/* Only free the buffer if we used it */
if (tx->tx_buffer_copy != NULL) {
- vfree(tx->tx_buffer_copy);
+ kgnilnd_vfree(tx->tx_buffer_copy, tx->tx_rdma_desc.length);
tx->tx_buffer_copy = NULL;
CDEBUG(D_MALLOC, "vfreed buffer2\n");
}
return sum;
}
-inline __u16
+__u16
kgnilnd_cksum_kiov(unsigned int nkiov, lnet_kiov_t *kiov,
unsigned int offset, unsigned int nob, int dump_blob)
{
LBUG();
}
/* only allow NAK on error and truncate to zero */
- LASSERTF(error <= 0, "error %d conn 0x%p, cookie "LPU64"\n",
+ LASSERTF(error <= 0, "error %d conn 0x%p, cookie %llu\n",
error, conn, cookie);
tx = kgnilnd_new_tx_msg(nak_type, source);
if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_ACK ||
tx->tx_msg.gnm_type == GNILND_MSG_GET_REQ) {
atomic64_add(bytes, &dev->gnd_rdmaq_bytes_out);
- GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to "LPD64"",
+ GNIDBG_TX(D_NETTRACE, tx, "rdma ++ %d to %lld",
bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
}
atomic64_sub(bytes, &dev->gnd_rdmaq_bytes_out);
LASSERTF(atomic64_read(&dev->gnd_rdmaq_bytes_out) >= 0,
"bytes_out negative! %ld\n", atomic64_read(&dev->gnd_rdmaq_bytes_out));
- GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to "LPD64"",
+ GNIDBG_TX(D_NETTRACE, tx, "rdma -- %d to %lld",
bytes, atomic64_read(&dev->gnd_rdmaq_bytes_out));
}
* GART resource, etc starvation handling */
if (rrc != GNI_RC_SUCCESS) {
GNIDBG_TX(D_NET, tx, "Can't map %d pages: dev %d "
- "phys %u pp %u, virt %u nob "LPU64"",
+ "phys %u pp %u, virt %u nob %llu",
tx->tx_phys_npages, dev->gnd_id,
dev->gnd_map_nphys, dev->gnd_map_physnop,
dev->gnd_map_nvirt, dev->gnd_map_virtnob);
NULL, flags, &tx->tx_map_key);
if (rrc != GNI_RC_SUCCESS) {
GNIDBG_TX(D_NET, tx, "Can't map %u bytes: dev %d "
- "phys %u pp %u, virt %u nob "LPU64"",
+ "phys %u pp %u, virt %u nob %llu",
tx->tx_nob, dev->gnd_id,
dev->gnd_map_nphys, dev->gnd_map_physnop,
dev->gnd_map_nvirt, dev->gnd_map_virtnob);
hold_timeout = GNILND_TIMEOUT2DEADMAN;
GNIDBG_TX(D_NET, tx,
- "dev %p delaying MDD release for %dms key "LPX64"."LPX64"",
+ "dev %p delaying MDD release for %dms key %#llx.%#llx",
tx->tx_conn->gnc_device, hold_timeout,
tx->tx_map_key.qword1, tx->tx_map_key.qword2);
}
}
/* lets us know if we can push this RDMA through now */
-inline int
+static int
kgnilnd_auth_rdma_bytes(kgn_device_t *dev, kgn_tx_t *tx)
{
long bytes_left;
}
/* don't create a connection if the peer is marked down */
- if (peer->gnp_down == GNILND_RCA_NODE_DOWN) {
+ if (peer->gnp_state != GNILND_PEER_UP) {
read_unlock(&kgnilnd_data.kgn_peer_conn_lock);
rc = -ENETRESET;
GOTO(no_peer, rc);
kgnilnd_add_peer_locked(target->nid, new_peer, &peer);
/* don't create a connection if the peer is not up */
- if (peer->gnp_down != GNILND_RCA_NODE_UP) {
+ if (peer->gnp_state != GNILND_PEER_UP) {
write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
rc = -ENETRESET;
GOTO(no_peer, rc);
if (tx->tx_buffer_copy == NULL) {
/* Allocate the largest copy buffer we will need, this will prevent us from overwriting data
* and require at most we allocate a few extra bytes. */
- tx->tx_buffer_copy = vmalloc(desc_nob);
+ tx->tx_buffer_copy = kgnilnd_vzalloc(desc_nob);
if (!tx->tx_buffer_copy) {
/* allocation of buffer failed nak the rdma */
rc = kgnilnd_mem_register(conn->gnc_device->gnd_handle, (__u64)tx->tx_buffer_copy, desc_nob, NULL, GNI_MEM_READWRITE, &tx->tx_buffer_copy_map_key);
if (rc != GNI_RC_SUCCESS) {
/* Registration Failed nak rdma and kill the tx. */
- vfree(tx->tx_buffer_copy);
+ kgnilnd_vfree(tx->tx_buffer_copy,
+ desc_nob);
tx->tx_buffer_copy = NULL;
kgnilnd_nak_rdma(tx->tx_conn, tx->tx_msg.gnm_type, -EFAULT, cookie, tx->tx_msg.gnm_srcnid);
kgnilnd_tx_done(tx, -EFAULT);
tx->tx_rdma_desc.remote_mem_hndl = sink->gnrd_key;
tx->tx_rdma_desc.length = desc_nob;
tx->tx_nob_rdma = nob;
- if (*kgnilnd_tunables.kgn_bte_dlvr_mode)
- tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_dlvr_mode;
+ if (post_type == GNI_POST_RDMA_PUT && *kgnilnd_tunables.kgn_bte_put_dlvr_mode)
+ tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_put_dlvr_mode;
+ if (post_type == GNI_POST_RDMA_GET && *kgnilnd_tunables.kgn_bte_get_dlvr_mode)
+ tx->tx_rdma_desc.dlvr_mode = *kgnilnd_tunables.kgn_bte_get_dlvr_mode;
/* prep final completion message */
kgnilnd_init_msg(&tx->tx_msg, type, tx->tx_msg.gnm_srcnid);
tx->tx_msg.gnm_u.completion.gncm_cookie = cookie;
tx, conn, conn->gnc_close_sent);
GNIDBG_TX(D_NET, tx, "Post RDMA type 0x%02x conn %p dlvr_mode "
- "0x%x cookie:"LPX64,
+ "0x%x cookie:%#llx",
type, conn, tx->tx_rdma_desc.dlvr_mode, cookie);
/* set CQ dedicated for RDMA */
kgnilnd_unmap_buffer(tx, 0);
if (tx->tx_buffer_copy != NULL) {
- vfree(tx->tx_buffer_copy);
+ kgnilnd_vfree(tx->tx_buffer_copy, desc_nob);
tx->tx_buffer_copy = NULL;
}
CERROR("Couldnt find matching peer %p or conn %p / %p\n",
peer, conn, found_conn);
if (found_conn) {
- CERROR("Unexpected connstamp "LPX64"("LPX64" expected)"
+ CERROR("Unexpected connstamp %#llx(%#llx expected)"
" from %s", rxmsg->gnm_connstamp,
found_conn->gnc_peer_connstamp,
libcfs_nid2str(peer->gnp_nid));
if (time_after_eq(now, newest_last_rx + timeout)) {
uint32_t level = D_CONSOLE|D_NETERROR;
- if (conn->gnc_peer->gnp_down == GNILND_RCA_NODE_DOWN) {
+ if (conn->gnc_peer->gnp_state == GNILND_PEER_DOWN) {
level = D_NET;
}
GNIDBG_CONN(level, conn,
conn->gnc_close_recvd = GNILND_CLOSE_INJECT1;
conn->gnc_peer_error = -ETIMEDOUT;
}
+
+ if (*kgnilnd_tunables.kgn_to_reconn_disable &&
+ rc == -ETIMEDOUT) {
+ peer->gnp_state = GNILND_PEER_TIMED_OUT;
+ CDEBUG(D_WARNING, "%s conn timed out, will "
+ "reconnect upon request from peer\n",
+ libcfs_nid2str(conn->gnc_peer->gnp_nid));
+ }
/* Once we mark closed, any of the scheduler threads could
* get it and move through before we hit the fail loc code */
kgnilnd_close_conn_locked(conn, rc);
/* Don't reconnect if we are still trying to clear out old conns.
* This prevents us sending traffic on the new mbox before ensuring we are done
* with the old one */
- reconnect = (peer->gnp_down == GNILND_RCA_NODE_UP) &&
+ reconnect = (peer->gnp_state == GNILND_PEER_UP) &&
(atomic_read(&peer->gnp_dirty_eps) == 0);
/* fast reconnect after a timeout */
"this is bad, somehow our credits didn't protect us"
" from CQ overrun\n");
LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_POST,
- "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
+ "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
event_data, GNI_CQ_GET_TYPE(event_data));
rrc = kgnilnd_get_completed(dev->gnd_snd_rdma_cqh, event_data,
if (rrc == GNI_RC_NOT_DONE) {
CDEBUG(D_INFO,
- "SMSG send CQ %d not ready (data "LPX64") "
+ "SMSG send CQ %d not ready (data %#llx) "
"processed %ld\n", dev->gnd_id, event_data,
num_processed);
return num_processed;
"this is bad, somehow our credits didn't "
"protect us from CQ overrun\n");
LASSERTF(GNI_CQ_GET_TYPE(event_data) == GNI_CQ_EVENT_TYPE_SMSG,
- "rrc %d, GNI_CQ_GET_TYPE("LPX64") = "LPX64"\n", rrc,
+ "rrc %d, GNI_CQ_GET_TYPE(%#llx) = %#llx\n", rrc,
event_data, GNI_CQ_GET_TYPE(event_data));
/* if SMSG couldn't handle an error, time for conn to die */
if (conn == NULL) {
/* Conn was destroyed? */
CDEBUG(D_NET,
- "SMSG CQID lookup "LPX64" failed\n",
+ "SMSG CQID lookup %#llx failed\n",
GNI_CQ_GET_INST_ID(event_data));
write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
continue;
kgnilnd_gl_mutex_unlock(&dev->gnd_cq_mutex);
if (rrc == GNI_RC_NOT_DONE) {
- CDEBUG(D_INFO, "SMSG RX CQ %d empty data "LPX64" "
+ CDEBUG(D_INFO, "SMSG RX CQ %d empty data %#llx "
"processed %ld\n",
dev->gnd_id, event_data, num_processed);
return num_processed;
/* sender should get error event too and take care
of failed transaction by re-transmitting */
if (rrc == GNI_RC_TRANSACTION_ERROR) {
- CDEBUG(D_NET, "SMSG RX CQ error "LPX64"\n", event_data);
+ CDEBUG(D_NET, "SMSG RX CQ error %#llx\n", event_data);
continue;
}
conn = kgnilnd_cqid2conn_locked(
GNI_CQ_GET_INST_ID(event_data));
if (conn == NULL) {
- CDEBUG(D_NET, "SMSG RX CQID lookup "LPU64" "
- "failed, dropping event "LPX64"\n",
+ CDEBUG(D_NET, "SMSG RX CQID lookup %llu "
+ "failed, dropping event %#llx\n",
GNI_CQ_GET_INST_ID(event_data),
event_data);
} else {
- CDEBUG(D_NET, "SMSG RX: CQID "LPU64" "
+ CDEBUG(D_NET, "SMSG RX: CQID %llu "
"conn %p->%s\n",
GNI_CQ_GET_INST_ID(event_data),
conn, conn->gnc_peer ?
GNITX_ASSERTF(tx, tx->tx_id.txe_smsg_id != 0,
"tx with zero id", NULL);
- CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie "LPX64"\n",
+ CDEBUG(D_NET, "sending regular msg: %p, type %s(0x%02x), cookie %#llx\n",
tx, kgnilnd_msgtype2str(tx->tx_msg.gnm_type),
tx->tx_msg.gnm_type, tx->tx_id.txe_cookie);
GNITX_ASSERTF(tx, ((tx->tx_id.txe_idx == ev_id.txe_idx) &&
(tx->tx_id.txe_cookie = cookie)),
"conn 0x%p->%s tx_ref_table hosed: wanted "
- "txe_cookie "LPX64" txe_idx %d "
- "found tx %p cookie "LPX64" txe_idx %d\n",
+ "txe_cookie %#llx txe_idx %d "
+ "found tx %p cookie %#llx txe_idx %d\n",
conn, libcfs_nid2str(conn->gnc_peer->gnp_nid),
cookie, ev_id.txe_idx,
tx, tx->tx_id.txe_cookie, tx->tx_id.txe_idx);
tx->tx_state, GNILND_TX_WAITING_REPLY,
libcfs_nid2str(conn->gnc_peer->gnp_nid));
} else {
- CWARN("Unmatched reply %02x, or %02x/"LPX64" from %s\n",
+ CWARN("Unmatched reply %02x, or %02x/%#llx from %s\n",
type1, type2, cookie, libcfs_nid2str(conn->gnc_peer->gnp_nid));
}
return tx;
tx->tx_state &= ~GNILND_TX_WAITING_REPLY;
if (rc == -EFAULT) {
- CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16"LPF64"u physnop %8d buffertype %#8x MemHandle "LPX64"."LPX64"x\n",
+ CDEBUG(D_NETERROR, "Error %d TX data: TX %p tx_id %x nob %16llu physnop %8d buffertype %#8x MemHandle %#llx.%#llxx\n",
rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
if(*kgnilnd_tunables.kgn_efault_lbug) {
}
if (msg->gnm_connstamp != conn->gnc_peer_connstamp) {
- GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp "LPX64"("LPX64
+ GNIDBG_MSG(D_NETERROR, msg, "Unexpected connstamp %#llx(%#llx"
" expected) from %s",
msg->gnm_connstamp, conn->gnc_peer_connstamp,
libcfs_nid2str(peer->gnp_nid));
} else {
GNIDBG_TX(log_retrans_level, tx,
"transient map failure #%d %d pages/%d bytes phys %u@%u "
- "virt %u@"LPU64" "
+ "virt %u@%llu "
"nq_map %d mdd# %d/%d GART %ld",
dev->gnd_map_attempt, tx->tx_phys_npages, tx->tx_nob,
dev->gnd_map_nphys, dev->gnd_map_physnop * PAGE_SIZE,