Can't use buffers with graphs because copy back to host fails #54
Labels
Comments
Bensuo
added
bug
reble
pushed a commit
that referenced
this issue
May 15, 2023
…callback
The `TypeSystemMap::m_mutex` guards against concurrent modifications
of members of `TypeSystemMap`. In particular, `m_map`.
`TypeSystemMap::ForEach` iterates through the entire `m_map` calling
a user-specified callback for each entry. This is all done while
`m_mutex` is locked. However, there's nothing that guarantees that
the callback itself won't call back into `TypeSystemMap` APIs on the
same thread. This lead to double-locking `m_mutex`, which is undefined
behaviour. We've seen this cause a deadlock in the swift plugin with
following backtrace:
```
int main() {
std::unique_ptr<int> up = std::make_unique<int>(5);
volatile int val = *up;
return val;
}
clang++ -std=c++2a -g -O1 main.cpp
./bin/lldb -o “br se -p return” -o run -o “v *up” -o “expr *up” -b
```
```
frame #4: std::lock_guard<std::mutex>::lock_guard
frame #5: lldb_private::TypeSystemMap::GetTypeSystemForLanguage <<<< Lock #2
frame #6: lldb_private::TypeSystemMap::GetTypeSystemForLanguage
frame #7: lldb_private::Target::GetScratchTypeSystemForLanguage
...
frame #26: lldb_private::SwiftASTContext::LoadLibraryUsingPaths
frame #27: lldb_private::SwiftASTContext::LoadModule
frame #30: swift::ModuleDecl::collectLinkLibraries
frame #31: lldb_private::SwiftASTContext::LoadModule
frame #34: lldb_private::SwiftASTContext::GetCompileUnitImportsImpl
frame #35: lldb_private::SwiftASTContext::PerformCompileUnitImports
frame #36: lldb_private::TypeSystemSwiftTypeRefForExpressions::GetSwiftASTContext
frame #37: lldb_private::TypeSystemSwiftTypeRefForExpressions::GetPersistentExpressionState
frame #38: lldb_private::Target::GetPersistentSymbol
frame #41: lldb_private::TypeSystemMap::ForEach <<<< Lock #1
frame #42: lldb_private::Target::GetPersistentSymbol
frame #43: lldb_private::IRExecutionUnit::FindInUserDefinedSymbols
frame #44: lldb_private::IRExecutionUnit::FindSymbol
frame #45: lldb_private::IRExecutionUnit::MemoryManager::GetSymbolAddressAndPresence
frame #46: lldb_private::IRExecutionUnit::MemoryManager::findSymbol
frame #47: non-virtual thunk to lldb_private::IRExecutionUnit::MemoryManager::findSymbol
frame #48: llvm::LinkingSymbolResolver::findSymbol
frame #49: llvm::LegacyJITSymbolResolver::lookup
frame #50: llvm::RuntimeDyldImpl::resolveExternalSymbols
frame #51: llvm::RuntimeDyldImpl::resolveRelocations
frame #52: llvm::MCJIT::finalizeLoadedModules
frame #53: llvm::MCJIT::finalizeObject
frame #54: lldb_private::IRExecutionUnit::ReportAllocations
frame #55: lldb_private::IRExecutionUnit::GetRunnableInfo
frame #56: lldb_private::ClangExpressionParser::PrepareForExecution
frame #57: lldb_private::ClangUserExpression::TryParse
frame #58: lldb_private::ClangUserExpression::Parse
```
Our solution is to simply iterate over a local copy of `m_map`.
**Testing**
* Confirmed on manual reproducer (would reproduce 100% of the time
before the patch)
Differential Revision: https://reviews.llvm.org/D149949
|
This issue is stale because it has been open 180 days with no activity. Remove stale label or comment or this will be automatically closed in 30 days. |
|
Closing as we've now forbidden buffer copy-back to host in the spec and the lazy queue property no longer exists |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Using buffers inside a command graph potentially causes a copy back to host to be triggered when the buffer is destroyed or synchronized in some other way. This presents issues for the current graphs implementation, which causes an error to be thrown similar to #24. The cause is slightly different in this case though:
This is the same problem preventing any other arbirtrary commands from executing on a queue which use the
lazy_queueproperty but likely requires a decent amount of work to support in the current POC. It also may not be feasible in the current implementation to indicate to the queue which submissions are intended for a graph and which are not without making changes to the PI interface by modifyingpi_queue, which seems like it may be required to solve this problem.The text was updated successfully, but these errors were encountered: