Ready… cancel… wait for it! (part 3)

A customer reported that their application was crashing in RPC, and they submitted a sample program which illustrated the same crash as their program. Their sample program was actually based on the AsyncRPC sample client program, which was nice, because it provided a mutually-known starting point. They made quite a few changes to the program, but this is the important one:

// old code:
// status = RpcAsyncCancelCall(&Async, FALSE);

// new code:
 status = RpcAsyncCancelCall(&Async, TRUE);

(It was actually more complicated than this, but this is the short version.)

The program was crashing for the same reason that Wednesday's I/O cancellation program was crashing: The program issued an asynchronous cancel and didn't wait for the cancel to complete. In this case, the crash occurred when the RPC call finally completed and RPC went about cleaning up the call based on the information in the now-freed RPC_ASYNC_STATE structure.

The error was probably caused by the not-very-helpful name for that last parameter to Rpc­Async­Cancel­Call: fAbort­Call, and the accompanying documentation which says, "In an abortive cancel (fAbort­Call is TRUE), the Rpc­Async­Cancel­Call function sends a cancel notification to the server and client side and the asynchronous call is canceled immediately, not waiting for a response from the server." Compare this to a nonabortive cancel, where "the Rpc­Async­Cancel­Call function notifies the server of the cancel and the client waits for the server to complete the call."

Obviously, it's faster if you don't wait for the server to respond, right? Let's pass TRUE, so that the function cancels the asynchronous call immediately without waiting for the server. Wow, look at how fast our program runs now!

Unfortunately, the documentation doesn't make it sufficiently clear that when you issue a cancellation, you still have to wait for the operation to complete before you can clean up all the resources associated with that operation. Another way of looking at that last parameter is to think of it as fAsync. If you pass fAsync = TRUE, then the Rpc­Async­Cancel­Call function issues the cancellation and returns before the operation completes. If you pass fAsync = FALSE, then the Rpc­Async­Cancel­Call function issues the cancellation and waits for the operation to complete before returning.

If you switch from a synchronous cancel to an asynchronous cancel, then you become responsible for keeping the RPC_ASYNC_STATE valid until the cancellation completes. In this case, the customer was using the Rpc­Notification­Type­Event notification type, which means that they need to wait for the Async.u.hEvent to become signaled before they can free the RPC_ASYNC_STATE.

The customer confirmed the fix and closed the support case. Another problem solved.

Three months later, the customer reopened the case, reporting that after they released a new version of their program with the aforementioned fix, they were nevertheless getting WinQual crashes which looked exactly like the ones that they were having before they applied the fix. It appears that the fix wasn't working.

Upon closer investigation, it turns out that the customer originally did apply the fix as recommended: They added a Wait­For­Single­Object(Async.u.hEvent, INFINITE) call before destroying the Async object to ensure that the cancellation was complete. However, they became frustrated that sometimes the cancellation would take a long time to complete, so they changed it to

WaitForSingleObject(Async.u.hEvent, 5000); // wait up to 5 seconds

The customer explained, "After the wait fails due to timeout, we just proceed as normal and call Rpc­Async­Complete­Call and free the the RPC_ASYNC_STATE. Is that wrong?"

Um, yeah. Changing the Wait­For­Single­Object from an infinite wait to one with a timeout means that you just reintroduced the bug that the Wait­For­Single­Object was originally supposed to fix! If the cancellation takes more than 5 seconds, then your code will continue and free the RPC_ASYNC_STATE, just like it did when you didn't wait at all.

"How long can I wait before assuming that the event will simply never get signaled?"

There is no such duration after which you can safely abandon the operation. Even if the event doesn't get signaled for 30 minutes (say because the computer is thrashing its guts out), it may get signaled at 30 minutes and 1 second.

"But we don't want our program to get stuck waiting for the server."

Great. It's fine to have your program continues running after issuing the cancellation, even if the RPC call hasn't completed. Just don't free the RPC_ASYNC_STATE until the call is complete. and if you set things up so that your completion event takes the form of a callback, you can just make the callback free the RPC__ASYNC_STATE. Then you don't have to keep track of the asynchronous call any more; the system will merely call you when it's finished, and then you can free the state structure.

Bonus RPC chatter: (For the purpose of this discussion, I'll use the term RPC operation instead of RPC call so we don't have confusion between function calls and RPC calls.) A colleague explained the lifetime of an RPC operation as follows:

Submit phase You call into the MIDL-generated stub. You cannot call Rpc­Async­Cancel­Call during the submit phase.
The stub does magic RPC stuff.
The stub returns control back to the caller.
Pending phase RPC is waiting for the response to the operation. The operation remains in this phase until the operation completes or is cancelled. You can call Rpc­Async­Cancel­Call to cancel the RPC operation and accelerate the transition to the Notified phase.
Notified phase RPC informs the application of the result of the operation in a manner described by the Notification­Type and RPC_ASYNC_NOTIFICATION_INFO members of the RPC_ASYNC_STATE structure. You can call Rpc­Async­Cancel­Call but it will have no effect since the operation is already complete.
Completion phase The application calls the Rpc­Async­Complete­Call function to clean up the resources used to track the RPC operation. You exit the completion phase when Rpc­Async­Complete­Call returns something other than RPC_S_ASYNC_CALL_PENDING. You cannot call Rpc­Async­Cancel­Call after Rpc­Async­Complete­Call indicates that the operation is complete, since that is the call that says "I'm all done!"
Comments (16)
  1. Alex Grigoriev says:
    1. Write some stuff in MIDL language.
    2. Issue RPC call.

    1. Profit!

    2. Go to ATM machine and enter your PIN number…

  2. Adam Rosenfield says:

    @Alex: I was going to mumble something about RAS syndrome, but you put it way better than I possibly could have.

  3. Tanveer Badar says:

    Do you crucify the people who utter things like "garbage collection", "managed memory" and CLR?

    P.S.: Lack of comments is astounding.

  4. Mathieu Garstecki says:

    @Tanveer: There must be a nitpickers' convention somewhere this week.

  5. jader3rd says:

    I wish there was an FxCop rule which forbid boolean parameters. They virtually never make sense to someone reading the code that calls the method.

  6. Mason Wheeler says:

    "fAbort­Call"?  I don't think I've ever seen an f prepended to a boolean parameter before.  What does it stand for?

  7. Paul M. Parks says:

    @Mason: "Flag"

  8. Joshua says:

    FYI, Garbage collection doesn't help here as garbage collection can't see pointers in kernel.

  9. Antonio Rodríguez says:

    @Joshua: Then simply rewrite the whole kernel in managed code… Ah, yes, I forgot this is Windows, not Singularity…

  10. Mr Pickles says:

    It's about time Microsoft did something about this.

  11. Joshua says:

    @Mr. Pickles: no, it isn't.

    These problems are normal to cancellable async. I/O. If you can't stomach it, don't use it. Memory's cheap. Use threads for cancellable sync. I/O.

    I know UNIX has another solution involving select(), but that's its own headache until you get it right.

  12. Gechurch says:

    Unbelievable! I'm used to seeing customers on Raymond's blog that miss something obvious, or do something dumb. But this customer has been told the problem in their code (they aren't waiting until the async call is cancelled before cleaning it up), then they go and changed the fixed code to have a timeout! And then they re-open a case with Microsoft!

    "Hey Microsoft, you know how we had that bug where we weren't waiting for the async event to cancel? Well your fixed worked for a while, but now that we've rewritten the code to only wait five seconds and then to clean up even if the call hasn't been cancelled the bug has come back. What could possibly be causing it?".


  13. Simon says:

    @jader3rd – absolutely correct. Valid for setting a property (assuming a language without native property syntax), but I'm even now in the middle of cleaning up some code featuring object constructors with up to 4 boolean parameters. Not exactly readable – I want to replace them with enums, assuming the settings cannot be removed entirely.

  14. Cheong says:

    @Joshua: I'm always hesitate to choose between 1) having asynchronous I/O; and 2) using synchronous I/O on a pool of threads. Is there any guideline on choosing which approach to use?

  15. Alex Grigoriev says:


    If a file handle is opened for synchronous IO (no OVERLAPPED flag), pool of threads is pointless, because only one IO can be in progress on such a handle.

  16. Cheong says:

    @Alex: My question comparing between pool of thread (or perheps better saying using background thread instead) is for the sake of responsiveness.

    But by what you said the result is clear.

Comments are closed.

Skip to main content