*
* You should have received a copy of the GNU General Public License
* version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
*
* GPL HEADER END
*/
* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2015, Intel Corporation.
+ * Copyright (c) 2012, 2016, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include "o2iblnd.h"
-static void kiblnd_peer_alive(kib_peer_t *peer);
-static void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
-static void kiblnd_check_sends(kib_conn_t *conn);
+#define MAX_CONN_RACES_BEFORE_ABORT 20
+
+static void kiblnd_peer_alive(kib_peer_ni_t *peer_ni);
+static void kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error);
static void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx,
int type, int body_nob);
static int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
static void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
static void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
static void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
+static void kiblnd_check_sends_locked(kib_conn_t *conn);
-static void
+void
kiblnd_tx_done (lnet_ni_t *ni, kib_tx_t *tx)
{
lnet_msg_t *lntmsg[2];
LASSERT (!in_interrupt());
LASSERT (!tx->tx_queued); /* mustn't be queued for sending */
LASSERT (tx->tx_sending == 0); /* mustn't be awaiting sent callback */
- LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer response */
+ LASSERT (!tx->tx_waiting); /* mustn't be awaiting peer_ni response */
LASSERT (tx->tx_pool != NULL);
kiblnd_unmap_tx(ni, tx);
kib_tx_t *tx;
kib_tx_poolset_t *tps;
- tps = net->ibn_tx_ps[lnet_cpt_of_nid(target)];
+ tps = net->ibn_tx_ps[lnet_cpt_of_nid(target, ni)];
node = kiblnd_pool_alloc_node(&tps->tps_poolset);
if (node == NULL)
return NULL;
conn->ibc_outstanding_credits++;
else
conn->ibc_reserved_credits++;
+ kiblnd_check_sends_locked(conn);
spin_unlock(&conn->ibc_lock);
- kiblnd_check_sends(conn);
out:
kiblnd_conn_decref(conn);
return rc;
if (tx == NULL) {
spin_unlock(&conn->ibc_lock);
- CWARN("Unmatched completion type %x cookie "LPX64" from %s\n",
+ CWARN("Unmatched completion type %x cookie %#llx from %s\n",
txtype, cookie, libcfs_nid2str(conn->ibc_peer->ibp_nid));
kiblnd_close_conn(conn, -EPROTO);
return;
!IBLND_OOB_CAPABLE(conn->ibc_version)) /* v1 only */
conn->ibc_outstanding_credits++;
+ kiblnd_check_sends_locked(conn);
spin_unlock(&conn->ibc_lock);
- kiblnd_check_sends(conn);
}
switch (msg->ibm_type) {
LASSERT (tx->tx_waiting);
/* CAVEAT EMPTOR: I could be racing with tx_complete, but...
- * (a) I can overwrite tx_msg since my peer has received it!
+ * (a) I can overwrite tx_msg since my peer_ni has received it!
* (b) tx_waiting set tells tx_complete() it's not done. */
tx->tx_nwrq = 0; /* overwrite PUT_REQ */
}
static int
-kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, int nob)
+kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
{
kib_hca_dev_t *hdev;
- __u64 *pages = tx->tx_pages;
kib_fmr_poolset_t *fps;
- int npages;
- int size;
int cpt;
int rc;
- int i;
LASSERT(tx->tx_pool != NULL);
LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
- hdev = tx->tx_pool->tpo_hdev;
-
- for (i = 0, npages = 0; i < rd->rd_nfrags; i++) {
- for (size = 0; size < rd->rd_frags[i].rf_nob;
- size += hdev->ibh_page_size) {
- pages[npages ++] = (rd->rd_frags[i].rf_addr &
- hdev->ibh_page_mask) + size;
- }
- }
-
+ hdev = tx->tx_pool->tpo_hdev;
cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
fps = net->ibn_fmr_ps[cpt];
- rc = kiblnd_fmr_pool_map(fps, pages, npages, 0, &tx->fmr);
- if (rc != 0) {
- CERROR ("Can't map %d pages: %d\n", npages, rc);
- return rc;
- }
+ rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
+ if (rc != 0) {
+ CERROR("Can't map %u pages: %d\n", nob, rc);
+ return rc;
+ }
- /* If rd is not tx_rd, it's going to get sent to a peer, who will need
+ /* If rd is not tx_rd, it's going to get sent to a peer_ni, who will need
* the rkey */
- rd->rd_key = (rd != tx->tx_rd) ? tx->fmr.fmr_pfmr->fmr->rkey :
- tx->fmr.fmr_pfmr->fmr->lkey;
+ rd->rd_key = tx->fmr.fmr_key;
rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
rd->rd_frags[0].rf_nob = nob;
rd->rd_nfrags = 1;
LASSERT(net != NULL);
- if (net->ibn_fmr_ps != NULL && tx->fmr.fmr_pfmr != NULL) {
+ if (net->ibn_fmr_ps != NULL)
kiblnd_fmr_pool_unmap(&tx->fmr, tx->tx_status);
- tx->fmr.fmr_pfmr = NULL;
- }
if (tx->tx_nfrags != 0) {
kiblnd_dma_unmap_sg(tx->tx_pool->tpo_hdev->ibh_ibdev,
static int
kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx, kib_rdma_desc_t *rd, int nfrags)
{
- kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
kib_net_t *net = ni->ni_data;
+ kib_hca_dev_t *hdev = net->ibn_dev->ibd_hdev;
struct ib_mr *mr = NULL;
- __u32 nob;
- int i;
+ __u32 nob;
+ int i;
- /* If rd is not tx_rd, it's going to get sent to a peer and I'm the
+ /* If rd is not tx_rd, it's going to get sent to a peer_ni and I'm the
* RDMA sink */
tx->tx_dmadir = (rd != tx->tx_rd) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
tx->tx_nfrags = nfrags;
nob += rd->rd_frags[i].rf_nob;
}
- mr = kiblnd_find_rd_dma_mr(hdev, rd,
+ mr = kiblnd_find_rd_dma_mr(ni, rd,
(tx->tx_conn != NULL) ?
tx->tx_conn->ibc_max_frags : -1);
if (mr != NULL) {
fragnob = min(fragnob, (int)PAGE_SIZE - page_offset);
sg_set_page(sg, page, fragnob, page_offset);
- sg++;
+ sg = sg_next(sg);
+ if (!sg) {
+ CERROR("lacking enough sg entries to map tx\n");
+ return -EFAULT;
+ }
if (offset + fragnob < iov->iov_len) {
offset += fragnob;
fragnob = min((int)(kiov->kiov_len - offset), nob);
- sg_set_page(sg, kiov->kiov_page, fragnob,
- kiov->kiov_offset + offset);
- sg++;
+ sg_set_page(sg, kiov->kiov_page, fragnob,
+ kiov->kiov_offset + offset);
+ sg = sg_next(sg);
+ if (!sg) {
+ CERROR("lacking enough sg entries to map tx\n");
+ return -EFAULT;
+ }
offset = 0;
kiov++;
kiblnd_post_tx_locked (kib_conn_t *conn, kib_tx_t *tx, int credit)
__must_hold(&conn->ibc_lock)
{
- kib_msg_t *msg = tx->tx_msg;
- kib_peer_t *peer = conn->ibc_peer;
- int ver = conn->ibc_version;
- int rc;
- int done;
- struct ib_send_wr *bad_wrq;
+ kib_msg_t *msg = tx->tx_msg;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
+ struct lnet_ni *ni = peer_ni->ibp_ni;
+ int ver = conn->ibc_version;
+ int rc;
+ int done;
LASSERT(tx->tx_queued);
/* We rely on this for QP sizing */
LASSERT(conn->ibc_credits >= 0);
LASSERT(conn->ibc_credits <= conn->ibc_queue_depth);
- if (conn->ibc_nsends_posted == IBLND_CONCURRENT_SENDS(ver)) {
+ if (conn->ibc_nsends_posted ==
+ kiblnd_concurrent_sends(ver, ni)) {
/* tx completions outstanding... */
CDEBUG(D_NET, "%s: posted enough\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
return -EAGAIN;
}
if (credit != 0 && conn->ibc_credits == 0) { /* no credits */
CDEBUG(D_NET, "%s: no credits\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
return -EAGAIN;
}
conn->ibc_credits == 1 && /* last credit reserved */
msg->ibm_type != IBLND_MSG_NOOP) { /* for NOOP */
CDEBUG(D_NET, "%s: not using last credit\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
return -EAGAIN;
}
(!kiblnd_need_noop(conn) || /* redundant NOOP */
(IBLND_OOB_CAPABLE(ver) && /* posted enough NOOP */
conn->ibc_noops_posted == IBLND_OOB_MSGS(ver)))) {
- /* OK to drop when posted enough NOOPs, since
- * kiblnd_check_sends will queue NOOP again when
- * posted NOOPs complete */
+ /* OK to drop when posted enough NOOPs, since
+ * kiblnd_check_sends_locked will queue NOOP again when
+ * posted NOOPs complete */
spin_unlock(&conn->ibc_lock);
- kiblnd_tx_done(peer->ibp_ni, tx);
+ kiblnd_tx_done(peer_ni->ibp_ni, tx);
spin_lock(&conn->ibc_lock);
CDEBUG(D_NET, "%s(%d): redundant or enough NOOP\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
conn->ibc_noops_posted);
return 0;
}
- kiblnd_pack_msg(peer->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
- peer->ibp_nid, conn->ibc_incarnation);
+ kiblnd_pack_msg(peer_ni->ibp_ni, msg, ver, conn->ibc_outstanding_credits,
+ peer_ni->ibp_nid, conn->ibc_incarnation);
conn->ibc_credits -= credit;
conn->ibc_outstanding_credits = 0;
/* close_conn will launch failover */
rc = -ENETDOWN;
} else {
- rc = ib_post_send(conn->ibc_cmid->qp,
- tx->tx_wrq, &bad_wrq);
- }
+ struct kib_fast_reg_descriptor *frd = tx->fmr.fmr_frd;
+ struct ib_send_wr *bad = &tx->tx_wrq[tx->tx_nwrq - 1].wr;
+ struct ib_send_wr *wr = &tx->tx_wrq[0].wr;
+
+ if (frd != NULL) {
+ if (!frd->frd_valid) {
+ wr = &frd->frd_inv_wr.wr;
+ wr->next = &frd->frd_fastreg_wr.wr;
+ } else {
+ wr = &frd->frd_fastreg_wr.wr;
+ }
+ frd->frd_fastreg_wr.wr.next = &tx->tx_wrq[0].wr;
+ }
+
+ LASSERTF(bad->wr_id == kiblnd_ptr2wreqid(tx, IBLND_WID_TX),
+ "bad wr_id %#llx, opc %d, flags %d, peer_ni: %s\n",
+ bad->wr_id, bad->opcode, bad->send_flags,
+ libcfs_nid2str(conn->ibc_peer->ibp_nid));
+
+ bad = NULL;
+ rc = ib_post_send(conn->ibc_cmid->qp, wr, &bad);
+ }
conn->ibc_last_send = jiffies;
if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
CERROR("Error %d posting transmit to %s\n",
- rc, libcfs_nid2str(peer->ibp_nid));
+ rc, libcfs_nid2str(peer_ni->ibp_nid));
else
CDEBUG(D_NET, "Error %d posting transmit to %s\n",
- rc, libcfs_nid2str(peer->ibp_nid));
+ rc, libcfs_nid2str(peer_ni->ibp_nid));
kiblnd_close_conn(conn, rc);
if (done)
- kiblnd_tx_done(peer->ibp_ni, tx);
+ kiblnd_tx_done(peer_ni->ibp_ni, tx);
spin_lock(&conn->ibc_lock);
}
static void
-kiblnd_check_sends (kib_conn_t *conn)
+kiblnd_check_sends_locked(kib_conn_t *conn)
{
int ver = conn->ibc_version;
lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
return;
}
- spin_lock(&conn->ibc_lock);
-
- LASSERT (conn->ibc_nsends_posted <= IBLND_CONCURRENT_SENDS(ver));
+ LASSERT(conn->ibc_nsends_posted <=
+ kiblnd_concurrent_sends(ver, ni));
LASSERT (!IBLND_OOB_CAPABLE(ver) ||
conn->ibc_noops_posted <= IBLND_OOB_MSGS(ver));
LASSERT (conn->ibc_reserved_credits >= 0);
if (kiblnd_post_tx_locked(conn, tx, credit) != 0)
break;
}
-
- spin_unlock(&conn->ibc_lock);
}
static void
if (failed) {
if (conn->ibc_state == IBLND_CONN_ESTABLISHED)
- CNETERR("Tx -> %s cookie "LPX64
+ CNETERR("Tx -> %s cookie %#llx"
" sending %d waiting %d: failed %d\n",
libcfs_nid2str(conn->ibc_peer->ibp_nid),
tx->tx_cookie, tx->tx_sending, tx->tx_waiting,
conn->ibc_noops_posted--;
if (failed) {
- tx->tx_waiting = 0; /* don't wait for peer */
+ tx->tx_waiting = 0; /* don't wait for peer_ni */
tx->tx_status = -EIO;
}
idle = (tx->tx_sending == 0) && /* This is the final callback */
- !tx->tx_waiting && /* Not waiting for peer */
+ !tx->tx_waiting && /* Not waiting for peer_ni */
!tx->tx_queued; /* Not re-queued (PUT_DONE) */
if (idle)
list_del(&tx->tx_list);
- kiblnd_conn_addref(conn); /* 1 ref for me.... */
-
+ kiblnd_check_sends_locked(conn);
spin_unlock(&conn->ibc_lock);
if (idle)
kiblnd_tx_done(conn->ibc_peer->ibp_ni, tx);
-
- kiblnd_check_sends(conn);
-
- kiblnd_conn_decref(conn); /* ...until here */
}
static void
kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob)
{
- kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
- struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
- struct ib_send_wr *wrq = &tx->tx_wrq[tx->tx_nwrq];
- int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
- struct ib_mr *mr = hdev->ibh_mrs;
+ kib_hca_dev_t *hdev = tx->tx_pool->tpo_hdev;
+ struct ib_sge *sge = &tx->tx_sge[tx->tx_nwrq];
+ struct ib_rdma_wr *wrq;
+ int nob = offsetof(kib_msg_t, ibm_u) + body_nob;
+ struct ib_mr *mr = hdev->ibh_mrs;
LASSERT(tx->tx_nwrq >= 0);
LASSERT(tx->tx_nwrq < IBLND_MAX_RDMA_FRAGS + 1);
sge->addr = tx->tx_msgaddr;
sge->length = nob;
- memset(wrq, 0, sizeof(*wrq));
+ wrq = &tx->tx_wrq[tx->tx_nwrq];
+ memset(wrq, 0, sizeof(*wrq));
- wrq->next = NULL;
- wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
- wrq->sg_list = sge;
- wrq->num_sge = 1;
- wrq->opcode = IB_WR_SEND;
- wrq->send_flags = IB_SEND_SIGNALED;
+ wrq->wr.next = NULL;
+ wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_TX);
+ wrq->wr.sg_list = sge;
+ wrq->wr.num_sge = 1;
+ wrq->wr.opcode = IB_WR_SEND;
+ wrq->wr.send_flags = IB_SEND_SIGNALED;
- tx->tx_nwrq++;
+ tx->tx_nwrq++;
}
static int
kib_msg_t *ibmsg = tx->tx_msg;
kib_rdma_desc_t *srcrd = tx->tx_rd;
struct ib_sge *sge = &tx->tx_sge[0];
- struct ib_send_wr *wrq = &tx->tx_wrq[0];
+ struct ib_rdma_wr *wrq;
int rc = resid;
int srcidx;
int dstidx;
}
if (tx->tx_nwrq >= conn->ibc_max_frags) {
- CERROR("RDMA has too many fragments for peer %s (%d), "
+ CERROR("RDMA has too many fragments for peer_ni %s (%d), "
"src idx/frags: %d/%d dst idx/frags: %d/%d\n",
libcfs_nid2str(conn->ibc_peer->ibp_nid),
conn->ibc_max_frags,
wrq = &tx->tx_wrq[tx->tx_nwrq];
- wrq->next = wrq + 1;
- wrq->wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
- wrq->sg_list = sge;
- wrq->num_sge = 1;
- wrq->opcode = IB_WR_RDMA_WRITE;
- wrq->send_flags = 0;
-
- wrq->wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
- wrq->wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
+ wrq->wr.next = &(wrq + 1)->wr;
+ wrq->wr.wr_id = kiblnd_ptr2wreqid(tx, IBLND_WID_RDMA);
+ wrq->wr.sg_list = sge;
+ wrq->wr.num_sge = 1;
+ wrq->wr.opcode = IB_WR_RDMA_WRITE;
+ wrq->wr.send_flags = 0;
+
+#ifdef HAVE_IB_RDMA_WR
+ wrq->remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
+ wrq->rkey = kiblnd_rd_frag_key(dstrd, dstidx);
+#else
+ wrq->wr.wr.rdma.remote_addr = kiblnd_rd_frag_addr(dstrd, dstidx);
+ wrq->wr.wr.rdma.rkey = kiblnd_rd_frag_key(dstrd, dstidx);
+#endif
srcidx = kiblnd_rd_consume_frag(srcrd, srcidx, wrknob);
dstidx = kiblnd_rd_consume_frag(dstrd, dstidx, wrknob);
{
spin_lock(&conn->ibc_lock);
kiblnd_queue_tx_locked(tx, conn);
+ kiblnd_check_sends_locked(conn);
spin_unlock(&conn->ibc_lock);
-
- kiblnd_check_sends(conn);
}
static int kiblnd_resolve_addr(struct rdma_cm_id *cmid,
return rc;
}
-void
-kiblnd_connect_peer (kib_peer_t *peer)
+static void
+kiblnd_connect_peer (kib_peer_ni_t *peer_ni)
{
struct rdma_cm_id *cmid;
kib_dev_t *dev;
- kib_net_t *net = peer->ibp_ni->ni_data;
+ kib_net_t *net = peer_ni->ibp_ni->ni_data;
struct sockaddr_in srcaddr;
struct sockaddr_in dstaddr;
int rc;
LASSERT (net != NULL);
- LASSERT (peer->ibp_connecting > 0);
+ LASSERT (peer_ni->ibp_connecting > 0);
+ LASSERT(!peer_ni->ibp_reconnecting);
- cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer, RDMA_PS_TCP,
+ cmid = kiblnd_rdma_create_id(kiblnd_cm_callback, peer_ni, RDMA_PS_TCP,
IB_QPT_RC);
if (IS_ERR(cmid)) {
CERROR("Can't create CMID for %s: %ld\n",
- libcfs_nid2str(peer->ibp_nid), PTR_ERR(cmid));
+ libcfs_nid2str(peer_ni->ibp_nid), PTR_ERR(cmid));
rc = PTR_ERR(cmid);
goto failed;
}
memset(&dstaddr, 0, sizeof(dstaddr));
dstaddr.sin_family = AF_INET;
dstaddr.sin_port = htons(*kiblnd_tunables.kib_service);
- dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer->ibp_nid));
+ dstaddr.sin_addr.s_addr = htonl(LNET_NIDADDR(peer_ni->ibp_nid));
- kiblnd_peer_addref(peer); /* cmid's ref */
+ kiblnd_peer_addref(peer_ni); /* cmid's ref */
if (*kiblnd_tunables.kib_use_priv_port) {
rc = kiblnd_resolve_addr(cmid, &srcaddr, &dstaddr,
if (rc != 0) {
/* Can't initiate address resolution: */
CERROR("Can't resolve addr for %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), rc);
+ libcfs_nid2str(peer_ni->ibp_nid), rc);
goto failed2;
}
LASSERT (cmid->device != NULL);
CDEBUG(D_NET, "%s: connection bound to %s:%pI4h:%s\n",
- libcfs_nid2str(peer->ibp_nid), dev->ibd_ifname,
+ libcfs_nid2str(peer_ni->ibp_nid), dev->ibd_ifname,
&dev->ibd_ifip, cmid->device->name);
- return;
+ return;
failed2:
- kiblnd_peer_decref(peer); /* cmid's ref */
- rdma_destroy_id(cmid);
+ kiblnd_peer_connect_failed(peer_ni, 1, rc);
+ kiblnd_peer_decref(peer_ni); /* cmid's ref */
+ rdma_destroy_id(cmid);
+ return;
failed:
- kiblnd_peer_connect_failed(peer, 1, rc);
+ kiblnd_peer_connect_failed(peer_ni, 1, rc);
+}
+
+bool
+kiblnd_reconnect_peer(kib_peer_ni_t *peer_ni)
+{
+ rwlock_t *glock = &kiblnd_data.kib_global_lock;
+ char *reason = NULL;
+ struct list_head txs;
+ unsigned long flags;
+
+ INIT_LIST_HEAD(&txs);
+
+ write_lock_irqsave(glock, flags);
+ if (peer_ni->ibp_reconnecting == 0) {
+ if (peer_ni->ibp_accepting)
+ reason = "accepting";
+ else if (peer_ni->ibp_connecting)
+ reason = "connecting";
+ else if (!list_empty(&peer_ni->ibp_conns))
+ reason = "connected";
+ else /* connected then closed */
+ reason = "closed";
+
+ goto no_reconnect;
+ }
+
+ LASSERT(!peer_ni->ibp_accepting && !peer_ni->ibp_connecting &&
+ list_empty(&peer_ni->ibp_conns));
+ peer_ni->ibp_reconnecting = 0;
+
+ if (!kiblnd_peer_active(peer_ni)) {
+ list_splice_init(&peer_ni->ibp_tx_queue, &txs);
+ reason = "unlinked";
+ goto no_reconnect;
+ }
+
+ peer_ni->ibp_connecting++;
+ peer_ni->ibp_reconnected++;
+
+ write_unlock_irqrestore(glock, flags);
+
+ kiblnd_connect_peer(peer_ni);
+ return true;
+
+ no_reconnect:
+ write_unlock_irqrestore(glock, flags);
+
+ CWARN("Abort reconnection of %s: %s\n",
+ libcfs_nid2str(peer_ni->ibp_nid), reason);
+ kiblnd_txlist_done(peer_ni->ibp_ni, &txs, -ECONNABORTED);
+ return false;
}
void
kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid)
{
- kib_peer_t *peer;
- kib_peer_t *peer2;
+ kib_peer_ni_t *peer_ni;
+ kib_peer_ni_t *peer2;
kib_conn_t *conn;
rwlock_t *g_lock = &kiblnd_data.kib_global_lock;
unsigned long flags;
LASSERT (tx == NULL || tx->tx_conn == NULL); /* only set when assigned a conn */
LASSERT (tx == NULL || tx->tx_nwrq > 0); /* work items have been set up */
- /* First time, just use a read lock since I expect to find my peer
+ /* First time, just use a read lock since I expect to find my peer_ni
* connected */
read_lock_irqsave(g_lock, flags);
- peer = kiblnd_find_peer_locked(nid);
- if (peer != NULL && !list_empty(&peer->ibp_conns)) {
- /* Found a peer with an established connection */
- conn = kiblnd_get_conn_locked(peer);
+ peer_ni = kiblnd_find_peer_locked(ni, nid);
+ if (peer_ni != NULL && !list_empty(&peer_ni->ibp_conns)) {
+ /* Found a peer_ni with an established connection */
+ conn = kiblnd_get_conn_locked(peer_ni);
kiblnd_conn_addref(conn); /* 1 ref for me... */
read_unlock_irqrestore(g_lock, flags);
/* Re-try with a write lock */
write_lock(g_lock);
- peer = kiblnd_find_peer_locked(nid);
- if (peer != NULL) {
- if (list_empty(&peer->ibp_conns)) {
- /* found a peer, but it's still connecting... */
- LASSERT (peer->ibp_connecting != 0 ||
- peer->ibp_accepting != 0);
+ peer_ni = kiblnd_find_peer_locked(ni, nid);
+ if (peer_ni != NULL) {
+ if (list_empty(&peer_ni->ibp_conns)) {
+ /* found a peer_ni, but it's still connecting... */
+ LASSERT(kiblnd_peer_connecting(peer_ni));
if (tx != NULL)
list_add_tail(&tx->tx_list,
- &peer->ibp_tx_queue);
+ &peer_ni->ibp_tx_queue);
write_unlock_irqrestore(g_lock, flags);
} else {
- conn = kiblnd_get_conn_locked(peer);
+ conn = kiblnd_get_conn_locked(peer_ni);
kiblnd_conn_addref(conn); /* 1 ref for me... */
write_unlock_irqrestore(g_lock, flags);
write_unlock_irqrestore(g_lock, flags);
- /* Allocate a peer ready to add to the peer table and retry */
- rc = kiblnd_create_peer(ni, &peer, nid);
+ /* Allocate a peer_ni ready to add to the peer_ni table and retry */
+ rc = kiblnd_create_peer(ni, &peer_ni, nid);
if (rc != 0) {
- CERROR("Can't create peer %s\n", libcfs_nid2str(nid));
+ CERROR("Can't create peer_ni %s\n", libcfs_nid2str(nid));
if (tx != NULL) {
tx->tx_status = -EHOSTUNREACH;
tx->tx_waiting = 0;
write_lock_irqsave(g_lock, flags);
- peer2 = kiblnd_find_peer_locked(nid);
+ peer2 = kiblnd_find_peer_locked(ni, nid);
if (peer2 != NULL) {
if (list_empty(&peer2->ibp_conns)) {
- /* found a peer, but it's still connecting... */
- LASSERT (peer2->ibp_connecting != 0 ||
- peer2->ibp_accepting != 0);
+ /* found a peer_ni, but it's still connecting... */
+ LASSERT(kiblnd_peer_connecting(peer2));
if (tx != NULL)
list_add_tail(&tx->tx_list,
&peer2->ibp_tx_queue);
kiblnd_conn_decref(conn); /* ...to here */
}
- kiblnd_peer_decref(peer);
+ kiblnd_peer_decref(peer_ni);
return;
}
- /* Brand new peer */
- LASSERT (peer->ibp_connecting == 0);
- peer->ibp_connecting = 1;
+ /* Brand new peer_ni */
+ LASSERT (peer_ni->ibp_connecting == 0);
+ peer_ni->ibp_connecting = 1;
/* always called with a ref on ni, which prevents ni being shutdown */
LASSERT (((kib_net_t *)ni->ni_data)->ibn_shutdown == 0);
if (tx != NULL)
- list_add_tail(&tx->tx_list, &peer->ibp_tx_queue);
+ list_add_tail(&tx->tx_list, &peer_ni->ibp_tx_queue);
- kiblnd_peer_addref(peer);
- list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
+ kiblnd_peer_addref(peer_ni);
+ list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
write_unlock_irqrestore(g_lock, flags);
- kiblnd_connect_peer(peer);
- kiblnd_peer_decref(peer);
+ kiblnd_connect_peer(peer_ni);
+ kiblnd_peer_decref(peer_ni);
}
int
CERROR("Can't setup PUT sink for %s: %d\n",
libcfs_nid2str(conn->ibc_peer->ibp_nid), rc);
kiblnd_tx_done(ni, tx);
- /* tell peer it's over */
+ /* tell peer_ni it's over */
kiblnd_send_completion(rx->rx_conn, IBLND_MSG_PUT_NAK, rc,
rxmsg->ibm_u.putreq.ibprm_cookie);
break;
}
static void
-kiblnd_peer_alive (kib_peer_t *peer)
+kiblnd_peer_alive (kib_peer_ni_t *peer_ni)
{
/* This is racy, but everyone's only writing cfs_time_current() */
- peer->ibp_last_alive = cfs_time_current();
+ peer_ni->ibp_last_alive = cfs_time_current();
smp_mb();
}
static void
-kiblnd_peer_notify (kib_peer_t *peer)
+kiblnd_peer_notify (kib_peer_ni_t *peer_ni)
{
int error = 0;
cfs_time_t last_alive = 0;
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
- if (list_empty(&peer->ibp_conns) &&
- peer->ibp_accepting == 0 &&
- peer->ibp_connecting == 0 &&
- peer->ibp_error != 0) {
- error = peer->ibp_error;
- peer->ibp_error = 0;
+ if (kiblnd_peer_idle(peer_ni) && peer_ni->ibp_error != 0) {
+ error = peer_ni->ibp_error;
+ peer_ni->ibp_error = 0;
- last_alive = peer->ibp_last_alive;
+ last_alive = peer_ni->ibp_last_alive;
}
read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
if (error != 0)
- lnet_notify(peer->ibp_ni,
- peer->ibp_nid, 0, last_alive);
+ lnet_notify(peer_ni->ibp_ni,
+ peer_ni->ibp_nid, 0, last_alive);
}
void
* connection to be finished off by the connd. Otherwise the connd is
* already dealing with it (either to set it up or tear it down).
* Caller holds kib_global_lock exclusively in irq context */
- kib_peer_t *peer = conn->ibc_peer;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
kib_dev_t *dev;
unsigned long flags;
list_empty(&conn->ibc_tx_queue_nocred) &&
list_empty(&conn->ibc_active_txs)) {
CDEBUG(D_NET, "closing conn to %s\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
} else {
CNETERR("Closing conn to %s: error %d%s%s%s%s%s\n",
- libcfs_nid2str(peer->ibp_nid), error,
+ libcfs_nid2str(peer_ni->ibp_nid), error,
list_empty(&conn->ibc_tx_queue) ? "" : "(sending)",
list_empty(&conn->ibc_tx_noops) ? "" : "(sending_noops)",
list_empty(&conn->ibc_tx_queue_rsrvd) ?
list_empty(&conn->ibc_active_txs) ? "" : "(waiting)");
}
- dev = ((kib_net_t *)peer->ibp_ni->ni_data)->ibn_dev;
+ dev = ((kib_net_t *)peer_ni->ibp_ni->ni_data)->ibn_dev;
list_del(&conn->ibc_list);
/* connd (see below) takes over ibc_list's ref */
- if (list_empty(&peer->ibp_conns) && /* no more conns */
- kiblnd_peer_active(peer)) { /* still in peer table */
- kiblnd_unlink_peer_locked(peer);
+ if (list_empty(&peer_ni->ibp_conns) && /* no more conns */
+ kiblnd_peer_active(peer_ni)) { /* still in peer_ni table */
+ kiblnd_unlink_peer_locked(peer_ni);
/* set/clear error on last conn */
- peer->ibp_error = conn->ibc_comms_error;
+ peer_ni->ibp_error = conn->ibc_comms_error;
}
kiblnd_set_conn_state(conn, IBLND_CONN_CLOSING);
}
static void
-kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error)
+kiblnd_peer_connect_failed(kib_peer_ni_t *peer_ni, int active, int error)
{
struct list_head zombies = LIST_HEAD_INIT(zombies);
unsigned long flags;
write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
if (active) {
- LASSERT (peer->ibp_connecting > 0);
- peer->ibp_connecting--;
+ LASSERT(peer_ni->ibp_connecting > 0);
+ peer_ni->ibp_connecting--;
} else {
- LASSERT (peer->ibp_accepting > 0);
- peer->ibp_accepting--;
+ LASSERT (peer_ni->ibp_accepting > 0);
+ peer_ni->ibp_accepting--;
}
- if (peer->ibp_connecting != 0 ||
- peer->ibp_accepting != 0) {
+ if (kiblnd_peer_connecting(peer_ni)) {
/* another connection attempt under way... */
write_unlock_irqrestore(&kiblnd_data.kib_global_lock,
flags);
return;
}
- if (list_empty(&peer->ibp_conns)) {
- /* Take peer's blocked transmits to complete with error */
- list_add(&zombies, &peer->ibp_tx_queue);
- list_del_init(&peer->ibp_tx_queue);
+ peer_ni->ibp_reconnected = 0;
+ if (list_empty(&peer_ni->ibp_conns)) {
+ /* Take peer_ni's blocked transmits to complete with error */
+ list_add(&zombies, &peer_ni->ibp_tx_queue);
+ list_del_init(&peer_ni->ibp_tx_queue);
- if (kiblnd_peer_active(peer))
- kiblnd_unlink_peer_locked(peer);
+ if (kiblnd_peer_active(peer_ni))
+ kiblnd_unlink_peer_locked(peer_ni);
- peer->ibp_error = error;
+ peer_ni->ibp_error = error;
} else {
/* Can't have blocked transmits if there are connections */
- LASSERT(list_empty(&peer->ibp_tx_queue));
+ LASSERT(list_empty(&peer_ni->ibp_tx_queue));
}
write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- kiblnd_peer_notify(peer);
+ kiblnd_peer_notify(peer_ni);
if (list_empty(&zombies))
return;
CNETERR("Deleting messages for %s: connection failed\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
- kiblnd_txlist_done(peer->ibp_ni, &zombies, -EHOSTUNREACH);
+ kiblnd_txlist_done(peer_ni->ibp_ni, &zombies, -EHOSTUNREACH);
}
static void
kiblnd_connreq_done(kib_conn_t *conn, int status)
{
- kib_peer_t *peer = conn->ibc_peer;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
kib_tx_t *tx;
struct list_head txs;
unsigned long flags;
active = (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
CDEBUG(D_NET,"%s: active(%d), version(%x), status(%d)\n",
- libcfs_nid2str(peer->ibp_nid), active,
+ libcfs_nid2str(peer_ni->ibp_nid), active,
conn->ibc_version, status);
LASSERT (!in_interrupt());
LASSERT ((conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT &&
- peer->ibp_connecting > 0) ||
+ peer_ni->ibp_connecting > 0) ||
(conn->ibc_state == IBLND_CONN_PASSIVE_WAIT &&
- peer->ibp_accepting > 0));
+ peer_ni->ibp_accepting > 0));
LIBCFS_FREE(conn->ibc_connvars, sizeof(*conn->ibc_connvars));
conn->ibc_connvars = NULL;
if (status != 0) {
/* failed to establish connection */
- kiblnd_peer_connect_failed(peer, active, status);
+ kiblnd_peer_connect_failed(peer_ni, active, status);
kiblnd_finalise_conn(conn);
return;
}
conn->ibc_last_send = jiffies;
kiblnd_set_conn_state(conn, IBLND_CONN_ESTABLISHED);
- kiblnd_peer_alive(peer);
+ kiblnd_peer_alive(peer_ni);
- /* Add conn to peer's list and nuke any dangling conns from a different
- * peer instance... */
+ /* Add conn to peer_ni's list and nuke any dangling conns from a different
+ * peer_ni instance... */
kiblnd_conn_addref(conn); /* +1 ref for ibc_list */
- list_add(&conn->ibc_list, &peer->ibp_conns);
+ list_add(&conn->ibc_list, &peer_ni->ibp_conns);
+ peer_ni->ibp_reconnected = 0;
if (active)
- peer->ibp_connecting--;
+ peer_ni->ibp_connecting--;
else
- peer->ibp_accepting--;
+ peer_ni->ibp_accepting--;
- if (peer->ibp_version == 0) {
- peer->ibp_version = conn->ibc_version;
- peer->ibp_incarnation = conn->ibc_incarnation;
+ if (peer_ni->ibp_version == 0) {
+ peer_ni->ibp_version = conn->ibc_version;
+ peer_ni->ibp_incarnation = conn->ibc_incarnation;
}
- if (peer->ibp_version != conn->ibc_version ||
- peer->ibp_incarnation != conn->ibc_incarnation) {
- kiblnd_close_stale_conns_locked(peer, conn->ibc_version,
+ if (peer_ni->ibp_version != conn->ibc_version ||
+ peer_ni->ibp_incarnation != conn->ibc_incarnation) {
+ kiblnd_close_stale_conns_locked(peer_ni, conn->ibc_version,
conn->ibc_incarnation);
- peer->ibp_version = conn->ibc_version;
- peer->ibp_incarnation = conn->ibc_incarnation;
+ peer_ni->ibp_version = conn->ibc_version;
+ peer_ni->ibp_incarnation = conn->ibc_incarnation;
}
/* grab pending txs while I have the lock */
- list_add(&txs, &peer->ibp_tx_queue);
- list_del_init(&peer->ibp_tx_queue);
+ list_add(&txs, &peer_ni->ibp_tx_queue);
+ list_del_init(&peer_ni->ibp_tx_queue);
- if (!kiblnd_peer_active(peer) || /* peer has been deleted */
+ if (!kiblnd_peer_active(peer_ni) || /* peer_ni has been deleted */
conn->ibc_comms_error != 0) { /* error has happened already */
- lnet_ni_t *ni = peer->ibp_ni;
+ lnet_ni_t *ni = peer_ni->ibp_ni;
/* start to shut down connection */
kiblnd_close_conn_locked(conn, -ECONNABORTED);
return;
}
- /* refcount taken by cmid is not reliable after I released the glock
- * because this connection is visible to other threads now, another
- * thread can find and close this connection right after I released
- * the glock, if kiblnd_cm_callback for RDMA_CM_EVENT_DISCONNECTED is
- * called, it can release the connection refcount taken by cmid.
- * It means the connection could be destroyed before I finish my
- * operations on it.
+ /* +1 ref for myself, this connection is visible to other threads
+ * now, refcount of peer:ibp_conns can be released by connection
+ * close from either a different thread, or the calling of
+ * kiblnd_check_sends_locked() below. See bz21911 for details.
*/
kiblnd_conn_addref(conn);
write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
kiblnd_queue_tx_locked(tx, conn);
}
+ kiblnd_check_sends_locked(conn);
spin_unlock(&conn->ibc_lock);
- kiblnd_check_sends(conn);
-
/* schedule blocked rxs */
kiblnd_handle_early_rxs(conn);
-
kiblnd_conn_decref(conn);
}
kib_msg_t *reqmsg = priv;
kib_msg_t *ackmsg;
kib_dev_t *ibdev;
- kib_peer_t *peer;
- kib_peer_t *peer2;
+ kib_peer_ni_t *peer_ni;
+ kib_peer_ni_t *peer2;
kib_conn_t *conn;
lnet_ni_t *ni = NULL;
kib_net_t *net = NULL;
if (*kiblnd_tunables.kib_require_priv_port &&
ntohs(peer_addr->sin_port) >= PROT_SOCK) {
__u32 ip = ntohl(peer_addr->sin_addr.s_addr);
- CERROR("Peer's port (%pI4h:%hu) is not privileged\n",
+ CERROR("peer_ni's port (%pI4h:%hu) is not privileged\n",
&ip, ntohs(peer_addr->sin_port));
- goto failed;
- }
+ goto failed;
+ }
- if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
- CERROR("Short connection request\n");
- goto failed;
- }
+ if (priv_nob < offsetof(kib_msg_t, ibm_type)) {
+ CERROR("Short connection request\n");
+ goto failed;
+ }
- /* Future protocol version compatibility support! If the
- * o2iblnd-specific protocol changes, or when LNET unifies
- * protocols over all LNDs, the initial connection will
- * negotiate a protocol version. I trap this here to avoid
- * console errors; the reject tells the peer which protocol I
- * speak. */
- if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
- reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
- goto failed;
- if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
- reqmsg->ibm_version != IBLND_MSG_VERSION &&
- reqmsg->ibm_version != IBLND_MSG_VERSION_1)
- goto failed;
- if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
- reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
- reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
- goto failed;
+ /* Future protocol version compatibility support! If the
+ * o2iblnd-specific protocol changes, or when LNET unifies
+ * protocols over all LNDs, the initial connection will
+ * negotiate a protocol version. I trap this here to avoid
+ * console errors; the reject tells the peer_ni which protocol I
+ * speak. */
+ if (reqmsg->ibm_magic == LNET_PROTO_MAGIC ||
+ reqmsg->ibm_magic == __swab32(LNET_PROTO_MAGIC))
+ goto failed;
+ if (reqmsg->ibm_magic == IBLND_MSG_MAGIC &&
+ reqmsg->ibm_version != IBLND_MSG_VERSION &&
+ reqmsg->ibm_version != IBLND_MSG_VERSION_1)
+ goto failed;
+ if (reqmsg->ibm_magic == __swab32(IBLND_MSG_MAGIC) &&
+ reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION) &&
+ reqmsg->ibm_version != __swab16(IBLND_MSG_VERSION_1))
+ goto failed;
- rc = kiblnd_unpack_msg(reqmsg, priv_nob);
- if (rc != 0) {
- CERROR("Can't parse connection request: %d\n", rc);
- goto failed;
- }
+ rc = kiblnd_unpack_msg(reqmsg, priv_nob);
+ if (rc != 0) {
+ CERROR("Can't parse connection request: %d\n", rc);
+ goto failed;
+ }
- nid = reqmsg->ibm_srcnid;
- ni = lnet_net2ni(LNET_NIDNET(reqmsg->ibm_dstnid));
+ nid = reqmsg->ibm_srcnid;
+ ni = lnet_nid2ni_addref(reqmsg->ibm_dstnid);
- if (ni != NULL) {
- net = (kib_net_t *)ni->ni_data;
- rej.ibr_incarnation = net->ibn_incarnation;
- }
+ if (ni != NULL) {
+ net = (kib_net_t *)ni->ni_data;
+ rej.ibr_incarnation = net->ibn_incarnation;
+ }
- if (ni == NULL || /* no matching net */
- ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
- net->ibn_dev != ibdev) { /* wrong device */
+ if (ni == NULL || /* no matching net */
+ ni->ni_nid != reqmsg->ibm_dstnid || /* right NET, wrong NID! */
+ net->ibn_dev != ibdev) { /* wrong device */
CERROR("Can't accept conn from %s on %s (%s:%d:%pI4h): "
- "bad dst nid %s\n", libcfs_nid2str(nid),
- ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
- ibdev->ibd_ifname, ibdev->ibd_nnets,
+ "bad dst nid %s\n", libcfs_nid2str(nid),
+ ni == NULL ? "NA" : libcfs_nid2str(ni->ni_nid),
+ ibdev->ibd_ifname, ibdev->ibd_nnets,
&ibdev->ibd_ifip,
- libcfs_nid2str(reqmsg->ibm_dstnid));
+ libcfs_nid2str(reqmsg->ibm_dstnid));
- goto failed;
- }
+ goto failed;
+ }
/* check time stamp as soon as possible */
- if (reqmsg->ibm_dststamp != 0 &&
- reqmsg->ibm_dststamp != net->ibn_incarnation) {
- CWARN("Stale connection request\n");
- rej.ibr_why = IBLND_REJECT_CONN_STALE;
- goto failed;
- }
+ if (reqmsg->ibm_dststamp != 0 &&
+ reqmsg->ibm_dststamp != net->ibn_incarnation) {
+ CWARN("Stale connection request\n");
+ rej.ibr_why = IBLND_REJECT_CONN_STALE;
+ goto failed;
+ }
- /* I can accept peer's version */
- version = reqmsg->ibm_version;
+ /* I can accept peer_ni's version */
+ version = reqmsg->ibm_version;
- if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
- CERROR("Unexpected connreq msg type: %x from %s\n",
- reqmsg->ibm_type, libcfs_nid2str(nid));
- goto failed;
- }
+ if (reqmsg->ibm_type != IBLND_MSG_CONNREQ) {
+ CERROR("Unexpected connreq msg type: %x from %s\n",
+ reqmsg->ibm_type, libcfs_nid2str(nid));
+ goto failed;
+ }
if (reqmsg->ibm_u.connparams.ibcp_queue_depth >
- IBLND_MSG_QUEUE_SIZE(version)) {
+ kiblnd_msg_queue_size(version, ni)) {
CERROR("Can't accept conn from %s, queue depth too large: "
" %d (<=%d wanted)\n",
libcfs_nid2str(nid),
reqmsg->ibm_u.connparams.ibcp_queue_depth,
- IBLND_MSG_QUEUE_SIZE(version));
+ kiblnd_msg_queue_size(version, ni));
if (version == IBLND_MSG_VERSION)
rej.ibr_why = IBLND_REJECT_MSG_QUEUE_SIZE;
}
if (reqmsg->ibm_u.connparams.ibcp_max_frags >
- IBLND_RDMA_FRAGS(version)) {
+ kiblnd_rdma_frags(version, ni)) {
CWARN("Can't accept conn from %s (version %x): "
"max_frags %d too large (%d wanted)\n",
- libcfs_nid2str(nid), version,
- reqmsg->ibm_u.connparams.ibcp_max_frags,
- IBLND_RDMA_FRAGS(version));
+ libcfs_nid2str(nid), version,
+ reqmsg->ibm_u.connparams.ibcp_max_frags,
+ kiblnd_rdma_frags(version, ni));
if (version >= IBLND_MSG_VERSION)
rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
goto failed;
} else if (reqmsg->ibm_u.connparams.ibcp_max_frags <
- IBLND_RDMA_FRAGS(version) && net->ibn_fmr_ps == NULL) {
+ kiblnd_rdma_frags(version, ni) &&
+ net->ibn_fmr_ps == NULL) {
CWARN("Can't accept conn from %s (version %x): "
"max_frags %d incompatible without FMR pool "
"(%d wanted)\n",
libcfs_nid2str(nid), version,
reqmsg->ibm_u.connparams.ibcp_max_frags,
- IBLND_RDMA_FRAGS(version));
+ kiblnd_rdma_frags(version, ni));
- if (version >= IBLND_MSG_VERSION)
+ if (version == IBLND_MSG_VERSION)
rej.ibr_why = IBLND_REJECT_RDMA_FRAGS;
goto failed;
goto failed;
}
- /* assume 'nid' is a new peer; create */
- rc = kiblnd_create_peer(ni, &peer, nid);
+ /* assume 'nid' is a new peer_ni; create */
+ rc = kiblnd_create_peer(ni, &peer_ni, nid);
if (rc != 0) {
- CERROR("Can't create peer for %s\n", libcfs_nid2str(nid));
+ CERROR("Can't create peer_ni for %s\n", libcfs_nid2str(nid));
rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
goto failed;
}
/* We have validated the peer's parameters so use those */
- peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
- peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
+ peer_ni->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
+ peer_ni->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
write_lock_irqsave(g_lock, flags);
- peer2 = kiblnd_find_peer_locked(nid);
+ peer2 = kiblnd_find_peer_locked(ni, nid);
if (peer2 != NULL) {
if (peer2->ibp_version == 0) {
peer2->ibp_version = version;
/* not the guy I've talked with */
if (peer2->ibp_incarnation != reqmsg->ibm_srcstamp ||
peer2->ibp_version != version) {
- kiblnd_close_peer_conns_locked(peer2, -ESTALE);
+ kiblnd_close_peer_conns_locked(peer2, -ESTALE);
+
+ if (kiblnd_peer_active(peer2)) {
+ peer2->ibp_incarnation = reqmsg->ibm_srcstamp;
+ peer2->ibp_version = version;
+ }
write_unlock_irqrestore(g_lock, flags);
- CWARN("Conn stale %s [old ver: %x, new ver: %x]\n",
- libcfs_nid2str(nid), peer2->ibp_version, version);
+ CWARN("Conn stale %s version %x/%x incarnation %llu/%llu\n",
+ libcfs_nid2str(nid), peer2->ibp_version, version,
+ peer2->ibp_incarnation, reqmsg->ibm_srcstamp);
- kiblnd_peer_decref(peer);
+ kiblnd_peer_decref(peer_ni);
rej.ibr_why = IBLND_REJECT_CONN_STALE;
goto failed;
}
- /* tie-break connection race in favour of the higher NID */
- if (peer2->ibp_connecting != 0 &&
- nid < ni->ni_nid) {
+ /* Tie-break connection race in favour of the higher NID.
+ * If we keep running into a race condition multiple times,
+ * we have to assume that the connection attempt with the
+ * higher NID is stuck in a connecting state and will never
+ * recover. As such, we pass through this if-block and let
+ * the lower NID connection win so we can move forward.
+ */
+ if (peer2->ibp_connecting != 0 &&
+ nid < ni->ni_nid && peer2->ibp_races <
+ MAX_CONN_RACES_BEFORE_ABORT) {
+ peer2->ibp_races++;
write_unlock_irqrestore(g_lock, flags);
- CWARN("Conn race %s\n", libcfs_nid2str(peer2->ibp_nid));
-
- kiblnd_peer_decref(peer);
- rej.ibr_why = IBLND_REJECT_CONN_RACE;
- goto failed;
- }
-
- peer2->ibp_accepting++;
- kiblnd_peer_addref(peer2);
+ CDEBUG(D_NET, "Conn race %s\n",
+ libcfs_nid2str(peer2->ibp_nid));
- /* Race with kiblnd_launch_tx (active connect) to create peer
+ kiblnd_peer_decref(peer_ni);
+ rej.ibr_why = IBLND_REJECT_CONN_RACE;
+ goto failed;
+ }
+ if (peer2->ibp_races >= MAX_CONN_RACES_BEFORE_ABORT)
+ CNETERR("Conn race %s: unresolved after %d attempts, letting lower NID win\n",
+ libcfs_nid2str(peer2->ibp_nid),
+ MAX_CONN_RACES_BEFORE_ABORT);
+ /*
+ * passive connection is allowed even this peer_ni is waiting for
+ * reconnection.
+ */
+ peer2->ibp_reconnecting = 0;
+ peer2->ibp_races = 0;
+ peer2->ibp_accepting++;
+ kiblnd_peer_addref(peer2);
+
+ /* Race with kiblnd_launch_tx (active connect) to create peer_ni
* so copy validated parameters since we now know what the
- * peer's limits are */
- peer2->ibp_max_frags = peer->ibp_max_frags;
- peer2->ibp_queue_depth = peer->ibp_queue_depth;
+ * peer_ni's limits are */
+ peer2->ibp_max_frags = peer_ni->ibp_max_frags;
+ peer2->ibp_queue_depth = peer_ni->ibp_queue_depth;
write_unlock_irqrestore(g_lock, flags);
- kiblnd_peer_decref(peer);
- peer = peer2;
+ kiblnd_peer_decref(peer_ni);
+ peer_ni = peer2;
} else {
- /* Brand new peer */
- LASSERT (peer->ibp_accepting == 0);
- LASSERT (peer->ibp_version == 0 &&
- peer->ibp_incarnation == 0);
+ /* Brand new peer_ni */
+ LASSERT (peer_ni->ibp_accepting == 0);
+ LASSERT (peer_ni->ibp_version == 0 &&
+ peer_ni->ibp_incarnation == 0);
- peer->ibp_accepting = 1;
- peer->ibp_version = version;
- peer->ibp_incarnation = reqmsg->ibm_srcstamp;
+ peer_ni->ibp_accepting = 1;
+ peer_ni->ibp_version = version;
+ peer_ni->ibp_incarnation = reqmsg->ibm_srcstamp;
/* I have a ref on ni that prevents it being shutdown */
LASSERT (net->ibn_shutdown == 0);
- kiblnd_peer_addref(peer);
- list_add_tail(&peer->ibp_list, kiblnd_nid2peerlist(nid));
+ kiblnd_peer_addref(peer_ni);
+ list_add_tail(&peer_ni->ibp_list, kiblnd_nid2peerlist(nid));
write_unlock_irqrestore(g_lock, flags);
}
- conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
+ conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_PASSIVE_WAIT, version);
if (conn == NULL) {
- kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
- kiblnd_peer_decref(peer);
+ kiblnd_peer_connect_failed(peer_ni, 0, -ENOMEM);
+ kiblnd_peer_decref(peer_ni);
rej.ibr_why = IBLND_REJECT_NO_RESOURCES;
goto failed;
}
return 0;
failed:
- if (ni != NULL)
- lnet_ni_decref(ni);
+ if (ni != NULL) {
+ rej.ibr_cp.ibcp_queue_depth =
+ kiblnd_msg_queue_size(version, ni);
+ rej.ibr_cp.ibcp_max_frags = kiblnd_rdma_frags(version, ni);
+ lnet_ni_decref(ni);
+ }
- rej.ibr_version = version;
- rej.ibr_cp.ibcp_queue_depth = IBLND_MSG_QUEUE_SIZE(version);
- rej.ibr_cp.ibcp_max_frags = IBLND_RDMA_FRAGS(version);
- kiblnd_reject(cmid, &rej);
+ rej.ibr_version = version;
+ kiblnd_reject(cmid, &rej);
- return -ECONNREFUSED;
+ return -ECONNREFUSED;
}
static void
-kiblnd_reconnect (kib_conn_t *conn, int version,
- __u64 incarnation, int why, kib_connparams_t *cp)
+kiblnd_check_reconnect(kib_conn_t *conn, int version,
+ __u64 incarnation, int why, kib_connparams_t *cp)
{
- kib_peer_t *peer = conn->ibc_peer;
+ rwlock_t *glock = &kiblnd_data.kib_global_lock;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
char *reason;
- int retry_now = 0;
+ int msg_size = IBLND_MSG_SIZE;
+ int frag_num = -1;
+ int queue_dep = -1;
+ bool reconnect;
unsigned long flags;
- LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
- LASSERT (peer->ibp_connecting > 0); /* 'conn' at least */
+ LASSERT(conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
+ LASSERT(peer_ni->ibp_connecting > 0); /* 'conn' at least */
+ LASSERT(!peer_ni->ibp_reconnecting);
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+ if (cp) {
+ msg_size = cp->ibcp_max_msg_size;
+ frag_num = cp->ibcp_max_frags;
+ queue_dep = cp->ibcp_queue_depth;
+ }
+ write_lock_irqsave(glock, flags);
/* retry connection if it's still needed and no other connection
* attempts (active or passive) are in progress
* NB: reconnect is still needed even when ibp_tx_queue is
* empty if ibp_version != version because reconnect may be
* initiated by kiblnd_query() */
- if ((!list_empty(&peer->ibp_tx_queue) ||
- peer->ibp_version != version) &&
- peer->ibp_connecting == 1 &&
- peer->ibp_accepting == 0) {
- if (why == IBLND_REJECT_CONN_RACE) {
- /* don't reconnect immediately, intensive reconnecting
- * may consume a lot of memory. kiblnd_destroy_conn
- * will reconnect after releasing all resources of
- * this connection */
- conn->ibc_conn_race = 1;
- } else {
- retry_now = 1;
- }
- peer->ibp_connecting++;
- peer->ibp_version = version;
- peer->ibp_incarnation = incarnation;
- }
-
- write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
-
- if (!retry_now)
- return;
+ reconnect = (!list_empty(&peer_ni->ibp_tx_queue) ||
+ peer_ni->ibp_version != version) &&
+ peer_ni->ibp_connecting == 1 &&
+ peer_ni->ibp_accepting == 0;
+ if (!reconnect) {
+ reason = "no need";
+ goto out;
+ }
switch (why) {
default:
reason = "Unknown";
break;
- case IBLND_REJECT_RDMA_FRAGS:
- if (!cp)
- goto failed;
+ case IBLND_REJECT_RDMA_FRAGS: {
+ struct lnet_ioctl_config_o2iblnd_tunables *tunables;
- if (conn->ibc_max_frags <= cp->ibcp_max_frags) {
- CNETERR("Unsupported max frags, peer supports %d\n",
- cp->ibcp_max_frags);
- goto failed;
- } else if (*kiblnd_tunables.kib_map_on_demand == 0) {
- CNETERR("map_on_demand must be enabled to support "
- "map_on_demand peers\n");
- goto failed;
+ if (!cp) {
+ reason = "can't negotiate max frags";
+ goto out;
+ }
+ tunables = &peer_ni->ibp_ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
+ if (!tunables->lnd_map_on_demand) {
+ reason = "map_on_demand must be enabled";
+ goto out;
+ }
+ if (conn->ibc_max_frags <= frag_num) {
+ reason = "unsupported max frags";
+ goto out;
}
- peer->ibp_max_frags = cp->ibcp_max_frags;
+ peer_ni->ibp_max_frags = frag_num;
reason = "rdma fragments";
break;
-
+ }
case IBLND_REJECT_MSG_QUEUE_SIZE:
- if (!cp)
- goto failed;
-
- if (conn->ibc_queue_depth <= cp->ibcp_queue_depth) {
- CNETERR("Unsupported queue depth, peer supports %d\n",
- cp->ibcp_queue_depth);
- goto failed;
+ if (!cp) {
+ reason = "can't negotiate queue depth";
+ goto out;
+ }
+ if (conn->ibc_queue_depth <= queue_dep) {
+ reason = "unsupported queue depth";
+ goto out;
}
- peer->ibp_queue_depth = cp->ibcp_queue_depth;
+ peer_ni->ibp_queue_depth = queue_dep;
reason = "queue depth";
break;
reason = "stale";
break;
+ case IBLND_REJECT_CONN_RACE:
+ reason = "conn race";
+ break;
+
case IBLND_REJECT_CONN_UNCOMPAT:
reason = "version negotiation";
break;
}
- CNETERR("%s: retrying (%s), %x, %x, "
- "queue_depth: %d, max_frags: %d, msg_size: %d\n",
- libcfs_nid2str(peer->ibp_nid),
- reason, IBLND_MSG_VERSION, version,
- conn->ibc_queue_depth, conn->ibc_max_frags,
- cp != NULL ? cp->ibcp_max_msg_size : IBLND_MSG_SIZE);
-
- kiblnd_connect_peer(peer);
- return;
-
- failed:
- write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
- peer->ibp_connecting--;
+ conn->ibc_reconnect = 1;
+ peer_ni->ibp_reconnecting = 1;
+ peer_ni->ibp_version = version;
+ if (incarnation != 0)
+ peer_ni->ibp_incarnation = incarnation;
+ out:
write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- return;
+ CNETERR("%s: %s (%s), %x, %x, msg_size: %d, queue_depth: %d/%d, max_frags: %d/%d\n",
+ libcfs_nid2str(peer_ni->ibp_nid),
+ reconnect ? "reconnect" : "don't reconnect",
+ reason, IBLND_MSG_VERSION, version, msg_size,
+ conn->ibc_queue_depth, queue_dep,
+ conn->ibc_max_frags, frag_num);
+ /*
+ * if conn::ibc_reconnect is TRUE, connd will reconnect to the peer_ni
+ * while destroying the zombie
+ */
}
static void
kiblnd_rejected (kib_conn_t *conn, int reason, void *priv, int priv_nob)
{
- kib_peer_t *peer = conn->ibc_peer;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
LASSERT (!in_interrupt());
LASSERT (conn->ibc_state == IBLND_CONN_ACTIVE_CONNECT);
switch (reason) {
case IB_CM_REJ_STALE_CONN:
- kiblnd_reconnect(conn, IBLND_MSG_VERSION, 0,
- IBLND_REJECT_CONN_STALE, NULL);
+ kiblnd_check_reconnect(conn, IBLND_MSG_VERSION, 0,
+ IBLND_REJECT_CONN_STALE, NULL);
break;
case IB_CM_REJ_INVALID_SERVICE_ID:
CNETERR("%s rejected: no listener at %d\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
*kiblnd_tunables.kib_service);
break;
* b) V2 will provide incarnation while rejecting me,
* -1 will be overwrote.
*
- * if I try to connect to a V1 peer with V2 protocol,
+ * if I try to connect to a V1 peer_ni with V2 protocol,
* it rejected me then upgrade to V2, I have no idea
* about the upgrading and try to reconnect with V1,
* in this case upgraded V2 can find out I'm trying to
if (rej->ibr_magic != IBLND_MSG_MAGIC &&
rej->ibr_magic != LNET_PROTO_MAGIC) {
CERROR("%s rejected: consumer defined fatal error\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
break;
}
if (rej->ibr_version != IBLND_MSG_VERSION &&
rej->ibr_version != IBLND_MSG_VERSION_1) {
CERROR("%s rejected: o2iblnd version %x error\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
rej->ibr_version);
break;
}
if (rej->ibr_why == IBLND_REJECT_FATAL &&
rej->ibr_version == IBLND_MSG_VERSION_1) {
- CDEBUG(D_NET, "rejected by old version peer %s: %x\n",
- libcfs_nid2str(peer->ibp_nid), rej->ibr_version);
+ CDEBUG(D_NET, "rejected by old version peer_ni %s: %x\n",
+ libcfs_nid2str(peer_ni->ibp_nid), rej->ibr_version);
if (conn->ibc_version != IBLND_MSG_VERSION_1)
rej->ibr_why = IBLND_REJECT_CONN_UNCOMPAT;
case IBLND_REJECT_CONN_UNCOMPAT:
case IBLND_REJECT_MSG_QUEUE_SIZE:
case IBLND_REJECT_RDMA_FRAGS:
- kiblnd_reconnect(conn, rej->ibr_version,
- incarnation, rej->ibr_why, cp);
+ kiblnd_check_reconnect(conn, rej->ibr_version,
+ incarnation, rej->ibr_why, cp);
break;
case IBLND_REJECT_NO_RESOURCES:
CERROR("%s rejected: o2iblnd no resources\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
break;
case IBLND_REJECT_FATAL:
CERROR("%s rejected: o2iblnd fatal error\n",
- libcfs_nid2str(peer->ibp_nid));
+ libcfs_nid2str(peer_ni->ibp_nid));
break;
default:
CERROR("%s rejected: o2iblnd reason %d\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
rej->ibr_why);
break;
}
/* fall through */
default:
CNETERR("%s rejected: reason %d, size %d\n",
- libcfs_nid2str(peer->ibp_nid), reason, priv_nob);
+ libcfs_nid2str(peer_ni->ibp_nid), reason, priv_nob);
break;
}
static void
kiblnd_check_connreply (kib_conn_t *conn, void *priv, int priv_nob)
{
- kib_peer_t *peer = conn->ibc_peer;
- lnet_ni_t *ni = peer->ibp_ni;
+ kib_peer_ni_t *peer_ni = conn->ibc_peer;
+ lnet_ni_t *ni = peer_ni->ibp_ni;
kib_net_t *net = ni->ni_data;
kib_msg_t *msg = priv;
int ver = conn->ibc_version;
if (rc != 0) {
CERROR("Can't unpack connack from %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), rc);
+ libcfs_nid2str(peer_ni->ibp_nid), rc);
goto failed;
}
if (msg->ibm_type != IBLND_MSG_CONNACK) {
CERROR("Unexpected message %d from %s\n",
- msg->ibm_type, libcfs_nid2str(peer->ibp_nid));
+ msg->ibm_type, libcfs_nid2str(peer_ni->ibp_nid));
rc = -EPROTO;
goto failed;
}
if (ver != msg->ibm_version) {
CERROR("%s replied version %x is different with "
"requested version %x\n",
- libcfs_nid2str(peer->ibp_nid), msg->ibm_version, ver);
+ libcfs_nid2str(peer_ni->ibp_nid), msg->ibm_version, ver);
rc = -EPROTO;
goto failed;
}
if (msg->ibm_u.connparams.ibcp_queue_depth >
conn->ibc_queue_depth) {
CERROR("%s has incompatible queue depth %d (<=%d wanted)\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
msg->ibm_u.connparams.ibcp_queue_depth,
conn->ibc_queue_depth);
rc = -EPROTO;
if (msg->ibm_u.connparams.ibcp_max_frags >
conn->ibc_max_frags) {
CERROR("%s has incompatible max_frags %d (<=%d wanted)\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
msg->ibm_u.connparams.ibcp_max_frags,
conn->ibc_max_frags);
rc = -EPROTO;
if (msg->ibm_u.connparams.ibcp_max_msg_size > IBLND_MSG_SIZE) {
CERROR("%s max message size %d too big (%d max)\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
msg->ibm_u.connparams.ibcp_max_msg_size,
IBLND_MSG_SIZE);
rc = -EPROTO;
if (rc != 0) {
CERROR("Bad connection reply from %s, rc = %d, "
"version: %x max_frags: %d\n",
- libcfs_nid2str(peer->ibp_nid), rc,
+ libcfs_nid2str(peer_ni->ibp_nid), rc,
msg->ibm_version, msg->ibm_u.connparams.ibcp_max_frags);
goto failed;
}
static int
kiblnd_active_connect (struct rdma_cm_id *cmid)
{
- kib_peer_t *peer = (kib_peer_t *)cmid->context;
+ kib_peer_ni_t *peer_ni = (kib_peer_ni_t *)cmid->context;
kib_conn_t *conn;
kib_msg_t *msg;
struct rdma_conn_param cp;
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
- incarnation = peer->ibp_incarnation;
- version = (peer->ibp_version == 0) ? IBLND_MSG_VERSION :
- peer->ibp_version;
+ incarnation = peer_ni->ibp_incarnation;
+ version = (peer_ni->ibp_version == 0) ? IBLND_MSG_VERSION :
+ peer_ni->ibp_version;
read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
- conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
+ conn = kiblnd_create_conn(peer_ni, cmid, IBLND_CONN_ACTIVE_CONNECT,
version);
if (conn == NULL) {
- kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
- kiblnd_peer_decref(peer); /* lose cmid's ref */
+ kiblnd_peer_connect_failed(peer_ni, 1, -ENOMEM);
+ kiblnd_peer_decref(peer_ni); /* lose cmid's ref */
return -ENOMEM;
}
/* conn "owns" cmid now, so I return success from here on to ensure the
* CM callback doesn't destroy cmid. conn also takes over cmid's ref
- * on peer */
+ * on peer_ni */
msg = &conn->ibc_connvars->cv_msg;
msg->ibm_u.connparams.ibcp_max_frags = conn->ibc_max_frags;
msg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
- kiblnd_pack_msg(peer->ibp_ni, msg, version,
- 0, peer->ibp_nid, incarnation);
+ kiblnd_pack_msg(peer_ni->ibp_ni, msg, version,
+ 0, peer_ni->ibp_nid, incarnation);
memset(&cp, 0, sizeof(cp));
cp.private_data = msg;
rc = rdma_connect(cmid, &cp);
if (rc != 0) {
CERROR("Can't connect to %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), rc);
+ libcfs_nid2str(peer_ni->ibp_nid), rc);
kiblnd_connreq_done(conn, rc);
kiblnd_conn_decref(conn);
}
int
kiblnd_cm_callback(struct rdma_cm_id *cmid, struct rdma_cm_event *event)
{
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
kib_conn_t *conn;
int rc;
return rc;
case RDMA_CM_EVENT_ADDR_ERROR:
- peer = (kib_peer_t *)cmid->context;
+ peer_ni = (kib_peer_ni_t *)cmid->context;
CNETERR("%s: ADDR ERROR %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
- kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
- kiblnd_peer_decref(peer);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
+ kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
+ kiblnd_peer_decref(peer_ni);
return -EHOSTUNREACH; /* rc != 0 destroys cmid */
case RDMA_CM_EVENT_ADDR_RESOLVED:
- peer = (kib_peer_t *)cmid->context;
+ peer_ni = (kib_peer_ni_t *)cmid->context;
CDEBUG(D_NET,"%s Addr resolved: %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
if (event->status != 0) {
CNETERR("Can't resolve address for %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
rc = event->status;
} else {
rc = rdma_resolve_route(
return 0;
/* Can't initiate route resolution */
CERROR("Can't resolve route for %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), rc);
+ libcfs_nid2str(peer_ni->ibp_nid), rc);
}
- kiblnd_peer_connect_failed(peer, 1, rc);
- kiblnd_peer_decref(peer);
+ kiblnd_peer_connect_failed(peer_ni, 1, rc);
+ kiblnd_peer_decref(peer_ni);
return rc; /* rc != 0 destroys cmid */
case RDMA_CM_EVENT_ROUTE_ERROR:
- peer = (kib_peer_t *)cmid->context;
+ peer_ni = (kib_peer_ni_t *)cmid->context;
CNETERR("%s: ROUTE ERROR %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
- kiblnd_peer_connect_failed(peer, 1, -EHOSTUNREACH);
- kiblnd_peer_decref(peer);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
+ kiblnd_peer_connect_failed(peer_ni, 1, -EHOSTUNREACH);
+ kiblnd_peer_decref(peer_ni);
return -EHOSTUNREACH; /* rc != 0 destroys cmid */
case RDMA_CM_EVENT_ROUTE_RESOLVED:
- peer = (kib_peer_t *)cmid->context;
+ peer_ni = (kib_peer_ni_t *)cmid->context;
CDEBUG(D_NET,"%s Route resolved: %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
if (event->status == 0)
return kiblnd_active_connect(cmid);
CNETERR("Can't resolve route for %s: %d\n",
- libcfs_nid2str(peer->ibp_nid), event->status);
- kiblnd_peer_connect_failed(peer, 1, event->status);
- kiblnd_peer_decref(peer);
+ libcfs_nid2str(peer_ni->ibp_nid), event->status);
+ kiblnd_peer_connect_failed(peer_ni, 1, event->status);
+ kiblnd_peer_decref(peer_ni);
return event->status; /* rc != 0 destroys cmid */
case RDMA_CM_EVENT_UNREACHABLE:
struct list_head checksends = LIST_HEAD_INIT(checksends);
struct list_head *peers = &kiblnd_data.kib_peers[idx];
struct list_head *ptmp;
- kib_peer_t *peer;
+ kib_peer_ni_t *peer_ni;
kib_conn_t *conn;
struct list_head *ctmp;
unsigned long flags;
- /* NB. We expect to have a look at all the peers and not find any
- * RDMAs to time out, so we just use a shared lock while we
- * take a look... */
+ /* NB. We expect to have a look at all the peers and not find any
+ * RDMAs to time out, so we just use a shared lock while we
+ * take a look... */
read_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
list_for_each(ptmp, peers) {
- peer = list_entry(ptmp, kib_peer_t, ibp_list);
+ peer_ni = list_entry(ptmp, kib_peer_ni_t, ibp_list);
- list_for_each(ctmp, &peer->ibp_conns) {
+ list_for_each(ctmp, &peer_ni->ibp_conns) {
int timedout;
int sendnoop;
if (timedout) {
CERROR("Timed out RDMA with %s (%lu): "
"c: %u, oc: %u, rc: %u\n",
- libcfs_nid2str(peer->ibp_nid),
+ libcfs_nid2str(peer_ni->ibp_nid),
cfs_duration_sec(cfs_time_current() -
- peer->ibp_last_alive),
+ peer_ni->ibp_last_alive),
conn->ibc_credits,
conn->ibc_outstanding_credits,
conn->ibc_reserved_credits);
conn = list_entry(checksends.next,
kib_conn_t, ibc_connd_list);
list_del(&conn->ibc_connd_list);
- kiblnd_check_sends(conn);
+
+ spin_lock(&conn->ibc_lock);
+ kiblnd_check_sends_locked(conn);
+ spin_unlock(&conn->ibc_lock);
+
kiblnd_conn_decref(conn);
}
}
kiblnd_peer_notify(conn->ibc_peer);
}
+/*
+ * High-water for reconnection to the same peer_ni, reconnection attempt should
+ * be delayed after trying more than KIB_RECONN_HIGH_RACE.
+ */
+#define KIB_RECONN_HIGH_RACE 10
+/*
+ * Allow connd to take a break and handle other things after consecutive
+ * reconnection attemps.
+ */
+#define KIB_RECONN_BREAK 100
+
int
kiblnd_connd (void *arg)
{
- wait_queue_t wait;
+ spinlock_t *lock= &kiblnd_data.kib_connd_lock;
+ wait_queue_t wait;
unsigned long flags;
kib_conn_t *conn;
int timeout;
init_waitqueue_entry(&wait, current);
kiblnd_data.kib_connd = current;
- spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+ spin_lock_irqsave(lock, flags);
- while (!kiblnd_data.kib_shutdown) {
+ while (!kiblnd_data.kib_shutdown) {
+ int reconn = 0;
dropped_lock = 0;
if (!list_empty(&kiblnd_data.kib_connd_zombies)) {
- conn = list_entry(kiblnd_data. \
- kib_connd_zombies.next,
- kib_conn_t, ibc_list);
+ kib_peer_ni_t *peer_ni = NULL;
+
+ conn = list_entry(kiblnd_data.kib_connd_zombies.next,
+ kib_conn_t, ibc_list);
list_del(&conn->ibc_list);
+ if (conn->ibc_reconnect) {
+ peer_ni = conn->ibc_peer;
+ kiblnd_peer_addref(peer_ni);
+ }
- spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
- flags);
+ spin_unlock_irqrestore(lock, flags);
dropped_lock = 1;
- kiblnd_destroy_conn(conn);
+ kiblnd_destroy_conn(conn, !peer_ni);
- spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+ spin_lock_irqsave(lock, flags);
+ if (!peer_ni)
+ continue;
+
+ conn->ibc_peer = peer_ni;
+ if (peer_ni->ibp_reconnected < KIB_RECONN_HIGH_RACE)
+ list_add_tail(&conn->ibc_list,
+ &kiblnd_data.kib_reconn_list);
+ else
+ list_add_tail(&conn->ibc_list,
+ &kiblnd_data.kib_reconn_wait);
}
if (!list_empty(&kiblnd_data.kib_connd_conns)) {
kib_conn_t, ibc_list);
list_del(&conn->ibc_list);
- spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock,
- flags);
+ spin_unlock_irqrestore(lock, flags);
dropped_lock = 1;
kiblnd_disconnect_conn(conn);
kiblnd_conn_decref(conn);
- spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+ spin_lock_irqsave(lock, flags);
}
+ while (reconn < KIB_RECONN_BREAK) {
+ if (kiblnd_data.kib_reconn_sec !=
+ ktime_get_real_seconds()) {
+ kiblnd_data.kib_reconn_sec = ktime_get_real_seconds();
+ list_splice_init(&kiblnd_data.kib_reconn_wait,
+ &kiblnd_data.kib_reconn_list);
+ }
+
+ if (list_empty(&kiblnd_data.kib_reconn_list))
+ break;
+
+ conn = list_entry(kiblnd_data.kib_reconn_list.next,
+ kib_conn_t, ibc_list);
+ list_del(&conn->ibc_list);
+
+ spin_unlock_irqrestore(lock, flags);
+ dropped_lock = 1;
+
+ reconn += kiblnd_reconnect_peer(conn->ibc_peer);
+ kiblnd_peer_decref(conn->ibc_peer);
+ LIBCFS_FREE(conn, sizeof(*conn));
+
+ spin_lock_irqsave(lock, flags);
+ }
+
/* careful with the jiffy wrap... */
timeout = (int)(deadline - jiffies);
if (timeout <= 0) {
const int p = 1;
int chunk = kiblnd_data.kib_peer_hash_size;
- spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+ spin_unlock_irqrestore(lock, flags);
dropped_lock = 1;
/* Time to check for RDMA timeouts on a few more
* peers: I do checks every 'p' seconds on a
- * proportion of the peer table and I need to check
+ * proportion of the peer_ni 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
}
deadline += msecs_to_jiffies(p * MSEC_PER_SEC);
- spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+ spin_lock_irqsave(lock, flags);
}
if (dropped_lock)
/* Nothing to do for 'timeout' */
set_current_state(TASK_INTERRUPTIBLE);
add_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
- spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+ spin_unlock_irqrestore(lock, flags);
schedule_timeout(timeout);
set_current_state(TASK_RUNNING);
remove_wait_queue(&kiblnd_data.kib_connd_waitq, &wait);
- spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags);
+ spin_lock_irqsave(lock, flags);
}
- spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);
+ spin_unlock_irqrestore(lock, flags);
kiblnd_thread_fini();
return 0;
case IB_EVENT_COMM_EST:
CDEBUG(D_NET, "%s established\n",
libcfs_nid2str(conn->ibc_peer->ibp_nid));
+ /* We received a packet but connection isn't established
+ * probably handshake packet was lost, so free to
+ * force make connection established */
+ rdma_notify(conn->ibc_cmid, IB_EVENT_COMM_EST);
return;
default:
static void
kiblnd_complete (struct ib_wc *wc)
{
- switch (kiblnd_wreqid2type(wc->wr_id)) {
- default:
- LBUG();
+ switch (kiblnd_wreqid2type(wc->wr_id)) {
+ default:
+ LBUG();
+
+ case IBLND_WID_MR:
+ if (wc->status != IB_WC_SUCCESS &&
+ wc->status != IB_WC_WR_FLUSH_ERR)
+ CNETERR("FastReg failed: %d\n", wc->status);
+ return;
case IBLND_WID_RDMA:
/* We only get RDMA completion notification if it fails. All
rc = cfs_cpt_bind(lnet_cpt_table(), sched->ibs_cpt);
if (rc != 0) {
- CWARN("Failed to bind on CPT %d, please verify whether "
- "all CPUs are healthy and reload modules if necessary, "
- "otherwise your system might under risk of low "
- "performance\n", sched->ibs_cpt);
+ CWARN("Unable to bind on CPU partition %d, please verify "
+ "whether all CPUs are healthy and reload modules if "
+ "necessary, otherwise your system might under risk of "
+ "low performance\n", sched->ibs_cpt);
}
spin_lock_irqsave(&sched->ibs_lock, flags);
spin_unlock_irqrestore(&sched->ibs_lock, flags);
+ wc.wr_id = IBLND_WID_INVAL;
+
rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
if (rc == 0) {
rc = ib_req_notify_cq(conn->ibc_cq,
rc = ib_poll_cq(conn->ibc_cq, 1, &wc);
}
+ if (unlikely(rc > 0 && wc.wr_id == IBLND_WID_INVAL)) {
+ LCONSOLE_ERROR(
+ "ib_poll_cq (rc: %d) returned invalid "
+ "wr_id, opcode %d, status: %d, "
+ "vendor_err: %d, conn: %s status: %d\n"
+ "please upgrade firmware and OFED or "
+ "contact vendor.\n", rc,
+ wc.opcode, wc.status, wc.vendor_err,
+ libcfs_nid2str(conn->ibc_peer->ibp_nid),
+ conn->ibc_state);
+ rc = -EINVAL;
+ }
+
if (rc < 0) {
CWARN("%s: ib_poll_cq failed: %d, "
"closing connection\n",