Hi,
In short the server stops reading data from the connection until it gets working threads to process them. (So some operations get queued in the kernel TCP/IP buffers but quickly the server stops responding until it can process new operations) We sometimes see that behavior when a very long write operation occurs and all the other threads collide on the database txn locks.
The only thing that is queued is the fact that there is network activity on a connection.  There is no explicit size limit to this queue. 
 But its size is always smaller than  nsslapd-maxthreadsperconn * "max number of connections" (which is itself limited to 64K or by the number of available file descriptors)
   
So the important parameters are:
nsslapd-threadnumber that determines how many threads can process operation.
And nsslapd-maxthreadsperconn that determines how many concurrent operations may run on a single connection.
The process "nofile" hard limit (usually set in systemd config file)