1/*
2 * Copyright (c) 2011-2015, 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 */
44
45/**
46 * @file
47 * Definition of a crossbar object.
48 */
49
50#include "mem/xbar.hh"
51
52#include "base/logging.hh"
53#include "base/trace.hh"
54#include "debug/AddrRanges.hh"
55#include "debug/Drain.hh"
56#include "debug/XBar.hh"
57
58BaseXBar::BaseXBar(const BaseXBarParams *p)
| 1/*
2 * Copyright (c) 2011-2015, 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 */
44
45/**
46 * @file
47 * Definition of a crossbar object.
48 */
49
50#include "mem/xbar.hh"
51
52#include "base/logging.hh"
53#include "base/trace.hh"
54#include "debug/AddrRanges.hh"
55#include "debug/Drain.hh"
56#include "debug/XBar.hh"
57
58BaseXBar::BaseXBar(const BaseXBarParams *p)
|
92 }
93}
94
95void
96BaseXBar::calcPacketTiming(PacketPtr pkt, Tick header_delay)
97{
98 // the crossbar will be called at a time that is not necessarily
99 // coinciding with its own clock, so start by determining how long
100 // until the next clock edge (could be zero)
101 Tick offset = clockEdge() - curTick();
102
103 // the header delay depends on the path through the crossbar, and
104 // we therefore rely on the caller to provide the actual
105 // value
106 pkt->headerDelay += offset + header_delay;
107
108 // note that we add the header delay to the existing value, and
109 // align it to the crossbar clock
110
111 // do a quick sanity check to ensure the timings are not being
112 // ignored, note that this specific value may cause problems for
113 // slower interconnects
114 panic_if(pkt->headerDelay > SimClock::Int::us,
115 "Encountered header delay exceeding 1 us\n");
116
117 if (pkt->hasData()) {
118 // the payloadDelay takes into account the relative time to
119 // deliver the payload of the packet, after the header delay,
120 // we take the maximum since the payload delay could already
121 // be longer than what this parcitular crossbar enforces.
122 pkt->payloadDelay = std::max<Tick>(pkt->payloadDelay,
123 divCeil(pkt->getSize(), width) *
124 clockPeriod());
125 }
126
127 // the payload delay is not paying for the clock offset as that is
128 // already done using the header delay, and the payload delay is
129 // also used to determine how long the crossbar layer is busy and
130 // thus regulates throughput
131}
132
133template <typename SrcType, typename DstType>
134BaseXBar::Layer<SrcType, DstType>::Layer(DstType& _port, BaseXBar& _xbar,
135 const std::string& _name) :
136 port(_port), xbar(_xbar), _name(_name), state(IDLE),
137 waitingForPeer(NULL), releaseEvent([this]{ releaseLayer(); }, name())
138{
139}
140
141template <typename SrcType, typename DstType>
142void BaseXBar::Layer<SrcType, DstType>::occupyLayer(Tick until)
143{
144 // ensure the state is busy at this point, as the layer should
145 // transition from idle as soon as it has decided to forward the
146 // packet to prevent any follow-on calls to sendTiming seeing an
147 // unoccupied layer
148 assert(state == BUSY);
149
150 // until should never be 0 as express snoops never occupy the layer
151 assert(until != 0);
152 xbar.schedule(releaseEvent, until);
153
154 // account for the occupied ticks
155 occupancy += until - curTick();
156
157 DPRINTF(BaseXBar, "The crossbar layer is now busy from tick %d to %d\n",
158 curTick(), until);
159}
160
161template <typename SrcType, typename DstType>
162bool
163BaseXBar::Layer<SrcType, DstType>::tryTiming(SrcType* src_port)
164{
165 // if we are in the retry state, we will not see anything but the
166 // retrying port (or in the case of the snoop ports the snoop
167 // response port that mirrors the actual slave port) as we leave
168 // this state again in zero time if the peer does not immediately
169 // call the layer when receiving the retry
170
171 // first we see if the layer is busy, next we check if the
172 // destination port is already engaged in a transaction waiting
173 // for a retry from the peer
174 if (state == BUSY || waitingForPeer != NULL) {
175 // the port should not be waiting already
176 assert(std::find(waitingForLayer.begin(), waitingForLayer.end(),
177 src_port) == waitingForLayer.end());
178
179 // put the port at the end of the retry list waiting for the
180 // layer to be freed up (and in the case of a busy peer, for
181 // that transaction to go through, and then the layer to free
182 // up)
183 waitingForLayer.push_back(src_port);
184 return false;
185 }
186
187 state = BUSY;
188
189 return true;
190}
191
192template <typename SrcType, typename DstType>
193void
194BaseXBar::Layer<SrcType, DstType>::succeededTiming(Tick busy_time)
195{
196 // we should have gone from idle or retry to busy in the tryTiming
197 // test
198 assert(state == BUSY);
199
200 // occupy the layer accordingly
201 occupyLayer(busy_time);
202}
203
204template <typename SrcType, typename DstType>
205void
206BaseXBar::Layer<SrcType, DstType>::failedTiming(SrcType* src_port,
207 Tick busy_time)
208{
209 // ensure no one got in between and tried to send something to
210 // this port
211 assert(waitingForPeer == NULL);
212
213 // if the source port is the current retrying one or not, we have
214 // failed in forwarding and should track that we are now waiting
215 // for the peer to send a retry
216 waitingForPeer = src_port;
217
218 // we should have gone from idle or retry to busy in the tryTiming
219 // test
220 assert(state == BUSY);
221
222 // occupy the bus accordingly
223 occupyLayer(busy_time);
224}
225
226template <typename SrcType, typename DstType>
227void
228BaseXBar::Layer<SrcType, DstType>::releaseLayer()
229{
230 // releasing the bus means we should now be idle
231 assert(state == BUSY);
232 assert(!releaseEvent.scheduled());
233
234 // update the state
235 state = IDLE;
236
237 // bus layer is now idle, so if someone is waiting we can retry
238 if (!waitingForLayer.empty()) {
239 // there is no point in sending a retry if someone is still
240 // waiting for the peer
241 if (waitingForPeer == NULL)
242 retryWaiting();
243 } else if (waitingForPeer == NULL && drainState() == DrainState::Draining) {
244 DPRINTF(Drain, "Crossbar done draining, signaling drain manager\n");
245 //If we weren't able to drain before, do it now.
246 signalDrainDone();
247 }
248}
249
250template <typename SrcType, typename DstType>
251void
252BaseXBar::Layer<SrcType, DstType>::retryWaiting()
253{
254 // this should never be called with no one waiting
255 assert(!waitingForLayer.empty());
256
257 // we always go to retrying from idle
258 assert(state == IDLE);
259
260 // update the state
261 state = RETRY;
262
263 // set the retrying port to the front of the retry list and pop it
264 // off the list
265 SrcType* retryingPort = waitingForLayer.front();
266 waitingForLayer.pop_front();
267
268 // tell the port to retry, which in some cases ends up calling the
269 // layer again
270 sendRetry(retryingPort);
271
272 // If the layer is still in the retry state, sendTiming wasn't
273 // called in zero time (e.g. the cache does this when a writeback
274 // is squashed)
275 if (state == RETRY) {
276 // update the state to busy and reset the retrying port, we
277 // have done our bit and sent the retry
278 state = BUSY;
279
280 // occupy the crossbar layer until the next clock edge
281 occupyLayer(xbar.clockEdge());
282 }
283}
284
285template <typename SrcType, typename DstType>
286void
287BaseXBar::Layer<SrcType, DstType>::recvRetry()
288{
289 // we should never get a retry without having failed to forward
290 // something to this port
291 assert(waitingForPeer != NULL);
292
293 // add the port where the failed packet originated to the front of
294 // the waiting ports for the layer, this allows us to call retry
295 // on the port immediately if the crossbar layer is idle
296 waitingForLayer.push_front(waitingForPeer);
297
298 // we are no longer waiting for the peer
299 waitingForPeer = NULL;
300
301 // if the layer is idle, retry this port straight away, if we
302 // are busy, then simply let the port wait for its turn
303 if (state == IDLE) {
304 retryWaiting();
305 } else {
306 assert(state == BUSY);
307 }
308}
309
310PortID
311BaseXBar::findPort(AddrRange addr_range)
312{
313 // we should never see any address lookups before we've got the
314 // ranges of all connected slave modules
315 assert(gotAllAddrRanges);
316
317 // Check the address map interval tree
318 auto i = portMap.contains(addr_range);
319 if (i != portMap.end()) {
320 return i->second;
321 }
322
323 // Check if this matches the default range
324 if (useDefaultRange) {
325 if (addr_range.isSubset(defaultRange)) {
326 DPRINTF(AddrRanges, " found addr %s on default\n",
327 addr_range.to_string());
328 return defaultPortID;
329 }
330 } else if (defaultPortID != InvalidPortID) {
331 DPRINTF(AddrRanges, "Unable to find destination for %s, "
332 "will use default port\n", addr_range.to_string());
333 return defaultPortID;
334 }
335
336 // we should use the range for the default port and it did not
337 // match, or the default port is not set
338 fatal("Unable to find destination for %s on %s\n", addr_range.to_string(),
339 name());
340}
341
342/** Function called by the port when the crossbar is receiving a range change.*/
343void
344BaseXBar::recvRangeChange(PortID master_port_id)
345{
346 DPRINTF(AddrRanges, "Received range change from slave port %s\n",
347 masterPorts[master_port_id]->getSlavePort().name());
348
349 // remember that we got a range from this master port and thus the
350 // connected slave module
351 gotAddrRanges[master_port_id] = true;
352
353 // update the global flag
354 if (!gotAllAddrRanges) {
355 // take a logical AND of all the ports and see if we got
356 // ranges from everyone
357 gotAllAddrRanges = true;
358 std::vector<bool>::const_iterator r = gotAddrRanges.begin();
359 while (gotAllAddrRanges && r != gotAddrRanges.end()) {
360 gotAllAddrRanges &= *r++;
361 }
362 if (gotAllAddrRanges)
363 DPRINTF(AddrRanges, "Got address ranges from all slaves\n");
364 }
365
366 // note that we could get the range from the default port at any
367 // point in time, and we cannot assume that the default range is
368 // set before the other ones are, so we do additional checks once
369 // all ranges are provided
370 if (master_port_id == defaultPortID) {
371 // only update if we are indeed checking ranges for the
372 // default port since the port might not have a valid range
373 // otherwise
374 if (useDefaultRange) {
375 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
376
377 if (ranges.size() != 1)
378 fatal("Crossbar %s may only have a single default range",
379 name());
380
381 defaultRange = ranges.front();
382 }
383 } else {
384 // the ports are allowed to update their address ranges
385 // dynamically, so remove any existing entries
386 if (gotAddrRanges[master_port_id]) {
387 for (auto p = portMap.begin(); p != portMap.end(); ) {
388 if (p->second == master_port_id)
389 // erasing invalidates the iterator, so advance it
390 // before the deletion takes place
391 portMap.erase(p++);
392 else
393 p++;
394 }
395 }
396
397 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
398
399 for (const auto& r: ranges) {
400 DPRINTF(AddrRanges, "Adding range %s for id %d\n",
401 r.to_string(), master_port_id);
402 if (portMap.insert(r, master_port_id) == portMap.end()) {
403 PortID conflict_id = portMap.intersects(r)->second;
404 fatal("%s has two ports responding within range "
405 "%s:\n\t%s\n\t%s\n",
406 name(),
407 r.to_string(),
408 masterPorts[master_port_id]->getSlavePort().name(),
409 masterPorts[conflict_id]->getSlavePort().name());
410 }
411 }
412 }
413
414 // if we have received ranges from all our neighbouring slave
415 // modules, go ahead and tell our connected master modules in
416 // turn, this effectively assumes a tree structure of the system
417 if (gotAllAddrRanges) {
418 DPRINTF(AddrRanges, "Aggregating address ranges\n");
419 xbarRanges.clear();
420
421 // start out with the default range
422 if (useDefaultRange) {
423 if (!gotAddrRanges[defaultPortID])
424 fatal("Crossbar %s uses default range, but none provided",
425 name());
426
427 xbarRanges.push_back(defaultRange);
428 DPRINTF(AddrRanges, "-- Adding default %s\n",
429 defaultRange.to_string());
430 }
431
432 // merge all interleaved ranges and add any range that is not
433 // a subset of the default range
434 std::vector<AddrRange> intlv_ranges;
435 for (const auto& r: portMap) {
436 // if the range is interleaved then save it for now
437 if (r.first.interleaved()) {
438 // if we already got interleaved ranges that are not
439 // part of the same range, then first do a merge
440 // before we add the new one
441 if (!intlv_ranges.empty() &&
442 !intlv_ranges.back().mergesWith(r.first)) {
443 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
444 intlv_ranges.size());
445 AddrRange merged_range(intlv_ranges);
446 // next decide if we keep the merged range or not
447 if (!(useDefaultRange &&
448 merged_range.isSubset(defaultRange))) {
449 xbarRanges.push_back(merged_range);
450 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
451 merged_range.to_string());
452 }
453 intlv_ranges.clear();
454 }
455 intlv_ranges.push_back(r.first);
456 } else {
457 // keep the current range if not a subset of the default
458 if (!(useDefaultRange &&
459 r.first.isSubset(defaultRange))) {
460 xbarRanges.push_back(r.first);
461 DPRINTF(AddrRanges, "-- Adding range %s\n",
462 r.first.to_string());
463 }
464 }
465 }
466
467 // if there is still interleaved ranges waiting to be merged,
468 // go ahead and do it
469 if (!intlv_ranges.empty()) {
470 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
471 intlv_ranges.size());
472 AddrRange merged_range(intlv_ranges);
473 if (!(useDefaultRange && merged_range.isSubset(defaultRange))) {
474 xbarRanges.push_back(merged_range);
475 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
476 merged_range.to_string());
477 }
478 }
479
480 // also check that no range partially intersects with the
481 // default range, this has to be done after all ranges are set
482 // as there are no guarantees for when the default range is
483 // update with respect to the other ones
484 if (useDefaultRange) {
485 for (const auto& r: xbarRanges) {
486 // see if the new range is partially
487 // overlapping the default range
488 if (r.intersects(defaultRange) &&
489 !r.isSubset(defaultRange))
490 fatal("Range %s intersects the " \
491 "default range of %s but is not a " \
492 "subset\n", r.to_string(), name());
493 }
494 }
495
496 // tell all our neighbouring master ports that our address
497 // ranges have changed
498 for (const auto& s: slavePorts)
499 s->sendRangeChange();
500 }
501}
502
503AddrRangeList
504BaseXBar::getAddrRanges() const
505{
506 // we should never be asked without first having sent a range
507 // change, and the latter is only done once we have all the ranges
508 // of the connected devices
509 assert(gotAllAddrRanges);
510
511 // at the moment, this never happens, as there are no cycles in
512 // the range queries and no devices on the master side of a crossbar
513 // (CPU, cache, bridge etc) actually care about the ranges of the
514 // ports they are connected to
515
516 DPRINTF(AddrRanges, "Received address range request\n");
517
518 return xbarRanges;
519}
520
521void
522BaseXBar::regStats()
523{
524 ClockedObject::regStats();
525
526 using namespace Stats;
527
528 transDist
529 .init(MemCmd::NUM_MEM_CMDS)
530 .name(name() + ".trans_dist")
531 .desc("Transaction distribution")
532 .flags(nozero);
533
534 // get the string representation of the commands
535 for (int i = 0; i < MemCmd::NUM_MEM_CMDS; i++) {
536 MemCmd cmd(i);
537 const std::string &cstr = cmd.toString();
538 transDist.subname(i, cstr);
539 }
540
541 pktCount
542 .init(slavePorts.size(), masterPorts.size())
543 .name(name() + ".pkt_count")
544 .desc("Packet count per connected master and slave (bytes)")
545 .flags(total | nozero | nonan);
546
547 pktSize
548 .init(slavePorts.size(), masterPorts.size())
549 .name(name() + ".pkt_size")
550 .desc("Cumulative packet size per connected master and slave (bytes)")
551 .flags(total | nozero | nonan);
552
553 // both the packet count and total size are two-dimensional
554 // vectors, indexed by slave port id and master port id, thus the
555 // neighbouring master and slave, they do not differentiate what
556 // came from the master and was forwarded to the slave (requests
557 // and snoop responses) and what came from the slave and was
558 // forwarded to the master (responses and snoop requests)
559 for (int i = 0; i < slavePorts.size(); i++) {
560 pktCount.subname(i, slavePorts[i]->getMasterPort().name());
561 pktSize.subname(i, slavePorts[i]->getMasterPort().name());
562 for (int j = 0; j < masterPorts.size(); j++) {
563 pktCount.ysubname(j, masterPorts[j]->getSlavePort().name());
564 pktSize.ysubname(j, masterPorts[j]->getSlavePort().name());
565 }
566 }
567}
568
569template <typename SrcType, typename DstType>
570DrainState
571BaseXBar::Layer<SrcType, DstType>::drain()
572{
573 //We should check that we're not "doing" anything, and that noone is
574 //waiting. We might be idle but have someone waiting if the device we
575 //contacted for a retry didn't actually retry.
576 if (state != IDLE) {
577 DPRINTF(Drain, "Crossbar not drained\n");
578 return DrainState::Draining;
579 } else {
580 return DrainState::Drained;
581 }
582}
583
584template <typename SrcType, typename DstType>
585void
586BaseXBar::Layer<SrcType, DstType>::regStats()
587{
588 using namespace Stats;
589
590 occupancy
591 .name(name() + ".occupancy")
592 .desc("Layer occupancy (ticks)")
593 .flags(nozero);
594
595 utilization
596 .name(name() + ".utilization")
597 .desc("Layer utilization (%)")
598 .precision(1)
599 .flags(nozero);
600
601 utilization = 100 * occupancy / simTicks;
602}
603
604/**
605 * Crossbar layer template instantiations. Could be removed with _impl.hh
606 * file, but since there are only two given options (MasterPort and
607 * SlavePort) it seems a bit excessive at this point.
608 */
609template class BaseXBar::Layer<SlavePort, MasterPort>;
610template class BaseXBar::Layer<MasterPort, SlavePort>;
| 92 }
93}
94
95void
96BaseXBar::calcPacketTiming(PacketPtr pkt, Tick header_delay)
97{
98 // the crossbar will be called at a time that is not necessarily
99 // coinciding with its own clock, so start by determining how long
100 // until the next clock edge (could be zero)
101 Tick offset = clockEdge() - curTick();
102
103 // the header delay depends on the path through the crossbar, and
104 // we therefore rely on the caller to provide the actual
105 // value
106 pkt->headerDelay += offset + header_delay;
107
108 // note that we add the header delay to the existing value, and
109 // align it to the crossbar clock
110
111 // do a quick sanity check to ensure the timings are not being
112 // ignored, note that this specific value may cause problems for
113 // slower interconnects
114 panic_if(pkt->headerDelay > SimClock::Int::us,
115 "Encountered header delay exceeding 1 us\n");
116
117 if (pkt->hasData()) {
118 // the payloadDelay takes into account the relative time to
119 // deliver the payload of the packet, after the header delay,
120 // we take the maximum since the payload delay could already
121 // be longer than what this parcitular crossbar enforces.
122 pkt->payloadDelay = std::max<Tick>(pkt->payloadDelay,
123 divCeil(pkt->getSize(), width) *
124 clockPeriod());
125 }
126
127 // the payload delay is not paying for the clock offset as that is
128 // already done using the header delay, and the payload delay is
129 // also used to determine how long the crossbar layer is busy and
130 // thus regulates throughput
131}
132
133template <typename SrcType, typename DstType>
134BaseXBar::Layer<SrcType, DstType>::Layer(DstType& _port, BaseXBar& _xbar,
135 const std::string& _name) :
136 port(_port), xbar(_xbar), _name(_name), state(IDLE),
137 waitingForPeer(NULL), releaseEvent([this]{ releaseLayer(); }, name())
138{
139}
140
141template <typename SrcType, typename DstType>
142void BaseXBar::Layer<SrcType, DstType>::occupyLayer(Tick until)
143{
144 // ensure the state is busy at this point, as the layer should
145 // transition from idle as soon as it has decided to forward the
146 // packet to prevent any follow-on calls to sendTiming seeing an
147 // unoccupied layer
148 assert(state == BUSY);
149
150 // until should never be 0 as express snoops never occupy the layer
151 assert(until != 0);
152 xbar.schedule(releaseEvent, until);
153
154 // account for the occupied ticks
155 occupancy += until - curTick();
156
157 DPRINTF(BaseXBar, "The crossbar layer is now busy from tick %d to %d\n",
158 curTick(), until);
159}
160
161template <typename SrcType, typename DstType>
162bool
163BaseXBar::Layer<SrcType, DstType>::tryTiming(SrcType* src_port)
164{
165 // if we are in the retry state, we will not see anything but the
166 // retrying port (or in the case of the snoop ports the snoop
167 // response port that mirrors the actual slave port) as we leave
168 // this state again in zero time if the peer does not immediately
169 // call the layer when receiving the retry
170
171 // first we see if the layer is busy, next we check if the
172 // destination port is already engaged in a transaction waiting
173 // for a retry from the peer
174 if (state == BUSY || waitingForPeer != NULL) {
175 // the port should not be waiting already
176 assert(std::find(waitingForLayer.begin(), waitingForLayer.end(),
177 src_port) == waitingForLayer.end());
178
179 // put the port at the end of the retry list waiting for the
180 // layer to be freed up (and in the case of a busy peer, for
181 // that transaction to go through, and then the layer to free
182 // up)
183 waitingForLayer.push_back(src_port);
184 return false;
185 }
186
187 state = BUSY;
188
189 return true;
190}
191
192template <typename SrcType, typename DstType>
193void
194BaseXBar::Layer<SrcType, DstType>::succeededTiming(Tick busy_time)
195{
196 // we should have gone from idle or retry to busy in the tryTiming
197 // test
198 assert(state == BUSY);
199
200 // occupy the layer accordingly
201 occupyLayer(busy_time);
202}
203
204template <typename SrcType, typename DstType>
205void
206BaseXBar::Layer<SrcType, DstType>::failedTiming(SrcType* src_port,
207 Tick busy_time)
208{
209 // ensure no one got in between and tried to send something to
210 // this port
211 assert(waitingForPeer == NULL);
212
213 // if the source port is the current retrying one or not, we have
214 // failed in forwarding and should track that we are now waiting
215 // for the peer to send a retry
216 waitingForPeer = src_port;
217
218 // we should have gone from idle or retry to busy in the tryTiming
219 // test
220 assert(state == BUSY);
221
222 // occupy the bus accordingly
223 occupyLayer(busy_time);
224}
225
226template <typename SrcType, typename DstType>
227void
228BaseXBar::Layer<SrcType, DstType>::releaseLayer()
229{
230 // releasing the bus means we should now be idle
231 assert(state == BUSY);
232 assert(!releaseEvent.scheduled());
233
234 // update the state
235 state = IDLE;
236
237 // bus layer is now idle, so if someone is waiting we can retry
238 if (!waitingForLayer.empty()) {
239 // there is no point in sending a retry if someone is still
240 // waiting for the peer
241 if (waitingForPeer == NULL)
242 retryWaiting();
243 } else if (waitingForPeer == NULL && drainState() == DrainState::Draining) {
244 DPRINTF(Drain, "Crossbar done draining, signaling drain manager\n");
245 //If we weren't able to drain before, do it now.
246 signalDrainDone();
247 }
248}
249
250template <typename SrcType, typename DstType>
251void
252BaseXBar::Layer<SrcType, DstType>::retryWaiting()
253{
254 // this should never be called with no one waiting
255 assert(!waitingForLayer.empty());
256
257 // we always go to retrying from idle
258 assert(state == IDLE);
259
260 // update the state
261 state = RETRY;
262
263 // set the retrying port to the front of the retry list and pop it
264 // off the list
265 SrcType* retryingPort = waitingForLayer.front();
266 waitingForLayer.pop_front();
267
268 // tell the port to retry, which in some cases ends up calling the
269 // layer again
270 sendRetry(retryingPort);
271
272 // If the layer is still in the retry state, sendTiming wasn't
273 // called in zero time (e.g. the cache does this when a writeback
274 // is squashed)
275 if (state == RETRY) {
276 // update the state to busy and reset the retrying port, we
277 // have done our bit and sent the retry
278 state = BUSY;
279
280 // occupy the crossbar layer until the next clock edge
281 occupyLayer(xbar.clockEdge());
282 }
283}
284
285template <typename SrcType, typename DstType>
286void
287BaseXBar::Layer<SrcType, DstType>::recvRetry()
288{
289 // we should never get a retry without having failed to forward
290 // something to this port
291 assert(waitingForPeer != NULL);
292
293 // add the port where the failed packet originated to the front of
294 // the waiting ports for the layer, this allows us to call retry
295 // on the port immediately if the crossbar layer is idle
296 waitingForLayer.push_front(waitingForPeer);
297
298 // we are no longer waiting for the peer
299 waitingForPeer = NULL;
300
301 // if the layer is idle, retry this port straight away, if we
302 // are busy, then simply let the port wait for its turn
303 if (state == IDLE) {
304 retryWaiting();
305 } else {
306 assert(state == BUSY);
307 }
308}
309
310PortID
311BaseXBar::findPort(AddrRange addr_range)
312{
313 // we should never see any address lookups before we've got the
314 // ranges of all connected slave modules
315 assert(gotAllAddrRanges);
316
317 // Check the address map interval tree
318 auto i = portMap.contains(addr_range);
319 if (i != portMap.end()) {
320 return i->second;
321 }
322
323 // Check if this matches the default range
324 if (useDefaultRange) {
325 if (addr_range.isSubset(defaultRange)) {
326 DPRINTF(AddrRanges, " found addr %s on default\n",
327 addr_range.to_string());
328 return defaultPortID;
329 }
330 } else if (defaultPortID != InvalidPortID) {
331 DPRINTF(AddrRanges, "Unable to find destination for %s, "
332 "will use default port\n", addr_range.to_string());
333 return defaultPortID;
334 }
335
336 // we should use the range for the default port and it did not
337 // match, or the default port is not set
338 fatal("Unable to find destination for %s on %s\n", addr_range.to_string(),
339 name());
340}
341
342/** Function called by the port when the crossbar is receiving a range change.*/
343void
344BaseXBar::recvRangeChange(PortID master_port_id)
345{
346 DPRINTF(AddrRanges, "Received range change from slave port %s\n",
347 masterPorts[master_port_id]->getSlavePort().name());
348
349 // remember that we got a range from this master port and thus the
350 // connected slave module
351 gotAddrRanges[master_port_id] = true;
352
353 // update the global flag
354 if (!gotAllAddrRanges) {
355 // take a logical AND of all the ports and see if we got
356 // ranges from everyone
357 gotAllAddrRanges = true;
358 std::vector<bool>::const_iterator r = gotAddrRanges.begin();
359 while (gotAllAddrRanges && r != gotAddrRanges.end()) {
360 gotAllAddrRanges &= *r++;
361 }
362 if (gotAllAddrRanges)
363 DPRINTF(AddrRanges, "Got address ranges from all slaves\n");
364 }
365
366 // note that we could get the range from the default port at any
367 // point in time, and we cannot assume that the default range is
368 // set before the other ones are, so we do additional checks once
369 // all ranges are provided
370 if (master_port_id == defaultPortID) {
371 // only update if we are indeed checking ranges for the
372 // default port since the port might not have a valid range
373 // otherwise
374 if (useDefaultRange) {
375 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
376
377 if (ranges.size() != 1)
378 fatal("Crossbar %s may only have a single default range",
379 name());
380
381 defaultRange = ranges.front();
382 }
383 } else {
384 // the ports are allowed to update their address ranges
385 // dynamically, so remove any existing entries
386 if (gotAddrRanges[master_port_id]) {
387 for (auto p = portMap.begin(); p != portMap.end(); ) {
388 if (p->second == master_port_id)
389 // erasing invalidates the iterator, so advance it
390 // before the deletion takes place
391 portMap.erase(p++);
392 else
393 p++;
394 }
395 }
396
397 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
398
399 for (const auto& r: ranges) {
400 DPRINTF(AddrRanges, "Adding range %s for id %d\n",
401 r.to_string(), master_port_id);
402 if (portMap.insert(r, master_port_id) == portMap.end()) {
403 PortID conflict_id = portMap.intersects(r)->second;
404 fatal("%s has two ports responding within range "
405 "%s:\n\t%s\n\t%s\n",
406 name(),
407 r.to_string(),
408 masterPorts[master_port_id]->getSlavePort().name(),
409 masterPorts[conflict_id]->getSlavePort().name());
410 }
411 }
412 }
413
414 // if we have received ranges from all our neighbouring slave
415 // modules, go ahead and tell our connected master modules in
416 // turn, this effectively assumes a tree structure of the system
417 if (gotAllAddrRanges) {
418 DPRINTF(AddrRanges, "Aggregating address ranges\n");
419 xbarRanges.clear();
420
421 // start out with the default range
422 if (useDefaultRange) {
423 if (!gotAddrRanges[defaultPortID])
424 fatal("Crossbar %s uses default range, but none provided",
425 name());
426
427 xbarRanges.push_back(defaultRange);
428 DPRINTF(AddrRanges, "-- Adding default %s\n",
429 defaultRange.to_string());
430 }
431
432 // merge all interleaved ranges and add any range that is not
433 // a subset of the default range
434 std::vector<AddrRange> intlv_ranges;
435 for (const auto& r: portMap) {
436 // if the range is interleaved then save it for now
437 if (r.first.interleaved()) {
438 // if we already got interleaved ranges that are not
439 // part of the same range, then first do a merge
440 // before we add the new one
441 if (!intlv_ranges.empty() &&
442 !intlv_ranges.back().mergesWith(r.first)) {
443 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
444 intlv_ranges.size());
445 AddrRange merged_range(intlv_ranges);
446 // next decide if we keep the merged range or not
447 if (!(useDefaultRange &&
448 merged_range.isSubset(defaultRange))) {
449 xbarRanges.push_back(merged_range);
450 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
451 merged_range.to_string());
452 }
453 intlv_ranges.clear();
454 }
455 intlv_ranges.push_back(r.first);
456 } else {
457 // keep the current range if not a subset of the default
458 if (!(useDefaultRange &&
459 r.first.isSubset(defaultRange))) {
460 xbarRanges.push_back(r.first);
461 DPRINTF(AddrRanges, "-- Adding range %s\n",
462 r.first.to_string());
463 }
464 }
465 }
466
467 // if there is still interleaved ranges waiting to be merged,
468 // go ahead and do it
469 if (!intlv_ranges.empty()) {
470 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
471 intlv_ranges.size());
472 AddrRange merged_range(intlv_ranges);
473 if (!(useDefaultRange && merged_range.isSubset(defaultRange))) {
474 xbarRanges.push_back(merged_range);
475 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
476 merged_range.to_string());
477 }
478 }
479
480 // also check that no range partially intersects with the
481 // default range, this has to be done after all ranges are set
482 // as there are no guarantees for when the default range is
483 // update with respect to the other ones
484 if (useDefaultRange) {
485 for (const auto& r: xbarRanges) {
486 // see if the new range is partially
487 // overlapping the default range
488 if (r.intersects(defaultRange) &&
489 !r.isSubset(defaultRange))
490 fatal("Range %s intersects the " \
491 "default range of %s but is not a " \
492 "subset\n", r.to_string(), name());
493 }
494 }
495
496 // tell all our neighbouring master ports that our address
497 // ranges have changed
498 for (const auto& s: slavePorts)
499 s->sendRangeChange();
500 }
501}
502
503AddrRangeList
504BaseXBar::getAddrRanges() const
505{
506 // we should never be asked without first having sent a range
507 // change, and the latter is only done once we have all the ranges
508 // of the connected devices
509 assert(gotAllAddrRanges);
510
511 // at the moment, this never happens, as there are no cycles in
512 // the range queries and no devices on the master side of a crossbar
513 // (CPU, cache, bridge etc) actually care about the ranges of the
514 // ports they are connected to
515
516 DPRINTF(AddrRanges, "Received address range request\n");
517
518 return xbarRanges;
519}
520
521void
522BaseXBar::regStats()
523{
524 ClockedObject::regStats();
525
526 using namespace Stats;
527
528 transDist
529 .init(MemCmd::NUM_MEM_CMDS)
530 .name(name() + ".trans_dist")
531 .desc("Transaction distribution")
532 .flags(nozero);
533
534 // get the string representation of the commands
535 for (int i = 0; i < MemCmd::NUM_MEM_CMDS; i++) {
536 MemCmd cmd(i);
537 const std::string &cstr = cmd.toString();
538 transDist.subname(i, cstr);
539 }
540
541 pktCount
542 .init(slavePorts.size(), masterPorts.size())
543 .name(name() + ".pkt_count")
544 .desc("Packet count per connected master and slave (bytes)")
545 .flags(total | nozero | nonan);
546
547 pktSize
548 .init(slavePorts.size(), masterPorts.size())
549 .name(name() + ".pkt_size")
550 .desc("Cumulative packet size per connected master and slave (bytes)")
551 .flags(total | nozero | nonan);
552
553 // both the packet count and total size are two-dimensional
554 // vectors, indexed by slave port id and master port id, thus the
555 // neighbouring master and slave, they do not differentiate what
556 // came from the master and was forwarded to the slave (requests
557 // and snoop responses) and what came from the slave and was
558 // forwarded to the master (responses and snoop requests)
559 for (int i = 0; i < slavePorts.size(); i++) {
560 pktCount.subname(i, slavePorts[i]->getMasterPort().name());
561 pktSize.subname(i, slavePorts[i]->getMasterPort().name());
562 for (int j = 0; j < masterPorts.size(); j++) {
563 pktCount.ysubname(j, masterPorts[j]->getSlavePort().name());
564 pktSize.ysubname(j, masterPorts[j]->getSlavePort().name());
565 }
566 }
567}
568
569template <typename SrcType, typename DstType>
570DrainState
571BaseXBar::Layer<SrcType, DstType>::drain()
572{
573 //We should check that we're not "doing" anything, and that noone is
574 //waiting. We might be idle but have someone waiting if the device we
575 //contacted for a retry didn't actually retry.
576 if (state != IDLE) {
577 DPRINTF(Drain, "Crossbar not drained\n");
578 return DrainState::Draining;
579 } else {
580 return DrainState::Drained;
581 }
582}
583
584template <typename SrcType, typename DstType>
585void
586BaseXBar::Layer<SrcType, DstType>::regStats()
587{
588 using namespace Stats;
589
590 occupancy
591 .name(name() + ".occupancy")
592 .desc("Layer occupancy (ticks)")
593 .flags(nozero);
594
595 utilization
596 .name(name() + ".utilization")
597 .desc("Layer utilization (%)")
598 .precision(1)
599 .flags(nozero);
600
601 utilization = 100 * occupancy / simTicks;
602}
603
604/**
605 * Crossbar layer template instantiations. Could be removed with _impl.hh
606 * file, but since there are only two given options (MasterPort and
607 * SlavePort) it seems a bit excessive at this point.
608 */
609template class BaseXBar::Layer<SlavePort, MasterPort>;
610template class BaseXBar::Layer<MasterPort, SlavePort>;
|