]> git.kaiwu.me - haproxy.git/commitdiff
WIP: haload sources
authorFrederic Lecaille <flecaille@haproxy.com>
Thu, 19 Mar 2026 17:27:48 +0000 (18:27 +0100)
committerFrederic Lecaille <flecaille@haproxy.com>
Tue, 5 May 2026 17:51:33 +0000 (19:51 +0200)
include/haproxy/haload.h [new file with mode: 0644]
include/haproxy/hldstream-t.h
src/haload.c [new file with mode: 0644]
src/haload_init.c [new file with mode: 0644]

diff --git a/include/haproxy/haload.h b/include/haproxy/haload.h
new file mode 100644 (file)
index 0000000..1aa207f
--- /dev/null
@@ -0,0 +1,59 @@
+#ifndef _HAPROXY_HALOAD_H
+#define _HAPROXY_HALOAD_H
+
+#include <import/ist.h>
+#include <haproxy/list-t.h>
+#include <haproxy/proxy-t.h>
+#include <haproxy/server-t.h>
+#include <haproxy/task-t.h>
+
+struct hld_path {
+       const char *path;
+       struct hld_path *next;
+};
+
+struct hld_url_cfg {
+       int ssl;
+       const char *addr;
+       const char *raw_addr;
+       char *ssl_opts;
+       struct server *srv;
+       struct hld_path *paths;
+       struct hld_url_cfg *next;
+};
+
+struct hld_url {
+       int nreqs;
+       int mreqs;
+       int flags;
+       struct hld_url_cfg *cfg;
+       struct hld_url *next;
+};
+
+/* haload header */
+struct hld_hdr {
+       struct ist name;
+       struct ist value;
+       struct list list;
+};
+
+extern const char *arg_host;
+extern const char *arg_conn_hdr;
+extern const char *arg_uri;
+extern const char *arg_path; // TO REMOVE
+extern struct list hld_hdrs;
+extern struct hld_url_cfg *hld_url_cfgs;
+
+extern struct proxy hld_proxy;
+extern int arg_dura;
+extern int arg_head;
+extern int arg_hscd;
+extern int arg_long;
+extern int arg_mreqs;
+extern int arg_reqs;
+extern int arg_rcon;
+extern int arg_slow;
+extern int arg_usr;
+extern int arg_wait;
+
+#endif /* _HAPROXY_HALOAD_H */
index 6c445b0df5c21019dc6f6a317a0db382e67b9dc3..2f0cda86eaa35b4fdc13e9615e37aca49608ed46 100644 (file)
@@ -23,6 +23,7 @@ struct hldstream {
        int flags;
        int state;
        unsigned long long to_send; /* number of body data bytes to send */
+       struct timeval req_date;
        struct list list;
 };
 
diff --git a/src/haload.c b/src/haload.c
new file mode 100644 (file)
index 0000000..9e43894
--- /dev/null
@@ -0,0 +1,1838 @@
+#include <openssl/ssl.h>
+
+#include <haproxy/api.h>
+#include <haproxy/dynbuf.h>
+#include <haproxy/errors.h>
+#include <haproxy/http.h>
+#include <haproxy/http_htx.h>
+#include <haproxy/htx.h>
+#include <haproxy/hldstream.h>
+#include <haproxy/haload.h>
+#include <haproxy/proxy.h>
+#include <haproxy/task.h>
+#include <haproxy/trace.h>
+#include <haproxy/protocol.h>
+#include <haproxy/quic_tp.h>
+#include <haproxy/server.h>
+#include <haproxy/session.h>
+#include <haproxy/stats.h>
+#include <haproxy/stconn.h>
+#include <haproxy/stream.h>
+
+#if 1
+#define DDPRINTF(x...) fprintf(x)
+#else
+#define DDPRINTF(x...) do {} while(0)
+#endif
+
+/* haload stream state flags */
+#define HLD_STRM_ST_IN_ALLOC     0x0001
+#define HLD_STRM_ST_OUT_ALLOC    0x0002
+#define HLD_STRM_ST_CONN_ERR     0x0004
+#define HLD_STRM_ST_HDRS_SENT    0x0008
+#define HLD_STRM_ST_REQ_TO_BUILD 0x0010
+#define HLD_STRM_ST_MUST_RECV    0x0020
+#define HLD_STRM_ST_GOT_RESP_SL  0x0040
+
+static inline struct hld_usr *hld_new_usr(int nreqs);
+
+struct hld_mtask {
+       struct task *t;
+       unsigned int show_time;
+} mtask;
+
+struct hld_freq_ctr {
+       uint32_t curr_sec; /* start date of current period (seconds from now.tv_sec) */
+       uint32_t curr_ctr; /* cumulated value for current period */
+       uint32_t prev_ctr; /* value for last period */
+
+};
+
+struct hld_thr_info {
+       struct timeval now;          // current time
+       struct hld_freq_ctr req_rate;    // thread's measured request rate
+       struct hld_freq_ctr conn_rate;   // thread's measured connection rate
+       uint32_t cur_req;            // number of active requests
+       uint32_t curconn;            // number of active connections
+       uint32_t maxconn;            // max number of active connections
+       uint32_t is_ssl;             // non-zero if SSL is used
+       uint64_t tot_conn;           // total conns attempted on this thread
+       //uint64_t tot_req;            // total requests started on this thread
+       uint64_t tot_done;           // total requests finished (successes+failures)
+       uint64_t tot_sent;           // total bytes sent on this thread
+       uint64_t tot_rcvd;           // total bytes received on this thread
+       uint64_t tot_serr;           // total socket errors on this thread
+       uint64_t tot_cerr;           // total connection errors on this thread
+       uint64_t tot_xerr;           // total xfer errors on this thread
+       uint64_t tot_perr;           // total protocol errors on this thread
+       uint64_t tot_cto;            // total connection timeouts on this thread
+       uint64_t tot_xto;            // total xfer timeouts on this thread
+       uint64_t tot_fbs;            // total number of ttfb samples
+       uint64_t tot_ttfb;           // total time-to-first-byte (us)
+       uint64_t tot_lbs;            // total number of ttlb samples
+       uint64_t tot_ttlb;           // total time-to-last-byte (us)
+       uint64_t *ttfb_pct;          // counts per ttfb value for percentile
+       uint64_t *ttlb_pct;          // counts per ttlb value for percentile
+       uint64_t tot_sc[5];          // total status codes on this thread: 1xx,2xx,3xx,4xx,5xx
+       //int epollfd;                 // poller's FD
+       int start_len;               // request's start line's length
+       char *start_line;            // copy of the request's start line to be sent
+       char *hdr_block;             // copy of the request's header block to be sent
+       int hdr_len;                 // request's header block's length
+       int ka_req_len;              // keep-alive request length
+       char *ka_req;                // fully assembled keep-alive request
+       char *cl_req;                // fully assembled close request
+       int cl_req_len;              // close request length
+       __attribute__((aligned(64))) union { } __pad;
+};
+
+struct hld_usr {
+       struct task *task;
+       struct session *sess;
+       struct list strms;
+       struct hld_url *urls;
+       struct hld_url *cur_url;
+       int flags;
+       int show_expire;
+};
+
+struct hld_thr_info *thrs_info;
+
+struct list hld_hdrs = LIST_HEAD_INIT(hld_hdrs);
+struct proxy hld_proxy;
+
+const char *arg_host;
+const char *arg_conn_hdr;
+const char *arg_uri;
+const char *arg_path;
+
+int arg_accu;          // more accurate req/time measurements in keep-alive
+int arg_dura;          // test duration in sec if non-nul
+int arg_head;          // use HEAD
+int arg_hscd;          // HTTP status code distribution
+int arg_long;          // long output format; 2=raw values
+int arg_mreqs = 1;     // max concurrent streams by connection
+int arg_rcon;          // max requests per conn
+int arg_reqs = -1;     // max total requests
+int arg_slow;          // slow start: delay in milliseconds
+int arg_usr = 1;       // number of users
+int arg_wait = 10000;  // I/O time out (ms)
+
+int conn_tid;
+int usr_cnt;           // user count
+int usr_reqs, min_reqs, mod_req;
+
+char *hld_args[MAX_LINE_ARGS + 1];
+
+/************ time manipulation functions ***************/
+
+struct timeval hld_start_date, hld_stop_date, hld_now;
+volatile uint32_t throttle = 0;  // pass to mul32hi() if not null.
+
+/* timeval is not set */
+#define TV_UNSET ((struct timeval){ .tv_sec = 0, .tv_usec = ~0 })
+
+/* make a timeval from <sec>, <usec> */
+__attribute__((unused))
+static inline struct timeval tv_set(time_t sec, suseconds_t usec)
+{
+       struct timeval ret = { .tv_sec = sec, .tv_usec = usec };
+       return ret;
+}
+
+/* used to unset a timeout */
+__attribute__((unused))
+static inline struct timeval tv_unset(void)
+{
+       return tv_set(0, ~0);
+}
+
+/* returns the interval in microseconds, which must be set */
+static inline uint64_t tv_us(const struct timeval tv)
+{
+       return tv.tv_sec * (uint64_t)1000000 + tv.tv_usec;
+}
+
+#if 0
+/* used to zero a timeval */
+static inline struct timeval tv_zero(void)
+{
+       return tv_set(0, 0);
+}
+
+/* returns true if the timeval is set */
+static inline int tv_isset(struct timeval tv)
+{
+       return tv.tv_usec != ~0;
+}
+
+/* returns true if <a> is before <b>, taking account unsets */
+static inline int tv_isbefore(const struct timeval a, const struct timeval b)
+{
+       return !tv_isset(b) ? 1 :
+              !tv_isset(a) ? 0 :
+              ( a.tv_sec < b.tv_sec || (a.tv_sec == b.tv_sec && a.tv_usec < b.tv_usec));
+}
+
+/* returns the lowest of the two timers, for use in delay computation */
+static inline struct timeval tv_min(const struct timeval a, const struct timeval b)
+{
+       if (tv_isbefore(a, b))
+               return a;
+       else
+               return b;
+}
+
+/* returns the normalized sum of the <from> plus <off> */
+static inline struct timeval tv_add(const struct timeval from, const struct timeval off)
+{
+       struct timeval ret;
+
+       ret.tv_sec  = from.tv_sec  + off.tv_sec;
+       ret.tv_usec = from.tv_usec + off.tv_usec;
+
+       if (ret.tv_usec >= 1000000) {
+               ret.tv_usec -= 1000000;
+               ret.tv_sec  += 1;
+       }
+       return ret;
+}
+
+/* returns the normalized sum of <from> plus <ms> milliseconds */
+static inline struct timeval tv_ms_add(const struct timeval from, unsigned int ms)
+{
+       struct timeval tv;
+
+       tv.tv_usec = from.tv_usec + (ms % 1000) * 1000;
+       tv.tv_sec  = from.tv_sec  + (ms / 1000);
+       if (tv.tv_usec >= 1000000) {
+               tv.tv_usec -= 1000000;
+               tv.tv_sec++;
+       }
+       return tv;
+}
+#endif
+
+/* returns the delay between <past> and <now> or zero if <past> is after <now> */
+__attribute__((unused))
+static inline struct timeval tv_diff(const struct timeval *past, const struct timeval *now)
+{
+       struct timeval ret = { .tv_sec = 0, .tv_usec = 0 };
+
+       if (tv_isbefore(past, now)) {
+               ret.tv_sec  = now->tv_sec  - past->tv_sec;
+               ret.tv_usec = now->tv_usec - past->tv_usec;
+
+               if ((signed)ret.tv_usec < 0) { // overflow
+                       ret.tv_usec += 1000000;
+                       ret.tv_sec  -= 1;
+               }
+       }
+       return ret;
+}
+
+#if 0
+/* returns the time remaining between <tv1> and <tv2>, or zero if passed */
+static inline struct timeval tv_remain(const struct timeval tv1, const struct timeval tv2)
+{
+       struct timeval tv;
+
+       tv.tv_usec = tv2.tv_usec - tv1.tv_usec;
+       tv.tv_sec  = tv2.tv_sec  - tv1.tv_sec;
+       if ((signed)tv.tv_sec > 0) {
+               if ((signed)tv.tv_usec < 0) {
+                       tv.tv_usec += 1000000;
+                       tv.tv_sec--;
+               }
+       } else if (tv.tv_sec == 0) {
+               if ((signed)tv.tv_usec < 0)
+                       tv.tv_usec = 0;
+       } else {
+               tv.tv_sec = 0;
+               tv.tv_usec = 0;
+       }
+       return tv;
+}
+
+/* returns the time remaining between <tv1> and <tv2> in milliseconds rounded
+ * up to the next millisecond, or zero if passed.
+ */
+static inline unsigned long tv_ms_remain(const struct timeval tv1, const struct timeval tv2)
+{
+       struct timeval tv;
+
+       tv = tv_remain(tv1, tv2);
+       return tv.tv_sec * 1000 + (tv.tv_usec + 999) / 1000;
+}
+
+
+/* Multiply the two 32-bit operands and shift the 64-bit result right 32 bits.
+ * This is used to compute fixed ratios by setting one of the operands to
+ * (2^32*ratio).
+ */
+static inline uint32_t mul32hi(uint32_t a, uint32_t b)
+{
+       return ((uint64_t)a * b + a - 1) >> 32;
+}
+#endif
+
+/* read a freq counter over a 1-second period and return the event rate/s */
+uint32_t hdl_read_freq_ctr(struct hld_freq_ctr *ctr, const struct timeval now)
+{
+       uint32_t curr, past;
+       uint32_t age;
+
+       age = now.tv_sec - ctr->curr_sec;
+       if (age > 1)
+               return 0;
+
+       curr = 0;
+       past = ctr->curr_ctr;
+       if (!age) {
+               curr = past;
+               past = ctr->prev_ctr;
+       }
+
+       if (past <= 1 && !curr)
+               return past; /* very low rate, avoid flapping */
+
+       return curr + mul32hi(past, (unsigned)(999999 - now.tv_usec) * 4294U);
+}
+
+/* returns the number of remaining events that can occur on this freq counter
+ * while respecting <freq> and taking into account that <pend> events are
+ * already known to be pending. Returns 0 if limit was reached.
+ */
+uint32_t hld_freq_ctr_remain(struct hld_freq_ctr *ctr, uint32_t freq,
+                             uint32_t pend, const struct timeval now)
+{
+       uint32_t curr, past;
+       uint32_t age;
+
+       curr = 0;
+       age = now.tv_sec - ctr->curr_sec;
+
+       if (age <= 1) {
+               past = ctr->curr_ctr;
+               if (!age) {
+                       curr = past;
+                       past = ctr->prev_ctr;
+               }
+               curr += mul32hi(past, (unsigned)(999999 - now.tv_usec) * 4294U);
+       }
+       curr += pend;
+
+       if (curr >= freq)
+               return 0;
+       return freq - curr;
+}
+
+/* return the expected wait time in ms before the next event may occur,
+ * respecting frequency <freq>, and assuming there may already be some pending
+ * events. It returns zero if we can proceed immediately, otherwise the wait
+ * time, which will be rounded down 1ms for better accuracy, with a minimum
+ * of one ms.
+ */
+uint32_t hld_next_event_delay(struct hld_freq_ctr *ctr, uint32_t freq,
+                              uint32_t pend, const struct timeval now)
+{
+       uint32_t curr, past;
+       uint32_t wait, age;
+
+       past = 0;
+       curr = 0;
+       age = now.tv_sec - ctr->curr_sec;
+
+       if (age <= 1) {
+               past = ctr->curr_ctr;
+               if (!age) {
+                       curr = past;
+                       past = ctr->prev_ctr;
+               }
+               curr += mul32hi(past, (unsigned)(999999 - now.tv_usec) * 4294U);
+       }
+       curr += pend;
+
+       if (curr < freq)
+               return 0;
+
+       /* too many events already, let's count how long to wait before they're
+        * processed.
+        */
+       curr = curr - freq; // number of events left after current period
+
+       /* each events takes 1/freq second or 1000/freq ms */
+
+       wait = curr * 1000 / freq;
+       if (!wait)
+               wait = 1;
+       return wait;
+}
+
+/* Rotate a frequency counter when current period is over. Must not be called
+ * during a valid period. It is important that it correctly initializes a null
+ * area.
+ */
+static inline void hld_rotate_freq_ctr(struct hld_freq_ctr *ctr,
+                                       const struct timeval now)
+{
+       ctr->prev_ctr = ctr->curr_ctr;
+       if (now.tv_sec - ctr->curr_sec != 1) {
+               /* we missed more than one second */
+               ctr->prev_ctr = 0;
+       }
+       ctr->curr_sec = now.tv_sec;
+       ctr->curr_ctr = 0; /* leave it at the end to help gcc optimize it away */
+}
+
+/* Update a frequency counter by <inc> incremental units. It is automatically
+ * rotated if the period is over. It is important that it correctly initializes
+ * a null area.
+ */
+__attribute__((unused))
+static inline void hdl_update_freq_ctr(struct hld_freq_ctr *ctr, uint32_t inc,
+                                       const struct timeval now)
+{
+       if (ctr->curr_sec == now.tv_sec) {
+               ctr->curr_ctr += inc;
+               return;
+       }
+       hld_rotate_freq_ctr(ctr, now);
+       ctr->curr_ctr = inc;
+}
+
+
+#define TRACE_SOURCE &trace_haload
+struct trace_source trace_haload;
+static void hld_trace(enum trace_level level, uint64_t mask, const struct trace_source *src,
+                         const struct ist where, const struct ist func,
+                         const void *a1, const void *a2, const void *a3, const void *a4);
+static inline void hldstream_free(struct hldstream **hs);
+
+static const struct name_desc hld_trace_logon_args[4] = {
+       /* arg1 */ { /* already used by the haload stream */ },
+       /* arg2 */ {
+               .name = "hld",
+               .desc = "haload",
+       },
+       /* arg3 */ { },
+       /* arg4 */ { }
+};
+
+static const struct trace_event hld_trace_events[] = {
+#define HLD_EV_MAIN_TASK  (1ULL << 0)
+       { .mask = HLD_EV_MAIN_TASK, .name = "mtask",      .desc = "haload main task" },
+#define HLD_EV_USR_TASK  (1ULL << 0)
+       { .mask = HLD_EV_USR_TASK,  .name = "usr_task",   .desc = "haload user task" },
+#define HLD_STRM_EV_TX     (1ULL << 1)
+       { .mask = HLD_STRM_EV_TX,     .name = "tx",        .desc = "haload stream sending" },
+#define HLD_STRM_EV_TX_BLK (1ULL << 2)
+       { .mask = HLD_STRM_EV_TX_BLK, .name = "tx_blk",    .desc = "haload stream sending blocked" },
+#define HLD_STRM_EV_RX     (1ULL << 3)
+       { .mask = HLD_STRM_EV_RX,     .name = "rx",        .desc = "haload stream receiving" },
+#define HLD_STRM_EV_RX_BLK (1ULL << 4)
+       { .mask = HLD_STRM_EV_RX_BLK, .name = "rx_blk",    .desc = "haload stream receiving blocked" },
+#define HLD_STRM_EV_TASK   (1ULL << 5)
+       { .mask = HLD_STRM_EV_TASK,   .name = "strm_task", .desc = "haload stream task" },
+#define HLD_STRM_EV_IO_CB  (1ULL << 6)
+       { .mask = HLD_STRM_EV_IO_CB,  .name = "io_cb",     .desc = "stconn i/o callback call" },
+};
+
+static const struct name_desc hld_trace_decoding[] = {
+#define HALOAD_VERB_CLEAN 1
+       { .name = "clean", .desc = "only user-friendly stuff, generally suitable for level \"user\"" },
+};
+
+struct trace_source trace_haload = {
+       .name = IST("haload"),
+       .desc = "haload benchmark tool",
+       /* TRACE()'s first argument is always a haload stream */
+       .arg_def = TRC_ARG1_HLDSTRM,
+       .default_cb = hld_trace,
+       .known_events = hld_trace_events,
+       .lockon_args = hld_trace_logon_args,
+       .decoding = hld_trace_decoding,
+       .report_events = ~0, /* report everything by default */
+};
+
+INITCALL1(STG_REGISTER, trace_register_source, TRACE_SOURCE);
+
+static void hld_trace(enum trace_level level, uint64_t mask, const struct trace_source *src,
+                         const struct ist where, const struct ist func,
+                         const void *a1, const void *a2, const void *a3, const void *a4)
+{
+       const struct hldstream *hs = a1;
+
+       if (!hs || src->verbosity < HALOAD_VERB_CLEAN)
+               return;
+
+       chunk_appendf(&trace_buf, " hs@%p conn@%p se@%p to_send=%u",
+                     hs, __sc_conn(hs->sc), __sc_endp(hs->sc), htxbuf(&hs->bo)->data);
+       if (hs->sc) {
+               struct connection *conn = sc_conn(hs->sc);
+               chunk_appendf(&trace_buf, " - conn=%p(0x%08x)", conn, conn ? conn->flags : 0);
+               chunk_appendf(&trace_buf, " sc=%p(0x%08x)", hs->sc, hs->sc->flags);
+       }
+}
+
+int hldstream_buf_available(void *target)
+{
+       struct hldstream *hs = target;
+
+       if ((hs->flags & HLD_STRM_ST_IN_ALLOC) && b_alloc(&hs->bi, DB_CHANNEL)) {
+               hs->flags &= ~HLD_STRM_ST_IN_ALLOC;
+               TRACE_STATE("unblocking stream, input buffer allocated",
+                           HLD_STRM_EV_RX|HLD_STRM_EV_RX_BLK, hs);
+               task_wakeup(hs->task, TASK_WOKEN_IO);
+               return 1;
+       }
+
+       if ((hs->flags & HLD_STRM_ST_OUT_ALLOC) && b_alloc(&hs->bo, DB_CHANNEL)) {
+               hs->flags &= ~HLD_STRM_ST_OUT_ALLOC;
+               TRACE_STATE("unblocking stream, ouput buffer allocated",
+                           HLD_STRM_EV_TX|HLD_STRM_EV_TX_BLK, hs);
+               task_wakeup(hs->task, TASK_WOKEN_IO);
+               return 1;
+       }
+
+       return 0;
+}
+
+/* Allocate a buffer. If it fails, it adds the stream in buffer wait queue */
+struct buffer *hldstream_get_buf(struct hldstream *hs, struct buffer *bptr)
+{
+       struct buffer *buf = NULL;
+
+       if (likely(!LIST_INLIST(&hs->buf_wait.list)) &&
+           unlikely((buf = b_alloc(bptr, DB_CHANNEL)) == NULL)) {
+               b_queue(DB_CHANNEL, &hs->buf_wait, hs, hldstream_buf_available);
+       }
+
+       return buf;
+}
+
+static inline struct buffer *hldstream_get_obuf(struct hldstream *hs)
+{
+       return hldstream_get_buf(hs, &hs->bo);
+}
+
+static inline struct buffer *hldstream_get_ibuf(struct hldstream *hs)
+{
+       return hldstream_get_buf(hs, &hs->bi);
+}
+
+/* Release a buffer, if any, and try to wake up entities waiting in the buffer
+ * wait queue.
+ */
+void hldstream_release_buf(struct hldstream *hs, struct buffer *bptr)
+{
+       if (bptr->size) {
+               b_free(bptr);
+               offer_buffers(hs->buf_wait.target, 1);
+       }
+}
+
+static inline void hldstream_release_ibuf(struct hldstream *hs)
+{
+       hldstream_release_buf(hs, &hs->bi);
+}
+
+static inline void hldstream_release_obuf(struct hldstream *hs)
+{
+       hldstream_release_buf(hs, &hs->bo);
+}
+
+static inline void hldstream_free(struct hldstream **hs)
+{
+       struct hldstream *h = *hs;
+
+       TRACE_PRINTF(TRACE_LEVEL_PROTO, HLD_STRM_EV_TASK, hs, 0, 0, 0,
+                    "freeing %p stream", h);
+       hldstream_release_ibuf(h);
+       hldstream_release_obuf(h);
+       ha_free(hs);
+       TRACE_LEAVE(HLD_STRM_EV_TASK);
+}
+
+
+/* Creates a new stream connector from a haload connection. There is no endpoint
+ * here, thus it will be created by sc_new(). So the SE_FL_DETACHED flag is set.
+ * It returns NULL on error. On success, the new stream connector is returned.
+ */
+struct stconn *sc_new_from_hldstream(struct hldstream *hs, unsigned int flags)
+{
+       struct stconn *sc;
+
+       sc = sc_new(NULL);
+       if (unlikely(!sc))
+               return NULL;
+
+       sc->flags |= flags;
+       sc_ep_set(sc, SE_FL_DETACHED);
+       sc->app = &hs->obj_type;
+       return sc;
+}
+
+/* reports a locally allocated string to represent a human-readable positive
+ * number on 4 characters (3 digits and a unit, which may be "." for ones) :
+ *   XXXu
+ *   XXuX
+ *   XuXX
+ */
+static const char *human_number(double x)
+{
+       static char str[5];
+       char unit = '.';
+
+       if (x < 0)
+               x = -x;
+
+       do {
+               if (x == 0.0 || x >= 1.0) break;
+               x *= 1000.0; unit = 'm';
+               if (x >= 1.0) break;
+               x *= 1000.0; unit = 'u';
+               if (x >= 1.0) break;
+               x *= 1000.0; unit = 'n';
+               if (x >= 1.0) break;
+               x *= 1000.0; unit = 'p';
+               if (x >= 1.0) break;
+               x *= 1000.0; unit = 'f';
+       } while (0);
+
+       do {
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'k';
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'M';
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'G';
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'T';
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'P';
+               if (x < 1000.0) break;
+               x /= 1000.0; unit = 'E';
+       } while (0);
+
+       if (x < 10.0)
+               snprintf(str, sizeof(str), "%d%c%02d", (int)x, unit, (int)((x - (int)x)*100));
+       else if (x < 100.0)
+               snprintf(str, sizeof(str), "%d%c%d",   (int)x, unit, (int)((x - (int)x)*10));
+       else
+               snprintf(str, sizeof(str), "%d%c",     (int)x, unit);
+       return str;
+}
+
+/* Builds a string from the time interval <us> (in microsecond), made of a 5
+ * digit value followed by a unit among 'n', 'u', 'm', 's' for "nanoseconds",
+ * "microseconds", "milliseconds", "seconds" respectively. Large values will
+ * stick to the seconds unit and will enlarge the output, though this is not
+ * expected to be a common case. This way the output can be converted back
+ * into integer values without too much hassle (e.g. for graphs). The string
+ * is locally allocated so this must not be used by multiple threads. Negative
+ * values are reported as "  -  ".
+ */
+static const char *short_delay_str(double us)
+{
+       static char str[20];
+       char unit;
+
+       if (us <= 0.0) {
+               return "   -  ";
+       }
+       else if (us < 1.0) {
+               us *= 1000.0;
+               unit = 'n';
+       }
+       else if (us < 1000.0) {
+               unit = 'u';
+       }
+       else if (us < 1000000.0) {
+               us /= 1000.0;
+               unit = 'm';
+       }
+       else {
+               us /= 1000000.0;
+               unit = 's';
+       }
+
+       if (us < 10.0)
+               snprintf(str, sizeof(str), "%1.3f%c", us, unit);
+       else if (us < 100.0)
+               snprintf(str, sizeof(str), "%2.2f%c", us, unit);
+       else if (us < 1000.0)
+               snprintf(str, sizeof(str), "%3.1f%c", us, unit);
+       else
+               snprintf(str, sizeof(str), "%5f%c", us, unit);
+       return str;
+}
+
+/* reports current date (now) and aggragated stats */
+void hld_summary(void)
+{
+       int th;
+       uint64_t cur_conn, tot_conn, tot_req, tot_err, tot_rcvd, bytes;
+       uint64_t tot_ttfb, tot_ttlb, tot_fbs, tot_lbs, tot_sc[5];
+       static uint64_t prev_totc, prev_totr, prev_totb;
+       static uint64_t prev_ttfb, prev_ttlb, prev_fbs, prev_lbs, prev_sc[5];
+       static struct timeval prev_date = TV_UNSET;
+       double interval;
+
+       cur_conn = tot_conn = tot_req = tot_err = tot_rcvd = 0;
+       tot_ttfb = tot_ttlb = tot_fbs = tot_lbs = 0;
+       tot_sc[0] = tot_sc[1] = tot_sc[2] = tot_sc[3] = tot_sc[4] = 0;
+
+       for (th = 0; th < global.nbthread; th++) {
+               cur_conn += HA_ATOMIC_LOAD(&thrs_info[th].curconn);
+               tot_conn += HA_ATOMIC_LOAD(&thrs_info[th].tot_conn);
+               tot_req  += HA_ATOMIC_LOAD(&thrs_info[th].tot_done);
+               tot_err  += HA_ATOMIC_LOAD(&thrs_info[th].tot_serr) +
+                           HA_ATOMIC_LOAD(&thrs_info[th].tot_cerr) +
+                           HA_ATOMIC_LOAD(&thrs_info[th].tot_xerr) +
+                           HA_ATOMIC_LOAD(&thrs_info[th].tot_perr);
+               tot_rcvd += HA_ATOMIC_LOAD(&thrs_info[th].tot_rcvd);
+               tot_ttfb += HA_ATOMIC_LOAD(&thrs_info[th].tot_ttfb);
+               tot_ttlb += HA_ATOMIC_LOAD(&thrs_info[th].tot_ttlb);
+               tot_fbs  += HA_ATOMIC_LOAD(&thrs_info[th].tot_fbs);
+               tot_lbs  += HA_ATOMIC_LOAD(&thrs_info[th].tot_lbs);
+               tot_sc[0]+= HA_ATOMIC_LOAD(&thrs_info[th].tot_sc[0]);
+               tot_sc[1]+= HA_ATOMIC_LOAD(&thrs_info[th].tot_sc[1]);
+               tot_sc[2]+= HA_ATOMIC_LOAD(&thrs_info[th].tot_sc[2]);
+               tot_sc[3]+= HA_ATOMIC_LOAD(&thrs_info[th].tot_sc[3]);
+               tot_sc[4]+= HA_ATOMIC_LOAD(&thrs_info[th].tot_sc[4]);
+       }
+
+#if 0
+       /* when called after having stopped, check if we need to dump a final
+        * line or not, to cover for the rare cases of the last thread
+        * finishing just after the last summary line
+        */
+       if (!(running & THR_COUNT) && (prev_date.tv_sec == now.tv_sec) &&
+            (prev_totc == tot_conn) && (prev_totr == tot_req) && (prev_totb == tot_rcvd))
+               return;
+#endif
+
+       if (tv_isset(&prev_date))
+               interval = tv_ms_remain(&prev_date, &hld_now) / 1000.0;
+       else
+               interval = 1.0;
+
+       printf("%10lu %5lu %8llu %8llu %14llu %6lu ",
+              arg_long ? (unsigned long)hld_now.tv_sec :
+              (unsigned long)(hld_now.tv_sec - hld_start_date.tv_sec),
+              (unsigned long)cur_conn,
+              (unsigned long long)tot_conn,
+              (unsigned long long)tot_req,
+              (unsigned long long)tot_rcvd,
+              (unsigned long)tot_err);
+
+       bytes = tot_rcvd - prev_totb;
+#if 0
+       if (arg_ovrp) {
+               long small_pkt = (bytes + (arg_ovrp - 1)) / arg_ovrp;
+               /* we need to account for overhead also on small packets and
+                * at minima once per response.
+                */
+               if (small_pkt < tot_req  - prev_totr)
+                       small_pkt = tot_req  - prev_totr;
+               bytes += small_pkt * arg_ovre;
+       }
+#endif
+
+       if (arg_long >= 2)
+               printf("%3u ", throttle ? mul32hi(100, throttle) : 100);
+
+       if (arg_long >= 2)
+               printf("%.1f ", (tot_conn - prev_totc) / interval);
+       else
+               printf("%s ", human_number((tot_conn - prev_totc) / interval));
+
+       if (arg_long >= 2)
+               printf("%.1f ", (tot_req  - prev_totr) / interval);
+       else
+               printf("%s ", human_number((tot_req  - prev_totr) / interval));
+
+       if (arg_long >= 2)
+               printf("%.1f ", bytes / interval);
+       else if (arg_long)
+               printf("%s ", human_number(bytes / interval));
+
+       if (arg_long >= 2)
+               printf("%.1f ", bytes * 8 / interval);
+       else
+               printf("%s ", human_number(bytes * 8 / interval));
+
+       if (arg_long >= 2) {
+               if (tot_fbs - prev_fbs)
+                       printf("%.1f ", (tot_ttfb - prev_ttfb) / (double)(tot_fbs - prev_fbs));
+               else
+                       printf("- ");
+       }
+       else
+               printf("%s ", tot_fbs == prev_fbs ? "   -  " :
+                      short_delay_str((tot_ttfb - prev_ttfb) / (double)(tot_fbs - prev_fbs)));
+
+       if (arg_long >= 2) {
+               if (tot_lbs - prev_lbs)
+                       printf("%.1f ", (tot_ttlb - prev_ttlb) / (double)(tot_lbs - prev_lbs));
+               else
+                       printf("- ");
+       }
+       else if (arg_long)
+               printf("%s ", tot_lbs == prev_lbs ? "   -  " :
+                      short_delay_str((tot_ttlb - prev_ttlb) / (double)(tot_lbs - prev_lbs)));
+
+       /* status codes distribution */
+       if (arg_hscd)
+               printf("%3llu %3llu %3llu %3llu %3llu ",
+                      (unsigned long long)(tot_sc[0] - prev_sc[0]),
+                      (unsigned long long)(tot_sc[1] - prev_sc[1]),
+                      (unsigned long long)(tot_sc[2] - prev_sc[2]),
+                      (unsigned long long)(tot_sc[3] - prev_sc[3]),
+                      (unsigned long long)(tot_sc[4] - prev_sc[4]));
+
+       putchar('\n');
+
+       prev_totc = tot_conn;
+       prev_totr = tot_req;
+       prev_totb = tot_rcvd;
+       prev_fbs  = tot_fbs;
+       prev_lbs  = tot_lbs;
+       prev_ttfb = tot_ttfb;
+       prev_ttlb = tot_ttlb;
+       prev_sc[0]= tot_sc[0];
+       prev_sc[1]= tot_sc[1];
+       prev_sc[2]= tot_sc[2];
+       prev_sc[3]= tot_sc[3];
+       prev_sc[4]= tot_sc[4];
+       prev_date = hld_now;
+}
+
+void update_throttle()
+{
+       int duration;
+       uint32_t ratio = 0;
+       uint32_t step, steps = 10, pos, base;
+
+       if (!arg_slow)
+               goto end;
+
+       duration = tv_ms_remain(&hld_start_date, &hld_now);
+       if (duration >= arg_slow)
+               goto end;
+
+       /* The ramp-up duration is cut into <steps> steps.
+        * Each step shows a ramp-up during the first quarter of its
+        * duration, and a stabilisation period during the last 3/4.
+        * For instance, with 4 steps, we have this:
+        *
+        *      ramp up
+        * |<-------------->|
+        * |             __________
+        * |         ___/:  :
+        * |     ___/   ::  :
+        * | ___/       ::  :
+        * |/           ::  :
+        * +------------++------------>
+        *
+        * Thus we have to determine the current step and the position within
+        * this step. In order to simplify this, we'll pretend there are 4
+        * times more steps and that only steps 0 mod 4 ramp up the load.
+        * The throttle is stable along the last 3 quarters of a step, at the
+        * base value of the next step.
+        */
+
+       step = (steps * 4) * duration / arg_slow;
+       if (step & 3) {
+               ratio = (uint64_t)0xffffffffU * (step / 4 + 1) / steps + 1;
+               goto end;
+       }
+
+       /* position in ms within the current step */
+       pos = duration - step * arg_slow / (steps * 4);
+
+       /* get a ratio out of it. We divide 4* the position by the step width
+        * (arg_slow/steps), and multiply this by 1/steps to get the relative
+        * height vs 100%. steps cancel each other.
+        */
+       pos = (uint64_t)0xffffffffU * pos * 4 / arg_slow;
+       base = (uint64_t)0xffffffffU * (step / 4) / steps;
+       ratio = base + pos;
+
+       //printf("base=%#x (%u)  pos=%#x (%u) tot=%#x (%u)\n",
+       //       base, mul32hi(100,base),
+       //       pos, mul32hi(100,pos),
+       //       ratio, mul32hi(100,ratio));
+
+       if (ratio < 1)
+               ratio = 1;
+ end:
+       throttle = ratio;
+}
+/* main task */
+static struct task *mtask_cb(struct task *t, void *context, unsigned int state)
+{
+       __attribute__((unused))
+       int i, nb_usr;
+       __attribute__((unused))
+       struct hld_usr *usr = context;
+
+       TRACE_ENTER(HLD_EV_MAIN_TASK);
+
+       if (tick_is_expired(mtask.show_time, now_ms)) {
+               hld_summary();
+               mtask.show_time = tick_add(now_ms, MS_TO_TICKS(1000));
+       }
+
+       gettimeofday(&hld_now, NULL);
+
+       /* users initializations */
+       DDPRINTF(stderr, "### %d arg_usr=%d usr_reqs=%d min_reqs=%d mod_req=%d\n",
+               tick_is_expired(t->expire, now_ms), arg_usr, usr_reqs, min_reqs, mod_req);
+
+       //for (i = 0; usr_cnt < arg_usr && i < 14000; i++, usr_cnt++) {
+       BUG_ON(usr_cnt > arg_usr);
+       nb_usr = MIN(arg_usr, arg_usr - usr_cnt);
+       nb_usr = MIN(80, nb_usr);
+       for (i = 0; i < nb_usr; i++, usr_cnt++) {
+               struct hld_usr *hu;
+               int req = min_reqs == -1 ? -1 :
+                       i < mod_req ? min_reqs + 1 : min_reqs;
+
+               hu = hld_new_usr(req);
+               if (!hu) {
+                       ha_alert("could not allocate a new haload user\n");
+                       goto leave;
+               }
+               fprintf(stderr, "U");
+       }
+
+       update_throttle();
+       if (usr_cnt <= arg_usr)
+               task_wakeup(t, TASK_WOKEN_IO);
+leave:
+       TRACE_LEAVE(HLD_EV_MAIN_TASK);
+       return t;
+}
+
+static int hldstream_build_http_req(struct hldstream *hs, struct ist path, int eom)
+{
+       int ret = 0;
+       struct buffer *buf;
+       struct htx *htx;
+       struct htx_sl *sl;
+       struct ist meth_ist;
+       struct hld_hdr *hdr;
+       unsigned int flags = HTX_SL_F_VER_11 | HTX_SL_F_XFER_LEN |
+               (!hs->to_send ? HTX_SL_F_BODYLESS : 0);
+
+       TRACE_ENTER(HLD_STRM_EV_TX, hs);
+       buf = hldstream_get_obuf(hs);
+       if (!buf) {
+               TRACE_STATE("waiting for ouput buffer", HLD_STRM_EV_TX|HLD_STRM_EV_TX_BLK, hs);
+               hs->flags |= HLD_STRM_ST_OUT_ALLOC;
+               goto leave;
+       }
+
+       htx = htx_from_buf(buf);
+       meth_ist = !arg_head ? ist("GET") : ist("HEAD");
+       sl = htx_add_stline(htx, HTX_BLK_REQ_SL, flags, meth_ist, path, ist("HTTP/1.1"));
+       if (!sl)
+               goto err;
+
+       sl->info.req.meth = !arg_head ? HTTP_METH_GET : HTTP_METH_HEAD;
+       list_for_each_entry(hdr, &hld_hdrs, list)
+               if (!htx_add_header(htx, hdr->name, hdr->value)) {
+                       TRACE_ERROR("could not add a header", HLD_STRM_EV_TX, hs);
+                       goto err;
+               }
+
+       if (!arg_host &&
+           !http_add_header(htx, ist("host"), ist(hs->url->cfg->raw_addr))) {
+               TRACE_ERROR("could not add host header", HLD_STRM_EV_TX, hs);
+               goto err;
+       }
+
+       if (arg_conn_hdr && !http_add_header(htx, ist("Connection"), ist("close"))) {
+               TRACE_ERROR("could not add connection header", HLD_STRM_EV_TX, hs);
+               goto err;
+       }
+
+       if (!htx_add_endof(htx, HTX_BLK_EOH))
+               goto err;
+
+       if (eom)
+               htx->flags |= HTX_FL_EOM;
+       htx_to_buf(htx, &hs->bo);
+ leave:
+       ret = 1;
+       TRACE_LEAVE(HLD_STRM_EV_TX, hs);
+       return ret;
+ err:
+       hs->flags |= HLD_STRM_ST_CONN_ERR;
+       TRACE_DEVEL("leaving on error", HLD_STRM_EV_TX, hs);
+       goto leave;
+}
+
+/* Send HTX data prepared for <hs> haload stream from <conn> connection */
+static int hldstream_htx_buf_snd(struct connection *conn, struct hldstream *hs)
+{
+       struct stconn *sc = hs->sc;
+       int ret = 0;
+       int nret;
+
+       TRACE_ENTER(HLD_STRM_EV_TX, hs);
+
+       if (!htxbuf(&hs->bo)->data) {
+               /* This is possible after having drained the body, so after
+                * having sent the response here when req_after_res=1.
+                */
+               ret = 1;
+               goto out;
+       }
+
+       //BUG_ON(!conn->mux || !conn->mux->snd_buf);
+       nret = conn->mux->snd_buf(hs->sc, &hs->bo, htxbuf(&hs->bo)->data, 0);
+       if (nret <= 0) {
+               if (hs->flags & HLD_STRM_ST_CONN_ERR ||
+                   conn->flags & CO_FL_ERROR || sc_ep_test(sc, SE_FL_ERROR)) {
+                       TRACE_DEVEL("connection error during send", HLD_STRM_EV_TX, hs);
+                       goto out;
+               }
+       }
+
+       hs->req_date = date;
+       /* The HTX data are not fully sent if the last HTX data
+        * were not fully transfered or if there are remaining data
+        * to send (->to_send > 0).
+        */
+       if (!htx_is_empty(htxbuf(&hs->bo))) {
+               TRACE_DEVEL("data not fully sent, wait", HLD_STRM_EV_TX, hs);
+               conn->mux->subscribe(sc, SUB_RETRY_SEND, &sc->wait_event);
+       }
+       else if (hs->to_send) {
+               TRACE_STATE("waking up task", HLD_STRM_EV_TX, hs);
+               task_wakeup(hs->task, TASK_WOKEN_IO);
+       }
+
+       ret = 1;
+ out:
+       if (htx_is_empty(htxbuf(&hs->bo)) || ret == 0) {
+               TRACE_DEVEL("releasing underlying buffer", HLD_STRM_EV_TX, hs);
+               hldstream_release_obuf(hs);
+       }
+
+       TRACE_LEAVE(HLD_STRM_EV_TX, hs);
+       return ret;
+}
+
+__attribute__((unused))
+static void hldstream_htx_buf_rcv(struct connection *conn,
+                                 struct hldstream *hs, int *fin)
+{
+       struct buffer *buf;
+       size_t max, read = 0, cur_read = 0;
+       int is_empty;
+       struct htx_sl *sl = NULL;
+       __attribute__((unused))
+       uint64_t ttfb, ttlb;     // time-to-first-byte, time-to-last-byte (in us)
+
+       TRACE_ENTER(HLD_STRM_EV_RX, hs);
+
+       *fin = 0;
+       if (hs->sc->wait_event.events & SUB_RETRY_RECV) {
+               TRACE_DEVEL("subscribed for RECV, waiting for data", HLD_STRM_EV_RX, hs);
+               goto leave;
+       }
+
+       if (sc_ep_test(hs->sc, SE_FL_EOS)) {
+               TRACE_STATE("end of stream", HLD_STRM_EV_RX, hs);
+               goto leave;
+       }
+
+       if (hs->flags & HLD_STRM_ST_IN_ALLOC) {
+               TRACE_STATE("waiting for input buffer", HLD_STRM_EV_RX, hs);
+               goto leave;
+       }
+
+       buf = hldstream_get_ibuf(hs);
+       if (!buf) {
+               TRACE_STATE("waiting for input buffer", HLD_STRM_EV_RX, hs);
+               hs->flags |= HLD_STRM_ST_IN_ALLOC;
+               goto leave;
+       }
+
+       while (sc_ep_test(hs->sc, SE_FL_RCV_MORE) ||
+              (!(conn->flags & CO_FL_ERROR) &&
+               !sc_ep_test(hs->sc, SE_FL_ERROR | SE_FL_EOS))) {
+               htx_reset(htxbuf(&hs->bi));
+               max = (IS_HTX_SC(hs->sc) ?
+                      htx_free_space(htxbuf(&hs->bi)) : b_room(&hs->bi));
+               sc_ep_clr(hs->sc, SE_FL_WANT_ROOM);
+               read = conn->mux->rcv_buf(hs->sc, &hs->bi, max, 0);
+               if (!(hs->flags & HLD_STRM_ST_GOT_RESP_SL) && read && !sl) {
+                       int status;
+                       sl = http_get_stline(htx_from_buf(&hs->bi));
+                       if (!sl) {
+                               TRACE_ERROR("start line not found", HLD_STRM_EV_RX, hs);
+                               hs->flags |= HLD_STRM_ST_CONN_ERR;
+                               goto leave;
+                       }
+
+                       status = sl->info.res.status;
+                       hs->flags |= HLD_STRM_ST_GOT_RESP_SL;
+                       TRACE_PRINTF(TRACE_LEVEL_PROTO, HLD_STRM_EV_RX, hs, 0, 0, 0,
+                                    "HTTP status: %d cur_read=%d",
+                                    status, (int)cur_read);
+                       thrs_info[tid].tot_sc[status * 41 / 4096 - 1]++;
+                       ttfb = tv_us(tv_diff(&hs->req_date, &date));
+                       thrs_info[tid].tot_fbs++;
+                       thrs_info[tid].tot_ttfb += ttfb;
+               }
+
+               cur_read += read;
+               if (!htx_expect_more(htxbuf(&hs->bi))) {
+                   *fin = 1;
+                   ttlb = tv_us(tv_diff(&hs->req_date, &date));
+                   thrs_info[tid].tot_lbs++;
+                   thrs_info[tid].tot_ttlb += ttlb;
+                   thrs_info[tid].tot_done++;
+                   break;
+               }
+
+               if (!read)
+                       break;
+       }
+
+       is_empty = (IS_HTX_SC(hs->sc) ?
+                   htx_is_empty(htxbuf(&hs->bi)) : !b_data(&hs->bi));
+       if (is_empty &&
+           ((conn->flags & CO_FL_ERROR) || sc_ep_test(hs->sc, SE_FL_ERROR))) {
+               /* Report network errors only if we got no other data. Otherwise
+                * we'll let the upper layers decide whether the response is OK
+                * or not. It is very common that an RST sent by the server is
+                * reported as an error just after the last data chunk.
+                */
+               TRACE_ERROR("connection error during recv", HLD_STRM_EV_RX, hs);
+               hs->flags |= HLD_STRM_ST_CONN_ERR;
+       }
+       else if (!read && !*fin && !sc_ep_test(hs->sc, SE_FL_ERROR | SE_FL_EOS)) {
+               TRACE_DEVEL("subscribing for read data", HLD_STRM_EV_RX, hs);
+               conn->mux->subscribe(hs->sc, SUB_RETRY_RECV, &hs->sc->wait_event);
+       }
+
+       thrs_info[tid].tot_rcvd += cur_read;
+ leave:
+       hldstream_release_ibuf(hs);
+       TRACE_PRINTF(TRACE_LEVEL_PROTO, HLD_STRM_EV_RX, hs, 0, 0, 0,
+                    "data received (%llu) read=%d *fin=%d",
+                    (unsigned long long)cur_read, (int)read, *fin);
+       TRACE_LEAVE(HLD_STRM_EV_RX, hs);
+}
+
+/* I/O handler wakeup from MUX */
+struct task *hld_io_cb(struct task *t, void *context, unsigned int state)
+{
+       struct stconn *sc = context;
+       struct hldstream *hs = __sc_hldstream(sc);
+       struct connection *conn; //= hs ? sc_conn(hs->sc) : NULL;
+
+       sc = hs->sc;
+       conn = sc_conn(sc);
+       if (sc_ep_test(sc, SE_FL_ERROR) || (conn && (conn->flags & CO_FL_ERROR))) {
+               TRACE_ERROR("connection error", HLD_STRM_EV_IO_CB, hs);
+               hs->flags |= HLD_STRM_ST_CONN_ERR;
+               task_wakeup(hs->task, TASK_WOKEN_IO);
+               goto err;
+       }
+
+       task_wakeup(hs->task, TASK_WOKEN_IO);
+ err:
+       return t;
+}
+
+static void hld_conn_destroy(struct connection *conn)
+{
+       int th, cur_conn = 0;
+       {
+               for (th = 0; th < global.nbthread; th++) {
+                       cur_conn += HA_ATOMIC_LOAD(&thrs_info[th].curconn);
+               }
+       }
+       fprintf(stderr, "%s: currconn=%u vs %u\n", __func__, thrs_info[tid].curconn, cur_conn);
+       BUG_ON(!thrs_info[tid].curconn);
+       thrs_info[tid].curconn--;
+}
+
+/* Try to reuse a connection from server <srv>, session <sess>, and
+ * stream connector <sc>.
+ * Always set the connection's <hash> to be reused, and return it
+ * at the <conn> address if found.
+ * Returns 1 if successful (no error, even if no connection was
+ * available to reuse), or 0 otherwise.
+ */
+static int hld_be_reuse_conn(struct connection **conn, int64_t *hash,
+                             struct stconn *sc, struct session *sess,
+                             struct server *srv)
+{
+       int ret;
+       struct sockaddr_storage dst;
+
+       /* Reset to ensure <conn> is always initialized */
+       *conn = NULL;
+       dst = srv->addr;
+       set_host_port(&dst, srv->svc_port);
+       *hash = be_calculate_conn_hash(srv, NULL, sess, NULL, &dst, IST_NULL);
+       ret = be_reuse_connection(*hash, sess, &hld_proxy, srv, sc, &srv->obj_type, 0);
+       if (ret == SF_ERR_INTERNAL) {
+               TRACE_ERROR("error during connection reuse", HLD_STRM_EV_TASK);
+               ret = 0;
+               goto leave;
+       }
+
+       if (ret == SF_ERR_NONE) {
+               TRACE_STATE("performed connection reuse", HLD_STRM_EV_TASK);
+               *conn = __sc_conn(sc);
+               conn_set_owner(*conn, sess, hld_conn_destroy);
+       }
+
+       ret = 1;
+ leave:
+       return ret;
+}
+
+static struct task *hld_strm_task(struct task *t, void *context, unsigned int state)
+{
+       struct hldstream *hs = context;
+       struct hld_url *url = hs->url;
+       struct connection *conn = sc_conn(hs->sc);
+       struct session *sess = hs->usr->sess;
+       struct server *srv = url->cfg->srv;
+       __attribute__((unused))
+       int ret, fin = 0;
+       __attribute__((unused))
+       int64_t hash;
+       __attribute__((unused))
+       struct sockaddr_storage dst;
+
+       TRACE_ENTER(HLD_STRM_EV_TASK, hs);
+
+       if (sc_ep_test(hs->sc, SE_FL_ERROR) || (conn && (conn->flags & CO_FL_ERROR))) {
+               TRACE_ERROR("connection error", HLD_STRM_EV_IO_CB, hs);
+               hs->flags |= HLD_STRM_ST_CONN_ERR;
+               goto err;
+       }
+
+       if (tick_is_expired(t->expire, now_ms)) {
+               TRACE_STATE("expired task", HLD_STRM_EV_TASK, hs);
+               hs->task = NULL;
+               t = NULL;
+               DDPRINTF(stderr, "@");
+               //task_wakeup(hs->usr->task, TASK_WOKEN_IO);
+               //goto leave;
+               goto err;
+       }
+
+       if (conn && conn->mux && conn->flags & CO_FL_WAIT_XPRT) {
+               TRACE_STATE("waiting for xprt, subscribing to send", HLD_STRM_EV_TASK, hs);
+               if (conn->mux->subscribe(hs->sc, SUB_RETRY_SEND, &hs->sc->wait_event) < 0) {
+                       TRACE_ERROR("send subscribing error", HLD_STRM_EV_TASK, hs);
+                       goto out;
+               }
+       }
+
+       if (!hs->conn) {
+               struct protocol *proto;
+               const struct mux_ops *mux_ops;
+               int status;
+
+               BUG_ON(conn);
+
+               hldstream_release_ibuf(hs);
+               hldstream_release_obuf(hs);
+
+               conn = conn_new(&srv->obj_type);
+               if (!conn) {
+                       TRACE_ERROR("stconn allocation error", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               conn->hash_node.key = hs->hash;
+               // VOIR la CB ici :
+               conn_set_owner(conn, sess, hld_conn_destroy);
+               BUG_ON(hs->sc->sedesc->sc != hs->sc);
+               if (sc_attach_mux(hs->sc, hs->sc->sedesc, conn) < 0) {
+                       TRACE_ERROR("mux attach error", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               conn->flags |= CO_FL_SSL_NO_CACHED_INFO;
+
+               if (!sockaddr_alloc(&conn->dst, NULL, 0)) {
+                       TRACE_ERROR("sockaddr allocation error", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               *conn->dst = srv->addr;
+               proto = protocol_lookup(conn->dst->ss_family,
+                                                               srv->addr_type.proto_type, srv->alt_proto);
+               set_host_port(conn->dst, srv->svc_port);
+
+               if (conn_prepare(conn, proto, srv->xprt) < 0) {
+                       TRACE_ERROR("xprt allocation error", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               /* Note that in case of connect() failure, the callback
+                * set by conn_set_owner() is called. Its role is to decrement
+                * the <currconn> counter. So, it must incremented here.
+                */
+               thrs_info[tid].curconn++;
+               thrs_info[tid].tot_conn++;
+
+               {
+                       int th, cur_conn = 0;
+                       for (th = 0; th < global.nbthread; th++) {
+                               cur_conn += HA_ATOMIC_LOAD(&thrs_info[th].curconn);
+                       }
+                       fprintf(stderr, "%s: currconn=%u vs %u\n", __func__, thrs_info[tid].curconn, cur_conn);
+               }
+               BUG_ON(!proto || !proto->connect);
+               /* XXX check the flags XXX */
+               status = proto->connect(conn, 0);
+               if (status != SF_ERR_NONE) {
+                       TRACE_ERROR("proto connect error", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               conn_set_private(conn);
+               session_add_conn(sess, conn);
+               conn->ctx = hs->sc;
+
+               if (conn_xprt_start(conn) < 0) {
+                       TRACE_ERROR("could not start xprt", HLD_STRM_EV_TASK, hs);
+                       goto err;
+               }
+
+               if (!conn_is_ssl(conn) || !srv->ssl_ctx.alpn_str) {
+                       mux_ops = conn_get_best_mux(conn, IST_NULL, PROTO_SIDE_BE, PROTO_MODE_HTTP);
+                       if (!mux_ops || conn_install_mux(conn, mux_ops, hs->sc, &hld_proxy, sess) < 0) {
+                               TRACE_ERROR("mux installation failed", HLD_STRM_EV_TASK, hs);
+                               goto err;
+                       }
+
+                       TRACE_PRINTF(TRACE_LEVEL_PROTO, HLD_STRM_EV_TASK, hs, 0, 0, 0,
+                                    "mux: %s", mux_ops->name);
+               }
+
+               hs->conn = conn;
+               if (conn->flags & CO_FL_WAIT_XPRT) {
+                       TRACE_STATE("waiting for xprt", HLD_STRM_EV_TASK, hs);
+                       if (conn->mux) {
+                               TRACE_STATE("subscribing to send", HLD_STRM_EV_TASK, hs);
+                               conn->mux->subscribe(hs->sc, SUB_RETRY_SEND, &hs->sc->wait_event);
+                       }
+               }
+
+               goto out;
+       }
+
+       if (hs->flags & HLD_STRM_ST_REQ_TO_BUILD) {
+               if (!hldstream_build_http_req(hs, ist(hs->path), 1))
+                       goto out;
+
+               hs->flags &= ~HLD_STRM_ST_REQ_TO_BUILD;
+       }
+
+       if (!hldstream_htx_buf_snd(conn, hs))
+               goto out;
+
+       hldstream_htx_buf_rcv(conn, hs, &fin);
+
+ out:
+       if (hs->flags & HLD_STRM_ST_CONN_ERR) {
+               TRACE_ERROR("haload stream error", HLD_STRM_EV_TASK, hs);
+               goto err;
+       }
+
+       if (fin) {
+               TRACE_STATE("end of stream", HLD_STRM_EV_TASK, hs);
+               goto done;
+       }
+
+       t->expire = tick_add(now_ms, MS_TO_TICKS(arg_wait));
+ leave:
+       TRACE_LEAVE(HLD_STRM_EV_TASK, hs);
+       return t;
+ done:
+       DDPRINTF(stderr, ".");
+       task_wakeup(hs->usr->task, TASK_WOKEN_IO);
+       hs->url->mreqs++;
+       LIST_DELETE(&hs->list);
+       sc_destroy(hs->sc);
+       hldstream_free(&hs);
+       t = NULL;
+       goto leave;
+ err:
+       thrs_info[tid].tot_perr++;
+       TRACE_DEVEL("leaving on error", HLD_STRM_EV_TASK, hs);
+       task_wakeup(hs->usr->task, TASK_WOKEN_IO);
+       LIST_DELETE(&hs->list);
+       sc_destroy(hs->sc);
+       hldstream_free(&hs);
+       t = NULL;
+       url->mreqs++;
+       url->nreqs = url->nreqs == -1 ? -1 : url->nreqs + 1;
+       goto leave;
+}
+
+/* Allocate a new haload stream.
+ * Return 1 if succeeded, 0 if not.
+ */
+static struct hldstream *hld_new_strm(struct hld_usr *usr,
+                                      struct hld_url *url,
+                                      struct hld_path *path)
+{
+       struct hldstream *hs;
+       struct stconn *sc;
+       struct task *t;
+       int64_t hash;
+       struct connection *conn;
+
+       hs = malloc(sizeof(*hs));
+       sc = sc_new_from_hldstream(hs, SC_FL_NONE);
+       t = task_new_here();
+       if (unlikely(!hs || !sc || !t)) {
+               TRACE_ERROR("could not allocate a new stconn", HLD_STRM_EV_TASK);
+               goto err;
+       }
+
+       /* Mandatory to make sc_attach_mux() identify this stream type */
+       hs->obj_type = OBJ_TYPE_HALOAD;
+       if (!hld_be_reuse_conn(&conn, &hash, sc, usr->sess, url->cfg->srv)) {
+               TRACE_ERROR("internal error during a connection reuse attempt",
+                           HLD_STRM_EV_TASK);
+               goto err;
+       }
+
+       t->context = hs;
+       t->process = hld_strm_task;
+       t->expire = tick_add(now_ms, MS_TO_TICKS(arg_wait));
+
+       hs->conn = conn;
+       hs->hash = hash;
+       hs->usr = usr;
+       hs->url = url;
+       hs->path = path->path;
+       hs->sc = sc;
+       hs->bi = hs->bo = BUF_NULL;
+       LIST_INIT(&hs->buf_wait.list);
+       hs->task = t;
+       if (!conn)
+               DDPRINTF(stderr, "n");
+       else
+               DDPRINTF(stderr, "y");
+       hs->flags = conn ? HLD_STRM_ST_REQ_TO_BUILD : HLD_STRM_ST_REQ_TO_BUILD;
+       hs->state = 0;
+       hs->to_send = 0;
+       hs->req_date = tv_unset();
+       LIST_APPEND(&usr->strms, &hs->list);
+       task_wakeup(t, TASK_WOKEN_INIT);
+       fprintf(stderr, "%s hs %p sc %p conn %p\n", __func__, hs, hs->sc, hs->conn);
+
+       return hs;
+ err:
+       task_destroy(t);
+       sc_destroy(sc);
+       free(hs);
+       return NULL;
+}
+
+static inline struct hld_url *hld_next_url(struct hld_url *list,
+                                           struct hld_url *cur)
+{
+       return cur->next ? cur->next : list;
+}
+
+static inline struct hld_path *hld_next_path(struct hld_path *list,
+                                             struct hld_path *cur)
+{
+       return cur->next ? cur->next : list;
+}
+
+static struct task *hld_usr_task(struct task *t, void *context, unsigned int state)
+{
+       struct hld_usr *usr = context;
+       struct hld_url *url, *urls = usr->cur_url, *first_url = urls;
+       struct hldstream *hs, *hsbak;
+       __attribute__((unused))
+       int nreqs, max_reqs = 20;
+       int remain = -1;
+       int connect_cnt = 0;
+
+       TRACE_ENTER(HLD_EV_USR_TASK);
+
+       list_for_each_entry_safe(hs, hsbak, &usr->strms, list) {
+               if (!tick_is_expired(hs->task->expire, now_ms))
+                       break;
+
+               DPRINTF(stderr, "$");
+               TRACE_STATE("expired task", HLD_EV_USR_TASK, hs);
+
+               hs->url->mreqs++;
+               LIST_DELETE(&hs->list);
+               task_destroy(hs->task);
+               hs->task = NULL;
+               sc_destroy(hs->sc);
+               hldstream_free(&hs);
+       }
+
+       for (url = urls; url; usr->cur_url = url = hld_next_url(urls, url)) {
+               struct hld_path *path, *paths = url->cfg->paths;
+
+               nreqs = url->nreqs >= 0 ? MIN(url->nreqs, url->mreqs) : url->mreqs;
+
+               DPRINTF(stdout, "\n%s: url:%s url == first_url? %d nreqs=%d remain=%d url->nreqs=%d url->mreqs=%d\n",
+                       __func__, url->cfg->addr, url == first_url, nreqs, remain, url->nreqs, url->mreqs);
+
+               for (path = paths; path && nreqs; path = hld_next_path(paths, path)) {
+                       struct hldstream *hs;
+
+                       DPRINTF(stdout, "\t%s: path:%s\n", __func__, path->path);
+                       if ((hs = hld_new_strm(usr, url, path)) == NULL) {
+                               TRACE_ERROR("could start a new stream task", HLD_EV_USR_TASK);
+                               goto out;
+                       }
+
+                       url->mreqs--;
+                       url->nreqs = url->nreqs == -1 ? -1 : url->nreqs - 1;
+                       nreqs--;
+                       BUG_ON(url->nreqs < -1 || url->mreqs < 0 || nreqs < 0);
+
+                       if (hs->conn) {
+                               remain = hs->conn->mux->avail_streams(hs->conn);
+                               //if (remain <= 2)
+                                       DDPRINTF(stderr, "-%d", remain);
+                               if (!remain)
+                                       break;
+                       }
+                       else {
+                               connect_cnt++;
+                               /* Connecting */
+                               remain = 0;
+                               break;
+                       }
+               }
+
+               if (!remain || hld_next_url(urls, url) == first_url)
+                       break;
+       }
+
+ out:
+       DDPRINTF(stderr, "L(%d:%d)", tid, connect_cnt);
+       TRACE_LEAVE(HLD_EV_USR_TASK);
+       return t;
+}
+
+/* Instantiate a haload user and wake up its underlying task */
+static inline struct hld_usr *hld_new_usr(int nreqs)
+{
+       struct hld_usr *usr;
+       struct hld_url_cfg *cfg;
+       struct hld_url *url, *urls = NULL, *next_url;
+       struct task *t;
+       struct session *sess;
+
+       //usr = malloc(sizeof(*usr));
+       usr = calloc(1, sizeof(*usr));
+       /* XXX check this: always thead 0 when calling task_new_anywhere() */
+       t = task_new_on(conn_tid++ % global.nbthread);
+       //t = task_new_anywhere();
+       sess = session_new(&hld_proxy, NULL, NULL);
+       if (!usr || !t || !sess) {
+               ha_alert("could not allocate a new user\n");
+               goto err;
+       }
+
+       t->process = hld_usr_task;
+       t->context = usr;
+       t->expire = TICK_ETERNITY;
+
+       usr->task = t;
+       usr->sess = sess;
+       usr->flags = 0;
+       usr->urls = NULL;
+       LIST_INIT(&usr->strms);
+
+       for (cfg = hld_url_cfgs; cfg; cfg = cfg->next) {
+               struct hld_url *url;
+
+               url = malloc(sizeof(*url));
+               if (!url)
+                       goto err;
+
+               url->nreqs = nreqs;
+               url->mreqs = arg_mreqs;
+               url->flags = 0;
+               url->cfg = cfg;
+               url->next = usr->urls;
+               usr->urls = url;
+       }
+
+       /* inverse the URLs order */
+       url = usr->urls;
+       while (url) {
+               next_url = url->next;
+               url->next = urls;
+               urls = url;
+               url = next_url;
+       }
+       usr->urls = urls;
+
+       usr->cur_url = usr->urls;
+       task_wakeup(t, TASK_WOKEN_INIT);
+       return usr;
+
+ err:
+       url = usr->urls;
+       while (url) {
+               next_url = url->next;
+               free(url);
+               url = next_url;
+       }
+       task_destroy(t);
+       free(usr);
+       return NULL;
+}
+
+static int hld_cfg_finalize(void)
+{
+       int ret = 0;
+       struct hld_url_cfg *cfg;
+       const char *errptr = NULL;
+
+       for (cfg = hld_url_cfgs; cfg; cfg = cfg->next) {
+               struct server *srv;
+               struct sockaddr_storage *sk;
+               int alt_proto, port;
+               char *errmsg = NULL;
+               int arg = sizeof(hld_args) / sizeof(*hld_args);
+
+               /* Same as _srv_parse_init() from here */
+               srv = new_server(&hld_proxy);
+               if (!srv) {
+                       ha_alert("could not allocate a new server\n");
+                       goto leave;
+               }
+
+               sk = str2sa_range(cfg->addr, &port, NULL, NULL, NULL, NULL,
+                                                 &srv->addr_type, &errmsg, NULL, NULL, &alt_proto,
+                                                 PA_O_PORT_OK | PA_O_STREAM | PA_O_DGRAM | PA_O_XPRT);
+               if (!sk) {
+                       ha_alert("%s\n", errmsg);
+                       ha_free(&errmsg);
+                       goto leave;
+               }
+
+               srv->id = strdup("haload");
+               srv->addr = *sk;
+               srv->svc_port = port;
+               srv->alt_proto = alt_proto;
+               srv->use_ssl = cfg->ssl;
+               srv->xprt = srv_is_quic(srv) ? xprt_get(XPRT_QUIC) :
+                       srv->use_ssl ? xprt_get(XPRT_SSL) : xprt_get(XPRT_RAW);
+
+               if (srv_is_quic(srv))
+                       quic_transport_params_init(&srv->quic_params, 0);
+
+               /* XXX Must this be done? XXX */
+               //srv_set_addr_desc(srv, 0);
+               srv_settings_init(srv);
+
+               if (cfg->ssl_opts) {
+                       size_t outlen = 256;
+                       int cur_arg = 0;
+                       char *outline = malloc(256);
+                       uint32_t err;
+                       int err_code;
+
+                       err = parse_line(cfg->ssl_opts, outline, &outlen, hld_args, &arg,
+                                        PARSE_OPT_ENV    | PARSE_OPT_DQUOTE  |
+                                        PARSE_OPT_SQUOTE | PARSE_OPT_BKSLASH |
+                                        PARSE_OPT_SHARP  | PARSE_OPT_WORD_EXPAND, &errptr);
+                       if (err) {
+                               ha_alert("ssl opts parsing error\n");
+                               goto leave;
+                       }
+
+                       err_code = _srv_parse_kw(srv, hld_args, &cur_arg, &hld_proxy, 0);
+                       if (err_code)
+                               goto leave;
+
+                       free(outline);
+               }
+               /* Should parse server keywords here with _srv_parse_kw() */
+
+               /* Same as _srv_parse_finalize() from here */
+               if (srv_is_quic(srv)) {
+                       if (!srv->use_ssl)
+                               srv->use_ssl = 1;
+
+                       if (!srv->ssl_ctx.alpn_str) {
+                               srv->ssl_ctx.alpn_str = strdup("\002h3");
+                               if (!srv->ssl_ctx.alpn_str) {
+                                       ha_alert("could not allocate a default alpn.\n");
+                                       goto leave;
+                               }
+
+                               srv->ssl_ctx.alpn_len = strlen(srv->ssl_ctx.alpn_str);
+                       }
+               }
+
+               if (!srv->mux_proto && srv_is_quic(srv))
+                       srv->mux_proto = get_mux_proto(ist("quic"));
+
+               if (srv->mux_proto) {
+                       int proto_mode = conn_pr_mode_to_proto_mode(hld_proxy.mode);
+                       const struct mux_proto_list *mux_ent;
+
+                       mux_ent = conn_get_best_mux_entry(srv->mux_proto->token,
+                                                         PROTO_SIDE_BE,
+                                                         srv_is_quic(srv), proto_mode);
+
+                       if (!mux_ent || !isteq(mux_ent->token, srv->mux_proto->token)) {
+                               ha_alert("MUX protocol is not usable for server.\n");
+                               goto leave;
+                       }
+                       else {
+                               if ((mux_ent->mux->flags & MX_FL_FRAMED) && !srv_is_quic(srv)) {
+                                       ha_alert("MUX protocol is incompatible with stream"
+                                                " transport used by server.\n");
+                                       goto leave;
+                               }
+                               else if (!(mux_ent->mux->flags & MX_FL_FRAMED) && srv_is_quic(srv)) {
+                                       ha_alert("MUX protocol is incompatible with framed"
+                                                " transport used by server.\n");
+                                       goto leave;
+                               }
+                       }
+               }
+
+               /* ensure minconn/maxconn consistency */
+               srv_minmax_conn_apply(srv);
+
+               if (srv->use_ssl) {
+                       if (xprt_get(XPRT_SSL) && xprt_get(XPRT_SSL)->prepare_srv) {
+                               if (xprt_get(XPRT_SSL)->prepare_srv(srv))
+                                       goto leave;
+                       }
+                       /* XXX TO CHECK XXX: in fact XPRT_SSL and XPRT_QUIC have the same
+                        * ->prepare_srv() callback.
+                        */
+                       else if (xprt_get(XPRT_QUIC) && xprt_get(XPRT_QUIC)->prepare_srv) {
+                               if (xprt_get(XPRT_QUIC)->prepare_srv(srv))
+                                       goto leave;
+                       }
+               }
+
+               if (srv_preinit(srv))
+                       goto leave;
+#if 0
+               /* XXX Must this be done? XXX */
+               if (!srv_alloc_lb(srv, &hld_proxy)) {
+                       ha_alert("Failed to initialize load-balancing data.\n");
+                       goto leave;
+               }
+#endif
+
+               if (!stats_allocate_proxy_counters_internal(&srv->extra_counters,
+                                                           COUNTERS_SV, STATS_PX_CAP_SRV,
+                                                           &srv->per_tgrp->extra_counters_storage,
+                                                           &srv->per_tgrp[1].extra_counters_storage -
+                                                           &srv->per_tgrp[0].extra_counters_storage)) {
+                       ha_alert("failed to allocate extra counters for server.\n");
+                       goto leave;
+               }
+
+               if (srv_postinit(srv))
+                       goto leave;
+
+               /* Attach the server to the URL */
+               cfg->srv = srv;
+       }
+
+       ret = 1;
+leave:
+       return ret;
+}
+
+#if 0
+/* this is in order to cleanly stop on Ctrl-C */
+void sigint_handler(int sig)
+{
+       /* claim we're done */
+       //__sync_fetch_and_or(&running, THR_DUR_OVER);
+       stop_date = tv_ms_add(now, 500);
+       /* make sure a second Ctrl-C really stops */
+       signal(SIGINT, SIG_DFL);
+}
+#endif
+
+static int hld_init(void)
+{
+       int ret = ERR_ALERT | ERR_FATAL;
+       char *errmsg = NULL;
+
+       if (arg_slow)
+               throttle = 1;
+
+       if (!hld_cfg_finalize())
+               goto leave;
+
+       usr_reqs = arg_reqs;
+       min_reqs = usr_reqs > 0 ? usr_reqs / arg_usr : -1;
+       mod_req = usr_reqs > 0 ? usr_reqs % arg_usr : -1;
+
+       mtask.t = task_new_anywhere();
+       if (mtask.t == NULL) {
+               ha_alert("could start main task\n");
+               goto leave;
+       }
+
+       if (arg_long >= 2)
+               printf("#_____time conns tot_conn  tot_req      tot_bytes"
+                      "    err thr cps rps Bps bps ttfb(us) ttlb(us)");
+       else if (arg_long)
+               printf("#     time conns tot_conn  tot_req      tot_bytes"
+                      "    err  cps  rps  Bps  bps   ttfb   ttlb");
+       else
+               printf("#     time conns tot_conn  tot_req      tot_bytes"
+                      "    err  cps  rps  bps   ttfb");
+       putchar('\n');
+
+       mtask.t->process = mtask_cb;
+       mtask.t->expire = TICK_ETERNITY;
+       mtask.show_time = tick_add(now_ms, MS_TO_TICKS(1000));
+       task_wakeup(mtask.t, TASK_WOKEN_INIT);
+
+       gettimeofday(&hld_start_date, NULL);
+       ret = ERR_NONE;
+ leave:
+       ha_free(&errmsg);
+       return ret;
+}
+REGISTER_POST_CHECK(hld_init);
+
+static int hld_alloc_thrs_info(void)
+{
+       fprintf(stderr, "NBTHREAD=%d\n", global.nbthread);
+       thrs_info = calloc(global.nbthread, sizeof(*thrs_info));
+       if (!thrs_info) {
+               ha_alert("failed to alloct threads information array.\n");
+               return -1;
+       }
+
+       return 1;
+}
+REGISTER_POST_CHECK(hld_alloc_thrs_info);
diff --git a/src/haload_init.c b/src/haload_init.c
new file mode 100644 (file)
index 0000000..de277a3
--- /dev/null
@@ -0,0 +1,476 @@
+#include <errno.h>
+
+#include <haproxy/connection.h>
+#include <haproxy/errors.h>
+#include <haproxy/global.h>
+#include <haproxy/hbuf.h>
+#include <haproxy/haload.h>
+#include <haproxy/proxy.h>
+#include <haproxy/version.h>
+#include <haproxy/server.h>
+
+static int hld_debug;
+struct hld_url_cfg *hld_url_cfgs;
+char *ssl_opts;
+
+__attribute__((unused))
+static const char *hld_cfg_dflt_str =
+       "defaults\n"
+               "\tmode http\n"
+               "\thttp-reuse never\n"
+               "\ttimeout connect 5s\n"
+               "\ttimeout server 40s\n";
+
+static void  hld_usage(char *name, int argc, int line)
+{
+       fprintf(stderr, "argc=%d\n", argc);
+       fprintf(stderr,
+               "%d: Usage : %s [opts]\n"
+               "        -d <time>      test duration in seconds (0)\n"
+               "        -l             enable long output format; double for raw values\n"
+               "        -n <reqs>      maximum total requests (-1)\n"
+               "        -r <reqs>      number of requests per connection (-1)\n"
+               "        -u <users>     number of users (1)\n"
+               "        -w <time>      I/O timeout in milliseconds (10000)\n"
+               "        -C             dump the configuration and exit\n"
+               "        -H \"foo:bar\"   add this header name and value\n"
+               "        -I             use HEAD instead of GET\n"
+               "        -v             shows version\n"
+               "        --traces       enable the traces for all the HTTP protocols\n"
+               "where <opts> may be any combination of:\n",
+               line, name);
+       exit(1);
+}
+
+static const char *hld_cfg_traces_str =
+       "traces\n"
+               "\ttrace ssl sink stdout level developer start now verbosity clean\n"
+               "\ttrace haload sink stdout level developer start now verbosity clean\n"
+               "\ttrace quic sink stdout level developer start now\n"
+               "\ttrace h1 sink stdout level developer start now verbosity minimal\n"
+               "\ttrace h2 sink stdout level developer start now verbosity minimal\n"
+               "\ttrace h3 sink stdout level developer start now verbosity minimal\n"
+               "\ttrace qmux sink stdout level developer start now verbosity minimal\n";
+
+static struct hld_hdr *hld_parse_hdr(char *hdr_str)
+{
+       struct hld_hdr *hdr= NULL;
+       char *value = strchr(hdr_str, ':');
+
+       if (value) {
+               *value++ = '\0';
+               if (!*value)
+                       value = NULL;
+       }
+
+       if (strcasecmp(hdr_str, "host") == 0)
+               arg_host = value;
+       else if (strcasecmp(hdr_str, "connection") == 0)
+               arg_conn_hdr = value;
+
+       hdr = malloc(sizeof(*hdr));
+       if (hdr) {
+               hdr->name = ist(hdr_str);
+               hdr->value = ist(value);
+       }
+
+       return hdr;
+}
+
+/* Allocate a URL from <url> command line argument without
+ * duplicating it, and append it to <hld_url_cfgs> list of URL.
+ * A URL is identified by its peer address and if it uses
+ * SSL or not. When a URL with the same peer address already
+ * exists, this function only add a new path to the list
+ * of paths attaached to the URL.
+ * Return the URL if succeeded, NULL if not.
+ */
+static struct hld_url_cfg *hld_alloc_url(char *url)
+{
+       int ssl = 0;
+       char *addr = NULL, *raw_addr = NULL, *path = NULL;
+       struct hld_url_cfg *hld_url_cfg = NULL;
+       struct hld_url_cfg *purl;
+       struct hld_path *p;
+
+       if (strncmp(url, "http://", 7) == 0)
+               addr = url + 7;
+       else if (strncmp(url, "https://", 8) == 0) {
+               ssl = 1;
+               addr = url + 8;
+       }
+       else
+               addr = url;
+
+       path = strchr(addr, '/');
+       if (path) {
+               char *new_path = strdup(path);
+               *path = '\0';
+               path = new_path;
+       }
+       else
+               path = strdup("/");
+
+       if (!path)
+               goto err;
+
+       for (purl = hld_url_cfgs; purl; purl = purl->next) {
+               /* XXX TODO: improve this check for QUIC XXX */
+               if (strcmp(purl->addr, addr) == 0 && purl->ssl == ssl) {
+                       /* Already existing URL with the same address. */
+                       hld_url_cfg = purl;
+
+                       p = malloc(sizeof(*p));
+                       if (!p)
+                               goto err;
+
+                       /* Append a new path to this URL */
+                       p->path = path;
+                       p->next = hld_url_cfg->paths;
+                       hld_url_cfg->paths = p;
+
+                       return hld_url_cfg;
+               }
+       }
+
+       addr = strdup(addr);
+       if (!addr)
+               goto err;
+
+       raw_addr = strchr(addr, '@');
+       if (!raw_addr)
+               raw_addr = strchr(addr, '+');
+
+       raw_addr = raw_addr ? raw_addr + 1: addr;
+
+       hld_url_cfg = malloc(sizeof(*hld_url_cfg));
+       p = malloc(sizeof(*p));
+       if (!hld_url_cfg || !p)
+               goto err;
+
+       p->path = path;
+       p->next = NULL;
+
+       hld_url_cfg->ssl = ssl;
+       hld_url_cfg->addr = addr;
+       hld_url_cfg->raw_addr = raw_addr;
+
+       if (ssl_opts) {
+               hld_url_cfg->ssl_opts = ssl_opts;
+               ssl_opts = NULL;
+       }
+       else
+               hld_url_cfg->ssl_opts = NULL;
+
+       hld_url_cfg->srv = NULL;
+       hld_url_cfg->paths = p;
+       /* Append this new URL to the list */
+       hld_url_cfg->next = hld_url_cfgs;
+       hld_url_cfgs = hld_url_cfg;
+
+       return hld_url_cfg;
+ err:
+       free(addr);
+       free(path);
+       return NULL;
+}
+
+static void hld_parse_long(int *val, char *opt, int *argc, char ***argv)
+{
+       char *endptr;
+
+       if (!*opt) {
+               ++*argv; --*argc;
+               if (*argc <= 0 || ***argv == '-')
+                       hld_usage(progname, *argc, __LINE__);
+
+               opt = **argv;
+       }
+
+       *val = strtol(opt, &endptr, 0);
+       if (endptr == opt || *val < 0)
+               hld_usage(progname, *argc, __LINE__);
+}
+
+static inline void hld_url_cfgs_inv(void)
+{
+       struct hld_url_cfg *urls = NULL, *url = hld_url_cfgs, *next_url;
+
+       /* inverse the URLs order */
+       while (url) {
+               struct hld_path *paths = NULL, *path = url->paths, *next_path;
+
+               /* inverse the paths order */
+               while (path) {
+                       next_path = path->next;
+                       path->next = paths;
+                       paths = path;
+                       path = next_path;
+               }
+               url->paths = paths;
+
+               next_url = url->next;
+               url->next = urls;
+               urls = url;
+               url = next_url;
+       }
+
+       hld_url_cfgs = urls;
+}
+
+
+void haproxy_init_args(int argc, char **argv)
+{
+       int err = 1, dump = 0;
+       struct hbuf buf = HBUF_NULL;  // cfgfile
+       struct hbuf gbuf = HBUF_NULL; // "global" section
+       struct hbuf tbuf = HBUF_NULL; // "traces" section
+
+       if (argc <= 1)
+               hld_usage(progname, argc, __LINE__);
+
+       if (hbuf_alloc(&gbuf) == NULL) {
+               ha_alert("failed to allocate a buffer.\n");
+               goto leave;
+       }
+
+       /* use 3MB of local cache per thread mainly for QUIC.
+        * Also trust the server certificates
+        */
+       hbuf_appendf(&gbuf, "global\n");
+       hbuf_appendf(&gbuf, "\ttune.memory.hot-size 3145728\n");
+       hbuf_appendf(&gbuf, "\tssl-server-verify none\n");
+
+       if (hbuf_alloc(&buf) == NULL) {
+               ha_alert("failed to allocate a buffer\n");
+               exit(1);
+       }
+
+       fileless_mode = 1;
+       client_mode = 1;
+       /* skip program name and start */
+       argc--; argv++;
+
+       while (argc > 0) {
+               if (**argv == '-') {
+                       char *opt = *argv + 1;
+
+                       //fprintf(stderr, "||||**argv='%c' argc=%d\n", **argv, argc);
+                       //fprintf(stderr, "====> *opt='%c'\n", *opt);
+                       if (*opt == '-') {
+                               /* long option */
+                               opt++;
+                               if (strcmp(opt, "traces") == 0) {
+                                       hld_debug = 1;
+#if 0
+                                       /* optional argument */
+                                       if (argc - 1 > 0 && **(argv + 1) != '-') {
+                                               fprintf(stderr, "trace arg: '%s'\n", *(argv + 1));
+                                               if (hbuf_is_null(&tbuf) && hbuf_alloc(&tbuf) == NULL) {
+                                                       ha_alert("failed to allocate a buffer.\n");
+                                                       goto leave;
+                                               }
+
+                                               hbuf_str_append(&tbuf, *(argv + 1));
+                                               argc--; argv++;
+                                       }
+#endif
+                               }
+                               else
+                                       hld_usage(progname, argc, __LINE__);
+                       }
+                       else if (*opt == 'd') {
+                               opt++;
+                               hld_parse_long(&arg_dura, opt, &argc, &argv);
+                       }
+                       else if (*opt == 'l') {
+                               arg_long++;
+                               while (*++opt && *opt == 'l')
+                                       arg_long++;
+                       }
+                       else if (*opt == 'm') {
+                               opt++;
+                               hld_parse_long(&arg_mreqs, opt, &argc, &argv);
+                       }
+                       else if (*opt == 'n') {
+                               opt++;
+                               hld_parse_long(&arg_reqs, opt, &argc, &argv);
+                       }
+                       else if (*opt == 'r') {
+                               opt++;
+                               hld_parse_long(&arg_rcon, opt, &argc, &argv);
+                       }
+                       else if (*opt == 's') {
+                               opt++;
+                               hld_parse_long(&arg_slow, opt, &argc, &argv);
+                               arg_slow *= 1000;
+                       }
+                       else if (*opt == 'u') {
+                               opt++;
+                               hld_parse_long(&arg_usr, opt, &argc, &argv);
+                       }
+                       else if (*opt == 'w') {
+                               opt++;
+                               hld_parse_long(&arg_wait, opt, &argc, &argv);
+                       }
+                       else if (*opt == 'C') {
+                               /* empty option */
+                               if (*(opt + 1))
+                                       hld_usage(progname, argc, __LINE__);
+
+                               dump = 1;
+                       }
+                       else if (*opt == 'H') {
+                               char *hdr_str;
+                               struct hld_hdr *hdr;
+
+                               opt++;
+                               if (!*opt) {
+                                       argv++; argc--;
+                                       if ((argc <= 0 || **argv == '-'))
+                                               hld_usage(progname, argc, __LINE__);
+
+                                       opt = *argv;
+                               }
+
+                               hdr_str = opt;
+                               hdr = hld_parse_hdr(hdr_str);
+                               if (!hdr) {
+                                       ha_alert("could not allocate a header\n");
+                                       goto leave;
+                               }
+
+                               LIST_APPEND(&hld_hdrs, &hdr->list);
+                       }
+                       else if (*opt == 'I') {
+                               /* empty option */
+                               if (*(opt + 1))
+                                       hld_usage(progname, argc, __LINE__);
+
+                               arg_head = 1;
+                       }
+                       else if (*opt == 'G') {
+                               argv++; argc--;
+                               if (argc <= 0 || **argv == '-')
+                                       hld_usage(progname, argc, __LINE__);
+
+                               hbuf_str_append(&gbuf, *argv);
+                       }
+                       else if (*opt == 'S') {
+                               opt++;
+                               if (!*opt) {
+                                       argv++, argc--;
+
+                                       if ((argc <= 0 || **argv == '-'))
+                                               hld_usage(progname, argc, __LINE__);
+
+                                       opt = *argv;
+                               }
+                               ssl_opts = strdup(opt);
+                       }
+                       else if (*opt == 'v') {
+                               /* empty option */
+                               if (*(opt + 1))
+                                       hld_usage(progname, argc, __LINE__);
+
+                               printf("haload version " HAPROXY_VERSION " released " HAPROXY_DATE "\n");
+                               exit(0);
+                       }
+                       else
+                               hld_usage(progname, argc, __LINE__);
+               }
+               else {
+                       struct hld_url_cfg *url;
+
+                       url = hld_alloc_url(*argv);
+                       if (!url) {
+                               ha_alert("could not parse a new URL\n");
+                               goto leave;
+                       }
+
+
+               }
+
+               argv++; argc--;
+               //fprintf(stderr, "///argc=%d argv='%s'\n", argc, *argv);
+       }
+
+       if (!hld_url_cfgs) {
+               ha_alert("no URL provided\n");
+               goto leave;
+       }
+
+       /* "global" section */
+       if (!hbuf_is_null(&gbuf))
+               hbuf_appendf(&buf, "%.*s\n", (int)gbuf.data, gbuf.area);
+       /* "traces" section */
+       if (hld_debug) {
+               hbuf_appendf(&buf, "%s", hld_cfg_traces_str);
+               if (!hbuf_is_null(&tbuf))
+                       hbuf_appendf(&buf, "%.*s\n", (int)tbuf.data, tbuf.area);
+       }
+
+       //hbuf_appendf(&buf, "%s\n", hld_cfg_dflt_str);
+
+       fileless_cfg.filename = strdup("haterm cfgfile");
+       fileless_cfg.content = strdup(buf.area);
+       if (!fileless_cfg.filename || !fileless_cfg.content) {
+               ha_alert("cfgfile strdup() failed.\n");
+               goto leave;
+       }
+
+       fileless_cfg.size = buf.data;
+
+       /* Config dump */
+       if (dump) {
+               fprintf(stdout, "%.*s", (int)fileless_cfg.size, fileless_cfg.content);
+               exit(0);
+       }
+
+       if (arg_reqs > 0 && arg_reqs < arg_usr) {
+               ha_alert("user count must not exceed request count\n");
+               goto leave;
+       }
+
+       /* Inverse the URLs and their paths */
+       hld_url_cfgs_inv();
+       err = 0;
+leave:
+       free_hbuf(&gbuf);
+       free_hbuf(&buf);
+       if (err)
+               exit(1);
+}
+
+/* Dummy argv copier function */
+char **copy_argv(int argc, char **argv)
+{
+       char **ret = calloc(1, sizeof(*ret));
+
+       if (ret)
+               *ret = strdup("");
+
+       return ret;
+}
+
+static int hld_pre_check(void)
+{
+       char *errmsg = NULL;
+
+       if (!setup_new_proxy(&hld_proxy, "<HALOAD-BE>",
+                            PR_CAP_FE | PR_CAP_BE | PR_CAP_INT, &errmsg)) {
+               ha_alert("could not setup internal proxy: %s\n", errmsg);
+               ha_free(&errmsg);
+               return ERR_FATAL;
+       }
+
+       hld_proxy.mode = PR_MODE_HTTP;
+       hld_proxy.next = proxies_list;
+    proxies_list = &hld_proxy;
+
+    hld_proxy.timeout.server = 20000;
+    hld_proxy.timeout.connect = 25000;
+
+       return ERR_NONE;
+}
+REGISTER_PRE_CHECK(hld_pre_check);