arch/tile/mm/fault.c

   1 /*
   2  * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3  *
   4  *   This program is free software; you can redistribute it and/or
   5  *   modify it under the terms of the GNU General Public License
   6  *   as published by the Free Software Foundation, version 2.
   7  *
   8  *   This program is distributed in the hope that it will be useful, but
   9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11  *   NON INFRINGEMENT.  See the GNU General Public License for
  12  *   more details.
  13  *
  14  * From i386 code copyright (C) 1995  Linus Torvalds
  15  */
  16
  17 #include <linux/signal.h>
  18 #include <linux/sched.h>
  19 #include <linux/kernel.h>
  20 #include <linux/errno.h>
  21 #include <linux/string.h>
  22 #include <linux/types.h>
  23 #include <linux/ptrace.h>
  24 #include <linux/mman.h>
  25 #include <linux/mm.h>
  26 #include <linux/smp.h>
  27 #include <linux/interrupt.h>
  28 #include <linux/init.h>
  29 #include <linux/tty.h>
  30 #include <linux/vt_kern.h>              /* For unblank_screen() */
  31 #include <linux/highmem.h>
  32 #include <linux/module.h>
  33 #include <linux/kprobes.h>
  34 #include <linux/hugetlb.h>
  35 #include <linux/syscalls.h>
  36 #include <linux/uaccess.h>
  37
  38 #include <asm/pgalloc.h>
  39 #include <asm/sections.h>
  40 #include <asm/traps.h>
  41 #include <asm/syscalls.h>
  42
  43 #include <arch/interrupts.h>
  44
  45 static noinline void force_sig_info_fault(const char *type, int si_signo,
  46                                           int si_code, unsigned long address,
  47                                           int fault_num,
  48                                           struct task_struct *tsk,
  49                                           struct pt_regs *regs)
  50 {
  51         siginfo_t info;
  52
  53         if (unlikely(tsk->pid < 2)) {
  54                 panic("Signal %d (code %d) at %#lx sent to %s!",
  55                       si_signo, si_code & 0xffff, address,
  56                       is_idle_task(tsk) ? "the idle task" : "init");
  57         }
  58
  59         info.si_signo = si_signo;
  60         info.si_errno = 0;
  61         info.si_code = si_code;
  62         info.si_addr = (void __user *)address;
  63         info.si_trapno = fault_num;
  64         trace_unhandled_signal(type, regs, address, si_signo);
  65         force_sig_info(si_signo, &info, tsk);
  66 }
  67
  68 #ifndef __tilegx__
  69 /*
  70  * Synthesize the fault a PL0 process would get by doing a word-load of
  71  * an unaligned address or a high kernel address.
  72  */
  73 SYSCALL_DEFINE1(cmpxchg_badaddr, unsigned long, address)
  74 {
  75         struct pt_regs *regs = current_pt_regs();
  76
  77         if (address >= PAGE_OFFSET)
  78                 force_sig_info_fault("atomic segfault", SIGSEGV, SEGV_MAPERR,
  79                                      address, INT_DTLB_MISS, current, regs);
  80         else
  81                 force_sig_info_fault("atomic alignment fault", SIGBUS,
  82                                      BUS_ADRALN, address,
  83                                      INT_UNALIGN_DATA, current, regs);
  84
  85         /*
  86          * Adjust pc to point at the actual instruction, which is unusual
  87          * for syscalls normally, but is appropriate when we are claiming
  88          * that a syscall swint1 caused a page fault or bus error.
  89          */
  90         regs->pc -= 8;
  91
  92         /*
  93          * Mark this as a caller-save interrupt, like a normal page fault,
  94          * so that when we go through the signal handler path we will
  95          * properly restore r0, r1, and r2 for the signal handler arguments.
  96          */
  97         regs->flags |= PT_FLAGS_CALLER_SAVES;
  98
  99         return 0;
 100 }
 101 #endif
 102
 103 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 104 {
 105         unsigned index = pgd_index(address);
 106         pgd_t *pgd_k;
 107         pud_t *pud, *pud_k;
 108         pmd_t *pmd, *pmd_k;
 109
 110         pgd += index;
 111         pgd_k = init_mm.pgd + index;
 112
 113         if (!pgd_present(*pgd_k))
 114                 return NULL;
 115
 116         pud = pud_offset(pgd, address);
 117         pud_k = pud_offset(pgd_k, address);
 118         if (!pud_present(*pud_k))
 119                 return NULL;
 120
 121         pmd = pmd_offset(pud, address);
 122         pmd_k = pmd_offset(pud_k, address);
 123         if (!pmd_present(*pmd_k))
 124                 return NULL;
 125         if (!pmd_present(*pmd)) {
 126                 set_pmd(pmd, *pmd_k);
 127                 arch_flush_lazy_mmu_mode();
 128         } else
 129                 BUG_ON(pmd_ptfn(*pmd) != pmd_ptfn(*pmd_k));
 130         return pmd_k;
 131 }
 132
 133 /*
 134  * Handle a fault on the vmalloc area.
 135  */
 136 static inline int vmalloc_fault(pgd_t *pgd, unsigned long address)
 137 {
 138         pmd_t *pmd_k;
 139         pte_t *pte_k;
 140
 141         /* Make sure we are in vmalloc area */
 142         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 143                 return -1;
 144
 145         /*
 146          * Synchronize this task's top level page-table
 147          * with the 'reference' page table.
 148          */
 149         pmd_k = vmalloc_sync_one(pgd, address);
 150         if (!pmd_k)
 151                 return -1;
 152         if (pmd_huge(*pmd_k))
 153                 return 0;   /* support TILE huge_vmap() API */
 154         pte_k = pte_offset_kernel(pmd_k, address);
 155         if (!pte_present(*pte_k))
 156                 return -1;
 157         return 0;
 158 }
 159
 160 /* Wait until this PTE has completed migration. */
 161 static void wait_for_migration(pte_t *pte)
 162 {
 163         if (pte_migrating(*pte)) {
 164                 /*
 165                  * Wait until the migrater fixes up this pte.
 166                  * We scale the loop count by the clock rate so we'll wait for
 167                  * a few seconds here.
 168                  */
 169                 int retries = 0;
 170                 int bound = get_clock_rate();
 171                 while (pte_migrating(*pte)) {
 172                         barrier();
 173                         if (++retries > bound)
 174                                 panic("Hit migrating PTE (%#llx) and"
 175                                       " page PFN %#lx still migrating",
 176                                       pte->val, pte_pfn(*pte));
 177                 }
 178         }
 179 }
 180
 181 /*
 182  * It's not generally safe to use "current" to get the page table pointer,
 183  * since we might be running an oprofile interrupt in the middle of a
 184  * task switch.
 185  */
 186 static pgd_t *get_current_pgd(void)
 187 {
 188         HV_Context ctx = hv_inquire_context();
 189         unsigned long pgd_pfn = ctx.page_table >> PAGE_SHIFT;
 190         struct page *pgd_page = pfn_to_page(pgd_pfn);
 191         BUG_ON(PageHighMem(pgd_page));
 192         return (pgd_t *) __va(ctx.page_table);
 193 }
 194
 195 /*
 196  * We can receive a page fault from a migrating PTE at any time.
 197  * Handle it by just waiting until the fault resolves.
 198  *
 199  * It's also possible to get a migrating kernel PTE that resolves
 200  * itself during the downcall from hypervisor to Linux.  We just check
 201  * here to see if the PTE seems valid, and if so we retry it.
 202  *
 203  * NOTE! We MUST NOT take any locks for this case.  We may be in an
 204  * interrupt or a critical region, and must do as little as possible.
 205  * Similarly, we can't use atomic ops here, since we may be handling a
 206  * fault caused by an atomic op access.
 207  *
 208  * If we find a migrating PTE while we're in an NMI context, and we're
 209  * at a PC that has a registered exception handler, we don't wait,
 210  * since this thread may (e.g.) have been interrupted while migrating
 211  * its own stack, which would then cause us to self-deadlock.
 212  */
 213 static int handle_migrating_pte(pgd_t *pgd, int fault_num,
 214                                 unsigned long address, unsigned long pc,
 215                                 int is_kernel_mode, int write)
 216 {
 217         pud_t *pud;
 218         pmd_t *pmd;
 219         pte_t *pte;
 220         pte_t pteval;
 221
 222         if (pgd_addr_invalid(address))
 223                 return 0;
 224
 225         pgd += pgd_index(address);
 226         pud = pud_offset(pgd, address);
 227         if (!pud || !pud_present(*pud))
 228                 return 0;
 229         pmd = pmd_offset(pud, address);
 230         if (!pmd || !pmd_present(*pmd))
 231                 return 0;
 232         pte = pmd_huge_page(*pmd) ? ((pte_t *)pmd) :
 233                 pte_offset_kernel(pmd, address);
 234         pteval = *pte;
 235         if (pte_migrating(pteval)) {
 236                 if (in_nmi() && search_exception_tables(pc))
 237                         return 0;
 238                 wait_for_migration(pte);
 239                 return 1;
 240         }
 241
 242         if (!is_kernel_mode || !pte_present(pteval))
 243                 return 0;
 244         if (fault_num == INT_ITLB_MISS) {
 245                 if (pte_exec(pteval))
 246                         return 1;
 247         } else if (write) {
 248                 if (pte_write(pteval))
 249                         return 1;
 250         } else {
 251                 if (pte_read(pteval))
 252                         return 1;
 253         }
 254
 255         return 0;
 256 }
 257
 258 /*
 259  * This routine is responsible for faulting in user pages.
 260  * It passes the work off to one of the appropriate routines.
 261  * It returns true if the fault was successfully handled.
 262  */
 263 static int handle_page_fault(struct pt_regs *regs,
 264                              int fault_num,
 265                              int is_page_fault,
 266                              unsigned long address,
 267                              int write)
 268 {
 269         struct task_struct *tsk;
 270         struct mm_struct *mm;
 271         struct vm_area_struct *vma;
 272         unsigned long stack_offset;
 273         int fault;
 274         int si_code;
 275         int is_kernel_mode;
 276         pgd_t *pgd;
 277         unsigned int flags;
 278
 279         /* on TILE, protection faults are always writes */
 280         if (!is_page_fault)
 281                 write = 1;
 282
 283         flags = (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
 284                  (write ? FAULT_FLAG_WRITE : 0));
 285
 286         is_kernel_mode = (EX1_PL(regs->ex1) != USER_PL);
 287
 288         tsk = validate_current();
 289
 290         /*
 291          * Check to see if we might be overwriting the stack, and bail
 292          * out if so.  The page fault code is a relatively likely
 293          * place to get trapped in an infinite regress, and once we
 294          * overwrite the whole stack, it becomes very hard to recover.
 295          */
 296         stack_offset = stack_pointer & (THREAD_SIZE-1);
 297         if (stack_offset < THREAD_SIZE / 8) {
 298                 pr_alert("Potential stack overrun: sp %#lx\n",
 299                        stack_pointer);
 300                 show_regs(regs);
 301                 pr_alert("Killing current process %d/%s\n",
 302                        tsk->pid, tsk->comm);
 303                 do_group_exit(SIGKILL);
 304         }
 305
 306         /*
 307          * Early on, we need to check for migrating PTE entries;
 308          * see homecache.c.  If we find a migrating PTE, we wait until
 309          * the backing page claims to be done migrating, then we proceed.
 310          * For kernel PTEs, we rewrite the PTE and return and retry.
 311          * Otherwise, we treat the fault like a normal "no PTE" fault,
 312          * rather than trying to patch up the existing PTE.
 313          */
 314         pgd = get_current_pgd();
 315         if (handle_migrating_pte(pgd, fault_num, address, regs->pc,
 316                                  is_kernel_mode, write))
 317                 return 1;
 318
 319         si_code = SEGV_MAPERR;
 320
 321         /*
 322          * We fault-in kernel-space virtual memory on-demand. The
 323          * 'reference' page table is init_mm.pgd.
 324          *
 325          * NOTE! We MUST NOT take any locks for this case. We may
 326          * be in an interrupt or a critical region, and should
 327          * only copy the information from the master page table,
 328          * nothing more.
 329          *
 330          * This verifies that the fault happens in kernel space
 331          * and that the fault was not a protection fault.
 332          */
 333         if (unlikely(address >= TASK_SIZE &&
 334                      !is_arch_mappable_range(address, 0))) {
 335                 if (is_kernel_mode && is_page_fault &&
 336                     vmalloc_fault(pgd, address) >= 0)
 337                         return 1;
 338                 /*
 339                  * Don't take the mm semaphore here. If we fixup a prefetch
 340                  * fault we could otherwise deadlock.
 341                  */
 342                 mm = NULL;  /* happy compiler */
 343                 vma = NULL;
 344                 goto bad_area_nosemaphore;
 345         }
 346
 347         /*
 348          * If we're trying to touch user-space addresses, we must
 349          * be either at PL0, or else with interrupts enabled in the
 350          * kernel, so either way we can re-enable interrupts here
 351          * unless we are doing atomic access to user space with
 352          * interrupts disabled.
 353          */
 354         if (!(regs->flags & PT_FLAGS_DISABLE_IRQ))
 355                 local_irq_enable();
 356
 357         mm = tsk->mm;
 358
 359         /*
 360          * If we're in an interrupt, have no user context or are running in an
 361          * atomic region then we must not take the fault.
 362          */
 363         if (in_atomic() || !mm) {
 364                 vma = NULL;  /* happy compiler */
 365                 goto bad_area_nosemaphore;
 366         }
 367
 368         /*
 369          * When running in the kernel we expect faults to occur only to
 370          * addresses in user space.  All other faults represent errors in the
 371          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 372          * erroneous fault occurring in a code path which already holds mmap_sem
 373          * we will deadlock attempting to validate the fault against the
 374          * address space.  Luckily the kernel only validly references user
 375          * space from well defined areas of code, which are listed in the
 376          * exceptions table.
 377          *
 378          * As the vast majority of faults will be valid we will only perform
 379          * the source reference check when there is a possibility of a deadlock.
 380          * Attempt to lock the address space, if we cannot we then validate the
 381          * source.  If this is invalid we can skip the address space check,
 382          * thus avoiding the deadlock.
 383          */
 384         if (!down_read_trylock(&mm->mmap_sem)) {
 385                 if (is_kernel_mode &&
 386                     !search_exception_tables(regs->pc)) {
 387                         vma = NULL;  /* happy compiler */
 388                         goto bad_area_nosemaphore;
 389                 }
 390
 391 retry:
 392                 down_read(&mm->mmap_sem);
 393         }
 394
 395         vma = find_vma(mm, address);
 396         if (!vma)
 397                 goto bad_area;
 398         if (vma->vm_start <= address)
 399                 goto good_area;
 400         if (!(vma->vm_flags & VM_GROWSDOWN))
 401                 goto bad_area;
 402         if (regs->sp < PAGE_OFFSET) {
 403                 /*
 404                  * accessing the stack below sp is always a bug.
 405                  */
 406                 if (address < regs->sp)
 407                         goto bad_area;
 408         }
 409         if (expand_stack(vma, address))
 410                 goto bad_area;
 411
 412 /*
 413  * Ok, we have a good vm_area for this memory access, so
 414  * we can handle it..
 415  */
 416 good_area:
 417         si_code = SEGV_ACCERR;
 418         if (fault_num == INT_ITLB_MISS) {
 419                 if (!(vma->vm_flags & VM_EXEC))
 420                         goto bad_area;
 421         } else if (write) {
 422 #ifdef TEST_VERIFY_AREA
 423                 if (!is_page_fault && regs->cs == KERNEL_CS)
 424                         pr_err("WP fault at "REGFMT"\n", regs->eip);
 425 #endif
 426                 if (!(vma->vm_flags & VM_WRITE))
 427                         goto bad_area;
 428         } else {
 429                 if (!is_page_fault || !(vma->vm_flags & VM_READ))
 430                         goto bad_area;
 431         }
 432
 433  survive:
 434         /*
 435          * If for any reason at all we couldn't handle the fault,
 436          * make sure we exit gracefully rather than endlessly redo
 437          * the fault.
 438          */
 439         fault = handle_mm_fault(mm, vma, address, flags);
 440
 441         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
 442                 return 0;
 443
 444         if (unlikely(fault & VM_FAULT_ERROR)) {
 445                 if (fault & VM_FAULT_OOM)
 446                         goto out_of_memory;
 447                 else if (fault & VM_FAULT_SIGBUS)
 448                         goto do_sigbus;
 449                 BUG();
 450         }
 451         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 452                 if (fault & VM_FAULT_MAJOR)
 453                         tsk->maj_flt++;
 454                 else
 455                         tsk->min_flt++;
 456                 if (fault & VM_FAULT_RETRY) {
 457                         flags &= ~FAULT_FLAG_ALLOW_RETRY;
 458                         flags |= FAULT_FLAG_TRIED;
 459
 460                          /*
 461                           * No need to up_read(&mm->mmap_sem) as we would
 462                           * have already released it in __lock_page_or_retry
 463                           * in mm/filemap.c.
 464                           */
 465                         goto retry;
 466                 }
 467         }
 468
 469 #if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
 470         /*
 471          * If this was an asynchronous fault,
 472          * restart the appropriate engine.
 473          */
 474         switch (fault_num) {
 475 #if CHIP_HAS_TILE_DMA()
 476         case INT_DMATLB_MISS:
 477         case INT_DMATLB_MISS_DWNCL:
 478         case INT_DMATLB_ACCESS:
 479         case INT_DMATLB_ACCESS_DWNCL:
 480                 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK);
 481                 break;
 482 #endif
 483 #if CHIP_HAS_SN_PROC()
 484         case INT_SNITLB_MISS:
 485         case INT_SNITLB_MISS_DWNCL:
 486                 __insn_mtspr(SPR_SNCTL,
 487                              __insn_mfspr(SPR_SNCTL) &
 488                              ~SPR_SNCTL__FRZPROC_MASK);
 489                 break;
 490 #endif
 491         }
 492 #endif
 493
 494         up_read(&mm->mmap_sem);
 495         return 1;
 496
 497 /*
 498  * Something tried to access memory that isn't in our memory map..
 499  * Fix it, but check if it's kernel or user first..
 500  */
 501 bad_area:
 502         up_read(&mm->mmap_sem);
 503
 504 bad_area_nosemaphore:
 505         /* User mode accesses just cause a SIGSEGV */
 506         if (!is_kernel_mode) {
 507                 /*
 508                  * It's possible to have interrupts off here.
 509                  */
 510                 local_irq_enable();
 511
 512                 force_sig_info_fault("segfault", SIGSEGV, si_code, address,
 513                                      fault_num, tsk, regs);
 514                 return 0;
 515         }
 516
 517 no_context:
 518         /* Are we prepared to handle this kernel fault?  */
 519         if (fixup_exception(regs))
 520                 return 0;
 521
 522 /*
 523  * Oops. The kernel tried to access some bad page. We'll have to
 524  * terminate things with extreme prejudice.
 525  */
 526
 527         bust_spinlocks(1);
 528
 529         /* FIXME: no lookup_address() yet */
 530 #ifdef SUPPORT_LOOKUP_ADDRESS
 531         if (fault_num == INT_ITLB_MISS) {
 532                 pte_t *pte = lookup_address(address);
 533
 534                 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 535                         pr_crit("kernel tried to execute"
 536                                " non-executable page - exploit attempt?"
 537                                " (uid: %d)\n", current->uid);
 538         }
 539 #endif
 540         if (address < PAGE_SIZE)
 541                 pr_alert("Unable to handle kernel NULL pointer dereference\n");
 542         else
 543                 pr_alert("Unable to handle kernel paging request\n");
 544         pr_alert(" at virtual address "REGFMT", pc "REGFMT"\n",
 545                  address, regs->pc);
 546
 547         show_regs(regs);
 548
 549         if (unlikely(tsk->pid < 2)) {
 550                 panic("Kernel page fault running %s!",
 551                       is_idle_task(tsk) ? "the idle task" : "init");
 552         }
 553
 554         /*
 555          * More FIXME: we should probably copy the i386 here and
 556          * implement a generic die() routine.  Not today.
 557          */
 558 #ifdef SUPPORT_DIE
 559         die("Oops", regs);
 560 #endif
 561         bust_spinlocks(1);
 562
 563         do_group_exit(SIGKILL);
 564
 565 /*
 566  * We ran out of memory, or some other thing happened to us that made
 567  * us unable to handle the page fault gracefully.
 568  */
 569 out_of_memory:
 570         up_read(&mm->mmap_sem);
 571         if (is_global_init(tsk)) {
 572                 yield();
 573                 down_read(&mm->mmap_sem);
 574                 goto survive;
 575         }
 576         if (is_kernel_mode)
 577                 goto no_context;
 578         pagefault_out_of_memory();
 579         return 0;
 580
 581 do_sigbus:
 582         up_read(&mm->mmap_sem);
 583
 584         /* Kernel mode? Handle exceptions or die */
 585         if (is_kernel_mode)
 586                 goto no_context;
 587
 588         force_sig_info_fault("bus error", SIGBUS, BUS_ADRERR, address,
 589                              fault_num, tsk, regs);
 590         return 0;
 591 }
 592
 593 #ifndef __tilegx__
 594
 595 /* We must release ICS before panicking or we won't get anywhere. */
 596 #define ics_panic(fmt, ...) do { \
 597         __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); \
 598         panic(fmt, __VA_ARGS__); \
 599 } while (0)
 600
 601 /*
 602  * When we take an ITLB or DTLB fault or access violation in the
 603  * supervisor while the critical section bit is set, the hypervisor is
 604  * reluctant to write new values into the EX_CONTEXT_K_x registers,
 605  * since that might indicate we have not yet squirreled the SPR
 606  * contents away and can thus safely take a recursive interrupt.
 607  * Accordingly, the hypervisor passes us the PC via SYSTEM_SAVE_K_2.
 608  *
 609  * Note that this routine is called before homecache_tlb_defer_enter(),
 610  * which means that we can properly unlock any atomics that might
 611  * be used there (good), but also means we must be very sensitive
 612  * to not touch any data structures that might be located in memory
 613  * that could migrate, as we could be entering the kernel on a dataplane
 614  * cpu that has been deferring kernel TLB updates.  This means, for
 615  * example, that we can't migrate init_mm or its pgd.
 616  */
 617 struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num,
 618                                       unsigned long address,
 619                                       unsigned long info)
 620 {
 621         unsigned long pc = info & ~1;
 622         int write = info & 1;
 623         pgd_t *pgd = get_current_pgd();
 624
 625         /* Retval is 1 at first since we will handle the fault fully. */
 626         struct intvec_state state = {
 627                 do_page_fault, fault_num, address, write, 1
 628         };
 629
 630         /* Validate that we are plausibly in the right routine. */
 631         if ((pc & 0x7) != 0 || pc < PAGE_OFFSET ||
 632             (fault_num != INT_DTLB_MISS &&
 633              fault_num != INT_DTLB_ACCESS)) {
 634                 unsigned long old_pc = regs->pc;
 635                 regs->pc = pc;
 636                 ics_panic("Bad ICS page fault args:"
 637                           " old PC %#lx, fault %d/%d at %#lx\n",
 638                           old_pc, fault_num, write, address);
 639         }
 640
 641         /* We might be faulting on a vmalloc page, so check that first. */
 642         if (fault_num != INT_DTLB_ACCESS && vmalloc_fault(pgd, address) >= 0)
 643                 return state;
 644
 645         /*
 646          * If we faulted with ICS set in sys_cmpxchg, we are providing
 647          * a user syscall service that should generate a signal on
 648          * fault.  We didn't set up a kernel stack on initial entry to
 649          * sys_cmpxchg, but instead had one set up by the fault, which
 650          * (because sys_cmpxchg never releases ICS) came to us via the
 651          * SYSTEM_SAVE_K_2 mechanism, and thus EX_CONTEXT_K_[01] are
 652          * still referencing the original user code.  We release the
 653          * atomic lock and rewrite pt_regs so that it appears that we
 654          * came from user-space directly, and after we finish the
 655          * fault we'll go back to user space and re-issue the swint.
 656          * This way the backtrace information is correct if we need to
 657          * emit a stack dump at any point while handling this.
 658          *
 659          * Must match register use in sys_cmpxchg().
 660          */
 661         if (pc >= (unsigned long) sys_cmpxchg &&
 662             pc < (unsigned long) __sys_cmpxchg_end) {
 663 #ifdef CONFIG_SMP
 664                 /* Don't unlock before we could have locked. */
 665                 if (pc >= (unsigned long)__sys_cmpxchg_grab_lock) {
 666                         int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
 667                         __atomic_fault_unlock(lock_ptr);
 668                 }
 669 #endif
 670                 regs->sp = regs->regs[27];
 671         }
 672
 673         /*
 674          * We can also fault in the atomic assembly, in which
 675          * case we use the exception table to do the first-level fixup.
 676          * We may re-fixup again in the real fault handler if it
 677          * turns out the faulting address is just bad, and not,
 678          * for example, migrating.
 679          */
 680         else if (pc >= (unsigned long) __start_atomic_asm_code &&
 681                    pc < (unsigned long) __end_atomic_asm_code) {
 682                 const struct exception_table_entry *fixup;
 683 #ifdef CONFIG_SMP
 684                 /* Unlock the atomic lock. */
 685                 int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]);
 686                 __atomic_fault_unlock(lock_ptr);
 687 #endif
 688                 fixup = search_exception_tables(pc);
 689                 if (!fixup)
 690                         ics_panic("ICS atomic fault not in table:"
 691                                   " PC %#lx, fault %d", pc, fault_num);
 692                 regs->pc = fixup->fixup;
 693                 regs->ex1 = PL_ICS_EX1(KERNEL_PL, 0);
 694         }
 695
 696         /*
 697          * Now that we have released the atomic lock (if necessary),
 698          * it's safe to spin if the PTE that caused the fault was migrating.
 699          */
 700         if (fault_num == INT_DTLB_ACCESS)
 701                 write = 1;
 702         if (handle_migrating_pte(pgd, fault_num, address, pc, 1, write))
 703                 return state;
 704
 705         /* Return zero so that we continue on with normal fault handling. */
 706         state.retval = 0;
 707         return state;
 708 }
 709
 710 #endif /* !__tilegx__ */
 711
 712 /*
 713  * This routine handles page faults.  It determines the address, and the
 714  * problem, and then passes it handle_page_fault() for normal DTLB and
 715  * ITLB issues, and for DMA or SN processor faults when we are in user
 716  * space.  For the latter, if we're in kernel mode, we just save the
 717  * interrupt away appropriately and return immediately.  We can't do
 718  * page faults for user code while in kernel mode.
 719  */
 720 void do_page_fault(struct pt_regs *regs, int fault_num,
 721                    unsigned long address, unsigned long write)
 722 {
 723         int is_page_fault;
 724
 725         /* This case should have been handled by do_page_fault_ics(). */
 726         BUG_ON(write & ~1);
 727
 728 #if CHIP_HAS_TILE_DMA()
 729         /*
 730          * If it's a DMA fault, suspend the transfer while we're
 731          * handling the miss; we'll restart after it's handled.  If we
 732          * don't suspend, it's possible that this process could swap
 733          * out and back in, and restart the engine since the DMA is
 734          * still 'running'.
 735          */
 736         if (fault_num == INT_DMATLB_MISS ||
 737             fault_num == INT_DMATLB_ACCESS ||
 738             fault_num == INT_DMATLB_MISS_DWNCL ||
 739             fault_num == INT_DMATLB_ACCESS_DWNCL) {
 740                 __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK);
 741                 while (__insn_mfspr(SPR_DMA_USER_STATUS) &
 742                        SPR_DMA_STATUS__BUSY_MASK)
 743                         ;
 744         }
 745 #endif
 746
 747         /* Validate fault num and decide if this is a first-time page fault. */
 748         switch (fault_num) {
 749         case INT_ITLB_MISS:
 750         case INT_DTLB_MISS:
 751 #if CHIP_HAS_TILE_DMA()
 752         case INT_DMATLB_MISS:
 753         case INT_DMATLB_MISS_DWNCL:
 754 #endif
 755 #if CHIP_HAS_SN_PROC()
 756         case INT_SNITLB_MISS:
 757         case INT_SNITLB_MISS_DWNCL:
 758 #endif
 759                 is_page_fault = 1;
 760                 break;
 761
 762         case INT_DTLB_ACCESS:
 763 #if CHIP_HAS_TILE_DMA()
 764         case INT_DMATLB_ACCESS:
 765         case INT_DMATLB_ACCESS_DWNCL:
 766 #endif
 767                 is_page_fault = 0;
 768                 break;
 769
 770         default:
 771                 panic("Bad fault number %d in do_page_fault", fault_num);
 772         }
 773
 774 #if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
 775         if (EX1_PL(regs->ex1) != USER_PL) {
 776                 struct async_tlb *async;
 777                 switch (fault_num) {
 778 #if CHIP_HAS_TILE_DMA()
 779                 case INT_DMATLB_MISS:
 780                 case INT_DMATLB_ACCESS:
 781                 case INT_DMATLB_MISS_DWNCL:
 782                 case INT_DMATLB_ACCESS_DWNCL:
 783                         async = &current->thread.dma_async_tlb;
 784                         break;
 785 #endif
 786 #if CHIP_HAS_SN_PROC()
 787                 case INT_SNITLB_MISS:
 788                 case INT_SNITLB_MISS_DWNCL:
 789                         async = &current->thread.sn_async_tlb;
 790                         break;
 791 #endif
 792                 default:
 793                         async = NULL;
 794                 }
 795                 if (async) {
 796
 797                         /*
 798                          * No vmalloc check required, so we can allow
 799                          * interrupts immediately at this point.
 800                          */
 801                         local_irq_enable();
 802
 803                         set_thread_flag(TIF_ASYNC_TLB);
 804                         if (async->fault_num != 0) {
 805                                 panic("Second async fault %d;"
 806                                       " old fault was %d (%#lx/%ld)",
 807                                       fault_num, async->fault_num,
 808                                       address, write);
 809                         }
 810                         BUG_ON(fault_num == 0);
 811                         async->fault_num = fault_num;
 812                         async->is_fault = is_page_fault;
 813                         async->is_write = write;
 814                         async->address = address;
 815                         return;
 816                 }
 817         }
 818 #endif
 819
 820         handle_page_fault(regs, fault_num, is_page_fault, address, write);
 821 }
 822
 823
 824 #if CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC()
 825 /*
 826  * Check an async_tlb structure to see if a deferred fault is waiting,
 827  * and if so pass it to the page-fault code.
 828  */
 829 static void handle_async_page_fault(struct pt_regs *regs,
 830                                     struct async_tlb *async)
 831 {
 832         if (async->fault_num) {
 833                 /*
 834                  * Clear async->fault_num before calling the page-fault
 835                  * handler so that if we re-interrupt before returning
 836                  * from the function we have somewhere to put the
 837                  * information from the new interrupt.
 838                  */
 839                 int fault_num = async->fault_num;
 840                 async->fault_num = 0;
 841                 handle_page_fault(regs, fault_num, async->is_fault,
 842                                   async->address, async->is_write);
 843         }
 844 }
 845
 846 /*
 847  * This routine effectively re-issues asynchronous page faults
 848  * when we are returning to user space.
 849  */
 850 void do_async_page_fault(struct pt_regs *regs)
 851 {
 852         /*
 853          * Clear thread flag early.  If we re-interrupt while processing
 854          * code here, we will reset it and recall this routine before
 855          * returning to user space.
 856          */
 857         clear_thread_flag(TIF_ASYNC_TLB);
 858
 859 #if CHIP_HAS_TILE_DMA()
 860         handle_async_page_fault(regs, &current->thread.dma_async_tlb);
 861 #endif
 862 #if CHIP_HAS_SN_PROC()
 863         handle_async_page_fault(regs, &current->thread.sn_async_tlb);
 864 #endif
 865 }
 866 #endif /* CHIP_HAS_TILE_DMA() || CHIP_HAS_SN_PROC() */
 867
 868
 869 void vmalloc_sync_all(void)
 870 {
 871 #ifdef __tilegx__
 872         /* Currently all L1 kernel pmd's are static and shared. */
 873         BUG_ON(pgd_index(VMALLOC_END) != pgd_index(VMALLOC_START));
 874 #else
 875         /*
 876          * Note that races in the updates of insync and start aren't
 877          * problematic: insync can only get set bits added, and updates to
 878          * start are only improving performance (without affecting correctness
 879          * if undone).
 880          */
 881         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 882         static unsigned long start = PAGE_OFFSET;
 883         unsigned long address;
 884
 885         BUILD_BUG_ON(PAGE_OFFSET & ~PGDIR_MASK);
 886         for (address = start; address >= PAGE_OFFSET; address += PGDIR_SIZE) {
 887                 if (!test_bit(pgd_index(address), insync)) {
 888                         unsigned long flags;
 889                         struct list_head *pos;
 890
 891                         spin_lock_irqsave(&pgd_lock, flags);
 892                         list_for_each(pos, &pgd_list)
 893                                 if (!vmalloc_sync_one(list_to_pgd(pos),
 894                                                                 address)) {
 895                                         /* Must be at first entry in list. */
 896                                         BUG_ON(pos != pgd_list.next);
 897                                         break;
 898                                 }
 899                         spin_unlock_irqrestore(&pgd_lock, flags);
 900                         if (pos != pgd_list.next)
 901                                 set_bit(pgd_index(address), insync);
 902                 }
 903                 if (address == start && test_bit(pgd_index(address), insync))
 904                         start = address + PGDIR_SIZE;
 905         }
 906 #endif
 907 }