1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-22 20:43:44 +02:00
llvm-mirror/lib/XRay/Trace.cpp

717 lines
29 KiB
C++
Raw Normal View History

//===- Trace.cpp - XRay Trace Loading implementation. ---------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// XRay log reader implementation.
//
//===----------------------------------------------------------------------===//
#include "llvm/XRay/Trace.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/DataExtractor.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/XRay/YAMLXRayRecord.h"
using namespace llvm;
using namespace llvm::xray;
using llvm::yaml::Input;
namespace {
using XRayRecordStorage =
std::aligned_storage<sizeof(XRayRecord), alignof(XRayRecord)>::type;
// Populates the FileHeader reference by reading the first 32 bytes of the file.
Error readBinaryFormatHeader(StringRef Data, XRayFileHeader &FileHeader) {
// FIXME: Maybe deduce whether the data is little or big-endian using some
// magic bytes in the beginning of the file?
// First 32 bytes of the file will always be the header. We assume a certain
// format here:
//
// (2) uint16 : version
// (2) uint16 : type
// (4) uint32 : bitfield
// (8) uint64 : cycle frequency
// (16) - : padding
DataExtractor HeaderExtractor(Data, true, 8);
uint32_t OffsetPtr = 0;
FileHeader.Version = HeaderExtractor.getU16(&OffsetPtr);
FileHeader.Type = HeaderExtractor.getU16(&OffsetPtr);
uint32_t Bitfield = HeaderExtractor.getU32(&OffsetPtr);
FileHeader.ConstantTSC = Bitfield & 1uL;
FileHeader.NonstopTSC = Bitfield & 1uL << 1;
FileHeader.CycleFrequency = HeaderExtractor.getU64(&OffsetPtr);
std::memcpy(&FileHeader.FreeFormData, Data.bytes_begin() + OffsetPtr, 16);
if (FileHeader.Version != 1 && FileHeader.Version != 2)
return make_error<StringError>(
Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version),
std::make_error_code(std::errc::invalid_argument));
return Error::success();
}
Error loadNaiveFormatLog(StringRef Data, XRayFileHeader &FileHeader,
std::vector<XRayRecord> &Records) {
if (Data.size() < 32)
return make_error<StringError>(
"Not enough bytes for an XRay log.",
std::make_error_code(std::errc::invalid_argument));
if (Data.size() - 32 == 0 || Data.size() % 32 != 0)
return make_error<StringError>(
"Invalid-sized XRay data.",
std::make_error_code(std::errc::invalid_argument));
if (auto E = readBinaryFormatHeader(Data, FileHeader))
return E;
// Each record after the header will be 32 bytes, in the following format:
//
// (2) uint16 : record type
// (1) uint8 : cpu id
// (1) uint8 : type
// (4) sint32 : function id
// (8) uint64 : tsc
// (4) uint32 : thread id
// (12) - : padding
for (auto S = Data.drop_front(32); !S.empty(); S = S.drop_front(32)) {
DataExtractor RecordExtractor(S, true, 8);
uint32_t OffsetPtr = 0;
switch (auto RecordType = RecordExtractor.getU16(&OffsetPtr)) {
case 0: { // Normal records.
Records.emplace_back();
auto &Record = Records.back();
Record.RecordType = RecordType;
Record.CPU = RecordExtractor.getU8(&OffsetPtr);
auto Type = RecordExtractor.getU8(&OffsetPtr);
switch (Type) {
case 0:
Record.Type = RecordTypes::ENTER;
break;
case 1:
Record.Type = RecordTypes::EXIT;
break;
case 2:
Record.Type = RecordTypes::TAIL_EXIT;
break;
case 3:
Record.Type = RecordTypes::ENTER_ARG;
break;
default:
return make_error<StringError>(
Twine("Unknown record type '") + Twine(int{Type}) + "'",
std::make_error_code(std::errc::executable_format_error));
}
Record.FuncId = RecordExtractor.getSigned(&OffsetPtr, sizeof(int32_t));
Record.TSC = RecordExtractor.getU64(&OffsetPtr);
Record.TId = RecordExtractor.getU32(&OffsetPtr);
break;
}
case 1: { // Arg payload record.
auto &Record = Records.back();
// Advance two bytes to avoid padding.
OffsetPtr += 2;
int32_t FuncId = RecordExtractor.getSigned(&OffsetPtr, sizeof(int32_t));
auto TId = RecordExtractor.getU32(&OffsetPtr);
if (Record.FuncId != FuncId || Record.TId != TId)
return make_error<StringError>(
Twine("Corrupted log, found arg payload following non-matching "
"function + thread record. Record for function ") +
Twine(Record.FuncId) + " != " + Twine(FuncId) + "; offset: " +
Twine(S.data() - Data.data()),
std::make_error_code(std::errc::executable_format_error));
// Advance another four bytes to avoid padding.
OffsetPtr += 4;
auto Arg = RecordExtractor.getU64(&OffsetPtr);
Record.CallArgs.push_back(Arg);
break;
}
default:
return make_error<StringError>(
Twine("Unknown record type == ") + Twine(RecordType),
std::make_error_code(std::errc::executable_format_error));
}
}
return Error::success();
}
/// When reading from a Flight Data Recorder mode log, metadata records are
/// sparse compared to packed function records, so we must maintain state as we
/// read through the sequence of entries. This allows the reader to denormalize
/// the CPUId and Thread Id onto each Function Record and transform delta
/// encoded TSC values into absolute encodings on each record.
struct FDRState {
uint16_t CPUId;
uint16_t ThreadId;
uint64_t BaseTSC;
/// Encode some of the state transitions for the FDR log reader as explicit
/// checks. These are expectations for the next Record in the stream.
enum class Token {
NEW_BUFFER_RECORD_OR_EOF,
WALLCLOCK_RECORD,
NEW_CPU_ID_RECORD,
FUNCTION_SEQUENCE,
SCAN_TO_END_OF_THREAD_BUF,
CUSTOM_EVENT_DATA,
CALL_ARGUMENT,
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
BUFFER_EXTENTS,
};
Token Expects;
// Each threads buffer may have trailing garbage to scan over, so we track our
// progress.
uint64_t CurrentBufferSize;
uint64_t CurrentBufferConsumed;
};
const char *fdrStateToTwine(const FDRState::Token &state) {
switch (state) {
case FDRState::Token::NEW_BUFFER_RECORD_OR_EOF:
return "NEW_BUFFER_RECORD_OR_EOF";
case FDRState::Token::WALLCLOCK_RECORD:
return "WALLCLOCK_RECORD";
case FDRState::Token::NEW_CPU_ID_RECORD:
return "NEW_CPU_ID_RECORD";
case FDRState::Token::FUNCTION_SEQUENCE:
return "FUNCTION_SEQUENCE";
case FDRState::Token::SCAN_TO_END_OF_THREAD_BUF:
return "SCAN_TO_END_OF_THREAD_BUF";
case FDRState::Token::CUSTOM_EVENT_DATA:
return "CUSTOM_EVENT_DATA";
case FDRState::Token::CALL_ARGUMENT:
return "CALL_ARGUMENT";
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
case FDRState::Token::BUFFER_EXTENTS:
return "BUFFER_EXTENTS";
}
return "UNKNOWN";
}
/// State transition when a NewBufferRecord is encountered.
Error processFDRNewBufferRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects != FDRState::Token::NEW_BUFFER_RECORD_OR_EOF)
return make_error<StringError>(
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
Twine("Malformed log. Read New Buffer record kind out of sequence; "
"expected: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
uint32_t OffsetPtr = 1; // 1 byte into record.
State.ThreadId = RecordExtractor.getU16(&OffsetPtr);
State.Expects = FDRState::Token::WALLCLOCK_RECORD;
return Error::success();
}
/// State transition when an EndOfBufferRecord is encountered.
Error processFDREndOfBufferRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects == FDRState::Token::NEW_BUFFER_RECORD_OR_EOF)
return make_error<StringError>(
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
Twine("Malformed log. Received EOB message without current buffer; "
"expected: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
State.Expects = FDRState::Token::SCAN_TO_END_OF_THREAD_BUF;
return Error::success();
}
/// State transition when a NewCPUIdRecord is encountered.
Error processFDRNewCPUIdRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects != FDRState::Token::FUNCTION_SEQUENCE &&
State.Expects != FDRState::Token::NEW_CPU_ID_RECORD)
return make_error<StringError>(
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
Twine("Malformed log. Read NewCPUId record kind out of sequence; "
"expected: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
uint32_t OffsetPtr = 1; // Read starting after the first byte.
State.CPUId = RecordExtractor.getU16(&OffsetPtr);
State.BaseTSC = RecordExtractor.getU64(&OffsetPtr);
State.Expects = FDRState::Token::FUNCTION_SEQUENCE;
return Error::success();
}
/// State transition when a TSCWrapRecord (overflow detection) is encountered.
Error processFDRTSCWrapRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects != FDRState::Token::FUNCTION_SEQUENCE)
return make_error<StringError>(
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
Twine("Malformed log. Read TSCWrap record kind out of sequence; "
"expecting: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
uint32_t OffsetPtr = 1; // Read starting after the first byte.
State.BaseTSC = RecordExtractor.getU64(&OffsetPtr);
return Error::success();
}
/// State transition when a WallTimeMarkerRecord is encountered.
Error processFDRWallTimeRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects != FDRState::Token::WALLCLOCK_RECORD)
return make_error<StringError>(
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
Twine("Malformed log. Read Wallclock record kind out of sequence; "
"expecting: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
// TODO: Someday, reconcile the TSC ticks to wall clock time for presentation
// purposes. For now, we're ignoring these records.
State.Expects = FDRState::Token::NEW_CPU_ID_RECORD;
return Error::success();
}
/// State transition when a CustomEventMarker is encountered.
Error processCustomEventMarker(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor,
size_t &RecordSize) {
// We can encounter a CustomEventMarker anywhere in the log, so we can handle
// it regardless of the expectation. However, we do set the expectation to
// read a set number of fixed bytes, as described in the metadata.
uint32_t OffsetPtr = 1; // Read after the first byte.
uint32_t DataSize = RecordExtractor.getU32(&OffsetPtr);
uint64_t TSC = RecordExtractor.getU64(&OffsetPtr);
// FIXME: Actually represent the record through the API. For now we only
// skip through the data.
(void)TSC;
RecordSize = 16 + DataSize;
return Error::success();
}
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
/// State transition when an BufferExtents record is encountered.
Error processBufferExtents(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor) {
if (State.Expects != FDRState::Token::BUFFER_EXTENTS)
return make_error<StringError>(
Twine("Malformed log. Buffer Extents unexpected; expected: ") +
fdrStateToTwine(State.Expects),
std::make_error_code(std::errc::executable_format_error));
uint32_t OffsetPtr = 1; // Read after the first byte.
State.CurrentBufferSize = RecordExtractor.getU64(&OffsetPtr);
State.Expects = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF;
return Error::success();
}
/// State transition when a CallArgumentRecord is encountered.
Error processFDRCallArgumentRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor,
std::vector<XRayRecord> &Records) {
uint32_t OffsetPtr = 1; // Read starting after the first byte.
auto &Enter = Records.back();
if (Enter.Type != RecordTypes::ENTER)
return make_error<StringError>(
"CallArgument needs to be right after a function entry",
std::make_error_code(std::errc::executable_format_error));
Enter.Type = RecordTypes::ENTER_ARG;
Enter.CallArgs.emplace_back(RecordExtractor.getU64(&OffsetPtr));
return Error::success();
}
/// Advances the state machine for reading the FDR record type by reading one
/// Metadata Record and updating the State appropriately based on the kind of
/// record encountered. The RecordKind is encoded in the first byte of the
/// Record, which the caller should pass in because they have already read it
/// to determine that this is a metadata record as opposed to a function record.
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
///
/// Beginning with Version 2 of the FDR log, we do not depend on the size of the
/// buffer, but rather use the extents to determine how far to read in the log
/// for this particular buffer.
Error processFDRMetadataRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor,
size_t &RecordSize,
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
std::vector<XRayRecord> &Records,
uint16_t Version) {
// The remaining 7 bits are the RecordKind enum.
uint8_t RecordKind = RecordFirstByte >> 1;
switch (RecordKind) {
case 0: // NewBuffer
if (auto E =
processFDRNewBufferRecord(State, RecordFirstByte, RecordExtractor))
return E;
break;
case 1: // EndOfBuffer
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
if (Version >= 2)
return make_error<StringError>(
"Since Version 2 of FDR logging, we no longer support EOB records.",
std::make_error_code(std::errc::executable_format_error));
if (auto E = processFDREndOfBufferRecord(State, RecordFirstByte,
RecordExtractor))
return E;
break;
case 2: // NewCPUId
if (auto E =
processFDRNewCPUIdRecord(State, RecordFirstByte, RecordExtractor))
return E;
break;
case 3: // TSCWrap
if (auto E =
processFDRTSCWrapRecord(State, RecordFirstByte, RecordExtractor))
return E;
break;
case 4: // WallTimeMarker
if (auto E =
processFDRWallTimeRecord(State, RecordFirstByte, RecordExtractor))
return E;
break;
case 5: // CustomEventMarker
if (auto E = processCustomEventMarker(State, RecordFirstByte,
RecordExtractor, RecordSize))
return E;
break;
case 6: // CallArgument
if (auto E = processFDRCallArgumentRecord(State, RecordFirstByte,
RecordExtractor, Records))
return E;
break;
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
case 7: // BufferExtents
if (auto E = processBufferExtents(State, RecordFirstByte, RecordExtractor))
return E;
break;
default:
// Widen the record type to uint16_t to prevent conversion to char.
return make_error<StringError>(
Twine("Illegal metadata record type: ")
.concat(Twine(static_cast<unsigned>(RecordKind))),
std::make_error_code(std::errc::executable_format_error));
}
return Error::success();
}
/// Reads a function record from an FDR format log, appending a new XRayRecord
/// to the vector being populated and updating the State with a new value
/// reference value to interpret TSC deltas.
///
/// The XRayRecord constructed includes information from the function record
/// processed here as well as Thread ID and CPU ID formerly extracted into
/// State.
Error processFDRFunctionRecord(FDRState &State, uint8_t RecordFirstByte,
DataExtractor &RecordExtractor,
std::vector<XRayRecord> &Records) {
switch (State.Expects) {
case FDRState::Token::NEW_BUFFER_RECORD_OR_EOF:
return make_error<StringError>(
"Malformed log. Received Function Record before new buffer setup.",
std::make_error_code(std::errc::executable_format_error));
case FDRState::Token::WALLCLOCK_RECORD:
return make_error<StringError>(
"Malformed log. Received Function Record when expecting wallclock.",
std::make_error_code(std::errc::executable_format_error));
case FDRState::Token::NEW_CPU_ID_RECORD:
return make_error<StringError>(
"Malformed log. Received Function Record before first CPU record.",
std::make_error_code(std::errc::executable_format_error));
default:
Records.emplace_back();
auto &Record = Records.back();
Record.RecordType = 0; // Record is type NORMAL.
// Strip off record type bit and use the next three bits.
uint8_t RecordType = (RecordFirstByte >> 1) & 0x07;
switch (RecordType) {
case static_cast<uint8_t>(RecordTypes::ENTER):
Record.Type = RecordTypes::ENTER;
break;
case static_cast<uint8_t>(RecordTypes::EXIT):
Record.Type = RecordTypes::EXIT;
break;
case static_cast<uint8_t>(RecordTypes::TAIL_EXIT):
Record.Type = RecordTypes::TAIL_EXIT;
break;
default:
// Cast to an unsigned integer to not interpret the record type as a char.
return make_error<StringError>(
Twine("Illegal function record type: ")
.concat(Twine(static_cast<unsigned>(RecordType))),
std::make_error_code(std::errc::executable_format_error));
}
Record.CPU = State.CPUId;
Record.TId = State.ThreadId;
// Back up to read first 32 bits, including the 4 we pulled RecordType
// and RecordKind out of. The remaining 28 are FunctionId.
uint32_t OffsetPtr = 0;
// Despite function Id being a signed int on XRayRecord,
// when it is written to an FDR format, the top bits are truncated,
// so it is effectively an unsigned value. When we shift off the
// top four bits, we want the shift to be logical, so we read as
// uint32_t.
uint32_t FuncIdBitField = RecordExtractor.getU32(&OffsetPtr);
Record.FuncId = FuncIdBitField >> 4;
// FunctionRecords have a 32 bit delta from the previous absolute TSC
// or TSC delta. If this would overflow, we should read a TSCWrap record
// with an absolute TSC reading.
uint64_t NewTSC = State.BaseTSC + RecordExtractor.getU32(&OffsetPtr);
State.BaseTSC = NewTSC;
Record.TSC = NewTSC;
}
return Error::success();
}
/// Reads a log in FDR mode for version 1 of this binary format. FDR mode is
/// defined as part of the compiler-rt project in xray_fdr_logging.h, and such
/// a log consists of the familiar 32 bit XRayHeader, followed by sequences of
/// of interspersed 16 byte Metadata Records and 8 byte Function Records.
///
/// The following is an attempt to document the grammar of the format, which is
/// parsed by this function for little-endian machines. Since the format makes
/// use of BitFields, when we support big-endian architectures, we will need to
/// adjust not only the endianness parameter to llvm's RecordExtractor, but also
/// the bit twiddling logic, which is consistent with the little-endian
/// convention that BitFields within a struct will first be packed into the
/// least significant bits the address they belong to.
///
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
/// We expect a format complying with the grammar in the following pseudo-EBNF
/// in Version 1 of the FDR log.
///
/// FDRLog: XRayFileHeader ThreadBuffer*
/// XRayFileHeader: 32 bytes to identify the log as FDR with machine metadata.
/// Includes BufferSize
/// ThreadBuffer: NewBuffer WallClockTime NewCPUId FunctionSequence EOB
/// BufSize: 8 byte unsigned integer indicating how large the buffer is.
/// NewBuffer: 16 byte metadata record with Thread Id.
/// WallClockTime: 16 byte metadata record with human readable time.
/// NewCPUId: 16 byte metadata record with CPUId and a 64 bit TSC reading.
/// EOB: 16 byte record in a thread buffer plus mem garbage to fill BufSize.
/// FunctionSequence: NewCPUId | TSCWrap | FunctionRecord
/// TSCWrap: 16 byte metadata record with a full 64 bit TSC reading.
/// FunctionRecord: 8 byte record with FunctionId, entry/exit, and TSC delta.
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
///
/// In Version 2, we make the following changes:
///
/// ThreadBuffer: BufferExtents NewBuffer WallClockTime NewCPUId
/// FunctionSequence
/// BufferExtents: 16 byte metdata record describing how many usable bytes are
/// in the buffer. This is measured from the start of the buffer
/// and must always be at least 48 (bytes).
/// EOB: *deprecated*
Error loadFDRLog(StringRef Data, XRayFileHeader &FileHeader,
std::vector<XRayRecord> &Records) {
if (Data.size() < 32)
return make_error<StringError>(
"Not enough bytes for an XRay log.",
std::make_error_code(std::errc::invalid_argument));
// For an FDR log, there are records sized 16 and 8 bytes.
// There actually may be no records if no non-trivial functions are
// instrumented.
if (Data.size() % 8 != 0)
return make_error<StringError>(
"Invalid-sized XRay data.",
std::make_error_code(std::errc::invalid_argument));
if (auto E = readBinaryFormatHeader(Data, FileHeader))
return E;
uint64_t BufferSize = 0;
{
StringRef ExtraDataRef(FileHeader.FreeFormData, 16);
DataExtractor ExtraDataExtractor(ExtraDataRef, true, 8);
uint32_t ExtraDataOffset = 0;
BufferSize = ExtraDataExtractor.getU64(&ExtraDataOffset);
}
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
FDRState::Token InitialExpectation;
switch (FileHeader.Version) {
case 1:
InitialExpectation = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF;
break;
case 2:
InitialExpectation = FDRState::Token::BUFFER_EXTENTS;
break;
default:
return make_error<StringError>(
Twine("Unsupported version '") + Twine(FileHeader.Version) + "'",
std::make_error_code(std::errc::executable_format_error));
}
FDRState State{0, 0, 0, InitialExpectation, BufferSize, 0};
// RecordSize will tell the loop how far to seek ahead based on the record
// type that we have just read.
size_t RecordSize = 0;
for (auto S = Data.drop_front(32); !S.empty(); S = S.drop_front(RecordSize)) {
DataExtractor RecordExtractor(S, true, 8);
uint32_t OffsetPtr = 0;
if (State.Expects == FDRState::Token::SCAN_TO_END_OF_THREAD_BUF) {
RecordSize = State.CurrentBufferSize - State.CurrentBufferConsumed;
if (S.size() < RecordSize) {
return make_error<StringError>(
Twine("Incomplete thread buffer. Expected at least ") +
Twine(RecordSize) + " bytes but found " + Twine(S.size()),
make_error_code(std::errc::invalid_argument));
}
State.CurrentBufferConsumed = 0;
State.Expects = FDRState::Token::NEW_BUFFER_RECORD_OR_EOF;
continue;
}
uint8_t BitField = RecordExtractor.getU8(&OffsetPtr);
bool isMetadataRecord = BitField & 0x01uL;
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
bool isBufferExtents =
(BitField >> 1) == 7; // BufferExtents record kind == 7
if (isMetadataRecord) {
RecordSize = 16;
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
if (auto E =
processFDRMetadataRecord(State, BitField, RecordExtractor,
RecordSize, Records, FileHeader.Version))
return E;
} else { // Process Function Record
RecordSize = 8;
if (auto E = processFDRFunctionRecord(State, BitField, RecordExtractor,
Records))
return E;
}
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
// The BufferExtents record is technically not part of the buffer, so we
// don't count the size of that record against the buffer's actual size.
if (!isBufferExtents)
State.CurrentBufferConsumed += RecordSize;
assert(State.CurrentBufferConsumed <= State.CurrentBufferSize);
if (FileHeader.Version == 2 &&
State.CurrentBufferSize == State.CurrentBufferConsumed) {
// In Version 2 of the log, we don't need to scan to the end of the thread
// buffer if we've already consumed all the bytes we need to.
State.Expects = FDRState::Token::BUFFER_EXTENTS;
State.CurrentBufferSize = BufferSize;
State.CurrentBufferConsumed = 0;
}
}
// Having iterated over everything we've been given, we've either consumed
// everything and ended up in the end state, or were told to skip the rest.
bool Finished = State.Expects == FDRState::Token::SCAN_TO_END_OF_THREAD_BUF &&
State.CurrentBufferSize == State.CurrentBufferConsumed;
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
if ((State.Expects != FDRState::Token::NEW_BUFFER_RECORD_OR_EOF &&
State.Expects != FDRState::Token::BUFFER_EXTENTS) &&
!Finished)
return make_error<StringError>(
Twine("Encountered EOF with unexpected state expectation ") +
fdrStateToTwine(State.Expects) +
". Remaining expected bytes in thread buffer total " +
Twine(State.CurrentBufferSize - State.CurrentBufferConsumed),
std::make_error_code(std::errc::executable_format_error));
return Error::success();
}
Error loadYAMLLog(StringRef Data, XRayFileHeader &FileHeader,
std::vector<XRayRecord> &Records) {
YAMLXRayTrace Trace;
Input In(Data);
In >> Trace;
if (In.error())
return make_error<StringError>("Failed loading YAML Data.", In.error());
FileHeader.Version = Trace.Header.Version;
FileHeader.Type = Trace.Header.Type;
FileHeader.ConstantTSC = Trace.Header.ConstantTSC;
FileHeader.NonstopTSC = Trace.Header.NonstopTSC;
FileHeader.CycleFrequency = Trace.Header.CycleFrequency;
if (FileHeader.Version != 1)
return make_error<StringError>(
Twine("Unsupported XRay file version: ") + Twine(FileHeader.Version),
std::make_error_code(std::errc::invalid_argument));
Records.clear();
std::transform(Trace.Records.begin(), Trace.Records.end(),
std::back_inserter(Records), [&](const YAMLXRayRecord &R) {
return XRayRecord{R.RecordType, R.CPU, R.Type, R.FuncId,
R.TSC, R.TId, R.CallArgs};
});
return Error::success();
}
} // namespace
Expected<Trace> llvm::xray::loadTraceFile(StringRef Filename, bool Sort) {
int Fd;
if (auto EC = sys::fs::openFileForRead(Filename, Fd)) {
return make_error<StringError>(
Twine("Cannot read log from '") + Filename + "'", EC);
}
uint64_t FileSize;
if (auto EC = sys::fs::file_size(Filename, FileSize)) {
return make_error<StringError>(
Twine("Cannot read log from '") + Filename + "'", EC);
}
if (FileSize < 4) {
return make_error<StringError>(
Twine("File '") + Filename + "' too small for XRay.",
std::make_error_code(std::errc::executable_format_error));
}
// Map the opened file into memory and use a StringRef to access it later.
std::error_code EC;
sys::fs::mapped_file_region MappedFile(
Fd, sys::fs::mapped_file_region::mapmode::readonly, FileSize, 0, EC);
if (EC) {
return make_error<StringError>(
Twine("Cannot read log from '") + Filename + "'", EC);
}
auto Data = StringRef(MappedFile.data(), MappedFile.size());
// Attempt to detect the file type using file magic. We have a slight bias
// towards the binary format, and we do this by making sure that the first 4
// bytes of the binary file is some combination of the following byte
// patterns: (observe the code loading them assumes they're little endian)
//
// 0x01 0x00 0x00 0x00 - version 1, "naive" format
// 0x01 0x00 0x01 0x00 - version 1, "flight data recorder" format
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
// 0x02 0x00 0x01 0x00 - version 2, "flight data recorder" format
//
// YAML files don't typically have those first four bytes as valid text so we
// try loading assuming YAML if we don't find these bytes.
//
// Only if we can't load either the binary or the YAML format will we yield an
// error.
StringRef Magic(MappedFile.data(), 4);
DataExtractor HeaderExtractor(Magic, true, 8);
uint32_t OffsetPtr = 0;
uint16_t Version = HeaderExtractor.getU16(&OffsetPtr);
uint16_t Type = HeaderExtractor.getU16(&OffsetPtr);
enum BinaryFormatType { NAIVE_FORMAT = 0, FLIGHT_DATA_RECORDER_FORMAT = 1 };
Trace T;
[XRay] Use optimistic logging model for FDR mode Summary: Before this change, the FDR mode implementation relied on at thread-exit handling to return buffers back to the (global) buffer queue. This introduces issues with the initialisation of the thread_local objects which, even through the use of pthread_setspecific(...) may eventually call into an allocation function. Similar to previous changes in this line, we're finding that there is a huge potential for deadlocks when initialising these thread-locals when the memory allocation implementation is also xray-instrumented. In this change, we limit the call to pthread_setspecific(...) to provide a non-null value to associate to the key created with pthread_key_create(...). While this doesn't completely eliminate the potential for the deadlock(s), it does allow us to still clean up at thread exit when we need to. The change is that we don't need to do more work when starting and ending a thread's lifetime. We also have a test to make sure that we actually can safely recycle the buffers in case we end up re-using the buffer(s) available from the queue on multiple thread entry/exits. This change cuts across both LLVM and compiler-rt to allow us to update both the XRay runtime implementation as well as the library support for loading these new versions of the FDR mode logging. Version 2 of the FDR logging implementation makes the following changes: * Introduction of a new 'BufferExtents' metadata record that's outside of the buffer's contents but are written before the actual buffer. This data is associated to the Buffer handed out by the BufferQueue rather than a record that occupies bytes in the actual buffer. * Removal of the "end of buffer" records. This is in-line with the changes we described above, to allow for optimistic logging without explicit record writing at thread exit. The optimistic logging model operates under the following assumptions: * Threads writing to the buffers will potentially race with the thread attempting to flush the log. To avoid this situation from occuring, we make sure that when we've finalized the logging implementation, that threads will see this finalization state on the next write, and either choose to not write records the thread would have written or write the record(s) in two phases -- first write the record(s), then update the extents metadata. * We change the buffer queue implementation so that once it's handed out a buffer to a thread, that we assume that buffer is marked "used" to be able to capture partial writes. None of this will be safe to handle if threads are racing to write the extents records and the reader thread is attempting to flush the log. The optimism comes from the finalization routine being required to complete before we attempt to flush the log. This is a fairly significant semantics change for the FDR implementation. This is why we've decided to update the version number for FDR mode logs. The tools, however, still need to be able to support older versions of the log until we finally deprecate those earlier versions. Reviewers: dblaikie, pelikan, kpw Subscribers: llvm-commits, hiraditya Differential Revision: https://reviews.llvm.org/D39526 llvm-svn: 318733
2017-11-21 08:16:57 +01:00
switch (Type) {
case NAIVE_FORMAT:
if (Version == 1 || Version == 2) {
if (auto E = loadNaiveFormatLog(Data, T.FileHeader, T.Records))
return std::move(E);
} else {
return make_error<StringError>(
Twine("Unsupported version for Basic/Naive Mode logging: ") +
Twine(Version),
std::make_error_code(std::errc::executable_format_error));
}
break;
case FLIGHT_DATA_RECORDER_FORMAT:
if (Version == 1 || Version == 2) {
if (auto E = loadFDRLog(Data, T.FileHeader, T.Records))
return std::move(E);
} else {
return make_error<StringError>(
Twine("Unsupported version for FDR Mode logging: ") + Twine(Version),
std::make_error_code(std::errc::executable_format_error));
}
break;
default:
if (auto E = loadYAMLLog(Data, T.FileHeader, T.Records))
return std::move(E);
}
if (Sort)
std::stable_sort(T.Records.begin(), T.Records.end(),
[&](const XRayRecord &L, const XRayRecord &R) {
return L.TSC < R.TSC;
});
return std::move(T);
}