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 && not match_types(sender->type_, receiver->type_))
154 receiver->unmatched_types_ = true;
155 if (sender->detached_)
156 receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
158 req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
159 XBT_DEBUG("match succeeded");
165 void Request::init_buffer(int count){
166 // FIXME Handle the case of a partial shared malloc.
167 // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
168 if (not smpi_process()->replaying() &&
169 ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED))) {
170 // This part handles the problem of non-contiguous memory
175 buf_ = xbt_malloc(count*type_->size());
176 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
177 type_->serialize(old_buf_, buf_, count);
183 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
185 auto ref = static_cast<MPI_Request>(a);
186 auto req = static_cast<MPI_Request>(b);
187 bool match = match_common(req, req, ref);
188 if (not match || ref->comm_ == MPI_COMM_UNINITIALIZED || ref->comm_->is_smp_comm())
191 if (ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_),
192 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_,
195 ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_),
196 ref->comm_->group()->rank(req->dst_), req->tag_),
197 req->src_, req->dst_, req->tag_);
198 ref->comm_->increment_received_messages_count(ref->comm_->group()->rank(req->src_),
199 ref->comm_->group()->rank(req->dst_), req->tag_);
200 if (ref->real_size_ > req->real_size_) {
201 ref->real_size_ = req->real_size_;
206 req->flags_ &= ~MPI_REQ_MATCHED;
207 ref->detached_sender_ = nullptr;
208 XBT_DEBUG("Refusing to match message, as its ID is not the one I expect. in comm %p, %u != %u, "
209 "from pid %ld to pid %ld, with tag %d",
211 ref->comm_->get_received_messages_count(ref->comm_->group()->rank(req->src_),
212 ref->comm_->group()->rank(req->dst_), req->tag_),
213 req->message_id_, req->src_, req->dst_, req->tag_);
218 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
220 auto ref = static_cast<MPI_Request>(a);
221 auto req = static_cast<MPI_Request>(b);
222 return match_common(req, ref, req);
225 void Request::print_request(const char* message) const
227 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
228 size_, src_, dst_, tag_, flags_);
231 /* factories, to hide the internal flags from the caller */
232 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
234 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
235 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
236 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
239 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
241 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
242 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
243 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
246 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
248 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
249 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
250 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
253 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
255 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
256 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
257 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
260 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
265 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
266 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
267 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
269 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
270 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
271 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
278 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
280 aid_t source = MPI_PROC_NULL;
281 if (src == MPI_ANY_SOURCE)
282 source = MPI_ANY_SOURCE;
283 else if (src != MPI_PROC_NULL)
284 source = comm->group()->actor(src);
285 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
287 simgrid::s4u::this_actor::get_pid(), tag, comm,
288 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
291 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
294 aid_t source = MPI_PROC_NULL;
295 if (src == MPI_ANY_SOURCE)
296 source = MPI_ANY_SOURCE;
297 else if (src != MPI_PROC_NULL)
298 source = comm->group()->actor(src);
301 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
302 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
303 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
306 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
307 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
308 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
313 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
315 aid_t source = MPI_PROC_NULL;
316 if (src == MPI_ANY_SOURCE)
317 source = MPI_ANY_SOURCE;
318 else if (src != MPI_PROC_NULL)
319 source = comm->group()->actor(src);
320 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
321 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
322 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
325 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
327 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
328 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
329 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
330 if(dst != MPI_PROC_NULL)
335 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
337 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
338 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
339 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
340 if(dst != MPI_PROC_NULL)
345 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
347 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
348 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
349 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
350 if(dst != MPI_PROC_NULL)
355 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
357 aid_t source = MPI_PROC_NULL;
358 if (src == MPI_ANY_SOURCE)
359 source = MPI_ANY_SOURCE;
360 else if (src != MPI_PROC_NULL)
361 source = comm->group()->actor(src);
362 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
363 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
364 if(src != MPI_PROC_NULL)
369 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
371 MPI_Request request = irecv(buf, count, datatype, src, tag, comm);
372 int retval = wait(&request,status);
377 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
379 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
380 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
381 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
383 if(dst != MPI_PROC_NULL)
385 wait(&request, MPI_STATUS_IGNORE);
389 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
391 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
392 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
393 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
394 if(dst != MPI_PROC_NULL)
396 wait(&request, MPI_STATUS_IGNORE);
400 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
402 auto request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
403 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
404 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
406 if(dst != MPI_PROC_NULL)
408 wait(&request,MPI_STATUS_IGNORE);
412 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
413 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
414 MPI_Comm comm, MPI_Status * status)
416 aid_t source = MPI_PROC_NULL;
417 if (src == MPI_ANY_SOURCE)
418 source = MPI_ANY_SOURCE;
419 else if (src != MPI_PROC_NULL)
420 source = comm->group()->actor(src);
421 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
423 std::array<MPI_Request, 2> requests;
424 std::array<MPI_Status, 2> stats;
425 if (aid_t myid = simgrid::s4u::this_actor::get_pid(); (destination == myid) && (source == myid)) {
426 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
427 if (status != MPI_STATUS_IGNORE) {
428 status->MPI_SOURCE = source;
429 status->MPI_TAG = recvtag;
430 status->MPI_ERROR = MPI_SUCCESS;
431 status->count = sendcount * sendtype->size();
435 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
436 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
437 startall(2, requests.data());
438 waitall(2, requests.data(), stats.data());
441 if(status != MPI_STATUS_IGNORE) {
442 // Copy receive status
447 void Request::start()
449 s4u::Mailbox* mailbox;
451 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
452 //reinitialize temporary buffer for persistent requests
453 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
455 init_buffer(real_size_/type_->size());
457 flags_ &= ~MPI_REQ_PREPARED;
458 flags_ &= ~MPI_REQ_FINISHED;
461 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
463 if ((flags_ & MPI_REQ_RECV) != 0) {
464 this->print_request("New recv");
466 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
468 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
469 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
472 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
473 flags_ |= MPI_REQ_PROBE;
475 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
476 mailbox = process->mailbox();
477 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
478 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
479 //begin with the more appropriate one : the small one.
480 mailbox = process->mailbox_small();
481 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
482 mailbox->get_cname());
483 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
485 if (action == nullptr) {
486 mailbox = process->mailbox();
487 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
488 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
489 if (action == nullptr) {
490 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
491 mailbox = process->mailbox_small();
494 XBT_DEBUG("yes there was something for us in the small mailbox");
497 mailbox = process->mailbox_small();
498 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
499 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
501 if (action == nullptr) {
502 XBT_DEBUG("No, nothing in the permanent receive mailbox");
503 mailbox = process->mailbox();
505 XBT_DEBUG("yes there was something for us in the small mailbox");
509 flags_ &= ~MPI_REQ_PROBE;
510 kernel::actor::CommIrecvSimcall observer{process->get_actor()->get_impl(),
512 static_cast<unsigned char*>(buf_),
515 process->replaying() ? &smpi_comm_null_copy_buffer_callback
516 : smpi_comm_copy_data_callback,
519 observer.set_tag(tag_);
521 action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::irecv(&observer); },
524 XBT_DEBUG("recv simcall posted");
526 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
528 } else { /* the RECV flag was not set, so this is a send */
529 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
530 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
531 if (TRACE_smpi_view_internals())
532 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
533 this->print_request("New send");
535 message_id_=comm_->get_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
536 comm_->increment_sent_messages_count(comm_->group()->rank(src_), comm_->group()->rank(dst_), tag_);
539 if ((flags_ & MPI_REQ_SSEND) == 0 &&
540 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
541 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
542 void *oldbuf = nullptr;
544 XBT_DEBUG("Send request %p is detached", this);
546 if (not(type_->flags() & DT_FLAG_DERIVED)) {
548 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
549 if (smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_), buf_))
550 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
552 //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
553 //so actually ... don't use manually attached buffer space.
554 buf = xbt_malloc(size_);
555 memcpy(buf,oldbuf,size_);
556 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
561 //if we are giving back the control to the user without waiting for completion, we have to inject timings
562 double sleeptime = 0.0;
563 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
564 // isend and send timings may be different
566 ((flags_ & MPI_REQ_ISEND) != 0)
567 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(
568 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(), simgrid::s4u::Actor::by_pid(dst_)->get_host())
569 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(
570 size_, simgrid::s4u::Actor::by_pid(src_)->get_host(),
571 simgrid::s4u::Actor::by_pid(dst_)->get_host());
575 simgrid::s4u::this_actor::sleep_for(sleeptime);
576 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
579 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
581 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
584 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
585 mailbox = process->mailbox();
586 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
587 bool is_probe = ((flags_ & MPI_REQ_PROBE) != 0);
588 flags_ |= MPI_REQ_PROBE;
590 mailbox = process->mailbox();
591 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
592 if (not mailbox->iprobe(1, &match_send, static_cast<void*>(this))) {
593 if ((flags_ & MPI_REQ_SSEND) == 0) {
594 mailbox = process->mailbox_small();
595 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
596 mailbox->get_cname());
598 mailbox = process->mailbox_small();
599 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
600 mailbox->get_cname());
601 if (not mailbox->iprobe(1, &match_send, static_cast<void*>(this))) {
602 XBT_DEBUG("No, we are first, send to large mailbox");
603 mailbox = process->mailbox();
607 XBT_DEBUG("Yes there was something for us in the large mailbox");
610 flags_ &= ~MPI_REQ_PROBE;
612 mailbox = process->mailbox();
613 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
616 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
617 kernel::actor::CommIsendSimcall observer{
618 simgrid::kernel::EngineImpl::get_instance()->get_actor_by_pid(src_), mailbox->get_impl(),
619 static_cast<double>(payload_size_), -1, static_cast<unsigned char*>(buf), real_size_, &match_send,
620 &xbt_free_f, // how to free the userdata if a detached send fails
621 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
622 // detach if msg size < eager/rdv switch limit
624 observer.set_tag(tag_);
625 action_ = kernel::actor::simcall_answered([&observer] { return kernel::activity::CommImpl::isend(&observer); },
627 XBT_DEBUG("send simcall posted");
629 /* FIXME: detached sends are not traceable (action_ == nullptr) */
630 if (action_ != nullptr) {
631 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
632 smpi_process()->get_tracing_category());
635 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
640 void Request::startall(int count, MPI_Request * requests)
642 if(requests== nullptr)
645 for(int i = 0; i < count; i++) {
646 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
647 requests[i]->start();
651 void Request::cancel()
653 this->flags_ |= MPI_REQ_CANCELLED;
654 if (this->action_ != nullptr)
655 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
658 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
659 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
660 // to avoid deadlocks if used as a break condition, such as
661 // while (MPI_Test(request, flag, status) && flag) dostuff...
662 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
663 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
664 xbt_assert(*request != MPI_REQUEST_NULL);
666 static int nsleeps = 1;
667 int ret = MPI_SUCCESS;
669 if(smpi_test_sleep > 0)
670 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
672 Status::empty(status);
675 if ((*request)->flags_ & MPI_REQ_NBC){
676 *flag = finish_nbc_requests(request, 1);
679 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
680 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
682 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
683 kernel::actor::ActivityTestSimcall observer{issuer, (*request)->action_.get()};
684 *flag = kernel::actor::simcall_answered(
685 [&observer] { return observer.get_activity()->test(observer.get_issuer()); }, &observer);
686 } catch (const Exception&) {
691 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
694 finish_wait(request, status); // may invalidate *request
695 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
696 MPI_Status tmp_status;
697 MPI_Status* mystatus;
698 if (status == MPI_STATUS_IGNORE) {
699 mystatus = &tmp_status;
700 Status::empty(mystatus);
704 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
706 nsleeps=1;//reset the number of sleeps we will do next time
707 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
708 *request = MPI_REQUEST_NULL;
709 } else if (smpi_cfg_grow_injected_times()) {
716 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
722 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
725 for (int i = 0; i < incount; i++) {
726 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
727 if (test(&requests[i], pstat, &flag) != MPI_SUCCESS)
731 if (status != MPI_STATUSES_IGNORE)
732 status[*count] = *pstat;
734 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
735 requests[i] = MPI_REQUEST_NULL;
741 if(count_dead==incount)*count=MPI_UNDEFINED;
743 return MPI_ERR_IN_STATUS;
748 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
750 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
751 comms.reserve(count);
754 int ret = MPI_SUCCESS;
755 *index = MPI_UNDEFINED;
757 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
758 for (int i = 0; i < count; i++) {
759 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
760 comms.push_back(requests[i]->action_.get());
764 if (not map.empty()) {
765 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
766 static int nsleeps = 1;
767 if(smpi_test_sleep > 0)
768 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
771 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
772 kernel::actor::ActivityTestanySimcall observer{issuer, comms};
773 i = kernel::actor::simcall_answered(
775 return kernel::activity::ActivityImpl::test_any(observer.get_issuer(), observer.get_activities());
778 } catch (const Exception&) {
779 XBT_DEBUG("Exception in testany");
783 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
785 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
786 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
789 finish_wait(&requests[*index],status);
790 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
791 MPI_Status tmp_status;
792 MPI_Status* mystatus;
793 if (status == MPI_STATUS_IGNORE) {
794 mystatus = &tmp_status;
795 Status::empty(mystatus);
799 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
802 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
803 *flag = finish_nbc_requests(&requests[*index] , 1);
806 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
807 requests[*index] = MPI_REQUEST_NULL;
808 XBT_DEBUG("Testany - returning with index %d", *index);
816 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
817 //all requests are null or inactive, return true
819 *index = MPI_UNDEFINED;
820 Status::empty(status);
826 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
829 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
833 for(int i=0; i<count; i++){
834 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
835 int ret = test(&requests[i], pstat, &flag);
841 if (ret != MPI_SUCCESS)
844 Status::empty(pstat);
846 if(status != MPI_STATUSES_IGNORE) {
851 return MPI_ERR_IN_STATUS;
856 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
858 //FIXME find another way to avoid busy waiting ?
859 // the issue here is that we have to wait on a nonexistent comm
861 iprobe(source, tag, comm, &flag, status);
862 XBT_DEBUG("Busy Waiting on probing : %d", flag);
866 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
867 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
868 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
869 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
870 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
871 static int nsleeps = 1;
872 double speed = s4u::this_actor::get_host()->get_speed();
873 double maxrate = smpi_cfg_iprobe_cpu_usage();
875 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
876 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
877 if (smpi_iprobe_sleep > 0) {
878 /** Compute the number of flops we will sleep **/
879 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
880 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
882 /* Not the entire CPU can be used when iprobing: This is important for
883 * the energy consumption caused by polling with iprobes.
884 * Note also that the number of flops that was
885 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
887 ->set_bound(maxrate * speed)
891 // behave like a receive, but don't do it
892 s4u::Mailbox* mailbox;
894 request->print_request("New iprobe");
895 // We have to test both mailboxes as we don't know if we will receive one or another
896 if (smpi_cfg_async_small_thresh() > 0) {
897 mailbox = smpi_process()->mailbox_small();
898 XBT_DEBUG("Trying to probe the perm recv mailbox");
899 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
902 if (request->action_ == nullptr){
903 mailbox = smpi_process()->mailbox();
904 XBT_DEBUG("trying to probe the other mailbox");
905 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
908 if (request->action_ != nullptr){
909 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
910 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
912 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
913 status->MPI_SOURCE = comm->group()->rank(req->src_);
914 status->MPI_TAG = req->tag_;
915 status->MPI_ERROR = MPI_SUCCESS;
916 status->count = req->real_size_;
918 nsleeps = 1;//reset the number of sleeps we will do next time
922 if (smpi_cfg_grow_injected_times())
926 xbt_assert(request == MPI_REQUEST_NULL);
929 int Request::finish_nbc_requests(MPI_Request* request, int test){
933 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
935 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
938 xbt_die("Failure when waiting on non blocking collective sub-requests");
940 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
941 for(auto& req: (*request)->nbc_requests_){
942 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
943 void * buf=req->buf_;
944 if((*request)->type_->flags() & DT_FLAG_DERIVED)
946 if(req->flags_ & MPI_REQ_RECV ){
947 if((*request)->op_!=MPI_OP_NULL){
948 int count=(*request)->size_/ (*request)->type_->size();
949 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
951 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
954 if(req!=MPI_REQUEST_NULL)
955 Request::unref(&req);
957 (*request)->nbc_requests_.clear();
962 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
964 MPI_Request req = *request;
965 Status::empty(status);
966 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
967 if (status!=MPI_STATUS_IGNORE)
969 if(req->detached_sender_ != nullptr)
970 unref(&(req->detached_sender_));
975 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
976 if (status != MPI_STATUS_IGNORE) {
977 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
978 Status::empty(status);
979 status->MPI_SOURCE = MPI_PROC_NULL;
981 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
982 status->MPI_SOURCE = req->comm_->group()->rank(src);
983 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
984 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
986 // this handles the case were size in receive differs from size in send
987 status->count = req->real_size_;
989 //detached send will be finished at the other end
990 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
991 req->print_request("Finishing");
992 MPI_Datatype datatype = req->type_;
994 // FIXME Handle the case of a partial shared malloc.
995 if (not smpi_process()->replaying() &&
996 (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))) {
997 if (smpi_switch_data_segment(simgrid::s4u::Actor::self(), req->old_buf_))
998 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
1000 if(datatype->flags() & DT_FLAG_DERIVED){
1001 // This part handles the problem of non-contiguous memory the unserialization at the reception
1002 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
1003 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
1004 xbt_free(req->buf_);
1006 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
1007 if (datatype->size() != 0) {
1008 int n = req->real_size_ / datatype->size();
1009 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
1011 xbt_free(req->buf_);
1018 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
1019 aid_t rank = simgrid::s4u::this_actor::get_pid();
1020 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
1021 TRACE_smpi_recv(src_traced, rank,req->tag_);
1023 if(req->detached_sender_ != nullptr){
1024 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
1025 simgrid::s4u::Host* dst_host = simgrid::s4u::Actor::by_pid(req->dst_)->get_host();
1026 if (double sleeptime = simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(
1027 req->real_size(), req->src_host_, dst_host);
1029 simgrid::s4u::this_actor::sleep_for(sleeptime);
1030 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
1032 unref(&(req->detached_sender_));
1034 if (req->flags_ & MPI_REQ_PERSISTENT)
1035 req->action_ = nullptr;
1036 req->flags_ |= MPI_REQ_FINISHED;
1038 if (req->truncated_ || req->unmatched_types_) {
1039 char error_string[MPI_MAX_ERROR_STRING];
1042 if(req->truncated_ )
1043 errkind = MPI_ERR_TRUNCATE;
1045 errkind = MPI_ERR_TYPE;
1046 PMPI_Error_string(errkind, error_string, &error_size);
1047 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
1048 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
1049 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1050 else if (err == MPI_ERRORS_ARE_FATAL)
1051 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
1053 err->call((req->comm_), errkind);
1054 if (err != MPI_ERRHANDLER_NULL)
1055 simgrid::smpi::Errhandler::unref(err);
1056 MC_assert(not MC_is_active()); /* Only fail in MC mode */
1058 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
1062 int Request::wait(MPI_Request * request, MPI_Status * status)
1064 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
1065 xbt_assert(*request != MPI_REQUEST_NULL);
1067 int ret=MPI_SUCCESS;
1069 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
1070 if (status != MPI_STATUS_IGNORE) {
1071 Status::empty(status);
1072 status->MPI_SOURCE = MPI_PROC_NULL;
1074 (*request)=MPI_REQUEST_NULL;
1078 (*request)->print_request("Waiting");
1079 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1080 Status::empty(status);
1084 if ((*request)->action_ != nullptr){
1086 // this is not a detached send
1087 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1088 kernel::actor::ActivityWaitSimcall observer{issuer, (*request)->action_.get(), -1};
1089 kernel::actor::simcall_blocking([issuer, &observer] { observer.get_activity()->wait_for(issuer, -1); },
1091 } catch (const CancelException&) {
1092 XBT_VERB("Request cancelled");
1096 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1097 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1098 ((*request)->generalized_funcs)->mutex->lock();
1099 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1100 ((*request)->generalized_funcs)->mutex->unlock();
1102 MPI_Status tmp_status;
1103 MPI_Status* mystatus;
1104 if (status == MPI_STATUS_IGNORE) {
1105 mystatus = &tmp_status;
1106 Status::empty(mystatus);
1110 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1113 if ((*request)->truncated_)
1114 ret = MPI_ERR_TRUNCATE;
1116 if ((*request)->flags_ & MPI_REQ_NBC)
1117 finish_nbc_requests(request, 0);
1119 finish_wait(request, status); // may invalidate *request
1120 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1121 *request = MPI_REQUEST_NULL;
1125 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1127 int index = MPI_UNDEFINED;
1130 // Wait for a request to complete
1131 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
1132 std::vector<int> map;
1133 XBT_DEBUG("Wait for one of %d", count);
1134 for(int i = 0; i < count; i++) {
1135 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1136 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1137 if (requests[i]->action_ != nullptr) {
1138 XBT_DEBUG("Waiting any %p ", requests[i]);
1139 comms.push_back(requests[i]->action_.get());
1142 // This is a finished detached request, let's return this one
1143 comms.clear(); // don't do the waitany call afterwards
1145 if (requests[index]->flags_ & MPI_REQ_NBC)
1146 finish_nbc_requests(&requests[index], 0);
1147 finish_wait(&requests[i], status); // cleanup if refcount = 0
1148 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1149 requests[i] = MPI_REQUEST_NULL; // set to null
1154 if (not comms.empty()) {
1155 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1158 kernel::actor::ActorImpl* issuer = kernel::actor::ActorImpl::self();
1159 kernel::actor::ActivityWaitanySimcall observer{issuer, comms, -1};
1160 i = kernel::actor::simcall_blocking(
1162 kernel::activity::ActivityImpl::wait_any_for(observer.get_issuer(), observer.get_activities(),
1163 observer.get_timeout());
1166 } catch (const CancelException&) {
1167 XBT_INFO("request cancelled");
1171 // not MPI_UNDEFINED, as this is a simix return code
1174 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1175 if ((requests[index] == MPI_REQUEST_NULL) ||
1176 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1177 finish_wait(&requests[index],status);
1178 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1179 requests[index] = MPI_REQUEST_NULL;
1186 if (index==MPI_UNDEFINED)
1187 Status::empty(status);
1192 static int sort_accumulates(const Request* a, const Request* b)
1194 return (a->tag() > b->tag());
1197 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1199 std::vector<MPI_Request> accumulates;
1202 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1203 int retvalue = MPI_SUCCESS;
1204 //tag invalid requests in the set
1205 if (status != MPI_STATUSES_IGNORE) {
1206 for (int c = 0; c < count; c++) {
1207 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1208 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1209 Status::empty(&status[c]);
1210 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1211 Status::empty(&status[c]);
1212 status[c].MPI_SOURCE = MPI_PROC_NULL;
1216 for (int c = 0; c < count; c++) {
1217 if (MC_is_active() || MC_record_replay_is_active()) {
1218 wait(&requests[c],pstat);
1221 index = waitany(count, requests, pstat);
1223 if (index == MPI_UNDEFINED)
1226 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1227 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1228 accumulates.push_back(requests[index]);
1229 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1230 requests[index] = MPI_REQUEST_NULL;
1232 if (status != MPI_STATUSES_IGNORE) {
1233 status[index] = *pstat;
1234 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1235 retvalue = MPI_ERR_IN_STATUS;
1239 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1240 for (auto& req : accumulates)
1241 finish_wait(&req, status);
1246 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1252 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1253 index = waitany(incount, requests, pstat);
1254 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1255 if(status != MPI_STATUSES_IGNORE) {
1256 status[count] = *pstat;
1258 indices[count] = index;
1260 for (int i = 0; i < incount; i++) {
1261 if (i != index && requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1262 test(&requests[i], pstat,&flag);
1265 if(status != MPI_STATUSES_IGNORE) {
1266 status[count] = *pstat;
1268 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1269 requests[i]=MPI_REQUEST_NULL;
1277 MPI_Request Request::f2c(int id)
1279 if(id==MPI_FORTRAN_REQUEST_NULL)
1280 return MPI_REQUEST_NULL;
1281 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1284 void Request::free_f(int id)
1286 if (id != MPI_FORTRAN_REQUEST_NULL) {
1287 F2C::lookup()->erase(id);
1291 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1293 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1294 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1298 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1304 if(req != MPI_REQUEST_NULL &&
1305 status != MPI_STATUS_IGNORE) {
1306 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1307 status->MPI_SOURCE = req->comm_->group()->rank(src);
1308 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1309 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1310 status->count = req->real_size_;
1315 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1316 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1318 *request = new Request();
1319 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1320 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1321 (*request)->refcount_ = 1;
1322 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1323 ((*request)->generalized_funcs)->query_fn=query_fn;
1324 ((*request)->generalized_funcs)->free_fn=free_fn;
1325 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1326 ((*request)->generalized_funcs)->extra_state=extra_state;
1327 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1328 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1332 int Request::grequest_complete(MPI_Request request)
1334 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1335 return MPI_ERR_REQUEST;
1336 request->generalized_funcs->mutex->lock();
1337 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1338 request->generalized_funcs->cond->notify_one();
1339 request->generalized_funcs->mutex->unlock();
1343 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1344 if (not reqs.empty()) {
1345 nbc_requests_ = reqs;
1346 Request::startall(reqs.size(), reqs.data());
1350 std::vector<MPI_Request> Request::get_nbc_requests() const
1352 return nbc_requests_;
1354 } // namespace simgrid::smpi