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 /*
  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 #ifdef CONFIG_ALTIVEC
 151         if (tsk->thread.regs) {
 152                 preempt_disable();
 153                 if (tsk->thread.regs->msr & MSR_VEC) {
 154 #ifdef CONFIG_SMP
 155                         BUG_ON(tsk != current);
 156 #endif
 157                         giveup_altivec(current);
 158                 }
 159                 preempt_enable();
 160         }
 161 #endif
 162 }
 163
 164 int dump_task_altivec(struct pt_regs *regs, elf_vrregset_t *vrregs)
 165 {
 166         flush_altivec_to_thread(current);
 167         memcpy(vrregs, &current->thread.vr[0], sizeof(*vrregs));
 168         return 1;
 169 }
 170
 171 #endif /* CONFIG_ALTIVEC */
 172
 173 struct task_struct *__switch_to(struct task_struct *prev,
 174                                 struct task_struct *new)
 175 {
 176         struct thread_struct *new_thread, *old_thread;
 177         unsigned long flags;
 178         struct task_struct *last;
 179
 180 #ifdef CONFIG_SMP
 181         /* avoid complexity of lazy save/restore of fpu
 182          * by just saving it every time we switch out if
 183          * this task used the fpu during the last quantum.
 184          *
 185          * If it tries to use the fpu again, it'll trap and
 186          * reload its fp regs.  So we don't have to do a restore
 187          * every switch, just a save.
 188          *  -- Cort
 189          */
 190         if (prev->thread.regs && (prev->thread.regs->msr & MSR_FP))
 191                 giveup_fpu(prev);
 192 #ifdef CONFIG_ALTIVEC
 193         if (prev->thread.regs && (prev->thread.regs->msr & MSR_VEC))
 194                 giveup_altivec(prev);
 195 #endif /* CONFIG_ALTIVEC */
 196 #endif /* CONFIG_SMP */
 197
 198 #if defined(CONFIG_ALTIVEC) && !defined(CONFIG_SMP)
 199         /* Avoid the trap.  On smp this this never happens since
 200          * we don't set last_task_used_altivec -- Cort
 201          */
 202         if (new->thread.regs && last_task_used_altivec == new)
 203                 new->thread.regs->msr |= MSR_VEC;
 204 #endif /* CONFIG_ALTIVEC */
 205
 206         flush_tlb_pending();
 207
 208         new_thread = &new->thread;
 209         old_thread = &current->thread;
 210
 211         local_irq_save(flags);
 212         last = _switch(old_thread, new_thread);
 213
 214         local_irq_restore(flags);
 215
 216         return last;
 217 }
 218
 219 void show_regs(struct pt_regs * regs)
 220 {
 221         int i;
 222         unsigned long trap;
 223
 224         printk("NIP: %016lX XER: %016lX LR: %016lX\n",
 225                regs->nip, regs->xer, regs->link);
 226         printk("REGS: %p TRAP: %04lx   %s  (%s)\n",
 227                regs, regs->trap, print_tainted(), UTS_RELEASE);
 228         printk("MSR: %016lx EE: %01x PR: %01x FP: %01x ME: %01x IR/DR: %01x%01x\n",
 229                regs->msr, regs->msr&MSR_EE ? 1 : 0, regs->msr&MSR_PR ? 1 : 0,
 230                regs->msr & MSR_FP ? 1 : 0,regs->msr&MSR_ME ? 1 : 0,
 231                regs->msr&MSR_IR ? 1 : 0,
 232                regs->msr&MSR_DR ? 1 : 0);
 233         trap = TRAP(regs);
 234         if (trap == 0x300 || trap == 0x380 || trap == 0x600)
 235                 printk("DAR: %016lx, DSISR: %016lx\n", regs->dar, regs->dsisr);
 236         printk("TASK: %p[%d] '%s' THREAD: %p",
 237                current, current->pid, current->comm, current->thread_info);
 238
 239 #ifdef CONFIG_SMP
 240         printk(" CPU: %d", smp_processor_id());
 241 #endif /* CONFIG_SMP */
 242
 243         for (i = 0; i < 32; i++) {
 244                 if ((i % 4) == 0) {
 245                         printk("\n" KERN_INFO "GPR%02d: ", i);
 246                 }
 247
 248                 printk("%016lX ", regs->gpr[i]);
 249                 if (i == 13 && !FULL_REGS(regs))
 250                         break;
 251         }
 252         printk("\n");
 253         /*
 254          * Lookup NIP late so we have the best change of getting the
 255          * above info out without failing
 256          */
 257         printk("NIP [%016lx] ", regs->nip);
 258         print_symbol("%s\n", regs->nip);
 259         printk("LR [%016lx] ", regs->link);
 260         print_symbol("%s\n", regs->link);
 261         show_stack(current, (unsigned long *)regs->gpr[1]);
 262 }
 263
 264 void exit_thread(void)
 265 {
 266 #ifndef CONFIG_SMP
 267         if (last_task_used_math == current)
 268                 last_task_used_math = NULL;
 269 #ifdef CONFIG_ALTIVEC
 270         if (last_task_used_altivec == current)
 271                 last_task_used_altivec = NULL;
 272 #endif /* CONFIG_ALTIVEC */
 273 #endif /* CONFIG_SMP */
 274 }
 275
 276 void flush_thread(void)
 277 {
 278         struct thread_info *t = current_thread_info();
 279
 280         if (t->flags & _TIF_ABI_PENDING)
 281                 t->flags ^= (_TIF_ABI_PENDING | _TIF_32BIT);
 282
 283 #ifndef CONFIG_SMP
 284         if (last_task_used_math == current)
 285                 last_task_used_math = NULL;
 286 #ifdef CONFIG_ALTIVEC
 287         if (last_task_used_altivec == current)
 288                 last_task_used_altivec = NULL;
 289 #endif /* CONFIG_ALTIVEC */
 290 #endif /* CONFIG_SMP */
 291 }
 292
 293 void
 294 release_thread(struct task_struct *t)
 295 {
 296 }
 297
 298
 299 /*
 300  * This gets called before we allocate a new thread and copy
 301  * the current task into it.
 302  */
 303 void prepare_to_copy(struct task_struct *tsk)
 304 {
 305         flush_fp_to_thread(current);
 306         flush_altivec_to_thread(current);
 307 }
 308
 309 /*
 310  * Copy a thread..
 311  */
 312 int
 313 copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 314             unsigned long unused, struct task_struct *p, struct pt_regs *regs)
 315 {
 316         struct pt_regs *childregs, *kregs;
 317         extern void ret_from_fork(void);
 318         unsigned long sp = (unsigned long)p->thread_info + THREAD_SIZE;
 319
 320         p->set_child_tid = p->clear_child_tid = NULL;
 321
 322         /* Copy registers */
 323         sp -= sizeof(struct pt_regs);
 324         childregs = (struct pt_regs *) sp;
 325         *childregs = *regs;
 326         if ((childregs->msr & MSR_PR) == 0) {
 327                 /* for kernel thread, set stackptr in new task */
 328                 childregs->gpr[1] = sp + sizeof(struct pt_regs);
 329                 p->thread.regs = NULL;  /* no user register state */
 330                 clear_ti_thread_flag(p->thread_info, TIF_32BIT);
 331 #ifdef CONFIG_PPC_ISERIES
 332                 set_ti_thread_flag(p->thread_info, TIF_RUN_LIGHT);
 333 #endif
 334         } else {
 335                 childregs->gpr[1] = usp;
 336                 p->thread.regs = childregs;
 337                 if (clone_flags & CLONE_SETTLS) {
 338                         if (test_thread_flag(TIF_32BIT))
 339                                 childregs->gpr[2] = childregs->gpr[6];
 340                         else
 341                                 childregs->gpr[13] = childregs->gpr[6];
 342                 }
 343         }
 344         childregs->gpr[3] = 0;  /* Result from fork() */
 345         sp -= STACK_FRAME_OVERHEAD;
 346
 347         /*
 348          * The way this works is that at some point in the future
 349          * some task will call _switch to switch to the new task.
 350          * That will pop off the stack frame created below and start
 351          * the new task running at ret_from_fork.  The new task will
 352          * do some house keeping and then return from the fork or clone
 353          * system call, using the stack frame created above.
 354          */
 355         sp -= sizeof(struct pt_regs);
 356         kregs = (struct pt_regs *) sp;
 357         sp -= STACK_FRAME_OVERHEAD;
 358         p->thread.ksp = sp;
 359
 360         /*
 361          * The PPC64 ABI makes use of a TOC to contain function
 362          * pointers.  The function (ret_from_except) is actually a pointer
 363          * to the TOC entry.  The first entry is a pointer to the actual
 364          * function.
 365          */
 366         kregs->nip = *((unsigned long *)ret_from_fork);
 367
 368         return 0;
 369 }
 370
 371 /*
 372  * Set up a thread for executing a new program
 373  */
 374 void start_thread(struct pt_regs *regs, unsigned long fdptr, unsigned long sp)
 375 {
 376         unsigned long entry, toc, load_addr = regs->gpr[2];
 377
 378         /* fdptr is a relocated pointer to the function descriptor for
 379          * the elf _start routine.  The first entry in the function
 380          * descriptor is the entry address of _start and the second
 381          * entry is the TOC value we need to use.
 382          */
 383         set_fs(USER_DS);
 384         __get_user(entry, (unsigned long __user *)fdptr);
 385         __get_user(toc, (unsigned long __user *)fdptr+1);
 386
 387         /* Check whether the e_entry function descriptor entries
 388          * need to be relocated before we can use them.
 389          */
 390         if ( load_addr != 0 ) {
 391                 entry += load_addr;
 392                 toc   += load_addr;
 393         }
 394
 395         regs->nip = entry;
 396         regs->gpr[1] = sp;
 397         regs->gpr[2] = toc;
 398         regs->msr = MSR_USER64;
 399 #ifndef CONFIG_SMP
 400         if (last_task_used_math == current)
 401                 last_task_used_math = 0;
 402 #endif /* CONFIG_SMP */
 403         memset(current->thread.fpr, 0, sizeof(current->thread.fpr));
 404         current->thread.fpscr = 0;
 405 #ifdef CONFIG_ALTIVEC
 406 #ifndef CONFIG_SMP
 407         if (last_task_used_altivec == current)
 408                 last_task_used_altivec = 0;
 409 #endif /* CONFIG_SMP */
 410         memset(current->thread.vr, 0, sizeof(current->thread.vr));
 411         current->thread.vscr.u[0] = 0;
 412         current->thread.vscr.u[1] = 0;
 413         current->thread.vscr.u[2] = 0;
 414         current->thread.vscr.u[3] = 0x00010000; /* Java mode disabled */
 415         current->thread.vrsave = 0;
 416         current->thread.used_vr = 0;
 417 #endif /* CONFIG_ALTIVEC */
 418 }
 419
 420 int set_fpexc_mode(struct task_struct *tsk, unsigned int val)
 421 {
 422         struct pt_regs *regs = tsk->thread.regs;
 423
 424         if (val > PR_FP_EXC_PRECISE)
 425                 return -EINVAL;
 426         tsk->thread.fpexc_mode = __pack_fe01(val);
 427         if (regs != NULL && (regs->msr & MSR_FP) != 0)
 428                 regs->msr = (regs->msr & ~(MSR_FE0|MSR_FE1))
 429                         | tsk->thread.fpexc_mode;
 430         return 0;
 431 }
 432
 433 int get_fpexc_mode(struct task_struct *tsk, unsigned long adr)
 434 {
 435         unsigned int val;
 436
 437         val = __unpack_fe01(tsk->thread.fpexc_mode);
 438         return put_user(val, (unsigned int __user *) adr);
 439 }
 440
 441 int sys_clone(unsigned long clone_flags, unsigned long p2, unsigned long p3,
 442               unsigned long p4, unsigned long p5, unsigned long p6,
 443               struct pt_regs *regs)
 444 {
 445         unsigned long parent_tidptr = 0;
 446         unsigned long child_tidptr = 0;
 447
 448         if (p2 == 0)
 449                 p2 = regs->gpr[1];      /* stack pointer for child */
 450
 451         if (clone_flags & (CLONE_PARENT_SETTID | CLONE_CHILD_SETTID |
 452                            CLONE_CHILD_CLEARTID)) {
 453                 parent_tidptr = p3;
 454                 child_tidptr = p5;
 455                 if (test_thread_flag(TIF_32BIT)) {
 456                         parent_tidptr &= 0xffffffff;
 457                         child_tidptr &= 0xffffffff;
 458                 }
 459         }
 460
 461         return do_fork(clone_flags & ~CLONE_IDLETASK, p2, regs, 0,
 462                     (int __user *)parent_tidptr, (int __user *)child_tidptr);
 463 }
 464
 465 int sys_fork(unsigned long p1, unsigned long p2, unsigned long p3,
 466              unsigned long p4, unsigned long p5, unsigned long p6,
 467              struct pt_regs *regs)
 468 {
 469         return do_fork(SIGCHLD, regs->gpr[1], regs, 0, NULL, NULL);
 470 }
 471
 472 int sys_vfork(unsigned long p1, unsigned long p2, unsigned long p3,
 473               unsigned long p4, unsigned long p5, unsigned long p6,
 474               struct pt_regs *regs)
 475 {
 476         return do_fork(CLONE_VFORK | CLONE_VM | SIGCHLD, regs->gpr[1], regs, 0,
 477                     NULL, NULL);
 478 }
 479
 480 int sys_execve(unsigned long a0, unsigned long a1, unsigned long a2,
 481                unsigned long a3, unsigned long a4, unsigned long a5,
 482                struct pt_regs *regs)
 483 {
 484         int error;
 485         char * filename;
 486
 487         filename = getname((char __user *) a0);
 488         error = PTR_ERR(filename);
 489         if (IS_ERR(filename))
 490                 goto out;
 491         flush_fp_to_thread(current);
 492         flush_altivec_to_thread(current);
 493         error = do_execve(filename, (char __user * __user *) a1,
 494                                     (char __user * __user *) a2, regs);
 495
 496         if (error == 0)
 497                 current->ptrace &= ~PT_DTRACE;
 498         putname(filename);
 499
 500 out:
 501         return error;
 502 }
 503
 504 static int kstack_depth_to_print = 64;
 505
 506 static int validate_sp(unsigned long sp, struct task_struct *p,
 507                        unsigned long nbytes)
 508 {
 509         unsigned long stack_page = (unsigned long)p->thread_info;
 510
 511         if (sp >= stack_page + sizeof(struct thread_struct)
 512             && sp <= stack_page + THREAD_SIZE - nbytes)
 513                 return 1;
 514
 515 #ifdef CONFIG_IRQSTACKS
 516         stack_page = (unsigned long) hardirq_ctx[task_cpu(p)];
 517         if (sp >= stack_page + sizeof(struct thread_struct)
 518             && sp <= stack_page + THREAD_SIZE - nbytes)
 519                 return 1;
 520
 521         stack_page = (unsigned long) softirq_ctx[task_cpu(p)];
 522         if (sp >= stack_page + sizeof(struct thread_struct)
 523             && sp <= stack_page + THREAD_SIZE - nbytes)
 524                 return 1;
 525 #endif
 526
 527         return 0;
 528 }
 529
 530 unsigned long get_wchan(struct task_struct *p)
 531 {
 532         unsigned long ip, sp;
 533         int count = 0;
 534
 535         if (!p || p == current || p->state == TASK_RUNNING)
 536                 return 0;
 537
 538         sp = p->thread.ksp;
 539         if (!validate_sp(sp, p, 112))
 540                 return 0;
 541
 542         do {
 543                 sp = *(unsigned long *)sp;
 544                 if (!validate_sp(sp, p, 112))
 545                         return 0;
 546                 if (count > 0) {
 547                         ip = *(unsigned long *)(sp + 16);
 548                         if (!in_sched_functions(ip))
 549                                 return ip;
 550                 }
 551         } while (count++ < 16);
 552         return 0;
 553 }
 554
 555 void show_stack(struct task_struct *p, unsigned long *_sp)
 556 {
 557         unsigned long ip, newsp, lr;
 558         int count = 0;
 559         unsigned long sp = (unsigned long)_sp;
 560         int firstframe = 1;
 561
 562         if (sp == 0) {
 563                 if (p) {
 564                         sp = p->thread.ksp;
 565                 } else {
 566                         sp = __get_SP();
 567                         p = current;
 568                 }
 569         }
 570
 571         lr = 0;
 572         printk("Call Trace:\n");
 573         do {
 574                 if (!validate_sp(sp, p, 112))
 575                         return;
 576
 577                 _sp = (unsigned long *) sp;
 578                 newsp = _sp[0];
 579                 ip = _sp[2];
 580                 if (!firstframe || ip != lr) {
 581                         printk("[%016lx] [%016lx] ", sp, ip);
 582                         print_symbol("%s", ip);
 583                         if (firstframe)
 584                                 printk(" (unreliable)");
 585                         printk("\n");
 586                 }
 587                 firstframe = 0;
 588
 589                 /*
 590                  * See if this is an exception frame.
 591                  * We look for the "regshere" marker in the current frame.
 592                  */
 593                 if (validate_sp(sp, p, sizeof(struct pt_regs) + 400)
 594                     && _sp[12] == 0x7265677368657265ul) {
 595                         struct pt_regs *regs = (struct pt_regs *)
 596                                 (sp + STACK_FRAME_OVERHEAD);
 597                         printk("--- Exception: %lx", regs->trap);
 598                         print_symbol(" at %s\n", regs->nip);
 599                         lr = regs->link;
 600                         print_symbol("    LR = %s\n", lr);
 601                         firstframe = 1;
 602                 }
 603
 604                 sp = newsp;
 605         } while (count++ < kstack_depth_to_print);
 606 }
 607
 608 void dump_stack(void)
 609 {
 610         show_stack(current, (unsigned long *)__get_SP());
 611 }
 612 EXPORT_SYMBOL(dump_stack);