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.
48 * Definition of a crossbar object.
51 #include "mem/coherent_xbar.hh"
53 #include "base/logging.hh"
54 #include "base/trace.hh"
55 #include "debug/AddrRanges.hh"
56 #include "debug/CoherentXBar.hh"
57 #include "sim/system.hh"
59 CoherentXBar::CoherentXBar(const CoherentXBarParams
*p
)
60 : BaseXBar(p
), system(p
->system
), snoopFilter(p
->snoop_filter
),
61 snoopResponseLatency(p
->snoop_response_latency
),
62 pointOfCoherency(p
->point_of_coherency
),
63 pointOfUnification(p
->point_of_unification
)
65 // create the ports based on the size of the master and slave
66 // vector ports, and the presence of the default port, the ports
67 // are enumerated starting from zero
68 for (int i
= 0; i
< p
->port_master_connection_count
; ++i
) {
69 std::string portName
= csprintf("%s.master[%d]", name(), i
);
70 MasterPort
* bp
= new CoherentXBarMasterPort(portName
, *this, i
);
71 masterPorts
.push_back(bp
);
72 reqLayers
.push_back(new ReqLayer(*bp
, *this,
73 csprintf(".reqLayer%d", i
)));
74 snoopLayers
.push_back(new SnoopRespLayer(*bp
, *this,
75 csprintf(".snoopLayer%d", i
)));
78 // see if we have a default slave device connected and if so add
79 // our corresponding master port
80 if (p
->port_default_connection_count
) {
81 defaultPortID
= masterPorts
.size();
82 std::string portName
= name() + ".default";
83 MasterPort
* bp
= new CoherentXBarMasterPort(portName
, *this,
85 masterPorts
.push_back(bp
);
86 reqLayers
.push_back(new ReqLayer(*bp
, *this, csprintf(".reqLayer%d",
88 snoopLayers
.push_back(new SnoopRespLayer(*bp
, *this,
89 csprintf(".snoopLayer%d",
93 // create the slave ports, once again starting at zero
94 for (int i
= 0; i
< p
->port_slave_connection_count
; ++i
) {
95 std::string portName
= csprintf("%s.slave[%d]", name(), i
);
96 QueuedSlavePort
* bp
= new CoherentXBarSlavePort(portName
, *this, i
);
97 slavePorts
.push_back(bp
);
98 respLayers
.push_back(new RespLayer(*bp
, *this,
99 csprintf(".respLayer%d", i
)));
100 snoopRespPorts
.push_back(new SnoopRespPort(*bp
, *this));
106 CoherentXBar::~CoherentXBar()
108 for (auto l
: reqLayers
)
110 for (auto l
: respLayers
)
112 for (auto l
: snoopLayers
)
114 for (auto p
: snoopRespPorts
)
123 // iterate over our slave ports and determine which of our
124 // neighbouring master ports are snooping and add them as snoopers
125 for (const auto& p
: slavePorts
) {
126 // check if the connected master port is snooping
127 if (p
->isSnooping()) {
128 DPRINTF(AddrRanges
, "Adding snooping master %s\n",
129 p
->getMasterPort().name());
130 snoopPorts
.push_back(p
);
134 if (snoopPorts
.empty())
135 warn("CoherentXBar %s has no snooping ports attached!\n", name());
137 // inform the snoop filter about the slave ports so it can create
138 // its own internal representation
140 snoopFilter
->setSlavePorts(slavePorts
);
144 CoherentXBar::recvTimingReq(PacketPtr pkt
, PortID slave_port_id
)
146 // determine the source port based on the id
147 SlavePort
*src_port
= slavePorts
[slave_port_id
];
149 // remember if the packet is an express snoop
150 bool is_express_snoop
= pkt
->isExpressSnoop();
151 bool cache_responding
= pkt
->cacheResponding();
152 // for normal requests, going downstream, the express snoop flag
153 // and the cache responding flag should always be the same
154 assert(is_express_snoop
== cache_responding
);
156 // determine the destination based on the address
157 PortID master_port_id
= findPort(pkt
->getAddr());
159 // test if the crossbar should be considered occupied for the current
160 // port, and exclude express snoops from the check
161 if (!is_express_snoop
&& !reqLayers
[master_port_id
]->tryTiming(src_port
)) {
162 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
163 src_port
->name(), pkt
->print());
167 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
168 src_port
->name(), pkt
->print());
170 // store size and command as they might be modified when
171 // forwarding the packet
172 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
173 unsigned int pkt_cmd
= pkt
->cmdToIndex();
175 // store the old header delay so we can restore it if needed
176 Tick old_header_delay
= pkt
->headerDelay
;
178 // a request sees the frontend and forward latency
179 Tick xbar_delay
= (frontendLatency
+ forwardLatency
) * clockPeriod();
181 // set the packet header and payload delay
182 calcPacketTiming(pkt
, xbar_delay
);
184 // determine how long to be crossbar layer is busy
185 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
187 // is this the destination point for this packet? (e.g. true if
188 // this xbar is the PoC for a cache maintenance operation to the
189 // PoC) otherwise the destination is any cache that can satisfy
191 const bool is_destination
= isDestination(pkt
);
193 const bool snoop_caches
= !system
->bypassCaches() &&
194 pkt
->cmd
!= MemCmd::WriteClean
;
196 assert(pkt
->snoopDelay
== 0);
198 if (pkt
->isClean() && !is_destination
) {
199 // before snooping we need to make sure that the memory
200 // below is not busy and the cache clean request can be
202 if (!masterPorts
[master_port_id
]->tryTiming(pkt
)) {
203 DPRINTF(CoherentXBar
, "%s: src %s packet %s RETRY\n", __func__
,
204 src_port
->name(), pkt
->print());
206 // update the layer state and schedule an idle event
207 reqLayers
[master_port_id
]->failedTiming(src_port
,
208 clockEdge(Cycles(1)));
214 // the packet is a memory-mapped request and should be
215 // broadcasted to our snoopers but the source
217 // check with the snoop filter where to forward this packet
218 auto sf_res
= snoopFilter
->lookupRequest(pkt
, *src_port
);
219 // the time required by a packet to be delivered through
220 // the xbar has to be charged also with to lookup latency
221 // of the snoop filter
222 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
223 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
224 __func__
, src_port
->name(), pkt
->print(),
225 sf_res
.first
.size(), sf_res
.second
);
227 if (pkt
->isEviction()) {
228 // for block-evicting packets, i.e. writebacks and
229 // clean evictions, there is no need to snoop up, as
230 // all we do is determine if the block is cached or
231 // not, instead just set it here based on the snoop
233 if (!sf_res
.first
.empty())
234 pkt
->setBlockCached();
236 forwardTiming(pkt
, slave_port_id
, sf_res
.first
);
239 forwardTiming(pkt
, slave_port_id
);
242 // add the snoop delay to our header delay, and then reset it
243 pkt
->headerDelay
+= pkt
->snoopDelay
;
247 // set up a sensible starting point
250 // remember if the packet will generate a snoop response by
251 // checking if a cache set the cacheResponding flag during the
253 const bool expect_snoop_resp
= !cache_responding
&& pkt
->cacheResponding();
254 bool expect_response
= pkt
->needsResponse() && !pkt
->cacheResponding();
256 const bool sink_packet
= sinkPacket(pkt
);
258 // in certain cases the crossbar is responsible for responding
259 bool respond_directly
= false;
260 // store the original address as an address mapper could possibly
261 // modify the address upon a sendTimingRequest
262 const Addr
addr(pkt
->getAddr());
264 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
267 // determine if we are forwarding the packet, or responding to
269 if (forwardPacket(pkt
)) {
270 // if we are passing on, rather than sinking, a packet to
271 // which an upstream cache has committed to responding,
272 // the line was needs writable, and the responding only
273 // had an Owned copy, so we need to immidiately let the
274 // downstream caches know, bypass any flow control
275 if (pkt
->cacheResponding()) {
276 pkt
->setExpressSnoop();
279 // make sure that the write request (e.g., WriteClean)
280 // will stop at the memory below if this crossbar is its
282 if (pkt
->isWrite() && is_destination
) {
283 pkt
->clearWriteThrough();
286 // since it is a normal request, attempt to send the packet
287 success
= masterPorts
[master_port_id
]->sendTimingReq(pkt
);
289 // no need to forward, turn this packet around and respond
291 assert(pkt
->needsResponse());
293 respond_directly
= true;
294 assert(!expect_snoop_resp
);
295 expect_response
= false;
299 if (snoopFilter
&& snoop_caches
) {
300 // Let the snoop filter know about the success of the send operation
301 snoopFilter
->finishRequest(!success
, addr
, pkt
->isSecure());
304 // check if we were successful in sending the packet onwards
306 // express snoops should never be forced to retry
307 assert(!is_express_snoop
);
309 // restore the header delay
310 pkt
->headerDelay
= old_header_delay
;
312 DPRINTF(CoherentXBar
, "%s: src %s packet %s RETRY\n", __func__
,
313 src_port
->name(), pkt
->print());
315 // update the layer state and schedule an idle event
316 reqLayers
[master_port_id
]->failedTiming(src_port
,
317 clockEdge(Cycles(1)));
319 // express snoops currently bypass the crossbar state entirely
320 if (!is_express_snoop
) {
321 // if this particular request will generate a snoop
323 if (expect_snoop_resp
) {
324 // we should never have an exsiting request outstanding
325 assert(outstandingSnoop
.find(pkt
->req
) ==
326 outstandingSnoop
.end());
327 outstandingSnoop
.insert(pkt
->req
);
329 // basic sanity check on the outstanding snoops
330 panic_if(outstandingSnoop
.size() > 512,
331 "Outstanding snoop requests exceeded 512\n");
334 // remember where to route the normal response to
335 if (expect_response
|| expect_snoop_resp
) {
336 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
337 routeTo
[pkt
->req
] = slave_port_id
;
339 panic_if(routeTo
.size() > 512,
340 "Routing table exceeds 512 packets\n");
343 // update the layer state and schedule an idle event
344 reqLayers
[master_port_id
]->succeededTiming(packetFinishTime
);
347 // stats updates only consider packets that were successfully sent
348 pktCount
[slave_port_id
][master_port_id
]++;
349 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
350 transDist
[pkt_cmd
]++;
352 if (is_express_snoop
) {
354 snoopTraffic
+= pkt_size
;
359 // queue the packet for deletion
360 pendingDelete
.reset(pkt
);
362 // normally we respond to the packet we just received if we need to
363 PacketPtr rsp_pkt
= pkt
;
364 PortID rsp_port_id
= slave_port_id
;
366 // If this is the destination of the cache clean operation the
367 // crossbar is responsible for responding. This crossbar will
368 // respond when the cache clean is complete. A cache clean
369 // is complete either:
370 // * direcly, if no cache above had a dirty copy of the block
371 // as indicated by the satisfied flag of the packet, or
372 // * when the crossbar has seen both the cache clean request
373 // (CleanSharedReq, CleanInvalidReq) and the corresponding
374 // write (WriteClean) which updates the block in the memory
377 ((pkt
->isClean() && pkt
->satisfied()) ||
378 pkt
->cmd
== MemCmd::WriteClean
) &&
380 PacketPtr deferred_rsp
= pkt
->isWrite() ? nullptr : pkt
;
381 auto cmo_lookup
= outstandingCMO
.find(pkt
->id
);
382 if (cmo_lookup
!= outstandingCMO
.end()) {
383 // the cache clean request has already reached this xbar
384 respond_directly
= true;
385 if (pkt
->isWrite()) {
386 rsp_pkt
= cmo_lookup
->second
;
389 // determine the destination
390 const auto route_lookup
= routeTo
.find(rsp_pkt
->req
);
391 assert(route_lookup
!= routeTo
.end());
392 rsp_port_id
= route_lookup
->second
;
393 assert(rsp_port_id
!= InvalidPortID
);
394 assert(rsp_port_id
< respLayers
.size());
395 // remove the request from the routing table
396 routeTo
.erase(route_lookup
);
398 outstandingCMO
.erase(cmo_lookup
);
400 respond_directly
= false;
401 outstandingCMO
.emplace(pkt
->id
, deferred_rsp
);
402 if (!pkt
->isWrite()) {
403 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
404 routeTo
[pkt
->req
] = slave_port_id
;
406 panic_if(routeTo
.size() > 512,
407 "Routing table exceeds 512 packets\n");
413 if (respond_directly
) {
414 assert(rsp_pkt
->needsResponse());
417 rsp_pkt
->makeResponse();
419 if (snoopFilter
&& !system
->bypassCaches()) {
420 // let the snoop filter inspect the response and update its state
421 snoopFilter
->updateResponse(rsp_pkt
, *slavePorts
[rsp_port_id
]);
424 // we send the response after the current packet, even if the
425 // response is not for this packet (e.g. cache clean operation
426 // where both the request and the write packet have to cross
427 // the destination xbar before the response is sent.)
428 Tick response_time
= clockEdge() + pkt
->headerDelay
;
429 rsp_pkt
->headerDelay
= 0;
431 slavePorts
[rsp_port_id
]->schedTimingResp(rsp_pkt
, response_time
);
438 CoherentXBar::recvTimingResp(PacketPtr pkt
, PortID master_port_id
)
440 // determine the source port based on the id
441 MasterPort
*src_port
= masterPorts
[master_port_id
];
443 // determine the destination
444 const auto route_lookup
= routeTo
.find(pkt
->req
);
445 assert(route_lookup
!= routeTo
.end());
446 const PortID slave_port_id
= route_lookup
->second
;
447 assert(slave_port_id
!= InvalidPortID
);
448 assert(slave_port_id
< respLayers
.size());
450 // test if the crossbar should be considered occupied for the
452 if (!respLayers
[slave_port_id
]->tryTiming(src_port
)) {
453 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
454 src_port
->name(), pkt
->print());
458 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
459 src_port
->name(), pkt
->print());
461 // store size and command as they might be modified when
462 // forwarding the packet
463 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
464 unsigned int pkt_cmd
= pkt
->cmdToIndex();
466 // a response sees the response latency
467 Tick xbar_delay
= responseLatency
* clockPeriod();
469 // set the packet header and payload delay
470 calcPacketTiming(pkt
, xbar_delay
);
472 // determine how long to be crossbar layer is busy
473 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
475 if (snoopFilter
&& !system
->bypassCaches()) {
476 // let the snoop filter inspect the response and update its state
477 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
480 // send the packet through the destination slave port and pay for
481 // any outstanding header delay
482 Tick latency
= pkt
->headerDelay
;
483 pkt
->headerDelay
= 0;
484 slavePorts
[slave_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
486 // remove the request from the routing table
487 routeTo
.erase(route_lookup
);
489 respLayers
[slave_port_id
]->succeededTiming(packetFinishTime
);
492 pktCount
[slave_port_id
][master_port_id
]++;
493 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
494 transDist
[pkt_cmd
]++;
500 CoherentXBar::recvTimingSnoopReq(PacketPtr pkt
, PortID master_port_id
)
502 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
503 masterPorts
[master_port_id
]->name(), pkt
->print());
505 // update stats here as we know the forwarding will succeed
506 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
507 transDist
[pkt
->cmdToIndex()]++;
509 snoopTraffic
+= pkt_size
;
511 // we should only see express snoops from caches
512 assert(pkt
->isExpressSnoop());
514 // set the packet header and payload delay, for now use forward latency
515 // @todo Assess the choice of latency further
516 calcPacketTiming(pkt
, forwardLatency
* clockPeriod());
518 // remember if a cache has already committed to responding so we
519 // can see if it changes during the snooping
520 const bool cache_responding
= pkt
->cacheResponding();
522 assert(pkt
->snoopDelay
== 0);
525 // let the Snoop Filter work its magic and guide probing
526 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
527 // the time required by a packet to be delivered through
528 // the xbar has to be charged also with to lookup latency
529 // of the snoop filter
530 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
531 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
532 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
533 sf_res
.first
.size(), sf_res
.second
);
535 // forward to all snoopers
536 forwardTiming(pkt
, InvalidPortID
, sf_res
.first
);
538 forwardTiming(pkt
, InvalidPortID
);
541 // add the snoop delay to our header delay, and then reset it
542 pkt
->headerDelay
+= pkt
->snoopDelay
;
545 // if we can expect a response, remember how to route it
546 if (!cache_responding
&& pkt
->cacheResponding()) {
547 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
548 routeTo
[pkt
->req
] = master_port_id
;
551 // a snoop request came from a connected slave device (one of
552 // our master ports), and if it is not coming from the slave
553 // device responsible for the address range something is
554 // wrong, hence there is nothing further to do as the packet
555 // would be going back to where it came from
556 assert(master_port_id
== findPort(pkt
->getAddr()));
560 CoherentXBar::recvTimingSnoopResp(PacketPtr pkt
, PortID slave_port_id
)
562 // determine the source port based on the id
563 SlavePort
* src_port
= slavePorts
[slave_port_id
];
565 // get the destination
566 const auto route_lookup
= routeTo
.find(pkt
->req
);
567 assert(route_lookup
!= routeTo
.end());
568 const PortID dest_port_id
= route_lookup
->second
;
569 assert(dest_port_id
!= InvalidPortID
);
571 // determine if the response is from a snoop request we
572 // created as the result of a normal request (in which case it
573 // should be in the outstandingSnoop), or if we merely forwarded
574 // someone else's snoop request
575 const bool forwardAsSnoop
= outstandingSnoop
.find(pkt
->req
) ==
576 outstandingSnoop
.end();
578 // test if the crossbar should be considered occupied for the
579 // current port, note that the check is bypassed if the response
580 // is being passed on as a normal response since this is occupying
581 // the response layer rather than the snoop response layer
582 if (forwardAsSnoop
) {
583 assert(dest_port_id
< snoopLayers
.size());
584 if (!snoopLayers
[dest_port_id
]->tryTiming(src_port
)) {
585 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
586 src_port
->name(), pkt
->print());
590 // get the master port that mirrors this slave port internally
591 MasterPort
* snoop_port
= snoopRespPorts
[slave_port_id
];
592 assert(dest_port_id
< respLayers
.size());
593 if (!respLayers
[dest_port_id
]->tryTiming(snoop_port
)) {
594 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
595 snoop_port
->name(), pkt
->print());
600 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
601 src_port
->name(), pkt
->print());
603 // store size and command as they might be modified when
604 // forwarding the packet
605 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
606 unsigned int pkt_cmd
= pkt
->cmdToIndex();
608 // responses are never express snoops
609 assert(!pkt
->isExpressSnoop());
611 // a snoop response sees the snoop response latency, and if it is
612 // forwarded as a normal response, the response latency
614 (forwardAsSnoop
? snoopResponseLatency
: responseLatency
) *
617 // set the packet header and payload delay
618 calcPacketTiming(pkt
, xbar_delay
);
620 // determine how long to be crossbar layer is busy
621 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
623 // forward it either as a snoop response or a normal response
624 if (forwardAsSnoop
) {
625 // this is a snoop response to a snoop request we forwarded,
626 // e.g. coming from the L1 and going to the L2, and it should
627 // be forwarded as a snoop response
630 // update the probe filter so that it can properly track the line
631 snoopFilter
->updateSnoopForward(pkt
, *slavePorts
[slave_port_id
],
632 *masterPorts
[dest_port_id
]);
635 bool success M5_VAR_USED
=
636 masterPorts
[dest_port_id
]->sendTimingSnoopResp(pkt
);
637 pktCount
[slave_port_id
][dest_port_id
]++;
638 pktSize
[slave_port_id
][dest_port_id
] += pkt_size
;
641 snoopLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
643 // we got a snoop response on one of our slave ports,
644 // i.e. from a coherent master connected to the crossbar, and
645 // since we created the snoop request as part of recvTiming,
646 // this should now be a normal response again
647 outstandingSnoop
.erase(pkt
->req
);
649 // this is a snoop response from a coherent master, hence it
650 // should never go back to where the snoop response came from,
651 // but instead to where the original request came from
652 assert(slave_port_id
!= dest_port_id
);
655 // update the probe filter so that it can properly track the line
656 snoopFilter
->updateSnoopResponse(pkt
, *slavePorts
[slave_port_id
],
657 *slavePorts
[dest_port_id
]);
660 DPRINTF(CoherentXBar
, "%s: src %s packet %s FWD RESP\n", __func__
,
661 src_port
->name(), pkt
->print());
663 // as a normal response, it should go back to a master through
664 // one of our slave ports, we also pay for any outstanding
666 Tick latency
= pkt
->headerDelay
;
667 pkt
->headerDelay
= 0;
668 slavePorts
[dest_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
670 respLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
673 // remove the request from the routing table
674 routeTo
.erase(route_lookup
);
677 transDist
[pkt_cmd
]++;
679 snoopTraffic
+= pkt_size
;
686 CoherentXBar::forwardTiming(PacketPtr pkt
, PortID exclude_slave_port_id
,
687 const std::vector
<QueuedSlavePort
*>& dests
)
689 DPRINTF(CoherentXBar
, "%s for %s\n", __func__
, pkt
->print());
691 // snoops should only happen if the system isn't bypassing caches
692 assert(!system
->bypassCaches());
696 for (const auto& p
: dests
) {
697 // we could have gotten this request from a snooping master
698 // (corresponding to our own slave port that is also in
699 // snoopPorts) and should not send it back to where it came
701 if (exclude_slave_port_id
== InvalidPortID
||
702 p
->getId() != exclude_slave_port_id
) {
703 // cache is not allowed to refuse snoop
704 p
->sendTimingSnoopReq(pkt
);
709 // Stats for fanout of this forward operation
710 snoopFanout
.sample(fanout
);
714 CoherentXBar::recvReqRetry(PortID master_port_id
)
716 // responses and snoop responses never block on forwarding them,
717 // so the retry will always be coming from a port to which we
718 // tried to forward a request
719 reqLayers
[master_port_id
]->recvRetry();
723 CoherentXBar::recvAtomic(PacketPtr pkt
, PortID slave_port_id
)
725 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
726 slavePorts
[slave_port_id
]->name(), pkt
->print());
728 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
729 unsigned int pkt_cmd
= pkt
->cmdToIndex();
731 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
732 Tick snoop_response_latency
= 0;
734 // is this the destination point for this packet? (e.g. true if
735 // this xbar is the PoC for a cache maintenance operation to the
736 // PoC) otherwise the destination is any cache that can satisfy
738 const bool is_destination
= isDestination(pkt
);
740 const bool snoop_caches
= !system
->bypassCaches() &&
741 pkt
->cmd
!= MemCmd::WriteClean
;
743 // forward to all snoopers but the source
744 std::pair
<MemCmd
, Tick
> snoop_result
;
746 // check with the snoop filter where to forward this packet
748 snoopFilter
->lookupRequest(pkt
, *slavePorts
[slave_port_id
]);
749 snoop_response_latency
+= sf_res
.second
* clockPeriod();
750 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
751 __func__
, slavePorts
[slave_port_id
]->name(), pkt
->print(),
752 sf_res
.first
.size(), sf_res
.second
);
754 // let the snoop filter know about the success of the send
755 // operation, and do it even before sending it onwards to
756 // avoid situations where atomic upward snoops sneak in
757 // between and change the filter state
758 snoopFilter
->finishRequest(false, pkt
->getAddr(), pkt
->isSecure());
760 if (pkt
->isEviction()) {
761 // for block-evicting packets, i.e. writebacks and
762 // clean evictions, there is no need to snoop up, as
763 // all we do is determine if the block is cached or
764 // not, instead just set it here based on the snoop
766 if (!sf_res
.first
.empty())
767 pkt
->setBlockCached();
769 snoop_result
= forwardAtomic(pkt
, slave_port_id
, InvalidPortID
,
773 snoop_result
= forwardAtomic(pkt
, slave_port_id
);
775 snoop_response_cmd
= snoop_result
.first
;
776 snoop_response_latency
+= snoop_result
.second
;
779 // set up a sensible default value
780 Tick response_latency
= 0;
782 const bool sink_packet
= sinkPacket(pkt
);
784 // even if we had a snoop response, we must continue and also
785 // perform the actual request at the destination
786 PortID master_port_id
= findPort(pkt
->getAddr());
789 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
792 if (forwardPacket(pkt
)) {
793 // make sure that the write request (e.g., WriteClean)
794 // will stop at the memory below if this crossbar is its
796 if (pkt
->isWrite() && is_destination
) {
797 pkt
->clearWriteThrough();
800 // forward the request to the appropriate destination
801 response_latency
= masterPorts
[master_port_id
]->sendAtomic(pkt
);
803 // if it does not need a response we sink the packet above
804 assert(pkt
->needsResponse());
810 // stats updates for the request
811 pktCount
[slave_port_id
][master_port_id
]++;
812 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
813 transDist
[pkt_cmd
]++;
816 // if lower levels have replied, tell the snoop filter
817 if (!system
->bypassCaches() && snoopFilter
&& pkt
->isResponse()) {
818 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
821 // if we got a response from a snooper, restore it here
822 if (snoop_response_cmd
!= MemCmd::InvalidCmd
) {
823 // no one else should have responded
824 assert(!pkt
->isResponse());
825 pkt
->cmd
= snoop_response_cmd
;
826 response_latency
= snoop_response_latency
;
829 // If this is the destination of the cache clean operation the
830 // crossbar is responsible for responding. This crossbar will
831 // respond when the cache clean is complete. An atomic cache clean
832 // is complete when the crossbars receives the cache clean
833 // request (CleanSharedReq, CleanInvalidReq), as either:
834 // * no cache above had a dirty copy of the block as indicated by
835 // the satisfied flag of the packet, or
836 // * the crossbar has already seen the corresponding write
837 // (WriteClean) which updates the block in the memory below.
838 if (pkt
->isClean() && isDestination(pkt
) && pkt
->satisfied()) {
839 auto it
= outstandingCMO
.find(pkt
->id
);
840 assert(it
!= outstandingCMO
.end());
841 // we are responding right away
842 outstandingCMO
.erase(it
);
843 } else if (pkt
->cmd
== MemCmd::WriteClean
&& isDestination(pkt
)) {
844 // if this is the destination of the operation, the xbar
845 // sends the responce to the cache clean operation only
846 // after having encountered the cache clean request
847 auto M5_VAR_USED ret
= outstandingCMO
.emplace(pkt
->id
, nullptr);
848 // in atomic mode we know that the WriteClean packet should
849 // precede the clean request
853 // add the response data
854 if (pkt
->isResponse()) {
855 pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
856 pkt_cmd
= pkt
->cmdToIndex();
859 pktCount
[slave_port_id
][master_port_id
]++;
860 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
861 transDist
[pkt_cmd
]++;
864 // @todo: Not setting header time
865 pkt
->payloadDelay
= response_latency
;
866 return response_latency
;
870 CoherentXBar::recvAtomicSnoop(PacketPtr pkt
, PortID master_port_id
)
872 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
873 masterPorts
[master_port_id
]->name(), pkt
->print());
875 // add the request snoop data
876 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
878 snoopTraffic
+= pkt_size
;
880 // forward to all snoopers
881 std::pair
<MemCmd
, Tick
> snoop_result
;
882 Tick snoop_response_latency
= 0;
884 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
885 snoop_response_latency
+= sf_res
.second
* clockPeriod();
886 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
887 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
888 sf_res
.first
.size(), sf_res
.second
);
889 snoop_result
= forwardAtomic(pkt
, InvalidPortID
, master_port_id
,
892 snoop_result
= forwardAtomic(pkt
, InvalidPortID
);
894 MemCmd snoop_response_cmd
= snoop_result
.first
;
895 snoop_response_latency
+= snoop_result
.second
;
897 if (snoop_response_cmd
!= MemCmd::InvalidCmd
)
898 pkt
->cmd
= snoop_response_cmd
;
900 // add the response snoop data
901 if (pkt
->isResponse()) {
905 // @todo: Not setting header time
906 pkt
->payloadDelay
= snoop_response_latency
;
907 return snoop_response_latency
;
910 std::pair
<MemCmd
, Tick
>
911 CoherentXBar::forwardAtomic(PacketPtr pkt
, PortID exclude_slave_port_id
,
912 PortID source_master_port_id
,
913 const std::vector
<QueuedSlavePort
*>& dests
)
915 // the packet may be changed on snoops, record the original
916 // command to enable us to restore it between snoops so that
917 // additional snoops can take place properly
918 MemCmd orig_cmd
= pkt
->cmd
;
919 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
920 Tick snoop_response_latency
= 0;
922 // snoops should only happen if the system isn't bypassing caches
923 assert(!system
->bypassCaches());
927 for (const auto& p
: dests
) {
928 // we could have gotten this request from a snooping master
929 // (corresponding to our own slave port that is also in
930 // snoopPorts) and should not send it back to where it came
932 if (exclude_slave_port_id
!= InvalidPortID
&&
933 p
->getId() == exclude_slave_port_id
)
936 Tick latency
= p
->sendAtomicSnoop(pkt
);
939 // in contrast to a functional access, we have to keep on
940 // going as all snoopers must be updated even if we get a
942 if (!pkt
->isResponse())
945 // response from snoop agent
946 assert(pkt
->cmd
!= orig_cmd
);
947 assert(pkt
->cacheResponding());
948 // should only happen once
949 assert(snoop_response_cmd
== MemCmd::InvalidCmd
);
950 // save response state
951 snoop_response_cmd
= pkt
->cmd
;
952 snoop_response_latency
= latency
;
955 // Handle responses by the snoopers and differentiate between
956 // responses to requests from above and snoops from below
957 if (source_master_port_id
!= InvalidPortID
) {
958 // Getting a response for a snoop from below
959 assert(exclude_slave_port_id
== InvalidPortID
);
960 snoopFilter
->updateSnoopForward(pkt
, *p
,
961 *masterPorts
[source_master_port_id
]);
963 // Getting a response for a request from above
964 assert(source_master_port_id
== InvalidPortID
);
965 snoopFilter
->updateSnoopResponse(pkt
, *p
,
966 *slavePorts
[exclude_slave_port_id
]);
969 // restore original packet state for remaining snoopers
974 snoopFanout
.sample(fanout
);
976 // the packet is restored as part of the loop and any potential
977 // snoop response is part of the returned pair
978 return std::make_pair(snoop_response_cmd
, snoop_response_latency
);
982 CoherentXBar::recvFunctional(PacketPtr pkt
, PortID slave_port_id
)
984 if (!pkt
->isPrint()) {
985 // don't do DPRINTFs on PrintReq as it clutters up the output
986 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
987 slavePorts
[slave_port_id
]->name(), pkt
->print());
990 if (!system
->bypassCaches()) {
991 // forward to all snoopers but the source
992 forwardFunctional(pkt
, slave_port_id
);
995 // there is no need to continue if the snooping has found what we
996 // were looking for and the packet is already a response
997 if (!pkt
->isResponse()) {
998 // since our slave ports are queued ports we need to check them as well
999 for (const auto& p
: slavePorts
) {
1000 // if we find a response that has the data, then the
1001 // downstream caches/memories may be out of date, so simply stop
1003 if (p
->checkFunctional(pkt
)) {
1004 if (pkt
->needsResponse())
1005 pkt
->makeResponse();
1010 PortID dest_id
= findPort(pkt
->getAddr());
1012 masterPorts
[dest_id
]->sendFunctional(pkt
);
1017 CoherentXBar::recvFunctionalSnoop(PacketPtr pkt
, PortID master_port_id
)
1019 if (!pkt
->isPrint()) {
1020 // don't do DPRINTFs on PrintReq as it clutters up the output
1021 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
1022 masterPorts
[master_port_id
]->name(), pkt
->print());
1025 for (const auto& p
: slavePorts
) {
1026 if (p
->checkFunctional(pkt
)) {
1027 if (pkt
->needsResponse())
1028 pkt
->makeResponse();
1033 // forward to all snoopers
1034 forwardFunctional(pkt
, InvalidPortID
);
1038 CoherentXBar::forwardFunctional(PacketPtr pkt
, PortID exclude_slave_port_id
)
1040 // snoops should only happen if the system isn't bypassing caches
1041 assert(!system
->bypassCaches());
1043 for (const auto& p
: snoopPorts
) {
1044 // we could have gotten this request from a snooping master
1045 // (corresponding to our own slave port that is also in
1046 // snoopPorts) and should not send it back to where it came
1048 if (exclude_slave_port_id
== InvalidPortID
||
1049 p
->getId() != exclude_slave_port_id
)
1050 p
->sendFunctionalSnoop(pkt
);
1052 // if we get a response we are done
1053 if (pkt
->isResponse()) {
1060 CoherentXBar::sinkPacket(const PacketPtr pkt
) const
1062 // we can sink the packet if:
1063 // 1) the crossbar is the point of coherency, and a cache is
1064 // responding after being snooped
1065 // 2) the crossbar is the point of coherency, and the packet is a
1066 // coherency packet (not a read or a write) that does not
1067 // require a response
1068 // 3) this is a clean evict or clean writeback, but the packet is
1069 // found in a cache above this crossbar
1070 // 4) a cache is responding after being snooped, and the packet
1071 // either does not need the block to be writable, or the cache
1072 // that has promised to respond (setting the cache responding
1073 // flag) is providing writable and thus had a Modified block,
1074 // and no further action is needed
1075 return (pointOfCoherency
&& pkt
->cacheResponding()) ||
1076 (pointOfCoherency
&& !(pkt
->isRead() || pkt
->isWrite()) &&
1077 !pkt
->needsResponse()) ||
1078 (pkt
->isCleanEviction() && pkt
->isBlockCached()) ||
1079 (pkt
->cacheResponding() &&
1080 (!pkt
->needsWritable() || pkt
->responderHadWritable()));
1084 CoherentXBar::forwardPacket(const PacketPtr pkt
)
1086 // we are forwarding the packet if:
1087 // 1) this is a cache clean request to the PoU/PoC and this
1088 // crossbar is above the PoU/PoC
1089 // 2) this is a read or a write
1090 // 3) this crossbar is above the point of coherency
1091 if (pkt
->isClean()) {
1092 return !isDestination(pkt
);
1094 return pkt
->isRead() || pkt
->isWrite() || !pointOfCoherency
;
1099 CoherentXBar::regStats()
1101 // register the stats of the base class and our layers
1102 BaseXBar::regStats();
1103 for (auto l
: reqLayers
)
1105 for (auto l
: respLayers
)
1107 for (auto l
: snoopLayers
)
1111 .name(name() + ".snoops")
1112 .desc("Total snoops (count)")
1116 .name(name() + ".snoopTraffic")
1117 .desc("Total snoop traffic (bytes)")
1121 .init(0, snoopPorts
.size(), 1)
1122 .name(name() + ".snoop_fanout")
1123 .desc("Request fanout histogram")
1128 CoherentXBarParams::create()
1130 return new CoherentXBar(this);