diff options
Diffstat (limited to 'erts/emulator/beam/dist.c')
| -rw-r--r-- | erts/emulator/beam/dist.c | 2013 |
1 files changed, 1388 insertions, 625 deletions
diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c index 9fddac8980..8593ad867a 100644 --- a/erts/emulator/beam/dist.c +++ b/erts/emulator/beam/dist.c @@ -103,25 +103,16 @@ dist_msg_dbg(ErtsDistExternal *edep, char *what, byte *buf, int sz) -#define PASS_THROUGH 'p' /* This code should go */ - int erts_is_alive; /* System must be blocked on change */ int erts_dist_buf_busy_limit; /* distribution trap functions */ -Export* dsend2_trap = NULL; -Export* dsend3_trap = NULL; -/*Export* dsend_nosuspend_trap = NULL;*/ -Export* dlink_trap = NULL; -Export* dunlink_trap = NULL; Export* dmonitor_node_trap = NULL; -Export* dgroup_leader_trap = NULL; -Export* dexit_trap = NULL; Export* dmonitor_p_trap = NULL; /* local variables */ - +static Export *dist_ctrl_put_data_trap; /* forward declarations */ @@ -129,9 +120,10 @@ static void clear_dist_entry(DistEntry*); static int dsig_send_ctl(ErtsDSigData* dsdp, Eterm ctl, int force_busy); static void send_nodes_mon_msgs(Process *, Eterm, Eterm, Eterm, Eterm); static void init_nodes_monitors(void); +static Sint abort_connection(DistEntry* dep, Uint32 conn_id); -static erts_smp_atomic_t no_caches; -static erts_smp_atomic_t no_nodes; +static erts_atomic_t no_caches; +static erts_atomic_t no_nodes; struct { Eterm reason; @@ -144,8 +136,8 @@ delete_cache(ErtsAtomCache *cache) { if (cache) { erts_free(ERTS_ALC_T_DCACHE, (void *) cache); - ASSERT(erts_smp_atomic_read_nob(&no_caches) > 0); - erts_smp_atomic_dec_nob(&no_caches); + ASSERT(erts_atomic_read_nob(&no_caches) > 0); + erts_atomic_dec_nob(&no_caches); } } @@ -156,14 +148,12 @@ create_cache(DistEntry *dep) int i; ErtsAtomCache *cp; - ERTS_SMP_LC_ASSERT( - is_internal_port(dep->cid) - && erts_lc_is_port_locked(erts_port_lookup_raw(dep->cid))); + ERTS_LC_ASSERT(is_nil(dep->cid)); ASSERT(!dep->cache); dep->cache = cp = (ErtsAtomCache*) erts_alloc(ERTS_ALC_T_DCACHE, sizeof(ErtsAtomCache)); - erts_smp_atomic_inc_nob(&no_caches); + erts_atomic_inc_nob(&no_caches); for (i = 0; i < sizeof(cp->in_arr)/sizeof(cp->in_arr[0]); i++) { cp->in_arr[i] = THE_NON_VALUE; cp->out_arr[i] = THE_NON_VALUE; @@ -172,15 +162,17 @@ create_cache(DistEntry *dep) Uint erts_dist_cache_size(void) { - return (Uint) erts_smp_atomic_read_mb(&no_caches)*sizeof(ErtsAtomCache); + return (Uint) erts_atomic_read_mb(&no_caches)*sizeof(ErtsAtomCache); } static ErtsProcList * -get_suspended_on_de(DistEntry *dep, Uint32 unset_qflgs) +get_suspended_on_de(DistEntry *dep, erts_aint32_t unset_qflgs) { - ERTS_SMP_LC_ASSERT(erts_smp_lc_mtx_is_locked(&dep->qlock)); - dep->qflgs &= ~unset_qflgs; - if (dep->qflgs & ERTS_DE_QFLG_EXIT) { + erts_aint32_t qflgs; + ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&dep->qlock)); + qflgs = erts_atomic32_read_band_acqb(&dep->qflgs, ~unset_qflgs); + qflgs &= ~unset_qflgs; + if (qflgs & ERTS_DE_QFLG_EXIT) { /* No resume when exit has been scheduled */ return NULL; } @@ -283,17 +275,15 @@ static void doit_monitor_net_exits(ErtsMonitor *mon, void *vnecp) watched = (is_atom(rmon->name) ? TUPLE2(lhp, rmon->name, dep->sysname) : rmon->u.pid); -#ifdef ERTS_SMP rp_locks |= ERTS_PROC_LOCKS_MSG_SEND; - erts_smp_proc_lock(rp, ERTS_PROC_LOCKS_MSG_SEND); -#endif + erts_proc_lock(rp, ERTS_PROC_LOCKS_MSG_SEND); erts_queue_monitor_message(rp, &rp_locks, mon->ref, am_process, watched, am_noconnection); erts_destroy_monitor(rmon); } UnUseTmpHeapNoproc(3); } - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); done: erts_destroy_monitor(mon); } @@ -342,7 +332,7 @@ static void doit_link_net_exits_sub(ErtsLink *sublnk, void *vlnecp) trace_proc(NULL, 0, rp, am_getting_unlinked, sublnk->pid); } } - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } done: erts_destroy_link(sublnk); @@ -384,7 +374,7 @@ static void doit_node_link_net_exits(ErtsLink *lnk, void *vnecp) rp = erts_proc_lookup(lnk->pid); if (!rp) goto done; - erts_smp_proc_lock(rp, rp_locks); + erts_proc_lock(rp, rp_locks); if (!ERTS_PROC_IS_EXITING(rp)) { rlnk = erts_remove_link(&ERTS_P_LINKS(rp), name); if (rlnk != NULL) { @@ -403,7 +393,7 @@ static void doit_node_link_net_exits(ErtsLink *lnk, void *vnecp) erts_queue_message(rp, rp_locks, msgp, tup, am_system); } } - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } done: erts_destroy_link(lnk); @@ -415,16 +405,16 @@ set_node_not_alive(void *unused) ErlHeapFragment *bp; Eterm nodename = erts_this_dist_entry->sysname; - ASSERT(erts_smp_atomic_read_nob(&no_nodes) == 0); + ASSERT(erts_atomic_read_nob(&no_nodes) == 0); - erts_smp_thr_progress_block(); + erts_thr_progress_block(); erts_set_this_node(am_Noname, 0); erts_is_alive = 0; send_nodes_mon_msgs(NULL, am_nodedown, nodename, am_visible, nodedown.reason); nodedown.reason = NIL; bp = nodedown.bp; nodedown.bp = NULL; - erts_smp_thr_progress_unblock(); + erts_thr_progress_unblock(); if (bp) free_message_buffer(bp); } @@ -432,7 +422,7 @@ set_node_not_alive(void *unused) static ERTS_INLINE void dec_no_nodes(void) { - erts_aint_t no = erts_smp_atomic_dec_read_mb(&no_nodes); + erts_aint_t no = erts_atomic_dec_read_mb(&no_nodes); ASSERT(no >= 0); ASSERT(erts_get_scheduler_id()); /* Need to be a scheduler */ if (no == 0) @@ -445,12 +435,40 @@ static ERTS_INLINE void inc_no_nodes(void) { #ifdef DEBUG - erts_aint_t no = erts_smp_atomic_read_nob(&no_nodes); + erts_aint_t no = erts_atomic_read_nob(&no_nodes); ASSERT(erts_is_alive ? no > 0 : no == 0); #endif - erts_smp_atomic_inc_mb(&no_nodes); + erts_atomic_inc_mb(&no_nodes); } - + +static void +kill_dist_ctrl_proc(void *vpid) +{ + Eterm pid = (Eterm) vpid; + ErtsProcLocks rp_locks = ERTS_PROC_LOCKS_XSIG_SEND; + Process *rp = erts_pid2proc(NULL, 0, pid, rp_locks); + if (rp) { + erts_send_exit_signal(NULL, rp->common.id, rp, &rp_locks, + am_kill, NIL, NULL, 0); + if (rp_locks) + erts_proc_unlock(rp, rp_locks); + } +} + +static void +schedule_kill_dist_ctrl_proc(Eterm pid) +{ + ErtsSchedulerData *esdp = erts_get_scheduler_data(); + int sched_id = 1; + if (!esdp || ERTS_SCHEDULER_IS_DIRTY(esdp)) + sched_id = 1; + else + sched_id = (int) esdp->no; + erts_schedule_misc_aux_work(sched_id, + kill_dist_ctrl_proc, + (void *) (UWord) pid); +} + /* * proc is currently running or exiting process. */ @@ -460,58 +478,83 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) if (dep == erts_this_dist_entry) { /* Net kernel has died (clean up!!) */ DistEntry *tdep; - int no_dist_port = 0; + int no_dist_ctrl; + int no_pending; Eterm nd_reason = (reason == am_no_network ? am_no_network : am_net_kernel_terminated); - erts_smp_rwmtx_rlock(&erts_dist_table_rwmtx); + int i = 0; + Eterm *dist_ctrl; + DistEntry** pending; + + ERTS_UNDEF(dist_ctrl, NULL); + ERTS_UNDEF(pending, NULL); + + erts_rwmtx_rlock(&erts_dist_table_rwmtx); - for (tdep = erts_hidden_dist_entries; tdep; tdep = tdep->next) - no_dist_port++; - for (tdep = erts_visible_dist_entries; tdep; tdep = tdep->next) - no_dist_port++; + no_dist_ctrl = (erts_no_of_hidden_dist_entries + + erts_no_of_visible_dist_entries); + no_pending = erts_no_of_pending_dist_entries; /* KILL all port controllers */ - if (no_dist_port == 0) - erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx); - else { - Eterm def_buf[128]; - int i = 0; - Eterm *dist_port; - - if (no_dist_port <= sizeof(def_buf)/sizeof(def_buf[0])) - dist_port = &def_buf[0]; - else - dist_port = erts_alloc(ERTS_ALC_T_TMP, - sizeof(Eterm)*no_dist_port); + if (no_dist_ctrl) { + dist_ctrl = erts_alloc(ERTS_ALC_T_TMP, + sizeof(Eterm)*no_dist_ctrl); for (tdep = erts_hidden_dist_entries; tdep; tdep = tdep->next) { - ASSERT(is_internal_port(tdep->cid)); - dist_port[i++] = tdep->cid; + ASSERT(is_internal_port(tdep->cid) || is_internal_pid(tdep->cid)); + ASSERT(i < no_dist_ctrl); + dist_ctrl[i++] = tdep->cid; } for (tdep = erts_visible_dist_entries; tdep; tdep = tdep->next) { - ASSERT(is_internal_port(tdep->cid)); - dist_port[i++] = tdep->cid; + ASSERT(is_internal_port(tdep->cid) || is_internal_pid(tdep->cid)); + ASSERT(i < no_dist_ctrl); + dist_ctrl[i++] = tdep->cid; } - erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx); - - for (i = 0; i < no_dist_port; i++) { - Port *prt = erts_port_lookup(dist_port[i], - ERTS_PORT_SFLGS_INVALID_LOOKUP); - if (!prt) - continue; - ASSERT(erts_atomic32_read_nob(&prt->state) - & ERTS_PORT_SFLG_DISTRIBUTION); + ASSERT(i == no_dist_ctrl); + } + if (no_pending) { + pending = erts_alloc(ERTS_ALC_T_TMP, sizeof(DistEntry*)*no_pending); + i = 0; + for (tdep = erts_pending_dist_entries; tdep; tdep = tdep->next) { + ASSERT(is_nil(tdep->cid)); + ASSERT(i < no_pending); + pending[i++] = tdep; + erts_ref_dist_entry(tdep); + } + ASSERT(i == no_pending); + } + erts_rwmtx_runlock(&erts_dist_table_rwmtx); + + if (no_dist_ctrl) { + for (i = 0; i < no_dist_ctrl; i++) { + if (is_internal_pid(dist_ctrl[i])) + schedule_kill_dist_ctrl_proc(dist_ctrl[i]); + else { + Port *prt = erts_port_lookup(dist_ctrl[i], + ERTS_PORT_SFLGS_INVALID_LOOKUP); + if (prt) { + ASSERT(erts_atomic32_read_nob(&prt->state) + & ERTS_PORT_SFLG_DISTRIBUTION); + + erts_port_exit(NULL, ERTS_PORT_SIG_FLG_FORCE_SCHED, + prt, dist_ctrl[i], nd_reason, NULL); + } + } + } + erts_free(ERTS_ALC_T_TMP, dist_ctrl); + } - erts_port_exit(NULL, ERTS_PORT_SIG_FLG_FORCE_SCHED, - prt, dist_port[i], nd_reason, NULL); - } + if (no_pending) { + for (i = 0; i < no_pending; i++) { + abort_connection(pending[i], pending[i]->connection_id); + erts_deref_dist_entry(pending[i]); + } + erts_free(ERTS_ALC_T_TMP, pending); + } - if (dist_port != &def_buf[0]) - erts_free(ERTS_ALC_T_TMP, dist_port); - } /* - * When last dist port exits, node will be taken + * When last dist ctrl exits, node will be taken * from alive to not alive. */ ASSERT(is_nil(nodedown.reason) && !nodedown.bp); @@ -528,52 +571,51 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) &nodedown.bp->off_heap); } } - else { /* Call from distribution port */ + else { /* Call from distribution controller (port/process) */ NetExitsContext nec = {dep}; ErtsLink *nlinks; ErtsLink *node_links; ErtsMonitor *monitors; Uint32 flags; - erts_smp_atomic_set_mb(&dep->dist_cmd_scheduled, 1); - erts_smp_de_rwlock(dep); + erts_atomic_set_mb(&dep->dist_cmd_scheduled, 1); + erts_de_rwlock(dep); - ERTS_SMP_LC_ASSERT(is_internal_port(dep->cid) - && erts_lc_is_port_locked(erts_port_lookup_raw(dep->cid))); + if (is_internal_port(dep->cid)) { + ERTS_LC_ASSERT(erts_lc_is_port_locked(erts_port_lookup_raw(dep->cid))); - if (erts_port_task_is_scheduled(&dep->dist_cmd)) - erts_port_task_abort(&dep->dist_cmd); + if (erts_port_task_is_scheduled(&dep->dist_cmd)) + erts_port_task_abort(&dep->dist_cmd); + } if (dep->status & ERTS_DE_SFLG_EXITING) { #ifdef DEBUG - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qflgs & ERTS_DE_QFLG_EXIT); - erts_smp_mtx_unlock(&dep->qlock); + ASSERT(erts_atomic32_read_nob(&dep->qflgs) & ERTS_DE_QFLG_EXIT); #endif } else { dep->status |= ERTS_DE_SFLG_EXITING; - erts_smp_mtx_lock(&dep->qlock); - ASSERT(!(dep->qflgs & ERTS_DE_QFLG_EXIT)); - dep->qflgs |= ERTS_DE_QFLG_EXIT; - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_lock(&dep->qlock); + ASSERT(!(erts_atomic32_read_nob(&dep->qflgs) & ERTS_DE_QFLG_EXIT)); + erts_atomic32_read_bor_relb(&dep->qflgs, ERTS_DE_QFLG_EXIT); + erts_mtx_unlock(&dep->qlock); } - erts_smp_de_links_lock(dep); + erts_de_links_lock(dep); monitors = dep->monitors; nlinks = dep->nlinks; node_links = dep->node_links; dep->monitors = NULL; dep->nlinks = NULL; dep->node_links = NULL; - erts_smp_de_links_unlock(dep); + erts_de_links_unlock(dep); nodename = dep->sysname; flags = dep->flags; erts_set_dist_entry_not_connected(dep); - erts_smp_de_rwunlock(dep); + erts_de_rwunlock(dep); erts_sweep_monitors(monitors, &doit_monitor_net_exits, (void *) &nec); erts_sweep_links(nlinks, &doit_link_net_exits, (void *) &nec); @@ -586,7 +628,6 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) reason == am_normal ? am_connection_closed : reason); clear_dist_entry(dep); - } dec_no_nodes(); @@ -607,19 +648,15 @@ void init_dist(void) nodedown.reason = NIL; nodedown.bp = NULL; - erts_smp_atomic_init_nob(&no_nodes, 0); - erts_smp_atomic_init_nob(&no_caches, 0); + erts_atomic_init_nob(&no_nodes, 0); + erts_atomic_init_nob(&no_caches, 0); /* Lookup/Install all references to trap functions */ - dsend2_trap = trap_function(am_dsend,2); - dsend3_trap = trap_function(am_dsend,3); - /* dsend_nosuspend_trap = trap_function(am_dsend_nosuspend,2);*/ - dlink_trap = trap_function(am_dlink,1); - dunlink_trap = trap_function(am_dunlink,1); dmonitor_node_trap = trap_function(am_dmonitor_node,3); - dgroup_leader_trap = trap_function(am_dgroup_leader,2); - dexit_trap = trap_function(am_dexit, 2); dmonitor_p_trap = trap_function(am_dmonitor_p, 2); + dist_ctrl_put_data_trap = erts_export_put(am_erts_internal, + am_dist_ctrl_put_data, + 2); } #define ErtsDistOutputBuf2Binary(OB) \ @@ -634,6 +671,7 @@ alloc_dist_obuf(Uint size) obuf = (ErtsDistOutputBuf *) &bin->orig_bytes[0]; #ifdef DEBUG obuf->dbg_pattern = ERTS_DIST_OUTPUT_BUF_DBG_PATTERN; + obuf->alloc_endp = obuf->data + size; ASSERT(bin == ErtsDistOutputBuf2Binary(obuf)); #endif return obuf; @@ -654,26 +692,11 @@ size_obuf(ErtsDistOutputBuf *obuf) return bin->orig_size; } -static void clear_dist_entry(DistEntry *dep) +static ErtsDistOutputBuf* clear_de_out_queues(DistEntry* dep) { - Sint obufsize = 0; - ErtsAtomCache *cache; - ErtsProcList *suspendees; ErtsDistOutputBuf *obuf; - erts_smp_de_rwlock(dep); - cache = dep->cache; - dep->cache = NULL; - -#ifdef DEBUG - erts_smp_de_links_lock(dep); - ASSERT(!dep->nlinks); - ASSERT(!dep->node_links); - ASSERT(!dep->monitors); - erts_smp_de_links_unlock(dep); -#endif - - erts_smp_mtx_lock(&dep->qlock); + ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&dep->qlock)); if (!dep->out_queue.last) obuf = dep->finalized_out_queue.first; @@ -682,21 +705,24 @@ static void clear_dist_entry(DistEntry *dep) obuf = dep->out_queue.first; } + if (dep->tmp_out_queue.first) { + dep->tmp_out_queue.last->next = obuf; + obuf = dep->tmp_out_queue.first; + } + dep->out_queue.first = NULL; dep->out_queue.last = NULL; + dep->tmp_out_queue.first = NULL; + dep->tmp_out_queue.last = NULL; dep->finalized_out_queue.first = NULL; dep->finalized_out_queue.last = NULL; - dep->status = 0; - suspendees = get_suspended_on_de(dep, ERTS_DE_QFLGS_ALL); - - erts_smp_mtx_unlock(&dep->qlock); - erts_smp_atomic_set_nob(&dep->dist_cmd_scheduled, 0); - dep->send = NULL; - erts_smp_de_rwunlock(dep); - erts_resume_processes(suspendees); + return obuf; +} - delete_cache(cache); +static void free_de_out_queues(DistEntry* dep, ErtsDistOutputBuf *obuf) +{ + Sint obufsize = 0; while (obuf) { ErtsDistOutputBuf *fobuf; @@ -707,13 +733,57 @@ static void clear_dist_entry(DistEntry *dep) } if (obufsize) { - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qsize >= obufsize); - dep->qsize -= obufsize; - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_lock(&dep->qlock); + ASSERT(erts_atomic_read_nob(&dep->qsize) >= obufsize); + erts_atomic_add_nob(&dep->qsize, + (erts_aint_t) -obufsize); + erts_mtx_unlock(&dep->qlock); } } +static void clear_dist_entry(DistEntry *dep) +{ + ErtsAtomCache *cache; + ErtsProcList *suspendees; + ErtsDistOutputBuf *obuf; + + erts_de_rwlock(dep); + erts_atomic_set_nob(&dep->input_handler, + (erts_aint_t) NIL); + cache = dep->cache; + dep->cache = NULL; + +#ifdef DEBUG + erts_de_links_lock(dep); + ASSERT(!dep->nlinks); + ASSERT(!dep->node_links); + ASSERT(!dep->monitors); + erts_de_links_unlock(dep); +#endif + + erts_mtx_lock(&dep->qlock); + + erts_atomic64_set_nob(&dep->in, 0); + erts_atomic64_set_nob(&dep->out, 0); + + obuf = clear_de_out_queues(dep); + dep->status = 0; + suspendees = get_suspended_on_de(dep, ERTS_DE_QFLGS_ALL); + + erts_mtx_unlock(&dep->qlock); + erts_atomic_set_nob(&dep->dist_cmd_scheduled, 0); + dep->send = NULL; + erts_de_rwunlock(dep); + + erts_resume_processes(suspendees); + + delete_cache(cache); + + free_de_out_queues(dep, obuf); + if (dep->transcode_ctx) + transcode_free_ctx(dep); +} + int erts_dsend_context_dtor(Binary* ctx_bin) { ErtsSendContext* ctx = ERTS_MAGIC_BIN_DATA(ctx_bin); @@ -729,8 +799,8 @@ int erts_dsend_context_dtor(Binary* ctx_bin) if (ctx->dss.phase >= ERTS_DSIG_SEND_PHASE_ALLOC && ctx->dss.obuf) { free_dist_obuf(ctx->dss.obuf); } - if (ctx->dep_to_deref) - erts_deref_dist_entry(ctx->dep_to_deref); + if (ctx->deref_dep) + erts_deref_dist_entry(ctx->dep); return 1; } @@ -815,9 +885,9 @@ erts_dsig_send_m_exit(ErtsDSigData *dsdp, Eterm watcher, Eterm watched, watched, watcher, ref, reason); #ifdef DEBUG - erts_smp_de_links_lock(dsdp->dep); + erts_de_links_lock(dsdp->dep); ASSERT(!erts_lookup_monitor(dsdp->dep->monitors, ref)); - erts_smp_de_links_unlock(dsdp->dep); + erts_de_links_unlock(dsdp->dep); #endif res = dsig_send_ctl(dsdp, ctl, 1); @@ -908,11 +978,30 @@ erts_dsig_send_msg(Eterm remote, Eterm message, ErtsSendContext* ctx) } #endif - if (token != NIL) - ctl = TUPLE4(&ctx->ctl_heap[0], - make_small(DOP_SEND_TT), am_Empty, remote, token); - else - ctl = TUPLE3(&ctx->ctl_heap[0], make_small(DOP_SEND), am_Empty, remote); + if (token != NIL) { + Eterm el1, el2; + if (ctx->dep->flags & DFLAG_SEND_SENDER) { + el1 = make_small(DOP_SEND_SENDER_TT); + el2 = sender->common.id; + } + else { + el1 = make_small(DOP_SEND_TT); + el2 = am_Empty; + } + ctl = TUPLE4(&ctx->ctl_heap[0], el1, el2, remote, token); + } + else { + Eterm el1, el2; + if (ctx->dep->flags & DFLAG_SEND_SENDER) { + el1 = make_small(DOP_SEND_SENDER); + el2 = sender->common.id; + } + else { + el1 = make_small(DOP_SEND); + el2 = am_Empty; + } + ctl = TUPLE3(&ctx->ctl_heap[0], el1, el2, remote); + } DTRACE6(message_send, sender_name, receiver_name, msize, tok_label, tok_lastcnt, tok_serial); DTRACE7(message_send_remote, sender_name, node_name, receiver_name, @@ -1149,22 +1238,25 @@ int erts_net_message(Port *prt, ErtsLink *lnk; Uint tuple_arity; int res; + Uint32 connection_id; #ifdef ERTS_DIST_MSG_DBG ErlDrvSizeT orig_len = len; #endif UseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); - ERTS_SMP_CHK_NO_PROC_LOCKS; + ERTS_CHK_NO_PROC_LOCKS; - ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt)); + ERTS_LC_ASSERT(!prt || erts_lc_is_port_locked(prt)); if (!erts_is_alive) { UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); return 0; } - if (hlen != 0) - goto data_error; + + + ASSERT(hlen == 0); + if (len == 0) { /* HANDLE TICK !!! */ UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); return 0; @@ -1183,30 +1275,31 @@ int erts_net_message(Port *prt, len--; } - if (len == 0) { - PURIFY_MSG("data error"); - goto data_error; - } - - res = erts_prepare_dist_ext(&ede, t, len, dep, dep->cache); + res = erts_prepare_dist_ext(&ede, t, len, dep, dep->cache, &connection_id); - if (res >= 0) - res = ctl_len = erts_decode_dist_ext_size(&ede); - else { + switch (res) { + case ERTS_PREP_DIST_EXT_CLOSED: + return 0; /* Connection not alive; ignore signal... */ + case ERTS_PREP_DIST_EXT_FAILED: #ifdef ERTS_DIST_MSG_DBG erts_fprintf(stderr, "DIST MSG DEBUG: erts_prepare_dist_ext() failed:\n"); bw(buf, orig_len); #endif - ctl_len = 0; - } - - if (res < 0) { + goto data_error; + case ERTS_PREP_DIST_EXT_SUCCESS: + ctl_len = erts_decode_dist_ext_size(&ede); + if (ctl_len < 0) { #ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, "DIST MSG DEBUG: erts_decode_dist_ext_size(CTL) failed:\n"); - bw(buf, orig_len); + erts_fprintf(stderr, "DIST MSG DEBUG: erts_decode_dist_ext_size(CTL) failed:\n"); + bw(buf, orig_len); #endif - PURIFY_MSG("data error"); - goto data_error; + PURIFY_MSG("data error"); + goto data_error; + } + break; + default: + ERTS_INTERNAL_ERROR("Unexpected result from erts_prepare_dist_ext()"); + break; } if (ctl_len > DIST_CTL_DEFAULT_SIZE) { @@ -1237,6 +1330,7 @@ int erts_net_message(Port *prt, } token_size = 0; + token = NIL; switch (type = unsigned_val(tuple[1])) { case DOP_LINK: @@ -1257,7 +1351,7 @@ int erts_net_message(Port *prt, /* This is tricky (we MUST force a distributed send) */ ErtsDSigData dsd; int code; - code = erts_dsig_prepare(&dsd, dep, NULL, ERTS_DSP_NO_LOCK, 0); + code = erts_dsig_prepare(&dsd, dep, NULL, 0, ERTS_DSP_NO_LOCK, 0, 0); if (code == ERTS_DSIG_PREP_CONNECTED) { code = erts_dsig_send_exit(&dsd, to, from, am_noproc); ASSERT(code == ERTS_DSIG_SEND_OK); @@ -1265,23 +1359,23 @@ int erts_net_message(Port *prt, break; } - erts_smp_de_links_lock(dep); + erts_de_links_lock(dep); res = erts_add_link(&ERTS_P_LINKS(rp), LINK_PID, from); if (res < 0) { /* It was already there! Lets skip the rest... */ - erts_smp_de_links_unlock(dep); - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK); + erts_de_links_unlock(dep); + erts_proc_unlock(rp, ERTS_PROC_LOCK_LINK); break; } lnk = erts_add_or_lookup_link(&(dep->nlinks), LINK_PID, rp->common.id); erts_add_link(&(ERTS_LINK_ROOT(lnk)), LINK_PID, from); - erts_smp_de_links_unlock(dep); + erts_de_links_unlock(dep); if (IS_TRACED_FL(rp, F_TRACE_PROCS)) trace_proc(NULL, 0, rp, am_getting_linked, from); - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK); + erts_proc_unlock(rp, ERTS_PROC_LOCK_LINK); break; case DOP_UNLINK: { @@ -1307,7 +1401,7 @@ int erts_net_message(Port *prt, trace_proc(NULL, 0, rp, am_getting_unlinked, from); } - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK); + erts_proc_unlock(rp, ERTS_PROC_LOCK_LINK); erts_remove_dist_link(&dld, to, from, dep); erts_destroy_dist_link(&dld); @@ -1349,7 +1443,7 @@ int erts_net_message(Port *prt, if (!rp) { ErtsDSigData dsd; int code; - code = erts_dsig_prepare(&dsd, dep, NULL, ERTS_DSP_NO_LOCK, 0); + code = erts_dsig_prepare(&dsd, dep, NULL, 0, ERTS_DSP_NO_LOCK, 0, 0); if (code == ERTS_DSIG_PREP_CONNECTED) { code = erts_dsig_send_m_exit(&dsd, watcher, watched, ref, am_noproc); @@ -1359,11 +1453,11 @@ int erts_net_message(Port *prt, else { if (is_atom(watched)) watched = rp->common.id; - erts_smp_de_links_lock(dep); + erts_de_links_lock(dep); erts_add_monitor(&(dep->monitors), MON_ORIGIN, ref, watched, name); erts_add_monitor(&ERTS_P_MONITORS(rp), MON_TARGET, ref, watcher, name); - erts_smp_de_links_unlock(dep); - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK); + erts_de_links_unlock(dep); + erts_proc_unlock(rp, ERTS_PROC_LOCK_LINK); } break; @@ -1385,9 +1479,9 @@ int erts_net_message(Port *prt, goto invalid_message; } - erts_smp_de_links_lock(dep); + erts_de_links_lock(dep); mon = erts_remove_monitor(&(dep->monitors),ref); - erts_smp_de_links_unlock(dep); + erts_de_links_unlock(dep); /* ASSERT(mon != NULL); can happen in case of broken dist message */ if (mon == NULL) { break; @@ -1401,7 +1495,7 @@ int erts_net_message(Port *prt, break; } mon = erts_remove_monitor(&ERTS_P_MONITORS(rp), ref); - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_LINK); + erts_proc_unlock(rp, ERTS_PROC_LOCK_LINK); ASSERT(mon != NULL); if (mon == NULL) { break; @@ -1462,42 +1556,56 @@ int erts_net_message(Port *prt, erts_queue_dist_message(rp, locks, ede_copy, token, from); if (locks) - erts_smp_proc_unlock(rp, locks); + erts_proc_unlock(rp, locks); } break; + case DOP_SEND_SENDER_TT: { + Uint xsize; case DOP_SEND_TT: + if (tuple_arity != 4) { goto invalid_message; } - - token_size = size_object(tuple[4]); - /* Fall through ... */ + + token = tuple[4]; + token_size = size_object(token); + xsize = ERTS_HEAP_FRAG_SIZE(token_size); + goto send_common; + + case DOP_SEND_SENDER: case DOP_SEND: + + token = NIL; + xsize = 0; + if (tuple_arity != 3) + goto invalid_message; + + send_common: + /* - * There is intentionally no testing of the cookie (it is always '') - * from R9B and onwards. + * If DOP_SEND_SENDER or DOP_SEND_SENDER_TT element 2 contains + * the sender pid (i.e. DFLAG_SEND_SENDER is set); otherwise, + * the atom '' (empty cookie). */ + ASSERT((type == DOP_SEND_SENDER || type == DOP_SEND_SENDER_TT) + ? (is_pid(tuple[2]) && (dep->flags & DFLAG_SEND_SENDER)) + : tuple[2] == am_Empty); + #ifdef ERTS_DIST_MSG_DBG dist_msg_dbg(&ede, "MSG", buf, orig_len); #endif - if (type != DOP_SEND_TT && tuple_arity != 3) { - goto invalid_message; - } to = tuple[3]; if (is_not_pid(to)) { goto invalid_message; } rp = erts_proc_lookup(to); if (rp) { - Uint xsize = type == DOP_SEND ? 0 : ERTS_HEAP_FRAG_SIZE(token_size); ErtsProcLocks locks = 0; ErtsDistExternal *ede_copy; ede_copy = erts_make_dist_ext_copy(&ede, xsize); - if (type == DOP_SEND) { - token = NIL; - } else { + if (is_not_nil(token)) { ErlHeapFragment *heap_frag; ErlOffHeap *ohp; ASSERT(xsize); @@ -1505,15 +1613,15 @@ int erts_net_message(Port *prt, ERTS_INIT_HEAP_FRAG(heap_frag, token_size, token_size); hp = heap_frag->mem; ohp = &heap_frag->off_heap; - token = tuple[4]; token = copy_struct(token, token_size, &hp, ohp); } - erts_queue_dist_message(rp, locks, ede_copy, token, tuple[2]); + erts_queue_dist_message(rp, locks, ede_copy, token, am_Empty); if (locks) - erts_smp_proc_unlock(rp, locks); + erts_proc_unlock(rp, locks); } break; + } case DOP_MONITOR_P_EXIT: { /* We are monitoring a process on the remote node which dies, we get @@ -1537,7 +1645,7 @@ int erts_net_message(Port *prt, goto invalid_message; } - erts_smp_de_links_lock(dep); + erts_de_links_lock(dep); sysname = dep->sysname; mon = erts_remove_monitor(&(dep->monitors), ref); /* @@ -1546,7 +1654,7 @@ int erts_net_message(Port *prt, * removed info about monitor. In this case, do nothing * and everything will be as it should. */ - erts_smp_de_links_unlock(dep); + erts_de_links_unlock(dep); if (mon == NULL) { break; } @@ -1560,7 +1668,7 @@ int erts_net_message(Port *prt, mon = erts_remove_monitor(&ERTS_P_MONITORS(rp), ref); if (mon == NULL) { - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); break; } UseTmpHeapNoproc(3); @@ -1571,7 +1679,7 @@ int erts_net_message(Port *prt, erts_queue_monitor_message(rp, &rp_locks, ref, am_process, watched, reason); - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); erts_destroy_monitor(mon); UnUseTmpHeapNoproc(3); break; @@ -1633,13 +1741,13 @@ int erts_net_message(Port *prt, if (xres >= 0 && IS_TRACED_FL(rp, F_TRACE_PROCS)) { /* We didn't exit the process and it is traced */ if (rp_locks & ERTS_PROC_LOCKS_XSIG_SEND) { - erts_smp_proc_unlock(rp, ERTS_PROC_LOCKS_XSIG_SEND); + erts_proc_unlock(rp, ERTS_PROC_LOCKS_XSIG_SEND); rp_locks &= ~ERTS_PROC_LOCKS_XSIG_SEND; } trace_proc(NULL, 0, rp, am_getting_unlinked, from); } } - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } erts_remove_dist_link(&dld, to, from, dep); if (lnk) @@ -1681,7 +1789,7 @@ int erts_net_message(Port *prt, token, NULL, 0); - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } break; } @@ -1699,7 +1807,7 @@ int erts_net_message(Port *prt, if (!rp) break; rp->group_leader = STORE_NC_IN_PROC(rp, from); - erts_smp_proc_unlock(rp, ERTS_PROC_LOCK_MAIN); + erts_proc_unlock(rp, ERTS_PROC_LOCK_MAIN); break; default: @@ -1711,7 +1819,7 @@ int erts_net_message(Port *prt, erts_free(ERTS_ALC_T_DCTRL_BUF, (void *) ctl); } UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); - ERTS_SMP_CHK_NO_PROC_LOCKS; + ERTS_CHK_NO_PROC_LOCKS; return 0; invalid_message: { @@ -1727,8 +1835,8 @@ decode_error: } data_error: UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); - erts_deliver_port_exit(prt, dep->cid, am_killed, 0, 1); - ERTS_SMP_CHK_NO_PROC_LOCKS; + erts_kill_dist_connection(dep, connection_id); + ERTS_CHK_NO_PROC_LOCKS; return -1; } @@ -1748,6 +1856,31 @@ static int dsig_send_ctl(ErtsDSigData* dsdp, Eterm ctl, int force_busy) return ret; } +static ERTS_INLINE void +notify_dist_data(Process *c_p, Eterm pid) +{ + Process *rp; + ErtsProcLocks rp_locks; + + ASSERT(erts_get_scheduler_data() + && !ERTS_SCHEDULER_IS_DIRTY(erts_get_scheduler_data())); + ASSERT(is_internal_pid(pid)); + + if (c_p && c_p->common.id == pid) { + rp = c_p; + rp_locks = ERTS_PROC_LOCK_MAIN; + } + else { + rp = erts_proc_lookup(pid); + rp_locks = 0; + } + + if (rp) { + ErtsMessage *mp = erts_alloc_message(0, NULL); + erts_queue_message(rp, rp_locks, mp, am_dist_data, am_system); + } +} + int erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) { @@ -1764,7 +1897,7 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) if (!ctx->c_p || dsdp->no_suspend) ctx->force_busy = 1; - ERTS_SMP_LC_ASSERT(!ctx->c_p + ERTS_LC_ASSERT(!ctx->c_p || (ERTS_PROC_LOCK_MAIN == erts_proc_lc_my_proc_locks(ctx->c_p))); @@ -1773,33 +1906,32 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) if (ctx->flags & DFLAG_DIST_HDR_ATOM_CACHE) { ctx->acmp = erts_get_atom_cache_map(ctx->c_p); - ctx->pass_through_size = 0; + ctx->max_finalize_prepend = 0; } else { ctx->acmp = NULL; - ctx->pass_through_size = 1; + ctx->max_finalize_prepend = 3; } #ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, ">>%s CTL: %T\n", ctx->pass_through_size ? "P" : " ", ctx->ctl); - if (is_value(ctx->msg)) - erts_fprintf(stderr, " MSG: %T\n", ctx->msg); + erts_fprintf(stderr, ">> CTL: %T\n", ctx->ctl); + if (is_value(ctx->msg)) + erts_fprintf(stderr, " MSG: %T\n", ctx->msg); #endif - ctx->data_size = ctx->pass_through_size; + ctx->data_size = ctx->max_finalize_prepend; erts_reset_atom_cache_map(ctx->acmp); erts_encode_dist_ext_size(ctx->ctl, ctx->flags, ctx->acmp, &ctx->data_size); - if (is_value(ctx->msg)) { - ctx->u.sc.wstack.wstart = NULL; - ctx->u.sc.flags = ctx->flags; - ctx->u.sc.level = 0; - ctx->phase = ERTS_DSIG_SEND_PHASE_MSG_SIZE; - } else { - ctx->phase = ERTS_DSIG_SEND_PHASE_ALLOC; - } - break; + if (is_non_value(ctx->msg)) { + ctx->phase = ERTS_DSIG_SEND_PHASE_ALLOC; + break; + } + ctx->u.sc.wstack.wstart = NULL; + ctx->u.sc.flags = ctx->flags; + ctx->u.sc.level = 0; + ctx->phase = ERTS_DSIG_SEND_PHASE_MSG_SIZE; case ERTS_DSIG_SEND_PHASE_MSG_SIZE: if (erts_encode_dist_ext_size_int(ctx->msg, ctx, &ctx->data_size)) { retval = ERTS_DSIG_SEND_CONTINUE; @@ -1814,23 +1946,24 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) ctx->data_size += ctx->dhdr_ext_size; ctx->obuf = alloc_dist_obuf(ctx->data_size); - ctx->obuf->ext_endp = &ctx->obuf->data[0] + ctx->pass_through_size + ctx->dhdr_ext_size; + ctx->obuf->ext_endp = &ctx->obuf->data[0] + ctx->max_finalize_prepend + ctx->dhdr_ext_size; /* Encode internal version of dist header */ ctx->obuf->extp = erts_encode_ext_dist_header_setup(ctx->obuf->ext_endp, ctx->acmp); /* Encode control message */ erts_encode_dist_ext(ctx->ctl, &ctx->obuf->ext_endp, ctx->flags, ctx->acmp, NULL, NULL); - if (is_value(ctx->msg)) { - ctx->u.ec.flags = ctx->flags; - ctx->u.ec.level = 0; - ctx->u.ec.wstack.wstart = NULL; - ctx->phase = ERTS_DSIG_SEND_PHASE_MSG_ENCODE; - } else { - ctx->phase = ERTS_DSIG_SEND_PHASE_FIN; - } - break; + if (is_non_value(ctx->msg)) { + ctx->obuf->msg_start = NULL; + ctx->phase = ERTS_DSIG_SEND_PHASE_FIN; + break; + } + ctx->u.ec.flags = ctx->flags; + ctx->u.ec.level = 0; + ctx->u.ec.wstack.wstart = NULL; + ctx->obuf->msg_start = ctx->obuf->ext_endp; - case ERTS_DSIG_SEND_PHASE_MSG_ENCODE: + ctx->phase = ERTS_DSIG_SEND_PHASE_MSG_ENCODE; + case ERTS_DSIG_SEND_PHASE_MSG_ENCODE: if (erts_encode_dist_ext(ctx->msg, &ctx->obuf->ext_endp, ctx->flags, ctx->acmp, &ctx->u.ec, &ctx->reds)) { retval = ERTS_DSIG_SEND_CONTINUE; goto done; @@ -1843,7 +1976,7 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) int resume = 0; ASSERT(ctx->obuf->extp < ctx->obuf->ext_endp); - ASSERT(&ctx->obuf->data[0] <= ctx->obuf->extp - ctx->pass_through_size); + ASSERT(&ctx->obuf->data[0] <= ctx->obuf->extp - ctx->max_finalize_prepend); ASSERT(ctx->obuf->ext_endp <= &ctx->obuf->data[0] + ctx->data_size); ctx->data_size = ctx->obuf->ext_endp - ctx->obuf->extp; @@ -1853,28 +1986,48 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) * and if so enqueue the signal and schedule it for send. */ ctx->obuf->next = NULL; - erts_smp_de_rlock(dep); + erts_de_rlock(dep); cid = dep->cid; - if (cid != dsdp->cid - || dep->connection_id != dsdp->connection_id - || dep->status & ERTS_DE_SFLG_EXITING) { + if (!(dep->status & (ERTS_DE_SFLG_PENDING | ERTS_DE_SFLG_CONNECTED)) + || dep->status & ERTS_DE_SFLG_EXITING + || dep->connection_id != dsdp->connection_id) { /* Not the same connection as when we started; drop message... */ - erts_smp_de_runlock(dep); + erts_de_runlock(dep); free_dist_obuf(ctx->obuf); } else { + Sint qsize; + erts_aint32_t qflgs; ErtsProcList *plp = NULL; - erts_smp_mtx_lock(&dep->qlock); - dep->qsize += size_obuf(ctx->obuf); - if (dep->qsize >= erts_dist_buf_busy_limit) - dep->qflgs |= ERTS_DE_QFLG_BUSY; - if (!ctx->force_busy && (dep->qflgs & ERTS_DE_QFLG_BUSY)) { - erts_smp_mtx_unlock(&dep->qlock); + Eterm notify_proc = NIL; + Sint obsz = size_obuf(ctx->obuf); + + erts_mtx_lock(&dep->qlock); + qsize = erts_atomic_add_read_nob(&dep->qsize, (erts_aint_t) obsz); + ASSERT(qsize >= obsz); + qflgs = erts_atomic32_read_nob(&dep->qflgs); + if (!(qflgs & ERTS_DE_QFLG_BUSY) && qsize >= erts_dist_buf_busy_limit) { + erts_atomic32_read_bor_relb(&dep->qflgs, ERTS_DE_QFLG_BUSY); + qflgs |= ERTS_DE_QFLG_BUSY; + } + if (qsize == obsz && (qflgs & ERTS_DE_QFLG_REQ_INFO)) { + /* Previously empty queue and info requested... */ + qflgs = erts_atomic32_read_band_mb(&dep->qflgs, + ~ERTS_DE_QFLG_REQ_INFO); + if (qflgs & ERTS_DE_QFLG_REQ_INFO) { + notify_proc = dep->cid; + ASSERT(is_internal_pid(notify_proc)); + } + /* else: requester will send itself the message... */ + qflgs &= ~ERTS_DE_QFLG_REQ_INFO; + } + if (!ctx->force_busy && (qflgs & ERTS_DE_QFLG_BUSY)) { + erts_mtx_unlock(&dep->qlock); plp = erts_proclist_create(ctx->c_p); erts_suspend(ctx->c_p, ERTS_PROC_LOCK_MAIN, NULL); suspended = 1; - erts_smp_mtx_lock(&dep->qlock); + erts_mtx_lock(&dep->qlock); } /* Enqueue obuf on dist entry */ @@ -1885,7 +2038,8 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) dep->out_queue.last = ctx->obuf; if (!ctx->force_busy) { - if (!(dep->qflgs & ERTS_DE_QFLG_BUSY)) { + qflgs = erts_atomic32_read_nob(&dep->qflgs); + if (!(qflgs & ERTS_DE_QFLG_BUSY)) { if (suspended) resume = 1; /* was busy when we started, but isn't now */ #ifdef USE_VM_PROBES @@ -1909,9 +2063,17 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) } } - erts_smp_mtx_unlock(&dep->qlock); - erts_schedule_dist_command(NULL, dep); - erts_smp_de_runlock(dep); + erts_mtx_unlock(&dep->qlock); + if (!(dep->status & ERTS_DE_SFLG_PENDING)) { + if (is_internal_port(dep->cid)) + erts_schedule_dist_command(NULL, dep); + } + else { + notify_proc = NIL; + } + erts_de_runlock(dep); + if (is_internal_pid(notify_proc)) + notify_dist_data(ctx->c_p, notify_proc); if (resume) { erts_resume(ctx->c_p, ERTS_PROC_LOCK_MAIN); @@ -1965,16 +2127,20 @@ static Uint dist_port_command(Port *prt, ErtsDistOutputBuf *obuf) { int fpe_was_unmasked; - Uint size = obuf->ext_endp - obuf->extp; + ErlDrvSizeT size; + char *bufp; - ERTS_SMP_CHK_NO_PROC_LOCKS; - ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt)); + ERTS_CHK_NO_PROC_LOCKS; + ERTS_LC_ASSERT(erts_lc_is_port_locked(prt)); - if (size > (Uint) INT_MAX) - erts_exit(ERTS_DUMP_EXIT, - "Absurdly large distribution output data buffer " - "(%beu bytes) passed.\n", - size); + if (!obuf) { + size = 0; + bufp = NULL; + } + else { + size = obuf->ext_endp - obuf->extp; + bufp = (char*) obuf->extp; + } #ifdef USE_VM_PROBES if (DTRACE_ENABLED(dist_output)) { @@ -1989,11 +2155,10 @@ dist_port_command(Port *prt, ErtsDistOutputBuf *obuf) remote_str, size); } #endif + prt->caller = NIL; fpe_was_unmasked = erts_block_fpe(); - (*prt->drv_ptr->output)((ErlDrvData) prt->drv_data, - (char*) obuf->extp, - (int) size); + (*prt->drv_ptr->output)((ErlDrvData) prt->drv_data, bufp, size); erts_unblock_fpe(fpe_was_unmasked); return size; } @@ -2002,33 +2167,41 @@ static Uint dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf) { int fpe_was_unmasked; - Uint size = obuf->ext_endp - obuf->extp; + ErlDrvSizeT size; SysIOVec iov[2]; ErlDrvBinary* bv[2]; ErlIOVec eiov; - ERTS_SMP_CHK_NO_PROC_LOCKS; - ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt)); - - if (size > (Uint) INT_MAX) - erts_exit(ERTS_DUMP_EXIT, - "Absurdly large distribution output data buffer " - "(%beu bytes) passed.\n", - size); + ERTS_CHK_NO_PROC_LOCKS; + ERTS_LC_ASSERT(erts_lc_is_port_locked(prt)); iov[0].iov_base = NULL; iov[0].iov_len = 0; bv[0] = NULL; - iov[1].iov_base = obuf->extp; - iov[1].iov_len = size; - bv[1] = Binary2ErlDrvBinary(ErtsDistOutputBuf2Binary(obuf)); + if (!obuf) { + size = 0; + eiov.vsize = 1; + } + else { + size = obuf->ext_endp - obuf->extp; + eiov.vsize = 2; + + iov[1].iov_base = obuf->extp; + iov[1].iov_len = size; + bv[1] = Binary2ErlDrvBinary(ErtsDistOutputBuf2Binary(obuf)); + } - eiov.vsize = 2; eiov.size = size; eiov.iov = iov; eiov.binv = bv; + if (size > (Uint) INT_MAX) + erts_exit(ERTS_DUMP_EXIT, + "Absurdly large distribution output data buffer " + "(%beu bytes) passed.\n", + size); + ASSERT(prt->drv_ptr->outputv); #ifdef USE_VM_PROBES @@ -2062,7 +2235,6 @@ dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf) #endif #define ERTS_PORT_REDS_DIST_CMD_START 5 -#define ERTS_PORT_REDS_DIST_CMD_FINALIZE 3 #define ERTS_PORT_REDS_DIST_CMD_EXIT 200 #define ERTS_PORT_REDS_DIST_CMD_RESUMED 5 #define ERTS_PORT_REDS_DIST_CMD_DATA(SZ) \ @@ -2071,37 +2243,36 @@ dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf) : ((((Sint) (SZ)) >> 10) & ((Sint) ERTS_PORT_REDS_MASK__))) int -erts_dist_command(Port *prt, int reds_limit) +erts_dist_command(Port *prt, int initial_reds) { - Sint reds = ERTS_PORT_REDS_DIST_CMD_START; + Sint reds = initial_reds - ERTS_PORT_REDS_DIST_CMD_START; Uint32 status; Uint32 flags; - Sint obufsize = 0; + Sint qsize, obufsize = 0; ErtsDistOutputQueue oq, foq; DistEntry *dep = prt->dist_entry; Uint (*send)(Port *prt, ErtsDistOutputBuf *obuf); erts_aint32_t sched_flags; ErtsSchedulerData *esdp = erts_get_scheduler_data(); - ERTS_SMP_LC_ASSERT(erts_lc_is_port_locked(prt)); + ERTS_LC_ASSERT(erts_lc_is_port_locked(prt)); - erts_smp_refc_inc(&dep->refc, 1); /* Otherwise dist_entry might be - removed if port command fails */ + erts_atomic_set_mb(&dep->dist_cmd_scheduled, 0); - erts_smp_atomic_set_mb(&dep->dist_cmd_scheduled, 0); - - erts_smp_de_rlock(dep); + erts_de_rlock(dep); flags = dep->flags; status = dep->status; send = dep->send; - erts_smp_de_runlock(dep); + erts_de_runlock(dep); if (status & ERTS_DE_SFLG_EXITING) { erts_deliver_port_exit(prt, prt->common.id, am_killed, 0, 1); - erts_deref_dist_entry(dep); - return reds + ERTS_PORT_REDS_DIST_CMD_EXIT; + reds -= ERTS_PORT_REDS_DIST_CMD_EXIT; + return initial_reds - reds; } + ASSERT(!(status & ERTS_DE_SFLG_PENDING)); + ASSERT(send); /* @@ -2112,42 +2283,42 @@ erts_dist_command(Port *prt, int reds_limit) * a mess. */ - erts_smp_mtx_lock(&dep->qlock); + erts_mtx_lock(&dep->qlock); oq.first = dep->out_queue.first; oq.last = dep->out_queue.last; dep->out_queue.first = NULL; dep->out_queue.last = NULL; - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_unlock(&dep->qlock); foq.first = dep->finalized_out_queue.first; foq.last = dep->finalized_out_queue.last; dep->finalized_out_queue.first = NULL; dep->finalized_out_queue.last = NULL; - sched_flags = erts_smp_atomic32_read_nob(&prt->sched.flags); + sched_flags = erts_atomic32_read_nob(&prt->sched.flags); - if (reds > reds_limit) + if (reds < 0) goto preempted; if (!(sched_flags & ERTS_PTS_FLG_BUSY_PORT) && foq.first) { int preempt = 0; do { - Uint size; - ErtsDistOutputBuf *fob; - - size = (*send)(prt, foq.first); - esdp->io.out += (Uint64) size; + Uint size; + ErtsDistOutputBuf *fob; + size = (*send)(prt, foq.first); + erts_atomic64_inc_nob(&dep->out); + esdp->io.out += (Uint64) size; #ifdef ERTS_RAW_DIST_MSG_DBG - erts_fprintf(stderr, ">> "); - bw(foq.first->extp, size); + erts_fprintf(stderr, ">> "); + bw(foq.first->extp, size); #endif - reds += ERTS_PORT_REDS_DIST_CMD_DATA(size); + reds -= ERTS_PORT_REDS_DIST_CMD_DATA(size); fob = foq.first; obufsize += size_obuf(fob); foq.first = foq.first->next; free_dist_obuf(fob); - sched_flags = erts_smp_atomic32_read_nob(&prt->sched.flags); - preempt = reds > reds_limit || (sched_flags & ERTS_PTS_FLG_EXIT); + sched_flags = erts_atomic32_read_nob(&prt->sched.flags); + preempt = reds < 0 || (sched_flags & ERTS_PTS_FLG_EXIT); if (sched_flags & ERTS_PTS_FLG_BUSY_PORT) break; } while (foq.first && !preempt); @@ -2160,79 +2331,71 @@ erts_dist_command(Port *prt, int reds_limit) if (sched_flags & ERTS_PTS_FLG_BUSY_PORT) { if (oq.first) { ErtsDistOutputBuf *ob; - int preempt; + ErtsDistOutputBuf *last_finalized = NULL; finalize_only: - preempt = 0; ob = oq.first; ASSERT(ob); do { - ob->extp = erts_encode_ext_dist_header_finalize(ob->extp, - dep->cache, - flags); - if (!(flags & DFLAG_DIST_HDR_ATOM_CACHE)) - *--ob->extp = PASS_THROUGH; /* Old node; 'pass through' - needed */ - ASSERT(&ob->data[0] <= ob->extp && ob->extp < ob->ext_endp); - reds += ERTS_PORT_REDS_DIST_CMD_FINALIZE; - preempt = reds > reds_limit; - if (preempt) - break; - ob = ob->next; + reds = erts_encode_ext_dist_header_finalize(ob, dep, flags, reds); + if (reds >= 0) { + last_finalized = ob; + ob = ob->next; + } } while (ob); - /* - * At least one buffer was finalized; if we got preempted, - * ob points to the last buffer that we finalized. - */ - if (foq.last) - foq.last->next = oq.first; - else - foq.first = oq.first; - if (!preempt) { - /* All buffers finalized */ - foq.last = oq.last; - oq.first = oq.last = NULL; - } - else { - /* Not all buffers finalized; split oq. */ - foq.last = ob; - oq.first = ob->next; - if (oq.first) - ob->next = NULL; - else - oq.last = NULL; - } - if (preempt) - goto preempted; + if (last_finalized) { + /* + * At least one buffer was finalized; if we got preempted, + * ob points to the next buffer to continue finalize. + */ + if (foq.last) + foq.last->next = oq.first; + else + foq.first = oq.first; + foq.last = last_finalized; + if (!ob) { + /* All buffers finalized */ + ASSERT(foq.last == oq.last); + ASSERT(foq.last->next == NULL); + oq.first = oq.last = NULL; + } + else { + /* Not all buffers finalized; split oq. */ + ASSERT(foq.last->next == ob); + foq.last->next = NULL; + oq.first = ob; + } + } + if (reds <= 0) + goto preempted; } } else { + int de_busy; int preempt = 0; while (oq.first && !preempt) { ErtsDistOutputBuf *fob; Uint size; - oq.first->extp - = erts_encode_ext_dist_header_finalize(oq.first->extp, - dep->cache, - flags); - reds += ERTS_PORT_REDS_DIST_CMD_FINALIZE; - if (!(flags & DFLAG_DIST_HDR_ATOM_CACHE)) - *--oq.first->extp = PASS_THROUGH; /* Old node; 'pass through' - needed */ + reds = erts_encode_ext_dist_header_finalize(oq.first, dep, flags, reds); + if (reds < 0) { + preempt = 1; + break; + } ASSERT(&oq.first->data[0] <= oq.first->extp && oq.first->extp < oq.first->ext_endp); size = (*send)(prt, oq.first); + erts_atomic64_inc_nob(&dep->out); esdp->io.out += (Uint64) size; #ifdef ERTS_RAW_DIST_MSG_DBG - erts_fprintf(stderr, ">> "); - bw(oq.first->extp, size); + erts_fprintf(stderr, ">> "); + bw(oq.first->extp, size); #endif - reds += ERTS_PORT_REDS_DIST_CMD_DATA(size); + reds -= ERTS_PORT_REDS_DIST_CMD_DATA(size); fob = oq.first; obufsize += size_obuf(fob); oq.first = oq.first->next; free_dist_obuf(fob); - sched_flags = erts_smp_atomic32_read_nob(&prt->sched.flags); - preempt = reds > reds_limit || (sched_flags & ERTS_PTS_FLG_EXIT); + sched_flags = erts_atomic32_read_nob(&prt->sched.flags); + preempt = reds <= 0 || (sched_flags & ERTS_PTS_FLG_EXIT); if ((sched_flags & ERTS_PTS_FLG_BUSY_PORT) && oq.first && !preempt) goto finalize_only; } @@ -2258,23 +2421,24 @@ erts_dist_command(Port *prt, int reds_limit) * dist entry in a non-busy state and resume suspended * processes. */ - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qsize >= obufsize); - dep->qsize -= obufsize; + erts_mtx_lock(&dep->qlock); + de_busy = !!(erts_atomic32_read_nob(&dep->qflgs) & ERTS_DE_QFLG_BUSY); + qsize = (Sint) erts_atomic_add_read_nob(&dep->qsize, + (erts_aint_t) -obufsize); + ASSERT(qsize >= 0); obufsize = 0; if (!(sched_flags & ERTS_PTS_FLG_BUSY_PORT) - && (dep->qflgs & ERTS_DE_QFLG_BUSY) - && dep->qsize < erts_dist_buf_busy_limit) { + && de_busy && qsize < erts_dist_buf_busy_limit) { ErtsProcList *suspendees; int resumed; suspendees = get_suspended_on_de(dep, ERTS_DE_QFLG_BUSY); - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_unlock(&dep->qlock); resumed = erts_resume_processes(suspendees); - reds += resumed*ERTS_PORT_REDS_DIST_CMD_RESUMED; + reds -= resumed*ERTS_PORT_REDS_DIST_CMD_RESUMED; } else - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_unlock(&dep->qlock); } ASSERT(!oq.first && !oq.last); @@ -2283,14 +2447,18 @@ erts_dist_command(Port *prt, int reds_limit) if (obufsize != 0) { ASSERT(obufsize > 0); - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qsize >= obufsize); - dep->qsize -= obufsize; - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_lock(&dep->qlock); +#ifdef DEBUG + qsize = (Sint) erts_atomic_add_read_nob(&dep->qsize, + (erts_aint_t) -obufsize); + ASSERT(qsize >= 0); +#else + erts_atomic_add_nob(&dep->qsize, (erts_aint_t) -obufsize); +#endif + erts_mtx_unlock(&dep->qlock); } - ASSERT(foq.first || !foq.last); - ASSERT(!foq.first || foq.last); + ASSERT(!!foq.first == !!foq.last); ASSERT(!dep->finalized_out_queue.first); ASSERT(!dep->finalized_out_queue.last); @@ -2300,12 +2468,10 @@ erts_dist_command(Port *prt, int reds_limit) } /* Avoid wrapping reduction counter... */ - if (reds > INT_MAX/2) - reds = INT_MAX/2; - - erts_deref_dist_entry(dep); + if (reds < INT_MIN/2) + reds = INT_MIN/2; - return reds; + return initial_reds - reds; preempted: /* @@ -2313,8 +2479,7 @@ erts_dist_command(Port *prt, int reds_limit) * since last call to driver. */ - ASSERT(oq.first || !oq.last); - ASSERT(!oq.first || oq.last); + ASSERT(!!oq.first == !!oq.last); if (sched_flags & ERTS_PTS_FLG_EXIT) { /* @@ -2338,12 +2503,6 @@ erts_dist_command(Port *prt, int reds_limit) foq.first = NULL; foq.last = NULL; - -#ifdef DEBUG - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qsize == obufsize); - erts_smp_mtx_unlock(&dep->qlock); -#endif } else { if (oq.first) { @@ -2351,14 +2510,14 @@ erts_dist_command(Port *prt, int reds_limit) * Unhandle buffers need to be put back first * in out_queue. */ - erts_smp_mtx_lock(&dep->qlock); - dep->qsize -= obufsize; + erts_mtx_lock(&dep->qlock); + erts_atomic_add_nob(&dep->qsize, -obufsize); obufsize = 0; oq.last->next = dep->out_queue.first; dep->out_queue.first = oq.first; if (!dep->out_queue.last) dep->out_queue.last = oq.last; - erts_smp_mtx_unlock(&dep->qlock); + erts_mtx_unlock(&dep->qlock); } erts_schedule_dist_command(prt, NULL); @@ -2366,6 +2525,369 @@ erts_dist_command(Port *prt, int reds_limit) goto done; } +#if 0 + +int +dist_data_finalize(Process *c_p, int reds_limit) +{ + int reds = 5; + DistEntry *dep = ; + ErtsDistOutputQueue oq, foq; + ErtsDistOutputBuf *ob; + int preempt; + + + erts_mtx_lock(&dep->qlock); + flags = dep->flags; + oq.first = dep->out_queue.first; + oq.last = dep->out_queue.last; + dep->out_queue.first = NULL; + dep->out_queue.last = NULL; + erts_mtx_unlock(&dep->qlock); + + if (!oq.first) { + ASSERT(!oq.last); + oq.first = dep->tmp_out_queue.first; + oq.last = dep->tmp_out_queue.last; + } + else { + ErtsDistOutputBuf *f, *l; + ASSERT(oq.last); + if (dep->tmp_out_queue.last) { + dep->tmp_out_queue.last->next = oq.first; + oq.first = dep->tmp_out_queue.first; + } + } + + if (!oq.first) { + /* Nothing to do... */ + ASSERT(!oq.last); + return reds; + } + + foq.first = dep->finalized_out_queue.first; + foq.last = dep->finalized_out_queue.last; + + preempt = 0; + ob = oq.first; + ASSERT(ob); + + do { + ob->extp = erts_encode_ext_dist_header_finalize(ob->extp, + dep->cache, + flags); + if (!(flags & DFLAG_DIST_HDR_ATOM_CACHE)) + *--ob->extp = PASS_THROUGH; /* Old node; 'pass through' + needed */ + ASSERT(&ob->data[0] <= ob->extp && ob->extp < ob->ext_endp); + reds += ERTS_PORT_REDS_DIST_CMD_FINALIZE; + preempt = reds > reds_limit; + if (preempt) + break; + ob = ob->next; + } while (ob); + /* + * At least one buffer was finalized; if we got preempted, + * ob points to the last buffer that we finalized. + */ + if (foq.last) + foq.last->next = oq.first; + else + foq.first = oq.first; + if (!preempt) { + /* All buffers finalized */ + foq.last = oq.last; + oq.first = oq.last = NULL; + } + else { + /* Not all buffers finalized; split oq. */ + foq.last = ob; + oq.first = ob->next; + if (oq.first) + ob->next = NULL; + else + oq.last = NULL; + } + + dep->finalized_out_queue.first = foq.first; + dep->finalized_out_queue.last = foq.last; + dep->tmp_out_queue.first = oq.first; + dep->tmp_out_queue.last = oq.last; + + return reds; +} + +#endif + +BIF_RETTYPE +dist_ctrl_get_data_notification_1(BIF_ALIST_1) +{ + DistEntry *dep = ERTS_PROC_GET_DIST_ENTRY(BIF_P); + erts_aint32_t qflgs; + erts_aint_t qsize; + Eterm receiver = NIL; + + if (!dep) + BIF_ERROR(BIF_P, EXC_NOTSUP); + + if (erts_dhandle_to_dist_entry(BIF_ARG_1) != dep) + BIF_ERROR(BIF_P, BADARG); + + /* + * Caller is the only one that can consume from this queue + * and the only one that can set the req-info flag... + */ + + erts_de_rlock(dep); + + ASSERT(dep->cid == BIF_P->common.id); + + qflgs = erts_atomic32_read_acqb(&dep->qflgs); + + if (!(qflgs & ERTS_DE_QFLG_REQ_INFO)) { + qsize = erts_atomic_read_acqb(&dep->qsize); + ASSERT(qsize >= 0); + if (qsize > 0) + receiver = BIF_P->common.id; /* Notify ourselves... */ + else { /* Empty queue; set req-info flag... */ + qflgs = erts_atomic32_read_bor_mb(&dep->qflgs, + ERTS_DE_QFLG_REQ_INFO); + qsize = erts_atomic_read_acqb(&dep->qsize); + ASSERT(qsize >= 0); + if (qsize > 0) { + qflgs = erts_atomic32_read_band_mb(&dep->qflgs, + ~ERTS_DE_QFLG_REQ_INFO); + if (qflgs & ERTS_DE_QFLG_REQ_INFO) + receiver = BIF_P->common.id; /* Notify ourselves... */ + /* else: someone else will notify us... */ + } + /* else: still empty queue... */ + } + } + /* else: Already requested... */ + + erts_de_runlock(dep); + + if (is_internal_pid(receiver)) + notify_dist_data(BIF_P, receiver); + + BIF_RET(am_ok); +} + +BIF_RETTYPE +dist_ctrl_put_data_2(BIF_ALIST_2) +{ + DistEntry *dep; + ErlDrvSizeT size; + Eterm input_handler; + + if (is_binary(BIF_ARG_2)) + size = binary_size(BIF_ARG_2); + else if (is_nil(BIF_ARG_2)) + size = 0; + else if (is_list(BIF_ARG_2)) + BIF_TRAP2(dist_ctrl_put_data_trap, + BIF_P, BIF_ARG_1, BIF_ARG_2); + else + BIF_ERROR(BIF_P, BADARG); + + dep = erts_dhandle_to_dist_entry(BIF_ARG_1); + if (!dep) + BIF_ERROR(BIF_P, BADARG); + + input_handler = (Eterm) erts_atomic_read_nob(&dep->input_handler); + + if (input_handler != BIF_P->common.id) + BIF_ERROR(BIF_P, EXC_NOTSUP); + + erts_atomic64_inc_nob(&dep->in); + + if (size != 0) { + byte *data, *temp_alloc = NULL; + + data = (byte *) erts_get_aligned_binary_bytes(BIF_ARG_2, &temp_alloc); + if (!data) + BIF_ERROR(BIF_P, BADARG); + + erts_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN); + + (void) erts_net_message(NULL, dep, NULL, 0, data, size); + /* + * We ignore any decode failures. On fatal failures the + * connection will be taken down by killing the + * distribution channel controller... + */ + + erts_proc_lock(BIF_P, ERTS_PROC_LOCK_MAIN); + + BUMP_REDS(BIF_P, 5); + + erts_free_aligned_binary_bytes(temp_alloc); + + } + + BIF_RET(am_ok); +} + +BIF_RETTYPE +dist_get_stat_1(BIF_ALIST_1) +{ + Sint64 read, write, pend; + Eterm res, *hp, **hpp; + Uint sz, *szp; + DistEntry *dep = erts_dhandle_to_dist_entry(BIF_ARG_1); + + if (!dep) + BIF_ERROR(BIF_P, BADARG); + + erts_de_rlock(dep); + + read = (Sint64) erts_atomic64_read_nob(&dep->in); + write = (Sint64) erts_atomic64_read_nob(&dep->out); + pend = (Sint64) erts_atomic_read_nob(&dep->qsize); + + erts_de_runlock(dep); + + sz = 0; + szp = &sz; + hpp = NULL; + + while (1) { + res = erts_bld_tuple(hpp, szp, 4, + am_ok, + erts_bld_sint64(hpp, szp, read), + erts_bld_sint64(hpp, szp, write), + pend ? am_true : am_false); + if (hpp) + break; + hp = HAlloc(BIF_P, sz); + hpp = &hp; + szp = NULL; + } + + BIF_RET(res); +} + +BIF_RETTYPE +dist_ctrl_input_handler_2(BIF_ALIST_2) +{ + DistEntry *dep = ERTS_PROC_GET_DIST_ENTRY(BIF_P); + + if (!dep) + BIF_ERROR(BIF_P, EXC_NOTSUP); + + if (erts_dhandle_to_dist_entry(BIF_ARG_1) != dep) + BIF_ERROR(BIF_P, BADARG); + + if (is_not_internal_pid(BIF_ARG_2)) + BIF_ERROR(BIF_P, BADARG); + + erts_atomic_set_nob(&dep->input_handler, + (erts_aint_t) BIF_ARG_2); + + BIF_RET(am_ok); +} + +BIF_RETTYPE +dist_ctrl_get_data_1(BIF_ALIST_1) +{ + DistEntry *dep = ERTS_PROC_GET_DIST_ENTRY(BIF_P); + const Sint initial_reds = ERTS_BIF_REDS_LEFT(BIF_P); + Sint reds = initial_reds; + ErtsDistOutputBuf *obuf; + Eterm *hp; + ProcBin *pb; + erts_aint_t qsize; + + if (!dep) + BIF_ERROR(BIF_P, EXC_NOTSUP); + + if (erts_dhandle_to_dist_entry(BIF_ARG_1) != dep) + BIF_ERROR(BIF_P, BADARG); + + erts_de_rlock(dep); + + if (dep->status & ERTS_DE_SFLG_EXITING) + goto return_none; + + ASSERT(dep->cid == BIF_P->common.id); + +#if 0 + if (dep->finalized_out_queue.first) { + obuf = dep->finalized_out_queue.first; + dep->finalized_out_queue.first = obuf->next; + if (!obuf->next) + dep->finalized_out_queue.last = NULL; + } + else +#endif + { + if (!dep->tmp_out_queue.first) { + ASSERT(!dep->tmp_out_queue.last); + ASSERT(!dep->transcode_ctx); + qsize = erts_atomic_read_acqb(&dep->qsize); + if (qsize > 0) { + erts_mtx_lock(&dep->qlock); + dep->tmp_out_queue.first = dep->out_queue.first; + dep->tmp_out_queue.last = dep->out_queue.last; + dep->out_queue.first = NULL; + dep->out_queue.last = NULL; + erts_mtx_unlock(&dep->qlock); + } + } + + if (!dep->tmp_out_queue.first) { + ASSERT(!dep->tmp_out_queue.last); + return_none: + erts_de_runlock(dep); + BIF_RET(am_none); + } + + obuf = dep->tmp_out_queue.first; + reds = erts_encode_ext_dist_header_finalize(obuf, dep, dep->flags, reds); + if (reds < 0) { + erts_de_runlock(dep); + ERTS_BIF_YIELD1(bif_export[BIF_dist_ctrl_get_data_1], + BIF_P, BIF_ARG_1); + } + + dep->tmp_out_queue.first = obuf->next; + if (!obuf->next) + dep->tmp_out_queue.last = NULL; + } + + erts_atomic64_inc_nob(&dep->out); + + erts_de_runlock(dep); + + hp = HAlloc(BIF_P, PROC_BIN_SIZE); + pb = (ProcBin *) (char *) hp; + pb->thing_word = HEADER_PROC_BIN; + pb->size = obuf->ext_endp - obuf->extp; + pb->next = MSO(BIF_P).first; + MSO(BIF_P).first = (struct erl_off_heap_header*) pb; + pb->val = ErtsDistOutputBuf2Binary(obuf); + pb->bytes = (byte*) obuf->extp; + pb->flags = 0; + + qsize = erts_atomic_add_read_nob(&dep->qsize, -size_obuf(obuf)); + ASSERT(qsize >= 0); + + if (qsize < erts_dist_buf_busy_limit/2 + && (erts_atomic32_read_acqb(&dep->qflgs) & ERTS_DE_QFLG_BUSY)) { + ErtsProcList *resume_procs = NULL; + erts_mtx_lock(&dep->qlock); + resume_procs = get_suspended_on_de(dep, ERTS_DE_QFLG_BUSY); + erts_mtx_unlock(&dep->qlock); + if (resume_procs) { + int resumed = erts_resume_processes(resume_procs); + reds -= resumed*ERTS_PORT_REDS_DIST_CMD_RESUMED; + } + } + + BIF_RET2(make_binary(pb), (initial_reds - reds)); +} + void erts_dist_port_not_busy(Port *prt) { @@ -2385,24 +2907,33 @@ erts_dist_port_not_busy(Port *prt) erts_schedule_dist_command(prt, NULL); } +static void kill_connection(DistEntry *dep) +{ + ERTS_LC_ASSERT(erts_lc_is_de_rwlocked(dep)); + ASSERT(dep->status == ERTS_DE_SFLG_CONNECTED); + + dep->status |= ERTS_DE_SFLG_EXITING; + erts_mtx_lock(&dep->qlock); + ASSERT(!(erts_atomic32_read_nob(&dep->qflgs) & ERTS_DE_QFLG_EXIT)); + erts_atomic32_read_bor_nob(&dep->qflgs, ERTS_DE_QFLG_EXIT); + erts_mtx_unlock(&dep->qlock); + + if (is_internal_port(dep->cid)) + erts_schedule_dist_command(NULL, dep); + else if (is_internal_pid(dep->cid)) + schedule_kill_dist_ctrl_proc(dep->cid); +} + void erts_kill_dist_connection(DistEntry *dep, Uint32 connection_id) { - erts_smp_de_rwlock(dep); - if (is_internal_port(dep->cid) - && connection_id == dep->connection_id - && !(dep->status & ERTS_DE_SFLG_EXITING)) { - - dep->status |= ERTS_DE_SFLG_EXITING; + erts_de_rwlock(dep); + if (connection_id == dep->connection_id + && dep->status == ERTS_DE_SFLG_CONNECTED) { - erts_smp_mtx_lock(&dep->qlock); - ASSERT(!(dep->qflgs & ERTS_DE_QFLG_EXIT)); - dep->qflgs |= ERTS_DE_QFLG_EXIT; - erts_smp_mtx_unlock(&dep->qlock); - - erts_schedule_dist_command(NULL, dep); + kill_connection(dep); } - erts_smp_de_rwunlock(dep); + erts_de_rwunlock(dep); } struct print_to_data { @@ -2517,9 +3048,6 @@ info_dist_entry(fmtfn_t to, void *arg, DistEntry *dep, int visible, int connecte } erts_print(to, arg, "Name: %T", dep->sysname); -#ifdef DEBUG - erts_print(to, arg, " (refc=%d)", erts_smp_refc_read(&dep->refc, 0)); -#endif erts_print(to, arg, "\n"); if (!connected && is_nil(dep->cid)) { if (dep->nlinks) { @@ -2563,6 +3091,10 @@ int distribution_info(fmtfn_t to, void *arg) /* Called by break handler */ info_dist_entry(to, arg, dep, 0, 1); } + for (dep = erts_pending_dist_entries; dep; dep = dep->next) { + info_dist_entry(to, arg, dep, 0, 0); + } + for (dep = erts_not_connected_dist_entries; dep; dep = dep->next) { if (dep != erts_this_dist_entry) { info_dist_entry(to, arg, dep, 0, 0); @@ -2585,7 +3117,6 @@ int distribution_info(fmtfn_t to, void *arg) /* Called by break handler */ monitor_node -- turn on/off node monitoring node controller only: - dist_exit/3 -- send exit signals from remote to local process dist_link/2 -- link a remote process to a local dist_unlink/2 -- unlink a remote from a local ****************************************************************************/ @@ -2595,15 +3126,6 @@ int distribution_info(fmtfn_t to, void *arg) /* Called by break handler */ /********************************************************************** ** Set the node name of current node fail if node already is set. ** setnode(name@host, Creation) - ** loads functions pointer to trap_functions from module erlang. - ** erlang:dsend/2 - ** erlang:dlink/1 - ** erlang:dunlink/1 - ** erlang:dmonitor_node/3 - ** erlang:dgroup_leader/2 - ** erlang:dexit/2 - ** -- are these needed ? - ** dexit/1 ***********************************************************************/ BIF_RETTYPE setnode_2(BIF_ALIST_2) @@ -2627,44 +3149,51 @@ BIF_RETTYPE setnode_2(BIF_ALIST_2) goto error; /* Check that all trap functions are defined !! */ - if (dsend2_trap->addressv[0] == NULL || - dsend3_trap->addressv[0] == NULL || - /* dsend_nosuspend_trap->address == NULL ||*/ - dlink_trap->addressv[0] == NULL || - dunlink_trap->addressv[0] == NULL || - dmonitor_node_trap->addressv[0] == NULL || - dgroup_leader_trap->addressv[0] == NULL || - dmonitor_p_trap->addressv[0] == NULL || - dexit_trap->addressv[0] == NULL) { + if (dmonitor_node_trap->addressv[0] == NULL || + dmonitor_p_trap->addressv[0] == NULL) { goto error; } - net_kernel = erts_whereis_process(BIF_P, ERTS_PROC_LOCK_MAIN, - am_net_kernel, ERTS_PROC_LOCK_MAIN, 0); - if (!net_kernel) + net_kernel = erts_whereis_process(BIF_P, + ERTS_PROC_LOCK_MAIN, + am_net_kernel, + ERTS_PROC_LOCK_MAIN|ERTS_PROC_LOCK_STATUS, + 0); + if (!net_kernel || ERTS_PROC_GET_DIST_ENTRY(net_kernel)) goto error; /* By setting F_DISTRIBUTION on net_kernel, - * do_net_exist will be called when net_kernel is terminated !! */ + * erts_do_net_exits will be called when net_kernel is terminated !! */ net_kernel->flags |= F_DISTRIBUTION; - if (net_kernel != BIF_P) - erts_smp_proc_unlock(net_kernel, ERTS_PROC_LOCK_MAIN); + erts_proc_unlock(net_kernel, + (ERTS_PROC_LOCK_STATUS + | ((net_kernel != BIF_P) + ? ERTS_PROC_LOCK_MAIN + : 0))); #ifdef DEBUG - erts_smp_rwmtx_rlock(&erts_dist_table_rwmtx); + erts_rwmtx_rlock(&erts_dist_table_rwmtx); ASSERT(!erts_visible_dist_entries && !erts_hidden_dist_entries); - erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx); + erts_rwmtx_runlock(&erts_dist_table_rwmtx); #endif - erts_smp_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN); - erts_smp_thr_progress_block(); + erts_proc_unlock(BIF_P, ERTS_PROC_LOCK_MAIN); + erts_thr_progress_block(); inc_no_nodes(); erts_set_this_node(BIF_ARG_1, (Uint32) creation); erts_is_alive = 1; send_nodes_mon_msgs(NULL, am_nodeup, BIF_ARG_1, am_visible, NIL); - erts_smp_thr_progress_unblock(); - erts_smp_proc_lock(BIF_P, ERTS_PROC_LOCK_MAIN); + erts_thr_progress_unblock(); + erts_proc_lock(BIF_P, ERTS_PROC_LOCK_MAIN); + + /* + * Note erts_this_dist_entry is changed by erts_set_this_node(), + * so we *need* to use the new one after erts_set_this_node() + * is called. + */ + erts_ref_dist_entry(erts_this_dist_entry); + ERTS_PROC_SET_DIST_ENTRY(net_kernel, erts_this_dist_entry); BIF_RET(am_true); @@ -2695,18 +3224,20 @@ BIF_RETTYPE setnode_3(BIF_ALIST_3) Eterm ic, oc; Eterm *tp; DistEntry *dep = NULL; + ErtsProcLocks proc_unlock = 0; + Process *proc; Port *pp = NULL; - - /* Prepare for success */ - ERTS_BIF_PREP_RET(ret, am_true); + Eterm notify_proc; + erts_aint32_t qflgs; /* * Check and pick out arguments */ if (!is_node_name_atom(BIF_ARG_1) || - is_not_internal_port(BIF_ARG_2) || - (erts_this_node->sysname == am_Noname)) { + !(is_internal_port(BIF_ARG_2) + || is_internal_pid(BIF_ARG_2)) + || (erts_this_node->sysname == am_Noname)) { goto badarg; } @@ -2725,21 +3256,25 @@ BIF_RETTYPE setnode_3(BIF_ALIST_3) if (!is_atom(ic) || !is_atom(oc)) goto badarg; - /* DFLAG_EXTENDED_REFERENCES is compulsory from R9 and forward */ - if (!(DFLAG_EXTENDED_REFERENCES & flags)) { + if (~flags & DFLAG_DIST_MANDATORY) { erts_dsprintf_buf_t *dsbufp = erts_create_logger_dsbuf(); erts_dsprintf(dsbufp, "%T", BIF_P->common.id); if (BIF_P->common.u.alive.reg) erts_dsprintf(dsbufp, " (%T)", BIF_P->common.u.alive.reg->name); erts_dsprintf(dsbufp, " attempted to enable connection to node %T " - "which is not able to handle extended references.\n", + "which does not support all mandatory capabilities.\n", BIF_ARG_1); erts_send_error_to_logger(BIF_P->group_leader, dsbufp); goto badarg; } /* + * ToDo: Should we not pass connection_id as well + * to make sure it's the right connection we commit. + */ + + /* * Arguments seem to be in order. */ @@ -2750,77 +3285,165 @@ BIF_RETTYPE setnode_3(BIF_ALIST_3) else if (!dep) goto system_limit; /* Should never happen!!! */ - pp = erts_id2port_sflgs(BIF_ARG_2, - BIF_P, - ERTS_PROC_LOCK_MAIN, - ERTS_PORT_SFLGS_INVALID_LOOKUP); - erts_smp_de_rwlock(dep); + if (is_internal_pid(BIF_ARG_2)) { + if (BIF_P->common.id == BIF_ARG_2) { + proc_unlock = 0; + proc = BIF_P; + } + else { + proc_unlock = ERTS_PROC_LOCK_MAIN; + proc = erts_pid2proc_not_running(BIF_P, ERTS_PROC_LOCK_MAIN, + BIF_ARG_2, proc_unlock); + } + erts_de_rwlock(dep); - if (!pp || (erts_atomic32_read_nob(&pp->state) - & ERTS_PORT_SFLG_EXITING)) - goto badarg; + if (!proc) + goto badarg; + else if (proc == ERTS_PROC_LOCK_BUSY) { + proc_unlock = 0; + goto yield; + } - if ((pp->drv_ptr->flags & ERL_DRV_FLAG_SOFT_BUSY) == 0) - goto badarg; + erts_proc_lock(proc, ERTS_PROC_LOCK_STATUS); + proc_unlock |= ERTS_PROC_LOCK_STATUS; - if (dep->cid == BIF_ARG_2 && pp->dist_entry == dep) - goto done; /* Already set */ + if (ERTS_PROC_GET_DIST_ENTRY(proc)) { + if (dep == ERTS_PROC_GET_DIST_ENTRY(proc) + && (proc->flags & F_DISTRIBUTION) + && dep->cid == BIF_ARG_2) { + ERTS_BIF_PREP_RET(ret, erts_make_dhandle(BIF_P, dep)); + goto done; + } + goto badarg; + } - if (dep->status & ERTS_DE_SFLG_EXITING) { - /* Suspend on dist entry waiting for the exit to finish */ - ErtsProcList *plp = erts_proclist_create(BIF_P); - plp->next = NULL; - erts_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, NULL); - erts_smp_mtx_lock(&dep->qlock); - erts_proclist_store_last(&dep->suspended, plp); - erts_smp_mtx_unlock(&dep->qlock); - goto yield; - } + if (dep->status & ERTS_DE_SFLG_EXITING) { + /* Suspend on dist entry waiting for the exit to finish */ + ErtsProcList *plp = erts_proclist_create(BIF_P); + plp->next = NULL; + erts_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, NULL); + erts_mtx_lock(&dep->qlock); + erts_proclist_store_last(&dep->suspended, plp); + erts_mtx_unlock(&dep->qlock); + goto yield; + } + if (dep->status != ERTS_DE_SFLG_PENDING) { + if (dep->status == 0) + erts_set_dist_entry_pending(dep); + else + goto badarg; + } - ASSERT(!(dep->status & ERTS_DE_SFLG_EXITING)); + if (is_not_nil(dep->cid)) + goto badarg; - if (pp->dist_entry || is_not_nil(dep->cid)) - goto badarg; + proc->flags |= F_DISTRIBUTION; + ERTS_PROC_SET_DIST_ENTRY(proc, dep); - erts_atomic32_read_bor_nob(&pp->state, ERTS_PORT_SFLG_DISTRIBUTION); + proc_unlock &= ~ERTS_PROC_LOCK_STATUS; + erts_proc_unlock(proc, ERTS_PROC_LOCK_STATUS); + + dep->send = NULL; /* Only for distr ports... */ - /* - * Dist-ports do not use the "busy port message queue" functionality, but - * instead use "busy dist entry" functionality. - */ - { - ErlDrvSizeT disable = ERL_DRV_BUSY_MSGQ_DISABLED; - erl_drv_busy_msgq_limits(ERTS_Port2ErlDrvPort(pp), &disable, NULL); } + else { - pp->dist_entry = dep; + pp = erts_id2port_sflgs(BIF_ARG_2, + BIF_P, + ERTS_PROC_LOCK_MAIN, + ERTS_PORT_SFLGS_INVALID_LOOKUP); + erts_de_rwlock(dep); - dep->version = version; - dep->creation = 0; + if (!pp || (erts_atomic32_read_nob(&pp->state) + & ERTS_PORT_SFLG_EXITING)) + goto badarg; - ASSERT(pp->drv_ptr->outputv || pp->drv_ptr->output); + if ((pp->drv_ptr->flags & ERL_DRV_FLAG_SOFT_BUSY) == 0) + goto badarg; -#if 1 - dep->send = (pp->drv_ptr->outputv - ? dist_port_commandv - : dist_port_command); -#else - dep->send = dist_port_command; -#endif - ASSERT(dep->send); + if (dep->cid == BIF_ARG_2 && pp->dist_entry == dep) { + ERTS_BIF_PREP_RET(ret, erts_make_dhandle(BIF_P, dep)); + goto done; /* Already set */ + } + + if (dep->status & ERTS_DE_SFLG_EXITING) { + /* Suspend on dist entry waiting for the exit to finish */ + ErtsProcList *plp = erts_proclist_create(BIF_P); + plp->next = NULL; + erts_suspend(BIF_P, ERTS_PROC_LOCK_MAIN, NULL); + erts_mtx_lock(&dep->qlock); + erts_proclist_store_last(&dep->suspended, plp); + erts_mtx_unlock(&dep->qlock); + goto yield; + } + if (dep->status != ERTS_DE_SFLG_PENDING) { + if (dep->status == 0) + erts_set_dist_entry_pending(dep); + else + goto badarg; + } + + if (pp->dist_entry || is_not_nil(dep->cid)) + goto badarg; + + erts_atomic32_read_bor_nob(&pp->state, ERTS_PORT_SFLG_DISTRIBUTION); + + pp->dist_entry = dep; + + ASSERT(pp->drv_ptr->outputv || pp->drv_ptr->output); + + dep->send = (pp->drv_ptr->outputv + ? dist_port_commandv + : dist_port_command); + ASSERT(dep->send); + + /* + * Dist-ports do not use the "busy port message queue" functionality, but + * instead use "busy dist entry" functionality. + */ + { + ErlDrvSizeT disable = ERL_DRV_BUSY_MSGQ_DISABLED; + erl_drv_busy_msgq_limits(ERTS_Port2ErlDrvPort(pp), &disable, NULL); + } + + } + + dep->version = version; + dep->creation = 0; #ifdef DEBUG - erts_smp_mtx_lock(&dep->qlock); - ASSERT(dep->qsize == 0); - erts_smp_mtx_unlock(&dep->qlock); + ASSERT(erts_atomic_read_nob(&dep->qsize) == 0 + || (dep->status & ERTS_DE_SFLG_PENDING)); #endif - erts_set_dist_entry_connected(dep, BIF_ARG_2, flags); - if (flags & DFLAG_DIST_HDR_ATOM_CACHE) create_cache(dep); - erts_smp_de_rwunlock(dep); + erts_set_dist_entry_connected(dep, BIF_ARG_2, flags); + + notify_proc = NIL; + if (erts_atomic_read_nob(&dep->qsize)) { + if (is_internal_port(dep->cid)) { + erts_schedule_dist_command(NULL, dep); + } + else { + qflgs = erts_atomic32_read_nob(&dep->qflgs); + if (qflgs & ERTS_DE_QFLG_REQ_INFO) { + qflgs = erts_atomic32_read_band_mb(&dep->qflgs, + ~ERTS_DE_QFLG_REQ_INFO); + if (qflgs & ERTS_DE_QFLG_REQ_INFO) { + notify_proc = dep->cid; + ASSERT(is_internal_pid(notify_proc)); + } + } + } + } + erts_de_rwunlock(dep); + if (is_internal_pid(notify_proc)) + notify_dist_data(BIF_P, notify_proc); + + ERTS_BIF_PREP_RET(ret, erts_make_dhandle(BIF_P, dep)); + dep = NULL; /* inc of refc transferred to port (dist_entry field) */ inc_no_nodes(); @@ -2833,13 +3456,16 @@ BIF_RETTYPE setnode_3(BIF_ALIST_3) done: if (dep && dep != erts_this_dist_entry) { - erts_smp_de_rwunlock(dep); + erts_de_rwunlock(dep); erts_deref_dist_entry(dep); } if (pp) erts_port_release(pp); + if (proc_unlock) + erts_proc_unlock(proc, proc_unlock); + return ret; yield: @@ -2856,83 +3482,168 @@ BIF_RETTYPE setnode_3(BIF_ALIST_3) goto done; } +BIF_RETTYPE erts_internal_new_connection_1(BIF_ALIST_1) +{ + DistEntry* dep; + Uint32 conn_id; + Eterm* hp; + Eterm dhandle; -/**********************************************************************/ -/* dist_exit(Local, Term, Remote) -> Bool */ + if (is_not_atom(BIF_ARG_1)) { + BIF_ERROR(BIF_P, BADARG); + } + dep = erts_find_or_insert_dist_entry(BIF_ARG_1); + + if (dep == erts_this_dist_entry) { + erts_deref_dist_entry(dep); + BIF_ERROR(BIF_P, BADARG); + } + + erts_de_rwlock(dep); -BIF_RETTYPE dist_exit_3(BIF_ALIST_3) + if (ERTS_DE_IS_CONNECTED(dep) || dep->status & ERTS_DE_SFLG_PENDING) + conn_id = dep->connection_id; + else if (dep->status == 0) { + erts_set_dist_entry_pending(dep); + conn_id = dep->connection_id; + } + else { + ASSERT(dep->status & ERTS_DE_SFLG_EXITING); + conn_id = (dep->connection_id + 1) & ERTS_DIST_CON_ID_MASK; + } + erts_de_rwunlock(dep); + hp = HAlloc(BIF_P, 3 + ERTS_MAGIC_REF_THING_SIZE); + dhandle = erts_build_dhandle(&hp, &BIF_P->off_heap, dep); + erts_deref_dist_entry(dep); + BIF_RET(TUPLE2(hp, make_small(conn_id), dhandle)); +} + +static Sint abort_connection(DistEntry* dep, Uint32 conn_id) { - Eterm local; - Eterm remote; - DistEntry *rdep; + erts_de_rwlock(dep); - local = BIF_ARG_1; - remote = BIF_ARG_3; + if (dep->connection_id != conn_id) + ; + else if (dep->status == ERTS_DE_SFLG_CONNECTED) { + kill_connection(dep); + } + else if (dep->status == ERTS_DE_SFLG_PENDING) { + NetExitsContext nec = {dep}; + ErtsLink *nlinks; + ErtsLink *node_links; + ErtsMonitor *monitors; + ErtsAtomCache *cache; + ErtsDistOutputBuf *obuf; + ErtsProcList *resume_procs; + Sint reds = 0; + + ASSERT(is_nil(dep->cid)); + + erts_de_links_lock(dep); + monitors = dep->monitors; + nlinks = dep->nlinks; + node_links = dep->node_links; + dep->monitors = NULL; + dep->nlinks = NULL; + dep->node_links = NULL; + erts_de_links_unlock(dep); + + cache = dep->cache; + dep->cache = NULL; + erts_mtx_lock(&dep->qlock); + obuf = dep->out_queue.first; + dep->out_queue.first = NULL; + dep->out_queue.last = NULL; + ASSERT(!dep->tmp_out_queue.first); + ASSERT(!dep->finalized_out_queue.first); + resume_procs = get_suspended_on_de(dep, ERTS_DE_QFLGS_ALL); + erts_mtx_unlock(&dep->qlock); + erts_atomic_set_nob(&dep->dist_cmd_scheduled, 0); + dep->send = NULL; + + erts_set_dist_entry_not_connected(dep); + + erts_de_rwunlock(dep); + + erts_sweep_monitors(monitors, &doit_monitor_net_exits, &nec); + erts_sweep_links(nlinks, &doit_link_net_exits, &nec); + erts_sweep_links(node_links, &doit_node_link_net_exits, &nec); + + if (resume_procs) { + int resumed = erts_resume_processes(resume_procs); + reds += resumed*ERTS_PORT_REDS_DIST_CMD_RESUMED; + } - /* Check that remote is a remote process */ - if (is_not_external_pid(remote)) - goto error; + delete_cache(cache); + free_de_out_queues(dep, obuf); + return reds; + } + erts_de_rwunlock(dep); + return 0; +} - rdep = external_dist_entry(remote); - - if(rdep == erts_this_dist_entry) - goto error; +BIF_RETTYPE erts_internal_abort_connection_2(BIF_ALIST_2) +{ + DistEntry* dep; + Eterm* tp; - /* Check that local is local */ - if (is_internal_pid(local)) { - Process *lp; - ErtsProcLocks lp_locks; - if (BIF_P->common.id == local) { - lp_locks = ERTS_PROC_LOCKS_ALL; - lp = BIF_P; - erts_smp_proc_lock(BIF_P, ERTS_PROC_LOCKS_ALL_MINOR); - } - else { - lp_locks = ERTS_PROC_LOCKS_XSIG_SEND; - lp = erts_pid2proc(BIF_P, ERTS_PROC_LOCK_MAIN, - local, lp_locks); - if (!lp) { - BIF_RET(am_true); /* ignore */ - } - } - - (void) erts_send_exit_signal(BIF_P, - remote, - lp, - &lp_locks, - BIF_ARG_2, - NIL, - NULL, - 0); -#ifdef ERTS_SMP - if (lp == BIF_P) - lp_locks &= ~ERTS_PROC_LOCK_MAIN; -#endif - erts_smp_proc_unlock(lp, lp_locks); - if (lp == BIF_P) { - erts_aint32_t state = erts_smp_atomic32_read_acqb(&BIF_P->state); - /* - * We may have exited current process and may have to take action. - */ - if (state & (ERTS_PSFLG_EXITING|ERTS_PSFLG_PENDING_EXIT)) { -#ifdef ERTS_SMP - if (state & ERTS_PSFLG_PENDING_EXIT) - erts_handle_pending_exit(BIF_P, ERTS_PROC_LOCK_MAIN); -#endif - ERTS_BIF_EXITED(BIF_P); - } - } + if (is_not_atom(BIF_ARG_1) || is_not_tuple_arity(BIF_ARG_2, 2)) { + BIF_ERROR(BIF_P, BADARG); } - else if (is_external_pid(local) - && external_dist_entry(local) == erts_this_dist_entry) { - BIF_RET(am_true); /* ignore */ + tp = tuple_val(BIF_ARG_2); + dep = erts_dhandle_to_dist_entry(tp[2]); + if (is_not_small(tp[1]) || dep != erts_find_dist_entry(BIF_ARG_1) + || dep == erts_this_dist_entry) { + BIF_ERROR(BIF_P, BADARG); + } + + if (dep) { + Sint reds = abort_connection(dep, unsigned_val(tp[1])); + BUMP_REDS(BIF_P, reds); } - else - goto error; BIF_RET(am_true); +} - error: - BIF_ERROR(BIF_P, BADARG); +int erts_auto_connect(DistEntry* dep, Process *proc, ErtsProcLocks proc_locks) +{ + erts_de_rwlock(dep); + if (dep->status != 0) { + erts_de_rwunlock(dep); + } + else { + Process* net_kernel; + ErtsProcLocks nk_locks = ERTS_PROC_LOCK_MSGQ; + Eterm *hp; + ErlOffHeap *ohp; + ErtsMessage *mp; + Eterm msg, dhandle; + Uint32 conn_id; + + erts_set_dist_entry_pending(dep); + conn_id = dep->connection_id; + erts_de_rwunlock(dep); + + net_kernel = erts_whereis_process(proc, proc_locks, + am_net_kernel, nk_locks, 0); + if (!net_kernel) { + abort_connection(dep, conn_id); + return 0; + } + + /* + * Send {auto_connect, Node, ConnId, DHandle} to net_kernel + */ + mp = erts_alloc_message_heap(net_kernel, &nk_locks, + 5 + ERTS_MAGIC_REF_THING_SIZE, + &hp, &ohp); + dhandle = erts_build_dhandle(&hp, ohp, dep); + msg = TUPLE4(hp, am_auto_connect, dep->sysname, make_small(conn_id), + dhandle); + erts_queue_message(net_kernel, nk_locks, mp, msg, proc->common.id); + erts_proc_unlock(net_kernel, nk_locks); + } + + return 1; } /**********************************************************************/ @@ -3004,13 +3715,15 @@ BIF_RETTYPE nodes_1(BIF_ALIST_1) length = 0; - erts_smp_rwmtx_rlock(&erts_dist_table_rwmtx); + erts_rwmtx_rlock(&erts_dist_table_rwmtx); ASSERT(erts_no_of_not_connected_dist_entries > 0); ASSERT(erts_no_of_hidden_dist_entries >= 0); + ASSERT(erts_no_of_pending_dist_entries >= 0); ASSERT(erts_no_of_visible_dist_entries >= 0); if(not_connected) - length += (erts_no_of_not_connected_dist_entries - 1); + length += ((erts_no_of_not_connected_dist_entries - 1) + + erts_no_of_pending_dist_entries); if(hidden) length += erts_no_of_hidden_dist_entries; if(visible) @@ -3021,7 +3734,7 @@ BIF_RETTYPE nodes_1(BIF_ALIST_1) result = NIL; if (length == 0) { - erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx); + erts_rwmtx_runlock(&erts_dist_table_rwmtx); goto done; } @@ -3030,13 +3743,18 @@ BIF_RETTYPE nodes_1(BIF_ALIST_1) #ifdef DEBUG endp = hp + length*2; #endif - if(not_connected) + if(not_connected) { for(dep = erts_not_connected_dist_entries; dep; dep = dep->next) { if (dep != erts_this_dist_entry) { result = CONS(hp, dep->sysname, result); hp += 2; } + } + for(dep = erts_pending_dist_entries; dep; dep = dep->next) { + result = CONS(hp, dep->sysname, result); + hp += 2; } + } if(hidden) for(dep = erts_hidden_dist_entries; dep; dep = dep->next) { result = CONS(hp, dep->sysname, result); @@ -3052,7 +3770,7 @@ BIF_RETTYPE nodes_1(BIF_ALIST_1) hp += 2; } ASSERT(endp == hp); - erts_smp_rwmtx_runlock(&erts_dist_table_rwmtx); + erts_rwmtx_runlock(&erts_dist_table_rwmtx); done: UnUseTmpHeap(2,BIF_P); @@ -3081,11 +3799,17 @@ monitor_node(Process* p, Eterm Node, Eterm Bool, Eterm Options) DistEntry *dep; ErtsLink *lnk; Eterm l; + int async_connect = 1; for (l = Options; l != NIL && is_list(l); l = CDR(list_val(l))) { Eterm t = CAR(list_val(l)); - /* allow_passive_connect the only available option right now */ - if (t != am_allow_passive_connect) { + if (t == am_allow_passive_connect) { + /* + * Handle this horrible feature by falling back on old synchronous + * auto-connect (if needed) + */ + async_connect = 0; + } else { BIF_ERROR(p, BADARG); } } @@ -3099,54 +3823,93 @@ monitor_node(Process* p, Eterm Node, Eterm Bool, Eterm Options) && (Node != erts_this_node->sysname))) { BIF_ERROR(p, BADARG); } - dep = erts_sysname_to_connected_dist_entry(Node); - if (!dep) { - do_trap: - BIF_TRAP3(dmonitor_node_trap, p, Node, Bool, Options); - } - if (dep == erts_this_dist_entry) - goto done; - - erts_smp_proc_lock(p, ERTS_PROC_LOCK_LINK); - erts_smp_de_rlock(dep); - if (ERTS_DE_IS_NOT_CONNECTED(dep)) { - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_LINK); - erts_smp_de_runlock(dep); - goto do_trap; - } - erts_smp_de_links_lock(dep); - erts_smp_de_runlock(dep); if (Bool == am_true) { - ASSERT(dep->cid != NIL); - lnk = erts_add_or_lookup_link(&(dep->node_links), LINK_NODE, - p->common.id); - ++ERTS_LINK_REFC(lnk); - lnk = erts_add_or_lookup_link(&ERTS_P_LINKS(p), LINK_NODE, Node); - ++ERTS_LINK_REFC(lnk); - } - else { - lnk = erts_lookup_link(dep->node_links, p->common.id); - if (lnk != NULL) { - if ((--ERTS_LINK_REFC(lnk)) == 0) { - erts_destroy_link(erts_remove_link(&(dep->node_links), - p->common.id)); - } - } - lnk = erts_lookup_link(ERTS_P_LINKS(p), Node); - if (lnk != NULL) { - if ((--ERTS_LINK_REFC(lnk)) == 0) { - erts_destroy_link(erts_remove_link(&ERTS_P_LINKS(p), - Node)); + ErtsDSigData dsd; + dsd.node = Node; + dep = erts_find_or_insert_dist_entry(Node); + if (dep == erts_this_dist_entry) + goto done; + + erts_proc_lock(p, ERTS_PROC_LOCK_LINK); + + switch (erts_dsig_prepare(&dsd, dep, p, + (ERTS_PROC_LOCK_MAIN | ERTS_PROC_LOCK_LINK), + ERTS_DSP_RLOCK, 0, async_connect)) { + case ERTS_DSIG_PREP_NOT_ALIVE: + case ERTS_DSIG_PREP_NOT_CONNECTED: + /* Trap to either send 'nodedown' or do passive connection attempt */ + trap: + erts_proc_unlock(p, ERTS_PROC_LOCK_LINK); + erts_deref_dist_entry(dep); + BIF_TRAP3(dmonitor_node_trap, p, Node, Bool, Options); + case ERTS_DSIG_PREP_PENDING: + if (!async_connect) { + /* + * Pending connection may fail, so we must trap + * to ensure passive connection attempt + */ + erts_de_runlock(dep); + goto trap; } - } + /*fall through*/ + case ERTS_DSIG_PREP_CONNECTED: + erts_de_links_lock(dep); + erts_de_runlock(dep); + lnk = erts_add_or_lookup_link(&(dep->node_links), LINK_NODE, + p->common.id); + ++ERTS_LINK_REFC(lnk); + lnk = erts_add_or_lookup_link(&ERTS_P_LINKS(p), LINK_NODE, Node); + ++ERTS_LINK_REFC(lnk); + erts_de_links_unlock(dep); + break; + default: + ERTS_ASSERT(! "Invalid dsig prepare result"); + } + erts_deref_dist_entry(dep); + } + else { /* Bool == false */ + dep = erts_sysname_to_connected_dist_entry(Node); + if (!dep) { + /* + * Before OTP-21 this case triggered auto-connect + * and a 'nodedown' message if that failed. + * Now it's a simple no-op which feels more reasonable. + */ + BIF_RET(am_true); + } + if (dep == erts_this_dist_entry) + goto done; + + erts_proc_lock(p, ERTS_PROC_LOCK_LINK); + erts_de_rlock(dep); + if (!(dep->status & (ERTS_DE_SFLG_PENDING | ERTS_DE_SFLG_CONNECTED))) { + erts_proc_unlock(p, ERTS_PROC_LOCK_LINK); + erts_de_runlock(dep); + goto done; + } + erts_de_links_lock(dep); + erts_de_runlock(dep); + lnk = erts_lookup_link(dep->node_links, p->common.id); + if (lnk != NULL) { + if ((--ERTS_LINK_REFC(lnk)) == 0) { + erts_destroy_link(erts_remove_link(&(dep->node_links), + p->common.id)); + } + } + lnk = erts_lookup_link(ERTS_P_LINKS(p), Node); + if (lnk != NULL) { + if ((--ERTS_LINK_REFC(lnk)) == 0) { + erts_destroy_link(erts_remove_link(&ERTS_P_LINKS(p), + Node)); + } + } + erts_de_links_unlock(dep); } - erts_smp_de_links_unlock(dep); - erts_smp_proc_unlock(p, ERTS_PROC_LOCK_LINK); + erts_proc_unlock(p, ERTS_PROC_LOCK_LINK); done: - erts_deref_dist_entry(dep); BIF_RET(am_true); } @@ -3175,9 +3938,9 @@ BIF_RETTYPE net_kernel_dflag_unicode_io_1(BIF_ALIST_1) if (de == erts_this_dist_entry) { BIF_RET(am_true); } - erts_smp_de_rlock(de); + erts_de_rlock(de); f = de->flags; - erts_smp_de_runlock(de); + erts_de_runlock(de); BIF_RET(((f & DFLAG_UNICODE_IO) ? am_true : am_false)); } @@ -3207,7 +3970,7 @@ struct ErtsNodesMonitor_ { Uint16 no; }; -static erts_smp_mtx_t nodes_monitors_mtx; +static erts_mtx_t nodes_monitors_mtx; static ErtsNodesMonitor *nodes_monitors; static ErtsNodesMonitor *nodes_monitors_end; @@ -3225,7 +3988,7 @@ static ErtsNodesMonitor *nodes_monitors_end; static void init_nodes_monitors(void) { - erts_smp_mtx_init(&nodes_monitors_mtx, "nodes_monitors", NIL, + erts_mtx_init(&nodes_monitors_mtx, "nodes_monitors", NIL, ERTS_LOCK_FLAGS_PROPERTY_STATIC | ERTS_LOCK_FLAGS_CATEGORY_DISTRIBUTION); nodes_monitors = NULL; nodes_monitors_end = NULL; @@ -3351,10 +4114,10 @@ send_nodes_mon_msgs(Process *c_p, Eterm what, Eterm node, Eterm type, Eterm reas } #endif - ERTS_SMP_LC_ASSERT(!c_p + ERTS_LC_ASSERT(!c_p || (erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN)); - erts_smp_mtx_lock(&nodes_monitors_mtx); + erts_mtx_lock(&nodes_monitors_mtx); for (nmp = nodes_monitors; nmp; nmp = nmp->next) { int i; @@ -3387,7 +4150,7 @@ send_nodes_mon_msgs(Process *c_p, Eterm what, Eterm node, Eterm type, Eterm reas if (rp) { if (rp == c_p) rp_locks &= ~ERTS_PROC_LOCK_MAIN; - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } rp = nmp->proc; @@ -3414,10 +4177,10 @@ send_nodes_mon_msgs(Process *c_p, Eterm what, Eterm node, Eterm type, Eterm reas if (rp) { if (rp == c_p) rp_locks &= ~ERTS_PROC_LOCK_MAIN; - erts_smp_proc_unlock(rp, rp_locks); + erts_proc_unlock(rp, rp_locks); } - erts_smp_mtx_unlock(&nodes_monitors_mtx); + erts_mtx_unlock(&nodes_monitors_mtx); } static Eterm @@ -3427,8 +4190,8 @@ insert_nodes_monitor(Process *c_p, Uint32 opts) Eterm res = am_false; ErtsNodesMonitor *xnmp, *nmp; - ERTS_SMP_LC_ASSERT(erts_smp_lc_mtx_is_locked(&nodes_monitors_mtx)); - ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); + ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&nodes_monitors_mtx)); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); xnmp = c_p->nodes_monitors; if (xnmp) { @@ -3512,8 +4275,8 @@ remove_nodes_monitors(Process *c_p, Uint32 opts, int all) Eterm res = am_false; ErtsNodesMonitor *nmp; - ERTS_SMP_LC_ASSERT(erts_smp_lc_mtx_is_locked(&nodes_monitors_mtx)); - ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); + ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&nodes_monitors_mtx)); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) & ERTS_PROC_LOCK_MAIN); nmp = c_p->nodes_monitors; ASSERT(!nmp || !nmp->prev || nmp->prev->proc != c_p); @@ -3555,23 +4318,23 @@ remove_nodes_monitors(Process *c_p, Uint32 opts, int all) void erts_delete_nodes_monitors(Process *c_p, ErtsProcLocks locks) { -#if defined(ERTS_ENABLE_LOCK_CHECK) && defined(ERTS_SMP) +#if defined(ERTS_ENABLE_LOCK_CHECK) if (c_p) { ErtsProcLocks might_unlock = locks & ~ERTS_PROC_LOCK_MAIN; if (might_unlock) erts_proc_lc_might_unlock(c_p, might_unlock); } #endif - if (erts_smp_mtx_trylock(&nodes_monitors_mtx) == EBUSY) { + if (erts_mtx_trylock(&nodes_monitors_mtx) == EBUSY) { ErtsProcLocks unlock_locks = locks & ~ERTS_PROC_LOCK_MAIN; if (c_p && unlock_locks) - erts_smp_proc_unlock(c_p, unlock_locks); - erts_smp_mtx_lock(&nodes_monitors_mtx); + erts_proc_unlock(c_p, unlock_locks); + erts_mtx_lock(&nodes_monitors_mtx); if (c_p && unlock_locks) - erts_smp_proc_lock(c_p, unlock_locks); + erts_proc_lock(c_p, unlock_locks); } remove_nodes_monitors(c_p, 0, 1); - erts_smp_mtx_unlock(&nodes_monitors_mtx); + erts_mtx_unlock(&nodes_monitors_mtx); } Eterm @@ -3582,7 +4345,7 @@ erts_monitor_nodes(Process *c_p, Eterm on, Eterm olist) Uint16 opts = (Uint16) 0; ASSERT(c_p); - ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN); if (on != am_true && on != am_false) return THE_NON_VALUE; @@ -3638,14 +4401,14 @@ erts_monitor_nodes(Process *c_p, Eterm on, Eterm olist) return THE_NON_VALUE; } - erts_smp_mtx_lock(&nodes_monitors_mtx); + erts_mtx_lock(&nodes_monitors_mtx); if (on == am_true) res = insert_nodes_monitor(c_p, opts); else res = remove_nodes_monitors(c_p, opts, 0); - erts_smp_mtx_unlock(&nodes_monitors_mtx); + erts_mtx_unlock(&nodes_monitors_mtx); return res; } @@ -3668,8 +4431,8 @@ erts_processes_monitoring_nodes(Process *c_p) #endif ASSERT(c_p); - ERTS_SMP_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN); - erts_smp_mtx_lock(&nodes_monitors_mtx); + ERTS_LC_ASSERT(erts_proc_lc_my_proc_locks(c_p) == ERTS_PROC_LOCK_MAIN); + erts_mtx_lock(&nodes_monitors_mtx); sz = 0; szp = &sz; @@ -3718,7 +4481,7 @@ erts_processes_monitoring_nodes(Process *c_p) ASSERT(hp == hend); - erts_smp_mtx_unlock(&nodes_monitors_mtx); + erts_mtx_unlock(&nodes_monitors_mtx); return res; } |