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