+/*** Fast reading of the whole pool ***/
+
+static struct fastbuf *obuck_rpf;
+static uns slurp_remains;
+static sh_off_t slurp_start, slurp_current, slurp_end;
+
+static int
+obuck_slurp_refill(struct fastbuf *f)
+{
+ if (!slurp_remains)
+ return 0;
+ uns l = bdirect_read_prepare(obuck_rpf, &f->buffer);
+ if (!l)
+ obuck_broken("Incomplete object", slurp_start);
+ l = MIN(l, slurp_remains);
+ /* XXX: This probably should be bdirect_read_commit_modified() in some cases,
+ * but it doesn't hurt since we aren't going to seek.
+ */
+ bdirect_read_commit(obuck_rpf, f->buffer + l);
+ slurp_remains -= l;
+ f->bptr = f->buffer;
+ f->bufend = f->bstop = f->buffer + l;
+ return 1;
+}
+
+struct fastbuf *
+obuck_slurp_pool(struct obuck_header *hdrp)
+{
+ static struct fastbuf limiter;
+ uns l;
+
+ do
+ {
+ if (!obuck_rpf)
+ {
+ obuck_lock_read();
+ obuck_rpf = bopen(obuck_name, O_RDONLY, obuck_slurp_buflen);
+ slurp_end = bfilesize(obuck_rpf);
+ obuck_lock_read_to_scan();
+ }
+ else
+ {
+ bsetpos(obuck_rpf, slurp_current - 4);
+ if (bgetl(obuck_rpf) != OBUCK_TRAILER)
+ obuck_broken("Missing trailer", slurp_start);
+ }
+ slurp_start = btell(obuck_rpf);
+ if (slurp_start < slurp_end)
+ l = bread(obuck_rpf, hdrp, sizeof(struct obuck_header));
+ else
+ l = 0;
+ if (!l)
+ {
+ bclose(obuck_rpf);
+ obuck_rpf = NULL;
+ obuck_unlock();
+ return NULL;
+ }
+ if (l != sizeof(struct obuck_header))
+ obuck_broken("Short header read", slurp_start);
+ if (hdrp->magic != OBUCK_MAGIC)
+ obuck_broken("Missing magic number", slurp_start);
+ slurp_current = slurp_start + obuck_bucket_size(hdrp->length);
+ }
+ while (hdrp->oid == OBUCK_OID_DELETED);
+ if (obuck_get_pos(hdrp->oid) != slurp_start)
+ obuck_broken("Invalid backlink", slurp_start);
+ slurp_remains = hdrp->length;
+ limiter.bptr = limiter.bstop = limiter.buffer = limiter.bufend = NULL;
+ limiter.name = "Bucket";
+ limiter.pos = 0;
+ limiter.refill = obuck_slurp_refill;
+ limiter.can_overwrite_buffer = obuck_rpf->can_overwrite_buffer;
+ return &limiter;
+}
+
+/*** Shakedown ***/
+
+static inline void
+shake_write(void *addr, int len, sh_off_t pos)
+{
+ int l = sh_pwrite(obuck_fd, addr, len, pos);
+ if (l != len)
+ {
+ if (l < 0)
+ die("obuck_shakedown write error: %m");
+ else
+ die("obuck_shakedown write error: disk full");
+ }
+}
+
+static inline void
+shake_sync(void)
+{
+ if (obuck_shake_security > 1)
+ fdatasync(obuck_fd);
+}
+
+static void
+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)
+{
+ struct obuck_header *bhdr;
+ int boff = 0;
+ int l;
+ oid_t old_oid;
+
+ /* First of all, the "normal" part -- everything that will be written in this pass */
+ DBG("Backing up first round of changes at position %Lx + %x", (long long) bpos, norm_size);
+ while (boff < norm_size)
+ {
+ /* This needn't be optimized for speed. */
+ bhdr = (struct obuck_header *) (norm_buf + boff);
+ ASSERT(bhdr->magic == OBUCK_MAGIC);
+ l = obuck_bucket_size(bhdr->length);
+ old_oid = bhdr->oid;
+ bhdr->oid = bpos >> OBUCK_SHIFT;
+ shake_write(bhdr, l, bpos);
+ bhdr->oid = old_oid;
+ boff += l;
+ bpos += l;
+ }
+
+ /* If we have an incomplete bucket at the end of the buffer, we must copy it as well. */
+ if (more_size)
+ {
+ DBG("Backing up fragment of size %x and %x more", frag_size, more_size);
+
+ /* First the part we already have in the buffer */
+ bhdr = (struct obuck_header *) fragment;
+ ASSERT(bhdr->magic == OBUCK_MAGIC);
+ old_oid = bhdr->oid;
+ bhdr->oid = bpos >> OBUCK_SHIFT;
+ shake_write(bhdr, frag_size, bpos);
+ bhdr->oid = old_oid;
+ bpos += frag_size;
+
+ /* And then the rest, using a small 64K buffer */
+ byte *auxbuf = alloca(65536);
+ l = 0;
+ while (l < more_size)
+ {
+ int j = MIN(more_size-l, 65536);
+ if (sh_pread(obuck_fd, auxbuf, j, frag_pos + frag_size + l) != j)
+ die("obuck_shakedown read error: %m");
+ shake_write(auxbuf, j, bpos);
+ bpos += j;
+ l += j;
+ }
+ }
+}
+
+static void
+shake_erase(sh_off_t start, sh_off_t end)
+{
+ if (start > end)
+ die("shake_erase called with negative length, that's a bug");
+ ASSERT(!(start & (OBUCK_ALIGN-1)) && !(end & (OBUCK_ALIGN-1)));
+ while (start < end)
+ {
+ u32 check = OBUCK_TRAILER;
+ obuck_hdr.magic = OBUCK_MAGIC;
+ obuck_hdr.oid = OBUCK_OID_DELETED;
+ uns len = MIN(0x40000000, end-start);
+ obuck_hdr.length = len - sizeof(obuck_hdr) - 4;
+ DBG("Erasing %08x bytes at %Lx", len, (long long) start);
+ shake_write(&obuck_hdr, sizeof(obuck_hdr), start);
+ start += len;
+ shake_write(&check, 4, start-4);
+ }
+}
+
+void
+obuck_shakedown(int (*kibitz)(struct obuck_header *old, oid_t new, byte *buck))
+{
+ byte *buf; /* Shakedown buffer and its size */
+ int buflen = ALIGN(obuck_shake_buflen, OBUCK_ALIGN);
+ byte *msg; /* Error message we will print */
+ sh_off_t rstart, wstart; /* Original and new position of buffer start */
+ sh_off_t r_bucket_start, w_bucket_start; /* Original and new position of the current bucket */
+ int roff, woff; /* Orig/new position of the current bucket relative to buffer start */
+ int rsize; /* Number of original bytes in the buffer */
+ int l; /* Raw size of the current bucket */
+ int changed = 0; /* "Something has been altered" flag */
+ int wrote_anything = 0; /* We already did a write to the bucket file */
+ struct obuck_header *rhdr, *whdr; /* Original and new address of header of the current bucket */
+ sh_off_t r_file_size; /* Original size of the bucket file */
+ int more; /* How much does the last bucket overlap the buffer */
+
+ buf = xmalloc(buflen);
+ rstart = wstart = 0;
+ roff = woff = rsize = 0;
+
+ /* We need to be the only accessor, all the object ID's are becoming invalid */
+ obuck_lock_write();
+ r_file_size = sh_seek(obuck_fd, 0, SEEK_END);
+ ASSERT(!(r_file_size & (OBUCK_ALIGN - 1)));
+ if (r_file_size >= (0x100000000 << OBUCK_SHIFT) - buflen)
+ die("Bucket file is too large for safe shakedown. Shaking down with Bucket.ShakeSecurity=0 will still work.");
+
+ DBG("Starting shakedown. Buffer size is %d, original length %Lx", buflen, (long long) r_file_size);
+
+ for(;;)
+ {
+ r_bucket_start = rstart + roff;
+ w_bucket_start = wstart + woff;
+ rhdr = (struct obuck_header *)(buf + roff);
+ whdr = (struct obuck_header *)(buf + woff);
+ if (roff == rsize)
+ {
+ more = 0;
+ goto next;
+ }
+ if (rhdr->magic != OBUCK_MAGIC ||
+ rhdr->oid != OBUCK_OID_DELETED && rhdr->oid != (oid_t)(r_bucket_start >> OBUCK_SHIFT))
+ {
+ msg = "header mismatch";
+ goto broken;
+ }
+ l = obuck_bucket_size(rhdr->length);
+ if (l > buflen)
+ {
+ if (rhdr->oid != OBUCK_OID_DELETED)
+ {
+ msg = "bucket longer than ShakeBufSize";
+ goto broken;
+ }
+ /* Empty buckets are allowed to be large, but we need to handle them extra */
+ DBG("Tricking around an extra-large empty bucket at %Lx + %x", (long long)r_bucket_start, l);
+ rsize = roff + l;
+ }
+ else
+ {
+ if (rsize - roff < l)
+ {
+ more = l - (rsize - roff);
+ goto next;
+ }
+ if (GET_U32((byte *)rhdr + l - 4) != OBUCK_TRAILER)
+ {
+ msg = "missing trailer";
+ goto broken;
+ }
+ }
+ if (rhdr->oid != OBUCK_OID_DELETED)
+ {
+ int status = kibitz(rhdr, w_bucket_start >> OBUCK_SHIFT, (byte *)(rhdr+1));
+ if (status)
+ {
+ int lnew = l;
+ if (status > 1)
+ {
+ /* Changed! Reconstruct the trailer. */
+ lnew = obuck_bucket_size(rhdr->length);
+ ASSERT(lnew <= l);
+ PUT_U32((byte *)rhdr + lnew - 4, OBUCK_TRAILER);
+ changed = 1;
+ }
+ whdr = (struct obuck_header *)(buf+woff);
+ if (rhdr != whdr)
+ memmove(whdr, rhdr, lnew);
+ whdr->oid = w_bucket_start >> OBUCK_SHIFT;
+ woff += lnew;
+ }
+ else
+ changed = 1;
+ }
+ else
+ {
+ kibitz(rhdr, OBUCK_OID_DELETED, NULL);
+ changed = 1;
+ }
+ roff += l;
+ continue;
+
+ next:
+ if (changed)
+ {
+ /* Write the new contents of the bucket file */
+ if (!wrote_anything)
+ {
+ if (obuck_shake_security)
+ {
+ /* But first write a backup at the end of the file to ensure nothing can be lost. */
+ shake_write_backup(r_file_size, buf, woff, buf+roff, rsize-roff, rstart+roff, more);
+ shake_sync();
+ }
+ wrote_anything = 1;
+ }
+ if (woff)
+ {
+ DBG("Write %Lx %x", wstart, woff);
+ shake_write(buf, woff, wstart);
+ shake_sync();
+ }
+ }
+ else
+ ASSERT(wstart == rstart);
+
+ /* In any case, update the write position */
+ wstart += woff;
+ woff = 0;
+
+ /* Skip what's been read and if there is any fragment at the end of the buffer, move it to the start */
+ rstart += roff;
+ if (more)
+ {
+ memmove(buf, buf+roff, rsize-roff);
+ rsize = rsize-roff;
+ }
+ else
+ rsize = 0;
+
+ /* And refill the buffer */
+ r_bucket_start = rstart+rsize; /* Also needed for error messages */
+ l = sh_pread(obuck_fd, buf+rsize, MIN(buflen-rsize, r_file_size - r_bucket_start), r_bucket_start);
+ DBG("Read %Lx %x (%x inherited)", (long long)r_bucket_start, l, rsize);
+ if (l < 0)
+ die("obuck_shakedown read error: %m");
+ if (!l)
+ {
+ if (!more)
+ break;
+ msg = "unexpected EOF";
+ goto broken;
+ }
+ if (l & (OBUCK_ALIGN-1))
+ {
+ msg = "garbage at the end of file";
+ goto broken;
+ }
+ rsize += l;
+ roff = 0;
+ }
+
+ DBG("Finished at position %Lx", (long long) wstart);
+ sh_ftruncate(obuck_fd, wstart);
+ shake_sync();
+
+ obuck_unlock();
+ xfree(buf);
+ return;
+
+ broken:
+ log(L_ERROR, "Error during object pool shakedown: %s (pos=%Ld, id=%x), gathering debris",
+ msg, (long long) r_bucket_start, (uns)(r_bucket_start >> OBUCK_SHIFT));
+ /*
+ * We can attempt to clean up the bucket file by erasing everything between the last
+ * byte written and the next byte to be read. If the secure mode is switched on, we can
+ * guarantee that no data are lost, only some might be duplicated.
+ */
+ shake_erase(wstart, rstart);
+ die("Fatal error during object pool shakedown");
+}
+