]> git.ipfire.org Git - thirdparty/haproxy.git/commitdiff
MEDIUM: mux-fcgi: Add the FCGI multiplexer
authorChristopher Faulet <cfaulet@haproxy.com>
Sun, 11 Aug 2019 21:11:30 +0000 (23:11 +0200)
committerChristopher Faulet <cfaulet@haproxy.com>
Tue, 17 Sep 2019 08:18:54 +0000 (10:18 +0200)
This multiplexer is only available on the backend side. It may handle
multiplexed connections if the FCGI application supports it. A FCGI application
must be configured on the backend to be used. If not redefined during the
request processing by the FCGI filter, this mux handles all mandatory
parameters.

There is a limitation on the way the requests are processed. The parameters must
be encoded into a uniq PARAMS record. It means, once encoded, all HTTP headers
and FCGI parameters must small enough to be store in a buffer. Otherwise, an
internal processing error is returned.

Makefile
src/mux_fcgi.c [new file with mode: 0644]

index 5996d802017fd172b811a91a32ddf0acb7c68995..94b0af5f4de65c213b7d6dc4abbc77b570e5fd29 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -787,7 +787,7 @@ OBJS = src/http_ana.o src/cfgparse-listen.o src/stream.o                      \
        src/protocol.o src/arg.o src/hpack-huff.o src/base64.o src/ring.o      \
        src/hash.o src/mailers.o src/activity.o src/version.o src/trace.o      \
        src/mworker.o src/mworker-prog.o src/debug.o src/wdt.o src/dict.o      \
-       src/xprt_handshake.o src/h1_htx.o src/fcgi.o src/fcgi-app.o
+       src/xprt_handshake.o src/h1_htx.o src/fcgi.o src/fcgi-app.o src/mux_fcgi.o
 
 EBTREE_OBJS = $(EBTREE_DIR)/ebtree.o $(EBTREE_DIR)/eb32sctree.o \
               $(EBTREE_DIR)/eb32tree.o $(EBTREE_DIR)/eb64tree.o \
diff --git a/src/mux_fcgi.c b/src/mux_fcgi.c
new file mode 100644 (file)
index 0000000..08d3806
--- /dev/null
@@ -0,0 +1,3424 @@
+/*
+ * FastCGI mux-demux for connections
+ *
+ * Copyright (C) 2019 HAProxy Technologies, Christopher Faulet <cfaulet@haproxy.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ */
+
+#include <common/cfgparse.h>
+#include <common/config.h>
+#include <common/fcgi.h>
+#include <common/h1.h>
+#include <common/htx.h>
+#include <common/initcall.h>
+#include <common/ist.h>
+#include <common/mini-clist.h>
+#include <common/net_helper.h>
+
+#include <types/proxy.h>
+#include <types/session.h>
+
+#include <proto/connection.h>
+#include <proto/fcgi-app.h>
+#include <proto/h1_htx.h>
+#include <proto/http_htx.h>
+#include <proto/log.h>
+#include <proto/session.h>
+#include <proto/ssl_sock.h>
+#include <proto/stream.h>
+#include <proto/stream_interface.h>
+
+/* FCGI Connection flags (32 bits) */
+#define FCGI_CF_NONE           0x00000000
+
+/* Flags indicating why writing to the mux is blockes */
+#define FCGI_CF_MUX_MALLOC      0x00000001 /* mux is blocked on lack connection's mux buffer */
+#define FCGI_CF_MUX_MFULL       0x00000002 /* mux is blocked on connection's mux buffer full */
+#define FCGI_CF_MUX_BLOCK_ANY   0x00000003 /* mux is blocked on connection's mux buffer full */
+
+/* Flags indicating why writing to the demux is blocked.
+ * The first two ones directly affect the ability for the mux to receive data
+ * from the connection. The other ones affect the mux's ability to demux
+ * received data.
+ */
+#define FCGI_CF_DEM_DALLOC      0x00000004  /* demux blocked on lack of connection's demux buffer */
+#define FCGI_CF_DEM_DFULL       0x00000008  /* demux blocked on connection's demux buffer full */
+#define FCGI_CF_DEM_MROOM       0x00000010  /* demux blocked on lack of room in mux buffer */
+#define FCGI_CF_DEM_SALLOC      0x00000020  /* demux blocked on lack of stream's rx buffer */
+#define FCGI_CF_DEM_SFULL       0x00000040  /* demux blocked on stream request buffer full */
+#define FCGI_CF_DEM_TOOMANY     0x00000080  /* demux blocked waiting for some conn_streams to leave */
+#define FCGI_CF_DEM_BLOCK_ANY   0x000000F0  /* aggregate of the demux flags above except DALLOC/DFULL */
+
+/* Other flags */
+#define FCGI_CF_MPXS_CONNS      0x00000100  /* connection multiplexing is supported */
+#define FCGI_CF_ABRTS_SENT      0x00000200  /* a record ABORT was successfully sent to all active streams */
+#define FCGI_CF_ABRTS_FAILED    0x00000400  /* failed to abort processing of all streams */
+#define FCGI_CF_WAIT_FOR_HS     0x00000800  /* We did check that at least a stream was waiting for handshake */
+#define FCGI_CF_KEEP_CONN       0x00001000  /* HAproxy is responsible to close the connection */
+#define FCGI_CF_GET_VALUES      0x00002000  /* retrieve settings */
+
+/* FCGI connection state (fcgi_conn->state) */
+enum fcgi_conn_st {
+       FCGI_CS_INIT = 0,    /* init done, waiting for sending GET_VALUES record */
+       FCGI_CS_SETTINGS,    /* GET_VALUES sent, waiting for the GET_VALUES_RESULT record */
+       FCGI_CS_RECORD_H,    /* GET_VALUES_RESULT received, waiting for a record header */
+       FCGI_CS_RECORD_D,    /* Record header OK, waiting for a record data */
+       FCGI_CS_RECORD_P,    /* Record processed, remains the padding */
+       FCGI_CS_CLOSED,      /* abort requests if necessary and  close the connection ASAP */
+       FCGI_CS_ENTRIES
+} __attribute__((packed));
+
+/* 32 buffers: one for the ring's root, rest for the mbuf itself */
+#define FCGI_C_MBUF_CNT 32
+
+/* FCGI connection descriptor */
+struct fcgi_conn {
+       struct connection *conn;
+
+       enum fcgi_conn_st state;              /* FCGI connection state */
+       int16_t max_id;                       /* highest ID known on this connection, <0 before mgmt records */
+       uint32_t streams_limit;               /* maximum number of concurrent streams the peer supports */
+       uint32_t flags;                      /* Connection flags: FCGI_CF_* */
+
+       int16_t  dsi;                        /* dmux stream ID (<0 = idle ) */
+       uint16_t drl;                        /* demux record length (if dsi >= 0) */
+       uint8_t  drt;                        /* demux record type (if dsi >= 0) */
+       uint8_t  drp;                        /* demux record padding (if dsi >= 0) */
+
+       struct buffer dbuf;                  /* demux buffer */
+       struct buffer mbuf[FCGI_C_MBUF_CNT]; /* mux buffers (ring) */
+
+       int timeout;                         /* idle timeout duration in ticks */
+       int shut_timeout;                    /* idle timeout duration in ticks after shutdown */
+       unsigned int nb_streams;             /* number of streams in the tree */
+       unsigned int nb_cs;                  /* number of attached conn_streams */
+       unsigned int nb_reserved;            /* number of reserved streams */
+       unsigned int stream_cnt;             /* total number of streams seen */
+
+       struct proxy *proxy;                 /* the proxy this connection was created for */
+       struct fcgi_app *app;                /* FCGI application used by this mux */
+       struct task *task;                   /* timeout management task */
+       struct eb_root streams_by_id;        /* all active streams by their ID */
+
+       struct list send_list;               /* list of blocked streams requesting to send */
+       struct list sending_list;            /* list of fcgi_strm scheduled to send data */
+
+       struct buffer_wait buf_wait;         /* Wait list for buffer allocation */
+       struct wait_event wait_event;        /* To be used if we're waiting for I/Os */
+};
+
+
+/* FCGI stream state, in fcgi_strm->state */
+enum fcgi_strm_st {
+       FCGI_SS_IDLE = 0,
+       FCGI_SS_OPEN,
+       FCGI_SS_HREM,     // half-closed(remote)
+       FCGI_SS_HLOC,     // half-closed(local)
+       FCGI_SS_ERROR,
+       FCGI_SS_CLOSED,
+       FCGI_SS_ENTRIES
+} __attribute__((packed));
+
+
+/* FCGI stream flags (32 bits) */
+#define FCGI_SF_NONE           0x00000000
+#define FCGI_SF_ES_RCVD        0x00000001 /* end-of-stream received (empty STDOUT or EDN_REQUEST record) */
+#define FCGI_SF_ES_SENT        0x00000002 /* end-of-strem sent (empty STDIN record) */
+#define FCGI_SF_ABRT_SENT      0x00000004 /* abort sent (ABORT_REQUEST record) */
+
+/* Stream flags indicating the reason the stream is blocked */
+#define FCGI_SF_BLK_MBUSY      0x00000010 /* blocked waiting for mux access (transient) */
+#define FCGI_SF_BLK_MROOM      0x00000020 /* blocked waiting for room in the mux */
+#define FCGI_SF_BLK_ANY        0x00000030 /* any of the reasons above */
+
+#define FCGI_SF_BEGIN_SENT     0x00000100  /* a BEGIN_REQUEST record was sent for this stream */
+#define FCGI_SF_OUTGOING_DATA  0x00000200  /* set whenever we've seen outgoing data */
+
+#define FCGI_SF_WANT_SHUTR     0x00001000  /* a stream couldn't shutr() (mux full/busy) */
+#define FCGI_SF_WANT_SHUTW     0x00002000  /* a stream couldn't shutw() (mux full/busy) */
+#define FCGI_SF_KILL_CONN      0x00004000  /* kill the whole connection with this stream */
+
+/* Other flags */
+#define FCGI_SF_HAVE_I_TLR     0x00010000 /* Set during input process to know the trailers were processed */
+
+/* FCGI stream descriptor */
+struct fcgi_strm {
+       struct conn_stream *cs;
+       struct session *sess;
+       struct fcgi_conn *fconn;
+
+       int32_t id;                   /* stream ID */
+
+       uint32_t flags;               /* Connection flags: FCGI_SF_* */
+       enum fcgi_strm_st state;      /* FCGI stream state */
+       int proto_status;             /* FCGI_PS_* */
+
+       struct h1m h1m;               /* response parser state for H1 */
+
+       struct buffer rxbuf;          /* receive buffer, always valid (buf_empty or real buffer) */
+
+       struct eb32_node by_id;       /* place in fcgi_conn's streams_by_id */
+       struct wait_event wait_event; /* Wait list, when we're attempting to send an ABORT but we can't send */
+       struct wait_event *recv_wait; /* Address of the wait_event the conn_stream associated is waiting on */
+       struct wait_event *send_wait; /* Address of the wait_event the conn_stream associated is waiting on */
+       struct list send_list;        /* To be used when adding in fcgi_conn->send_list */
+       struct list sending_list;     /* To be used when adding in fcgi_conn->sending_list */
+};
+
+/* Flags representing all default FCGI parameters */
+#define FCGI_SP_CGI_GATEWAY    0x00000001
+#define FCGI_SP_DOC_ROOT       0x00000002
+#define FCGI_SP_SCRIPT_NAME    0x00000004
+#define FCGI_SP_PATH_INFO      0x00000008
+#define FCGI_SP_REQ_URI        0x00000010
+#define FCGI_SP_REQ_METH       0x00000020
+#define FCGI_SP_REQ_QS         0x00000040
+#define FCGI_SP_SRV_PORT       0x00000080
+#define FCGI_SP_SRV_PROTO      0x00000100
+#define FCGI_SP_SRV_NAME       0x00000200
+#define FCGI_SP_REM_ADDR       0x00000400
+#define FCGI_SP_REM_PORT       0x00000800
+#define FCGI_SP_SCRIPT_FILE    0x00001000
+#define FCGI_SP_PATH_TRANS     0x00002000
+#define FCGI_SP_CONT_LEN       0x00004000
+#define FCGI_SP_HTTPS          0x00008000
+#define FCGI_SP_MASK           0x0000FFFF
+#define FCGI_SP_URI_MASK       (FCGI_SP_SCRIPT_NAME|FCGI_SP_PATH_INFO|FCGI_SP_REQ_QS)
+
+/* FCGI parameters used when PARAMS record is sent */
+struct fcgi_strm_params {
+       uint32_t mask;
+       struct ist docroot;
+       struct ist scriptname;
+       struct ist pathinfo;
+       struct ist meth;
+       struct ist uri;
+       struct ist vsn;
+       struct ist qs;
+       struct ist srv_name;
+       struct ist srv_port;
+       struct ist rem_addr;
+       struct ist rem_port;
+       struct ist cont_len;
+       int https;
+       struct buffer *p;
+};
+
+/* Maximum amount of data we're OK with re-aligning for buffer optimizations */
+#define MAX_DATA_REALIGN 1024
+
+/* FCGI connection and stream pools */
+DECLARE_STATIC_POOL(pool_head_fcgi_conn, "fcgi_conn", sizeof(struct fcgi_conn));
+DECLARE_STATIC_POOL(pool_head_fcgi_strm, "fcgi_strm", sizeof(struct fcgi_strm));
+
+static struct task *fcgi_timeout_task(struct task *t, void *context, unsigned short state);
+static int fcgi_process(struct fcgi_conn *fconn);
+static struct task *fcgi_io_cb(struct task *t, void *ctx, unsigned short state);
+static inline struct fcgi_strm *fcgi_conn_st_by_id(struct fcgi_conn *fconn, int id);
+static struct task *fcgi_deferred_shut(struct task *t, void *ctx, unsigned short state);
+static struct fcgi_strm *fcgi_conn_stream_new(struct fcgi_conn *fconn, struct conn_stream *cs, struct session *sess);
+static void fcgi_strm_alert(struct fcgi_strm *fstrm);
+static int fcgi_strm_send_abort(struct fcgi_conn *fconn, struct fcgi_strm *fstrm);
+
+/* a dmumy management stream */
+static const struct fcgi_strm *fcgi_mgmt_stream = &(const struct fcgi_strm){
+       .cs        = NULL,
+       .fconn     = NULL,
+       .state     = FCGI_SS_CLOSED,
+       .flags     = FCGI_SF_NONE,
+       .id        = 0,
+};
+
+/* and a dummy idle stream for use with any unknown stream */
+static const struct fcgi_strm *fcgi_unknown_stream = &(const struct fcgi_strm){
+       .cs        = NULL,
+       .fconn     = NULL,
+       .state     = FCGI_SS_IDLE,
+       .flags     = FCGI_SF_NONE,
+       .id        = 0,
+};
+
+
+/*****************************************************/
+/* functions below are for dynamic buffer management */
+/*****************************************************/
+
+/* Indicates whether or not the we may call the fcgi_recv() function to attempt
+ * to receive data into the buffer and/or demux pending data. The condition is
+ * a bit complex due to some API limits for now. The rules are the following :
+ *   - if an error or a shutdown was detected on the connection and the buffer
+ *     is empty, we must not attempt to receive
+ *   - if the demux buf failed to be allocated, we must not try to receive and
+ *     we know there is nothing pending
+ *   - if no flag indicates a blocking condition, we may attempt to receive,
+ *     regardless of whether the demux buffer is full or not, so that only
+ *     de demux part decides whether or not to block. This is needed because
+ *     the connection API indeed prevents us from re-enabling receipt that is
+ *     already enabled in a polled state, so we must always immediately stop
+ *     as soon as the demux can't proceed so as never to hit an end of read
+ *     with data pending in the buffers.
+ *   - otherwise must may not attempt
+ */
+static inline int fcgi_recv_allowed(const struct fcgi_conn *fconn)
+{
+       if (b_data(&fconn->dbuf) == 0 &&
+           (fconn->state == FCGI_CS_CLOSED ||
+            fconn->conn->flags & CO_FL_ERROR ||
+            conn_xprt_read0_pending(fconn->conn)))
+               return 0;
+
+       if (!(fconn->flags & FCGI_CF_DEM_DALLOC) &&
+           !(fconn->flags & FCGI_CF_DEM_BLOCK_ANY))
+               return 1;
+
+       return 0;
+}
+
+/* Restarts reading on the connection if it was not enabled */
+static inline void fcgi_conn_restart_reading(const struct fcgi_conn *fconn, int consider_buffer)
+{
+       if (!fcgi_recv_allowed(fconn))
+               return;
+       if ((!consider_buffer || !b_data(&fconn->dbuf)) &&
+           (fconn->wait_event.events & SUB_RETRY_RECV))
+               return;
+       tasklet_wakeup(fconn->wait_event.tasklet);
+}
+
+
+/* Tries to grab a buffer and to re-enable processing on mux <target>. The
+ * fcgi_conn flags are used to figure what buffer was requested. It returns 1 if
+ * the allocation succeeds, in which case the connection is woken up, or 0 if
+ * it's impossible to wake up and we prefer to be woken up later.
+ */
+static int fcgi_buf_available(void *target)
+{
+       struct fcgi_conn *fconn = target;
+       struct fcgi_strm *fstrm;
+
+       if ((fconn->flags & FCGI_CF_DEM_DALLOC) && b_alloc_margin(&fconn->dbuf, 0)) {
+               fconn->flags &= ~FCGI_CF_DEM_DALLOC;
+               fcgi_conn_restart_reading(fconn, 1);
+               return 1;
+       }
+
+       if ((fconn->flags & FCGI_CF_MUX_MALLOC) && b_alloc_margin(br_tail(fconn->mbuf), 0)) {
+               fconn->flags &= ~FCGI_CF_MUX_MALLOC;
+
+               if (fconn->flags & FCGI_CF_DEM_MROOM) {
+                       fconn->flags &= ~FCGI_CF_DEM_MROOM;
+                       fcgi_conn_restart_reading(fconn, 1);
+               }
+               return 1;
+       }
+
+       if ((fconn->flags & FCGI_CF_DEM_SALLOC) &&
+           (fstrm = fcgi_conn_st_by_id(fconn, fconn->dsi)) && fstrm->cs &&
+           b_alloc_margin(&fstrm->rxbuf, 0)) {
+               fconn->flags &= ~FCGI_CF_DEM_SALLOC;
+               fcgi_conn_restart_reading(fconn, 1);
+               return 1;
+       }
+
+       return 0;
+}
+
+static inline struct buffer *fcgi_get_buf(struct fcgi_conn *fconn, struct buffer *bptr)
+{
+       struct buffer *buf = NULL;
+
+       if (likely(!LIST_ADDED(&fconn->buf_wait.list)) &&
+           unlikely((buf = b_alloc_margin(bptr, 0)) == NULL)) {
+               fconn->buf_wait.target = fconn;
+               fconn->buf_wait.wakeup_cb = fcgi_buf_available;
+               HA_SPIN_LOCK(BUF_WQ_LOCK, &buffer_wq_lock);
+               LIST_ADDQ(&buffer_wq, &fconn->buf_wait.list);
+               HA_SPIN_UNLOCK(BUF_WQ_LOCK, &buffer_wq_lock);
+               __conn_xprt_stop_recv(fconn->conn);
+       }
+       return buf;
+}
+
+static inline void fcgi_release_buf(struct fcgi_conn *fconn, struct buffer *bptr)
+{
+       if (bptr->size) {
+               b_free(bptr);
+               offer_buffers(NULL, tasks_run_queue);
+       }
+}
+
+static inline void fcgi_release_mbuf(struct fcgi_conn *fconn)
+{
+       struct buffer *buf;
+       unsigned int count = 0;
+
+       while (b_size(buf = br_head_pick(fconn->mbuf))) {
+               b_free(buf);
+               count++;
+       }
+       if (count)
+               offer_buffers(NULL, tasks_run_queue);
+}
+
+/* Returns the number of allocatable outgoing streams for the connection taking
+ * the number reserved streams into account.
+ */
+static inline int fcgi_streams_left(const struct fcgi_conn *fconn)
+{
+       int ret;
+
+       ret = (unsigned int)(0x7FFF - fconn->max_id) - fconn->nb_reserved - 1;
+       if (ret < 0)
+               ret = 0;
+       return ret;
+}
+
+/* Returns the number of streams in use on a connection to figure if it's
+ * idle or not. We check nb_cs and not nb_streams as the caller will want
+ * to know if it was the last one after a detach().
+ */
+static int fcgi_used_streams(struct connection *conn)
+{
+       struct fcgi_conn *fconn = conn->ctx;
+
+       return fconn->nb_cs;
+}
+
+/* Returns the number of concurrent streams available on the connection */
+static int fcgi_avail_streams(struct connection *conn)
+{
+       struct server *srv = objt_server(conn->target);
+       struct fcgi_conn *fconn = conn->ctx;
+       int ret1, ret2;
+
+       /* Don't open new stream if the connection is closed */
+       if (fconn->state == FCGI_CS_CLOSED)
+               return 0;
+
+       /* May be negative if this setting has changed */
+       ret1 = (fconn->streams_limit - fconn->nb_streams);
+
+       /* we must also consider the limit imposed by stream IDs */
+       ret2 = fcgi_streams_left(fconn);
+       ret1 = MIN(ret1, ret2);
+       if (ret1 > 0 && srv && srv->max_reuse >= 0) {
+               ret2 = ((fconn->stream_cnt <= srv->max_reuse) ? srv->max_reuse - fconn->stream_cnt + 1: 0);
+               ret1 = MIN(ret1, ret2);
+       }
+       return ret1;
+}
+
+/*****************************************************************/
+/* functions below are dedicated to the mux setup and management */
+/*****************************************************************/
+
+/* Initializes the mux once it's attached. Only outgoing connections are
+ * supported. So the context is already initialized before installing the
+ * mux. <input> is always used as Input buffer and may contain data. It is the
+ * caller responsibility to not reuse it anymore. Returns < 0 on error.
+ */
+static int fcgi_init(struct connection *conn, struct proxy *px, struct session *sess,
+                    struct buffer *input)
+{
+       struct fcgi_conn *fconn;
+       struct fcgi_strm *fstrm;
+       struct fcgi_app *app = get_px_fcgi_app(px);
+       struct task *t = NULL;
+
+       if (!app)
+               goto fail_conn;
+
+       fconn = pool_alloc(pool_head_fcgi_conn);
+       if (!fconn)
+               goto fail_conn;
+
+       fconn->shut_timeout = fconn->timeout = px->timeout.server;
+       if (tick_isset(px->timeout.serverfin))
+               fconn->shut_timeout = px->timeout.serverfin;
+
+       fconn->flags = FCGI_CF_NONE;
+
+       /* Retrieve usefull info from the FCGI app */
+       if (app->flags & FCGI_APP_FL_KEEP_CONN)
+               fconn->flags |= FCGI_CF_KEEP_CONN;
+       if (app->flags & FCGI_APP_FL_GET_VALUES)
+               fconn->flags |= FCGI_CF_GET_VALUES;
+       if (app->flags & FCGI_APP_FL_MPXS_CONNS)
+               fconn->flags |= FCGI_CF_MPXS_CONNS;
+
+       fconn->proxy = px;
+       fconn->app = app;
+       fconn->task = NULL;
+       if (tick_isset(fconn->timeout)) {
+               t = task_new(tid_bit);
+               if (!t)
+                       goto fail;
+
+               fconn->task = t;
+               t->process = fcgi_timeout_task;
+               t->context = fconn;
+               t->expire = tick_add(now_ms, fconn->timeout);
+       }
+
+       fconn->wait_event.tasklet = tasklet_new();
+       if (!fconn->wait_event.tasklet)
+               goto fail;
+       fconn->wait_event.tasklet->process = fcgi_io_cb;
+       fconn->wait_event.tasklet->context = fconn;
+       fconn->wait_event.events = 0;
+
+       /* Initialise the context. */
+       fconn->state = FCGI_CS_INIT;
+       fconn->conn = conn;
+       fconn->streams_limit = app->maxreqs;
+       fconn->max_id = -1;
+       fconn->nb_streams = 0;
+       fconn->nb_cs = 0;
+       fconn->nb_reserved = 0;
+       fconn->stream_cnt = 0;
+
+       fconn->dbuf = *input;
+       fconn->dsi = -1;
+
+       br_init(fconn->mbuf, sizeof(fconn->mbuf) / sizeof(fconn->mbuf[0]));
+       fconn->streams_by_id = EB_ROOT;
+       LIST_INIT(&fconn->send_list);
+       LIST_INIT(&fconn->sending_list);
+       LIST_INIT(&fconn->buf_wait.list);
+
+       if (t)
+               task_queue(t);
+
+       /* FIXME: this is temporary, for outgoing connections we need to
+        * immediately allocate a stream until the code is modified so that the
+        * caller calls ->attach(). For now the outgoing cs is stored as
+        * conn->ctx by the caller.
+        */
+       fstrm = fcgi_conn_stream_new(fconn, conn->ctx, sess);
+       if (!fstrm)
+               goto fail;
+
+       conn->ctx = fconn;
+
+       /* Repare to read something */
+       fcgi_conn_restart_reading(fconn, 1);
+       return 0;
+
+  fail:
+       task_destroy(t);
+       if (fconn->wait_event.tasklet)
+               tasklet_free(fconn->wait_event.tasklet);
+       pool_free(pool_head_fcgi_conn, fconn);
+  fail_conn:
+       return -1;
+}
+
+/* Returns the next allocatable outgoing stream ID for the FCGI connection, or
+ * -1 if no more is allocatable.
+ */
+static inline int32_t fcgi_conn_get_next_sid(const struct fcgi_conn *fconn)
+{
+       int32_t id = (fconn->max_id + 1) | 1;
+
+       if ((id & 0x80000000U))
+               id = -1;
+       return id;
+}
+
+/* Returns the stream associated with id <id> or NULL if not found */
+static inline struct fcgi_strm *fcgi_conn_st_by_id(struct fcgi_conn *fconn, int id)
+{
+       struct eb32_node *node;
+
+       if (id == 0)
+               return (struct fcgi_strm *)fcgi_mgmt_stream;
+
+       if (id > fconn->max_id)
+               return (struct fcgi_strm *)fcgi_unknown_stream;
+
+       node = eb32_lookup(&fconn->streams_by_id, id);
+       if (!node)
+               return (struct fcgi_strm *)fcgi_unknown_stream;
+       return container_of(node, struct fcgi_strm, by_id);
+}
+
+
+/* Release function. This one should be called to free all resources allocated
+ * to the mux.
+ */
+static void fcgi_release(struct fcgi_conn *fconn)
+{
+       struct connection *conn = NULL;;
+
+       if (fconn) {
+               /* The connection must be attached to this mux to be released */
+               if (fconn->conn && fconn->conn->ctx == fconn)
+                       conn = fconn->conn;
+
+               if (LIST_ADDED(&fconn->buf_wait.list)) {
+                       HA_SPIN_LOCK(BUF_WQ_LOCK, &buffer_wq_lock);
+                       LIST_DEL(&fconn->buf_wait.list);
+                       HA_SPIN_UNLOCK(BUF_WQ_LOCK, &buffer_wq_lock);
+               }
+
+               fcgi_release_buf(fconn, &fconn->dbuf);
+               fcgi_release_mbuf(fconn);
+
+               if (fconn->task) {
+                       fconn->task->context = NULL;
+                       task_wakeup(fconn->task, TASK_WOKEN_OTHER);
+                       fconn->task = NULL;
+               }
+               if (fconn->wait_event.tasklet)
+                       tasklet_free(fconn->wait_event.tasklet);
+               if (fconn->wait_event.events != 0)
+                       conn->xprt->unsubscribe(conn, conn->xprt_ctx, fconn->wait_event.events,
+                                               &fconn->wait_event);
+       }
+
+       if (conn) {
+               conn->mux = NULL;
+               conn->ctx = NULL;
+
+               conn_stop_tracking(conn);
+               conn_full_close(conn);
+               if (conn->destroy_cb)
+                       conn->destroy_cb(conn);
+               conn_free(conn);
+       }
+}
+
+
+/* Retruns true if the FCGI connection must be release */
+static inline int fcgi_conn_is_dead(struct fcgi_conn *fconn)
+{
+       if (eb_is_empty(&fconn->streams_by_id) &&               /* don't close if streams exist */
+           (!(fconn->flags & FCGI_CF_KEEP_CONN) ||             /* don't keep the connection alive */
+            (fconn->conn->flags & CO_FL_ERROR) ||              /* errors close immediately */
+            (fconn->state == FCGI_CS_CLOSED && !fconn->task) ||/* a timeout stroke earlier */
+            (!(fconn->conn->owner)) ||                         /* Nobody's left to take care of the connection, drop it now */
+            (!br_data(fconn->mbuf) &&                          /* mux buffer empty, also process clean events below */
+             conn_xprt_read0_pending(fconn->conn))))
+               return 1;
+       return 0;
+}
+
+
+/********************************************************/
+/* functions below are for the FCGI protocol processing */
+/********************************************************/
+
+/* Returns the stream if of stream <fstrm> or 0 if <strm> is NULL */
+static inline int fcgi_strm_id(const struct fcgi_strm *fstrm)
+{
+       return (fstrm ? fstrm->id : 0);
+}
+
+/* Marks an error on the stream. */
+static inline void fcgi_strm_error(struct fcgi_strm *fstrm)
+{
+       if (fstrm->id && fstrm->state != FCGI_SS_ERROR) {
+               if (fstrm->state < FCGI_SS_ERROR)
+                       fstrm->state = FCGI_SS_ERROR;
+               if (fstrm->cs)
+                       cs_set_error(fstrm->cs);
+       }
+}
+
+/* Attempts to notify the data layer of recv availability */
+static void fcgi_strm_notify_recv(struct fcgi_strm *fstrm)
+{
+       struct wait_event *sw;
+
+       if (fstrm->recv_wait) {
+               sw = fstrm->recv_wait;
+               sw->events &= ~SUB_RETRY_RECV;
+               tasklet_wakeup(sw->tasklet);
+               fstrm->recv_wait = NULL;
+       }
+}
+
+/* Attempts to notify the data layer of send availability */
+static void fcgi_strm_notify_send(struct fcgi_strm *fstrm)
+{
+       struct wait_event *sw;
+
+       if (fstrm->send_wait && !LIST_ADDED(&fstrm->sending_list)) {
+               sw = fstrm->send_wait;
+               sw->events &= ~SUB_RETRY_SEND;
+               LIST_ADDQ(&fstrm->fconn->sending_list, &fstrm->sending_list);
+               tasklet_wakeup(sw->tasklet);
+       }
+}
+
+/* Alerts the data layer, trying to wake it up by all means, following
+ * this sequence :
+ *   - if the fcgi stream' data layer is subscribed to recv, then it's woken up
+ *     for recv
+ *   - if its subscribed to send, then it's woken up for send
+ *   - if it was subscribed to neither, its ->wake() callback is called
+ * It is safe to call this function with a closed stream which doesn't have a
+ * conn_stream anymore.
+ */
+static void fcgi_strm_alert(struct fcgi_strm *fstrm)
+{
+       if (fstrm->recv_wait || fstrm->send_wait) {
+               fcgi_strm_notify_recv(fstrm);
+               fcgi_strm_notify_send(fstrm);
+       }
+       else if (fstrm->cs && fstrm->cs->data_cb->wake != NULL)
+               fstrm->cs->data_cb->wake(fstrm->cs);
+}
+
+/* Writes the 16-bit record size <len> at address <record> */
+static inline void fcgi_set_record_size(void *record, uint16_t len)
+{
+       uint8_t *out = (record + 4);
+
+       *out       = (len >> 8);
+       *(out + 1) = (len & 0xff);
+}
+
+/* Writes the 16-bit stream id <id> at address <record> */
+static inline void fcgi_set_record_id(void *record, uint16_t id)
+{
+       uint8_t *out = (record + 2);
+
+       *out       = (id >> 8);
+       *(out + 1) = (id & 0xff);
+}
+
+/* Marks a FCGI stream as CLOSED and decrement the number of active streams for
+ * its connection if the stream was not yet closed. Please use this exclusively
+ * before closing a stream to ensure stream count is well maintained.
+ */
+static inline void fcgi_strm_close(struct fcgi_strm *fstrm)
+{
+       if (fstrm->state != FCGI_SS_CLOSED) {
+               fstrm->fconn->nb_streams--;
+               if (!fstrm->id)
+                       fstrm->fconn->nb_reserved--;
+               if (fstrm->cs) {
+                       if (!(fstrm->cs->flags & CS_FL_EOS) && !b_data(&fstrm->rxbuf))
+                               fcgi_strm_notify_recv(fstrm);
+               }
+       }
+       fstrm->state = FCGI_SS_CLOSED;
+}
+
+/* Detaches a FCGI stream from its FCGI connection and releases it to the
+ * fcgi_strm pool.
+ */
+static void fcgi_strm_destroy(struct fcgi_strm *fstrm)
+{
+       fcgi_strm_close(fstrm);
+       eb32_delete(&fstrm->by_id);
+       if (b_size(&fstrm->rxbuf)) {
+               b_free(&fstrm->rxbuf);
+               offer_buffers(NULL, tasks_run_queue);
+       }
+       if (fstrm->send_wait != NULL)
+               fstrm->send_wait->events &= ~SUB_RETRY_SEND;
+       if (fstrm->recv_wait != NULL)
+               fstrm->recv_wait->events &= ~SUB_RETRY_RECV;
+       /* There's no need to explicitly call unsubscribe here, the only
+        * reference left would be in the fconn send_list/fctl_list, and if
+        * we're in it, we're getting out anyway
+        */
+       LIST_DEL_INIT(&fstrm->send_list);
+       if (LIST_ADDED(&fstrm->sending_list)) {
+               tasklet_remove_from_tasklet_list(fstrm->send_wait->tasklet);
+               LIST_DEL_INIT(&fstrm->sending_list);
+       }
+       tasklet_free(fstrm->wait_event.tasklet);
+       pool_free(pool_head_fcgi_strm, fstrm);
+}
+
+/* Allocates a new stream <id> for connection <fconn> and adds it into fconn's
+ * stream tree. In case of error, nothing is added and NULL is returned. The
+ * causes of errors can be any failed memory allocation. The caller is
+ * responsible for checking if the connection may support an extra stream prior
+ * to calling this function.
+ */
+static struct fcgi_strm *fcgi_strm_new(struct fcgi_conn *fconn, int id)
+{
+       struct fcgi_strm *fstrm;
+
+       fstrm = pool_alloc(pool_head_fcgi_strm);
+       if (!fstrm)
+               goto out;
+
+       fstrm->wait_event.tasklet = tasklet_new();
+       if (!fstrm->wait_event.tasklet) {
+               pool_free(pool_head_fcgi_strm, fstrm);
+               goto out;
+       }
+       fstrm->send_wait = NULL;
+       fstrm->recv_wait = NULL;
+       fstrm->wait_event.tasklet->process = fcgi_deferred_shut;
+       fstrm->wait_event.tasklet->context = fstrm;
+       fstrm->wait_event.events = 0;
+       LIST_INIT(&fstrm->send_list);
+       LIST_INIT(&fstrm->sending_list);
+       fstrm->fconn = fconn;
+       fstrm->cs = NULL;
+       fstrm->flags = FCGI_SF_NONE;
+       fstrm->proto_status = 0;
+       fstrm->state = FCGI_SS_IDLE;
+       fstrm->rxbuf = BUF_NULL;
+
+       h1m_init_res(&fstrm->h1m);
+       fstrm->h1m.err_pos = -1; // don't care about errors on the request path
+       fstrm->h1m.flags |= (H1_MF_NO_PHDR|H1_MF_CLEAN_CONN_HDR);
+
+       fstrm->by_id.key = fstrm->id = id;
+       if (id > 0)
+               fconn->max_id = id;
+       else
+               fconn->nb_reserved++;
+
+       eb32_insert(&fconn->streams_by_id, &fstrm->by_id);
+       fconn->nb_streams++;
+       fconn->stream_cnt++;
+
+       return fstrm;
+
+  out:
+       return NULL;
+}
+
+/* Allocates a new stream associated to conn_stream <cs> on the FCGI connection
+ * <fconn> and returns it, or NULL in case of memory allocation error or if the
+ * highest possible stream ID was reached.
+ */
+static struct fcgi_strm *fcgi_conn_stream_new(struct fcgi_conn *fconn, struct conn_stream *cs,
+                                             struct session *sess)
+{
+       struct fcgi_strm *fstrm = NULL;
+
+       if (fconn->nb_streams >= fconn->streams_limit)
+               goto out;
+
+       if (fcgi_streams_left(fconn) < 1)
+               goto out;
+
+       /* Defer choosing the ID until we send the first message to create the stream */
+       fstrm = fcgi_strm_new(fconn, 0);
+       if (!fstrm)
+               goto out;
+
+       fstrm->cs = cs;
+       fstrm->sess = sess;
+       cs->ctx = fstrm;
+       fconn->nb_cs++;
+
+  out:
+       return fstrm;
+}
+
+/* Wakes a specific stream and assign its conn_stream some CS_FL_* flags among
+ * CS_FL_ERR_PENDING and CS_FL_ERROR if needed. The stream's state is
+ * automatically updated accordingly. If the stream is orphaned, it is
+ * destroyed.
+ */
+static void fcgi_strm_wake_one_stream(struct fcgi_strm *fstrm)
+{
+       if (!fstrm->cs) {
+               /* this stream was already orphaned */
+               fcgi_strm_destroy(fstrm);
+               return;
+       }
+
+       if (conn_xprt_read0_pending(fstrm->fconn->conn)) {
+               if (fstrm->state == FCGI_SS_OPEN)
+                       fstrm->state = FCGI_SS_HREM;
+               else if (fstrm->state == FCGI_SS_HLOC)
+                       fcgi_strm_close(fstrm);
+       }
+
+       if ((fstrm->fconn->state == FCGI_CS_CLOSED || fstrm->fconn->conn->flags & CO_FL_ERROR)) {
+               fstrm->cs->flags |= CS_FL_ERR_PENDING;
+               if (fstrm->cs->flags & CS_FL_EOS)
+                       fstrm->cs->flags |= CS_FL_ERROR;
+               if (fstrm->state < FCGI_SS_ERROR)
+                       fstrm->state = FCGI_SS_ERROR;
+       }
+
+       fcgi_strm_alert(fstrm);
+}
+
+/* Wakes unassigned streams (ID == 0) attached to the connection. */
+static void fcgi_wake_unassigned_streams(struct fcgi_conn *fconn)
+{
+       struct eb32_node *node;
+       struct fcgi_strm *fstrm;
+
+       node = eb32_lookup(&fconn->streams_by_id, 0);
+       while (node) {
+               fstrm = container_of(node, struct fcgi_strm, by_id);
+               if (fstrm->id > 0)
+                       break;
+               node = eb32_next(node);
+               fcgi_strm_wake_one_stream(fstrm);
+       }
+}
+
+/* Wakes the streams attached to the connection, whose id is greater than <last>
+ * or unassigned.
+ */
+static void fcgi_wake_some_streams(struct fcgi_conn *fconn, int last)
+{
+       struct eb32_node *node;
+       struct fcgi_strm *fstrm;
+
+       /* Wake all streams with ID > last */
+       node = eb32_lookup_ge(&fconn->streams_by_id, last + 1);
+       while (node) {
+               fstrm = container_of(node, struct fcgi_strm, by_id);
+               node = eb32_next(node);
+               fcgi_strm_wake_one_stream(fstrm);
+       }
+       fcgi_wake_unassigned_streams(fconn);
+}
+
+static int fcgi_set_default_param(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                 struct htx *htx, struct htx_sl *sl,
+                                 struct fcgi_strm_params *params)
+{
+       struct connection *cli_conn = objt_conn(fstrm->sess->origin);
+       struct ist p;
+
+       if (!sl)
+               goto error;
+
+       if (!(params->mask & FCGI_SP_DOC_ROOT))
+               params->docroot = fconn->app->docroot;
+
+       if (!(params->mask & FCGI_SP_REQ_METH)) {
+               p  = htx_sl_req_meth(sl);
+               params->meth = ist2(b_tail(params->p), p.len);
+               chunk_memcat(params->p, p.ptr, p.len);
+       }
+       if (!(params->mask & FCGI_SP_REQ_URI)) {
+               p  = htx_sl_req_uri(sl);
+               params->uri = ist2(b_tail(params->p), p.len);
+               chunk_memcat(params->p, p.ptr, p.len);
+       }
+       if (!(params->mask & FCGI_SP_SRV_PROTO)) {
+               p  = htx_sl_req_vsn(sl);
+               params->vsn = ist2(b_tail(params->p), p.len);
+               chunk_memcat(params->p, p.ptr, p.len);
+       }
+       if (!(params->mask & FCGI_SP_SRV_PORT)) {
+               char *end;
+               int port = 0;
+               if (conn_get_dst(cli_conn))
+                       port = get_host_port(cli_conn->dst);
+               end = ultoa_o(port, b_tail(params->p), b_room(params->p));
+               if (!end)
+                       goto error;
+               params->srv_port = ist2(b_tail(params->p), end - b_tail(params->p));
+               params->p->data += params->srv_port.len;
+       }
+       if (!(params->mask & FCGI_SP_SRV_NAME)) {
+               /* If no Host header found, use the server address to fill
+                * srv_name */
+               if (!istlen(params->srv_name)) {
+                       char *ptr = NULL;
+
+                       if (conn_get_dst(cli_conn))
+                               if (addr_to_str(cli_conn->dst, b_tail(params->p), b_room(params->p)) != -1)
+                                       ptr = b_tail(params->p);
+                       if (ptr) {
+                               params->srv_name = ist2(ptr, strlen(ptr));
+                               params->p->data += params->srv_name.len;
+                       }
+               }
+       }
+       if (!(params->mask & FCGI_SP_REM_ADDR)) {
+               char *ptr = NULL;
+
+               if (conn_get_src(cli_conn))
+                       if (addr_to_str(cli_conn->src, b_tail(params->p), b_room(params->p)) != -1)
+                               ptr = b_tail(params->p);
+               if (ptr) {
+                       params->rem_addr = ist2(ptr, strlen(ptr));
+                       params->p->data += params->rem_addr.len;
+               }
+       }
+       if (!(params->mask & FCGI_SP_REM_PORT)) {
+               char *end;
+               int port = 0;
+               if (conn_get_src(cli_conn))
+                       port = get_host_port(cli_conn->src);
+               end = ultoa_o(port, b_tail(params->p), b_room(params->p));
+               if (!end)
+                       goto error;
+               params->rem_port = ist2(b_tail(params->p), end - b_tail(params->p));
+               params->p->data += params->rem_port.len;
+       }
+       if (!(params->mask & FCGI_SP_CONT_LEN)) {
+               struct htx_blk *blk;
+               enum htx_blk_type type;
+               char *end;
+               size_t len = 0;
+
+               for (blk = htx_get_head_blk(htx); blk; blk = htx_get_next_blk(htx, blk)) {
+                       type = htx_get_blk_type(blk);
+
+                       if (type == HTX_BLK_EOM || type == HTX_BLK_TLR || type == HTX_BLK_EOT)
+                               break;
+                       if (type == HTX_BLK_DATA)
+                               len += htx_get_blksz(blk);
+               }
+               end = ultoa_o(len, b_tail(params->p), b_room(params->p));
+               if (!end)
+                       goto error;
+               params->cont_len = ist2(b_tail(params->p), end - b_tail(params->p));
+               params->p->data += params->cont_len.len;
+       }
+       if (!(params->mask & FCGI_SP_HTTPS)) {
+               params->https = ssl_sock_is_ssl(cli_conn);
+       }
+       if ((params->mask & FCGI_SP_URI_MASK) != FCGI_SP_URI_MASK) {
+               /* one of scriptname, pathinfo or query_string is no set */
+               struct ist path = http_get_path(params->uri);
+               int len;
+
+               /* Decode the path. it must first be copied to keep the URI
+                * untouched.
+                */
+               chunk_memcat(params->p, path.ptr, path.len);
+               path.ptr = b_tail(params->p) - path.len;
+               path.ptr[path.len] = '\0';
+               len = url_decode(path.ptr);
+               if (len < 0)
+                       goto error;
+               path.len = len;
+
+               /* No scrit_name set but no valid path ==> error */
+               if (!(params->mask & FCGI_SP_SCRIPT_NAME) && !istlen(path))
+                       goto error;
+
+               /* Find limit between the path and the query-string */
+               for (len = 0; len < path.len && *(path.ptr + len) != '?'; len++);
+
+               /* If there is a query-string, Set it if not already set */
+               if (!(params->mask & FCGI_SP_REQ_QS) && len < path.len)
+                       params->qs = ist2(path.ptr+len+1, path.len-len-1);
+
+               /* If the script_name is set, don't try to deduce the path_info
+                * too. The opposite is not true.
+                */
+               if (params->mask & FCGI_SP_SCRIPT_NAME) {
+                       params->mask |= FCGI_SP_PATH_INFO;
+                       goto end;
+               }
+
+               /* script_name not set, preset it with the path for now */
+               params->scriptname = ist2(path.ptr, len);
+
+               /* If there is no regex to match the pathinfo, just to the last
+                * part and see if the index must be used.
+                */
+               if (!fconn->app->pathinfo_re)
+                       goto check_index;
+
+               /* The regex does not match, just to the last part and see if
+                * the index must be used.
+                */
+               if (!regex_exec_match2(fconn->app->pathinfo_re, path.ptr, len, MAX_MATCH, pmatch, 0))
+                       goto check_index;
+
+               /* We must have at least 2 captures, otherwise we do nothing and
+                * jump to the last part. Only first 2 ones will be considered
+                */
+               if (pmatch[1].rm_so == -1 || pmatch[1].rm_eo == -1 ||
+                   pmatch[2].rm_so == -1 || pmatch[2].rm_eo == -1)
+                       goto check_index;
+
+               /* Finally we can set the script_name and the path_info */
+               params->scriptname = ist2(path.ptr + pmatch[1].rm_so, pmatch[1].rm_eo - pmatch[1].rm_so);
+               params->pathinfo   = ist2(path.ptr + pmatch[2].rm_so, pmatch[2].rm_eo - pmatch[2].rm_so);
+
+         check_index:
+               len = params->scriptname.len;
+               /* the script_name if finished by a '/' so we can add the index
+                * part, if any.
+                */
+               if (istlen(fconn->app->index) && params->scriptname.ptr[len-1] == '/') {
+                       struct ist sn = params->scriptname;
+
+                       params->scriptname = ist2(b_tail(params->p), len+fconn->app->index.len);
+                       chunk_memcat(params->p, sn.ptr, sn.len);
+                       chunk_memcat(params->p, fconn->app->index.ptr, fconn->app->index.len);
+               }
+       }
+
+  end:
+       return 1;
+  error:
+       return 0;
+}
+
+static int fcgi_encode_default_param(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                    struct fcgi_strm_params *params, struct buffer *outbuf, int flag)
+{
+       struct fcgi_param p;
+
+       if (params->mask & flag)
+               return 1;
+
+       chunk_reset(&trash);
+
+       switch (flag) {
+               case FCGI_SP_CGI_GATEWAY:
+                       p.n = ist("GATEWAY_INTERFACE");
+                       p.v = ist("CGI/1.1");
+                       goto encode;
+               case FCGI_SP_DOC_ROOT:
+                       p.n = ist("DOCUMENT_ROOT");
+                       p.v = params->docroot;
+                       goto encode;
+               case FCGI_SP_SCRIPT_NAME:
+                       p.n = ist("SCRIPT_NAME");
+                       p.v = params->scriptname;
+                       goto encode;
+               case FCGI_SP_PATH_INFO:
+                       p.n = ist("PATH_INFO");
+                       p.v = params->pathinfo;
+                       goto encode;
+               case FCGI_SP_REQ_URI:
+                       p.n = ist("REQUEST_URI");
+                       p.v = params->uri;
+                       goto encode;
+               case FCGI_SP_REQ_METH:
+                       p.n = ist("REQUEST_METHOD");
+                       p.v = params->meth;
+                       goto encode;
+               case FCGI_SP_REQ_QS:
+                       p.n = ist("QUERY_STRING");
+                       p.v = params->qs;
+                       goto encode;
+               case FCGI_SP_SRV_NAME:
+                       p.n = ist("SERVER_NAME");
+                       p.v = params->srv_name;
+                       goto encode;
+               case FCGI_SP_SRV_PORT:
+                       p.n = ist("SERVER_PORT");
+                       p.v = params->srv_port;
+                       goto encode;
+               case FCGI_SP_SRV_PROTO:
+                       p.n = ist("SERVER_PROTOCOL");
+                       p.v = params->vsn;
+                       goto encode;
+               case FCGI_SP_REM_ADDR:
+                       p.n = ist("REMOTE_ADDR");
+                       p.v = params->rem_addr;
+                       goto encode;
+               case FCGI_SP_REM_PORT:
+                       p.n = ist("REMOTE_PORT");
+                       p.v = params->rem_port;
+                       goto encode;
+               case FCGI_SP_SCRIPT_FILE:
+                       p.n = ist("SCRIPT_FILENAME");
+                       chunk_memcat(&trash, params->docroot.ptr, params->docroot.len);
+                       chunk_memcat(&trash, params->scriptname.ptr, params->scriptname.len);
+                       p.v = ist2(b_head(&trash), b_data(&trash));
+                       goto encode;
+               case FCGI_SP_PATH_TRANS:
+                       if (!istlen(params->pathinfo))
+                               goto skip;
+                       p.n = ist("PATH_TRANSLATED");
+                       chunk_memcat(&trash, params->docroot.ptr, params->docroot.len);
+                       chunk_memcat(&trash, params->pathinfo.ptr, params->pathinfo.len);
+                       p.v = ist2(b_head(&trash), b_data(&trash));
+                       goto encode;
+               case FCGI_SP_CONT_LEN:
+                       p.n = ist("CONTENT_LENGTH");
+                       p.v = params->cont_len;
+                       goto encode;
+               case FCGI_SP_HTTPS:
+                       if (!params->https)
+                               goto skip;
+                       p.n = ist("HTTPS");
+                       p.v = ist("on");
+                       goto encode;
+               default:
+                       goto skip;
+       }
+
+  encode:
+       if (!istlen(p.v))
+               goto skip;
+       if (!fcgi_encode_param(outbuf, &p))
+               return 0;
+  skip:
+       params->mask |= flag;
+       return 1;
+}
+
+/* Sends a GET_VALUES record. Returns > 0 on success, 0 if it couldn't do
+ * anything. It is highly unexpected, but if the record is larger than a buffer
+ * and cannot be encoded in one time, an error is triggered and the connection is
+ * closed. GET_VALUES record cannot be split.
+ */
+static int fcgi_conn_send_get_values(struct fcgi_conn *fconn)
+{
+       struct buffer outbuf;
+       struct buffer *mbuf;
+       struct fcgi_param max_reqs   = { .n = ist("FCGI_MAX_REQS"),   .v = ist("")};
+       struct fcgi_param mpxs_conns = { .n = ist("FCGI_MPXS_CONNS"), .v = ist("")};
+       int ret;
+
+       mbuf = br_tail(fconn->mbuf);
+  retry:
+       if (!fcgi_get_buf(fconn, mbuf)) {
+               fconn->flags |= FCGI_CF_MUX_MALLOC;
+               fconn->flags |= FCGI_CF_DEM_MROOM;
+               return 0;
+       }
+
+       while (1) {
+               outbuf = b_make(b_tail(mbuf), b_contig_space(mbuf), 0, 0);
+               if (outbuf.size >= 8 || !b_space_wraps(mbuf))
+                       break;
+         realign_again:
+               b_slow_realign(mbuf, trash.area, b_data(mbuf));
+       }
+
+       if (outbuf.size < 8)
+               goto full;
+
+       /* vsn: 1(FCGI_VERSION), type: (9)FCGI_GET_VALUES, id: 0x0000,
+        *  len: 0x0000 (fill later), padding: 0x00, rsv: 0x00 */
+       memcpy(outbuf.area, "\x01\x09\x00\x00\x00\x00\x00\x00", 8);
+       outbuf.data = 8;
+
+       /* Note: Don't send the param FCGI_MAX_CONNS because its value cannot be
+        *       handled by HAProxy.
+        */
+       if (!fcgi_encode_param(&outbuf, &max_reqs) || !fcgi_encode_param(&outbuf, &mpxs_conns))
+               goto full;
+
+       /* update the record's size now */
+       fcgi_set_record_size(outbuf.area, outbuf.data - 8);
+       b_add(mbuf, outbuf.data);
+       ret = 1;
+
+  end:
+       return ret;
+  full:
+       /* Too large to be encoded. For GET_VALUES records, it is an error */
+       if (!b_data(mbuf))
+               goto fail;
+
+       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+               goto retry;
+       fconn->flags |= FCGI_CF_MUX_MFULL;
+       fconn->flags |= FCGI_CF_DEM_MROOM;
+       ret = 0;
+       goto end;
+  fail:
+       fconn->state = FCGI_CS_CLOSED;
+       ret = 0;
+       goto end;
+}
+
+/* Processes a GET_VALUES_RESULT record. Returns > 0 on success, 0 if it
+ * couldn't do anything. It is highly unexpected, but if the record is larger
+ * than a buffer and cannot be decoded in one time, an error is triggered and
+ * the connection is closed. GET_VALUES_RESULT record cannot be split.
+ */
+static int fcgi_conn_handle_values_result(struct fcgi_conn *fconn)
+{
+       struct buffer inbuf;
+       struct buffer *dbuf;
+       size_t offset;
+
+       dbuf = &fconn->dbuf;
+
+       /* Record too large to be fully decoded */
+       if (b_size(dbuf) < (fconn->drl + fconn->drp))
+               goto fail;
+
+       /* process full record only */
+       if (b_data(dbuf) < (fconn->drl + fconn->drp))
+               return 0;
+
+       if (unlikely(b_contig_data(dbuf, b_head_ofs(dbuf)) < fconn->drl)) {
+               /* Realign the dmux buffer if the record wraps. It is unexpected
+                * at this stage because it should be the first record received
+                * from the FCGI application.
+                */
+               b_slow_realign(dbuf, trash.area, 0);
+       }
+
+       inbuf = b_make(b_head(dbuf), b_data(dbuf), 0, fconn->drl);
+
+       for (offset = 0; offset < b_data(&inbuf); ) {
+               struct fcgi_param p;
+               size_t ret;
+
+               ret = fcgi_aligned_decode_param(&inbuf, offset, &p);
+               if (!ret) {
+                       /* name or value too large to be decoded at once */
+                       goto fail;
+               }
+               offset += ret;
+
+               if (isteqi(p.n, ist("FCGI_MPXS_CONNS"))) {
+                       if (isteq(p.v, ist("1")))
+                               fconn->flags |= FCGI_CF_MPXS_CONNS;
+                       else
+                               fconn->flags &= ~FCGI_CF_MPXS_CONNS;
+               }
+               else if (isteqi(p.n, ist("FCGI_MAX_REQS"))) {
+                       fconn->streams_limit = strl2ui(p.v.ptr, p.v.len);
+               }
+               /*
+                * Ignore all other params
+                */
+       }
+
+       /* Reset the number of concurrent streams supported if the FCGI
+        * application does not support connection multiplexing
+        */
+       if (!(fconn->flags & FCGI_CF_MPXS_CONNS))
+               fconn->streams_limit = 1;
+
+       /* We must be sure to have read exactly the announced record length, no
+        * more no less
+        */
+       if (offset != fconn->drl)
+               goto fail;
+
+       b_del(&fconn->dbuf, fconn->drl + fconn->drp);
+       fconn->drl = 0;
+       fconn->drp = 0;
+       fconn->state = FCGI_CS_RECORD_H;
+       fcgi_wake_unassigned_streams(fconn);
+       return 1;
+  fail:
+       fconn->state = FCGI_CS_CLOSED;
+       return 0;
+}
+
+/* Sends an ABORT_REQUEST record for each active streams. Closed streams are
+ * excluded, as the streams which already received the end-of-stream. It returns
+ * > 0 if the record was sent tp all streams. Otherwise it returns 0.
+ */
+static int fcgi_conn_send_aborts(struct fcgi_conn *fconn)
+{
+       struct eb32_node *node;
+       struct fcgi_strm *fstrm;
+
+       node = eb32_lookup_ge(&fconn->streams_by_id, 1);
+       while (node) {
+               fstrm = container_of(node, struct fcgi_strm, by_id);
+               node = eb32_next(node);
+               if (fstrm->state != FCGI_SS_CLOSED &&
+                   !(fstrm->flags & (FCGI_SF_ES_RCVD|FCGI_SF_ABRT_SENT)) &&
+                   !fcgi_strm_send_abort(fconn, fstrm))
+                       return 0;
+       }
+       fconn->flags |= FCGI_CF_ABRTS_SENT;
+       return 1;
+}
+
+/* Sends a BEGIN_REQUEST record. It returns > 0 on success, 0 if it couldn't do
+ * anything. BEGIN_REQUEST record cannot be split. So we wait to have enough
+ * space to proceed. It is small enough to be encoded in an empty buffer.
+ */
+static int fcgi_strm_send_begin_request(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       struct buffer outbuf;
+       struct buffer *mbuf;
+       struct fcgi_begin_request rec = { .role = FCGI_RESPONDER, .flags = 0};
+       int ret;
+
+       mbuf = br_tail(fconn->mbuf);
+  retry:
+       if (!fcgi_get_buf(fconn, mbuf)) {
+               fconn->flags |= FCGI_CF_MUX_MALLOC;
+               fconn->flags |= FCGI_CF_DEM_MROOM;
+               return 0;
+       }
+
+       while (1) {
+               outbuf = b_make(b_tail(mbuf), b_contig_space(mbuf), 0, 0);
+               if (outbuf.size >= 8 || !b_space_wraps(mbuf))
+                       break;
+         realign_again:
+               b_slow_realign(mbuf, trash.area, b_data(mbuf));
+       }
+
+       if (outbuf.size < 8)
+               goto full;
+
+       /* vsn: 1(FCGI_VERSION), type: (1)FCGI_BEGIN_REQUEST, id: fstrm->id,
+        *  len: 0x0008, padding: 0x00, rsv: 0x00 */
+       memcpy(outbuf.area, "\x01\x01\x00\x00\x00\x08\x00\x00", 8);
+       fcgi_set_record_id(outbuf.area, fstrm->id);
+       outbuf.data = 8;
+
+       if (fconn->flags & FCGI_CF_KEEP_CONN)
+               rec.flags |= FCGI_KEEP_CONN;
+       if (!fcgi_encode_begin_request(&outbuf, &rec))
+               goto full;
+
+       /* commit the record */
+       b_add(mbuf, outbuf.data);
+       fstrm->flags |= FCGI_SF_BEGIN_SENT;
+       fstrm->state = FCGI_SS_OPEN;
+
+       ret = 1;
+
+  end:
+       return ret;
+  full:
+       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+               goto retry;
+       fconn->flags |= FCGI_CF_MUX_MFULL;
+       fstrm->flags |= FCGI_SF_BLK_MROOM;
+       ret = 0;
+       goto end;
+}
+
+/* Sends an empty record of type <rtype>. It returns > 0 on success, 0 if it
+ * couldn't do anything. Empty record cannot be split. So we wait to have enough
+ * space to proceed. It is small enough to be encoded in an empty buffer.
+ */
+static int fcgi_strm_send_empty_record(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                      enum fcgi_record_type rtype)
+{
+       struct buffer outbuf;
+       struct buffer *mbuf;
+       int ret;
+
+       mbuf = br_tail(fconn->mbuf);
+  retry:
+       if (!fcgi_get_buf(fconn, mbuf)) {
+               fconn->flags |= FCGI_CF_MUX_MALLOC;
+               fconn->flags |= FCGI_CF_DEM_MROOM;
+               return 0;
+       }
+
+       while (1) {
+               outbuf = b_make(b_tail(mbuf), b_contig_space(mbuf), 0, 0);
+               if (outbuf.size >= 8 || !b_space_wraps(mbuf))
+                       break;
+         realign_again:
+               b_slow_realign(mbuf, trash.area, b_data(mbuf));
+       }
+
+       if (outbuf.size < 8)
+               goto full;
+
+       /* vsn: 1(FCGI_VERSION), type: rtype, id: fstrm->id,
+        *  len: 0x0000, padding: 0x00, rsv: 0x00 */
+       memcpy(outbuf.area, "\x01\x05\x00\x00\x00\x00\x00\x00", 8);
+       outbuf.area[1] = rtype;
+       fcgi_set_record_id(outbuf.area, fstrm->id);
+       outbuf.data = 8;
+
+       /* commit the record */
+       b_add(mbuf, outbuf.data);
+       ret = 1;
+
+  end:
+       return ret;
+  full:
+       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+               goto retry;
+       fconn->flags |= FCGI_CF_MUX_MFULL;
+       fstrm->flags |= FCGI_SF_BLK_MROOM;
+       ret = 0;
+       goto end;
+}
+
+
+/* Sends an empty PARAMS record. It relies on fcgi_strm_send_empty_record(). It
+ * marks the end of params.
+ */
+static int fcgi_strm_send_empty_params(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       return fcgi_strm_send_empty_record(fconn, fstrm, FCGI_PARAMS);
+}
+
+/* Sends an empty STDIN record. It relies on fcgi_strm_send_empty_record(). It
+ * marks the end of input. On success, all the request was successfully sent.
+ */
+static int fcgi_strm_send_empty_stdin(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       int ret;
+
+       ret = fcgi_strm_send_empty_record(fconn, fstrm, FCGI_STDIN);
+       if (ret)
+               fstrm->flags |= FCGI_SF_ES_SENT;
+       return ret;
+}
+
+/* Sends an ABORT_REQUEST record. It relies on fcgi_strm_send_empty_record(). It
+ * stops the request processing.
+ */
+static int fcgi_strm_send_abort(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       int ret;
+
+       ret = fcgi_strm_send_empty_record(fconn, fstrm, FCGI_ABORT_REQUEST);
+       if (ret)
+               fstrm->flags |= FCGI_SF_ABRT_SENT;
+       return ret;
+}
+
+/* Sends a PARAMS record. Returns > 0 on success, 0 if it couldn't do
+ * anything. If there are too much K/V params to be encoded in a PARAMS record,
+ * several records are sent. However, a K/V param cannot be split between 2
+ * records.
+ */
+static size_t fcgi_strm_send_params(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                   struct htx *htx)
+{
+       struct buffer outbuf;
+       struct buffer *mbuf;
+       struct htx_blk *blk;
+       struct htx_sl *sl = NULL;
+       struct fcgi_strm_params params;
+       size_t total = 0;
+
+       memset(&params, 0, sizeof(params));
+       params.p = get_trash_chunk();
+
+       mbuf = br_tail(fconn->mbuf);
+  retry:
+       if (!fcgi_get_buf(fconn, mbuf)) {
+               fconn->flags |= FCGI_CF_MUX_MALLOC;
+               fconn->flags |= FCGI_CF_DEM_MROOM;
+               return 0;
+       }
+
+       while (1) {
+               outbuf = b_make(b_tail(mbuf), b_contig_space(mbuf), 0, 0);
+               if (outbuf.size >= 8 || !b_space_wraps(mbuf))
+                       break;
+         realign_again:
+               b_slow_realign(mbuf, trash.area, b_data(mbuf));
+       }
+
+       if (outbuf.size < 8)
+               goto full;
+
+       /* vsn: 1(FCGI_VERSION), type: (4)FCGI_PARAMS, id: fstrm->id,
+        *  len: 0x0000 (fill later), padding: 0x00, rsv: 0x00 */
+       memcpy(outbuf.area, "\x01\x04\x00\x00\x00\x00\x00\x00", 8);
+       fcgi_set_record_id(outbuf.area, fstrm->id);
+       outbuf.data = 8;
+
+       blk = htx_get_head_blk(htx);
+       while (blk) {
+               enum htx_blk_type type;
+               uint32_t size = htx_get_blksz(blk);
+               struct fcgi_param p;
+
+               type = htx_get_blk_type(blk);
+               switch (type) {
+                       case HTX_BLK_REQ_SL:
+                               sl = htx_get_blk_ptr(htx, blk);
+                               if (sl->info.req.meth == HTTP_METH_HEAD)
+                                       fstrm->h1m.flags |= H1_MF_METH_HEAD;
+                               if (sl->flags & HTX_SL_F_VER_11)
+                                       fstrm->h1m.flags |= H1_MF_VER_11;
+                               break;
+
+                       case HTX_BLK_HDR:
+                               p.n = htx_get_blk_name(htx, blk);
+                               p.v = htx_get_blk_value(htx, blk);
+
+                               if (istmatch(p.n, ist(":fcgi-"))) {
+                                       p.n.ptr += 6;
+                                       p.n.len -= 6;
+                                       if (isteq(p.n, ist("gateway_interface")))
+                                               params.mask |= FCGI_SP_CGI_GATEWAY;
+                                       else if (isteq(p.n, ist("document_root"))) {
+                                               params.mask |= FCGI_SP_DOC_ROOT;
+                                               params.docroot = p.v;
+                                       }
+                                       else if (isteq(p.n, ist("script_name"))) {
+                                               params.mask |= FCGI_SP_SCRIPT_NAME;
+                                               params.scriptname = p.v;
+                                       }
+                                       else if (isteq(p.n, ist("path_info"))) {
+                                               params.mask |= FCGI_SP_PATH_INFO;
+                                               params.pathinfo = p.v;
+                                       }
+                                       else if (isteq(p.n, ist("request_uri"))) {
+                                               params.mask |= FCGI_SP_REQ_URI;
+                                               params.uri = p.v;
+                                       }
+                                       else if (isteq(p.n, ist("request_meth")))
+                                               params.mask |= FCGI_SP_REQ_METH;
+                                       else if (isteq(p.n, ist("query_string")))
+                                               params.mask |= FCGI_SP_REQ_QS;
+                                       else if (isteq(p.n, ist("server_name")))
+                                               params.mask |= FCGI_SP_SRV_NAME;
+                                       else if (isteq(p.n, ist("server_port")))
+                                               params.mask |= FCGI_SP_SRV_PORT;
+                                       else if (isteq(p.n, ist("server_protocol")))
+                                               params.mask |= FCGI_SP_SRV_PROTO;
+                                       else if (isteq(p.n, ist("remote_addr")))
+                                               params.mask |= FCGI_SP_REM_ADDR;
+                                       else if (isteq(p.n, ist("remote_port")))
+                                               params.mask |= FCGI_SP_REM_PORT;
+                                       else if (isteq(p.n, ist("script_filename")))
+                                               params.mask |= FCGI_SP_SCRIPT_FILE;
+                                       else if (isteq(p.n, ist("path_translated")))
+                                               params.mask |= FCGI_SP_PATH_TRANS;
+                                       else if (isteq(p.n, ist("https")))
+                                               params.mask |= FCGI_SP_HTTPS;
+                               }
+                               else if (isteq(p.n, ist("content-length"))) {
+                                       p.n = ist("CONTENT_LENGTH");
+                                       params.mask |= FCGI_SP_CONT_LEN;
+                               }
+                               else if (isteq(p.n, ist("content-type")))
+                                       p.n = ist("CONTENT_TYPE");
+                               else {
+                                       if (isteq(p.n, ist("host")))
+                                               params.srv_name = p.v;
+
+                                       memcpy(trash.area, "http_", 5);
+                                       memcpy(trash.area+5, p.n.ptr, p.n.len);
+                                       p.n = ist2(trash.area, p.n.len+5);
+                               }
+
+                               if (!fcgi_encode_param(&outbuf, &p)) {
+                                       if (b_space_wraps(mbuf))
+                                               goto realign_again;
+                                       if (outbuf.data == 8)
+                                               goto full;
+                                       goto done;
+                               }
+                               break;
+
+                       case HTX_BLK_EOH:
+                               goto done;
+
+                       default:
+                               break;
+               }
+               total += size;
+               blk = htx_remove_blk(htx, blk);
+       }
+
+  done:
+       if (!fcgi_set_default_param(fconn, fstrm, htx, sl, &params))
+               goto error;
+
+       if (!fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_CGI_GATEWAY) ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_DOC_ROOT)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_SCRIPT_NAME) ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_PATH_INFO)   ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_REQ_URI)     ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_REQ_METH)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_REQ_QS)      ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_SRV_NAME)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_SRV_PORT)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_SRV_PROTO)   ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_REM_ADDR)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_REM_PORT)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_SCRIPT_FILE) ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_PATH_TRANS)  ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_CONT_LEN)    ||
+           !fcgi_encode_default_param(fconn, fstrm, &params, &outbuf, FCGI_SP_HTTPS))
+               goto error;
+
+       /* update the record's size */
+       fcgi_set_record_size(outbuf.area, outbuf.data - 8);
+       b_add(mbuf, outbuf.data);
+
+  end:
+       return total;
+  full:
+       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+               goto retry;
+       fconn->flags |= FCGI_CF_MUX_MFULL;
+       fstrm->flags |= FCGI_SF_BLK_MROOM;
+       if (total)
+               goto error;
+       goto end;
+
+  error:
+       htx->flags |= HTX_FL_PROCESSING_ERROR;
+       fcgi_strm_error(fstrm);
+       goto end;
+}
+
+/* Sends a STDIN record. Returns > 0 on success, 0 if it couldn't do
+ * anything. STDIN records contain the request body.
+ */
+static size_t fcgi_strm_send_stdin(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                  struct htx *htx, size_t count, struct buffer *buf)
+{
+       struct buffer outbuf;
+       struct buffer *mbuf;
+       struct htx_blk *blk;
+       enum htx_blk_type type;
+       uint32_t size;
+       size_t total = 0;
+
+       if (!count)
+               goto end;
+
+       mbuf = br_tail(fconn->mbuf);
+  retry:
+       if (!fcgi_get_buf(fconn, mbuf)) {
+               fconn->flags |= FCGI_CF_MUX_MALLOC;
+               fconn->flags |= FCGI_CF_DEM_MROOM;
+               return 0;
+       }
+
+       /* Perform some optimizations to reduce the number of buffer copies.
+        * First, if the mux's buffer is empty and the htx area contains exactly
+        * one data block of the same size as the requested count, and this
+        * count fits within the record size, then it's possible to simply swap
+        * the caller's buffer with the mux's output buffer and adjust offsets
+        * and length to match the entire DATA HTX block in the middle. In this
+        * case we perform a true zero-copy operation from end-to-end. This is
+        * the situation that happens all the time with large files. Second, if
+        * this is not possible, but the mux's output buffer is empty, we still
+        * have an opportunity to avoid the copy to the intermediary buffer, by
+        * making the intermediary buffer's area point to the output buffer's
+        * area. In this case we want to skip the HTX header to make sure that
+        * copies remain aligned and that this operation remains possible all
+        * the time. This goes for headers, data blocks and any data extracted
+        * from the HTX blocks.
+        */
+       blk  = htx_get_head_blk(htx);
+       if (!blk)
+               goto end;
+       type = htx_get_blk_type(blk);
+       size = htx_get_blksz(blk);
+       if (unlikely(size == count && htx_nbblks(htx) == 1 && type == HTX_BLK_DATA)) {
+               void *old_area = mbuf->area;
+
+               if (b_data(mbuf)) {
+                       /* Too bad there are data left there. We're willing to memcpy/memmove
+                        * up to 1/4 of the buffer, which means that it's OK to copy a large
+                        * record into a buffer containing few data if it needs to be realigned,
+                        * and that it's also OK to copy few data without realigning. Otherwise
+                        * we'll pretend the mbuf is full and wait for it to become empty.
+                        */
+                       if (size + 8 <= b_room(mbuf) &&
+                           (b_data(mbuf) <= b_size(mbuf) / 4 ||
+                            (size <= b_size(mbuf) / 4 && size + 8 <= b_contig_space(mbuf))))
+                               goto copy;
+
+                       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+                               goto retry;
+
+                       fconn->flags |= FCGI_CF_MUX_MFULL;
+                       fstrm->flags |= FCGI_SF_BLK_MROOM;
+                       goto end;
+               }
+
+               /* map a FCGI record to the HTX block so that we can put the
+                * record header there.
+                */
+               *mbuf = b_make(buf->area, buf->size, sizeof(struct htx) + blk->addr - 8, size + 8);
+               outbuf.area = b_head(mbuf);
+
+               /* prepend a FCGI record header just before the DATA block */
+               memcpy(outbuf.area, "\x01\x05\x00\x00\x00\x00\x00\x00", 8);
+               fcgi_set_record_id(outbuf.area, fstrm->id);
+               fcgi_set_record_size(outbuf.area, size);
+
+               /* and exchange with our old area */
+               buf->area = old_area;
+               buf->data = buf->head = 0;
+               total += size;
+               goto end;
+       }
+
+  copy:
+       while (1) {
+               outbuf = b_make(b_tail(mbuf), b_contig_space(mbuf), 0, 0);
+               if (outbuf.size >= 8 || !b_space_wraps(mbuf))
+                       break;
+         realign_again:
+               b_slow_realign(mbuf, trash.area, b_data(mbuf));
+       }
+
+       if (outbuf.size < 8)
+               goto full;
+
+       /* vsn: 1(FCGI_VERSION), type: (5)FCGI_STDIN, id: fstrm->id,
+        *  len: 0x0000 (fill later), padding: 0x00, rsv: 0x00 */
+       memcpy(outbuf.area, "\x01\x05\x00\x00\x00\x00\x00\x00", 8);
+       fcgi_set_record_id(outbuf.area, fstrm->id);
+       outbuf.data = 8;
+
+       blk = htx_get_head_blk(htx);
+       while (blk && count) {
+               enum htx_blk_type type = htx_get_blk_type(blk);
+               uint32_t size = htx_get_blksz(blk);
+               struct ist v;
+
+               switch (type) {
+                       case HTX_BLK_DATA:
+                               v = htx_get_blk_value(htx, blk);
+                               if (v.len > count)
+                                       v.len = count;
+
+                               if (v.len > b_room(&outbuf)) {
+                                       /* It doesn't fit at once. If it at least fits once split and
+                                        * the amount of data to move is low, let's defragment the
+                                        * buffer now.
+                                        */
+                                       if (b_space_wraps(mbuf) &&
+                                           b_data(&outbuf) + v.len <= b_room(mbuf) &&
+                                           b_data(mbuf) <= MAX_DATA_REALIGN)
+                                               goto realign_again;
+                                       v.len = b_room(&outbuf);
+                               }
+                               if (!v.len || !chunk_memcat(&outbuf, v.ptr, v.len)) {
+                                       if (outbuf.data == 8)
+                                               goto full;
+                                       goto done;
+                               }
+                               if (v.len != size) {
+                                       total += v.len;
+                                       count -= v.len;
+                                       htx_cut_data_blk(htx, blk, v.len);
+                                       goto done;
+                               }
+                               break;
+
+                       case HTX_BLK_EOM:
+                               goto done;
+
+                       default:
+                               break;
+               }
+               total += size;
+               count -= size;
+               blk = htx_remove_blk(htx, blk);
+       }
+
+  done:
+       /* update the record's size */
+       fcgi_set_record_size(outbuf.area, outbuf.data - 8);
+       b_add(mbuf, outbuf.data);
+
+  end:
+       return total;
+  full:
+       if ((mbuf = br_tail_add(fconn->mbuf)) != NULL)
+               goto retry;
+       fconn->flags |= FCGI_CF_MUX_MFULL;
+       fstrm->flags |= FCGI_SF_BLK_MROOM;
+       goto end;
+}
+
+/* Processes a STDOUT record. Returns > 0 on success, 0 if it couldn't do
+ * anything. STDOUT records contain the entire response. All the content is
+ * copied in the stream's rxbuf. The parsing will be handled in fcgi_rcv_buf().
+ */
+static int fcgi_strm_handle_stdout(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       struct buffer *dbuf;
+       size_t ret;
+       size_t max;
+
+       dbuf = &fconn->dbuf;
+
+       /* Only padding remains */
+       if (fconn->state == FCGI_CS_RECORD_P)
+               goto end_transfer;
+
+       if (b_data(dbuf) < (fconn->drl + fconn->drp) &&
+           b_size(dbuf) > (fconn->drl + fconn->drp) &&
+           buf_room_for_htx_data(dbuf))
+               goto fail; // incomplete record
+
+       if (!fcgi_get_buf(fconn, &fstrm->rxbuf)) {
+               fconn->flags |= FCGI_CF_DEM_SALLOC;
+               return 0;
+       }
+
+       /*max = MIN(b_room(&fstrm->rxbuf), fconn->drl);*/
+       max = buf_room_for_htx_data(&fstrm->rxbuf);
+       if (!b_data(&fstrm->rxbuf))
+               fstrm->rxbuf.head = sizeof(struct htx);
+       if (max > fconn->drl)
+               max = fconn->drl;
+
+       ret = b_xfer(&fstrm->rxbuf, dbuf, max);
+       if (!ret)
+               goto fail;
+       fconn->drl -= ret;
+
+       if (!buf_room_for_htx_data(&fstrm->rxbuf))
+               fconn->flags |= FCGI_CF_DEM_SFULL;
+
+       if (fconn->drl)
+               goto fail;
+
+  end_transfer:
+       fconn->drl += fconn->drp;
+       fconn->drp = 0;
+       ret = MIN(b_data(&fconn->dbuf), fconn->drl);
+       b_del(&fconn->dbuf, ret);
+       fconn->drl -= ret;
+       if (fconn->drl)
+               goto fail;
+
+       fconn->state = FCGI_CS_RECORD_H;
+       return 1;
+  fail:
+       return 0;
+}
+
+
+/* Processes an empty STDOUT. Returns > 0 on success, 0 if it couldn't do
+ * anything. It only skip the padding in fact, there is no payload for such
+ * records. It makrs the end of the response.
+ */
+static int fcgi_strm_handle_empty_stdout(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       int ret;
+
+       fconn->state = FCGI_CS_RECORD_P;
+       fconn->drl += fconn->drp;
+       fconn->drp = 0;
+       ret = MIN(b_data(&fconn->dbuf), fconn->drl);
+       b_del(&fconn->dbuf, ret);
+       fconn->drl -= ret;
+       if (fconn->drl)
+               return 0;
+       fconn->state = FCGI_CS_RECORD_H;
+       fstrm->state |= FCGI_SF_ES_RCVD;
+       return 1;
+}
+
+/* Processes a STDERR record. Returns > 0 on success, 0 if it couldn't do
+ * anything.
+ */
+static int fcgi_strm_handle_stderr(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       struct buffer *dbuf;
+       struct buffer tag;
+       size_t ret;
+
+       dbuf = &fconn->dbuf;
+
+       /* Only padding remains */
+       if (fconn->state == FCGI_CS_RECORD_P)
+               goto end_transfer;
+
+       if (b_data(dbuf) < (fconn->drl + fconn->drp) &&
+           b_size(dbuf) > (fconn->drl + fconn->drp) &&
+           buf_room_for_htx_data(dbuf))
+               goto fail; // incomplete record
+
+       chunk_reset(&trash);
+       ret = b_xfer(&trash, dbuf, MIN(b_room(&trash), fconn->drl));
+       if (!ret)
+               goto fail;
+       fconn->drl -= ret;
+
+       trash.area[ret]   = '\n';
+       trash.area[ret+1] = '\0';
+       tag.area = fconn->app->name; tag.data = strlen(fconn->app->name);
+       app_log(&fconn->app->logsrvs, &tag, LOG_ERR, trash.area);
+
+       if (fconn->drl)
+               goto fail;
+
+  end_transfer:
+       fconn->drl += fconn->drp;
+       fconn->drp = 0;
+       ret = MIN(b_data(&fconn->dbuf), fconn->drl);
+       b_del(&fconn->dbuf, ret);
+       fconn->drl -= ret;
+       if (fconn->drl)
+               goto fail;
+       fconn->state = FCGI_CS_RECORD_H;
+       return 1;
+  fail:
+       return 0;
+}
+
+/* Processes an END_REQUEST record. Returns > 0 on success, 0 if it couldn't do
+ * anything. If the empty STDOUT record is not already received, this one marks
+ * the end of the response. It is highly unexpected, but if the record is larger
+ * than a buffer and cannot be decoded in one time, an error is triggered and
+ * the connection is closed. END_REQUEST record cannot be split.
+ */
+static int fcgi_strm_handle_end_request(struct fcgi_conn *fconn, struct fcgi_strm *fstrm)
+{
+       struct buffer inbuf;
+       struct buffer *dbuf;
+       struct fcgi_end_request endreq;
+
+       dbuf = &fconn->dbuf;
+
+       /* Record too large to be fully decoded */
+       if (b_size(dbuf) < (fconn->drl + fconn->drp))
+               goto fail;
+
+       /* process full record only */
+       if (b_data(dbuf) < (fconn->drl + fconn->drp))
+               return 0;
+
+       if (unlikely(b_contig_data(dbuf, b_head_ofs(dbuf)) < fconn->drl)) {
+               /* Realign the dmux buffer if the record wraps. It is unexpected
+                * at this stage because it should be the first record received
+                * from the FCGI application.
+                */
+               b_slow_realign(dbuf, trash.area, 0);
+       }
+
+       inbuf = b_make(b_head(dbuf), b_data(dbuf), 0, fconn->drl);
+
+       if (!fcgi_decode_end_request(&inbuf, 0, &endreq))
+               goto fail;
+
+       fstrm->flags |= FCGI_SF_ES_RCVD;
+       fstrm->proto_status = endreq.errcode;
+       fcgi_strm_close(fstrm);
+
+       b_del(&fconn->dbuf, fconn->drl + fconn->drp);
+       fconn->drl = 0;
+       fconn->drp = 0;
+       fconn->state = FCGI_CS_RECORD_H;
+       return 1;
+
+  fail:
+       fcgi_strm_error(fstrm);
+       return 0;
+}
+
+/* process Rx records to be demultiplexed */
+static void fcgi_process_demux(struct fcgi_conn *fconn)
+{
+       struct fcgi_strm *fstrm = NULL, *tmp_fstrm;
+       struct fcgi_header hdr;
+       int ret;
+
+       if (fconn->state == FCGI_CS_CLOSED)
+               return;
+
+       if (unlikely(fconn->state < FCGI_CS_RECORD_H)) {
+               if (fconn->state == FCGI_CS_INIT)
+                               return;
+               if (fconn->state == FCGI_CS_SETTINGS) {
+                       /* ensure that what is pending is a valid GET_VALUES_RESULT record. */
+                       ret = fcgi_decode_record_hdr(&fconn->dbuf, 0, &hdr);
+                       if (!ret)
+                               goto fail;
+                       b_del(&fconn->dbuf, ret);
+
+                       if (hdr.id || (hdr.type != FCGI_GET_VALUES_RESULT && hdr.type != FCGI_UNKNOWN_TYPE)) {
+                               fconn->state = FCGI_CS_CLOSED;
+                               goto fail;
+                       }
+                       goto new_record;
+               }
+       }
+
+       /* process as many incoming frames as possible below */
+       while (b_data(&fconn->dbuf)) {
+
+               if (fconn->state == FCGI_CS_CLOSED)
+                       break;
+
+               if (fconn->state == FCGI_CS_RECORD_H) {
+                       ret = fcgi_decode_record_hdr(&fconn->dbuf, 0, &hdr);
+                       if (!ret)
+                               break;
+                       b_del(&fconn->dbuf, ret);
+
+                 new_record:
+                       fconn->dsi = hdr.id;
+                       fconn->drt = hdr.type;
+                       fconn->drl = hdr.len;
+                       fconn->drp = hdr.padding;
+                       fconn->state = FCGI_CS_RECORD_D;
+               }
+
+               /* Only FCGI_CS_RECORD_D or FCGI_CS_RECORD_P */
+               tmp_fstrm = fcgi_conn_st_by_id(fconn, fconn->dsi);
+
+               if (tmp_fstrm != fstrm && fstrm && fstrm->cs &&
+                   (b_data(&fstrm->rxbuf) ||
+                    conn_xprt_read0_pending(fconn->conn) ||
+                    fstrm->state == FCGI_SS_CLOSED ||
+                    (fstrm->flags & FCGI_SF_ES_RCVD) ||
+                    (fstrm->cs->flags & (CS_FL_ERROR|CS_FL_ERR_PENDING|CS_FL_EOS)))) {
+                       /* we may have to signal the upper layers */
+                       fstrm->cs->flags |= CS_FL_RCV_MORE;
+                       fcgi_strm_notify_recv(fstrm);
+               }
+               fstrm = tmp_fstrm;
+
+               if (fstrm->state == FCGI_SS_CLOSED && fconn->dsi != 0) {
+                       /* ignore all record for closed streams */
+                       goto ignore_record;
+               }
+               if (fstrm->state == FCGI_SS_IDLE) {
+                       /* ignore all record for unknown streams */
+                       goto ignore_record;
+               }
+
+               switch (fconn->drt) {
+                       case FCGI_GET_VALUES_RESULT:
+                               ret = fcgi_conn_handle_values_result(fconn);
+                               break;
+
+                       case FCGI_STDOUT:
+                               if (fstrm->flags & FCGI_SF_ES_RCVD)
+                                       goto ignore_record;
+
+                               if (fconn->drl)
+                                       ret = fcgi_strm_handle_stdout(fconn, fstrm);
+                               else
+                                       ret = fcgi_strm_handle_empty_stdout(fconn, fstrm);
+                               break;
+
+                       case FCGI_STDERR:
+                               ret = fcgi_strm_handle_stderr(fconn, fstrm);
+                               break;
+
+                       case FCGI_END_REQUEST:
+                               ret = fcgi_strm_handle_end_request(fconn, fstrm);
+                               break;
+
+                       /* implement all extra frame types here */
+                       default:
+                         ignore_record:
+                               /* drop records that we ignore. They may be
+                                * larger than the buffer so we drain all of
+                                * their contents until we reach the end.
+                                */
+                               fconn->state = FCGI_CS_RECORD_P;
+                               fconn->drl += fconn->drp;
+                               fconn->drp = 0;
+                               ret = MIN(b_data(&fconn->dbuf), fconn->drl);
+                               b_del(&fconn->dbuf, ret);
+                               fconn->drl -= ret;
+                               ret = (fconn->drl == 0);
+               }
+
+               /* error or missing data condition met above ? */
+               if (ret <= 0)
+                       break;
+
+               if (fconn->state != FCGI_CS_RECORD_H && !(fconn->drl+fconn->drp))
+                       fconn->state = FCGI_CS_RECORD_H;
+       }
+
+ fail:
+       /* we can go here on missing data, blocked response or error */
+       if (fstrm && fstrm->cs &&
+           (b_data(&fstrm->rxbuf) ||
+            conn_xprt_read0_pending(fconn->conn) ||
+            fstrm->state == FCGI_SS_CLOSED ||
+            (fstrm->flags & FCGI_SF_ES_RCVD) ||
+            (fstrm->cs->flags & (CS_FL_ERROR|CS_FL_ERR_PENDING|CS_FL_EOS)))) {
+               /* we may have to signal the upper layers */
+               fstrm->cs->flags |= CS_FL_RCV_MORE;
+               fcgi_strm_notify_recv(fstrm);
+       }
+
+       fcgi_conn_restart_reading(fconn, 0);
+}
+
+/* process Tx records from streams to be multiplexed. Returns > 0 if it reached
+ * the end.
+ */
+static int fcgi_process_mux(struct fcgi_conn *fconn)
+{
+       struct fcgi_strm *fstrm, *fstrm_back;
+
+       if (unlikely(fconn->state < FCGI_CS_RECORD_H)) {
+               if (unlikely(fconn->state == FCGI_CS_INIT)) {
+                       if (!(fconn->flags & FCGI_CF_GET_VALUES)) {
+                               fconn->state = FCGI_CS_RECORD_H;
+                               fcgi_wake_unassigned_streams(fconn);
+                               goto mux;
+                       }
+                       if (unlikely(!fcgi_conn_send_get_values(fconn)))
+                               goto fail;
+                       fconn->state = FCGI_CS_SETTINGS;
+               }
+               /* need to wait for the other side */
+               if (fconn->state < FCGI_CS_RECORD_H)
+                       return 1;
+       }
+
+  mux:
+       list_for_each_entry_safe(fstrm, fstrm_back, &fconn->send_list, send_list) {
+               if (fconn->state == FCGI_CS_CLOSED || fconn->flags & FCGI_CF_MUX_BLOCK_ANY)
+                       break;
+
+               if (LIST_ADDED(&fstrm->sending_list))
+                       continue;
+
+               /* For some reason, the upper layer failed to subsribe again,
+                * so remove it from the send_list
+                */
+               if (!fstrm->send_wait) {
+                       LIST_DEL_INIT(&fstrm->send_list);
+                       continue;
+               }
+               fstrm->flags &= ~FCGI_SF_BLK_ANY;
+               fstrm->send_wait->events &= ~SUB_RETRY_SEND;
+               LIST_ADDQ(&fconn->sending_list, &fstrm->sending_list);
+               tasklet_wakeup(fstrm->send_wait->tasklet);
+       }
+
+ fail:
+       if (fconn->state == FCGI_CS_CLOSED) {
+               if (fconn->stream_cnt - fconn->nb_reserved > 0) {
+                       fcgi_conn_send_aborts(fconn);
+                       if (fconn->flags & FCGI_CF_MUX_BLOCK_ANY)
+                               return 0;
+               }
+       }
+       return 1;
+}
+
+
+/* Attempt to read data, and subscribe if none available.
+ * The function returns 1 if data has been received, otherwise zero.
+ */
+static int fcgi_recv(struct fcgi_conn *fconn)
+{
+       struct connection *conn = fconn->conn;
+       struct buffer *buf;
+       int max;
+       size_t ret;
+
+       if (fconn->wait_event.events & SUB_RETRY_RECV)
+               return (b_data(&fconn->dbuf));
+
+       if (!fcgi_recv_allowed(fconn))
+               return 1;
+
+       buf = fcgi_get_buf(fconn, &fconn->dbuf);
+       if (!buf) {
+               fconn->flags |= FCGI_CF_DEM_DALLOC;
+               return 0;
+       }
+
+       b_realign_if_empty(buf);
+       if (!b_data(buf)) {
+               /* try to pre-align the buffer like the
+                * rxbufs will be to optimize memory copies. We'll make
+                * sure that the frame header lands at the end of the
+                * HTX block to alias it upon recv. We cannot use the
+                * head because rcv_buf() will realign the buffer if
+                * it's empty. Thus we cheat and pretend we already
+                * have a few bytes there.
+                */
+               max = buf_room_for_htx_data(buf) + (fconn->state == FCGI_CS_RECORD_H ? 8 : 0);
+               buf->head = sizeof(struct htx) - (fconn->state == FCGI_CS_RECORD_H ? 8 : 0);
+       }
+       else
+               max = buf_room_for_htx_data(buf);
+
+       ret = max ? conn->xprt->rcv_buf(conn, conn->xprt_ctx, buf, max, 0) : 0;
+
+       if (max && !ret && fcgi_recv_allowed(fconn))
+               conn->xprt->subscribe(conn, conn->xprt_ctx, SUB_RETRY_RECV, &fconn->wait_event);
+
+       if (!b_data(buf)) {
+               fcgi_release_buf(fconn, &fconn->dbuf);
+               return (conn->flags & CO_FL_ERROR || conn_xprt_read0_pending(conn));
+       }
+
+       if (ret == max)
+               fconn->flags |= FCGI_CF_DEM_DFULL;
+
+       return !!ret || (conn->flags & CO_FL_ERROR) || conn_xprt_read0_pending(conn);
+}
+
+
+/* Try to send data if possible.
+ * The function returns 1 if data have been sent, otherwise zero.
+ */
+static int fcgi_send(struct fcgi_conn *fconn)
+{
+       struct connection *conn = fconn->conn;
+       int done;
+       int sent = 0;
+
+       if (conn->flags & CO_FL_ERROR)
+               return 1;
+
+
+       if (conn->flags & (CO_FL_HANDSHAKE|CO_FL_WAIT_L4_CONN|CO_FL_WAIT_L6_CONN)) {
+               /* a handshake was requested */
+               goto schedule;
+       }
+
+       /* This loop is quite simple : it tries to fill as much as it can from
+        * pending streams into the existing buffer until it's reportedly full
+        * or the end of send requests is reached. Then it tries to send this
+        * buffer's contents out, marks it not full if at least one byte could
+        * be sent, and tries again.
+        *
+        * The snd_buf() function normally takes a "flags" argument which may
+        * be made of a combination of CO_SFL_MSG_MORE to indicate that more
+        * data immediately comes and CO_SFL_STREAMER to indicate that the
+        * connection is streaming lots of data (used to increase TLS record
+        * size at the expense of latency). The former can be sent any time
+        * there's a buffer full flag, as it indicates at least one stream
+        * attempted to send and failed so there are pending data. An
+        * alternative would be to set it as long as there's an active stream
+        * but that would be problematic for ACKs until we have an absolute
+        * guarantee that all waiters have at least one byte to send. The
+        * latter should possibly not be set for now.
+        */
+
+       done = 0;
+       while (!done) {
+               unsigned int flags = 0;
+               unsigned int released = 0;
+               struct buffer *buf;
+
+               /* fill as much as we can into the current buffer */
+               while (((fconn->flags & (FCGI_CF_MUX_MFULL|FCGI_CF_MUX_MALLOC)) == 0) && !done)
+                       done = fcgi_process_mux(fconn);
+
+               if (fconn->flags & FCGI_CF_MUX_MALLOC)
+                       done = 1; // we won't go further without extra buffers
+
+               if (conn->flags & CO_FL_ERROR)
+                       break;
+
+               if (fconn->flags & (FCGI_CF_MUX_MFULL | FCGI_CF_DEM_MROOM))
+                       flags |= CO_SFL_MSG_MORE;
+
+               for (buf = br_head(fconn->mbuf); b_size(buf); buf = br_del_head(fconn->mbuf)) {
+                       if (b_data(buf)) {
+                               int ret;
+
+                               ret = conn->xprt->snd_buf(conn, conn->xprt_ctx, buf, b_data(buf), flags);
+                               if (!ret) {
+                                       done = 1;
+                                       break;
+                               }
+                               sent = 1;
+                               b_del(buf, ret);
+                               if (b_data(buf)) {
+                                       done = 1;
+                                       break;
+                               }
+                       }
+                       b_free(buf);
+                       released++;
+               }
+
+               if (released)
+                       offer_buffers(NULL, tasks_run_queue);
+
+               /* wrote at least one byte, the buffer is not full anymore */
+               fconn->flags &= ~(FCGI_CF_MUX_MFULL | FCGI_CF_DEM_MROOM);
+       }
+
+       if (conn->flags & CO_FL_SOCK_WR_SH) {
+               /* output closed, nothing to send, clear the buffer to release it */
+               b_reset(br_tail(fconn->mbuf));
+       }
+       /* We're not full anymore, so we can wake any task that are waiting
+        * for us.
+        */
+       if (!(fconn->flags & (FCGI_CF_MUX_MFULL | FCGI_CF_DEM_MROOM))) {
+               struct fcgi_strm *fstrm;
+
+               list_for_each_entry(fstrm, &fconn->send_list, send_list) {
+                       if (fconn->state == FCGI_CS_CLOSED || fconn->flags & FCGI_CF_MUX_BLOCK_ANY)
+                               break;
+
+                       if (LIST_ADDED(&fstrm->sending_list))
+                               continue;
+
+                       /* For some reason, the upper layer failed to subsribe again,
+                        * so remove it from the send_list
+                        */
+                       if (!fstrm->send_wait) {
+                               LIST_DEL_INIT(&fstrm->send_list);
+                               continue;
+                       }
+                       fstrm->flags &= ~FCGI_SF_BLK_ANY;
+                       fstrm->send_wait->events &= ~SUB_RETRY_SEND;
+                       tasklet_wakeup(fstrm->send_wait->tasklet);
+                       LIST_ADDQ(&fconn->sending_list, &fstrm->sending_list);
+               }
+       }
+       /* We're done, no more to send */
+       if (!br_data(fconn->mbuf))
+               return sent;
+schedule:
+       if (!(conn->flags & CO_FL_ERROR) && !(fconn->wait_event.events & SUB_RETRY_SEND))
+               conn->xprt->subscribe(conn, conn->xprt_ctx, SUB_RETRY_SEND, &fconn->wait_event);
+
+       return sent;
+}
+
+/* this is the tasklet referenced in fconn->wait_event.tasklet */
+static struct task *fcgi_io_cb(struct task *t, void *ctx, unsigned short status)
+{
+       struct fcgi_conn *fconn = ctx;
+       int ret = 0;
+
+       if (!(fconn->wait_event.events & SUB_RETRY_SEND))
+               ret = fcgi_send(fconn);
+       if (!(fconn->wait_event.events & SUB_RETRY_RECV))
+               ret |= fcgi_recv(fconn);
+       if (ret || b_data(&fconn->dbuf))
+               fcgi_process(fconn);
+       return NULL;
+}
+
+/* callback called on any event by the connection handler.
+ * It applies changes and returns zero, or < 0 if it wants immediate
+ * destruction of the connection (which normally doesn not happen in FCGI).
+ */
+static int fcgi_process(struct fcgi_conn *fconn)
+{
+       struct connection *conn = fconn->conn;
+
+       if (b_data(&fconn->dbuf) && !(fconn->flags & FCGI_CF_DEM_BLOCK_ANY)) {
+               fcgi_process_demux(fconn);
+
+               if (fconn->state == FCGI_CS_CLOSED || conn->flags & CO_FL_ERROR)
+                       b_reset(&fconn->dbuf);
+
+               if (buf_room_for_htx_data(&fconn->dbuf))
+                       fconn->flags &= ~FCGI_CF_DEM_DFULL;
+       }
+       fcgi_send(fconn);
+
+       if (unlikely(fconn->proxy->state == PR_STSTOPPED)) {
+               /* frontend is stopping, reload likely in progress, let's try
+                * to announce a graceful shutdown if not yet done. We don't
+                * care if it fails, it will be tried again later.
+                */
+               if (!(fconn->flags & (FCGI_CF_ABRTS_SENT|FCGI_CF_ABRTS_FAILED))) {
+                       if (fconn->stream_cnt - fconn->nb_reserved > 0)
+                               fcgi_conn_send_aborts(fconn);
+               }
+       }
+
+       /*
+        * If we received early data, and the handshake is done, wake
+        * any stream that was waiting for it.
+        */
+       if (!(fconn->flags & FCGI_CF_WAIT_FOR_HS) &&
+           (conn->flags & (CO_FL_EARLY_SSL_HS | CO_FL_HANDSHAKE | CO_FL_EARLY_DATA)) == CO_FL_EARLY_DATA) {
+               struct eb32_node *node;
+               struct fcgi_strm *fstrm;
+
+               fconn->flags |= FCGI_CF_WAIT_FOR_HS;
+               node = eb32_lookup_ge(&fconn->streams_by_id, 1);
+
+               while (node) {
+                       fstrm = container_of(node, struct fcgi_strm, by_id);
+                       if (fstrm->cs && fstrm->cs->flags & CS_FL_WAIT_FOR_HS)
+                               fcgi_strm_notify_recv(fstrm);
+                       node = eb32_next(node);
+               }
+       }
+
+       if ((conn->flags & CO_FL_ERROR) || conn_xprt_read0_pending(conn) ||
+           fconn->state == FCGI_CS_CLOSED || (fconn->flags & FCGI_CF_ABRTS_FAILED) ||
+           eb_is_empty(&fconn->streams_by_id)) {
+               fcgi_wake_some_streams(fconn, 0);
+
+               if (eb_is_empty(&fconn->streams_by_id)) {
+                       /* no more stream, kill the connection now */
+                       fcgi_release(fconn);
+                       return -1;
+               }
+       }
+
+       if (!b_data(&fconn->dbuf))
+               fcgi_release_buf(fconn, &fconn->dbuf);
+
+       if ((conn->flags & CO_FL_SOCK_WR_SH) ||
+           fconn->state == FCGI_CS_CLOSED  || (fconn->flags & FCGI_CF_ABRTS_FAILED) ||
+           (!br_data(fconn->mbuf) && ((fconn->flags & FCGI_CF_MUX_BLOCK_ANY) || LIST_ISEMPTY(&fconn->send_list))))
+               fcgi_release_mbuf(fconn);
+
+       if (fconn->task) {
+               fconn->task->expire = tick_add(now_ms, (fconn->state == FCGI_CS_CLOSED ? fconn->shut_timeout : fconn->timeout));
+               task_queue(fconn->task);
+       }
+
+       fcgi_send(fconn);
+       return 0;
+}
+
+
+/* wake-up function called by the connection layer (mux_ops.wake) */
+static int fcgi_wake(struct connection *conn)
+{
+       struct fcgi_conn *fconn = conn->ctx;
+
+       return (fcgi_process(fconn));
+}
+
+/* Connection timeout management. The principle is that if there's no receipt
+ * nor sending for a certain amount of time, the connection is closed. If the
+ * MUX buffer still has lying data or is not allocatable, the connection is
+ * immediately killed. If it's allocatable and empty, we attempt to send a
+ * GOAWAY frame.
+ */
+static struct task *fcgi_timeout_task(struct task *t, void *context, unsigned short state)
+{
+       struct fcgi_conn *fconn = context;
+       int expired = tick_is_expired(t->expire, now_ms);
+
+       if (!expired && fconn)
+               return t;
+
+       task_destroy(t);
+
+       if (!fconn) {
+               /* resources were already deleted */
+               return NULL;
+       }
+
+       fconn->task = NULL;
+       fconn->state = FCGI_CS_CLOSED;
+       fcgi_wake_some_streams(fconn, 0);
+
+       if (br_data(fconn->mbuf)) {
+               /* don't even try to send aborts, the buffer is stuck */
+               fconn->flags |= FCGI_CF_ABRTS_FAILED;
+               goto end;
+       }
+
+       /* try to send but no need to insist */
+       if (!fcgi_conn_send_aborts(fconn))
+               fconn->flags |= FCGI_CF_ABRTS_FAILED;
+
+       if (br_data(fconn->mbuf) && !(fconn->flags & FCGI_CF_ABRTS_FAILED) &&
+           conn_xprt_ready(fconn->conn)) {
+               unsigned int released = 0;
+               struct buffer *buf;
+
+               for (buf = br_head(fconn->mbuf); b_size(buf); buf = br_del_head(fconn->mbuf)) {
+                       if (b_data(buf)) {
+                               int ret = fconn->conn->xprt->snd_buf(fconn->conn, fconn->conn->xprt_ctx,
+                                                                    buf, b_data(buf), 0);
+                               if (!ret)
+                                       break;
+                               b_del(buf, ret);
+                               if (b_data(buf))
+                                       break;
+                               b_free(buf);
+                               released++;
+                       }
+               }
+
+               if (released)
+                       offer_buffers(NULL, tasks_run_queue);
+       }
+
+  end:
+       /* either we can release everything now or it will be done later once
+        * the last stream closes.
+        */
+       if (eb_is_empty(&fconn->streams_by_id))
+               fcgi_release(fconn);
+
+       return NULL;
+}
+
+
+/*******************************************/
+/* functions below are used by the streams */
+/*******************************************/
+
+/* Append the description of what is present in error snapshot <es> into <out>.
+ * The description must be small enough to always fit in a buffer. The output
+ * buffer may be the trash so the trash must not be used inside this function.
+ */
+static void fcgi_show_error_snapshot(struct buffer *out, const struct error_snapshot *es)
+{
+       chunk_appendf(out,
+                     "  FCGI connection flags 0x%08x, FCGI stream flags 0x%08x\n"
+                     "  H1 msg state %s(%d), H1 msg flags 0x%08x\n"
+                     "  H1 chunk len %lld bytes, H1 body len %lld bytes :\n",
+                     es->ctx.h1.c_flags, es->ctx.h1.s_flags,
+                     h1m_state_str(es->ctx.h1.state), es->ctx.h1.state,
+                     es->ctx.h1.m_flags, es->ctx.h1.m_clen, es->ctx.h1.m_blen);
+}
+/*
+ * Capture a bad response and archive it in the proxy's structure.  By default
+ * it tries to report the error position as h1m->err_pos. However if this one is
+ * not set, it will then report h1m->next, which is the last known parsing
+ * point. The function is able to deal with wrapping buffers. It always displays
+ * buffers as a contiguous area starting at buf->p. The direction is determined
+ * thanks to the h1m's flags.
+ */
+static void fcgi_strm_capture_bad_message(struct fcgi_conn *fconn, struct fcgi_strm *fstrm,
+                                         struct h1m *h1m, struct buffer *buf)
+{
+       struct session *sess = fstrm->sess;
+       struct proxy *proxy = fconn->proxy;
+       struct proxy *other_end = sess->fe;
+       union error_snapshot_ctx ctx;
+
+       /* http-specific part now */
+       ctx.h1.state   = h1m->state;
+       ctx.h1.c_flags = fconn->flags;
+       ctx.h1.s_flags = fstrm->flags;
+       ctx.h1.m_flags = h1m->flags;
+       ctx.h1.m_clen  = h1m->curr_len;
+       ctx.h1.m_blen  = h1m->body_len;
+
+       proxy_capture_error(proxy, 1, other_end, fconn->conn->target, sess, buf, 0, 0,
+                           (h1m->err_pos >= 0) ? h1m->err_pos : h1m->next,
+                           &ctx, fcgi_show_error_snapshot);
+}
+
+static size_t fcgi_strm_parse_headers(struct fcgi_strm *fstrm, struct h1m *h1m, struct htx *htx,
+                                     struct buffer *buf, size_t *ofs, size_t max)
+{
+       int ret;
+
+       ret = h1_parse_msg_hdrs(h1m, NULL, htx, buf, *ofs, max);
+       if (!ret) {
+               if (htx->flags & HTX_FL_PARSING_ERROR) {
+                       fcgi_strm_error(fstrm);
+                       fcgi_strm_capture_bad_message(fstrm->fconn, fstrm, h1m, buf);
+               }
+               goto end;
+       }
+
+       *ofs += ret;
+  end:
+       return ret;
+
+}
+
+static size_t fcgi_strm_parse_data(struct fcgi_strm *fstrm, struct h1m *h1m, struct htx *htx,
+                                  struct buffer *buf, size_t *ofs, size_t max, struct buffer *htxbuf)
+{
+       int ret;
+
+       ret = h1_parse_msg_data(h1m, htx, buf, *ofs, max, htxbuf);
+       if (ret <= 0) {
+               if (htx->flags & HTX_FL_PARSING_ERROR) {
+                       fcgi_strm_error(fstrm);
+                       fcgi_strm_capture_bad_message(fstrm->fconn, fstrm, h1m, buf);
+               }
+               goto end;
+       }
+       *ofs += ret;
+  end:
+       return ret;
+}
+
+static size_t fcgi_strm_parse_trailers(struct fcgi_strm *fstrm, struct h1m *h1m, struct htx *htx,
+                                      struct buffer *buf, size_t *ofs, size_t max)
+{
+       int ret;
+
+       ret = h1_parse_msg_tlrs(h1m, htx, buf, *ofs, max);
+       if (ret <= 0) {
+               if (htx->flags & HTX_FL_PARSING_ERROR) {
+                       fcgi_strm_error(fstrm);
+                       fcgi_strm_capture_bad_message(fstrm->fconn, fstrm, h1m, buf);
+               }
+               goto end;
+       }
+       *ofs += ret;
+       fstrm->flags |= FCGI_SF_HAVE_I_TLR;
+  end:
+       return ret;
+}
+
+static size_t fcgi_strm_add_eom(struct fcgi_strm *fstrm, struct h1m *h1m, struct htx *htx,
+                               size_t max)
+{
+       if (max < sizeof(struct htx_blk) + 1 || !htx_add_endof(htx, HTX_BLK_EOM))
+               return 0;
+
+       h1m->state = H1_MSG_DONE;
+       return (sizeof(struct htx_blk) + 1);
+}
+
+static size_t fcgi_strm_parse_response(struct fcgi_strm *fstrm, struct buffer *buf, size_t count)
+{
+       struct htx *htx;
+       struct h1m *h1m = &fstrm->h1m;
+       size_t ret, data, total = 0;
+
+       htx = htx_from_buf(buf);
+       data = htx->data;
+       if (fstrm->state == FCGI_SS_ERROR)
+               goto end;
+
+       do {
+               size_t used = htx_used_space(htx);
+
+               if (h1m->state <= H1_MSG_LAST_LF) {
+                       ret = fcgi_strm_parse_headers(fstrm, h1m, htx, &fstrm->rxbuf, &total, count);
+                       if (!ret)
+                               break;
+                       if ((h1m->flags & (H1_MF_VER_11|H1_MF_XFER_LEN)) == H1_MF_VER_11) {
+                               struct htx_blk *blk = htx_get_head_blk(htx);
+                               struct htx_sl *sl;
+
+                               if (!blk)
+                                       break;
+                               sl = htx_get_blk_ptr(htx, blk);
+                               sl->flags |= HTX_SL_F_XFER_LEN;
+                               htx->extra = 0;
+                       }
+               }
+               else if (h1m->state < H1_MSG_TRAILERS) {
+                       ret = fcgi_strm_parse_data(fstrm, h1m, htx, &fstrm->rxbuf, &total, count, buf);
+                       htx = htx_from_buf(buf);
+                       if (!ret)
+                               break;
+               }
+               else if (h1m->state == H1_MSG_TRAILERS) {
+                       if (!(fstrm->flags & FCGI_SF_HAVE_I_TLR)) {
+                               ret = fcgi_strm_parse_trailers(fstrm, h1m, htx, &fstrm->rxbuf, &total, count);
+                               if (!ret)
+                                       break;
+                       }
+                       else if (!fcgi_strm_add_eom(fstrm, h1m, htx, count))
+                               break;
+               }
+               else if (h1m->state == H1_MSG_DONE) {
+                       if (b_data(&fstrm->rxbuf) > total) {
+                               htx->flags |= HTX_FL_PARSING_ERROR;
+                               fcgi_strm_error(fstrm);
+                       }
+                       break;
+               }
+               else if (h1m->state == H1_MSG_TUNNEL) {
+                       ret = fcgi_strm_parse_data(fstrm, h1m, htx, &fstrm->rxbuf, &total, count, buf);
+                       htx = htx_from_buf(buf);
+                       if (fstrm->state != FCGI_SS_ERROR &&
+                           (fstrm->flags & FCGI_SF_ES_RCVD) && b_data(&fstrm->rxbuf) == total) {
+                               if ((h1m->flags & (H1_MF_VER_11|H1_MF_XFER_LEN)) == H1_MF_VER_11) {
+                                       if (!fcgi_strm_add_eom(fstrm, h1m, htx, count))
+                                               break;
+                               }
+                               else {
+                                       h1m->state = H1_MSG_DONE;
+                                       break;
+                               }
+                       }
+                       if (!ret)
+                               break;
+               }
+               else {
+                       htx->flags |= HTX_FL_PROCESSING_ERROR;
+                       fcgi_strm_error(fstrm);
+                       break;
+               }
+
+               count -= htx_used_space(htx) - used;
+       } while (fstrm->state != FCGI_SS_ERROR/* /\*fstrm->state == FCGI_SS_OPEN && *\/count */);
+
+       if (fstrm->state == FCGI_SS_ERROR) {
+               b_reset(&fstrm->rxbuf);
+               htx_to_buf(htx, buf);
+               return 0;
+       }
+
+       b_del(&fstrm->rxbuf, total);
+
+  end:
+       htx_to_buf(htx, buf);
+       ret = htx->data - data;
+       return ret;
+}
+
+/*
+ * Attach a new stream to a connection
+ * (Used for outgoing connections)
+ */
+static struct conn_stream *fcgi_attach(struct connection *conn, struct session *sess)
+{
+       struct conn_stream *cs;
+       struct fcgi_strm *fstrm;
+       struct fcgi_conn *fconn = conn->ctx;
+
+       cs = cs_new(conn);
+       if (!cs)
+               return NULL;
+       fstrm = fcgi_conn_stream_new(fconn, cs, sess);
+       if (!fstrm) {
+               cs_free(cs);
+               return NULL;
+       }
+       return cs;
+}
+
+/* Retrieves the first valid conn_stream from this connection, or returns NULL.
+ * We have to scan because we may have some orphan streams. It might be
+ * beneficial to scan backwards from the end to reduce the likeliness to find
+ * orphans.
+ */
+static const struct conn_stream *fcgi_get_first_cs(const struct connection *conn)
+{
+       struct fcgi_conn *fconn = conn->ctx;
+       struct fcgi_strm *fstrm;
+       struct eb32_node *node;
+
+       node = eb32_first(&fconn->streams_by_id);
+       while (node) {
+               fstrm = container_of(node, struct fcgi_strm, by_id);
+               if (fstrm->cs)
+                       return fstrm->cs;
+               node = eb32_next(node);
+       }
+       return NULL;
+}
+
+/*
+ * Destroy the mux and the associated connection, if it is no longer used
+ */
+static void fcgi_destroy(void *ctx)
+{
+       struct fcgi_conn *fconn = ctx;
+
+       if (eb_is_empty(&fconn->streams_by_id) || !fconn->conn || fconn->conn->ctx != fconn)
+               fcgi_release(fconn);
+}
+
+/*
+ * Detach the stream from the connection and possibly release the connection.
+ */
+static void fcgi_detach(struct conn_stream *cs)
+{
+       struct fcgi_strm *fstrm = cs->ctx;
+       struct fcgi_conn *fconn;
+       struct session *sess;
+
+       cs->ctx = NULL;
+       if (!fstrm)
+               return;
+
+       /* The stream is about to die, so no need to attempt to run its task */
+       if (LIST_ADDED(&fstrm->sending_list)  &&
+           fstrm->send_wait != &fstrm->wait_event) {
+               tasklet_remove_from_tasklet_list(fstrm->send_wait->tasklet);
+               LIST_DEL_INIT(&fstrm->sending_list);
+               /*
+                * At this point, the stream_interface is supposed to have called
+                * fcgi_unsubscribe(), so the only way there's still a
+                * subscription that came from the stream_interface (as we
+                * can subscribe ourself, in fcgi_do_shutw() and fcgi_do_shutr(),
+                * without the stream_interface involved) is that we subscribed
+                * for sending, we woke the tasklet up and removed the
+                * SUB_RETRY_SEND flag, so the stream_interface would not
+                * know it has to unsubscribe for send, but the tasklet hasn't
+                * run yet. Make sure to handle that by explicitely setting
+                * send_wait to NULL, as nothing else will do it for us.
+                */
+               fstrm->send_wait = NULL;
+       }
+
+       sess = fstrm->sess;
+       fconn = fstrm->fconn;
+       fstrm->cs = NULL;
+       fconn->nb_cs--;
+
+       if (fstrm->proto_status == FCGI_PS_CANT_MPX_CONN) {
+               fconn->flags &= ~FCGI_CF_MPXS_CONNS;
+               fconn->streams_limit = 1;
+       }
+       else if (fstrm->proto_status == FCGI_PS_OVERLOADED ||
+                fstrm->proto_status == FCGI_PS_UNKNOWN_ROLE) {
+               fconn->flags &= ~FCGI_CF_KEEP_CONN;
+               fconn->state = FCGI_CS_CLOSED;
+       }
+
+       /* this stream may be blocked waiting for some data to leave, so orphan
+        * it in this case.
+        */
+       if (!(cs->conn->flags & CO_FL_ERROR) &&
+           (fconn->state != FCGI_CS_CLOSED) &&
+           (fstrm->flags & (FCGI_SF_BLK_MBUSY|FCGI_SF_BLK_MROOM)) && (fstrm->send_wait || fstrm->recv_wait))
+               return;
+
+       if ((fconn->flags & FCGI_CF_DEM_BLOCK_ANY && fstrm->id == fconn->dsi)) {
+               /* unblock the connection if it was blocked on this stream. */
+               fconn->flags &= ~FCGI_CF_DEM_BLOCK_ANY;
+               fcgi_conn_restart_reading(fconn, 1);
+       }
+
+       fcgi_strm_destroy(fstrm);
+
+       if (!(fconn->conn->flags & (CO_FL_ERROR|CO_FL_SOCK_RD_SH|CO_FL_SOCK_WR_SH)) &&
+           !(fconn->flags & FCGI_CF_KEEP_CONN)) {
+               if (!fconn->conn->owner) {
+                       fconn->conn->owner = sess;
+                       if (!session_add_conn(sess, fconn->conn, fconn->conn->target)) {
+                               fconn->conn->owner = NULL;
+                               if (eb_is_empty(&fconn->streams_by_id)) {
+                                       if (!srv_add_to_idle_list(objt_server(fconn->conn->target), fconn->conn))
+                                               /* The server doesn't want it, let's kill the connection right away */
+                                               fconn->conn->mux->destroy(fconn->conn);
+                                       return;
+                               }
+                       }
+               }
+               if (eb_is_empty(&fconn->streams_by_id)) {
+                       if (session_check_idle_conn(fconn->conn->owner, fconn->conn) != 0)
+                               /* At this point either the connection is destroyed,
+                                  or it's been added to the server idle list, just stop */
+                               return;
+               }
+               /* Never ever allow to reuse a connection from a non-reuse backend */
+               if ((fconn->proxy->options & PR_O_REUSE_MASK) == PR_O_REUSE_NEVR)
+                       fconn->conn->flags |= CO_FL_PRIVATE;
+               if (!LIST_ADDED(&fconn->conn->list) && fconn->nb_streams < fconn->streams_limit) {
+                       struct server *srv = objt_server(fconn->conn->target);
+
+                       if (srv) {
+                               if (fconn->conn->flags & CO_FL_PRIVATE)
+                                       LIST_ADD(&srv->priv_conns[tid], &fconn->conn->list);
+                               else
+                                       LIST_ADD(&srv->idle_conns[tid], &fconn->conn->list);
+                       }
+               }
+       }
+
+       /* We don't want to close right now unless we're removing the last
+        * stream, and either the connection is in error, or it reached the ID
+        * already specified in a GOAWAY frame received or sent (as seen by
+        * last_sid >= 0).
+        */
+       if (fcgi_conn_is_dead(fconn)) {
+               /* no more stream will come, kill it now */
+               fcgi_release(fconn);
+       }
+       else if (fconn->task) {
+               fconn->task->expire = tick_add(now_ms, (fconn->state == FCGI_CS_CLOSED ? fconn->shut_timeout : fconn->timeout));
+               task_queue(fconn->task);
+       }
+}
+
+
+/* Performs a synchronous or asynchronous shutr(). */
+static void fcgi_do_shutr(struct fcgi_strm *fstrm)
+{
+       struct fcgi_conn *fconn = fstrm->fconn;
+       struct wait_event *sw = &fstrm->wait_event;
+
+       if (fstrm->state == FCGI_SS_CLOSED)
+               goto done;
+
+       /* a connstream may require us to immediately kill the whole connection
+        * for example because of a "tcp-request content reject" rule that is
+        * normally used to limit abuse.
+        */
+       if ((fstrm->flags & FCGI_SF_KILL_CONN) &&
+           !(fconn->flags & (FCGI_CF_ABRTS_SENT|FCGI_CF_ABRTS_FAILED)))
+               fconn->state = FCGI_CS_CLOSED;
+       else if (fstrm->flags & FCGI_SF_BEGIN_SENT) {
+               if (!(fstrm->flags & (FCGI_SF_ES_SENT|FCGI_SF_ABRT_SENT)) &&
+                   !fcgi_strm_send_abort(fconn, fstrm))
+                       goto add_to_list;
+       }
+
+       fcgi_strm_close(fstrm);
+
+       if (!(fconn->wait_event.events & SUB_RETRY_SEND))
+               tasklet_wakeup(fconn->wait_event.tasklet);
+  done:
+       fstrm->flags &= ~FCGI_SF_WANT_SHUTR;
+       return;
+
+  add_to_list:
+       if (!LIST_ADDED(&fstrm->send_list)) {
+               sw->events |= SUB_RETRY_SEND;
+               if (fstrm->flags & (FCGI_SF_BLK_MBUSY|FCGI_SF_BLK_MROOM)) {
+                       fstrm->send_wait = sw;
+                       LIST_ADDQ(&fconn->send_list, &fstrm->send_list);
+               }
+       }
+       /* Let the handler know we want shutr */
+       fstrm->flags |= FCGI_SF_WANT_SHUTR;
+       return;
+}
+
+/* Performs a synchronous or asynchronous shutw(). */
+static void fcgi_do_shutw(struct fcgi_strm *fstrm)
+{
+       struct fcgi_conn *fconn = fstrm->fconn;
+       struct wait_event *sw = &fstrm->wait_event;
+
+       if (fstrm->state != FCGI_SS_HLOC || fstrm->state == FCGI_SS_CLOSED)
+               goto done;
+
+       if (fstrm->state != FCGI_SS_ERROR && (fstrm->flags & FCGI_SF_BEGIN_SENT)) {
+               if (!(fstrm->flags & (FCGI_SF_ES_SENT|FCGI_SF_ABRT_SENT)) &&
+                   !fcgi_strm_send_abort(fconn, fstrm))
+                       goto add_to_list;
+
+               if (fstrm->state == FCGI_SS_HREM)
+                       fcgi_strm_close(fstrm);
+               else
+                       fstrm->state = FCGI_SS_HLOC;
+       } else {
+               /* a connstream may require us to immediately kill the whole connection
+                * for example because of a "tcp-request content reject" rule that is
+                * normally used to limit abuse.
+                */
+               if ((fstrm->flags & FCGI_SF_KILL_CONN) &&
+                   !(fconn->flags & (FCGI_CF_ABRTS_SENT|FCGI_CF_ABRTS_FAILED)))
+                       fconn->state = FCGI_CS_CLOSED;
+
+               fcgi_strm_close(fstrm);
+       }
+
+       if (!(fconn->wait_event.events & SUB_RETRY_SEND))
+               tasklet_wakeup(fconn->wait_event.tasklet);
+  done:
+       fstrm->flags &= ~FCGI_SF_WANT_SHUTW;
+       return;
+
+  add_to_list:
+       if (!LIST_ADDED(&fstrm->send_list)) {
+               sw->events |= SUB_RETRY_SEND;
+               if (fstrm->flags & (FCGI_SF_BLK_MBUSY|FCGI_SF_BLK_MROOM)) {
+                       fstrm->send_wait = sw;
+                       LIST_ADDQ(&fconn->send_list, &fstrm->send_list);
+               }
+       }
+       /* let the handler know we want to shutw */
+       fstrm->flags |= FCGI_SF_WANT_SHUTW;
+       return;
+}
+
+/* This is the tasklet referenced in fstrm->wait_event.tasklet, it is used for
+ * deferred shutdowns when the fcgi_detach() was done but the mux buffer was full
+ * and prevented the last frame from being emitted.
+ */
+static struct task *fcgi_deferred_shut(struct task *t, void *ctx, unsigned short state)
+{
+       struct fcgi_strm *fstrm = ctx;
+       struct fcgi_conn *fconn = fstrm->fconn;
+
+       LIST_DEL_INIT(&fstrm->sending_list);
+       if (fstrm->flags & FCGI_SF_WANT_SHUTW)
+               fcgi_do_shutw(fstrm);
+
+       if (fstrm->flags & FCGI_SF_WANT_SHUTR)
+               fcgi_do_shutr(fstrm);
+
+       if (!(fstrm->flags & (FCGI_SF_WANT_SHUTR|FCGI_SF_WANT_SHUTW))) {
+               /* We're done trying to send, remove ourself from the send_list */
+               LIST_DEL_INIT(&fstrm->send_list);
+
+               if (!fstrm->cs) {
+                       fcgi_strm_destroy(fstrm);
+                       if (fcgi_conn_is_dead(fconn))
+                               fcgi_release(fconn);
+               }
+       }
+
+       return NULL;
+}
+
+/* shutr() called by the conn_stream (mux_ops.shutr) */
+static void fcgi_shutr(struct conn_stream *cs, enum cs_shr_mode mode)
+{
+       struct fcgi_strm *fstrm = cs->ctx;
+
+       if (cs->flags & CS_FL_KILL_CONN)
+               fstrm->flags |= FCGI_SF_KILL_CONN;
+
+       if (!mode)
+               return;
+
+       fcgi_do_shutr(fstrm);
+}
+
+/* shutw() called by the conn_stream (mux_ops.shutw) */
+static void fcgi_shutw(struct conn_stream *cs, enum cs_shw_mode mode)
+{
+       struct fcgi_strm *fstrm = cs->ctx;
+
+       if (cs->flags & CS_FL_KILL_CONN)
+               fstrm->flags |= FCGI_SF_KILL_CONN;
+
+       fcgi_do_shutw(fstrm);
+}
+
+/* Called from the upper layer, to subscribe to events, such as being able to send.
+ * The <param> argument here is supposed to be a pointer to a wait_event struct
+ * which will be passed to fstrm->recv_wait or fstrm->send_wait depending on the
+ * event_type. The event_type must only be a combination of SUB_RETRY_RECV and
+ * SUB_RETRY_SEND, other values will lead to -1 being returned. It always
+ * returns 0 except for the error above.
+ */
+static int fcgi_subscribe(struct conn_stream *cs, int event_type, void *param)
+{
+       struct wait_event *sw;
+       struct fcgi_strm *fstrm = cs->ctx;
+       struct fcgi_conn *fconn = fstrm->fconn;
+
+       if (event_type & SUB_RETRY_RECV) {
+               sw = param;
+               BUG_ON(fstrm->recv_wait != NULL || (sw->events & SUB_RETRY_RECV));
+               sw->events |= SUB_RETRY_RECV;
+               fstrm->recv_wait = sw;
+               event_type &= ~SUB_RETRY_RECV;
+       }
+       if (event_type & SUB_RETRY_SEND) {
+               sw = param;
+               BUG_ON(fstrm->send_wait != NULL || (sw->events & SUB_RETRY_SEND));
+               sw->events |= SUB_RETRY_SEND;
+               fstrm->send_wait = sw;
+               if (!LIST_ADDED(&fstrm->send_list))
+                       LIST_ADDQ(&fconn->send_list, &fstrm->send_list);
+               event_type &= ~SUB_RETRY_SEND;
+       }
+       if (event_type != 0)
+               return -1;
+       return 0;
+}
+
+/* Called from the upper layer, to unsubscribe some events (undo fcgi_subscribe).
+ * The <param> argument here is supposed to be a pointer to the same wait_event
+ * struct that was passed to fcgi_subscribe() otherwise nothing will be changed.
+ * It always returns zero.
+ */
+static int fcgi_unsubscribe(struct conn_stream *cs, int event_type, void *param)
+{
+       struct wait_event *sw;
+       struct fcgi_strm *fstrm = cs->ctx;
+
+       if (event_type & SUB_RETRY_RECV) {
+               sw = param;
+               BUG_ON(fstrm->recv_wait != sw);
+               sw->events &= ~SUB_RETRY_RECV;
+               fstrm->recv_wait = NULL;
+       }
+       if (event_type & SUB_RETRY_SEND) {
+               sw = param;
+               BUG_ON(fstrm->send_wait != sw);
+               LIST_DEL(&fstrm->send_list);
+               LIST_INIT(&fstrm->send_list);
+               sw->events &= ~SUB_RETRY_SEND;
+               /* We were about to send, make sure it does not happen */
+               if (LIST_ADDED(&fstrm->sending_list) && fstrm->send_wait != &fstrm->wait_event) {
+                       tasklet_remove_from_tasklet_list(fstrm->send_wait->tasklet);
+                       LIST_DEL_INIT(&fstrm->sending_list);
+               }
+               fstrm->send_wait = NULL;
+       }
+       return 0;
+}
+
+/* Called from the upper layer, to receive data */
+static size_t fcgi_rcv_buf(struct conn_stream *cs, struct buffer *buf, size_t count, int flags)
+{
+       struct fcgi_strm *fstrm = cs->ctx;
+       struct fcgi_conn *fconn = fstrm->fconn;
+       size_t ret = 0;
+
+       if (!(fconn->flags & FCGI_CF_DEM_SALLOC))
+               ret = fcgi_strm_parse_response(fstrm, buf, count);
+
+       if (b_data(&fstrm->rxbuf))
+               cs->flags |= (CS_FL_RCV_MORE | CS_FL_WANT_ROOM);
+       else {
+               cs->flags &= ~(CS_FL_RCV_MORE | CS_FL_WANT_ROOM);
+               if (fstrm->state == FCGI_SS_ERROR || fstrm->h1m.state == H1_MSG_DONE) {
+                       cs->flags |= CS_FL_EOI;
+                       if (!(fstrm->h1m.flags & (H1_MF_VER_11|H1_MF_XFER_LEN)))
+                               cs->flags |= CS_FL_EOS;
+               }
+               if (conn_xprt_read0_pending(fconn->conn))
+                       cs->flags |= CS_FL_EOS;
+               if (cs->flags & CS_FL_ERR_PENDING)
+                       cs->flags |= CS_FL_ERROR;
+               fcgi_release_buf(fconn, &fstrm->rxbuf);
+       }
+
+       if (ret && fconn->dsi == fstrm->id) {
+               /* demux is blocking on this stream's buffer */
+               fconn->flags &= ~FCGI_CF_DEM_SFULL;
+               fcgi_conn_restart_reading(fconn, 1);
+       }
+
+       return ret;
+}
+
+
+/* stops all senders of this connection for example when the mux buffer is full.
+ * They are moved from the sending_list to send_list.
+ */
+static void fcgi_stop_senders(struct fcgi_conn *fconn)
+{
+       struct fcgi_strm *fstrm, *fstrm_back;
+
+       list_for_each_entry_safe(fstrm, fstrm_back, &fconn->sending_list, sending_list) {
+               LIST_DEL_INIT(&fstrm->sending_list);
+               tasklet_remove_from_tasklet_list(fstrm->send_wait->tasklet);
+               fstrm->send_wait->events |= SUB_RETRY_SEND;
+       }
+}
+
+
+/* Called from the upper layer, to send data from buffer <buf> for no more than
+ * <count> bytes. Returns the number of bytes effectively sent. Some status
+ * flags may be updated on the conn_stream.
+ */
+static size_t fcgi_snd_buf(struct conn_stream *cs, struct buffer *buf, size_t count, int flags)
+{
+       struct fcgi_strm *fstrm = cs->ctx;
+       struct fcgi_conn *fconn = fstrm->fconn;
+       size_t total = 0;
+       size_t ret;
+       struct htx *htx = NULL;
+       struct htx_sl *sl;
+       struct htx_blk *blk;
+       uint32_t bsize;
+
+       /* If we were not just woken because we wanted to send but couldn't,
+        * and there's somebody else that is waiting to send, do nothing,
+        * we will subscribe later and be put at the end of the list
+        */
+       if (!LIST_ADDED(&fstrm->sending_list) && !LIST_ISEMPTY(&fconn->send_list))
+               return 0;
+       LIST_DEL_INIT(&fstrm->sending_list);
+
+       /* We couldn't set it to NULL before, because we needed it in case
+        * we had to cancel the tasklet
+        */
+       fstrm->send_wait = NULL;
+
+       if (fconn->state < FCGI_CS_RECORD_H)
+               return 0;
+
+       htx = htxbuf(buf);
+       if (fstrm->id == 0) {
+               int32_t id = fcgi_conn_get_next_sid(fconn);
+
+               if (id < 0) {
+                       fcgi_strm_close(fstrm);
+                       cs->flags |= CS_FL_ERROR;
+                       return 0;
+               }
+
+               eb32_delete(&fstrm->by_id);
+               fstrm->by_id.key = fstrm->id = id;
+               fconn->max_id = id;
+               fconn->nb_reserved--;
+               eb32_insert(&fconn->streams_by_id, &fstrm->by_id);
+
+
+               /* Check if length of the body is known or if the message is
+                * full. Otherwise, the request is invalid.
+                */
+               sl = http_get_stline(htx);
+               if (!sl || (!(sl->flags & HTX_SL_F_CLEN) && (htx_get_tail_type(htx) != HTX_BLK_EOM))) {
+                       htx->flags |= HTX_FL_PARSING_ERROR;
+                       fcgi_strm_error(fstrm);
+                       goto done;
+               }
+       }
+
+       if (!(fstrm->flags & FCGI_SF_BEGIN_SENT)) {
+               if (!fcgi_strm_send_begin_request(fconn, fstrm))
+                       goto done;
+       }
+
+       if (!(fstrm->flags & FCGI_SF_OUTGOING_DATA) && count)
+               fstrm->flags |= FCGI_SF_OUTGOING_DATA;
+
+       while (fstrm->state < FCGI_SS_HLOC && !(fconn->flags & FCGI_SF_BLK_ANY) &&
+              count && !htx_is_empty(htx)) {
+               blk = htx_get_head_blk(htx);
+               bsize = htx_get_blksz(blk);
+
+               switch (htx_get_blk_type(blk)) {
+                       case HTX_BLK_REQ_SL:
+                       case HTX_BLK_HDR:
+                               ret = fcgi_strm_send_params(fconn, fstrm, htx);
+                               if (!ret) {
+                                       goto done;
+                               }
+                               total += ret;
+                               count -= ret;
+                               break;
+
+                       case HTX_BLK_EOH:
+                               ret = fcgi_strm_send_empty_params(fconn, fstrm);
+                               if (!ret)
+                                       goto done;
+                               goto remove_blk;
+
+                       case HTX_BLK_DATA:
+                               ret = fcgi_strm_send_stdin(fconn, fstrm, htx, count, buf);
+                               if (ret > 0) {
+                                       htx = htx_from_buf(buf);
+                                       total += ret;
+                                       count -= ret;
+                                       if (ret < bsize)
+                                               goto done;
+                               }
+                               break;
+
+                       case HTX_BLK_EOM:
+                               ret = fcgi_strm_send_empty_stdin(fconn, fstrm);
+                               if (!ret)
+                                       goto done;
+                               goto remove_blk;
+
+                       default:
+                         remove_blk:
+                               htx_remove_blk(htx, blk);
+                               total += bsize;
+                               count -= bsize;
+                               break;
+               }
+       }
+
+  done:
+       if (fstrm->state >= FCGI_SS_HLOC) {
+               /* trim any possibly pending data after we close (extra CR-LF,
+                * unprocessed trailers, abnormal extra data, ...)
+                */
+               total += count;
+               count = 0;
+       }
+
+       if (fstrm->state == FCGI_SS_ERROR) {
+               cs_set_error(cs);
+               if (!(fstrm->flags & FCGI_SF_BEGIN_SENT) || fcgi_strm_send_abort(fconn, fstrm))
+                       fcgi_strm_close(fstrm);
+       }
+
+       if (htx)
+               htx_to_buf(htx, buf);
+
+       /* The mux is full, cancel the pending tasks */
+       if ((fconn->flags & FCGI_CF_MUX_BLOCK_ANY) || (fstrm->flags & FCGI_SF_BLK_MBUSY))
+               fcgi_stop_senders(fconn);
+
+       if (total > 0) {
+               if (!(fconn->wait_event.events & SUB_RETRY_SEND))
+                       tasklet_wakeup(fconn->wait_event.tasklet);
+
+               /* Ok we managed to send something, leave the send_list */
+               LIST_DEL_INIT(&fstrm->send_list);
+       }
+       return total;
+}
+
+/* for debugging with CLI's "show fd" command */
+static void fcgi_show_fd(struct buffer *msg, struct connection *conn)
+{
+       struct fcgi_conn *fconn = conn->ctx;
+       struct fcgi_strm *fstrm = NULL;
+       struct eb32_node *node;
+       int send_cnt = 0;
+       int tree_cnt = 0;
+       int orph_cnt = 0;
+       struct buffer *hmbuf, *tmbuf;
+
+       if (!fconn)
+               return;
+
+       list_for_each_entry(fstrm, &fconn->send_list, send_list)
+               send_cnt++;
+
+       fstrm = NULL;
+       node = eb32_first(&fconn->streams_by_id);
+       while (node) {
+               fstrm = container_of(node, struct fcgi_strm, by_id);
+               tree_cnt++;
+               if (!fstrm->cs)
+                       orph_cnt++;
+               node = eb32_next(node);
+       }
+
+       hmbuf = br_head(fconn->mbuf);
+       tmbuf = br_tail(fconn->mbuf);
+       chunk_appendf(msg, " fconn.st0=%d .maxid=%d .flg=0x%04x .nbst=%u"
+                     " .nbcs=%u .send_cnt=%d .tree_cnt=%d .orph_cnt=%d .sub=%d "
+                     ".dsi=%d .dbuf=%u@%p+%u/%u .mbuf=[%u..%u|%u],h=[%u@%p+%u/%u],t=[%u@%p+%u/%u]",
+                     fconn->state, fconn->max_id, fconn->flags,
+                     fconn->nb_streams, fconn->nb_cs, send_cnt, tree_cnt, orph_cnt,
+                     fconn->wait_event.events, fconn->dsi,
+                     (unsigned int)b_data(&fconn->dbuf), b_orig(&fconn->dbuf),
+                     (unsigned int)b_head_ofs(&fconn->dbuf), (unsigned int)b_size(&fconn->dbuf),
+                     br_head_idx(fconn->mbuf), br_tail_idx(fconn->mbuf), br_size(fconn->mbuf),
+                     (unsigned int)b_data(hmbuf), b_orig(hmbuf),
+                     (unsigned int)b_head_ofs(hmbuf), (unsigned int)b_size(hmbuf),
+                     (unsigned int)b_data(tmbuf), b_orig(tmbuf),
+                     (unsigned int)b_head_ofs(tmbuf), (unsigned int)b_size(tmbuf));
+
+       if (fstrm) {
+               chunk_appendf(msg, " last_fstrm=%p .id=%d .flg=0x%04x .rxbuf=%u@%p+%u/%u .cs=%p",
+                             fstrm, fstrm->id, fstrm->flags,
+                             (unsigned int)b_data(&fstrm->rxbuf), b_orig(&fstrm->rxbuf),
+                             (unsigned int)b_head_ofs(&fstrm->rxbuf), (unsigned int)b_size(&fstrm->rxbuf),
+                             fstrm->cs);
+               if (fstrm->cs)
+                       chunk_appendf(msg, " .cs.flg=0x%08x .cs.data=%p",
+                                     fstrm->cs->flags, fstrm->cs->data);
+       }
+}
+
+/****************************************/
+/* MUX initialization and instanciation */
+/****************************************/
+
+/* The mux operations */
+static const struct mux_ops mux_fcgi_ops = {
+       .init          = fcgi_init,
+       .wake          = fcgi_wake,
+       .attach        = fcgi_attach,
+       .get_first_cs  = fcgi_get_first_cs,
+       .detach        = fcgi_detach,
+       .destroy       = fcgi_destroy,
+       .avail_streams = fcgi_avail_streams,
+       .used_streams  = fcgi_used_streams,
+       .rcv_buf       = fcgi_rcv_buf,
+       .snd_buf       = fcgi_snd_buf,
+       .subscribe     = fcgi_subscribe,
+       .unsubscribe   = fcgi_unsubscribe,
+       .shutr         = fcgi_shutr,
+       .shutw         = fcgi_shutw,
+       .show_fd       = fcgi_show_fd,
+       .flags         = MX_FL_HTX,
+       .name          = "FCGI",
+};
+
+
+/* this mux registers FCGI proto */
+static struct mux_proto_list mux_proto_fcgi =
+{ .token = IST("fcgi"), .mode = PROTO_MODE_HTTP, .side = PROTO_SIDE_BE, .mux = &mux_fcgi_ops };
+
+INITCALL1(STG_REGISTER, register_mux_proto, &mux_proto_fcgi);
+
+/*
+ * Local variables:
+ *  c-indent-level: 8
+ *  c-basic-offset: 8
+ * End:
+ */