linux/arch/m68k/sun3/mmu_emu.c
<<
>>
Prefs
   1/*
   2** Tablewalk MMU emulator
   3**
   4** by Toshiyasu Morita
   5**
   6** Started 1/16/98 @ 2:22 am
   7*/
   8
   9#include <linux/mman.h>
  10#include <linux/mm.h>
  11#include <linux/kernel.h>
  12#include <linux/ptrace.h>
  13#include <linux/delay.h>
  14#include <linux/bootmem.h>
  15#include <linux/bitops.h>
  16#include <linux/module.h>
  17
  18#include <asm/setup.h>
  19#include <asm/traps.h>
  20#include <asm/system.h>
  21#include <asm/uaccess.h>
  22#include <asm/page.h>
  23#include <asm/pgtable.h>
  24#include <asm/sun3mmu.h>
  25#include <asm/segment.h>
  26#include <asm/oplib.h>
  27#include <asm/mmu_context.h>
  28#include <asm/dvma.h>
  29
  30
  31#undef DEBUG_MMU_EMU
  32#define DEBUG_PROM_MAPS
  33
  34/*
  35** Defines
  36*/
  37
  38#define CONTEXTS_NUM            8
  39#define SEGMAPS_PER_CONTEXT_NUM 2048
  40#define PAGES_PER_SEGMENT       16
  41#define PMEGS_NUM               256
  42#define PMEG_MASK               0xFF
  43
  44/*
  45** Globals
  46*/
  47
  48unsigned long m68k_vmalloc_end;
  49EXPORT_SYMBOL(m68k_vmalloc_end);
  50
  51unsigned long pmeg_vaddr[PMEGS_NUM];
  52unsigned char pmeg_alloc[PMEGS_NUM];
  53unsigned char pmeg_ctx[PMEGS_NUM];
  54
  55/* pointers to the mm structs for each task in each
  56   context. 0xffffffff is a marker for kernel context */
  57static struct mm_struct *ctx_alloc[CONTEXTS_NUM] = {
  58    [0] = (struct mm_struct *)0xffffffff
  59};
  60
  61/* has this context been mmdrop'd? */
  62static unsigned char ctx_avail = CONTEXTS_NUM-1;
  63
  64/* array of pages to be marked off for the rom when we do mem_init later */
  65/* 256 pages lets the rom take up to 2mb of physical ram..  I really
  66   hope it never wants mote than that. */
  67unsigned long rom_pages[256];
  68
  69/* Print a PTE value in symbolic form. For debugging. */
  70void print_pte (pte_t pte)
  71{
  72#if 0
  73        /* Verbose version. */
  74        unsigned long val = pte_val (pte);
  75        printk (" pte=%lx [addr=%lx",
  76                val, (val & SUN3_PAGE_PGNUM_MASK) << PAGE_SHIFT);
  77        if (val & SUN3_PAGE_VALID)      printk (" valid");
  78        if (val & SUN3_PAGE_WRITEABLE)  printk (" write");
  79        if (val & SUN3_PAGE_SYSTEM)     printk (" sys");
  80        if (val & SUN3_PAGE_NOCACHE)    printk (" nocache");
  81        if (val & SUN3_PAGE_ACCESSED)   printk (" accessed");
  82        if (val & SUN3_PAGE_MODIFIED)   printk (" modified");
  83        switch (val & SUN3_PAGE_TYPE_MASK) {
  84                case SUN3_PAGE_TYPE_MEMORY: printk (" memory"); break;
  85                case SUN3_PAGE_TYPE_IO:     printk (" io");     break;
  86                case SUN3_PAGE_TYPE_VME16:  printk (" vme16");  break;
  87                case SUN3_PAGE_TYPE_VME32:  printk (" vme32");  break;
  88        }
  89        printk ("]\n");
  90#else
  91        /* Terse version. More likely to fit on a line. */
  92        unsigned long val = pte_val (pte);
  93        char flags[7], *type;
  94
  95        flags[0] = (val & SUN3_PAGE_VALID)     ? 'v' : '-';
  96        flags[1] = (val & SUN3_PAGE_WRITEABLE) ? 'w' : '-';
  97        flags[2] = (val & SUN3_PAGE_SYSTEM)    ? 's' : '-';
  98        flags[3] = (val & SUN3_PAGE_NOCACHE)   ? 'x' : '-';
  99        flags[4] = (val & SUN3_PAGE_ACCESSED)  ? 'a' : '-';
 100        flags[5] = (val & SUN3_PAGE_MODIFIED)  ? 'm' : '-';
 101        flags[6] = '\0';
 102
 103        switch (val & SUN3_PAGE_TYPE_MASK) {
 104                case SUN3_PAGE_TYPE_MEMORY: type = "memory"; break;
 105                case SUN3_PAGE_TYPE_IO:     type = "io"    ; break;
 106                case SUN3_PAGE_TYPE_VME16:  type = "vme16" ; break;
 107                case SUN3_PAGE_TYPE_VME32:  type = "vme32" ; break;
 108                default: type = "unknown?"; break;
 109        }
 110
 111        printk (" pte=%08lx [%07lx %s %s]\n",
 112                val, (val & SUN3_PAGE_PGNUM_MASK) << PAGE_SHIFT, flags, type);
 113#endif
 114}
 115
 116/* Print the PTE value for a given virtual address. For debugging. */
 117void print_pte_vaddr (unsigned long vaddr)
 118{
 119        printk (" vaddr=%lx [%02lx]", vaddr, sun3_get_segmap (vaddr));
 120        print_pte (__pte (sun3_get_pte (vaddr)));
 121}
 122
 123/*
 124 * Initialise the MMU emulator.
 125 */
 126void mmu_emu_init(unsigned long bootmem_end)
 127{
 128        unsigned long seg, num;
 129        int i,j;
 130
 131        memset(rom_pages, 0, sizeof(rom_pages));
 132        memset(pmeg_vaddr, 0, sizeof(pmeg_vaddr));
 133        memset(pmeg_alloc, 0, sizeof(pmeg_alloc));
 134        memset(pmeg_ctx, 0, sizeof(pmeg_ctx));
 135
 136        /* pmeg align the end of bootmem, adding another pmeg,
 137         * later bootmem allocations will likely need it */
 138        bootmem_end = (bootmem_end + (2 * SUN3_PMEG_SIZE)) & ~SUN3_PMEG_MASK;
 139
 140        /* mark all of the pmegs used thus far as reserved */
 141        for (i=0; i < __pa(bootmem_end) / SUN3_PMEG_SIZE ; ++i)
 142                pmeg_alloc[i] = 2;
 143
 144
 145        /* I'm thinking that most of the top pmeg's are going to be
 146           used for something, and we probably shouldn't risk it */
 147        for(num = 0xf0; num <= 0xff; num++)
 148                pmeg_alloc[num] = 2;
 149
 150        /* liberate all existing mappings in the rest of kernel space */
 151        for(seg = bootmem_end; seg < 0x0f800000; seg += SUN3_PMEG_SIZE) {
 152                i = sun3_get_segmap(seg);
 153
 154                if(!pmeg_alloc[i]) {
 155#ifdef DEBUG_MMU_EMU
 156                        printk("freed: ");
 157                        print_pte_vaddr (seg);
 158#endif
 159                        sun3_put_segmap(seg, SUN3_INVALID_PMEG);
 160                }
 161        }
 162
 163        j = 0;
 164        for (num=0, seg=0x0F800000; seg<0x10000000; seg+=16*PAGE_SIZE) {
 165                if (sun3_get_segmap (seg) != SUN3_INVALID_PMEG) {
 166#ifdef DEBUG_PROM_MAPS
 167                        for(i = 0; i < 16; i++) {
 168                                printk ("mapped:");
 169                                print_pte_vaddr (seg + (i*PAGE_SIZE));
 170                                break;
 171                        }
 172#endif
 173                        // the lowest mapping here is the end of our
 174                        // vmalloc region
 175                        if (!m68k_vmalloc_end)
 176                                m68k_vmalloc_end = seg;
 177
 178                        // mark the segmap alloc'd, and reserve any
 179                        // of the first 0xbff pages the hardware is
 180                        // already using...  does any sun3 support > 24mb?
 181                        pmeg_alloc[sun3_get_segmap(seg)] = 2;
 182                }
 183        }
 184
 185        dvma_init();
 186
 187
 188        /* blank everything below the kernel, and we've got the base
 189           mapping to start all the contexts off with... */
 190        for(seg = 0; seg < PAGE_OFFSET; seg += SUN3_PMEG_SIZE)
 191                sun3_put_segmap(seg, SUN3_INVALID_PMEG);
 192
 193        set_fs(MAKE_MM_SEG(3));
 194        for(seg = 0; seg < 0x10000000; seg += SUN3_PMEG_SIZE) {
 195                i = sun3_get_segmap(seg);
 196                for(j = 1; j < CONTEXTS_NUM; j++)
 197                        (*(romvec->pv_setctxt))(j, (void *)seg, i);
 198        }
 199        set_fs(KERNEL_DS);
 200
 201}
 202
 203/* erase the mappings for a dead context.  Uses the pg_dir for hints
 204   as the pmeg tables proved somewhat unreliable, and unmapping all of
 205   TASK_SIZE was much slower and no more stable. */
 206/* todo: find a better way to keep track of the pmegs used by a
 207   context for when they're cleared */
 208void clear_context(unsigned long context)
 209{
 210     unsigned char oldctx;
 211     unsigned long i;
 212
 213     if(context) {
 214             if(!ctx_alloc[context])
 215                     panic("clear_context: context not allocated\n");
 216
 217             ctx_alloc[context]->context = SUN3_INVALID_CONTEXT;
 218             ctx_alloc[context] = (struct mm_struct *)0;
 219             ctx_avail++;
 220     }
 221
 222     oldctx = sun3_get_context();
 223
 224     sun3_put_context(context);
 225
 226     for(i = 0; i < SUN3_INVALID_PMEG; i++) {
 227             if((pmeg_ctx[i] == context) && (pmeg_alloc[i] == 1)) {
 228                     sun3_put_segmap(pmeg_vaddr[i], SUN3_INVALID_PMEG);
 229                     pmeg_ctx[i] = 0;
 230                     pmeg_alloc[i] = 0;
 231                     pmeg_vaddr[i] = 0;
 232             }
 233     }
 234
 235     sun3_put_context(oldctx);
 236}
 237
 238/* gets an empty context.  if full, kills the next context listed to
 239   die first */
 240/* This context invalidation scheme is, well, totally arbitrary, I'm
 241   sure it could be much more intelligent...  but it gets the job done
 242   for now without much overhead in making it's decision. */
 243/* todo: come up with optimized scheme for flushing contexts */
 244unsigned long get_free_context(struct mm_struct *mm)
 245{
 246        unsigned long new = 1;
 247        static unsigned char next_to_die = 1;
 248
 249        if(!ctx_avail) {
 250                /* kill someone to get our context */
 251                new = next_to_die;
 252                clear_context(new);
 253                next_to_die = (next_to_die + 1) & 0x7;
 254                if(!next_to_die)
 255                        next_to_die++;
 256        } else {
 257                while(new < CONTEXTS_NUM) {
 258                        if(ctx_alloc[new])
 259                                new++;
 260                        else
 261                                break;
 262                }
 263                // check to make sure one was really free...
 264                if(new == CONTEXTS_NUM)
 265                        panic("get_free_context: failed to find free context");
 266        }
 267
 268        ctx_alloc[new] = mm;
 269        ctx_avail--;
 270
 271        return new;
 272}
 273
 274/*
 275 * Dynamically select a `spare' PMEG and use it to map virtual `vaddr' in
 276 * `context'. Maintain internal PMEG management structures. This doesn't
 277 * actually map the physical address, but does clear the old mappings.
 278 */
 279//todo: better allocation scheme? but is extra complexity worthwhile?
 280//todo: only clear old entries if necessary? how to tell?
 281
 282inline void mmu_emu_map_pmeg (int context, int vaddr)
 283{
 284        static unsigned char curr_pmeg = 128;
 285        int i;
 286
 287        /* Round address to PMEG boundary. */
 288        vaddr &= ~SUN3_PMEG_MASK;
 289
 290        /* Find a spare one. */
 291        while (pmeg_alloc[curr_pmeg] == 2)
 292                ++curr_pmeg;
 293
 294
 295#ifdef DEBUG_MMU_EMU
 296printk("mmu_emu_map_pmeg: pmeg %x to context %d vaddr %x\n",
 297       curr_pmeg, context, vaddr);
 298#endif
 299
 300        /* Invalidate old mapping for the pmeg, if any */
 301        if (pmeg_alloc[curr_pmeg] == 1) {
 302                sun3_put_context(pmeg_ctx[curr_pmeg]);
 303                sun3_put_segmap (pmeg_vaddr[curr_pmeg], SUN3_INVALID_PMEG);
 304                sun3_put_context(context);
 305        }
 306
 307        /* Update PMEG management structures. */
 308        // don't take pmeg's away from the kernel...
 309        if(vaddr >= PAGE_OFFSET) {
 310                /* map kernel pmegs into all contexts */
 311                unsigned char i;
 312
 313                for(i = 0; i < CONTEXTS_NUM; i++) {
 314                        sun3_put_context(i);
 315                        sun3_put_segmap (vaddr, curr_pmeg);
 316                }
 317                sun3_put_context(context);
 318                pmeg_alloc[curr_pmeg] = 2;
 319                pmeg_ctx[curr_pmeg] = 0;
 320
 321        }
 322        else {
 323                pmeg_alloc[curr_pmeg] = 1;
 324                pmeg_ctx[curr_pmeg] = context;
 325                sun3_put_segmap (vaddr, curr_pmeg);
 326
 327        }
 328        pmeg_vaddr[curr_pmeg] = vaddr;
 329
 330        /* Set hardware mapping and clear the old PTE entries. */
 331        for (i=0; i<SUN3_PMEG_SIZE; i+=SUN3_PTE_SIZE)
 332                sun3_put_pte (vaddr + i, SUN3_PAGE_SYSTEM);
 333
 334        /* Consider a different one next time. */
 335        ++curr_pmeg;
 336}
 337
 338/*
 339 * Handle a pagefault at virtual address `vaddr'; check if there should be a
 340 * page there (specifically, whether the software pagetables indicate that
 341 * there is). This is necessary due to the limited size of the second-level
 342 * Sun3 hardware pagetables (256 groups of 16 pages). If there should be a
 343 * mapping present, we select a `spare' PMEG and use it to create a mapping.
 344 * `read_flag' is nonzero for a read fault; zero for a write. Returns nonzero
 345 * if we successfully handled the fault.
 346 */
 347//todo: should we bump minor pagefault counter? if so, here or in caller?
 348//todo: possibly inline this into bus_error030 in <asm/buserror.h> ?
 349
 350// kernel_fault is set when a kernel page couldn't be demand mapped,
 351// and forces another try using the kernel page table.  basically a
 352// hack so that vmalloc would work correctly.
 353
 354int mmu_emu_handle_fault (unsigned long vaddr, int read_flag, int kernel_fault)
 355{
 356        unsigned long segment, offset;
 357        unsigned char context;
 358        pte_t *pte;
 359        pgd_t * crp;
 360
 361        if(current->mm == NULL) {
 362                crp = swapper_pg_dir;
 363                context = 0;
 364        } else {
 365                context = current->mm->context;
 366                if(kernel_fault)
 367                        crp = swapper_pg_dir;
 368                else
 369                        crp = current->mm->pgd;
 370        }
 371
 372#ifdef DEBUG_MMU_EMU
 373        printk ("mmu_emu_handle_fault: vaddr=%lx type=%s crp=%p\n",
 374                vaddr, read_flag ? "read" : "write", crp);
 375#endif
 376
 377        segment = (vaddr >> SUN3_PMEG_SIZE_BITS) & 0x7FF;
 378        offset  = (vaddr >> SUN3_PTE_SIZE_BITS) & 0xF;
 379
 380#ifdef DEBUG_MMU_EMU
 381        printk ("mmu_emu_handle_fault: segment=%lx offset=%lx\n", segment, offset);
 382#endif
 383
 384        pte = (pte_t *) pgd_val (*(crp + segment));
 385
 386//todo: next line should check for valid pmd properly.
 387        if (!pte) {
 388//                printk ("mmu_emu_handle_fault: invalid pmd\n");
 389                return 0;
 390        }
 391
 392        pte = (pte_t *) __va ((unsigned long)(pte + offset));
 393
 394        /* Make sure this is a valid page */
 395        if (!(pte_val (*pte) & SUN3_PAGE_VALID))
 396                return 0;
 397
 398        /* Make sure there's a pmeg allocated for the page */
 399        if (sun3_get_segmap (vaddr&~SUN3_PMEG_MASK) == SUN3_INVALID_PMEG)
 400                mmu_emu_map_pmeg (context, vaddr);
 401
 402        /* Write the pte value to hardware MMU */
 403        sun3_put_pte (vaddr&PAGE_MASK, pte_val (*pte));
 404
 405        /* Update software copy of the pte value */
 406// I'm not sure this is necessary. If this is required, we ought to simply
 407// copy this out when we reuse the PMEG or at some other convenient time.
 408// Doing it here is fairly meaningless, anyway, as we only know about the
 409// first access to a given page. --m
 410        if (!read_flag) {
 411                if (pte_val (*pte) & SUN3_PAGE_WRITEABLE)
 412                        pte_val (*pte) |= (SUN3_PAGE_ACCESSED
 413                                           | SUN3_PAGE_MODIFIED);
 414                else
 415                        return 0;       /* Write-protect error. */
 416        } else
 417                pte_val (*pte) |= SUN3_PAGE_ACCESSED;
 418
 419#ifdef DEBUG_MMU_EMU
 420        printk ("seg:%d crp:%p ->", get_fs().seg, crp);
 421        print_pte_vaddr (vaddr);
 422        printk ("\n");
 423#endif
 424
 425        return 1;
 426}
 427