lighttpd1.4/src/connections.c

#include "first.h"

#include "base.h"
#include "buffer.h"
#include "burl.h"       /* HTTP_PARSEOPT_HEADER_STRICT */
#include "chunk.h"
#include "log.h"
#include "connections.h"
#include "fdevent.h"
#include "h2.h"
#include "http_header.h"

#include "reqpool.h"
#include "request.h"
#include "response.h"
#include "network.h"
#include "stat_cache.h"

#include "plugin.h"

#include "sock_addr_cache.h"

#include <sys/stat.h>

#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
#include <string.h>

#ifdef HAVE_SYS_FILIO_H
# include <sys/filio.h>
#endif

#include "sys-socket.h"

#define HTTP_LINGER_TIMEOUT 5

#define connection_set_state(r, n) ((r)->state = (n))

__attribute_cold__
static void connection_set_state_error(request_st * const r, const request_state_t state) {
    connection_set_state(r, state);
}

__attribute_cold__
static connection *connection_init(server *srv);

static void connection_reset(connection *con);


static connection *connections_get_new_connection(server *srv) {
	connections * const conns = &srv->conns;
	size_t i;

	if (conns->size == conns->used) {
		conns->size += srv->max_conns >= 128 ? 128 : srv->max_conns > 16 ? 16 : srv->max_conns;
		conns->ptr = realloc(conns->ptr, sizeof(*conns->ptr) * conns->size);
		force_assert(NULL != conns->ptr);

		for (i = conns->used; i < conns->size; i++) {
			conns->ptr[i] = connection_init(srv);
			connection_reset(conns->ptr[i]);
		}
	}

	conns->ptr[conns->used]->ndx = conns->used;
	return conns->ptr[conns->used++];
}

static void connection_del(server *srv, connection *con) {
	connections * const conns = &srv->conns;

	if (-1 == con->ndx) return;
	uint32_t i = (uint32_t)con->ndx;

	/* not last element */

	if (i != --conns->used) {
		connection * const temp = conns->ptr[i];
		conns->ptr[i] = conns->ptr[conns->used];
		conns->ptr[conns->used] = temp;

		conns->ptr[i]->ndx = i;
		conns->ptr[conns->used]->ndx = -1;
	}

	con->ndx = -1;
#if 0
	fprintf(stderr, "%s.%d: del: (%d)", __FILE__, __LINE__, conns->used);
	for (i = 0; i < conns->used; i++) {
		fprintf(stderr, "%d ", conns->ptr[i]->fd);
	}
	fprintf(stderr, "\n");
#endif
}

static void connection_close(connection *con) {
	if (con->fd < 0) con->fd = -con->fd;

	plugins_call_handle_connection_close(con);

	server * const srv = con->srv;
	request_st * const r = &con->request;
	request_reset_ex(r); /*(r->conf.* is still valid below)*/
	connection_set_state(r, CON_STATE_CONNECT);

	chunkqueue_reset(con->read_queue);
	con->request_count = 0;
	con->is_ssl_sock = 0;
	con->revents_err = 0;

	fdevent_fdnode_event_del(srv->ev, con->fdn);
	fdevent_unregister(srv->ev, con->fd);
	con->fdn = NULL;
#ifdef __WIN32
	if (0 == closesocket(con->fd))
#else
	if (0 == close(con->fd))
#endif
		--srv->cur_fds;
	else
		log_perror(r->conf.errh, __FILE__, __LINE__,
		  "(warning) close: %d", con->fd);

	if (r->conf.log_state_handling) {
		log_error(r->conf.errh, __FILE__, __LINE__,
		  "connection closed for fd %d", con->fd);
	}
	con->fd = -1;

	connection_del(srv, con);
}

static void connection_read_for_eos_plain(connection * const con) {
	/* we have to do the linger_on_close stuff regardless
	 * of r->keep_alive; even non-keepalive sockets
	 * may still have unread data, and closing before reading
	 * it will make the client not see all our output.
	 */
	ssize_t len;
	const int type = sock_addr_get_family(&con->dst_addr);
	char buf[16384];
	do {
		len = fdevent_socket_read_discard(con->fd, buf, sizeof(buf),
						  type, SOCK_STREAM);
	} while (len > 0 || (len < 0 && errno == EINTR));

	if (len < 0 && errno == EAGAIN) return;
      #if defined(EWOULDBLOCK) && EWOULDBLOCK != EAGAIN
	if (len < 0 && errno == EWOULDBLOCK) return;
      #endif

	/* 0 == len || (len < 0 && (errno is a non-recoverable error)) */
		con->close_timeout_ts = log_epoch_secs - (HTTP_LINGER_TIMEOUT+1);
}

static void connection_read_for_eos_ssl(connection * const con) {
	if (con->network_read(con, con->read_queue, MAX_READ_LIMIT) < 0)
		con->close_timeout_ts = log_epoch_secs - (HTTP_LINGER_TIMEOUT+1);
	chunkqueue_reset(con->read_queue);
}

static void connection_read_for_eos(connection * const con) {
	!con->is_ssl_sock
	  ? connection_read_for_eos_plain(con)
	  : connection_read_for_eos_ssl(con);
}

static void connection_handle_close_state(connection *con) {
	connection_read_for_eos(con);

	if (log_epoch_secs - con->close_timeout_ts > HTTP_LINGER_TIMEOUT) {
		connection_close(con);
	}
}

static void connection_handle_shutdown(connection *con) {
	plugins_call_handle_connection_shut_wr(con);

	connection_reset(con);
	++con->srv->con_closed;

	/* close the connection */
	if (con->fd >= 0
	    && (con->is_ssl_sock || 0 == shutdown(con->fd, SHUT_WR))) {
		con->close_timeout_ts = log_epoch_secs;

		request_st * const r = &con->request;
		connection_set_state(r, CON_STATE_CLOSE);
		if (r->conf.log_state_handling) {
			log_error(r->conf.errh, __FILE__, __LINE__,
			  "shutdown for fd %d", con->fd);
		}
	} else {
		connection_close(con);
	}
}

__attribute_cold__
static void connection_fdwaitqueue_append(connection *con) {
    connection_list_append(&con->srv->fdwaitqueue, con);
}


static void connection_handle_response_end_state(request_st * const r, connection * const con) {
	if (r->http_version > HTTP_VERSION_1_1) {
		h2_retire_con(r, con);
		r->keep_alive = 0;
		/* set a status so that mod_accesslog, mod_rrdtool hooks are called
		 * in plugins_call_handle_request_done() (XXX: or set to 0 to omit) */
		r->http_status = 100; /* XXX: what if con->state == CON_STATE_ERROR? */
	}

	/* call request_done hook if http_status set (e.g. to log request) */
	/* (even if error, connection dropped, as long as http_status is set) */
	if (r->http_status) plugins_call_handle_request_done(r);

	if (r->state != CON_STATE_ERROR) ++con->srv->con_written;

	if (r->reqbody_length != r->reqbody_queue.bytes_in
	    || r->state == CON_STATE_ERROR) {
		/* request body may not have been read completely */
		r->keep_alive = 0;
		/* clean up failed partial write of 1xx intermediate responses*/
		if (&r->write_queue != con->write_queue) { /*(for HTTP/1.1)*/
			chunkqueue_free(con->write_queue);
			con->write_queue = &r->write_queue;
		}
	}

        if (r->keep_alive) {
		request_reset(r);
		config_reset_config(r);
		con->is_readable = 1; /* potentially trigger optimistic read */
		/*(accounting used by mod_accesslog for HTTP/1.0 and HTTP/1.1)*/
		r->bytes_read_ckpt = con->bytes_read;
		r->bytes_written_ckpt = con->bytes_written;
#if 0
		r->start_hp.tv_sec = con->read_idle_ts = log_epoch_secs;
#endif
		connection_set_state(r, CON_STATE_REQUEST_START);
	} else {
		connection_handle_shutdown(con);
	}
}


__attribute_pure__
static off_t
connection_write_throttled (const connection * const con, off_t max_bytes)
{
    const request_config * const restrict rconf = &con->request.conf;
    if (0 == rconf->global_bytes_per_second && 0 == rconf->bytes_per_second)
        return max_bytes;

    if (rconf->global_bytes_per_second) {
        off_t limit = (off_t)rconf->global_bytes_per_second
                    - *(rconf->global_bytes_per_second_cnt_ptr);
        if (max_bytes > limit)
            max_bytes = limit;
    }

    if (rconf->bytes_per_second) {
        off_t limit = (off_t)rconf->bytes_per_second
                    - con->bytes_written_cur_second;
        if (max_bytes > limit)
            max_bytes = limit;
    }

    return max_bytes > 0 ? max_bytes : 0; /*(0 == reached traffic limit)*/
}


static off_t
connection_write_throttle (connection * const con, off_t max_bytes)
{
    /*assert(max_bytes > 0);*/
    max_bytes = connection_write_throttled(con, max_bytes);
    if (0 == max_bytes) con->traffic_limit_reached = 1;
    return max_bytes;
}


static int
connection_write_chunkqueue (connection * const con, chunkqueue * const restrict cq, off_t max_bytes)
{
    /*assert(!chunkqueue_is_empty(cq));*//* checked by callers */

    con->write_request_ts = log_epoch_secs;

    max_bytes = connection_write_throttle(con, max_bytes);
    if (0 == max_bytes) return 1;

    off_t written = cq->bytes_out;
    int ret;

  #ifdef TCP_CORK
    int corked = 0;
  #endif

    /* walk chunkqueue up to first FILE_CHUNK (if present)
     * This may incur memory load misses for pointer chasing, but effectively
     * preloads part of the chunkqueue, something which used to be a side effect
     * of a previous (less efficient) version of chunkqueue_length() which
     * walked the entire chunkqueue (on each and every call).  The loads here
     * make a measurable difference in performance in underlying call to
     * con->network_write() */
    if (cq->first->next) {
        const chunk *c = cq->first;
        while (c->type == MEM_CHUNK && NULL != (c = c->next)) ;
      #ifdef TCP_CORK
        /* Linux: put a cork into socket as we want to combine write() calls
         * but only if we really have multiple chunks including non-MEM_CHUNK
         * (or if multiple chunks and TLS), and only if TCP socket */
        if (NULL != c || con->is_ssl_sock) {
            const int sa_family = sock_addr_get_family(&con->srv_socket->addr);
            if (sa_family == AF_INET || sa_family == AF_INET6) {
                corked = 1;
                (void)setsockopt(con->fd, IPPROTO_TCP, TCP_CORK,
                                 &corked, sizeof(corked));
            }
        }
      #endif
    }

    ret = con->network_write(con, cq, max_bytes);
    if (ret >= 0) {
        ret = chunkqueue_is_empty(cq) ? 0 : 1;
    }

  #ifdef TCP_CORK
    if (corked) {
        corked = 0;
        (void)setsockopt(con->fd, IPPROTO_TCP, TCP_CORK,
                         &corked, sizeof(corked));
    }
  #endif

    written = cq->bytes_out - written;
    con->bytes_written += written;
    con->bytes_written_cur_second += written;
    request_st * const r = &con->request;
    if (r->conf.global_bytes_per_second_cnt_ptr)
        *(r->conf.global_bytes_per_second_cnt_ptr) += written;

    return ret;
}


static int
connection_write_1xx_info (request_st * const r, connection * const con)
{
    /* (Note: prior 1xx intermediate responses may be present in cq) */
    /* (Note: also choosing not to update con->write_request_ts
     *  which differs from connection_write_chunkqueue()) */
    chunkqueue * const cq = con->write_queue;
    off_t written = cq->bytes_out;

    int rc = con->network_write(con, cq, MAX_WRITE_LIMIT);

    written = cq->bytes_out - written;
    con->bytes_written += written;
    con->bytes_written_cur_second += written;
    if (r->conf.global_bytes_per_second_cnt_ptr)
        *(r->conf.global_bytes_per_second_cnt_ptr) += written;

    if (rc < 0) {
        connection_set_state_error(r, CON_STATE_ERROR);
        return 0; /* error */
    }

    if (!chunkqueue_is_empty(cq)) { /* partial write (unlikely) */
        con->is_writable = 0;
        if (cq == &r->write_queue) {
            /* save partial write of 1xx in separate chunkqueue
             * Note: sending of remainder of 1xx might be delayed
             * until next set of response headers are sent */
            con->write_queue = chunkqueue_init(NULL);
            chunkqueue_append_chunkqueue(con->write_queue, cq);
        }
    }

  #if 0
    /* XXX: accounting inconsistency
     * 1xx is not currently included in r->resp_header_len,
     * so mod_accesslog reporting of %b or %B (FORMAT_BYTES_OUT_NO_HEADER)
     * reports all bytes out minus len of final response headers,
     * but including 1xx intermediate responses.  If 1xx intermediate
     * responses were included in r->resp_header_len, then there are a
     * few places in the code which must be adjusted to use r->resp_header_done
     * instead of (0 == r->resp_header_len) as flag that final response was set
     * (Doing the following would "discard" the 1xx len from bytes_out)
     */
    r->write_queue.bytes_in = r->write_queue.bytes_out = 0;
  #endif

    return 1; /* success */
}


int
connection_send_1xx (request_st * const r, connection * const con)
{
    /* Make best effort to send HTTP/1.1 1xx intermediate */
    /* (Note: if other modules set response headers *before* the
     *  handle_response_start hook, and the backends subsequently sends 1xx,
     *  then the response headers are sent here with 1xx and might be cleared
     *  by caller (http_response_parse_headers() and http_response_check_1xx()),
     *  instead of being sent with the final response.
     *  (e.g. mod_magnet setting response headers, then backend sending 103)) */

    chunkqueue * const cq = con->write_queue; /*(bypass r->write_queue)*/

    buffer * const b = chunkqueue_append_buffer_open(cq);
    buffer_copy_string_len(b, CONST_STR_LEN("HTTP/1.1 "));
    http_status_append(b, r->http_status);
    for (uint32_t i = 0; i < r->resp_headers.used; ++i) {
        const data_string * const ds = (data_string *)r->resp_headers.data[i];

        if (buffer_string_is_empty(&ds->value)) continue;
        if (buffer_string_is_empty(&ds->key)) continue;

        buffer_append_string_len(b, CONST_STR_LEN("\r\n"));
        buffer_append_string_buffer(b, &ds->key);
        buffer_append_string_len(b, CONST_STR_LEN(": "));
        buffer_append_string_buffer(b, &ds->value);
    }
    buffer_append_string_len(b, CONST_STR_LEN("\r\n\r\n"));
    chunkqueue_append_buffer_commit(cq);

    if (con->traffic_limit_reached)
        return 1; /* success; send later if throttled */

    return connection_write_1xx_info(r, con);
}


static int
connection_write_100_continue (request_st * const r, connection * const con)
{
    /* Make best effort to send "HTTP/1.1 100 Continue" */
    static const char http_100_continue[] = "HTTP/1.1 100 Continue\r\n\r\n";

    if (con->traffic_limit_reached)
        return 1; /* success; skip sending if throttled */

    chunkqueue * const cq = con->write_queue; /*(bypass r->write_queue)*/
    chunkqueue_append_mem(cq, http_100_continue, sizeof(http_100_continue)-1);
    return connection_write_1xx_info(r, con);
}


static int connection_handle_write(request_st * const r, connection * const con) {
	/*assert(!chunkqueue_is_empty(cq));*//* checked by callers */

	if (!con->is_writable) return CON_STATE_WRITE;
	int rc = connection_write_chunkqueue(con, con->write_queue, MAX_WRITE_LIMIT);
	switch (rc) {
	case 0:
		if (r->resp_body_finished) {
			connection_set_state(r, CON_STATE_RESPONSE_END);
			return CON_STATE_RESPONSE_END;
		}
		break;
	case -1: /* error on our side */
		log_error(r->conf.errh, __FILE__, __LINE__,
		  "connection closed: write failed on fd %d", con->fd);
		connection_set_state_error(r, CON_STATE_ERROR);
		return CON_STATE_ERROR;
	case -2: /* remote close */
		connection_set_state_error(r, CON_STATE_ERROR);
		return CON_STATE_ERROR;
	case 1:
		/* do not spin trying to send HTTP/2 server Connection Preface
		 * while waiting for TLS negotiation to complete */
		if (con->write_queue->bytes_out)
			con->is_writable = 0;

		/* not finished yet -> WRITE */
		break;
	}

	return CON_STATE_WRITE; /*(state did not change)*/
}

static int connection_handle_write_state(request_st * const r, connection * const con) {
    do {
        /* only try to write if we have something in the queue */
        if (!chunkqueue_is_empty(&r->write_queue)) {
            if (r->http_version <= HTTP_VERSION_1_1) {
                int rc = connection_handle_write(r, con);
                if (rc != CON_STATE_WRITE) return rc;
            }
        } else if (r->resp_body_finished) {
            connection_set_state(r, CON_STATE_RESPONSE_END);
            return CON_STATE_RESPONSE_END;
        }

        if (r->handler_module && !r->resp_body_finished) {
            const plugin * const p = r->handler_module;
            int rc = p->handle_subrequest(r, p->data);
            switch(rc) {
            case HANDLER_WAIT_FOR_EVENT:
            case HANDLER_FINISHED:
            case HANDLER_GO_ON:
                break;
            case HANDLER_WAIT_FOR_FD:
                /* (In addition to waiting for dispatch from fdwaitqueue,
                 *  HTTP/2 connections may retry more frequently after any
                 *  activity occurs on connection or on other streams) */
                connection_fdwaitqueue_append(con);
                break;
            case HANDLER_COMEBACK:
            default:
                log_error(r->conf.errh, __FILE__, __LINE__,
                  "unexpected subrequest handler ret-value: %d %d",
                  con->fd, rc);
                __attribute_fallthrough__
            case HANDLER_ERROR:
                connection_set_state_error(r, CON_STATE_ERROR);
                return CON_STATE_ERROR;
            }
        }
    } while (r->http_version <= HTTP_VERSION_1_1
             && (!chunkqueue_is_empty(&r->write_queue)
                 ? con->is_writable
                 : r->resp_body_finished));

    return CON_STATE_WRITE;
}


__attribute_cold__
static connection *connection_init(server *srv) {
	connection * const con = calloc(1, sizeof(*con));
	force_assert(NULL != con);

	con->fd = 0;
	con->ndx = -1;
	con->bytes_written = 0;
	con->bytes_read = 0;

	con->dst_addr_buf = buffer_init();
	con->srv  = srv;
	con->plugin_slots = srv->plugin_slots;
	con->config_data_base = srv->config_data_base;

	request_st * const r = &con->request;
	request_init_data(r, con, srv);
	config_reset_config(r);
	con->write_queue = &r->write_queue;
	con->read_queue = &r->read_queue;

	/* init plugin-specific per-connection structures */
	con->plugin_ctx = calloc(1, (srv->plugins.used + 1) * sizeof(void *));
	force_assert(NULL != con->plugin_ctx);

	return con;
}


void connections_free(server *srv) {
	connections * const conns = &srv->conns;
	for (uint32_t i = 0; i < conns->size; ++i) {
		connection *con = conns->ptr[i];
		request_st * const r = &con->request;

		connection_reset(con);
		if (con->write_queue != &r->write_queue)
			chunkqueue_free(con->write_queue);
		if (con->read_queue != &r->read_queue)
			chunkqueue_free(con->read_queue);
		request_free_data(r);

		free(con->plugin_ctx);
		buffer_free(con->dst_addr_buf);

		free(con);
	}

	free(conns->ptr);
	conns->ptr = NULL;
}


static void connection_reset(connection *con) {
	request_st * const r = &con->request;
	request_reset(r);
	config_reset_config(r);
	r->bytes_read_ckpt = 0;
	r->bytes_written_ckpt = 0;
	con->is_readable = 1;

	con->bytes_written = 0;
	con->bytes_written_cur_second = 0;
	con->bytes_read = 0;
}


__attribute_cold__
static chunk *
connection_discard_blank_line (chunkqueue * const cq, uint32_t header_len)
{
    /*(separate func only to be able to mark with compiler hint as cold)*/
    chunkqueue_mark_written(cq, header_len);
    return cq->first; /* refresh c after chunkqueue_mark_written() */
}


static chunk * connection_read_header_more(connection *con, chunkqueue *cq, chunk *c, const size_t olen) {
    /*(should not be reached by HTTP/2 streams)*/
    /*if (r->http_version == HTTP_VERSION_2) return NULL;*/
    /*(However, new connections over TLS may become HTTP/2 connections via ALPN
     * and return from this routine with r->http_version == HTTP_VERSION_2) */

    if ((NULL == c || NULL == c->next) && con->is_readable) {