scsi: iscsi: Run recv path from workqueue

We don't always want to run the recv path from the network softirq because
when we have to have multiple sessions sharing the same CPUs, some sessions
can eat up the NAPI softirq budget and affect other sessions or users.

Allow us to queue the recv handling to the iscsi workqueue so we can have
the scheduler/wq code try to balance the work and CPU use across all
sessions' worker threads.

Note: It wasn't the original intent of the change but a nice side effect is
that for some workloads/configs we get a nice performance boost. For a
simple read heavy test:

  fio --direct=1 --filename=/dev/dm-0  --rw=randread --bs=256K
    --ioengine=libaio --iodepth=128 --numjobs=4

where the iscsi threads, fio jobs, and rps_cpus share CPUs we see a 32%
throughput boost. We also see increases for small I/O IOPs tests but it's
not as high.

Link: https://lore.kernel.org/r/20220616224557.115234-4-michael.christie@oracle.com
Reviewed-by: Lee Duncan <lduncan@suse.com>
Signed-off-by: Mike Christie <michael.christie@oracle.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
index da1dc345b8739b1bd183cbae0cd94a55144d1df4..10d7f2b7dd0e1fa44a89711a49c8cf941a257a47 100644 (file)
--- a/drivers/scsi/iscsi_tcp.c
+++ b/drivers/scsi/iscsi_tcp.c
@@ -52,6 +52,10 @@ static struct iscsi_transport iscsi_sw_tcp_transport;
 static unsigned int iscsi_max_lun = ~0;
 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
 
+static bool iscsi_recv_from_iscsi_q;
+module_param_named(recv_from_iscsi_q, iscsi_recv_from_iscsi_q, bool, 0644);
+MODULE_PARM_DESC(recv_from_iscsi_q, "Set to true to read iSCSI data/headers from the iscsi_q workqueue. The default is false which will perform reads from the network softirq context.");
+
 static int iscsi_sw_tcp_dbg;
 module_param_named(debug_iscsi_tcp, iscsi_sw_tcp_dbg, int,
                   S_IRUGO | S_IWUSR);
@@ -122,20 +126,13 @@ static inline int iscsi_sw_sk_state_check(struct sock *sk)
        return 0;
 }
 
-static void iscsi_sw_tcp_data_ready(struct sock *sk)
+static void iscsi_sw_tcp_recv_data(struct iscsi_conn *conn)
 {
-       struct iscsi_conn *conn;
-       struct iscsi_tcp_conn *tcp_conn;
+       struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
+       struct iscsi_sw_tcp_conn *tcp_sw_conn = tcp_conn->dd_data;
+       struct sock *sk = tcp_sw_conn->sock->sk;
        read_descriptor_t rd_desc;
 
-       read_lock_bh(&sk->sk_callback_lock);
-       conn = sk->sk_user_data;
-       if (!conn) {
-               read_unlock_bh(&sk->sk_callback_lock);
-               return;
-       }
-       tcp_conn = conn->dd_data;
-
        /*
         * Use rd_desc to pass 'conn' to iscsi_tcp_recv.
         * We set count to 1 because we want the network layer to
@@ -144,13 +141,48 @@ static void iscsi_sw_tcp_data_ready(struct sock *sk)
         */
        rd_desc.arg.data = conn;
        rd_desc.count = 1;
-       tcp_read_sock(sk, &rd_desc, iscsi_sw_tcp_recv);
 
-       iscsi_sw_sk_state_check(sk);
+       tcp_read_sock(sk, &rd_desc, iscsi_sw_tcp_recv);
 
        /* If we had to (atomically) map a highmem page,
         * unmap it now. */
        iscsi_tcp_segment_unmap(&tcp_conn->in.segment);
+
+       iscsi_sw_sk_state_check(sk);
+}
+
+static void iscsi_sw_tcp_recv_data_work(struct work_struct *work)
+{
+       struct iscsi_conn *conn = container_of(work, struct iscsi_conn,
+                                              recvwork);
+       struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
+       struct iscsi_sw_tcp_conn *tcp_sw_conn = tcp_conn->dd_data;
+       struct sock *sk = tcp_sw_conn->sock->sk;
+
+       lock_sock(sk);
+       iscsi_sw_tcp_recv_data(conn);
+       release_sock(sk);
+}
+
+static void iscsi_sw_tcp_data_ready(struct sock *sk)
+{
+       struct iscsi_sw_tcp_conn *tcp_sw_conn;
+       struct iscsi_tcp_conn *tcp_conn;
+       struct iscsi_conn *conn;
+
+       read_lock_bh(&sk->sk_callback_lock);
+       conn = sk->sk_user_data;
+       if (!conn) {
+               read_unlock_bh(&sk->sk_callback_lock);
+               return;
+       }
+       tcp_conn = conn->dd_data;
+       tcp_sw_conn = tcp_conn->dd_data;
+
+       if (tcp_sw_conn->queue_recv)
+               iscsi_conn_queue_recv(conn);
+       else
+               iscsi_sw_tcp_recv_data(conn);
        read_unlock_bh(&sk->sk_callback_lock);
 }
 
@@ -276,6 +308,9 @@ static int iscsi_sw_tcp_xmit_segment(struct iscsi_tcp_conn *tcp_conn,
                if (segment->total_copied + segment->size < segment->total_size)
                        flags |= MSG_MORE;
 
+               if (tcp_sw_conn->queue_recv)
+                       flags |= MSG_DONTWAIT;
+
                /* Use sendpage if we can; else fall back to sendmsg */
                if (!segment->data) {
                        sg = segment->sg;
@@ -557,6 +592,8 @@ iscsi_sw_tcp_conn_create(struct iscsi_cls_session *cls_session,
        conn = cls_conn->dd_data;
        tcp_conn = conn->dd_data;
        tcp_sw_conn = tcp_conn->dd_data;
+       INIT_WORK(&conn->recvwork, iscsi_sw_tcp_recv_data_work);
+       tcp_sw_conn->queue_recv = iscsi_recv_from_iscsi_q;
 
        tfm = crypto_alloc_ahash("crc32c", 0, CRYPTO_ALG_ASYNC);
        if (IS_ERR(tfm))
@@ -610,6 +647,8 @@ static void iscsi_sw_tcp_release_conn(struct iscsi_conn *conn)
        iscsi_sw_tcp_conn_restore_callbacks(conn);
        sock_put(sock->sk);
 
+       iscsi_suspend_rx(conn);
+
        spin_lock_bh(&session->frwd_lock);
        tcp_sw_conn->sock = NULL;
        spin_unlock_bh(&session->frwd_lock);

diff --git a/drivers/scsi/iscsi_tcp.h b/drivers/scsi/iscsi_tcp.h
index 791453195099c0f67a4299d5e4ec0596e5df8a0a..850a018aefb9bb6c41aed65f047f42834e751641 100644 (file)
--- a/drivers/scsi/iscsi_tcp.h
+++ b/drivers/scsi/iscsi_tcp.h
@@ -28,6 +28,8 @@ struct iscsi_sw_tcp_send {
 
 struct iscsi_sw_tcp_conn {
        struct socket           *sock;
+       struct work_struct      recvwork;
+       bool                    queue_recv;
 
        struct iscsi_sw_tcp_send out;
        /* old values for socket callbacks */