On Thu, Jan 2, 2020 at 1:13 PM Gabriel Krisman Bertazi
<krisman@xxxxxxxxxxxxx> wrote:
>
> Khazhismel Kumykov <khazhy@xxxxxxxxxx> writes:
>
> > On Thu, Dec 26, 2019 at 3:48 PM Gabriel Krisman Bertazi
> > <krisman@xxxxxxxxxxxxx> wrote:
> >>
> >> From: Bharath Ravi <rbharath@xxxxxxxxxx>
> >>
> >> Connection failure processing depends on a daemon being present to (at
> >> least) stop the connection and start recovery. This is a problem on a
> >> multipath scenario, where if the daemon failed for whatever reason, the
> >> SCSI path is never marked as down, multipath won't perform the
> >> failover and IO to the device will be forever waiting for that
> >> connection to come back.
> >>
> >> This patch performs the connection failure entirely inside the kernel.
> >> This way, the failover can happen and pending IO can continue even if
> >> the daemon is dead. Once the daemon comes alive again, it can execute
> >> recovery procedures if applicable.
> >>
> >> Changes since v2:
> >> - Don't hold rx_mutex for too long at once
> >>
> >> Changes since v1:
> >> - Remove module parameter.
> >> - Always do kernel-side stop work.
> >> - Block recovery timeout handler if system is dying.
> >> - send a CONN_TERM stop if the system is dying.
> >>
> >> Cc: Mike Christie <mchristi@xxxxxxxxxx>
> >> Cc: Lee Duncan <LDuncan@xxxxxxxx>
> >> Cc: Bart Van Assche <bvanassche@xxxxxxx>
> >> Co-developed-by: Dave Clausen <dclausen@xxxxxxxxxx>
> >> Signed-off-by: Dave Clausen <dclausen@xxxxxxxxxx>
> >> Co-developed-by: Nick Black <nlb@xxxxxxxxxx>
> >> Signed-off-by: Nick Black <nlb@xxxxxxxxxx>
> >> Co-developed-by: Vaibhav Nagarnaik <vnagarnaik@xxxxxxxxxx>
> >> Signed-off-by: Vaibhav Nagarnaik <vnagarnaik@xxxxxxxxxx>
> >> Co-developed-by: Anatol Pomazau <anatol@xxxxxxxxxx>
> >> Signed-off-by: Anatol Pomazau <anatol@xxxxxxxxxx>
> >> Co-developed-by: Tahsin Erdogan <tahsin@xxxxxxxxxx>
> >> Signed-off-by: Tahsin Erdogan <tahsin@xxxxxxxxxx>
> >> Co-developed-by: Frank Mayhar <fmayhar@xxxxxxxxxx>
> >> Signed-off-by: Frank Mayhar <fmayhar@xxxxxxxxxx>
> >> Co-developed-by: Junho Ryu <jayr@xxxxxxxxxx>
> >> Signed-off-by: Junho Ryu <jayr@xxxxxxxxxx>
> >> Co-developed-by: Khazhismel Kumykov <khazhy@xxxxxxxxxx>
> >> Signed-off-by: Khazhismel Kumykov <khazhy@xxxxxxxxxx>
> >> Signed-off-by: Bharath Ravi <rbharath@xxxxxxxxxx>
> >> Co-developed-by: Gabriel Krisman Bertazi <krisman@xxxxxxxxxxxxx>
> >> Signed-off-by: Gabriel Krisman Bertazi <krisman@xxxxxxxxxxxxx>
> >> ---
> >> drivers/scsi/scsi_transport_iscsi.c | 63 +++++++++++++++++++++++++++++
> >> include/scsi/scsi_transport_iscsi.h | 1 +
> >> 2 files changed, 64 insertions(+)
> >>
> >> diff --git a/drivers/scsi/scsi_transport_iscsi.c b/drivers/scsi/scsi_transport_iscsi.c
> >> index 271afea654e2..c6db6ded60a1 100644
> >> --- a/drivers/scsi/scsi_transport_iscsi.c
> >> +++ b/drivers/scsi/scsi_transport_iscsi.c
> >> @@ -86,6 +86,12 @@ struct iscsi_internal {
> >> struct transport_container session_cont;
> >> };
> >>
> >> +/* Worker to perform connection failure on unresponsive connections
> >> + * completely in kernel space.
> >> + */
> >> +static void stop_conn_work_fn(struct work_struct *work);
> >> +static DECLARE_WORK(stop_conn_work, stop_conn_work_fn);
> >> +
> >> static atomic_t iscsi_session_nr; /* sysfs session id for next new session */
> >> static struct workqueue_struct *iscsi_eh_timer_workq;
> >>
> >> @@ -1611,6 +1617,7 @@ static DEFINE_MUTEX(rx_queue_mutex);
> >> static LIST_HEAD(sesslist);
> >> static DEFINE_SPINLOCK(sesslock);
> >> static LIST_HEAD(connlist);
> >> +static LIST_HEAD(connlist_err);
> >> static DEFINE_SPINLOCK(connlock);
> >>
> >> static uint32_t iscsi_conn_get_sid(struct iscsi_cls_conn *conn)
> >> @@ -2247,6 +2254,7 @@ iscsi_create_conn(struct iscsi_cls_session *session, int dd_size, uint32_t cid)
> >>
> >> mutex_init(&conn->ep_mutex);
> >> INIT_LIST_HEAD(&conn->conn_list);
> >> + INIT_LIST_HEAD(&conn->conn_list_err);
> >> conn->transport = transport;
> >> conn->cid = cid;
> >>
> >> @@ -2293,6 +2301,7 @@ int iscsi_destroy_conn(struct iscsi_cls_conn *conn)
> >>
> >> spin_lock_irqsave(&connlock, flags);
> >> list_del(&conn->conn_list);
> >> + list_del(&conn->conn_list_err);
> >> spin_unlock_irqrestore(&connlock, flags);
> >>
> >> transport_unregister_device(&conn->dev);
> >> @@ -2407,6 +2416,51 @@ int iscsi_offload_mesg(struct Scsi_Host *shost,
> >> }
> >> EXPORT_SYMBOL_GPL(iscsi_offload_mesg);
> >>
> >> +static void stop_conn_work_fn(struct work_struct *work)
> >> +{
> >> + struct iscsi_cls_conn *conn, *tmp;
> >> + unsigned long flags;
> >> + LIST_HEAD(recovery_list);
> >> +
> >> + spin_lock_irqsave(&connlock, flags);
> >> + if (list_empty(&connlist_err)) {
> >> + spin_unlock_irqrestore(&connlock, flags);
> >> + return;
> >> + }
> >> + list_splice_init(&connlist_err, &recovery_list);
> >> + spin_unlock_irqrestore(&connlock, flags);
> >> +
> >> + list_for_each_entry_safe(conn, tmp, &recovery_list, conn_list_err) {
> >> + uint32_t sid = iscsi_conn_get_sid(conn);
> >> + struct iscsi_cls_session *session;
> >> +
> >> + mutex_lock(&rx_queue_mutex);
> > This worried me a bit, but it seems we won't destroy_conn while it's
> > on the err list - cool.
> >> +
> >> + session = iscsi_session_lookup(sid);
> >> + if (session) {
> >> + if (system_state != SYSTEM_RUNNING) {
> >> + session->recovery_tmo = 0;
> >> + conn->transport->stop_conn(conn,
> >> + STOP_CONN_TERM);
> >> + } else {
> >> + conn->transport->stop_conn(conn,
> >> + STOP_CONN_RECOVER);
> >> + }
> >> + }
> >> +
> >> + list_del_init(&conn->conn_list_err);
> >> +
> >> + mutex_unlock(&rx_queue_mutex);
> >> +
> >> + /* we don't want to hold rx_queue_mutex for too long,
> >> + * for instance if many conns failed at the same time,
> >> + * since this stall other iscsi maintenance operations.
> >> + * Give other users a chance to proceed.
> >> + */
> >> + cond_resched();
> >> + }
> >> +}
> >> +
> >> void iscsi_conn_error_event(struct iscsi_cls_conn *conn, enum iscsi_err error)
> >> {
> >> struct nlmsghdr *nlh;
> >> @@ -2414,6 +2468,12 @@ void iscsi_conn_error_event(struct iscsi_cls_conn *conn, enum iscsi_err error)
> >> struct iscsi_uevent *ev;
> >> struct iscsi_internal *priv;
> >> int len = nlmsg_total_size(sizeof(*ev));
> >> + unsigned long flags;
> >> +
> >> + spin_lock_irqsave(&connlock, flags);
> >> + list_add(&conn->conn_list_err, &connlist_err);
> >> + spin_unlock_irqrestore(&connlock, flags);
> >> + queue_work(system_unbound_wq, &stop_conn_work);
> >>
> >> priv = iscsi_if_transport_lookup(conn->transport);
> >> if (!priv)
> >> @@ -2748,6 +2808,9 @@ iscsi_if_destroy_conn(struct iscsi_transport *transport, struct iscsi_uevent *ev
> >> if (!conn)
> >> return -EINVAL;
> >>
> >> + if (!list_empty(&conn->conn_list_err))
> > Does this check need to be under connlock?
>
> My understanding is that it is not necessary, since it is serialized
> against the conn removal itself, through the rx_mutex, it seemed safe to
> do the verification lockless.
>
> It can only race with the insertion, in which case, it will be safely
> removed from the dispatch list here, under rx_mutex, and the worker will
> detect and skipped it.
My worry is the splice, which is under only connlock, not rx_mutex, which
might lead to UB if we're checking empty while modifying the list_head ?
>
> --
> Gabriel Krisman Bertazi
Attachment:
smime.p7s
Description: S/MIME Cryptographic Signature
