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/EngineImpl.hpp"
19 #include "src/kernel/activity/CommImpl.hpp"
20 #include "src/kernel/actor/ActorImpl.hpp"
21 #include "src/kernel/actor/SimcallObserver.hpp"
22 #include "src/mc/mc_replay.hpp"
23 #include "src/smpi/include/smpi_actor.hpp"
28 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
30 static simgrid::config::Flag<double> smpi_iprobe_sleep(
31 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
32 static simgrid::config::Flag<double> smpi_test_sleep(
33 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
35 std::vector<s_smpi_factor_t> smpi_ois_values;
37 extern std::function<void(simgrid::kernel::activity::CommImpl*, void*, size_t)> smpi_comm_copy_data_callback;
39 namespace simgrid::smpi {
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 if (auto src_process = simgrid::s4u::Actor::by_pid(src))
64 src_host_ = src_process->get_host();
66 unmatched_types_ = false;
69 if (flags & MPI_REQ_PERSISTENT)
82 void Request::unref(MPI_Request* request)
84 xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
86 (*request)->refcount_--;
87 if ((*request)->refcount_ < 0) {
88 (*request)->print_request("wrong refcount");
89 xbt_die("Whoops, wrong refcount");
91 if ((*request)->refcount_ == 0) {
92 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
93 ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
95 Comm::unref((*request)->comm_);
96 Datatype::unref((*request)->type_);
98 if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
99 Op::unref(&(*request)->op_);
101 (*request)->print_request("Destroying");
102 F2C::free_f((*request)->f2c_id());
104 *request = MPI_REQUEST_NULL;
106 (*request)->print_request("Decrementing");
110 bool Request::match_types(MPI_Datatype stype, MPI_Datatype rtype){
112 if ((stype == rtype) ||
113 //byte and packed always match with anything
114 (stype == MPI_PACKED || rtype == MPI_PACKED || stype == MPI_BYTE || rtype == MPI_BYTE) ||
115 //complex datatypes - we don't properly match these yet, as it would mean checking each subtype recursively.
116 (stype->flags() & DT_FLAG_DERIVED || rtype->flags() & DT_FLAG_DERIVED) ||
117 //duplicated datatypes, check if underlying is ok
118 (stype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype->duplicated_datatype(), rtype)) ||
119 (rtype->duplicated_datatype()!=MPI_DATATYPE_NULL && match_types(stype, rtype->duplicated_datatype())))
122 XBT_WARN("Mismatched datatypes : sending %s and receiving %s", stype->name().c_str(), rtype->name().c_str());
127 bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
129 xbt_assert(sender, "Cannot match against null sender");
130 xbt_assert(receiver, "Cannot match against null receiver");
131 XBT_DEBUG("Trying to match %s of sender src %ld against %ld, tag %d against %d, id %d against %d",
132 (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
133 sender->comm_->id(), receiver->comm_->id());
135 if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED ||
136 receiver->comm_->id() == sender->comm_->id()) &&
137 ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) ||
138 receiver->src_ == sender->src_) &&
139 ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) {
140 // we match, we can transfer some values
141 if (receiver->src_ == MPI_ANY_SOURCE) {
142 receiver->real_src_ = sender->src_;
143 receiver->src_host_ = sender->src_host_;
145 if (receiver->tag_ == MPI_ANY_TAG)
146 receiver->real_tag_ = sender->tag_;
147 if ((receiver->flags_ & MPI_REQ_PROBE) == 0 && receiver->real_size_ < sender->real_size_) {
148 XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_,
150 receiver->truncated_ = true;
152 //0-sized datatypes/counts should not interfere and match
153 if ( sender->real_size_ != 0 && receiver->real_size_ != 0 &&
154 !match_types(sender->type_, receiver->type_))
155 receiver->unmatched_types_ = true;
156 if (sender->detached_)
157 receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
159 req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
160 XBT_DEBUG("match succeeded");
166 void Request::init_buffer(int count){
167 // FIXME Handle the case of a partial shared malloc.
168 // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
169 if (not smpi_process()->replaying() &&
170 ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED))) {
171 // This part handles the problem of non-contiguous memory
176 buf_ = xbt_malloc(count*type_->size());
177 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
178 type_->serialize(old_buf_, buf_, count);
184 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
186 auto ref = static_cast<MPI_Request>(a);
187 auto req = static_cast<MPI_Request>(b);
188 bool match = match_common(req, req, ref);
189 if (not 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_),
193 ref->comm_->group()->rank(req->dst_), req->tag_) == req->message_id_) {
194 if (((ref->flags_ & MPI_REQ_PROBE) == 0) && ((req->flags_ & MPI_REQ_PROBE) == 0)) {
195 XBT_DEBUG("increasing count in comm %p, which was %u from pid %ld, to pid %ld with tag %d", ref->comm_,
196 ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_),
197 ref->comm_->group()->rank(req->dst_), req->tag_),
198 req->src_, req->dst_, req->tag_);
199 ref->comm_->increment_received_messages_count(ref->comm_->group()->rank(req->src_),
200 ref->comm_->group()->rank(req->dst_), req->tag_);
201 if (ref->real_size_ > req->real_size_) {
202 ref->real_size_ = req->real_size_;
207 req->flags_ &= ~MPI_REQ_MATCHED;
208 ref->detached_sender_ = nullptr;
209 XBT_DEBUG("Refusing to match message, as its ID is not the one I expect. in comm %p, %u != %u, "
210 "from pid %ld to pid %ld, with tag %d",
212 ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_),
213 ref->comm_->group()->rank(req->dst_), req->tag_),
214 req->message_id_, req->src_, req->dst_, req->tag_);
219 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
221 auto ref = static_cast<MPI_Request>(a);
222 auto req = static_cast<MPI_Request>(b);
223 return match_common(req, ref, req);
226 void Request::print_request(const char* message) const
228 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
229 size_, src_, dst_, tag_, flags_);
232 /* factories, to hide the internal flags from the caller */
233 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
235 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
236 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
237 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
240 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
242 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
243 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
244 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
247 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
249 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
250 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
251 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
254 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
256 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
257 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
258 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
261 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
266 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
267 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
268 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
270 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
271 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
272 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
279 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
281 aid_t source = MPI_PROC_NULL;
282 if (src == MPI_ANY_SOURCE)
283 source = MPI_ANY_SOURCE;
284 else if (src != MPI_PROC_NULL)
285 source = comm->group()->actor(src);
286 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
288 simgrid::s4u::this_actor::get_pid(), tag, comm,
289 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
292 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
295 aid_t source = MPI_PROC_NULL;
296 if (src == MPI_ANY_SOURCE)
297 source = MPI_ANY_SOURCE;
298 else if (src != MPI_PROC_NULL)
299 source = comm->group()->actor(src);
302 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
303 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
304 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
307 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
308 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
309 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
314 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
316 aid_t source = MPI_PROC_NULL;
317 if (src == MPI_ANY_SOURCE)
318 source = MPI_ANY_SOURCE;
319 else if (src != MPI_PROC_NULL)
320 source = comm->group()->actor(src);
321 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
322 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
323 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
326 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
328 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
329 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
330 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
331 if(dst != MPI_PROC_NULL)
336 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
338 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
339 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
340 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
341 if(dst != MPI_PROC_NULL)
346 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
348 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
349 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
350 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
351 if(dst != MPI_PROC_NULL)
356 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
358 aid_t source = MPI_PROC_NULL;
359 if (src == MPI_ANY_SOURCE)
360 source = MPI_ANY_SOURCE;
361 else if (src != MPI_PROC_NULL)
362 source = comm->group()->actor(src);
363 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
364 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
365 if(src != MPI_PROC_NULL)
370 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
372 MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
373 int retval = wait(&request,status);
378 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
380 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
381 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
382 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
384 if(dst != MPI_PROC_NULL)
386 wait(&request, MPI_STATUS_IGNORE);
390 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
392 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
393 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
394 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
395 if(dst != MPI_PROC_NULL)
397 wait(&request, MPI_STATUS_IGNORE);
401 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
403 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
404 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
405 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
407 if(dst != MPI_PROC_NULL)
409 wait(&request,MPI_STATUS_IGNORE);
413 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
414 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
415 MPI_Comm comm, MPI_Status * status)
417 aid_t source = MPI_PROC_NULL;
418 if (src == MPI_ANY_SOURCE)
419 source = MPI_ANY_SOURCE;
420 else if (src != MPI_PROC_NULL)
421 source = comm->group()->actor(src);
422 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
424 std::array<MPI_Request, 2> requests;
425 std::array<MPI_Status, 2> stats;
426 if (aid_t myid = simgrid::s4u::this_actor::get_pid(); (destination == myid) && (source == myid)) {
427 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
428 if (status != MPI_STATUS_IGNORE) {
429 status->MPI_SOURCE = source;
430 status->MPI_TAG = recvtag;
431 status->MPI_ERROR = MPI_SUCCESS;
432 status->count = sendcount * sendtype->size();
436 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
437 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
438 startall(2, requests.data());
439 waitall(2, requests.data(), stats.data());
442 if(status != MPI_STATUS_IGNORE) {
443 // Copy receive status
448 void Request::start()
450 s4u::Mailbox* mailbox;
452 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
453 //reinitialize temporary buffer for persistent requests
454 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
456 init_buffer(real_size_/type_->size());
458 flags_ &= ~MPI_REQ_PREPARED;
459 flags_ &= ~MPI_REQ_FINISHED;
462 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
464 if ((flags_ & MPI_REQ_RECV) != 0) {
465 this->print_request("New recv");
467 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
469 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
470 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
473 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
474 flags_ |= MPI_REQ_PROBE;
476 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
477 mailbox = process->mailbox();
478 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
479 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
480 //begin with the more appropriate one : the small one.
481 mailbox = process->mailbox_small();
482 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
483 mailbox->get_cname());
484 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
486 if (action == nullptr) {
487 mailbox = process->mailbox();
488 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
489 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
490 if (action == nullptr) {
491 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
492 mailbox = process->mailbox_small();
495 XBT_DEBUG("yes there was something for us in the small mailbox");
498 mailbox = process->mailbox_small();
499 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
500 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
502 if (action == nullptr) {
503 XBT_DEBUG("No, nothing in the permanent receive mailbox");
504 mailbox = process->mailbox();
506 XBT_DEBUG("yes there was something for us in the small mailbox");
510 flags_ &= ~MPI_REQ_PROBE;
511 kernel::actor::CommIrecvSimcall observer{process->get_actor()->get_impl(),
513 static_cast<unsigned char*>(buf_),
516 process->replaying() ? &smpi_comm_null_copy_buffer_callback
517 : smpi_comm_copy_data_callback,
520 observer.set_tag(tag_);
522 action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::irecv(&observer); },
525 XBT_DEBUG("recv simcall posted");
527 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
529 } else { /* the RECV flag was not set, so this is a send */
530 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
531 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
532 if (TRACE_smpi_view_internals())
533 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
534 this->print_request("New send");
536 message_id_=comm_->get_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
537 comm_->increment_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
540 if ((flags_ & MPI_REQ_SSEND) == 0 &&
541 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
542 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
543 void *oldbuf = nullptr;
545 XBT_DEBUG("Send request %p is detached", this);
547 if (not(type_->flags() & DT_FLAG_DERIVED)) {
549 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
550 if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
551 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
553 //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
554 //so actually ... don't use manually attached buffer space.
555 buf = xbt_malloc(size_);
556 memcpy(buf,oldbuf,size_);
557 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
562 //if we are giving back the control to the user without waiting for completion, we have to inject timings
563 double sleeptime = 0.0;
564 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
565 // isend and send timings may be different
567 ((flags_ & MPI_REQ_ISEND) != 0)
568 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(
569 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(), simgrid::s4u::Actor::by_pid(dst_)->get_host())
570 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(
571 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(),
572 simgrid::s4u::Actor::by_pid(dst_)->get_host());
576 simgrid::s4u::this_actor::sleep_for(sleeptime);
577 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
580 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
582 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
585 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
586 mailbox = process->mailbox();
587 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
588 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
589 flags_ |= MPI_REQ_PROBE;
591 mailbox = process->mailbox();
592 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
593 if (not mailbox->iprobe(1, &match_send, static_cast<void*>(this))) {
594 if ((flags_ & MPI_REQ_SSEND) == 0) {
595 mailbox = process->mailbox_small();
596 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
597 mailbox->get_cname());
599 mailbox = process->mailbox_small();
600 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
601 mailbox->get_cname());
602 if (not mailbox->iprobe(1, &match_send, static_cast<void*>(this))) {
603 XBT_DEBUG("No, we are first, send to large mailbox");
604 mailbox = process->mailbox();
608 XBT_DEBUG("Yes there was something for us in the large mailbox");
611 flags_ &= ~MPI_REQ_PROBE;
613 mailbox = process->mailbox();
614 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
617 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
618 kernel::actor::CommIsendSimcall observer{
619 simgrid::kernel::EngineImpl::get_instance()->get_actor_by_pid(src_), mailbox->get_impl(),
620 static_cast<double>(payload_size_), -1, static_cast<unsigned char*>(buf), real_size_, &match_send,
621 &xbt_free_f, // how to free the userdata if a detached send fails
622 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
623 // detach if msg size < eager/rdv switch limit
625 observer.set_tag(tag_);
626 action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::isend(&observer); },
628 XBT_DEBUG("send simcall posted");
630 /* FIXME: detached sends are not traceable (action_ == nullptr) */
631 if (action_ != nullptr) {
632 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
633 smpi_process()->get_tracing_category());
636 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
641 void Request::startall(int count, MPI_Request * requests)
643 if(requests== nullptr)
646 for(int i = 0; i < count; i++) {
647 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
648 requests[i]->start();
652 void Request::cancel()
654 this->flags_ |= MPI_REQ_CANCELLED;
655 if (this->action_ != nullptr)
656 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
659 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
660 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
661 // to avoid deadlocks if used as a break condition, such as
662 // while (MPI_Test(request, flag, status) && flag) dostuff...
663 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
664 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
665 xbt_assert(*request != MPI_REQUEST_NULL);
667 static int nsleeps = 1;
668 int ret = MPI_SUCCESS;
670 if(smpi_test_sleep > 0)
671 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
673 Status::empty(status);
676 if ((*request)->flags_ & MPI_REQ_NBC){
677 *flag = finish_nbc_requests(request, 1);
680 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
681 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
683 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
684 kernel::actor::ActivityTestSimcall observer{issuer, (*request)->action_.get()};
685 *flag = kernel::actor::simcall_answered(
686 [&observer] { return observer.get_activity()->test(observer.get_issuer()); }, &observer);
687 } catch (const Exception&) {
692 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
695 finish_wait(request, status); // may invalidate *request
696 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
697 MPI_Status tmp_status;
698 MPI_Status* mystatus;
699 if (status == MPI_STATUS_IGNORE) {
700 mystatus = &tmp_status;
701 Status::empty(mystatus);
705 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
707 nsleeps=1;//reset the number of sleeps we will do next time
708 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
709 *request = MPI_REQUEST_NULL;
710 } else if (smpi_cfg_grow_injected_times()) {
717 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
723 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
726 for (int i = 0; i < incount; i++) {
727 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
728 if (test(&requests[i], pstat, &flag) != MPI_SUCCESS)
732 if (status != MPI_STATUSES_IGNORE)
733 status[*count] = *pstat;
735 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
736 requests[i] = MPI_REQUEST_NULL;
742 if(count_dead==incount)*count=MPI_UNDEFINED;
744 return MPI_ERR_IN_STATUS;
749 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
751 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
752 comms.reserve(count);
755 int ret = MPI_SUCCESS;
756 *index = MPI_UNDEFINED;
758 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
759 for (int i = 0; i < count; i++) {
760 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
761 comms.push_back(requests[i]->action_.get());
765 if (not map.empty()) {
766 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
767 static int nsleeps = 1;
768 if(smpi_test_sleep > 0)
769 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
772 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
773 kernel::actor::ActivityTestanySimcall observer{issuer, comms};
774 i = kernel::actor::simcall_answered(
776 return kernel::activity::ActivityImpl::test_any(observer.get_issuer(), observer.get_activities());
779 } catch (const Exception&) {
780 XBT_DEBUG("Exception in testany");
784 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
786 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
787 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
790 finish_wait(&requests[*index],status);
791 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
792 MPI_Status tmp_status;
793 MPI_Status* mystatus;
794 if (status == MPI_STATUS_IGNORE) {
795 mystatus = &tmp_status;
796 Status::empty(mystatus);
800 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
803 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
804 *flag = finish_nbc_requests(&requests[*index] , 1);
807 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
808 requests[*index] = MPI_REQUEST_NULL;
809 XBT_DEBUG("Testany - returning with index %d", *index);
817 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
818 //all requests are null or inactive, return true
820 *index = MPI_UNDEFINED;
821 Status::empty(status);
827 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
830 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
834 for(int i=0; i<count; i++){
835 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
836 int ret = test(&requests[i], pstat, &flag);
842 if (ret != MPI_SUCCESS)
845 Status::empty(pstat);
847 if(status != MPI_STATUSES_IGNORE) {
852 return MPI_ERR_IN_STATUS;
857 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
859 //FIXME find another way to avoid busy waiting ?
860 // the issue here is that we have to wait on a nonexistent comm
862 iprobe(source, tag, comm, &flag, status);
863 XBT_DEBUG("Busy Waiting on probing : %d", flag);
867 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
868 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
869 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
870 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
871 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
872 static int nsleeps = 1;
873 double speed = s4u::this_actor::get_host()->get_speed();
874 double maxrate = smpi_cfg_iprobe_cpu_usage();
876 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
877 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
878 if (smpi_iprobe_sleep > 0) {
879 /** Compute the number of flops we will sleep **/
880 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
881 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
883 /* Not the entire CPU can be used when iprobing: This is important for
884 * the energy consumption caused by polling with iprobes.
885 * Note also that the number of flops that was
886 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
888 ->set_bound(maxrate * speed)
892 // behave like a receive, but don't do it
893 s4u::Mailbox* mailbox;
895 request->print_request("New iprobe");
896 // We have to test both mailboxes as we don't know if we will receive one or another
897 if (smpi_cfg_async_small_thresh() > 0) {
898 mailbox = smpi_process()->mailbox_small();
899 XBT_DEBUG("Trying to probe the perm recv mailbox");
900 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
903 if (request->action_ == nullptr){
904 mailbox = smpi_process()->mailbox();
905 XBT_DEBUG("trying to probe the other mailbox");
906 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
909 if (request->action_ != nullptr){
910 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
911 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
913 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
914 status->MPI_SOURCE = comm->group()->rank(req->src_);
915 status->MPI_TAG = req->tag_;
916 status->MPI_ERROR = MPI_SUCCESS;
917 status->count = req->real_size_;
919 nsleeps = 1;//reset the number of sleeps we will do next time
923 if (smpi_cfg_grow_injected_times())
927 xbt_assert(request == MPI_REQUEST_NULL);
930 int Request::finish_nbc_requests(MPI_Request* request, int test){
934 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
936 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
939 xbt_die("Failure when waiting on non blocking collective sub-requests");
941 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
942 for(auto& req: (*request)->nbc_requests_){
943 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
944 void * buf=req->buf_;
945 if((*request)->type_->flags() & DT_FLAG_DERIVED)
947 if(req->flags_ & MPI_REQ_RECV ){
948 if((*request)->op_!=MPI_OP_NULL){
949 int count=(*request)->size_/ (*request)->type_->size();
950 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
952 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
955 if(req!=MPI_REQUEST_NULL)
956 Request::unref(&req);
958 (*request)->nbc_requests_.clear();
963 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
965 MPI_Request req = *request;
966 Status::empty(status);
967 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
968 if (status!=MPI_STATUS_IGNORE)
970 if(req->detached_sender_ != nullptr)
971 unref(&(req->detached_sender_));
976 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
977 if (status != MPI_STATUS_IGNORE) {
978 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
979 Status::empty(status);
980 status->MPI_SOURCE = MPI_PROC_NULL;
982 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
983 status->MPI_SOURCE = req->comm_->group()->rank(src);
984 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
985 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
987 // this handles the case were size in receive differs from size in send
988 status->count = req->real_size_;
990 //detached send will be finished at the other end
991 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
992 req->print_request("Finishing");
993 MPI_Datatype datatype = req->type_;
995 // FIXME Handle the case of a partial shared malloc.
996 if (not smpi_process()->replaying() &&
997 (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
998 if (smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
999 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
1001 if(datatype->flags() & DT_FLAG_DERIVED){
1002 // This part handles the problem of non-contiguous memory the unserialization at the reception
1003 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
1004 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
1005 xbt_free(req->buf_);
1007 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
1008 if (datatype->size() != 0) {
1009 int n = req->real_size_ / datatype->size();
1010 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
1012 xbt_free(req->buf_);
1019 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
1020 aid_t rank = simgrid::s4u::this_actor::get_pid();
1021 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
1022 TRACE_smpi_recv(src_traced, rank,req->tag_);
1024 if(req->detached_sender_ != nullptr){
1025 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
1026 simgrid::s4u::Host* dst_host = simgrid::s4u::Actor::by_pid(req->dst_)->get_host();
1027 if (double sleeptime = simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(
1028 req->real_size(), req->src_host_, dst_host);
1030 simgrid::s4u::this_actor::sleep_for(sleeptime);
1031 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
1033 unref(&(req->detached_sender_));
1035 if (req->flags_ & MPI_REQ_PERSISTENT)
1036 req->action_ = nullptr;
1037 req->flags_ |= MPI_REQ_FINISHED;
1039 if (req->truncated_ || req->unmatched_types_) {
1040 char error_string[MPI_MAX_ERROR_STRING];
1043 if(req->truncated_ )
1044 errkind = MPI_ERR_TRUNCATE;
1046 errkind = MPI_ERR_TYPE;
1047 PMPI_Error_string(errkind, error_string, &error_size);
1048 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
1049 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
1050 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1051 else if (err == MPI_ERRORS_ARE_FATAL)
1052 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1054 err->call((req->comm_), errkind);
1055 if (err != MPI_ERRHANDLER_NULL)
1056 simgrid::smpi::Errhandler::unref(err);
1057 MC_assert(not MC_is_active()); /* Only fail in MC mode */
1059 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
1063 int Request::wait(MPI_Request * request, MPI_Status * status)
1065 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
1066 xbt_assert(*request != MPI_REQUEST_NULL);
1068 int ret=MPI_SUCCESS;
1070 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
1071 if (status != MPI_STATUS_IGNORE) {
1072 Status::empty(status);
1073 status->MPI_SOURCE = MPI_PROC_NULL;
1075 (*request)=MPI_REQUEST_NULL;
1079 (*request)->print_request("Waiting");
1080 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1081 Status::empty(status);
1085 if ((*request)->action_ != nullptr){
1087 // this is not a detached send
1088 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1089 kernel::actor::ActivityWaitSimcall observer{issuer, (*request)->action_.get(), -1};
1090 kernel::actor::simcall_blocking([issuer, &observer] { observer.get_activity()->wait_for(issuer, -1); },
1092 } catch (const CancelException&) {
1093 XBT_VERB("Request cancelled");
1097 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1098 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1099 ((*request)->generalized_funcs)->mutex->lock();
1100 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1101 ((*request)->generalized_funcs)->mutex->unlock();
1103 MPI_Status tmp_status;
1104 MPI_Status* mystatus;
1105 if (status == MPI_STATUS_IGNORE) {
1106 mystatus = &tmp_status;
1107 Status::empty(mystatus);
1111 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1114 if ((*request)->truncated_)
1115 ret = MPI_ERR_TRUNCATE;
1117 if ((*request)->flags_ & MPI_REQ_NBC)
1118 finish_nbc_requests(request, 0);
1120 finish_wait(request, status); // may invalidate *request
1121 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1122 *request = MPI_REQUEST_NULL;
1126 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1128 int index = MPI_UNDEFINED;
1131 // Wait for a request to complete
1132 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
1133 std::vector<int> map;
1134 XBT_DEBUG("Wait for one of %d", count);
1135 for(int i = 0; i < count; i++) {
1136 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1137 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1138 if (requests[i]->action_ != nullptr) {
1139 XBT_DEBUG("Waiting any %p ", requests[i]);
1140 comms.push_back(requests[i]->action_.get());
1143 // This is a finished detached request, let's return this one
1144 comms.clear(); // don't do the waitany call afterwards
1146 if (requests[index]->flags_ & MPI_REQ_NBC)
1147 finish_nbc_requests(&requests[index], 0);
1148 finish_wait(&requests[i], status); // cleanup if refcount = 0
1149 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1150 requests[i] = MPI_REQUEST_NULL; // set to null
1155 if (not comms.empty()) {
1156 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1159 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1160 kernel::actor::ActivityWaitanySimcall observer{issuer, comms, -1};
1161 i = kernel::actor::simcall_blocking(
1163 kernel::activity::ActivityImpl::wait_any_for(observer.get_issuer(), observer.get_activities(),
1164 observer.get_timeout());
1167 } catch (const CancelException&) {
1168 XBT_INFO("request cancelled");
1172 // not MPI_UNDEFINED, as this is a simix return code
1175 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1176 if ((requests[index] == MPI_REQUEST_NULL) ||
1177 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1178 finish_wait(&requests[index],status);
1179 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1180 requests[index] = MPI_REQUEST_NULL;
1187 if (index==MPI_UNDEFINED)
1188 Status::empty(status);
1193 static int sort_accumulates(const Request* a, const Request* b)
1195 return (a->tag() > b->tag());
1198 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1200 std::vector<MPI_Request> accumulates;
1203 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1204 int retvalue = MPI_SUCCESS;
1205 //tag invalid requests in the set
1206 if (status != MPI_STATUSES_IGNORE) {
1207 for (int c = 0; c < count; c++) {
1208 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1209 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1210 Status::empty(&status[c]);
1211 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1212 Status::empty(&status[c]);
1213 status[c].MPI_SOURCE = MPI_PROC_NULL;
1217 for (int c = 0; c < count; c++) {
1218 if (MC_is_active() || MC_record_replay_is_active()) {
1219 wait(&requests[c],pstat);
1222 index = waitany(count, requests, pstat);
1224 if (index == MPI_UNDEFINED)
1227 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1228 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1229 accumulates.push_back(requests[index]);
1230 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1231 requests[index] = MPI_REQUEST_NULL;
1233 if (status != MPI_STATUSES_IGNORE) {
1234 status[index] = *pstat;
1235 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1236 retvalue = MPI_ERR_IN_STATUS;
1240 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1241 for (auto& req : accumulates)
1242 finish_wait(&req, status);
1247 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1253 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1254 index = waitany(incount, requests, pstat);
1255 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1256 if(status != MPI_STATUSES_IGNORE) {
1257 status[count] = *pstat;
1259 indices[count] = index;
1261 for (int i = 0; i < incount; i++) {
1262 if (i != index && requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1263 test(&requests[i], pstat,&flag);
1266 if(status != MPI_STATUSES_IGNORE) {
1267 status[count] = *pstat;
1269 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1270 requests[i]=MPI_REQUEST_NULL;
1278 MPI_Request Request::f2c(int id)
1280 if(id==MPI_FORTRAN_REQUEST_NULL)
1281 return MPI_REQUEST_NULL;
1282 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1285 void Request::free_f(int id)
1287 if (id != MPI_FORTRAN_REQUEST_NULL) {
1288 F2C::lookup()->erase(id);
1292 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1296 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1297 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1301 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1307 if(req != MPI_REQUEST_NULL &&
1308 status != MPI_STATUS_IGNORE) {
1309 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1310 status->MPI_SOURCE = req->comm_->group()->rank(src);
1311 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1312 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1313 status->count = req->real_size_;
1318 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1319 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1321 *request = new Request();
1322 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1323 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1324 (*request)->refcount_ = 1;
1325 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1326 ((*request)->generalized_funcs)->query_fn=query_fn;
1327 ((*request)->generalized_funcs)->free_fn=free_fn;
1328 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1329 ((*request)->generalized_funcs)->extra_state=extra_state;
1330 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1331 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1335 int Request::grequest_complete(MPI_Request request)
1337 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1338 return MPI_ERR_REQUEST;
1339 request->generalized_funcs->mutex->lock();
1340 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1341 request->generalized_funcs->cond->notify_one();
1342 request->generalized_funcs->mutex->unlock();
1346 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1347 if (not reqs.empty()) {
1348 nbc_requests_ = reqs;
1349 Request::startall(reqs.size(), reqs.data());
1353 std::vector<MPI_Request> Request::get_nbc_requests() const
1355 return nbc_requests_;
1357 } // namespace simgrid::smpi