src/dir/buffer.c

   1 /*
   2  * Copyright 2000, International Business Machines Corporation and others.
   3  * All Rights Reserved.
   4  *
   5  * This software has been released under the terms of the IBM Public
   6  * License.  For details, see the LICENSE file in the top-level source
   7  * directory or online at http://www.openafs.org/dl/license10.html
   8  */
   9
  10 #include <afsconfig.h>
  11 #include <afs/param.h>
  12
  13 #include <roken.h>
  14 #include <afs/opr.h>
  15
  16 #include <lock.h>
  17
  18 #include "dir.h"
  19
  20 #ifdef AFS_64BIT_IOPS_ENV
  21 #define BUFFER_FID_SIZE (9*sizeof(int) + 2*sizeof(char*))
  22 #else
  23 #define BUFFER_FID_SIZE (6*sizeof(int) + 2*sizeof(char*))
  24 #endif
  25
  26 struct buffer {
  27     /* fid is used for Unique cache key + i/o addressing.
  28      * fid size is based on 4 + size of inode and size of pointer
  29      */
  30     char fid[BUFFER_FID_SIZE];
  31     afs_int32 page;
  32     afs_int32 accesstime;
  33     struct buffer *hashNext;
  34     void *data;
  35     char lockers;
  36     char dirty;
  37     char hashIndex;
  38     struct Lock lock;
  39 };
  40
  41 static_inline dir_file_t
  42 bufferDir(struct buffer *b)
  43 {
  44     return (dir_file_t) &b->fid;
  45 }
  46
  47 struct Lock afs_bufferLock;
  48
  49 /* page size */
  50 #define BUFFER_PAGE_SIZE 2048
  51 /* log page size */
  52 #define LOGPS 11
  53 /* page hash table size */
  54 #define PHSIZE 32
  55 /* The hash table should be somewhat efficient even if there are only
  56  * a few partitions (less than 32). So the hash for the fileserver is now
  57  * based on the volume id. This means this macro is dependent upon the
  58  * layout of DirHandle in viced/viced.h, vol/salvage.h and volser/salvage.h.
  59  */
  60 #define pHash(fid) (((afs_int32 *)fid)[0] & (PHSIZE-1))
  61 #define vHash(vid) (vid & (PHSIZE-1))
  62
  63 /* admittedly system dependent, this is the maximum signed 32-bit value */
  64 #define BUFFER_LONG_MAX   2147483647
  65 #ifndef NULL
  66 #define NULL 0
  67 #endif
  68
  69 static struct buffer **Buffers;
  70
  71 char *BufferData;
  72
  73 static struct buffer *phTable[PHSIZE];  /* page hash table */
  74 static struct buffer *LastBuffer;
  75 int nbuffers;
  76 int timecounter;
  77 static int calls = 0, ios = 0;
  78
  79 struct buffer *newslot(dir_file_t dir, afs_int32 apage,
  80                        struct buffer *lp);
  81
  82 /* XXX - This sucks. The correct prototypes for these functions are ...
  83  *
  84  * extern void FidZero(DirHandle *);
  85  * extern int  FidEq(DirHandle *a, DirHandle *b);
  86  * extern int  ReallyRead(DirHandle *a, int block, char *data);
  87  */
  88
  89 extern void FidZero(dir_file_t);
  90 extern int FidEq(dir_file_t, dir_file_t);
  91 extern int ReallyRead(dir_file_t, int block, char *data);
  92 extern int ReallyWrite(dir_file_t, int block, char *data);
  93 extern void FidZap(dir_file_t);
  94 extern int  FidVolEq(dir_file_t, afs_int32 vid);
  95 extern void FidCpy(dir_file_t, dir_file_t fromfile);
  96
  97 int
  98 DStat(int *abuffers, int *acalls, int *aios)
  99 {
 100     *abuffers = nbuffers;
 101     *acalls = calls;
 102     *aios = ios;
 103     return 0;
 104 }
 105
 106 /**
 107  * initialize the directory package.
 108  *
 109  * @param[in] abuffers  size of directory buffer cache
 110  *
 111  * @return operation status
 112  *    @retval 0 success
 113  */
 114 void
 115 DInit(int abuffers)
 116 {
 117     /* Initialize the venus buffer system. */
 118     int i, tsize;
 119     struct buffer *tb;
 120     char *tp;
 121
 122     Lock_Init(&afs_bufferLock);
 123     /* Align each element of Buffers on a doubleword boundary */
 124     tsize = (sizeof(struct buffer) + 7) & ~7;
 125     tp = malloc(abuffers * tsize);
 126     Buffers = malloc(abuffers * sizeof(struct buffer *));
 127     BufferData = malloc(abuffers * BUFFER_PAGE_SIZE);
 128     timecounter = 0;
 129     LastBuffer = (struct buffer *)tp;
 130     nbuffers = abuffers;
 131     for (i = 0; i < PHSIZE; i++)
 132         phTable[i] = 0;
 133     for (i = 0; i < abuffers; i++) {
 134         /* Fill in each buffer with an empty indication. */
 135         tb = (struct buffer *)tp;
 136         Buffers[i] = tb;
 137         tp += tsize;
 138         FidZero(bufferDir(tb));
 139         tb->accesstime = tb->lockers = 0;
 140         tb->data = &BufferData[BUFFER_PAGE_SIZE * i];
 141         tb->hashIndex = 0;
 142         tb->dirty = 0;
 143         Lock_Init(&tb->lock);
 144     }
 145     return;
 146 }
 147
 148 /**
 149  * read a page out of a directory object.
 150  *
 151  * @param[in] fid   directory object fid
 152  * @param[in] page  page in hash table to be read
 153  *
 154  * @return pointer to requested page in directory cache
 155  *    @retval NULL read failed
 156  */
 157 int
 158 DRead(dir_file_t fid, int page, struct DirBuffer *entry)
 159 {
 160     /* Read a page from the disk. */
 161     struct buffer *tb, *tb2, **bufhead;
 162
 163     memset(entry, 0, sizeof(struct DirBuffer));
 164
 165     ObtainWriteLock(&afs_bufferLock);
 166     calls++;
 167
 168 #define bufmatch(tb,fid) (tb->page == page && FidEq(bufferDir(tb), fid))
 169 #define buf_Front(head,parent,p) {(parent)->hashNext = (p)->hashNext; (p)->hashNext= *(head);*(head)=(p);}
 170
 171     /* this apparently-complicated-looking code is simply an example of
 172      * a little bit of loop unrolling, and is a standard linked-list
 173      * traversal trick. It saves a few assignments at the the expense
 174      * of larger code size.  This could be simplified by better use of
 175      * macros.  With the use of these LRU queues, the old one-cache is
 176      * probably obsolete.
 177      */
 178     if ((tb = phTable[pHash(fid)])) {   /* ASSMT HERE */
 179         if (bufmatch(tb, fid)) {
 180             ObtainWriteLock(&tb->lock);
 181             tb->lockers++;
 182             ReleaseWriteLock(&afs_bufferLock);
 183             tb->accesstime = ++timecounter;
 184             ReleaseWriteLock(&tb->lock);
 185             entry->buffer = tb;
 186             entry->data = tb->data;
 187             return 0;
 188         } else {
 189             bufhead = &(phTable[pHash(fid)]);
 190             while ((tb2 = tb->hashNext)) {
 191                 if (bufmatch(tb2, fid)) {
 192                     buf_Front(bufhead, tb, tb2);
 193                     ObtainWriteLock(&tb2->lock);
 194                     tb2->lockers++;
 195                     ReleaseWriteLock(&afs_bufferLock);
 196                     tb2->accesstime = ++timecounter;
 197                     ReleaseWriteLock(&tb2->lock);
 198                     entry->buffer = tb2;
 199                     entry->data = tb2->data;
 200                     return 0;
 201                 }
 202                 if ((tb = tb2->hashNext)) {     /* ASSIGNMENT HERE! */
 203                     if (bufmatch(tb, fid)) {
 204                         buf_Front(bufhead, tb2, tb);
 205                         ObtainWriteLock(&tb->lock);
 206                         tb->lockers++;
 207                         ReleaseWriteLock(&afs_bufferLock);
 208                         tb->accesstime = ++timecounter;
 209                         ReleaseWriteLock(&tb->lock);
 210                         entry->buffer = tb;
 211                         entry->data = tb->data;
 212                         return 0;
 213                     }
 214                 } else
 215                     break;
 216             }
 217         }
 218     } else
 219         tb2 = NULL;
 220
 221     /* can't find it */
 222     /* The last thing we looked at was either tb or tb2 (or nothing). That
 223      * is at least the oldest buffer on one particular hash chain, so it's
 224      * a pretty good place to start looking for the truly oldest buffer.
 225      */
 226     tb = newslot(fid, page, (tb ? tb : tb2));
 227     ios++;
 228     ObtainWriteLock(&tb->lock);
 229     tb->lockers++;
 230     ReleaseWriteLock(&afs_bufferLock);
 231     if (ReallyRead(bufferDir(tb), tb->page, tb->data)) {
 232         tb->lockers--;
 233         FidZap(bufferDir(tb));  /* disaster */
 234         ReleaseWriteLock(&tb->lock);
 235         return EIO;
 236     }
 237     /* Note that findslot sets the page field in the buffer equal to
 238      * what it is searching for.
 239      */
 240     ReleaseWriteLock(&tb->lock);
 241     entry->buffer = tb;
 242     entry->data = tb->data;
 243     return 0;
 244 }
 245
 246
 247 static int
 248 FixupBucket(struct buffer *ap)
 249 {
 250     struct buffer **lp, *tp;
 251     int i;
 252
 253     /* first try to get it out of its current hash bucket, in which it might not be */
 254     i = ap->hashIndex;
 255     lp = &phTable[i];
 256     for (tp = *lp; tp; tp = tp->hashNext) {
 257         if (tp == ap) {
 258             *lp = tp->hashNext;
 259             break;
 260         }
 261         lp = &tp->hashNext;
 262     }
 263     /* now figure the new hash bucket */
 264     i = pHash(ap);
 265     ap->hashIndex = i;          /* remember where we are for deletion */
 266     ap->hashNext = phTable[i];  /* add us to the list */
 267     phTable[i] = ap;            /* at the front, since it's LRU */
 268     return 0;
 269 }
 270
 271 struct buffer *
 272 newslot(dir_file_t dir, afs_int32 apage, struct buffer *lp)
 273 {
 274     /* Find a usable buffer slot */
 275     afs_int32 i;
 276     afs_int32 lt;
 277     struct buffer **tbp;
 278
 279     if (lp && (lp->lockers == 0)) {
 280         lt = lp->accesstime;
 281     } else {
 282         lp = 0;
 283         lt = BUFFER_LONG_MAX;
 284     }
 285
 286     tbp = Buffers;
 287     for (i = 0; i < nbuffers; i++, tbp++) {
 288         if ((*tbp)->lockers == 0) {
 289             if ((*tbp)->accesstime < lt) {
 290                 lp = (*tbp);
 291                 lt = (*tbp)->accesstime;
 292             }
 293         }
 294     }
 295
 296     /* There are no unlocked buffers */
 297     if (lp == 0) {
 298         if (lt < 0)
 299             Die("accesstime counter wrapped");
 300         else
 301             Die("all buffers locked");
 302     }
 303
 304     /* We do not need to lock the buffer here because it has no lockers
 305      * and the afs_bufferLock prevents other threads from zapping this
 306      * buffer while we are writing it out */
 307     if (lp->dirty) {
 308         if (ReallyWrite(bufferDir(lp), lp->page, lp->data))
 309             Die("writing bogus buffer");
 310         lp->dirty = 0;
 311     }
 312
 313     /* Now fill in the header. */
 314     FidZap(bufferDir(lp));
 315     FidCpy(bufferDir(lp), dir); /* set this */
 316     memset(lp->data, 0, BUFFER_PAGE_SIZE);  /* Don't leak stale data. */
 317     lp->page = apage;
 318     lp->accesstime = ++timecounter;
 319
 320     FixupBucket(lp);            /* move to the right hash bucket */
 321
 322     return lp;
 323 }
 324
 325 /* Release a buffer, specifying whether or not the buffer has been modified
 326  * by the locker. */
 327 void
 328 DRelease(struct DirBuffer *entry, int flag)
 329 {
 330     struct buffer *bp;
 331
 332     bp = (struct buffer *) entry->buffer;
 333     if (bp == NULL)
 334         return;
 335     ObtainWriteLock(&bp->lock);
 336     bp->lockers--;
 337     if (flag)
 338         bp->dirty = 1;
 339     ReleaseWriteLock(&bp->lock);
 340 }
 341
 342 /* Return the byte within a file represented by a buffer pointer. */
 343 int
 344 DVOffset(struct DirBuffer *entry)
 345 {
 346     struct buffer *bp;
 347
 348     bp = entry->buffer;
 349     return BUFFER_PAGE_SIZE * bp->page + (char *)entry->data - (char *)bp->data;
 350 }
 351
 352 void
 353 DZap(dir_file_t dir)
 354 {
 355     /* Destroy all buffers pertaining to a particular fid. */
 356     struct buffer *tb;
 357     ObtainReadLock(&afs_bufferLock);
 358     for (tb = phTable[pHash(dir)]; tb; tb = tb->hashNext)
 359         if (FidEq(bufferDir(tb), dir)) {
 360             ObtainWriteLock(&tb->lock);
 361             FidZap(bufferDir(tb));
 362             tb->dirty = 0;
 363             ReleaseWriteLock(&tb->lock);
 364         }
 365     ReleaseReadLock(&afs_bufferLock);
 366 }
 367
 368 int
 369 DFlushVolume(afs_int32 vid)
 370 {
 371     /* Flush all data and release all inode handles for a particular volume */
 372     struct buffer *tb;
 373     int code, rcode = 0;
 374     ObtainReadLock(&afs_bufferLock);
 375     for (tb = phTable[vHash(vid)]; tb; tb = tb->hashNext)
 376         if (FidVolEq(bufferDir(tb), vid)) {
 377             ObtainWriteLock(&tb->lock);
 378             if (tb->dirty) {
 379                 code = ReallyWrite(bufferDir(tb), tb->page, tb->data);
 380                 if (code && !rcode)
 381                     rcode = code;
 382                 tb->dirty = 0;
 383             }
 384             FidZap(bufferDir(tb));
 385             ReleaseWriteLock(&tb->lock);
 386         }
 387     ReleaseReadLock(&afs_bufferLock);
 388     return rcode;
 389 }
 390
 391 int
 392 DFlushEntry(dir_file_t fid)
 393 {
 394     /* Flush pages modified by one entry. */
 395     struct buffer *tb;
 396     int code;
 397
 398     ObtainReadLock(&afs_bufferLock);
 399     for (tb = phTable[pHash(fid)]; tb; tb = tb->hashNext)
 400         if (FidEq(bufferDir(tb), fid) && tb->dirty) {
 401             ObtainWriteLock(&tb->lock);
 402             if (tb->dirty) {
 403                 code = ReallyWrite(bufferDir(tb), tb->page, tb->data);
 404                 if (code) {
 405                     ReleaseWriteLock(&tb->lock);
 406                     ReleaseReadLock(&afs_bufferLock);
 407                     return code;
 408                 }
 409                 tb->dirty = 0;
 410             }
 411             ReleaseWriteLock(&tb->lock);
 412         }
 413     ReleaseReadLock(&afs_bufferLock);
 414     return 0;
 415 }
 416
 417 int
 418 DFlush(void)
 419 {
 420     /* Flush all the modified buffers. */
 421     int i;
 422     struct buffer **tbp;
 423     afs_int32 code, rcode;
 424
 425     rcode = 0;
 426     tbp = Buffers;
 427     ObtainReadLock(&afs_bufferLock);
 428     for (i = 0; i < nbuffers; i++, tbp++) {
 429         if ((*tbp)->dirty) {
 430             ObtainWriteLock(&(*tbp)->lock);
 431             (*tbp)->lockers++;
 432             ReleaseReadLock(&afs_bufferLock);
 433             if ((*tbp)->dirty) {
 434                 code = ReallyWrite(bufferDir(*tbp), (*tbp)->page, (*tbp)->data);
 435                 if (!code)
 436                     (*tbp)->dirty = 0;  /* Clear the dirty flag */
 437                 if (code && !rcode) {
 438                     rcode = code;
 439                 }
 440             }
 441             (*tbp)->lockers--;
 442             ReleaseWriteLock(&(*tbp)->lock);
 443             ObtainReadLock(&afs_bufferLock);
 444         }
 445     }
 446     ReleaseReadLock(&afs_bufferLock);
 447     return rcode;
 448 }
 449
 450 /* Same as read, only do *not* even try to read the page,
 451  * since it probably doesn't exist.
 452  */
 453 int
 454 DNew(dir_file_t dir, int page, struct DirBuffer *entry)
 455 {
 456     struct buffer *tb;
 457
 458     memset(entry,0, sizeof(struct DirBuffer));
 459
 460     ObtainWriteLock(&afs_bufferLock);
 461     if ((tb = newslot(dir, page, 0)) == 0) {
 462         ReleaseWriteLock(&afs_bufferLock);
 463         return EIO;
 464     }
 465     ObtainWriteLock(&tb->lock);
 466     tb->lockers++;
 467     ReleaseWriteLock(&afs_bufferLock);
 468     ReleaseWriteLock(&tb->lock);
 469
 470     entry->buffer = tb;
 471     entry->data = tb->data;
 472
 473     return 0;
 474 }