1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Page table allocation functions 4 * 5 * Copyright IBM Corp. 2016 6 * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com> 7 */ 8 9 #include <linux/sysctl.h> 10 #include <linux/slab.h> 11 #include <linux/mm.h> 12 #include <asm/mmu_context.h> 13 #include <asm/pgalloc.h> 14 #include <asm/gmap.h> 15 #include <asm/tlb.h> 16 #include <asm/tlbflush.h> 17 18 #ifdef CONFIG_PGSTE 19 20 static int page_table_allocate_pgste_min = 0; 21 static int page_table_allocate_pgste_max = 1; 22 int page_table_allocate_pgste = 0; 23 EXPORT_SYMBOL(page_table_allocate_pgste); 24 25 static struct ctl_table page_table_sysctl[] = { 26 { 27 .procname = "allocate_pgste", 28 .data = &page_table_allocate_pgste, 29 .maxlen = sizeof(int), 30 .mode = S_IRUGO | S_IWUSR, 31 .proc_handler = proc_dointvec, 32 .extra1 = &page_table_allocate_pgste_min, 33 .extra2 = &page_table_allocate_pgste_max, 34 }, 35 { } 36 }; 37 38 static struct ctl_table page_table_sysctl_dir[] = { 39 { 40 .procname = "vm", 41 .maxlen = 0, 42 .mode = 0555, 43 .child = page_table_sysctl, 44 }, 45 { } 46 }; 47 48 static int __init page_table_register_sysctl(void) 49 { 50 return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM; 51 } 52 __initcall(page_table_register_sysctl); 53 54 #endif /* CONFIG_PGSTE */ 55 56 unsigned long *crst_table_alloc(struct mm_struct *mm) 57 { 58 struct page *page = alloc_pages(GFP_KERNEL, 2); 59 60 if (!page) 61 return NULL; 62 arch_set_page_dat(page, 2); 63 return (unsigned long *) page_to_phys(page); 64 } 65 66 void crst_table_free(struct mm_struct *mm, unsigned long *table) 67 { 68 free_pages((unsigned long) table, 2); 69 } 70 71 static void __crst_table_upgrade(void *arg) 72 { 73 struct mm_struct *mm = arg; 74 75 if (current->active_mm == mm) 76 set_user_asce(mm); 77 __tlb_flush_local(); 78 } 79 80 int crst_table_upgrade(struct mm_struct *mm, unsigned long end) 81 { 82 unsigned long *table, *pgd; 83 int rc, notify; 84 85 /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */ 86 VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE); 87 rc = 0; 88 notify = 0; 89 while (mm->context.asce_limit < end) { 90 table = crst_table_alloc(mm); 91 if (!table) { 92 rc = -ENOMEM; 93 break; 94 } 95 spin_lock_bh(&mm->page_table_lock); 96 pgd = (unsigned long *) mm->pgd; 97 if (mm->context.asce_limit == _REGION2_SIZE) { 98 crst_table_init(table, _REGION2_ENTRY_EMPTY); 99 p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd); 100 mm->pgd = (pgd_t *) table; 101 mm->context.asce_limit = _REGION1_SIZE; 102 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 103 _ASCE_USER_BITS | _ASCE_TYPE_REGION2; 104 } else { 105 crst_table_init(table, _REGION1_ENTRY_EMPTY); 106 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd); 107 mm->pgd = (pgd_t *) table; 108 mm->context.asce_limit = -PAGE_SIZE; 109 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 110 _ASCE_USER_BITS | _ASCE_TYPE_REGION1; 111 } 112 notify = 1; 113 spin_unlock_bh(&mm->page_table_lock); 114 } 115 if (notify) 116 on_each_cpu(__crst_table_upgrade, mm, 0); 117 return rc; 118 } 119 120 void crst_table_downgrade(struct mm_struct *mm) 121 { 122 pgd_t *pgd; 123 124 /* downgrade should only happen from 3 to 2 levels (compat only) */ 125 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE); 126 127 if (current->active_mm == mm) { 128 clear_user_asce(); 129 __tlb_flush_mm(mm); 130 } 131 132 pgd = mm->pgd; 133 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN); 134 mm->context.asce_limit = _REGION3_SIZE; 135 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 136 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT; 137 crst_table_free(mm, (unsigned long *) pgd); 138 139 if (current->active_mm == mm) 140 set_user_asce(mm); 141 } 142 143 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits) 144 { 145 unsigned int old, new; 146 147 do { 148 old = atomic_read(v); 149 new = old ^ bits; 150 } while (atomic_cmpxchg(v, old, new) != old); 151 return new; 152 } 153 154 #ifdef CONFIG_PGSTE 155 156 struct page *page_table_alloc_pgste(struct mm_struct *mm) 157 { 158 struct page *page; 159 u64 *table; 160 161 page = alloc_page(GFP_KERNEL); 162 if (page) { 163 table = (u64 *)page_to_phys(page); 164 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 165 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 166 } 167 return page; 168 } 169 170 void page_table_free_pgste(struct page *page) 171 { 172 __free_page(page); 173 } 174 175 #endif /* CONFIG_PGSTE */ 176 177 /* 178 * page table entry allocation/free routines. 179 */ 180 unsigned long *page_table_alloc(struct mm_struct *mm) 181 { 182 unsigned long *table; 183 struct page *page; 184 unsigned int mask, bit; 185 186 /* Try to get a fragment of a 4K page as a 2K page table */ 187 if (!mm_alloc_pgste(mm)) { 188 table = NULL; 189 spin_lock_bh(&mm->context.lock); 190 if (!list_empty(&mm->context.pgtable_list)) { 191 page = list_first_entry(&mm->context.pgtable_list, 192 struct page, lru); 193 mask = atomic_read(&page->_mapcount); 194 mask = (mask | (mask >> 4)) & 3; 195 if (mask != 3) { 196 table = (unsigned long *) page_to_phys(page); 197 bit = mask & 1; /* =1 -> second 2K */ 198 if (bit) 199 table += PTRS_PER_PTE; 200 atomic_xor_bits(&page->_mapcount, 1U << bit); 201 list_del(&page->lru); 202 } 203 } 204 spin_unlock_bh(&mm->context.lock); 205 if (table) 206 return table; 207 } 208 /* Allocate a fresh page */ 209 page = alloc_page(GFP_KERNEL); 210 if (!page) 211 return NULL; 212 if (!pgtable_page_ctor(page)) { 213 __free_page(page); 214 return NULL; 215 } 216 arch_set_page_dat(page, 0); 217 /* Initialize page table */ 218 table = (unsigned long *) page_to_phys(page); 219 if (mm_alloc_pgste(mm)) { 220 /* Return 4K page table with PGSTEs */ 221 atomic_set(&page->_mapcount, 3); 222 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 223 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 224 } else { 225 /* Return the first 2K fragment of the page */ 226 atomic_set(&page->_mapcount, 1); 227 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE); 228 spin_lock_bh(&mm->context.lock); 229 list_add(&page->lru, &mm->context.pgtable_list); 230 spin_unlock_bh(&mm->context.lock); 231 } 232 return table; 233 } 234 235 void page_table_free(struct mm_struct *mm, unsigned long *table) 236 { 237 struct page *page; 238 unsigned int bit, mask; 239 240 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 241 if (!mm_alloc_pgste(mm)) { 242 /* Free 2K page table fragment of a 4K page */ 243 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)); 244 spin_lock_bh(&mm->context.lock); 245 mask = atomic_xor_bits(&page->_mapcount, 1U << bit); 246 if (mask & 3) 247 list_add(&page->lru, &mm->context.pgtable_list); 248 else 249 list_del(&page->lru); 250 spin_unlock_bh(&mm->context.lock); 251 if (mask != 0) 252 return; 253 } 254 255 pgtable_page_dtor(page); 256 atomic_set(&page->_mapcount, -1); 257 __free_page(page); 258 } 259 260 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table, 261 unsigned long vmaddr) 262 { 263 struct mm_struct *mm; 264 struct page *page; 265 unsigned int bit, mask; 266 267 mm = tlb->mm; 268 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 269 if (mm_alloc_pgste(mm)) { 270 gmap_unlink(mm, table, vmaddr); 271 table = (unsigned long *) (__pa(table) | 3); 272 tlb_remove_table(tlb, table); 273 return; 274 } 275 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)); 276 spin_lock_bh(&mm->context.lock); 277 mask = atomic_xor_bits(&page->_mapcount, 0x11U << bit); 278 if (mask & 3) 279 list_add_tail(&page->lru, &mm->context.pgtable_list); 280 else 281 list_del(&page->lru); 282 spin_unlock_bh(&mm->context.lock); 283 table = (unsigned long *) (__pa(table) | (1U << bit)); 284 tlb_remove_table(tlb, table); 285 } 286 287 static void __tlb_remove_table(void *_table) 288 { 289 unsigned int mask = (unsigned long) _table & 3; 290 void *table = (void *)((unsigned long) _table ^ mask); 291 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 292 293 switch (mask) { 294 case 0: /* pmd, pud, or p4d */ 295 free_pages((unsigned long) table, 2); 296 break; 297 case 1: /* lower 2K of a 4K page table */ 298 case 2: /* higher 2K of a 4K page table */ 299 if (atomic_xor_bits(&page->_mapcount, mask << 4) != 0) 300 break; 301 /* fallthrough */ 302 case 3: /* 4K page table with pgstes */ 303 pgtable_page_dtor(page); 304 atomic_set(&page->_mapcount, -1); 305 __free_page(page); 306 break; 307 } 308 } 309 310 static void tlb_remove_table_smp_sync(void *arg) 311 { 312 /* Simply deliver the interrupt */ 313 } 314 315 static void tlb_remove_table_one(void *table) 316 { 317 /* 318 * This isn't an RCU grace period and hence the page-tables cannot be 319 * assumed to be actually RCU-freed. 320 * 321 * It is however sufficient for software page-table walkers that rely 322 * on IRQ disabling. See the comment near struct mmu_table_batch. 323 */ 324 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 325 __tlb_remove_table(table); 326 } 327 328 static void tlb_remove_table_rcu(struct rcu_head *head) 329 { 330 struct mmu_table_batch *batch; 331 int i; 332 333 batch = container_of(head, struct mmu_table_batch, rcu); 334 335 for (i = 0; i < batch->nr; i++) 336 __tlb_remove_table(batch->tables[i]); 337 338 free_page((unsigned long)batch); 339 } 340 341 void tlb_table_flush(struct mmu_gather *tlb) 342 { 343 struct mmu_table_batch **batch = &tlb->batch; 344 345 if (*batch) { 346 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 347 *batch = NULL; 348 } 349 } 350 351 void tlb_remove_table(struct mmu_gather *tlb, void *table) 352 { 353 struct mmu_table_batch **batch = &tlb->batch; 354 355 tlb->mm->context.flush_mm = 1; 356 if (*batch == NULL) { 357 *batch = (struct mmu_table_batch *) 358 __get_free_page(GFP_NOWAIT | __GFP_NOWARN); 359 if (*batch == NULL) { 360 __tlb_flush_mm_lazy(tlb->mm); 361 tlb_remove_table_one(table); 362 return; 363 } 364 (*batch)->nr = 0; 365 } 366 (*batch)->tables[(*batch)->nr++] = table; 367 if ((*batch)->nr == MAX_TABLE_BATCH) 368 tlb_flush_mmu(tlb); 369 } 370 371 /* 372 * Base infrastructure required to generate basic asces, region, segment, 373 * and page tables that do not make use of enhanced features like EDAT1. 374 */ 375 376 static struct kmem_cache *base_pgt_cache; 377 378 static unsigned long base_pgt_alloc(void) 379 { 380 u64 *table; 381 382 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL); 383 if (table) 384 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 385 return (unsigned long) table; 386 } 387 388 static void base_pgt_free(unsigned long table) 389 { 390 kmem_cache_free(base_pgt_cache, (void *) table); 391 } 392 393 static unsigned long base_crst_alloc(unsigned long val) 394 { 395 unsigned long table; 396 397 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER); 398 if (table) 399 crst_table_init((unsigned long *)table, val); 400 return table; 401 } 402 403 static void base_crst_free(unsigned long table) 404 { 405 free_pages(table, CRST_ALLOC_ORDER); 406 } 407 408 #define BASE_ADDR_END_FUNC(NAME, SIZE) \ 409 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \ 410 unsigned long end) \ 411 { \ 412 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \ 413 \ 414 return (next - 1) < (end - 1) ? next : end; \ 415 } 416 417 BASE_ADDR_END_FUNC(page, _PAGE_SIZE) 418 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE) 419 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE) 420 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE) 421 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE) 422 423 static inline unsigned long base_lra(unsigned long address) 424 { 425 unsigned long real; 426 427 asm volatile( 428 " lra %0,0(%1)\n" 429 : "=d" (real) : "a" (address) : "cc"); 430 return real; 431 } 432 433 static int base_page_walk(unsigned long origin, unsigned long addr, 434 unsigned long end, int alloc) 435 { 436 unsigned long *pte, next; 437 438 if (!alloc) 439 return 0; 440 pte = (unsigned long *) origin; 441 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT; 442 do { 443 next = base_page_addr_end(addr, end); 444 *pte = base_lra(addr); 445 } while (pte++, addr = next, addr < end); 446 return 0; 447 } 448 449 static int base_segment_walk(unsigned long origin, unsigned long addr, 450 unsigned long end, int alloc) 451 { 452 unsigned long *ste, next, table; 453 int rc; 454 455 ste = (unsigned long *) origin; 456 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT; 457 do { 458 next = base_segment_addr_end(addr, end); 459 if (*ste & _SEGMENT_ENTRY_INVALID) { 460 if (!alloc) 461 continue; 462 table = base_pgt_alloc(); 463 if (!table) 464 return -ENOMEM; 465 *ste = table | _SEGMENT_ENTRY; 466 } 467 table = *ste & _SEGMENT_ENTRY_ORIGIN; 468 rc = base_page_walk(table, addr, next, alloc); 469 if (rc) 470 return rc; 471 if (!alloc) 472 base_pgt_free(table); 473 cond_resched(); 474 } while (ste++, addr = next, addr < end); 475 return 0; 476 } 477 478 static int base_region3_walk(unsigned long origin, unsigned long addr, 479 unsigned long end, int alloc) 480 { 481 unsigned long *rtte, next, table; 482 int rc; 483 484 rtte = (unsigned long *) origin; 485 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT; 486 do { 487 next = base_region3_addr_end(addr, end); 488 if (*rtte & _REGION_ENTRY_INVALID) { 489 if (!alloc) 490 continue; 491 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 492 if (!table) 493 return -ENOMEM; 494 *rtte = table | _REGION3_ENTRY; 495 } 496 table = *rtte & _REGION_ENTRY_ORIGIN; 497 rc = base_segment_walk(table, addr, next, alloc); 498 if (rc) 499 return rc; 500 if (!alloc) 501 base_crst_free(table); 502 } while (rtte++, addr = next, addr < end); 503 return 0; 504 } 505 506 static int base_region2_walk(unsigned long origin, unsigned long addr, 507 unsigned long end, int alloc) 508 { 509 unsigned long *rste, next, table; 510 int rc; 511 512 rste = (unsigned long *) origin; 513 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT; 514 do { 515 next = base_region2_addr_end(addr, end); 516 if (*rste & _REGION_ENTRY_INVALID) { 517 if (!alloc) 518 continue; 519 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 520 if (!table) 521 return -ENOMEM; 522 *rste = table | _REGION2_ENTRY; 523 } 524 table = *rste & _REGION_ENTRY_ORIGIN; 525 rc = base_region3_walk(table, addr, next, alloc); 526 if (rc) 527 return rc; 528 if (!alloc) 529 base_crst_free(table); 530 } while (rste++, addr = next, addr < end); 531 return 0; 532 } 533 534 static int base_region1_walk(unsigned long origin, unsigned long addr, 535 unsigned long end, int alloc) 536 { 537 unsigned long *rfte, next, table; 538 int rc; 539 540 rfte = (unsigned long *) origin; 541 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT; 542 do { 543 next = base_region1_addr_end(addr, end); 544 if (*rfte & _REGION_ENTRY_INVALID) { 545 if (!alloc) 546 continue; 547 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 548 if (!table) 549 return -ENOMEM; 550 *rfte = table | _REGION1_ENTRY; 551 } 552 table = *rfte & _REGION_ENTRY_ORIGIN; 553 rc = base_region2_walk(table, addr, next, alloc); 554 if (rc) 555 return rc; 556 if (!alloc) 557 base_crst_free(table); 558 } while (rfte++, addr = next, addr < end); 559 return 0; 560 } 561 562 /** 563 * base_asce_free - free asce and tables returned from base_asce_alloc() 564 * @asce: asce to be freed 565 * 566 * Frees all region, segment, and page tables that were allocated with a 567 * corresponding base_asce_alloc() call. 568 */ 569 void base_asce_free(unsigned long asce) 570 { 571 unsigned long table = asce & _ASCE_ORIGIN; 572 573 if (!asce) 574 return; 575 switch (asce & _ASCE_TYPE_MASK) { 576 case _ASCE_TYPE_SEGMENT: 577 base_segment_walk(table, 0, _REGION3_SIZE, 0); 578 break; 579 case _ASCE_TYPE_REGION3: 580 base_region3_walk(table, 0, _REGION2_SIZE, 0); 581 break; 582 case _ASCE_TYPE_REGION2: 583 base_region2_walk(table, 0, _REGION1_SIZE, 0); 584 break; 585 case _ASCE_TYPE_REGION1: 586 base_region1_walk(table, 0, -_PAGE_SIZE, 0); 587 break; 588 } 589 base_crst_free(table); 590 } 591 592 static int base_pgt_cache_init(void) 593 { 594 static DEFINE_MUTEX(base_pgt_cache_mutex); 595 unsigned long sz = _PAGE_TABLE_SIZE; 596 597 if (base_pgt_cache) 598 return 0; 599 mutex_lock(&base_pgt_cache_mutex); 600 if (!base_pgt_cache) 601 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL); 602 mutex_unlock(&base_pgt_cache_mutex); 603 return base_pgt_cache ? 0 : -ENOMEM; 604 } 605 606 /** 607 * base_asce_alloc - create kernel mapping without enhanced DAT features 608 * @addr: virtual start address of kernel mapping 609 * @num_pages: number of consecutive pages 610 * 611 * Generate an asce, including all required region, segment and page tables, 612 * that can be used to access the virtual kernel mapping. The difference is 613 * that the returned asce does not make use of any enhanced DAT features like 614 * e.g. large pages. This is required for some I/O functions that pass an 615 * asce, like e.g. some service call requests. 616 * 617 * Note: the returned asce may NEVER be attached to any cpu. It may only be 618 * used for I/O requests. tlb entries that might result because the 619 * asce was attached to a cpu won't be cleared. 620 */ 621 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages) 622 { 623 unsigned long asce, table, end; 624 int rc; 625 626 if (base_pgt_cache_init()) 627 return 0; 628 end = addr + num_pages * PAGE_SIZE; 629 if (end <= _REGION3_SIZE) { 630 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 631 if (!table) 632 return 0; 633 rc = base_segment_walk(table, addr, end, 1); 634 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH; 635 } else if (end <= _REGION2_SIZE) { 636 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 637 if (!table) 638 return 0; 639 rc = base_region3_walk(table, addr, end, 1); 640 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH; 641 } else if (end <= _REGION1_SIZE) { 642 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 643 if (!table) 644 return 0; 645 rc = base_region2_walk(table, addr, end, 1); 646 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH; 647 } else { 648 table = base_crst_alloc(_REGION1_ENTRY_EMPTY); 649 if (!table) 650 return 0; 651 rc = base_region1_walk(table, addr, end, 1); 652 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH; 653 } 654 if (rc) { 655 base_asce_free(asce); 656 asce = 0; 657 } 658 return asce; 659 } 660