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         preempt_disable();
  70
  71         xen_enable_sysenter();
  72         xen_enable_syscall();
  73
  74         cpu = smp_processor_id();
  75         smp_store_cpu_info(cpu);
  76         cpu_data(cpu).x86_max_cores = 1;
  77         set_cpu_sibling_map(cpu);
  78
  79         xen_setup_cpu_clockevents();
  80
  81         cpu_set(cpu, cpu_online_map);
  82         x86_write_percpu(cpu_state, CPU_ONLINE);
  83         wmb();
  84
  85         /* We can take interrupts now: we're officially "up". */
  86         local_irq_enable();
  87
  88         wmb();                  /* make sure everything is out */
  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_CPUS; i++) {
 158                 rc = HYPERVISOR_vcpu_op(VCPUOP_is_up, i, NULL);
 159                 if (rc >= 0) {
 160                         num_processors++;
 161                         cpu_set(i, cpu_possible_map);
 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(&per_cpu_var(gdt_page));
 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         smp_store_cpu_info(0);
 183         cpu_data(0).x86_max_cores = 1;
 184         set_cpu_sibling_map(0);
 185
 186         if (xen_smp_intr_init(0))
 187                 BUG();
 188
 189         xen_cpu_initialized_map = cpumask_of_cpu(0);
 190
 191         /* Restrict the possible_map according to max_cpus. */
 192         while ((num_possible_cpus() > 1) && (num_possible_cpus() > max_cpus)) {
 193                 for (cpu = NR_CPUS - 1; !cpu_possible(cpu); cpu--)
 194                         continue;
 195                 cpu_clear(cpu, cpu_possible_map);
 196         }
 197
 198         for_each_possible_cpu (cpu) {
 199                 struct task_struct *idle;
 200
 201                 if (cpu == 0)
 202                         continue;
 203
 204                 idle = fork_idle(cpu);
 205                 if (IS_ERR(idle))
 206                         panic("failed fork for CPU %d", cpu);
 207
 208                 cpu_set(cpu, cpu_present_map);
 209         }
 210
 211         //init_xenbus_allowed_cpumask();
 212 }
 213
 214 static __cpuinit int
 215 cpu_initialize_context(unsigned int cpu, struct task_struct *idle)
 216 {
 217         struct vcpu_guest_context *ctxt;
 218         struct desc_struct *gdt;
 219
 220         if (cpu_test_and_set(cpu, xen_cpu_initialized_map))
 221                 return 0;
 222
 223         ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
 224         if (ctxt == NULL)
 225                 return -ENOMEM;
 226
 227         gdt = get_cpu_gdt_table(cpu);
 228
 229         ctxt->flags = VGCF_IN_KERNEL;
 230         ctxt->user_regs.ds = __USER_DS;
 231         ctxt->user_regs.es = __USER_DS;
 232         ctxt->user_regs.ss = __KERNEL_DS;
 233 #ifdef CONFIG_X86_32
 234         ctxt->user_regs.fs = __KERNEL_PERCPU;
 235 #endif
 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 & ~PAGE_MASK);
 246         make_lowmem_page_readonly(gdt);
 247
 248         ctxt->gdt_frames[0] = virt_to_mfn(gdt);
 249         ctxt->gdt_ents      = GDT_ENTRIES;
 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 #ifdef CONFIG_X86_32
 258         ctxt->event_callback_cs     = __KERNEL_CS;
 259         ctxt->failsafe_callback_cs  = __KERNEL_CS;
 260 #endif
 261         ctxt->event_callback_eip    = (unsigned long)xen_hypervisor_callback;
 262         ctxt->failsafe_callback_eip = (unsigned long)xen_failsafe_callback;
 263
 264         per_cpu(xen_cr3, cpu) = __pa(swapper_pg_dir);
 265         ctxt->ctrlreg[3] = xen_pfn_to_cr3(virt_to_mfn(swapper_pg_dir));
 266
 267         if (HYPERVISOR_vcpu_op(VCPUOP_initialise, cpu, ctxt))
 268                 BUG();
 269
 270         kfree(ctxt);
 271         return 0;
 272 }
 273
 274 static int __cpuinit xen_cpu_up(unsigned int cpu)
 275 {
 276         struct task_struct *idle = idle_task(cpu);
 277         int rc;
 278
 279 #if 0
 280         rc = cpu_up_check(cpu);
 281         if (rc)
 282                 return rc;
 283 #endif
 284
 285 #ifdef CONFIG_X86_64
 286         /* Allocate node local memory for AP pdas */
 287         WARN_ON(cpu == 0);
 288         if (cpu > 0) {
 289                 rc = get_local_pda(cpu);
 290                 if (rc)
 291                         return rc;
 292         }
 293 #endif
 294
 295 #ifdef CONFIG_X86_32
 296         init_gdt(cpu);
 297         per_cpu(current_task, cpu) = idle;
 298         irq_ctx_init(cpu);
 299 #else
 300         cpu_pda(cpu)->pcurrent = idle;
 301         clear_tsk_thread_flag(idle, TIF_FORK);
 302 #endif
 303         xen_setup_timer(cpu);
 304
 305         per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
 306
 307         /* make sure interrupts start blocked */
 308         per_cpu(xen_vcpu, cpu)->evtchn_upcall_mask = 1;
 309
 310         rc = cpu_initialize_context(cpu, idle);
 311         if (rc)
 312                 return rc;
 313
 314         if (num_online_cpus() == 1)
 315                 alternatives_smp_switch(1);
 316
 317         rc = xen_smp_intr_init(cpu);
 318         if (rc)
 319                 return rc;
 320
 321         rc = HYPERVISOR_vcpu_op(VCPUOP_up, cpu, NULL);
 322         BUG_ON(rc);
 323
 324         while(per_cpu(cpu_state, cpu) != CPU_ONLINE) {
 325                 HYPERVISOR_sched_op(SCHEDOP_yield, 0);
 326                 barrier();
 327         }
 328
 329         return 0;
 330 }
 331
 332 static void xen_smp_cpus_done(unsigned int max_cpus)
 333 {
 334 }
 335
 336 static void stop_self(void *v)
 337 {
 338         int cpu = smp_processor_id();
 339
 340         /* make sure we're not pinning something down */
 341         load_cr3(swapper_pg_dir);
 342         /* should set up a minimal gdt */
 343
 344         HYPERVISOR_vcpu_op(VCPUOP_down, cpu, NULL);
 345         BUG();
 346 }
 347
 348 static void xen_smp_send_stop(void)
 349 {
 350         smp_call_function(stop_self, NULL, 0);
 351 }
 352
 353 static void xen_smp_send_reschedule(int cpu)
 354 {
 355         xen_send_IPI_one(cpu, XEN_RESCHEDULE_VECTOR);
 356 }
 357
 358 static void xen_send_IPI_mask(cpumask_t mask, enum ipi_vector vector)
 359 {
 360         unsigned cpu;
 361
 362         cpus_and(mask, mask, cpu_online_map);
 363
 364         for_each_cpu_mask(cpu, mask)
 365                 xen_send_IPI_one(cpu, vector);
 366 }
 367
 368 static void xen_smp_send_call_function_ipi(cpumask_t mask)
 369 {
 370         int cpu;
 371
 372         xen_send_IPI_mask(mask, XEN_CALL_FUNCTION_VECTOR);
 373
 374         /* Make sure other vcpus get a chance to run if they need to. */
 375         for_each_cpu_mask(cpu, mask) {
 376                 if (xen_vcpu_stolen(cpu)) {
 377                         HYPERVISOR_sched_op(SCHEDOP_yield, 0);
 378                         break;
 379                 }
 380         }
 381 }
 382
 383 static void xen_smp_send_call_function_single_ipi(int cpu)
 384 {
 385         xen_send_IPI_mask(cpumask_of_cpu(cpu), XEN_CALL_FUNCTION_SINGLE_VECTOR);
 386 }
 387
 388 static irqreturn_t xen_call_function_interrupt(int irq, void *dev_id)
 389 {
 390         irq_enter();
 391         generic_smp_call_function_interrupt();
 392 #ifdef CONFIG_X86_32
 393         __get_cpu_var(irq_stat).irq_call_count++;
 394 #else
 395         add_pda(irq_call_count, 1);
 396 #endif
 397         irq_exit();
 398
 399         return IRQ_HANDLED;
 400 }
 401
 402 static irqreturn_t xen_call_function_single_interrupt(int irq, void *dev_id)
 403 {
 404         irq_enter();
 405         generic_smp_call_function_single_interrupt();
 406 #ifdef CONFIG_X86_32
 407         __get_cpu_var(irq_stat).irq_call_count++;
 408 #else
 409         add_pda(irq_call_count, 1);
 410 #endif
 411         irq_exit();
 412
 413         return IRQ_HANDLED;
 414 }
 415
 416 static const struct smp_ops xen_smp_ops __initdata = {
 417         .smp_prepare_boot_cpu = xen_smp_prepare_boot_cpu,
 418         .smp_prepare_cpus = xen_smp_prepare_cpus,
 419         .cpu_up = xen_cpu_up,
 420         .smp_cpus_done = xen_smp_cpus_done,
 421
 422         .smp_send_stop = xen_smp_send_stop,
 423         .smp_send_reschedule = xen_smp_send_reschedule,
 424
 425         .send_call_func_ipi = xen_smp_send_call_function_ipi,
 426         .send_call_func_single_ipi = xen_smp_send_call_function_single_ipi,
 427 };
 428
 429 void __init xen_smp_init(void)
 430 {
 431         smp_ops = xen_smp_ops;
 432         xen_fill_possible_map();
 433 }