On Fri, 14 Jun 2024 09:09:58 GMT, Daniel Jeliński <djelin...@openjdk.org> wrote:
>> Hi >> >> This change is to improve TLS 1.3 session resumption by allowing a TLS >> server to send more than one resumption ticket per connection and clients to >> store more. Resumption is a quick way to use an existing TLS session to >> establish another session by avoiding the long TLS full handshake process. >> In TLS 1.2 and below, clients can repeatedly resume a session by using the >> session ID from an established connection. In TLS 1.3, a one-time >> "resumption ticket" is sent by the server after the TLS connection has been >> established. The server may send multiple resumption tickets to help >> clients that rapidly resume connections. If the client does not have >> another resumption ticket, it must go through the full TLS handshake again. >> The current implementation in JDK 23 and below, only sends and store one >> resumption ticket. >> >> The number of resumption tickets a server can send should be configurable by >> the application developer or administrator. [RFC >> 8446](https://www.rfc-editor.org/rfc/rfc8446) does not specify a default >> value. A system property called `jdk.tls.server.newSessionTicketCount` >> allows the user to change the number of resumption tickets sent by the >> server. If this property is not set or given an invalid value, the default >> value of 3 is used. Further details are in the CSR. >> >> A large portion of the changeset is on the client side by changing the >> caching system used by TLS. It creates a new `CacheEntry<>` type called >> `QueueCacheEntry<>` that will store multiple values for a Map entry. > > src/java.base/share/classes/sun/security/ssl/NewSessionTicket.java line 397: > >> 395: * and server are on different machines. >> 396: */ >> 397: Thread nstThread = Thread.ofVirtual().name("NST").start(() >> -> { > > Please don't use threads during handshake. There is no alternative that I have found for this synchronization/timing situation. We certainly don't want a `sleep()` call and NSTs are not send/ack situation. If the client ignores the NST, that is fine. Hung thread paranoia is the only reason I put the `join()` in the code below as when this finish isn't critical. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/19465#discussion_r1640129880