brintos

brintos / linux-shallow public Read only

0
0
Text · 34.1 KiB · 5cde275 Raw
1336 lines · c
1/*2 * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.5 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.6 *7 * This software is available to you under a choice of one of two8 * licenses.  You may choose to be licensed under the terms of the GNU9 * General Public License (GPL) Version 2, available from the file10 * COPYING in the main directory of this source tree, or the11 * OpenIB.org BSD license below:12 *13 *     Redistribution and use in source and binary forms, with or14 *     without modification, are permitted provided that the following15 *     conditions are met:16 *17 *      - Redistributions of source code must retain the above18 *        copyright notice, this list of conditions and the following19 *        disclaimer.20 *21 *      - Redistributions in binary form must reproduce the above22 *        copyright notice, this list of conditions and the following23 *        disclaimer in the documentation and/or other materials24 *        provided with the distribution.25 *26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE33 * SOFTWARE.34 */35 36#include <linux/delay.h>37#include <linux/moduleparam.h>38#include <linux/dma-mapping.h>39#include <linux/slab.h>40 41#include <linux/ip.h>42#include <linux/tcp.h>43#include <rdma/ib_cache.h>44 45#include "ipoib.h"46 47#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA48static int data_debug_level;49 50module_param(data_debug_level, int, 0644);51MODULE_PARM_DESC(data_debug_level,52		 "Enable data path debug tracing if > 0");53#endif54 55struct ipoib_ah *ipoib_create_ah(struct net_device *dev,56				 struct ib_pd *pd, struct rdma_ah_attr *attr)57{58	struct ipoib_ah *ah;59	struct ib_ah *vah;60 61	ah = kmalloc(sizeof(*ah), GFP_KERNEL);62	if (!ah)63		return ERR_PTR(-ENOMEM);64 65	ah->dev       = dev;66	ah->last_send = 0;67	kref_init(&ah->ref);68 69	vah = rdma_create_ah(pd, attr, RDMA_CREATE_AH_SLEEPABLE);70	if (IS_ERR(vah)) {71		kfree(ah);72		ah = (struct ipoib_ah *)vah;73	} else {74		ah->ah = vah;75		ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah);76	}77 78	return ah;79}80 81void ipoib_free_ah(struct kref *kref)82{83	struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref);84	struct ipoib_dev_priv *priv = ipoib_priv(ah->dev);85 86	unsigned long flags;87 88	spin_lock_irqsave(&priv->lock, flags);89	list_add_tail(&ah->list, &priv->dead_ahs);90	spin_unlock_irqrestore(&priv->lock, flags);91}92 93static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv,94				  u64 mapping[IPOIB_UD_RX_SG])95{96	ib_dma_unmap_single(priv->ca, mapping[0],97			    IPOIB_UD_BUF_SIZE(priv->max_ib_mtu),98			    DMA_FROM_DEVICE);99}100 101static int ipoib_ib_post_receive(struct net_device *dev, int id)102{103	struct ipoib_dev_priv *priv = ipoib_priv(dev);104	int ret;105 106	priv->rx_wr.wr_id   = id | IPOIB_OP_RECV;107	priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0];108	priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1];109 110 111	ret = ib_post_recv(priv->qp, &priv->rx_wr, NULL);112	if (unlikely(ret)) {113		ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret);114		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping);115		dev_kfree_skb_any(priv->rx_ring[id].skb);116		priv->rx_ring[id].skb = NULL;117	}118 119	return ret;120}121 122static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id)123{124	struct ipoib_dev_priv *priv = ipoib_priv(dev);125	struct sk_buff *skb;126	int buf_size;127	u64 *mapping;128 129	buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu);130 131	skb = dev_alloc_skb(buf_size + IPOIB_HARD_LEN);132	if (unlikely(!skb))133		return NULL;134 135	/*136	 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is137	 * 64 bytes aligned138	 */139	skb_reserve(skb, sizeof(struct ipoib_pseudo_header));140 141	mapping = priv->rx_ring[id].mapping;142	mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size,143				       DMA_FROM_DEVICE);144	if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0])))145		goto error;146 147	priv->rx_ring[id].skb = skb;148	return skb;149error:150	dev_kfree_skb_any(skb);151	return NULL;152}153 154static int ipoib_ib_post_receives(struct net_device *dev)155{156	struct ipoib_dev_priv *priv = ipoib_priv(dev);157	int i;158 159	for (i = 0; i < ipoib_recvq_size; ++i) {160		if (!ipoib_alloc_rx_skb(dev, i)) {161			ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);162			return -ENOMEM;163		}164		if (ipoib_ib_post_receive(dev, i)) {165			ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i);166			return -EIO;167		}168	}169 170	return 0;171}172 173static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)174{175	struct ipoib_dev_priv *priv = ipoib_priv(dev);176	unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV;177	struct sk_buff *skb;178	u64 mapping[IPOIB_UD_RX_SG];179	union ib_gid *dgid;180	union ib_gid *sgid;181 182	ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n",183		       wr_id, wc->status);184 185	if (unlikely(wr_id >= ipoib_recvq_size)) {186		ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n",187			   wr_id, ipoib_recvq_size);188		return;189	}190 191	skb  = priv->rx_ring[wr_id].skb;192 193	if (unlikely(wc->status != IB_WC_SUCCESS)) {194		if (wc->status != IB_WC_WR_FLUSH_ERR)195			ipoib_warn(priv,196				   "failed recv event (status=%d, wrid=%d vend_err %#x)\n",197				   wc->status, wr_id, wc->vendor_err);198		ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping);199		dev_kfree_skb_any(skb);200		priv->rx_ring[wr_id].skb = NULL;201		return;202	}203 204	memcpy(mapping, priv->rx_ring[wr_id].mapping,205	       IPOIB_UD_RX_SG * sizeof(*mapping));206 207	/*208	 * If we can't allocate a new RX buffer, dump209	 * this packet and reuse the old buffer.210	 */211	if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) {212		++dev->stats.rx_dropped;213		goto repost;214	}215 216	ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",217		       wc->byte_len, wc->slid);218 219	ipoib_ud_dma_unmap_rx(priv, mapping);220 221	skb_put(skb, wc->byte_len);222 223	/* First byte of dgid signals multicast when 0xff */224	dgid = &((struct ib_grh *)skb->data)->dgid;225 226	if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff)227		skb->pkt_type = PACKET_HOST;228	else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0)229		skb->pkt_type = PACKET_BROADCAST;230	else231		skb->pkt_type = PACKET_MULTICAST;232 233	sgid = &((struct ib_grh *)skb->data)->sgid;234 235	/*236	 * Drop packets that this interface sent, ie multicast packets237	 * that the HCA has replicated.238	 */239	if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) {240		int need_repost = 1;241 242		if ((wc->wc_flags & IB_WC_GRH) &&243		    sgid->global.interface_id != priv->local_gid.global.interface_id)244			need_repost = 0;245 246		if (need_repost) {247			dev_kfree_skb_any(skb);248			goto repost;249		}250	}251 252	skb_pull(skb, IB_GRH_BYTES);253 254	skb->protocol = ((struct ipoib_header *) skb->data)->proto;255	skb_add_pseudo_hdr(skb);256 257	++dev->stats.rx_packets;258	dev->stats.rx_bytes += skb->len;259	if (skb->pkt_type == PACKET_MULTICAST)260		dev->stats.multicast++;261 262	skb->dev = dev;263	if ((dev->features & NETIF_F_RXCSUM) &&264			likely(wc->wc_flags & IB_WC_IP_CSUM_OK))265		skb->ip_summed = CHECKSUM_UNNECESSARY;266 267	napi_gro_receive(&priv->recv_napi, skb);268 269repost:270	if (unlikely(ipoib_ib_post_receive(dev, wr_id)))271		ipoib_warn(priv, "ipoib_ib_post_receive failed "272			   "for buf %d\n", wr_id);273}274 275int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req)276{277	struct sk_buff *skb = tx_req->skb;278	u64 *mapping = tx_req->mapping;279	int i;280	int off;281 282	if (skb_headlen(skb)) {283		mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb),284					       DMA_TO_DEVICE);285		if (unlikely(ib_dma_mapping_error(ca, mapping[0])))286			return -EIO;287 288		off = 1;289	} else290		off = 0;291 292	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {293		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];294		mapping[i + off] = ib_dma_map_page(ca,295						 skb_frag_page(frag),296						 skb_frag_off(frag),297						 skb_frag_size(frag),298						 DMA_TO_DEVICE);299		if (unlikely(ib_dma_mapping_error(ca, mapping[i + off])))300			goto partial_error;301	}302	return 0;303 304partial_error:305	for (; i > 0; --i) {306		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1];307 308		ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE);309	}310 311	if (off)312		ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE);313 314	return -EIO;315}316 317void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv,318			struct ipoib_tx_buf *tx_req)319{320	struct sk_buff *skb = tx_req->skb;321	u64 *mapping = tx_req->mapping;322	int i;323	int off;324 325	if (skb_headlen(skb)) {326		ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb),327				    DMA_TO_DEVICE);328		off = 1;329	} else330		off = 0;331 332	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {333		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];334 335		ib_dma_unmap_page(priv->ca, mapping[i + off],336				  skb_frag_size(frag), DMA_TO_DEVICE);337	}338}339 340/*341 * As the result of a completion error the QP Can be transferred to SQE states.342 * The function checks if the (send)QP is in SQE state and343 * moves it back to RTS state, that in order to have it functional again.344 */345static void ipoib_qp_state_validate_work(struct work_struct *work)346{347	struct ipoib_qp_state_validate *qp_work =348		container_of(work, struct ipoib_qp_state_validate, work);349 350	struct ipoib_dev_priv *priv = qp_work->priv;351	struct ib_qp_attr qp_attr;352	struct ib_qp_init_attr query_init_attr;353	int ret;354 355	ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr);356	if (ret) {357		ipoib_warn(priv, "%s: Failed to query QP ret: %d\n",358			   __func__, ret);359		goto free_res;360	}361	pr_info("%s: QP: 0x%x is in state: %d\n",362		__func__, priv->qp->qp_num, qp_attr.qp_state);363 364	/* currently support only in SQE->RTS transition*/365	if (qp_attr.qp_state == IB_QPS_SQE) {366		qp_attr.qp_state = IB_QPS_RTS;367 368		ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE);369		if (ret) {370			pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n",371				ret, priv->qp->qp_num);372			goto free_res;373		}374		pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n",375			__func__, priv->qp->qp_num);376	} else {377		pr_warn("QP (%d) will stay in state: %d\n",378			priv->qp->qp_num, qp_attr.qp_state);379	}380 381free_res:382	kfree(qp_work);383}384 385static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)386{387	struct ipoib_dev_priv *priv = ipoib_priv(dev);388	unsigned int wr_id = wc->wr_id;389	struct ipoib_tx_buf *tx_req;390 391	ipoib_dbg_data(priv, "send completion: id %d, status: %d\n",392		       wr_id, wc->status);393 394	if (unlikely(wr_id >= ipoib_sendq_size)) {395		ipoib_warn(priv, "send completion event with wrid %d (> %d)\n",396			   wr_id, ipoib_sendq_size);397		return;398	}399 400	tx_req = &priv->tx_ring[wr_id];401 402	ipoib_dma_unmap_tx(priv, tx_req);403 404	++dev->stats.tx_packets;405	dev->stats.tx_bytes += tx_req->skb->len;406 407	dev_kfree_skb_any(tx_req->skb);408 409	++priv->tx_tail;410	++priv->global_tx_tail;411 412	if (unlikely(netif_queue_stopped(dev) &&413		     ((priv->global_tx_head - priv->global_tx_tail) <=414		      ipoib_sendq_size >> 1) &&415		     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))416		netif_wake_queue(dev);417 418	if (wc->status != IB_WC_SUCCESS &&419	    wc->status != IB_WC_WR_FLUSH_ERR) {420		struct ipoib_qp_state_validate *qp_work;421		ipoib_warn(priv,422			   "failed send event (status=%d, wrid=%d vend_err %#x)\n",423			   wc->status, wr_id, wc->vendor_err);424		qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC);425		if (!qp_work)426			return;427 428		INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work);429		qp_work->priv = priv;430		queue_work(priv->wq, &qp_work->work);431	}432}433 434static int poll_tx(struct ipoib_dev_priv *priv)435{436	int n, i;437	struct ib_wc *wc;438 439	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);440	for (i = 0; i < n; ++i) {441		wc = priv->send_wc + i;442		if (wc->wr_id & IPOIB_OP_CM)443			ipoib_cm_handle_tx_wc(priv->dev, priv->send_wc + i);444		else445			ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i);446	}447	return n == MAX_SEND_CQE;448}449 450int ipoib_rx_poll(struct napi_struct *napi, int budget)451{452	struct ipoib_dev_priv *priv =453		container_of(napi, struct ipoib_dev_priv, recv_napi);454	struct net_device *dev = priv->dev;455	int done;456	int t;457	int n, i;458 459	done  = 0;460 461poll_more:462	while (done < budget) {463		int max = (budget - done);464 465		t = min(IPOIB_NUM_WC, max);466		n = ib_poll_cq(priv->recv_cq, t, priv->ibwc);467 468		for (i = 0; i < n; i++) {469			struct ib_wc *wc = priv->ibwc + i;470 471			if (wc->wr_id & IPOIB_OP_RECV) {472				++done;473				if (wc->wr_id & IPOIB_OP_CM)474					ipoib_cm_handle_rx_wc(dev, wc);475				else476					ipoib_ib_handle_rx_wc(dev, wc);477			} else {478				pr_warn("%s: Got unexpected wqe id\n", __func__);479			}480		}481 482		if (n != t)483			break;484	}485 486	if (done < budget) {487		napi_complete(napi);488		if (unlikely(ib_req_notify_cq(priv->recv_cq,489					      IB_CQ_NEXT_COMP |490					      IB_CQ_REPORT_MISSED_EVENTS)) &&491		    napi_schedule(napi))492			goto poll_more;493	}494 495	return done;496}497 498int ipoib_tx_poll(struct napi_struct *napi, int budget)499{500	struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv,501						   send_napi);502	struct net_device *dev = priv->dev;503	int n, i;504	struct ib_wc *wc;505 506poll_more:507	n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc);508 509	for (i = 0; i < n; i++) {510		wc = priv->send_wc + i;511		if (wc->wr_id & IPOIB_OP_CM)512			ipoib_cm_handle_tx_wc(dev, wc);513		else514			ipoib_ib_handle_tx_wc(dev, wc);515	}516 517	if (n < budget) {518		napi_complete(napi);519		if (unlikely(ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |520					      IB_CQ_REPORT_MISSED_EVENTS)) &&521		    napi_schedule(napi))522			goto poll_more;523	}524	return n < 0 ? 0 : n;525}526 527void ipoib_ib_rx_completion(struct ib_cq *cq, void *ctx_ptr)528{529	struct ipoib_dev_priv *priv = ctx_ptr;530 531	napi_schedule(&priv->recv_napi);532}533 534/* The function will force napi_schedule */535void ipoib_napi_schedule_work(struct work_struct *work)536{537	struct ipoib_dev_priv *priv =538		container_of(work, struct ipoib_dev_priv, reschedule_napi_work);539	bool ret;540 541	do {542		ret = napi_schedule(&priv->send_napi);543		if (!ret)544			msleep(3);545	} while (!ret && netif_queue_stopped(priv->dev) &&546		 test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags));547}548 549void ipoib_ib_tx_completion(struct ib_cq *cq, void *ctx_ptr)550{551	struct ipoib_dev_priv *priv = ctx_ptr;552	bool ret;553 554	ret = napi_schedule(&priv->send_napi);555	/*556	 * if the queue is closed the driver must be able to schedule napi,557	 * otherwise we can end with closed queue forever, because no new558	 * packets to send and napi callback might not get new event after559	 * its re-arm of the napi.560	 */561	if (!ret && netif_queue_stopped(priv->dev))562		schedule_work(&priv->reschedule_napi_work);563}564 565static inline int post_send(struct ipoib_dev_priv *priv,566			    unsigned int wr_id,567			    struct ib_ah *address, u32 dqpn,568			    struct ipoib_tx_buf *tx_req,569			    void *head, int hlen)570{571	struct sk_buff *skb = tx_req->skb;572 573	ipoib_build_sge(priv, tx_req);574 575	priv->tx_wr.wr.wr_id	= wr_id;576	priv->tx_wr.remote_qpn	= dqpn;577	priv->tx_wr.ah		= address;578 579	if (head) {580		priv->tx_wr.mss		= skb_shinfo(skb)->gso_size;581		priv->tx_wr.header	= head;582		priv->tx_wr.hlen	= hlen;583		priv->tx_wr.wr.opcode	= IB_WR_LSO;584	} else585		priv->tx_wr.wr.opcode	= IB_WR_SEND;586 587	return ib_post_send(priv->qp, &priv->tx_wr.wr, NULL);588}589 590int ipoib_send(struct net_device *dev, struct sk_buff *skb,591	       struct ib_ah *address, u32 dqpn)592{593	struct ipoib_dev_priv *priv = ipoib_priv(dev);594	struct ipoib_tx_buf *tx_req;595	int hlen, rc;596	void *phead;597	unsigned int usable_sge = priv->max_send_sge - !!skb_headlen(skb);598 599	if (skb_is_gso(skb)) {600		hlen = skb_tcp_all_headers(skb);601		phead = skb->data;602		if (unlikely(!skb_pull(skb, hlen))) {603			ipoib_warn(priv, "linear data too small\n");604			++dev->stats.tx_dropped;605			++dev->stats.tx_errors;606			dev_kfree_skb_any(skb);607			return -1;608		}609	} else {610		if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) {611			ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",612				   skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN);613			++dev->stats.tx_dropped;614			++dev->stats.tx_errors;615			ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu);616			return -1;617		}618		phead = NULL;619		hlen  = 0;620	}621	if (skb_shinfo(skb)->nr_frags > usable_sge) {622		if (skb_linearize(skb) < 0) {623			ipoib_warn(priv, "skb could not be linearized\n");624			++dev->stats.tx_dropped;625			++dev->stats.tx_errors;626			dev_kfree_skb_any(skb);627			return -1;628		}629		/* Does skb_linearize return ok without reducing nr_frags? */630		if (skb_shinfo(skb)->nr_frags > usable_sge) {631			ipoib_warn(priv, "too many frags after skb linearize\n");632			++dev->stats.tx_dropped;633			++dev->stats.tx_errors;634			dev_kfree_skb_any(skb);635			return -1;636		}637	}638 639	ipoib_dbg_data(priv,640		       "sending packet, length=%d address=%p dqpn=0x%06x\n",641		       skb->len, address, dqpn);642 643	/*644	 * We put the skb into the tx_ring _before_ we call post_send()645	 * because it's entirely possible that the completion handler will646	 * run before we execute anything after the post_send().  That647	 * means we have to make sure everything is properly recorded and648	 * our state is consistent before we call post_send().649	 */650	tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)];651	tx_req->skb = skb;652	if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {653		++dev->stats.tx_errors;654		dev_kfree_skb_any(skb);655		return -1;656	}657 658	if (skb->ip_summed == CHECKSUM_PARTIAL)659		priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM;660	else661		priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;662	/* increase the tx_head after send success, but use it for queue state */663	if ((priv->global_tx_head - priv->global_tx_tail) ==664	    ipoib_sendq_size - 1) {665		ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n");666		netif_stop_queue(dev);667	}668 669	skb_orphan(skb);670	skb_dst_drop(skb);671 672	if (netif_queue_stopped(dev))673		if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |674				     IB_CQ_REPORT_MISSED_EVENTS) < 0)675			ipoib_warn(priv, "request notify on send CQ failed\n");676 677	rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1),678		       address, dqpn, tx_req, phead, hlen);679	if (unlikely(rc)) {680		ipoib_warn(priv, "post_send failed, error %d\n", rc);681		++dev->stats.tx_errors;682		ipoib_dma_unmap_tx(priv, tx_req);683		dev_kfree_skb_any(skb);684		if (netif_queue_stopped(dev))685			netif_wake_queue(dev);686		rc = 0;687	} else {688		netif_trans_update(dev);689 690		rc = priv->tx_head;691		++priv->tx_head;692		++priv->global_tx_head;693	}694	return rc;695}696 697static void ipoib_reap_dead_ahs(struct ipoib_dev_priv *priv)698{699	struct ipoib_ah *ah, *tah;700	unsigned long flags;701 702	netif_tx_lock_bh(priv->dev);703	spin_lock_irqsave(&priv->lock, flags);704 705	list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list)706		if ((int) priv->tx_tail - (int) ah->last_send >= 0) {707			list_del(&ah->list);708			rdma_destroy_ah(ah->ah, 0);709			kfree(ah);710		}711 712	spin_unlock_irqrestore(&priv->lock, flags);713	netif_tx_unlock_bh(priv->dev);714}715 716void ipoib_reap_ah(struct work_struct *work)717{718	struct ipoib_dev_priv *priv =719		container_of(work, struct ipoib_dev_priv, ah_reap_task.work);720 721	ipoib_reap_dead_ahs(priv);722 723	if (!test_bit(IPOIB_STOP_REAPER, &priv->flags))724		queue_delayed_work(priv->wq, &priv->ah_reap_task,725				   round_jiffies_relative(HZ));726}727 728static void ipoib_start_ah_reaper(struct ipoib_dev_priv *priv)729{730	clear_bit(IPOIB_STOP_REAPER, &priv->flags);731	queue_delayed_work(priv->wq, &priv->ah_reap_task,732			   round_jiffies_relative(HZ));733}734 735static void ipoib_stop_ah_reaper(struct ipoib_dev_priv *priv)736{737	set_bit(IPOIB_STOP_REAPER, &priv->flags);738	cancel_delayed_work(&priv->ah_reap_task);739	/*740	 * After ipoib_stop_ah_reaper() we always go through741	 * ipoib_reap_dead_ahs() which ensures the work is really stopped and742	 * does a final flush out of the dead_ah's list743	 */744}745 746static int recvs_pending(struct net_device *dev)747{748	struct ipoib_dev_priv *priv = ipoib_priv(dev);749	int pending = 0;750	int i;751 752	for (i = 0; i < ipoib_recvq_size; ++i)753		if (priv->rx_ring[i].skb)754			++pending;755 756	return pending;757}758 759static void check_qp_movement_and_print(struct ipoib_dev_priv *priv,760					struct ib_qp *qp,761					enum ib_qp_state new_state)762{763	struct ib_qp_attr qp_attr;764	struct ib_qp_init_attr query_init_attr;765	int ret;766 767	ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr);768	if (ret) {769		ipoib_warn(priv, "%s: Failed to query QP\n", __func__);770		return;771	}772	/* print according to the new-state and the previous state.*/773	if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET)774		ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n");775	else776		ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n",777			   new_state, qp_attr.qp_state);778}779 780static void ipoib_napi_enable(struct net_device *dev)781{782	struct ipoib_dev_priv *priv = ipoib_priv(dev);783 784	napi_enable(&priv->recv_napi);785	napi_enable(&priv->send_napi);786}787 788static void ipoib_napi_disable(struct net_device *dev)789{790	struct ipoib_dev_priv *priv = ipoib_priv(dev);791 792	napi_disable(&priv->recv_napi);793	napi_disable(&priv->send_napi);794}795 796int ipoib_ib_dev_stop_default(struct net_device *dev)797{798	struct ipoib_dev_priv *priv = ipoib_priv(dev);799	struct ib_qp_attr qp_attr;800	unsigned long begin;801	struct ipoib_tx_buf *tx_req;802	int i;803 804	if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))805		ipoib_napi_disable(dev);806 807	ipoib_cm_dev_stop(dev);808 809	/*810	 * Move our QP to the error state and then reinitialize in811	 * when all work requests have completed or have been flushed.812	 */813	qp_attr.qp_state = IB_QPS_ERR;814	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))815		check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR);816 817	/* Wait for all sends and receives to complete */818	begin = jiffies;819 820	while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) {821		if (time_after(jiffies, begin + 5 * HZ)) {822			ipoib_warn(priv,823				   "timing out; %d sends %d receives not completed\n",824				   priv->tx_head - priv->tx_tail,825				   recvs_pending(dev));826 827			/*828			 * assume the HW is wedged and just free up829			 * all our pending work requests.830			 */831			while ((int)priv->tx_tail - (int)priv->tx_head < 0) {832				tx_req = &priv->tx_ring[priv->tx_tail &833							(ipoib_sendq_size - 1)];834				ipoib_dma_unmap_tx(priv, tx_req);835				dev_kfree_skb_any(tx_req->skb);836				++priv->tx_tail;837				++priv->global_tx_tail;838			}839 840			for (i = 0; i < ipoib_recvq_size; ++i) {841				struct ipoib_rx_buf *rx_req;842 843				rx_req = &priv->rx_ring[i];844				if (!rx_req->skb)845					continue;846				ipoib_ud_dma_unmap_rx(priv,847						      priv->rx_ring[i].mapping);848				dev_kfree_skb_any(rx_req->skb);849				rx_req->skb = NULL;850			}851 852			goto timeout;853		}854 855		ipoib_drain_cq(dev);856 857		usleep_range(1000, 2000);858	}859 860	ipoib_dbg(priv, "All sends and receives done.\n");861 862timeout:863	qp_attr.qp_state = IB_QPS_RESET;864	if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE))865		ipoib_warn(priv, "Failed to modify QP to RESET state\n");866 867	ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP);868 869	return 0;870}871 872int ipoib_ib_dev_open_default(struct net_device *dev)873{874	struct ipoib_dev_priv *priv = ipoib_priv(dev);875	int ret;876 877	ret = ipoib_init_qp(dev);878	if (ret) {879		ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret);880		return -1;881	}882 883	ret = ipoib_ib_post_receives(dev);884	if (ret) {885		ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret);886		goto out;887	}888 889	ret = ipoib_cm_dev_open(dev);890	if (ret) {891		ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret);892		goto out;893	}894 895	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))896		ipoib_napi_enable(dev);897 898	return 0;899out:900	return -1;901}902 903int ipoib_ib_dev_open(struct net_device *dev)904{905	struct ipoib_dev_priv *priv = ipoib_priv(dev);906 907	ipoib_pkey_dev_check_presence(dev);908 909	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {910		ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey,911			   (!(priv->pkey & 0x7fff) ? "Invalid" : "not found"));912		return -1;913	}914 915	ipoib_start_ah_reaper(priv);916	if (priv->rn_ops->ndo_open(dev)) {917		pr_warn("%s: Failed to open dev\n", dev->name);918		goto dev_stop;919	}920 921	set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);922 923	return 0;924 925dev_stop:926	ipoib_stop_ah_reaper(priv);927	return -1;928}929 930void ipoib_ib_dev_stop(struct net_device *dev)931{932	struct ipoib_dev_priv *priv = ipoib_priv(dev);933 934	priv->rn_ops->ndo_stop(dev);935 936	clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags);937	ipoib_stop_ah_reaper(priv);938}939 940void ipoib_pkey_dev_check_presence(struct net_device *dev)941{942	struct ipoib_dev_priv *priv = ipoib_priv(dev);943	struct rdma_netdev *rn = netdev_priv(dev);944 945	if (!(priv->pkey & 0x7fff) ||946	    ib_find_pkey(priv->ca, priv->port, priv->pkey,947			 &priv->pkey_index)) {948		clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);949	} else {950		if (rn->set_id)951			rn->set_id(dev, priv->pkey_index);952		set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);953	}954}955 956void ipoib_ib_dev_up(struct net_device *dev)957{958	struct ipoib_dev_priv *priv = ipoib_priv(dev);959 960	ipoib_pkey_dev_check_presence(dev);961 962	if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) {963		ipoib_dbg(priv, "PKEY is not assigned.\n");964		return;965	}966 967	set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);968 969	ipoib_mcast_start_thread(dev);970}971 972void ipoib_ib_dev_down(struct net_device *dev)973{974	struct ipoib_dev_priv *priv = ipoib_priv(dev);975 976	ipoib_dbg(priv, "downing ib_dev\n");977 978	clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);979	netif_carrier_off(dev);980 981	ipoib_mcast_stop_thread(dev);982	ipoib_mcast_dev_flush(dev);983 984	ipoib_flush_paths(dev);985}986 987void ipoib_drain_cq(struct net_device *dev)988{989	struct ipoib_dev_priv *priv = ipoib_priv(dev);990	int i, n;991 992	/*993	 * We call completion handling routines that expect to be994	 * called from the BH-disabled NAPI poll context, so disable995	 * BHs here too.996	 */997	local_bh_disable();998 999	do {1000		n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc);1001		for (i = 0; i < n; ++i) {1002			/*1003			 * Convert any successful completions to flush1004			 * errors to avoid passing packets up the1005			 * stack after bringing the device down.1006			 */1007			if (priv->ibwc[i].status == IB_WC_SUCCESS)1008				priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR;1009 1010			if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) {1011				if (priv->ibwc[i].wr_id & IPOIB_OP_CM)1012					ipoib_cm_handle_rx_wc(dev, priv->ibwc + i);1013				else1014					ipoib_ib_handle_rx_wc(dev, priv->ibwc + i);1015			} else {1016				pr_warn("%s: Got unexpected wqe id\n", __func__);1017			}1018		}1019	} while (n == IPOIB_NUM_WC);1020 1021	while (poll_tx(priv))1022		; /* nothing */1023 1024	local_bh_enable();1025}1026 1027/*1028 * Takes whatever value which is in pkey index 0 and updates priv->pkey1029 * returns 0 if the pkey value was changed.1030 */1031static inline int update_parent_pkey(struct ipoib_dev_priv *priv)1032{1033	int result;1034	u16 prev_pkey;1035 1036	prev_pkey = priv->pkey;1037	result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);1038	if (result) {1039		ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n",1040			   priv->port, result);1041		return result;1042	}1043 1044	priv->pkey |= 0x8000;1045 1046	if (prev_pkey != priv->pkey) {1047		ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n",1048			  prev_pkey, priv->pkey);1049		/*1050		 * Update the pkey in the broadcast address, while making sure to set1051		 * the full membership bit, so that we join the right broadcast group.1052		 */1053		priv->dev->broadcast[8] = priv->pkey >> 8;1054		priv->dev->broadcast[9] = priv->pkey & 0xff;1055		return 0;1056	}1057 1058	return 1;1059}1060/*1061 * returns 0 if pkey value was found in a different slot.1062 */1063static inline int update_child_pkey(struct ipoib_dev_priv *priv)1064{1065	u16 old_index = priv->pkey_index;1066 1067	priv->pkey_index = 0;1068	ipoib_pkey_dev_check_presence(priv->dev);1069 1070	if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) &&1071	    (old_index == priv->pkey_index))1072		return 1;1073	return 0;1074}1075 1076/*1077 * returns true if the device address of the ipoib interface has changed and the1078 * new address is a valid one (i.e in the gid table), return false otherwise.1079 */1080static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv)1081{1082	union ib_gid search_gid;1083	union ib_gid gid0;1084	int err;1085	u16 index;1086	u32 port;1087	bool ret = false;1088 1089	if (rdma_query_gid(priv->ca, priv->port, 0, &gid0))1090		return false;1091 1092	netif_addr_lock_bh(priv->dev);1093 1094	/* The subnet prefix may have changed, update it now so we won't have1095	 * to do it later1096	 */1097	priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix;1098	dev_addr_mod(priv->dev, 4, (u8 *)&gid0.global.subnet_prefix,1099		     sizeof(gid0.global.subnet_prefix));1100	search_gid.global.subnet_prefix = gid0.global.subnet_prefix;1101 1102	search_gid.global.interface_id = priv->local_gid.global.interface_id;1103 1104	netif_addr_unlock_bh(priv->dev);1105 1106	err = ib_find_gid(priv->ca, &search_gid, &port, &index);1107 1108	netif_addr_lock_bh(priv->dev);1109 1110	if (search_gid.global.interface_id !=1111	    priv->local_gid.global.interface_id)1112		/* There was a change while we were looking up the gid, bail1113		 * here and let the next work sort this out1114		 */1115		goto out;1116 1117	/* The next section of code needs some background:1118	 * Per IB spec the port GUID can't change if the HCA is powered on.1119	 * port GUID is the basis for GID at index 0 which is the basis for1120	 * the default device address of a ipoib interface.1121	 *1122	 * so it seems the flow should be:1123	 * if user_changed_dev_addr && gid in gid tbl1124	 *	set bit dev_addr_set1125	 *	return true1126	 * else1127	 *	return false1128	 *1129	 * The issue is that there are devices that don't follow the spec,1130	 * they change the port GUID when the HCA is powered, so in order1131	 * not to break userspace applications, We need to check if the1132	 * user wanted to control the device address and we assume that1133	 * if he sets the device address back to be based on GID index 0,1134	 * he no longer wishs to control it.1135	 *1136	 * If the user doesn't control the device address,1137	 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means1138	 * the port GUID has changed and GID at index 0 has changed1139	 * so we need to change priv->local_gid and priv->dev->dev_addr1140	 * to reflect the new GID.1141	 */1142	if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) {1143		if (!err && port == priv->port) {1144			set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);1145			if (index == 0)1146				clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL,1147					  &priv->flags);1148			else1149				set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags);1150			ret = true;1151		} else {1152			ret = false;1153		}1154	} else {1155		if (!err && port == priv->port) {1156			ret = true;1157		} else {1158			if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) {1159				memcpy(&priv->local_gid, &gid0,1160				       sizeof(priv->local_gid));1161				dev_addr_mod(priv->dev, 4, (u8 *)&gid0,1162					     sizeof(priv->local_gid));1163				ret = true;1164			}1165		}1166	}1167 1168out:1169	netif_addr_unlock_bh(priv->dev);1170 1171	return ret;1172}1173 1174static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv,1175				enum ipoib_flush_level level,1176				int nesting)1177{1178	struct ipoib_dev_priv *cpriv;1179	struct net_device *dev = priv->dev;1180	int result;1181 1182	down_read_nested(&priv->vlan_rwsem, nesting);1183 1184	/*1185	 * Flush any child interfaces too -- they might be up even if1186	 * the parent is down.1187	 */1188	list_for_each_entry(cpriv, &priv->child_intfs, list)1189		__ipoib_ib_dev_flush(cpriv, level, nesting + 1);1190 1191	up_read(&priv->vlan_rwsem);1192 1193	if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) &&1194	    level != IPOIB_FLUSH_HEAVY) {1195		/* Make sure the dev_addr is set even if not flushing */1196		if (level == IPOIB_FLUSH_LIGHT)1197			ipoib_dev_addr_changed_valid(priv);1198		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n");1199		return;1200	}1201 1202	if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {1203		/* interface is down. update pkey and leave. */1204		if (level == IPOIB_FLUSH_HEAVY) {1205			if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))1206				update_parent_pkey(priv);1207			else1208				update_child_pkey(priv);1209		} else if (level == IPOIB_FLUSH_LIGHT)1210			ipoib_dev_addr_changed_valid(priv);1211		ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n");1212		return;1213	}1214 1215	if (level == IPOIB_FLUSH_HEAVY) {1216		/* child devices chase their origin pkey value, while non-child1217		 * (parent) devices should always takes what present in pkey index 01218		 */1219		if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {1220			result = update_child_pkey(priv);1221			if (result) {1222				/* restart QP only if P_Key index is changed */1223				ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n");1224				return;1225			}1226 1227		} else {1228			result = update_parent_pkey(priv);1229			/* restart QP only if P_Key value changed */1230			if (result) {1231				ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n");1232				return;1233			}1234		}1235	}1236 1237	if (level == IPOIB_FLUSH_LIGHT) {1238		int oper_up;1239		ipoib_mark_paths_invalid(dev);1240		/* Set IPoIB operation as down to prevent races between:1241		 * the flush flow which leaves MCG and on the fly joins1242		 * which can happen during that time. mcast restart task1243		 * should deal with join requests we missed.1244		 */1245		oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags);1246		ipoib_mcast_dev_flush(dev);1247		if (oper_up)1248			set_bit(IPOIB_FLAG_OPER_UP, &priv->flags);1249		ipoib_reap_dead_ahs(priv);1250	}1251 1252	if (level >= IPOIB_FLUSH_NORMAL)1253		ipoib_ib_dev_down(dev);1254 1255	if (level == IPOIB_FLUSH_HEAVY) {1256		if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags))1257			ipoib_ib_dev_stop(dev);1258 1259		if (ipoib_ib_dev_open(dev))1260			return;1261 1262		if (netif_queue_stopped(dev))1263			netif_start_queue(dev);1264	}1265 1266	/*1267	 * The device could have been brought down between the start and when1268	 * we get here, don't bring it back up if it's not configured up1269	 */1270	if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) {1271		if (level >= IPOIB_FLUSH_NORMAL)1272			ipoib_ib_dev_up(dev);1273		if (ipoib_dev_addr_changed_valid(priv))1274			ipoib_mcast_restart_task(&priv->restart_task);1275	}1276}1277 1278void ipoib_ib_dev_flush_light(struct work_struct *work)1279{1280	struct ipoib_dev_priv *priv =1281		container_of(work, struct ipoib_dev_priv, flush_light);1282 1283	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0);1284}1285 1286void ipoib_ib_dev_flush_normal(struct work_struct *work)1287{1288	struct ipoib_dev_priv *priv =1289		container_of(work, struct ipoib_dev_priv, flush_normal);1290 1291	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0);1292}1293 1294void ipoib_ib_dev_flush_heavy(struct work_struct *work)1295{1296	struct ipoib_dev_priv *priv =1297		container_of(work, struct ipoib_dev_priv, flush_heavy);1298 1299	rtnl_lock();1300	__ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0);1301	rtnl_unlock();1302}1303 1304void ipoib_ib_dev_cleanup(struct net_device *dev)1305{1306	struct ipoib_dev_priv *priv = ipoib_priv(dev);1307 1308	ipoib_dbg(priv, "cleaning up ib_dev\n");1309	/*1310	 * We must make sure there are no more (path) completions1311	 * that may wish to touch priv fields that are no longer valid1312	 */1313	ipoib_flush_paths(dev);1314 1315	ipoib_mcast_stop_thread(dev);1316	ipoib_mcast_dev_flush(dev);1317 1318	/*1319	 * All of our ah references aren't free until after1320	 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and1321	 * the neighbor garbage collection is stopped and reaped.1322	 * That should all be done now, so make a final ah flush.1323	 */1324	ipoib_reap_dead_ahs(priv);1325 1326	clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags);1327 1328	priv->rn_ops->ndo_uninit(dev);1329 1330	if (priv->pd) {1331		ib_dealloc_pd(priv->pd);1332		priv->pd = NULL;1333	}1334}1335 1336