Disk: ReadDriver: Add unified cache

Previously, we had to copy an entire block out of the
old cache every time we wanted to read even a single
byte from it.

This ended up being a fairly significant performance
issue, in addition to the fact that the caching code
was duplicated.
This commit is contained in:
Paul Hollinsky
2022-04-14 18:26:44 -04:00
parent d45d708446
commit 103f938d69
10 changed files with 316 additions and 245 deletions
+48 -3
View File
@@ -9,14 +9,25 @@ optional<uint64_t> ReadDriver::readLogicalDisk(Communication& com, device_eventh
if(amount == 0)
return 0;
optional<uint64_t> ret;
pos += vsaOffset;
// First read from the cache
optional<uint64_t> ret = readFromCache(pos, into, amount);
if(ret == amount) // Full cache hit, we're done
return ret;
const uint64_t totalAmount = amount;
if(ret.has_value()) { // Partial cache hit
pos += *ret;
into += *ret;
amount -= *ret;
}
// Read into here if we can't read directly into the user buffer
// That would be the case either if we don't want some at the
// beginning or end of the block.
std::vector<uint8_t> alignedReadBuffer;
pos += vsaOffset;
const uint32_t idealBlockSize = getBlockSizeBounds().second;
const uint64_t startBlock = pos / idealBlockSize;
const uint32_t posWithinFirstBlock = static_cast<uint32_t>(pos % idealBlockSize);
@@ -36,7 +47,7 @@ optional<uint64_t> ReadDriver::readLogicalDisk(Communication& com, device_eventh
const uint32_t posWithinCurrentBlock = (blocksProcessed ? 0 : posWithinFirstBlock);
uint32_t curAmt = idealBlockSize - posWithinCurrentBlock;
const auto amountLeft = amount - ret.value_or(0);
const auto amountLeft = totalAmount - ret.value_or(0);
if(curAmt > amountLeft)
curAmt = static_cast<uint32_t>(amountLeft);
@@ -65,7 +76,41 @@ optional<uint64_t> ReadDriver::readLogicalDisk(Communication& com, device_eventh
ret.emplace();
*ret += std::min<uint64_t>(*readAmount, curAmt);
blocksProcessed++;
if(blocksProcessed == blocks) {
// Last block, add to the cache
if(useAlignedReadBuffer) {
cache = std::move(alignedReadBuffer);
} else {
if(cache.size() != idealBlockSize)
cache.resize(idealBlockSize);
memcpy(cache.data(), into + intoOffset, idealBlockSize);
}
cachePos = currentBlock * idealBlockSize;
cachedAt = std::chrono::steady_clock::now();
}
}
return ret;
}
void ReadDriver::invalidateCache(uint64_t pos, uint64_t amount) {
if(pos <= cachePos + cache.size() && pos + amount >= cachePos)
cache.clear();
}
optional<uint64_t> ReadDriver::readFromCache(uint64_t pos, uint8_t* into, uint64_t amount, std::chrono::milliseconds staleAfter) {
if(cache.empty())
return nullopt; // Nothing in the cache
if(cachedAt + staleAfter < std::chrono::steady_clock::now())
return nullopt; // Cache is stale
if(pos > cachePos + cache.size() || pos < cachePos)
return nullopt; // Cache miss
const auto cacheOffset = pos - cachePos;
const auto copyAmount = std::min<uint64_t>(cache.size() - cacheOffset, amount);
memcpy(into, cache.data() + cacheOffset, static_cast<size_t>(copyAmount));
return copyAmount;
}
+5
View File
@@ -75,6 +75,7 @@ optional<uint64_t> WriteDriver::writeLogicalDisk(Communication& com, device_even
if(bytesTransferred == RetryAtomic) {
// The user may want to log these events in order to see how many atomic misses they are getting
report(APIEvent::Type::AtomicOperationRetried, APIEvent::Severity::EventInfo);
readDriver.invalidateCache(currentBlock * idealBlockSize, idealBlockSize);
continue;
}
@@ -93,5 +94,9 @@ optional<uint64_t> WriteDriver::writeLogicalDisk(Communication& com, device_even
blocksProcessed++;
}
// No matter how much succeeded, to be safe, we'll invalidate anything
// we may have even tried to write, since it may have succeeded without
// notifying, etc.
readDriver.invalidateCache(pos, amount);
return ret;
}
+141 -148
View File
@@ -37,177 +37,170 @@ optional<uint64_t> ExtExtractorDiskReadDriver::attemptReadLogicalDiskAligned(Com
uint64_t pos, uint8_t* into, uint64_t amount, std::chrono::milliseconds timeout) {
static std::shared_ptr<MessageFilter> NeoMemorySDRead = std::make_shared<MessageFilter>(Network::NetID::NeoMemorySDRead);
if(cachePos != pos || std::chrono::steady_clock::now() > cachedAt + CacheTime) {
uint64_t sector = pos / SectorSize;
uint64_t sector = pos / SectorSize;
uint64_t largeSectorCount = amount / SectorSize;
uint32_t sectorCount = uint32_t(largeSectorCount);
if (largeSectorCount != uint64_t(sectorCount))
return nullopt;
uint64_t largeSectorCount = amount / SectorSize;
uint32_t sectorCount = uint32_t(largeSectorCount);
if (largeSectorCount != uint64_t(sectorCount))
return nullopt;
// The cache does not have this data, go get it
std::mutex m;
std::condition_variable cv;
uint16_t receiving = 0; // How much are we about to get before another header or completion
uint64_t received = 0;
uint16_t receivedCurrent = 0;
size_t skipping = 0;
std::vector<uint8_t> header;
std::unique_lock<std::mutex> lk(m);
bool error = !com.redirectRead([&](std::vector<uint8_t>&& data) {
std::unique_lock<std::mutex> lk2(m);
if(error) {
lk2.unlock();
cv.notify_all();
return;
}
std::mutex m;
std::condition_variable cv;
uint16_t receiving = 0; // How much are we about to get before another header or completion
uint64_t received = 0;
uint16_t receivedCurrent = 0;
size_t skipping = 0;
std::vector<uint8_t> header;
std::unique_lock<std::mutex> lk(m);
bool error = !com.redirectRead([&](std::vector<uint8_t>&& data) {
std::unique_lock<std::mutex> lk2(m);
if(error) {
lk2.unlock();
cv.notify_all();
return;
}
if(skipping > data.size()) {
skipping -= data.size();
return;
}
size_t offset = skipping;
skipping = 0;
while(offset < data.size()) {
size_t left = data.size() - offset;
if(skipping > data.size()) {
skipping -= data.size();
return;
}
size_t offset = skipping;
skipping = 0;
while(offset < data.size()) {
size_t left = data.size() - offset;
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Going to process " << left << " bytes" << std::endl;
#endif
if(header.size() != HeaderLength) {
if(header.empty() && left && data[offset] != 0xaa) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Incorrect header " << int(data[offset]) << ' ' << int(offset) << std::endl;
#endif
error = true;
lk2.unlock();
cv.notify_all();
return;
}
// Did we get a correct header and at least one byte of data?
const auto begin = data.begin() + offset;
int32_t headerLeft = int32_t(HeaderLength - header.size());
if(int32_t(left) < headerLeft) {
// Not enough data here, grab what header we can and continue
header.insert(header.end(), begin, data.end());
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << int(left) << " bytes of header at " << offset << " (incomplete " <<
header.size() << ')' << std::endl;
#endif
return;
}
header.insert(header.end(), begin, begin + headerLeft);
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Going to process " << left << " bytes" << std::endl;
std::cout << "Got " << int(headerLeft) << " bytes of header at " << offset << " (complete " <<
header.size() << ')' << std::endl;
#endif
if(header.size() != HeaderLength) {
if(header.empty() && left && data[offset] != 0xaa) {
offset += headerLeft;
if(header[1] == uint8_t(Network::NetID::RED)) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got extended response " << int(offset) << std::endl;
#endif
// This is the extended command response, not all devices send this
// If we got it, we need to figure out how much more data to ignore
uint16_t length = (header[2] + (header[3] << 8));
// Try for another header after this, regardless how much we choose
// to skip and how we skip it
header.clear();
if(length <= 6) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Incorrect header " << int(data[offset]) << ' ' << int(offset) << std::endl;
std::cout << "Incorrect extended response length " << int(length) << ' ' << int(offset) << std::endl;
#endif
error = true;
lk2.unlock();
cv.notify_all();
return;
}
// Did we get a correct header and at least one byte of data?
const auto begin = data.begin() + offset;
int32_t headerLeft = int32_t(HeaderLength - header.size());
if(int32_t(left) < headerLeft) {
// Not enough data here, grab what header we can and continue
header.insert(header.end(), begin, data.end());
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << int(left) << " bytes of header at " << offset << " (incomplete " <<
header.size() << ')' << std::endl;
#endif
length -= 7;
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Skipping " << int(length) << ' ' << int(left) << std::endl;
#endif
if(left < length) {
skipping = length - left;
return;
}
header.insert(header.end(), begin, begin + headerLeft);
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << int(headerLeft) << " bytes of header at " << offset << " (complete " <<
header.size() << ')' << std::endl;
#endif
offset += headerLeft;
if(header[1] == uint8_t(Network::NetID::RED)) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got extended response " << int(offset) << std::endl;
#endif
// This is the extended command response, not all devices send this
// If we got it, we need to figure out how much more data to ignore
uint16_t length = (header[2] + (header[3] << 8));
// Try for another header after this, regardless how much we choose
// to skip and how we skip it
header.clear();
if(length <= 6) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Incorrect extended response length " << int(length) << ' ' << int(offset) << std::endl;
#endif
error = true;
lk2.unlock();
cv.notify_all();
return;
}
length -= 7;
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Skipping " << int(length) << ' ' << int(left) << std::endl;
#endif
if(left < length) {
skipping = length - left;
return;
}
offset += length;
continue;
}
// The device tells us how much it's sending us before the next header
receiving = (header[5] | (header[6] << 8));
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Started packet of size " << receiving << " bytes" << std::endl;
#endif
offset += length;
continue;
}
left = data.size() - offset;
auto count = uint16_t(std::min<uint64_t>(std::min<uint64_t>(receiving - receivedCurrent, left), amount - received));
// The device tells us how much it's sending us before the next header
receiving = (header[5] | (header[6] << 8));
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "With " << int(left) << " bytes " << int(offset) << std::endl;
std::cout << "Started packet of size " << receiving << " bytes" << std::endl;
#endif
memcpy(cache.data() + received, data.data() + offset, count);
received += count;
receivedCurrent += count;
offset += count;
if(amount == received) {
if(receivedCurrent % 2 == 0)
offset++;
header.clear(); // Now we will need another header
lk2.unlock();
cv.notify_all();
lk2.lock();
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Finished!" << std::endl;
#endif
}
else if(receivedCurrent == receiving) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << count << " bytes, " << receivedCurrent << " byte packet " << received <<
" complete of " << amount << std::endl;
#endif
if(receivedCurrent % 2 == 0)
offset++;
header.clear(); // Now we will need another header
receivedCurrent = 0;
} else {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << count << " bytes, incomplete (of " << receiving << " bytes)" << std::endl;
#endif
}
}
});
Lifetime clearRedirect([&com, &lk] { lk.unlock(); com.clearRedirectRead(); });
if(error)
return nullopt;
left = data.size() - offset;
auto count = uint16_t(std::min<uint64_t>(std::min<uint64_t>(receiving - receivedCurrent, left), amount - received));
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "With " << int(left) << " bytes " << int(offset) << std::endl;
#endif
memcpy(into + received, data.data() + offset, count);
received += count;
receivedCurrent += count;
offset += count;
error = !com.sendCommand(ExtendedCommand::Extract, {
uint8_t(sector & 0xff),
uint8_t((sector >> 8) & 0xff),
uint8_t((sector >> 16) & 0xff),
uint8_t((sector >> 24) & 0xff),
uint8_t((sector >> 32) & 0xff),
uint8_t((sector >> 40) & 0xff),
uint8_t((sector >> 48) & 0xff),
uint8_t((sector >> 56) & 0xff),
uint8_t(sectorCount & 0xff),
uint8_t((sectorCount >> 8) & 0xff),
uint8_t((sectorCount >> 16) & 0xff),
uint8_t((sectorCount >> 24) & 0xff),
});
if(error)
return nullopt;
if(amount == received) {
if(receivedCurrent % 2 == 0)
offset++;
header.clear(); // Now we will need another header
lk2.unlock();
cv.notify_all();
lk2.lock();
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Finished!" << std::endl;
#endif
}
else if(receivedCurrent == receiving) {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << count << " bytes, " << receivedCurrent << " byte packet " << received <<
" complete of " << amount << std::endl;
#endif
if(receivedCurrent % 2 == 0)
offset++;
header.clear(); // Now we will need another header
receivedCurrent = 0;
} else {
#ifdef ICSNEO_EXTENDED_EXTRACTOR_DEBUG_PRINTS
std::cout << "Got " << count << " bytes, incomplete (of " << receiving << " bytes)" << std::endl;
#endif
}
}
});
Lifetime clearRedirect([&com, &lk] { lk.unlock(); com.clearRedirectRead(); });
bool hitTimeout = !cv.wait_for(lk, timeout, [&]() { return error || amount == received; });
if(hitTimeout || error)
return nullopt;
if(error)
return nullopt;
cachedAt = std::chrono::steady_clock::now();
cachePos = pos;
}
error = !com.sendCommand(ExtendedCommand::Extract, {
uint8_t(sector & 0xff),
uint8_t((sector >> 8) & 0xff),
uint8_t((sector >> 16) & 0xff),
uint8_t((sector >> 24) & 0xff),
uint8_t((sector >> 32) & 0xff),
uint8_t((sector >> 40) & 0xff),
uint8_t((sector >> 48) & 0xff),
uint8_t((sector >> 56) & 0xff),
uint8_t(sectorCount & 0xff),
uint8_t((sectorCount >> 8) & 0xff),
uint8_t((sectorCount >> 16) & 0xff),
uint8_t((sectorCount >> 24) & 0xff),
});
if(error)
return nullopt;
bool hitTimeout = !cv.wait_for(lk, timeout, [&]() { return error || amount == received; });
if(hitTimeout || error)
return nullopt;
memcpy(into, cache.data(), size_t(amount));
return amount;
}
+21 -32
View File
@@ -15,38 +15,31 @@ optional<uint64_t> NeoMemoryDiskDriver::readLogicalDiskAligned(Communication& co
if(amount != SectorSize)
return nullopt;
if(cachePos != pos || std::chrono::steady_clock::now() > cachedAt + CacheTime) {
// The cache does not have this data, go get it
const uint64_t currentSector = pos / SectorSize;
auto msg = com.waitForMessageSync([&currentSector, &com] {
return com.sendCommand(Command::NeoReadMemory, {
MemoryTypeSD,
uint8_t(currentSector & 0xFF),
uint8_t((currentSector >> 8) & 0xFF),
uint8_t((currentSector >> 16) & 0xFF),
uint8_t((currentSector >> 24) & 0xFF),
uint8_t(SectorSize & 0xFF),
uint8_t((SectorSize >> 8) & 0xFF),
uint8_t((SectorSize >> 16) & 0xFF),
uint8_t((SectorSize >> 24) & 0xFF)
});
}, NeoMemorySDRead, timeout);
const uint64_t currentSector = pos / SectorSize;
auto msg = com.waitForMessageSync([&currentSector, &com] {
return com.sendCommand(Command::NeoReadMemory, {
MemoryTypeSD,
uint8_t(currentSector & 0xFF),
uint8_t((currentSector >> 8) & 0xFF),
uint8_t((currentSector >> 16) & 0xFF),
uint8_t((currentSector >> 24) & 0xFF),
uint8_t(SectorSize & 0xFF),
uint8_t((SectorSize >> 8) & 0xFF),
uint8_t((SectorSize >> 16) & 0xFF),
uint8_t((SectorSize >> 24) & 0xFF)
});
}, NeoMemorySDRead, timeout);
if(!msg)
return 0;
if(!msg)
return 0;
const auto sdmsg = std::dynamic_pointer_cast<NeoReadMemorySDMessage>(msg);
if(!sdmsg || sdmsg->data.size() != SectorSize) {
report(APIEvent::Type::PacketDecodingError, APIEvent::Severity::Error);
return nullopt;
}
memcpy(cache.data(), sdmsg->data.data(), SectorSize);
cachedAt = std::chrono::steady_clock::now();
cachePos = pos;
const auto sdmsg = std::dynamic_pointer_cast<NeoReadMemorySDMessage>(msg);
if(!sdmsg || sdmsg->data.size() != SectorSize) {
report(APIEvent::Type::PacketDecodingError, APIEvent::Severity::Error);
return nullopt;
}
memcpy(into, cache.data(), SectorSize);
memcpy(into, sdmsg->data.data(), SectorSize);
return SectorSize;
}
@@ -61,10 +54,6 @@ optional<uint64_t> NeoMemoryDiskDriver::writeLogicalDiskAligned(Communication& c
if(amount != SectorSize)
return nullopt;
// Clear the cache if we're writing to the cached sector
if(pos == cachePos)
cachedAt = std::chrono::time_point<std::chrono::steady_clock>();
// Requesting an atomic operation, but neoMemory does not support it
// Continue on anyway but warn the caller
if(atomicBuf != nullptr)
+37 -47
View File
@@ -19,60 +19,50 @@ optional<uint64_t> PlasionDiskReadDriver::readLogicalDiskAligned(Communication&
if(pos % getBlockSizeBounds().first != 0)
return nullopt;
if(cachePos != pos || std::chrono::steady_clock::now() > cachedAt + CacheTime) {
uint64_t largeSector = pos / SectorSize;
uint32_t sector = uint32_t(largeSector);
if (largeSector != uint64_t(sector))
return nullopt;
uint64_t largeSector = pos / SectorSize;
uint32_t sector = uint32_t(largeSector);
if (largeSector != uint64_t(sector))
return nullopt;
// The cache does not have this data, go get it
std::mutex m;
std::condition_variable cv;
uint32_t copied = 0;
bool error = false;
std::mutex m;
std::condition_variable cv;
uint32_t copied = 0;
bool error = false;
std::unique_lock<std::mutex> lk(m);
auto cb = com.addMessageCallback(MessageCallback([&](std::shared_ptr<Message> msg) {
std::unique_lock<std::mutex> lk(m);
auto cb = com.addMessageCallback(MessageCallback([&](std::shared_ptr<Message> msg) {
std::unique_lock<std::mutex> lk(m);
const auto sdmsg = std::dynamic_pointer_cast<NeoReadMemorySDMessage>(msg);
if(!sdmsg || cache.size() < copied + sdmsg->data.size()) {
error = true;
lk.unlock();
cv.notify_all();
return;
}
const auto sdmsg = std::dynamic_pointer_cast<NeoReadMemorySDMessage>(msg);
if(!sdmsg || amount < copied + sdmsg->data.size()) {
error = true;
lk.unlock();
cv.notify_all();
return;
}
// Invalidate the cache here in case we fail half-way through
cachedAt = std::chrono::steady_clock::time_point();
memcpy(into + copied, sdmsg->data.data(), sdmsg->data.size());
copied += uint32_t(sdmsg->data.size());
if(copied == amount) {
lk.unlock();
cv.notify_all();
}
}, NeoMemorySDRead));
memcpy(cache.data() + copied, sdmsg->data.data(), sdmsg->data.size());
copied += uint32_t(sdmsg->data.size());
if(copied == amount) {
lk.unlock();
cv.notify_all();
}
}, NeoMemorySDRead));
com.rawWrite({
uint8_t(MultiChannelCommunication::CommandType::HostPC_from_SDCC1),
uint8_t(sector & 0xFF),
uint8_t((sector >> 8) & 0xFF),
uint8_t((sector >> 16) & 0xFF),
uint8_t((sector >> 24) & 0xFF),
uint8_t(amount & 0xFF),
uint8_t((amount >> 8) & 0xFF),
});
com.rawWrite({
uint8_t(MultiChannelCommunication::CommandType::HostPC_from_SDCC1),
uint8_t(sector & 0xFF),
uint8_t((sector >> 8) & 0xFF),
uint8_t((sector >> 16) & 0xFF),
uint8_t((sector >> 24) & 0xFF),
uint8_t(amount & 0xFF),
uint8_t((amount >> 8) & 0xFF),
});
bool hitTimeout = !cv.wait_for(lk, timeout, [&copied, &error, &amount] { return error || copied == amount; });
com.removeMessageCallback(cb);
bool hitTimeout = !cv.wait_for(lk, timeout, [&copied, &error, &amount] { return error || copied == amount; });
com.removeMessageCallback(cb);
if(hitTimeout)
return nullopt;
if(hitTimeout)
return nullopt;
cachedAt = std::chrono::steady_clock::now();
cachePos = pos;
}
memcpy(into, cache.data(), size_t(amount));
return amount;
}