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