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), refreshEvent(this),
60 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), refreshDueAt(0), refreshState(REF_IDLE), prevArrival(0),
85 nextReqTime(0), idleStartTick(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 // some basic sanity checks
136 if (tREFI <= tRP || tREFI <= tRFC) {
137 fatal("tREFI (%d) must be larger than tRP (%d) and tRFC (%d)\n",
138 tREFI, tRP, tRFC);
139 }
140}
141
142void
143DRAMCtrl::init()
144{
145 if (!port.isConnected()) {
146 fatal("DRAMCtrl %s is unconnected!\n", name());
147 } else {
148 port.sendRangeChange();
149 }
150}
151
152void
153DRAMCtrl::startup()
154{
155 // update the start tick for the precharge accounting to the
156 // current tick
157 idleStartTick = curTick();
158
159 // shift the bus busy time sufficiently far ahead that we never
160 // have to worry about negative values when computing the time for
161 // the next request, this will add an insignificant bubble at the
162 // start of simulation
163 busBusyUntil = curTick() + tRP + tRCD + tCL;
164
165 // print the configuration of the controller
166 printParams();
167
168 // kick off the refresh, and give ourselves enough time to
169 // precharge
170 schedule(refreshEvent, curTick() + tREFI - tRP);
171}
172
173Tick
174DRAMCtrl::recvAtomic(PacketPtr pkt)
175{
176 DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
177
178 // do the actual memory access and turn the packet into a response
179 access(pkt);
180
181 Tick latency = 0;
182 if (!pkt->memInhibitAsserted() && pkt->hasData()) {
183 // this value is not supposed to be accurate, just enough to
184 // keep things going, mimic a closed page
185 latency = tRP + tRCD + tCL;
186 }
187 return latency;
188}
189
190bool
191DRAMCtrl::readQueueFull(unsigned int neededEntries) const
192{
193 DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
194 readBufferSize, readQueue.size() + respQueue.size(),
195 neededEntries);
196
197 return
198 (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
199}
200
201bool
202DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
203{
204 DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
205 writeBufferSize, writeQueue.size(), neededEntries);
206 return (writeQueue.size() + neededEntries) > writeBufferSize;
207}
208
209DRAMCtrl::DRAMPacket*
210DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
211 bool isRead)
212{
213 // decode the address based on the address mapping scheme, with
214 // Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
215 // channel, respectively
216 uint8_t rank;
217 uint8_t bank;
218 uint16_t row;
219
220 // truncate the address to the access granularity
221 Addr addr = dramPktAddr / burstSize;
222
223 // we have removed the lowest order address bits that denote the
224 // position within the column
225 if (addrMapping == Enums::RoRaBaChCo) {
226 // the lowest order bits denote the column to ensure that
227 // sequential cache lines occupy the same row
228 addr = addr / columnsPerRowBuffer;
229
230 // take out the channel part of the address
231 addr = addr / channels;
232
233 // after the channel bits, get the bank bits to interleave
234 // over the banks
235 bank = addr % banksPerRank;
236 addr = addr / banksPerRank;
237
238 // after the bank, we get the rank bits which thus interleaves
239 // over the ranks
240 rank = addr % ranksPerChannel;
241 addr = addr / ranksPerChannel;
242
243 // lastly, get the row bits
244 row = addr % rowsPerBank;
245 addr = addr / rowsPerBank;
246 } else if (addrMapping == Enums::RoRaBaCoCh) {
247 // take out the channel part of the address
248 addr = addr / channels;
249
250 // next, the column
251 addr = addr / columnsPerRowBuffer;
252
253 // after the column bits, we get the bank bits to interleave
254 // over the banks
255 bank = addr % banksPerRank;
256 addr = addr / banksPerRank;
257
258 // after the bank, we get the rank bits which thus interleaves
259 // over the ranks
260 rank = addr % ranksPerChannel;
261 addr = addr / ranksPerChannel;
262
263 // lastly, get the row bits
264 row = addr % rowsPerBank;
265 addr = addr / rowsPerBank;
266 } else if (addrMapping == Enums::RoCoRaBaCh) {
267 // optimise for closed page mode and utilise maximum
268 // parallelism of the DRAM (at the cost of power)
269
270 // take out the channel part of the address, not that this has
271 // to match with how accesses are interleaved between the
272 // controllers in the address mapping
273 addr = addr / channels;
274
275 // start with the bank bits, as this provides the maximum
276 // opportunity for parallelism between requests
277 bank = addr % banksPerRank;
278 addr = addr / banksPerRank;
279
280 // next get the rank bits
281 rank = addr % ranksPerChannel;
282 addr = addr / ranksPerChannel;
283
284 // next the column bits which we do not need to keep track of
285 // and simply skip past
286 addr = addr / columnsPerRowBuffer;
287
288 // lastly, get the row bits
289 row = addr % rowsPerBank;
290 addr = addr / rowsPerBank;
291 } else
292 panic("Unknown address mapping policy chosen!");
293
294 assert(rank < ranksPerChannel);
295 assert(bank < banksPerRank);
296 assert(row < rowsPerBank);
297
298 DPRINTF(DRAM, "Address: %lld Rank %d Bank %d Row %d\n",
299 dramPktAddr, rank, bank, row);
300
301 // create the corresponding DRAM packet with the entry time and
302 // ready time set to the current tick, the latter will be updated
303 // later
304 uint16_t bank_id = banksPerRank * rank + bank;
305 return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
306 size, banks[rank][bank]);
307}
308
309void
310DRAMCtrl::addToReadQueue(PacketPtr pkt, unsigned int pktCount)
311{
312 // only add to the read queue here. whenever the request is
313 // eventually done, set the readyTime, and call schedule()
314 assert(!pkt->isWrite());
315
316 assert(pktCount != 0);
317
318 // if the request size is larger than burst size, the pkt is split into
319 // multiple DRAM packets
320 // Note if the pkt starting address is not aligened to burst size, the
321 // address of first DRAM packet is kept unaliged. Subsequent DRAM packets
322 // are aligned to burst size boundaries. This is to ensure we accurately
323 // check read packets against packets in write queue.
324 Addr addr = pkt->getAddr();
325 unsigned pktsServicedByWrQ = 0;
326 BurstHelper* burst_helper = NULL;
327 for (int cnt = 0; cnt < pktCount; ++cnt) {
328 unsigned size = std::min((addr | (burstSize - 1)) + 1,
329 pkt->getAddr() + pkt->getSize()) - addr;
330 readPktSize[ceilLog2(size)]++;
331 readBursts++;
332
333 // First check write buffer to see if the data is already at
334 // the controller
335 bool foundInWrQ = false;
336 for (auto i = writeQueue.begin(); i != writeQueue.end(); ++i) {
337 // check if the read is subsumed in the write entry we are
338 // looking at
339 if ((*i)->addr <= addr &&
340 (addr + size) <= ((*i)->addr + (*i)->size)) {
341 foundInWrQ = true;
342 servicedByWrQ++;
343 pktsServicedByWrQ++;
344 DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
345 "write queue\n", addr, size);
346 bytesReadWrQ += burstSize;
347 break;
348 }
349 }
350
351 // If not found in the write q, make a DRAM packet and
352 // push it onto the read queue
353 if (!foundInWrQ) {
354
355 // Make the burst helper for split packets
356 if (pktCount > 1 && burst_helper == NULL) {
357 DPRINTF(DRAM, "Read to addr %lld translates to %d "
358 "dram requests\n", pkt->getAddr(), pktCount);
359 burst_helper = new BurstHelper(pktCount);
360 }
361
362 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, true);
363 dram_pkt->burstHelper = burst_helper;
364
365 assert(!readQueueFull(1));
366 rdQLenPdf[readQueue.size() + respQueue.size()]++;
367
368 DPRINTF(DRAM, "Adding to read queue\n");
369
370 readQueue.push_back(dram_pkt);
371
372 // Update stats
373 avgRdQLen = readQueue.size() + respQueue.size();
374 }
375
376 // Starting address of next dram pkt (aligend to burstSize boundary)
377 addr = (addr | (burstSize - 1)) + 1;
378 }
379
380 // If all packets are serviced by write queue, we send the repsonse back
381 if (pktsServicedByWrQ == pktCount) {
382 accessAndRespond(pkt, frontendLatency);
383 return;
384 }
385
386 // Update how many split packets are serviced by write queue
387 if (burst_helper != NULL)
388 burst_helper->burstsServiced = pktsServicedByWrQ;
389
390 // If we are not already scheduled to get a request out of the
391 // queue, do so now
392 if (!nextReqEvent.scheduled()) {
393 DPRINTF(DRAM, "Request scheduled immediately\n");
394 schedule(nextReqEvent, curTick());
395 }
396}
397
398void
399DRAMCtrl::addToWriteQueue(PacketPtr pkt, unsigned int pktCount)
400{
401 // only add to the write queue here. whenever the request is
402 // eventually done, set the readyTime, and call schedule()
403 assert(pkt->isWrite());
404
405 // if the request size is larger than burst size, the pkt is split into
406 // multiple DRAM packets
407 Addr addr = pkt->getAddr();
408 for (int cnt = 0; cnt < pktCount; ++cnt) {
409 unsigned size = std::min((addr | (burstSize - 1)) + 1,
410 pkt->getAddr() + pkt->getSize()) - addr;
411 writePktSize[ceilLog2(size)]++;
412 writeBursts++;
413
414 // see if we can merge with an existing item in the write
415 // queue and keep track of whether we have merged or not so we
416 // can stop at that point and also avoid enqueueing a new
417 // request
418 bool merged = false;
419 auto w = writeQueue.begin();
420
421 while(!merged && w != writeQueue.end()) {
422 // either of the two could be first, if they are the same
423 // it does not matter which way we go
424 if ((*w)->addr >= addr) {
425 // the existing one starts after the new one, figure
426 // out where the new one ends with respect to the
427 // existing one
428 if ((addr + size) >= ((*w)->addr + (*w)->size)) {
429 // check if the existing one is completely
430 // subsumed in the new one
431 DPRINTF(DRAM, "Merging write covering existing burst\n");
432 merged = true;
433 // update both the address and the size
434 (*w)->addr = addr;
435 (*w)->size = size;
436 } else if ((addr + size) >= (*w)->addr &&
437 ((*w)->addr + (*w)->size - addr) <= burstSize) {
438 // the new one is just before or partially
439 // overlapping with the existing one, and together
440 // they fit within a burst
441 DPRINTF(DRAM, "Merging write before existing burst\n");
442 merged = true;
443 // the existing queue item needs to be adjusted with
444 // respect to both address and size
445 (*w)->size = (*w)->addr + (*w)->size - addr;
446 (*w)->addr = addr;
447 }
448 } else {
449 // the new one starts after the current one, figure
450 // out where the existing one ends with respect to the
451 // new one
452 if (((*w)->addr + (*w)->size) >= (addr + size)) {
453 // check if the new one is completely subsumed in the
454 // existing one
455 DPRINTF(DRAM, "Merging write into existing burst\n");
456 merged = true;
457 // no adjustments necessary
458 } else if (((*w)->addr + (*w)->size) >= addr &&
459 (addr + size - (*w)->addr) <= burstSize) {
460 // the existing one is just before or partially
461 // overlapping with the new one, and together
462 // they fit within a burst
463 DPRINTF(DRAM, "Merging write after existing burst\n");
464 merged = true;
465 // the address is right, and only the size has
466 // to be adjusted
467 (*w)->size = addr + size - (*w)->addr;
468 }
469 }
470 ++w;
471 }
472
473 // if the item was not merged we need to create a new write
474 // and enqueue it
475 if (!merged) {
476 DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
477
478 assert(writeQueue.size() < writeBufferSize);
479 wrQLenPdf[writeQueue.size()]++;
480
481 DPRINTF(DRAM, "Adding to write queue\n");
482
483 writeQueue.push_back(dram_pkt);
484
485 // Update stats
486 avgWrQLen = writeQueue.size();
487 } else {
488 // keep track of the fact that this burst effectively
489 // disappeared as it was merged with an existing one
490 mergedWrBursts++;
491 }
492
493 // Starting address of next dram pkt (aligend to burstSize boundary)
494 addr = (addr | (burstSize - 1)) + 1;
495 }
496
497 // we do not wait for the writes to be send to the actual memory,
498 // but instead take responsibility for the consistency here and
499 // snoop the write queue for any upcoming reads
500 // @todo, if a pkt size is larger than burst size, we might need a
501 // different front end latency
502 accessAndRespond(pkt, frontendLatency);
503
504 // If we are not already scheduled to get a request out of the
505 // queue, do so now
506 if (!nextReqEvent.scheduled()) {
507 DPRINTF(DRAM, "Request scheduled immediately\n");
508 schedule(nextReqEvent, curTick());
509 }
510}
511
512void
513DRAMCtrl::printParams() const
514{
515 // Sanity check print of important parameters
516 DPRINTF(DRAM,
517 "Memory controller %s physical organization\n" \
518 "Number of devices per rank %d\n" \
519 "Device bus width (in bits) %d\n" \
520 "DRAM data bus burst (bytes) %d\n" \
521 "Row buffer size (bytes) %d\n" \
522 "Columns per row buffer %d\n" \
523 "Rows per bank %d\n" \
524 "Banks per rank %d\n" \
525 "Ranks per channel %d\n" \
526 "Total mem capacity (bytes) %u\n",
527 name(), devicesPerRank, deviceBusWidth, burstSize, rowBufferSize,
528 columnsPerRowBuffer, rowsPerBank, banksPerRank, ranksPerChannel,
529 rowBufferSize * rowsPerBank * banksPerRank * ranksPerChannel);
530
531 string scheduler = memSchedPolicy == Enums::fcfs ? "FCFS" : "FR-FCFS";
532 string address_mapping = addrMapping == Enums::RoRaBaChCo ? "RoRaBaChCo" :
533 (addrMapping == Enums::RoRaBaCoCh ? "RoRaBaCoCh" : "RoCoRaBaCh");
534 string page_policy = pageMgmt == Enums::open ? "OPEN" :
535 (pageMgmt == Enums::open_adaptive ? "OPEN (adaptive)" :
536 (pageMgmt == Enums::close_adaptive ? "CLOSE (adaptive)" : "CLOSE"));
537
538 DPRINTF(DRAM,
539 "Memory controller %s characteristics\n" \
540 "Read buffer size %d\n" \
541 "Write buffer size %d\n" \
542 "Write high thresh %d\n" \
543 "Write low thresh %d\n" \
544 "Scheduler %s\n" \
545 "Address mapping %s\n" \
546 "Page policy %s\n",
547 name(), readBufferSize, writeBufferSize, writeHighThreshold,
548 writeLowThreshold, scheduler, address_mapping, page_policy);
549
550 DPRINTF(DRAM, "Memory controller %s timing specs\n" \
551 "tRCD %d ticks\n" \
552 "tCL %d ticks\n" \
553 "tRP %d ticks\n" \
554 "tBURST %d ticks\n" \
555 "tRFC %d ticks\n" \
556 "tREFI %d ticks\n" \
557 "tWTR %d ticks\n" \
558 "tRTW %d ticks\n" \
559 "tXAW (%d) %d ticks\n",
560 name(), tRCD, tCL, tRP, tBURST, tRFC, tREFI, tWTR,
561 tRTW, activationLimit, tXAW);
562}
563
564void
565DRAMCtrl::printQs() const {
566 DPRINTF(DRAM, "===READ QUEUE===\n\n");
567 for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
568 DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
569 }
570 DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
571 for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
572 DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
573 }
574 DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
575 for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
576 DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
577 }
578}
579
580bool
581DRAMCtrl::recvTimingReq(PacketPtr pkt)
582{
583 /// @todo temporary hack to deal with memory corruption issues until
584 /// 4-phase transactions are complete
585 for (int x = 0; x < pendingDelete.size(); x++)
586 delete pendingDelete[x];
587 pendingDelete.clear();
588
589 // This is where we enter from the outside world
590 DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
591 pkt->cmdString(), pkt->getAddr(), pkt->getSize());
592
593 // simply drop inhibited packets for now
594 if (pkt->memInhibitAsserted()) {
595 DPRINTF(DRAM, "Inhibited packet -- Dropping it now\n");
596 pendingDelete.push_back(pkt);
597 return true;
598 }
599
600 // Calc avg gap between requests
601 if (prevArrival != 0) {
602 totGap += curTick() - prevArrival;
603 }
604 prevArrival = curTick();
605
606
607 // Find out how many dram packets a pkt translates to
608 // If the burst size is equal or larger than the pkt size, then a pkt
609 // translates to only one dram packet. Otherwise, a pkt translates to
610 // multiple dram packets
611 unsigned size = pkt->getSize();
612 unsigned offset = pkt->getAddr() & (burstSize - 1);
613 unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
614
615 // check local buffers and do not accept if full
616 if (pkt->isRead()) {
617 assert(size != 0);
618 if (readQueueFull(dram_pkt_count)) {
619 DPRINTF(DRAM, "Read queue full, not accepting\n");
620 // remember that we have to retry this port
621 retryRdReq = true;
622 numRdRetry++;
623 return false;
624 } else {
625 addToReadQueue(pkt, dram_pkt_count);
626 readReqs++;
627 bytesReadSys += size;
628 }
629 } else if (pkt->isWrite()) {
630 assert(size != 0);
631 if (writeQueueFull(dram_pkt_count)) {
632 DPRINTF(DRAM, "Write queue full, not accepting\n");
633 // remember that we have to retry this port
634 retryWrReq = true;
635 numWrRetry++;
636 return false;
637 } else {
638 addToWriteQueue(pkt, dram_pkt_count);
639 writeReqs++;
640 bytesWrittenSys += size;
641 }
642 } else {
643 DPRINTF(DRAM,"Neither read nor write, ignore timing\n");
644 neitherReadNorWrite++;
645 accessAndRespond(pkt, 1);
646 }
647
648 return true;
649}
650
651void
652DRAMCtrl::processRespondEvent()
653{
654 DPRINTF(DRAM,
655 "processRespondEvent(): Some req has reached its readyTime\n");
656
657 DRAMPacket* dram_pkt = respQueue.front();
658
659 if (dram_pkt->burstHelper) {
660 // it is a split packet
661 dram_pkt->burstHelper->burstsServiced++;
662 if (dram_pkt->burstHelper->burstsServiced ==
663 dram_pkt->burstHelper->burstCount) {
664 // we have now serviced all children packets of a system packet
665 // so we can now respond to the requester
666 // @todo we probably want to have a different front end and back
667 // end latency for split packets
668 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
669 delete dram_pkt->burstHelper;
670 dram_pkt->burstHelper = NULL;
671 }
672 } else {
673 // it is not a split packet
674 accessAndRespond(dram_pkt->pkt, frontendLatency + backendLatency);
675 }
676
677 delete respQueue.front();
678 respQueue.pop_front();
679
680 if (!respQueue.empty()) {
681 assert(respQueue.front()->readyTime >= curTick());
682 assert(!respondEvent.scheduled());
683 schedule(respondEvent, respQueue.front()->readyTime);
684 } else {
685 // if there is nothing left in any queue, signal a drain
686 if (writeQueue.empty() && readQueue.empty() &&
687 drainManager) {
688 drainManager->signalDrainDone();
689 drainManager = NULL;
690 }
691 }
692
693 // We have made a location in the queue available at this point,
694 // so if there is a read that was forced to wait, retry now
695 if (retryRdReq) {
696 retryRdReq = false;
697 port.sendRetry();
698 }
699}
700
701void
702DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue)
703{
704 // This method does the arbitration between requests. The chosen
705 // packet is simply moved to the head of the queue. The other
706 // methods know that this is the place to look. For example, with
707 // FCFS, this method does nothing
708 assert(!queue.empty());
709
710 if (queue.size() == 1) {
711 DPRINTF(DRAM, "Single request, nothing to do\n");
712 return;
713 }
714
715 if (memSchedPolicy == Enums::fcfs) {
716 // Do nothing, since the correct request is already head
717 } else if (memSchedPolicy == Enums::frfcfs) {
718 reorderQueue(queue);
719 } else
720 panic("No scheduling policy chosen\n");
721}
722
723void
724DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue)
725{
726 // Only determine this when needed
727 uint64_t earliest_banks = 0;
728
729 // Search for row hits first, if no row hit is found then schedule the
730 // packet to one of the earliest banks available
731 bool found_earliest_pkt = false;
732 auto selected_pkt_it = queue.begin();
733
734 for (auto i = queue.begin(); i != queue.end() ; ++i) {
735 DRAMPacket* dram_pkt = *i;
736 const Bank& bank = dram_pkt->bankRef;
737 // Check if it is a row hit
738 if (bank.openRow == dram_pkt->row) {
739 DPRINTF(DRAM, "Row buffer hit\n");
740 selected_pkt_it = i;
741 break;
742 } else if (!found_earliest_pkt) {
743 // No row hit, go for first ready
744 if (earliest_banks == 0)
745 earliest_banks = minBankFreeAt(queue);
746
747 // Bank is ready or is the first available bank
748 if (bank.freeAt <= curTick() ||
749 bits(earliest_banks, dram_pkt->bankId, dram_pkt->bankId)) {
750 // Remember the packet to be scheduled to one of the earliest
751 // banks available
752 selected_pkt_it = i;
753 found_earliest_pkt = true;
754 }
755 }
756 }
757
758 DRAMPacket* selected_pkt = *selected_pkt_it;
759 queue.erase(selected_pkt_it);
760 queue.push_front(selected_pkt);
761}
762
763void
764DRAMCtrl::accessAndRespond(PacketPtr pkt, Tick static_latency)
765{
766 DPRINTF(DRAM, "Responding to Address %lld.. ",pkt->getAddr());
767
768 bool needsResponse = pkt->needsResponse();
769 // do the actual memory access which also turns the packet into a
770 // response
771 access(pkt);
772
773 // turn packet around to go back to requester if response expected
774 if (needsResponse) {
775 // access already turned the packet into a response
776 assert(pkt->isResponse());
777
778 // @todo someone should pay for this
779 pkt->busFirstWordDelay = pkt->busLastWordDelay = 0;
780
781 // queue the packet in the response queue to be sent out after
782 // the static latency has passed
783 port.schedTimingResp(pkt, curTick() + static_latency);
784 } else {
785 // @todo the packet is going to be deleted, and the DRAMPacket
786 // is still having a pointer to it
787 pendingDelete.push_back(pkt);
788 }
789
790 DPRINTF(DRAM, "Done\n");
791
792 return;
793}
794
795pair<Tick, Tick>
796DRAMCtrl::estimateLatency(DRAMPacket* dram_pkt, Tick inTime)
797{
798 // If a request reaches a bank at tick 'inTime', how much time
799 // *after* that does it take to finish the request, depending
800 // on bank status and page open policy. Note that this method
801 // considers only the time taken for the actual read or write
802 // to complete, NOT any additional time thereafter for tRAS or
803 // tRP.
804 Tick accLat = 0;
805 Tick bankLat = 0;
806 rowHitFlag = false;
807 Tick potentialActTick;
808
809 const Bank& bank = dram_pkt->bankRef;
810 // open-page policy or close_adaptive policy
811 if (pageMgmt == Enums::open || pageMgmt == Enums::open_adaptive ||
812 pageMgmt == Enums::close_adaptive) {
813 if (bank.openRow == dram_pkt->row) {
814 // When we have a row-buffer hit,
815 // we don't care about tRAS having expired or not,
816 // but do care about bank being free for access
817 rowHitFlag = true;
818
819 // When a series of requests arrive to the same row,
820 // DDR systems are capable of streaming data continuously
821 // at maximum bandwidth (subject to tCCD). Here, we approximate
822 // this condition, and assume that if whenever a bank is already
823 // busy and a new request comes in, it can be completed with no
824 // penalty beyond waiting for the existing read to complete.
825 if (bank.freeAt > inTime) {
826 accLat += bank.freeAt - inTime;
827 bankLat += 0;
828 } else {
829 // CAS latency only
830 accLat += tCL;
831 bankLat += tCL;
832 }
833
834 } else {
835 // Row-buffer miss, need to close existing row
836 // once tRAS has expired, then open the new one,
837 // then add cas latency.
838 Tick freeTime = std::max(bank.tRASDoneAt, bank.freeAt);
839
840 if (freeTime > inTime)
841 accLat += freeTime - inTime;
842
843 // If the there is no open row (open adaptive), then there
844 // is no precharge delay, otherwise go with tRP
845 Tick precharge_delay = bank.openRow == Bank::NO_ROW ? 0 : tRP;
846
847 //The bank is free, and you may be able to activate
848 potentialActTick = inTime + accLat + precharge_delay;
849 if (potentialActTick < bank.actAllowedAt)
850 accLat += bank.actAllowedAt - potentialActTick;
851
852 accLat += precharge_delay + tRCD + tCL;
853 bankLat += precharge_delay + tRCD + tCL;
854 }
855 } else if (pageMgmt == Enums::close) {
856 // With a close page policy, no notion of
857 // bank.tRASDoneAt
858 if (bank.freeAt > inTime)
859 accLat += bank.freeAt - inTime;
860
861 //The bank is free, and you may be able to activate
862 potentialActTick = inTime + accLat;
863 if (potentialActTick < bank.actAllowedAt)
864 accLat += bank.actAllowedAt - potentialActTick;
865
866 // page already closed, simply open the row, and
867 // add cas latency
868 accLat += tRCD + tCL;
869 bankLat += tRCD + tCL;
870 } else
871 panic("No page management policy chosen\n");
872
873 DPRINTF(DRAM, "Returning < %lld, %lld > from estimateLatency()\n",
874 bankLat, accLat);
875
876 return make_pair(bankLat, accLat);
877}
878
879void
880DRAMCtrl::recordActivate(Tick act_tick, uint8_t rank, uint8_t bank,
881 uint16_t row)
882{
883 assert(0 <= rank && rank < ranksPerChannel);
884 assert(actTicks[rank].size() == activationLimit);
885
886 DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
887
888 // idleStartTick is the tick when all the banks were
889 // precharged. Thus, the difference between act_tick and
890 // idleStartTick gives the time for which the DRAM is in an idle
891 // state with all banks precharged. Note that we may end up
892 // "changing history" by scheduling an activation before an
893 // already scheduled precharge, effectively canceling it out.
894 if (numBanksActive == 0 && act_tick > idleStartTick) {
895 prechargeAllTime += act_tick - idleStartTick;
896 }
897
898 // update the open row
899 assert(banks[rank][bank].openRow == Bank::NO_ROW);
900 banks[rank][bank].openRow = row;
901
902 // start counting anew, this covers both the case when we
903 // auto-precharged, and when this access is forced to
904 // precharge
905 banks[rank][bank].bytesAccessed = 0;
906 banks[rank][bank].rowAccesses = 0;
907
908 ++numBanksActive;
909 assert(numBanksActive <= banksPerRank * ranksPerChannel);
910
911 DPRINTF(DRAM, "Activate bank at tick %lld, now got %d active\n",
912 act_tick, numBanksActive);
913
914 // start by enforcing tRRD
915 for(int i = 0; i < banksPerRank; i++) {
916 // next activate must not happen before tRRD
917 banks[rank][i].actAllowedAt = act_tick + tRRD;
918 }
919 // tRC should be added to activation tick of the bank currently accessed,
920 // where tRC = tRAS + tRP, this is just for a check as actAllowedAt for same
921 // bank is already captured by bank.freeAt and bank.tRASDoneAt
922 banks[rank][bank].actAllowedAt = act_tick + tRAS + tRP;
923
924 // next, we deal with tXAW, if the activation limit is disabled
925 // then we are done
926 if (actTicks[rank].empty())
927 return;
928
929 // sanity check
930 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
931 // @todo For now, stick with a warning
932 warn("Got %d activates in window %d (%d - %d) which is smaller "
933 "than %d\n", activationLimit, act_tick - actTicks[rank].back(),
934 act_tick, actTicks[rank].back(), tXAW);
935 }
936
937 // shift the times used for the book keeping, the last element
938 // (highest index) is the oldest one and hence the lowest value
939 actTicks[rank].pop_back();
940
941 // record an new activation (in the future)
942 actTicks[rank].push_front(act_tick);
943
944 // cannot activate more than X times in time window tXAW, push the
945 // next one (the X + 1'st activate) to be tXAW away from the
946 // oldest in our window of X
947 if (actTicks[rank].back() && (act_tick - actTicks[rank].back()) < tXAW) {
948 DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate no earlier "
949 "than %d\n", activationLimit, actTicks[rank].back() + tXAW);
950 for(int j = 0; j < banksPerRank; j++)
951 // next activate must not happen before end of window
952 banks[rank][j].actAllowedAt = actTicks[rank].back() + tXAW;
953 }
954}
955
956void
957DRAMCtrl::prechargeBank(Bank& bank, Tick free_at)
958{
959 // make sure the bank has an open row
960 assert(bank.openRow != Bank::NO_ROW);
961
962 // sample the bytes per activate here since we are closing
963 // the page
964 bytesPerActivate.sample(bank.bytesAccessed);
965
966 bank.openRow = Bank::NO_ROW;
967
968 bank.freeAt = free_at;
969
970 assert(numBanksActive != 0);
971 --numBanksActive;
972
973 DPRINTF(DRAM, "Precharged bank, done at tick %lld, now got %d active\n",
974 bank.freeAt, numBanksActive);
975
976 // if we reached zero, then special conditions apply as we track
977 // if all banks are precharged for the power models
978 if (numBanksActive == 0) {
979 idleStartTick = std::max(idleStartTick, bank.freeAt);
980 DPRINTF(DRAM, "All banks precharged at tick: %ld\n",
981 idleStartTick);
982 }
983}
984
985void
986DRAMCtrl::doDRAMAccess(DRAMPacket* dram_pkt)
987{
988
989 DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
990 dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
991
992 // estimate the bank and access latency
993 pair<Tick, Tick> lat = estimateLatency(dram_pkt, curTick());
994 Tick bankLat = lat.first;
995 Tick accessLat = lat.second;
996 Tick actTick;
997
998 // This request was woken up at this time based on a prior call
999 // to estimateLatency(). However, between then and now, both the
1000 // accessLatency and/or busBusyUntil may have changed. We need
1001 // to correct for that.
1002
1003 Tick addDelay = (curTick() + accessLat < busBusyUntil) ?
1004 busBusyUntil - (curTick() + accessLat) : 0;
1005
1006 Bank& bank = dram_pkt->bankRef;
1007
1008 // Update bank state
1009 if (pageMgmt == Enums::open || pageMgmt == Enums::open_adaptive ||
1010 pageMgmt == Enums::close_adaptive) {
1011
1012 if (rowHitFlag) {
1013 bank.freeAt = curTick() + addDelay + accessLat;
1014 } else {
1015 // If there is a page open, precharge it.
1016 if (bank.openRow != Bank::NO_ROW) {
1017 prechargeBank(bank, std::max(std::max(bank.freeAt,
1018 bank.tRASDoneAt),
1019 curTick()) + tRP);
1020 }
1021
1022 // Any precharge is already part of the latency
1023 // estimation, so update the bank free time
1024 bank.freeAt = curTick() + addDelay + accessLat;
1025
1026 // any waiting for banks account for in freeAt
1027 actTick = bank.freeAt - tCL - tRCD;
1028
1029 // If you activated a new row do to this access, the next access
1030 // will have to respect tRAS for this bank
1031 bank.tRASDoneAt = actTick + tRAS;
1032
1033 recordActivate(actTick, dram_pkt->rank, dram_pkt->bank,
1034 dram_pkt->row);
1035 }
1036
1037 // increment the bytes accessed and the accesses per row
1038 bank.bytesAccessed += burstSize;
1039 ++bank.rowAccesses;
1040
1041 // if we reached the max, then issue with an auto-precharge
1042 bool auto_precharge = bank.rowAccesses == maxAccessesPerRow;
1043
1044 // if we did not hit the limit, we might still want to
1045 // auto-precharge
1046 if (!auto_precharge &&
1047 (pageMgmt == Enums::open_adaptive ||
1048 pageMgmt == Enums::close_adaptive)) {
1049 // a twist on the open and close page policies:
1050 // 1) open_adaptive page policy does not blindly keep the
1051 // page open, but close it if there are no row hits, and there
1052 // are bank conflicts in the queue
1053 // 2) close_adaptive page policy does not blindly close the
1054 // page, but closes it only if there are no row hits in the queue.
1055 // In this case, only force an auto precharge when there
1056 // are no same page hits in the queue
1057 bool got_more_hits = false;
1058 bool got_bank_conflict = false;
1059
1060 // either look at the read queue or write queue
1061 const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
1062 writeQueue;
1063 auto p = queue.begin();
1064 // make sure we are not considering the packet that we are
1065 // currently dealing with (which is the head of the queue)
1066 ++p;
1067
1068 // keep on looking until we have found required condition or
1069 // reached the end
1070 while (!(got_more_hits &&
1071 (got_bank_conflict || pageMgmt == Enums::close_adaptive)) &&
1072 p != queue.end()) {
1073 bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
1074 (dram_pkt->bank == (*p)->bank);
1075 bool same_row = dram_pkt->row == (*p)->row;
1076 got_more_hits |= same_rank_bank && same_row;
1077 got_bank_conflict |= same_rank_bank && !same_row;
1078 ++p;
1079 }
1080
1081 // auto pre-charge when either
1082 // 1) open_adaptive policy, we have not got any more hits, and
1083 // have a bank conflict
1084 // 2) close_adaptive policy and we have not got any more hits
1085 auto_precharge = !got_more_hits &&
1086 (got_bank_conflict || pageMgmt == Enums::close_adaptive);
1087 }
1088
1089 // if this access should use auto-precharge, then we are
1090 // closing the row
1091 if (auto_precharge) {
1092 prechargeBank(bank, std::max(bank.freeAt, bank.tRASDoneAt) + tRP);
1093
1094 DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
1095 }
1096
1097 DPRINTF(DRAM, "doDRAMAccess::bank.freeAt is %lld\n", bank.freeAt);
1098 } else if (pageMgmt == Enums::close) {
1099 actTick = curTick() + addDelay + accessLat - tRCD - tCL;
1100 recordActivate(actTick, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
1101
1102 bank.freeAt = actTick + tRCD + tCL;
1103 bank.tRASDoneAt = actTick + tRAS;
1104
1105 // sample the relevant values when precharging
1106 bank.bytesAccessed = burstSize;
1107 bank.rowAccesses = 1;
1108
1109 prechargeBank(bank, std::max(bank.freeAt, bank.tRASDoneAt) + tRP);
1110 DPRINTF(DRAM, "doDRAMAccess::bank.freeAt is %lld\n", bank.freeAt);
1111 } else
1112 panic("No page management policy chosen\n");
1113
1114 // Update request parameters
1115 dram_pkt->readyTime = curTick() + addDelay + accessLat + tBURST;
1116
1117
1118 DPRINTF(DRAM, "Req %lld: curtick is %lld accessLat is %d " \
1119 "readytime is %lld busbusyuntil is %lld. " \
1120 "Scheduling at readyTime\n", dram_pkt->addr,
1121 curTick(), accessLat, dram_pkt->readyTime, busBusyUntil);
1122
1123 // Make sure requests are not overlapping on the databus
1124 assert(dram_pkt->readyTime - busBusyUntil >= tBURST);
1125
1126 // Update bus state
1127 busBusyUntil = dram_pkt->readyTime;
1128
1129 DPRINTF(DRAM,"Access time is %lld\n",
1130 dram_pkt->readyTime - dram_pkt->entryTime);
1131
1132 // Update the minimum timing between the requests, this is a
1133 // conservative estimate of when we have to schedule the next
1134 // request to not introduce any unecessary bubbles. In most cases
1135 // we will wake up sooner than we have to.
1136 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1137
1138 // Update the stats and schedule the next request
1139 if (dram_pkt->isRead) {
1140 ++readsThisTime;
1141 if (rowHitFlag)
1142 readRowHits++;
1143 bytesReadDRAM += burstSize;
1144 perBankRdBursts[dram_pkt->bankId]++;
1145
1146 // Update latency stats
1147 totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
1148 totBankLat += bankLat;
1149 totBusLat += tBURST;
1150 totQLat += dram_pkt->readyTime - dram_pkt->entryTime - bankLat -
1151 tBURST;
1152 } else {
1153 ++writesThisTime;
1154 if (rowHitFlag)
1155 writeRowHits++;
1156 bytesWritten += burstSize;
1157 perBankWrBursts[dram_pkt->bankId]++;
1158 }
1159}
1160
1161void
1162DRAMCtrl::moveToRespQ()
1163{
1164 // Remove from read queue
1165 DRAMPacket* dram_pkt = readQueue.front();
1166 readQueue.pop_front();
1167
1168 // sanity check
1169 assert(dram_pkt->size <= burstSize);
1170
1171 // Insert into response queue sorted by readyTime
1172 // It will be sent back to the requestor at its
1173 // readyTime
1174 if (respQueue.empty()) {
1175 respQueue.push_front(dram_pkt);
1176 assert(!respondEvent.scheduled());
1177 assert(dram_pkt->readyTime >= curTick());
1178 schedule(respondEvent, dram_pkt->readyTime);
1179 } else {
1180 bool done = false;
1181 auto i = respQueue.begin();
1182 while (!done && i != respQueue.end()) {
1183 if ((*i)->readyTime > dram_pkt->readyTime) {
1184 respQueue.insert(i, dram_pkt);
1185 done = true;
1186 }
1187 ++i;
1188 }
1189
1190 if (!done)
1191 respQueue.push_back(dram_pkt);
1192
1193 assert(respondEvent.scheduled());
1194
1195 if (respQueue.front()->readyTime < respondEvent.when()) {
1196 assert(respQueue.front()->readyTime >= curTick());
1197 reschedule(respondEvent, respQueue.front()->readyTime);
1198 }
1199 }
1200}
1201
1202void
1203DRAMCtrl::processNextReqEvent()
1204{
1205 if (busState == READ_TO_WRITE) {
1206 DPRINTF(DRAM, "Switching to writes after %d reads with %d reads "
1207 "waiting\n", readsThisTime, readQueue.size());
1208
1209 // sample and reset the read-related stats as we are now
1210 // transitioning to writes, and all reads are done
1211 rdPerTurnAround.sample(readsThisTime);
1212 readsThisTime = 0;
1213
1214 // now proceed to do the actual writes
1215 busState = WRITE;
1216 } else if (busState == WRITE_TO_READ) {
1217 DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
1218 "waiting\n", writesThisTime, writeQueue.size());
1219
1220 wrPerTurnAround.sample(writesThisTime);
1221 writesThisTime = 0;
1222
1223 busState = READ;
1224 }
1225
1226 if (refreshState != REF_IDLE) {
1227 // if a refresh waiting for this event loop to finish, then hand
1228 // over now, and do not schedule a new nextReqEvent
1229 if (refreshState == REF_DRAIN) {
1230 DPRINTF(DRAM, "Refresh drain done, now precharging\n");
1231
1232 refreshState = REF_PRE;
1233
1234 // hand control back to the refresh event loop
1235 schedule(refreshEvent, curTick());
1236 }
1237
1238 // let the refresh finish before issuing any further requests
1239 return;
1240 }
1241
1242 // when we get here it is either a read or a write
1243 if (busState == READ) {
1244
1245 // track if we should switch or not
1246 bool switch_to_writes = false;
1247
1248 if (readQueue.empty()) {
1249 // In the case there is no read request to go next,
1250 // trigger writes if we have passed the low threshold (or
1251 // if we are draining)
1252 if (!writeQueue.empty() &&
1253 (drainManager || writeQueue.size() > writeLowThreshold)) {
1254
1255 switch_to_writes = true;
1256 } else {
1257 // check if we are drained
1258 if (respQueue.empty () && drainManager) {
1259 drainManager->signalDrainDone();
1260 drainManager = NULL;
1261 }
1262
1263 // nothing to do, not even any point in scheduling an
1264 // event for the next request
1265 return;
1266 }
1267 } else {
1268 // Figure out which read request goes next, and move it to the
1269 // front of the read queue
1270 chooseNext(readQueue);
1271
1272 doDRAMAccess(readQueue.front());
1273
1274 // At this point we're done dealing with the request
1275 // It will be moved to a separate response queue with a
1276 // correct readyTime, and eventually be sent back at that
1277 // time
1278 moveToRespQ();
1279
1280 // we have so many writes that we have to transition
1281 if (writeQueue.size() > writeHighThreshold) {
1282 switch_to_writes = true;
1283 }
1284 }
1285
1286 // switching to writes, either because the read queue is empty
1287 // and the writes have passed the low threshold (or we are
1288 // draining), or because the writes hit the hight threshold
1289 if (switch_to_writes) {
1290 // transition to writing
1291 busState = READ_TO_WRITE;
1292
1293 // add a bubble to the data bus, as defined by the
1294 // tRTW parameter
1295 busBusyUntil += tRTW;
1296
1297 // update the minimum timing between the requests,
1298 // this shifts us back in time far enough to do any
1299 // bank preparation
1300 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1301 }
1302 } else {
1303 chooseNext(writeQueue);
1304 DRAMPacket* dram_pkt = writeQueue.front();
1305 // sanity check
1306 assert(dram_pkt->size <= burstSize);
1307 doDRAMAccess(dram_pkt);
1308
1309 writeQueue.pop_front();
1310 delete dram_pkt;
1311
1312 // If we emptied the write queue, or got sufficiently below the
1313 // threshold (using the minWritesPerSwitch as the hysteresis) and
1314 // are not draining, or we have reads waiting and have done enough
1315 // writes, then switch to reads.
1316 if (writeQueue.empty() ||
1317 (writeQueue.size() + minWritesPerSwitch < writeLowThreshold &&
1318 !drainManager) ||
1319 (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
1320 // turn the bus back around for reads again
1321 busState = WRITE_TO_READ;
1322
1323 // note that the we switch back to reads also in the idle
1324 // case, which eventually will check for any draining and
1325 // also pause any further scheduling if there is really
1326 // nothing to do
1327
1328 // here we get a bit creative and shift the bus busy time not
1329 // just the tWTR, but also a CAS latency to capture the fact
1330 // that we are allowed to prepare a new bank, but not issue a
1331 // read command until after tWTR, in essence we capture a
1332 // bubble on the data bus that is tWTR + tCL
1333 busBusyUntil += tWTR + tCL;
1334
1335 // update the minimum timing between the requests, this shifts
1336 // us back in time far enough to do any bank preparation
1337 nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
1338 }
1339 }
1340
1341 schedule(nextReqEvent, std::max(nextReqTime, curTick()));
1342
1343 // If there is space available and we have writes waiting then let
1344 // them retry. This is done here to ensure that the retry does not
1345 // cause a nextReqEvent to be scheduled before we do so as part of
1346 // the next request processing
1347 if (retryWrReq && writeQueue.size() < writeBufferSize) {
1348 retryWrReq = false;
1349 port.sendRetry();
1350 }
1351}
1352
1353uint64_t
1354DRAMCtrl::minBankFreeAt(const deque<DRAMPacket*>& queue) const
1355{
1356 uint64_t bank_mask = 0;
1357 Tick freeAt = MaxTick;
1358
1359 // detemrine if we have queued transactions targetting the
1360 // bank in question
1361 vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
1362 for (auto p = queue.begin(); p != queue.end(); ++p) {
1363 got_waiting[(*p)->bankId] = true;
1364 }
1365
1366 for (int i = 0; i < ranksPerChannel; i++) {
1367 for (int j = 0; j < banksPerRank; j++) {
1368 // if we have waiting requests for the bank, and it is
1369 // amongst the first available, update the mask
1370 if (got_waiting[i * banksPerRank + j] &&
1371 banks[i][j].freeAt <= freeAt) {
1372 // reset bank mask if new minimum is found
1373 if (banks[i][j].freeAt < freeAt)
1374 bank_mask = 0;
1375 // set the bit corresponding to the available bank
1376 uint8_t bit_index = i * ranksPerChannel + j;
1377 replaceBits(bank_mask, bit_index, bit_index, 1);
1378 freeAt = banks[i][j].freeAt;
1379 }
1380 }
1381 }
1382 return bank_mask;
1383}
1384
1385void
1386DRAMCtrl::processRefreshEvent()
1387{
1388 // when first preparing the refresh, remember when it was due
1389 if (refreshState == REF_IDLE) {
1390 // remember when the refresh is due
1391 refreshDueAt = curTick();
1392
1393 // proceed to drain
1394 refreshState = REF_DRAIN;
1395
1396 DPRINTF(DRAM, "Refresh due\n");
1397 }
1398
1399 // let any scheduled read or write go ahead, after which it will
1400 // hand control back to this event loop
1401 if (refreshState == REF_DRAIN) {
1402 if (nextReqEvent.scheduled()) {
1403 // hand control over to the request loop until it is
1404 // evaluated next
1405 DPRINTF(DRAM, "Refresh awaiting draining\n");
1406
1407 return;
1408 } else {
1409 refreshState = REF_PRE;
1410 }
1411 }
1412
1413 // at this point, ensure that all banks are precharged
1414 if (refreshState == REF_PRE) {
1415 DPRINTF(DRAM, "Precharging all\n");
1416
1417 // precharge any active bank
1418 for (int i = 0; i < ranksPerChannel; i++) {
1419 for (int j = 0; j < banksPerRank; j++) {
1420 if (banks[i][j].openRow != Bank::NO_ROW) {
1421 // respect both causality and any existing bank
1422 // constraints
1423 Tick free_at = std::max(std::max(banks[i][j].freeAt,
1424 banks[i][j].tRASDoneAt),
1425 curTick()) + tRP;
1426
1427 prechargeBank(banks[i][j], free_at);
1428 }
1429 }
1430 }
1431
1432 if (numBanksActive != 0)
1433 panic("Refresh scheduled with %d active banks\n", numBanksActive);
1434
1435 // advance the state
1436 refreshState = REF_RUN;
1437
1438 // call ourselves in the future
1439 schedule(refreshEvent, std::max(curTick(), idleStartTick));
1440 return;
1441 }
1442
1443 // last but not least we perform the actual refresh
1444 if (refreshState == REF_RUN) {
1445 // should never get here with any banks active
1446 assert(numBanksActive == 0);
1447
1448 Tick banksFree = curTick() + tRFC;
1449
1450 for (int i = 0; i < ranksPerChannel; i++) {
1451 for (int j = 0; j < banksPerRank; j++) {
1452 banks[i][j].freeAt = banksFree;
1453 }
1454 }
1455
1456 // make sure we did not wait so long that we cannot make up
1457 // for it
1458 if (refreshDueAt + tREFI < banksFree) {
1459 fatal("Refresh was delayed so long we cannot catch up\n");
1460 }
1461
1462 // compensate for the delay in actually performing the refresh
1463 // when scheduling the next one
1464 schedule(refreshEvent, refreshDueAt + tREFI - tRP);
1465
1466 // back to business as usual
1467 refreshState = REF_IDLE;
1468
1469 // we are now refreshing until tRFC is done
1470 idleStartTick = banksFree;
1471
1472 // kick the normal request processing loop into action again
1473 // as early as possible, i.e. when the request is done, the
1474 // scheduling of this event also prevents any new requests
1475 // from going ahead before the scheduled point in time
1476 nextReqTime = banksFree;
1477 schedule(nextReqEvent, nextReqTime);
1478 }
1479}
1480
1481void
1482DRAMCtrl::regStats()
1483{
1484 using namespace Stats;
1485
1486 AbstractMemory::regStats();
1487
1488 readReqs
1489 .name(name() + ".readReqs")
1490 .desc("Number of read requests accepted");
1491
1492 writeReqs
1493 .name(name() + ".writeReqs")
1494 .desc("Number of write requests accepted");
1495
1496 readBursts
1497 .name(name() + ".readBursts")
1498 .desc("Number of DRAM read bursts, "
1499 "including those serviced by the write queue");
1500
1501 writeBursts
1502 .name(name() + ".writeBursts")
1503 .desc("Number of DRAM write bursts, "
1504 "including those merged in the write queue");
1505
1506 servicedByWrQ
1507 .name(name() + ".servicedByWrQ")
1508 .desc("Number of DRAM read bursts serviced by the write queue");
1509
1510 mergedWrBursts
1511 .name(name() + ".mergedWrBursts")
1512 .desc("Number of DRAM write bursts merged with an existing one");
1513
1514 neitherReadNorWrite
1515 .name(name() + ".neitherReadNorWriteReqs")
1516 .desc("Number of requests that are neither read nor write");
1517
1518 perBankRdBursts
1519 .init(banksPerRank * ranksPerChannel)
1520 .name(name() + ".perBankRdBursts")
1521 .desc("Per bank write bursts");
1522
1523 perBankWrBursts
1524 .init(banksPerRank * ranksPerChannel)
1525 .name(name() + ".perBankWrBursts")
1526 .desc("Per bank write bursts");
1527
1528 avgRdQLen
1529 .name(name() + ".avgRdQLen")
1530 .desc("Average read queue length when enqueuing")
1531 .precision(2);
1532
1533 avgWrQLen
1534 .name(name() + ".avgWrQLen")
1535 .desc("Average write queue length when enqueuing")
1536 .precision(2);
1537
1538 totQLat
1539 .name(name() + ".totQLat")
1540 .desc("Total ticks spent queuing");
1541
1542 totBankLat
1543 .name(name() + ".totBankLat")
1544 .desc("Total ticks spent accessing banks");
1545
1546 totBusLat
1547 .name(name() + ".totBusLat")
1548 .desc("Total ticks spent in databus transfers");
1549
1550 totMemAccLat
1551 .name(name() + ".totMemAccLat")
1552 .desc("Total ticks spent from burst creation until serviced "
1553 "by the DRAM");
1554
1555 avgQLat
1556 .name(name() + ".avgQLat")
1557 .desc("Average queueing delay per DRAM burst")
1558 .precision(2);
1559
1560 avgQLat = totQLat / (readBursts - servicedByWrQ);
1561
1562 avgBankLat
1563 .name(name() + ".avgBankLat")
1564 .desc("Average bank access latency per DRAM burst")
1565 .precision(2);
1566
1567 avgBankLat = totBankLat / (readBursts - servicedByWrQ);
1568
1569 avgBusLat
1570 .name(name() + ".avgBusLat")
1571 .desc("Average bus latency per DRAM burst")
1572 .precision(2);
1573
1574 avgBusLat = totBusLat / (readBursts - servicedByWrQ);
1575
1576 avgMemAccLat
1577 .name(name() + ".avgMemAccLat")
1578 .desc("Average memory access latency per DRAM burst")
1579 .precision(2);
1580
1581 avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
1582
1583 numRdRetry
1584 .name(name() + ".numRdRetry")
1585 .desc("Number of times read queue was full causing retry");
1586
1587 numWrRetry
1588 .name(name() + ".numWrRetry")
1589 .desc("Number of times write queue was full causing retry");
1590
1591 readRowHits
1592 .name(name() + ".readRowHits")
1593 .desc("Number of row buffer hits during reads");
1594
1595 writeRowHits
1596 .name(name() + ".writeRowHits")
1597 .desc("Number of row buffer hits during writes");
1598
1599 readRowHitRate
1600 .name(name() + ".readRowHitRate")
1601 .desc("Row buffer hit rate for reads")
1602 .precision(2);
1603
1604 readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
1605
1606 writeRowHitRate
1607 .name(name() + ".writeRowHitRate")
1608 .desc("Row buffer hit rate for writes")
1609 .precision(2);
1610
1611 writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
1612
1613 readPktSize
1614 .init(ceilLog2(burstSize) + 1)
1615 .name(name() + ".readPktSize")
1616 .desc("Read request sizes (log2)");
1617
1618 writePktSize
1619 .init(ceilLog2(burstSize) + 1)
1620 .name(name() + ".writePktSize")
1621 .desc("Write request sizes (log2)");
1622
1623 rdQLenPdf
1624 .init(readBufferSize)
1625 .name(name() + ".rdQLenPdf")
1626 .desc("What read queue length does an incoming req see");
1627
1628 wrQLenPdf
1629 .init(writeBufferSize)
1630 .name(name() + ".wrQLenPdf")
1631 .desc("What write queue length does an incoming req see");
1632
1633 bytesPerActivate
1634 .init(maxAccessesPerRow)
1635 .name(name() + ".bytesPerActivate")
1636 .desc("Bytes accessed per row activation")
1637 .flags(nozero);
1638
1639 rdPerTurnAround
1640 .init(readBufferSize)
1641 .name(name() + ".rdPerTurnAround")
1642 .desc("Reads before turning the bus around for writes")
1643 .flags(nozero);
1644
1645 wrPerTurnAround
1646 .init(writeBufferSize)
1647 .name(name() + ".wrPerTurnAround")
1648 .desc("Writes before turning the bus around for reads")
1649 .flags(nozero);
1650
1651 bytesReadDRAM
1652 .name(name() + ".bytesReadDRAM")
1653 .desc("Total number of bytes read from DRAM");
1654
1655 bytesReadWrQ
1656 .name(name() + ".bytesReadWrQ")
1657 .desc("Total number of bytes read from write queue");
1658
1659 bytesWritten
1660 .name(name() + ".bytesWritten")
1661 .desc("Total number of bytes written to DRAM");
1662
1663 bytesReadSys
1664 .name(name() + ".bytesReadSys")
1665 .desc("Total read bytes from the system interface side");
1666
1667 bytesWrittenSys
1668 .name(name() + ".bytesWrittenSys")
1669 .desc("Total written bytes from the system interface side");
1670
1671 avgRdBW
1672 .name(name() + ".avgRdBW")
1673 .desc("Average DRAM read bandwidth in MiByte/s")
1674 .precision(2);
1675
1676 avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
1677
1678 avgWrBW
1679 .name(name() + ".avgWrBW")
1680 .desc("Average achieved write bandwidth in MiByte/s")
1681 .precision(2);
1682
1683 avgWrBW = (bytesWritten / 1000000) / simSeconds;
1684
1685 avgRdBWSys
1686 .name(name() + ".avgRdBWSys")
1687 .desc("Average system read bandwidth in MiByte/s")
1688 .precision(2);
1689
1690 avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
1691
1692 avgWrBWSys
1693 .name(name() + ".avgWrBWSys")
1694 .desc("Average system write bandwidth in MiByte/s")
1695 .precision(2);
1696
1697 avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
1698
1699 peakBW
1700 .name(name() + ".peakBW")
1701 .desc("Theoretical peak bandwidth in MiByte/s")
1702 .precision(2);
1703
1704 peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
1705
1706 busUtil
1707 .name(name() + ".busUtil")
1708 .desc("Data bus utilization in percentage")
1709 .precision(2);
1710
1711 busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
1712
1713 totGap
1714 .name(name() + ".totGap")
1715 .desc("Total gap between requests");
1716
1717 avgGap
1718 .name(name() + ".avgGap")
1719 .desc("Average gap between requests")
1720 .precision(2);
1721
1722 avgGap = totGap / (readReqs + writeReqs);
1723
1724 // Stats for DRAM Power calculation based on Micron datasheet
1725 busUtilRead
1726 .name(name() + ".busUtilRead")
1727 .desc("Data bus utilization in percentage for reads")
1728 .precision(2);
1729
1730 busUtilRead = avgRdBW / peakBW * 100;
1731
1732 busUtilWrite
1733 .name(name() + ".busUtilWrite")
1734 .desc("Data bus utilization in percentage for writes")
1735 .precision(2);
1736
1737 busUtilWrite = avgWrBW / peakBW * 100;
1738
1739 pageHitRate
1740 .name(name() + ".pageHitRate")
1741 .desc("Row buffer hit rate, read and write combined")
1742 .precision(2);
1743
1744 pageHitRate = (writeRowHits + readRowHits) /
1745 (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
1746
1747 prechargeAllPercent
1748 .name(name() + ".prechargeAllPercent")
1749 .desc("Percentage of time for which DRAM has all the banks in "
1750 "precharge state")
1751 .precision(2);
1752
1753 prechargeAllPercent = prechargeAllTime / simTicks * 100;
1754}
1755
1756void
1757DRAMCtrl::recvFunctional(PacketPtr pkt)
1758{
1759 // rely on the abstract memory
1760 functionalAccess(pkt);
1761}
1762
1763BaseSlavePort&
1764DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
1765{
1766 if (if_name != "port") {
1767 return MemObject::getSlavePort(if_name, idx);
1768 } else {
1769 return port;
1770 }
1771}
1772
1773unsigned int
1774DRAMCtrl::drain(DrainManager *dm)
1775{
1776 unsigned int count = port.drain(dm);
1777
1778 // if there is anything in any of our internal queues, keep track
1779 // of that as well
1780 if (!(writeQueue.empty() && readQueue.empty() &&
1781 respQueue.empty())) {
1782 DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
1783 " resp: %d\n", writeQueue.size(), readQueue.size(),
1784 respQueue.size());
1785 ++count;
1786 drainManager = dm;
1787
1788 // the only part that is not drained automatically over time
1789 // is the write queue, thus kick things into action if needed
1790 if (!writeQueue.empty() && !nextReqEvent.scheduled()) {
1791 schedule(nextReqEvent, curTick());
1792 }
1793 }
1794
1795 if (count)
1796 setDrainState(Drainable::Draining);
1797 else
1798 setDrainState(Drainable::Drained);
1799 return count;
1800}
1801
1802DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
1803 : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
1804 memory(_memory)
1805{ }
1806
1807AddrRangeList
1808DRAMCtrl::MemoryPort::getAddrRanges() const
1809{
1810 AddrRangeList ranges;
1811 ranges.push_back(memory.getAddrRange());
1812 return ranges;
1813}
1814
1815void
1816DRAMCtrl::MemoryPort::recvFunctional(PacketPtr pkt)
1817{
1818 pkt->pushLabel(memory.name());
1819
1820 if (!queue.checkFunctional(pkt)) {
1821 // Default implementation of SimpleTimingPort::recvFunctional()
1822 // calls recvAtomic() and throws away the latency; we can save a
1823 // little here by just not calculating the latency.
1824 memory.recvFunctional(pkt);
1825 }
1826
1827 pkt->popLabel();
1828}
1829
1830Tick
1831DRAMCtrl::MemoryPort::recvAtomic(PacketPtr pkt)
1832{
1833 return memory.recvAtomic(pkt);
1834}
1835
1836bool
1837DRAMCtrl::MemoryPort::recvTimingReq(PacketPtr pkt)
1838{
1839 // pass it to the memory controller
1840 return memory.recvTimingReq(pkt);
1841}
1842
1843DRAMCtrl*
1844DRAMCtrlParams::create()
1845{
1846 return new DRAMCtrl(this);
1847}