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