[Lldb-commits] [PATCH] D76814: Preserve ThreadPlanStacks for unreported threads

Mon Mar 30 07:00:37 PDT 2020

labath added inline comments.

================
Comment at: lldb/source/Target/TargetProperties.td:183
     Desc<"A path to a python OS plug-in module file that contains a OperatingSystemPlugIn class.">;
+  def PluginReportsAllThreads: Property<"plugin-reports-all-threads", "Boolean">,
+    Global,
----------------
jingham wrote:
> labath wrote:
> > jingham wrote:
> > > labath wrote:
> > > > If this is relevant only for os plugins, then it would be good to reflect that in the name as well.
> > > I thought about that.  In the future a regular Process plugin might decide it was too expensive to report all threads as well.  There's nothing in this patch that wouldn't "just work" with that case as well.  Leaving the OS out was meant to indicate this was about how the Process plugin OR any of its helpers (e.g. the OS Plugin) produces threads.
> > Well, I am hoping that if we ever extend this support to the regular process plugins, we will implement it in a way where the plugin itself can tell us whether it is operating/supporting this mode (which I guess would involve a new gdb-remote packet and some specification of what exactly should happen when it gets sent), instead of relying on the user to set this correctly.
> > 
> > I mean, in an ideal world this is I would want to happen with the python plugins as well, but it seems that we are stuck with some existing plugins which already do that. However, I wouldn't want to plan on the same thing happening again. :)
> Right, I put this in mostly to help backwards compatibility.  For instance, another OS Plugin I know about handles some older cooperative threading scheme.  That one does report all threads on each stop.  I didn't want to force them to do anything to keep their plugin working as well as it did before.  That's also why I set the default to true here.
> 
> Even when we have plugins that actually support not reporting all threads, you could imagine somebody having a Process plugin that supports both modes - particularly early in the development of its support for this feature, and in some corner cases the "doesn't report all threads" mode has some subtle problem.  Having this setting will allow people to get the slower but more assuredly correct behavior till it works 100% reliably.  So I still think the setting has some value.
> 
> But I agree, going forward there should be some kind of handshake between the ThreadPlanStackMap and the Process Plugin, either a "I've reported all threads now" which could trigger a prune, or a "Is TID X still alive" which the generic code could use to balance the cost of keeping outdated stacks alive against when we want to ask about all threads.
All of this sounds fine, and I wouldn't mind having a setting like that, even after the support for partial thread lists is considered "stable". However, that sounds like a different setting to me -- that's like a directive to the process plugin about how it should behave, whereas this setting is more like a statement of fact about what the plugin does.

The setting could be later repurposed to do both, but it's not clear to me whether that is useful. Like, since we already support plugin-specific settings, the plugin which decides to implement both modes of operation could expose that as a custom setting. That way, one wouldn't get the impression that this setting applies to any process plugin...

================
Comment at: lldb/test/API/functionalities/plugins/python_os_plugin/stepping_plugin_threads/main.cpp:40
+  std::thread thread_1(thread_func);
+  g_cv.wait(main_lock);
+  g_value = 1;
----------------
spurious wake up danger here too

================
Comment at: lldb/test/API/functionalities/thread_plan/TestThreadPlanCommands.py:23
+
+    def check_list_output(self, command, active_plans = [], completed_plans = [], discarded_plans = []):
+        # Check the "thread plan list" output against a list of active & completed and discarded plans.
----------------
jingham wrote:
> labath wrote:
> > jingham wrote:
> > > labath wrote:
> > > > this looks like a model example for using `self.filecheck`
> > > I don't see that.  Do you mean write a file check matching file with the contents of the three arrays in this function, and then run file check on that?  Or did you mean convert the call sites into embedded patterns and then call file check on myself?  But then I'd have to also embed the headers in the body such that they came in the right order for all the tests.  Neither of these seem like attractive options to me.  It doesn't seem like it will be hard to maintain this little checker, and I don't see what I would gain by using file check instead.
> > What I meant was doing doing something like:
> > ```
> > self.filecheck("thread plan list %d"%(current_id), __file__, "--check-prefix=CHECK1")
> > # CHECK1:      Active plan stack:
> > # CHECK1-NEXT: Element 0: Base thread plan
> > # CHECK1-NEXT: Element 1: Stepping over line main.c
> > # CHECK1-EMPTY:
> > ```
> > instead of `self.check_list_output(..., [], ["Stepping over line main.c"])`
> > 
> > The main advantage of that I see is in the error messages it produces. If anything goes wrong, the most likely error message we're going to get is "9 != 10, Too many elements in match arrays".  The second most likely error message is "False is not True, Didn't find active plan Stepping over line main.c". If the same happens while using filecheck, it would print the string it is trying to match, the line it is trying to match it against, and the line of a "possible intended match", which makes it much easier to figure out what is going on without debugging the test. Now you could embed all of that into this matcher function, but then you're sort of reinventing FileCheck.
> > 
> > The second advantage is that it is easier to see what is the "expected" output supposed to be when it is layout out like this, instead of embedded in the logic of the matcher. That may introduce some amount of repetition for specifying the common parts of the output, but that is not necessarily bad (because it's more readable). If there is a huge amount of repetition, then there are ways to avoid that via using multiple prefixes:
> > ```
> > FileCheck --check-prefixes=ALL,FIRST
> > ...
> > FileCheck --check-prefixes=ALL,SECOND
> > # ALL: Active plan stack:
> > # ALL: Element 0: Base thread plan
> > # FIRST: Element 1: Stepping over line main.c
> > # SECOND: Element 1: ???
> > ```
> > But that again makes it harder to see the expected output, so it should be balanced against that.
> Yeah, if you don't mind, I'm going to keep it as it is.  To do it the way you suggest you either would have to repeat all the header stuff over and over or make some compound section for the filecheck part of the test with FIRST and SECOND etc for the tests.   That's sort of okay for a few cases but if you have more that gets hard to follow and edit without making mistakes.  
> 
> The way it is now, you say explicitly what you want to see in the the categories at the check site, which seems really clear to me.  And the processing code is not terribly difficult, so I don't think maintaining it will be a problem.
Well, I wouldn't say I don't mind, but I am not interested in making a protracted debate out of this. But if you are going to stick to this, then I think you should do something to make the error messages coming out of here more reasonable. The least you could do is use `self.runCmd` instead of handle command (so that the command output is visible in trace mode at least), and replace all `assertTrue(needle in haystack)` with `assertIn(needle, haystack)` (to make it easier to see what is being matched against what).

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D76814/new/

https://reviews.llvm.org/D76814