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 = smp_processor_id();
  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         set_cpu_online(cpu, true);
  79         percpu_write(cpu_state, CPU_ONLINE);
  80         wmb();
  81
  82         /* We can take interrupts now: we're officially "up". */
  83         local_irq_enable();
  84
  85         wmb();                  /* make sure everything is out */
  86 }
  87
  88 static void __cpuinit cpu_bringup_and_idle(void)
  89 {
  90         cpu_bringup();
  91         cpu_idle();
  92 }
  93
  94 static int xen_smp_intr_init(unsigned int cpu)
  95 {
  96         int rc;
  97         const char *resched_name, *callfunc_name, *debug_name;
  98
  99         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
 100         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
 101                                     cpu,
 102                                     xen_reschedule_interrupt,
 103                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 104                                     resched_name,
 105                                     NULL);
 106         if (rc < 0)
 107                 goto fail;
 108         per_cpu(xen_resched_irq, cpu) = rc;
 109
 110         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 111         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 112                                     cpu,
 113                                     xen_call_function_interrupt,
 114                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 115                                     callfunc_name,
 116                                     NULL);
 117         if (rc < 0)
 118                 goto fail;
 119         per_cpu(xen_callfunc_irq, cpu) = rc;
 120
 121         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 122         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 123                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
 124                                      debug_name, NULL);
 125         if (rc < 0)
 126                 goto fail;
 127         per_cpu(xen_debug_irq, cpu) = rc;
 128
 129         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 130         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 131                                     cpu,
 132                                     xen_call_function_single_interrupt,
 133                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 134                                     callfunc_name,
 135                                     NULL);
 136         if (rc < 0)
 137                 goto fail;
 138         per_cpu(xen_callfuncsingle_irq, cpu) = rc;
 139
 140         return 0;
 141
 142  fail:
 143         if (per_cpu(xen_resched_irq, cpu) >= 0)
 144                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
 145         if (per_cpu(xen_callfunc_irq, cpu) >= 0)
 146                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
 147         if (per_cpu(xen_debug_irq, cpu) >= 0)
 148                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
 149         if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0)
 150                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu),
 151                                        NULL);
 152
 153         return rc;
 154 }
 155
 156 static void __init xen_fill_possible_map(void)
 157 {
 158         int i, rc;
 159
 160         if (xen_initial_domain())
 161                 return;
 162
 163         for (i = 0; i < nr_cpu_ids; i++) {
 164                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 165                 if (rc >= 0) {
 166                         num_processors++;
 167                         set_cpu_possible(i, true);
 168                 }
 169         }
 170 }
 171
 172 static void __init xen_filter_cpu_maps(void)
 173 {
 174         int i, rc;
 175         unsigned int subtract = 0;
 176
 177         if (!xen_initial_domain())
 178                 return;
 179
 180         num_processors = 0;
 181         disabled_cpus = 0;
 182         for (i = 0; i < nr_cpu_ids; i++) {
 183                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 184                 if (rc >= 0) {
 185                         num_processors++;
 186                         set_cpu_possible(i, true);
 187                 } else {
 188                         set_cpu_possible(i, false);
 189                         set_cpu_present(i, false);
 190                         subtract++;
 191                 }
 192         }
 193 #ifdef CONFIG_HOTPLUG_CPU
 194         /* This is akin to using 'nr_cpus' on the Linux command line.
 195          * Which is OK as when we use 'dom0_max_vcpus=X' we can only
 196          * have up to X, while nr_cpu_ids is greater than X. This
 197          * normally is not a problem, except when CPU hotplugging
 198          * is involved and then there might be more than X CPUs
 199          * in the guest - which will not work as there is no
 200          * hypercall to expand the max number of VCPUs an already
 201          * running guest has. So cap it up to X. */
 202         if (subtract)
 203                 nr_cpu_ids = nr_cpu_ids - subtract;
 204 #endif
 205
 206 }
 207
 208 static void __init xen_smp_prepare_boot_cpu(void)
 209 {
 210         BUG_ON(smp_processor_id() != 0);
 211         native_smp_prepare_boot_cpu();
 212
 213         /* We've switched to the "real" per-cpu gdt, so make sure the
 214            old memory can be recycled */
 215         make_lowmem_page_readwrite(xen_initial_gdt);
 216
 217         xen_filter_cpu_maps();
 218         xen_setup_vcpu_info_placement();
 219 }
 220
 221 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 222 {
 223         unsigned cpu;
 224         unsigned int i;
 225
 226         if (skip_ioapic_setup) {
 227                 char *m = (max_cpus == 0) ?
 228                         "The nosmp parameter is incompatible with Xen; " \
 229                         "use Xen dom0_max_vcpus=1 parameter" :
 230                         "The noapic parameter is incompatible with Xen";
 231
 232                 xen_raw_printk(m);
 233                 panic(m);
 234         }
 235         xen_init_lock_cpu(0);
 236
 237         smp_store_cpu_info(0);
 238         cpu_data(0).x86_max_cores = 1;
 239
 240         for_each_possible_cpu(i) {
 241                 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
 242                 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
 243                 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
 244         }
 245         set_cpu_sibling_map(0);
 246
 247         if (xen_smp_intr_init(0))
 248                 BUG();
 249
 250         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 251                 panic("could not allocate xen_cpu_initialized_map\n");
 252
 253         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 254
 255         /* Restrict the possible_map according to max_cpus. */
 256         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 257                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 258                         continue;
 259                 set_cpu_possible(cpu, false);
 260         }
 261
 262         for_each_possible_cpu (cpu) {
 263                 struct task_struct *idle;
 264
 265                 if (cpu == 0)
 266                         continue;
 267
 268                 idle = fork_idle(cpu);
 269                 if (IS_ERR(idle))
 270                         panic("failed fork for CPU %d", cpu);
 271
 272                 set_cpu_present(cpu, true);
 273         }
 274 }
 275
 276 static int __cpuinit
 277 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 278 {
 279         struct vcpu_guest_context *ctxt;
 280         struct desc_struct *gdt;
 281         unsigned long gdt_mfn;
 282
 283         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 284                 return 0;
 285
 286         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 287         if (ctxt == NULL)
 288                 return -ENOMEM;
 289
 290         gdt = get_cpu_gdt_table(cpu);
 291
 292         ctxt->flags = VGCF_IN_KERNEL;
 293         ctxt->user_regs.ds = __USER_DS;
 294         ctxt->user_regs.es = __USER_DS;
 295         ctxt->user_regs.ss = __KERNEL_DS;
 296 #ifdef CONFIG_X86_32
 297         ctxt->user_regs.fs = __KERNEL_PERCPU;
 298         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 299 #else
 300         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 301 #endif
 302         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 303         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 304
 305         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 306
 307         xen_copy_trap_info(ctxt->trap_ctxt);
 308
 309         ctxt->ldt_ents = 0;
 310
 311         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 312
 313         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 314         make_lowmem_page_readonly(gdt);
 315         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 316
 317         ctxt->gdt_frames[0] = gdt_mfn;
 318         ctxt->gdt_ents      = GDT_ENTRIES;
 319
 320         ctxt->user_regs.cs = __KERNEL_CS;
 321         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 322
 323         ctxt->kernel_ss = __KERNEL_DS;
 324         ctxt->kernel_sp = idle->thread.sp0;
 325
 326 #ifdef CONFIG_X86_32
 327         ctxt->event_callback_cs     = __KERNEL_CS;
 328         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 329 #endif
 330         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 331         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 332
 333         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 334         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 335
 336         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 337                 BUG();
 338
 339         kfree(ctxt);
 340         return 0;
 341 }
 342
 343 static int __cpuinit xen_cpu_up(unsigned int cpu)
 344 {
 345         struct task_struct *idle = idle_task(cpu);
 346         int rc;
 347
 348         per_cpu(current_task, cpu) = idle;
 349 #ifdef CONFIG_X86_32
 350         irq_ctx_init(cpu);
 351 #else
 352         clear_tsk_thread_flag(idle, TIF_FORK);
 353         per_cpu(kernel_stack, cpu) =
 354                 (unsigned long)task_stack_page(idle) -
 355                 KERNEL_STACK_OFFSET + THREAD_SIZE;
 356 #endif
 357         xen_setup_runstate_info(cpu);
 358         xen_setup_timer(cpu);
 359         xen_init_lock_cpu(cpu);
 360
 361         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 362
 363         /* make sure interrupts start blocked */
 364         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 365
 366         rc = cpu_initialize_context(cpu, idle);
 367         if (rc)
 368                 return rc;
 369
 370         if (num_online_cpus() == 1)
 371                 alternatives_smp_switch(1);
 372
 373         rc = xen_smp_intr_init(cpu);
 374         if (rc)
 375                 return rc;
 376
 377         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 378         BUG_ON(rc);
 379
 380         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
 381                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 382                 barrier();
 383         }
 384
 385         return 0;
 386 }
 387
 388 static void xen_smp_cpus_done(unsigned int max_cpus)
 389 {
 390 }
 391
 392 #ifdef CONFIG_HOTPLUG_CPU
 393 static int xen_cpu_disable(void)
 394 {
 395         unsigned int cpu = smp_processor_id();
 396         if (cpu == 0)
 397                 return -EBUSY;
 398
 399         cpu_disable_common();
 400
 401         load_cr3(swapper_pg_dir);
 402         return 0;
 403 }
 404
 405 static void xen_cpu_die(unsigned int cpu)
 406 {
 407         while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
 408                 current->state = TASK_UNINTERRUPTIBLE;
 409                 schedule_timeout(HZ/10);
 410         }
 411         unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
 412         unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
 413         unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
 414         unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
 415         xen_uninit_lock_cpu(cpu);
 416         xen_teardown_timer(cpu);
 417
 418         if (num_online_cpus() == 1)
 419                 alternatives_smp_switch(0);
 420 }
 421
 422 static void __cpuinit xen_play_dead(void) /* used only with HOTPLUG_CPU */
 423 {
 424         play_dead_common();
 425         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 426         cpu_bringup();
 427         /*
 428          * Balance out the preempt calls - as we are running in cpu_idle
 429          * loop which has been called at bootup from cpu_bringup_and_idle.
 430          * The cpucpu_bringup_and_idle called cpu_bringup which made a
 431          * preempt_disable() So this preempt_enable will balance it out.
 432          */
 433         preempt_enable();
 434 }
 435
 436 #else /* !CONFIG_HOTPLUG_CPU */
 437 static int xen_cpu_disable(void)
 438 {
 439         return -ENOSYS;
 440 }
 441
 442 static void xen_cpu_die(unsigned int cpu)
 443 {
 444         BUG();
 445 }
 446
 447 static void xen_play_dead(void)
 448 {
 449         BUG();
 450 }
 451
 452 #endif
 453 static void stop_self(void *v)
 454 {
 455         int cpu = smp_processor_id();
 456
 457         /* make sure we're not pinning something down */
 458         load_cr3(swapper_pg_dir);
 459         /* should set up a minimal gdt */
 460
 461         set_cpu_online(cpu, false);
 462
 463         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 464         BUG();
 465 }
 466
 467 static void xen_stop_other_cpus(int wait)
 468 {
 469         smp_call_function(stop_self, NULL, wait);
 470 }
 471
 472 static void xen_smp_send_reschedule(int cpu)
 473 {
 474         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 475 }
 476
 477 static void xen_send_IPI_mask(const struct cpumask *mask,
 478                               enum ipi_vector vector)
 479 {
 480         unsigned cpu;
 481
 482         for_each_cpu_and(cpu, mask, cpu_online_mask)
 483                 xen_send_IPI_one(cpu, vector);
 484 }
 485
 486 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 487 {
 488         int cpu;
 489
 490         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 491
 492         /* Make sure other vcpus get a chance to run if they need to. */
 493         for_each_cpu(cpu, mask) {
 494                 if (xen_vcpu_stolen(cpu)) {
 495                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 496                         break;
 497                 }
 498         }
 499 }
 500
 501 static void xen_smp_send_call_function_single_ipi(int cpu)
 502 {
 503         xen_send_IPI_mask(cpumask_of(cpu),
 504                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 505 }
 506
 507 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 508 {
 509         irq_enter();
 510         generic_smp_call_function_interrupt();
 511         inc_irq_stat(irq_call_count);
 512         irq_exit();
 513
 514         return IRQ_HANDLED;
 515 }
 516
 517 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 518 {
 519         irq_enter();
 520         generic_smp_call_function_single_interrupt();
 521         inc_irq_stat(irq_call_count);
 522         irq_exit();
 523
 524         return IRQ_HANDLED;
 525 }
 526
 527 static const struct smp_ops xen_smp_ops __initconst = {
 528         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 529         .smp_prepare_cpus = xen_smp_prepare_cpus,
 530         .smp_cpus_done = xen_smp_cpus_done,
 531
 532         .cpu_up = xen_cpu_up,
 533         .cpu_die = xen_cpu_die,
 534         .cpu_disable = xen_cpu_disable,
 535         .play_dead = xen_play_dead,
 536
 537         .stop_other_cpus = xen_stop_other_cpus,
 538         .smp_send_reschedule = xen_smp_send_reschedule,
 539
 540         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 541         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 542 };
 543
 544 void __init xen_smp_init(void)
 545 {
 546         smp_ops = xen_smp_ops;
 547         xen_fill_possible_map();
 548         xen_init_spinlocks();
 549 }
 550
 551 static void __init xen_hvm_smp_prepare_cpus(unsigned int max_cpus)
 552 {
 553         native_smp_prepare_cpus(max_cpus);
 554         WARN_ON(xen_smp_intr_init(0));
 555
 556         xen_init_lock_cpu(0);
 557 }
 558
 559 static int __cpuinit xen_hvm_cpu_up(unsigned int cpu)
 560 {
 561         int rc;
 562         rc = native_cpu_up(cpu);
 563         WARN_ON (xen_smp_intr_init(cpu));
 564         return rc;
 565 }
 566
 567 static void xen_hvm_cpu_die(unsigned int cpu)
 568 {
 569         unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
 570         unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
 571         unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
 572         unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
 573         native_cpu_die(cpu);
 574 }
 575
 576 void __init xen_hvm_smp_init(void)
 577 {
 578         if (!xen_have_vector_callback)
 579                 return;
 580         smp_ops.smp_prepare_cpus = xen_hvm_smp_prepare_cpus;
 581         smp_ops.smp_send_reschedule = xen_smp_send_reschedule;
 582         smp_ops.cpu_up = xen_hvm_cpu_up;
 583         smp_ops.cpu_die = xen_hvm_cpu_die;
 584         smp_ops.send_call_func_ipi = xen_smp_send_call_function_ipi;
 585         smp_ops.send_call_func_single_ipi = xen_smp_send_call_function_single_ipi;
 586 }