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         /*
  91          * identify_cpu() may have set logical_pkg_id to -1 due
  92          * to incorrect phys_proc_id. Let's re-comupte it.
  93          */
  94         topology_update_package_map(apic->cpu_present_to_apicid(cpu), cpu);
  95
  96         xen_setup_cpu_clockevents();
  97
  98         notify_cpu_starting(cpu);
  99
 100         set_cpu_online(cpu, true);
 101
 102         cpu_set_state_online(cpu);  /* Implies full memory barrier. */
 103
 104         /* We can take interrupts now: we're officially "up". */
 105         local_irq_enable();
 106 }
 107
 108 /*
 109  * Note: cpu parameter is only relevant for PVH. The reason for passing it
 110  * is we can't do smp_processor_id until the percpu segments are loaded, for
 111  * which we need the cpu number! So we pass it in rdi as first parameter.
 112  */
 113 asmlinkage __visible void cpu_bringup_and_idle(int cpu)
 114 {
 115 #ifdef CONFIG_XEN_PVH
 116         if (xen_feature(XENFEAT_auto_translated_physmap) &&
 117             xen_feature(XENFEAT_supervisor_mode_kernel))
 118                 xen_pvh_secondary_vcpu_init(cpu);
 119 #endif
 120         cpu_bringup();
 121         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 122 }
 123
 124 static void xen_smp_intr_free(unsigned int cpu)
 125 {
 126         if (per_cpu(xen_resched_irq, cpu).irq >= 0) {
 127                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu).irq, NULL);
 128                 per_cpu(xen_resched_irq, cpu).irq = -1;
 129                 kfree(per_cpu(xen_resched_irq, cpu).name);
 130                 per_cpu(xen_resched_irq, cpu).name = NULL;
 131         }
 132         if (per_cpu(xen_callfunc_irq, cpu).irq >= 0) {
 133                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu).irq, NULL);
 134                 per_cpu(xen_callfunc_irq, cpu).irq = -1;
 135                 kfree(per_cpu(xen_callfunc_irq, cpu).name);
 136                 per_cpu(xen_callfunc_irq, cpu).name = NULL;
 137         }
 138         if (per_cpu(xen_debug_irq, cpu).irq >= 0) {
 139                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu).irq, NULL);
 140                 per_cpu(xen_debug_irq, cpu).irq = -1;
 141                 kfree(per_cpu(xen_debug_irq, cpu).name);
 142                 per_cpu(xen_debug_irq, cpu).name = NULL;
 143         }
 144         if (per_cpu(xen_callfuncsingle_irq, cpu).irq >= 0) {
 145                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu).irq,
 146                                        NULL);
 147                 per_cpu(xen_callfuncsingle_irq, cpu).irq = -1;
 148                 kfree(per_cpu(xen_callfuncsingle_irq, cpu).name);
 149                 per_cpu(xen_callfuncsingle_irq, cpu).name = NULL;
 150         }
 151         if (xen_hvm_domain())
 152                 return;
 153
 154         if (per_cpu(xen_irq_work, cpu).irq >= 0) {
 155                 unbind_from_irqhandler(per_cpu(xen_irq_work, cpu).irq, NULL);
 156                 per_cpu(xen_irq_work, cpu).irq = -1;
 157                 kfree(per_cpu(xen_irq_work, cpu).name);
 158                 per_cpu(xen_irq_work, cpu).name = NULL;
 159         }
 160
 161         if (per_cpu(xen_pmu_irq, cpu).irq >= 0) {
 162                 unbind_from_irqhandler(per_cpu(xen_pmu_irq, cpu).irq, NULL);
 163                 per_cpu(xen_pmu_irq, cpu).irq = -1;
 164                 kfree(per_cpu(xen_pmu_irq, cpu).name);
 165                 per_cpu(xen_pmu_irq, cpu).name = NULL;
 166         }
 167 };
 168 static int xen_smp_intr_init(unsigned int cpu)
 169 {
 170         int rc;
 171         char *resched_name, *callfunc_name, *debug_name, *pmu_name;
 172
 173         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
 174         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
 175                                     cpu,
 176                                     xen_reschedule_interrupt,
 177                                     IRQF_PERCPU|IRQF_NOBALANCING,
 178                                     resched_name,
 179                                     NULL);
 180         if (rc < 0)
 181                 goto fail;
 182         per_cpu(xen_resched_irq, cpu).irq = rc;
 183         per_cpu(xen_resched_irq, cpu).name = resched_name;
 184
 185         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 186         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 187                                     cpu,
 188                                     xen_call_function_interrupt,
 189                                     IRQF_PERCPU|IRQF_NOBALANCING,
 190                                     callfunc_name,
 191                                     NULL);
 192         if (rc < 0)
 193                 goto fail;
 194         per_cpu(xen_callfunc_irq, cpu).irq = rc;
 195         per_cpu(xen_callfunc_irq, cpu).name = callfunc_name;
 196
 197         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 198         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 199                                      IRQF_PERCPU | IRQF_NOBALANCING,
 200                                      debug_name, NULL);
 201         if (rc < 0)
 202                 goto fail;
 203         per_cpu(xen_debug_irq, cpu).irq = rc;
 204         per_cpu(xen_debug_irq, cpu).name = debug_name;
 205
 206         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 207         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 208                                     cpu,
 209                                     xen_call_function_single_interrupt,
 210                                     IRQF_PERCPU|IRQF_NOBALANCING,
 211                                     callfunc_name,
 212                                     NULL);
 213         if (rc < 0)
 214                 goto fail;
 215         per_cpu(xen_callfuncsingle_irq, cpu).irq = rc;
 216         per_cpu(xen_callfuncsingle_irq, cpu).name = callfunc_name;
 217
 218         /*
 219          * The IRQ worker on PVHVM goes through the native path and uses the
 220          * IPI mechanism.
 221          */
 222         if (xen_hvm_domain())
 223                 return 0;
 224
 225         callfunc_name = kasprintf(GFP_KERNEL, "irqwork%d", cpu);
 226         rc = bind_ipi_to_irqhandler(XEN_IRQ_WORK_VECTOR,
 227                                     cpu,
 228                                     xen_irq_work_interrupt,
 229                                     IRQF_PERCPU|IRQF_NOBALANCING,
 230                                     callfunc_name,
 231                                     NULL);
 232         if (rc < 0)
 233                 goto fail;
 234         per_cpu(xen_irq_work, cpu).irq = rc;
 235         per_cpu(xen_irq_work, cpu).name = callfunc_name;
 236
 237         if (is_xen_pmu(cpu)) {
 238                 pmu_name = kasprintf(GFP_KERNEL, "pmu%d", cpu);
 239                 rc = bind_virq_to_irqhandler(VIRQ_XENPMU, cpu,
 240                                              xen_pmu_irq_handler,
 241                                              IRQF_PERCPU|IRQF_NOBALANCING,
 242                                              pmu_name, NULL);
 243                 if (rc < 0)
 244                         goto fail;
 245                 per_cpu(xen_pmu_irq, cpu).irq = rc;
 246                 per_cpu(xen_pmu_irq, cpu).name = pmu_name;
 247         }
 248
 249         return 0;
 250
 251  fail:
 252         xen_smp_intr_free(cpu);
 253         return rc;
 254 }
 255
 256 static void __init xen_fill_possible_map(void)
 257 {
 258         int i, rc;
 259
 260         if (xen_initial_domain())
 261                 return;
 262
 263         for (i = 0; i < nr_cpu_ids; i++) {
 264                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 265                 if (rc >= 0) {
 266                         num_processors++;
 267                         set_cpu_possible(i, true);
 268                 }
 269         }
 270 }
 271
 272 static void __init xen_filter_cpu_maps(void)
 273 {
 274         int i, rc;
 275         unsigned int subtract = 0;
 276
 277         if (!xen_initial_domain())
 278                 return;
 279
 280         num_processors = 0;
 281         disabled_cpus = 0;
 282         for (i = 0; i < nr_cpu_ids; i++) {
 283                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 284                 if (rc >= 0) {
 285                         num_processors++;
 286                         set_cpu_possible(i, true);
 287                 } else {
 288                         set_cpu_possible(i, false);
 289                         set_cpu_present(i, false);
 290                         subtract++;
 291                 }
 292         }
 293 #ifdef CONFIG_HOTPLUG_CPU
 294         /* This is akin to using 'nr_cpus' on the Linux command line.
 295          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 296          * have up to X, while nr_cpu_ids is greater than X. This
 297          * normally is not a problem, except when CPU hotplugging
 298          * is involved and then there might be more than X CPUs
 299          * in the guest - which will not work as there is no
 300          * hypercall to expand the max number of VCPUs an already
 301          * running guest has. So cap it up to X. */
 302         if (subtract)
 303                 nr_cpu_ids = nr_cpu_ids - subtract;
 304 #endif
 305
 306 }
 307
 308 static void __init xen_smp_prepare_boot_cpu(void)
 309 {
 310         BUG_ON(smp_processor_id() != 0);
 311         native_smp_prepare_boot_cpu();
 312
 313         if (xen_pv_domain()) {
 314                 if (!xen_feature(XENFEAT_writable_page_tables))
 315                         /* We've switched to the "real" per-cpu gdt, so make
 316                          * sure the old memory can be recycled. */
 317                         make_lowmem_page_readwrite(xen_initial_gdt);
 318
 319 #ifdef CONFIG_X86_32
 320                 /*
 321                  * Xen starts us with XEN_FLAT_RING1_DS, but linux code
 322                  * expects __USER_DS
 323                  */
 324                 loadsegment(ds, __USER_DS);
 325                 loadsegment(es, __USER_DS);
 326 #endif
 327
 328                 xen_filter_cpu_maps();
 329                 xen_setup_vcpu_info_placement();
 330         }
 331
 332         /*
 333          * Setup vcpu_info for boot CPU.
 334          */
 335         if (xen_hvm_domain())
 336                 xen_vcpu_setup(0);
 337
 338         /*
 339          * The alternative logic (which patches the unlock/lock) runs before
 340          * the smp bootup up code is activated. Hence we need to set this up
 341          * the core kernel is being patched. Otherwise we will have only
 342          * modules patched but not core code.
 343          */
 344         xen_init_spinlocks();
 345 }
 346
 347 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 348 {
 349         unsigned cpu;
 350         unsigned int i;
 351
 352         if (skip_ioapic_setup) {
 353                 char *m = (max_cpus == 0) ?
 354                         "The nosmp parameter is incompatible with Xen; " \
 355                         "use Xen dom0_max_vcpus=1 parameter" :
 356                         "The noapic parameter is incompatible with Xen";
 357
 358                 xen_raw_printk(m);
 359                 panic(m);
 360         }
 361         xen_init_lock_cpu(0);
 362
 363         smp_store_boot_cpu_info();
 364         cpu_data(0).x86_max_cores = 1;
 365
 366         for_each_possible_cpu(i) {
 367                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 368                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 369                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 370         }
 371         set_cpu_sibling_map(0);
 372
 373         xen_pmu_init(0);
 374
 375         if (xen_smp_intr_init(0))
 376                 BUG();
 377
 378         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 379                 panic("could not allocate xen_cpu_initialized_map\n");
 380
 381         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 382
 383         /* Restrict the possible_map according to max_cpus. */
 384         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 385                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 386                         continue;
 387                 set_cpu_possible(cpu, false);
 388         }
 389
 390         for_each_possible_cpu(cpu)
 391                 set_cpu_present(cpu, true);
 392 }
 393
 394 static int
 395 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 396 {
 397         struct vcpu_guest_context *ctxt;
 398         struct desc_struct *gdt;
 399         unsigned long gdt_mfn;
 400
 401         /* used to tell cpu_init() that it can proceed with initialization */
 402         cpumask_set_cpu(cpu, cpu_callout_mask);
 403         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 404                 return 0;
 405
 406         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 407         if (ctxt == NULL)
 408                 return -ENOMEM;
 409
 410         gdt = get_cpu_gdt_table(cpu);
 411
 412 #ifdef CONFIG_X86_32
 413         /* Note: PVH is not yet supported on x86_32. */
 414         ctxt->user_regs.fs = __KERNEL_PERCPU;
 415         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 416 #endif
 417         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 418
 419         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 420                 ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 421                 ctxt->flags = VGCF_IN_KERNEL;
 422                 ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 423                 ctxt->user_regs.ds = __USER_DS;
 424                 ctxt->user_regs.es = __USER_DS;
 425                 ctxt->user_regs.ss = __KERNEL_DS;
 426
 427                 xen_copy_trap_info(ctxt->trap_ctxt);
 428
 429                 ctxt->ldt_ents = 0;
 430
 431                 BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 432
 433                 gdt_mfn = arbitrary_virt_to_mfn(gdt);
 434                 make_lowmem_page_readonly(gdt);
 435                 make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 436
 437                 ctxt->gdt_frames[0] = gdt_mfn;
 438                 ctxt->gdt_ents      = GDT_ENTRIES;
 439
 440                 ctxt->kernel_ss = __KERNEL_DS;
 441                 ctxt->kernel_sp = idle->thread.sp0;
 442
 443 #ifdef CONFIG_X86_32
 444                 ctxt->event_callback_cs     = __KERNEL_CS;
 445                 ctxt->failsafe_callback_cs  = __KERNEL_CS;
 446 #else
 447                 ctxt->gs_base_kernel = per_cpu_offset(cpu);
 448 #endif
 449                 ctxt->event_callback_eip    =
 450                                         (unsigned long)xen_hypervisor_callback;
 451                 ctxt->failsafe_callback_eip =
 452                                         (unsigned long)xen_failsafe_callback;
 453                 ctxt->user_regs.cs = __KERNEL_CS;
 454                 per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 455         }
 456 #ifdef CONFIG_XEN_PVH
 457         else {
 458                 /*
 459                  * The vcpu comes on kernel page tables which have the NX pte
 460                  * bit set. This means before DS/SS is touched, NX in
 461                  * EFER must be set. Hence the following assembly glue code.
 462                  */
 463                 ctxt->user_regs.eip = (unsigned long)xen_pvh_early_cpu_init;
 464                 ctxt->user_regs.rdi = cpu;
 465                 ctxt->user_regs.rsi = true;  /* entry == true */
 466         }
 467 #endif
 468         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 469         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_gfn(swapper_pg_dir));
 470         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, xen_vcpu_nr(cpu), ctxt))
 471                 BUG();
 472
 473         kfree(ctxt);
 474         return 0;
 475 }
 476
 477 static int xen_cpu_up(unsigned int cpu, struct task_struct *idle)
 478 {
 479         int rc;
 480
 481         common_cpu_up(cpu, idle);
 482
 483         xen_setup_runstate_info(cpu);
 484         xen_setup_timer(cpu);
 485         xen_init_lock_cpu(cpu);
 486
 487         /*
 488          * PV VCPUs are always successfully taken down (see 'while' loop
 489          * in xen_cpu_die()), so -EBUSY is an error.
 490          */
 491         rc = cpu_check_up_prepare(cpu);
 492         if (rc)
 493                 return rc;
 494
 495         /* make sure interrupts start blocked */
 496         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 497
 498         rc = cpu_initialize_context(cpu, idle);
 499         if (rc)
 500                 return rc;
 501
 502         xen_pmu_init(cpu);
 503
 504         rc = xen_smp_intr_init(cpu);
 505         if (rc)
 506                 return rc;
 507
 508         rc = HYPERVISOR_vcpu_op(VCPUOP_up, xen_vcpu_nr(cpu), NULL);
 509         BUG_ON(rc);
 510
 511         while (cpu_report_state(cpu) != CPU_ONLINE)
 512                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 513
 514         return 0;
 515 }
 516
 517 static void xen_smp_cpus_done(unsigned int max_cpus)
 518 {
 519 }
 520
 521 #ifdef CONFIG_HOTPLUG_CPU
 522 static int xen_cpu_disable(void)
 523 {
 524         unsigned int cpu = smp_processor_id();
 525         if (cpu == 0)
 526                 return -EBUSY;
 527
 528         cpu_disable_common();
 529
 530         load_cr3(swapper_pg_dir);
 531         return 0;
 532 }
 533
 534 static void xen_cpu_die(unsigned int cpu)
 535 {
 536         while (xen_pv_domain() && HYPERVISOR_vcpu_op(VCPUOP_is_up,
 537                                                      xen_vcpu_nr(cpu), NULL)) {
 538                 __set_current_state(TASK_UNINTERRUPTIBLE);
 539                 schedule_timeout(HZ/10);
 540         }
 541
 542         if (common_cpu_die(cpu) == 0) {
 543                 xen_smp_intr_free(cpu);
 544                 xen_uninit_lock_cpu(cpu);
 545                 xen_teardown_timer(cpu);
 546                 xen_pmu_finish(cpu);
 547         }
 548 }
 549
 550 static void xen_play_dead(void) /* used only with HOTPLUG_CPU */
 551 {
 552         play_dead_common();
 553         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
 554         cpu_bringup();
 555         /*
 556          * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
 557          * clears certain data that the cpu_idle loop (which called us
 558          * and that we return from) expects. The only way to get that
 559          * data back is to call:
 560          */
 561         tick_nohz_idle_enter();
 562
 563         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 564 }
 565
 566 #else /* !CONFIG_HOTPLUG_CPU */
 567 static int xen_cpu_disable(void)
 568 {
 569         return -ENOSYS;
 570 }
 571
 572 static void xen_cpu_die(unsigned int cpu)
 573 {
 574         BUG();
 575 }
 576
 577 static void xen_play_dead(void)
 578 {
 579         BUG();
 580 }
 581
 582 #endif
 583 static void stop_self(void *v)
 584 {
 585         int cpu = smp_processor_id();
 586
 587         /* make sure we're not pinning something down */
 588         load_cr3(swapper_pg_dir);
 589         /* should set up a minimal gdt */
 590
 591         set_cpu_online(cpu, false);
 592
 593         HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(cpu), NULL);
 594         BUG();
 595 }
 596
 597 static void xen_stop_other_cpus(int wait)
 598 {
 599         smp_call_function(stop_self, NULL, wait);
 600 }
 601
 602 static void xen_smp_send_reschedule(int cpu)
 603 {
 604         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 605 }
 606
 607 static void __xen_send_IPI_mask(const struct cpumask *mask,
 608                               int vector)
 609 {
 610         unsigned cpu;
 611
 612         for_each_cpu_and(cpu, mask, cpu_online_mask)
 613                 xen_send_IPI_one(cpu, vector);
 614 }
 615
 616 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 617 {
 618         int cpu;
 619
 620         __xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 621
 622         /* Make sure other vcpus get a chance to run if they need to. */
 623         for_each_cpu(cpu, mask) {
 624                 if (xen_vcpu_stolen(cpu)) {
 625                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 626                         break;
 627                 }
 628         }
 629 }
 630
 631 static void xen_smp_send_call_function_single_ipi(int cpu)
 632 {
 633         __xen_send_IPI_mask(cpumask_of(cpu),
 634                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 635 }
 636
 637 static inline int xen_map_vector(int vector)
 638 {
 639         int xen_vector;
 640
 641         switch (vector) {
 642         case RESCHEDULE_VECTOR:
 643                 xen_vector = XEN_RESCHEDULE_VECTOR;
 644                 break;
 645         case CALL_FUNCTION_VECTOR:
 646                 xen_vector = XEN_CALL_FUNCTION_VECTOR;
 647                 break;
 648         case CALL_FUNCTION_SINGLE_VECTOR:
 649                 xen_vector = XEN_CALL_FUNCTION_SINGLE_VECTOR;
 650                 break;
 651         case IRQ_WORK_VECTOR:
 652                 xen_vector = XEN_IRQ_WORK_VECTOR;
 653                 break;
 654 #ifdef CONFIG_X86_64
 655         case NMI_VECTOR:
 656         case APIC_DM_NMI: /* Some use that instead of NMI_VECTOR */
 657                 xen_vector = XEN_NMI_VECTOR;
 658                 break;
 659 #endif
 660         default:
 661                 xen_vector = -1;
 662                 printk(KERN_ERR "xen: vector 0x%x is not implemented\n",
 663                         vector);
 664         }
 665
 666         return xen_vector;
 667 }
 668
 669 void xen_send_IPI_mask(const struct cpumask *mask,
 670                               int vector)
 671 {
 672         int xen_vector = xen_map_vector(vector);
 673
 674         if (xen_vector >= 0)
 675                 __xen_send_IPI_mask(mask, xen_vector);
 676 }
 677
 678 void xen_send_IPI_all(int vector)
 679 {
 680         int xen_vector = xen_map_vector(vector);
 681
 682         if (xen_vector >= 0)
 683                 __xen_send_IPI_mask(cpu_online_mask, xen_vector);
 684 }
 685
 686 void xen_send_IPI_self(int vector)
 687 {
 688         int xen_vector = xen_map_vector(vector);
 689
 690         if (xen_vector >= 0)
 691                 xen_send_IPI_one(smp_processor_id(), xen_vector);
 692 }
 693
 694 void xen_send_IPI_mask_allbutself(const struct cpumask *mask,
 695                                 int vector)
 696 {
 697         unsigned cpu;
 698         unsigned int this_cpu = smp_processor_id();
 699         int xen_vector = xen_map_vector(vector);
 700
 701         if (!(num_online_cpus() > 1) || (xen_vector < 0))
 702                 return;
 703
 704         for_each_cpu_and(cpu, mask, cpu_online_mask) {
 705                 if (this_cpu == cpu)
 706                         continue;
 707
 708                 xen_send_IPI_one(cpu, xen_vector);
 709         }
 710 }
 711
 712 void xen_send_IPI_allbutself(int vector)
 713 {
 714         xen_send_IPI_mask_allbutself(cpu_online_mask, vector);
 715 }
 716
 717 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 718 {
 719         irq_enter();
 720         generic_smp_call_function_interrupt();
 721         inc_irq_stat(irq_call_count);
 722         irq_exit();
 723
 724         return IRQ_HANDLED;
 725 }
 726
 727 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 728 {
 729         irq_enter();
 730         generic_smp_call_function_single_interrupt();
 731         inc_irq_stat(irq_call_count);
 732         irq_exit();
 733
 734         return IRQ_HANDLED;
 735 }
 736
 737 static irqreturn_t xen_irq_work_interrupt(int irq, void *dev_id)
 738 {
 739         irq_enter();
 740         irq_work_run();
 741         inc_irq_stat(apic_irq_work_irqs);
 742         irq_exit();
 743
 744         return IRQ_HANDLED;
 745 }
 746
 747 static const struct smp_ops xen_smp_ops __initconst = {
 748         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 749         .smp_prepare_cpus = xen_smp_prepare_cpus,
 750         .smp_cpus_done = xen_smp_cpus_done,
 751
 752         .cpu_up = xen_cpu_up,
 753         .cpu_die = xen_cpu_die,
 754         .cpu_disable = xen_cpu_disable,
 755         .play_dead = xen_play_dead,
 756
 757         .stop_other_cpus = xen_stop_other_cpus,
 758         .smp_send_reschedule = xen_smp_send_reschedule,
 759
 760         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 761         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 762 };
 763
 764 void __init xen_smp_init(void)
 765 {
 766         smp_ops = xen_smp_ops;
 767         xen_fill_possible_map();
 768 }
 769
 770 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
 771 {
 772         native_smp_prepare_cpus(max_cpus);
 773         WARN_ON(xen_smp_intr_init(0));
 774
 775         xen_init_lock_cpu(0);
 776 }
 777
 778 static int xen_hvm_cpu_up(unsigned int cpu, struct task_struct *tidle)
 779 {
 780         int rc;
 781
 782         /*
 783          * This can happen if CPU was offlined earlier and
 784          * offlining timed out in common_cpu_die().
 785          */
 786         if (cpu_report_state(cpu) == CPU_DEAD_FROZEN) {
 787                 xen_smp_intr_free(cpu);
 788                 xen_uninit_lock_cpu(cpu);
 789         }
 790
 791         /*
 792          * xen_smp_intr_init() needs to run before native_cpu_up()
 793          * so that IPI vectors are set up on the booting CPU before
 794          * it is marked online in native_cpu_up().
 795         */
 796         rc = xen_smp_intr_init(cpu);
 797         WARN_ON(rc);
 798         if (!rc)
 799                 rc =  native_cpu_up(cpu, tidle);
 800
 801         /*
 802          * We must initialize the slowpath CPU kicker _after_ the native
 803          * path has executed. If we initialized it before none of the
 804          * unlocker IPI kicks would reach the booting CPU as the booting
 805          * CPU had not set itself 'online' in cpu_online_mask. That mask
 806          * is checked when IPIs are sent (on HVM at least).
 807          */
 808         xen_init_lock_cpu(cpu);
 809         return rc;
 810 }
 811
 812 void __init xen_hvm_smp_init(void)
 813 {
 814         if (!xen_have_vector_callback)
 815                 return;
 816         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
 817         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
 818         smp_ops.cpu_up = xen_hvm_cpu_up;
 819         smp_ops.cpu_die = xen_cpu_die;
 820         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
 821         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
 822         smp_ops.smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu;
 823 }