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/version.h>
  38
  39 #include <asm/pgtable.h>
  40 #include <asm/uaccess.h>
  41 #include <asm/system.h>
  42 #include <asm/io.h>
  43 #include <asm/processor.h>
  44 #include <asm/mmu.h>
  45 #include <asm/mmu_context.h>
  46 #include <asm/prom.h>
  47 #include <asm/ppcdebug.h>
  48 #include <asm/machdep.h>
  49 #include <asm/iSeries/HvCallHpt.h>
  50 #include <asm/hardirq.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 void enable_kernel_fp(void)
  69 {
  70 #ifdef CONFIG_SMP
  71         if (current->thread.regs && (current->thread.regs->msr & MSR_FP))
  72                 giveup_fpu(current);
  73         else
  74                 giveup_fpu(NULL);       /* just enables FP for kernel */
  75 #else
  76         giveup_fpu(last_task_used_math);
  77 #endif /* CONFIG_SMP */
  78 }
  79 EXPORT_SYMBOL(enable_kernel_fp);
  80
  81 int dump_task_fpu(struct task_struct *tsk, elf_fpregset_t *fpregs)
  82 {
  83         struct pt_regs *regs = tsk->thread.regs;
  84
  85         if (!regs)
  86                 return 0;
  87         if (tsk == current && (regs->msr & MSR_FP))
  88                 giveup_fpu(current);
  89
  90         memcpy(fpregs, &tsk->thread.fpr[0], sizeof(*fpregs));
  91
  92         return 1;
  93 }
  94
  95 #ifdef CONFIG_ALTIVEC
  96
  97 void enable_kernel_altivec(void)
  98 {
  99 #ifdef CONFIG_SMP
 100         if (current->thread.regs && (current->thread.regs->msr & MSR_VEC))
 101                 giveup_altivec(current);
 102         else
 103                 giveup_altivec(NULL);   /* just enables FP for kernel */
 104 #else
 105         giveup_altivec(last_task_used_altivec);
 106 #endif /* CONFIG_SMP */
 107 }
 108 EXPORT_SYMBOL(enable_kernel_altivec);
 109
 110 int dump_task_altivec(struct pt_regs *regs, elf_vrregset_t *vrregs)
 111 {
 112         if (regs->msr & MSR_VEC)
 113                 giveup_altivec(current);
 114         memcpy(vrregs, &current->thread.vr[0], sizeof(*vrregs));
 115         return 1;
 116 }
 117
 118 #endif /* CONFIG_ALTIVEC */
 119
 120 struct task_struct *__switch_to(struct task_struct *prev,
 121                                 struct task_struct *new)
 122 {
 123         struct thread_struct *new_thread, *old_thread;
 124         unsigned long flags;
 125         struct task_struct *last;
 126
 127 #ifdef CONFIG_SMP
 128         /* avoid complexity of lazy save/restore of fpu
 129          * by just saving it every time we switch out if
 130          * this task used the fpu during the last quantum.
 131          *
 132          * If it tries to use the fpu again, it'll trap and
 133          * reload its fp regs.  So we don't have to do a restore
 134          * every switch, just a save.
 135          *  -- Cort
 136          */
 137         if (prev->thread.regs && (prev->thread.regs->msr & MSR_FP))
 138                 giveup_fpu(prev);
 139 #ifdef CONFIG_ALTIVEC
 140         if (prev->thread.regs && (prev->thread.regs->msr & MSR_VEC))
 141                 giveup_altivec(prev);
 142 #endif /* CONFIG_ALTIVEC */
 143 #endif /* CONFIG_SMP */
 144
 145 #if defined(CONFIG_ALTIVEC) && !defined(CONFIG_SMP)
 146         /* Avoid the trap.  On smp this this never happens since
 147          * we don't set last_task_used_altivec -- Cort
 148          */
 149         if (new->thread.regs && last_task_used_altivec == new)
 150                 new->thread.regs->msr |= MSR_VEC;
 151 #endif /* CONFIG_ALTIVEC */
 152
 153         flush_tlb_pending();
 154
 155         new_thread = &new->thread;
 156         old_thread = &current->thread;
 157
 158         local_irq_save(flags);
 159         last = _switch(old_thread, new_thread);
 160
 161         local_irq_restore(flags);
 162
 163         return last;
 164 }
 165
 166 void show_regs(struct pt_regs * regs)
 167 {
 168         int i;
 169
 170         printk("NIP: %016lX XER: %016lX LR: %016lX\n",
 171                regs->nip, regs->xer, regs->link);
 172         printk("REGS: %p TRAP: %04lx   %s  (%s)\n",
 173                regs, regs->trap, print_tainted(), UTS_RELEASE);
 174         printk("MSR: %016lx EE: %01x PR: %01x FP: %01x ME: %01x IR/DR: %01x%01x\n",
 175                regs->msr, regs->msr&MSR_EE ? 1 : 0, regs->msr&MSR_PR ? 1 : 0,
 176                regs->msr & MSR_FP ? 1 : 0,regs->msr&MSR_ME ? 1 : 0,
 177                regs->msr&MSR_IR ? 1 : 0,
 178                regs->msr&MSR_DR ? 1 : 0);
 179         if (regs->trap == 0x300 || regs->trap == 0x380 || regs->trap == 0x600)
 180                 printk("DAR: %016lx, DSISR: %016lx\n", regs->dar, regs->dsisr);
 181         printk("TASK: %p[%d] '%s' THREAD: %p",
 182                current, current->pid, current->comm, current->thread_info);
 183
 184 #ifdef CONFIG_SMP
 185         printk(" CPU: %d", smp_processor_id());
 186 #endif /* CONFIG_SMP */
 187
 188         for (i = 0; i < 32; i++) {
 189                 if ((i % 4) == 0) {
 190                         printk("\n" KERN_INFO "GPR%02d: ", i);
 191                 }
 192
 193                 printk("%016lX ", regs->gpr[i]);
 194         }
 195         printk("\n");
 196         /*
 197          * Lookup NIP late so we have the best change of getting the
 198          * above info out without failing
 199          */
 200         printk("NIP [%016lx] ", regs->nip);
 201         print_symbol("%s\n", regs->nip);
 202         printk("LR [%016lx] ", regs->link);
 203         print_symbol("%s\n", regs->link);
 204         show_stack(current, (unsigned long *)regs->gpr[1]);
 205 }
 206
 207 void exit_thread(void)
 208 {
 209 #ifndef CONFIG_SMP
 210         if (last_task_used_math == current)
 211                 last_task_used_math = NULL;
 212 #ifdef CONFIG_ALTIVEC
 213         if (last_task_used_altivec == current)
 214                 last_task_used_altivec = NULL;
 215 #endif /* CONFIG_ALTIVEC */
 216 #endif /* CONFIG_SMP */
 217 }
 218
 219 void flush_thread(void)
 220 {
 221         struct thread_info *t = current_thread_info();
 222
 223         if (t->flags & _TIF_ABI_PENDING)
 224                 t->flags ^= (_TIF_ABI_PENDING | _TIF_32BIT);
 225
 226 #ifndef CONFIG_SMP
 227         if (last_task_used_math == current)
 228                 last_task_used_math = NULL;
 229 #ifdef CONFIG_ALTIVEC
 230         if (last_task_used_altivec == current)
 231                 last_task_used_altivec = NULL;
 232 #endif /* CONFIG_ALTIVEC */
 233 #endif /* CONFIG_SMP */
 234 }
 235
 236 void
 237 release_thread(struct task_struct *t)
 238 {
 239 }
 240
 241
 242 /*
 243  * This gets called before we allocate a new thread and copy
 244  * the current task into it.
 245  */
 246 void prepare_to_copy(struct task_struct *tsk)
 247 {
 248         struct pt_regs *regs = tsk->thread.regs;
 249
 250         if (regs == NULL)
 251                 return;
 252         if (regs->msr & MSR_FP)
 253                 giveup_fpu(current);
 254 #ifdef CONFIG_ALTIVEC
 255         if (regs->msr & MSR_VEC)
 256                 giveup_altivec(current);
 257 #endif /* CONFIG_ALTIVEC */
 258 }
 259
 260 /*
 261  * Copy a thread..
 262  */
 263 int
 264 copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 265             unsigned long unused, struct task_struct *p, struct pt_regs *regs)
 266 {
 267         struct pt_regs *childregs, *kregs;
 268         extern void ret_from_fork(void);
 269         unsigned long sp = (unsigned long)p->thread_info + THREAD_SIZE;
 270
 271         p->set_child_tid = p->clear_child_tid = NULL;
 272
 273         /* Copy registers */
 274         sp -= sizeof(struct pt_regs);
 275         childregs = (struct pt_regs *) sp;
 276         *childregs = *regs;
 277         if ((childregs->msr & MSR_PR) == 0) {
 278                 /* for kernel thread, set stackptr in new task */
 279                 childregs->gpr[1] = sp + sizeof(struct pt_regs);
 280                 p->thread.regs = NULL;  /* no user register state */
 281                 clear_ti_thread_flag(p->thread_info, TIF_32BIT);
 282 #ifdef CONFIG_PPC_ISERIES
 283                 set_ti_thread_flag(p->thread_info, TIF_RUN_LIGHT);
 284 #endif
 285         } else {
 286                 childregs->gpr[1] = usp;
 287                 p->thread.regs = childregs;
 288                 if (clone_flags & CLONE_SETTLS) {
 289                         if (test_thread_flag(TIF_32BIT))
 290                                 childregs->gpr[2] = childregs->gpr[6];
 291                         else
 292                                 childregs->gpr[13] = childregs->gpr[6];
 293                 }
 294         }
 295         childregs->gpr[3] = 0;  /* Result from fork() */
 296         sp -= STACK_FRAME_OVERHEAD;
 297
 298         /*
 299          * The way this works is that at some point in the future
 300          * some task will call _switch to switch to the new task.
 301          * That will pop off the stack frame created below and start
 302          * the new task running at ret_from_fork.  The new task will
 303          * do some house keeping and then return from the fork or clone
 304          * system call, using the stack frame created above.
 305          */
 306         sp -= sizeof(struct pt_regs);
 307         kregs = (struct pt_regs *) sp;
 308         sp -= STACK_FRAME_OVERHEAD;
 309         p->thread.ksp = sp;
 310
 311         /*
 312          * The PPC64 ABI makes use of a TOC to contain function
 313          * pointers.  The function (ret_from_except) is actually a pointer
 314          * to the TOC entry.  The first entry is a pointer to the actual
 315          * function.
 316          */
 317         kregs->nip = *((unsigned long *)ret_from_fork);
 318
 319         return 0;
 320 }
 321
 322 /*
 323  * Set up a thread for executing a new program
 324  */
 325 void start_thread(struct pt_regs *regs, unsigned long fdptr, unsigned long sp)
 326 {
 327         unsigned long entry, toc, load_addr = regs->gpr[2];
 328
 329         /* fdptr is a relocated pointer to the function descriptor for
 330          * the elf _start routine.  The first entry in the function
 331          * descriptor is the entry address of _start and the second
 332          * entry is the TOC value we need to use.
 333          */
 334         set_fs(USER_DS);
 335         __get_user(entry, (unsigned long __user *)fdptr);
 336         __get_user(toc, (unsigned long __user *)fdptr+1);
 337
 338         /* Check whether the e_entry function descriptor entries
 339          * need to be relocated before we can use them.
 340          */
 341         if ( load_addr != 0 ) {
 342                 entry += load_addr;
 343                 toc   += load_addr;
 344         }
 345
 346         regs->nip = entry;
 347         regs->gpr[1] = sp;
 348         regs->gpr[2] = toc;
 349         regs->msr = MSR_USER64;
 350 #ifndef CONFIG_SMP
 351         if (last_task_used_math == current)
 352                 last_task_used_math = 0;
 353 #endif /* CONFIG_SMP */
 354         memset(current->thread.fpr, 0, sizeof(current->thread.fpr));
 355         current->thread.fpscr = 0;
 356 #ifdef CONFIG_ALTIVEC
 357 #ifndef CONFIG_SMP
 358         if (last_task_used_altivec == current)
 359                 last_task_used_altivec = 0;
 360 #endif /* CONFIG_SMP */
 361         memset(current->thread.vr, 0, sizeof(current->thread.vr));
 362         current->thread.vscr.u[0] = 0;
 363         current->thread.vscr.u[1] = 0;
 364         current->thread.vscr.u[2] = 0;
 365         current->thread.vscr.u[3] = 0x00010000; /* Java mode disabled */
 366         current->thread.vrsave = 0;
 367         current->thread.used_vr = 0;
 368 #endif /* CONFIG_ALTIVEC */
 369 }
 370
 371 int set_fpexc_mode(struct task_struct *tsk, unsigned int val)
 372 {
 373         struct pt_regs *regs = tsk->thread.regs;
 374
 375         if (val > PR_FP_EXC_PRECISE)
 376                 return -EINVAL;
 377         tsk->thread.fpexc_mode = __pack_fe01(val);
 378         if (regs != NULL && (regs->msr & MSR_FP) != 0)
 379                 regs->msr = (regs->msr & ~(MSR_FE0|MSR_FE1))
 380                         | tsk->thread.fpexc_mode;
 381         return 0;
 382 }
 383
 384 int get_fpexc_mode(struct task_struct *tsk, unsigned long adr)
 385 {
 386         unsigned int val;
 387
 388         val = __unpack_fe01(tsk->thread.fpexc_mode);
 389         return put_user(val, (unsigned int __user *) adr);
 390 }
 391
 392 int sys_clone(unsigned long clone_flags, unsigned long p2, unsigned long p3,
 393               unsigned long p4, unsigned long p5, unsigned long p6,
 394               struct pt_regs *regs)
 395 {
 396         unsigned long parent_tidptr = 0;
 397         unsigned long child_tidptr = 0;
 398
 399         if (p2 == 0)
 400                 p2 = regs->gpr[1];      /* stack pointer for child */
 401
 402         if (clone_flags & (CLONE_PARENT_SETTID | CLONE_CHILD_SETTID |
 403                            CLONE_CHILD_CLEARTID)) {
 404                 parent_tidptr = p3;
 405                 child_tidptr = p5;
 406                 if (test_thread_flag(TIF_32BIT)) {
 407                         parent_tidptr &= 0xffffffff;
 408                         child_tidptr &= 0xffffffff;
 409                 }
 410         }
 411
 412         return do_fork(clone_flags & ~CLONE_IDLETASK, p2, regs, 0,
 413                     (int __user *)parent_tidptr, (int __user *)child_tidptr);
 414 }
 415
 416 int sys_fork(unsigned long p1, unsigned long p2, unsigned long p3,
 417              unsigned long p4, unsigned long p5, unsigned long p6,
 418              struct pt_regs *regs)
 419 {
 420         return do_fork(SIGCHLD, regs->gpr[1], regs, 0, NULL, NULL);
 421 }
 422
 423 int sys_vfork(unsigned long p1, unsigned long p2, unsigned long p3,
 424               unsigned long p4, unsigned long p5, unsigned long p6,
 425               struct pt_regs *regs)
 426 {
 427         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->gpr[1], regs, 0,
 428                     NULL, NULL);
 429 }
 430
 431 int sys_execve(unsigned long a0, unsigned long a1, unsigned long a2,
 432                unsigned long a3, unsigned long a4, unsigned long a5,
 433                struct pt_regs *regs)
 434 {
 435         int error;
 436         char * filename;
 437
 438         filename = getname((char __user *) a0);
 439         error = PTR_ERR(filename);
 440         if (IS_ERR(filename))
 441                 goto out;
 442         if (regs->msr & MSR_FP)
 443                 giveup_fpu(current);
 444 #ifdef CONFIG_ALTIVEC
 445         if (regs->msr & MSR_VEC)
 446                 giveup_altivec(current);
 447 #endif /* CONFIG_ALTIVEC */
 448         error = do_execve(filename, (char __user * __user *) a1,
 449                                     (char __user * __user *) a2, regs);
 450
 451         if (error == 0)
 452                 current->ptrace &= ~PT_DTRACE;
 453         putname(filename);
 454
 455 out:
 456         return error;
 457 }
 458
 459 static int kstack_depth_to_print = 64;
 460
 461 static int validate_sp(unsigned long sp, struct task_struct *p,
 462                        unsigned long nbytes)
 463 {
 464         unsigned long stack_page = (unsigned long)p->thread_info;
 465
 466         if (sp >= stack_page + sizeof(struct thread_struct)
 467             && sp <= stack_page + THREAD_SIZE - nbytes)
 468                 return 1;
 469
 470 #ifdef CONFIG_IRQSTACKS
 471         stack_page = (unsigned long) hardirq_ctx[task_cpu(p)];
 472         if (sp >= stack_page + sizeof(struct thread_struct)
 473             && sp <= stack_page + THREAD_SIZE - nbytes)
 474                 return 1;
 475
 476         stack_page = (unsigned long) softirq_ctx[task_cpu(p)];
 477         if (sp >= stack_page + sizeof(struct thread_struct)
 478             && sp <= stack_page + THREAD_SIZE - nbytes)
 479                 return 1;
 480 #endif
 481
 482         return 0;
 483 }
 484
 485 unsigned long get_wchan(struct task_struct *p)
 486 {
 487         unsigned long ip, sp;
 488         int count = 0;
 489
 490         if (!p || p == current || p->state == TASK_RUNNING)
 491                 return 0;
 492
 493         sp = p->thread.ksp;
 494         if (!validate_sp(sp, p, 112))
 495                 return 0;
 496
 497         do {
 498                 sp = *(unsigned long *)sp;
 499                 if (!validate_sp(sp, p, 112))
 500                         return 0;
 501                 if (count > 0) {
 502                         ip = *(unsigned long *)(sp + 16);
 503                         if (!in_sched_functions(ip))
 504                                 return ip;
 505                 }
 506         } while (count++ < 16);
 507         return 0;
 508 }
 509
 510 void show_stack(struct task_struct *p, unsigned long *_sp)
 511 {
 512         unsigned long ip, newsp, lr;
 513         int count = 0;
 514         unsigned long sp = (unsigned long)_sp;
 515         int firstframe = 1;
 516
 517         if (sp == 0) {
 518                 if (p) {
 519                         sp = p->thread.ksp;
 520                 } else {
 521                         sp = __get_SP();
 522                         p = current;
 523                 }
 524         }
 525
 526         lr = 0;
 527         printk("Call Trace:\n");
 528         do {
 529                 if (!validate_sp(sp, p, 112))
 530                         return;
 531
 532                 _sp = (unsigned long *) sp;
 533                 newsp = _sp[0];
 534                 ip = _sp[2];
 535                 if (!firstframe || ip != lr) {
 536                         printk("[%016lx] [%016lx] ", sp, ip);
 537                         print_symbol("%s", ip);
 538                         if (firstframe)
 539                                 printk(" (unreliable)");
 540                         printk("\n");
 541                 }
 542                 firstframe = 0;
 543
 544                 /*
 545                  * See if this is an exception frame.
 546                  * We look for the "regshere" marker in the current frame.
 547                  */
 548                 if (validate_sp(sp, p, sizeof(struct pt_regs) + 400)
 549                     && _sp[12] == 0x7265677368657265ul) {
 550                         struct pt_regs *regs = (struct pt_regs *)
 551                                 (sp + STACK_FRAME_OVERHEAD);
 552                         printk("--- Exception: %lx", regs->trap);
 553                         print_symbol(" at %s\n", regs->nip);
 554                         lr = regs->link;
 555                         print_symbol("    LR = %s\n", lr);
 556                         firstframe = 1;
 557                 }
 558
 559                 sp = newsp;
 560         } while (count++ < kstack_depth_to_print);
 561 }
 562
 563 void dump_stack(void)
 564 {
 565         show_stack(current, (unsigned long *)__get_SP());
 566 }
 567 EXPORT_SYMBOL(dump_stack);