*/
#include "socknal.h"
-
-int ksocknal_io_timeout = SOCKNAL_IO_TIMEOUT;
-#if SOCKNAL_ZC
-int ksocknal_do_zc = 1;
-int ksocknal_zc_min_frag = SOCKNAL_ZC_MIN_FRAG;
+#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+# include <linux/syscalls.h>
#endif
/*
*
*/
int
-ksocknal_read(nal_cb_t *nal, void *private, void *dst_addr,
- user_ptr src_addr, size_t len)
-{
- CDEBUG(D_NET, LPX64": reading %ld bytes from %p -> %p\n",
- nal->ni.nid, (long)len, src_addr, dst_addr);
-
- memcpy( dst_addr, src_addr, len );
- return 0;
-}
-
-int
-ksocknal_write(nal_cb_t *nal, void *private, user_ptr dst_addr,
- void *src_addr, size_t len)
-{
- CDEBUG(D_NET, LPX64": writing %ld bytes from %p -> %p\n",
- nal->ni.nid, (long)len, src_addr, dst_addr);
-
- memcpy( dst_addr, src_addr, len );
- return 0;
-}
-
-int
-ksocknal_callback (nal_cb_t * nal, void *private, lib_eq_t *eq,
- ptl_event_t *ev)
-{
- CDEBUG(D_NET, LPX64": callback eq %p ev %p\n",
- nal->ni.nid, eq, ev);
-
- if (eq->event_callback != NULL)
- eq->event_callback(ev);
-
- return 0;
-}
-
-void *
-ksocknal_malloc(nal_cb_t *nal, size_t len)
-{
- void *buf;
-
- PORTAL_ALLOC(buf, len);
-
- if (buf != NULL)
- memset(buf, 0, len);
-
- return (buf);
-}
-
-void
-ksocknal_free(nal_cb_t *nal, void *buf, size_t len)
-{
- PORTAL_FREE(buf, len);
-}
-
-void
-ksocknal_printf(nal_cb_t *nal, const char *fmt, ...)
-{
- va_list ap;
- char msg[256];
-
- va_start (ap, fmt);
- vsnprintf (msg, sizeof (msg), fmt, ap); /* sprint safely */
- va_end (ap);
-
- msg[sizeof (msg) - 1] = 0; /* ensure terminated */
-
- CDEBUG (D_NET, "%s", msg);
-}
-
-void
-ksocknal_cli(nal_cb_t *nal, unsigned long *flags)
-{
- ksock_nal_data_t *data = nal->nal_data;
-
- spin_lock(&data->ksnd_nal_cb_lock);
-}
-
-void
-ksocknal_sti(nal_cb_t *nal, unsigned long *flags)
-{
- ksock_nal_data_t *data;
- data = nal->nal_data;
-
- spin_unlock(&data->ksnd_nal_cb_lock);
-}
-
-int
-ksocknal_dist(nal_cb_t *nal, ptl_nid_t nid, unsigned long *dist)
+ksocknal_dist(lib_nal_t *nal, ptl_nid_t nid, unsigned long *dist)
{
/* I would guess that if ksocknal_get_peer (nid) == NULL,
and we're not routing, then 'nid' is very distant :) */
- if ( nal->ni.nid == nid ) {
+ if (nal->libnal_ni.ni_pid.nid == nid) {
*dist = 0;
} else {
*dist = 1;
return 0;
}
-ksock_ltx_t *
-ksocknal_get_ltx (int may_block)
-{
- unsigned long flags;
- ksock_ltx_t *ltx = NULL;
-
- for (;;) {
- spin_lock_irqsave (&ksocknal_data.ksnd_idle_ltx_lock, flags);
-
- if (!list_empty (&ksocknal_data.ksnd_idle_ltx_list)) {
- ltx = list_entry(ksocknal_data.ksnd_idle_ltx_list.next,
- ksock_ltx_t, ltx_tx.tx_list);
- list_del (<x->ltx_tx.tx_list);
- ksocknal_data.ksnd_active_ltxs++;
- break;
- }
-
- if (!may_block) {
- if (!list_empty(&ksocknal_data.ksnd_idle_nblk_ltx_list)) {
- ltx = list_entry(ksocknal_data.ksnd_idle_nblk_ltx_list.next,
- ksock_ltx_t, ltx_tx.tx_list);
- list_del (<x->ltx_tx.tx_list);
- ksocknal_data.ksnd_active_ltxs++;
- }
- break;
- }
-
- spin_unlock_irqrestore(&ksocknal_data.ksnd_idle_ltx_lock,
- flags);
-
- wait_event (ksocknal_data.ksnd_idle_ltx_waitq,
- !list_empty (&ksocknal_data.ksnd_idle_ltx_list));
- }
-
- spin_unlock_irqrestore (&ksocknal_data.ksnd_idle_ltx_lock, flags);
-
- return (ltx);
-}
-
void
-ksocknal_put_ltx (ksock_ltx_t *ltx)
+ksocknal_free_ltx (ksock_ltx_t *ltx)
{
- unsigned long flags;
-
- spin_lock_irqsave (&ksocknal_data.ksnd_idle_ltx_lock, flags);
-
- ksocknal_data.ksnd_active_ltxs--;
- list_add_tail (<x->ltx_tx.tx_list, ltx->ltx_idle);
-
- /* normal tx desc => wakeup anyone blocking for one */
- if (ltx->ltx_idle == &ksocknal_data.ksnd_idle_ltx_list &&
- waitqueue_active (&ksocknal_data.ksnd_idle_ltx_waitq))
- wake_up (&ksocknal_data.ksnd_idle_ltx_waitq);
-
- spin_unlock_irqrestore (&ksocknal_data.ksnd_idle_ltx_lock, flags);
+ atomic_dec(&ksocknal_data.ksnd_nactive_ltxs);
+ PORTAL_FREE(ltx, ltx->ltx_desc_size);
}
-#if SOCKNAL_ZC
+#if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
struct page *
ksocknal_kvaddr_to_page (unsigned long vaddr)
{
{
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 = !list_empty (&conn->ksnc_tx_queue) |
- (tx->tx_niov > 1) |
- (tx->tx_nkiov > 1);
-#if SOCKNAL_ZC
+#if (SOCKNAL_ZC && SOCKNAL_VADDR_ZC)
+ unsigned long vaddr = (unsigned long)iov->iov_base
int offset = vaddr & (PAGE_SIZE - 1);
- int zcsize = MIN (fragsize, PAGE_SIZE - offset);
+ int zcsize = MIN (iov->iov_len, PAGE_SIZE - offset);
struct page *page;
#endif
+ 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);
+ * or leave them alone. */
LASSERT (tx->tx_niov > 0);
-#if SOCKNAL_ZC
- if (ksocknal_do_zc &&
+#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)) &&
- zcsize >= ksocknal_zc_min_frag &&
(page = ksocknal_kvaddr_to_page (vaddr)) != NULL) {
+ int msgflg = MSG_DONTWAIT;
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);
+ if (!list_empty (&conn->ksnc_tx_queue) ||
+ zcsize < tx->tx_resid)
+ msgflg |= MSG_MORE;
+
+ rc = tcp_sendpage_zccd(sock, page, offset, zcsize, msgflg, &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};
+#if SOCKNAL_SINGLE_FRAG_TX
+ struct iovec scratch;
+ struct iovec *scratchiov = &scratch;
+ int niov = 1;
+#else
+ struct iovec *scratchiov = conn->ksnc_tx_scratch_iov;
+ int niov = tx->tx_niov;
+#endif
struct msghdr msg = {
.msg_name = NULL,
.msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
+ .msg_iov = scratchiov,
+ .msg_iovlen = niov,
.msg_control = NULL,
.msg_controllen = 0,
- .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
+ .msg_flags = MSG_DONTWAIT
};
mm_segment_t oldmm = get_fs();
+ int i;
+
+ for (nob = i = 0; i < niov; i++) {
+ scratchiov[i] = tx->tx_iov[i];
+ nob += scratchiov[i].iov_len;
+ }
+
+ if (!list_empty(&conn->ksnc_tx_queue) ||
+ nob < tx->tx_resid)
+ msg.msg_flags |= MSG_MORE;
set_fs (KERNEL_DS);
- rc = sock_sendmsg(sock, &msg, fragsize);
+ rc = sock_sendmsg(sock, &msg, nob);
set_fs (oldmm);
}
- if (rc <= 0)
+ if (rc <= 0) /* sent nothing? */
return (rc);
- tx->tx_resid -= rc;
+ nob = rc;
+ LASSERT (nob <= tx->tx_resid);
+ tx->tx_resid -= nob;
- if (rc < iov->iov_len) {
- /* didn't send whole iov entry... */
- iov->iov_base = (void *)(vaddr + rc);
- iov->iov_len -= rc;
- /* ...but did we send everything we tried to send? */
- return ((rc == fragsize) ? 1 : -EAGAIN);
- }
+ /* "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);
+ }
- tx->tx_iov++;
- tx->tx_niov--;
- return (1);
+ nob -= iov->iov_len;
+ tx->tx_iov = ++iov;
+ tx->tx_niov--;
+ } while (nob != 0);
+
+ return (rc);
}
int
{
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 = !list_empty (&conn->ksnc_tx_queue) |
- (tx->tx_nkiov > 1);
int rc;
-
+ int nob;
+
/* 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);
+ * or leave them alone. */
LASSERT (tx->tx_niov == 0);
LASSERT (tx->tx_nkiov > 0);
#if SOCKNAL_ZC
- if (ksocknal_do_zc &&
+ if (kiov->kiov_len >= 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)) &&
- fragsize >= ksocknal_zc_min_frag) {
+ (sock->sk->route_caps & (NETIF_F_IP_CSUM | NETIF_F_NO_CSUM | NETIF_F_HW_CSUM))) {
+ struct page *page = kiov->kiov_page;
+ int offset = kiov->kiov_offset;
+ int fragsize = kiov->kiov_len;
+ int msgflg = MSG_DONTWAIT;
CDEBUG(D_NET, "page %p + offset %x for %d\n",
- page, offset, fragsize);
+ page, offset, kiov->kiov_len);
+
+ if (!list_empty(&conn->ksnc_tx_queue) ||
+ fragsize < tx->tx_resid)
+ msgflg |= MSG_MORE;
- rc = tcp_sendpage_zccd(sock, page, offset, fragsize,
- more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT,
+ rc = tcp_sendpage_zccd(sock, page, offset, fragsize, msgflg,
&tx->tx_zccd);
} else
#endif
{
- char *addr = ((char *)kmap (page)) + offset;
- struct iovec fragiov = {.iov_base = addr,
- .iov_len = fragsize};
+#if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
+ struct iovec scratch;
+ struct iovec *scratchiov = &scratch;
+ int niov = 1;
+#else
+#warning "XXX risk of kmap deadlock on multiple frags..."
+ struct iovec *scratchiov = conn->ksnc_tx_scratch_iov;
+ int niov = tx->tx_nkiov;
+#endif
struct msghdr msg = {
.msg_name = NULL,
.msg_namelen = 0,
- .msg_iov = &fragiov,
- .msg_iovlen = 1,
+ .msg_iov = scratchiov,
+ .msg_iovlen = niov,
.msg_control = NULL,
.msg_controllen = 0,
- .msg_flags = more ? (MSG_DONTWAIT | MSG_MORE) : MSG_DONTWAIT
+ .msg_flags = MSG_DONTWAIT
};
mm_segment_t oldmm = get_fs();
+ int i;
+ for (nob = i = 0; i < niov; i++) {
+ scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
+ kiov[i].kiov_offset;
+ nob += scratchiov[i].iov_len = kiov[i].kiov_len;
+ }
+
+ if (!list_empty(&conn->ksnc_tx_queue) ||
+ nob < tx->tx_resid)
+ msg.msg_flags |= MSG_DONTWAIT;
+
set_fs (KERNEL_DS);
- rc = sock_sendmsg(sock, &msg, fragsize);
+ rc = sock_sendmsg(sock, &msg, nob);
set_fs (oldmm);
- kunmap (page);
+
+ for (i = 0; i < niov; i++)
+ kunmap(kiov[i].kiov_page);
}
- if (rc <= 0)
+ if (rc <= 0) /* sent nothing? */
return (rc);
- tx->tx_resid -= rc;
-
- if (rc < fragsize) {
- /* didn't send whole frag */
- kiov->kiov_offset = offset + rc;
- kiov->kiov_len = fragsize - rc;
- return (-EAGAIN);
- }
+ nob = rc;
+ LASSERT (nob <= tx->tx_resid);
+ tx->tx_resid -= nob;
+
+ 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);
- /* everything went */
- LASSERT (rc == fragsize);
- tx->tx_kiov++;
- tx->tx_nkiov--;
- return (1);
+ return (rc);
}
int
-ksocknal_sendmsg (ksock_conn_t *conn, ksock_tx_t *tx)
+ksocknal_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
{
- /* Return 0 on success, < 0 on error.
- * caller checks tx_resid to determine progress/completion */
int rc;
- ENTRY;
+ int bufnob;
if (ksocknal_data.ksnd_stall_tx != 0) {
set_current_state (TASK_UNINTERRUPTIBLE);
schedule_timeout (ksocknal_data.ksnd_stall_tx * HZ);
}
- rc = ksocknal_getconnsock (conn);
- if (rc != 0)
- return (rc);
-
- for (;;) {
- LASSERT (tx->tx_resid != 0);
+ LASSERT (tx->tx_resid != 0);
- if (conn->ksnc_closing) {
- rc = -ESHUTDOWN;
- break;
- }
+ rc = ksocknal_getconnsock (conn);
+ if (rc != 0) {
+ LASSERT (conn->ksnc_closing);
+ return (-ESHUTDOWN);
+ }
- if (tx->tx_niov != 0)
+ do {
+ if (ksocknal_data.ksnd_enomem_tx > 0) {
+ /* testing... */
+ ksocknal_data.ksnd_enomem_tx--;
+ rc = -EAGAIN;
+ } else if (tx->tx_niov != 0) {
rc = ksocknal_send_iov (conn, tx);
- else
+ } else {
rc = ksocknal_send_kiov (conn, tx);
+ }
- if (rc <= 0) { /* error or socket full? */
- /* NB: rc == 0 and rc == -EAGAIN both mean try
- * again later (linux stack returns -EAGAIN for
- * this, but Adaptech TOE returns 0) */
- if (rc == -EAGAIN)
- rc = 0;
- break;
+ bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
+ 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 = jiffies +
+ ksocknal_tunables.ksnd_io_timeout * HZ;
+ conn->ksnc_peer->ksnp_last_alive = jiffies;
+ conn->ksnc_tx_bufnob = bufnob;
+ mb();
}
- if (tx->tx_resid == 0) { /* sent everything */
- rc = 0;
+ 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 (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
+ !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;
+
+ } while (tx->tx_resid != 0);
ksocknal_putconnsock (conn);
- RETURN (rc);
+ 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)
{
+#if SOCKNAL_SINGLE_FRAG_RX
+ struct iovec scratch;
+ struct iovec *scratchiov = &scratch;
+ int niov = 1;
+#else
+ struct iovec *scratchiov = conn->ksnc_rx_scratch_iov;
+ int niov = conn->ksnc_rx_niov;
+#endif
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_iov = scratchiov,
+ .msg_iovlen = niov,
.msg_control = NULL,
.msg_controllen = 0,
.msg_flags = 0
};
mm_segment_t oldmm = get_fs();
+ int nob;
+ int i;
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 (conn->ksnc_rx_niov > 0);
- LASSERT (fragsize <= conn->ksnc_rx_nob_wanted);
-
+ * or leave them alone. */
+ LASSERT (niov > 0);
+
+ for (nob = i = 0; i < niov; i++) {
+ scratchiov[i] = iov[i];
+ nob += scratchiov[i].iov_len;
+ }
+ LASSERT (nob <= 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............................^ */
+ rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
+ /* NB this is just a boolean..........................^ */
set_fs (oldmm);
if (rc <= 0)
return (rc);
- conn->ksnc_rx_nob_wanted -= rc;
- conn->ksnc_rx_nob_left -= rc;
+ /* received something... */
+ nob = rc;
+
+ 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;
+
+ conn->ksnc_rx_nob_wanted -= nob;
+ conn->ksnc_rx_nob_left -= nob;
+
+ do {
+ LASSERT (conn->ksnc_rx_niov > 0);
- if (rc < fragsize) {
- iov->iov_base = (void *)(vaddr + rc);
- iov->iov_len = fragsize - rc;
- return (-EAGAIN);
- }
+ 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);
- conn->ksnc_rx_iov++;
- conn->ksnc_rx_niov--;
- return (1);
+ return (rc);
}
int
ksocknal_recv_kiov (ksock_conn_t *conn)
{
+#if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
+ struct iovec scratch;
+ struct iovec *scratchiov = &scratch;
+ int niov = 1;
+#else
+#warning "XXX risk of kmap deadlock on multiple frags..."
+ struct iovec *scratchiov = conn->ksnc_rx_scratch_iov;
+ int niov = conn->ksnc_rx_nkiov;
+#endif
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_iov = scratchiov,
+ .msg_iovlen = niov,
.msg_control = NULL,
.msg_controllen = 0,
.msg_flags = 0
};
mm_segment_t oldmm = get_fs();
+ int nob;
+ int i;
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);
LASSERT (conn->ksnc_rx_nkiov > 0);
- LASSERT (offset + fragsize <= PAGE_SIZE);
-
+
+ /* NB we can't trust socket ops to either consume our iovs
+ * or leave them alone. */
+ for (nob = i = 0; i < niov; i++) {
+ scratchiov[i].iov_base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
+ nob += scratchiov[i].iov_len = kiov[i].kiov_len;
+ }
+ LASSERT (nob <= 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............................^ */
+ rc = sock_recvmsg (conn->ksnc_sock, &msg, nob, MSG_DONTWAIT);
+ /* NB this is just a boolean.......................^ */
set_fs (oldmm);
- kunmap (page);
-
+
+ for (i = 0; i < niov; i++)
+ kunmap(kiov[i].kiov_page);
+
if (rc <= 0)
return (rc);
- conn->ksnc_rx_nob_wanted -= rc;
- conn->ksnc_rx_nob_left -= rc;
+ /* received something... */
+ nob = rc;
+
+ 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;
+
+ conn->ksnc_rx_nob_wanted -= nob;
+ conn->ksnc_rx_nob_left -= nob;
+
+ do {
+ LASSERT (conn->ksnc_rx_nkiov > 0);
- if (rc < fragsize) {
- kiov->kiov_offset = offset + rc;
- kiov->kiov_len = fragsize - rc;
- return (-EAGAIN);
- }
+ 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);
- conn->ksnc_rx_kiov++;
- conn->ksnc_rx_nkiov--;
- return (1);
+ return 1;
}
int
-ksocknal_recvmsg (ksock_conn_t *conn)
+ksocknal_receive (ksock_conn_t *conn)
{
/* Return 1 on success, 0 on EOF, < 0 on error.
* Caller checks ksnc_rx_nob_wanted to determine
}
rc = ksocknal_getconnsock (conn);
- if (rc != 0)
- return (rc);
+ if (rc != 0) {
+ LASSERT (conn->ksnc_closing);
+ return (-ESHUTDOWN);
+ }
for (;;) {
- if (conn->ksnc_closing) {
- rc = -ESHUTDOWN;
- break;
- }
-
if (conn->ksnc_rx_niov != 0)
rc = ksocknal_recv_iov (conn);
else
rc = ksocknal_recv_kiov (conn);
-
+
if (rc <= 0) {
/* error/EOF or partial receive */
- if (rc == -EAGAIN)
+ if (rc == -EAGAIN) {
rc = 1;
+ } else if (rc == 0 && conn->ksnc_rx_started) {
+ /* EOF in the middle of a message */
+ rc = -EPROTO;
+ }
break;
}
+ /* Completed a fragment */
+
if (conn->ksnc_rx_nob_wanted == 0) {
+ /* Completed a message segment (header or payload) */
+ if ((ksocknal_tunables.ksnd_eager_ack & conn->ksnc_type) != 0 &&
+ (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);
+ }
rc = 1;
break;
}
spin_lock_irqsave (&sched->kss_lock, flags);
- list_del (&tx->tx_list); /* remove from kss_zctxpending_list */
list_add_tail (&tx->tx_list, &sched->kss_zctxdone_list);
- if (waitqueue_active (&sched->kss_waitq))
- wake_up (&sched->kss_waitq);
+ wake_up (&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 (tx->tx_isfwd) { /* was a forwarded packet? */
kpr_fwd_done (&ksocknal_data.ksnd_router,
- KSOCK_TX_2_KPR_FWD_DESC (tx), 0);
+ KSOCK_TX_2_KPR_FWD_DESC (tx),
+ (tx->tx_resid == 0) ? 0 : -ECONNABORTED);
EXIT;
return;
}
/* local send */
ltx = KSOCK_TX_2_KSOCK_LTX (tx);
- lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie);
+ lib_finalize (&ksocknal_lib, ltx->ltx_private, ltx->ltx_cookie,
+ (tx->tx_resid == 0) ? PTL_OK : PTL_FAIL);
- ksocknal_put_ltx (ltx);
+ ksocknal_free_ltx (ltx);
EXIT;
}
{
#if SOCKNAL_ZC
if (atomic_read (&tx->tx_zccd.zccd_count) != 1) {
- unsigned long flags;
ksock_conn_t *conn = tx->tx_conn;
- ksock_sched_t *sched = conn->ksnc_scheduler;
/* 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);
- /* Stash it for timeout...
- * NB We have to hold a lock to stash the tx, and we have
- * stash it before we zcc_put(), but we have to _not_ hold
- * this lock when we zcc_put(), otherwise we could deadlock
- * if it turns out to be the last put. Aaaaarrrrggghhh! */
- spin_lock_irqsave (&sched->kss_lock, flags);
- list_add_tail (&tx->tx_list, &conn->ksnc_tx_pending);
- spin_unlock_irqrestore (&sched->kss_lock, flags);
-
/* ...then drop the initial ref on zccd, so the zero copy
* callback can occur */
zccd_put (&tx->tx_zccd);
ksocknal_tx_done (tx, 0);
}
-void
-ksocknal_process_transmit (ksock_sched_t *sched, unsigned long *irq_flags)
+int
+ksocknal_process_transmit (ksock_conn_t *conn, ksock_tx_t *tx)
{
- ksock_conn_t *conn;
- ksock_tx_t *tx;
- int rc;
-
- LASSERT (!list_empty (&sched->kss_tx_conns));
- conn = list_entry(sched->kss_tx_conns.next, ksock_conn_t, ksnc_tx_list);
- list_del (&conn->ksnc_tx_list);
+ unsigned long flags;
+ int rc;
+
+ rc = ksocknal_transmit (conn, tx);
- 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);
- /* assume transmit will complete now, so dequeue while I've got lock */
- list_del (&tx->tx_list);
+ CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
- spin_unlock_irqrestore (&sched->kss_lock, *irq_flags);
+ if (tx->tx_resid == 0) {
+ /* Sent everything OK */
+ LASSERT (rc == 0);
- LASSERT (tx->tx_resid > 0);
+ ksocknal_tx_launched (tx);
+ return (0);
+ }
- conn->ksnc_tx_ready = 0;/* write_space may race with me and set ready */
- mb(); /* => clear BEFORE trying to write */
+ if (rc == -EAGAIN)
+ return (rc);
- rc = ksocknal_sendmsg (conn, tx);
+ if (rc == -ENOMEM) {
+ static int counter;
- CDEBUG (D_NET, "send(%d) %d\n", tx->tx_resid, rc);
+ counter++; /* exponential backoff warnings */
+ if ((counter & (-counter)) == counter)
+ CWARN("%d ENOMEM tx %p\n", counter, conn);
- if (rc != 0) {
- if (ksocknal_close_conn_unlocked (conn)) {
- /* I'm the first to close */
- CERROR ("[%p] Error %d on write to "LPX64" ip %08x:%d\n",
- conn, rc, conn->ksnc_peer->ksnp_nid,
- conn->ksnc_ipaddr, conn->ksnc_port);
- }
- ksocknal_tx_launched (tx);
- spin_lock_irqsave (&sched->kss_lock, *irq_flags);
+ /* Queue on ksnd_enomem_conns for retry after a timeout */
+ spin_lock_irqsave(&ksocknal_data.ksnd_reaper_lock, flags);
- } else if (tx->tx_resid == 0) {
+ /* enomem list takes over scheduler's ref... */
+ 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,
+ ksocknal_data.ksnd_reaper_waketime))
+ wake_up (&ksocknal_data.ksnd_reaper_waitq);
+
+ spin_unlock_irqrestore(&ksocknal_data.ksnd_reaper_lock, flags);
+ return (rc);
+ }
- /* everything went; assume more can go, and avoid
- * write_space locking */
- conn->ksnc_tx_ready = 1;
+ /* Actual error */
+ LASSERT (rc < 0);
- ksocknal_tx_launched (tx);
- spin_lock_irqsave (&sched->kss_lock, *irq_flags);
- } else {
- spin_lock_irqsave (&sched->kss_lock, *irq_flags);
+ if (!conn->ksnc_closing)
+ CERROR("[%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);
- /* back onto HEAD of tx_queue */
- list_add (&tx->tx_list, &conn->ksnc_tx_queue);
- }
+ ksocknal_close_conn_and_siblings (conn, rc);
+ ksocknal_tx_launched (tx);
- /* no space to write, or nothing to write? */
- if (!conn->ksnc_tx_ready ||
- list_empty (&conn->ksnc_tx_queue)) {
- /* mark not scheduled */
- conn->ksnc_tx_scheduled = 0;
- /* drop scheduler's ref */
- ksocknal_put_conn (conn);
- } else {
- /* stay scheduled */
- list_add_tail (&conn->ksnc_tx_list, &sched->kss_tx_conns);
- }
+ return (rc);
}
void
/* called holding write lock on ksnd_global_lock */
- LASSERT (route->ksnr_conn == NULL);
- LASSERT (!route->ksnr_deleted && !route->ksnr_connecting);
+ 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 == 0);
- route->ksnr_connecting = 1;
+ 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 */
spin_lock_irqsave (&ksocknal_data.ksnd_autoconnectd_lock, flags);
list_add_tail (&route->ksnr_connect_list,
&ksocknal_data.ksnd_autoconnectd_routes);
-
- if (waitqueue_active (&ksocknal_data.ksnd_autoconnectd_waitq))
- wake_up (&ksocknal_data.ksnd_autoconnectd_waitq);
+ wake_up (&ksocknal_data.ksnd_autoconnectd_waitq);
spin_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
}
ksock_peer_t *
ksocknal_find_target_peer_locked (ksock_tx_t *tx, ptl_nid_t nid)
{
+ char ipbuf[PTL_NALFMT_SIZE];
ptl_nid_t target_nid;
int rc;
ksock_peer_t *peer = ksocknal_find_peer_locked (nid);
-
+
if (peer != NULL)
return (peer);
-
+
if (tx->tx_isfwd) {
CERROR ("Can't send packet to "LPX64
- ": routed target is not a peer\n", nid);
+ " %s: routed target is not a peer\n",
+ nid, portals_nid2str(SOCKNAL, nid, ipbuf));
return (NULL);
}
-
- rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, &target_nid);
+
+ rc = kpr_lookup (&ksocknal_data.ksnd_router, nid, tx->tx_nob,
+ &target_nid);
if (rc != 0) {
- CERROR ("Can't route to "LPX64": router error %d\n", nid, rc);
+ CERROR ("Can't route to "LPX64" %s: router error %d\n",
+ nid, portals_nid2str(SOCKNAL, nid, ipbuf), rc);
return (NULL);
}
if (peer != NULL)
return (peer);
- CERROR ("Can't send packet to "LPX64": no peer entry\n", target_nid);
+ CERROR ("Can't send packet to "LPX64" %s: no peer entry\n",
+ target_nid, portals_nid2str(SOCKNAL, target_nid, ipbuf));
return (NULL);
}
ksock_conn_t *
-ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
+ksocknal_find_conn_locked (ksock_tx_t *tx, ksock_peer_t *peer)
{
struct list_head *tmp;
- ksock_conn_t *conn = NULL;
+ ksock_conn_t *typed = NULL;
+ 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);
-
+ 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;
+#endif
LASSERT (!c->ksnc_closing);
-
- if (conn == NULL ||
- atomic_read (&conn->ksnc_tx_nob) >
- atomic_read (&c->ksnc_tx_nob))
- conn = c;
+
+ if (fallback == NULL || nob < fnob) {
+ fallback = c;
+ fnob = nob;
+ }
+
+ if (!ksocknal_tunables.ksnd_typed_conns)
+ continue;
+
+ switch (c->ksnc_type) {
+ default:
+ LBUG();
+ case SOCKNAL_CONN_ANY:
+ break;
+ case SOCKNAL_CONN_BULK_IN:
+ continue;
+ case SOCKNAL_CONN_BULK_OUT:
+ if (tx->tx_nob < ksocknal_tunables.ksnd_min_bulk)
+ continue;
+ break;
+ case SOCKNAL_CONN_CONTROL:
+ if (tx->tx_nob >= ksocknal_tunables.ksnd_min_bulk)
+ continue;
+ break;
+ }
+
+ if (typed == NULL || nob < tnob) {
+ typed = c;
+ tnob = nob;
+ }
}
- return (conn);
+ /* prefer the typed selection */
+ 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
unsigned long flags;
ksock_sched_t *sched = conn->ksnc_scheduler;
- /* called holding global lock (read or irq-write) */
-
- CDEBUG (D_NET, "Sending to "LPX64" on port %d\n",
- conn->ksnc_peer->ksnp_nid, conn->ksnc_port);
+ /* called holding global lock (read or irq-write) and caller may
+ * not have dropped this lock between finding conn and calling me,
+ * so we don't need the {get,put}connsock dance to deref
+ * 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),
+ conn->ksnc_port);
atomic_add (tx->tx_nob, &conn->ksnc_tx_nob);
- tx->tx_resid = tx->tx_nob;
tx->tx_conn = conn;
#if SOCKNAL_ZC
/* NB this sets 1 ref on zccd, so the callback can only occur after
* I've released this ref. */
#endif
-
spin_lock_irqsave (&sched->kss_lock, flags);
-
- tx->tx_deadline = jiffies_64 + ksocknal_io_timeout;
- list_add_tail (&tx->tx_list, &conn->ksnc_tx_queue);
+
+ if (list_empty(&conn->ksnc_tx_queue) &&
+ conn->ksnc_sock->sk->sk_wmem_queued == 0) {
+ /* First packet starts the timeout */
+ conn->ksnc_tx_deadline = jiffies +
+ ksocknal_tunables.ksnd_io_timeout * HZ;
+ 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 */
list_add_tail (&conn->ksnc_tx_list,
&sched->kss_tx_conns);
conn->ksnc_tx_scheduled = 1;
- if (waitqueue_active (&sched->kss_waitq))
- wake_up (&sched->kss_waitq);
+ wake_up (&sched->kss_waitq);
}
spin_unlock_irqrestore (&sched->kss_lock, flags);
{
struct list_head *tmp;
ksock_route_t *route;
+ int bits;
list_for_each (tmp, &peer->ksnp_routes) {
route = list_entry (tmp, ksock_route_t, ksnr_list);
+ bits = route->ksnr_connected;
+
+ /* All typed connections established? */
+ if ((bits & KSNR_TYPED_ROUTES) == KSNR_TYPED_ROUTES)
+ continue;
+
+ /* Untyped connection established? */
+ if ((bits & (1 << SOCKNAL_CONN_ANY)) != 0)
+ continue;
+
+ /* connection being established? */
+ if (route->ksnr_connecting != 0)
+ continue;
+
+ /* too soon to retry this guy? */
+ if (!time_after_eq (jiffies, route->ksnr_timeout))
+ continue;
- if (route->ksnr_conn == NULL && /* not connected */
- !route->ksnr_connecting && /* not connecting */
- route->ksnr_timeout <= jiffies_64) /* OK to retry */
- return (route);
+ return (route);
}
return (NULL);
list_for_each (tmp, &peer->ksnp_routes) {
route = list_entry (tmp, ksock_route_t, ksnr_list);
- if (route->ksnr_connecting)
+ if (route->ksnr_connecting != 0)
return (route);
}
tx->tx_nob, tx->tx_niov, tx->tx_nkiov);
tx->tx_conn = NULL; /* only set when assigned a conn */
+ tx->tx_resid = tx->tx_nob;
+ 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 (PTL_FAIL);
+ return (-EHOSTUNREACH);
}
- /* Any routes need to be connected? (need write lock if so) */
- if (ksocknal_find_connectable_route_locked (peer) == NULL) {
+ if (ksocknal_find_connectable_route_locked(peer) == NULL) {
conn = ksocknal_find_conn_locked (tx, peer);
if (conn != NULL) {
+ /* I've got no autoconnect routes that need to be
+ * connecting and I do have an actual connection... */
ksocknal_queue_tx_locked (tx, conn);
read_unlock (g_lock);
- return (PTL_OK);
+ return (0);
}
}
-
- /* need a write lock now to change peer state... */
-
- 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);
- return (PTL_FAIL);
+#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 */
- /* I may launch autoconnects, now we're write locked... */
- while ((route = ksocknal_find_connectable_route_locked (peer)) != NULL)
+ for (;;) {
+ /* launch any/all autoconnections that need it */
+ route = ksocknal_find_connectable_route_locked (peer);
+ if (route == NULL)
+ break;
+
ksocknal_launch_autoconnect_locked (route);
+ }
conn = ksocknal_find_conn_locked (tx, peer);
if (conn != NULL) {
+ /* Connection exists; queue message on it */
ksocknal_queue_tx_locked (tx, conn);
write_unlock_irqrestore (g_lock, flags);
- return (PTL_OK);
+ return (0);
}
-
- if (ksocknal_find_connecting_route_locked (peer) == NULL) {
- /* no routes actually connecting now */
+
+ route = ksocknal_find_connecting_route_locked (peer);
+ if (route != NULL) {
+ /* At least 1 connection is being established; queue the
+ * message... */
+ list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
write_unlock_irqrestore (g_lock, flags);
- return (PTL_FAIL);
+ return (0);
}
-
- list_add_tail (&tx->tx_list, &peer->ksnp_tx_queue);
-
+
write_unlock_irqrestore (g_lock, flags);
- return (PTL_OK);
+ return (-EHOSTUNREACH);
}
-ksock_ltx_t *
-ksocknal_setup_hdr (nal_cb_t *nal, void *private, lib_msg_t *cookie,
- ptl_hdr_t *hdr, int type)
+ptl_err_t
+ksocknal_sendmsg(lib_nal_t *nal,
+ void *private,
+ lib_msg_t *cookie,
+ ptl_hdr_t *hdr,
+ int type,
+ ptl_nid_t nid,
+ ptl_pid_t pid,
+ unsigned int payload_niov,
+ struct iovec *payload_iov,
+ ptl_kiov_t *payload_kiov,
+ size_t payload_offset,
+ size_t payload_nob)
{
ksock_ltx_t *ltx;
+ int desc_size;
+ int rc;
+
+ /* NB 'private' is different depending on what we're sending.
+ * Just ignore it... */
+
+ CDEBUG(D_NET, "sending "LPSZ" bytes in %d frags to nid:"LPX64
+ " pid %d\n", payload_nob, payload_niov, nid , pid);
- /* I may not block for a transmit descriptor if I might block the
- * receiver, or an interrupt handler. */
- ltx = ksocknal_get_ltx (!(type == PTL_MSG_ACK ||
- type == PTL_MSG_REPLY ||
- in_interrupt ()));
+ LASSERT (payload_nob == 0 || payload_niov > 0);
+ LASSERT (payload_niov <= PTL_MD_MAX_IOV);
+
+ /* It must be OK to kmap() if required */
+ 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) {
+ /* Can't block if in interrupt or responding to an incoming
+ * message */
+ PORTAL_ALLOC_ATOMIC(ltx, desc_size);
+ } else {
+ PORTAL_ALLOC(ltx, desc_size);
+ }
+
if (ltx == NULL) {
- CERROR ("Can't allocate tx desc\n");
- return (NULL);
+ CERROR("Can't allocate tx desc type %d size %d %s\n",
+ type, desc_size, in_interrupt() ? "(intr)" : "");
+ return (PTL_NO_SPACE);
}
- /* Init local send packet (storage for hdr, finalize() args) */
+ 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;
- /* Init common ltx_tx */
ltx->ltx_tx.tx_isfwd = 0;
- ltx->ltx_tx.tx_nob = sizeof (*hdr);
+ ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_nob;
- /* We always have 1 mapped frag for the header */
- ltx->ltx_tx.tx_niov = 1;
- ltx->ltx_tx.tx_iov = <x->ltx_iov_space.hdr;
- ltx->ltx_tx.tx_iov[0].iov_base = <x->ltx_hdr;
- ltx->ltx_tx.tx_iov[0].iov_len = sizeof (ltx->ltx_hdr);
+ if (payload_iov != NULL) {
+ /* payload is all mapped */
+ ltx->ltx_tx.tx_kiov = NULL;
+ ltx->ltx_tx.tx_nkiov = 0;
- ltx->ltx_tx.tx_kiov = NULL;
- ltx->ltx_tx.tx_nkiov = 0;
+ ltx->ltx_tx.tx_niov =
+ 1 + lib_extract_iov(payload_niov, <x->ltx_iov[1],
+ payload_niov, payload_iov,
+ payload_offset, payload_nob);
+ } else {
+ /* payload is all pages */
+ ltx->ltx_tx.tx_niov = 1;
+
+ ltx->ltx_tx.tx_kiov = ltx->ltx_kiov;
+ ltx->ltx_tx.tx_nkiov =
+ lib_extract_kiov(payload_niov, ltx->ltx_kiov,
+ payload_niov, payload_kiov,
+ payload_offset, payload_nob);
+ }
- return (ltx);
+ rc = ksocknal_launch_packet(<x->ltx_tx, nid);
+ if (rc == 0)
+ return (PTL_OK);
+
+ ksocknal_free_ltx(ltx);
+ return (PTL_FAIL);
}
-int
-ksocknal_send (nal_cb_t *nal, void *private, lib_msg_t *cookie,
+ptl_err_t
+ksocknal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie,
ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
unsigned int payload_niov, struct iovec *payload_iov,
- size_t payload_len)
+ size_t payload_offset, size_t payload_len)
{
- ksock_ltx_t *ltx;
- int rc;
-
- /* NB 'private' is different depending on what we're sending.
- * Just ignore it until we can rely on it
- */
-
- CDEBUG(D_NET,
- "sending "LPSZ" bytes in %d mapped frags to nid: "LPX64
- " pid %d\n", payload_len, payload_niov, nid, pid);
-
- ltx = ksocknal_setup_hdr (nal, private, cookie, hdr, type);
- if (ltx == NULL)
- return (PTL_FAIL);
-
- /* append the payload_iovs to the one pointing at the header */
- LASSERT (ltx->ltx_tx.tx_niov == 1 && ltx->ltx_tx.tx_nkiov == 0);
- LASSERT (payload_niov <= PTL_MD_MAX_IOV);
-
- memcpy (ltx->ltx_tx.tx_iov + 1, payload_iov,
- payload_niov * sizeof (*payload_iov));
- ltx->ltx_tx.tx_niov = 1 + payload_niov;
- ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_len;
-
- rc = ksocknal_launch_packet (<x->ltx_tx, nid);
- if (rc != PTL_OK)
- ksocknal_put_ltx (ltx);
-
- return (rc);
+ return (ksocknal_sendmsg(nal, private, cookie,
+ hdr, type, nid, pid,
+ payload_niov, payload_iov, NULL,
+ payload_offset, payload_len));
}
-int
-ksocknal_send_pages (nal_cb_t *nal, void *private, lib_msg_t *cookie,
+ptl_err_t
+ksocknal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie,
ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
- unsigned int payload_niov, ptl_kiov_t *payload_iov, size_t payload_len)
+ unsigned int payload_niov, ptl_kiov_t *payload_kiov,
+ size_t payload_offset, size_t payload_len)
{
- ksock_ltx_t *ltx;
- int rc;
-
- /* NB 'private' is different depending on what we're sending.
- * Just ignore it until we can rely on it */
-
- CDEBUG(D_NET,
- "sending "LPSZ" bytes in %d mapped frags to nid: "LPX64" pid %d\n",
- payload_len, payload_niov, nid, pid);
-
- ltx = ksocknal_setup_hdr (nal, private, cookie, hdr, type);
- if (ltx == NULL)
- return (PTL_FAIL);
-
- LASSERT (ltx->ltx_tx.tx_niov == 1 && ltx->ltx_tx.tx_nkiov == 0);
- LASSERT (payload_niov <= PTL_MD_MAX_IOV);
-
- ltx->ltx_tx.tx_kiov = ltx->ltx_iov_space.payload.kiov;
- memcpy (ltx->ltx_tx.tx_kiov, payload_iov,
- payload_niov * sizeof (*payload_iov));
- ltx->ltx_tx.tx_nkiov = payload_niov;
- ltx->ltx_tx.tx_nob = sizeof (*hdr) + payload_len;
-
- rc = ksocknal_launch_packet (<x->ltx_tx, nid);
- if (rc != PTL_OK)
- ksocknal_put_ltx (ltx);
-
- return (rc);
+ return (ksocknal_sendmsg(nal, private, cookie,
+ hdr, type, nid, pid,
+ payload_niov, NULL, payload_kiov,
+ payload_offset, payload_len));
}
void
ksocknal_fwd_packet (void *arg, kpr_fwd_desc_t *fwd)
{
ptl_nid_t nid = fwd->kprfd_gateway_nid;
- ksock_tx_t *tx = (ksock_tx_t *)&fwd->kprfd_scratch;
+ 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);
/* I'm the gateway; must be the last hop */
- if (nid == ksocknal_lib.ni.nid)
+ if (nid == ksocknal_lib.libnal_ni.ni_pid.nid)
nid = fwd->kprfd_target_nid;
- tx->tx_isfwd = 1; /* This is a forwarding packet */
- tx->tx_nob = fwd->kprfd_nob;
- tx->tx_niov = fwd->kprfd_niov;
- tx->tx_iov = fwd->kprfd_iov;
- tx->tx_nkiov = 0;
- tx->tx_kiov = NULL;
- tx->tx_hdr = (ptl_hdr_t *)fwd->kprfd_iov[0].iov_base;
+ /* setup iov for hdr */
+ ftx->ftx_iov.iov_base = fwd->kprfd_hdr;
+ ftx->ftx_iov.iov_len = sizeof(ptl_hdr_t);
- rc = ksocknal_launch_packet (tx, nid);
- if (rc != 0) {
- /* FIXME, could pass a better completion error */
- kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, -EHOSTUNREACH);
- }
+ ftx->ftx_tx.tx_isfwd = 1; /* This is a forwarding packet */
+ ftx->ftx_tx.tx_nob = sizeof(ptl_hdr_t) + fwd->kprfd_nob;
+ ftx->ftx_tx.tx_niov = 1;
+ ftx->ftx_tx.tx_iov = &ftx->ftx_iov;
+ ftx->ftx_tx.tx_nkiov = fwd->kprfd_niov;
+ ftx->ftx_tx.tx_kiov = fwd->kprfd_kiov;
+
+ rc = ksocknal_launch_packet (&ftx->ftx_tx, nid);
+ if (rc != 0)
+ kpr_fwd_done (&ksocknal_data.ksnd_router, fwd, rc);
}
int
ksocknal_thread_start (int (*fn)(void *arg), void *arg)
{
- long pid = kernel_thread (fn, arg, 0);
+ long pid = 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_pages[0]);
+ ptl_hdr_t *hdr = &fmb->fmb_hdr;
ksock_conn_t *conn = NULL;
ksock_sched_t *sched;
unsigned long flags;
+ char ipbuf[PTL_NALFMT_SIZE];
+ char ipbuf2[PTL_NALFMT_SIZE];
if (error != 0)
- CERROR("Failed to route packet from "LPX64" to "LPX64": %d\n",
- NTOH__u64(hdr->src_nid), NTOH__u64(hdr->dest_nid),
+ CERROR("Failed to route packet from "
+ LPX64" %s to "LPX64" %s: %d\n",
+ 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 (&fmp->fmp_lock, flags);
list_add (&fmb->fmb_list, &fmp->fmp_idle_fmbs);
+ fmp->fmp_nactive_fmbs--;
if (!list_empty (&fmp->fmp_blocked_conns)) {
conn = list_entry (fmb->fmb_pool->fmp_blocked_conns.next,
spin_unlock_irqrestore (&fmp->fmp_lock, flags);
- /* drop peer ref taken on init */
- ksocknal_put_peer (fmb->fmb_peer);
-
if (conn == NULL)
return;
spin_lock_irqsave (&sched->kss_lock, flags);
list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
-
- if (waitqueue_active (&sched->kss_waitq))
- wake_up (&sched->kss_waitq);
+ wake_up (&sched->kss_waitq);
spin_unlock_irqrestore (&sched->kss_lock, flags);
}
ksocknal_get_idle_fmb (ksock_conn_t *conn)
{
int payload_nob = conn->ksnc_rx_nob_left;
- int packet_nob = sizeof (ptl_hdr_t) + payload_nob;
unsigned long flags;
ksock_fmb_pool_t *pool;
ksock_fmb_t *fmb;
LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_GET_FMB);
- LASSERT (ksocknal_data.ksnd_fmbs != NULL);
+ LASSERT (kpr_routing(&ksocknal_data.ksnd_router));
- if (packet_nob <= SOCKNAL_SMALL_FWD_PAGES * PAGE_SIZE)
+ if (payload_nob <= SOCKNAL_SMALL_FWD_PAGES * PAGE_SIZE)
pool = &ksocknal_data.ksnd_small_fmp;
else
pool = &ksocknal_data.ksnd_large_fmp;
fmb = list_entry(pool->fmp_idle_fmbs.next,
ksock_fmb_t, fmb_list);
list_del (&fmb->fmb_list);
+ pool->fmp_nactive_fmbs++;
spin_unlock_irqrestore (&pool->fmp_lock, flags);
return (fmb);
int
ksocknal_init_fmb (ksock_conn_t *conn, ksock_fmb_t *fmb)
{
- int payload_nob = conn->ksnc_rx_nob_left;
- int packet_nob = sizeof (ptl_hdr_t) + payload_nob;
- ptl_nid_t dest_nid = NTOH__u64 (conn->ksnc_hdr.dest_nid);
- int niov; /* at least the header */
- int nob;
+ int payload_nob = conn->ksnc_rx_nob_left;
+ ptl_nid_t dest_nid = le64_to_cpu(conn->ksnc_hdr.dest_nid);
+ int niov = 0;
+ int nob = payload_nob;
LASSERT (conn->ksnc_rx_scheduled);
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 (packet_nob <= fmb->fmb_npages * PAGE_SIZE);
+ LASSERT (payload_nob <= fmb->fmb_pool->fmp_buff_pages * PAGE_SIZE);
LASSERT (sizeof (ptl_hdr_t) < PAGE_SIZE);
-
- /* Got a forwarding buffer; copy the header we just read into the
- * forwarding buffer. If there's payload, start reading reading it
- * into the buffer, otherwise the forwarding buffer can be kicked
- * off immediately.
- *
- * NB fmb->fmb_iov spans the WHOLE packet.
- * conn->ksnc_rx_iov spans just the payload.
- */
- fmb->fmb_iov[0].iov_base = page_address (fmb->fmb_pages[0]);
-
- /* copy header */
- memcpy (fmb->fmb_iov[0].iov_base, &conn->ksnc_hdr, sizeof (ptl_hdr_t));
+ LASSERT (fmb->fmb_kiov[0].kiov_offset == 0);
/* Take a ref on the conn's peer to prevent module unload before
- * forwarding completes. NB we ref peer and not conn since because
- * all refs on conn after it has been closed must remove themselves
- * in finite time */
+ * forwarding completes. */
fmb->fmb_peer = conn->ksnc_peer;
atomic_inc (&conn->ksnc_peer->ksnp_refcount);
- if (payload_nob == 0) { /* got complete packet already */
- CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d fwd_start (immediate)\n",
- conn, NTOH__u64 (conn->ksnc_hdr.src_nid),
- dest_nid, packet_nob);
+ /* Copy the header we just read into the forwarding buffer. If
+ * there's payload, start reading reading it into the buffer,
+ * otherwise the forwarding buffer can be kicked off
+ * immediately. */
+ fmb->fmb_hdr = conn->ksnc_hdr;
- fmb->fmb_iov[0].iov_len = sizeof (ptl_hdr_t);
+ 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;
+ niov++;
+ }
- kpr_fwd_init (&fmb->fmb_fwd, dest_nid,
- packet_nob, 1, fmb->fmb_iov,
- ksocknal_fmb_callback, fmb);
+ kpr_fwd_init(&fmb->fmb_fwd, dest_nid, &fmb->fmb_hdr,
+ payload_nob, niov, fmb->fmb_kiov,
+ ksocknal_fmb_callback, fmb);
+
+ if (payload_nob == 0) { /* got complete packet already */
+ CDEBUG (D_NET, "%p "LPX64"->"LPX64" fwd_start (immediate)\n",
+ conn, le64_to_cpu(conn->ksnc_hdr.src_nid), dest_nid);
- /* forward it now */
kpr_fwd_start (&ksocknal_data.ksnd_router, &fmb->fmb_fwd);
ksocknal_new_packet (conn, 0); /* on to next packet */
return (1);
}
- niov = 1;
- if (packet_nob <= PAGE_SIZE) { /* whole packet fits in first page */
- fmb->fmb_iov[0].iov_len = packet_nob;
- } else {
- fmb->fmb_iov[0].iov_len = PAGE_SIZE;
- nob = packet_nob - PAGE_SIZE;
-
- do {
- LASSERT (niov < fmb->fmb_npages);
- fmb->fmb_iov[niov].iov_base =
- page_address (fmb->fmb_pages[niov]);
- fmb->fmb_iov[niov].iov_len = MIN (PAGE_SIZE, nob);
- nob -= PAGE_SIZE;
- niov++;
- } while (nob > 0);
- }
-
- kpr_fwd_init (&fmb->fmb_fwd, dest_nid,
- packet_nob, niov, fmb->fmb_iov,
- ksocknal_fmb_callback, fmb);
-
conn->ksnc_cookie = fmb; /* stash fmb for later */
conn->ksnc_rx_state = SOCKNAL_RX_BODY_FWD; /* read in the payload */
- conn->ksnc_rx_deadline = jiffies_64 + ksocknal_io_timeout; /* start timeout */
- /* payload is desc's iov-ed buffer, but skipping the hdr */
- LASSERT (niov <= sizeof (conn->ksnc_rx_iov_space) /
- sizeof (struct iovec));
-
- conn->ksnc_rx_iov = (struct iovec *)&conn->ksnc_rx_iov_space;
- conn->ksnc_rx_iov[0].iov_base =
- (void *)(((unsigned long)fmb->fmb_iov[0].iov_base) +
- sizeof (ptl_hdr_t));
- conn->ksnc_rx_iov[0].iov_len =
- fmb->fmb_iov[0].iov_len - sizeof (ptl_hdr_t);
-
- if (niov > 1)
- memcpy(&conn->ksnc_rx_iov[1], &fmb->fmb_iov[1],
- (niov - 1) * sizeof (struct iovec));
-
- conn->ksnc_rx_niov = niov;
+ /* 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_niov = 0;
+ 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);
- int body_len = NTOH__u32 (PTL_HDR_LENGTH(&conn->ksnc_hdr));
+ 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];
CDEBUG (D_NET, "%p "LPX64"->"LPX64" %d parsing header\n", conn,
- NTOH__u64 (conn->ksnc_hdr.src_nid),
- dest_nid, conn->ksnc_rx_nob_left);
+ src_nid, dest_nid, conn->ksnc_rx_nob_left);
LASSERT (conn->ksnc_rx_state == SOCKNAL_RX_HEADER);
LASSERT (conn->ksnc_rx_scheduled);
if (body_len < 0) { /* length corrupt (overflow) */
- CERROR("dropping packet from "LPX64" for "LPX64": packet "
- "size %d illegal\n", NTOH__u64 (conn->ksnc_hdr.src_nid),
- dest_nid, body_len);
+ CERROR("dropping packet from "LPX64" (%s) for "LPX64" (%s): "
+ "packet size %d illegal\n",
+ src_nid, portals_nid2str(TCPNAL, src_nid, str),
+ dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
+ body_len);
ksocknal_new_packet (conn, 0); /* on to new packet */
- ksocknal_close_conn_unlocked (conn); /* give up on conn */
return;
}
- if (ksocknal_data.ksnd_fmbs == NULL) { /* not forwarding */
- CERROR("dropping packet from "LPX64" for "LPX64": not "
- "forwarding\n", conn->ksnc_hdr.src_nid,
- conn->ksnc_hdr.dest_nid);
+ if (!kpr_routing(&ksocknal_data.ksnd_router)) { /* not forwarding */
+ CERROR("dropping packet from "LPX64" (%s) for "LPX64
+ " (%s): not forwarding\n",
+ src_nid, portals_nid2str(TCPNAL, src_nid, str),
+ dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
/* on to new packet (skip this one's body) */
ksocknal_new_packet (conn, body_len);
return;
}
- if (body_len > SOCKNAL_MAX_FWD_PAYLOAD) { /* too big to forward */
- CERROR ("dropping packet from "LPX64" for "LPX64
- ": packet size %d too big\n", conn->ksnc_hdr.src_nid,
- conn->ksnc_hdr.dest_nid, body_len);
+ if (body_len > PTL_MTU) { /* too big to forward */
+ CERROR ("dropping packet from "LPX64" (%s) for "LPX64
+ "(%s): packet size %d too big\n",
+ src_nid, portals_nid2str(TCPNAL, src_nid, str),
+ dest_nid, portals_nid2str(TCPNAL, dest_nid, str2),
+ body_len);
/* on to new packet (skip this one's body) */
ksocknal_new_packet (conn, body_len);
return;
/* should have gone direct */
peer = ksocknal_get_peer (conn->ksnc_hdr.dest_nid);
if (peer != NULL) {
- CERROR ("dropping packet from "LPX64" for "LPX64
- ": target is a peer\n", conn->ksnc_hdr.src_nid,
- conn->ksnc_hdr.dest_nid);
+ CERROR ("dropping packet from "LPX64" (%s) for "LPX64
+ "(%s): target is a peer\n",
+ src_nid, portals_nid2str(TCPNAL, src_nid, str),
+ dest_nid, portals_nid2str(TCPNAL, dest_nid, str2));
ksocknal_put_peer (peer); /* drop ref from get above */
/* on to next packet (skip this one's body) */
int skipped;
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);
return (0);
}
-void
-ksocknal_process_receive (ksock_sched_t *sched, unsigned long *irq_flags)
+int
+ksocknal_process_receive (ksock_conn_t *conn)
{
- ksock_conn_t *conn;
ksock_fmb_t *fmb;
int rc;
-
- /* NB: sched->ksnc_lock lock held */
-
- LASSERT (!list_empty (&sched->kss_rx_conns));
- conn = list_entry(sched->kss_rx_conns.next, ksock_conn_t, ksnc_rx_list);
- list_del (&conn->ksnc_rx_list);
-
- spin_unlock_irqrestore (&sched->kss_lock, *irq_flags);
-
- CDEBUG(D_NET, "sched %p conn %p\n", sched, conn);
+
LASSERT (atomic_read (&conn->ksnc_refcount) > 0);
- LASSERT (conn->ksnc_rx_scheduled);
- LASSERT (conn->ksnc_rx_ready);
/* doesn't need a forwarding buffer */
if (conn->ksnc_rx_state != SOCKNAL_RX_GET_FMB)
get_fmb:
fmb = ksocknal_get_idle_fmb (conn);
- if (fmb == NULL) { /* conn descheduled waiting for idle fmb */
- spin_lock_irqsave (&sched->kss_lock, *irq_flags);
- return;
+ if (fmb == NULL) {
+ /* conn descheduled waiting for idle fmb */
+ return (0);
}
- if (ksocknal_init_fmb (conn, fmb)) /* packet forwarded ? */
- goto out; /* come back later for next packet */
+ if (ksocknal_init_fmb (conn, fmb)) {
+ /* packet forwarded */
+ return (0);
+ }
try_read:
/* NB: sched lock NOT held */
LASSERT (conn->ksnc_rx_nob_wanted > 0);
- conn->ksnc_rx_ready = 0;/* data ready may race with me and set ready */
- mb(); /* => clear BEFORE trying to read */
-
- rc = ksocknal_recvmsg(conn);
+ rc = ksocknal_receive(conn);
if (rc <= 0) {
- if (ksocknal_close_conn_unlocked (conn)) {
- /* I'm the first to close */
- if (rc < 0)
- CERROR ("[%p] Error %d on read from "LPX64" ip %08x:%d\n",
- conn, rc, conn->ksnc_peer->ksnp_nid,
- conn->ksnc_ipaddr, conn->ksnc_port);
- else
- CERROR ("[%p] EOF from "LPX64" ip %08x:%d\n",
- conn, conn->ksnc_peer->ksnp_nid,
- conn->ksnc_ipaddr, conn->ksnc_port);
- }
- goto out;
- }
+ LASSERT (rc != -EAGAIN);
- if (conn->ksnc_rx_nob_wanted != 0) /* short read */
- goto out; /* try again later */
+ if (rc == 0)
+ CWARN ("[%p] EOF from "LPX64" ip %d.%d.%d.%d:%d\n",
+ conn, conn->ksnc_peer->ksnp_nid,
+ HIPQUAD(conn->ksnc_ipaddr),
+ conn->ksnc_port);
+ else if (!conn->ksnc_closing)
+ CERROR ("[%p] Error %d on read from "LPX64
+ " ip %d.%d.%d.%d:%d\n",
+ conn, rc, conn->ksnc_peer->ksnp_nid,
+ HIPQUAD(conn->ksnc_ipaddr),
+ conn->ksnc_port);
- /* got all I wanted, assume there's more - prevent data_ready locking */
- conn->ksnc_rx_ready = 1;
+ ksocknal_close_conn_and_siblings (conn, rc);
+ return (rc == 0 ? -ESHUTDOWN : rc);
+ }
+ if (conn->ksnc_rx_nob_wanted != 0) {
+ /* 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) != ksocknal_lib.ni.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);
switch (conn->ksnc_rx_state) {
case SOCKNAL_RX_HEADER: /* skipped (zero payload) */
- goto out; /* => come back later */
+ return (0); /* => come back later */
case SOCKNAL_RX_SLOP: /* skipping packet's body */
goto try_read; /* => go read it */
case SOCKNAL_RX_GET_FMB: /* forwarding */
}
/* sets wanted_len, iovs etc */
- lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
+ rc = lib_parse(&ksocknal_lib, &conn->ksnc_hdr, conn);
- /* start timeout (lib is waiting for finalize) */
- conn->ksnc_rx_deadline = jiffies_64 + ksocknal_io_timeout;
+ if (rc != PTL_OK) {
+ /* I just received garbage: give up on this conn */
+ ksocknal_close_conn_and_siblings (conn, rc);
+ return (-EPROTO);
+ }
if (conn->ksnc_rx_nob_wanted != 0) { /* need to get payload? */
conn->ksnc_rx_state = SOCKNAL_RX_BODY;
case SOCKNAL_RX_BODY:
/* payload all received */
- conn->ksnc_rx_deadline = 0; /* cancel timeout */
- lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie);
+ lib_finalize(&ksocknal_lib, NULL, conn->ksnc_cookie, PTL_OK);
/* Fall through */
case SOCKNAL_RX_SLOP:
/* starting new packet? */
if (ksocknal_new_packet (conn, conn->ksnc_rx_nob_left))
- goto out; /* come back later */
+ return (0); /* come back later */
goto try_read; /* try to finish reading slop now */
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);
- /* cancel timeout (only needed it while fmb allocated) */
- conn->ksnc_rx_deadline = 0;
-
/* forward the packet. NB ksocknal_init_fmb() put fmb into
* conn->ksnc_cookie */
fmb = (ksock_fmb_t *)conn->ksnc_cookie;
LASSERT (conn->ksnc_rx_nob_left == 0);
ksocknal_new_packet (conn, 0); /* on to next packet */
- goto out; /* (later) */
+ return (0); /* (later) */
default:
break;
/* Not Reached */
LBUG ();
-
- out:
- spin_lock_irqsave (&sched->kss_lock, *irq_flags);
-
- /* no data there to read? */
- if (!conn->ksnc_rx_ready) {
- /* let socket callback schedule again */
- conn->ksnc_rx_scheduled = 0;
- /* drop scheduler's ref */
- ksocknal_put_conn (conn);
- } else {
- /* stay scheduled */
- list_add_tail (&conn->ksnc_rx_list, &sched->kss_rx_conns);
- }
+ return (-EINVAL); /* keep gcc happy */
}
-int
-ksocknal_recv (nal_cb_t *nal, void *private, lib_msg_t *msg,
- unsigned int niov, struct iovec *iov, size_t mlen, size_t rlen)
+ptl_err_t
+ksocknal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg,
+ unsigned int niov, struct iovec *iov,
+ size_t offset, size_t mlen, size_t rlen)
{
ksock_conn_t *conn = (ksock_conn_t *)private;
conn->ksnc_rx_nkiov = 0;
conn->ksnc_rx_kiov = NULL;
- conn->ksnc_rx_niov = niov;
conn->ksnc_rx_iov = conn->ksnc_rx_iov_space.iov;
- memcpy (conn->ksnc_rx_iov, iov, niov * sizeof (*iov));
+ conn->ksnc_rx_niov =
+ lib_extract_iov(PTL_MD_MAX_IOV, conn->ksnc_rx_iov,
+ niov, iov, offset, mlen);
LASSERT (mlen ==
lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
- return (rlen);
+ return (PTL_OK);
}
-int
-ksocknal_recv_pages (nal_cb_t *nal, void *private, lib_msg_t *msg,
- unsigned int niov, ptl_kiov_t *kiov, size_t mlen, size_t rlen)
+ptl_err_t
+ksocknal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg,
+ unsigned int niov, ptl_kiov_t *kiov,
+ size_t offset, size_t mlen, size_t rlen)
{
ksock_conn_t *conn = (ksock_conn_t *)private;
conn->ksnc_rx_niov = 0;
conn->ksnc_rx_iov = NULL;
- conn->ksnc_rx_nkiov = niov;
conn->ksnc_rx_kiov = conn->ksnc_rx_iov_space.kiov;
- memcpy (conn->ksnc_rx_kiov, kiov, niov * sizeof (*kiov));
+ conn->ksnc_rx_nkiov =
+ lib_extract_kiov(PTL_MD_MAX_IOV, conn->ksnc_rx_kiov,
+ niov, kiov, offset, mlen);
LASSERT (mlen ==
lib_iov_nob (conn->ksnc_rx_niov, conn->ksnc_rx_iov) +
lib_kiov_nob (conn->ksnc_rx_nkiov, conn->ksnc_rx_kiov));
- return (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;
+ ksock_conn_t *conn;
+ ksock_tx_t *tx;
unsigned long flags;
int rc;
int nloops = 0;
int id = sched - ksocknal_data.ksnd_schedulers;
char name[16];
- snprintf (name, sizeof (name),"ksocknald[%d]", id);
+ snprintf (name, sizeof (name),"ksocknald_%02d", id);
kportal_daemonize (name);
kportal_blockallsigs ();
#if (CONFIG_SMP && CPU_AFFINITY)
- if ((cpu_online_map & (1 << id)) != 0)
- current->cpus_allowed = (1 << id);
- else
- CERROR ("Can't set CPU affinity for %s\n", name);
+ 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 to %d\n", name, id);
+ }
#endif /* CONFIG_SMP && CPU_AFFINITY */
spin_lock_irqsave (&sched->kss_lock, flags);
/* Ensure I progress everything semi-fairly */
if (!list_empty (&sched->kss_rx_conns)) {
+ conn = list_entry(sched->kss_rx_conns.next,
+ ksock_conn_t, ksnc_rx_list);
+ list_del(&conn->ksnc_rx_list);
+
+ LASSERT(conn->ksnc_rx_scheduled);
+ LASSERT(conn->ksnc_rx_ready);
+
+ /* clear rx_ready in case receive isn't complete.
+ * Do it BEFORE we call process_recv, since
+ * data_ready can set it any time after we release
+ * 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 */
+ LASSERT(conn->ksnc_rx_scheduled);
+
+ /* 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.
+ * It will get queued for my attention when
+ * one becomes available (and it might just
+ * already have been!). Meanwhile my ref
+ * on it stays put. */
+ } else if (conn->ksnc_rx_ready) {
+ /* reschedule for rx */
+ list_add_tail (&conn->ksnc_rx_list,
+ &sched->kss_rx_conns);
+ } else {
+ conn->ksnc_rx_scheduled = 0;
+ /* drop my ref */
+ ksocknal_put_conn(conn);
+ }
+
did_something = 1;
- /* drops & regains kss_lock */
- ksocknal_process_receive (sched, &flags);
}
if (!list_empty (&sched->kss_tx_conns)) {
+ 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
+ * kss_lock. */
+ conn->ksnc_tx_ready = 0;
+ spin_unlock_irqrestore (&sched->kss_lock, flags);
+
+ rc = ksocknal_process_transmit(conn, tx);
+
+ spin_lock_irqsave (&sched->kss_lock, flags);
+
+ if (rc == -ENOMEM || rc == -EAGAIN) {
+ /* Incomplete send: replace tx on HEAD of tx_queue */
+ list_add (&tx->tx_list, &conn->ksnc_tx_queue);
+ } else {
+ /* Complete send; assume space for more */
+ conn->ksnc_tx_ready = 1;
+ }
+
+ if (rc == -ENOMEM) {
+ /* Do nothing; after a short timeout, this
+ * conn will be reposted on kss_tx_conns. */
+ } else if (conn->ksnc_tx_ready &&
+ !list_empty (&conn->ksnc_tx_queue)) {
+ /* reschedule for tx */
+ list_add_tail (&conn->ksnc_tx_list,
+ &sched->kss_tx_conns);
+ } else {
+ conn->ksnc_tx_scheduled = 0;
+ /* drop my ref */
+ ksocknal_put_conn (conn);
+ }
+
did_something = 1;
- /* drops and regains kss_lock */
- ksocknal_process_transmit (sched, &flags);
}
#if SOCKNAL_ZC
if (!list_empty (&sched->kss_zctxdone_list)) {
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();
read_lock (&ksocknal_data.ksnd_global_lock);
conn = sk->sk_user_data;
- if (conn == NULL) { /* raced with ksocknal_close_sock */
+ if (conn == NULL) { /* raced with ksocknal_terminate_conn */
LASSERT (sk->sk_data_ready != &ksocknal_data_ready);
sk->sk_data_ready (sk, n);
- } else if (!conn->ksnc_rx_ready) { /* new news */
- /* Set ASAP in case of concurrent calls to me */
- conn->ksnc_rx_ready = 1;
-
+ } else {
sched = conn->ksnc_scheduler;
spin_lock_irqsave (&sched->kss_lock, flags);
- /* Set again (process_receive may have cleared while I blocked for the lock) */
conn->ksnc_rx_ready = 1;
if (!conn->ksnc_rx_scheduled) { /* not being progressed */
/* extra ref for scheduler */
atomic_inc (&conn->ksnc_refcount);
- if (waitqueue_active (&sched->kss_waitq))
- wake_up (&sched->kss_waitq);
+ wake_up (&sched->kss_waitq);
}
spin_unlock_irqrestore (&sched->kss_lock, flags);
(conn == NULL) ? "" : (list_empty (&conn->ksnc_tx_queue) ?
" empty" : " queued"));
- if (conn == NULL) { /* raced with ksocknal_close_sock */
+ if (conn == NULL) { /* raced with ksocknal_terminate_conn */
LASSERT (sk->sk_write_space != &ksocknal_write_space);
sk->sk_write_space (sk);
- } else if (tcp_wspace(sk) >= SOCKNAL_TX_LOW_WATER(sk)) { /* got enough space */
- clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
-
- if (!conn->ksnc_tx_ready) { /* new news */
- /* Set ASAP in case of concurrent calls to me */
- conn->ksnc_tx_ready = 1;
- sched = conn->ksnc_scheduler;
+ read_unlock (&ksocknal_data.ksnd_global_lock);
+ return;
+ }
- spin_lock_irqsave (&sched->kss_lock, flags);
+ if (tcp_wspace(sk) >= SOCKNAL_TX_LOW_WATER(sk)) { /* got enough space */
+ sched = conn->ksnc_scheduler;
- /* Set again (process_transmit may have
- cleared while I blocked for the lock) */
- conn->ksnc_tx_ready = 1;
+ spin_lock_irqsave (&sched->kss_lock, flags);
- 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);
+ clear_bit (SOCK_NOSPACE, &sk->sk_socket->flags);
+ conn->ksnc_tx_ready = 1;
- if (waitqueue_active (&sched->kss_waitq))
- wake_up (&sched->kss_waitq);
- }
+ 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);
- spin_unlock_irqrestore (&sched->kss_lock, flags);
+ wake_up (&sched->kss_waitq);
}
+
+ spin_unlock_irqrestore (&sched->kss_lock, flags);
}
read_unlock (&ksocknal_data.ksnd_global_lock);
}
int
-ksocknal_exchange_nids (struct socket *sock, ptl_nid_t nid)
+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;
+ int i;
+ int rc;
+
+ LASSERT (conn->ksnc_type != SOCKNAL_CONN_NONE);
+ LASSERT (nipaddrs <= SOCKNAL_MAX_INTERFACES);
+
+ /* No need for getconnsock/putconnsock */
+ LASSERT (!conn->ksnc_closing);
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);
- memset (&hdr, 0, sizeof (hdr));
- hmv->magic = __cpu_to_le32 (PORTALS_PROTO_MAGIC);
- hmv->version_major = __cpu_to_le32 (PORTALS_PROTO_VERSION_MAJOR);
- hmv->version_minor = __cpu_to_le32 (PORTALS_PROTO_VERSION_MINOR);
-
- hdr.src_nid = __cpu_to_le64 (ksocknal_lib.ni.nid);
- hdr.type = __cpu_to_le32 (PTL_MSG_HELLO);
-
- /* Assume sufficient socket buffering for this message */
- rc = ksocknal_sock_write (sock, &hdr, sizeof (hdr));
+ 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);
+
+ /* Receiver is eager */
+ rc = ksocknal_sock_write (sock, &hdr, sizeof(hdr));
if (rc != 0) {
- CERROR ("Error %d sending HELLO to "LPX64"\n", rc, nid);
+ 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;
+
+ 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"\n", rc, nid);
+ 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"\n",
- __cpu_to_le32 (hmv->magic), PORTALS_PROTO_MAGIC, nid);
- return (-EINVAL);
+
+ 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"\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);
- return (-EINVAL);
+ HIPQUAD(conn->ksnc_ipaddr));
+ return (-EPROTO);
}
- LASSERT (PORTALS_PROTO_VERSION_MAJOR == 0);
- /* version 0 sends magic/version as the dest_nid of a 'hello' header,
- * so read the rest of it in now... */
+#if (PORTALS_PROTO_VERSION_MAJOR != 1)
+# error "This code only understands protocol version 1.x"
+#endif
+ /* 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"\n",
- rc, nid);
+ 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) ||
- PTL_HDR_LENGTH (&hdr) != __cpu_to_le32 (0)) {
- CERROR ("Expecting a HELLO hdr with 0 payload,"
- " but got type %d with %d payload from "LPX64"\n",
- __le32_to_cpu (hdr.type),
- __le32_to_cpu (PTL_HDR_LENGTH (&hdr)), nid);
- return (-EINVAL);
+ 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"
+ "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"@%u.%u.%u.%u "
+ "but expecting "LPX64"\n",
+ le64_to_cpu (hdr.src_nid),
+ HIPQUAD(conn->ksnc_ipaddr), *nid);
+ return (-EPROTO);
+ }
+
+ type = __le32_to_cpu(hdr.msg.hello.type);
+
+ if (conn->ksnc_type == SOCKNAL_CONN_NONE) {
+ /* I've accepted this connection; peer determines type */
+ 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 (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);
+
+ nips = __le32_to_cpu (hdr.payload_length) / sizeof (__u32);
+
+ 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));
}
+
+ if (nips == 0)
+ return (0);
- if (__le64_to_cpu (hdr.src_nid) != nid) {
- CERROR ("Connected to nid "LPX64", but expecting "LPX64"\n",
- __le64_to_cpu (hdr.src_nid), nid);
- return (-EINVAL);
+ rc = ksocknal_sock_read (sock, ipaddrs, nips * sizeof(*ipaddrs));
+ if (rc != 0) {
+ CERROR ("Error %d reading IPs from "LPX64"@%u.%u.%u.%u\n",
+ rc, *nid, HIPQUAD(conn->ksnc_ipaddr));
+ return (rc);
}
- return (0);
+ for (i = 0; i < nips; i++) {
+ ipaddrs[i] = __le32_to_cpu(ipaddrs[i]);
+
+ if (ipaddrs[i] == 0) {
+ CERROR("Zero IP[%d] from "LPX64"@%u.%u.%u.%u\n",
+ i, *nid, HIPQUAD(conn->ksnc_ipaddr));
+ return (-EPROTO);
+ }
+ }
+
+ return (nips);
}
int
-ksocknal_set_linger (struct socket *sock)
+ksocknal_get_conn_tunables (ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
+{
+ mm_segment_t oldmm = get_fs ();
+ struct socket *sock = conn->ksnc_sock;
+ int len;
+ int rc;
+
+ rc = ksocknal_getconnsock (conn);
+ if (rc != 0) {
+ LASSERT (conn->ksnc_closing);
+ *txmem = *rxmem = *nagle = 0;
+ return (-ESHUTDOWN);
+ }
+
+ set_fs (KERNEL_DS);
+
+ len = sizeof(*txmem);
+ rc = sock_getsockopt(sock, SOL_SOCKET, SO_SNDBUF,
+ (char *)txmem, &len);
+ if (rc == 0) {
+ len = sizeof(*rxmem);
+ rc = sock_getsockopt(sock, SOL_SOCKET, SO_RCVBUF,
+ (char *)rxmem, &len);
+ }
+ if (rc == 0) {
+ len = sizeof(*nagle);
+ rc = sock->ops->getsockopt(sock, SOL_TCP, TCP_NODELAY,
+ (char *)nagle, &len);
+ }
+
+ set_fs (oldmm);
+ ksocknal_putconnsock (conn);
+
+ if (rc == 0)
+ *nagle = !*nagle;
+ else
+ *txmem = *rxmem = *nagle = 0;
+
+ return (rc);
+}
+
+int
+ksocknal_setup_sock (struct socket *sock)
{
mm_segment_t oldmm = get_fs ();
int rc;
int option;
+ int keep_idle;
+ int keep_intvl;
+ int keep_count;
+ int do_keepalive;
struct linger linger;
+ sock->sk->sk_allocation = GFP_NOFS;
+
/* 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,
+ 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);
}
-
+
option = -1;
set_fs (KERNEL_DS);
rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_LINGER2,
CERROR ("Can't set SO_LINGER2: %d\n", rc);
return (rc);
}
+
+ if (!ksocknal_tunables.ksnd_nagle) {
+ 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);
+ return (rc);
+ }
+ }
+
+ if (ksocknal_tunables.ksnd_buffer_size > 0) {
+ option = ksocknal_tunables.ksnd_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",
+ option, rc);
+ 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",
+ option, rc);
+ return (rc);
+ }
+ }
+
+ /* snapshot tunables */
+ keep_idle = ksocknal_tunables.ksnd_keepalive_idle;
+ keep_count = ksocknal_tunables.ksnd_keepalive_count;
+ keep_intvl = ksocknal_tunables.ksnd_keepalive_intvl;
+ do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
+
+ option = (do_keepalive ? 1 : 0);
+ 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);
+ }
+
+ if (!do_keepalive)
+ return (0);
+
+ set_fs (KERNEL_DS);
+ rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPIDLE,
+ (char *)&keep_idle, sizeof (keep_idle));
+ set_fs (oldmm);
+ if (rc != 0) {
+ CERROR ("Can't set TCP_KEEPIDLE: %d\n", rc);
+ return (rc);
+ }
+
+ set_fs (KERNEL_DS);
+ rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPINTVL,
+ (char *)&keep_intvl, sizeof (keep_intvl));
+ set_fs (oldmm);
+ if (rc != 0) {
+ CERROR ("Can't set TCP_KEEPINTVL: %d\n", rc);
+ return (rc);
+ }
+
+ set_fs (KERNEL_DS);
+ rc = sock->ops->setsockopt (sock, SOL_TCP, TCP_KEEPCNT,
+ (char *)&keep_count, sizeof (keep_count));
+ set_fs (oldmm);
+ if (rc != 0) {
+ CERROR ("Can't set TCP_KEEPCNT: %d\n", rc);
+ return (rc);
+ }
+
return (0);
}
-int
-ksocknal_connect_peer (ksock_route_t *route)
+static int
+ksocknal_connect_sock(struct socket **sockp, int *may_retry,
+ ksock_route_t *route, int local_port)
{
- struct sockaddr_in peer_addr;
- mm_segment_t oldmm = get_fs();
- __u64 n;
- struct timeval tv;
- int fd;
+ struct sockaddr_in locaddr;
+ struct sockaddr_in srvaddr;
struct socket *sock;
int rc;
+ int option;
+ mm_segment_t oldmm = get_fs();
+ struct timeval tv;
+
+ memset(&locaddr, 0, sizeof(locaddr));
+ locaddr.sin_family = AF_INET;
+ locaddr.sin_port = htons(local_port);
+ locaddr.sin_addr.s_addr =
+ (route->ksnr_myipaddr != 0) ? htonl(route->ksnr_myipaddr)
+ : INADDR_ANY;
+
+ memset (&srvaddr, 0, sizeof (srvaddr));
+ srvaddr.sin_family = AF_INET;
+ srvaddr.sin_port = htons (route->ksnr_port);
+ srvaddr.sin_addr.s_addr = htonl (route->ksnr_ipaddr);
+
+ *may_retry = 0;
rc = sock_create (PF_INET, SOCK_STREAM, 0, &sock);
+ *sockp = sock;
if (rc != 0) {
CERROR ("Can't create autoconnect socket: %d\n", rc);
return (rc);
}
/* Ugh; have to map_fd for compatibility with sockets passed in
- * from userspace. And we actually need the refcounting that
- * this gives you :) */
+ * from userspace. And we actually need the sock->file refcounting
+ * that this gives you :) */
- fd = sock_map_fd (sock);
- if (fd < 0) {
+ rc = sock_map_fd (sock);
+ if (rc < 0) {
sock_release (sock);
- CERROR ("sock_map_fd error %d\n", fd);
- return (fd);
+ CERROR ("sock_map_fd error %d\n", rc);
+ return (rc);
}
- /* NB the fd now owns the ref on sock->file */
+ /* NB the file descriptor (rc) now owns the ref on sock->file */
LASSERT (sock->file != NULL);
LASSERT (file_count(sock->file) == 1);
-
+
+ get_file(sock->file); /* extra ref makes sock->file */
+ sys_close(rc); /* survive this close */
+
+ /* Still got a single ref on sock->file */
+ LASSERT (file_count(sock->file) == 1);
+
/* Set the socket timeouts, so our connection attempt completes in
* finite time */
- tv.tv_sec = ksocknal_io_timeout / HZ;
- n = ksocknal_io_timeout % HZ;
- n = n * 1000000 + HZ - 1;
- do_div (n, HZ);
- tv.tv_usec = n;
+ 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 (in HZ): %d\n",
- ksocknal_io_timeout, rc);
- goto out;
+ CERROR ("Can't set send timeout %d: %d\n",
+ ksocknal_tunables.ksnd_io_timeout, rc);
+ goto failed;
}
set_fs (KERNEL_DS);
(char *)&tv, sizeof (tv));
set_fs (oldmm);
if (rc != 0) {
- CERROR ("Can't set receive timeout %d (in HZ): %d\n",
- ksocknal_io_timeout, rc);
- goto out;
+ CERROR ("Can't set receive timeout %d: %d\n",
+ ksocknal_tunables.ksnd_io_timeout, rc);
+ goto failed;
}
- if (route->ksnr_nonagel) {
- 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 nagel: %d\n", rc);
- goto out;
- }
+ set_fs (KERNEL_DS);
+ option = 1;
+ rc = sock_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
+ (char *)&option, sizeof (option));
+ set_fs (oldmm);
+ if (rc != 0) {
+ CERROR("Can't set SO_REUSEADDR for socket: %d\n", rc);
+ goto failed;
}
-
- 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;
- }
- 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;
- }
+ rc = sock->ops->bind(sock,
+ (struct sockaddr *)&locaddr, sizeof(locaddr));
+ if (rc == -EADDRINUSE) {
+ CDEBUG(D_NET, "Port %d already in use\n", local_port);
+ *may_retry = 1;
+ goto failed;
}
-
- 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"\n", rc,
- route->ksnr_peer->ksnp_nid);
- goto out;
- }
+ CERROR("Error trying to bind to reserved port %d: %d\n",
+ local_port, rc);
+ goto failed;
+ }
+
+ rc = sock->ops->connect(sock,
+ (struct sockaddr *)&srvaddr, sizeof(srvaddr),
+ sock->file->f_flags);
+ if (rc == 0)
+ return 0;
+
+ /* EADDRNOTAVAIL probably means we're already connected to the same
+ * peer/port on the same local port on a differently typed
+ * connection. Let our caller retry with a different local
+ * port... */
+ *may_retry = (rc == -EADDRNOTAVAIL);
+
+ CDEBUG(*may_retry ? D_NET : D_ERROR,
+ "Error %d connecting %u.%u.%u.%u/%d -> %u.%u.%u.%u/%d\n", rc,
+ HIPQUAD(route->ksnr_myipaddr), local_port,
+ HIPQUAD(route->ksnr_ipaddr), route->ksnr_port);
+
+ failed:
+ fput(sock->file);
+ return rc;
+}
+
+int
+ksocknal_connect_peer (ksock_route_t *route, int type)
+{
+ struct socket *sock;
+ int rc;
+ int port;
+ int may_retry;
- if (route->ksnr_xchange_nids) {
- rc = ksocknal_exchange_nids (sock, route->ksnr_peer->ksnp_nid);
- if (rc != 0)
- goto out;
- }
+ /* 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. */
- rc = ksocknal_create_conn (route->ksnr_peer->ksnp_nid,
- route, sock, route->ksnr_irq_affinity);
- 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);
+ for (port = 1023; port > 512; --port) {
+
+ rc = ksocknal_connect_sock(&sock, &may_retry, route, port);
+
+ if (rc == 0) {
+ rc = ksocknal_create_conn(route, sock, type);
+ fput(sock->file);
+ return rc;
+ }
+
+ 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
ksock_peer_t *peer;
unsigned long flags;
int rc;
+ int type;
- rc = ksocknal_connect_peer (route);
- if (rc == 0) {
+ for (;;) {
+ for (type = 0; type < SOCKNAL_CONN_NTYPES; type++)
+ if ((route->ksnr_connecting & (1 << type)) != 0)
+ break;
+ LASSERT (type < SOCKNAL_CONN_NTYPES);
+
+ rc = ksocknal_connect_peer (route, type);
+ if (rc != 0)
+ break;
+
/* successfully autoconnected: create_conn did the
- * route/conn binding and scheduled any blocked packets,
- * so there's nothing left to do now. */
- return;
+ * route/conn binding and scheduled any blocked packets */
+
+ if (route->ksnr_connecting == 0) {
+ /* No more connections required */
+ return;
+ }
}
+ /* Connection attempt failed */
+
write_lock_irqsave (&ksocknal_data.ksnd_global_lock, flags);
peer = route->ksnr_peer;
route->ksnr_connecting = 0;
+ /* This is a retry rather than a new connection */
LASSERT (route->ksnr_retry_interval != 0);
- route->ksnr_timeout = jiffies_64 + route->ksnr_retry_interval;
+ route->ksnr_timeout = jiffies + 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));
}
+#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)) {
+ char ipbuf[PTL_NALFMT_SIZE];
+ char ipbuf2[PTL_NALFMT_SIZE];
tx = list_entry (zombies.next, ksock_tx_t, tx_list);
-
- CERROR ("Deleting packet type %d len %d ("LPX64"->"LPX64")\n",
- NTOH__u32 (tx->tx_hdr->type),
- NTOH__u32 (PTL_HDR_LENGTH(tx->tx_hdr)),
- NTOH__u64 (tx->tx_hdr->src_nid),
- NTOH__u64 (tx->tx_hdr->dest_nid));
+
+ CERROR ("Deleting packet type %d len %d ("LPX64" %s->"LPX64" %s)\n",
+ 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,
+ le64_to_cpu(tx->tx_hdr->src_nid),
+ ipbuf),
+ le64_to_cpu (tx->tx_hdr->dest_nid),
+ portals_nid2str(SOCKNAL,
+ le64_to_cpu(tx->tx_hdr->src_nid),
+ ipbuf2));
list_del (&tx->tx_list);
/* complete now */
ksock_route_t *route;
int rc;
- snprintf (name, sizeof (name), "ksocknal_ad[%ld]", id);
+ snprintf (name, sizeof (name), "ksocknal_ad%02ld", id);
kportal_daemonize (name);
kportal_blockallsigs ();
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);
- spin_lock_irqsave (&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_unlock_irqrestore (&ksocknal_data.ksnd_autoconnectd_lock, flags);
ksocknal_find_timed_out_conn (ksock_peer_t *peer)
{
/* We're called with a shared lock on ksnd_global_lock */
- unsigned long flags;
ksock_conn_t *conn;
struct list_head *ctmp;
- ksock_tx_t *tx;
- struct list_head *ttmp;
- ksock_sched_t *sched;
list_for_each (ctmp, &peer->ksnp_conns) {
conn = list_entry (ctmp, ksock_conn_t, ksnc_list);
- sched = conn->ksnc_scheduler;
-
- if (conn->ksnc_rx_deadline != 0 &&
- conn->ksnc_rx_deadline <= jiffies_64)
- goto timed_out;
- spin_lock_irqsave (&sched->kss_lock, flags);
-
- list_for_each (ttmp, &conn->ksnc_tx_queue) {
- tx = list_entry (ttmp, ksock_tx_t, tx_list);
- LASSERT (tx->tx_deadline != 0);
-
- if (tx->tx_deadline <= jiffies_64)
- goto timed_out_locked;
+ /* Don't need the {get,put}connsock dance to deref ksnc_sock... */
+ LASSERT (!conn->ksnc_closing);
+
+ if (conn->ksnc_sock->sk->sk_err != 0) {
+ /* Something (e.g. failed keepalive) set the socket error */
+ atomic_inc (&conn->ksnc_refcount);
+ CERROR ("Socket error %d: "LPX64" %p %d.%d.%d.%d\n",
+ conn->ksnc_sock->sk->sk_err, peer->ksnp_nid,
+ conn, HIPQUAD(conn->ksnc_ipaddr));
+ return (conn);
}
-#if SOCKNAL_ZC
- list_for_each (ttmp, &conn->ksnc_tx_pending) {
- tx = list_entry (ttmp, ksock_tx_t, tx_list);
- LASSERT (tx->tx_deadline != 0);
- if (tx->tx_deadline <= jiffies_64)
- goto timed_out_locked;
+ if (conn->ksnc_rx_started &&
+ time_after_eq (jiffies, conn->ksnc_rx_deadline)) {
+ /* Timed out incomplete incoming message */
+ atomic_inc (&conn->ksnc_refcount);
+ CERROR ("Timed out RX from "LPX64" %p %d.%d.%d.%d\n",
+ peer->ksnp_nid,conn,HIPQUAD(conn->ksnc_ipaddr));
+ return (conn);
}
-#endif
- spin_unlock_irqrestore (&sched->kss_lock, flags);
- continue;
- timed_out_locked:
- spin_unlock_irqrestore (&sched->kss_lock, flags);
- timed_out:
- atomic_inc (&conn->ksnc_refcount);
- 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
+ * 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,
+ list_empty (&conn->ksnc_tx_queue) ? "" : "Q ",
+ conn->ksnc_sock->sk->sk_wmem_queued, conn,
+ HIPQUAD(conn->ksnc_ipaddr));
+ return (conn);
+ }
}
return (NULL);
}
void
-ksocknal_check_peer_timeouts (struct list_head *peers)
+ksocknal_check_peer_timeouts (int idx)
{
+ struct list_head *peers = &ksocknal_data.ksnd_peers[idx];
struct list_head *ptmp;
ksock_peer_t *peer;
ksock_conn_t *conn;
if (conn != NULL) {
read_unlock (&ksocknal_data.ksnd_global_lock);
- if (ksocknal_close_conn_unlocked (conn)) {
- /* I actually closed... */
- CERROR ("Timeout out conn->"LPX64" ip %x:%d\n",
- peer->ksnp_nid, 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! */
wait_queue_t wait;
unsigned long flags;
ksock_conn_t *conn;
+ ksock_sched_t *sched;
+ struct list_head enomem_conns;
+ int nenomem_conns;
int timeout;
+ int i;
int peer_index = 0;
- __u64 deadline = jiffies_64;
+ unsigned long deadline = jiffies;
kportal_daemonize ("ksocknal_reaper");
kportal_blockallsigs ();
+ INIT_LIST_HEAD(&enomem_conns);
init_waitqueue_entry (&wait, current);
spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
spin_lock_irqsave (&ksocknal_data.ksnd_reaper_lock, flags);
continue;
}
-
+
+ if (!list_empty (&ksocknal_data.ksnd_enomem_conns)) {
+ list_add(&enomem_conns, &ksocknal_data.ksnd_enomem_conns);
+ list_del_init(&ksocknal_data.ksnd_enomem_conns);
+ }
+
spin_unlock_irqrestore (&ksocknal_data.ksnd_reaper_lock, flags);
- while ((timeout = deadline - jiffies_64) <= 0) {
- /* Time to check for timeouts on a few more peers */
- ksocknal_check_peer_timeouts (&ksocknal_data.ksnd_peers[peer_index]);
+ /* reschedule all the connections that stalled with ENOMEM... */
+ nenomem_conns = 0;
+ while (!list_empty (&enomem_conns)) {
+ conn = list_entry (enomem_conns.next,
+ ksock_conn_t, ksnc_tx_list);
+ list_del (&conn->ksnc_tx_list);
- peer_index = (peer_index + 1) % SOCKNAL_PEER_HASH_SIZE;
- deadline += HZ;
+ sched = conn->ksnc_scheduler;
+
+ spin_lock_irqsave (&sched->kss_lock, flags);
+
+ 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);
+
+ spin_unlock_irqrestore (&sched->kss_lock, flags);
+ nenomem_conns++;
+ }
+
+ /* careful with the jiffy wrap... */
+ while ((timeout = (int)(deadline - jiffies)) <= 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
+ * within a timeout interval, to ensure I detect a
+ * timeout on any connection within (n+1)/n times the
+ * timeout interval. */
+
+ if (ksocknal_tunables.ksnd_io_timeout > 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) %
+ ksocknal_data.ksnd_peer_hash_size;
+ }
+
+ deadline += p * HZ;
}
- add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
+ if (nenomem_conns != 0) {
+ /* Reduce my timeout if I rescheduled ENOMEM conns.
+ * This also prevents me getting woken immediately
+ * if any go back on my enomem list. */
+ timeout = SOCKNAL_ENOMEM_RETRY;
+ }
+ ksocknal_data.ksnd_reaper_waketime = jiffies + timeout;
+
set_current_state (TASK_INTERRUPTIBLE);
+ add_wait_queue (&ksocknal_data.ksnd_reaper_waitq, &wait);
if (!ksocknal_data.ksnd_shuttingdown &&
list_empty (&ksocknal_data.ksnd_deathrow_conns) &&
return (0);
}
-nal_cb_t ksocknal_lib = {
- nal_data: &ksocknal_data, /* NAL private data */
- cb_send: ksocknal_send,
- cb_send_pages: ksocknal_send_pages,
- cb_recv: ksocknal_recv,
- cb_recv_pages: ksocknal_recv_pages,
- cb_read: ksocknal_read,
- cb_write: ksocknal_write,
- cb_callback: ksocknal_callback,
- cb_malloc: ksocknal_malloc,
- cb_free: ksocknal_free,
- cb_printf: ksocknal_printf,
- cb_cli: ksocknal_cli,
- cb_sti: ksocknal_sti,
- cb_dist: ksocknal_dist
+lib_nal_t ksocknal_lib = {
+ libnal_data: &ksocknal_data, /* NAL private data */
+ libnal_send: ksocknal_send,
+ libnal_send_pages: ksocknal_send_pages,
+ libnal_recv: ksocknal_recv,
+ libnal_recv_pages: ksocknal_recv_pages,
+ libnal_dist: ksocknal_dist
};