arch/s390/mm/pgtable.c

   1 /*
   2  *    Copyright IBM Corp. 2007, 2011
   3  *    Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/kernel.h>
   8 #include <linux/errno.h>
   9 #include <linux/gfp.h>
  10 #include <linux/mm.h>
  11 #include <linux/swap.h>
  12 #include <linux/smp.h>
  13 #include <linux/highmem.h>
  14 #include <linux/pagemap.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/module.h>
  17 #include <linux/quicklist.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/slab.h>
  20
  21 #include <asm/pgtable.h>
  22 #include <asm/pgalloc.h>
  23 #include <asm/tlb.h>
  24 #include <asm/tlbflush.h>
  25 #include <asm/mmu_context.h>
  26
  27 #ifndef CONFIG_64BIT
  28 #define ALLOC_ORDER     1
  29 #define FRAG_MASK       0x0f
  30 #else
  31 #define ALLOC_ORDER     2
  32 #define FRAG_MASK       0x03
  33 #endif
  34
  35
  36 unsigned long *crst_table_alloc(struct mm_struct *mm)
  37 {
  38         struct page *page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
  39
  40         if (!page)
  41                 return NULL;
  42         return (unsigned long *) page_to_phys(page);
  43 }
  44
  45 void crst_table_free(struct mm_struct *mm, unsigned long *table)
  46 {
  47         free_pages((unsigned long) table, ALLOC_ORDER);
  48 }
  49
  50 #ifdef CONFIG_64BIT
  51 int crst_table_upgrade(struct mm_struct *mm, unsigned long limit)
  52 {
  53         unsigned long *table, *pgd;
  54         unsigned long entry;
  55
  56         BUG_ON(limit > (1UL << 53));
  57 repeat:
  58         table = crst_table_alloc(mm);
  59         if (!table)
  60                 return -ENOMEM;
  61         spin_lock_bh(&mm->page_table_lock);
  62         if (mm->context.asce_limit < limit) {
  63                 pgd = (unsigned long *) mm->pgd;
  64                 if (mm->context.asce_limit <= (1UL << 31)) {
  65                         entry = _REGION3_ENTRY_EMPTY;
  66                         mm->context.asce_limit = 1UL << 42;
  67                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
  68                                                 _ASCE_USER_BITS |
  69                                                 _ASCE_TYPE_REGION3;
  70                 } else {
  71                         entry = _REGION2_ENTRY_EMPTY;
  72                         mm->context.asce_limit = 1UL << 53;
  73                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
  74                                                 _ASCE_USER_BITS |
  75                                                 _ASCE_TYPE_REGION2;
  76                 }
  77                 crst_table_init(table, entry);
  78                 pgd_populate(mm, (pgd_t *) table, (pud_t *) pgd);
  79                 mm->pgd = (pgd_t *) table;
  80                 mm->task_size = mm->context.asce_limit;
  81                 table = NULL;
  82         }
  83         spin_unlock_bh(&mm->page_table_lock);
  84         if (table)
  85                 crst_table_free(mm, table);
  86         if (mm->context.asce_limit < limit)
  87                 goto repeat;
  88         return 0;
  89 }
  90
  91 void crst_table_downgrade(struct mm_struct *mm, unsigned long limit)
  92 {
  93         pgd_t *pgd;
  94
  95         while (mm->context.asce_limit > limit) {
  96                 pgd = mm->pgd;
  97                 switch (pgd_val(*pgd) & _REGION_ENTRY_TYPE_MASK) {
  98                 case _REGION_ENTRY_TYPE_R2:
  99                         mm->context.asce_limit = 1UL << 42;
 100                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
 101                                                 _ASCE_USER_BITS |
 102                                                 _ASCE_TYPE_REGION3;
 103                         break;
 104                 case _REGION_ENTRY_TYPE_R3:
 105                         mm->context.asce_limit = 1UL << 31;
 106                         mm->context.asce_bits = _ASCE_TABLE_LENGTH |
 107                                                 _ASCE_USER_BITS |
 108                                                 _ASCE_TYPE_SEGMENT;
 109                         break;
 110                 default:
 111                         BUG();
 112                 }
 113                 mm->pgd = (pgd_t *) (pgd_val(*pgd) & _REGION_ENTRY_ORIGIN);
 114                 mm->task_size = mm->context.asce_limit;
 115                 crst_table_free(mm, (unsigned long *) pgd);
 116         }
 117 }
 118 #endif
 119
 120 #ifdef CONFIG_PGSTE
 121
 122 /**
 123  * gmap_alloc - allocate a guest address space
 124  * @mm: pointer to the parent mm_struct
 125  *
 126  * Returns a guest address space structure.
 127  */
 128 struct gmap *gmap_alloc(struct mm_struct *mm)
 129 {
 130         struct gmap *gmap;
 131         struct page *page;
 132         unsigned long *table;
 133
 134         gmap = kzalloc(sizeof(struct gmap), GFP_KERNEL);
 135         if (!gmap)
 136                 goto out;
 137         INIT_LIST_HEAD(&gmap->crst_list);
 138         gmap->mm = mm;
 139         page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
 140         if (!page)
 141                 goto out_free;
 142         list_add(&page->lru, &gmap->crst_list);
 143         table = (unsigned long *) page_to_phys(page);
 144         crst_table_init(table, _REGION1_ENTRY_EMPTY);
 145         gmap->table = table;
 146         gmap->asce = _ASCE_TYPE_REGION1 | _ASCE_TABLE_LENGTH |
 147                      _ASCE_USER_BITS | __pa(table);
 148         list_add(&gmap->list, &mm->context.gmap_list);
 149         return gmap;
 150
 151 out_free:
 152         kfree(gmap);
 153 out:
 154         return NULL;
 155 }
 156 EXPORT_SYMBOL_GPL(gmap_alloc);
 157
 158 static int gmap_unlink_segment(struct gmap *gmap, unsigned long *table)
 159 {
 160         struct gmap_pgtable *mp;
 161         struct gmap_rmap *rmap;
 162         struct page *page;
 163
 164         if (*table & _SEGMENT_ENTRY_INV)
 165                 return 0;
 166         page = pfn_to_page(*table >> PAGE_SHIFT);
 167         mp = (struct gmap_pgtable *) page->index;
 168         list_for_each_entry(rmap, &mp->mapper, list) {
 169                 if (rmap->entry != table)
 170                         continue;
 171                 list_del(&rmap->list);
 172                 kfree(rmap);
 173                 break;
 174         }
 175         *table = _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | mp->vmaddr;
 176         return 1;
 177 }
 178
 179 static void gmap_flush_tlb(struct gmap *gmap)
 180 {
 181         if (MACHINE_HAS_IDTE)
 182                 __tlb_flush_idte((unsigned long) gmap->table |
 183                                  _ASCE_TYPE_REGION1);
 184         else
 185                 __tlb_flush_global();
 186 }
 187
 188 /**
 189  * gmap_free - free a guest address space
 190  * @gmap: pointer to the guest address space structure
 191  */
 192 void gmap_free(struct gmap *gmap)
 193 {
 194         struct page *page, *next;
 195         unsigned long *table;
 196         int i;
 197
 198
 199         /* Flush tlb. */
 200         if (MACHINE_HAS_IDTE)
 201                 __tlb_flush_idte((unsigned long) gmap->table |
 202                                  _ASCE_TYPE_REGION1);
 203         else
 204                 __tlb_flush_global();
 205
 206         /* Free all segment & region tables. */
 207         down_read(&gmap->mm->mmap_sem);
 208         spin_lock(&gmap->mm->page_table_lock);
 209         list_for_each_entry_safe(page, next, &gmap->crst_list, lru) {
 210                 table = (unsigned long *) page_to_phys(page);
 211                 if ((*table & _REGION_ENTRY_TYPE_MASK) == 0)
 212                         /* Remove gmap rmap structures for segment table. */
 213                         for (i = 0; i < PTRS_PER_PMD; i++, table++)
 214                                 gmap_unlink_segment(gmap, table);
 215                 __free_pages(page, ALLOC_ORDER);
 216         }
 217         spin_unlock(&gmap->mm->page_table_lock);
 218         up_read(&gmap->mm->mmap_sem);
 219         list_del(&gmap->list);
 220         kfree(gmap);
 221 }
 222 EXPORT_SYMBOL_GPL(gmap_free);
 223
 224 /**
 225  * gmap_enable - switch primary space to the guest address space
 226  * @gmap: pointer to the guest address space structure
 227  */
 228 void gmap_enable(struct gmap *gmap)
 229 {
 230         S390_lowcore.gmap = (unsigned long) gmap;
 231 }
 232 EXPORT_SYMBOL_GPL(gmap_enable);
 233
 234 /**
 235  * gmap_disable - switch back to the standard primary address space
 236  * @gmap: pointer to the guest address space structure
 237  */
 238 void gmap_disable(struct gmap *gmap)
 239 {
 240         S390_lowcore.gmap = 0UL;
 241 }
 242 EXPORT_SYMBOL_GPL(gmap_disable);
 243
 244 /*
 245  * gmap_alloc_table is assumed to be called with mmap_sem held
 246  */
 247 static int gmap_alloc_table(struct gmap *gmap,
 248                                unsigned long *table, unsigned long init)
 249 {
 250         struct page *page;
 251         unsigned long *new;
 252
 253         /* since we dont free the gmap table until gmap_free we can unlock */
 254         spin_unlock(&gmap->mm->page_table_lock);
 255         page = alloc_pages(GFP_KERNEL, ALLOC_ORDER);
 256         spin_lock(&gmap->mm->page_table_lock);
 257         if (!page)
 258                 return -ENOMEM;
 259         new = (unsigned long *) page_to_phys(page);
 260         crst_table_init(new, init);
 261         if (*table & _REGION_ENTRY_INV) {
 262                 list_add(&page->lru, &gmap->crst_list);
 263                 *table = (unsigned long) new | _REGION_ENTRY_LENGTH |
 264                         (*table & _REGION_ENTRY_TYPE_MASK);
 265         } else
 266                 __free_pages(page, ALLOC_ORDER);
 267         return 0;
 268 }
 269
 270 /**
 271  * gmap_unmap_segment - unmap segment from the guest address space
 272  * @gmap: pointer to the guest address space structure
 273  * @addr: address in the guest address space
 274  * @len: length of the memory area to unmap
 275  *
 276  * Returns 0 if the unmap succeded, -EINVAL if not.
 277  */
 278 int gmap_unmap_segment(struct gmap *gmap, unsigned long to, unsigned long len)
 279 {
 280         unsigned long *table;
 281         unsigned long off;
 282         int flush;
 283
 284         if ((to | len) & (PMD_SIZE - 1))
 285                 return -EINVAL;
 286         if (len == 0 || to + len < to)
 287                 return -EINVAL;
 288
 289         flush = 0;
 290         down_read(&gmap->mm->mmap_sem);
 291         spin_lock(&gmap->mm->page_table_lock);
 292         for (off = 0; off < len; off += PMD_SIZE) {
 293                 /* Walk the guest addr space page table */
 294                 table = gmap->table + (((to + off) >> 53) & 0x7ff);
 295                 if (*table & _REGION_ENTRY_INV)
 296                         goto out;
 297                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 298                 table = table + (((to + off) >> 42) & 0x7ff);
 299                 if (*table & _REGION_ENTRY_INV)
 300                         goto out;
 301                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 302                 table = table + (((to + off) >> 31) & 0x7ff);
 303                 if (*table & _REGION_ENTRY_INV)
 304                         goto out;
 305                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 306                 table = table + (((to + off) >> 20) & 0x7ff);
 307
 308                 /* Clear segment table entry in guest address space. */
 309                 flush |= gmap_unlink_segment(gmap, table);
 310                 *table = _SEGMENT_ENTRY_INV;
 311         }
 312 out:
 313         spin_unlock(&gmap->mm->page_table_lock);
 314         up_read(&gmap->mm->mmap_sem);
 315         if (flush)
 316                 gmap_flush_tlb(gmap);
 317         return 0;
 318 }
 319 EXPORT_SYMBOL_GPL(gmap_unmap_segment);
 320
 321 /**
 322  * gmap_mmap_segment - map a segment to the guest address space
 323  * @gmap: pointer to the guest address space structure
 324  * @from: source address in the parent address space
 325  * @to: target address in the guest address space
 326  *
 327  * Returns 0 if the mmap succeded, -EINVAL or -ENOMEM if not.
 328  */
 329 int gmap_map_segment(struct gmap *gmap, unsigned long from,
 330                      unsigned long to, unsigned long len)
 331 {
 332         unsigned long *table;
 333         unsigned long off;
 334         int flush;
 335
 336         if ((from | to | len) & (PMD_SIZE - 1))
 337                 return -EINVAL;
 338         if (len == 0 || from + len > PGDIR_SIZE ||
 339             from + len < from || to + len < to)
 340                 return -EINVAL;
 341
 342         flush = 0;
 343         down_read(&gmap->mm->mmap_sem);
 344         spin_lock(&gmap->mm->page_table_lock);
 345         for (off = 0; off < len; off += PMD_SIZE) {
 346                 /* Walk the gmap address space page table */
 347                 table = gmap->table + (((to + off) >> 53) & 0x7ff);
 348                 if ((*table & _REGION_ENTRY_INV) &&
 349                     gmap_alloc_table(gmap, table, _REGION2_ENTRY_EMPTY))
 350                         goto out_unmap;
 351                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 352                 table = table + (((to + off) >> 42) & 0x7ff);
 353                 if ((*table & _REGION_ENTRY_INV) &&
 354                     gmap_alloc_table(gmap, table, _REGION3_ENTRY_EMPTY))
 355                         goto out_unmap;
 356                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 357                 table = table + (((to + off) >> 31) & 0x7ff);
 358                 if ((*table & _REGION_ENTRY_INV) &&
 359                     gmap_alloc_table(gmap, table, _SEGMENT_ENTRY_EMPTY))
 360                         goto out_unmap;
 361                 table = (unsigned long *) (*table & _REGION_ENTRY_ORIGIN);
 362                 table = table + (((to + off) >> 20) & 0x7ff);
 363
 364                 /* Store 'from' address in an invalid segment table entry. */
 365                 flush |= gmap_unlink_segment(gmap, table);
 366                 *table = _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | (from + off);
 367         }
 368         spin_unlock(&gmap->mm->page_table_lock);
 369         up_read(&gmap->mm->mmap_sem);
 370         if (flush)
 371                 gmap_flush_tlb(gmap);
 372         return 0;
 373
 374 out_unmap:
 375         spin_unlock(&gmap->mm->page_table_lock);
 376         up_read(&gmap->mm->mmap_sem);
 377         gmap_unmap_segment(gmap, to, len);
 378         return -ENOMEM;
 379 }
 380 EXPORT_SYMBOL_GPL(gmap_map_segment);
 381
 382 static unsigned long *gmap_table_walk(unsigned long address, struct gmap *gmap)
 383 {
 384         unsigned long *table;
 385
 386         table = gmap->table + ((address >> 53) & 0x7ff);
 387         if (unlikely(*table & _REGION_ENTRY_INV))
 388                 return ERR_PTR(-EFAULT);
 389         table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 390         table = table + ((address >> 42) & 0x7ff);
 391         if (unlikely(*table & _REGION_ENTRY_INV))
 392                 return ERR_PTR(-EFAULT);
 393         table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 394         table = table + ((address >> 31) & 0x7ff);
 395         if (unlikely(*table & _REGION_ENTRY_INV))
 396                 return ERR_PTR(-EFAULT);
 397         table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 398         table = table + ((address >> 20) & 0x7ff);
 399         return table;
 400 }
 401
 402 /**
 403  * __gmap_translate - translate a guest address to a user space address
 404  * @address: guest address
 405  * @gmap: pointer to guest mapping meta data structure
 406  *
 407  * Returns user space address which corresponds to the guest address or
 408  * -EFAULT if no such mapping exists.
 409  * This function does not establish potentially missing page table entries.
 410  * The mmap_sem of the mm that belongs to the address space must be held
 411  * when this function gets called.
 412  */
 413 unsigned long __gmap_translate(unsigned long address, struct gmap *gmap)
 414 {
 415         unsigned long *segment_ptr, vmaddr, segment;
 416         struct gmap_pgtable *mp;
 417         struct page *page;
 418
 419         current->thread.gmap_addr = address;
 420         segment_ptr = gmap_table_walk(address, gmap);
 421         if (IS_ERR(segment_ptr))
 422                 return PTR_ERR(segment_ptr);
 423         /* Convert the gmap address to an mm address. */
 424         segment = *segment_ptr;
 425         if (!(segment & _SEGMENT_ENTRY_INV)) {
 426                 page = pfn_to_page(segment >> PAGE_SHIFT);
 427                 mp = (struct gmap_pgtable *) page->index;
 428                 return mp->vmaddr | (address & ~PMD_MASK);
 429         } else if (segment & _SEGMENT_ENTRY_RO) {
 430                 vmaddr = segment & _SEGMENT_ENTRY_ORIGIN;
 431                 return vmaddr | (address & ~PMD_MASK);
 432         }
 433         return -EFAULT;
 434 }
 435 EXPORT_SYMBOL_GPL(__gmap_translate);
 436
 437 /**
 438  * gmap_translate - translate a guest address to a user space address
 439  * @address: guest address
 440  * @gmap: pointer to guest mapping meta data structure
 441  *
 442  * Returns user space address which corresponds to the guest address or
 443  * -EFAULT if no such mapping exists.
 444  * This function does not establish potentially missing page table entries.
 445  */
 446 unsigned long gmap_translate(unsigned long address, struct gmap *gmap)
 447 {
 448         unsigned long rc;
 449
 450         down_read(&gmap->mm->mmap_sem);
 451         rc = __gmap_translate(address, gmap);
 452         up_read(&gmap->mm->mmap_sem);
 453         return rc;
 454 }
 455 EXPORT_SYMBOL_GPL(gmap_translate);
 456
 457 static int gmap_connect_pgtable(unsigned long segment,
 458                                 unsigned long *segment_ptr,
 459                                 struct gmap *gmap)
 460 {
 461         unsigned long vmaddr;
 462         struct vm_area_struct *vma;
 463         struct gmap_pgtable *mp;
 464         struct gmap_rmap *rmap;
 465         struct mm_struct *mm;
 466         struct page *page;
 467         pgd_t *pgd;
 468         pud_t *pud;
 469         pmd_t *pmd;
 470
 471         mm = gmap->mm;
 472         vmaddr = segment & _SEGMENT_ENTRY_ORIGIN;
 473         vma = find_vma(mm, vmaddr);
 474         if (!vma || vma->vm_start > vmaddr)
 475                 return -EFAULT;
 476         /* Walk the parent mm page table */
 477         pgd = pgd_offset(mm, vmaddr);
 478         pud = pud_alloc(mm, pgd, vmaddr);
 479         if (!pud)
 480                 return -ENOMEM;
 481         pmd = pmd_alloc(mm, pud, vmaddr);
 482         if (!pmd)
 483                 return -ENOMEM;
 484         if (!pmd_present(*pmd) &&
 485             __pte_alloc(mm, vma, pmd, vmaddr))
 486                 return -ENOMEM;
 487         /* pmd now points to a valid segment table entry. */
 488         rmap = kmalloc(sizeof(*rmap), GFP_KERNEL|__GFP_REPEAT);
 489         if (!rmap)
 490                 return -ENOMEM;
 491         /* Link gmap segment table entry location to page table. */
 492         page = pmd_page(*pmd);
 493         mp = (struct gmap_pgtable *) page->index;
 494         rmap->entry = segment_ptr;
 495         spin_lock(&mm->page_table_lock);
 496         if (*segment_ptr == segment) {
 497                 list_add(&rmap->list, &mp->mapper);
 498                 /* Set gmap segment table entry to page table. */
 499                 *segment_ptr = pmd_val(*pmd) & PAGE_MASK;
 500                 rmap = NULL;
 501         }
 502         spin_unlock(&mm->page_table_lock);
 503         kfree(rmap);
 504         return 0;
 505 }
 506
 507 static void gmap_disconnect_pgtable(struct mm_struct *mm, unsigned long *table)
 508 {
 509         struct gmap_rmap *rmap, *next;
 510         struct gmap_pgtable *mp;
 511         struct page *page;
 512         int flush;
 513
 514         flush = 0;
 515         spin_lock(&mm->page_table_lock);
 516         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 517         mp = (struct gmap_pgtable *) page->index;
 518         list_for_each_entry_safe(rmap, next, &mp->mapper, list) {
 519                 *rmap->entry =
 520                         _SEGMENT_ENTRY_INV | _SEGMENT_ENTRY_RO | mp->vmaddr;
 521                 list_del(&rmap->list);
 522                 kfree(rmap);
 523                 flush = 1;
 524         }
 525         spin_unlock(&mm->page_table_lock);
 526         if (flush)
 527                 __tlb_flush_global();
 528 }
 529
 530 /*
 531  * this function is assumed to be called with mmap_sem held
 532  */
 533 unsigned long __gmap_fault(unsigned long address, struct gmap *gmap)
 534 {
 535         unsigned long *segment_ptr, segment;
 536         struct gmap_pgtable *mp;
 537         struct page *page;
 538         int rc;
 539
 540         current->thread.gmap_addr = address;
 541         segment_ptr = gmap_table_walk(address, gmap);
 542         if (IS_ERR(segment_ptr))
 543                 return -EFAULT;
 544         /* Convert the gmap address to an mm address. */
 545         while (1) {
 546                 segment = *segment_ptr;
 547                 if (!(segment & _SEGMENT_ENTRY_INV)) {
 548                         /* Page table is present */
 549                         page = pfn_to_page(segment >> PAGE_SHIFT);
 550                         mp = (struct gmap_pgtable *) page->index;
 551                         return mp->vmaddr | (address & ~PMD_MASK);
 552                 }
 553                 if (!(segment & _SEGMENT_ENTRY_RO))
 554                         /* Nothing mapped in the gmap address space. */
 555                         break;
 556                 rc = gmap_connect_pgtable(segment, segment_ptr, gmap);
 557                 if (rc)
 558                         return rc;
 559         }
 560         return -EFAULT;
 561 }
 562
 563 unsigned long gmap_fault(unsigned long address, struct gmap *gmap)
 564 {
 565         unsigned long rc;
 566
 567         down_read(&gmap->mm->mmap_sem);
 568         rc = __gmap_fault(address, gmap);
 569         up_read(&gmap->mm->mmap_sem);
 570
 571         return rc;
 572 }
 573 EXPORT_SYMBOL_GPL(gmap_fault);
 574
 575 void gmap_discard(unsigned long from, unsigned long to, struct gmap *gmap)
 576 {
 577
 578         unsigned long *table, address, size;
 579         struct vm_area_struct *vma;
 580         struct gmap_pgtable *mp;
 581         struct page *page;
 582
 583         down_read(&gmap->mm->mmap_sem);
 584         address = from;
 585         while (address < to) {
 586                 /* Walk the gmap address space page table */
 587                 table = gmap->table + ((address >> 53) & 0x7ff);
 588                 if (unlikely(*table & _REGION_ENTRY_INV)) {
 589                         address = (address + PMD_SIZE) & PMD_MASK;
 590                         continue;
 591                 }
 592                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 593                 table = table + ((address >> 42) & 0x7ff);
 594                 if (unlikely(*table & _REGION_ENTRY_INV)) {
 595                         address = (address + PMD_SIZE) & PMD_MASK;
 596                         continue;
 597                 }
 598                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 599                 table = table + ((address >> 31) & 0x7ff);
 600                 if (unlikely(*table & _REGION_ENTRY_INV)) {
 601                         address = (address + PMD_SIZE) & PMD_MASK;
 602                         continue;
 603                 }
 604                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 605                 table = table + ((address >> 20) & 0x7ff);
 606                 if (unlikely(*table & _SEGMENT_ENTRY_INV)) {
 607                         address = (address + PMD_SIZE) & PMD_MASK;
 608                         continue;
 609                 }
 610                 page = pfn_to_page(*table >> PAGE_SHIFT);
 611                 mp = (struct gmap_pgtable *) page->index;
 612                 vma = find_vma(gmap->mm, mp->vmaddr);
 613                 size = min(to - address, PMD_SIZE - (address & ~PMD_MASK));
 614                 zap_page_range(vma, mp->vmaddr | (address & ~PMD_MASK),
 615                                size, NULL);
 616                 address = (address + PMD_SIZE) & PMD_MASK;
 617         }
 618         up_read(&gmap->mm->mmap_sem);
 619 }
 620 EXPORT_SYMBOL_GPL(gmap_discard);
 621
 622 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm,
 623                                                     unsigned long vmaddr)
 624 {
 625         struct page *page;
 626         unsigned long *table;
 627         struct gmap_pgtable *mp;
 628
 629         page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
 630         if (!page)
 631                 return NULL;
 632         mp = kmalloc(sizeof(*mp), GFP_KERNEL|__GFP_REPEAT);
 633         if (!mp) {
 634                 __free_page(page);
 635                 return NULL;
 636         }
 637         pgtable_page_ctor(page);
 638         mp->vmaddr = vmaddr & PMD_MASK;
 639         INIT_LIST_HEAD(&mp->mapper);
 640         page->index = (unsigned long) mp;
 641         atomic_set(&page->_mapcount, 3);
 642         table = (unsigned long *) page_to_phys(page);
 643         clear_table(table, _PAGE_TYPE_EMPTY, PAGE_SIZE/2);
 644         clear_table(table + PTRS_PER_PTE, 0, PAGE_SIZE/2);
 645         return table;
 646 }
 647
 648 static inline void page_table_free_pgste(unsigned long *table)
 649 {
 650         struct page *page;
 651         struct gmap_pgtable *mp;
 652
 653         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 654         mp = (struct gmap_pgtable *) page->index;
 655         BUG_ON(!list_empty(&mp->mapper));
 656         pgtable_page_dtor(page);
 657         atomic_set(&page->_mapcount, -1);
 658         kfree(mp);
 659         __free_page(page);
 660 }
 661
 662 #else /* CONFIG_PGSTE */
 663
 664 static inline unsigned long *page_table_alloc_pgste(struct mm_struct *mm,
 665                                                     unsigned long vmaddr)
 666 {
 667         return NULL;
 668 }
 669
 670 static inline void page_table_free_pgste(unsigned long *table)
 671 {
 672 }
 673
 674 static inline void gmap_disconnect_pgtable(struct mm_struct *mm,
 675                                            unsigned long *table)
 676 {
 677 }
 678
 679 #endif /* CONFIG_PGSTE */
 680
 681 static inline unsigned int atomic_xor_bits(atomic_t *v, unsigned int bits)
 682 {
 683         unsigned int old, new;
 684
 685         do {
 686                 old = atomic_read(v);
 687                 new = old ^ bits;
 688         } while (atomic_cmpxchg(v, old, new) != old);
 689         return new;
 690 }
 691
 692 /*
 693  * page table entry allocation/free routines.
 694  */
 695 unsigned long *page_table_alloc(struct mm_struct *mm, unsigned long vmaddr)
 696 {
 697         unsigned long *uninitialized_var(table);
 698         struct page *uninitialized_var(page);
 699         unsigned int mask, bit;
 700
 701         if (mm_has_pgste(mm))
 702                 return page_table_alloc_pgste(mm, vmaddr);
 703         /* Allocate fragments of a 4K page as 1K/2K page table */
 704         spin_lock_bh(&mm->context.list_lock);
 705         mask = FRAG_MASK;
 706         if (!list_empty(&mm->context.pgtable_list)) {
 707                 page = list_first_entry(&mm->context.pgtable_list,
 708                                         struct page, lru);
 709                 table = (unsigned long *) page_to_phys(page);
 710                 mask = atomic_read(&page->_mapcount);
 711                 mask = mask | (mask >> 4);
 712         }
 713         if ((mask & FRAG_MASK) == FRAG_MASK) {
 714                 spin_unlock_bh(&mm->context.list_lock);
 715                 page = alloc_page(GFP_KERNEL|__GFP_REPEAT);
 716                 if (!page)
 717                         return NULL;
 718                 pgtable_page_ctor(page);
 719                 atomic_set(&page->_mapcount, 1);
 720                 table = (unsigned long *) page_to_phys(page);
 721                 clear_table(table, _PAGE_TYPE_EMPTY, PAGE_SIZE);
 722                 spin_lock_bh(&mm->context.list_lock);
 723                 list_add(&page->lru, &mm->context.pgtable_list);
 724         } else {
 725                 for (bit = 1; mask & bit; bit <<= 1)
 726                         table += PTRS_PER_PTE;
 727                 mask = atomic_xor_bits(&page->_mapcount, bit);
 728                 if ((mask & FRAG_MASK) == FRAG_MASK)
 729                         list_del(&page->lru);
 730         }
 731         spin_unlock_bh(&mm->context.list_lock);
 732         return table;
 733 }
 734
 735 void page_table_free(struct mm_struct *mm, unsigned long *table)
 736 {
 737         struct page *page;
 738         unsigned int bit, mask;
 739
 740         if (mm_has_pgste(mm)) {
 741                 gmap_disconnect_pgtable(mm, table);
 742                 return page_table_free_pgste(table);
 743         }
 744         /* Free 1K/2K page table fragment of a 4K page */
 745         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 746         bit = 1 << ((__pa(table) & ~PAGE_MASK)/(PTRS_PER_PTE*sizeof(pte_t)));
 747         spin_lock_bh(&mm->context.list_lock);
 748         if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
 749                 list_del(&page->lru);
 750         mask = atomic_xor_bits(&page->_mapcount, bit);
 751         if (mask & FRAG_MASK)
 752                 list_add(&page->lru, &mm->context.pgtable_list);
 753         spin_unlock_bh(&mm->context.list_lock);
 754         if (mask == 0) {
 755                 pgtable_page_dtor(page);
 756                 atomic_set(&page->_mapcount, -1);
 757                 __free_page(page);
 758         }
 759 }
 760
 761 static void __page_table_free_rcu(void *table, unsigned bit)
 762 {
 763         struct page *page;
 764
 765         if (bit == FRAG_MASK)
 766                 return page_table_free_pgste(table);
 767         /* Free 1K/2K page table fragment of a 4K page */
 768         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 769         if (atomic_xor_bits(&page->_mapcount, bit) == 0) {
 770                 pgtable_page_dtor(page);
 771                 atomic_set(&page->_mapcount, -1);
 772                 __free_page(page);
 773         }
 774 }
 775
 776 void page_table_free_rcu(struct mmu_gather *tlb, unsigned long *table)
 777 {
 778         struct mm_struct *mm;
 779         struct page *page;
 780         unsigned int bit, mask;
 781
 782         mm = tlb->mm;
 783         if (mm_has_pgste(mm)) {
 784                 gmap_disconnect_pgtable(mm, table);
 785                 table = (unsigned long *) (__pa(table) | FRAG_MASK);
 786                 tlb_remove_table(tlb, table);
 787                 return;
 788         }
 789         bit = 1 << ((__pa(table) & ~PAGE_MASK) / (PTRS_PER_PTE*sizeof(pte_t)));
 790         page = pfn_to_page(__pa(table) >> PAGE_SHIFT);
 791         spin_lock_bh(&mm->context.list_lock);
 792         if ((atomic_read(&page->_mapcount) & FRAG_MASK) != FRAG_MASK)
 793                 list_del(&page->lru);
 794         mask = atomic_xor_bits(&page->_mapcount, bit | (bit << 4));
 795         if (mask & FRAG_MASK)
 796                 list_add_tail(&page->lru, &mm->context.pgtable_list);
 797         spin_unlock_bh(&mm->context.list_lock);
 798         table = (unsigned long *) (__pa(table) | (bit << 4));
 799         tlb_remove_table(tlb, table);
 800 }
 801
 802 void __tlb_remove_table(void *_table)
 803 {
 804         const unsigned long mask = (FRAG_MASK << 4) | FRAG_MASK;
 805         void *table = (void *)((unsigned long) _table & ~mask);
 806         unsigned type = (unsigned long) _table & mask;
 807
 808         if (type)
 809                 __page_table_free_rcu(table, type);
 810         else
 811                 free_pages((unsigned long) table, ALLOC_ORDER);
 812 }
 813
 814 static void tlb_remove_table_smp_sync(void *arg)
 815 {
 816         /* Simply deliver the interrupt */
 817 }
 818
 819 static void tlb_remove_table_one(void *table)
 820 {
 821         /*
 822          * This isn't an RCU grace period and hence the page-tables cannot be
 823          * assumed to be actually RCU-freed.
 824          *
 825          * It is however sufficient for software page-table walkers that rely
 826          * on IRQ disabling. See the comment near struct mmu_table_batch.
 827          */
 828         smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
 829         __tlb_remove_table(table);
 830 }
 831
 832 static void tlb_remove_table_rcu(struct rcu_head *head)
 833 {
 834         struct mmu_table_batch *batch;
 835         int i;
 836
 837         batch = container_of(head, struct mmu_table_batch, rcu);
 838
 839         for (i = 0; i < batch->nr; i++)
 840                 __tlb_remove_table(batch->tables[i]);
 841
 842         free_page((unsigned long)batch);
 843 }
 844
 845 void tlb_table_flush(struct mmu_gather *tlb)
 846 {
 847         struct mmu_table_batch **batch = &tlb->batch;
 848
 849         if (*batch) {
 850                 __tlb_flush_mm(tlb->mm);
 851                 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
 852                 *batch = NULL;
 853         }
 854 }
 855
 856 void tlb_remove_table(struct mmu_gather *tlb, void *table)
 857 {
 858         struct mmu_table_batch **batch = &tlb->batch;
 859
 860         if (*batch == NULL) {
 861                 *batch = (struct mmu_table_batch *)
 862                         __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
 863                 if (*batch == NULL) {
 864                         __tlb_flush_mm(tlb->mm);
 865                         tlb_remove_table_one(table);
 866                         return;
 867                 }
 868                 (*batch)->nr = 0;
 869         }
 870         (*batch)->tables[(*batch)->nr++] = table;
 871         if ((*batch)->nr == MAX_TABLE_BATCH)
 872                 tlb_table_flush(tlb);
 873 }
 874
 875 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 876 void thp_split_vma(struct vm_area_struct *vma)
 877 {
 878         unsigned long addr;
 879         struct page *page;
 880
 881         for (addr = vma->vm_start; addr < vma->vm_end; addr += PAGE_SIZE) {
 882                 page = follow_page(vma, addr, FOLL_SPLIT);
 883         }
 884 }
 885
 886 void thp_split_mm(struct mm_struct *mm)
 887 {
 888         struct vm_area_struct *vma = mm->mmap;
 889
 890         while (vma != NULL) {
 891                 thp_split_vma(vma);
 892                 vma->vm_flags &= ~VM_HUGEPAGE;
 893                 vma->vm_flags |= VM_NOHUGEPAGE;
 894                 vma = vma->vm_next;
 895         }
 896 }
 897 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 898
 899 /*
 900  * switch on pgstes for its userspace process (for kvm)
 901  */
 902 int s390_enable_sie(void)
 903 {
 904         struct task_struct *tsk = current;
 905         struct mm_struct *mm, *old_mm;
 906
 907         /* Do we have switched amode? If no, we cannot do sie */
 908         if (s390_user_mode == HOME_SPACE_MODE)
 909                 return -EINVAL;
 910
 911         /* Do we have pgstes? if yes, we are done */
 912         if (mm_has_pgste(tsk->mm))
 913                 return 0;
 914
 915         /* lets check if we are allowed to replace the mm */
 916         task_lock(tsk);
 917         if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
 918 #ifdef CONFIG_AIO
 919             !hlist_empty(&tsk->mm->ioctx_list) ||
 920 #endif
 921             tsk->mm != tsk->active_mm) {
 922                 task_unlock(tsk);
 923                 return -EINVAL;
 924         }
 925         task_unlock(tsk);
 926
 927         /* we copy the mm and let dup_mm create the page tables with_pgstes */
 928         tsk->mm->context.alloc_pgste = 1;
 929         /* make sure that both mms have a correct rss state */
 930         sync_mm_rss(tsk->mm);
 931         mm = dup_mm(tsk);
 932         tsk->mm->context.alloc_pgste = 0;
 933         if (!mm)
 934                 return -ENOMEM;
 935
 936 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 937         /* split thp mappings and disable thp for future mappings */
 938         thp_split_mm(mm);
 939         mm->def_flags |= VM_NOHUGEPAGE;
 940 #endif
 941
 942         /* Now lets check again if something happened */
 943         task_lock(tsk);
 944         if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
 945 #ifdef CONFIG_AIO
 946             !hlist_empty(&tsk->mm->ioctx_list) ||
 947 #endif
 948             tsk->mm != tsk->active_mm) {
 949                 mmput(mm);
 950                 task_unlock(tsk);
 951                 return -EINVAL;
 952         }
 953
 954         /* ok, we are alone. No ptrace, no threads, etc. */
 955         old_mm = tsk->mm;
 956         tsk->mm = tsk->active_mm = mm;
 957         preempt_disable();
 958         update_mm(mm, tsk);
 959         atomic_inc(&mm->context.attach_count);
 960         atomic_dec(&old_mm->context.attach_count);
 961         cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
 962         preempt_enable();
 963         task_unlock(tsk);
 964         mmput(old_mm);
 965         return 0;
 966 }
 967 EXPORT_SYMBOL_GPL(s390_enable_sie);
 968
 969 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 970 int pmdp_clear_flush_young(struct vm_area_struct *vma, unsigned long address,
 971                            pmd_t *pmdp)
 972 {
 973         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 974         /* No need to flush TLB
 975          * On s390 reference bits are in storage key and never in TLB */
 976         return pmdp_test_and_clear_young(vma, address, pmdp);
 977 }
 978
 979 int pmdp_set_access_flags(struct vm_area_struct *vma,
 980                           unsigned long address, pmd_t *pmdp,
 981                           pmd_t entry, int dirty)
 982 {
 983         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 984
 985         if (pmd_same(*pmdp, entry))
 986                 return 0;
 987         pmdp_invalidate(vma, address, pmdp);
 988         set_pmd_at(vma->vm_mm, address, pmdp, entry);
 989         return 1;
 990 }
 991
 992 static void pmdp_splitting_flush_sync(void *arg)
 993 {
 994         /* Simply deliver the interrupt */
 995 }
 996
 997 void pmdp_splitting_flush(struct vm_area_struct *vma, unsigned long address,
 998                           pmd_t *pmdp)
 999 {
1000         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1001         if (!test_and_set_bit(_SEGMENT_ENTRY_SPLIT_BIT,
1002                               (unsigned long *) pmdp)) {
1003                 /* need to serialize against gup-fast (IRQ disabled) */
1004                 smp_call_function(pmdp_splitting_flush_sync, NULL, 1);
1005         }
1006 }
1007
1008 void pgtable_trans_huge_deposit(struct mm_struct *mm, pgtable_t pgtable)
1009 {
1010         struct list_head *lh = (struct list_head *) pgtable;
1011
1012         assert_spin_locked(&mm->page_table_lock);
1013
1014         /* FIFO */
1015         if (!mm->pmd_huge_pte)
1016                 INIT_LIST_HEAD(lh);
1017         else
1018                 list_add(lh, (struct list_head *) mm->pmd_huge_pte);
1019         mm->pmd_huge_pte = pgtable;
1020 }
1021
1022 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm)
1023 {
1024         struct list_head *lh;
1025         pgtable_t pgtable;
1026         pte_t *ptep;
1027
1028         assert_spin_locked(&mm->page_table_lock);
1029
1030         /* FIFO */
1031         pgtable = mm->pmd_huge_pte;
1032         lh = (struct list_head *) pgtable;
1033         if (list_empty(lh))
1034                 mm->pmd_huge_pte = NULL;
1035         else {
1036                 mm->pmd_huge_pte = (pgtable_t) lh->next;
1037                 list_del(lh);
1038         }
1039         ptep = (pte_t *) pgtable;
1040         pte_val(*ptep) = _PAGE_TYPE_EMPTY;
1041         ptep++;
1042         pte_val(*ptep) = _PAGE_TYPE_EMPTY;
1043         return pgtable;
1044 }
1045 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */