arch/x86/kernel/process.c

   1 #include <linux/errno.h>
   2 #include <linux/kernel.h>
   3 #include <linux/mm.h>
   4 #include <linux/smp.h>
   5 #include <linux/prctl.h>
   6 #include <linux/slab.h>
   7 #include <linux/sched.h>
   8 #include <linux/module.h>
   9 #include <linux/pm.h>
  10 #include <linux/clockchips.h>
  11 #include <linux/random.h>
  12 #include <trace/events/power.h>
  13 #include <asm/system.h>
  14 #include <asm/apic.h>
  15 #include <asm/syscalls.h>
  16 #include <asm/idle.h>
  17 #include <asm/uaccess.h>
  18 #include <asm/i387.h>
  19 #include <asm/ds.h>
  20 #include <asm/debugreg.h>
  21 #include <asm/hw_breakpoint.h>
  22
  23 unsigned long idle_halt;
  24 EXPORT_SYMBOL(idle_halt);
  25 unsigned long idle_nomwait;
  26 EXPORT_SYMBOL(idle_nomwait);
  27
  28 struct kmem_cache *task_xstate_cachep;
  29
  30 int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
  31 {
  32         *dst = *src;
  33         if (src->thread.xstate) {
  34                 dst->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
  35                                                       GFP_KERNEL);
  36                 if (!dst->thread.xstate)
  37                         return -ENOMEM;
  38                 WARN_ON((unsigned long)dst->thread.xstate & 15);
  39                 memcpy(dst->thread.xstate, src->thread.xstate, xstate_size);
  40         }
  41         return 0;
  42 }
  43
  44 void free_thread_xstate(struct task_struct *tsk)
  45 {
  46         if (tsk->thread.xstate) {
  47                 kmem_cache_free(task_xstate_cachep, tsk->thread.xstate);
  48                 tsk->thread.xstate = NULL;
  49         }
  50         if (unlikely(test_tsk_thread_flag(tsk, TIF_DEBUG)))
  51                 flush_thread_hw_breakpoint(tsk);
  52
  53         WARN(tsk->thread.ds_ctx, "leaking DS context\n");
  54 }
  55
  56 void free_thread_info(struct thread_info *ti)
  57 {
  58         free_thread_xstate(ti->task);
  59         free_pages((unsigned long)ti, get_order(THREAD_SIZE));
  60 }
  61
  62 void arch_task_cache_init(void)
  63 {
  64         task_xstate_cachep =
  65                 kmem_cache_create("task_xstate", xstate_size,
  66                                   __alignof__(union thread_xstate),
  67                                   SLAB_PANIC | SLAB_NOTRACK, NULL);
  68 }
  69
  70 /*
  71  * Free current thread data structures etc..
  72  */
  73 void exit_thread(void)
  74 {
  75         struct task_struct *me = current;
  76         struct thread_struct *t = &me->thread;
  77         unsigned long *bp = t->io_bitmap_ptr;
  78
  79         if (bp) {
  80                 struct tss_struct *tss = &per_cpu(init_tss, get_cpu());
  81
  82                 t->io_bitmap_ptr = NULL;
  83                 clear_thread_flag(TIF_IO_BITMAP);
  84                 /*
  85                  * Careful, clear this in the TSS too:
  86                  */
  87                 memset(tss->io_bitmap, 0xff, t->io_bitmap_max);
  88                 t->io_bitmap_max = 0;
  89                 put_cpu();
  90                 kfree(bp);
  91         }
  92 }
  93
  94 void flush_thread(void)
  95 {
  96         struct task_struct *tsk = current;
  97
  98 #ifdef CONFIG_X86_64
  99         if (test_tsk_thread_flag(tsk, TIF_ABI_PENDING)) {
 100                 clear_tsk_thread_flag(tsk, TIF_ABI_PENDING);
 101                 if (test_tsk_thread_flag(tsk, TIF_IA32)) {
 102                         clear_tsk_thread_flag(tsk, TIF_IA32);
 103                 } else {
 104                         set_tsk_thread_flag(tsk, TIF_IA32);
 105                         current_thread_info()->status |= TS_COMPAT;
 106                 }
 107         }
 108 #endif
 109
 110         clear_tsk_thread_flag(tsk, TIF_DEBUG);
 111
 112         if (unlikely(test_tsk_thread_flag(tsk, TIF_DEBUG)))
 113                 flush_thread_hw_breakpoint(tsk);
 114         memset(tsk->thread.tls_array, 0, sizeof(tsk->thread.tls_array));
 115         /*
 116          * Forget coprocessor state..
 117          */
 118         tsk->fpu_counter = 0;
 119         clear_fpu(tsk);
 120         clear_used_math();
 121 }
 122
 123 static void hard_disable_TSC(void)
 124 {
 125         write_cr4(read_cr4() | X86_CR4_TSD);
 126 }
 127
 128 void disable_TSC(void)
 129 {
 130         preempt_disable();
 131         if (!test_and_set_thread_flag(TIF_NOTSC))
 132                 /*
 133                  * Must flip the CPU state synchronously with
 134                  * TIF_NOTSC in the current running context.
 135                  */
 136                 hard_disable_TSC();
 137         preempt_enable();
 138 }
 139
 140 static void hard_enable_TSC(void)
 141 {
 142         write_cr4(read_cr4() & ~X86_CR4_TSD);
 143 }
 144
 145 static void enable_TSC(void)
 146 {
 147         preempt_disable();
 148         if (test_and_clear_thread_flag(TIF_NOTSC))
 149                 /*
 150                  * Must flip the CPU state synchronously with
 151                  * TIF_NOTSC in the current running context.
 152                  */
 153                 hard_enable_TSC();
 154         preempt_enable();
 155 }
 156
 157 int get_tsc_mode(unsigned long adr)
 158 {
 159         unsigned int val;
 160
 161         if (test_thread_flag(TIF_NOTSC))
 162                 val = PR_TSC_SIGSEGV;
 163         else
 164                 val = PR_TSC_ENABLE;
 165
 166         return put_user(val, (unsigned int __user *)adr);
 167 }
 168
 169 int set_tsc_mode(unsigned int val)
 170 {
 171         if (val == PR_TSC_SIGSEGV)
 172                 disable_TSC();
 173         else if (val == PR_TSC_ENABLE)
 174                 enable_TSC();
 175         else
 176                 return -EINVAL;
 177
 178         return 0;
 179 }
 180
 181 void __switch_to_xtra(struct task_struct *prev_p, struct task_struct *next_p,
 182                       struct tss_struct *tss)
 183 {
 184         struct thread_struct *prev, *next;
 185
 186         prev = &prev_p->thread;
 187         next = &next_p->thread;
 188
 189         if (test_tsk_thread_flag(next_p, TIF_DS_AREA_MSR) ||
 190             test_tsk_thread_flag(prev_p, TIF_DS_AREA_MSR))
 191                 ds_switch_to(prev_p, next_p);
 192         else if (next->debugctlmsr != prev->debugctlmsr)
 193                 update_debugctlmsr(next->debugctlmsr);
 194
 195         if (test_tsk_thread_flag(prev_p, TIF_NOTSC) ^
 196             test_tsk_thread_flag(next_p, TIF_NOTSC)) {
 197                 /* prev and next are different */
 198                 if (test_tsk_thread_flag(next_p, TIF_NOTSC))
 199                         hard_disable_TSC();
 200                 else
 201                         hard_enable_TSC();
 202         }
 203
 204         if (test_tsk_thread_flag(next_p, TIF_IO_BITMAP)) {
 205                 /*
 206                  * Copy the relevant range of the IO bitmap.
 207                  * Normally this is 128 bytes or less:
 208                  */
 209                 memcpy(tss->io_bitmap, next->io_bitmap_ptr,
 210                        max(prev->io_bitmap_max, next->io_bitmap_max));
 211         } else if (test_tsk_thread_flag(prev_p, TIF_IO_BITMAP)) {
 212                 /*
 213                  * Clear any possible leftover bits:
 214                  */
 215                 memset(tss->io_bitmap, 0xff, prev->io_bitmap_max);
 216         }
 217 }
 218
 219 int sys_fork(struct pt_regs *regs)
 220 {
 221         return do_fork(SIGCHLD, regs->sp, regs, 0, NULL, NULL);
 222 }
 223
 224 /*
 225  * This is trivial, and on the face of it looks like it
 226  * could equally well be done in user mode.
 227  *
 228  * Not so, for quite unobvious reasons - register pressure.
 229  * In user mode vfork() cannot have a stack frame, and if
 230  * done by calling the "clone()" system call directly, you
 231  * do not have enough call-clobbered registers to hold all
 232  * the information you need.
 233  */
 234 int sys_vfork(struct pt_regs *regs)
 235 {
 236         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->sp, regs, 0,
 237                        NULL, NULL);
 238 }
 239
 240
 241 /*
 242  * Idle related variables and functions
 243  */
 244 unsigned long boot_option_idle_override = 0;
 245 EXPORT_SYMBOL(boot_option_idle_override);
 246
 247 /*
 248  * Powermanagement idle function, if any..
 249  */
 250 void (*pm_idle)(void);
 251 EXPORT_SYMBOL(pm_idle);
 252
 253 #ifdef CONFIG_X86_32
 254 /*
 255  * This halt magic was a workaround for ancient floppy DMA
 256  * wreckage. It should be safe to remove.
 257  */
 258 static int hlt_counter;
 259 void disable_hlt(void)
 260 {
 261         hlt_counter++;
 262 }
 263 EXPORT_SYMBOL(disable_hlt);
 264
 265 void enable_hlt(void)
 266 {
 267         hlt_counter--;
 268 }
 269 EXPORT_SYMBOL(enable_hlt);
 270
 271 static inline int hlt_use_halt(void)
 272 {
 273         return (!hlt_counter && boot_cpu_data.hlt_works_ok);
 274 }
 275 #else
 276 static inline int hlt_use_halt(void)
 277 {
 278         return 1;
 279 }
 280 #endif
 281
 282 /*
 283  * We use this if we don't have any better
 284  * idle routine..
 285  */
 286 void default_idle(void)
 287 {
 288         if (hlt_use_halt()) {
 289                 trace_power_start(POWER_CSTATE, 1);
 290                 current_thread_info()->status &= ~TS_POLLING;
 291                 /*
 292                  * TS_POLLING-cleared state must be visible before we
 293                  * test NEED_RESCHED:
 294                  */
 295                 smp_mb();
 296
 297                 if (!need_resched())
 298                         safe_halt();    /* enables interrupts racelessly */
 299                 else
 300                         local_irq_enable();
 301                 current_thread_info()->status |= TS_POLLING;
 302         } else {
 303                 local_irq_enable();
 304                 /* loop is done by the caller */
 305                 cpu_relax();
 306         }
 307 }
 308 #ifdef CONFIG_APM_MODULE
 309 EXPORT_SYMBOL(default_idle);
 310 #endif
 311
 312 void stop_this_cpu(void *dummy)
 313 {
 314         local_irq_disable();
 315         /*
 316          * Remove this CPU:
 317          */
 318         set_cpu_online(smp_processor_id(), false);
 319         disable_local_APIC();
 320
 321         for (;;) {
 322                 if (hlt_works(smp_processor_id()))
 323                         halt();
 324         }
 325 }
 326
 327 static void do_nothing(void *unused)
 328 {
 329 }
 330
 331 /*
 332  * cpu_idle_wait - Used to ensure that all the CPUs discard old value of
 333  * pm_idle and update to new pm_idle value. Required while changing pm_idle
 334  * handler on SMP systems.
 335  *
 336  * Caller must have changed pm_idle to the new value before the call. Old
 337  * pm_idle value will not be used by any CPU after the return of this function.
 338  */
 339 void cpu_idle_wait(void)
 340 {
 341         smp_mb();
 342         /* kick all the CPUs so that they exit out of pm_idle */
 343         smp_call_function(do_nothing, NULL, 1);
 344 }
 345 EXPORT_SYMBOL_GPL(cpu_idle_wait);
 346
 347 /*
 348  * This uses new MONITOR/MWAIT instructions on P4 processors with PNI,
 349  * which can obviate IPI to trigger checking of need_resched.
 350  * We execute MONITOR against need_resched and enter optimized wait state
 351  * through MWAIT. Whenever someone changes need_resched, we would be woken
 352  * up from MWAIT (without an IPI).
 353  *
 354  * New with Core Duo processors, MWAIT can take some hints based on CPU
 355  * capability.
 356  */
 357 void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
 358 {
 359         trace_power_start(POWER_CSTATE, (ax>>4)+1);
 360         if (!need_resched()) {
 361                 if (cpu_has(&current_cpu_data, X86_FEATURE_CLFLUSH_MONITOR))
 362                         clflush((void *)&current_thread_info()->flags);
 363
 364                 __monitor((void *)&current_thread_info()->flags, 0, 0);
 365                 smp_mb();
 366                 if (!need_resched())
 367                         __mwait(ax, cx);
 368         }
 369 }
 370
 371 /* Default MONITOR/MWAIT with no hints, used for default C1 state */
 372 static void mwait_idle(void)
 373 {
 374         if (!need_resched()) {
 375                 trace_power_start(POWER_CSTATE, 1);
 376                 if (cpu_has(&current_cpu_data, X86_FEATURE_CLFLUSH_MONITOR))
 377                         clflush((void *)&current_thread_info()->flags);
 378
 379                 __monitor((void *)&current_thread_info()->flags, 0, 0);
 380                 smp_mb();
 381                 if (!need_resched())
 382                         __sti_mwait(0, 0);
 383                 else
 384                         local_irq_enable();
 385         } else
 386                 local_irq_enable();
 387 }
 388
 389 /*
 390  * On SMP it's slightly faster (but much more power-consuming!)
 391  * to poll the ->work.need_resched flag instead of waiting for the
 392  * cross-CPU IPI to arrive. Use this option with caution.
 393  */
 394 static void poll_idle(void)
 395 {
 396         trace_power_start(POWER_CSTATE, 0);
 397         local_irq_enable();
 398         while (!need_resched())
 399                 cpu_relax();
 400         trace_power_end(0);
 401 }
 402
 403 /*
 404  * mwait selection logic:
 405  *
 406  * It depends on the CPU. For AMD CPUs that support MWAIT this is
 407  * wrong. Family 0x10 and 0x11 CPUs will enter C1 on HLT. Powersavings
 408  * then depend on a clock divisor and current Pstate of the core. If
 409  * all cores of a processor are in halt state (C1) the processor can
 410  * enter the C1E (C1 enhanced) state. If mwait is used this will never
 411  * happen.
 412  *
 413  * idle=mwait overrides this decision and forces the usage of mwait.
 414  */
 415 static int __cpuinitdata force_mwait;
 416
 417 #define MWAIT_INFO                      0x05
 418 #define MWAIT_ECX_EXTENDED_INFO         0x01
 419 #define MWAIT_EDX_C1                    0xf0
 420
 421 static int __cpuinit mwait_usable(const struct cpuinfo_x86 *c)
 422 {
 423         u32 eax, ebx, ecx, edx;
 424
 425         if (force_mwait)
 426                 return 1;
 427
 428         if (c->cpuid_level < MWAIT_INFO)
 429                 return 0;
 430
 431         cpuid(MWAIT_INFO, &eax, &ebx, &ecx, &edx);
 432         /* Check, whether EDX has extended info about MWAIT */
 433         if (!(ecx & MWAIT_ECX_EXTENDED_INFO))
 434                 return 1;
 435
 436         /*
 437          * edx enumeratios MONITOR/MWAIT extensions. Check, whether
 438          * C1  supports MWAIT
 439          */
 440         return (edx & MWAIT_EDX_C1);
 441 }
 442
 443 /*
 444  * Check for AMD CPUs, which have potentially C1E support
 445  */
 446 static int __cpuinit check_c1e_idle(const struct cpuinfo_x86 *c)
 447 {
 448         if (c->x86_vendor != X86_VENDOR_AMD)
 449                 return 0;
 450
 451         if (c->x86 < 0x0F)
 452                 return 0;
 453
 454         /* Family 0x0f models < rev F do not have C1E */
 455         if (c->x86 == 0x0f && c->x86_model < 0x40)
 456                 return 0;
 457
 458         return 1;
 459 }
 460
 461 static cpumask_var_t c1e_mask;
 462 static int c1e_detected;
 463
 464 void c1e_remove_cpu(int cpu)
 465 {
 466         if (c1e_mask != NULL)
 467                 cpumask_clear_cpu(cpu, c1e_mask);
 468 }
 469
 470 /*
 471  * C1E aware idle routine. We check for C1E active in the interrupt
 472  * pending message MSR. If we detect C1E, then we handle it the same
 473  * way as C3 power states (local apic timer and TSC stop)
 474  */
 475 static void c1e_idle(void)
 476 {
 477         if (need_resched())
 478                 return;
 479
 480         if (!c1e_detected) {
 481                 u32 lo, hi;
 482
 483                 rdmsr(MSR_K8_INT_PENDING_MSG, lo, hi);
 484                 if (lo & K8_INTP_C1E_ACTIVE_MASK) {
 485                         c1e_detected = 1;
 486                         if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 487                                 mark_tsc_unstable("TSC halt in AMD C1E");
 488                         printk(KERN_INFO "System has AMD C1E enabled\n");
 489                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_AMDC1E);
 490                 }
 491         }
 492
 493         if (c1e_detected) {
 494                 int cpu = smp_processor_id();
 495
 496                 if (!cpumask_test_cpu(cpu, c1e_mask)) {
 497                         cpumask_set_cpu(cpu, c1e_mask);
 498                         /*
 499                          * Force broadcast so ACPI can not interfere.
 500                          */
 501                         clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_FORCE,
 502                                            &cpu);
 503                         printk(KERN_INFO "Switch to broadcast mode on CPU%d\n",
 504                                cpu);
 505                 }
 506                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 507
 508                 default_idle();
 509
 510                 /*
 511                  * The switch back from broadcast mode needs to be
 512                  * called with interrupts disabled.
 513                  */
 514                  local_irq_disable();
 515                  clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 516                  local_irq_enable();
 517         } else
 518                 default_idle();
 519 }
 520
 521 void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
 522 {
 523 #ifdef CONFIG_SMP
 524         if (pm_idle == poll_idle && smp_num_siblings > 1) {
 525                 printk(KERN_WARNING "WARNING: polling idle and HT enabled,"
 526                         " performance may degrade.\n");
 527         }
 528 #endif
 529         if (pm_idle)
 530                 return;
 531
 532         if (cpu_has(c, X86_FEATURE_MWAIT) && mwait_usable(c)) {
 533                 /*
 534                  * One CPU supports mwait => All CPUs supports mwait
 535                  */
 536                 printk(KERN_INFO "using mwait in idle threads.\n");
 537                 pm_idle = mwait_idle;
 538         } else if (check_c1e_idle(c)) {
 539                 printk(KERN_INFO "using C1E aware idle routine\n");
 540                 pm_idle = c1e_idle;
 541         } else
 542                 pm_idle = default_idle;
 543 }
 544
 545 void __init init_c1e_mask(void)
 546 {
 547         /* If we're using c1e_idle, we need to allocate c1e_mask. */
 548         if (pm_idle == c1e_idle)
 549                 zalloc_cpumask_var(&c1e_mask, GFP_KERNEL);
 550 }
 551
 552 static int __init idle_setup(char *str)
 553 {
 554         if (!str)
 555                 return -EINVAL;
 556
 557         if (!strcmp(str, "poll")) {
 558                 printk("using polling idle threads.\n");
 559                 pm_idle = poll_idle;
 560         } else if (!strcmp(str, "mwait"))
 561                 force_mwait = 1;
 562         else if (!strcmp(str, "halt")) {
 563                 /*
 564                  * When the boot option of idle=halt is added, halt is
 565                  * forced to be used for CPU idle. In such case CPU C2/C3
 566                  * won't be used again.
 567                  * To continue to load the CPU idle driver, don't touch
 568                  * the boot_option_idle_override.
 569                  */
 570                 pm_idle = default_idle;
 571                 idle_halt = 1;
 572                 return 0;
 573         } else if (!strcmp(str, "nomwait")) {
 574                 /*
 575                  * If the boot option of "idle=nomwait" is added,
 576                  * it means that mwait will be disabled for CPU C2/C3
 577                  * states. In such case it won't touch the variable
 578                  * of boot_option_idle_override.
 579                  */
 580                 idle_nomwait = 1;
 581                 return 0;
 582         } else
 583                 return -1;
 584
 585         boot_option_idle_override = 1;
 586         return 0;
 587 }
 588 early_param("idle", idle_setup);
 589
 590 unsigned long arch_align_stack(unsigned long sp)
 591 {
 592         if (!(current->personality & ADDR_NO_RANDOMIZE) && randomize_va_space)
 593                 sp -= get_random_int() % 8192;
 594         return sp & ~0xf;
 595 }
 596
 597 unsigned long arch_randomize_brk(struct mm_struct *mm)
 598 {
 599         unsigned long range_end = mm->brk + 0x02000000;
 600         return randomize_range(mm->brk, range_end, 0) ? : mm->brk;
 601 }
 602