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 spinlock_t kmap_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
  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) {
 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                 bounce_copy_vec(tovec, vfrom);
 304         }
 305 }
 306
 307 static void bounce_end_io(struct bio *bio, mempool_t *pool)
 308 {
 309         struct bio *bio_orig = bio->bi_private;
 310         struct bio_vec *bvec, *org_vec;
 311         int i, err = 0;
 312
 313         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 314                 err = -EIO;
 315
 316         /*
 317          * free up bounce indirect pages used
 318          */
 319         bio_for_each_segment(bvec, bio, i) {
 320                 org_vec = bio_orig->bi_io_vec + i;
 321                 if (bvec->bv_page == org_vec->bv_page)
 322                         continue;
 323
 324                 mempool_free(bvec->bv_page, pool);
 325         }
 326
 327         bio_endio(bio_orig, bio_orig->bi_size, err);
 328         bio_put(bio);
 329 }
 330
 331 static int bounce_end_io_write(struct bio *bio, unsigned int bytes_done,int err)
 332 {
 333         if (bio->bi_size)
 334                 return 1;
 335
 336         bounce_end_io(bio, page_pool);
 337         return 0;
 338 }
 339
 340 static int bounce_end_io_write_isa(struct bio *bio, unsigned int bytes_done, int err)
 341 {
 342         if (bio->bi_size)
 343                 return 1;
 344
 345         bounce_end_io(bio, isa_page_pool);
 346         return 0;
 347 }
 348
 349 static void __bounce_end_io_read(struct bio *bio, mempool_t *pool)
 350 {
 351         struct bio *bio_orig = bio->bi_private;
 352
 353         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 354                 copy_to_high_bio_irq(bio_orig, bio);
 355
 356         bounce_end_io(bio, pool);
 357 }
 358
 359 static int bounce_end_io_read(struct bio *bio, unsigned int bytes_done, int err)
 360 {
 361         if (bio->bi_size)
 362                 return 1;
 363
 364         __bounce_end_io_read(bio, page_pool);
 365         return 0;
 366 }
 367
 368 static int bounce_end_io_read_isa(struct bio *bio, unsigned int bytes_done, int err)
 369 {
 370         if (bio->bi_size)
 371                 return 1;
 372
 373         __bounce_end_io_read(bio, isa_page_pool);
 374         return 0;
 375 }
 376
 377 static void __blk_queue_bounce(request_queue_t *q, struct bio **bio_orig,
 378                         mempool_t *pool)
 379 {
 380         struct page *page;
 381         struct bio *bio = NULL;
 382         int i, rw = bio_data_dir(*bio_orig);
 383         struct bio_vec *to, *from;
 384
 385         bio_for_each_segment(from, *bio_orig, i) {
 386                 page = from->bv_page;
 387
 388                 /*
 389                  * is destination page below bounce pfn?
 390                  */
 391                 if (page_to_pfn(page) < q->bounce_pfn)
 392                         continue;
 393
 394                 /*
 395                  * irk, bounce it
 396                  */
 397                 if (!bio)
 398                         bio = bio_alloc(GFP_NOIO, (*bio_orig)->bi_vcnt);
 399
 400                 to = bio->bi_io_vec + i;
 401
 402                 to->bv_page = mempool_alloc(pool, q->bounce_gfp);
 403                 to->bv_len = from->bv_len;
 404                 to->bv_offset = from->bv_offset;
 405
 406                 if (rw == WRITE) {
 407                         char *vto, *vfrom;
 408
 409                         vto = page_address(to->bv_page) + to->bv_offset;
 410                         vfrom = kmap(from->bv_page) + from->bv_offset;
 411                         memcpy(vto, vfrom, to->bv_len);
 412                         kunmap(from->bv_page);
 413                 }
 414         }
 415
 416         /*
 417          * no pages bounced
 418          */
 419         if (!bio)
 420                 return;
 421
 422         /*
 423          * at least one page was bounced, fill in possible non-highmem
 424          * pages
 425          */
 426         bio_for_each_segment(from, *bio_orig, i) {
 427                 to = bio_iovec_idx(bio, i);
 428                 if (!to->bv_page) {
 429                         to->bv_page = from->bv_page;
 430                         to->bv_len = from->bv_len;
 431                         to->bv_offset = from->bv_offset;
 432                 }
 433         }
 434
 435         bio->bi_bdev = (*bio_orig)->bi_bdev;
 436         bio->bi_flags |= (1 << BIO_BOUNCED);
 437         bio->bi_sector = (*bio_orig)->bi_sector;
 438         bio->bi_rw = (*bio_orig)->bi_rw;
 439
 440         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
 441         bio->bi_idx = (*bio_orig)->bi_idx;
 442         bio->bi_size = (*bio_orig)->bi_size;
 443
 444         if (pool == page_pool) {
 445                 bio->bi_end_io = bounce_end_io_write;
 446                 if (rw == READ)
 447                         bio->bi_end_io = bounce_end_io_read;
 448         } else {
 449                 bio->bi_end_io = bounce_end_io_write_isa;
 450                 if (rw == READ)
 451                         bio->bi_end_io = bounce_end_io_read_isa;
 452         }
 453
 454         bio->bi_private = *bio_orig;
 455         *bio_orig = bio;
 456 }
 457
 458 void blk_queue_bounce(request_queue_t *q, struct bio **bio_orig)
 459 {
 460         mempool_t *pool;
 461
 462         /*
 463          * for non-isa bounce case, just check if the bounce pfn is equal
 464          * to or bigger than the highest pfn in the system -- in that case,
 465          * don't waste time iterating over bio segments
 466          */
 467         if (!(q->bounce_gfp & GFP_DMA)) {
 468                 if (q->bounce_pfn >= blk_max_pfn)
 469                         return;
 470                 pool = page_pool;
 471         } else {
 472                 BUG_ON(!isa_page_pool);
 473                 pool = isa_page_pool;
 474         }
 475
 476         /*
 477          * slow path
 478          */
 479         __blk_queue_bounce(q, bio_orig, pool);
 480 }
 481
 482 EXPORT_SYMBOL(blk_queue_bounce);
 483
 484 #if defined(HASHED_PAGE_VIRTUAL)
 485
 486 #define PA_HASH_ORDER   7
 487
 488 /*
 489  * Describes one page->virtual association
 490  */
 491 struct page_address_map {
 492         struct page *page;
 493         void *virtual;
 494         struct list_head list;
 495 };
 496
 497 /*
 498  * page_address_map freelist, allocated from page_address_maps.
 499  */
 500 static struct list_head page_address_pool;      /* freelist */
 501 static spinlock_t pool_lock;                    /* protects page_address_pool */
 502
 503 /*
 504  * Hash table bucket
 505  */
 506 static struct page_address_slot {
 507         struct list_head lh;                    /* List of page_address_maps */
 508         spinlock_t lock;                        /* Protect this bucket's list */
 509 } ____cacheline_aligned_in_smp page_address_htable[1<<PA_HASH_ORDER];
 510
 511 static struct page_address_slot *page_slot(struct page *page)
 512 {
 513         return &page_address_htable[hash_ptr(page, PA_HASH_ORDER)];
 514 }
 515
 516 void *page_address(struct page *page)
 517 {
 518         unsigned long flags;
 519         void *ret;
 520         struct page_address_slot *pas;
 521
 522         if (!PageHighMem(page))
 523                 return lowmem_page_address(page);
 524
 525         pas = page_slot(page);
 526         ret = NULL;
 527         spin_lock_irqsave(&pas->lock, flags);
 528         if (!list_empty(&pas->lh)) {
 529                 struct page_address_map *pam;
 530
 531                 list_for_each_entry(pam, &pas->lh, list) {
 532                         if (pam->page == page) {
 533                                 ret = pam->virtual;
 534                                 goto done;
 535                         }
 536                 }
 537         }
 538 done:
 539         spin_unlock_irqrestore(&pas->lock, flags);
 540         return ret;
 541 }
 542
 543 EXPORT_SYMBOL(page_address);
 544
 545 void set_page_address(struct page *page, void *virtual)
 546 {
 547         unsigned long flags;
 548         struct page_address_slot *pas;
 549         struct page_address_map *pam;
 550
 551         BUG_ON(!PageHighMem(page));
 552
 553         pas = page_slot(page);
 554         if (virtual) {          /* Add */
 555                 BUG_ON(list_empty(&page_address_pool));
 556
 557                 spin_lock_irqsave(&pool_lock, flags);
 558                 pam = list_entry(page_address_pool.next,
 559                                 struct page_address_map, list);
 560                 list_del(&pam->list);
 561                 spin_unlock_irqrestore(&pool_lock, flags);
 562
 563                 pam->page = page;
 564                 pam->virtual = virtual;
 565
 566                 spin_lock_irqsave(&pas->lock, flags);
 567                 list_add_tail(&pam->list, &pas->lh);
 568                 spin_unlock_irqrestore(&pas->lock, flags);
 569         } else {                /* Remove */
 570                 spin_lock_irqsave(&pas->lock, flags);
 571                 list_for_each_entry(pam, &pas->lh, list) {
 572                         if (pam->page == page) {
 573                                 list_del(&pam->list);
 574                                 spin_unlock_irqrestore(&pas->lock, flags);
 575                                 spin_lock_irqsave(&pool_lock, flags);
 576                                 list_add_tail(&pam->list, &page_address_pool);
 577                                 spin_unlock_irqrestore(&pool_lock, flags);
 578                                 goto done;
 579                         }
 580                 }
 581                 spin_unlock_irqrestore(&pas->lock, flags);
 582         }
 583 done:
 584         return;
 585 }
 586
 587 static struct page_address_map page_address_maps[LAST_PKMAP];
 588
 589 void __init page_address_init(void)
 590 {
 591         int i;
 592
 593         INIT_LIST_HEAD(&page_address_pool);
 594         for (i = 0; i < ARRAY_SIZE(page_address_maps); i++)
 595                 list_add(&page_address_maps[i].list, &page_address_pool);
 596         for (i = 0; i < ARRAY_SIZE(page_address_htable); i++) {
 597                 INIT_LIST_HEAD(&page_address_htable[i].lh);
 598                 spin_lock_init(&page_address_htable[i].lock);
 599         }
 600         spin_lock_init(&pool_lock);
 601 }
 602
 603 #endif  /* defined(CONFIG_HIGHMEM) && !defined(WANT_PAGE_VIRTUAL) */