2 * Copyright (c) 2011-2018 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(
75 new SnoopRespLayer(*bp
, *this, 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));
104 CoherentXBar::~CoherentXBar()
106 for (auto l
: reqLayers
)
108 for (auto l
: respLayers
)
110 for (auto l
: snoopLayers
)
112 for (auto p
: snoopRespPorts
)
121 // iterate over our slave ports and determine which of our
122 // neighbouring master ports are snooping and add them as snoopers
123 for (const auto& p
: slavePorts
) {
124 // check if the connected master port is snooping
125 if (p
->isSnooping()) {
126 DPRINTF(AddrRanges
, "Adding snooping master %s\n",
127 p
->getMasterPort().name());
128 snoopPorts
.push_back(p
);
132 if (snoopPorts
.empty())
133 warn("CoherentXBar %s has no snooping ports attached!\n", name());
135 // inform the snoop filter about the slave ports so it can create
136 // its own internal representation
138 snoopFilter
->setSlavePorts(slavePorts
);
142 CoherentXBar::recvTimingReq(PacketPtr pkt
, PortID slave_port_id
)
144 // determine the source port based on the id
145 SlavePort
*src_port
= slavePorts
[slave_port_id
];
147 // remember if the packet is an express snoop
148 bool is_express_snoop
= pkt
->isExpressSnoop();
149 bool cache_responding
= pkt
->cacheResponding();
150 // for normal requests, going downstream, the express snoop flag
151 // and the cache responding flag should always be the same
152 assert(is_express_snoop
== cache_responding
);
154 // determine the destination based on the destination address range
155 PortID master_port_id
= findPort(pkt
->getAddrRange());
157 // test if the crossbar should be considered occupied for the current
158 // port, and exclude express snoops from the check
159 if (!is_express_snoop
&& !reqLayers
[master_port_id
]->tryTiming(src_port
)) {
160 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
161 src_port
->name(), pkt
->print());
165 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
166 src_port
->name(), pkt
->print());
168 // store size and command as they might be modified when
169 // forwarding the packet
170 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
171 unsigned int pkt_cmd
= pkt
->cmdToIndex();
173 // store the old header delay so we can restore it if needed
174 Tick old_header_delay
= pkt
->headerDelay
;
176 // a request sees the frontend and forward latency
177 Tick xbar_delay
= (frontendLatency
+ forwardLatency
) * clockPeriod();
179 // set the packet header and payload delay
180 calcPacketTiming(pkt
, xbar_delay
);
182 // determine how long to be crossbar layer is busy
183 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
185 // is this the destination point for this packet? (e.g. true if
186 // this xbar is the PoC for a cache maintenance operation to the
187 // PoC) otherwise the destination is any cache that can satisfy
189 const bool is_destination
= isDestination(pkt
);
191 const bool snoop_caches
= !system
->bypassCaches() &&
192 pkt
->cmd
!= MemCmd::WriteClean
;
194 assert(pkt
->snoopDelay
== 0);
196 if (pkt
->isClean() && !is_destination
) {
197 // before snooping we need to make sure that the memory
198 // below is not busy and the cache clean request can be
200 if (!masterPorts
[master_port_id
]->tryTiming(pkt
)) {
201 DPRINTF(CoherentXBar
, "%s: src %s packet %s RETRY\n", __func__
,
202 src_port
->name(), pkt
->print());
204 // update the layer state and schedule an idle event
205 reqLayers
[master_port_id
]->failedTiming(src_port
,
206 clockEdge(Cycles(1)));
212 // the packet is a memory-mapped request and should be
213 // broadcasted to our snoopers but the source
215 // check with the snoop filter where to forward this packet
216 auto sf_res
= snoopFilter
->lookupRequest(pkt
, *src_port
);
217 // the time required by a packet to be delivered through
218 // the xbar has to be charged also with to lookup latency
219 // of the snoop filter
220 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
221 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
222 __func__
, src_port
->name(), pkt
->print(),
223 sf_res
.first
.size(), sf_res
.second
);
225 if (pkt
->isEviction()) {
226 // for block-evicting packets, i.e. writebacks and
227 // clean evictions, there is no need to snoop up, as
228 // all we do is determine if the block is cached or
229 // not, instead just set it here based on the snoop
231 if (!sf_res
.first
.empty())
232 pkt
->setBlockCached();
234 forwardTiming(pkt
, slave_port_id
, sf_res
.first
);
237 forwardTiming(pkt
, slave_port_id
);
240 // add the snoop delay to our header delay, and then reset it
241 pkt
->headerDelay
+= pkt
->snoopDelay
;
245 // set up a sensible starting point
248 // remember if the packet will generate a snoop response by
249 // checking if a cache set the cacheResponding flag during the
251 const bool expect_snoop_resp
= !cache_responding
&& pkt
->cacheResponding();
252 bool expect_response
= pkt
->needsResponse() && !pkt
->cacheResponding();
254 const bool sink_packet
= sinkPacket(pkt
);
256 // in certain cases the crossbar is responsible for responding
257 bool respond_directly
= false;
258 // store the original address as an address mapper could possibly
259 // modify the address upon a sendTimingRequest
260 const Addr
addr(pkt
->getAddr());
262 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
265 // determine if we are forwarding the packet, or responding to
267 if (forwardPacket(pkt
)) {
268 // if we are passing on, rather than sinking, a packet to
269 // which an upstream cache has committed to responding,
270 // the line was needs writable, and the responding only
271 // had an Owned copy, so we need to immidiately let the
272 // downstream caches know, bypass any flow control
273 if (pkt
->cacheResponding()) {
274 pkt
->setExpressSnoop();
277 // make sure that the write request (e.g., WriteClean)
278 // will stop at the memory below if this crossbar is its
280 if (pkt
->isWrite() && is_destination
) {
281 pkt
->clearWriteThrough();
284 // since it is a normal request, attempt to send the packet
285 success
= masterPorts
[master_port_id
]->sendTimingReq(pkt
);
287 // no need to forward, turn this packet around and respond
289 assert(pkt
->needsResponse());
291 respond_directly
= true;
292 assert(!expect_snoop_resp
);
293 expect_response
= false;
297 if (snoopFilter
&& snoop_caches
) {
298 // Let the snoop filter know about the success of the send operation
299 snoopFilter
->finishRequest(!success
, addr
, pkt
->isSecure());
302 // check if we were successful in sending the packet onwards
304 // express snoops should never be forced to retry
305 assert(!is_express_snoop
);
307 // restore the header delay
308 pkt
->headerDelay
= old_header_delay
;
310 DPRINTF(CoherentXBar
, "%s: src %s packet %s RETRY\n", __func__
,
311 src_port
->name(), pkt
->print());
313 // update the layer state and schedule an idle event
314 reqLayers
[master_port_id
]->failedTiming(src_port
,
315 clockEdge(Cycles(1)));
317 // express snoops currently bypass the crossbar state entirely
318 if (!is_express_snoop
) {
319 // if this particular request will generate a snoop
321 if (expect_snoop_resp
) {
322 // we should never have an exsiting request outstanding
323 assert(outstandingSnoop
.find(pkt
->req
) ==
324 outstandingSnoop
.end());
325 outstandingSnoop
.insert(pkt
->req
);
327 // basic sanity check on the outstanding snoops
328 panic_if(outstandingSnoop
.size() > 512,
329 "Outstanding snoop requests exceeded 512\n");
332 // remember where to route the normal response to
333 if (expect_response
|| expect_snoop_resp
) {
334 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
335 routeTo
[pkt
->req
] = slave_port_id
;
337 panic_if(routeTo
.size() > 512,
338 "Routing table exceeds 512 packets\n");
341 // update the layer state and schedule an idle event
342 reqLayers
[master_port_id
]->succeededTiming(packetFinishTime
);
345 // stats updates only consider packets that were successfully sent
346 pktCount
[slave_port_id
][master_port_id
]++;
347 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
348 transDist
[pkt_cmd
]++;
350 if (is_express_snoop
) {
352 snoopTraffic
+= pkt_size
;
357 // queue the packet for deletion
358 pendingDelete
.reset(pkt
);
360 // normally we respond to the packet we just received if we need to
361 PacketPtr rsp_pkt
= pkt
;
362 PortID rsp_port_id
= slave_port_id
;
364 // If this is the destination of the cache clean operation the
365 // crossbar is responsible for responding. This crossbar will
366 // respond when the cache clean is complete. A cache clean
367 // is complete either:
368 // * direcly, if no cache above had a dirty copy of the block
369 // as indicated by the satisfied flag of the packet, or
370 // * when the crossbar has seen both the cache clean request
371 // (CleanSharedReq, CleanInvalidReq) and the corresponding
372 // write (WriteClean) which updates the block in the memory
375 ((pkt
->isClean() && pkt
->satisfied()) ||
376 pkt
->cmd
== MemCmd::WriteClean
) &&
378 PacketPtr deferred_rsp
= pkt
->isWrite() ? nullptr : pkt
;
379 auto cmo_lookup
= outstandingCMO
.find(pkt
->id
);
380 if (cmo_lookup
!= outstandingCMO
.end()) {
381 // the cache clean request has already reached this xbar
382 respond_directly
= true;
383 if (pkt
->isWrite()) {
384 rsp_pkt
= cmo_lookup
->second
;
387 // determine the destination
388 const auto route_lookup
= routeTo
.find(rsp_pkt
->req
);
389 assert(route_lookup
!= routeTo
.end());
390 rsp_port_id
= route_lookup
->second
;
391 assert(rsp_port_id
!= InvalidPortID
);
392 assert(rsp_port_id
< respLayers
.size());
393 // remove the request from the routing table
394 routeTo
.erase(route_lookup
);
396 outstandingCMO
.erase(cmo_lookup
);
398 respond_directly
= false;
399 outstandingCMO
.emplace(pkt
->id
, deferred_rsp
);
400 if (!pkt
->isWrite()) {
401 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
402 routeTo
[pkt
->req
] = slave_port_id
;
404 panic_if(routeTo
.size() > 512,
405 "Routing table exceeds 512 packets\n");
411 if (respond_directly
) {
412 assert(rsp_pkt
->needsResponse());
415 rsp_pkt
->makeResponse();
417 if (snoopFilter
&& !system
->bypassCaches()) {
418 // let the snoop filter inspect the response and update its state
419 snoopFilter
->updateResponse(rsp_pkt
, *slavePorts
[rsp_port_id
]);
422 // we send the response after the current packet, even if the
423 // response is not for this packet (e.g. cache clean operation
424 // where both the request and the write packet have to cross
425 // the destination xbar before the response is sent.)
426 Tick response_time
= clockEdge() + pkt
->headerDelay
;
427 rsp_pkt
->headerDelay
= 0;
429 slavePorts
[rsp_port_id
]->schedTimingResp(rsp_pkt
, response_time
);
436 CoherentXBar::recvTimingResp(PacketPtr pkt
, PortID master_port_id
)
438 // determine the source port based on the id
439 MasterPort
*src_port
= masterPorts
[master_port_id
];
441 // determine the destination
442 const auto route_lookup
= routeTo
.find(pkt
->req
);
443 assert(route_lookup
!= routeTo
.end());
444 const PortID slave_port_id
= route_lookup
->second
;
445 assert(slave_port_id
!= InvalidPortID
);
446 assert(slave_port_id
< respLayers
.size());
448 // test if the crossbar should be considered occupied for the
450 if (!respLayers
[slave_port_id
]->tryTiming(src_port
)) {
451 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
452 src_port
->name(), pkt
->print());
456 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
457 src_port
->name(), pkt
->print());
459 // store size and command as they might be modified when
460 // forwarding the packet
461 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
462 unsigned int pkt_cmd
= pkt
->cmdToIndex();
464 // a response sees the response latency
465 Tick xbar_delay
= responseLatency
* clockPeriod();
467 // set the packet header and payload delay
468 calcPacketTiming(pkt
, xbar_delay
);
470 // determine how long to be crossbar layer is busy
471 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
473 if (snoopFilter
&& !system
->bypassCaches()) {
474 // let the snoop filter inspect the response and update its state
475 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
478 // send the packet through the destination slave port and pay for
479 // any outstanding header delay
480 Tick latency
= pkt
->headerDelay
;
481 pkt
->headerDelay
= 0;
482 slavePorts
[slave_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
484 // remove the request from the routing table
485 routeTo
.erase(route_lookup
);
487 respLayers
[slave_port_id
]->succeededTiming(packetFinishTime
);
490 pktCount
[slave_port_id
][master_port_id
]++;
491 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
492 transDist
[pkt_cmd
]++;
498 CoherentXBar::recvTimingSnoopReq(PacketPtr pkt
, PortID master_port_id
)
500 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
501 masterPorts
[master_port_id
]->name(), pkt
->print());
503 // update stats here as we know the forwarding will succeed
504 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
505 transDist
[pkt
->cmdToIndex()]++;
507 snoopTraffic
+= pkt_size
;
509 // we should only see express snoops from caches
510 assert(pkt
->isExpressSnoop());
512 // set the packet header and payload delay, for now use forward latency
513 // @todo Assess the choice of latency further
514 calcPacketTiming(pkt
, forwardLatency
* clockPeriod());
516 // remember if a cache has already committed to responding so we
517 // can see if it changes during the snooping
518 const bool cache_responding
= pkt
->cacheResponding();
520 assert(pkt
->snoopDelay
== 0);
523 // let the Snoop Filter work its magic and guide probing
524 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
525 // the time required by a packet to be delivered through
526 // the xbar has to be charged also with to lookup latency
527 // of the snoop filter
528 pkt
->headerDelay
+= sf_res
.second
* clockPeriod();
529 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
530 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
531 sf_res
.first
.size(), sf_res
.second
);
533 // forward to all snoopers
534 forwardTiming(pkt
, InvalidPortID
, sf_res
.first
);
536 forwardTiming(pkt
, InvalidPortID
);
539 // add the snoop delay to our header delay, and then reset it
540 pkt
->headerDelay
+= pkt
->snoopDelay
;
543 // if we can expect a response, remember how to route it
544 if (!cache_responding
&& pkt
->cacheResponding()) {
545 assert(routeTo
.find(pkt
->req
) == routeTo
.end());
546 routeTo
[pkt
->req
] = master_port_id
;
549 // a snoop request came from a connected slave device (one of
550 // our master ports), and if it is not coming from the slave
551 // device responsible for the address range something is
552 // wrong, hence there is nothing further to do as the packet
553 // would be going back to where it came from
554 assert(findPort(pkt
->getAddrRange()) == master_port_id
);
558 CoherentXBar::recvTimingSnoopResp(PacketPtr pkt
, PortID slave_port_id
)
560 // determine the source port based on the id
561 SlavePort
* src_port
= slavePorts
[slave_port_id
];
563 // get the destination
564 const auto route_lookup
= routeTo
.find(pkt
->req
);
565 assert(route_lookup
!= routeTo
.end());
566 const PortID dest_port_id
= route_lookup
->second
;
567 assert(dest_port_id
!= InvalidPortID
);
569 // determine if the response is from a snoop request we
570 // created as the result of a normal request (in which case it
571 // should be in the outstandingSnoop), or if we merely forwarded
572 // someone else's snoop request
573 const bool forwardAsSnoop
= outstandingSnoop
.find(pkt
->req
) ==
574 outstandingSnoop
.end();
576 // test if the crossbar should be considered occupied for the
577 // current port, note that the check is bypassed if the response
578 // is being passed on as a normal response since this is occupying
579 // the response layer rather than the snoop response layer
580 if (forwardAsSnoop
) {
581 assert(dest_port_id
< snoopLayers
.size());
582 if (!snoopLayers
[dest_port_id
]->tryTiming(src_port
)) {
583 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
584 src_port
->name(), pkt
->print());
588 // get the master port that mirrors this slave port internally
589 MasterPort
* snoop_port
= snoopRespPorts
[slave_port_id
];
590 assert(dest_port_id
< respLayers
.size());
591 if (!respLayers
[dest_port_id
]->tryTiming(snoop_port
)) {
592 DPRINTF(CoherentXBar
, "%s: src %s packet %s BUSY\n", __func__
,
593 snoop_port
->name(), pkt
->print());
598 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
599 src_port
->name(), pkt
->print());
601 // store size and command as they might be modified when
602 // forwarding the packet
603 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
604 unsigned int pkt_cmd
= pkt
->cmdToIndex();
606 // responses are never express snoops
607 assert(!pkt
->isExpressSnoop());
609 // a snoop response sees the snoop response latency, and if it is
610 // forwarded as a normal response, the response latency
612 (forwardAsSnoop
? snoopResponseLatency
: responseLatency
) *
615 // set the packet header and payload delay
616 calcPacketTiming(pkt
, xbar_delay
);
618 // determine how long to be crossbar layer is busy
619 Tick packetFinishTime
= clockEdge(Cycles(1)) + pkt
->payloadDelay
;
621 // forward it either as a snoop response or a normal response
622 if (forwardAsSnoop
) {
623 // this is a snoop response to a snoop request we forwarded,
624 // e.g. coming from the L1 and going to the L2, and it should
625 // be forwarded as a snoop response
628 // update the probe filter so that it can properly track the line
629 snoopFilter
->updateSnoopForward(pkt
, *slavePorts
[slave_port_id
],
630 *masterPorts
[dest_port_id
]);
633 bool success M5_VAR_USED
=
634 masterPorts
[dest_port_id
]->sendTimingSnoopResp(pkt
);
635 pktCount
[slave_port_id
][dest_port_id
]++;
636 pktSize
[slave_port_id
][dest_port_id
] += pkt_size
;
639 snoopLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
641 // we got a snoop response on one of our slave ports,
642 // i.e. from a coherent master connected to the crossbar, and
643 // since we created the snoop request as part of recvTiming,
644 // this should now be a normal response again
645 outstandingSnoop
.erase(pkt
->req
);
647 // this is a snoop response from a coherent master, hence it
648 // should never go back to where the snoop response came from,
649 // but instead to where the original request came from
650 assert(slave_port_id
!= dest_port_id
);
653 // update the probe filter so that it can properly track the line
654 snoopFilter
->updateSnoopResponse(pkt
, *slavePorts
[slave_port_id
],
655 *slavePorts
[dest_port_id
]);
658 DPRINTF(CoherentXBar
, "%s: src %s packet %s FWD RESP\n", __func__
,
659 src_port
->name(), pkt
->print());
661 // as a normal response, it should go back to a master through
662 // one of our slave ports, we also pay for any outstanding
664 Tick latency
= pkt
->headerDelay
;
665 pkt
->headerDelay
= 0;
666 slavePorts
[dest_port_id
]->schedTimingResp(pkt
, curTick() + latency
);
668 respLayers
[dest_port_id
]->succeededTiming(packetFinishTime
);
671 // remove the request from the routing table
672 routeTo
.erase(route_lookup
);
675 transDist
[pkt_cmd
]++;
677 snoopTraffic
+= pkt_size
;
684 CoherentXBar::forwardTiming(PacketPtr pkt
, PortID exclude_slave_port_id
,
685 const std::vector
<QueuedSlavePort
*>& dests
)
687 DPRINTF(CoherentXBar
, "%s for %s\n", __func__
, pkt
->print());
689 // snoops should only happen if the system isn't bypassing caches
690 assert(!system
->bypassCaches());
694 for (const auto& p
: dests
) {
695 // we could have gotten this request from a snooping master
696 // (corresponding to our own slave port that is also in
697 // snoopPorts) and should not send it back to where it came
699 if (exclude_slave_port_id
== InvalidPortID
||
700 p
->getId() != exclude_slave_port_id
) {
701 // cache is not allowed to refuse snoop
702 p
->sendTimingSnoopReq(pkt
);
707 // Stats for fanout of this forward operation
708 snoopFanout
.sample(fanout
);
712 CoherentXBar::recvReqRetry(PortID master_port_id
)
714 // responses and snoop responses never block on forwarding them,
715 // so the retry will always be coming from a port to which we
716 // tried to forward a request
717 reqLayers
[master_port_id
]->recvRetry();
721 CoherentXBar::recvAtomicBackdoor(PacketPtr pkt
, PortID slave_port_id
,
722 MemBackdoorPtr
*backdoor
)
724 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
725 slavePorts
[slave_port_id
]->name(), pkt
->print());
727 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
728 unsigned int pkt_cmd
= pkt
->cmdToIndex();
730 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
731 Tick snoop_response_latency
= 0;
733 // is this the destination point for this packet? (e.g. true if
734 // this xbar is the PoC for a cache maintenance operation to the
735 // PoC) otherwise the destination is any cache that can satisfy
737 const bool is_destination
= isDestination(pkt
);
739 const bool snoop_caches
= !system
->bypassCaches() &&
740 pkt
->cmd
!= MemCmd::WriteClean
;
742 // forward to all snoopers but the source
743 std::pair
<MemCmd
, Tick
> snoop_result
;
745 // check with the snoop filter where to forward this packet
747 snoopFilter
->lookupRequest(pkt
, *slavePorts
[slave_port_id
]);
748 snoop_response_latency
+= sf_res
.second
* clockPeriod();
749 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
750 __func__
, slavePorts
[slave_port_id
]->name(), pkt
->print(),
751 sf_res
.first
.size(), sf_res
.second
);
753 // let the snoop filter know about the success of the send
754 // operation, and do it even before sending it onwards to
755 // avoid situations where atomic upward snoops sneak in
756 // between and change the filter state
757 snoopFilter
->finishRequest(false, pkt
->getAddr(), pkt
->isSecure());
759 if (pkt
->isEviction()) {
760 // for block-evicting packets, i.e. writebacks and
761 // clean evictions, there is no need to snoop up, as
762 // all we do is determine if the block is cached or
763 // not, instead just set it here based on the snoop
765 if (!sf_res
.first
.empty())
766 pkt
->setBlockCached();
768 snoop_result
= forwardAtomic(pkt
, slave_port_id
, InvalidPortID
,
772 snoop_result
= forwardAtomic(pkt
, slave_port_id
);
774 snoop_response_cmd
= snoop_result
.first
;
775 snoop_response_latency
+= snoop_result
.second
;
778 // set up a sensible default value
779 Tick response_latency
= 0;
781 const bool sink_packet
= sinkPacket(pkt
);
783 // even if we had a snoop response, we must continue and also
784 // perform the actual request at the destination
785 PortID master_port_id
= findPort(pkt
->getAddrRange());
788 DPRINTF(CoherentXBar
, "%s: Not forwarding %s\n", __func__
,
791 if (forwardPacket(pkt
)) {
792 // make sure that the write request (e.g., WriteClean)
793 // will stop at the memory below if this crossbar is its
795 if (pkt
->isWrite() && is_destination
) {
796 pkt
->clearWriteThrough();
799 // forward the request to the appropriate destination
800 auto master
= masterPorts
[master_port_id
];
801 response_latency
= backdoor
?
802 master
->sendAtomicBackdoor(pkt
, *backdoor
) :
803 master
->sendAtomic(pkt
);
805 // if it does not need a response we sink the packet above
806 assert(pkt
->needsResponse());
812 // stats updates for the request
813 pktCount
[slave_port_id
][master_port_id
]++;
814 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
815 transDist
[pkt_cmd
]++;
818 // if lower levels have replied, tell the snoop filter
819 if (!system
->bypassCaches() && snoopFilter
&& pkt
->isResponse()) {
820 snoopFilter
->updateResponse(pkt
, *slavePorts
[slave_port_id
]);
823 // if we got a response from a snooper, restore it here
824 if (snoop_response_cmd
!= MemCmd::InvalidCmd
) {
825 // no one else should have responded
826 assert(!pkt
->isResponse());
827 pkt
->cmd
= snoop_response_cmd
;
828 response_latency
= snoop_response_latency
;
831 // If this is the destination of the cache clean operation the
832 // crossbar is responsible for responding. This crossbar will
833 // respond when the cache clean is complete. An atomic cache clean
834 // is complete when the crossbars receives the cache clean
835 // request (CleanSharedReq, CleanInvalidReq), as either:
836 // * no cache above had a dirty copy of the block as indicated by
837 // the satisfied flag of the packet, or
838 // * the crossbar has already seen the corresponding write
839 // (WriteClean) which updates the block in the memory below.
840 if (pkt
->isClean() && isDestination(pkt
) && pkt
->satisfied()) {
841 auto it
= outstandingCMO
.find(pkt
->id
);
842 assert(it
!= outstandingCMO
.end());
843 // we are responding right away
844 outstandingCMO
.erase(it
);
845 } else if (pkt
->cmd
== MemCmd::WriteClean
&& isDestination(pkt
)) {
846 // if this is the destination of the operation, the xbar
847 // sends the responce to the cache clean operation only
848 // after having encountered the cache clean request
849 auto M5_VAR_USED ret
= outstandingCMO
.emplace(pkt
->id
, nullptr);
850 // in atomic mode we know that the WriteClean packet should
851 // precede the clean request
855 // add the response data
856 if (pkt
->isResponse()) {
857 pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
858 pkt_cmd
= pkt
->cmdToIndex();
861 pktCount
[slave_port_id
][master_port_id
]++;
862 pktSize
[slave_port_id
][master_port_id
] += pkt_size
;
863 transDist
[pkt_cmd
]++;
866 // @todo: Not setting header time
867 pkt
->payloadDelay
= response_latency
;
868 return response_latency
;
872 CoherentXBar::recvAtomicSnoop(PacketPtr pkt
, PortID master_port_id
)
874 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
875 masterPorts
[master_port_id
]->name(), pkt
->print());
877 // add the request snoop data
878 unsigned int pkt_size
= pkt
->hasData() ? pkt
->getSize() : 0;
880 snoopTraffic
+= pkt_size
;
882 // forward to all snoopers
883 std::pair
<MemCmd
, Tick
> snoop_result
;
884 Tick snoop_response_latency
= 0;
886 auto sf_res
= snoopFilter
->lookupSnoop(pkt
);
887 snoop_response_latency
+= sf_res
.second
* clockPeriod();
888 DPRINTF(CoherentXBar
, "%s: src %s packet %s SF size: %i lat: %i\n",
889 __func__
, masterPorts
[master_port_id
]->name(), pkt
->print(),
890 sf_res
.first
.size(), sf_res
.second
);
891 snoop_result
= forwardAtomic(pkt
, InvalidPortID
, master_port_id
,
894 snoop_result
= forwardAtomic(pkt
, InvalidPortID
);
896 MemCmd snoop_response_cmd
= snoop_result
.first
;
897 snoop_response_latency
+= snoop_result
.second
;
899 if (snoop_response_cmd
!= MemCmd::InvalidCmd
)
900 pkt
->cmd
= snoop_response_cmd
;
902 // add the response snoop data
903 if (pkt
->isResponse()) {
907 // @todo: Not setting header time
908 pkt
->payloadDelay
= snoop_response_latency
;
909 return snoop_response_latency
;
912 std::pair
<MemCmd
, Tick
>
913 CoherentXBar::forwardAtomic(PacketPtr pkt
, PortID exclude_slave_port_id
,
914 PortID source_master_port_id
,
915 const std::vector
<QueuedSlavePort
*>& dests
)
917 // the packet may be changed on snoops, record the original
918 // command to enable us to restore it between snoops so that
919 // additional snoops can take place properly
920 MemCmd orig_cmd
= pkt
->cmd
;
921 MemCmd snoop_response_cmd
= MemCmd::InvalidCmd
;
922 Tick snoop_response_latency
= 0;
924 // snoops should only happen if the system isn't bypassing caches
925 assert(!system
->bypassCaches());
929 for (const auto& p
: dests
) {
930 // we could have gotten this request from a snooping master
931 // (corresponding to our own slave port that is also in
932 // snoopPorts) and should not send it back to where it came
934 if (exclude_slave_port_id
!= InvalidPortID
&&
935 p
->getId() == exclude_slave_port_id
)
938 Tick latency
= p
->sendAtomicSnoop(pkt
);
941 // in contrast to a functional access, we have to keep on
942 // going as all snoopers must be updated even if we get a
944 if (!pkt
->isResponse())
947 // response from snoop agent
948 assert(pkt
->cmd
!= orig_cmd
);
949 assert(pkt
->cacheResponding());
950 // should only happen once
951 assert(snoop_response_cmd
== MemCmd::InvalidCmd
);
952 // save response state
953 snoop_response_cmd
= pkt
->cmd
;
954 snoop_response_latency
= latency
;
957 // Handle responses by the snoopers and differentiate between
958 // responses to requests from above and snoops from below
959 if (source_master_port_id
!= InvalidPortID
) {
960 // Getting a response for a snoop from below
961 assert(exclude_slave_port_id
== InvalidPortID
);
962 snoopFilter
->updateSnoopForward(pkt
, *p
,
963 *masterPorts
[source_master_port_id
]);
965 // Getting a response for a request from above
966 assert(source_master_port_id
== InvalidPortID
);
967 snoopFilter
->updateSnoopResponse(pkt
, *p
,
968 *slavePorts
[exclude_slave_port_id
]);
971 // restore original packet state for remaining snoopers
976 snoopFanout
.sample(fanout
);
978 // the packet is restored as part of the loop and any potential
979 // snoop response is part of the returned pair
980 return std::make_pair(snoop_response_cmd
, snoop_response_latency
);
984 CoherentXBar::recvFunctional(PacketPtr pkt
, PortID slave_port_id
)
986 if (!pkt
->isPrint()) {
987 // don't do DPRINTFs on PrintReq as it clutters up the output
988 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
989 slavePorts
[slave_port_id
]->name(), pkt
->print());
992 if (!system
->bypassCaches()) {
993 // forward to all snoopers but the source
994 forwardFunctional(pkt
, slave_port_id
);
997 // there is no need to continue if the snooping has found what we
998 // were looking for and the packet is already a response
999 if (!pkt
->isResponse()) {
1000 // since our slave ports are queued ports we need to check them as well
1001 for (const auto& p
: slavePorts
) {
1002 // if we find a response that has the data, then the
1003 // downstream caches/memories may be out of date, so simply stop
1005 if (p
->trySatisfyFunctional(pkt
)) {
1006 if (pkt
->needsResponse())
1007 pkt
->makeResponse();
1012 PortID dest_id
= findPort(pkt
->getAddrRange());
1014 masterPorts
[dest_id
]->sendFunctional(pkt
);
1019 CoherentXBar::recvFunctionalSnoop(PacketPtr pkt
, PortID master_port_id
)
1021 if (!pkt
->isPrint()) {
1022 // don't do DPRINTFs on PrintReq as it clutters up the output
1023 DPRINTF(CoherentXBar
, "%s: src %s packet %s\n", __func__
,
1024 masterPorts
[master_port_id
]->name(), pkt
->print());
1027 for (const auto& p
: slavePorts
) {
1028 if (p
->trySatisfyFunctional(pkt
)) {
1029 if (pkt
->needsResponse())
1030 pkt
->makeResponse();
1035 // forward to all snoopers
1036 forwardFunctional(pkt
, InvalidPortID
);
1040 CoherentXBar::forwardFunctional(PacketPtr pkt
, PortID exclude_slave_port_id
)
1042 // snoops should only happen if the system isn't bypassing caches
1043 assert(!system
->bypassCaches());
1045 for (const auto& p
: snoopPorts
) {
1046 // we could have gotten this request from a snooping master
1047 // (corresponding to our own slave port that is also in
1048 // snoopPorts) and should not send it back to where it came
1050 if (exclude_slave_port_id
== InvalidPortID
||
1051 p
->getId() != exclude_slave_port_id
)
1052 p
->sendFunctionalSnoop(pkt
);
1054 // if we get a response we are done
1055 if (pkt
->isResponse()) {
1062 CoherentXBar::sinkPacket(const PacketPtr pkt
) const
1064 // we can sink the packet if:
1065 // 1) the crossbar is the point of coherency, and a cache is
1066 // responding after being snooped
1067 // 2) the crossbar is the point of coherency, and the packet is a
1068 // coherency packet (not a read or a write) that does not
1069 // require a response
1070 // 3) this is a clean evict or clean writeback, but the packet is
1071 // found in a cache above this crossbar
1072 // 4) a cache is responding after being snooped, and the packet
1073 // either does not need the block to be writable, or the cache
1074 // that has promised to respond (setting the cache responding
1075 // flag) is providing writable and thus had a Modified block,
1076 // and no further action is needed
1077 return (pointOfCoherency
&& pkt
->cacheResponding()) ||
1078 (pointOfCoherency
&& !(pkt
->isRead() || pkt
->isWrite()) &&
1079 !pkt
->needsResponse()) ||
1080 (pkt
->isCleanEviction() && pkt
->isBlockCached()) ||
1081 (pkt
->cacheResponding() &&
1082 (!pkt
->needsWritable() || pkt
->responderHadWritable()));
1086 CoherentXBar::forwardPacket(const PacketPtr pkt
)
1088 // we are forwarding the packet if:
1089 // 1) this is a cache clean request to the PoU/PoC and this
1090 // crossbar is above the PoU/PoC
1091 // 2) this is a read or a write
1092 // 3) this crossbar is above the point of coherency
1093 if (pkt
->isClean()) {
1094 return !isDestination(pkt
);
1096 return pkt
->isRead() || pkt
->isWrite() || !pointOfCoherency
;
1101 CoherentXBar::regStats()
1103 // register the stats of the base class and our layers
1104 BaseXBar::regStats();
1105 for (auto l
: reqLayers
)
1107 for (auto l
: respLayers
)
1109 for (auto l
: snoopLayers
)
1113 .name(name() + ".snoops")
1114 .desc("Total snoops (count)")
1118 .name(name() + ".snoopTraffic")
1119 .desc("Total snoop traffic (bytes)")
1123 .init(0, snoopPorts
.size(), 1)
1124 .name(name() + ".snoop_fanout")
1125 .desc("Request fanout histogram")
1130 CoherentXBarParams::create()
1132 return new CoherentXBar(this);