log_get.c

这是国外的resip协议栈

		}
		nlsn.file = lp->lsn.file;
		nlsn.offset = lp->lsn.offset - lp->len;
		break;
	case DB_SET:				/* Set log record. */
		nlsn = *alsn;
		break;
	default:
		DB_ASSERT(0);
		ret = EINVAL;
		goto err;
	}

	if (0) {				/* Move to the next file. */
next_file:	++nlsn.file;
		nlsn.offset = 0;
	}

	/*
	 * The above switch statement should have set nlsn to the lsn of
	 * the requested record.
	 */

	if (CRYPTO_ON(dbenv)) {
		hdr.size = HDR_CRYPTO_SZ;
		is_hmac = 1;
	} else {
		hdr.size = HDR_NORMAL_SZ;
		is_hmac = 0;
	}
	/* Check to see if the record is in the cursor's buffer. */
	if ((ret = __log_c_incursor(logc, &nlsn, &hdr, &rp)) != 0)
		goto err;
	if (rp != NULL)
		goto cksum;

	/*
	 * Look to see if we're moving backward in the log with the last record
	 * coming from the disk -- it means the record can't be in the region's
	 * buffer.  Else, check the region's buffer.
	 *
	 * If the record isn't in the region's buffer, then either logs are
	 * in-memory, and we're done, or  we're going to have to read the
	 * record from disk.  We want to make a point of not reading past the
	 * end of the logical log (after recovery, there may be data after the
	 * end of the logical log, not to mention the log file may have been
	 * pre-allocated).  So, zero out last_lsn, and initialize it inside
	 * __log_c_inregion -- if it's still zero when we check it in
	 * __log_c_ondisk, that's OK, it just means the logical end of the log
	 * isn't an issue for this request.
	 */
	ZERO_LSN(last_lsn);
	if (!F_ISSET(logc, DB_LOG_DISK) ||
	    log_compare(&nlsn, &logc->c_lsn) > 0) {
		F_CLR(logc, DB_LOG_DISK);

		if ((ret = __log_c_inregion(logc,
		    &nlsn, &rlock, &last_lsn, &hdr, &rp)) != 0)
			goto err;
		if (rp != NULL)
			goto cksum;
		if (lp->db_log_inmemory)
			goto nohdr;
	}

	/*
	 * We have to read from an on-disk file to retrieve the record.
	 * If we ever can't retrieve the record at offset 0, we're done,
	 * return EOF/DB_NOTFOUND.
	 *
	 * Discard the region lock if we're still holding it, the on-disk
	 * reading routines don't need it.
	 */
	if (rlock == L_ACQUIRED) {
		rlock = L_NONE;
		R_UNLOCK(dbenv, &dblp->reginfo);
	}
	if ((ret = __log_c_ondisk(
	    logc, &nlsn, &last_lsn, flags, &hdr, &rp, &eof)) != 0)
		goto err;
	if (eof) {
		/*
		 * Only DB_NEXT automatically moves to the next file, and
		 * it only happens once.
		 */
		if (flags != DB_NEXT || nlsn.offset == 0)
			return (DB_NOTFOUND);
		goto next_file;
	}
	F_SET(logc, DB_LOG_DISK);

cksum:	/*
	 * Discard the region lock if we're still holding it.  (The path to
	 * get here is that we acquired the lock because of the caller's
	 * flag argument, but we found the record in the cursor's buffer.
	 * Improbable, but it's easy to avoid.
	 */
	if (rlock == L_ACQUIRED) {
		rlock = L_NONE;
		R_UNLOCK(dbenv, &dblp->reginfo);
	}

	/*
	 * Checksum: there are two types of errors -- a configuration error
	 * or a checksum mismatch.  The former is always bad.  The latter is
	 * OK if we're searching for the end of the log, and very, very bad
	 * if we're reading random log records.
	 */
	db_cipher = dbenv->crypto_handle;
	if ((ret = __db_check_chksum(dbenv, db_cipher,
	    hdr.chksum, rp + hdr.size, hdr.len - hdr.size, is_hmac)) != 0) {
		if (F_ISSET(logc, DB_LOG_SILENT_ERR)) {
			if (ret == 0 || ret == -1)
				ret = EIO;
		} else if (ret == -1) {
			__db_err(dbenv,
		    "DB_LOGC->get: log record LSN %lu/%lu: checksum mismatch",
			    (u_long)nlsn.file, (u_long)nlsn.offset);
			__db_err(dbenv,
		    "DB_LOGC->get: catastrophic recovery may be required");
			ret = __db_panic(dbenv, DB_RUNRECOVERY);
		}
		goto err;
	}

	/*
	 * If we got a 0-length record, that means we're in the midst of
	 * some bytes that got 0'd as the result of a vtruncate.  We're
	 * going to have to retry.
	 */
	if (hdr.len == 0) {
nohdr:		switch (flags) {
		case DB_FIRST:
		case DB_NEXT:
			/* Zero'd records always indicate the end of a file. */
			goto next_file;
		case DB_LAST:
		case DB_PREV:
			/*
			 * We should never get here.  If we recover a log
			 * file with 0's at the end, we'll treat the 0'd
			 * headers as the end of log and ignore them.  If
			 * we're reading backwards from another file, then
			 * the first record in that new file should have its
			 * prev field set correctly.
			 */
			__db_err(dbenv,
		"Encountered zero length records while traversing backwards");
			DB_ASSERT(0);
			ret = __db_panic(dbenv, DB_RUNRECOVERY);
			goto err;
		case DB_SET:
		default:
			/* Return the 0-length record. */
			break;
		}
	}

	/* Copy the record into the user's DBT. */
	if ((ret = __db_retcopy(dbenv, dbt, rp + hdr.size,
	    (u_int32_t)(hdr.len - hdr.size),
	    &logc->c_dbt.data, &logc->c_dbt.ulen)) != 0)
		goto err;

	if (CRYPTO_ON(dbenv)) {
		if ((ret = db_cipher->decrypt(dbenv, db_cipher->data,
		    hdr.iv, dbt->data, hdr.len - hdr.size)) != 0) {
			ret = EAGAIN;
			goto err;
		}
		/*
		 * Return the original log record size to the user,
		 * even though we've allocated more than that, possibly.
		 * The log record is decrypted in the user dbt, not in
		 * the buffer, so we must do this here after decryption,
		 * not adjust the len passed to the __db_retcopy call.
		 */
		dbt->size = hdr.orig_size;
	}

	/* Update the cursor and the returned LSN. */
	*alsn = nlsn;
	logc->c_lsn = nlsn;
	logc->c_len = hdr.len;
	logc->c_prev = hdr.prev;

err:	if (rlock == L_ACQUIRED)
		R_UNLOCK(dbenv, &dblp->reginfo);

	return (ret);
}

/*
 * __log_c_incursor --
 *	Check to see if the requested record is in the cursor's buffer.
 */
static int
__log_c_incursor(logc, lsn, hdr, pp)
	DB_LOGC *logc;
	DB_LSN *lsn;
	HDR *hdr;
	u_int8_t **pp;
{
	u_int8_t *p;
	int eof;

	*pp = NULL;

	/*
	 * Test to see if the requested LSN could be part of the cursor's
	 * buffer.
	 *
	 * The record must be part of the same file as the cursor's buffer.
	 * The record must start at a byte offset equal to or greater than
	 * the cursor buffer.
	 * The record must not start at a byte offset after the cursor
	 * buffer's end.
	 */
	if (logc->bp_lsn.file != lsn->file)
		return (0);
	if (logc->bp_lsn.offset > lsn->offset)
		return (0);
	if (logc->bp_lsn.offset + logc->bp_rlen <= lsn->offset + hdr->size)
		return (0);

	/*
	 * Read the record's header and check if the record is entirely held
	 * in the buffer.  If the record is not entirely held, get it again.
	 * (The only advantage in having part of the record locally is that
	 * we might avoid a system call because we already have the HDR in
	 * memory.)
	 *
	 * If the header check fails for any reason, it must be because the
	 * LSN is bogus.  Fail hard.
	 */
	p = logc->bp + (lsn->offset - logc->bp_lsn.offset);
	memcpy(hdr, p, hdr->size);
	if (__log_c_hdrchk(logc, lsn, hdr, &eof))
		return (DB_NOTFOUND);
	if (eof || logc->bp_lsn.offset + logc->bp_rlen < lsn->offset + hdr->len)
		return (0);

	*pp = p;				/* Success. */

	return (0);
}

/*
 * __log_c_inregion --
 *	Check to see if the requested record is in the region's buffer.
 */
static int
__log_c_inregion(logc, lsn, rlockp, last_lsn, hdr, pp)
	DB_LOGC *logc;
	DB_LSN *lsn, *last_lsn;
	RLOCK *rlockp;
	HDR *hdr;
	u_int8_t **pp;
{
	DB_ENV *dbenv;
	DB_LOG *dblp;
	LOG *lp;
	size_t b_region, len, nr;
	u_int32_t b_disk;
	int eof, ret;
	u_int8_t *p;

	dbenv = logc->dbenv;
	dblp = dbenv->lg_handle;
	lp = ((DB_LOG *)logc->dbenv->lg_handle)->reginfo.primary;

	ret = 0;
	b_region = 0;
	*pp = NULL;

	/* If we haven't yet acquired the log region lock, do so. */
	if (*rlockp == L_NONE) {
		*rlockp = L_ACQUIRED;
		R_LOCK(dbenv, &dblp->reginfo);
	}

	/*
	 * The routines to read from disk must avoid reading past the logical
	 * end of the log, so pass that information back to it.
	 *
	 * Since they're reading directly from the disk, they must also avoid
	 * reading past the offset we've written out.  If the log was
	 * truncated, it's possible that there are zeroes or garbage on
	 * disk after this offset, and the logical end of the log can
	 * come later than this point if the log buffer isn't empty.
	 */
	*last_lsn = lp->lsn;
	if (!lp->db_log_inmemory && last_lsn->offset > lp->w_off)
		last_lsn->offset = lp->w_off;

	/*
	 * Test to see if the requested LSN could be part of the region's
	 * buffer.
	 *
	 * During recovery, we read the log files getting the information to
	 * initialize the region.  In that case, the region's lsn field will
	 * not yet have been filled in, use only the disk.
	 *
	 * The record must not start at a byte offset after the region buffer's
	 * end, since that means the request is for a record after the end of
	 * the log.  Do this test even if the region's buffer is empty -- after
	 * recovery, the log files may continue past the declared end-of-log,
	 * and the disk reading routine will incorrectly attempt to read the
	 * remainder of the log.
	 *
	 * Otherwise, test to see if the region's buffer actually has what we
	 * want:
	 *
	 * The buffer must have some useful content.
	 * The record must be in the same file as the region's buffer and must
	 * start at a byte offset equal to or greater than the region's buffer.
	 */
	if (IS_ZERO_LSN(lp->lsn))
		return (0);
	if (log_compare(lsn, &lp->lsn) >= 0)
		return (DB_NOTFOUND);
	else if (lp->db_log_inmemory) {
		if ((ret = __log_inmem_lsnoff(dblp, lsn, &b_region)) != 0)
			return (ret);
	} else if (lp->b_off == 0 || log_compare(lsn, &lp->f_lsn) < 0)
		return (0);

	/*
	 * The current contents of the cursor's buffer will be useless for a
	 * future call, we're about to overwrite it -- trash it rather than
	 * try and make it look correct.
	 */
	logc->bp_rlen = 0;

	/*
	 * If the requested LSN is greater than the region buffer's first
	 * byte, we know the entire record is in the buffer on a good LSN.
	 *
	 * If we're given a bad LSN, the "entire" record might not be in
	 * our buffer in order to fail at the chksum.  __log_c_hdrchk made
	 * sure our dest buffer fits, via bp_maxrec, but we also need to
	 * make sure we don't run off the end of this buffer, the src.
	 *
	 * There is one case where the header check can fail: on a scan through
	 * in-memory logs, when we reach the end of a file we can read an empty
	 * heady.  In that case, it's safe to return zero, here: it will be
	 * caught in our caller.  Otherwise, the LSN is bogus.  Fail hard.
	 */
	if (lp->db_log_inmemory || log_compare(lsn, &lp->f_lsn) > 0) {
		if (!lp->db_log_inmemory)
			b_region = lsn->offset - lp->w_off;
		__log_inmem_copyout(dblp, b_region, hdr, hdr->size);
		if (__log_c_hdrchk(logc, lsn, hdr, &eof) != 0)
			return (DB_NOTFOUND);
		if (eof)
			return (0);
		if (lp->db_log_inmemory) {
			if (RINGBUF_LEN(lp, b_region, lp->b_off) < hdr->len)
				return (DB_NOTFOUND);
		} else if (lsn->offset + hdr->len > lp->w_off + lp->buffer_size)
			return (DB_NOTFOUND);
		if (logc->bp_size <= hdr->len) {
			len = (size_t)DB_ALIGN(hdr->len * 2, 128);
			if ((ret =
			    __os_realloc(logc->dbenv, len, &logc->bp)) != 0)
				 return (ret);
			logc->bp_size = (u_int32_t)len;
		}
		__log_inmem_copyout(dblp, b_region, logc->bp, hdr->len);
		*pp = logc->bp;
		return (0);
	}

	DB_ASSERT(!lp->db_log_inmemory);

	/*
	 * There's a partial record, that is, the requested record starts
	 * in a log file and finishes in the region buffer.  We have to
	 * find out how many bytes of the record are in the region buffer
	 * so we can copy them out into the cursor buffer.  First, check
	 * to see if the requested record is the only record in the region
	 * buffer, in which case we should copy the entire region buffer.
	 *
	 * Else, walk back through the region's buffer to find the first LSN
	 * after the record that crosses the buffer boundary -- we can detect
	 * that LSN, because its "prev" field will reference the record we
	 * want.  The bytes we need to copy from the region buffer are the
	 * bytes up to the record we find.  The bytes we'll need to allocate
	 * to hold the log record are the bytes between the two offsets.
	 */
	b_disk = lp->w_off - lsn->offset;
	if (lp->b_off <= lp->len)
		b_region = (u_int32_t)lp->b_off;
	else