Commit | Line | Data |
---|---|---|
920dae64 AT |
1 | // ========== Copyright Header Begin ========================================== |
2 | // | |
3 | // OpenSPARC T2 Processor File: MemorySync.cc | |
4 | // Copyright (c) 2006 Sun Microsystems, Inc. All Rights Reserved. | |
5 | // DO NOT ALTER OR REMOVE COPYRIGHT NOTICES. | |
6 | // | |
7 | // The above named program is free software; you can redistribute it and/or | |
8 | // modify it under the terms of the GNU General Public | |
9 | // License version 2 as published by the Free Software Foundation. | |
10 | // | |
11 | // The above named program is distributed in the hope that it will be | |
12 | // useful, but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
14 | // General Public License for more details. | |
15 | // | |
16 | // You should have received a copy of the GNU General Public | |
17 | // License along with this work; if not, write to the Free Software | |
18 | // Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA. | |
19 | // | |
20 | // ========== Copyright Header End ============================================ | |
21 | /************************************************************************ | |
22 | ** | |
23 | ** Copyright (C) 2002, Sun Microsystems, Inc. | |
24 | ** | |
25 | ** Sun considers its source code as an unpublished, proprietary | |
26 | ** trade secret and it is available only under strict license provisions. | |
27 | ** This copyright notice is placed here only to protect Sun in the event | |
28 | ** the source is deemed a published work. Disassembly, decompilation, | |
29 | ** or other means of reducing the object code to human readable form | |
30 | ** is prohibited by the license agreement under which this code is | |
31 | ** provided to the user or company in possession of this copy." | |
32 | ** | |
33 | *************************************************************************/ | |
34 | #include "MemorySync.h" | |
35 | #include <sstream> | |
36 | #include <cstdlib> | |
37 | #include <time.h> | |
38 | #include "MemoryTransaction.h" | |
39 | #include "SS_Strand.h" | |
40 | ||
41 | ||
42 | using namespace std; | |
43 | //////////////////////////////////////////////// | |
44 | typedef void (MemorySync::* mem_type)(MemoryTransaction&); | |
45 | ||
46 | // static variable | |
47 | MemorySync* MemorySync::msyncObj = NULL; | |
48 | // static function, dump buffer content | |
49 | string MemorySync::dumpBuffers() | |
50 | { | |
51 | if (MemorySync::msyncObj) { | |
52 | return MemorySync::msyncObj->toString(); | |
53 | } | |
54 | else { | |
55 | return string(""); | |
56 | } | |
57 | } | |
58 | ||
59 | MemorySync::MemorySync() : addrTrans(NULL), socket(NULL) | |
60 | { | |
61 | } | |
62 | ||
63 | MemorySync::MemorySync(int max_strands, int strands_per_core, int cores_per_cpu, int memDebug, int tsoChecker, int callback) : | |
64 | memDebug_(memDebug), | |
65 | addrTrans(NULL) | |
66 | { | |
67 | int i; | |
68 | ||
69 | // keep this in sas.log, so it is easier to check whether msync is enabled | |
70 | // or not. | |
71 | time_t clock; | |
72 | time(&clock); | |
73 | ||
74 | rif_.setMaxStrands(max_strands); | |
75 | rif_.setCoreStrands(strands_per_core); | |
76 | rif_.setCpuCores(cores_per_cpu); | |
77 | mab_.setRieslingInterface(&rif_); | |
78 | ||
79 | ldb_ = new LoadStoreBuffer[MAX_STRANDS]; | |
80 | stb_ = new LoadStoreBuffer[MAX_STRANDS]; | |
81 | retStb_ = new LoadStoreBuffer[MAX_STRANDS]; | |
82 | rmoStb_ = new LoadStoreBuffer[MAX_STRANDS]; | |
83 | ifb_ = new LoadStoreBuffer[MAX_STRANDS]; | |
84 | ||
85 | iseq_ = new uint64_t[MAX_STRANDS]; | |
86 | ||
87 | for (i = 0; i < MAX_STRANDS; i++) { | |
88 | ldb_[i].setBufName((char*) "LoadBuffer"); | |
89 | stb_[i].setBufName((char*) "StoreBuffer"); | |
90 | retStb_[i].setBufName((char*) "RetiredBuffer"); | |
91 | rmoStb_[i].setBufName((char*) "RMOBuffer"); | |
92 | ifb_[i].setBufName((char*) "FetchBuffer"); | |
93 | iseq_[i] = 1; | |
94 | } | |
95 | ||
96 | /** | |
97 | * set Memory preMemoryAccess and postMemoryAccess function pointer | |
98 | */ | |
99 | if (callback > 0) | |
100 | { | |
101 | SS_Memory::memory.msync_object = this; | |
102 | SS_Memory::memory.msync_pre_access = pre_memory_access; | |
103 | SS_Memory::memory.msync_post_access = post_memory_access; | |
104 | ||
105 | SS_Io::io.msync_object = this; | |
106 | SS_Io::io.msync_pre_access = pre_memory_access; | |
107 | SS_Io::io.msync_post_access = post_memory_access; | |
108 | } | |
109 | ||
110 | MemorySyncMessage::debugLevel = memDebug; | |
111 | if ((tsoChecker > 0) && (callback > 0)) { | |
112 | // tsoChecker can become heavy in a multi-strand environment, use it | |
113 | // with caution. By default tsoChecker is off, use | |
114 | // -sas_run_args=-DTSO_CHECKER to enable it. | |
115 | tsoChecker_ = new TsoChecker(); | |
116 | tsoChecker_->init(max_strands); | |
117 | } | |
118 | else { | |
119 | tsoChecker_ = NULL; | |
120 | } | |
121 | ||
122 | if (MemorySync::msyncObj == NULL) { | |
123 | MemorySync::msyncObj = this; | |
124 | } | |
125 | ||
126 | // by default, only core_0 of node_0 is enabled and is able to produce | |
127 | // cache invalidation request. | |
128 | inv_vec_mask[0] = 0x1; | |
129 | for (int i = 1; i < INV_VEC_SIZE; i++) | |
130 | inv_vec_mask[i] = 0x0; | |
131 | } | |
132 | ||
133 | //////////////////////////////////////////////// | |
134 | ||
135 | MemorySync::MemorySync( const MemorySync & orig ) | |
136 | { | |
137 | ||
138 | } | |
139 | ||
140 | //////////////////////////////////////////////// | |
141 | ||
142 | MemorySync::~MemorySync() | |
143 | { | |
144 | } | |
145 | ||
146 | //////////////////////////////////////////////// | |
147 | ||
148 | const MemorySync & | |
149 | MemorySync::operator=( const MemorySync & rhs ) | |
150 | { | |
151 | // Replace the following line with your function body. | |
152 | // RIESLING_THROW_DOMAIN_ERROR( "Unimplemented function." ); | |
153 | ||
154 | return *this; | |
155 | } | |
156 | ||
157 | //////////////////////////////////////////////// | |
158 | ||
159 | bool | |
160 | MemorySync::operator==( const MemorySync & rhs ) const | |
161 | { | |
162 | // Replace the following line with your function body. | |
163 | // RIESLING_THROW_DOMAIN_ERROR( "Unimplemented function." ); | |
164 | return false; | |
165 | } | |
166 | ||
167 | //////////////////////////////////////////////// | |
168 | ||
169 | string | |
170 | MemorySync::toString() const | |
171 | { | |
172 | ostringstream os; | |
173 | ||
174 | os << "mab: " << mab_.toString(); | |
175 | for (int i = 0; i < MAX_STRANDS; i++) { | |
176 | if (ldb_[i].size() > 0) { | |
177 | os << "ldb[" << i << "]: " << ldb_[i].toString(); | |
178 | } | |
179 | if (stb_[i].size() > 0) { | |
180 | os << "stb[" << i << "]: " << stb_[i].toString(); | |
181 | } | |
182 | if (retStb_[i].size() > 0) { | |
183 | os << "retStb[" << i << "]: " << retStb_[i].toString(); | |
184 | } | |
185 | if (rmoStb_[i].size() > 0) { | |
186 | os << "rmoStb[" << i << "]: " << rmoStb_[i].toString(); | |
187 | } | |
188 | if (ifb_[i].size() > 0) { | |
189 | os << "ifb[" << i << "]: " << ifb_[i].toString(); | |
190 | } | |
191 | } | |
192 | ||
193 | return os.str(); | |
194 | } | |
195 | ||
196 | /****************************************************************************** | |
197 | * This implementation assumes each load sends only one LoadIssue command. | |
198 | * If an environment prefers to send 8 LoadIssues for a Block-Load and 2 | |
199 | * LoadIssues for a Quad-Load, then the code can be simplied to handle all | |
200 | * types of load like the default case. | |
201 | ******************************************************************************/ | |
202 | void | |
203 | MemorySync::handleLoadIssue(LoadStoreCmd& cmd) | |
204 | { | |
205 | if (addrTrans) | |
206 | { | |
207 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
208 | } | |
209 | list<LoadStoreEntry>::iterator ii; | |
210 | LoadStoreEntry* entry; | |
211 | LoadStoreBuffer& ldb = ldb_[cmd.getThrdId()]; | |
212 | ||
213 | #ifdef N2MODEL | |
214 | /****************************************************************************** | |
215 | * In general, the RTL should send the data source along with the LoadData. | |
216 | * However, N2 made a special request to also check STB bypassing in Memory | |
217 | * Sync model for double check RTL function. It was told to check this STB | |
218 | * bypassing when receiving LoadData command. However, it turns out that the | |
219 | * RTL actually checks STB bypassing at LoadIssue. (5/17/04). Hence, this | |
220 | * function is moved to here. This is one case of drawback that a model is too | |
221 | * microarchitecture-dependent. Accompany with this change, they also change | |
222 | * their STB bypass check criteria. Hence, my code. | |
223 | ******************************************************************************/ | |
224 | list<LoadStoreEntry>::iterator ste; | |
225 | LoadStoreBuffer& stb = stb_[cmd.getThrdId()]; | |
226 | ||
227 | if (!cmd.isIO() && | |
228 | cmd.getItype() != ITYPE_BLOCK_LOAD && | |
229 | cmd.getItype() != ITYPE_DOUBLE_LOAD && // added 6/1/04 TPS | |
230 | cmd.getItype() != ITYPE_QUAD_LOAD && | |
231 | cmd.getItype() != ITYPE_ATOMIC ) { | |
232 | ste = stb.findN2DataBypassing(cmd.getAddr(), cmd.getSizeV()); | |
233 | if (ste != stb.end()) { // found in STB | |
234 | // if (cmd.getDsrc() != DSRC_STB) { | |
235 | // MS_ERROR("MemorySync found the data in STB; while RTL's data is from %s", mmdsrc[cmd.getDsrc()]); | |
236 | // } | |
237 | cmd.setDsrc(DSRC_STB); | |
238 | } else { // else find data from L2_MEMORY | |
239 | // if (cmd.getDsrc() == DSRC_STB) { | |
240 | // MS_ERROR("RTL indicates data is from STB; while MemorySync does not find the data in STB"); | |
241 | // } | |
242 | cmd.setDsrc(DSRC_NONE); | |
243 | } | |
244 | } | |
245 | #endif | |
246 | ||
247 | // uint64_t iseq = iSeq(cmd.getThrdId()); | |
248 | uint64_t iseq; | |
249 | ||
250 | switch (cmd.getItype()) { | |
251 | case ITYPE_BLOCK_LOAD: | |
252 | cmd.setSize(8); | |
253 | for (int i = 0; i < 8; i++) { | |
254 | entry = new LoadStoreEntry(cmd); | |
255 | entry->setIseq(iSeq(cmd.getThrdId())); // each entry set different iseq | |
256 | ii = ldb.pushBack(*entry); | |
257 | cmd.setAddr(cmd.getAddr() + 8ull); | |
258 | MSYNC_DEBUG(2, "Last Enter to LDB %s", ii->toString().c_str()); | |
259 | } | |
260 | break; | |
261 | case ITYPE_QUAD_LOAD: | |
262 | iseq = iSeq(cmd.getThrdId()); | |
263 | cmd.setSize(8); | |
264 | for (int i = 0; i < 2; i++) { | |
265 | entry = new LoadStoreEntry(cmd); | |
266 | entry->setIseq(iseq); // each entry set same iseq | |
267 | ii = ldb.pushBack(*entry); | |
268 | cmd.setAddr(cmd.getAddr() + 8ull); | |
269 | MSYNC_DEBUG(2, "Last Enter to LDB %s", ii->toString().c_str()); | |
270 | } | |
271 | break; | |
272 | default: | |
273 | entry = new LoadStoreEntry(cmd); | |
274 | entry->setIseq(iSeq(cmd.getThrdId())); | |
275 | ii = ldb.pushBack(*entry); | |
276 | MSYNC_DEBUG(2, "Last Enter to LDB %s", ii->toString().c_str()); | |
277 | break; | |
278 | } | |
279 | } | |
280 | ||
281 | void | |
282 | MemorySync::handleLoadData(LoadStoreCmd& cmd) | |
283 | { | |
284 | if (addrTrans) | |
285 | { | |
286 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
287 | } | |
288 | list<MemoryAccessEntry>::iterator mlink; | |
289 | list<LoadStoreEntry>::iterator lde; | |
290 | list<LoadStoreEntry>::iterator ste; | |
291 | MemoryAccessEntry* mae; | |
292 | int nMaePerLoadData; | |
293 | uint32_t cid = cmd.getCoreId(); | |
294 | uint32_t tid = cmd.getThrdId(); | |
295 | enum DATA_SRC dsrc = cmd.getDsrc(); | |
296 | ||
297 | LoadStoreBuffer& ldb = ldb_[tid]; | |
298 | ||
299 | /****************************************************************************** | |
300 | * Find corresponding load instruction in the Load Buffer | |
301 | * - Note that <address> and <size> are generally not countable since if the | |
302 | * access is to L2, usually only address bits necessary to access cache lines | |
303 | * are observable in the RTL, so does the size parameter. | |
304 | * - <id> is certainly the most dependable way of matching. However, machines | |
305 | * that only allow one outstanding load may consider sending this info over | |
306 | * a un-necessary effort. | |
307 | * | |
308 | * This implementation searches the load buffer for a LoadData by finding the | |
309 | * oldest one that has not yet been associated with a LoadData and whose <id> | |
310 | * and <address> match with the <id> and <address> of the LoadData, where the | |
311 | * <address> comparison ignores the 3 least significant bits. | |
312 | * | |
313 | * In case the <id> is not implemented, then the mechanism still works as long | |
314 | * as the access to the same cache line (or relax more, the same 8-byte block) | |
315 | * results in in-order LoadData. Machines allowing only one outstanding load | |
316 | * certainly meet this condition. I believe many machines that allows multiple | |
317 | * outstanding loads should also meet this criteria. Note that the environment | |
318 | * that does not implement <id> should always set this field 0. | |
319 | ******************************************************************************/ | |
320 | lde = ldb.find1stNoLink (cmd.getId(), cmd.getAddr(), BLK_ADDR_MASK); | |
321 | if (lde == ldb_[tid].end()) | |
322 | { | |
323 | MS_ERROR("LoadData failed to find matched LoadIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
324 | return; | |
325 | } | |
326 | ||
327 | /* Assign size info to the cmd from load buffer so that the MEM_LOAD_DATA | |
328 | can have exact data size. This satisfies two purposes: (1) N2 requests | |
329 | to remove <size> info from LoadData command since it is hard for them | |
330 | to probe correct <size> info in multi-core environment. (6/22/04) (2) | |
331 | the TsoChecker needs to have correct <size> info for load. */ | |
332 | cmd.setSize(lde->getSize()); | |
333 | cmd.setSizeV(lde->getSizeV()); | |
334 | ||
335 | #ifdef N2MODEL_REMOVE | |
336 | /****************************************************************************** | |
337 | * In general, the RTL should send the data source along with the LoadData. | |
338 | * However, N2 made a special request to only indicate whether the data is | |
339 | * from L1$ or not. If the data is not from L1$, then the Memory Sync Model | |
340 | * will have to find out whether the data is in the store buffer. Also, in | |
341 | * N2, one LoadData represent 16 bytes of data for block load and quad load. | |
342 | * ^^^^^^^^^^^ This check moves to LoadIssue. See note in LoadIssue. 5/17/04 | |
343 | ******************************************************************************/ | |
344 | /****************************************************************************** | |
345 | * Four types of accesses in N2 will never get data from the store buffer. | |
346 | * They are IO, Block Load, Qual Load, and atomic accesses. | |
347 | ******************************************************************************/ | |
348 | if (dsrc == DSRC_L2_MEMORY && | |
349 | !cmd.isIO() && | |
350 | lde->getItype() != ITYPE_BLOCK_LOAD && | |
351 | lde->getItype() != ITYPE_QUAD_LOAD && | |
352 | lde->getItype() != ITYPE_ATOMIC ) { | |
353 | ste = stb_[tid].findDataBypassing(cmd.getAddr(), cmd.getSizeV()); | |
354 | if (ste != stb_[tid].end()) { // find in STB | |
355 | dsrc = DSRC_STB; | |
356 | } // else find data from L2_MEMORY | |
357 | } | |
358 | #endif | |
359 | ||
360 | #ifdef N2MODEL | |
361 | /* If the STB hit at LoadIssue, then the LoadData data source is ignored. */ | |
362 | if (lde->getDsrc() == DSRC_STB) { | |
363 | if (dsrc != DSRC_STB) { | |
364 | MS_ERROR("MemorySync found the data in STB; while RTL's data is from %s", mmdsrc[cmd.getDsrc()]); | |
365 | return; | |
366 | } | |
367 | } else { // else find data from L2_MEMORY | |
368 | if (dsrc == DSRC_STB) { | |
369 | MS_ERROR("RTL indicates data is from STB; while MemorySync does not find the data in STB"); | |
370 | return; | |
371 | } | |
372 | } | |
373 | ||
374 | if (lde->getItype() == ITYPE_BLOCK_LOAD || | |
375 | lde->getItype() == ITYPE_QUAD_LOAD) { | |
376 | nMaePerLoadData = 2; | |
377 | } else { | |
378 | nMaePerLoadData = 1; | |
379 | } | |
380 | #else | |
381 | nMaePerLoadData = 1; | |
382 | #endif | |
383 | ||
384 | mae = new MemoryAccessEntry (cmd); | |
385 | mae->setItype(lde->getItype()); | |
386 | mae->setDsrc(dsrc); | |
387 | mae->setIseq(lde->getIseq()); | |
388 | mlink = mab_.pushBack(*mae); | |
389 | lde->setLink(mlink); // load buffer entry has link to the correponding MAB entry | |
390 | ||
391 | /************************************** | |
392 | * if (LDDATA & dsrc=3), | |
393 | * then scan MAB for any STCOM to this PA. STCOM inv_vect does not matter in this case. | |
394 | * | |
395 | * if (LDDATA & dsrc=2) | |
396 | * then scan MAB for STCOM | |
397 | * if (STCOM inv_vect[mycore]=0), then use this Store | |
398 | * else if (STCOM inv_vect[mycore]=1) and (STINV received), then use this Store | |
399 | * else if (STCOM inv_vect[mycore]=1) and (STINV not received), then can't use this Store, keep scanning MAB. | |
400 | * | |
401 | * where mycore is the core number of the LDDATA. | |
402 | */ | |
403 | ||
404 | switch (dsrc) { | |
405 | case DSRC_STB: | |
406 | { | |
407 | LoadStoreBuffer& stb = stb_[tid]; | |
408 | LoadStoreBuffer& rstb = retStb_[tid]; | |
409 | ||
410 | /* find data from store buffer */ | |
411 | ||
412 | /****************************************************************************** | |
413 | * STB bypassing is checked different for N1 and N2 | |
414 | * (also see comments in how to model stores) | |
415 | * o N1 checks both stb_ and retStb_. It checks retired_buf_ becuase there | |
416 | * is still a timing gap allowed for bypassing after a StoreAck is received | |
417 | * in N1. This implementation does not check rmoStb_, however. The assumption | |
418 | * is that no STB bypassing of the rmo store data is possible. In theory, this | |
419 | * assumption is not always true. The timing gap mentioned above should also | |
420 | * possible in RMO case. In general, if a load would like to use the data of | |
421 | * a RMO store, a membar instruction should be introduced. If this is strictly | |
422 | * followed, then there is no any problem not to check rmoStb_. Diags may | |
423 | * fail to guarantee this, the implementation takes the risk any way. | |
424 | * o N2 checks only stb_. It must be there since it is checked before by the | |
425 | * program. | |
426 | ******************************************************************************/ | |
427 | ste = stb.findDataBypassing(cmd.getAddr(), cmd.getSizeV()); | |
428 | if (ste == stb.end()) { // not in STB | |
429 | ste = rstb.findDataBypassing(cmd.getAddr(), cmd.getSizeV()); // try retired_stb | |
430 | if (ste == rstb.end()) | |
431 | { | |
432 | MS_ERROR("LoadData failed to find matched STB bypassing entry. tid=%d PA=%llx", tid, cmd.getAddr()); | |
433 | return; | |
434 | } | |
435 | ||
436 | MSYNC_DEBUG(1, "Store Bypassing gets data from Retired Buffer %s", | |
437 | ste->toString().c_str()); | |
438 | } | |
439 | /* set data in the memory access entry */ | |
440 | #ifdef N2MODEL | |
441 | if (!ste->isDataValid()) | |
442 | { | |
443 | MS_ERROR("LoadData finds Store Buffer data is not ready. tid=%d PA=%llx", tid, cmd.getAddr()); | |
444 | return; | |
445 | } | |
446 | #endif | |
447 | // In N1 swerver-memory.cc environment, the data may not be ready at this time | |
448 | if (ste->isDataValid()) { | |
449 | mlink->setData(ste->getData()); | |
450 | } else { // N2 should not go this path | |
451 | mlink->setData(0ull); // set arbitrary data | |
452 | ste->setLink2(mlink); // when data is ready, it can also set the MAB entry | |
453 | } | |
454 | mlink->setDsrcMid(mlink->getId()); | |
455 | break; | |
456 | } | |
457 | case DSRC_L1: | |
458 | { | |
459 | list<MemoryAccessEntry>::iterator dsrcMlink; | |
460 | ||
461 | dsrcMlink = mab_.findL1DataEntry(cmd); | |
462 | if (dsrcMlink == mab_.end()) { | |
463 | mlink->setData(rif_.readMemory(cid, tid, cmd.getAddr() & ADDR_MASK, 8)); // read aligned 8 bytes | |
464 | // mlink->setDsrcMid(mlink->getId()); | |
465 | mlink->setDsrcMid(mab_.begin()->getId() - 1); // source is from before MAB head | |
466 | } else { | |
467 | mlink->setData(dsrcMlink->getData()); | |
468 | mlink->setDsrcMid(dsrcMlink->getDsrcMid()); | |
469 | } | |
470 | break; | |
471 | } | |
472 | case DSRC_L2_MEMORY: | |
473 | // if it is a I/O address, let's defer the data setting til the | |
474 | // corresponding load instruction is invoked, where we may pick up a | |
475 | // (follow-me) CSR_READ | |
476 | if ((cmd.getAddr() & IO_ADDR_BIT_MASK) != IO_ADDR_BIT_MASK) | |
477 | { | |
478 | // not an I/O address, should set the value here so that entries | |
479 | // after this may use it. | |
480 | mlink->setData(getL2Data(mab_.end(), cmd)); | |
481 | } | |
482 | mlink->setDsrcMid(mlink->getId()); | |
483 | ||
484 | if (nMaePerLoadData > 1) { | |
485 | for (int i = 1; i < nMaePerLoadData; i++) { | |
486 | lde++; | |
487 | cmd.setAddr(cmd.getAddr() + 8); // must be 8 byte per access if nMaePerLoadData > 1 | |
488 | if (lde == ldb_[tid].end()) | |
489 | { | |
490 | MS_ERROR("LoadData failed to find matched LoadIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
491 | return; | |
492 | } | |
493 | MSYNC_DEBUG(2, "lde->getAddr()=%llx, cmd.getAddr()=%llx", | |
494 | lde->getAddr(), cmd.getAddr()); | |
495 | if (lde->getAddr() != cmd.getAddr()) | |
496 | { | |
497 | MS_ERROR("Mismatch address between LoadData and LoadIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
498 | return; | |
499 | } | |
500 | ||
501 | mae = new MemoryAccessEntry (cmd); | |
502 | mae->setItype(lde->getItype()); | |
503 | mae->setDsrc(dsrc); | |
504 | // if it is not a I/O address, ditto | |
505 | if ((cmd.getAddr() & IO_ADDR_BIT_MASK) != IO_ADDR_BIT_MASK) | |
506 | { | |
507 | // ditto | |
508 | mae->setData(getL2Data(mlink, cmd)); | |
509 | } | |
510 | mae->setDsrcMid(mae->getId()); | |
511 | mae->setIseq(lde->getIseq()); | |
512 | mlink = mab_.pushBack(*mae); | |
513 | lde->setLink(mlink); // load buffer entry has link to the correponding MAB entry | |
514 | } | |
515 | } | |
516 | break; | |
517 | default: | |
518 | MS_ERROR("LoadData receives unknown data source. tid=%d PA=%llx", tid, cmd.getAddr()); | |
519 | return; | |
520 | } | |
521 | } | |
522 | ||
523 | void | |
524 | MemorySync::handleLoadFill (LoadStoreCmd& cmd) | |
525 | { | |
526 | if (addrTrans) | |
527 | { | |
528 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
529 | } | |
530 | list<MemoryAccessEntry>::iterator mlink, newlink; | |
531 | MemoryAccessEntry entry (cmd); | |
532 | ||
533 | mlink = mab_.findLoadFillSrc(cmd); | |
534 | if (mlink == mab_.end()) | |
535 | { | |
536 | MS_ERROR("LoadFill failed to find matched LoadData. tid=%d PA=%llx", cmd.getThrdId(), cmd.getAddr()); | |
537 | return; | |
538 | } | |
539 | // //TODO this assumes the corresponding load entry is popped by LOAD_POP, | |
540 | // // but it can be an error instead, so need a better solution. | |
541 | // if (mlink == mab_.end()) return; | |
542 | ||
543 | entry.setData(mlink->getData()); | |
544 | entry.setExecuted(true); // set for being retired from MAB when it reaches to head | |
545 | entry.setLink(mlink); | |
546 | entry.setDsrcMid(mlink->getId()); | |
547 | newlink = mab_.pushBack(entry); | |
548 | mlink->setLink(newlink); | |
549 | ||
550 | /* Also get the other part of L1$ line data, TPS 4/22/04 */ | |
551 | MemoryAccessEntry* e; | |
552 | uint64_t line_addr_mask = ~((uint64_t) (DCACHE_LINE_SIZE - 1)); | |
553 | uint64_t addr = cmd.getAddr() & line_addr_mask; | |
554 | ||
555 | for (int i = 0; i < DCACHE_LINE_SIZE/8; i++, addr += 8) { | |
556 | if ((mlink->getAddr() & ADDR_MASK) == (addr & ADDR_MASK)) { | |
557 | continue; | |
558 | } | |
559 | ||
560 | cmd.setAddr(addr); | |
561 | e = new MemoryAccessEntry (cmd); | |
562 | e->setData(getL2Data(mlink, cmd)); // the same line must get data from the same time | |
563 | e->setExecuted(true); // set true so that when it reach to head, it can be removed | |
564 | e->setLink(mlink); // set to mlink indicating its search starts from here | |
565 | e->setDsrcMid(mlink->getId()); | |
566 | mab_.pushBack(*e); | |
567 | } | |
568 | } | |
569 | ||
570 | /****************************************************************************** | |
571 | * %%% Store Behavior | |
572 | * | |
573 | * %% General Implementation Note | |
574 | * | |
575 | * - StoreAck | |
576 | * By definition, StoreAck is sent when the corresponding StoreIssue is to be | |
577 | * removed from the Store Buffer so that no STB bypassing from this entry is | |
578 | * possible. | |
579 | * | |
580 | * However, N1 still has a timing window that the STB bypassing is allowed | |
581 | * after the StoreAck is sent. Hence, a retStb_ is added to temporarily | |
582 | * hold the Acked entries. | |
583 | * | |
584 | * N2 does obey this rule, its behavior is described below: | |
585 | * 1. Normal store | |
586 | * In Niagara 2, the design of the StoreAck can be abstracted as follows: | |
587 | * StoreAcks are queued, and are released from the queue in order. If a | |
588 | * StoreAck also causes store update or invalidation, the StoreAck will be | |
589 | * held in the queue until its update/invalidation is done. | |
590 | * | |
591 | * Upon a StoreAck is released from the queue, a StoreAck is sent to the | |
592 | * Memory Sync model followed by, if any, a StoreUpdate or StoreInv. Hence, | |
593 | * the Memory Sync model should see in-order StoreAcks. | |
594 | * | |
595 | * For normal store, StoreAck is always after StoreCommit. | |
596 | * | |
597 | * 2. RMO Store (Block Store, stores using *_BLOCK_INIT_ST_QUAD_LDD_*) | |
598 | * By definition, these stores do not check data dependence, and follow RMO | |
599 | * memory consistency model. They leave the StoreBuffer when they are issued. | |
600 | * A StoreAck will be sent to reflect this fact. Hence, this type of stores | |
601 | * should have StoreAck before its StoreCommit. A data structure, rmoStb_ | |
602 | * is introduced to hold these stores after they are removed from the store | |
603 | * buffer so that the StoreCommit can still find the matched entry to extract | |
604 | * some needed info from StoreIssue. | |
605 | * | |
606 | * - Order Between STEP and StoreAck | |
607 | * Note that this implementation does not assume the order of a STEP and its | |
608 | * corresonding StoreAck. Different CPU can have different implementation. | |
609 | * The following are some cases: | |
610 | * > Normal stores usually have their STEP before StoreAck. | |
611 | * > Atomic alwyas have STEP after StoreAck in N2, but may have STEP before | |
612 | * StoreAck in N1. | |
613 | * > RMO Stores could have StoreAck before or after STEP. | |
614 | * | |
615 | * | |
616 | * %% MemorySync Store Model | |
617 | * | |
618 | * - When a StoreIssue is received, an entry is place into store buffer | |
619 | * | |
620 | * - When a StoreAck is received, the corresponding StoreIssue must be at the head | |
621 | * of the STB. The entry is then moved to | |
622 | * o rmoStb_[tid], if cmd.isRMOstore() is true | |
623 | * o retStb_[tid], otherwise | |
624 | * | |
625 | * - When a StoreCommit is received, the program searches for both stb_[tid] | |
626 | * (must be in-order) and rmo_stb[tid] (may be out-of-order). Normal stores | |
627 | * must obey TSO, hence, N1/N2 cannot issue next store, unless in the same | |
628 | * cache line, until a StoreAck is received. Therefore, the StoreCommit must | |
629 | * be in-order. | |
630 | * However, RMO stores follow RMO memory consistency model. A RMO store can | |
631 | * be issued from the store buffer without waiting for the commit of a prior | |
632 | * RMO store. As a result, the MemorySync may see out-of-order StoreCommit | |
633 | * for RMO stores since stores access different banks may commit out-of issue | |
634 | * order. Nevertheless, accesses to the same bank should still follow issue order. | |
635 | * | |
636 | * In summay, in either case, StoreCommit for the same cache line should be | |
637 | * in-order, which is the assumption made to match StoreCommit-StoreIssue in | |
638 | * handleStoreCommit() method. | |
639 | * | |
640 | * - When a STEP (callback) is received, the program searches for both stb_[tid] | |
641 | * (must be in-order) and rmo_stb[tid] (out-of-order or in-order ?). | |
642 | * | |
643 | * - When a StoreUpdate/StoreInv is received, it finds the StoreCommit that | |
644 | * initiates this StoreUpdate/StoreInv. No need to match it with entry in | |
645 | * either stb_, rmoStb_, or retStb_. | |
646 | * | |
647 | * => a rmoStb_ entry is removed when it has issued to L2 and is executed | |
648 | * => a retire_stb_ entry is removed when a new entry is entered and the | |
649 | * retire_stb_ size is greater than 2 (2 is a arbitrary number since this | |
650 | * retire_stb_ is for N1 compatibility to keep the entry enough time for | |
651 | * STB bypassing) | |
652 | ******************************************************************************/ | |
653 | ||
654 | void | |
655 | MemorySync::handleStoreIssue(LoadStoreCmd& cmd) | |
656 | { | |
657 | if (addrTrans) | |
658 | { | |
659 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
660 | } | |
661 | list<LoadStoreEntry>::iterator ii; | |
662 | LoadStoreEntry ste (cmd); | |
663 | ||
664 | #ifndef N1MODEL // <data> is expected, however, N1 does not have data at this moment | |
665 | ste.setState(LS_TDATA); | |
666 | #endif | |
667 | if (cmd.getItype() == ITYPE_ATOMIC) { // assume atomic-store always directly | |
668 | ste.setIseq(getIseq(cmd.getThrdId()) - 1); // follows atomic-load | |
669 | } else { | |
670 | ste.setIseq(iSeq(cmd.getThrdId())); | |
671 | } | |
672 | ii = stb_[cmd.getThrdId()].pushBack(ste); | |
673 | ||
674 | MSYNC_DEBUG(2, "Last Enter to STB %s", ii->toString().c_str()); | |
675 | } | |
676 | ||
677 | void | |
678 | MemorySync::handleStoreCommit (LoadStoreCmd& cmd) | |
679 | { | |
680 | if (addrTrans) | |
681 | { | |
682 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
683 | } | |
684 | int tid = cmd.getThrdId(); | |
685 | list<MemoryAccessEntry>::iterator mlink; | |
686 | list<LoadStoreEntry>::iterator slink; | |
687 | LoadStoreBuffer& stb = stb_[tid]; | |
688 | LoadStoreBuffer& rmostb = rmoStb_[tid]; | |
689 | MemoryAccessEntry entry (cmd); | |
690 | uint64_t mdata; // mab data | |
691 | uint64_t sdata; // stb data | |
692 | uint8_t mszv; // size vector | |
693 | uint8_t sszv; // size vector | |
694 | bool zeroCacheLine; | |
695 | ||
696 | /****************************************************************************** | |
697 | * Similar to the reason in LoadData, when matching the StoreCommit to | |
698 | * StoreIssue, the implementation searches the store buffer by finding the | |
699 | * oldest one that has not yet been associated with a StoreCommit and whose <id> | |
700 | * and <address> match with the <id> and <address> of the StoreCommit, where the | |
701 | * <address> comparison ignores the 3 least significant bits. | |
702 | * | |
703 | * When the <id> is not implemented, this mechanism works as long as the access | |
704 | * to the same cache line (or relax more, the same 8-byte block) results in | |
705 | * in-order StoreCommit. This condition should be true for cacheable access. | |
706 | * For non-cacheable access, then whether the condition is met depends on the | |
707 | * implementation decision. The guess is yes. For I/O, Sun's CPU design will | |
708 | * most likely to enforce SYNC before and after. Note that the environment | |
709 | * that does not implement <id> should always set this field 0. | |
710 | ******************************************************************************/ | |
711 | ||
712 | #ifdef N3MODEL // N1 assume RMO store's StoreAck is after StoreCommit as other store | |
713 | slink = stb.find1stNoLink (cmd.getId(), cmd.getAddr(), BLK_ADDR_MASK); | |
714 | if (slink == stb.end()) | |
715 | { | |
716 | MS_ERROR("StoreCommit failed to find matched StoreIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
717 | return; | |
718 | } | |
719 | slink->setState(LS_ISSUE); | |
720 | ||
721 | #else // N2's RMO store has StoreAck before StoreCommit | |
722 | slink = rmostb.find1stNoLink (cmd.getId(), cmd.getAddr(), BLK_ADDR_MASK); | |
723 | if (slink == rmostb.end()) { | |
724 | slink = stb.find1stNoLink (cmd.getId(), cmd.getAddr(), BLK_ADDR_MASK); | |
725 | if (slink == stb.end()) | |
726 | { | |
727 | MS_ERROR("StoreCommit failed to find matched StoreIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
728 | return; | |
729 | } | |
730 | slink->setState(LS_ISSUE); | |
731 | } else { | |
732 | entry.setAcked(true); | |
733 | } | |
734 | #endif | |
735 | /****************************************************************************** | |
736 | * Special handling for BLK_INIT store that in a certain case will zero the | |
737 | * whole case line. This type of instruction is supported in N1 and N2. | |
738 | ******************************************************************************/ | |
739 | if ((slink->getItype() == ITYPE_STORE_INIT) && | |
740 | !slink->isIO() && ((slink->getAddr() & 0x3f) == 0) && !cmd.isL2hit()) { | |
741 | zeroCacheLine = true; | |
742 | } else { | |
743 | zeroCacheLine = false; | |
744 | } | |
745 | ||
746 | /* merge data and saved the merged data in the correponding memory access entry */ | |
747 | sdata = slink->getData(); | |
748 | ||
749 | /* Origianl, sizeV == 0 indicates CAS comparison false. Change to use a separate flag | |
750 | to indicate this status. 9/10/04 */ | |
751 | // sszv = (cmd.getSizeV() == 0) ? 0 : slink->getSizeV(); // size_vector = 0 in CMD => cas comparison false | |
752 | sszv = (!cmd.isSwitchData()) ? 0 : slink->getSizeV(); // !isSwitchData() => cas comparison false | |
753 | ||
754 | entry.setMerged(false); | |
755 | if (zeroCacheLine) { // special case for STORE_INIT | |
756 | mdata = merge (0ull, sdata, sszv); | |
757 | } else { | |
758 | if (slink->getSize() == 8 && sszv == 0xff) { // no merge needed | |
759 | mdata = sdata; | |
760 | } else { // sszv = 0 still need to get correct data in the entry | |
761 | mdata = getL2Data(mab_.end(), cmd); | |
762 | mdata = merge (mdata, sdata, sszv); | |
763 | // keep merge-related data in case we need to repeat it (with a different | |
764 | // mdata) | |
765 | entry.setMerged(true); | |
766 | entry.setOrigData(sdata); | |
767 | entry.setOrigSizeV(sszv); | |
768 | } | |
769 | } | |
770 | entry.setData(mdata); | |
771 | entry.setExecuted(slink->isExecuted()); | |
772 | ||
773 | if (sszv == 0) entry.setSizeV(0); // This is possible in N1 for CAS | |
774 | ||
775 | entry.setItype(slink->getItype()); | |
776 | entry.setIseq(slink->getIseq()); | |
777 | mlink = mab_.pushBack(entry); | |
778 | mlink->setLink(mlink); // point to itself | |
779 | slink->setLink(mlink); // STB entry's link pointed to the STORE_COMMIT in MAB | |
780 | // if inv_vec is not zero, re-adjust it to mask out sniper related | |
781 | // invalidation, because sniper does not generate related inv command, | |
782 | // otheriwse the inv_vec bit(s) related to sniper will stay on, and the | |
783 | // store_commit will stay in mab forever. | |
784 | if (mlink->getInv()) | |
785 | { | |
786 | uint32_t mask = 0; | |
787 | for (int i=0; i<4; i++) | |
788 | { | |
789 | mask |= (((uint32_t)inv_vec_mask[i]) << (i*8)); | |
790 | } | |
791 | mlink->setInv(mlink->getInv() & mask); | |
792 | } | |
793 | ||
794 | if (zeroCacheLine) { // Add StoreCommit entries with 0 data for the l2 cache line | |
795 | MemoryAccessEntry* zero_entry; | |
796 | for (int i = 1; i < L2CACHE_LINE_SIZE/8; i++) { | |
797 | zero_entry = new MemoryAccessEntry (cmd); | |
798 | zero_entry->setAddr(slink->getAddr() + (i << 3)); | |
799 | zero_entry->setData(0ull); | |
800 | zero_entry->setExecuted(true); | |
801 | zero_entry->setSizeV(0xff); | |
802 | zero_entry->setInv(0); | |
803 | zero_entry->setAcked(true); | |
804 | zero_entry->setItype(ITYPE_STORE_INIT); | |
805 | zero_entry->setIseq(slink->getIseq()); | |
806 | mlink = mab_.pushBack(*zero_entry); | |
807 | mlink->setLink(mlink); // point to itself | |
808 | } | |
809 | } | |
810 | ||
811 | /* Manage the rmoStb_ */ | |
812 | if (slink->isRMOstore() && slink->isExecuted()) { | |
813 | rmostb.erase(slink); | |
814 | } | |
815 | ||
816 | /* Added to speed up the drain of MAB - 6/3/04 TPS */ | |
817 | vector<MemoryAccessEntry> wdbuf; | |
818 | ||
819 | mab_.popFront(&wdbuf, tsoChecker_); | |
820 | bufWriteMemory(wdbuf); | |
821 | } | |
822 | ||
823 | void | |
824 | MemorySync::handleStoreAck (LoadStoreCmd& cmd) | |
825 | { | |
826 | if (addrTrans) | |
827 | { | |
828 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
829 | } | |
830 | uint32_t tid = cmd.getThrdId(); | |
831 | LoadStoreBuffer& stb = stb_[tid]; | |
832 | list<LoadStoreEntry>::iterator ii; | |
833 | ||
834 | MSYNC_DEBUG(2, "stb=%s", stb.toString().c_str()); | |
835 | ||
836 | /****************************************************************************** | |
837 | * In Niagara 1, the StoreAck is issued in-order, but they may | |
838 | * Note that in N1, rmo stores, BLK_INIT store and BLOCK_STORE, are Acked same | |
839 | * as other stores. Since rmo stores can be issued to L2 without waiting for | |
840 | * the ACK before issuing another one, their StoreAcks can be out-of-order if | |
841 | * they are from different banks. | |
842 | * In N2, StoreAck for a rmo store is sent when it is released from the store | |
843 | * buffer. Hence, it is still in order. | |
844 | * | |
845 | * The N1MODEL may have a potential problem. STORE_ACK CMD has only address, | |
846 | * N1's swerver-memory.cc uses address and size for searching. So, a <size> | |
847 | * field may be need in my specification. However, as the development of this | |
848 | * code, N1 is not a target to support. | |
849 | ******************************************************************************/ | |
850 | #ifdef N1MODEL // cannot just search by address, entries may have same address | |
851 | ii = stb.findMatchO2Y(cmd.getAddr(), cmd.isRMOstore(), FULL_ADDR_MASK); | |
852 | if (ii == stb.end()) | |
853 | { | |
854 | MS_ERROR("StoreAck failed to find matched StoreIssue. tid=%d PA=%llx", tid, cmd.getAddr()); | |
855 | return; | |
856 | } | |
857 | #else // N2 does not send address, and guarantees to send StoreAck in order | |
858 | ii = stb.begin(); | |
859 | if (ii == stb.end()) | |
860 | { | |
861 | MS_ERROR("StoreAck sees empty STB. tid=%d PA=%llx", tid, cmd.getAddr()); | |
862 | return; | |
863 | } | |
864 | #endif | |
865 | ||
866 | #ifdef N1MODEL // N2's StoreAck does not have address | |
867 | MSYNC_DEBUG (2, "addr-stb=%#llx addr-cmd=%#llx", ii->getAddr(), cmd.getAddr()); | |
868 | if (ii->getAddr() != cmd.getAddr()) | |
869 | { | |
870 | MS_ERROR("StoreAck is issued out-of-order, should be in-order. tid=%d PA=%llx", tid, cmd.getAddr()); | |
871 | return; | |
872 | } | |
873 | #endif | |
874 | ||
875 | /* In n1 swerer-memory.cc, if the atomic, block store, or store init needs | |
876 | not to invalidate the L1$, then they are also copied to the thrd_retired_buf. | |
877 | Removing the conditions below for compatibility. For cases that not copying | |
878 | into the thrd_retired_buf, the code is inserted with the RetireBuffer.popBack() | |
879 | to remove the entry. */ | |
880 | // if (ii->isExecuted() && ii->getItype() != ITYPE_ATOMIC && | |
881 | // ii->getItype() != ITYPE_BLOCK_STORE && ii->getItype() != ITYPE_STORE_INIT) { | |
882 | ||
883 | if (cmd.isRMOstore()) { | |
884 | /* No state change for RMO store whose StoreAck is before StoreCommit */ | |
885 | if (!ii->isRMOstore()) | |
886 | { | |
887 | MS_ERROR("RMO StoreAck find non-RMO entry in STB. tid=%d PA=%llx", tid, cmd.getAddr()); | |
888 | return; | |
889 | } | |
890 | rmoStb_[tid].pushBack(*ii); | |
891 | } else { | |
892 | ii->setState(LS_ACK); | |
893 | if (!ii->isExecuted() && (ii->getItype() != ITYPE_ATOMIC)) | |
894 | { | |
895 | MS_ERROR("StoreAck before SSTEP for non-atomic non-RMO store. tid=%d PA=%llx", tid, cmd.getAddr()); | |
896 | return; | |
897 | } | |
898 | retStb_[tid].pushBack(*ii); // move to retired store buffer, for N1 | |
899 | if (retStb_[tid].size() > 2) { | |
900 | retStb_[tid].popFront(); | |
901 | } | |
902 | } | |
903 | stb.erase(ii); | |
904 | ||
905 | /* Add to speed-up the drain of MAB */ | |
906 | vector<MemoryAccessEntry> wdbuf; | |
907 | ||
908 | mab_.popFront(&wdbuf, tsoChecker_); | |
909 | bufWriteMemory(wdbuf); | |
910 | ||
911 | /* StoreAck is not invloved in determining if a StoreCommit can be removed, | |
912 | hence, there is no need to call handleStoreInv(cmd) | |
913 | */ | |
914 | } | |
915 | ||
916 | void | |
917 | MemorySync::handleStoreInv (LoadStoreCmd& cmd) | |
918 | { | |
919 | if (addrTrans) | |
920 | { | |
921 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
922 | } | |
923 | uint32_t tid = cmd.getThrdId(); | |
924 | LoadStoreBuffer& rstb = retStb_[tid]; | |
925 | LoadStoreBuffer& rmostb = rmoStb_[tid]; | |
926 | list<LoadStoreEntry>::iterator rse, rme; | |
927 | ||
928 | /****************************************************************************** | |
929 | * StoreInv command does not provide enough info for us to check matching | |
930 | * store buffer entry as that is done in StoreUpate. StoreUpdate can only | |
931 | * happens on normal store, hence, we only have to check retStb_. But, | |
932 | * StoreInv can also caused by RMO store. The StoreInv API is not required | |
933 | * to differentiate RMO store from others. Therefore, in general a StoreInv | |
934 | * can match with StoreIssue in either retStb_ or rmoStb_, or both. | |
935 | * It is possible to implement this check. The intruction type can be obtained | |
936 | * by matched StoreCommit. Another issue is that the code does not required | |
937 | * a store in a store buffer waiting the StoreUpdate/StoreInv before it can | |
938 | * be removed. So, just remove the check. | |
939 | ******************************************************************************/ | |
940 | // if (cmd.getThrdId() == cmd.getSrcTid()) { // same thread | |
941 | ||
942 | // #ifdef N2MODEL | |
943 | // /* N2's StoreInv set 3 LSBs to 0, hence cannot compare exact address. | |
944 | // However, it must be Acked and Inv in order. */ | |
945 | // rse = --(rstb.end()); // note that (rsb.end())-- does not work | |
946 | // MS_ASSERT((cmd.getAddr() & ADDR_MASK) == (rse->getAddr() & ADDR_MASK), "StoreInv finds last StoreAck having different address", cmd.getAddr()); | |
947 | // #else | |
948 | // rse = rstb.findMatchO2Y(cmd.getAddr(), LS_ACK); // update or inv to the same core should be in-order | |
949 | // MS_ASSERT(rse != rstb.end(), "StoreInv failed to find matched one in ther RetiredBuffer", cmd.getAddr()); | |
950 | // #endif | |
951 | // rse->setState(LS_INV); | |
952 | // } | |
953 | ||
954 | completeStoreInvStoreUpdate(cmd); | |
955 | } | |
956 | ||
957 | /******************************************************************************** | |
958 | * StoreUpdate command provides two purposes. One is exact the same as StoreInv | |
959 | * that prevent the corresponding StoreCommit from removing from the MAB too | |
960 | * early. The other is to provide data for the following LoadData with L1 hit. | |
961 | * Recall that the LoadData with L1 hit will search for its data from the same | |
962 | * core LoadFill or StoreUpdate record in the MAB. | |
963 | ********************************************************************************/ | |
964 | ||
965 | void | |
966 | MemorySync::handleStoreUpdate (LoadStoreCmd& cmd) | |
967 | { | |
968 | if (addrTrans) | |
969 | { | |
970 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
971 | } | |
972 | uint32_t tid = cmd.getThrdId(); | |
973 | LoadStoreBuffer& rstb = retStb_[tid]; | |
974 | list<LoadStoreEntry>::iterator ii; | |
975 | ||
976 | /****************************************************************************** | |
977 | * The following code does check. It is not functionally critical. In SUN's | |
978 | * implementation, usually the block stores (and in N2 BLK_INIT stores) do not | |
979 | * update L1$, instead, they invalidate L1$. Hence, only retStb_ needs to | |
980 | * perform search for matching StoreIssue. | |
981 | * | |
982 | * In addition, Niagara 2 sends StoreAck and StoreUpdate or StoreInv (same cid) | |
983 | * in the same cycle. In N2, the design of the StoreAck/StoreUpdate/StoreInv can | |
984 | * be abstracted as follows: | |
985 | * StoreAcks are queued, and are released from the queue in order. If a | |
986 | * StoreAck also causes store update or invalidation, the StoreAck will be | |
987 | * held in the queue until its update/invalidation is done. | |
988 | * | |
989 | * Niagara 1, however, does not queue the StoreAck. Hence, the StoreUpdate (and | |
990 | * StoreInv) may not related to the last StoreAck since not all StoreAcks cause | |
991 | * StoreUpdate/StoreInv. | |
992 | ******************************************************************************/ | |
993 | #ifdef N2MODEL | |
994 | /* N2's StoreUpdate set 3 LSBs to 0, hence cannot compare exact address. | |
995 | However, it must be Acked and Updated in order. */ | |
996 | ii = --(rstb.end()); // note that (rsb.end())-- does not work | |
997 | // cerr << "MemSync: " << ii->toString(); | |
998 | if ((cmd.getAddr() & ADDR_MASK) != (ii->getAddr() & ADDR_MASK)) | |
999 | { | |
1000 | MS_ERROR("StoreUpdate finds last StoreAck having different address. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1001 | return; | |
1002 | } | |
1003 | #else | |
1004 | /****************************************************************************** | |
1005 | * In general, searching for state LS_ACK is not ok since as discussed in the | |
1006 | * handleStoreInv StoreInv is not appropriate to perform such check, and thus | |
1007 | * the state of the corresponding StoreIssue remains at state LS_ACK. As a | |
1008 | * result, the search may find wrong entry. The longer the retire_stb_, the | |
1009 | * more likely the error. So far, the retire_stb_ maximum size is set to 2. | |
1010 | ******************************************************************************/ | |
1011 | ii = rstb.findMatchO2Y(cmd.getAddr(), LS_ACK, BLK_ADDR_MASK); // update or inv to the same core should be in-order | |
1012 | if (ii == rstb.end()) | |
1013 | { | |
1014 | MS_ERROR("StoreUpdate failed to find matched one in the RetiredBuffer. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1015 | return; | |
1016 | } | |
1017 | #endif | |
1018 | ii->setState(LS_UPDATE); | |
1019 | ||
1020 | completeStoreInvStoreUpdate(cmd); | |
1021 | } | |
1022 | ||
1023 | void | |
1024 | MemorySync::completeStoreInvStoreUpdate (LoadStoreCmd& cmd) | |
1025 | { | |
1026 | list<MemoryAccessEntry>::iterator srclink; | |
1027 | list<MemoryAccessEntry>::iterator newlink; | |
1028 | vector<MemoryAccessEntry> wdbuf; | |
1029 | MemoryAccessEntry mae (cmd); | |
1030 | uint32_t cbit = 1 << cmd.getCoreId(); | |
1031 | ||
1032 | srclink = mab_.findStoreInvStoreUpdateSrc (cmd); | |
1033 | ||
1034 | #ifdef N1MODEL // it seems that in N1 IO access does not have StoreCommit signal | |
1035 | if (cmd.isIO() && srclink == mab_.end()) { | |
1036 | return; | |
1037 | } | |
1038 | #endif | |
1039 | ||
1040 | if (srclink == mab_.end()) | |
1041 | { | |
1042 | MS_ERROR("StoreInv/Update failed to find matched StoreCommit. tid=%d PA=%llx", cmd.getSrcTid(), cmd.getAddr()); | |
1043 | return; | |
1044 | } | |
1045 | MSYNC_DEBUG(4, "StoreInv-StoreUpdate %s", srclink->toString().c_str()); | |
1046 | srclink->addCinv(cbit); // record invalidated/updated core | |
1047 | ||
1048 | mae.setData(srclink->getData()); // although only StoreUpdate needs data, it's ok for StoreInv | |
1049 | mae.setExecuted(true); // required for retirement from MAB when it reaches to head | |
1050 | ||
1051 | // initially, a StoreCommit will have a link to itself | |
1052 | if (!srclink->isLinkValid()) | |
1053 | { | |
1054 | MS_ERROR("StoreCommit does not have a valid link"); | |
1055 | return; | |
1056 | } | |
1057 | mae.setLink(srclink->getLink()); // set link chain for debugging purpose | |
1058 | mae.setDsrcMid(srclink->getId()); // data source MAE id | |
1059 | srclink->setGlobal(mae.getId()); // Global will be set the last StoreInv or StoreUpdate, eventually | |
1060 | newlink = mab_.pushBack(mae); | |
1061 | srclink->setLink(newlink); | |
1062 | ||
1063 | /* Add to speed-up the drain of MAB */ | |
1064 | mab_.popFront(&wdbuf, tsoChecker_); | |
1065 | bufWriteMemory(wdbuf); | |
1066 | } | |
1067 | ||
1068 | void | |
1069 | MemorySync::handleEvict (LoadStoreCmd& cmd) | |
1070 | { | |
1071 | if (addrTrans) | |
1072 | { | |
1073 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1074 | } | |
1075 | /* Note that if the D$ line size is different from the I$ line size, | |
1076 | then the following code may have to be modified. Although N2 has | |
1077 | differernt line size of D$ and I$, it is ok to just use the line | |
1078 | size of D$ since the N2 is decided not to handle I-Fetch side of | |
1079 | operation. */ | |
1080 | /* For MemorySync to function correctly, only one EVICT entry per | |
1081 | EVICT command is enough since it functions as a fence. However, | |
1082 | for TSOchecker to fucntion correctly, each 8-byte needs to have | |
1083 | a EVICT so that the TSOchecker can depend on this to retire store | |
1084 | node in the memory access history structure. */ | |
1085 | if (cmd.getInv() == 0) | |
1086 | { | |
1087 | MS_ERROR("MEM_EVICT has inv_vec==0x0. cmd=%s", cmd.toString().c_str()); | |
1088 | return; | |
1089 | } | |
1090 | uint64_t base = cmd.getAddr() & (~(DCACHE_LINE_SIZE - 1)); | |
1091 | uint64_t addr; | |
1092 | for (int i = 0; i < DCACHE_LINE_SIZE/8; i++) | |
1093 | { | |
1094 | addr = base + (8 * i); | |
1095 | cmd.setAddr(addr); | |
1096 | cmd.setSrcBank ((addr & L2_BANK_ADDR_BITS) >> L2_BANK_ADDR_SFT); | |
1097 | list<MemoryAccessEntry>::iterator ii = mab_.findDmaStoreStart(cmd, DMA_EVICT); | |
1098 | if (ii != mab_.end()) | |
1099 | { | |
1100 | // a remote evict for a dma_store, store its inv in the dma_store_start, | |
1101 | // it will be picked up when the corresponding dma_store is issued later | |
1102 | ii->setInv((ii->getInv() | cmd.getInv())); | |
1103 | } | |
1104 | else | |
1105 | { | |
1106 | ii = mab_.findDmaStoreEntry(cmd, DMA_EVICT); | |
1107 | if (ii != mab_.end()) | |
1108 | { | |
1109 | // find one, this evict is for a dma_store entry, it is mainly used to | |
1110 | // fill in the dma_store's inv_vec field. No need to add this entry | |
1111 | // to MAB. the dma_store may contain remote evict, so OR them together | |
1112 | ii->setInv((ii->getInv() | cmd.getInv())); | |
1113 | // we should see no more evict associated with this entry | |
1114 | ii->setInvSet(true); | |
1115 | // let inv/cinv decide when this entry can be removed from MAB | |
1116 | ii->setExecuted(true); | |
1117 | //ii->setSrcBank((cmd.getAddr() & L2_BANK_ADDR_BITS) >> L2_BANK_ADDR_SFT); | |
1118 | MSYNC_DEBUG(4, "EVICT DMA_STORE %s", ii->toString().c_str()); | |
1119 | } | |
1120 | else | |
1121 | { | |
1122 | // normal MEM_EVICT | |
1123 | list<MemoryAccessEntry>::iterator newlink; | |
1124 | MemoryAccessEntry* mae = new MemoryAccessEntry (cmd); | |
1125 | //mae->setAddr(addr); | |
1126 | mae->setData(0ull); // Evict does not set real data | |
1127 | mae->setExecuted(true); // needed for retirement from MAB | |
1128 | mae->setAcked(true); | |
1129 | //mae->setSrcBank ((addr & L2_BANK_ADDR_BITS) >> L2_BANK_ADDR_SFT); | |
1130 | mae->setSizeV(0xff); | |
1131 | newlink = mab_.pushBack(*mae); | |
1132 | newlink->setLink(newlink); // set link to itself | |
1133 | } | |
1134 | } | |
1135 | } | |
1136 | } | |
1137 | ||
1138 | void | |
1139 | MemorySync::handleEvictInv (LoadStoreCmd& cmd) | |
1140 | { | |
1141 | if (addrTrans) | |
1142 | { | |
1143 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1144 | } | |
1145 | // evict_inv does not have addr value, so it is always 0 to begin with. | |
1146 | uint64_t base = cmd.getAddr() & (~(DCACHE_LINE_SIZE - 1)); | |
1147 | uint64_t addr; | |
1148 | uint32_t cbit = 1 << cmd.getCoreId(); | |
1149 | /** | |
1150 | * Note using DCACHE_LINE_SIZE implies only one EvictInv can be sent | |
1151 | * from each core assocaited with one Evict command. | |
1152 | */ | |
1153 | for (int i = 0; i < DCACHE_LINE_SIZE/8; i++) | |
1154 | { | |
1155 | addr = base + (8 * i); | |
1156 | cmd.setAddr(addr); | |
1157 | // for evict_inv, srcBank is provided by testbench, this can be a problem | |
1158 | // when partial bank mode is used. ===> it will not, testbench uses static | |
1159 | // bank calculation too. | |
1160 | list<MemoryAccessEntry>::iterator ii = mab_.findDmaStoreStart(cmd, DMA_EVICT_INV); | |
1161 | if (ii != mab_.end()) | |
1162 | { | |
1163 | // a remote evict_inv for a dma_store, update the cinv in the | |
1164 | // dma_store_start, it will be picked up when the corresponding | |
1165 | // dma_store is issued later | |
1166 | ii->addCinv(cbit); | |
1167 | } | |
1168 | else | |
1169 | { | |
1170 | ii = mab_.findDmaStoreEntry(cmd, DMA_EVICT_INV); | |
1171 | if (ii != mab_.end()) | |
1172 | { | |
1173 | // find one, this evict_inv is for a dma_store entry, mask the | |
1174 | // corresponding bit in inv_vec. | |
1175 | ii->addCinv(cbit); | |
1176 | MSYNC_DEBUG(4, "EVICT_INC DMA_STORE %s", ii->toString().c_str()); | |
1177 | } | |
1178 | else | |
1179 | { | |
1180 | // normal MEM_EVICT_INV | |
1181 | list<MemoryAccessEntry>::iterator srclink; | |
1182 | list<MemoryAccessEntry>::iterator newlink; | |
1183 | MemoryAccessEntry* mae; | |
1184 | srclink = mab_.findEvictInvSrc(cmd); | |
1185 | if (srclink == mab_.end()) | |
1186 | { | |
1187 | MS_ERROR("EvictInv failed to find matched Evict. tid=%d PA=%llx", cmd.getThrdId(), cmd.getAddr()); | |
1188 | return; | |
1189 | } | |
1190 | srclink->addCinv(cbit); | |
1191 | ||
1192 | mae = new MemoryAccessEntry(cmd); | |
1193 | mae->setExecuted(true); // required for retirement from MAB when it reaches to head | |
1194 | //assert (srclink->isLinkValid()); // initially, an Evict will have a link to itself | |
1195 | mae->setLink(srclink->getLink()); // set link chain for debugging purpose | |
1196 | srclink->setGlobal(mae->getId()); // set dinv | |
1197 | newlink = mab_.pushBack(*mae); | |
1198 | srclink->setLink(newlink); | |
1199 | } | |
1200 | } | |
1201 | } | |
1202 | } | |
1203 | ||
1204 | /****************************************************************************** | |
1205 | To allow DMA accesses from I/O. It occurres when a SIU Store (from ENET | |
1206 | or PCI) is seen in the L2 or when Bench back-door stores to memory. | |
1207 | ******************************************************************************/ | |
1208 | void | |
1209 | MemorySync::handleStoreSlam(StoreIssueCmd& cmd) | |
1210 | { | |
1211 | if (addrTrans) | |
1212 | { | |
1213 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1214 | } | |
1215 | uint64_t paddr = cmd.getAddr(); | |
1216 | if (cmd.getSizeV() != 0xff) { | |
1217 | // need to read in data for merging | |
1218 | uint64_t srcData = rif_.readMemory(0, 0, (paddr & ADDR_MASK), 8); | |
1219 | uint64_t tarData = align8byte(cmd.getData(), cmd.getSize(), (paddr & ~ADDR_MASK)); | |
1220 | uint64_t newData = merge(srcData, tarData, cmd.getSizeV()); | |
1221 | rif_.slamMemory(0, 0, (paddr & ADDR_MASK), newData, 8); | |
1222 | } | |
1223 | else { | |
1224 | if ((paddr & 0x7ULL) != 0x0) | |
1225 | { | |
1226 | MS_ERROR("Address 0x%llx is not 8-byte aligned", paddr); | |
1227 | return; | |
1228 | } | |
1229 | rif_.slamMemory(0, 0, paddr, cmd.getData(), 8); | |
1230 | } | |
1231 | } | |
1232 | ||
1233 | //============================================================================= | |
1234 | // signal the beginning of a dma_store operation, there is just one | |
1235 | // dma_store_start, regardless of the dma_store is 8 bytes or 64 bytes. | |
1236 | // For a 64 bytes dma_store, we will see 8 dma_store commands. | |
1237 | //============================================================================= | |
1238 | void | |
1239 | MemorySync::handleDmaStoreStart(LoadStoreCmd& cmd) | |
1240 | { | |
1241 | if (addrTrans) | |
1242 | { | |
1243 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1244 | } | |
1245 | int totalSize = cmd.getTsize(); | |
1246 | uint64_t addr = cmd.getAddr(); | |
1247 | for (int i = 0; i < totalSize/8; i++) | |
1248 | { | |
1249 | // create one DMA_STORE_START, in MAB, for every 8-byte DMA_STORE | |
1250 | cmd.setAddr(addr); | |
1251 | MemoryAccessEntry* mae = new MemoryAccessEntry(cmd); | |
1252 | mae->setEntryType(MEM_DMA_STORE_START); | |
1253 | mab_.pushBack(*mae); | |
1254 | addr += 8; | |
1255 | } | |
1256 | } | |
1257 | ||
1258 | /****************************************************************************** | |
1259 | Similar to MEM_SLAM, but allow inv_vec to handle L1 conflict. | |
1260 | ******************************************************************************/ | |
1261 | void | |
1262 | MemorySync::handleDmaStore(LoadStoreCmd& cmd) | |
1263 | { | |
1264 | if (addrTrans) | |
1265 | { | |
1266 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1267 | } | |
1268 | // treat it as a store_commit with evict & evict_inv | |
1269 | // we will have one evict to fill in the inv_vec, then one or more | |
1270 | // evict_inv to mask each able bit in inv_vec. | |
1271 | // 2/14/06, all DMA_STORE must go through MAB | |
1272 | list<MemoryAccessEntry>::iterator mlink; | |
1273 | uint64_t mdata; | |
1274 | uint64_t sdata; | |
1275 | uint8_t sszv; | |
1276 | int tid = cmd.getThrdId(); | |
1277 | StoreCommitCmd commit(cmd.getThrdId(), 0, 0, cmd.getAddr(), cmd.getSizeV(), true, (cmd.getSizeV()==0?false:true), cmd.getCycle()); | |
1278 | // each DMA_STORE operation can be either 8 bytes or 64 bytes, but | |
1279 | // each DMA_STORE command only represent 8 bytes, we need the total | |
1280 | // size informaiton to handle the corresponding EVICT and EVICT_INV | |
1281 | // commands, as there will be only one EVICT command for each | |
1282 | // DMA_STORE operation (which can have 1 or 8 DMA_STORE commands) | |
1283 | commit.setTsize(cmd.getTsize()); | |
1284 | commit.setData(cmd.getData()); | |
1285 | MemoryAccessEntry entry(commit); | |
1286 | sdata = commit.getData(); | |
1287 | sszv = commit.getSizeV(); | |
1288 | entry.setMerged(false); | |
1289 | if ((commit.getSize() == 8) && (sszv == 0xff)) | |
1290 | { | |
1291 | // no merge needed | |
1292 | mdata = sdata; | |
1293 | } | |
1294 | else | |
1295 | { | |
1296 | // sszv = 0 still need to get correct data in the entry | |
1297 | mdata = getL2Data(mab_.end(), commit); | |
1298 | mdata = merge(mdata, sdata, sszv); | |
1299 | // keep merge-related data in case we need to repeat it (with a | |
1300 | // different mdata) | |
1301 | entry.setMerged(true); | |
1302 | entry.setOrigData(sdata); | |
1303 | entry.setOrigSizeV(sszv); | |
1304 | } | |
1305 | entry.setData(mdata); | |
1306 | list<MemoryAccessEntry>::iterator ii = mab_.findDmaStoreStart(cmd, DMA_STORE); | |
1307 | if (ii != mab_.end()) | |
1308 | { | |
1309 | // there is a dma_store_start associated with this dma_store, use the | |
1310 | // inv and cinv there. mark the dma_store_start as executed so that it | |
1311 | // can be removed from MAB | |
1312 | ii->setExecuted(true); | |
1313 | entry.setInv(ii->getInv()); | |
1314 | entry.setCinv(ii->getCinv()); | |
1315 | } | |
1316 | if (cmd.getInv() == 0) | |
1317 | { | |
1318 | // there is no local evict associated with the dma_store, so set inv | |
1319 | // setting as done, also set this entry as executed, and let inv/cinv | |
1320 | // decide when this entry's data can be committed to memory. | |
1321 | entry.setInvSet(true); | |
1322 | entry.setExecuted(true); | |
1323 | } | |
1324 | entry.setItype(ITYPE_STORE); | |
1325 | entry.setIseq(iSeq(tid)); | |
1326 | entry.setDmaStore(true); | |
1327 | mlink = mab_.pushBack(entry); | |
1328 | mlink->setLink(mlink); // point to itself | |
1329 | ||
1330 | // buffer for writing data back to memory | |
1331 | vector<MemoryAccessEntry> wdbuf; | |
1332 | mab_.popFront(&wdbuf, tsoChecker_); | |
1333 | /* write STORE_COMMIT data to memory */ | |
1334 | bufWriteMemory(wdbuf); | |
1335 | } | |
1336 | ||
1337 | /************************************************************************************ | |
1338 | Reset the state back to LS_ACK. This implementation assumes only the | |
1339 | cacheable request can have FetchFill where the state is set to LS_ACK. | |
1340 | In general, this is inaccurate since LS_ACK in Fetch means instruction | |
1341 | is received. This may be ok if only cacheable instruction can have | |
1342 | self-modifying-code. | |
1343 | ||
1344 | In general, a fetch has | |
1345 | - FetchIssue: instruction fetch issue | |
1346 | - FetchData: instruction access gets data from memory system | |
1347 | - FetchFill: line fill to L1 | |
1348 | - FetchAck: instruction data come back to fetch buffer | |
1349 | ||
1350 | So far, only FetchData and FetchFill is implemented. | |
1351 | *************************************************************************************/ | |
1352 | ||
1353 | void | |
1354 | MemorySync::handleFetchIssue(LoadStoreCmd& cmd) | |
1355 | { | |
1356 | if (addrTrans) | |
1357 | { | |
1358 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1359 | } | |
1360 | list<LoadStoreEntry>::iterator ii; | |
1361 | LoadStoreEntry* entry; | |
1362 | uint32_t tid = cmd.getThrdId(); | |
1363 | ||
1364 | entry = new LoadStoreEntry(cmd); | |
1365 | ii = ifb_[tid].pushBack(*entry); | |
1366 | MSYNC_DEBUG(2, "Last Enter to IFB %s", ii->toString().c_str()); | |
1367 | } | |
1368 | ||
1369 | void | |
1370 | MemorySync::handleFetchData(LoadStoreCmd& cmd) | |
1371 | { | |
1372 | if (addrTrans) | |
1373 | { | |
1374 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1375 | } | |
1376 | list<MemoryAccessEntry>::iterator mlink; | |
1377 | list<LoadStoreEntry>::iterator ife; | |
1378 | vector<MemoryAccessEntry> wdbuf; | |
1379 | MemoryAccessEntry* mae; | |
1380 | uint32_t tid = cmd.getThrdId(); | |
1381 | enum DATA_SRC dsrc = cmd.getDsrc(); | |
1382 | LoadStoreBuffer& ifb = ifb_[tid]; | |
1383 | ||
1384 | // ife = ifb.findMatchO2Y (cmd.getId(), cmd.getAddr(), cmd.getSizeV()); | |
1385 | ife = ifb.findMatchO2Y (cmd.getAddr(), LS_NEW, BLK_ADDR_MASK); | |
1386 | if (ife == ifb.end()) | |
1387 | { | |
1388 | MS_ERROR("Matched FetchIssue expected. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1389 | return; | |
1390 | } | |
1391 | ||
1392 | mae = new MemoryAccessEntry (cmd); | |
1393 | mae->setItype(ITYPE_FETCH); | |
1394 | ||
1395 | /* Reset the state back to LS_ACK. This implementation assumes only the | |
1396 | cacheable request can have FetchFill where the state is set to LS_ACK. | |
1397 | In general, this is inaccurate since LS_ACK in Fetch means instruction | |
1398 | is received. This may be ok if only cacheable instruction can have | |
1399 | self-modifying-code. | |
1400 | ||
1401 | In general, a fetch has | |
1402 | - FetchIssue: instruction fetch issue | |
1403 | - FetchData: instruction access gets data from memory system | |
1404 | - FetchFill: line fill to L1 | |
1405 | - FetchAck: instruction data come back to fetch buffer | |
1406 | ||
1407 | So far, only FetchIssue, FetchData and FetchFill is implemented. | |
1408 | */ | |
1409 | ||
1410 | switch (dsrc) { | |
1411 | case DSRC_L1: | |
1412 | mae->setData(getL1Instr(cmd)); | |
1413 | break; | |
1414 | case DSRC_L2_MEMORY: | |
1415 | MSYNC_DEBUG(2, "mab_=%s", mab_.toString().c_str()); | |
1416 | mae->setData(getL2Data(mab_.end(), cmd)); | |
1417 | break; | |
1418 | case DSRC_IO: // access instruction directly from Riesling memory | |
1419 | mae->setData(rif_.readMemory(cmd.getCoreId(), cmd.getThrdId(), cmd.getAddr() & ADDR_MASK, 8)); | |
1420 | break; | |
1421 | default: | |
1422 | MS_ERROR("Wrong dsrc value %d", dsrc); | |
1423 | return; | |
1424 | } | |
1425 | mlink = mab_.pushBack(*mae); | |
1426 | ife->setLink(mlink); | |
1427 | ife->setState(LS_RDATA); | |
1428 | } | |
1429 | ||
1430 | /** | |
1431 | LoadFill must bring in the whole L1 cache line, TPS 4/22/04 | |
1432 | */ | |
1433 | void | |
1434 | MemorySync::handleFetchFill (LoadStoreCmd& cmd) | |
1435 | { | |
1436 | if (addrTrans) | |
1437 | { | |
1438 | cmd.setAddr(addrTrans(cmd.getAddr())); | |
1439 | } | |
1440 | list<MemoryAccessEntry>::iterator mlink, newlink; | |
1441 | list<LoadStoreEntry>::iterator ife; | |
1442 | MemoryAccessEntry entry (cmd); | |
1443 | uint32_t tid = cmd.getThrdId(); | |
1444 | LoadStoreBuffer& ifb = ifb_[tid]; | |
1445 | ||
1446 | /*********************************************************************************** | |
1447 | * The design assumes FetchFill of the same 8-byte block should come back in order. | |
1448 | * In addition, only fetches whose data source is L2/Memory and is cacheable to L1 | |
1449 | * can have FetchFill. The following do-while statement searches for such entry in | |
1450 | * the fetch buffer. Entries that fetch the same 8-byte block but do not need | |
1451 | * FetchFill are assigned LS_ACK state. This is to indicate that those entries are | |
1452 | * considered done with fetched instruction backed to the fetch buffer already. | |
1453 | * (Recall that FetchAck is not implemented so far and it maybe never needed.) | |
1454 | * Note that after an entry is assigned LS_ACK, the search in the next iteration | |
1455 | * will skip it. | |
1456 | ***********************************************************************************/ | |
1457 | do { | |
1458 | MSYNC_DEBUG(2, "ifb=%s", ifb.toString().c_str()); | |
1459 | ife = ifb.findMatchO2Y(cmd.getAddr(), LS_RDATA, BLK_ADDR_MASK); | |
1460 | if (ife == ifb.end()) | |
1461 | { | |
1462 | MS_ERROR("FetchFill failed to find matched entry in FetchBuffer. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1463 | return; | |
1464 | } | |
1465 | if (!ife->isLinkValid()) | |
1466 | { | |
1467 | MS_ERROR("FetchFill's matched FetchBuffer entry does not have FetchData. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1468 | return; | |
1469 | } | |
1470 | ife->setState(LS_ACK); // this could be the one found, or one does not need FetchFill | |
1471 | } while (!(ife->getLink()->getDsrc() == DSRC_L2_MEMORY && ife->getLink()->isCacheL1())); | |
1472 | ||
1473 | /* The following search in FetchBuffer is actually not needed. This provides | |
1474 | additional layer of check to see if its link is the same as that found | |
1475 | from the MAB search. When the protocol stablizes, can determine if to | |
1476 | remove this segment of code */ | |
1477 | ||
1478 | mlink = mab_.findFetchFillSrc(cmd); | |
1479 | MSYNC_DEBUG(2, "ife=%s", (ife->toString()).c_str()); | |
1480 | MSYNC_DEBUG(2, "mlink=%s", (mlink->toString()).c_str()); | |
1481 | if (mlink == mab_.end()) | |
1482 | { | |
1483 | MS_ERROR("FetchFill failed to find matched FetchData. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1484 | return; | |
1485 | } | |
1486 | if (ife->getLink() != mlink) | |
1487 | { | |
1488 | MS_ERROR("FetchFill's matched FetchBuffer entry's FetchData mismatch. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1489 | return; | |
1490 | } | |
1491 | ||
1492 | entry.setData(mlink->getData()); | |
1493 | entry.setLink(mlink); | |
1494 | entry.setExecuted(true);// set for being retired from MAB when it reaches to head | |
1495 | newlink = mab_.pushBack(entry); | |
1496 | mlink->setLink(newlink); | |
1497 | ||
1498 | /* Also get the other part of L1$ line data, TPS 5/4/04 */ | |
1499 | MemoryAccessEntry* e; | |
1500 | uint64_t line_addr_mask = ~((uint64_t) (ICACHE_LINE_SIZE - 1)); | |
1501 | uint64_t addr = cmd.getAddr() & line_addr_mask; | |
1502 | uint64_t maeAddr = mlink->getAddr(); | |
1503 | ||
1504 | for (int i = 0; i < ICACHE_LINE_SIZE/8; i++, addr += 8) { | |
1505 | if ((maeAddr & ADDR_MASK) == (addr & ADDR_MASK)) { | |
1506 | continue; | |
1507 | } | |
1508 | ||
1509 | cmd.setAddr(addr); | |
1510 | e = new MemoryAccessEntry (cmd); | |
1511 | e->setData(getL2Data(mlink, cmd)); | |
1512 | e->setExecuted(true); // set true so that when it reach to head, it can be removed | |
1513 | e->setLink(mlink); // set to mlink indicating its data got from here | |
1514 | mab_.pushBack(*e); | |
1515 | } | |
1516 | } | |
1517 | ||
1518 | void | |
1519 | MemorySync::removeRetiredStore (uint64_t addr, uint32_t tid) | |
1520 | { | |
1521 | LoadStoreBuffer& rstb = retStb_[tid]; | |
1522 | list<LoadStoreEntry>::iterator ii; | |
1523 | ||
1524 | /* must be the first entry of the STB */ | |
1525 | ii = rstb.begin(); | |
1526 | ||
1527 | if (rstb.getBufPtr()->size() == 0) | |
1528 | { | |
1529 | MS_ERROR("Attempt to remove entry in empty RetiredBuffer. tid=%d PA=%llx", tid, addr); | |
1530 | return; | |
1531 | } | |
1532 | // if (rstb.getBufPtr()->size() == 0) { | |
1533 | // cerr << "MemModel: WARNING[removeRetiredStore()] rstb.size() == 0" << endl; | |
1534 | // return; | |
1535 | // } | |
1536 | ||
1537 | MSYNC_DEBUG (1, "Remove RetiredBuffer Entry tid=%d addr=%#llx addr_buf=%#llx executed=%d", | |
1538 | tid, addr, ii->getAddr(), (int) ii->isExecuted()); | |
1539 | MSYNC_DEBUG (2, "%s", toString().c_str()); | |
1540 | ||
1541 | if (ii->getAddr() != addr) | |
1542 | { | |
1543 | MS_ERROR("Remove RetiredBuffer head entry mismatches with the address. tid=%d PA=%llx", tid, addr); | |
1544 | return; | |
1545 | } | |
1546 | ||
1547 | /* For CAS instruction, if the store part is not sent then the retired instruction | |
1548 | may not be executed. The following code assumes that when an instruction is to | |
1549 | be removed from this buffer, it must have been executed. | |
1550 | */ | |
1551 | if (!ii->isExecuted()) { | |
1552 | if (ii->getItype() != ITYPE_ATOMIC) | |
1553 | { | |
1554 | MS_ERROR("Found non-executed instr. in RetiredBuffer to be removed. tid=%d PA=%llx", tid, addr); | |
1555 | return; | |
1556 | } | |
1557 | ii->setExecuted(true); | |
1558 | if (ii->isLinkValid()) { | |
1559 | ii->getLink()->setExecuted(true); | |
1560 | } | |
1561 | } | |
1562 | ||
1563 | rstb.popFront(); | |
1564 | } | |
1565 | ||
1566 | ||
1567 | /************************************************************************************ | |
1568 | * preMemoryAccess() | |
1569 | * - takes the data to be written by the reference model, | |
1570 | * - checks that with the one in Memory Sync Model, and | |
1571 | * - sets flag so that the reference model won't update its memory. | |
1572 | * | |
1573 | * The MemoryTransaction is used to store the data. The data is in | |
1574 | * 8-byte chunks, and each is arranged in big endian fashion. If the | |
1575 | * size is less than 8, only data[0] is used and the data is shifted | |
1576 | * to the right (toward least significant byte). | |
1577 | ************************************************************************************/ | |
1578 | ||
1579 | void MemorySync::pre_memory_access( void* msync_obj, MemoryTransaction &xact ) | |
1580 | { | |
1581 | MemorySync* self = (MemorySync*)msync_obj; | |
1582 | if (self) | |
1583 | self->preMemoryAccess(xact); | |
1584 | } | |
1585 | ||
1586 | void MemorySync::preMemoryAccess ( MemoryTransaction& xact) | |
1587 | { | |
1588 | ||
1589 | if (MemorySyncMessage::skipCallback == 1) { | |
1590 | //cerr << "ERROR: MSYNC-SKIP: T" << dec << rif_.getTid(xact) << " skip preMemoryAccess()\n"; | |
1591 | return; | |
1592 | } | |
1593 | ||
1594 | list<LoadStoreEntry>::iterator ii; | |
1595 | uint64_t mask; | |
1596 | uint64_t rdata[8]; | |
1597 | uint64_t rpa = xact.paddr(); | |
1598 | if (addrTrans) | |
1599 | { | |
1600 | rpa = addrTrans(rpa); | |
1601 | } | |
1602 | uint64_t rva = xact.vaddr(); | |
1603 | uint8_t rsize = xact.size(); | |
1604 | int nrdata = 0; | |
1605 | ||
1606 | int tid; | |
1607 | int i,j; | |
1608 | ||
1609 | if ((rpa & 0xff00000000ULL) == 0x9000000000ULL) { | |
1610 | // Bench is not sending any ldst_sync messages if PA[39:32]=0x90. | |
1611 | // This is true for ST, STXA and LD, LDXA | |
1612 | return; | |
1613 | } | |
1614 | ||
1615 | tid = xact.getStrand(); | |
1616 | ||
1617 | if (tid >= MAX_STRANDS) { // indicates the xact is not normal instruction access | |
1618 | return; | |
1619 | } | |
1620 | ||
1621 | if (xact.referenceType() == MemoryTransaction::INSTR) { | |
1622 | return; | |
1623 | } | |
1624 | ||
1625 | // if (xact.noOperationXact()) { | |
1626 | // return; | |
1627 | // } | |
1628 | /********************************************************************************** | |
1629 | * In Riesling, atomic is performed in two transactions, one read and one write. | |
1630 | * This following checks return normal load as well as the load part of an atomic | |
1631 | **********************************************************************************/ | |
1632 | // if (xact.readXact()) { | |
1633 | // return; | |
1634 | // } | |
1635 | ||
1636 | if (!xact.writeXact()) { | |
1637 | return; | |
1638 | } | |
1639 | ||
1640 | if (xact.internalXact() || xact.tablewalk()) { | |
1641 | return; | |
1642 | } | |
1643 | ||
1644 | /* store or atomic instructions */ | |
1645 | LoadStoreBuffer& stb = stb_[tid]; | |
1646 | ||
1647 | MSYNC_DEBUG(1, "STEP (preMemAcc) tid=%d va=%#llx pa=%#llx size=%d atomic=%d", tid, rva, rpa, rsize, (int) xact.atomicXact()); | |
1648 | ||
1649 | if (rsize > 8) { | |
1650 | uint8_t buf[64]; | |
1651 | if (rsize != 64 && rsize != 16) { | |
1652 | MS_ERROR("Unsupported store data size %d", xact.size()); | |
1653 | return; | |
1654 | } | |
1655 | for (i = 0; i < rsize/8; i++) { | |
1656 | rdata[i] = xact.getData(i); | |
1657 | nrdata = 8; | |
1658 | } | |
1659 | } else { | |
1660 | rdata[0] = xact.getData(0); | |
1661 | nrdata = 1; | |
1662 | } | |
1663 | ||
1664 | bool atomic = xact.atomicXact(); | |
1665 | uint64_t addr; | |
1666 | LoadStoreBuffer& rstb = retStb_[tid]; | |
1667 | LoadStoreBuffer& rmostb = rmoStb_[tid]; | |
1668 | /********************************************************************************** | |
1669 | * The assumption made is only atomic and rmo stores can have StoreAck issued | |
1670 | * before SSTEP. Therefore, for these two types of instructions, search matched | |
1671 | * StoreIssue must begin with either retStb_ (atomic) or rm_stb_ (rmo). | |
1672 | **********************************************************************************/ | |
1673 | LoadStoreBuffer& ret_or_rmo_stb = (atomic) ? rstb : rmostb; | |
1674 | ||
1675 | MSYNC_DEBUG(2, "rstb=%s", rstb.toString().c_str()); | |
1676 | MSYNC_DEBUG(2, "stb=%s", stb.toString().c_str()); | |
1677 | MSYNC_DEBUG(2, "rmostb=%s", rmostb.toString().c_str()); | |
1678 | ||
1679 | for (i = 0; i < nrdata; i++) { | |
1680 | addr = rpa + (i << 3); | |
1681 | ||
1682 | ii = ret_or_rmo_stb.find1stNonExe(); | |
1683 | if (ii == ret_or_rmo_stb.end()) { | |
1684 | ii = stb.find1stNonExe(); | |
1685 | //MS_ASSERT(ii != stb.end(), "STEP failed to find match StoreIssue in STB. tid=%d PA=%llx", tid, addr); | |
1686 | if (ii == stb.end()) { | |
1687 | if (atomic) { | |
1688 | MS_ERROR("STEP (store part of an atomic instr) failed to find match StoreIssue in STB. tid=%d PA=%llx", tid, addr); | |
1689 | return; | |
1690 | } | |
1691 | else { | |
1692 | MS_ERROR("STEP failed to find match StoreIssue in STB. tid=%d PA=%llx", tid, addr); | |
1693 | return; | |
1694 | } | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | /* Note that RTL does not check store address and data, MemorySync performs | |
1699 | this additional check for completion */ | |
1700 | if ((addr & ADDR_MASK) != (ii->getAddr() & ADDR_MASK)) { | |
1701 | MS_ERROR (" Store Address Mismatch (bits 2-0 ignored) cid=%d tid=%d va=%#llx pa-ref=%#llx pa-rtl=%#llx", | |
1702 | tid/NSTRANDS_PER_CORE, tid, rva + (i << 3), addr, ii->getAddr()); | |
1703 | return; | |
1704 | } | |
1705 | if ((addr & ADDR_MASK) != (ii->getAddr() & ADDR_MASK)) | |
1706 | { | |
1707 | MS_ERROR("STEP's address mismatches with 1st non-executed StoreIssue entry. tid=%d PA=%llx", tid, addr); | |
1708 | return; | |
1709 | } | |
1710 | ||
1711 | // #ifdef N2MODEL | |
1712 | // if (rif_.isPartialStore(xact.asi())) { | |
1713 | // rdata[i] = align8byte (rdata[i], rsize, addr & ~ADDR_MASK); | |
1714 | // } else { | |
1715 | // rdata[i] = align8byte (rdata[i], ii->getSize(), ii->getAddr() & ~ADDR_MASK); | |
1716 | // } | |
1717 | // #else | |
1718 | // rdata[i] = align8byte (rdata[i], ii->getSize(), ii->getAddr() & ~ADDR_MASK); | |
1719 | // #endif | |
1720 | ||
1721 | /* It should be fine to use xact info to aling data */ | |
1722 | rdata[i] = align8byte (rdata[i], rsize, addr & ~ADDR_MASK); | |
1723 | /* size_vector exists only in MemorySync model */ | |
1724 | mask = byteMask(ii->getVbyte()); | |
1725 | ||
1726 | MSYNC_DEBUG(1, "STEP (preMemAcc) cid=%d tid=%d iseq=%lld va=%#llx pa=%#llx data-ref=%#llx data-rtl=%#llx mask=%#llx", | |
1727 | tid/NSTRANDS_PER_CORE, tid, ii->getIseq(), rva + (i << 3), addr, rdata[i], | |
1728 | ii->getData(), mask); | |
1729 | ||
1730 | /********************************************************************************** | |
1731 | * Note that in N1 swerver-memory.cc environment, store data is not always known | |
1732 | * before the SSTEP command. (Data is known @PCX_L2 stage that can happen after | |
1733 | * SSTEP). In this case, the swerver-memory.cc saves the data obtained at SSTEP | |
1734 | * to thrd_write_buf, i.e., STB. So, a load with STB Bypass can still get the | |
1735 | * correct data. | |
1736 | * | |
1737 | * In order to make this memory sync model work in that environment, the data in | |
1738 | * the STB is also updated, the state then is changed to LS_RDATA. | |
1739 | **********************************************************************************/ | |
1740 | if (ii->isDataValid()) { // indicates data is set, in N2, StoreIssue should set state to LS_TDATA | |
1741 | if ((rdata[i] & mask) != (ii->getData() & mask)) { | |
1742 | MS_ERROR (" (Store Data) cid=%d tid=%d va=%#llx pa=%#llx data-ref=%#llx data-rtl=%#llx mask=%#llx", | |
1743 | tid/NSTRANDS_PER_CORE, tid, rva + (i << 3), addr, | |
1744 | rdata[i], ii->getData(), mask); | |
1745 | return; | |
1746 | } | |
1747 | } else { // in N2, this should not happen | |
1748 | ii->setData(rdata[i]); // ii points to the store buffer entry | |
1749 | ii->setState(LS_RDATA); // set data from Reference Model | |
1750 | /* Does the following statements needed for N1, definitely not needed for N2? TPS, 6/16/04 */ | |
1751 | // if (ii->isLink2Valid()) { | |
1752 | // ii->getLink2()->setData(data); | |
1753 | // } | |
1754 | } | |
1755 | ||
1756 | ii->setExecuted(true); | |
1757 | if (ii->isLinkValid()) { | |
1758 | (ii->getLink())->setExecuted(true); | |
1759 | } | |
1760 | ||
1761 | /* Manage the rmoStb_ */ | |
1762 | if (ii->isRMOstore() && ii->isLinkValid()) { | |
1763 | rmostb.erase(ii); | |
1764 | } | |
1765 | ||
1766 | /* Two conditions must be satisfied before a store buffer entry is removed: | |
1767 | 1). must be executed by the reference model, 2). must be acked | |
1768 | Except for atomic instruction, all stores will have STEP before ACK. | |
1769 | If this happens, perform popFront() here. Others will perform popFront() | |
1770 | when they receive ACK. | |
1771 | */ | |
1772 | } | |
1773 | xact.access(MemoryTransaction::WRITE | MemoryTransaction::NOP); | |
1774 | ||
1775 | /* in order not to alter the memory access, no mae_.popFront() and write memory here. */ | |
1776 | /* Don't know why the above statement is here? Add anyway in order to speed up the mab | |
1777 | retirement so that the self-modifying code will not cause problem - 6/3/04 */ | |
1778 | ||
1779 | vector<MemoryAccessEntry> wdbuf; // buffer for writing data back to memory | |
1780 | ||
1781 | mab_.popFront(&wdbuf, tsoChecker_); | |
1782 | /* write STORE_COMMIT data to memory */ | |
1783 | bufWriteMemory(wdbuf); | |
1784 | ||
1785 | return; | |
1786 | } | |
1787 | ||
1788 | /************************************************************************************ | |
1789 | * postMemoryAccess() | |
1790 | * - takes the data stored in the memory sync model, and | |
1791 | * - puts the data back to the MemoryTrasaction so that the reference | |
1792 | * model actually read data provided by the memory sync model | |
1793 | * | |
1794 | * The MemoryTransaction is used to store the data. The data is in | |
1795 | * 8-byte chunks, and each is arranged in big endian fashion. If the | |
1796 | * size is less than 8, only data[0] is used and the data is shifted | |
1797 | * to the right (toward least significant byte). | |
1798 | ************************************************************************************/ | |
1799 | ||
1800 | void MemorySync::post_memory_access( void* msync_obj, MemoryTransaction &xact ) | |
1801 | { | |
1802 | MemorySync* self = (MemorySync*)msync_obj; | |
1803 | if (self) | |
1804 | self->postMemoryAccess(xact); | |
1805 | } | |
1806 | ||
1807 | void MemorySync::postMemoryAccess (MemoryTransaction& xact) | |
1808 | { | |
1809 | if (MemorySyncMessage::skipCallback == 1) { | |
1810 | //cerr << "ERROR: MSYNC-SKIP: T" << dec << rif_.getTid(xact) << " skip postMemoryAccess()\n"; | |
1811 | return; | |
1812 | } | |
1813 | ||
1814 | vector<MemoryAccessEntry> wdbuf; // buffer for writing data back to memory | |
1815 | list<MemoryAccessEntry>::iterator mii; | |
1816 | list<LoadStoreEntry>::iterator ii; | |
1817 | uint64_t mask; | |
1818 | uint64_t data; | |
1819 | uint64_t rpa = xact.paddr(); | |
1820 | if (addrTrans) | |
1821 | { | |
1822 | rpa = addrTrans(rpa); | |
1823 | } | |
1824 | uint64_t rva = xact.vaddr(); | |
1825 | uint8_t rsize = xact.size(); | |
1826 | ||
1827 | int tid; | |
1828 | int iter; | |
1829 | int i,j; | |
1830 | ||
1831 | if ((rpa & 0xff00000000ULL) == 0x9000000000ULL) { | |
1832 | // Bench is not sending any ldst_sync messages if PA[39:32]=0x90. | |
1833 | // This is true for ST, STXA and LD, LDXA | |
1834 | return; | |
1835 | } | |
1836 | ||
1837 | tid = xact.getStrand(); | |
1838 | ||
1839 | if (tid >= MAX_STRANDS) { // indicates the xact is not normal instruction access | |
1840 | return; | |
1841 | } | |
1842 | ||
1843 | if (xact.referenceType() == MemoryTransaction::INSTR) { | |
1844 | MSYNC_DEBUG(1, "I-Fetch tid=%d va=%#llx pa=%#llx iword=%#010llx size=%d", | |
1845 | tid, rva, rpa, xact.getData(), rsize); | |
1846 | ||
1847 | /* As of 5/18/04, only N1 sends Fetch related command */ | |
1848 | #ifdef N1MODEL | |
1849 | if (rva == 0 && rpa == 0) return; // remove glitch during RESET_INTERRUPT | |
1850 | ||
1851 | LoadStoreCmd cmd; | |
1852 | list<LoadStoreEntry>::iterator ife, ii; | |
1853 | list<MemoryAccessEntry>::iterator mae; | |
1854 | LoadStoreBuffer& ifb = ifb_[tid]; | |
1855 | ||
1856 | cmd.setAddr(rpa); | |
1857 | ||
1858 | // ife = ifb.findMatchY2O(rpa, cmd.sz2szv(rsize, rpa & ~ADDR_MASK)); | |
1859 | ife = ifb.find1stNonExeFetch (rpa); | |
1860 | // MS_ASSERT(ife != ifb.end(), "I-Fetch cannot find matched entry in FetchBuffer", cmd.getAddr()); | |
1861 | if (ife == ifb.end()) { | |
1862 | if (MSYNC_DEBUGLevel != 88) { | |
1863 | MemorySyncMessage::warning("I-Fetch cannot find matched entry in FetchBuffer tid=%d va=%#llx pa=%#llx iword=%#010llx", tid, rva, rpa, xact.getData()); | |
1864 | } | |
1865 | else { | |
1866 | MemorySyncMessage::warning("I-Fetch cannot find matched entry in FetchBuffer tid=%d va=%#llx pa=%#llx iword='%s'", tid, rva, rpa, Hv_InstructionWord::disassemble(xact.getData()).c_str()); | |
1867 | } | |
1868 | return; // use the one in nas-memory | |
1869 | } | |
1870 | // if (ife != (ii = ifb.find1stNonExe())) { | |
1871 | if (ife != (ii = ifb.begin())) { | |
1872 | for (ii; ii != ife;) { | |
1873 | ii->setExecuted(true); | |
1874 | // MS_ASSERT(ii->isLinkValid(), "I-Fetch entry in FetchBuffer does not have FetchData", cmd.getAddr()); | |
1875 | if (ii->isLinkValid()) // some fetch issue may not be real? | |
1876 | ii->getLink()->setExecuted(true); | |
1877 | MSYNC_DEBUG(1, "I-Fetch removes non-used Fetch Entries tid=%d pa=%#llx size=%d", | |
1878 | tid, ii->getAddr(), ii->getSize()); | |
1879 | ii++; | |
1880 | ifb.popFront(); | |
1881 | } | |
1882 | } | |
1883 | ||
1884 | //#ifdef N1MODEL // There is cases in N1 that FetchIssue does not have FetchData, timing? | |
1885 | if (!ife->isLinkValid()) return; | |
1886 | //#endif | |
1887 | if (!ife->isLinkValid()) | |
1888 | { | |
1889 | MS_ERROR("I-Fetch entry in FetchBuffer does not have FetchData. tid=%d PA=%llx", tid, cmd.getAddr()); | |
1890 | return; | |
1891 | } | |
1892 | ife->setExecuted(true); | |
1893 | mae = ife->getLink(); | |
1894 | mae->setExecuted(true); | |
1895 | data = align2addr(mae->getData(), 4, (rpa & ~ADDR_MASK)); | |
1896 | if (data != xact.getData() && !xact.noWriteXact()) { | |
1897 | MemorySyncMessage::warning("I-Fetch detects on-the-fly modified code tid=%d va=%#llx pa=%#llx i-ref=%#10llx i-rtl=%#10llx", | |
1898 | tid, rva, rpa, xact.getData(), data); | |
1899 | xact.setData(data); | |
1900 | } | |
1901 | ifb.popFront(); | |
1902 | mab_.popFront(&wdbuf, tsoChecker_); | |
1903 | bufWriteMemory(wdbuf); | |
1904 | #endif // #ifdef N1MODEL | |
1905 | return; | |
1906 | } | |
1907 | ||
1908 | // if (xact.noOperationXact()) { | |
1909 | // return; | |
1910 | // } | |
1911 | ||
1912 | /********************************************************************************** | |
1913 | * In Riesling, atomic is performed in two transactions, one read and one write. | |
1914 | * This following checks return normal store as well as the store part of an atomic | |
1915 | **********************************************************************************/ | |
1916 | // if (xact.writeXact()) { | |
1917 | // return; | |
1918 | // } | |
1919 | ||
1920 | if (!xact.readXact()) { | |
1921 | return; | |
1922 | } | |
1923 | ||
1924 | if (xact.internalXact()) { | |
1925 | return; | |
1926 | } | |
1927 | ||
1928 | if (xact.tablewalk()) { | |
1929 | // accessing TSB, look for matching ST_COMMIT | |
1930 | uint64_t data = mab_.getL2Data(mab_.end(), tid/NSTRANDS_PER_CORE, tid, rpa, true); | |
1931 | xact.setData(0, data); | |
1932 | return; | |
1933 | } | |
1934 | ||
1935 | /********************************************************************************** | |
1936 | * In Riesling, partial store sends load and store transaction. | |
1937 | * The first load should be ignored in this model, the data check in store part | |
1938 | * should check only those indicated by the size_vector. | |
1939 | **********************************************************************************/ | |
1940 | //TODO why do we comment this out? 11/22/06 | |
1941 | //if (rif_.isPartialStore(xact.asi())) { | |
1942 | // return; | |
1943 | //} | |
1944 | ||
1945 | /* read and atomic instructions */ | |
1946 | ||
1947 | LoadStoreBuffer& ldb = ldb_[tid]; | |
1948 | ||
1949 | MSYNC_DEBUG(1, "STEP (postMemAcc) tid=%d va=%#llx pa=%#llx size=%d atomic=%d", tid, rva, rpa, rsize, (int) xact.atomicXact()); | |
1950 | ||
1951 | int sz; | |
1952 | if (rsize > 8) { | |
1953 | if (rsize != 64 && rsize != 16) { | |
1954 | MS_ERROR("Unsupported load data size %d", xact.size()); | |
1955 | return; | |
1956 | } | |
1957 | iter = rsize / 8; | |
1958 | sz = 8; // each one equals 8 bytes | |
1959 | } else { | |
1960 | iter = 1; | |
1961 | sz = rsize; | |
1962 | } | |
1963 | uint64_t addr; | |
1964 | for (int i = 0; i < iter; i++) { | |
1965 | addr = rpa + (i << 3); | |
1966 | ||
1967 | if ((addr & IO_ADDR_BIT_MASK) == IO_ADDR_BIT_MASK) { // IO address | |
1968 | ii = ldb.find1stNonExeMatchedAddr(addr); // IO can be out-of-order | |
1969 | } else { | |
1970 | ii = ldb.find1stNonExe(); | |
1971 | } | |
1972 | ||
1973 | MSYNC_DEBUG (4, "ldb=%s", ldb.toString().c_str()); | |
1974 | ||
1975 | if (ii == ldb.end()) | |
1976 | { | |
1977 | MS_ERROR("STEP failed to find LoadIssue entry (possibly DUT took trap & Riesling did not). tid=%d PA=%llx", tid, addr); | |
1978 | return; | |
1979 | } | |
1980 | /* Note that RTL does not check load address, MemorySync performs | |
1981 | this additional check for completion */ | |
1982 | if (addr != ii->getAddr()) { | |
1983 | MS_ERROR (" Load Address Mismatch cid=%d tid=%d va=%#llx pa-ref=%#llx pa-rtl=%#llx", | |
1984 | tid/NSTRANDS_PER_CORE, tid, rva + (i << 3), addr, ii->getAddr()); | |
1985 | return; | |
1986 | } | |
1987 | if (addr != ii->getAddr()) | |
1988 | { | |
1989 | MS_ERROR("STEP's address mismatches with the 1st non-executed Load. tid=%d PA=%llx", tid, addr); | |
1990 | return; | |
1991 | } | |
1992 | if (!ii->isLinkValid()) | |
1993 | { | |
1994 | MS_ERROR("STEP's corresponding LoadData has not yet been issued. tid=%d PA=%llx", tid, addr); | |
1995 | return; | |
1996 | } | |
1997 | if (xact.atomicXact()) { // this assert is to make sure we can use xact.atomic() later | |
1998 | if (ii->getItype() != ITYPE_ATOMIC) | |
1999 | { | |
2000 | MS_ERROR("STEP (postMemAcc) atomic matches non-atomic load entry. tid=%d PA=%llx", tid, addr); | |
2001 | return; | |
2002 | } | |
2003 | } | |
2004 | ||
2005 | // If Load to I/O space (PA[39]=1), Riesling will see the following | |
2006 | // messages | |
2007 | // LDISSUE | |
2008 | // LDDATA | |
2009 | // [CSR_READ] [Optional] | |
2010 | // SSTEP | |
2011 | // | |
2012 | // Riesling will have to queue up the CSR_READ messages from the Bench. | |
2013 | // When Riesling sees an SSTEP for an Load to I/O, it must first | |
2014 | // process the LDDATA normally. Then, if the oldest CSR_READ in the | |
2015 | // queue has a matching PA, use it to override the LDDATA value and | |
2016 | // pop it from the queue. If oldest CSR_READ does not have a matching | |
2017 | // PA, then use LDDATA value and do not pop from queue. | |
2018 | ||
2019 | // xact contains data from either memory or I/O address follow-me, we | |
2020 | // cannot tell which here. | |
2021 | ||
2022 | //TODO so we can be in trouble here, e.g. | |
2023 | // LDISSUE | |
2024 | // LDDATA | |
2025 | // STCOMMIT | |
2026 | // SSTEP | |
2027 | // in this case the data from memory (there is no csr_read follow-me) can | |
2028 | // be different to the value in LDDATA, yet we cannot tell whether the | |
2029 | // value is from follow-me or not. | |
2030 | // xact.getData() will conduct data re-arrangement according to the | |
2031 | // related instruction, we don't want that here, we only want the original | |
2032 | // raw data, so use no-fault version of getData(). | |
2033 | //TODO there is no getDataNF(), is getData() good enough? | |
2034 | uint64_t memData = xact.getData(i); | |
2035 | // if the load entry's data comes from a store_commit entry, make sure | |
2036 | // that entry is indeed executed, it may happen that the entry is voided | |
2037 | // due to error injection | |
2038 | //===> shouldn't we mark it as 'popped' in such case? 2/15/06 | |
2039 | if (ii->getLink()->getDsrc() == DSRC_L2_MEMORY) { | |
2040 | if (!xact.followmeXact()) { | |
2041 | // if the xact data is coming from a follow-me, don't do another | |
2042 | // read here, we may pop out another follow-me (intended for the | |
2043 | // next load) | |
2044 | //ii->getLink()->setData(mab_.getL2Data(ii->getLink(), tid/NSTRANDS_PER_CORE, tid, addr, true)); | |
2045 | ii->getLink()->setData(mab_.getL2Data(ii->getLink(), tid/NSTRANDS_PER_CORE, tid, addr)); | |
2046 | } | |
2047 | } | |
2048 | // now we should have the correct value | |
2049 | data = ii->getLink()->getData(); | |
2050 | uint64_t ldData = data; | |
2051 | if (xact.followmeXact()) { | |
2052 | //TODO if memData is not from csr_read follow-me, this is wrong, | |
2053 | // do we have a way to tell? 10/26/05 | |
2054 | // memData from xact should already be aligned to address/size, | |
2055 | // no need to go through align2addr() again. | |
2056 | data = memData; | |
2057 | } | |
2058 | else { | |
2059 | data = align2addr(data, sz, (addr & ~ADDR_MASK)); | |
2060 | } | |
2061 | ||
2062 | MSYNC_DEBUG(1, "STEP (postMemAcc) cid=%d tid=%d iseq=%lld va=%#llx pa=%#llx data-ref=%#llx (aligned=%#llx) data-rtl=%#llx size=%d ---> use data=%#llx", tid/NSTRANDS_PER_CORE, tid, ii->getIseq(), rva + (i << 3), addr, memData, align2addr(memData, sz, (addr & ~ADDR_MASK)), align2addr(ldData, sz, (addr & ~ADDR_MASK)), sz, data); | |
2063 | ||
2064 | if (!xact.noWriteXact()) { | |
2065 | xact.setData(i, data); | |
2066 | } | |
2067 | else { | |
2068 | // although the noWriteXact() indicates we must use the data in xact, | |
2069 | // the data still has to follow aligning rules. | |
2070 | //---> memData alignment should be handled in xact already, no | |
2071 | // need to do it again. | |
2072 | //xact.setData(i, align2addr(memData, sz, (addr & ~ADDR_MASK))); | |
2073 | xact.setData(i, memData); | |
2074 | } | |
2075 | ii->setExecuted(true); | |
2076 | ii->getLink()->setExecuted(true); | |
2077 | mii = ii->getLink(); | |
2078 | ldb.erase(ii); | |
2079 | } | |
2080 | ||
2081 | /********************************************************************************** | |
2082 | * The following code is to handle CAS instruction | |
2083 | * Two issues related to the CAS: | |
2084 | * 1. The Memory Sync model needs to know the comparison results so that it knows | |
2085 | * which data to compare | |
2086 | * The solution employed is 1) the StoreIssue of a CAS instruction sends r[rd] | |
2087 | * in the <data> field, and 2) if mismatch, reset the StoreCommit size_vector | |
2088 | * to zero. | |
2089 | * | |
2090 | * N1 is different. In ni environment, the atomic STEP is sent before the | |
2091 | * STORE_COMMIT. Hence, for N1, CAS mismatch will reset | |
2092 | * the <size_vector> of the StoreIssue to 0, rather than StoreCommit. However, | |
2093 | * the StoreCommit will then copy this value to its field. | |
2094 | * | |
2095 | * 2. The Riesling does not send a second write memory transaction if the CAS | |
2096 | * comparison results in a false. | |
2097 | * In this case, the the executed_ bit of the StoreIssue and StoreCommit must be | |
2098 | * set in this method. (Normally, the executed_ bit of a StoreIssue or StoreCommit | |
2099 | * is set in preMemoryAccess() when a write memory transaction is issued by the | |
2100 | * Riesling model. | |
2101 | * | |
2102 | * Therefore, in this atomic-load transaction, if the size-vector is 0, it | |
2103 | * indicates no write transaction, and the execute_ bit will be set accordingly. | |
2104 | **********************************************************************************/ | |
2105 | if (xact.atomicXact()) { | |
2106 | list<LoadStoreEntry>::iterator slink; | |
2107 | LoadStoreBuffer& stb = stb_[tid]; | |
2108 | LoadStoreBuffer& rstb = retStb_[tid]; | |
2109 | ||
2110 | /* find the corresponding StoreIssue */ | |
2111 | slink = rstb.find1stNonExe(); // must step in order | |
2112 | if (slink == rstb.end()) { | |
2113 | slink = stb.find1stNonExe(); | |
2114 | if (slink == stb.end()) | |
2115 | { | |
2116 | MS_ERROR("STEP(at) failed to find non-executed StoreIssue entry (possibly DUT took trap & Riesling did not). tid=%d PA=%llx", tid, addr); | |
2117 | return; | |
2118 | } | |
2119 | } | |
2120 | if (rpa != slink->getAddr()) | |
2121 | { | |
2122 | MS_ERROR("STEP(at) mis-matches addr with the 1st non-executed StoreIssue. tid=%d PA=%llx", tid, addr); | |
2123 | return; | |
2124 | } | |
2125 | if (slink->getItype() != ITYPE_ATOMIC) | |
2126 | { | |
2127 | MS_ERROR("STEP(at) found 1st non-executed StoreIssue non-atomic. tid=%d PA=%llx", tid, addr); | |
2128 | return; | |
2129 | } | |
2130 | ||
2131 | /* set the executed_ bit if needed */ | |
2132 | #ifdef N1MODEL | |
2133 | MSYNC_DEBUG(2, "PostMemAccess atomic sizeV=%#x", (int) slink->getSizeV()); | |
2134 | if (slink->getSizeV() == 0) { // CAS comparison is false => no write necessary | |
2135 | slink->setExecuted(true); | |
2136 | if (slink->isLinkValid()) { | |
2137 | slink->getLink()->setExecuted(true); | |
2138 | /* Fill the store part the same data written info as it load part --- TPS 9/10/04 */ | |
2139 | slink->getLink()->setData(mii->getData()); | |
2140 | slink->getLink()->setSizeV(mii->getSizeV()); | |
2141 | } | |
2142 | } | |
2143 | #else | |
2144 | if (!slink->isLinkValid()) | |
2145 | { | |
2146 | MS_ERROR("STEP(at) misses StoreCommit. tid=%d PA=%llx", tid, addr); | |
2147 | return; | |
2148 | } | |
2149 | if (slink->getLink()->getSizeV() == 0) { // CAS comparison is false => no write necessary | |
2150 | slink->setExecuted(true); | |
2151 | slink->getLink()->setExecuted(true); | |
2152 | /* Fill the store part the same data written info as it load part --- TPS 9/10/04 */ | |
2153 | slink->getLink()->setData(mii->getData()); | |
2154 | slink->getLink()->setSizeV(mii->getSizeV()); | |
2155 | } | |
2156 | #endif | |
2157 | MSYNC_DEBUG(2, "slink=%s", (slink->toString()).c_str()); | |
2158 | } | |
2159 | ||
2160 | mab_.popFront(&wdbuf, tsoChecker_); | |
2161 | /* write STORE_COMMIT data to memory */ | |
2162 | bufWriteMemory(wdbuf); | |
2163 | ||
2164 | return; | |
2165 | } | |
2166 | ||
2167 | /* write STORE_COMMIT data to memory */ | |
2168 | void | |
2169 | MemorySync::bufWriteMemory (vector<MemoryAccessEntry>& wdbuf) | |
2170 | { | |
2171 | vector<MemoryAccessEntry>::iterator wi; | |
2172 | int qsize = wdbuf.size(); | |
2173 | int i = 0; | |
2174 | ||
2175 | wi = wdbuf.begin(); | |
2176 | while (i < qsize) { | |
2177 | if (wi->getSizeV() != 0 && // (sizeV == 0) => no need to store | |
2178 | !(wi->getItype() == ITYPE_ATOMIC && !wi->isSwitchData())) { // cas false, no need to store | |
2179 | rif_.writeMemory(wi->getCoreId(), wi->getThrdId(), (wi->getAddr() & ADDR_MASK), wi->getData(), 8); | |
2180 | } | |
2181 | i++; | |
2182 | wi++; | |
2183 | } | |
2184 | wdbuf.clear(); | |
2185 | } | |
2186 | ||
2187 | void | |
2188 | MemorySync::setTestBenchData (uint32_t tid, uint64_t addr, uint64_t data, bool rmo) | |
2189 | { | |
2190 | if (addrTrans) | |
2191 | { | |
2192 | addr = addrTrans(addr); | |
2193 | } | |
2194 | LoadStoreBuffer& stb = stb_[tid]; | |
2195 | LoadStoreBuffer& rmostb = rmoStb_[tid]; | |
2196 | list<LoadStoreEntry>::iterator ii; | |
2197 | ||
2198 | if (rmo) { | |
2199 | ii = rmostb.findNeedTDataO2Y(addr); | |
2200 | if (ii == rmostb.end()) { | |
2201 | ii = stb.findNeedTDataO2Y(addr); | |
2202 | if (ii == rmostb.end()) | |
2203 | { | |
2204 | MS_ERROR("setTestBenchData failed to find RMO entry in (RMO)STB. tid=%d PA=%llx", tid, addr); | |
2205 | return; | |
2206 | } | |
2207 | if (!ii->isRMOstore()) | |
2208 | { | |
2209 | MS_ERROR("setTestBenchData expected match-address entry in STB to be RMO. tid=%d PA=%llx", tid, addr); | |
2210 | return; | |
2211 | } | |
2212 | } | |
2213 | } else { | |
2214 | ii = stb.findNeedTDataO2Y(addr); | |
2215 | if (ii == stb.end()) | |
2216 | { | |
2217 | MS_ERROR("setTestBenchData failed to find entry in (RMO)STB. tid=%d PA=%llx", tid, addr); | |
2218 | return; | |
2219 | } | |
2220 | } | |
2221 | ||
2222 | if (ii->isDataValid()) { | |
2223 | // can only be RDATA, Reference Model data set in STEP | |
2224 | if (ii->getState() != LS_RDATA) | |
2225 | { | |
2226 | MS_ERROR("The entry is not in LS_RDATA state"); | |
2227 | return; | |
2228 | } | |
2229 | uint64_t mask = byteMask(ii->getVbyte()); | |
2230 | if ((data & mask) != (ii->getData() & mask)) { | |
2231 | MS_ERROR (" (Store Data) cid=%d tid=%d pa=%#llx data-ref=%#llx data-rtl=%#llx mask=%#x", | |
2232 | ii->getCoreId(), ii->getThrdId(), ii->getAddr(), | |
2233 | ii->getData(), data, mask); | |
2234 | return; | |
2235 | } | |
2236 | } else { | |
2237 | ii->setData(data); | |
2238 | /* Does the following statements needed for N1, definitely not needed for N2? TPS, 6/16/04 */ | |
2239 | // if (ii->isLink2Valid()) { | |
2240 | // ii->getLink2()->setData(data); | |
2241 | // } | |
2242 | } | |
2243 | ii->setState(LS_TDATA); | |
2244 | } | |
2245 | ||
2246 | /* assume data is aligned to 8 byte chunk, the result ha | |
2247 | 1. data shift to the least significant bytes, also | |
2248 | 2. make the higher byte outside the size range 0, which is | |
2249 | required when this function is used in postMemoryAccess() | |
2250 | and provide data back to Riesling Reference Model since | |
2251 | Riesling zero out all bytes not in the request data size | |
2252 | range. | |
2253 | */ | |
2254 | uint64_t | |
2255 | MemorySync::align2addr (uint64_t data, uint8_t size, uint32_t addr_offset) | |
2256 | { | |
2257 | uint64_t result; | |
2258 | int sft; | |
2259 | ||
2260 | switch (size) { | |
2261 | case 8: | |
2262 | assert ((addr_offset & 0x7ULL) == 0); | |
2263 | result = data; | |
2264 | break; | |
2265 | case 4: | |
2266 | assert ((addr_offset & 0x3ULL) == 0); | |
2267 | sft = 32 - ((addr_offset & 0x4ULL) << 3); | |
2268 | result = (data >> sft) & 0xffffffffULL; | |
2269 | break; | |
2270 | case 2: | |
2271 | assert ((addr_offset & 0x1ULL) == 0); | |
2272 | sft = 48 - ((addr_offset & 0x6ULL) << 3); | |
2273 | result = (data >> sft) & 0xffffULL; | |
2274 | break; | |
2275 | case 1: | |
2276 | sft = 56 - (addr_offset << 3); | |
2277 | result = (data >> sft) & 0xffULL; | |
2278 | break; | |
2279 | default: | |
2280 | assert(0); | |
2281 | result = data; | |
2282 | break; | |
2283 | } | |
2284 | return (result); | |
2285 | } | |
2286 | ||
2287 | /* assume data is at the least significant bytes */ | |
2288 | uint64_t | |
2289 | MemorySync::align8byte (uint64_t data, uint8_t size, uint32_t addr_offset) | |
2290 | { | |
2291 | uint64_t result; | |
2292 | int sft; | |
2293 | ||
2294 | switch (size) { | |
2295 | case 64: | |
2296 | case 16: | |
2297 | case 8: | |
2298 | assert ((addr_offset & 0x7ULL) == 0); | |
2299 | result = data; | |
2300 | break; | |
2301 | case 4: | |
2302 | assert ((addr_offset & 0x3ULL) == 0); | |
2303 | sft = 32 - ((addr_offset & 0x4ULL) << 3); | |
2304 | result = data << sft; | |
2305 | break; | |
2306 | case 2: | |
2307 | assert ((addr_offset & 0x1ULL) == 0); | |
2308 | sft = 48 - ((addr_offset & 0x6ULL) << 3); | |
2309 | result = data << sft; | |
2310 | break; | |
2311 | case 1: | |
2312 | sft = 56 - (addr_offset << 3); | |
2313 | result = data << sft; | |
2314 | break; | |
2315 | case 0: | |
2316 | result = data; | |
2317 | break; | |
2318 | default: | |
2319 | assert(0); | |
2320 | result = data; | |
2321 | break; | |
2322 | } | |
2323 | return (result); | |
2324 | } | |
2325 | ||
2326 | uint64_t | |
2327 | MemorySync::merge (uint64_t todata, uint64_t fromdata, uint8_t mgvec) | |
2328 | { | |
2329 | uint64_t data; | |
2330 | uint64_t byteMask1 = byteMask(~mgvec); | |
2331 | uint64_t byteMask2 = byteMask(mgvec); | |
2332 | data = (todata & byteMask1) | (fromdata & byteMask2); | |
2333 | ||
2334 | MSYNC_DEBUG(4, "merge: todata=0x%llx fdata=0x%llx merge=0x%x result=0x%llx", | |
2335 | todata, fromdata, (int) mgvec, data); | |
2336 | ||
2337 | return (data); | |
2338 | } | |
2339 | ||
2340 | ||
2341 | uint64_t | |
2342 | MemorySync::byteMask(const uint8_t vbyte) { | |
2343 | uint64_t mask = 0ull; | |
2344 | uint8_t bitSelector = 0x80; // 10000000 | |
2345 | ||
2346 | for (int i = 0; i < 8; i++) { | |
2347 | mask = mask << 8; | |
2348 | mask = ((vbyte & bitSelector) == 0) ? mask : mask | 0xffull; | |
2349 | bitSelector >>= 1; | |
2350 | } | |
2351 | // cerr << "byteMask: in=0x" << hex << (int) vbyte << " out=0x" << hex << mask << endl; | |
2352 | return (mask); | |
2353 | } | |
2354 | ||
2355 | uint64_t | |
2356 | MemorySync::getL1Data(LoadStoreCmd& cmd) | |
2357 | { | |
2358 | list<MemoryAccessEntry>::iterator mae; | |
2359 | ||
2360 | mae = mab_.findL1DataEntry(cmd); | |
2361 | if (mae == mab_.end()) { | |
2362 | return (rif_.readMemory(cmd.getCoreId(), cmd.getThrdId(), cmd.getAddr() & ADDR_MASK, 8)); // read aligned 8 bytes | |
2363 | } else { | |
2364 | return (mae->getData()); | |
2365 | } | |
2366 | } | |
2367 | ||
2368 | uint64_t | |
2369 | MemorySync::getL2Data(list<MemoryAccessEntry>::iterator from, LoadStoreCmd& cmd) | |
2370 | { | |
2371 | list<MemoryAccessEntry>::iterator mae; | |
2372 | ||
2373 | mae = mab_.findL2DataEntry(from, cmd); | |
2374 | if (mae == mab_.end()) { | |
2375 | return (rif_.readMemory(cmd.getCoreId(), cmd.getThrdId(), cmd.getAddr() & ADDR_MASK, 8)); // read aligned 8 bytes | |
2376 | } else { | |
2377 | return (mae->getData()); | |
2378 | } | |
2379 | } | |
2380 | ||
2381 | uint64_t | |
2382 | MemorySync::getL1Instr(LoadStoreCmd& cmd) | |
2383 | { | |
2384 | list<MemoryAccessEntry>::iterator mae; | |
2385 | uint32_t tid = cmd.getThrdId(); | |
2386 | ||
2387 | mae = mab_.findL1InstrEntry(cmd); | |
2388 | if (mae == mab_.end()) { | |
2389 | return (rif_.readMemory(cmd.getCoreId(), cmd.getThrdId(), cmd.getAddr() & ADDR_MASK, 8)); // read aligned 8 bytes | |
2390 | } else { | |
2391 | return (mae->getData()); | |
2392 | } | |
2393 | } | |
2394 | ||
2395 | //============================================================================= | |
2396 | //============================================================================= | |
2397 | void | |
2398 | MemorySync::nullMsyncCallback() | |
2399 | { | |
2400 | // turn off msync related callbacks | |
2401 | MemorySyncMessage::skipCallback = 1; | |
2402 | } | |
2403 | ||
2404 | //============================================================================= | |
2405 | //============================================================================= | |
2406 | void | |
2407 | MemorySync::flushMsyncCallback(int tid) | |
2408 | { | |
2409 | MSYNC_DEBUG(1, "flushMsyncCallback( tid=%d )", tid); | |
2410 | ||
2411 | ldb_[tid].empty(); | |
2412 | ||
2413 | //TODO not sure if we should empty the following buffers as well | |
2414 | //stb_[tid].empty(); | |
2415 | //ifb_[tid].empty(); | |
2416 | ||
2417 | mab_.empty(tid); | |
2418 | // we rely on the entry in MAB to determine which entry needs to be removed | |
2419 | // from retStb and rmoStb, so the "mab_.empty" must come first. | |
2420 | retStb_[tid].markPop(); | |
2421 | rmoStb_[tid].markPop(); | |
2422 | } | |
2423 | ||
2424 | //============================================================================= | |
2425 | //============================================================================= | |
2426 | void | |
2427 | MemorySync::handleLoadPop(int tid) | |
2428 | { | |
2429 | MSYNC_DEBUG(1, "handleLoadPop( MEM_LD_POP, tid=%d )", tid); | |
2430 | ||
2431 | list<LoadStoreEntry>::iterator ii = ldb_[tid].queryBack(); | |
2432 | if (ii == ldb_[tid].end()) | |
2433 | { | |
2434 | MS_ERROR("No load entry to be popped"); | |
2435 | return; | |
2436 | } | |
2437 | int count; | |
2438 | switch ((ii)->getItype()) { | |
2439 | case ITYPE_BLOCK_LOAD: | |
2440 | count = 8; | |
2441 | break; | |
2442 | case ITYPE_QUAD_LOAD: | |
2443 | count = 2; | |
2444 | break; | |
2445 | default: | |
2446 | count = 1; | |
2447 | break; | |
2448 | } | |
2449 | ||
2450 | if (count > ldb_[tid].size()) | |
2451 | { | |
2452 | MS_ERROR("Load buffer does not have enough entries to be popped, ldb.size=%d pop=%d", ldb_[tid].size(), count); | |
2453 | return; | |
2454 | } | |
2455 | ||
2456 | ldb_[tid].popBack(count); | |
2457 | mab_.popBack(tid, count); | |
2458 | } | |
2459 | ||
2460 | //============================================================================= | |
2461 | //============================================================================= | |
2462 | void | |
2463 | MemorySync::handleStorePop(int tid) | |
2464 | { | |
2465 | MSYNC_DEBUG(1, "handleStorePop( MEM_ST_POP, tid=%d )", tid); | |
2466 | ||
2467 | list<LoadStoreEntry>::iterator ii = stb_[tid].queryBack(); | |
2468 | if (ii == stb_[tid].end()) | |
2469 | { | |
2470 | MS_ERROR("No store entry to be popped"); | |
2471 | return; | |
2472 | } | |
2473 | int count; | |
2474 | switch ((ii)->getItype()) { | |
2475 | case ITYPE_BLOCK_STORE: | |
2476 | count = 8; | |
2477 | break; | |
2478 | default: | |
2479 | count = 1; | |
2480 | break; | |
2481 | } | |
2482 | ||
2483 | if (count > stb_[tid].size()) | |
2484 | { | |
2485 | MS_ERROR("Store buffer does not have enough entries to be popped, stb.size=%d pop=%d", stb_[tid].size(), count); | |
2486 | return; | |
2487 | } | |
2488 | ||
2489 | stb_[tid].popBack(count); | |
2490 | //TODO do we need to pop store entries in MAB? | |
2491 | } | |
2492 | ||
2493 | /****************************************************************************** | |
2494 | compare a 64-byte data block in memory with the data provided by testbench, | |
2495 | throw error if miscompare. | |
2496 | ******************************************************************************/ | |
2497 | void | |
2498 | MemorySync::handleMemoryCheck(uint64_t paddr, uint64_t* data, int dataSize) | |
2499 | { | |
2500 | if (addrTrans) | |
2501 | { | |
2502 | paddr = addrTrans(paddr); | |
2503 | } | |
2504 | if ((paddr % (dataSize*8)) != 0) | |
2505 | { | |
2506 | MS_ERROR("paddr=%#llx is not %d-byte aligned", paddr, (dataSize*8)); | |
2507 | return; | |
2508 | } | |
2509 | ||
2510 | for (int i = 0; i < dataSize; i++) { | |
2511 | // DMA_STORE always goes through msync, so MEM_CHECK should go through | |
2512 | // msync to find matched entry. | |
2513 | //uint64_t memData = rif_.readMemory(0, 0, (paddr+8*i), 8); | |
2514 | uint64_t memData = mab_.getL2Data(mab_.end(), 0, 0, (paddr+8*i), true); | |
2515 | if (data[i] != memData) | |
2516 | { | |
2517 | MS_ERROR("memory-check mismatch at addr=%#llx, RTL=%#llx, riesling=%#llx", (paddr+8*i), data[i], memData); | |
2518 | return; | |
2519 | } | |
2520 | } | |
2521 | } | |
2522 | ||
2523 | //============================================================================= | |
2524 | //============================================================================= | |
2525 | void | |
2526 | MemorySync::flushAll() | |
2527 | { | |
2528 | // flush out all entries | |
2529 | mab_.empty(); | |
2530 | for (int i = 0; i < MAX_STRANDS; i++) | |
2531 | { | |
2532 | ldb_[i].empty(); | |
2533 | stb_[i].empty(); | |
2534 | retStb_[i].empty(); | |
2535 | rmoStb_[i].empty(); | |
2536 | ifb_[i].empty(); | |
2537 | } | |
2538 | } | |
2539 | ||
2540 | //============================================================================= | |
2541 | // sniper is used by testbench to generate load/store traffic from | |
2542 | // simulated core(s), the goal is to test multi-core load/store traffic | |
2543 | // before multi-core RTL is available. The faked load/store traffic will | |
2544 | // be communicated with riesling through pli-msync commands. With real | |
2545 | // load/store, there are real load/store instructions to match/process them, | |
2546 | // but that won't be the case for sniper-generated load/store, so a new pli | |
2547 | // command SSTEP_SNIPER is added, it will be used to match up sniper-generated | |
2548 | // load/store events in msync, the SSTEP_SNIPER command will not cause any | |
2549 | // cpu architecture state change. | |
2550 | //============================================================================= | |
2551 | void MemorySync::handleSniper(int tid, uint64_t addr, INSTR_TYPE itype, uint64_t data) | |
2552 | { | |
2553 | MSYNC_DEBUG(4, "SNIPER: STEP (postMemAcc) tid=%d pa=%#llx itype=%d data=%#llx", tid, addr, (int)itype, data); | |
2554 | ||
2555 | if ((itype == ITYPE_LOAD) || | |
2556 | (itype == ITYPE_BLOCK_LOAD) || | |
2557 | (itype == ITYPE_DOUBLE_LOAD) || | |
2558 | (itype == ITYPE_QUAD_LOAD) || | |
2559 | (itype == ITYPE_ATOMIC)) | |
2560 | { | |
2561 | list<LoadStoreEntry>::iterator ii; | |
2562 | LoadStoreBuffer& ldb = ldb_[tid]; | |
2563 | if ((addr & IO_ADDR_BIT_MASK) == IO_ADDR_BIT_MASK) | |
2564 | { // IO address, IO can be out-of-order | |
2565 | ii = ldb.find1stNonExeMatchedAddr(addr); | |
2566 | } | |
2567 | else | |
2568 | { | |
2569 | ii = ldb.find1stNonExe(); | |
2570 | } | |
2571 | MSYNC_DEBUG (4, "ldb=%s", ldb.toString().c_str()); | |
2572 | if (ii == ldb.end()) | |
2573 | { | |
2574 | MS_ERROR("SNIPER: STEP failed to find LoadIssue entry (possibly DUT took trap & Riesling did not). tid=%d PA=%llx", tid, addr); | |
2575 | return; | |
2576 | } | |
2577 | // Note that RTL does not check load address, MemorySync performs | |
2578 | // this additional check for completion | |
2579 | if (addr != ii->getAddr()) | |
2580 | { | |
2581 | MS_ERROR("SNIPER: STEP's address mismatches with the 1st non-executed Load. tid=%d PA=%llx", tid, addr); | |
2582 | return; | |
2583 | } | |
2584 | if (!ii->isLinkValid()) | |
2585 | { | |
2586 | MS_ERROR("SNIPER: STEP's corresponding LoadData has not yet been issued. tid=%d PA=%llx", tid, addr); | |
2587 | return; | |
2588 | } | |
2589 | if (itype == ITYPE_ATOMIC) | |
2590 | { | |
2591 | // this assert is to make sure we can check atomic later | |
2592 | if (ii->getItype() != ITYPE_ATOMIC) | |
2593 | { | |
2594 | MS_ERROR("SNIPER: STEP (postMemAcc) atomic matches non-atomic load entry. tid=%d PA=%llx", tid, addr); | |
2595 | return; | |
2596 | } | |
2597 | } | |
2598 | if (ii->getLink()->getDsrc() == DSRC_L2_MEMORY) | |
2599 | { | |
2600 | ii->getLink()->setData(mab_.getL2Data(ii->getLink(), tid/NSTRANDS_PER_CORE, tid, addr)); | |
2601 | } | |
2602 | // msync data does not match the one provided by testbench | |
2603 | if (ii->getData() != data) | |
2604 | { | |
2605 | MS_ERROR("SNIPER: load data (%#llx) mismatches with testbench data (%#llx). tid=%d PA=%llx", ii->getData(), data, tid, addr); | |
2606 | return; | |
2607 | } | |
2608 | ii->setExecuted(true); | |
2609 | ii->getLink()->setExecuted(true); | |
2610 | list<MemoryAccessEntry>::iterator mii = ii->getLink(); | |
2611 | ldb.erase(ii); | |
2612 | ||
2613 | if (itype == ITYPE_ATOMIC) | |
2614 | { | |
2615 | list<LoadStoreEntry>::iterator slink; | |
2616 | LoadStoreBuffer& stb = stb_[tid]; | |
2617 | LoadStoreBuffer& rstb = retStb_[tid]; | |
2618 | ||
2619 | // find the corresponding StoreIssue | |
2620 | slink = rstb.find1stNonExe(); // must step in order | |
2621 | if (slink == rstb.end()) | |
2622 | { | |
2623 | slink = stb.find1stNonExe(); | |
2624 | if (slink == stb.end()) | |
2625 | { | |
2626 | MS_ERROR("SNIPER: STEP(at) failed to find non-executed StoreIssue entry (possibly DUT took trap & Riesling did not). tid=%d PA=%llx", tid, addr); | |
2627 | return; | |
2628 | } | |
2629 | } | |
2630 | if (addr != slink->getAddr()) | |
2631 | { | |
2632 | MS_ERROR("SNIPER: STEP(at) mis-matches addr with the 1st non-executed StoreIssue. tid=%d PA=%llx", tid, addr); | |
2633 | return; | |
2634 | } | |
2635 | if (slink->getItype() != ITYPE_ATOMIC) | |
2636 | { | |
2637 | MS_ERROR("SNIPER: STEP(at) found 1st non-executed StoreIssue non-atomic. tid=%d PA=%llx", tid, addr); | |
2638 | return; | |
2639 | } | |
2640 | if (!slink->isLinkValid()) | |
2641 | { | |
2642 | MS_ERROR("SNIPER: STEP(at) misses StoreCommit. tid=%d PA=%llx", tid, addr); | |
2643 | return; | |
2644 | } | |
2645 | if (slink->getLink()->getSizeV() == 0) | |
2646 | { | |
2647 | // CAS comparison is false => no write necessary | |
2648 | slink->setExecuted(true); | |
2649 | slink->getLink()->setExecuted(true); | |
2650 | // Fill the store part the same data written info as it load part | |
2651 | // --- TPS 9/10/04 | |
2652 | slink->getLink()->setData(mii->getData()); | |
2653 | slink->getLink()->setSizeV(mii->getSizeV()); | |
2654 | } | |
2655 | MSYNC_DEBUG(2, "slink=%s", (slink->toString()).c_str()); | |
2656 | } | |
2657 | // process completed entries in MAB, write STORE_COMMIT data to memory | |
2658 | // if ready | |
2659 | vector<MemoryAccessEntry> wdbuf; | |
2660 | mab_.popFront(&wdbuf, tsoChecker_); | |
2661 | bufWriteMemory(wdbuf); | |
2662 | } | |
2663 | ||
2664 | if ((itype == ITYPE_STORE) || | |
2665 | (itype == ITYPE_BLOCK_STORE) || | |
2666 | (itype == ITYPE_STORE_INIT) || | |
2667 | (itype == ITYPE_ATOMIC)) | |
2668 | { | |
2669 | list<LoadStoreEntry>::iterator ii; | |
2670 | LoadStoreBuffer& stb = stb_[tid]; | |
2671 | LoadStoreBuffer& rstb = retStb_[tid]; | |
2672 | LoadStoreBuffer& rmostb = rmoStb_[tid]; | |
2673 | LoadStoreBuffer& ret_or_rmo_stb = (itype == ITYPE_ATOMIC) ? rstb : rmostb; | |
2674 | MSYNC_DEBUG(2, "SNIPER: rstb=%s", rstb.toString().c_str()); | |
2675 | MSYNC_DEBUG(2, "SNIPER: stb=%s", stb.toString().c_str()); | |
2676 | MSYNC_DEBUG(2, "SNIPER: rmostb=%s", rmostb.toString().c_str()); | |
2677 | ii = ret_or_rmo_stb.find1stNonExe(); | |
2678 | if (ii == ret_or_rmo_stb.end()) | |
2679 | { | |
2680 | ii = stb.find1stNonExe(); | |
2681 | if (ii == stb.end()) | |
2682 | { | |
2683 | if (itype == ITYPE_ATOMIC) | |
2684 | { | |
2685 | MS_ERROR("SNIPER: STEP (store part of an atomic instr) failed to find match StoreIssue in STB. tid=%d PA=%llx", tid, addr); | |
2686 | return; | |
2687 | } | |
2688 | else | |
2689 | { | |
2690 | MS_ERROR("SNIPER: STEP failed to find match StoreIssue in STB. tid=%d PA=%llx", tid, addr); | |
2691 | return; | |
2692 | } | |
2693 | } | |
2694 | } | |
2695 | if ((addr & ADDR_MASK) != (ii->getAddr() & ADDR_MASK)) | |
2696 | { | |
2697 | MS_ERROR("SNIPER: STEP's address mismatches with 1st non-executed StoreIssue entry. tid=%d PA=%llx", tid, addr); | |
2698 | return; | |
2699 | } | |
2700 | ii->setExecuted(true); | |
2701 | if (ii->isLinkValid()) | |
2702 | { | |
2703 | (ii->getLink())->setExecuted(true); | |
2704 | } | |
2705 | if (ii->isRMOstore() && ii->isLinkValid()) | |
2706 | { | |
2707 | rmostb.erase(ii); | |
2708 | } | |
2709 | vector<MemoryAccessEntry> wdbuf; | |
2710 | mab_.popFront(&wdbuf, tsoChecker_); | |
2711 | // write STORE_COMMIT data to memory | |
2712 | bufWriteMemory(wdbuf); | |
2713 | } | |
2714 | } | |
2715 | ||
2716 | void MemorySync::setCoreEnable(int node_index, uint64_t data) | |
2717 | { | |
2718 | uint8_t bits = 0; | |
2719 | for (int i=0; i<8; i++) | |
2720 | { | |
2721 | if (data & 0xff) | |
2722 | { | |
2723 | bits |= (0x1 << i); | |
2724 | } | |
2725 | data = data >> 8; | |
2726 | } | |
2727 | inv_vec_mask[node_index] = bits; | |
2728 | } |