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 31 #include <asm/head.h> 32 #include <asm/page.h> 33 #include <asm/pgalloc.h> 34 #include <asm/pgtable.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_page_impl(struct page *page) 199 { 200 BUG_ON(tlb_type == hypervisor); 201 #ifdef CONFIG_DEBUG_DCFLUSH 202 atomic_inc(&dcpage_flushes); 203 #endif 204 205 #ifdef DCACHE_ALIASING_POSSIBLE 206 __flush_dcache_page(page_address(page), 207 ((tlb_type == spitfire) && 208 page_mapping_file(page) != NULL)); 209 #else 210 if (page_mapping_file(page) != NULL && 211 tlb_type == spitfire) 212 __flush_icache_page(__pa(page_address(page))); 213 #endif 214 } 215 216 #define PG_dcache_dirty PG_arch_1 217 #define PG_dcache_cpu_shift 32UL 218 #define PG_dcache_cpu_mask \ 219 ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL) 220 221 #define dcache_dirty_cpu(page) \ 222 (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask) 223 224 static inline void set_dcache_dirty(struct page *page, int this_cpu) 225 { 226 unsigned long mask = this_cpu; 227 unsigned long non_cpu_bits; 228 229 non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift); 230 mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty); 231 232 __asm__ __volatile__("1:\n\t" 233 "ldx [%2], %%g7\n\t" 234 "and %%g7, %1, %%g1\n\t" 235 "or %%g1, %0, %%g1\n\t" 236 "casx [%2], %%g7, %%g1\n\t" 237 "cmp %%g7, %%g1\n\t" 238 "bne,pn %%xcc, 1b\n\t" 239 " nop" 240 : /* no outputs */ 241 : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags) 242 : "g1", "g7"); 243 } 244 245 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu) 246 { 247 unsigned long mask = (1UL << PG_dcache_dirty); 248 249 __asm__ __volatile__("! test_and_clear_dcache_dirty\n" 250 "1:\n\t" 251 "ldx [%2], %%g7\n\t" 252 "srlx %%g7, %4, %%g1\n\t" 253 "and %%g1, %3, %%g1\n\t" 254 "cmp %%g1, %0\n\t" 255 "bne,pn %%icc, 2f\n\t" 256 " andn %%g7, %1, %%g1\n\t" 257 "casx [%2], %%g7, %%g1\n\t" 258 "cmp %%g7, %%g1\n\t" 259 "bne,pn %%xcc, 1b\n\t" 260 " nop\n" 261 "2:" 262 : /* no outputs */ 263 : "r" (cpu), "r" (mask), "r" (&page->flags), 264 "i" (PG_dcache_cpu_mask), 265 "i" (PG_dcache_cpu_shift) 266 : "g1", "g7"); 267 } 268 269 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte) 270 { 271 unsigned long tsb_addr = (unsigned long) ent; 272 273 if (tlb_type == cheetah_plus || tlb_type == hypervisor) 274 tsb_addr = __pa(tsb_addr); 275 276 __tsb_insert(tsb_addr, tag, pte); 277 } 278 279 unsigned long _PAGE_ALL_SZ_BITS __read_mostly; 280 281 static void flush_dcache(unsigned long pfn) 282 { 283 struct page *page; 284 285 page = pfn_to_page(pfn); 286 if (page) { 287 unsigned long pg_flags; 288 289 pg_flags = page->flags; 290 if (pg_flags & (1UL << PG_dcache_dirty)) { 291 int cpu = ((pg_flags >> PG_dcache_cpu_shift) & 292 PG_dcache_cpu_mask); 293 int this_cpu = get_cpu(); 294 295 /* This is just to optimize away some function calls 296 * in the SMP case. 297 */ 298 if (cpu == this_cpu) 299 flush_dcache_page_impl(page); 300 else 301 smp_flush_dcache_page_impl(page, cpu); 302 303 clear_dcache_dirty_cpu(page, cpu); 304 305 put_cpu(); 306 } 307 } 308 } 309 310 /* mm->context.lock must be held */ 311 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index, 312 unsigned long tsb_hash_shift, unsigned long address, 313 unsigned long tte) 314 { 315 struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb; 316 unsigned long tag; 317 318 if (unlikely(!tsb)) 319 return; 320 321 tsb += ((address >> tsb_hash_shift) & 322 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL)); 323 tag = (address >> 22UL); 324 tsb_insert(tsb, tag, tte); 325 } 326 327 #ifdef CONFIG_HUGETLB_PAGE 328 static void __init add_huge_page_size(unsigned long size) 329 { 330 unsigned int order; 331 332 if (size_to_hstate(size)) 333 return; 334 335 order = ilog2(size) - PAGE_SHIFT; 336 hugetlb_add_hstate(order); 337 } 338 339 static int __init hugetlbpage_init(void) 340 { 341 add_huge_page_size(1UL << HPAGE_64K_SHIFT); 342 add_huge_page_size(1UL << HPAGE_SHIFT); 343 add_huge_page_size(1UL << HPAGE_256MB_SHIFT); 344 add_huge_page_size(1UL << HPAGE_2GB_SHIFT); 345 346 return 0; 347 } 348 349 arch_initcall(hugetlbpage_init); 350 351 static void __init pud_huge_patch(void) 352 { 353 struct pud_huge_patch_entry *p; 354 unsigned long addr; 355 356 p = &__pud_huge_patch; 357 addr = p->addr; 358 *(unsigned int *)addr = p->insn; 359 360 __asm__ __volatile__("flush %0" : : "r" (addr)); 361 } 362 363 static int __init setup_hugepagesz(char *string) 364 { 365 unsigned long long hugepage_size; 366 unsigned int hugepage_shift; 367 unsigned short hv_pgsz_idx; 368 unsigned int hv_pgsz_mask; 369 int rc = 0; 370 371 hugepage_size = memparse(string, &string); 372 hugepage_shift = ilog2(hugepage_size); 373 374 switch (hugepage_shift) { 375 case HPAGE_16GB_SHIFT: 376 hv_pgsz_mask = HV_PGSZ_MASK_16GB; 377 hv_pgsz_idx = HV_PGSZ_IDX_16GB; 378 pud_huge_patch(); 379 break; 380 case HPAGE_2GB_SHIFT: 381 hv_pgsz_mask = HV_PGSZ_MASK_2GB; 382 hv_pgsz_idx = HV_PGSZ_IDX_2GB; 383 break; 384 case HPAGE_256MB_SHIFT: 385 hv_pgsz_mask = HV_PGSZ_MASK_256MB; 386 hv_pgsz_idx = HV_PGSZ_IDX_256MB; 387 break; 388 case HPAGE_SHIFT: 389 hv_pgsz_mask = HV_PGSZ_MASK_4MB; 390 hv_pgsz_idx = HV_PGSZ_IDX_4MB; 391 break; 392 case HPAGE_64K_SHIFT: 393 hv_pgsz_mask = HV_PGSZ_MASK_64K; 394 hv_pgsz_idx = HV_PGSZ_IDX_64K; 395 break; 396 default: 397 hv_pgsz_mask = 0; 398 } 399 400 if ((hv_pgsz_mask & cpu_pgsz_mask) == 0U) { 401 hugetlb_bad_size(); 402 pr_err("hugepagesz=%llu not supported by MMU.\n", 403 hugepage_size); 404 goto out; 405 } 406 407 add_huge_page_size(hugepage_size); 408 rc = 1; 409 410 out: 411 return rc; 412 } 413 __setup("hugepagesz=", setup_hugepagesz); 414 #endif /* CONFIG_HUGETLB_PAGE */ 415 416 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep) 417 { 418 struct mm_struct *mm; 419 unsigned long flags; 420 bool is_huge_tsb; 421 pte_t pte = *ptep; 422 423 if (tlb_type != hypervisor) { 424 unsigned long pfn = pte_pfn(pte); 425 426 if (pfn_valid(pfn)) 427 flush_dcache(pfn); 428 } 429 430 mm = vma->vm_mm; 431 432 /* Don't insert a non-valid PTE into the TSB, we'll deadlock. */ 433 if (!pte_accessible(mm, pte)) 434 return; 435 436 spin_lock_irqsave(&mm->context.lock, flags); 437 438 is_huge_tsb = false; 439 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 440 if (mm->context.hugetlb_pte_count || mm->context.thp_pte_count) { 441 unsigned long hugepage_size = PAGE_SIZE; 442 443 if (is_vm_hugetlb_page(vma)) 444 hugepage_size = huge_page_size(hstate_vma(vma)); 445 446 if (hugepage_size >= PUD_SIZE) { 447 unsigned long mask = 0x1ffc00000UL; 448 449 /* Transfer bits [32:22] from address to resolve 450 * at 4M granularity. 451 */ 452 pte_val(pte) &= ~mask; 453 pte_val(pte) |= (address & mask); 454 } else if (hugepage_size >= PMD_SIZE) { 455 /* We are fabricating 8MB pages using 4MB 456 * real hw pages. 457 */ 458 pte_val(pte) |= (address & (1UL << REAL_HPAGE_SHIFT)); 459 } 460 461 if (hugepage_size >= PMD_SIZE) { 462 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, 463 REAL_HPAGE_SHIFT, address, pte_val(pte)); 464 is_huge_tsb = true; 465 } 466 } 467 #endif 468 if (!is_huge_tsb) 469 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT, 470 address, pte_val(pte)); 471 472 spin_unlock_irqrestore(&mm->context.lock, flags); 473 } 474 475 void flush_dcache_page(struct page *page) 476 { 477 struct address_space *mapping; 478 int this_cpu; 479 480 if (tlb_type == hypervisor) 481 return; 482 483 /* Do not bother with the expensive D-cache flush if it 484 * is merely the zero page. The 'bigcore' testcase in GDB 485 * causes this case to run millions of times. 486 */ 487 if (page == ZERO_PAGE(0)) 488 return; 489 490 this_cpu = get_cpu(); 491 492 mapping = page_mapping_file(page); 493 if (mapping && !mapping_mapped(mapping)) { 494 int dirty = test_bit(PG_dcache_dirty, &page->flags); 495 if (dirty) { 496 int dirty_cpu = dcache_dirty_cpu(page); 497 498 if (dirty_cpu == this_cpu) 499 goto out; 500 smp_flush_dcache_page_impl(page, dirty_cpu); 501 } 502 set_dcache_dirty(page, this_cpu); 503 } else { 504 /* We could delay the flush for the !page_mapping 505 * case too. But that case is for exec env/arg 506 * pages and those are %99 certainly going to get 507 * faulted into the tlb (and thus flushed) anyways. 508 */ 509 flush_dcache_page_impl(page); 510 } 511 512 out: 513 put_cpu(); 514 } 515 EXPORT_SYMBOL(flush_dcache_page); 516 517 void __kprobes flush_icache_range(unsigned long start, unsigned long end) 518 { 519 /* Cheetah and Hypervisor platform cpus have coherent I-cache. */ 520 if (tlb_type == spitfire) { 521 unsigned long kaddr; 522 523 /* This code only runs on Spitfire cpus so this is 524 * why we can assume _PAGE_PADDR_4U. 525 */ 526 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) { 527 unsigned long paddr, mask = _PAGE_PADDR_4U; 528 529 if (kaddr >= PAGE_OFFSET) 530 paddr = kaddr & mask; 531 else { 532 pgd_t *pgdp = pgd_offset_k(kaddr); 533 pud_t *pudp = pud_offset(pgdp, kaddr); 534 pmd_t *pmdp = pmd_offset(pudp, kaddr); 535 pte_t *ptep = pte_offset_kernel(pmdp, kaddr); 536 537 paddr = pte_val(*ptep) & mask; 538 } 539 __flush_icache_page(paddr); 540 } 541 } 542 } 543 EXPORT_SYMBOL(flush_icache_range); 544 545 void mmu_info(struct seq_file *m) 546 { 547 static const char *pgsz_strings[] = { 548 "8K", "64K", "512K", "4MB", "32MB", 549 "256MB", "2GB", "16GB", 550 }; 551 int i, printed; 552 553 if (tlb_type == cheetah) 554 seq_printf(m, "MMU Type\t: Cheetah\n"); 555 else if (tlb_type == cheetah_plus) 556 seq_printf(m, "MMU Type\t: Cheetah+\n"); 557 else if (tlb_type == spitfire) 558 seq_printf(m, "MMU Type\t: Spitfire\n"); 559 else if (tlb_type == hypervisor) 560 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n"); 561 else 562 seq_printf(m, "MMU Type\t: ???\n"); 563 564 seq_printf(m, "MMU PGSZs\t: "); 565 printed = 0; 566 for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) { 567 if (cpu_pgsz_mask & (1UL << i)) { 568 seq_printf(m, "%s%s", 569 printed ? "," : "", pgsz_strings[i]); 570 printed++; 571 } 572 } 573 seq_putc(m, '\n'); 574 575 #ifdef CONFIG_DEBUG_DCFLUSH 576 seq_printf(m, "DCPageFlushes\t: %d\n", 577 atomic_read(&dcpage_flushes)); 578 #ifdef CONFIG_SMP 579 seq_printf(m, "DCPageFlushesXC\t: %d\n", 580 atomic_read(&dcpage_flushes_xcall)); 581 #endif /* CONFIG_SMP */ 582 #endif /* CONFIG_DEBUG_DCFLUSH */ 583 } 584 585 struct linux_prom_translation prom_trans[512] __read_mostly; 586 unsigned int prom_trans_ents __read_mostly; 587 588 unsigned long kern_locked_tte_data; 589 590 /* The obp translations are saved based on 8k pagesize, since obp can 591 * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS -> 592 * HI_OBP_ADDRESS range are handled in ktlb.S. 593 */ 594 static inline int in_obp_range(unsigned long vaddr) 595 { 596 return (vaddr >= LOW_OBP_ADDRESS && 597 vaddr < HI_OBP_ADDRESS); 598 } 599 600 static int cmp_ptrans(const void *a, const void *b) 601 { 602 const struct linux_prom_translation *x = a, *y = b; 603 604 if (x->virt > y->virt) 605 return 1; 606 if (x->virt < y->virt) 607 return -1; 608 return 0; 609 } 610 611 /* Read OBP translations property into 'prom_trans[]'. */ 612 static void __init read_obp_translations(void) 613 { 614 int n, node, ents, first, last, i; 615 616 node = prom_finddevice("/virtual-memory"); 617 n = prom_getproplen(node, "translations"); 618 if (unlikely(n == 0 || n == -1)) { 619 prom_printf("prom_mappings: Couldn't get size.\n"); 620 prom_halt(); 621 } 622 if (unlikely(n > sizeof(prom_trans))) { 623 prom_printf("prom_mappings: Size %d is too big.\n", n); 624 prom_halt(); 625 } 626 627 if ((n = prom_getproperty(node, "translations", 628 (char *)&prom_trans[0], 629 sizeof(prom_trans))) == -1) { 630 prom_printf("prom_mappings: Couldn't get property.\n"); 631 prom_halt(); 632 } 633 634 n = n / sizeof(struct linux_prom_translation); 635 636 ents = n; 637 638 sort(prom_trans, ents, sizeof(struct linux_prom_translation), 639 cmp_ptrans, NULL); 640 641 /* Now kick out all the non-OBP entries. */ 642 for (i = 0; i < ents; i++) { 643 if (in_obp_range(prom_trans[i].virt)) 644 break; 645 } 646 first = i; 647 for (; i < ents; i++) { 648 if (!in_obp_range(prom_trans[i].virt)) 649 break; 650 } 651 last = i; 652 653 for (i = 0; i < (last - first); i++) { 654 struct linux_prom_translation *src = &prom_trans[i + first]; 655 struct linux_prom_translation *dest = &prom_trans[i]; 656 657 *dest = *src; 658 } 659 for (; i < ents; i++) { 660 struct linux_prom_translation *dest = &prom_trans[i]; 661 dest->virt = dest->size = dest->data = 0x0UL; 662 } 663 664 prom_trans_ents = last - first; 665 666 if (tlb_type == spitfire) { 667 /* Clear diag TTE bits. */ 668 for (i = 0; i < prom_trans_ents; i++) 669 prom_trans[i].data &= ~0x0003fe0000000000UL; 670 } 671 672 /* Force execute bit on. */ 673 for (i = 0; i < prom_trans_ents; i++) 674 prom_trans[i].data |= (tlb_type == hypervisor ? 675 _PAGE_EXEC_4V : _PAGE_EXEC_4U); 676 } 677 678 static void __init hypervisor_tlb_lock(unsigned long vaddr, 679 unsigned long pte, 680 unsigned long mmu) 681 { 682 unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu); 683 684 if (ret != 0) { 685 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: " 686 "errors with %lx\n", vaddr, 0, pte, mmu, ret); 687 prom_halt(); 688 } 689 } 690 691 static unsigned long kern_large_tte(unsigned long paddr); 692 693 static void __init remap_kernel(void) 694 { 695 unsigned long phys_page, tte_vaddr, tte_data; 696 int i, tlb_ent = sparc64_highest_locked_tlbent(); 697 698 tte_vaddr = (unsigned long) KERNBASE; 699 phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB; 700 tte_data = kern_large_tte(phys_page); 701 702 kern_locked_tte_data = tte_data; 703 704 /* Now lock us into the TLBs via Hypervisor or OBP. */ 705 if (tlb_type == hypervisor) { 706 for (i = 0; i < num_kernel_image_mappings; i++) { 707 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU); 708 hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU); 709 tte_vaddr += 0x400000; 710 tte_data += 0x400000; 711 } 712 } else { 713 for (i = 0; i < num_kernel_image_mappings; i++) { 714 prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr); 715 prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr); 716 tte_vaddr += 0x400000; 717 tte_data += 0x400000; 718 } 719 sparc64_highest_unlocked_tlb_ent = tlb_ent - i; 720 } 721 if (tlb_type == cheetah_plus) { 722 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 | 723 CTX_CHEETAH_PLUS_NUC); 724 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC; 725 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0; 726 } 727 } 728 729 730 static void __init inherit_prom_mappings(void) 731 { 732 /* Now fixup OBP's idea about where we really are mapped. */ 733 printk("Remapping the kernel... "); 734 remap_kernel(); 735 printk("done.\n"); 736 } 737 738 void prom_world(int enter) 739 { 740 if (!enter) 741 set_fs(get_fs()); 742 743 __asm__ __volatile__("flushw"); 744 } 745 746 void __flush_dcache_range(unsigned long start, unsigned long end) 747 { 748 unsigned long va; 749 750 if (tlb_type == spitfire) { 751 int n = 0; 752 753 for (va = start; va < end; va += 32) { 754 spitfire_put_dcache_tag(va & 0x3fe0, 0x0); 755 if (++n >= 512) 756 break; 757 } 758 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) { 759 start = __pa(start); 760 end = __pa(end); 761 for (va = start; va < end; va += 32) 762 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 763 "membar #Sync" 764 : /* no outputs */ 765 : "r" (va), 766 "i" (ASI_DCACHE_INVALIDATE)); 767 } 768 } 769 EXPORT_SYMBOL(__flush_dcache_range); 770 771 /* get_new_mmu_context() uses "cache + 1". */ 772 DEFINE_SPINLOCK(ctx_alloc_lock); 773 unsigned long tlb_context_cache = CTX_FIRST_VERSION; 774 #define MAX_CTX_NR (1UL << CTX_NR_BITS) 775 #define CTX_BMAP_SLOTS BITS_TO_LONGS(MAX_CTX_NR) 776 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR); 777 DEFINE_PER_CPU(struct mm_struct *, per_cpu_secondary_mm) = {0}; 778 779 static void mmu_context_wrap(void) 780 { 781 unsigned long old_ver = tlb_context_cache & CTX_VERSION_MASK; 782 unsigned long new_ver, new_ctx, old_ctx; 783 struct mm_struct *mm; 784 int cpu; 785 786 bitmap_zero(mmu_context_bmap, 1 << CTX_NR_BITS); 787 788 /* Reserve kernel context */ 789 set_bit(0, mmu_context_bmap); 790 791 new_ver = (tlb_context_cache & CTX_VERSION_MASK) + CTX_FIRST_VERSION; 792 if (unlikely(new_ver == 0)) 793 new_ver = CTX_FIRST_VERSION; 794 tlb_context_cache = new_ver; 795 796 /* 797 * Make sure that any new mm that are added into per_cpu_secondary_mm, 798 * are going to go through get_new_mmu_context() path. 799 */ 800 mb(); 801 802 /* 803 * Updated versions to current on those CPUs that had valid secondary 804 * contexts 805 */ 806 for_each_online_cpu(cpu) { 807 /* 808 * If a new mm is stored after we took this mm from the array, 809 * it will go into get_new_mmu_context() path, because we 810 * already bumped the version in tlb_context_cache. 811 */ 812 mm = per_cpu(per_cpu_secondary_mm, cpu); 813 814 if (unlikely(!mm || mm == &init_mm)) 815 continue; 816 817 old_ctx = mm->context.sparc64_ctx_val; 818 if (likely((old_ctx & CTX_VERSION_MASK) == old_ver)) { 819 new_ctx = (old_ctx & ~CTX_VERSION_MASK) | new_ver; 820 set_bit(new_ctx & CTX_NR_MASK, mmu_context_bmap); 821 mm->context.sparc64_ctx_val = new_ctx; 822 } 823 } 824 } 825 826 /* Caller does TLB context flushing on local CPU if necessary. 827 * The caller also ensures that CTX_VALID(mm->context) is false. 828 * 829 * We must be careful about boundary cases so that we never 830 * let the user have CTX 0 (nucleus) or we ever use a CTX 831 * version of zero (and thus NO_CONTEXT would not be caught 832 * by version mis-match tests in mmu_context.h). 833 * 834 * Always invoked with interrupts disabled. 835 */ 836 void get_new_mmu_context(struct mm_struct *mm) 837 { 838 unsigned long ctx, new_ctx; 839 unsigned long orig_pgsz_bits; 840 841 spin_lock(&ctx_alloc_lock); 842 retry: 843 /* wrap might have happened, test again if our context became valid */ 844 if (unlikely(CTX_VALID(mm->context))) 845 goto out; 846 orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK); 847 ctx = (tlb_context_cache + 1) & CTX_NR_MASK; 848 new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx); 849 if (new_ctx >= (1 << CTX_NR_BITS)) { 850 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1); 851 if (new_ctx >= ctx) { 852 mmu_context_wrap(); 853 goto retry; 854 } 855 } 856 if (mm->context.sparc64_ctx_val) 857 cpumask_clear(mm_cpumask(mm)); 858 mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63)); 859 new_ctx |= (tlb_context_cache & CTX_VERSION_MASK); 860 tlb_context_cache = new_ctx; 861 mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits; 862 out: 863 spin_unlock(&ctx_alloc_lock); 864 } 865 866 static int numa_enabled = 1; 867 static int numa_debug; 868 869 static int __init early_numa(char *p) 870 { 871 if (!p) 872 return 0; 873 874 if (strstr(p, "off")) 875 numa_enabled = 0; 876 877 if (strstr(p, "debug")) 878 numa_debug = 1; 879 880 return 0; 881 } 882 early_param("numa", early_numa); 883 884 #define numadbg(f, a...) \ 885 do { if (numa_debug) \ 886 printk(KERN_INFO f, ## a); \ 887 } while (0) 888 889 static void __init find_ramdisk(unsigned long phys_base) 890 { 891 #ifdef CONFIG_BLK_DEV_INITRD 892 if (sparc_ramdisk_image || sparc_ramdisk_image64) { 893 unsigned long ramdisk_image; 894 895 /* Older versions of the bootloader only supported a 896 * 32-bit physical address for the ramdisk image 897 * location, stored at sparc_ramdisk_image. Newer 898 * SILO versions set sparc_ramdisk_image to zero and 899 * provide a full 64-bit physical address at 900 * sparc_ramdisk_image64. 901 */ 902 ramdisk_image = sparc_ramdisk_image; 903 if (!ramdisk_image) 904 ramdisk_image = sparc_ramdisk_image64; 905 906 /* Another bootloader quirk. The bootloader normalizes 907 * the physical address to KERNBASE, so we have to 908 * factor that back out and add in the lowest valid 909 * physical page address to get the true physical address. 910 */ 911 ramdisk_image -= KERNBASE; 912 ramdisk_image += phys_base; 913 914 numadbg("Found ramdisk at physical address 0x%lx, size %u\n", 915 ramdisk_image, sparc_ramdisk_size); 916 917 initrd_start = ramdisk_image; 918 initrd_end = ramdisk_image + sparc_ramdisk_size; 919 920 memblock_reserve(initrd_start, sparc_ramdisk_size); 921 922 initrd_start += PAGE_OFFSET; 923 initrd_end += PAGE_OFFSET; 924 } 925 #endif 926 } 927 928 struct node_mem_mask { 929 unsigned long mask; 930 unsigned long match; 931 }; 932 static struct node_mem_mask node_masks[MAX_NUMNODES]; 933 static int num_node_masks; 934 935 #ifdef CONFIG_NEED_MULTIPLE_NODES 936 937 struct mdesc_mlgroup { 938 u64 node; 939 u64 latency; 940 u64 match; 941 u64 mask; 942 }; 943 944 static struct mdesc_mlgroup *mlgroups; 945 static int num_mlgroups; 946 947 int numa_cpu_lookup_table[NR_CPUS]; 948 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES]; 949 950 struct mdesc_mblock { 951 u64 base; 952 u64 size; 953 u64 offset; /* RA-to-PA */ 954 }; 955 static struct mdesc_mblock *mblocks; 956 static int num_mblocks; 957 958 static struct mdesc_mblock * __init addr_to_mblock(unsigned long addr) 959 { 960 struct mdesc_mblock *m = NULL; 961 int i; 962 963 for (i = 0; i < num_mblocks; i++) { 964 m = &mblocks[i]; 965 966 if (addr >= m->base && 967 addr < (m->base + m->size)) { 968 break; 969 } 970 } 971 972 return m; 973 } 974 975 static u64 __init memblock_nid_range_sun4u(u64 start, u64 end, int *nid) 976 { 977 int prev_nid, new_nid; 978 979 prev_nid = NUMA_NO_NODE; 980 for ( ; start < end; start += PAGE_SIZE) { 981 for (new_nid = 0; new_nid < num_node_masks; new_nid++) { 982 struct node_mem_mask *p = &node_masks[new_nid]; 983 984 if ((start & p->mask) == p->match) { 985 if (prev_nid == NUMA_NO_NODE) 986 prev_nid = new_nid; 987 break; 988 } 989 } 990 991 if (new_nid == num_node_masks) { 992 prev_nid = 0; 993 WARN_ONCE(1, "addr[%Lx] doesn't match a NUMA node rule. Some memory will be owned by node 0.", 994 start); 995 break; 996 } 997 998 if (prev_nid != new_nid) 999 break; 1000 } 1001 *nid = prev_nid; 1002 1003 return start > end ? end : start; 1004 } 1005 1006 static u64 __init memblock_nid_range(u64 start, u64 end, int *nid) 1007 { 1008 u64 ret_end, pa_start, m_mask, m_match, m_end; 1009 struct mdesc_mblock *mblock; 1010 int _nid, i; 1011 1012 if (tlb_type != hypervisor) 1013 return memblock_nid_range_sun4u(start, end, nid); 1014 1015 mblock = addr_to_mblock(start); 1016 if (!mblock) { 1017 WARN_ONCE(1, "memblock_nid_range: Can't find mblock addr[%Lx]", 1018 start); 1019 1020 _nid = 0; 1021 ret_end = end; 1022 goto done; 1023 } 1024 1025 pa_start = start + mblock->offset; 1026 m_match = 0; 1027 m_mask = 0; 1028 1029 for (_nid = 0; _nid < num_node_masks; _nid++) { 1030 struct node_mem_mask *const m = &node_masks[_nid]; 1031 1032 if ((pa_start & m->mask) == m->match) { 1033 m_match = m->match; 1034 m_mask = m->mask; 1035 break; 1036 } 1037 } 1038 1039 if (num_node_masks == _nid) { 1040 /* We could not find NUMA group, so default to 0, but lets 1041 * search for latency group, so we could calculate the correct 1042 * end address that we return 1043 */ 1044 _nid = 0; 1045 1046 for (i = 0; i < num_mlgroups; i++) { 1047 struct mdesc_mlgroup *const m = &mlgroups[i]; 1048 1049 if ((pa_start & m->mask) == m->match) { 1050 m_match = m->match; 1051 m_mask = m->mask; 1052 break; 1053 } 1054 } 1055 1056 if (i == num_mlgroups) { 1057 WARN_ONCE(1, "memblock_nid_range: Can't find latency group addr[%Lx]", 1058 start); 1059 1060 ret_end = end; 1061 goto done; 1062 } 1063 } 1064 1065 /* 1066 * Each latency group has match and mask, and each memory block has an 1067 * offset. An address belongs to a latency group if its address matches 1068 * the following formula: ((addr + offset) & mask) == match 1069 * It is, however, slow to check every single page if it matches a 1070 * particular latency group. As optimization we calculate end value by 1071 * using bit arithmetics. 1072 */ 1073 m_end = m_match + (1ul << __ffs(m_mask)) - mblock->offset; 1074 m_end += pa_start & ~((1ul << fls64(m_mask)) - 1); 1075 ret_end = m_end > end ? end : m_end; 1076 1077 done: 1078 *nid = _nid; 1079 return ret_end; 1080 } 1081 #endif 1082 1083 /* This must be invoked after performing all of the necessary 1084 * memblock_set_node() calls for 'nid'. We need to be able to get 1085 * correct data from get_pfn_range_for_nid(). 1086 */ 1087 static void __init allocate_node_data(int nid) 1088 { 1089 struct pglist_data *p; 1090 unsigned long start_pfn, end_pfn; 1091 #ifdef CONFIG_NEED_MULTIPLE_NODES 1092 1093 NODE_DATA(nid) = memblock_alloc_node(sizeof(struct pglist_data), 1094 SMP_CACHE_BYTES, nid); 1095 if (!NODE_DATA(nid)) { 1096 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid); 1097 prom_halt(); 1098 } 1099 1100 NODE_DATA(nid)->node_id = nid; 1101 #endif 1102 1103 p = NODE_DATA(nid); 1104 1105 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn); 1106 p->node_start_pfn = start_pfn; 1107 p->node_spanned_pages = end_pfn - start_pfn; 1108 } 1109 1110 static void init_node_masks_nonnuma(void) 1111 { 1112 #ifdef CONFIG_NEED_MULTIPLE_NODES 1113 int i; 1114 #endif 1115 1116 numadbg("Initializing tables for non-numa.\n"); 1117 1118 node_masks[0].mask = 0; 1119 node_masks[0].match = 0; 1120 num_node_masks = 1; 1121 1122 #ifdef CONFIG_NEED_MULTIPLE_NODES 1123 for (i = 0; i < NR_CPUS; i++) 1124 numa_cpu_lookup_table[i] = 0; 1125 1126 cpumask_setall(&numa_cpumask_lookup_table[0]); 1127 #endif 1128 } 1129 1130 #ifdef CONFIG_NEED_MULTIPLE_NODES 1131 struct pglist_data *node_data[MAX_NUMNODES]; 1132 1133 EXPORT_SYMBOL(numa_cpu_lookup_table); 1134 EXPORT_SYMBOL(numa_cpumask_lookup_table); 1135 EXPORT_SYMBOL(node_data); 1136 1137 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio, 1138 u32 cfg_handle) 1139 { 1140 u64 arc; 1141 1142 mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) { 1143 u64 target = mdesc_arc_target(md, arc); 1144 const u64 *val; 1145 1146 val = mdesc_get_property(md, target, 1147 "cfg-handle", NULL); 1148 if (val && *val == cfg_handle) 1149 return 0; 1150 } 1151 return -ENODEV; 1152 } 1153 1154 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp, 1155 u32 cfg_handle) 1156 { 1157 u64 arc, candidate, best_latency = ~(u64)0; 1158 1159 candidate = MDESC_NODE_NULL; 1160 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1161 u64 target = mdesc_arc_target(md, arc); 1162 const char *name = mdesc_node_name(md, target); 1163 const u64 *val; 1164 1165 if (strcmp(name, "pio-latency-group")) 1166 continue; 1167 1168 val = mdesc_get_property(md, target, "latency", NULL); 1169 if (!val) 1170 continue; 1171 1172 if (*val < best_latency) { 1173 candidate = target; 1174 best_latency = *val; 1175 } 1176 } 1177 1178 if (candidate == MDESC_NODE_NULL) 1179 return -ENODEV; 1180 1181 return scan_pio_for_cfg_handle(md, candidate, cfg_handle); 1182 } 1183 1184 int of_node_to_nid(struct device_node *dp) 1185 { 1186 const struct linux_prom64_registers *regs; 1187 struct mdesc_handle *md; 1188 u32 cfg_handle; 1189 int count, nid; 1190 u64 grp; 1191 1192 /* This is the right thing to do on currently supported 1193 * SUN4U NUMA platforms as well, as the PCI controller does 1194 * not sit behind any particular memory controller. 1195 */ 1196 if (!mlgroups) 1197 return -1; 1198 1199 regs = of_get_property(dp, "reg", NULL); 1200 if (!regs) 1201 return -1; 1202 1203 cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff; 1204 1205 md = mdesc_grab(); 1206 1207 count = 0; 1208 nid = NUMA_NO_NODE; 1209 mdesc_for_each_node_by_name(md, grp, "group") { 1210 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) { 1211 nid = count; 1212 break; 1213 } 1214 count++; 1215 } 1216 1217 mdesc_release(md); 1218 1219 return nid; 1220 } 1221 1222 static void __init add_node_ranges(void) 1223 { 1224 struct memblock_region *reg; 1225 unsigned long prev_max; 1226 1227 memblock_resized: 1228 prev_max = memblock.memory.max; 1229 1230 for_each_memblock(memory, reg) { 1231 unsigned long size = reg->size; 1232 unsigned long start, end; 1233 1234 start = reg->base; 1235 end = start + size; 1236 while (start < end) { 1237 unsigned long this_end; 1238 int nid; 1239 1240 this_end = memblock_nid_range(start, end, &nid); 1241 1242 numadbg("Setting memblock NUMA node nid[%d] " 1243 "start[%lx] end[%lx]\n", 1244 nid, start, this_end); 1245 1246 memblock_set_node(start, this_end - start, 1247 &memblock.memory, nid); 1248 if (memblock.memory.max != prev_max) 1249 goto memblock_resized; 1250 start = this_end; 1251 } 1252 } 1253 } 1254 1255 static int __init grab_mlgroups(struct mdesc_handle *md) 1256 { 1257 unsigned long paddr; 1258 int count = 0; 1259 u64 node; 1260 1261 mdesc_for_each_node_by_name(md, node, "memory-latency-group") 1262 count++; 1263 if (!count) 1264 return -ENOENT; 1265 1266 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mlgroup), 1267 SMP_CACHE_BYTES); 1268 if (!paddr) 1269 return -ENOMEM; 1270 1271 mlgroups = __va(paddr); 1272 num_mlgroups = count; 1273 1274 count = 0; 1275 mdesc_for_each_node_by_name(md, node, "memory-latency-group") { 1276 struct mdesc_mlgroup *m = &mlgroups[count++]; 1277 const u64 *val; 1278 1279 m->node = node; 1280 1281 val = mdesc_get_property(md, node, "latency", NULL); 1282 m->latency = *val; 1283 val = mdesc_get_property(md, node, "address-match", NULL); 1284 m->match = *val; 1285 val = mdesc_get_property(md, node, "address-mask", NULL); 1286 m->mask = *val; 1287 1288 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] " 1289 "match[%llx] mask[%llx]\n", 1290 count - 1, m->node, m->latency, m->match, m->mask); 1291 } 1292 1293 return 0; 1294 } 1295 1296 static int __init grab_mblocks(struct mdesc_handle *md) 1297 { 1298 unsigned long paddr; 1299 int count = 0; 1300 u64 node; 1301 1302 mdesc_for_each_node_by_name(md, node, "mblock") 1303 count++; 1304 if (!count) 1305 return -ENOENT; 1306 1307 paddr = memblock_phys_alloc(count * sizeof(struct mdesc_mblock), 1308 SMP_CACHE_BYTES); 1309 if (!paddr) 1310 return -ENOMEM; 1311 1312 mblocks = __va(paddr); 1313 num_mblocks = count; 1314 1315 count = 0; 1316 mdesc_for_each_node_by_name(md, node, "mblock") { 1317 struct mdesc_mblock *m = &mblocks[count++]; 1318 const u64 *val; 1319 1320 val = mdesc_get_property(md, node, "base", NULL); 1321 m->base = *val; 1322 val = mdesc_get_property(md, node, "size", NULL); 1323 m->size = *val; 1324 val = mdesc_get_property(md, node, 1325 "address-congruence-offset", NULL); 1326 1327 /* The address-congruence-offset property is optional. 1328 * Explicity zero it be identifty this. 1329 */ 1330 if (val) 1331 m->offset = *val; 1332 else 1333 m->offset = 0UL; 1334 1335 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n", 1336 count - 1, m->base, m->size, m->offset); 1337 } 1338 1339 return 0; 1340 } 1341 1342 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md, 1343 u64 grp, cpumask_t *mask) 1344 { 1345 u64 arc; 1346 1347 cpumask_clear(mask); 1348 1349 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) { 1350 u64 target = mdesc_arc_target(md, arc); 1351 const char *name = mdesc_node_name(md, target); 1352 const u64 *id; 1353 1354 if (strcmp(name, "cpu")) 1355 continue; 1356 id = mdesc_get_property(md, target, "id", NULL); 1357 if (*id < nr_cpu_ids) 1358 cpumask_set_cpu(*id, mask); 1359 } 1360 } 1361 1362 static struct mdesc_mlgroup * __init find_mlgroup(u64 node) 1363 { 1364 int i; 1365 1366 for (i = 0; i < num_mlgroups; i++) { 1367 struct mdesc_mlgroup *m = &mlgroups[i]; 1368 if (m->node == node) 1369 return m; 1370 } 1371 return NULL; 1372 } 1373 1374 int __node_distance(int from, int to) 1375 { 1376 if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) { 1377 pr_warn("Returning default NUMA distance value for %d->%d\n", 1378 from, to); 1379 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE; 1380 } 1381 return numa_latency[from][to]; 1382 } 1383 EXPORT_SYMBOL(__node_distance); 1384 1385 static int __init find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp) 1386 { 1387 int i; 1388 1389 for (i = 0; i < MAX_NUMNODES; i++) { 1390 struct node_mem_mask *n = &node_masks[i]; 1391 1392 if ((grp->mask == n->mask) && (grp->match == n->match)) 1393 break; 1394 } 1395 return i; 1396 } 1397 1398 static void __init find_numa_latencies_for_group(struct mdesc_handle *md, 1399 u64 grp, int index) 1400 { 1401 u64 arc; 1402 1403 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1404 int tnode; 1405 u64 target = mdesc_arc_target(md, arc); 1406 struct mdesc_mlgroup *m = find_mlgroup(target); 1407 1408 if (!m) 1409 continue; 1410 tnode = find_best_numa_node_for_mlgroup(m); 1411 if (tnode == MAX_NUMNODES) 1412 continue; 1413 numa_latency[index][tnode] = m->latency; 1414 } 1415 } 1416 1417 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp, 1418 int index) 1419 { 1420 struct mdesc_mlgroup *candidate = NULL; 1421 u64 arc, best_latency = ~(u64)0; 1422 struct node_mem_mask *n; 1423 1424 mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) { 1425 u64 target = mdesc_arc_target(md, arc); 1426 struct mdesc_mlgroup *m = find_mlgroup(target); 1427 if (!m) 1428 continue; 1429 if (m->latency < best_latency) { 1430 candidate = m; 1431 best_latency = m->latency; 1432 } 1433 } 1434 if (!candidate) 1435 return -ENOENT; 1436 1437 if (num_node_masks != index) { 1438 printk(KERN_ERR "Inconsistent NUMA state, " 1439 "index[%d] != num_node_masks[%d]\n", 1440 index, num_node_masks); 1441 return -EINVAL; 1442 } 1443 1444 n = &node_masks[num_node_masks++]; 1445 1446 n->mask = candidate->mask; 1447 n->match = candidate->match; 1448 1449 numadbg("NUMA NODE[%d]: mask[%lx] match[%lx] (latency[%llx])\n", 1450 index, n->mask, n->match, candidate->latency); 1451 1452 return 0; 1453 } 1454 1455 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp, 1456 int index) 1457 { 1458 cpumask_t mask; 1459 int cpu; 1460 1461 numa_parse_mdesc_group_cpus(md, grp, &mask); 1462 1463 for_each_cpu(cpu, &mask) 1464 numa_cpu_lookup_table[cpu] = index; 1465 cpumask_copy(&numa_cpumask_lookup_table[index], &mask); 1466 1467 if (numa_debug) { 1468 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index); 1469 for_each_cpu(cpu, &mask) 1470 printk("%d ", cpu); 1471 printk("]\n"); 1472 } 1473 1474 return numa_attach_mlgroup(md, grp, index); 1475 } 1476 1477 static int __init numa_parse_mdesc(void) 1478 { 1479 struct mdesc_handle *md = mdesc_grab(); 1480 int i, j, err, count; 1481 u64 node; 1482 1483 node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups"); 1484 if (node == MDESC_NODE_NULL) { 1485 mdesc_release(md); 1486 return -ENOENT; 1487 } 1488 1489 err = grab_mblocks(md); 1490 if (err < 0) 1491 goto out; 1492 1493 err = grab_mlgroups(md); 1494 if (err < 0) 1495 goto out; 1496 1497 count = 0; 1498 mdesc_for_each_node_by_name(md, node, "group") { 1499 err = numa_parse_mdesc_group(md, node, count); 1500 if (err < 0) 1501 break; 1502 count++; 1503 } 1504 1505 count = 0; 1506 mdesc_for_each_node_by_name(md, node, "group") { 1507 find_numa_latencies_for_group(md, node, count); 1508 count++; 1509 } 1510 1511 /* Normalize numa latency matrix according to ACPI SLIT spec. */ 1512 for (i = 0; i < MAX_NUMNODES; i++) { 1513 u64 self_latency = numa_latency[i][i]; 1514 1515 for (j = 0; j < MAX_NUMNODES; j++) { 1516 numa_latency[i][j] = 1517 (numa_latency[i][j] * LOCAL_DISTANCE) / 1518 self_latency; 1519 } 1520 } 1521 1522 add_node_ranges(); 1523 1524 for (i = 0; i < num_node_masks; i++) { 1525 allocate_node_data(i); 1526 node_set_online(i); 1527 } 1528 1529 err = 0; 1530 out: 1531 mdesc_release(md); 1532 return err; 1533 } 1534 1535 static int __init numa_parse_jbus(void) 1536 { 1537 unsigned long cpu, index; 1538 1539 /* NUMA node id is encoded in bits 36 and higher, and there is 1540 * a 1-to-1 mapping from CPU ID to NUMA node ID. 1541 */ 1542 index = 0; 1543 for_each_present_cpu(cpu) { 1544 numa_cpu_lookup_table[cpu] = index; 1545 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu)); 1546 node_masks[index].mask = ~((1UL << 36UL) - 1UL); 1547 node_masks[index].match = cpu << 36UL; 1548 1549 index++; 1550 } 1551 num_node_masks = index; 1552 1553 add_node_ranges(); 1554 1555 for (index = 0; index < num_node_masks; index++) { 1556 allocate_node_data(index); 1557 node_set_online(index); 1558 } 1559 1560 return 0; 1561 } 1562 1563 static int __init numa_parse_sun4u(void) 1564 { 1565 if (tlb_type == cheetah || tlb_type == cheetah_plus) { 1566 unsigned long ver; 1567 1568 __asm__ ("rdpr %%ver, %0" : "=r" (ver)); 1569 if ((ver >> 32UL) == __JALAPENO_ID || 1570 (ver >> 32UL) == __SERRANO_ID) 1571 return numa_parse_jbus(); 1572 } 1573 return -1; 1574 } 1575 1576 static int __init bootmem_init_numa(void) 1577 { 1578 int i, j; 1579 int err = -1; 1580 1581 numadbg("bootmem_init_numa()\n"); 1582 1583 /* Some sane defaults for numa latency values */ 1584 for (i = 0; i < MAX_NUMNODES; i++) { 1585 for (j = 0; j < MAX_NUMNODES; j++) 1586 numa_latency[i][j] = (i == j) ? 1587 LOCAL_DISTANCE : REMOTE_DISTANCE; 1588 } 1589 1590 if (numa_enabled) { 1591 if (tlb_type == hypervisor) 1592 err = numa_parse_mdesc(); 1593 else 1594 err = numa_parse_sun4u(); 1595 } 1596 return err; 1597 } 1598 1599 #else 1600 1601 static int bootmem_init_numa(void) 1602 { 1603 return -1; 1604 } 1605 1606 #endif 1607 1608 static void __init bootmem_init_nonnuma(void) 1609 { 1610 unsigned long top_of_ram = memblock_end_of_DRAM(); 1611 unsigned long total_ram = memblock_phys_mem_size(); 1612 1613 numadbg("bootmem_init_nonnuma()\n"); 1614 1615 printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n", 1616 top_of_ram, total_ram); 1617 printk(KERN_INFO "Memory hole size: %ldMB\n", 1618 (top_of_ram - total_ram) >> 20); 1619 1620 init_node_masks_nonnuma(); 1621 memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0); 1622 allocate_node_data(0); 1623 node_set_online(0); 1624 } 1625 1626 static unsigned long __init bootmem_init(unsigned long phys_base) 1627 { 1628 unsigned long end_pfn; 1629 1630 end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT; 1631 max_pfn = max_low_pfn = end_pfn; 1632 min_low_pfn = (phys_base >> PAGE_SHIFT); 1633 1634 if (bootmem_init_numa() < 0) 1635 bootmem_init_nonnuma(); 1636 1637 /* Dump memblock with node info. */ 1638 memblock_dump_all(); 1639 1640 /* XXX cpu notifier XXX */ 1641 1642 sparse_memory_present_with_active_regions(MAX_NUMNODES); 1643 sparse_init(); 1644 1645 return end_pfn; 1646 } 1647 1648 static struct linux_prom64_registers pall[MAX_BANKS] __initdata; 1649 static int pall_ents __initdata; 1650 1651 static unsigned long max_phys_bits = 40; 1652 1653 bool kern_addr_valid(unsigned long addr) 1654 { 1655 pgd_t *pgd; 1656 pud_t *pud; 1657 pmd_t *pmd; 1658 pte_t *pte; 1659 1660 if ((long)addr < 0L) { 1661 unsigned long pa = __pa(addr); 1662 1663 if ((pa >> max_phys_bits) != 0UL) 1664 return false; 1665 1666 return pfn_valid(pa >> PAGE_SHIFT); 1667 } 1668 1669 if (addr >= (unsigned long) KERNBASE && 1670 addr < (unsigned long)&_end) 1671 return true; 1672 1673 pgd = pgd_offset_k(addr); 1674 if (pgd_none(*pgd)) 1675 return 0; 1676 1677 pud = pud_offset(pgd, addr); 1678 if (pud_none(*pud)) 1679 return 0; 1680 1681 if (pud_large(*pud)) 1682 return pfn_valid(pud_pfn(*pud)); 1683 1684 pmd = pmd_offset(pud, addr); 1685 if (pmd_none(*pmd)) 1686 return 0; 1687 1688 if (pmd_large(*pmd)) 1689 return pfn_valid(pmd_pfn(*pmd)); 1690 1691 pte = pte_offset_kernel(pmd, addr); 1692 if (pte_none(*pte)) 1693 return 0; 1694 1695 return pfn_valid(pte_pfn(*pte)); 1696 } 1697 EXPORT_SYMBOL(kern_addr_valid); 1698 1699 static unsigned long __ref kernel_map_hugepud(unsigned long vstart, 1700 unsigned long vend, 1701 pud_t *pud) 1702 { 1703 const unsigned long mask16gb = (1UL << 34) - 1UL; 1704 u64 pte_val = vstart; 1705 1706 /* Each PUD is 8GB */ 1707 if ((vstart & mask16gb) || 1708 (vend - vstart <= mask16gb)) { 1709 pte_val ^= kern_linear_pte_xor[2]; 1710 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE; 1711 1712 return vstart + PUD_SIZE; 1713 } 1714 1715 pte_val ^= kern_linear_pte_xor[3]; 1716 pte_val |= _PAGE_PUD_HUGE; 1717 1718 vend = vstart + mask16gb + 1UL; 1719 while (vstart < vend) { 1720 pud_val(*pud) = pte_val; 1721 1722 pte_val += PUD_SIZE; 1723 vstart += PUD_SIZE; 1724 pud++; 1725 } 1726 return vstart; 1727 } 1728 1729 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend, 1730 bool guard) 1731 { 1732 if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE) 1733 return true; 1734 1735 return false; 1736 } 1737 1738 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart, 1739 unsigned long vend, 1740 pmd_t *pmd) 1741 { 1742 const unsigned long mask256mb = (1UL << 28) - 1UL; 1743 const unsigned long mask2gb = (1UL << 31) - 1UL; 1744 u64 pte_val = vstart; 1745 1746 /* Each PMD is 8MB */ 1747 if ((vstart & mask256mb) || 1748 (vend - vstart <= mask256mb)) { 1749 pte_val ^= kern_linear_pte_xor[0]; 1750 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE; 1751 1752 return vstart + PMD_SIZE; 1753 } 1754 1755 if ((vstart & mask2gb) || 1756 (vend - vstart <= mask2gb)) { 1757 pte_val ^= kern_linear_pte_xor[1]; 1758 pte_val |= _PAGE_PMD_HUGE; 1759 vend = vstart + mask256mb + 1UL; 1760 } else { 1761 pte_val ^= kern_linear_pte_xor[2]; 1762 pte_val |= _PAGE_PMD_HUGE; 1763 vend = vstart + mask2gb + 1UL; 1764 } 1765 1766 while (vstart < vend) { 1767 pmd_val(*pmd) = pte_val; 1768 1769 pte_val += PMD_SIZE; 1770 vstart += PMD_SIZE; 1771 pmd++; 1772 } 1773 1774 return vstart; 1775 } 1776 1777 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend, 1778 bool guard) 1779 { 1780 if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE) 1781 return true; 1782 1783 return false; 1784 } 1785 1786 static unsigned long __ref kernel_map_range(unsigned long pstart, 1787 unsigned long pend, pgprot_t prot, 1788 bool use_huge) 1789 { 1790 unsigned long vstart = PAGE_OFFSET + pstart; 1791 unsigned long vend = PAGE_OFFSET + pend; 1792 unsigned long alloc_bytes = 0UL; 1793 1794 if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) { 1795 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n", 1796 vstart, vend); 1797 prom_halt(); 1798 } 1799 1800 while (vstart < vend) { 1801 unsigned long this_end, paddr = __pa(vstart); 1802 pgd_t *pgd = pgd_offset_k(vstart); 1803 pud_t *pud; 1804 pmd_t *pmd; 1805 pte_t *pte; 1806 1807 if (pgd_none(*pgd)) { 1808 pud_t *new; 1809 1810 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1811 PAGE_SIZE); 1812 if (!new) 1813 goto err_alloc; 1814 alloc_bytes += PAGE_SIZE; 1815 pgd_populate(&init_mm, pgd, new); 1816 } 1817 pud = pud_offset(pgd, vstart); 1818 if (pud_none(*pud)) { 1819 pmd_t *new; 1820 1821 if (kernel_can_map_hugepud(vstart, vend, use_huge)) { 1822 vstart = kernel_map_hugepud(vstart, vend, pud); 1823 continue; 1824 } 1825 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1826 PAGE_SIZE); 1827 if (!new) 1828 goto err_alloc; 1829 alloc_bytes += PAGE_SIZE; 1830 pud_populate(&init_mm, pud, new); 1831 } 1832 1833 pmd = pmd_offset(pud, vstart); 1834 if (pmd_none(*pmd)) { 1835 pte_t *new; 1836 1837 if (kernel_can_map_hugepmd(vstart, vend, use_huge)) { 1838 vstart = kernel_map_hugepmd(vstart, vend, pmd); 1839 continue; 1840 } 1841 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, 1842 PAGE_SIZE); 1843 if (!new) 1844 goto err_alloc; 1845 alloc_bytes += PAGE_SIZE; 1846 pmd_populate_kernel(&init_mm, pmd, new); 1847 } 1848 1849 pte = pte_offset_kernel(pmd, vstart); 1850 this_end = (vstart + PMD_SIZE) & PMD_MASK; 1851 if (this_end > vend) 1852 this_end = vend; 1853 1854 while (vstart < this_end) { 1855 pte_val(*pte) = (paddr | pgprot_val(prot)); 1856 1857 vstart += PAGE_SIZE; 1858 paddr += PAGE_SIZE; 1859 pte++; 1860 } 1861 } 1862 1863 return alloc_bytes; 1864 1865 err_alloc: 1866 panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n", 1867 __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE); 1868 return -ENOMEM; 1869 } 1870 1871 static void __init flush_all_kernel_tsbs(void) 1872 { 1873 int i; 1874 1875 for (i = 0; i < KERNEL_TSB_NENTRIES; i++) { 1876 struct tsb *ent = &swapper_tsb[i]; 1877 1878 ent->tag = (1UL << TSB_TAG_INVALID_BIT); 1879 } 1880 #ifndef CONFIG_DEBUG_PAGEALLOC 1881 for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) { 1882 struct tsb *ent = &swapper_4m_tsb[i]; 1883 1884 ent->tag = (1UL << TSB_TAG_INVALID_BIT); 1885 } 1886 #endif 1887 } 1888 1889 extern unsigned int kvmap_linear_patch[1]; 1890 1891 static void __init kernel_physical_mapping_init(void) 1892 { 1893 unsigned long i, mem_alloced = 0UL; 1894 bool use_huge = true; 1895 1896 #ifdef CONFIG_DEBUG_PAGEALLOC 1897 use_huge = false; 1898 #endif 1899 for (i = 0; i < pall_ents; i++) { 1900 unsigned long phys_start, phys_end; 1901 1902 phys_start = pall[i].phys_addr; 1903 phys_end = phys_start + pall[i].reg_size; 1904 1905 mem_alloced += kernel_map_range(phys_start, phys_end, 1906 PAGE_KERNEL, use_huge); 1907 } 1908 1909 printk("Allocated %ld bytes for kernel page tables.\n", 1910 mem_alloced); 1911 1912 kvmap_linear_patch[0] = 0x01000000; /* nop */ 1913 flushi(&kvmap_linear_patch[0]); 1914 1915 flush_all_kernel_tsbs(); 1916 1917 __flush_tlb_all(); 1918 } 1919 1920 #ifdef CONFIG_DEBUG_PAGEALLOC 1921 void __kernel_map_pages(struct page *page, int numpages, int enable) 1922 { 1923 unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT; 1924 unsigned long phys_end = phys_start + (numpages * PAGE_SIZE); 1925 1926 kernel_map_range(phys_start, phys_end, 1927 (enable ? PAGE_KERNEL : __pgprot(0)), false); 1928 1929 flush_tsb_kernel_range(PAGE_OFFSET + phys_start, 1930 PAGE_OFFSET + phys_end); 1931 1932 /* we should perform an IPI and flush all tlbs, 1933 * but that can deadlock->flush only current cpu. 1934 */ 1935 __flush_tlb_kernel_range(PAGE_OFFSET + phys_start, 1936 PAGE_OFFSET + phys_end); 1937 } 1938 #endif 1939 1940 unsigned long __init find_ecache_flush_span(unsigned long size) 1941 { 1942 int i; 1943 1944 for (i = 0; i < pavail_ents; i++) { 1945 if (pavail[i].reg_size >= size) 1946 return pavail[i].phys_addr; 1947 } 1948 1949 return ~0UL; 1950 } 1951 1952 unsigned long PAGE_OFFSET; 1953 EXPORT_SYMBOL(PAGE_OFFSET); 1954 1955 unsigned long VMALLOC_END = 0x0000010000000000UL; 1956 EXPORT_SYMBOL(VMALLOC_END); 1957 1958 unsigned long sparc64_va_hole_top = 0xfffff80000000000UL; 1959 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL; 1960 1961 static void __init setup_page_offset(void) 1962 { 1963 if (tlb_type == cheetah || tlb_type == cheetah_plus) { 1964 /* Cheetah/Panther support a full 64-bit virtual 1965 * address, so we can use all that our page tables 1966 * support. 1967 */ 1968 sparc64_va_hole_top = 0xfff0000000000000UL; 1969 sparc64_va_hole_bottom = 0x0010000000000000UL; 1970 1971 max_phys_bits = 42; 1972 } else if (tlb_type == hypervisor) { 1973 switch (sun4v_chip_type) { 1974 case SUN4V_CHIP_NIAGARA1: 1975 case SUN4V_CHIP_NIAGARA2: 1976 /* T1 and T2 support 48-bit virtual addresses. */ 1977 sparc64_va_hole_top = 0xffff800000000000UL; 1978 sparc64_va_hole_bottom = 0x0000800000000000UL; 1979 1980 max_phys_bits = 39; 1981 break; 1982 case SUN4V_CHIP_NIAGARA3: 1983 /* T3 supports 48-bit virtual addresses. */ 1984 sparc64_va_hole_top = 0xffff800000000000UL; 1985 sparc64_va_hole_bottom = 0x0000800000000000UL; 1986 1987 max_phys_bits = 43; 1988 break; 1989 case SUN4V_CHIP_NIAGARA4: 1990 case SUN4V_CHIP_NIAGARA5: 1991 case SUN4V_CHIP_SPARC64X: 1992 case SUN4V_CHIP_SPARC_M6: 1993 /* T4 and later support 52-bit virtual addresses. */ 1994 sparc64_va_hole_top = 0xfff8000000000000UL; 1995 sparc64_va_hole_bottom = 0x0008000000000000UL; 1996 max_phys_bits = 47; 1997 break; 1998 case SUN4V_CHIP_SPARC_M7: 1999 case SUN4V_CHIP_SPARC_SN: 2000 /* M7 and later support 52-bit virtual addresses. */ 2001 sparc64_va_hole_top = 0xfff8000000000000UL; 2002 sparc64_va_hole_bottom = 0x0008000000000000UL; 2003 max_phys_bits = 49; 2004 break; 2005 case SUN4V_CHIP_SPARC_M8: 2006 default: 2007 /* M8 and later support 54-bit virtual addresses. 2008 * However, restricting M8 and above VA bits to 53 2009 * as 4-level page table cannot support more than 2010 * 53 VA bits. 2011 */ 2012 sparc64_va_hole_top = 0xfff0000000000000UL; 2013 sparc64_va_hole_bottom = 0x0010000000000000UL; 2014 max_phys_bits = 51; 2015 break; 2016 } 2017 } 2018 2019 if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) { 2020 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n", 2021 max_phys_bits); 2022 prom_halt(); 2023 } 2024 2025 PAGE_OFFSET = sparc64_va_hole_top; 2026 VMALLOC_END = ((sparc64_va_hole_bottom >> 1) + 2027 (sparc64_va_hole_bottom >> 2)); 2028 2029 pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n", 2030 PAGE_OFFSET, max_phys_bits); 2031 pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n", 2032 VMALLOC_START, VMALLOC_END); 2033 pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n", 2034 VMEMMAP_BASE, VMEMMAP_BASE << 1); 2035 } 2036 2037 static void __init tsb_phys_patch(void) 2038 { 2039 struct tsb_ldquad_phys_patch_entry *pquad; 2040 struct tsb_phys_patch_entry *p; 2041 2042 pquad = &__tsb_ldquad_phys_patch; 2043 while (pquad < &__tsb_ldquad_phys_patch_end) { 2044 unsigned long addr = pquad->addr; 2045 2046 if (tlb_type == hypervisor) 2047 *(unsigned int *) addr = pquad->sun4v_insn; 2048 else 2049 *(unsigned int *) addr = pquad->sun4u_insn; 2050 wmb(); 2051 __asm__ __volatile__("flush %0" 2052 : /* no outputs */ 2053 : "r" (addr)); 2054 2055 pquad++; 2056 } 2057 2058 p = &__tsb_phys_patch; 2059 while (p < &__tsb_phys_patch_end) { 2060 unsigned long addr = p->addr; 2061 2062 *(unsigned int *) addr = p->insn; 2063 wmb(); 2064 __asm__ __volatile__("flush %0" 2065 : /* no outputs */ 2066 : "r" (addr)); 2067 2068 p++; 2069 } 2070 } 2071 2072 /* Don't mark as init, we give this to the Hypervisor. */ 2073 #ifndef CONFIG_DEBUG_PAGEALLOC 2074 #define NUM_KTSB_DESCR 2 2075 #else 2076 #define NUM_KTSB_DESCR 1 2077 #endif 2078 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR]; 2079 2080 /* The swapper TSBs are loaded with a base sequence of: 2081 * 2082 * sethi %uhi(SYMBOL), REG1 2083 * sethi %hi(SYMBOL), REG2 2084 * or REG1, %ulo(SYMBOL), REG1 2085 * or REG2, %lo(SYMBOL), REG2 2086 * sllx REG1, 32, REG1 2087 * or REG1, REG2, REG1 2088 * 2089 * When we use physical addressing for the TSB accesses, we patch the 2090 * first four instructions in the above sequence. 2091 */ 2092 2093 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa) 2094 { 2095 unsigned long high_bits, low_bits; 2096 2097 high_bits = (pa >> 32) & 0xffffffff; 2098 low_bits = (pa >> 0) & 0xffffffff; 2099 2100 while (start < end) { 2101 unsigned int *ia = (unsigned int *)(unsigned long)*start; 2102 2103 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10); 2104 __asm__ __volatile__("flush %0" : : "r" (ia)); 2105 2106 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10); 2107 __asm__ __volatile__("flush %0" : : "r" (ia + 1)); 2108 2109 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff); 2110 __asm__ __volatile__("flush %0" : : "r" (ia + 2)); 2111 2112 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff); 2113 __asm__ __volatile__("flush %0" : : "r" (ia + 3)); 2114 2115 start++; 2116 } 2117 } 2118 2119 static void ktsb_phys_patch(void) 2120 { 2121 extern unsigned int __swapper_tsb_phys_patch; 2122 extern unsigned int __swapper_tsb_phys_patch_end; 2123 unsigned long ktsb_pa; 2124 2125 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE); 2126 patch_one_ktsb_phys(&__swapper_tsb_phys_patch, 2127 &__swapper_tsb_phys_patch_end, ktsb_pa); 2128 #ifndef CONFIG_DEBUG_PAGEALLOC 2129 { 2130 extern unsigned int __swapper_4m_tsb_phys_patch; 2131 extern unsigned int __swapper_4m_tsb_phys_patch_end; 2132 ktsb_pa = (kern_base + 2133 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE)); 2134 patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch, 2135 &__swapper_4m_tsb_phys_patch_end, ktsb_pa); 2136 } 2137 #endif 2138 } 2139 2140 static void __init sun4v_ktsb_init(void) 2141 { 2142 unsigned long ktsb_pa; 2143 2144 /* First KTSB for PAGE_SIZE mappings. */ 2145 ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE); 2146 2147 switch (PAGE_SIZE) { 2148 case 8 * 1024: 2149 default: 2150 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K; 2151 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K; 2152 break; 2153 2154 case 64 * 1024: 2155 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K; 2156 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K; 2157 break; 2158 2159 case 512 * 1024: 2160 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K; 2161 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K; 2162 break; 2163 2164 case 4 * 1024 * 1024: 2165 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB; 2166 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB; 2167 break; 2168 } 2169 2170 ktsb_descr[0].assoc = 1; 2171 ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES; 2172 ktsb_descr[0].ctx_idx = 0; 2173 ktsb_descr[0].tsb_base = ktsb_pa; 2174 ktsb_descr[0].resv = 0; 2175 2176 #ifndef CONFIG_DEBUG_PAGEALLOC 2177 /* Second KTSB for 4MB/256MB/2GB/16GB mappings. */ 2178 ktsb_pa = (kern_base + 2179 ((unsigned long)&swapper_4m_tsb[0] - KERNBASE)); 2180 2181 ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB; 2182 ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB | 2183 HV_PGSZ_MASK_256MB | 2184 HV_PGSZ_MASK_2GB | 2185 HV_PGSZ_MASK_16GB) & 2186 cpu_pgsz_mask); 2187 ktsb_descr[1].assoc = 1; 2188 ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES; 2189 ktsb_descr[1].ctx_idx = 0; 2190 ktsb_descr[1].tsb_base = ktsb_pa; 2191 ktsb_descr[1].resv = 0; 2192 #endif 2193 } 2194 2195 void sun4v_ktsb_register(void) 2196 { 2197 unsigned long pa, ret; 2198 2199 pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE); 2200 2201 ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa); 2202 if (ret != 0) { 2203 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: " 2204 "errors with %lx\n", pa, ret); 2205 prom_halt(); 2206 } 2207 } 2208 2209 static void __init sun4u_linear_pte_xor_finalize(void) 2210 { 2211 #ifndef CONFIG_DEBUG_PAGEALLOC 2212 /* This is where we would add Panther support for 2213 * 32MB and 256MB pages. 2214 */ 2215 #endif 2216 } 2217 2218 static void __init sun4v_linear_pte_xor_finalize(void) 2219 { 2220 unsigned long pagecv_flag; 2221 2222 /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead 2223 * enables MCD error. Do not set bit 9 on M7 processor. 2224 */ 2225 switch (sun4v_chip_type) { 2226 case SUN4V_CHIP_SPARC_M7: 2227 case SUN4V_CHIP_SPARC_M8: 2228 case SUN4V_CHIP_SPARC_SN: 2229 pagecv_flag = 0x00; 2230 break; 2231 default: 2232 pagecv_flag = _PAGE_CV_4V; 2233 break; 2234 } 2235 #ifndef CONFIG_DEBUG_PAGEALLOC 2236 if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) { 2237 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^ 2238 PAGE_OFFSET; 2239 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag | 2240 _PAGE_P_4V | _PAGE_W_4V); 2241 } else { 2242 kern_linear_pte_xor[1] = kern_linear_pte_xor[0]; 2243 } 2244 2245 if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) { 2246 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^ 2247 PAGE_OFFSET; 2248 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag | 2249 _PAGE_P_4V | _PAGE_W_4V); 2250 } else { 2251 kern_linear_pte_xor[2] = kern_linear_pte_xor[1]; 2252 } 2253 2254 if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) { 2255 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^ 2256 PAGE_OFFSET; 2257 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag | 2258 _PAGE_P_4V | _PAGE_W_4V); 2259 } else { 2260 kern_linear_pte_xor[3] = kern_linear_pte_xor[2]; 2261 } 2262 #endif 2263 } 2264 2265 /* paging_init() sets up the page tables */ 2266 2267 static unsigned long last_valid_pfn; 2268 2269 static void sun4u_pgprot_init(void); 2270 static void sun4v_pgprot_init(void); 2271 2272 static phys_addr_t __init available_memory(void) 2273 { 2274 phys_addr_t available = 0ULL; 2275 phys_addr_t pa_start, pa_end; 2276 u64 i; 2277 2278 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start, 2279 &pa_end, NULL) 2280 available = available + (pa_end - pa_start); 2281 2282 return available; 2283 } 2284 2285 #define _PAGE_CACHE_4U (_PAGE_CP_4U | _PAGE_CV_4U) 2286 #define _PAGE_CACHE_4V (_PAGE_CP_4V | _PAGE_CV_4V) 2287 #define __DIRTY_BITS_4U (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U) 2288 #define __DIRTY_BITS_4V (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V) 2289 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R) 2290 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R) 2291 2292 /* We need to exclude reserved regions. This exclusion will include 2293 * vmlinux and initrd. To be more precise the initrd size could be used to 2294 * compute a new lower limit because it is freed later during initialization. 2295 */ 2296 static void __init reduce_memory(phys_addr_t limit_ram) 2297 { 2298 phys_addr_t avail_ram = available_memory(); 2299 phys_addr_t pa_start, pa_end; 2300 u64 i; 2301 2302 if (limit_ram >= avail_ram) 2303 return; 2304 2305 for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start, 2306 &pa_end, NULL) { 2307 phys_addr_t region_size = pa_end - pa_start; 2308 phys_addr_t clip_start = pa_start; 2309 2310 avail_ram = avail_ram - region_size; 2311 /* Are we consuming too much? */ 2312 if (avail_ram < limit_ram) { 2313 phys_addr_t give_back = limit_ram - avail_ram; 2314 2315 region_size = region_size - give_back; 2316 clip_start = clip_start + give_back; 2317 } 2318 2319 memblock_remove(clip_start, region_size); 2320 2321 if (avail_ram <= limit_ram) 2322 break; 2323 i = 0UL; 2324 } 2325 } 2326 2327 void __init paging_init(void) 2328 { 2329 unsigned long end_pfn, shift, phys_base; 2330 unsigned long real_end, i; 2331 2332 setup_page_offset(); 2333 2334 /* These build time checkes make sure that the dcache_dirty_cpu() 2335 * page->flags usage will work. 2336 * 2337 * When a page gets marked as dcache-dirty, we store the 2338 * cpu number starting at bit 32 in the page->flags. Also, 2339 * functions like clear_dcache_dirty_cpu use the cpu mask 2340 * in 13-bit signed-immediate instruction fields. 2341 */ 2342 2343 /* 2344 * Page flags must not reach into upper 32 bits that are used 2345 * for the cpu number 2346 */ 2347 BUILD_BUG_ON(NR_PAGEFLAGS > 32); 2348 2349 /* 2350 * The bit fields placed in the high range must not reach below 2351 * the 32 bit boundary. Otherwise we cannot place the cpu field 2352 * at the 32 bit boundary. 2353 */ 2354 BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH + 2355 ilog2(roundup_pow_of_two(NR_CPUS)) > 32); 2356 2357 BUILD_BUG_ON(NR_CPUS > 4096); 2358 2359 kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB; 2360 kern_size = (unsigned long)&_end - (unsigned long)KERNBASE; 2361 2362 /* Invalidate both kernel TSBs. */ 2363 memset(swapper_tsb, 0x40, sizeof(swapper_tsb)); 2364 #ifndef CONFIG_DEBUG_PAGEALLOC 2365 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb)); 2366 #endif 2367 2368 /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde 2369 * bit on M7 processor. This is a conflicting usage of the same 2370 * bit. Enabling TTE.cv on M7 would turn on Memory Corruption 2371 * Detection error on all pages and this will lead to problems 2372 * later. Kernel does not run with MCD enabled and hence rest 2373 * of the required steps to fully configure memory corruption 2374 * detection are not taken. We need to ensure TTE.mcde is not 2375 * set on M7 processor. Compute the value of cacheability 2376 * flag for use later taking this into consideration. 2377 */ 2378 switch (sun4v_chip_type) { 2379 case SUN4V_CHIP_SPARC_M7: 2380 case SUN4V_CHIP_SPARC_M8: 2381 case SUN4V_CHIP_SPARC_SN: 2382 page_cache4v_flag = _PAGE_CP_4V; 2383 break; 2384 default: 2385 page_cache4v_flag = _PAGE_CACHE_4V; 2386 break; 2387 } 2388 2389 if (tlb_type == hypervisor) 2390 sun4v_pgprot_init(); 2391 else 2392 sun4u_pgprot_init(); 2393 2394 if (tlb_type == cheetah_plus || 2395 tlb_type == hypervisor) { 2396 tsb_phys_patch(); 2397 ktsb_phys_patch(); 2398 } 2399 2400 if (tlb_type == hypervisor) 2401 sun4v_patch_tlb_handlers(); 2402 2403 /* Find available physical memory... 2404 * 2405 * Read it twice in order to work around a bug in openfirmware. 2406 * The call to grab this table itself can cause openfirmware to 2407 * allocate memory, which in turn can take away some space from 2408 * the list of available memory. Reading it twice makes sure 2409 * we really do get the final value. 2410 */ 2411 read_obp_translations(); 2412 read_obp_memory("reg", &pall[0], &pall_ents); 2413 read_obp_memory("available", &pavail[0], &pavail_ents); 2414 read_obp_memory("available", &pavail[0], &pavail_ents); 2415 2416 phys_base = 0xffffffffffffffffUL; 2417 for (i = 0; i < pavail_ents; i++) { 2418 phys_base = min(phys_base, pavail[i].phys_addr); 2419 memblock_add(pavail[i].phys_addr, pavail[i].reg_size); 2420 } 2421 2422 memblock_reserve(kern_base, kern_size); 2423 2424 find_ramdisk(phys_base); 2425 2426 if (cmdline_memory_size) 2427 reduce_memory(cmdline_memory_size); 2428 2429 memblock_allow_resize(); 2430 memblock_dump_all(); 2431 2432 set_bit(0, mmu_context_bmap); 2433 2434 shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE); 2435 2436 real_end = (unsigned long)_end; 2437 num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB); 2438 printk("Kernel: Using %d locked TLB entries for main kernel image.\n", 2439 num_kernel_image_mappings); 2440 2441 /* Set kernel pgd to upper alias so physical page computations 2442 * work. 2443 */ 2444 init_mm.pgd += ((shift) / (sizeof(pgd_t))); 2445 2446 memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir)); 2447 2448 inherit_prom_mappings(); 2449 2450 /* Ok, we can use our TLB miss and window trap handlers safely. */ 2451 setup_tba(); 2452 2453 __flush_tlb_all(); 2454 2455 prom_build_devicetree(); 2456 of_populate_present_mask(); 2457 #ifndef CONFIG_SMP 2458 of_fill_in_cpu_data(); 2459 #endif 2460 2461 if (tlb_type == hypervisor) { 2462 sun4v_mdesc_init(); 2463 mdesc_populate_present_mask(cpu_all_mask); 2464 #ifndef CONFIG_SMP 2465 mdesc_fill_in_cpu_data(cpu_all_mask); 2466 #endif 2467 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask); 2468 2469 sun4v_linear_pte_xor_finalize(); 2470 2471 sun4v_ktsb_init(); 2472 sun4v_ktsb_register(); 2473 } else { 2474 unsigned long impl, ver; 2475 2476 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K | 2477 HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB); 2478 2479 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver)); 2480 impl = ((ver >> 32) & 0xffff); 2481 if (impl == PANTHER_IMPL) 2482 cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB | 2483 HV_PGSZ_MASK_256MB); 2484 2485 sun4u_linear_pte_xor_finalize(); 2486 } 2487 2488 /* Flush the TLBs and the 4M TSB so that the updated linear 2489 * pte XOR settings are realized for all mappings. 2490 */ 2491 __flush_tlb_all(); 2492 #ifndef CONFIG_DEBUG_PAGEALLOC 2493 memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb)); 2494 #endif 2495 __flush_tlb_all(); 2496 2497 /* Setup bootmem... */ 2498 last_valid_pfn = end_pfn = bootmem_init(phys_base); 2499 2500 kernel_physical_mapping_init(); 2501 2502 { 2503 unsigned long max_zone_pfns[MAX_NR_ZONES]; 2504 2505 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 2506 2507 max_zone_pfns[ZONE_NORMAL] = end_pfn; 2508 2509 free_area_init_nodes(max_zone_pfns); 2510 } 2511 2512 printk("Booting Linux...\n"); 2513 } 2514 2515 int page_in_phys_avail(unsigned long paddr) 2516 { 2517 int i; 2518 2519 paddr &= PAGE_MASK; 2520 2521 for (i = 0; i < pavail_ents; i++) { 2522 unsigned long start, end; 2523 2524 start = pavail[i].phys_addr; 2525 end = start + pavail[i].reg_size; 2526 2527 if (paddr >= start && paddr < end) 2528 return 1; 2529 } 2530 if (paddr >= kern_base && paddr < (kern_base + kern_size)) 2531 return 1; 2532 #ifdef CONFIG_BLK_DEV_INITRD 2533 if (paddr >= __pa(initrd_start) && 2534 paddr < __pa(PAGE_ALIGN(initrd_end))) 2535 return 1; 2536 #endif 2537 2538 return 0; 2539 } 2540 2541 static void __init register_page_bootmem_info(void) 2542 { 2543 #ifdef CONFIG_NEED_MULTIPLE_NODES 2544 int i; 2545 2546 for_each_online_node(i) 2547 if (NODE_DATA(i)->node_spanned_pages) 2548 register_page_bootmem_info_node(NODE_DATA(i)); 2549 #endif 2550 } 2551 void __init mem_init(void) 2552 { 2553 high_memory = __va(last_valid_pfn << PAGE_SHIFT); 2554 2555 memblock_free_all(); 2556 2557 /* 2558 * Must be done after boot memory is put on freelist, because here we 2559 * might set fields in deferred struct pages that have not yet been 2560 * initialized, and memblock_free_all() initializes all the reserved 2561 * deferred pages for us. 2562 */ 2563 register_page_bootmem_info(); 2564 2565 /* 2566 * Set up the zero page, mark it reserved, so that page count 2567 * is not manipulated when freeing the page from user ptes. 2568 */ 2569 mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0); 2570 if (mem_map_zero == NULL) { 2571 prom_printf("paging_init: Cannot alloc zero page.\n"); 2572 prom_halt(); 2573 } 2574 mark_page_reserved(mem_map_zero); 2575 2576 mem_init_print_info(NULL); 2577 2578 if (tlb_type == cheetah || tlb_type == cheetah_plus) 2579 cheetah_ecache_flush_init(); 2580 } 2581 2582 void free_initmem(void) 2583 { 2584 unsigned long addr, initend; 2585 int do_free = 1; 2586 2587 /* If the physical memory maps were trimmed by kernel command 2588 * line options, don't even try freeing this initmem stuff up. 2589 * The kernel image could have been in the trimmed out region 2590 * and if so the freeing below will free invalid page structs. 2591 */ 2592 if (cmdline_memory_size) 2593 do_free = 0; 2594 2595 /* 2596 * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes. 2597 */ 2598 addr = PAGE_ALIGN((unsigned long)(__init_begin)); 2599 initend = (unsigned long)(__init_end) & PAGE_MASK; 2600 for (; addr < initend; addr += PAGE_SIZE) { 2601 unsigned long page; 2602 2603 page = (addr + 2604 ((unsigned long) __va(kern_base)) - 2605 ((unsigned long) KERNBASE)); 2606 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE); 2607 2608 if (do_free) 2609 free_reserved_page(virt_to_page(page)); 2610 } 2611 } 2612 2613 #ifdef CONFIG_BLK_DEV_INITRD 2614 void free_initrd_mem(unsigned long start, unsigned long end) 2615 { 2616 free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM, 2617 "initrd"); 2618 } 2619 #endif 2620 2621 pgprot_t PAGE_KERNEL __read_mostly; 2622 EXPORT_SYMBOL(PAGE_KERNEL); 2623 2624 pgprot_t PAGE_KERNEL_LOCKED __read_mostly; 2625 pgprot_t PAGE_COPY __read_mostly; 2626 2627 pgprot_t PAGE_SHARED __read_mostly; 2628 EXPORT_SYMBOL(PAGE_SHARED); 2629 2630 unsigned long pg_iobits __read_mostly; 2631 2632 unsigned long _PAGE_IE __read_mostly; 2633 EXPORT_SYMBOL(_PAGE_IE); 2634 2635 unsigned long _PAGE_E __read_mostly; 2636 EXPORT_SYMBOL(_PAGE_E); 2637 2638 unsigned long _PAGE_CACHE __read_mostly; 2639 EXPORT_SYMBOL(_PAGE_CACHE); 2640 2641 #ifdef CONFIG_SPARSEMEM_VMEMMAP 2642 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend, 2643 int node, struct vmem_altmap *altmap) 2644 { 2645 unsigned long pte_base; 2646 2647 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U | 2648 _PAGE_CP_4U | _PAGE_CV_4U | 2649 _PAGE_P_4U | _PAGE_W_4U); 2650 if (tlb_type == hypervisor) 2651 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V | 2652 page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V); 2653 2654 pte_base |= _PAGE_PMD_HUGE; 2655 2656 vstart = vstart & PMD_MASK; 2657 vend = ALIGN(vend, PMD_SIZE); 2658 for (; vstart < vend; vstart += PMD_SIZE) { 2659 pgd_t *pgd = vmemmap_pgd_populate(vstart, node); 2660 unsigned long pte; 2661 pud_t *pud; 2662 pmd_t *pmd; 2663 2664 if (!pgd) 2665 return -ENOMEM; 2666 2667 pud = vmemmap_pud_populate(pgd, vstart, node); 2668 if (!pud) 2669 return -ENOMEM; 2670 2671 pmd = pmd_offset(pud, vstart); 2672 pte = pmd_val(*pmd); 2673 if (!(pte & _PAGE_VALID)) { 2674 void *block = vmemmap_alloc_block(PMD_SIZE, node); 2675 2676 if (!block) 2677 return -ENOMEM; 2678 2679 pmd_val(*pmd) = pte_base | __pa(block); 2680 } 2681 } 2682 2683 return 0; 2684 } 2685 2686 void vmemmap_free(unsigned long start, unsigned long end, 2687 struct vmem_altmap *altmap) 2688 { 2689 } 2690 #endif /* CONFIG_SPARSEMEM_VMEMMAP */ 2691 2692 static void prot_init_common(unsigned long page_none, 2693 unsigned long page_shared, 2694 unsigned long page_copy, 2695 unsigned long page_readonly, 2696 unsigned long page_exec_bit) 2697 { 2698 PAGE_COPY = __pgprot(page_copy); 2699 PAGE_SHARED = __pgprot(page_shared); 2700 2701 protection_map[0x0] = __pgprot(page_none); 2702 protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit); 2703 protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit); 2704 protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit); 2705 protection_map[0x4] = __pgprot(page_readonly); 2706 protection_map[0x5] = __pgprot(page_readonly); 2707 protection_map[0x6] = __pgprot(page_copy); 2708 protection_map[0x7] = __pgprot(page_copy); 2709 protection_map[0x8] = __pgprot(page_none); 2710 protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit); 2711 protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit); 2712 protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit); 2713 protection_map[0xc] = __pgprot(page_readonly); 2714 protection_map[0xd] = __pgprot(page_readonly); 2715 protection_map[0xe] = __pgprot(page_shared); 2716 protection_map[0xf] = __pgprot(page_shared); 2717 } 2718 2719 static void __init sun4u_pgprot_init(void) 2720 { 2721 unsigned long page_none, page_shared, page_copy, page_readonly; 2722 unsigned long page_exec_bit; 2723 int i; 2724 2725 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID | 2726 _PAGE_CACHE_4U | _PAGE_P_4U | 2727 __ACCESS_BITS_4U | __DIRTY_BITS_4U | 2728 _PAGE_EXEC_4U); 2729 PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID | 2730 _PAGE_CACHE_4U | _PAGE_P_4U | 2731 __ACCESS_BITS_4U | __DIRTY_BITS_4U | 2732 _PAGE_EXEC_4U | _PAGE_L_4U); 2733 2734 _PAGE_IE = _PAGE_IE_4U; 2735 _PAGE_E = _PAGE_E_4U; 2736 _PAGE_CACHE = _PAGE_CACHE_4U; 2737 2738 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U | 2739 __ACCESS_BITS_4U | _PAGE_E_4U); 2740 2741 #ifdef CONFIG_DEBUG_PAGEALLOC 2742 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET; 2743 #else 2744 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^ 2745 PAGE_OFFSET; 2746 #endif 2747 kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U | 2748 _PAGE_P_4U | _PAGE_W_4U); 2749 2750 for (i = 1; i < 4; i++) 2751 kern_linear_pte_xor[i] = kern_linear_pte_xor[0]; 2752 2753 _PAGE_ALL_SZ_BITS = (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U | 2754 _PAGE_SZ64K_4U | _PAGE_SZ8K_4U | 2755 _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U); 2756 2757 2758 page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U; 2759 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2760 __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U); 2761 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2762 __ACCESS_BITS_4U | _PAGE_EXEC_4U); 2763 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U | 2764 __ACCESS_BITS_4U | _PAGE_EXEC_4U); 2765 2766 page_exec_bit = _PAGE_EXEC_4U; 2767 2768 prot_init_common(page_none, page_shared, page_copy, page_readonly, 2769 page_exec_bit); 2770 } 2771 2772 static void __init sun4v_pgprot_init(void) 2773 { 2774 unsigned long page_none, page_shared, page_copy, page_readonly; 2775 unsigned long page_exec_bit; 2776 int i; 2777 2778 PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID | 2779 page_cache4v_flag | _PAGE_P_4V | 2780 __ACCESS_BITS_4V | __DIRTY_BITS_4V | 2781 _PAGE_EXEC_4V); 2782 PAGE_KERNEL_LOCKED = PAGE_KERNEL; 2783 2784 _PAGE_IE = _PAGE_IE_4V; 2785 _PAGE_E = _PAGE_E_4V; 2786 _PAGE_CACHE = page_cache4v_flag; 2787 2788 #ifdef CONFIG_DEBUG_PAGEALLOC 2789 kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET; 2790 #else 2791 kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^ 2792 PAGE_OFFSET; 2793 #endif 2794 kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V | 2795 _PAGE_W_4V); 2796 2797 for (i = 1; i < 4; i++) 2798 kern_linear_pte_xor[i] = kern_linear_pte_xor[0]; 2799 2800 pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V | 2801 __ACCESS_BITS_4V | _PAGE_E_4V); 2802 2803 _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V | 2804 _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V | 2805 _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V | 2806 _PAGE_SZ64K_4V | _PAGE_SZ8K_4V); 2807 2808 page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag; 2809 page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2810 __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V); 2811 page_copy = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2812 __ACCESS_BITS_4V | _PAGE_EXEC_4V); 2813 page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag | 2814 __ACCESS_BITS_4V | _PAGE_EXEC_4V); 2815 2816 page_exec_bit = _PAGE_EXEC_4V; 2817 2818 prot_init_common(page_none, page_shared, page_copy, page_readonly, 2819 page_exec_bit); 2820 } 2821 2822 unsigned long pte_sz_bits(unsigned long sz) 2823 { 2824 if (tlb_type == hypervisor) { 2825 switch (sz) { 2826 case 8 * 1024: 2827 default: 2828 return _PAGE_SZ8K_4V; 2829 case 64 * 1024: 2830 return _PAGE_SZ64K_4V; 2831 case 512 * 1024: 2832 return _PAGE_SZ512K_4V; 2833 case 4 * 1024 * 1024: 2834 return _PAGE_SZ4MB_4V; 2835 } 2836 } else { 2837 switch (sz) { 2838 case 8 * 1024: 2839 default: 2840 return _PAGE_SZ8K_4U; 2841 case 64 * 1024: 2842 return _PAGE_SZ64K_4U; 2843 case 512 * 1024: 2844 return _PAGE_SZ512K_4U; 2845 case 4 * 1024 * 1024: 2846 return _PAGE_SZ4MB_4U; 2847 } 2848 } 2849 } 2850 2851 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size) 2852 { 2853 pte_t pte; 2854 2855 pte_val(pte) = page | pgprot_val(pgprot_noncached(prot)); 2856 pte_val(pte) |= (((unsigned long)space) << 32); 2857 pte_val(pte) |= pte_sz_bits(page_size); 2858 2859 return pte; 2860 } 2861 2862 static unsigned long kern_large_tte(unsigned long paddr) 2863 { 2864 unsigned long val; 2865 2866 val = (_PAGE_VALID | _PAGE_SZ4MB_4U | 2867 _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U | 2868 _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U); 2869 if (tlb_type == hypervisor) 2870 val = (_PAGE_VALID | _PAGE_SZ4MB_4V | 2871 page_cache4v_flag | _PAGE_P_4V | 2872 _PAGE_EXEC_4V | _PAGE_W_4V); 2873 2874 return val | paddr; 2875 } 2876 2877 /* If not locked, zap it. */ 2878 void __flush_tlb_all(void) 2879 { 2880 unsigned long pstate; 2881 int i; 2882 2883 __asm__ __volatile__("flushw\n\t" 2884 "rdpr %%pstate, %0\n\t" 2885 "wrpr %0, %1, %%pstate" 2886 : "=r" (pstate) 2887 : "i" (PSTATE_IE)); 2888 if (tlb_type == hypervisor) { 2889 sun4v_mmu_demap_all(); 2890 } else if (tlb_type == spitfire) { 2891 for (i = 0; i < 64; i++) { 2892 /* Spitfire Errata #32 workaround */ 2893 /* NOTE: Always runs on spitfire, so no 2894 * cheetah+ page size encodings. 2895 */ 2896 __asm__ __volatile__("stxa %0, [%1] %2\n\t" 2897 "flush %%g6" 2898 : /* No outputs */ 2899 : "r" (0), 2900 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU)); 2901 2902 if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) { 2903 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 2904 "membar #Sync" 2905 : /* no outputs */ 2906 : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU)); 2907 spitfire_put_dtlb_data(i, 0x0UL); 2908 } 2909 2910 /* Spitfire Errata #32 workaround */ 2911 /* NOTE: Always runs on spitfire, so no 2912 * cheetah+ page size encodings. 2913 */ 2914 __asm__ __volatile__("stxa %0, [%1] %2\n\t" 2915 "flush %%g6" 2916 : /* No outputs */ 2917 : "r" (0), 2918 "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU)); 2919 2920 if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) { 2921 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t" 2922 "membar #Sync" 2923 : /* no outputs */ 2924 : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU)); 2925 spitfire_put_itlb_data(i, 0x0UL); 2926 } 2927 } 2928 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) { 2929 cheetah_flush_dtlb_all(); 2930 cheetah_flush_itlb_all(); 2931 } 2932 __asm__ __volatile__("wrpr %0, 0, %%pstate" 2933 : : "r" (pstate)); 2934 } 2935 2936 pte_t *pte_alloc_one_kernel(struct mm_struct *mm) 2937 { 2938 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO); 2939 pte_t *pte = NULL; 2940 2941 if (page) 2942 pte = (pte_t *) page_address(page); 2943 2944 return pte; 2945 } 2946 2947 pgtable_t pte_alloc_one(struct mm_struct *mm) 2948 { 2949 struct page *page = alloc_page(GFP_KERNEL | __GFP_ZERO); 2950 if (!page) 2951 return NULL; 2952 if (!pgtable_page_ctor(page)) { 2953 free_unref_page(page); 2954 return NULL; 2955 } 2956 return (pte_t *) page_address(page); 2957 } 2958 2959 void pte_free_kernel(struct mm_struct *mm, pte_t *pte) 2960 { 2961 free_page((unsigned long)pte); 2962 } 2963 2964 static void __pte_free(pgtable_t pte) 2965 { 2966 struct page *page = virt_to_page(pte); 2967 2968 pgtable_page_dtor(page); 2969 __free_page(page); 2970 } 2971 2972 void pte_free(struct mm_struct *mm, pgtable_t pte) 2973 { 2974 __pte_free(pte); 2975 } 2976 2977 void pgtable_free(void *table, bool is_page) 2978 { 2979 if (is_page) 2980 __pte_free(table); 2981 else 2982 kmem_cache_free(pgtable_cache, table); 2983 } 2984 2985 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 2986 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr, 2987 pmd_t *pmd) 2988 { 2989 unsigned long pte, flags; 2990 struct mm_struct *mm; 2991 pmd_t entry = *pmd; 2992 2993 if (!pmd_large(entry) || !pmd_young(entry)) 2994 return; 2995 2996 pte = pmd_val(entry); 2997 2998 /* Don't insert a non-valid PMD into the TSB, we'll deadlock. */ 2999 if (!(pte & _PAGE_VALID)) 3000 return; 3001 3002 /* We are fabricating 8MB pages using 4MB real hw pages. */ 3003 pte |= (addr & (1UL << REAL_HPAGE_SHIFT)); 3004 3005 mm = vma->vm_mm; 3006 3007 spin_lock_irqsave(&mm->context.lock, flags); 3008 3009 if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) 3010 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT, 3011 addr, pte); 3012 3013 spin_unlock_irqrestore(&mm->context.lock, flags); 3014 } 3015 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 3016 3017 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE) 3018 static void context_reload(void *__data) 3019 { 3020 struct mm_struct *mm = __data; 3021 3022 if (mm == current->mm) 3023 load_secondary_context(mm); 3024 } 3025 3026 void hugetlb_setup(struct pt_regs *regs) 3027 { 3028 struct mm_struct *mm = current->mm; 3029 struct tsb_config *tp; 3030 3031 if (faulthandler_disabled() || !mm) { 3032 const struct exception_table_entry *entry; 3033 3034 entry = search_exception_tables(regs->tpc); 3035 if (entry) { 3036 regs->tpc = entry->fixup; 3037 regs->tnpc = regs->tpc + 4; 3038 return; 3039 } 3040 pr_alert("Unexpected HugeTLB setup in atomic context.\n"); 3041 die_if_kernel("HugeTSB in atomic", regs); 3042 } 3043 3044 tp = &mm->context.tsb_block[MM_TSB_HUGE]; 3045 if (likely(tp->tsb == NULL)) 3046 tsb_grow(mm, MM_TSB_HUGE, 0); 3047 3048 tsb_context_switch(mm); 3049 smp_tsb_sync(mm); 3050 3051 /* On UltraSPARC-III+ and later, configure the second half of 3052 * the Data-TLB for huge pages. 3053 */ 3054 if (tlb_type == cheetah_plus) { 3055 bool need_context_reload = false; 3056 unsigned long ctx; 3057 3058 spin_lock_irq(&ctx_alloc_lock); 3059 ctx = mm->context.sparc64_ctx_val; 3060 ctx &= ~CTX_PGSZ_MASK; 3061 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT; 3062 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT; 3063 3064 if (ctx != mm->context.sparc64_ctx_val) { 3065 /* When changing the page size fields, we 3066 * must perform a context flush so that no 3067 * stale entries match. This flush must 3068 * occur with the original context register 3069 * settings. 3070 */ 3071 do_flush_tlb_mm(mm); 3072 3073 /* Reload the context register of all processors 3074 * also executing in this address space. 3075 */ 3076 mm->context.sparc64_ctx_val = ctx; 3077 need_context_reload = true; 3078 } 3079 spin_unlock_irq(&ctx_alloc_lock); 3080 3081 if (need_context_reload) 3082 on_each_cpu(context_reload, mm, 0); 3083 } 3084 } 3085 #endif 3086 3087 static struct resource code_resource = { 3088 .name = "Kernel code", 3089 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3090 }; 3091 3092 static struct resource data_resource = { 3093 .name = "Kernel data", 3094 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3095 }; 3096 3097 static struct resource bss_resource = { 3098 .name = "Kernel bss", 3099 .flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM 3100 }; 3101 3102 static inline resource_size_t compute_kern_paddr(void *addr) 3103 { 3104 return (resource_size_t) (addr - KERNBASE + kern_base); 3105 } 3106 3107 static void __init kernel_lds_init(void) 3108 { 3109 code_resource.start = compute_kern_paddr(_text); 3110 code_resource.end = compute_kern_paddr(_etext - 1); 3111 data_resource.start = compute_kern_paddr(_etext); 3112 data_resource.end = compute_kern_paddr(_edata - 1); 3113 bss_resource.start = compute_kern_paddr(__bss_start); 3114 bss_resource.end = compute_kern_paddr(_end - 1); 3115 } 3116 3117 static int __init report_memory(void) 3118 { 3119 int i; 3120 struct resource *res; 3121 3122 kernel_lds_init(); 3123 3124 for (i = 0; i < pavail_ents; i++) { 3125 res = kzalloc(sizeof(struct resource), GFP_KERNEL); 3126 3127 if (!res) { 3128 pr_warn("Failed to allocate source.\n"); 3129 break; 3130 } 3131 3132 res->name = "System RAM"; 3133 res->start = pavail[i].phys_addr; 3134 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1; 3135 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM; 3136 3137 if (insert_resource(&iomem_resource, res) < 0) { 3138 pr_warn("Resource insertion failed.\n"); 3139 break; 3140 } 3141 3142 insert_resource(res, &code_resource); 3143 insert_resource(res, &data_resource); 3144 insert_resource(res, &bss_resource); 3145 } 3146 3147 return 0; 3148 } 3149 arch_initcall(report_memory); 3150 3151 #ifdef CONFIG_SMP 3152 #define do_flush_tlb_kernel_range smp_flush_tlb_kernel_range 3153 #else 3154 #define do_flush_tlb_kernel_range __flush_tlb_kernel_range 3155 #endif 3156 3157 void flush_tlb_kernel_range(unsigned long start, unsigned long end) 3158 { 3159 if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) { 3160 if (start < LOW_OBP_ADDRESS) { 3161 flush_tsb_kernel_range(start, LOW_OBP_ADDRESS); 3162 do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS); 3163 } 3164 if (end > HI_OBP_ADDRESS) { 3165 flush_tsb_kernel_range(HI_OBP_ADDRESS, end); 3166 do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end); 3167 } 3168 } else { 3169 flush_tsb_kernel_range(start, end); 3170 do_flush_tlb_kernel_range(start, end); 3171 } 3172 } 3173 3174 void copy_user_highpage(struct page *to, struct page *from, 3175 unsigned long vaddr, struct vm_area_struct *vma) 3176 { 3177 char *vfrom, *vto; 3178 3179 vfrom = kmap_atomic(from); 3180 vto = kmap_atomic(to); 3181 copy_user_page(vto, vfrom, vaddr, to); 3182 kunmap_atomic(vto); 3183 kunmap_atomic(vfrom); 3184 3185 /* If this page has ADI enabled, copy over any ADI tags 3186 * as well 3187 */ 3188 if (vma->vm_flags & VM_SPARC_ADI) { 3189 unsigned long pfrom, pto, i, adi_tag; 3190 3191 pfrom = page_to_phys(from); 3192 pto = page_to_phys(to); 3193 3194 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) { 3195 asm volatile("ldxa [%1] %2, %0\n\t" 3196 : "=r" (adi_tag) 3197 : "r" (i), "i" (ASI_MCD_REAL)); 3198 asm volatile("stxa %0, [%1] %2\n\t" 3199 : 3200 : "r" (adi_tag), "r" (pto), 3201 "i" (ASI_MCD_REAL)); 3202 pto += adi_blksize(); 3203 } 3204 asm volatile("membar #Sync\n\t"); 3205 } 3206 } 3207 EXPORT_SYMBOL(copy_user_highpage); 3208 3209 void copy_highpage(struct page *to, struct page *from) 3210 { 3211 char *vfrom, *vto; 3212 3213 vfrom = kmap_atomic(from); 3214 vto = kmap_atomic(to); 3215 copy_page(vto, vfrom); 3216 kunmap_atomic(vto); 3217 kunmap_atomic(vfrom); 3218 3219 /* If this platform is ADI enabled, copy any ADI tags 3220 * as well 3221 */ 3222 if (adi_capable()) { 3223 unsigned long pfrom, pto, i, adi_tag; 3224 3225 pfrom = page_to_phys(from); 3226 pto = page_to_phys(to); 3227 3228 for (i = pfrom; i < (pfrom + PAGE_SIZE); i += adi_blksize()) { 3229 asm volatile("ldxa [%1] %2, %0\n\t" 3230 : "=r" (adi_tag) 3231 : "r" (i), "i" (ASI_MCD_REAL)); 3232 asm volatile("stxa %0, [%1] %2\n\t" 3233 : 3234 : "r" (adi_tag), "r" (pto), 3235 "i" (ASI_MCD_REAL)); 3236 pto += adi_blksize(); 3237 } 3238 asm volatile("membar #Sync\n\t"); 3239 } 3240 } 3241 EXPORT_SYMBOL(copy_highpage); 3242