[PATCH v4 1/8] server: Actually set initial status in set_async_direct_result handler.

[Jinoh Kang]

On 3/5/22 04:04, Jinoh Kang wrote:
> On 3/5/22 02:47, Zebediah Figura wrote:
>> Right, but I think it's possible to trigger this assert now. For a rather pathological case, consider a client which calls set_async_direct_result on an async which is currently being serviced by a device driver and has not yet reached get_next_device_request.
> 
> For that to be possible, something has to terminate the async so that set_async_direct_result does not reject it.
> The easiest way to do so is to cancel it, or make it timeout somehow.
> 
> In fact, I think we have discovered an _existing_ bug: the client can use set_async_direct_result to interfere with cancelled asyncs (device-backed or not).
> 
> Further investigation revealed some other bugs, which is tangentially related:
> 
> 1. An async may time out before the client calls set_async_direct_result to report back the status of initial I/O.  In this case, the timeout is ignored.
> 2. There is a time window where alerted asyncs ignore IoCancel/IoCancelEx requests.  If async_terminate( async, STATUS_ALERTED ) is called, the async will ignore any cancellation requests until the APC_ASYNC_IO routine returns.
>    Note that this bug has existed *before* set_async_direct_result, and I think this one may need to be addressed first.
> 
> So what I think we need here is a mechanism to tell alerted asyncs apart from irreversibly-terminated asyncs.
> This will solve all of the problems above:
> 
> - It's no longer possible to call set_async_direct_result on a device async in the first place, since it will now reject irreversibly-terminated asyncs while still accepting alerted asyncs (which is not possible for device asyncs).
> - Timeout and cancellation works as expected: set_async_direct_result will treat timed out asyncs as irreversibly-terminated.  (the client needs to be modified to handle this case explicitly, though.)

Would it still be worth working on this?  I couldn't find the bug report for this, though.

-- 
Sincerely,
Jinoh Kang