arch/powerpc/kernel/process.c

   1 /*
   2  *  Derived from "arch/i386/kernel/process.c"
   3  *    Copyright (C) 1995  Linus Torvalds
   4  *
   5  *  Updated and modified by Cort Dougan (cort@cs.nmt.edu) and
   6  *  Paul Mackerras (paulus@cs.anu.edu.au)
   7  *
   8  *  PowerPC version
   9  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
  10  *
  11  *  This program is free software; you can redistribute it and/or
  12  *  modify it under the terms of the GNU General Public License
  13  *  as published by the Free Software Foundation; either version
  14  *  2 of the License, or (at your option) any later version.
  15  */
  16
  17 #include <linux/errno.h>
  18 #include <linux/sched.h>
  19 #include <linux/kernel.h>
  20 #include <linux/mm.h>
  21 #include <linux/smp.h>
  22 #include <linux/smp_lock.h>
  23 #include <linux/stddef.h>
  24 #include <linux/unistd.h>
  25 #include <linux/ptrace.h>
  26 #include <linux/slab.h>
  27 #include <linux/user.h>
  28 #include <linux/elf.h>
  29 #include <linux/init.h>
  30 #include <linux/prctl.h>
  31 #include <linux/init_task.h>
  32 #include <linux/module.h>
  33 #include <linux/kallsyms.h>
  34 #include <linux/mqueue.h>
  35 #include <linux/hardirq.h>
  36 #include <linux/utsname.h>
  37
  38 #include <asm/pgtable.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/system.h>
  41 #include <asm/io.h>
  42 #include <asm/processor.h>
  43 #include <asm/mmu.h>
  44 #include <asm/prom.h>
  45 #include <asm/machdep.h>
  46 #include <asm/time.h>
  47 #include <asm/syscalls.h>
  48 #ifdef CONFIG_PPC64
  49 #include <asm/firmware.h>
  50 #endif
  51
  52 extern unsigned long _get_SP(void);
  53
  54 #ifndef CONFIG_SMP
  55 struct task_struct *last_task_used_math = NULL;
  56 struct task_struct *last_task_used_altivec = NULL;
  57 struct task_struct *last_task_used_spe = NULL;
  58 #endif
  59
  60 /*
  61  * Make sure the floating-point register state in the
  62  * the thread_struct is up to date for task tsk.
  63  */
  64 void flush_fp_to_thread(struct task_struct *tsk)
  65 {
  66         if (tsk->thread.regs) {
  67                 /*
  68                  * We need to disable preemption here because if we didn't,
  69                  * another process could get scheduled after the regs->msr
  70                  * test but before we have finished saving the FP registers
  71                  * to the thread_struct.  That process could take over the
  72                  * FPU, and then when we get scheduled again we would store
  73                  * bogus values for the remaining FP registers.
  74                  */
  75                 preempt_disable();
  76                 if (tsk->thread.regs->msr & MSR_FP) {
  77 #ifdef CONFIG_SMP
  78                         /*
  79                          * This should only ever be called for current or
  80                          * for a stopped child process.  Since we save away
  81                          * the FP register state on context switch on SMP,
  82                          * there is something wrong if a stopped child appears
  83                          * to still have its FP state in the CPU registers.
  84                          */
  85                         BUG_ON(tsk != current);
  86 #endif
  87                         giveup_fpu(current);
  88                 }
  89                 preempt_enable();
  90         }
  91 }
  92
  93 void enable_kernel_fp(void)
  94 {
  95         WARN_ON(preemptible());
  96
  97 #ifdef CONFIG_SMP
  98         if (current->thread.regs && (current->thread.regs->msr & MSR_FP))
  99                 giveup_fpu(current);
 100         else
 101                 giveup_fpu(NULL);       /* just enables FP for kernel */
 102 #else
 103         giveup_fpu(last_task_used_math);
 104 #endif /* CONFIG_SMP */
 105 }
 106 EXPORT_SYMBOL(enable_kernel_fp);
 107
 108 int dump_task_fpu(struct task_struct *tsk, elf_fpregset_t *fpregs)
 109 {
 110         if (!tsk->thread.regs)
 111                 return 0;
 112         flush_fp_to_thread(current);
 113
 114         memcpy(fpregs, &tsk->thread.fpr[0], sizeof(*fpregs));
 115
 116         return 1;
 117 }
 118
 119 #ifdef CONFIG_ALTIVEC
 120 void enable_kernel_altivec(void)
 121 {
 122         WARN_ON(preemptible());
 123
 124 #ifdef CONFIG_SMP
 125         if (current->thread.regs && (current->thread.regs->msr & MSR_VEC))
 126                 giveup_altivec(current);
 127         else
 128                 giveup_altivec(NULL);   /* just enable AltiVec for kernel - force */
 129 #else
 130         giveup_altivec(last_task_used_altivec);
 131 #endif /* CONFIG_SMP */
 132 }
 133 EXPORT_SYMBOL(enable_kernel_altivec);
 134
 135 /*
 136  * Make sure the VMX/Altivec register state in the
 137  * the thread_struct is up to date for task tsk.
 138  */
 139 void flush_altivec_to_thread(struct task_struct *tsk)
 140 {
 141         if (tsk->thread.regs) {
 142                 preempt_disable();
 143                 if (tsk->thread.regs->msr & MSR_VEC) {
 144 #ifdef CONFIG_SMP
 145                         BUG_ON(tsk != current);
 146 #endif
 147                         giveup_altivec(current);
 148                 }
 149                 preempt_enable();
 150         }
 151 }
 152
 153 int dump_task_altivec(struct pt_regs *regs, elf_vrregset_t *vrregs)
 154 {
 155         flush_altivec_to_thread(current);
 156         memcpy(vrregs, &current->thread.vr[0], sizeof(*vrregs));
 157         return 1;
 158 }
 159 #endif /* CONFIG_ALTIVEC */
 160
 161 #ifdef CONFIG_SPE
 162
 163 void enable_kernel_spe(void)
 164 {
 165         WARN_ON(preemptible());
 166
 167 #ifdef CONFIG_SMP
 168         if (current->thread.regs && (current->thread.regs->msr & MSR_SPE))
 169                 giveup_spe(current);
 170         else
 171                 giveup_spe(NULL);       /* just enable SPE for kernel - force */
 172 #else
 173         giveup_spe(last_task_used_spe);
 174 #endif /* __SMP __ */
 175 }
 176 EXPORT_SYMBOL(enable_kernel_spe);
 177
 178 void flush_spe_to_thread(struct task_struct *tsk)
 179 {
 180         if (tsk->thread.regs) {
 181                 preempt_disable();
 182                 if (tsk->thread.regs->msr & MSR_SPE) {
 183 #ifdef CONFIG_SMP
 184                         BUG_ON(tsk != current);
 185 #endif
 186                         giveup_spe(current);
 187                 }
 188                 preempt_enable();
 189         }
 190 }
 191
 192 int dump_spe(struct pt_regs *regs, elf_vrregset_t *evrregs)
 193 {
 194         flush_spe_to_thread(current);
 195         /* We copy u32 evr[32] + u64 acc + u32 spefscr -> 35 */
 196         memcpy(evrregs, &current->thread.evr[0], sizeof(u32) * 35);
 197         return 1;
 198 }
 199 #endif /* CONFIG_SPE */
 200
 201 #ifndef CONFIG_SMP
 202 /*
 203  * If we are doing lazy switching of CPU state (FP, altivec or SPE),
 204  * and the current task has some state, discard it.
 205  */
 206 void discard_lazy_cpu_state(void)
 207 {
 208         preempt_disable();
 209         if (last_task_used_math == current)
 210                 last_task_used_math = NULL;
 211 #ifdef CONFIG_ALTIVEC
 212         if (last_task_used_altivec == current)
 213                 last_task_used_altivec = NULL;
 214 #endif /* CONFIG_ALTIVEC */
 215 #ifdef CONFIG_SPE
 216         if (last_task_used_spe == current)
 217                 last_task_used_spe = NULL;
 218 #endif
 219         preempt_enable();
 220 }
 221 #endif /* CONFIG_SMP */
 222
 223 #ifdef CONFIG_PPC_MERGE         /* XXX for now */
 224 int set_dabr(unsigned long dabr)
 225 {
 226         if (ppc_md.set_dabr)
 227                 return ppc_md.set_dabr(dabr);
 228
 229         mtspr(SPRN_DABR, dabr);
 230         return 0;
 231 }
 232 #endif
 233
 234 #ifdef CONFIG_PPC64
 235 DEFINE_PER_CPU(struct cpu_usage, cpu_usage_array);
 236 static DEFINE_PER_CPU(unsigned long, current_dabr);
 237 #endif
 238
 239 struct task_struct *__switch_to(struct task_struct *prev,
 240         struct task_struct *new)
 241 {
 242         struct thread_struct *new_thread, *old_thread;
 243         unsigned long flags;
 244         struct task_struct *last;
 245
 246 #ifdef CONFIG_SMP
 247         /* avoid complexity of lazy save/restore of fpu
 248          * by just saving it every time we switch out if
 249          * this task used the fpu during the last quantum.
 250          *
 251          * If it tries to use the fpu again, it'll trap and
 252          * reload its fp regs.  So we don't have to do a restore
 253          * every switch, just a save.
 254          *  -- Cort
 255          */
 256         if (prev->thread.regs && (prev->thread.regs->msr & MSR_FP))
 257                 giveup_fpu(prev);
 258 #ifdef CONFIG_ALTIVEC
 259         /*
 260          * If the previous thread used altivec in the last quantum
 261          * (thus changing altivec regs) then save them.
 262          * We used to check the VRSAVE register but not all apps
 263          * set it, so we don't rely on it now (and in fact we need
 264          * to save & restore VSCR even if VRSAVE == 0).  -- paulus
 265          *
 266          * On SMP we always save/restore altivec regs just to avoid the
 267          * complexity of changing processors.
 268          *  -- Cort
 269          */
 270         if (prev->thread.regs && (prev->thread.regs->msr & MSR_VEC))
 271                 giveup_altivec(prev);
 272 #endif /* CONFIG_ALTIVEC */
 273 #ifdef CONFIG_SPE
 274         /*
 275          * If the previous thread used spe in the last quantum
 276          * (thus changing spe regs) then save them.
 277          *
 278          * On SMP we always save/restore spe regs just to avoid the
 279          * complexity of changing processors.
 280          */
 281         if ((prev->thread.regs && (prev->thread.regs->msr & MSR_SPE)))
 282                 giveup_spe(prev);
 283 #endif /* CONFIG_SPE */
 284
 285 #else  /* CONFIG_SMP */
 286 #ifdef CONFIG_ALTIVEC
 287         /* Avoid the trap.  On smp this this never happens since
 288          * we don't set last_task_used_altivec -- Cort
 289          */
 290         if (new->thread.regs && last_task_used_altivec == new)
 291                 new->thread.regs->msr |= MSR_VEC;
 292 #endif /* CONFIG_ALTIVEC */
 293 #ifdef CONFIG_SPE
 294         /* Avoid the trap.  On smp this this never happens since
 295          * we don't set last_task_used_spe
 296          */
 297         if (new->thread.regs && last_task_used_spe == new)
 298                 new->thread.regs->msr |= MSR_SPE;
 299 #endif /* CONFIG_SPE */
 300
 301 #endif /* CONFIG_SMP */
 302
 303 #ifdef CONFIG_PPC64     /* for now */
 304         if (unlikely(__get_cpu_var(current_dabr) != new->thread.dabr)) {
 305                 set_dabr(new->thread.dabr);
 306                 __get_cpu_var(current_dabr) = new->thread.dabr;
 307         }
 308
 309         flush_tlb_pending();
 310 #endif
 311
 312         new_thread = &new->thread;
 313         old_thread = &current->thread;
 314
 315 #ifdef CONFIG_PPC64
 316         /*
 317          * Collect processor utilization data per process
 318          */
 319         if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
 320                 struct cpu_usage *cu = &__get_cpu_var(cpu_usage_array);
 321                 long unsigned start_tb, current_tb;
 322                 start_tb = old_thread->start_tb;
 323                 cu->current_tb = current_tb = mfspr(SPRN_PURR);
 324                 old_thread->accum_tb += (current_tb - start_tb);
 325                 new_thread->start_tb = current_tb;
 326         }
 327 #endif
 328
 329         local_irq_save(flags);
 330
 331         account_system_vtime(current);
 332         account_process_vtime(current);
 333         calculate_steal_time();
 334
 335         last = _switch(old_thread, new_thread);
 336
 337         local_irq_restore(flags);
 338
 339         return last;
 340 }
 341
 342 static int instructions_to_print = 16;
 343
 344 #ifdef CONFIG_PPC64
 345 #define BAD_PC(pc)      ((REGION_ID(pc) != KERNEL_REGION_ID) && \
 346                          (REGION_ID(pc) != VMALLOC_REGION_ID))
 347 #else
 348 #define BAD_PC(pc)      ((pc) < KERNELBASE)
 349 #endif
 350
 351 static void show_instructions(struct pt_regs *regs)
 352 {
 353         int i;
 354         unsigned long pc = regs->nip - (instructions_to_print * 3 / 4 *
 355                         sizeof(int));
 356
 357         printk("Instruction dump:");
 358
 359         for (i = 0; i < instructions_to_print; i++) {
 360                 int instr;
 361
 362                 if (!(i % 8))
 363                         printk("\n");
 364
 365                 /* We use __get_user here *only* to avoid an OOPS on a
 366                  * bad address because the pc *should* only be a
 367                  * kernel address.
 368                  */
 369                 if (BAD_PC(pc) || __get_user(instr, (unsigned int __user *)pc)) {
 370                         printk("XXXXXXXX ");
 371                 } else {
 372                         if (regs->nip == pc)
 373                                 printk("<%08x> ", instr);
 374                         else
 375                                 printk("%08x ", instr);
 376                 }
 377
 378                 pc += sizeof(int);
 379         }
 380
 381         printk("\n");
 382 }
 383
 384 static struct regbit {
 385         unsigned long bit;
 386         const char *name;
 387 } msr_bits[] = {
 388         {MSR_EE,        "EE"},
 389         {MSR_PR,        "PR"},
 390         {MSR_FP,        "FP"},
 391         {MSR_ME,        "ME"},
 392         {MSR_IR,        "IR"},
 393         {MSR_DR,        "DR"},
 394         {0,             NULL}
 395 };
 396
 397 static void printbits(unsigned long val, struct regbit *bits)
 398 {
 399         const char *sep = "";
 400
 401         printk("<");
 402         for (; bits->bit; ++bits)
 403                 if (val & bits->bit) {
 404                         printk("%s%s", sep, bits->name);
 405                         sep = ",";
 406                 }
 407         printk(">");
 408 }
 409
 410 #ifdef CONFIG_PPC64
 411 #define REG             "%016lX"
 412 #define REGS_PER_LINE   4
 413 #define LAST_VOLATILE   13
 414 #else
 415 #define REG             "%08lX"
 416 #define REGS_PER_LINE   8
 417 #define LAST_VOLATILE   12
 418 #endif
 419
 420 void show_regs(struct pt_regs * regs)
 421 {
 422         int i, trap;
 423
 424         printk("NIP: "REG" LR: "REG" CTR: "REG"\n",
 425                regs->nip, regs->link, regs->ctr);
 426         printk("REGS: %p TRAP: %04lx   %s  (%s)\n",
 427                regs, regs->trap, print_tainted(), system_utsname.release);
 428         printk("MSR: "REG" ", regs->msr);
 429         printbits(regs->msr, msr_bits);
 430         printk("  CR: %08lX  XER: %08lX\n", regs->ccr, regs->xer);
 431         trap = TRAP(regs);
 432         if (trap == 0x300 || trap == 0x600)
 433                 printk("DAR: "REG", DSISR: "REG"\n", regs->dar, regs->dsisr);
 434         printk("TASK = %p[%d,#%u] '%s' THREAD: %p",
 435                current, current->pid, current->xid,
 436                current->comm, task_thread_info(current));
 437
 438 #ifdef CONFIG_SMP
 439         printk(" CPU: %d", smp_processor_id());
 440 #endif /* CONFIG_SMP */
 441
 442         for (i = 0;  i < 32;  i++) {
 443                 if ((i % REGS_PER_LINE) == 0)
 444                         printk("\n" KERN_INFO "GPR%02d: ", i);
 445                 printk(REG " ", regs->gpr[i]);
 446                 if (i == LAST_VOLATILE && !FULL_REGS(regs))
 447                         break;
 448         }
 449         printk("\n");
 450 #ifdef CONFIG_KALLSYMS
 451         /*
 452          * Lookup NIP late so we have the best change of getting the
 453          * above info out without failing
 454          */
 455         printk("NIP ["REG"] ", regs->nip);
 456         print_symbol("%s\n", regs->nip);
 457         printk("LR ["REG"] ", regs->link);
 458         print_symbol("%s\n", regs->link);
 459 #endif
 460         show_stack(current, (unsigned long *) regs->gpr[1]);
 461         if (!user_mode(regs))
 462                 show_instructions(regs);
 463 }
 464
 465 void exit_thread(void)
 466 {
 467         discard_lazy_cpu_state();
 468 }
 469
 470 void flush_thread(void)
 471 {
 472 #ifdef CONFIG_PPC64
 473         struct thread_info *t = current_thread_info();
 474
 475         if (t->flags & _TIF_ABI_PENDING)
 476                 t->flags ^= (_TIF_ABI_PENDING | _TIF_32BIT);
 477 #endif
 478
 479         discard_lazy_cpu_state();
 480
 481 #ifdef CONFIG_PPC64     /* for now */
 482         if (current->thread.dabr) {
 483                 current->thread.dabr = 0;
 484                 set_dabr(0);
 485         }
 486 #endif
 487 }
 488
 489 void
 490 release_thread(struct task_struct *t)
 491 {
 492 }
 493
 494 /*
 495  * This gets called before we allocate a new thread and copy
 496  * the current task into it.
 497  */
 498 void prepare_to_copy(struct task_struct *tsk)
 499 {
 500         flush_fp_to_thread(current);
 501         flush_altivec_to_thread(current);
 502         flush_spe_to_thread(current);
 503 }
 504
 505 /*
 506  * Copy a thread..
 507  */
 508 int copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 509                 unsigned long unused, struct task_struct *p,
 510                 struct pt_regs *regs)
 511 {
 512         struct pt_regs *childregs, *kregs;
 513         extern void ret_from_fork(void);
 514         unsigned long sp = (unsigned long)task_stack_page(p) + THREAD_SIZE;
 515
 516         CHECK_FULL_REGS(regs);
 517         /* Copy registers */
 518         sp -= sizeof(struct pt_regs);
 519         childregs = (struct pt_regs *) sp;
 520         *childregs = *regs;
 521         if ((childregs->msr & MSR_PR) == 0) {
 522                 /* for kernel thread, set `current' and stackptr in new task */
 523                 childregs->gpr[1] = sp + sizeof(struct pt_regs);
 524 #ifdef CONFIG_PPC32
 525                 childregs->gpr[2] = (unsigned long) p;
 526 #else
 527                 clear_tsk_thread_flag(p, TIF_32BIT);
 528 #endif
 529                 p->thread.regs = NULL;  /* no user register state */
 530         } else {
 531                 childregs->gpr[1] = usp;
 532                 p->thread.regs = childregs;
 533                 if (clone_flags & CLONE_SETTLS) {
 534 #ifdef CONFIG_PPC64
 535                         if (!test_thread_flag(TIF_32BIT))
 536                                 childregs->gpr[13] = childregs->gpr[6];
 537                         else
 538 #endif
 539                                 childregs->gpr[2] = childregs->gpr[6];
 540                 }
 541         }
 542         childregs->gpr[3] = 0;  /* Result from fork() */
 543         sp -= STACK_FRAME_OVERHEAD;
 544
 545         /*
 546          * The way this works is that at some point in the future
 547          * some task will call _switch to switch to the new task.
 548          * That will pop off the stack frame created below and start
 549          * the new task running at ret_from_fork.  The new task will
 550          * do some house keeping and then return from the fork or clone
 551          * system call, using the stack frame created above.
 552          */
 553         sp -= sizeof(struct pt_regs);
 554         kregs = (struct pt_regs *) sp;
 555         sp -= STACK_FRAME_OVERHEAD;
 556         p->thread.ksp = sp;
 557
 558 #ifdef CONFIG_PPC64
 559         if (cpu_has_feature(CPU_FTR_SLB)) {
 560                 unsigned long sp_vsid = get_kernel_vsid(sp);
 561                 unsigned long llp = mmu_psize_defs[mmu_linear_psize].sllp;
 562
 563                 sp_vsid <<= SLB_VSID_SHIFT;
 564                 sp_vsid |= SLB_VSID_KERNEL | llp;
 565                 p->thread.ksp_vsid = sp_vsid;
 566         }
 567
 568         /*
 569          * The PPC64 ABI makes use of a TOC to contain function
 570          * pointers.  The function (ret_from_except) is actually a pointer
 571          * to the TOC entry.  The first entry is a pointer to the actual
 572          * function.
 573          */
 574         kregs->nip = *((unsigned long *)ret_from_fork);
 575 #else
 576         kregs->nip = (unsigned long)ret_from_fork;
 577         p->thread.last_syscall = -1;
 578 #endif
 579
 580         return 0;
 581 }
 582
 583 /*
 584  * Set up a thread for executing a new program
 585  */
 586 void start_thread(struct pt_regs *regs, unsigned long start, unsigned long sp)
 587 {
 588 #ifdef CONFIG_PPC64
 589         unsigned long load_addr = regs->gpr[2]; /* saved by ELF_PLAT_INIT */
 590 #endif
 591
 592         set_fs(USER_DS);
 593
 594         /*
 595          * If we exec out of a kernel thread then thread.regs will not be
 596          * set.  Do it now.
 597          */
 598         if (!current->thread.regs) {
 599                 struct pt_regs *regs = task_stack_page(current) + THREAD_SIZE;
 600                 current->thread.regs = regs - 1;
 601         }
 602
 603         memset(regs->gpr, 0, sizeof(regs->gpr));
 604         regs->ctr = 0;
 605         regs->link = 0;
 606         regs->xer = 0;
 607         regs->ccr = 0;
 608         regs->gpr[1] = sp;
 609
 610 #ifdef CONFIG_PPC32
 611         regs->mq = 0;
 612         regs->nip = start;
 613         regs->msr = MSR_USER;
 614 #else
 615         if (!test_thread_flag(TIF_32BIT)) {
 616                 unsigned long entry, toc;
 617
 618                 /* start is a relocated pointer to the function descriptor for
 619                  * the elf _start routine.  The first entry in the function
 620                  * descriptor is the entry address of _start and the second
 621                  * entry is the TOC value we need to use.
 622                  */
 623                 __get_user(entry, (unsigned long __user *)start);
 624                 __get_user(toc, (unsigned long __user *)start+1);
 625
 626                 /* Check whether the e_entry function descriptor entries
 627                  * need to be relocated before we can use them.
 628                  */
 629                 if (load_addr != 0) {
 630                         entry += load_addr;
 631                         toc   += load_addr;
 632                 }
 633                 regs->nip = entry;
 634                 regs->gpr[2] = toc;
 635                 regs->msr = MSR_USER64;
 636         } else {
 637                 regs->nip = start;
 638                 regs->gpr[2] = 0;
 639                 regs->msr = MSR_USER32;
 640         }
 641 #endif
 642
 643         discard_lazy_cpu_state();
 644         memset(current->thread.fpr, 0, sizeof(current->thread.fpr));
 645         current->thread.fpscr.val = 0;
 646 #ifdef CONFIG_ALTIVEC
 647         memset(current->thread.vr, 0, sizeof(current->thread.vr));
 648         memset(&current->thread.vscr, 0, sizeof(current->thread.vscr));
 649         current->thread.vscr.u[3] = 0x00010000; /* Java mode disabled */
 650         current->thread.vrsave = 0;
 651         current->thread.used_vr = 0;
 652 #endif /* CONFIG_ALTIVEC */
 653 #ifdef CONFIG_SPE
 654         memset(current->thread.evr, 0, sizeof(current->thread.evr));
 655         current->thread.acc = 0;
 656         current->thread.spefscr = 0;
 657         current->thread.used_spe = 0;
 658 #endif /* CONFIG_SPE */
 659 }
 660
 661 #define PR_FP_ALL_EXCEPT (PR_FP_EXC_DIV | PR_FP_EXC_OVF | PR_FP_EXC_UND \
 662                 | PR_FP_EXC_RES | PR_FP_EXC_INV)
 663
 664 int set_fpexc_mode(struct task_struct *tsk, unsigned int val)
 665 {
 666         struct pt_regs *regs = tsk->thread.regs;
 667
 668         /* This is a bit hairy.  If we are an SPE enabled  processor
 669          * (have embedded fp) we store the IEEE exception enable flags in
 670          * fpexc_mode.  fpexc_mode is also used for setting FP exception
 671          * mode (asyn, precise, disabled) for 'Classic' FP. */
 672         if (val & PR_FP_EXC_SW_ENABLE) {
 673 #ifdef CONFIG_SPE
 674                 tsk->thread.fpexc_mode = val &
 675                         (PR_FP_EXC_SW_ENABLE | PR_FP_ALL_EXCEPT);
 676                 return 0;
 677 #else
 678                 return -EINVAL;
 679 #endif
 680         }
 681
 682         /* on a CONFIG_SPE this does not hurt us.  The bits that
 683          * __pack_fe01 use do not overlap with bits used for
 684          * PR_FP_EXC_SW_ENABLE.  Additionally, the MSR[FE0,FE1] bits
 685          * on CONFIG_SPE implementations are reserved so writing to
 686          * them does not change anything */
 687         if (val > PR_FP_EXC_PRECISE)
 688                 return -EINVAL;
 689         tsk->thread.fpexc_mode = __pack_fe01(val);
 690         if (regs != NULL && (regs->msr & MSR_FP) != 0)
 691                 regs->msr = (regs->msr & ~(MSR_FE0|MSR_FE1))
 692                         | tsk->thread.fpexc_mode;
 693         return 0;
 694 }
 695
 696 int get_fpexc_mode(struct task_struct *tsk, unsigned long adr)
 697 {
 698         unsigned int val;
 699
 700         if (tsk->thread.fpexc_mode & PR_FP_EXC_SW_ENABLE)
 701 #ifdef CONFIG_SPE
 702                 val = tsk->thread.fpexc_mode;
 703 #else
 704                 return -EINVAL;
 705 #endif
 706         else
 707                 val = __unpack_fe01(tsk->thread.fpexc_mode);
 708         return put_user(val, (unsigned int __user *) adr);
 709 }
 710
 711 int set_endian(struct task_struct *tsk, unsigned int val)
 712 {
 713         struct pt_regs *regs = tsk->thread.regs;
 714
 715         if ((val == PR_ENDIAN_LITTLE && !cpu_has_feature(CPU_FTR_REAL_LE)) ||
 716             (val == PR_ENDIAN_PPC_LITTLE && !cpu_has_feature(CPU_FTR_PPC_LE)))
 717                 return -EINVAL;
 718
 719         if (regs == NULL)
 720                 return -EINVAL;
 721
 722         if (val == PR_ENDIAN_BIG)
 723                 regs->msr &= ~MSR_LE;
 724         else if (val == PR_ENDIAN_LITTLE || val == PR_ENDIAN_PPC_LITTLE)
 725                 regs->msr |= MSR_LE;
 726         else
 727                 return -EINVAL;
 728
 729         return 0;
 730 }
 731
 732 int get_endian(struct task_struct *tsk, unsigned long adr)
 733 {
 734         struct pt_regs *regs = tsk->thread.regs;
 735         unsigned int val;
 736
 737         if (!cpu_has_feature(CPU_FTR_PPC_LE) &&
 738             !cpu_has_feature(CPU_FTR_REAL_LE))
 739                 return -EINVAL;
 740
 741         if (regs == NULL)
 742                 return -EINVAL;
 743
 744         if (regs->msr & MSR_LE) {
 745                 if (cpu_has_feature(CPU_FTR_REAL_LE))
 746                         val = PR_ENDIAN_LITTLE;
 747                 else
 748                         val = PR_ENDIAN_PPC_LITTLE;
 749         } else
 750                 val = PR_ENDIAN_BIG;
 751
 752         return put_user(val, (unsigned int __user *)adr);
 753 }
 754
 755 int set_unalign_ctl(struct task_struct *tsk, unsigned int val)
 756 {
 757         tsk->thread.align_ctl = val;
 758         return 0;
 759 }
 760
 761 int get_unalign_ctl(struct task_struct *tsk, unsigned long adr)
 762 {
 763         return put_user(tsk->thread.align_ctl, (unsigned int __user *)adr);
 764 }
 765
 766 #define TRUNC_PTR(x)    ((typeof(x))(((unsigned long)(x)) & 0xffffffff))
 767
 768 int sys_clone(unsigned long clone_flags, unsigned long usp,
 769               int __user *parent_tidp, void __user *child_threadptr,
 770               int __user *child_tidp, int p6,
 771               struct pt_regs *regs)
 772 {
 773         CHECK_FULL_REGS(regs);
 774         if (usp == 0)
 775                 usp = regs->gpr[1];     /* stack pointer for child */
 776 #ifdef CONFIG_PPC64
 777         if (test_thread_flag(TIF_32BIT)) {
 778                 parent_tidp = TRUNC_PTR(parent_tidp);
 779                 child_tidp = TRUNC_PTR(child_tidp);
 780         }
 781 #endif
 782         return do_fork(clone_flags, usp, regs, 0, parent_tidp, child_tidp);
 783 }
 784
 785 int sys_fork(unsigned long p1, unsigned long p2, unsigned long p3,
 786              unsigned long p4, unsigned long p5, unsigned long p6,
 787              struct pt_regs *regs)
 788 {
 789         CHECK_FULL_REGS(regs);
 790         return do_fork(SIGCHLD, regs->gpr[1], regs, 0, NULL, NULL);
 791 }
 792
 793 int sys_vfork(unsigned long p1, unsigned long p2, unsigned long p3,
 794               unsigned long p4, unsigned long p5, unsigned long p6,
 795               struct pt_regs *regs)
 796 {
 797         CHECK_FULL_REGS(regs);
 798         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->gpr[1],
 799                         regs, 0, NULL, NULL);
 800 }
 801
 802 int sys_execve(unsigned long a0, unsigned long a1, unsigned long a2,
 803                unsigned long a3, unsigned long a4, unsigned long a5,
 804                struct pt_regs *regs)
 805 {
 806         int error;
 807         char *filename;
 808
 809         filename = getname((char __user *) a0);
 810         error = PTR_ERR(filename);
 811         if (IS_ERR(filename))
 812                 goto out;
 813         flush_fp_to_thread(current);
 814         flush_altivec_to_thread(current);
 815         flush_spe_to_thread(current);
 816         error = do_execve(filename, (char __user * __user *) a1,
 817                           (char __user * __user *) a2, regs);
 818         if (error == 0) {
 819                 task_lock(current);
 820                 current->ptrace &= ~PT_DTRACE;
 821                 task_unlock(current);
 822         }
 823         putname(filename);
 824 out:
 825         return error;
 826 }
 827
 828 int validate_sp(unsigned long sp, struct task_struct *p,
 829                        unsigned long nbytes)
 830 {
 831         unsigned long stack_page = (unsigned long)task_stack_page(p);
 832
 833         if (sp >= stack_page + sizeof(struct thread_struct)
 834             && sp <= stack_page + THREAD_SIZE - nbytes)
 835                 return 1;
 836
 837 #ifdef CONFIG_IRQSTACKS
 838         stack_page = (unsigned long) hardirq_ctx[task_cpu(p)];
 839         if (sp >= stack_page + sizeof(struct thread_struct)
 840             && sp <= stack_page + THREAD_SIZE - nbytes)
 841                 return 1;
 842
 843         stack_page = (unsigned long) softirq_ctx[task_cpu(p)];
 844         if (sp >= stack_page + sizeof(struct thread_struct)
 845             && sp <= stack_page + THREAD_SIZE - nbytes)
 846                 return 1;
 847 #endif
 848
 849         return 0;
 850 }
 851
 852 #ifdef CONFIG_PPC64
 853 #define MIN_STACK_FRAME 112     /* same as STACK_FRAME_OVERHEAD, in fact */
 854 #define FRAME_LR_SAVE   2
 855 #define INT_FRAME_SIZE  (sizeof(struct pt_regs) + STACK_FRAME_OVERHEAD + 288)
 856 #define REGS_MARKER     0x7265677368657265ul
 857 #define FRAME_MARKER    12
 858 #else
 859 #define MIN_STACK_FRAME 16
 860 #define FRAME_LR_SAVE   1
 861 #define INT_FRAME_SIZE  (sizeof(struct pt_regs) + STACK_FRAME_OVERHEAD)
 862 #define REGS_MARKER     0x72656773ul
 863 #define FRAME_MARKER    2
 864 #endif
 865
 866 EXPORT_SYMBOL(validate_sp);
 867
 868 unsigned long get_wchan(struct task_struct *p)
 869 {
 870         unsigned long ip, sp;
 871         int count = 0;
 872
 873         if (!p || p == current || p->state == TASK_RUNNING)
 874                 return 0;
 875
 876         sp = p->thread.ksp;
 877         if (!validate_sp(sp, p, MIN_STACK_FRAME))
 878                 return 0;
 879
 880         do {
 881                 sp = *(unsigned long *)sp;
 882                 if (!validate_sp(sp, p, MIN_STACK_FRAME))
 883                         return 0;
 884                 if (count > 0) {
 885                         ip = ((unsigned long *)sp)[FRAME_LR_SAVE];
 886                         if (!in_sched_functions(ip))
 887                                 return ip;
 888                 }
 889         } while (count++ < 16);
 890         return 0;
 891 }
 892
 893 static int kstack_depth_to_print = 64;
 894
 895 void show_stack(struct task_struct *tsk, unsigned long *stack)
 896 {
 897         unsigned long sp, ip, lr, newsp;
 898         int count = 0;
 899         int firstframe = 1;
 900
 901         sp = (unsigned long) stack;
 902         if (tsk == NULL)
 903                 tsk = current;
 904         if (sp == 0) {
 905                 if (tsk == current)
 906                         asm("mr %0,1" : "=r" (sp));
 907                 else
 908                         sp = tsk->thread.ksp;
 909         }
 910
 911         lr = 0;
 912         printk("Call Trace:\n");
 913         do {
 914                 if (!validate_sp(sp, tsk, MIN_STACK_FRAME))
 915                         return;
 916
 917                 stack = (unsigned long *) sp;
 918                 newsp = stack[0];
 919                 ip = stack[FRAME_LR_SAVE];
 920                 if (!firstframe || ip != lr) {
 921                         printk("["REG"] ["REG"] ", sp, ip);
 922                         print_symbol("%s", ip);
 923                         if (firstframe)
 924                                 printk(" (unreliable)");
 925                         printk("\n");
 926                 }
 927                 firstframe = 0;
 928
 929                 /*
 930                  * See if this is an exception frame.
 931                  * We look for the "regshere" marker in the current frame.
 932                  */
 933                 if (validate_sp(sp, tsk, INT_FRAME_SIZE)
 934                     && stack[FRAME_MARKER] == REGS_MARKER) {
 935                         struct pt_regs *regs = (struct pt_regs *)
 936                                 (sp + STACK_FRAME_OVERHEAD);
 937                         printk("--- Exception: %lx", regs->trap);
 938                         print_symbol(" at %s\n", regs->nip);
 939                         lr = regs->link;
 940                         print_symbol("    LR = %s\n", lr);
 941                         firstframe = 1;
 942                 }
 943
 944                 sp = newsp;
 945         } while (count++ < kstack_depth_to_print);
 946 }
 947
 948 void dump_stack(void)
 949 {
 950         show_stack(current, NULL);
 951 }
 952 EXPORT_SYMBOL(dump_stack);
 953
 954 #ifdef CONFIG_PPC64
 955 void ppc64_runlatch_on(void)
 956 {
 957         unsigned long ctrl;
 958
 959         if (cpu_has_feature(CPU_FTR_CTRL) && !test_thread_flag(TIF_RUNLATCH)) {
 960                 HMT_medium();
 961
 962                 ctrl = mfspr(SPRN_CTRLF);
 963                 ctrl |= CTRL_RUNLATCH;
 964                 mtspr(SPRN_CTRLT, ctrl);
 965
 966                 set_thread_flag(TIF_RUNLATCH);
 967         }
 968 }
 969
 970 void ppc64_runlatch_off(void)
 971 {
 972         unsigned long ctrl;
 973
 974         if (cpu_has_feature(CPU_FTR_CTRL) && test_thread_flag(TIF_RUNLATCH)) {
 975                 HMT_medium();
 976
 977                 clear_thread_flag(TIF_RUNLATCH);
 978
 979                 ctrl = mfspr(SPRN_CTRLF);
 980                 ctrl &= ~CTRL_RUNLATCH;
 981                 mtspr(SPRN_CTRLT, ctrl);
 982         }
 983 }
 984 #endif