*/
#include "socknal.h"
-#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
-# include <linux/syscalls.h>
-#endif
/*
* LIB functions follow
PORTAL_FREE(ltx, ltx->ltx_desc_size);
}
-#if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
-struct page *
-ksocknal_kvaddr_to_page (unsigned long vaddr)
-{
- struct page *page;
-
- if (vaddr >= VMALLOC_START &&
- vaddr < VMALLOC_END)
- page = vmalloc_to_page ((void *)vaddr);
-#if CONFIG_HIGHMEM
- else if (vaddr >= PKMAP_BASE &&
- vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE))
- page = vmalloc_to_page ((void *)vaddr);
- /* in 2.4 ^ just walks the page tables */
-#endif
- else
- page = virt_to_page (vaddr);
-
- if (page == NULL ||
- !VALID_PAGE (page))
- return (NULL);
-
- return (page);
-}
-#endif
-
int
ksocknal_send_iov (ksock_conn_t *conn, ksock_tx_t *tx)
-{
- struct socket *sock = conn->ksnc_sock;
+{
struct iovec *iov = tx->tx_iov;
- int fragsize = iov->iov_len;
- unsigned long vaddr = (unsigned long)iov->iov_base;
- int more = (tx->tx_niov > 1) ||
- (tx->tx_nkiov > 0) ||
- (!list_empty (&conn->ksnc_tx_queue));
-#if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
- int offset = vaddr & (PAGE_SIZE - 1);
- int zcsize = MIN (fragsize, PAGE_SIZE - offset);
- struct page *page;
-#endif
- int rc;
+ int nob;
+ int rc;
- /* NB we can't trust socket ops to either consume our iovs
- * or leave them alone, so we only send 1 frag at a time. */
- LASSERT (fragsize <= tx->tx_resid);
LASSERT (tx->tx_niov > 0);
-
-#if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
- if (zcsize >= ksocknal_data.ksnd_zc_min_frag &&
- (sock->sk->route_caps & NETIF_F_SG) &&
- (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM)) &&
- (page = ksocknal_kvaddr_to_page (vaddr)) != NULL) {
-
- CDEBUG(D_NET, "vaddr %p, page %p->%p + offset %x for %d\n",
- (void *)vaddr, page, page_address(page), offset, zcsize);
-
- if (fragsize > zcsize) {
- more = 1;
- fragsize = zcsize;
- }
- rc = tcp_sendpage_zccd(sock, page, offset, zcsize,
- more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
- &tx->tx_zccd);
- } else
-#endif
- {
- /* NB don't pass tx's iov; sendmsg may or may not update it */
- struct iovec fragiov = { .iov_base = (void *)vaddr,
- .iov_len = fragsize};
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
- };
- mm_segment_t oldmm = get_fs();
-
- set_fs (KERNEL_DS);
- rc = sock_sendmsg(sock, &msg, fragsize);
- set_fs (oldmm);
- }
+ /* Never touch tx->tx_iov inside ksocknal_lib_send_iov() */
+ rc = ksocknal_lib_send_iov(conn, tx);
- if (rc > 0) {
- tx->tx_resid -= rc;
+ if (rc <= 0) /* sent nothing? */
+ return (rc);
+
+ nob = rc;
+ LASSERT (nob <= tx->tx_resid);
+ tx->tx_resid -= nob;
+
+ /* "consume" iov */
+ do {
+ LASSERT (tx->tx_niov > 0);
+
+ if (nob < iov->iov_len) {
+ iov->iov_base = (void *)(((unsigned long)(iov->iov_base)) + nob);
+ iov->iov_len -= nob;
+ return (rc);
+ }
+
+ nob -= iov->iov_len;
+ tx->tx_iov = ++iov;
+ tx->tx_niov--;
+ } while (nob != 0);
- if (rc < iov->iov_len) {
- /* didn't send whole iov entry... */
- iov->iov_base = (void *)(vaddr + rc);
- iov->iov_len -= rc;
- } else {
- tx->tx_iov++;
- tx->tx_niov--;
- }
- }
-
return (rc);
}
int
ksocknal_send_kiov (ksock_conn_t *conn, ksock_tx_t *tx)
-{
- struct socket *sock = conn->ksnc_sock;
+{
ptl_kiov_t *kiov = tx->tx_kiov;
- int fragsize = kiov->kiov_len;
- struct page *page = kiov->kiov_page;
- int offset = kiov->kiov_offset;
- int more = (tx->tx_nkiov > 1) ||
- (!list_empty (&conn->ksnc_tx_queue));
- int rc;
+ int nob;
+ int rc;
- /* NB we can't trust socket ops to either consume our iovs
- * or leave them alone, so we only send 1 frag at a time. */
- LASSERT (fragsize <= tx->tx_resid);
- LASSERT (offset + fragsize <= PAGE_SIZE);
- LASSERT (tx->tx_niov == 0);
+ LASSERT (tx->tx_niov == 0);
LASSERT (tx->tx_nkiov > 0);
-#if SOCKNAL_ZC
- if (fragsize >= ksocknal_tunables.ksnd_zc_min_frag &&
- (sock->sk->route_caps & NETIF_F_SG) &&
- (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM))) {
+ /* Never touch tx->tx_kiov inside ksocknal_lib_send_kiov() */
+ rc = ksocknal_lib_send_kiov(conn, tx);
- CDEBUG(D_NET, "page %p + offset %x for %d\n",
- page, offset, fragsize);
+ if (rc <= 0) /* sent nothing? */
+ return (rc);
- rc = tcp_sendpage_zccd(sock, page, offset, fragsize,
- more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
- &tx->tx_zccd);
- } else
-#endif
- {
- char *addr = ((char *)kmap (page)) + offset;
- struct iovec fragiov = {.iov_base = addr,
- .iov_len = fragsize};
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
- };
- mm_segment_t oldmm = get_fs();
-
- set_fs (KERNEL_DS);
- rc = sock_sendmsg(sock, &msg, fragsize);
- set_fs (oldmm);
-
- kunmap (page);
- }
-
- if (rc > 0) {
- tx->tx_resid -= rc;
-
- if (rc < fragsize) {
- kiov->kiov_offset = offset + rc;
- kiov->kiov_len = fragsize - rc;
- } else {
- tx->tx_kiov++;
- tx->tx_nkiov--;
- }
- }
+ nob = rc;
+ LASSERT (nob <= tx->tx_resid);
+ tx->tx_resid -= nob;
+
+ /* "consume" kiov */
+ do {
+ LASSERT(tx->tx_nkiov > 0);
+
+ if (nob < kiov->kiov_len) {
+ kiov->kiov_offset += nob;
+ kiov->kiov_len -= nob;
+ return rc;
+ }
+
+ nob -= kiov->kiov_len;
+ tx->tx_kiov = ++kiov;
+ tx->tx_nkiov--;
+ } while (nob != 0);
return (rc);
}
ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
{
int rc;
-
+ int bufnob;
+
if (ksocknal_data.ksnd_stall_tx != 0) {
set_current_state (TASK_UNINTERRUPTIBLE);
- schedule_timeout (ksocknal_data.ksnd_stall_tx * HZ);
+ schedule_timeout (cfs_time_seconds(ksocknal_data.ksnd_stall_tx));
}
LASSERT (tx->tx_resid != 0);
rc = ksocknal_send_kiov (conn, tx);
}
- if (rc <= 0) {
- /* Didn't write anything.
- *
- * NB: rc == 0 and rc == -EAGAIN both mean try
- * again later (linux stack returns -EAGAIN for
- * this, but Adaptech TOE returns 0).
- *
- * Also, sends never fail with -ENOMEM, just
- * -EAGAIN, but with the added bonus that we can't
- * expect write_space() to call us back to tell us
- * when to try sending again. We use the
- * SOCK_NOSPACE flag to diagnose... */
-
- LASSERT(rc != -ENOMEM);
-
- if (rc == 0 || rc == -EAGAIN) {
- if (test_bit(SOCK_NOSPACE,
- &conn->ksnc_sock->flags)) {
- rc = -EAGAIN;
- } else {
- static int counter;
-
- counter++;
- if ((counter & (-counter)) == counter)
- CWARN("%d ENOMEM tx %p\n",
- counter, conn);
- rc = -ENOMEM;
- }
+ bufnob = SOCK_WMEM_QUEUED(conn->ksnc_sock);
+ if (rc > 0) /* sent something? */
+ conn->ksnc_tx_bufnob += rc; /* account it */
+
+ if (bufnob < conn->ksnc_tx_bufnob) {
+ /* allocated send buffer bytes < computed; infer
+ * something got ACKed */
+ conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
+ conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
+ conn->ksnc_tx_bufnob = bufnob;
+ mb();
+ }
+
+ if (rc <= 0) { /* Didn't write anything? */
+ unsigned long flags;
+ ksock_sched_t *sched;
+
+ if (rc == 0) /* some stacks return 0 instead of -EAGAIN */
+ rc = -EAGAIN;
+
+ if (rc != -EAGAIN)
+ break;
+
+ /* Check if EAGAIN is due to memory pressure */
+
+ sched = conn->ksnc_scheduler;
+ spin_lock_irqsave(&sched->kss_lock, flags);
+
+ 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;
}
+
+ spin_unlock_irqrestore(&sched->kss_lock, flags);
break;
}
+ /* socket's wmem_queued now includes 'rc' bytes */
+ atomic_sub (rc, &conn->ksnc_tx_nob);
rc = 0;
- /* Consider the connection alive since we managed to chuck
- * more data into it. Really, we'd like to consider it
- * alive only when the peer ACKs something, but
- * write_space() only gets called back while SOCK_NOSPACE
- * is set. Instead, we presume peer death has occurred if
- * the socket doesn't drain within a timout */
- conn->ksnc_tx_deadline = jiffies +
- ksocknal_tunables.ksnd_io_timeout * HZ;
- conn->ksnc_peer->ksnp_last_alive = jiffies;
-
} while (tx->tx_resid != 0);
ksocknal_putconnsock (conn);
return (rc);
}
-void
-ksocknal_eager_ack (ksock_conn_t *conn)
-{
- int opt = 1;
- mm_segment_t oldmm = get_fs();
- struct socket *sock = conn->ksnc_sock;
-
- /* Remind the socket to ACK eagerly. If I don't, the socket might
- * think I'm about to send something it could piggy-back the ACK
- * on, introducing delay in completing zero-copy sends in my
- * peer. */
-
- set_fs(KERNEL_DS);
- sock->ops->setsockopt (sock, SOL_TCP, TCP_QUICKACK,
- (char *)&opt, sizeof (opt));
- set_fs(oldmm);
-}
-
int
ksocknal_recv_iov (ksock_conn_t *conn)
-{
+{
struct iovec *iov = conn->ksnc_rx_iov;
- int fragsize = iov->iov_len;
- unsigned long vaddr = (unsigned long)iov->iov_base;
- struct iovec fragiov = { .iov_base = (void *)vaddr,
- .iov_len = fragsize};
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = 0
- };
- mm_segment_t oldmm = get_fs();
- int rc;
-
- /* NB we can't trust socket ops to either consume our iovs
- * or leave them alone, so we only receive 1 frag at a time. */
+ int nob;
+ int rc;
+
LASSERT (conn->ksnc_rx_niov > 0);
- LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
- set_fs (KERNEL_DS);
- rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
- /* NB this is just a boolean............................^ */
- set_fs (oldmm);
+ /* Never touch conn->ksnc_rx_iov or change connection
+ * status inside ksocknal_lib_recv_iov */
+ rc = ksocknal_lib_recv_iov(conn);
- if (rc <= 0)
- return (rc);
+ if (rc <= 0)
+ return (rc);
- /* received something... */
- conn->ksnc_peer->ksnp_last_alive = jiffies;
- conn->ksnc_rx_deadline = jiffies +
- ksocknal_tunables.ksnd_io_timeout * HZ;
- mb(); /* order with setting rx_started */
- conn->ksnc_rx_started = 1;
+ /* received something... */
+ nob = rc;
- conn->ksnc_rx_nob_wanted -= rc;
- conn->ksnc_rx_nob_left -= rc;
-
- if (rc < fragsize) {
- iov->iov_base = (void *)(vaddr + rc);
- iov->iov_len = fragsize - rc;
- return (-EAGAIN);
- }
+ conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
+ conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout);
+ mb(); /* order with setting rx_started */
+ conn->ksnc_rx_started = 1;
- conn->ksnc_rx_iov++;
- conn->ksnc_rx_niov--;
- return (1);
+ conn->ksnc_rx_nob_wanted -= nob;
+ conn->ksnc_rx_nob_left -= nob;
+
+ do {
+ LASSERT (conn->ksnc_rx_niov > 0);
+
+ if (nob < iov->iov_len) {
+ iov->iov_len -= nob;
+ iov->iov_base = (void *)(((unsigned long)iov->iov_base) + nob);
+ return (-EAGAIN);
+ }
+
+ nob -= iov->iov_len;
+ conn->ksnc_rx_iov = ++iov;
+ conn->ksnc_rx_niov--;
+ } while (nob != 0);
+
+ return (rc);
}
int
ksocknal_recv_kiov (ksock_conn_t *conn)
{
ptl_kiov_t *kiov = conn->ksnc_rx_kiov;
- struct page *page = kiov->kiov_page;
- int offset = kiov->kiov_offset;
- int fragsize = kiov->kiov_len;
- unsigned long vaddr = ((unsigned long)kmap (page)) + offset;
- struct iovec fragiov = { .iov_base = (void *)vaddr,
- .iov_len = fragsize};
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = 0
- };
- mm_segment_t oldmm = get_fs();
- int rc;
-
- /* NB we can't trust socket ops to either consume our iovs
- * or leave them alone, so we only receive 1 frag at a time. */
- LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
+ int nob;
+ int rc;
LASSERT (conn->ksnc_rx_nkiov > 0);
- LASSERT (offset + fragsize <= PAGE_SIZE);
- set_fs (KERNEL_DS);
- rc = sock_recvmsg (conn->ksnc_sock, &msg, fragsize, MSG_DONTWAIT);
- /* NB this is just a boolean............................^ */
- set_fs (oldmm);
+ /* Never touch conn->ksnc_rx_kiov or change connection
+ * status inside ksocknal_lib_recv_iov */
+ rc = ksocknal_lib_recv_kiov(conn);
- kunmap (page);
-
- if (rc <= 0)
- return (rc);
-
- /* received something... */
- conn->ksnc_peer->ksnp_last_alive = jiffies;
- conn->ksnc_rx_deadline = jiffies +
- ksocknal_tunables.ksnd_io_timeout * HZ;
- mb(); /* order with setting rx_started */
+ if (rc <= 0)
+ return (rc);
+
+ /* received something... */
+ nob = rc;
+
+ conn->ksnc_peer->ksnp_last_alive = cfs_time_current();
+ conn->ksnc_rx_deadline = cfs_time_shift (ksocknal_tunables.ksnd_io_timeout);
+ mb(); /* order with setting rx_started */
conn->ksnc_rx_started = 1;
- conn->ksnc_rx_nob_wanted -= rc;
- conn->ksnc_rx_nob_left -= rc;
-
- if (rc < fragsize) {
- kiov->kiov_offset = offset + rc;
- kiov->kiov_len = fragsize - rc;
- return (-EAGAIN);
- }
+ conn->ksnc_rx_nob_wanted -= nob;
+ conn->ksnc_rx_nob_left -= nob;
- conn->ksnc_rx_kiov++;
- conn->ksnc_rx_nkiov--;
- return (1);
+ do {
+ LASSERT (conn->ksnc_rx_nkiov > 0);
+
+ if (nob < kiov->kiov_len) {
+ kiov->kiov_offset += nob;
+ kiov->kiov_len -= nob;
+ return -EAGAIN;
+ }
+
+ nob -= kiov->kiov_len;
+ conn->ksnc_rx_kiov = ++kiov;
+ conn->ksnc_rx_nkiov--;
+ } while (nob != 0);
+
+ return 1;
}
int
* progress/completion. */
int rc;
ENTRY;
-
+
if (ksocknal_data.ksnd_stall_rx != 0) {
set_current_state (TASK_UNINTERRUPTIBLE);
- schedule_timeout (ksocknal_data.ksnd_stall_rx * HZ);
+ schedule_timeout(cfs_time_seconds (ksocknal_data.ksnd_stall_rx));
}
rc = ksocknal_getconnsock (conn);
(conn->ksnc_rx_state == SOCKNAL_RX_BODY ||
conn->ksnc_rx_state == SOCKNAL_RX_BODY_FWD)) {
/* Remind the socket to ack eagerly... */
- ksocknal_eager_ack(conn);
+ ksocknal_lib_eager_ack(conn);
}
rc = 1;
break;
spin_lock_irqsave (&sched->kss_lock, flags);
list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list);
- wake_up (&sched->kss_waitq);
+ cfs_waitq_signal (&sched->kss_waitq);
spin_unlock_irqrestore (&sched->kss_lock, flags);
EXIT;
ENTRY;
if (tx->tx_conn != NULL) {
- /* This tx got queued on a conn; do the accounting... */
- atomic_sub (tx->tx_nob, &tx->tx_conn->ksnc_tx_nob);
#if SOCKNAL_ZC
/* zero copy completion isn't always from
* process_transmit() so it needs to keep a ref on
#if SOCKNAL_ZC
if (atomic_read (&tx->tx_zccd.zccd_count) != 1) {
ksock_conn_t *conn = tx->tx_conn;
-
+
/* zccd skbufs are still in-flight. First take a ref on
* conn, so it hangs about for ksocknal_tx_done... */
atomic_inc (&conn->ksnc_refcount);
{
unsigned long flags;
int rc;
-
+
rc = ksocknal_transmit (conn, tx);
CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
return (rc);
if (rc == -ENOMEM) {
+ static int counter;
+
+ counter++; /* exponential backoff warnings */
+ if ((counter & (-counter)) == counter)
+ CWARN("%d ENOMEM tx %p (%u allocated)\n",
+ counter, conn, atomic_read(&portal_kmemory));
+
/* Queue on ksnd_enomem_conns for retry after a timeout */
spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags);
LASSERT (conn->ksnc_tx_scheduled);
list_add_tail(&conn->ksnc_tx_list,
&ksocknal_data.ksnd_enomem_conns);
- if (!time_after_eq(jiffies + SOCKNAL_ENOMEM_RETRY,
+ if (!cfs_time_aftereq(cfs_time_add(cfs_time_current(),
+ SOCKNAL_ENOMEM_RETRY),
ksocknal_data.ksnd_reaper_waketime))
- wake_up (&ksocknal_data.ksnd_reaper_waitq);
-
+ cfs_waitq_signal (&ksocknal_data.ksnd_reaper_waitq);
+
spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags);
return (rc);
}
/* Actual error */
LASSERT (rc < 0);
- if (!conn->ksnc_closing)
- CERROR("[%p] Error %d on write to "LPX64
+ if (!conn->ksnc_closing) {
+ switch (rc) {
+ case -ECONNRESET:
+ LCONSOLE_WARN("Host %u.%u.%u.%u reset our connection "
+ "while we were sending data; it may have "
+ "rebooted.\n",
+ HIPQUAD(conn->ksnc_ipaddr));
+ break;
+ default:
+ LCONSOLE_WARN("There was an unexpected network error "
+ "while writing to %u.%u.%u.%u: %d.\n",
+ HIPQUAD(conn->ksnc_ipaddr), rc);
+ break;
+ }
+ CDEBUG(D_HA, "[%p] Error %d on write to "LPX64
" ip %d.%d.%d.%d:%d\n", conn, rc,
conn->ksnc_peer->ksnp_nid,
HIPQUAD(conn->ksnc_ipaddr),
conn->ksnc_port);
+ }
ksocknal_close_conn_and_siblings (conn, rc);
ksocknal_tx_launched (tx);
unsigned long flags;
/* called holding write lock on ksnd_global_lock */
-
- LASSERT (!route->ksnr_deleted);
- LASSERT ((route->ksnr_connected & (1 << SOCKNAL_CONN_ANY)) == 0);
- LASSERT ((route->ksnr_connected & KSNR_TYPED_ROUTES) != KSNR_TYPED_ROUTES);
LASSERT (!route->ksnr_connecting);
-
- if (ksocknal_tunables.ksnd_typed_conns)
- route->ksnr_connecting =
- KSNR_TYPED_ROUTES & ~route->ksnr_connected;
- else
- route->ksnr_connecting = (1 << SOCKNAL_CONN_ANY);
- atomic_inc (&route->ksnr_refcount); /* extra ref for asynchd */
-
+ route->ksnr_connecting = 1; /* scheduling conn for autoconnectd */
+ atomic_inc (&route->ksnr_refcount); /* extra ref for autoconnectd */
+
spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
-
+
list_add_tail (&route->ksnr_connect_list,
&ksocknal_data.ksnd_autoconnectd_routes);
- wake_up (&ksocknal_data.ksnd_autoconnectd_waitq);
-
+ cfs_waitq_signal (&ksocknal_data.ksnd_autoconnectd_waitq);
+
spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
}
int tnob = 0;
ksock_conn_t *fallback = NULL;
int fnob = 0;
+ ksock_conn_t *conn;
- /* Find the conn with the shortest tx queue */
list_for_each (tmp, &peer->ksnp_conns) {
ksock_conn_t *c = list_entry(tmp, ksock_conn_t, ksnc_list);
+#if SOCKNAL_ROUND_ROBIN
+ const int nob = 0;
+#else
int nob = atomic_read(&c->ksnc_tx_nob) +
- c->ksnc_sock->sk->sk_wmem_queued;
-
+ SOCK_WMEM_QUEUED(c->ksnc_sock);
+#endif
LASSERT (!c->ksnc_closing);
if (fallback == NULL || nob < fnob) {
switch (c->ksnc_type) {
default:
+ CERROR("ksnc_type bad: %u\n", c->ksnc_type);
LBUG();
case SOCKNAL_CONN_ANY:
break;
}
/* prefer the typed selection */
- return ((typed != NULL) ? typed : fallback);
+ conn = (typed != NULL) ? typed : fallback;
+
+#if SOCKNAL_ROUND_ROBIN
+ if (conn != NULL) {
+ /* round-robin all else being equal */
+ list_del (&conn->ksnc_list);
+ list_add_tail (&conn->ksnc_list, &peer->ksnp_conns);
+ }
+#endif
+ return conn;
}
void
* ksnc_sock... */
LASSERT(!conn->ksnc_closing);
LASSERT(tx->tx_resid == tx->tx_nob);
-
+
CDEBUG (D_NET, "Sending to "LPX64" ip %d.%d.%d.%d:%d\n",
conn->ksnc_peer->ksnp_nid,
HIPQUAD(conn->ksnc_ipaddr),
#endif
spin_lock_irqsave (&sched->kss_lock, flags);
- conn->ksnc_tx_deadline = jiffies +
- ksocknal_tunables.ksnd_io_timeout * HZ;
- mb(); /* order with list_add_tail */
+ if (list_empty(&conn->ksnc_tx_queue) &&
+ SOCK_WMEM_QUEUED(conn->ksnc_sock) == 0) {
+ /* First packet starts the timeout */
+ conn->ksnc_tx_deadline = cfs_time_shift(ksocknal_tunables.ksnd_io_timeout);
+ conn->ksnc_tx_bufnob = 0;
+ mb(); /* order with adding to tx_queue */
+ }
list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
-
+
if (conn->ksnc_tx_ready && /* able to send */
!conn->ksnc_tx_scheduled) { /* not scheduled to send */
/* +1 ref for scheduler */
list_add_tail (&conn->ksnc_tx_list,
&sched->kss_tx_conns);
conn->ksnc_tx_scheduled = 1;
- wake_up (&sched->kss_waitq);
+ cfs_waitq_signal (&sched->kss_waitq);
}
spin_unlock_irqrestore (&sched->kss_lock, flags);
{
struct list_head *tmp;
ksock_route_t *route;
- ksock_route_t *first_lazy = NULL;
- int found_connecting_or_connected = 0;
int bits;
-
+
list_for_each (tmp, &peer->ksnp_routes) {
route = list_entry (tmp, ksock_route_t, ksnr_list);
bits = route->ksnr_connected;
-
- if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES ||
- (bits & (1 << SOCKNAL_CONN_ANY)) != 0 ||
- route->ksnr_connecting != 0) {
- /* All typed connections have been established, or
- * an untyped connection has been established, or
- * connections are currently being established */
- found_connecting_or_connected = 1;
- continue;
+
+ if (ksocknal_tunables.ksnd_typed_conns) {
+ /* All typed connections established? */
+ if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES)
+ continue;
+ } else {
+ /* Untyped connection established? */
+ if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0)
+ continue;
}
+ /* connection being established? */
+ if (route->ksnr_connecting)
+ continue;
+
/* too soon to retry this guy? */
- if (!time_after_eq (jiffies, route->ksnr_timeout))
+ if (!cfs_time_aftereq (cfs_time_current(), route->ksnr_timeout))
continue;
-
- /* eager routes always want to be connected */
- if (route->ksnr_eager)
- return (route);
- if (first_lazy == NULL)
- first_lazy = route;
+ return (route);
}
-
- /* No eager routes need to be connected. If some connection has
- * already been established, or is being established there's nothing to
- * do. Otherwise we return the first lazy route we found. If it fails
- * to connect, it will go to the end of the list. */
- if (!list_empty (&peer->ksnp_conns) ||
- found_connecting_or_connected)
- return (NULL);
-
- return (first_lazy);
+ return (NULL);
}
ksock_route_t *
list_for_each (tmp, &peer->ksnp_routes) {
route = list_entry (tmp, ksock_route_t, ksnr_list);
-
- if (route->ksnr_connecting != 0)
+
+ if (route->ksnr_connecting)
return (route);
}
-
+
return (NULL);
}
ksock_conn_t *conn;
ksock_route_t *route;
rwlock_t *g_lock;
-
+
/* Ensure the frags we've been given EXACTLY match the number of
* bytes we want to send. Many TCP/IP stacks disregard any total
* size parameters passed to them and just look at the frags.
tx->tx_hdr = (ptl_hdr_t *)tx->tx_iov[0].iov_base;
g_lock = &ksocknal_data.ksnd_global_lock;
+#if !SOCKNAL_ROUND_ROBIN
read_lock (g_lock);
-
+
peer = ksocknal_find_target_peer_locked (tx, nid);
if (peer == NULL) {
read_unlock (g_lock);
return (0);
}
}
-
- /* Making one or more connections; I'll need a write lock... */
- atomic_inc (&peer->ksnp_refcount); /* +1 ref for me while I unlock */
+ /* I'll need a write lock... */
read_unlock (g_lock);
- write_lock_irqsave (g_lock, flags);
-
- if (peer->ksnp_closing) { /* peer deleted as I blocked! */
- write_unlock_irqrestore (g_lock, flags);
- ksocknal_put_peer (peer);
+#endif
+ write_lock_irqsave(g_lock, flags);
+
+ peer = ksocknal_find_target_peer_locked (tx, nid);
+ if (peer == NULL) {
+ write_unlock_irqrestore(g_lock, flags);
return (-EHOSTUNREACH);
}
- ksocknal_put_peer (peer); /* drop ref I got above */
for (;;) {
/* launch any/all autoconnections that need it */
write_unlock_irqrestore (g_lock, flags);
return (0);
}
-
+
write_unlock_irqrestore (g_lock, flags);
return (-EHOSTUNREACH);
}
LASSERT (payload_kiov == NULL || !in_interrupt ());
/* payload is either all vaddrs or all pages */
LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
-
+
if (payload_iov != NULL)
desc_size = offsetof(ksock_ltx_t, ltx_iov[1 + payload_niov]);
else
desc_size = offsetof(ksock_ltx_t, ltx_kiov[payload_niov]);
-
+
if (in_interrupt() ||
type == PTL_MSG_ACK ||
type == PTL_MSG_REPLY) {
} else {
PORTAL_ALLOC(ltx, desc_size);
}
-
+
if (ltx == NULL) {
CERROR("Can't allocate tx desc type %d size %d %s\n",
type, desc_size, in_interrupt() ? "(intr)" : "");
atomic_inc(&ksocknal_data.ksnd_nactive_ltxs);
ltx->ltx_desc_size = desc_size;
-
+
/* We always have 1 mapped frag for the header */
ltx->ltx_tx.tx_iov = ltx->ltx_iov;
ltx->ltx_iov[0].iov_base = <x->ltx_hdr;
ltx->ltx_iov[0].iov_len = sizeof(*hdr);
ltx->ltx_hdr = *hdr;
-
+
ltx->ltx_private = private;
ltx->ltx_cookie = cookie;
-
+
ltx->ltx_tx.tx_isfwd = 0;
ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
rc = ksocknal_launch_packet(<x->ltx_tx, nid);
if (rc == 0)
return (PTL_OK);
-
+
ksocknal_free_ltx(ltx);
return (PTL_FAIL);
}
ptl_nid_t nid = fwd->kprfd_gateway_nid;
ksock_ftx_t *ftx = (ksock_ftx_t *)&fwd->kprfd_scratch;
int rc;
-
+
CDEBUG (D_NET, "Forwarding [%p] -> "LPX64" ("LPX64"))\n", fwd,
fwd->kprfd_gateway_nid, fwd->kprfd_target_nid);
int
ksocknal_thread_start (int (*fn)(void *arg), void *arg)
{
- long pid = kernel_thread (fn, arg, 0);
+ long pid = cfs_kernel_thread (fn, arg, 0);
+ unsigned long flags;
if (pid < 0)
return ((int)pid);
- atomic_inc (&ksocknal_data.ksnd_nthreads);
+ write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
+ ksocknal_data.ksnd_nthreads++;
+ write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
return (0);
}
void
ksocknal_thread_fini (void)
{
- atomic_dec (&ksocknal_data.ksnd_nthreads);
+ unsigned long flags;
+
+ write_lock_irqsave(&ksocknal_data.ksnd_global_lock, flags);
+ ksocknal_data.ksnd_nthreads--;
+ write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
}
void
{
ksock_fmb_t *fmb = (ksock_fmb_t *)arg;
ksock_fmb_pool_t *fmp = fmb->fmb_pool;
- ptl_hdr_t *hdr = (ptl_hdr_t *)page_address(fmb->fmb_kiov[0].kiov_page);
+ ptl_hdr_t *hdr = &fmb->fmb_hdr;
ksock_conn_t *conn = NULL;
ksock_sched_t *sched;
unsigned long flags;
if (error != 0)
CERROR("Failed to route packet from "
LPX64" %s to "LPX64" %s: %d\n",
- NTOH__u64(hdr->src_nid),
- portals_nid2str(SOCKNAL, NTOH__u64(hdr->src_nid), ipbuf),
- NTOH__u64(hdr->dest_nid),
- portals_nid2str(SOCKNAL, NTOH__u64(hdr->dest_nid), ipbuf2),
+ le64_to_cpu(hdr->src_nid),
+ portals_nid2str(SOCKNAL, le64_to_cpu(hdr->src_nid), ipbuf),
+ le64_to_cpu(hdr->dest_nid),
+ portals_nid2str(SOCKNAL, le64_to_cpu(hdr->dest_nid), ipbuf2),
error);
else
CDEBUG (D_NET, "routed packet from "LPX64" to "LPX64": OK\n",
- NTOH__u64 (hdr->src_nid), NTOH__u64 (hdr->dest_nid));
+ le64_to_cpu(hdr->src_nid), le64_to_cpu(hdr->dest_nid));
/* drop peer ref taken on init */
ksocknal_put_peer (fmb->fmb_peer);
spin_lock_irqsave (&sched->kss_lock, flags);
list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
- wake_up (&sched->kss_waitq);
+ cfs_waitq_signal (&sched->kss_waitq);
spin_unlock_irqrestore (&sched->kss_lock, flags);
}
LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
- if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * PAGE_SIZE)
+ if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * CFS_PAGE_SIZE)
pool = &ksocknal_data.ksnd_small_fmp;
else
pool = &ksocknal_data.ksnd_large_fmp;
ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
{
int payload_nob = conn->ksnc_rx_nob_left;
- ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
+ ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
int niov = 0;
int nob = payload_nob;
LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
LASSERT (conn->ksnc_rx_nob_wanted == conn->ksnc_rx_nob_left);
LASSERT (payload_nob >= 0);
- LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * PAGE_SIZE);
- LASSERT (sizeof (ptl_hdr_t) < PAGE_SIZE);
+ LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * CFS_PAGE_SIZE);
+ LASSERT (sizeof (ptl_hdr_t) < CFS_PAGE_SIZE);
LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
/* Take a ref on the conn's peer to prevent module unload before
while (nob > 0) {
LASSERT (niov < fmb->fmb_pool->fmp_buff_pages);
LASSERT (fmb->fmb_kiov[niov].kiov_offset == 0);
- fmb->fmb_kiov[niov].kiov_len = MIN (PAGE_SIZE, nob);
- nob -= PAGE_SIZE;
+ fmb->fmb_kiov[niov].kiov_len = MIN (CFS_PAGE_SIZE, nob);
+ nob -= CFS_PAGE_SIZE;
niov++;
}
if (payload_nob == 0) { /* got complete packet already */
CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
- conn, NTOH__u64 (conn->ksnc_hdr.src_nid), dest_nid);
+ conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid);
kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
conn->ksnc_cookie = fmb; /* stash fmb for later */
conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
-
+
/* Set up conn->ksnc_rx_kiov to read the payload into fmb's kiov-ed
* buffer */
LASSERT (niov <= sizeof(conn->ksnc_rx_iov_space)/sizeof(ptl_kiov_t));
conn->ksnc_rx_nkiov = niov;
conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
memcpy(conn->ksnc_rx_kiov, fmb->fmb_kiov, niov * sizeof(ptl_kiov_t));
-
+
CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d reading body\n", conn,
- NTOH__u64 (conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
+ le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid, payload_nob);
return (0);
}
ksocknal_fwd_parse (ksock_conn_t *conn)
{
ksock_peer_t *peer;
- ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
- ptl_nid_t src_nid = NTOH__u64 (conn->ksnc_hdr.src_nid);
- int body_len = NTOH__u32 (conn->ksnc_hdr.payload_length);
+ ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
+ ptl_nid_t src_nid = le64_to_cpu(conn->ksnc_hdr.src_nid);
+ int body_len = le32_to_cpu(conn->ksnc_hdr.payload_length);
char str[PTL_NALFMT_SIZE];
char str2[PTL_NALFMT_SIZE];
if (nob_to_skip == 0) { /* right at next packet boundary now */
conn->ksnc_rx_started = 0;
mb (); /* racing with timeout thread */
-
+
conn->ksnc_rx_state = SOCKNAL_RX_HEADER;
conn->ksnc_rx_nob_wanted = sizeof (ptl_hdr_t);
conn->ksnc_rx_nob_left = sizeof (ptl_hdr_t);
{
ksock_fmb_t *fmb;
int rc;
-
+
LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
/* doesn't need a forwarding buffer */
/* short read */
return (-EAGAIN);
}
-
+
switch (conn->ksnc_rx_state) {
case SOCKNAL_RX_HEADER:
- if (conn->ksnc_hdr.type != HTON__u32(PTL_MSG_HELLO) &&
- NTOH__u64(conn->ksnc_hdr.dest_nid) !=
+ if (conn->ksnc_hdr.type != cpu_to_le32(PTL_MSG_HELLO) &&
+ le64_to_cpu(conn->ksnc_hdr.dest_nid) !=
ksocknal_lib.libnal_ni.ni_pid.nid) {
/* This packet isn't for me */
ksocknal_fwd_parse (conn);
case SOCKNAL_RX_BODY_FWD:
/* payload all received */
CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (got body)\n",
- conn, NTOH__u64 (conn->ksnc_hdr.src_nid),
- NTOH__u64 (conn->ksnc_hdr.dest_nid),
+ conn, le64_to_cpu(conn->ksnc_hdr.src_nid),
+ le64_to_cpu(conn->ksnc_hdr.dest_nid),
conn->ksnc_rx_nob_left);
/* forward the packet. NB ksocknal_init_fmb() put fmb into
LASSERT (mlen <= rlen);
LASSERT (niov <= PTL_MD_MAX_IOV);
-
+
conn->ksnc_cookie = msg;
conn->ksnc_rx_nob_wanted = mlen;
conn->ksnc_rx_nob_left = rlen;
LASSERT (mlen <= rlen);
LASSERT (niov <= PTL_MD_MAX_IOV);
-
+
conn->ksnc_cookie = msg;
conn->ksnc_rx_nob_wanted = mlen;
conn->ksnc_rx_nob_left = rlen;
return (PTL_OK);
}
+static inline int
+ksocknal_sched_cansleep(ksock_sched_t *sched)
+{
+ unsigned long flags;
+ int rc;
+
+ spin_lock_irqsave(&sched->kss_lock, flags);
+
+ rc = (!ksocknal_data.ksnd_shuttingdown &&
+#if SOCKNAL_ZC
+ list_empty(&sched->kss_zctxdone_list) &&
+#endif
+ list_empty(&sched->kss_rx_conns) &&
+ list_empty(&sched->kss_tx_conns));
+
+ spin_unlock_irqrestore(&sched->kss_lock, flags);
+ return (rc);
+}
+
int ksocknal_scheduler (void *arg)
{
ksock_sched_t *sched = (ksock_sched_t *)arg;
kportal_blockallsigs ();
#if (CONFIG_SMP && CPU_AFFINITY)
- if ((cpu_online_map & (1 << id)) != 0) {
-#if 1
- current->cpus_allowed = (1 << id);
-#else
- set_cpus_allowed (current, 1<<id);
-#endif
+ id = ksocknal_sched2cpu(id);
+ if (cpu_online(id)) {
+ cpumask_t m;
+ cpu_set(id, m);
+ set_cpus_allowed(current, m);
} else {
- CERROR ("Can't set CPU affinity for %s\n", name);
+ CERROR ("Can't set CPU affinity for %s to %d\n", name, id);
}
#endif /* CONFIG_SMP && CPU_AFFINITY */
-
+
spin_lock_irqsave (&sched->kss_lock, flags);
while (!ksocknal_data.ksnd_shuttingdown) {
* kss_lock. */
conn->ksnc_rx_ready = 0;
spin_unlock_irqrestore(&sched->kss_lock, flags);
-
+
rc = ksocknal_process_receive(conn);
-
+
spin_lock_irqsave(&sched->kss_lock, flags);
/* I'm the only one that can clear this flag */
/* Did process_receive get everything it wanted? */
if (rc == 0)
conn->ksnc_rx_ready = 1;
-
+
if (conn->ksnc_rx_state == SOCKNAL_RX_FMB_SLEEP ||
conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB) {
/* Conn blocked for a forwarding buffer.
conn = list_entry(sched->kss_tx_conns.next,
ksock_conn_t, ksnc_tx_list);
list_del (&conn->ksnc_tx_list);
-
+
LASSERT(conn->ksnc_tx_scheduled);
LASSERT(conn->ksnc_tx_ready);
LASSERT(!list_empty(&conn->ksnc_tx_queue));
-
+
tx = list_entry(conn->ksnc_tx_queue.next,
ksock_tx_t, tx_list);
/* dequeue now so empty list => more to send */
list_del(&tx->tx_list);
-
+
/* Clear tx_ready in case send isn't complete. Do
* it BEFORE we call process_transmit, since
* write_space can set it any time after we release
/* drop my ref */
ksocknal_put_conn (conn);
}
-
+
did_something = 1;
}
#if SOCKNAL_ZC
nloops = 0;
if (!did_something) { /* wait for something to do */
-#if SOCKNAL_ZC
- rc = wait_event_interruptible (sched->kss_waitq,
- ksocknal_data.ksnd_shuttingdown ||
- !list_empty(&sched->kss_rx_conns) ||
- !list_empty(&sched->kss_tx_conns) ||
- !list_empty(&sched->kss_zctxdone_list));
-#else
rc = wait_event_interruptible (sched->kss_waitq,
- ksocknal_data.ksnd_shuttingdown ||
- !list_empty(&sched->kss_rx_conns) ||
- !list_empty(&sched->kss_tx_conns));
-#endif
+ !ksocknal_sched_cansleep(sched));
LASSERT (rc == 0);
} else
our_cond_resched();
return (0);
}
-void
-ksocknal_data_ready (struct sock *sk, int n)
+/*
+ * Add connection to kss_rx_conns of scheduler
+ * and wakeup the scheduler.
+ */
+void ksocknal_read_callback (ksock_conn_t *conn)
{
+ ksock_sched_t *sched;
unsigned long flags;
- ksock_conn_t *conn;
- ksock_sched_t *sched;
ENTRY;
- /* interleave correctly with closing sockets... */
- read_lock (&ksocknal_data.ksnd_global_lock);
-
- conn = sk->sk_user_data;
- if (conn == NULL) { /* raced with ksocknal_terminate_conn */
- LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
- sk->sk_data_ready (sk, n);
- } else {
- sched = conn->ksnc_scheduler;
-
- spin_lock_irqsave (&sched->kss_lock, flags);
+ sched = conn->ksnc_scheduler;
- conn->ksnc_rx_ready = 1;
+ spin_lock_irqsave (&sched->kss_lock, flags);
- if (!conn->ksnc_rx_scheduled) { /* not being progressed */
- list_add_tail(&conn->ksnc_rx_list,
- &sched->kss_rx_conns);
- conn->ksnc_rx_scheduled = 1;
- /* extra ref for scheduler */
- atomic_inc (&conn->ksnc_refcount);
+ conn->ksnc_rx_ready = 1;
- wake_up (&sched->kss_waitq);
- }
+ if (!conn->ksnc_rx_scheduled) { /* not being progressed */
+ list_add_tail(&conn->ksnc_rx_list,
+ &sched->kss_rx_conns);
+ conn->ksnc_rx_scheduled = 1;
+ /* extra ref for scheduler */
+ atomic_inc (&conn->ksnc_refcount);
- spin_unlock_irqrestore (&sched->kss_lock, flags);
- }
-
- read_unlock (&ksocknal_data.ksnd_global_lock);
+ cfs_waitq_signal (&sched->kss_waitq);
+ }
+ spin_unlock_irqrestore (&sched->kss_lock, flags);
EXIT;
-}
+}
-void
-ksocknal_write_space (struct sock *sk)
-{
+/*
+ * Add connection to kss_tx_conns of scheduler
+ * and wakeup the scheduler.
+ */
+void ksocknal_write_callback (ksock_conn_t *conn)
+{
+ ksock_sched_t *sched;
unsigned long flags;
- ksock_conn_t *conn;
- ksock_sched_t *sched;
-
- /* interleave correctly with closing sockets... */
- read_lock (&ksocknal_data.ksnd_global_lock);
-
- conn = sk->sk_user_data;
-
- CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
- sk, tcp_wspace(sk), SOCKNAL_TX_LOW_WATER(sk), conn,
- (conn == NULL) ? "" : (conn->ksnc_tx_ready ?
- " ready" : " blocked"),
- (conn == NULL) ? "" : (conn->ksnc_tx_scheduled ?
- " scheduled" : " idle"),
- (conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
- " empty" : " queued"));
-
- if (conn == NULL) { /* raced with ksocknal_terminate_conn */
- LASSERT (sk->sk_write_space != &ksocknal_write_space);
- sk->sk_write_space (sk);
-
- read_unlock (&ksocknal_data.ksnd_global_lock);
- return;
- }
-
- if (tcp_wspace(sk) >= SOCKNAL_TX_LOW_WATER(sk)) { /* got enough space */
- clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
+ ENTRY;
- sched = conn->ksnc_scheduler;
+ sched = conn->ksnc_scheduler;
- spin_lock_irqsave (&sched->kss_lock, flags);
+ spin_lock_irqsave (&sched->kss_lock, flags);
- conn->ksnc_tx_ready = 1;
+ conn->ksnc_tx_ready = 1;
- if (!conn->ksnc_tx_scheduled && // not being progressed
- !list_empty(&conn->ksnc_tx_queue)){//packets to send
- list_add_tail (&conn->ksnc_tx_list,
- &sched->kss_tx_conns);
- conn->ksnc_tx_scheduled = 1;
- /* extra ref for scheduler */
- atomic_inc (&conn->ksnc_refcount);
+ if (!conn->ksnc_tx_scheduled && // not being progressed
+ !list_empty(&conn->ksnc_tx_queue)){//packets to send
+ list_add_tail (&conn->ksnc_tx_list,
+ &sched->kss_tx_conns);
+ conn->ksnc_tx_scheduled = 1;
+ /* extra ref for scheduler */
+ atomic_inc (&conn->ksnc_refcount);
- wake_up (&sched->kss_waitq);
- }
+ cfs_waitq_signal (&sched->kss_waitq);
+ }
- spin_unlock_irqrestore (&sched->kss_lock, flags);
- }
+ spin_unlock_irqrestore (&sched->kss_lock, flags);
- read_unlock (&ksocknal_data.ksnd_global_lock);
+ EXIT;
}
int
ksocknal_sock_write (struct socket *sock, void *buffer, int nob)
{
- int rc;
- mm_segment_t oldmm = get_fs();
-
- while (nob > 0) {
- struct iovec iov = {
- .iov_base = buffer,
- .iov_len = nob
- };
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &iov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = 0
- };
-
- set_fs (KERNEL_DS);
- rc = sock_sendmsg (sock, &msg, iov.iov_len);
- set_fs (oldmm);
-
- if (rc < 0)
- return (rc);
-
- if (rc == 0) {
- CERROR ("Unexpected zero rc\n");
- return (-ECONNABORTED);
- }
-
- buffer = ((char *)buffer) + rc;
- nob -= rc;
- }
-
- return (0);
+ return ksocknal_lib_sock_write(sock, buffer, nob);
}
int
ksocknal_sock_read (struct socket *sock, void *buffer, int nob)
{
- int rc;
- mm_segment_t oldmm = get_fs();
-
- while (nob > 0) {
- struct iovec iov = {
- .iov_base = buffer,
- .iov_len = nob
- };
- struct msghdr msg = {
- .msg_name = NULL,
- .msg_namelen = 0,
- .msg_iov = &iov,
- .msg_iovlen = 1,
- .msg_control = NULL,
- .msg_controllen = 0,
- .msg_flags = 0
- };
-
- set_fs (KERNEL_DS);
- rc = sock_recvmsg (sock, &msg, iov.iov_len, 0);
- set_fs (oldmm);
-
- if (rc < 0)
- return (rc);
-
- if (rc == 0)
- return (-ECONNABORTED);
-
- buffer = ((char *)buffer) + rc;
- nob -= rc;
- }
-
- return (0);
+ return ksocknal_lib_sock_read(sock, buffer, nob);
}
int
-ksocknal_hello (struct socket *sock, ptl_nid_t *nid, int *type,
- __u64 *incarnation)
+ksocknal_send_hello (ksock_conn_t *conn, __u32 *ipaddrs, int nipaddrs)
{
- int rc;
+ /* CAVEAT EMPTOR: this byte flips 'ipaddrs' */
+ struct socket *sock = conn->ksnc_sock;
ptl_hdr_t hdr;
ptl_magicversion_t *hmv = (ptl_magicversion_t *)&hdr.dest_nid;
- char ipbuf[PTL_NALFMT_SIZE];
- char ipbuf2[PTL_NALFMT_SIZE];
+ int i;
+ int rc;
- LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
+ LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE);
+ LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES);
- memset (&hdr, 0, sizeof (hdr));
- hmv->magic = __cpu_to_le32 (PORTALS_PROTO_MAGIC);
- hmv->version_major = __cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
- hmv->version_minor = __cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
+ /* No need for getconnsock/putconnsock */
+ LASSERT (!conn->ksnc_closing);
- hdr.src_nid = __cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid);
- hdr.type = __cpu_to_le32 (PTL_MSG_HELLO);
+ LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
+ hmv->magic = cpu_to_le32 (PORTALS_PROTO_MAGIC);
+ hmv->version_major = cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR);
+ hmv->version_minor = cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR);
- hdr.msg.hello.type = __cpu_to_le32 (*type);
+ hdr.src_nid = cpu_to_le64 (ksocknal_lib.libnal_ni.ni_pid.nid);
+ hdr.type = cpu_to_le32 (PTL_MSG_HELLO);
+ hdr.payload_length = cpu_to_le32 (nipaddrs * sizeof(*ipaddrs));
+
+ hdr.msg.hello.type = cpu_to_le32 (conn->ksnc_type);
hdr.msg.hello.incarnation =
- __cpu_to_le64 (ksocknal_data.ksnd_incarnation);
+ cpu_to_le64 (ksocknal_data.ksnd_incarnation);
- /* Assume sufficient socket buffering for this message */
- rc = ksocknal_sock_write (sock, &hdr, sizeof (hdr));
+ /* Receiver is eager */
+ rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr));
if (rc != 0) {
- CERROR ("Error %d sending HELLO to "LPX64" %s\n",
- rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
+ CERROR ("Error %d sending HELLO hdr to %u.%u.%u.%u/%d\n",
+ rc, HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
return (rc);
}
+ if (nipaddrs == 0)
+ return (0);
+
+ for (i = 0; i < nipaddrs; i++) {
+ ipaddrs[i] = __cpu_to_le32 (ipaddrs[i]);
+ }
+
+ rc = ksocknal_sock_write (sock, ipaddrs, nipaddrs * sizeof(*ipaddrs));
+ if (rc != 0)
+ CERROR ("Error %d sending HELLO payload (%d)"
+ " to %u.%u.%u.%u/%d\n", rc, nipaddrs,
+ HIPQUAD(conn->ksnc_ipaddr), conn->ksnc_port);
+ return (rc);
+}
+
+int
+ksocknal_invert_type(int type)
+{
+ switch (type)
+ {
+ case SOCKNAL_CONN_ANY:
+ case SOCKNAL_CONN_CONTROL:
+ return (type);
+ case SOCKNAL_CONN_BULK_IN:
+ return SOCKNAL_CONN_BULK_OUT;
+ case SOCKNAL_CONN_BULK_OUT:
+ return SOCKNAL_CONN_BULK_IN;
+ default:
+ return (SOCKNAL_CONN_NONE);
+ }
+}
+
+int
+ksocknal_recv_hello (ksock_conn_t *conn, ptl_nid_t *nid,
+ __u64 *incarnation, __u32 *ipaddrs)
+{
+ struct socket *sock = conn->ksnc_sock;
+ int rc;
+ int nips;
+ int i;
+ int type;
+ ptl_hdr_t hdr;
+ ptl_magicversion_t *hmv;
+ char ipbuf[PTL_NALFMT_SIZE];
+
+ hmv = (ptl_magicversion_t *)&hdr.dest_nid;
+ LASSERT (sizeof (*hmv) == sizeof (hdr.dest_nid));
+
rc = ksocknal_sock_read (sock, hmv, sizeof (*hmv));
if (rc != 0) {
- CERROR ("Error %d reading HELLO from "LPX64" %s\n",
- rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
+ CERROR ("Error %d reading HELLO from %u.%u.%u.%u\n",
+ rc, HIPQUAD(conn->ksnc_ipaddr));
return (rc);
}
- if (hmv->magic != __le32_to_cpu (PORTALS_PROTO_MAGIC)) {
- CERROR ("Bad magic %#08x (%#08x expected) from "LPX64" %s\n",
- __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC, *nid,
- portals_nid2str(SOCKNAL, *nid, ipbuf));
+ if (hmv->magic != le32_to_cpu (PORTALS_PROTO_MAGIC)) {
+ CERROR ("Bad magic %#08x (%#08x expected) from %u.%u.%u.%u\n",
+ __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC,
+ HIPQUAD(conn->ksnc_ipaddr));
return (-EPROTO);
}
- if (hmv->version_major != __cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
- hmv->version_minor != __cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
+ if (hmv->version_major != cpu_to_le16 (PORTALS_PROTO_VERSION_MAJOR) ||
+ hmv->version_minor != cpu_to_le16 (PORTALS_PROTO_VERSION_MINOR)) {
CERROR ("Incompatible protocol version %d.%d (%d.%d expected)"
- " from "LPX64" %s\n",
- __le16_to_cpu (hmv->version_major),
- __le16_to_cpu (hmv->version_minor),
+ " from %u.%u.%u.%u\n",
+ le16_to_cpu (hmv->version_major),
+ le16_to_cpu (hmv->version_minor),
PORTALS_PROTO_VERSION_MAJOR,
PORTALS_PROTO_VERSION_MINOR,
- *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
+ HIPQUAD(conn->ksnc_ipaddr));
return (-EPROTO);
}
-#if (PORTALS_PROTO_VERSION_MAJOR != 0)
-# error "This code only understands protocol version 0.x"
+#if (PORTALS_PROTO_VERSION_MAJOR != 1)
+# error "This code only understands protocol version 1.x"
#endif
- /* version 0 sends magic/version as the dest_nid of a 'hello' header,
- * so read the rest of it in now... */
+ /* version 1 sends magic/version as the dest_nid of a 'hello'
+ * header, followed by payload full of interface IP addresses.
+ * Read the rest of it in now... */
rc = ksocknal_sock_read (sock, hmv + 1, sizeof (hdr) - sizeof (*hmv));
if (rc != 0) {
- CERROR ("Error %d reading rest of HELLO hdr from "LPX64" %s\n",
- rc, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf));
+ CERROR ("Error %d reading rest of HELLO hdr from %u.%u.%u.%u\n",
+ rc, HIPQUAD(conn->ksnc_ipaddr));
return (rc);
}
/* ...and check we got what we expected */
- if (hdr.type != __cpu_to_le32 (PTL_MSG_HELLO) ||
- hdr.payload_length != __cpu_to_le32 (0)) {
- CERROR ("Expecting a HELLO hdr with 0 payload,"
- " but got type %d with %d payload from "LPX64" %s\n",
- __le32_to_cpu (hdr.type),
- __le32_to_cpu (hdr.payload_length), *nid,
- portals_nid2str(SOCKNAL, *nid, ipbuf));
+ if (hdr.type != cpu_to_le32 (PTL_MSG_HELLO)) {
+ CERROR ("Expecting a HELLO hdr,"
+ " but got type %d from %u.%u.%u.%u\n",
+ le32_to_cpu (hdr.type),
+ HIPQUAD(conn->ksnc_ipaddr));
return (-EPROTO);
}
- if (__le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
- CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY\n");
+ if (le64_to_cpu(hdr.src_nid) == PTL_NID_ANY) {
+ CERROR("Expecting a HELLO hdr with a NID, but got PTL_NID_ANY"
+ "from %u.%u.%u.%u\n", HIPQUAD(conn->ksnc_ipaddr));
return (-EPROTO);
}
if (*nid == PTL_NID_ANY) { /* don't know peer's nid yet */
- *nid = __le64_to_cpu(hdr.src_nid);
- } else if (*nid != __le64_to_cpu (hdr.src_nid)) {
- CERROR ("Connected to nid "LPX64" %s, but expecting "LPX64" %s\n",
- __le64_to_cpu (hdr.src_nid),
- portals_nid2str(SOCKNAL,
- __le64_to_cpu(hdr.src_nid),
- ipbuf),
- *nid, portals_nid2str(SOCKNAL, *nid, ipbuf2));
+ *nid = le64_to_cpu(hdr.src_nid);
+ } else if (*nid != le64_to_cpu (hdr.src_nid)) {
+ LCONSOLE_ERROR("Connected successfully to nid "LPX64" on host "
+ "%u.%u.%u.%u, but they claimed they were nid "
+ LPX64" (%s); please check your Lustre "
+ "configuration.\n",
+ *nid, HIPQUAD(conn->ksnc_ipaddr),
+ le64_to_cpu(hdr.src_nid),
+ portals_nid2str(SOCKNAL,
+ le64_to_cpu(hdr.src_nid),
+ ipbuf));
+
return (-EPROTO);
}
- if (*type == SOCKNAL_CONN_NONE) {
+ type = __le32_to_cpu(hdr.msg.hello.type);
+
+ if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
/* I've accepted this connection; peer determines type */
- *type = __le32_to_cpu(hdr.msg.hello.type);
- switch (*type) {
- case SOCKNAL_CONN_ANY:
- case SOCKNAL_CONN_CONTROL:
- break;
- case SOCKNAL_CONN_BULK_IN:
- *type = SOCKNAL_CONN_BULK_OUT;
- break;
- case SOCKNAL_CONN_BULK_OUT:
- *type = SOCKNAL_CONN_BULK_IN;
- break;
- default:
- CERROR ("Unexpected type %d from "LPX64" %s\n",
- *type, *nid,
- portals_nid2str(SOCKNAL, *nid, ipbuf));
+ conn->ksnc_type = ksocknal_invert_type(type);
+ if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
+ CERROR ("Unexpected type %d from "LPX64"@%u.%u.%u.%u\n",
+ type, *nid, HIPQUAD(conn->ksnc_ipaddr));
return (-EPROTO);
}
- } else if (__le32_to_cpu(hdr.msg.hello.type) != SOCKNAL_CONN_NONE) {
- CERROR ("Mismatched types: me %d "LPX64" %s %d\n",
- *type, *nid, portals_nid2str(SOCKNAL, *nid, ipbuf),
- __le32_to_cpu(hdr.msg.hello.type));
+ } else if (ksocknal_invert_type(type) != conn->ksnc_type) {
+ CERROR ("Mismatched types: me %d, "LPX64"@%u.%u.%u.%u %d\n",
+ conn->ksnc_type, *nid, HIPQUAD(conn->ksnc_ipaddr),
+ le32_to_cpu(hdr.msg.hello.type));
return (-EPROTO);
}
- *incarnation = __le64_to_cpu(hdr.msg.hello.incarnation);
+ *incarnation = le64_to_cpu(hdr.msg.hello.incarnation);
- return (0);
-}
-
-int
-ksocknal_setup_sock (struct socket *sock)
-{
- mm_segment_t oldmm = get_fs ();
- int rc;
- int option;
- struct linger linger;
+ nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32);
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
- sock->sk->sk_allocation = GFP_NOFS;
-#else
- sock->sk->allocation = GFP_NOFS;
-#endif
-
- /* Ensure this socket aborts active sends immediately when we close
- * it. */
-
- linger.l_onoff = 0;
- linger.l_linger = 0;
-
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_LINGER,
- (char *)&linger, sizeof (linger));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set SO_LINGER: %d\n", rc);
- return (rc);
+ if (nips > SOCKNAL_MAX_INTERFACES ||
+ nips * sizeof(__u32) != __le32_to_cpu (hdr.payload_length)) {
+ CERROR("Bad payload length %d from "LPX64"@%u.%u.%u.%u\n",
+ __le32_to_cpu (hdr.payload_length),
+ *nid, HIPQUAD(conn->ksnc_ipaddr));
}
- option = -1;
- set_fs (KERNEL_DS);
- rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
+ if (nips == 0)
+ return (0);
+
+ rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs));
if (rc != 0) {
- CERROR ("Can't set SO_LINGER2: %d\n", rc);
+ CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n",
+ rc, *nid, HIPQUAD(conn->ksnc_ipaddr));
return (rc);
}
-#if SOCKNAL_USE_KEEPALIVES
- /* Keepalives: If 3/4 of the timeout elapses, start probing every
- * second until the timeout elapses. */
+ for (i = 0; i < nips; i++) {
+ ipaddrs[i] = __le32_to_cpu(ipaddrs[i]);
- option = (ksocknal_tunables.ksnd_io_timeout * 3) / 4;
- set_fs (KERNEL_DS);
- rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
- return (rc);
- }
-
- option = 1;
- set_fs (KERNEL_DS);
- rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
- return (rc);
- }
-
- option = ksocknal_tunables.ksnd_io_timeout / 4;
- set_fs (KERNEL_DS);
- rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
- return (rc);
+ if (ipaddrs[i] == 0) {
+ CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n",
+ i, *nid, HIPQUAD(conn->ksnc_ipaddr));
+ return (-EPROTO);
+ }
}
- option = 1;
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_KEEPALIVE,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set SO_KEEPALIVE: %d\n", rc);
- return (rc);
- }
-#endif
- return (0);
+ return (nips);
}
int
ksocknal_connect_peer (ksock_route_t *route, int type)
{
- struct sockaddr_in peer_addr;
- mm_segment_t oldmm = get_fs();
- struct timeval tv;
- int fd;
struct socket *sock;
int rc;
- char ipbuf[PTL_NALFMT_SIZE];
+ int port;
+ int may_retry;
- rc = sock_create (PF_INET, SOCK_STREAM, 0, &sock);
- if (rc != 0) {
- CERROR ("Can't create autoconnect socket: %d\n", rc);
- return (rc);
- }
+ /* Iterate through reserved ports. When typed connections are
+ * used, we will need to bind to multiple ports, but we only know
+ * this at connect time. But, by that time we've already called
+ * bind() so we need a new socket. */
- /* Ugh; have to map_fd for compatibility with sockets passed in
- * from userspace. And we actually need the sock->file refcounting
- * that this gives you :) */
+ for (port = 1023; port > 512; --port) {
- fd = sock_map_fd (sock);
- if (fd < 0) {
- sock_release (sock);
- CERROR ("sock_map_fd error %d\n", fd);
- return (fd);
- }
+ rc = ksocknal_lib_connect_sock(&sock, &may_retry, route, port);
- /* NB the fd now owns the ref on sock->file */
- LASSERT (sock->file != NULL);
- LASSERT (file_count(sock->file) == 1);
-
- /* Set the socket timeouts, so our connection attempt completes in
- * finite time */
- tv.tv_sec = ksocknal_tunables.ksnd_io_timeout;
- tv.tv_usec = 0;
-
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDTIMEO,
- (char *)&tv, sizeof (tv));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set send timeout %d: %d\n",
- ksocknal_tunables.ksnd_io_timeout, rc);
- goto out;
- }
-
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVTIMEO,
- (char *)&tv, sizeof (tv));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set receive timeout %d: %d\n",
- ksocknal_tunables.ksnd_io_timeout, rc);
- goto out;
- }
-
- {
- int option = 1;
-
- set_fs (KERNEL_DS);
- rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_NODELAY,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't disable nagle: %d\n", rc);
- goto out;
- }
- }
-
- if (route->ksnr_buffer_size != 0) {
- int option = route->ksnr_buffer_size;
-
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_SNDBUF,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set send buffer %d: %d\n",
- route->ksnr_buffer_size, rc);
- goto out;
+ if (rc == 0) {
+ rc = ksocknal_create_conn(route, sock, type);
+ cfs_put_file(KSN_SOCK2FILE(sock));
+ return rc;
}
- set_fs (KERNEL_DS);
- rc = sock_setsockopt (sock, SOL_SOCKET, SO_RCVBUF,
- (char *)&option, sizeof (option));
- set_fs (oldmm);
- if (rc != 0) {
- CERROR ("Can't set receive buffer %d: %d\n",
- route->ksnr_buffer_size, rc);
- goto out;
- }
- }
-
- memset (&peer_addr, 0, sizeof (peer_addr));
- peer_addr.sin_family = AF_INET;
- peer_addr.sin_port = htons (route->ksnr_port);
- peer_addr.sin_addr.s_addr = htonl (route->ksnr_ipaddr);
-
- rc = sock->ops->connect (sock, (struct sockaddr *)&peer_addr,
- sizeof (peer_addr), sock->file->f_flags);
- if (rc != 0) {
- CERROR ("Error %d connecting to "LPX64" %s\n", rc,
- route->ksnr_peer->ksnp_nid,
- portals_nid2str(SOCKNAL,
- route->ksnr_peer->ksnp_nid,
- ipbuf));
- goto out;
- }
-
- rc = ksocknal_create_conn (route, sock, route->ksnr_irq_affinity, type);
- if (rc == 0) {
- /* Take an extra ref on sock->file to compensate for the
- * upcoming close which will lose fd's ref on it. */
- get_file (sock->file);
+ if (!may_retry)
+ return rc;
}
- out:
- sys_close (fd);
- return (rc);
+ CERROR("Out of ports trying to bind to a reserved port\n");
+ return (-EADDRINUSE);
}
void
ksocknal_autoconnect (ksock_route_t *route)
{
- LIST_HEAD (zombies);
+ CFS_LIST_HEAD (zombies);
ksock_tx_t *tx;
ksock_peer_t *peer;
unsigned long flags;
- int rc;
int type;
-
+ int rc = 0;
+
+ write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
+
for (;;) {
- for (type = 0; type < SOCKNAL_CONN_NTYPES; type++)
- if ((route->ksnr_connecting & (1 << type)) != 0)
- break;
- LASSERT (type < SOCKNAL_CONN_NTYPES);
+ if (!ksocknal_tunables.ksnd_typed_conns) {
+ if ((route->ksnr_connected & (1<<SOCKNAL_CONN_ANY)) == 0)
+ type = SOCKNAL_CONN_ANY;
+ else
+ break; /* got connected while route queued */
+ } else {
+ if ((route->ksnr_connected & (1<<SOCKNAL_CONN_CONTROL)) == 0)
+ type = SOCKNAL_CONN_CONTROL;
+ else if ((route->ksnr_connected & (1<<SOCKNAL_CONN_BULK_IN)) == 0)
+ type = SOCKNAL_CONN_BULK_IN;
+ else if ((route->ksnr_connected & (1<<SOCKNAL_CONN_BULK_OUT)) == 0)
+ type = SOCKNAL_CONN_BULK_OUT;
+ else
+ break; /* got connected while route queued */
+ }
- rc = ksocknal_connect_peer (route, type);
+ write_unlock_irqrestore(&ksocknal_data.ksnd_global_lock, flags);
+ rc = ksocknal_connect_peer (route, type);
if (rc != 0)
- break;
-
- /* successfully autoconnected: create_conn did the
- * route/conn binding and scheduled any blocked packets */
+ goto failed;
- if (route->ksnr_connecting == 0) {
- /* No more connections required */
- return;
- }
+ write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
}
- /* Connection attempt failed */
+ LASSERT (route->ksnr_connecting);
+ route->ksnr_connecting = 0;
+ write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
+ return;
+
+ failed:
+ switch (rc) {
+ /* "normal" errors */
+ case -ECONNREFUSED:
+ LCONSOLE_ERROR("Connection was refused by host %u.%u.%u.%u on "
+ "port %d; check that Lustre is running on that "
+ "node.\n",
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+ break;
+ case -EHOSTUNREACH:
+ case -ENETUNREACH:
+ LCONSOLE_ERROR("Host %u.%u.%u.%u was unreachable; the network "
+ "or that node may be down, or Lustre may be "
+ "misconfigured.\n",
+ HIPQUAD(route->ksnr_ipaddr));
+ break;
+ case -ETIMEDOUT:
+ LCONSOLE_ERROR("Connecting to host %u.%u.%u.%u on port %d took "
+ "too long; that node may be hung or "
+ "experiencing high load.\n",
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+ break;
+ /* errors that should be rare */
+ case -EPROTO:
+ LCONSOLE_ERROR("Protocol error connecting to host %u.%u.%u.%u "
+ "on port %d: Is it running a compatible version"
+ " of Lustre?\n",
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+ break;
+ case -EADDRINUSE:
+ LCONSOLE_ERROR("No privileged ports available to connect to "
+ "host %u.%u.%u.%u on port %d\n",
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+ break;
+ default:
+ LCONSOLE_ERROR("Unexpected error %d connecting to "
+ "host %u.%u.%u.%u on port %d\n", rc,
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+ break;
+ }
write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
peer = route->ksnr_peer;
+ LASSERT (route->ksnr_connecting);
route->ksnr_connecting = 0;
/* This is a retry rather than a new connection */
LASSERT (route->ksnr_retry_interval != 0);
- route->ksnr_timeout = jiffies + route->ksnr_retry_interval;
+ route->ksnr_timeout = cfs_time_add(cfs_time_current(),
+ route->ksnr_retry_interval);
route->ksnr_retry_interval = MIN (route->ksnr_retry_interval * 2,
SOCKNAL_MAX_RECONNECT_INTERVAL);
} while (!list_empty (&peer->ksnp_tx_queue));
}
- /* make this route least-favourite for re-selection */
+#if 0 /* irrelevent with only eager routes */
if (!route->ksnr_deleted) {
+ /* make this route least-favourite for re-selection */
list_del(&route->ksnr_list);
list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
}
-
+#endif
write_unlock_irqrestore (&ksocknal_data.ksnd_global_lock, flags);
while (!list_empty (&zombies)) {
tx = list_entry (zombies.next, ksock_tx_t, tx_list);
CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
- NTOH__u32 (tx->tx_hdr->type),
- NTOH__u32 (tx->tx_hdr->payload_length),
- NTOH__u64 (tx->tx_hdr->src_nid),
+ le32_to_cpu (tx->tx_hdr->type),
+ le32_to_cpu (tx->tx_hdr->payload_length),
+ le64_to_cpu (tx->tx_hdr->src_nid),
portals_nid2str(SOCKNAL,
- NTOH__u64(tx->tx_hdr->src_nid),
+ le64_to_cpu(tx->tx_hdr->src_nid),
ipbuf),
- NTOH__u64 (tx->tx_hdr->dest_nid),
+ le64_to_cpu (tx->tx_hdr->dest_nid),
portals_nid2str(SOCKNAL,
- NTOH__u64(tx->tx_hdr->src_nid),
+ le64_to_cpu(tx->tx_hdr->src_nid),
ipbuf2));
list_del (&tx->tx_list);
if (!list_empty (&ksocknal_data.ksnd_autoconnectd_routes)) {
route = list_entry (ksocknal_data.ksnd_autoconnectd_routes.next,
ksock_route_t, ksnr_connect_list);
-
+
list_del (&route->ksnr_connect_list);
spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
ksocknal_autoconnect (route);
ksocknal_put_route (route);
- spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
+ spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock,
+ flags);
continue;
}
-
- spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
- rc = wait_event_interruptible (ksocknal_data.ksnd_autoconnectd_waitq,
- ksocknal_data.ksnd_shuttingdown ||
- !list_empty (&ksocknal_data.ksnd_autoconnectd_routes));
+ spin_unlock_irqrestore(&ksocknal_data.ksnd_autoconnectd_lock,
+ flags);
+
+ rc = wait_event_interruptible(ksocknal_data.ksnd_autoconnectd_waitq,
+ ksocknal_data.ksnd_shuttingdown ||
+ !list_empty(&ksocknal_data.ksnd_autoconnectd_routes));
- spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
+ spin_lock_irqsave(&ksocknal_data.ksnd_autoconnectd_lock, flags);
}
spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
/* We're called with a shared lock on ksnd_global_lock */
ksock_conn_t *conn;
struct list_head *ctmp;
- ksock_sched_t *sched;
list_for_each (ctmp, &peer->ksnp_conns) {
conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
- sched = conn->ksnc_scheduler;
/* Don't need the {get,put}connsock dance to deref ksnc_sock... */
LASSERT (!conn->ksnc_closing);
-
+
+ if (SOCK_ERROR(conn->ksnc_sock) != 0) {
+ atomic_inc (&conn->ksnc_refcount);
+
+ switch (SOCK_ERROR(conn->ksnc_sock)) {
+ case ECONNRESET:
+ LCONSOLE_WARN("A connection with %u.%u.%u.%u "
+ "was reset; they may have "
+ "rebooted.\n",
+ HIPQUAD(conn->ksnc_ipaddr));
+ break;
+ case ETIMEDOUT:
+ LCONSOLE_WARN("A connection with %u.%u.%u.%u "
+ "timed out; the network or that "
+ "node may be down.\n",
+ HIPQUAD(conn->ksnc_ipaddr));
+ break;
+ default:
+ LCONSOLE_WARN("An unexpected network error "
+ "occurred with %u.%u.%u.%u: %d\n",
+ HIPQUAD(conn->ksnc_ipaddr),
+ SOCK_ERROR(conn->ksnc_sock));
+ break;
+ }
+
+ /* Something (e.g. failed keepalive) set the socket error */
+ CDEBUG(D_HA,"Socket error %d: "LPX64" %p %d.%d.%d.%d\n",
+ SOCK_ERROR(conn->ksnc_sock), peer->ksnp_nid,
+ conn, HIPQUAD(conn->ksnc_ipaddr));
+
+ return (conn);
+ }
+
if (conn->ksnc_rx_started &&
- time_after_eq (jiffies, conn->ksnc_rx_deadline)) {
+ cfs_time_aftereq (cfs_time_current(),
+ conn->ksnc_rx_deadline)) {
/* Timed out incomplete incoming message */
atomic_inc (&conn->ksnc_refcount);
+ LCONSOLE_ERROR("A timeout occurred receiving data from "
+ "%u.%u.%u.%u; the network or that node "
+ "may be down.\n",
+ HIPQUAD(conn->ksnc_ipaddr));
CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
- peer->ksnp_nid, conn, HIPQUAD(conn->ksnc_ipaddr));
+ peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr));
return (conn);
}
-
+
if ((!list_empty (&conn->ksnc_tx_queue) ||
- conn->ksnc_sock->sk->sk_wmem_queued != 0) &&
- time_after_eq (jiffies, conn->ksnc_tx_deadline)) {
- /* Timed out messages queued for sending, or
- * messages buffered in the socket's send buffer */
+ SOCK_WMEM_QUEUED(conn->ksnc_sock) != 0) &&
+ cfs_time_aftereq (cfs_time_current(),
+ conn->ksnc_tx_deadline)) {
+ /* Timed out messages queued for sending or
+ * buffered in the socket's send buffer */
atomic_inc (&conn->ksnc_refcount);
- CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
- peer->ksnp_nid,
+ LCONSOLE_ERROR("A timeout occurred sending data to "
+ "%u.%u.%u.%u; the network or that node "
+ "may be down.\n",
+ HIPQUAD(conn->ksnc_ipaddr));
+ CERROR ("Timed out TX to "LPX64" %s%d %p %d.%d.%d.%d\n",
+ peer->ksnp_nid,
list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
- conn->ksnc_sock->sk->sk_wmem_queued, conn,
+ SOCK_WMEM_QUEUED(conn->ksnc_sock), conn,
HIPQUAD(conn->ksnc_ipaddr));
return (conn);
}
list_for_each (ptmp, peers) {
peer = list_entry (ptmp, ksock_peer_t, ksnp_list);
conn = ksocknal_find_timed_out_conn (peer);
-
+
if (conn != NULL) {
read_unlock (&ksocknal_data.ksnd_global_lock);
- CERROR ("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
- peer->ksnp_nid,
- HIPQUAD(conn->ksnc_ipaddr),
- conn->ksnc_port);
+ CERROR("Timeout out conn->"LPX64" ip %d.%d.%d.%d:%d\n",
+ peer->ksnp_nid, HIPQUAD(conn->ksnc_ipaddr),
+ conn->ksnc_port);
ksocknal_close_conn_and_siblings (conn, -ETIMEDOUT);
-
+
/* NB we won't find this one again, but we can't
* just proceed with the next peer, since we dropped
* ksnd_global_lock and it might be dead already! */
int
ksocknal_reaper (void *arg)
{
- wait_queue_t wait;
+ cfs_waitlink_t wait;
unsigned long flags;
ksock_conn_t *conn;
ksock_sched_t *sched;
struct list_head enomem_conns;
int nenomem_conns;
- int timeout;
+ cfs_duration_t timeout;
int i;
int peer_index = 0;
- unsigned long deadline = jiffies;
-
+ cfs_time_t deadline = cfs_time_current();
+
kportal_daemonize ("ksocknal_reaper");
kportal_blockallsigs ();
- INIT_LIST_HEAD(&enomem_conns);
- init_waitqueue_entry (&wait, current);
+ CFS_INIT_LIST_HEAD(&enomem_conns);
+ cfs_waitlink_init (&wait);
spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
conn = list_entry (ksocknal_data.ksnd_deathrow_conns.next,
ksock_conn_t, ksnc_list);
list_del (&conn->ksnc_list);
-
+
spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
ksocknal_terminate_conn (conn);
conn = list_entry (ksocknal_data.ksnd_zombie_conns.next,
ksock_conn_t, ksnc_list);
list_del (&conn->ksnc_list);
-
+
spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
ksocknal_destroy_conn (conn);
LASSERT (conn->ksnc_tx_scheduled);
conn->ksnc_tx_ready = 1;
- list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
- wake_up (&sched->kss_waitq);
+ list_add_tail(&conn->ksnc_tx_list,&sched->kss_tx_conns);
+ cfs_waitq_signal (&sched->kss_waitq);
spin_unlock_irqrestore (&sched->kss_lock, flags);
nenomem_conns++;
}
-
+
/* careful with the jiffy wrap... */
- while ((timeout = (int)(deadline - jiffies)) <= 0) {
+ while ((timeout = cfs_time_sub(deadline,
+ cfs_time_current())) <= 0) {
const int n = 4;
const int p = 1;
int chunk = ksocknal_data.ksnd_peer_hash_size;
-
+
/* Time to check for timeouts on a few more peers: I do
* checks every 'p' seconds on a proportion of the peer
* table and I need to check every connection 'n' times
* timeout interval. */
if (ksocknal_tunables.ksnd_io_timeout > n * p)
- chunk = (chunk * n * p) /
+ chunk = (chunk * n * p) /
ksocknal_tunables.ksnd_io_timeout;
if (chunk == 0)
chunk = 1;
for (i = 0; i < chunk; i++) {
ksocknal_check_peer_timeouts (peer_index);
- peer_index = (peer_index + 1) %
+ peer_index = (peer_index + 1) %
ksocknal_data.ksnd_peer_hash_size;
}
- deadline += p * HZ;
+ deadline = cfs_time_add(deadline, cfs_time_seconds(p));
}
if (nenomem_conns != 0) {
* if any go back on my enomem list. */
timeout = SOCKNAL_ENOMEM_RETRY;
}
- ksocknal_data.ksnd_reaper_waketime = jiffies + timeout;
+ ksocknal_data.ksnd_reaper_waketime =
+ cfs_time_add(cfs_time_current(), timeout);
set_current_state (TASK_INTERRUPTIBLE);
- add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
+ cfs_waitq_add (&ksocknal_data.ksnd_reaper_waitq, &wait);
if (!ksocknal_data.ksnd_shuttingdown &&
list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
list_empty (&ksocknal_data.ksnd_zombie_conns))
- schedule_timeout (timeout);
+ cfs_waitq_timedwait (&wait, timeout);
set_current_state (TASK_RUNNING);
- remove_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
+ cfs_waitq_del (&ksocknal_data.ksnd_reaper_waitq, &wait);
spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
}