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 <trace/power.h>
  12 #include <asm/system.h>
  13 #include <asm/apic.h>
  14 #include <asm/idle.h>
  15 #include <asm/uaccess.h>
  16 #include <asm/i387.h>
  17
  18 unsigned long idle_halt;
  19 EXPORT_SYMBOL(idle_halt);
  20 unsigned long idle_nomwait;
  21 EXPORT_SYMBOL(idle_nomwait);
  22
  23 struct kmem_cache *task_xstate_cachep;
  24
  25 DEFINE_TRACE(power_start);
  26 DEFINE_TRACE(power_end);
  27
  28 int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
  29 {
  30         *dst = *src;
  31         if (src->thread.xstate) {
  32                 dst->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
  33                                                       GFP_KERNEL);
  34                 if (!dst->thread.xstate)
  35                         return -ENOMEM;
  36                 WARN_ON((unsigned long)dst->thread.xstate & 15);
  37                 memcpy(dst->thread.xstate, src->thread.xstate, xstate_size);
  38         }
  39         return 0;
  40 }
  41
  42 void free_thread_xstate(struct task_struct *tsk)
  43 {
  44         if (tsk->thread.xstate) {
  45                 kmem_cache_free(task_xstate_cachep, tsk->thread.xstate);
  46                 tsk->thread.xstate = NULL;
  47         }
  48 }
  49
  50 void free_thread_info(struct thread_info *ti)
  51 {
  52         free_thread_xstate(ti->task);
  53         free_pages((unsigned long)ti, get_order(THREAD_SIZE));
  54 }
  55
  56 void arch_task_cache_init(void)
  57 {
  58         task_xstate_cachep =
  59                 kmem_cache_create("task_xstate", xstate_size,
  60                                   __alignof__(union thread_xstate),
  61                                   SLAB_PANIC, NULL);
  62 }
  63
  64 /*
  65  * Free current thread data structures etc..
  66  */
  67 void exit_thread(void)
  68 {
  69         struct task_struct *me = current;
  70         struct thread_struct *t = &me->thread;
  71         unsigned long *bp = t->io_bitmap_ptr;
  72
  73         if (bp) {
  74                 struct tss_struct *tss = &per_cpu(init_tss, get_cpu());
  75
  76                 t->io_bitmap_ptr = NULL;
  77                 clear_thread_flag(TIF_IO_BITMAP);
  78                 /*
  79                  * Careful, clear this in the TSS too:
  80                  */
  81                 memset(tss->io_bitmap, 0xff, t->io_bitmap_max);
  82                 t->io_bitmap_max = 0;
  83                 put_cpu();
  84                 kfree(bp);
  85         }
  86
  87         ds_exit_thread(current);
  88 }
  89
  90 void flush_thread(void)
  91 {
  92         struct task_struct *tsk = current;
  93
  94 #ifdef CONFIG_X86_64
  95         if (test_tsk_thread_flag(tsk, TIF_ABI_PENDING)) {
  96                 clear_tsk_thread_flag(tsk, TIF_ABI_PENDING);
  97                 if (test_tsk_thread_flag(tsk, TIF_IA32)) {
  98                         clear_tsk_thread_flag(tsk, TIF_IA32);
  99                 } else {
 100                         set_tsk_thread_flag(tsk, TIF_IA32);
 101                         current_thread_info()->status |= TS_COMPAT;
 102                 }
 103         }
 104 #endif
 105
 106         clear_tsk_thread_flag(tsk, TIF_DEBUG);
 107
 108         tsk->thread.debugreg0 = 0;
 109         tsk->thread.debugreg1 = 0;
 110         tsk->thread.debugreg2 = 0;
 111         tsk->thread.debugreg3 = 0;
 112         tsk->thread.debugreg6 = 0;
 113         tsk->thread.debugreg7 = 0;
 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(next_p, TIF_DEBUG)) {
 196                 set_debugreg(next->debugreg0, 0);
 197                 set_debugreg(next->debugreg1, 1);
 198                 set_debugreg(next->debugreg2, 2);
 199                 set_debugreg(next->debugreg3, 3);
 200                 /* no 4 and 5 */
 201                 set_debugreg(next->debugreg6, 6);
 202                 set_debugreg(next->debugreg7, 7);
 203         }
 204
 205         if (test_tsk_thread_flag(prev_p, TIF_NOTSC) ^
 206             test_tsk_thread_flag(next_p, TIF_NOTSC)) {
 207                 /* prev and next are different */
 208                 if (test_tsk_thread_flag(next_p, TIF_NOTSC))
 209                         hard_disable_TSC();
 210                 else
 211                         hard_enable_TSC();
 212         }
 213
 214         if (test_tsk_thread_flag(next_p, TIF_IO_BITMAP)) {
 215                 /*
 216                  * Copy the relevant range of the IO bitmap.
 217                  * Normally this is 128 bytes or less:
 218                  */
 219                 memcpy(tss->io_bitmap, next->io_bitmap_ptr,
 220                        max(prev->io_bitmap_max, next->io_bitmap_max));
 221         } else if (test_tsk_thread_flag(prev_p, TIF_IO_BITMAP)) {
 222                 /*
 223                  * Clear any possible leftover bits:
 224                  */
 225                 memset(tss->io_bitmap, 0xff, prev->io_bitmap_max);
 226         }
 227 }
 228
 229 int sys_fork(struct pt_regs *regs)
 230 {
 231         return do_fork(SIGCHLD, regs->sp, regs, 0, NULL, NULL);
 232 }
 233
 234 /*
 235  * This is trivial, and on the face of it looks like it
 236  * could equally well be done in user mode.
 237  *
 238  * Not so, for quite unobvious reasons - register pressure.
 239  * In user mode vfork() cannot have a stack frame, and if
 240  * done by calling the "clone()" system call directly, you
 241  * do not have enough call-clobbered registers to hold all
 242  * the information you need.
 243  */
 244 int sys_vfork(struct pt_regs *regs)
 245 {
 246         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->sp, regs, 0,
 247                        NULL, NULL);
 248 }
 249
 250
 251 /*
 252  * Idle related variables and functions
 253  */
 254 unsigned long boot_option_idle_override = 0;
 255 EXPORT_SYMBOL(boot_option_idle_override);
 256
 257 /*
 258  * Powermanagement idle function, if any..
 259  */
 260 void (*pm_idle)(void);
 261 EXPORT_SYMBOL(pm_idle);
 262
 263 #ifdef CONFIG_X86_32
 264 /*
 265  * This halt magic was a workaround for ancient floppy DMA
 266  * wreckage. It should be safe to remove.
 267  */
 268 static int hlt_counter;
 269 void disable_hlt(void)
 270 {
 271         hlt_counter++;
 272 }
 273 EXPORT_SYMBOL(disable_hlt);
 274
 275 void enable_hlt(void)
 276 {
 277         hlt_counter--;
 278 }
 279 EXPORT_SYMBOL(enable_hlt);
 280
 281 static inline int hlt_use_halt(void)
 282 {
 283         return (!hlt_counter && boot_cpu_data.hlt_works_ok);
 284 }
 285 #else
 286 static inline int hlt_use_halt(void)
 287 {
 288         return 1;
 289 }
 290 #endif
 291
 292 /*
 293  * We use this if we don't have any better
 294  * idle routine..
 295  */
 296 void default_idle(void)
 297 {
 298         if (hlt_use_halt()) {
 299                 struct power_trace it;
 300
 301                 trace_power_start(&it, POWER_CSTATE, 1);
 302                 current_thread_info()->status &= ~TS_POLLING;
 303                 /*
 304                  * TS_POLLING-cleared state must be visible before we
 305                  * test NEED_RESCHED:
 306                  */
 307                 smp_mb();
 308
 309                 if (!need_resched())
 310                         safe_halt();    /* enables interrupts racelessly */
 311                 else
 312                         local_irq_enable();
 313                 current_thread_info()->status |= TS_POLLING;
 314                 trace_power_end(&it);
 315         } else {
 316                 local_irq_enable();
 317                 /* loop is done by the caller */
 318                 cpu_relax();
 319         }
 320 }
 321 #ifdef CONFIG_APM_MODULE
 322 EXPORT_SYMBOL(default_idle);
 323 #endif
 324
 325 void stop_this_cpu(void *dummy)
 326 {
 327         local_irq_disable();
 328         /*
 329          * Remove this CPU:
 330          */
 331         cpu_clear(smp_processor_id(), cpu_online_map);
 332         disable_local_APIC();
 333
 334         for (;;) {
 335                 if (hlt_works(smp_processor_id()))
 336                         halt();
 337         }
 338 }
 339
 340 static void do_nothing(void *unused)
 341 {
 342 }
 343
 344 /*
 345  * cpu_idle_wait - Used to ensure that all the CPUs discard old value of
 346  * pm_idle and update to new pm_idle value. Required while changing pm_idle
 347  * handler on SMP systems.
 348  *
 349  * Caller must have changed pm_idle to the new value before the call. Old
 350  * pm_idle value will not be used by any CPU after the return of this function.
 351  */
 352 void cpu_idle_wait(void)
 353 {
 354         smp_mb();
 355         /* kick all the CPUs so that they exit out of pm_idle */
 356         smp_call_function(do_nothing, NULL, 1);
 357 }
 358 EXPORT_SYMBOL_GPL(cpu_idle_wait);
 359
 360 /*
 361  * This uses new MONITOR/MWAIT instructions on P4 processors with PNI,
 362  * which can obviate IPI to trigger checking of need_resched.
 363  * We execute MONITOR against need_resched and enter optimized wait state
 364  * through MWAIT. Whenever someone changes need_resched, we would be woken
 365  * up from MWAIT (without an IPI).
 366  *
 367  * New with Core Duo processors, MWAIT can take some hints based on CPU
 368  * capability.
 369  */
 370 void mwait_idle_with_hints(unsigned long ax, unsigned long cx)
 371 {
 372         struct power_trace it;
 373
 374         trace_power_start(&it, POWER_CSTATE, (ax>>4)+1);
 375         if (!need_resched()) {
 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                         __mwait(ax, cx);
 383         }
 384         trace_power_end(&it);
 385 }
 386
 387 /* Default MONITOR/MWAIT with no hints, used for default C1 state */
 388 static void mwait_idle(void)
 389 {
 390         struct power_trace it;
 391         if (!need_resched()) {
 392                 trace_power_start(&it, POWER_CSTATE, 1);
 393                 if (cpu_has(&current_cpu_data, X86_FEATURE_CLFLUSH_MONITOR))
 394                         clflush((void *)&current_thread_info()->flags);
 395
 396                 __monitor((void *)&current_thread_info()->flags, 0, 0);
 397                 smp_mb();
 398                 if (!need_resched())
 399                         __sti_mwait(0, 0);
 400                 else
 401                         local_irq_enable();
 402                 trace_power_end(&it);
 403         } else
 404                 local_irq_enable();
 405 }
 406
 407 /*
 408  * On SMP it's slightly faster (but much more power-consuming!)
 409  * to poll the ->work.need_resched flag instead of waiting for the
 410  * cross-CPU IPI to arrive. Use this option with caution.
 411  */
 412 static void poll_idle(void)
 413 {
 414         struct power_trace it;
 415
 416         trace_power_start(&it, POWER_CSTATE, 0);
 417         local_irq_enable();
 418         while (!need_resched())
 419                 cpu_relax();
 420         trace_power_end(&it);
 421 }
 422
 423 /*
 424  * mwait selection logic:
 425  *
 426  * It depends on the CPU. For AMD CPUs that support MWAIT this is
 427  * wrong. Family 0x10 and 0x11 CPUs will enter C1 on HLT. Powersavings
 428  * then depend on a clock divisor and current Pstate of the core. If
 429  * all cores of a processor are in halt state (C1) the processor can
 430  * enter the C1E (C1 enhanced) state. If mwait is used this will never
 431  * happen.
 432  *
 433  * idle=mwait overrides this decision and forces the usage of mwait.
 434  */
 435 static int __cpuinitdata force_mwait;
 436
 437 #define MWAIT_INFO                      0x05
 438 #define MWAIT_ECX_EXTENDED_INFO         0x01
 439 #define MWAIT_EDX_C1                    0xf0
 440
 441 static int __cpuinit mwait_usable(const struct cpuinfo_x86 *c)
 442 {
 443         u32 eax, ebx, ecx, edx;
 444
 445         if (force_mwait)
 446                 return 1;
 447
 448         if (c->cpuid_level < MWAIT_INFO)
 449                 return 0;
 450
 451         cpuid(MWAIT_INFO, &eax, &ebx, &ecx, &edx);
 452         /* Check, whether EDX has extended info about MWAIT */
 453         if (!(ecx & MWAIT_ECX_EXTENDED_INFO))
 454                 return 1;
 455
 456         /*
 457          * edx enumeratios MONITOR/MWAIT extensions. Check, whether
 458          * C1  supports MWAIT
 459          */
 460         return (edx & MWAIT_EDX_C1);
 461 }
 462
 463 /*
 464  * Check for AMD CPUs, which have potentially C1E support
 465  */
 466 static int __cpuinit check_c1e_idle(const struct cpuinfo_x86 *c)
 467 {
 468         if (c->x86_vendor != X86_VENDOR_AMD)
 469                 return 0;
 470
 471         if (c->x86 < 0x0F)
 472                 return 0;
 473
 474         /* Family 0x0f models < rev F do not have C1E */
 475         if (c->x86 == 0x0f && c->x86_model < 0x40)
 476                 return 0;
 477
 478         return 1;
 479 }
 480
 481 static cpumask_t c1e_mask = CPU_MASK_NONE;
 482 static int c1e_detected;
 483
 484 void c1e_remove_cpu(int cpu)
 485 {
 486         cpu_clear(cpu, c1e_mask);
 487 }
 488
 489 /*
 490  * C1E aware idle routine. We check for C1E active in the interrupt
 491  * pending message MSR. If we detect C1E, then we handle it the same
 492  * way as C3 power states (local apic timer and TSC stop)
 493  */
 494 static void c1e_idle(void)
 495 {
 496         if (need_resched())
 497                 return;
 498
 499         if (!c1e_detected) {
 500                 u32 lo, hi;
 501
 502                 rdmsr(MSR_K8_INT_PENDING_MSG, lo, hi);
 503                 if (lo & K8_INTP_C1E_ACTIVE_MASK) {
 504                         c1e_detected = 1;
 505                         if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 506                                 mark_tsc_unstable("TSC halt in AMD C1E");
 507                         printk(KERN_INFO "System has AMD C1E enabled\n");
 508                         set_cpu_cap(&boot_cpu_data, X86_FEATURE_AMDC1E);
 509                 }
 510         }
 511
 512         if (c1e_detected) {
 513                 int cpu = smp_processor_id();
 514
 515                 if (!cpu_isset(cpu, c1e_mask)) {
 516                         cpu_set(cpu, c1e_mask);
 517                         /*
 518                          * Force broadcast so ACPI can not interfere. Needs
 519                          * to run with interrupts enabled as it uses
 520                          * smp_function_call.
 521                          */
 522                         local_irq_enable();
 523                         clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_FORCE,
 524                                            &cpu);
 525                         printk(KERN_INFO "Switch to broadcast mode on CPU%d\n",
 526                                cpu);
 527                         local_irq_disable();
 528                 }
 529                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 530
 531                 default_idle();
 532
 533                 /*
 534                  * The switch back from broadcast mode needs to be
 535                  * called with interrupts disabled.
 536                  */
 537                  local_irq_disable();
 538                  clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 539                  local_irq_enable();
 540         } else
 541                 default_idle();
 542 }
 543
 544 void __cpuinit select_idle_routine(const struct cpuinfo_x86 *c)
 545 {
 546 #ifdef CONFIG_SMP
 547         if (pm_idle == poll_idle && smp_num_siblings > 1) {
 548                 printk(KERN_WARNING "WARNING: polling idle and HT enabled,"
 549                         " performance may degrade.\n");
 550         }
 551 #endif
 552         if (pm_idle)
 553                 return;
 554
 555         if (cpu_has(c, X86_FEATURE_MWAIT) && mwait_usable(c)) {
 556                 /*
 557                  * One CPU supports mwait => All CPUs supports mwait
 558                  */
 559                 printk(KERN_INFO "using mwait in idle threads.\n");
 560                 pm_idle = mwait_idle;
 561         } else if (check_c1e_idle(c)) {
 562                 printk(KERN_INFO "using C1E aware idle routine\n");
 563                 pm_idle = c1e_idle;
 564         } else
 565                 pm_idle = default_idle;
 566 }
 567
 568 static int __init idle_setup(char *str)
 569 {
 570         if (!str)
 571                 return -EINVAL;
 572
 573         if (!strcmp(str, "poll")) {
 574                 printk("using polling idle threads.\n");
 575                 pm_idle = poll_idle;
 576         } else if (!strcmp(str, "mwait"))
 577                 force_mwait = 1;
 578         else if (!strcmp(str, "halt")) {
 579                 /*
 580                  * When the boot option of idle=halt is added, halt is
 581                  * forced to be used for CPU idle. In such case CPU C2/C3
 582                  * won't be used again.
 583                  * To continue to load the CPU idle driver, don't touch
 584                  * the boot_option_idle_override.
 585                  */
 586                 pm_idle = default_idle;
 587                 idle_halt = 1;
 588                 return 0;
 589         } else if (!strcmp(str, "nomwait")) {
 590                 /*
 591                  * If the boot option of "idle=nomwait" is added,
 592                  * it means that mwait will be disabled for CPU C2/C3
 593                  * states. In such case it won't touch the variable
 594                  * of boot_option_idle_override.
 595                  */
 596                 idle_nomwait = 1;
 597                 return 0;
 598         } else
 599                 return -1;
 600
 601         boot_option_idle_override = 1;
 602         return 0;
 603 }
 604 early_param("idle", idle_setup);
 605