forked from OSchip/llvm-project
[ThreadPlan] Reflow docs to fit the 80 column limit, NFC
This commit is contained in:
parent
ba21376883
commit
34d56b05fd
|
@ -23,310 +23,260 @@
|
|||
namespace lldb_private {
|
||||
|
||||
// ThreadPlan:
|
||||
//
|
||||
// This is the pure virtual base class for thread plans.
|
||||
//
|
||||
// The thread plans provide the "atoms" of behavior that
|
||||
// all the logical process control, either directly from commands or through
|
||||
// more complex composite plans will rely on.
|
||||
// The thread plans provide the "atoms" of behavior that all the logical
|
||||
// process control, either directly from commands or through more complex
|
||||
// composite plans will rely on.
|
||||
//
|
||||
// Plan Stack:
|
||||
//
|
||||
// The thread maintaining a thread plan stack, and you program the actions of a
|
||||
// particular thread
|
||||
// by pushing plans onto the plan stack.
|
||||
// There is always a "Current" plan, which is the top of the plan stack,
|
||||
// though in some cases
|
||||
// The thread maintaining a thread plan stack, and you program the actions of
|
||||
// a particular thread by pushing plans onto the plan stack. There is always
|
||||
// a "Current" plan, which is the top of the plan stack, though in some cases
|
||||
// a plan may defer to plans higher in the stack for some piece of information
|
||||
// (let us define that the plan stack grows downwards).
|
||||
//
|
||||
// The plan stack is never empty, there is always a Base Plan which persists
|
||||
// through the life
|
||||
// of the running process.
|
||||
// through the life of the running process.
|
||||
//
|
||||
//
|
||||
// Creating Plans:
|
||||
//
|
||||
// The thread plan is generally created and added to the plan stack through the
|
||||
// QueueThreadPlanFor... API
|
||||
// in lldb::Thread. Those API's will return the plan that performs the named
|
||||
// operation in a manner
|
||||
// appropriate for the current process. The plans in lldb/source/Target are
|
||||
// generic
|
||||
// The thread plan is generally created and added to the plan stack through
|
||||
// the QueueThreadPlanFor... API in lldb::Thread. Those API's will return the
|
||||
// plan that performs the named operation in a manner appropriate for the
|
||||
// current process. The plans in lldb/source/Target are generic
|
||||
// implementations, but a Process plugin can override them.
|
||||
//
|
||||
// ValidatePlan is then called. If it returns false, the plan is unshipped.
|
||||
// This is a little
|
||||
// convenience which keeps us from having to error out of the constructor.
|
||||
// This is a little convenience which keeps us from having to error out of the
|
||||
// constructor.
|
||||
//
|
||||
// Then the plan is added to the plan stack. When the plan is added to the
|
||||
// plan stack its DidPush
|
||||
// will get called. This is useful if a plan wants to push any additional
|
||||
// plans as it is constructed,
|
||||
// since you need to make sure you're already on the stack before you push
|
||||
// additional plans.
|
||||
// plan stack its DidPush will get called. This is useful if a plan wants to
|
||||
// push any additional plans as it is constructed, since you need to make sure
|
||||
// you're already on the stack before you push additional plans.
|
||||
//
|
||||
// Completed Plans:
|
||||
//
|
||||
// When the target process stops the plans are queried, among other things, for
|
||||
// whether their job is done.
|
||||
// If it is they are moved from the plan stack to the Completed Plan stack in
|
||||
// reverse order from their position
|
||||
// on the plan stack (since multiple plans may be done at a given stop.) This
|
||||
// is used primarily so that
|
||||
// the lldb::Thread::StopInfo for the thread can be set properly. If one plan
|
||||
// pushes another to achieve part of
|
||||
// its job, but it doesn't want that sub-plan to be the one that sets the
|
||||
// StopInfo, then call SetPrivate on the
|
||||
// sub-plan when you create it, and the Thread will pass over that plan in
|
||||
// reporting the reason for the stop.
|
||||
// When the target process stops the plans are queried, among other things,
|
||||
// for whether their job is done. If it is they are moved from the plan stack
|
||||
// to the Completed Plan stack in reverse order from their position on the
|
||||
// plan stack (since multiple plans may be done at a given stop.) This is
|
||||
// used primarily so that the lldb::Thread::StopInfo for the thread can be set
|
||||
// properly. If one plan pushes another to achieve part of its job, but it
|
||||
// doesn't want that sub-plan to be the one that sets the StopInfo, then call
|
||||
// SetPrivate on the sub-plan when you create it, and the Thread will pass
|
||||
// over that plan in reporting the reason for the stop.
|
||||
//
|
||||
// Discarded plans:
|
||||
//
|
||||
// Your plan may also get discarded, i.e. moved from the plan stack to the
|
||||
// "discarded plan stack". This can
|
||||
// happen, for instance, if the plan is calling a function and the function
|
||||
// call crashes and you want
|
||||
// to unwind the attempt to call. So don't assume that your plan will always
|
||||
// successfully stop. Which leads to:
|
||||
// "discarded plan stack". This can happen, for instance, if the plan is
|
||||
// calling a function and the function call crashes and you want to unwind the
|
||||
// attempt to call. So don't assume that your plan will always successfully
|
||||
// stop. Which leads to:
|
||||
//
|
||||
// Cleaning up after your plans:
|
||||
//
|
||||
// When the plan is moved from the plan stack its WillPop method is always
|
||||
// called, no matter why. Once it is
|
||||
// moved off the plan stack it is done, and won't get a chance to run again.
|
||||
// So you should
|
||||
// undo anything that affects target state in this method. But be sure to
|
||||
// leave the plan able to correctly
|
||||
// fill the StopInfo, however.
|
||||
// N.B. Don't wait to do clean up target state till the destructor, since that
|
||||
// will usually get called when
|
||||
// called, no matter why. Once it is moved off the plan stack it is done, and
|
||||
// won't get a chance to run again. So you should undo anything that affects
|
||||
// target state in this method. But be sure to leave the plan able to
|
||||
// correctly fill the StopInfo, however. N.B. Don't wait to do clean up
|
||||
// target state till the destructor, since that will usually get called when
|
||||
// the target resumes, and you want to leave the target state correct for new
|
||||
// plans in the time between when
|
||||
// your plan gets unshipped and the next resume.
|
||||
// plans in the time between when your plan gets unshipped and the next
|
||||
// resume.
|
||||
//
|
||||
// Thread State Checkpoint:
|
||||
//
|
||||
// Note that calling functions on target process (ThreadPlanCallFunction) changes
|
||||
// current thread state. The function can be called either by direct user demand or
|
||||
// internally, for example lldb allocates memory on device to calculate breakpoint
|
||||
// condition expression - on Linux it is performed by calling mmap on device.
|
||||
// ThreadStateCheckpoint saves Thread state (stop info and completed
|
||||
// plan stack) to restore it after completing function call.
|
||||
// Note that calling functions on target process (ThreadPlanCallFunction)
|
||||
// changes current thread state. The function can be called either by direct
|
||||
// user demand or internally, for example lldb allocates memory on device to
|
||||
// calculate breakpoint condition expression - on Linux it is performed by
|
||||
// calling mmap on device. ThreadStateCheckpoint saves Thread state (stop
|
||||
// info and completed plan stack) to restore it after completing function
|
||||
// call.
|
||||
//
|
||||
// Over the lifetime of the plan, various methods of the ThreadPlan are then
|
||||
// called in response to changes of state in
|
||||
// the process we are debugging as follows:
|
||||
// called in response to changes of state in the process we are debugging as
|
||||
// follows:
|
||||
//
|
||||
// Resuming:
|
||||
//
|
||||
// When the target process is about to be restarted, the plan's WillResume
|
||||
// method is called,
|
||||
// giving the plan a chance to prepare for the run. If WillResume returns
|
||||
// false, then the
|
||||
// process is not restarted. Be sure to set an appropriate error value in the
|
||||
// Process if
|
||||
// you have to do this. Note, ThreadPlans actually implement DoWillResume,
|
||||
// WillResume wraps that call.
|
||||
// method is called, giving the plan a chance to prepare for the run. If
|
||||
// WillResume returns false, then the process is not restarted. Be sure to
|
||||
// set an appropriate error value in the Process if you have to do this.
|
||||
// Note, ThreadPlans actually implement DoWillResume, WillResume wraps that
|
||||
// call.
|
||||
//
|
||||
// Next the "StopOthers" method of all the threads are polled, and if one
|
||||
// thread's Current plan
|
||||
// returns "true" then only that thread gets to run. If more than one returns
|
||||
// "true" the threads that want to run solo
|
||||
// get run one by one round robin fashion. Otherwise all are let to run.
|
||||
// thread's Current plan returns "true" then only that thread gets to run. If
|
||||
// more than one returns "true" the threads that want to run solo get run one
|
||||
// by one round robin fashion. Otherwise all are let to run.
|
||||
//
|
||||
// Note, the way StopOthers is implemented, the base class implementation just
|
||||
// asks the previous plan. So if your plan
|
||||
// has no opinion about whether it should run stopping others or not, just
|
||||
// don't implement StopOthers, and the parent
|
||||
// will be asked.
|
||||
// asks the previous plan. So if your plan has no opinion about whether it
|
||||
// should run stopping others or not, just don't implement StopOthers, and the
|
||||
// parent will be asked.
|
||||
//
|
||||
// Finally, for each thread that is running, it run state is set to the return
|
||||
// of RunState from the
|
||||
// thread's Current plan.
|
||||
// of RunState from the thread's Current plan.
|
||||
//
|
||||
// Responding to a stop:
|
||||
//
|
||||
// When the target process stops, the plan is called in the following stages:
|
||||
//
|
||||
// First the thread asks the Current Plan if it can handle this stop by calling
|
||||
// PlanExplainsStop.
|
||||
// If the Current plan answers "true" then it is asked if the stop should
|
||||
// percolate all the way to the
|
||||
// user by calling the ShouldStop method. If the current plan doesn't explain
|
||||
// the stop, then we query up
|
||||
// the plan stack for a plan that does explain the stop. The plan that does
|
||||
// explain the stop then needs to
|
||||
// figure out what to do about the plans below it in the stack. If the stop is
|
||||
// recoverable, then the plan that
|
||||
// understands it can just do what it needs to set up to restart, and then
|
||||
// continue.
|
||||
// Otherwise, the plan that understood the stop should call DiscardPlanStack to
|
||||
// clean up the stack below it.
|
||||
// Note, plans actually implement DoPlanExplainsStop, the result is cached in
|
||||
// PlanExplainsStop so the DoPlanExplainsStop
|
||||
// itself will only get called once per stop.
|
||||
// First the thread asks the Current Plan if it can handle this stop by
|
||||
// calling PlanExplainsStop. If the Current plan answers "true" then it is
|
||||
// asked if the stop should percolate all the way to the user by calling the
|
||||
// ShouldStop method. If the current plan doesn't explain the stop, then we
|
||||
// query up the plan stack for a plan that does explain the stop. The plan
|
||||
// that does explain the stop then needs to figure out what to do about the
|
||||
// plans below it in the stack. If the stop is recoverable, then the plan
|
||||
// that understands it can just do what it needs to set up to restart, and
|
||||
// then continue. Otherwise, the plan that understood the stop should call
|
||||
// DiscardPlanStack to clean up the stack below it. Note, plans actually
|
||||
// implement DoPlanExplainsStop, the result is cached in PlanExplainsStop so
|
||||
// the DoPlanExplainsStop itself will only get called once per stop.
|
||||
//
|
||||
// Master plans:
|
||||
//
|
||||
// In the normal case, when we decide to stop, we will collapse the plan stack
|
||||
// up to the point of the plan that understood
|
||||
// the stop reason. However, if a plan wishes to stay on the stack after an
|
||||
// event it didn't directly handle
|
||||
// it can designate itself a "Master" plan by responding true to IsMasterPlan,
|
||||
// and then if it wants not to be
|
||||
// discarded, it can return false to OkayToDiscard, and it and all its dependent
|
||||
// plans will be preserved when
|
||||
// we resume execution.
|
||||
// In the normal case, when we decide to stop, we will collapse the plan
|
||||
// stack up to the point of the plan that understood the stop reason.
|
||||
// However, if a plan wishes to stay on the stack after an event it didn't
|
||||
// directly handle it can designate itself a "Master" plan by responding true
|
||||
// to IsMasterPlan, and then if it wants not to be discarded, it can return
|
||||
// false to OkayToDiscard, and it and all its dependent plans will be
|
||||
// preserved when we resume execution.
|
||||
//
|
||||
// The other effect of being a master plan is that when the Master plan is done
|
||||
// , if it has set "OkayToDiscard" to false,
|
||||
// then it will be popped & execution will stop and return to the user.
|
||||
// Remember that if OkayToDiscard is false, the
|
||||
// plan will be popped and control will be given to the next plan above it on
|
||||
// the stack So setting OkayToDiscard to
|
||||
// false means the user will regain control when the MasterPlan is completed.
|
||||
// The other effect of being a master plan is that when the Master plan is
|
||||
// done , if it has set "OkayToDiscard" to false, then it will be popped &
|
||||
// execution will stop and return to the user. Remember that if OkayToDiscard
|
||||
// is false, the plan will be popped and control will be given to the next
|
||||
// plan above it on the stack So setting OkayToDiscard to false means the
|
||||
// user will regain control when the MasterPlan is completed.
|
||||
//
|
||||
// Between these two controls this allows things like: a MasterPlan/DontDiscard
|
||||
// Step Over to hit a breakpoint, stop and
|
||||
// return control to the user, but then when the user continues, the step out
|
||||
// succeeds.
|
||||
// Even more tricky, when the breakpoint is hit, the user can continue to step
|
||||
// in/step over/etc, and finally when they
|
||||
// continue, they will finish up the Step Over.
|
||||
// Between these two controls this allows things like: a
|
||||
// MasterPlan/DontDiscard Step Over to hit a breakpoint, stop and return
|
||||
// control to the user, but then when the user continues, the step out
|
||||
// succeeds. Even more tricky, when the breakpoint is hit, the user can
|
||||
// continue to step in/step over/etc, and finally when they continue, they
|
||||
// will finish up the Step Over.
|
||||
//
|
||||
// FIXME: MasterPlan & OkayToDiscard aren't really orthogonal. MasterPlan
|
||||
// designation means that this plan controls
|
||||
// it's fate and the fate of plans below it. OkayToDiscard tells whether the
|
||||
// MasterPlan wants to stay on the stack. I
|
||||
// originally thought "MasterPlan-ness" would need to be a fixed characteristic
|
||||
// of a ThreadPlan, in which case you needed
|
||||
// the extra control. But that doesn't seem to be true. So we should be able
|
||||
// to convert to only MasterPlan status to mean
|
||||
// the current "MasterPlan/DontDiscard". Then no plans would be MasterPlans by
|
||||
// default, and you would set the ones you
|
||||
// designation means that this plan controls it's fate and the fate of plans
|
||||
// below it. OkayToDiscard tells whether the MasterPlan wants to stay on the
|
||||
// stack. I originally thought "MasterPlan-ness" would need to be a fixed
|
||||
// characteristic of a ThreadPlan, in which case you needed the extra control.
|
||||
// But that doesn't seem to be true. So we should be able to convert to only
|
||||
// MasterPlan status to mean the current "MasterPlan/DontDiscard". Then no
|
||||
// plans would be MasterPlans by default, and you would set the ones you
|
||||
// wanted to be "user level" in this way.
|
||||
//
|
||||
//
|
||||
// Actually Stopping:
|
||||
//
|
||||
// If a plan says responds "true" to ShouldStop, then it is asked if it's job
|
||||
// is complete by calling
|
||||
// MischiefManaged. If that returns true, the plan is popped from the plan
|
||||
// stack and added to the
|
||||
// Completed Plan Stack. Then the next plan in the stack is asked if it
|
||||
// ShouldStop, and it returns "true",
|
||||
// it is asked if it is done, and if yes popped, and so on till we reach a plan
|
||||
// that is not done.
|
||||
// is complete by calling MischiefManaged. If that returns true, the plan is
|
||||
// popped from the plan stack and added to the Completed Plan Stack. Then the
|
||||
// next plan in the stack is asked if it ShouldStop, and it returns "true",
|
||||
// it is asked if it is done, and if yes popped, and so on till we reach a
|
||||
// plan that is not done.
|
||||
//
|
||||
// Since you often know in the ShouldStop method whether your plan is complete,
|
||||
// as a convenience you can call
|
||||
// SetPlanComplete and the ThreadPlan implementation of MischiefManaged will
|
||||
// return "true", without your having
|
||||
// to redo the calculation when your sub-classes MischiefManaged is called. If
|
||||
// you call SetPlanComplete, you can
|
||||
// later use IsPlanComplete to determine whether the plan is complete. This is
|
||||
// only a convenience for sub-classes,
|
||||
// Since you often know in the ShouldStop method whether your plan is
|
||||
// complete, as a convenience you can call SetPlanComplete and the ThreadPlan
|
||||
// implementation of MischiefManaged will return "true", without your having
|
||||
// to redo the calculation when your sub-classes MischiefManaged is called.
|
||||
// If you call SetPlanComplete, you can later use IsPlanComplete to determine
|
||||
// whether the plan is complete. This is only a convenience for sub-classes,
|
||||
// the logic in lldb::Thread will only call MischiefManaged.
|
||||
//
|
||||
// One slightly tricky point is you have to be careful using SetPlanComplete in
|
||||
// PlanExplainsStop because you
|
||||
// are not guaranteed that PlanExplainsStop for a plan will get called before
|
||||
// ShouldStop gets called. If your sub-plan
|
||||
// One slightly tricky point is you have to be careful using SetPlanComplete
|
||||
// in PlanExplainsStop because you are not guaranteed that PlanExplainsStop
|
||||
// for a plan will get called before ShouldStop gets called. If your sub-plan
|
||||
// explained the stop and then popped itself, only your ShouldStop will get
|
||||
// called.
|
||||
//
|
||||
// If ShouldStop for any thread returns "true", then the WillStop method of the
|
||||
// Current plan of
|
||||
// all threads will be called, the stop event is placed on the Process's public
|
||||
// broadcaster, and
|
||||
// control returns to the upper layers of the debugger.
|
||||
// If ShouldStop for any thread returns "true", then the WillStop method of
|
||||
// the Current plan of all threads will be called, the stop event is placed on
|
||||
// the Process's public broadcaster, and control returns to the upper layers
|
||||
// of the debugger.
|
||||
//
|
||||
// Reporting the stop:
|
||||
//
|
||||
// When the process stops, the thread is given a StopReason, in the form of a
|
||||
// StopInfo object. If there is a completed
|
||||
// plan corresponding to the stop, then the "actual" stop reason can be
|
||||
// suppressed, and instead a StopInfoThreadPlan
|
||||
// object will be cons'ed up from the top completed plan in the stack.
|
||||
// However, if the plan doesn't want to be
|
||||
// the stop reason, then it can call SetPlanComplete and pass in "false" for
|
||||
// the "success" parameter. In that case,
|
||||
// the real stop reason will be used instead. One example of this is the
|
||||
// "StepRangeStepIn" thread plan. If it stops
|
||||
// because of a crash or breakpoint hit, it wants to unship itself, because it
|
||||
// isn't so useful to have step in keep going
|
||||
// after a breakpoint hit. But it can't be the reason for the stop or no-one
|
||||
// would see that they had hit a breakpoint.
|
||||
// StopInfo object. If there is a completed plan corresponding to the stop,
|
||||
// then the "actual" stop reason can be suppressed, and instead a
|
||||
// StopInfoThreadPlan object will be cons'ed up from the top completed plan in
|
||||
// the stack. However, if the plan doesn't want to be the stop reason, then
|
||||
// it can call SetPlanComplete and pass in "false" for the "success"
|
||||
// parameter. In that case, the real stop reason will be used instead. One
|
||||
// example of this is the "StepRangeStepIn" thread plan. If it stops because
|
||||
// of a crash or breakpoint hit, it wants to unship itself, because it isn't
|
||||
// so useful to have step in keep going after a breakpoint hit. But it can't
|
||||
// be the reason for the stop or no-one would see that they had hit a
|
||||
// breakpoint.
|
||||
//
|
||||
// Cleaning up the plan stack:
|
||||
//
|
||||
// One of the complications of MasterPlans is that you may get past the limits
|
||||
// of a plan without triggering it to clean
|
||||
// itself up. For instance, if you are doing a MasterPlan StepOver, and hit a
|
||||
// breakpoint in a called function, then
|
||||
// step over enough times to step out of the initial StepOver range, each of
|
||||
// the step overs will explain the stop &
|
||||
// take themselves off the stack, but control would never be returned to the
|
||||
// original StepOver. Eventually, the user
|
||||
// will continue, and when that continue stops, the old stale StepOver plan
|
||||
// that was left on the stack will get woken
|
||||
// up and notice it is done. But that can leave junk on the stack for a while.
|
||||
// To avoid that, the plans implement a
|
||||
// "IsPlanStale" method, that can check whether it is relevant anymore. On
|
||||
// stop, after the regular plan negotiation,
|
||||
// the remaining plan stack is consulted and if any plan says it is stale, it
|
||||
// and the plans below it are discarded from
|
||||
// the stack.
|
||||
// of a plan without triggering it to clean itself up. For instance, if you
|
||||
// are doing a MasterPlan StepOver, and hit a breakpoint in a called function,
|
||||
// then step over enough times to step out of the initial StepOver range, each
|
||||
// of the step overs will explain the stop & take themselves off the stack,
|
||||
// but control would never be returned to the original StepOver. Eventually,
|
||||
// the user will continue, and when that continue stops, the old stale
|
||||
// StepOver plan that was left on the stack will get woken up and notice it is
|
||||
// done. But that can leave junk on the stack for a while. To avoid that, the
|
||||
// plans implement a "IsPlanStale" method, that can check whether it is
|
||||
// relevant anymore. On stop, after the regular plan negotiation, the
|
||||
// remaining plan stack is consulted and if any plan says it is stale, it and
|
||||
// the plans below it are discarded from the stack.
|
||||
//
|
||||
// Automatically Resuming:
|
||||
//
|
||||
// If ShouldStop for all threads returns "false", then the target process will
|
||||
// resume. This then cycles back to
|
||||
// Resuming above.
|
||||
// resume. This then cycles back to Resuming above.
|
||||
//
|
||||
// Reporting eStateStopped events when the target is restarted:
|
||||
//
|
||||
// If a plan decides to auto-continue the target by returning "false" from
|
||||
// ShouldStop, then it will be asked
|
||||
// whether the Stopped event should still be reported. For instance, if you
|
||||
// hit a breakpoint that is a User set
|
||||
// breakpoint, but the breakpoint callback said to continue the target process,
|
||||
// you might still want to inform
|
||||
// the upper layers of lldb that the stop had happened.
|
||||
// The way this works is every thread gets to vote on whether to report the
|
||||
// stop. If all votes are eVoteNoOpinion,
|
||||
// then the thread list will decide what to do (at present it will pretty much
|
||||
// always suppress these stopped events.)
|
||||
// If there is an eVoteYes, then the event will be reported regardless of the
|
||||
// other votes. If there is an eVoteNo
|
||||
// and no eVoteYes's, then the event won't be reported.
|
||||
// ShouldStop, then it will be asked whether the Stopped event should still be
|
||||
// reported. For instance, if you hit a breakpoint that is a User set
|
||||
// breakpoint, but the breakpoint callback said to continue the target
|
||||
// process, you might still want to inform the upper layers of lldb that the
|
||||
// stop had happened. The way this works is every thread gets to vote on
|
||||
// whether to report the stop. If all votes are eVoteNoOpinion, then the
|
||||
// thread list will decide what to do (at present it will pretty much always
|
||||
// suppress these stopped events.) If there is an eVoteYes, then the event
|
||||
// will be reported regardless of the other votes. If there is an eVoteNo and
|
||||
// no eVoteYes's, then the event won't be reported.
|
||||
//
|
||||
// One other little detail here, sometimes a plan will push another plan onto
|
||||
// the plan stack to do some part of
|
||||
// the first plan's job, and it would be convenient to tell that plan how it
|
||||
// should respond to ShouldReportStop.
|
||||
// the plan stack to do some part of the first plan's job, and it would be
|
||||
// convenient to tell that plan how it should respond to ShouldReportStop.
|
||||
// You can do that by setting the stop_vote in the child plan when you create
|
||||
// it.
|
||||
//
|
||||
// Suppressing the initial eStateRunning event:
|
||||
//
|
||||
// The private process running thread will take care of ensuring that only one
|
||||
// "eStateRunning" event will be
|
||||
// delivered to the public Process broadcaster per public eStateStopped event.
|
||||
// However there are some cases
|
||||
// where the public state of this process is eStateStopped, but a thread plan
|
||||
// needs to restart the target, but
|
||||
// doesn't want the running event to be publicly broadcast. The obvious
|
||||
// example of this is running functions
|
||||
// by hand as part of expression evaluation. To suppress the running event
|
||||
// return eVoteNo from ShouldReportStop,
|
||||
// to force a running event to be reported return eVoteYes, in general though
|
||||
// you should return eVoteNoOpinion
|
||||
// which will allow the ThreadList to figure out the right thing to do.
|
||||
// The run_vote argument to the constructor works like stop_vote, and is a way
|
||||
// for a plan to instruct a sub-plan
|
||||
// on how to respond to ShouldReportStop.
|
||||
//
|
||||
// "eStateRunning" event will be delivered to the public Process broadcaster
|
||||
// per public eStateStopped event. However there are some cases where the
|
||||
// public state of this process is eStateStopped, but a thread plan needs to
|
||||
// restart the target, but doesn't want the running event to be publicly
|
||||
// broadcast. The obvious example of this is running functions by hand as
|
||||
// part of expression evaluation. To suppress the running event return
|
||||
// eVoteNo from ShouldReportStop, to force a running event to be reported
|
||||
// return eVoteYes, in general though you should return eVoteNoOpinion which
|
||||
// will allow the ThreadList to figure out the right thing to do. The
|
||||
// run_vote argument to the constructor works like stop_vote, and is a way for
|
||||
// a plan to instruct a sub-plan on how to respond to ShouldReportStop.
|
||||
|
||||
class ThreadPlan : public std::enable_shared_from_this<ThreadPlan>,
|
||||
public UserID {
|
||||
|
|
Loading…
Reference in New Issue