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