arch/mips/mm/init.c

   1 /*
   2  * This file is subject to the terms and conditions of the GNU General Public
   3  * License.  See the file "COPYING" in the main directory of this archive
   4  * for more details.
   5  *
   6  * Copyright (C) 1994 - 2000 Ralf Baechle
   7  * Copyright (C) 1999, 2000 Silicon Graphics, Inc.
   8  * Kevin D. Kissell, kevink@mips.com and Carsten Langgaard, carstenl@mips.com
   9  * Copyright (C) 2000 MIPS Technologies, Inc.  All rights reserved.
  10  */
  11 #include <linux/bug.h>
  12 #include <linux/init.h>
  13 #include <linux/export.h>
  14 #include <linux/signal.h>
  15 #include <linux/sched.h>
  16 #include <linux/smp.h>
  17 #include <linux/kernel.h>
  18 #include <linux/errno.h>
  19 #include <linux/string.h>
  20 #include <linux/types.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/ptrace.h>
  23 #include <linux/mman.h>
  24 #include <linux/mm.h>
  25 #include <linux/memblock.h>
  26 #include <linux/highmem.h>
  27 #include <linux/swap.h>
  28 #include <linux/proc_fs.h>
  29 #include <linux/pfn.h>
  30 #include <linux/hardirq.h>
  31 #include <linux/gfp.h>
  32 #include <linux/kcore.h>
  33 #include <linux/initrd.h>
  34
  35 #include <asm/bootinfo.h>
  36 #include <asm/cachectl.h>
  37 #include <asm/cpu.h>
  38 #include <asm/dma.h>
  39 #include <asm/maar.h>
  40 #include <asm/mmu_context.h>
  41 #include <asm/sections.h>
  42 #include <asm/pgalloc.h>
  43 #include <asm/tlb.h>
  44 #include <asm/fixmap.h>
  45
  46 /*
  47  * We have up to 8 empty zeroed pages so we can map one of the right colour
  48  * when needed.  This is necessary only on R4000 / R4400 SC and MC versions
  49  * where we have to avoid VCED / VECI exceptions for good performance at
  50  * any price.  Since page is never written to after the initialization we
  51  * don't have to care about aliases on other CPUs.
  52  */
  53 unsigned long empty_zero_page, zero_page_mask;
  54 EXPORT_SYMBOL_GPL(empty_zero_page);
  55 EXPORT_SYMBOL(zero_page_mask);
  56
  57 /*
  58  * Not static inline because used by IP27 special magic initialization code
  59  */
  60 void setup_zero_pages(void)
  61 {
  62         unsigned int order, i;
  63         struct page *page;
  64
  65         if (cpu_has_vce)
  66                 order = 3;
  67         else
  68                 order = 0;
  69
  70         empty_zero_page = __get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
  71         if (!empty_zero_page)
  72                 panic("Oh boy, that early out of memory?");
  73
  74         page = virt_to_page((void *)empty_zero_page);
  75         split_page(page, order);
  76         for (i = 0; i < (1 << order); i++, page++)
  77                 mark_page_reserved(page);
  78
  79         zero_page_mask = ((PAGE_SIZE << order) - 1) & PAGE_MASK;
  80 }
  81
  82 static void *__kmap_pgprot(struct page *page, unsigned long addr, pgprot_t prot)
  83 {
  84         enum fixed_addresses idx;
  85         unsigned int old_mmid;
  86         unsigned long vaddr, flags, entrylo;
  87         unsigned long old_ctx;
  88         pte_t pte;
  89         int tlbidx;
  90
  91         BUG_ON(Page_dcache_dirty(page));
  92
  93         preempt_disable();
  94         pagefault_disable();
  95         idx = (addr >> PAGE_SHIFT) & (FIX_N_COLOURS - 1);
  96         idx += in_interrupt() ? FIX_N_COLOURS : 0;
  97         vaddr = __fix_to_virt(FIX_CMAP_END - idx);
  98         pte = mk_pte(page, prot);
  99 #if defined(CONFIG_XPA)
 100         entrylo = pte_to_entrylo(pte.pte_high);
 101 #elif defined(CONFIG_PHYS_ADDR_T_64BIT) && defined(CONFIG_CPU_MIPS32)
 102         entrylo = pte.pte_high;
 103 #else
 104         entrylo = pte_to_entrylo(pte_val(pte));
 105 #endif
 106
 107         local_irq_save(flags);
 108         old_ctx = read_c0_entryhi();
 109         write_c0_entryhi(vaddr & (PAGE_MASK << 1));
 110         write_c0_entrylo0(entrylo);
 111         write_c0_entrylo1(entrylo);
 112         if (cpu_has_mmid) {
 113                 old_mmid = read_c0_memorymapid();
 114                 write_c0_memorymapid(MMID_KERNEL_WIRED);
 115         }
 116 #ifdef CONFIG_XPA
 117         if (cpu_has_xpa) {
 118                 entrylo = (pte.pte_low & _PFNX_MASK);
 119                 writex_c0_entrylo0(entrylo);
 120                 writex_c0_entrylo1(entrylo);
 121         }
 122 #endif
 123         tlbidx = num_wired_entries();
 124         write_c0_wired(tlbidx + 1);
 125         write_c0_index(tlbidx);
 126         mtc0_tlbw_hazard();
 127         tlb_write_indexed();
 128         tlbw_use_hazard();
 129         write_c0_entryhi(old_ctx);
 130         if (cpu_has_mmid)
 131                 write_c0_memorymapid(old_mmid);
 132         local_irq_restore(flags);
 133
 134         return (void*) vaddr;
 135 }
 136
 137 void *kmap_coherent(struct page *page, unsigned long addr)
 138 {
 139         return __kmap_pgprot(page, addr, PAGE_KERNEL);
 140 }
 141
 142 void *kmap_noncoherent(struct page *page, unsigned long addr)
 143 {
 144         return __kmap_pgprot(page, addr, PAGE_KERNEL_NC);
 145 }
 146
 147 void kunmap_coherent(void)
 148 {
 149         unsigned int wired;
 150         unsigned long flags, old_ctx;
 151
 152         local_irq_save(flags);
 153         old_ctx = read_c0_entryhi();
 154         wired = num_wired_entries() - 1;
 155         write_c0_wired(wired);
 156         write_c0_index(wired);
 157         write_c0_entryhi(UNIQUE_ENTRYHI(wired));
 158         write_c0_entrylo0(0);
 159         write_c0_entrylo1(0);
 160         mtc0_tlbw_hazard();
 161         tlb_write_indexed();
 162         tlbw_use_hazard();
 163         write_c0_entryhi(old_ctx);
 164         local_irq_restore(flags);
 165         pagefault_enable();
 166         preempt_enable();
 167 }
 168
 169 void copy_user_highpage(struct page *to, struct page *from,
 170         unsigned long vaddr, struct vm_area_struct *vma)
 171 {
 172         void *vfrom, *vto;
 173
 174         vto = kmap_atomic(to);
 175         if (cpu_has_dc_aliases &&
 176             page_mapcount(from) && !Page_dcache_dirty(from)) {
 177                 vfrom = kmap_coherent(from, vaddr);
 178                 copy_page(vto, vfrom);
 179                 kunmap_coherent();
 180         } else {
 181                 vfrom = kmap_atomic(from);
 182                 copy_page(vto, vfrom);
 183                 kunmap_atomic(vfrom);
 184         }
 185         if ((!cpu_has_ic_fills_f_dc) ||
 186             pages_do_alias((unsigned long)vto, vaddr & PAGE_MASK))
 187                 flush_data_cache_page((unsigned long)vto);
 188         kunmap_atomic(vto);
 189         /* Make sure this page is cleared on other CPU's too before using it */
 190         smp_wmb();
 191 }
 192
 193 void copy_to_user_page(struct vm_area_struct *vma,
 194         struct page *page, unsigned long vaddr, void *dst, const void *src,
 195         unsigned long len)
 196 {
 197         if (cpu_has_dc_aliases &&
 198             page_mapcount(page) && !Page_dcache_dirty(page)) {
 199                 void *vto = kmap_coherent(page, vaddr) + (vaddr & ~PAGE_MASK);
 200                 memcpy(vto, src, len);
 201                 kunmap_coherent();
 202         } else {
 203                 memcpy(dst, src, len);
 204                 if (cpu_has_dc_aliases)
 205                         SetPageDcacheDirty(page);
 206         }
 207         if (vma->vm_flags & VM_EXEC)
 208                 flush_cache_page(vma, vaddr, page_to_pfn(page));
 209 }
 210
 211 void copy_from_user_page(struct vm_area_struct *vma,
 212         struct page *page, unsigned long vaddr, void *dst, const void *src,
 213         unsigned long len)
 214 {
 215         if (cpu_has_dc_aliases &&
 216             page_mapcount(page) && !Page_dcache_dirty(page)) {
 217                 void *vfrom = kmap_coherent(page, vaddr) + (vaddr & ~PAGE_MASK);
 218                 memcpy(dst, vfrom, len);
 219                 kunmap_coherent();
 220         } else {
 221                 memcpy(dst, src, len);
 222                 if (cpu_has_dc_aliases)
 223                         SetPageDcacheDirty(page);
 224         }
 225 }
 226 EXPORT_SYMBOL_GPL(copy_from_user_page);
 227
 228 void __init fixrange_init(unsigned long start, unsigned long end,
 229         pgd_t *pgd_base)
 230 {
 231 #ifdef CONFIG_HIGHMEM
 232         pgd_t *pgd;
 233         pud_t *pud;
 234         pmd_t *pmd;
 235         pte_t *pte;
 236         int i, j, k;
 237         unsigned long vaddr;
 238
 239         vaddr = start;
 240         i = pgd_index(vaddr);
 241         j = pud_index(vaddr);
 242         k = pmd_index(vaddr);
 243         pgd = pgd_base + i;
 244
 245         for ( ; (i < PTRS_PER_PGD) && (vaddr < end); pgd++, i++) {
 246                 pud = (pud_t *)pgd;
 247                 for ( ; (j < PTRS_PER_PUD) && (vaddr < end); pud++, j++) {
 248                         pmd = (pmd_t *)pud;
 249                         for (; (k < PTRS_PER_PMD) && (vaddr < end); pmd++, k++) {
 250                                 if (pmd_none(*pmd)) {
 251                                         pte = (pte_t *) memblock_alloc_low(PAGE_SIZE,
 252                                                                            PAGE_SIZE);
 253                                         if (!pte)
 254                                                 panic("%s: Failed to allocate %lu bytes align=%lx\n",
 255                                                       __func__, PAGE_SIZE,
 256                                                       PAGE_SIZE);
 257
 258                                         set_pmd(pmd, __pmd((unsigned long)pte));
 259                                         BUG_ON(pte != pte_offset_kernel(pmd, 0));
 260                                 }
 261                                 vaddr += PMD_SIZE;
 262                         }
 263                         k = 0;
 264                 }
 265                 j = 0;
 266         }
 267 #endif
 268 }
 269
 270 struct maar_walk_info {
 271         struct maar_config cfg[16];
 272         unsigned int num_cfg;
 273 };
 274
 275 static int maar_res_walk(unsigned long start_pfn, unsigned long nr_pages,
 276                          void *data)
 277 {
 278         struct maar_walk_info *wi = data;
 279         struct maar_config *cfg = &wi->cfg[wi->num_cfg];
 280         unsigned int maar_align;
 281
 282         /* MAAR registers hold physical addresses right shifted by 4 bits */
 283         maar_align = BIT(MIPS_MAAR_ADDR_SHIFT + 4);
 284
 285         /* Fill in the MAAR config entry */
 286         cfg->lower = ALIGN(PFN_PHYS(start_pfn), maar_align);
 287         cfg->upper = ALIGN_DOWN(PFN_PHYS(start_pfn + nr_pages), maar_align) - 1;
 288         cfg->attrs = MIPS_MAAR_S;
 289
 290         /* Ensure we don't overflow the cfg array */
 291         if (!WARN_ON(wi->num_cfg >= ARRAY_SIZE(wi->cfg)))
 292                 wi->num_cfg++;
 293
 294         return 0;
 295 }
 296
 297
 298 unsigned __weak platform_maar_init(unsigned num_pairs)
 299 {
 300         unsigned int num_configured;
 301         struct maar_walk_info wi;
 302
 303         wi.num_cfg = 0;
 304         walk_system_ram_range(0, max_pfn, &wi, maar_res_walk);
 305
 306         num_configured = maar_config(wi.cfg, wi.num_cfg, num_pairs);
 307         if (num_configured < wi.num_cfg)
 308                 pr_warn("Not enough MAAR pairs (%u) for all memory regions (%u)\n",
 309                         num_pairs, wi.num_cfg);
 310
 311         return num_configured;
 312 }
 313
 314 void maar_init(void)
 315 {
 316         unsigned num_maars, used, i;
 317         phys_addr_t lower, upper, attr;
 318         static struct {
 319                 struct maar_config cfgs[3];
 320                 unsigned used;
 321         } recorded = { { { 0 } }, 0 };
 322
 323         if (!cpu_has_maar)
 324                 return;
 325
 326         /* Detect the number of MAARs */
 327         write_c0_maari(~0);
 328         back_to_back_c0_hazard();
 329         num_maars = read_c0_maari() + 1;
 330
 331         /* MAARs should be in pairs */
 332         WARN_ON(num_maars % 2);
 333
 334         /* Set MAARs using values we recorded already */
 335         if (recorded.used) {
 336                 used = maar_config(recorded.cfgs, recorded.used, num_maars / 2);
 337                 BUG_ON(used != recorded.used);
 338         } else {
 339                 /* Configure the required MAARs */
 340                 used = platform_maar_init(num_maars / 2);
 341         }
 342
 343         /* Disable any further MAARs */
 344         for (i = (used * 2); i < num_maars; i++) {
 345                 write_c0_maari(i);
 346                 back_to_back_c0_hazard();
 347                 write_c0_maar(0);
 348                 back_to_back_c0_hazard();
 349         }
 350
 351         if (recorded.used)
 352                 return;
 353
 354         pr_info("MAAR configuration:\n");
 355         for (i = 0; i < num_maars; i += 2) {
 356                 write_c0_maari(i);
 357                 back_to_back_c0_hazard();
 358                 upper = read_c0_maar();
 359 #ifdef CONFIG_XPA
 360                 upper |= (phys_addr_t)readx_c0_maar() << MIPS_MAARX_ADDR_SHIFT;
 361 #endif
 362
 363                 write_c0_maari(i + 1);
 364                 back_to_back_c0_hazard();
 365                 lower = read_c0_maar();
 366 #ifdef CONFIG_XPA
 367                 lower |= (phys_addr_t)readx_c0_maar() << MIPS_MAARX_ADDR_SHIFT;
 368 #endif
 369
 370                 attr = lower & upper;
 371                 lower = (lower & MIPS_MAAR_ADDR) << 4;
 372                 upper = ((upper & MIPS_MAAR_ADDR) << 4) | 0xffff;
 373
 374                 pr_info("  [%d]: ", i / 2);
 375                 if ((attr & MIPS_MAAR_V) != MIPS_MAAR_V) {
 376                         pr_cont("disabled\n");
 377                         continue;
 378                 }
 379
 380                 pr_cont("%pa-%pa", &lower, &upper);
 381
 382                 if (attr & MIPS_MAAR_S)
 383                         pr_cont(" speculate");
 384
 385                 pr_cont("\n");
 386
 387                 /* Record the setup for use on secondary CPUs */
 388                 if (used <= ARRAY_SIZE(recorded.cfgs)) {
 389                         recorded.cfgs[recorded.used].lower = lower;
 390                         recorded.cfgs[recorded.used].upper = upper;
 391                         recorded.cfgs[recorded.used].attrs = attr;
 392                         recorded.used++;
 393                 }
 394         }
 395 }
 396
 397 #ifndef CONFIG_NUMA
 398 void __init paging_init(void)
 399 {
 400         unsigned long max_zone_pfns[MAX_NR_ZONES];
 401
 402         pagetable_init();
 403
 404 #ifdef CONFIG_ZONE_DMA
 405         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 406 #endif
 407 #ifdef CONFIG_ZONE_DMA32
 408         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 409 #endif
 410         max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
 411 #ifdef CONFIG_HIGHMEM
 412         max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
 413
 414         if (cpu_has_dc_aliases && max_low_pfn != highend_pfn) {
 415                 printk(KERN_WARNING "This processor doesn't support highmem."
 416                        " %ldk highmem ignored\n",
 417                        (highend_pfn - max_low_pfn) << (PAGE_SHIFT - 10));
 418                 max_zone_pfns[ZONE_HIGHMEM] = max_low_pfn;
 419         }
 420 #endif
 421
 422         free_area_init(max_zone_pfns);
 423 }
 424
 425 #ifdef CONFIG_64BIT
 426 static struct kcore_list kcore_kseg0;
 427 #endif
 428
 429 static inline void __init mem_init_free_highmem(void)
 430 {
 431 #ifdef CONFIG_HIGHMEM
 432         unsigned long tmp;
 433
 434         if (cpu_has_dc_aliases)
 435                 return;
 436
 437         for (tmp = highstart_pfn; tmp < highend_pfn; tmp++) {
 438                 struct page *page = pfn_to_page(tmp);
 439
 440                 if (!memblock_is_memory(PFN_PHYS(tmp)))
 441                         SetPageReserved(page);
 442                 else
 443                         free_highmem_page(page);
 444         }
 445 #endif
 446 }
 447
 448 void __init mem_init(void)
 449 {
 450         /*
 451          * When _PFN_SHIFT is greater than PAGE_SHIFT we won't have enough PTE
 452          * bits to hold a full 32b physical address on MIPS32 systems.
 453          */
 454         BUILD_BUG_ON(IS_ENABLED(CONFIG_32BIT) && (_PFN_SHIFT > PAGE_SHIFT));
 455
 456 #ifdef CONFIG_HIGHMEM
 457         max_mapnr = highend_pfn ? highend_pfn : max_low_pfn;
 458 #else
 459         max_mapnr = max_low_pfn;
 460 #endif
 461         high_memory = (void *) __va(max_low_pfn << PAGE_SHIFT);
 462
 463         maar_init();
 464         memblock_free_all();
 465         setup_zero_pages();     /* Setup zeroed pages.  */
 466         mem_init_free_highmem();
 467
 468 #ifdef CONFIG_64BIT
 469         if ((unsigned long) &_text > (unsigned long) CKSEG0)
 470                 /* The -4 is a hack so that user tools don't have to handle
 471                    the overflow.  */
 472                 kclist_add(&kcore_kseg0, (void *) CKSEG0,
 473                                 0x80000000 - 4, KCORE_TEXT);
 474 #endif
 475 }
 476 #endif /* !CONFIG_NUMA */
 477
 478 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
 479 {
 480         unsigned long pfn;
 481
 482         for (pfn = PFN_UP(begin); pfn < PFN_DOWN(end); pfn++) {
 483                 struct page *page = pfn_to_page(pfn);
 484                 void *addr = phys_to_virt(PFN_PHYS(pfn));
 485
 486                 memset(addr, POISON_FREE_INITMEM, PAGE_SIZE);
 487                 free_reserved_page(page);
 488         }
 489         printk(KERN_INFO "Freeing %s: %ldk freed\n", what, (end - begin) >> 10);
 490 }
 491
 492 void (*free_init_pages_eva)(void *begin, void *end) = NULL;
 493
 494 void __weak __init prom_free_prom_memory(void)
 495 {
 496         /* nothing to do */
 497 }
 498
 499 void __ref free_initmem(void)
 500 {
 501         prom_free_prom_memory();
 502         /*
 503          * Let the platform define a specific function to free the
 504          * init section since EVA may have used any possible mapping
 505          * between virtual and physical addresses.
 506          */
 507         if (free_init_pages_eva)
 508                 free_init_pages_eva((void *)&__init_begin, (void *)&__init_end);
 509         else
 510                 free_initmem_default(POISON_FREE_INITMEM);
 511 }
 512
 513 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 514 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 515 EXPORT_SYMBOL(__per_cpu_offset);
 516
 517 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 518 {
 519         return node_distance(cpu_to_node(from), cpu_to_node(to));
 520 }
 521
 522 static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size,
 523                                        size_t align)
 524 {
 525         return memblock_alloc_try_nid(size, align, __pa(MAX_DMA_ADDRESS),
 526                                       MEMBLOCK_ALLOC_ACCESSIBLE,
 527                                       cpu_to_node(cpu));
 528 }
 529
 530 static void __init pcpu_fc_free(void *ptr, size_t size)
 531 {
 532         memblock_free_early(__pa(ptr), size);
 533 }
 534
 535 void __init setup_per_cpu_areas(void)
 536 {
 537         unsigned long delta;
 538         unsigned int cpu;
 539         int rc;
 540
 541         /*
 542          * Always reserve area for module percpu variables.  That's
 543          * what the legacy allocator did.
 544          */
 545         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
 546                                     PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
 547                                     pcpu_cpu_distance,
 548                                     pcpu_fc_alloc, pcpu_fc_free);
 549         if (rc < 0)
 550                 panic("Failed to initialize percpu areas.");
 551
 552         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 553         for_each_possible_cpu(cpu)
 554                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
 555 }
 556 #endif
 557
 558 #ifndef CONFIG_MIPS_PGD_C0_CONTEXT
 559 unsigned long pgd_current[NR_CPUS];
 560 #endif
 561
 562 /*
 563  * Align swapper_pg_dir in to 64K, allows its address to be loaded
 564  * with a single LUI instruction in the TLB handlers.  If we used
 565  * __aligned(64K), its size would get rounded up to the alignment
 566  * size, and waste space.  So we place it in its own section and align
 567  * it in the linker script.
 568  */
 569 pgd_t swapper_pg_dir[PTRS_PER_PGD] __section(".bss..swapper_pg_dir");
 570 #ifndef __PAGETABLE_PUD_FOLDED
 571 pud_t invalid_pud_table[PTRS_PER_PUD] __page_aligned_bss;
 572 #endif
 573 #ifndef __PAGETABLE_PMD_FOLDED
 574 pmd_t invalid_pmd_table[PTRS_PER_PMD] __page_aligned_bss;
 575 EXPORT_SYMBOL_GPL(invalid_pmd_table);
 576 #endif
 577 pte_t invalid_pte_table[PTRS_PER_PTE] __page_aligned_bss;
 578 EXPORT_SYMBOL(invalid_pte_table);