root/mm/kmalloc.c

/* */

DEFINITIONS

1. kmalloc_init
2. get_order
3. kmalloc
4. kfree_s

   1 /*
   2  *  linux/mm/kmalloc.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds & Roger Wolff.
   5  *
   6  *  Written by R.E. Wolff Sept/Oct '93.
   7  *
   8  */
   9 
  10 /*
  11  * Modified by Alex Bligh (alex@cconcepts.co.uk) 4 Apr 1994 to use multiple
  12  * pages. So for 'page' throughout, read 'area'.
  13  */
  14 
  15 #include <linux/mm.h>
  16 #include <asm/system.h>
  17 #include <linux/delay.h>
  18 
  19 #define GFP_LEVEL_MASK 0xf
  20 
  21 /* I want this low enough for a while to catch errors.
  22    I want this number to be increased in the near future:
  23         loadable device drivers should use this function to get memory */
  24 
  25 #define MAX_KMALLOC_K ((PAGE_SIZE<<(NUM_AREA_ORDERS-1))>>10)
  26 
  27 
  28 /* This defines how many times we should try to allocate a free page before
  29    giving up. Normally this shouldn't happen at all. */
  30 #define MAX_GET_FREE_PAGE_TRIES 4
  31 
  32 
  33 /* Private flags. */
  34 
  35 #define MF_USED 0xffaa0055
  36 #define MF_FREE 0x0055ffaa
  37 
  38 
  39 /* 
  40  * Much care has gone into making these routines in this file reentrant.
  41  *
  42  * The fancy bookkeeping of nbytesmalloced and the like are only used to
  43  * report them to the user (oooohhhhh, aaaaahhhhh....) are not 
  44  * protected by cli(). (If that goes wrong. So what?)
  45  *
  46  * These routines restore the interrupt status to allow calling with ints
  47  * off. 
  48  */
  49 
  50 /* 
  51  * A block header. This is in front of every malloc-block, whether free or not.
  52  */
  53 struct block_header {
  54         unsigned long bh_flags;
  55         union {
  56                 unsigned long ubh_length;
  57                 struct block_header *fbh_next;
  58         } vp;
  59 };
  60 
  61 
  62 #define bh_length vp.ubh_length
  63 #define bh_next   vp.fbh_next
  64 #define BH(p) ((struct block_header *)(p))
  65 
  66 
  67 /* 
  68  * The page descriptor is at the front of every page that malloc has in use. 
  69  */
  70 struct page_descriptor {
  71         struct page_descriptor *next;
  72         struct block_header *firstfree;
  73         int order;
  74         int nfree;
  75 };
  76 
  77 
  78 #define PAGE_DESC(p) ((struct page_descriptor *)(((unsigned long)(p)) & PAGE_MASK))
  79 
  80 
  81 /*
  82  * A size descriptor describes a specific class of malloc sizes.
  83  * Each class of sizes has its own freelist.
  84  */
  85 struct size_descriptor {
  86         struct page_descriptor *firstfree;
  87         struct page_descriptor *dmafree; /* DMA-able memory */
  88         int size;
  89         int nblocks;
  90 
  91         int nmallocs;
  92         int nfrees;
  93         int nbytesmalloced;
  94         int npages;
  95         unsigned long gfporder; /* number of pages in the area required */
  96 };
  97 
  98 /*
  99  * For now it is unsafe to allocate bucket sizes between n & n=16 where n is
 100  * 4096 * any power of two
 101  */
 102 
 103 struct size_descriptor sizes[] = { 
 104         { NULL, NULL,  32,127, 0,0,0,0, 0},
 105         { NULL, NULL,  64, 63, 0,0,0,0, 0 },
 106         { NULL, NULL, 128, 31, 0,0,0,0, 0 },
 107         { NULL, NULL, 252, 16, 0,0,0,0, 0 },
 108         { NULL, NULL, 508,  8, 0,0,0,0, 0 },
 109         { NULL, NULL,1020,  4, 0,0,0,0, 0 },
 110         { NULL, NULL,2040,  2, 0,0,0,0, 0 },
 111         { NULL, NULL,4096-16,  1, 0,0,0,0, 0 },
 112         { NULL, NULL,8192-16,  1, 0,0,0,0, 1 },
 113         { NULL, NULL,16384-16,  1, 0,0,0,0, 2 },
 114         { NULL, NULL,32768-16,  1, 0,0,0,0, 3 },
 115         { NULL, NULL,65536-16,  1, 0,0,0,0, 4 },
 116         { NULL, NULL,131072-16,  1, 0,0,0,0, 5 },
 117         { NULL, NULL,   0,  0, 0,0,0,0, 0 }
 118 };
 119 
 120 
 121 #define NBLOCKS(order)          (sizes[order].nblocks)
 122 #define BLOCKSIZE(order)        (sizes[order].size)
 123 #define AREASIZE(order)         (PAGE_SIZE<<(sizes[order].gfporder))
 124 
 125 
 126 long kmalloc_init (long start_mem,long end_mem)
     /*  */
 127 {
 128         int order;
 129 
 130 /* 
 131  * Check the static info array. Things will blow up terribly if it's
 132  * incorrect. This is a late "compile time" check.....
 133  */
 134 for (order = 0;BLOCKSIZE(order);order++)
 135     {
 136     if ((NBLOCKS (order)*BLOCKSIZE(order) + sizeof (struct page_descriptor)) >
 137         AREASIZE(order)) 
 138         {
 139         printk ("Cannot use %d bytes out of %d in order = %d block mallocs\n",
 140                 (int) (NBLOCKS (order) * BLOCKSIZE(order) + 
 141                         sizeof (struct page_descriptor)),
 142                 (int) AREASIZE(order),
 143                 BLOCKSIZE (order));
 144         panic ("This only happens if someone messes with kmalloc");
 145         }
 146     }
 147 return start_mem;
 148 }
 149 
 150 
 151 
 152 int get_order (int size)
     /*  */
 153 {
 154         int order;
 155 
 156         /* Add the size of the header */
 157         size += sizeof (struct block_header); 
 158         for (order = 0;BLOCKSIZE(order);order++)
 159                 if (size <= BLOCKSIZE (order))
 160                         return order; 
 161         return -1;
 162 }
 163 
 164 void * kmalloc (size_t size, int priority)
     /*  */
 165 {
 166         unsigned long flags;
 167         int order,tries,i,sz;
 168         int dma_flag;
 169         struct block_header *p;
 170         struct page_descriptor *page;
 171 
 172         dma_flag = (priority & GFP_DMA);
 173         priority &= GFP_LEVEL_MASK;
 174           
 175 /* Sanity check... */
 176         if (intr_count && priority != GFP_ATOMIC) {
 177                 static int count = 0;
 178                 if (++count < 5) {
 179                         printk("kmalloc called nonatomically from interrupt %p\n",
 180                                 __builtin_return_address(0));
 181                         priority = GFP_ATOMIC;
 182                 }
 183         }
 184 
 185 order = get_order (size);
 186 if (order < 0)
 187     {
 188     printk ("kmalloc of too large a block (%d bytes).\n",(int) size);
 189     return (NULL);
 190     }
 191 
 192 save_flags(flags);
 193 
 194 /* It seems VERY unlikely to me that it would be possible that this 
 195    loop will get executed more than once. */
 196 tries = MAX_GET_FREE_PAGE_TRIES; 
 197 while (tries --)
 198     {
 199     /* Try to allocate a "recently" freed memory block */
 200     cli ();
 201     if ((page = (dma_flag ? sizes[order].dmafree : sizes[order].firstfree)) &&
 202         (p    =  page->firstfree))
 203         {
 204         if (p->bh_flags == MF_FREE)
 205             {
 206             page->firstfree = p->bh_next;
 207             page->nfree--;
 208             if (!page->nfree)
 209                 {
 210                 sizes[order].firstfree = page->next;
 211                 page->next = NULL;
 212                 }
 213             restore_flags(flags);
 214 
 215             sizes [order].nmallocs++;
 216             sizes [order].nbytesmalloced += size;
 217             p->bh_flags =  MF_USED; /* As of now this block is officially in use */
 218             p->bh_length = size;
 219             return p+1; /* Pointer arithmetic: increments past header */
 220             }
 221         printk ("Problem: block on freelist at %08lx isn't free.\n",(long)p);
 222         return (NULL);
 223         }
 224     restore_flags(flags);
 225 
 226 
 227     /* Now we're in trouble: We need to get a new free page..... */
 228 
 229     sz = BLOCKSIZE(order); /* sz is the size of the blocks we're dealing with */
 230 
 231     /* This can be done with ints on: This is private to this invocation */
 232     if (dma_flag)
 233       page = (struct page_descriptor *) __get_dma_pages (priority & GFP_LEVEL_MASK, sizes[order].gfporder);
 234     else
 235       page = (struct page_descriptor *) __get_free_pages (priority & GFP_LEVEL_MASK, sizes[order].gfporder);
 236 
 237     if (!page) {
 238         static unsigned long last = 0;
 239         if (last + 10*HZ < jiffies) {
 240                 last = jiffies;
 241                 printk ("Couldn't get a free page.....\n");
 242         }
 243         return NULL;
 244     }
 245 #if 0
 246     printk ("Got page %08x to use for %d byte mallocs....",(long)page,sz);
 247 #endif
 248     sizes[order].npages++;
 249 
 250     /* Loop for all but last block: */
 251     for (i=NBLOCKS(order),p=BH (page+1);i > 1;i--,p=p->bh_next) 
 252         {
 253         p->bh_flags = MF_FREE;
 254         p->bh_next = BH ( ((long)p)+sz);
 255         }
 256     /* Last block: */
 257     p->bh_flags = MF_FREE;
 258     p->bh_next = NULL;
 259 
 260     page->order = order;
 261     page->nfree = NBLOCKS(order); 
 262     page->firstfree = BH(page+1);
 263 #if 0
 264     printk ("%d blocks per page\n",page->nfree);
 265 #endif
 266     /* Now we're going to muck with the "global" freelist for this size:
 267        this should be uninterruptible */
 268     cli ();
 269     /* 
 270      * sizes[order].firstfree used to be NULL, otherwise we wouldn't be
 271      * here, but you never know.... 
 272      */
 273     if (dma_flag) {
 274       page->next = sizes[order].dmafree;
 275       sizes[order].dmafree = page;
 276     } else {
 277       page->next = sizes[order].firstfree;
 278       sizes[order].firstfree = page;
 279     }
 280     restore_flags(flags);
 281     }
 282 
 283 /* Pray that printk won't cause this to happen again :-) */
 284 
 285 printk ("Hey. This is very funny. I tried %d times to allocate a whole\n"
 286         "new page for an object only %d bytes long, but some other process\n"
 287         "beat me to actually allocating it. Also note that this 'error'\n"
 288         "message is soooo very long to catch your attention. I'd appreciate\n"
 289         "it if you'd be so kind as to report what conditions caused this to\n"
 290         "the author of this kmalloc: wolff@dutecai.et.tudelft.nl.\n"
 291         "(Executive summary: This can't happen)\n", 
 292                 MAX_GET_FREE_PAGE_TRIES,
 293                 (int) size);
 294 return NULL;
 295 }
 296 
 297 void kfree_s (void *ptr,int size)
     /*  */
 298 {
 299 unsigned long flags;
 300 int order;
 301 register struct block_header *p=((struct block_header *)ptr) -1;
 302 struct page_descriptor *page,*pg2;
 303 
 304 page = PAGE_DESC (p);
 305 order = page->order;
 306 if ((order < 0) || 
 307     (order > sizeof (sizes)/sizeof (sizes[0])) ||
 308     (((long)(page->next)) & ~PAGE_MASK) ||
 309     (p->bh_flags != MF_USED))
 310     {
 311     printk ("kfree of non-kmalloced memory: %p, next= %p, order=%d\n",
 312                 p, page->next, page->order);
 313     return;
 314     }
 315 if (size &&
 316     size != p->bh_length)
 317     {
 318     printk ("Trying to free pointer at %p with wrong size: %d instead of %lu.\n",
 319         p,size,p->bh_length);
 320     return;
 321     }
 322 size = p->bh_length;
 323 p->bh_flags = MF_FREE; /* As of now this block is officially free */
 324 save_flags(flags);
 325 cli ();
 326 p->bh_next = page->firstfree;
 327 page->firstfree = p;
 328 page->nfree ++;
 329 
 330 if (page->nfree == 1)
 331    { /* Page went from full to one free block: put it on the freelist.  Do not bother
 332       trying to put it on the DMA list. */
 333    if (page->next)
 334         {
 335         printk ("Page %p already on freelist dazed and confused....\n", page);
 336         }
 337    else
 338         {
 339         page->next = sizes[order].firstfree;
 340         sizes[order].firstfree = page;
 341         }
 342    }
 343 
 344 /* If page is completely free, free it */
 345 if (page->nfree == NBLOCKS (page->order))
 346     {
 347 #if 0
 348     printk ("Freeing page %08x.\n", (long)page);
 349 #endif
 350     if (sizes[order].firstfree == page)
 351         {
 352         sizes[order].firstfree = page->next;
 353         }
 354     else if (sizes[order].dmafree == page)
 355         {
 356         sizes[order].dmafree = page->next;
 357         }
 358     else
 359         {
 360         for (pg2=sizes[order].firstfree;
 361                 (pg2 != NULL) && (pg2->next != page);
 362                         pg2=pg2->next)
 363             /* Nothing */;
 364         if (!pg2)
 365           for (pg2=sizes[order].dmafree;
 366                (pg2 != NULL) && (pg2->next != page);
 367                pg2=pg2->next)
 368             /* Nothing */;
 369         if (pg2 != NULL)
 370             pg2->next = page->next;
 371         else
 372             printk ("Ooops. page %p doesn't show on freelist.\n", page);
 373         }
 374 /* FIXME: I'm sure we should do something with npages here (like npages--) */
 375     free_pages ((long)page, sizes[order].gfporder);
 376     }
 377 restore_flags(flags);
 378 
 379 /* FIXME: ?? Are these increment & decrement operations guaranteed to be
 380  *           atomic? Could an IRQ not occur between the read & the write?
 381  *           Maybe yes on a x86 with GCC...??
 382  */
 383 sizes[order].nfrees++;      /* Noncritical (monitoring) admin stuff */
 384 sizes[order].nbytesmalloced -= size;
 385 }

/* */