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)
59 : MemObject(p),
60 frontendLatency(p->frontend_latency),
61 forwardLatency(p->forward_latency),
62 responseLatency(p->response_latency),
63 width(p->width),
64 gotAddrRanges(p->port_default_connection_count +
65 p->port_master_connection_count, false),
66 gotAllAddrRanges(false), defaultPortID(InvalidPortID),
67 useDefaultRange(p->use_default_range)
68{}
69
70BaseXBar::~BaseXBar()
71{
72 for (auto m: masterPorts)
73 delete m;
74
75 for (auto s: slavePorts)
76 delete s;
77}
78
79void
80BaseXBar::init()
81{
82}
83
84Port &
85BaseXBar::getPort(const std::string &if_name, PortID idx)
86{
87 if (if_name == "master" && idx < masterPorts.size()) {
88 // the master port index translates directly to the vector position
89 return *masterPorts[idx];
90 } else if (if_name == "default") {
91 return *masterPorts[defaultPortID];
92 } else if (if_name == "slave" && idx < slavePorts.size()) {
93 // the slave port index translates directly to the vector position
94 return *slavePorts[idx];
95 } else {
96 return MemObject::getPort(if_name, idx);
97 }
98}
99
100void
101BaseXBar::calcPacketTiming(PacketPtr pkt, Tick header_delay)
102{
103 // the crossbar will be called at a time that is not necessarily
104 // coinciding with its own clock, so start by determining how long
105 // until the next clock edge (could be zero)
106 Tick offset = clockEdge() - curTick();
107
108 // the header delay depends on the path through the crossbar, and
109 // we therefore rely on the caller to provide the actual
110 // value
111 pkt->headerDelay += offset + header_delay;
112
113 // note that we add the header delay to the existing value, and
114 // align it to the crossbar clock
115
116 // do a quick sanity check to ensure the timings are not being
117 // ignored, note that this specific value may cause problems for
118 // slower interconnects
119 panic_if(pkt->headerDelay > SimClock::Int::us,
120 "Encountered header delay exceeding 1 us\n");
121
122 if (pkt->hasData()) {
123 // the payloadDelay takes into account the relative time to
124 // deliver the payload of the packet, after the header delay,
125 // we take the maximum since the payload delay could already
126 // be longer than what this parcitular crossbar enforces.
127 pkt->payloadDelay = std::max<Tick>(pkt->payloadDelay,
128 divCeil(pkt->getSize(), width) *
129 clockPeriod());
130 }
131
132 // the payload delay is not paying for the clock offset as that is
133 // already done using the header delay, and the payload delay is
134 // also used to determine how long the crossbar layer is busy and
135 // thus regulates throughput
136}
137
138template <typename SrcType, typename DstType>
139BaseXBar::Layer::Layer(DstType& _port, BaseXBar& _xbar,
140 const std::string& _name) :
141 port(_port), xbar(_xbar), _name(_name), state(IDLE),
142 waitingForPeer(NULL), releaseEvent([this]{ releaseLayer(); }, name())
143{
144}
145
146template <typename SrcType, typename DstType>
147void BaseXBar::Layer::occupyLayer(Tick until)
148{
149 // ensure the state is busy at this point, as the layer should
150 // transition from idle as soon as it has decided to forward the
151 // packet to prevent any follow-on calls to sendTiming seeing an
152 // unoccupied layer
153 assert(state == BUSY);
154
155 // until should never be 0 as express snoops never occupy the layer
156 assert(until != 0);
157 xbar.schedule(releaseEvent, until);
158
159 // account for the occupied ticks
160 occupancy += until - curTick();
161
162 DPRINTF(BaseXBar, "The crossbar layer is now busy from tick %d to %d\n",
163 curTick(), until);
164}
165
166template <typename SrcType, typename DstType>
167bool
168BaseXBar::Layer::tryTiming(SrcType* src_port)
169{
170 // if we are in the retry state, we will not see anything but the
171 // retrying port (or in the case of the snoop ports the snoop
172 // response port that mirrors the actual slave port) as we leave
173 // this state again in zero time if the peer does not immediately
174 // call the layer when receiving the retry
175
176 // first we see if the layer is busy, next we check if the
177 // destination port is already engaged in a transaction waiting
178 // for a retry from the peer
179 if (state == BUSY || waitingForPeer != NULL) {
180 // the port should not be waiting already
181 assert(std::find(waitingForLayer.begin(), waitingForLayer.end(),
182 src_port) == waitingForLayer.end());
183
184 // put the port at the end of the retry list waiting for the
185 // layer to be freed up (and in the case of a busy peer, for
186 // that transaction to go through, and then the layer to free
187 // up)
188 waitingForLayer.push_back(src_port);
189 return false;
190 }
191
192 state = BUSY;
193
194 return true;
195}
196
197template <typename SrcType, typename DstType>
198void
199BaseXBar::Layer::succeededTiming(Tick busy_time)
200{
201 // we should have gone from idle or retry to busy in the tryTiming
202 // test
203 assert(state == BUSY);
204
205 // occupy the layer accordingly
206 occupyLayer(busy_time);
207}
208
209template <typename SrcType, typename DstType>
210void
211BaseXBar::Layer::failedTiming(SrcType* src_port,
212 Tick busy_time)
213{
214 // ensure no one got in between and tried to send something to
215 // this port
216 assert(waitingForPeer == NULL);
217
218 // if the source port is the current retrying one or not, we have
219 // failed in forwarding and should track that we are now waiting
220 // for the peer to send a retry
221 waitingForPeer = src_port;
222
223 // we should have gone from idle or retry to busy in the tryTiming
224 // test
225 assert(state == BUSY);
226
227 // occupy the bus accordingly
228 occupyLayer(busy_time);
229}
230
231template <typename SrcType, typename DstType>
232void
233BaseXBar::Layer::releaseLayer()
234{
235 // releasing the bus means we should now be idle
236 assert(state == BUSY);
237 assert(!releaseEvent.scheduled());
238
239 // update the state
240 state = IDLE;
241
242 // bus layer is now idle, so if someone is waiting we can retry
243 if (!waitingForLayer.empty()) {
244 // there is no point in sending a retry if someone is still
245 // waiting for the peer
246 if (waitingForPeer == NULL)
247 retryWaiting();
248 } else if (waitingForPeer == NULL && drainState() == DrainState::Draining) {
249 DPRINTF(Drain, "Crossbar done draining, signaling drain manager\n");
250 //If we weren't able to drain before, do it now.
251 signalDrainDone();
252 }
253}
254
255template <typename SrcType, typename DstType>
256void
257BaseXBar::Layer::retryWaiting()
258{
259 // this should never be called with no one waiting
260 assert(!waitingForLayer.empty());
261
262 // we always go to retrying from idle
263 assert(state == IDLE);
264
265 // update the state
266 state = RETRY;
267
268 // set the retrying port to the front of the retry list and pop it
269 // off the list
270 SrcType* retryingPort = waitingForLayer.front();
271 waitingForLayer.pop_front();
272
273 // tell the port to retry, which in some cases ends up calling the
274 // layer again
275 sendRetry(retryingPort);
276
277 // If the layer is still in the retry state, sendTiming wasn't
278 // called in zero time (e.g. the cache does this when a writeback
279 // is squashed)
280 if (state == RETRY) {
281 // update the state to busy and reset the retrying port, we
282 // have done our bit and sent the retry
283 state = BUSY;
284
285 // occupy the crossbar layer until the next clock edge
286 occupyLayer(xbar.clockEdge());
287 }
288}
289
290template <typename SrcType, typename DstType>
291void
292BaseXBar::Layer::recvRetry()
293{
294 // we should never get a retry without having failed to forward
295 // something to this port
296 assert(waitingForPeer != NULL);
297
298 // add the port where the failed packet originated to the front of
299 // the waiting ports for the layer, this allows us to call retry
300 // on the port immediately if the crossbar layer is idle
301 waitingForLayer.push_front(waitingForPeer);
302
303 // we are no longer waiting for the peer
304 waitingForPeer = NULL;
305
306 // if the layer is idle, retry this port straight away, if we
307 // are busy, then simply let the port wait for its turn
308 if (state == IDLE) {
309 retryWaiting();
310 } else {
311 assert(state == BUSY);
312 }
313}
314
315PortID
316BaseXBar::findPort(AddrRange addr_range)
317{
318 // we should never see any address lookups before we've got the
319 // ranges of all connected slave modules
320 assert(gotAllAddrRanges);
321
322 // Check the address map interval tree
323 auto i = portMap.contains(addr_range);
324 if (i != portMap.end()) {
325 return i->second;
326 }
327
328 // Check if this matches the default range
329 if (useDefaultRange) {
330 if (addr_range.isSubset(defaultRange)) {
331 DPRINTF(AddrRanges, " found addr %s on default\n",
332 addr_range.to_string());
333 return defaultPortID;
334 }
335 } else if (defaultPortID != InvalidPortID) {
336 DPRINTF(AddrRanges, "Unable to find destination for %s, "
337 "will use default port\n", addr_range.to_string());
338 return defaultPortID;
339 }
340
341 // we should use the range for the default port and it did not
342 // match, or the default port is not set
343 fatal("Unable to find destination for %s on %s\n", addr_range.to_string(),
344 name());
345}
346
347/** Function called by the port when the crossbar is receiving a range change.*/
348void
349BaseXBar::recvRangeChange(PortID master_port_id)
350{
351 DPRINTF(AddrRanges, "Received range change from slave port %s\n",
352 masterPorts[master_port_id]->getSlavePort().name());
353
354 // remember that we got a range from this master port and thus the
355 // connected slave module
356 gotAddrRanges[master_port_id] = true;
357
358 // update the global flag
359 if (!gotAllAddrRanges) {
360 // take a logical AND of all the ports and see if we got
361 // ranges from everyone
362 gotAllAddrRanges = true;
363 std::vector<bool>::const_iterator r = gotAddrRanges.begin();
364 while (gotAllAddrRanges && r != gotAddrRanges.end()) {
365 gotAllAddrRanges &= *r++;
366 }
367 if (gotAllAddrRanges)
368 DPRINTF(AddrRanges, "Got address ranges from all slaves\n");
369 }
370
371 // note that we could get the range from the default port at any
372 // point in time, and we cannot assume that the default range is
373 // set before the other ones are, so we do additional checks once
374 // all ranges are provided
375 if (master_port_id == defaultPortID) {
376 // only update if we are indeed checking ranges for the
377 // default port since the port might not have a valid range
378 // otherwise
379 if (useDefaultRange) {
380 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
381
382 if (ranges.size() != 1)
383 fatal("Crossbar %s may only have a single default range",
384 name());
385
386 defaultRange = ranges.front();
387 }
388 } else {
389 // the ports are allowed to update their address ranges
390 // dynamically, so remove any existing entries
391 if (gotAddrRanges[master_port_id]) {
392 for (auto p = portMap.begin(); p != portMap.end(); ) {
393 if (p->second == master_port_id)
394 // erasing invalidates the iterator, so advance it
395 // before the deletion takes place
396 portMap.erase(p++);
397 else
398 p++;
399 }
400 }
401
402 AddrRangeList ranges = masterPorts[master_port_id]->getAddrRanges();
403
404 for (const auto& r: ranges) {
405 DPRINTF(AddrRanges, "Adding range %s for id %d\n",
406 r.to_string(), master_port_id);
407 if (portMap.insert(r, master_port_id) == portMap.end()) {
408 PortID conflict_id = portMap.intersects(r)->second;
409 fatal("%s has two ports responding within range "
410 "%s:\n\t%s\n\t%s\n",
411 name(),
412 r.to_string(),
413 masterPorts[master_port_id]->getSlavePort().name(),
414 masterPorts[conflict_id]->getSlavePort().name());
415 }
416 }
417 }
418
419 // if we have received ranges from all our neighbouring slave
420 // modules, go ahead and tell our connected master modules in
421 // turn, this effectively assumes a tree structure of the system
422 if (gotAllAddrRanges) {
423 DPRINTF(AddrRanges, "Aggregating address ranges\n");
424 xbarRanges.clear();
425
426 // start out with the default range
427 if (useDefaultRange) {
428 if (!gotAddrRanges[defaultPortID])
429 fatal("Crossbar %s uses default range, but none provided",
430 name());
431
432 xbarRanges.push_back(defaultRange);
433 DPRINTF(AddrRanges, "-- Adding default %s\n",
434 defaultRange.to_string());
435 }
436
437 // merge all interleaved ranges and add any range that is not
438 // a subset of the default range
439 std::vector<AddrRange> intlv_ranges;
440 for (const auto& r: portMap) {
441 // if the range is interleaved then save it for now
442 if (r.first.interleaved()) {
443 // if we already got interleaved ranges that are not
444 // part of the same range, then first do a merge
445 // before we add the new one
446 if (!intlv_ranges.empty() &&
447 !intlv_ranges.back().mergesWith(r.first)) {
448 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
449 intlv_ranges.size());
450 AddrRange merged_range(intlv_ranges);
451 // next decide if we keep the merged range or not
452 if (!(useDefaultRange &&
453 merged_range.isSubset(defaultRange))) {
454 xbarRanges.push_back(merged_range);
455 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
456 merged_range.to_string());
457 }
458 intlv_ranges.clear();
459 }
460 intlv_ranges.push_back(r.first);
461 } else {
462 // keep the current range if not a subset of the default
463 if (!(useDefaultRange &&
464 r.first.isSubset(defaultRange))) {
465 xbarRanges.push_back(r.first);
466 DPRINTF(AddrRanges, "-- Adding range %s\n",
467 r.first.to_string());
468 }
469 }
470 }
471
472 // if there is still interleaved ranges waiting to be merged,
473 // go ahead and do it
474 if (!intlv_ranges.empty()) {
475 DPRINTF(AddrRanges, "-- Merging range from %d ranges\n",
476 intlv_ranges.size());
477 AddrRange merged_range(intlv_ranges);
478 if (!(useDefaultRange && merged_range.isSubset(defaultRange))) {
479 xbarRanges.push_back(merged_range);
480 DPRINTF(AddrRanges, "-- Adding merged range %s\n",
481 merged_range.to_string());
482 }
483 }
484
485 // also check that no range partially intersects with the
486 // default range, this has to be done after all ranges are set
487 // as there are no guarantees for when the default range is
488 // update with respect to the other ones
489 if (useDefaultRange) {
490 for (const auto& r: xbarRanges) {
491 // see if the new range is partially
492 // overlapping the default range
493 if (r.intersects(defaultRange) &&
494 !r.isSubset(defaultRange))
495 fatal("Range %s intersects the " \
496 "default range of %s but is not a " \
497 "subset\n", r.to_string(), name());
498 }
499 }
500
501 // tell all our neighbouring master ports that our address
502 // ranges have changed
503 for (const auto& s: slavePorts)
504 s->sendRangeChange();
505 }
506}
507
508AddrRangeList
509BaseXBar::getAddrRanges() const
510{
511 // we should never be asked without first having sent a range
512 // change, and the latter is only done once we have all the ranges
513 // of the connected devices
514 assert(gotAllAddrRanges);
515
516 // at the moment, this never happens, as there are no cycles in
517 // the range queries and no devices on the master side of a crossbar
518 // (CPU, cache, bridge etc) actually care about the ranges of the
519 // ports they are connected to
520
521 DPRINTF(AddrRanges, "Received address range request\n");
522
523 return xbarRanges;
524}
525
526void
527BaseXBar::regStats()
528{
529 ClockedObject::regStats();
530
531 using namespace Stats;
532
533 transDist
534 .init(MemCmd::NUM_MEM_CMDS)
535 .name(name() + ".trans_dist")
536 .desc("Transaction distribution")
537 .flags(nozero);
538
539 // get the string representation of the commands
540 for (int i = 0; i < MemCmd::NUM_MEM_CMDS; i++) {
541 MemCmd cmd(i);
542 const std::string &cstr = cmd.toString();
543 transDist.subname(i, cstr);
544 }
545
546 pktCount
547 .init(slavePorts.size(), masterPorts.size())
548 .name(name() + ".pkt_count")
549 .desc("Packet count per connected master and slave (bytes)")
550 .flags(total | nozero | nonan);
551
552 pktSize
553 .init(slavePorts.size(), masterPorts.size())
554 .name(name() + ".pkt_size")
555 .desc("Cumulative packet size per connected master and slave (bytes)")
556 .flags(total | nozero | nonan);
557
558 // both the packet count and total size are two-dimensional
559 // vectors, indexed by slave port id and master port id, thus the
560 // neighbouring master and slave, they do not differentiate what
561 // came from the master and was forwarded to the slave (requests
562 // and snoop responses) and what came from the slave and was
563 // forwarded to the master (responses and snoop requests)
564 for (int i = 0; i < slavePorts.size(); i++) {
565 pktCount.subname(i, slavePorts[i]->getMasterPort().name());
566 pktSize.subname(i, slavePorts[i]->getMasterPort().name());
567 for (int j = 0; j < masterPorts.size(); j++) {
568 pktCount.ysubname(j, masterPorts[j]->getSlavePort().name());
569 pktSize.ysubname(j, masterPorts[j]->getSlavePort().name());
570 }
571 }
572}
573
574template <typename SrcType, typename DstType>
575DrainState
576BaseXBar::Layer::drain()
577{
578 //We should check that we're not "doing" anything, and that noone is
579 //waiting. We might be idle but have someone waiting if the device we
580 //contacted for a retry didn't actually retry.
581 if (state != IDLE) {
582 DPRINTF(Drain, "Crossbar not drained\n");
583 return DrainState::Draining;
584 } else {
585 return DrainState::Drained;
586 }
587}
588
589template <typename SrcType, typename DstType>
590void
591BaseXBar::Layer::regStats()
592{
593 using namespace Stats;
594
595 occupancy
596 .name(name() + ".occupancy")
597 .desc("Layer occupancy (ticks)")
598 .flags(nozero);
599
600 utilization
601 .name(name() + ".utilization")
602 .desc("Layer utilization (%)")
603 .precision(1)
604 .flags(nozero);
605
606 utilization = 100 * occupancy / simTicks;
607}
608
609/**
610 * Crossbar layer template instantiations. Could be removed with _impl.hh
611 * file, but since there are only two given options (MasterPort and
612 * SlavePort) it seems a bit excessive at this point.
613 */
614template class BaseXBar::Layer;
615template class BaseXBar::Layer;