arch/ppc64/kernel/process.c

   1 /*
   2  *  linux/arch/ppc64/kernel/process.c
   3  *
   4  *  Derived from "arch/i386/kernel/process.c"
   5  *    Copyright (C) 1995  Linus Torvalds
   6  *
   7  *  Updated and modified by Cort Dougan (cort@cs.nmt.edu) and
   8  *  Paul Mackerras (paulus@cs.anu.edu.au)
   9  *
  10  *  PowerPC version
  11  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
  12  *
  13  *  This program is free software; you can redistribute it and/or
  14  *  modify it under the terms of the GNU General Public License
  15  *  as published by the Free Software Foundation; either version
  16  *  2 of the License, or (at your option) any later version.
  17  */
  18
  19 #include <linux/config.h>
  20 #include <linux/module.h>
  21 #include <linux/errno.h>
  22 #include <linux/sched.h>
  23 #include <linux/kernel.h>
  24 #include <linux/mm.h>
  25 #include <linux/smp.h>
  26 #include <linux/smp_lock.h>
  27 #include <linux/stddef.h>
  28 #include <linux/unistd.h>
  29 #include <linux/slab.h>
  30 #include <linux/user.h>
  31 #include <linux/elf.h>
  32 #include <linux/init.h>
  33 #include <linux/init_task.h>
  34 #include <linux/prctl.h>
  35 #include <linux/ptrace.h>
  36 #include <linux/kallsyms.h>
  37 #include <linux/interrupt.h>
  38 #include <linux/version.h>
  39
  40 #include <asm/pgtable.h>
  41 #include <asm/uaccess.h>
  42 #include <asm/system.h>
  43 #include <asm/io.h>
  44 #include <asm/processor.h>
  45 #include <asm/mmu.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/prom.h>
  48 #include <asm/ppcdebug.h>
  49 #include <asm/machdep.h>
  50 #include <asm/iSeries/HvCallHpt.h>
  51 #include <asm/cputable.h>
  52 #include <asm/sections.h>
  53 #include <asm/tlbflush.h>
  54
  55 #ifndef CONFIG_SMP
  56 struct task_struct *last_task_used_math = NULL;
  57 struct task_struct *last_task_used_altivec = NULL;
  58 #endif
  59
  60 struct mm_struct ioremap_mm = {
  61         .pgd            = ioremap_dir,
  62         .mm_users       = ATOMIC_INIT(2),
  63         .mm_count       = ATOMIC_INIT(1),
  64         .cpu_vm_mask    = CPU_MASK_ALL,
  65         .page_table_lock = SPIN_LOCK_UNLOCKED,
  66 };
  67
  68 /*
  69  * Make sure the floating-point register state in the
  70  * the thread_struct is up to date for task tsk.
  71  */
  72 void flush_fp_to_thread(struct task_struct *tsk)
  73 {
  74         if (tsk->thread.regs) {
  75                 /*
  76                  * We need to disable preemption here because if we didn't,
  77                  * another process could get scheduled after the regs->msr
  78                  * test but before we have finished saving the FP registers
  79                  * to the thread_struct.  That process could take over the
  80                  * FPU, and then when we get scheduled again we would store
  81                  * bogus values for the remaining FP registers.
  82                  */
  83                 preempt_disable();
  84                 if (tsk->thread.regs->msr & MSR_FP) {
  85 #ifdef CONFIG_SMP
  86                         /*
  87                          * This should only ever be called for current or
  88                          * for a stopped child process.  Since we save away
  89                          * the FP register state on context switch on SMP,
  90                          * there is something wrong if a stopped child appears
  91                          * to still have its FP state in the CPU registers.
  92                          */
  93                         BUG_ON(tsk != current);
  94 #endif
  95                         giveup_fpu(current);
  96                 }
  97                 preempt_enable();
  98         }
  99 }
 100
 101 void enable_kernel_fp(void)
 102 {
 103         WARN_ON(preemptible());
 104
 105 #ifdef CONFIG_SMP
 106         if (current->thread.regs && (current->thread.regs->msr & MSR_FP))
 107                 giveup_fpu(current);
 108         else
 109                 giveup_fpu(NULL);       /* just enables FP for kernel */
 110 #else
 111         giveup_fpu(last_task_used_math);
 112 #endif /* CONFIG_SMP */
 113 }
 114 EXPORT_SYMBOL(enable_kernel_fp);
 115
 116 int dump_task_fpu(struct task_struct *tsk, elf_fpregset_t *fpregs)
 117 {
 118         if (!tsk->thread.regs)
 119                 return 0;
 120         flush_fp_to_thread(current);
 121
 122         memcpy(fpregs, &tsk->thread.fpr[0], sizeof(*fpregs));
 123
 124         return 1;
 125 }
 126
 127 #ifdef CONFIG_ALTIVEC
 128
 129 void enable_kernel_altivec(void)
 130 {
 131         WARN_ON(preemptible());
 132
 133 #ifdef CONFIG_SMP
 134         if (current->thread.regs && (current->thread.regs->msr & MSR_VEC))
 135                 giveup_altivec(current);
 136         else
 137                 giveup_altivec(NULL);   /* just enables FP for kernel */
 138 #else
 139         giveup_altivec(last_task_used_altivec);
 140 #endif /* CONFIG_SMP */
 141 }
 142 EXPORT_SYMBOL(enable_kernel_altivec);
 143
 144 /*
 145  * Make sure the VMX/Altivec register state in the
 146  * the thread_struct is up to date for task tsk.
 147  */
 148 void flush_altivec_to_thread(struct task_struct *tsk)
 149 {
 150         if (tsk->thread.regs) {
 151                 preempt_disable();
 152                 if (tsk->thread.regs->msr & MSR_VEC) {
 153 #ifdef CONFIG_SMP
 154                         BUG_ON(tsk != current);
 155 #endif
 156                         giveup_altivec(current);
 157                 }
 158                 preempt_enable();
 159         }
 160 }
 161
 162 int dump_task_altivec(struct pt_regs *regs, elf_vrregset_t *vrregs)
 163 {
 164         flush_altivec_to_thread(current);
 165         memcpy(vrregs, &current->thread.vr[0], sizeof(*vrregs));
 166         return 1;
 167 }
 168
 169 #endif /* CONFIG_ALTIVEC */
 170
 171 struct task_struct *__switch_to(struct task_struct *prev,
 172                                 struct task_struct *new)
 173 {
 174         struct thread_struct *new_thread, *old_thread;
 175         unsigned long flags;
 176         struct task_struct *last;
 177
 178 #ifdef CONFIG_SMP
 179         /* avoid complexity of lazy save/restore of fpu
 180          * by just saving it every time we switch out if
 181          * this task used the fpu during the last quantum.
 182          *
 183          * If it tries to use the fpu again, it'll trap and
 184          * reload its fp regs.  So we don't have to do a restore
 185          * every switch, just a save.
 186          *  -- Cort
 187          */
 188         if (prev->thread.regs && (prev->thread.regs->msr & MSR_FP))
 189                 giveup_fpu(prev);
 190 #ifdef CONFIG_ALTIVEC
 191         if (prev->thread.regs && (prev->thread.regs->msr & MSR_VEC))
 192                 giveup_altivec(prev);
 193 #endif /* CONFIG_ALTIVEC */
 194 #endif /* CONFIG_SMP */
 195
 196 #if defined(CONFIG_ALTIVEC) && !defined(CONFIG_SMP)
 197         /* Avoid the trap.  On smp this this never happens since
 198          * we don't set last_task_used_altivec -- Cort
 199          */
 200         if (new->thread.regs && last_task_used_altivec == new)
 201                 new->thread.regs->msr |= MSR_VEC;
 202 #endif /* CONFIG_ALTIVEC */
 203
 204         flush_tlb_pending();
 205
 206         new_thread = &new->thread;
 207         old_thread = &current->thread;
 208
 209         local_irq_save(flags);
 210         last = _switch(old_thread, new_thread);
 211
 212         local_irq_restore(flags);
 213
 214         return last;
 215 }
 216
 217 void show_regs(struct pt_regs * regs)
 218 {
 219         int i;
 220         unsigned long trap;
 221
 222         printk("NIP: %016lX XER: %016lX LR: %016lX\n",
 223                regs->nip, regs->xer, regs->link);
 224         printk("REGS: %p TRAP: %04lx   %s  (%s)\n",
 225                regs, regs->trap, print_tainted(), UTS_RELEASE);
 226         printk("MSR: %016lx EE: %01x PR: %01x FP: %01x ME: %01x IR/DR: %01x%01x\n",
 227                regs->msr, regs->msr&MSR_EE ? 1 : 0, regs->msr&MSR_PR ? 1 : 0,
 228                regs->msr & MSR_FP ? 1 : 0,regs->msr&MSR_ME ? 1 : 0,
 229                regs->msr&MSR_IR ? 1 : 0,
 230                regs->msr&MSR_DR ? 1 : 0);
 231         trap = TRAP(regs);
 232         if (trap == 0x300 || trap == 0x380 || trap == 0x600)
 233                 printk("DAR: %016lx, DSISR: %016lx\n", regs->dar, regs->dsisr);
 234         printk("TASK: %p[%d] '%s' THREAD: %p",
 235                current, current->pid, current->comm, current->thread_info);
 236
 237 #ifdef CONFIG_SMP
 238         printk(" CPU: %d", smp_processor_id());
 239 #endif /* CONFIG_SMP */
 240
 241         for (i = 0; i < 32; i++) {
 242                 if ((i % 4) == 0) {
 243                         printk("\n" KERN_INFO "GPR%02d: ", i);
 244                 }
 245
 246                 printk("%016lX ", regs->gpr[i]);
 247                 if (i == 13 && !FULL_REGS(regs))
 248                         break;
 249         }
 250         printk("\n");
 251         /*
 252          * Lookup NIP late so we have the best change of getting the
 253          * above info out without failing
 254          */
 255         printk("NIP [%016lx] ", regs->nip);
 256         print_symbol("%s\n", regs->nip);
 257         printk("LR [%016lx] ", regs->link);
 258         print_symbol("%s\n", regs->link);
 259         show_stack(current, (unsigned long *)regs->gpr[1]);
 260 }
 261 EXPORT_SYMBOL_GPL(show_regs);
 262
 263 void exit_thread(void)
 264 {
 265 #ifndef CONFIG_SMP
 266         if (last_task_used_math == current)
 267                 last_task_used_math = NULL;
 268 #ifdef CONFIG_ALTIVEC
 269         if (last_task_used_altivec == current)
 270                 last_task_used_altivec = NULL;
 271 #endif /* CONFIG_ALTIVEC */
 272 #endif /* CONFIG_SMP */
 273 }
 274
 275 void flush_thread(void)
 276 {
 277         struct thread_info *t = current_thread_info();
 278
 279         if (t->flags & _TIF_ABI_PENDING)
 280                 t->flags ^= (_TIF_ABI_PENDING | _TIF_32BIT);
 281
 282 #ifndef CONFIG_SMP
 283         if (last_task_used_math == current)
 284                 last_task_used_math = NULL;
 285 #ifdef CONFIG_ALTIVEC
 286         if (last_task_used_altivec == current)
 287                 last_task_used_altivec = NULL;
 288 #endif /* CONFIG_ALTIVEC */
 289 #endif /* CONFIG_SMP */
 290 }
 291
 292 void
 293 release_thread(struct task_struct *t)
 294 {
 295 }
 296
 297
 298 /*
 299  * This gets called before we allocate a new thread and copy
 300  * the current task into it.
 301  */
 302 void prepare_to_copy(struct task_struct *tsk)
 303 {
 304         flush_fp_to_thread(current);
 305         flush_altivec_to_thread(current);
 306 }
 307
 308 /*
 309  * Copy a thread..
 310  */
 311 int
 312 copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 313             unsigned long unused, struct task_struct *p, struct pt_regs *regs)
 314 {
 315         struct pt_regs *childregs, *kregs;
 316         extern void ret_from_fork(void);
 317         unsigned long sp = (unsigned long)p->thread_info + THREAD_SIZE;
 318
 319         /* Copy registers */
 320         sp -= sizeof(struct pt_regs);
 321         childregs = (struct pt_regs *) sp;
 322         *childregs = *regs;
 323         if ((childregs->msr & MSR_PR) == 0) {
 324                 /* for kernel thread, set stackptr in new task */
 325                 childregs->gpr[1] = sp + sizeof(struct pt_regs);
 326                 p->thread.regs = NULL;  /* no user register state */
 327                 clear_ti_thread_flag(p->thread_info, TIF_32BIT);
 328 #ifdef CONFIG_PPC_ISERIES
 329                 set_ti_thread_flag(p->thread_info, TIF_RUN_LIGHT);
 330 #endif
 331         } else {
 332                 childregs->gpr[1] = usp;
 333                 p->thread.regs = childregs;
 334                 if (clone_flags & CLONE_SETTLS) {
 335                         if (test_thread_flag(TIF_32BIT))
 336                                 childregs->gpr[2] = childregs->gpr[6];
 337                         else
 338                                 childregs->gpr[13] = childregs->gpr[6];
 339                 }
 340         }
 341         childregs->gpr[3] = 0;  /* Result from fork() */
 342         sp -= STACK_FRAME_OVERHEAD;
 343
 344         /*
 345          * The way this works is that at some point in the future
 346          * some task will call _switch to switch to the new task.
 347          * That will pop off the stack frame created below and start
 348          * the new task running at ret_from_fork.  The new task will
 349          * do some house keeping and then return from the fork or clone
 350          * system call, using the stack frame created above.
 351          */
 352         sp -= sizeof(struct pt_regs);
 353         kregs = (struct pt_regs *) sp;
 354         sp -= STACK_FRAME_OVERHEAD;
 355         p->thread.ksp = sp;
 356         if (cur_cpu_spec->cpu_features & CPU_FTR_SLB) {
 357                 unsigned long sp_vsid = get_kernel_vsid(sp);
 358
 359                 sp_vsid <<= SLB_VSID_SHIFT;
 360                 sp_vsid |= SLB_VSID_KERNEL;
 361                 if (cur_cpu_spec->cpu_features & CPU_FTR_16M_PAGE)
 362                         sp_vsid |= SLB_VSID_L;
 363
 364                 p->thread.ksp_vsid = sp_vsid;
 365         }
 366
 367         /*
 368          * The PPC64 ABI makes use of a TOC to contain function
 369          * pointers.  The function (ret_from_except) is actually a pointer
 370          * to the TOC entry.  The first entry is a pointer to the actual
 371          * function.
 372          */
 373         kregs->nip = *((unsigned long *)ret_from_fork);
 374
 375         return 0;
 376 }
 377
 378 /*
 379  * Set up a thread for executing a new program
 380  */
 381 void start_thread(struct pt_regs *regs, unsigned long fdptr, unsigned long sp)
 382 {
 383         unsigned long entry, toc, load_addr = regs->gpr[2];
 384
 385         /* fdptr is a relocated pointer to the function descriptor for
 386          * the elf _start routine.  The first entry in the function
 387          * descriptor is the entry address of _start and the second
 388          * entry is the TOC value we need to use.
 389          */
 390         set_fs(USER_DS);
 391         __get_user(entry, (unsigned long __user *)fdptr);
 392         __get_user(toc, (unsigned long __user *)fdptr+1);
 393
 394         /* Check whether the e_entry function descriptor entries
 395          * need to be relocated before we can use them.
 396          */
 397         if (load_addr != 0) {
 398                 entry += load_addr;
 399                 toc   += load_addr;
 400         }
 401
 402         /*
 403          * If we exec out of a kernel thread then thread.regs will not be
 404          * set. Do it now.
 405          */
 406         if (!current->thread.regs) {
 407                 unsigned long childregs = (unsigned long)current->thread_info +
 408                                                 THREAD_SIZE;
 409                 childregs -= sizeof(struct pt_regs);
 410                 current->thread.regs = (struct pt_regs *)childregs;
 411         }
 412
 413         regs->nip = entry;
 414         regs->gpr[1] = sp;
 415         regs->gpr[2] = toc;
 416         regs->msr = MSR_USER64;
 417 #ifndef CONFIG_SMP
 418         if (last_task_used_math == current)
 419                 last_task_used_math = 0;
 420 #endif /* CONFIG_SMP */
 421         memset(current->thread.fpr, 0, sizeof(current->thread.fpr));
 422         current->thread.fpscr = 0;
 423 #ifdef CONFIG_ALTIVEC
 424 #ifndef CONFIG_SMP
 425         if (last_task_used_altivec == current)
 426                 last_task_used_altivec = 0;
 427 #endif /* CONFIG_SMP */
 428         memset(current->thread.vr, 0, sizeof(current->thread.vr));
 429         current->thread.vscr.u[0] = 0;
 430         current->thread.vscr.u[1] = 0;
 431         current->thread.vscr.u[2] = 0;
 432         current->thread.vscr.u[3] = 0x00010000; /* Java mode disabled */
 433         current->thread.vrsave = 0;
 434         current->thread.used_vr = 0;
 435 #endif /* CONFIG_ALTIVEC */
 436 }
 437
 438 int set_fpexc_mode(struct task_struct *tsk, unsigned int val)
 439 {
 440         struct pt_regs *regs = tsk->thread.regs;
 441
 442         if (val > PR_FP_EXC_PRECISE)
 443                 return -EINVAL;
 444         tsk->thread.fpexc_mode = __pack_fe01(val);
 445         if (regs != NULL && (regs->msr & MSR_FP) != 0)
 446                 regs->msr = (regs->msr & ~(MSR_FE0|MSR_FE1))
 447                         | tsk->thread.fpexc_mode;
 448         return 0;
 449 }
 450
 451 int get_fpexc_mode(struct task_struct *tsk, unsigned long adr)
 452 {
 453         unsigned int val;
 454
 455         val = __unpack_fe01(tsk->thread.fpexc_mode);
 456         return put_user(val, (unsigned int __user *) adr);
 457 }
 458
 459 int sys_clone(unsigned long clone_flags, unsigned long p2, unsigned long p3,
 460               unsigned long p4, unsigned long p5, unsigned long p6,
 461               struct pt_regs *regs)
 462 {
 463         unsigned long parent_tidptr = 0;
 464         unsigned long child_tidptr = 0;
 465
 466         if (p2 == 0)
 467                 p2 = regs->gpr[1];      /* stack pointer for child */
 468
 469         if (clone_flags & (CLONE_PARENT_SETTID | CLONE_CHILD_SETTID |
 470                            CLONE_CHILD_CLEARTID)) {
 471                 parent_tidptr = p3;
 472                 child_tidptr = p5;
 473                 if (test_thread_flag(TIF_32BIT)) {
 474                         parent_tidptr &= 0xffffffff;
 475                         child_tidptr &= 0xffffffff;
 476                 }
 477         }
 478
 479         return do_fork(clone_flags, p2, regs, 0,
 480                     (int __user *)parent_tidptr, (int __user *)child_tidptr);
 481 }
 482
 483 int sys_fork(unsigned long p1, unsigned long p2, unsigned long p3,
 484              unsigned long p4, unsigned long p5, unsigned long p6,
 485              struct pt_regs *regs)
 486 {
 487         return do_fork(SIGCHLD, regs->gpr[1], regs, 0, NULL, NULL);
 488 }
 489
 490 int sys_vfork(unsigned long p1, unsigned long p2, unsigned long p3,
 491               unsigned long p4, unsigned long p5, unsigned long p6,
 492               struct pt_regs *regs)
 493 {
 494         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->gpr[1], regs, 0,
 495                     NULL, NULL);
 496 }
 497
 498 int sys_execve(unsigned long a0, unsigned long a1, unsigned long a2,
 499                unsigned long a3, unsigned long a4, unsigned long a5,
 500                struct pt_regs *regs)
 501 {
 502         int error;
 503         char * filename;
 504
 505         filename = getname((char __user *) a0);
 506         error = PTR_ERR(filename);
 507         if (IS_ERR(filename))
 508                 goto out;
 509         flush_fp_to_thread(current);
 510         flush_altivec_to_thread(current);
 511         error = do_execve(filename, (char __user * __user *) a1,
 512                                     (char __user * __user *) a2, regs);
 513
 514         if (error == 0) {
 515                 task_lock(current);
 516                 current->ptrace &= ~PT_DTRACE;
 517                 task_unlock(current);
 518         }
 519         putname(filename);
 520
 521 out:
 522         return error;
 523 }
 524
 525 static int kstack_depth_to_print = 64;
 526
 527 static int validate_sp(unsigned long sp, struct task_struct *p,
 528                        unsigned long nbytes)
 529 {
 530         unsigned long stack_page = (unsigned long)p->thread_info;
 531
 532         if (sp >= stack_page + sizeof(struct thread_struct)
 533             && sp <= stack_page + THREAD_SIZE - nbytes)
 534                 return 1;
 535
 536 #ifdef CONFIG_IRQSTACKS
 537         stack_page = (unsigned long) hardirq_ctx[task_cpu(p)];
 538         if (sp >= stack_page + sizeof(struct thread_struct)
 539             && sp <= stack_page + THREAD_SIZE - nbytes)
 540                 return 1;
 541
 542         stack_page = (unsigned long) softirq_ctx[task_cpu(p)];
 543         if (sp >= stack_page + sizeof(struct thread_struct)
 544             && sp <= stack_page + THREAD_SIZE - nbytes)
 545                 return 1;
 546 #endif
 547
 548         return 0;
 549 }
 550
 551 unsigned long get_wchan(struct task_struct *p)
 552 {
 553         unsigned long ip, sp;
 554         int count = 0;
 555
 556         if (!p || p == current || p->state == TASK_RUNNING)
 557                 return 0;
 558
 559         sp = p->thread.ksp;
 560         if (!validate_sp(sp, p, 112))
 561                 return 0;
 562
 563         do {
 564                 sp = *(unsigned long *)sp;
 565                 if (!validate_sp(sp, p, 112))
 566                         return 0;
 567                 if (count > 0) {
 568                         ip = *(unsigned long *)(sp + 16);
 569                         if (!in_sched_functions(ip))
 570                                 return ip;
 571                 }
 572         } while (count++ < 16);
 573         return 0;
 574 }
 575
 576 void show_stack(struct task_struct *p, unsigned long *_sp)
 577 {
 578         unsigned long ip, newsp, lr;
 579         int count = 0;
 580         unsigned long sp = (unsigned long)_sp;
 581         int firstframe = 1;
 582
 583         if (sp == 0) {
 584                 if (p) {
 585                         sp = p->thread.ksp;
 586                 } else {
 587                         sp = __get_SP();
 588                         p = current;
 589                 }
 590         }
 591
 592         lr = 0;
 593         printk("Call Trace:\n");
 594         do {
 595                 if (!validate_sp(sp, p, 112))
 596                         return;
 597
 598                 _sp = (unsigned long *) sp;
 599                 newsp = _sp[0];
 600                 ip = _sp[2];
 601                 if (!firstframe || ip != lr) {
 602                         printk("[%016lx] [%016lx] ", sp, ip);
 603                         print_symbol("%s", ip);
 604                         if (firstframe)
 605                                 printk(" (unreliable)");
 606                         printk("\n");
 607                 }
 608                 firstframe = 0;
 609
 610                 /*
 611                  * See if this is an exception frame.
 612                  * We look for the "regshere" marker in the current frame.
 613                  */
 614                 if (validate_sp(sp, p, sizeof(struct pt_regs) + 400)
 615                     && _sp[12] == 0x7265677368657265ul) {
 616                         struct pt_regs *regs = (struct pt_regs *)
 617                                 (sp + STACK_FRAME_OVERHEAD);
 618                         printk("--- Exception: %lx", regs->trap);
 619                         print_symbol(" at %s\n", regs->nip);
 620                         lr = regs->link;
 621                         print_symbol("    LR = %s\n", lr);
 622                         firstframe = 1;
 623                 }
 624
 625                 sp = newsp;
 626         } while (count++ < kstack_depth_to_print);
 627 }
 628
 629 void dump_stack(void)
 630 {
 631         show_stack(current, (unsigned long *)__get_SP());
 632 }
 633 EXPORT_SYMBOL(dump_stack);