1 /* 2 * SPARC64 Huge TLB page support. 3 * 4 * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net) 5 */ 6 7 #include <linux/fs.h> 8 #include <linux/mm.h> 9 #include <linux/sched/mm.h> 10 #include <linux/hugetlb.h> 11 #include <linux/pagemap.h> 12 #include <linux/sysctl.h> 13 14 #include <asm/mman.h> 15 #include <asm/pgalloc.h> 16 #include <asm/pgtable.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_2GB_SHIFT: 147 hugepage_size = _PAGE_SZ2GB_4V; 148 pte_val(entry) |= _PAGE_PMD_HUGE; 149 break; 150 case HPAGE_256MB_SHIFT: 151 hugepage_size = _PAGE_SZ256MB_4V; 152 pte_val(entry) |= _PAGE_PMD_HUGE; 153 break; 154 case HPAGE_SHIFT: 155 pte_val(entry) |= _PAGE_PMD_HUGE; 156 break; 157 case HPAGE_64K_SHIFT: 158 hugepage_size = _PAGE_SZ64K_4V; 159 break; 160 default: 161 WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift); 162 } 163 164 pte_val(entry) = pte_val(entry) | hugepage_size; 165 return entry; 166 } 167 168 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift) 169 { 170 if (tlb_type == hypervisor) 171 return sun4v_hugepage_shift_to_tte(entry, shift); 172 else 173 return sun4u_hugepage_shift_to_tte(entry, shift); 174 } 175 176 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma, 177 struct page *page, int writeable) 178 { 179 unsigned int shift = huge_page_shift(hstate_vma(vma)); 180 181 return hugepage_shift_to_tte(entry, shift); 182 } 183 184 static unsigned int sun4v_huge_tte_to_shift(pte_t entry) 185 { 186 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V; 187 unsigned int shift; 188 189 switch (tte_szbits) { 190 case _PAGE_SZ2GB_4V: 191 shift = HPAGE_2GB_SHIFT; 192 break; 193 case _PAGE_SZ256MB_4V: 194 shift = HPAGE_256MB_SHIFT; 195 break; 196 case _PAGE_SZ4MB_4V: 197 shift = REAL_HPAGE_SHIFT; 198 break; 199 case _PAGE_SZ64K_4V: 200 shift = HPAGE_64K_SHIFT; 201 break; 202 default: 203 shift = PAGE_SHIFT; 204 break; 205 } 206 return shift; 207 } 208 209 static unsigned int sun4u_huge_tte_to_shift(pte_t entry) 210 { 211 unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U; 212 unsigned int shift; 213 214 switch (tte_szbits) { 215 case _PAGE_SZ256MB_4U: 216 shift = HPAGE_256MB_SHIFT; 217 break; 218 case _PAGE_SZ4MB_4U: 219 shift = REAL_HPAGE_SHIFT; 220 break; 221 case _PAGE_SZ64K_4U: 222 shift = HPAGE_64K_SHIFT; 223 break; 224 default: 225 shift = PAGE_SHIFT; 226 break; 227 } 228 return shift; 229 } 230 231 static unsigned int huge_tte_to_shift(pte_t entry) 232 { 233 unsigned long shift; 234 235 if (tlb_type == hypervisor) 236 shift = sun4v_huge_tte_to_shift(entry); 237 else 238 shift = sun4u_huge_tte_to_shift(entry); 239 240 if (shift == PAGE_SHIFT) 241 WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n", 242 pte_val(entry)); 243 244 return shift; 245 } 246 247 static unsigned long huge_tte_to_size(pte_t pte) 248 { 249 unsigned long size = 1UL << huge_tte_to_shift(pte); 250 251 if (size == REAL_HPAGE_SIZE) 252 size = HPAGE_SIZE; 253 return size; 254 } 255 256 pte_t *huge_pte_alloc(struct mm_struct *mm, 257 unsigned long addr, unsigned long sz) 258 { 259 pgd_t *pgd; 260 pud_t *pud; 261 pmd_t *pmd; 262 pte_t *pte = NULL; 263 264 pgd = pgd_offset(mm, addr); 265 pud = pud_alloc(mm, pgd, addr); 266 if (pud) { 267 pmd = pmd_alloc(mm, pud, addr); 268 if (!pmd) 269 return NULL; 270 271 if (sz >= PMD_SIZE) 272 pte = (pte_t *)pmd; 273 else 274 pte = pte_alloc_map(mm, pmd, addr); 275 } 276 277 return pte; 278 } 279 280 pte_t *huge_pte_offset(struct mm_struct *mm, 281 unsigned long addr, unsigned long sz) 282 { 283 pgd_t *pgd; 284 pud_t *pud; 285 pmd_t *pmd; 286 pte_t *pte = NULL; 287 288 pgd = pgd_offset(mm, addr); 289 if (!pgd_none(*pgd)) { 290 pud = pud_offset(pgd, addr); 291 if (!pud_none(*pud)) { 292 pmd = pmd_offset(pud, addr); 293 if (!pmd_none(*pmd)) { 294 if (is_hugetlb_pmd(*pmd)) 295 pte = (pte_t *)pmd; 296 else 297 pte = pte_offset_map(pmd, addr); 298 } 299 } 300 } 301 302 return pte; 303 } 304 305 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, 306 pte_t *ptep, pte_t entry) 307 { 308 unsigned int i, nptes, orig_shift, shift; 309 unsigned long size; 310 pte_t orig; 311 312 size = huge_tte_to_size(entry); 313 shift = size >= HPAGE_SIZE ? PMD_SHIFT : PAGE_SHIFT; 314 nptes = size >> shift; 315 316 if (!pte_present(*ptep) && pte_present(entry)) 317 mm->context.hugetlb_pte_count += nptes; 318 319 addr &= ~(size - 1); 320 orig = *ptep; 321 orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig); 322 323 for (i = 0; i < nptes; i++) 324 ptep[i] = __pte(pte_val(entry) + (i << shift)); 325 326 maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift); 327 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 328 if (size == HPAGE_SIZE) 329 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0, 330 orig_shift); 331 } 332 333 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, 334 pte_t *ptep) 335 { 336 unsigned int i, nptes, hugepage_shift; 337 unsigned long size; 338 pte_t entry; 339 340 entry = *ptep; 341 size = huge_tte_to_size(entry); 342 if (size >= HPAGE_SIZE) 343 nptes = size >> PMD_SHIFT; 344 else 345 nptes = size >> PAGE_SHIFT; 346 347 hugepage_shift = pte_none(entry) ? PAGE_SHIFT : 348 huge_tte_to_shift(entry); 349 350 if (pte_present(entry)) 351 mm->context.hugetlb_pte_count -= nptes; 352 353 addr &= ~(size - 1); 354 for (i = 0; i < nptes; i++) 355 ptep[i] = __pte(0UL); 356 357 maybe_tlb_batch_add(mm, addr, ptep, entry, 0, hugepage_shift); 358 /* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */ 359 if (size == HPAGE_SIZE) 360 maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0, 361 hugepage_shift); 362 363 return entry; 364 } 365 366 int pmd_huge(pmd_t pmd) 367 { 368 return !pmd_none(pmd) && 369 (pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID; 370 } 371 372 int pud_huge(pud_t pud) 373 { 374 return 0; 375 } 376 377 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, 378 unsigned long addr) 379 { 380 pgtable_t token = pmd_pgtable(*pmd); 381 382 pmd_clear(pmd); 383 pte_free_tlb(tlb, token, addr); 384 atomic_long_dec(&tlb->mm->nr_ptes); 385 } 386 387 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud, 388 unsigned long addr, unsigned long end, 389 unsigned long floor, unsigned long ceiling) 390 { 391 pmd_t *pmd; 392 unsigned long next; 393 unsigned long start; 394 395 start = addr; 396 pmd = pmd_offset(pud, addr); 397 do { 398 next = pmd_addr_end(addr, end); 399 if (pmd_none(*pmd)) 400 continue; 401 if (is_hugetlb_pmd(*pmd)) 402 pmd_clear(pmd); 403 else 404 hugetlb_free_pte_range(tlb, pmd, addr); 405 } while (pmd++, addr = next, addr != end); 406 407 start &= PUD_MASK; 408 if (start < floor) 409 return; 410 if (ceiling) { 411 ceiling &= PUD_MASK; 412 if (!ceiling) 413 return; 414 } 415 if (end - 1 > ceiling - 1) 416 return; 417 418 pmd = pmd_offset(pud, start); 419 pud_clear(pud); 420 pmd_free_tlb(tlb, pmd, start); 421 mm_dec_nr_pmds(tlb->mm); 422 } 423 424 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, 425 unsigned long addr, unsigned long end, 426 unsigned long floor, unsigned long ceiling) 427 { 428 pud_t *pud; 429 unsigned long next; 430 unsigned long start; 431 432 start = addr; 433 pud = pud_offset(pgd, addr); 434 do { 435 next = pud_addr_end(addr, end); 436 if (pud_none_or_clear_bad(pud)) 437 continue; 438 hugetlb_free_pmd_range(tlb, pud, addr, next, floor, 439 ceiling); 440 } while (pud++, addr = next, addr != end); 441 442 start &= PGDIR_MASK; 443 if (start < floor) 444 return; 445 if (ceiling) { 446 ceiling &= PGDIR_MASK; 447 if (!ceiling) 448 return; 449 } 450 if (end - 1 > ceiling - 1) 451 return; 452 453 pud = pud_offset(pgd, start); 454 pgd_clear(pgd); 455 pud_free_tlb(tlb, pud, start); 456 } 457 458 void hugetlb_free_pgd_range(struct mmu_gather *tlb, 459 unsigned long addr, unsigned long end, 460 unsigned long floor, unsigned long ceiling) 461 { 462 pgd_t *pgd; 463 unsigned long next; 464 465 addr &= PMD_MASK; 466 if (addr < floor) { 467 addr += PMD_SIZE; 468 if (!addr) 469 return; 470 } 471 if (ceiling) { 472 ceiling &= PMD_MASK; 473 if (!ceiling) 474 return; 475 } 476 if (end - 1 > ceiling - 1) 477 end -= PMD_SIZE; 478 if (addr > end - 1) 479 return; 480 481 pgd = pgd_offset(tlb->mm, addr); 482 do { 483 next = pgd_addr_end(addr, end); 484 if (pgd_none_or_clear_bad(pgd)) 485 continue; 486 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling); 487 } while (pgd++, addr = next, addr != end); 488 } 489