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