dram_ctrl.hh revision 9974 
1/*
2 * Copyright (c) 2012 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 * SimpleDRAM declaration
48 */
49
50#ifndef __MEM_SIMPLE_DRAM_HH__
51#define __MEM_SIMPLE_DRAM_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/SimpleDRAM.hh"
62#include "sim/eventq.hh"
63
64/**
65 * The simple DRAM is a basic single-channel memory controller aiming
66 * to mimic a high-level DRAM controller and the most important timing
67 * 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 SimpleDRAM : 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        SimpleDRAM& memory;
95
96      public:
97
98        MemoryPort(const std::string& name, SimpleDRAM& _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     * Remember that a row buffer hit occured
126     */
127    bool rowHitFlag;
128
129    /**
130     * Use this flag to shutoff reads, i.e. do not schedule any reads
131     * beyond those already done so that we can turn the bus around
132     * and do a few writes, or refresh, or whatever
133     */
134    bool stopReads;
135
136    /** List to keep track of activate ticks */
137    std::vector<std::deque<Tick>> actTicks;
138
139    /**
140     * A basic class to track the bank state indirectly via times
141     * "freeAt" and "tRASDoneAt" and what page is currently open. The
142     * bank also keeps track of how many bytes have been accessed in
143     * the open row since it was opened.
144     */
145    class Bank
146    {
147
148      public:
149
150        static const uint32_t INVALID_ROW = -1;
151
152        uint32_t openRow;
153
154        Tick freeAt;
155        Tick tRASDoneAt;
156        Tick actAllowedAt;
157
158        uint32_t bytesAccessed;
159
160        Bank() :
161            openRow(INVALID_ROW), freeAt(0), tRASDoneAt(0), actAllowedAt(0),
162            bytesAccessed(0)
163        { }
164    };
165
166    /**
167     * A burst helper helps organize and manage a packet that is larger than
168     * the DRAM burst size. A system packet that is larger than the burst size
169     * is split into multiple DRAM packets and all those DRAM packets point to
170     * a single burst helper such that we know when the whole packet is served.
171     */
172    class BurstHelper {
173
174      public:
175
176        /** Number of DRAM bursts requred for a system packet **/
177        const unsigned int burstCount;
178
179        /** Number of DRAM bursts serviced so far for a system packet **/
180        unsigned int burstsServiced;
181
182        BurstHelper(unsigned int _burstCount)
183            : burstCount(_burstCount), burstsServiced(0)
184            { }
185    };
186
187    /**
188     * A DRAM packet stores packets along with the timestamp of when
189     * the packet entered the queue, and also the decoded address.
190     */
191    class DRAMPacket {
192
193      public:
194
195        /** When did request enter the controller */
196        const Tick entryTime;
197
198        /** When will request leave the controller */
199        Tick readyTime;
200
201        /** This comes from the outside world */
202        const PacketPtr pkt;
203
204        const bool isRead;
205
206        /** Will be populated by address decoder */
207        const uint8_t rank;
208        const uint8_t bank;
209        const uint16_t row;
210
211        /**
212         * Bank id is calculated considering banks in all the ranks
213         * eg: 2 ranks each with 8 banks, then bankId = 0 --> rank0, bank0 and
214         * bankId = 8 --> rank1, bank0
215         */
216        const uint16_t bankId;
217
218        /**
219         * The starting address of the DRAM packet.
220         * This address could be unaligned to burst size boundaries. The
221         * reason is to keep the address offset so we can accurately check
222         * incoming read packets with packets in the write queue.
223         */
224        Addr addr;
225
226        /**
227         * The size of this dram packet in bytes
228         * It is always equal or smaller than DRAM burst size
229         */
230        unsigned int size;
231
232        /**
233         * A pointer to the BurstHelper if this DRAMPacket is a split packet
234         * If not a split packet (common case), this is set to NULL
235         */
236        BurstHelper* burstHelper;
237        Bank& bankRef;
238
239        DRAMPacket(PacketPtr _pkt, bool is_read, uint8_t _rank, uint8_t _bank,
240                   uint16_t _row, uint16_t bank_id, Addr _addr,
241                   unsigned int _size, Bank& bank_ref)
242            : entryTime(curTick()), readyTime(curTick()),
243              pkt(_pkt), isRead(is_read), rank(_rank), bank(_bank), row(_row),
244              bankId(bank_id), addr(_addr), size(_size), burstHelper(NULL),
245              bankRef(bank_ref)
246        { }
247
248    };
249
250    /**
251     * Bunch of things requires to setup "events" in gem5
252     * When event "writeEvent" occurs for example, the method
253     * processWriteEvent is called; no parameters are allowed
254     * in these methods
255     */
256    void processWriteEvent();
257    EventWrapper<SimpleDRAM, &SimpleDRAM::processWriteEvent> writeEvent;
258
259    void processRespondEvent();
260    EventWrapper<SimpleDRAM, &SimpleDRAM::processRespondEvent> respondEvent;
261
262    void processRefreshEvent();
263    EventWrapper<SimpleDRAM, &SimpleDRAM::processRefreshEvent> refreshEvent;
264
265    void processNextReqEvent();
266    EventWrapper<SimpleDRAM,&SimpleDRAM::processNextReqEvent> nextReqEvent;
267
268
269    /**
270     * Check if the read queue has room for more entries
271     *
272     * @param pktCount The number of entries needed in the read queue
273     * @return true if read queue is full, false otherwise
274     */
275    bool readQueueFull(unsigned int pktCount) const;
276
277    /**
278     * Check if the write queue has room for more entries
279     *
280     * @param pktCount The number of entries needed in the write queue
281     * @return true if write queue is full, false otherwise
282     */
283    bool writeQueueFull(unsigned int pktCount) const;
284
285    /**
286     * When a new read comes in, first check if the write q has a
287     * pending request to the same address.\ If not, decode the
288     * address to populate rank/bank/row, create one or mutliple
289     * "dram_pkt", and push them to the back of the read queue.\
290     * If this is the only
291     * read request in the system, schedule an event to start
292     * servicing it.
293     *
294     * @param pkt The request packet from the outside world
295     * @param pktCount The number of DRAM bursts the pkt
296     * translate to. If pkt size is larger then one full burst,
297     * then pktCount is greater than one.
298     */
299    void addToReadQueue(PacketPtr pkt, unsigned int pktCount);
300
301    /**
302     * Decode the incoming pkt, create a dram_pkt and push to the
303     * back of the write queue. \If the write q length is more than
304     * the threshold specified by the user, ie the queue is beginning
305     * to get full, stop reads, and start draining writes.
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 addToWriteQueue(PacketPtr pkt, unsigned int pktCount);
313
314    /**
315     * Actually do the DRAM access - figure out the latency it
316     * will take to service the req based on bank state, channel state etc
317     * and then update those states to account for this request.\ Based
318     * on this, update the packet's "readyTime" and move it to the
319     * response q from where it will eventually go back to the outside
320     * world.
321     *
322     * @param pkt The DRAM packet created from the outside world pkt
323     */
324    void doDRAMAccess(DRAMPacket* dram_pkt);
325
326    /**
327     * Check when the channel is free to turnaround, add turnaround
328     * delay and schedule a whole bunch of writes.
329     */
330    void triggerWrites();
331
332    /**
333     * When a packet reaches its "readyTime" in the response Q,
334     * use the "access()" method in AbstractMemory to actually
335     * create the response packet, and send it back to the outside
336     * world requestor.
337     *
338     * @param pkt The packet from the outside world
339     * @param static_latency Static latency to add before sending the packet
340     */
341    void accessAndRespond(PacketPtr pkt, Tick static_latency);
342
343    /**
344     * Address decoder to figure out physical mapping onto ranks,
345     * banks, and rows. This function is called multiple times on the same
346     * system packet if the pakcet is larger than burst of the memory. The
347     * dramPktAddr is used for the offset within the packet.
348     *
349     * @param pkt The packet from the outside world
350     * @param dramPktAddr The starting address of the DRAM packet
351     * @param size The size of the DRAM packet in bytes
352     * @param isRead Is the request for a read or a write to DRAM
353     * @return A DRAMPacket pointer with the decoded information
354     */
355    DRAMPacket* decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned int size, bool isRead);
356
357    /**
358     * The memory schduler/arbiter - picks which read request needs to
359     * go next, based on the specified policy such as FCFS or FR-FCFS
360     * and moves it to the head of the read queue.
361     *
362     * @return True if a request was chosen and false if queue is empty
363     */
364    bool chooseNextRead();
365
366    /**
367     * Calls chooseNextReq() to pick the right request, then calls
368     * doDRAMAccess on that request in order to actually service
369     * that request
370     */
371    void scheduleNextReq();
372
373    /**
374     *Looks at the state of the banks, channels, row buffer hits etc
375     * to estimate how long a request will take to complete.
376     *
377     * @param dram_pkt The request for which we want to estimate latency
378     * @param inTime The tick at which you want to probe the memory
379     *
380     * @return A pair of ticks, one indicating how many ticks *after*
381     *         inTime the request require, and the other indicating how
382     *         much of that was just the bank access time, ignoring the
383     *         ticks spent simply waiting for resources to become free
384     */
385    std::pair<Tick, Tick> estimateLatency(DRAMPacket* dram_pkt, Tick inTime);
386
387    /**
388     * Move the request at the head of the read queue to the response
389     * queue, sorting by readyTime.\ If it is the only packet in the
390     * response queue, schedule a respond event to send it back to the
391     * outside world
392     */
393    void moveToRespQ();
394
395    /**
396     * Scheduling policy within the write queue
397     */
398    void chooseNextWrite();
399
400    /**
401     * For FR-FCFS policy reorder the read/write queue depending on row buffer
402     * hits and earliest banks available in DRAM
403     */
404    void reorderQueue(std::deque<DRAMPacket*>& queue);
405
406    /**
407     * Looking at all banks, determine the moment in time when they
408     * are all free.
409     *
410     * @return The tick when all banks are free
411     */
412    Tick maxBankFreeAt() const;
413
414    /**
415     * Find which are the earliest available banks for the enqueued
416     * requests. Assumes maximum of 64 banks per DIMM
417     *
418     * @param Queued requests to consider
419     * @return One-hot encoded mask of bank indices
420     */
421    uint64_t minBankFreeAt(const std::deque<DRAMPacket*>& queue) const;
422
423    /**
424     * Keep track of when row activations happen, in order to enforce
425     * the maximum number of activations in the activation window. The
426     * method updates the time that the banks become available based
427     * on the current limits.
428     */
429    void recordActivate(Tick act_tick, uint8_t rank, uint8_t bank);
430
431    void printParams() const;
432    void printQs() const;
433
434    /**
435     * The controller's main read and write queues
436     */
437    std::deque<DRAMPacket*> readQueue;
438    std::deque<DRAMPacket*> writeQueue;
439
440    /**
441     * Response queue where read packets wait after we're done working
442     * with them, but it's not time to send the response yet. The
443     * responses are stored seperately mostly to keep the code clean
444     * and help with events scheduling. For all logical purposes such
445     * as sizing the read queue, this and the main read queue need to
446     * be added together.
447     */
448    std::deque<DRAMPacket*> respQueue;
449
450    /**
451     * If we need to drain, keep the drain manager around until we're
452     * done here.
453     */
454    DrainManager *drainManager;
455
456    /**
457     * Multi-dimensional vector of banks, first dimension is ranks,
458     * second is bank
459     */
460    std::vector<std::vector<Bank> > banks;
461
462    /**
463     * The following are basic design parameters of the memory
464     * controller, and are initialized based on parameter values.
465     * The rowsPerBank is determined based on the capacity, number of
466     * ranks and banks, the burst size, and the row buffer size.
467     */
468    const uint32_t deviceBusWidth;
469    const uint32_t burstLength;
470    const uint32_t deviceRowBufferSize;
471    const uint32_t devicesPerRank;
472    const uint32_t burstSize;
473    const uint32_t rowBufferSize;
474    const uint32_t ranksPerChannel;
475    const uint32_t banksPerRank;
476    const uint32_t channels;
477    uint32_t rowsPerBank;
478    uint32_t columnsPerRowBuffer;
479    const uint32_t readBufferSize;
480    const uint32_t writeBufferSize;
481    const double writeHighThresholdPerc;
482    uint32_t writeHighThreshold;
483    const double writeLowThresholdPerc;
484    uint32_t writeLowThreshold;
485
486    /**
487     * Basic memory timing parameters initialized based on parameter
488     * values.
489     */
490    const Tick tWTR;
491    const Tick tBURST;
492    const Tick tRCD;
493    const Tick tCL;
494    const Tick tRP;
495    const Tick tRAS;
496    const Tick tRFC;
497    const Tick tREFI;
498    const Tick tRRD;
499    const Tick tXAW;
500    const uint32_t activationLimit;
501
502    /**
503     * Memory controller configuration initialized based on parameter
504     * values.
505     */
506    Enums::MemSched memSchedPolicy;
507    Enums::AddrMap addrMapping;
508    Enums::PageManage pageMgmt;
509
510    /**
511     * Pipeline latency of the controller frontend. The frontend
512     * contribution is added to writes (that complete when they are in
513     * the write buffer) and reads that are serviced the write buffer.
514     */
515    const Tick frontendLatency;
516
517    /**
518     * Pipeline latency of the backend and PHY. Along with the
519     * frontend contribution, this latency is added to reads serviced
520     * by the DRAM.
521     */
522    const Tick backendLatency;
523
524    /**
525     * Till when has the main data bus been spoken for already?
526     */
527    Tick busBusyUntil;
528
529    Tick writeStartTime;
530    Tick prevArrival;
531    int numReqs;
532
533    // Tracks number of writes done to meet the write threshold
534    uint32_t numWritesThisTime;
535
536    // The absolute soonest you have to start thinking about the
537    // next request is the longest access time that can occur before
538    // busBusyUntil. Assuming you need to precharge,
539    // open a new row, and access, it is tRP + tRCD + tCL
540    Tick newTime;
541
542    // All statistics that the model needs to capture
543    Stats::Scalar readReqs;
544    Stats::Scalar writeReqs;
545    Stats::Scalar readBursts;
546    Stats::Scalar writeBursts;
547    Stats::Scalar bytesRead;
548    Stats::Scalar bytesWritten;
549    Stats::Scalar bytesConsumedRd;
550    Stats::Scalar bytesConsumedWr;
551    Stats::Scalar servicedByWrQ;
552    Stats::Scalar neitherReadNorWrite;
553    Stats::Vector perBankRdReqs;
554    Stats::Vector perBankWrReqs;
555    Stats::Scalar numRdRetry;
556    Stats::Scalar numWrRetry;
557    Stats::Scalar totGap;
558    Stats::Vector readPktSize;
559    Stats::Vector writePktSize;
560    Stats::Vector rdQLenPdf;
561    Stats::Vector wrQLenPdf;
562    Stats::Histogram bytesPerActivate;
563
564    // Latencies summed over all requests
565    Stats::Scalar totQLat;
566    Stats::Scalar totMemAccLat;
567    Stats::Scalar totBusLat;
568    Stats::Scalar totBankLat;
569
570    // Average latencies per request
571    Stats::Formula avgQLat;
572    Stats::Formula avgBankLat;
573    Stats::Formula avgBusLat;
574    Stats::Formula avgMemAccLat;
575
576    // Average bandwidth
577    Stats::Formula avgRdBW;
578    Stats::Formula avgWrBW;
579    Stats::Formula avgConsumedRdBW;
580    Stats::Formula avgConsumedWrBW;
581    Stats::Formula peakBW;
582    Stats::Formula busUtil;
583
584    // Average queue lengths
585    Stats::Average avgRdQLen;
586    Stats::Average avgWrQLen;
587
588    // Row hit count and rate
589    Stats::Scalar readRowHits;
590    Stats::Scalar writeRowHits;
591    Stats::Formula readRowHitRate;
592    Stats::Formula writeRowHitRate;
593    Stats::Formula avgGap;
594
595    /** @todo this is a temporary workaround until the 4-phase code is
596     * committed. upstream caches needs this packet until true is returned, so
597     * hold onto it for deletion until a subsequent call
598     */
599    std::vector<PacketPtr> pendingDelete;
600
601  public:
602
603    void regStats();
604
605    SimpleDRAM(const SimpleDRAMParams* p);
606
607    unsigned int drain(DrainManager* dm);
608
609    virtual BaseSlavePort& getSlavePort(const std::string& if_name,
610                                        PortID idx = InvalidPortID);
611
612    virtual void init();
613    virtual void startup();
614
615  protected:
616
617    Tick recvAtomic(PacketPtr pkt);
618    void recvFunctional(PacketPtr pkt);
619    bool recvTimingReq(PacketPtr pkt);
620
621};
622
623#endif //__MEM_SIMPLE_DRAM_HH__
624