SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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, xen_resched_irq);
  39 static DEFINE_PER_CPU(int, xen_callfunc_irq);
  40 static DEFINE_PER_CPU(int, xen_callfuncsingle_irq);
  41 static DEFINE_PER_CPU(int, xen_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         set_cpu_online(cpu, true);
  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(xen_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(xen_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(xen_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(xen_callfuncsingle_irq, cpu) = rc;
 137
 138         return 0;
 139
 140  fail:
 141         if (per_cpu(xen_resched_irq, cpu) >= 0)
 142                 unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
 143         if (per_cpu(xen_callfunc_irq, cpu) >= 0)
 144                 unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
 145         if (per_cpu(xen_debug_irq, cpu) >= 0)
 146                 unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
 147         if (per_cpu(xen_callfuncsingle_irq, cpu) >= 0)
 148                 unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu),
 149                                        NULL);
 150
 151         return rc;
 152 }
 153
 154 static void __init xen_fill_possible_map(void)
 155 {
 156         int i, rc;
 157
 158         for (i = 0; i < nr_cpu_ids; i++) {
 159                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 160                 if (rc >= 0) {
 161                         num_processors++;
 162                         set_cpu_possible(i, true);
 163                 }
 164         }
 165 }
 166
 167 static void __init xen_smp_prepare_boot_cpu(void)
 168 {
 169         BUG_ON(smp_processor_id() != 0);
 170         native_smp_prepare_boot_cpu();
 171
 172         /* We've switched to the "real" per-cpu gdt, so make sure the
 173            old memory can be recycled */
 174         make_lowmem_page_readwrite(xen_initial_gdt);
 175
 176         xen_setup_vcpu_info_placement();
 177 }
 178
 179 static void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 180 {
 181         unsigned cpu;
 182
 183         xen_init_lock_cpu(0);
 184
 185         smp_store_cpu_info(0);
 186         cpu_data(0).x86_max_cores = 1;
 187         set_cpu_sibling_map(0);
 188
 189         if (xen_smp_intr_init(0))
 190                 BUG();
 191
 192         if (!alloc_cpumask_var(&xen_cpu_initialized_map, GFP_KERNEL))
 193                 panic("could not allocate xen_cpu_initialized_map\n");
 194
 195         cpumask_copy(xen_cpu_initialized_map, cpumask_of(0));
 196
 197         /* Restrict the possible_map according to max_cpus. */
 198         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 199                 for (cpu = nr_cpu_ids - 1; !cpu_possible(cpu); cpu--)
 200                         continue;
 201                 set_cpu_possible(cpu, false);
 202         }
 203
 204         for_each_possible_cpu (cpu) {
 205                 struct task_struct *idle;
 206
 207                 if (cpu == 0)
 208                         continue;
 209
 210                 idle = fork_idle(cpu);
 211                 if (IS_ERR(idle))
 212                         panic("failed fork for CPU %d", cpu);
 213
 214                 set_cpu_present(cpu, true);
 215         }
 216 }
 217
 218 static __cpuinit int
 219 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 220 {
 221         struct vcpu_guest_context *ctxt;
 222         struct desc_struct *gdt;
 223         unsigned long gdt_mfn;
 224
 225         if (cpumask_test_and_set_cpu(cpu, xen_cpu_initialized_map))
 226                 return 0;
 227
 228         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 229         if (ctxt == NULL)
 230                 return -ENOMEM;
 231
 232         gdt = get_cpu_gdt_table(cpu);
 233
 234         ctxt->flags = VGCF_IN_KERNEL;
 235         ctxt->user_regs.ds = __USER_DS;
 236         ctxt->user_regs.es = __USER_DS;
 237         ctxt->user_regs.ss = __KERNEL_DS;
 238 #ifdef CONFIG_X86_32
 239         ctxt->user_regs.fs = __KERNEL_PERCPU;
 240         ctxt->user_regs.gs = __KERNEL_STACK_CANARY;
 241 #else
 242         ctxt->gs_base_kernel = per_cpu_offset(cpu);
 243 #endif
 244         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 245         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 246
 247         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 248
 249         xen_copy_trap_info(ctxt->trap_ctxt);
 250
 251         ctxt->ldt_ents = 0;
 252
 253         BUG_ON((unsigned long)gdt & ~PAGE_MASK);
 254
 255         gdt_mfn = arbitrary_virt_to_mfn(gdt);
 256         make_lowmem_page_readonly(gdt);
 257         make_lowmem_page_readonly(mfn_to_virt(gdt_mfn));
 258
 259         ctxt->gdt_frames[0] = gdt_mfn;
 260         ctxt->gdt_ents      = GDT_ENTRIES;
 261
 262         ctxt->user_regs.cs = __KERNEL_CS;
 263         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 264
 265         ctxt->kernel_ss = __KERNEL_DS;
 266         ctxt->kernel_sp = idle->thread.sp0;
 267
 268 #ifdef CONFIG_X86_32
 269         ctxt->event_callback_cs     = __KERNEL_CS;
 270         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 271 #endif
 272         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 273         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 274
 275         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 276         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 277
 278         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 279                 BUG();
 280
 281         kfree(ctxt);
 282         return 0;
 283 }
 284
 285 static int __cpuinit xen_cpu_up(unsigned int cpu)
 286 {
 287         struct task_struct *idle = idle_task(cpu);
 288         int rc;
 289
 290         per_cpu(current_task, cpu) = idle;
 291 #ifdef CONFIG_X86_32
 292         irq_ctx_init(cpu);
 293 #else
 294         clear_tsk_thread_flag(idle, TIF_FORK);
 295         per_cpu(kernel_stack, cpu) =
 296                 (unsigned long)task_stack_page(idle) -
 297                 KERNEL_STACK_OFFSET + THREAD_SIZE;
 298 #endif
 299         xen_setup_runstate_info(cpu);
 300         xen_setup_timer(cpu);
 301         xen_init_lock_cpu(cpu);
 302
 303         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 304
 305         /* make sure interrupts start blocked */
 306         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 307
 308         rc = cpu_initialize_context(cpu, idle);
 309         if (rc)
 310                 return rc;
 311
 312         if (num_online_cpus() == 1)
 313                 alternatives_smp_switch(1);
 314
 315         rc = xen_smp_intr_init(cpu);
 316         if (rc)
 317                 return rc;
 318
 319         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 320         BUG_ON(rc);
 321
 322         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
 323                 HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 324                 barrier();
 325         }
 326
 327         return 0;
 328 }
 329
 330 static void xen_smp_cpus_done(unsigned int max_cpus)
 331 {
 332 }
 333
 334 #ifdef CONFIG_HOTPLUG_CPU
 335 static int xen_cpu_disable(void)
 336 {
 337         unsigned int cpu = smp_processor_id();
 338         if (cpu == 0)
 339                 return -EBUSY;
 340
 341         cpu_disable_common();
 342
 343         load_cr3(swapper_pg_dir);
 344         return 0;
 345 }
 346
 347 static void xen_cpu_die(unsigned int cpu)
 348 {
 349         while (HYPERVISOR_vcpu_op(VCPUOP_is_up, cpu, NULL)) {
 350                 current->state = TASK_UNINTERRUPTIBLE;
 351                 schedule_timeout(HZ/10);
 352         }
 353         unbind_from_irqhandler(per_cpu(xen_resched_irq, cpu), NULL);
 354         unbind_from_irqhandler(per_cpu(xen_callfunc_irq, cpu), NULL);
 355         unbind_from_irqhandler(per_cpu(xen_debug_irq, cpu), NULL);
 356         unbind_from_irqhandler(per_cpu(xen_callfuncsingle_irq, cpu), NULL);
 357         xen_uninit_lock_cpu(cpu);
 358         xen_teardown_timer(cpu);
 359
 360         if (num_online_cpus() == 1)
 361                 alternatives_smp_switch(0);
 362 }
 363
 364 static void __cpuinit xen_play_dead(void) /* used only with CPU_HOTPLUG */
 365 {
 366         play_dead_common();
 367         HYPERVISOR_vcpu_op(VCPUOP_down, smp_processor_id(), NULL);
 368         cpu_bringup();
 369 }
 370
 371 #else /* !CONFIG_HOTPLUG_CPU */
 372 static int xen_cpu_disable(void)
 373 {
 374         return -ENOSYS;
 375 }
 376
 377 static void xen_cpu_die(unsigned int cpu)
 378 {
 379         BUG();
 380 }
 381
 382 static void xen_play_dead(void)
 383 {
 384         BUG();
 385 }
 386
 387 #endif
 388 static void stop_self(void *v)
 389 {
 390         int cpu = smp_processor_id();
 391
 392         /* make sure we're not pinning something down */
 393         load_cr3(swapper_pg_dir);
 394         /* should set up a minimal gdt */
 395
 396         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 397         BUG();
 398 }
 399
 400 static void xen_smp_send_stop(void)
 401 {
 402         smp_call_function(stop_self, NULL, 0);
 403 }
 404
 405 static void xen_smp_send_reschedule(int cpu)
 406 {
 407         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 408 }
 409
 410 static void xen_send_IPI_mask(const struct cpumask *mask,
 411                               enum ipi_vector vector)
 412 {
 413         unsigned cpu;
 414
 415         for_each_cpu_and(cpu, mask, cpu_online_mask)
 416                 xen_send_IPI_one(cpu, vector);
 417 }
 418
 419 static void xen_smp_send_call_function_ipi(const struct cpumask *mask)
 420 {
 421         int cpu;
 422
 423         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 424
 425         /* Make sure other vcpus get a chance to run if they need to. */
 426         for_each_cpu(cpu, mask) {
 427                 if (xen_vcpu_stolen(cpu)) {
 428                         HYPERVISOR_sched_op(SCHEDOP_yield, NULL);
 429                         break;
 430                 }
 431         }
 432 }
 433
 434 static void xen_smp_send_call_function_single_ipi(int cpu)
 435 {
 436         xen_send_IPI_mask(cpumask_of(cpu),
 437                           XEN_CALL_FUNCTION_SINGLE_VECTOR);
 438 }
 439
 440 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 441 {
 442         irq_enter();
 443         generic_smp_call_function_interrupt();
 444         inc_irq_stat(irq_call_count);
 445         irq_exit();
 446
 447         return IRQ_HANDLED;
 448 }
 449
 450 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 451 {
 452         irq_enter();
 453         generic_smp_call_function_single_interrupt();
 454         inc_irq_stat(irq_call_count);
 455         irq_exit();
 456
 457         return IRQ_HANDLED;
 458 }
 459
 460 static const struct smp_ops xen_smp_ops __initdata = {
 461         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 462         .smp_prepare_cpus = xen_smp_prepare_cpus,
 463         .smp_cpus_done = xen_smp_cpus_done,
 464
 465         .cpu_up = xen_cpu_up,
 466         .cpu_die = xen_cpu_die,
 467         .cpu_disable = xen_cpu_disable,
 468         .play_dead = xen_play_dead,
 469
 470         .smp_send_stop = xen_smp_send_stop,
 471         .smp_send_reschedule = xen_smp_send_reschedule,
 472
 473         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 474         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 475 };
 476
 477 void __init xen_smp_init(void)
 478 {
 479         smp_ops = xen_smp_ops;
 480         xen_fill_possible_map();
 481         xen_init_spinlocks();
 482 }