1/*
| 1/*
|
2 * Copyright (c) 2011-2017 ARM Limited
| 2 * Copyright (c) 2011-2018 ARM Limited
|
3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2006 The Regents of The University of Michigan
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Ali Saidi
41 * Andreas Hansson
42 * William Wang
43 * Nikos Nikoleris
44 */
45
46/**
47 * @file
48 * Definition of a crossbar object.
49 */
50
51#include "mem/coherent_xbar.hh"
52
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"
58
59CoherentXBar::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)
64{
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)));
76 }
77
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,
84 defaultPortID);
85 masterPorts.push_back(bp);
86 reqLayers.push_back(new ReqLayer(*bp, *this, csprintf(".reqLayer%d",
87 defaultPortID)));
88 snoopLayers.push_back(new SnoopRespLayer(*bp, *this,
89 csprintf(".snoopLayer%d",
90 defaultPortID)));
91 }
92
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));
101 }
102}
103
104CoherentXBar::~CoherentXBar()
105{
106 for (auto l: reqLayers)
107 delete l;
108 for (auto l: respLayers)
109 delete l;
110 for (auto l: snoopLayers)
111 delete l;
112 for (auto p: snoopRespPorts)
113 delete p;
114}
115
116void
117CoherentXBar::init()
118{
119 BaseXBar::init();
120
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);
129 }
130 }
131
132 if (snoopPorts.empty())
133 warn("CoherentXBar %s has no snooping ports attached!\n", name());
134
135 // inform the snoop filter about the slave ports so it can create
136 // its own internal representation
137 if (snoopFilter)
138 snoopFilter->setSlavePorts(slavePorts);
139}
140
141bool
142CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
143{
144 // determine the source port based on the id
145 SlavePort *src_port = slavePorts[slave_port_id];
146
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);
153
| 3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder. You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2006 The Regents of The University of Michigan
15 * All rights reserved.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions are
19 * met: redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer;
21 * redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution;
24 * neither the name of the copyright holders nor the names of its
25 * contributors may be used to endorse or promote products derived from
26 * this software without specific prior written permission.
27 *
28 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
29 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
30 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
31 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
32 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
33 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
34 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
35 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
36 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
37 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
38 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
39 *
40 * Authors: Ali Saidi
41 * Andreas Hansson
42 * William Wang
43 * Nikos Nikoleris
44 */
45
46/**
47 * @file
48 * Definition of a crossbar object.
49 */
50
51#include "mem/coherent_xbar.hh"
52
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"
58
59CoherentXBar::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)
64{
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)));
76 }
77
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,
84 defaultPortID);
85 masterPorts.push_back(bp);
86 reqLayers.push_back(new ReqLayer(*bp, *this, csprintf(".reqLayer%d",
87 defaultPortID)));
88 snoopLayers.push_back(new SnoopRespLayer(*bp, *this,
89 csprintf(".snoopLayer%d",
90 defaultPortID)));
91 }
92
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));
101 }
102}
103
104CoherentXBar::~CoherentXBar()
105{
106 for (auto l: reqLayers)
107 delete l;
108 for (auto l: respLayers)
109 delete l;
110 for (auto l: snoopLayers)
111 delete l;
112 for (auto p: snoopRespPorts)
113 delete p;
114}
115
116void
117CoherentXBar::init()
118{
119 BaseXBar::init();
120
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);
129 }
130 }
131
132 if (snoopPorts.empty())
133 warn("CoherentXBar %s has no snooping ports attached!\n", name());
134
135 // inform the snoop filter about the slave ports so it can create
136 // its own internal representation
137 if (snoopFilter)
138 snoopFilter->setSlavePorts(slavePorts);
139}
140
141bool
142CoherentXBar::recvTimingReq(PacketPtr pkt, PortID slave_port_id)
143{
144 // determine the source port based on the id
145 SlavePort *src_port = slavePorts[slave_port_id];
146
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);
153
|
154 // determine the destination based on the address
155 PortID master_port_id = findPort(pkt->getAddr());
| 154 // determine the destination based on the destination address range
155 AddrRange addr_range = RangeSize(pkt->getAddr(), pkt->getSize());
156 PortID master_port_id = findPort(addr_range);
|
156
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());
162 return false;
163 }
164
165 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
166 src_port->name(), pkt->print());
167
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();
172
173 // store the old header delay so we can restore it if needed
174 Tick old_header_delay = pkt->headerDelay;
175
176 // a request sees the frontend and forward latency
177 Tick xbar_delay = (frontendLatency + forwardLatency) * clockPeriod();
178
179 // set the packet header and payload delay
180 calcPacketTiming(pkt, xbar_delay);
181
182 // determine how long to be crossbar layer is busy
183 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
184
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
188 // the request
189 const bool is_destination = isDestination(pkt);
190
191 const bool snoop_caches = !system->bypassCaches() &&
192 pkt->cmd != MemCmd::WriteClean;
193 if (snoop_caches) {
194 assert(pkt->snoopDelay == 0);
195
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
199 // forwarded to it
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());
203
204 // update the layer state and schedule an idle event
205 reqLayers[master_port_id]->failedTiming(src_port,
206 clockEdge(Cycles(1)));
207 return false;
208 }
209 }
210
211
212 // the packet is a memory-mapped request and should be
213 // broadcasted to our snoopers but the source
214 if (snoopFilter) {
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);
224
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
230 // filter result
231 if (!sf_res.first.empty())
232 pkt->setBlockCached();
233 } else {
234 forwardTiming(pkt, slave_port_id, sf_res.first);
235 }
236 } else {
237 forwardTiming(pkt, slave_port_id);
238 }
239
240 // add the snoop delay to our header delay, and then reset it
241 pkt->headerDelay += pkt->snoopDelay;
242 pkt->snoopDelay = 0;
243 }
244
245 // set up a sensible starting point
246 bool success = true;
247
248 // remember if the packet will generate a snoop response by
249 // checking if a cache set the cacheResponding flag during the
250 // snooping above
251 const bool expect_snoop_resp = !cache_responding && pkt->cacheResponding();
252 bool expect_response = pkt->needsResponse() && !pkt->cacheResponding();
253
254 const bool sink_packet = sinkPacket(pkt);
255
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());
261 if (sink_packet) {
262 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__,
263 pkt->print());
264 } else {
265 // determine if we are forwarding the packet, or responding to
266 // it
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();
275 }
276
277 // make sure that the write request (e.g., WriteClean)
278 // will stop at the memory below if this crossbar is its
279 // destination
280 if (pkt->isWrite() && is_destination) {
281 pkt->clearWriteThrough();
282 }
283
284 // since it is a normal request, attempt to send the packet
285 success = masterPorts[master_port_id]->sendTimingReq(pkt);
286 } else {
287 // no need to forward, turn this packet around and respond
288 // directly
289 assert(pkt->needsResponse());
290
291 respond_directly = true;
292 assert(!expect_snoop_resp);
293 expect_response = false;
294 }
295 }
296
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());
300 }
301
302 // check if we were successful in sending the packet onwards
303 if (!success) {
304 // express snoops should never be forced to retry
305 assert(!is_express_snoop);
306
307 // restore the header delay
308 pkt->headerDelay = old_header_delay;
309
310 DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__,
311 src_port->name(), pkt->print());
312
313 // update the layer state and schedule an idle event
314 reqLayers[master_port_id]->failedTiming(src_port,
315 clockEdge(Cycles(1)));
316 } else {
317 // express snoops currently bypass the crossbar state entirely
318 if (!is_express_snoop) {
319 // if this particular request will generate a snoop
320 // response
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);
326
327 // basic sanity check on the outstanding snoops
328 panic_if(outstandingSnoop.size() > 512,
329 "Outstanding snoop requests exceeded 512\n");
330 }
331
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;
336
337 panic_if(routeTo.size() > 512,
338 "Routing table exceeds 512 packets\n");
339 }
340
341 // update the layer state and schedule an idle event
342 reqLayers[master_port_id]->succeededTiming(packetFinishTime);
343 }
344
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]++;
349
350 if (is_express_snoop) {
351 snoops++;
352 snoopTraffic += pkt_size;
353 }
354 }
355
356 if (sink_packet)
357 // queue the packet for deletion
358 pendingDelete.reset(pkt);
359
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;
363
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
373 // below.
374 if (success &&
375 ((pkt->isClean() && pkt->satisfied()) ||
376 pkt->cmd == MemCmd::WriteClean) &&
377 is_destination) {
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;
385 assert(rsp_pkt);
386
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);
395 }
396 outstandingCMO.erase(cmo_lookup);
397 } else {
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;
403
404 panic_if(routeTo.size() > 512,
405 "Routing table exceeds 512 packets\n");
406 }
407 }
408 }
409
410
411 if (respond_directly) {
412 assert(rsp_pkt->needsResponse());
413 assert(success);
414
415 rsp_pkt->makeResponse();
416
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]);
420 }
421
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;
428
429 slavePorts[rsp_port_id]->schedTimingResp(rsp_pkt, response_time);
430 }
431
432 return success;
433}
434
435bool
436CoherentXBar::recvTimingResp(PacketPtr pkt, PortID master_port_id)
437{
438 // determine the source port based on the id
439 MasterPort *src_port = masterPorts[master_port_id];
440
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());
447
448 // test if the crossbar should be considered occupied for the
449 // current port
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());
453 return false;
454 }
455
456 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
457 src_port->name(), pkt->print());
458
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();
463
464 // a response sees the response latency
465 Tick xbar_delay = responseLatency * clockPeriod();
466
467 // set the packet header and payload delay
468 calcPacketTiming(pkt, xbar_delay);
469
470 // determine how long to be crossbar layer is busy
471 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
472
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]);
476 }
477
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);
483
484 // remove the request from the routing table
485 routeTo.erase(route_lookup);
486
487 respLayers[slave_port_id]->succeededTiming(packetFinishTime);
488
489 // stats updates
490 pktCount[slave_port_id][master_port_id]++;
491 pktSize[slave_port_id][master_port_id] += pkt_size;
492 transDist[pkt_cmd]++;
493
494 return true;
495}
496
497void
498CoherentXBar::recvTimingSnoopReq(PacketPtr pkt, PortID master_port_id)
499{
500 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
501 masterPorts[master_port_id]->name(), pkt->print());
502
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()]++;
506 snoops++;
507 snoopTraffic += pkt_size;
508
509 // we should only see express snoops from caches
510 assert(pkt->isExpressSnoop());
511
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());
515
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();
519
520 assert(pkt->snoopDelay == 0);
521
522 if (snoopFilter) {
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);
532
533 // forward to all snoopers
534 forwardTiming(pkt, InvalidPortID, sf_res.first);
535 } else {
536 forwardTiming(pkt, InvalidPortID);
537 }
538
539 // add the snoop delay to our header delay, and then reset it
540 pkt->headerDelay += pkt->snoopDelay;
541 pkt->snoopDelay = 0;
542
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;
547 }
548
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
| 157
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());
163 return false;
164 }
165
166 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
167 src_port->name(), pkt->print());
168
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();
173
174 // store the old header delay so we can restore it if needed
175 Tick old_header_delay = pkt->headerDelay;
176
177 // a request sees the frontend and forward latency
178 Tick xbar_delay = (frontendLatency + forwardLatency) * clockPeriod();
179
180 // set the packet header and payload delay
181 calcPacketTiming(pkt, xbar_delay);
182
183 // determine how long to be crossbar layer is busy
184 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
185
186 // is this the destination point for this packet? (e.g. true if
187 // this xbar is the PoC for a cache maintenance operation to the
188 // PoC) otherwise the destination is any cache that can satisfy
189 // the request
190 const bool is_destination = isDestination(pkt);
191
192 const bool snoop_caches = !system->bypassCaches() &&
193 pkt->cmd != MemCmd::WriteClean;
194 if (snoop_caches) {
195 assert(pkt->snoopDelay == 0);
196
197 if (pkt->isClean() && !is_destination) {
198 // before snooping we need to make sure that the memory
199 // below is not busy and the cache clean request can be
200 // forwarded to it
201 if (!masterPorts[master_port_id]->tryTiming(pkt)) {
202 DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__,
203 src_port->name(), pkt->print());
204
205 // update the layer state and schedule an idle event
206 reqLayers[master_port_id]->failedTiming(src_port,
207 clockEdge(Cycles(1)));
208 return false;
209 }
210 }
211
212
213 // the packet is a memory-mapped request and should be
214 // broadcasted to our snoopers but the source
215 if (snoopFilter) {
216 // check with the snoop filter where to forward this packet
217 auto sf_res = snoopFilter->lookupRequest(pkt, *src_port);
218 // the time required by a packet to be delivered through
219 // the xbar has to be charged also with to lookup latency
220 // of the snoop filter
221 pkt->headerDelay += sf_res.second * clockPeriod();
222 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n",
223 __func__, src_port->name(), pkt->print(),
224 sf_res.first.size(), sf_res.second);
225
226 if (pkt->isEviction()) {
227 // for block-evicting packets, i.e. writebacks and
228 // clean evictions, there is no need to snoop up, as
229 // all we do is determine if the block is cached or
230 // not, instead just set it here based on the snoop
231 // filter result
232 if (!sf_res.first.empty())
233 pkt->setBlockCached();
234 } else {
235 forwardTiming(pkt, slave_port_id, sf_res.first);
236 }
237 } else {
238 forwardTiming(pkt, slave_port_id);
239 }
240
241 // add the snoop delay to our header delay, and then reset it
242 pkt->headerDelay += pkt->snoopDelay;
243 pkt->snoopDelay = 0;
244 }
245
246 // set up a sensible starting point
247 bool success = true;
248
249 // remember if the packet will generate a snoop response by
250 // checking if a cache set the cacheResponding flag during the
251 // snooping above
252 const bool expect_snoop_resp = !cache_responding && pkt->cacheResponding();
253 bool expect_response = pkt->needsResponse() && !pkt->cacheResponding();
254
255 const bool sink_packet = sinkPacket(pkt);
256
257 // in certain cases the crossbar is responsible for responding
258 bool respond_directly = false;
259 // store the original address as an address mapper could possibly
260 // modify the address upon a sendTimingRequest
261 const Addr addr(pkt->getAddr());
262 if (sink_packet) {
263 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__,
264 pkt->print());
265 } else {
266 // determine if we are forwarding the packet, or responding to
267 // it
268 if (forwardPacket(pkt)) {
269 // if we are passing on, rather than sinking, a packet to
270 // which an upstream cache has committed to responding,
271 // the line was needs writable, and the responding only
272 // had an Owned copy, so we need to immidiately let the
273 // downstream caches know, bypass any flow control
274 if (pkt->cacheResponding()) {
275 pkt->setExpressSnoop();
276 }
277
278 // make sure that the write request (e.g., WriteClean)
279 // will stop at the memory below if this crossbar is its
280 // destination
281 if (pkt->isWrite() && is_destination) {
282 pkt->clearWriteThrough();
283 }
284
285 // since it is a normal request, attempt to send the packet
286 success = masterPorts[master_port_id]->sendTimingReq(pkt);
287 } else {
288 // no need to forward, turn this packet around and respond
289 // directly
290 assert(pkt->needsResponse());
291
292 respond_directly = true;
293 assert(!expect_snoop_resp);
294 expect_response = false;
295 }
296 }
297
298 if (snoopFilter && snoop_caches) {
299 // Let the snoop filter know about the success of the send operation
300 snoopFilter->finishRequest(!success, addr, pkt->isSecure());
301 }
302
303 // check if we were successful in sending the packet onwards
304 if (!success) {
305 // express snoops should never be forced to retry
306 assert(!is_express_snoop);
307
308 // restore the header delay
309 pkt->headerDelay = old_header_delay;
310
311 DPRINTF(CoherentXBar, "%s: src %s packet %s RETRY\n", __func__,
312 src_port->name(), pkt->print());
313
314 // update the layer state and schedule an idle event
315 reqLayers[master_port_id]->failedTiming(src_port,
316 clockEdge(Cycles(1)));
317 } else {
318 // express snoops currently bypass the crossbar state entirely
319 if (!is_express_snoop) {
320 // if this particular request will generate a snoop
321 // response
322 if (expect_snoop_resp) {
323 // we should never have an exsiting request outstanding
324 assert(outstandingSnoop.find(pkt->req) ==
325 outstandingSnoop.end());
326 outstandingSnoop.insert(pkt->req);
327
328 // basic sanity check on the outstanding snoops
329 panic_if(outstandingSnoop.size() > 512,
330 "Outstanding snoop requests exceeded 512\n");
331 }
332
333 // remember where to route the normal response to
334 if (expect_response || expect_snoop_resp) {
335 assert(routeTo.find(pkt->req) == routeTo.end());
336 routeTo[pkt->req] = slave_port_id;
337
338 panic_if(routeTo.size() > 512,
339 "Routing table exceeds 512 packets\n");
340 }
341
342 // update the layer state and schedule an idle event
343 reqLayers[master_port_id]->succeededTiming(packetFinishTime);
344 }
345
346 // stats updates only consider packets that were successfully sent
347 pktCount[slave_port_id][master_port_id]++;
348 pktSize[slave_port_id][master_port_id] += pkt_size;
349 transDist[pkt_cmd]++;
350
351 if (is_express_snoop) {
352 snoops++;
353 snoopTraffic += pkt_size;
354 }
355 }
356
357 if (sink_packet)
358 // queue the packet for deletion
359 pendingDelete.reset(pkt);
360
361 // normally we respond to the packet we just received if we need to
362 PacketPtr rsp_pkt = pkt;
363 PortID rsp_port_id = slave_port_id;
364
365 // If this is the destination of the cache clean operation the
366 // crossbar is responsible for responding. This crossbar will
367 // respond when the cache clean is complete. A cache clean
368 // is complete either:
369 // * direcly, if no cache above had a dirty copy of the block
370 // as indicated by the satisfied flag of the packet, or
371 // * when the crossbar has seen both the cache clean request
372 // (CleanSharedReq, CleanInvalidReq) and the corresponding
373 // write (WriteClean) which updates the block in the memory
374 // below.
375 if (success &&
376 ((pkt->isClean() && pkt->satisfied()) ||
377 pkt->cmd == MemCmd::WriteClean) &&
378 is_destination) {
379 PacketPtr deferred_rsp = pkt->isWrite() ? nullptr : pkt;
380 auto cmo_lookup = outstandingCMO.find(pkt->id);
381 if (cmo_lookup != outstandingCMO.end()) {
382 // the cache clean request has already reached this xbar
383 respond_directly = true;
384 if (pkt->isWrite()) {
385 rsp_pkt = cmo_lookup->second;
386 assert(rsp_pkt);
387
388 // determine the destination
389 const auto route_lookup = routeTo.find(rsp_pkt->req);
390 assert(route_lookup != routeTo.end());
391 rsp_port_id = route_lookup->second;
392 assert(rsp_port_id != InvalidPortID);
393 assert(rsp_port_id < respLayers.size());
394 // remove the request from the routing table
395 routeTo.erase(route_lookup);
396 }
397 outstandingCMO.erase(cmo_lookup);
398 } else {
399 respond_directly = false;
400 outstandingCMO.emplace(pkt->id, deferred_rsp);
401 if (!pkt->isWrite()) {
402 assert(routeTo.find(pkt->req) == routeTo.end());
403 routeTo[pkt->req] = slave_port_id;
404
405 panic_if(routeTo.size() > 512,
406 "Routing table exceeds 512 packets\n");
407 }
408 }
409 }
410
411
412 if (respond_directly) {
413 assert(rsp_pkt->needsResponse());
414 assert(success);
415
416 rsp_pkt->makeResponse();
417
418 if (snoopFilter && !system->bypassCaches()) {
419 // let the snoop filter inspect the response and update its state
420 snoopFilter->updateResponse(rsp_pkt, *slavePorts[rsp_port_id]);
421 }
422
423 // we send the response after the current packet, even if the
424 // response is not for this packet (e.g. cache clean operation
425 // where both the request and the write packet have to cross
426 // the destination xbar before the response is sent.)
427 Tick response_time = clockEdge() + pkt->headerDelay;
428 rsp_pkt->headerDelay = 0;
429
430 slavePorts[rsp_port_id]->schedTimingResp(rsp_pkt, response_time);
431 }
432
433 return success;
434}
435
436bool
437CoherentXBar::recvTimingResp(PacketPtr pkt, PortID master_port_id)
438{
439 // determine the source port based on the id
440 MasterPort *src_port = masterPorts[master_port_id];
441
442 // determine the destination
443 const auto route_lookup = routeTo.find(pkt->req);
444 assert(route_lookup != routeTo.end());
445 const PortID slave_port_id = route_lookup->second;
446 assert(slave_port_id != InvalidPortID);
447 assert(slave_port_id < respLayers.size());
448
449 // test if the crossbar should be considered occupied for the
450 // current port
451 if (!respLayers[slave_port_id]->tryTiming(src_port)) {
452 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__,
453 src_port->name(), pkt->print());
454 return false;
455 }
456
457 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
458 src_port->name(), pkt->print());
459
460 // store size and command as they might be modified when
461 // forwarding the packet
462 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
463 unsigned int pkt_cmd = pkt->cmdToIndex();
464
465 // a response sees the response latency
466 Tick xbar_delay = responseLatency * clockPeriod();
467
468 // set the packet header and payload delay
469 calcPacketTiming(pkt, xbar_delay);
470
471 // determine how long to be crossbar layer is busy
472 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
473
474 if (snoopFilter && !system->bypassCaches()) {
475 // let the snoop filter inspect the response and update its state
476 snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
477 }
478
479 // send the packet through the destination slave port and pay for
480 // any outstanding header delay
481 Tick latency = pkt->headerDelay;
482 pkt->headerDelay = 0;
483 slavePorts[slave_port_id]->schedTimingResp(pkt, curTick() + latency);
484
485 // remove the request from the routing table
486 routeTo.erase(route_lookup);
487
488 respLayers[slave_port_id]->succeededTiming(packetFinishTime);
489
490 // stats updates
491 pktCount[slave_port_id][master_port_id]++;
492 pktSize[slave_port_id][master_port_id] += pkt_size;
493 transDist[pkt_cmd]++;
494
495 return true;
496}
497
498void
499CoherentXBar::recvTimingSnoopReq(PacketPtr pkt, PortID master_port_id)
500{
501 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
502 masterPorts[master_port_id]->name(), pkt->print());
503
504 // update stats here as we know the forwarding will succeed
505 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
506 transDist[pkt->cmdToIndex()]++;
507 snoops++;
508 snoopTraffic += pkt_size;
509
510 // we should only see express snoops from caches
511 assert(pkt->isExpressSnoop());
512
513 // set the packet header and payload delay, for now use forward latency
514 // @todo Assess the choice of latency further
515 calcPacketTiming(pkt, forwardLatency * clockPeriod());
516
517 // remember if a cache has already committed to responding so we
518 // can see if it changes during the snooping
519 const bool cache_responding = pkt->cacheResponding();
520
521 assert(pkt->snoopDelay == 0);
522
523 if (snoopFilter) {
524 // let the Snoop Filter work its magic and guide probing
525 auto sf_res = snoopFilter->lookupSnoop(pkt);
526 // the time required by a packet to be delivered through
527 // the xbar has to be charged also with to lookup latency
528 // of the snoop filter
529 pkt->headerDelay += sf_res.second * clockPeriod();
530 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n",
531 __func__, masterPorts[master_port_id]->name(), pkt->print(),
532 sf_res.first.size(), sf_res.second);
533
534 // forward to all snoopers
535 forwardTiming(pkt, InvalidPortID, sf_res.first);
536 } else {
537 forwardTiming(pkt, InvalidPortID);
538 }
539
540 // add the snoop delay to our header delay, and then reset it
541 pkt->headerDelay += pkt->snoopDelay;
542 pkt->snoopDelay = 0;
543
544 // if we can expect a response, remember how to route it
545 if (!cache_responding && pkt->cacheResponding()) {
546 assert(routeTo.find(pkt->req) == routeTo.end());
547 routeTo[pkt->req] = master_port_id;
548 }
549
550 // a snoop request came from a connected slave device (one of
551 // our master ports), and if it is not coming from the slave
552 // device responsible for the address range something is
553 // wrong, hence there is nothing further to do as the packet
554 // would be going back to where it came from
|
554 assert(master_port_id == findPort(pkt->getAddr()));
| 555 AddrRange addr_range M5_VAR_USED =
556 RangeSize(pkt->getAddr(), pkt->getSize());
557 assert(findPort(addr_range) == master_port_id);
|
555}
556
557bool
558CoherentXBar::recvTimingSnoopResp(PacketPtr pkt, PortID slave_port_id)
559{
560 // determine the source port based on the id
561 SlavePort* src_port = slavePorts[slave_port_id];
562
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);
568
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();
575
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());
585 return false;
586 }
587 } else {
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());
594 return false;
595 }
596 }
597
598 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
599 src_port->name(), pkt->print());
600
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();
605
606 // responses are never express snoops
607 assert(!pkt->isExpressSnoop());
608
609 // a snoop response sees the snoop response latency, and if it is
610 // forwarded as a normal response, the response latency
611 Tick xbar_delay =
612 (forwardAsSnoop ? snoopResponseLatency : responseLatency) *
613 clockPeriod();
614
615 // set the packet header and payload delay
616 calcPacketTiming(pkt, xbar_delay);
617
618 // determine how long to be crossbar layer is busy
619 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
620
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
626
627 if (snoopFilter) {
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]);
631 }
632
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;
637 assert(success);
638
639 snoopLayers[dest_port_id]->succeededTiming(packetFinishTime);
640 } else {
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);
646
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);
651
652 if (snoopFilter) {
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]);
656 }
657
658 DPRINTF(CoherentXBar, "%s: src %s packet %s FWD RESP\n", __func__,
659 src_port->name(), pkt->print());
660
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
663 // header latency
664 Tick latency = pkt->headerDelay;
665 pkt->headerDelay = 0;
666 slavePorts[dest_port_id]->schedTimingResp(pkt, curTick() + latency);
667
668 respLayers[dest_port_id]->succeededTiming(packetFinishTime);
669 }
670
671 // remove the request from the routing table
672 routeTo.erase(route_lookup);
673
674 // stats updates
675 transDist[pkt_cmd]++;
676 snoops++;
677 snoopTraffic += pkt_size;
678
679 return true;
680}
681
682
683void
684CoherentXBar::forwardTiming(PacketPtr pkt, PortID exclude_slave_port_id,
685 const std::vector<QueuedSlavePort*>& dests)
686{
687 DPRINTF(CoherentXBar, "%s for %s\n", __func__, pkt->print());
688
689 // snoops should only happen if the system isn't bypassing caches
690 assert(!system->bypassCaches());
691
692 unsigned fanout = 0;
693
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
698 // from
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);
703 fanout++;
704 }
705 }
706
707 // Stats for fanout of this forward operation
708 snoopFanout.sample(fanout);
709}
710
711void
712CoherentXBar::recvReqRetry(PortID master_port_id)
713{
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();
718}
719
720Tick
721CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id)
722{
723 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
724 slavePorts[slave_port_id]->name(), pkt->print());
725
726 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
727 unsigned int pkt_cmd = pkt->cmdToIndex();
728
729 MemCmd snoop_response_cmd = MemCmd::InvalidCmd;
730 Tick snoop_response_latency = 0;
731
732 // is this the destination point for this packet? (e.g. true if
733 // this xbar is the PoC for a cache maintenance operation to the
734 // PoC) otherwise the destination is any cache that can satisfy
735 // the request
736 const bool is_destination = isDestination(pkt);
737
738 const bool snoop_caches = !system->bypassCaches() &&
739 pkt->cmd != MemCmd::WriteClean;
740 if (snoop_caches) {
741 // forward to all snoopers but the source
742 std::pair<MemCmd, Tick> snoop_result;
743 if (snoopFilter) {
744 // check with the snoop filter where to forward this packet
745 auto sf_res =
746 snoopFilter->lookupRequest(pkt, *slavePorts[slave_port_id]);
747 snoop_response_latency += sf_res.second * clockPeriod();
748 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n",
749 __func__, slavePorts[slave_port_id]->name(), pkt->print(),
750 sf_res.first.size(), sf_res.second);
751
752 // let the snoop filter know about the success of the send
753 // operation, and do it even before sending it onwards to
754 // avoid situations where atomic upward snoops sneak in
755 // between and change the filter state
756 snoopFilter->finishRequest(false, pkt->getAddr(), pkt->isSecure());
757
758 if (pkt->isEviction()) {
759 // for block-evicting packets, i.e. writebacks and
760 // clean evictions, there is no need to snoop up, as
761 // all we do is determine if the block is cached or
762 // not, instead just set it here based on the snoop
763 // filter result
764 if (!sf_res.first.empty())
765 pkt->setBlockCached();
766 } else {
767 snoop_result = forwardAtomic(pkt, slave_port_id, InvalidPortID,
768 sf_res.first);
769 }
770 } else {
771 snoop_result = forwardAtomic(pkt, slave_port_id);
772 }
773 snoop_response_cmd = snoop_result.first;
774 snoop_response_latency += snoop_result.second;
775 }
776
777 // set up a sensible default value
778 Tick response_latency = 0;
779
780 const bool sink_packet = sinkPacket(pkt);
781
782 // even if we had a snoop response, we must continue and also
783 // perform the actual request at the destination
| 558}
559
560bool
561CoherentXBar::recvTimingSnoopResp(PacketPtr pkt, PortID slave_port_id)
562{
563 // determine the source port based on the id
564 SlavePort* src_port = slavePorts[slave_port_id];
565
566 // get the destination
567 const auto route_lookup = routeTo.find(pkt->req);
568 assert(route_lookup != routeTo.end());
569 const PortID dest_port_id = route_lookup->second;
570 assert(dest_port_id != InvalidPortID);
571
572 // determine if the response is from a snoop request we
573 // created as the result of a normal request (in which case it
574 // should be in the outstandingSnoop), or if we merely forwarded
575 // someone else's snoop request
576 const bool forwardAsSnoop = outstandingSnoop.find(pkt->req) ==
577 outstandingSnoop.end();
578
579 // test if the crossbar should be considered occupied for the
580 // current port, note that the check is bypassed if the response
581 // is being passed on as a normal response since this is occupying
582 // the response layer rather than the snoop response layer
583 if (forwardAsSnoop) {
584 assert(dest_port_id < snoopLayers.size());
585 if (!snoopLayers[dest_port_id]->tryTiming(src_port)) {
586 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__,
587 src_port->name(), pkt->print());
588 return false;
589 }
590 } else {
591 // get the master port that mirrors this slave port internally
592 MasterPort* snoop_port = snoopRespPorts[slave_port_id];
593 assert(dest_port_id < respLayers.size());
594 if (!respLayers[dest_port_id]->tryTiming(snoop_port)) {
595 DPRINTF(CoherentXBar, "%s: src %s packet %s BUSY\n", __func__,
596 snoop_port->name(), pkt->print());
597 return false;
598 }
599 }
600
601 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
602 src_port->name(), pkt->print());
603
604 // store size and command as they might be modified when
605 // forwarding the packet
606 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
607 unsigned int pkt_cmd = pkt->cmdToIndex();
608
609 // responses are never express snoops
610 assert(!pkt->isExpressSnoop());
611
612 // a snoop response sees the snoop response latency, and if it is
613 // forwarded as a normal response, the response latency
614 Tick xbar_delay =
615 (forwardAsSnoop ? snoopResponseLatency : responseLatency) *
616 clockPeriod();
617
618 // set the packet header and payload delay
619 calcPacketTiming(pkt, xbar_delay);
620
621 // determine how long to be crossbar layer is busy
622 Tick packetFinishTime = clockEdge(Cycles(1)) + pkt->payloadDelay;
623
624 // forward it either as a snoop response or a normal response
625 if (forwardAsSnoop) {
626 // this is a snoop response to a snoop request we forwarded,
627 // e.g. coming from the L1 and going to the L2, and it should
628 // be forwarded as a snoop response
629
630 if (snoopFilter) {
631 // update the probe filter so that it can properly track the line
632 snoopFilter->updateSnoopForward(pkt, *slavePorts[slave_port_id],
633 *masterPorts[dest_port_id]);
634 }
635
636 bool success M5_VAR_USED =
637 masterPorts[dest_port_id]->sendTimingSnoopResp(pkt);
638 pktCount[slave_port_id][dest_port_id]++;
639 pktSize[slave_port_id][dest_port_id] += pkt_size;
640 assert(success);
641
642 snoopLayers[dest_port_id]->succeededTiming(packetFinishTime);
643 } else {
644 // we got a snoop response on one of our slave ports,
645 // i.e. from a coherent master connected to the crossbar, and
646 // since we created the snoop request as part of recvTiming,
647 // this should now be a normal response again
648 outstandingSnoop.erase(pkt->req);
649
650 // this is a snoop response from a coherent master, hence it
651 // should never go back to where the snoop response came from,
652 // but instead to where the original request came from
653 assert(slave_port_id != dest_port_id);
654
655 if (snoopFilter) {
656 // update the probe filter so that it can properly track the line
657 snoopFilter->updateSnoopResponse(pkt, *slavePorts[slave_port_id],
658 *slavePorts[dest_port_id]);
659 }
660
661 DPRINTF(CoherentXBar, "%s: src %s packet %s FWD RESP\n", __func__,
662 src_port->name(), pkt->print());
663
664 // as a normal response, it should go back to a master through
665 // one of our slave ports, we also pay for any outstanding
666 // header latency
667 Tick latency = pkt->headerDelay;
668 pkt->headerDelay = 0;
669 slavePorts[dest_port_id]->schedTimingResp(pkt, curTick() + latency);
670
671 respLayers[dest_port_id]->succeededTiming(packetFinishTime);
672 }
673
674 // remove the request from the routing table
675 routeTo.erase(route_lookup);
676
677 // stats updates
678 transDist[pkt_cmd]++;
679 snoops++;
680 snoopTraffic += pkt_size;
681
682 return true;
683}
684
685
686void
687CoherentXBar::forwardTiming(PacketPtr pkt, PortID exclude_slave_port_id,
688 const std::vector<QueuedSlavePort*>& dests)
689{
690 DPRINTF(CoherentXBar, "%s for %s\n", __func__, pkt->print());
691
692 // snoops should only happen if the system isn't bypassing caches
693 assert(!system->bypassCaches());
694
695 unsigned fanout = 0;
696
697 for (const auto& p: dests) {
698 // we could have gotten this request from a snooping master
699 // (corresponding to our own slave port that is also in
700 // snoopPorts) and should not send it back to where it came
701 // from
702 if (exclude_slave_port_id == InvalidPortID ||
703 p->getId() != exclude_slave_port_id) {
704 // cache is not allowed to refuse snoop
705 p->sendTimingSnoopReq(pkt);
706 fanout++;
707 }
708 }
709
710 // Stats for fanout of this forward operation
711 snoopFanout.sample(fanout);
712}
713
714void
715CoherentXBar::recvReqRetry(PortID master_port_id)
716{
717 // responses and snoop responses never block on forwarding them,
718 // so the retry will always be coming from a port to which we
719 // tried to forward a request
720 reqLayers[master_port_id]->recvRetry();
721}
722
723Tick
724CoherentXBar::recvAtomic(PacketPtr pkt, PortID slave_port_id)
725{
726 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
727 slavePorts[slave_port_id]->name(), pkt->print());
728
729 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
730 unsigned int pkt_cmd = pkt->cmdToIndex();
731
732 MemCmd snoop_response_cmd = MemCmd::InvalidCmd;
733 Tick snoop_response_latency = 0;
734
735 // is this the destination point for this packet? (e.g. true if
736 // this xbar is the PoC for a cache maintenance operation to the
737 // PoC) otherwise the destination is any cache that can satisfy
738 // the request
739 const bool is_destination = isDestination(pkt);
740
741 const bool snoop_caches = !system->bypassCaches() &&
742 pkt->cmd != MemCmd::WriteClean;
743 if (snoop_caches) {
744 // forward to all snoopers but the source
745 std::pair<MemCmd, Tick> snoop_result;
746 if (snoopFilter) {
747 // check with the snoop filter where to forward this packet
748 auto sf_res =
749 snoopFilter->lookupRequest(pkt, *slavePorts[slave_port_id]);
750 snoop_response_latency += sf_res.second * clockPeriod();
751 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n",
752 __func__, slavePorts[slave_port_id]->name(), pkt->print(),
753 sf_res.first.size(), sf_res.second);
754
755 // let the snoop filter know about the success of the send
756 // operation, and do it even before sending it onwards to
757 // avoid situations where atomic upward snoops sneak in
758 // between and change the filter state
759 snoopFilter->finishRequest(false, pkt->getAddr(), pkt->isSecure());
760
761 if (pkt->isEviction()) {
762 // for block-evicting packets, i.e. writebacks and
763 // clean evictions, there is no need to snoop up, as
764 // all we do is determine if the block is cached or
765 // not, instead just set it here based on the snoop
766 // filter result
767 if (!sf_res.first.empty())
768 pkt->setBlockCached();
769 } else {
770 snoop_result = forwardAtomic(pkt, slave_port_id, InvalidPortID,
771 sf_res.first);
772 }
773 } else {
774 snoop_result = forwardAtomic(pkt, slave_port_id);
775 }
776 snoop_response_cmd = snoop_result.first;
777 snoop_response_latency += snoop_result.second;
778 }
779
780 // set up a sensible default value
781 Tick response_latency = 0;
782
783 const bool sink_packet = sinkPacket(pkt);
784
785 // even if we had a snoop response, we must continue and also
786 // perform the actual request at the destination
|
784 PortID master_port_id = findPort(pkt->getAddr());
| 787 AddrRange addr_range = RangeSize(pkt->getAddr(), pkt->getSize());
788 PortID master_port_id = findPort(addr_range);
|
785
786 if (sink_packet) {
787 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__,
788 pkt->print());
789 } else {
790 if (forwardPacket(pkt)) {
791 // make sure that the write request (e.g., WriteClean)
792 // will stop at the memory below if this crossbar is its
793 // destination
794 if (pkt->isWrite() && is_destination) {
795 pkt->clearWriteThrough();
796 }
797
798 // forward the request to the appropriate destination
799 response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
800 } else {
801 // if it does not need a response we sink the packet above
802 assert(pkt->needsResponse());
803
804 pkt->makeResponse();
805 }
806 }
807
808 // stats updates for the request
809 pktCount[slave_port_id][master_port_id]++;
810 pktSize[slave_port_id][master_port_id] += pkt_size;
811 transDist[pkt_cmd]++;
812
813
814 // if lower levels have replied, tell the snoop filter
815 if (!system->bypassCaches() && snoopFilter && pkt->isResponse()) {
816 snoopFilter->updateResponse(pkt, *slavePorts[slave_port_id]);
817 }
818
819 // if we got a response from a snooper, restore it here
820 if (snoop_response_cmd != MemCmd::InvalidCmd) {
821 // no one else should have responded
822 assert(!pkt->isResponse());
823 pkt->cmd = snoop_response_cmd;
824 response_latency = snoop_response_latency;
825 }
826
827 // If this is the destination of the cache clean operation the
828 // crossbar is responsible for responding. This crossbar will
829 // respond when the cache clean is complete. An atomic cache clean
830 // is complete when the crossbars receives the cache clean
831 // request (CleanSharedReq, CleanInvalidReq), as either:
832 // * no cache above had a dirty copy of the block as indicated by
833 // the satisfied flag of the packet, or
834 // * the crossbar has already seen the corresponding write
835 // (WriteClean) which updates the block in the memory below.
836 if (pkt->isClean() && isDestination(pkt) && pkt->satisfied()) {
837 auto it = outstandingCMO.find(pkt->id);
838 assert(it != outstandingCMO.end());
839 // we are responding right away
840 outstandingCMO.erase(it);
841 } else if (pkt->cmd == MemCmd::WriteClean && isDestination(pkt)) {
842 // if this is the destination of the operation, the xbar
843 // sends the responce to the cache clean operation only
844 // after having encountered the cache clean request
845 auto M5_VAR_USED ret = outstandingCMO.emplace(pkt->id, nullptr);
846 // in atomic mode we know that the WriteClean packet should
847 // precede the clean request
848 assert(ret.second);
849 }
850
851 // add the response data
852 if (pkt->isResponse()) {
853 pkt_size = pkt->hasData() ? pkt->getSize() : 0;
854 pkt_cmd = pkt->cmdToIndex();
855
856 // stats updates
857 pktCount[slave_port_id][master_port_id]++;
858 pktSize[slave_port_id][master_port_id] += pkt_size;
859 transDist[pkt_cmd]++;
860 }
861
862 // @todo: Not setting header time
863 pkt->payloadDelay = response_latency;
864 return response_latency;
865}
866
867Tick
868CoherentXBar::recvAtomicSnoop(PacketPtr pkt, PortID master_port_id)
869{
870 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
871 masterPorts[master_port_id]->name(), pkt->print());
872
873 // add the request snoop data
874 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
875 snoops++;
876 snoopTraffic += pkt_size;
877
878 // forward to all snoopers
879 std::pair<MemCmd, Tick> snoop_result;
880 Tick snoop_response_latency = 0;
881 if (snoopFilter) {
882 auto sf_res = snoopFilter->lookupSnoop(pkt);
883 snoop_response_latency += sf_res.second * clockPeriod();
884 DPRINTF(CoherentXBar, "%s: src %s packet %s SF size: %i lat: %i\n",
885 __func__, masterPorts[master_port_id]->name(), pkt->print(),
886 sf_res.first.size(), sf_res.second);
887 snoop_result = forwardAtomic(pkt, InvalidPortID, master_port_id,
888 sf_res.first);
889 } else {
890 snoop_result = forwardAtomic(pkt, InvalidPortID);
891 }
892 MemCmd snoop_response_cmd = snoop_result.first;
893 snoop_response_latency += snoop_result.second;
894
895 if (snoop_response_cmd != MemCmd::InvalidCmd)
896 pkt->cmd = snoop_response_cmd;
897
898 // add the response snoop data
899 if (pkt->isResponse()) {
900 snoops++;
901 }
902
903 // @todo: Not setting header time
904 pkt->payloadDelay = snoop_response_latency;
905 return snoop_response_latency;
906}
907
908std::pair<MemCmd, Tick>
909CoherentXBar::forwardAtomic(PacketPtr pkt, PortID exclude_slave_port_id,
910 PortID source_master_port_id,
911 const std::vector<QueuedSlavePort*>& dests)
912{
913 // the packet may be changed on snoops, record the original
914 // command to enable us to restore it between snoops so that
915 // additional snoops can take place properly
916 MemCmd orig_cmd = pkt->cmd;
917 MemCmd snoop_response_cmd = MemCmd::InvalidCmd;
918 Tick snoop_response_latency = 0;
919
920 // snoops should only happen if the system isn't bypassing caches
921 assert(!system->bypassCaches());
922
923 unsigned fanout = 0;
924
925 for (const auto& p: dests) {
926 // we could have gotten this request from a snooping master
927 // (corresponding to our own slave port that is also in
928 // snoopPorts) and should not send it back to where it came
929 // from
930 if (exclude_slave_port_id != InvalidPortID &&
931 p->getId() == exclude_slave_port_id)
932 continue;
933
934 Tick latency = p->sendAtomicSnoop(pkt);
935 fanout++;
936
937 // in contrast to a functional access, we have to keep on
938 // going as all snoopers must be updated even if we get a
939 // response
940 if (!pkt->isResponse())
941 continue;
942
943 // response from snoop agent
944 assert(pkt->cmd != orig_cmd);
945 assert(pkt->cacheResponding());
946 // should only happen once
947 assert(snoop_response_cmd == MemCmd::InvalidCmd);
948 // save response state
949 snoop_response_cmd = pkt->cmd;
950 snoop_response_latency = latency;
951
952 if (snoopFilter) {
953 // Handle responses by the snoopers and differentiate between
954 // responses to requests from above and snoops from below
955 if (source_master_port_id != InvalidPortID) {
956 // Getting a response for a snoop from below
957 assert(exclude_slave_port_id == InvalidPortID);
958 snoopFilter->updateSnoopForward(pkt, *p,
959 *masterPorts[source_master_port_id]);
960 } else {
961 // Getting a response for a request from above
962 assert(source_master_port_id == InvalidPortID);
963 snoopFilter->updateSnoopResponse(pkt, *p,
964 *slavePorts[exclude_slave_port_id]);
965 }
966 }
967 // restore original packet state for remaining snoopers
968 pkt->cmd = orig_cmd;
969 }
970
971 // Stats for fanout
972 snoopFanout.sample(fanout);
973
974 // the packet is restored as part of the loop and any potential
975 // snoop response is part of the returned pair
976 return std::make_pair(snoop_response_cmd, snoop_response_latency);
977}
978
979void
980CoherentXBar::recvFunctional(PacketPtr pkt, PortID slave_port_id)
981{
982 if (!pkt->isPrint()) {
983 // don't do DPRINTFs on PrintReq as it clutters up the output
984 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
985 slavePorts[slave_port_id]->name(), pkt->print());
986 }
987
988 if (!system->bypassCaches()) {
989 // forward to all snoopers but the source
990 forwardFunctional(pkt, slave_port_id);
991 }
992
993 // there is no need to continue if the snooping has found what we
994 // were looking for and the packet is already a response
995 if (!pkt->isResponse()) {
996 // since our slave ports are queued ports we need to check them as well
997 for (const auto& p : slavePorts) {
998 // if we find a response that has the data, then the
999 // downstream caches/memories may be out of date, so simply stop
1000 // here
1001 if (p->checkFunctional(pkt)) {
1002 if (pkt->needsResponse())
1003 pkt->makeResponse();
1004 return;
1005 }
1006 }
1007
| 789
790 if (sink_packet) {
791 DPRINTF(CoherentXBar, "%s: Not forwarding %s\n", __func__,
792 pkt->print());
793 } else {
794 if (forwardPacket(pkt)) {
795 // make sure that the write request (e.g., WriteClean)
796 // will stop at the memory below if this crossbar is its
797 // destination
798 if (pkt->isWrite() && is_destination) {
799 pkt->clearWriteThrough();
800 }
801
802 // forward the request to the appropriate destination
803 response_latency = masterPorts[master_port_id]->sendAtomic(pkt);
804 } else {
805 // if it does not need a response we sink the packet above
806 assert(pkt->needsResponse());
807
808 pkt->makeResponse();
809 }
810 }
811
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]++;
816
817
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]);
821 }
822
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;
829 }
830
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
852 assert(ret.second);
853 }
854
855 // add the response data
856 if (pkt->isResponse()) {
857 pkt_size = pkt->hasData() ? pkt->getSize() : 0;
858 pkt_cmd = pkt->cmdToIndex();
859
860 // stats updates
861 pktCount[slave_port_id][master_port_id]++;
862 pktSize[slave_port_id][master_port_id] += pkt_size;
863 transDist[pkt_cmd]++;
864 }
865
866 // @todo: Not setting header time
867 pkt->payloadDelay = response_latency;
868 return response_latency;
869}
870
871Tick
872CoherentXBar::recvAtomicSnoop(PacketPtr pkt, PortID master_port_id)
873{
874 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
875 masterPorts[master_port_id]->name(), pkt->print());
876
877 // add the request snoop data
878 unsigned int pkt_size = pkt->hasData() ? pkt->getSize() : 0;
879 snoops++;
880 snoopTraffic += pkt_size;
881
882 // forward to all snoopers
883 std::pair<MemCmd, Tick> snoop_result;
884 Tick snoop_response_latency = 0;
885 if (snoopFilter) {
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,
892 sf_res.first);
893 } else {
894 snoop_result = forwardAtomic(pkt, InvalidPortID);
895 }
896 MemCmd snoop_response_cmd = snoop_result.first;
897 snoop_response_latency += snoop_result.second;
898
899 if (snoop_response_cmd != MemCmd::InvalidCmd)
900 pkt->cmd = snoop_response_cmd;
901
902 // add the response snoop data
903 if (pkt->isResponse()) {
904 snoops++;
905 }
906
907 // @todo: Not setting header time
908 pkt->payloadDelay = snoop_response_latency;
909 return snoop_response_latency;
910}
911
912std::pair<MemCmd, Tick>
913CoherentXBar::forwardAtomic(PacketPtr pkt, PortID exclude_slave_port_id,
914 PortID source_master_port_id,
915 const std::vector<QueuedSlavePort*>& dests)
916{
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;
923
924 // snoops should only happen if the system isn't bypassing caches
925 assert(!system->bypassCaches());
926
927 unsigned fanout = 0;
928
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
933 // from
934 if (exclude_slave_port_id != InvalidPortID &&
935 p->getId() == exclude_slave_port_id)
936 continue;
937
938 Tick latency = p->sendAtomicSnoop(pkt);
939 fanout++;
940
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
943 // response
944 if (!pkt->isResponse())
945 continue;
946
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;
955
956 if (snoopFilter) {
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]);
964 } else {
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]);
969 }
970 }
971 // restore original packet state for remaining snoopers
972 pkt->cmd = orig_cmd;
973 }
974
975 // Stats for fanout
976 snoopFanout.sample(fanout);
977
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);
981}
982
983void
984CoherentXBar::recvFunctional(PacketPtr pkt, PortID slave_port_id)
985{
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());
990 }
991
992 if (!system->bypassCaches()) {
993 // forward to all snoopers but the source
994 forwardFunctional(pkt, slave_port_id);
995 }
996
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
1004 // here
1005 if (p->checkFunctional(pkt)) {
1006 if (pkt->needsResponse())
1007 pkt->makeResponse();
1008 return;
1009 }
1010 }
1011
|
1008 PortID dest_id = findPort(pkt->getAddr());
| 1012 PortID dest_id = findPort(RangeSize(pkt->getAddr(), pkt->getSize()));
|
1009
1010 masterPorts[dest_id]->sendFunctional(pkt);
1011 }
1012}
1013
1014void
1015CoherentXBar::recvFunctionalSnoop(PacketPtr pkt, PortID master_port_id)
1016{
1017 if (!pkt->isPrint()) {
1018 // don't do DPRINTFs on PrintReq as it clutters up the output
1019 DPRINTF(CoherentXBar, "%s: src %s packet %s\n", __func__,
1020 masterPorts[master_port_id]->name(), pkt->print());
1021 }
1022
1023 for (const auto& p : slavePorts) {
1024 if (p->checkFunctional(pkt)) {
1025 if (pkt->needsResponse())
1026 pkt->makeResponse();
1027 return;
1028 }
1029 }
1030
1031 // forward to all snoopers
1032 forwardFunctional(pkt, InvalidPortID);
1033}
1034
1035void
1036CoherentXBar::forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id)
1037{
1038 // snoops should only happen if the system isn't bypassing caches
1039 assert(!system->bypassCaches());
1040
1041 for (const auto& p: snoopPorts) {
1042 // we could have gotten this request from a snooping master
1043 // (corresponding to our own slave port that is also in
1044 // snoopPorts) and should not send it back to where it came
1045 // from
1046 if (exclude_slave_port_id == InvalidPortID ||
1047 p->getId() != exclude_slave_port_id)
1048 p->sendFunctionalSnoop(pkt);
1049
1050 // if we get a response we are done
1051 if (pkt->isResponse()) {
1052 break;
1053 }
1054 }
1055}
1056
1057bool
1058CoherentXBar::sinkPacket(const PacketPtr pkt) const
1059{
1060 // we can sink the packet if:
1061 // 1) the crossbar is the point of coherency, and a cache is
1062 // responding after being snooped
1063 // 2) the crossbar is the point of coherency, and the packet is a
1064 // coherency packet (not a read or a write) that does not
1065 // require a response
1066 // 3) this is a clean evict or clean writeback, but the packet is
1067 // found in a cache above this crossbar
1068 // 4) a cache is responding after being snooped, and the packet
1069 // either does not need the block to be writable, or the cache
1070 // that has promised to respond (setting the cache responding
1071 // flag) is providing writable and thus had a Modified block,
1072 // and no further action is needed
1073 return (pointOfCoherency && pkt->cacheResponding()) ||
1074 (pointOfCoherency && !(pkt->isRead() || pkt->isWrite()) &&
1075 !pkt->needsResponse()) ||
1076 (pkt->isCleanEviction() && pkt->isBlockCached()) ||
1077 (pkt->cacheResponding() &&
1078 (!pkt->needsWritable() || pkt->responderHadWritable()));
1079}
1080
1081bool
1082CoherentXBar::forwardPacket(const PacketPtr pkt)
1083{
1084 // we are forwarding the packet if:
1085 // 1) this is a cache clean request to the PoU/PoC and this
1086 // crossbar is above the PoU/PoC
1087 // 2) this is a read or a write
1088 // 3) this crossbar is above the point of coherency
1089 if (pkt->isClean()) {
1090 return !isDestination(pkt);
1091 }
1092 return pkt->isRead() || pkt->isWrite() || !pointOfCoherency;
1093}
1094
1095
1096void
1097CoherentXBar::regStats()
1098{
1099 // register the stats of the base class and our layers
1100 BaseXBar::regStats();
1101 for (auto l: reqLayers)
1102 l->regStats();
1103 for (auto l: respLayers)
1104 l->regStats();
1105 for (auto l: snoopLayers)
1106 l->regStats();
1107
1108 snoops
1109 .name(name() + ".snoops")
1110 .desc("Total snoops (count)")
1111 ;
1112
1113 snoopTraffic
1114 .name(name() + ".snoopTraffic")
1115 .desc("Total snoop traffic (bytes)")
1116 ;
1117
1118 snoopFanout
1119 .init(0, snoopPorts.size(), 1)
1120 .name(name() + ".snoop_fanout")
1121 .desc("Request fanout histogram")
1122 ;
1123}
1124
1125CoherentXBar *
1126CoherentXBarParams::create()
1127{
1128 return new CoherentXBar(this);
1129}
| 1013
1014 masterPorts[dest_id]->sendFunctional(pkt);
1015 }
1016}
1017
1018void
1019CoherentXBar::recvFunctionalSnoop(PacketPtr pkt, PortID master_port_id)
1020{
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());
1025 }
1026
1027 for (const auto& p : slavePorts) {
1028 if (p->checkFunctional(pkt)) {
1029 if (pkt->needsResponse())
1030 pkt->makeResponse();
1031 return;
1032 }
1033 }
1034
1035 // forward to all snoopers
1036 forwardFunctional(pkt, InvalidPortID);
1037}
1038
1039void
1040CoherentXBar::forwardFunctional(PacketPtr pkt, PortID exclude_slave_port_id)
1041{
1042 // snoops should only happen if the system isn't bypassing caches
1043 assert(!system->bypassCaches());
1044
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
1049 // from
1050 if (exclude_slave_port_id == InvalidPortID ||
1051 p->getId() != exclude_slave_port_id)
1052 p->sendFunctionalSnoop(pkt);
1053
1054 // if we get a response we are done
1055 if (pkt->isResponse()) {
1056 break;
1057 }
1058 }
1059}
1060
1061bool
1062CoherentXBar::sinkPacket(const PacketPtr pkt) const
1063{
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()));
1083}
1084
1085bool
1086CoherentXBar::forwardPacket(const PacketPtr pkt)
1087{
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);
1095 }
1096 return pkt->isRead() || pkt->isWrite() || !pointOfCoherency;
1097}
1098
1099
1100void
1101CoherentXBar::regStats()
1102{
1103 // register the stats of the base class and our layers
1104 BaseXBar::regStats();
1105 for (auto l: reqLayers)
1106 l->regStats();
1107 for (auto l: respLayers)
1108 l->regStats();
1109 for (auto l: snoopLayers)
1110 l->regStats();
1111
1112 snoops
1113 .name(name() + ".snoops")
1114 .desc("Total snoops (count)")
1115 ;
1116
1117 snoopTraffic
1118 .name(name() + ".snoopTraffic")
1119 .desc("Total snoop traffic (bytes)")
1120 ;
1121
1122 snoopFanout
1123 .init(0, snoopPorts.size(), 1)
1124 .name(name() + ".snoop_fanout")
1125 .desc("Request fanout histogram")
1126 ;
1127}
1128
1129CoherentXBar *
1130CoherentXBarParams::create()
1131{
1132 return new CoherentXBar(this);
1133}
|