No comments on this specifically, but +1 to reducing shared_ptr usage in general. We use way too many, and often it feels like shared_ptr is used as a way to avoid having to think about ownership, which leads to more problems than it solves imo
We do a pretty good job of using shared pointers where they are needed, so I don’t agree with the above statement. We have strong memory models in LLDB and certain things need to keep certain things alive. LLVM is quite the wild west with regard to memory management, so I don’t want to base any changes on it. We just need to use std::weak_ptr when needed. There should be no two items that contain strong references to each other.
In general think about what should keep things alive. If I have a SBModule variable, I should expect that it won’t let my module go away. If I have a SBTarget, I would expect it to keep the target around as long as I have a reference. Things that belong to any of these items, like breakpoints, watchpoints, compile units, functions, shouldn’t keep the module or target around.
I’ve been debugging the newly added TestStepOverBreakpoint.py, which
has been failing on windows, for a very prosaic reason: after the test
completes, we are unable to run “make clean” because lldb still holds
the file open.
After some debugging, I’ve found that this happens because the test
case stores the SBBreakpoint object in a member variable of the python
test case class. The breakpoint class ends up transitively holding a
reference to the main executable module, which prevents the module
from being garbage-collected when the target is destroyed.
For reference, the full ownership chain is something like:
StepOverBreakpointTestCase(python) => SBBreakpoint => Breakpoint =>
BreakpointLocation => LLVMUserExpression => IRExecutionUnit =>
SymbolContext => Module.
To get the test working, we need to break this chain somewhere. A
couple of places I see are:
- BreakpointLocation: Remove the compiled expression reference when
the target is destroyed (AFAICS, it is used as a cache to avoid
recomputing the expression every time. It can be theoretically
recomputed if needed, but that shouldn’t be necessary as the target is
We should absolutely be clearing these useless expressions when the process is killed. Compiling expressions take 300-500 ms at the very least and we do cache the breakpoint expressions in each location, which is good for the lifetime of the process, but these locations should go away when the process goes away or execs.
- SBBreakpoint: make SBBreakpoint hold a weak_ptr to the Breakpoint
object. When the target is destroyed, the SBBreakpoint object becomes
invalid (One doesn’t cannot do anything useful with the breakpoint
once the target has been deleted anyway).
This is also valid. If the breakpoint is still around, you will be able to use it as the weak_ptr will make a shared pointer with a valid value, and if it isn’t it won’t be able to do anything anyway.
- StepOverBreakpointTestCase: Have the test not store the breakpoints
in the test case object. Basically, declare that this is not a bug,
and it’s the users responsibility to clean up necessary objects.
It would be nice to avoid this.
Any thoughts on what is the appropriate solution here?
So I vote for:
1 - remove all breakpoint expressions when the target stops and the locations becomes unresolved.
2 - SBBreakpoint, SBBreakpointLocation, SBWatchpoint and can switch over to using std::weak_ptr
The other fix is that might be worth it is to make a SymbolContextRef, kind of like we have the ExecutionContext (with shared pointers) and ExecutionContextRef (with weak pointers). We would then have things like IRExecutionUnit switch over to use a SymbolContextRef. We should also look for other places where people are storing SymbolContext as member variables and possibly switch them over to use SymbolContextRef.
lldb::TargetWP target_wp; ///< The Target for a given query
lldb::ModuleWP module_wp; ///< The Module for a given query
CompileUnit *comp_unit; ///< The CompileUnit for a given query
Function *function; ///< The Function for a given query
Block *block; ///< The Block for a given query
LineEntry line_entry; ///< The LineEntry for a given query
Symbol *symbol; ///< The Symbol for a given query
Variable *variable; ///< The global variable matching the given query
Then a SymbolContext would add a constructor that takes a SymbolContextRef:
SymbolContextRef sym_ctx_ref = …;
Anytime you want to use a SymbolContextRef, you first must turn it into a SymbolContext:
If the module is gone, then we need to NULL out comp_unit, function, block, symbol and variable (since they are owned by the Module) and clear the line_entry. If the module is still around, then these things can stay as is and the local SymbolContext now has a strong reference to the module and or target.
IRExecutionUnit currently uses the symbol context for name lookups, so I am not even sure if it is a good idea for it to store one permanently, but if it we use SymbolContextRef, then I don’t care as much.
So in general think about who owns what and what should keep things alive. In this case it sounds like making SBBreakpoint, SBBreakpointLocation, and SBWatchpoint use std::weak_ptr would solve this part of this issue, but I still don’t like a IRExecutionUnit being able to keep a module around, so we should probably make a SymbolContextRef. I am guessing if you switch SBBreakpoint over to use std::weak_ptr all the expressions will go away on their own.