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_minmax, 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 mm_inc_nr_puds(mm); 105 } else { 106 crst_table_init(table, _REGION1_ENTRY_EMPTY); 107 pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd); 108 mm->pgd = (pgd_t *) table; 109 mm->context.asce_limit = -PAGE_SIZE; 110 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 111 _ASCE_USER_BITS | _ASCE_TYPE_REGION1; 112 } 113 notify = 1; 114 spin_unlock_bh(&mm->page_table_lock); 115 } 116 if (notify) 117 on_each_cpu(__crst_table_upgrade, mm, 0); 118 return rc; 119 } 120 121 void crst_table_downgrade(struct mm_struct *mm) 122 { 123 pgd_t *pgd; 124 125 /* downgrade should only happen from 3 to 2 levels (compat only) */ 126 VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE); 127 128 if (current->active_mm == mm) { 129 clear_user_asce(); 130 __tlb_flush_mm(mm); 131 } 132 133 pgd = mm->pgd; 134 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN); 135 mm->context.asce_limit = _REGION3_SIZE; 136 mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH | 137 _ASCE_USER_BITS | _ASCE_TYPE_SEGMENT; 138 crst_table_free(mm, (unsigned long *) pgd); 139 140 if (current->active_mm == mm) 141 set_user_asce(mm); 142 } 143 144 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits) 145 { 146 unsigned int old, new; 147 148 do { 149 old = atomic_read(v); 150 new = old ^ bits; 151 } while (atomic_cmpxchg(v, old, new) != old); 152 return new; 153 } 154 155 #ifdef CONFIG_PGSTE 156 157 struct page *page_table_alloc_pgste(struct mm_struct *mm) 158 { 159 struct page *page; 160 u64 *table; 161 162 page = alloc_page(GFP_KERNEL); 163 if (page) { 164 table = (u64 *)page_to_phys(page); 165 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 166 memset64(table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 167 } 168 return page; 169 } 170 171 void page_table_free_pgste(struct page *page) 172 { 173 __free_page(page); 174 } 175 176 #endif /* CONFIG_PGSTE */ 177 178 /* 179 * page table entry allocation/free routines. 180 */ 181 unsigned long *page_table_alloc(struct mm_struct *mm) 182 { 183 unsigned long *table; 184 struct page *page; 185 unsigned int mask, bit; 186 187 /* Try to get a fragment of a 4K page as a 2K page table */ 188 if (!mm_alloc_pgste(mm)) { 189 table = NULL; 190 spin_lock_bh(&mm->context.lock); 191 if (!list_empty(&mm->context.pgtable_list)) { 192 page = list_first_entry(&mm->context.pgtable_list, 193 struct page, lru); 194 mask = atomic_read(&page->_refcount) >> 24; 195 mask = (mask | (mask >> 4)) & 3; 196 if (mask != 3) { 197 table = (unsigned long *) page_to_phys(page); 198 bit = mask & 1; /* =1 -> second 2K */ 199 if (bit) 200 table += PTRS_PER_PTE; 201 atomic_xor_bits(&page->_refcount, 202 1U << (bit + 24)); 203 list_del(&page->lru); 204 } 205 } 206 spin_unlock_bh(&mm->context.lock); 207 if (table) 208 return table; 209 } 210 /* Allocate a fresh page */ 211 page = alloc_page(GFP_KERNEL); 212 if (!page) 213 return NULL; 214 if (!pgtable_page_ctor(page)) { 215 __free_page(page); 216 return NULL; 217 } 218 arch_set_page_dat(page, 0); 219 /* Initialize page table */ 220 table = (unsigned long *) page_to_phys(page); 221 if (mm_alloc_pgste(mm)) { 222 /* Return 4K page table with PGSTEs */ 223 atomic_xor_bits(&page->_refcount, 3 << 24); 224 memset64((u64 *)table, _PAGE_INVALID, PTRS_PER_PTE); 225 memset64((u64 *)table + PTRS_PER_PTE, 0, PTRS_PER_PTE); 226 } else { 227 /* Return the first 2K fragment of the page */ 228 atomic_xor_bits(&page->_refcount, 1 << 24); 229 memset64((u64 *)table, _PAGE_INVALID, 2 * PTRS_PER_PTE); 230 spin_lock_bh(&mm->context.lock); 231 list_add(&page->lru, &mm->context.pgtable_list); 232 spin_unlock_bh(&mm->context.lock); 233 } 234 return table; 235 } 236 237 void page_table_free(struct mm_struct *mm, unsigned long *table) 238 { 239 struct page *page; 240 unsigned int bit, mask; 241 242 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 243 if (!mm_alloc_pgste(mm)) { 244 /* Free 2K page table fragment of a 4K page */ 245 bit = (__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)); 246 spin_lock_bh(&mm->context.lock); 247 mask = atomic_xor_bits(&page->_refcount, 1U << (bit + 24)); 248 mask >>= 24; 249 if (mask & 3) 250 list_add(&page->lru, &mm->context.pgtable_list); 251 else 252 list_del(&page->lru); 253 spin_unlock_bh(&mm->context.lock); 254 if (mask != 0) 255 return; 256 } else { 257 atomic_xor_bits(&page->_refcount, 3U << 24); 258 } 259 260 pgtable_page_dtor(page); 261 __free_page(page); 262 } 263 264 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table, 265 unsigned long vmaddr) 266 { 267 struct mm_struct *mm; 268 struct page *page; 269 unsigned int bit, mask; 270 271 mm = tlb->mm; 272 page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 273 if (mm_alloc_pgste(mm)) { 274 gmap_unlink(mm, table, vmaddr); 275 table = (unsigned long *) (__pa(table) | 3); 276 tlb_remove_table(tlb, table); 277 return; 278 } 279 bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)); 280 spin_lock_bh(&mm->context.lock); 281 mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24)); 282 mask >>= 24; 283 if (mask & 3) 284 list_add_tail(&page->lru, &mm->context.pgtable_list); 285 else 286 list_del(&page->lru); 287 spin_unlock_bh(&mm->context.lock); 288 table = (unsigned long *) (__pa(table) | (1U << bit)); 289 tlb_remove_table(tlb, table); 290 } 291 292 static void __tlb_remove_table(void *_table) 293 { 294 unsigned int mask = (unsigned long) _table & 3; 295 void *table = (void *)((unsigned long) _table ^ mask); 296 struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT); 297 298 switch (mask) { 299 case 0: /* pmd, pud, or p4d */ 300 free_pages((unsigned long) table, 2); 301 break; 302 case 1: /* lower 2K of a 4K page table */ 303 case 2: /* higher 2K of a 4K page table */ 304 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24)); 305 mask >>= 24; 306 if (mask != 0) 307 break; 308 /* fallthrough */ 309 case 3: /* 4K page table with pgstes */ 310 if (mask & 3) 311 atomic_xor_bits(&page->_refcount, 3 << 24); 312 pgtable_page_dtor(page); 313 __free_page(page); 314 break; 315 } 316 } 317 318 static void tlb_remove_table_smp_sync(void *arg) 319 { 320 /* Simply deliver the interrupt */ 321 } 322 323 static void tlb_remove_table_one(void *table) 324 { 325 /* 326 * This isn't an RCU grace period and hence the page-tables cannot be 327 * assumed to be actually RCU-freed. 328 * 329 * It is however sufficient for software page-table walkers that rely 330 * on IRQ disabling. See the comment near struct mmu_table_batch. 331 */ 332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1); 333 __tlb_remove_table(table); 334 } 335 336 static void tlb_remove_table_rcu(struct rcu_head *head) 337 { 338 struct mmu_table_batch *batch; 339 int i; 340 341 batch = container_of(head, struct mmu_table_batch, rcu); 342 343 for (i = 0; i < batch->nr; i++) 344 __tlb_remove_table(batch->tables[i]); 345 346 free_page((unsigned long)batch); 347 } 348 349 void tlb_table_flush(struct mmu_gather *tlb) 350 { 351 struct mmu_table_batch **batch = &tlb->batch; 352 353 if (*batch) { 354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); 355 *batch = NULL; 356 } 357 } 358 359 void tlb_remove_table(struct mmu_gather *tlb, void *table) 360 { 361 struct mmu_table_batch **batch = &tlb->batch; 362 363 tlb->mm->context.flush_mm = 1; 364 if (*batch == NULL) { 365 *batch = (struct mmu_table_batch *) 366 __get_free_page(GFP_NOWAIT | __GFP_NOWARN); 367 if (*batch == NULL) { 368 __tlb_flush_mm_lazy(tlb->mm); 369 tlb_remove_table_one(table); 370 return; 371 } 372 (*batch)->nr = 0; 373 } 374 (*batch)->tables[(*batch)->nr++] = table; 375 if ((*batch)->nr == MAX_TABLE_BATCH) 376 tlb_flush_mmu(tlb); 377 } 378 379 /* 380 * Base infrastructure required to generate basic asces, region, segment, 381 * and page tables that do not make use of enhanced features like EDAT1. 382 */ 383 384 static struct kmem_cache *base_pgt_cache; 385 386 static unsigned long base_pgt_alloc(void) 387 { 388 u64 *table; 389 390 table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL); 391 if (table) 392 memset64(table, _PAGE_INVALID, PTRS_PER_PTE); 393 return (unsigned long) table; 394 } 395 396 static void base_pgt_free(unsigned long table) 397 { 398 kmem_cache_free(base_pgt_cache, (void *) table); 399 } 400 401 static unsigned long base_crst_alloc(unsigned long val) 402 { 403 unsigned long table; 404 405 table = __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER); 406 if (table) 407 crst_table_init((unsigned long *)table, val); 408 return table; 409 } 410 411 static void base_crst_free(unsigned long table) 412 { 413 free_pages(table, CRST_ALLOC_ORDER); 414 } 415 416 #define BASE_ADDR_END_FUNC(NAME, SIZE) \ 417 static inline unsigned long base_##NAME##_addr_end(unsigned long addr, \ 418 unsigned long end) \ 419 { \ 420 unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1); \ 421 \ 422 return (next - 1) < (end - 1) ? next : end; \ 423 } 424 425 BASE_ADDR_END_FUNC(page, _PAGE_SIZE) 426 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE) 427 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE) 428 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE) 429 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE) 430 431 static inline unsigned long base_lra(unsigned long address) 432 { 433 unsigned long real; 434 435 asm volatile( 436 " lra %0,0(%1)\n" 437 : "=d" (real) : "a" (address) : "cc"); 438 return real; 439 } 440 441 static int base_page_walk(unsigned long origin, unsigned long addr, 442 unsigned long end, int alloc) 443 { 444 unsigned long *pte, next; 445 446 if (!alloc) 447 return 0; 448 pte = (unsigned long *) origin; 449 pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT; 450 do { 451 next = base_page_addr_end(addr, end); 452 *pte = base_lra(addr); 453 } while (pte++, addr = next, addr < end); 454 return 0; 455 } 456 457 static int base_segment_walk(unsigned long origin, unsigned long addr, 458 unsigned long end, int alloc) 459 { 460 unsigned long *ste, next, table; 461 int rc; 462 463 ste = (unsigned long *) origin; 464 ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT; 465 do { 466 next = base_segment_addr_end(addr, end); 467 if (*ste & _SEGMENT_ENTRY_INVALID) { 468 if (!alloc) 469 continue; 470 table = base_pgt_alloc(); 471 if (!table) 472 return -ENOMEM; 473 *ste = table | _SEGMENT_ENTRY; 474 } 475 table = *ste & _SEGMENT_ENTRY_ORIGIN; 476 rc = base_page_walk(table, addr, next, alloc); 477 if (rc) 478 return rc; 479 if (!alloc) 480 base_pgt_free(table); 481 cond_resched(); 482 } while (ste++, addr = next, addr < end); 483 return 0; 484 } 485 486 static int base_region3_walk(unsigned long origin, unsigned long addr, 487 unsigned long end, int alloc) 488 { 489 unsigned long *rtte, next, table; 490 int rc; 491 492 rtte = (unsigned long *) origin; 493 rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT; 494 do { 495 next = base_region3_addr_end(addr, end); 496 if (*rtte & _REGION_ENTRY_INVALID) { 497 if (!alloc) 498 continue; 499 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 500 if (!table) 501 return -ENOMEM; 502 *rtte = table | _REGION3_ENTRY; 503 } 504 table = *rtte & _REGION_ENTRY_ORIGIN; 505 rc = base_segment_walk(table, addr, next, alloc); 506 if (rc) 507 return rc; 508 if (!alloc) 509 base_crst_free(table); 510 } while (rtte++, addr = next, addr < end); 511 return 0; 512 } 513 514 static int base_region2_walk(unsigned long origin, unsigned long addr, 515 unsigned long end, int alloc) 516 { 517 unsigned long *rste, next, table; 518 int rc; 519 520 rste = (unsigned long *) origin; 521 rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT; 522 do { 523 next = base_region2_addr_end(addr, end); 524 if (*rste & _REGION_ENTRY_INVALID) { 525 if (!alloc) 526 continue; 527 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 528 if (!table) 529 return -ENOMEM; 530 *rste = table | _REGION2_ENTRY; 531 } 532 table = *rste & _REGION_ENTRY_ORIGIN; 533 rc = base_region3_walk(table, addr, next, alloc); 534 if (rc) 535 return rc; 536 if (!alloc) 537 base_crst_free(table); 538 } while (rste++, addr = next, addr < end); 539 return 0; 540 } 541 542 static int base_region1_walk(unsigned long origin, unsigned long addr, 543 unsigned long end, int alloc) 544 { 545 unsigned long *rfte, next, table; 546 int rc; 547 548 rfte = (unsigned long *) origin; 549 rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT; 550 do { 551 next = base_region1_addr_end(addr, end); 552 if (*rfte & _REGION_ENTRY_INVALID) { 553 if (!alloc) 554 continue; 555 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 556 if (!table) 557 return -ENOMEM; 558 *rfte = table | _REGION1_ENTRY; 559 } 560 table = *rfte & _REGION_ENTRY_ORIGIN; 561 rc = base_region2_walk(table, addr, next, alloc); 562 if (rc) 563 return rc; 564 if (!alloc) 565 base_crst_free(table); 566 } while (rfte++, addr = next, addr < end); 567 return 0; 568 } 569 570 /** 571 * base_asce_free - free asce and tables returned from base_asce_alloc() 572 * @asce: asce to be freed 573 * 574 * Frees all region, segment, and page tables that were allocated with a 575 * corresponding base_asce_alloc() call. 576 */ 577 void base_asce_free(unsigned long asce) 578 { 579 unsigned long table = asce & _ASCE_ORIGIN; 580 581 if (!asce) 582 return; 583 switch (asce & _ASCE_TYPE_MASK) { 584 case _ASCE_TYPE_SEGMENT: 585 base_segment_walk(table, 0, _REGION3_SIZE, 0); 586 break; 587 case _ASCE_TYPE_REGION3: 588 base_region3_walk(table, 0, _REGION2_SIZE, 0); 589 break; 590 case _ASCE_TYPE_REGION2: 591 base_region2_walk(table, 0, _REGION1_SIZE, 0); 592 break; 593 case _ASCE_TYPE_REGION1: 594 base_region1_walk(table, 0, -_PAGE_SIZE, 0); 595 break; 596 } 597 base_crst_free(table); 598 } 599 600 static int base_pgt_cache_init(void) 601 { 602 static DEFINE_MUTEX(base_pgt_cache_mutex); 603 unsigned long sz = _PAGE_TABLE_SIZE; 604 605 if (base_pgt_cache) 606 return 0; 607 mutex_lock(&base_pgt_cache_mutex); 608 if (!base_pgt_cache) 609 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL); 610 mutex_unlock(&base_pgt_cache_mutex); 611 return base_pgt_cache ? 0 : -ENOMEM; 612 } 613 614 /** 615 * base_asce_alloc - create kernel mapping without enhanced DAT features 616 * @addr: virtual start address of kernel mapping 617 * @num_pages: number of consecutive pages 618 * 619 * Generate an asce, including all required region, segment and page tables, 620 * that can be used to access the virtual kernel mapping. The difference is 621 * that the returned asce does not make use of any enhanced DAT features like 622 * e.g. large pages. This is required for some I/O functions that pass an 623 * asce, like e.g. some service call requests. 624 * 625 * Note: the returned asce may NEVER be attached to any cpu. It may only be 626 * used for I/O requests. tlb entries that might result because the 627 * asce was attached to a cpu won't be cleared. 628 */ 629 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages) 630 { 631 unsigned long asce, table, end; 632 int rc; 633 634 if (base_pgt_cache_init()) 635 return 0; 636 end = addr + num_pages * PAGE_SIZE; 637 if (end <= _REGION3_SIZE) { 638 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY); 639 if (!table) 640 return 0; 641 rc = base_segment_walk(table, addr, end, 1); 642 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH; 643 } else if (end <= _REGION2_SIZE) { 644 table = base_crst_alloc(_REGION3_ENTRY_EMPTY); 645 if (!table) 646 return 0; 647 rc = base_region3_walk(table, addr, end, 1); 648 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH; 649 } else if (end <= _REGION1_SIZE) { 650 table = base_crst_alloc(_REGION2_ENTRY_EMPTY); 651 if (!table) 652 return 0; 653 rc = base_region2_walk(table, addr, end, 1); 654 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH; 655 } else { 656 table = base_crst_alloc(_REGION1_ENTRY_EMPTY); 657 if (!table) 658 return 0; 659 rc = base_region1_walk(table, addr, end, 1); 660 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH; 661 } 662 if (rc) { 663 base_asce_free(asce); 664 asce = 0; 665 } 666 return asce; 667 } 668