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