2 * Copyright (c) 2011-2017 ARM Limited
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.
14 * Copyright (c) 2006 The Regents of The University of Michigan
15 * All rights reserved.
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.
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.
47 * Definition of a crossbar object.
50 #include "mem/coherent_xbar.hh"
52 #include "base/logging.hh"
53 #include "base/trace.hh"
54 #include "debug/AddrRanges.hh"
55 #include "debug/CoherentXBar.hh"
56 #include "sim/system.hh"
58 CoherentXBar::CoherentXBar(const CoherentXBarParams
*p
)
59 : BaseXBar(p
), system(p
->system
), snoopFilter(p
->snoop_filter
),
60 snoopResponseLatency(p
->snoop_response_latency
),
61 pointOfCoherency(p
->point_of_coherency
),
62 pointOfUnification(p
->point_of_unification
)
64 // create the ports based on the size of the master and slave
65 // vector ports, and the presence of the default port, the ports
66 // are enumerated starting from zero
67 for (int i
= 0; i
< p
->port_master_connection_count
; ++i
) {
68 std::string portName
= csprintf("%s.master[%d]", name(), i
);
69 MasterPort
* bp
= new CoherentXBarMasterPort(portName
, *this, i
);
70 masterPorts
.push_back(bp
);
71 reqLayers
.push_back(new ReqLayer(*bp
, *this,
72 csprintf(".reqLayer%d", i
)));
73 snoopLayers
.push_back(new SnoopRespLayer(*bp
, *this,
74 csprintf(".snoopLayer%d", i
)));
77 // see if we have a default slave device connected and if so add
78 // our corresponding master port
79 if (p
->port_default_connection_count
) {
80 defaultPortID
= masterPorts
.size();
81 std::string portName
= name() + ".default";
82 MasterPort
* bp
= new CoherentXBarMasterPort(portName
, *this,
84 masterPorts
.push_back(bp
);
85 reqLayers
.push_back(new ReqLayer(*bp
, *this, csprintf(".reqLayer%d",
87 snoopLayers
.push_back(new SnoopRespLayer(*bp
, *this,
88 csprintf(".snoopLayer%d",
92 // create the slave ports, once again starting at zero
93 for (int i
= 0; i
< p
->port_slave_connection_count
; ++i
) {
94 std::string portName
= csprintf("%s.slave[%d]", name(), i
);
95 QueuedSlavePort
* bp
= new CoherentXBarSlavePort(portName
, *this, i
);
96 slavePorts
.push_back(bp
);
97 respLayers
.push_back(new RespLayer(*bp
, *this,
98 csprintf(".respLayer%d", i
)));
99 snoopRespPorts
.push_back(new SnoopRespPort(*bp
, *this));
105 CoherentXBar::~CoherentXBar()
107 for (auto l
: reqLayers
)
109 for (auto l
: respLayers
)
111 for (auto l
: snoopLayers
)
113 for (auto p
: snoopRespPorts
)
122 // iterate over our slave ports and determine which of our
123 // neighbouring master ports are snooping and add them as snoopers
124 for (const auto& p
: slavePorts
) {
125 // check if the connected master port is snooping
126 if (p
->isSnooping()) {
127 DPRINTF(AddrRanges
, "Adding snooping master %s\n",
128 p
->getMasterPort().name());
129 snoopPorts
.push_back(p
);
133 if (snoopPorts
.empty())
134 warn("CoherentXBar %s has no snooping ports attached!\n", name());
136 // inform the snoop filter about the slave ports so it can create
137 // its own internal representation
139 snoopFilter
->setSlavePorts(slavePorts
);
143 CoherentXBar::recvTimingReq(PacketPtr pkt
, PortID slave_port_id
)
145 // determine the source port based on the id
146 SlavePort
*src_port
= slavePorts
[slave_port_id
];
148 // remember if the packet is an express snoop
149 bool is_express_snoop
= pkt
->isExpressSnoop();
150 bool cache_responding
= pkt
->cacheResponding();
151 // for normal requests, going downstream, the express snoop flag
152 // and the cache responding flag should always be the same
153 assert(is_express_snoop
== cache_responding
);
155 // determine the destination based on the address
156 PortID master_port_id
= findPort(pkt
->getAddr());
158 // test if the crossbar should be considered occupied for the current
159 // port, and exclude express snoops from the check
160 if (!is_express_snoop
&& !reqLayers
[master_port_id
]->tryTiming(src_port
)) {
161 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
162 src_port
->name(), pkt
->print());
166 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
167 src_port
->name(), pkt
->print());
169 // store size and command as they might be modified when
170 // forwarding the packet
171 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
172 unsigned int pkt_cmd
= pkt
->cmdToIndex();
174 // store the old header delay so we can restore it if needed
175 Tick old_header_delay
= pkt
->headerDelay
;
177 // a request sees the frontend and forward latency
178 Tick xbar_delay
= (frontendLatency
+ forwardLatency
) * clockPeriod();
180 // set the packet header and payload delay
181 calcPacketTiming(pkt
, xbar_delay
);
183 // determine how long to be crossbar layer is busy
184 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
186 const bool snoop_caches
= !system
->bypassCaches() &&
187 pkt
->cmd
!= MemCmd::WriteClean
;
189 assert(pkt
->snoopDelay
== 0);
191 // the packet is a memory-mapped request and should be
192 // broadcasted to our snoopers but the source
194 // check with the snoop filter where to forward this packet
195 auto sf_res
= snoopFilter
->lookupRequest(pkt
, *src_port
);
196 // the time required by a packet to be delivered through
197 // the xbar has to be charged also with to lookup latency
198 // of the snoop filter
199 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
200 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
201 __func__
, src_port
->name(), pkt
->print(),
202 sf_res
.first
.size(), sf_res
.second
);
204 if (pkt
->isEviction()) {
205 // for block-evicting packets, i.e. writebacks and
206 // clean evictions, there is no need to snoop up, as
207 // all we do is determine if the block is cached or
208 // not, instead just set it here based on the snoop
210 if (!sf_res
.first
.empty())
211 pkt
->setBlockCached();
213 forwardTiming(pkt
, slave_port_id
, sf_res
.first
);
216 forwardTiming(pkt
, slave_port_id
);
219 // add the snoop delay to our header delay, and then reset it
220 pkt
->headerDelay
+= pkt
->snoopDelay
;
224 // set up a sensible starting point
227 // remember if the packet will generate a snoop response by
228 // checking if a cache set the cacheResponding flag during the
230 const bool expect_snoop_resp
= !cache_responding
&& pkt
->cacheResponding();
231 bool expect_response
= pkt
->needsResponse() && !pkt
->cacheResponding();
233 const bool sink_packet
= sinkPacket(pkt
);
235 // in certain cases the crossbar is responsible for responding
236 bool respond_directly
= false;
237 // store the original address as an address mapper could possibly
238 // modify the address upon a sendTimingRequest
239 const Addr
addr(pkt
->getAddr());
241 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
244 // determine if we are forwarding the packet, or responding to
246 if (!pointOfCoherency
|| pkt
->isRead() || pkt
->isWrite()) {
247 // if we are passing on, rather than sinking, a packet to
248 // which an upstream cache has committed to responding,
249 // the line was needs writable, and the responding only
250 // had an Owned copy, so we need to immidiately let the
251 // downstream caches know, bypass any flow control
252 if (pkt
->cacheResponding()) {
253 pkt
->setExpressSnoop();
256 // since it is a normal request, attempt to send the packet
257 success
= masterPorts
[master_port_id
]->sendTimingReq(pkt
);
259 // no need to forward, turn this packet around and respond
261 assert(pkt
->needsResponse());
263 respond_directly
= true;
264 assert(!expect_snoop_resp
);
265 expect_response
= false;
269 if (snoopFilter
&& snoop_caches
) {
270 // Let the snoop filter know about the success of the send operation
271 snoopFilter
->finishRequest(!success
, addr
, pkt
->isSecure());
274 // check if we were successful in sending the packet onwards
276 // express snoops should never be forced to retry
277 assert(!is_express_snoop
);
279 // restore the header delay
280 pkt
->headerDelay
= old_header_delay
;
282 DPRINTF(CoherentXBar
, "%s: src %s packet %s RETRY\n", __func__
,
283 src_port
->name(), pkt
->print());
285 // update the layer state and schedule an idle event
286 reqLayers
[master_port_id
]->failedTiming(src_port
,
287 clockEdge(Cycles(1)));
289 // express snoops currently bypass the crossbar state entirely
290 if (!is_express_snoop
) {
291 // if this particular request will generate a snoop
293 if (expect_snoop_resp
) {
294 // we should never have an exsiting request outstanding
295 assert(outstandingSnoop
.find(pkt
->req
) ==
296 outstandingSnoop
.end());
297 outstandingSnoop
.insert(pkt
->req
);
299 // basic sanity check on the outstanding snoops
300 panic_if(outstandingSnoop
.size() > 512,
301 "Outstanding snoop requests exceeded 512\n");
304 // remember where to route the normal response to
305 if (expect_response
|| expect_snoop_resp
) {
306 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
307 routeTo
[pkt
->req
] = slave_port_id
;
309 panic_if(routeTo
.size() > 512,
310 "Routing table exceeds 512 packets\n");
313 // update the layer state and schedule an idle event
314 reqLayers
[master_port_id
]->succeededTiming(packetFinishTime
);
317 // stats updates only consider packets that were successfully sent
318 pktCount
[slave_port_id
][master_port_id
]++;
319 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
320 transDist
[pkt_cmd
]++;
322 if (is_express_snoop
) {
324 snoopTraffic
+= pkt_size
;
329 // queue the packet for deletion
330 pendingDelete
.reset(pkt
);
332 if (respond_directly
) {
333 assert(pkt
->needsResponse());
338 if (snoopFilter
&& !system
->bypassCaches()) {
339 // let the snoop filter inspect the response and update its state
340 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
343 Tick response_time
= clockEdge() + pkt
->headerDelay
;
344 pkt
->headerDelay
= 0;
346 slavePorts
[slave_port_id
]->schedTimingResp(pkt
, response_time
);
353 CoherentXBar::recvTimingResp(PacketPtr pkt
, PortID master_port_id
)
355 // determine the source port based on the id
356 MasterPort
*src_port
= masterPorts
[master_port_id
];
358 // determine the destination
359 const auto route_lookup
= routeTo
.find(pkt
->req
);
360 assert(route_lookup
!= routeTo
.end());
361 const PortID slave_port_id
= route_lookup
->second
;
362 assert(slave_port_id
!= InvalidPortID
);
363 assert(slave_port_id
< respLayers
.size());
365 // test if the crossbar should be considered occupied for the
367 if (!respLayers
[slave_port_id
]->tryTiming(src_port
)) {
368 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
369 src_port
->name(), pkt
->print());
373 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
374 src_port
->name(), pkt
->print());
376 // store size and command as they might be modified when
377 // forwarding the packet
378 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
379 unsigned int pkt_cmd
= pkt
->cmdToIndex();
381 // a response sees the response latency
382 Tick xbar_delay
= responseLatency
* clockPeriod();
384 // set the packet header and payload delay
385 calcPacketTiming(pkt
, xbar_delay
);
387 // determine how long to be crossbar layer is busy
388 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
390 if (snoopFilter
&& !system
->bypassCaches()) {
391 // let the snoop filter inspect the response and update its state
392 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
395 // send the packet through the destination slave port and pay for
396 // any outstanding header delay
397 Tick latency
= pkt
->headerDelay
;
398 pkt
->headerDelay
= 0;
399 slavePorts
[slave_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
401 // remove the request from the routing table
402 routeTo
.erase(route_lookup
);
404 respLayers
[slave_port_id
]->succeededTiming(packetFinishTime
);
407 pktCount
[slave_port_id
][master_port_id
]++;
408 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
409 transDist
[pkt_cmd
]++;
415 CoherentXBar::recvTimingSnoopReq(PacketPtr pkt
, PortID master_port_id
)
417 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
418 masterPorts
[master_port_id
]->name(), pkt
->print());
420 // update stats here as we know the forwarding will succeed
421 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
422 transDist
[pkt
->cmdToIndex()]++;
424 snoopTraffic
+= pkt_size
;
426 // we should only see express snoops from caches
427 assert(pkt
->isExpressSnoop());
429 // set the packet header and payload delay, for now use forward latency
430 // @todo Assess the choice of latency further
431 calcPacketTiming(pkt
, forwardLatency
* clockPeriod());
433 // remember if a cache has already committed to responding so we
434 // can see if it changes during the snooping
435 const bool cache_responding
= pkt
->cacheResponding();
437 assert(pkt
->snoopDelay
== 0);
440 // let the Snoop Filter work its magic and guide probing
441 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
442 // the time required by a packet to be delivered through
443 // the xbar has to be charged also with to lookup latency
444 // of the snoop filter
445 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
446 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
447 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
448 sf_res
.first
.size(), sf_res
.second
);
450 // forward to all snoopers
451 forwardTiming(pkt
, InvalidPortID
, sf_res
.first
);
453 forwardTiming(pkt
, InvalidPortID
);
456 // add the snoop delay to our header delay, and then reset it
457 pkt
->headerDelay
+= pkt
->snoopDelay
;
460 // if we can expect a response, remember how to route it
461 if (!cache_responding
&& pkt
->cacheResponding()) {
462 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
463 routeTo
[pkt
->req
] = master_port_id
;
466 // a snoop request came from a connected slave device (one of
467 // our master ports), and if it is not coming from the slave
468 // device responsible for the address range something is
469 // wrong, hence there is nothing further to do as the packet
470 // would be going back to where it came from
471 assert(master_port_id
== findPort(pkt
->getAddr()));
475 CoherentXBar::recvTimingSnoopResp(PacketPtr pkt
, PortID slave_port_id
)
477 // determine the source port based on the id
478 SlavePort
* src_port
= slavePorts
[slave_port_id
];
480 // get the destination
481 const auto route_lookup
= routeTo
.find(pkt
->req
);
482 assert(route_lookup
!= routeTo
.end());
483 const PortID dest_port_id
= route_lookup
->second
;
484 assert(dest_port_id
!= InvalidPortID
);
486 // determine if the response is from a snoop request we
487 // created as the result of a normal request (in which case it
488 // should be in the outstandingSnoop), or if we merely forwarded
489 // someone else's snoop request
490 const bool forwardAsSnoop
= outstandingSnoop
.find(pkt
->req
) ==
491 outstandingSnoop
.end();
493 // test if the crossbar should be considered occupied for the
494 // current port, note that the check is bypassed if the response
495 // is being passed on as a normal response since this is occupying
496 // the response layer rather than the snoop response layer
497 if (forwardAsSnoop
) {
498 assert(dest_port_id
< snoopLayers
.size());
499 if (!snoopLayers
[dest_port_id
]->tryTiming(src_port
)) {
500 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
501 src_port
->name(), pkt
->print());
505 // get the master port that mirrors this slave port internally
506 MasterPort
* snoop_port
= snoopRespPorts
[slave_port_id
];
507 assert(dest_port_id
< respLayers
.size());
508 if (!respLayers
[dest_port_id
]->tryTiming(snoop_port
)) {
509 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
510 snoop_port
->name(), pkt
->print());
515 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
516 src_port
->name(), pkt
->print());
518 // store size and command as they might be modified when
519 // forwarding the packet
520 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
521 unsigned int pkt_cmd
= pkt
->cmdToIndex();
523 // responses are never express snoops
524 assert(!pkt
->isExpressSnoop());
526 // a snoop response sees the snoop response latency, and if it is
527 // forwarded as a normal response, the response latency
529 (forwardAsSnoop
? snoopResponseLatency
: responseLatency
) *
532 // set the packet header and payload delay
533 calcPacketTiming(pkt
, xbar_delay
);
535 // determine how long to be crossbar layer is busy
536 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
538 // forward it either as a snoop response or a normal response
539 if (forwardAsSnoop
) {
540 // this is a snoop response to a snoop request we forwarded,
541 // e.g. coming from the L1 and going to the L2, and it should
542 // be forwarded as a snoop response
545 // update the probe filter so that it can properly track the line
546 snoopFilter
->updateSnoopForward(pkt
, *slavePorts
[slave_port_id
],
547 *masterPorts
[dest_port_id
]);
550 bool success M5_VAR_USED
=
551 masterPorts
[dest_port_id
]->sendTimingSnoopResp(pkt
);
552 pktCount
[slave_port_id
][dest_port_id
]++;
553 pktSize
[slave_port_id
][dest_port_id
] += pkt_size
;
556 snoopLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
558 // we got a snoop response on one of our slave ports,
559 // i.e. from a coherent master connected to the crossbar, and
560 // since we created the snoop request as part of recvTiming,
561 // this should now be a normal response again
562 outstandingSnoop
.erase(pkt
->req
);
564 // this is a snoop response from a coherent master, hence it
565 // should never go back to where the snoop response came from,
566 // but instead to where the original request came from
567 assert(slave_port_id
!= dest_port_id
);
570 // update the probe filter so that it can properly track the line
571 snoopFilter
->updateSnoopResponse(pkt
, *slavePorts
[slave_port_id
],
572 *slavePorts
[dest_port_id
]);
575 DPRINTF(CoherentXBar
, "%s: src %s packet %s FWD RESP\n", __func__
,
576 src_port
->name(), pkt
->print());
578 // as a normal response, it should go back to a master through
579 // one of our slave ports, we also pay for any outstanding
581 Tick latency
= pkt
->headerDelay
;
582 pkt
->headerDelay
= 0;
583 slavePorts
[dest_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
585 respLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
588 // remove the request from the routing table
589 routeTo
.erase(route_lookup
);
592 transDist
[pkt_cmd
]++;
594 snoopTraffic
+= pkt_size
;
601 CoherentXBar::forwardTiming(PacketPtr pkt
, PortID exclude_slave_port_id
,
602 const std::vector
<QueuedSlavePort
*>& dests
)
604 DPRINTF(CoherentXBar
, "%s for %s\n", __func__
, pkt
->print());
606 // snoops should only happen if the system isn't bypassing caches
607 assert(!system
->bypassCaches());
611 for (const auto& p
: dests
) {
612 // we could have gotten this request from a snooping master
613 // (corresponding to our own slave port that is also in
614 // snoopPorts) and should not send it back to where it came
616 if (exclude_slave_port_id
== InvalidPortID
||
617 p
->getId() != exclude_slave_port_id
) {
618 // cache is not allowed to refuse snoop
619 p
->sendTimingSnoopReq(pkt
);
624 // Stats for fanout of this forward operation
625 snoopFanout
.sample(fanout
);
629 CoherentXBar::recvReqRetry(PortID master_port_id
)
631 // responses and snoop responses never block on forwarding them,
632 // so the retry will always be coming from a port to which we
633 // tried to forward a request
634 reqLayers
[master_port_id
]->recvRetry();
638 CoherentXBar::recvAtomic(PacketPtr pkt
, PortID slave_port_id
)
640 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
641 slavePorts
[slave_port_id
]->name(), pkt
->print());
643 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
644 unsigned int pkt_cmd
= pkt
->cmdToIndex();
646 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
647 Tick snoop_response_latency
= 0;
649 const bool snoop_caches
= !system
->bypassCaches() &&
650 pkt
->cmd
!= MemCmd::WriteClean
;
652 // forward to all snoopers but the source
653 std::pair
<MemCmd
, Tick
> snoop_result
;
655 // check with the snoop filter where to forward this packet
657 snoopFilter
->lookupRequest(pkt
, *slavePorts
[slave_port_id
]);
658 snoop_response_latency
+= sf_res
.second
* clockPeriod();
659 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
660 __func__
, slavePorts
[slave_port_id
]->name(), pkt
->print(),
661 sf_res
.first
.size(), sf_res
.second
);
663 // let the snoop filter know about the success of the send
664 // operation, and do it even before sending it onwards to
665 // avoid situations where atomic upward snoops sneak in
666 // between and change the filter state
667 snoopFilter
->finishRequest(false, pkt
->getAddr(), pkt
->isSecure());
669 if (pkt
->isEviction()) {
670 // for block-evicting packets, i.e. writebacks and
671 // clean evictions, there is no need to snoop up, as
672 // all we do is determine if the block is cached or
673 // not, instead just set it here based on the snoop
675 if (!sf_res
.first
.empty())
676 pkt
->setBlockCached();
678 snoop_result
= forwardAtomic(pkt
, slave_port_id
, InvalidPortID
,
682 snoop_result
= forwardAtomic(pkt
, slave_port_id
);
684 snoop_response_cmd
= snoop_result
.first
;
685 snoop_response_latency
+= snoop_result
.second
;
688 // set up a sensible default value
689 Tick response_latency
= 0;
691 const bool sink_packet
= sinkPacket(pkt
);
693 // even if we had a snoop response, we must continue and also
694 // perform the actual request at the destination
695 PortID master_port_id
= findPort(pkt
->getAddr());
698 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
701 if (!pointOfCoherency
|| pkt
->isRead() || pkt
->isWrite()) {
702 // forward the request to the appropriate destination
703 response_latency
= masterPorts
[master_port_id
]->sendAtomic(pkt
);
705 // if it does not need a response we sink the packet above
706 assert(pkt
->needsResponse());
712 // stats updates for the request
713 pktCount
[slave_port_id
][master_port_id
]++;
714 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
715 transDist
[pkt_cmd
]++;
718 // if lower levels have replied, tell the snoop filter
719 if (!system
->bypassCaches() && snoopFilter
&& pkt
->isResponse()) {
720 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
723 // if we got a response from a snooper, restore it here
724 if (snoop_response_cmd
!= MemCmd::InvalidCmd
) {
725 // no one else should have responded
726 assert(!pkt
->isResponse());
727 pkt
->cmd
= snoop_response_cmd
;
728 response_latency
= snoop_response_latency
;
731 // add the response data
732 if (pkt
->isResponse()) {
733 pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
734 pkt_cmd
= pkt
->cmdToIndex();
737 pktCount
[slave_port_id
][master_port_id
]++;
738 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
739 transDist
[pkt_cmd
]++;
742 // @todo: Not setting header time
743 pkt
->payloadDelay
= response_latency
;
744 return response_latency
;
748 CoherentXBar::recvAtomicSnoop(PacketPtr pkt
, PortID master_port_id
)
750 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
751 masterPorts
[master_port_id
]->name(), pkt
->print());
753 // add the request snoop data
754 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
756 snoopTraffic
+= pkt_size
;
758 // forward to all snoopers
759 std::pair
<MemCmd
, Tick
> snoop_result
;
760 Tick snoop_response_latency
= 0;
762 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
763 snoop_response_latency
+= sf_res
.second
* clockPeriod();
764 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
765 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
766 sf_res
.first
.size(), sf_res
.second
);
767 snoop_result
= forwardAtomic(pkt
, InvalidPortID
, master_port_id
,
770 snoop_result
= forwardAtomic(pkt
, InvalidPortID
);
772 MemCmd snoop_response_cmd
= snoop_result
.first
;
773 snoop_response_latency
+= snoop_result
.second
;
775 if (snoop_response_cmd
!= MemCmd::InvalidCmd
)
776 pkt
->cmd
= snoop_response_cmd
;
778 // add the response snoop data
779 if (pkt
->isResponse()) {
783 // @todo: Not setting header time
784 pkt
->payloadDelay
= snoop_response_latency
;
785 return snoop_response_latency
;
788 std::pair
<MemCmd
, Tick
>
789 CoherentXBar::forwardAtomic(PacketPtr pkt
, PortID exclude_slave_port_id
,
790 PortID source_master_port_id
,
791 const std::vector
<QueuedSlavePort
*>& dests
)
793 // the packet may be changed on snoops, record the original
794 // command to enable us to restore it between snoops so that
795 // additional snoops can take place properly
796 MemCmd orig_cmd
= pkt
->cmd
;
797 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
798 Tick snoop_response_latency
= 0;
800 // snoops should only happen if the system isn't bypassing caches
801 assert(!system
->bypassCaches());
805 for (const auto& p
: dests
) {
806 // we could have gotten this request from a snooping master
807 // (corresponding to our own slave port that is also in
808 // snoopPorts) and should not send it back to where it came
810 if (exclude_slave_port_id
!= InvalidPortID
&&
811 p
->getId() == exclude_slave_port_id
)
814 Tick latency
= p
->sendAtomicSnoop(pkt
);
817 // in contrast to a functional access, we have to keep on
818 // going as all snoopers must be updated even if we get a
820 if (!pkt
->isResponse())
823 // response from snoop agent
824 assert(pkt
->cmd
!= orig_cmd
);
825 assert(pkt
->cacheResponding());
826 // should only happen once
827 assert(snoop_response_cmd
== MemCmd::InvalidCmd
);
828 // save response state
829 snoop_response_cmd
= pkt
->cmd
;
830 snoop_response_latency
= latency
;
833 // Handle responses by the snoopers and differentiate between
834 // responses to requests from above and snoops from below
835 if (source_master_port_id
!= InvalidPortID
) {
836 // Getting a response for a snoop from below
837 assert(exclude_slave_port_id
== InvalidPortID
);
838 snoopFilter
->updateSnoopForward(pkt
, *p
,
839 *masterPorts
[source_master_port_id
]);
841 // Getting a response for a request from above
842 assert(source_master_port_id
== InvalidPortID
);
843 snoopFilter
->updateSnoopResponse(pkt
, *p
,
844 *slavePorts
[exclude_slave_port_id
]);
847 // restore original packet state for remaining snoopers
852 snoopFanout
.sample(fanout
);
854 // the packet is restored as part of the loop and any potential
855 // snoop response is part of the returned pair
856 return std::make_pair(snoop_response_cmd
, snoop_response_latency
);
860 CoherentXBar::recvFunctional(PacketPtr pkt
, PortID slave_port_id
)
862 if (!pkt
->isPrint()) {
863 // don't do DPRINTFs on PrintReq as it clutters up the output
864 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
865 slavePorts
[slave_port_id
]->name(), pkt
->print());
868 if (!system
->bypassCaches()) {
869 // forward to all snoopers but the source
870 forwardFunctional(pkt
, slave_port_id
);
873 // there is no need to continue if the snooping has found what we
874 // were looking for and the packet is already a response
875 if (!pkt
->isResponse()) {
876 // since our slave ports are queued ports we need to check them as well
877 for (const auto& p
: slavePorts
) {
878 // if we find a response that has the data, then the
879 // downstream caches/memories may be out of date, so simply stop
881 if (p
->checkFunctional(pkt
)) {
882 if (pkt
->needsResponse())
888 PortID dest_id
= findPort(pkt
->getAddr());
890 masterPorts
[dest_id
]->sendFunctional(pkt
);
895 CoherentXBar::recvFunctionalSnoop(PacketPtr pkt
, PortID master_port_id
)
897 if (!pkt
->isPrint()) {
898 // don't do DPRINTFs on PrintReq as it clutters up the output
899 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
900 masterPorts
[master_port_id
]->name(), pkt
->print());
903 for (const auto& p
: slavePorts
) {
904 if (p
->checkFunctional(pkt
)) {
905 if (pkt
->needsResponse())
911 // forward to all snoopers
912 forwardFunctional(pkt
, InvalidPortID
);
916 CoherentXBar::forwardFunctional(PacketPtr pkt
, PortID exclude_slave_port_id
)
918 // snoops should only happen if the system isn't bypassing caches
919 assert(!system
->bypassCaches());
921 for (const auto& p
: snoopPorts
) {
922 // we could have gotten this request from a snooping master
923 // (corresponding to our own slave port that is also in
924 // snoopPorts) and should not send it back to where it came
926 if (exclude_slave_port_id
== InvalidPortID
||
927 p
->getId() != exclude_slave_port_id
)
928 p
->sendFunctionalSnoop(pkt
);
930 // if we get a response we are done
931 if (pkt
->isResponse()) {
938 CoherentXBar::sinkPacket(const PacketPtr pkt
) const
940 // we can sink the packet if:
941 // 1) the crossbar is the point of coherency, and a cache is
942 // responding after being snooped
943 // 2) the crossbar is the point of coherency, and the packet is a
944 // coherency packet (not a read or a write) that does not
945 // require a response
946 // 3) this is a clean evict or clean writeback, but the packet is
947 // found in a cache above this crossbar
948 // 4) a cache is responding after being snooped, and the packet
949 // either does not need the block to be writable, or the cache
950 // that has promised to respond (setting the cache responding
951 // flag) is providing writable and thus had a Modified block,
952 // and no further action is needed
953 return (pointOfCoherency
&& pkt
->cacheResponding()) ||
954 (pointOfCoherency
&& !(pkt
->isRead() || pkt
->isWrite()) &&
955 !pkt
->needsResponse()) ||
956 (pkt
->isCleanEviction() && pkt
->isBlockCached()) ||
957 (pkt
->cacheResponding() &&
958 (!pkt
->needsWritable() || pkt
->responderHadWritable()));
962 CoherentXBar::regStats()
964 // register the stats of the base class and our layers
965 BaseXBar::regStats();
966 for (auto l
: reqLayers
)
968 for (auto l
: respLayers
)
970 for (auto l
: snoopLayers
)
974 .name(name() + ".snoops")
975 .desc("Total snoops (count)")
979 .name(name() + ".snoopTraffic")
980 .desc("Total snoop traffic (bytes)")
984 .init(0, snoopPorts
.size(), 1)
985 .name(name() + ".snoop_fanout")
986 .desc("Request fanout histogram")
991 CoherentXBarParams::create()
993 return new CoherentXBar(this);