dram_ctrl.hh revision 10247
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 basic single-channel memory controller
66 * aiming to mimic a high-level DRAM controller and the most important
67 * timing constraints associated with the DRAM. The focus is really on
68 * modelling the impact on the system rather than the DRAM itself,
69 * hence the focus is on the controller model and not on the
70 * memory. By adhering to the correct timing constraints, ultimately
71 * there is no need for a memory model in addition to the controller
72 * model.
73 *
74 * As a basic design principle, this controller is not cycle callable,
75 * but instead uses events to decide when new decisions can be made,
76 * when resources become available, when things are to be considered
77 * done, and when to send things back. Through these simple
78 * principles, we achieve a performant model that is not
79 * cycle-accurate, but enables us to evaluate the system impact of a
80 * wide range of memory technologies, and also collect statistics
81 * about the use of the memory.
82 */
83class DRAMCtrl : public AbstractMemory
84{
85
86  private:
87
88    // For now, make use of a queued slave port to avoid dealing with
89    // flow control for the responses being sent back
90    class MemoryPort : public QueuedSlavePort
91    {
92
93        SlavePacketQueue queue;
94        DRAMCtrl& memory;
95
96      public:
97
98        MemoryPort(const std::string& name, DRAMCtrl& _memory);
99
100      protected:
101
102        Tick recvAtomic(PacketPtr pkt);
103
104        void recvFunctional(PacketPtr pkt);
105
106        bool recvTimingReq(PacketPtr);
107
108        virtual AddrRangeList getAddrRanges() const;
109
110    };
111
112    /**
113     * Our incoming port, for a multi-ported controller add a crossbar
114     * in front of it
115     */
116    MemoryPort port;
117
118    /**
119     * Remember if we have to retry a request when available.
120     */
121    bool retryRdReq;
122    bool retryWrReq;
123
124    /**
125     * Bus state used to control the read/write switching and drive
126     * the scheduling of the next request.
127     */
128    enum BusState {
129        READ = 0,
130        READ_TO_WRITE,
131        WRITE,
132        WRITE_TO_READ
133    };
134
135    BusState busState;
136
137    /** List to keep track of activate ticks */
138    std::vector<std::deque<Tick>> actTicks;
139
140    /**
141     * A basic class to track the bank state, i.e. what row is
142     * currently open (if any), when is the bank free to accept a new
143     * column (read/write) command, when can it be precharged, and
144     * when can it be activated.
145     *
146     * The bank also keeps track of how many bytes have been accessed
147     * in the open row since it was opened.
148     */
149    class Bank
150    {
151
152      public:
153
154        static const uint32_t NO_ROW = -1;
155
156        uint32_t openRow;
157        uint8_t rank;
158        uint8_t bank;
159
160        Tick colAllowedAt;
161        Tick preAllowedAt;
162        Tick actAllowedAt;
163
164        uint32_t rowAccesses;
165        uint32_t bytesAccessed;
166
167        Bank() :
168            openRow(NO_ROW), rank(0), bank(0),
169            colAllowedAt(0), preAllowedAt(0), actAllowedAt(0),
170            rowAccesses(0), bytesAccessed(0)
171        { }
172    };
173
174    /**
175     * A burst helper helps organize and manage a packet that is larger than
176     * the DRAM burst size. A system packet that is larger than the burst size
177     * is split into multiple DRAM packets and all those DRAM packets point to
178     * a single burst helper such that we know when the whole packet is served.
179     */
180    class BurstHelper {
181
182      public:
183
184        /** Number of DRAM bursts requred for a system packet **/
185        const unsigned int burstCount;
186
187        /** Number of DRAM bursts serviced so far for a system packet **/
188        unsigned int burstsServiced;
189
190        BurstHelper(unsigned int _burstCount)
191            : burstCount(_burstCount), burstsServiced(0)
192            { }
193    };
194
195    /**
196     * A DRAM packet stores packets along with the timestamp of when
197     * the packet entered the queue, and also the decoded address.
198     */
199    class DRAMPacket {
200
201      public:
202
203        /** When did request enter the controller */
204        const Tick entryTime;
205
206        /** When will request leave the controller */
207        Tick readyTime;
208
209        /** This comes from the outside world */
210        const PacketPtr pkt;
211
212        const bool isRead;
213
214        /** Will be populated by address decoder */
215        const uint8_t rank;
216        const uint8_t bank;
217        const uint32_t row;
218
219        /**
220         * Bank id is calculated considering banks in all the ranks
221         * eg: 2 ranks each with 8 banks, then bankId = 0 --> rank0, bank0 and
222         * bankId = 8 --> rank1, bank0
223         */
224        const uint16_t bankId;
225
226        /**
227         * The starting address of the DRAM packet.
228         * This address could be unaligned to burst size boundaries. The
229         * reason is to keep the address offset so we can accurately check
230         * incoming read packets with packets in the write queue.
231         */
232        Addr addr;
233
234        /**
235         * The size of this dram packet in bytes
236         * It is always equal or smaller than DRAM burst size
237         */
238        unsigned int size;
239
240        /**
241         * A pointer to the BurstHelper if this DRAMPacket is a split packet
242         * If not a split packet (common case), this is set to NULL
243         */
244        BurstHelper* burstHelper;
245        Bank& bankRef;
246
247        DRAMPacket(PacketPtr _pkt, bool is_read, uint8_t _rank, uint8_t _bank,
248                   uint32_t _row, uint16_t bank_id, Addr _addr,
249                   unsigned int _size, Bank& bank_ref)
250            : entryTime(curTick()), readyTime(curTick()),
251              pkt(_pkt), isRead(is_read), rank(_rank), bank(_bank), row(_row),
252              bankId(bank_id), addr(_addr), size(_size), burstHelper(NULL),
253              bankRef(bank_ref)
254        { }
255
256    };
257
258    /**
259     * Bunch of things requires to setup "events" in gem5
260     * When event "respondEvent" occurs for example, the method
261     * processRespondEvent is called; no parameters are allowed
262     * in these methods
263     */
264    void processNextReqEvent();
265    EventWrapper<DRAMCtrl,&DRAMCtrl::processNextReqEvent> nextReqEvent;
266
267    void processRespondEvent();
268    EventWrapper<DRAMCtrl, &DRAMCtrl::processRespondEvent> respondEvent;
269
270    void processActivateEvent();
271    EventWrapper<DRAMCtrl, &DRAMCtrl::processActivateEvent> activateEvent;
272
273    void processPrechargeEvent();
274    EventWrapper<DRAMCtrl, &DRAMCtrl::processPrechargeEvent> prechargeEvent;
275
276    void processRefreshEvent();
277    EventWrapper<DRAMCtrl, &DRAMCtrl::processRefreshEvent> refreshEvent;
278
279    void processPowerEvent();
280    EventWrapper<DRAMCtrl,&DRAMCtrl::processPowerEvent> powerEvent;
281
282    /**
283     * Check if the read queue has room for more entries
284     *
285     * @param pktCount The number of entries needed in the read queue
286     * @return true if read queue is full, false otherwise
287     */
288    bool readQueueFull(unsigned int pktCount) const;
289
290    /**
291     * Check if the write queue has room for more entries
292     *
293     * @param pktCount The number of entries needed in the write queue
294     * @return true if write queue is full, false otherwise
295     */
296    bool writeQueueFull(unsigned int pktCount) const;
297
298    /**
299     * When a new read comes in, first check if the write q has a
300     * pending request to the same address.\ If not, decode the
301     * address to populate rank/bank/row, create one or mutliple
302     * "dram_pkt", and push them to the back of the read queue.\
303     * If this is the only
304     * read request in the system, schedule an event to start
305     * servicing it.
306     *
307     * @param pkt The request packet from the outside world
308     * @param pktCount The number of DRAM bursts the pkt
309     * translate to. If pkt size is larger then one full burst,
310     * then pktCount is greater than one.
311     */
312    void addToReadQueue(PacketPtr pkt, unsigned int pktCount);
313
314    /**
315     * Decode the incoming pkt, create a dram_pkt and push to the
316     * back of the write queue. \If the write q length is more than
317     * the threshold specified by the user, ie the queue is beginning
318     * to get full, stop reads, and start draining writes.
319     *
320     * @param pkt The request packet from the outside world
321     * @param pktCount The number of DRAM bursts the pkt
322     * translate to. If pkt size is larger then one full burst,
323     * then pktCount is greater than one.
324     */
325    void addToWriteQueue(PacketPtr pkt, unsigned int pktCount);
326
327    /**
328     * Actually do the DRAM access - figure out the latency it
329     * will take to service the req based on bank state, channel state etc
330     * and then update those states to account for this request.\ Based
331     * on this, update the packet's "readyTime" and move it to the
332     * response q from where it will eventually go back to the outside
333     * world.
334     *
335     * @param pkt The DRAM packet created from the outside world pkt
336     */
337    void doDRAMAccess(DRAMPacket* dram_pkt);
338
339    /**
340     * When a packet reaches its "readyTime" in the response Q,
341     * use the "access()" method in AbstractMemory to actually
342     * create the response packet, and send it back to the outside
343     * world requestor.
344     *
345     * @param pkt The packet from the outside world
346     * @param static_latency Static latency to add before sending the packet
347     */
348    void accessAndRespond(PacketPtr pkt, Tick static_latency);
349
350    /**
351     * Address decoder to figure out physical mapping onto ranks,
352     * banks, and rows. This function is called multiple times on the same
353     * system packet if the pakcet is larger than burst of the memory. The
354     * dramPktAddr is used for the offset within the packet.
355     *
356     * @param pkt The packet from the outside world
357     * @param dramPktAddr The starting address of the DRAM packet
358     * @param size The size of the DRAM packet in bytes
359     * @param isRead Is the request for a read or a write to DRAM
360     * @return A DRAMPacket pointer with the decoded information
361     */
362    DRAMPacket* decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned int size,
363                           bool isRead);
364
365    /**
366     * The memory schduler/arbiter - picks which request needs to
367     * go next, based on the specified policy such as FCFS or FR-FCFS
368     * and moves it to the head of the queue.
369     */
370    void chooseNext(std::deque<DRAMPacket*>& queue);
371
372    /**
373     * For FR-FCFS policy reorder the read/write queue depending on row buffer
374     * hits and earliest banks available in DRAM
375     */
376    void reorderQueue(std::deque<DRAMPacket*>& queue);
377
378    /**
379     * Find which are the earliest banks ready to issue an activate
380     * for the enqueued requests. Assumes maximum of 64 banks per DIMM
381     *
382     * @param Queued requests to consider
383     * @return One-hot encoded mask of bank indices
384     */
385    uint64_t minBankActAt(const std::deque<DRAMPacket*>& queue) const;
386
387    /**
388     * Keep track of when row activations happen, in order to enforce
389     * the maximum number of activations in the activation window. The
390     * method updates the time that the banks become available based
391     * on the current limits.
392     *
393     * @param bank Reference to the bank
394     * @param act_tick Time when the activation takes place
395     * @param row Index of the row
396     */
397    void activateBank(Bank& bank, Tick act_tick, uint32_t row);
398
399    /**
400     * Precharge a given bank and also update when the precharge is
401     * done. This will also deal with any stats related to the
402     * accesses to the open page.
403     *
404     * @param bank_ref The bank to precharge
405     * @param pre_at Time when the precharge takes place
406     * @param trace Is this an auto precharge then do not add to trace
407     */
408    void prechargeBank(Bank& bank_ref, Tick pre_at,  bool trace = true);
409
410    /**
411     * Used for debugging to observe the contents of the queues.
412     */
413    void printQs() const;
414
415    /**
416     * The controller's main read and write queues
417     */
418    std::deque<DRAMPacket*> readQueue;
419    std::deque<DRAMPacket*> writeQueue;
420
421    /**
422     * Response queue where read packets wait after we're done working
423     * with them, but it's not time to send the response yet. The
424     * responses are stored seperately mostly to keep the code clean
425     * and help with events scheduling. For all logical purposes such
426     * as sizing the read queue, this and the main read queue need to
427     * be added together.
428     */
429    std::deque<DRAMPacket*> respQueue;
430
431    /**
432     * If we need to drain, keep the drain manager around until we're
433     * done here.
434     */
435    DrainManager *drainManager;
436
437    /**
438     * Multi-dimensional vector of banks, first dimension is ranks,
439     * second is bank
440     */
441    std::vector<std::vector<Bank> > banks;
442
443    /**
444     * The following are basic design parameters of the memory
445     * controller, and are initialized based on parameter values.
446     * The rowsPerBank is determined based on the capacity, number of
447     * ranks and banks, the burst size, and the row buffer size.
448     */
449    const uint32_t deviceBusWidth;
450    const uint32_t burstLength;
451    const uint32_t deviceRowBufferSize;
452    const uint32_t devicesPerRank;
453    const uint32_t burstSize;
454    const uint32_t rowBufferSize;
455    const uint32_t columnsPerRowBuffer;
456    const uint32_t ranksPerChannel;
457    const uint32_t banksPerRank;
458    const uint32_t channels;
459    uint32_t rowsPerBank;
460    const uint32_t readBufferSize;
461    const uint32_t writeBufferSize;
462    const uint32_t writeHighThreshold;
463    const uint32_t writeLowThreshold;
464    const uint32_t minWritesPerSwitch;
465    uint32_t writesThisTime;
466    uint32_t readsThisTime;
467
468    /**
469     * Basic memory timing parameters initialized based on parameter
470     * values.
471     */
472    const Tick tCK;
473    const Tick tWTR;
474    const Tick tRTW;
475    const Tick tBURST;
476    const Tick tRCD;
477    const Tick tCL;
478    const Tick tRP;
479    const Tick tRAS;
480    const Tick tWR;
481    const Tick tRTP;
482    const Tick tRFC;
483    const Tick tREFI;
484    const Tick tRRD;
485    const Tick tXAW;
486    const uint32_t activationLimit;
487
488    /**
489     * Memory controller configuration initialized based on parameter
490     * values.
491     */
492    Enums::MemSched memSchedPolicy;
493    Enums::AddrMap addrMapping;
494    Enums::PageManage pageMgmt;
495
496    /**
497     * Max column accesses (read and write) per row, before forefully
498     * closing it.
499     */
500    const uint32_t maxAccessesPerRow;
501
502    /**
503     * Pipeline latency of the controller frontend. The frontend
504     * contribution is added to writes (that complete when they are in
505     * the write buffer) and reads that are serviced the write buffer.
506     */
507    const Tick frontendLatency;
508
509    /**
510     * Pipeline latency of the backend and PHY. Along with the
511     * frontend contribution, this latency is added to reads serviced
512     * by the DRAM.
513     */
514    const Tick backendLatency;
515
516    /**
517     * Till when has the main data bus been spoken for already?
518     */
519    Tick busBusyUntil;
520
521    /**
522     * Keep track of when a refresh is due.
523     */
524    Tick refreshDueAt;
525
526    /**
527     * The refresh state is used to control the progress of the
528     * refresh scheduling. When normal operation is in progress the
529     * refresh state is idle. From there, it progresses to the refresh
530     * drain state once tREFI has passed. The refresh drain state
531     * captures the DRAM row active state, as it will stay there until
532     * all ongoing accesses complete. Thereafter all banks are
533     * precharged, and lastly, the DRAM is refreshed.
534     */
535    enum RefreshState {
536        REF_IDLE = 0,
537        REF_DRAIN,
538        REF_PRE,
539        REF_RUN
540    };
541
542    RefreshState refreshState;
543
544    /**
545     * The power state captures the different operational states of
546     * the DRAM and interacts with the bus read/write state machine,
547     * and the refresh state machine. In the idle state all banks are
548     * precharged. From there we either go to an auto refresh (as
549     * determined by the refresh state machine), or to a precharge
550     * power down mode. From idle the memory can also go to the active
551     * state (with one or more banks active), and in turn from there
552     * to active power down. At the moment we do not capture the deep
553     * power down and self-refresh state.
554     */
555    enum PowerState {
556        PWR_IDLE = 0,
557        PWR_REF,
558        PWR_PRE_PDN,
559        PWR_ACT,
560        PWR_ACT_PDN
561    };
562
563    /**
564     * Since we are taking decisions out of order, we need to keep
565     * track of what power transition is happening at what time, such
566     * that we can go back in time and change history. For example, if
567     * we precharge all banks and schedule going to the idle state, we
568     * might at a later point decide to activate a bank before the
569     * transition to idle would have taken place.
570     */
571    PowerState pwrStateTrans;
572
573    /**
574     * Current power state.
575     */
576    PowerState pwrState;
577
578    /**
579     * Schedule a power state transition in the future, and
580     * potentially override an already scheduled transition.
581     *
582     * @param pwr_state Power state to transition to
583     * @param tick Tick when transition should take place
584     */
585    void schedulePowerEvent(PowerState pwr_state, Tick tick);
586
587    Tick prevArrival;
588
589    /**
590     * The soonest you have to start thinking about the next request
591     * is the longest access time that can occur before
592     * busBusyUntil. Assuming you need to precharge, open a new row,
593     * and access, it is tRP + tRCD + tCL.
594     */
595    Tick nextReqTime;
596
597    // All statistics that the model needs to capture
598    Stats::Scalar readReqs;
599    Stats::Scalar writeReqs;
600    Stats::Scalar readBursts;
601    Stats::Scalar writeBursts;
602    Stats::Scalar bytesReadDRAM;
603    Stats::Scalar bytesReadWrQ;
604    Stats::Scalar bytesWritten;
605    Stats::Scalar bytesReadSys;
606    Stats::Scalar bytesWrittenSys;
607    Stats::Scalar servicedByWrQ;
608    Stats::Scalar mergedWrBursts;
609    Stats::Scalar neitherReadNorWrite;
610    Stats::Vector perBankRdBursts;
611    Stats::Vector perBankWrBursts;
612    Stats::Scalar numRdRetry;
613    Stats::Scalar numWrRetry;
614    Stats::Scalar totGap;
615    Stats::Vector readPktSize;
616    Stats::Vector writePktSize;
617    Stats::Vector rdQLenPdf;
618    Stats::Vector wrQLenPdf;
619    Stats::Histogram bytesPerActivate;
620    Stats::Histogram rdPerTurnAround;
621    Stats::Histogram wrPerTurnAround;
622
623    // Latencies summed over all requests
624    Stats::Scalar totQLat;
625    Stats::Scalar totMemAccLat;
626    Stats::Scalar totBusLat;
627
628    // Average latencies per request
629    Stats::Formula avgQLat;
630    Stats::Formula avgBusLat;
631    Stats::Formula avgMemAccLat;
632
633    // Average bandwidth
634    Stats::Formula avgRdBW;
635    Stats::Formula avgWrBW;
636    Stats::Formula avgRdBWSys;
637    Stats::Formula avgWrBWSys;
638    Stats::Formula peakBW;
639    Stats::Formula busUtil;
640    Stats::Formula busUtilRead;
641    Stats::Formula busUtilWrite;
642
643    // Average queue lengths
644    Stats::Average avgRdQLen;
645    Stats::Average avgWrQLen;
646
647    // Row hit count and rate
648    Stats::Scalar readRowHits;
649    Stats::Scalar writeRowHits;
650    Stats::Formula readRowHitRate;
651    Stats::Formula writeRowHitRate;
652    Stats::Formula avgGap;
653
654    // DRAM Power Calculation
655    Stats::Formula pageHitRate;
656    Stats::Vector pwrStateTime;
657
658    // Track when we transitioned to the current power state
659    Tick pwrStateTick;
660
661    // To track number of banks which are currently active
662    unsigned int numBanksActive;
663
664    /** @todo this is a temporary workaround until the 4-phase code is
665     * committed. upstream caches needs this packet until true is returned, so
666     * hold onto it for deletion until a subsequent call
667     */
668    std::vector<PacketPtr> pendingDelete;
669
670  public:
671
672    void regStats();
673
674    DRAMCtrl(const DRAMCtrlParams* p);
675
676    unsigned int drain(DrainManager* dm);
677
678    virtual BaseSlavePort& getSlavePort(const std::string& if_name,
679                                        PortID idx = InvalidPortID);
680
681    virtual void init();
682    virtual void startup();
683
684  protected:
685
686    Tick recvAtomic(PacketPtr pkt);
687    void recvFunctional(PacketPtr pkt);
688    bool recvTimingReq(PacketPtr pkt);
689
690};
691
692#endif //__MEM_DRAM_CTRL_HH__
693