xref: /linux/drivers/infiniband/hw/mlx4/mad.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_sa.h>
36 #include <rdma/ib_cache.h>
37 
38 #include <linux/random.h>
39 #include <linux/mlx4/cmd.h>
40 #include <linux/gfp.h>
41 #include <rdma/ib_pma.h>
42 
43 #include "mlx4_ib.h"
44 
45 enum {
46 	MLX4_IB_VENDOR_CLASS1 = 0x9,
47 	MLX4_IB_VENDOR_CLASS2 = 0xa
48 };
49 
50 #define MLX4_TUN_SEND_WRID_SHIFT 34
51 #define MLX4_TUN_QPN_SHIFT 32
52 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
53 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
54 
55 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
56 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
57 
58  /* Port mgmt change event handling */
59 
60 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
61 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
62 #define NUM_IDX_IN_PKEY_TBL_BLK 32
63 #define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
64 #define GUID_TBL_BLK_NUM_ENTRIES 8
65 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
66 
67 struct mlx4_mad_rcv_buf {
68 	struct ib_grh grh;
69 	u8 payload[256];
70 } __packed;
71 
72 struct mlx4_mad_snd_buf {
73 	u8 payload[256];
74 } __packed;
75 
76 struct mlx4_tunnel_mad {
77 	struct ib_grh grh;
78 	struct mlx4_ib_tunnel_header hdr;
79 	struct ib_mad mad;
80 } __packed;
81 
82 struct mlx4_rcv_tunnel_mad {
83 	struct mlx4_rcv_tunnel_hdr hdr;
84 	struct ib_grh grh;
85 	struct ib_mad mad;
86 } __packed;
87 
88 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
89 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
90 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
91 				int block, u32 change_bitmap);
92 
93 __be64 mlx4_ib_gen_node_guid(void)
94 {
95 #define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
96 	return cpu_to_be64(NODE_GUID_HI | prandom_u32());
97 }
98 
99 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
100 {
101 	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
102 		cpu_to_be64(0xff00000000000000LL);
103 }
104 
105 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
106 		 int port, const struct ib_wc *in_wc,
107 		 const struct ib_grh *in_grh,
108 		 const void *in_mad, void *response_mad)
109 {
110 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
111 	void *inbox;
112 	int err;
113 	u32 in_modifier = port;
114 	u8 op_modifier = 0;
115 
116 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
117 	if (IS_ERR(inmailbox))
118 		return PTR_ERR(inmailbox);
119 	inbox = inmailbox->buf;
120 
121 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
122 	if (IS_ERR(outmailbox)) {
123 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
124 		return PTR_ERR(outmailbox);
125 	}
126 
127 	memcpy(inbox, in_mad, 256);
128 
129 	/*
130 	 * Key check traps can't be generated unless we have in_wc to
131 	 * tell us where to send the trap.
132 	 */
133 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
134 		op_modifier |= 0x1;
135 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
136 		op_modifier |= 0x2;
137 	if (mlx4_is_mfunc(dev->dev) &&
138 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
139 		op_modifier |= 0x8;
140 
141 	if (in_wc) {
142 		struct {
143 			__be32		my_qpn;
144 			u32		reserved1;
145 			__be32		rqpn;
146 			u8		sl;
147 			u8		g_path;
148 			u16		reserved2[2];
149 			__be16		pkey;
150 			u32		reserved3[11];
151 			u8		grh[40];
152 		} *ext_info;
153 
154 		memset(inbox + 256, 0, 256);
155 		ext_info = inbox + 256;
156 
157 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
158 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
159 		ext_info->sl     = in_wc->sl << 4;
160 		ext_info->g_path = in_wc->dlid_path_bits |
161 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
162 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
163 
164 		if (in_grh)
165 			memcpy(ext_info->grh, in_grh, 40);
166 
167 		op_modifier |= 0x4;
168 
169 		in_modifier |= in_wc->slid << 16;
170 	}
171 
172 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
173 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
174 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
175 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
176 
177 	if (!err)
178 		memcpy(response_mad, outmailbox->buf, 256);
179 
180 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
181 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
182 
183 	return err;
184 }
185 
186 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
187 {
188 	struct ib_ah *new_ah;
189 	struct ib_ah_attr ah_attr;
190 	unsigned long flags;
191 
192 	if (!dev->send_agent[port_num - 1][0])
193 		return;
194 
195 	memset(&ah_attr, 0, sizeof ah_attr);
196 	ah_attr.dlid     = lid;
197 	ah_attr.sl       = sl;
198 	ah_attr.port_num = port_num;
199 
200 	new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
201 			      &ah_attr);
202 	if (IS_ERR(new_ah))
203 		return;
204 
205 	spin_lock_irqsave(&dev->sm_lock, flags);
206 	if (dev->sm_ah[port_num - 1])
207 		ib_destroy_ah(dev->sm_ah[port_num - 1]);
208 	dev->sm_ah[port_num - 1] = new_ah;
209 	spin_unlock_irqrestore(&dev->sm_lock, flags);
210 }
211 
212 /*
213  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
214  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
215  */
216 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
217 		      u16 prev_lid)
218 {
219 	struct ib_port_info *pinfo;
220 	u16 lid;
221 	__be16 *base;
222 	u32 bn, pkey_change_bitmap;
223 	int i;
224 
225 
226 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
227 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
228 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
229 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
230 		switch (mad->mad_hdr.attr_id) {
231 		case IB_SMP_ATTR_PORT_INFO:
232 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
233 			lid = be16_to_cpu(pinfo->lid);
234 
235 			update_sm_ah(dev, port_num,
236 				     be16_to_cpu(pinfo->sm_lid),
237 				     pinfo->neighbormtu_mastersmsl & 0xf);
238 
239 			if (pinfo->clientrereg_resv_subnetto & 0x80)
240 				handle_client_rereg_event(dev, port_num);
241 
242 			if (prev_lid != lid)
243 				handle_lid_change_event(dev, port_num);
244 			break;
245 
246 		case IB_SMP_ATTR_PKEY_TABLE:
247 			if (!mlx4_is_mfunc(dev->dev)) {
248 				mlx4_ib_dispatch_event(dev, port_num,
249 						       IB_EVENT_PKEY_CHANGE);
250 				break;
251 			}
252 
253 			/* at this point, we are running in the master.
254 			 * Slaves do not receive SMPs.
255 			 */
256 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
257 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
258 			pkey_change_bitmap = 0;
259 			for (i = 0; i < 32; i++) {
260 				pr_debug("PKEY[%d] = x%x\n",
261 					 i + bn*32, be16_to_cpu(base[i]));
262 				if (be16_to_cpu(base[i]) !=
263 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
264 					pkey_change_bitmap |= (1 << i);
265 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
266 						be16_to_cpu(base[i]);
267 				}
268 			}
269 			pr_debug("PKEY Change event: port=%d, "
270 				 "block=0x%x, change_bitmap=0x%x\n",
271 				 port_num, bn, pkey_change_bitmap);
272 
273 			if (pkey_change_bitmap) {
274 				mlx4_ib_dispatch_event(dev, port_num,
275 						       IB_EVENT_PKEY_CHANGE);
276 				if (!dev->sriov.is_going_down)
277 					__propagate_pkey_ev(dev, port_num, bn,
278 							    pkey_change_bitmap);
279 			}
280 			break;
281 
282 		case IB_SMP_ATTR_GUID_INFO:
283 			/* paravirtualized master's guid is guid 0 -- does not change */
284 			if (!mlx4_is_master(dev->dev))
285 				mlx4_ib_dispatch_event(dev, port_num,
286 						       IB_EVENT_GID_CHANGE);
287 			/*if master, notify relevant slaves*/
288 			if (mlx4_is_master(dev->dev) &&
289 			    !dev->sriov.is_going_down) {
290 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
291 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
292 								    (u8 *)(&((struct ib_smp *)mad)->data));
293 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
294 								     (u8 *)(&((struct ib_smp *)mad)->data));
295 			}
296 			break;
297 
298 		default:
299 			break;
300 		}
301 }
302 
303 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
304 				int block, u32 change_bitmap)
305 {
306 	int i, ix, slave, err;
307 	int have_event = 0;
308 
309 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
310 		if (slave == mlx4_master_func_num(dev->dev))
311 			continue;
312 		if (!mlx4_is_slave_active(dev->dev, slave))
313 			continue;
314 
315 		have_event = 0;
316 		for (i = 0; i < 32; i++) {
317 			if (!(change_bitmap & (1 << i)))
318 				continue;
319 			for (ix = 0;
320 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
321 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
322 				    [ix] == i + 32 * block) {
323 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
324 					pr_debug("propagate_pkey_ev: slave %d,"
325 						 " port %d, ix %d (%d)\n",
326 						 slave, port_num, ix, err);
327 					have_event = 1;
328 					break;
329 				}
330 			}
331 			if (have_event)
332 				break;
333 		}
334 	}
335 }
336 
337 static void node_desc_override(struct ib_device *dev,
338 			       struct ib_mad *mad)
339 {
340 	unsigned long flags;
341 
342 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
343 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
344 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
345 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
346 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
347 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
348 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
349 	}
350 }
351 
352 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
353 {
354 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
355 	struct ib_mad_send_buf *send_buf;
356 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
357 	int ret;
358 	unsigned long flags;
359 
360 	if (agent) {
361 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
362 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
363 					      IB_MGMT_BASE_VERSION);
364 		if (IS_ERR(send_buf))
365 			return;
366 		/*
367 		 * We rely here on the fact that MLX QPs don't use the
368 		 * address handle after the send is posted (this is
369 		 * wrong following the IB spec strictly, but we know
370 		 * it's OK for our devices).
371 		 */
372 		spin_lock_irqsave(&dev->sm_lock, flags);
373 		memcpy(send_buf->mad, mad, sizeof *mad);
374 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
375 			ret = ib_post_send_mad(send_buf, NULL);
376 		else
377 			ret = -EINVAL;
378 		spin_unlock_irqrestore(&dev->sm_lock, flags);
379 
380 		if (ret)
381 			ib_free_send_mad(send_buf);
382 	}
383 }
384 
385 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
386 							     struct ib_sa_mad *sa_mad)
387 {
388 	int ret = 0;
389 
390 	/* dispatch to different sa handlers */
391 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
392 	case IB_SA_ATTR_MC_MEMBER_REC:
393 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
394 		break;
395 	default:
396 		break;
397 	}
398 	return ret;
399 }
400 
401 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
402 {
403 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
404 	int i;
405 
406 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
407 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
408 			return i;
409 	}
410 	return -1;
411 }
412 
413 
414 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
415 				   u8 port, u16 pkey, u16 *ix)
416 {
417 	int i, ret;
418 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
419 	u16 slot_pkey;
420 
421 	if (slave == mlx4_master_func_num(dev->dev))
422 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
423 
424 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
425 
426 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
427 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
428 			continue;
429 
430 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
431 
432 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
433 		if (ret)
434 			continue;
435 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
436 			if (slot_pkey & 0x8000) {
437 				*ix = (u16) pkey_ix;
438 				return 0;
439 			} else {
440 				/* take first partial pkey index found */
441 				if (partial_ix == 0xFF)
442 					partial_ix = pkey_ix;
443 			}
444 		}
445 	}
446 
447 	if (partial_ix < 0xFF) {
448 		*ix = (u16) partial_ix;
449 		return 0;
450 	}
451 
452 	return -EINVAL;
453 }
454 
455 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
456 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
457 			  struct ib_grh *grh, struct ib_mad *mad)
458 {
459 	struct ib_sge list;
460 	struct ib_send_wr wr, *bad_wr;
461 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
462 	struct mlx4_ib_demux_pv_qp *tun_qp;
463 	struct mlx4_rcv_tunnel_mad *tun_mad;
464 	struct ib_ah_attr attr;
465 	struct ib_ah *ah;
466 	struct ib_qp *src_qp = NULL;
467 	unsigned tun_tx_ix = 0;
468 	int dqpn;
469 	int ret = 0;
470 	u16 tun_pkey_ix;
471 	u16 cached_pkey;
472 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
473 
474 	if (dest_qpt > IB_QPT_GSI)
475 		return -EINVAL;
476 
477 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
478 
479 	/* check if proxy qp created */
480 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
481 		return -EAGAIN;
482 
483 	if (!dest_qpt)
484 		tun_qp = &tun_ctx->qp[0];
485 	else
486 		tun_qp = &tun_ctx->qp[1];
487 
488 	/* compute P_Key index to put in tunnel header for slave */
489 	if (dest_qpt) {
490 		u16 pkey_ix;
491 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
492 		if (ret)
493 			return -EINVAL;
494 
495 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
496 		if (ret)
497 			return -EINVAL;
498 		tun_pkey_ix = pkey_ix;
499 	} else
500 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
501 
502 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
503 
504 	/* get tunnel tx data buf for slave */
505 	src_qp = tun_qp->qp;
506 
507 	/* create ah. Just need an empty one with the port num for the post send.
508 	 * The driver will set the force loopback bit in post_send */
509 	memset(&attr, 0, sizeof attr);
510 	attr.port_num = port;
511 	if (is_eth) {
512 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
513 		attr.ah_flags = IB_AH_GRH;
514 	}
515 	ah = ib_create_ah(tun_ctx->pd, &attr);
516 	if (IS_ERR(ah))
517 		return -ENOMEM;
518 
519 	/* allocate tunnel tx buf after pass failure returns */
520 	spin_lock(&tun_qp->tx_lock);
521 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
522 	    (MLX4_NUM_TUNNEL_BUFS - 1))
523 		ret = -EAGAIN;
524 	else
525 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
526 	spin_unlock(&tun_qp->tx_lock);
527 	if (ret)
528 		goto out;
529 
530 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
531 	if (tun_qp->tx_ring[tun_tx_ix].ah)
532 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
533 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
534 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
535 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
536 				   sizeof (struct mlx4_rcv_tunnel_mad),
537 				   DMA_TO_DEVICE);
538 
539 	/* copy over to tunnel buffer */
540 	if (grh)
541 		memcpy(&tun_mad->grh, grh, sizeof *grh);
542 	memcpy(&tun_mad->mad, mad, sizeof *mad);
543 
544 	/* adjust tunnel data */
545 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
546 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
547 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
548 
549 	if (is_eth) {
550 		u16 vlan = 0;
551 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
552 						NULL)) {
553 			/* VST mode */
554 			if (vlan != wc->vlan_id)
555 				/* Packet vlan is not the VST-assigned vlan.
556 				 * Drop the packet.
557 				 */
558 				goto out;
559 			 else
560 				/* Remove the vlan tag before forwarding
561 				 * the packet to the VF.
562 				 */
563 				vlan = 0xffff;
564 		} else {
565 			vlan = wc->vlan_id;
566 		}
567 
568 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
569 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
570 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
571 	} else {
572 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
573 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
574 	}
575 
576 	ib_dma_sync_single_for_device(&dev->ib_dev,
577 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
578 				      sizeof (struct mlx4_rcv_tunnel_mad),
579 				      DMA_TO_DEVICE);
580 
581 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
582 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
583 	list.lkey = tun_ctx->pd->local_dma_lkey;
584 
585 	wr.wr.ud.ah = ah;
586 	wr.wr.ud.port_num = port;
587 	wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
588 	wr.wr.ud.remote_qpn = dqpn;
589 	wr.next = NULL;
590 	wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
591 	wr.sg_list = &list;
592 	wr.num_sge = 1;
593 	wr.opcode = IB_WR_SEND;
594 	wr.send_flags = IB_SEND_SIGNALED;
595 
596 	ret = ib_post_send(src_qp, &wr, &bad_wr);
597 out:
598 	if (ret)
599 		ib_destroy_ah(ah);
600 	return ret;
601 }
602 
603 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
604 			struct ib_wc *wc, struct ib_grh *grh,
605 			struct ib_mad *mad)
606 {
607 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
608 	int err;
609 	int slave;
610 	u8 *slave_id;
611 	int is_eth = 0;
612 
613 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
614 		is_eth = 0;
615 	else
616 		is_eth = 1;
617 
618 	if (is_eth) {
619 		if (!(wc->wc_flags & IB_WC_GRH)) {
620 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
621 			return -EINVAL;
622 		}
623 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
624 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
625 			return -EINVAL;
626 		}
627 		if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
628 			mlx4_ib_warn(ibdev, "failed matching grh\n");
629 			return -ENOENT;
630 		}
631 		if (slave >= dev->dev->caps.sqp_demux) {
632 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
633 				     slave, dev->dev->caps.sqp_demux);
634 			return -ENOENT;
635 		}
636 
637 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
638 			return 0;
639 
640 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
641 		if (err)
642 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
643 				 slave, err);
644 		return 0;
645 	}
646 
647 	/* Initially assume that this mad is for us */
648 	slave = mlx4_master_func_num(dev->dev);
649 
650 	/* See if the slave id is encoded in a response mad */
651 	if (mad->mad_hdr.method & 0x80) {
652 		slave_id = (u8 *) &mad->mad_hdr.tid;
653 		slave = *slave_id;
654 		if (slave != 255) /*255 indicates the dom0*/
655 			*slave_id = 0; /* remap tid */
656 	}
657 
658 	/* If a grh is present, we demux according to it */
659 	if (wc->wc_flags & IB_WC_GRH) {
660 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
661 		if (slave < 0) {
662 			mlx4_ib_warn(ibdev, "failed matching grh\n");
663 			return -ENOENT;
664 		}
665 	}
666 	/* Class-specific handling */
667 	switch (mad->mad_hdr.mgmt_class) {
668 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
669 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
670 		/* 255 indicates the dom0 */
671 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
672 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
673 				return -EPERM;
674 			/* for a VF. drop unsolicited MADs */
675 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
676 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
677 					     slave, mad->mad_hdr.mgmt_class,
678 					     mad->mad_hdr.method);
679 				return -EINVAL;
680 			}
681 		}
682 		break;
683 	case IB_MGMT_CLASS_SUBN_ADM:
684 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
685 					     (struct ib_sa_mad *) mad))
686 			return 0;
687 		break;
688 	case IB_MGMT_CLASS_CM:
689 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
690 			return 0;
691 		break;
692 	case IB_MGMT_CLASS_DEVICE_MGMT:
693 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
694 			return 0;
695 		break;
696 	default:
697 		/* Drop unsupported classes for slaves in tunnel mode */
698 		if (slave != mlx4_master_func_num(dev->dev)) {
699 			pr_debug("dropping unsupported ingress mad from class:%d "
700 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
701 			return 0;
702 		}
703 	}
704 	/*make sure that no slave==255 was not handled yet.*/
705 	if (slave >= dev->dev->caps.sqp_demux) {
706 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
707 			     slave, dev->dev->caps.sqp_demux);
708 		return -ENOENT;
709 	}
710 
711 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
712 	if (err)
713 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
714 			 slave, err);
715 	return 0;
716 }
717 
718 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
719 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
720 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
721 {
722 	u16 slid, prev_lid = 0;
723 	int err;
724 	struct ib_port_attr pattr;
725 
726 	if (in_wc && in_wc->qp->qp_num) {
727 		pr_debug("received MAD: slid:%d sqpn:%d "
728 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
729 			in_wc->slid, in_wc->src_qp,
730 			in_wc->dlid_path_bits,
731 			in_wc->qp->qp_num,
732 			in_wc->wc_flags,
733 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
734 			be16_to_cpu(in_mad->mad_hdr.attr_id));
735 		if (in_wc->wc_flags & IB_WC_GRH) {
736 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
737 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
738 				 be64_to_cpu(in_grh->sgid.global.interface_id));
739 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
740 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
741 				 be64_to_cpu(in_grh->dgid.global.interface_id));
742 		}
743 	}
744 
745 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
746 
747 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
748 		forward_trap(to_mdev(ibdev), port_num, in_mad);
749 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
750 	}
751 
752 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
753 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
754 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
755 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
756 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
757 			return IB_MAD_RESULT_SUCCESS;
758 
759 		/*
760 		 * Don't process SMInfo queries -- the SMA can't handle them.
761 		 */
762 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
763 			return IB_MAD_RESULT_SUCCESS;
764 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
765 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
766 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
767 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
768 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
769 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
770 			return IB_MAD_RESULT_SUCCESS;
771 	} else
772 		return IB_MAD_RESULT_SUCCESS;
773 
774 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
775 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
776 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
777 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
778 	    !ib_query_port(ibdev, port_num, &pattr))
779 		prev_lid = pattr.lid;
780 
781 	err = mlx4_MAD_IFC(to_mdev(ibdev),
782 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
783 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
784 			   MLX4_MAD_IFC_NET_VIEW,
785 			   port_num, in_wc, in_grh, in_mad, out_mad);
786 	if (err)
787 		return IB_MAD_RESULT_FAILURE;
788 
789 	if (!out_mad->mad_hdr.status) {
790 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
791 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
792 		/* slaves get node desc from FW */
793 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
794 			node_desc_override(ibdev, out_mad);
795 	}
796 
797 	/* set return bit in status of directed route responses */
798 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
799 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
800 
801 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
802 		/* no response for trap repress */
803 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
804 
805 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
806 }
807 
808 static void edit_counter(struct mlx4_counter *cnt,
809 					struct ib_pma_portcounters *pma_cnt)
810 {
811 	ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
812 			     (be64_to_cpu(cnt->tx_bytes) >> 2));
813 	ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
814 			     (be64_to_cpu(cnt->rx_bytes) >> 2));
815 	ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
816 			     be64_to_cpu(cnt->tx_frames));
817 	ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
818 			     be64_to_cpu(cnt->rx_frames));
819 }
820 
821 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
822 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
823 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
824 {
825 	struct mlx4_counter counter_stats;
826 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
827 	int err;
828 
829 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
830 		return -EINVAL;
831 
832 	memset(&counter_stats, 0, sizeof(counter_stats));
833 	err = mlx4_get_counter_stats(dev->dev,
834 				     dev->counters[port_num - 1].index,
835 				     &counter_stats, 0);
836 	if (err)
837 		err = IB_MAD_RESULT_FAILURE;
838 	else {
839 		memset(out_mad->data, 0, sizeof out_mad->data);
840 		switch (counter_stats.counter_mode & 0xf) {
841 		case 0:
842 			edit_counter(&counter_stats,
843 				     (void *)(out_mad->data + 40));
844 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
845 			break;
846 		default:
847 			err = IB_MAD_RESULT_FAILURE;
848 		}
849 	}
850 
851 	return err;
852 }
853 
854 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
855 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
856 			const struct ib_mad_hdr *in, size_t in_mad_size,
857 			struct ib_mad_hdr *out, size_t *out_mad_size,
858 			u16 *out_mad_pkey_index)
859 {
860 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
861 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
862 	struct ib_mad *out_mad = (struct ib_mad *)out;
863 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
864 
865 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
866 			 *out_mad_size != sizeof(*out_mad)))
867 		return IB_MAD_RESULT_FAILURE;
868 
869 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
870 	 * queries, should be called only by VFs and for that specific purpose
871 	 */
872 	if (link == IB_LINK_LAYER_INFINIBAND) {
873 		if (mlx4_is_slave(dev->dev) &&
874 		    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
875 		    in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS)
876 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
877 						in_grh, in_mad, out_mad);
878 
879 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
880 				      in_grh, in_mad, out_mad);
881 	}
882 
883 	if (link == IB_LINK_LAYER_ETHERNET)
884 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
885 					in_grh, in_mad, out_mad);
886 
887 	return -EINVAL;
888 }
889 
890 static void send_handler(struct ib_mad_agent *agent,
891 			 struct ib_mad_send_wc *mad_send_wc)
892 {
893 	if (mad_send_wc->send_buf->context[0])
894 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
895 	ib_free_send_mad(mad_send_wc->send_buf);
896 }
897 
898 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
899 {
900 	struct ib_mad_agent *agent;
901 	int p, q;
902 	int ret;
903 	enum rdma_link_layer ll;
904 
905 	for (p = 0; p < dev->num_ports; ++p) {
906 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
907 		for (q = 0; q <= 1; ++q) {
908 			if (ll == IB_LINK_LAYER_INFINIBAND) {
909 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
910 							      q ? IB_QPT_GSI : IB_QPT_SMI,
911 							      NULL, 0, send_handler,
912 							      NULL, NULL, 0);
913 				if (IS_ERR(agent)) {
914 					ret = PTR_ERR(agent);
915 					goto err;
916 				}
917 				dev->send_agent[p][q] = agent;
918 			} else
919 				dev->send_agent[p][q] = NULL;
920 		}
921 	}
922 
923 	return 0;
924 
925 err:
926 	for (p = 0; p < dev->num_ports; ++p)
927 		for (q = 0; q <= 1; ++q)
928 			if (dev->send_agent[p][q])
929 				ib_unregister_mad_agent(dev->send_agent[p][q]);
930 
931 	return ret;
932 }
933 
934 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
935 {
936 	struct ib_mad_agent *agent;
937 	int p, q;
938 
939 	for (p = 0; p < dev->num_ports; ++p) {
940 		for (q = 0; q <= 1; ++q) {
941 			agent = dev->send_agent[p][q];
942 			if (agent) {
943 				dev->send_agent[p][q] = NULL;
944 				ib_unregister_mad_agent(agent);
945 			}
946 		}
947 
948 		if (dev->sm_ah[p])
949 			ib_destroy_ah(dev->sm_ah[p]);
950 	}
951 }
952 
953 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
954 {
955 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
956 
957 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
958 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
959 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
960 }
961 
962 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
963 {
964 	/* re-configure the alias-guid and mcg's */
965 	if (mlx4_is_master(dev->dev)) {
966 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
967 
968 		if (!dev->sriov.is_going_down) {
969 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
970 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
971 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
972 		}
973 	}
974 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
975 }
976 
977 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
978 			      struct mlx4_eqe *eqe)
979 {
980 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
981 			    GET_MASK_FROM_EQE(eqe));
982 }
983 
984 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
985 				      u32 guid_tbl_blk_num, u32 change_bitmap)
986 {
987 	struct ib_smp *in_mad  = NULL;
988 	struct ib_smp *out_mad  = NULL;
989 	u16 i;
990 
991 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
992 		return;
993 
994 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
995 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
996 	if (!in_mad || !out_mad) {
997 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
998 		goto out;
999 	}
1000 
1001 	guid_tbl_blk_num  *= 4;
1002 
1003 	for (i = 0; i < 4; i++) {
1004 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1005 			continue;
1006 		memset(in_mad, 0, sizeof *in_mad);
1007 		memset(out_mad, 0, sizeof *out_mad);
1008 
1009 		in_mad->base_version  = 1;
1010 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1011 		in_mad->class_version = 1;
1012 		in_mad->method        = IB_MGMT_METHOD_GET;
1013 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1014 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1015 
1016 		if (mlx4_MAD_IFC(dev,
1017 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1018 				 port_num, NULL, NULL, in_mad, out_mad)) {
1019 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1020 			goto out;
1021 		}
1022 
1023 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1024 						    port_num,
1025 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1026 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1027 						     port_num,
1028 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1029 	}
1030 
1031 out:
1032 	kfree(in_mad);
1033 	kfree(out_mad);
1034 	return;
1035 }
1036 
1037 void handle_port_mgmt_change_event(struct work_struct *work)
1038 {
1039 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1040 	struct mlx4_ib_dev *dev = ew->ib_dev;
1041 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1042 	u8 port = eqe->event.port_mgmt_change.port;
1043 	u32 changed_attr;
1044 	u32 tbl_block;
1045 	u32 change_bitmap;
1046 
1047 	switch (eqe->subtype) {
1048 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1049 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1050 
1051 		/* Update the SM ah - This should be done before handling
1052 		   the other changed attributes so that MADs can be sent to the SM */
1053 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1054 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1055 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1056 			update_sm_ah(dev, port, lid, sl);
1057 		}
1058 
1059 		/* Check if it is a lid change event */
1060 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1061 			handle_lid_change_event(dev, port);
1062 
1063 		/* Generate GUID changed event */
1064 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1065 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1066 			/*if master, notify all slaves*/
1067 			if (mlx4_is_master(dev->dev))
1068 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1069 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1070 		}
1071 
1072 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1073 			handle_client_rereg_event(dev, port);
1074 		break;
1075 
1076 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1077 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1078 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1079 			propagate_pkey_ev(dev, port, eqe);
1080 		break;
1081 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1082 		/* paravirtualized master's guid is guid 0 -- does not change */
1083 		if (!mlx4_is_master(dev->dev))
1084 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1085 		/*if master, notify relevant slaves*/
1086 		else if (!dev->sriov.is_going_down) {
1087 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1088 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1089 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1090 		}
1091 		break;
1092 	default:
1093 		pr_warn("Unsupported subtype 0x%x for "
1094 			"Port Management Change event\n", eqe->subtype);
1095 	}
1096 
1097 	kfree(ew);
1098 }
1099 
1100 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1101 			    enum ib_event_type type)
1102 {
1103 	struct ib_event event;
1104 
1105 	event.device		= &dev->ib_dev;
1106 	event.element.port_num	= port_num;
1107 	event.event		= type;
1108 
1109 	ib_dispatch_event(&event);
1110 }
1111 
1112 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1113 {
1114 	unsigned long flags;
1115 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1116 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1117 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1118 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1119 		queue_work(ctx->wq, &ctx->work);
1120 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1121 }
1122 
1123 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1124 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1125 				  int index)
1126 {
1127 	struct ib_sge sg_list;
1128 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1129 	int size;
1130 
1131 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1132 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1133 
1134 	sg_list.addr = tun_qp->ring[index].map;
1135 	sg_list.length = size;
1136 	sg_list.lkey = ctx->pd->local_dma_lkey;
1137 
1138 	recv_wr.next = NULL;
1139 	recv_wr.sg_list = &sg_list;
1140 	recv_wr.num_sge = 1;
1141 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1142 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1143 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1144 				      size, DMA_FROM_DEVICE);
1145 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1146 }
1147 
1148 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1149 		int slave, struct ib_sa_mad *sa_mad)
1150 {
1151 	int ret = 0;
1152 
1153 	/* dispatch to different sa handlers */
1154 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1155 	case IB_SA_ATTR_MC_MEMBER_REC:
1156 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1157 		break;
1158 	default:
1159 		break;
1160 	}
1161 	return ret;
1162 }
1163 
1164 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1165 {
1166 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1167 
1168 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1169 }
1170 
1171 
1172 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1173 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1174 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1175 			 u8 *s_mac, struct ib_mad *mad)
1176 {
1177 	struct ib_sge list;
1178 	struct ib_send_wr wr, *bad_wr;
1179 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1180 	struct mlx4_ib_demux_pv_qp *sqp;
1181 	struct mlx4_mad_snd_buf *sqp_mad;
1182 	struct ib_ah *ah;
1183 	struct ib_qp *send_qp = NULL;
1184 	unsigned wire_tx_ix = 0;
1185 	int ret = 0;
1186 	u16 wire_pkey_ix;
1187 	int src_qpnum;
1188 	u8 sgid_index;
1189 
1190 
1191 	sqp_ctx = dev->sriov.sqps[port-1];
1192 
1193 	/* check if proxy qp created */
1194 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1195 		return -EAGAIN;
1196 
1197 	if (dest_qpt == IB_QPT_SMI) {
1198 		src_qpnum = 0;
1199 		sqp = &sqp_ctx->qp[0];
1200 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1201 	} else {
1202 		src_qpnum = 1;
1203 		sqp = &sqp_ctx->qp[1];
1204 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1205 	}
1206 
1207 	send_qp = sqp->qp;
1208 
1209 	/* create ah */
1210 	sgid_index = attr->grh.sgid_index;
1211 	attr->grh.sgid_index = 0;
1212 	ah = ib_create_ah(sqp_ctx->pd, attr);
1213 	if (IS_ERR(ah))
1214 		return -ENOMEM;
1215 	attr->grh.sgid_index = sgid_index;
1216 	to_mah(ah)->av.ib.gid_index = sgid_index;
1217 	/* get rid of force-loopback bit */
1218 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1219 	spin_lock(&sqp->tx_lock);
1220 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1221 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1222 		ret = -EAGAIN;
1223 	else
1224 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1225 	spin_unlock(&sqp->tx_lock);
1226 	if (ret)
1227 		goto out;
1228 
1229 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1230 	if (sqp->tx_ring[wire_tx_ix].ah)
1231 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1232 	sqp->tx_ring[wire_tx_ix].ah = ah;
1233 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1234 				   sqp->tx_ring[wire_tx_ix].buf.map,
1235 				   sizeof (struct mlx4_mad_snd_buf),
1236 				   DMA_TO_DEVICE);
1237 
1238 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1239 
1240 	ib_dma_sync_single_for_device(&dev->ib_dev,
1241 				      sqp->tx_ring[wire_tx_ix].buf.map,
1242 				      sizeof (struct mlx4_mad_snd_buf),
1243 				      DMA_TO_DEVICE);
1244 
1245 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1246 	list.length = sizeof (struct mlx4_mad_snd_buf);
1247 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1248 
1249 	wr.wr.ud.ah = ah;
1250 	wr.wr.ud.port_num = port;
1251 	wr.wr.ud.pkey_index = wire_pkey_ix;
1252 	wr.wr.ud.remote_qkey = qkey;
1253 	wr.wr.ud.remote_qpn = remote_qpn;
1254 	wr.next = NULL;
1255 	wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1256 	wr.sg_list = &list;
1257 	wr.num_sge = 1;
1258 	wr.opcode = IB_WR_SEND;
1259 	wr.send_flags = IB_SEND_SIGNALED;
1260 	if (s_mac)
1261 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1262 
1263 
1264 	ret = ib_post_send(send_qp, &wr, &bad_wr);
1265 out:
1266 	if (ret)
1267 		ib_destroy_ah(ah);
1268 	return ret;
1269 }
1270 
1271 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1272 {
1273 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1274 		return slave;
1275 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1276 }
1277 
1278 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1279 				    struct ib_ah_attr *ah_attr)
1280 {
1281 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1282 		ah_attr->grh.sgid_index = slave;
1283 	else
1284 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1285 }
1286 
1287 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1288 {
1289 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1290 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1291 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1292 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1293 	struct mlx4_ib_ah ah;
1294 	struct ib_ah_attr ah_attr;
1295 	u8 *slave_id;
1296 	int slave;
1297 	int port;
1298 
1299 	/* Get slave that sent this packet */
1300 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1301 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1302 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1303 	    wc->src_qp & 0x4) {
1304 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1305 		return;
1306 	}
1307 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1308 	if (slave != ctx->slave) {
1309 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1310 			     "belongs to another slave\n", wc->src_qp);
1311 		return;
1312 	}
1313 
1314 	/* Map transaction ID */
1315 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1316 				   sizeof (struct mlx4_tunnel_mad),
1317 				   DMA_FROM_DEVICE);
1318 	switch (tunnel->mad.mad_hdr.method) {
1319 	case IB_MGMT_METHOD_SET:
1320 	case IB_MGMT_METHOD_GET:
1321 	case IB_MGMT_METHOD_REPORT:
1322 	case IB_SA_METHOD_GET_TABLE:
1323 	case IB_SA_METHOD_DELETE:
1324 	case IB_SA_METHOD_GET_MULTI:
1325 	case IB_SA_METHOD_GET_TRACE_TBL:
1326 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1327 		if (*slave_id) {
1328 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1329 				     "class:%d slave:%d\n", *slave_id,
1330 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1331 			return;
1332 		} else
1333 			*slave_id = slave;
1334 	default:
1335 		/* nothing */;
1336 	}
1337 
1338 	/* Class-specific handling */
1339 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1340 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1341 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1342 		if (slave != mlx4_master_func_num(dev->dev) &&
1343 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1344 			return;
1345 		break;
1346 	case IB_MGMT_CLASS_SUBN_ADM:
1347 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1348 			      (struct ib_sa_mad *) &tunnel->mad))
1349 			return;
1350 		break;
1351 	case IB_MGMT_CLASS_CM:
1352 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1353 			      (struct ib_mad *) &tunnel->mad))
1354 			return;
1355 		break;
1356 	case IB_MGMT_CLASS_DEVICE_MGMT:
1357 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1358 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1359 			return;
1360 		break;
1361 	default:
1362 		/* Drop unsupported classes for slaves in tunnel mode */
1363 		if (slave != mlx4_master_func_num(dev->dev)) {
1364 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1365 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1366 			return;
1367 		}
1368 	}
1369 
1370 	/* We are using standard ib_core services to send the mad, so generate a
1371 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1372 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1373 	ah.ibah.device = ctx->ib_dev;
1374 
1375 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1376 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1377 	if (port < 0)
1378 		return;
1379 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1380 
1381 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1382 	if (ah_attr.ah_flags & IB_AH_GRH)
1383 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1384 
1385 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1386 	ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1387 	/* if slave have default vlan use it */
1388 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1389 				    &ah_attr.vlan_id, &ah_attr.sl);
1390 
1391 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1392 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1393 			     IB_QPT_SMI : IB_QPT_GSI,
1394 			     be16_to_cpu(tunnel->hdr.pkey_index),
1395 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1396 			     be32_to_cpu(tunnel->hdr.qkey),
1397 			     &ah_attr, wc->smac, &tunnel->mad);
1398 }
1399 
1400 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1401 				 enum ib_qp_type qp_type, int is_tun)
1402 {
1403 	int i;
1404 	struct mlx4_ib_demux_pv_qp *tun_qp;
1405 	int rx_buf_size, tx_buf_size;
1406 
1407 	if (qp_type > IB_QPT_GSI)
1408 		return -EINVAL;
1409 
1410 	tun_qp = &ctx->qp[qp_type];
1411 
1412 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1413 			       GFP_KERNEL);
1414 	if (!tun_qp->ring)
1415 		return -ENOMEM;
1416 
1417 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1418 				  sizeof (struct mlx4_ib_tun_tx_buf),
1419 				  GFP_KERNEL);
1420 	if (!tun_qp->tx_ring) {
1421 		kfree(tun_qp->ring);
1422 		tun_qp->ring = NULL;
1423 		return -ENOMEM;
1424 	}
1425 
1426 	if (is_tun) {
1427 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1428 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1429 	} else {
1430 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1431 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1432 	}
1433 
1434 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1435 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1436 		if (!tun_qp->ring[i].addr)
1437 			goto err;
1438 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1439 							tun_qp->ring[i].addr,
1440 							rx_buf_size,
1441 							DMA_FROM_DEVICE);
1442 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1443 			kfree(tun_qp->ring[i].addr);
1444 			goto err;
1445 		}
1446 	}
1447 
1448 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1449 		tun_qp->tx_ring[i].buf.addr =
1450 			kmalloc(tx_buf_size, GFP_KERNEL);
1451 		if (!tun_qp->tx_ring[i].buf.addr)
1452 			goto tx_err;
1453 		tun_qp->tx_ring[i].buf.map =
1454 			ib_dma_map_single(ctx->ib_dev,
1455 					  tun_qp->tx_ring[i].buf.addr,
1456 					  tx_buf_size,
1457 					  DMA_TO_DEVICE);
1458 		if (ib_dma_mapping_error(ctx->ib_dev,
1459 					 tun_qp->tx_ring[i].buf.map)) {
1460 			kfree(tun_qp->tx_ring[i].buf.addr);
1461 			goto tx_err;
1462 		}
1463 		tun_qp->tx_ring[i].ah = NULL;
1464 	}
1465 	spin_lock_init(&tun_qp->tx_lock);
1466 	tun_qp->tx_ix_head = 0;
1467 	tun_qp->tx_ix_tail = 0;
1468 	tun_qp->proxy_qpt = qp_type;
1469 
1470 	return 0;
1471 
1472 tx_err:
1473 	while (i > 0) {
1474 		--i;
1475 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1476 				    tx_buf_size, DMA_TO_DEVICE);
1477 		kfree(tun_qp->tx_ring[i].buf.addr);
1478 	}
1479 	kfree(tun_qp->tx_ring);
1480 	tun_qp->tx_ring = NULL;
1481 	i = MLX4_NUM_TUNNEL_BUFS;
1482 err:
1483 	while (i > 0) {
1484 		--i;
1485 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1486 				    rx_buf_size, DMA_FROM_DEVICE);
1487 		kfree(tun_qp->ring[i].addr);
1488 	}
1489 	kfree(tun_qp->ring);
1490 	tun_qp->ring = NULL;
1491 	return -ENOMEM;
1492 }
1493 
1494 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1495 				     enum ib_qp_type qp_type, int is_tun)
1496 {
1497 	int i;
1498 	struct mlx4_ib_demux_pv_qp *tun_qp;
1499 	int rx_buf_size, tx_buf_size;
1500 
1501 	if (qp_type > IB_QPT_GSI)
1502 		return;
1503 
1504 	tun_qp = &ctx->qp[qp_type];
1505 	if (is_tun) {
1506 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1507 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1508 	} else {
1509 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1510 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1511 	}
1512 
1513 
1514 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1515 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1516 				    rx_buf_size, DMA_FROM_DEVICE);
1517 		kfree(tun_qp->ring[i].addr);
1518 	}
1519 
1520 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1521 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1522 				    tx_buf_size, DMA_TO_DEVICE);
1523 		kfree(tun_qp->tx_ring[i].buf.addr);
1524 		if (tun_qp->tx_ring[i].ah)
1525 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1526 	}
1527 	kfree(tun_qp->tx_ring);
1528 	kfree(tun_qp->ring);
1529 }
1530 
1531 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1532 {
1533 	struct mlx4_ib_demux_pv_ctx *ctx;
1534 	struct mlx4_ib_demux_pv_qp *tun_qp;
1535 	struct ib_wc wc;
1536 	int ret;
1537 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1538 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1539 
1540 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1541 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1542 		if (wc.status == IB_WC_SUCCESS) {
1543 			switch (wc.opcode) {
1544 			case IB_WC_RECV:
1545 				mlx4_ib_multiplex_mad(ctx, &wc);
1546 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1547 							     wc.wr_id &
1548 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1549 				if (ret)
1550 					pr_err("Failed reposting tunnel "
1551 					       "buf:%lld\n", wc.wr_id);
1552 				break;
1553 			case IB_WC_SEND:
1554 				pr_debug("received tunnel send completion:"
1555 					 "wrid=0x%llx, status=0x%x\n",
1556 					 wc.wr_id, wc.status);
1557 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1558 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1559 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1560 					= NULL;
1561 				spin_lock(&tun_qp->tx_lock);
1562 				tun_qp->tx_ix_tail++;
1563 				spin_unlock(&tun_qp->tx_lock);
1564 
1565 				break;
1566 			default:
1567 				break;
1568 			}
1569 		} else  {
1570 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1571 				 " status = %d, wrid = 0x%llx\n",
1572 				 ctx->slave, wc.status, wc.wr_id);
1573 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1574 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1575 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1576 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1577 					= NULL;
1578 				spin_lock(&tun_qp->tx_lock);
1579 				tun_qp->tx_ix_tail++;
1580 				spin_unlock(&tun_qp->tx_lock);
1581 			}
1582 		}
1583 	}
1584 }
1585 
1586 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1587 {
1588 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1589 
1590 	/* It's worse than that! He's dead, Jim! */
1591 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1592 	       event->event, sqp->port);
1593 }
1594 
1595 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1596 			    enum ib_qp_type qp_type, int create_tun)
1597 {
1598 	int i, ret;
1599 	struct mlx4_ib_demux_pv_qp *tun_qp;
1600 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1601 	struct ib_qp_attr attr;
1602 	int qp_attr_mask_INIT;
1603 
1604 	if (qp_type > IB_QPT_GSI)
1605 		return -EINVAL;
1606 
1607 	tun_qp = &ctx->qp[qp_type];
1608 
1609 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1610 	qp_init_attr.init_attr.send_cq = ctx->cq;
1611 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1612 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1613 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1614 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1615 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1616 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1617 	if (create_tun) {
1618 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1619 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1620 		qp_init_attr.port = ctx->port;
1621 		qp_init_attr.slave = ctx->slave;
1622 		qp_init_attr.proxy_qp_type = qp_type;
1623 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1624 			   IB_QP_QKEY | IB_QP_PORT;
1625 	} else {
1626 		qp_init_attr.init_attr.qp_type = qp_type;
1627 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1628 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1629 	}
1630 	qp_init_attr.init_attr.port_num = ctx->port;
1631 	qp_init_attr.init_attr.qp_context = ctx;
1632 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1633 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1634 	if (IS_ERR(tun_qp->qp)) {
1635 		ret = PTR_ERR(tun_qp->qp);
1636 		tun_qp->qp = NULL;
1637 		pr_err("Couldn't create %s QP (%d)\n",
1638 		       create_tun ? "tunnel" : "special", ret);
1639 		return ret;
1640 	}
1641 
1642 	memset(&attr, 0, sizeof attr);
1643 	attr.qp_state = IB_QPS_INIT;
1644 	ret = 0;
1645 	if (create_tun)
1646 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1647 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1648 					      &attr.pkey_index);
1649 	if (ret || !create_tun)
1650 		attr.pkey_index =
1651 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1652 	attr.qkey = IB_QP1_QKEY;
1653 	attr.port_num = ctx->port;
1654 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1655 	if (ret) {
1656 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1657 		       create_tun ? "tunnel" : "special", ret);
1658 		goto err_qp;
1659 	}
1660 	attr.qp_state = IB_QPS_RTR;
1661 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1662 	if (ret) {
1663 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1664 		       create_tun ? "tunnel" : "special", ret);
1665 		goto err_qp;
1666 	}
1667 	attr.qp_state = IB_QPS_RTS;
1668 	attr.sq_psn = 0;
1669 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1670 	if (ret) {
1671 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1672 		       create_tun ? "tunnel" : "special", ret);
1673 		goto err_qp;
1674 	}
1675 
1676 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1677 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1678 		if (ret) {
1679 			pr_err(" mlx4_ib_post_pv_buf error"
1680 			       " (err = %d, i = %d)\n", ret, i);
1681 			goto err_qp;
1682 		}
1683 	}
1684 	return 0;
1685 
1686 err_qp:
1687 	ib_destroy_qp(tun_qp->qp);
1688 	tun_qp->qp = NULL;
1689 	return ret;
1690 }
1691 
1692 /*
1693  * IB MAD completion callback for real SQPs
1694  */
1695 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1696 {
1697 	struct mlx4_ib_demux_pv_ctx *ctx;
1698 	struct mlx4_ib_demux_pv_qp *sqp;
1699 	struct ib_wc wc;
1700 	struct ib_grh *grh;
1701 	struct ib_mad *mad;
1702 
1703 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1704 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1705 
1706 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1707 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1708 		if (wc.status == IB_WC_SUCCESS) {
1709 			switch (wc.opcode) {
1710 			case IB_WC_SEND:
1711 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1712 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1713 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1714 					= NULL;
1715 				spin_lock(&sqp->tx_lock);
1716 				sqp->tx_ix_tail++;
1717 				spin_unlock(&sqp->tx_lock);
1718 				break;
1719 			case IB_WC_RECV:
1720 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1721 						(sqp->ring[wc.wr_id &
1722 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1723 				grh = &(((struct mlx4_mad_rcv_buf *)
1724 						(sqp->ring[wc.wr_id &
1725 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1726 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1727 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1728 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1729 					pr_err("Failed reposting SQP "
1730 					       "buf:%lld\n", wc.wr_id);
1731 				break;
1732 			default:
1733 				BUG_ON(1);
1734 				break;
1735 			}
1736 		} else  {
1737 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1738 				 " status = %d, wrid = 0x%llx\n",
1739 				 ctx->slave, wc.status, wc.wr_id);
1740 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1741 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1742 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1743 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1744 					= NULL;
1745 				spin_lock(&sqp->tx_lock);
1746 				sqp->tx_ix_tail++;
1747 				spin_unlock(&sqp->tx_lock);
1748 			}
1749 		}
1750 	}
1751 }
1752 
1753 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1754 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1755 {
1756 	struct mlx4_ib_demux_pv_ctx *ctx;
1757 
1758 	*ret_ctx = NULL;
1759 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1760 	if (!ctx) {
1761 		pr_err("failed allocating pv resource context "
1762 		       "for port %d, slave %d\n", port, slave);
1763 		return -ENOMEM;
1764 	}
1765 
1766 	ctx->ib_dev = &dev->ib_dev;
1767 	ctx->port = port;
1768 	ctx->slave = slave;
1769 	*ret_ctx = ctx;
1770 	return 0;
1771 }
1772 
1773 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1774 {
1775 	if (dev->sriov.demux[port - 1].tun[slave]) {
1776 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1777 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1778 	}
1779 }
1780 
1781 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1782 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1783 {
1784 	int ret, cq_size;
1785 	struct ib_cq_init_attr cq_attr = {};
1786 
1787 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1788 		return -EEXIST;
1789 
1790 	ctx->state = DEMUX_PV_STATE_STARTING;
1791 	/* have QP0 only if link layer is IB */
1792 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1793 	    IB_LINK_LAYER_INFINIBAND)
1794 		ctx->has_smi = 1;
1795 
1796 	if (ctx->has_smi) {
1797 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1798 		if (ret) {
1799 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1800 			goto err_out;
1801 		}
1802 	}
1803 
1804 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1805 	if (ret) {
1806 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1807 		goto err_out_qp0;
1808 	}
1809 
1810 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1811 	if (ctx->has_smi)
1812 		cq_size *= 2;
1813 
1814 	cq_attr.cqe = cq_size;
1815 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1816 			       NULL, ctx, &cq_attr);
1817 	if (IS_ERR(ctx->cq)) {
1818 		ret = PTR_ERR(ctx->cq);
1819 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1820 		goto err_buf;
1821 	}
1822 
1823 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1824 	if (IS_ERR(ctx->pd)) {
1825 		ret = PTR_ERR(ctx->pd);
1826 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1827 		goto err_cq;
1828 	}
1829 
1830 	if (ctx->has_smi) {
1831 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1832 		if (ret) {
1833 			pr_err("Couldn't create %s QP0 (%d)\n",
1834 			       create_tun ? "tunnel for" : "",  ret);
1835 			goto err_pd;
1836 		}
1837 	}
1838 
1839 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1840 	if (ret) {
1841 		pr_err("Couldn't create %s QP1 (%d)\n",
1842 		       create_tun ? "tunnel for" : "",  ret);
1843 		goto err_qp0;
1844 	}
1845 
1846 	if (create_tun)
1847 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1848 	else
1849 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1850 
1851 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1852 
1853 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1854 	if (ret) {
1855 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1856 		goto err_wq;
1857 	}
1858 	ctx->state = DEMUX_PV_STATE_ACTIVE;
1859 	return 0;
1860 
1861 err_wq:
1862 	ctx->wq = NULL;
1863 	ib_destroy_qp(ctx->qp[1].qp);
1864 	ctx->qp[1].qp = NULL;
1865 
1866 
1867 err_qp0:
1868 	if (ctx->has_smi)
1869 		ib_destroy_qp(ctx->qp[0].qp);
1870 	ctx->qp[0].qp = NULL;
1871 
1872 err_pd:
1873 	ib_dealloc_pd(ctx->pd);
1874 	ctx->pd = NULL;
1875 
1876 err_cq:
1877 	ib_destroy_cq(ctx->cq);
1878 	ctx->cq = NULL;
1879 
1880 err_buf:
1881 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1882 
1883 err_out_qp0:
1884 	if (ctx->has_smi)
1885 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1886 err_out:
1887 	ctx->state = DEMUX_PV_STATE_DOWN;
1888 	return ret;
1889 }
1890 
1891 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1892 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1893 {
1894 	if (!ctx)
1895 		return;
1896 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
1897 		ctx->state = DEMUX_PV_STATE_DOWNING;
1898 		if (flush)
1899 			flush_workqueue(ctx->wq);
1900 		if (ctx->has_smi) {
1901 			ib_destroy_qp(ctx->qp[0].qp);
1902 			ctx->qp[0].qp = NULL;
1903 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1904 		}
1905 		ib_destroy_qp(ctx->qp[1].qp);
1906 		ctx->qp[1].qp = NULL;
1907 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1908 		ib_dealloc_pd(ctx->pd);
1909 		ctx->pd = NULL;
1910 		ib_destroy_cq(ctx->cq);
1911 		ctx->cq = NULL;
1912 		ctx->state = DEMUX_PV_STATE_DOWN;
1913 	}
1914 }
1915 
1916 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1917 				  int port, int do_init)
1918 {
1919 	int ret = 0;
1920 
1921 	if (!do_init) {
1922 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1923 		/* for master, destroy real sqp resources */
1924 		if (slave == mlx4_master_func_num(dev->dev))
1925 			destroy_pv_resources(dev, slave, port,
1926 					     dev->sriov.sqps[port - 1], 1);
1927 		/* destroy the tunnel qp resources */
1928 		destroy_pv_resources(dev, slave, port,
1929 				     dev->sriov.demux[port - 1].tun[slave], 1);
1930 		return 0;
1931 	}
1932 
1933 	/* create the tunnel qp resources */
1934 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1935 				  dev->sriov.demux[port - 1].tun[slave]);
1936 
1937 	/* for master, create the real sqp resources */
1938 	if (!ret && slave == mlx4_master_func_num(dev->dev))
1939 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1940 					  dev->sriov.sqps[port - 1]);
1941 	return ret;
1942 }
1943 
1944 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1945 {
1946 	struct mlx4_ib_demux_work *dmxw;
1947 
1948 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1949 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1950 			       dmxw->do_init);
1951 	kfree(dmxw);
1952 	return;
1953 }
1954 
1955 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1956 				       struct mlx4_ib_demux_ctx *ctx,
1957 				       int port)
1958 {
1959 	char name[12];
1960 	int ret = 0;
1961 	int i;
1962 
1963 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1964 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1965 	if (!ctx->tun)
1966 		return -ENOMEM;
1967 
1968 	ctx->dev = dev;
1969 	ctx->port = port;
1970 	ctx->ib_dev = &dev->ib_dev;
1971 
1972 	for (i = 0;
1973 	     i < min(dev->dev->caps.sqp_demux,
1974 	     (u16)(dev->dev->persist->num_vfs + 1));
1975 	     i++) {
1976 		struct mlx4_active_ports actv_ports =
1977 			mlx4_get_active_ports(dev->dev, i);
1978 
1979 		if (!test_bit(port - 1, actv_ports.ports))
1980 			continue;
1981 
1982 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1983 		if (ret) {
1984 			ret = -ENOMEM;
1985 			goto err_mcg;
1986 		}
1987 	}
1988 
1989 	ret = mlx4_ib_mcg_port_init(ctx);
1990 	if (ret) {
1991 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1992 		goto err_mcg;
1993 	}
1994 
1995 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
1996 	ctx->wq = create_singlethread_workqueue(name);
1997 	if (!ctx->wq) {
1998 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
1999 		ret = -ENOMEM;
2000 		goto err_wq;
2001 	}
2002 
2003 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2004 	ctx->ud_wq = create_singlethread_workqueue(name);
2005 	if (!ctx->ud_wq) {
2006 		pr_err("Failed to create up/down WQ for port %d\n", port);
2007 		ret = -ENOMEM;
2008 		goto err_udwq;
2009 	}
2010 
2011 	return 0;
2012 
2013 err_udwq:
2014 	destroy_workqueue(ctx->wq);
2015 	ctx->wq = NULL;
2016 
2017 err_wq:
2018 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2019 err_mcg:
2020 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2021 		free_pv_object(dev, i, port);
2022 	kfree(ctx->tun);
2023 	ctx->tun = NULL;
2024 	return ret;
2025 }
2026 
2027 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2028 {
2029 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2030 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2031 		flush_workqueue(sqp_ctx->wq);
2032 		if (sqp_ctx->has_smi) {
2033 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2034 			sqp_ctx->qp[0].qp = NULL;
2035 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2036 		}
2037 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2038 		sqp_ctx->qp[1].qp = NULL;
2039 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2040 		ib_dealloc_pd(sqp_ctx->pd);
2041 		sqp_ctx->pd = NULL;
2042 		ib_destroy_cq(sqp_ctx->cq);
2043 		sqp_ctx->cq = NULL;
2044 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2045 	}
2046 }
2047 
2048 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2049 {
2050 	int i;
2051 	if (ctx) {
2052 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2053 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2054 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2055 			if (!ctx->tun[i])
2056 				continue;
2057 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2058 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2059 		}
2060 		flush_workqueue(ctx->wq);
2061 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2062 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2063 			free_pv_object(dev, i, ctx->port);
2064 		}
2065 		kfree(ctx->tun);
2066 		destroy_workqueue(ctx->ud_wq);
2067 		destroy_workqueue(ctx->wq);
2068 	}
2069 }
2070 
2071 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2072 {
2073 	int i;
2074 
2075 	if (!mlx4_is_master(dev->dev))
2076 		return;
2077 	/* initialize or tear down tunnel QPs for the master */
2078 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2079 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2080 	return;
2081 }
2082 
2083 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2084 {
2085 	int i = 0;
2086 	int err;
2087 
2088 	if (!mlx4_is_mfunc(dev->dev))
2089 		return 0;
2090 
2091 	dev->sriov.is_going_down = 0;
2092 	spin_lock_init(&dev->sriov.going_down_lock);
2093 	mlx4_ib_cm_paravirt_init(dev);
2094 
2095 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2096 
2097 	if (mlx4_is_slave(dev->dev)) {
2098 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2099 		return 0;
2100 	}
2101 
2102 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2103 		if (i == mlx4_master_func_num(dev->dev))
2104 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2105 		else
2106 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2107 	}
2108 
2109 	err = mlx4_ib_init_alias_guid_service(dev);
2110 	if (err) {
2111 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2112 		goto paravirt_err;
2113 	}
2114 	err = mlx4_ib_device_register_sysfs(dev);
2115 	if (err) {
2116 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2117 		goto sysfs_err;
2118 	}
2119 
2120 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2121 		     dev->dev->caps.sqp_demux);
2122 	for (i = 0; i < dev->num_ports; i++) {
2123 		union ib_gid gid;
2124 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2125 		if (err)
2126 			goto demux_err;
2127 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2128 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2129 				      &dev->sriov.sqps[i]);
2130 		if (err)
2131 			goto demux_err;
2132 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2133 		if (err)
2134 			goto free_pv;
2135 	}
2136 	mlx4_ib_master_tunnels(dev, 1);
2137 	return 0;
2138 
2139 free_pv:
2140 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2141 demux_err:
2142 	while (--i >= 0) {
2143 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2144 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2145 	}
2146 	mlx4_ib_device_unregister_sysfs(dev);
2147 
2148 sysfs_err:
2149 	mlx4_ib_destroy_alias_guid_service(dev);
2150 
2151 paravirt_err:
2152 	mlx4_ib_cm_paravirt_clean(dev, -1);
2153 
2154 	return err;
2155 }
2156 
2157 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2158 {
2159 	int i;
2160 	unsigned long flags;
2161 
2162 	if (!mlx4_is_mfunc(dev->dev))
2163 		return;
2164 
2165 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2166 	dev->sriov.is_going_down = 1;
2167 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2168 	if (mlx4_is_master(dev->dev)) {
2169 		for (i = 0; i < dev->num_ports; i++) {
2170 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2171 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2172 			kfree(dev->sriov.sqps[i]);
2173 			dev->sriov.sqps[i] = NULL;
2174 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2175 		}
2176 
2177 		mlx4_ib_cm_paravirt_clean(dev, -1);
2178 		mlx4_ib_destroy_alias_guid_service(dev);
2179 		mlx4_ib_device_unregister_sysfs(dev);
2180 	}
2181 }
2182