1 /* Copyright (c) 2007-2022. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include "smpi_request.hpp"
10 #include "simgrid/Exception.hpp"
11 #include "simgrid/s4u/ConditionVariable.hpp"
12 #include "simgrid/s4u/Exec.hpp"
13 #include "simgrid/s4u/Mutex.hpp"
14 #include "smpi_comm.hpp"
15 #include "smpi_datatype.hpp"
16 #include "smpi_host.hpp"
17 #include "smpi_op.hpp"
18 #include "src/kernel/activity/CommImpl.hpp"
19 #include "src/kernel/actor/ActorImpl.hpp"
20 #include "src/kernel/actor/SimcallObserver.hpp"
21 #include "src/mc/mc_replay.hpp"
22 #include "src/smpi/include/smpi_actor.hpp"
27 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
29 static simgrid::config::Flag<double> smpi_iprobe_sleep(
30 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
31 static simgrid::config::Flag<double> smpi_test_sleep(
32 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
34 std::vector<s_smpi_factor_t> smpi_ois_values;
36 extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
41 Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
42 unsigned flags, MPI_Op op)
43 : buf_(const_cast<void*>(buf))
46 , size_(datatype->size() * count)
56 if(op != MPI_REPLACE && op != MPI_OP_NULL)
60 detached_sender_ = nullptr;
62 // get src_host if it's available (src is valid)
63 auto src_process = simgrid::s4u::Actor::by_pid(src);
65 src_host_ = src_process->get_host();
67 unmatched_types_ = false;
70 if (flags & MPI_REQ_PERSISTENT)
83 void Request::unref(MPI_Request* request)
85 xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
87 (*request)->refcount_--;
88 if ((*request)->refcount_ < 0) {
89 (*request)->print_request("wrong refcount");
90 xbt_die("Whoops, wrong refcount");
92 if ((*request)->refcount_ == 0) {
93 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
94 ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
96 Comm::unref((*request)->comm_);
97 Datatype::unref((*request)->type_);
99 if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
100 Op::unref(&(*request)->op_);
102 (*request)->print_request("Destroying");
103 F2C::free_f((*request)->f2c_id());
105 *request = MPI_REQUEST_NULL;
107 (*request)->print_request("Decrementing");
111 bool Request::match_types(MPI_Datatype stype, MPI_Datatype rtype){
113 if ((stype == rtype) ||
114 //byte and packed always match with anything
115 (stype == MPI_PACKED || rtype == MPI_PACKED || stype == MPI_BYTE || rtype == MPI_BYTE) ||
116 //complex datatypes - we don't properly match these yet, as it would mean checking each subtype recursively.
117 (stype->flags() & DT_FLAG_DERIVED || rtype->flags() & DT_FLAG_DERIVED) ||
118 //duplicated datatypes, check if underlying is ok
119 (stype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype->duplicated_datatype(), rtype)) ||
120 (rtype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype, rtype->duplicated_datatype())))
123 XBT_WARN("Mismatched datatypes : sending %s and receiving %s", stype->name().c_str(), rtype->name().c_str());
128 bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
130 xbt_assert(sender, "Cannot match against null sender");
131 xbt_assert(receiver, "Cannot match against null receiver");
132 XBT_DEBUG("Trying to match %s of sender src %ld against %ld, tag %d against %d, id %d against %d",
133 (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
134 sender->comm_->id(), receiver->comm_->id());
136 if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED ||
137 receiver->comm_->id() == sender->comm_->id()) &&
138 ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) ||
139 receiver->src_ == sender->src_) &&
140 ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) {
141 // we match, we can transfer some values
142 if (receiver->src_ == MPI_ANY_SOURCE) {
143 receiver->real_src_ = sender->src_;
144 receiver->src_host_ = sender->src_host_;
146 if (receiver->tag_ == MPI_ANY_TAG)
147 receiver->real_tag_ = sender->tag_;
148 if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){
149 if (receiver->real_size_ < sender->real_size_){
150 XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
151 receiver->truncated_ = true;
154 //0-sized datatypes/counts should not interfere and match
155 if ( sender->real_size_ != 0 && receiver->real_size_ != 0 &&
156 !match_types(sender->type_, receiver->type_))
157 receiver->unmatched_types_ = true;
158 if (sender->detached_)
159 receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
161 req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
162 XBT_DEBUG("match succeeded");
168 void Request::init_buffer(int count){
169 // FIXME Handle the case of a partial shared malloc.
170 // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
171 if (not smpi_process()->replaying() &&
172 ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED))) {
173 // This part handles the problem of non-contiguous memory
178 buf_ = xbt_malloc(count*type_->size());
179 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
180 type_->serialize(old_buf_, buf_, count);
186 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
188 auto ref = static_cast<MPI_Request>(a);
189 auto req = static_cast<MPI_Request>(b);
190 bool match = match_common(req, req, ref);
191 if (match && (ref->comm_ != MPI_COMM_UNINITIALIZED) && !ref->comm_->is_smp_comm()){
192 if (ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), ref->comm_->group()->rank(req->dst_), req->tag_) == req->message_id_ ){
193 if (((ref->flags_ & MPI_REQ_PROBE) == 0 ) && ((req->flags_ & MPI_REQ_PROBE) == 0)){
194 XBT_DEBUG("increasing count in comm %p, which was %u from pid %ld, to pid %ld with tag %d", ref->comm_, ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), ref->comm_->group()->rank(req->dst_), req->tag_), req->src_, req->dst_, req->tag_);
195 ref->comm_->increment_received_messages_count(ref->comm_->group()->rank(req->src_), ref->comm_->group()->rank(req->dst_), req->tag_);
196 if (ref->real_size_ > req->real_size_){
197 ref->real_size_=req->real_size_;
202 req->flags_ &= ~MPI_REQ_MATCHED;
203 ref->detached_sender_=nullptr;
204 XBT_DEBUG("Refusing to match message, as its ID is not the one I expect. in comm %p, %u != %u, from pid %ld to pid %ld, with tag %d",ref->comm_, ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_), ref->comm_->group()->rank(req->dst_), req->tag_), req->message_id_ , req->src_, req->dst_, req->tag_);
210 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
212 auto ref = static_cast<MPI_Request>(a);
213 auto req = static_cast<MPI_Request>(b);
214 return match_common(req, ref, req);
217 void Request::print_request(const char* message) const
219 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
220 size_, src_, dst_, tag_, flags_);
223 /* factories, to hide the internal flags from the caller */
224 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
226 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
227 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
228 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
231 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
233 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
234 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
235 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
238 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
240 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
241 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
242 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
245 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
247 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
248 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
249 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
252 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
257 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
258 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
259 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
261 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
262 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
263 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
270 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
272 aid_t source = MPI_PROC_NULL;
273 if (src == MPI_ANY_SOURCE)
274 source = MPI_ANY_SOURCE;
275 else if (src != MPI_PROC_NULL)
276 source = comm->group()->actor(src);
277 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
279 simgrid::s4u::this_actor::get_pid(), tag, comm,
280 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
283 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
286 aid_t source = MPI_PROC_NULL;
287 if (src == MPI_ANY_SOURCE)
288 source = MPI_ANY_SOURCE;
289 else if (src != MPI_PROC_NULL)
290 source = comm->group()->actor(src);
293 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
294 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
295 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
298 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
299 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
300 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
305 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
307 aid_t source = MPI_PROC_NULL;
308 if (src == MPI_ANY_SOURCE)
309 source = MPI_ANY_SOURCE;
310 else if (src != MPI_PROC_NULL)
311 source = comm->group()->actor(src);
312 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
313 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
314 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
317 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
319 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
320 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
321 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
322 if(dst != MPI_PROC_NULL)
327 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
329 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
330 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
331 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
332 if(dst != MPI_PROC_NULL)
337 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
339 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
340 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
341 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
342 if(dst != MPI_PROC_NULL)
347 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
349 aid_t source = MPI_PROC_NULL;
350 if (src == MPI_ANY_SOURCE)
351 source = MPI_ANY_SOURCE;
352 else if (src != MPI_PROC_NULL)
353 source = comm->group()->actor(src);
354 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
355 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
356 if(src != MPI_PROC_NULL)
361 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
363 MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
364 int retval = wait(&request,status);
369 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
371 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
372 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
373 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
375 if(dst != MPI_PROC_NULL)
377 wait(&request, MPI_STATUS_IGNORE);
381 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
383 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
384 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
385 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
386 if(dst != MPI_PROC_NULL)
388 wait(&request, MPI_STATUS_IGNORE);
392 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
394 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
395 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
396 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
398 if(dst != MPI_PROC_NULL)
400 wait(&request,MPI_STATUS_IGNORE);
404 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
405 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
406 MPI_Comm comm, MPI_Status * status)
408 aid_t source = MPI_PROC_NULL;
409 if (src == MPI_ANY_SOURCE)
410 source = MPI_ANY_SOURCE;
411 else if (src != MPI_PROC_NULL)
412 source = comm->group()->actor(src);
413 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
415 std::array<MPI_Request, 2> requests;
416 std::array<MPI_Status, 2> stats;
417 aid_t myid = simgrid::s4u::this_actor::get_pid();
418 if ((destination == myid) && (source == myid)) {
419 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
420 if (status != MPI_STATUS_IGNORE) {
421 status->MPI_SOURCE = source;
422 status->MPI_TAG = recvtag;
423 status->MPI_ERROR = MPI_SUCCESS;
424 status->count = sendcount * sendtype->size();
428 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
429 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
430 startall(2, requests.data());
431 waitall(2, requests.data(), stats.data());
434 if(status != MPI_STATUS_IGNORE) {
435 // Copy receive status
440 void Request::start()
442 s4u::Mailbox* mailbox;
444 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
445 //reinitialize temporary buffer for persistent requests
446 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
448 init_buffer(real_size_/type_->size());
450 flags_ &= ~MPI_REQ_PREPARED;
451 flags_ &= ~MPI_REQ_FINISHED;
454 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
456 if ((flags_ & MPI_REQ_RECV) != 0) {
457 this->print_request("New recv");
459 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
461 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
462 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
465 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
466 flags_ |= MPI_REQ_PROBE;
468 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
469 mailbox = process->mailbox();
470 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
471 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
472 //begin with the more appropriate one : the small one.
473 mailbox = process->mailbox_small();
474 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
475 mailbox->get_cname());
476 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
478 if (action == nullptr) {
479 mailbox = process->mailbox();
480 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
481 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
482 if (action == nullptr) {
483 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
484 mailbox = process->mailbox_small();
487 XBT_DEBUG("yes there was something for us in the small mailbox");
490 mailbox = process->mailbox_small();
491 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
492 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
494 if (action == nullptr) {
495 XBT_DEBUG("No, nothing in the permanent receive mailbox");
496 mailbox = process->mailbox();
498 XBT_DEBUG("yes there was something for us in the small mailbox");
502 flags_ &= ~MPI_REQ_PROBE;
504 action_ = simcall_comm_irecv(
505 process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
506 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
507 XBT_DEBUG("recv simcall posted");
509 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
511 } else { /* the RECV flag was not set, so this is a send */
512 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
513 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
514 if (TRACE_smpi_view_internals())
515 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
516 this->print_request("New send");
518 message_id_=comm_->get_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
519 comm_->increment_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
522 if ((flags_ & MPI_REQ_SSEND) == 0 &&
523 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
524 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
525 void *oldbuf = nullptr;
527 XBT_DEBUG("Send request %p is detached", this);
529 if (not(type_->flags() & DT_FLAG_DERIVED)) {
531 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
532 if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
533 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
535 //we need this temporary buffer even for bsend, as it will be released in the copy callback and we don't have a way to differentiate it
536 //so actually ... don't use manually attached buffer space.
537 buf = xbt_malloc(size_);
538 memcpy(buf,oldbuf,size_);
539 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
544 //if we are giving back the control to the user without waiting for completion, we have to inject timings
545 double sleeptime = 0.0;
546 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
547 // isend and send timings may be different
549 ((flags_ & MPI_REQ_ISEND) != 0)
550 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(
551 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(), simgrid::s4u::Actor::by_pid(dst_)->get_host())
552 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(
553 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(),
554 simgrid::s4u::Actor::by_pid(dst_)->get_host());
558 simgrid::s4u::this_actor::sleep_for(sleeptime);
559 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
562 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
564 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
567 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
568 mailbox = process->mailbox();
569 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
570 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
571 flags_ |= MPI_REQ_PROBE;
573 mailbox = process->mailbox();
574 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
575 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
576 if (action == nullptr) {
577 if ((flags_ & MPI_REQ_SSEND) == 0) {
578 mailbox = process->mailbox_small();
579 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
580 mailbox->get_cname());
582 mailbox = process->mailbox_small();
583 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
584 mailbox->get_cname());
585 action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
586 if (action == nullptr) {
587 XBT_DEBUG("No, we are first, send to large mailbox");
588 mailbox = process->mailbox();
592 XBT_DEBUG("Yes there was something for us in the large mailbox");
595 flags_ &= ~MPI_REQ_PROBE;
597 mailbox = process->mailbox();
598 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
601 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
602 action_ = simcall_comm_isend(
603 simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_,
605 &xbt_free_f, // how to free the userdata if a detached send fails
606 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
607 // detach if msg size < eager/rdv switch limit
609 XBT_DEBUG("send simcall posted");
611 /* FIXME: detached sends are not traceable (action_ == nullptr) */
612 if (action_ != nullptr) {
613 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
614 smpi_process()->get_tracing_category());
617 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
622 void Request::startall(int count, MPI_Request * requests)
624 if(requests== nullptr)
627 for(int i = 0; i < count; i++) {
628 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
629 requests[i]->start();
633 void Request::cancel()
635 this->flags_ |= MPI_REQ_CANCELLED;
636 if (this->action_ != nullptr)
637 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
640 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
641 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
642 // to avoid deadlocks if used as a break condition, such as
643 // while (MPI_Test(request, flag, status) && flag) dostuff...
644 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
645 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
646 xbt_assert(*request != MPI_REQUEST_NULL);
648 static int nsleeps = 1;
649 int ret = MPI_SUCCESS;
651 if(smpi_test_sleep > 0)
652 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
654 Status::empty(status);
657 if ((*request)->flags_ & MPI_REQ_NBC){
658 *flag = finish_nbc_requests(request, 1);
661 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
662 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
664 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
665 kernel::actor::ActivityTestSimcall observer{issuer, (*request)->action_.get()};
666 *flag = kernel::actor::simcall_blocking([&observer] { observer.get_activity()->test(observer.get_issuer()); },
668 } catch (const Exception&) {
673 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
676 finish_wait(request, status); // may invalidate *request
677 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
678 MPI_Status tmp_status;
679 MPI_Status* mystatus;
680 if (status == MPI_STATUS_IGNORE) {
681 mystatus = &tmp_status;
682 Status::empty(mystatus);
686 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
688 nsleeps=1;//reset the number of sleeps we will do next time
689 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
690 *request = MPI_REQUEST_NULL;
691 } else if (smpi_cfg_grow_injected_times()) {
698 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
704 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
707 for (int i = 0; i < incount; i++) {
708 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
709 int ret = test(&requests[i], pstat, &flag);
714 if (status != MPI_STATUSES_IGNORE)
715 status[*count] = *pstat;
717 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
718 requests[i] = MPI_REQUEST_NULL;
724 if(count_dead==incount)*count=MPI_UNDEFINED;
726 return MPI_ERR_IN_STATUS;
731 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
733 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
734 comms.reserve(count);
737 int ret = MPI_SUCCESS;
738 *index = MPI_UNDEFINED;
740 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
741 for (int i = 0; i < count; i++) {
742 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
743 comms.push_back(requests[i]->action_.get());
747 if (not map.empty()) {
748 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
749 static int nsleeps = 1;
750 if(smpi_test_sleep > 0)
751 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
754 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
755 kernel::actor::ActivityTestanySimcall observer{issuer, comms};
756 i = kernel::actor::simcall_blocking(
757 [&observer] { kernel::activity::ActivityImpl::test_any(observer.get_issuer(), observer.get_activities()); },
759 } catch (const Exception&) {
760 XBT_DEBUG("Exception in testany");
764 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
766 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
767 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
770 finish_wait(&requests[*index],status);
771 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
772 MPI_Status tmp_status;
773 MPI_Status* mystatus;
774 if (status == MPI_STATUS_IGNORE) {
775 mystatus = &tmp_status;
776 Status::empty(mystatus);
780 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
783 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
784 *flag = finish_nbc_requests(&requests[*index] , 1);
787 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
788 requests[*index] = MPI_REQUEST_NULL;
789 XBT_DEBUG("Testany - returning with index %d", *index);
797 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
798 //all requests are null or inactive, return true
800 *index = MPI_UNDEFINED;
801 Status::empty(status);
807 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
810 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
814 for(int i=0; i<count; i++){
815 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
816 int ret = test(&requests[i], pstat, &flag);
822 if (ret != MPI_SUCCESS)
825 Status::empty(pstat);
827 if(status != MPI_STATUSES_IGNORE) {
832 return MPI_ERR_IN_STATUS;
837 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
839 //FIXME find another way to avoid busy waiting ?
840 // the issue here is that we have to wait on a nonexistent comm
842 iprobe(source, tag, comm, &flag, status);
843 XBT_DEBUG("Busy Waiting on probing : %d", flag);
847 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
848 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
849 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
850 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
851 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
852 static int nsleeps = 1;
853 double speed = s4u::this_actor::get_host()->get_speed();
854 double maxrate = smpi_cfg_iprobe_cpu_usage();
856 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
857 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
858 if (smpi_iprobe_sleep > 0) {
859 /** Compute the number of flops we will sleep **/
860 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
861 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
863 /* Not the entire CPU can be used when iprobing: This is important for
864 * the energy consumption caused by polling with iprobes.
865 * Note also that the number of flops that was
866 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
868 ->set_bound(maxrate * speed)
872 // behave like a receive, but don't do it
873 s4u::Mailbox* mailbox;
875 request->print_request("New iprobe");
876 // We have to test both mailboxes as we don't know if we will receive one or another
877 if (smpi_cfg_async_small_thresh() > 0) {
878 mailbox = smpi_process()->mailbox_small();
879 XBT_DEBUG("Trying to probe the perm recv mailbox");
880 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
883 if (request->action_ == nullptr){
884 mailbox = smpi_process()->mailbox();
885 XBT_DEBUG("trying to probe the other mailbox");
886 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
889 if (request->action_ != nullptr){
890 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
891 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
893 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
894 status->MPI_SOURCE = comm->group()->rank(req->src_);
895 status->MPI_TAG = req->tag_;
896 status->MPI_ERROR = MPI_SUCCESS;
897 status->count = req->real_size_;
899 nsleeps = 1;//reset the number of sleeps we will do next time
903 if (smpi_cfg_grow_injected_times())
907 xbt_assert(request == MPI_REQUEST_NULL);
910 int Request::finish_nbc_requests(MPI_Request* request, int test){
914 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
916 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
919 xbt_die("Failure when waiting on non blocking collective sub-requests");
921 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
922 for(auto& req: (*request)->nbc_requests_){
923 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
924 void * buf=req->buf_;
925 if((*request)->type_->flags() & DT_FLAG_DERIVED)
927 if(req->flags_ & MPI_REQ_RECV ){
928 if((*request)->op_!=MPI_OP_NULL){
929 int count=(*request)->size_/ (*request)->type_->size();
930 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
932 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
935 if(req!=MPI_REQUEST_NULL)
936 Request::unref(&req);
938 (*request)->nbc_requests_.clear();
943 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
945 MPI_Request req = *request;
946 Status::empty(status);
947 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
948 if (status!=MPI_STATUS_IGNORE)
950 if(req->detached_sender_ != nullptr)
951 unref(&(req->detached_sender_));
956 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
957 if (status != MPI_STATUS_IGNORE) {
958 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
959 Status::empty(status);
960 status->MPI_SOURCE = MPI_PROC_NULL;
962 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
963 status->MPI_SOURCE = req->comm_->group()->rank(src);
964 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
965 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
967 // this handles the case were size in receive differs from size in send
968 status->count = req->real_size_;
970 //detached send will be finished at the other end
971 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
972 req->print_request("Finishing");
973 MPI_Datatype datatype = req->type_;
975 // FIXME Handle the case of a partial shared malloc.
976 if (not smpi_process()->replaying() &&
977 (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
978 if (smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
979 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
981 if(datatype->flags() & DT_FLAG_DERIVED){
982 // This part handles the problem of non-contiguous memory the unserialization at the reception
983 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
984 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
987 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
988 if (datatype->size() != 0) {
989 int n = req->real_size_ / datatype->size();
990 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
999 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
1000 aid_t rank = simgrid::s4u::this_actor::get_pid();
1001 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
1002 TRACE_smpi_recv(src_traced, rank,req->tag_);
1004 if(req->detached_sender_ != nullptr){
1005 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
1006 simgrid::s4u::Host* dst_host = simgrid::s4u::Actor::by_pid(req->dst_)->get_host();
1007 double sleeptime = simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(
1008 req->real_size(), req->src_host_, dst_host);
1009 if (sleeptime > 0.0) {
1010 simgrid::s4u::this_actor::sleep_for(sleeptime);
1011 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
1013 unref(&(req->detached_sender_));
1015 if (req->flags_ & MPI_REQ_PERSISTENT)
1016 req->action_ = nullptr;
1017 req->flags_ |= MPI_REQ_FINISHED;
1019 if (req->truncated_ || req->unmatched_types_) {
1020 char error_string[MPI_MAX_ERROR_STRING];
1023 if(req->truncated_ )
1024 errkind = MPI_ERR_TRUNCATE;
1026 errkind = MPI_ERR_TYPE;
1027 PMPI_Error_string(errkind, error_string, &error_size);
1028 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
1029 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
1030 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1031 else if (err == MPI_ERRORS_ARE_FATAL)
1032 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1034 err->call((req->comm_), errkind);
1035 if (err != MPI_ERRHANDLER_NULL)
1036 simgrid::smpi::Errhandler::unref(err);
1037 MC_assert(not MC_is_active()); /* Only fail in MC mode */
1039 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
1043 int Request::wait(MPI_Request * request, MPI_Status * status)
1045 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
1046 xbt_assert(*request != MPI_REQUEST_NULL);
1048 int ret=MPI_SUCCESS;
1050 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
1051 if (status != MPI_STATUS_IGNORE) {
1052 Status::empty(status);
1053 status->MPI_SOURCE = MPI_PROC_NULL;
1055 (*request)=MPI_REQUEST_NULL;
1059 (*request)->print_request("Waiting");
1060 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1061 Status::empty(status);
1065 if ((*request)->action_ != nullptr){
1067 // this is not a detached send
1068 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1069 kernel::actor::ActivityWaitSimcall observer{issuer, (*request)->action_.get(), -1};
1070 kernel::actor::simcall_blocking(
1071 [&observer] { observer.get_activity()->wait_for(observer.get_issuer(), observer.get_timeout()); },
1073 } catch (const CancelException&) {
1074 XBT_VERB("Request cancelled");
1078 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1079 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1080 ((*request)->generalized_funcs)->mutex->lock();
1081 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1082 ((*request)->generalized_funcs)->mutex->unlock();
1084 MPI_Status tmp_status;
1085 MPI_Status* mystatus;
1086 if (status == MPI_STATUS_IGNORE) {
1087 mystatus = &tmp_status;
1088 Status::empty(mystatus);
1092 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1095 if ((*request)->truncated_)
1096 ret = MPI_ERR_TRUNCATE;
1098 if ((*request)->flags_ & MPI_REQ_NBC)
1099 finish_nbc_requests(request, 0);
1101 finish_wait(request, status); // may invalidate *request
1102 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1103 *request = MPI_REQUEST_NULL;
1107 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1109 int index = MPI_UNDEFINED;
1112 // Wait for a request to complete
1113 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
1114 std::vector<int> map;
1115 XBT_DEBUG("Wait for one of %d", count);
1116 for(int i = 0; i < count; i++) {
1117 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1118 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1119 if (requests[i]->action_ != nullptr) {
1120 XBT_DEBUG("Waiting any %p ", requests[i]);
1121 comms.push_back(requests[i]->action_.get());
1124 // This is a finished detached request, let's return this one
1125 comms.clear(); // don't do the waitany call afterwards
1127 if (requests[index]->flags_ & MPI_REQ_NBC)
1128 finish_nbc_requests(&requests[index], 0);
1129 finish_wait(&requests[i], status); // cleanup if refcount = 0
1130 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1131 requests[i] = MPI_REQUEST_NULL; // set to null
1136 if (not comms.empty()) {
1137 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1140 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1141 kernel::actor::ActivityWaitanySimcall observer{issuer, comms, -1};
1142 i = kernel::actor::simcall_blocking(
1144 kernel::activity::ActivityImpl::wait_any_for(observer.get_issuer(), observer.get_activities(),
1145 observer.get_timeout());
1148 } catch (const CancelException&) {
1149 XBT_INFO("request cancelled");
1153 // not MPI_UNDEFINED, as this is a simix return code
1156 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1157 if ((requests[index] == MPI_REQUEST_NULL) ||
1158 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1159 finish_wait(&requests[index],status);
1160 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1161 requests[index] = MPI_REQUEST_NULL;
1168 if (index==MPI_UNDEFINED)
1169 Status::empty(status);
1174 static int sort_accumulates(const Request* a, const Request* b)
1176 return (a->tag() > b->tag());
1179 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1181 std::vector<MPI_Request> accumulates;
1184 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1185 int retvalue = MPI_SUCCESS;
1186 //tag invalid requests in the set
1187 if (status != MPI_STATUSES_IGNORE) {
1188 for (int c = 0; c < count; c++) {
1189 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1190 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1191 Status::empty(&status[c]);
1192 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1193 Status::empty(&status[c]);
1194 status[c].MPI_SOURCE = MPI_PROC_NULL;
1198 for (int c = 0; c < count; c++) {
1199 if (MC_is_active() || MC_record_replay_is_active()) {
1200 wait(&requests[c],pstat);
1203 index = waitany(count, requests, pstat);
1205 if (index == MPI_UNDEFINED)
1208 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1209 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1210 accumulates.push_back(requests[index]);
1211 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1212 requests[index] = MPI_REQUEST_NULL;
1214 if (status != MPI_STATUSES_IGNORE) {
1215 status[index] = *pstat;
1216 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1217 retvalue = MPI_ERR_IN_STATUS;
1221 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1222 for (auto& req : accumulates)
1223 finish_wait(&req, status);
1228 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1234 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1235 index = waitany(incount, requests, pstat);
1236 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1237 if(status != MPI_STATUSES_IGNORE) {
1238 status[count] = *pstat;
1240 indices[count] = index;
1242 for (int i = 0; i < incount; i++) {
1243 if (i != index && requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1244 test(&requests[i], pstat,&flag);
1247 if(status != MPI_STATUSES_IGNORE) {
1248 status[count] = *pstat;
1250 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1251 requests[i]=MPI_REQUEST_NULL;
1259 MPI_Request Request::f2c(int id)
1261 if(id==MPI_FORTRAN_REQUEST_NULL)
1262 return MPI_REQUEST_NULL;
1263 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1266 void Request::free_f(int id)
1268 if (id != MPI_FORTRAN_REQUEST_NULL) {
1269 F2C::lookup()->erase(id);
1273 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1277 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1278 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1282 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1288 if(req != MPI_REQUEST_NULL &&
1289 status != MPI_STATUS_IGNORE) {
1290 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1291 status->MPI_SOURCE = req->comm_->group()->rank(src);
1292 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1293 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1294 status->count = req->real_size_;
1299 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1300 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1302 *request = new Request();
1303 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1304 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1305 (*request)->refcount_ = 1;
1306 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1307 ((*request)->generalized_funcs)->query_fn=query_fn;
1308 ((*request)->generalized_funcs)->free_fn=free_fn;
1309 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1310 ((*request)->generalized_funcs)->extra_state=extra_state;
1311 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1312 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1316 int Request::grequest_complete(MPI_Request request)
1318 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1319 return MPI_ERR_REQUEST;
1320 request->generalized_funcs->mutex->lock();
1321 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1322 request->generalized_funcs->cond->notify_one();
1323 request->generalized_funcs->mutex->unlock();
1327 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1328 if (not reqs.empty()) {
1329 nbc_requests_ = reqs;
1330 Request::startall(reqs.size(), reqs.data());
1334 std::vector<MPI_Request> Request::get_nbc_requests() const
1336 return nbc_requests_;