1 files changed, 42 insertions, 23 deletions

diff --git a/tools/gator/daemon/PerfBuffer.cpp b/tools/gator/daemon/PerfBuffer.cpp
index 5fad583f7bd0..f127c996d43b 100644
--- a/tools/gator/daemon/PerfBuffer.cpp
+++ b/tools/gator/daemon/PerfBuffer.cpp
@@ -20,6 +20,7 @@ PerfBuffer::PerfBuffer() {
 	for (int cpu = 0; cpu < ARRAY_LENGTH(mBuf); ++cpu) {
 		mBuf[cpu] = MAP_FAILED;
 		mDiscard[cpu] = false;
+		mFds[cpu] = -1;
 	}
 }
 
@@ -31,8 +32,8 @@ PerfBuffer::~PerfBuffer() {
 	}
 }
 
-bool PerfBuffer::useFd(const int cpu, const int fd, const int groupFd) {
-	if (fd == groupFd) {
+bool PerfBuffer::useFd(const int cpu, const int fd) {
+	if (mFds[cpu] < 0) {
 		if (mBuf[cpu] != MAP_FAILED) {
 			logg->logMessage("%s(%s:%i): cpu %i already online or not correctly cleaned up", __FUNCTION__, __FILE__, __LINE__, cpu);
 			return false;
@@ -44,6 +45,7 @@ bool PerfBuffer::useFd(const int cpu, const int fd, const int groupFd) {
 			logg->logMessage("%s(%s:%i): mmap failed", __FUNCTION__, __FILE__, __LINE__);
 			return false;
 		}
+		mFds[cpu] = fd;
 
 		// Check the version
 		struct perf_event_mmap_page *pemp = static_cast<struct perf_event_mmap_page *>(mBuf[cpu]);
@@ -57,7 +59,7 @@ bool PerfBuffer::useFd(const int cpu, const int fd, const int groupFd) {
 			return false;
 		}
 
-		if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, groupFd) < 0) {
+		if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, mFds[cpu]) < 0) {
 			logg->logMessage("%s(%s:%i): ioctl failed", __FUNCTION__, __FILE__, __LINE__);
 			return false;
 		}
@@ -89,6 +91,41 @@ bool PerfBuffer::isEmpty() {
 	return true;
 }
 
+static void compressAndSend(const int cpu, const __u64 head, __u64 tail, const uint8_t *const b, Sender *const sender) {
+	// Pick a big size but something smaller than the chunkSize in Sender::writeData which is 100k
+	char buf[1<<16];
+	int writePos = 0;
+	const int typeLength = gSessionData->mLocalCapture ? 0 : 1;
+
+	while (head > tail) {
+		writePos = 0;
+		if (!gSessionData->mLocalCapture) {
+			buf[writePos++] = RESPONSE_APC_DATA;
+		}
+		// Reserve space for size
+		writePos += sizeof(uint32_t);
+		Buffer::packInt(buf, sizeof(buf), writePos, FRAME_PERF);
+		Buffer::packInt(buf, sizeof(buf), writePos, cpu);
+
+		while (head > tail) {
+			const int count = reinterpret_cast<const struct perf_event_header *>(b + (tail & BUF_MASK))->size/sizeof(uint64_t);
+			// Can this whole message be written as Streamline assumes events are not split between frames
+			if (sizeof(buf) <= writePos + count*Buffer::MAXSIZE_PACK64) {
+				break;
+			}
+			for (int i = 0; i < count; ++i) {
+				// Must account for message size
+				Buffer::packInt64(buf, sizeof(buf), writePos, *reinterpret_cast<const uint64_t *>(b + (tail & BUF_MASK)));
+				tail += sizeof(uint64_t);
+			}
+		}
+
+		// Write size
+		Buffer::writeLEInt(reinterpret_cast<unsigned char *>(buf + typeLength), writePos - typeLength - sizeof(uint32_t));
+		sender->writeData(buf, writePos, RESPONSE_APC_DATA);
+	}
+}
+
 bool PerfBuffer::send(Sender *const sender) {
 	for (int cpu = 0; cpu < gSessionData->mCores; ++cpu) {
 		if (mBuf[cpu] == MAP_FAILED) {
@@ -102,26 +139,7 @@ bool PerfBuffer::send(Sender *const sender) {
 
 		if (head > tail) {
 			const uint8_t *const b = static_cast<uint8_t *>(mBuf[cpu]) + gSessionData->mPageSize;
-			const int offset = gSessionData->mLocalCapture ? 1 : 0;
-			unsigned char header[7];
-			header[0] = RESPONSE_APC_DATA;
-			Buffer::writeLEInt(header + 1, head - tail + sizeof(header) - 5);
-			// Should use real packing functions
-			header[5] = FRAME_PERF;
-			header[6] = cpu;
-
-			// Write header
-			sender->writeData(reinterpret_cast<const char *>(&header) + offset, sizeof(header) - offset, RESPONSE_APC_DATA);
-
-			// Write data
-			if ((head & ~BUF_MASK) == (tail & ~BUF_MASK)) {
-				// Not wrapped
-				sender->writeData(reinterpret_cast<const char *>(b + (tail & BUF_MASK)), head - tail, RESPONSE_APC_DATA);
-			} else {
-				// Wrapped
-				sender->writeData(reinterpret_cast<const char *>(b + (tail & BUF_MASK)), BUF_SIZE - (tail & BUF_MASK), RESPONSE_APC_DATA);
-				sender->writeData(reinterpret_cast<const char *>(b), head & BUF_MASK, RESPONSE_APC_DATA);
-			}
+			compressAndSend(cpu, head, tail, b, sender);
 
 			// Update tail with the data read
 			pemp->data_tail = head;
@@ -131,6 +149,7 @@ bool PerfBuffer::send(Sender *const sender) {
 			munmap(mBuf[cpu], gSessionData->mPageSize + BUF_SIZE);
 			mBuf[cpu] = MAP_FAILED;
 			mDiscard[cpu] = false;
+			mFds[cpu] = -1;
 			logg->logMessage("%s(%s:%i): Unmaped cpu %i", __FUNCTION__, __FILE__, __LINE__, cpu);
 		}
 	}