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
 262 void exit_thread(void)
 263 {
 264 #ifndef CONFIG_SMP
 265         if (last_task_used_math == current)
 266                 last_task_used_math = NULL;
 267 #ifdef CONFIG_ALTIVEC
 268         if (last_task_used_altivec == current)
 269                 last_task_used_altivec = NULL;
 270 #endif /* CONFIG_ALTIVEC */
 271 #endif /* CONFIG_SMP */
 272 }
 273
 274 void flush_thread(void)
 275 {
 276         struct thread_info *t = current_thread_info();
 277
 278         if (t->flags & _TIF_ABI_PENDING)
 279                 t->flags ^= (_TIF_ABI_PENDING | _TIF_32BIT);
 280
 281 #ifndef CONFIG_SMP
 282         if (last_task_used_math == current)
 283                 last_task_used_math = NULL;
 284 #ifdef CONFIG_ALTIVEC
 285         if (last_task_used_altivec == current)
 286                 last_task_used_altivec = NULL;
 287 #endif /* CONFIG_ALTIVEC */
 288 #endif /* CONFIG_SMP */
 289 }
 290
 291 void
 292 release_thread(struct task_struct *t)
 293 {
 294 }
 295
 296
 297 /*
 298  * This gets called before we allocate a new thread and copy
 299  * the current task into it.
 300  */
 301 void prepare_to_copy(struct task_struct *tsk)
 302 {
 303         flush_fp_to_thread(current);
 304         flush_altivec_to_thread(current);
 305 }
 306
 307 /*
 308  * Copy a thread..
 309  */
 310 int
 311 copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 312             unsigned long unused, struct task_struct *p, struct pt_regs *regs)
 313 {
 314         struct pt_regs *childregs, *kregs;
 315         extern void ret_from_fork(void);
 316         unsigned long sp = (unsigned long)p->thread_info + THREAD_SIZE;
 317
 318         p->set_child_tid = p->clear_child_tid = NULL;
 319
 320         /* Copy registers */
 321         sp -= sizeof(struct pt_regs);
 322         childregs = (struct pt_regs *) sp;
 323         *childregs = *regs;
 324         if ((childregs->msr & MSR_PR) == 0) {
 325                 /* for kernel thread, set stackptr in new task */
 326                 childregs->gpr[1] = sp + sizeof(struct pt_regs);
 327                 p->thread.regs = NULL;  /* no user register state */
 328                 clear_ti_thread_flag(p->thread_info, TIF_32BIT);
 329 #ifdef CONFIG_PPC_ISERIES
 330                 set_ti_thread_flag(p->thread_info, TIF_RUN_LIGHT);
 331 #endif
 332         } else {
 333                 childregs->gpr[1] = usp;
 334                 p->thread.regs = childregs;
 335                 if (clone_flags & CLONE_SETTLS) {
 336                         if (test_thread_flag(TIF_32BIT))
 337                                 childregs->gpr[2] = childregs->gpr[6];
 338                         else
 339                                 childregs->gpr[13] = childregs->gpr[6];
 340                 }
 341         }
 342         childregs->gpr[3] = 0;  /* Result from fork() */
 343         sp -= STACK_FRAME_OVERHEAD;
 344
 345         /*
 346          * The way this works is that at some point in the future
 347          * some task will call _switch to switch to the new task.
 348          * That will pop off the stack frame created below and start
 349          * the new task running at ret_from_fork.  The new task will
 350          * do some house keeping and then return from the fork or clone
 351          * system call, using the stack frame created above.
 352          */
 353         sp -= sizeof(struct pt_regs);
 354         kregs = (struct pt_regs *) sp;
 355         sp -= STACK_FRAME_OVERHEAD;
 356         p->thread.ksp = sp;
 357         if (cur_cpu_spec->cpu_features & CPU_FTR_SLB) {
 358                 unsigned long sp_vsid = get_kernel_vsid(sp);
 359
 360                 sp_vsid <<= SLB_VSID_SHIFT;
 361                 sp_vsid |= SLB_VSID_KERNEL;
 362                 if (cur_cpu_spec->cpu_features & CPU_FTR_16M_PAGE)
 363                         sp_vsid |= SLB_VSID_L;
 364
 365                 p->thread.ksp_vsid = sp_vsid;
 366         }
 367
 368         /*
 369          * The PPC64 ABI makes use of a TOC to contain function
 370          * pointers.  The function (ret_from_except) is actually a pointer
 371          * to the TOC entry.  The first entry is a pointer to the actual
 372          * function.
 373          */
 374         kregs->nip = *((unsigned long *)ret_from_fork);
 375
 376         return 0;
 377 }
 378
 379 /*
 380  * Set up a thread for executing a new program
 381  */
 382 void start_thread(struct pt_regs *regs, unsigned long fdptr, unsigned long sp)
 383 {
 384         unsigned long entry, toc, load_addr = regs->gpr[2];
 385
 386         /* fdptr is a relocated pointer to the function descriptor for
 387          * the elf _start routine.  The first entry in the function
 388          * descriptor is the entry address of _start and the second
 389          * entry is the TOC value we need to use.
 390          */
 391         set_fs(USER_DS);
 392         __get_user(entry, (unsigned long __user *)fdptr);
 393         __get_user(toc, (unsigned long __user *)fdptr+1);
 394
 395         /* Check whether the e_entry function descriptor entries
 396          * need to be relocated before we can use them.
 397          */
 398         if (load_addr != 0) {
 399                 entry += load_addr;
 400                 toc   += load_addr;
 401         }
 402
 403         /*
 404          * If we exec out of a kernel thread then thread.regs will not be
 405          * set. Do it now.
 406          */
 407         if (!current->thread.regs) {
 408                 unsigned long childregs = (unsigned long)current->thread_info +
 409                                                 THREAD_SIZE;
 410                 childregs -= sizeof(struct pt_regs);
 411                 current->thread.regs = (struct pt_regs *)childregs;
 412         }
 413
 414         regs->nip = entry;
 415         regs->gpr[1] = sp;
 416         regs->gpr[2] = toc;
 417         regs->msr = MSR_USER64;
 418 #ifndef CONFIG_SMP
 419         if (last_task_used_math == current)
 420                 last_task_used_math = 0;
 421 #endif /* CONFIG_SMP */
 422         memset(current->thread.fpr, 0, sizeof(current->thread.fpr));
 423         current->thread.fpscr = 0;
 424 #ifdef CONFIG_ALTIVEC
 425 #ifndef CONFIG_SMP
 426         if (last_task_used_altivec == current)
 427                 last_task_used_altivec = 0;
 428 #endif /* CONFIG_SMP */
 429         memset(current->thread.vr, 0, sizeof(current->thread.vr));
 430         current->thread.vscr.u[0] = 0;
 431         current->thread.vscr.u[1] = 0;
 432         current->thread.vscr.u[2] = 0;
 433         current->thread.vscr.u[3] = 0x00010000; /* Java mode disabled */
 434         current->thread.vrsave = 0;
 435         current->thread.used_vr = 0;
 436 #endif /* CONFIG_ALTIVEC */
 437 }
 438
 439 int set_fpexc_mode(struct task_struct *tsk, unsigned int val)
 440 {
 441         struct pt_regs *regs = tsk->thread.regs;
 442
 443         if (val > PR_FP_EXC_PRECISE)
 444                 return -EINVAL;
 445         tsk->thread.fpexc_mode = __pack_fe01(val);
 446         if (regs != NULL && (regs->msr & MSR_FP) != 0)
 447                 regs->msr = (regs->msr & ~(MSR_FE0|MSR_FE1))
 448                         | tsk->thread.fpexc_mode;
 449         return 0;
 450 }
 451
 452 int get_fpexc_mode(struct task_struct *tsk, unsigned long adr)
 453 {
 454         unsigned int val;
 455
 456         val = __unpack_fe01(tsk->thread.fpexc_mode);
 457         return put_user(val, (unsigned int __user *) adr);
 458 }
 459
 460 int sys_clone(unsigned long clone_flags, unsigned long p2, unsigned long p3,
 461               unsigned long p4, unsigned long p5, unsigned long p6,
 462               struct pt_regs *regs)
 463 {
 464         unsigned long parent_tidptr = 0;
 465         unsigned long child_tidptr = 0;
 466
 467         if (p2 == 0)
 468                 p2 = regs->gpr[1];      /* stack pointer for child */
 469
 470         if (clone_flags & (CLONE_PARENT_SETTID | CLONE_CHILD_SETTID |
 471                            CLONE_CHILD_CLEARTID)) {
 472                 parent_tidptr = p3;
 473                 child_tidptr = p5;
 474                 if (test_thread_flag(TIF_32BIT)) {
 475                         parent_tidptr &= 0xffffffff;
 476                         child_tidptr &= 0xffffffff;
 477                 }
 478         }
 479
 480         return do_fork(clone_flags, p2, regs, 0,
 481                     (int __user *)parent_tidptr, (int __user *)child_tidptr);
 482 }
 483
 484 int sys_fork(unsigned long p1, unsigned long p2, unsigned long p3,
 485              unsigned long p4, unsigned long p5, unsigned long p6,
 486              struct pt_regs *regs)
 487 {
 488         return do_fork(SIGCHLD, regs->gpr[1], regs, 0, NULL, NULL);
 489 }
 490
 491 int sys_vfork(unsigned long p1, unsigned long p2, unsigned long p3,
 492               unsigned long p4, unsigned long p5, unsigned long p6,
 493               struct pt_regs *regs)
 494 {
 495         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->gpr[1], regs, 0,
 496                     NULL, NULL);
 497 }
 498
 499 int sys_execve(unsigned long a0, unsigned long a1, unsigned long a2,
 500                unsigned long a3, unsigned long a4, unsigned long a5,
 501                struct pt_regs *regs)
 502 {
 503         int error;
 504         char * filename;
 505
 506         filename = getname((char __user *) a0);
 507         error = PTR_ERR(filename);
 508         if (IS_ERR(filename))
 509                 goto out;
 510         flush_fp_to_thread(current);
 511         flush_altivec_to_thread(current);
 512         error = do_execve(filename, (char __user * __user *) a1,
 513                                     (char __user * __user *) a2, regs);
 514
 515         if (error == 0)
 516                 current->ptrace &= ~PT_DTRACE;
 517         putname(filename);
 518
 519 out:
 520         return error;
 521 }
 522
 523 static int kstack_depth_to_print = 64;
 524
 525 static int validate_sp(unsigned long sp, struct task_struct *p,
 526                        unsigned long nbytes)
 527 {
 528         unsigned long stack_page = (unsigned long)p->thread_info;
 529
 530         if (sp >= stack_page + sizeof(struct thread_struct)
 531             && sp <= stack_page + THREAD_SIZE - nbytes)
 532                 return 1;
 533
 534 #ifdef CONFIG_IRQSTACKS
 535         stack_page = (unsigned long) hardirq_ctx[task_cpu(p)];
 536         if (sp >= stack_page + sizeof(struct thread_struct)
 537             && sp <= stack_page + THREAD_SIZE - nbytes)
 538                 return 1;
 539
 540         stack_page = (unsigned long) softirq_ctx[task_cpu(p)];
 541         if (sp >= stack_page + sizeof(struct thread_struct)
 542             && sp <= stack_page + THREAD_SIZE - nbytes)
 543                 return 1;
 544 #endif
 545
 546         return 0;
 547 }
 548
 549 unsigned long get_wchan(struct task_struct *p)
 550 {
 551         unsigned long ip, sp;
 552         int count = 0;
 553
 554         if (!p || p == current || p->state == TASK_RUNNING)
 555                 return 0;
 556
 557         sp = p->thread.ksp;
 558         if (!validate_sp(sp, p, 112))
 559                 return 0;
 560
 561         do {
 562                 sp = *(unsigned long *)sp;
 563                 if (!validate_sp(sp, p, 112))
 564                         return 0;
 565                 if (count > 0) {
 566                         ip = *(unsigned long *)(sp + 16);
 567                         if (!in_sched_functions(ip))
 568                                 return ip;
 569                 }
 570         } while (count++ < 16);
 571         return 0;
 572 }
 573
 574 void show_stack(struct task_struct *p, unsigned long *_sp)
 575 {
 576         unsigned long ip, newsp, lr;
 577         int count = 0;
 578         unsigned long sp = (unsigned long)_sp;
 579         int firstframe = 1;
 580
 581         if (sp == 0) {
 582                 if (p) {
 583                         sp = p->thread.ksp;
 584                 } else {
 585                         sp = __get_SP();
 586                         p = current;
 587                 }
 588         }
 589
 590         lr = 0;
 591         printk("Call Trace:\n");
 592         do {
 593                 if (!validate_sp(sp, p, 112))
 594                         return;
 595
 596                 _sp = (unsigned long *) sp;
 597                 newsp = _sp[0];
 598                 ip = _sp[2];
 599                 if (!firstframe || ip != lr) {
 600                         printk("[%016lx] [%016lx] ", sp, ip);
 601                         print_symbol("%s", ip);
 602                         if (firstframe)
 603                                 printk(" (unreliable)");
 604                         printk("\n");
 605                 }
 606                 firstframe = 0;
 607
 608                 /*
 609                  * See if this is an exception frame.
 610                  * We look for the "regshere" marker in the current frame.
 611                  */
 612                 if (validate_sp(sp, p, sizeof(struct pt_regs) + 400)
 613                     && _sp[12] == 0x7265677368657265ul) {
 614                         struct pt_regs *regs = (struct pt_regs *)
 615                                 (sp + STACK_FRAME_OVERHEAD);
 616                         printk("--- Exception: %lx", regs->trap);
 617                         print_symbol(" at %s\n", regs->nip);
 618                         lr = regs->link;
 619                         print_symbol("    LR = %s\n", lr);
 620                         firstframe = 1;
 621                 }
 622
 623                 sp = newsp;
 624         } while (count++ < kstack_depth_to_print);
 625 }
 626
 627 void dump_stack(void)
 628 {
 629         show_stack(current, (unsigned long *)__get_SP());
 630 }
 631 EXPORT_SYMBOL(dump_stack);