/**
* Close and destroy a transport handle
*/
void rd_kafka_transport_close (rd_kafka_transport_t *rktrans) {
#if WITH_SSL
if (rktrans->rktrans_ssl) {
SSL_shutdown(rktrans->rktrans_ssl);
SSL_free(rktrans->rktrans_ssl);
}
#endif
#if WITH_SASL
if (rktrans->rktrans_sasl.conn)
sasl_dispose(&rktrans->rktrans_sasl.conn);
#endif
if (rktrans->rktrans_recv_buf)
rd_kafka_buf_destroy(rktrans->rktrans_recv_buf);
if (rktrans->rktrans_s != -1) {
#ifndef _MSC_VER
close(rktrans->rktrans_s);
#else
closesocket(rktrans->rktrans_s);
#endif
}
rd_free(rktrans);
}
void rd_kafka_op_destroy (rd_kafka_op_t *rko) {
/* Decrease refcount on rkbuf to eventually rd_free the shared buffer */
if (rko->rko_rkbuf)
rd_kafka_buf_destroy(rko->rko_rkbuf);
else if (rko->rko_payload && rko->rko_flags & RD_KAFKA_OP_F_FREE) {
if (rko->rko_free_cb)
rko->rko_free_cb(rko->rko_payload);
else
rd_free(rko->rko_payload);
}
if (rko->rko_rkt)
rd_kafka_topic_destroy0(rd_kafka_topic_a2s(rko->rko_rkt));
if (rko->rko_rktp)
rd_kafka_toppar_destroy(rko->rko_rktp);
if (rko->rko_metadata)
rd_kafka_metadata_destroy(rko->rko_metadata);
if (rko->rko_replyq)
rd_kafka_q_destroy(rko->rko_replyq);
if (rd_atomic32_sub(&rd_kafka_op_cnt, 1) < 0)
rd_kafka_assert(NULL, !*"rd_kafka_op_cnt < 0");
rd_free(rko);
}
/**
* Handle received frame from broker.
*/
static int rd_kafka_sasl_handle_recv (rd_kafka_transport_t *rktrans,
rd_kafka_buf_t *rkbuf,
char *errstr, int errstr_size) {
int r;
rd_rkb_dbg(rktrans->rktrans_rkb, SECURITY, "SASL",
"Received SASL frame from broker (%"PRIdsz" bytes)",
rkbuf ? rkbuf->rkbuf_len : 0);
if (rktrans->rktrans_sasl.complete && (!rkbuf || rkbuf->rkbuf_len == 0))
goto auth_successful;
do {
sasl_interact_t *interact = NULL;
const char *out;
unsigned int outlen;
r = sasl_client_step(rktrans->rktrans_sasl.conn,
rkbuf && rkbuf->rkbuf_len > 0 ?
rkbuf->rkbuf_rbuf : NULL,
rkbuf ? rkbuf->rkbuf_len : 0,
&interact,
&out, &outlen);
if (rkbuf) {
rd_kafka_buf_destroy(rkbuf);
rkbuf = NULL;
}
if (r >= 0) {
/* Note: outlen may be 0 here for an empty response */
if (rd_kafka_sasl_send(rktrans, out, outlen,
errstr, errstr_size) == -1)
return -1;
}
if (r == SASL_INTERACT)
rd_rkb_dbg(rktrans->rktrans_rkb, SECURITY, "SASL",
"SASL_INTERACT: %lu %s, %s, %s, %p",
interact->id,
interact->challenge,
interact->prompt,
interact->defresult,
interact->result);
} while (r == SASL_INTERACT);
if (r == SASL_CONTINUE)
return 0; /* Wait for more data from broker */
else if (r != SASL_OK) {
rd_snprintf(errstr, errstr_size,
"SASL handshake failed (step): %s",
sasl_errdetail(rktrans->rktrans_sasl.conn));
return -1;
}
/* Authentication successful */
auth_successful:
if (rktrans->rktrans_rkb->rkb_rk->rk_conf.debug &
RD_KAFKA_DBG_SECURITY) {
const char *user, *mech, *authsrc;
if (sasl_getprop(rktrans->rktrans_sasl.conn, SASL_USERNAME,
(const void **)&user) != SASL_OK)
user = "******";
if (sasl_getprop(rktrans->rktrans_sasl.conn, SASL_MECHNAME,
(const void **)&mech) != SASL_OK)
mech = "(unknown)";
if (sasl_getprop(rktrans->rktrans_sasl.conn, SASL_AUTHSOURCE,
(const void **)&authsrc) != SASL_OK)
authsrc = "(unknown)";
rd_rkb_dbg(rktrans->rktrans_rkb, SECURITY, "SASL",
"Authenticated as %s using %s (%s)",
user, mech, authsrc);
}
rd_kafka_broker_connect_up(rktrans->rktrans_rkb);
return 0;
}
/**
* @brief Decompress MessageSet, pass the uncompressed MessageSet to
* the MessageSet reader.
*/
static rd_kafka_resp_err_t
rd_kafka_msgset_reader_decompress (rd_kafka_msgset_reader_t *msetr,
int MsgVersion, int Attributes,
int64_t Timestamp, int64_t Offset,
const void *compressed,
size_t compressed_size) {
struct iovec iov = { .iov_base = NULL, .iov_len = 0 };
rd_kafka_toppar_t *rktp = msetr->msetr_rktp;
int codec = Attributes & RD_KAFKA_MSG_ATTR_COMPRESSION_MASK;
rd_kafka_resp_err_t err = RD_KAFKA_RESP_ERR_NO_ERROR;
rd_kafka_buf_t *rkbufz;
switch (codec)
{
#if WITH_ZLIB
case RD_KAFKA_COMPRESSION_GZIP:
{
uint64_t outlenx = 0;
/* Decompress Message payload */
iov.iov_base = rd_gz_decompress(compressed, (int)compressed_size,
&outlenx);
if (unlikely(!iov.iov_base)) {
rd_rkb_dbg(msetr->msetr_rkb, MSG, "GZIP",
"Failed to decompress Gzip "
"message at offset %"PRId64
" of %"PRIusz" bytes: "
"ignoring message",
Offset, compressed_size);
err = RD_KAFKA_RESP_ERR__BAD_COMPRESSION;
goto err;
}
iov.iov_len = (size_t)outlenx;
}
break;
#endif
#if WITH_SNAPPY
case RD_KAFKA_COMPRESSION_SNAPPY:
{
const char *inbuf = compressed;
size_t inlen = compressed_size;
int r;
static const unsigned char snappy_java_magic[] =
{ 0x82, 'S','N','A','P','P','Y', 0 };
static const size_t snappy_java_hdrlen = 8+4+4;
/* snappy-java adds its own header (SnappyCodec)
* which is not compatible with the official Snappy
* implementation.
* 8: magic, 4: version, 4: compatible
* followed by any number of chunks:
* 4: length
* ...: snappy-compressed data. */
if (likely(inlen > snappy_java_hdrlen + 4 &&
!memcmp(inbuf, snappy_java_magic, 8))) {
/* snappy-java framing */
char errstr[128];
inbuf = inbuf + snappy_java_hdrlen;
inlen -= snappy_java_hdrlen;
iov.iov_base = rd_kafka_snappy_java_uncompress(
inbuf, inlen,
&iov.iov_len,
errstr, sizeof(errstr));
if (unlikely(!iov.iov_base)) {
rd_rkb_dbg(msetr->msetr_rkb, MSG, "SNAPPY",
"%s [%"PRId32"]: "
"Snappy decompression for message "
"at offset %"PRId64" failed: %s: "
"ignoring message",
rktp->rktp_rkt->rkt_topic->str,
rktp->rktp_partition,
Offset, errstr);
err = RD_KAFKA_RESP_ERR__BAD_COMPRESSION;
goto err;
}
} else {
/* No framing */
/* Acquire uncompressed length */
if (unlikely(!rd_kafka_snappy_uncompressed_length(
inbuf, inlen, &iov.iov_len))) {
rd_rkb_dbg(msetr->msetr_rkb, MSG, "SNAPPY",
"Failed to get length of Snappy "
"compressed payload "
"for message at offset %"PRId64
" (%"PRIusz" bytes): "
"ignoring message",
Offset, inlen);
err = RD_KAFKA_RESP_ERR__BAD_COMPRESSION;
goto err;
}
/* Allocate output buffer for uncompressed data */
iov.iov_base = rd_malloc(iov.iov_len);
if (unlikely(!iov.iov_base)) {
rd_rkb_dbg(msetr->msetr_rkb, MSG, "SNAPPY",
"Failed to allocate Snappy "
"decompress buffer of size %"PRIusz
"for message at offset %"PRId64
" (%"PRIusz" bytes): %s: "
"ignoring message",
iov.iov_len, Offset, inlen,
rd_strerror(errno));
err = RD_KAFKA_RESP_ERR__CRIT_SYS_RESOURCE;
goto err;
}
/* Uncompress to outbuf */
if (unlikely((r = rd_kafka_snappy_uncompress(
inbuf, inlen, iov.iov_base)))) {
rd_rkb_dbg(msetr->msetr_rkb, MSG, "SNAPPY",
"Failed to decompress Snappy "
"payload for message at offset "
"%"PRId64" (%"PRIusz" bytes): %s: "
"ignoring message",
Offset, inlen,
rd_strerror(-r/*negative errno*/));
rd_free(iov.iov_base);
err = RD_KAFKA_RESP_ERR__BAD_COMPRESSION;
goto err;
}
}
}
break;
#endif
case RD_KAFKA_COMPRESSION_LZ4:
{
err = rd_kafka_lz4_decompress(msetr->msetr_rkb,
/* Proper HC? */
MsgVersion >= 1 ? 1 : 0,
Offset,
/* @warning Will modify compressed
* if no proper HC */
(char *)compressed,
compressed_size,
&iov.iov_base, &iov.iov_len);
if (err)
goto err;
}
break;
default:
rd_rkb_dbg(msetr->msetr_rkb, MSG, "CODEC",
"%s [%"PRId32"]: Message at offset %"PRId64
" with unsupported "
"compression codec 0x%x: message ignored",
rktp->rktp_rkt->rkt_topic->str,
rktp->rktp_partition,
Offset, (int)codec);
err = RD_KAFKA_RESP_ERR__NOT_IMPLEMENTED;
goto err;
}
rd_assert(iov.iov_base);
/*
* Decompression successful
*/
/* Create a new buffer pointing to the uncompressed
* allocated buffer (outbuf) and let messages keep a reference to
* this new buffer. */
rkbufz = rd_kafka_buf_new_shadow(iov.iov_base, iov.iov_len, rd_free);
rkbufz->rkbuf_rkb = msetr->msetr_rkbuf->rkbuf_rkb;
rd_kafka_broker_keep(rkbufz->rkbuf_rkb);
/* In MsgVersion v0..1 the decompressed data contains
* an inner MessageSet, pass it to a new MessageSet reader.
*
* For MsgVersion v2 the decompressed data are the list of messages.
*/
if (MsgVersion <= 1) {
/* Pass decompressed data (inner Messageset)
* to new instance of the MessageSet parser. */
rd_kafka_msgset_reader_t inner_msetr;
rd_kafka_msgset_reader_init(&inner_msetr,
rkbufz,
msetr->msetr_rktp,
msetr->msetr_tver,
&msetr->msetr_rkq);
if (MsgVersion == 1) {
/* postproc() will convert relative to
* absolute offsets */
inner_msetr.msetr_relative_offsets = 1;
inner_msetr.msetr_outer.offset = Offset;
/* Apply single LogAppendTime timestamp for
* all messages. */
if (Attributes & RD_KAFKA_MSG_ATTR_LOG_APPEND_TIME) {
inner_msetr.msetr_outer.tstype =
RD_KAFKA_TIMESTAMP_LOG_APPEND_TIME;
inner_msetr.msetr_outer.timestamp = Timestamp;
}
}
/* Parse the inner MessageSet */
err = rd_kafka_msgset_reader_run(&inner_msetr);
} else {
/* MsgVersion 2 */
rd_kafka_buf_t *orig_rkbuf = msetr->msetr_rkbuf;
/* Temporarily replace read buffer with uncompressed buffer */
msetr->msetr_rkbuf = rkbufz;
/* Read messages */
err = rd_kafka_msgset_reader_msgs_v2(msetr);
/* Restore original buffer */
msetr->msetr_rkbuf = orig_rkbuf;
}
/* Loose our refcnt of the uncompressed rkbuf.
* Individual messages/rko's will have their own reference. */
rd_kafka_buf_destroy(rkbufz);
return err;
err:
/* Enqueue error messsage:
* Create op and push on temporary queue. */
rd_kafka_q_op_err(&msetr->msetr_rkq, RD_KAFKA_OP_CONSUMER_ERR,
err, msetr->msetr_tver->version, rktp, Offset,
"Decompression (codec 0x%x) of message at %"PRIu64
" of %"PRIu64" bytes failed: %s",
codec, Offset, compressed_size, rd_kafka_err2str(err));
return err;
}
/**
* @brief Message parser for MsgVersion v0..1
*
* @returns RD_KAFKA_RESP_ERR_NO_ERROR on success or on single-message errors,
* or any other error code when the MessageSet parser should stop
* parsing (such as for partial Messages).
*/
static rd_kafka_resp_err_t
rd_kafka_msgset_reader_msg_v0_1 (rd_kafka_msgset_reader_t *msetr) {
rd_kafka_buf_t *rkbuf = msetr->msetr_rkbuf;
rd_kafka_toppar_t *rktp = msetr->msetr_rktp;
rd_kafka_broker_t *rkb = msetr->msetr_rkb;
struct {
int64_t Offset; /* MessageSet header */
int32_t MessageSize; /* MessageSet header */
uint32_t Crc;
int8_t MagicByte; /* MsgVersion */
int8_t Attributes;
int64_t Timestamp; /* v1 */
} hdr; /* Message header */
rd_kafkap_bytes_t Key;
rd_kafkap_bytes_t Value;
int32_t Value_len;
rd_kafka_op_t *rko;
size_t hdrsize = 6; /* Header size following MessageSize */
rd_slice_t crc_slice;
rd_kafka_msg_t *rkm;
int relative_offsets = 0;
const char *reloff_str = "";
/* Only log decoding errors if protocol debugging enabled. */
int log_decode_errors = (rkbuf->rkbuf_rkb->rkb_rk->rk_conf.debug &
RD_KAFKA_DBG_PROTOCOL) ? LOG_DEBUG : 0;
size_t message_end;
rd_kafka_buf_read_i64(rkbuf, &hdr.Offset);
rd_kafka_buf_read_i32(rkbuf, &hdr.MessageSize);
message_end = rd_slice_offset(&rkbuf->rkbuf_reader) + hdr.MessageSize;
rd_kafka_buf_read_i32(rkbuf, &hdr.Crc);
if (!rd_slice_narrow_copy_relative(&rkbuf->rkbuf_reader, &crc_slice,
hdr.MessageSize - 4))
rd_kafka_buf_check_len(rkbuf, hdr.MessageSize - 4);
rd_kafka_buf_read_i8(rkbuf, &hdr.MagicByte);
rd_kafka_buf_read_i8(rkbuf, &hdr.Attributes);
if (hdr.MagicByte == 1) { /* MsgVersion */
rd_kafka_buf_read_i64(rkbuf, &hdr.Timestamp);
hdrsize += 8;
/* MsgVersion 1 has relative offsets for compressed MessageSets*/
if (!(hdr.Attributes & RD_KAFKA_MSG_ATTR_COMPRESSION_MASK) &&
msetr->msetr_relative_offsets) {
relative_offsets = 1;
reloff_str = "relative ";
}
} else
hdr.Timestamp = 0;
/* Verify MessageSize */
if (unlikely(hdr.MessageSize < (ssize_t)hdrsize))
rd_kafka_buf_parse_fail(rkbuf,
"Message at %soffset %"PRId64
" MessageSize %"PRId32
" < hdrsize %"PRIusz,
reloff_str,
hdr.Offset, hdr.MessageSize, hdrsize);
/* Early check for partial messages */
rd_kafka_buf_check_len(rkbuf, hdr.MessageSize - hdrsize);
if (rkb->rkb_rk->rk_conf.check_crcs) {
/* Verify CRC32 if desired. */
uint32_t calc_crc;
calc_crc = rd_slice_crc32(&crc_slice);
rd_dassert(rd_slice_remains(&crc_slice) == 0);
if (unlikely(hdr.Crc != calc_crc)) {
/* Propagate CRC error to application and
* continue with next message. */
rd_kafka_q_op_err(&msetr->msetr_rkq,
RD_KAFKA_OP_CONSUMER_ERR,
RD_KAFKA_RESP_ERR__BAD_MSG,
msetr->msetr_tver->version,
rktp,
hdr.Offset,
"Message at %soffset %"PRId64
" (%"PRId32" bytes) "
"failed CRC32 check "
"(original 0x%"PRIx32" != "
"calculated 0x%"PRIx32")",
reloff_str, hdr.Offset,
hdr.MessageSize, hdr.Crc, calc_crc);
rd_kafka_buf_skip_to(rkbuf, message_end);
rd_atomic64_add(&rkb->rkb_c.rx_err, 1);
/* Continue with next message */
return RD_KAFKA_RESP_ERR_NO_ERROR;
}
}
/* Extract key */
rd_kafka_buf_read_bytes(rkbuf, &Key);
/* Extract Value */
rd_kafka_buf_read_bytes(rkbuf, &Value);
Value_len = RD_KAFKAP_BYTES_LEN(&Value);
/* MessageSets may contain offsets earlier than we
* requested (compressed MessageSets in particular),
* drop the earlier messages.
* Note: the inner offset may only be trusted for
* absolute offsets. KIP-31 introduced
* ApiVersion 2 that maintains relative offsets
* of compressed messages and the base offset
* in the outer message is the offset of
* the *LAST* message in the MessageSet.
* This requires us to assign offsets
* after all messages have been read from
* the messageset, and it also means
* we cant perform this offset check here
* in that case. */
if (!relative_offsets &&
hdr.Offset < rktp->rktp_offsets.fetch_offset)
return RD_KAFKA_RESP_ERR_NO_ERROR; /* Continue with next msg */
/* Handle compressed MessageSet */
if (unlikely(hdr.Attributes & RD_KAFKA_MSG_ATTR_COMPRESSION_MASK))
return rd_kafka_msgset_reader_decompress(
msetr, hdr.MagicByte, hdr.Attributes, hdr.Timestamp,
hdr.Offset, Value.data, Value_len);
/* Pure uncompressed message, this is the innermost
* handler after all compression and cascaded
* MessageSets have been peeled off. */
/* Create op/message container for message. */
rko = rd_kafka_op_new_fetch_msg(&rkm, rktp, msetr->msetr_tver->version,
rkbuf,
hdr.Offset,
(size_t)RD_KAFKAP_BYTES_LEN(&Key),
RD_KAFKAP_BYTES_IS_NULL(&Key) ?
NULL : Key.data,
(size_t)RD_KAFKAP_BYTES_LEN(&Value),
RD_KAFKAP_BYTES_IS_NULL(&Value) ?
NULL : Value.data);
/* Assign message timestamp.
* If message was in a compressed MessageSet and the outer/wrapper
* Message.Attribute had a LOG_APPEND_TIME set, use the
* outer timestamp */
if (msetr->msetr_outer.tstype == RD_KAFKA_TIMESTAMP_LOG_APPEND_TIME) {
rkm->rkm_timestamp = msetr->msetr_outer.timestamp;
rkm->rkm_tstype = msetr->msetr_outer.tstype;
} else if (hdr.MagicByte >= 1 && hdr.Timestamp) {
rkm->rkm_timestamp = hdr.Timestamp;
if (hdr.Attributes & RD_KAFKA_MSG_ATTR_LOG_APPEND_TIME)
rkm->rkm_tstype = RD_KAFKA_TIMESTAMP_LOG_APPEND_TIME;
else
rkm->rkm_tstype = RD_KAFKA_TIMESTAMP_CREATE_TIME;
}
/* Enqueue message on temporary queue */
rd_kafka_q_enq(&msetr->msetr_rkq, rko);
msetr->msetr_msgcnt++;
return RD_KAFKA_RESP_ERR_NO_ERROR; /* Continue */
err_parse:
/* Count all parse errors as partial message errors. */
rd_atomic64_add(&msetr->msetr_rkb->rkb_c.rx_partial, 1);
return rkbuf->rkbuf_err;
}
/**
* Length framed receive handling.
* Currently only supports a the following framing:
* [int32_t:big_endian_length_of_payload][payload]
*
* To be used on POLLIN event, will return:
* -1: on fatal error (errstr will be updated, *rkbufp remains unset)
* 0: still waiting for data (*rkbufp remains unset)
* 1: data complete, (buffer returned in *rkbufp)
*/
int rd_kafka_transport_framed_recvmsg (rd_kafka_transport_t *rktrans,
rd_kafka_buf_t **rkbufp,
char *errstr, size_t errstr_size) {
rd_kafka_buf_t *rkbuf = rktrans->rktrans_recv_buf;
ssize_t r;
const int log_decode_errors = 0;
/* States:
* !rktrans_recv_buf: initial state; set up buf to receive header.
* rkbuf_len == 0: awaiting header
* rkbuf_len > 0: awaiting payload
*/
if (!rkbuf) {
rkbuf = rd_kafka_buf_new(NULL, 1, 4);
/* Point read buffer to main buffer. */
rkbuf->rkbuf_rbuf = rkbuf->rkbuf_buf;
rd_kafka_buf_push(rkbuf, rkbuf->rkbuf_buf, 4);
rktrans->rktrans_recv_buf = rkbuf;
}
r = rd_kafka_transport_recvmsg(rktrans, &rkbuf->rkbuf_msg,
errstr, errstr_size);
if (r == 0)
return 0;
else if (r == -1)
return -1;
rkbuf->rkbuf_wof += r;
if (rkbuf->rkbuf_len == 0) {
/* Frame length not known yet. */
int32_t frame_len;
if (rkbuf->rkbuf_wof < sizeof(frame_len)) {
/* Wait for entire frame header. */
return 0;
}
/* Reader header: payload length */
rd_kafka_buf_read_i32(rkbuf, &frame_len);
if (frame_len < 0 ||
frame_len > rktrans->rktrans_rkb->
rkb_rk->rk_conf.recv_max_msg_size) {
rd_snprintf(errstr, errstr_size,
"Invalid frame size %"PRId32, frame_len);
return -1;
}
rkbuf->rkbuf_len = frame_len;
if (frame_len == 0) {
/* Payload is empty, we're done. */
rktrans->rktrans_recv_buf = NULL;
*rkbufp = rkbuf;
return 1;
}
/* Allocate memory to hold entire frame payload in contigious
* memory. */
rd_kafka_buf_alloc_recvbuf(rkbuf, frame_len);
/* Try reading directly, there is probably more data available*/
return rd_kafka_transport_framed_recvmsg(rktrans, rkbufp,
errstr, errstr_size);
}
if (rkbuf->rkbuf_wof == rkbuf->rkbuf_len) {
/* Payload is complete. */
rktrans->rktrans_recv_buf = NULL;
*rkbufp = rkbuf;
return 1;
}
/* Wait for more data */
return 0;
err:
if (rkbuf)
rd_kafka_buf_destroy(rkbuf);
rd_snprintf(errstr, errstr_size, "Frame header parsing failed");
return -1;
}