Skip to content
  • Shlomo Pongratz's avatar
    [SCSI] libiscsi: Reduce locking contention in fast path · 659743b0
    Shlomo Pongratz authored
    
    
    Replace the session lock with two locks, a forward lock and
    a backwards lock named frwd_lock and back_lock respectively.
    
    The forward lock protects resources that change while sending a
    request to the target, such as cmdsn, queued_cmdsn, and allocating
    task from the commands' pool with kfifo_out.
    
    The backward lock protects resources that change while processing
    a response or in error path, such as cmdsn_exp, cmdsn_max, and
    returning tasks to the commands' pool with kfifo_in.
    
    Under a steady state fast-path situation, that is when one
    or more processes/threads submit IO to an iscsi device and
    a single kernel upcall (e.g softirq) is dealing with processing
    of responses without errors, this patch eliminates the contention
    between the queuecommand()/request response/scsi_done() flows
    associated with iscsi sessions.
    
    Between the forward and the backward locks exists a strict locking
    hierarchy. The mutual exclusion zone protected by the forward lock can
    enclose the mutual exclusion zone protected by the backward lock but not
    vice versa.
    
    For example, in iscsi_conn_teardown or in iscsi_xmit_data when there is
    a failure and __iscsi_put_task is called, the backward lock is taken while
    the forward lock is still taken. On the other hand, if in the RX path a nop
    is to be sent, for example in iscsi_handle_reject or __iscsi_complete_pdu
    than the forward lock is released and the backward lock is taken for the
    duration of iscsi_send_nopout, later the backward lock is released and the
    forward lock is retaken.
    
    libiscsi_tcp uses two kernel fifos the r2t pool and the r2t queue.
    
    The insertion and deletion from these queues didn't corespond to the
    assumption taken by the new forward/backwards session locking paradigm.
    
    That is, in iscsi_tcp_clenup_task which belongs to the RX (backwards)
    path, r2t is taken out from r2t queue and inserted to the r2t pool.
    In iscsi_tcp_get_curr_r2t which belong to the TX (forward) path, r2t
    is also inserted to the r2t pool and another r2t is pulled from r2t
    queue.
    
    Only in iscsi_tcp_r2t_rsp which is called in the RX path but can requeue
    to the TX path, r2t is taken from the r2t pool and inserted to the r2t
    queue.
    
    In order to cope with this situation, two spin locks were added,
    pool2queue and queue2pool. The former protects extracting from the
    r2t pool and inserting to the r2t queue, and the later protects the
    extracing from the r2t queue and inserting to the r2t pool.
    
    Signed-off-by: default avatarShlomo Pongratz <shlomop@mellanox.com>
    Signed-off-by: default avatarOr Gerlitz <ogerlitz@mellanox.com>
    [minor fix up to apply cleanly and compile fix]
    Signed-off-by: default avatarMike Christie <michaelc@cs.wisc.edu>
    Signed-off-by: default avatarJames Bottomley <JBottomley@Parallels.com>
    659743b0