From f2c4f6f83deecba0c2527e520f0f18fba7d84815 Mon Sep 17 00:00:00 2001 From: Lukas Larsson Date: Wed, 26 Sep 2018 11:55:01 +0200 Subject: erts: Implement fragmentation of distrubution messages --- erts/emulator/beam/dist.c | 929 ++++++++++++++++++++++++++++++++-------------- 1 file changed, 653 insertions(+), 276 deletions(-) (limited to 'erts/emulator/beam/dist.c') diff --git a/erts/emulator/beam/dist.c b/erts/emulator/beam/dist.c index 91a9481d84..78674848a1 100644 --- a/erts/emulator/beam/dist.c +++ b/erts/emulator/beam/dist.c @@ -55,15 +55,21 @@ */ #if 0 #define ERTS_DIST_MSG_DBG +FILE *dbg_file; #endif #if 0 +/* Enable this to print the dist debug messages to a file instead */ +#define ERTS_DIST_MSG_DBG_FILE "/tmp/dist_dbg.%d" +#endif +#if 0 +/* Enable this to print the raw bytes sent and received */ #define ERTS_RAW_DIST_MSG_DBG #endif #if defined(ERTS_DIST_MSG_DBG) || defined(ERTS_RAW_DIST_MSG_DBG) static void bw(byte *buf, ErlDrvSizeT sz) { - bin_write(ERTS_PRINT_STDERR, NULL, buf, sz); + bin_write(ERTS_PRINT_FILE, dbg_file, buf, sz); } #endif @@ -71,37 +77,93 @@ static void bw(byte *buf, ErlDrvSizeT sz) static void dist_msg_dbg(ErtsDistExternal *edep, char *what, byte *buf, int sz) { - ErtsHeapFactory factory; - byte *extp = edep->extp; + byte *extp = edep->data->extp; Eterm msg; Sint ctl_len; - Sint size = ctl_len = erts_decode_dist_ext_size(edep); + Sint size = ctl_len = erts_decode_dist_ext_size(edep, 0); if (size < 0) { - erts_fprintf(stderr, + erts_fprintf(dbg_file, "DIST MSG DEBUG: erts_decode_dist_ext_size(%s) failed:\n", what); bw(buf, sz); } else { - ErlHeapFragment *mbuf = new_message_buffer(size); - erts_factory_static_init(&factory, mbuf->mem, ctl_len, &mbuf->off_heap); - msg = erts_decode_dist_ext(&factory, edep); + ErtsHeapFactory factory; + ErtsMessage *mbuf = erts_factory_message_create(&factory, NULL, 0, ctl_len); + /* Set mbuf msg to NIL as erts_factory_undo will fail otherwise */ + ERL_MESSAGE_TERM(mbuf) = NIL; + msg = erts_decode_dist_ext(&factory, edep, 0); if (is_value(msg)) - erts_fprintf(stderr, " %s: %.80T\n", what, msg); + erts_fprintf(dbg_file, " %s: %.80T\n", what, msg); else { - erts_fprintf(stderr, + erts_fprintf(dbg_file, "DIST MSG DEBUG: erts_decode_dist_ext(%s) failed:\n", what); bw(buf, sz); } - free_message_buffer(mbuf); - edep->extp = extp; + erts_factory_undo(&factory); + edep->data->extp = extp; } } +static char *erts_dop_to_string(enum dop dop) { + if (dop == DOP_LINK) + return "LINK"; + if (dop == DOP_SEND) + return "SEND"; + if (dop == DOP_EXIT) + return "EXIT"; + if (dop == DOP_UNLINK) + return "UNLINK"; + if (dop == DOP_REG_SEND) + return "REG_SEND"; + if (dop == DOP_GROUP_LEADER) + return "GROUP_LEADER"; + if (dop == DOP_EXIT2) + return "EXIT2"; + if (dop == DOP_SEND_TT) + return "SEND_TT"; + if (dop == DOP_EXIT_TT) + return "EXIT_TT"; + if (dop == DOP_REG_SEND_TT) + return "REG_SEND_TT"; + if (dop == DOP_EXIT2_TT) + return "EXIT2_TT"; + if (dop == DOP_MONITOR_P) + return "MONITOR_P"; + if (dop == DOP_DEMONITOR_P) + return "DEMONITOR_P"; + if (dop == DOP_MONITOR_P_EXIT) + return "MONITOR_P_EXIT"; + if (dop == DOP_SEND_SENDER) + return "SEND_SENDER"; + if (dop == DOP_SEND_SENDER_TT) + return "SEND_SENDER_TT"; + if (dop == DOP_PAYLOAD_EXIT) + return "PAYLOAD_EXIT"; + if (dop == DOP_PAYLOAD_EXIT_TT) + return "PAYLOAD_EXIT_TT"; + if (dop == DOP_PAYLOAD_EXIT2) + return "PAYLOAD_EXIT2"; + if (dop == DOP_PAYLOAD_EXIT2_TT) + return "PAYLOAD_EXIT2_TT"; + if (dop == DOP_PAYLOAD_MONITOR_P_EXIT) + return "PAYLOAD_MONITOR_P_EXIT"; + ASSERT(0); + return "UNKNOWN"; +} + #endif +#if defined(VALGRIND) +#include +#include +# define PURIFY_MSG(msg) \ + VALGRIND_PRINTF("%s, line %d: %s", __FILE__, __LINE__, msg) +#else +# define PURIFY_MSG(msg) +#endif int erts_is_alive; /* System must be blocked on change */ int erts_dist_buf_busy_limit; @@ -115,13 +177,16 @@ static Export *dist_ctrl_put_data_trap; /* forward declarations */ -static int dsig_send_exit(ErtsDSigData* dsdp, Eterm ctl, Eterm msg, int force_busy); +static int dsig_send_exit(ErtsDSigData* dsdp, Eterm ctl, Eterm msg, Eterm from, int force_busy); 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 ErtsDistOutputBuf* clear_de_out_queues(DistEntry*); static void free_de_out_queues(DistEntry*, ErtsDistOutputBuf*); +int erts_dist_seq_tree_foreach_delete_yielding(DistSeqNode **root, + void **vyspp, + Sint limit); static erts_atomic_t no_caches; static erts_atomic_t no_nodes; @@ -142,7 +207,6 @@ delete_cache(ErtsAtomCache *cache) } } - static void create_cache(DistEntry *dep) { @@ -207,12 +271,14 @@ typedef enum { ERTS_CML_CLEANUP_STATE_LINKS, ERTS_CML_CLEANUP_STATE_MONITORS, ERTS_CML_CLEANUP_STATE_ONAME_MONITORS, + ERTS_CML_CLEANUP_STATE_SEQUENCES, ERTS_CML_CLEANUP_STATE_NODE_MONITORS -} ErtsConMonLnkCleaupState; +} ErtsConMonLnkSeqCleanupState; typedef struct { - ErtsConMonLnkCleaupState state; + ErtsConMonLnkSeqCleanupState state; ErtsMonLnkDist *dist; + DistSeqNode *seq; void *yield_state; int trigger_node_monitors; Eterm nodename; @@ -220,12 +286,12 @@ typedef struct { Eterm reason; ErlOffHeap oh; Eterm heap[1]; -} ErtsConMonLnkCleanup; +} ErtsConMonLnkSeqCleanup; static void -con_monitor_link_cleanup(void *vcmlcp) +con_monitor_link_seq_cleanup(void *vcmlcp) { - ErtsConMonLnkCleanup *cmlcp = vcmlcp; + ErtsConMonLnkSeqCleanup *cmlcp = vcmlcp; ErtsMonLnkDist *dist = cmlcp->dist; ErtsSchedulerData *esdp; int reds = CONTEXT_REDS; @@ -262,6 +328,15 @@ con_monitor_link_cleanup(void *vcmlcp) cmlcp->dist = NULL; erts_mon_link_dist_dec_refc(dist); + ASSERT(!cmlcp->yield_state); + cmlcp->state = ERTS_CML_CLEANUP_STATE_SEQUENCES; + case ERTS_CML_CLEANUP_STATE_SEQUENCES: + reds = erts_dist_seq_tree_foreach_delete_yielding(&cmlcp->seq, + &cmlcp->yield_state, + reds); + if (reds <= 0) + break; + ASSERT(!cmlcp->yield_state); cmlcp->state = ERTS_CML_CLEANUP_STATE_NODE_MONITORS; case ERTS_CML_CLEANUP_STATE_NODE_MONITORS: @@ -282,22 +357,23 @@ con_monitor_link_cleanup(void *vcmlcp) esdp = erts_get_scheduler_data(); ASSERT(esdp && esdp->type == ERTS_SCHED_NORMAL); erts_schedule_misc_aux_work((int) esdp->no, - con_monitor_link_cleanup, + con_monitor_link_seq_cleanup, (void *) cmlcp); } static void -schedule_con_monitor_link_cleanup(ErtsMonLnkDist *dist, - Eterm nodename, - Eterm visability, - Eterm reason) +schedule_con_monitor_link_seq_cleanup(ErtsMonLnkDist *dist, + DistSeqNode *seq, + Eterm nodename, + Eterm visability, + Eterm reason) { - if (dist || is_value(nodename)) { + if (dist || is_value(nodename) || seq) { ErtsSchedulerData *esdp; - ErtsConMonLnkCleanup *cmlcp; + ErtsConMonLnkSeqCleanup *cmlcp; Uint rsz, size; - size = sizeof(ErtsConMonLnkCleanup); + size = sizeof(ErtsConMonLnkSeqCleanup); if (is_non_value(reason) || is_immed(reason)) { rsz = 0; @@ -324,6 +400,8 @@ schedule_con_monitor_link_cleanup(ErtsMonLnkDist *dist, erts_mtx_unlock(&dist->mtx); } + cmlcp->seq = seq; + cmlcp->trigger_node_monitors = is_value(nodename); cmlcp->nodename = nodename; cmlcp->visability = visability; @@ -337,13 +415,13 @@ schedule_con_monitor_link_cleanup(ErtsMonLnkDist *dist, esdp = erts_get_scheduler_data(); ASSERT(esdp && esdp->type == ERTS_SCHED_NORMAL); erts_schedule_misc_aux_work((int) esdp->no, - con_monitor_link_cleanup, + con_monitor_link_seq_cleanup, (void *) cmlcp); } } /* -** A full node name constists of a "n@h" +** A full node name consists of a "n@h" ** ** n must be a valid node name: string of ([a-z][A-Z][0-9]_-)+ ** @@ -560,6 +638,7 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) } else { /* Call from distribution controller (port/process) */ ErtsMonLnkDist *mld; + DistSeqNode *sequences; ErtsAtomCache *cache; ErtsProcList *suspendees; ErtsDistOutputBuf *obuf; @@ -589,6 +668,9 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) mld = dep->mld; dep->mld = NULL; + sequences = dep->sequences; + dep->sequences = NULL; + nodename = dep->sysname; flags = dep->flags; @@ -612,14 +694,15 @@ int erts_do_net_exits(DistEntry *dep, Eterm reason) erts_de_rwunlock(dep); - schedule_con_monitor_link_cleanup(mld, - nodename, - (flags & DFLAG_PUBLISHED - ? am_visible - : am_hidden), - (reason == am_normal - ? am_connection_closed - : reason)); + schedule_con_monitor_link_seq_cleanup(mld, + sequences, + nodename, + (flags & DFLAG_PUBLISHED + ? am_visible + : am_hidden), + (reason == am_normal + ? am_connection_closed + : reason)); erts_resume_processes(suspendees); @@ -653,6 +736,16 @@ void init_dist(void) { init_nodes_monitors(); +#ifdef ERTS_DIST_MSG_DBG_FILE + { + char buff[255]; + sprintf(buff, ERTS_DIST_MSG_DBG_FILE, getpid()); + dbg_file = fopen(buff,"w+"); + } +#elif defined (ERTS_DIST_MSG_DBG) + dbg_file = stderr; +#endif + nodedown.reason = NIL; nodedown.bp = NULL; @@ -676,21 +769,27 @@ void init_dist(void) } } -#define ErtsDistOutputBuf2Binary(OB) \ - ((Binary *) (((char *) (OB)) - offsetof(Binary, orig_bytes))) +#define ErtsDistOutputBuf2Binary(OB) OB->bin static ERTS_INLINE ErtsDistOutputBuf * -alloc_dist_obuf(Uint size) +alloc_dist_obuf(Uint size, Uint headers) { + int i; ErtsDistOutputBuf *obuf; - Uint obuf_size = sizeof(ErtsDistOutputBuf)+sizeof(byte)*(size-1); + Uint obuf_size = sizeof(ErtsDistOutputBuf)*(headers) + + sizeof(byte)*size; Binary *bin = erts_bin_drv_alloc(obuf_size); - obuf = (ErtsDistOutputBuf *) &bin->orig_bytes[0]; + obuf = (ErtsDistOutputBuf *) &bin->orig_bytes[size]; + erts_refc_add(&bin->intern.refc, headers - 1, 1); + for (i = 0; i < headers; i++) { + obuf[i].bin = bin; + obuf[i].extp = (byte *)&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)); + obuf[i].dbg_pattern = ERTS_DIST_OUTPUT_BUF_DBG_PATTERN; + obuf[i].alloc_endp = obuf->extp + size; + ASSERT(bin == ErtsDistOutputBuf2Binary(obuf)); #endif + } return obuf; } @@ -705,8 +804,8 @@ free_dist_obuf(ErtsDistOutputBuf *obuf) static ERTS_INLINE Sint size_obuf(ErtsDistOutputBuf *obuf) { - Binary *bin = ErtsDistOutputBuf2Binary(obuf); - return bin->orig_size; + return sizeof(ErtsDistOutputBuf) + (obuf->ext_endp - obuf->ext_start) + + (obuf->hdr_endp - obuf->hdrp); } static ErtsDistOutputBuf* clear_de_out_queues(DistEntry* dep) @@ -771,7 +870,9 @@ int erts_dsend_context_dtor(Binary* ctx_bin) default:; } if (ctx->dss.phase >= ERTS_DSIG_SEND_PHASE_ALLOC && ctx->dss.obuf) { - free_dist_obuf(ctx->dss.obuf); + int i; + for (i = 0; i < ctx->dss.fragments; i++) + free_dist_obuf(&ctx->dss.obuf[i]); } if (ctx->deref_dep) erts_deref_dist_entry(ctx->dep); @@ -845,8 +946,8 @@ erts_dsig_send_unlink(ErtsDSigData *dsdp, Eterm local, Eterm remote) /* A local process that's being monitored by a remote one exits. We send: {DOP_MONITOR_P_EXIT, Local pid or name, Remote pid, ref, reason} */ int -erts_dsig_send_m_exit(ErtsDSigData *dsdp, Eterm watcher, Eterm watched, - Eterm ref, Eterm reason) +erts_dsig_send_m_exit(ErtsDSigData *dsdp, Eterm from, Eterm watcher, Eterm watched, + Eterm ref, Eterm reason) { Eterm ctl, msg; DeclareTmpHeapNoproc(ctl_heap,6); @@ -871,7 +972,7 @@ erts_dsig_send_m_exit(ErtsDSigData *dsdp, Eterm watcher, Eterm watched, msg = THE_NON_VALUE; } - res = dsig_send_exit(dsdp, ctl, msg, 1); + res = dsig_send_exit(dsdp, ctl, msg, from, 1); UnUseTmpHeapNoproc(6); return res; } @@ -1031,8 +1132,7 @@ erts_dsig_send_msg(Eterm remote, Eterm message, ErtsSendContext* ctx) } int -erts_dsig_send_reg_msg(Eterm remote_name, Eterm message, - ErtsSendContext* ctx) +erts_dsig_send_reg_msg(Eterm remote_name, Eterm message, ErtsSendContext* ctx) { Eterm ctl; Eterm token = NIL; @@ -1149,13 +1249,13 @@ erts_dsig_send_exit_tt(ErtsDSigData *dsdp, Eterm local, Eterm remote, #endif DTRACE7(process_exit_signal_remote, sender_name, node_name, remote_name, reason_str, tok_label, tok_lastcnt, tok_serial); - res = dsig_send_exit(dsdp, ctl, msg, 1); + res = dsig_send_exit(dsdp, ctl, msg, local, 1); UnUseTmpHeapNoproc(6); return res; } int -erts_dsig_send_exit(ErtsDSigData *dsdp, Eterm local, Eterm remote, Eterm reason) +erts_dsig_send_exit(ErtsDSigData *dsdp, Eterm from, Eterm local, Eterm remote, Eterm reason) { DeclareTmpHeapNoproc(ctl_heap,5); int res; @@ -1170,7 +1270,7 @@ erts_dsig_send_exit(ErtsDSigData *dsdp, Eterm local, Eterm remote, Eterm reason) ctl = TUPLE4(&ctl_heap[0], make_small(DOP_EXIT), local, remote, reason); msg = THE_NON_VALUE; } - res = dsig_send_exit(dsdp, ctl, msg, 1); + res = dsig_send_exit(dsdp, ctl, msg, from, 1); UnUseTmpHeapNoproc(5); return res; } @@ -1193,7 +1293,7 @@ erts_dsig_send_exit2(ErtsDSigData *dsdp, Eterm local, Eterm remote, Eterm reason msg = THE_NON_VALUE; } - res = dsig_send_exit(dsdp, ctl, msg, 0); + res = dsig_send_exit(dsdp, ctl, msg, local, 0); UnUseTmpHeapNoproc(5); return res; } @@ -1215,17 +1315,123 @@ erts_dsig_send_group_leader(ErtsDSigData *dsdp, Eterm leader, Eterm remote) return res; } -#if defined(PURIFY) -# define PURIFY_MSG(msg) \ - purify_printf("%s, line %d: %s", __FILE__, __LINE__, msg) -#elif defined(VALGRIND) -#include -#include -# define PURIFY_MSG(msg) \ - VALGRIND_PRINTF("%s, line %d: %s", __FILE__, __LINE__, msg) -#else -# define PURIFY_MSG(msg) -#endif +struct dist_sequences { + ErlHeapFragment hfrag; + struct dist_sequences *parent; + struct dist_sequences *left; + struct dist_sequences *right; + char is_red; + + Uint64 seq_id; + int cnt; + Sint ctl_len; +}; + +#define ERTS_RBT_PREFIX dist_seq +#define ERTS_RBT_T DistSeqNode +#define ERTS_RBT_KEY_T Uint +#define ERTS_RBT_FLAGS_T int +#define ERTS_RBT_INIT_EMPTY_TNODE(T) \ + do { \ + (T)->parent = NULL; \ + (T)->left = NULL; \ + (T)->right = NULL; \ + (T)->is_red = 0; \ + } while(0) +#define ERTS_RBT_IS_RED(T) ((T)->is_red) +#define ERTS_RBT_SET_RED(T) ((T)->is_red = 1) +#define ERTS_RBT_IS_BLACK(T) (!ERTS_RBT_IS_RED(T)) +#define ERTS_RBT_SET_BLACK(T) ((T)->is_red = 0) +#define ERTS_RBT_GET_FLAGS(T) ((T)->is_red) +#define ERTS_RBT_SET_FLAGS(T, F) ((T)->is_red = F) +#define ERTS_RBT_GET_PARENT(T) ((T)->parent) +#define ERTS_RBT_SET_PARENT(T, P) ((T)->parent = P) +#define ERTS_RBT_GET_RIGHT(T) ((T)->right) +#define ERTS_RBT_SET_RIGHT(T, R) ((T)->right = (R)) +#define ERTS_RBT_GET_LEFT(T) ((T)->left) +#define ERTS_RBT_SET_LEFT(T, L) ((T)->left = (L)) +#define ERTS_RBT_GET_KEY(T) ((T)->seq_id) +#define ERTS_RBT_IS_LT(KX, KY) (KX < KY) +#define ERTS_RBT_IS_EQ(KX, KY) (KX == KY) +#define ERTS_RBT_WANT_DELETE +#define ERTS_RBT_WANT_LOOKUP_INSERT +#define ERTS_RBT_WANT_LOOKUP +#define ERTS_RBT_WANT_FOREACH +#define ERTS_RBT_WANT_FOREACH_DESTROY_YIELDING + +#include "erl_rbtree.h" + +struct erts_dist_seq_tree_foreach_iter_arg { + int (*func)(ErtsDistExternal *, void *, Sint); + void *arg; +}; + +static int +erts_dist_seq_tree_foreach_iter(DistSeqNode *seq, void *arg, Sint reds) +{ + struct erts_dist_seq_tree_foreach_iter_arg *state = arg; + return state->func(erts_get_dist_ext(&seq->hfrag), state->arg, reds); +} + +void +erts_dist_seq_tree_foreach(DistEntry *dep, int (*func)(ErtsDistExternal *, void *, Sint), void *arg) +{ + struct erts_dist_seq_tree_foreach_iter_arg state; + state.func = func; + state.arg = arg; + dist_seq_rbt_foreach(dep->sequences, erts_dist_seq_tree_foreach_iter, &state); +} + +static int dist_seq_cleanup(DistSeqNode *seq, void *unused, Sint reds) +{ + erts_free_dist_ext_copy(erts_get_dist_ext(&seq->hfrag)); + free_message_buffer(&seq->hfrag); + return 1; +} + +typedef struct { + DistSeqNode *root; + dist_seq_rbt_yield_state_t rbt_ystate; +} DistSeqNodeYieldState; + +int +erts_dist_seq_tree_foreach_delete_yielding(DistSeqNode **root, + void **vyspp, + Sint limit) +{ + DistSeqNodeYieldState ys = {*root, ERTS_RBT_YIELD_STAT_INITER}; + DistSeqNodeYieldState *ysp; + int res; + + ysp = (DistSeqNodeYieldState *) *vyspp; + if (!ysp) { + *root = NULL; + ysp = &ys; + } + res = dist_seq_rbt_foreach_destroy_yielding(&ysp->root, + dist_seq_cleanup, + NULL, + &ysp->rbt_ystate, + limit); + if (res > 0) { + if (ysp != &ys) + erts_free(ERTS_ALC_T_ML_YIELD_STATE, ysp); + *vyspp = NULL; + } + else { + + if (ysp == &ys) { + ysp = erts_alloc(ERTS_ALC_T_SEQ_YIELD_STATE, + sizeof(DistSeqNodeYieldState)); + sys_memcpy((void *) ysp, (void *) &ys, + sizeof(DistSeqNodeYieldState)); + } + + *vyspp = (void *) ysp; + } + + return res; +} /* ** Input from distribution port. @@ -1246,7 +1452,9 @@ int erts_net_message(Port *prt, byte *buf, ErlDrvSizeT len) { - ErtsDistExternal ede; + ErtsDistExternal ede, *edep = &ede; + ErtsDistExternalData ede_data; + ErlHeapFragment *ede_hfrag = NULL; Sint ctl_len; Eterm arg; Eterm from, to; @@ -1285,10 +1493,12 @@ int erts_net_message(Port *prt, } #ifdef ERTS_RAW_DIST_MSG_DBG - erts_fprintf(stderr, "<< "); + erts_fprintf(dbg_file, "RECV: "); bw(buf, len); #endif + ede.data = &ede_data; + res = erts_prepare_dist_ext(&ede, buf, len, bin, dep, conn_id, dep->cache); switch (res) { @@ -1296,51 +1506,164 @@ int erts_net_message(Port *prt, 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"); + erts_fprintf(dbg_file, "DIST MSG DEBUG: erts_prepare_dist_ext() failed:\n"); bw(buf, orig_len); #endif goto data_error; case ERTS_PREP_DIST_EXT_SUCCESS: - ctl_len = erts_decode_dist_ext_size(&ede); + ctl_len = erts_decode_dist_ext_size(&ede, 1); if (ctl_len < 0) { #ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, "DIST MSG DEBUG: erts_decode_dist_ext_size(CTL) failed:\n"); + erts_fprintf(dbg_file, "DIST MSG DEBUG: erts_decode_dist_ext_size(CTL) failed:\n"); bw(buf, orig_len); #endif PURIFY_MSG("data error"); goto data_error; } + + /* A non-fragmented message */ + if (!ede.data->seq_id) { + if (ctl_len > DIST_CTL_DEFAULT_SIZE) { + ctl = erts_alloc(ERTS_ALC_T_DCTRL_BUF, ctl_len * sizeof(Eterm)); + } + + erts_factory_tmp_init(&factory, ctl, ctl_len, ERTS_ALC_T_DCTRL_BUF); + break; + } else { + DistSeqNode *seq; + Uint sz = erts_dist_ext_size(&ede); + Uint used_sz = ctl_len * sizeof(Eterm); + + /* We calculate the size of the heap fragment to be allocated. + The used_size part has to be larger that the ctl data and the + DistSeqNode. */ + if (used_sz + (sizeof(ErlHeapFragment) - sizeof(Eterm)) < sizeof(DistSeqNode)) + used_sz = sizeof(DistSeqNode) - (sizeof(ErlHeapFragment) - sizeof(Eterm)); + + seq = (DistSeqNode *)new_message_buffer((sz + used_sz) / sizeof(Eterm)); + seq->hfrag.used_size = used_sz / sizeof(Eterm); + + seq->ctl_len = ctl_len; + seq->seq_id = ede.data->seq_id; + seq->cnt = ede.data->frag_id; + if (dist_seq_rbt_lookup_insert(&dep->sequences, seq) != NULL) { + free_message_buffer(&seq->hfrag); + goto data_error; + } + + erts_make_dist_ext_copy(&ede, erts_get_dist_ext(&seq->hfrag)); + + if (ede.data->frag_id > 1) { + seq->cnt--; + return 0; + } + } + + /* fall through, the first fragment in the sequence was the last fragment */ + case ERTS_PREP_DIST_EXT_FRAG_CONT: { + DistSeqNode *seq = dist_seq_rbt_lookup(dep->sequences, ede.data->seq_id); + + if (!seq) + goto data_error; + + /* If we did a fall-though we already did this */ + if (res == ERTS_PREP_DIST_EXT_FRAG_CONT) + erts_dist_ext_frag(&ede_data, erts_get_dist_ext(&seq->hfrag)); + + /* Verify that the fragments have arrived in the correct order */ + if (seq->cnt != ede.data->frag_id) + goto data_error; + + seq->cnt--; + + /* Check if this was the last fragment */ + if (ede.data->frag_id > 1) + return 0; + + /* Last fragment arrived, time to dispatch the signal */ + dist_seq_rbt_delete(&dep->sequences, seq); + ctl_len = seq->ctl_len; + + /* Now that we no longer need the DistSeqNode we re-use the heapfragment + to decode the ctl msg into. We don't need the ctl message to be in + the heapfragment, but we decode into the heapfragment speculatively + in case there is a trace token that we need. */ + erts_factory_heap_frag_init(&factory, &seq->hfrag); + edep = erts_get_dist_ext(&seq->hfrag); + ede_hfrag = &seq->hfrag; + + /* If the sequence consisted of more than 1 fragment we create one large + binary out of all of the fragments. This because erts_decode_ext + cannot handle a segmented buffer. + TODO: Move this copy to as late as possible, preferably in in the + erts_decode_dist_ext in the receiving process. + */ + if (edep->data->frag_id > 1) { + Uint sz = 0; + Binary *bin; + int i; + byte *ep; + + for (i = 0; i < edep->data->frag_id; i++) + sz += edep->data[i].ext_endp - edep->data[i].extp; + + bin = erts_bin_nrml_alloc(sz); + ep = (byte*)bin->orig_bytes; + + for (i = 0; i < edep->data->frag_id; i++) { + sys_memcpy(ep, edep->data[i].extp, edep->data[i].ext_endp - edep->data[i].extp); + ep += edep->data[i].ext_endp - edep->data[i].extp; + erts_bin_release(edep->data[i].binp); + edep->data[i].binp = NULL; + edep->data[i].extp = NULL; + edep->data[i].ext_endp = NULL; + } + + edep->data->frag_id = 1; + edep->data->extp = (byte*)bin->orig_bytes; + edep->data->ext_endp = ep; + edep->data->binp = bin; + } + break; + } default: ERTS_INTERNAL_ERROR("Unexpected result from erts_prepare_dist_ext()"); break; } - if (ctl_len > DIST_CTL_DEFAULT_SIZE) { - ctl = erts_alloc(ERTS_ALC_T_DCTRL_BUF, ctl_len * sizeof(Eterm)); - } - - erts_factory_tmp_init(&factory, ctl, ctl_len, ERTS_ALC_T_DCTRL_BUF); - arg = erts_decode_dist_ext(&factory, &ede); + arg = erts_decode_dist_ext(&factory, edep, 1); if (is_non_value(arg)) { #ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, "DIST MSG DEBUG: erts_decode_dist_ext(CTL) failed:\n"); + erts_fprintf(dbg_file, "DIST MSG DEBUG: erts_decode_dist_ext(CTL) failed:\n"); bw(buf, orig_len); #endif PURIFY_MSG("data error"); goto decode_error; } -#ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, "<< CTL: %.80T\n", arg); -#endif + /* Fill the unused part of the hfrag with a bignum header */ + if (ede_hfrag && ede_hfrag->mem + ede_hfrag->used_size > factory.hp) { + Uint slot = factory.hp - ede_hfrag->mem; + ede_hfrag->mem[slot] = make_pos_bignum_header(ede_hfrag->used_size - slot - 1); + } if (is_not_tuple(arg) || (tuple = tuple_val(arg), (tuple_arity = arityval(*tuple)) < 1) || is_not_small(tuple[1])) { +#ifdef ERTS_DIST_MSG_DBG + if (is_tuple(arg) && arityval(*tuple) > 1) + erts_fprintf(dbg_file, "RECV: CTL: %s: %.80T\n", + erts_dop_to_string(unsigned_val(tuple[1])), arg); +#endif goto invalid_message; } +#ifdef ERTS_DIST_MSG_DBG + erts_fprintf(dbg_file, "RECV: CTL: %s: %.80T\n", + erts_dop_to_string(unsigned_val(tuple[1])), arg); +#endif + token = NIL; switch (type = unsigned_val(tuple[1])) { @@ -1385,7 +1708,7 @@ int erts_net_message(Port *prt, 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); + code = erts_dsig_send_exit(&dsd, to, to, from, am_noproc); ASSERT(code == ERTS_DSIG_SEND_OK); } @@ -1475,7 +1798,7 @@ int erts_net_message(Port *prt, 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, + code = erts_dsig_send_m_exit(&dsd, pid, watcher, watched, ref, am_noproc); ASSERT(code == ERTS_DSIG_SEND_OK); } @@ -1549,7 +1872,7 @@ int erts_net_message(Port *prt, } #ifdef ERTS_DIST_MSG_DBG - dist_msg_dbg(&ede, "MSG", buf, orig_len); + dist_msg_dbg(edep, "MSG", buf, orig_len); #endif from = tuple[2]; @@ -1560,7 +1883,6 @@ int erts_net_message(Port *prt, rp = erts_whereis_process(NULL, 0, to, 0, 0); if (rp) { ErtsProcLocks locks = 0; - ErtsDistExternal *ede_copy; if (type == DOP_REG_SEND) { token = NIL; @@ -1568,13 +1890,14 @@ int erts_net_message(Port *prt, token = tuple[5]; } - ede_copy = erts_make_dist_ext_copy(&ede, &token); - - erts_queue_dist_message(rp, locks, ede_copy, token, from); + erts_queue_dist_message(rp, locks, edep, ede_hfrag, token, from); if (locks) erts_proc_unlock(rp, locks); - } + } else if (ede_hfrag) { + erts_free_dist_ext_copy(erts_get_dist_ext(ede_hfrag)); + free_message_buffer(ede_hfrag); + } break; case DOP_SEND_SENDER_TT: { @@ -1606,7 +1929,7 @@ int erts_net_message(Port *prt, : tuple[2] == am_Empty); #ifdef ERTS_DIST_MSG_DBG - dist_msg_dbg(&ede, "MSG", buf, orig_len); + dist_msg_dbg(edep, "MSG", buf, orig_len); #endif to = tuple[3]; if (is_not_pid(to)) { @@ -1615,20 +1938,19 @@ int erts_net_message(Port *prt, rp = erts_proc_lookup(to); if (rp) { ErtsProcLocks locks = 0; - ErtsDistExternal *ede_copy; - - ede_copy = erts_make_dist_ext_copy(&ede, &token); - erts_queue_dist_message(rp, locks, ede_copy, token, am_Empty); + erts_queue_dist_message(rp, locks, edep, ede_hfrag, token, am_Empty); if (locks) erts_proc_unlock(rp, locks); - } + } else if (ede_hfrag) { + erts_free_dist_ext_copy(erts_get_dist_ext(ede_hfrag)); + free_message_buffer(ede_hfrag); + } break; } case DOP_PAYLOAD_MONITOR_P_EXIT: case DOP_MONITOR_P_EXIT: { - ErtsDistExternal *ede_copy = NULL; /* We are monitoring a process on the remote node which dies, we get {DOP_MONITOR_P_EXIT, Remote pid or name, Local pid, ref, reason} */ @@ -1663,19 +1985,18 @@ int erts_net_message(Port *prt, goto invalid_message; } + if (!erts_proc_lookup(watcher)) break; /* Process not alive */ + if (reason == THE_NON_VALUE) { #ifdef ERTS_DIST_MSG_DBG - dist_msg_dbg(&ede, "MSG", buf, orig_len); + dist_msg_dbg(edep, "MSG", buf, orig_len); #endif - if (!erts_proc_lookup(watcher)) break; /* Process not alive */ - - ede_copy = erts_make_dist_ext_copy(&ede, &token); } - erts_proc_sig_send_dist_monitor_down(dep, ref, watched, - watcher, ede_copy, reason); + erts_proc_sig_send_dist_monitor_down( + dep, ref, watched, watcher, edep, ede_hfrag, reason); break; } @@ -1683,7 +2004,6 @@ int erts_net_message(Port *prt, case DOP_PAYLOAD_EXIT_TT: case DOP_EXIT_TT: case DOP_EXIT: { - ErtsDistExternal *ede_copy = NULL; /* 'from', which 'to' is linked to, died */ from = tuple[2]; @@ -1720,19 +2040,16 @@ int erts_net_message(Port *prt, goto invalid_message; } - if (reason == THE_NON_VALUE) { + if (!erts_proc_lookup(to)) break; /* Process not alive */ + if (reason == THE_NON_VALUE) { #ifdef ERTS_DIST_MSG_DBG - dist_msg_dbg(&ede, "MSG", buf, orig_len); + dist_msg_dbg(edep, "MSG", buf, orig_len); #endif - - if (!erts_proc_lookup(to)) break; /* Process not alive */ - - ede_copy = erts_make_dist_ext_copy(&ede, &token); } erts_proc_sig_send_dist_link_exit(dep, - from, to, ede_copy, + from, to, edep, ede_hfrag, reason, token); break; } @@ -1740,7 +2057,6 @@ int erts_net_message(Port *prt, case DOP_PAYLOAD_EXIT2: case DOP_EXIT2_TT: case DOP_EXIT2: { - ErtsDistExternal *ede_copy = NULL; /* 'from' is send an exit signal to 'to' */ from = tuple[2]; @@ -1777,18 +2093,15 @@ int erts_net_message(Port *prt, goto invalid_message; } - if (reason == THE_NON_VALUE) { + if (!erts_proc_lookup(to)) break; /* Process not alive */ + if (reason == THE_NON_VALUE) { #ifdef ERTS_DIST_MSG_DBG - dist_msg_dbg(&ede, "MSG", buf, orig_len); + dist_msg_dbg(edep, "MSG", buf, orig_len); #endif - - if (!erts_proc_lookup(to)) break; /* Process not alive */ - - ede_copy = erts_make_dist_ext_copy(&ede, &token); } - erts_proc_sig_send_dist_exit(dep, from, to, ede_copy, reason, token); + erts_proc_sig_send_dist_exit(dep, from, to, edep, ede_hfrag, reason, token); break; } case DOP_GROUP_LEADER: @@ -1804,13 +2117,15 @@ int erts_net_message(Port *prt, (void) erts_proc_sig_send_group_leader(NULL, to, from, NIL); break; - default: + default: goto invalid_message; } - erts_factory_close(&factory); - if (ctl != ctl_default) { - erts_free(ERTS_ALC_T_DCTRL_BUF, (void *) ctl); + if (ede_hfrag == NULL) { + erts_factory_close(&factory); + if (ctl != ctl_default) { + erts_free(ERTS_ALC_T_DCTRL_BUF, (void *) ctl); + } } UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); ERTS_CHK_NO_PROC_LOCKS; @@ -1823,9 +2138,14 @@ int erts_net_message(Port *prt, } decode_error: PURIFY_MSG("data error"); - erts_factory_close(&factory); - if (ctl != ctl_default) { - erts_free(ERTS_ALC_T_DCTRL_BUF, (void *) ctl); + if (ede_hfrag == NULL) { + erts_factory_close(&factory); + if (ctl != ctl_default) { + erts_free(ERTS_ALC_T_DCTRL_BUF, (void *) ctl); + } + } else { + erts_free_dist_ext_copy(erts_get_dist_ext(ede_hfrag)); + free_message_buffer(ede_hfrag); } data_error: UnUseTmpHeapNoproc(DIST_CTL_DEFAULT_SIZE); @@ -1834,18 +2154,17 @@ data_error: return -1; } -static int dsig_send_exit(ErtsDSigData* dsdp, Eterm ctl, Eterm msg, int force_busy) +static int dsig_send_exit(ErtsDSigData* dsdp, Eterm ctl, Eterm msg, Eterm from, int force_busy) { struct erts_dsig_send_context ctx; int ret; ctx.ctl = ctl; ctx.msg = msg; + ctx.from = from; ctx.force_busy = force_busy; ctx.force_encode = 1; ctx.phase = ERTS_DSIG_SEND_PHASE_INIT; -#ifdef DEBUG ctx.reds = 1; /* provoke assert below (no reduction count with force_encode) */ -#endif ret = erts_dsig_send(dsdp, &ctx); ASSERT(ret != ERTS_DSIG_SEND_CONTINUE); return ret; @@ -1857,12 +2176,11 @@ static int dsig_send_ctl(ErtsDSigData* dsdp, Eterm ctl, int force_busy) int ret; ctx.ctl = ctl; ctx.msg = THE_NON_VALUE; + ctx.from = THE_NON_VALUE; ctx.force_busy = force_busy; ctx.force_encode = 1; ctx.phase = ERTS_DSIG_SEND_PHASE_INIT; -#ifdef DEBUG ctx.reds = 1; /* provoke assert below (no reduction count without msg) */ -#endif ret = erts_dsig_send(dsdp, &ctx); ASSERT(ret != ERTS_DSIG_SEND_CONTINUE); return ret; @@ -1926,9 +2244,11 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) } #ifdef ERTS_DIST_MSG_DBG - erts_fprintf(stderr, ">> CTL: %.80T\n", ctx->ctl); + erts_fprintf(dbg_file, "SEND: CTL: %s: %.80T\n", + erts_dop_to_string(unsigned_val(tuple_val(ctx->ctl)[1])), + ctx->ctl); if (is_value(ctx->msg)) - erts_fprintf(stderr, " MSG: %.80T\n", ctx->msg); + erts_fprintf(dbg_file, " MSG: %.160T\n", ctx->msg); #endif ctx->data_size = ctx->max_finalize_prepend; @@ -1958,16 +2278,32 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) case ERTS_DSIG_SEND_PHASE_ALLOC: erts_finalize_atom_cache_map(ctx->acmp, ctx->flags); - ctx->dhdr_ext_size = erts_encode_ext_dist_header_size(ctx->acmp); - ctx->data_size += ctx->dhdr_ext_size; + if (ctx->flags & DFLAG_FRAGMENTS && is_value(ctx->msg) && is_not_immed(ctx->msg)) { + /* Calculate the max number of fragments that are needed */ + ASSERT(is_pid(ctx->from) && + "from has to be a pid because it is used as sequence id"); + ctx->fragments = ctx->data_size / ERTS_DIST_FRAGMENT_SIZE + 1; + } else + ctx->fragments = 1; - ctx->obuf = alloc_dist_obuf(ctx->data_size); - ctx->obuf->ext_endp = &ctx->obuf->data[0] + ctx->max_finalize_prepend + ctx->dhdr_ext_size; + ctx->dhdr_ext_size = erts_encode_ext_dist_header_size(ctx->acmp, ctx->fragments); + + ctx->obuf = alloc_dist_obuf( + ctx->dhdr_ext_size + ctx->data_size + + (ctx->fragments-1) * ERTS_DIST_FRAGMENT_HEADER_SIZE, + ctx->fragments); + ctx->obuf->ext_start = &ctx->obuf->extp[0]; + ctx->obuf->ext_endp = &ctx->obuf->extp[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); + ctx->obuf->extp = erts_encode_ext_dist_header_setup( + ctx->obuf->ext_endp, ctx->acmp, ctx->fragments, ctx->from); /* Encode control message */ erts_encode_dist_ext(ctx->ctl, &ctx->obuf->ext_endp, ctx->flags, ctx->acmp, NULL, NULL); + + ctx->obuf->hdrp = NULL; + ctx->obuf->hdr_endp = NULL; + if (is_non_value(ctx->msg)) { ctx->obuf->msg_start = NULL; ctx->phase = ERTS_DSIG_SEND_PHASE_FIN; @@ -1988,33 +2324,91 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) goto done; } } else { - erts_encode_dist_ext(ctx->msg, &ctx->obuf->ext_endp, ctx->flags, ctx->acmp, NULL, NULL); + erts_encode_dist_ext(ctx->msg, &ctx->obuf->ext_endp, ctx->flags, + ctx->acmp, NULL, NULL); } - ctx->phase = ERTS_DSIG_SEND_PHASE_FIN; + ctx->phase = ERTS_DSIG_SEND_PHASE_FIN; case ERTS_DSIG_SEND_PHASE_FIN: { - DistEntry *dep = dsdp->dep; - int suspended = 0; - int resume = 0; ASSERT(ctx->obuf->extp < ctx->obuf->ext_endp); - ASSERT(&ctx->obuf->data[0] <= ctx->obuf->extp - ctx->max_finalize_prepend); - ASSERT(ctx->obuf->ext_endp <= &ctx->obuf->data[0] + ctx->data_size); + ASSERT(((byte*)&ctx->obuf->bin->orig_bytes[0]+obuf_list_size) <= ctx->obuf->extp - ctx->max_finalize_prepend); + ASSERT(ctx->obuf->ext_endp <= ((byte*)ctx->obuf->bin->orig_bytes+obuf_list_size) + ctx->data_size + ctx->dhdr_ext_size); ctx->data_size = ctx->obuf->ext_endp - ctx->obuf->extp; - if (ctx->data_size > (Uint) INT_MAX) { - free_dist_obuf(ctx->obuf); - ctx->obuf = NULL; - retval = ERTS_DSIG_SEND_TOO_LRG; - goto done; - } ctx->obuf->hopefull_flags = ctx->u.ec.hopefull_flags; - /* + + if (ctx->fragments > 1) { + int fin_fragments; + int i; + byte *msg = ctx->obuf->msg_start, + *msg_end = ctx->obuf->ext_endp, + *hdrp = msg_end; + + ASSERT((ctx->obuf->hopefull_flags & ctx->flags) == ctx->obuf->hopefull_flags); + ASSERT(get_int64(ctx->obuf->extp + 1 + 1 + 8) == ctx->fragments); + + /* Now that encoding is done we know how large the term will + be so we adjust the number of fragments to send. Note that + this can mean that only 1 fragment is sent. */ + fin_fragments = (ctx->obuf->ext_endp - ctx->obuf->msg_start + ERTS_DIST_FRAGMENT_SIZE-1) / + ERTS_DIST_FRAGMENT_SIZE - 1; + + /* Update the frag_id in the DIST_FRAG_HEADER */ + put_int64(fin_fragments+1, ctx->obuf->extp + 1 + 1 + 8); + + if (fin_fragments > 0) + msg += ERTS_DIST_FRAGMENT_SIZE; + else + msg = msg_end; + ctx->obuf->next = &ctx->obuf[1]; + ctx->obuf->ext_endp = msg; + + /* Loop through all fragments, updating the output buffers + to be correct and also writing the DIST_FRAG_CONT header. */ + for (i = 1; i < fin_fragments + 1; i++) { + ctx->obuf[i].hopefull_flags = 0; + ctx->obuf[i].extp = msg; + ctx->obuf[i].ext_start = msg; + if (msg + ERTS_DIST_FRAGMENT_SIZE > msg_end) + ctx->obuf[i].ext_endp = msg_end; + else { + msg += ERTS_DIST_FRAGMENT_SIZE; + ctx->obuf[i].ext_endp = msg; + } + ASSERT(ctx->obuf[i].ext_endp > ctx->obuf[i].extp); + ctx->obuf[i].hdrp = erts_encode_ext_dist_header_fragment( + &hdrp, fin_fragments - i + 1, ctx->from); + ctx->obuf[i].hdr_endp = hdrp; + ctx->obuf[i].next = &ctx->obuf[i+1]; + } + /* If the initial fragment calculation was incorrect we free the + remaining output buffers. */ + for (; i < ctx->fragments; i++) { + free_dist_obuf(&ctx->obuf[i]); + } + if (!ctx->force_encode && !ctx->force_busy) + ctx->reds -= ctx->fragments; + ctx->fragments = fin_fragments + 1; + } + + ctx->phase = ERTS_DSIG_SEND_PHASE_SEND; + + if (ctx->reds <= 0) { + retval = ERTS_DSIG_SEND_CONTINUE; + goto done; + } + } + case ERTS_DSIG_SEND_PHASE_SEND: { + /* * Signal encoded; now verify that the connection still exists, * and if so enqueue the signal and schedule it for send. */ - ctx->obuf->next = NULL; + DistEntry *dep = dsdp->dep; + int suspended = 0; + int resume = 0; + int i; erts_de_rlock(dep); cid = dep->cid; if (dep->state == ERTS_DE_STATE_EXITING @@ -2022,14 +2416,38 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) || dep->connection_id != dsdp->connection_id) { /* Not the same connection as when we started; drop message... */ erts_de_runlock(dep); - free_dist_obuf(ctx->obuf); + for (i = 0; i < ctx->fragments; i++) + free_dist_obuf(&ctx->obuf[i]); + ctx->fragments = 0; } else { - Sint qsize; + Sint qsize = erts_atomic_read_nob(&dep->qsize); erts_aint32_t qflgs; ErtsProcList *plp = NULL; Eterm notify_proc = NIL; - Sint obsz = size_obuf(ctx->obuf); + Sint obsz; + int fragments; + + /* Calculate how many fragments to send. This depends on + the available space in the distr queue and the amount + of remaining reductions. */ + for (fragments = 0, obsz = 0; + fragments < ctx->fragments && + ((ctx->reds > 0 && (qsize + obsz) < erts_dist_buf_busy_limit) || + ctx->force_encode || ctx->force_busy); + fragments++) { +#ifdef DEBUG + int reds = 100; +#else + int reds = 10; +#endif + if (!ctx->force_encode && !ctx->force_busy) + ctx->reds -= reds; + obsz += size_obuf(&ctx->obuf[fragments]); + } + + ASSERT(fragments == ctx->fragments || + (!ctx->force_encode && !ctx->force_busy)); erts_mtx_lock(&dep->qlock); qsize = erts_atomic_add_read_nob(&dep->qsize, (erts_aint_t) obsz); @@ -2059,12 +2477,25 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) erts_mtx_lock(&dep->qlock); } - /* Enqueue obuf on dist entry */ - if (dep->out_queue.last) - dep->out_queue.last->next = ctx->obuf; - else - dep->out_queue.first = ctx->obuf; - dep->out_queue.last = ctx->obuf; + if (fragments > 1) { + if (!ctx->obuf->hdrp) { + ASSERT(get_int64(ctx->obuf->extp + 10) == ctx->fragments); + } else { + ASSERT(get_int64(ctx->obuf->hdrp + 10) == ctx->fragments); + } + } + + if (fragments) { + ctx->obuf[fragments-1].next = NULL; + if (dep->out_queue.last) + dep->out_queue.last->next = ctx->obuf; + else + dep->out_queue.first = ctx->obuf; + dep->out_queue.last = &ctx->obuf[fragments-1]; + + ctx->fragments -= fragments; + ctx->obuf = &ctx->obuf[fragments]; + } if (!ctx->force_busy) { qflgs = erts_atomic32_read_nob(&dep->qflgs); @@ -2113,6 +2544,13 @@ erts_dsig_send(ErtsDSigData *dsdp, struct erts_dsig_send_context* ctx) * erroneously scheduled when it shouldn't be. */ } + /* More fragments left to be sent, yield and re-schedule */ + if (ctx->fragments) { + retval = ERTS_DSIG_SEND_CONTINUE; + if (!resume && erts_system_monitor_flags.busy_dist_port) + monitor_generic(ctx->c_p, am_busy_dist_port, cid); + goto done; + } } ctx->obuf = NULL; @@ -2197,9 +2635,9 @@ static Uint dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf) { int fpe_was_unmasked; - ErlDrvSizeT size; - SysIOVec iov[2]; - ErlDrvBinary* bv[2]; + ErlDrvSizeT size = 0; + SysIOVec iov[3]; + ErlDrvBinary* bv[3]; ErlIOVec eiov; ERTS_CHK_NO_PROC_LOCKS; @@ -2214,12 +2652,31 @@ dist_port_commandv(Port *prt, ErtsDistOutputBuf *obuf) eiov.vsize = 1; } else { - size = obuf->ext_endp - obuf->extp; + int i = 1; eiov.vsize = 2; - iov[1].iov_base = obuf->extp; - iov[1].iov_len = size; - bv[1] = Binary2ErlDrvBinary(ErtsDistOutputBuf2Binary(obuf)); + if (obuf->hdrp) { + eiov.vsize = 3; + iov[i].iov_base = obuf->hdrp; + iov[i].iov_len = obuf->hdr_endp - obuf->hdrp; + size += iov[i].iov_len; + bv[i] = Binary2ErlDrvBinary(ErtsDistOutputBuf2Binary(obuf)); +#ifdef ERTS_RAW_DIST_MSG_DBG + erts_fprintf(dbg_file, "SEND: "); + bw(iov[i].iov_base, iov[i].iov_len); +#endif + i++; + + } + + iov[i].iov_base = obuf->extp; + iov[i].iov_len = obuf->ext_endp - obuf->extp; +#ifdef ERTS_RAW_DIST_MSG_DBG + erts_fprintf(dbg_file, "SEND: "); + bw(iov[i].iov_base, iov[i].iov_len); +#endif + size += iov[i].iov_len; + bv[i] = Binary2ErlDrvBinary(ErtsDistOutputBuf2Binary(obuf)); } eiov.size = size; @@ -2326,6 +2783,23 @@ erts_dist_command(Port *prt, int initial_reds) dep->finalized_out_queue.first = NULL; dep->finalized_out_queue.last = NULL; +#ifdef DEBUG + { + Uint sz = 0; + ErtsDistOutputBuf *curr = oq.first; + while (curr) { + sz += size_obuf(curr); + curr = curr->next; + } + curr = foq.first; + while (curr) { + sz += size_obuf(curr); + curr = curr->next; + } + ASSERT(sz <= erts_atomic_read_nob(&dep->qsize)); + } +#endif + sched_flags = erts_atomic32_read_nob(&prt->sched.flags); if (reds < 0) @@ -2339,10 +2813,6 @@ erts_dist_command(Port *prt, int initial_reds) 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); -#endif reds -= ERTS_PORT_REDS_DIST_CMD_DATA(size); fob = foq.first; obufsize += size_obuf(fob); @@ -2411,15 +2881,11 @@ erts_dist_command(Port *prt, int initial_reds) preempt = 1; break; } - ASSERT(&oq.first->data[0] <= oq.first->extp - && oq.first->extp <= oq.first->ext_endp); + ASSERT(oq.first->bin->orig_bytes <= (char*)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); -#endif reds -= ERTS_PORT_REDS_DIST_CMD_DATA(size); fob = oq.first; obufsize += size_obuf(fob); @@ -2556,100 +3022,6 @@ erts_dist_command(Port *prt, int initial_reds) 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) { @@ -2998,6 +3370,10 @@ static void kill_connection(DistEntry *dep) void erts_kill_dist_connection(DistEntry *dep, Uint32 conn_id) { +#ifdef ERTS_DIST_MSG_DBG + erts_fprintf(dbg_file, "INTR: kill dist conn to %T:%u\n", + dep->sysname, conn_id); +#endif erts_de_rwlock(dep); if (conn_id == dep->connection_id && dep->state == ERTS_DE_STATE_CONNECTED) { @@ -3659,8 +4035,9 @@ Sint erts_abort_connection_rwunlock(DistEntry* dep) erts_set_dist_entry_not_connected(dep); erts_de_rwunlock(dep); - schedule_con_monitor_link_cleanup(mld, THE_NON_VALUE, - THE_NON_VALUE, THE_NON_VALUE); + schedule_con_monitor_link_seq_cleanup( + mld, NULL, THE_NON_VALUE, + THE_NON_VALUE, THE_NON_VALUE); if (resume_procs) { int resumed = erts_resume_processes(resume_procs); -- cgit v1.2.3