[PATCH] mieq: Provide better adaptability and diagnostics during mieq overflow

[Jeremy Huddleston]

On Oct 16, 2011, at 8:44 AM, Keith Packard wrote:

> On Sat, 15 Oct 2011 23:09:30 -0700, Jeremy Huddleston <jeremyhu at apple.com> wrote:
> 
>> +	if (new_queue == NULL)
>> +		FatalError("Unable to allocate memory for the event queue.\n");
> 
> This function should be boolean and return whether the queue was
> resized, instead of failing when out of memory.

Ok.

>> +	miEventQueue.tail = (miEventQueue.tail - miEventQueue.head) % miEventQueue.q_size;
> 
> Shouldn't this just be miEventQueue.q_size? If not, head and tail are
> both 'int' type, so the subtraction can be negative, in which case the %
> operator is undefined (thanks, dmr!).

No.  miEventQueue.q_size is the number of buckets.  (miEventQueue.tail - miEventQueue.head) is the number of filled buckets (modulo number of buckets).

% is undefined for negative numbers?  WTF?  That's the first time I've heard that.  It's certainly well defined in mathematics.

> 
> 
>> +			ErrorF("[mi] EQ overflowing.  Increasing queue size to %lu to accomidate.\n", miEventQueue.q_size << 1);
> 
> accommodate.

That was removed in v2 already.

> 
>> +			if (miEventQueue.dropped == 1) {
>> +				ErrorF("[mi] EQ overflowing.  Maximum queue size reached.  Events will be discarded.\n");
>> +	           	xorg_backtrace();
>> +			} else if (miEventQueue.dropped % DROP_BACKTRACE_FREQUENCY == 0 &&
>> +			           miEventQueue.dropped / DROP_BACKTRACE_FREQUENCY <= DROP_BACKTRACE_MAX) {
>> +				ErrorF("[mi] EQ overflow continuing.  %lu events have been dropped.\n", miEventQueue.dropped);
>> +				if (miEventQueue.dropped / DROP_BACKTRACE_FREQUENCY == DROP_BACKTRACE_MAX) {
>> +					ErrorF("[mi] No further overflow reports will be reported until the clog is cleared.\n");
>> +				}
>> +	           	xorg_backtrace();
> 
> I'm not sure we need to be even this verbose; just a single warning when
> the queue is blocked and another one when it starts working again should
> be sufficient. I fear flooding the disk with errors if this happens
> because your cat falls asleep on the keyboard while watching dancing
> mice on the screen.

That's why we shut up after DROP_BACKTRACE_MAX reports.  What about lowering DROP_BACKTRACE_MAX to 5 rather than 20?