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