On Tue, Dec 07, 2021 at 01:59:08PM -0600, Mike Christie wrote: > On 11/11/21 7:37 AM, Maurizio Lombardi wrote: > > When the connection reinstatement is performed, the target driver > > executes a complex scheme of complete()/wait_for_completion() that is not > > really needed. > > > > Considering that: > > > > 1) The callers of iscsit_connection_reinstatement_rcfr() and > > iscsit_cause_connection_reinstatement() hold a reference > > to the conn structure.> > 2) iscsit_close_connection() will sleep when calling > > iscsit_check_conn_usage_count() until the conn structure's refcount > > reaches zero. > > > > we can optimize the driver the following way: > > > > * The threads that must sleep until the connection is closed > > will all wait for the "conn_wait_comp" completion, > > iscsit_close_connection() will then call complete_all() to wake them up. > > No need to have multiple completion structures. > > > > * The conn_post_wait_comp completion is not necessary and can be removed > > because iscsit_close_connection() sleeps until all the other threads > > release the conn structure. > > (see the iscsit_check_conn_usage_count() function) > > > > V2: do not set connection_reinstatement to 1 in iscsit_close_connection(), > > leave iscsit_cause_connection_reinstatement() deal with reentrancy. > > > > What was the issue with setting it to 1? > At first I thought that setting it to 1 could introduce a race condition, then I realized that it's not the case, setting it to 1 simply has no effect and therefore the assignment can be removed. Maurizio
