1 /* Copyright (c) 2007-2021. 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/Exec.hpp"
12 #include "simgrid/s4u/Mutex.hpp"
13 #include "simgrid/s4u/ConditionVariable.hpp"
14 #include "smpi_comm.hpp"
15 #include "smpi_datatype.hpp"
16 #include "smpi_host.hpp"
17 #include "smpi_op.hpp"
18 #include "src/kernel/activity/CommImpl.hpp"
19 #include "src/mc/mc_replay.hpp"
20 #include "src/smpi/include/smpi_actor.hpp"
25 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (request)");
27 static simgrid::config::Flag<double> smpi_iprobe_sleep(
28 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
29 static simgrid::config::Flag<double> smpi_test_sleep(
30 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
32 std::vector<s_smpi_factor_t> smpi_ois_values;
34 extern void (*smpi_comm_copy_data_callback)(simgrid::kernel::activity::CommImpl*, void*, size_t);
39 Request::Request(const void* buf, int count, MPI_Datatype datatype, aid_t src, aid_t dst, int tag, MPI_Comm comm,
40 unsigned flags, MPI_Op op)
41 : buf_(const_cast<void*>(buf))
43 , size_(datatype->size() * count)
53 if(op != MPI_REPLACE && op != MPI_OP_NULL)
57 detached_sender_ = nullptr;
62 if (flags & MPI_REQ_PERSISTENT)
74 void Request::unref(MPI_Request* request)
76 xbt_assert(*request != MPI_REQUEST_NULL, "freeing an already free request");
78 (*request)->refcount_--;
79 if ((*request)->refcount_ < 0) {
80 (*request)->print_request("wrong refcount");
81 xbt_die("Whoops, wrong refcount");
83 if ((*request)->refcount_ == 0) {
84 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
85 ((*request)->generalized_funcs)->free_fn(((*request)->generalized_funcs)->extra_state);
87 Comm::unref((*request)->comm_);
88 Datatype::unref((*request)->type_);
90 if ((*request)->op_ != MPI_REPLACE && (*request)->op_ != MPI_OP_NULL)
91 Op::unref(&(*request)->op_);
93 (*request)->print_request("Destroying");
94 F2C::free_f((*request)->f2c_id());
96 *request = MPI_REQUEST_NULL;
98 (*request)->print_request("Decrementing");
102 bool Request::match_common(MPI_Request req, MPI_Request sender, MPI_Request receiver)
104 xbt_assert(sender, "Cannot match against null sender");
105 xbt_assert(receiver, "Cannot match against null receiver");
106 XBT_DEBUG("Trying to match %s of sender src %ld against %ld, tag %d against %d, id %d against %d",
107 (req == receiver ? "send" : "recv"), sender->src_, receiver->src_, sender->tag_, receiver->tag_,
108 sender->comm_->id(), receiver->comm_->id());
110 if ((receiver->comm_->id() == MPI_UNDEFINED || sender->comm_->id() == MPI_UNDEFINED ||
111 receiver->comm_->id() == sender->comm_->id()) &&
112 ((receiver->src_ == MPI_ANY_SOURCE && (receiver->comm_->group()->rank(sender->src_) != MPI_UNDEFINED)) ||
113 receiver->src_ == sender->src_) &&
114 ((receiver->tag_ == MPI_ANY_TAG && sender->tag_ >= 0) || receiver->tag_ == sender->tag_)) {
115 // we match, we can transfer some values
116 if (receiver->src_ == MPI_ANY_SOURCE)
117 receiver->real_src_ = sender->src_;
118 if (receiver->tag_ == MPI_ANY_TAG)
119 receiver->real_tag_ = sender->tag_;
120 if ((receiver->flags_ & MPI_REQ_PROBE) == 0 ){
121 if (receiver->real_size_ < sender->real_size_){
122 XBT_DEBUG("Truncating message - should not happen: receiver size : %zu < sender size : %zu", receiver->real_size_, sender->real_size_);
123 receiver->truncated_ = true;
124 } else if (receiver->real_size_ > sender->real_size_){
125 receiver->real_size_=sender->real_size_;
128 if (sender->detached_)
129 receiver->detached_sender_ = sender; // tie the sender to the receiver, as it is detached and has to be freed in
131 req->flags_ |= MPI_REQ_MATCHED; // mark as impossible to cancel anymore
132 XBT_DEBUG("match succeeded");
138 void Request::init_buffer(int count){
139 void *old_buf = nullptr;
140 // FIXME Handle the case of a partial shared malloc.
141 // This part handles the problem of non-contiguous memory (for the unserialization at the reception)
142 if ((((flags_ & MPI_REQ_RECV) != 0) && ((flags_ & MPI_REQ_ACCUMULATE) != 0)) || (type_->flags() & DT_FLAG_DERIVED)) {
143 // This part handles the problem of non-contiguous memory
148 buf_ = xbt_malloc(count*type_->size());
149 if ((type_->flags() & DT_FLAG_DERIVED) && ((flags_ & MPI_REQ_SEND) != 0)) {
150 type_->serialize(old_buf, buf_, count);
157 bool Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl*)
159 auto ref = static_cast<MPI_Request>(a);
160 auto req = static_cast<MPI_Request>(b);
161 return match_common(req, req, ref);
164 bool Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl*)
166 auto ref = static_cast<MPI_Request>(a);
167 auto req = static_cast<MPI_Request>(b);
168 return match_common(req, ref, req);
171 void Request::print_request(const char* message) const
173 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %ld, dst = %ld, tag = %d, flags = %x]", message, this, buf_,
174 size_, src_, dst_, tag_, flags_);
177 /* factories, to hide the internal flags from the caller */
178 MPI_Request Request::bsend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
180 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
181 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
182 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED | MPI_REQ_BSEND);
185 MPI_Request Request::send_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
187 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
188 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
189 MPI_REQ_PERSISTENT | MPI_REQ_SEND | MPI_REQ_PREPARED);
192 MPI_Request Request::ssend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
194 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
195 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
196 MPI_REQ_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
199 MPI_Request Request::isend_init(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
201 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
202 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
203 MPI_REQ_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
206 MPI_Request Request::rma_send_init(const void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
209 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
211 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
212 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
213 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED);
215 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, comm->group()->actor(src),
216 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
217 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_PREPARED |
224 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
226 aid_t source = MPI_PROC_NULL;
227 if (src == MPI_ANY_SOURCE)
228 source = MPI_ANY_SOURCE;
229 else if (src != MPI_PROC_NULL)
230 source = comm->group()->actor(src);
231 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
233 simgrid::s4u::this_actor::get_pid(), tag, comm,
234 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
237 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
240 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
241 aid_t source = MPI_PROC_NULL;
242 if (src == MPI_ANY_SOURCE)
243 source = MPI_ANY_SOURCE;
244 else if (src != MPI_PROC_NULL)
245 source = comm->group()->actor(src);
247 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
248 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
249 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
252 new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, source,
253 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
254 MPI_REQ_RMA | MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED | MPI_REQ_ACCUMULATE, op);
259 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
261 aid_t source = MPI_PROC_NULL;
262 if (src == MPI_ANY_SOURCE)
263 source = MPI_ANY_SOURCE;
264 else if (src != MPI_PROC_NULL)
265 source = comm->group()->actor(src);
266 return new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
267 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
268 MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PREPARED);
271 MPI_Request Request::ibsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
273 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
274 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
275 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
276 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND | MPI_REQ_BSEND);
277 if(dst != MPI_PROC_NULL)
282 MPI_Request Request::isend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
284 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
285 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
286 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
287 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SEND);
288 if(dst != MPI_PROC_NULL)
293 MPI_Request Request::issend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
295 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
296 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
297 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
298 MPI_REQ_NON_PERSISTENT | MPI_REQ_ISEND | MPI_REQ_SSEND | MPI_REQ_SEND);
299 if(dst != MPI_PROC_NULL)
304 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
306 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
307 aid_t source = MPI_PROC_NULL;
308 if (src == MPI_ANY_SOURCE)
309 source = MPI_ANY_SOURCE;
310 else if (src != MPI_PROC_NULL)
311 source = comm->group()->actor(src);
312 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype,
313 source, simgrid::s4u::this_actor::get_pid(), tag, comm,
314 MPI_REQ_NON_PERSISTENT | MPI_REQ_RECV);
315 if(src != MPI_PROC_NULL)
320 int Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
322 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
323 request = irecv(buf, count, datatype, src, tag, comm);
324 int retval = wait(&request,status);
329 void Request::bsend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
331 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
332 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
333 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
334 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND | MPI_REQ_BSEND);
336 if(dst != MPI_PROC_NULL)
338 wait(&request, MPI_STATUS_IGNORE);
342 void Request::send(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
344 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
345 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
346 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
347 MPI_REQ_NON_PERSISTENT | MPI_REQ_SEND);
348 if(dst != MPI_PROC_NULL)
350 wait(&request, MPI_STATUS_IGNORE);
354 void Request::ssend(const void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
356 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
357 request = new Request(buf == MPI_BOTTOM ? nullptr : buf, count, datatype, simgrid::s4u::this_actor::get_pid(),
358 dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL, tag, comm,
359 MPI_REQ_NON_PERSISTENT | MPI_REQ_SSEND | MPI_REQ_SEND);
361 if(dst != MPI_PROC_NULL)
363 wait(&request,MPI_STATUS_IGNORE);
367 void Request::sendrecv(const void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
368 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
369 MPI_Comm comm, MPI_Status * status)
371 aid_t source = MPI_PROC_NULL;
372 if (src == MPI_ANY_SOURCE)
373 source = MPI_ANY_SOURCE;
374 else if (src != MPI_PROC_NULL)
375 source = comm->group()->actor(src);
376 aid_t destination = dst != MPI_PROC_NULL ? comm->group()->actor(dst) : MPI_PROC_NULL;
378 std::array<MPI_Request, 2> requests;
379 std::array<MPI_Status, 2> stats;
380 aid_t myid = simgrid::s4u::this_actor::get_pid();
381 if ((destination == myid) && (source == myid)) {
382 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
383 if (status != MPI_STATUS_IGNORE) {
384 status->MPI_SOURCE = source;
385 status->MPI_TAG = recvtag;
386 status->MPI_ERROR = MPI_SUCCESS;
387 status->count = sendcount * sendtype->size();
391 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
392 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
393 startall(2, requests.data());
394 waitall(2, requests.data(), stats.data());
397 if(status != MPI_STATUS_IGNORE) {
398 // Copy receive status
403 void Request::start()
405 s4u::Mailbox* mailbox;
407 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
408 //reinitialize temporary buffer for persistent requests
409 if(real_size_ > 0 && flags_ & MPI_REQ_FINISHED){
411 init_buffer(real_size_/type_->size());
413 flags_ &= ~MPI_REQ_PREPARED;
414 flags_ &= ~MPI_REQ_FINISHED;
417 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
419 if ((flags_ & MPI_REQ_RECV) != 0) {
420 this->print_request("New recv");
422 simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
424 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
425 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
428 if (smpi_cfg_async_small_thresh() == 0 && (flags_ & MPI_REQ_RMA) == 0) {
429 mailbox = process->mailbox();
430 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) {
431 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
432 //begin with the more appropriate one : the small one.
433 mailbox = process->mailbox_small();
434 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %s (in case of SSEND)?",
435 mailbox->get_cname());
436 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
438 if (action == nullptr) {
439 mailbox = process->mailbox();
440 XBT_DEBUG("No, nothing in the small mailbox test the other one : %s", mailbox->get_cname());
441 action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
442 if (action == nullptr) {
443 XBT_DEBUG("Still nothing, switch back to the small mailbox : %s", mailbox->get_cname());
444 mailbox = process->mailbox_small();
447 XBT_DEBUG("yes there was something for us in the large mailbox");
450 mailbox = process->mailbox_small();
451 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
452 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(0, &match_recv, static_cast<void*>(this));
454 if (action == nullptr) {
455 XBT_DEBUG("No, nothing in the permanent receive mailbox");
456 mailbox = process->mailbox();
458 XBT_DEBUG("yes there was something for us in the small mailbox");
462 action_ = simcall_comm_irecv(
463 process->get_actor()->get_impl(), mailbox->get_impl(), buf_, &real_size_, &match_recv,
464 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
465 XBT_DEBUG("recv simcall posted");
467 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
469 } else { /* the RECV flag was not set, so this is a send */
470 const simgrid::smpi::ActorExt* process = smpi_process_remote(simgrid::s4u::Actor::by_pid(dst_));
471 xbt_assert(process, "Actor pid=%ld is gone??", dst_);
472 if (TRACE_smpi_view_internals())
473 TRACE_smpi_send(src_, src_, dst_, tag_, size_);
474 this->print_request("New send");
477 if ((flags_ & MPI_REQ_SSEND) == 0 &&
478 ((flags_ & MPI_REQ_RMA) != 0 || (flags_ & MPI_REQ_BSEND) != 0 ||
479 static_cast<int>(size_) < smpi_cfg_detached_send_thresh())) {
480 void *oldbuf = nullptr;
482 XBT_DEBUG("Send request %p is detached", this);
484 if (not(type_->flags() & DT_FLAG_DERIVED)) {
486 if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
487 if ((smpi_cfg_privatization() != SmpiPrivStrategies::NONE) &&
488 (static_cast<char*>(buf_) >= smpi_data_exe_start) &&
489 (static_cast<char*>(buf_) < smpi_data_exe_start + smpi_data_exe_size)) {
490 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
491 smpi_switch_data_segment(simgrid::s4u::Actor::by_pid(src_));
493 //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
494 //so actually ... don't use manually attached buffer space.
495 buf = xbt_malloc(size_);
496 memcpy(buf,oldbuf,size_);
497 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
502 //if we are giving back the control to the user without waiting for completion, we have to inject timings
503 double sleeptime = 0.0;
504 if (detached_ || ((flags_ & (MPI_REQ_ISEND | MPI_REQ_SSEND)) != 0)) { // issend should be treated as isend
505 // isend and send timings may be different
506 sleeptime = ((flags_ & MPI_REQ_ISEND) != 0)
507 ? simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->oisend(size_)
508 : simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->osend(size_);
512 simgrid::s4u::this_actor::sleep_for(sleeptime);
513 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
516 simgrid::s4u::MutexPtr mut = process->mailboxes_mutex();
518 if (smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)
521 if (not(smpi_cfg_async_small_thresh() != 0 || (flags_ & MPI_REQ_RMA) != 0)) {
522 mailbox = process->mailbox();
523 } else if (((flags_ & MPI_REQ_RMA) != 0) || static_cast<int>(size_) < smpi_cfg_async_small_thresh()) { // eager mode
524 mailbox = process->mailbox();
525 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %s?", mailbox->get_cname());
526 simgrid::kernel::activity::ActivityImplPtr action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
527 if (action == nullptr) {
528 if ((flags_ & MPI_REQ_SSEND) == 0) {
529 mailbox = process->mailbox_small();
530 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %s",
531 mailbox->get_cname());
533 mailbox = process->mailbox_small();
534 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %s?",
535 mailbox->get_cname());
536 action = mailbox->iprobe(1, &match_send, static_cast<void*>(this));
537 if (action == nullptr) {
538 XBT_DEBUG("No, we are first, send to large mailbox");
539 mailbox = process->mailbox();
543 XBT_DEBUG("Yes there was something for us in the large mailbox");
546 mailbox = process->mailbox();
547 XBT_DEBUG("Send request %p is in the large mailbox %s (buf: %p)", this, mailbox->get_cname(), buf_);
550 size_t payload_size_ = size_ + 16;//MPI enveloppe size (tag+dest+communicator)
551 action_ = simcall_comm_isend(
552 simgrid::kernel::actor::ActorImpl::by_pid(src_), mailbox->get_impl(), payload_size_, -1.0, buf, real_size_,
554 &xbt_free_f, // how to free the userdata if a detached send fails
555 process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this,
556 // detach if msg size < eager/rdv switch limit
558 XBT_DEBUG("send simcall posted");
560 /* FIXME: detached sends are not traceable (action_ == nullptr) */
561 if (action_ != nullptr) {
562 boost::static_pointer_cast<kernel::activity::CommImpl>(action_)->set_tracing_category(
563 smpi_process()->get_tracing_category());
566 if (smpi_cfg_async_small_thresh() != 0 || ((flags_ & MPI_REQ_RMA) != 0))
571 void Request::startall(int count, MPI_Request * requests)
573 if(requests== nullptr)
576 for(int i = 0; i < count; i++) {
577 if(requests[i]->src_ != MPI_PROC_NULL && requests[i]->dst_ != MPI_PROC_NULL)
578 requests[i]->start();
582 void Request::cancel()
584 this->flags_ |= MPI_REQ_CANCELLED;
585 if (this->action_ != nullptr)
586 (boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(this->action_))->cancel();
589 int Request::test(MPI_Request * request, MPI_Status * status, int* flag) {
590 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
591 // to avoid deadlocks if used as a break condition, such as
592 // while (MPI_Test(request, flag, status) && flag) dostuff...
593 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
594 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
595 xbt_assert(*request != MPI_REQUEST_NULL);
597 static int nsleeps = 1;
598 int ret = MPI_SUCCESS;
600 if(smpi_test_sleep > 0)
601 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
603 Status::empty(status);
606 if ((*request)->flags_ & MPI_REQ_NBC){
607 *flag = finish_nbc_requests(request, 1);
610 if (((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) == 0) {
611 if ((*request)->action_ != nullptr && ((*request)->flags_ & MPI_REQ_CANCELLED) == 0){
613 *flag = simcall_comm_test((*request)->action_.get());
614 } catch (const Exception&) {
619 if (((*request)->flags_ & MPI_REQ_GENERALIZED) && not((*request)->flags_ & MPI_REQ_COMPLETE))
622 finish_wait(request, status); // may invalidate *request
623 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_GENERALIZED)){
624 MPI_Status tmp_status;
625 MPI_Status* mystatus;
626 if (status == MPI_STATUS_IGNORE) {
627 mystatus = &tmp_status;
628 Status::empty(mystatus);
632 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
634 nsleeps=1;//reset the number of sleeps we will do next time
635 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & MPI_REQ_PERSISTENT) == 0)
636 *request = MPI_REQUEST_NULL;
637 } else if (smpi_cfg_grow_injected_times()) {
644 int Request::testsome(int incount, MPI_Request requests[], int *count, int *indices, MPI_Status status[])
650 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
653 for (int i = 0; i < incount; i++) {
654 if (requests[i] != MPI_REQUEST_NULL && not (requests[i]->flags_ & MPI_REQ_FINISHED)) {
655 int ret = test(&requests[i], pstat, &flag);
660 if (status != MPI_STATUSES_IGNORE)
661 status[*count] = *pstat;
663 if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
664 requests[i] = MPI_REQUEST_NULL;
670 if(count_dead==incount)*count=MPI_UNDEFINED;
672 return MPI_ERR_IN_STATUS;
677 int Request::testany(int count, MPI_Request requests[], int *index, int* flag, MPI_Status * status)
679 std::vector<simgrid::kernel::activity::CommImpl*> comms;
680 comms.reserve(count);
684 int ret = MPI_SUCCESS;
685 *index = MPI_UNDEFINED;
687 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
688 for(i = 0; i < count; i++) {
689 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
690 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
694 if (not map.empty()) {
695 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
696 static int nsleeps = 1;
697 if(smpi_test_sleep > 0)
698 simgrid::s4u::this_actor::sleep_for(nsleeps * smpi_test_sleep);
700 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
701 } catch (const Exception&) {
702 XBT_DEBUG("Exception in testany");
706 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
708 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED) &&
709 not(requests[*index]->flags_ & MPI_REQ_COMPLETE)) {
712 finish_wait(&requests[*index],status);
713 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_GENERALIZED)){
714 MPI_Status tmp_status;
715 MPI_Status* mystatus;
716 if (status == MPI_STATUS_IGNORE) {
717 mystatus = &tmp_status;
718 Status::empty(mystatus);
722 ret=(requests[*index]->generalized_funcs)->query_fn((requests[*index]->generalized_funcs)->extra_state, mystatus);
725 if (requests[*index] != MPI_REQUEST_NULL && requests[*index]->flags_ & MPI_REQ_NBC){
726 *flag = finish_nbc_requests(&requests[*index] , 1);
729 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & MPI_REQ_NON_PERSISTENT))
730 requests[*index] = MPI_REQUEST_NULL;
731 XBT_DEBUG("Testany - returning with index %d", *index);
739 XBT_DEBUG("Testany on inactive handles, returning flag=1 but empty status");
740 //all requests are null or inactive, return true
742 *index = MPI_UNDEFINED;
743 Status::empty(status);
749 int Request::testall(int count, MPI_Request requests[], int* outflag, MPI_Status status[])
752 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
756 for(int i=0; i<count; i++){
757 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED)) {
758 int ret = test(&requests[i], pstat, &flag);
764 if (ret != MPI_SUCCESS)
767 Status::empty(pstat);
769 if(status != MPI_STATUSES_IGNORE) {
774 return MPI_ERR_IN_STATUS;
779 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
781 //FIXME find another way to avoid busy waiting ?
782 // the issue here is that we have to wait on a nonexistent comm
784 iprobe(source, tag, comm, &flag, status);
785 XBT_DEBUG("Busy Waiting on probing : %d", flag);
789 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
790 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
791 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
792 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
793 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
794 static int nsleeps = 1;
795 double speed = s4u::this_actor::get_host()->get_speed();
796 double maxrate = smpi_cfg_iprobe_cpu_usage();
798 new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->actor(source),
799 simgrid::s4u::this_actor::get_pid(), tag, comm, MPI_REQ_PERSISTENT | MPI_REQ_RECV | MPI_REQ_PROBE);
800 if (smpi_iprobe_sleep > 0) {
801 /** Compute the number of flops we will sleep **/
802 s4u::this_actor::exec_init(/*nsleeps: See comment above */ nsleeps *
803 /*(seconds * flop/s -> total flops)*/ smpi_iprobe_sleep * speed * maxrate)
805 /* Not the entire CPU can be used when iprobing: This is important for
806 * the energy consumption caused by polling with iprobes.
807 * Note also that the number of flops that was
808 * computed above contains a maxrate factor and is hence reduced (maxrate < 1)
810 ->set_bound(maxrate*speed)
814 // behave like a receive, but don't do it
815 s4u::Mailbox* mailbox;
817 request->print_request("New iprobe");
818 // We have to test both mailboxes as we don't know if we will receive one or another
819 if (smpi_cfg_async_small_thresh() > 0) {
820 mailbox = smpi_process()->mailbox_small();
821 XBT_DEBUG("Trying to probe the perm recv mailbox");
822 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
825 if (request->action_ == nullptr){
826 mailbox = smpi_process()->mailbox();
827 XBT_DEBUG("trying to probe the other mailbox");
828 request->action_ = mailbox->iprobe(0, &match_recv, static_cast<void*>(request));
831 if (request->action_ != nullptr){
832 kernel::activity::CommImplPtr sync_comm = boost::static_pointer_cast<kernel::activity::CommImpl>(request->action_);
833 const Request* req = static_cast<MPI_Request>(sync_comm->src_data_);
835 if (status != MPI_STATUS_IGNORE && (req->flags_ & MPI_REQ_PREPARED) == 0) {
836 status->MPI_SOURCE = comm->group()->rank(req->src_);
837 status->MPI_TAG = req->tag_;
838 status->MPI_ERROR = MPI_SUCCESS;
839 status->count = req->real_size_;
841 nsleeps = 1;//reset the number of sleeps we will do next time
845 if (smpi_cfg_grow_injected_times())
849 xbt_assert(request == MPI_REQUEST_NULL);
852 int Request::finish_nbc_requests(MPI_Request* request, int test){
856 ret = waitall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), MPI_STATUSES_IGNORE);
858 ret = testall((*request)->nbc_requests_.size(), (*request)->nbc_requests_.data(), &flag, MPI_STATUSES_IGNORE);
861 xbt_die("Failure when waiting on non blocking collective sub-requests");
863 XBT_DEBUG("Finishing non blocking collective request with %zu sub-requests", (*request)->nbc_requests_.size());
864 for(auto& req: (*request)->nbc_requests_){
865 if((*request)->buf_!=nullptr && req!=MPI_REQUEST_NULL){//reduce case
866 void * buf=req->buf_;
867 if((*request)->type_->flags() & DT_FLAG_DERIVED)
869 if(req->flags_ & MPI_REQ_RECV ){
870 if((*request)->op_!=MPI_OP_NULL){
871 int count=(*request)->size_/ (*request)->type_->size();
872 (*request)->op_->apply(buf, (*request)->buf_, &count, (*request)->type_);
874 smpi_free_tmp_buffer(static_cast<unsigned char*>(buf));
877 if(req!=MPI_REQUEST_NULL)
878 Request::unref(&req);
880 (*request)->nbc_requests_.clear();
885 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
887 MPI_Request req = *request;
888 Status::empty(status);
889 if((req->flags_ & MPI_REQ_CANCELLED) != 0 && (req->flags_ & MPI_REQ_MATCHED) == 0) {
890 if (status!=MPI_STATUS_IGNORE)
892 if(req->detached_sender_ != nullptr)
893 unref(&(req->detached_sender_));
898 if ((req->flags_ & (MPI_REQ_PREPARED | MPI_REQ_GENERALIZED | MPI_REQ_FINISHED)) == 0) {
899 if (status != MPI_STATUS_IGNORE) {
900 if (req->src_== MPI_PROC_NULL || req->dst_== MPI_PROC_NULL){
901 Status::empty(status);
902 status->MPI_SOURCE = MPI_PROC_NULL;
904 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
905 status->MPI_SOURCE = req->comm_->group()->rank(src);
906 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
907 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
909 // this handles the case were size in receive differs from size in send
910 status->count = req->real_size_;
912 //detached send will be finished at the other end
913 if (not(req->detached_ && ((req->flags_ & MPI_REQ_SEND) != 0))) {
914 req->print_request("Finishing");
915 MPI_Datatype datatype = req->type_;
917 // FIXME Handle the case of a partial shared malloc.
918 if (((req->flags_ & MPI_REQ_ACCUMULATE) != 0) ||
919 (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
920 if (not smpi_process()->replaying() && smpi_cfg_privatization() != SmpiPrivStrategies::NONE &&
921 static_cast<char*>(req->old_buf_) >= smpi_data_exe_start &&
922 static_cast<char*>(req->old_buf_) < smpi_data_exe_start + smpi_data_exe_size) {
923 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
924 smpi_switch_data_segment(simgrid::s4u::Actor::self());
927 if(datatype->flags() & DT_FLAG_DERIVED){
928 // This part handles the problem of non-contiguous memory the unserialization at the reception
929 if ((req->flags_ & MPI_REQ_RECV) && datatype->size() != 0)
930 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
933 } else if (req->flags_ & MPI_REQ_RECV) { // apply op on contiguous buffer for accumulate
934 if (datatype->size() != 0) {
935 int n = req->real_size_ / datatype->size();
936 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
945 if (TRACE_smpi_view_internals() && ((req->flags_ & MPI_REQ_RECV) != 0)) {
946 aid_t rank = simgrid::s4u::this_actor::get_pid();
947 aid_t src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
948 TRACE_smpi_recv(src_traced, rank,req->tag_);
950 if(req->detached_sender_ != nullptr){
951 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
953 simgrid::s4u::Actor::self()->get_host()->extension<simgrid::smpi::Host>()->orecv(req->real_size());
954 if (sleeptime > 0.0) {
955 simgrid::s4u::this_actor::sleep_for(sleeptime);
956 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
958 unref(&(req->detached_sender_));
960 if (req->flags_ & MPI_REQ_PERSISTENT)
961 req->action_ = nullptr;
962 req->flags_ |= MPI_REQ_FINISHED;
964 if (req->truncated_) {
965 char error_string[MPI_MAX_ERROR_STRING];
967 PMPI_Error_string(MPI_ERR_TRUNCATE, error_string, &error_size);
968 MPI_Errhandler err = (req->comm_) ? (req->comm_)->errhandler() : MPI_ERRHANDLER_NULL;
969 if (err == MPI_ERRHANDLER_NULL || err == MPI_ERRORS_RETURN)
970 XBT_WARN("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
971 else if (err == MPI_ERRORS_ARE_FATAL)
972 xbt_die("recv - returned %.*s instead of MPI_SUCCESS", error_size, error_string);
974 err->call((req->comm_), MPI_ERR_TRUNCATE);
975 if (err != MPI_ERRHANDLER_NULL)
976 simgrid::smpi::Errhandler::unref(err);
977 MC_assert(not MC_is_active()); /* Only fail in MC mode */
979 if(req->src_ != MPI_PROC_NULL && req->dst_ != MPI_PROC_NULL)
983 int Request::wait(MPI_Request * request, MPI_Status * status)
985 // assume that *request is not MPI_REQUEST_NULL (filtered in PMPI_Wait before)
986 xbt_assert(*request != MPI_REQUEST_NULL);
990 if((*request)->src_ == MPI_PROC_NULL || (*request)->dst_ == MPI_PROC_NULL){
991 if (status != MPI_STATUS_IGNORE) {
992 Status::empty(status);
993 status->MPI_SOURCE = MPI_PROC_NULL;
995 (*request)=MPI_REQUEST_NULL;
999 (*request)->print_request("Waiting");
1000 if ((*request)->flags_ & (MPI_REQ_PREPARED | MPI_REQ_FINISHED)) {
1001 Status::empty(status);
1005 if ((*request)->action_ != nullptr){
1007 // this is not a detached send
1008 simcall_comm_wait((*request)->action_.get(), -1.0);
1009 } catch (const CancelException&) {
1010 XBT_VERB("Request cancelled");
1014 if ((*request)->flags_ & MPI_REQ_GENERALIZED) {
1015 if (not((*request)->flags_ & MPI_REQ_COMPLETE)) {
1016 ((*request)->generalized_funcs)->mutex->lock();
1017 ((*request)->generalized_funcs)->cond->wait(((*request)->generalized_funcs)->mutex);
1018 ((*request)->generalized_funcs)->mutex->unlock();
1020 MPI_Status tmp_status;
1021 MPI_Status* mystatus;
1022 if (status == MPI_STATUS_IGNORE) {
1023 mystatus = &tmp_status;
1024 Status::empty(mystatus);
1028 ret = ((*request)->generalized_funcs)->query_fn(((*request)->generalized_funcs)->extra_state, mystatus);
1031 if ((*request)->truncated_)
1032 ret = MPI_ERR_TRUNCATE;
1034 if ((*request)->flags_ & MPI_REQ_NBC)
1035 finish_nbc_requests(request, 0);
1037 finish_wait(request, status); // may invalidate *request
1038 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & MPI_REQ_NON_PERSISTENT) != 0))
1039 *request = MPI_REQUEST_NULL;
1043 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
1045 int index = MPI_UNDEFINED;
1048 // Wait for a request to complete
1049 std::vector<simgrid::kernel::activity::CommImpl*> comms;
1050 std::vector<int> map;
1051 XBT_DEBUG("Wait for one of %d", count);
1052 for(int i = 0; i < count; i++) {
1053 if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & MPI_REQ_PREPARED) &&
1054 not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1055 if (requests[i]->action_ != nullptr) {
1056 XBT_DEBUG("Waiting any %p ", requests[i]);
1057 comms.push_back(static_cast<simgrid::kernel::activity::CommImpl*>(requests[i]->action_.get()));
1060 // This is a finished detached request, let's return this one
1061 comms.clear(); // don't do the waitany call afterwards
1063 if (requests[index] != MPI_REQUEST_NULL && (requests[index])->flags_ & MPI_REQ_NBC)
1064 finish_nbc_requests(&requests[index], 0);
1065 finish_wait(&requests[i], status); // cleanup if refcount = 0
1066 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1067 requests[i] = MPI_REQUEST_NULL; // set to null
1072 if (not comms.empty()) {
1073 XBT_DEBUG("Enter waitany for %zu comms", comms.size());
1076 i = simcall_comm_waitany(comms.data(), comms.size(), -1);
1077 } catch (const CancelException&) {
1078 XBT_INFO("request cancelled");
1082 // not MPI_UNDEFINED, as this is a simix return code
1085 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
1086 if ((requests[index] == MPI_REQUEST_NULL) ||
1087 (not((requests[index]->flags_ & MPI_REQ_ACCUMULATE) && (requests[index]->flags_ & MPI_REQ_RECV)))) {
1088 finish_wait(&requests[index],status);
1089 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1090 requests[index] = MPI_REQUEST_NULL;
1097 if (index==MPI_UNDEFINED)
1098 Status::empty(status);
1103 static int sort_accumulates(const Request* a, const Request* b)
1105 return (a->tag() > b->tag());
1108 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
1110 std::vector<MPI_Request> accumulates;
1113 MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
1114 int retvalue = MPI_SUCCESS;
1115 //tag invalid requests in the set
1116 if (status != MPI_STATUSES_IGNORE) {
1117 for (int c = 0; c < count; c++) {
1118 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL ||
1119 (requests[c]->flags_ & MPI_REQ_PREPARED)) {
1120 Status::empty(&status[c]);
1121 } else if (requests[c]->src_ == MPI_PROC_NULL) {
1122 Status::empty(&status[c]);
1123 status[c].MPI_SOURCE = MPI_PROC_NULL;
1127 for (int c = 0; c < count; c++) {
1128 if (MC_is_active() || MC_record_replay_is_active()) {
1129 wait(&requests[c],pstat);
1132 index = waitany(count, requests, pstat);
1134 if (index == MPI_UNDEFINED)
1137 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_RECV) &&
1138 (requests[index]->flags_ & MPI_REQ_ACCUMULATE))
1139 accumulates.push_back(requests[index]);
1140 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & MPI_REQ_NON_PERSISTENT))
1141 requests[index] = MPI_REQUEST_NULL;
1143 if (status != MPI_STATUSES_IGNORE) {
1144 status[index] = *pstat;
1145 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
1146 retvalue = MPI_ERR_IN_STATUS;
1150 if (not accumulates.empty()) {
1151 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
1152 for (auto& req : accumulates) {
1153 finish_wait(&req, status);
1160 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
1166 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
1167 index = waitany(incount, requests, pstat);
1168 if(index==MPI_UNDEFINED) return MPI_UNDEFINED;
1169 if(status != MPI_STATUSES_IGNORE) {
1170 status[count] = *pstat;
1172 indices[count] = index;
1174 for (int i = 0; i < incount; i++) {
1175 if (i!=index && requests[i] != MPI_REQUEST_NULL
1176 && not(requests[i]->flags_ & MPI_REQ_FINISHED)) {
1177 test(&requests[i], pstat,&flag);
1180 if(status != MPI_STATUSES_IGNORE) {
1181 status[count] = *pstat;
1183 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & MPI_REQ_NON_PERSISTENT))
1184 requests[i]=MPI_REQUEST_NULL;
1192 MPI_Request Request::f2c(int id)
1194 if(id==MPI_FORTRAN_REQUEST_NULL)
1195 return MPI_REQUEST_NULL;
1196 return static_cast<MPI_Request>(F2C::lookup()->at(id));
1199 void Request::free_f(int id)
1201 if (id != MPI_FORTRAN_REQUEST_NULL) {
1202 F2C::lookup()->erase(id);
1206 int Request::get_status(const Request* req, int* flag, MPI_Status* status)
1210 if(req != MPI_REQUEST_NULL && req->action_ != nullptr) {
1211 req->iprobe(req->comm_->group()->rank(req->src_), req->tag_, req->comm_, flag, status);
1215 if (req != MPI_REQUEST_NULL && (req->flags_ & MPI_REQ_GENERALIZED) && not(req->flags_ & MPI_REQ_COMPLETE)) {
1221 if(req != MPI_REQUEST_NULL &&
1222 status != MPI_STATUS_IGNORE) {
1223 aid_t src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
1224 status->MPI_SOURCE = req->comm_->group()->rank(src);
1225 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
1226 status->MPI_ERROR = req->truncated_ ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
1227 status->count = req->real_size_;
1232 int Request::grequest_start(MPI_Grequest_query_function* query_fn, MPI_Grequest_free_function* free_fn,
1233 MPI_Grequest_cancel_function* cancel_fn, void* extra_state, MPI_Request* request)
1235 *request = new Request();
1236 (*request)->flags_ |= MPI_REQ_GENERALIZED;
1237 (*request)->flags_ |= MPI_REQ_PERSISTENT;
1238 (*request)->refcount_ = 1;
1239 ((*request)->generalized_funcs) = std::make_unique<smpi_mpi_generalized_request_funcs_t>();
1240 ((*request)->generalized_funcs)->query_fn=query_fn;
1241 ((*request)->generalized_funcs)->free_fn=free_fn;
1242 ((*request)->generalized_funcs)->cancel_fn=cancel_fn;
1243 ((*request)->generalized_funcs)->extra_state=extra_state;
1244 ((*request)->generalized_funcs)->cond = simgrid::s4u::ConditionVariable::create();
1245 ((*request)->generalized_funcs)->mutex = simgrid::s4u::Mutex::create();
1249 int Request::grequest_complete(MPI_Request request)
1251 if ((not(request->flags_ & MPI_REQ_GENERALIZED)) || request->generalized_funcs->mutex == nullptr)
1252 return MPI_ERR_REQUEST;
1253 request->generalized_funcs->mutex->lock();
1254 request->flags_ |= MPI_REQ_COMPLETE; // in case wait would be called after complete
1255 request->generalized_funcs->cond->notify_one();
1256 request->generalized_funcs->mutex->unlock();
1260 void Request::start_nbc_requests(std::vector<MPI_Request> reqs){
1261 if (not reqs.empty()) {
1262 nbc_requests_ = reqs;
1263 Request::startall(reqs.size(), reqs.data());
1267 std::vector<MPI_Request> Request::get_nbc_requests() const
1269 return nbc_requests_;