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  * This does not handle HOTPLUG_CPU yet.
  16  */
  17 #include <linux/sched.h>
  18 #include <linux/err.h>
  19 #include <linux/smp.h>
  20
  21 #include <asm/paravirt.h>
  22 #include <asm/desc.h>
  23 #include <asm/pgtable.h>
  24 #include <asm/cpu.h>
  25
  26 #include <xen/interface/xen.h>
  27 #include <xen/interface/vcpu.h>
  28
  29 #include <asm/xen/interface.h>
  30 #include <asm/xen/hypercall.h>
  31
  32 #include <xen/page.h>
  33 #include <xen/events.h>
  34
  35 #include "xen-ops.h"
  36 #include "mmu.h"
  37
  38 static cpumask_t xen_cpu_initialized_map;
  39 static DEFINE_PER_CPU(int, resched_irq) = -1;
  40 static DEFINE_PER_CPU(int, callfunc_irq) = -1;
  41 static DEFINE_PER_CPU(int, debug_irq) = -1;
  42
  43 /*
  44  * Structure and data for smp_call_function(). This is designed to minimise
  45  * static memory requirements. It also looks cleaner.
  46  */
  47 static DEFINE_SPINLOCK(call_lock);
  48
  49 struct call_data_struct {
  50         void (*func) (void *info);
  51         void *info;
  52         atomic_t started;
  53         atomic_t finished;
  54         int wait;
  55 };
  56
  57 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  58
  59 static struct call_data_struct *call_data;
  60
  61 /*
  62  * Reschedule call back. Nothing to do,
  63  * all the work is done automatically when
  64  * we return from the interrupt.
  65  */
  66 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  67 {
  68         return IRQ_HANDLED;
  69 }
  70
  71 static __cpuinit void cpu_bringup_and_idle(void)
  72 {
  73         int cpu = smp_processor_id();
  74
  75         cpu_init();
  76         xen_enable_sysenter();
  77
  78         preempt_disable();
  79         per_cpu(cpu_state, cpu) = CPU_ONLINE;
  80
  81         xen_setup_cpu_clockevents();
  82
  83         /* We can take interrupts now: we're officially "up". */
  84         local_irq_enable();
  85
  86         wmb();                  /* make sure everything is out */
  87         cpu_idle();
  88 }
  89
  90 static int xen_smp_intr_init(unsigned int cpu)
  91 {
  92         int rc;
  93         const char *resched_name, *callfunc_name, *debug_name;
  94
  95         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
  96         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
  97                                     cpu,
  98                                     xen_reschedule_interrupt,
  99                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 100                                     resched_name,
 101                                     NULL);
 102         if (rc < 0)
 103                 goto fail;
 104         per_cpu(resched_irq, cpu) = rc;
 105
 106         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 107         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 108                                     cpu,
 109                                     xen_call_function_interrupt,
 110                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 111                                     callfunc_name,
 112                                     NULL);
 113         if (rc < 0)
 114                 goto fail;
 115         per_cpu(callfunc_irq, cpu) = rc;
 116
 117         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 118         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 119                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
 120                                      debug_name, NULL);
 121         if (rc < 0)
 122                 goto fail;
 123         per_cpu(debug_irq, cpu) = rc;
 124
 125         return 0;
 126
 127  fail:
 128         if (per_cpu(resched_irq, cpu) >= 0)
 129                 unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
 130         if (per_cpu(callfunc_irq, cpu) >= 0)
 131                 unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
 132         if (per_cpu(debug_irq, cpu) >= 0)
 133                 unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
 134         return rc;
 135 }
 136
 137 void __init xen_fill_possible_map(void)
 138 {
 139         int i, rc;
 140
 141         for (i = 0; i < NR_CPUS; i++) {
 142                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 143                 if (rc >= 0)
 144                         cpu_set(i, cpu_possible_map);
 145         }
 146 }
 147
 148 void __init xen_smp_prepare_boot_cpu(void)
 149 {
 150         int cpu;
 151
 152         BUG_ON(smp_processor_id() != 0);
 153         native_smp_prepare_boot_cpu();
 154
 155         /* We've switched to the "real" per-cpu gdt, so make sure the
 156            old memory can be recycled */
 157         make_lowmem_page_readwrite(&per_cpu__gdt_page);
 158
 159         for_each_possible_cpu(cpu) {
 160                 cpus_clear(per_cpu(cpu_sibling_map, cpu));
 161                 /*
 162                  * cpu_core_map lives in a per cpu area that is cleared
 163                  * when the per cpu array is allocated.
 164                  *
 165                  * cpus_clear(per_cpu(cpu_core_map, cpu));
 166                  */
 167         }
 168
 169         xen_setup_vcpu_info_placement();
 170 }
 171
 172 void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 173 {
 174         unsigned cpu;
 175
 176         for_each_possible_cpu(cpu) {
 177                 cpus_clear(per_cpu(cpu_sibling_map, cpu));
 178                 /*
 179                  * cpu_core_ map will be zeroed when the per
 180                  * cpu area is allocated.
 181                  *
 182                  * cpus_clear(per_cpu(cpu_core_map, cpu));
 183                  */
 184         }
 185
 186         smp_store_cpu_info(0);
 187         set_cpu_sibling_map(0);
 188
 189         if (xen_smp_intr_init(0))
 190                 BUG();
 191
 192         xen_cpu_initialized_map = cpumask_of_cpu(0);
 193
 194         /* Restrict the possible_map according to max_cpus. */
 195         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 196                 for (cpu = NR_CPUS - 1; !cpu_possible(cpu); cpu--)
 197                         continue;
 198                 cpu_clear(cpu, cpu_possible_map);
 199         }
 200
 201         for_each_possible_cpu (cpu) {
 202                 struct task_struct *idle;
 203
 204                 if (cpu == 0)
 205                         continue;
 206
 207                 idle = fork_idle(cpu);
 208                 if (IS_ERR(idle))
 209                         panic("failed fork for CPU %d", cpu);
 210
 211                 cpu_set(cpu, cpu_present_map);
 212         }
 213
 214         //init_xenbus_allowed_cpumask();
 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 gdt_page *gdt = &per_cpu(gdt_page, cpu);
 222
 223         if (cpu_test_and_set(cpu, xen_cpu_initialized_map))
 224                 return 0;
 225
 226         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 227         if (ctxt == NULL)
 228                 return -ENOMEM;
 229
 230         ctxt->flags = VGCF_IN_KERNEL;
 231         ctxt->user_regs.ds = __USER_DS;
 232         ctxt->user_regs.es = __USER_DS;
 233         ctxt->user_regs.fs = __KERNEL_PERCPU;
 234         ctxt->user_regs.gs = 0;
 235         ctxt->user_regs.ss = __KERNEL_DS;
 236         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 237         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 238
 239         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 240
 241         xen_copy_trap_info(ctxt->trap_ctxt);
 242
 243         ctxt->ldt_ents = 0;
 244
 245         BUG_ON((unsigned long)gdt->gdt & ~PAGE_MASK);
 246         make_lowmem_page_readonly(gdt->gdt);
 247
 248         ctxt->gdt_frames[0] = virt_to_mfn(gdt->gdt);
 249         ctxt->gdt_ents      = ARRAY_SIZE(gdt->gdt);
 250
 251         ctxt->user_regs.cs = __KERNEL_CS;
 252         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 253
 254         ctxt->kernel_ss = __KERNEL_DS;
 255         ctxt->kernel_sp = idle->thread.sp0;
 256
 257         ctxt->event_callback_cs     = __KERNEL_CS;
 258         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 259         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 260         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 261
 262         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 263         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 264
 265         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 266                 BUG();
 267
 268         kfree(ctxt);
 269         return 0;
 270 }
 271
 272 int __cpuinit xen_cpu_up(unsigned int cpu)
 273 {
 274         struct task_struct *idle = idle_task(cpu);
 275         int rc;
 276
 277 #if 0
 278         rc = cpu_up_check(cpu);
 279         if (rc)
 280                 return rc;
 281 #endif
 282
 283         init_gdt(cpu);
 284         per_cpu(current_task, cpu) = idle;
 285         irq_ctx_init(cpu);
 286         xen_setup_timer(cpu);
 287
 288         /* make sure interrupts start blocked */
 289         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 290
 291         rc = cpu_initialize_context(cpu, idle);
 292         if (rc)
 293                 return rc;
 294
 295         if (num_online_cpus() == 1)
 296                 alternatives_smp_switch(1);
 297
 298         rc = xen_smp_intr_init(cpu);
 299         if (rc)
 300                 return rc;
 301
 302         smp_store_cpu_info(cpu);
 303         set_cpu_sibling_map(cpu);
 304         /* This must be done before setting cpu_online_map */
 305         wmb();
 306
 307         cpu_set(cpu, cpu_online_map);
 308
 309         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 310         BUG_ON(rc);
 311
 312         return 0;
 313 }
 314
 315 void xen_smp_cpus_done(unsigned int max_cpus)
 316 {
 317 }
 318
 319 static void stop_self(void *v)
 320 {
 321         int cpu = smp_processor_id();
 322
 323         /* make sure we're not pinning something down */
 324         load_cr3(swapper_pg_dir);
 325         /* should set up a minimal gdt */
 326
 327         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 328         BUG();
 329 }
 330
 331 void xen_smp_send_stop(void)
 332 {
 333         smp_call_function(stop_self, NULL, 0, 0);
 334 }
 335
 336 void xen_smp_send_reschedule(int cpu)
 337 {
 338         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 339 }
 340
 341
 342 static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector)
 343 {
 344         unsigned cpu;
 345
 346         cpus_and(mask, mask, cpu_online_map);
 347
 348         for_each_cpu_mask(cpu, mask)
 349                 xen_send_IPI_one(cpu, vector);
 350 }
 351
 352 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 353 {
 354         void (*func) (void *info) = call_data->func;
 355         void *info = call_data->info;
 356         int wait = call_data->wait;
 357
 358         /*
 359          * Notify initiating CPU that I've grabbed the data and am
 360          * about to execute the function
 361          */
 362         mb();
 363         atomic_inc(&call_data->started);
 364         /*
 365          * At this point the info structure may be out of scope unless wait==1
 366          */
 367         irq_enter();
 368         (*func)(info);
 369         __get_cpu_var(irq_stat).irq_call_count++;
 370         irq_exit();
 371
 372         if (wait) {
 373                 mb();           /* commit everything before setting finished */
 374                 atomic_inc(&call_data->finished);
 375         }
 376
 377         return IRQ_HANDLED;
 378 }
 379
 380 int xen_smp_call_function_mask(cpumask_t mask, void (*func)(void *),
 381                                void *info, int wait)
 382 {
 383         struct call_data_struct data;
 384         int cpus, cpu;
 385         bool yield;
 386
 387         /* Holding any lock stops cpus from going down. */
 388         spin_lock(&call_lock);
 389
 390         cpu_clear(smp_processor_id(), mask);
 391
 392         cpus = cpus_weight(mask);
 393         if (!cpus) {
 394                 spin_unlock(&call_lock);
 395                 return 0;
 396         }
 397
 398         /* Can deadlock when called with interrupts disabled */
 399         WARN_ON(irqs_disabled());
 400
 401         data.func = func;
 402         data.info = info;
 403         atomic_set(&data.started, 0);
 404         data.wait = wait;
 405         if (wait)
 406                 atomic_set(&data.finished, 0);
 407
 408         call_data = &data;
 409         mb();                   /* write everything before IPI */
 410
 411         /* Send a message to other CPUs and wait for them to respond */
 412         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 413
 414         /* Make sure other vcpus get a chance to run if they need to. */
 415         yield = false;
 416         for_each_cpu_mask(cpu, mask)
 417                 if (xen_vcpu_stolen(cpu))
 418                         yield = true;
 419
 420         if (yield)
 421                 HYPERVISOR_sched_op(SCHEDOP_yield, 0);
 422
 423         /* Wait for response */
 424         while (atomic_read(&data.started) != cpus ||
 425                (wait && atomic_read(&data.finished) != cpus))
 426                 cpu_relax();
 427
 428         spin_unlock(&call_lock);
 429
 430         return 0;
 431 }