1/*
2 * Copyright (c) 2010-2013 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/Drain.hh"
49#include "mem/dram_ctrl.hh"
50#include "sim/system.hh"
51
52using namespace std;
53
54DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
55 AbstractMemory(p),
56 port(name() + ".port", *this),
57 retryRdReq(false), retryWrReq(false),
58 rowHitFlag(false), stopReads(false),
59 writeEvent(this), respondEvent(this),
60 refreshEvent(this), nextReqEvent(this), drainManager(NULL),
61 deviceBusWidth(p->device_bus_width), burstLength(p->burst_length),
62 deviceRowBufferSize(p->device_rowbuffer_size),
63 devicesPerRank(p->devices_per_rank),
64 burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8),
65 rowBufferSize(devicesPerRank * deviceRowBufferSize),
66 columnsPerRowBuffer(rowBufferSize / burstSize),
67 ranksPerChannel(p->ranks_per_channel),
68 banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0),
69 readBufferSize(p->read_buffer_size),
70 writeBufferSize(p->write_buffer_size),
71 writeHighThreshold(writeBufferSize * p->write_high_thresh_perc / 100.0),
72 writeLowThreshold(writeBufferSize * p->write_low_thresh_perc / 100.0),
73 minWritesPerSwitch(p->min_writes_per_switch), writesThisTime(0),
74 tWTR(p->tWTR), tBURST(p->tBURST),
75 tRCD(p->tRCD), tCL(p->tCL), tRP(p->tRP), tRAS(p->tRAS),
76 tRFC(p->tRFC), tREFI(p->tREFI), tRRD(p->tRRD),
77 tXAW(p->tXAW), activationLimit(p->activation_limit),
78 memSchedPolicy(p->mem_sched_policy), addrMapping(p->addr_mapping),
79 pageMgmt(p->page_policy),
80 maxAccessesPerRow(p->max_accesses_per_row),
81 frontendLatency(p->static_frontend_latency),
82 backendLatency(p->static_backend_latency),
83 busBusyUntil(0), prevArrival(0),
84 newTime(0), startTickPrechargeAll(0), numBanksActive(0)
85{
86 // create the bank states based on the dimensions of the ranks and
87 // banks
88 banks.resize(ranksPerChannel);
89 actTicks.resize(ranksPerChannel);
90 for (size_t c = 0; c < ranksPerChannel; ++c) {
91 banks[c].resize(banksPerRank);
92 actTicks[c].resize(activationLimit, 0);
93 }
94
95 // perform a basic check of the write thresholds
96 if (p->write_low_thresh_perc >= p->write_high_thresh_perc)
97 fatal("Write buffer low threshold %d must be smaller than the "
98 "high threshold %d\n", p->write_low_thresh_perc,
99 p->write_high_thresh_perc);
100
101 // determine the rows per bank by looking at the total capacity
102 uint64_t capacity = ULL(1) << ceilLog2(AbstractMemory::size());
103
104 DPRINTF(DRAM, "Memory capacity %lld (%lld) bytes\n", capacity,
105 AbstractMemory::size());
106
107 DPRINTF(DRAM, "Row buffer size %d bytes with %d columns per row buffer\n",
108 rowBufferSize, columnsPerRowBuffer);
109
110 rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel);
111
112 if (range.interleaved()) {
113 if (channels != range.stripes())
114 fatal("%s has %d interleaved address stripes but %d channel(s)\n",
115 name(), range.stripes(), channels);
116
117 if (addrMapping == Enums::RoRaBaChCo) {
118 if (rowBufferSize != range.granularity()) {
119 fatal("Interleaving of %s doesn't match RoRaBaChCo "
120 "address map\n", name());
121 }
122 } else if (addrMapping == Enums::RoRaBaCoCh) {
123 if (system()->cacheLineSize() != range.granularity()) {
124 fatal("Interleaving of %s doesn't match RoRaBaCoCh "
125 "address map\n", name());
126 }
127 } else if (addrMapping == Enums::RoCoRaBaCh) {
128 if (system()->cacheLineSize() != range.granularity())
129 fatal("Interleaving of %s doesn't match RoCoRaBaCh "
130 "address map\n", name());
131 }
132 }
133}
134
135void
136DRAMCtrl::init()
137{
138 if (!port.isConnected()) {
139 fatal("DRAMCtrl %s is unconnected!\n", name());
140 } else {
141 port.sendRangeChange();
142 }
143}
144
145void
146DRAMCtrl::startup()
147{
148 // update the start tick for the precharge accounting to the
149 // current tick
150 startTickPrechargeAll = curTick();
151
152 // print the configuration of the controller
153 printParams();
154
155 // kick off the refresh
156 schedule(refreshEvent, curTick() + tREFI);
157}
158
159Tick
160DRAMCtrl::recvAtomic(PacketPtr pkt)
161{
162 DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
163
164 // do the actual memory access and turn the packet into a response
165 access(pkt);
166
167 Tick latency = 0;
168 if (!pkt->memInhibitAsserted() && pkt->hasData()) {
169 // this value is not supposed to be accurate, just enough to
170 // keep things going, mimic a closed page
171 latency = tRP + tRCD + tCL;
172 }
173 return latency;
174}
175
176bool
177DRAMCtrl::readQueueFull(unsigned int neededEntries) const
178{
179 DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
180 readBufferSize, readQueue.size() + respQueue.size(),
181 neededEntries);
182
183 return
184 (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
185}
186
187bool
188DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
189{
190 DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
191 writeBufferSize, writeQueue.size(), neededEntries);
192 return (writeQueue.size() + neededEntries) > writeBufferSize;
193}
194
195DRAMCtrl::DRAMPacket*
196DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
197 bool isRead)
198{
199 // decode the address based on the address mapping scheme, with
200 // Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
201 // channel, respectively
202 uint8_t rank;
203 uint8_t bank;
204 uint16_t row;
205
206 // truncate the address to the access granularity
207 Addr addr = dramPktAddr / burstSize;
208
209 // we have removed the lowest order address bits that denote the
210 // position within the column
211 if (addrMapping == Enums::RoRaBaChCo) {
212 // the lowest order bits denote the column to ensure that
213 // sequential cache lines occupy the same row
214 addr = addr / columnsPerRowBuffer;
215
216 // take out the channel part of the address
217 addr = addr / channels;
218
219 // after the channel bits, get the bank bits to interleave
220 // over the banks
221 bank = addr % banksPerRank;
222 addr = addr / banksPerRank;
223
224 // after the bank, we get the rank bits which thus interleaves
225 // over the ranks
226 rank = addr % ranksPerChannel;
227 addr = addr / ranksPerChannel;
228
229 // lastly, get the row bits
230 row = addr % rowsPerBank;
231 addr = addr / rowsPerBank;
232 } else if (addrMapping == Enums::RoRaBaCoCh) {
233 // take out the channel part of the address
234 addr = addr / channels;
235
236 // next, the column
237 addr = addr / columnsPerRowBuffer;
238
239 // after the column bits, we get the bank bits to interleave
240 // over the banks
241 bank = addr % banksPerRank;
242 addr = addr / banksPerRank;
243
244 // after the bank, we get the rank bits which thus interleaves
245 // over the ranks
246 rank = addr % ranksPerChannel;
247 addr = addr / ranksPerChannel;
248
249 // lastly, get the row bits
250 row = addr % rowsPerBank;
251 addr = addr / rowsPerBank;
252 } else if (addrMapping == Enums::RoCoRaBaCh) {
253 // optimise for closed page mode and utilise maximum
254 // parallelism of the DRAM (at the cost of power)
255
256 // take out the channel part of the address, not that this has
257 // to match with how accesses are interleaved between the
258 // controllers in the address mapping
259 addr = addr / channels;
260
261 // start with the bank bits, as this provides the maximum
262 // opportunity for parallelism between requests
263 bank = addr % banksPerRank;
264 addr = addr / banksPerRank;
265
266 // next get the rank bits
267 rank = addr % ranksPerChannel;
268 addr = addr / ranksPerChannel;
269
270 // next the column bits which we do not need to keep track of
271 // and simply skip past
272 addr = addr / columnsPerRowBuffer;
273
274 // lastly, get the row bits
275 row = addr % rowsPerBank;
276 addr = addr / rowsPerBank;
277 } else
278 panic("Unknown address mapping policy chosen!");
279
280 assert(rank < ranksPerChannel);
281 assert(bank < banksPerRank);
282 assert(row < rowsPerBank);
283
284 DPRINTF(DRAM, "Address: %lld Rank %d Bank %d Row %d\n",
285 dramPktAddr, rank, bank, row);
286
287 // create the corresponding DRAM packet with the entry time and
288 // ready time set to the current tick, the latter will be updated
289 // later
290 uint16_t bank_id = banksPerRank * rank + bank;
291 return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
292 size, banks[rank][bank]);
293}
294
295void
296DRAMCtrl::addToReadQueue(PacketPtr pkt, unsigned int pktCount)
297{
298 // only add to the read queue here. whenever the request is
299 // eventually done, set the readyTime, and call schedule()
300 assert(!pkt->isWrite());
301
302 assert(pktCount != 0);
303
304 // if the request size is larger than burst size, the pkt is split into
305 // multiple DRAM packets
306 // Note if the pkt starting address is not aligened to burst size, the
307 // address of first DRAM packet is kept unaliged. Subsequent DRAM packets
308 // are aligned to burst size boundaries. This is to ensure we accurately
309 // check read packets against packets in write queue.
310 Addr addr = pkt->getAddr();
311 unsigned pktsServicedByWrQ = 0;
312 BurstHelper* burst_helper = NULL;
313 for (int cnt = 0; cnt < pktCount; ++cnt) {
314 unsigned size = std::min((addr | (burstSize - 1)) + 1,
315 pkt->getAddr() + pkt->getSize()) - addr;
316 readPktSize[ceilLog2(size)]++;
317 readBursts++;
318
319 // First check write buffer to see if the data is already at
320 // the controller
321 bool foundInWrQ = false;
322 for (auto i = writeQueue.begin(); i != writeQueue.end(); ++i) {
323 // check if the read is subsumed in the write entry we are
324 // looking at
325 if ((*i)->addr <= addr &&
326 (addr + size) <= ((*i)->addr + (*i)->size)) {
327 foundInWrQ = true;
328 servicedByWrQ++;
329 pktsServicedByWrQ++;
330 DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
331 "write queue\n", addr, size);
332 bytesReadWrQ += burstSize;
333 break;
334 }
335 }
336
337 // If not found in the write q, make a DRAM packet and
338 // push it onto the read queue
339 if (!foundInWrQ) {
340
341 // Make the burst helper for split packets
342 if (pktCount > 1 && burst_helper == NULL) {
343 DPRINTF(DRAM, "Read to addr %lld translates to %d "
344 "dram requests\n", pkt->getAddr(), pktCount);
345 burst_helper = new BurstHelper(pktCount);
346 }
347
348 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, true);
349 dram_pkt->burstHelper = burst_helper;
350
351 assert(!readQueueFull(1));
352 rdQLenPdf[readQueue.size() + respQueue.size()]++;
353
354 DPRINTF(DRAM, "Adding to read queue\n");
355
356 readQueue.push_back(dram_pkt);
357
358 // Update stats
359 avgRdQLen = readQueue.size() + respQueue.size();
360 }
361
362 // Starting address of next dram pkt (aligend to burstSize boundary)
363 addr = (addr | (burstSize - 1)) + 1;
364 }
365
366 // If all packets are serviced by write queue, we send the repsonse back
367 if (pktsServicedByWrQ == pktCount) {
368 accessAndRespond(pkt, frontendLatency);
369 return;
370 }
371
372 // Update how many split packets are serviced by write queue
373 if (burst_helper != NULL)
374 burst_helper->burstsServiced = pktsServicedByWrQ;
375
376 // If we are not already scheduled to get the read request out of
377 // the queue, do so now
378 if (!nextReqEvent.scheduled() && !stopReads) {
379 DPRINTF(DRAM, "Request scheduled immediately\n");
380 schedule(nextReqEvent, curTick());
381 }
382}
383
384void
385DRAMCtrl::processWriteEvent()
386{
387 assert(!writeQueue.empty());
388
389 DPRINTF(DRAM, "Beginning DRAM Write\n");
390 Tick temp1 M5_VAR_USED = std::max(curTick(), busBusyUntil);
391 Tick temp2 M5_VAR_USED = std::max(curTick(), maxBankFreeAt());
392
393 chooseNextWrite();
394 DRAMPacket* dram_pkt = writeQueue.front();
395 // sanity check
396 assert(dram_pkt->size <= burstSize);
397 doDRAMAccess(dram_pkt);
398
399 writeQueue.pop_front();
400 delete dram_pkt;
401
402 ++writesThisTime;
403
404 DPRINTF(DRAM, "Writing, bus busy for %lld ticks, banks busy "
405 "for %lld ticks\n", busBusyUntil - temp1, maxBankFreeAt() - temp2);
406
407 // If we emptied the write queue, or got below the threshold and
408 // are not draining, or we have reads waiting and have done enough
409 // writes, then switch to reads. The retry above could already
410 // have caused it to be scheduled, so first check
411 if (writeQueue.empty() ||
412 (writeQueue.size() < writeLowThreshold && !drainManager) ||
413 (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
414 // turn the bus back around for reads again
415 busBusyUntil += tWTR;
416 stopReads = false;
417 writesThisTime = 0;
418
419 if (!nextReqEvent.scheduled())
420 schedule(nextReqEvent, busBusyUntil);
421 } else {
422 assert(!writeEvent.scheduled());
423 DPRINTF(DRAM, "Next write scheduled at %lld\n", newTime);
424 schedule(writeEvent, newTime);
425 }
426
427 if (retryWrReq) {
428 retryWrReq = false;
429 port.sendRetry();
430 }
431
432 // if there is nothing left in any queue, signal a drain
433 if (writeQueue.empty() && readQueue.empty() &&
434 respQueue.empty () && drainManager) {
435 drainManager->signalDrainDone();
436 drainManager = NULL;
437 }
438}
439
440
441void
442DRAMCtrl::triggerWrites()
443{
444 DPRINTF(DRAM, "Writes triggered at %lld\n", curTick());
445 // Flag variable to stop any more read scheduling
446 stopReads = true;
447
448 Tick write_start_time = std::max(busBusyUntil, curTick()) + tWTR;
449
450 DPRINTF(DRAM, "Writes scheduled at %lld\n", write_start_time);
451
452 assert(write_start_time >= curTick());
453 assert(!writeEvent.scheduled());
454 schedule(writeEvent, write_start_time);
455}
456
457void
458DRAMCtrl::addToWriteQueue(PacketPtr pkt, unsigned int pktCount)
459{
460 // only add to the write queue here. whenever the request is
461 // eventually done, set the readyTime, and call schedule()
462 assert(pkt->isWrite());
463
464 // if the request size is larger than burst size, the pkt is split into
465 // multiple DRAM packets
466 Addr addr = pkt->getAddr();
467 for (int cnt = 0; cnt < pktCount; ++cnt) {
468 unsigned size = std::min((addr | (burstSize - 1)) + 1,
469 pkt->getAddr() + pkt->getSize()) - addr;
470 writePktSize[ceilLog2(size)]++;
471 writeBursts++;
472
473 // see if we can merge with an existing item in the write
474 // queue and keep track of whether we have merged or not so we
475 // can stop at that point and also avoid enqueueing a new
476 // request
477 bool merged = false;
478 auto w = writeQueue.begin();
479
480 while(!merged && w != writeQueue.end()) {
481 // either of the two could be first, if they are the same
482 // it does not matter which way we go
483 if ((*w)->addr >= addr) {
484 // the existing one starts after the new one, figure
485 // out where the new one ends with respect to the
486 // existing one
487 if ((addr + size) >= ((*w)->addr + (*w)->size)) {
488 // check if the existing one is completely
489 // subsumed in the new one
490 DPRINTF(DRAM, "Merging write covering existing burst\n");
491 merged = true;
492 // update both the address and the size
493 (*w)->addr = addr;
494 (*w)->size = size;
495 } else if ((addr + size) >= (*w)->addr &&
496 ((*w)->addr + (*w)->size - addr) <= burstSize) {
497 // the new one is just before or partially
498 // overlapping with the existing one, and together
499 // they fit within a burst
500 DPRINTF(DRAM, "Merging write before existing burst\n");
501 merged = true;
502 // the existing queue item needs to be adjusted with
503 // respect to both address and size
504 (*w)->size = (*w)->addr + (*w)->size - addr;
505 (*w)->addr = addr;
506 }
507 } else {
508 // the new one starts after the current one, figure
509 // out where the existing one ends with respect to the
510 // new one
511 if (((*w)->addr + (*w)->size) >= (addr + size)) {
512 // check if the new one is completely subsumed in the
513 // existing one
514 DPRINTF(DRAM, "Merging write into existing burst\n");
515 merged = true;
516 // no adjustments necessary
517 } else if (((*w)->addr + (*w)->size) >= addr &&
518 (addr + size - (*w)->addr) <= burstSize) {
519 // the existing one is just before or partially
520 // overlapping with the new one, and together
521 // they fit within a burst
522 DPRINTF(DRAM, "Merging write after existing burst\n");
523 merged = true;
524 // the address is right, and only the size has
525 // to be adjusted
526 (*w)->size = addr + size - (*w)->addr;
527 }
528 }
529 ++w;
530 }
531
532 // if the item was not merged we need to create a new write
533 // and enqueue it
534 if (!merged) {
535 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
536
537 assert(writeQueue.size() < writeBufferSize);
538 wrQLenPdf[writeQueue.size()]++;
539
540 DPRINTF(DRAM, "Adding to write queue\n");
541
542 writeQueue.push_back(dram_pkt);
543
544 // Update stats
545 avgWrQLen = writeQueue.size();
546 } else {
547 // keep track of the fact that this burst effectively
548 // disappeared as it was merged with an existing one
549 mergedWrBursts++;
550 }
551
552 // Starting address of next dram pkt (aligend to burstSize boundary)
553 addr = (addr | (burstSize - 1)) + 1;
554 }
555
556 // we do not wait for the writes to be send to the actual memory,
557 // but instead take responsibility for the consistency here and
558 // snoop the write queue for any upcoming reads
559 // @todo, if a pkt size is larger than burst size, we might need a
560 // different front end latency
561 accessAndRespond(pkt, frontendLatency);
562
563 // If your write buffer is starting to fill up, drain it!
564 if (writeQueue.size() >= writeHighThreshold && !stopReads){
565 triggerWrites();
566 }
567}
568
569void
570DRAMCtrl::printParams() const
571{
572 // Sanity check print of important parameters
573 DPRINTF(DRAM,
574 "Memory controller %s physical organization\n" \
575 "Number of devices per rank %d\n" \
576 "Device bus width (in bits) %d\n" \
577 "DRAM data bus burst (bytes) %d\n" \
578 "Row buffer size (bytes) %d\n" \
579 "Columns per row buffer %d\n" \
580 "Rows per bank %d\n" \
581 "Banks per rank %d\n" \
582 "Ranks per channel %d\n" \
583 "Total mem capacity (bytes) %u\n",
584 name(), devicesPerRank, deviceBusWidth, burstSize, rowBufferSize,
585 columnsPerRowBuffer, rowsPerBank, banksPerRank, ranksPerChannel,
586 rowBufferSize * rowsPerBank * banksPerRank * ranksPerChannel);
587
588 string scheduler = memSchedPolicy == Enums::fcfs ? "FCFS" : "FR-FCFS";
589 string address_mapping = addrMapping == Enums::RoRaBaChCo ? "RoRaBaChCo" :
590 (addrMapping == Enums::RoRaBaCoCh ? "RoRaBaCoCh" : "RoCoRaBaCh");
591 string page_policy = pageMgmt == Enums::open ? "OPEN" :
592 (pageMgmt == Enums::open_adaptive ? "OPEN (adaptive)" :
593 (pageMgmt == Enums::close_adaptive ? "CLOSE (adaptive)" : "CLOSE"));
594
595 DPRINTF(DRAM,
596 "Memory controller %s characteristics\n" \
597 "Read buffer size %d\n" \
598 "Write buffer size %d\n" \
599 "Write high thresh %d\n" \
600 "Write low thresh %d\n" \
601 "Scheduler %s\n" \
602 "Address mapping %s\n" \
603 "Page policy %s\n",
604 name(), readBufferSize, writeBufferSize, writeHighThreshold,
605 writeLowThreshold, scheduler, address_mapping, page_policy);
606
607 DPRINTF(DRAM, "Memory controller %s timing specs\n" \
608 "tRCD %d ticks\n" \
609 "tCL %d ticks\n" \
610 "tRP %d ticks\n" \
611 "tBURST %d ticks\n" \
612 "tRFC %d ticks\n" \
613 "tREFI %d ticks\n" \
614 "tWTR %d ticks\n" \
615 "tXAW (%d) %d ticks\n",
616 name(), tRCD, tCL, tRP, tBURST, tRFC, tREFI, tWTR,
617 activationLimit, tXAW);
618}
619
620void
621DRAMCtrl::printQs() const {
622 DPRINTF(DRAM, "===READ QUEUE===\n\n");
623 for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
624 DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
625 }
626 DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
627 for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
628 DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
629 }
630 DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
631 for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
632 DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
633 }
634}
635
636bool
637DRAMCtrl::recvTimingReq(PacketPtr pkt)
638{
639 /// @todo temporary hack to deal with memory corruption issues until
640 /// 4-phase transactions are complete
641 for (int x = 0; x < pendingDelete.size(); x++)
642 delete pendingDelete[x];
643 pendingDelete.clear();
644
645 // This is where we enter from the outside world
646 DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
647 pkt->cmdString(), pkt->getAddr(), pkt->getSize());
648
649 // simply drop inhibited packets for now
650 if (pkt->memInhibitAsserted()) {
651 DPRINTF(DRAM, "Inhibited packet -- Dropping it now\n");
652 pendingDelete.push_back(pkt);
653 return true;
654 }
655
656 // Calc avg gap between requests
657 if (prevArrival != 0) {
658 totGap += curTick() - prevArrival;
659 }
660 prevArrival = curTick();
661
662
663 // Find out how many dram packets a pkt translates to
664 // If the burst size is equal or larger than the pkt size, then a pkt
665 // translates to only one dram packet. Otherwise, a pkt translates to
666 // multiple dram packets
667 unsigned size = pkt->getSize();
668 unsigned offset = pkt->getAddr() & (burstSize - 1);
669 unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
670
671 // check local buffers and do not accept if full
672 if (pkt->isRead()) {
673 assert(size != 0);
674 if (readQueueFull(dram_pkt_count)) {
675 DPRINTF(DRAM, "Read queue full, not accepting\n");
676 // remember that we have to retry this port
677 retryRdReq = true;
678 numRdRetry++;
679 return false;
680 } else {
681 addToReadQueue(pkt, dram_pkt_count);
682 readReqs++;
683 bytesReadSys += size;
684 }
685 } else if (pkt->isWrite()) {
686 assert(size != 0);
687 if (writeQueueFull(dram_pkt_count)) {
688 DPRINTF(DRAM, "Write queue full, not accepting\n");
689 // remember that we have to retry this port
690 retryWrReq = true;
691 numWrRetry++;
692 return false;
693 } else {
694 addToWriteQueue(pkt, dram_pkt_count);
695 writeReqs++;
696 bytesWrittenSys += size;
697 }
698 } else {
699 DPRINTF(DRAM,"Neither read nor write, ignore timing\n");
700 neitherReadNorWrite++;
701 accessAndRespond(pkt, 1);
702 }
703
704 return true;
705}
706
707void
708DRAMCtrl::processRespondEvent()
709{
710 DPRINTF(DRAM,
711 "processRespondEvent(): Some req has reached its readyTime\n");
712
713 DRAMPacket* dram_pkt = respQueue.front();
714
715 if (dram_pkt->burstHelper) {
716 // it is a split packet
717 dram_pkt->burstHelper->burstsServiced++;
718 if (dram_pkt->burstHelper->burstsServiced ==
719 dram_pkt->burstHelper->burstCount) {
720 // we have now serviced all children packets of a system packet
721 // so we can now respond to the requester
722 // @todo we probably want to have a different front end and back
723 // end latency for split packets
724 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
725 delete dram_pkt->burstHelper;
726 dram_pkt->burstHelper = NULL;
727 }
728 } else {
729 // it is not a split packet
730 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
731 }
732
733 delete respQueue.front();
734 respQueue.pop_front();
735
736 if (!respQueue.empty()) {
737 assert(respQueue.front()->readyTime >= curTick());
738 assert(!respondEvent.scheduled());
739 schedule(respondEvent, respQueue.front()->readyTime);
740 } else {
741 // if there is nothing left in any queue, signal a drain
742 if (writeQueue.empty() && readQueue.empty() &&
743 drainManager) {
744 drainManager->signalDrainDone();
745 drainManager = NULL;
746 }
747 }
748
749 // We have made a location in the queue available at this point,
750 // so if there is a read that was forced to wait, retry now
751 if (retryRdReq) {
752 retryRdReq = false;
753 port.sendRetry();
754 }
755}
756
757void
758DRAMCtrl::chooseNextWrite()
759{
760 // This method does the arbitration between write requests. The
761 // chosen packet is simply moved to the head of the write
762 // queue. The other methods know that this is the place to
763 // look. For example, with FCFS, this method does nothing
764 assert(!writeQueue.empty());
765
766 if (writeQueue.size() == 1) {
767 DPRINTF(DRAM, "Single write request, nothing to do\n");
768 return;
769 }
770
771 if (memSchedPolicy == Enums::fcfs) {
772 // Do nothing, since the correct request is already head
773 } else if (memSchedPolicy == Enums::frfcfs) {
774 reorderQueue(writeQueue);
775 } else
776 panic("No scheduling policy chosen\n");
777
778 DPRINTF(DRAM, "Selected next write request\n");
779}
780
781bool
782DRAMCtrl::chooseNextRead()
783{
784 // This method does the arbitration between read requests. The
785 // chosen packet is simply moved to the head of the queue. The
786 // other methods know that this is the place to look. For example,
787 // with FCFS, this method does nothing
788 if (readQueue.empty()) {
789 DPRINTF(DRAM, "No read request to select\n");
790 return false;
791 }
792
793 // If there is only one request then there is nothing left to do
794 if (readQueue.size() == 1)
795 return true;
796
797 if (memSchedPolicy == Enums::fcfs) {
798 // Do nothing, since the request to serve is already the first
799 // one in the read queue
800 } else if (memSchedPolicy == Enums::frfcfs) {
801 reorderQueue(readQueue);
802 } else
803 panic("No scheduling policy chosen!\n");
804
805 DPRINTF(DRAM, "Selected next read request\n");
806 return true;
807}
808
809void
810DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue)
811{
812 // Only determine this when needed
813 uint64_t earliest_banks = 0;
814
815 // Search for row hits first, if no row hit is found then schedule the
816 // packet to one of the earliest banks available
817 bool found_earliest_pkt = false;
818 auto selected_pkt_it = queue.begin();
819
820 for (auto i = queue.begin(); i != queue.end() ; ++i) {
821 DRAMPacket* dram_pkt = *i;
822 const Bank& bank = dram_pkt->bankRef;
823 // Check if it is a row hit
824 if (bank.openRow == dram_pkt->row) {
825 DPRINTF(DRAM, "Row buffer hit\n");
826 selected_pkt_it = i;
827 break;
828 } else if (!found_earliest_pkt) {
829 // No row hit, go for first ready
830 if (earliest_banks == 0)
831 earliest_banks = minBankFreeAt(queue);
832
833 // Bank is ready or is the first available bank
834 if (bank.freeAt <= curTick() ||
835 bits(earliest_banks, dram_pkt->bankId, dram_pkt->bankId)) {
836 // Remember the packet to be scheduled to one of the earliest
837 // banks available
838 selected_pkt_it = i;
839 found_earliest_pkt = true;
840 }
841 }
842 }
843
844 DRAMPacket* selected_pkt = *selected_pkt_it;
845 queue.erase(selected_pkt_it);
846 queue.push_front(selected_pkt);
847}
848
849void
850DRAMCtrl::accessAndRespond(PacketPtr pkt, Tick static_latency)
851{
852 DPRINTF(DRAM, "Responding to Address %lld.. ",pkt->getAddr());
853
854 bool needsResponse = pkt->needsResponse();
855 // do the actual memory access which also turns the packet into a
856 // response
857 access(pkt);
858
859 // turn packet around to go back to requester if response expected
860 if (needsResponse) {
861 // access already turned the packet into a response
862 assert(pkt->isResponse());
863
864 // @todo someone should pay for this
865 pkt->busFirstWordDelay = pkt->busLastWordDelay = 0;
866
867 // queue the packet in the response queue to be sent out after
868 // the static latency has passed
869 port.schedTimingResp(pkt, curTick() + static_latency);
870 } else {
871 // @todo the packet is going to be deleted, and the DRAMPacket
872 // is still having a pointer to it
873 pendingDelete.push_back(pkt);
874 }
875
876 DPRINTF(DRAM, "Done\n");
877
878 return;
879}
880
881pair<Tick, Tick>
882DRAMCtrl::estimateLatency(DRAMPacket* dram_pkt, Tick inTime)
883{
884 // If a request reaches a bank at tick 'inTime', how much time
885 // *after* that does it take to finish the request, depending
886 // on bank status and page open policy. Note that this method
887 // considers only the time taken for the actual read or write
888 // to complete, NOT any additional time thereafter for tRAS or
889 // tRP.
890 Tick accLat = 0;
891 Tick bankLat = 0;
892 rowHitFlag = false;
893 Tick potentialActTick;
894
895 const Bank& bank = dram_pkt->bankRef;
896 // open-page policy or close_adaptive policy
897 if (pageMgmt == Enums::open || pageMgmt == Enums::open_adaptive ||
898 pageMgmt == Enums::close_adaptive) {
899 if (bank.openRow == dram_pkt->row) {
900 // When we have a row-buffer hit,
901 // we don't care about tRAS having expired or not,
902 // but do care about bank being free for access
903 rowHitFlag = true;
904
905 // When a series of requests arrive to the same row,
906 // DDR systems are capable of streaming data continuously
907 // at maximum bandwidth (subject to tCCD). Here, we approximate
908 // this condition, and assume that if whenever a bank is already
909 // busy and a new request comes in, it can be completed with no
910 // penalty beyond waiting for the existing read to complete.
911 if (bank.freeAt > inTime) {
912 accLat += bank.freeAt - inTime;
913 bankLat += 0;
914 } else {
915 // CAS latency only
916 accLat += tCL;
917 bankLat += tCL;
918 }
919
920 } else {
921 // Row-buffer miss, need to close existing row
922 // once tRAS has expired, then open the new one,
923 // then add cas latency.
924 Tick freeTime = std::max(bank.tRASDoneAt, bank.freeAt);
925
926 if (freeTime > inTime)
927 accLat += freeTime - inTime;
928
929 // If the there is no open row (open adaptive), then there
930 // is no precharge delay, otherwise go with tRP
931 Tick precharge_delay = bank.openRow == -1 ? 0 : tRP;
932
933 //The bank is free, and you may be able to activate
934 potentialActTick = inTime + accLat + precharge_delay;
935 if (potentialActTick < bank.actAllowedAt)
936 accLat += bank.actAllowedAt - potentialActTick;
937
938 accLat += precharge_delay + tRCD + tCL;
939 bankLat += precharge_delay + tRCD + tCL;
940 }
941 } else if (pageMgmt == Enums::close) {
942 // With a close page policy, no notion of
943 // bank.tRASDoneAt
944 if (bank.freeAt > inTime)
945 accLat += bank.freeAt - inTime;
946
947 //The bank is free, and you may be able to activate
948 potentialActTick = inTime + accLat;
949 if (potentialActTick < bank.actAllowedAt)
950 accLat += bank.actAllowedAt - potentialActTick;
951
952 // page already closed, simply open the row, and
953 // add cas latency
954 accLat += tRCD + tCL;
955 bankLat += tRCD + tCL;
956 } else
957 panic("No page management policy chosen\n");
958
959 DPRINTF(DRAM, "Returning < %lld, %lld > from estimateLatency()\n",
960 bankLat, accLat);
961
962 return make_pair(bankLat, accLat);
963}
964
965void
966DRAMCtrl::processNextReqEvent()
967{
968 scheduleNextReq();
969}
970
971void
972DRAMCtrl::recordActivate(Tick act_tick, uint8_t rank, uint8_t bank)
973{
974 assert(0 <= rank && rank < ranksPerChannel);
975 assert(actTicks[rank].size() == activationLimit);
976
977 DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
978
979 // Tracking accesses after all banks are precharged.
980 // startTickPrechargeAll: is the tick when all the banks were again
981 // precharged. The difference between act_tick and startTickPrechargeAll
982 // gives the time for which DRAM doesn't get any accesses after refreshing
983 // or after a page is closed in closed-page or open-adaptive-page policy.
984 if ((numBanksActive == 0) && (act_tick > startTickPrechargeAll)) {
985 prechargeAllTime += act_tick - startTickPrechargeAll;
986 }
987
988 // No need to update number of active banks for closed-page policy as only 1
989 // bank will be activated at any given point, which will be instatntly
990 // precharged
991 if (pageMgmt == Enums::open || pageMgmt == Enums::open_adaptive ||
992 pageMgmt == Enums::close_adaptive)
993 ++numBanksActive;
994
995 // start by enforcing tRRD
996 for(int i = 0; i < banksPerRank; i++) {
997 // next activate must not happen before tRRD
998 banks[rank][i].actAllowedAt = act_tick + tRRD;
999 }
1000 // tRC should be added to activation tick of the bank currently accessed,
1001 // where tRC = tRAS + tRP, this is just for a check as actAllowedAt for same
1002 // bank is already captured by bank.freeAt and bank.tRASDoneAt
1003 banks[rank][bank].actAllowedAt = act_tick + tRAS + tRP;
1004
1005 // next, we deal with tXAW, if the activation limit is disabled
1006 // then we are done
1007 if (actTicks[rank].empty())
1008 return;
1009
1010 // sanity check
1011 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
1012 // @todo For now, stick with a warning
1013 warn("Got %d activates in window %d (%d - %d) which is smaller "
1014 "than %d\n", activationLimit, act_tick - actTicks[rank].back(),
1015 act_tick, actTicks[rank].back(), tXAW);
1016 }
1017
1018 // shift the times used for the book keeping, the last element
1019 // (highest index) is the oldest one and hence the lowest value
1020 actTicks[rank].pop_back();
1021
1022 // record an new activation (in the future)
1023 actTicks[rank].push_front(act_tick);
1024
1025 // cannot activate more than X times in time window tXAW, push the
1026 // next one (the X + 1'st activate) to be tXAW away from the
1027 // oldest in our window of X
1028 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
1029 DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate no earlier "
1030 "than %d\n", activationLimit, actTicks[rank].back() + tXAW);
1031 for(int j = 0; j < banksPerRank; j++)
1032 // next activate must not happen before end of window
1033 banks[rank][j].actAllowedAt = actTicks[rank].back() + tXAW;
1034 }
1035}
1036
1037void
1038DRAMCtrl::doDRAMAccess(DRAMPacket* dram_pkt)
1039{
1040
1041 DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
1042 dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
1043
1044 // estimate the bank and access latency
1045 pair<Tick, Tick> lat = estimateLatency(dram_pkt, curTick());
1046 Tick bankLat = lat.first;
1047 Tick accessLat = lat.second;
1048 Tick actTick;
1049
1050 // This request was woken up at this time based on a prior call
1051 // to estimateLatency(). However, between then and now, both the
1052 // accessLatency and/or busBusyUntil may have changed. We need
1053 // to correct for that.
1054
1055 Tick addDelay = (curTick() + accessLat < busBusyUntil) ?
1056 busBusyUntil - (curTick() + accessLat) : 0;
1057
1058 Bank& bank = dram_pkt->bankRef;
1059
1060 // Update bank state
1061 if (pageMgmt == Enums::open || pageMgmt == Enums::open_adaptive ||
1062 pageMgmt == Enums::close_adaptive) {
1063 bank.freeAt = curTick() + addDelay + accessLat;
1064
1065 // If you activated a new row do to this access, the next access
1066 // will have to respect tRAS for this bank.
1067 if (!rowHitFlag) {
1068 // any waiting for banks account for in freeAt
1069 actTick = bank.freeAt - tCL - tRCD;
1070 bank.tRASDoneAt = actTick + tRAS;
1071 recordActivate(actTick, dram_pkt->rank, dram_pkt->bank);
1072
1073 // if we closed an open row as a result of this access,
1074 // then sample the number of bytes accessed before
1075 // resetting it
1076 if (bank.openRow != -1)
1077 bytesPerActivate.sample(bank.bytesAccessed);
1078
1079 // update the open row
1080 bank.openRow = dram_pkt->row;
1081
1082 // start counting anew, this covers both the case when we
1083 // auto-precharged, and when this access is forced to
1084 // precharge
1085 bank.bytesAccessed = 0;
1086 bank.rowAccesses = 0;
1087 }
1088
1089 // increment the bytes accessed and the accesses per row
1090 bank.bytesAccessed += burstSize;
1091 ++bank.rowAccesses;
1092
1093 // if we reached the max, then issue with an auto-precharge
1094 bool auto_precharge = bank.rowAccesses == maxAccessesPerRow;
1095
1096 // if we did not hit the limit, we might still want to
1097 // auto-precharge
1098 if (!auto_precharge &&
1099 (pageMgmt == Enums::open_adaptive ||
1100 pageMgmt == Enums::close_adaptive)) {
1101 // a twist on the open and close page policies:
1102 // 1) open_adaptive page policy does not blindly keep the
1103 // page open, but close it if there are no row hits, and there
1104 // are bank conflicts in the queue
1105 // 2) close_adaptive page policy does not blindly close the
1106 // page, but closes it only if there are no row hits in the queue.
1107 // In this case, only force an auto precharge when there
1108 // are no same page hits in the queue
1109 bool got_more_hits = false;
1110 bool got_bank_conflict = false;
1111
1112 // either look at the read queue or write queue
1113 const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
1114 writeQueue;
1115 auto p = queue.begin();
1116 // make sure we are not considering the packet that we are
1117 // currently dealing with (which is the head of the queue)
1118 ++p;
1119
1120 // keep on looking until we have found required condition or
1121 // reached the end
1122 while (!(got_more_hits &&
1123 (got_bank_conflict || pageMgmt == Enums::close_adaptive)) &&
1124 p != queue.end()) {
1125 bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
1126 (dram_pkt->bank == (*p)->bank);
1127 bool same_row = dram_pkt->row == (*p)->row;
1128 got_more_hits |= same_rank_bank && same_row;
1129 got_bank_conflict |= same_rank_bank && !same_row;
1130 ++p;
1131 }
1132
1133 // auto pre-charge when either
1134 // 1) open_adaptive policy, we have not got any more hits, and
1135 // have a bank conflict
1136 // 2) close_adaptive policy and we have not got any more hits
1137 auto_precharge = !got_more_hits &&
1138 (got_bank_conflict || pageMgmt == Enums::close_adaptive);
1139 }
1140
1141 // if this access should use auto-precharge, then we are
1142 // closing the row
1143 if (auto_precharge) {
1144 bank.openRow = -1;
1145 bank.freeAt = std::max(bank.freeAt, bank.tRASDoneAt) + tRP;
1146 --numBanksActive;
1147 if (numBanksActive == 0) {
1148 startTickPrechargeAll = std::max(startTickPrechargeAll,
1149 bank.freeAt);
1150 DPRINTF(DRAM, "All banks precharged at tick: %ld\n",
1151 startTickPrechargeAll);
1152 }
1153
1154 // sample the bytes per activate here since we are closing
1155 // the page
1156 bytesPerActivate.sample(bank.bytesAccessed);
1157
1158 DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
1159 }
1160
1161 DPRINTF(DRAM, "doDRAMAccess::bank.freeAt is %lld\n", bank.freeAt);
1162 } else if (pageMgmt == Enums::close) {
1163 actTick = curTick() + addDelay + accessLat - tRCD - tCL;
1164 recordActivate(actTick, dram_pkt->rank, dram_pkt->bank);
1165
1166 // If the DRAM has a very quick tRAS, bank can be made free
1167 // after consecutive tCL,tRCD,tRP times. In general, however,
1168 // an additional wait is required to respect tRAS.
1169 bank.freeAt = std::max(actTick + tRAS + tRP,
1170 actTick + tRCD + tCL + tRP);
1171 DPRINTF(DRAM, "doDRAMAccess::bank.freeAt is %lld\n", bank.freeAt);
1172 bytesPerActivate.sample(burstSize);
1173 startTickPrechargeAll = std::max(startTickPrechargeAll, bank.freeAt);
1174 } else
1175 panic("No page management policy chosen\n");
1176
1177 // Update request parameters
1178 dram_pkt->readyTime = curTick() + addDelay + accessLat + tBURST;
1179
1180
1181 DPRINTF(DRAM, "Req %lld: curtick is %lld accessLat is %d " \
1182 "readytime is %lld busbusyuntil is %lld. " \
1183 "Scheduling at readyTime\n", dram_pkt->addr,
1184 curTick(), accessLat, dram_pkt->readyTime, busBusyUntil);
1185
1186 // Make sure requests are not overlapping on the databus
1187 assert (dram_pkt->readyTime - busBusyUntil >= tBURST);
1188
1189 // Update bus state
1190 busBusyUntil = dram_pkt->readyTime;
1191
1192 DPRINTF(DRAM,"Access time is %lld\n",
1193 dram_pkt->readyTime - dram_pkt->entryTime);
1194
1195 // Update the minimum timing between the requests
1196 newTime = (busBusyUntil > tRP + tRCD + tCL) ?
1197 std::max(busBusyUntil - (tRP + tRCD + tCL), curTick()) : curTick();
1198
1199 // Update the access related stats
1200 if (dram_pkt->isRead) {
1201 if (rowHitFlag)
1202 readRowHits++;
1203 bytesReadDRAM += burstSize;
1204 perBankRdBursts[dram_pkt->bankId]++;
1205 } else {
1206 if (rowHitFlag)
1207 writeRowHits++;
1208 bytesWritten += burstSize;
1209 perBankWrBursts[dram_pkt->bankId]++;
1210
1211 // At this point, commonality between reads and writes ends.
1212 // For writes, we are done since we long ago responded to the
1213 // requestor.
1214 return;
1215 }
1216
1217 // Update latency stats
1218 totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
1219 totBankLat += bankLat;
1220 totBusLat += tBURST;
1221 totQLat += dram_pkt->readyTime - dram_pkt->entryTime - bankLat - tBURST;
1222
1223
1224 // At this point we're done dealing with the request
1225 // It will be moved to a separate response queue with a
1226 // correct readyTime, and eventually be sent back at that
1227 //time
1228 moveToRespQ();
1229
1230 // Schedule the next read event
1231 if (!nextReqEvent.scheduled() && !stopReads) {
1232 schedule(nextReqEvent, newTime);
1233 } else {
1234 if (newTime < nextReqEvent.when())
1235 reschedule(nextReqEvent, newTime);
1236 }
1237}
1238
1239void
1240DRAMCtrl::moveToRespQ()
1241{
1242 // Remove from read queue
1243 DRAMPacket* dram_pkt = readQueue.front();
1244 readQueue.pop_front();
1245
1246 // sanity check
1247 assert(dram_pkt->size <= burstSize);
1248
1249 // Insert into response queue sorted by readyTime
1250 // It will be sent back to the requestor at its
1251 // readyTime
1252 if (respQueue.empty()) {
1253 respQueue.push_front(dram_pkt);
1254 assert(!respondEvent.scheduled());
1255 assert(dram_pkt->readyTime >= curTick());
1256 schedule(respondEvent, dram_pkt->readyTime);
1257 } else {
1258 bool done = false;
1259 auto i = respQueue.begin();
1260 while (!done && i != respQueue.end()) {
1261 if ((*i)->readyTime > dram_pkt->readyTime) {
1262 respQueue.insert(i, dram_pkt);
1263 done = true;
1264 }
1265 ++i;
1266 }
1267
1268 if (!done)
1269 respQueue.push_back(dram_pkt);
1270
1271 assert(respondEvent.scheduled());
1272
1273 if (respQueue.front()->readyTime < respondEvent.when()) {
1274 assert(respQueue.front()->readyTime >= curTick());
1275 reschedule(respondEvent, respQueue.front()->readyTime);
1276 }
1277 }
1278}
1279
1280void
1281DRAMCtrl::scheduleNextReq()
1282{
1283 DPRINTF(DRAM, "Reached scheduleNextReq()\n");
1284
1285 // Figure out which read request goes next, and move it to the
1286 // front of the read queue
1287 if (!chooseNextRead()) {
1288 // In the case there is no read request to go next, trigger
1289 // writes if we have passed the low threshold (or if we are
1290 // draining)
1291 if (!writeQueue.empty() && !writeEvent.scheduled() &&
1292 (writeQueue.size() > writeLowThreshold || drainManager))
1293 triggerWrites();
1294 } else {
1295 doDRAMAccess(readQueue.front());
1296 }
1297}
1298
1299Tick
1300DRAMCtrl::maxBankFreeAt() const
1301{
1302 Tick banksFree = 0;
1303
1304 for(int i = 0; i < ranksPerChannel; i++)
1305 for(int j = 0; j < banksPerRank; j++)
1306 banksFree = std::max(banks[i][j].freeAt, banksFree);
1307
1308 return banksFree;
1309}
1310
1311uint64_t
1312DRAMCtrl::minBankFreeAt(const deque<DRAMPacket*>& queue) const
1313{
1314 uint64_t bank_mask = 0;
1315 Tick freeAt = MaxTick;
1316
1317 // detemrine if we have queued transactions targetting the
1318 // bank in question
1319 vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
1320 for (auto p = queue.begin(); p != queue.end(); ++p) {
1321 got_waiting[(*p)->bankId] = true;
1322 }
1323
1324 for (int i = 0; i < ranksPerChannel; i++) {
1325 for (int j = 0; j < banksPerRank; j++) {
1326 // if we have waiting requests for the bank, and it is
1327 // amongst the first available, update the mask
1328 if (got_waiting[i * banksPerRank + j] &&
1329 banks[i][j].freeAt <= freeAt) {
1330 // reset bank mask if new minimum is found
1331 if (banks[i][j].freeAt < freeAt)
1332 bank_mask = 0;
1333 // set the bit corresponding to the available bank
1334 uint8_t bit_index = i * ranksPerChannel + j;
1335 replaceBits(bank_mask, bit_index, bit_index, 1);
1336 freeAt = banks[i][j].freeAt;
1337 }
1338 }
1339 }
1340 return bank_mask;
1341}
1342
1343void
1344DRAMCtrl::processRefreshEvent()
1345{
1346 DPRINTF(DRAM, "Refreshing at tick %ld\n", curTick());
1347
1348 Tick banksFree = std::max(curTick(), maxBankFreeAt()) + tRFC;
1349
1350 for(int i = 0; i < ranksPerChannel; i++)
1351 for(int j = 0; j < banksPerRank; j++) {
1352 banks[i][j].freeAt = banksFree;
1353 banks[i][j].openRow = -1;
1354 }
1355
1356 // updating startTickPrechargeAll, isprechargeAll
1357 numBanksActive = 0;
1358 startTickPrechargeAll = banksFree;
1359
1360 schedule(refreshEvent, curTick() + tREFI);
1361}
1362
1363void
1364DRAMCtrl::regStats()
1365{
1366 using namespace Stats;
1367
1368 AbstractMemory::regStats();
1369
1370 readReqs
1371 .name(name() + ".readReqs")
1372 .desc("Number of read requests accepted");
1373
1374 writeReqs
1375 .name(name() + ".writeReqs")
1376 .desc("Number of write requests accepted");
1377
1378 readBursts
1379 .name(name() + ".readBursts")
1380 .desc("Number of DRAM read bursts, "
1381 "including those serviced by the write queue");
1382
1383 writeBursts
1384 .name(name() + ".writeBursts")
1385 .desc("Number of DRAM write bursts, "
1386 "including those merged in the write queue");
1387
1388 servicedByWrQ
1389 .name(name() + ".servicedByWrQ")
1390 .desc("Number of DRAM read bursts serviced by the write queue");
1391
1392 mergedWrBursts
1393 .name(name() + ".mergedWrBursts")
1394 .desc("Number of DRAM write bursts merged with an existing one");
1395
1396 neitherReadNorWrite
1397 .name(name() + ".neitherReadNorWriteReqs")
1398 .desc("Number of requests that are neither read nor write");
1399
1400 perBankRdBursts
1401 .init(banksPerRank * ranksPerChannel)
1402 .name(name() + ".perBankRdBursts")
1403 .desc("Per bank write bursts");
1404
1405 perBankWrBursts
1406 .init(banksPerRank * ranksPerChannel)
1407 .name(name() + ".perBankWrBursts")
1408 .desc("Per bank write bursts");
1409
1410 avgRdQLen
1411 .name(name() + ".avgRdQLen")
1412 .desc("Average read queue length when enqueuing")
1413 .precision(2);
1414
1415 avgWrQLen
1416 .name(name() + ".avgWrQLen")
1417 .desc("Average write queue length when enqueuing")
1418 .precision(2);
1419
1420 totQLat
1421 .name(name() + ".totQLat")
1422 .desc("Total ticks spent queuing");
1423
1424 totBankLat
1425 .name(name() + ".totBankLat")
1426 .desc("Total ticks spent accessing banks");
1427
1428 totBusLat
1429 .name(name() + ".totBusLat")
1430 .desc("Total ticks spent in databus transfers");
1431
1432 totMemAccLat
1433 .name(name() + ".totMemAccLat")
1434 .desc("Total ticks spent from burst creation until serviced "
1435 "by the DRAM");
1436
1437 avgQLat
1438 .name(name() + ".avgQLat")
1439 .desc("Average queueing delay per DRAM burst")
1440 .precision(2);
1441
1442 avgQLat = totQLat / (readBursts - servicedByWrQ);
1443
1444 avgBankLat
1445 .name(name() + ".avgBankLat")
1446 .desc("Average bank access latency per DRAM burst")
1447 .precision(2);
1448
1449 avgBankLat = totBankLat / (readBursts - servicedByWrQ);
1450
1451 avgBusLat
1452 .name(name() + ".avgBusLat")
1453 .desc("Average bus latency per DRAM burst")
1454 .precision(2);
1455
1456 avgBusLat = totBusLat / (readBursts - servicedByWrQ);
1457
1458 avgMemAccLat
1459 .name(name() + ".avgMemAccLat")
1460 .desc("Average memory access latency per DRAM burst")
1461 .precision(2);
1462
1463 avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
1464
1465 numRdRetry
1466 .name(name() + ".numRdRetry")
1467 .desc("Number of times read queue was full causing retry");
1468
1469 numWrRetry
1470 .name(name() + ".numWrRetry")
1471 .desc("Number of times write queue was full causing retry");
1472
1473 readRowHits
1474 .name(name() + ".readRowHits")
1475 .desc("Number of row buffer hits during reads");
1476
1477 writeRowHits
1478 .name(name() + ".writeRowHits")
1479 .desc("Number of row buffer hits during writes");
1480
1481 readRowHitRate
1482 .name(name() + ".readRowHitRate")
1483 .desc("Row buffer hit rate for reads")
1484 .precision(2);
1485
1486 readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
1487
1488 writeRowHitRate
1489 .name(name() + ".writeRowHitRate")
1490 .desc("Row buffer hit rate for writes")
1491 .precision(2);
1492
1493 writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
1494
1495 readPktSize
1496 .init(ceilLog2(burstSize) + 1)
1497 .name(name() + ".readPktSize")
1498 .desc("Read request sizes (log2)");
1499
1500 writePktSize
1501 .init(ceilLog2(burstSize) + 1)
1502 .name(name() + ".writePktSize")
1503 .desc("Write request sizes (log2)");
1504
1505 rdQLenPdf
1506 .init(readBufferSize)
1507 .name(name() + ".rdQLenPdf")
1508 .desc("What read queue length does an incoming req see");
1509
1510 wrQLenPdf
1511 .init(writeBufferSize)
1512 .name(name() + ".wrQLenPdf")
1513 .desc("What write queue length does an incoming req see");
1514
1515 bytesPerActivate
1516 .init(maxAccessesPerRow)
1517 .name(name() + ".bytesPerActivate")
1518 .desc("Bytes accessed per row activation")
1519 .flags(nozero);
1520
1521 bytesReadDRAM
1522 .name(name() + ".bytesReadDRAM")
1523 .desc("Total number of bytes read from DRAM");
1524
1525 bytesReadWrQ
1526 .name(name() + ".bytesReadWrQ")
1527 .desc("Total number of bytes read from write queue");
1528
1529 bytesWritten
1530 .name(name() + ".bytesWritten")
1531 .desc("Total number of bytes written to DRAM");
1532
1533 bytesReadSys
1534 .name(name() + ".bytesReadSys")
1535 .desc("Total read bytes from the system interface side");
1536
1537 bytesWrittenSys
1538 .name(name() + ".bytesWrittenSys")
1539 .desc("Total written bytes from the system interface side");
1540
1541 avgRdBW
1542 .name(name() + ".avgRdBW")
1543 .desc("Average DRAM read bandwidth in MiByte/s")
1544 .precision(2);
1545
1546 avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
1547
1548 avgWrBW
1549 .name(name() + ".avgWrBW")
1550 .desc("Average achieved write bandwidth in MiByte/s")
1551 .precision(2);
1552
1553 avgWrBW = (bytesWritten / 1000000) / simSeconds;
1554
1555 avgRdBWSys
1556 .name(name() + ".avgRdBWSys")
1557 .desc("Average system read bandwidth in MiByte/s")
1558 .precision(2);
1559
1560 avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
1561
1562 avgWrBWSys
1563 .name(name() + ".avgWrBWSys")
1564 .desc("Average system write bandwidth in MiByte/s")
1565 .precision(2);
1566
1567 avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
1568
1569 peakBW
1570 .name(name() + ".peakBW")
1571 .desc("Theoretical peak bandwidth in MiByte/s")
1572 .precision(2);
1573
1574 peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
1575
1576 busUtil
1577 .name(name() + ".busUtil")
1578 .desc("Data bus utilization in percentage")
1579 .precision(2);
1580
1581 busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
1582
1583 totGap
1584 .name(name() + ".totGap")
1585 .desc("Total gap between requests");
1586
1587 avgGap
1588 .name(name() + ".avgGap")
1589 .desc("Average gap between requests")
1590 .precision(2);
1591
1592 avgGap = totGap / (readReqs + writeReqs);
1593
1594 // Stats for DRAM Power calculation based on Micron datasheet
1595 busUtilRead
1596 .name(name() + ".busUtilRead")
1597 .desc("Data bus utilization in percentage for reads")
1598 .precision(2);
1599
1600 busUtilRead = avgRdBW / peakBW * 100;
1601
1602 busUtilWrite
1603 .name(name() + ".busUtilWrite")
1604 .desc("Data bus utilization in percentage for writes")
1605 .precision(2);
1606
1607 busUtilWrite = avgWrBW / peakBW * 100;
1608
1609 pageHitRate
1610 .name(name() + ".pageHitRate")
1611 .desc("Row buffer hit rate, read and write combined")
1612 .precision(2);
1613
1614 pageHitRate = (writeRowHits + readRowHits) /
1615 (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
1616
1617 prechargeAllPercent
1618 .name(name() + ".prechargeAllPercent")
1619 .desc("Percentage of time for which DRAM has all the banks in "
1620 "precharge state")
1621 .precision(2);
1622
1623 prechargeAllPercent = prechargeAllTime / simTicks * 100;
1624}
1625
1626void
1627DRAMCtrl::recvFunctional(PacketPtr pkt)
1628{
1629 // rely on the abstract memory
1630 functionalAccess(pkt);
1631}
1632
1633BaseSlavePort&
1634DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
1635{
1636 if (if_name != "port") {
1637 return MemObject::getSlavePort(if_name, idx);
1638 } else {
1639 return port;
1640 }
1641}
1642
1643unsigned int
1644DRAMCtrl::drain(DrainManager *dm)
1645{
1646 unsigned int count = port.drain(dm);
1647
1648 // if there is anything in any of our internal queues, keep track
1649 // of that as well
1650 if (!(writeQueue.empty() && readQueue.empty() &&
1651 respQueue.empty())) {
1652 DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
1653 " resp: %d\n", writeQueue.size(), readQueue.size(),
1654 respQueue.size());
1655 ++count;
1656 drainManager = dm;
1657 // the only part that is not drained automatically over time
1658 // is the write queue, thus trigger writes if there are any
1659 // waiting and no reads waiting, otherwise wait until the
1660 // reads are done
1661 if (readQueue.empty() && !writeQueue.empty() &&
1662 !writeEvent.scheduled())
1663 triggerWrites();
1664 }
1665
1666 if (count)
1667 setDrainState(Drainable::Draining);
1668 else
1669 setDrainState(Drainable::Drained);
1670 return count;
1671}
1672
1673DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
1674 : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
1675 memory(_memory)
1676{ }
1677
1678AddrRangeList
1679DRAMCtrl::MemoryPort::getAddrRanges() const
1680{
1681 AddrRangeList ranges;
1682 ranges.push_back(memory.getAddrRange());
1683 return ranges;
1684}
1685
1686void
1687DRAMCtrl::MemoryPort::recvFunctional(PacketPtr pkt)
1688{
1689 pkt->pushLabel(memory.name());
1690
1691 if (!queue.checkFunctional(pkt)) {
1692 // Default implementation of SimpleTimingPort::recvFunctional()
1693 // calls recvAtomic() and throws away the latency; we can save a
1694 // little here by just not calculating the latency.
1695 memory.recvFunctional(pkt);
1696 }
1697
1698 pkt->popLabel();
1699}
1700
1701Tick
1702DRAMCtrl::MemoryPort::recvAtomic(PacketPtr pkt)
1703{
1704 return memory.recvAtomic(pkt);
1705}
1706
1707bool
1708DRAMCtrl::MemoryPort::recvTimingReq(PacketPtr pkt)
1709{
1710 // pass it to the memory controller
1711 return memory.recvTimingReq(pkt);
1712}
1713
1714DRAMCtrl*
1715DRAMCtrlParams::create()
1716{
1717 return new DRAMCtrl(this);
1718}