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/smp.h>
  18
  19 #include <asm/paravirt.h>
  20 #include <asm/desc.h>
  21 #include <asm/pgtable.h>
  22 #include <asm/cpu.h>
  23
  24 #include <xen/interface/xen.h>
  25 #include <xen/interface/vcpu.h>
  26
  27 #include <asm/xen/interface.h>
  28 #include <asm/xen/hypercall.h>
  29
  30 #include <xen/page.h>
  31 #include <xen/events.h>
  32
  33 #include "xen-ops.h"
  34 #include "mmu.h"
  35
  36 cpumask_var_t xen_cpu_initialized_map;
  37
  38 static DEFINE_PER_CPU(int, resched_irq);
  39 static DEFINE_PER_CPU(int, callfunc_irq);
  40 static DEFINE_PER_CPU(int, callfuncsingle_irq);
  41 static DEFINE_PER_CPU(int, debug_irq) = -1;
  42
  43 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  44 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
  45
  46 /*
  47  * Reschedule call back. Nothing to do,
  48  * all the work is done automatically when
  49  * we return from the interrupt.
  50  */
  51 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  52 {
  53         inc_irq_stat(irq_resched_count);
  54
  55         return IRQ_HANDLED;
  56 }
  57
  58 static __cpuinit void cpu_bringup(void)
  59 {
  60         int cpu = smp_processor_id();
  61
  62         cpu_init();
  63         touch_softlockup_watchdog();
  64         preempt_disable();
  65
  66         xen_enable_sysenter();
  67         xen_enable_syscall();
  68
  69         cpu = smp_processor_id();
  70         smp_store_cpu_info(cpu);
  71         cpu_data(cpu).x86_max_cores = 1;
  72         set_cpu_sibling_map(cpu);
  73
  74         xen_setup_cpu_clockevents();
  75
  76         cpu_set(cpu, cpu_online_map);
  77         percpu_write(cpu_state, CPU_ONLINE);
  78         wmb();
  79
  80         /* We can take interrupts now: we're officially "up". */
  81         local_irq_enable();
  82
  83         wmb();                  /* make sure everything is out */
  84 }
  85
  86 static __cpuinit void cpu_bringup_and_idle(void)
  87 {
  88         cpu_bringup();
  89         cpu_idle();
  90 }
  91
  92 static int xen_smp_intr_init(unsigned int cpu)
  93 {
  94         int rc;
  95         const char *resched_name, *callfunc_name, *debug_name;
  96
  97         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
  98         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
  99                                     cpu,
 100                                     xen_reschedule_interrupt,
 101                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 102                                     resched_name,
 103                                     NULL);
 104         if (rc < 0)
 105                 goto fail;
 106         per_cpu(resched_irq, cpu) = rc;
 107
 108         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 109         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 110                                     cpu,
 111                                     xen_call_function_interrupt,
 112                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 113                                     callfunc_name,
 114                                     NULL);
 115         if (rc < 0)
 116                 goto fail;
 117         per_cpu(callfunc_irq, cpu) = rc;
 118
 119         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 120         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 121                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
 122                                      debug_name, NULL);
 123         if (rc < 0)
 124                 goto fail;
 125         per_cpu(debug_irq, cpu) = rc;
 126
 127         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 128         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 129                                     cpu,
 130                                     xen_call_function_single_interrupt,
 131                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 132                                     callfunc_name,
 133                                     NULL);
 134         if (rc < 0)
 135                 goto fail;
 136         per_cpu(callfuncsingle_irq, cpu) = rc;
 137
 138         return 0;
 139
 140  fail:
 141         if (per_cpu(resched_irq, cpu) >= 0)
 142                 unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
 143         if (per_cpu(callfunc_irq, cpu) >= 0)
 144                 unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
 145         if (per_cpu(debug_irq, cpu) >= 0)
 146                 unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
 147         if (per_cpu(callfuncsingle_irq, cpu) >= 0)
 148                 unbind_from_irqhandler(per_cpu(callfuncsingle_irq, cpu), NULL);
 149
 150         return rc;
 151 }
 152
 153 static void __init xen_fill_possible_map(void)
 154 {
 155         int i, rc;
 156
 157         for (i = 0; i < nr_cpu_ids; i++) {
 158                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 159                 if (rc >= 0) {
 160                         num_processors++;
 161                         set_cpu_possible(i, true);
 162                 }
 163         }
 164 }
 165
 166 static void __init xen_smp_prepare_boot_cpu(void)
 167 {
 168         BUG_ON(smp_processor_id() != 0);
 169         native_smp_prepare_boot_cpu();
 170
 171         /* We've switched to the "real" per-cpu gdt, so make sure the
 172            old memory can be recycled */
 173         make_lowmem_page_readwrite(xen_initial_gdt);
 174
 175         xen_setup_vcpu_info_placement();
 176 }
 177
 178 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 179 {
 180         unsigned cpu;
 181
 182         xen_init_lock_cpu(0);
 183
 184         smp_store_cpu_info(0);
 185         cpu_data(0).x86_max_cores = 1;
 186         set_cpu_sibling_map(0);
 187
 188         if (xen_smp_intr_init(0))
 189                 BUG();
 190
 191         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 192                 panic("could not allocate xen_cpu_initialized_map\n");
 193
 194         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 195
 196         /* Restrict the possible_map according to max_cpus. */
 197         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 198                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 199                         continue;
 200                 set_cpu_possible(cpu, false);
 201         }
 202
 203         for_each_possible_cpu (cpu) {
 204                 struct task_struct *idle;
 205
 206                 if (cpu == 0)
 207                         continue;
 208
 209                 idle = fork_idle(cpu);
 210                 if (IS_ERR(idle))
 211                         panic("failed fork for CPU %d", cpu);
 212
 213                 set_cpu_present(cpu, true);
 214         }
 215 }
 216
 217 static __cpuinit int
 218 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 219 {
 220         struct vcpu_guest_context *ctxt;
 221         struct desc_struct *gdt;
 222         unsigned long gdt_mfn;
 223
 224         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 225                 return 0;
 226
 227         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 228         if (ctxt == NULL)
 229                 return -ENOMEM;
 230
 231         gdt = get_cpu_gdt_table(cpu);
 232
 233         ctxt->flags = VGCF_IN_KERNEL;
 234         ctxt->user_regs.ds = __USER_DS;
 235         ctxt->user_regs.es = __USER_DS;
 236         ctxt->user_regs.ss = __KERNEL_DS;
 237 #ifdef CONFIG_X86_32
 238         ctxt->user_regs.fs = __KERNEL_PERCPU;
 239 #else
 240         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 241 #endif
 242         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 243         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 244
 245         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 246
 247         xen_copy_trap_info(ctxt->trap_ctxt);
 248
 249         ctxt->ldt_ents = 0;
 250
 251         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 252
 253         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 254         make_lowmem_page_readonly(gdt);
 255         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 256
 257         ctxt->gdt_frames[0] = gdt_mfn;
 258         ctxt->gdt_ents      = GDT_ENTRIES;
 259
 260         ctxt->user_regs.cs = __KERNEL_CS;
 261         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 262
 263         ctxt->kernel_ss = __KERNEL_DS;
 264         ctxt->kernel_sp = idle->thread.sp0;
 265
 266 #ifdef CONFIG_X86_32
 267         ctxt->event_callback_cs     = __KERNEL_CS;
 268         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 269 #endif
 270         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 271         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 272
 273         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 274         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 275
 276         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 277                 BUG();
 278
 279         kfree(ctxt);
 280         return 0;
 281 }
 282
 283 static int __cpuinit xen_cpu_up(unsigned int cpu)
 284 {
 285         struct task_struct *idle = idle_task(cpu);
 286         int rc;
 287
 288         per_cpu(current_task, cpu) = idle;
 289 #ifdef CONFIG_X86_32
 290         irq_ctx_init(cpu);
 291 #else
 292         clear_tsk_thread_flag(idle, TIF_FORK);
 293         per_cpu(kernel_stack, cpu) =
 294                 (unsigned long)task_stack_page(idle) -
 295                 KERNEL_STACK_OFFSET + THREAD_SIZE;
 296 #endif
 297         xen_setup_timer(cpu);
 298         xen_init_lock_cpu(cpu);
 299
 300         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 301
 302         /* make sure interrupts start blocked */
 303         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 304
 305         rc = cpu_initialize_context(cpu, idle);
 306         if (rc)
 307                 return rc;
 308
 309         if (num_online_cpus() == 1)
 310                 alternatives_smp_switch(1);
 311
 312         rc = xen_smp_intr_init(cpu);
 313         if (rc)
 314                 return rc;
 315
 316         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 317         BUG_ON(rc);
 318
 319         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
 320                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 321                 barrier();
 322         }
 323
 324         return 0;
 325 }
 326
 327 static void xen_smp_cpus_done(unsigned int max_cpus)
 328 {
 329 }
 330
 331 #ifdef CONFIG_HOTPLUG_CPU
 332 static int xen_cpu_disable(void)
 333 {
 334         unsigned int cpu = smp_processor_id();
 335         if (cpu == 0)
 336                 return -EBUSY;
 337
 338         cpu_disable_common();
 339
 340         load_cr3(swapper_pg_dir);
 341         return 0;
 342 }
 343
 344 static void xen_cpu_die(unsigned int cpu)
 345 {
 346         while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
 347                 current->state = TASK_UNINTERRUPTIBLE;
 348                 schedule_timeout(HZ/10);
 349         }
 350         unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
 351         unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
 352         unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
 353         unbind_from_irqhandler(per_cpu(callfuncsingle_irq, cpu), NULL);
 354         xen_uninit_lock_cpu(cpu);
 355         xen_teardown_timer(cpu);
 356
 357         if (num_online_cpus() == 1)
 358                 alternatives_smp_switch(0);
 359 }
 360
 361 static void __cpuinit xen_play_dead(void) /* used only with CPU_HOTPLUG */
 362 {
 363         play_dead_common();
 364         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 365         cpu_bringup();
 366 }
 367
 368 #else /* !CONFIG_HOTPLUG_CPU */
 369 static int xen_cpu_disable(void)
 370 {
 371         return -ENOSYS;
 372 }
 373
 374 static void xen_cpu_die(unsigned int cpu)
 375 {
 376         BUG();
 377 }
 378
 379 static void xen_play_dead(void)
 380 {
 381         BUG();
 382 }
 383
 384 #endif
 385 static void stop_self(void *v)
 386 {
 387         int cpu = smp_processor_id();
 388
 389         /* make sure we're not pinning something down */
 390         load_cr3(swapper_pg_dir);
 391         /* should set up a minimal gdt */
 392
 393         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 394         BUG();
 395 }
 396
 397 static void xen_smp_send_stop(void)
 398 {
 399         smp_call_function(stop_self, NULL, 0);
 400 }
 401
 402 static void xen_smp_send_reschedule(int cpu)
 403 {
 404         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 405 }
 406
 407 static void xen_send_IPI_mask(const struct cpumask *mask,
 408                               enum ipi_vector vector)
 409 {
 410         unsigned cpu;
 411
 412         for_each_cpu_and(cpu, mask, cpu_online_mask)
 413                 xen_send_IPI_one(cpu, vector);
 414 }
 415
 416 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 417 {
 418         int cpu;
 419
 420         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 421
 422         /* Make sure other vcpus get a chance to run if they need to. */
 423         for_each_cpu(cpu, mask) {
 424                 if (xen_vcpu_stolen(cpu)) {
 425                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 426                         break;
 427                 }
 428         }
 429 }
 430
 431 static void xen_smp_send_call_function_single_ipi(int cpu)
 432 {
 433         xen_send_IPI_mask(cpumask_of(cpu),
 434                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 435 }
 436
 437 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 438 {
 439         irq_enter();
 440         generic_smp_call_function_interrupt();
 441         inc_irq_stat(irq_call_count);
 442         irq_exit();
 443
 444         return IRQ_HANDLED;
 445 }
 446
 447 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 448 {
 449         irq_enter();
 450         generic_smp_call_function_single_interrupt();
 451         inc_irq_stat(irq_call_count);
 452         irq_exit();
 453
 454         return IRQ_HANDLED;
 455 }
 456
 457 static const struct smp_ops xen_smp_ops __initdata = {
 458         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 459         .smp_prepare_cpus = xen_smp_prepare_cpus,
 460         .smp_cpus_done = xen_smp_cpus_done,
 461
 462         .cpu_up = xen_cpu_up,
 463         .cpu_die = xen_cpu_die,
 464         .cpu_disable = xen_cpu_disable,
 465         .play_dead = xen_play_dead,
 466
 467         .smp_send_stop = xen_smp_send_stop,
 468         .smp_send_reschedule = xen_smp_send_reschedule,
 469
 470         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 471         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 472 };
 473
 474 void __init xen_smp_init(void)
 475 {
 476         smp_ops = xen_smp_ops;
 477         xen_fill_possible_map();
 478         xen_init_spinlocks();
 479 }