1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * srmmu.c: SRMMU specific routines for memory management. 4 * 5 * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu) 6 * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com) 7 * Copyright (C) 1996 Eddie C. Dost (ecd@skynet.be) 8 * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz) 9 * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org) 10 */ 11 12 #include <linux/seq_file.h> 13 #include <linux/spinlock.h> 14 #include <linux/memblock.h> 15 #include <linux/pagemap.h> 16 #include <linux/vmalloc.h> 17 #include <linux/kdebug.h> 18 #include <linux/export.h> 19 #include <linux/kernel.h> 20 #include <linux/init.h> 21 #include <linux/log2.h> 22 #include <linux/gfp.h> 23 #include <linux/fs.h> 24 #include <linux/mm.h> 25 26 #include <asm/mmu_context.h> 27 #include <asm/cacheflush.h> 28 #include <asm/tlbflush.h> 29 #include <asm/io-unit.h> 30 #include <asm/pgalloc.h> 31 #include <asm/pgtable.h> 32 #include <asm/bitext.h> 33 #include <asm/vaddrs.h> 34 #include <asm/cache.h> 35 #include <asm/traps.h> 36 #include <asm/oplib.h> 37 #include <asm/mbus.h> 38 #include <asm/page.h> 39 #include <asm/asi.h> 40 #include <asm/smp.h> 41 #include <asm/io.h> 42 43 /* Now the cpu specific definitions. */ 44 #include <asm/turbosparc.h> 45 #include <asm/tsunami.h> 46 #include <asm/viking.h> 47 #include <asm/swift.h> 48 #include <asm/leon.h> 49 #include <asm/mxcc.h> 50 #include <asm/ross.h> 51 52 #include "mm_32.h" 53 54 enum mbus_module srmmu_modtype; 55 static unsigned int hwbug_bitmask; 56 int vac_cache_size; 57 EXPORT_SYMBOL(vac_cache_size); 58 int vac_line_size; 59 60 extern struct resource sparc_iomap; 61 62 extern unsigned long last_valid_pfn; 63 64 static pgd_t *srmmu_swapper_pg_dir; 65 66 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops; 67 EXPORT_SYMBOL(sparc32_cachetlb_ops); 68 69 #ifdef CONFIG_SMP 70 const struct sparc32_cachetlb_ops *local_ops; 71 72 #define FLUSH_BEGIN(mm) 73 #define FLUSH_END 74 #else 75 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) { 76 #define FLUSH_END } 77 #endif 78 79 int flush_page_for_dma_global = 1; 80 81 char *srmmu_name; 82 83 ctxd_t *srmmu_ctx_table_phys; 84 static ctxd_t *srmmu_context_table; 85 86 int viking_mxcc_present; 87 static DEFINE_SPINLOCK(srmmu_context_spinlock); 88 89 static int is_hypersparc; 90 91 static int srmmu_cache_pagetables; 92 93 /* these will be initialized in srmmu_nocache_calcsize() */ 94 static unsigned long srmmu_nocache_size; 95 static unsigned long srmmu_nocache_end; 96 97 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */ 98 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4) 99 100 /* The context table is a nocache user with the biggest alignment needs. */ 101 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS) 102 103 void *srmmu_nocache_pool; 104 static struct bit_map srmmu_nocache_map; 105 106 static inline int srmmu_pmd_none(pmd_t pmd) 107 { return !(pmd_val(pmd) & 0xFFFFFFF); } 108 109 /* XXX should we hyper_flush_whole_icache here - Anton */ 110 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp) 111 { 112 pte_t pte; 113 114 pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4))); 115 set_pte((pte_t *)ctxp, pte); 116 } 117 118 /* 119 * Locations of MSI Registers. 120 */ 121 #define MSI_MBUS_ARBEN 0xe0001008 /* MBus Arbiter Enable register */ 122 123 /* 124 * Useful bits in the MSI Registers. 125 */ 126 #define MSI_ASYNC_MODE 0x80000000 /* Operate the MSI asynchronously */ 127 128 static void msi_set_sync(void) 129 { 130 __asm__ __volatile__ ("lda [%0] %1, %%g3\n\t" 131 "andn %%g3, %2, %%g3\n\t" 132 "sta %%g3, [%0] %1\n\t" : : 133 "r" (MSI_MBUS_ARBEN), 134 "i" (ASI_M_CTL), "r" (MSI_ASYNC_MODE) : "g3"); 135 } 136 137 void pmd_set(pmd_t *pmdp, pte_t *ptep) 138 { 139 unsigned long ptp = __nocache_pa(ptep) >> 4; 140 set_pte((pte_t *)&pmd_val(*pmdp), __pte(SRMMU_ET_PTD | ptp)); 141 } 142 143 /* 144 * size: bytes to allocate in the nocache area. 145 * align: bytes, number to align at. 146 * Returns the virtual address of the allocated area. 147 */ 148 static void *__srmmu_get_nocache(int size, int align) 149 { 150 int offset, minsz = 1 << SRMMU_NOCACHE_BITMAP_SHIFT; 151 unsigned long addr; 152 153 if (size < minsz) { 154 printk(KERN_ERR "Size 0x%x too small for nocache request\n", 155 size); 156 size = minsz; 157 } 158 if (size & (minsz - 1)) { 159 printk(KERN_ERR "Size 0x%x unaligned in nocache request\n", 160 size); 161 size += minsz - 1; 162 } 163 BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX); 164 165 offset = bit_map_string_get(&srmmu_nocache_map, 166 size >> SRMMU_NOCACHE_BITMAP_SHIFT, 167 align >> SRMMU_NOCACHE_BITMAP_SHIFT); 168 if (offset == -1) { 169 printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n", 170 size, (int) srmmu_nocache_size, 171 srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT); 172 return NULL; 173 } 174 175 addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT); 176 return (void *)addr; 177 } 178 179 void *srmmu_get_nocache(int size, int align) 180 { 181 void *tmp; 182 183 tmp = __srmmu_get_nocache(size, align); 184 185 if (tmp) 186 memset(tmp, 0, size); 187 188 return tmp; 189 } 190 191 void srmmu_free_nocache(void *addr, int size) 192 { 193 unsigned long vaddr; 194 int offset; 195 196 vaddr = (unsigned long)addr; 197 if (vaddr < SRMMU_NOCACHE_VADDR) { 198 printk("Vaddr %lx is smaller than nocache base 0x%lx\n", 199 vaddr, (unsigned long)SRMMU_NOCACHE_VADDR); 200 BUG(); 201 } 202 if (vaddr + size > srmmu_nocache_end) { 203 printk("Vaddr %lx is bigger than nocache end 0x%lx\n", 204 vaddr, srmmu_nocache_end); 205 BUG(); 206 } 207 if (!is_power_of_2(size)) { 208 printk("Size 0x%x is not a power of 2\n", size); 209 BUG(); 210 } 211 if (size < SRMMU_NOCACHE_BITMAP_SHIFT) { 212 printk("Size 0x%x is too small\n", size); 213 BUG(); 214 } 215 if (vaddr & (size - 1)) { 216 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size); 217 BUG(); 218 } 219 220 offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT; 221 size = size >> SRMMU_NOCACHE_BITMAP_SHIFT; 222 223 bit_map_clear(&srmmu_nocache_map, offset, size); 224 } 225 226 static void srmmu_early_allocate_ptable_skeleton(unsigned long start, 227 unsigned long end); 228 229 /* Return how much physical memory we have. */ 230 static unsigned long __init probe_memory(void) 231 { 232 unsigned long total = 0; 233 int i; 234 235 for (i = 0; sp_banks[i].num_bytes; i++) 236 total += sp_banks[i].num_bytes; 237 238 return total; 239 } 240 241 /* 242 * Reserve nocache dynamically proportionally to the amount of 243 * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002 244 */ 245 static void __init srmmu_nocache_calcsize(void) 246 { 247 unsigned long sysmemavail = probe_memory() / 1024; 248 int srmmu_nocache_npages; 249 250 srmmu_nocache_npages = 251 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256; 252 253 /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */ 254 // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256; 255 if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES) 256 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES; 257 258 /* anything above 1280 blows up */ 259 if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES) 260 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES; 261 262 srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE; 263 srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size; 264 } 265 266 static void __init srmmu_nocache_init(void) 267 { 268 void *srmmu_nocache_bitmap; 269 unsigned int bitmap_bits; 270 pgd_t *pgd; 271 p4d_t *p4d; 272 pud_t *pud; 273 pmd_t *pmd; 274 pte_t *pte; 275 unsigned long paddr, vaddr; 276 unsigned long pteval; 277 278 bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT; 279 280 srmmu_nocache_pool = memblock_alloc_or_panic(srmmu_nocache_size, 281 SRMMU_NOCACHE_ALIGN_MAX); 282 memset(srmmu_nocache_pool, 0, srmmu_nocache_size); 283 284 srmmu_nocache_bitmap = 285 memblock_alloc_or_panic(BITS_TO_LONGS(bitmap_bits) * sizeof(long), 286 SMP_CACHE_BYTES); 287 bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits); 288 289 srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE); 290 memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE); 291 init_mm.pgd = srmmu_swapper_pg_dir; 292 293 srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end); 294 295 paddr = __pa((unsigned long)srmmu_nocache_pool); 296 vaddr = SRMMU_NOCACHE_VADDR; 297 298 while (vaddr < srmmu_nocache_end) { 299 pgd = pgd_offset_k(vaddr); 300 p4d = p4d_offset(pgd, vaddr); 301 pud = pud_offset(p4d, vaddr); 302 pmd = pmd_offset(__nocache_fix(pud), vaddr); 303 pte = pte_offset_kernel(__nocache_fix(pmd), vaddr); 304 305 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV); 306 307 if (srmmu_cache_pagetables) 308 pteval |= SRMMU_CACHE; 309 310 set_pte(__nocache_fix(pte), __pte(pteval)); 311 312 vaddr += PAGE_SIZE; 313 paddr += PAGE_SIZE; 314 } 315 316 flush_cache_all(); 317 flush_tlb_all(); 318 } 319 320 pgd_t *get_pgd_fast(void) 321 { 322 pgd_t *pgd = NULL; 323 324 pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE); 325 if (pgd) { 326 pgd_t *init = pgd_offset_k(0); 327 memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t)); 328 memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD, 329 (PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t)); 330 } 331 332 return pgd; 333 } 334 335 /* 336 * Hardware needs alignment to 256 only, but we align to whole page size 337 * to reduce fragmentation problems due to the buddy principle. 338 * XXX Provide actual fragmentation statistics in /proc. 339 * 340 * Alignments up to the page size are the same for physical and virtual 341 * addresses of the nocache area. 342 */ 343 pgtable_t pte_alloc_one(struct mm_struct *mm) 344 { 345 pte_t *ptep; 346 struct page *page; 347 348 if (!(ptep = pte_alloc_one_kernel(mm))) 349 return NULL; 350 page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT); 351 spin_lock(&mm->page_table_lock); 352 if (page_ref_inc_return(page) == 2 && 353 !pagetable_pte_ctor(mm, page_ptdesc(page))) { 354 page_ref_dec(page); 355 ptep = NULL; 356 } 357 spin_unlock(&mm->page_table_lock); 358 359 return ptep; 360 } 361 362 void pte_free(struct mm_struct *mm, pgtable_t ptep) 363 { 364 struct page *page; 365 366 page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT); 367 spin_lock(&mm->page_table_lock); 368 if (page_ref_dec_return(page) == 1) 369 pagetable_dtor(page_ptdesc(page)); 370 spin_unlock(&mm->page_table_lock); 371 372 srmmu_free_nocache(ptep, SRMMU_PTE_TABLE_SIZE); 373 } 374 375 /* context handling - a dynamically sized pool is used */ 376 #define NO_CONTEXT -1 377 378 struct ctx_list { 379 struct ctx_list *next; 380 struct ctx_list *prev; 381 unsigned int ctx_number; 382 struct mm_struct *ctx_mm; 383 }; 384 385 static struct ctx_list *ctx_list_pool; 386 static struct ctx_list ctx_free; 387 static struct ctx_list ctx_used; 388 389 /* At boot time we determine the number of contexts */ 390 static int num_contexts; 391 392 static inline void remove_from_ctx_list(struct ctx_list *entry) 393 { 394 entry->next->prev = entry->prev; 395 entry->prev->next = entry->next; 396 } 397 398 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry) 399 { 400 entry->next = head; 401 (entry->prev = head->prev)->next = entry; 402 head->prev = entry; 403 } 404 #define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry) 405 #define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry) 406 407 408 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm) 409 { 410 struct ctx_list *ctxp; 411 412 ctxp = ctx_free.next; 413 if (ctxp != &ctx_free) { 414 remove_from_ctx_list(ctxp); 415 add_to_used_ctxlist(ctxp); 416 mm->context = ctxp->ctx_number; 417 ctxp->ctx_mm = mm; 418 return; 419 } 420 ctxp = ctx_used.next; 421 if (ctxp->ctx_mm == old_mm) 422 ctxp = ctxp->next; 423 if (ctxp == &ctx_used) 424 panic("out of mmu contexts"); 425 flush_cache_mm(ctxp->ctx_mm); 426 flush_tlb_mm(ctxp->ctx_mm); 427 remove_from_ctx_list(ctxp); 428 add_to_used_ctxlist(ctxp); 429 ctxp->ctx_mm->context = NO_CONTEXT; 430 ctxp->ctx_mm = mm; 431 mm->context = ctxp->ctx_number; 432 } 433 434 static inline void free_context(int context) 435 { 436 struct ctx_list *ctx_old; 437 438 ctx_old = ctx_list_pool + context; 439 remove_from_ctx_list(ctx_old); 440 add_to_free_ctxlist(ctx_old); 441 } 442 443 static void __init sparc_context_init(int numctx) 444 { 445 int ctx; 446 unsigned long size; 447 448 size = numctx * sizeof(struct ctx_list); 449 ctx_list_pool = memblock_alloc_or_panic(size, SMP_CACHE_BYTES); 450 451 for (ctx = 0; ctx < numctx; ctx++) { 452 struct ctx_list *clist; 453 454 clist = (ctx_list_pool + ctx); 455 clist->ctx_number = ctx; 456 clist->ctx_mm = NULL; 457 } 458 ctx_free.next = ctx_free.prev = &ctx_free; 459 ctx_used.next = ctx_used.prev = &ctx_used; 460 for (ctx = 0; ctx < numctx; ctx++) 461 add_to_free_ctxlist(ctx_list_pool + ctx); 462 } 463 464 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, 465 struct task_struct *tsk) 466 { 467 unsigned long flags; 468 469 if (mm->context == NO_CONTEXT) { 470 spin_lock_irqsave(&srmmu_context_spinlock, flags); 471 alloc_context(old_mm, mm); 472 spin_unlock_irqrestore(&srmmu_context_spinlock, flags); 473 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd); 474 } 475 476 if (sparc_cpu_model == sparc_leon) 477 leon_switch_mm(); 478 479 if (is_hypersparc) 480 hyper_flush_whole_icache(); 481 482 srmmu_set_context(mm->context); 483 } 484 485 /* Low level IO area allocation on the SRMMU. */ 486 static inline void srmmu_mapioaddr(unsigned long physaddr, 487 unsigned long virt_addr, int bus_type) 488 { 489 pgd_t *pgdp; 490 p4d_t *p4dp; 491 pud_t *pudp; 492 pmd_t *pmdp; 493 pte_t *ptep; 494 unsigned long tmp; 495 496 physaddr &= PAGE_MASK; 497 pgdp = pgd_offset_k(virt_addr); 498 p4dp = p4d_offset(pgdp, virt_addr); 499 pudp = pud_offset(p4dp, virt_addr); 500 pmdp = pmd_offset(pudp, virt_addr); 501 ptep = pte_offset_kernel(pmdp, virt_addr); 502 tmp = (physaddr >> 4) | SRMMU_ET_PTE; 503 504 /* I need to test whether this is consistent over all 505 * sun4m's. The bus_type represents the upper 4 bits of 506 * 36-bit physical address on the I/O space lines... 507 */ 508 tmp |= (bus_type << 28); 509 tmp |= SRMMU_PRIV; 510 __flush_page_to_ram(virt_addr); 511 set_pte(ptep, __pte(tmp)); 512 } 513 514 void srmmu_mapiorange(unsigned int bus, unsigned long xpa, 515 unsigned long xva, unsigned int len) 516 { 517 while (len != 0) { 518 len -= PAGE_SIZE; 519 srmmu_mapioaddr(xpa, xva, bus); 520 xva += PAGE_SIZE; 521 xpa += PAGE_SIZE; 522 } 523 flush_tlb_all(); 524 } 525 526 static inline void srmmu_unmapioaddr(unsigned long virt_addr) 527 { 528 pgd_t *pgdp; 529 p4d_t *p4dp; 530 pud_t *pudp; 531 pmd_t *pmdp; 532 pte_t *ptep; 533 534 535 pgdp = pgd_offset_k(virt_addr); 536 p4dp = p4d_offset(pgdp, virt_addr); 537 pudp = pud_offset(p4dp, virt_addr); 538 pmdp = pmd_offset(pudp, virt_addr); 539 ptep = pte_offset_kernel(pmdp, virt_addr); 540 541 /* No need to flush uncacheable page. */ 542 __pte_clear(ptep); 543 } 544 545 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len) 546 { 547 while (len != 0) { 548 len -= PAGE_SIZE; 549 srmmu_unmapioaddr(virt_addr); 550 virt_addr += PAGE_SIZE; 551 } 552 flush_tlb_all(); 553 } 554 555 /* tsunami.S */ 556 extern void tsunami_flush_cache_all(void); 557 extern void tsunami_flush_cache_mm(struct mm_struct *mm); 558 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end); 559 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page); 560 extern void tsunami_flush_page_to_ram(unsigned long page); 561 extern void tsunami_flush_page_for_dma(unsigned long page); 562 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr); 563 extern void tsunami_flush_tlb_all(void); 564 extern void tsunami_flush_tlb_mm(struct mm_struct *mm); 565 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end); 566 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page); 567 extern void tsunami_setup_blockops(void); 568 569 /* swift.S */ 570 extern void swift_flush_cache_all(void); 571 extern void swift_flush_cache_mm(struct mm_struct *mm); 572 extern void swift_flush_cache_range(struct vm_area_struct *vma, 573 unsigned long start, unsigned long end); 574 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page); 575 extern void swift_flush_page_to_ram(unsigned long page); 576 extern void swift_flush_page_for_dma(unsigned long page); 577 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr); 578 extern void swift_flush_tlb_all(void); 579 extern void swift_flush_tlb_mm(struct mm_struct *mm); 580 extern void swift_flush_tlb_range(struct vm_area_struct *vma, 581 unsigned long start, unsigned long end); 582 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page); 583 584 #if 0 /* P3: deadwood to debug precise flushes on Swift. */ 585 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page) 586 { 587 int cctx, ctx1; 588 589 page &= PAGE_MASK; 590 if ((ctx1 = vma->vm_mm->context) != -1) { 591 cctx = srmmu_get_context(); 592 /* Is context # ever different from current context? P3 */ 593 if (cctx != ctx1) { 594 printk("flush ctx %02x curr %02x\n", ctx1, cctx); 595 srmmu_set_context(ctx1); 596 swift_flush_page(page); 597 __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : : 598 "r" (page), "i" (ASI_M_FLUSH_PROBE)); 599 srmmu_set_context(cctx); 600 } else { 601 /* Rm. prot. bits from virt. c. */ 602 /* swift_flush_cache_all(); */ 603 /* swift_flush_cache_page(vma, page); */ 604 swift_flush_page(page); 605 606 __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : : 607 "r" (page), "i" (ASI_M_FLUSH_PROBE)); 608 /* same as above: srmmu_flush_tlb_page() */ 609 } 610 } 611 } 612 #endif 613 614 /* 615 * The following are all MBUS based SRMMU modules, and therefore could 616 * be found in a multiprocessor configuration. On the whole, these 617 * chips seems to be much more touchy about DVMA and page tables 618 * with respect to cache coherency. 619 */ 620 621 /* viking.S */ 622 extern void viking_flush_cache_all(void); 623 extern void viking_flush_cache_mm(struct mm_struct *mm); 624 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start, 625 unsigned long end); 626 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page); 627 extern void viking_flush_page_to_ram(unsigned long page); 628 extern void viking_flush_page_for_dma(unsigned long page); 629 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr); 630 extern void viking_flush_page(unsigned long page); 631 extern void viking_mxcc_flush_page(unsigned long page); 632 extern void viking_flush_tlb_all(void); 633 extern void viking_flush_tlb_mm(struct mm_struct *mm); 634 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, 635 unsigned long end); 636 extern void viking_flush_tlb_page(struct vm_area_struct *vma, 637 unsigned long page); 638 extern void sun4dsmp_flush_tlb_all(void); 639 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm); 640 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, 641 unsigned long end); 642 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma, 643 unsigned long page); 644 645 /* hypersparc.S */ 646 extern void hypersparc_flush_cache_all(void); 647 extern void hypersparc_flush_cache_mm(struct mm_struct *mm); 648 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end); 649 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page); 650 extern void hypersparc_flush_page_to_ram(unsigned long page); 651 extern void hypersparc_flush_page_for_dma(unsigned long page); 652 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr); 653 extern void hypersparc_flush_tlb_all(void); 654 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm); 655 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end); 656 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page); 657 extern void hypersparc_setup_blockops(void); 658 659 /* 660 * NOTE: All of this startup code assumes the low 16mb (approx.) of 661 * kernel mappings are done with one single contiguous chunk of 662 * ram. On small ram machines (classics mainly) we only get 663 * around 8mb mapped for us. 664 */ 665 666 static void __init early_pgtable_allocfail(char *type) 667 { 668 prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type); 669 prom_halt(); 670 } 671 672 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start, 673 unsigned long end) 674 { 675 pgd_t *pgdp; 676 p4d_t *p4dp; 677 pud_t *pudp; 678 pmd_t *pmdp; 679 pte_t *ptep; 680 681 while (start < end) { 682 pgdp = pgd_offset_k(start); 683 p4dp = p4d_offset(pgdp, start); 684 pudp = pud_offset(p4dp, start); 685 if (pud_none(*__nocache_fix(pudp))) { 686 pmdp = __srmmu_get_nocache( 687 SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE); 688 if (pmdp == NULL) 689 early_pgtable_allocfail("pmd"); 690 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE); 691 pud_set(__nocache_fix(pudp), pmdp); 692 } 693 pmdp = pmd_offset(__nocache_fix(pudp), start); 694 if (srmmu_pmd_none(*__nocache_fix(pmdp))) { 695 ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE); 696 if (ptep == NULL) 697 early_pgtable_allocfail("pte"); 698 memset(__nocache_fix(ptep), 0, PTE_SIZE); 699 pmd_set(__nocache_fix(pmdp), ptep); 700 } 701 if (start > (0xffffffffUL - PMD_SIZE)) 702 break; 703 start = (start + PMD_SIZE) & PMD_MASK; 704 } 705 } 706 707 static void __init srmmu_allocate_ptable_skeleton(unsigned long start, 708 unsigned long end) 709 { 710 pgd_t *pgdp; 711 p4d_t *p4dp; 712 pud_t *pudp; 713 pmd_t *pmdp; 714 pte_t *ptep; 715 716 while (start < end) { 717 pgdp = pgd_offset_k(start); 718 p4dp = p4d_offset(pgdp, start); 719 pudp = pud_offset(p4dp, start); 720 if (pud_none(*pudp)) { 721 pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE); 722 if (pmdp == NULL) 723 early_pgtable_allocfail("pmd"); 724 memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE); 725 pud_set((pud_t *)pgdp, pmdp); 726 } 727 pmdp = pmd_offset(pudp, start); 728 if (srmmu_pmd_none(*pmdp)) { 729 ptep = __srmmu_get_nocache(PTE_SIZE, 730 PTE_SIZE); 731 if (ptep == NULL) 732 early_pgtable_allocfail("pte"); 733 memset(ptep, 0, PTE_SIZE); 734 pmd_set(pmdp, ptep); 735 } 736 if (start > (0xffffffffUL - PMD_SIZE)) 737 break; 738 start = (start + PMD_SIZE) & PMD_MASK; 739 } 740 } 741 742 /* These flush types are not available on all chips... */ 743 static inline unsigned long srmmu_probe(unsigned long vaddr) 744 { 745 unsigned long retval; 746 747 if (sparc_cpu_model != sparc_leon) { 748 749 vaddr &= PAGE_MASK; 750 __asm__ __volatile__("lda [%1] %2, %0\n\t" : 751 "=r" (retval) : 752 "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE)); 753 } else { 754 retval = leon_swprobe(vaddr, NULL); 755 } 756 return retval; 757 } 758 759 /* 760 * This is much cleaner than poking around physical address space 761 * looking at the prom's page table directly which is what most 762 * other OS's do. Yuck... this is much better. 763 */ 764 static void __init srmmu_inherit_prom_mappings(unsigned long start, 765 unsigned long end) 766 { 767 unsigned long probed; 768 unsigned long addr; 769 pgd_t *pgdp; 770 p4d_t *p4dp; 771 pud_t *pudp; 772 pmd_t *pmdp; 773 pte_t *ptep; 774 int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */ 775 776 while (start <= end) { 777 if (start == 0) 778 break; /* probably wrap around */ 779 if (start == 0xfef00000) 780 start = KADB_DEBUGGER_BEGVM; 781 probed = srmmu_probe(start); 782 if (!probed) { 783 /* continue probing until we find an entry */ 784 start += PAGE_SIZE; 785 continue; 786 } 787 788 /* A red snapper, see what it really is. */ 789 what = 0; 790 addr = start - PAGE_SIZE; 791 792 if (!(start & ~(PMD_MASK))) { 793 if (srmmu_probe(addr + PMD_SIZE) == probed) 794 what = 1; 795 } 796 797 if (!(start & ~(PGDIR_MASK))) { 798 if (srmmu_probe(addr + PGDIR_SIZE) == probed) 799 what = 2; 800 } 801 802 pgdp = pgd_offset_k(start); 803 p4dp = p4d_offset(pgdp, start); 804 pudp = pud_offset(p4dp, start); 805 if (what == 2) { 806 *__nocache_fix(pgdp) = __pgd(probed); 807 start += PGDIR_SIZE; 808 continue; 809 } 810 if (pud_none(*__nocache_fix(pudp))) { 811 pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, 812 SRMMU_PMD_TABLE_SIZE); 813 if (pmdp == NULL) 814 early_pgtable_allocfail("pmd"); 815 memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE); 816 pud_set(__nocache_fix(pudp), pmdp); 817 } 818 pmdp = pmd_offset(__nocache_fix(pudp), start); 819 if (what == 1) { 820 *(pmd_t *)__nocache_fix(pmdp) = __pmd(probed); 821 start += PMD_SIZE; 822 continue; 823 } 824 if (srmmu_pmd_none(*__nocache_fix(pmdp))) { 825 ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE); 826 if (ptep == NULL) 827 early_pgtable_allocfail("pte"); 828 memset(__nocache_fix(ptep), 0, PTE_SIZE); 829 pmd_set(__nocache_fix(pmdp), ptep); 830 } 831 ptep = pte_offset_kernel(__nocache_fix(pmdp), start); 832 *__nocache_fix(ptep) = __pte(probed); 833 start += PAGE_SIZE; 834 } 835 } 836 837 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID) 838 839 /* Create a third-level SRMMU 16MB page mapping. */ 840 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base) 841 { 842 pgd_t *pgdp = pgd_offset_k(vaddr); 843 unsigned long big_pte; 844 845 big_pte = KERNEL_PTE(phys_base >> 4); 846 *__nocache_fix(pgdp) = __pgd(big_pte); 847 } 848 849 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */ 850 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry) 851 { 852 unsigned long pstart = (sp_banks[sp_entry].base_addr & PGDIR_MASK); 853 unsigned long vstart = (vbase & PGDIR_MASK); 854 unsigned long vend = PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes); 855 /* Map "low" memory only */ 856 const unsigned long min_vaddr = PAGE_OFFSET; 857 const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM; 858 859 if (vstart < min_vaddr || vstart >= max_vaddr) 860 return vstart; 861 862 if (vend > max_vaddr || vend < min_vaddr) 863 vend = max_vaddr; 864 865 while (vstart < vend) { 866 do_large_mapping(vstart, pstart); 867 vstart += PGDIR_SIZE; pstart += PGDIR_SIZE; 868 } 869 return vstart; 870 } 871 872 static void __init map_kernel(void) 873 { 874 int i; 875 876 if (phys_base > 0) { 877 do_large_mapping(PAGE_OFFSET, phys_base); 878 } 879 880 for (i = 0; sp_banks[i].num_bytes != 0; i++) { 881 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i); 882 } 883 } 884 885 void (*poke_srmmu)(void) = NULL; 886 887 void __init arch_zone_limits_init(unsigned long *max_zone_pfns) 888 { 889 max_zone_pfns[ZONE_DMA] = max_low_pfn; 890 max_zone_pfns[ZONE_NORMAL] = max_low_pfn; 891 max_zone_pfns[ZONE_HIGHMEM] = highend_pfn; 892 } 893 894 void __init srmmu_paging_init(void) 895 { 896 int i; 897 phandle cpunode; 898 char node_str[128]; 899 pgd_t *pgd; 900 p4d_t *p4d; 901 pud_t *pud; 902 pmd_t *pmd; 903 pte_t *pte; 904 unsigned long pages_avail; 905 906 init_mm.context = (unsigned long) NO_CONTEXT; 907 sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */ 908 909 if (sparc_cpu_model == sun4d) 910 num_contexts = 65536; /* We know it is Viking */ 911 else { 912 /* Find the number of contexts on the srmmu. */ 913 cpunode = prom_getchild(prom_root_node); 914 num_contexts = 0; 915 while (cpunode != 0) { 916 prom_getstring(cpunode, "device_type", node_str, sizeof(node_str)); 917 if (!strcmp(node_str, "cpu")) { 918 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8); 919 break; 920 } 921 cpunode = prom_getsibling(cpunode); 922 } 923 } 924 925 if (!num_contexts) { 926 prom_printf("Something wrong, can't find cpu node in paging_init.\n"); 927 prom_halt(); 928 } 929 930 pages_avail = 0; 931 last_valid_pfn = bootmem_init(&pages_avail); 932 933 srmmu_nocache_calcsize(); 934 srmmu_nocache_init(); 935 srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE)); 936 map_kernel(); 937 938 /* ctx table has to be physically aligned to its size */ 939 srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t)); 940 srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table); 941 942 for (i = 0; i < num_contexts; i++) 943 srmmu_ctxd_set(__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir); 944 945 flush_cache_all(); 946 srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys); 947 #ifdef CONFIG_SMP 948 /* Stop from hanging here... */ 949 local_ops->tlb_all(); 950 #else 951 flush_tlb_all(); 952 #endif 953 poke_srmmu(); 954 955 srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END); 956 srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END); 957 958 srmmu_allocate_ptable_skeleton( 959 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP); 960 srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END); 961 962 pgd = pgd_offset_k(PKMAP_BASE); 963 p4d = p4d_offset(pgd, PKMAP_BASE); 964 pud = pud_offset(p4d, PKMAP_BASE); 965 pmd = pmd_offset(pud, PKMAP_BASE); 966 pte = pte_offset_kernel(pmd, PKMAP_BASE); 967 pkmap_page_table = pte; 968 969 flush_cache_all(); 970 flush_tlb_all(); 971 972 sparc_context_init(num_contexts); 973 } 974 975 void mmu_info(struct seq_file *m) 976 { 977 seq_printf(m, 978 "MMU type\t: %s\n" 979 "contexts\t: %d\n" 980 "nocache total\t: %ld\n" 981 "nocache used\t: %d\n", 982 srmmu_name, 983 num_contexts, 984 srmmu_nocache_size, 985 srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT); 986 } 987 988 int init_new_context(struct task_struct *tsk, struct mm_struct *mm) 989 { 990 mm->context = NO_CONTEXT; 991 return 0; 992 } 993 994 void destroy_context(struct mm_struct *mm) 995 { 996 unsigned long flags; 997 998 if (mm->context != NO_CONTEXT) { 999 flush_cache_mm(mm); 1000 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir); 1001 flush_tlb_mm(mm); 1002 spin_lock_irqsave(&srmmu_context_spinlock, flags); 1003 free_context(mm->context); 1004 spin_unlock_irqrestore(&srmmu_context_spinlock, flags); 1005 mm->context = NO_CONTEXT; 1006 } 1007 } 1008 1009 /* Init various srmmu chip types. */ 1010 static void __init srmmu_is_bad(void) 1011 { 1012 prom_printf("Could not determine SRMMU chip type.\n"); 1013 prom_halt(); 1014 } 1015 1016 static void __init init_vac_layout(void) 1017 { 1018 phandle nd; 1019 int cache_lines; 1020 char node_str[128]; 1021 #ifdef CONFIG_SMP 1022 int cpu = 0; 1023 unsigned long max_size = 0; 1024 unsigned long min_line_size = 0x10000000; 1025 #endif 1026 1027 nd = prom_getchild(prom_root_node); 1028 while ((nd = prom_getsibling(nd)) != 0) { 1029 prom_getstring(nd, "device_type", node_str, sizeof(node_str)); 1030 if (!strcmp(node_str, "cpu")) { 1031 vac_line_size = prom_getint(nd, "cache-line-size"); 1032 if (vac_line_size == -1) { 1033 prom_printf("can't determine cache-line-size, halting.\n"); 1034 prom_halt(); 1035 } 1036 cache_lines = prom_getint(nd, "cache-nlines"); 1037 if (cache_lines == -1) { 1038 prom_printf("can't determine cache-nlines, halting.\n"); 1039 prom_halt(); 1040 } 1041 1042 vac_cache_size = cache_lines * vac_line_size; 1043 #ifdef CONFIG_SMP 1044 if (vac_cache_size > max_size) 1045 max_size = vac_cache_size; 1046 if (vac_line_size < min_line_size) 1047 min_line_size = vac_line_size; 1048 //FIXME: cpus not contiguous!! 1049 cpu++; 1050 if (cpu >= nr_cpu_ids || !cpu_online(cpu)) 1051 break; 1052 #else 1053 break; 1054 #endif 1055 } 1056 } 1057 if (nd == 0) { 1058 prom_printf("No CPU nodes found, halting.\n"); 1059 prom_halt(); 1060 } 1061 #ifdef CONFIG_SMP 1062 vac_cache_size = max_size; 1063 vac_line_size = min_line_size; 1064 #endif 1065 printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n", 1066 (int)vac_cache_size, (int)vac_line_size); 1067 } 1068 1069 static void poke_hypersparc(void) 1070 { 1071 volatile unsigned long clear; 1072 unsigned long mreg = srmmu_get_mmureg(); 1073 1074 hyper_flush_unconditional_combined(); 1075 1076 mreg &= ~(HYPERSPARC_CWENABLE); 1077 mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE); 1078 mreg |= (HYPERSPARC_CMODE); 1079 1080 srmmu_set_mmureg(mreg); 1081 1082 #if 0 /* XXX I think this is bad news... -DaveM */ 1083 hyper_clear_all_tags(); 1084 #endif 1085 1086 put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE); 1087 hyper_flush_whole_icache(); 1088 clear = srmmu_get_faddr(); 1089 clear = srmmu_get_fstatus(); 1090 } 1091 1092 static const struct sparc32_cachetlb_ops hypersparc_ops = { 1093 .cache_all = hypersparc_flush_cache_all, 1094 .cache_mm = hypersparc_flush_cache_mm, 1095 .cache_page = hypersparc_flush_cache_page, 1096 .cache_range = hypersparc_flush_cache_range, 1097 .tlb_all = hypersparc_flush_tlb_all, 1098 .tlb_mm = hypersparc_flush_tlb_mm, 1099 .tlb_page = hypersparc_flush_tlb_page, 1100 .tlb_range = hypersparc_flush_tlb_range, 1101 .page_to_ram = hypersparc_flush_page_to_ram, 1102 .sig_insns = hypersparc_flush_sig_insns, 1103 .page_for_dma = hypersparc_flush_page_for_dma, 1104 }; 1105 1106 static void __init init_hypersparc(void) 1107 { 1108 srmmu_name = "ROSS HyperSparc"; 1109 srmmu_modtype = HyperSparc; 1110 1111 init_vac_layout(); 1112 1113 is_hypersparc = 1; 1114 sparc32_cachetlb_ops = &hypersparc_ops; 1115 1116 poke_srmmu = poke_hypersparc; 1117 1118 hypersparc_setup_blockops(); 1119 } 1120 1121 static void poke_swift(void) 1122 { 1123 unsigned long mreg; 1124 1125 /* Clear any crap from the cache or else... */ 1126 swift_flush_cache_all(); 1127 1128 /* Enable I & D caches */ 1129 mreg = srmmu_get_mmureg(); 1130 mreg |= (SWIFT_IE | SWIFT_DE); 1131 /* 1132 * The Swift branch folding logic is completely broken. At 1133 * trap time, if things are just right, if can mistakenly 1134 * think that a trap is coming from kernel mode when in fact 1135 * it is coming from user mode (it mis-executes the branch in 1136 * the trap code). So you see things like crashme completely 1137 * hosing your machine which is completely unacceptable. Turn 1138 * this shit off... nice job Fujitsu. 1139 */ 1140 mreg &= ~(SWIFT_BF); 1141 srmmu_set_mmureg(mreg); 1142 } 1143 1144 static const struct sparc32_cachetlb_ops swift_ops = { 1145 .cache_all = swift_flush_cache_all, 1146 .cache_mm = swift_flush_cache_mm, 1147 .cache_page = swift_flush_cache_page, 1148 .cache_range = swift_flush_cache_range, 1149 .tlb_all = swift_flush_tlb_all, 1150 .tlb_mm = swift_flush_tlb_mm, 1151 .tlb_page = swift_flush_tlb_page, 1152 .tlb_range = swift_flush_tlb_range, 1153 .page_to_ram = swift_flush_page_to_ram, 1154 .sig_insns = swift_flush_sig_insns, 1155 .page_for_dma = swift_flush_page_for_dma, 1156 }; 1157 1158 #define SWIFT_MASKID_ADDR 0x10003018 1159 static void __init init_swift(void) 1160 { 1161 unsigned long swift_rev; 1162 1163 __asm__ __volatile__("lda [%1] %2, %0\n\t" 1164 "srl %0, 0x18, %0\n\t" : 1165 "=r" (swift_rev) : 1166 "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS)); 1167 srmmu_name = "Fujitsu Swift"; 1168 switch (swift_rev) { 1169 case 0x11: 1170 case 0x20: 1171 case 0x23: 1172 case 0x30: 1173 srmmu_modtype = Swift_lots_o_bugs; 1174 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN); 1175 /* 1176 * Gee george, I wonder why Sun is so hush hush about 1177 * this hardware bug... really braindamage stuff going 1178 * on here. However I think we can find a way to avoid 1179 * all of the workaround overhead under Linux. Basically, 1180 * any page fault can cause kernel pages to become user 1181 * accessible (the mmu gets confused and clears some of 1182 * the ACC bits in kernel ptes). Aha, sounds pretty 1183 * horrible eh? But wait, after extensive testing it appears 1184 * that if you use pgd_t level large kernel pte's (like the 1185 * 4MB pages on the Pentium) the bug does not get tripped 1186 * at all. This avoids almost all of the major overhead. 1187 * Welcome to a world where your vendor tells you to, 1188 * "apply this kernel patch" instead of "sorry for the 1189 * broken hardware, send it back and we'll give you 1190 * properly functioning parts" 1191 */ 1192 break; 1193 case 0x25: 1194 case 0x31: 1195 srmmu_modtype = Swift_bad_c; 1196 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN; 1197 /* 1198 * You see Sun allude to this hardware bug but never 1199 * admit things directly, they'll say things like, 1200 * "the Swift chip cache problems" or similar. 1201 */ 1202 break; 1203 default: 1204 srmmu_modtype = Swift_ok; 1205 break; 1206 } 1207 1208 sparc32_cachetlb_ops = &swift_ops; 1209 flush_page_for_dma_global = 0; 1210 1211 /* 1212 * Are you now convinced that the Swift is one of the 1213 * biggest VLSI abortions of all time? Bravo Fujitsu! 1214 * Fujitsu, the !#?!%$'d up processor people. I bet if 1215 * you examined the microcode of the Swift you'd find 1216 * XXX's all over the place. 1217 */ 1218 poke_srmmu = poke_swift; 1219 } 1220 1221 static void turbosparc_flush_cache_all(void) 1222 { 1223 flush_user_windows(); 1224 turbosparc_idflash_clear(); 1225 } 1226 1227 static void turbosparc_flush_cache_mm(struct mm_struct *mm) 1228 { 1229 FLUSH_BEGIN(mm) 1230 flush_user_windows(); 1231 turbosparc_idflash_clear(); 1232 FLUSH_END 1233 } 1234 1235 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end) 1236 { 1237 FLUSH_BEGIN(vma->vm_mm) 1238 flush_user_windows(); 1239 turbosparc_idflash_clear(); 1240 FLUSH_END 1241 } 1242 1243 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page) 1244 { 1245 FLUSH_BEGIN(vma->vm_mm) 1246 flush_user_windows(); 1247 if (vma->vm_flags & VM_EXEC) 1248 turbosparc_flush_icache(); 1249 turbosparc_flush_dcache(); 1250 FLUSH_END 1251 } 1252 1253 /* TurboSparc is copy-back, if we turn it on, but this does not work. */ 1254 static void turbosparc_flush_page_to_ram(unsigned long page) 1255 { 1256 #ifdef TURBOSPARC_WRITEBACK 1257 volatile unsigned long clear; 1258 1259 if (srmmu_probe(page)) 1260 turbosparc_flush_page_cache(page); 1261 clear = srmmu_get_fstatus(); 1262 #endif 1263 } 1264 1265 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr) 1266 { 1267 } 1268 1269 static void turbosparc_flush_page_for_dma(unsigned long page) 1270 { 1271 turbosparc_flush_dcache(); 1272 } 1273 1274 static void turbosparc_flush_tlb_all(void) 1275 { 1276 srmmu_flush_whole_tlb(); 1277 } 1278 1279 static void turbosparc_flush_tlb_mm(struct mm_struct *mm) 1280 { 1281 FLUSH_BEGIN(mm) 1282 srmmu_flush_whole_tlb(); 1283 FLUSH_END 1284 } 1285 1286 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end) 1287 { 1288 FLUSH_BEGIN(vma->vm_mm) 1289 srmmu_flush_whole_tlb(); 1290 FLUSH_END 1291 } 1292 1293 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page) 1294 { 1295 FLUSH_BEGIN(vma->vm_mm) 1296 srmmu_flush_whole_tlb(); 1297 FLUSH_END 1298 } 1299 1300 1301 static void poke_turbosparc(void) 1302 { 1303 unsigned long mreg = srmmu_get_mmureg(); 1304 unsigned long ccreg; 1305 1306 /* Clear any crap from the cache or else... */ 1307 turbosparc_flush_cache_all(); 1308 /* Temporarily disable I & D caches */ 1309 mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); 1310 mreg &= ~(TURBOSPARC_PCENABLE); /* Don't check parity */ 1311 srmmu_set_mmureg(mreg); 1312 1313 ccreg = turbosparc_get_ccreg(); 1314 1315 #ifdef TURBOSPARC_WRITEBACK 1316 ccreg |= (TURBOSPARC_SNENABLE); /* Do DVMA snooping in Dcache */ 1317 ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE); 1318 /* Write-back D-cache, emulate VLSI 1319 * abortion number three, not number one */ 1320 #else 1321 /* For now let's play safe, optimize later */ 1322 ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE); 1323 /* Do DVMA snooping in Dcache, Write-thru D-cache */ 1324 ccreg &= ~(TURBOSPARC_uS2); 1325 /* Emulate VLSI abortion number three, not number one */ 1326 #endif 1327 1328 switch (ccreg & 7) { 1329 case 0: /* No SE cache */ 1330 case 7: /* Test mode */ 1331 break; 1332 default: 1333 ccreg |= (TURBOSPARC_SCENABLE); 1334 } 1335 turbosparc_set_ccreg(ccreg); 1336 1337 mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */ 1338 mreg |= (TURBOSPARC_ICSNOOP); /* Icache snooping on */ 1339 srmmu_set_mmureg(mreg); 1340 } 1341 1342 static const struct sparc32_cachetlb_ops turbosparc_ops = { 1343 .cache_all = turbosparc_flush_cache_all, 1344 .cache_mm = turbosparc_flush_cache_mm, 1345 .cache_page = turbosparc_flush_cache_page, 1346 .cache_range = turbosparc_flush_cache_range, 1347 .tlb_all = turbosparc_flush_tlb_all, 1348 .tlb_mm = turbosparc_flush_tlb_mm, 1349 .tlb_page = turbosparc_flush_tlb_page, 1350 .tlb_range = turbosparc_flush_tlb_range, 1351 .page_to_ram = turbosparc_flush_page_to_ram, 1352 .sig_insns = turbosparc_flush_sig_insns, 1353 .page_for_dma = turbosparc_flush_page_for_dma, 1354 }; 1355 1356 static void __init init_turbosparc(void) 1357 { 1358 srmmu_name = "Fujitsu TurboSparc"; 1359 srmmu_modtype = TurboSparc; 1360 sparc32_cachetlb_ops = &turbosparc_ops; 1361 poke_srmmu = poke_turbosparc; 1362 } 1363 1364 static void poke_tsunami(void) 1365 { 1366 unsigned long mreg = srmmu_get_mmureg(); 1367 1368 tsunami_flush_icache(); 1369 tsunami_flush_dcache(); 1370 mreg &= ~TSUNAMI_ITD; 1371 mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB); 1372 srmmu_set_mmureg(mreg); 1373 } 1374 1375 static const struct sparc32_cachetlb_ops tsunami_ops = { 1376 .cache_all = tsunami_flush_cache_all, 1377 .cache_mm = tsunami_flush_cache_mm, 1378 .cache_page = tsunami_flush_cache_page, 1379 .cache_range = tsunami_flush_cache_range, 1380 .tlb_all = tsunami_flush_tlb_all, 1381 .tlb_mm = tsunami_flush_tlb_mm, 1382 .tlb_page = tsunami_flush_tlb_page, 1383 .tlb_range = tsunami_flush_tlb_range, 1384 .page_to_ram = tsunami_flush_page_to_ram, 1385 .sig_insns = tsunami_flush_sig_insns, 1386 .page_for_dma = tsunami_flush_page_for_dma, 1387 }; 1388 1389 static void __init init_tsunami(void) 1390 { 1391 /* 1392 * Tsunami's pretty sane, Sun and TI actually got it 1393 * somewhat right this time. Fujitsu should have 1394 * taken some lessons from them. 1395 */ 1396 1397 srmmu_name = "TI Tsunami"; 1398 srmmu_modtype = Tsunami; 1399 sparc32_cachetlb_ops = &tsunami_ops; 1400 poke_srmmu = poke_tsunami; 1401 1402 tsunami_setup_blockops(); 1403 } 1404 1405 static void poke_viking(void) 1406 { 1407 unsigned long mreg = srmmu_get_mmureg(); 1408 static int smp_catch; 1409 1410 if (viking_mxcc_present) { 1411 unsigned long mxcc_control = mxcc_get_creg(); 1412 1413 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE); 1414 mxcc_control &= ~(MXCC_CTL_RRC); 1415 mxcc_set_creg(mxcc_control); 1416 1417 /* 1418 * We don't need memory parity checks. 1419 * XXX This is a mess, have to dig out later. ecd. 1420 viking_mxcc_turn_off_parity(&mreg, &mxcc_control); 1421 */ 1422 1423 /* We do cache ptables on MXCC. */ 1424 mreg |= VIKING_TCENABLE; 1425 } else { 1426 unsigned long bpreg; 1427 1428 mreg &= ~(VIKING_TCENABLE); 1429 if (smp_catch++) { 1430 /* Must disable mixed-cmd mode here for other cpu's. */ 1431 bpreg = viking_get_bpreg(); 1432 bpreg &= ~(VIKING_ACTION_MIX); 1433 viking_set_bpreg(bpreg); 1434 1435 /* Just in case PROM does something funny. */ 1436 msi_set_sync(); 1437 } 1438 } 1439 1440 mreg |= VIKING_SPENABLE; 1441 mreg |= (VIKING_ICENABLE | VIKING_DCENABLE); 1442 mreg |= VIKING_SBENABLE; 1443 mreg &= ~(VIKING_ACENABLE); 1444 srmmu_set_mmureg(mreg); 1445 } 1446 1447 static struct sparc32_cachetlb_ops viking_ops __ro_after_init = { 1448 .cache_all = viking_flush_cache_all, 1449 .cache_mm = viking_flush_cache_mm, 1450 .cache_page = viking_flush_cache_page, 1451 .cache_range = viking_flush_cache_range, 1452 .tlb_all = viking_flush_tlb_all, 1453 .tlb_mm = viking_flush_tlb_mm, 1454 .tlb_page = viking_flush_tlb_page, 1455 .tlb_range = viking_flush_tlb_range, 1456 .page_to_ram = viking_flush_page_to_ram, 1457 .sig_insns = viking_flush_sig_insns, 1458 .page_for_dma = viking_flush_page_for_dma, 1459 }; 1460 1461 #ifdef CONFIG_SMP 1462 /* On sun4d the cpu broadcasts local TLB flushes, so we can just 1463 * perform the local TLB flush and all the other cpus will see it. 1464 * But, unfortunately, there is a bug in the sun4d XBUS backplane 1465 * that requires that we add some synchronization to these flushes. 1466 * 1467 * The bug is that the fifo which keeps track of all the pending TLB 1468 * broadcasts in the system is an entry or two too small, so if we 1469 * have too many going at once we'll overflow that fifo and lose a TLB 1470 * flush resulting in corruption. 1471 * 1472 * Our workaround is to take a global spinlock around the TLB flushes, 1473 * which guarentees we won't ever have too many pending. It's a big 1474 * hammer, but a semaphore like system to make sure we only have N TLB 1475 * flushes going at once will require SMP locking anyways so there's 1476 * no real value in trying any harder than this. 1477 */ 1478 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = { 1479 .cache_all = viking_flush_cache_all, 1480 .cache_mm = viking_flush_cache_mm, 1481 .cache_page = viking_flush_cache_page, 1482 .cache_range = viking_flush_cache_range, 1483 .tlb_all = sun4dsmp_flush_tlb_all, 1484 .tlb_mm = sun4dsmp_flush_tlb_mm, 1485 .tlb_page = sun4dsmp_flush_tlb_page, 1486 .tlb_range = sun4dsmp_flush_tlb_range, 1487 .page_to_ram = viking_flush_page_to_ram, 1488 .sig_insns = viking_flush_sig_insns, 1489 .page_for_dma = viking_flush_page_for_dma, 1490 }; 1491 #endif 1492 1493 static void __init init_viking(void) 1494 { 1495 unsigned long mreg = srmmu_get_mmureg(); 1496 1497 /* Ahhh, the viking. SRMMU VLSI abortion number two... */ 1498 if (mreg & VIKING_MMODE) { 1499 srmmu_name = "TI Viking"; 1500 viking_mxcc_present = 0; 1501 msi_set_sync(); 1502 1503 /* 1504 * We need this to make sure old viking takes no hits 1505 * on its cache for dma snoops to workaround the 1506 * "load from non-cacheable memory" interrupt bug. 1507 * This is only necessary because of the new way in 1508 * which we use the IOMMU. 1509 */ 1510 viking_ops.page_for_dma = viking_flush_page; 1511 #ifdef CONFIG_SMP 1512 viking_sun4d_smp_ops.page_for_dma = viking_flush_page; 1513 #endif 1514 flush_page_for_dma_global = 0; 1515 } else { 1516 srmmu_name = "TI Viking/MXCC"; 1517 viking_mxcc_present = 1; 1518 srmmu_cache_pagetables = 1; 1519 } 1520 1521 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *) 1522 &viking_ops; 1523 #ifdef CONFIG_SMP 1524 if (sparc_cpu_model == sun4d) 1525 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *) 1526 &viking_sun4d_smp_ops; 1527 #endif 1528 1529 poke_srmmu = poke_viking; 1530 } 1531 1532 /* Probe for the srmmu chip version. */ 1533 static void __init get_srmmu_type(void) 1534 { 1535 unsigned long mreg, psr; 1536 unsigned long mod_typ, mod_rev, psr_typ, psr_vers; 1537 1538 srmmu_modtype = SRMMU_INVAL_MOD; 1539 hwbug_bitmask = 0; 1540 1541 mreg = srmmu_get_mmureg(); psr = get_psr(); 1542 mod_typ = (mreg & 0xf0000000) >> 28; 1543 mod_rev = (mreg & 0x0f000000) >> 24; 1544 psr_typ = (psr >> 28) & 0xf; 1545 psr_vers = (psr >> 24) & 0xf; 1546 1547 /* First, check for sparc-leon. */ 1548 if (sparc_cpu_model == sparc_leon) { 1549 init_leon(); 1550 return; 1551 } 1552 1553 /* Second, check for HyperSparc or Cypress. */ 1554 if (mod_typ == 1) { 1555 switch (mod_rev) { 1556 case 7: 1557 /* UP or MP Hypersparc */ 1558 init_hypersparc(); 1559 break; 1560 case 0: 1561 case 2: 1562 case 10: 1563 case 11: 1564 case 12: 1565 case 13: 1566 case 14: 1567 case 15: 1568 default: 1569 prom_printf("Sparc-Linux Cypress support does not longer exit.\n"); 1570 prom_halt(); 1571 break; 1572 } 1573 return; 1574 } 1575 1576 /* Now Fujitsu TurboSparc. It might happen that it is 1577 * in Swift emulation mode, so we will check later... 1578 */ 1579 if (psr_typ == 0 && psr_vers == 5) { 1580 init_turbosparc(); 1581 return; 1582 } 1583 1584 /* Next check for Fujitsu Swift. */ 1585 if (psr_typ == 0 && psr_vers == 4) { 1586 phandle cpunode; 1587 char node_str[128]; 1588 1589 /* Look if it is not a TurboSparc emulating Swift... */ 1590 cpunode = prom_getchild(prom_root_node); 1591 while ((cpunode = prom_getsibling(cpunode)) != 0) { 1592 prom_getstring(cpunode, "device_type", node_str, sizeof(node_str)); 1593 if (!strcmp(node_str, "cpu")) { 1594 if (!prom_getintdefault(cpunode, "psr-implementation", 1) && 1595 prom_getintdefault(cpunode, "psr-version", 1) == 5) { 1596 init_turbosparc(); 1597 return; 1598 } 1599 break; 1600 } 1601 } 1602 1603 init_swift(); 1604 return; 1605 } 1606 1607 /* Now the Viking family of srmmu. */ 1608 if (psr_typ == 4 && 1609 ((psr_vers == 0) || 1610 ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) { 1611 init_viking(); 1612 return; 1613 } 1614 1615 /* Finally the Tsunami. */ 1616 if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) { 1617 init_tsunami(); 1618 return; 1619 } 1620 1621 /* Oh well */ 1622 srmmu_is_bad(); 1623 } 1624 1625 #ifdef CONFIG_SMP 1626 /* Local cross-calls. */ 1627 static void smp_flush_page_for_dma(unsigned long page) 1628 { 1629 xc1(local_ops->page_for_dma, page); 1630 local_ops->page_for_dma(page); 1631 } 1632 1633 static void smp_flush_cache_all(void) 1634 { 1635 xc0(local_ops->cache_all); 1636 local_ops->cache_all(); 1637 } 1638 1639 static void smp_flush_tlb_all(void) 1640 { 1641 xc0(local_ops->tlb_all); 1642 local_ops->tlb_all(); 1643 } 1644 1645 static bool any_other_mm_cpus(struct mm_struct *mm) 1646 { 1647 return cpumask_any_but(mm_cpumask(mm), smp_processor_id()) < nr_cpu_ids; 1648 } 1649 1650 static void smp_flush_cache_mm(struct mm_struct *mm) 1651 { 1652 if (mm->context != NO_CONTEXT) { 1653 if (any_other_mm_cpus(mm)) 1654 xc1(local_ops->cache_mm, (unsigned long)mm); 1655 local_ops->cache_mm(mm); 1656 } 1657 } 1658 1659 static void smp_flush_tlb_mm(struct mm_struct *mm) 1660 { 1661 if (mm->context != NO_CONTEXT) { 1662 if (any_other_mm_cpus(mm)) { 1663 xc1(local_ops->tlb_mm, (unsigned long)mm); 1664 if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm) 1665 cpumask_copy(mm_cpumask(mm), 1666 cpumask_of(smp_processor_id())); 1667 } 1668 local_ops->tlb_mm(mm); 1669 } 1670 } 1671 1672 static void smp_flush_cache_range(struct vm_area_struct *vma, 1673 unsigned long start, 1674 unsigned long end) 1675 { 1676 struct mm_struct *mm = vma->vm_mm; 1677 1678 if (mm->context != NO_CONTEXT) { 1679 if (any_other_mm_cpus(mm)) 1680 xc3(local_ops->cache_range, (unsigned long)vma, start, 1681 end); 1682 local_ops->cache_range(vma, start, end); 1683 } 1684 } 1685 1686 static void smp_flush_tlb_range(struct vm_area_struct *vma, 1687 unsigned long start, 1688 unsigned long end) 1689 { 1690 struct mm_struct *mm = vma->vm_mm; 1691 1692 if (mm->context != NO_CONTEXT) { 1693 if (any_other_mm_cpus(mm)) 1694 xc3(local_ops->tlb_range, (unsigned long)vma, start, 1695 end); 1696 local_ops->tlb_range(vma, start, end); 1697 } 1698 } 1699 1700 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page) 1701 { 1702 struct mm_struct *mm = vma->vm_mm; 1703 1704 if (mm->context != NO_CONTEXT) { 1705 if (any_other_mm_cpus(mm)) 1706 xc2(local_ops->cache_page, (unsigned long)vma, page); 1707 local_ops->cache_page(vma, page); 1708 } 1709 } 1710 1711 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page) 1712 { 1713 struct mm_struct *mm = vma->vm_mm; 1714 1715 if (mm->context != NO_CONTEXT) { 1716 if (any_other_mm_cpus(mm)) 1717 xc2(local_ops->tlb_page, (unsigned long)vma, page); 1718 local_ops->tlb_page(vma, page); 1719 } 1720 } 1721 1722 static void smp_flush_page_to_ram(unsigned long page) 1723 { 1724 /* Current theory is that those who call this are the one's 1725 * who have just dirtied their cache with the pages contents 1726 * in kernel space, therefore we only run this on local cpu. 1727 * 1728 * XXX This experiment failed, research further... -DaveM 1729 */ 1730 #if 1 1731 xc1(local_ops->page_to_ram, page); 1732 #endif 1733 local_ops->page_to_ram(page); 1734 } 1735 1736 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr) 1737 { 1738 if (any_other_mm_cpus(mm)) 1739 xc2(local_ops->sig_insns, (unsigned long)mm, insn_addr); 1740 local_ops->sig_insns(mm, insn_addr); 1741 } 1742 1743 static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = { 1744 .cache_all = smp_flush_cache_all, 1745 .cache_mm = smp_flush_cache_mm, 1746 .cache_page = smp_flush_cache_page, 1747 .cache_range = smp_flush_cache_range, 1748 .tlb_all = smp_flush_tlb_all, 1749 .tlb_mm = smp_flush_tlb_mm, 1750 .tlb_page = smp_flush_tlb_page, 1751 .tlb_range = smp_flush_tlb_range, 1752 .page_to_ram = smp_flush_page_to_ram, 1753 .sig_insns = smp_flush_sig_insns, 1754 .page_for_dma = smp_flush_page_for_dma, 1755 }; 1756 #endif 1757 1758 /* Load up routines and constants for sun4m and sun4d mmu */ 1759 void __init load_mmu(void) 1760 { 1761 /* Functions */ 1762 get_srmmu_type(); 1763 1764 #ifdef CONFIG_SMP 1765 /* El switcheroo... */ 1766 local_ops = sparc32_cachetlb_ops; 1767 1768 if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) { 1769 smp_cachetlb_ops.tlb_all = local_ops->tlb_all; 1770 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm; 1771 smp_cachetlb_ops.tlb_range = local_ops->tlb_range; 1772 smp_cachetlb_ops.tlb_page = local_ops->tlb_page; 1773 } 1774 1775 if (poke_srmmu == poke_viking) { 1776 /* Avoid unnecessary cross calls. */ 1777 smp_cachetlb_ops.cache_all = local_ops->cache_all; 1778 smp_cachetlb_ops.cache_mm = local_ops->cache_mm; 1779 smp_cachetlb_ops.cache_range = local_ops->cache_range; 1780 smp_cachetlb_ops.cache_page = local_ops->cache_page; 1781 1782 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram; 1783 smp_cachetlb_ops.sig_insns = local_ops->sig_insns; 1784 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma; 1785 } 1786 1787 /* It really is const after this point. */ 1788 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *) 1789 &smp_cachetlb_ops; 1790 #endif 1791 1792 if (sparc_cpu_model != sun4d) 1793 ld_mmu_iommu(); 1794 #ifdef CONFIG_SMP 1795 if (sparc_cpu_model == sun4d) 1796 sun4d_init_smp(); 1797 else if (sparc_cpu_model == sparc_leon) 1798 leon_init_smp(); 1799 else 1800 sun4m_init_smp(); 1801 #endif 1802 } 1803