mm/highmem.c

   1 /*
   2  * High memory handling common code and variables.
   3  *
   4  * (C) 1999 Andrea Arcangeli, SuSE GmbH, andrea@suse.de
   5  *          Gerhard Wichert, Siemens AG, Gerhard.Wichert@pdb.siemens.de
   6  *
   7  *
   8  * Redesigned the x86 32-bit VM architecture to deal with
   9  * 64-bit physical space. With current x86 CPUs this
  10  * means up to 64 Gigabytes physical RAM.
  11  *
  12  * Rewrote high memory support to move the page cache into
  13  * high memory. Implemented permanent (schedulable) kmaps
  14  * based on Linus' idea.
  15  *
  16  * Copyright (C) 1999 Ingo Molnar <mingo@redhat.com>
  17  */
  18
  19 #include <linux/mm.h>
  20 #include <linux/module.h>
  21 #include <linux/swap.h>
  22 #include <linux/bio.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/mempool.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/init.h>
  27 #include <linux/hash.h>
  28 #include <linux/highmem.h>
  29 #include <asm/tlbflush.h>
  30
  31 static mempool_t *page_pool, *isa_page_pool;
  32
  33 static void *page_pool_alloc_isa(gfp_t gfp_mask, void *data)
  34 {
  35         return alloc_page(gfp_mask | GFP_DMA);
  36 }
  37
  38 static void page_pool_free(void *page, void *data)
  39 {
  40         __free_page(page);
  41 }
  42
  43 /*
  44  * Virtual_count is not a pure "count".
  45  *  0 means that it is not mapped, and has not been mapped
  46  *    since a TLB flush - it is usable.
  47  *  1 means that there are no users, but it has been mapped
  48  *    since the last TLB flush - so we can't use it.
  49  *  n means that there are (n-1) current users of it.
  50  */
  51 #ifdef CONFIG_HIGHMEM
  52
  53 static void *page_pool_alloc(gfp_t gfp_mask, void *data)
  54 {
  55         return alloc_page(gfp_mask);
  56 }
  57
  58 static int pkmap_count[LAST_PKMAP];
  59 static unsigned int last_pkmap_nr;
  60 static  __cacheline_aligned_in_smp DEFINE_SPINLOCK(kmap_lock);
  61
  62 pte_t * pkmap_page_table;
  63
  64 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  65
  66 static void flush_all_zero_pkmaps(void)
  67 {
  68         int i;
  69
  70         flush_cache_kmaps();
  71
  72         for (i = 0; i < LAST_PKMAP; i++) {
  73                 struct page *page;
  74
  75                 /*
  76                  * zero means we don't have anything to do,
  77                  * >1 means that it is still in use. Only
  78                  * a count of 1 means that it is free but
  79                  * needs to be unmapped
  80                  */
  81                 if (pkmap_count[i] != 1)
  82                         continue;
  83                 pkmap_count[i] = 0;
  84
  85                 /* sanity check */
  86                 if (pte_none(pkmap_page_table[i]))
  87                         BUG();
  88
  89                 /*
  90                  * Don't need an atomic fetch-and-clear op here;
  91                  * no-one has the page mapped, and cannot get at
  92                  * its virtual address (and hence PTE) without first
  93                  * getting the kmap_lock (which is held here).
  94                  * So no dangers, even with speculative execution.
  95                  */
  96                 page = pte_page(pkmap_page_table[i]);
  97                 pte_clear(&init_mm, (unsigned long)page_address(page),
  98                           &pkmap_page_table[i]);
  99
 100                 set_page_address(page, NULL);
 101         }
 102         flush_tlb_kernel_range(PKMAP_ADDR(0), PKMAP_ADDR(LAST_PKMAP));
 103 }
 104
 105 static inline unsigned long map_new_virtual(struct page *page)
 106 {
 107         unsigned long vaddr;
 108         int count;
 109
 110 start:
 111         count = LAST_PKMAP;
 112         /* Find an empty entry */
 113         for (;;) {
 114                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
 115                 if (!last_pkmap_nr) {
 116                         flush_all_zero_pkmaps();
 117                         count = LAST_PKMAP;
 118                 }
 119                 if (!pkmap_count[last_pkmap_nr])
 120                         break;  /* Found a usable entry */
 121                 if (--count)
 122                         continue;
 123
 124                 /*
 125                  * Sleep for somebody else to unmap their entries
 126                  */
 127                 {
 128                         DECLARE_WAITQUEUE(wait, current);
 129
 130                         __set_current_state(TASK_UNINTERRUPTIBLE);
 131                         add_wait_queue(&pkmap_map_wait, &wait);
 132                         spin_unlock(&kmap_lock);
 133                         schedule();
 134                         remove_wait_queue(&pkmap_map_wait, &wait);
 135                         spin_lock(&kmap_lock);
 136
 137                         /* Somebody else might have mapped it while we slept */
 138                         if (page_address(page))
 139                                 return (unsigned long)page_address(page);
 140
 141                         /* Re-start */
 142                         goto start;
 143                 }
 144         }
 145         vaddr = PKMAP_ADDR(last_pkmap_nr);
 146         set_pte_at(&init_mm, vaddr,
 147                    &(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 148
 149         pkmap_count[last_pkmap_nr] = 1;
 150         set_page_address(page, (void *)vaddr);
 151
 152         return vaddr;
 153 }
 154
 155 #ifdef CONFIG_XEN
 156 void kmap_flush_unused(void)
 157 {
 158         spin_lock(&kmap_lock);
 159         flush_all_zero_pkmaps();
 160         spin_unlock(&kmap_lock);
 161 }
 162
 163 EXPORT_SYMBOL(kmap_flush_unused);
 164 #endif
 165
 166 void fastcall *kmap_high(struct page *page)
 167 {
 168         unsigned long vaddr;
 169
 170         /*
 171          * For highmem pages, we can't trust "virtual" until
 172          * after we have the lock.
 173          *
 174          * We cannot call this from interrupts, as it may block
 175          */
 176         spin_lock(&kmap_lock);
 177         vaddr = (unsigned long)page_address(page);
 178         if (!vaddr)
 179                 vaddr = map_new_virtual(page);
 180         pkmap_count[PKMAP_NR(vaddr)]++;
 181         if (pkmap_count[PKMAP_NR(vaddr)] < 2)
 182                 BUG();
 183         spin_unlock(&kmap_lock);
 184         return (void*) vaddr;
 185 }
 186
 187 EXPORT_SYMBOL(kmap_high);
 188
 189 void fastcall kunmap_high(struct page *page)
 190 {
 191         unsigned long vaddr;
 192         unsigned long nr;
 193         int need_wakeup;
 194
 195         spin_lock(&kmap_lock);
 196         vaddr = (unsigned long)page_address(page);
 197         if (!vaddr)
 198                 BUG();
 199         nr = PKMAP_NR(vaddr);
 200
 201         /*
 202          * A count must never go down to zero
 203          * without a TLB flush!
 204          */
 205         need_wakeup = 0;
 206         switch (--pkmap_count[nr]) {
 207         case 0:
 208                 BUG();
 209         case 1:
 210                 /*
 211                  * Avoid an unnecessary wake_up() function call.
 212                  * The common case is pkmap_count[] == 1, but
 213                  * no waiters.
 214                  * The tasks queued in the wait-queue are guarded
 215                  * by both the lock in the wait-queue-head and by
 216                  * the kmap_lock.  As the kmap_lock is held here,
 217                  * no need for the wait-queue-head's lock.  Simply
 218                  * test if the queue is empty.
 219                  */
 220                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 221         }
 222         spin_unlock(&kmap_lock);
 223
 224         /* do wake-up, if needed, race-free outside of the spin lock */
 225         if (need_wakeup)
 226                 wake_up(&pkmap_map_wait);
 227 }
 228
 229 EXPORT_SYMBOL(kunmap_high);
 230
 231 #define POOL_SIZE       64
 232
 233 static __init int init_emergency_pool(void)
 234 {
 235         struct sysinfo i;
 236         si_meminfo(&i);
 237         si_swapinfo(&i);
 238
 239         if (!i.totalhigh)
 240                 return 0;
 241
 242         page_pool = mempool_create(POOL_SIZE, page_pool_alloc, page_pool_free, NULL);
 243         if (!page_pool)
 244                 BUG();
 245         printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
 246
 247         return 0;
 248 }
 249
 250 __initcall(init_emergency_pool);
 251
 252 /*
 253  * highmem version, map in to vec
 254  */
 255 static void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
 256 {
 257         unsigned long flags;
 258         unsigned char *vto;
 259
 260         local_irq_save(flags);
 261         vto = kmap_atomic(to->bv_page, KM_BOUNCE_READ);
 262         memcpy(vto + to->bv_offset, vfrom, to->bv_len);
 263         kunmap_atomic(vto, KM_BOUNCE_READ);
 264         local_irq_restore(flags);
 265 }
 266
 267 #else /* CONFIG_HIGHMEM */
 268
 269 #define bounce_copy_vec(to, vfrom)      \
 270         memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
 271
 272 #endif
 273
 274 #define ISA_POOL_SIZE   16
 275
 276 /*
 277  * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
 278  * as the max address, so check if the pool has already been created.
 279  */
 280 int init_emergency_isa_pool(void)
 281 {
 282         if (isa_page_pool)
 283                 return 0;
 284
 285         isa_page_pool = mempool_create(ISA_POOL_SIZE, page_pool_alloc_isa, page_pool_free, NULL);
 286         if (!isa_page_pool)
 287                 BUG();
 288
 289         printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
 290         return 0;
 291 }
 292
 293 /*
 294  * Simple bounce buffer support for highmem pages. Depending on the
 295  * queue gfp mask set, *to may or may not be a highmem page. kmap it
 296  * always, it will do the Right Thing
 297  */
 298 static void copy_to_high_bio_irq(struct bio *to, struct bio *from)
 299 {
 300         unsigned char *vfrom;
 301         struct bio_vec *tovec, *fromvec;
 302         int i;
 303
 304         __bio_for_each_segment(tovec, to, i, 0) {
 305                 fromvec = from->bi_io_vec + i;
 306
 307                 /*
 308                  * not bounced
 309                  */
 310                 if (tovec->bv_page == fromvec->bv_page)
 311                         continue;
 312
 313                 /*
 314                  * fromvec->bv_offset and fromvec->bv_len might have been
 315                  * modified by the block layer, so use the original copy,
 316                  * bounce_copy_vec already uses tovec->bv_len
 317                  */
 318                 vfrom = page_address(fromvec->bv_page) + tovec->bv_offset;
 319
 320                 flush_dcache_page(tovec->bv_page);
 321                 bounce_copy_vec(tovec, vfrom);
 322         }
 323 }
 324
 325 static void bounce_end_io(struct bio *bio, mempool_t *pool, int err)
 326 {
 327         struct bio *bio_orig = bio->bi_private;
 328         struct bio_vec *bvec, *org_vec;
 329         int i;
 330
 331         if (test_bit(BIO_EOPNOTSUPP, &bio->bi_flags))
 332                 set_bit(BIO_EOPNOTSUPP, &bio_orig->bi_flags);
 333
 334         /*
 335          * free up bounce indirect pages used
 336          */
 337         __bio_for_each_segment(bvec, bio, i, 0) {
 338                 org_vec = bio_orig->bi_io_vec + i;
 339                 if (bvec->bv_page == org_vec->bv_page)
 340                         continue;
 341
 342                 mempool_free(bvec->bv_page, pool);
 343                 dec_page_state(nr_bounce);
 344         }
 345
 346         bio_endio(bio_orig, bio_orig->bi_size, err);
 347         bio_put(bio);
 348 }
 349
 350 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done,int err)
 351 {
 352         if (bio->bi_size)
 353                 return 1;
 354
 355         bounce_end_io(bio, page_pool, err);
 356         return 0;
 357 }
 358
 359 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
 360 {
 361         if (bio->bi_size)
 362                 return 1;
 363
 364         bounce_end_io(bio, isa_page_pool, err);
 365         return 0;
 366 }
 367
 368 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool, int err)
 369 {
 370         struct bio *bio_orig = bio->bi_private;
 371
 372         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 373                 copy_to_high_bio_irq(bio_orig, bio);
 374
 375         bounce_end_io(bio, pool, err);
 376 }
 377
 378 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
 379 {
 380         if (bio->bi_size)
 381                 return 1;
 382
 383         __bounce_end_io_read(bio, page_pool, err);
 384         return 0;
 385 }
 386
 387 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
 388 {
 389         if (bio->bi_size)
 390                 return 1;
 391
 392         __bounce_end_io_read(bio, isa_page_pool, err);
 393         return 0;
 394 }
 395
 396 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
 397                         mempool_t *pool)
 398 {
 399         struct page *page;
 400         struct bio *bio = NULL;
 401         int i, rw = bio_data_dir(*bio_orig);
 402         struct bio_vec *to, *from;
 403
 404         bio_for_each_segment(from, *bio_orig, i) {
 405                 page = from->bv_page;
 406
 407                 /*
 408                  * is destination page below bounce pfn?
 409                  */
 410                 if (page_to_pfn(page) < q->bounce_pfn)
 411                         continue;
 412
 413                 /*
 414                  * irk, bounce it
 415                  */
 416                 if (!bio)
 417                         bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
 418
 419                 to = bio->bi_io_vec + i;
 420
 421                 to->bv_page = mempool_alloc(pool, q->bounce_gfp);
 422                 to->bv_len = from->bv_len;
 423                 to->bv_offset = from->bv_offset;
 424                 inc_page_state(nr_bounce);
 425
 426                 if (rw == WRITE) {
 427                         char *vto, *vfrom;
 428
 429                         flush_dcache_page(from->bv_page);
 430                         vto = page_address(to->bv_page) + to->bv_offset;
 431                         vfrom = kmap(from->bv_page) + from->bv_offset;
 432                         memcpy(vto, vfrom, to->bv_len);
 433                         kunmap(from->bv_page);
 434                 }
 435         }
 436
 437         /*
 438          * no pages bounced
 439          */
 440         if (!bio)
 441                 return;
 442
 443         /*
 444          * at least one page was bounced, fill in possible non-highmem
 445          * pages
 446          */
 447         __bio_for_each_segment(from, *bio_orig, i, 0) {
 448                 to = bio_iovec_idx(bio, i);
 449                 if (!to->bv_page) {
 450                         to->bv_page = from->bv_page;
 451                         to->bv_len = from->bv_len;
 452                         to->bv_offset = from->bv_offset;
 453                 }
 454         }
 455
 456         bio->bi_bdev = (*bio_orig)->bi_bdev;
 457         bio->bi_flags |= (1 << BIO_BOUNCED);
 458         bio->bi_sector = (*bio_orig)->bi_sector;
 459         bio->bi_rw = (*bio_orig)->bi_rw;
 460
 461         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
 462         bio->bi_idx = (*bio_orig)->bi_idx;
 463         bio->bi_size = (*bio_orig)->bi_size;
 464
 465         if (pool == page_pool) {
 466                 bio->bi_end_io = bounce_end_io_write;
 467                 if (rw == READ)
 468                         bio->bi_end_io = bounce_end_io_read;
 469         } else {
 470                 bio->bi_end_io = bounce_end_io_write_isa;
 471                 if (rw == READ)
 472                         bio->bi_end_io = bounce_end_io_read_isa;
 473         }
 474
 475         bio->bi_private = *bio_orig;
 476         *bio_orig = bio;
 477 }
 478
 479 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
 480 {
 481         mempool_t *pool;
 482
 483         /*
 484          * for non-isa bounce case, just check if the bounce pfn is equal
 485          * to or bigger than the highest pfn in the system -- in that case,
 486          * don't waste time iterating over bio segments
 487          */
 488         if (!(q->bounce_gfp & GFP_DMA)) {
 489                 if (q->bounce_pfn >= blk_max_pfn)
 490                         return;
 491                 pool = page_pool;
 492         } else {
 493                 BUG_ON(!isa_page_pool);
 494                 pool = isa_page_pool;
 495         }
 496
 497         /*
 498          * slow path
 499          */
 500         __blk_queue_bounce(q, bio_orig, pool);
 501 }
 502
 503 EXPORT_SYMBOL(blk_queue_bounce);
 504
 505 #if defined(HASHED_PAGE_VIRTUAL)
 506
 507 #define PA_HASH_ORDER   7
 508
 509 /*
 510  * Describes one page->virtual association
 511  */
 512 struct page_address_map {
 513         struct page *page;
 514         void *virtual;
 515         struct list_head list;
 516 };
 517
 518 /*
 519  * page_address_map freelist, allocated from page_address_maps.
 520  */
 521 static struct list_head page_address_pool;      /* freelist */
 522 static spinlock_t pool_lock;                    /* protects page_address_pool */
 523
 524 /*
 525  * Hash table bucket
 526  */
 527 static struct page_address_slot {
 528         struct list_head lh;                    /* List of page_address_maps */
 529         spinlock_t lock;                        /* Protect this bucket's list */
 530 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 531
 532 static struct page_address_slot *page_slot(struct page *page)
 533 {
 534         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 535 }
 536
 537 void *page_address(struct page *page)
 538 {
 539         unsigned long flags;
 540         void *ret;
 541         struct page_address_slot *pas;
 542
 543         if (!PageHighMem(page))
 544                 return lowmem_page_address(page);
 545
 546         pas = page_slot(page);
 547         ret = NULL;
 548         spin_lock_irqsave(&pas->lock, flags);
 549         if (!list_empty(&pas->lh)) {
 550                 struct page_address_map *pam;
 551
 552                 list_for_each_entry(pam, &pas->lh, list) {
 553                         if (pam->page == page) {
 554                                 ret = pam->virtual;
 555                                 goto done;
 556                         }
 557                 }
 558         }
 559 done:
 560         spin_unlock_irqrestore(&pas->lock, flags);
 561         return ret;
 562 }
 563
 564 EXPORT_SYMBOL(page_address);
 565
 566 void set_page_address(struct page *page, void *virtual)
 567 {
 568         unsigned long flags;
 569         struct page_address_slot *pas;
 570         struct page_address_map *pam;
 571
 572         BUG_ON(!PageHighMem(page));
 573
 574         pas = page_slot(page);
 575         if (virtual) {          /* Add */
 576                 BUG_ON(list_empty(&page_address_pool));
 577
 578                 spin_lock_irqsave(&pool_lock, flags);
 579                 pam = list_entry(page_address_pool.next,
 580                                 struct page_address_map, list);
 581                 list_del(&pam->list);
 582                 spin_unlock_irqrestore(&pool_lock, flags);
 583
 584                 pam->page = page;
 585                 pam->virtual = virtual;
 586
 587                 spin_lock_irqsave(&pas->lock, flags);
 588                 list_add_tail(&pam->list, &pas->lh);
 589                 spin_unlock_irqrestore(&pas->lock, flags);
 590         } else {                /* Remove */
 591                 spin_lock_irqsave(&pas->lock, flags);
 592                 list_for_each_entry(pam, &pas->lh, list) {
 593                         if (pam->page == page) {
 594                                 list_del(&pam->list);
 595                                 spin_unlock_irqrestore(&pas->lock, flags);
 596                                 spin_lock_irqsave(&pool_lock, flags);
 597                                 list_add_tail(&pam->list, &page_address_pool);
 598                                 spin_unlock_irqrestore(&pool_lock, flags);
 599                                 goto done;
 600                         }
 601                 }
 602                 spin_unlock_irqrestore(&pas->lock, flags);
 603         }
 604 done:
 605         return;
 606 }
 607
 608 static struct page_address_map page_address_maps[LAST_PKMAP];
 609
 610 void __init page_address_init(void)
 611 {
 612         int i;
 613
 614         INIT_LIST_HEAD(&page_address_pool);
 615         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
 616                 list_add(&page_address_maps[i].list, &page_address_pool);
 617         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 618                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 619                 spin_lock_init(&page_address_htable[i].lock);
 620         }
 621         spin_lock_init(&pool_lock);
 622 }
 623
 624 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */