physical.cc revision 5222:bb733a878f85
1/*
2 * Copyright (c) 2001-2005 The Regents of The University of Michigan
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met: redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer;
9 * redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution;
12 * neither the name of the copyright holders nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Authors: Ron Dreslinski
29 *          Ali Saidi
30 */
31
32#include <sys/types.h>
33#include <sys/mman.h>
34#include <errno.h>
35#include <fcntl.h>
36#include <unistd.h>
37#include <zlib.h>
38
39#include <iostream>
40#include <string>
41
42#include "arch/isa_traits.hh"
43#include "base/misc.hh"
44#include "config/full_system.hh"
45#include "mem/packet_access.hh"
46#include "mem/physical.hh"
47#include "sim/eventq.hh"
48#include "sim/host.hh"
49
50using namespace std;
51using namespace TheISA;
52
53PhysicalMemory::PhysicalMemory(const Params *p)
54    : MemObject(p), pmemAddr(NULL), lat(p->latency)
55{
56    if (params()->range.size() % TheISA::PageBytes != 0)
57        panic("Memory Size not divisible by page size\n");
58
59    int map_flags = MAP_ANON | MAP_PRIVATE;
60    pmemAddr = (uint8_t *)mmap(NULL, params()->range.size(),
61                               PROT_READ | PROT_WRITE, map_flags, -1, 0);
62
63    if (pmemAddr == (void *)MAP_FAILED) {
64        perror("mmap");
65        fatal("Could not mmap!\n");
66    }
67
68    //If requested, initialize all the memory to 0
69    if (p->zero)
70        memset(pmemAddr, 0, p->range.size());
71
72    pagePtr = 0;
73}
74
75void
76PhysicalMemory::init()
77{
78    if (ports.size() == 0) {
79        fatal("PhysicalMemory object %s is unconnected!", name());
80    }
81
82    for (PortIterator pi = ports.begin(); pi != ports.end(); ++pi) {
83        if (*pi)
84            (*pi)->sendStatusChange(Port::RangeChange);
85    }
86}
87
88PhysicalMemory::~PhysicalMemory()
89{
90    if (pmemAddr)
91        munmap((char*)pmemAddr, params()->range.size());
92    //Remove memPorts?
93}
94
95Addr
96PhysicalMemory::new_page()
97{
98    Addr return_addr = pagePtr << LogVMPageSize;
99    return_addr += start();
100
101    ++pagePtr;
102    return return_addr;
103}
104
105int
106PhysicalMemory::deviceBlockSize()
107{
108    //Can accept anysize request
109    return 0;
110}
111
112Tick
113PhysicalMemory::calculateLatency(PacketPtr pkt)
114{
115    return lat;
116}
117
118
119
120// Add load-locked to tracking list.  Should only be called if the
121// operation is a load and the LOCKED flag is set.
122void
123PhysicalMemory::trackLoadLocked(PacketPtr pkt)
124{
125    Request *req = pkt->req;
126    Addr paddr = LockedAddr::mask(req->getPaddr());
127
128    // first we check if we already have a locked addr for this
129    // xc.  Since each xc only gets one, we just update the
130    // existing record with the new address.
131    list<LockedAddr>::iterator i;
132
133    for (i = lockedAddrList.begin(); i != lockedAddrList.end(); ++i) {
134        if (i->matchesContext(req)) {
135            DPRINTF(LLSC, "Modifying lock record: cpu %d thread %d addr %#x\n",
136                    req->getCpuNum(), req->getThreadNum(), paddr);
137            i->addr = paddr;
138            return;
139        }
140    }
141
142    // no record for this xc: need to allocate a new one
143    DPRINTF(LLSC, "Adding lock record: cpu %d thread %d addr %#x\n",
144            req->getCpuNum(), req->getThreadNum(), paddr);
145    lockedAddrList.push_front(LockedAddr(req));
146}
147
148
149// Called on *writes* only... both regular stores and
150// store-conditional operations.  Check for conventional stores which
151// conflict with locked addresses, and for success/failure of store
152// conditionals.
153bool
154PhysicalMemory::checkLockedAddrList(PacketPtr pkt)
155{
156    Request *req = pkt->req;
157    Addr paddr = LockedAddr::mask(req->getPaddr());
158    bool isLocked = pkt->isLocked();
159
160    // Initialize return value.  Non-conditional stores always
161    // succeed.  Assume conditional stores will fail until proven
162    // otherwise.
163    bool success = !isLocked;
164
165    // Iterate over list.  Note that there could be multiple matching
166    // records, as more than one context could have done a load locked
167    // to this location.
168    list<LockedAddr>::iterator i = lockedAddrList.begin();
169
170    while (i != lockedAddrList.end()) {
171
172        if (i->addr == paddr) {
173            // we have a matching address
174
175            if (isLocked && i->matchesContext(req)) {
176                // it's a store conditional, and as far as the memory
177                // system can tell, the requesting context's lock is
178                // still valid.
179                DPRINTF(LLSC, "StCond success: cpu %d thread %d addr %#x\n",
180                        req->getCpuNum(), req->getThreadNum(), paddr);
181                success = true;
182            }
183
184            // Get rid of our record of this lock and advance to next
185            DPRINTF(LLSC, "Erasing lock record: cpu %d thread %d addr %#x\n",
186                    i->cpuNum, i->threadNum, paddr);
187            i = lockedAddrList.erase(i);
188        }
189        else {
190            // no match: advance to next record
191            ++i;
192        }
193    }
194
195    if (isLocked) {
196        req->setExtraData(success ? 1 : 0);
197    }
198
199    return success;
200}
201
202
203#if TRACING_ON
204
205#define CASE(A, T)                                                      \
206  case sizeof(T):                                                       \
207    DPRINTF(MemoryAccess, A " of size %i on address 0x%x data 0x%x\n",  \
208            pkt->getSize(), pkt->getAddr(), pkt->get<T>());             \
209  break
210
211
212#define TRACE_PACKET(A)                                                 \
213    do {                                                                \
214        switch (pkt->getSize()) {                                       \
215          CASE(A, uint64_t);                                            \
216          CASE(A, uint32_t);                                            \
217          CASE(A, uint16_t);                                            \
218          CASE(A, uint8_t);                                             \
219          default:                                                      \
220            DPRINTF(MemoryAccess, A " of size %i on address 0x%x\n",    \
221                    pkt->getSize(), pkt->getAddr());                    \
222        }                                                               \
223    } while (0)
224
225#else
226
227#define TRACE_PACKET(A)
228
229#endif
230
231Tick
232PhysicalMemory::doAtomicAccess(PacketPtr pkt)
233{
234    assert(pkt->getAddr() >= start() &&
235           pkt->getAddr() + pkt->getSize() <= start() + size());
236
237    if (pkt->memInhibitAsserted()) {
238        DPRINTF(MemoryAccess, "mem inhibited on 0x%x: not responding\n",
239                pkt->getAddr());
240        return 0;
241    }
242
243    uint8_t *hostAddr = pmemAddr + pkt->getAddr() - start();
244
245    if (pkt->cmd == MemCmd::SwapReq) {
246        IntReg overwrite_val;
247        bool overwrite_mem;
248        uint64_t condition_val64;
249        uint32_t condition_val32;
250
251        assert(sizeof(IntReg) >= pkt->getSize());
252
253        overwrite_mem = true;
254        // keep a copy of our possible write value, and copy what is at the
255        // memory address into the packet
256        std::memcpy(&overwrite_val, pkt->getPtr<uint8_t>(), pkt->getSize());
257        std::memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
258
259        if (pkt->req->isCondSwap()) {
260            if (pkt->getSize() == sizeof(uint64_t)) {
261                condition_val64 = pkt->req->getExtraData();
262                overwrite_mem = !std::memcmp(&condition_val64, hostAddr,
263                                             sizeof(uint64_t));
264            } else if (pkt->getSize() == sizeof(uint32_t)) {
265                condition_val32 = (uint32_t)pkt->req->getExtraData();
266                overwrite_mem = !std::memcmp(&condition_val32, hostAddr,
267                                             sizeof(uint32_t));
268            } else
269                panic("Invalid size for conditional read/write\n");
270        }
271
272        if (overwrite_mem)
273            std::memcpy(hostAddr, &overwrite_val, pkt->getSize());
274
275        TRACE_PACKET("Read/Write");
276    } else if (pkt->isRead()) {
277        assert(!pkt->isWrite());
278        if (pkt->isLocked()) {
279            trackLoadLocked(pkt);
280        }
281        memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
282        TRACE_PACKET("Read");
283    } else if (pkt->isWrite()) {
284        if (writeOK(pkt)) {
285            memcpy(hostAddr, pkt->getPtr<uint8_t>(), pkt->getSize());
286            TRACE_PACKET("Write");
287        }
288    } else if (pkt->isInvalidate()) {
289        //upgrade or invalidate
290        if (pkt->needsResponse()) {
291            pkt->makeAtomicResponse();
292        }
293    } else {
294        panic("unimplemented");
295    }
296
297    if (pkt->needsResponse()) {
298        pkt->makeAtomicResponse();
299    }
300    return calculateLatency(pkt);
301}
302
303
304void
305PhysicalMemory::doFunctionalAccess(PacketPtr pkt)
306{
307    warn("addr %#x >= %#x AND %#x <= %#x",
308         pkt->getAddr(), start(), pkt->getAddr() + pkt->getSize(), start() + size());
309
310    assert(pkt->getAddr() >= start() &&
311           pkt->getAddr() + pkt->getSize() <= start() + size());
312
313
314    uint8_t *hostAddr = pmemAddr + pkt->getAddr() - start();
315
316    if (pkt->cmd == MemCmd::ReadReq) {
317        memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
318        TRACE_PACKET("Read");
319    } else if (pkt->cmd == MemCmd::WriteReq) {
320        memcpy(hostAddr, pkt->getPtr<uint8_t>(), pkt->getSize());
321        TRACE_PACKET("Write");
322    } else {
323        panic("PhysicalMemory: unimplemented functional command %s",
324              pkt->cmdString());
325    }
326
327    pkt->makeAtomicResponse();
328}
329
330
331Port *
332PhysicalMemory::getPort(const std::string &if_name, int idx)
333{
334    // Accept request for "functional" port for backwards compatibility
335    // with places where this function is called from C++.  I'd prefer
336    // to move all these into Python someday.
337    if (if_name == "functional") {
338        return new MemoryPort(csprintf("%s-functional", name()), this);
339    }
340
341    if (if_name != "port") {
342        panic("PhysicalMemory::getPort: unknown port %s requested", if_name);
343    }
344
345    if (idx >= ports.size()) {
346        ports.resize(idx+1);
347    }
348
349    if (ports[idx] != NULL) {
350        panic("PhysicalMemory::getPort: port %d already assigned", idx);
351    }
352
353    MemoryPort *port =
354        new MemoryPort(csprintf("%s-port%d", name(), idx), this);
355
356    ports[idx] = port;
357    return port;
358}
359
360
361void
362PhysicalMemory::recvStatusChange(Port::Status status)
363{
364}
365
366PhysicalMemory::MemoryPort::MemoryPort(const std::string &_name,
367                                       PhysicalMemory *_memory)
368    : SimpleTimingPort(_name), memory(_memory)
369{ }
370
371void
372PhysicalMemory::MemoryPort::recvStatusChange(Port::Status status)
373{
374    memory->recvStatusChange(status);
375}
376
377void
378PhysicalMemory::MemoryPort::getDeviceAddressRanges(AddrRangeList &resp,
379                                                   bool &snoop)
380{
381    memory->getAddressRanges(resp, snoop);
382}
383
384void
385PhysicalMemory::getAddressRanges(AddrRangeList &resp, bool &snoop)
386{
387    snoop = false;
388    resp.clear();
389    resp.push_back(RangeSize(start(), params()->range.size()));
390}
391
392int
393PhysicalMemory::MemoryPort::deviceBlockSize()
394{
395    return memory->deviceBlockSize();
396}
397
398Tick
399PhysicalMemory::MemoryPort::recvAtomic(PacketPtr pkt)
400{
401    return memory->doAtomicAccess(pkt);
402}
403
404void
405PhysicalMemory::MemoryPort::recvFunctional(PacketPtr pkt)
406{
407    if (!checkFunctional(pkt)) {
408        // Default implementation of SimpleTimingPort::recvFunctional()
409        // calls recvAtomic() and throws away the latency; we can save a
410        // little here by just not calculating the latency.
411        memory->doFunctionalAccess(pkt);
412    }
413}
414
415unsigned int
416PhysicalMemory::drain(Event *de)
417{
418    int count = 0;
419    for (PortIterator pi = ports.begin(); pi != ports.end(); ++pi) {
420        count += (*pi)->drain(de);
421    }
422
423    if (count)
424        changeState(Draining);
425    else
426        changeState(Drained);
427    return count;
428}
429
430void
431PhysicalMemory::serialize(ostream &os)
432{
433    gzFile compressedMem;
434    string filename = name() + ".physmem";
435
436    SERIALIZE_SCALAR(filename);
437
438    // write memory file
439    string thefile = Checkpoint::dir() + "/" + filename.c_str();
440    int fd = creat(thefile.c_str(), 0664);
441    if (fd < 0) {
442        perror("creat");
443        fatal("Can't open physical memory checkpoint file '%s'\n", filename);
444    }
445
446    compressedMem = gzdopen(fd, "wb");
447    if (compressedMem == NULL)
448        fatal("Insufficient memory to allocate compression state for %s\n",
449                filename);
450
451    if (gzwrite(compressedMem, pmemAddr, params()->range.size()) !=
452        params()->range.size()) {
453        fatal("Write failed on physical memory checkpoint file '%s'\n",
454              filename);
455    }
456
457    if (gzclose(compressedMem))
458        fatal("Close failed on physical memory checkpoint file '%s'\n",
459              filename);
460}
461
462void
463PhysicalMemory::unserialize(Checkpoint *cp, const string &section)
464{
465    gzFile compressedMem;
466    long *tempPage;
467    long *pmem_current;
468    uint64_t curSize;
469    uint32_t bytesRead;
470    const int chunkSize = 16384;
471
472
473    string filename;
474
475    UNSERIALIZE_SCALAR(filename);
476
477    filename = cp->cptDir + "/" + filename;
478
479    // mmap memoryfile
480    int fd = open(filename.c_str(), O_RDONLY);
481    if (fd < 0) {
482        perror("open");
483        fatal("Can't open physical memory checkpoint file '%s'", filename);
484    }
485
486    compressedMem = gzdopen(fd, "rb");
487    if (compressedMem == NULL)
488        fatal("Insufficient memory to allocate compression state for %s\n",
489                filename);
490
491    // unmap file that was mmaped in the constructor
492    // This is done here to make sure that gzip and open don't muck with our
493    // nice large space of memory before we reallocate it
494    munmap((char*)pmemAddr, params()->range.size());
495
496    pmemAddr = (uint8_t *)mmap(NULL, params()->range.size(),
497        PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
498
499    if (pmemAddr == (void *)MAP_FAILED) {
500        perror("mmap");
501        fatal("Could not mmap physical memory!\n");
502    }
503
504    curSize = 0;
505    tempPage = (long*)malloc(chunkSize);
506    if (tempPage == NULL)
507        fatal("Unable to malloc memory to read file %s\n", filename);
508
509    /* Only copy bytes that are non-zero, so we don't give the VM system hell */
510    while (curSize < params()->range.size()) {
511        bytesRead = gzread(compressedMem, tempPage, chunkSize);
512        if (bytesRead != chunkSize &&
513            bytesRead != params()->range.size() - curSize)
514            fatal("Read failed on physical memory checkpoint file '%s'"
515                  " got %d bytes, expected %d or %d bytes\n",
516                  filename, bytesRead, chunkSize,
517                  params()->range.size() - curSize);
518
519        assert(bytesRead % sizeof(long) == 0);
520
521        for (int x = 0; x < bytesRead/sizeof(long); x++)
522        {
523             if (*(tempPage+x) != 0) {
524                 pmem_current = (long*)(pmemAddr + curSize + x * sizeof(long));
525                 *pmem_current = *(tempPage+x);
526             }
527        }
528        curSize += bytesRead;
529    }
530
531    free(tempPage);
532
533    if (gzclose(compressedMem))
534        fatal("Close failed on physical memory checkpoint file '%s'\n",
535              filename);
536
537}
538
539PhysicalMemory *
540PhysicalMemoryParams::create()
541{
542    return new PhysicalMemory(this);
543}
544