__u32 sd_send_case;
};
+static inline bool
+lnet_msg_is_response(struct lnet_msg *msg)
+{
+ return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
+}
+
+static inline bool
+lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
+{
+ if (md_options & LNET_MD_NO_TRACK_RESPONSE)
+ /* Explicitly disabled in MD options */
+ return false;
+
+ if (md_options & LNET_MD_TRACK_RESPONSE)
+ /* Explicity enabled in MD options */
+ return true;
+
+ if (lnet_response_tracking == 3)
+ /* Enabled for all message types */
+ return true;
+
+ if (msg_type == LNET_MSG_PUT)
+ return lnet_response_tracking == 2;
+
+ if (msg_type == LNET_MSG_GET)
+ return lnet_response_tracking == 1;
+
+ return false;
+}
+
static inline struct lnet_comm_count *
get_stats_counts(struct lnet_element_stats *stats,
enum lnet_stats_type stats_type)
struct lnet_test_peer *tp;
struct list_head *el;
struct list_head *next;
- struct list_head cull;
+ LIST_HEAD(cull);
/* NB: use lnet_net_lock(0) to serialize operations on test peers */
if (threshold != 0) {
return 0;
}
- /* removing entries */
- INIT_LIST_HEAD(&cull);
-
lnet_net_lock(0);
list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
if (tp->tp_threshold == 0 || /* needs culling anyway */
nid == LNET_NID_ANY || /* removing all entries */
tp->tp_nid == nid) { /* matched this one */
- list_del(&tp->tp_list);
- list_add(&tp->tp_list, &cull);
+ list_move(&tp->tp_list, &cull);
}
}
struct lnet_test_peer *tp;
struct list_head *el;
struct list_head *next;
- struct list_head cull;
- int fail = 0;
-
- INIT_LIST_HEAD(&cull);
+ LIST_HEAD(cull);
+ int fail = 0;
/* NB: use lnet_net_lock(0) to serialize operations on test peers */
lnet_net_lock(0);
/* only cull zombies on outgoing tests,
* since we may be at interrupt priority on
* incoming messages. */
- list_del(&tp->tp_list);
- list_add(&tp->tp_list, &cull);
+ list_move(&tp->tp_list, &cull);
}
continue;
}
if (outgoing &&
tp->tp_threshold == 0) {
/* see above */
- list_del(&tp->tp_list);
- list_add(&tp->tp_list, &cull);
+ list_move(&tp->tp_list, &cull);
}
}
break;
unsigned int nob)
{
/* NB diov, siov are READ-ONLY */
- unsigned int this_nob;
+ unsigned int this_nob;
if (nob == 0)
return;
do {
LASSERT(ndiov > 0);
LASSERT(nsiov > 0);
- this_nob = MIN(diov->iov_len - doffset,
- siov->iov_len - soffset);
- this_nob = MIN(this_nob, nob);
+ this_nob = min3((unsigned int)diov->iov_len - doffset,
+ (unsigned int)siov->iov_len - soffset,
+ nob);
memcpy((char *)diov->iov_base + doffset,
(char *)siov->iov_base + soffset, this_nob);
}
EXPORT_SYMBOL(lnet_copy_iov2iov);
-int
-lnet_extract_iov(int dst_niov, struct kvec *dst,
- int src_niov, struct kvec *src,
- unsigned int offset, unsigned int len)
-{
- /* Initialise 'dst' to the subset of 'src' starting at 'offset',
- * for exactly 'len' bytes, and return the number of entries.
- * NB not destructive to 'src' */
- unsigned int frag_len;
- unsigned int niov;
-
- if (len == 0) /* no data => */
- return (0); /* no frags */
-
- LASSERT(src_niov > 0);
- while (offset >= src->iov_len) { /* skip initial frags */
- offset -= src->iov_len;
- src_niov--;
- src++;
- LASSERT(src_niov > 0);
- }
-
- niov = 1;
- for (;;) {
- LASSERT(src_niov > 0);
- LASSERT((int)niov <= dst_niov);
-
- frag_len = src->iov_len - offset;
- dst->iov_base = ((char *)src->iov_base) + offset;
-
- if (len <= frag_len) {
- dst->iov_len = len;
- return (niov);
- }
-
- dst->iov_len = frag_len;
-
- len -= frag_len;
- dst++;
- src++;
- niov++;
- src_niov--;
- offset = 0;
- }
-}
-EXPORT_SYMBOL(lnet_extract_iov);
-
-
unsigned int
-lnet_kiov_nob(unsigned int niov, lnet_kiov_t *kiov)
+lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
{
unsigned int nob = 0;
LASSERT(niov == 0 || kiov != NULL);
while (niov-- > 0)
- nob += (kiov++)->kiov_len;
+ nob += (kiov++)->bv_len;
return (nob);
}
EXPORT_SYMBOL(lnet_kiov_nob);
void
-lnet_copy_kiov2kiov(unsigned int ndiov, lnet_kiov_t *diov, unsigned int doffset,
- unsigned int nsiov, lnet_kiov_t *siov, unsigned int soffset,
+lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
+ unsigned int doffset,
+ unsigned int nsiov, struct bio_vec *siov,
+ unsigned int soffset,
unsigned int nob)
{
/* NB diov, siov are READ-ONLY */
LASSERT (!in_interrupt ());
LASSERT (ndiov > 0);
- while (doffset >= diov->kiov_len) {
- doffset -= diov->kiov_len;
+ while (doffset >= diov->bv_len) {
+ doffset -= diov->bv_len;
diov++;
ndiov--;
LASSERT(ndiov > 0);
}
LASSERT(nsiov > 0);
- while (soffset >= siov->kiov_len) {
- soffset -= siov->kiov_len;
+ while (soffset >= siov->bv_len) {
+ soffset -= siov->bv_len;
siov++;
nsiov--;
LASSERT(nsiov > 0);
do {
LASSERT(ndiov > 0);
LASSERT(nsiov > 0);
- this_nob = MIN(diov->kiov_len - doffset,
- siov->kiov_len - soffset);
- this_nob = MIN(this_nob, nob);
+ this_nob = min3(diov->bv_len - doffset,
+ siov->bv_len - soffset,
+ nob);
if (daddr == NULL)
- daddr = ((char *)kmap(diov->kiov_page)) +
- diov->kiov_offset + doffset;
+ daddr = ((char *)kmap(diov->bv_page)) +
+ diov->bv_offset + doffset;
if (saddr == NULL)
- saddr = ((char *)kmap(siov->kiov_page)) +
- siov->kiov_offset + soffset;
+ saddr = ((char *)kmap(siov->bv_page)) +
+ siov->bv_offset + soffset;
/* Vanishing risk of kmap deadlock when mapping 2 pages.
* However in practice at least one of the kiovs will be mapped
memcpy (daddr, saddr, this_nob);
nob -= this_nob;
- if (diov->kiov_len > doffset + this_nob) {
+ if (diov->bv_len > doffset + this_nob) {
daddr += this_nob;
doffset += this_nob;
} else {
- kunmap(diov->kiov_page);
+ kunmap(diov->bv_page);
daddr = NULL;
diov++;
ndiov--;
doffset = 0;
}
- if (siov->kiov_len > soffset + this_nob) {
+ if (siov->bv_len > soffset + this_nob) {
saddr += this_nob;
soffset += this_nob;
} else {
- kunmap(siov->kiov_page);
+ kunmap(siov->bv_page);
saddr = NULL;
siov++;
nsiov--;
} while (nob > 0);
if (daddr != NULL)
- kunmap(diov->kiov_page);
+ kunmap(diov->bv_page);
if (saddr != NULL)
- kunmap(siov->kiov_page);
+ kunmap(siov->bv_page);
}
EXPORT_SYMBOL(lnet_copy_kiov2kiov);
void
lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
- unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
+ unsigned int nkiov, struct bio_vec *kiov,
+ unsigned int kiovoffset,
unsigned int nob)
{
/* NB iov, kiov are READ-ONLY */
}
LASSERT(nkiov > 0);
- while (kiovoffset >= kiov->kiov_len) {
- kiovoffset -= kiov->kiov_len;
+ while (kiovoffset >= kiov->bv_len) {
+ kiovoffset -= kiov->bv_len;
kiov++;
nkiov--;
LASSERT(nkiov > 0);
do {
LASSERT(niov > 0);
LASSERT(nkiov > 0);
- this_nob = MIN(iov->iov_len - iovoffset,
- kiov->kiov_len - kiovoffset);
- this_nob = MIN(this_nob, nob);
+ this_nob = min3((unsigned int)iov->iov_len - iovoffset,
+ (unsigned int)kiov->bv_len - kiovoffset,
+ nob);
if (addr == NULL)
- addr = ((char *)kmap(kiov->kiov_page)) +
- kiov->kiov_offset + kiovoffset;
+ addr = ((char *)kmap(kiov->bv_page)) +
+ kiov->bv_offset + kiovoffset;
memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
nob -= this_nob;
iovoffset = 0;
}
- if (kiov->kiov_len > kiovoffset + this_nob) {
+ if (kiov->bv_len > kiovoffset + this_nob) {
addr += this_nob;
kiovoffset += this_nob;
} else {
- kunmap(kiov->kiov_page);
+ kunmap(kiov->bv_page);
addr = NULL;
kiov++;
nkiov--;
} while (nob > 0);
if (addr != NULL)
- kunmap(kiov->kiov_page);
+ kunmap(kiov->bv_page);
}
EXPORT_SYMBOL(lnet_copy_kiov2iov);
void
-lnet_copy_iov2kiov(unsigned int nkiov, lnet_kiov_t *kiov, unsigned int kiovoffset,
+lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
+ unsigned int kiovoffset,
unsigned int niov, struct kvec *iov, unsigned int iovoffset,
unsigned int nob)
{
LASSERT (!in_interrupt ());
LASSERT (nkiov > 0);
- while (kiovoffset >= kiov->kiov_len) {
- kiovoffset -= kiov->kiov_len;
+ while (kiovoffset >= kiov->bv_len) {
+ kiovoffset -= kiov->bv_len;
kiov++;
nkiov--;
LASSERT(nkiov > 0);
do {
LASSERT(nkiov > 0);
LASSERT(niov > 0);
- this_nob = MIN(kiov->kiov_len - kiovoffset,
- iov->iov_len - iovoffset);
- this_nob = MIN(this_nob, nob);
+ this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
+ (unsigned int)iov->iov_len - iovoffset,
+ nob);
if (addr == NULL)
- addr = ((char *)kmap(kiov->kiov_page)) +
- kiov->kiov_offset + kiovoffset;
+ addr = ((char *)kmap(kiov->bv_page)) +
+ kiov->bv_offset + kiovoffset;
memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
nob -= this_nob;
- if (kiov->kiov_len > kiovoffset + this_nob) {
+ if (kiov->bv_len > kiovoffset + this_nob) {
addr += this_nob;
kiovoffset += this_nob;
} else {
- kunmap(kiov->kiov_page);
+ kunmap(kiov->bv_page);
addr = NULL;
kiov++;
nkiov--;
} while (nob > 0);
if (addr != NULL)
- kunmap(kiov->kiov_page);
+ kunmap(kiov->bv_page);
}
EXPORT_SYMBOL(lnet_copy_iov2kiov);
int
-lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
- int src_niov, lnet_kiov_t *src,
+lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
+ int src_niov, struct bio_vec *src,
unsigned int offset, unsigned int len)
{
/* Initialise 'dst' to the subset of 'src' starting at 'offset',
return (0); /* no frags */
LASSERT(src_niov > 0);
- while (offset >= src->kiov_len) { /* skip initial frags */
- offset -= src->kiov_len;
+ while (offset >= src->bv_len) { /* skip initial frags */
+ offset -= src->bv_len;
src_niov--;
src++;
LASSERT(src_niov > 0);
LASSERT(src_niov > 0);
LASSERT((int)niov <= dst_niov);
- frag_len = src->kiov_len - offset;
- dst->kiov_page = src->kiov_page;
- dst->kiov_offset = src->kiov_offset + offset;
+ frag_len = src->bv_len - offset;
+ dst->bv_page = src->bv_page;
+ dst->bv_offset = src->bv_offset + offset;
if (len <= frag_len) {
- dst->kiov_len = len;
- LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
+ dst->bv_len = len;
+ LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
return niov;
}
- dst->kiov_len = frag_len;
- LASSERT(dst->kiov_offset + dst->kiov_len <= PAGE_SIZE);
+ dst->bv_len = frag_len;
+ LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
len -= frag_len;
dst++;
int delayed, unsigned int offset, unsigned int mlen,
unsigned int rlen)
{
- unsigned int niov = 0;
+ unsigned int niov = 0;
struct kvec *iov = NULL;
- lnet_kiov_t *kiov = NULL;
- int rc;
+ struct bio_vec *kiov = NULL;
+ int rc;
LASSERT (!in_interrupt ());
LASSERT (mlen == 0 || msg != NULL);
if (mlen != 0) {
niov = msg->msg_niov;
- iov = msg->msg_iov;
kiov = msg->msg_kiov;
LASSERT (niov > 0);
}
rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
- niov, iov, kiov, offset, mlen,
+ niov, kiov, offset, mlen,
rlen);
if (rc < 0)
lnet_finalize(msg, rc);
LASSERT(!msg->msg_routing);
LASSERT(md != NULL);
LASSERT(msg->msg_niov == 0);
- LASSERT(msg->msg_iov == NULL);
LASSERT(msg->msg_kiov == NULL);
msg->msg_niov = md->md_niov;
- if ((md->md_options & LNET_MD_KIOV) != 0)
- msg->msg_kiov = md->md_iov.kiov;
- else
- msg->msg_iov = md->md_iov.iov;
+ msg->msg_kiov = md->md_kiov;
}
void
static void
lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
{
- void *priv = msg->msg_private;
+ void *priv = msg->msg_private;
int rc;
- LASSERT (!in_interrupt ());
- LASSERT (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND ||
- (msg->msg_txcredit && msg->msg_peertxcredit));
+ LASSERT(!in_interrupt());
+ LASSERT(ni->ni_nid == LNET_NID_LO_0 ||
+ (msg->msg_txcredit && msg->msg_peertxcredit));
rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
if (rc < 0) {
return rc;
}
-/*
- * This function can be called from two paths:
- * 1. when sending a message
- * 2. when decommiting a message (lnet_msg_decommit_tx())
- * In both these cases the peer_ni should have it's reference count
- * acquired by the caller and therefore it is safe to drop the spin
- * lock before calling lnd_query()
- */
-static void
-lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
-{
- time64_t last_alive = 0;
- int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
-
- LASSERT(lnet_peer_aliveness_enabled(lp));
- LASSERT(ni->ni_net->net_lnd->lnd_query != NULL);
-
- lnet_net_unlock(cpt);
- (ni->ni_net->net_lnd->lnd_query)(ni, lp->lpni_nid, &last_alive);
- lnet_net_lock(cpt);
-
- lp->lpni_last_query = ktime_get_seconds();
-
- if (last_alive != 0) /* NI has updated timestamp */
- lp->lpni_last_alive = last_alive;
-}
-
-/* NB: always called with lnet_net_lock held */
-static inline int
-lnet_peer_is_alive(struct lnet_peer_ni *lp, time64_t now)
+static bool
+lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
{
- int alive;
time64_t deadline;
- LASSERT (lnet_peer_aliveness_enabled(lp));
+ deadline = lpni->lpni_last_alive +
+ lpni->lpni_net->net_tunables.lct_peer_timeout;
/*
- * Trust lnet_notify() if it has more recent aliveness news, but
- * ignore the initial assumed death (see lnet_peers_start_down()).
+ * assume peer_ni is alive as long as we're within the configured
+ * peer timeout
*/
- spin_lock(&lp->lpni_lock);
- if (!lp->lpni_alive && lp->lpni_alive_count > 0 &&
- lp->lpni_timestamp >= lp->lpni_last_alive) {
- spin_unlock(&lp->lpni_lock);
- return 0;
- }
+ if (deadline > now)
+ return false;
- deadline = lp->lpni_last_alive +
- lp->lpni_net->net_tunables.lct_peer_timeout;
- alive = deadline > now;
-
- /*
- * Update obsolete lp_alive except for routers assumed to be dead
- * initially, because router checker would update aliveness in this
- * case, and moreover lpni_last_alive at peer creation is assumed.
- */
- if (alive && !lp->lpni_alive &&
- !(lnet_isrouter(lp) && lp->lpni_alive_count == 0)) {
- spin_unlock(&lp->lpni_lock);
- lnet_notify_locked(lp, 0, 1, lp->lpni_last_alive);
- } else {
- spin_unlock(&lp->lpni_lock);
- }
-
- return alive;
+ return true;
}
-
/* NB: returns 1 when alive, 0 when dead, negative when error;
* may drop the lnet_net_lock */
static int
-lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp,
+lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
struct lnet_msg *msg)
{
time64_t now = ktime_get_seconds();
- if (!lnet_peer_aliveness_enabled(lp))
+ if (!lnet_peer_aliveness_enabled(lpni))
return -ENODEV;
- if (lnet_peer_is_alive(lp, now))
- return 1;
-
/*
* If we're resending a message, let's attempt to send it even if
* the peer is down to fulfill our resend quota on the message
if (msg->msg_retry_count > 0)
return 1;
- /*
- * Peer appears dead, but we should avoid frequent NI queries (at
- * most once per lnet_queryinterval seconds).
- */
- if (lp->lpni_last_query != 0) {
- static const int lnet_queryinterval = 1;
- time64_t next_query;
-
- next_query = lp->lpni_last_query + lnet_queryinterval;
-
- if (now < next_query) {
- if (lp->lpni_alive)
- CWARN("Unexpected aliveness of peer %s: "
- "%lld < %lld (%d/%d)\n",
- libcfs_nid2str(lp->lpni_nid),
- now, next_query,
- lnet_queryinterval,
- lp->lpni_net->net_tunables.lct_peer_timeout);
- return 0;
- }
- }
-
- /* query NI for latest aliveness news */
- lnet_ni_query_locked(ni, lp);
+ /* try and send recovery messages irregardless */
+ if (msg->msg_recovery)
+ return 1;
- if (lnet_peer_is_alive(lp, now))
+ /* always send any responses */
+ if (lnet_msg_is_response(msg))
return 1;
- lnet_notify_locked(lp, 0, 0, lp->lpni_last_alive);
- return 0;
+ if (!lnet_is_peer_deadline_passed(lpni, now))
+ return true;
+
+ return lnet_is_peer_ni_alive(lpni);
}
/**
LASSERT(!do_send || msg->msg_tx_delayed);
LASSERT(!msg->msg_receiving);
LASSERT(msg->msg_tx_committed);
+
/* can't get here if we're sending to the loopback interface */
- LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
+ if (the_lnet.ln_loni)
+ LASSERT(lp->lpni_nid != the_lnet.ln_loni->ni_nid);
/* NB 'lp' is always the next hop */
if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
* sets do_recv FALSE and I don't do the unlock/send/lock bit.
* I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
* received or OK to receive */
- struct lnet_peer_ni *lp = msg->msg_rxpeer;
+ struct lnet_peer_ni *lpni = msg->msg_rxpeer;
+ struct lnet_peer *lp;
struct lnet_rtrbufpool *rbp;
struct lnet_rtrbuf *rb;
- LASSERT (msg->msg_iov == NULL);
- LASSERT (msg->msg_kiov == NULL);
- LASSERT (msg->msg_niov == 0);
- LASSERT (msg->msg_routing);
- LASSERT (msg->msg_receiving);
- LASSERT (!msg->msg_sending);
+ LASSERT(msg->msg_kiov == NULL);
+ LASSERT(msg->msg_niov == 0);
+ LASSERT(msg->msg_routing);
+ LASSERT(msg->msg_receiving);
+ LASSERT(!msg->msg_sending);
+ LASSERT(lpni->lpni_peer_net);
+ LASSERT(lpni->lpni_peer_net->lpn_peer);
+
+ lp = lpni->lpni_peer_net->lpn_peer;
/* non-lnet_parse callers only receive delayed messages */
LASSERT(!do_recv || msg->msg_rx_delayed);
if (!msg->msg_peerrtrcredit) {
- spin_lock(&lp->lpni_lock);
- LASSERT((lp->lpni_rtrcredits < 0) ==
- !list_empty(&lp->lpni_rtrq));
+ /* lpni_lock protects the credit manipulation */
+ spin_lock(&lpni->lpni_lock);
msg->msg_peerrtrcredit = 1;
- lp->lpni_rtrcredits--;
- if (lp->lpni_rtrcredits < lp->lpni_minrtrcredits)
- lp->lpni_minrtrcredits = lp->lpni_rtrcredits;
+ lpni->lpni_rtrcredits--;
+ if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
+ lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
- if (lp->lpni_rtrcredits < 0) {
+ if (lpni->lpni_rtrcredits < 0) {
+ spin_unlock(&lpni->lpni_lock);
/* must have checked eager_recv before here */
LASSERT(msg->msg_rx_ready_delay);
msg->msg_rx_delayed = 1;
- list_add_tail(&msg->msg_list, &lp->lpni_rtrq);
- spin_unlock(&lp->lpni_lock);
+ /* lp_lock protects the lp_rtrq */
+ spin_lock(&lp->lp_lock);
+ list_add_tail(&msg->msg_list, &lp->lp_rtrq);
+ spin_unlock(&lp->lp_lock);
return LNET_CREDIT_WAIT;
}
- spin_unlock(&lp->lpni_lock);
+ spin_unlock(&lpni->lpni_lock);
}
rbp = lnet_msg2bufpool(msg);
void
lnet_return_rx_credits_locked(struct lnet_msg *msg)
{
- struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
+ struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
+ struct lnet_peer *lp;
struct lnet_ni *rxni = msg->msg_rxni;
struct lnet_msg *msg2;
routing_off:
if (msg->msg_peerrtrcredit) {
+ LASSERT(rxpeerni);
+ LASSERT(rxpeerni->lpni_peer_net);
+ LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
+
/* give back peer router credits */
msg->msg_peerrtrcredit = 0;
- spin_lock(&rxpeer->lpni_lock);
- LASSERT((rxpeer->lpni_rtrcredits < 0) ==
- !list_empty(&rxpeer->lpni_rtrq));
+ spin_lock(&rxpeerni->lpni_lock);
+ rxpeerni->lpni_rtrcredits++;
+ spin_unlock(&rxpeerni->lpni_lock);
- rxpeer->lpni_rtrcredits++;
+ lp = rxpeerni->lpni_peer_net->lpn_peer;
+ spin_lock(&lp->lp_lock);
/* drop all messages which are queued to be routed on that
* peer. */
if (!the_lnet.ln_routing) {
- struct list_head drop;
- INIT_LIST_HEAD(&drop);
- list_splice_init(&rxpeer->lpni_rtrq, &drop);
- spin_unlock(&rxpeer->lpni_lock);
+ LIST_HEAD(drop);
+ list_splice_init(&lp->lp_rtrq, &drop);
+ spin_unlock(&lp->lp_lock);
lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
- } else if (rxpeer->lpni_rtrcredits <= 0) {
- msg2 = list_entry(rxpeer->lpni_rtrq.next,
+ } else if (!list_empty(&lp->lp_rtrq)) {
+ int msg2_cpt;
+
+ msg2 = list_entry(lp->lp_rtrq.next,
struct lnet_msg, msg_list);
list_del(&msg2->msg_list);
- spin_unlock(&rxpeer->lpni_lock);
+ msg2_cpt = msg2->msg_rx_cpt;
+ spin_unlock(&lp->lp_lock);
+ /*
+ * messages on the lp_rtrq can be from any NID in
+ * the peer, which means they might have different
+ * cpts. We need to make sure we lock the right
+ * one.
+ */
+ if (msg2_cpt != msg->msg_rx_cpt) {
+ lnet_net_unlock(msg->msg_rx_cpt);
+ lnet_net_lock(msg2_cpt);
+ }
(void) lnet_post_routed_recv_locked(msg2, 1);
+ if (msg2_cpt != msg->msg_rx_cpt) {
+ lnet_net_unlock(msg2_cpt);
+ lnet_net_lock(msg->msg_rx_cpt);
+ }
} else {
- spin_unlock(&rxpeer->lpni_lock);
+ spin_unlock(&lp->lp_lock);
}
}
if (rxni != NULL) {
msg->msg_rxni = NULL;
lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
}
- if (rxpeer != NULL) {
+ if (rxpeerni != NULL) {
msg->msg_rxpeer = NULL;
- lnet_peer_ni_decref_locked(rxpeer);
+ lnet_peer_ni_decref_locked(rxpeerni);
+ }
+}
+
+static struct lnet_peer_ni *
+lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
+ struct lnet_peer *peer,
+ struct lnet_peer_ni *best_lpni,
+ struct lnet_peer_net *peer_net)
+{
+ /*
+ * Look at the peer NIs for the destination peer that connect
+ * to the chosen net. If a peer_ni is preferred when using the
+ * best_ni to communicate, we use that one. If there is no
+ * preferred peer_ni, or there are multiple preferred peer_ni,
+ * the available transmit credits are used. If the transmit
+ * credits are equal, we round-robin over the peer_ni.
+ */
+ struct lnet_peer_ni *lpni = NULL;
+ int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
+ INT_MIN;
+ int best_lpni_healthv = (best_lpni) ?
+ atomic_read(&best_lpni->lpni_healthv) : 0;
+ bool best_lpni_is_preferred = false;
+ bool lpni_is_preferred;
+ int lpni_healthv;
+ __u32 lpni_sel_prio;
+ __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
+
+ while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
+ /*
+ * if the best_ni we've chosen aleady has this lpni
+ * preferred, then let's use it
+ */
+ if (best_ni) {
+ lpni_is_preferred = lnet_peer_is_pref_nid_locked(lpni,
+ best_ni->ni_nid);
+ CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
+ libcfs_nid2str(best_ni->ni_nid),
+ lpni_is_preferred);
+ } else {
+ lpni_is_preferred = false;
+ }
+
+ lpni_healthv = atomic_read(&lpni->lpni_healthv);
+ lpni_sel_prio = lpni->lpni_sel_priority;
+
+ if (best_lpni)
+ CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
+ libcfs_nid2str(lpni->lpni_nid),
+ libcfs_nid2str(best_lpni->lpni_nid),
+ lpni_healthv, best_lpni_healthv,
+ lpni_sel_prio, best_sel_prio,
+ lpni->lpni_txcredits, best_lpni_credits,
+ lpni->lpni_seq, best_lpni->lpni_seq);
+ else
+ goto select_lpni;
+
+ /* pick the healthiest peer ni */
+ if (lpni_healthv < best_lpni_healthv)
+ continue;
+ else if (lpni_healthv > best_lpni_healthv) {
+ if (best_lpni_is_preferred)
+ best_lpni_is_preferred = false;
+ goto select_lpni;
+ }
+
+ if (lpni_sel_prio > best_sel_prio)
+ continue;
+ else if (lpni_sel_prio < best_sel_prio) {
+ if (best_lpni_is_preferred)
+ best_lpni_is_preferred = false;
+ goto select_lpni;
+ }
+
+ /* if this is a preferred peer use it */
+ if (!best_lpni_is_preferred && lpni_is_preferred) {
+ best_lpni_is_preferred = true;
+ goto select_lpni;
+ } else if (best_lpni_is_preferred && !lpni_is_preferred) {
+ /* this is not the preferred peer so let's ignore
+ * it.
+ */
+ continue;
+ }
+
+ if (lpni->lpni_txcredits < best_lpni_credits)
+ /* We already have a peer that has more credits
+ * available than this one. No need to consider
+ * this peer further.
+ */
+ continue;
+ else if (lpni->lpni_txcredits > best_lpni_credits)
+ goto select_lpni;
+
+ /* The best peer found so far and the current peer
+ * have the same number of available credits let's
+ * make sure to select between them using Round Robin
+ */
+ if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
+ continue;
+select_lpni:
+ best_lpni_is_preferred = lpni_is_preferred;
+ best_lpni_healthv = lpni_healthv;
+ best_sel_prio = lpni_sel_prio;
+ best_lpni = lpni;
+ best_lpni_credits = lpni->lpni_txcredits;
+ }
+
+ /* if we still can't find a peer ni then we can't reach it */
+ if (!best_lpni) {
+ __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
+ LNET_NIDNET(dst_nid);
+ CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
+ libcfs_net2str(net_id));
+ return NULL;
+ }
+
+ CDEBUG(D_NET, "sd_best_lpni = %s\n",
+ libcfs_nid2str(best_lpni->lpni_nid));
+
+ return best_lpni;
+}
+
+/*
+ * Prerequisite: the best_ni should already be set in the sd
+ * Find the best lpni.
+ * If the net id is provided then restrict lpni selection on
+ * that particular net.
+ * Otherwise find any reachable lpni. When dealing with an MR
+ * gateway and it has multiple lpnis which we can use
+ * we want to select the best one from the list of reachable
+ * ones.
+ */
+static inline struct lnet_peer_ni *
+lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
+ struct lnet_peer *peer, __u32 net_id)
+{
+ struct lnet_peer_net *peer_net;
+
+ /* find the best_lpni on any local network */
+ if (net_id == LNET_NET_ANY) {
+ struct lnet_peer_ni *best_lpni = NULL;
+ struct lnet_peer_net *lpn;
+ list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
+ /* no net specified find any reachable peer ni */
+ if (!lnet_islocalnet_locked(lpn->lpn_net_id))
+ continue;
+ best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
+ best_lpni, lpn);
+ }
+
+ return best_lpni;
}
+ /* restrict on the specified net */
+ peer_net = lnet_peer_get_net_locked(peer, net_id);
+ if (peer_net)
+ return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
+
+ return NULL;
}
static int
-lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
+lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
{
- if (p1->lpni_txqnob < p2->lpni_txqnob)
+ if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
return 1;
- if (p1->lpni_txqnob > p2->lpni_txqnob)
+ if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
return -1;
- if (p1->lpni_txcredits > p2->lpni_txcredits)
+ if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
return 1;
- if (p1->lpni_txcredits < p2->lpni_txcredits)
+ if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
return -1;
return 0;
}
+/* Compare route priorities and hop counts */
static int
lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
{
- struct lnet_peer_ni *p1 = r1->lr_gateway;
- struct lnet_peer_ni *p2 = r2->lr_gateway;
int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
- int rc;
if (r1->lr_priority < r2->lr_priority)
return 1;
if (r1_hops > r2_hops)
return -1;
- rc = lnet_compare_peers(p1, p2);
- if (rc)
- return rc;
-
- if (r1->lr_seq - r2->lr_seq <= 0)
- return 1;
-
- return -1;
+ return 0;
}
-static struct lnet_peer_ni *
-lnet_find_route_locked(struct lnet_net *net, __u32 remote_net,
- lnet_nid_t rtr_nid, struct lnet_route **use_route,
- struct lnet_route **prev_route)
+static struct lnet_route *
+lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
+ struct lnet_peer_ni *remote_lpni,
+ struct lnet_route **prev_route,
+ struct lnet_peer_ni **gwni)
{
- struct lnet_remotenet *rnet;
- struct lnet_route *route;
- struct lnet_route *best_route;
- struct lnet_route *last_route;
- struct lnet_peer_ni *lpni_best;
- struct lnet_peer_ni *lp;
- int rc;
-
- /* If @rtr_nid is not LNET_NID_ANY, return the gateway with
- * rtr_nid nid, otherwise find the best gateway I can use */
+ struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
+ struct lnet_route *best_route;
+ struct lnet_route *last_route;
+ struct lnet_route *route;
+ int rc;
+ bool best_rte_is_preferred = false;
+ lnet_nid_t gw_pnid;
- rnet = lnet_find_rnet_locked(remote_net);
- if (rnet == NULL)
- return NULL;
+ CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
+ libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
- lpni_best = NULL;
best_route = last_route = NULL;
list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
- lp = route->lr_gateway;
-
if (!lnet_is_route_alive(route))
continue;
+ gw_pnid = route->lr_gateway->lp_primary_nid;
- if (net != NULL && lp->lpni_net != net)
- continue;
+ /* no protection on below fields, but it's harmless */
+ if (last_route && (last_route->lr_seq - route->lr_seq < 0))
+ last_route = route;
- if (lp->lpni_nid == rtr_nid) /* it's pre-determined router */
- return lp;
+ /* if the best route found is in the preferred list then
+ * tag it as preferred and use it later on. But if we
+ * didn't find any routes which are on the preferred list
+ * then just use the best route possible.
+ */
+ rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
+
+ if (!best_route || (rc && !best_rte_is_preferred)) {
+ /* Restrict the selection of the router NI on the
+ * src_net provided. If the src_net is LNET_NID_ANY,
+ * then select the best interface available.
+ */
+ lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
+ route->lr_gateway,
+ src_net);
+ if (!lpni) {
+ CDEBUG(D_NET,
+ "Gateway %s does not have a peer NI on net %s\n",
+ libcfs_nid2str(gw_pnid),
+ libcfs_net2str(src_net));
+ continue;
+ }
+ }
+
+ if (rc && !best_rte_is_preferred) {
+ /* This is the first preferred route we found,
+ * so it beats any route found previously
+ */
+ best_route = route;
+ if (!last_route)
+ last_route = route;
+ best_gw_ni = lpni;
+ best_rte_is_preferred = true;
+ CDEBUG(D_NET, "preferred gw = %s\n",
+ libcfs_nid2str(gw_pnid));
+ continue;
+ } else if ((!rc) && best_rte_is_preferred)
+ /* The best route we found so far is in the preferred
+ * list, so it beats any non-preferred route
+ */
+ continue;
- if (lpni_best == NULL) {
+ if (!best_route) {
best_route = last_route = route;
- lpni_best = lp;
+ best_gw_ni = lpni;
continue;
}
- /* no protection on below fields, but it's harmless */
- if (last_route->lr_seq - route->lr_seq < 0)
- last_route = route;
-
rc = lnet_compare_routes(route, best_route);
- if (rc < 0)
+ if (rc == -1)
continue;
- best_route = route;
- lpni_best = lp;
- }
+ /* Restrict the selection of the router NI on the
+ * src_net provided. If the src_net is LNET_NID_ANY,
+ * then select the best interface available.
+ */
+ lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
+ route->lr_gateway,
+ src_net);
+ if (!lpni) {
+ CDEBUG(D_NET,
+ "Gateway %s does not have a peer NI on net %s\n",
+ libcfs_nid2str(gw_pnid),
+ libcfs_net2str(src_net));
+ continue;
+ }
+
+ if (rc == 1) {
+ best_route = route;
+ best_gw_ni = lpni;
+ continue;
+ }
+
+ rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
+ if (rc == -1)
+ continue;
- if (best_route != NULL) {
- *use_route = best_route;
- *prev_route = last_route;
+ if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
+ best_route = route;
+ best_gw_ni = lpni;
+ continue;
+ }
}
- return lpni_best;
+
+ *prev_route = last_route;
+ *gwni = best_gw_ni;
+
+ return best_route;
}
static struct lnet_ni *
unsigned int shortest_distance;
int best_credits;
int best_healthv;
+ __u32 best_sel_prio;
/*
* If there is no peer_ni that we can send to on this network,
return best_ni;
if (best_ni == NULL) {
+ best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
shortest_distance = UINT_MAX;
best_credits = INT_MIN;
best_healthv = 0;
best_ni->ni_dev_cpt);
best_credits = atomic_read(&best_ni->ni_tx_credits);
best_healthv = atomic_read(&best_ni->ni_healthv);
+ best_sel_prio = best_ni->ni_sel_priority;
}
while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
int ni_credits;
int ni_healthv;
int ni_fatal;
+ __u32 ni_sel_prio;
ni_credits = atomic_read(&ni->ni_tx_credits);
ni_healthv = atomic_read(&ni->ni_healthv);
ni_fatal = atomic_read(&ni->ni_fatal_error_on);
+ ni_sel_prio = ni->ni_sel_priority;
/*
* calculate the distance from the CPT on which
md_cpt,
ni->ni_dev_cpt);
- CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
- libcfs_nid2str(ni->ni_nid), ni_credits, distance,
- ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
- : "not seleced", best_credits, shortest_distance,
- (best_ni) ? best_ni->ni_seq : 0);
-
/*
* All distances smaller than the NUMA range
* are treated equally.
* Select on health, shorter distance, available
* credits, then round-robin.
*/
- if (ni_fatal) {
+ if (ni_fatal)
continue;
- } else if (ni_healthv < best_healthv) {
+
+ if (best_ni)
+ CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u]\n",
+ libcfs_nid2str(ni->ni_nid), ni_credits, distance,
+ ni->ni_seq, ni_sel_prio,
+ (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
+ : "not selected", best_credits, shortest_distance,
+ (best_ni) ? best_ni->ni_seq : 0,
+ best_sel_prio);
+ else
+ goto select_ni;
+
+ if (ni_healthv < best_healthv)
continue;
- } else if (ni_healthv > best_healthv) {
- best_healthv = ni_healthv;
- /*
- * If we're going to prefer this ni because it's
- * the healthiest, then we should set the
- * shortest_distance in the algorithm in case
- * there are multiple NIs with the same health but
- * different distances.
- */
- if (distance < shortest_distance)
- shortest_distance = distance;
- } else if (distance > shortest_distance) {
+ else if (ni_healthv > best_healthv)
+ goto select_ni;
+
+ if (ni_sel_prio > best_sel_prio)
+ continue;
+ else if (ni_sel_prio < best_sel_prio)
+ goto select_ni;
+
+ if (distance > shortest_distance)
continue;
- } else if (distance < shortest_distance) {
- shortest_distance = distance;
- } else if (ni_credits < best_credits) {
+ else if (distance < shortest_distance)
+ goto select_ni;
+
+ if (ni_credits < best_credits)
continue;
- } else if (ni_credits == best_credits) {
- if (best_ni && best_ni->ni_seq <= ni->ni_seq)
- continue;
- }
+ else if (ni_credits > best_credits)
+ goto select_ni;
+
+ if (best_ni && best_ni->ni_seq <= ni->ni_seq)
+ continue;
+
+select_ni:
+ best_sel_prio = ni_sel_prio;
+ shortest_distance = distance;
+ best_healthv = ni_healthv;
best_ni = ni;
best_credits = ni_credits;
}
struct lnet_msg *msg = sd->sd_msg;
int cpt = sd->sd_cpt;
+ if (the_lnet.ln_state != LNET_STATE_RUNNING)
+ return -ESHUTDOWN;
+
/* No send credit hassles with LOLND */
lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
msg->msg_hdr.dest_nid = cpu_to_le64(the_lnet.ln_loni->ni_nid);
__u32 routing = send_case & REMOTE_DST;
struct lnet_rsp_tracker *rspt;
- /*
- * Increment sequence number of the selected peer so that we
- * pick the next one in Round Robin.
+ /* Increment sequence number of the selected peer, peer net,
+ * local ni and local net so that we pick the next ones
+ * in Round Robin.
*/
best_lpni->lpni_seq++;
+ best_lpni->lpni_peer_net->lpn_seq++;
+ best_ni->ni_seq++;
+ best_ni->ni_net->net_seq++;
+
+ CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
+ libcfs_nid2str(best_ni->ni_nid),
+ best_ni->ni_seq, best_ni->ni_net->net_seq,
+ atomic_read(&best_ni->ni_tx_credits),
+ best_ni->ni_sel_priority,
+ libcfs_nid2str(best_lpni->lpni_nid),
+ best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
+ best_lpni->lpni_txcredits,
+ best_lpni->lpni_sel_priority);
/*
* grab a reference on the peer_ni so it sticks around even if
rc = lnet_post_send_locked(msg, 0);
if (!rc)
- CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) : %s try# %d\n",
+ CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
libcfs_nid2str(msg->msg_hdr.src_nid),
libcfs_nid2str(msg->msg_txni->ni_nid),
libcfs_nid2str(sd->sd_src_nid),
libcfs_nid2str(msg->msg_hdr.dest_nid),
libcfs_nid2str(sd->sd_dst_nid),
libcfs_nid2str(msg->msg_txpeer->lpni_nid),
+ libcfs_nid2str(sd->sd_rtr_nid),
lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
return rc;
}
-static struct lnet_peer_ni *
-lnet_select_peer_ni(struct lnet_send_data *sd, struct lnet_peer *peer,
- struct lnet_peer_net *peer_net)
-{
- /*
- * Look at the peer NIs for the destination peer that connect
- * to the chosen net. If a peer_ni is preferred when using the
- * best_ni to communicate, we use that one. If there is no
- * preferred peer_ni, or there are multiple preferred peer_ni,
- * the available transmit credits are used. If the transmit
- * credits are equal, we round-robin over the peer_ni.
- */
- struct lnet_peer_ni *lpni = NULL;
- struct lnet_peer_ni *best_lpni = NULL;
- struct lnet_ni *best_ni = sd->sd_best_ni;
- lnet_nid_t dst_nid = sd->sd_dst_nid;
- int best_lpni_credits = INT_MIN;
- bool preferred = false;
- bool ni_is_pref;
- int best_lpni_healthv = 0;
- int lpni_healthv;
-
- while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
- /*
- * if the best_ni we've chosen aleady has this lpni
- * preferred, then let's use it
- */
- ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
- best_ni->ni_nid);
-
- lpni_healthv = atomic_read(&lpni->lpni_healthv);
-
- CDEBUG(D_NET, "%s ni_is_pref = %d\n",
- libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
-
- if (best_lpni)
- CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
- libcfs_nid2str(lpni->lpni_nid),
- lpni->lpni_txcredits, best_lpni_credits,
- lpni->lpni_seq, best_lpni->lpni_seq);
-
- /* pick the healthiest peer ni */
- if (lpni_healthv < best_lpni_healthv) {
- continue;
- } else if (lpni_healthv > best_lpni_healthv) {
- best_lpni_healthv = lpni_healthv;
- /* if this is a preferred peer use it */
- } else if (!preferred && ni_is_pref) {
- preferred = true;
- } else if (preferred && !ni_is_pref) {
- /*
- * this is not the preferred peer so let's ignore
- * it.
- */
- continue;
- } else if (lpni->lpni_txcredits < best_lpni_credits) {
- /*
- * We already have a peer that has more credits
- * available than this one. No need to consider
- * this peer further.
- */
- continue;
- } else if (lpni->lpni_txcredits == best_lpni_credits) {
- /*
- * The best peer found so far and the current peer
- * have the same number of available credits let's
- * make sure to select between them using Round
- * Robin
- */
- if (best_lpni) {
- if (best_lpni->lpni_seq <= lpni->lpni_seq)
- continue;
- }
- }
-
- best_lpni = lpni;
- best_lpni_credits = lpni->lpni_txcredits;
- }
-
- /* if we still can't find a peer ni then we can't reach it */
- if (!best_lpni) {
- __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
- LNET_NIDNET(dst_nid);
- CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
- libcfs_net2str(net_id));
- return NULL;
- }
-
- CDEBUG(D_NET, "sd_best_lpni = %s\n",
- libcfs_nid2str(best_lpni->lpni_nid));
-
- return best_lpni;
-}
-
-/*
- * Prerequisite: the best_ni should already be set in the sd
- */
-static inline struct lnet_peer_ni *
-lnet_find_best_lpni_on_net(struct lnet_send_data *sd, struct lnet_peer *peer,
- __u32 net_id)
-{
- struct lnet_peer_net *peer_net;
-
- /*
- * The gateway is Multi-Rail capable so now we must select the
- * proper peer_ni
- */
- peer_net = lnet_peer_get_net_locked(peer, net_id);
-
- if (!peer_net) {
- CERROR("gateway peer %s has no NI on net %s\n",
- libcfs_nid2str(peer->lp_primary_nid),
- libcfs_net2str(net_id));
- return NULL;
- }
-
- return lnet_select_peer_ni(sd, peer, peer_net);
-}
-
static inline void
-lnet_set_non_mr_pref_nid(struct lnet_send_data *sd)
+lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
+ struct lnet_msg *msg)
{
- if (sd->sd_send_case & NMR_DST &&
- sd->sd_msg->msg_type != LNET_MSG_REPLY &&
- sd->sd_msg->msg_type != LNET_MSG_ACK &&
- sd->sd_best_lpni->lpni_pref_nnids == 0) {
+ if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
+ !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
- libcfs_nid2str(sd->sd_best_ni->ni_nid),
- libcfs_nid2str(sd->sd_best_lpni->lpni_nid));
- lnet_peer_ni_set_non_mr_pref_nid(sd->sd_best_lpni,
- sd->sd_best_ni->ni_nid);
+ libcfs_nid2str(lni->ni_nid),
+ libcfs_nid2str(lpni->lpni_nid));
+ lnet_peer_ni_set_non_mr_pref_nid(lpni, lni->ni_nid);
}
}
return -EINVAL;
}
- /*
- * the preferred NID will only be set for NMR peers
- */
- lnet_set_non_mr_pref_nid(sd);
+ lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
return lnet_handle_send(sd);
}
* Local Destination
* MR Peer
*
- * Run the selection algorithm on the peer NIs unless we're sending
- * a response, in this case just send to the destination
+ * Don't run the selection algorithm on the peer NIs. By specifying the
+ * local NID, we're also saying that we should always use the destination NID
+ * provided. This handles the case where we should be using the same
+ * destination NID for the all the messages which belong to the same RPC
+ * request.
*/
static int
lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
return -EINVAL;
}
- /*
- * only run the selection algorithm to pick the peer_ni if we're
- * sending a GET or a PUT. Responses are sent to the same
- * destination NID provided.
- */
- if (!(sd->sd_send_case & SND_RESP)) {
- sd->sd_best_lpni =
- lnet_find_best_lpni_on_net(sd, sd->sd_peer,
- sd->sd_best_ni->ni_net->net_id);
- }
-
if (sd->sd_best_lpni &&
sd->sd_best_lpni->lpni_nid == the_lnet.ln_loni->ni_nid)
return lnet_handle_lo_send(sd);
lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
struct lnet_peer *peer,
struct lnet_peer_net *peer_net,
- int cpt,
- bool incr_seq)
+ int cpt)
{
struct lnet_net *local_net;
struct lnet_ni *best_ni;
best_ni = lnet_get_best_ni(local_net, cur_best_ni,
peer, peer_net, cpt);
- if (incr_seq && best_ni)
- best_ni->ni_seq++;
-
return best_ni;
}
static int
-lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni,
- struct lnet_msg *msg, lnet_nid_t rtr_nid,
+lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
int cpt)
{
struct lnet_peer *peer;
- lnet_nid_t primary_nid;
+ struct lnet_peer_ni *new_lpni;
int rc;
lnet_peer_ni_addref_locked(lpni);
+ peer = lpni->lpni_peer_net->lpn_peer;
+
+ if (lnet_peer_gw_discovery(peer)) {
+ lnet_peer_ni_decref_locked(lpni);
+ return 0;
+ }
+
+ if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
+ lnet_peer_ni_decref_locked(lpni);
+ return 0;
+ }
+
rc = lnet_discover_peer_locked(lpni, cpt, false);
if (rc) {
lnet_peer_ni_decref_locked(lpni);
return rc;
}
- /* The peer may have changed. */
- peer = lpni->lpni_peer_net->lpn_peer;
- /* queue message and return */
- msg->msg_rtr_nid_param = rtr_nid;
- msg->msg_sending = 0;
- msg->msg_txpeer = NULL;
+
+ new_lpni = lnet_find_peer_ni_locked(lpni->lpni_nid);
+ if (!new_lpni) {
+ lnet_peer_ni_decref_locked(lpni);
+ return -ENOENT;
+ }
+
+ peer = new_lpni->lpni_peer_net->lpn_peer;
spin_lock(&peer->lp_lock);
- list_add_tail(&msg->msg_list, &peer->lp_dc_pendq);
+ if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
+ /* The peer NI did not change and the peer is up to date.
+ * Nothing more to do.
+ */
+ spin_unlock(&peer->lp_lock);
+ lnet_peer_ni_decref_locked(lpni);
+ lnet_peer_ni_decref_locked(new_lpni);
+ return 0;
+ }
spin_unlock(&peer->lp_lock);
+
+ /* Either the peer NI changed during discovery, or the peer isn't up
+ * to date. In both cases we want to queue the message on the
+ * (possibly new) peer's pending queue and queue the peer for discovery
+ */
+ msg->msg_sending = 0;
+ msg->msg_txpeer = NULL;
+ lnet_net_unlock(cpt);
+ lnet_peer_queue_message(peer, msg);
+ lnet_net_lock(cpt);
+
lnet_peer_ni_decref_locked(lpni);
- primary_nid = peer->lp_primary_nid;
+ lnet_peer_ni_decref_locked(new_lpni);
CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
- msg, libcfs_nid2str(primary_nid));
+ msg, libcfs_nid2str(peer->lp_primary_nid));
+
+ return LNET_DC_WAIT;
+}
+
+static int
+lnet_handle_find_routed_path(struct lnet_send_data *sd,
+ lnet_nid_t dst_nid,
+ struct lnet_peer_ni **gw_lpni,
+ struct lnet_peer **gw_peer)
+{
+ int rc;
+ __u32 local_lnet;
+ struct lnet_peer *gw;
+ struct lnet_peer *lp;
+ struct lnet_peer_net *lpn;
+ struct lnet_peer_net *best_lpn = NULL;
+ struct lnet_remotenet *rnet, *best_rnet = NULL;
+ struct lnet_route *best_route = NULL;
+ struct lnet_route *last_route = NULL;
+ struct lnet_peer_ni *lpni = NULL;
+ struct lnet_peer_ni *gwni = NULL;
+ bool route_found = false;
+ lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
+ (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
+ LNET_NID_ANY;
+ int best_lpn_healthv = 0;
+ __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
+
+ CDEBUG(D_NET, "using src nid %s for route restriction\n",
+ libcfs_nid2str(src_nid));
+
+ /* If a router nid was specified then we are replying to a GET or
+ * sending an ACK. In this case we use the gateway associated with the
+ * specified router nid.
+ */
+ if (sd->sd_rtr_nid != LNET_NID_ANY) {
+ gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
+ if (gwni) {
+ gw = gwni->lpni_peer_net->lpn_peer;
+ lnet_peer_ni_decref_locked(gwni);
+ if (gw->lp_rtr_refcount) {
+ local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
+ route_found = true;
+ }
+ } else {
+ CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
+ libcfs_nid2str(sd->sd_rtr_nid));
+ }
+ }
+
+ if (!route_found) {
+ if (sd->sd_msg->msg_routing) {
+ /* If I'm routing this message then I need to find the
+ * next hop based on the destination NID
+ */
+ best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
+ if (!best_rnet) {
+ CERROR("Unable to route message to %s - Route table may be misconfigured\n",
+ libcfs_nid2str(sd->sd_dst_nid));
+ return -EHOSTUNREACH;
+ }
+ } else {
+ /* we've already looked up the initial lpni using
+ * dst_nid
+ */
+ lpni = sd->sd_best_lpni;
+ /* the peer tree must be in existence */
+ LASSERT(lpni && lpni->lpni_peer_net &&
+ lpni->lpni_peer_net->lpn_peer);
+ lp = lpni->lpni_peer_net->lpn_peer;
+
+ list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
+ /* is this remote network reachable? */
+ rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
+ if (!rnet)
+ continue;
+
+ if (!best_lpn) {
+ best_lpn = lpn;
+ best_rnet = rnet;
+ }
+
+ /* select the preferred peer net */
+ if (best_lpn_healthv > lpn->lpn_healthv)
+ continue;
+ else if (best_lpn_healthv < lpn->lpn_healthv)
+ goto use_lpn;
+
+ if (best_lpn_sel_prio < lpn->lpn_sel_priority)
+ continue;
+ else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
+ goto use_lpn;
+
+ if (best_lpn->lpn_seq <= lpn->lpn_seq)
+ continue;
+use_lpn:
+ best_lpn_healthv = lpn->lpn_healthv;
+ best_lpn_sel_prio = lpn->lpn_sel_priority;
+ best_lpn = lpn;
+ best_rnet = rnet;
+ }
+
+ if (!best_lpn) {
+ CERROR("peer %s has no available nets\n",
+ libcfs_nid2str(sd->sd_dst_nid));
+ return -EHOSTUNREACH;
+ }
+
+ sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
+ sd->sd_dst_nid,
+ lp,
+ best_lpn->lpn_net_id);
+ if (!sd->sd_best_lpni) {
+ CERROR("peer %s is unreachable\n",
+ libcfs_nid2str(sd->sd_dst_nid));
+ return -EHOSTUNREACH;
+ }
+
+ /* We're attempting to round robin over the remote peer
+ * NI's so update the final destination we selected
+ */
+ sd->sd_final_dst_lpni = sd->sd_best_lpni;
- return LNET_DC_WAIT;
-}
+ /* Increment the sequence number of the remote lpni so
+ * we can round robin over the different interfaces of
+ * the remote lpni
+ */
+ sd->sd_best_lpni->lpni_seq++;
+ }
-static int
-lnet_handle_find_routed_path(struct lnet_send_data *sd,
- lnet_nid_t dst_nid,
- struct lnet_peer_ni **gw_lpni,
- struct lnet_peer **gw_peer)
-{
- struct lnet_route *best_route = NULL;
- struct lnet_route *last_route = NULL;
- struct lnet_peer_ni *gw;
- lnet_nid_t src_nid = sd->sd_src_nid;
-
- gw = lnet_find_route_locked(NULL, LNET_NIDNET(dst_nid),
- sd->sd_rtr_nid, &best_route, &last_route);
- if (!gw) {
- CERROR("no route to %s from %s\n",
- libcfs_nid2str(dst_nid), libcfs_nid2str(src_nid));
- return -EHOSTUNREACH;
- }
+ /*
+ * find the best route. Restrict the selection on the net of the
+ * local NI if we've already picked the local NI to send from.
+ * Otherwise, let's pick any route we can find and then find
+ * a local NI we can reach the route's gateway on. Any route we
+ * select will be reachable by virtue of the restriction we have
+ * when adding a route.
+ */
+ best_route = lnet_find_route_locked(best_rnet,
+ LNET_NIDNET(src_nid),
+ sd->sd_best_lpni,
+ &last_route, &gwni);
+
+ if (!best_route) {
+ CERROR("no route to %s from %s\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(src_nid));
+ return -EHOSTUNREACH;
+ }
- /* get the peer of the gw_ni */
- LASSERT(gw->lpni_peer_net);
- LASSERT(gw->lpni_peer_net->lpn_peer);
+ if (!gwni) {
+ CERROR("Internal Error. Route expected to %s from %s\n",
+ libcfs_nid2str(dst_nid),
+ libcfs_nid2str(src_nid));
+ return -EFAULT;
+ }
- *gw_peer = gw->lpni_peer_net->lpn_peer;
+ gw = best_route->lr_gateway;
+ LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
+ local_lnet = best_route->lr_lnet;
+ }
/*
* Discover this gateway if it hasn't already been discovered.
* This means we might delay the message until discovery has
* completed
*/
- if (lnet_msg_discovery(sd->sd_msg) &&
- !lnet_peer_is_uptodate(*gw_peer)) {
- sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
- return lnet_initiate_peer_discovery(gw, sd->sd_msg,
- sd->sd_rtr_nid, sd->sd_cpt);
- }
+ sd->sd_msg->msg_src_nid_param = sd->sd_src_nid;
+ rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
+ if (rc)
+ return rc;
if (!sd->sd_best_ni)
- sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, *gw_peer,
- gw->lpni_peer_net,
- sd->sd_md_cpt,
- true);
+ sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw,
+ lnet_peer_get_net_locked(gw,
+ local_lnet),
+ sd->sd_md_cpt);
if (!sd->sd_best_ni) {
- CERROR("Internal Error. Expected local ni on %s "
- "but non found :%s\n",
- libcfs_net2str(gw->lpni_peer_net->lpn_net_id),
+ CERROR("Internal Error. Expected local ni on %s but non found :%s\n",
+ libcfs_net2str(local_lnet),
libcfs_nid2str(sd->sd_src_nid));
return -EFAULT;
}
- /*
- * if gw is MR let's find its best peer_ni
- */
- if (lnet_peer_is_multi_rail(*gw_peer)) {
- gw = lnet_find_best_lpni_on_net(sd, *gw_peer,
- sd->sd_best_ni->ni_net->net_id);
- /*
- * We've already verified that the gw has an NI on that
- * desired net, but we're not finding it. Something is
- * wrong.
- */
- if (!gw) {
- CERROR("Internal Error. Route expected to %s from %s\n",
- libcfs_nid2str(dst_nid),
- libcfs_nid2str(src_nid));
- return -EFAULT;
- }
- }
-
- *gw_lpni = gw;
+ *gw_lpni = gwni;
+ *gw_peer = gw;
/*
- * increment the route sequence number since now we're sure we're
- * going to use it
+ * increment the sequence numbers since now we're sure we're
+ * going to use this path
*/
- LASSERT(best_route && last_route);
- best_route->lr_seq = last_route->lr_seq + 1;
+ if (sd->sd_rtr_nid == LNET_NID_ANY) {
+ LASSERT(best_route && last_route);
+ best_route->lr_seq = last_route->lr_seq + 1;
+ if (best_lpn)
+ best_lpn->lpn_seq++;
+ }
return 0;
}
if (sd->sd_send_case & NMR_DST)
/*
- * since the final destination is non-MR let's set its preferred
- * NID before we send
- */
- lnet_set_non_mr_pref_nid(sd);
+ * since the final destination is non-MR let's set its preferred
+ * NID before we send
+ */
+ lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
+ sd->sd_msg);
/*
* We're going to send to the gw found so let's set its
}
struct lnet_ni *
-lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt)
+lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
+ bool discovery)
{
- struct lnet_peer_net *peer_net = NULL;
+ struct lnet_peer_net *lpn = NULL;
+ struct lnet_peer_net *best_lpn = NULL;
+ struct lnet_net *net = NULL;
+ struct lnet_net *best_net = NULL;
struct lnet_ni *best_ni = NULL;
+ int best_lpn_healthv = 0;
+ int best_net_healthv = 0;
+ int net_healthv;
+ __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
+ __u32 lpn_sel_prio;
+ __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
+ __u32 net_sel_prio;
+ bool exit = false;
/*
* The peer can have multiple interfaces, some of them can be on
*/
/* go through all the peer nets and find the best_ni */
- list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_peer_nets) {
+ list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
/*
* The peer's list of nets can contain non-local nets. We
* want to only examine the local ones.
*/
- if (!lnet_get_net_locked(peer_net->lpn_net_id))
+ net = lnet_get_net_locked(lpn->lpn_net_id);
+ if (!net)
+ continue;
+
+ lpn_sel_prio = lpn->lpn_sel_priority;
+ net_healthv = lnet_get_net_healthv_locked(net);
+ net_sel_prio = net->net_sel_priority;
+
+ /*
+ * if this is a discovery message and lp_disc_net_id is
+ * specified then use that net to send the discovery on.
+ */
+ if (peer->lp_disc_net_id == lpn->lpn_net_id &&
+ discovery) {
+ exit = true;
+ goto select_lpn;
+ }
+
+ if (!best_lpn)
+ goto select_lpn;
+
+ /* always select the lpn with the best health */
+ if (best_lpn_healthv > lpn->lpn_healthv)
+ continue;
+ else if (best_lpn_healthv < lpn->lpn_healthv)
+ goto select_lpn;
+
+ /* select the preferred peer and local nets */
+ if (best_lpn_sel_prio < lpn_sel_prio)
+ continue;
+ else if (best_lpn_sel_prio > lpn_sel_prio)
+ goto select_lpn;
+
+ if (best_net_healthv > net_healthv)
+ continue;
+ else if (best_net_healthv < net_healthv)
+ goto select_lpn;
+
+ if (best_net_sel_prio < net_sel_prio)
+ continue;
+ else if (best_net_sel_prio > net_sel_prio)
+ goto select_lpn;
+
+ if (best_lpn->lpn_seq < lpn->lpn_seq)
+ continue;
+ else if (best_lpn->lpn_seq > lpn->lpn_seq)
+ goto select_lpn;
+
+ /* round robin over the local networks */
+ if (best_net->net_seq <= net->net_seq)
continue;
- best_ni = lnet_find_best_ni_on_spec_net(best_ni, peer,
- peer_net, md_cpt, false);
+
+select_lpn:
+ best_net_healthv = net_healthv;
+ best_net_sel_prio = net_sel_prio;
+ best_lpn_healthv = lpn->lpn_healthv;
+ best_lpn_sel_prio = lpn_sel_prio;
+ best_lpn = lpn;
+ best_net = net;
+
+ if (exit)
+ break;
}
- if (best_ni)
- /* increment sequence number so we can round robin */
- best_ni->ni_seq++;
+ if (best_lpn) {
+ /* Select the best NI on the same net as best_lpn chosen
+ * above
+ */
+ best_ni = lnet_find_best_ni_on_spec_net(NULL, peer,
+ best_lpn, md_cpt);
+ }
return best_ni;
}
static struct lnet_ni *
-lnet_find_existing_preferred_best_ni(struct lnet_send_data *sd)
+lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
{
struct lnet_ni *best_ni = NULL;
- struct lnet_peer_net *peer_net;
- struct lnet_peer *peer = sd->sd_peer;
- struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
- struct lnet_peer_ni *lpni;
- int cpt = sd->sd_cpt;
+ struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
+ struct lnet_peer_ni *lpni_entry;
/*
* We must use a consistent source address when sending to a
* So we need to pick the NI the peer prefers for this
* particular network.
*/
-
- /* Get the target peer_ni */
- peer_net = lnet_peer_get_net_locked(peer,
- LNET_NIDNET(best_lpni->lpni_nid));
- LASSERT(peer_net != NULL);
- list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
- lpni_peer_nis) {
- if (lpni->lpni_pref_nnids == 0)
+ LASSERT(peer_net);
+ list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
+ lpni_peer_nis) {
+ if (lpni_entry->lpni_pref_nnids == 0)
continue;
- LASSERT(lpni->lpni_pref_nnids == 1);
- best_ni = lnet_nid2ni_locked(
- lpni->lpni_pref.nid, cpt);
+ LASSERT(lpni_entry->lpni_pref_nnids == 1);
+ best_ni = lnet_nid2ni_locked(lpni_entry->lpni_pref.nid, cpt);
break;
}
* particular network.
*/
- best_ni = lnet_find_existing_preferred_best_ni(sd);
+ best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
+ sd->sd_cpt);
/* if best_ni is still not set just pick one */
if (!best_ni) {
best_ni =
lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
sd->sd_best_lpni->lpni_peer_net,
- sd->sd_md_cpt, true);
+ sd->sd_md_cpt);
/* If there is no best_ni we don't have a route */
if (!best_ni) {
CERROR("no path to %s from net %s\n",
sd->sd_best_ni = best_ni;
/* Set preferred NI if necessary. */
- lnet_set_non_mr_pref_nid(sd);
+ lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
return 0;
}
static int
lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
{
- int rc;
+ int rc = 0;
/* sd->sd_best_lpni is already set to the final destination */
return -EFAULT;
}
- rc = lnet_select_preferred_best_ni(sd);
+ if (sd->sd_msg->msg_routing) {
+ /* If I'm forwarding this message then I can choose any NI
+ * on the destination peer net
+ */
+ sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
+ sd->sd_peer,
+ sd->sd_best_lpni->lpni_peer_net,
+ sd->sd_md_cpt);
+ if (!sd->sd_best_ni) {
+ CERROR("Unable to forward message to %s. No local NI available\n",
+ libcfs_nid2str(sd->sd_dst_nid));
+ rc = -EHOSTUNREACH;
+ }
+ } else
+ rc = lnet_select_preferred_best_ni(sd);
+
if (!rc)
rc = lnet_handle_send(sd);
sd->sd_best_ni =
lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
sd->sd_best_lpni->lpni_peer_net,
- sd->sd_md_cpt, true);
+ sd->sd_md_cpt);
if (!sd->sd_best_ni) {
/*
* networks.
*/
sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
- sd->sd_md_cpt);
+ sd->sd_md_cpt,
+ lnet_msg_discovery(sd->sd_msg));
if (sd->sd_best_ni) {
sd->sd_best_lpni =
- lnet_find_best_lpni_on_net(sd, sd->sd_peer,
- sd->sd_best_ni->ni_net->net_id);
+ lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
+ sd->sd_peer,
+ sd->sd_best_ni->ni_net->net_id);
/*
* if we're successful in selecting a peer_ni on the local
return rc;
/*
- * TODO; One possible enhancement is to run the selection
- * algorithm on the peer. However for remote peers the credits are
- * not decremented, so we'll be basically going over the peer NIs
- * in round robin. An MR router will run the selection algorithm
- * on the next-hop interfaces.
+ * Now that we must route to the destination, we must consider the
+ * MR case, where the destination has multiple interfaces, some of
+ * which we can route to and others we do not. For this reason we
+ * need to select the destination which we can route to and if
+ * there are multiple, we need to round robin.
*/
rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
&gw_peer);
struct lnet_peer *gw_peer = NULL;
/*
- * Let's set if we have a preferred NI to talk to this NMR peer
+ * Let's see if we have a preferred NI to talk to this NMR peer
*/
- sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd);
+ sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
+ sd->sd_cpt);
/*
* find the router and that'll find the best NI if we didn't find
* set the best_ni we've chosen as the preferred one for
* this peer
*/
- lnet_set_non_mr_pref_nid(sd);
+ lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
/* we'll be sending to the gw */
sd->sd_best_lpni = gw_lpni;
lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
struct lnet_msg *msg, lnet_nid_t rtr_nid)
{
- struct lnet_peer_ni *lpni;
- struct lnet_peer *peer;
- struct lnet_send_data send_data;
- int cpt, rc;
- int md_cpt;
- __u32 send_case = 0;
+ struct lnet_peer_ni *lpni;
+ struct lnet_peer *peer;
+ struct lnet_send_data send_data;
+ int cpt, rc;
+ int md_cpt;
+ __u32 send_case = 0;
+ bool final_hop;
+ bool mr_forwarding_allowed;
memset(&send_data, 0, sizeof(send_data));
*/
send_data.sd_msg = msg;
send_data.sd_cpt = cpt;
- if (LNET_NETTYP(LNET_NIDNET(dst_nid)) == LOLND) {
+ if (dst_nid == LNET_NID_LO_0) {
rc = lnet_handle_lo_send(&send_data);
lnet_net_unlock(cpt);
return rc;
}
/*
- * Cache the original src_nid. If we need to resend the message
- * then we'll need to know whether the src_nid was originally
+ * Cache the original src_nid and rtr_nid. If we need to resend the
+ * message then we'll need to know whether the src_nid was originally
* specified for this message. If it was originally specified,
* then we need to keep using the same src_nid since it's
- * continuing the same sequence of messages.
+ * continuing the same sequence of messages. Similarly, rtr_nid will
+ * affect our choice of next hop.
*/
msg->msg_src_nid_param = src_nid;
+ msg->msg_rtr_nid_param = rtr_nid;
/*
- * Now that we have a peer_ni, check if we want to discover
- * the peer. Traffic to the LNET_RESERVED_PORTAL should not
- * trigger discovery.
+ * If necessary, perform discovery on the peer that owns this peer_ni.
+ * Note, this can result in the ownership of this peer_ni changing
+ * to another peer object.
*/
- peer = lpni->lpni_peer_net->lpn_peer;
- if (lnet_msg_discovery(msg) && !lnet_peer_is_uptodate(peer)) {
- rc = lnet_initiate_peer_discovery(lpni, msg, rtr_nid, cpt);
+ rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
+ if (rc) {
lnet_peer_ni_decref_locked(lpni);
lnet_net_unlock(cpt);
return rc;
}
lnet_peer_ni_decref_locked(lpni);
+ peer = lpni->lpni_peer_net->lpn_peer;
+
/*
* Identify the different send cases
*/
else
send_case |= REMOTE_DST;
+ final_hop = false;
+ if (msg->msg_routing && (send_case & LOCAL_DST))
+ final_hop = true;
+
+ /* Determine whether to allow MR forwarding for this message.
+ * NB: MR forwarding is allowed if the message originator and the
+ * destination are both MR capable, and the destination lpni that was
+ * originally chosen by the originator is unhealthy or down.
+ * We check the MR capability of the destination further below
+ */
+ mr_forwarding_allowed = false;
+ if (final_hop) {
+ struct lnet_peer *src_lp;
+ struct lnet_peer_ni *src_lpni;
+
+ src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
+ LNET_NID_ANY, cpt);
+ /* We don't fail the send if we hit any errors here. We'll just
+ * try to send it via non-multi-rail criteria
+ */
+ if (!IS_ERR(src_lpni)) {
+ /* Drop ref taken by lnet_nid2peerni_locked() */
+ lnet_peer_ni_decref_locked(src_lpni);
+ src_lp = lpni->lpni_peer_net->lpn_peer;
+ if (lnet_peer_is_multi_rail(src_lp) &&
+ !lnet_is_peer_ni_alive(lpni))
+ mr_forwarding_allowed = true;
+
+ }
+ CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
+ mr_forwarding_allowed ? "allowed" : "not allowed");
+ }
+
/*
- * if this is a non-MR peer or if we're recovering a peer ni then
- * let's consider this an NMR case so we can hit the destination
- * NID.
+ * Deal with the peer as NMR in the following cases:
+ * 1. the peer is NMR
+ * 2. We're trying to recover a specific peer NI
+ * 3. I'm a router sending to the final destination and MR forwarding is
+ * not allowed for this message (as determined above).
+ * In this case the source of the message would've
+ * already selected the final destination so my job
+ * is to honor the selection.
*/
- if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery)
+ if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
+ (final_hop && !mr_forwarding_allowed))
send_case |= NMR_DST;
else
send_case |= MR_DST;
- if (msg->msg_type == LNET_MSG_REPLY ||
- msg->msg_type == LNET_MSG_ACK)
+ if (lnet_msg_is_response(msg))
send_case |= SND_RESP;
/* assign parameters to the send_data */
LASSERT(!msg->msg_tx_committed);
rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
- if (rc < 0)
+ if (rc < 0) {
+ if (rc == -EHOSTUNREACH)
+ msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
+ else
+ msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
return rc;
+ }
if (rc == LNET_CREDIT_OK)
lnet_ni_send(msg->msg_txni, msg);
return;
rspt = md->md_rspt_ptr;
- md->md_rspt_ptr = NULL;
/* debug code */
LASSERT(rspt->rspt_cpt == cpt);
- /*
- * invalidate the handle to indicate that a response has been
- * received, which will then lead the monitor thread to clean up
- * the rspt block.
- */
- LNetInvalidateMDHandle(&rspt->rspt_mdh);
+ md->md_rspt_ptr = NULL;
+
+ if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
+ /*
+ * The monitor thread has invalidated this handle because the
+ * response timed out, but it failed to lookup the MD. That
+ * means this response tracker is on the zombie list. We can
+ * safely remove it under the resource lock (held by caller) and
+ * free the response tracker block.
+ */
+ list_del(&rspt->rspt_on_list);
+ lnet_rspt_free(rspt, cpt);
+ } else {
+ /*
+ * invalidate the handle to indicate that a response has been
+ * received, which will then lead the monitor thread to clean up
+ * the rspt block.
+ */
+ LNetInvalidateMDHandle(&rspt->rspt_mdh);
+ }
+}
+
+void
+lnet_clean_zombie_rstqs(void)
+{
+ struct lnet_rsp_tracker *rspt, *tmp;
+ int i;
+
+ cfs_cpt_for_each(i, lnet_cpt_table()) {
+ list_for_each_entry_safe(rspt, tmp,
+ the_lnet.ln_mt_zombie_rstqs[i],
+ rspt_on_list) {
+ list_del(&rspt->rspt_on_list);
+ lnet_rspt_free(rspt, i);
+ }
+ }
+
+ cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
}
static void
-lnet_finalize_expired_responses(bool force)
+lnet_finalize_expired_responses(void)
{
struct lnet_libmd *md;
- struct list_head local_queue;
struct lnet_rsp_tracker *rspt, *tmp;
+ ktime_t now;
int i;
if (the_lnet.ln_mt_rstq == NULL)
return;
cfs_cpt_for_each(i, lnet_cpt_table()) {
- INIT_LIST_HEAD(&local_queue);
+ LIST_HEAD(local_queue);
lnet_net_lock(i);
if (!the_lnet.ln_mt_rstq[i]) {
list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
lnet_net_unlock(i);
+ now = ktime_get();
+
list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
/*
* The rspt mdh will be invalidated when a response
lnet_res_lock(i);
if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
lnet_res_unlock(i);
- list_del_init(&rspt->rspt_on_list);
+ list_del(&rspt->rspt_on_list);
lnet_rspt_free(rspt, i);
continue;
}
- if (ktime_compare(ktime_get(), rspt->rspt_deadline) >= 0 ||
- force) {
+ if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
+ the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
struct lnet_peer_ni *lpni;
lnet_nid_t nid;
md = lnet_handle2md(&rspt->rspt_mdh);
if (!md) {
+ /* MD has been queued for unlink, but
+ * rspt hasn't been detached (Note we've
+ * checked above that the rspt_mdh is
+ * valid). Since we cannot lookup the MD
+ * we're unable to detach the rspt
+ * ourselves. Thus, move the rspt to the
+ * zombie list where we'll wait for
+ * either:
+ * 1. The remaining operations on the
+ * MD to complete. In this case the
+ * final operation will result in
+ * lnet_msg_detach_md()->
+ * lnet_detach_rsp_tracker() where
+ * we will clean up this response
+ * tracker.
+ * 2. LNet to shutdown. In this case
+ * we'll wait until after all LND Nets
+ * have shutdown and then we can
+ * safely free any remaining response
+ * tracker blocks on the zombie list.
+ * Note: We need to hold the resource
+ * lock when adding to the zombie list
+ * because we may have concurrent access
+ * with lnet_detach_rsp_tracker().
+ */
LNetInvalidateMDHandle(&rspt->rspt_mdh);
+ list_move(&rspt->rspt_on_list,
+ the_lnet.ln_mt_zombie_rstqs[i]);
lnet_res_unlock(i);
- list_del_init(&rspt->rspt_on_list);
- lnet_rspt_free(rspt, i);
continue;
}
LASSERT(md->md_rspt_ptr == rspt);
md->md_rspt_ptr = NULL;
lnet_res_unlock(i);
- lnet_net_lock(i);
- the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
- lnet_net_unlock(i);
-
- list_del_init(&rspt->rspt_on_list);
+ LNetMDUnlink(rspt->rspt_mdh);
nid = rspt->rspt_next_hop_nid;
- CNETERR("Response timed out: md = %p: nid = %s\n",
- md, libcfs_nid2str(nid));
- LNetMDUnlink(rspt->rspt_mdh);
+ list_del(&rspt->rspt_on_list);
lnet_rspt_free(rspt, i);
+ /* If we're shutting down we just want to clean
+ * up the rspt blocks
+ */
+ if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
+ continue;
+
+ lnet_net_lock(i);
+ the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
+ lnet_net_unlock(i);
+
+ CDEBUG(D_NET,
+ "Response timeout: md = %p: nid = %s\n",
+ md, libcfs_nid2str(nid));
+
/*
* If there is a timeout on the response
* from the next hop decrement its health
}
}
- lnet_net_lock(i);
- if (!list_empty(&local_queue))
+ if (!list_empty(&local_queue)) {
+ lnet_net_lock(i);
list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
- lnet_net_unlock(i);
+ lnet_net_unlock(i);
+ }
}
}
lnet_finalize(msg, -EFAULT);
lnet_net_lock(cpt);
} else {
- struct lnet_peer *peer;
int rc;
- lnet_nid_t src_nid = LNET_NID_ANY;
-
- /*
- * if this message is not being routed and the
- * peer is non-MR then we must use the same
- * src_nid that was used in the original send.
- * Otherwise if we're routing the message (IE
- * we're a router) then we can use any of our
- * local interfaces. It doesn't matter to the
- * final destination.
- */
- peer = lpni->lpni_peer_net->lpn_peer;
- if (!msg->msg_routing &&
- !lnet_peer_is_multi_rail(peer))
- src_nid = le64_to_cpu(msg->msg_hdr.src_nid);
- /*
- * If we originally specified a src NID, then we
- * must attempt to reuse it in the resend as well.
- */
- if (msg->msg_src_nid_param != LNET_NID_ANY)
- src_nid = msg->msg_src_nid_param;
lnet_peer_ni_decref_locked(lpni);
lnet_net_unlock(cpt);
CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
- libcfs_nid2str(src_nid),
+ libcfs_nid2str(msg->msg_src_nid_param),
libcfs_id2str(msg->msg_target),
lnet_msgtyp2str(msg->msg_type),
msg->msg_recovery,
msg->msg_retry_count);
- rc = lnet_send(src_nid, msg, LNET_NID_ANY);
+ rc = lnet_send(msg->msg_src_nid_param, msg,
+ msg->msg_rtr_nid_param);
if (rc) {
CERROR("Error sending %s to %s: %d\n",
lnet_msgtyp2str(msg->msg_type),
lnet_recover_local_nis(void)
{
struct lnet_mt_event_info *ev_info;
- struct list_head processed_list;
- struct list_head local_queue;
+ LIST_HEAD(processed_list);
+ LIST_HEAD(local_queue);
struct lnet_handle_md mdh;
struct lnet_ni *tmp;
struct lnet_ni *ni;
int healthv;
int rc;
- INIT_LIST_HEAD(&local_queue);
- INIT_LIST_HEAD(&processed_list);
-
/*
* splice the recovery queue on a local queue. We will iterate
* through the local queue and update it as needed. Once we're
ev_info->mt_type = MT_TYPE_LOCAL_NI;
ev_info->mt_nid = nid;
rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
- ev_info, the_lnet.ln_mt_eqh, true);
+ ev_info, the_lnet.ln_mt_handler,
+ true);
/* lookup the nid again */
lnet_net_lock(0);
ni = lnet_nid2ni_locked(nid, 0);
lnet_net_unlock(0);
}
-static struct list_head **
-lnet_create_array_of_queues(void)
-{
- struct list_head **qs;
- struct list_head *q;
- int i;
-
- qs = cfs_percpt_alloc(lnet_cpt_table(),
- sizeof(struct list_head));
- if (!qs) {
- CERROR("Failed to allocate queues\n");
- return NULL;
- }
-
- cfs_percpt_for_each(q, i, qs)
- INIT_LIST_HEAD(q);
-
- return qs;
-}
-
static int
lnet_resendqs_create(void)
{
lnet_clean_resendqs(void)
{
struct lnet_msg *msg, *tmp;
- struct list_head msgs;
+ LIST_HEAD(msgs);
int i;
- INIT_LIST_HEAD(&msgs);
-
cfs_cpt_for_each(i, lnet_cpt_table()) {
lnet_net_lock(i);
list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
lnet_recover_peer_nis(void)
{
struct lnet_mt_event_info *ev_info;
- struct list_head processed_list;
- struct list_head local_queue;
+ LIST_HEAD(processed_list);
+ LIST_HEAD(local_queue);
struct lnet_handle_md mdh;
struct lnet_peer_ni *lpni;
struct lnet_peer_ni *tmp;
lnet_nid_t nid;
int healthv;
int rc;
-
- INIT_LIST_HEAD(&local_queue);
- INIT_LIST_HEAD(&processed_list);
+ time64_t now;
/*
* Always use cpt 0 for locking across all interactions with
&local_queue);
lnet_net_unlock(0);
+ now = ktime_get_seconds();
+
list_for_each_entry_safe(lpni, tmp, &local_queue,
lpni_recovery) {
/*
ev_info->mt_type = MT_TYPE_PEER_NI;
ev_info->mt_nid = nid;
rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
- ev_info, the_lnet.ln_mt_eqh, true);
+ ev_info, the_lnet.ln_mt_handler,
+ true);
lnet_net_lock(0);
/*
* lnet_find_peer_ni_locked() grabs a refcount for
}
lpni->lpni_recovery_ping_mdh = mdh;
- /*
- * While we're unlocked the lpni could've been
- * readded on the recovery queue. In this case we
- * don't need to add it to the local queue, since
- * it's already on there and the thread that added
- * it would've incremented the refcount on the
- * peer, which means we need to decref the refcount
- * that was implicitly grabbed by find_peer_ni_locked.
- * Otherwise, if the lpni is still not on
- * the recovery queue, then we'll add it to the
- * processed list.
- */
- if (list_empty(&lpni->lpni_recovery))
- list_add_tail(&lpni->lpni_recovery, &processed_list);
- else
- lnet_peer_ni_decref_locked(lpni);
- lnet_net_unlock(0);
- spin_lock(&lpni->lpni_lock);
- if (rc)
+ lnet_peer_ni_add_to_recoveryq_locked(lpni,
+ &processed_list,
+ now);
+ if (rc) {
+ spin_lock(&lpni->lpni_lock);
lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
- }
- spin_unlock(&lpni->lpni_lock);
+ spin_unlock(&lpni->lpni_lock);
+ }
+
+ /* Drop the ref taken by lnet_find_peer_ni_locked() */
+ lnet_peer_ni_decref_locked(lpni);
+ lnet_net_unlock(0);
+ } else
+ spin_unlock(&lpni->lpni_lock);
}
list_splice_init(&processed_list, &local_queue);
int interval;
time64_t now;
+ wait_for_completion(&the_lnet.ln_started);
/*
* The monitor thread takes care of the following:
* 1. Checks the aliveness of routers
* 4. Checks if there are any NIs on the remote recovery queue
* and pings them.
*/
- cfs_block_allsigs();
-
while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
now = ktime_get_real_seconds();
lnet_resend_pending_msgs();
if (now >= rsp_timeout) {
- lnet_finalize_expired_responses(false);
+ lnet_finalize_expired_responses();
rsp_timeout = now + (lnet_transaction_timeout / 2);
}
* if we wake up every 1 second? Although, we've seen
* cases where we get a complaint that an idle thread
* is waking up unnecessarily.
+ *
+ * Take into account the current net_count when you wake
+ * up for alive router checking, since we need to check
+ * possibly as many networks as we have configured.
*/
interval = min(lnet_recovery_interval,
- lnet_transaction_timeout / 2);
- wait_event_interruptible_timeout(the_lnet.ln_mt_waitq,
- false,
- cfs_time_seconds(interval));
+ min((unsigned int) alive_router_check_interval /
+ lnet_current_net_count,
+ lnet_transaction_timeout / 2));
+ wait_for_completion_interruptible_timeout(
+ &the_lnet.ln_mt_wait_complete,
+ cfs_time_seconds(interval));
+ /* Must re-init the completion before testing anything,
+ * including ln_mt_state.
+ */
+ reinit_completion(&the_lnet.ln_mt_wait_complete);
}
- /* clean up the router checker */
- lnet_prune_rc_data(1);
-
/* Shutting down */
lnet_net_lock(LNET_LOCK_EX);
the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
int
lnet_send_ping(lnet_nid_t dest_nid,
struct lnet_handle_md *mdh, int nnis,
- void *user_data, struct lnet_handle_eq eqh, bool recovery)
+ void *user_data, lnet_handler_t handler, bool recovery)
{
struct lnet_md md = { NULL };
struct lnet_process_id id;
md.length = LNET_PING_INFO_SIZE(nnis);
md.threshold = 2; /* GET/REPLY */
md.max_size = 0;
- md.options = LNET_MD_TRUNCATE;
+ md.options = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
md.user_ptr = user_data;
- md.eq_handle = eqh;
+ md.handler = handler;
- rc = LNetMDBind(md, LNET_UNLINK, mdh);
+ rc = LNetMDBind(&md, LNET_UNLINK, mdh);
if (rc) {
lnet_ping_buffer_decref(pbuf);
CERROR("Can't bind MD: %d\n", rc);
static void
lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
- int status, bool unlink_event)
+ int status, bool send, bool unlink_event)
{
lnet_nid_t nid = ev_info->mt_nid;
return;
}
lnet_ni_lock(ni);
- ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
+ if (!send || (send && status != 0))
+ ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
if (status)
ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
lnet_ni_unlock(ni);
* In the peer case, it'll naturally be incremented
*/
if (!unlink_event)
- lnet_inc_healthv(&ni->ni_healthv);
+ lnet_inc_healthv(&ni->ni_healthv,
+ lnet_health_sensitivity);
} else {
struct lnet_peer_ni *lpni;
int cpt;
return;
}
spin_lock(&lpni->lpni_lock);
- lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
+ if (!send || (send && status != 0))
+ lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
if (status)
lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
spin_unlock(&lpni->lpni_lock);
void
lnet_mt_event_handler(struct lnet_event *event)
{
- struct lnet_mt_event_info *ev_info = event->md.user_ptr;
+ struct lnet_mt_event_info *ev_info = event->md_user_ptr;
struct lnet_ping_buffer *pbuf;
/* TODO: remove assert */
case LNET_EVENT_UNLINK:
CDEBUG(D_NET, "%s recovery ping unlinked\n",
libcfs_nid2str(ev_info->mt_nid));
+ /* fallthrough */
case LNET_EVENT_REPLY:
- lnet_handle_recovery_reply(ev_info, event->status,
+ lnet_handle_recovery_reply(ev_info, event->status, false,
event->type == LNET_EVENT_UNLINK);
break;
case LNET_EVENT_SEND:
libcfs_nid2str(ev_info->mt_nid),
(event->status) ? "unsuccessfully" :
"successfully", event->status);
+ lnet_handle_recovery_reply(ev_info, event->status, true, false);
break;
default:
CERROR("Unexpected event: %d\n", event->type);
}
if (event->unlinked) {
LIBCFS_FREE(ev_info, sizeof(*ev_info));
- pbuf = LNET_PING_INFO_TO_BUFFER(event->md.start);
+ pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
lnet_ping_buffer_decref(pbuf);
}
}
static void
lnet_rsp_tracker_clean(void)
{
- lnet_finalize_expired_responses(true);
+ lnet_finalize_expired_responses();
cfs_percpt_free(the_lnet.ln_mt_rstq);
the_lnet.ln_mt_rstq = NULL;
if (rc)
goto clean_queues;
- /* Pre monitor thread start processing */
- rc = lnet_router_pre_mt_start();
- if (rc)
- goto free_mem;
-
sema_init(&the_lnet.ln_mt_signal, 0);
lnet_net_lock(LNET_LOCK_EX);
goto clean_thread;
}
- /* post monitor thread start processing */
- lnet_router_post_mt_start();
-
return 0;
clean_thread:
/* block until event callback signals exit */
down(&the_lnet.ln_mt_signal);
/* clean up */
- lnet_router_cleanup();
-free_mem:
lnet_net_lock(LNET_LOCK_EX);
the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
lnet_net_unlock(LNET_LOCK_EX);
lnet_clean_local_ni_recoveryq();
lnet_clean_peer_ni_recoveryq();
lnet_clean_resendqs();
- LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
+ the_lnet.ln_mt_handler = NULL;
return rc;
clean_queues:
lnet_rsp_tracker_clean();
lnet_net_unlock(LNET_LOCK_EX);
/* tell the monitor thread that we're shutting down */
- wake_up(&the_lnet.ln_mt_waitq);
+ complete(&the_lnet.ln_mt_wait_complete);
/* block until monitor thread signals that it's done */
down(&the_lnet.ln_mt_signal);
LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
/* perform cleanup tasks */
- lnet_router_cleanup();
lnet_rsp_tracker_clean();
lnet_clean_local_ni_recoveryq();
lnet_clean_peer_ni_recoveryq();
lnet_clean_resendqs();
-
- return;
}
void
struct lnet_hdr *hdr = &msg->msg_hdr;
struct lnet_process_id src = {0};
struct lnet_libmd *md;
- int rlength;
- int mlength;
+ unsigned int rlength;
+ unsigned int mlength;
int cpt;
cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
LASSERT(md->md_offset == 0);
rlength = hdr->payload_length;
- mlength = MIN(rlength, (int)md->md_length);
+ mlength = min(rlength, md->md_length);
if (mlength < rlength &&
(md->md_options & LNET_MD_TRUNCATE) == 0) {
}
}
-void
-lnet_print_hdr(struct lnet_hdr *hdr)
-{
- struct lnet_process_id src = {
- .nid = hdr->src_nid,
- .pid = hdr->src_pid,
- };
- struct lnet_process_id dst = {
- .nid = hdr->dest_nid,
- .pid = hdr->dest_pid,
- };
- char *type_str = lnet_msgtyp2str(hdr->type);
-
- CWARN("P3 Header at %p of type %s\n", hdr, type_str);
- CWARN(" From %s\n", libcfs_id2str(src));
- CWARN(" To %s\n", libcfs_id2str(dst));
-
- switch (hdr->type) {
- default:
- break;
-
- case LNET_MSG_PUT:
- CWARN(" Ptl index %d, ack md %#llx.%#llx, "
- "match bits %llu\n",
- hdr->msg.put.ptl_index,
- hdr->msg.put.ack_wmd.wh_interface_cookie,
- hdr->msg.put.ack_wmd.wh_object_cookie,
- hdr->msg.put.match_bits);
- CWARN(" Length %d, offset %d, hdr data %#llx\n",
- hdr->payload_length, hdr->msg.put.offset,
- hdr->msg.put.hdr_data);
- break;
-
- case LNET_MSG_GET:
- CWARN(" Ptl index %d, return md %#llx.%#llx, "
- "match bits %llu\n", hdr->msg.get.ptl_index,
- hdr->msg.get.return_wmd.wh_interface_cookie,
- hdr->msg.get.return_wmd.wh_object_cookie,
- hdr->msg.get.match_bits);
- CWARN(" Length %d, src offset %d\n",
- hdr->msg.get.sink_length,
- hdr->msg.get.src_offset);
- break;
-
- case LNET_MSG_ACK:
- CWARN(" dst md %#llx.%#llx, "
- "manipulated length %d\n",
- hdr->msg.ack.dst_wmd.wh_interface_cookie,
- hdr->msg.ack.dst_wmd.wh_object_cookie,
- hdr->msg.ack.mlength);
- break;
-
- case LNET_MSG_REPLY:
- CWARN(" dst md %#llx.%#llx, "
- "length %d\n",
- hdr->msg.reply.dst_wmd.wh_interface_cookie,
- hdr->msg.reply.dst_wmd.wh_object_cookie,
- hdr->payload_length);
- }
-
-}
-
int
lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
void *private, int rdma_req)
{
- int rc = 0;
- int cpt;
- int for_me;
- struct lnet_msg *msg;
- lnet_pid_t dest_pid;
- lnet_nid_t dest_nid;
- lnet_nid_t src_nid;
struct lnet_peer_ni *lpni;
- __u32 payload_length;
- __u32 type;
+ struct lnet_msg *msg;
+ __u32 payload_length;
+ lnet_pid_t dest_pid;
+ lnet_nid_t dest_nid;
+ lnet_nid_t src_nid;
+ bool push = false;
+ int for_me;
+ __u32 type;
+ int rc = 0;
+ int cpt;
LASSERT (!in_interrupt ());
}
if (the_lnet.ln_routing &&
- ni->ni_last_alive != ktime_get_real_seconds()) {
- /* NB: so far here is the only place to set NI status to "up */
+ ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
lnet_ni_lock(ni);
- ni->ni_last_alive = ktime_get_real_seconds();
- if (ni->ni_status != NULL &&
- ni->ni_status->ns_status == LNET_NI_STATUS_DOWN)
- ni->ni_status->ns_status = LNET_NI_STATUS_UP;
+ spin_lock(&ni->ni_net->net_lock);
+ ni->ni_net->net_last_alive = ktime_get_real_seconds();
+ spin_unlock(&ni->ni_net->net_lock);
+ push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
lnet_ni_unlock(ni);
}
+ if (push)
+ lnet_push_update_to_peers(1);
+
/* Regard a bad destination NID as a protocol error. Senders should
* know what they're doing; if they don't they're misconfigured, buggy
* or malicious so we chop them off at the knees :) */
}
if (!list_empty(&the_lnet.ln_drop_rules) &&
- lnet_drop_rule_match(hdr, NULL)) {
- CDEBUG(D_NET, "%s, src %s, dst %s: Dropping %s to simulate"
- "silent message loss\n",
+ lnet_drop_rule_match(hdr, ni->ni_nid, NULL)) {
+ CDEBUG(D_NET,
+ "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
libcfs_nid2str(from_nid), libcfs_nid2str(src_nid),
libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
goto drop;
}
- if (lnet_drop_asym_route && for_me &&
- LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
- struct lnet_net *net;
- struct lnet_remotenet *rnet;
- bool found = true;
-
- /* we are dealing with a routed message,
- * so see if route to reach src_nid goes through from_nid
- */
- lnet_net_lock(cpt);
- net = lnet_get_net_locked(LNET_NIDNET(ni->ni_nid));
- if (!net) {
- lnet_net_unlock(cpt);
- CERROR("net %s not found\n",
- libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
- return -EPROTO;
- }
-
- rnet = lnet_find_rnet_locked(LNET_NIDNET(src_nid));
- if (rnet) {
- struct lnet_peer_ni *gw = NULL;
- struct lnet_route *route;
-
- list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
- found = false;
- gw = route->lr_gateway;
- if (gw->lpni_net != net)
- continue;
- if (gw->lpni_nid == from_nid) {
- found = true;
- break;
- }
- }
- }
- lnet_net_unlock(cpt);
- if (!found) {
- /* we would not use from_nid to route a message to
- * src_nid
- * => asymmetric routing detected but forbidden
- */
- CERROR("%s, src %s: Dropping asymmetrical route %s\n",
- libcfs_nid2str(from_nid),
- libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
- goto drop;
- }
- }
-
msg = lnet_msg_alloc();
if (msg == NULL) {
CERROR("%s, src %s: Dropping %s (out of memory)\n",
return 0;
goto drop;
}
+
+ if (lnet_drop_asym_route && for_me &&
+ LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
+ __u32 src_net_id = LNET_NIDNET(src_nid);
+ struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
+ struct lnet_route *route;
+ bool found = false;
+
+ list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
+ if (route->lr_net == src_net_id) {
+ found = true;
+ break;
+ }
+ }
+ if (!found) {
+ lnet_net_unlock(cpt);
+ /* we would not use from_nid to route a message to
+ * src_nid
+ * => asymmetric routing detected but forbidden
+ */
+ CERROR("%s, src %s: Dropping asymmetrical route %s\n",
+ libcfs_nid2str(from_nid),
+ libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
+ lnet_msg_free(msg);
+ goto drop;
+ }
+ }
+
+ lpni->lpni_last_alive = ktime_get_seconds();
+
msg->msg_rxpeer = lpni;
msg->msg_rxni = ni;
lnet_ni_addref_locked(ni, cpt);
/* Multi-Rail: Primary NID of source. */
msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
- if (lnet_isrouter(msg->msg_rxpeer)) {
- lnet_peer_set_alive(msg->msg_rxpeer);
- if (avoid_asym_router_failure &&
- LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
- /* received a remote message from router, update
- * remote NI status on this router.
- * NB: multi-hop routed message will be ignored.
- */
- lnet_router_ni_update_locked(msg->msg_rxpeer,
- LNET_NIDNET(src_nid));
- }
- }
+ /*
+ * mark the status of this lpni as UP since we received a message
+ * from it. The ping response reports back the ns_status which is
+ * marked on the remote as up or down and we cache it here.
+ */
+ msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
lnet_msg_commit(msg, cpt);
struct lnet_libmd *md, struct lnet_handle_md mdh)
{
s64 timeout_ns;
- bool new_entry = true;
struct lnet_rsp_tracker *local_rspt;
/*
* we already have an rspt attached to the md, so we'll
* update the deadline on that one.
*/
- LIBCFS_FREE(rspt, sizeof(*rspt));
- new_entry = false;
+ lnet_rspt_free(rspt, cpt);
} else {
/* new md */
rspt->rspt_mdh = mdh;
* list in order to expire all the older entries first.
*/
lnet_net_lock(cpt);
- if (!new_entry && !list_empty(&local_rspt->rspt_on_list))
- list_del_init(&local_rspt->rspt_on_list);
- list_add_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
+ list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
lnet_net_unlock(cpt);
lnet_res_unlock(cpt);
}
libcfs_id2str(target));
return -ENOMEM;
}
- msg->msg_vmflush = !!memory_pressure_get();
+ msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
cpt = lnet_cpt_of_cookie(mdh.cookie);
md->md_me->me_portal);
lnet_res_unlock(cpt);
- LIBCFS_FREE(rspt, sizeof(*rspt));
+ if (rspt)
+ lnet_rspt_free(rspt, cpt);
+
lnet_msg_free(msg);
return -ENOENT;
}
lnet_build_msg_event(msg, LNET_EVENT_SEND);
- if (ack == LNET_ACK_REQ)
+ if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
+ md->md_options))
lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+ else if (rspt)
+ lnet_rspt_free(rspt, cpt);
if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
CFS_FAIL_ONCE))
lnet_res_unlock(cpt);
lnet_msg_free(msg);
- LIBCFS_FREE(rspt, sizeof(*rspt));
+ lnet_rspt_free(rspt, cpt);
return -ENOENT;
}
lnet_build_msg_event(msg, LNET_EVENT_SEND);
- lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+ if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
+ lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+ else
+ lnet_rspt_free(rspt, cpt);
rc = lnet_send(self, msg, LNET_NID_ANY);
if (rc < 0) {
int
LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
{
- struct list_head *e;
+ struct list_head *e;
struct lnet_ni *ni = NULL;
struct lnet_remotenet *rnet;
- __u32 dstnet = LNET_NIDNET(dstnid);
- int hops;
- int cpt;
- __u32 order = 2;
- struct list_head *rn_list;
+ __u32 dstnet = LNET_NIDNET(dstnid);
+ int hops;
+ int cpt;
+ __u32 order = 2;
+ struct list_head *rn_list;
/* if !local_nid_dist_zero, I don't return a distance of 0 ever
* (when lustre sees a distance of 0, it substitutes 0@lo), so I
if (srcnidp != NULL)
*srcnidp = dstnid;
if (orderp != NULL) {
- if (LNET_NETTYP(LNET_NIDNET(dstnid)) == LOLND)
+ if (dstnid == LNET_NID_LO_0)
*orderp = 0;
else
*orderp = 1;
* current net namespace.
* If not, assign order above 0xffff0000,
* to make this ni not a priority. */
- if (!net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
- order += 0xffff0000;
-
+ if (current->nsproxy &&
+ !net_eq(ni->ni_net_ns, current->nsproxy->net_ns))
+ order += 0xffff0000;
if (srcnidp != NULL)
*srcnidp = ni->ni_nid;
if (orderp != NULL)
LASSERT(shortest != NULL);
hops = shortest_hops;
if (srcnidp != NULL) {
- ni = lnet_get_next_ni_locked(
- shortest->lr_gateway->lpni_net,
- NULL);
+ struct lnet_net *net;
+ net = lnet_get_net_locked(shortest->lr_lnet);
+ LASSERT(net);
+ ni = lnet_get_next_ni_locked(net, NULL);
*srcnidp = ni->ni_nid;
}
if (orderp != NULL)