arch/s390/mm/pgalloc.c

   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 int page_table_allocate_pgste = 0;
  21 EXPORT_SYMBOL(page_table_allocate_pgste);
  22
  23 static struct ctl_table page_table_sysctl[] = {
  24         {
  25                 .procname       = "allocate_pgste",
  26                 .data           = &page_table_allocate_pgste,
  27                 .maxlen         = sizeof(int),
  28                 .mode           = S_IRUGO | S_IWUSR,
  29                 .proc_handler   = proc_dointvec_minmax,
  30                 .extra1         = SYSCTL_ZERO,
  31                 .extra2         = SYSCTL_ONE,
  32         },
  33         { }
  34 };
  35
  36 static struct ctl_table page_table_sysctl_dir[] = {
  37         {
  38                 .procname       = "vm",
  39                 .maxlen         = 0,
  40                 .mode           = 0555,
  41                 .child          = page_table_sysctl,
  42         },
  43         { }
  44 };
  45
  46 static int __init page_table_register_sysctl(void)
  47 {
  48         return register_sysctl_table(page_table_sysctl_dir) ? 0 : -ENOMEM;
  49 }
  50 __initcall(page_table_register_sysctl);
  51
  52 #endif /* CONFIG_PGSTE */
  53
  54 unsigned long *crst_table_alloc(struct mm_struct *mm)
  55 {
  56         struct page *page = alloc_pages(GFP_KERNEL, 2);
  57
  58         if (!page)
  59                 return NULL;
  60         arch_set_page_dat(page, 2);
  61         return (unsigned long *) page_to_phys(page);
  62 }
  63
  64 void crst_table_free(struct mm_struct *mm, unsigned long *table)
  65 {
  66         free_pages((unsigned long) table, 2);
  67 }
  68
  69 static void __crst_table_upgrade(void *arg)
  70 {
  71         struct mm_struct *mm = arg;
  72
  73         if (current->active_mm == mm)
  74                 set_user_asce(mm);
  75         __tlb_flush_local();
  76 }
  77
  78 int crst_table_upgrade(struct mm_struct *mm, unsigned long end)
  79 {
  80         unsigned long *table, *pgd;
  81         int rc, notify;
  82
  83         /* upgrade should only happen from 3 to 4, 3 to 5, or 4 to 5 levels */
  84         VM_BUG_ON(mm->context.asce_limit < _REGION2_SIZE);
  85         rc = 0;
  86         notify = 0;
  87         while (mm->context.asce_limit < end) {
  88                 table = crst_table_alloc(mm);
  89                 if (!table) {
  90                         rc = -ENOMEM;
  91                         break;
  92                 }
  93                 spin_lock_bh(&mm->page_table_lock);
  94                 pgd = (unsigned long *) mm->pgd;
  95                 if (mm->context.asce_limit == _REGION2_SIZE) {
  96                         crst_table_init(table, _REGION2_ENTRY_EMPTY);
  97                         p4d_populate(mm, (p4d_t *) table, (pud_t *) pgd);
  98                         mm->pgd = (pgd_t *) table;
  99                         mm->context.asce_limit = _REGION1_SIZE;
 100                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 101                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION2;
 102                         mm_inc_nr_puds(mm);
 103                 } else {
 104                         crst_table_init(table, _REGION1_ENTRY_EMPTY);
 105                         pgd_populate(mm, (pgd_t *) table, (p4d_t *) pgd);
 106                         mm->pgd = (pgd_t *) table;
 107                         mm->context.asce_limit = -PAGE_SIZE;
 108                         mm->context.asce = __pa(mm->pgd) | _ASCE_TABLE_LENGTH |
 109                                 _ASCE_USER_BITS | _ASCE_TYPE_REGION1;
 110                 }
 111                 notify = 1;
 112                 spin_unlock_bh(&mm->page_table_lock);
 113         }
 114         if (notify)
 115                 on_each_cpu(__crst_table_upgrade, mm, 0);
 116         return rc;
 117 }
 118
 119 void crst_table_downgrade(struct mm_struct *mm)
 120 {
 121         pgd_t *pgd;
 122
 123         /* downgrade should only happen from 3 to 2 levels (compat only) */
 124         VM_BUG_ON(mm->context.asce_limit != _REGION2_SIZE);
 125
 126         if (current->active_mm == mm) {
 127                 clear_user_asce();
 128                 __tlb_flush_mm(mm);
 129         }
 130
 131         pgd = mm->pgd;
 132         mm_dec_nr_pmds(mm);
 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_pte_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         } else {
 256                 atomic_xor_bits(&page->_refcount, 3U << 24);
 257         }
 258
 259         pgtable_pte_page_dtor(page);
 260         __free_page(page);
 261 }
 262
 263 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table,
 264                          unsigned long vmaddr)
 265 {
 266         struct mm_struct *mm;
 267         struct page *page;
 268         unsigned int bit, mask;
 269
 270         mm = tlb->mm;
 271         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 272         if (mm_alloc_pgste(mm)) {
 273                 gmap_unlink(mm, table, vmaddr);
 274                 table = (unsigned long *) (__pa(table) | 3);
 275                 tlb_remove_table(tlb, table);
 276                 return;
 277         }
 278         bit = (__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t));
 279         spin_lock_bh(&mm->context.lock);
 280         mask = atomic_xor_bits(&page->_refcount, 0x11U << (bit + 24));
 281         mask >>= 24;
 282         if (mask & 3)
 283                 list_add_tail(&page->lru, &mm->context.pgtable_list);
 284         else
 285                 list_del(&page->lru);
 286         spin_unlock_bh(&mm->context.lock);
 287         table = (unsigned long *) (__pa(table) | (1U << bit));
 288         tlb_remove_table(tlb, table);
 289 }
 290
 291 void __tlb_remove_table(void *_table)
 292 {
 293         unsigned int mask = (unsigned long) _table & 3;
 294         void *table = (void *)((unsigned long) _table ^ mask);
 295         struct page *page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 296
 297         switch (mask) {
 298         case 0:         /* pmd, pud, or p4d */
 299                 free_pages((unsigned long) table, 2);
 300                 break;
 301         case 1:         /* lower 2K of a 4K page table */
 302         case 2:         /* higher 2K of a 4K page table */
 303                 mask = atomic_xor_bits(&page->_refcount, mask << (4 + 24));
 304                 mask >>= 24;
 305                 if (mask != 0)
 306                         break;
 307                 /* fallthrough */
 308         case 3:         /* 4K page table with pgstes */
 309                 if (mask & 3)
 310                         atomic_xor_bits(&page->_refcount, 3 << 24);
 311                 pgtable_pte_page_dtor(page);
 312                 __free_page(page);
 313                 break;
 314         }
 315 }
 316
 317 /*
 318  * Base infrastructure required to generate basic asces, region, segment,
 319  * and page tables that do not make use of enhanced features like EDAT1.
 320  */
 321
 322 static struct kmem_cache *base_pgt_cache;
 323
 324 static unsigned long base_pgt_alloc(void)
 325 {
 326         u64 *table;
 327
 328         table = kmem_cache_alloc(base_pgt_cache, GFP_KERNEL);
 329         if (table)
 330                 memset64(table, _PAGE_INVALID, PTRS_PER_PTE);
 331         return (unsigned long) table;
 332 }
 333
 334 static void base_pgt_free(unsigned long table)
 335 {
 336         kmem_cache_free(base_pgt_cache, (void *) table);
 337 }
 338
 339 static unsigned long base_crst_alloc(unsigned long val)
 340 {
 341         unsigned long table;
 342
 343         table =  __get_free_pages(GFP_KERNEL, CRST_ALLOC_ORDER);
 344         if (table)
 345                 crst_table_init((unsigned long *)table, val);
 346         return table;
 347 }
 348
 349 static void base_crst_free(unsigned long table)
 350 {
 351         free_pages(table, CRST_ALLOC_ORDER);
 352 }
 353
 354 #define BASE_ADDR_END_FUNC(NAME, SIZE)                                  \
 355 static inline unsigned long base_##NAME##_addr_end(unsigned long addr,  \
 356                                                    unsigned long end)   \
 357 {                                                                       \
 358         unsigned long next = (addr + (SIZE)) & ~((SIZE) - 1);           \
 359                                                                         \
 360         return (next - 1) < (end - 1) ? next : end;                     \
 361 }
 362
 363 BASE_ADDR_END_FUNC(page,    _PAGE_SIZE)
 364 BASE_ADDR_END_FUNC(segment, _SEGMENT_SIZE)
 365 BASE_ADDR_END_FUNC(region3, _REGION3_SIZE)
 366 BASE_ADDR_END_FUNC(region2, _REGION2_SIZE)
 367 BASE_ADDR_END_FUNC(region1, _REGION1_SIZE)
 368
 369 static inline unsigned long base_lra(unsigned long address)
 370 {
 371         unsigned long real;
 372
 373         asm volatile(
 374                 "       lra     %0,0(%1)\n"
 375                 : "=d" (real) : "a" (address) : "cc");
 376         return real;
 377 }
 378
 379 static int base_page_walk(unsigned long origin, unsigned long addr,
 380                           unsigned long end, int alloc)
 381 {
 382         unsigned long *pte, next;
 383
 384         if (!alloc)
 385                 return 0;
 386         pte = (unsigned long *) origin;
 387         pte += (addr & _PAGE_INDEX) >> _PAGE_SHIFT;
 388         do {
 389                 next = base_page_addr_end(addr, end);
 390                 *pte = base_lra(addr);
 391         } while (pte++, addr = next, addr < end);
 392         return 0;
 393 }
 394
 395 static int base_segment_walk(unsigned long origin, unsigned long addr,
 396                              unsigned long end, int alloc)
 397 {
 398         unsigned long *ste, next, table;
 399         int rc;
 400
 401         ste = (unsigned long *) origin;
 402         ste += (addr & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 403         do {
 404                 next = base_segment_addr_end(addr, end);
 405                 if (*ste & _SEGMENT_ENTRY_INVALID) {
 406                         if (!alloc)
 407                                 continue;
 408                         table = base_pgt_alloc();
 409                         if (!table)
 410                                 return -ENOMEM;
 411                         *ste = table | _SEGMENT_ENTRY;
 412                 }
 413                 table = *ste & _SEGMENT_ENTRY_ORIGIN;
 414                 rc = base_page_walk(table, addr, next, alloc);
 415                 if (rc)
 416                         return rc;
 417                 if (!alloc)
 418                         base_pgt_free(table);
 419                 cond_resched();
 420         } while (ste++, addr = next, addr < end);
 421         return 0;
 422 }
 423
 424 static int base_region3_walk(unsigned long origin, unsigned long addr,
 425                              unsigned long end, int alloc)
 426 {
 427         unsigned long *rtte, next, table;
 428         int rc;
 429
 430         rtte = (unsigned long *) origin;
 431         rtte += (addr & _REGION3_INDEX) >> _REGION3_SHIFT;
 432         do {
 433                 next = base_region3_addr_end(addr, end);
 434                 if (*rtte & _REGION_ENTRY_INVALID) {
 435                         if (!alloc)
 436                                 continue;
 437                         table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
 438                         if (!table)
 439                                 return -ENOMEM;
 440                         *rtte = table | _REGION3_ENTRY;
 441                 }
 442                 table = *rtte & _REGION_ENTRY_ORIGIN;
 443                 rc = base_segment_walk(table, addr, next, alloc);
 444                 if (rc)
 445                         return rc;
 446                 if (!alloc)
 447                         base_crst_free(table);
 448         } while (rtte++, addr = next, addr < end);
 449         return 0;
 450 }
 451
 452 static int base_region2_walk(unsigned long origin, unsigned long addr,
 453                              unsigned long end, int alloc)
 454 {
 455         unsigned long *rste, next, table;
 456         int rc;
 457
 458         rste = (unsigned long *) origin;
 459         rste += (addr & _REGION2_INDEX) >> _REGION2_SHIFT;
 460         do {
 461                 next = base_region2_addr_end(addr, end);
 462                 if (*rste & _REGION_ENTRY_INVALID) {
 463                         if (!alloc)
 464                                 continue;
 465                         table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
 466                         if (!table)
 467                                 return -ENOMEM;
 468                         *rste = table | _REGION2_ENTRY;
 469                 }
 470                 table = *rste & _REGION_ENTRY_ORIGIN;
 471                 rc = base_region3_walk(table, addr, next, alloc);
 472                 if (rc)
 473                         return rc;
 474                 if (!alloc)
 475                         base_crst_free(table);
 476         } while (rste++, addr = next, addr < end);
 477         return 0;
 478 }
 479
 480 static int base_region1_walk(unsigned long origin, unsigned long addr,
 481                              unsigned long end, int alloc)
 482 {
 483         unsigned long *rfte, next, table;
 484         int rc;
 485
 486         rfte = (unsigned long *) origin;
 487         rfte += (addr & _REGION1_INDEX) >> _REGION1_SHIFT;
 488         do {
 489                 next = base_region1_addr_end(addr, end);
 490                 if (*rfte & _REGION_ENTRY_INVALID) {
 491                         if (!alloc)
 492                                 continue;
 493                         table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
 494                         if (!table)
 495                                 return -ENOMEM;
 496                         *rfte = table | _REGION1_ENTRY;
 497                 }
 498                 table = *rfte & _REGION_ENTRY_ORIGIN;
 499                 rc = base_region2_walk(table, addr, next, alloc);
 500                 if (rc)
 501                         return rc;
 502                 if (!alloc)
 503                         base_crst_free(table);
 504         } while (rfte++, addr = next, addr < end);
 505         return 0;
 506 }
 507
 508 /**
 509  * base_asce_free - free asce and tables returned from base_asce_alloc()
 510  * @asce: asce to be freed
 511  *
 512  * Frees all region, segment, and page tables that were allocated with a
 513  * corresponding base_asce_alloc() call.
 514  */
 515 void base_asce_free(unsigned long asce)
 516 {
 517         unsigned long table = asce & _ASCE_ORIGIN;
 518
 519         if (!asce)
 520                 return;
 521         switch (asce & _ASCE_TYPE_MASK) {
 522         case _ASCE_TYPE_SEGMENT:
 523                 base_segment_walk(table, 0, _REGION3_SIZE, 0);
 524                 break;
 525         case _ASCE_TYPE_REGION3:
 526                 base_region3_walk(table, 0, _REGION2_SIZE, 0);
 527                 break;
 528         case _ASCE_TYPE_REGION2:
 529                 base_region2_walk(table, 0, _REGION1_SIZE, 0);
 530                 break;
 531         case _ASCE_TYPE_REGION1:
 532                 base_region1_walk(table, 0, -_PAGE_SIZE, 0);
 533                 break;
 534         }
 535         base_crst_free(table);
 536 }
 537
 538 static int base_pgt_cache_init(void)
 539 {
 540         static DEFINE_MUTEX(base_pgt_cache_mutex);
 541         unsigned long sz = _PAGE_TABLE_SIZE;
 542
 543         if (base_pgt_cache)
 544                 return 0;
 545         mutex_lock(&base_pgt_cache_mutex);
 546         if (!base_pgt_cache)
 547                 base_pgt_cache = kmem_cache_create("base_pgt", sz, sz, 0, NULL);
 548         mutex_unlock(&base_pgt_cache_mutex);
 549         return base_pgt_cache ? 0 : -ENOMEM;
 550 }
 551
 552 /**
 553  * base_asce_alloc - create kernel mapping without enhanced DAT features
 554  * @addr: virtual start address of kernel mapping
 555  * @num_pages: number of consecutive pages
 556  *
 557  * Generate an asce, including all required region, segment and page tables,
 558  * that can be used to access the virtual kernel mapping. The difference is
 559  * that the returned asce does not make use of any enhanced DAT features like
 560  * e.g. large pages. This is required for some I/O functions that pass an
 561  * asce, like e.g. some service call requests.
 562  *
 563  * Note: the returned asce may NEVER be attached to any cpu. It may only be
 564  *       used for I/O requests. tlb entries that might result because the
 565  *       asce was attached to a cpu won't be cleared.
 566  */
 567 unsigned long base_asce_alloc(unsigned long addr, unsigned long num_pages)
 568 {
 569         unsigned long asce, table, end;
 570         int rc;
 571
 572         if (base_pgt_cache_init())
 573                 return 0;
 574         end = addr + num_pages * PAGE_SIZE;
 575         if (end <= _REGION3_SIZE) {
 576                 table = base_crst_alloc(_SEGMENT_ENTRY_EMPTY);
 577                 if (!table)
 578                         return 0;
 579                 rc = base_segment_walk(table, addr, end, 1);
 580                 asce = table | _ASCE_TYPE_SEGMENT | _ASCE_TABLE_LENGTH;
 581         } else if (end <= _REGION2_SIZE) {
 582                 table = base_crst_alloc(_REGION3_ENTRY_EMPTY);
 583                 if (!table)
 584                         return 0;
 585                 rc = base_region3_walk(table, addr, end, 1);
 586                 asce = table | _ASCE_TYPE_REGION3 | _ASCE_TABLE_LENGTH;
 587         } else if (end <= _REGION1_SIZE) {
 588                 table = base_crst_alloc(_REGION2_ENTRY_EMPTY);
 589                 if (!table)
 590                         return 0;
 591                 rc = base_region2_walk(table, addr, end, 1);
 592                 asce = table | _ASCE_TYPE_REGION2 | _ASCE_TABLE_LENGTH;
 593         } else {
 594                 table = base_crst_alloc(_REGION1_ENTRY_EMPTY);
 595                 if (!table)
 596                         return 0;
 597                 rc = base_region1_walk(table, addr, end, 1);
 598                 asce = table | _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH;
 599         }
 600         if (rc) {
 601                 base_asce_free(asce);
 602                 asce = 0;
 603         }
 604         return asce;
 605 }