snoop_filter.cc revision 11603
1/*
2 * Copyright (c) 2013-2016 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 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions are
16 * met: redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer;
18 * redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution;
21 * neither the name of the copyright holders nor the names of its
22 * contributors may be used to endorse or promote products derived from
23 * this software without specific prior written permission.
24 *
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
29 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
30 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
31 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
32 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
33 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
34 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
35 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 *
37 * Authors: Stephan Diestelhorst
38 */
39
40/**
41 * @file
42 * Implementation of a snoop filter.
43 */
44
45#include "base/misc.hh"
46#include "base/trace.hh"
47#include "debug/SnoopFilter.hh"
48#include "mem/snoop_filter.hh"
49#include "sim/system.hh"
50
51void
52SnoopFilter::eraseIfNullEntry(SnoopFilterCache::iterator& sf_it)
53{
54    SnoopItem& sf_item = sf_it->second;
55    if (!(sf_item.requested | sf_item.holder)) {
56        cachedLocations.erase(sf_it);
57        DPRINTF(SnoopFilter, "%s:   Removed SF entry.\n",
58                __func__);
59    }
60}
61
62std::pair<SnoopFilter::SnoopList, Cycles>
63SnoopFilter::lookupRequest(const Packet* cpkt, const SlavePort& slave_port)
64{
65    DPRINTF(SnoopFilter, "%s: packet src %s addr 0x%x cmd %s\n",
66            __func__, slave_port.name(), cpkt->getAddr(), cpkt->cmdString());
67
68    // check if the packet came from a cache
69    bool allocate = !cpkt->req->isUncacheable() && slave_port.isSnooping() &&
70        cpkt->fromCache();
71    Addr line_addr = cpkt->getBlockAddr(linesize);
72    SnoopMask req_port = portToMask(slave_port);
73    reqLookupResult = cachedLocations.find(line_addr);
74    bool is_hit = (reqLookupResult != cachedLocations.end());
75
76    // If the snoop filter has no entry, and we should not allocate,
77    // do not create a new snoop filter entry, simply return a NULL
78    // portlist.
79    if (!is_hit && !allocate)
80        return snoopDown(lookupLatency);
81
82    // If no hit in snoop filter create a new element and update iterator
83    if (!is_hit)
84        reqLookupResult = cachedLocations.emplace(line_addr, SnoopItem()).first;
85    SnoopItem& sf_item = reqLookupResult->second;
86    SnoopMask interested = sf_item.holder | sf_item.requested;
87
88    // Store unmodified value of snoop filter item in temp storage in
89    // case we need to revert because of a send retry in
90    // updateRequest.
91    retryItem = sf_item;
92
93    totRequests++;
94    if (is_hit) {
95        // Single bit set -> value is a power of two
96        if (isPow2(interested))
97            hitSingleRequests++;
98        else
99            hitMultiRequests++;
100    }
101
102    DPRINTF(SnoopFilter, "%s:   SF value %x.%x\n",
103            __func__, sf_item.requested, sf_item.holder);
104
105    // If we are not allocating, we are done
106    if (!allocate)
107        return snoopSelected(maskToPortList(interested & ~req_port),
108                             lookupLatency);
109
110    if (cpkt->needsResponse()) {
111        if (!cpkt->cacheResponding()) {
112            // Max one request per address per port
113            panic_if(sf_item.requested & req_port, "double request :( " \
114                     "SF value %x.%x\n", sf_item.requested, sf_item.holder);
115
116            // Mark in-flight requests to distinguish later on
117            sf_item.requested |= req_port;
118            DPRINTF(SnoopFilter, "%s:   new SF value %x.%x\n",
119                    __func__,  sf_item.requested, sf_item.holder);
120        } else {
121            // NOTE: The memInhibit might have been asserted by a cache closer
122            // to the CPU, already -> the response will not be seen by this
123            // filter -> we do not need to keep the in-flight request, but make
124            // sure that we know that that cluster has a copy
125            panic_if(!(sf_item.holder & req_port), "Need to hold the value!");
126            DPRINTF(SnoopFilter,
127                    "%s: not marking request. SF value %x.%x\n",
128                    __func__,  sf_item.requested, sf_item.holder);
129        }
130    } else { // if (!cpkt->needsResponse())
131        assert(cpkt->isEviction());
132        // make sure that the sender actually had the line
133        panic_if(!(sf_item.holder & req_port), "requester %x is not a " \
134                 "holder :( SF value %x.%x\n", req_port,
135                 sf_item.requested, sf_item.holder);
136        // CleanEvicts and Writebacks -> the sender and all caches above
137        // it may not have the line anymore.
138        if (!cpkt->isBlockCached()) {
139            sf_item.holder &= ~req_port;
140            DPRINTF(SnoopFilter, "%s:   new SF value %x.%x\n",
141                    __func__,  sf_item.requested, sf_item.holder);
142        }
143    }
144
145    return snoopSelected(maskToPortList(interested & ~req_port), lookupLatency);
146}
147
148void
149SnoopFilter::finishRequest(bool will_retry, const Addr addr)
150{
151    if (reqLookupResult != cachedLocations.end()) {
152        // since we rely on the caller, do a basic check to ensure
153        // that finishRequest is being called following lookupRequest
154        assert(reqLookupResult->first == (addr & ~(Addr(linesize - 1))));
155        if (will_retry) {
156            // Undo any changes made in lookupRequest to the snoop filter
157            // entry if the request will come again. retryItem holds
158            // the previous value of the snoopfilter entry.
159            reqLookupResult->second = retryItem;
160
161            DPRINTF(SnoopFilter, "%s:   restored SF value %x.%x\n",
162                    __func__,  retryItem.requested, retryItem.holder);
163        }
164
165        eraseIfNullEntry(reqLookupResult);
166    }
167}
168
169std::pair<SnoopFilter::SnoopList, Cycles>
170SnoopFilter::lookupSnoop(const Packet* cpkt)
171{
172    DPRINTF(SnoopFilter, "%s: packet addr 0x%x cmd %s\n",
173            __func__, cpkt->getAddr(), cpkt->cmdString());
174
175    assert(cpkt->isRequest());
176
177    Addr line_addr = cpkt->getBlockAddr(linesize);
178    auto sf_it = cachedLocations.find(line_addr);
179    bool is_hit = (sf_it != cachedLocations.end());
180
181    panic_if(!is_hit && (cachedLocations.size() >= maxEntryCount),
182             "snoop filter exceeded capacity of %d cache blocks\n",
183             maxEntryCount);
184
185    // If the snoop filter has no entry, simply return a NULL
186    // portlist, there is no point creating an entry only to remove it
187    // later
188    if (!is_hit)
189        return snoopDown(lookupLatency);
190
191    SnoopItem& sf_item = sf_it->second;
192
193    DPRINTF(SnoopFilter, "%s:   old SF value %x.%x\n",
194            __func__, sf_item.requested, sf_item.holder);
195
196    SnoopMask interested = (sf_item.holder | sf_item.requested);
197
198    totSnoops++;
199    // Single bit set -> value is a power of two
200    if (isPow2(interested))
201        hitSingleSnoops++;
202    else
203        hitMultiSnoops++;
204
205    // ReadEx and Writes require both invalidation and exlusivity, while reads
206    // require neither. Writebacks on the other hand require exclusivity but
207    // not the invalidation. Previously Writebacks did not generate upward
208    // snoops so this was never an aissue. Now that Writebacks generate snoops
209    // we need to special case for Writebacks.
210    assert(cpkt->isWriteback() || cpkt->req->isUncacheable() ||
211           (cpkt->isInvalidate() == cpkt->needsWritable()));
212    if (cpkt->isInvalidate() && !sf_item.requested) {
213        // Early clear of the holder, if no other request is currently going on
214        // @todo: This should possibly be updated even though we do not filter
215        // upward snoops
216        sf_item.holder = 0;
217    }
218
219    eraseIfNullEntry(sf_it);
220    DPRINTF(SnoopFilter, "%s:   new SF value %x.%x interest: %x \n",
221            __func__, sf_item.requested, sf_item.holder, interested);
222
223    return snoopSelected(maskToPortList(interested), lookupLatency);
224}
225
226void
227SnoopFilter::updateSnoopResponse(const Packet* cpkt,
228                                 const SlavePort& rsp_port,
229                                 const SlavePort& req_port)
230{
231    DPRINTF(SnoopFilter, "%s: packet rsp %s req %s addr 0x%x cmd %s\n",
232            __func__, rsp_port.name(), req_port.name(), cpkt->getAddr(),
233            cpkt->cmdString());
234
235    assert(cpkt->isResponse());
236    assert(cpkt->cacheResponding());
237
238    // if this snoop response is due to an uncacheable request, or is
239    // being turned into a normal response, there is nothing more to
240    // do
241    if (cpkt->req->isUncacheable() || !req_port.isSnooping()) {
242        return;
243    }
244
245    Addr line_addr = cpkt->getBlockAddr(linesize);
246    SnoopMask rsp_mask = portToMask(rsp_port);
247    SnoopMask req_mask = portToMask(req_port);
248    SnoopItem& sf_item = cachedLocations[line_addr];
249
250    DPRINTF(SnoopFilter, "%s:   old SF value %x.%x\n",
251            __func__,  sf_item.requested, sf_item.holder);
252
253    // The source should have the line
254    panic_if(!(sf_item.holder & rsp_mask), "SF value %x.%x does not have "\
255             "the line\n", sf_item.requested, sf_item.holder);
256
257    // The destination should have had a request in
258    panic_if(!(sf_item.requested & req_mask), "SF value %x.%x missing "\
259             "the original request\n",  sf_item.requested, sf_item.holder);
260
261    // If the snoop response has no sharers the line is passed in
262    // Modified state, and we know that there are no other copies, or
263    // they will all be invalidated imminently
264    if (!cpkt->hasSharers()) {
265        DPRINTF(SnoopFilter,
266                "%s: dropping %x because non-shared snoop "
267                "response SF val: %x.%x\n", __func__,  rsp_mask,
268                sf_item.requested, sf_item.holder);
269        sf_item.holder = 0;
270    }
271    assert(!cpkt->isWriteback());
272    // @todo Deal with invalidating responses
273    sf_item.holder |=  req_mask;
274    sf_item.requested &= ~req_mask;
275    assert(sf_item.requested | sf_item.holder);
276    DPRINTF(SnoopFilter, "%s:   new SF value %x.%x\n",
277            __func__, sf_item.requested, sf_item.holder);
278}
279
280void
281SnoopFilter::updateSnoopForward(const Packet* cpkt,
282        const SlavePort& rsp_port, const MasterPort& req_port)
283{
284    DPRINTF(SnoopFilter, "%s: packet rsp %s req %s addr 0x%x cmd %s\n",
285            __func__, rsp_port.name(), req_port.name(), cpkt->getAddr(),
286            cpkt->cmdString());
287
288    assert(cpkt->isResponse());
289    assert(cpkt->cacheResponding());
290
291    Addr line_addr = cpkt->getBlockAddr(linesize);
292    auto sf_it = cachedLocations.find(line_addr);
293    bool is_hit = sf_it != cachedLocations.end();
294
295    // Nothing to do if it is not a hit
296    if (!is_hit)
297        return;
298
299    SnoopItem& sf_item = sf_it->second;
300
301    DPRINTF(SnoopFilter, "%s:   old SF value %x.%x\n",
302            __func__,  sf_item.requested, sf_item.holder);
303
304    // If the snoop response has no sharers the line is passed in
305    // Modified state, and we know that there are no other copies, or
306    // they will all be invalidated imminently
307    if (!cpkt->hasSharers()) {
308        sf_item.holder = 0;
309    }
310    DPRINTF(SnoopFilter, "%s:   new SF value %x.%x\n",
311            __func__, sf_item.requested, sf_item.holder);
312    eraseIfNullEntry(sf_it);
313
314}
315
316void
317SnoopFilter::updateResponse(const Packet* cpkt, const SlavePort& slave_port)
318{
319    DPRINTF(SnoopFilter, "%s: packet src %s addr 0x%x cmd %s\n",
320            __func__, slave_port.name(), cpkt->getAddr(), cpkt->cmdString());
321
322    assert(cpkt->isResponse());
323
324    // we only allocate if the packet actually came from a cache, but
325    // start by checking if the port is snooping
326    if (cpkt->req->isUncacheable() || !slave_port.isSnooping())
327        return;
328
329    // next check if we actually allocated an entry
330    Addr line_addr = cpkt->getBlockAddr(linesize);
331    auto sf_it = cachedLocations.find(line_addr);
332    if (sf_it == cachedLocations.end())
333        return;
334
335    SnoopMask slave_mask = portToMask(slave_port);
336    SnoopItem& sf_item = sf_it->second;
337
338    DPRINTF(SnoopFilter, "%s:   old SF value %x.%x\n",
339            __func__,  sf_item.requested, sf_item.holder);
340
341    // Make sure we have seen the actual request, too
342    panic_if(!(sf_item.requested & slave_mask), "SF value %x.%x missing "\
343             "request bit\n", sf_item.requested, sf_item.holder);
344
345    // Update the residency of the cache line. If the response has no
346    // sharers we know that the line has been invalidated in all
347    // branches that are not where we are responding to.
348     if (!cpkt->hasSharers())
349        sf_item.holder = 0;
350    sf_item.holder |=  slave_mask;
351    sf_item.requested &= ~slave_mask;
352    assert(sf_item.holder | sf_item.requested);
353    DPRINTF(SnoopFilter, "%s:   new SF value %x.%x\n",
354            __func__, sf_item.requested, sf_item.holder);
355}
356
357void
358SnoopFilter::regStats()
359{
360    SimObject::regStats();
361
362    totRequests
363        .name(name() + ".tot_requests")
364        .desc("Total number of requests made to the snoop filter.");
365
366    hitSingleRequests
367        .name(name() + ".hit_single_requests")
368        .desc("Number of requests hitting in the snoop filter with a single "\
369              "holder of the requested data.");
370
371    hitMultiRequests
372        .name(name() + ".hit_multi_requests")
373        .desc("Number of requests hitting in the snoop filter with multiple "\
374              "(>1) holders of the requested data.");
375
376    totSnoops
377        .name(name() + ".tot_snoops")
378        .desc("Total number of snoops made to the snoop filter.");
379
380    hitSingleSnoops
381        .name(name() + ".hit_single_snoops")
382        .desc("Number of snoops hitting in the snoop filter with a single "\
383              "holder of the requested data.");
384
385    hitMultiSnoops
386        .name(name() + ".hit_multi_snoops")
387        .desc("Number of snoops hitting in the snoop filter with multiple "\
388              "(>1) holders of the requested data.");
389}
390
391SnoopFilter *
392SnoopFilterParams::create()
393{
394    return new SnoopFilter(this);
395}
396