The per-PSB packet decoding logic was wrong because it was assuming that pt_insn_get_sync_offset was being udpated after every PSB. Silly me, that is not true. It returns the offset of the PSB packet after invoking pt_insn_sync_forward regardless of how many PSBs are visited later. Instead, I'm now following the approach described in https://github.com/intel/libipt/blob/master/doc/howto_libipt.md#parallel-decode for parallel decoding, which is basically what we need.
A nasty error that happened because of this is that when we had two PSBs (A and B), the following was happening
1. PSB A was processed all the way up to the end of the trace, which includes PSB B.
2. PSB B was then processed until the end of the trace.
The instructions emitted by step 2. were also emitted as part of step 1. so our trace had duplicated chunks. This problem becomes worse when you many PSBs.
As part of making sure this diff is correct, I added some other features that are very useful.
- Added a "synchronization point" event to the TraceCursor, so we can inspect when PSBs are emitted.
- Removed the single-thread decoder. Now the per-cpu decoder and single-thread decoder use the same code paths.
- Use the query decoder to fetch PSBs and timestamps. It turns out that the pt_insn_sync_forward of the instruction decoder can move past several PSBs (this means that we could skip some TSCs). On the other hand, the pt_query_sync_forward method doesn't skip PSBs, so we can get more accurate sync events and timing information.
- Turned LibiptDecoder into PSBBlockDecoder, which decodes single PSB blocks. It is the fundamental processing unit for decoding.
- Added many comments, asserts and improved error handling for clarity.
- Improved DecodeSystemWideTraceForThread so that a TSC is emitted always before a cpu change event. This was a bug that was annoying me before.
- SplitTraceInContinuousExecutions and FindLowestTSCInTrace are now using the query decoder, which can identify precisely each PSB along with their TSCs.
- Added an "only-events" option to the trace dumper to inspect only events.
I did extensive testing and I think we should have an in-house testing CI. The LLVM buildbots are not capable of supporting testing post-mortem traces of hundreds of megabytes. I'll leave that for later, but at least for now the current tests were able to catch most of the issues I encountered when doing this task.
A sample output of a program that I was single stepping is the following. You can see that only one PSB is emitted even though stepping happened!
```
thread #1: tid = 3578223
0: (event) trace synchronization point [offset = 0x0xef0]
a.out`main + 20 at main.cpp:29:20
1: 0x0000000000402479 leaq -0x1210(%rbp), %rax
2: (event) software disabled tracing
3: 0x0000000000402480 movq %rax, %rdi
4: (event) software disabled tracing
5: (event) software disabled tracing
6: 0x0000000000402483 callq 0x403bd4 ; std::vector<int, std::allocator<int>>::vector at stl_vector.h:391:7
7: (event) software disabled tracing
a.out`std::vector<int, std::allocator<int>>::vector() at stl_vector.h:391:7
8: 0x0000000000403bd4 pushq %rbp
9: (event) software disabled tracing
10: 0x0000000000403bd5 movq %rsp, %rbp
11: (event) software disabled tracing
```
This is another trace of a long program with a few PSBs.
```
(lldb) thread trace dump instructions -E -f thread #1: tid = 3603082
0: (event) trace synchronization point [offset = 0x0x80]
47417: (event) software disabled tracing
129231: (event) trace synchronization point [offset = 0x0x800]
146747: (event) software disabled tracing
246076: (event) software disabled tracing
259068: (event) trace synchronization point [offset = 0x0xf78]
259276: (event) software disabled tracing
259278: (event) software disabled tracing
no more data
```
Differential Revision: https://reviews.llvm.org/D131630
Pavel Labath taught me that clang-format sorts headers automatically
using llvm's rules, and it's better not to have spaces between
So in this diff I'm removing those spaces and formatting them as well.
I used `clang-format -i` to format these files.
Add bindings for the `TraceCursor` to allow for programatic traversal of
traces.
This diff adds bindings for all public `TraceCursor` methods except
`GetHwClock` and also adds `SBTrace::CreateNewCursor`. A new unittest
has been added to TestTraceLoad.py that uses the new `SBTraceCursor` API
to test that the sequential and random access APIs of the `TraceCursor`
are equivalent.
This diff depends on D130925.
Test Plan:
`ninja lldb-dotest && ./bin/lldb-dotest -p TestTraceLoad`
Differential Revision: https://reviews.llvm.org/D130930
Thanks to fredzhou@fb.com for coming up with this feature.
When tracing in per-cpu mode, we have information of in which cpu we are execution each instruction, which comes from the context switch trace. This diff makes this information available as a `cpu changed event`, which an additional accessor in the cursor `GetCPU()`. As cpu changes are very infrequent, any consumer should listen to cpu change events instead of querying the actual cpu of a trace item. Once a cpu change event is seen, the consumer can invoke GetCPU() to get that information. Also, it's possible to invoke GetCPU() on an arbitrary instruction item, which will return the last cpu seen. However, this call is O(logn) and should be used sparingly.
Manually tested with a sample program that starts on cpu 52, then goes to 18, and then goes back to 52.
Differential Revision: https://reviews.llvm.org/D129340
We want to include events with metadata, like context switches, and this
requires the API to handle events with payloads (e.g. information about
such context switches). Besides this, we want to support multiple
similar events between two consecutive instructions, like multiple
context switches. However, the current implementation is not good for this because
we are defining events as bitmask enums associated with specific
instructions. Thus, we need to decouple instructions from events and
make events actual items in the trace, just like instructions and
errors.
- Add accessors in the TraceCursor to know if an item is an event or not
- Modify from the TraceDumper all the way to DecodedThread to support
- Renamed the paused event to disabled.
- Improved the tsc handling logic. I was using an API for getting the tsc from libipt, but that was an overkill that should be used when not processing events manually, but as we are already processing events, we can more easily get the tscs.
event items. Fortunately this simplified many things
- As part of this refactor, I also fixed and long stating issue, which is that some non decoding errors were being inserted in the decoded thread. I changed this so that TraceIntelPT::Decode returns an error if the decoder couldn't be set up proplerly. Then, errors within a trace are actual anomalies found in between instrutions.
All test pass
Differential Revision: https://reviews.llvm.org/D128576
This is the final functional patch to support intel pt decoding per cpu.
It works by doing the following:
- First, all context switches are split by tid and sorted in order. This produces a list of continuous executes per thread per core.
- Then, all intel pt subtraces are split by PSB boundaries and assigned to individual thread continuous executions on the same core by doing simple TSC-based comparisons.
- With this, we have, per thread, a sorted list of continuous executions each one with a list of intel pt subtraces. Up to this point, this is really fast because no instructions were actually decoded.
- Then, each thread can be decoded by traversing their continuous executions and intel pt subtraces. An advantage of having these continuous executions is that we can identify if a continuous exexecution doesn't have intel pt data, and thus has a gap in it. We can later to more sofisticated comparisons to identify if within a continuous execution there are gaps.
I'm adding a test as well.
Differential Revision: https://reviews.llvm.org/D126394
A trace might contain events traced during the target's execution. For
example, a thread might be paused for some period of time due to context
switches or breakpoints, which actually force a context switch. Not only
that, a trace might be paused because the CPU decides to trace only a
specific part of the target, like the address filtering provided by
intel pt, which will cause pause events. Besides this case, other kinds
of events might exist.
This patch adds the method `TraceCursor::GetEvents()`` that returns the
list of events that happened right before the instruction being pointed
at by the cursor. Some refactors were done to make this change simpler.
Besides this new API, the instruction dumper now supports the -e flag
which shows pause events, like in the following example, where pauses
happened due to breakpoints.
```
thread #1: tid = 2717361
a.out`main + 20 at main.cpp:27:20
0: 0x00000000004023d9 leaq -0x1200(%rbp), %rax
[paused]
1: 0x00000000004023e0 movq %rax, %rdi
[paused]
2: 0x00000000004023e3 callq 0x403a62 ; std::vector<int, std::allocator<int> >::vector at stl_vector.h:391:7
a.out`std::vector<int, std::allocator<int> >::vector() at stl_vector.h:391:7
3: 0x0000000000403a62 pushq %rbp
4: 0x0000000000403a63 movq %rsp, %rbp
```
The `dump info` command has also been updated and now it shows the
number of instructions that have associated events.
Differential Revision: https://reviews.llvm.org/D123982
Something ugly I did was to report the trace buffer size to the DecodedThread,
which is later used as part of the `dump info` command. Instead of doing that,
we can just directly ask the trace for the raw buffer and print its size.
I thought about not asking for the entire trace but instead just for its size,
but in this case, as our traces as not extremely big, I prefer to ask for the
entire trace, ensuring it could be fetched, and then print its size.
Differential Revision: https://reviews.llvm.org/D123358
I'm adding two new classes that can be used to measure the duration of long
tasks as process and thread level, e.g. decoding, fetching data from
lldb-server, etc. In this first patch, I'm using it to measure the time it takes
to decode each thread, which is printed out with the `dump info` command. In a
later patch I'll start adding process-level tasks and I might move these
classes to the upper Trace level, instead of having them in the intel-pt
plugin. I might need to do that anyway in the future when we have to
measure HTR. For now, I want to keep the impact of this change minimal.
With it, I was able to generate the following info of a very big trace:
```
(lldb) thread trace dump info Trace technology: intel-pt
thread #1: tid = 616081
Total number of instructions: 9729366
Memory usage:
Raw trace size: 1024 KiB
Total approximate memory usage (excluding raw trace): 123517.34 KiB
Average memory usage per instruction (excluding raw trace): 13.00 bytes
Timing:
Decoding instructions: 1.62s
Errors:
Number of TSC decoding errors: 0
```
As seen above, it took 1.62 seconds to decode 9.7M instructions. This is great
news, as we don't need to do any optimization work in this area.
Differential Revision: https://reviews.llvm.org/D123357
In order to support quick arbitrary access to instructions in the trace, we need
each instruction to have an id. It could be an index or any other value that the
trace plugin defines.
This will be useful for reverse debugging or for creating callstacks, as each
frame will need an instruction id associated with them.
I've updated the `thread trace dump instructions` command accordingly. It now
prints the instruction id instead of relative offset. I've also added a new --id
argument that allows starting the dump from an arbitrary position.
Differential Revision: https://reviews.llvm.org/D122254
A problem that I introduced in the decoder is that I was considering TSC decoding
errors as actual instruction errors, which mean that the trace has a gap. This is
wrong because a TSC decoding error doesn't mean that there's a gap in the trace.
Instead, now I'm just counting how many of these errors happened and I'm using
the `dump info` command to check for this number.
Besides that, I refactored the decoder a little bit to make it simpler, more
readable, and to handle TSCs in a cleaner way.
Differential Revision: https://reviews.llvm.org/D122867
Storing timestamps (TSCs) in a more efficient map at the decoded thread level to speed up TSC lookup, as well as reduce the amount of memory used by each decoded instruction. Also introduced TSC range which keeps the current timestamp valid for all subsequent instructions until the next timestamp is emitted.
Differential Revision: https://reviews.llvm.org/D122603
Now the decoded thread has Append methods that provide more flexibility
in terms of the underlying data structure that represents the
instructions. In this case, we are able to represent the sporadic errors
as map and thus reduce the size of each instruction.
Differential Revision: https://reviews.llvm.org/D122293
When the user types that command 'thread trace dump info' and there's a running Trace session in LLDB, a raw trace in bytes should be printed; the command 'thread trace dump info all' should print the info for all the threads.
Original Author: hanbingwang
Reviewed By: clayborg, wallace
Differential Revision: https://reviews.llvm.org/D105717
D104422 added the interface for TraceCursor, which is the main way to traverse instructions in a trace. This diff implements the corresponding cursor class for Intel PT and deletes the now obsolete code.
Besides that, the logic for the "thread trace dump instructions" was adapted to use this cursor (pretty much I ended up moving code from Trace.cpp to TraceCursor.cpp). The command by default traverses the instructions backwards, and if the user passes --forwards, then it's not forwards. More information about that is in the Options.td file.
Regarding the Intel PT cursor. All Intel PT cursors for the same thread share the same DecodedThread instance. I'm not yet implementing lazy decoding because we don't need it. That'll be for later. For the time being, the entire thread trace is decoded when the first cursor for that thread is requested.
Differential Revision: https://reviews.llvm.org/D105531
This implements the interactive trace start and stop methods.
This diff ended up being much larger than I anticipated because, by doing it, I found that I had implemented in the beginning many things in a non optimal way. In any case, the code is much better now.
There's a lot of boilerplate code due to the gdb-remote protocol, but the main changes are:
- New tracing packets: jLLDBTraceStop, jLLDBTraceStart, jLLDBTraceGetBinaryData. The gdb-remote packet definitions are quite comprehensive.
- Implementation of the "process trace start|stop" and "thread trace start|stop" commands.
- Implementaiton of an API in Trace.h to interact with live traces.
- Created an IntelPTDecoder for live threads, that use the debugger's stop id as checkpoint for its internal cache.
- Added a functionality to stop the process in case "process tracing" is enabled and a new thread can't traced.
- Added tests
I have some ideas to unify the code paths for post mortem and live threads, but I'll do that in another diff.
Differential Revision: https://reviews.llvm.org/D91679
Depends on D89408.
This diff finally implements trace decoding!
The current interface is
$ trace load /path/to/trace/session/file.json
$ thread trace dump instructions
thread #1: tid = 3842849, total instructions = 22
[ 0] 0x40052d
[ 1] 0x40052d
...
[19] 0x400521
$ # simply enter, which is a repeat command
[20] 0x40052d
[21] 0x400529
...
This doesn't do any disassembly, which will be done in the next diff.
Changes:
- Added an IntelPTDecoder class, that is a wrapper for libipt, which is the actual library that performs the decoding.
- Added TraceThreadDecoder class that decodes traces and memoizes the result to avoid repeating the decoding step.
- Added a DecodedThread class, which represents the output from decoding and that for the time being only stores the list of reconstructed instructions. Later it'll contain the function call hierarchy, which will enable reconstructing backtraces.
- Added basic APIs for accessing the trace in Trace.h:
- GetInstructionCount, which counts the number of instructions traced for a given thread
- IsTraceFailed, which returns an Error if decoding a thread failed
- ForEachInstruction, which iterates on the instructions traced for a given thread, concealing the internal storage of threads, as plug-ins can decide to generate the instructions on the fly or to store them all in a vector, like I do.
- DumpTraceInstructions was updated to print the instructions or show an error message if decoding was impossible.
- Tests included
Differential Revision: https://reviews.llvm.org/D89283
Walter Erquinigo