[Xorp-cvs] XORP cvs commit: xorp/libxorp

[Bruce M Simpson]

Pavlin Radoslavov wrote:
> 	However, inside SelectorList::Node::add_okay() Step 2, the original
> 	code selected the first _empty_ bin to store the information which
> 	seems inconsistent with the above-mentioned assumption.
>   

This issue was exposed in my testing for bug 799 on Friday, as you 
probably saw. The code was using an invalid SEL_XX_IDX of -1 in some 
situations.

> 	
> 	After the fix, READ I/O registration goes into [SEL_RD_IDX],
> 	WRITE I/O registration goes into [SEL_WR_IDX], and so on.
> 	The rest of the Selector code is not modified: it still looks
> 	for various masks in all array bins which is sub-optimal.
> 	
> 	A side note: internally, we don't need the concept of
> 	SelectorMask, because outside of the EventLoop we can
> 	register only for a particular event type: IOT_READ, IOT_WRITE, etc.
> 	The only exception is IOT_ANY, and that one is used only
> 	when remove_ioevent_cb() is called to remove all callbacks for
> 	a file descriptor.
> 	If SelectorMask is removed, this will simplify the implementation,
> 	and will speed-up the eventloop processing.
>   

Yes, that can probably go... One of the goals for Windows refactoring 
was to avoid modifying the UNIX code path if at all possible.

thanks
BMS