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 cpumask_t xen_cpu_initialized_map;
  39
  40 static DEFINE_PER_CPU(int, resched_irq);
  41 static DEFINE_PER_CPU(int, callfunc_irq);
  42 static DEFINE_PER_CPU(int, callfuncsingle_irq);
  43 static DEFINE_PER_CPU(int, debug_irq) = -1;
  44
  45 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id);
  46 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id);
  47
  48 /*
  49  * Reschedule call back. Nothing to do,
  50  * all the work is done automatically when
  51  * we return from the interrupt.
  52  */
  53 static irqreturn_t xen_reschedule_interrupt(int irq, void *dev_id)
  54 {
  55 #ifdef CONFIG_X86_32
  56         __get_cpu_var(irq_stat).irq_resched_count++;
  57 #else
  58         add_pda(irq_resched_count, 1);
  59 #endif
  60
  61         return IRQ_HANDLED;
  62 }
  63
  64 static __cpuinit void cpu_bringup_and_idle(void)
  65 {
  66         int cpu = smp_processor_id();
  67
  68         cpu_init();
  69         xen_enable_sysenter();
  70
  71         preempt_disable();
  72         per_cpu(cpu_state, cpu) = CPU_ONLINE;
  73
  74         xen_setup_cpu_clockevents();
  75
  76         /* We can take interrupts now: we're officially "up". */
  77         local_irq_enable();
  78
  79         wmb();                  /* make sure everything is out */
  80         cpu_idle();
  81 }
  82
  83 static int xen_smp_intr_init(unsigned int cpu)
  84 {
  85         int rc;
  86         const char *resched_name, *callfunc_name, *debug_name;
  87
  88         resched_name = kasprintf(GFP_KERNEL, "resched%d", cpu);
  89         rc = bind_ipi_to_irqhandler(XEN_RESCHEDULE_VECTOR,
  90                                     cpu,
  91                                     xen_reschedule_interrupt,
  92                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
  93                                     resched_name,
  94                                     NULL);
  95         if (rc < 0)
  96                 goto fail;
  97         per_cpu(resched_irq, cpu) = rc;
  98
  99         callfunc_name = kasprintf(GFP_KERNEL, "callfunc%d", cpu);
 100         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_VECTOR,
 101                                     cpu,
 102                                     xen_call_function_interrupt,
 103                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 104                                     callfunc_name,
 105                                     NULL);
 106         if (rc < 0)
 107                 goto fail;
 108         per_cpu(callfunc_irq, cpu) = rc;
 109
 110         debug_name = kasprintf(GFP_KERNEL, "debug%d", cpu);
 111         rc = bind_virq_to_irqhandler(VIRQ_DEBUG, cpu, xen_debug_interrupt,
 112                                      IRQF_DISABLED | IRQF_PERCPU | IRQF_NOBALANCING,
 113                                      debug_name, NULL);
 114         if (rc < 0)
 115                 goto fail;
 116         per_cpu(debug_irq, cpu) = rc;
 117
 118         callfunc_name = kasprintf(GFP_KERNEL, "callfuncsingle%d", cpu);
 119         rc = bind_ipi_to_irqhandler(XEN_CALL_FUNCTION_SINGLE_VECTOR,
 120                                     cpu,
 121                                     xen_call_function_single_interrupt,
 122                                     IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING,
 123                                     callfunc_name,
 124                                     NULL);
 125         if (rc < 0)
 126                 goto fail;
 127         per_cpu(callfuncsingle_irq, cpu) = rc;
 128
 129         return 0;
 130
 131  fail:
 132         if (per_cpu(resched_irq, cpu) >= 0)
 133                 unbind_from_irqhandler(per_cpu(resched_irq, cpu), NULL);
 134         if (per_cpu(callfunc_irq, cpu) >= 0)
 135                 unbind_from_irqhandler(per_cpu(callfunc_irq, cpu), NULL);
 136         if (per_cpu(debug_irq, cpu) >= 0)
 137                 unbind_from_irqhandler(per_cpu(debug_irq, cpu), NULL);
 138         if (per_cpu(callfuncsingle_irq, cpu) >= 0)
 139                 unbind_from_irqhandler(per_cpu(callfuncsingle_irq, cpu), NULL);
 140
 141         return rc;
 142 }
 143
 144 void __init xen_fill_possible_map(void)
 145 {
 146         int i, rc;
 147
 148         for (i = 0; i < NR_CPUS; i++) {
 149                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 150                 if (rc >= 0)
 151                         cpu_set(i, cpu_possible_map);
 152         }
 153 }
 154
 155 void __init xen_smp_prepare_boot_cpu(void)
 156 {
 157         int cpu;
 158
 159         BUG_ON(smp_processor_id() != 0);
 160         native_smp_prepare_boot_cpu();
 161
 162         /* We've switched to the "real" per-cpu gdt, so make sure the
 163            old memory can be recycled */
 164         make_lowmem_page_readwrite(&per_cpu__gdt_page);
 165
 166         for_each_possible_cpu(cpu) {
 167                 cpus_clear(per_cpu(cpu_sibling_map, cpu));
 168                 /*
 169                  * cpu_core_map lives in a per cpu area that is cleared
 170                  * when the per cpu array is allocated.
 171                  *
 172                  * cpus_clear(per_cpu(cpu_core_map, cpu));
 173                  */
 174         }
 175
 176         xen_setup_vcpu_info_placement();
 177 }
 178
 179 void __init xen_smp_prepare_cpus(unsigned int max_cpus)
 180 {
 181         unsigned cpu;
 182
 183         for_each_possible_cpu(cpu) {
 184                 cpus_clear(per_cpu(cpu_sibling_map, cpu));
 185                 /*
 186                  * cpu_core_ map will be zeroed when the per
 187                  * cpu area is allocated.
 188                  *
 189                  * cpus_clear(per_cpu(cpu_core_map, cpu));
 190                  */
 191         }
 192
 193         smp_store_cpu_info(0);
 194         set_cpu_sibling_map(0);
 195
 196         if (xen_smp_intr_init(0))
 197                 BUG();
 198
 199         xen_cpu_initialized_map = cpumask_of_cpu(0);
 200
 201         /* Restrict the possible_map according to max_cpus. */
 202         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 203                 for (cpu = NR_CPUS - 1; !cpu_possible(cpu); cpu--)
 204                         continue;
 205                 cpu_clear(cpu, cpu_possible_map);
 206         }
 207
 208         for_each_possible_cpu (cpu) {
 209                 struct task_struct *idle;
 210
 211                 if (cpu == 0)
 212                         continue;
 213
 214                 idle = fork_idle(cpu);
 215                 if (IS_ERR(idle))
 216                         panic("failed fork for CPU %d", cpu);
 217
 218                 cpu_set(cpu, cpu_present_map);
 219         }
 220
 221         //init_xenbus_allowed_cpumask();
 222 }
 223
 224 static __cpuinit int
 225 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 226 {
 227         struct vcpu_guest_context *ctxt;
 228         struct gdt_page *gdt = &per_cpu(gdt_page, cpu);
 229
 230         if (cpu_test_and_set(cpu, xen_cpu_initialized_map))
 231                 return 0;
 232
 233         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 234         if (ctxt == NULL)
 235                 return -ENOMEM;
 236
 237         ctxt->flags = VGCF_IN_KERNEL;
 238         ctxt->user_regs.ds = __USER_DS;
 239         ctxt->user_regs.es = __USER_DS;
 240         ctxt->user_regs.fs = __KERNEL_PERCPU;
 241         ctxt->user_regs.gs = 0;
 242         ctxt->user_regs.ss = __KERNEL_DS;
 243         ctxt->user_regs.eip = (unsigned long)cpu_bringup_and_idle;
 244         ctxt->user_regs.eflags = 0x1000; /* IOPL_RING1 */
 245
 246         memset(&ctxt->fpu_ctxt, 0, sizeof(ctxt->fpu_ctxt));
 247
 248         xen_copy_trap_info(ctxt->trap_ctxt);
 249
 250         ctxt->ldt_ents = 0;
 251
 252         BUG_ON((unsigned long)gdt->gdt & ~PAGE_MASK);
 253         make_lowmem_page_readonly(gdt->gdt);
 254
 255         ctxt->gdt_frames[0] = virt_to_mfn(gdt->gdt);
 256         ctxt->gdt_ents      = ARRAY_SIZE(gdt->gdt);
 257
 258         ctxt->user_regs.cs = __KERNEL_CS;
 259         ctxt->user_regs.esp = idle->thread.sp0 - sizeof(struct pt_regs);
 260
 261         ctxt->kernel_ss = __KERNEL_DS;
 262         ctxt->kernel_sp = idle->thread.sp0;
 263
 264         ctxt->event_callback_cs     = __KERNEL_CS;
 265         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 266         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 267         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 268
 269         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 270         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 271
 272         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 273                 BUG();
 274
 275         kfree(ctxt);
 276         return 0;
 277 }
 278
 279 int __cpuinit xen_cpu_up(unsigned int cpu)
 280 {
 281         struct task_struct *idle = idle_task(cpu);
 282         int rc;
 283
 284 #if 0
 285         rc = cpu_up_check(cpu);
 286         if (rc)
 287                 return rc;
 288 #endif
 289
 290         init_gdt(cpu);
 291         per_cpu(current_task, cpu) = idle;
 292         irq_ctx_init(cpu);
 293         xen_setup_timer(cpu);
 294
 295         /* make sure interrupts start blocked */
 296         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 297
 298         rc = cpu_initialize_context(cpu, idle);
 299         if (rc)
 300                 return rc;
 301
 302         if (num_online_cpus() == 1)
 303                 alternatives_smp_switch(1);
 304
 305         rc = xen_smp_intr_init(cpu);
 306         if (rc)
 307                 return rc;
 308
 309         smp_store_cpu_info(cpu);
 310         set_cpu_sibling_map(cpu);
 311         /* This must be done before setting cpu_online_map */
 312         wmb();
 313
 314         cpu_set(cpu, cpu_online_map);
 315
 316         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 317         BUG_ON(rc);
 318
 319         return 0;
 320 }
 321
 322 void xen_smp_cpus_done(unsigned int max_cpus)
 323 {
 324 }
 325
 326 static void stop_self(void *v)
 327 {
 328         int cpu = smp_processor_id();
 329
 330         /* make sure we're not pinning something down */
 331         load_cr3(swapper_pg_dir);
 332         /* should set up a minimal gdt */
 333
 334         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 335         BUG();
 336 }
 337
 338 void xen_smp_send_stop(void)
 339 {
 340         smp_call_function(stop_self, NULL, 0);
 341 }
 342
 343 void xen_smp_send_reschedule(int cpu)
 344 {
 345         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 346 }
 347
 348 static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector)
 349 {
 350         unsigned cpu;
 351
 352         cpus_and(mask, mask, cpu_online_map);
 353
 354         for_each_cpu_mask(cpu, mask)
 355                 xen_send_IPI_one(cpu, vector);
 356 }
 357
 358 void xen_smp_send_call_function_ipi(cpumask_t mask)
 359 {
 360         int cpu;
 361
 362         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 363
 364         /* Make sure other vcpus get a chance to run if they need to. */
 365         for_each_cpu_mask(cpu, mask) {
 366                 if (xen_vcpu_stolen(cpu)) {
 367                         HYPERVISOR_sched_op(SCHEDOP_yield, 0);
 368                         break;
 369                 }
 370         }
 371 }
 372
 373 void xen_smp_send_call_function_single_ipi(int cpu)
 374 {
 375         xen_send_IPI_mask(cpumask_of_cpu(cpu), XEN_CALL_FUNCTION_SINGLE_VECTOR);
 376 }
 377
 378 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 379 {
 380         irq_enter();
 381         generic_smp_call_function_interrupt();
 382         __get_cpu_var(irq_stat).irq_call_count++;
 383         irq_exit();
 384
 385         return IRQ_HANDLED;
 386 }
 387
 388 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 389 {
 390         irq_enter();
 391         generic_smp_call_function_single_interrupt();
 392         __get_cpu_var(irq_stat).irq_call_count++;
 393         irq_exit();
 394
 395         return IRQ_HANDLED;
 396 }