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