in this commit the optimisation added in line 891

/* Adjust cpy_len to ensure zero-copy operation for remaining parts of current message */
if (client->msg_idx >= MQTT_VAR_HEADER_BUFFER_LEN) {
  if (cpy_len > (p->len - in_offset))
    cpy_len = p->len - in_offset;
}

this cause to infintie loop if p is chained (two or more pbufs), because when whe move on next pbuf
cpy_len become zero and we never processed more data. here is the log:

mqtt_parse_incoming: msg_idx: 9596, cpy_len: 124, remaining 598149
mqtt_parse_incoming: msg_idx: 9720, cpy_len: 124, remaining 598025
mqtt_parse_incoming: msg_idx: 9844, cpy_len: 124, remaining 597901
mqtt_parse_incoming: msg_idx: 9968, cpy_len: 124, remaining 597777
mqtt_parse_incoming: msg_idx: 10092, cpy_len: 124, remaining 597653
mqtt_parse_incoming: msg_idx: 10216, cpy_len: 124, remaining 597529
mqtt_parse_incoming: msg_idx: 10340, cpy_len: 124, remaining 597405
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 60, remaining 597345
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 0, remaining 597345
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 0, remaining 597345
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 0, remaining 597345
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 0, remaining 597345
mqtt_parse_incoming: msg_idx: 10400, cpy_len: 0, remaining 597345
[inline loop]

Proper version of this must look like (same as check in pbuf_get_contiguous)

/* Adjust cpy_len to ensure zero-copy operation for remaining parts of current message */
q = pbuf_skip(p, in_offset, &out_offset);
  if (q->len < (out_offset + cpy_len)) {
    cpy_len = q->len - out_offset;
}

with this fix, I successfully received large files without fails.

The next is optimization part (my proposals)

Since zero copy always perform so pbuf_get_contiguous never use our rx_buffer

var_hdr_payload = (u8_t*)pbuf_get_contiguous(p, client->rx_buffer + fixed_hdr_len,
                                                buffer_space, cpy_len, in_offset);

it may be replaced with

var_hdr_payload = (u8_t*)pbuf_get_contiguous(p, NULL, 0, cpy_len, in_offset);

More over, since our rx_buffer is not used, it is not necessary to limit cpy_len to available space in buffer

/* Limit to available space in buffer */
buffer_space = MQTT_VAR_HEADER_BUFFER_LEN - fixed_hdr_len;
if (cpy_len > buffer_space) {
   cpy_len = buffer_space;
}

this leads to the following that bigger parts of data process directly from pbufs in mqtt_message_received and user app (not by peaces).

However there is a LWIP_ERROR in mqtt_message_received that check data fits in rx_buffer

LWIP_ERROR("buffer length mismatch", fixed_hdr_len + length <= MQTT_VAR_HEADER_BUFFER_LEN,
             return MQTT_CONNECT_DISCONNECTED);

I found out that no problems occur to process bigger data.
Also docs mention that app callbacks data valid only before callback returns without limiting its size.
So I deleted this check and ran some tests.

Everything works ok. I received larger peaces of data in app callback
and overall download time of large file (~600K) decreased from 5min to 1min.