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