1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * arch/sparc64/mm/init.c 4 * 5 * Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu) 6 * Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz) 7 */ 8 9 #include <linux/extable.h> 10 #include <linux/kernel.h> 11 #include <linux/sched.h> 12 #include <linux/string.h> 13 #include <linux/init.h> 14 #include <linux/memblock.h> 15 #include <linux/mm.h> 16 #include <linux/hugetlb.h> 17 #include <linux/initrd.h> 18 #include <linux/swap.h> 19 #include <linux/pagemap.h> 20 #include <linux/poison.h> 21 #include <linux/fs.h> 22 #include <linux/seq_file.h> 23 #include <linux/kprobes.h> 24 #include <linux/cache.h> 25 #include <linux/sort.h> 26 #include <linux/ioport.h> 27 #include <linux/percpu.h> 28 #include <linux/mmzone.h> 29 #include <linux/gfp.h> 30 #include <linux/bootmem_info.h> 31 32 #include <asm/head.h> 33 #include <asm/page.h> 34 #include <asm/pgalloc.h> 35 #include <asm/oplib.h> 36 #include <asm/iommu.h> 37 #include <asm/io.h> 38 #include <linux/uaccess.h> 39 #include <asm/mmu_context.h> 40 #include <asm/tlbflush.h> 41 #include <asm/dma.h> 42 #include <asm/starfire.h> 43 #include <asm/tlb.h> 44 #include <asm/spitfire.h> 45 #include <asm/sections.h> 46 #include <asm/tsb.h> 47 #include <asm/hypervisor.h> 48 #include <asm/prom.h> 49 #include <asm/mdesc.h> 50 #include <asm/cpudata.h> 51 #include <asm/setup.h> 52 #include <asm/irq.h> 53 54 #include "init_64.h" 55 56 unsigned long kern_linear_pte_xor[4] __read_mostly; 57 static unsigned long page_cache4v_flag; 58 59 /* A bitmap, two bits for every 256MB of physical memory. These two 60 * bits determine what page size we use for kernel linear 61 * translations. They form an index into kern_linear_pte_xor[]. The 62 * value in the indexed slot is XOR'd with the TLB miss virtual 63 * address to form the resulting TTE. The mapping is: 64 * 65 * 0 ==> 4MB 66 * 1 ==> 256MB 67 * 2 ==> 2GB 68 * 3 ==> 16GB 69 * 70 * All sun4v chips support 256MB pages. Only SPARC-T4 and later 71 * support 2GB pages, and hopefully future cpus will support the 16GB 72 * pages as well. For slots 2 and 3, we encode a 256MB TTE xor there 73 * if these larger page sizes are not supported by the cpu. 74 * 75 * It would be nice to determine this from the machine description 76 * 'cpu' properties, but we need to have this table setup before the 77 * MDESC is initialized. 78 */ 79 80 #ifndef CONFIG_DEBUG_PAGEALLOC 81 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings. 82 * Space is allocated for this right after the trap table in 83 * arch/sparc64/kernel/head.S 84 */ 85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES]; 86 #endif 87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES]; 88 89 static unsigned long cpu_pgsz_mask; 90 91 #define MAX_BANKS 1024 92 93 static struct linux_prom64_registers pavail[MAX_BANKS]; 94 static int pavail_ents; 95 96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES]; 97 98 static int cmp_p64(const void *a, const void *b) 99 { 100 const struct linux_prom64_registers *x = a, *y = b; 101 102 if (x->phys_addr > y->phys_addr) 103 return 1; 104 if (x->phys_addr < y->phys_addr) 105 return -1; 106 return 0; 107 } 108 109 static void __init read_obp_memory(const char *property, 110 struct linux_prom64_registers *regs, 111 int *num_ents) 112 { 113 phandle node = prom_finddevice("/memory"); 114 int prop_size = prom_getproplen(node, property); 115 int ents, ret, i; 116 117 ents = prop_size / sizeof(struct linux_prom64_registers); 118 if (ents > MAX_BANKS) { 119 prom_printf("The machine has more %s property entries than " 120 "this kernel can support (%d).\n", 121 property, MAX_BANKS); 122 prom_halt(); 123 } 124 125 ret = prom_getproperty(node, property, (char *) regs, prop_size); 126 if (ret == -1) { 127 prom_printf("Couldn't get %s property from /memory.\n", 128 property); 129 prom_halt(); 130 } 131 132 /* Sanitize what we got from the firmware, by page aligning 133 * everything. 134 */ 135 for (i = 0; i < ents; i++) { 136 unsigned long base, size; 137 138 base = regs[i].phys_addr; 139 size = regs[i].reg_size; 140 141 size &= PAGE_MASK; 142 if (base & ~PAGE_MASK) { 143 unsigned long new_base = PAGE_ALIGN(base); 144 145 size -= new_base - base; 146 if ((long) size < 0L) 147 size = 0UL; 148 base = new_base; 149 } 150 if (size == 0UL) { 151 /* If it is empty, simply get rid of it. 152 * This simplifies the logic of the other 153 * functions that process these arrays. 154 */ 155 memmove(®s[i], ®s[i + 1], 156 (ents - i - 1) * sizeof(regs[0])); 157 i--; 158 ents--; 159 continue; 160 } 161 regs[i].phys_addr = base; 162 regs[i].reg_size = size; 163 } 164 165 *num_ents = ents; 166 167 sort(regs, ents, sizeof(struct linux_prom64_registers), 168 cmp_p64, NULL); 169 } 170 171 /* Kernel physical address base and size in bytes. */ 172 unsigned long kern_base __read_mostly; 173 unsigned long kern_size __read_mostly; 174 175 /* Initial ramdisk setup */ 176 extern unsigned long sparc_ramdisk_image64; 177 extern unsigned int sparc_ramdisk_image; 178 extern unsigned int sparc_ramdisk_size; 179 180 struct page *mem_map_zero __read_mostly; 181 EXPORT_SYMBOL(mem_map_zero); 182 183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly; 184 185 unsigned long sparc64_kern_pri_context __read_mostly; 186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly; 187 unsigned long sparc64_kern_sec_context __read_mostly; 188 189 int num_kernel_image_mappings; 190 191 #ifdef CONFIG_DEBUG_DCFLUSH 192 atomic_t dcpage_flushes = ATOMIC_INIT(0); 193 #ifdef CONFIG_SMP 194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0); 195 #endif 196 #endif 197 198 inline void flush_dcache_folio_impl(struct folio *folio) 199 { 200 unsigned int i, nr = folio_nr_pages(folio); 201 202 BUG_ON(tlb_type == hypervisor); 203 #ifdef CONFIG_DEBUG_DCFLUSH 204 atomic_inc(&dcpage_flushes); 205 #endif 206 207 #ifdef DCACHE_ALIASING_POSSIBLE 208 for (i = 0; i < nr; i++) 209 __flush_dcache_page(folio_address(folio) + i * PAGE_SIZE, 210 ((tlb_type == spitfire) && 211 folio_flush_mapping(folio) != NULL)); 212 #else 213 if (folio_flush_mapping(folio) != NULL && 214 tlb_type == spitfire) { 215 for (i = 0; i < nr; i++) 216 __flush_icache_page((pfn + i) * PAGE_SIZE); 217 } 218 #endif 219 } 220 221 #define PG_dcache_dirty PG_arch_1 222 #define PG_dcache_cpu_shift 32UL 223 #define PG_dcache_cpu_mask \ 224 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL) 225 226 #define dcache_dirty_cpu(folio) \ 227 (((folio)->flags.f >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask) 228 229 static inline void set_dcache_dirty(struct folio *folio, int this_cpu) 230 { 231 unsigned long mask = this_cpu; 232 unsigned long non_cpu_bits; 233 234 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift); 235 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty); 236 237 __asm__ __volatile__("1:\n\t" 238 "ldx [%2], %%g7\n\t" 239 "and %%g7, %1, %%g1\n\t" 240 "or %%g1, %0, %%g1\n\t" 241 "casx [%2], %%g7, %%g1\n\t" 242 "cmp %%g7, %%g1\n\t" 243 "bne,pn %%xcc, 1b\n\t" 244 " nop" 245 : /* no outputs */ 246 : "r" (mask), "r" (non_cpu_bits), "r" (&folio->flags.f) 247 : "g1", "g7"); 248 } 249 250 static inline void clear_dcache_dirty_cpu(struct folio *folio, unsigned long cpu) 251 { 252 unsigned long mask = (1UL << PG_dcache_dirty); 253 254 __asm__ __volatile__("! test_and_clear_dcache_dirty\n" 255 "1:\n\t" 256 "ldx [%2], %%g7\n\t" 257 "srlx %%g7, %4, %%g1\n\t" 258 "and %%g1, %3, %%g1\n\t" 259 "cmp %%g1, %0\n\t" 260 "bne,pn %%icc, 2f\n\t" 261 " andn %%g7, %1, %%g1\n\t" 262 "casx [%2], %%g7, %%g1\n\t" 263 "cmp %%g7, %%g1\n\t" 264 "bne,pn %%xcc, 1b\n\t" 265 " nop\n" 266 "2:" 267 : /* no outputs */ 268 : "r" (cpu), "r" (mask), "r" (&folio->flags.f), 269 "i" (PG_dcache_cpu_mask), 270 "i" (PG_dcache_cpu_shift) 271 : "g1", "g7"); 272 } 273 274 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte) 275 { 276 unsigned long tsb_addr = (unsigned long) ent; 277 278 if (tlb_type == cheetah_plus || tlb_type == hypervisor) 279 tsb_addr = __pa(tsb_addr); 280 281 __tsb_insert(tsb_addr, tag, pte); 282 } 283 284 unsigned long _PAGE_ALL_SZ_BITS __read_mostly; 285 286 static void flush_dcache(unsigned long pfn) 287 { 288 struct page *page; 289 290 page = pfn_to_page(pfn); 291 if (page) { 292 struct folio *folio = page_folio(page); 293 unsigned long pg_flags; 294 295 pg_flags = folio->flags.f; 296 if (pg_flags & (1UL << PG_dcache_dirty)) { 297 int cpu = ((pg_flags >> PG_dcache_cpu_shift) & 298 PG_dcache_cpu_mask); 299 int this_cpu = get_cpu(); 300 301 /* This is just to optimize away some function calls 302 * in the SMP case. 303 */ 304 if (cpu == this_cpu) 305 flush_dcache_folio_impl(folio); 306 else 307 smp_flush_dcache_folio_impl(folio, cpu); 308 309 clear_dcache_dirty_cpu(folio, cpu); 310 311 put_cpu(); 312 } 313 } 314 } 315 316 /* mm->context.lock must be held */ 317 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index, 318 unsigned long tsb_hash_shift, unsigned long address, 319 unsigned long tte) 320 { 321 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb; 322 unsigned long tag; 323 324 if (unlikely(!tsb)) 325 return; 326 327 tsb += ((address >> tsb_hash_shift) & 328 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL)); 329 tag = (address >> 22UL); 330 tsb_insert(tsb, tag, tte); 331 } 332 333 #ifdef CONFIG_HUGETLB_PAGE 334 static int __init hugetlbpage_init(void) 335 { 336 hugetlb_add_hstate(HPAGE_64K_SHIFT - PAGE_SHIFT); 337 hugetlb_add_hstate(HPAGE_SHIFT - PAGE_SHIFT); 338 hugetlb_add_hstate(HPAGE_256MB_SHIFT - PAGE_SHIFT); 339 hugetlb_add_hstate(HPAGE_2GB_SHIFT - PAGE_SHIFT); 340 341 return 0; 342 } 343 344 arch_initcall(hugetlbpage_init); 345 346 static void __init pud_huge_patch(void) 347 { 348 struct pud_huge_patch_entry *p; 349 unsigned long addr; 350 351 p = &__pud_huge_patch; 352 addr = p->addr; 353 *(unsigned int *)addr = p->insn; 354 355 __asm__ __volatile__("flush %0" : : "r" (addr)); 356 } 357 358 bool __init arch_hugetlb_valid_size(unsigned long size) 359 { 360 unsigned int hugepage_shift = ilog2(size); 361 unsigned short hv_pgsz_idx; 362 unsigned int hv_pgsz_mask; 363 364 switch (hugepage_shift) { 365 case HPAGE_16GB_SHIFT: 366 hv_pgsz_mask = HV_PGSZ_MASK_16GB; 367 hv_pgsz_idx = HV_PGSZ_IDX_16GB; 368 pud_huge_patch(); 369 break; 370 case HPAGE_2GB_SHIFT: 371 hv_pgsz_mask = HV_PGSZ_MASK_2GB; 372 hv_pgsz_idx = HV_PGSZ_IDX_2GB; 373 break; 374 case HPAGE_256MB_SHIFT: 375 hv_pgsz_mask = HV_PGSZ_MASK_256MB; 376 hv_pgsz_idx = HV_PGSZ_IDX_256MB; 377 break; 378 case HPAGE_SHIFT: 379 hv_pgsz_mask = HV_PGSZ_MASK_4MB; 380 hv_pgsz_idx = HV_PGSZ_IDX_4MB; 381 break; 382 case HPAGE_64K_SHIFT: 383 hv_pgsz_mask = HV_PGSZ_MASK_64K; 384 hv_pgsz_idx = HV_PGSZ_IDX_64K; 385 break; 386 default: 387 hv_pgsz_mask = 0; 388 } 389 390 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) 391 return false; 392 393 return true; 394 } 395 #endif /* CONFIG_HUGETLB_PAGE */ 396 397 void update_mmu_cache_range(struct vm_fault *vmf, struct vm_area_struct *vma, 398 unsigned long address, pte_t *ptep, unsigned int nr) 399 { 400 struct mm_struct *mm; 401 unsigned long flags; 402 bool is_huge_tsb; 403 pte_t pte = *ptep; 404 unsigned int i; 405 406 if (tlb_type != hypervisor) { 407 unsigned long pfn = pte_pfn(pte); 408 409 if (pfn_valid(pfn)) 410 flush_dcache(pfn); 411 } 412 413 mm = vma->vm_mm; 414 415 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */ 416 if (!pte_accessible(mm, pte)) 417 return; 418 419 spin_lock_irqsave(&mm->context.lock, flags); 420 421 is_huge_tsb = false; 422 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 423 if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) { 424 unsigned long hugepage_size = PAGE_SIZE; 425 426 if (is_vm_hugetlb_page(vma)) 427 hugepage_size = huge_page_size(hstate_vma(vma)); 428 429 if (hugepage_size >= PUD_SIZE) { 430 unsigned long mask = 0x1ffc00000UL; 431 432 /* Transfer bits [32:22] from address to resolve 433 * at 4M granularity. 434 */ 435 pte_val(pte) &= ~mask; 436 pte_val(pte) |= (address & mask); 437 } else if (hugepage_size >= PMD_SIZE) { 438 /* We are fabricating 8MB pages using 4MB 439 * real hw pages. 440 */ 441 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT)); 442 } 443 444 if (hugepage_size >= PMD_SIZE) { 445 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, 446 REAL_HPAGE_SHIFT, address, pte_val(pte)); 447 is_huge_tsb = true; 448 } 449 } 450 #endif 451 if (!is_huge_tsb) { 452 for (i = 0; i < nr; i++) { 453 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT, 454 address, pte_val(pte)); 455 address += PAGE_SIZE; 456 pte_val(pte) += PAGE_SIZE; 457 } 458 } 459 460 spin_unlock_irqrestore(&mm->context.lock, flags); 461 } 462 463 void flush_dcache_folio(struct folio *folio) 464 { 465 unsigned long pfn = folio_pfn(folio); 466 struct address_space *mapping; 467 int this_cpu; 468 469 if (tlb_type == hypervisor) 470 return; 471 472 /* Do not bother with the expensive D-cache flush if it 473 * is merely the zero page. The 'bigcore' testcase in GDB 474 * causes this case to run millions of times. 475 */ 476 if (is_zero_pfn(pfn)) 477 return; 478 479 this_cpu = get_cpu(); 480 481 mapping = folio_flush_mapping(folio); 482 if (mapping && !mapping_mapped(mapping)) { 483 bool dirty = test_bit(PG_dcache_dirty, &folio->flags.f); 484 if (dirty) { 485 int dirty_cpu = dcache_dirty_cpu(folio); 486 487 if (dirty_cpu == this_cpu) 488 goto out; 489 smp_flush_dcache_folio_impl(folio, dirty_cpu); 490 } 491 set_dcache_dirty(folio, this_cpu); 492 } else { 493 /* We could delay the flush for the !folio_mapping 494 * case too. But that case is for exec env/arg 495 * pages and those are %99 certainly going to get 496 * faulted into the tlb (and thus flushed) anyways. 497 */ 498 flush_dcache_folio_impl(folio); 499 } 500 501 out: 502 put_cpu(); 503 } 504 EXPORT_SYMBOL(flush_dcache_folio); 505 506 void __kprobes flush_icache_range(unsigned long start, unsigned long end) 507 { 508 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */ 509 if (tlb_type == spitfire) { 510 unsigned long kaddr; 511 512 /* This code only runs on Spitfire cpus so this is 513 * why we can assume _PAGE_PADDR_4U. 514 */ 515 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) { 516 unsigned long paddr, mask = _PAGE_PADDR_4U; 517 518 if (kaddr >= PAGE_OFFSET) 519 paddr = kaddr & mask; 520 else { 521 pte_t *ptep = virt_to_kpte(kaddr); 522 523 paddr = pte_val(*ptep) & mask; 524 } 525 __flush_icache_page(paddr); 526 } 527 } 528 } 529 EXPORT_SYMBOL(flush_icache_range); 530 531 void mmu_info(struct seq_file *m) 532 { 533 static const char *pgsz_strings[] = { 534 "8K", "64K", "512K", "4MB", "32MB", 535 "256MB", "2GB", "16GB", 536 }; 537 int i, printed; 538 539 if (tlb_type == cheetah) 540 seq_printf(m, "MMU Type\t: Cheetah\n"); 541 else if (tlb_type == cheetah_plus) 542 seq_printf(m, "MMU Type\t: Cheetah+\n"); 543 else if (tlb_type == spitfire) 544 seq_printf(m, "MMU Type\t: Spitfire\n"); 545 else if (tlb_type == hypervisor) 546 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n"); 547 else 548 seq_printf(m, "MMU Type\t: ???\n"); 549 550 seq_printf(m, "MMU PGSZs\t: "); 551 printed = 0; 552 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) { 553 if (cpu_pgsz_mask & (1UL << i)) { 554 seq_printf(m, "%s%s", 555 printed ? "," : "", pgsz_strings[i]); 556 printed++; 557 } 558 } 559 seq_putc(m, '\n'); 560 561 #ifdef CONFIG_DEBUG_DCFLUSH 562 seq_printf(m, "DCPageFlushes\t: %d\n", 563 atomic_read(&dcpage_flushes)); 564 #ifdef CONFIG_SMP 565 seq_printf(m, "DCPageFlushesXC\t: %d\n", 566 atomic_read(&dcpage_flushes_xcall)); 567 #endif /* CONFIG_SMP */ 568 #endif /* CONFIG_DEBUG_DCFLUSH */ 569 } 570 571 struct linux_prom_translation prom_trans[512] __read_mostly; 572 unsigned int prom_trans_ents __read_mostly; 573 574 unsigned long kern_locked_tte_data; 575 576 /* The obp translations are saved based on 8k pagesize, since obp can 577 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS -> 578 * HI_OBP_ADDRESS range are handled in ktlb.S. 579 */ 580 static inline int in_obp_range(unsigned long vaddr) 581 { 582 return (vaddr >= LOW_OBP_ADDRESS && 583 vaddr < HI_OBP_ADDRESS); 584 } 585 586 static int cmp_ptrans(const void *a, const void *b) 587 { 588 const struct linux_prom_translation *x = a, *y = b; 589 590 if (x->virt > y->virt) 591 return 1; 592 if (x->virt < y->virt) 593 return -1; 594 return 0; 595 } 596 597 /* Read OBP translations property into 'prom_trans[]'. */ 598 static void __init read_obp_translations(void) 599 { 600 int n, node, ents, first, last, i; 601 602 node = prom_finddevice("/virtual-memory"); 603 n = prom_getproplen(node, "translations"); 604 if (unlikely(n == 0 || n == -1)) { 605 prom_printf("prom_mappings: Couldn't get size.\n"); 606 prom_halt(); 607 } 608 if (unlikely(n > sizeof(prom_trans))) { 609 prom_printf("prom_mappings: Size %d is too big.\n", n); 610 prom_halt(); 611 } 612 613 if ((n = prom_getproperty(node, "translations", 614 (char *)&prom_trans[0], 615 sizeof(prom_trans))) == -1) { 616 prom_printf("prom_mappings: Couldn't get property.\n"); 617 prom_halt(); 618 } 619 620 n = n / sizeof(struct linux_prom_translation); 621 622 ents = n; 623 624 sort(prom_trans, ents, sizeof(struct linux_prom_translation), 625 cmp_ptrans, NULL); 626 627 /* Now kick out all the non-OBP entries. */ 628 for (i = 0; i < ents; i++) { 629 if (in_obp_range(prom_trans[i].virt)) 630 break; 631 } 632 first = i; 633 for (; i < ents; i++) { 634 if (!in_obp_range(prom_trans[i].virt)) 635 break; 636 } 637 last = i; 638 639 for (i = 0; i < (last - first); i++) { 640 struct linux_prom_translation *src = &prom_trans[i + first]; 641 struct linux_prom_translation *dest = &prom_trans[i]; 642 643 *dest = *src; 644 } 645 for (; i < ents; i++) { 646 struct linux_prom_translation *dest = &prom_trans[i]; 647 dest->virt = dest->size = dest->data = 0x0UL; 648 } 649 650 prom_trans_ents = last - first; 651 652 if (tlb_type == spitfire) { 653 /* Clear diag TTE bits. */ 654 for (i = 0; i < prom_trans_ents; i++) 655 prom_trans[i].data &= ~0x0003fe0000000000UL; 656 } 657 658 /* Force execute bit on. */ 659 for (i = 0; i < prom_trans_ents; i++) 660 prom_trans[i].data |= (tlb_type == hypervisor ? 661 _PAGE_EXEC_4V : _PAGE_EXEC_4U); 662 } 663 664 static void __init hypervisor_tlb_lock(unsigned long vaddr, 665 unsigned long pte, 666 unsigned long mmu) 667 { 668 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu); 669 670 if (ret != 0) { 671 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: " 672 "errors with %lx\n", vaddr, 0, pte, mmu, ret); 673 prom_halt(); 674 } 675 } 676 677 static unsigned long kern_large_tte(unsigned long paddr); 678 679 static void __init remap_kernel(void) 680 { 681 unsigned long phys_page, tte_vaddr, tte_data; 682 int i, tlb_ent = sparc64_highest_locked_tlbent(); 683 684 tte_vaddr = (unsigned long) KERNBASE; 685 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB; 686 tte_data = kern_large_tte(phys_page); 687 688 kern_locked_tte_data = tte_data; 689 690 /* Now lock us into the TLBs via Hypervisor or OBP. */ 691 if (tlb_type == hypervisor) { 692 for (i = 0; i < num_kernel_image_mappings; i++) { 693 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU); 694 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU); 695 tte_vaddr += 0x400000; 696 tte_data += 0x400000; 697 } 698 } else { 699 for (i = 0; i < num_kernel_image_mappings; i++) { 700 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr); 701 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr); 702 tte_vaddr += 0x400000; 703 tte_data += 0x400000; 704 } 705 sparc64_highest_unlocked_tlb_ent = tlb_ent - i; 706 } 707 if (tlb_type == cheetah_plus) { 708 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 | 709 CTX_CHEETAH_PLUS_NUC); 710 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC; 711 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0; 712 } 713 } 714 715 716 static void __init inherit_prom_mappings(void) 717 { 718 /* Now fixup OBP's idea about where we really are mapped. */ 719 printk("Remapping the kernel... "); 720 remap_kernel(); 721 printk("done.\n"); 722 } 723 724 void prom_world(int enter) 725 { 726 /* 727 * No need to change the address space any more, just flush 728 * the register windows 729 */ 730 __asm__ __volatile__("flushw"); 731 } 732 733 void __flush_dcache_range(unsigned long start, unsigned long end) 734 { 735 unsigned long va; 736 737 if (tlb_type == spitfire) { 738 int n = 0; 739 740 for (va = start; va < end; va += 32) { 741 spitfire_put_dcache_tag(va & 0x3fe0, 0x0); 742 if (++n >= 512) 743 break; 744 } 745 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) { 746 start = __pa(start); 747 end = __pa(end); 748 for (va = start; va < end; va += 32) 749 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 750 "membar #Sync" 751 : /* no outputs */ 752 : "r" (va), 753 "i" (ASI_DCACHE_INVALIDATE)); 754 } 755 } 756 EXPORT_SYMBOL(__flush_dcache_range); 757 758 /* get_new_mmu_context() uses "cache + 1". */ 759 DEFINE_SPINLOCK(ctx_alloc_lock); 760 unsigned long tlb_context_cache = CTX_FIRST_VERSION; 761 #define MAX_CTX_NR (1UL << CTX_NR_BITS) 762 #define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR) 763 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR); 764 DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0}; 765 766 static void mmu_context_wrap(void) 767 { 768 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK; 769 unsigned long new_ver, new_ctx, old_ctx; 770 struct mm_struct *mm; 771 int cpu; 772 773 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS); 774 775 /* Reserve kernel context */ 776 set_bit(0, mmu_context_bmap); 777 778 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION; 779 if (unlikely(new_ver == 0)) 780 new_ver = CTX_FIRST_VERSION; 781 tlb_context_cache = new_ver; 782 783 /* 784 * Make sure that any new mm that are added into per_cpu_secondary_mm, 785 * are going to go through get_new_mmu_context() path. 786 */ 787 mb(); 788 789 /* 790 * Updated versions to current on those CPUs that had valid secondary 791 * contexts 792 */ 793 for_each_online_cpu(cpu) { 794 /* 795 * If a new mm is stored after we took this mm from the array, 796 * it will go into get_new_mmu_context() path, because we 797 * already bumped the version in tlb_context_cache. 798 */ 799 mm = per_cpu(per_cpu_secondary_mm, cpu); 800 801 if (unlikely(!mm || mm == &init_mm)) 802 continue; 803 804 old_ctx = mm->context.sparc64_ctx_val; 805 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) { 806 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver; 807 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap); 808 mm->context.sparc64_ctx_val = new_ctx; 809 } 810 } 811 } 812 813 /* Caller does TLB context flushing on local CPU if necessary. 814 * The caller also ensures that CTX_VALID(mm->context) is false. 815 * 816 * We must be careful about boundary cases so that we never 817 * let the user have CTX 0 (nucleus) or we ever use a CTX 818 * version of zero (and thus NO_CONTEXT would not be caught 819 * by version mis-match tests in mmu_context.h). 820 * 821 * Always invoked with interrupts disabled. 822 */ 823 void get_new_mmu_context(struct mm_struct *mm) 824 { 825 unsigned long ctx, new_ctx; 826 unsigned long orig_pgsz_bits; 827 828 spin_lock(&ctx_alloc_lock); 829 retry: 830 /* wrap might have happened, test again if our context became valid */ 831 if (unlikely(CTX_VALID(mm->context))) 832 goto out; 833 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK); 834 ctx = (tlb_context_cache + 1) & CTX_NR_MASK; 835 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx); 836 if (new_ctx >= (1 << CTX_NR_BITS)) { 837 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1); 838 if (new_ctx >= ctx) { 839 mmu_context_wrap(); 840 goto retry; 841 } 842 } 843 if (mm->context.sparc64_ctx_val) 844 cpumask_clear(mm_cpumask(mm)); 845 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63)); 846 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK); 847 tlb_context_cache = new_ctx; 848 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits; 849 out: 850 spin_unlock(&ctx_alloc_lock); 851 } 852 853 static int numa_enabled = 1; 854 static int numa_debug; 855 856 static int __init early_numa(char *p) 857 { 858 if (!p) 859 return 0; 860 861 if (strstr(p, "off")) 862 numa_enabled = 0; 863 864 if (strstr(p, "debug")) 865 numa_debug = 1; 866 867 return 0; 868 } 869 early_param("numa", early_numa); 870 871 #define numadbg(f, a...) \ 872 do { if (numa_debug) \ 873 printk(KERN_INFO f, ## a); \ 874 } while (0) 875 876 static void __init find_ramdisk(unsigned long phys_base) 877 { 878 #ifdef CONFIG_BLK_DEV_INITRD 879 if (sparc_ramdisk_image || sparc_ramdisk_image64) { 880 unsigned long ramdisk_image; 881 882 /* Older versions of the bootloader only supported a 883 * 32-bit physical address for the ramdisk image 884 * location, stored at sparc_ramdisk_image. Newer 885 * SILO versions set sparc_ramdisk_image to zero and 886 * provide a full 64-bit physical address at 887 * sparc_ramdisk_image64. 888 */ 889 ramdisk_image = sparc_ramdisk_image; 890 if (!ramdisk_image) 891 ramdisk_image = sparc_ramdisk_image64; 892 893 /* Another bootloader quirk. The bootloader normalizes 894 * the physical address to KERNBASE, so we have to 895 * factor that back out and add in the lowest valid 896 * physical page address to get the true physical address. 897 */ 898 ramdisk_image -= KERNBASE; 899 ramdisk_image += phys_base; 900 901 numadbg("Found ramdisk at physical address 0x%lx, size %u\n", 902 ramdisk_image, sparc_ramdisk_size); 903 904 initrd_start = ramdisk_image; 905 initrd_end = ramdisk_image + sparc_ramdisk_size; 906 907 memblock_reserve(initrd_start, sparc_ramdisk_size); 908 909 initrd_start += PAGE_OFFSET; 910 initrd_end += PAGE_OFFSET; 911 } 912 #endif 913 } 914 915 struct node_mem_mask { 916 unsigned long mask; 917 unsigned long match; 918 }; 919 static struct node_mem_mask node_masks[MAX_NUMNODES]; 920 static int num_node_masks; 921 922 #ifdef CONFIG_NUMA 923 924 struct mdesc_mlgroup { 925 u64 node; 926 u64 latency; 927 u64 match; 928 u64 mask; 929 }; 930 931 static struct mdesc_mlgroup *mlgroups; 932 static int num_mlgroups; 933 934 int numa_cpu_lookup_table[NR_CPUS]; 935 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES]; 936 937 struct mdesc_mblock { 938 u64 base; 939 u64 size; 940 u64 offset; /* RA-to-PA */ 941 }; 942 static struct mdesc_mblock *mblocks; 943 static int num_mblocks; 944 945 static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr) 946 { 947 struct mdesc_mblock *m = NULL; 948 int i; 949 950 for (i = 0; i < num_mblocks; i++) { 951 m = &mblocks[i]; 952 953 if (addr >= m->base && 954 addr < (m->base + m->size)) { 955 break; 956 } 957 } 958 959 return m; 960 } 961 962 static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid) 963 { 964 int prev_nid, new_nid; 965 966 prev_nid = NUMA_NO_NODE; 967 for ( ; start < end; start += PAGE_SIZE) { 968 for (new_nid = 0; new_nid < num_node_masks; new_nid++) { 969 struct node_mem_mask *p = &node_masks[new_nid]; 970 971 if ((start & p->mask) == p->match) { 972 if (prev_nid == NUMA_NO_NODE) 973 prev_nid = new_nid; 974 break; 975 } 976 } 977 978 if (new_nid == num_node_masks) { 979 prev_nid = 0; 980 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.", 981 start); 982 break; 983 } 984 985 if (prev_nid != new_nid) 986 break; 987 } 988 *nid = prev_nid; 989 990 return start > end ? end : start; 991 } 992 993 static u64 __init memblock_nid_range(u64 start, u64 end, int *nid) 994 { 995 u64 ret_end, pa_start, m_mask, m_match, m_end; 996 struct mdesc_mblock *mblock; 997 int _nid, i; 998 999 if (tlb_type != hypervisor) 1000 return memblock_nid_range_sun4u(start, end, nid); 1001 1002 mblock = addr_to_mblock(start); 1003 if (!mblock) { 1004 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]", 1005 start); 1006 1007 _nid = 0; 1008 ret_end = end; 1009 goto done; 1010 } 1011 1012 pa_start = start + mblock->offset; 1013 m_match = 0; 1014 m_mask = 0; 1015 1016 for (_nid = 0; _nid < num_node_masks; _nid++) { 1017 struct node_mem_mask *const m = &node_masks[_nid]; 1018 1019 if ((pa_start & m->mask) == m->match) { 1020 m_match = m->match; 1021 m_mask = m->mask; 1022 break; 1023 } 1024 } 1025 1026 if (num_node_masks == _nid) { 1027 /* We could not find NUMA group, so default to 0, but lets 1028 * search for latency group, so we could calculate the correct 1029 * end address that we return 1030 */ 1031 _nid = 0; 1032 1033 for (i = 0; i < num_mlgroups; i++) { 1034 struct mdesc_mlgroup *const m = &mlgroups[i]; 1035 1036 if ((pa_start & m->mask) == m->match) { 1037 m_match = m->match; 1038 m_mask = m->mask; 1039 break; 1040 } 1041 } 1042 1043 if (i == num_mlgroups) { 1044 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]", 1045 start); 1046 1047 ret_end = end; 1048 goto done; 1049 } 1050 } 1051 1052 /* 1053 * Each latency group has match and mask, and each memory block has an 1054 * offset. An address belongs to a latency group if its address matches 1055 * the following formula: ((addr + offset) & mask) == match 1056 * It is, however, slow to check every single page if it matches a 1057 * particular latency group. As optimization we calculate end value by 1058 * using bit arithmetics. 1059 */ 1060 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset; 1061 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1); 1062 ret_end = m_end > end ? end : m_end; 1063 1064 done: 1065 *nid = _nid; 1066 return ret_end; 1067 } 1068 #endif 1069 1070 /* This must be invoked after performing all of the necessary 1071 * memblock_set_node() calls for 'nid'. We need to be able to get 1072 * correct data from get_pfn_range_for_nid(). 1073 */ 1074 static void __init allocate_node_data(int nid) 1075 { 1076 struct pglist_data *p; 1077 unsigned long start_pfn, end_pfn; 1078 1079 #ifdef CONFIG_NUMA 1080 alloc_node_data(nid); 1081 1082 NODE_DATA(nid)->node_id = nid; 1083 #endif 1084 1085 p = NODE_DATA(nid); 1086 1087 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 1088 p->node_start_pfn = start_pfn; 1089 p->node_spanned_pages = end_pfn - start_pfn; 1090 } 1091 1092 static void init_node_masks_nonnuma(void) 1093 { 1094 #ifdef CONFIG_NUMA 1095 int i; 1096 #endif 1097 1098 numadbg("Initializing tables for non-numa.\n"); 1099 1100 node_masks[0].mask = 0; 1101 node_masks[0].match = 0; 1102 num_node_masks = 1; 1103 1104 #ifdef CONFIG_NUMA 1105 for (i = 0; i < NR_CPUS; i++) 1106 numa_cpu_lookup_table[i] = 0; 1107 1108 cpumask_setall(&numa_cpumask_lookup_table[0]); 1109 #endif 1110 } 1111 1112 #ifdef CONFIG_NUMA 1113 1114 EXPORT_SYMBOL(numa_cpu_lookup_table); 1115 EXPORT_SYMBOL(numa_cpumask_lookup_table); 1116 1117 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio, 1118 u32 cfg_handle) 1119 { 1120 u64 arc; 1121 1122 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) { 1123 u64 target = mdesc_arc_target(md, arc); 1124 const u64 *val; 1125 1126 val = mdesc_get_property(md, target, 1127 "cfg-handle", NULL); 1128 if (val && *val == cfg_handle) 1129 return 0; 1130 } 1131 return -ENODEV; 1132 } 1133 1134 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp, 1135 u32 cfg_handle) 1136 { 1137 u64 arc, candidate, best_latency = ~(u64)0; 1138 1139 candidate = MDESC_NODE_NULL; 1140 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1141 u64 target = mdesc_arc_target(md, arc); 1142 const char *name = mdesc_node_name(md, target); 1143 const u64 *val; 1144 1145 if (strcmp(name, "pio-latency-group")) 1146 continue; 1147 1148 val = mdesc_get_property(md, target, "latency", NULL); 1149 if (!val) 1150 continue; 1151 1152 if (*val < best_latency) { 1153 candidate = target; 1154 best_latency = *val; 1155 } 1156 } 1157 1158 if (candidate == MDESC_NODE_NULL) 1159 return -ENODEV; 1160 1161 return scan_pio_for_cfg_handle(md, candidate, cfg_handle); 1162 } 1163 1164 int of_node_to_nid(struct device_node *dp) 1165 { 1166 const struct linux_prom64_registers *regs; 1167 struct mdesc_handle *md; 1168 u32 cfg_handle; 1169 int count, nid; 1170 u64 grp; 1171 1172 /* This is the right thing to do on currently supported 1173 * SUN4U NUMA platforms as well, as the PCI controller does 1174 * not sit behind any particular memory controller. 1175 */ 1176 if (!mlgroups) 1177 return -1; 1178 1179 regs = of_get_property(dp, "reg", NULL); 1180 if (!regs) 1181 return -1; 1182 1183 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff; 1184 1185 md = mdesc_grab(); 1186 1187 count = 0; 1188 nid = NUMA_NO_NODE; 1189 mdesc_for_each_node_by_name(md, grp, "group") { 1190 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) { 1191 nid = count; 1192 break; 1193 } 1194 count++; 1195 } 1196 1197 mdesc_release(md); 1198 1199 return nid; 1200 } 1201 1202 static void __init add_node_ranges(void) 1203 { 1204 phys_addr_t start, end; 1205 unsigned long prev_max; 1206 u64 i; 1207 1208 memblock_resized: 1209 prev_max = memblock.memory.max; 1210 1211 for_each_mem_range(i, &start, &end) { 1212 while (start < end) { 1213 unsigned long this_end; 1214 int nid; 1215 1216 this_end = memblock_nid_range(start, end, &nid); 1217 1218 numadbg("Setting memblock NUMA node nid[%d] " 1219 "start[%llx] end[%lx]\n", 1220 nid, start, this_end); 1221 1222 memblock_set_node(start, this_end - start, 1223 &memblock.memory, nid); 1224 if (memblock.memory.max != prev_max) 1225 goto memblock_resized; 1226 start = this_end; 1227 } 1228 } 1229 } 1230 1231 static int __init grab_mlgroups(struct mdesc_handle *md) 1232 { 1233 unsigned long paddr; 1234 int count = 0; 1235 u64 node; 1236 1237 mdesc_for_each_node_by_name(md, node, "memory-latency-group") 1238 count++; 1239 if (!count) 1240 return -ENOENT; 1241 1242 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mlgroup), 1243 SMP_CACHE_BYTES); 1244 if (!paddr) 1245 return -ENOMEM; 1246 1247 mlgroups = __va(paddr); 1248 num_mlgroups = count; 1249 1250 count = 0; 1251 mdesc_for_each_node_by_name(md, node, "memory-latency-group") { 1252 struct mdesc_mlgroup *m = &mlgroups[count++]; 1253 const u64 *val; 1254 1255 m->node = node; 1256 1257 val = mdesc_get_property(md, node, "latency", NULL); 1258 m->latency = *val; 1259 val = mdesc_get_property(md, node, "address-match", NULL); 1260 m->match = *val; 1261 val = mdesc_get_property(md, node, "address-mask", NULL); 1262 m->mask = *val; 1263 1264 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] " 1265 "match[%llx] mask[%llx]\n", 1266 count - 1, m->node, m->latency, m->match, m->mask); 1267 } 1268 1269 return 0; 1270 } 1271 1272 static int __init grab_mblocks(struct mdesc_handle *md) 1273 { 1274 unsigned long paddr; 1275 int count = 0; 1276 u64 node; 1277 1278 mdesc_for_each_node_by_name(md, node, "mblock") 1279 count++; 1280 if (!count) 1281 return -ENOENT; 1282 1283 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mblock), 1284 SMP_CACHE_BYTES); 1285 if (!paddr) 1286 return -ENOMEM; 1287 1288 mblocks = __va(paddr); 1289 num_mblocks = count; 1290 1291 count = 0; 1292 mdesc_for_each_node_by_name(md, node, "mblock") { 1293 struct mdesc_mblock *m = &mblocks[count++]; 1294 const u64 *val; 1295 1296 val = mdesc_get_property(md, node, "base", NULL); 1297 m->base = *val; 1298 val = mdesc_get_property(md, node, "size", NULL); 1299 m->size = *val; 1300 val = mdesc_get_property(md, node, 1301 "address-congruence-offset", NULL); 1302 1303 /* The address-congruence-offset property is optional. 1304 * Explicity zero it be identifty this. 1305 */ 1306 if (val) 1307 m->offset = *val; 1308 else 1309 m->offset = 0UL; 1310 1311 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n", 1312 count - 1, m->base, m->size, m->offset); 1313 } 1314 1315 return 0; 1316 } 1317 1318 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md, 1319 u64 grp, cpumask_t *mask) 1320 { 1321 u64 arc; 1322 1323 cpumask_clear(mask); 1324 1325 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) { 1326 u64 target = mdesc_arc_target(md, arc); 1327 const char *name = mdesc_node_name(md, target); 1328 const u64 *id; 1329 1330 if (strcmp(name, "cpu")) 1331 continue; 1332 id = mdesc_get_property(md, target, "id", NULL); 1333 if (*id < nr_cpu_ids) 1334 cpumask_set_cpu(*id, mask); 1335 } 1336 } 1337 1338 static struct mdesc_mlgroup * __init find_mlgroup(u64 node) 1339 { 1340 int i; 1341 1342 for (i = 0; i < num_mlgroups; i++) { 1343 struct mdesc_mlgroup *m = &mlgroups[i]; 1344 if (m->node == node) 1345 return m; 1346 } 1347 return NULL; 1348 } 1349 1350 int __node_distance(int from, int to) 1351 { 1352 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) { 1353 pr_warn("Returning default NUMA distance value for %d->%d\n", 1354 from, to); 1355 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE; 1356 } 1357 return numa_latency[from][to]; 1358 } 1359 EXPORT_SYMBOL(__node_distance); 1360 1361 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp) 1362 { 1363 int i; 1364 1365 for (i = 0; i < MAX_NUMNODES; i++) { 1366 struct node_mem_mask *n = &node_masks[i]; 1367 1368 if ((grp->mask == n->mask) && (grp->match == n->match)) 1369 break; 1370 } 1371 return i; 1372 } 1373 1374 static void __init find_numa_latencies_for_group(struct mdesc_handle *md, 1375 u64 grp, int index) 1376 { 1377 u64 arc; 1378 1379 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1380 int tnode; 1381 u64 target = mdesc_arc_target(md, arc); 1382 struct mdesc_mlgroup *m = find_mlgroup(target); 1383 1384 if (!m) 1385 continue; 1386 tnode = find_best_numa_node_for_mlgroup(m); 1387 if (tnode == MAX_NUMNODES) 1388 continue; 1389 numa_latency[index][tnode] = m->latency; 1390 } 1391 } 1392 1393 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp, 1394 int index) 1395 { 1396 struct mdesc_mlgroup *candidate = NULL; 1397 u64 arc, best_latency = ~(u64)0; 1398 struct node_mem_mask *n; 1399 1400 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1401 u64 target = mdesc_arc_target(md, arc); 1402 struct mdesc_mlgroup *m = find_mlgroup(target); 1403 if (!m) 1404 continue; 1405 if (m->latency < best_latency) { 1406 candidate = m; 1407 best_latency = m->latency; 1408 } 1409 } 1410 if (!candidate) 1411 return -ENOENT; 1412 1413 if (num_node_masks != index) { 1414 printk(KERN_ERR "Inconsistent NUMA state, " 1415 "index[%d] != num_node_masks[%d]\n", 1416 index, num_node_masks); 1417 return -EINVAL; 1418 } 1419 1420 n = &node_masks[num_node_masks++]; 1421 1422 n->mask = candidate->mask; 1423 n->match = candidate->match; 1424 1425 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n", 1426 index, n->mask, n->match, candidate->latency); 1427 1428 return 0; 1429 } 1430 1431 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp, 1432 int index) 1433 { 1434 cpumask_t mask; 1435 int cpu; 1436 1437 numa_parse_mdesc_group_cpus(md, grp, &mask); 1438 1439 for_each_cpu(cpu, &mask) 1440 numa_cpu_lookup_table[cpu] = index; 1441 cpumask_copy(&numa_cpumask_lookup_table[index], &mask); 1442 1443 if (numa_debug) { 1444 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index); 1445 for_each_cpu(cpu, &mask) 1446 printk("%d ", cpu); 1447 printk("]\n"); 1448 } 1449 1450 return numa_attach_mlgroup(md, grp, index); 1451 } 1452 1453 static int __init numa_parse_mdesc(void) 1454 { 1455 struct mdesc_handle *md = mdesc_grab(); 1456 int i, j, err, count; 1457 u64 node; 1458 1459 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups"); 1460 if (node == MDESC_NODE_NULL) { 1461 mdesc_release(md); 1462 return -ENOENT; 1463 } 1464 1465 err = grab_mblocks(md); 1466 if (err < 0) 1467 goto out; 1468 1469 err = grab_mlgroups(md); 1470 if (err < 0) 1471 goto out; 1472 1473 count = 0; 1474 mdesc_for_each_node_by_name(md, node, "group") { 1475 err = numa_parse_mdesc_group(md, node, count); 1476 if (err < 0) 1477 break; 1478 count++; 1479 } 1480 1481 count = 0; 1482 mdesc_for_each_node_by_name(md, node, "group") { 1483 find_numa_latencies_for_group(md, node, count); 1484 count++; 1485 } 1486 1487 /* Normalize numa latency matrix according to ACPI SLIT spec. */ 1488 for (i = 0; i < MAX_NUMNODES; i++) { 1489 u64 self_latency = numa_latency[i][i]; 1490 1491 for (j = 0; j < MAX_NUMNODES; j++) { 1492 numa_latency[i][j] = 1493 (numa_latency[i][j] * LOCAL_DISTANCE) / 1494 self_latency; 1495 } 1496 } 1497 1498 add_node_ranges(); 1499 1500 for (i = 0; i < num_node_masks; i++) { 1501 allocate_node_data(i); 1502 node_set_online(i); 1503 } 1504 1505 err = 0; 1506 out: 1507 mdesc_release(md); 1508 return err; 1509 } 1510 1511 static int __init numa_parse_jbus(void) 1512 { 1513 unsigned long cpu, index; 1514 1515 /* NUMA node id is encoded in bits 36 and higher, and there is 1516 * a 1-to-1 mapping from CPU ID to NUMA node ID. 1517 */ 1518 index = 0; 1519 for_each_present_cpu(cpu) { 1520 numa_cpu_lookup_table[cpu] = index; 1521 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu)); 1522 node_masks[index].mask = ~((1UL << 36UL) - 1UL); 1523 node_masks[index].match = cpu << 36UL; 1524 1525 index++; 1526 } 1527 num_node_masks = index; 1528 1529 add_node_ranges(); 1530 1531 for (index = 0; index < num_node_masks; index++) { 1532 allocate_node_data(index); 1533 node_set_online(index); 1534 } 1535 1536 return 0; 1537 } 1538 1539 static int __init numa_parse_sun4u(void) 1540 { 1541 if (tlb_type == cheetah || tlb_type == cheetah_plus) { 1542 unsigned long ver; 1543 1544 __asm__ ("rdpr %%ver, %0" : "=r" (ver)); 1545 if ((ver >> 32UL) == __JALAPENO_ID || 1546 (ver >> 32UL) == __SERRANO_ID) 1547 return numa_parse_jbus(); 1548 } 1549 return -1; 1550 } 1551 1552 static int __init bootmem_init_numa(void) 1553 { 1554 int i, j; 1555 int err = -1; 1556 1557 numadbg("bootmem_init_numa()\n"); 1558 1559 /* Some sane defaults for numa latency values */ 1560 for (i = 0; i < MAX_NUMNODES; i++) { 1561 for (j = 0; j < MAX_NUMNODES; j++) 1562 numa_latency[i][j] = (i == j) ? 1563 LOCAL_DISTANCE : REMOTE_DISTANCE; 1564 } 1565 1566 if (numa_enabled) { 1567 if (tlb_type == hypervisor) 1568 err = numa_parse_mdesc(); 1569 else 1570 err = numa_parse_sun4u(); 1571 } 1572 return err; 1573 } 1574 1575 #else 1576 1577 static int bootmem_init_numa(void) 1578 { 1579 return -1; 1580 } 1581 1582 #endif 1583 1584 static void __init bootmem_init_nonnuma(void) 1585 { 1586 unsigned long top_of_ram = memblock_end_of_DRAM(); 1587 unsigned long total_ram = memblock_phys_mem_size(); 1588 1589 numadbg("bootmem_init_nonnuma()\n"); 1590 1591 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n", 1592 top_of_ram, total_ram); 1593 printk(KERN_INFO "Memory hole size: %ldMB\n", 1594 (top_of_ram - total_ram) >> 20); 1595 1596 init_node_masks_nonnuma(); 1597 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0); 1598 allocate_node_data(0); 1599 node_set_online(0); 1600 } 1601 1602 static unsigned long __init bootmem_init(unsigned long phys_base) 1603 { 1604 unsigned long end_pfn; 1605 1606 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; 1607 max_pfn = max_low_pfn = end_pfn; 1608 min_low_pfn = (phys_base >> PAGE_SHIFT); 1609 1610 if (bootmem_init_numa() < 0) 1611 bootmem_init_nonnuma(); 1612 1613 /* Dump memblock with node info. */ 1614 memblock_dump_all(); 1615 1616 /* XXX cpu notifier XXX */ 1617 1618 return end_pfn; 1619 } 1620 1621 static struct linux_prom64_registers pall[MAX_BANKS] __initdata; 1622 static int pall_ents __initdata; 1623 1624 static unsigned long max_phys_bits = 40; 1625 1626 bool kern_addr_valid(unsigned long addr) 1627 { 1628 pgd_t *pgd; 1629 p4d_t *p4d; 1630 pud_t *pud; 1631 pmd_t *pmd; 1632 pte_t *pte; 1633 1634 if ((long)addr < 0L) { 1635 unsigned long pa = __pa(addr); 1636 1637 if ((pa >> max_phys_bits) != 0UL) 1638 return false; 1639 1640 return pfn_valid(pa >> PAGE_SHIFT); 1641 } 1642 1643 if (addr >= (unsigned long) KERNBASE && 1644 addr < (unsigned long)&_end) 1645 return true; 1646 1647 pgd = pgd_offset_k(addr); 1648 if (pgd_none(*pgd)) 1649 return false; 1650 1651 p4d = p4d_offset(pgd, addr); 1652 if (p4d_none(*p4d)) 1653 return false; 1654 1655 pud = pud_offset(p4d, addr); 1656 if (pud_none(*pud)) 1657 return false; 1658 1659 if (pud_leaf(*pud)) 1660 return pfn_valid(pud_pfn(*pud)); 1661 1662 pmd = pmd_offset(pud, addr); 1663 if (pmd_none(*pmd)) 1664 return false; 1665 1666 if (pmd_leaf(*pmd)) 1667 return pfn_valid(pmd_pfn(*pmd)); 1668 1669 pte = pte_offset_kernel(pmd, addr); 1670 if (pte_none(*pte)) 1671 return false; 1672 1673 return pfn_valid(pte_pfn(*pte)); 1674 } 1675 1676 static unsigned long __ref kernel_map_hugepud(unsigned long vstart, 1677 unsigned long vend, 1678 pud_t *pud) 1679 { 1680 const unsigned long mask16gb = (1UL << 34) - 1UL; 1681 u64 pte_val = vstart; 1682 1683 /* Each PUD is 8GB */ 1684 if ((vstart & mask16gb) || 1685 (vend - vstart <= mask16gb)) { 1686 pte_val ^= kern_linear_pte_xor[2]; 1687 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE; 1688 1689 return vstart + PUD_SIZE; 1690 } 1691 1692 pte_val ^= kern_linear_pte_xor[3]; 1693 pte_val |= _PAGE_PUD_HUGE; 1694 1695 vend = vstart + mask16gb + 1UL; 1696 while (vstart < vend) { 1697 pud_val(*pud) = pte_val; 1698 1699 pte_val += PUD_SIZE; 1700 vstart += PUD_SIZE; 1701 pud++; 1702 } 1703 return vstart; 1704 } 1705 1706 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend, 1707 bool guard) 1708 { 1709 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE) 1710 return true; 1711 1712 return false; 1713 } 1714 1715 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart, 1716 unsigned long vend, 1717 pmd_t *pmd) 1718 { 1719 const unsigned long mask256mb = (1UL << 28) - 1UL; 1720 const unsigned long mask2gb = (1UL << 31) - 1UL; 1721 u64 pte_val = vstart; 1722 1723 /* Each PMD is 8MB */ 1724 if ((vstart & mask256mb) || 1725 (vend - vstart <= mask256mb)) { 1726 pte_val ^= kern_linear_pte_xor[0]; 1727 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE; 1728 1729 return vstart + PMD_SIZE; 1730 } 1731 1732 if ((vstart & mask2gb) || 1733 (vend - vstart <= mask2gb)) { 1734 pte_val ^= kern_linear_pte_xor[1]; 1735 pte_val |= _PAGE_PMD_HUGE; 1736 vend = vstart + mask256mb + 1UL; 1737 } else { 1738 pte_val ^= kern_linear_pte_xor[2]; 1739 pte_val |= _PAGE_PMD_HUGE; 1740 vend = vstart + mask2gb + 1UL; 1741 } 1742 1743 while (vstart < vend) { 1744 pmd_val(*pmd) = pte_val; 1745 1746 pte_val += PMD_SIZE; 1747 vstart += PMD_SIZE; 1748 pmd++; 1749 } 1750 1751 return vstart; 1752 } 1753 1754 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend, 1755 bool guard) 1756 { 1757 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE) 1758 return true; 1759 1760 return false; 1761 } 1762 1763 static unsigned long __ref kernel_map_range(unsigned long pstart, 1764 unsigned long pend, pgprot_t prot, 1765 bool use_huge) 1766 { 1767 unsigned long vstart = PAGE_OFFSET + pstart; 1768 unsigned long vend = PAGE_OFFSET + pend; 1769 unsigned long alloc_bytes = 0UL; 1770 1771 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) { 1772 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n", 1773 vstart, vend); 1774 prom_halt(); 1775 } 1776 1777 while (vstart < vend) { 1778 unsigned long this_end, paddr = __pa(vstart); 1779 pgd_t *pgd = pgd_offset_k(vstart); 1780 p4d_t *p4d; 1781 pud_t *pud; 1782 pmd_t *pmd; 1783 pte_t *pte; 1784 1785 if (pgd_none(*pgd)) { 1786 pud_t *new; 1787 1788 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1789 PAGE_SIZE); 1790 if (!new) 1791 goto err_alloc; 1792 alloc_bytes += PAGE_SIZE; 1793 pgd_populate(&init_mm, pgd, new); 1794 } 1795 1796 p4d = p4d_offset(pgd, vstart); 1797 if (p4d_none(*p4d)) { 1798 pud_t *new; 1799 1800 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1801 PAGE_SIZE); 1802 if (!new) 1803 goto err_alloc; 1804 alloc_bytes += PAGE_SIZE; 1805 p4d_populate(&init_mm, p4d, new); 1806 } 1807 1808 pud = pud_offset(p4d, vstart); 1809 if (pud_none(*pud)) { 1810 pmd_t *new; 1811 1812 if (kernel_can_map_hugepud(vstart, vend, use_huge)) { 1813 vstart = kernel_map_hugepud(vstart, vend, pud); 1814 continue; 1815 } 1816 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1817 PAGE_SIZE); 1818 if (!new) 1819 goto err_alloc; 1820 alloc_bytes += PAGE_SIZE; 1821 pud_populate(&init_mm, pud, new); 1822 } 1823 1824 pmd = pmd_offset(pud, vstart); 1825 if (pmd_none(*pmd)) { 1826 pte_t *new; 1827 1828 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) { 1829 vstart = kernel_map_hugepmd(vstart, vend, pmd); 1830 continue; 1831 } 1832 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1833 PAGE_SIZE); 1834 if (!new) 1835 goto err_alloc; 1836 alloc_bytes += PAGE_SIZE; 1837 pmd_populate_kernel(&init_mm, pmd, new); 1838 } 1839 1840 pte = pte_offset_kernel(pmd, vstart); 1841 this_end = (vstart + PMD_SIZE) & PMD_MASK; 1842 if (this_end > vend) 1843 this_end = vend; 1844 1845 while (vstart < this_end) { 1846 pte_val(*pte) = (paddr | pgprot_val(prot)); 1847 1848 vstart += PAGE_SIZE; 1849 paddr += PAGE_SIZE; 1850 pte++; 1851 } 1852 } 1853 1854 return alloc_bytes; 1855 1856 err_alloc: 1857 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n", 1858 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE); 1859 return -ENOMEM; 1860 } 1861 1862 static void __init flush_all_kernel_tsbs(void) 1863 { 1864 int i; 1865 1866 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) { 1867 struct tsb *ent = &swapper_tsb[i]; 1868 1869 ent->tag = (1UL << TSB_TAG_INVALID_BIT); 1870 } 1871 #ifndef CONFIG_DEBUG_PAGEALLOC 1872 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) { 1873 struct tsb *ent = &swapper_4m_tsb[i]; 1874 1875 ent->tag = (1UL << TSB_TAG_INVALID_BIT); 1876 } 1877 #endif 1878 } 1879 1880 extern unsigned int kvmap_linear_patch[1]; 1881 1882 static void __init kernel_physical_mapping_init(void) 1883 { 1884 unsigned long i, mem_alloced = 0UL; 1885 bool use_huge = true; 1886 1887 #ifdef CONFIG_DEBUG_PAGEALLOC 1888 use_huge = false; 1889 #endif 1890 for (i = 0; i < pall_ents; i++) { 1891 unsigned long phys_start, phys_end; 1892 1893 phys_start = pall[i].phys_addr; 1894 phys_end = phys_start + pall[i].reg_size; 1895 1896 mem_alloced += kernel_map_range(phys_start, phys_end, 1897 PAGE_KERNEL, use_huge); 1898 } 1899 1900 printk("Allocated %ld bytes for kernel page tables.\n", 1901 mem_alloced); 1902 1903 kvmap_linear_patch[0] = 0x01000000; /* nop */ 1904 flushi(&kvmap_linear_patch[0]); 1905 1906 flush_all_kernel_tsbs(); 1907 1908 __flush_tlb_all(); 1909 } 1910 1911 #ifdef CONFIG_DEBUG_PAGEALLOC 1912 void __kernel_map_pages(struct page *page, int numpages, int enable) 1913 { 1914 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT; 1915 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE); 1916 1917 kernel_map_range(phys_start, phys_end, 1918 (enable ? PAGE_KERNEL : __pgprot(0)), false); 1919 1920 flush_tsb_kernel_range(PAGE_OFFSET + phys_start, 1921 PAGE_OFFSET + phys_end); 1922 1923 /* we should perform an IPI and flush all tlbs, 1924 * but that can deadlock->flush only current cpu. 1925 */ 1926 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start, 1927 PAGE_OFFSET + phys_end); 1928 } 1929 #endif 1930 1931 unsigned long __init find_ecache_flush_span(unsigned long size) 1932 { 1933 int i; 1934 1935 for (i = 0; i < pavail_ents; i++) { 1936 if (pavail[i].reg_size >= size) 1937 return pavail[i].phys_addr; 1938 } 1939 1940 return ~0UL; 1941 } 1942 1943 unsigned long PAGE_OFFSET; 1944 EXPORT_SYMBOL(PAGE_OFFSET); 1945 1946 unsigned long VMALLOC_END = 0x0000010000000000UL; 1947 EXPORT_SYMBOL(VMALLOC_END); 1948 1949 unsigned long sparc64_va_hole_top = 0xfffff80000000000UL; 1950 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL; 1951 1952 static void __init setup_page_offset(void) 1953 { 1954 if (tlb_type == cheetah || tlb_type == cheetah_plus) { 1955 /* Cheetah/Panther support a full 64-bit virtual 1956 * address, so we can use all that our page tables 1957 * support. 1958 */ 1959 sparc64_va_hole_top = 0xfff0000000000000UL; 1960 sparc64_va_hole_bottom = 0x0010000000000000UL; 1961 1962 max_phys_bits = 42; 1963 } else if (tlb_type == hypervisor) { 1964 switch (sun4v_chip_type) { 1965 case SUN4V_CHIP_NIAGARA1: 1966 case SUN4V_CHIP_NIAGARA2: 1967 /* T1 and T2 support 48-bit virtual addresses. */ 1968 sparc64_va_hole_top = 0xffff800000000000UL; 1969 sparc64_va_hole_bottom = 0x0000800000000000UL; 1970 1971 max_phys_bits = 39; 1972 break; 1973 case SUN4V_CHIP_NIAGARA3: 1974 /* T3 supports 48-bit virtual addresses. */ 1975 sparc64_va_hole_top = 0xffff800000000000UL; 1976 sparc64_va_hole_bottom = 0x0000800000000000UL; 1977 1978 max_phys_bits = 43; 1979 break; 1980 case SUN4V_CHIP_NIAGARA4: 1981 case SUN4V_CHIP_NIAGARA5: 1982 case SUN4V_CHIP_SPARC64X: 1983 case SUN4V_CHIP_SPARC_M6: 1984 /* T4 and later support 52-bit virtual addresses. */ 1985 sparc64_va_hole_top = 0xfff8000000000000UL; 1986 sparc64_va_hole_bottom = 0x0008000000000000UL; 1987 max_phys_bits = 47; 1988 break; 1989 case SUN4V_CHIP_SPARC_M7: 1990 case SUN4V_CHIP_SPARC_SN: 1991 /* M7 and later support 52-bit virtual addresses. */ 1992 sparc64_va_hole_top = 0xfff8000000000000UL; 1993 sparc64_va_hole_bottom = 0x0008000000000000UL; 1994 max_phys_bits = 49; 1995 break; 1996 case SUN4V_CHIP_SPARC_M8: 1997 default: 1998 /* M8 and later support 54-bit virtual addresses. 1999 * However, restricting M8 and above VA bits to 53 2000 * as 4-level page table cannot support more than 2001 * 53 VA bits. 2002 */ 2003 sparc64_va_hole_top = 0xfff0000000000000UL; 2004 sparc64_va_hole_bottom = 0x0010000000000000UL; 2005 max_phys_bits = 51; 2006 break; 2007 } 2008 } 2009 2010 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) { 2011 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n", 2012 max_phys_bits); 2013 prom_halt(); 2014 } 2015 2016 PAGE_OFFSET = sparc64_va_hole_top; 2017 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) + 2018 (sparc64_va_hole_bottom >> 2)); 2019 2020 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n", 2021 PAGE_OFFSET, max_phys_bits); 2022 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n", 2023 VMALLOC_START, VMALLOC_END); 2024 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n", 2025 VMEMMAP_BASE, VMEMMAP_BASE << 1); 2026 } 2027 2028 static void __init tsb_phys_patch(void) 2029 { 2030 struct tsb_ldquad_phys_patch_entry *pquad; 2031 struct tsb_phys_patch_entry *p; 2032 2033 pquad = &__tsb_ldquad_phys_patch; 2034 while (pquad < &__tsb_ldquad_phys_patch_end) { 2035 unsigned long addr = pquad->addr; 2036 2037 if (tlb_type == hypervisor) 2038 *(unsigned int *) addr = pquad->sun4v_insn; 2039 else 2040 *(unsigned int *) addr = pquad->sun4u_insn; 2041 wmb(); 2042 __asm__ __volatile__("flush %0" 2043 : /* no outputs */ 2044 : "r" (addr)); 2045 2046 pquad++; 2047 } 2048 2049 p = &__tsb_phys_patch; 2050 while (p < &__tsb_phys_patch_end) { 2051 unsigned long addr = p->addr; 2052 2053 *(unsigned int *) addr = p->insn; 2054 wmb(); 2055 __asm__ __volatile__("flush %0" 2056 : /* no outputs */ 2057 : "r" (addr)); 2058 2059 p++; 2060 } 2061 } 2062 2063 /* Don't mark as init, we give this to the Hypervisor. */ 2064 #ifndef CONFIG_DEBUG_PAGEALLOC 2065 #define NUM_KTSB_DESCR 2 2066 #else 2067 #define NUM_KTSB_DESCR 1 2068 #endif 2069 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR]; 2070 2071 /* The swapper TSBs are loaded with a base sequence of: 2072 * 2073 * sethi %uhi(SYMBOL), REG1 2074 * sethi %hi(SYMBOL), REG2 2075 * or REG1, %ulo(SYMBOL), REG1 2076 * or REG2, %lo(SYMBOL), REG2 2077 * sllx REG1, 32, REG1 2078 * or REG1, REG2, REG1 2079 * 2080 * When we use physical addressing for the TSB accesses, we patch the 2081 * first four instructions in the above sequence. 2082 */ 2083 2084 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa) 2085 { 2086 unsigned long high_bits, low_bits; 2087 2088 high_bits = (pa >> 32) & 0xffffffff; 2089 low_bits = (pa >> 0) & 0xffffffff; 2090 2091 while (start < end) { 2092 unsigned int *ia = (unsigned int *)(unsigned long)*start; 2093 2094 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10); 2095 __asm__ __volatile__("flush %0" : : "r" (ia)); 2096 2097 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10); 2098 __asm__ __volatile__("flush %0" : : "r" (ia + 1)); 2099 2100 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff); 2101 __asm__ __volatile__("flush %0" : : "r" (ia + 2)); 2102 2103 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff); 2104 __asm__ __volatile__("flush %0" : : "r" (ia + 3)); 2105 2106 start++; 2107 } 2108 } 2109 2110 static void ktsb_phys_patch(void) 2111 { 2112 extern unsigned int __swapper_tsb_phys_patch; 2113 extern unsigned int __swapper_tsb_phys_patch_end; 2114 unsigned long ktsb_pa; 2115 2116 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE); 2117 patch_one_ktsb_phys(&__swapper_tsb_phys_patch, 2118 &__swapper_tsb_phys_patch_end, ktsb_pa); 2119 #ifndef CONFIG_DEBUG_PAGEALLOC 2120 { 2121 extern unsigned int __swapper_4m_tsb_phys_patch; 2122 extern unsigned int __swapper_4m_tsb_phys_patch_end; 2123 ktsb_pa = (kern_base + 2124 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE)); 2125 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch, 2126 &__swapper_4m_tsb_phys_patch_end, ktsb_pa); 2127 } 2128 #endif 2129 } 2130 2131 static void __init sun4v_ktsb_init(void) 2132 { 2133 unsigned long ktsb_pa; 2134 2135 /* First KTSB for PAGE_SIZE mappings. */ 2136 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE); 2137 2138 switch (PAGE_SIZE) { 2139 case 8 * 1024: 2140 default: 2141 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K; 2142 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K; 2143 break; 2144 2145 case 64 * 1024: 2146 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K; 2147 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K; 2148 break; 2149 2150 case 512 * 1024: 2151 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K; 2152 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K; 2153 break; 2154 2155 case 4 * 1024 * 1024: 2156 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB; 2157 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB; 2158 break; 2159 } 2160 2161 ktsb_descr[0].assoc = 1; 2162 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES; 2163 ktsb_descr[0].ctx_idx = 0; 2164 ktsb_descr[0].tsb_base = ktsb_pa; 2165 ktsb_descr[0].resv = 0; 2166 2167 #ifndef CONFIG_DEBUG_PAGEALLOC 2168 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */ 2169 ktsb_pa = (kern_base + 2170 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE)); 2171 2172 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB; 2173 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB | 2174 HV_PGSZ_MASK_256MB | 2175 HV_PGSZ_MASK_2GB | 2176 HV_PGSZ_MASK_16GB) & 2177 cpu_pgsz_mask); 2178 ktsb_descr[1].assoc = 1; 2179 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES; 2180 ktsb_descr[1].ctx_idx = 0; 2181 ktsb_descr[1].tsb_base = ktsb_pa; 2182 ktsb_descr[1].resv = 0; 2183 #endif 2184 } 2185 2186 void sun4v_ktsb_register(void) 2187 { 2188 unsigned long pa, ret; 2189 2190 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE); 2191 2192 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa); 2193 if (ret != 0) { 2194 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: " 2195 "errors with %lx\n", pa, ret); 2196 prom_halt(); 2197 } 2198 } 2199 2200 static void __init sun4u_linear_pte_xor_finalize(void) 2201 { 2202 #ifndef CONFIG_DEBUG_PAGEALLOC 2203 /* This is where we would add Panther support for 2204 * 32MB and 256MB pages. 2205 */ 2206 #endif 2207 } 2208 2209 static void __init sun4v_linear_pte_xor_finalize(void) 2210 { 2211 unsigned long pagecv_flag; 2212 2213 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead 2214 * enables MCD error. Do not set bit 9 on M7 processor. 2215 */ 2216 switch (sun4v_chip_type) { 2217 case SUN4V_CHIP_SPARC_M7: 2218 case SUN4V_CHIP_SPARC_M8: 2219 case SUN4V_CHIP_SPARC_SN: 2220 pagecv_flag = 0x00; 2221 break; 2222 default: 2223 pagecv_flag = _PAGE_CV_4V; 2224 break; 2225 } 2226 #ifndef CONFIG_DEBUG_PAGEALLOC 2227 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) { 2228 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^ 2229 PAGE_OFFSET; 2230 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag | 2231 _PAGE_P_4V | _PAGE_W_4V); 2232 } else { 2233 kern_linear_pte_xor[1] = kern_linear_pte_xor[0]; 2234 } 2235 2236 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) { 2237 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^ 2238 PAGE_OFFSET; 2239 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag | 2240 _PAGE_P_4V | _PAGE_W_4V); 2241 } else { 2242 kern_linear_pte_xor[2] = kern_linear_pte_xor[1]; 2243 } 2244 2245 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) { 2246 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^ 2247 PAGE_OFFSET; 2248 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag | 2249 _PAGE_P_4V | _PAGE_W_4V); 2250 } else { 2251 kern_linear_pte_xor[3] = kern_linear_pte_xor[2]; 2252 } 2253 #endif 2254 } 2255 2256 /* paging_init() sets up the page tables */ 2257 2258 static unsigned long last_valid_pfn; 2259 2260 static void sun4u_pgprot_init(void); 2261 static void sun4v_pgprot_init(void); 2262 2263 #define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U) 2264 #define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V) 2265 #define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U) 2266 #define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V) 2267 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R) 2268 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R) 2269 2270 /* We need to exclude reserved regions. This exclusion will include 2271 * vmlinux and initrd. To be more precise the initrd size could be used to 2272 * compute a new lower limit because it is freed later during initialization. 2273 */ 2274 static void __init reduce_memory(phys_addr_t limit_ram) 2275 { 2276 limit_ram += memblock_reserved_size(); 2277 memblock_enforce_memory_limit(limit_ram); 2278 } 2279 2280 void __init arch_zone_limits_init(unsigned long *max_zone_pfns) 2281 { 2282 max_zone_pfns[ZONE_NORMAL] = last_valid_pfn; 2283 } 2284 2285 void __init paging_init(void) 2286 { 2287 unsigned long end_pfn, shift, phys_base; 2288 unsigned long real_end, i; 2289 2290 setup_page_offset(); 2291 2292 /* These build time checkes make sure that the dcache_dirty_cpu() 2293 * folio->flags usage will work. 2294 * 2295 * When a page gets marked as dcache-dirty, we store the 2296 * cpu number starting at bit 32 in the folio->flags. Also, 2297 * functions like clear_dcache_dirty_cpu use the cpu mask 2298 * in 13-bit signed-immediate instruction fields. 2299 */ 2300 2301 /* 2302 * Page flags must not reach into upper 32 bits that are used 2303 * for the cpu number 2304 */ 2305 BUILD_BUG_ON(NR_PAGEFLAGS > 32); 2306 2307 /* 2308 * The bit fields placed in the high range must not reach below 2309 * the 32 bit boundary. Otherwise we cannot place the cpu field 2310 * at the 32 bit boundary. 2311 */ 2312 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH + 2313 ilog2(roundup_pow_of_two(NR_CPUS)) > 32); 2314 2315 BUILD_BUG_ON(NR_CPUS > 4096); 2316 2317 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB; 2318 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE; 2319 2320 /* Invalidate both kernel TSBs. */ 2321 memset(swapper_tsb, 0x40, sizeof(swapper_tsb)); 2322 #ifndef CONFIG_DEBUG_PAGEALLOC 2323 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb)); 2324 #endif 2325 2326 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde 2327 * bit on M7 processor. This is a conflicting usage of the same 2328 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption 2329 * Detection error on all pages and this will lead to problems 2330 * later. Kernel does not run with MCD enabled and hence rest 2331 * of the required steps to fully configure memory corruption 2332 * detection are not taken. We need to ensure TTE.mcde is not 2333 * set on M7 processor. Compute the value of cacheability 2334 * flag for use later taking this into consideration. 2335 */ 2336 switch (sun4v_chip_type) { 2337 case SUN4V_CHIP_SPARC_M7: 2338 case SUN4V_CHIP_SPARC_M8: 2339 case SUN4V_CHIP_SPARC_SN: 2340 page_cache4v_flag = _PAGE_CP_4V; 2341 break; 2342 default: 2343 page_cache4v_flag = _PAGE_CACHE_4V; 2344 break; 2345 } 2346 2347 if (tlb_type == hypervisor) 2348 sun4v_pgprot_init(); 2349 else 2350 sun4u_pgprot_init(); 2351 2352 if (tlb_type == cheetah_plus || 2353 tlb_type == hypervisor) { 2354 tsb_phys_patch(); 2355 ktsb_phys_patch(); 2356 } 2357 2358 if (tlb_type == hypervisor) 2359 sun4v_patch_tlb_handlers(); 2360 2361 /* Find available physical memory... 2362 * 2363 * Read it twice in order to work around a bug in openfirmware. 2364 * The call to grab this table itself can cause openfirmware to 2365 * allocate memory, which in turn can take away some space from 2366 * the list of available memory. Reading it twice makes sure 2367 * we really do get the final value. 2368 */ 2369 read_obp_translations(); 2370 read_obp_memory("reg", &pall[0], &pall_ents); 2371 read_obp_memory("available", &pavail[0], &pavail_ents); 2372 read_obp_memory("available", &pavail[0], &pavail_ents); 2373 2374 phys_base = 0xffffffffffffffffUL; 2375 for (i = 0; i < pavail_ents; i++) { 2376 phys_base = min(phys_base, pavail[i].phys_addr); 2377 memblock_add(pavail[i].phys_addr, pavail[i].reg_size); 2378 } 2379 2380 memblock_reserve(kern_base, kern_size); 2381 2382 find_ramdisk(phys_base); 2383 2384 if (cmdline_memory_size) 2385 reduce_memory(cmdline_memory_size); 2386 2387 memblock_allow_resize(); 2388 memblock_dump_all(); 2389 2390 set_bit(0, mmu_context_bmap); 2391 2392 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE); 2393 2394 real_end = (unsigned long)_end; 2395 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB); 2396 printk("Kernel: Using %d locked TLB entries for main kernel image.\n", 2397 num_kernel_image_mappings); 2398 2399 /* Set kernel pgd to upper alias so physical page computations 2400 * work. 2401 */ 2402 init_mm.pgd += ((shift) / (sizeof(pgd_t))); 2403 2404 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir)); 2405 2406 inherit_prom_mappings(); 2407 2408 /* Ok, we can use our TLB miss and window trap handlers safely. */ 2409 setup_tba(); 2410 2411 __flush_tlb_all(); 2412 2413 prom_build_devicetree(); 2414 of_populate_present_mask(); 2415 #ifndef CONFIG_SMP 2416 of_fill_in_cpu_data(); 2417 #endif 2418 2419 if (tlb_type == hypervisor) { 2420 sun4v_mdesc_init(); 2421 mdesc_populate_present_mask(cpu_all_mask); 2422 #ifndef CONFIG_SMP 2423 mdesc_fill_in_cpu_data(cpu_all_mask); 2424 #endif 2425 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask); 2426 2427 sun4v_linear_pte_xor_finalize(); 2428 2429 sun4v_ktsb_init(); 2430 sun4v_ktsb_register(); 2431 } else { 2432 unsigned long impl, ver; 2433 2434 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K | 2435 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB); 2436 2437 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver)); 2438 impl = ((ver >> 32) & 0xffff); 2439 if (impl == PANTHER_IMPL) 2440 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB | 2441 HV_PGSZ_MASK_256MB); 2442 2443 sun4u_linear_pte_xor_finalize(); 2444 } 2445 2446 /* Flush the TLBs and the 4M TSB so that the updated linear 2447 * pte XOR settings are realized for all mappings. 2448 */ 2449 __flush_tlb_all(); 2450 #ifndef CONFIG_DEBUG_PAGEALLOC 2451 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb)); 2452 #endif 2453 __flush_tlb_all(); 2454 2455 /* Setup bootmem... */ 2456 last_valid_pfn = end_pfn = bootmem_init(phys_base); 2457 2458 kernel_physical_mapping_init(); 2459 2460 printk("Booting Linux...\n"); 2461 } 2462 2463 int page_in_phys_avail(unsigned long paddr) 2464 { 2465 int i; 2466 2467 paddr &= PAGE_MASK; 2468 2469 for (i = 0; i < pavail_ents; i++) { 2470 unsigned long start, end; 2471 2472 start = pavail[i].phys_addr; 2473 end = start + pavail[i].reg_size; 2474 2475 if (paddr >= start && paddr < end) 2476 return 1; 2477 } 2478 if (paddr >= kern_base && paddr < (kern_base + kern_size)) 2479 return 1; 2480 #ifdef CONFIG_BLK_DEV_INITRD 2481 if (paddr >= __pa(initrd_start) && 2482 paddr < __pa(PAGE_ALIGN(initrd_end))) 2483 return 1; 2484 #endif 2485 2486 return 0; 2487 } 2488 2489 static void __init register_page_bootmem_info(void) 2490 { 2491 #ifdef CONFIG_NUMA 2492 int i; 2493 2494 for_each_online_node(i) 2495 if (NODE_DATA(i)->node_spanned_pages) 2496 register_page_bootmem_info_node(NODE_DATA(i)); 2497 #endif 2498 } 2499 void __init mem_init(void) 2500 { 2501 /* 2502 * Must be done after boot memory is put on freelist, because here we 2503 * might set fields in deferred struct pages that have not yet been 2504 * initialized, and memblock_free_all() initializes all the reserved 2505 * deferred pages for us. 2506 */ 2507 register_page_bootmem_info(); 2508 2509 /* 2510 * Set up the zero page, mark it reserved, so that page count 2511 * is not manipulated when freeing the page from user ptes. 2512 */ 2513 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0); 2514 if (mem_map_zero == NULL) { 2515 prom_printf("paging_init: Cannot alloc zero page.\n"); 2516 prom_halt(); 2517 } 2518 mark_page_reserved(mem_map_zero); 2519 2520 2521 if (tlb_type == cheetah || tlb_type == cheetah_plus) 2522 cheetah_ecache_flush_init(); 2523 } 2524 2525 void free_initmem(void) 2526 { 2527 unsigned long addr, initend; 2528 int do_free = 1; 2529 2530 /* If the physical memory maps were trimmed by kernel command 2531 * line options, don't even try freeing this initmem stuff up. 2532 * The kernel image could have been in the trimmed out region 2533 * and if so the freeing below will free invalid page structs. 2534 */ 2535 if (cmdline_memory_size) 2536 do_free = 0; 2537 2538 /* 2539 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes. 2540 */ 2541 addr = PAGE_ALIGN((unsigned long)(__init_begin)); 2542 initend = (unsigned long)(__init_end) & PAGE_MASK; 2543 for (; addr < initend; addr += PAGE_SIZE) { 2544 unsigned long page; 2545 2546 page = (addr + 2547 ((unsigned long) __va(kern_base)) - 2548 ((unsigned long) KERNBASE)); 2549 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE); 2550 2551 if (do_free) 2552 free_reserved_page(virt_to_page(page)); 2553 } 2554 } 2555 2556 pgprot_t PAGE_KERNEL __read_mostly; 2557 EXPORT_SYMBOL(PAGE_KERNEL); 2558 2559 pgprot_t PAGE_KERNEL_LOCKED __read_mostly; 2560 pgprot_t PAGE_COPY __read_mostly; 2561 2562 pgprot_t PAGE_SHARED __read_mostly; 2563 EXPORT_SYMBOL(PAGE_SHARED); 2564 2565 unsigned long pg_iobits __read_mostly; 2566 2567 unsigned long _PAGE_IE __read_mostly; 2568 EXPORT_SYMBOL(_PAGE_IE); 2569 2570 unsigned long _PAGE_E __read_mostly; 2571 EXPORT_SYMBOL(_PAGE_E); 2572 2573 unsigned long _PAGE_CACHE __read_mostly; 2574 EXPORT_SYMBOL(_PAGE_CACHE); 2575 2576 #ifdef CONFIG_SPARSEMEM_VMEMMAP 2577 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend, 2578 int node, struct vmem_altmap *altmap) 2579 { 2580 unsigned long pte_base; 2581 2582 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U | 2583 _PAGE_CP_4U | _PAGE_CV_4U | 2584 _PAGE_P_4U | _PAGE_W_4U); 2585 if (tlb_type == hypervisor) 2586 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V | 2587 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V); 2588 2589 pte_base |= _PAGE_PMD_HUGE; 2590 2591 vstart = vstart & PMD_MASK; 2592 vend = ALIGN(vend, PMD_SIZE); 2593 for (; vstart < vend; vstart += PMD_SIZE) { 2594 pgd_t *pgd = vmemmap_pgd_populate(vstart, node); 2595 unsigned long pte; 2596 p4d_t *p4d; 2597 pud_t *pud; 2598 pmd_t *pmd; 2599 2600 if (!pgd) 2601 return -ENOMEM; 2602 2603 p4d = vmemmap_p4d_populate(pgd, vstart, node); 2604 if (!p4d) 2605 return -ENOMEM; 2606 2607 pud = vmemmap_pud_populate(p4d, vstart, node); 2608 if (!pud) 2609 return -ENOMEM; 2610 2611 pmd = pmd_offset(pud, vstart); 2612 pte = pmd_val(*pmd); 2613 if (!(pte & _PAGE_VALID)) { 2614 void *block = vmemmap_alloc_block(PMD_SIZE, node); 2615 2616 if (!block) 2617 return -ENOMEM; 2618 2619 pmd_val(*pmd) = pte_base | __pa(block); 2620 } 2621 } 2622 2623 return 0; 2624 } 2625 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 2626 2627 /* These are actually filled in at boot time by sun4{u,v}_pgprot_init() */ 2628 static pgprot_t protection_map[16] __ro_after_init; 2629 2630 static void prot_init_common(unsigned long page_none, 2631 unsigned long page_shared, 2632 unsigned long page_copy, 2633 unsigned long page_readonly, 2634 unsigned long page_exec_bit) 2635 { 2636 PAGE_COPY = __pgprot(page_copy); 2637 PAGE_SHARED = __pgprot(page_shared); 2638 2639 protection_map[0x0] = __pgprot(page_none); 2640 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit); 2641 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit); 2642 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit); 2643 protection_map[0x4] = __pgprot(page_readonly); 2644 protection_map[0x5] = __pgprot(page_readonly); 2645 protection_map[0x6] = __pgprot(page_copy); 2646 protection_map[0x7] = __pgprot(page_copy); 2647 protection_map[0x8] = __pgprot(page_none); 2648 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit); 2649 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit); 2650 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit); 2651 protection_map[0xc] = __pgprot(page_readonly); 2652 protection_map[0xd] = __pgprot(page_readonly); 2653 protection_map[0xe] = __pgprot(page_shared); 2654 protection_map[0xf] = __pgprot(page_shared); 2655 } 2656 2657 static void __init sun4u_pgprot_init(void) 2658 { 2659 unsigned long page_none, page_shared, page_copy, page_readonly; 2660 unsigned long page_exec_bit; 2661 int i; 2662 2663 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID | 2664 _PAGE_CACHE_4U | _PAGE_P_4U | 2665 __ACCESS_BITS_4U | __DIRTY_BITS_4U | 2666 _PAGE_EXEC_4U); 2667 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID | 2668 _PAGE_CACHE_4U | _PAGE_P_4U | 2669 __ACCESS_BITS_4U | __DIRTY_BITS_4U | 2670 _PAGE_EXEC_4U | _PAGE_L_4U); 2671 2672 _PAGE_IE = _PAGE_IE_4U; 2673 _PAGE_E = _PAGE_E_4U; 2674 _PAGE_CACHE = _PAGE_CACHE_4U; 2675 2676 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U | 2677 __ACCESS_BITS_4U | _PAGE_E_4U); 2678 2679 #ifdef CONFIG_DEBUG_PAGEALLOC 2680 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET; 2681 #else 2682 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^ 2683 PAGE_OFFSET; 2684 #endif 2685 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U | 2686 _PAGE_P_4U | _PAGE_W_4U); 2687 2688 for (i = 1; i < 4; i++) 2689 kern_linear_pte_xor[i] = kern_linear_pte_xor[0]; 2690 2691 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U | 2692 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U | 2693 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U); 2694 2695 2696 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U; 2697 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2698 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U); 2699 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2700 __ACCESS_BITS_4U | _PAGE_EXEC_4U); 2701 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2702 __ACCESS_BITS_4U | _PAGE_EXEC_4U); 2703 2704 page_exec_bit = _PAGE_EXEC_4U; 2705 2706 prot_init_common(page_none, page_shared, page_copy, page_readonly, 2707 page_exec_bit); 2708 } 2709 2710 static void __init sun4v_pgprot_init(void) 2711 { 2712 unsigned long page_none, page_shared, page_copy, page_readonly; 2713 unsigned long page_exec_bit; 2714 int i; 2715 2716 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID | 2717 page_cache4v_flag | _PAGE_P_4V | 2718 __ACCESS_BITS_4V | __DIRTY_BITS_4V | 2719 _PAGE_EXEC_4V); 2720 PAGE_KERNEL_LOCKED = PAGE_KERNEL; 2721 2722 _PAGE_IE = _PAGE_IE_4V; 2723 _PAGE_E = _PAGE_E_4V; 2724 _PAGE_CACHE = page_cache4v_flag; 2725 2726 #ifdef CONFIG_DEBUG_PAGEALLOC 2727 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET; 2728 #else 2729 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^ 2730 PAGE_OFFSET; 2731 #endif 2732 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V | 2733 _PAGE_W_4V); 2734 2735 for (i = 1; i < 4; i++) 2736 kern_linear_pte_xor[i] = kern_linear_pte_xor[0]; 2737 2738 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V | 2739 __ACCESS_BITS_4V | _PAGE_E_4V); 2740 2741 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V | 2742 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V | 2743 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V | 2744 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V); 2745 2746 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag; 2747 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2748 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V); 2749 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2750 __ACCESS_BITS_4V | _PAGE_EXEC_4V); 2751 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2752 __ACCESS_BITS_4V | _PAGE_EXEC_4V); 2753 2754 page_exec_bit = _PAGE_EXEC_4V; 2755 2756 prot_init_common(page_none, page_shared, page_copy, page_readonly, 2757 page_exec_bit); 2758 } 2759 2760 unsigned long pte_sz_bits(unsigned long sz) 2761 { 2762 if (tlb_type == hypervisor) { 2763 switch (sz) { 2764 case 8 * 1024: 2765 default: 2766 return _PAGE_SZ8K_4V; 2767 case 64 * 1024: 2768 return _PAGE_SZ64K_4V; 2769 case 512 * 1024: 2770 return _PAGE_SZ512K_4V; 2771 case 4 * 1024 * 1024: 2772 return _PAGE_SZ4MB_4V; 2773 } 2774 } else { 2775 switch (sz) { 2776 case 8 * 1024: 2777 default: 2778 return _PAGE_SZ8K_4U; 2779 case 64 * 1024: 2780 return _PAGE_SZ64K_4U; 2781 case 512 * 1024: 2782 return _PAGE_SZ512K_4U; 2783 case 4 * 1024 * 1024: 2784 return _PAGE_SZ4MB_4U; 2785 } 2786 } 2787 } 2788 2789 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size) 2790 { 2791 pte_t pte; 2792 2793 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot)); 2794 pte_val(pte) |= (((unsigned long)space) << 32); 2795 pte_val(pte) |= pte_sz_bits(page_size); 2796 2797 return pte; 2798 } 2799 2800 static unsigned long kern_large_tte(unsigned long paddr) 2801 { 2802 unsigned long val; 2803 2804 val = (_PAGE_VALID | _PAGE_SZ4MB_4U | 2805 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U | 2806 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U); 2807 if (tlb_type == hypervisor) 2808 val = (_PAGE_VALID | _PAGE_SZ4MB_4V | 2809 page_cache4v_flag | _PAGE_P_4V | 2810 _PAGE_EXEC_4V | _PAGE_W_4V); 2811 2812 return val | paddr; 2813 } 2814 2815 /* If not locked, zap it. */ 2816 void __flush_tlb_all(void) 2817 { 2818 unsigned long pstate; 2819 int i; 2820 2821 __asm__ __volatile__("flushw\n\t" 2822 "rdpr %%pstate, %0\n\t" 2823 "wrpr %0, %1, %%pstate" 2824 : "=r" (pstate) 2825 : "i" (PSTATE_IE)); 2826 if (tlb_type == hypervisor) { 2827 sun4v_mmu_demap_all(); 2828 } else if (tlb_type == spitfire) { 2829 for (i = 0; i < 64; i++) { 2830 /* Spitfire Errata #32 workaround */ 2831 /* NOTE: Always runs on spitfire, so no 2832 * cheetah+ page size encodings. 2833 */ 2834 __asm__ __volatile__("stxa %0, [%1] %2\n\t" 2835 "flush %%g6" 2836 : /* No outputs */ 2837 : "r" (0), 2838 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU)); 2839 2840 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) { 2841 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 2842 "membar #Sync" 2843 : /* no outputs */ 2844 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU)); 2845 spitfire_put_dtlb_data(i, 0x0UL); 2846 } 2847 2848 /* Spitfire Errata #32 workaround */ 2849 /* NOTE: Always runs on spitfire, so no 2850 * cheetah+ page size encodings. 2851 */ 2852 __asm__ __volatile__("stxa %0, [%1] %2\n\t" 2853 "flush %%g6" 2854 : /* No outputs */ 2855 : "r" (0), 2856 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU)); 2857 2858 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) { 2859 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 2860 "membar #Sync" 2861 : /* no outputs */ 2862 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU)); 2863 spitfire_put_itlb_data(i, 0x0UL); 2864 } 2865 } 2866 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) { 2867 cheetah_flush_dtlb_all(); 2868 cheetah_flush_itlb_all(); 2869 } 2870 __asm__ __volatile__("wrpr %0, 0, %%pstate" 2871 : : "r" (pstate)); 2872 } 2873 2874 static pte_t *__pte_alloc_one(struct mm_struct *mm) 2875 { 2876 struct ptdesc *ptdesc = pagetable_alloc(GFP_KERNEL | __GFP_ZERO, 0); 2877 2878 if (!ptdesc) 2879 return NULL; 2880 if (!pagetable_pte_ctor(mm, ptdesc)) { 2881 pagetable_free(ptdesc); 2882 return NULL; 2883 } 2884 return ptdesc_address(ptdesc); 2885 } 2886 2887 pte_t *pte_alloc_one_kernel(struct mm_struct *mm) 2888 { 2889 return __pte_alloc_one(mm); 2890 } 2891 2892 pgtable_t pte_alloc_one(struct mm_struct *mm) 2893 { 2894 return __pte_alloc_one(mm); 2895 } 2896 2897 static void __pte_free(pgtable_t pte) 2898 { 2899 struct ptdesc *ptdesc = virt_to_ptdesc(pte); 2900 2901 pagetable_dtor(ptdesc); 2902 pagetable_free(ptdesc); 2903 } 2904 2905 void pte_free_kernel(struct mm_struct *mm, pte_t *pte) 2906 { 2907 __pte_free(pte); 2908 } 2909 2910 void pte_free(struct mm_struct *mm, pgtable_t pte) 2911 { 2912 __pte_free(pte); 2913 } 2914 2915 void pgtable_free(void *table, bool is_page) 2916 { 2917 if (is_page) 2918 __pte_free(table); 2919 else 2920 kmem_cache_free(pgtable_cache, table); 2921 } 2922 2923 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 2924 static void pte_free_now(struct rcu_head *head) 2925 { 2926 struct page *page; 2927 2928 page = container_of(head, struct page, rcu_head); 2929 __pte_free((pgtable_t)page_address(page)); 2930 } 2931 2932 void pte_free_defer(struct mm_struct *mm, pgtable_t pgtable) 2933 { 2934 struct page *page; 2935 2936 page = virt_to_page(pgtable); 2937 call_rcu(&page->rcu_head, pte_free_now); 2938 } 2939 2940 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr, 2941 pmd_t *pmd) 2942 { 2943 unsigned long pte, flags; 2944 struct mm_struct *mm; 2945 pmd_t entry = *pmd; 2946 2947 if (!pmd_leaf(entry) || !pmd_young(entry)) 2948 return; 2949 2950 pte = pmd_val(entry); 2951 2952 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */ 2953 if (!(pte & _PAGE_VALID)) 2954 return; 2955 2956 /* We are fabricating 8MB pages using 4MB real hw pages. */ 2957 pte |= (addr & (1UL << REAL_HPAGE_SHIFT)); 2958 2959 mm = vma->vm_mm; 2960 2961 spin_lock_irqsave(&mm->context.lock, flags); 2962 2963 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) 2964 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT, 2965 addr, pte); 2966 2967 spin_unlock_irqrestore(&mm->context.lock, flags); 2968 } 2969 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 2970 2971 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 2972 static void context_reload(void *__data) 2973 { 2974 struct mm_struct *mm = __data; 2975 2976 if (mm == current->mm) 2977 load_secondary_context(mm); 2978 } 2979 2980 void hugetlb_setup(struct pt_regs *regs) 2981 { 2982 struct mm_struct *mm = current->mm; 2983 struct tsb_config *tp; 2984 2985 if (faulthandler_disabled() || !mm) { 2986 const struct exception_table_entry *entry; 2987 2988 entry = search_exception_tables(regs->tpc); 2989 if (entry) { 2990 regs->tpc = entry->fixup; 2991 regs->tnpc = regs->tpc + 4; 2992 return; 2993 } 2994 pr_alert("Unexpected HugeTLB setup in atomic context.\n"); 2995 die_if_kernel("HugeTSB in atomic", regs); 2996 } 2997 2998 tp = &mm->context.tsb_block[MM_TSB_HUGE]; 2999 if (likely(tp->tsb == NULL)) 3000 tsb_grow(mm, MM_TSB_HUGE, 0); 3001 3002 tsb_context_switch(mm); 3003 smp_tsb_sync(mm); 3004 3005 /* On UltraSPARC-III+ and later, configure the second half of 3006 * the Data-TLB for huge pages. 3007 */ 3008 if (tlb_type == cheetah_plus) { 3009 bool need_context_reload = false; 3010 unsigned long ctx; 3011 3012 spin_lock_irq(&ctx_alloc_lock); 3013 ctx = mm->context.sparc64_ctx_val; 3014 ctx &= ~CTX_PGSZ_MASK; 3015 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT; 3016 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT; 3017 3018 if (ctx != mm->context.sparc64_ctx_val) { 3019 /* When changing the page size fields, we 3020 * must perform a context flush so that no 3021 * stale entries match. This flush must 3022 * occur with the original context register 3023 * settings. 3024 */ 3025 do_flush_tlb_mm(mm); 3026 3027 /* Reload the context register of all processors 3028 * also executing in this address space. 3029 */ 3030 mm->context.sparc64_ctx_val = ctx; 3031 need_context_reload = true; 3032 } 3033 spin_unlock_irq(&ctx_alloc_lock); 3034 3035 if (need_context_reload) 3036 on_each_cpu(context_reload, mm, 0); 3037 } 3038 } 3039 #endif 3040 3041 static struct resource code_resource = { 3042 .name = "Kernel code", 3043 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3044 }; 3045 3046 static struct resource data_resource = { 3047 .name = "Kernel data", 3048 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3049 }; 3050 3051 static struct resource bss_resource = { 3052 .name = "Kernel bss", 3053 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3054 }; 3055 3056 static inline resource_size_t compute_kern_paddr(void *addr) 3057 { 3058 return (resource_size_t) (addr - KERNBASE + kern_base); 3059 } 3060 3061 static void __init kernel_lds_init(void) 3062 { 3063 code_resource.start = compute_kern_paddr(_text); 3064 code_resource.end = compute_kern_paddr(_etext - 1); 3065 data_resource.start = compute_kern_paddr(_etext); 3066 data_resource.end = compute_kern_paddr(_edata - 1); 3067 bss_resource.start = compute_kern_paddr(__bss_start); 3068 bss_resource.end = compute_kern_paddr(_end - 1); 3069 } 3070 3071 static int __init report_memory(void) 3072 { 3073 int i; 3074 struct resource *res; 3075 3076 kernel_lds_init(); 3077 3078 for (i = 0; i < pavail_ents; i++) { 3079 res = kzalloc(sizeof(struct resource), GFP_KERNEL); 3080 3081 if (!res) { 3082 pr_warn("Failed to allocate source.\n"); 3083 break; 3084 } 3085 3086 res->name = "System RAM"; 3087 res->start = pavail[i].phys_addr; 3088 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1; 3089 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM; 3090 3091 if (insert_resource(&iomem_resource, res) < 0) { 3092 pr_warn("Resource insertion failed.\n"); 3093 break; 3094 } 3095 3096 insert_resource(res, &code_resource); 3097 insert_resource(res, &data_resource); 3098 insert_resource(res, &bss_resource); 3099 } 3100 3101 return 0; 3102 } 3103 arch_initcall(report_memory); 3104 3105 #ifdef CONFIG_SMP 3106 #define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range 3107 #else 3108 #define do_flush_tlb_kernel_range __flush_tlb_kernel_range 3109 #endif 3110 3111 void flush_tlb_kernel_range(unsigned long start, unsigned long end) 3112 { 3113 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) { 3114 if (start < LOW_OBP_ADDRESS) { 3115 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS); 3116 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS); 3117 } 3118 if (end > HI_OBP_ADDRESS) { 3119 flush_tsb_kernel_range(HI_OBP_ADDRESS, end); 3120 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end); 3121 } 3122 } else { 3123 flush_tsb_kernel_range(start, end); 3124 do_flush_tlb_kernel_range(start, end); 3125 } 3126 } 3127 3128 void copy_user_highpage(struct page *to, struct page *from, 3129 unsigned long vaddr, struct vm_area_struct *vma) 3130 { 3131 char *vfrom, *vto; 3132 3133 vfrom = kmap_atomic(from); 3134 vto = kmap_atomic(to); 3135 copy_user_page(vto, vfrom, vaddr, to); 3136 kunmap_atomic(vto); 3137 kunmap_atomic(vfrom); 3138 3139 /* If this page has ADI enabled, copy over any ADI tags 3140 * as well 3141 */ 3142 if (vma->vm_flags & VM_SPARC_ADI) { 3143 unsigned long pfrom, pto, i, adi_tag; 3144 3145 pfrom = page_to_phys(from); 3146 pto = page_to_phys(to); 3147 3148 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) { 3149 asm volatile("ldxa [%1] %2, %0\n\t" 3150 : "=r" (adi_tag) 3151 : "r" (i), "i" (ASI_MCD_REAL)); 3152 asm volatile("stxa %0, [%1] %2\n\t" 3153 : 3154 : "r" (adi_tag), "r" (pto), 3155 "i" (ASI_MCD_REAL)); 3156 pto += adi_blksize(); 3157 } 3158 asm volatile("membar #Sync\n\t"); 3159 } 3160 } 3161 EXPORT_SYMBOL(copy_user_highpage); 3162 3163 void copy_highpage(struct page *to, struct page *from) 3164 { 3165 char *vfrom, *vto; 3166 3167 vfrom = kmap_atomic(from); 3168 vto = kmap_atomic(to); 3169 copy_page(vto, vfrom); 3170 kunmap_atomic(vto); 3171 kunmap_atomic(vfrom); 3172 3173 /* If this platform is ADI enabled, copy any ADI tags 3174 * as well 3175 */ 3176 if (adi_capable()) { 3177 unsigned long pfrom, pto, i, adi_tag; 3178 3179 pfrom = page_to_phys(from); 3180 pto = page_to_phys(to); 3181 3182 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) { 3183 asm volatile("ldxa [%1] %2, %0\n\t" 3184 : "=r" (adi_tag) 3185 : "r" (i), "i" (ASI_MCD_REAL)); 3186 asm volatile("stxa %0, [%1] %2\n\t" 3187 : 3188 : "r" (adi_tag), "r" (pto), 3189 "i" (ASI_MCD_REAL)); 3190 pto += adi_blksize(); 3191 } 3192 asm volatile("membar #Sync\n\t"); 3193 } 3194 } 3195 EXPORT_SYMBOL(copy_highpage); 3196 3197 pgprot_t vm_get_page_prot(vm_flags_t vm_flags) 3198 { 3199 unsigned long prot = pgprot_val(protection_map[vm_flags & 3200 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]); 3201 3202 if (vm_flags & VM_SPARC_ADI) 3203 prot |= _PAGE_MCD_4V; 3204 3205 return __pgprot(prot); 3206 } 3207 EXPORT_SYMBOL(vm_get_page_prot); 3208