1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * SPARC64 Huge TLB page support. 4 * 5 * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net) 6 */ 7 8 #include <linux/fs.h> 9 #include <linux/mm.h> 10 #include <linux/sched/mm.h> 11 #include <linux/hugetlb.h> 12 #include <linux/pagemap.h> 13 #include <linux/sysctl.h> 14 15 #include <asm/mman.h> 16 #include <asm/pgalloc.h> 17 #include <asm/tlb.h> 18 #include <asm/tlbflush.h> 19 #include <asm/cacheflush.h> 20 #include <asm/mmu_context.h> 21 22 /* Slightly simplified from the non-hugepage variant because by 23 * definition we don't have to worry about any page coloring stuff 24 */ 25 26 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp, 27 unsigned long addr, 28 unsigned long len, 29 unsigned long pgoff, 30 unsigned long flags) 31 { 32 struct hstate *h = hstate_file(filp); 33 unsigned long task_size = TASK_SIZE; 34 struct vm_unmapped_area_info info; 35 36 if (test_thread_flag(TIF_32BIT)) 37 task_size = STACK_TOP32; 38 39 info.flags = 0; 40 info.length = len; 41 info.low_limit = TASK_UNMAPPED_BASE; 42 info.high_limit = min(task_size, VA_EXCLUDE_START); 43 info.align_mask = PAGE_MASK & ~huge_page_mask(h); 44 info.align_offset = 0; 45 addr = vm_unmapped_area(&info); 46 47 if ((addr & ~PAGE_MASK) && task_size > VA_EXCLUDE_END) { 48 VM_BUG_ON(addr != -ENOMEM); 49 info.low_limit = VA_EXCLUDE_END; 50 info.high_limit = task_size; 51 addr = vm_unmapped_area(&info); 52 } 53 54 return addr; 55 } 56 57 static unsigned long 58 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, 59 const unsigned long len, 60 const unsigned long pgoff, 61 const unsigned long flags) 62 { 63 struct hstate *h = hstate_file(filp); 64 struct mm_struct *mm = current->mm; 65 unsigned long addr = addr0; 66 struct vm_unmapped_area_info info; 67 68 /* This should only ever run for 32-bit processes. */ 69 BUG_ON(!test_thread_flag(TIF_32BIT)); 70 71 info.flags = VM_UNMAPPED_AREA_TOPDOWN; 72 info.length = len; 73 info.low_limit = PAGE_SIZE; 74 info.high_limit = mm->mmap_base; 75 info.align_mask = PAGE_MASK & ~huge_page_mask(h); 76 info.align_offset = 0; 77 addr = vm_unmapped_area(&info); 78 79 /* 80 * A failed mmap() very likely causes application failure, 81 * so fall back to the bottom-up function here. This scenario 82 * can happen with large stack limits and large mmap() 83 * allocations. 84 */ 85 if (addr & ~PAGE_MASK) { 86 VM_BUG_ON(addr != -ENOMEM); 87 info.flags = 0; 88 info.low_limit = TASK_UNMAPPED_BASE; 89 info.high_limit = STACK_TOP32; 90 addr = vm_unmapped_area(&info); 91 } 92 93 return addr; 94 } 95 96 unsigned long 97 hugetlb_get_unmapped_area(struct file *file, unsigned long addr, 98 unsigned long len, unsigned long pgoff, unsigned long flags) 99 { 100 struct hstate *h = hstate_file(file); 101 struct mm_struct *mm = current->mm; 102 struct vm_area_struct *vma; 103 unsigned long task_size = TASK_SIZE; 104 105 if (test_thread_flag(TIF_32BIT)) 106 task_size = STACK_TOP32; 107 108 if (len & ~huge_page_mask(h)) 109 return -EINVAL; 110 if (len > task_size) 111 return -ENOMEM; 112 113 if (flags & MAP_FIXED) { 114 if (prepare_hugepage_range(file, addr, len)) 115 return -EINVAL; 116 return addr; 117 } 118 119 if (addr) { 120 addr = ALIGN(addr, huge_page_size(h)); 121 vma = find_vma(mm, addr); 122 if (task_size - len >= addr && 123 (!vma || addr + len <= vm_start_gap(vma))) 124 return addr; 125 } 126 if (mm->get_unmapped_area == arch_get_unmapped_area) 127 return hugetlb_get_unmapped_area_bottomup(file, addr, len, 128 pgoff, flags); 129 else 130 return hugetlb_get_unmapped_area_topdown(file, addr, len, 131 pgoff, flags); 132 } 133 134 static pte_t sun4u_hugepage_shift_to_tte(pte_t entry, unsigned int shift) 135 { 136 return entry; 137 } 138 139 static pte_t sun4v_hugepage_shift_to_tte(pte_t entry, unsigned int shift) 140 { 141 unsigned long hugepage_size = _PAGE_SZ4MB_4V; 142 143 pte_val(entry) = pte_val(entry) & ~_PAGE_SZALL_4V; 144 145 switch (shift) { 146 case HPAGE_16GB_SHIFT: 147 hugepage_size = _PAGE_SZ16GB_4V; 148 pte_val(entry) |= _PAGE_PUD_HUGE; 149 break; 150 case HPAGE_2GB_SHIFT: 151 hugepage_size = _PAGE_SZ2GB_4V; 152 pte_val(entry) |= _PAGE_PMD_HUGE; 153 break; 154 case HPAGE_256MB_SHIFT: 155 hugepage_size = _PAGE_SZ256MB_4V; 156 pte_val(entry) |= _PAGE_PMD_HUGE; 157 break; 158 case HPAGE_SHIFT: 159 pte_val(entry) |= _PAGE_PMD_HUGE; 160 break; 161 case HPAGE_64K_SHIFT: 162 hugepage_size = _PAGE_SZ64K_4V; 163 break; 164 default: 165 WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift); 166 } 167 168 pte_val(entry) = pte_val(entry) | hugepage_size; 169 return entry; 170 } 171 172 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift) 173 { 174 if (tlb_type == hypervisor) 175 return sun4v_hugepage_shift_to_tte(entry, shift); 176 else 177 return sun4u_hugepage_shift_to_tte(entry, shift); 178 } 179 180 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma, 181 struct page *page, int writeable) 182 { 183 unsigned int shift = huge_page_shift(hstate_vma(vma)); 184 pte_t pte; 185 186 pte = hugepage_shift_to_tte(entry, shift); 187 188 #ifdef CONFIG_SPARC64 189 /* If this vma has ADI enabled on it, turn on TTE.mcd 190 */ 191 if (vma->vm_flags & VM_SPARC_ADI) 192 return pte_mkmcd(pte); 193 else 194 return pte_mknotmcd(pte); 195 #else 196 return pte; 197 #endif 198 } 199 200 static unsigned int sun4v_huge_tte_to_shift(pte_t entry) 201 { 202 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V; 203 unsigned int shift; 204 205 switch (tte_szbits) { 206 case _PAGE_SZ16GB_4V: 207 shift = HPAGE_16GB_SHIFT; 208 break; 209 case _PAGE_SZ2GB_4V: 210 shift = HPAGE_2GB_SHIFT; 211 break; 212 case _PAGE_SZ256MB_4V: 213 shift = HPAGE_256MB_SHIFT; 214 break; 215 case _PAGE_SZ4MB_4V: 216 shift = REAL_HPAGE_SHIFT; 217 break; 218 case _PAGE_SZ64K_4V: 219 shift = HPAGE_64K_SHIFT; 220 break; 221 default: 222 shift = PAGE_SHIFT; 223 break; 224 } 225 return shift; 226 } 227 228 static unsigned int sun4u_huge_tte_to_shift(pte_t entry) 229 { 230 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U; 231 unsigned int shift; 232 233 switch (tte_szbits) { 234 case _PAGE_SZ256MB_4U: 235 shift = HPAGE_256MB_SHIFT; 236 break; 237 case _PAGE_SZ4MB_4U: 238 shift = REAL_HPAGE_SHIFT; 239 break; 240 case _PAGE_SZ64K_4U: 241 shift = HPAGE_64K_SHIFT; 242 break; 243 default: 244 shift = PAGE_SHIFT; 245 break; 246 } 247 return shift; 248 } 249 250 static unsigned long tte_to_shift(pte_t entry) 251 { 252 if (tlb_type == hypervisor) 253 return sun4v_huge_tte_to_shift(entry); 254 255 return sun4u_huge_tte_to_shift(entry); 256 } 257 258 static unsigned int huge_tte_to_shift(pte_t entry) 259 { 260 unsigned long shift = tte_to_shift(entry); 261 262 if (shift == PAGE_SHIFT) 263 WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n", 264 pte_val(entry)); 265 266 return shift; 267 } 268 269 static unsigned long huge_tte_to_size(pte_t pte) 270 { 271 unsigned long size = 1UL << huge_tte_to_shift(pte); 272 273 if (size == REAL_HPAGE_SIZE) 274 size = HPAGE_SIZE; 275 return size; 276 } 277 278 unsigned long pud_leaf_size(pud_t pud) { return 1UL << tte_to_shift(*(pte_t *)&pud); } 279 unsigned long pmd_leaf_size(pmd_t pmd) { return 1UL << tte_to_shift(*(pte_t *)&pmd); } 280 unsigned long pte_leaf_size(pte_t pte) { return 1UL << tte_to_shift(pte); } 281 282 pte_t *huge_pte_alloc(struct mm_struct *mm, 283 unsigned long addr, unsigned long sz) 284 { 285 pgd_t *pgd; 286 p4d_t *p4d; 287 pud_t *pud; 288 pmd_t *pmd; 289 290 pgd = pgd_offset(mm, addr); 291 p4d = p4d_offset(pgd, addr); 292 pud = pud_alloc(mm, p4d, addr); 293 if (!pud) 294 return NULL; 295 if (sz >= PUD_SIZE) 296 return (pte_t *)pud; 297 pmd = pmd_alloc(mm, pud, addr); 298 if (!pmd) 299 return NULL; 300 if (sz >= PMD_SIZE) 301 return (pte_t *)pmd; 302 return pte_alloc_map(mm, pmd, addr); 303 } 304 305 pte_t *huge_pte_offset(struct mm_struct *mm, 306 unsigned long addr, unsigned long sz) 307 { 308 pgd_t *pgd; 309 p4d_t *p4d; 310 pud_t *pud; 311 pmd_t *pmd; 312 313 pgd = pgd_offset(mm, addr); 314 if (pgd_none(*pgd)) 315 return NULL; 316 p4d = p4d_offset(pgd, addr); 317 if (p4d_none(*p4d)) 318 return NULL; 319 pud = pud_offset(p4d, addr); 320 if (pud_none(*pud)) 321 return NULL; 322 if (is_hugetlb_pud(*pud)) 323 return (pte_t *)pud; 324 pmd = pmd_offset(pud, addr); 325 if (pmd_none(*pmd)) 326 return NULL; 327 if (is_hugetlb_pmd(*pmd)) 328 return (pte_t *)pmd; 329 return pte_offset_map(pmd, addr); 330 } 331 332 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, 333 pte_t *ptep, pte_t entry) 334 { 335 unsigned int nptes, orig_shift, shift; 336 unsigned long i, size; 337 pte_t orig; 338 339 size = huge_tte_to_size(entry); 340 341 shift = PAGE_SHIFT; 342 if (size >= PUD_SIZE) 343 shift = PUD_SHIFT; 344 else if (size >= PMD_SIZE) 345 shift = PMD_SHIFT; 346 else 347 shift = PAGE_SHIFT; 348 349 nptes = size >> shift; 350 351 if (!pte_present(*ptep) && pte_present(entry)) 352 mm->context.hugetlb_pte_count += nptes; 353 354 addr &= ~(size - 1); 355 orig = *ptep; 356 orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig); 357 358 for (i = 0; i < nptes; i++) 359 ptep[i] = __pte(pte_val(entry) + (i << shift)); 360 361 maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift); 362 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 363 if (size == HPAGE_SIZE) 364 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0, 365 orig_shift); 366 } 367 368 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, 369 pte_t *ptep) 370 { 371 unsigned int i, nptes, orig_shift, shift; 372 unsigned long size; 373 pte_t entry; 374 375 entry = *ptep; 376 size = huge_tte_to_size(entry); 377 378 shift = PAGE_SHIFT; 379 if (size >= PUD_SIZE) 380 shift = PUD_SHIFT; 381 else if (size >= PMD_SIZE) 382 shift = PMD_SHIFT; 383 else 384 shift = PAGE_SHIFT; 385 386 nptes = size >> shift; 387 orig_shift = pte_none(entry) ? PAGE_SHIFT : huge_tte_to_shift(entry); 388 389 if (pte_present(entry)) 390 mm->context.hugetlb_pte_count -= nptes; 391 392 addr &= ~(size - 1); 393 for (i = 0; i < nptes; i++) 394 ptep[i] = __pte(0UL); 395 396 maybe_tlb_batch_add(mm, addr, ptep, entry, 0, orig_shift); 397 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 398 if (size == HPAGE_SIZE) 399 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0, 400 orig_shift); 401 402 return entry; 403 } 404 405 int pmd_huge(pmd_t pmd) 406 { 407 return !pmd_none(pmd) && 408 (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID; 409 } 410 411 int pud_huge(pud_t pud) 412 { 413 return !pud_none(pud) && 414 (pud_val(pud) & (_PAGE_VALID|_PAGE_PUD_HUGE)) != _PAGE_VALID; 415 } 416 417 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 418 unsigned long addr) 419 { 420 pgtable_t token = pmd_pgtable(*pmd); 421 422 pmd_clear(pmd); 423 pte_free_tlb(tlb, token, addr); 424 mm_dec_nr_ptes(tlb->mm); 425 } 426 427 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 428 unsigned long addr, unsigned long end, 429 unsigned long floor, unsigned long ceiling) 430 { 431 pmd_t *pmd; 432 unsigned long next; 433 unsigned long start; 434 435 start = addr; 436 pmd = pmd_offset(pud, addr); 437 do { 438 next = pmd_addr_end(addr, end); 439 if (pmd_none(*pmd)) 440 continue; 441 if (is_hugetlb_pmd(*pmd)) 442 pmd_clear(pmd); 443 else 444 hugetlb_free_pte_range(tlb, pmd, addr); 445 } while (pmd++, addr = next, addr != end); 446 447 start &= PUD_MASK; 448 if (start < floor) 449 return; 450 if (ceiling) { 451 ceiling &= PUD_MASK; 452 if (!ceiling) 453 return; 454 } 455 if (end - 1 > ceiling - 1) 456 return; 457 458 pmd = pmd_offset(pud, start); 459 pud_clear(pud); 460 pmd_free_tlb(tlb, pmd, start); 461 mm_dec_nr_pmds(tlb->mm); 462 } 463 464 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d, 465 unsigned long addr, unsigned long end, 466 unsigned long floor, unsigned long ceiling) 467 { 468 pud_t *pud; 469 unsigned long next; 470 unsigned long start; 471 472 start = addr; 473 pud = pud_offset(p4d, addr); 474 do { 475 next = pud_addr_end(addr, end); 476 if (pud_none_or_clear_bad(pud)) 477 continue; 478 if (is_hugetlb_pud(*pud)) 479 pud_clear(pud); 480 else 481 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, 482 ceiling); 483 } while (pud++, addr = next, addr != end); 484 485 start &= PGDIR_MASK; 486 if (start < floor) 487 return; 488 if (ceiling) { 489 ceiling &= PGDIR_MASK; 490 if (!ceiling) 491 return; 492 } 493 if (end - 1 > ceiling - 1) 494 return; 495 496 pud = pud_offset(p4d, start); 497 p4d_clear(p4d); 498 pud_free_tlb(tlb, pud, start); 499 mm_dec_nr_puds(tlb->mm); 500 } 501 502 void hugetlb_free_pgd_range(struct mmu_gather *tlb, 503 unsigned long addr, unsigned long end, 504 unsigned long floor, unsigned long ceiling) 505 { 506 pgd_t *pgd; 507 p4d_t *p4d; 508 unsigned long next; 509 510 addr &= PMD_MASK; 511 if (addr < floor) { 512 addr += PMD_SIZE; 513 if (!addr) 514 return; 515 } 516 if (ceiling) { 517 ceiling &= PMD_MASK; 518 if (!ceiling) 519 return; 520 } 521 if (end - 1 > ceiling - 1) 522 end -= PMD_SIZE; 523 if (addr > end - 1) 524 return; 525 526 pgd = pgd_offset(tlb->mm, addr); 527 p4d = p4d_offset(pgd, addr); 528 do { 529 next = p4d_addr_end(addr, end); 530 if (p4d_none_or_clear_bad(p4d)) 531 continue; 532 hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling); 533 } while (p4d++, addr = next, addr != end); 534 } 535