linux/arch/m68k/sun3/mmu_emu.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3** Tablewalk MMU emulator
   4**
   5** by Toshiyasu Morita
   6**
   7** Started 1/16/98 @ 2:22 am
   8*/
   9
  10#include <linux/init.h>
  11#include <linux/mman.h>
  12#include <linux/mm.h>
  13#include <linux/kernel.h>
  14#include <linux/ptrace.h>
  15#include <linux/delay.h>
  16#include <linux/memblock.h>
  17#include <linux/bitops.h>
  18#include <linux/module.h>
  19#include <linux/sched/mm.h>
  20
  21#include <asm/setup.h>
  22#include <asm/traps.h>
  23#include <linux/uaccess.h>
  24#include <asm/page.h>
  25#include <asm/sun3mmu.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        pr_cont(" pte=%lx [addr=%lx",
  76                val, (val & SUN3_PAGE_PGNUM_MASK) << PAGE_SHIFT);
  77        if (val & SUN3_PAGE_VALID)      pr_cont(" valid");
  78        if (val & SUN3_PAGE_WRITEABLE)  pr_cont(" write");
  79        if (val & SUN3_PAGE_SYSTEM)     pr_cont(" sys");
  80        if (val & SUN3_PAGE_NOCACHE)    pr_cont(" nocache");
  81        if (val & SUN3_PAGE_ACCESSED)   pr_cont(" accessed");
  82        if (val & SUN3_PAGE_MODIFIED)   pr_cont(" modified");
  83        switch (val & SUN3_PAGE_TYPE_MASK) {
  84                case SUN3_PAGE_TYPE_MEMORY: pr_cont(" memory"); break;
  85                case SUN3_PAGE_TYPE_IO:     pr_cont(" io");     break;
  86                case SUN3_PAGE_TYPE_VME16:  pr_cont(" vme16");  break;
  87                case SUN3_PAGE_TYPE_VME32:  pr_cont(" vme32");  break;
  88        }
  89        pr_cont("]\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        pr_cont(" 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        pr_cont(" 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 __init 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                        pr_info("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                                pr_info("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_fc(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_fc(USER_DATA);
 200}
 201
 202/* erase the mappings for a dead context.  Uses the pg_dir for hints
 203   as the pmeg tables proved somewhat unreliable, and unmapping all of
 204   TASK_SIZE was much slower and no more stable. */
 205/* todo: find a better way to keep track of the pmegs used by a
 206   context for when they're cleared */
 207void clear_context(unsigned long context)
 208{
 209     unsigned char oldctx;
 210     unsigned long i;
 211
 212     if(context) {
 213             if(!ctx_alloc[context])
 214                     panic("clear_context: context not allocated\n");
 215
 216             ctx_alloc[context]->context = SUN3_INVALID_CONTEXT;
 217             ctx_alloc[context] = (struct mm_struct *)0;
 218             ctx_avail++;
 219     }
 220
 221     oldctx = sun3_get_context();
 222
 223     sun3_put_context(context);
 224
 225     for(i = 0; i < SUN3_INVALID_PMEG; i++) {
 226             if((pmeg_ctx[i] == context) && (pmeg_alloc[i] == 1)) {
 227                     sun3_put_segmap(pmeg_vaddr[i], SUN3_INVALID_PMEG);
 228                     pmeg_ctx[i] = 0;
 229                     pmeg_alloc[i] = 0;
 230                     pmeg_vaddr[i] = 0;
 231             }
 232     }
 233
 234     sun3_put_context(oldctx);
 235}
 236
 237/* gets an empty context.  if full, kills the next context listed to
 238   die first */
 239/* This context invalidation scheme is, well, totally arbitrary, I'm
 240   sure it could be much more intelligent...  but it gets the job done
 241   for now without much overhead in making it's decision. */
 242/* todo: come up with optimized scheme for flushing contexts */
 243unsigned long get_free_context(struct mm_struct *mm)
 244{
 245        unsigned long new = 1;
 246        static unsigned char next_to_die = 1;
 247
 248        if(!ctx_avail) {
 249                /* kill someone to get our context */
 250                new = next_to_die;
 251                clear_context(new);
 252                next_to_die = (next_to_die + 1) & 0x7;
 253                if(!next_to_die)
 254                        next_to_die++;
 255        } else {
 256                while(new < CONTEXTS_NUM) {
 257                        if(ctx_alloc[new])
 258                                new++;
 259                        else
 260                                break;
 261                }
 262                // check to make sure one was really free...
 263                if(new == CONTEXTS_NUM)
 264                        panic("get_free_context: failed to find free context");
 265        }
 266
 267        ctx_alloc[new] = mm;
 268        ctx_avail--;
 269
 270        return new;
 271}
 272
 273/*
 274 * Dynamically select a `spare' PMEG and use it to map virtual `vaddr' in
 275 * `context'. Maintain internal PMEG management structures. This doesn't
 276 * actually map the physical address, but does clear the old mappings.
 277 */
 278//todo: better allocation scheme? but is extra complexity worthwhile?
 279//todo: only clear old entries if necessary? how to tell?
 280
 281inline void mmu_emu_map_pmeg (int context, int vaddr)
 282{
 283        static unsigned char curr_pmeg = 128;
 284        int i;
 285
 286        /* Round address to PMEG boundary. */
 287        vaddr &= ~SUN3_PMEG_MASK;
 288
 289        /* Find a spare one. */
 290        while (pmeg_alloc[curr_pmeg] == 2)
 291                ++curr_pmeg;
 292
 293
 294#ifdef DEBUG_MMU_EMU
 295        pr_info("mmu_emu_map_pmeg: pmeg %x to context %d vaddr %x\n",
 296                curr_pmeg, context, vaddr);
 297#endif
 298
 299        /* Invalidate old mapping for the pmeg, if any */
 300        if (pmeg_alloc[curr_pmeg] == 1) {
 301                sun3_put_context(pmeg_ctx[curr_pmeg]);
 302                sun3_put_segmap (pmeg_vaddr[curr_pmeg], SUN3_INVALID_PMEG);
 303                sun3_put_context(context);
 304        }
 305
 306        /* Update PMEG management structures. */
 307        // don't take pmeg's away from the kernel...
 308        if(vaddr >= PAGE_OFFSET) {
 309                /* map kernel pmegs into all contexts */
 310                unsigned char i;
 311
 312                for(i = 0; i < CONTEXTS_NUM; i++) {
 313                        sun3_put_context(i);
 314                        sun3_put_segmap (vaddr, curr_pmeg);
 315                }
 316                sun3_put_context(context);
 317                pmeg_alloc[curr_pmeg] = 2;
 318                pmeg_ctx[curr_pmeg] = 0;
 319
 320        }
 321        else {
 322                pmeg_alloc[curr_pmeg] = 1;
 323                pmeg_ctx[curr_pmeg] = context;
 324                sun3_put_segmap (vaddr, curr_pmeg);
 325
 326        }
 327        pmeg_vaddr[curr_pmeg] = vaddr;
 328
 329        /* Set hardware mapping and clear the old PTE entries. */
 330        for (i=0; i<SUN3_PMEG_SIZE; i+=SUN3_PTE_SIZE)
 331                sun3_put_pte (vaddr + i, SUN3_PAGE_SYSTEM);
 332
 333        /* Consider a different one next time. */
 334        ++curr_pmeg;
 335}
 336
 337/*
 338 * Handle a pagefault at virtual address `vaddr'; check if there should be a
 339 * page there (specifically, whether the software pagetables indicate that
 340 * there is). This is necessary due to the limited size of the second-level
 341 * Sun3 hardware pagetables (256 groups of 16 pages). If there should be a
 342 * mapping present, we select a `spare' PMEG and use it to create a mapping.
 343 * `read_flag' is nonzero for a read fault; zero for a write. Returns nonzero
 344 * if we successfully handled the fault.
 345 */
 346//todo: should we bump minor pagefault counter? if so, here or in caller?
 347//todo: possibly inline this into bus_error030 in <asm/buserror.h> ?
 348
 349// kernel_fault is set when a kernel page couldn't be demand mapped,
 350// and forces another try using the kernel page table.  basically a
 351// hack so that vmalloc would work correctly.
 352
 353int mmu_emu_handle_fault (unsigned long vaddr, int read_flag, int kernel_fault)
 354{
 355        unsigned long segment, offset;
 356        unsigned char context;
 357        pte_t *pte;
 358        pgd_t * crp;
 359
 360        if(current->mm == NULL) {
 361                crp = swapper_pg_dir;
 362                context = 0;
 363        } else {
 364                context = current->mm->context;
 365                if(kernel_fault)
 366                        crp = swapper_pg_dir;
 367                else
 368                        crp = current->mm->pgd;
 369        }
 370
 371#ifdef DEBUG_MMU_EMU
 372        pr_info("mmu_emu_handle_fault: vaddr=%lx type=%s crp=%p\n",
 373                vaddr, read_flag ? "read" : "write", crp);
 374#endif
 375
 376        segment = (vaddr >> SUN3_PMEG_SIZE_BITS) & 0x7FF;
 377        offset  = (vaddr >> SUN3_PTE_SIZE_BITS) & 0xF;
 378
 379#ifdef DEBUG_MMU_EMU
 380        pr_info("mmu_emu_handle_fault: segment=%lx offset=%lx\n", segment,
 381                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//                pr_info("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        pr_info("seg:%ld crp:%p ->", get_fs().seg, crp);
 421        print_pte_vaddr (vaddr);
 422        pr_cont("\n");
 423#endif
 424
 425        return 1;
 426}
 427