2 * Sherlock Library -- Object Buckets
4 * (c) 2001--2004 Martin Mares <mj@ucw.cz>
5 * (c) 2004 Robert Spalek <robert@ucw.cz>
7 * This software may be freely distributed and used according to the terms
8 * of the GNU Lesser General Public License.
14 #include "lib/bucket.h"
15 #include "lib/fastbuf.h"
27 static struct obuck_header obuck_hdr, obuck_create_hdr;
28 static sh_off_t bucket_find_pos;
29 static struct fastbuf *obuck_write_fb;
31 /*** Configuration ***/
33 byte *obuck_name = "not/configured";
34 static uns obuck_io_buflen = 65536;
35 static int obuck_shake_buflen = 1048576;
36 static uns obuck_shake_security;
37 static uns obuck_slurp_buflen = 65536;
39 static struct cfitem obuck_config[] = {
40 { "Buckets", CT_SECTION, NULL },
41 { "BucketFile", CT_STRING, &obuck_name },
42 { "BufSize", CT_INT, &obuck_io_buflen },
43 { "ShakeBufSize", CT_INT, &obuck_shake_buflen },
44 { "ShakeSecurity", CT_INT, &obuck_shake_security },
45 { "SlurpBufSize", CT_INT, &obuck_slurp_buflen },
46 { NULL, CT_STOP, NULL }
49 static void CONSTRUCTOR obuck_init_config(void)
51 cf_register(obuck_config);
54 /*** Internal operations ***/
57 obuck_broken(char *msg, sh_off_t pos)
59 die("Object pool corrupted: %s (pos=%Lx)", msg, (long long) pos);
63 * We need several types of locks:
65 * Read lock reading parts of bucket file
66 * Write lock any write operations
67 * Append lock appending to the end of the file
68 * Scan lock reading parts which we are certain they exist
70 * Multiple read and scan locks can co-exist together.
71 * Scan locks can co-exist with an append lock.
72 * There can be at most one write/append lock at a time.
74 * These lock types map to a pair of normal read-write locks which
75 * we represent as fcntl() locks on the first and second byte of the
76 * bucket file. [We cannot use flock() since it happily permits
77 * locking a shared fd (e.g., after fork()) multiple times at it also
78 * doesn't offer multiple locks on a single file.]
82 * Write <write> <write>
88 obuck_do_lock(int type, int start, int len)
93 fl.l_whence = SEEK_SET;
96 if (fcntl(obuck_fd, F_SETLKW, &fl) < 0)
97 die("fcntl lock: %m");
101 obuck_lock_read(void)
103 obuck_do_lock(F_RDLCK, 0, 2);
107 obuck_lock_write(void)
109 obuck_do_lock(F_WRLCK, 0, 2);
113 obuck_lock_append(void)
115 obuck_do_lock(F_WRLCK, 0, 1);
119 obuck_lock_read_to_scan(void)
121 obuck_do_lock(F_UNLCK, 0, 1);
127 obuck_do_lock(F_UNLCK, 0, 2);
130 /*** FastIO emulation ***/
138 #define FB_BUCKET(f) ((struct fb_bucket *)(f)->is_fastbuf)
140 static int obuck_fb_count;
143 obuck_fb_close(struct fastbuf *f)
149 /* We need to use pread/pwrite since we work on fd's shared between processes */
152 obuck_fb_refill(struct fastbuf *f)
154 uns remains, bufsize, size, datasize;
156 remains = FB_BUCKET(f)->bucket_size - (uns)f->pos;
157 bufsize = f->bufend - f->buffer;
160 sh_off_t start = FB_BUCKET(f)->start_pos;
161 sh_off_t pos = start + sizeof(struct obuck_header) + f->pos;
162 if (remains <= bufsize)
165 size = start + obuck_bucket_size(FB_BUCKET(f)->bucket_size) - pos;
168 size = datasize = bufsize;
169 int l = sh_pread(obuck_fd, f->buffer, size, pos);
171 die("Error reading bucket: %m");
172 if ((unsigned) l != size)
173 obuck_broken("Short read", FB_BUCKET(f)->start_pos);
175 f->bstop = f->buffer + datasize;
179 if (GET_U32(f->buffer + size - 4) != OBUCK_TRAILER)
180 obuck_broken("Missing trailer", FB_BUCKET(f)->start_pos);
186 obuck_fb_spout(struct fastbuf *f)
188 int l = f->bptr - f->buffer;
193 int z = sh_pwrite(obuck_fd, c, l, FB_BUCKET(f)->start_pos + sizeof(struct obuck_header) + f->pos);
195 die("Error writing bucket: %m");
203 /*** Exported functions ***/
206 obuck_init(int writeable)
210 obuck_fd = sh_open(obuck_name, (writeable ? O_RDWR | O_CREAT : O_RDONLY), 0666);
212 die("Unable to open bucket file %s: %m", obuck_name);
214 size = sh_seek(obuck_fd, 0, SEEK_END);
217 /* If the bucket pool is not empty, check consistency of its end */
219 if (sh_pread(obuck_fd, &check, 4, size-4) != 4 ||
220 check != OBUCK_TRAILER)
221 obuck_broken("Missing trailer of last object", size - 4);
231 log(L_ERROR, "Bug: Unbalanced bucket opens/closes: %d streams remain", obuck_fb_count);
233 log(L_ERROR, "Bug: Forgot to close bucket write stream");
240 bflush(obuck_write_fb);
247 bucket_find_pos = obuck_get_pos(oid);
248 if (sh_pread(obuck_fd, &obuck_hdr, sizeof(obuck_hdr), bucket_find_pos) != sizeof(obuck_hdr))
249 obuck_broken("Short header read", bucket_find_pos);
250 if (obuck_hdr.magic != OBUCK_MAGIC)
251 obuck_broken("Missing magic number", bucket_find_pos);
252 if (obuck_hdr.oid == OBUCK_OID_DELETED)
253 obuck_broken("Access to deleted bucket", bucket_find_pos);
254 if (obuck_hdr.oid != oid)
255 obuck_broken("Invalid backlink", bucket_find_pos);
259 obuck_find_by_oid(struct obuck_header *hdrp)
261 oid_t oid = hdrp->oid;
263 ASSERT(oid < OBUCK_OID_FIRST_SPECIAL);
267 memcpy(hdrp, &obuck_hdr, sizeof(obuck_hdr));
271 obuck_find_first(struct obuck_header *hdrp, int full)
275 return obuck_find_next(hdrp, full);
279 obuck_find_next(struct obuck_header *hdrp, int full)
286 bucket_find_pos += obuck_bucket_size(obuck_hdr.length);
288 c = sh_pread(obuck_fd, &obuck_hdr, sizeof(obuck_hdr), bucket_find_pos);
292 if (c != sizeof(obuck_hdr))
293 obuck_broken("Short header read", bucket_find_pos);
294 if (obuck_hdr.magic != OBUCK_MAGIC)
295 obuck_broken("Missing magic number", bucket_find_pos);
296 if (obuck_hdr.oid != OBUCK_OID_DELETED || full)
298 memcpy(hdrp, &obuck_hdr, sizeof(obuck_hdr));
308 uns official_buflen = ALIGN(MIN(obuck_hdr.length, obuck_io_buflen), OBUCK_ALIGN);
309 uns real_buflen = official_buflen + OBUCK_ALIGN;
311 b = xmalloc(sizeof(struct fb_bucket) + real_buflen);
312 b->buffer = b->bptr = b->bstop = FB_BUCKET(b)->buffer;
313 b->bufend = b->buffer + official_buflen;
314 b->name = "bucket-read";
316 b->refill = obuck_fb_refill;
319 b->close = obuck_fb_close;
321 b->can_overwrite_buffer = 2;
322 FB_BUCKET(b)->start_pos = bucket_find_pos;
323 FB_BUCKET(b)->bucket_size = obuck_hdr.length;
329 obuck_predict_last_oid(void)
331 sh_off_t size = sh_seek(obuck_fd, 0, SEEK_END);
332 return (oid_t)(size >> OBUCK_SHIFT);
336 obuck_create(u32 type)
338 ASSERT(!obuck_write_fb);
341 sh_off_t start = sh_seek(obuck_fd, 0, SEEK_END);
342 if (start & (OBUCK_ALIGN - 1))
343 obuck_broken("Misaligned file", start);
344 obuck_create_hdr.magic = OBUCK_INCOMPLETE_MAGIC;
345 obuck_create_hdr.oid = start >> OBUCK_SHIFT;
346 obuck_create_hdr.length = 0;
347 obuck_create_hdr.type = type;
349 struct fastbuf *b = xmalloc(sizeof(struct fb_bucket) + obuck_io_buflen);
351 b->buffer = FB_BUCKET(b)->buffer;
352 b->bptr = b->bstop = b->buffer;
353 b->bufend = b->buffer + obuck_io_buflen;
354 b->pos = -(int)sizeof(obuck_create_hdr);
355 b->name = "bucket-write";
357 b->spout = obuck_fb_spout;
361 b->can_overwrite_buffer = 0;
362 FB_BUCKET(b)->start_pos = start;
363 FB_BUCKET(b)->bucket_size = 0;
364 bwrite(b, &obuck_create_hdr, sizeof(obuck_create_hdr));
370 obuck_create_end(struct fastbuf *b, struct obuck_header *hdrp)
372 ASSERT(b == obuck_write_fb);
373 obuck_write_fb = NULL;
375 obuck_create_hdr.magic = OBUCK_MAGIC;
376 obuck_create_hdr.length = btell(b);
377 int pad = (OBUCK_ALIGN - sizeof(obuck_create_hdr) - obuck_create_hdr.length - 4) & (OBUCK_ALIGN - 1);
380 bputl(b, OBUCK_TRAILER);
382 ASSERT(!((FB_BUCKET(b)->start_pos + sizeof(obuck_create_hdr) + b->pos) & (OBUCK_ALIGN - 1)));
383 if (sh_pwrite(obuck_fd, &obuck_create_hdr, sizeof(obuck_create_hdr), FB_BUCKET(b)->start_pos) != sizeof(obuck_create_hdr))
384 die("Bucket header update failed: %m");
386 memcpy(hdrp, &obuck_create_hdr, sizeof(obuck_create_hdr));
391 obuck_delete(oid_t oid)
395 obuck_hdr.oid = OBUCK_OID_DELETED;
396 sh_pwrite(obuck_fd, &obuck_hdr, sizeof(obuck_hdr), bucket_find_pos);
400 /*** Fast reading of the whole pool ***/
402 static struct fastbuf *obuck_rpf;
403 static uns slurp_remains;
404 static sh_off_t slurp_start, slurp_current, slurp_end;
407 obuck_slurp_refill(struct fastbuf *f)
413 l = bdirect_read_prepare(obuck_rpf, &f->buffer);
415 obuck_broken("Incomplete object", slurp_start);
416 l = MIN(l, slurp_remains);
417 bdirect_read_commit(obuck_rpf, f->buffer + l);
420 f->bufend = f->bstop = f->buffer + l;
425 obuck_slurp_pool(struct obuck_header *hdrp)
427 static struct fastbuf limiter;
435 obuck_rpf = bopen(obuck_name, O_RDONLY, obuck_slurp_buflen);
436 bseek(obuck_rpf, 0, SEEK_END);
437 slurp_end = btell(obuck_rpf);
438 bsetpos(obuck_rpf, 0);
439 obuck_lock_read_to_scan();
443 bsetpos(obuck_rpf, slurp_current - 4);
444 if (bgetl(obuck_rpf) != OBUCK_TRAILER)
445 obuck_broken("Missing trailer", slurp_start);
447 slurp_start = btell(obuck_rpf);
448 if (slurp_start < slurp_end)
449 l = bread(obuck_rpf, hdrp, sizeof(struct obuck_header));
459 if (l != sizeof(struct obuck_header))
460 obuck_broken("Short header read", slurp_start);
461 if (hdrp->magic != OBUCK_MAGIC)
462 obuck_broken("Missing magic number", slurp_start);
463 slurp_current = slurp_start + obuck_bucket_size(hdrp->length);
465 while (hdrp->oid == OBUCK_OID_DELETED);
466 if (obuck_get_pos(hdrp->oid) != slurp_start)
467 obuck_broken("Invalid backlink", slurp_start);
468 slurp_remains = hdrp->length;
469 limiter.bptr = limiter.bstop = limiter.buffer = limiter.bufend = NULL;
470 limiter.name = "Bucket";
472 limiter.refill = obuck_slurp_refill;
479 shake_write(void *addr, int len, sh_off_t pos)
481 int l = sh_pwrite(obuck_fd, addr, len, pos);
485 die("obuck_shakedown write error: %m");
487 die("obuck_shakedown write error: disk full");
494 if (obuck_shake_security > 1)
499 shake_write_backup(sh_off_t bpos, byte *norm_buf, int norm_size, byte *fragment, int frag_size, sh_off_t frag_pos, int more_size)
501 struct obuck_header *bhdr;
506 /* First of all, the "normal" part -- everything that will be written in this pass */
507 DBG("Backing up first round of changes at position %Lx + %x", (long long) bpos, norm_size);
508 while (boff < norm_size)
510 /* This needn't be optimized for speed. */
511 bhdr = (struct obuck_header *) (norm_buf + boff);
512 ASSERT(bhdr->magic == OBUCK_MAGIC);
513 l = obuck_bucket_size(bhdr->length);
515 bhdr->oid = bpos >> OBUCK_SHIFT;
516 shake_write(bhdr, l, bpos);
522 /* If we have an incomplete bucket at the end of the buffer, we must copy it as well. */
525 DBG("Backing up fragment of size %x and %x more", frag_size, more_size);
527 /* First the part we already have in the buffer */
528 bhdr = (struct obuck_header *) fragment;
529 ASSERT(bhdr->magic == OBUCK_MAGIC);
531 bhdr->oid = bpos >> OBUCK_SHIFT;
532 shake_write(bhdr, frag_size, bpos);
536 /* And then the rest, using a small 64K buffer */
537 byte *auxbuf = alloca(65536);
539 while (l < more_size)
541 int j = MIN(more_size-l, 65536);
542 if (sh_pread(obuck_fd, auxbuf, j, frag_pos + frag_size + l) != j)
543 die("obuck_shakedown read error: %m");
544 shake_write(auxbuf, j, bpos);
552 shake_erase(sh_off_t start, sh_off_t end)
555 die("shake_erase called with negative length, that's a bug");
556 ASSERT(!(start & (OBUCK_ALIGN-1)) && !(end & (OBUCK_ALIGN-1)));
559 u32 check = OBUCK_TRAILER;
560 obuck_hdr.magic = OBUCK_MAGIC;
561 obuck_hdr.oid = OBUCK_OID_DELETED;
562 uns len = MIN(0x40000000, end-start);
563 obuck_hdr.length = len - sizeof(obuck_hdr) - 4;
564 DBG("Erasing %08x bytes at %Lx", len, (long long) start);
565 shake_write(&obuck_hdr, sizeof(obuck_hdr), start);
567 shake_write(&check, 4, start-4);
572 obuck_shakedown(int (*kibitz)(struct obuck_header *old, oid_t new, byte *buck))
574 byte *buf; /* Shakedown buffer and its size */
575 int buflen = ALIGN(obuck_shake_buflen, OBUCK_ALIGN);
576 byte *msg; /* Error message we will print */
577 sh_off_t rstart, wstart; /* Original and new position of buffer start */
578 sh_off_t r_bucket_start, w_bucket_start; /* Original and new position of the current bucket */
579 int roff, woff; /* Orig/new position of the current bucket relative to buffer start */
580 int rsize; /* Number of original bytes in the buffer */
581 int l; /* Raw size of the current bucket */
582 int changed = 0; /* "Something has been altered" flag */
583 int wrote_anything = 0; /* We already did a write to the bucket file */
584 struct obuck_header *rhdr, *whdr; /* Original and new address of header of the current bucket */
585 sh_off_t r_file_size; /* Original size of the bucket file */
586 int more; /* How much does the last bucket overlap the buffer */
588 buf = xmalloc(buflen);
590 roff = woff = rsize = 0;
592 /* We need to be the only accessor, all the object ID's are becoming invalid */
594 r_file_size = sh_seek(obuck_fd, 0, SEEK_END);
595 ASSERT(!(r_file_size & (OBUCK_ALIGN - 1)));
596 if (r_file_size >= (0x100000000 << OBUCK_SHIFT) - buflen)
597 die("Bucket file is too large for safe shakedown. Shaking down with Bucket.ShakeSecurity=0 will still work.");
599 DBG("Starting shakedown. Buffer size is %d, original length %Lx", buflen, (long long) r_file_size);
603 r_bucket_start = rstart + roff;
604 w_bucket_start = wstart + woff;
605 rhdr = (struct obuck_header *)(buf + roff);
606 whdr = (struct obuck_header *)(buf + woff);
612 if (rhdr->magic != OBUCK_MAGIC ||
613 rhdr->oid != OBUCK_OID_DELETED && rhdr->oid != (oid_t)(r_bucket_start >> OBUCK_SHIFT))
615 msg = "header mismatch";
618 l = obuck_bucket_size(rhdr->length);
621 if (rhdr->oid != OBUCK_OID_DELETED)
623 msg = "bucket longer than ShakeBufSize";
626 /* Empty buckets are allowed to be large, but we need to handle them extra */
627 DBG("Tricking around an extra-large empty bucket at %Lx + %x", (long long)r_bucket_start, l);
632 if (rsize - roff < l)
634 more = l - (rsize - roff);
637 if (GET_U32((byte *)rhdr + l - 4) != OBUCK_TRAILER)
639 msg = "missing trailer";
643 if (rhdr->oid != OBUCK_OID_DELETED)
645 int status = kibitz(rhdr, w_bucket_start >> OBUCK_SHIFT, (byte *)(rhdr+1));
651 /* Changed! Reconstruct the trailer. */
652 lnew = obuck_bucket_size(rhdr->length);
654 PUT_U32((byte *)rhdr + lnew - 4, OBUCK_TRAILER);
657 whdr = (struct obuck_header *)(buf+woff);
659 memmove(whdr, rhdr, lnew);
660 whdr->oid = w_bucket_start >> OBUCK_SHIFT;
668 kibitz(rhdr, OBUCK_OID_DELETED, NULL);
677 /* Write the new contents of the bucket file */
680 if (obuck_shake_security)
682 /* But first write a backup at the end of the file to ensure nothing can be lost. */
683 shake_write_backup(r_file_size, buf, woff, buf+roff, rsize-roff, rstart+roff, more);
690 DBG("Write %Lx %x", wstart, woff);
691 shake_write(buf, woff, wstart);
696 ASSERT(wstart == rstart);
698 /* In any case, update the write position */
702 /* Skip what's been read and if there is any fragment at the end of the buffer, move it to the start */
706 memmove(buf, buf+roff, rsize-roff);
712 /* And refill the buffer */
713 r_bucket_start = rstart+rsize; /* Also needed for error messages */
714 l = sh_pread(obuck_fd, buf+rsize, MIN(buflen-rsize, r_file_size - r_bucket_start), r_bucket_start);
715 DBG("Read %Lx %x (%x inherited)", (long long)r_bucket_start, l, rsize);
717 die("obuck_shakedown read error: %m");
722 msg = "unexpected EOF";
725 if (l & (OBUCK_ALIGN-1))
727 msg = "garbage at the end of file";
734 DBG("Finished at position %Lx", (long long) wstart);
735 sh_ftruncate(obuck_fd, wstart);
743 log(L_ERROR, "Error during object pool shakedown: %s (pos=%Ld, id=%x), gathering debris",
744 msg, (long long) r_bucket_start, (uns)(r_bucket_start >> OBUCK_SHIFT));
746 * We can attempt to clean up the bucket file by erasing everything between the last
747 * byte written and the next byte to be read. If the secure mode is switched on, we can
748 * guarantee that no data are lost, only some might be duplicated.
750 shake_erase(wstart, rstart);
751 die("Fatal error during object pool shakedown");
761 #define LEN(i) ((259309*(i))%MAXLEN)
763 static int test_kibitz(struct obuck_header *h, oid_t new, byte *buck)
768 int main(int argc, char **argv)
771 unsigned int i, j, cnt;
772 struct obuck_header h;
776 if (cf_getopt(argc, argv, CF_SHORT_OPTS, CF_NO_LONG_OPTS, NULL) >= 0 ||
779 fputs("This program supports only the following command-line arguments:\n" CF_USAGE, stderr);
785 for(j=0; j<COUNT; j++)
787 b = obuck_create(BUCKET_TYPE_PLAIN);
788 for(i=0; i<LEN(j); i++)
789 bputc(b, (i+j) % 256);
790 obuck_create_end(b, &h);
791 printf("Writing %08x %d\n", h.oid, h.length);
794 for(j=0; j<COUNT; j++)
795 if (j % 100 < KILLPERC)
797 printf("Deleting %08x\n", ids[j]);
798 obuck_delete(ids[j]);
801 for(j=0; j<COUNT; j++)
802 if (j % 100 >= KILLPERC)
806 obuck_find_by_oid(&h);
808 printf("Reading %08x %d\n", h.oid, h.length);
809 if (h.length != LEN(j))
810 die("Invalid length");
811 for(i=0; i<h.length; i++)
812 if ((unsigned) bgetc(b) != (i+j) % 256)
813 die("Contents mismatch");
818 obuck_shakedown(test_kibitz);
819 if (obuck_find_first(&h, 0))
822 printf("<<< %08x\t%d\n", h.oid, h.length);
825 while (obuck_find_next(&h, 0));
827 die("Walk mismatch");