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 pte_t pte; 186 187 pte = hugepage_shift_to_tte(entry, shift); 188 189 #ifdef CONFIG_SPARC64 190 /* If this vma has ADI enabled on it, turn on TTE.mcd 191 */ 192 if (vma->vm_flags & VM_SPARC_ADI) 193 return pte_mkmcd(pte); 194 else 195 return pte_mknotmcd(pte); 196 #else 197 return pte; 198 #endif 199 } 200 201 static unsigned int sun4v_huge_tte_to_shift(pte_t entry) 202 { 203 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V; 204 unsigned int shift; 205 206 switch (tte_szbits) { 207 case _PAGE_SZ16GB_4V: 208 shift = HPAGE_16GB_SHIFT; 209 break; 210 case _PAGE_SZ2GB_4V: 211 shift = HPAGE_2GB_SHIFT; 212 break; 213 case _PAGE_SZ256MB_4V: 214 shift = HPAGE_256MB_SHIFT; 215 break; 216 case _PAGE_SZ4MB_4V: 217 shift = REAL_HPAGE_SHIFT; 218 break; 219 case _PAGE_SZ64K_4V: 220 shift = HPAGE_64K_SHIFT; 221 break; 222 default: 223 shift = PAGE_SHIFT; 224 break; 225 } 226 return shift; 227 } 228 229 static unsigned int sun4u_huge_tte_to_shift(pte_t entry) 230 { 231 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U; 232 unsigned int shift; 233 234 switch (tte_szbits) { 235 case _PAGE_SZ256MB_4U: 236 shift = HPAGE_256MB_SHIFT; 237 break; 238 case _PAGE_SZ4MB_4U: 239 shift = REAL_HPAGE_SHIFT; 240 break; 241 case _PAGE_SZ64K_4U: 242 shift = HPAGE_64K_SHIFT; 243 break; 244 default: 245 shift = PAGE_SHIFT; 246 break; 247 } 248 return shift; 249 } 250 251 static unsigned int huge_tte_to_shift(pte_t entry) 252 { 253 unsigned long shift; 254 255 if (tlb_type == hypervisor) 256 shift = sun4v_huge_tte_to_shift(entry); 257 else 258 shift = sun4u_huge_tte_to_shift(entry); 259 260 if (shift == PAGE_SHIFT) 261 WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n", 262 pte_val(entry)); 263 264 return shift; 265 } 266 267 static unsigned long huge_tte_to_size(pte_t pte) 268 { 269 unsigned long size = 1UL << huge_tte_to_shift(pte); 270 271 if (size == REAL_HPAGE_SIZE) 272 size = HPAGE_SIZE; 273 return size; 274 } 275 276 pte_t *huge_pte_alloc(struct mm_struct *mm, 277 unsigned long addr, unsigned long sz) 278 { 279 pgd_t *pgd; 280 pud_t *pud; 281 pmd_t *pmd; 282 283 pgd = pgd_offset(mm, addr); 284 pud = pud_alloc(mm, pgd, addr); 285 if (!pud) 286 return NULL; 287 if (sz >= PUD_SIZE) 288 return (pte_t *)pud; 289 pmd = pmd_alloc(mm, pud, addr); 290 if (!pmd) 291 return NULL; 292 if (sz >= PMD_SIZE) 293 return (pte_t *)pmd; 294 return pte_alloc_map(mm, pmd, addr); 295 } 296 297 pte_t *huge_pte_offset(struct mm_struct *mm, 298 unsigned long addr, unsigned long sz) 299 { 300 pgd_t *pgd; 301 pud_t *pud; 302 pmd_t *pmd; 303 304 pgd = pgd_offset(mm, addr); 305 if (pgd_none(*pgd)) 306 return NULL; 307 pud = pud_offset(pgd, addr); 308 if (pud_none(*pud)) 309 return NULL; 310 if (is_hugetlb_pud(*pud)) 311 return (pte_t *)pud; 312 pmd = pmd_offset(pud, addr); 313 if (pmd_none(*pmd)) 314 return NULL; 315 if (is_hugetlb_pmd(*pmd)) 316 return (pte_t *)pmd; 317 return pte_offset_map(pmd, addr); 318 } 319 320 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, 321 pte_t *ptep, pte_t entry) 322 { 323 unsigned int nptes, orig_shift, shift; 324 unsigned long i, size; 325 pte_t orig; 326 327 size = huge_tte_to_size(entry); 328 329 shift = PAGE_SHIFT; 330 if (size >= PUD_SIZE) 331 shift = PUD_SHIFT; 332 else if (size >= PMD_SIZE) 333 shift = PMD_SHIFT; 334 else 335 shift = PAGE_SHIFT; 336 337 nptes = size >> shift; 338 339 if (!pte_present(*ptep) && pte_present(entry)) 340 mm->context.hugetlb_pte_count += nptes; 341 342 addr &= ~(size - 1); 343 orig = *ptep; 344 orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig); 345 346 for (i = 0; i < nptes; i++) 347 ptep[i] = __pte(pte_val(entry) + (i << shift)); 348 349 maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift); 350 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 351 if (size == HPAGE_SIZE) 352 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0, 353 orig_shift); 354 } 355 356 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, 357 pte_t *ptep) 358 { 359 unsigned int i, nptes, orig_shift, shift; 360 unsigned long size; 361 pte_t entry; 362 363 entry = *ptep; 364 size = huge_tte_to_size(entry); 365 366 shift = PAGE_SHIFT; 367 if (size >= PUD_SIZE) 368 shift = PUD_SHIFT; 369 else if (size >= PMD_SIZE) 370 shift = PMD_SHIFT; 371 else 372 shift = PAGE_SHIFT; 373 374 nptes = size >> shift; 375 orig_shift = pte_none(entry) ? PAGE_SHIFT : huge_tte_to_shift(entry); 376 377 if (pte_present(entry)) 378 mm->context.hugetlb_pte_count -= nptes; 379 380 addr &= ~(size - 1); 381 for (i = 0; i < nptes; i++) 382 ptep[i] = __pte(0UL); 383 384 maybe_tlb_batch_add(mm, addr, ptep, entry, 0, orig_shift); 385 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 386 if (size == HPAGE_SIZE) 387 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0, 388 orig_shift); 389 390 return entry; 391 } 392 393 int pmd_huge(pmd_t pmd) 394 { 395 return !pmd_none(pmd) && 396 (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID; 397 } 398 399 int pud_huge(pud_t pud) 400 { 401 return !pud_none(pud) && 402 (pud_val(pud) & (_PAGE_VALID|_PAGE_PUD_HUGE)) != _PAGE_VALID; 403 } 404 405 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 406 unsigned long addr) 407 { 408 pgtable_t token = pmd_pgtable(*pmd); 409 410 pmd_clear(pmd); 411 pte_free_tlb(tlb, token, addr); 412 mm_dec_nr_ptes(tlb->mm); 413 } 414 415 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 416 unsigned long addr, unsigned long end, 417 unsigned long floor, unsigned long ceiling) 418 { 419 pmd_t *pmd; 420 unsigned long next; 421 unsigned long start; 422 423 start = addr; 424 pmd = pmd_offset(pud, addr); 425 do { 426 next = pmd_addr_end(addr, end); 427 if (pmd_none(*pmd)) 428 continue; 429 if (is_hugetlb_pmd(*pmd)) 430 pmd_clear(pmd); 431 else 432 hugetlb_free_pte_range(tlb, pmd, addr); 433 } while (pmd++, addr = next, addr != end); 434 435 start &= PUD_MASK; 436 if (start < floor) 437 return; 438 if (ceiling) { 439 ceiling &= PUD_MASK; 440 if (!ceiling) 441 return; 442 } 443 if (end - 1 > ceiling - 1) 444 return; 445 446 pmd = pmd_offset(pud, start); 447 pud_clear(pud); 448 pmd_free_tlb(tlb, pmd, start); 449 mm_dec_nr_pmds(tlb->mm); 450 } 451 452 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 453 unsigned long addr, unsigned long end, 454 unsigned long floor, unsigned long ceiling) 455 { 456 pud_t *pud; 457 unsigned long next; 458 unsigned long start; 459 460 start = addr; 461 pud = pud_offset(pgd, addr); 462 do { 463 next = pud_addr_end(addr, end); 464 if (pud_none_or_clear_bad(pud)) 465 continue; 466 if (is_hugetlb_pud(*pud)) 467 pud_clear(pud); 468 else 469 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, 470 ceiling); 471 } while (pud++, addr = next, addr != end); 472 473 start &= PGDIR_MASK; 474 if (start < floor) 475 return; 476 if (ceiling) { 477 ceiling &= PGDIR_MASK; 478 if (!ceiling) 479 return; 480 } 481 if (end - 1 > ceiling - 1) 482 return; 483 484 pud = pud_offset(pgd, start); 485 pgd_clear(pgd); 486 pud_free_tlb(tlb, pud, start); 487 mm_dec_nr_puds(tlb->mm); 488 } 489 490 void hugetlb_free_pgd_range(struct mmu_gather *tlb, 491 unsigned long addr, unsigned long end, 492 unsigned long floor, unsigned long ceiling) 493 { 494 pgd_t *pgd; 495 unsigned long next; 496 497 addr &= PMD_MASK; 498 if (addr < floor) { 499 addr += PMD_SIZE; 500 if (!addr) 501 return; 502 } 503 if (ceiling) { 504 ceiling &= PMD_MASK; 505 if (!ceiling) 506 return; 507 } 508 if (end - 1 > ceiling - 1) 509 end -= PMD_SIZE; 510 if (addr > end - 1) 511 return; 512 513 pgd = pgd_offset(tlb->mm, addr); 514 do { 515 next = pgd_addr_end(addr, end); 516 if (pgd_none_or_clear_bad(pgd)) 517 continue; 518 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); 519 } while (pgd++, addr = next, addr != end); 520 } 521