669
 670 void xen_send_IPI_self(int vector)
 671 {
 672         int xen_vector = xen_map_vector(vector);
 673
 674         if (xen_vector >= 0)
 675                 xen_send_IPI_one(smp_processor_id(), xen_vector);
 676 }
 677
 678 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
 679                                 int vector)
 680 {
 681         unsigned cpu;
 682         unsigned int this_cpu = smp_processor_id();
 683         int xen_vector = xen_map_vector(vector);
 684
 685         if (!(num_online_cpus() > 1) || (xen_vector < 0))
 686                 return;
 687
 688         for_each_cpu_and(cpu, mask, cpu_online_mask) {
 689                 if (this_cpu == cpu)
 690                         continue;
 691
 692                 xen_send_IPI_one(cpu, xen_vector);
 693         }
 694 }
 695
 696 void xen_send_IPI_allbutself(int vector)
 697 {
 698         xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
 699 }
 700
 701 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 702 {
 703         irq_enter();
 704         generic_smp_call_function_interrupt();
 705         inc_irq_stat(irq_call_count);
 706         irq_exit();
 707
 708         return IRQ_HANDLED;
 709 }
 710
 711 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 712 {
 713         irq_enter();
 714         generic_smp_call_function_single_interrupt();
 715         inc_irq_stat(irq_call_count);
 716         irq_exit();
 717
 718         return IRQ_HANDLED;
 719 }
 720
 721 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 722 {
 723         irq_enter();
 724         irq_work_run();
 725         inc_irq_stat(apic_irq_work_irqs);
 726         irq_exit();
 727
 728         return IRQ_HANDLED;
 729 }
 730
 731 static const struct smp_ops xen_smp_ops __initconst = {
 732         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 733         .smp_prepare_cpus = xen_smp_prepare_cpus,
 734         .smp_cpus_done = xen_smp_cpus_done,
 735
 736         .cpu_up = xen_cpu_up,
 737         .cpu_die = xen_cpu_die,
 738         .cpu_disable = xen_cpu_disable,
 739         .play_dead = xen_play_dead,
 740
 741         .stop_other_cpus = xen_stop_other_cpus,
 742         .smp_send_reschedule = xen_smp_send_reschedule,
 743
 744         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 745         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 746 };
 747
 748 void __init xen_smp_init(void)
 749 {
 750         smp_ops = xen_smp_ops;
 751         xen_fill_possible_map();
 752 }
 753
 754 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
 755 {
 756         native_smp_prepare_cpus(max_cpus);
 757         WARN_ON(xen_smp_intr_init(0));
 758
 759         xen_init_lock_cpu(0);
 760 }
 761
 762 static int xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle)
 763 {
 764         int rc;
 765
 766         /*
 767          * This can happen if CPU was offlined earlier and
 768          * offlining timed out in common_cpu_die().
 769          */
 770         if (cpu_report_state(cpu) == CPU_DEAD_FROZEN) {
 771                 xen_smp_intr_free(cpu);
 772                 xen_uninit_lock_cpu(cpu);
 773         }
 774
 775         /*
 776          * xen_smp_intr_init() needs to run before native_cpu_up()
 777          * so that IPI vectors are set up on the booting CPU before
 778          * it is marked online in native_cpu_up().
 779         */
 780         rc = xen_smp_intr_init(cpu);
 781         WARN_ON(rc);
 782         if (!rc)
 783                 rc =  native_cpu_up(cpu, tidle);
 784
 785         /*
 786          * We must initialize the slowpath CPU kicker _after_ the native
 787          * path has executed. If we initialized it before none of the
 788          * unlocker IPI kicks would reach the booting CPU as the booting
 789          * CPU had not set itself 'online' in cpu_online_mask. That mask
 790          * is checked when IPIs are sent (on HVM at least).
 791          */
 792         xen_init_lock_cpu(cpu);
 793         return rc;
 794 }
 795
 796 void __init xen_hvm_smp_init(void)
 797 {
 798         if (!xen_have_vector_callback)
 799                 return;
 800         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
 801         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
 802         smp_ops.cpu_up = xen_hvm_cpu_up;
 803         smp_ops.cpu_die = xen_cpu_die;
 804         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
 805         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
 806         smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;
 807 }