1/*
2 * Copyright (c) 2012-2013, 2015-2016, 2018 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) 2003-2005 The Regents of The University of Michigan
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: Erik Hallnor
41 * Steve Reinhardt
42 * Ron Dreslinski
43 * Andreas Hansson
44 * Nikos Nikoleris
45 */
46
47/**
48 * @file
49 * Declares a basic cache interface BaseCache.
50 */
51
52#ifndef __MEM_CACHE_BASE_HH__
53#define __MEM_CACHE_BASE_HH__
54
55#include <cassert>
56#include <cstdint>
57#include <string>
58
59#include "base/addr_range.hh"
60#include "base/statistics.hh"
61#include "base/trace.hh"
62#include "base/types.hh"
63#include "debug/Cache.hh"
64#include "debug/CachePort.hh"
65#include "enums/Clusivity.hh"
66#include "mem/cache/cache_blk.hh"
67#include "mem/cache/mshr_queue.hh"
68#include "mem/cache/tags/base.hh"
69#include "mem/cache/write_queue.hh"
70#include "mem/cache/write_queue_entry.hh"
71#include "mem/mem_object.hh"
72#include "mem/packet.hh"
73#include "mem/packet_queue.hh"
74#include "mem/qport.hh"
75#include "mem/request.hh"
76#include "params/WriteAllocator.hh"
77#include "sim/eventq.hh"
78#include "sim/probe/probe.hh"
79#include "sim/serialize.hh"
80#include "sim/sim_exit.hh"
81#include "sim/system.hh"
82
83class BaseMasterPort;
84class BasePrefetcher;
85class BaseSlavePort;
86class MSHR;
87class MasterPort;
88class QueueEntry;
89struct BaseCacheParams;
90
91/**
92 * A basic cache interface. Implements some common functions for speed.
93 */
94class BaseCache : public MemObject
95{
96 protected:
97 /**
98 * Indexes to enumerate the MSHR queues.
99 */
100 enum MSHRQueueIndex {
101 MSHRQueue_MSHRs,
102 MSHRQueue_WriteBuffer
103 };
104
105 public:
106 /**
107 * Reasons for caches to be blocked.
108 */
109 enum BlockedCause {
110 Blocked_NoMSHRs = MSHRQueue_MSHRs,
111 Blocked_NoWBBuffers = MSHRQueue_WriteBuffer,
112 Blocked_NoTargets,
113 NUM_BLOCKED_CAUSES
114 };
115
116 protected:
117
118 /**
119 * A cache master port is used for the memory-side port of the
120 * cache, and in addition to the basic timing port that only sends
121 * response packets through a transmit list, it also offers the
122 * ability to schedule and send request packets (requests &
123 * writebacks). The send event is scheduled through schedSendEvent,
124 * and the sendDeferredPacket of the timing port is modified to
125 * consider both the transmit list and the requests from the MSHR.
126 */
127 class CacheMasterPort : public QueuedMasterPort
128 {
129
130 public:
131
132 /**
133 * Schedule a send of a request packet (from the MSHR). Note
134 * that we could already have a retry outstanding.
135 */
136 void schedSendEvent(Tick time)
137 {
138 DPRINTF(CachePort, "Scheduling send event at %llu\n", time);
139 reqQueue.schedSendEvent(time);
140 }
141
142 protected:
143
144 CacheMasterPort(const std::string &_name, BaseCache *_cache,
145 ReqPacketQueue &_reqQueue,
146 SnoopRespPacketQueue &_snoopRespQueue) :
147 QueuedMasterPort(_name, _cache, _reqQueue, _snoopRespQueue)
148 { }
149
150 /**
151 * Memory-side port always snoops.
152 *
153 * @return always true
154 */
155 virtual bool isSnooping() const { return true; }
156 };
157
158 /**
159 * Override the default behaviour of sendDeferredPacket to enable
160 * the memory-side cache port to also send requests based on the
161 * current MSHR status. This queue has a pointer to our specific
162 * cache implementation and is used by the MemSidePort.
163 */
164 class CacheReqPacketQueue : public ReqPacketQueue
165 {
166
167 protected:
168
169 BaseCache &cache;
170 SnoopRespPacketQueue &snoopRespQueue;
171
172 public:
173
174 CacheReqPacketQueue(BaseCache &cache, MasterPort &port,
175 SnoopRespPacketQueue &snoop_resp_queue,
176 const std::string &label) :
177 ReqPacketQueue(cache, port, label), cache(cache),
178 snoopRespQueue(snoop_resp_queue) { }
179
180 /**
181 * Override the normal sendDeferredPacket and do not only
182 * consider the transmit list (used for responses), but also
183 * requests.
184 */
185 virtual void sendDeferredPacket();
186
187 /**
188 * Check if there is a conflicting snoop response about to be
189 * send out, and if so simply stall any requests, and schedule
190 * a send event at the same time as the next snoop response is
191 * being sent out.
192 */
193 bool checkConflictingSnoop(Addr addr)
194 {
195 if (snoopRespQueue.hasAddr(addr)) {
196 DPRINTF(CachePort, "Waiting for snoop response to be "
197 "sent\n");
198 Tick when = snoopRespQueue.deferredPacketReadyTime();
199 schedSendEvent(when);
200 return true;
201 }
202 return false;
203 }
204 };
205
206
207 /**
208 * The memory-side port extends the base cache master port with
209 * access functions for functional, atomic and timing snoops.
210 */
211 class MemSidePort : public CacheMasterPort
212 {
213 private:
214
215 /** The cache-specific queue. */
216 CacheReqPacketQueue _reqQueue;
217
218 SnoopRespPacketQueue _snoopRespQueue;
219
220 // a pointer to our specific cache implementation
221 BaseCache *cache;
222
223 protected:
224
225 virtual void recvTimingSnoopReq(PacketPtr pkt);
226
227 virtual bool recvTimingResp(PacketPtr pkt);
228
229 virtual Tick recvAtomicSnoop(PacketPtr pkt);
230
231 virtual void recvFunctionalSnoop(PacketPtr pkt);
232
233 public:
234
235 MemSidePort(const std::string &_name, BaseCache *_cache,
236 const std::string &_label);
237 };
238
239 /**
240 * A cache slave port is used for the CPU-side port of the cache,
241 * and it is basically a simple timing port that uses a transmit
242 * list for responses to the CPU (or connected master). In
243 * addition, it has the functionality to block the port for
244 * incoming requests. If blocked, the port will issue a retry once
245 * unblocked.
246 */
247 class CacheSlavePort : public QueuedSlavePort
248 {
249
250 public:
251
252 /** Do not accept any new requests. */
253 void setBlocked();
254
255 /** Return to normal operation and accept new requests. */
256 void clearBlocked();
257
258 bool isBlocked() const { return blocked; }
259
260 protected:
261
262 CacheSlavePort(const std::string &_name, BaseCache *_cache,
263 const std::string &_label);
264
265 /** A normal packet queue used to store responses. */
266 RespPacketQueue queue;
267
268 bool blocked;
269
270 bool mustSendRetry;
271
272 private:
273
274 void processSendRetry();
275
276 EventFunctionWrapper sendRetryEvent;
277
278 };
279
280 /**
281 * The CPU-side port extends the base cache slave port with access
282 * functions for functional, atomic and timing requests.
283 */
284 class CpuSidePort : public CacheSlavePort
285 {
286 private:
287
288 // a pointer to our specific cache implementation
289 BaseCache *cache;
290
291 protected:
292 virtual bool recvTimingSnoopResp(PacketPtr pkt) override;
293
294 virtual bool tryTiming(PacketPtr pkt) override;
295
296 virtual bool recvTimingReq(PacketPtr pkt) override;
297
298 virtual Tick recvAtomic(PacketPtr pkt) override;
299
300 virtual void recvFunctional(PacketPtr pkt) override;
301
302 virtual AddrRangeList getAddrRanges() const override;
303
304 public:
305
306 CpuSidePort(const std::string &_name, BaseCache *_cache,
307 const std::string &_label);
308
309 };
310
311 CpuSidePort cpuSidePort;
312 MemSidePort memSidePort;
313
314 protected:
315
316 /** Miss status registers */
317 MSHRQueue mshrQueue;
318
319 /** Write/writeback buffer */
320 WriteQueue writeBuffer;
321
322 /** Tag and data Storage */
323 BaseTags *tags;
324
325 /** Prefetcher */
326 BasePrefetcher *prefetcher;
327
327 /**
328 * Notify the prefetcher on every access, not just misses.
329 */
330 const bool prefetchOnAccess;
328 /** To probe when a cache hit occurs */
329 ProbePointArg<PacketPtr> *ppHit;
330
331 /** To probe when a cache miss occurs */
332 ProbePointArg<PacketPtr> *ppMiss;
333
334 /**
335 * The writeAllocator drive optimizations for streaming writes.
336 * It first determines whether a WriteReq MSHR should be delayed,
337 * thus ensuring that we wait longer in cases when we are write
338 * coalescing and allowing all the bytes of the line to be written
339 * before the MSHR packet is sent downstream. This works in unison
340 * with the tracking in the MSHR to check if the entire line is
341 * written. The write mode also affects the behaviour on filling
342 * any whole-line writes. Normally the cache allocates the line
343 * when receiving the InvalidateResp, but after seeing enough
344 * consecutive lines we switch to using the tempBlock, and thus
345 * end up not allocating the line, and instead turning the
346 * whole-line write into a writeback straight away.
347 */
348 WriteAllocator * const writeAllocator;
349
350 /**
351 * Temporary cache block for occasional transitory use. We use
352 * the tempBlock to fill when allocation fails (e.g., when there
353 * is an outstanding request that accesses the victim block) or
354 * when we want to avoid allocation (e.g., exclusive caches)
355 */
356 TempCacheBlk *tempBlock;
357
358 /**
359 * Upstream caches need this packet until true is returned, so
360 * hold it for deletion until a subsequent call
361 */
362 std::unique_ptr<Packet> pendingDelete;
363
364 /**
365 * Mark a request as in service (sent downstream in the memory
366 * system), effectively making this MSHR the ordering point.
367 */
368 void markInService(MSHR *mshr, bool pending_modified_resp)
369 {
370 bool wasFull = mshrQueue.isFull();
371 mshrQueue.markInService(mshr, pending_modified_resp);
372
373 if (wasFull && !mshrQueue.isFull()) {
374 clearBlocked(Blocked_NoMSHRs);
375 }
376 }
377
378 void markInService(WriteQueueEntry *entry)
379 {
380 bool wasFull = writeBuffer.isFull();
381 writeBuffer.markInService(entry);
382
383 if (wasFull && !writeBuffer.isFull()) {
384 clearBlocked(Blocked_NoWBBuffers);
385 }
386 }
387
388 /**
389 * Determine whether we should allocate on a fill or not. If this
390 * cache is mostly inclusive with regards to the upstream cache(s)
391 * we always allocate (for any non-forwarded and cacheable
392 * requests). In the case of a mostly exclusive cache, we allocate
393 * on fill if the packet did not come from a cache, thus if we:
394 * are dealing with a whole-line write (the latter behaves much
395 * like a writeback), the original target packet came from a
396 * non-caching source, or if we are performing a prefetch or LLSC.
397 *
398 * @param cmd Command of the incoming requesting packet
399 * @return Whether we should allocate on the fill
400 */
401 inline bool allocOnFill(MemCmd cmd) const
402 {
403 return clusivity == Enums::mostly_incl ||
404 cmd == MemCmd::WriteLineReq ||
405 cmd == MemCmd::ReadReq ||
406 cmd == MemCmd::WriteReq ||
407 cmd.isPrefetch() ||
408 cmd.isLLSC();
409 }
410
411 /**
412 * Regenerate block address using tags.
413 * Block address regeneration depends on whether we're using a temporary
414 * block or not.
415 *
416 * @param blk The block to regenerate address.
417 * @return The block's address.
418 */
419 Addr regenerateBlkAddr(CacheBlk* blk);
420
421 /**
422 * Does all the processing necessary to perform the provided request.
423 * @param pkt The memory request to perform.
424 * @param blk The cache block to be updated.
425 * @param lat The latency of the access.
426 * @param writebacks List for any writebacks that need to be performed.
427 * @return Boolean indicating whether the request was satisfied.
428 */
429 virtual bool access(PacketPtr pkt, CacheBlk *&blk, Cycles &lat,
430 PacketList &writebacks);
431
432 /*
433 * Handle a timing request that hit in the cache
434 *
435 * @param ptk The request packet
436 * @param blk The referenced block
437 * @param request_time The tick at which the block lookup is compete
438 */
439 virtual void handleTimingReqHit(PacketPtr pkt, CacheBlk *blk,
440 Tick request_time);
441
442 /*
443 * Handle a timing request that missed in the cache
444 *
445 * Implementation specific handling for different cache
446 * implementations
447 *
448 * @param ptk The request packet
449 * @param blk The referenced block
450 * @param forward_time The tick at which we can process dependent requests
451 * @param request_time The tick at which the block lookup is compete
452 */
453 virtual void handleTimingReqMiss(PacketPtr pkt, CacheBlk *blk,
454 Tick forward_time,
455 Tick request_time) = 0;
456
457 /*
458 * Handle a timing request that missed in the cache
459 *
460 * Common functionality across different cache implementations
461 *
462 * @param ptk The request packet
463 * @param blk The referenced block
464 * @param mshr Any existing mshr for the referenced cache block
465 * @param forward_time The tick at which we can process dependent requests
466 * @param request_time The tick at which the block lookup is compete
467 */
468 void handleTimingReqMiss(PacketPtr pkt, MSHR *mshr, CacheBlk *blk,
469 Tick forward_time, Tick request_time);
470
471 /**
472 * Performs the access specified by the request.
473 * @param pkt The request to perform.
474 */
475 virtual void recvTimingReq(PacketPtr pkt);
476
477 /**
478 * Handling the special case of uncacheable write responses to
479 * make recvTimingResp less cluttered.
480 */
481 void handleUncacheableWriteResp(PacketPtr pkt);
482
483 /**
484 * Service non-deferred MSHR targets using the received response
485 *
486 * Iterates through the list of targets that can be serviced with
487 * the current response. Any writebacks that need to performed
488 * must be appended to the writebacks parameter.
489 *
490 * @param mshr The MSHR that corresponds to the reponse
491 * @param pkt The response packet
492 * @param blk The reference block
493 * @param writebacks List of writebacks that need to be performed
494 */
495 virtual void serviceMSHRTargets(MSHR *mshr, const PacketPtr pkt,
496 CacheBlk *blk, PacketList& writebacks) = 0;
497
498 /**
499 * Handles a response (cache line fill/write ack) from the bus.
500 * @param pkt The response packet
501 */
502 virtual void recvTimingResp(PacketPtr pkt);
503
504 /**
505 * Snoops bus transactions to maintain coherence.
506 * @param pkt The current bus transaction.
507 */
508 virtual void recvTimingSnoopReq(PacketPtr pkt) = 0;
509
510 /**
511 * Handle a snoop response.
512 * @param pkt Snoop response packet
513 */
514 virtual void recvTimingSnoopResp(PacketPtr pkt) = 0;
515
516 /**
517 * Handle a request in atomic mode that missed in this cache
518 *
519 * Creates a downstream request, sends it to the memory below and
520 * handles the response. As we are in atomic mode all operations
521 * are performed immediately.
522 *
523 * @param pkt The packet with the requests
524 * @param blk The referenced block
525 * @param writebacks A list with packets for any performed writebacks
526 * @return Cycles for handling the request
527 */
528 virtual Cycles handleAtomicReqMiss(PacketPtr pkt, CacheBlk *&blk,
529 PacketList &writebacks) = 0;
530
531 /**
532 * Performs the access specified by the request.
533 * @param pkt The request to perform.
534 * @return The number of ticks required for the access.
535 */
536 virtual Tick recvAtomic(PacketPtr pkt);
537
538 /**
539 * Snoop for the provided request in the cache and return the estimated
540 * time taken.
541 * @param pkt The memory request to snoop
542 * @return The number of ticks required for the snoop.
543 */
544 virtual Tick recvAtomicSnoop(PacketPtr pkt) = 0;
545
546 /**
547 * Performs the access specified by the request.
548 *
549 * @param pkt The request to perform.
550 * @param fromCpuSide from the CPU side port or the memory side port
551 */
552 virtual void functionalAccess(PacketPtr pkt, bool from_cpu_side);
553
554 /**
555 * Handle doing the Compare and Swap function for SPARC.
556 */
557 void cmpAndSwap(CacheBlk *blk, PacketPtr pkt);
558
559 /**
560 * Return the next queue entry to service, either a pending miss
561 * from the MSHR queue, a buffered write from the write buffer, or
562 * something from the prefetcher. This function is responsible
563 * for prioritizing among those sources on the fly.
564 */
565 QueueEntry* getNextQueueEntry();
566
567 /**
568 * Insert writebacks into the write buffer
569 */
570 virtual void doWritebacks(PacketList& writebacks, Tick forward_time) = 0;
571
572 /**
573 * Send writebacks down the memory hierarchy in atomic mode
574 */
575 virtual void doWritebacksAtomic(PacketList& writebacks) = 0;
576
577 /**
578 * Create an appropriate downstream bus request packet.
579 *
580 * Creates a new packet with the request to be send to the memory
581 * below, or nullptr if the current request in cpu_pkt should just
582 * be forwarded on.
583 *
584 * @param cpu_pkt The miss packet that needs to be satisfied.
585 * @param blk The referenced block, can be nullptr.
586 * @param needs_writable Indicates that the block must be writable
587 * even if the request in cpu_pkt doesn't indicate that.
588 * @param is_whole_line_write True if there are writes for the
589 * whole line
590 * @return A packet send to the memory below
591 */
592 virtual PacketPtr createMissPacket(PacketPtr cpu_pkt, CacheBlk *blk,
593 bool needs_writable,
594 bool is_whole_line_write) const = 0;
595
596 /**
597 * Determine if clean lines should be written back or not. In
598 * cases where a downstream cache is mostly inclusive we likely
599 * want it to act as a victim cache also for lines that have not
600 * been modified. Hence, we cannot simply drop the line (or send a
601 * clean evict), but rather need to send the actual data.
602 */
603 const bool writebackClean;
604
605 /**
606 * Writebacks from the tempBlock, resulting on the response path
607 * in atomic mode, must happen after the call to recvAtomic has
608 * finished (for the right ordering of the packets). We therefore
609 * need to hold on to the packets, and have a method and an event
610 * to send them.
611 */
612 PacketPtr tempBlockWriteback;
613
614 /**
615 * Send the outstanding tempBlock writeback. To be called after
616 * recvAtomic finishes in cases where the block we filled is in
617 * fact the tempBlock, and now needs to be written back.
618 */
619 void writebackTempBlockAtomic() {
620 assert(tempBlockWriteback != nullptr);
621 PacketList writebacks{tempBlockWriteback};
622 doWritebacksAtomic(writebacks);
623 tempBlockWriteback = nullptr;
624 }
625
626 /**
627 * An event to writeback the tempBlock after recvAtomic
628 * finishes. To avoid other calls to recvAtomic getting in
629 * between, we create this event with a higher priority.
630 */
631 EventFunctionWrapper writebackTempBlockAtomicEvent;
632
633 /**
634 * Perform any necessary updates to the block and perform any data
635 * exchange between the packet and the block. The flags of the
636 * packet are also set accordingly.
637 *
638 * @param pkt Request packet from upstream that hit a block
639 * @param blk Cache block that the packet hit
640 * @param deferred_response Whether this request originally missed
641 * @param pending_downgrade Whether the writable flag is to be removed
642 */
643 virtual void satisfyRequest(PacketPtr pkt, CacheBlk *blk,
644 bool deferred_response = false,
645 bool pending_downgrade = false);
646
647 /**
648 * Maintain the clusivity of this cache by potentially
649 * invalidating a block. This method works in conjunction with
650 * satisfyRequest, but is separate to allow us to handle all MSHR
651 * targets before potentially dropping a block.
652 *
653 * @param from_cache Whether we have dealt with a packet from a cache
654 * @param blk The block that should potentially be dropped
655 */
656 void maintainClusivity(bool from_cache, CacheBlk *blk);
657
658 /**
659 * Handle a fill operation caused by a received packet.
660 *
661 * Populates a cache block and handles all outstanding requests for the
662 * satisfied fill request. This version takes two memory requests. One
663 * contains the fill data, the other is an optional target to satisfy.
664 * Note that the reason we return a list of writebacks rather than
665 * inserting them directly in the write buffer is that this function
666 * is called by both atomic and timing-mode accesses, and in atomic
667 * mode we don't mess with the write buffer (we just perform the
668 * writebacks atomically once the original request is complete).
669 *
670 * @param pkt The memory request with the fill data.
671 * @param blk The cache block if it already exists.
672 * @param writebacks List for any writebacks that need to be performed.
673 * @param allocate Whether to allocate a block or use the temp block
674 * @return Pointer to the new cache block.
675 */
676 CacheBlk *handleFill(PacketPtr pkt, CacheBlk *blk,
677 PacketList &writebacks, bool allocate);
678
679 /**
680 * Allocate a new block and perform any necessary writebacks
681 *
682 * Find a victim block and if necessary prepare writebacks for any
683 * existing data. May return nullptr if there are no replaceable
684 * blocks. If a replaceable block is found, it inserts the new block in
685 * its place. The new block, however, is not set as valid yet.
686 *
687 * @param pkt Packet holding the address to update
688 * @param writebacks A list of writeback packets for the evicted blocks
689 * @return the allocated block
690 */
691 CacheBlk *allocateBlock(const PacketPtr pkt, PacketList &writebacks);
692 /**
693 * Evict a cache block.
694 *
695 * Performs a writeback if necesssary and invalidates the block
696 *
697 * @param blk Block to invalidate
698 * @return A packet with the writeback, can be nullptr
699 */
700 M5_NODISCARD virtual PacketPtr evictBlock(CacheBlk *blk) = 0;
701
702 /**
703 * Evict a cache block.
704 *
705 * Performs a writeback if necesssary and invalidates the block
706 *
707 * @param blk Block to invalidate
708 * @param writebacks Return a list of packets with writebacks
709 */
710 void evictBlock(CacheBlk *blk, PacketList &writebacks);
711
712 /**
713 * Invalidate a cache block.
714 *
715 * @param blk Block to invalidate
716 */
717 void invalidateBlock(CacheBlk *blk);
718
719 /**
720 * Create a writeback request for the given block.
721 *
722 * @param blk The block to writeback.
723 * @return The writeback request for the block.
724 */
725 PacketPtr writebackBlk(CacheBlk *blk);
726
727 /**
728 * Create a writeclean request for the given block.
729 *
730 * Creates a request that writes the block to the cache below
731 * without evicting the block from the current cache.
732 *
733 * @param blk The block to write clean.
734 * @param dest The destination of the write clean operation.
735 * @param id Use the given packet id for the write clean operation.
736 * @return The generated write clean packet.
737 */
738 PacketPtr writecleanBlk(CacheBlk *blk, Request::Flags dest, PacketId id);
739
740 /**
741 * Write back dirty blocks in the cache using functional accesses.
742 */
743 virtual void memWriteback() override;
744
745 /**
746 * Invalidates all blocks in the cache.
747 *
748 * @warn Dirty cache lines will not be written back to
749 * memory. Make sure to call functionalWriteback() first if you
750 * want the to write them to memory.
751 */
752 virtual void memInvalidate() override;
753
754 /**
755 * Determine if there are any dirty blocks in the cache.
756 *
757 * @return true if at least one block is dirty, false otherwise.
758 */
759 bool isDirty() const;
760
761 /**
762 * Determine if an address is in the ranges covered by this
763 * cache. This is useful to filter snoops.
764 *
765 * @param addr Address to check against
766 *
767 * @return If the address in question is in range
768 */
769 bool inRange(Addr addr) const;
770
771 /**
772 * Find next request ready time from among possible sources.
773 */
774 Tick nextQueueReadyTime() const;
775
776 /** Block size of this cache */
777 const unsigned blkSize;
778
779 /**
780 * The latency of tag lookup of a cache. It occurs when there is
781 * an access to the cache.
782 */
783 const Cycles lookupLatency;
784
785 /**
786 * The latency of data access of a cache. It occurs when there is
787 * an access to the cache.
788 */
789 const Cycles dataLatency;
790
791 /**
792 * This is the forward latency of the cache. It occurs when there
793 * is a cache miss and a request is forwarded downstream, in
794 * particular an outbound miss.
795 */
796 const Cycles forwardLatency;
797
798 /** The latency to fill a cache block */
799 const Cycles fillLatency;
800
801 /**
802 * The latency of sending reponse to its upper level cache/core on
803 * a linefill. The responseLatency parameter captures this
804 * latency.
805 */
806 const Cycles responseLatency;
807
808 /** The number of targets for each MSHR. */
809 const int numTarget;
810
811 /** Do we forward snoops from mem side port through to cpu side port? */
812 bool forwardSnoops;
813
814 /**
815 * Clusivity with respect to the upstream cache, determining if we
816 * fill into both this cache and the cache above on a miss. Note
817 * that we currently do not support strict clusivity policies.
818 */
819 const Enums::Clusivity clusivity;
820
821 /**
822 * Is this cache read only, for example the instruction cache, or
823 * table-walker cache. A cache that is read only should never see
824 * any writes, and should never get any dirty data (and hence
825 * never have to do any writebacks).
826 */
827 const bool isReadOnly;
828
829 /**
830 * Bit vector of the blocking reasons for the access path.
831 * @sa #BlockedCause
832 */
833 uint8_t blocked;
834
835 /** Increasing order number assigned to each incoming request. */
836 uint64_t order;
837
838 /** Stores time the cache blocked for statistics. */
839 Cycles blockedCycle;
840
841 /** Pointer to the MSHR that has no targets. */
842 MSHR *noTargetMSHR;
843
844 /** The number of misses to trigger an exit event. */
845 Counter missCount;
846
847 /**
848 * The address range to which the cache responds on the CPU side.
849 * Normally this is all possible memory addresses. */
850 const AddrRangeList addrRanges;
851
852 public:
853 /** System we are currently operating in. */
854 System *system;
855
856 // Statistics
857 /**
858 * @addtogroup CacheStatistics
859 * @{
860 */
861
862 /** Number of hits per thread for each type of command.
863 @sa Packet::Command */
864 Stats::Vector hits[MemCmd::NUM_MEM_CMDS];
865 /** Number of hits for demand accesses. */
866 Stats::Formula demandHits;
867 /** Number of hit for all accesses. */
868 Stats::Formula overallHits;
869
870 /** Number of misses per thread for each type of command.
871 @sa Packet::Command */
872 Stats::Vector misses[MemCmd::NUM_MEM_CMDS];
873 /** Number of misses for demand accesses. */
874 Stats::Formula demandMisses;
875 /** Number of misses for all accesses. */
876 Stats::Formula overallMisses;
877
878 /**
879 * Total number of cycles per thread/command spent waiting for a miss.
880 * Used to calculate the average miss latency.
881 */
882 Stats::Vector missLatency[MemCmd::NUM_MEM_CMDS];
883 /** Total number of cycles spent waiting for demand misses. */
884 Stats::Formula demandMissLatency;
885 /** Total number of cycles spent waiting for all misses. */
886 Stats::Formula overallMissLatency;
887
888 /** The number of accesses per command and thread. */
889 Stats::Formula accesses[MemCmd::NUM_MEM_CMDS];
890 /** The number of demand accesses. */
891 Stats::Formula demandAccesses;
892 /** The number of overall accesses. */
893 Stats::Formula overallAccesses;
894
895 /** The miss rate per command and thread. */
896 Stats::Formula missRate[MemCmd::NUM_MEM_CMDS];
897 /** The miss rate of all demand accesses. */
898 Stats::Formula demandMissRate;
899 /** The miss rate for all accesses. */
900 Stats::Formula overallMissRate;
901
902 /** The average miss latency per command and thread. */
903 Stats::Formula avgMissLatency[MemCmd::NUM_MEM_CMDS];
904 /** The average miss latency for demand misses. */
905 Stats::Formula demandAvgMissLatency;
906 /** The average miss latency for all misses. */
907 Stats::Formula overallAvgMissLatency;
908
909 /** The total number of cycles blocked for each blocked cause. */
910 Stats::Vector blocked_cycles;
911 /** The number of times this cache blocked for each blocked cause. */
912 Stats::Vector blocked_causes;
913
914 /** The average number of cycles blocked for each blocked cause. */
915 Stats::Formula avg_blocked;
916
917 /** The number of times a HW-prefetched block is evicted w/o reference. */
918 Stats::Scalar unusedPrefetches;
919
920 /** Number of blocks written back per thread. */
921 Stats::Vector writebacks;
922
923 /** Number of misses that hit in the MSHRs per command and thread. */
924 Stats::Vector mshr_hits[MemCmd::NUM_MEM_CMDS];
925 /** Demand misses that hit in the MSHRs. */
926 Stats::Formula demandMshrHits;
927 /** Total number of misses that hit in the MSHRs. */
928 Stats::Formula overallMshrHits;
929
930 /** Number of misses that miss in the MSHRs, per command and thread. */
931 Stats::Vector mshr_misses[MemCmd::NUM_MEM_CMDS];
932 /** Demand misses that miss in the MSHRs. */
933 Stats::Formula demandMshrMisses;
934 /** Total number of misses that miss in the MSHRs. */
935 Stats::Formula overallMshrMisses;
936
937 /** Number of misses that miss in the MSHRs, per command and thread. */
938 Stats::Vector mshr_uncacheable[MemCmd::NUM_MEM_CMDS];
939 /** Total number of misses that miss in the MSHRs. */
940 Stats::Formula overallMshrUncacheable;
941
942 /** Total cycle latency of each MSHR miss, per command and thread. */
943 Stats::Vector mshr_miss_latency[MemCmd::NUM_MEM_CMDS];
944 /** Total cycle latency of demand MSHR misses. */
945 Stats::Formula demandMshrMissLatency;
946 /** Total cycle latency of overall MSHR misses. */
947 Stats::Formula overallMshrMissLatency;
948
949 /** Total cycle latency of each MSHR miss, per command and thread. */
950 Stats::Vector mshr_uncacheable_lat[MemCmd::NUM_MEM_CMDS];
951 /** Total cycle latency of overall MSHR misses. */
952 Stats::Formula overallMshrUncacheableLatency;
953
954#if 0
955 /** The total number of MSHR accesses per command and thread. */
956 Stats::Formula mshrAccesses[MemCmd::NUM_MEM_CMDS];
957 /** The total number of demand MSHR accesses. */
958 Stats::Formula demandMshrAccesses;
959 /** The total number of MSHR accesses. */
960 Stats::Formula overallMshrAccesses;
961#endif
962
963 /** The miss rate in the MSHRs pre command and thread. */
964 Stats::Formula mshrMissRate[MemCmd::NUM_MEM_CMDS];
965 /** The demand miss rate in the MSHRs. */
966 Stats::Formula demandMshrMissRate;
967 /** The overall miss rate in the MSHRs. */
968 Stats::Formula overallMshrMissRate;
969
970 /** The average latency of an MSHR miss, per command and thread. */
971 Stats::Formula avgMshrMissLatency[MemCmd::NUM_MEM_CMDS];
972 /** The average latency of a demand MSHR miss. */
973 Stats::Formula demandAvgMshrMissLatency;
974 /** The average overall latency of an MSHR miss. */
975 Stats::Formula overallAvgMshrMissLatency;
976
977 /** The average latency of an MSHR miss, per command and thread. */
978 Stats::Formula avgMshrUncacheableLatency[MemCmd::NUM_MEM_CMDS];
979 /** The average overall latency of an MSHR miss. */
980 Stats::Formula overallAvgMshrUncacheableLatency;
981
982 /** Number of replacements of valid blocks. */
983 Stats::Scalar replacements;
984
985 /**
986 * @}
987 */
988
989 /**
990 * Register stats for this object.
991 */
992 void regStats() override;
993
994 /** Registers probes. */
995 void regProbePoints() override;
996
997 public:
998 BaseCache(const BaseCacheParams *p, unsigned blk_size);
999 ~BaseCache();
1000
1001 void init() override;
1002
1003 BaseMasterPort &getMasterPort(const std::string &if_name,
1004 PortID idx = InvalidPortID) override;
1005 BaseSlavePort &getSlavePort(const std::string &if_name,
1006 PortID idx = InvalidPortID) override;
1007
1008 /**
1009 * Query block size of a cache.
1010 * @return The block size
1011 */
1012 unsigned
1013 getBlockSize() const
1014 {
1015 return blkSize;
1016 }
1017
1018 const AddrRangeList &getAddrRanges() const { return addrRanges; }
1019
1020 MSHR *allocateMissBuffer(PacketPtr pkt, Tick time, bool sched_send = true)
1021 {
1022 MSHR *mshr = mshrQueue.allocate(pkt->getBlockAddr(blkSize), blkSize,
1023 pkt, time, order++,
1024 allocOnFill(pkt->cmd));
1025
1026 if (mshrQueue.isFull()) {
1027 setBlocked((BlockedCause)MSHRQueue_MSHRs);
1028 }
1029
1030 if (sched_send) {
1031 // schedule the send
1032 schedMemSideSendEvent(time);
1033 }
1034
1035 return mshr;
1036 }
1037
1038 void allocateWriteBuffer(PacketPtr pkt, Tick time)
1039 {
1040 // should only see writes or clean evicts here
1041 assert(pkt->isWrite() || pkt->cmd == MemCmd::CleanEvict);
1042
1043 Addr blk_addr = pkt->getBlockAddr(blkSize);
1044
1045 WriteQueueEntry *wq_entry =
1046 writeBuffer.findMatch(blk_addr, pkt->isSecure());
1047 if (wq_entry && !wq_entry->inService) {
1048 DPRINTF(Cache, "Potential to merge writeback %s", pkt->print());
1049 }
1050
1051 writeBuffer.allocate(blk_addr, blkSize, pkt, time, order++);
1052
1053 if (writeBuffer.isFull()) {
1054 setBlocked((BlockedCause)MSHRQueue_WriteBuffer);
1055 }
1056
1057 // schedule the send
1058 schedMemSideSendEvent(time);
1059 }
1060
1061 /**
1062 * Returns true if the cache is blocked for accesses.
1063 */
1064 bool isBlocked() const
1065 {
1066 return blocked != 0;
1067 }
1068
1069 /**
1070 * Marks the access path of the cache as blocked for the given cause. This
1071 * also sets the blocked flag in the slave interface.
1072 * @param cause The reason for the cache blocking.
1073 */
1074 void setBlocked(BlockedCause cause)
1075 {
1076 uint8_t flag = 1 << cause;
1077 if (blocked == 0) {
1078 blocked_causes[cause]++;
1079 blockedCycle = curCycle();
1080 cpuSidePort.setBlocked();
1081 }
1082 blocked |= flag;
1083 DPRINTF(Cache,"Blocking for cause %d, mask=%d\n", cause, blocked);
1084 }
1085
1086 /**
1087 * Marks the cache as unblocked for the given cause. This also clears the
1088 * blocked flags in the appropriate interfaces.
1089 * @param cause The newly unblocked cause.
1090 * @warning Calling this function can cause a blocked request on the bus to
1091 * access the cache. The cache must be in a state to handle that request.
1092 */
1093 void clearBlocked(BlockedCause cause)
1094 {
1095 uint8_t flag = 1 << cause;
1096 blocked &= ~flag;
1097 DPRINTF(Cache,"Unblocking for cause %d, mask=%d\n", cause, blocked);
1098 if (blocked == 0) {
1099 blocked_cycles[cause] += curCycle() - blockedCycle;
1100 cpuSidePort.clearBlocked();
1101 }
1102 }
1103
1104 /**
1105 * Schedule a send event for the memory-side port. If already
1106 * scheduled, this may reschedule the event at an earlier
1107 * time. When the specified time is reached, the port is free to
1108 * send either a response, a request, or a prefetch request.
1109 *
1110 * @param time The time when to attempt sending a packet.
1111 */
1112 void schedMemSideSendEvent(Tick time)
1113 {
1114 memSidePort.schedSendEvent(time);
1115 }
1116
1117 bool inCache(Addr addr, bool is_secure) const {
1118 return tags->findBlock(addr, is_secure);
1119 }
1120
1121 bool inMissQueue(Addr addr, bool is_secure) const {
1122 return mshrQueue.findMatch(addr, is_secure);
1123 }
1124
1125 void incMissCount(PacketPtr pkt)
1126 {
1127 assert(pkt->req->masterId() < system->maxMasters());
1128 misses[pkt->cmdToIndex()][pkt->req->masterId()]++;
1129 pkt->req->incAccessDepth();
1130 if (missCount) {
1131 --missCount;
1132 if (missCount == 0)
1133 exitSimLoop("A cache reached the maximum miss count");
1134 }
1135 }
1136 void incHitCount(PacketPtr pkt)
1137 {
1138 assert(pkt->req->masterId() < system->maxMasters());
1139 hits[pkt->cmdToIndex()][pkt->req->masterId()]++;
1140
1141 }
1142
1143 /**
1144 * Checks if the cache is coalescing writes
1145 *
1146 * @return True if the cache is coalescing writes
1147 */
1148 bool coalesce() const;
1149
1150
1151 /**
1152 * Cache block visitor that writes back dirty cache blocks using
1153 * functional writes.
1154 */
1155 void writebackVisitor(CacheBlk &blk);
1156
1157 /**
1158 * Cache block visitor that invalidates all blocks in the cache.
1159 *
1160 * @warn Dirty cache lines will not be written back to memory.
1161 */
1162 void invalidateVisitor(CacheBlk &blk);
1163
1164 /**
1165 * Take an MSHR, turn it into a suitable downstream packet, and
1166 * send it out. This construct allows a queue entry to choose a suitable
1167 * approach based on its type.
1168 *
1169 * @param mshr The MSHR to turn into a packet and send
1170 * @return True if the port is waiting for a retry
1171 */
1172 virtual bool sendMSHRQueuePacket(MSHR* mshr);
1173
1174 /**
1175 * Similar to sendMSHR, but for a write-queue entry
1176 * instead. Create the packet, and send it, and if successful also
1177 * mark the entry in service.
1178 *
1179 * @param wq_entry The write-queue entry to turn into a packet and send
1180 * @return True if the port is waiting for a retry
1181 */
1182 bool sendWriteQueuePacket(WriteQueueEntry* wq_entry);
1183
1184 /**
1185 * Serialize the state of the caches
1186 *
1187 * We currently don't support checkpointing cache state, so this panics.
1188 */
1189 void serialize(CheckpointOut &cp) const override;
1190 void unserialize(CheckpointIn &cp) override;
1178
1191};
1192
1193/**
1194 * The write allocator inspects write packets and detects streaming
1195 * patterns. The write allocator supports a single stream where writes
1196 * are expected to access consecutive locations and keeps track of
1197 * size of the area covered by the concecutive writes in byteCount.
1198 *
1199 * 1) When byteCount has surpassed the coallesceLimit the mode
1200 * switches from ALLOCATE to COALESCE where writes should be delayed
1201 * until the whole block is written at which point a single packet
1202 * (whole line write) can service them.
1203 *
1204 * 2) When byteCount has also exceeded the noAllocateLimit (whole
1205 * line) we switch to NO_ALLOCATE when writes should not allocate in
1206 * the cache but rather send a whole line write to the memory below.
1207 */
1208class WriteAllocator : public SimObject {
1209 public:
1210 WriteAllocator(const WriteAllocatorParams *p) :
1211 SimObject(p),
1212 coalesceLimit(p->coalesce_limit * p->block_size),
1213 noAllocateLimit(p->no_allocate_limit * p->block_size),
1214 delayThreshold(p->delay_threshold)
1215 {
1216 reset();
1217 }
1218
1219 /**
1220 * Should writes be coalesced? This is true if the mode is set to
1221 * NO_ALLOCATE.
1222 *
1223 * @return return true if the cache should coalesce writes.
1224 */
1225 bool coalesce() const {
1226 return mode != WriteMode::ALLOCATE;
1227 }
1228
1229 /**
1230 * Should writes allocate?
1231 *
1232 * @return return true if the cache should not allocate for writes.
1233 */
1234 bool allocate() const {
1235 return mode != WriteMode::NO_ALLOCATE;
1236 }
1237
1238 /**
1239 * Reset the write allocator state, meaning that it allocates for
1240 * writes and has not recorded any information about qualifying
1241 * writes that might trigger a switch to coalescing and later no
1242 * allocation.
1243 */
1244 void reset() {
1245 mode = WriteMode::ALLOCATE;
1246 byteCount = 0;
1247 nextAddr = 0;
1248 }
1249
1250 /**
1251 * Access whether we need to delay the current write.
1252 *
1253 * @param blk_addr The block address the packet writes to
1254 * @return true if the current packet should be delayed
1255 */
1256 bool delay(Addr blk_addr) {
1257 if (delayCtr[blk_addr] > 0) {
1258 --delayCtr[blk_addr];
1259 return true;
1260 } else {
1261 return false;
1262 }
1263 }
1264
1265 /**
1266 * Clear delay counter for the input block
1267 *
1268 * @param blk_addr The accessed cache block
1269 */
1270 void resetDelay(Addr blk_addr) {
1271 delayCtr.erase(blk_addr);
1272 }
1273
1274 /**
1275 * Update the write mode based on the current write
1276 * packet. This method compares the packet's address with any
1277 * current stream, and updates the tracking and the mode
1278 * accordingly.
1279 *
1280 * @param write_addr Start address of the write request
1281 * @param write_size Size of the write request
1282 * @param blk_addr The block address that this packet writes to
1283 */
1284 void updateMode(Addr write_addr, unsigned write_size, Addr blk_addr);
1285
1286 private:
1287 /**
1288 * The current mode for write coalescing and allocation, either
1289 * normal operation (ALLOCATE), write coalescing (COALESCE), or
1290 * write coalescing without allocation (NO_ALLOCATE).
1291 */
1292 enum class WriteMode : char {
1293 ALLOCATE,
1294 COALESCE,
1295 NO_ALLOCATE,
1296 };
1297 WriteMode mode;
1298
1299 /** Address to match writes against to detect streams. */
1300 Addr nextAddr;
1301
1302 /**
1303 * Bytes written contiguously. Saturating once we no longer
1304 * allocate.
1305 */
1306 uint32_t byteCount;
1307
1308 /**
1309 * Limits for when to switch between the different write modes.
1310 */
1311 const uint32_t coalesceLimit;
1312 const uint32_t noAllocateLimit;
1313 /**
1314 * The number of times the allocator will delay an WriteReq MSHR.
1315 */
1316 const uint32_t delayThreshold;
1317
1318 /**
1319 * Keep track of the number of times the allocator has delayed an
1320 * WriteReq MSHR.
1321 */
1322 std::unordered_map<Addr, Counter> delayCtr;
1323};
1324
1325#endif //__MEM_CACHE_BASE_HH__