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/highmem.h>
  22 #include <linux/swap.h>
  23 #include <linux/slab.h>
  24
  25 /*
  26  * Virtual_count is not a pure "count".
  27  *  0 means that it is not mapped, and has not been mapped
  28  *    since a TLB flush - it is usable.
  29  *  1 means that there are no users, but it has been mapped
  30  *    since the last TLB flush - so we can't use it.
  31  *  n means that there are (n-1) current users of it.
  32  */
  33 static int pkmap_count[LAST_PKMAP];
  34 static unsigned int last_pkmap_nr;
  35 static spinlock_cacheline_t kmap_lock_cacheline = {SPIN_LOCK_UNLOCKED};
  36 #define kmap_lock  kmap_lock_cacheline.lock
  37
  38 pte_t * pkmap_page_table;
  39
  40 static DECLARE_WAIT_QUEUE_HEAD(pkmap_map_wait);
  41
  42 static void flush_all_zero_pkmaps(void)
  43 {
  44         int i;
  45
  46         flush_cache_all();
  47
  48         for (i = 0; i < LAST_PKMAP; i++) {
  49                 struct page *page;
  50
  51                 /*
  52                  * zero means we don't have anything to do,
  53                  * >1 means that it is still in use. Only
  54                  * a count of 1 means that it is free but
  55                  * needs to be unmapped
  56                  */
  57                 if (pkmap_count[i] != 1)
  58                         continue;
  59                 pkmap_count[i] = 0;
  60
  61                 /* sanity check */
  62                 if (pte_none(pkmap_page_table[i]))
  63                         BUG();
  64
  65                 /*
  66                  * Don't need an atomic fetch-and-clear op here;
  67                  * no-one has the page mapped, and cannot get at
  68                  * its virtual address (and hence PTE) without first
  69                  * getting the kmap_lock (which is held here).
  70                  * So no dangers, even with speculative execution.
  71                  */
  72                 page = pte_page(pkmap_page_table[i]);
  73                 pte_clear(&pkmap_page_table[i]);
  74
  75                 page->virtual = NULL;
  76         }
  77         flush_tlb_all();
  78 }
  79
  80 static inline unsigned long map_new_virtual(struct page *page, int nonblocking)
  81 {
  82         unsigned long vaddr;
  83         int count;
  84
  85 start:
  86         count = LAST_PKMAP;
  87         /* Find an empty entry */
  88         for (;;) {
  89                 last_pkmap_nr = (last_pkmap_nr + 1) & LAST_PKMAP_MASK;
  90                 if (!last_pkmap_nr) {
  91                         flush_all_zero_pkmaps();
  92                         count = LAST_PKMAP;
  93                 }
  94                 if (!pkmap_count[last_pkmap_nr])
  95                         break;  /* Found a usable entry */
  96                 if (--count)
  97                         continue;
  98
  99                 if (nonblocking)
 100                         return 0;
 101
 102                 /*
 103                  * Sleep for somebody else to unmap their entries
 104                  */
 105                 {
 106                         DECLARE_WAITQUEUE(wait, current);
 107
 108                         current->state = TASK_UNINTERRUPTIBLE;
 109                         add_wait_queue(&pkmap_map_wait, &wait);
 110                         spin_unlock(&kmap_lock);
 111                         schedule();
 112                         remove_wait_queue(&pkmap_map_wait, &wait);
 113                         spin_lock(&kmap_lock);
 114
 115                         /* Somebody else might have mapped it while we slept */
 116                         if (page->virtual)
 117                                 return (unsigned long) page->virtual;
 118
 119                         /* Re-start */
 120                         goto start;
 121                 }
 122         }
 123         vaddr = PKMAP_ADDR(last_pkmap_nr);
 124         set_pte(&(pkmap_page_table[last_pkmap_nr]), mk_pte(page, kmap_prot));
 125
 126         pkmap_count[last_pkmap_nr] = 1;
 127         page->virtual = (void *) vaddr;
 128
 129         return vaddr;
 130 }
 131
 132 void fastcall *kmap_high(struct page *page, int nonblocking)
 133 {
 134         unsigned long vaddr;
 135
 136         /*
 137          * For highmem pages, we can't trust "virtual" until
 138          * after we have the lock.
 139          *
 140          * We cannot call this from interrupts, as it may block
 141          */
 142         spin_lock(&kmap_lock);
 143         vaddr = (unsigned long) page->virtual;
 144         if (!vaddr) {
 145                 vaddr = map_new_virtual(page, nonblocking);
 146                 if (!vaddr)
 147                         goto out;
 148         }
 149         pkmap_count[PKMAP_NR(vaddr)]++;
 150         if (pkmap_count[PKMAP_NR(vaddr)] < 2)
 151                 BUG();
 152  out:
 153         spin_unlock(&kmap_lock);
 154         return (void*) vaddr;
 155 }
 156
 157 void fastcall kunmap_high(struct page *page)
 158 {
 159         unsigned long vaddr;
 160         unsigned long nr;
 161         int need_wakeup;
 162
 163         spin_lock(&kmap_lock);
 164         vaddr = (unsigned long) page->virtual;
 165         if (!vaddr)
 166                 BUG();
 167         nr = PKMAP_NR(vaddr);
 168
 169         /*
 170          * A count must never go down to zero
 171          * without a TLB flush!
 172          */
 173         need_wakeup = 0;
 174         switch (--pkmap_count[nr]) {
 175         case 0:
 176                 BUG();
 177         case 1:
 178                 /*
 179                  * Avoid an unnecessary wake_up() function call.
 180                  * The common case is pkmap_count[] == 1, but
 181                  * no waiters.
 182                  * The tasks queued in the wait-queue are guarded
 183                  * by both the lock in the wait-queue-head and by
 184                  * the kmap_lock.  As the kmap_lock is held here,
 185                  * no need for the wait-queue-head's lock.  Simply
 186                  * test if the queue is empty.
 187                  */
 188                 need_wakeup = waitqueue_active(&pkmap_map_wait);
 189         }
 190         spin_unlock(&kmap_lock);
 191
 192         /* do wake-up, if needed, race-free outside of the spin lock */
 193         if (need_wakeup)
 194                 wake_up(&pkmap_map_wait);
 195 }
 196
 197 #define POOL_SIZE 32
 198
 199 /*
 200  * This lock gets no contention at all, normally.
 201  */
 202 static spinlock_t emergency_lock = SPIN_LOCK_UNLOCKED;
 203
 204 int nr_emergency_pages;
 205 static LIST_HEAD(emergency_pages);
 206
 207 int nr_emergency_bhs;
 208 static LIST_HEAD(emergency_bhs);
 209
 210 /*
 211  * Simple bounce buffer support for highmem pages.
 212  * This will be moved to the block layer in 2.5.
 213  */
 214
 215 static inline void copy_from_high_bh (struct buffer_head *to,
 216                          struct buffer_head *from)
 217 {
 218         struct page *p_from;
 219         char *vfrom;
 220
 221         p_from = from->b_page;
 222
 223         vfrom = kmap_atomic(p_from, KM_USER0);
 224         memcpy(to->b_data, vfrom + bh_offset(from), to->b_size);
 225         kunmap_atomic(vfrom, KM_USER0);
 226 }
 227
 228 static inline void copy_to_high_bh_irq (struct buffer_head *to,
 229                          struct buffer_head *from)
 230 {
 231         struct page *p_to;
 232         char *vto;
 233         unsigned long flags;
 234
 235         p_to = to->b_page;
 236         __save_flags(flags);
 237         __cli();
 238         vto = kmap_atomic(p_to, KM_BOUNCE_READ);
 239         memcpy(vto + bh_offset(to), from->b_data, to->b_size);
 240         kunmap_atomic(vto, KM_BOUNCE_READ);
 241         __restore_flags(flags);
 242 }
 243
 244 static inline void bounce_end_io (struct buffer_head *bh, int uptodate)
 245 {
 246         struct page *page;
 247         struct buffer_head *bh_orig = (struct buffer_head *)(bh->b_private);
 248         unsigned long flags;
 249
 250         bh_orig->b_end_io(bh_orig, uptodate);
 251
 252         page = bh->b_page;
 253
 254         spin_lock_irqsave(&emergency_lock, flags);
 255         if (nr_emergency_pages >= POOL_SIZE)
 256                 __free_page(page);
 257         else {
 258                 /*
 259                  * We are abusing page->list to manage
 260                  * the highmem emergency pool:
 261                  */
 262                 list_add(&page->list, &emergency_pages);
 263                 nr_emergency_pages++;
 264         }
 265
 266         if (nr_emergency_bhs >= POOL_SIZE) {
 267 #ifdef HIGHMEM_DEBUG
 268                 /* Don't clobber the constructed slab cache */
 269                 init_waitqueue_head(&bh->b_wait);
 270 #endif
 271                 kmem_cache_free(bh_cachep, bh);
 272         } else {
 273                 /*
 274                  * Ditto in the bh case, here we abuse b_inode_buffers:
 275                  */
 276                 list_add(&bh->b_inode_buffers, &emergency_bhs);
 277                 nr_emergency_bhs++;
 278         }
 279         spin_unlock_irqrestore(&emergency_lock, flags);
 280 }
 281
 282 static __init int init_emergency_pool(void)
 283 {
 284         struct sysinfo i;
 285         si_meminfo(&i);
 286         si_swapinfo(&i);
 287
 288         if (!i.totalhigh)
 289                 return 0;
 290
 291         spin_lock_irq(&emergency_lock);
 292         while (nr_emergency_pages < POOL_SIZE) {
 293                 struct page * page = alloc_page(GFP_ATOMIC);
 294                 if (!page) {
 295                         printk("couldn't refill highmem emergency pages");
 296                         break;
 297                 }
 298                 list_add(&page->list, &emergency_pages);
 299                 nr_emergency_pages++;
 300         }
 301         while (nr_emergency_bhs < POOL_SIZE) {
 302                 struct buffer_head * bh = kmem_cache_alloc(bh_cachep, SLAB_ATOMIC);
 303                 if (!bh) {
 304                         printk("couldn't refill highmem emergency bhs");
 305                         break;
 306                 }
 307                 list_add(&bh->b_inode_buffers, &emergency_bhs);
 308                 nr_emergency_bhs++;
 309         }
 310         spin_unlock_irq(&emergency_lock);
 311         printk("allocated %d pages and %d bhs reserved for the highmem bounces\n",
 312                nr_emergency_pages, nr_emergency_bhs);
 313
 314         return 0;
 315 }
 316
 317 __initcall(init_emergency_pool);
 318
 319 static void bounce_end_io_write (struct buffer_head *bh, int uptodate)
 320 {
 321         bounce_end_io(bh, uptodate);
 322 }
 323
 324 static void bounce_end_io_read (struct buffer_head *bh, int uptodate)
 325 {
 326         struct buffer_head *bh_orig = (struct buffer_head *)(bh->b_private);
 327
 328         if (uptodate)
 329                 copy_to_high_bh_irq(bh_orig, bh);
 330         bounce_end_io(bh, uptodate);
 331 }
 332
 333 struct page *alloc_bounce_page (void)
 334 {
 335         struct list_head *tmp;
 336         struct page *page;
 337
 338         page = alloc_page(GFP_NOHIGHIO);
 339         if (page)
 340                 return page;
 341         /*
 342          * No luck. First, kick the VM so it doesn't idle around while
 343          * we are using up our emergency rations.
 344          */
 345         wakeup_bdflush();
 346
 347 repeat_alloc:
 348         /*
 349          * Try to allocate from the emergency pool.
 350          */
 351         tmp = &emergency_pages;
 352         spin_lock_irq(&emergency_lock);
 353         if (!list_empty(tmp)) {
 354                 page = list_entry(tmp->next, struct page, list);
 355                 list_del(tmp->next);
 356                 nr_emergency_pages--;
 357         }
 358         spin_unlock_irq(&emergency_lock);
 359         if (page)
 360                 return page;
 361
 362         /* we need to wait I/O completion */
 363         run_task_queue(&tq_disk);
 364
 365         yield();
 366         goto repeat_alloc;
 367 }
 368
 369 struct buffer_head *alloc_bounce_bh (void)
 370 {
 371         struct list_head *tmp;
 372         struct buffer_head *bh;
 373
 374         bh = kmem_cache_alloc(bh_cachep, SLAB_NOHIGHIO);
 375         if (bh)
 376                 return bh;
 377         /*
 378          * No luck. First, kick the VM so it doesn't idle around while
 379          * we are using up our emergency rations.
 380          */
 381         wakeup_bdflush();
 382
 383 repeat_alloc:
 384         /*
 385          * Try to allocate from the emergency pool.
 386          */
 387         tmp = &emergency_bhs;
 388         spin_lock_irq(&emergency_lock);
 389         if (!list_empty(tmp)) {
 390                 bh = list_entry(tmp->next, struct buffer_head, b_inode_buffers);
 391                 list_del(tmp->next);
 392                 nr_emergency_bhs--;
 393         }
 394         spin_unlock_irq(&emergency_lock);
 395         if (bh)
 396                 return bh;
 397
 398         /* we need to wait I/O completion */
 399         run_task_queue(&tq_disk);
 400
 401         yield();
 402         goto repeat_alloc;
 403 }
 404
 405 struct buffer_head * create_bounce(int rw, struct buffer_head * bh_orig)
 406 {
 407         struct page *page;
 408         struct buffer_head *bh;
 409
 410         if (!PageHighMem(bh_orig->b_page))
 411                 return bh_orig;
 412
 413         bh = alloc_bounce_bh();
 414         /*
 415          * This is wasteful for 1k buffers, but this is a stopgap measure
 416          * and we are being ineffective anyway. This approach simplifies
 417          * things immensly. On boxes with more than 4GB RAM this should
 418          * not be an issue anyway.
 419          */
 420         page = alloc_bounce_page();
 421
 422         set_bh_page(bh, page, 0);
 423
 424         bh->b_next = NULL;
 425         bh->b_blocknr = bh_orig->b_blocknr;
 426         bh->b_size = bh_orig->b_size;
 427         bh->b_list = -1;
 428         bh->b_dev = bh_orig->b_dev;
 429         bh->b_count = bh_orig->b_count;
 430         bh->b_rdev = bh_orig->b_rdev;
 431         bh->b_state = bh_orig->b_state;
 432 #ifdef HIGHMEM_DEBUG
 433         bh->b_flushtime = jiffies;
 434         bh->b_next_free = NULL;
 435         bh->b_prev_free = NULL;
 436         /* bh->b_this_page */
 437         bh->b_reqnext = NULL;
 438         bh->b_pprev = NULL;
 439 #endif
 440         /* bh->b_page */
 441         if (rw == WRITE) {
 442                 bh->b_end_io = bounce_end_io_write;
 443                 copy_from_high_bh(bh, bh_orig);
 444         } else
 445                 bh->b_end_io = bounce_end_io_read;
 446         bh->b_private = (void *)bh_orig;
 447         bh->b_rsector = bh_orig->b_rsector;
 448 #ifdef HIGHMEM_DEBUG
 449         memset(&bh->b_wait, -1, sizeof(bh->b_wait));
 450 #endif
 451
 452         return bh;
 453 }
 454