#include "socklnd.h"
# if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
-static cfs_sysctl_table_t ksocknal_ctl_table[23];
-cfs_sysctl_table_t ksocknal_top_ctl_table[] = {
- {
- .ctl_name = 200,
- .procname = "socknal",
- .data = NULL,
- .maxlen = 0,
- .mode = 0555,
- .child = ksocknal_ctl_table
- },
- { 0 }
+#ifndef HAVE_SYSCTL_UNNUMBERED
+
+enum {
+ SOCKLND_TIMEOUT = 1,
+ SOCKLND_CREDITS,
+ SOCKLND_PEER_CREDITS,
+ SOCKLND_NCONNDS,
+ SOCKLND_RECONNECTS_MIN,
+ SOCKLND_RECONNECTS_MAX,
+ SOCKLND_EAGER_ACK,
+ SOCKLND_ZERO_COPY,
+ SOCKLND_TYPED,
+ SOCKLND_BULK_MIN,
+ SOCKLND_RX_BUFFER_SIZE,
+ SOCKLND_TX_BUFFER_SIZE,
+ SOCKLND_NAGLE,
+ SOCKLND_IRQ_AFFINITY,
+ SOCKLND_ROUND_ROBIN,
+ SOCKLND_KEEPALIVE,
+ SOCKLND_KEEPALIVE_IDLE,
+ SOCKLND_KEEPALIVE_COUNT,
+ SOCKLND_KEEPALIVE_INTVL,
+ SOCKLND_BACKOFF_INIT,
+ SOCKLND_BACKOFF_MAX,
+ SOCKLND_PROTOCOL,
+ SOCKLND_ZERO_COPY_RECV,
+ SOCKLND_ZERO_COPY_RECV_MIN_NFRAGS
};
+#else
-int
-ksocknal_lib_tunables_init ()
-{
- int i = 0;
- int j = 1;
-
- if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags < 2)
- *ksocknal_tunables.ksnd_zc_recv_min_nfrags = 2;
-
- if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags > LNET_MAX_IOV)
- *ksocknal_tunables.ksnd_zc_recv_min_nfrags = LNET_MAX_IOV;
+#define SOCKLND_TIMEOUT CTL_UNNUMBERED
+#define SOCKLND_CREDITS CTL_UNNUMBERED
+#define SOCKLND_PEER_CREDITS CTL_UNNUMBERED
+#define SOCKLND_NCONNDS CTL_UNNUMBERED
+#define SOCKLND_RECONNECTS_MIN CTL_UNNUMBERED
+#define SOCKLND_RECONNECTS_MAX CTL_UNNUMBERED
+#define SOCKLND_EAGER_ACK CTL_UNNUMBERED
+#define SOCKLND_ZERO_COPY CTL_UNNUMBERED
+#define SOCKLND_TYPED CTL_UNNUMBERED
+#define SOCKLND_BULK_MIN CTL_UNNUMBERED
+#define SOCKLND_RX_BUFFER_SIZE CTL_UNNUMBERED
+#define SOCKLND_TX_BUFFER_SIZE CTL_UNNUMBERED
+#define SOCKLND_NAGLE CTL_UNNUMBERED
+#define SOCKLND_IRQ_AFFINITY CTL_UNNUMBERED
+#define SOCKLND_ROUND_ROBIN CTL_UNNUMBERED
+#define SOCKLND_KEEPALIVE CTL_UNNUMBERED
+#define SOCKLND_KEEPALIVE_IDLE CTL_UNNUMBERED
+#define SOCKLND_KEEPALIVE_COUNT CTL_UNNUMBERED
+#define SOCKLND_KEEPALIVE_INTVL CTL_UNNUMBERED
+#define SOCKLND_BACKOFF_INIT CTL_UNNUMBERED
+#define SOCKLND_BACKOFF_MAX CTL_UNNUMBERED
+#define SOCKLND_PROTOCOL CTL_UNNUMBERED
+#define SOCKLND_ZERO_COPY_RECV CTL_UNNUMBERED
+#define SOCKLND_ZERO_COPY_RECV_MIN_NFRAGS CTL_UNNUMBERED
+#endif
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+static cfs_sysctl_table_t ksocknal_ctl_table[] = {
+ {
+ .ctl_name = SOCKLND_TIMEOUT,
.procname = "timeout",
- .data = ksocknal_tunables.ksnd_timeout,
+ .data = &ksocknal_tunables.ksnd_timeout,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_CREDITS,
.procname = "credits",
- .data = ksocknal_tunables.ksnd_credits,
+ .data = &ksocknal_tunables.ksnd_credits,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_PEER_CREDITS,
.procname = "peer_credits",
- .data = ksocknal_tunables.ksnd_peercredits,
+ .data = &ksocknal_tunables.ksnd_peercredits,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_NCONNDS,
.procname = "nconnds",
- .data = ksocknal_tunables.ksnd_nconnds,
+ .data = &ksocknal_tunables.ksnd_nconnds,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_RECONNECTS_MIN,
.procname = "min_reconnectms",
- .data = ksocknal_tunables.ksnd_min_reconnectms,
+ .data = &ksocknal_tunables.ksnd_min_reconnectms,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_RECONNECTS_MAX,
.procname = "max_reconnectms",
- .data = ksocknal_tunables.ksnd_max_reconnectms,
+ .data = &ksocknal_tunables.ksnd_max_reconnectms,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_EAGER_ACK,
.procname = "eager_ack",
- .data = ksocknal_tunables.ksnd_eager_ack,
+ .data = &ksocknal_tunables.ksnd_eager_ack,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_ZERO_COPY,
.procname = "zero_copy",
- .data = ksocknal_tunables.ksnd_zc_min_frag,
+ .data = &ksocknal_tunables.ksnd_zc_min_payload,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_ZERO_COPY_RECV,
.procname = "zero_copy_recv",
- .data = ksocknal_tunables.ksnd_zc_recv,
+ .data = &ksocknal_tunables.ksnd_zc_recv,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
- .procname = "zero_copy_recv_min_nfrags",
- .data = ksocknal_tunables.ksnd_zc_recv_min_nfrags,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+
+ {
+ .ctl_name = SOCKLND_ZERO_COPY_RECV_MIN_NFRAGS,
+ .procname = "zero_copy_recv",
+ .data = &ksocknal_tunables.ksnd_zc_recv_min_nfrags,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_TYPED,
.procname = "typed",
- .data = ksocknal_tunables.ksnd_typed_conns,
+ .data = &ksocknal_tunables.ksnd_typed_conns,
.maxlen = sizeof (int),
.mode = 0444,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_BULK_MIN,
.procname = "min_bulk",
- .data = ksocknal_tunables.ksnd_min_bulk,
+ .data = &ksocknal_tunables.ksnd_min_bulk,
.maxlen = sizeof (int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_RX_BUFFER_SIZE,
.procname = "rx_buffer_size",
- .data = ksocknal_tunables.ksnd_rx_buffer_size,
+ .data = &ksocknal_tunables.ksnd_rx_buffer_size,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_TX_BUFFER_SIZE,
.procname = "tx_buffer_size",
- .data = ksocknal_tunables.ksnd_tx_buffer_size,
+ .data = &ksocknal_tunables.ksnd_tx_buffer_size,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_NAGLE,
.procname = "nagle",
- .data = ksocknal_tunables.ksnd_nagle,
+ .data = &ksocknal_tunables.ksnd_nagle,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
#ifdef CPU_AFFINITY
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ {
+ .ctl_name = SOCKLND_IRQ_AFFINITY,
.procname = "irq_affinity",
- .data = ksocknal_tunables.ksnd_irq_affinity,
+ .data = &ksocknal_tunables.ksnd_irq_affinity,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
#endif
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
- .procname = "keepalive_idle",
- .data = ksocknal_tunables.ksnd_keepalive_idle,
+ {
+ .ctl_name = SOCKLND_ROUND_ROBIN,
+ .procname = "round_robin",
+ .data = &ksocknal_tunables.ksnd_round_robin,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
- .procname = "keepalive_count",
- .data = ksocknal_tunables.ksnd_keepalive_count,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_KEEPALIVE,
+ .procname = "keepalive",
+ .data = &ksocknal_tunables.ksnd_keepalive,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_KEEPALIVE_IDLE,
+ .procname = "keepalive_idle",
+ .data = &ksocknal_tunables.ksnd_keepalive_idle,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_KEEPALIVE_COUNT,
+ .procname = "keepalive_count",
+ .data = &ksocknal_tunables.ksnd_keepalive_count,
+ .maxlen = sizeof(int),
+ .mode = 0644,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_KEEPALIVE_INTVL,
.procname = "keepalive_intvl",
- .data = ksocknal_tunables.ksnd_keepalive_intvl,
+ .data = &ksocknal_tunables.ksnd_keepalive_intvl,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
#ifdef SOCKNAL_BACKOFF
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ {
+ .ctl_name = SOCKLND_BACKOFF_INIT,
.procname = "backoff_init",
- .data = ksocknal_tunables.ksnd_backoff_init,
+ .data = &ksocknal_tunables.ksnd_backoff_init,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
+ {
+ .ctl_name = SOCKLND_BACKOFF_MAX,
.procname = "backoff_max",
- .data = ksocknal_tunables.ksnd_backoff_max,
+ .data = &ksocknal_tunables.ksnd_backoff_max,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
#endif
#if SOCKNAL_VERSION_DEBUG
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) {
- .ctl_name = j++,
+ {
+ .ctl_name = SOCKLND_PROTOCOL,
.procname = "protocol",
- .data = ksocknal_tunables.ksnd_protocol,
+ .data = &ksocknal_tunables.ksnd_protocol,
.maxlen = sizeof(int),
.mode = 0644,
- .proc_handler = &proc_dointvec
- };
+ .proc_handler = &proc_dointvec,
+ .strategy = &sysctl_intvec,
+ },
#endif
- ksocknal_ctl_table[i++] = (cfs_sysctl_table_t) { 0 };
+ {0}
+};
- LASSERT (j == i);
- LASSERT (i <= sizeof(ksocknal_ctl_table)/sizeof(ksocknal_ctl_table[0]));
+
+cfs_sysctl_table_t ksocknal_top_ctl_table[] = {
+ {
+ .ctl_name = CTL_SOCKLND,
+ .procname = "socknal",
+ .data = NULL,
+ .maxlen = 0,
+ .mode = 0555,
+ .child = ksocknal_ctl_table
+ },
+ { 0 }
+};
+
+int
+ksocknal_lib_tunables_init ()
+{
+ if (!*ksocknal_tunables.ksnd_typed_conns) {
+ int rc = -EINVAL;
+#if SOCKNAL_VERSION_DEBUG
+ if (*ksocknal_tunables.ksnd_protocol < 3)
+ rc = 0;
+#endif
+ if (rc != 0) {
+ CERROR("Protocol V3.x MUST have typed connections\n");
+ return rc;
+ }
+ }
+
+ if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags < 2)
+ *ksocknal_tunables.ksnd_zc_recv_min_nfrags = 2;
+ if (*ksocknal_tunables.ksnd_zc_recv_min_nfrags > LNET_MAX_IOV)
+ *ksocknal_tunables.ksnd_zc_recv_min_nfrags = LNET_MAX_IOV;
ksocknal_tunables.ksnd_sysctl =
cfs_register_sysctl_table(ksocknal_top_ctl_table, 0);
info = &ksocknal_data.ksnd_irqinfo[irq];
- write_lock_bh (&ksocknal_data.ksnd_global_lock);
+ cfs_write_lock_bh (&ksocknal_data.ksnd_global_lock);
LASSERT (info->ksni_valid);
bind = !info->ksni_bound;
info->ksni_bound = 1;
- write_unlock_bh (&ksocknal_data.ksnd_global_lock);
+ cfs_write_unlock_bh (&ksocknal_data.ksnd_global_lock);
if (!bind) /* bound already */
return;
}
int
-ksocknal_lib_zc_capable(struct socket *sock)
+ksocknal_lib_zc_capable(ksock_conn_t *conn)
{
- int caps = sock->sk->sk_route_caps;
+ int caps = conn->ksnc_sock->sk->sk_route_caps;
+
+ if (conn->ksnc_proto == &ksocknal_protocol_v1x)
+ return 0;
/* ZC if the socket supports scatter/gather and doesn't need software
* checksums */
ksocknal_lib_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
{
struct socket *sock = conn->ksnc_sock;
- lnet_kiov_t *kiov = tx->tx_kiov;
+ lnet_kiov_t *kiov = tx->tx_kiov;
int rc;
int nob;
+ /* Not NOOP message */
+ LASSERT (tx->tx_lnetmsg != NULL);
+
/* NB we can't trust socket ops to either consume our iovs
* or leave them alone. */
-
- if (kiov->kiov_len >= *ksocknal_tunables.ksnd_zc_min_frag &&
- tx->tx_msg.ksm_zc_req_cookie != 0) {
+ if (tx->tx_msg.ksm_zc_cookies[0] != 0) {
/* Zero copy is enabled */
struct sock *sk = sock->sk;
struct page *page = kiov->kiov_page;
msg.msg_iovlen = niov;
}
-
LASSERT (nob <= conn->ksnc_rx_nob_wanted);
set_fs (KERNEL_DS);
/* interleave correctly with closing sockets... */
LASSERT(!in_irq());
- read_lock (&ksocknal_data.ksnd_global_lock);
+ cfs_read_lock (&ksocknal_data.ksnd_global_lock);
conn = sk->sk_user_data;
if (conn == NULL) { /* raced with ksocknal_terminate_conn */
} else
ksocknal_read_callback(conn);
- read_unlock (&ksocknal_data.ksnd_global_lock);
+ cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
EXIT;
}
/* interleave correctly with closing sockets... */
LASSERT(!in_irq());
- read_lock (&ksocknal_data.ksnd_global_lock);
+ cfs_read_lock (&ksocknal_data.ksnd_global_lock);
conn = sk->sk_user_data;
wspace = SOCKNAL_WSPACE(sk);
LASSERT (sk->sk_write_space != &ksocknal_write_space);
sk->sk_write_space (sk);
- read_unlock (&ksocknal_data.ksnd_global_lock);
+ cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
return;
}
clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
}
- read_unlock (&ksocknal_data.ksnd_global_lock);
+ cfs_read_unlock (&ksocknal_data.ksnd_global_lock);
}
void
return ;
}
+
+int
+ksocknal_lib_memory_pressure(ksock_conn_t *conn)
+{
+ int rc = 0;
+ ksock_sched_t *sched;
+
+ sched = conn->ksnc_scheduler;
+ cfs_spin_lock_bh (&sched->kss_lock);
+
+ if (!SOCK_TEST_NOSPACE(conn->ksnc_sock) &&
+ !conn->ksnc_tx_ready) {
+ /* SOCK_NOSPACE is set when the socket fills
+ * and cleared in the write_space callback
+ * (which also sets ksnc_tx_ready). If
+ * SOCK_NOSPACE and ksnc_tx_ready are BOTH
+ * zero, I didn't fill the socket and
+ * write_space won't reschedule me, so I
+ * return -ENOMEM to get my caller to retry
+ * after a timeout */
+ rc = -ENOMEM;
+ }
+
+ cfs_spin_unlock_bh (&sched->kss_lock);
+
+ return rc;
+}
+
+__u64
+ksocknal_lib_new_incarnation(void)
+{
+ struct timeval tv;
+
+ /* The incarnation number is the time this module loaded and it
+ * identifies this particular instance of the socknal. Hopefully
+ * we won't be able to reboot more frequently than 1MHz for the
+ * forseeable future :) */
+
+ do_gettimeofday(&tv);
+
+ return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
+}
+
+int
+ksocknal_lib_bind_thread_to_cpu(int id)
+{
+#if defined(CONFIG_SMP) && defined(CPU_AFFINITY)
+ id = ksocknal_sched2cpu(id);
+ if (cpu_online(id)) {
+ cpumask_t m = CPU_MASK_NONE;
+ cpu_set(id, m);
+ set_cpus_allowed(current, m);
+ return 0;
+ }
+
+ return -1;
+
+#else
+ return 0;
+#endif
+}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff