dram_ctrl.hh revision 10287:4966471a1ba1
1/*
2 * Copyright (c) 2012-2014 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/**
46 * @file
47 * DRAMCtrl declaration
48 */
49
50#ifndef __MEM_DRAM_CTRL_HH__
51#define __MEM_DRAM_CTRL_HH__
52
53#include <deque>
54
55#include "base/statistics.hh"
56#include "enums/AddrMap.hh"
57#include "enums/MemSched.hh"
58#include "enums/PageManage.hh"
59#include "mem/abstract_mem.hh"
60#include "mem/qport.hh"
61#include "params/DRAMCtrl.hh"
62#include "sim/eventq.hh"
63
64/**
65 * The DRAM controller is a single-channel memory controller capturing
66 * the most important timing constraints associated with a
67 * contemporary DRAM. For multi-channel memory systems, the controller
68 * is combined with a crossbar model, with the channel address
69 * interleaving taking part in the crossbar.
70 *
71 * As a basic design principle, this controller
72 * model is not cycle callable, but instead uses events to: 1) decide
73 * when new decisions can be made, 2) when resources become available,
74 * 3) when things are to be considered done, and 4) when to send
75 * things back. Through these simple principles, the model delivers
76 * high performance, and lots of flexibility, allowing users to
77 * evaluate the system impact of a wide range of memory technologies,
78 * such as DDR3/4, LPDDR2/3/4, WideIO1/2, HBM and HMC.
79 *
80 * For more details, please see Hansson et al, "Simulating DRAM
81 * controllers for future system architecture exploration",
82 * Proc. ISPASS, 2014. If you use this model as part of your research
83 * please cite the paper.
84 */
85class DRAMCtrl : public AbstractMemory
86{
87
88  private:
89
90    // For now, make use of a queued slave port to avoid dealing with
91    // flow control for the responses being sent back
92    class MemoryPort : public QueuedSlavePort
93    {
94
95        SlavePacketQueue queue;
96        DRAMCtrl& memory;
97
98      public:
99
100        MemoryPort(const std::string& name, DRAMCtrl& _memory);
101
102      protected:
103
104        Tick recvAtomic(PacketPtr pkt);
105
106        void recvFunctional(PacketPtr pkt);
107
108        bool recvTimingReq(PacketPtr);
109
110        virtual AddrRangeList getAddrRanges() const;
111
112    };
113
114    /**
115     * Our incoming port, for a multi-ported controller add a crossbar
116     * in front of it
117     */
118    MemoryPort port;
119
120    /**
121     * Remember if we have to retry a request when available.
122     */
123    bool retryRdReq;
124    bool retryWrReq;
125
126    /**
127     * Bus state used to control the read/write switching and drive
128     * the scheduling of the next request.
129     */
130    enum BusState {
131        READ = 0,
132        READ_TO_WRITE,
133        WRITE,
134        WRITE_TO_READ
135    };
136
137    BusState busState;
138
139    /** List to keep track of activate ticks */
140    std::vector<std::deque<Tick>> actTicks;
141
142    /**
143     * A basic class to track the bank state, i.e. what row is
144     * currently open (if any), when is the bank free to accept a new
145     * column (read/write) command, when can it be precharged, and
146     * when can it be activated.
147     *
148     * The bank also keeps track of how many bytes have been accessed
149     * in the open row since it was opened.
150     */
151    class Bank
152    {
153
154      public:
155
156        static const uint32_t NO_ROW = -1;
157
158        uint32_t openRow;
159        uint8_t rank;
160        uint8_t bank;
161
162        Tick colAllowedAt;
163        Tick preAllowedAt;
164        Tick actAllowedAt;
165
166        uint32_t rowAccesses;
167        uint32_t bytesAccessed;
168
169        Bank() :
170            openRow(NO_ROW), rank(0), bank(0),
171            colAllowedAt(0), preAllowedAt(0), actAllowedAt(0),
172            rowAccesses(0), bytesAccessed(0)
173        { }
174    };
175
176    /**
177     * A burst helper helps organize and manage a packet that is larger than
178     * the DRAM burst size. A system packet that is larger than the burst size
179     * is split into multiple DRAM packets and all those DRAM packets point to
180     * a single burst helper such that we know when the whole packet is served.
181     */
182    class BurstHelper {
183
184      public:
185
186        /** Number of DRAM bursts requred for a system packet **/
187        const unsigned int burstCount;
188
189        /** Number of DRAM bursts serviced so far for a system packet **/
190        unsigned int burstsServiced;
191
192        BurstHelper(unsigned int _burstCount)
193            : burstCount(_burstCount), burstsServiced(0)
194            { }
195    };
196
197    /**
198     * A DRAM packet stores packets along with the timestamp of when
199     * the packet entered the queue, and also the decoded address.
200     */
201    class DRAMPacket {
202
203      public:
204
205        /** When did request enter the controller */
206        const Tick entryTime;
207
208        /** When will request leave the controller */
209        Tick readyTime;
210
211        /** This comes from the outside world */
212        const PacketPtr pkt;
213
214        const bool isRead;
215
216        /** Will be populated by address decoder */
217        const uint8_t rank;
218        const uint8_t bank;
219        const uint32_t row;
220
221        /**
222         * Bank id is calculated considering banks in all the ranks
223         * eg: 2 ranks each with 8 banks, then bankId = 0 --> rank0, bank0 and
224         * bankId = 8 --> rank1, bank0
225         */
226        const uint16_t bankId;
227
228        /**
229         * The starting address of the DRAM packet.
230         * This address could be unaligned to burst size boundaries. The
231         * reason is to keep the address offset so we can accurately check
232         * incoming read packets with packets in the write queue.
233         */
234        Addr addr;
235
236        /**
237         * The size of this dram packet in bytes
238         * It is always equal or smaller than DRAM burst size
239         */
240        unsigned int size;
241
242        /**
243         * A pointer to the BurstHelper if this DRAMPacket is a split packet
244         * If not a split packet (common case), this is set to NULL
245         */
246        BurstHelper* burstHelper;
247        Bank& bankRef;
248
249        DRAMPacket(PacketPtr _pkt, bool is_read, uint8_t _rank, uint8_t _bank,
250                   uint32_t _row, uint16_t bank_id, Addr _addr,
251                   unsigned int _size, Bank& bank_ref)
252            : entryTime(curTick()), readyTime(curTick()),
253              pkt(_pkt), isRead(is_read), rank(_rank), bank(_bank), row(_row),
254              bankId(bank_id), addr(_addr), size(_size), burstHelper(NULL),
255              bankRef(bank_ref)
256        { }
257
258    };
259
260    /**
261     * Bunch of things requires to setup "events" in gem5
262     * When event "respondEvent" occurs for example, the method
263     * processRespondEvent is called; no parameters are allowed
264     * in these methods
265     */
266    void processNextReqEvent();
267    EventWrapper<DRAMCtrl,&DRAMCtrl::processNextReqEvent> nextReqEvent;
268
269    void processRespondEvent();
270    EventWrapper<DRAMCtrl, &DRAMCtrl::processRespondEvent> respondEvent;
271
272    void processActivateEvent();
273    EventWrapper<DRAMCtrl, &DRAMCtrl::processActivateEvent> activateEvent;
274
275    void processPrechargeEvent();
276    EventWrapper<DRAMCtrl, &DRAMCtrl::processPrechargeEvent> prechargeEvent;
277
278    void processRefreshEvent();
279    EventWrapper<DRAMCtrl, &DRAMCtrl::processRefreshEvent> refreshEvent;
280
281    void processPowerEvent();
282    EventWrapper<DRAMCtrl,&DRAMCtrl::processPowerEvent> powerEvent;
283
284    /**
285     * Check if the read queue has room for more entries
286     *
287     * @param pktCount The number of entries needed in the read queue
288     * @return true if read queue is full, false otherwise
289     */
290    bool readQueueFull(unsigned int pktCount) const;
291
292    /**
293     * Check if the write queue has room for more entries
294     *
295     * @param pktCount The number of entries needed in the write queue
296     * @return true if write queue is full, false otherwise
297     */
298    bool writeQueueFull(unsigned int pktCount) const;
299
300    /**
301     * When a new read comes in, first check if the write q has a
302     * pending request to the same address.\ If not, decode the
303     * address to populate rank/bank/row, create one or mutliple
304     * "dram_pkt", and push them to the back of the read queue.\
305     * If this is the only
306     * read request in the system, schedule an event to start
307     * servicing it.
308     *
309     * @param pkt The request packet from the outside world
310     * @param pktCount The number of DRAM bursts the pkt
311     * translate to. If pkt size is larger then one full burst,
312     * then pktCount is greater than one.
313     */
314    void addToReadQueue(PacketPtr pkt, unsigned int pktCount);
315
316    /**
317     * Decode the incoming pkt, create a dram_pkt and push to the
318     * back of the write queue. \If the write q length is more than
319     * the threshold specified by the user, ie the queue is beginning
320     * to get full, stop reads, and start draining writes.
321     *
322     * @param pkt The request packet from the outside world
323     * @param pktCount The number of DRAM bursts the pkt
324     * translate to. If pkt size is larger then one full burst,
325     * then pktCount is greater than one.
326     */
327    void addToWriteQueue(PacketPtr pkt, unsigned int pktCount);
328
329    /**
330     * Actually do the DRAM access - figure out the latency it
331     * will take to service the req based on bank state, channel state etc
332     * and then update those states to account for this request.\ Based
333     * on this, update the packet's "readyTime" and move it to the
334     * response q from where it will eventually go back to the outside
335     * world.
336     *
337     * @param pkt The DRAM packet created from the outside world pkt
338     */
339    void doDRAMAccess(DRAMPacket* dram_pkt);
340
341    /**
342     * When a packet reaches its "readyTime" in the response Q,
343     * use the "access()" method in AbstractMemory to actually
344     * create the response packet, and send it back to the outside
345     * world requestor.
346     *
347     * @param pkt The packet from the outside world
348     * @param static_latency Static latency to add before sending the packet
349     */
350    void accessAndRespond(PacketPtr pkt, Tick static_latency);
351
352    /**
353     * Address decoder to figure out physical mapping onto ranks,
354     * banks, and rows. This function is called multiple times on the same
355     * system packet if the pakcet is larger than burst of the memory. The
356     * dramPktAddr is used for the offset within the packet.
357     *
358     * @param pkt The packet from the outside world
359     * @param dramPktAddr The starting address of the DRAM packet
360     * @param size The size of the DRAM packet in bytes
361     * @param isRead Is the request for a read or a write to DRAM
362     * @return A DRAMPacket pointer with the decoded information
363     */
364    DRAMPacket* decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned int size,
365                           bool isRead);
366
367    /**
368     * The memory schduler/arbiter - picks which request needs to
369     * go next, based on the specified policy such as FCFS or FR-FCFS
370     * and moves it to the head of the queue.
371     */
372    void chooseNext(std::deque<DRAMPacket*>& queue);
373
374    /**
375     * For FR-FCFS policy reorder the read/write queue depending on row buffer
376     * hits and earliest banks available in DRAM
377     */
378    void reorderQueue(std::deque<DRAMPacket*>& queue);
379
380    /**
381     * Find which are the earliest banks ready to issue an activate
382     * for the enqueued requests. Assumes maximum of 64 banks per DIMM
383     *
384     * @param Queued requests to consider
385     * @return One-hot encoded mask of bank indices
386     */
387    uint64_t minBankActAt(const std::deque<DRAMPacket*>& queue) const;
388
389    /**
390     * Keep track of when row activations happen, in order to enforce
391     * the maximum number of activations in the activation window. The
392     * method updates the time that the banks become available based
393     * on the current limits.
394     *
395     * @param bank Reference to the bank
396     * @param act_tick Time when the activation takes place
397     * @param row Index of the row
398     */
399    void activateBank(Bank& bank, Tick act_tick, uint32_t row);
400
401    /**
402     * Precharge a given bank and also update when the precharge is
403     * done. This will also deal with any stats related to the
404     * accesses to the open page.
405     *
406     * @param bank_ref The bank to precharge
407     * @param pre_at Time when the precharge takes place
408     * @param trace Is this an auto precharge then do not add to trace
409     */
410    void prechargeBank(Bank& bank_ref, Tick pre_at,  bool trace = true);
411
412    /**
413     * Used for debugging to observe the contents of the queues.
414     */
415    void printQs() const;
416
417    /**
418     * The controller's main read and write queues
419     */
420    std::deque<DRAMPacket*> readQueue;
421    std::deque<DRAMPacket*> writeQueue;
422
423    /**
424     * Response queue where read packets wait after we're done working
425     * with them, but it's not time to send the response yet. The
426     * responses are stored seperately mostly to keep the code clean
427     * and help with events scheduling. For all logical purposes such
428     * as sizing the read queue, this and the main read queue need to
429     * be added together.
430     */
431    std::deque<DRAMPacket*> respQueue;
432
433    /**
434     * If we need to drain, keep the drain manager around until we're
435     * done here.
436     */
437    DrainManager *drainManager;
438
439    /**
440     * Multi-dimensional vector of banks, first dimension is ranks,
441     * second is bank
442     */
443    std::vector<std::vector<Bank> > banks;
444
445    /**
446     * The following are basic design parameters of the memory
447     * controller, and are initialized based on parameter values.
448     * The rowsPerBank is determined based on the capacity, number of
449     * ranks and banks, the burst size, and the row buffer size.
450     */
451    const uint32_t deviceBusWidth;
452    const uint32_t burstLength;
453    const uint32_t deviceRowBufferSize;
454    const uint32_t devicesPerRank;
455    const uint32_t burstSize;
456    const uint32_t rowBufferSize;
457    const uint32_t columnsPerRowBuffer;
458    const uint32_t columnsPerStripe;
459    const uint32_t ranksPerChannel;
460    const uint32_t banksPerRank;
461    const uint32_t channels;
462    uint32_t rowsPerBank;
463    const uint32_t readBufferSize;
464    const uint32_t writeBufferSize;
465    const uint32_t writeHighThreshold;
466    const uint32_t writeLowThreshold;
467    const uint32_t minWritesPerSwitch;
468    uint32_t writesThisTime;
469    uint32_t readsThisTime;
470
471    /**
472     * Basic memory timing parameters initialized based on parameter
473     * values.
474     */
475    const Tick M5_CLASS_VAR_USED tCK;
476    const Tick tWTR;
477    const Tick tRTW;
478    const Tick tBURST;
479    const Tick tRCD;
480    const Tick tCL;
481    const Tick tRP;
482    const Tick tRAS;
483    const Tick tWR;
484    const Tick tRTP;
485    const Tick tRFC;
486    const Tick tREFI;
487    const Tick tRRD;
488    const Tick tXAW;
489    const uint32_t activationLimit;
490
491    /**
492     * Memory controller configuration initialized based on parameter
493     * values.
494     */
495    Enums::MemSched memSchedPolicy;
496    Enums::AddrMap addrMapping;
497    Enums::PageManage pageMgmt;
498
499    /**
500     * Max column accesses (read and write) per row, before forefully
501     * closing it.
502     */
503    const uint32_t maxAccessesPerRow;
504
505    /**
506     * Pipeline latency of the controller frontend. The frontend
507     * contribution is added to writes (that complete when they are in
508     * the write buffer) and reads that are serviced the write buffer.
509     */
510    const Tick frontendLatency;
511
512    /**
513     * Pipeline latency of the backend and PHY. Along with the
514     * frontend contribution, this latency is added to reads serviced
515     * by the DRAM.
516     */
517    const Tick backendLatency;
518
519    /**
520     * Till when has the main data bus been spoken for already?
521     */
522    Tick busBusyUntil;
523
524    /**
525     * Keep track of when a refresh is due.
526     */
527    Tick refreshDueAt;
528
529    /**
530     * The refresh state is used to control the progress of the
531     * refresh scheduling. When normal operation is in progress the
532     * refresh state is idle. From there, it progresses to the refresh
533     * drain state once tREFI has passed. The refresh drain state
534     * captures the DRAM row active state, as it will stay there until
535     * all ongoing accesses complete. Thereafter all banks are
536     * precharged, and lastly, the DRAM is refreshed.
537     */
538    enum RefreshState {
539        REF_IDLE = 0,
540        REF_DRAIN,
541        REF_PRE,
542        REF_RUN
543    };
544
545    RefreshState refreshState;
546
547    /**
548     * The power state captures the different operational states of
549     * the DRAM and interacts with the bus read/write state machine,
550     * and the refresh state machine. In the idle state all banks are
551     * precharged. From there we either go to an auto refresh (as
552     * determined by the refresh state machine), or to a precharge
553     * power down mode. From idle the memory can also go to the active
554     * state (with one or more banks active), and in turn from there
555     * to active power down. At the moment we do not capture the deep
556     * power down and self-refresh state.
557     */
558    enum PowerState {
559        PWR_IDLE = 0,
560        PWR_REF,
561        PWR_PRE_PDN,
562        PWR_ACT,
563        PWR_ACT_PDN
564    };
565
566    /**
567     * Since we are taking decisions out of order, we need to keep
568     * track of what power transition is happening at what time, such
569     * that we can go back in time and change history. For example, if
570     * we precharge all banks and schedule going to the idle state, we
571     * might at a later point decide to activate a bank before the
572     * transition to idle would have taken place.
573     */
574    PowerState pwrStateTrans;
575
576    /**
577     * Current power state.
578     */
579    PowerState pwrState;
580
581    /**
582     * Schedule a power state transition in the future, and
583     * potentially override an already scheduled transition.
584     *
585     * @param pwr_state Power state to transition to
586     * @param tick Tick when transition should take place
587     */
588    void schedulePowerEvent(PowerState pwr_state, Tick tick);
589
590    Tick prevArrival;
591
592    /**
593     * The soonest you have to start thinking about the next request
594     * is the longest access time that can occur before
595     * busBusyUntil. Assuming you need to precharge, open a new row,
596     * and access, it is tRP + tRCD + tCL.
597     */
598    Tick nextReqTime;
599
600    // All statistics that the model needs to capture
601    Stats::Scalar readReqs;
602    Stats::Scalar writeReqs;
603    Stats::Scalar readBursts;
604    Stats::Scalar writeBursts;
605    Stats::Scalar bytesReadDRAM;
606    Stats::Scalar bytesReadWrQ;
607    Stats::Scalar bytesWritten;
608    Stats::Scalar bytesReadSys;
609    Stats::Scalar bytesWrittenSys;
610    Stats::Scalar servicedByWrQ;
611    Stats::Scalar mergedWrBursts;
612    Stats::Scalar neitherReadNorWrite;
613    Stats::Vector perBankRdBursts;
614    Stats::Vector perBankWrBursts;
615    Stats::Scalar numRdRetry;
616    Stats::Scalar numWrRetry;
617    Stats::Scalar totGap;
618    Stats::Vector readPktSize;
619    Stats::Vector writePktSize;
620    Stats::Vector rdQLenPdf;
621    Stats::Vector wrQLenPdf;
622    Stats::Histogram bytesPerActivate;
623    Stats::Histogram rdPerTurnAround;
624    Stats::Histogram wrPerTurnAround;
625
626    // Latencies summed over all requests
627    Stats::Scalar totQLat;
628    Stats::Scalar totMemAccLat;
629    Stats::Scalar totBusLat;
630
631    // Average latencies per request
632    Stats::Formula avgQLat;
633    Stats::Formula avgBusLat;
634    Stats::Formula avgMemAccLat;
635
636    // Average bandwidth
637    Stats::Formula avgRdBW;
638    Stats::Formula avgWrBW;
639    Stats::Formula avgRdBWSys;
640    Stats::Formula avgWrBWSys;
641    Stats::Formula peakBW;
642    Stats::Formula busUtil;
643    Stats::Formula busUtilRead;
644    Stats::Formula busUtilWrite;
645
646    // Average queue lengths
647    Stats::Average avgRdQLen;
648    Stats::Average avgWrQLen;
649
650    // Row hit count and rate
651    Stats::Scalar readRowHits;
652    Stats::Scalar writeRowHits;
653    Stats::Formula readRowHitRate;
654    Stats::Formula writeRowHitRate;
655    Stats::Formula avgGap;
656
657    // DRAM Power Calculation
658    Stats::Formula pageHitRate;
659    Stats::Vector pwrStateTime;
660
661    // Track when we transitioned to the current power state
662    Tick pwrStateTick;
663
664    // To track number of banks which are currently active
665    unsigned int numBanksActive;
666
667    /** @todo this is a temporary workaround until the 4-phase code is
668     * committed. upstream caches needs this packet until true is returned, so
669     * hold onto it for deletion until a subsequent call
670     */
671    std::vector<PacketPtr> pendingDelete;
672
673  public:
674
675    void regStats();
676
677    DRAMCtrl(const DRAMCtrlParams* p);
678
679    unsigned int drain(DrainManager* dm);
680
681    virtual BaseSlavePort& getSlavePort(const std::string& if_name,
682                                        PortID idx = InvalidPortID);
683
684    virtual void init();
685    virtual void startup();
686
687  protected:
688
689    Tick recvAtomic(PacketPtr pkt);
690    void recvFunctional(PacketPtr pkt);
691    bool recvTimingReq(PacketPtr pkt);
692
693};
694
695#endif //__MEM_DRAM_CTRL_HH__
696