X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lnet%2Fklnds%2Fgnilnd%2Fgnilnd_cb.c;h=705a341c895a3a9a57a80c6bececc65e5b48304e;hp=381aa647823740eff78e2f156b96cd6af26ca999;hb=aab1d832130ee5c181cf7e0e5aa555244d150b00;hpb=381060a6244dfba4819fa81f2b928beb12a39350

diff --git a/lnet/klnds/gnilnd/gnilnd_cb.c b/lnet/klnds/gnilnd/gnilnd_cb.c
index 381aa64..705a341 100644
--- a/lnet/klnds/gnilnd/gnilnd_cb.c
+++ b/lnet/klnds/gnilnd/gnilnd_cb.c
@@ -23,6 +23,7 @@
  *
  */
 
+#include <asm/page.h>
 #include <linux/nmi.h>
 #include "gnilnd.h"
 
@@ -228,7 +229,7 @@ kgnilnd_free_tx(kgn_tx_t *tx)
 
 	/* we only allocate this if we need to */
 	if (tx->tx_phys != NULL) {
-		cfs_mem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
+		kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
 		CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
 		       LNET_MAX_IOV * sizeof(gni_mem_segment_t), tx->tx_phys);
 	}
@@ -242,9 +243,8 @@ kgnilnd_free_tx(kgn_tx_t *tx)
 #if 0
 	KGNILND_POISON(tx, 0x5a, sizeof(kgn_tx_t));
 #endif
-	cfs_mem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
-	CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n",
-	       sizeof(*tx), tx);
+	CDEBUG(D_MALLOC, "slab-freed 'tx': %lu at %p.\n", sizeof(*tx), tx);
+	kmem_cache_free(kgnilnd_data.kgn_tx_cache, tx);
 }
 
 kgn_tx_t *
@@ -255,7 +255,7 @@ kgnilnd_alloc_tx (void)
 	if (CFS_FAIL_CHECK(CFS_FAIL_GNI_ALLOC_TX))
 		return tx;
 
-	tx = cfs_mem_cache_alloc(kgnilnd_data.kgn_tx_cache, CFS_ALLOC_ATOMIC);
+	tx = kmem_cache_alloc(kgnilnd_data.kgn_tx_cache, GFP_ATOMIC);
 	if (tx == NULL) {
 		CERROR("failed to allocate tx\n");
 		return NULL;
@@ -500,7 +500,6 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io
 	 * gni_smsg_send to send that as the payload */
 
 	LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
-	LASSERT(nob >= 0);
 
 	if (nob == 0) {
 		tx->tx_buffer = NULL;
@@ -522,7 +521,7 @@ kgnilnd_setup_immediate_buffer(kgn_tx_t *tx, unsigned int niov, struct iovec *io
 			 * than kiov_len, we will also have a whole at the end of that page
 			 * which isn't allowed */
 			if ((kiov[i].kiov_offset != 0 && i > 0) ||
-			    (kiov[i].kiov_offset + kiov[i].kiov_len != CFS_PAGE_SIZE && i < niov - 1)) {
+			    (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1)) {
 				CNETERR("Can't make payload contiguous in I/O VM:"
 				       "page %d, offset %u, nob %u, kiov_offset %u kiov_len %u \n",
 				       i, offset, nob, kiov->kiov_offset, kiov->kiov_len);
@@ -640,8 +639,8 @@ kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
 	LASSERT(tx->tx_buftype == GNILND_BUF_NONE);
 
 	/* only allocate this if we are going to use it */
-	tx->tx_phys = cfs_mem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
-					      CFS_ALLOC_ATOMIC);
+	tx->tx_phys = kmem_cache_alloc(kgnilnd_data.kgn_tx_phys_cache,
+					      GFP_ATOMIC);
 	if (tx->tx_phys == NULL) {
 		CERROR("failed to allocate tx_phys\n");
 		rc = -ENOMEM;
@@ -711,7 +710,7 @@ kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
 			       "nkiov %u offset %u\n",
 		      kiov->kiov_page, kiov->kiov_offset, kiov->kiov_len, nob, nkiov, offset);
 
-		phys->address = lnet_page2phys(kiov->kiov_page);
+		phys->address = page_to_phys(kiov->kiov_page);
 		phys++;
 		kiov++;
 		nkiov--;
@@ -729,7 +728,7 @@ kgnilnd_setup_phys_buffer(kgn_tx_t *tx, int nkiov, lnet_kiov_t *kiov,
 
 error:
 	if (tx->tx_phys != NULL) {
-		cfs_mem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
+		kmem_cache_free(kgnilnd_data.kgn_tx_phys_cache, tx->tx_phys);
 		CDEBUG(D_MALLOC, "slab-freed 'tx_phys': %lu at %p.\n",
 		       sizeof(*tx->tx_phys), tx->tx_phys);
 		tx->tx_phys = NULL;
@@ -1099,7 +1098,7 @@ kgnilnd_add_purgatory_tx(kgn_tx_t *tx)
 	if (tx->tx_buffer_copy)
 		gmp->gmp_map_key = tx->tx_buffer_copy_map_key;
 	else
-	gmp->gmp_map_key = tx->tx_map_key;
+		gmp->gmp_map_key = tx->tx_map_key;
 
 	atomic_inc(&conn->gnc_device->gnd_n_mdd_held);
 
@@ -1177,8 +1176,8 @@ kgnilnd_unmap_buffer(kgn_tx_t *tx, int error)
 			rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, 0);
 			LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
 		} else {
-		rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
-		LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
+			rrc = kgnilnd_mem_deregister(dev->gnd_handle, &tx->tx_map_key, hold_timeout);
+			LASSERTF(rrc == GNI_RC_SUCCESS, "rrc %d\n", rrc);
 		}
 
 		tx->tx_buftype--;
@@ -1209,7 +1208,7 @@ kgnilnd_tx_done(kgn_tx_t *tx, int completion)
 		       libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
 		       tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
 		       kgnilnd_tx_state2str(tx->tx_list_state),
-		       cfs_duration_sec((long)jiffies - tx->tx_qtime));
+		       cfs_duration_sec((unsigned long)jiffies - tx->tx_qtime));
 	}
 
 	/* The error codes determine if we hold onto the MDD */
@@ -1697,7 +1696,7 @@ kgnilnd_queue_rdma(kgn_conn_t *conn, kgn_tx_t *tx)
 void
 kgnilnd_queue_tx(kgn_conn_t *conn, kgn_tx_t *tx)
 {
-	int            rc;
+	int            rc = 0;
 	int            add_tail = 1;
 
 	/* set the tx_id here, we delay it until we have an actual conn
@@ -1760,6 +1759,7 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
 	kgn_peer_t      *new_peer = NULL;
 	kgn_conn_t      *conn = NULL;
 	int              rc;
+	int              node_state;
 
 	ENTRY;
 
@@ -1800,6 +1800,8 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
 
 	CFS_RACE(CFS_FAIL_GNI_FIND_TARGET);
 
+	node_state = kgnilnd_get_node_state(LNET_NIDADDR(target->nid));
+
 	/* NB - this will not block during normal operations -
 	 * the only writer of this is in the startup/shutdown path. */
 	rc = down_read_trylock(&kgnilnd_data.kgn_net_rw_sem);
@@ -1811,7 +1813,7 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
 	/* ignore previous peer entirely - we cycled the lock, so we
 	 * will create new peer and at worst drop it if peer is still
 	 * in the tables */
-	rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net);
+	rc = kgnilnd_create_peer_safe(&new_peer, target->nid, net, node_state);
 	if (rc != 0) {
 		up_read(&kgnilnd_data.kgn_net_rw_sem);
 		GOTO(no_peer, rc);
@@ -1824,7 +1826,20 @@ kgnilnd_launch_tx(kgn_tx_t *tx, kgn_net_t *net, lnet_process_id_t *target)
 	 * if we don't find it, add our new one to the list */
 	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) {
+		write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
+		rc = -ENETRESET;
+		GOTO(no_peer, rc);
+	}
+
 	conn = kgnilnd_find_or_create_conn_locked(peer);
+
+	if (CFS_FAIL_CHECK(CFS_FAIL_GNI_DGRAM_DROP_TX)) {
+		write_unlock(&kgnilnd_data.kgn_peer_conn_lock);
+		GOTO(no_peer, rc);
+	}
+
 	if (conn != NULL) {
 		/* oh hey, found a conn now... magical */
 		kgnilnd_queue_tx(conn, tx);
@@ -1990,7 +2005,7 @@ kgnilnd_alloc_rx(void)
 {
 	kgn_rx_t	*rx;
 
-	rx = cfs_mem_cache_alloc(kgnilnd_data.kgn_rx_cache, CFS_ALLOC_ATOMIC);
+	rx = kmem_cache_alloc(kgnilnd_data.kgn_rx_cache, GFP_ATOMIC);
 	if (rx == NULL) {
 		CERROR("failed to allocate rx\n");
 		return NULL;
@@ -2037,6 +2052,7 @@ kgnilnd_consume_rx(kgn_rx_t *rx)
 	/* if we are eager, free the cache alloc'd msg */
 	if (unlikely(rx->grx_eager)) {
 		LIBCFS_FREE(rxmsg, sizeof(*rxmsg) + *kgnilnd_tunables.kgn_max_immediate);
+		atomic_dec(&kgnilnd_data.kgn_neager_allocs);
 
 		/* release ref from eager_recv */
 		kgnilnd_conn_decref(conn);
@@ -2045,7 +2061,7 @@ kgnilnd_consume_rx(kgn_rx_t *rx)
 		kgnilnd_release_msg(conn);
 	}
 
-	cfs_mem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
+	kmem_cache_free(kgnilnd_data.kgn_rx_cache, rx);
 	CDEBUG(D_MALLOC, "slab-freed 'rx': %lu at %p.\n",
 	       sizeof(*rx), rx);
 
@@ -2342,6 +2358,15 @@ kgnilnd_eager_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
 
 	/* we have no credits or buffers for this message, so copy it
 	 * somewhere for a later kgnilnd_recv */
+	if (atomic_read(&kgnilnd_data.kgn_neager_allocs) >=
+			*kgnilnd_tunables.kgn_eager_credits) {
+		CERROR("Out of eager credits to %s\n",
+			libcfs_nid2str(conn->gnc_peer->gnp_nid));
+		return -ENOMEM;
+	}
+
+	atomic_inc(&kgnilnd_data.kgn_neager_allocs);
+
 	LIBCFS_ALLOC(eagermsg, sizeof(*eagermsg) + *kgnilnd_tunables.kgn_max_immediate);
 	if (eagermsg == NULL) {
 		kgnilnd_conn_decref(conn);
@@ -2515,7 +2540,7 @@ kgnilnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
 		tx->tx_msg.gnm_u.putack.gnpam_desc.gnrd_nob = mlen;
 
 		tx->tx_lntmsg[0] = lntmsg; /* finalize this on RDMA_DONE */
-
+		tx->tx_qtime = jiffies;
 		/* we only queue from kgnilnd_recv - we might get called from other contexts
 		 * and we don't want to block the mutex in those cases */
 
@@ -2946,7 +2971,6 @@ kgnilnd_reaper(void *arg)
 	struct timer_list  timer;
 	DEFINE_WAIT(wait);
 
-	cfs_daemonize("kgnilnd_rpr");
 	cfs_block_allsigs();
 
 	/* all gnilnd threads need to run fairly urgently */
@@ -3043,7 +3067,7 @@ kgnilnd_recv_bte_get(kgn_tx_t *tx) {
 		lnet_copy_flat2kiov(
 			niov, kiov, offset,
 			nob,
-			tx->tx_buffer_copy, tx->tx_offset, nob);
+			tx->tx_buffer_copy + tx->tx_offset, 0, nob);
 	} else {
 		memcpy(tx->tx_buffer, tx->tx_buffer_copy + tx->tx_offset, nob);
 	}
@@ -3063,6 +3087,9 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev)
 	long                   num_processed = 0;
 	kgn_conn_t            *conn = NULL;
 	kgn_tx_t              *tx = NULL;
+	kgn_rdma_desc_t       *rdesc;
+	unsigned int           rnob;
+	__u64                  rcookie;
 
 	for (;;) {
 		/* make sure we don't keep looping if we need to reset */
@@ -3183,31 +3210,28 @@ kgnilnd_check_rdma_cq(kgn_device_t *dev)
 
 		if (tx->tx_msg.gnm_type == GNILND_MSG_PUT_DONE ||
 		    tx->tx_msg.gnm_type == GNILND_MSG_GET_DONE_REV) {
-			if (should_retry) {
-				kgnilnd_rdma(tx, tx->tx_msg.gnm_type,
-					     &tx->tx_putinfo.gnpam_desc,
-					     tx->tx_putinfo.gnpam_desc.gnrd_nob,
-					     tx->tx_putinfo.gnpam_dst_cookie);
-			} else {
-				kgnilnd_nak_rdma(conn, tx->tx_msg.gnm_type,
-						-EFAULT,
-						tx->tx_putinfo.gnpam_dst_cookie,
-						tx->tx_msg.gnm_srcnid);
-				kgnilnd_tx_done(tx, -EFAULT);
-			}
+			rdesc    = &tx->tx_putinfo.gnpam_desc;
+			rnob     = tx->tx_putinfo.gnpam_desc.gnrd_nob;
+			rcookie  = tx->tx_putinfo.gnpam_dst_cookie;
 		} else {
-			if (should_retry) {
-				kgnilnd_rdma(tx, tx->tx_msg.gnm_type,
-					     &tx->tx_getinfo.gngm_desc,
-					     tx->tx_lntmsg[0]->msg_len,
-					     tx->tx_getinfo.gngm_cookie);
-			} else {
-				kgnilnd_nak_rdma(conn, tx->tx_msg.gnm_type,
-						-EFAULT,
-						tx->tx_getinfo.gngm_cookie,
-						tx->tx_msg.gnm_srcnid);
-				kgnilnd_tx_done(tx, -EFAULT);
-			}
+			rdesc    = &tx->tx_getinfo.gngm_desc;
+			rnob     = tx->tx_lntmsg[0]->msg_len;
+			rcookie  = tx->tx_getinfo.gngm_cookie;
+		}
+
+		if (should_retry) {
+			kgnilnd_rdma(tx,
+				     tx->tx_msg.gnm_type,
+				     rdesc,
+				     rnob, rcookie);
+		} else {
+			kgnilnd_nak_rdma(conn,
+					 tx->tx_msg.gnm_type,
+					 -EFAULT,
+					 rcookie,
+					 tx->tx_msg.gnm_srcnid);
+			kgnilnd_tx_done(tx, -EFAULT);
+			kgnilnd_close_conn(conn, -ECOMM);
 		}
 
 		/* drop ref from kgnilnd_validate_tx_ev_id */
@@ -3581,7 +3605,7 @@ kgnilnd_send_mapped_tx(kgn_tx_t *tx, int try_map_if_full)
 	case GNILND_MSG_PUT_DONE_REV:
 		kgnilnd_rdma(tx, GNILND_MSG_PUT_DONE_REV,
 			     &tx->tx_getinfo.gngm_desc,
-			     tx->tx_lntmsg[0]->msg_len,
+			     tx->tx_nob,
 			     tx->tx_getinfo.gngm_cookie);
 		break;
 	case GNILND_MSG_GET_ACK_REV:
@@ -3910,6 +3934,13 @@ kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
 {
 	int             complete = 0;
 	kgn_conn_t      *conn = tx->tx_conn;
+	__u64 nob = tx->tx_nob;
+	__u32 physnop = tx->tx_phys_npages;
+	int   id = tx->tx_id.txe_smsg_id;
+	int buftype = tx->tx_buftype;
+	gni_mem_handle_t hndl;
+	hndl.qword1 = tx->tx_map_key.qword1;
+	hndl.qword2 = tx->tx_map_key.qword2;
 
 	spin_lock(&conn->gnc_list_lock);
 
@@ -3919,6 +3950,22 @@ kgnilnd_complete_tx(kgn_tx_t *tx, int rc)
 	tx->tx_rc = rc;
 	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",
+			rc, tx, id, nob, physnop, buftype, hndl.qword1, hndl.qword2);
+
+		if(*kgnilnd_tunables.kgn_efault_lbug) {
+			GNIDBG_TOMSG(D_NETERROR, &tx->tx_msg,
+			"error %d on tx 0x%p->%s id %u/%d state %s age %ds",
+			rc, tx, conn ?
+			libcfs_nid2str(conn->gnc_peer->gnp_nid) : "<?>",
+			tx->tx_id.txe_smsg_id, tx->tx_id.txe_idx,
+			kgnilnd_tx_state2str(tx->tx_list_state),
+			cfs_duration_sec((unsigned long) jiffies - tx->tx_qtime));
+			LBUG();
+		}
+	}
+
 	if (!(tx->tx_state & GNILND_TX_WAITING_COMPLETION)) {
 		kgnilnd_tx_del_state_locked(tx, NULL, conn, GNILND_TX_ALLOCD);
 		/* sample under lock as follow on steps require gnc_list_lock
@@ -4850,15 +4897,12 @@ kgnilnd_scheduler(void *arg)
 {
 	int               threadno = (long)arg;
 	kgn_device_t		*dev;
-	char			name[16];
 	int			busy_loops = 0;
 	unsigned long	  deadline = 0;
 	DEFINE_WAIT(wait);
 
 	dev = &kgnilnd_data.kgn_devices[(threadno + 1) % kgnilnd_data.kgn_ndevs];
 
-	snprintf(name, sizeof(name), "kgnilnd_sd_%02d", threadno);
-	cfs_daemonize(name);
 	cfs_block_allsigs();
 
 	/* all gnilnd threads need to run fairly urgently */