Ruby Memory Vector: Allow more than 4GB of memory
[gem5.git] / src / mem / physical.cc
index e8b978ec8fe1126fb793d6aa6489869b7658e5b3..5f92976f9dedc35ed7b45c32bd6d91dd8cb23eef 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2010 ARM Limited
+ * Copyright (c) 2012 ARM Limited
  * All rights reserved
  *
  * The license below extends only to copyright in the software and shall
@@ -11,9 +11,6 @@
  * unmodified and in its entirety in all distributions of the software,
  * modified or unmodified, in source code or in binary form.
  *
- * Copyright (c) 2001-2005 The Regents of The University of Michigan
- * All rights reserved.
- *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions are
  * met: redistributions of source code must retain the above copyright
  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  *
- * Authors: Ron Dreslinski
- *          Ali Saidi
+ * Authors: Andreas Hansson
  */
 
-#include <sys/mman.h>
-#include <sys/types.h>
-#include <sys/user.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <zlib.h>
-
-#include <cerrno>
-#include <cstdio>
-#include <iostream>
-#include <string>
-
-#include "arch/isa_traits.hh"
-#include "arch/registers.hh"
-#include "base/intmath.hh"
-#include "base/misc.hh"
-#include "base/random.hh"
-#include "base/types.hh"
-#include "config/the_isa.hh"
-#include "debug/LLSC.hh"
-#include "debug/MemoryAccess.hh"
-#include "mem/packet_access.hh"
+#include "debug/BusAddrRanges.hh"
 #include "mem/physical.hh"
-#include "sim/eventq.hh"
 
 using namespace std;
-using namespace TheISA;
-
-PhysicalMemory::PhysicalMemory(const Params *p)
-    : MemObject(p), pmemAddr(NULL), lat(p->latency), lat_var(p->latency_var),
-      _size(params()->range.size()), _start(params()->range.start)
-{
-    if (size() % TheISA::PageBytes != 0)
-        panic("Memory Size not divisible by page size\n");
-
-    if (params()->null)
-        return;
-
-
-    if (params()->file == "") {
-        int map_flags = MAP_ANON | MAP_PRIVATE;
-        pmemAddr = (uint8_t *)mmap(NULL, size(),
-                                   PROT_READ | PROT_WRITE, map_flags, -1, 0);
-    } else {
-        int map_flags = MAP_PRIVATE;
-        int fd = open(params()->file.c_str(), O_RDONLY);
-        _size = lseek(fd, 0, SEEK_END);
-        lseek(fd, 0, SEEK_SET);
-        pmemAddr = (uint8_t *)mmap(NULL, roundUp(size(), sysconf(_SC_PAGESIZE)),
-                                   PROT_READ | PROT_WRITE, map_flags, fd, 0);
-    }
-
-    if (pmemAddr == (void *)MAP_FAILED) {
-        perror("mmap");
-        if (params()->file == "")
-            fatal("Could not mmap!\n");
-        else
-            fatal("Could not find file: %s\n", params()->file);
-    }
-
-    //If requested, initialize all the memory to 0
-    if (p->zero)
-        memset(pmemAddr, 0, size());
-}
-
-void
-PhysicalMemory::init()
-{
-    if (ports.size() == 0) {
-        fatal("PhysicalMemory object %s is unconnected!", name());
-    }
-
-    for (PortIterator pi = ports.begin(); pi != ports.end(); ++pi) {
-        if (*pi)
-            (*pi)->sendStatusChange(Port::RangeChange);
-    }
-}
 
-PhysicalMemory::~PhysicalMemory()
+PhysicalMemory::PhysicalMemory(const vector<AbstractMemory*>& _memories) :
+    size(0)
 {
-    if (pmemAddr)
-        munmap((char*)pmemAddr, size());
-}
-
-unsigned
-PhysicalMemory::deviceBlockSize() const
-{
-    //Can accept anysize request
-    return 0;
-}
-
-Tick
-PhysicalMemory::calculateLatency(PacketPtr pkt)
-{
-    Tick latency = lat;
-    if (lat_var != 0)
-        latency += random_mt.random<Tick>(0, lat_var);
-    return latency;
-}
-
-
-
-// Add load-locked to tracking list.  Should only be called if the
-// operation is a load and the LLSC flag is set.
-void
-PhysicalMemory::trackLoadLocked(PacketPtr pkt)
-{
-    Request *req = pkt->req;
-    Addr paddr = LockedAddr::mask(req->getPaddr());
+    for (vector<AbstractMemory*>::const_iterator m = _memories.begin();
+         m != _memories.end(); ++m) {
+        // only add the memory if it is part of the global address map
+        if ((*m)->isInAddrMap()) {
+            memories.push_back(*m);
 
-    // first we check if we already have a locked addr for this
-    // xc.  Since each xc only gets one, we just update the
-    // existing record with the new address.
-    list<LockedAddr>::iterator i;
+            // calculate the total size once and for all
+            size += (*m)->size();
 
-    for (i = lockedAddrList.begin(); i != lockedAddrList.end(); ++i) {
-        if (i->matchesContext(req)) {
-            DPRINTF(LLSC, "Modifying lock record: context %d addr %#x\n",
-                    req->contextId(), paddr);
-            i->addr = paddr;
-            return;
+            // add the range to our interval tree and make sure it does not
+            // intersect an existing range
+            if (addrMap.insert((*m)->getAddrRange(), *m) == addrMap.end())
+                fatal("Memory address range for %s is overlapping\n",
+                      (*m)->name());
         }
+        DPRINTF(BusAddrRanges,
+                "Skipping memory %s that is not in global address map\n",
+                (*m)->name());
     }
-
-    // no record for this xc: need to allocate a new one
-    DPRINTF(LLSC, "Adding lock record: context %d addr %#x\n",
-            req->contextId(), paddr);
-    lockedAddrList.push_front(LockedAddr(req));
+    rangeCache.invalidate();
 }
 
-
-// Called on *writes* only... both regular stores and
-// store-conditional operations.  Check for conventional stores which
-// conflict with locked addresses, and for success/failure of store
-// conditionals.
 bool
-PhysicalMemory::checkLockedAddrList(PacketPtr pkt)
-{
-    Request *req = pkt->req;
-    Addr paddr = LockedAddr::mask(req->getPaddr());
-    bool isLLSC = pkt->isLLSC();
-
-    // Initialize return value.  Non-conditional stores always
-    // succeed.  Assume conditional stores will fail until proven
-    // otherwise.
-    bool success = !isLLSC;
-
-    // Iterate over list.  Note that there could be multiple matching
-    // records, as more than one context could have done a load locked
-    // to this location.
-    list<LockedAddr>::iterator i = lockedAddrList.begin();
-
-    while (i != lockedAddrList.end()) {
-
-        if (i->addr == paddr) {
-            // we have a matching address
-
-            if (isLLSC && i->matchesContext(req)) {
-                // it's a store conditional, and as far as the memory
-                // system can tell, the requesting context's lock is
-                // still valid.
-                DPRINTF(LLSC, "StCond success: context %d addr %#x\n",
-                        req->contextId(), paddr);
-                success = true;
-            }
-
-            // Get rid of our record of this lock and advance to next
-            DPRINTF(LLSC, "Erasing lock record: context %d addr %#x\n",
-                    i->contextId, paddr);
-            i = lockedAddrList.erase(i);
-        }
-        else {
-            // no match: advance to next record
-            ++i;
+PhysicalMemory::isMemAddr(Addr addr) const
+{
+    // see if the address is within the last matched range
+    if (addr != rangeCache) {
+        // lookup in the interval tree
+        range_map<Addr, AbstractMemory*>::const_iterator r =
+            addrMap.find(addr);
+        if (r == addrMap.end()) {
+            // not in the cache, and not in the tree
+            return false;
         }
+        // the range is in the tree, update the cache
+        rangeCache = r->first;
     }
 
-    if (isLLSC) {
-        req->setExtraData(success ? 1 : 0);
-    }
+    assert(addrMap.find(addr) != addrMap.end());
 
-    return success;
+    // either matched the cache or found in the tree
+    return true;
 }
 
-
-#if TRACING_ON
-
-#define CASE(A, T)                                                      \
-  case sizeof(T):                                                       \
-    DPRINTF(MemoryAccess,"%s of size %i on address 0x%x data 0x%x\n",   \
-            A, pkt->getSize(), pkt->getAddr(), pkt->get<T>());          \
-  break
-
-
-#define TRACE_PACKET(A)                                                 \
-    do {                                                                \
-        switch (pkt->getSize()) {                                       \
-          CASE(A, uint64_t);                                            \
-          CASE(A, uint32_t);                                            \
-          CASE(A, uint16_t);                                            \
-          CASE(A, uint8_t);                                             \
-          default:                                                      \
-            DPRINTF(MemoryAccess, "%s of size %i on address 0x%x\n",    \
-                    A, pkt->getSize(), pkt->getAddr());                 \
-            DDUMP(MemoryAccess, pkt->getPtr<uint8_t>(), pkt->getSize());\
-        }                                                               \
-    } while (0)
-
-#else
-
-#define TRACE_PACKET(A)
-
-#endif
-
-Tick
-PhysicalMemory::doAtomicAccess(PacketPtr pkt)
+AddrRangeList
+PhysicalMemory::getConfAddrRanges() const
 {
-    assert(pkt->getAddr() >= start() &&
-           pkt->getAddr() + pkt->getSize() <= start() + size());
-
-    if (pkt->memInhibitAsserted()) {
-        DPRINTF(MemoryAccess, "mem inhibited on 0x%x: not responding\n",
-                pkt->getAddr());
-        return 0;
-    }
-
-    uint8_t *hostAddr = pmemAddr + pkt->getAddr() - start();
-
-    if (pkt->cmd == MemCmd::SwapReq) {
-        IntReg overwrite_val;
-        bool overwrite_mem;
-        uint64_t condition_val64;
-        uint32_t condition_val32;
-
-        if (!pmemAddr)
-            panic("Swap only works if there is real memory (i.e. null=False)");
-        assert(sizeof(IntReg) >= pkt->getSize());
-
-        overwrite_mem = true;
-        // keep a copy of our possible write value, and copy what is at the
-        // memory address into the packet
-        std::memcpy(&overwrite_val, pkt->getPtr<uint8_t>(), pkt->getSize());
-        std::memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
-
-        if (pkt->req->isCondSwap()) {
-            if (pkt->getSize() == sizeof(uint64_t)) {
-                condition_val64 = pkt->req->getExtraData();
-                overwrite_mem = !std::memcmp(&condition_val64, hostAddr,
-                                             sizeof(uint64_t));
-            } else if (pkt->getSize() == sizeof(uint32_t)) {
-                condition_val32 = (uint32_t)pkt->req->getExtraData();
-                overwrite_mem = !std::memcmp(&condition_val32, hostAddr,
-                                             sizeof(uint32_t));
-            } else
-                panic("Invalid size for conditional read/write\n");
+    // this could be done once in the constructor, but since it is unlikely to
+    // be called more than once the iteration should not be a problem
+    AddrRangeList ranges;
+    for (vector<AbstractMemory*>::const_iterator m = memories.begin();
+         m != memories.end(); ++m) {
+        if ((*m)->isConfReported()) {
+            ranges.push_back((*m)->getAddrRange());
         }
-
-        if (overwrite_mem)
-            std::memcpy(hostAddr, &overwrite_val, pkt->getSize());
-
-        assert(!pkt->req->isInstFetch());
-        TRACE_PACKET("Read/Write");
-    } else if (pkt->isRead()) {
-        assert(!pkt->isWrite());
-        if (pkt->isLLSC()) {
-            trackLoadLocked(pkt);
-        }
-        if (pmemAddr)
-            memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
-        TRACE_PACKET(pkt->req->isInstFetch() ? "IFetch" : "Read");
-    } else if (pkt->isWrite()) {
-        if (writeOK(pkt)) {
-            if (pmemAddr)
-                memcpy(hostAddr, pkt->getPtr<uint8_t>(), pkt->getSize());
-            assert(!pkt->req->isInstFetch());
-            TRACE_PACKET("Write");
-        }
-    } else if (pkt->isInvalidate()) {
-        //upgrade or invalidate
-        if (pkt->needsResponse()) {
-            pkt->makeAtomicResponse();
-        }
-    } else {
-        panic("unimplemented");
     }
 
-    if (pkt->needsResponse()) {
-        pkt->makeAtomicResponse();
-    }
-    return calculateLatency(pkt);
+    return ranges;
 }
 
-
 void
-PhysicalMemory::doFunctionalAccess(PacketPtr pkt)
+PhysicalMemory::access(PacketPtr pkt)
 {
-    assert(pkt->getAddr() >= start() &&
-           pkt->getAddr() + pkt->getSize() <= start() + size());
-
-
-    uint8_t *hostAddr = pmemAddr + pkt->getAddr() - start();
-
-    if (pkt->isRead()) {
-        if (pmemAddr)
-            memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
-        TRACE_PACKET("Read");
-        pkt->makeAtomicResponse();
-    } else if (pkt->isWrite()) {
-        if (pmemAddr)
-            memcpy(hostAddr, pkt->getPtr<uint8_t>(), pkt->getSize());
-        TRACE_PACKET("Write");
-        pkt->makeAtomicResponse();
-    } else if (pkt->isPrint()) {
-        Packet::PrintReqState *prs =
-            dynamic_cast<Packet::PrintReqState*>(pkt->senderState);
-        // Need to call printLabels() explicitly since we're not going
-        // through printObj().
-        prs->printLabels();
-        // Right now we just print the single byte at the specified address.
-        ccprintf(prs->os, "%s%#x\n", prs->curPrefix(), *hostAddr);
-    } else {
-        panic("PhysicalMemory: unimplemented functional command %s",
-              pkt->cmdString());
-    }
-}
-
-
-Port *
-PhysicalMemory::getPort(const std::string &if_name, int idx)
-{
-    // Accept request for "functional" port for backwards compatibility
-    // with places where this function is called from C++.  I'd prefer
-    // to move all these into Python someday.
-    if (if_name == "functional") {
-        return new MemoryPort(csprintf("%s-functional", name()), this);
-    }
-
-    if (if_name != "port") {
-        panic("PhysicalMemory::getPort: unknown port %s requested", if_name);
-    }
-
-    if (idx >= (int)ports.size()) {
-        ports.resize(idx + 1);
-    }
-
-    if (ports[idx] != NULL) {
-        panic("PhysicalMemory::getPort: port %d already assigned", idx);
-    }
-
-    MemoryPort *port =
-        new MemoryPort(csprintf("%s-port%d", name(), idx), this);
-
-    ports[idx] = port;
-    return port;
-}
-
-
-void
-PhysicalMemory::recvStatusChange(Port::Status status)
-{
-}
-
-PhysicalMemory::MemoryPort::MemoryPort(const std::string &_name,
-                                       PhysicalMemory *_memory)
-    : SimpleTimingPort(_name, _memory), memory(_memory)
-{ }
-
-void
-PhysicalMemory::MemoryPort::recvStatusChange(Port::Status status)
-{
-    memory->recvStatusChange(status);
-}
-
-void
-PhysicalMemory::MemoryPort::getDeviceAddressRanges(AddrRangeList &resp,
-                                                   bool &snoop)
-{
-    memory->getAddressRanges(resp, snoop);
-}
-
-void
-PhysicalMemory::getAddressRanges(AddrRangeList &resp, bool &snoop)
-{
-    snoop = false;
-    resp.clear();
-    resp.push_back(RangeSize(start(), size()));
-}
-
-unsigned
-PhysicalMemory::MemoryPort::deviceBlockSize() const
-{
-    return memory->deviceBlockSize();
-}
-
-Tick
-PhysicalMemory::MemoryPort::recvAtomic(PacketPtr pkt)
-{
-    return memory->doAtomicAccess(pkt);
-}
-
-void
-PhysicalMemory::MemoryPort::recvFunctional(PacketPtr pkt)
-{
-    pkt->pushLabel(memory->name());
-
-    if (!checkFunctional(pkt)) {
-        // Default implementation of SimpleTimingPort::recvFunctional()
-        // calls recvAtomic() and throws away the latency; we can save a
-        // little here by just not calculating the latency.
-        memory->doFunctionalAccess(pkt);
-    }
-
-    pkt->popLabel();
-}
-
-unsigned int
-PhysicalMemory::drain(Event *de)
-{
-    int count = 0;
-    for (PortIterator pi = ports.begin(); pi != ports.end(); ++pi) {
-        count += (*pi)->drain(de);
-    }
-
-    if (count)
-        changeState(Draining);
-    else
-        changeState(Drained);
-    return count;
+    assert(pkt->isRequest());
+    Addr addr = pkt->getAddr();
+    range_map<Addr, AbstractMemory*>::const_iterator m = addrMap.find(addr);
+    assert(m != addrMap.end());
+    m->second->access(pkt);
 }
 
 void
-PhysicalMemory::serialize(ostream &os)
-{
-    if (!pmemAddr)
-        return;
-
-    gzFile compressedMem;
-    string filename = name() + ".physmem";
-
-    SERIALIZE_SCALAR(filename);
-    SERIALIZE_SCALAR(_size);
-
-    // write memory file
-    string thefile = Checkpoint::dir() + "/" + filename.c_str();
-    int fd = creat(thefile.c_str(), 0664);
-    if (fd < 0) {
-        perror("creat");
-        fatal("Can't open physical memory checkpoint file '%s'\n", filename);
-    }
-
-    compressedMem = gzdopen(fd, "wb");
-    if (compressedMem == NULL)
-        fatal("Insufficient memory to allocate compression state for %s\n",
-                filename);
-
-    if (gzwrite(compressedMem, pmemAddr, size()) != (int)size()) {
-        fatal("Write failed on physical memory checkpoint file '%s'\n",
-              filename);
-    }
-
-    if (gzclose(compressedMem))
-        fatal("Close failed on physical memory checkpoint file '%s'\n",
-              filename);
-
-    list<LockedAddr>::iterator i = lockedAddrList.begin();
-
-    vector<Addr> lal_addr;
-    vector<int> lal_cid;
-    while (i != lockedAddrList.end()) {
-        lal_addr.push_back(i->addr);
-        lal_cid.push_back(i->contextId);
-        i++;
-    }
-    arrayParamOut(os, "lal_addr", lal_addr);
-    arrayParamOut(os, "lal_cid", lal_cid);
-}
-
-void
-PhysicalMemory::unserialize(Checkpoint *cp, const string &section)
-{
-    if (!pmemAddr)
-        return;
-
-    gzFile compressedMem;
-    long *tempPage;
-    long *pmem_current;
-    uint64_t curSize;
-    uint32_t bytesRead;
-    const uint32_t chunkSize = 16384;
-
-    string filename;
-
-    UNSERIALIZE_SCALAR(filename);
-
-    filename = cp->cptDir + "/" + filename;
-
-    // mmap memoryfile
-    int fd = open(filename.c_str(), O_RDONLY);
-    if (fd < 0) {
-        perror("open");
-        fatal("Can't open physical memory checkpoint file '%s'", filename);
-    }
-
-    compressedMem = gzdopen(fd, "rb");
-    if (compressedMem == NULL)
-        fatal("Insufficient memory to allocate compression state for %s\n",
-                filename);
-
-    // unmap file that was mmapped in the constructor
-    // This is done here to make sure that gzip and open don't muck with our
-    // nice large space of memory before we reallocate it
-    munmap((char*)pmemAddr, size());
-
-    UNSERIALIZE_SCALAR(_size);
-    if (size() > params()->range.size())
-        fatal("Memory size has changed!\n");
-
-    pmemAddr = (uint8_t *)mmap(NULL, size(),
-        PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, -1, 0);
-
-    if (pmemAddr == (void *)MAP_FAILED) {
-        perror("mmap");
-        fatal("Could not mmap physical memory!\n");
-    }
-
-    curSize = 0;
-    tempPage = (long*)malloc(chunkSize);
-    if (tempPage == NULL)
-        fatal("Unable to malloc memory to read file %s\n", filename);
-
-    /* Only copy bytes that are non-zero, so we don't give the VM system hell */
-    while (curSize < size()) {
-        bytesRead = gzread(compressedMem, tempPage, chunkSize);
-        if (bytesRead == 0)
-            break;
-
-        assert(bytesRead % sizeof(long) == 0);
-
-        for (uint32_t x = 0; x < bytesRead / sizeof(long); x++)
-        {
-             if (*(tempPage+x) != 0) {
-                 pmem_current = (long*)(pmemAddr + curSize + x * sizeof(long));
-                 *pmem_current = *(tempPage+x);
-             }
-        }
-        curSize += bytesRead;
-    }
-
-    free(tempPage);
-
-    if (gzclose(compressedMem))
-        fatal("Close failed on physical memory checkpoint file '%s'\n",
-              filename);
-
-    vector<Addr> lal_addr;
-    vector<int> lal_cid;
-    arrayParamIn(cp, section, "lal_addr", lal_addr);
-    arrayParamIn(cp, section, "lal_cid", lal_cid);
-    for(int i = 0; i < lal_addr.size(); i++)
-        lockedAddrList.push_front(LockedAddr(lal_addr[i], lal_cid[i]));
-}
-
-PhysicalMemory *
-PhysicalMemoryParams::create()
+PhysicalMemory::functionalAccess(PacketPtr pkt)
 {
-    return new PhysicalMemory(this);
+    assert(pkt->isRequest());
+    Addr addr = pkt->getAddr();
+    range_map<Addr, AbstractMemory*>::const_iterator m = addrMap.find(addr);
+    assert(m != addrMap.end());
+    m->second->functionalAccess(pkt);
 }