mm/bootmem.c

   1 /*
   2  *  linux/mm/bootmem.c
   3  *
   4  *  Copyright (C) 1999 Ingo Molnar
   5  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
   6  *
   7  *  simple boot-time physical memory area allocator and
   8  *  free memory collector. It's used to deal with reserved
   9  *  system memory and memory holes as well.
  10  */
  11
  12 #include <linux/mm.h>
  13 #include <linux/kernel_stat.h>
  14 #include <linux/swap.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/init.h>
  17 #include <linux/bootmem.h>
  18 #include <linux/mmzone.h>
  19 #include <linux/module.h>
  20 #include <asm/dma.h>
  21 #include <asm/io.h>
  22 #include "internal.h"
  23
  24 /*
  25  * Access to this subsystem has to be serialized externally. (this is
  26  * true for the boot process anyway)
  27  */
  28 unsigned long max_low_pfn;
  29 unsigned long min_low_pfn;
  30 unsigned long max_pfn;
  31
  32 EXPORT_SYMBOL(max_pfn);         /* This is exported so
  33                                  * dma_get_required_mask(), which uses
  34                                  * it, can be an inline function */
  35
  36 /* return the number of _pages_ that will be allocated for the boot bitmap */
  37 unsigned long __init bootmem_bootmap_pages (unsigned long pages)
  38 {
  39         unsigned long mapsize;
  40
  41         mapsize = (pages+7)/8;
  42         mapsize = (mapsize + ~PAGE_MASK) & PAGE_MASK;
  43         mapsize >>= PAGE_SHIFT;
  44
  45         return mapsize;
  46 }
  47
  48 /*
  49  * Called once to set up the allocator itself.
  50  */
  51 static unsigned long __init init_bootmem_core (pg_data_t *pgdat,
  52         unsigned long mapstart, unsigned long start, unsigned long end)
  53 {
  54         bootmem_data_t *bdata = pgdat->bdata;
  55         unsigned long mapsize = ((end - start)+7)/8;
  56
  57         pgdat->pgdat_next = pgdat_list;
  58         pgdat_list = pgdat;
  59
  60         mapsize = (mapsize + (sizeof(long) - 1UL)) & ~(sizeof(long) - 1UL);
  61         bdata->node_bootmem_map = phys_to_virt(mapstart << PAGE_SHIFT);
  62         bdata->node_boot_start = (start << PAGE_SHIFT);
  63         bdata->node_low_pfn = end;
  64
  65         /*
  66          * Initially all pages are reserved - setup_arch() has to
  67          * register free RAM areas explicitly.
  68          */
  69         memset(bdata->node_bootmem_map, 0xff, mapsize);
  70
  71         return mapsize;
  72 }
  73
  74 /*
  75  * Marks a particular physical memory range as unallocatable. Usable RAM
  76  * might be used for boot-time allocations - or it might get added
  77  * to the free page pool later on.
  78  */
  79 static void __init reserve_bootmem_core(bootmem_data_t *bdata, unsigned long addr, unsigned long size)
  80 {
  81         unsigned long i;
  82         /*
  83          * round up, partially reserved pages are considered
  84          * fully reserved.
  85          */
  86         unsigned long sidx = (addr - bdata->node_boot_start)/PAGE_SIZE;
  87         unsigned long eidx = (addr + size - bdata->node_boot_start +
  88                                                         PAGE_SIZE-1)/PAGE_SIZE;
  89         unsigned long end = (addr + size + PAGE_SIZE-1)/PAGE_SIZE;
  90
  91         BUG_ON(!size);
  92         BUG_ON(sidx >= eidx);
  93         BUG_ON((addr >> PAGE_SHIFT) >= bdata->node_low_pfn);
  94         BUG_ON(end > bdata->node_low_pfn);
  95
  96         for (i = sidx; i < eidx; i++)
  97                 if (test_and_set_bit(i, bdata->node_bootmem_map)) {
  98 #ifdef CONFIG_DEBUG_BOOTMEM
  99                         printk("hm, page %08lx reserved twice.\n", i*PAGE_SIZE);
 100 #endif
 101                 }
 102 }
 103
 104 static void __init free_bootmem_core(bootmem_data_t *bdata, unsigned long addr, unsigned long size)
 105 {
 106         unsigned long i;
 107         unsigned long start;
 108         /*
 109          * round down end of usable mem, partially free pages are
 110          * considered reserved.
 111          */
 112         unsigned long sidx;
 113         unsigned long eidx = (addr + size - bdata->node_boot_start)/PAGE_SIZE;
 114         unsigned long end = (addr + size)/PAGE_SIZE;
 115
 116         BUG_ON(!size);
 117         BUG_ON(end > bdata->node_low_pfn);
 118
 119         if (addr < bdata->last_success)
 120                 bdata->last_success = addr;
 121
 122         /*
 123          * Round up the beginning of the address.
 124          */
 125         start = (addr + PAGE_SIZE-1) / PAGE_SIZE;
 126         sidx = start - (bdata->node_boot_start/PAGE_SIZE);
 127
 128         for (i = sidx; i < eidx; i++) {
 129                 if (unlikely(!test_and_clear_bit(i, bdata->node_bootmem_map)))
 130                         BUG();
 131         }
 132 }
 133
 134 /*
 135  * We 'merge' subsequent allocations to save space. We might 'lose'
 136  * some fraction of a page if allocations cannot be satisfied due to
 137  * size constraints on boxes where there is physical RAM space
 138  * fragmentation - in these cases (mostly large memory boxes) this
 139  * is not a problem.
 140  *
 141  * On low memory boxes we get it right in 100% of the cases.
 142  *
 143  * alignment has to be a power of 2 value.
 144  *
 145  * NOTE:  This function is _not_ reentrant.
 146  */
 147 static void * __init
 148 __alloc_bootmem_core(struct bootmem_data *bdata, unsigned long size,
 149                 unsigned long align, unsigned long goal)
 150 {
 151         unsigned long offset, remaining_size, areasize, preferred;
 152         unsigned long i, start = 0, incr, eidx;
 153         void *ret;
 154
 155         if(!size) {
 156                 printk("__alloc_bootmem_core(): zero-sized request\n");
 157                 BUG();
 158         }
 159         BUG_ON(align & (align-1));
 160
 161         eidx = bdata->node_low_pfn - (bdata->node_boot_start >> PAGE_SHIFT);
 162         offset = 0;
 163         if (align &&
 164             (bdata->node_boot_start & (align - 1UL)) != 0)
 165                 offset = (align - (bdata->node_boot_start & (align - 1UL)));
 166         offset >>= PAGE_SHIFT;
 167
 168         /*
 169          * We try to allocate bootmem pages above 'goal'
 170          * first, then we try to allocate lower pages.
 171          */
 172         if (goal && (goal >= bdata->node_boot_start) &&
 173             ((goal >> PAGE_SHIFT) < bdata->node_low_pfn)) {
 174                 preferred = goal - bdata->node_boot_start;
 175
 176                 if (bdata->last_success >= preferred)
 177                         preferred = bdata->last_success;
 178         } else
 179                 preferred = 0;
 180
 181         preferred = ((preferred + align - 1) & ~(align - 1)) >> PAGE_SHIFT;
 182         preferred += offset;
 183         areasize = (size+PAGE_SIZE-1)/PAGE_SIZE;
 184         incr = align >> PAGE_SHIFT ? : 1;
 185
 186 restart_scan:
 187         for (i = preferred; i < eidx; i += incr) {
 188                 unsigned long j;
 189                 i = find_next_zero_bit(bdata->node_bootmem_map, eidx, i);
 190                 i = ALIGN(i, incr);
 191                 if (test_bit(i, bdata->node_bootmem_map))
 192                         continue;
 193                 for (j = i + 1; j < i + areasize; ++j) {
 194                         if (j >= eidx)
 195                                 goto fail_block;
 196                         if (test_bit (j, bdata->node_bootmem_map))
 197                                 goto fail_block;
 198                 }
 199                 start = i;
 200                 goto found;
 201         fail_block:
 202                 i = ALIGN(j, incr);
 203         }
 204
 205         if (preferred > offset) {
 206                 preferred = offset;
 207                 goto restart_scan;
 208         }
 209         return NULL;
 210
 211 found:
 212         bdata->last_success = start << PAGE_SHIFT;
 213         BUG_ON(start >= eidx);
 214
 215         /*
 216          * Is the next page of the previous allocation-end the start
 217          * of this allocation's buffer? If yes then we can 'merge'
 218          * the previous partial page with this allocation.
 219          */
 220         if (align < PAGE_SIZE &&
 221             bdata->last_offset && bdata->last_pos+1 == start) {
 222                 offset = (bdata->last_offset+align-1) & ~(align-1);
 223                 BUG_ON(offset > PAGE_SIZE);
 224                 remaining_size = PAGE_SIZE-offset;
 225                 if (size < remaining_size) {
 226                         areasize = 0;
 227                         /* last_pos unchanged */
 228                         bdata->last_offset = offset+size;
 229                         ret = phys_to_virt(bdata->last_pos*PAGE_SIZE + offset +
 230                                                 bdata->node_boot_start);
 231                 } else {
 232                         remaining_size = size - remaining_size;
 233                         areasize = (remaining_size+PAGE_SIZE-1)/PAGE_SIZE;
 234                         ret = phys_to_virt(bdata->last_pos*PAGE_SIZE + offset +
 235                                                 bdata->node_boot_start);
 236                         bdata->last_pos = start+areasize-1;
 237                         bdata->last_offset = remaining_size;
 238                 }
 239                 bdata->last_offset &= ~PAGE_MASK;
 240         } else {
 241                 bdata->last_pos = start + areasize - 1;
 242                 bdata->last_offset = size & ~PAGE_MASK;
 243                 ret = phys_to_virt(start * PAGE_SIZE + bdata->node_boot_start);
 244         }
 245
 246         /*
 247          * Reserve the area now:
 248          */
 249         for (i = start; i < start+areasize; i++)
 250                 if (unlikely(test_and_set_bit(i, bdata->node_bootmem_map)))
 251                         BUG();
 252         memset(ret, 0, size);
 253         return ret;
 254 }
 255
 256 static unsigned long __init free_all_bootmem_core(pg_data_t *pgdat)
 257 {
 258         struct page *page;
 259         unsigned long pfn;
 260         bootmem_data_t *bdata = pgdat->bdata;
 261         unsigned long i, count, total = 0;
 262         unsigned long idx;
 263         unsigned long *map;
 264         int gofast = 0;
 265
 266         BUG_ON(!bdata->node_bootmem_map);
 267
 268         count = 0;
 269         /* first extant page of the node */
 270         pfn = bdata->node_boot_start >> PAGE_SHIFT;
 271         idx = bdata->node_low_pfn - (bdata->node_boot_start >> PAGE_SHIFT);
 272         map = bdata->node_bootmem_map;
 273         /* Check physaddr is O(LOG2(BITS_PER_LONG)) page aligned */
 274         if (bdata->node_boot_start == 0 ||
 275             ffs(bdata->node_boot_start) - PAGE_SHIFT > ffs(BITS_PER_LONG))
 276                 gofast = 1;
 277         for (i = 0; i < idx; ) {
 278                 unsigned long v = ~map[i / BITS_PER_LONG];
 279
 280                 if (gofast && v == ~0UL) {
 281                         int j, order;
 282
 283                         page = pfn_to_page(pfn);
 284                         count += BITS_PER_LONG;
 285                         __ClearPageReserved(page);
 286                         order = ffs(BITS_PER_LONG) - 1;
 287                         set_page_refs(page, order);
 288                         for (j = 1; j < BITS_PER_LONG; j++) {
 289                                 if (j + 16 < BITS_PER_LONG)
 290                                         prefetchw(page + j + 16);
 291                                 __ClearPageReserved(page + j);
 292                         }
 293                         __free_pages(page, order);
 294                         i += BITS_PER_LONG;
 295                         page += BITS_PER_LONG;
 296                 } else if (v) {
 297                         unsigned long m;
 298
 299                         page = pfn_to_page(pfn);
 300                         for (m = 1; m && i < idx; m<<=1, page++, i++) {
 301                                 if (v & m) {
 302                                         count++;
 303                                         __ClearPageReserved(page);
 304                                         set_page_refs(page, 0);
 305                                         __free_page(page);
 306                                 }
 307                         }
 308                 } else {
 309                         i+=BITS_PER_LONG;
 310                 }
 311                 pfn += BITS_PER_LONG;
 312         }
 313         total += count;
 314
 315         /*
 316          * Now free the allocator bitmap itself, it's not
 317          * needed anymore:
 318          */
 319         page = virt_to_page(bdata->node_bootmem_map);
 320         count = 0;
 321         for (i = 0; i < ((bdata->node_low_pfn-(bdata->node_boot_start >> PAGE_SHIFT))/8 + PAGE_SIZE-1)/PAGE_SIZE; i++,page++) {
 322                 count++;
 323                 __ClearPageReserved(page);
 324                 set_page_count(page, 1);
 325                 __free_page(page);
 326         }
 327         total += count;
 328         bdata->node_bootmem_map = NULL;
 329
 330         return total;
 331 }
 332
 333 unsigned long __init init_bootmem_node (pg_data_t *pgdat, unsigned long freepfn, unsigned long startpfn, unsigned long endpfn)
 334 {
 335         return(init_bootmem_core(pgdat, freepfn, startpfn, endpfn));
 336 }
 337
 338 void __init reserve_bootmem_node (pg_data_t *pgdat, unsigned long physaddr, unsigned long size)
 339 {
 340         reserve_bootmem_core(pgdat->bdata, physaddr, size);
 341 }
 342
 343 void __init free_bootmem_node (pg_data_t *pgdat, unsigned long physaddr, unsigned long size)
 344 {
 345         free_bootmem_core(pgdat->bdata, physaddr, size);
 346 }
 347
 348 unsigned long __init free_all_bootmem_node (pg_data_t *pgdat)
 349 {
 350         return(free_all_bootmem_core(pgdat));
 351 }
 352
 353 unsigned long __init init_bootmem (unsigned long start, unsigned long pages)
 354 {
 355         max_low_pfn = pages;
 356         min_low_pfn = start;
 357         return(init_bootmem_core(NODE_DATA(0), start, 0, pages));
 358 }
 359
 360 #ifndef CONFIG_HAVE_ARCH_BOOTMEM_NODE
 361 void __init reserve_bootmem (unsigned long addr, unsigned long size)
 362 {
 363         reserve_bootmem_core(NODE_DATA(0)->bdata, addr, size);
 364 }
 365 #endif /* !CONFIG_HAVE_ARCH_BOOTMEM_NODE */
 366
 367 void __init free_bootmem (unsigned long addr, unsigned long size)
 368 {
 369         free_bootmem_core(NODE_DATA(0)->bdata, addr, size);
 370 }
 371
 372 unsigned long __init free_all_bootmem (void)
 373 {
 374         return(free_all_bootmem_core(NODE_DATA(0)));
 375 }
 376
 377 void * __init __alloc_bootmem (unsigned long size, unsigned long align, unsigned long goal)
 378 {
 379         pg_data_t *pgdat = pgdat_list;
 380         void *ptr;
 381
 382         for_each_pgdat(pgdat)
 383                 if ((ptr = __alloc_bootmem_core(pgdat->bdata, size,
 384                                                 align, goal)))
 385                         return(ptr);
 386
 387         /*
 388          * Whoops, we cannot satisfy the allocation request.
 389          */
 390         printk(KERN_ALERT "bootmem alloc of %lu bytes failed!\n", size);
 391         panic("Out of memory");
 392         return NULL;
 393 }
 394
 395 void * __init __alloc_bootmem_node (pg_data_t *pgdat, unsigned long size, unsigned long align, unsigned long goal)
 396 {
 397         void *ptr;
 398
 399         ptr = __alloc_bootmem_core(pgdat->bdata, size, align, goal);
 400         if (ptr)
 401                 return (ptr);
 402
 403         return __alloc_bootmem(size, align, goal);
 404 }
 405