Implementer Notes.txt   [plain text]


February 2002:

The mDNSResponder code has a slight architectural change to improve
efficiency.

The mDNSResponder code previously called ScheduleNextTask() after every
operation, to calculate the time at which it needed to be called back to
perform its next timed operation. When the workload is light, and
protocol operations are rare and far apart, this makes sense.

However, on networks where there is a lot of mDNS traffic (or the CPU is
slow), this leads to the following anomolous behaviour: mDNSResponder
spends a lot of CPU time working out what to do next, when what it needs
to do next should be obvious: Finish processing the big backlog of
packets that have been received.

To remedy this, mDNSResponder now only executes ScheduleNextTask() when
there is no other obvious work waiting to be done. However, the
mDNSResponder code does not have direct access to this knowledge. Only
the platform layer below knows whether there are packets waiting to be
processed. Only the client layer above knows whether it is in the
process of performing a long sequence of back-to-back mDNS API calls.

This means that the new architecture places an additional responsibility
on the client layer and/or platform support layer. As long as they have
immediate work to do, they should call the appropriate mDNSCore routines
to accomplish that work. With each call, mDNSCore will do only what it
immediately has to do to satisfy the call. Any optional work will be
deferred. As soon as there is no more immediate work to do, the calling
layer MUST call mDNS_Execute(). Failure to call mDNS_Execute() will lead
to unreliable or incorrect operation.

The value returned from mDNS_Execute() is the next time (in absolute
platform time units) at which mDNS_Execute() MUST be called again to
perform its next necessary operation (e.g. transmitting its next
scheduled query packet, etc.) Note that the time returned is an absolute
time, not the time *interval* between now and the next required call.
For OS APIs that work in terms of intervals instead of absolute times,
mDNSPlatformTimeNow() must be subtracted from the absolute time to get
the interval between now and the next event.

In a single-threaded application using a blocking select() call as its
main synchronization point, this means that you should call
mDNS_Execute() before calling select(), and the timeout value you pass
to select() MUST NOT be larger than that indicated by the result
returned from mDNS_Execute(). After the blocking select() call returns,
you should do whatever work you have to do, and then, if mDNS packets
were received, or mDNS API calls were made, be sure to call
mDNS_Execute() again, and if necessary adjust your timeout value
accordingly, before going back into the select() call.

In an asynchronous or interrupt-driven application, there are three
places that should call mDNS_Execute():

1. After delivering received packets, the platform support layer should
call mDNS_Execute(), and use the value returned to set the platform
callback timer to fire at the indicated time.

2. After making any mDNS API call or series of calls, the client layer
should call mDNS_Execute(), and use the value returned to set the
platform callback timer to fire at the indicated time.

3. When the platform callback timer fires, it should call mDNS_Execute()
(to allow mDNSCore to perform its necessary work) and then the timer
routine use the result returned to reset itself to fire at the right
time for the next scheduled event.