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/pagemap.h>
  21 #include <linux/mempool.h>
  22 #include <linux/blkdev.h>
  23
  24 static mempool_t *page_pool, *isa_page_pool;
  25
  26 static void *page_pool_alloc(int gfp_mask, void *data)
  27 {
  28         int gfp = gfp_mask | (int) (long) data;
  29
  30         return alloc_page(gfp);
  31 }
  32
  33 static void page_pool_free(void *page, void *data)
  34 {
  35         __free_page(page);
  36 }
  37
  38 /*
  39  * Virtual_count is not a pure "count".
  40  *  0 means that it is not mapped, and has not been mapped
  41  *    since a TLB flush - it is usable.
  42  *  1 means that there are no users, but it has been mapped
  43  *    since the last TLB flush - so we can't use it.
  44  *  n means that there are (n-1) current users of it.
  45  */
  46 #ifdef CONFIG_HIGHMEM
  47 static int pkmap_count[LAST_PKMAP];
  48 static unsigned int last_pkmap_nr;
  49 static spinlock_t kmap_lock __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
  50
  51 pte_t * pkmap_page_table;
  52
  53 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  54
  55 static void flush_all_zero_pkmaps(void)
  56 {
  57         int i;
  58
  59         flush_cache_all();
  60
  61         for (i = 0; i < LAST_PKMAP; i++) {
  62                 struct page *page;
  63
  64                 /*
  65                  * zero means we don't have anything to do,
  66                  * >1 means that it is still in use. Only
  67                  * a count of 1 means that it is free but
  68                  * needs to be unmapped
  69                  */
  70                 if (pkmap_count[i] != 1)
  71                         continue;
  72                 pkmap_count[i] = 0;
  73
  74                 /* sanity check */
  75                 if (pte_none(pkmap_page_table[i]))
  76                         BUG();
  77
  78                 /*
  79                  * Don't need an atomic fetch-and-clear op here;
  80                  * no-one has the page mapped, and cannot get at
  81                  * its virtual address (and hence PTE) without first
  82                  * getting the kmap_lock (which is held here).
  83                  * So no dangers, even with speculative execution.
  84                  */
  85                 page = pte_page(pkmap_page_table[i]);
  86                 pte_clear(&pkmap_page_table[i]);
  87
  88                 page->virtual = NULL;
  89         }
  90         flush_tlb_all();
  91 }
  92
  93 static inline unsigned long map_new_virtual(struct page *page)
  94 {
  95         unsigned long vaddr;
  96         int count;
  97
  98 start:
  99         count = LAST_PKMAP;
 100         /* Find an empty entry */
 101         for (;;) {
 102                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
 103                 if (!last_pkmap_nr) {
 104                         flush_all_zero_pkmaps();
 105                         count = LAST_PKMAP;
 106                 }
 107                 if (!pkmap_count[last_pkmap_nr])
 108                         break;  /* Found a usable entry */
 109                 if (--count)
 110                         continue;
 111
 112                 /*
 113                  * Sleep for somebody else to unmap their entries
 114                  */
 115                 {
 116                         DECLARE_WAITQUEUE(wait, current);
 117
 118                         current->state = TASK_UNINTERRUPTIBLE;
 119                         add_wait_queue(&pkmap_map_wait, &wait);
 120                         spin_unlock(&kmap_lock);
 121                         schedule();
 122                         remove_wait_queue(&pkmap_map_wait, &wait);
 123                         spin_lock(&kmap_lock);
 124
 125                         /* Somebody else might have mapped it while we slept */
 126                         if (page->virtual)
 127                                 return (unsigned long) page->virtual;
 128
 129                         /* Re-start */
 130                         goto start;
 131                 }
 132         }
 133         vaddr = PKMAP_ADDR(last_pkmap_nr);
 134         set_pte(&(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 135
 136         pkmap_count[last_pkmap_nr] = 1;
 137         page->virtual = (void *) vaddr;
 138
 139         return vaddr;
 140 }
 141
 142 void *kmap_high(struct page *page)
 143 {
 144         unsigned long vaddr;
 145
 146         /*
 147          * For highmem pages, we can't trust "virtual" until
 148          * after we have the lock.
 149          *
 150          * We cannot call this from interrupts, as it may block
 151          */
 152         spin_lock(&kmap_lock);
 153         vaddr = (unsigned long) page->virtual;
 154         if (!vaddr)
 155                 vaddr = map_new_virtual(page);
 156         pkmap_count[PKMAP_NR(vaddr)]++;
 157         if (pkmap_count[PKMAP_NR(vaddr)] < 2)
 158                 BUG();
 159         spin_unlock(&kmap_lock);
 160         return (void*) vaddr;
 161 }
 162
 163 void kunmap_high(struct page *page)
 164 {
 165         unsigned long vaddr;
 166         unsigned long nr;
 167         int need_wakeup;
 168
 169         spin_lock(&kmap_lock);
 170         vaddr = (unsigned long) page->virtual;
 171         if (!vaddr)
 172                 BUG();
 173         nr = PKMAP_NR(vaddr);
 174
 175         /*
 176          * A count must never go down to zero
 177          * without a TLB flush!
 178          */
 179         need_wakeup = 0;
 180         switch (--pkmap_count[nr]) {
 181         case 0:
 182                 BUG();
 183         case 1:
 184                 /*
 185                  * Avoid an unnecessary wake_up() function call.
 186                  * The common case is pkmap_count[] == 1, but
 187                  * no waiters.
 188                  * The tasks queued in the wait-queue are guarded
 189                  * by both the lock in the wait-queue-head and by
 190                  * the kmap_lock.  As the kmap_lock is held here,
 191                  * no need for the wait-queue-head's lock.  Simply
 192                  * test if the queue is empty.
 193                  */
 194                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 195         }
 196         spin_unlock(&kmap_lock);
 197
 198         /* do wake-up, if needed, race-free outside of the spin lock */
 199         if (need_wakeup)
 200                 wake_up(&pkmap_map_wait);
 201 }
 202
 203 #define POOL_SIZE       64
 204
 205 static __init int init_emergency_pool(void)
 206 {
 207         struct sysinfo i;
 208         si_meminfo(&i);
 209         si_swapinfo(&i);
 210
 211         if (!i.totalhigh)
 212                 return 0;
 213
 214         page_pool = mempool_create(POOL_SIZE, page_pool_alloc, page_pool_free, NULL);
 215         if (!page_pool)
 216                 BUG();
 217         printk("highmem bounce pool size: %d pages\n", POOL_SIZE);
 218
 219         return 0;
 220 }
 221
 222 __initcall(init_emergency_pool);
 223
 224 /*
 225  * highmem version, map in to vec
 226  */
 227 static inline void bounce_copy_vec(struct bio_vec *to, unsigned char *vfrom)
 228 {
 229         unsigned long flags;
 230         unsigned char *vto;
 231
 232         local_irq_save(flags);
 233         vto = kmap_atomic(to->bv_page, KM_BOUNCE_READ);
 234         memcpy(vto + to->bv_offset, vfrom, to->bv_len);
 235         kunmap_atomic(vto, KM_BOUNCE_READ);
 236         local_irq_restore(flags);
 237 }
 238
 239 #else /* CONFIG_HIGHMEM */
 240
 241 #define bounce_copy_vec(to, vfrom)      \
 242         memcpy(page_address((to)->bv_page) + (to)->bv_offset, vfrom, (to)->bv_len)
 243
 244 #endif
 245
 246 #define ISA_POOL_SIZE   16
 247
 248 /*
 249  * gets called "every" time someone init's a queue with BLK_BOUNCE_ISA
 250  * as the max address, so check if the pool has already been created.
 251  */
 252 int init_emergency_isa_pool(void)
 253 {
 254         if (isa_page_pool)
 255                 return 0;
 256
 257         isa_page_pool = mempool_create(ISA_POOL_SIZE, page_pool_alloc, page_pool_free, (void *) __GFP_DMA);
 258         if (!isa_page_pool)
 259                 BUG();
 260
 261         printk("isa bounce pool size: %d pages\n", ISA_POOL_SIZE);
 262         return 0;
 263 }
 264
 265 /*
 266  * Simple bounce buffer support for highmem pages. Depending on the
 267  * queue gfp mask set, *to may or may not be a highmem page. kmap it
 268  * always, it will do the Right Thing
 269  */
 270 static inline void copy_to_high_bio_irq(struct bio *to, struct bio *from)
 271 {
 272         unsigned char *vfrom;
 273         struct bio_vec *tovec, *fromvec;
 274         int i;
 275
 276         __bio_for_each_segment(tovec, to, i, 0) {
 277                 fromvec = from->bi_io_vec + i;
 278
 279                 /*
 280                  * not bounced
 281                  */
 282                 if (tovec->bv_page == fromvec->bv_page)
 283                         continue;
 284
 285                 vfrom = page_address(fromvec->bv_page) + fromvec->bv_offset;
 286
 287                 bounce_copy_vec(tovec, vfrom);
 288         }
 289 }
 290
 291 static inline void bounce_end_io(struct bio *bio, mempool_t *pool)
 292 {
 293         struct bio *bio_orig = bio->bi_private;
 294         struct bio_vec *bvec, *org_vec;
 295         int i;
 296
 297         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
 298                 goto out_eio;
 299
 300         set_bit(BIO_UPTODATE, &bio_orig->bi_flags);
 301
 302         /*
 303          * free up bounce indirect pages used
 304          */
 305         __bio_for_each_segment(bvec, bio, i, 0) {
 306                 org_vec = bio_orig->bi_io_vec + i;
 307                 if (bvec->bv_page == org_vec->bv_page)
 308                         continue;
 309
 310                 mempool_free(bvec->bv_page, pool);
 311         }
 312
 313 out_eio:
 314         bio_orig->bi_end_io(bio_orig);
 315         bio_put(bio);
 316 }
 317
 318 static void bounce_end_io_write(struct bio *bio)
 319 {
 320         bounce_end_io(bio, page_pool);
 321 }
 322
 323 static void bounce_end_io_write_isa(struct bio *bio)
 324 {
 325         bounce_end_io(bio, isa_page_pool);
 326 }
 327
 328 static inline void __bounce_end_io_read(struct bio *bio, mempool_t *pool)
 329 {
 330         struct bio *bio_orig = bio->bi_private;
 331
 332         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
 333                 copy_to_high_bio_irq(bio_orig, bio);
 334
 335         bounce_end_io(bio, pool);
 336 }
 337
 338 static void bounce_end_io_read(struct bio *bio)
 339 {
 340         __bounce_end_io_read(bio, page_pool);
 341 }
 342
 343 static void bounce_end_io_read_isa(struct bio *bio)
 344 {
 345         return __bounce_end_io_read(bio, isa_page_pool);
 346 }
 347
 348 void create_bounce(unsigned long pfn, int gfp, struct bio **bio_orig)
 349 {
 350         struct page *page;
 351         struct bio *bio = NULL;
 352         int i, rw = bio_data_dir(*bio_orig), bio_gfp;
 353         struct bio_vec *to, *from;
 354         mempool_t *pool;
 355
 356         BUG_ON((*bio_orig)->bi_idx);
 357
 358         /*
 359          * for non-isa bounce case, just check if the bounce pfn is equal
 360          * to or bigger than the highest pfn in the system -- in that case,
 361          * don't waste time iterating over bio segments
 362          */
 363         if (!(gfp & GFP_DMA)) {
 364                 if (pfn >= blk_max_pfn)
 365                         return;
 366
 367                 bio_gfp = GFP_NOHIGHIO;
 368                 pool = page_pool;
 369         } else {
 370                 BUG_ON(!isa_page_pool);
 371                 bio_gfp = GFP_NOIO;
 372                 pool = isa_page_pool;
 373         }
 374
 375         bio_for_each_segment(from, *bio_orig, i) {
 376                 page = from->bv_page;
 377
 378                 /*
 379                  * is destination page below bounce pfn?
 380                  */
 381                 if ((page - page->zone->zone_mem_map) + (page->zone->zone_start_paddr >> PAGE_SHIFT) < pfn)
 382                         continue;
 383
 384                 /*
 385                  * irk, bounce it
 386                  */
 387                 if (!bio)
 388                         bio = bio_alloc(bio_gfp, (*bio_orig)->bi_vcnt);
 389
 390                 to = bio->bi_io_vec + i;
 391
 392                 to->bv_page = mempool_alloc(pool, gfp);
 393                 to->bv_len = from->bv_len;
 394                 to->bv_offset = from->bv_offset;
 395
 396                 if (rw & WRITE) {
 397                         char *vto, *vfrom;
 398
 399                         vto = page_address(to->bv_page) + to->bv_offset;
 400                         vfrom = kmap(from->bv_page) + from->bv_offset;
 401                         memcpy(vto, vfrom, to->bv_len);
 402                         kunmap(from->bv_page);
 403                 }
 404         }
 405
 406         /*
 407          * no pages bounced
 408          */
 409         if (!bio)
 410                 return;
 411
 412         /*
 413          * at least one page was bounced, fill in possible non-highmem
 414          * pages
 415          */
 416         bio_for_each_segment(from, *bio_orig, i) {
 417                 to = &bio->bi_io_vec[i];
 418                 if (!to->bv_page) {
 419                         to->bv_page = from->bv_page;
 420                         to->bv_len = from->bv_len;
 421                         to->bv_offset = to->bv_offset;
 422                 }
 423         }
 424
 425         bio->bi_dev = (*bio_orig)->bi_dev;
 426         bio->bi_sector = (*bio_orig)->bi_sector;
 427         bio->bi_rw = (*bio_orig)->bi_rw;
 428
 429         bio->bi_vcnt = (*bio_orig)->bi_vcnt;
 430         bio->bi_idx = 0;
 431         bio->bi_size = (*bio_orig)->bi_size;
 432
 433         if (pool == page_pool) {
 434                 if (rw & WRITE)
 435                         bio->bi_end_io = bounce_end_io_write;
 436                 else
 437                         bio->bi_end_io = bounce_end_io_read;
 438         } else {
 439                 if (rw & WRITE)
 440                         bio->bi_end_io = bounce_end_io_write_isa;
 441                 else
 442                         bio->bi_end_io = bounce_end_io_read_isa;
 443         }
 444
 445         bio->bi_private = *bio_orig;
 446         *bio_orig = bio;
 447 }