1/*
2 * Copyright (c) 2010-2014 ARM Limited
3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2013 Amin Farmahini-Farahani
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Andreas Hansson
41 * Ani Udipi
42 * Neha Agarwal
43 */
44
45#include "base/bitfield.hh"
46#include "base/trace.hh"
47#include "debug/DRAM.hh"
48#include "debug/DRAMPower.hh"
49#include "debug/DRAMState.hh"
50#include "debug/Drain.hh"
51#include "mem/dram_ctrl.hh"
52#include "sim/system.hh"
53
54using namespace std;
55
56DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
57 AbstractMemory(p),
58 port(name() + ".port", *this),
59 retryRdReq(false), retryWrReq(false),
60 busState(READ),
61 nextReqEvent(this), respondEvent(this), activateEvent(this),
62 prechargeEvent(this), refreshEvent(this), powerEvent(this),
63 drainManager(NULL),
64 deviceBusWidth(p->device_bus_width), burstLength(p->burst_length),
65 deviceRowBufferSize(p->device_rowbuffer_size),
66 devicesPerRank(p->devices_per_rank),
67 burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8),
68 rowBufferSize(devicesPerRank * deviceRowBufferSize),
69 columnsPerRowBuffer(rowBufferSize / burstSize),
70 ranksPerChannel(p->ranks_per_channel),
71 banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0),
72 readBufferSize(p->read_buffer_size),
73 writeBufferSize(p->write_buffer_size),
74 writeHighThreshold(writeBufferSize * p->write_high_thresh_perc / 100.0),
75 writeLowThreshold(writeBufferSize * p->write_low_thresh_perc / 100.0),
76 minWritesPerSwitch(p->min_writes_per_switch),
77 writesThisTime(0), readsThisTime(0),
78 tCK(p->tCK), tWTR(p->tWTR), tRTW(p->tRTW), tBURST(p->tBURST),
79 tRCD(p->tRCD), tCL(p->tCL), tRP(p->tRP), tRAS(p->tRAS), tWR(p->tWR),
80 tRTP(p->tRTP), tRFC(p->tRFC), tREFI(p->tREFI), tRRD(p->tRRD),
81 tXAW(p->tXAW), activationLimit(p->activation_limit),
82 memSchedPolicy(p->mem_sched_policy), addrMapping(p->addr_mapping),
83 pageMgmt(p->page_policy),
84 maxAccessesPerRow(p->max_accesses_per_row),
85 frontendLatency(p->static_frontend_latency),
86 backendLatency(p->static_backend_latency),
87 busBusyUntil(0), refreshDueAt(0), refreshState(REF_IDLE),
88 pwrStateTrans(PWR_IDLE), pwrState(PWR_IDLE), prevArrival(0),
89 nextReqTime(0), pwrStateTick(0), numBanksActive(0)
90{
91 // create the bank states based on the dimensions of the ranks and
92 // banks
93 banks.resize(ranksPerChannel);
94 actTicks.resize(ranksPerChannel);
95 for (size_t c = 0; c < ranksPerChannel; ++c) {
96 banks[c].resize(banksPerRank);
97 actTicks[c].resize(activationLimit, 0);
98 }
99
100 // set the bank indices
101 for (int r = 0; r < ranksPerChannel; r++) {
102 for (int b = 0; b < banksPerRank; b++) {
103 banks[r][b].rank = r;
104 banks[r][b].bank = b;
105 }
106 }
107
108 // perform a basic check of the write thresholds
109 if (p->write_low_thresh_perc >= p->write_high_thresh_perc)
110 fatal("Write buffer low threshold %d must be smaller than the "
111 "high threshold %d\n", p->write_low_thresh_perc,
112 p->write_high_thresh_perc);
113
114 // determine the rows per bank by looking at the total capacity
115 uint64_t capacity = ULL(1) << ceilLog2(AbstractMemory::size());
116
117 DPRINTF(DRAM, "Memory capacity %lld (%lld) bytes\n", capacity,
118 AbstractMemory::size());
119
120 DPRINTF(DRAM, "Row buffer size %d bytes with %d columns per row buffer\n",
121 rowBufferSize, columnsPerRowBuffer);
122
123 rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel);
124
125 if (range.interleaved()) {
126 if (channels != range.stripes())
127 fatal("%s has %d interleaved address stripes but %d channel(s)\n",
128 name(), range.stripes(), channels);
129
130 if (addrMapping == Enums::RoRaBaChCo) {
131 if (rowBufferSize != range.granularity()) {
132 fatal("Interleaving of %s doesn't match RoRaBaChCo "
133 "address map\n", name());
134 }
135 } else if (addrMapping == Enums::RoRaBaCoCh) {
136 if (system()->cacheLineSize() != range.granularity()) {
137 fatal("Interleaving of %s doesn't match RoRaBaCoCh "
138 "address map\n", name());
139 }
140 } else if (addrMapping == Enums::RoCoRaBaCh) {
141 if (system()->cacheLineSize() != range.granularity())
142 fatal("Interleaving of %s doesn't match RoCoRaBaCh "
143 "address map\n", name());
144 }
145 }
146
147 // some basic sanity checks
148 if (tREFI <= tRP || tREFI <= tRFC) {
149 fatal("tREFI (%d) must be larger than tRP (%d) and tRFC (%d)\n",
150 tREFI, tRP, tRFC);
151 }
152}
153
154void
155DRAMCtrl::init()
156{
157 if (!port.isConnected()) {
158 fatal("DRAMCtrl %s is unconnected!\n", name());
159 } else {
160 port.sendRangeChange();
161 }
162}
163
164void
165DRAMCtrl::startup()
166{
167 // update the start tick for the precharge accounting to the
168 // current tick
169 pwrStateTick = curTick();
170
171 // shift the bus busy time sufficiently far ahead that we never
172 // have to worry about negative values when computing the time for
173 // the next request, this will add an insignificant bubble at the
174 // start of simulation
175 busBusyUntil = curTick() + tRP + tRCD + tCL;
176
177 // kick off the refresh, and give ourselves enough time to
178 // precharge
179 schedule(refreshEvent, curTick() + tREFI - tRP);
180}
181
182Tick
183DRAMCtrl::recvAtomic(PacketPtr pkt)
184{
185 DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
186
187 // do the actual memory access and turn the packet into a response
188 access(pkt);
189
190 Tick latency = 0;
191 if (!pkt->memInhibitAsserted() && pkt->hasData()) {
192 // this value is not supposed to be accurate, just enough to
193 // keep things going, mimic a closed page
194 latency = tRP + tRCD + tCL;
195 }
196 return latency;
197}
198
199bool
200DRAMCtrl::readQueueFull(unsigned int neededEntries) const
201{
202 DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
203 readBufferSize, readQueue.size() + respQueue.size(),
204 neededEntries);
205
206 return
207 (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
208}
209
210bool
211DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
212{
213 DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
214 writeBufferSize, writeQueue.size(), neededEntries);
215 return (writeQueue.size() + neededEntries) > writeBufferSize;
216}
217
218DRAMCtrl::DRAMPacket*
219DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
220 bool isRead)
221{
222 // decode the address based on the address mapping scheme, with
223 // Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
224 // channel, respectively
225 uint8_t rank;
226 uint8_t bank;
227 // use a 64-bit unsigned during the computations as the row is
228 // always the top bits, and check before creating the DRAMPacket
229 uint64_t row;
230
231 // truncate the address to the access granularity
232 Addr addr = dramPktAddr / burstSize;
233
234 // we have removed the lowest order address bits that denote the
235 // position within the column
236 if (addrMapping == Enums::RoRaBaChCo) {
237 // the lowest order bits denote the column to ensure that
238 // sequential cache lines occupy the same row
239 addr = addr / columnsPerRowBuffer;
240
241 // take out the channel part of the address
242 addr = addr / channels;
243
244 // after the channel bits, get the bank bits to interleave
245 // over the banks
246 bank = addr % banksPerRank;
247 addr = addr / banksPerRank;
248
249 // after the bank, we get the rank bits which thus interleaves
250 // over the ranks
251 rank = addr % ranksPerChannel;
252 addr = addr / ranksPerChannel;
253
254 // lastly, get the row bits
255 row = addr % rowsPerBank;
256 addr = addr / rowsPerBank;
257 } else if (addrMapping == Enums::RoRaBaCoCh) {
258 // take out the channel part of the address
259 addr = addr / channels;
260
261 // next, the column
262 addr = addr / columnsPerRowBuffer;
263
264 // after the column bits, we get the bank bits to interleave
265 // over the banks
266 bank = addr % banksPerRank;
267 addr = addr / banksPerRank;
268
269 // after the bank, we get the rank bits which thus interleaves
270 // over the ranks
271 rank = addr % ranksPerChannel;
272 addr = addr / ranksPerChannel;
273
274 // lastly, get the row bits
275 row = addr % rowsPerBank;
276 addr = addr / rowsPerBank;
277 } else if (addrMapping == Enums::RoCoRaBaCh) {
278 // optimise for closed page mode and utilise maximum
279 // parallelism of the DRAM (at the cost of power)
280
281 // take out the channel part of the address, not that this has
282 // to match with how accesses are interleaved between the
283 // controllers in the address mapping
284 addr = addr / channels;
285
286 // start with the bank bits, as this provides the maximum
287 // opportunity for parallelism between requests
288 bank = addr % banksPerRank;
289 addr = addr / banksPerRank;
290
291 // next get the rank bits
292 rank = addr % ranksPerChannel;
293 addr = addr / ranksPerChannel;
294
295 // next the column bits which we do not need to keep track of
296 // and simply skip past
297 addr = addr / columnsPerRowBuffer;
298
299 // lastly, get the row bits
300 row = addr % rowsPerBank;
301 addr = addr / rowsPerBank;
302 } else
303 panic("Unknown address mapping policy chosen!");
304
305 assert(rank < ranksPerChannel);
306 assert(bank < banksPerRank);
307 assert(row < rowsPerBank);
308 assert(row < Bank::NO_ROW);
309
310 DPRINTF(DRAM, "Address: %lld Rank %d Bank %d Row %d\n",
311 dramPktAddr, rank, bank, row);
312
313 // create the corresponding DRAM packet with the entry time and
314 // ready time set to the current tick, the latter will be updated
315 // later
316 uint16_t bank_id = banksPerRank * rank + bank;
317 return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
318 size, banks[rank][bank]);
319}
320
321void
322DRAMCtrl::addToReadQueue(PacketPtr pkt, unsigned int pktCount)
323{
324 // only add to the read queue here. whenever the request is
325 // eventually done, set the readyTime, and call schedule()
326 assert(!pkt->isWrite());
327
328 assert(pktCount != 0);
329
330 // if the request size is larger than burst size, the pkt is split into
331 // multiple DRAM packets
332 // Note if the pkt starting address is not aligened to burst size, the
333 // address of first DRAM packet is kept unaliged. Subsequent DRAM packets
334 // are aligned to burst size boundaries. This is to ensure we accurately
335 // check read packets against packets in write queue.
336 Addr addr = pkt->getAddr();
337 unsigned pktsServicedByWrQ = 0;
338 BurstHelper* burst_helper = NULL;
339 for (int cnt = 0; cnt < pktCount; ++cnt) {
340 unsigned size = std::min((addr | (burstSize - 1)) + 1,
341 pkt->getAddr() + pkt->getSize()) - addr;
342 readPktSize[ceilLog2(size)]++;
343 readBursts++;
344
345 // First check write buffer to see if the data is already at
346 // the controller
347 bool foundInWrQ = false;
348 for (auto i = writeQueue.begin(); i != writeQueue.end(); ++i) {
349 // check if the read is subsumed in the write entry we are
350 // looking at
351 if ((*i)->addr <= addr &&
352 (addr + size) <= ((*i)->addr + (*i)->size)) {
353 foundInWrQ = true;
354 servicedByWrQ++;
355 pktsServicedByWrQ++;
356 DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
357 "write queue\n", addr, size);
358 bytesReadWrQ += burstSize;
359 break;
360 }
361 }
362
363 // If not found in the write q, make a DRAM packet and
364 // push it onto the read queue
365 if (!foundInWrQ) {
366
367 // Make the burst helper for split packets
368 if (pktCount > 1 && burst_helper == NULL) {
369 DPRINTF(DRAM, "Read to addr %lld translates to %d "
370 "dram requests\n", pkt->getAddr(), pktCount);
371 burst_helper = new BurstHelper(pktCount);
372 }
373
374 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, true);
375 dram_pkt->burstHelper = burst_helper;
376
377 assert(!readQueueFull(1));
378 rdQLenPdf[readQueue.size() + respQueue.size()]++;
379
380 DPRINTF(DRAM, "Adding to read queue\n");
381
382 readQueue.push_back(dram_pkt);
383
384 // Update stats
385 avgRdQLen = readQueue.size() + respQueue.size();
386 }
387
388 // Starting address of next dram pkt (aligend to burstSize boundary)
389 addr = (addr | (burstSize - 1)) + 1;
390 }
391
392 // If all packets are serviced by write queue, we send the repsonse back
393 if (pktsServicedByWrQ == pktCount) {
394 accessAndRespond(pkt, frontendLatency);
395 return;
396 }
397
398 // Update how many split packets are serviced by write queue
399 if (burst_helper != NULL)
400 burst_helper->burstsServiced = pktsServicedByWrQ;
401
402 // If we are not already scheduled to get a request out of the
403 // queue, do so now
404 if (!nextReqEvent.scheduled()) {
405 DPRINTF(DRAM, "Request scheduled immediately\n");
406 schedule(nextReqEvent, curTick());
407 }
408}
409
410void
411DRAMCtrl::addToWriteQueue(PacketPtr pkt, unsigned int pktCount)
412{
413 // only add to the write queue here. whenever the request is
414 // eventually done, set the readyTime, and call schedule()
415 assert(pkt->isWrite());
416
417 // if the request size is larger than burst size, the pkt is split into
418 // multiple DRAM packets
419 Addr addr = pkt->getAddr();
420 for (int cnt = 0; cnt < pktCount; ++cnt) {
421 unsigned size = std::min((addr | (burstSize - 1)) + 1,
422 pkt->getAddr() + pkt->getSize()) - addr;
423 writePktSize[ceilLog2(size)]++;
424 writeBursts++;
425
426 // see if we can merge with an existing item in the write
427 // queue and keep track of whether we have merged or not so we
428 // can stop at that point and also avoid enqueueing a new
429 // request
430 bool merged = false;
431 auto w = writeQueue.begin();
432
433 while(!merged && w != writeQueue.end()) {
434 // either of the two could be first, if they are the same
435 // it does not matter which way we go
436 if ((*w)->addr >= addr) {
437 // the existing one starts after the new one, figure
438 // out where the new one ends with respect to the
439 // existing one
440 if ((addr + size) >= ((*w)->addr + (*w)->size)) {
441 // check if the existing one is completely
442 // subsumed in the new one
443 DPRINTF(DRAM, "Merging write covering existing burst\n");
444 merged = true;
445 // update both the address and the size
446 (*w)->addr = addr;
447 (*w)->size = size;
448 } else if ((addr + size) >= (*w)->addr &&
449 ((*w)->addr + (*w)->size - addr) <= burstSize) {
450 // the new one is just before or partially
451 // overlapping with the existing one, and together
452 // they fit within a burst
453 DPRINTF(DRAM, "Merging write before existing burst\n");
454 merged = true;
455 // the existing queue item needs to be adjusted with
456 // respect to both address and size
457 (*w)->size = (*w)->addr + (*w)->size - addr;
458 (*w)->addr = addr;
459 }
460 } else {
461 // the new one starts after the current one, figure
462 // out where the existing one ends with respect to the
463 // new one
464 if (((*w)->addr + (*w)->size) >= (addr + size)) {
465 // check if the new one is completely subsumed in the
466 // existing one
467 DPRINTF(DRAM, "Merging write into existing burst\n");
468 merged = true;
469 // no adjustments necessary
470 } else if (((*w)->addr + (*w)->size) >= addr &&
471 (addr + size - (*w)->addr) <= burstSize) {
472 // the existing one is just before or partially
473 // overlapping with the new one, and together
474 // they fit within a burst
475 DPRINTF(DRAM, "Merging write after existing burst\n");
476 merged = true;
477 // the address is right, and only the size has
478 // to be adjusted
479 (*w)->size = addr + size - (*w)->addr;
480 }
481 }
482 ++w;
483 }
484
485 // if the item was not merged we need to create a new write
486 // and enqueue it
487 if (!merged) {
488 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
489
490 assert(writeQueue.size() < writeBufferSize);
491 wrQLenPdf[writeQueue.size()]++;
492
493 DPRINTF(DRAM, "Adding to write queue\n");
494
495 writeQueue.push_back(dram_pkt);
496
497 // Update stats
498 avgWrQLen = writeQueue.size();
499 } else {
500 // keep track of the fact that this burst effectively
501 // disappeared as it was merged with an existing one
502 mergedWrBursts++;
503 }
504
505 // Starting address of next dram pkt (aligend to burstSize boundary)
506 addr = (addr | (burstSize - 1)) + 1;
507 }
508
509 // we do not wait for the writes to be send to the actual memory,
510 // but instead take responsibility for the consistency here and
511 // snoop the write queue for any upcoming reads
512 // @todo, if a pkt size is larger than burst size, we might need a
513 // different front end latency
514 accessAndRespond(pkt, frontendLatency);
515
516 // If we are not already scheduled to get a request out of the
517 // queue, do so now
518 if (!nextReqEvent.scheduled()) {
519 DPRINTF(DRAM, "Request scheduled immediately\n");
520 schedule(nextReqEvent, curTick());
521 }
522}
523
524void
525DRAMCtrl::printQs() const {
526 DPRINTF(DRAM, "===READ QUEUE===\n\n");
527 for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
528 DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
529 }
530 DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
531 for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
532 DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
533 }
534 DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
535 for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
536 DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
537 }
538}
539
540bool
541DRAMCtrl::recvTimingReq(PacketPtr pkt)
542{
543 /// @todo temporary hack to deal with memory corruption issues until
544 /// 4-phase transactions are complete
545 for (int x = 0; x < pendingDelete.size(); x++)
546 delete pendingDelete[x];
547 pendingDelete.clear();
548
549 // This is where we enter from the outside world
550 DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
551 pkt->cmdString(), pkt->getAddr(), pkt->getSize());
552
553 // simply drop inhibited packets for now
554 if (pkt->memInhibitAsserted()) {
555 DPRINTF(DRAM, "Inhibited packet -- Dropping it now\n");
556 pendingDelete.push_back(pkt);
557 return true;
558 }
559
560 // Calc avg gap between requests
561 if (prevArrival != 0) {
562 totGap += curTick() - prevArrival;
563 }
564 prevArrival = curTick();
565
566
567 // Find out how many dram packets a pkt translates to
568 // If the burst size is equal or larger than the pkt size, then a pkt
569 // translates to only one dram packet. Otherwise, a pkt translates to
570 // multiple dram packets
571 unsigned size = pkt->getSize();
572 unsigned offset = pkt->getAddr() & (burstSize - 1);
573 unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
574
575 // check local buffers and do not accept if full
576 if (pkt->isRead()) {
577 assert(size != 0);
578 if (readQueueFull(dram_pkt_count)) {
579 DPRINTF(DRAM, "Read queue full, not accepting\n");
580 // remember that we have to retry this port
581 retryRdReq = true;
582 numRdRetry++;
583 return false;
584 } else {
585 addToReadQueue(pkt, dram_pkt_count);
586 readReqs++;
587 bytesReadSys += size;
588 }
589 } else if (pkt->isWrite()) {
590 assert(size != 0);
591 if (writeQueueFull(dram_pkt_count)) {
592 DPRINTF(DRAM, "Write queue full, not accepting\n");
593 // remember that we have to retry this port
594 retryWrReq = true;
595 numWrRetry++;
596 return false;
597 } else {
598 addToWriteQueue(pkt, dram_pkt_count);
599 writeReqs++;
600 bytesWrittenSys += size;
601 }
602 } else {
603 DPRINTF(DRAM,"Neither read nor write, ignore timing\n");
604 neitherReadNorWrite++;
605 accessAndRespond(pkt, 1);
606 }
607
608 return true;
609}
610
611void
612DRAMCtrl::processRespondEvent()
613{
614 DPRINTF(DRAM,
615 "processRespondEvent(): Some req has reached its readyTime\n");
616
617 DRAMPacket* dram_pkt = respQueue.front();
618
619 if (dram_pkt->burstHelper) {
620 // it is a split packet
621 dram_pkt->burstHelper->burstsServiced++;
622 if (dram_pkt->burstHelper->burstsServiced ==
623 dram_pkt->burstHelper->burstCount) {
624 // we have now serviced all children packets of a system packet
625 // so we can now respond to the requester
626 // @todo we probably want to have a different front end and back
627 // end latency for split packets
628 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
629 delete dram_pkt->burstHelper;
630 dram_pkt->burstHelper = NULL;
631 }
632 } else {
633 // it is not a split packet
634 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
635 }
636
637 delete respQueue.front();
638 respQueue.pop_front();
639
640 if (!respQueue.empty()) {
641 assert(respQueue.front()->readyTime >= curTick());
642 assert(!respondEvent.scheduled());
643 schedule(respondEvent, respQueue.front()->readyTime);
644 } else {
645 // if there is nothing left in any queue, signal a drain
646 if (writeQueue.empty() && readQueue.empty() &&
647 drainManager) {
648 drainManager->signalDrainDone();
649 drainManager = NULL;
650 }
651 }
652
653 // We have made a location in the queue available at this point,
654 // so if there is a read that was forced to wait, retry now
655 if (retryRdReq) {
656 retryRdReq = false;
657 port.sendRetry();
658 }
659}
660
661void
662DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue)
663{
664 // This method does the arbitration between requests. The chosen
665 // packet is simply moved to the head of the queue. The other
666 // methods know that this is the place to look. For example, with
667 // FCFS, this method does nothing
668 assert(!queue.empty());
669
670 if (queue.size() == 1) {
671 DPRINTF(DRAM, "Single request, nothing to do\n");
672 return;
673 }
674
675 if (memSchedPolicy == Enums::fcfs) {
676 // Do nothing, since the correct request is already head
677 } else if (memSchedPolicy == Enums::frfcfs) {
678 reorderQueue(queue);
679 } else
680 panic("No scheduling policy chosen\n");
681}
682
683void
684DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue)
685{
686 // Only determine this when needed
687 uint64_t earliest_banks = 0;
688
689 // Search for row hits first, if no row hit is found then schedule the
690 // packet to one of the earliest banks available
691 bool found_earliest_pkt = false;
692 auto selected_pkt_it = queue.begin();
693
694 for (auto i = queue.begin(); i != queue.end() ; ++i) {
695 DRAMPacket* dram_pkt = *i;
696 const Bank& bank = dram_pkt->bankRef;
697 // Check if it is a row hit
698 if (bank.openRow == dram_pkt->row) {
699 // FCFS within the hits
700 DPRINTF(DRAM, "Row buffer hit\n");
701 selected_pkt_it = i;
702 break;
703 } else if (!found_earliest_pkt) {
704 // No row hit, go for first ready
705 if (earliest_banks == 0)
706 earliest_banks = minBankActAt(queue);
707
708 // simplistic approximation of when the bank can issue an
709 // activate, this is calculated in minBankActAt and could
710 // be cached
711 Tick act_at = bank.openRow == Bank::NO_ROW ?
712 bank.actAllowedAt :
713 std::max(bank.preAllowedAt, curTick()) + tRP;
714
715 // Bank is ready or is the first available bank
716 if (act_at <= curTick() ||
717 bits(earliest_banks, dram_pkt->bankId, dram_pkt->bankId)) {
718 // Remember the packet to be scheduled to one of the earliest
719 // banks available, FCFS amongst the earliest banks
720 selected_pkt_it = i;
721 found_earliest_pkt = true;
722 }
723 }
724 }
725
726 DRAMPacket* selected_pkt = *selected_pkt_it;
727 queue.erase(selected_pkt_it);
728 queue.push_front(selected_pkt);
729}
730
731void
732DRAMCtrl::accessAndRespond(PacketPtr pkt, Tick static_latency)
733{
734 DPRINTF(DRAM, "Responding to Address %lld.. ",pkt->getAddr());
735
736 bool needsResponse = pkt->needsResponse();
737 // do the actual memory access which also turns the packet into a
738 // response
739 access(pkt);
740
741 // turn packet around to go back to requester if response expected
742 if (needsResponse) {
743 // access already turned the packet into a response
744 assert(pkt->isResponse());
745
746 // @todo someone should pay for this
747 pkt->busFirstWordDelay = pkt->busLastWordDelay = 0;
748
749 // queue the packet in the response queue to be sent out after
750 // the static latency has passed
751 port.schedTimingResp(pkt, curTick() + static_latency);
752 } else {
753 // @todo the packet is going to be deleted, and the DRAMPacket
754 // is still having a pointer to it
755 pendingDelete.push_back(pkt);
756 }
757
758 DPRINTF(DRAM, "Done\n");
759
760 return;
761}
762
763void
764DRAMCtrl::activateBank(Bank& bank, Tick act_tick, uint32_t row)
765{
766 // get the rank index from the bank
767 uint8_t rank = bank.rank;
768
769 assert(actTicks[rank].size() == activationLimit);
770
771 DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
772
773 // update the open row
774 assert(bank.openRow == Bank::NO_ROW);
775 bank.openRow = row;
776
777 // start counting anew, this covers both the case when we
778 // auto-precharged, and when this access is forced to
779 // precharge
780 bank.bytesAccessed = 0;
781 bank.rowAccesses = 0;
782
783 ++numBanksActive;
784 assert(numBanksActive <= banksPerRank * ranksPerChannel);
785
786 DPRINTF(DRAM, "Activate bank %d, rank %d at tick %lld, now got %d active\n",
787 bank.bank, bank.rank, act_tick, numBanksActive);
788
789 DPRINTF(DRAMPower, "%llu,ACT,%d,%d\n", divCeil(act_tick, tCK), bank.bank,
790 bank.rank);
791
792 // The next access has to respect tRAS for this bank
793 bank.preAllowedAt = act_tick + tRAS;
794
795 // Respect the row-to-column command delay
796 bank.colAllowedAt = act_tick + tRCD;
797
798 // start by enforcing tRRD
799 for(int i = 0; i < banksPerRank; i++) {
800 // next activate to any bank in this rank must not happen
801 // before tRRD
802 banks[rank][i].actAllowedAt = std::max(act_tick + tRRD,
803 banks[rank][i].actAllowedAt);
804 }
805
806 // next, we deal with tXAW, if the activation limit is disabled
807 // then we are done
808 if (actTicks[rank].empty())
809 return;
810
811 // sanity check
812 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
813 panic("Got %d activates in window %d (%llu - %llu) which is smaller "
814 "than %llu\n", activationLimit, act_tick - actTicks[rank].back(),
815 act_tick, actTicks[rank].back(), tXAW);
816 }
817
818 // shift the times used for the book keeping, the last element
819 // (highest index) is the oldest one and hence the lowest value
820 actTicks[rank].pop_back();
821
822 // record an new activation (in the future)
823 actTicks[rank].push_front(act_tick);
824
825 // cannot activate more than X times in time window tXAW, push the
826 // next one (the X + 1'st activate) to be tXAW away from the
827 // oldest in our window of X
828 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
829 DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate no earlier "
830 "than %llu\n", activationLimit, actTicks[rank].back() + tXAW);
831 for(int j = 0; j < banksPerRank; j++)
832 // next activate must not happen before end of window
833 banks[rank][j].actAllowedAt =
834 std::max(actTicks[rank].back() + tXAW,
835 banks[rank][j].actAllowedAt);
836 }
837
838 // at the point when this activate takes place, make sure we
839 // transition to the active power state
840 if (!activateEvent.scheduled())
841 schedule(activateEvent, act_tick);
842 else if (activateEvent.when() > act_tick)
843 // move it sooner in time
844 reschedule(activateEvent, act_tick);
845}
846
847void
848DRAMCtrl::processActivateEvent()
849{
850 // we should transition to the active state as soon as any bank is active
851 if (pwrState != PWR_ACT)
852 // note that at this point numBanksActive could be back at
853 // zero again due to a precharge scheduled in the future
854 schedulePowerEvent(PWR_ACT, curTick());
855}
856
857void
858DRAMCtrl::prechargeBank(Bank& bank, Tick pre_at, bool trace)
859{
860 // make sure the bank has an open row
861 assert(bank.openRow != Bank::NO_ROW);
862
863 // sample the bytes per activate here since we are closing
864 // the page
865 bytesPerActivate.sample(bank.bytesAccessed);
866
867 bank.openRow = Bank::NO_ROW;
868
869 // no precharge allowed before this one
870 bank.preAllowedAt = pre_at;
871
872 Tick pre_done_at = pre_at + tRP;
873
874 bank.actAllowedAt = std::max(bank.actAllowedAt, pre_done_at);
875
876 assert(numBanksActive != 0);
877 --numBanksActive;
878
879 DPRINTF(DRAM, "Precharging bank %d, rank %d at tick %lld, now got "
880 "%d active\n", bank.bank, bank.rank, pre_at, numBanksActive);
881
882 if (trace)
883 DPRINTF(DRAMPower, "%llu,PRE,%d,%d\n", divCeil(pre_at, tCK),
884 bank.bank, bank.rank);
885
886 // if we look at the current number of active banks we might be
887 // tempted to think the DRAM is now idle, however this can be
888 // undone by an activate that is scheduled to happen before we
889 // would have reached the idle state, so schedule an event and
890 // rather check once we actually make it to the point in time when
891 // the (last) precharge takes place
892 if (!prechargeEvent.scheduled())
893 schedule(prechargeEvent, pre_done_at);
894 else if (prechargeEvent.when() < pre_done_at)
895 reschedule(prechargeEvent, pre_done_at);
896}
897
898void
899DRAMCtrl::processPrechargeEvent()
900{
901 // if we reached zero, then special conditions apply as we track
902 // if all banks are precharged for the power models
903 if (numBanksActive == 0) {
904 // we should transition to the idle state when the last bank
905 // is precharged
906 schedulePowerEvent(PWR_IDLE, curTick());
907 }
908}
909
910void
911DRAMCtrl::doDRAMAccess(DRAMPacket* dram_pkt)
912{
913 DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
914 dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
915
916 // get the bank
917 Bank& bank = dram_pkt->bankRef;
918
919 // for the state we need to track if it is a row hit or not
920 bool row_hit = true;
921
922 // respect any constraints on the command (e.g. tRCD or tCCD)
923 Tick cmd_at = std::max(bank.colAllowedAt, curTick());
924
925 // Determine the access latency and update the bank state
926 if (bank.openRow == dram_pkt->row) {
927 // nothing to do
928 } else {
929 row_hit = false;
930
931 // If there is a page open, precharge it.
932 if (bank.openRow != Bank::NO_ROW) {
933 prechargeBank(bank, std::max(bank.preAllowedAt, curTick()));
934 }
935
936 // next we need to account for the delay in activating the
937 // page
938 Tick act_tick = std::max(bank.actAllowedAt, curTick());
939
940 // Record the activation and deal with all the global timing
941 // constraints caused be a new activation (tRRD and tXAW)
942 activateBank(bank, act_tick, dram_pkt->row);
943
944 // issue the command as early as possible
945 cmd_at = bank.colAllowedAt;
946 }
947
948 // we need to wait until the bus is available before we can issue
949 // the command
950 cmd_at = std::max(cmd_at, busBusyUntil - tCL);
951
952 // update the packet ready time
953 dram_pkt->readyTime = cmd_at + tCL + tBURST;
954
955 // only one burst can use the bus at any one point in time
956 assert(dram_pkt->readyTime - busBusyUntil >= tBURST);
957
958 // not strictly necessary, but update the time for the next
959 // read/write (add a max with tCCD here)
960 bank.colAllowedAt = cmd_at + tBURST;
961
962 // If this is a write, we also need to respect the write recovery
963 // time before a precharge, in the case of a read, respect the
964 // read to precharge constraint
965 bank.preAllowedAt = std::max(bank.preAllowedAt,
966 dram_pkt->isRead ? cmd_at + tRTP :
967 dram_pkt->readyTime + tWR);
968
969 // increment the bytes accessed and the accesses per row
970 bank.bytesAccessed += burstSize;
971 ++bank.rowAccesses;
972
973 // if we reached the max, then issue with an auto-precharge
974 bool auto_precharge = pageMgmt == Enums::close ||
975 bank.rowAccesses == maxAccessesPerRow;
976
977 // if we did not hit the limit, we might still want to
978 // auto-precharge
979 if (!auto_precharge &&
980 (pageMgmt == Enums::open_adaptive ||
981 pageMgmt == Enums::close_adaptive)) {
982 // a twist on the open and close page policies:
983 // 1) open_adaptive page policy does not blindly keep the
984 // page open, but close it if there are no row hits, and there
985 // are bank conflicts in the queue
986 // 2) close_adaptive page policy does not blindly close the
987 // page, but closes it only if there are no row hits in the queue.
988 // In this case, only force an auto precharge when there
989 // are no same page hits in the queue
990 bool got_more_hits = false;
991 bool got_bank_conflict = false;
992
993 // either look at the read queue or write queue
994 const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
995 writeQueue;
996 auto p = queue.begin();
997 // make sure we are not considering the packet that we are
998 // currently dealing with (which is the head of the queue)
999 ++p;
1000
1001 // keep on looking until we have found required condition or
1002 // reached the end
1003 while (!(got_more_hits &&
1004 (got_bank_conflict || pageMgmt == Enums::close_adaptive)) &&
1005 p != queue.end()) {
1006 bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
1007 (dram_pkt->bank == (*p)->bank);
1008 bool same_row = dram_pkt->row == (*p)->row;
1009 got_more_hits |= same_rank_bank && same_row;
1010 got_bank_conflict |= same_rank_bank && !same_row;
1011 ++p;
1012 }
1013
1014 // auto pre-charge when either
1015 // 1) open_adaptive policy, we have not got any more hits, and
1016 // have a bank conflict
1017 // 2) close_adaptive policy and we have not got any more hits
1018 auto_precharge = !got_more_hits &&
1019 (got_bank_conflict || pageMgmt == Enums::close_adaptive);
1020 }
1021
1022 // DRAMPower trace command to be written
1023 std::string mem_cmd = dram_pkt->isRead ? "RD" : "WR";
1024
1025 // if this access should use auto-precharge, then we are
1026 // closing the row
1027 if (auto_precharge) {
1028 prechargeBank(bank, std::max(curTick(), bank.preAllowedAt), false);
1029
1030 mem_cmd.append("A");
1031
1032 DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
1033 }
1034
1035 // Update bus state
1036 busBusyUntil = dram_pkt->readyTime;
1037
1038 DPRINTF(DRAM, "Access to %lld, ready at %lld bus busy until %lld.\n",
1039 dram_pkt->addr, dram_pkt->readyTime, busBusyUntil);
1040
1041 DPRINTF(DRAMPower, "%llu,%s,%d,%d\n", divCeil(cmd_at, tCK), mem_cmd,
1042 dram_pkt->bank, dram_pkt->rank);
1043
1044 // Update the minimum timing between the requests, this is a
1045 // conservative estimate of when we have to schedule the next
1046 // request to not introduce any unecessary bubbles. In most cases
1047 // we will wake up sooner than we have to.
1048 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1049
1050 // Update the stats and schedule the next request
1051 if (dram_pkt->isRead) {
1052 ++readsThisTime;
1053 if (row_hit)
1054 readRowHits++;
1055 bytesReadDRAM += burstSize;
1056 perBankRdBursts[dram_pkt->bankId]++;
1057
1058 // Update latency stats
1059 totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
1060 totBusLat += tBURST;
1061 totQLat += cmd_at - dram_pkt->entryTime;
1062 } else {
1063 ++writesThisTime;
1064 if (row_hit)
1065 writeRowHits++;
1066 bytesWritten += burstSize;
1067 perBankWrBursts[dram_pkt->bankId]++;
1068 }
1069}
1070
1071void
1072DRAMCtrl::processNextReqEvent()
1073{
1074 if (busState == READ_TO_WRITE) {
1075 DPRINTF(DRAM, "Switching to writes after %d reads with %d reads "
1076 "waiting\n", readsThisTime, readQueue.size());
1077
1078 // sample and reset the read-related stats as we are now
1079 // transitioning to writes, and all reads are done
1080 rdPerTurnAround.sample(readsThisTime);
1081 readsThisTime = 0;
1082
1083 // now proceed to do the actual writes
1084 busState = WRITE;
1085 } else if (busState == WRITE_TO_READ) {
1086 DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
1087 "waiting\n", writesThisTime, writeQueue.size());
1088
1089 wrPerTurnAround.sample(writesThisTime);
1090 writesThisTime = 0;
1091
1092 busState = READ;
1093 }
1094
1095 if (refreshState != REF_IDLE) {
1096 // if a refresh waiting for this event loop to finish, then hand
1097 // over now, and do not schedule a new nextReqEvent
1098 if (refreshState == REF_DRAIN) {
1099 DPRINTF(DRAM, "Refresh drain done, now precharging\n");
1100
1101 refreshState = REF_PRE;
1102
1103 // hand control back to the refresh event loop
1104 schedule(refreshEvent, curTick());
1105 }
1106
1107 // let the refresh finish before issuing any further requests
1108 return;
1109 }
1110
1111 // when we get here it is either a read or a write
1112 if (busState == READ) {
1113
1114 // track if we should switch or not
1115 bool switch_to_writes = false;
1116
1117 if (readQueue.empty()) {
1118 // In the case there is no read request to go next,
1119 // trigger writes if we have passed the low threshold (or
1120 // if we are draining)
1121 if (!writeQueue.empty() &&
1122 (drainManager || writeQueue.size() > writeLowThreshold)) {
1123
1124 switch_to_writes = true;
1125 } else {
1126 // check if we are drained
1127 if (respQueue.empty () && drainManager) {
1128 drainManager->signalDrainDone();
1129 drainManager = NULL;
1130 }
1131
1132 // nothing to do, not even any point in scheduling an
1133 // event for the next request
1134 return;
1135 }
1136 } else {
1137 // Figure out which read request goes next, and move it to the
1138 // front of the read queue
1139 chooseNext(readQueue);
1140
1141 DRAMPacket* dram_pkt = readQueue.front();
1142
1143 doDRAMAccess(dram_pkt);
1144
1145 // At this point we're done dealing with the request
1146 readQueue.pop_front();
1147
1148 // sanity check
1149 assert(dram_pkt->size <= burstSize);
1150 assert(dram_pkt->readyTime >= curTick());
1151
1152 // Insert into response queue. It will be sent back to the
1153 // requestor at its readyTime
1154 if (respQueue.empty()) {
1155 assert(!respondEvent.scheduled());
1156 schedule(respondEvent, dram_pkt->readyTime);
1157 } else {
1158 assert(respQueue.back()->readyTime <= dram_pkt->readyTime);
1159 assert(respondEvent.scheduled());
1160 }
1161
1162 respQueue.push_back(dram_pkt);
1163
1164 // we have so many writes that we have to transition
1165 if (writeQueue.size() > writeHighThreshold) {
1166 switch_to_writes = true;
1167 }
1168 }
1169
1170 // switching to writes, either because the read queue is empty
1171 // and the writes have passed the low threshold (or we are
1172 // draining), or because the writes hit the hight threshold
1173 if (switch_to_writes) {
1174 // transition to writing
1175 busState = READ_TO_WRITE;
1176
1177 // add a bubble to the data bus, as defined by the
1178 // tRTW parameter
1179 busBusyUntil += tRTW;
1180
1181 // update the minimum timing between the requests,
1182 // this shifts us back in time far enough to do any
1183 // bank preparation
1184 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1185 }
1186 } else {
1187 chooseNext(writeQueue);
1188 DRAMPacket* dram_pkt = writeQueue.front();
1189 // sanity check
1190 assert(dram_pkt->size <= burstSize);
1191 doDRAMAccess(dram_pkt);
1192
1193 writeQueue.pop_front();
1194 delete dram_pkt;
1195
1196 // If we emptied the write queue, or got sufficiently below the
1197 // threshold (using the minWritesPerSwitch as the hysteresis) and
1198 // are not draining, or we have reads waiting and have done enough
1199 // writes, then switch to reads.
1200 if (writeQueue.empty() ||
1201 (writeQueue.size() + minWritesPerSwitch < writeLowThreshold &&
1202 !drainManager) ||
1203 (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
1204 // turn the bus back around for reads again
1205 busState = WRITE_TO_READ;
1206
1207 // note that the we switch back to reads also in the idle
1208 // case, which eventually will check for any draining and
1209 // also pause any further scheduling if there is really
1210 // nothing to do
1211
1212 // here we get a bit creative and shift the bus busy time not
1213 // just the tWTR, but also a CAS latency to capture the fact
1214 // that we are allowed to prepare a new bank, but not issue a
1215 // read command until after tWTR, in essence we capture a
1216 // bubble on the data bus that is tWTR + tCL
1217 busBusyUntil += tWTR + tCL;
1218
1219 // update the minimum timing between the requests, this shifts
1220 // us back in time far enough to do any bank preparation
1221 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1222 }
1223 }
1224
1225 schedule(nextReqEvent, std::max(nextReqTime, curTick()));
1226
1227 // If there is space available and we have writes waiting then let
1228 // them retry. This is done here to ensure that the retry does not
1229 // cause a nextReqEvent to be scheduled before we do so as part of
1230 // the next request processing
1231 if (retryWrReq && writeQueue.size() < writeBufferSize) {
1232 retryWrReq = false;
1233 port.sendRetry();
1234 }
1235}
1236
1237uint64_t
1238DRAMCtrl::minBankActAt(const deque<DRAMPacket*>& queue) const
1239{
1240 uint64_t bank_mask = 0;
1241 Tick min_act_at = MaxTick;
1242
1243 // deterimne if we have queued transactions targetting a
1244 // bank in question
1245 vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
1246 for (auto p = queue.begin(); p != queue.end(); ++p) {
1247 got_waiting[(*p)->bankId] = true;
1248 }
1249
1250 for (int i = 0; i < ranksPerChannel; i++) {
1251 for (int j = 0; j < banksPerRank; j++) {
1252 uint8_t bank_id = i * banksPerRank + j;
1253
1254 // if we have waiting requests for the bank, and it is
1255 // amongst the first available, update the mask
1256 if (got_waiting[bank_id]) {
1257 // simplistic approximation of when the bank can issue
1258 // an activate, ignoring any rank-to-rank switching
1259 // cost
1260 Tick act_at = banks[i][j].openRow == Bank::NO_ROW ?
1261 banks[i][j].actAllowedAt :
1262 std::max(banks[i][j].preAllowedAt, curTick()) + tRP;
1263
1264 if (act_at <= min_act_at) {
1265 // reset bank mask if new minimum is found
1266 if (act_at < min_act_at)
1267 bank_mask = 0;
1268 // set the bit corresponding to the available bank
1269 replaceBits(bank_mask, bank_id, bank_id, 1);
1270 min_act_at = act_at;
1271 }
1272 }
1273 }
1274 }
1275
1276 return bank_mask;
1277}
1278
1279void
1280DRAMCtrl::processRefreshEvent()
1281{
1282 // when first preparing the refresh, remember when it was due
1283 if (refreshState == REF_IDLE) {
1284 // remember when the refresh is due
1285 refreshDueAt = curTick();
1286
1287 // proceed to drain
1288 refreshState = REF_DRAIN;
1289
1290 DPRINTF(DRAM, "Refresh due\n");
1291 }
1292
1293 // let any scheduled read or write go ahead, after which it will
1294 // hand control back to this event loop
1295 if (refreshState == REF_DRAIN) {
1296 if (nextReqEvent.scheduled()) {
1297 // hand control over to the request loop until it is
1298 // evaluated next
1299 DPRINTF(DRAM, "Refresh awaiting draining\n");
1300
1301 return;
1302 } else {
1303 refreshState = REF_PRE;
1304 }
1305 }
1306
1307 // at this point, ensure that all banks are precharged
1308 if (refreshState == REF_PRE) {
1309 // precharge any active bank if we are not already in the idle
1310 // state
1311 if (pwrState != PWR_IDLE) {
1312 // at the moment, we use a precharge all even if there is
1313 // only a single bank open
1314 DPRINTF(DRAM, "Precharging all\n");
1315
1316 // first determine when we can precharge
1317 Tick pre_at = curTick();
1318 for (int i = 0; i < ranksPerChannel; i++) {
1319 for (int j = 0; j < banksPerRank; j++) {
1320 // respect both causality and any existing bank
1321 // constraints, some banks could already have a
1322 // (auto) precharge scheduled
1323 pre_at = std::max(banks[i][j].preAllowedAt, pre_at);
1324 }
1325 }
1326
1327 // make sure all banks are precharged, and for those that
1328 // already are, update their availability
1329 Tick act_allowed_at = pre_at + tRP;
1330
1331 for (int i = 0; i < ranksPerChannel; i++) {
1332 for (int j = 0; j < banksPerRank; j++) {
1333 if (banks[i][j].openRow != Bank::NO_ROW) {
1334 prechargeBank(banks[i][j], pre_at, false);
1335 } else {
1336 banks[i][j].actAllowedAt =
1337 std::max(banks[i][j].actAllowedAt, act_allowed_at);
1338 banks[i][j].preAllowedAt =
1339 std::max(banks[i][j].preAllowedAt, pre_at);
1340 }
1341 }
1342
1343 // at the moment this affects all ranks
1344 DPRINTF(DRAMPower, "%llu,PREA,0,%d\n", divCeil(pre_at, tCK),
1345 i);
1346 }
1347 } else {
1348 DPRINTF(DRAM, "All banks already precharged, starting refresh\n");
1349
1350 // go ahead and kick the power state machine into gear if
1351 // we are already idle
1352 schedulePowerEvent(PWR_REF, curTick());
1353 }
1354
1355 refreshState = REF_RUN;
1356 assert(numBanksActive == 0);
1357
1358 // wait for all banks to be precharged, at which point the
1359 // power state machine will transition to the idle state, and
1360 // automatically move to a refresh, at that point it will also
1361 // call this method to get the refresh event loop going again
1362 return;
1363 }
1364
1365 // last but not least we perform the actual refresh
1366 if (refreshState == REF_RUN) {
1367 // should never get here with any banks active
1368 assert(numBanksActive == 0);
1369 assert(pwrState == PWR_REF);
1370
1371 Tick ref_done_at = curTick() + tRFC;
1372
1373 for (int i = 0; i < ranksPerChannel; i++) {
1374 for (int j = 0; j < banksPerRank; j++) {
1375 banks[i][j].actAllowedAt = ref_done_at;
1376 }
1377
1378 // at the moment this affects all ranks
1379 DPRINTF(DRAMPower, "%llu,REF,0,%d\n", divCeil(curTick(), tCK), i);
1380 }
1381
1382 // make sure we did not wait so long that we cannot make up
1383 // for it
1384 if (refreshDueAt + tREFI < ref_done_at) {
1385 fatal("Refresh was delayed so long we cannot catch up\n");
1386 }
1387
1388 // compensate for the delay in actually performing the refresh
1389 // when scheduling the next one
1390 schedule(refreshEvent, refreshDueAt + tREFI - tRP);
1391
1392 assert(!powerEvent.scheduled());
1393
1394 // move to the idle power state once the refresh is done, this
1395 // will also move the refresh state machine to the refresh
1396 // idle state
1397 schedulePowerEvent(PWR_IDLE, ref_done_at);
1398
1399 DPRINTF(DRAMState, "Refresh done at %llu and next refresh at %llu\n",
1400 ref_done_at, refreshDueAt + tREFI);
1401 }
1402}
1403
1404void
1405DRAMCtrl::schedulePowerEvent(PowerState pwr_state, Tick tick)
1406{
1407 // respect causality
1408 assert(tick >= curTick());
1409
1410 if (!powerEvent.scheduled()) {
1411 DPRINTF(DRAMState, "Scheduling power event at %llu to state %d\n",
1412 tick, pwr_state);
1413
1414 // insert the new transition
1415 pwrStateTrans = pwr_state;
1416
1417 schedule(powerEvent, tick);
1418 } else {
1419 panic("Scheduled power event at %llu to state %d, "
1420 "with scheduled event at %llu to %d\n", tick, pwr_state,
1421 powerEvent.when(), pwrStateTrans);
1422 }
1423}
1424
1425void
1426DRAMCtrl::processPowerEvent()
1427{
1428 // remember where we were, and for how long
1429 Tick duration = curTick() - pwrStateTick;
1430 PowerState prev_state = pwrState;
1431
1432 // update the accounting
1433 pwrStateTime[prev_state] += duration;
1434
1435 pwrState = pwrStateTrans;
1436 pwrStateTick = curTick();
1437
1438 if (pwrState == PWR_IDLE) {
1439 DPRINTF(DRAMState, "All banks precharged\n");
1440
1441 // if we were refreshing, make sure we start scheduling requests again
1442 if (prev_state == PWR_REF) {
1443 DPRINTF(DRAMState, "Was refreshing for %llu ticks\n", duration);
1444 assert(pwrState == PWR_IDLE);
1445
1446 // kick things into action again
1447 refreshState = REF_IDLE;
1448 assert(!nextReqEvent.scheduled());
1449 schedule(nextReqEvent, curTick());
1450 } else {
1451 assert(prev_state == PWR_ACT);
1452
1453 // if we have a pending refresh, and are now moving to
1454 // the idle state, direclty transition to a refresh
1455 if (refreshState == REF_RUN) {
1456 // there should be nothing waiting at this point
1457 assert(!powerEvent.scheduled());
1458
1459 // update the state in zero time and proceed below
1460 pwrState = PWR_REF;
1461 }
1462 }
1463 }
1464
1465 // we transition to the refresh state, let the refresh state
1466 // machine know of this state update and let it deal with the
1467 // scheduling of the next power state transition as well as the
1468 // following refresh
1469 if (pwrState == PWR_REF) {
1470 DPRINTF(DRAMState, "Refreshing\n");
1471 // kick the refresh event loop into action again, and that
1472 // in turn will schedule a transition to the idle power
1473 // state once the refresh is done
1474 assert(refreshState == REF_RUN);
1475 processRefreshEvent();
1476 }
1477}
1478
1479void
1480DRAMCtrl::regStats()
1481{
1482 using namespace Stats;
1483
1484 AbstractMemory::regStats();
1485
1486 readReqs
1487 .name(name() + ".readReqs")
1488 .desc("Number of read requests accepted");
1489
1490 writeReqs
1491 .name(name() + ".writeReqs")
1492 .desc("Number of write requests accepted");
1493
1494 readBursts
1495 .name(name() + ".readBursts")
1496 .desc("Number of DRAM read bursts, "
1497 "including those serviced by the write queue");
1498
1499 writeBursts
1500 .name(name() + ".writeBursts")
1501 .desc("Number of DRAM write bursts, "
1502 "including those merged in the write queue");
1503
1504 servicedByWrQ
1505 .name(name() + ".servicedByWrQ")
1506 .desc("Number of DRAM read bursts serviced by the write queue");
1507
1508 mergedWrBursts
1509 .name(name() + ".mergedWrBursts")
1510 .desc("Number of DRAM write bursts merged with an existing one");
1511
1512 neitherReadNorWrite
1513 .name(name() + ".neitherReadNorWriteReqs")
1514 .desc("Number of requests that are neither read nor write");
1515
1516 perBankRdBursts
1517 .init(banksPerRank * ranksPerChannel)
1518 .name(name() + ".perBankRdBursts")
1519 .desc("Per bank write bursts");
1520
1521 perBankWrBursts
1522 .init(banksPerRank * ranksPerChannel)
1523 .name(name() + ".perBankWrBursts")
1524 .desc("Per bank write bursts");
1525
1526 avgRdQLen
1527 .name(name() + ".avgRdQLen")
1528 .desc("Average read queue length when enqueuing")
1529 .precision(2);
1530
1531 avgWrQLen
1532 .name(name() + ".avgWrQLen")
1533 .desc("Average write queue length when enqueuing")
1534 .precision(2);
1535
1536 totQLat
1537 .name(name() + ".totQLat")
1538 .desc("Total ticks spent queuing");
1539
1540 totBusLat
1541 .name(name() + ".totBusLat")
1542 .desc("Total ticks spent in databus transfers");
1543
1544 totMemAccLat
1545 .name(name() + ".totMemAccLat")
1546 .desc("Total ticks spent from burst creation until serviced "
1547 "by the DRAM");
1548
1549 avgQLat
1550 .name(name() + ".avgQLat")
1551 .desc("Average queueing delay per DRAM burst")
1552 .precision(2);
1553
1554 avgQLat = totQLat / (readBursts - servicedByWrQ);
1555
1556 avgBusLat
1557 .name(name() + ".avgBusLat")
1558 .desc("Average bus latency per DRAM burst")
1559 .precision(2);
1560
1561 avgBusLat = totBusLat / (readBursts - servicedByWrQ);
1562
1563 avgMemAccLat
1564 .name(name() + ".avgMemAccLat")
1565 .desc("Average memory access latency per DRAM burst")
1566 .precision(2);
1567
1568 avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
1569
1570 numRdRetry
1571 .name(name() + ".numRdRetry")
1572 .desc("Number of times read queue was full causing retry");
1573
1574 numWrRetry
1575 .name(name() + ".numWrRetry")
1576 .desc("Number of times write queue was full causing retry");
1577
1578 readRowHits
1579 .name(name() + ".readRowHits")
1580 .desc("Number of row buffer hits during reads");
1581
1582 writeRowHits
1583 .name(name() + ".writeRowHits")
1584 .desc("Number of row buffer hits during writes");
1585
1586 readRowHitRate
1587 .name(name() + ".readRowHitRate")
1588 .desc("Row buffer hit rate for reads")
1589 .precision(2);
1590
1591 readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
1592
1593 writeRowHitRate
1594 .name(name() + ".writeRowHitRate")
1595 .desc("Row buffer hit rate for writes")
1596 .precision(2);
1597
1598 writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
1599
1600 readPktSize
1601 .init(ceilLog2(burstSize) + 1)
1602 .name(name() + ".readPktSize")
1603 .desc("Read request sizes (log2)");
1604
1605 writePktSize
1606 .init(ceilLog2(burstSize) + 1)
1607 .name(name() + ".writePktSize")
1608 .desc("Write request sizes (log2)");
1609
1610 rdQLenPdf
1611 .init(readBufferSize)
1612 .name(name() + ".rdQLenPdf")
1613 .desc("What read queue length does an incoming req see");
1614
1615 wrQLenPdf
1616 .init(writeBufferSize)
1617 .name(name() + ".wrQLenPdf")
1618 .desc("What write queue length does an incoming req see");
1619
1620 bytesPerActivate
1621 .init(maxAccessesPerRow)
1622 .name(name() + ".bytesPerActivate")
1623 .desc("Bytes accessed per row activation")
1624 .flags(nozero);
1625
1626 rdPerTurnAround
1627 .init(readBufferSize)
1628 .name(name() + ".rdPerTurnAround")
1629 .desc("Reads before turning the bus around for writes")
1630 .flags(nozero);
1631
1632 wrPerTurnAround
1633 .init(writeBufferSize)
1634 .name(name() + ".wrPerTurnAround")
1635 .desc("Writes before turning the bus around for reads")
1636 .flags(nozero);
1637
1638 bytesReadDRAM
1639 .name(name() + ".bytesReadDRAM")
1640 .desc("Total number of bytes read from DRAM");
1641
1642 bytesReadWrQ
1643 .name(name() + ".bytesReadWrQ")
1644 .desc("Total number of bytes read from write queue");
1645
1646 bytesWritten
1647 .name(name() + ".bytesWritten")
1648 .desc("Total number of bytes written to DRAM");
1649
1650 bytesReadSys
1651 .name(name() + ".bytesReadSys")
1652 .desc("Total read bytes from the system interface side");
1653
1654 bytesWrittenSys
1655 .name(name() + ".bytesWrittenSys")
1656 .desc("Total written bytes from the system interface side");
1657
1658 avgRdBW
1659 .name(name() + ".avgRdBW")
1660 .desc("Average DRAM read bandwidth in MiByte/s")
1661 .precision(2);
1662
1663 avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
1664
1665 avgWrBW
1666 .name(name() + ".avgWrBW")
1667 .desc("Average achieved write bandwidth in MiByte/s")
1668 .precision(2);
1669
1670 avgWrBW = (bytesWritten / 1000000) / simSeconds;
1671
1672 avgRdBWSys
1673 .name(name() + ".avgRdBWSys")
1674 .desc("Average system read bandwidth in MiByte/s")
1675 .precision(2);
1676
1677 avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
1678
1679 avgWrBWSys
1680 .name(name() + ".avgWrBWSys")
1681 .desc("Average system write bandwidth in MiByte/s")
1682 .precision(2);
1683
1684 avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
1685
1686 peakBW
1687 .name(name() + ".peakBW")
1688 .desc("Theoretical peak bandwidth in MiByte/s")
1689 .precision(2);
1690
1691 peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
1692
1693 busUtil
1694 .name(name() + ".busUtil")
1695 .desc("Data bus utilization in percentage")
1696 .precision(2);
1697
1698 busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
1699
1700 totGap
1701 .name(name() + ".totGap")
1702 .desc("Total gap between requests");
1703
1704 avgGap
1705 .name(name() + ".avgGap")
1706 .desc("Average gap between requests")
1707 .precision(2);
1708
1709 avgGap = totGap / (readReqs + writeReqs);
1710
1711 // Stats for DRAM Power calculation based on Micron datasheet
1712 busUtilRead
1713 .name(name() + ".busUtilRead")
1714 .desc("Data bus utilization in percentage for reads")
1715 .precision(2);
1716
1717 busUtilRead = avgRdBW / peakBW * 100;
1718
1719 busUtilWrite
1720 .name(name() + ".busUtilWrite")
1721 .desc("Data bus utilization in percentage for writes")
1722 .precision(2);
1723
1724 busUtilWrite = avgWrBW / peakBW * 100;
1725
1726 pageHitRate
1727 .name(name() + ".pageHitRate")
1728 .desc("Row buffer hit rate, read and write combined")
1729 .precision(2);
1730
1731 pageHitRate = (writeRowHits + readRowHits) /
1732 (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
1733
1734 pwrStateTime
1735 .init(5)
1736 .name(name() + ".memoryStateTime")
1737 .desc("Time in different power states");
1738 pwrStateTime.subname(0, "IDLE");
1739 pwrStateTime.subname(1, "REF");
1740 pwrStateTime.subname(2, "PRE_PDN");
1741 pwrStateTime.subname(3, "ACT");
1742 pwrStateTime.subname(4, "ACT_PDN");
1743}
1744
1745void
1746DRAMCtrl::recvFunctional(PacketPtr pkt)
1747{
1748 // rely on the abstract memory
1749 functionalAccess(pkt);
1750}
1751
1752BaseSlavePort&
1753DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
1754{
1755 if (if_name != "port") {
1756 return MemObject::getSlavePort(if_name, idx);
1757 } else {
1758 return port;
1759 }
1760}
1761
1762unsigned int
1763DRAMCtrl::drain(DrainManager *dm)
1764{
1765 unsigned int count = port.drain(dm);
1766
1767 // if there is anything in any of our internal queues, keep track
1768 // of that as well
1769 if (!(writeQueue.empty() && readQueue.empty() &&
1770 respQueue.empty())) {
1771 DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
1772 " resp: %d\n", writeQueue.size(), readQueue.size(),
1773 respQueue.size());
1774 ++count;
1775 drainManager = dm;
1776
1777 // the only part that is not drained automatically over time
1778 // is the write queue, thus kick things into action if needed
1779 if (!writeQueue.empty() && !nextReqEvent.scheduled()) {
1780 schedule(nextReqEvent, curTick());
1781 }
1782 }
1783
1784 if (count)
1785 setDrainState(Drainable::Draining);
1786 else
1787 setDrainState(Drainable::Drained);
1788 return count;
1789}
1790
1791DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
1792 : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
1793 memory(_memory)
1794{ }
1795
1796AddrRangeList
1797DRAMCtrl::MemoryPort::getAddrRanges() const
1798{
1799 AddrRangeList ranges;
1800 ranges.push_back(memory.getAddrRange());
1801 return ranges;
1802}
1803
1804void
1805DRAMCtrl::MemoryPort::recvFunctional(PacketPtr pkt)
1806{
1807 pkt->pushLabel(memory.name());
1808
1809 if (!queue.checkFunctional(pkt)) {
1810 // Default implementation of SimpleTimingPort::recvFunctional()
1811 // calls recvAtomic() and throws away the latency; we can save a
1812 // little here by just not calculating the latency.
1813 memory.recvFunctional(pkt);
1814 }
1815
1816 pkt->popLabel();
1817}
1818
1819Tick
1820DRAMCtrl::MemoryPort::recvAtomic(PacketPtr pkt)
1821{
1822 return memory.recvAtomic(pkt);
1823}
1824
1825bool
1826DRAMCtrl::MemoryPort::recvTimingReq(PacketPtr pkt)
1827{
1828 // pass it to the memory controller
1829 return memory.recvTimingReq(pkt);
1830}
1831
1832DRAMCtrl*
1833DRAMCtrlParams::create()
1834{
1835 return new DRAMCtrl(this);
1836}