http://en.wikipedia.org/wiki/Ricart-Agrawala_algorithm
From that description it seems quite clear to me.
Any node requesting exclusivity will only proceed to seize it when and iff
all other nodes have unanimously agreed that it may do so, this is determined
by the requestor by the successful receipt of all other nodes' permission grant
messages.
When any node receives another node's request it will issue a permission
grant response to that node when and iff
the recipient node is not itself presently holding exclusivity, and
the recipient node has no pending and temporally precedent request for
exclusivity for itself.
Thus the exclusivity grants are serviced on a "first requested, first served"
basis with all nodes en masse ceding the spot at the head of the request
queue to the node with the first timestamped pending request.
A pending request cannot be granted until there's ubiquity of assent, so
only nodes temporarily withholding their assent due to their own
exclusivity or precedent request will delay (beyond round-trip communication
latency time) another node's request being granted.
Situations to be concerned about are handling of simultaneous requests
(wrt. the timestamp) from multiple nodes, as well as handling network/node failure
situations since any failed node / link could prohibit all other nodes from
acquiring exclusivity absent assent from unreachable nodes.
Simultaneous requests could just be handled by all nodes ceding precedence
to their peers and then reissuing their own requests after some small random
or delay. Or nodes could take precedence in such simultaneous request
situations according to some fixed node associated priority value (node number etc).