xref: /linux/drivers/infiniband/hw/mlx4/mad.c (revision 96ba9dd65788a0bd2a7d1e57ec78b7642f0ccc25)
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, struct ib_wc *in_wc, struct ib_grh *in_grh,
107 		 void *in_mad, void *response_mad)
108 {
109 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
110 	void *inbox;
111 	int err;
112 	u32 in_modifier = port;
113 	u8 op_modifier = 0;
114 
115 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
116 	if (IS_ERR(inmailbox))
117 		return PTR_ERR(inmailbox);
118 	inbox = inmailbox->buf;
119 
120 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
121 	if (IS_ERR(outmailbox)) {
122 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
123 		return PTR_ERR(outmailbox);
124 	}
125 
126 	memcpy(inbox, in_mad, 256);
127 
128 	/*
129 	 * Key check traps can't be generated unless we have in_wc to
130 	 * tell us where to send the trap.
131 	 */
132 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
133 		op_modifier |= 0x1;
134 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
135 		op_modifier |= 0x2;
136 	if (mlx4_is_mfunc(dev->dev) &&
137 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
138 		op_modifier |= 0x8;
139 
140 	if (in_wc) {
141 		struct {
142 			__be32		my_qpn;
143 			u32		reserved1;
144 			__be32		rqpn;
145 			u8		sl;
146 			u8		g_path;
147 			u16		reserved2[2];
148 			__be16		pkey;
149 			u32		reserved3[11];
150 			u8		grh[40];
151 		} *ext_info;
152 
153 		memset(inbox + 256, 0, 256);
154 		ext_info = inbox + 256;
155 
156 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
157 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
158 		ext_info->sl     = in_wc->sl << 4;
159 		ext_info->g_path = in_wc->dlid_path_bits |
160 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
161 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
162 
163 		if (in_grh)
164 			memcpy(ext_info->grh, in_grh, 40);
165 
166 		op_modifier |= 0x4;
167 
168 		in_modifier |= in_wc->slid << 16;
169 	}
170 
171 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
172 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
173 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
174 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
175 
176 	if (!err)
177 		memcpy(response_mad, outmailbox->buf, 256);
178 
179 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
180 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
181 
182 	return err;
183 }
184 
185 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
186 {
187 	struct ib_ah *new_ah;
188 	struct ib_ah_attr ah_attr;
189 	unsigned long flags;
190 
191 	if (!dev->send_agent[port_num - 1][0])
192 		return;
193 
194 	memset(&ah_attr, 0, sizeof ah_attr);
195 	ah_attr.dlid     = lid;
196 	ah_attr.sl       = sl;
197 	ah_attr.port_num = port_num;
198 
199 	new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
200 			      &ah_attr);
201 	if (IS_ERR(new_ah))
202 		return;
203 
204 	spin_lock_irqsave(&dev->sm_lock, flags);
205 	if (dev->sm_ah[port_num - 1])
206 		ib_destroy_ah(dev->sm_ah[port_num - 1]);
207 	dev->sm_ah[port_num - 1] = new_ah;
208 	spin_unlock_irqrestore(&dev->sm_lock, flags);
209 }
210 
211 /*
212  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
213  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
214  */
215 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
216 		      u16 prev_lid)
217 {
218 	struct ib_port_info *pinfo;
219 	u16 lid;
220 	__be16 *base;
221 	u32 bn, pkey_change_bitmap;
222 	int i;
223 
224 
225 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
226 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
227 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
228 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
229 		switch (mad->mad_hdr.attr_id) {
230 		case IB_SMP_ATTR_PORT_INFO:
231 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
232 			lid = be16_to_cpu(pinfo->lid);
233 
234 			update_sm_ah(dev, port_num,
235 				     be16_to_cpu(pinfo->sm_lid),
236 				     pinfo->neighbormtu_mastersmsl & 0xf);
237 
238 			if (pinfo->clientrereg_resv_subnetto & 0x80)
239 				handle_client_rereg_event(dev, port_num);
240 
241 			if (prev_lid != lid)
242 				handle_lid_change_event(dev, port_num);
243 			break;
244 
245 		case IB_SMP_ATTR_PKEY_TABLE:
246 			if (!mlx4_is_mfunc(dev->dev)) {
247 				mlx4_ib_dispatch_event(dev, port_num,
248 						       IB_EVENT_PKEY_CHANGE);
249 				break;
250 			}
251 
252 			/* at this point, we are running in the master.
253 			 * Slaves do not receive SMPs.
254 			 */
255 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
256 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
257 			pkey_change_bitmap = 0;
258 			for (i = 0; i < 32; i++) {
259 				pr_debug("PKEY[%d] = x%x\n",
260 					 i + bn*32, be16_to_cpu(base[i]));
261 				if (be16_to_cpu(base[i]) !=
262 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
263 					pkey_change_bitmap |= (1 << i);
264 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
265 						be16_to_cpu(base[i]);
266 				}
267 			}
268 			pr_debug("PKEY Change event: port=%d, "
269 				 "block=0x%x, change_bitmap=0x%x\n",
270 				 port_num, bn, pkey_change_bitmap);
271 
272 			if (pkey_change_bitmap) {
273 				mlx4_ib_dispatch_event(dev, port_num,
274 						       IB_EVENT_PKEY_CHANGE);
275 				if (!dev->sriov.is_going_down)
276 					__propagate_pkey_ev(dev, port_num, bn,
277 							    pkey_change_bitmap);
278 			}
279 			break;
280 
281 		case IB_SMP_ATTR_GUID_INFO:
282 			/* paravirtualized master's guid is guid 0 -- does not change */
283 			if (!mlx4_is_master(dev->dev))
284 				mlx4_ib_dispatch_event(dev, port_num,
285 						       IB_EVENT_GID_CHANGE);
286 			/*if master, notify relevant slaves*/
287 			if (mlx4_is_master(dev->dev) &&
288 			    !dev->sriov.is_going_down) {
289 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
290 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
291 								    (u8 *)(&((struct ib_smp *)mad)->data));
292 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
293 								     (u8 *)(&((struct ib_smp *)mad)->data));
294 			}
295 			break;
296 
297 		default:
298 			break;
299 		}
300 }
301 
302 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
303 				int block, u32 change_bitmap)
304 {
305 	int i, ix, slave, err;
306 	int have_event = 0;
307 
308 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
309 		if (slave == mlx4_master_func_num(dev->dev))
310 			continue;
311 		if (!mlx4_is_slave_active(dev->dev, slave))
312 			continue;
313 
314 		have_event = 0;
315 		for (i = 0; i < 32; i++) {
316 			if (!(change_bitmap & (1 << i)))
317 				continue;
318 			for (ix = 0;
319 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
320 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
321 				    [ix] == i + 32 * block) {
322 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
323 					pr_debug("propagate_pkey_ev: slave %d,"
324 						 " port %d, ix %d (%d)\n",
325 						 slave, port_num, ix, err);
326 					have_event = 1;
327 					break;
328 				}
329 			}
330 			if (have_event)
331 				break;
332 		}
333 	}
334 }
335 
336 static void node_desc_override(struct ib_device *dev,
337 			       struct ib_mad *mad)
338 {
339 	unsigned long flags;
340 
341 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
342 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
343 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
344 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
345 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
346 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
347 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
348 	}
349 }
350 
351 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
352 {
353 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
354 	struct ib_mad_send_buf *send_buf;
355 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
356 	int ret;
357 	unsigned long flags;
358 
359 	if (agent) {
360 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
361 					      IB_MGMT_MAD_DATA, GFP_ATOMIC);
362 		if (IS_ERR(send_buf))
363 			return;
364 		/*
365 		 * We rely here on the fact that MLX QPs don't use the
366 		 * address handle after the send is posted (this is
367 		 * wrong following the IB spec strictly, but we know
368 		 * it's OK for our devices).
369 		 */
370 		spin_lock_irqsave(&dev->sm_lock, flags);
371 		memcpy(send_buf->mad, mad, sizeof *mad);
372 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
373 			ret = ib_post_send_mad(send_buf, NULL);
374 		else
375 			ret = -EINVAL;
376 		spin_unlock_irqrestore(&dev->sm_lock, flags);
377 
378 		if (ret)
379 			ib_free_send_mad(send_buf);
380 	}
381 }
382 
383 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
384 							     struct ib_sa_mad *sa_mad)
385 {
386 	int ret = 0;
387 
388 	/* dispatch to different sa handlers */
389 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
390 	case IB_SA_ATTR_MC_MEMBER_REC:
391 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
392 		break;
393 	default:
394 		break;
395 	}
396 	return ret;
397 }
398 
399 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
400 {
401 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
402 	int i;
403 
404 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
405 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
406 			return i;
407 	}
408 	return -1;
409 }
410 
411 
412 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
413 				   u8 port, u16 pkey, u16 *ix)
414 {
415 	int i, ret;
416 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
417 	u16 slot_pkey;
418 
419 	if (slave == mlx4_master_func_num(dev->dev))
420 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
421 
422 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
423 
424 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
425 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
426 			continue;
427 
428 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
429 
430 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
431 		if (ret)
432 			continue;
433 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
434 			if (slot_pkey & 0x8000) {
435 				*ix = (u16) pkey_ix;
436 				return 0;
437 			} else {
438 				/* take first partial pkey index found */
439 				if (partial_ix == 0xFF)
440 					partial_ix = pkey_ix;
441 			}
442 		}
443 	}
444 
445 	if (partial_ix < 0xFF) {
446 		*ix = (u16) partial_ix;
447 		return 0;
448 	}
449 
450 	return -EINVAL;
451 }
452 
453 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
454 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
455 			  struct ib_grh *grh, struct ib_mad *mad)
456 {
457 	struct ib_sge list;
458 	struct ib_send_wr wr, *bad_wr;
459 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
460 	struct mlx4_ib_demux_pv_qp *tun_qp;
461 	struct mlx4_rcv_tunnel_mad *tun_mad;
462 	struct ib_ah_attr attr;
463 	struct ib_ah *ah;
464 	struct ib_qp *src_qp = NULL;
465 	unsigned tun_tx_ix = 0;
466 	int dqpn;
467 	int ret = 0;
468 	u16 tun_pkey_ix;
469 	u16 cached_pkey;
470 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
471 
472 	if (dest_qpt > IB_QPT_GSI)
473 		return -EINVAL;
474 
475 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
476 
477 	/* check if proxy qp created */
478 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
479 		return -EAGAIN;
480 
481 	/* QP0 forwarding only for Dom0 */
482 	if (!dest_qpt && (mlx4_master_func_num(dev->dev) != slave))
483 		return -EINVAL;
484 
485 	if (!dest_qpt)
486 		tun_qp = &tun_ctx->qp[0];
487 	else
488 		tun_qp = &tun_ctx->qp[1];
489 
490 	/* compute P_Key index to put in tunnel header for slave */
491 	if (dest_qpt) {
492 		u16 pkey_ix;
493 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
494 		if (ret)
495 			return -EINVAL;
496 
497 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
498 		if (ret)
499 			return -EINVAL;
500 		tun_pkey_ix = pkey_ix;
501 	} else
502 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
503 
504 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
505 
506 	/* get tunnel tx data buf for slave */
507 	src_qp = tun_qp->qp;
508 
509 	/* create ah. Just need an empty one with the port num for the post send.
510 	 * The driver will set the force loopback bit in post_send */
511 	memset(&attr, 0, sizeof attr);
512 	attr.port_num = port;
513 	if (is_eth) {
514 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
515 		attr.ah_flags = IB_AH_GRH;
516 	}
517 	ah = ib_create_ah(tun_ctx->pd, &attr);
518 	if (IS_ERR(ah))
519 		return -ENOMEM;
520 
521 	/* allocate tunnel tx buf after pass failure returns */
522 	spin_lock(&tun_qp->tx_lock);
523 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
524 	    (MLX4_NUM_TUNNEL_BUFS - 1))
525 		ret = -EAGAIN;
526 	else
527 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
528 	spin_unlock(&tun_qp->tx_lock);
529 	if (ret)
530 		goto out;
531 
532 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
533 	if (tun_qp->tx_ring[tun_tx_ix].ah)
534 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
535 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
536 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
537 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
538 				   sizeof (struct mlx4_rcv_tunnel_mad),
539 				   DMA_TO_DEVICE);
540 
541 	/* copy over to tunnel buffer */
542 	if (grh)
543 		memcpy(&tun_mad->grh, grh, sizeof *grh);
544 	memcpy(&tun_mad->mad, mad, sizeof *mad);
545 
546 	/* adjust tunnel data */
547 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
548 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
549 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
550 
551 	if (is_eth) {
552 		u16 vlan = 0;
553 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
554 						NULL)) {
555 			/* VST mode */
556 			if (vlan != wc->vlan_id)
557 				/* Packet vlan is not the VST-assigned vlan.
558 				 * Drop the packet.
559 				 */
560 				goto out;
561 			 else
562 				/* Remove the vlan tag before forwarding
563 				 * the packet to the VF.
564 				 */
565 				vlan = 0xffff;
566 		} else {
567 			vlan = wc->vlan_id;
568 		}
569 
570 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
571 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
572 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
573 	} else {
574 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
575 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
576 	}
577 
578 	ib_dma_sync_single_for_device(&dev->ib_dev,
579 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
580 				      sizeof (struct mlx4_rcv_tunnel_mad),
581 				      DMA_TO_DEVICE);
582 
583 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
584 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
585 	list.lkey = tun_ctx->mr->lkey;
586 
587 	wr.wr.ud.ah = ah;
588 	wr.wr.ud.port_num = port;
589 	wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
590 	wr.wr.ud.remote_qpn = dqpn;
591 	wr.next = NULL;
592 	wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
593 	wr.sg_list = &list;
594 	wr.num_sge = 1;
595 	wr.opcode = IB_WR_SEND;
596 	wr.send_flags = IB_SEND_SIGNALED;
597 
598 	ret = ib_post_send(src_qp, &wr, &bad_wr);
599 out:
600 	if (ret)
601 		ib_destroy_ah(ah);
602 	return ret;
603 }
604 
605 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
606 			struct ib_wc *wc, struct ib_grh *grh,
607 			struct ib_mad *mad)
608 {
609 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
610 	int err;
611 	int slave;
612 	u8 *slave_id;
613 	int is_eth = 0;
614 
615 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
616 		is_eth = 0;
617 	else
618 		is_eth = 1;
619 
620 	if (is_eth) {
621 		if (!(wc->wc_flags & IB_WC_GRH)) {
622 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
623 			return -EINVAL;
624 		}
625 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
626 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
627 			return -EINVAL;
628 		}
629 		if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
630 			mlx4_ib_warn(ibdev, "failed matching grh\n");
631 			return -ENOENT;
632 		}
633 		if (slave >= dev->dev->caps.sqp_demux) {
634 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
635 				     slave, dev->dev->caps.sqp_demux);
636 			return -ENOENT;
637 		}
638 
639 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
640 			return 0;
641 
642 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
643 		if (err)
644 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
645 				 slave, err);
646 		return 0;
647 	}
648 
649 	/* Initially assume that this mad is for us */
650 	slave = mlx4_master_func_num(dev->dev);
651 
652 	/* See if the slave id is encoded in a response mad */
653 	if (mad->mad_hdr.method & 0x80) {
654 		slave_id = (u8 *) &mad->mad_hdr.tid;
655 		slave = *slave_id;
656 		if (slave != 255) /*255 indicates the dom0*/
657 			*slave_id = 0; /* remap tid */
658 	}
659 
660 	/* If a grh is present, we demux according to it */
661 	if (wc->wc_flags & IB_WC_GRH) {
662 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
663 		if (slave < 0) {
664 			mlx4_ib_warn(ibdev, "failed matching grh\n");
665 			return -ENOENT;
666 		}
667 	}
668 	/* Class-specific handling */
669 	switch (mad->mad_hdr.mgmt_class) {
670 	case IB_MGMT_CLASS_SUBN_ADM:
671 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
672 					     (struct ib_sa_mad *) mad))
673 			return 0;
674 		break;
675 	case IB_MGMT_CLASS_CM:
676 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
677 			return 0;
678 		break;
679 	case IB_MGMT_CLASS_DEVICE_MGMT:
680 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
681 			return 0;
682 		break;
683 	default:
684 		/* Drop unsupported classes for slaves in tunnel mode */
685 		if (slave != mlx4_master_func_num(dev->dev)) {
686 			pr_debug("dropping unsupported ingress mad from class:%d "
687 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
688 			return 0;
689 		}
690 	}
691 	/*make sure that no slave==255 was not handled yet.*/
692 	if (slave >= dev->dev->caps.sqp_demux) {
693 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
694 			     slave, dev->dev->caps.sqp_demux);
695 		return -ENOENT;
696 	}
697 
698 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
699 	if (err)
700 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
701 			 slave, err);
702 	return 0;
703 }
704 
705 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
706 			struct ib_wc *in_wc, struct ib_grh *in_grh,
707 			struct ib_mad *in_mad, struct ib_mad *out_mad)
708 {
709 	u16 slid, prev_lid = 0;
710 	int err;
711 	struct ib_port_attr pattr;
712 
713 	if (in_wc && in_wc->qp->qp_num) {
714 		pr_debug("received MAD: slid:%d sqpn:%d "
715 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
716 			in_wc->slid, in_wc->src_qp,
717 			in_wc->dlid_path_bits,
718 			in_wc->qp->qp_num,
719 			in_wc->wc_flags,
720 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
721 			be16_to_cpu(in_mad->mad_hdr.attr_id));
722 		if (in_wc->wc_flags & IB_WC_GRH) {
723 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
724 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
725 				 be64_to_cpu(in_grh->sgid.global.interface_id));
726 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
727 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
728 				 be64_to_cpu(in_grh->dgid.global.interface_id));
729 		}
730 	}
731 
732 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
733 
734 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
735 		forward_trap(to_mdev(ibdev), port_num, in_mad);
736 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
737 	}
738 
739 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
740 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
741 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
742 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
743 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
744 			return IB_MAD_RESULT_SUCCESS;
745 
746 		/*
747 		 * Don't process SMInfo queries -- the SMA can't handle them.
748 		 */
749 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
750 			return IB_MAD_RESULT_SUCCESS;
751 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
752 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
753 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
754 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
755 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
756 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
757 			return IB_MAD_RESULT_SUCCESS;
758 	} else
759 		return IB_MAD_RESULT_SUCCESS;
760 
761 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
762 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
763 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
764 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
765 	    !ib_query_port(ibdev, port_num, &pattr))
766 		prev_lid = pattr.lid;
767 
768 	err = mlx4_MAD_IFC(to_mdev(ibdev),
769 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
770 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
771 			   MLX4_MAD_IFC_NET_VIEW,
772 			   port_num, in_wc, in_grh, in_mad, out_mad);
773 	if (err)
774 		return IB_MAD_RESULT_FAILURE;
775 
776 	if (!out_mad->mad_hdr.status) {
777 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
778 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
779 		/* slaves get node desc from FW */
780 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
781 			node_desc_override(ibdev, out_mad);
782 	}
783 
784 	/* set return bit in status of directed route responses */
785 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
786 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
787 
788 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
789 		/* no response for trap repress */
790 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
791 
792 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
793 }
794 
795 static void edit_counter(struct mlx4_counter *cnt,
796 					struct ib_pma_portcounters *pma_cnt)
797 {
798 	pma_cnt->port_xmit_data = cpu_to_be32((be64_to_cpu(cnt->tx_bytes)>>2));
799 	pma_cnt->port_rcv_data  = cpu_to_be32((be64_to_cpu(cnt->rx_bytes)>>2));
800 	pma_cnt->port_xmit_packets = cpu_to_be32(be64_to_cpu(cnt->tx_frames));
801 	pma_cnt->port_rcv_packets  = cpu_to_be32(be64_to_cpu(cnt->rx_frames));
802 }
803 
804 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
805 			struct ib_wc *in_wc, struct ib_grh *in_grh,
806 			struct ib_mad *in_mad, struct ib_mad *out_mad)
807 {
808 	struct mlx4_cmd_mailbox *mailbox;
809 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
810 	int err;
811 	u32 inmod = dev->counters[port_num - 1] & 0xffff;
812 	u8 mode;
813 
814 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
815 		return -EINVAL;
816 
817 	mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
818 	if (IS_ERR(mailbox))
819 		return IB_MAD_RESULT_FAILURE;
820 
821 	err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
822 			   MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
823 			   MLX4_CMD_WRAPPED);
824 	if (err)
825 		err = IB_MAD_RESULT_FAILURE;
826 	else {
827 		memset(out_mad->data, 0, sizeof out_mad->data);
828 		mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
829 		switch (mode & 0xf) {
830 		case 0:
831 			edit_counter(mailbox->buf,
832 						(void *)(out_mad->data + 40));
833 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
834 			break;
835 		default:
836 			err = IB_MAD_RESULT_FAILURE;
837 		}
838 	}
839 
840 	mlx4_free_cmd_mailbox(dev->dev, mailbox);
841 
842 	return err;
843 }
844 
845 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
846 			struct ib_wc *in_wc, struct ib_grh *in_grh,
847 			struct ib_mad *in_mad, struct ib_mad *out_mad)
848 {
849 	switch (rdma_port_get_link_layer(ibdev, port_num)) {
850 	case IB_LINK_LAYER_INFINIBAND:
851 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
852 				      in_grh, in_mad, out_mad);
853 	case IB_LINK_LAYER_ETHERNET:
854 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
855 					  in_grh, in_mad, out_mad);
856 	default:
857 		return -EINVAL;
858 	}
859 }
860 
861 static void send_handler(struct ib_mad_agent *agent,
862 			 struct ib_mad_send_wc *mad_send_wc)
863 {
864 	if (mad_send_wc->send_buf->context[0])
865 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
866 	ib_free_send_mad(mad_send_wc->send_buf);
867 }
868 
869 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
870 {
871 	struct ib_mad_agent *agent;
872 	int p, q;
873 	int ret;
874 	enum rdma_link_layer ll;
875 
876 	for (p = 0; p < dev->num_ports; ++p) {
877 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
878 		for (q = 0; q <= 1; ++q) {
879 			if (ll == IB_LINK_LAYER_INFINIBAND) {
880 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
881 							      q ? IB_QPT_GSI : IB_QPT_SMI,
882 							      NULL, 0, send_handler,
883 							      NULL, NULL);
884 				if (IS_ERR(agent)) {
885 					ret = PTR_ERR(agent);
886 					goto err;
887 				}
888 				dev->send_agent[p][q] = agent;
889 			} else
890 				dev->send_agent[p][q] = NULL;
891 		}
892 	}
893 
894 	return 0;
895 
896 err:
897 	for (p = 0; p < dev->num_ports; ++p)
898 		for (q = 0; q <= 1; ++q)
899 			if (dev->send_agent[p][q])
900 				ib_unregister_mad_agent(dev->send_agent[p][q]);
901 
902 	return ret;
903 }
904 
905 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
906 {
907 	struct ib_mad_agent *agent;
908 	int p, q;
909 
910 	for (p = 0; p < dev->num_ports; ++p) {
911 		for (q = 0; q <= 1; ++q) {
912 			agent = dev->send_agent[p][q];
913 			if (agent) {
914 				dev->send_agent[p][q] = NULL;
915 				ib_unregister_mad_agent(agent);
916 			}
917 		}
918 
919 		if (dev->sm_ah[p])
920 			ib_destroy_ah(dev->sm_ah[p]);
921 	}
922 }
923 
924 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
925 {
926 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
927 
928 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
929 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
930 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
931 }
932 
933 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
934 {
935 	/* re-configure the alias-guid and mcg's */
936 	if (mlx4_is_master(dev->dev)) {
937 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
938 
939 		if (!dev->sriov.is_going_down) {
940 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
941 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
942 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
943 		}
944 	}
945 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
946 }
947 
948 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
949 			      struct mlx4_eqe *eqe)
950 {
951 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
952 			    GET_MASK_FROM_EQE(eqe));
953 }
954 
955 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
956 				      u32 guid_tbl_blk_num, u32 change_bitmap)
957 {
958 	struct ib_smp *in_mad  = NULL;
959 	struct ib_smp *out_mad  = NULL;
960 	u16 i;
961 
962 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
963 		return;
964 
965 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
966 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
967 	if (!in_mad || !out_mad) {
968 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
969 		goto out;
970 	}
971 
972 	guid_tbl_blk_num  *= 4;
973 
974 	for (i = 0; i < 4; i++) {
975 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
976 			continue;
977 		memset(in_mad, 0, sizeof *in_mad);
978 		memset(out_mad, 0, sizeof *out_mad);
979 
980 		in_mad->base_version  = 1;
981 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
982 		in_mad->class_version = 1;
983 		in_mad->method        = IB_MGMT_METHOD_GET;
984 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
985 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
986 
987 		if (mlx4_MAD_IFC(dev,
988 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
989 				 port_num, NULL, NULL, in_mad, out_mad)) {
990 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
991 			goto out;
992 		}
993 
994 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
995 						    port_num,
996 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
997 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
998 						     port_num,
999 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1000 	}
1001 
1002 out:
1003 	kfree(in_mad);
1004 	kfree(out_mad);
1005 	return;
1006 }
1007 
1008 void handle_port_mgmt_change_event(struct work_struct *work)
1009 {
1010 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1011 	struct mlx4_ib_dev *dev = ew->ib_dev;
1012 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1013 	u8 port = eqe->event.port_mgmt_change.port;
1014 	u32 changed_attr;
1015 	u32 tbl_block;
1016 	u32 change_bitmap;
1017 
1018 	switch (eqe->subtype) {
1019 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1020 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1021 
1022 		/* Update the SM ah - This should be done before handling
1023 		   the other changed attributes so that MADs can be sent to the SM */
1024 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1025 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1026 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1027 			update_sm_ah(dev, port, lid, sl);
1028 		}
1029 
1030 		/* Check if it is a lid change event */
1031 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1032 			handle_lid_change_event(dev, port);
1033 
1034 		/* Generate GUID changed event */
1035 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1036 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1037 			/*if master, notify all slaves*/
1038 			if (mlx4_is_master(dev->dev))
1039 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1040 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1041 		}
1042 
1043 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1044 			handle_client_rereg_event(dev, port);
1045 		break;
1046 
1047 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1048 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1049 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1050 			propagate_pkey_ev(dev, port, eqe);
1051 		break;
1052 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1053 		/* paravirtualized master's guid is guid 0 -- does not change */
1054 		if (!mlx4_is_master(dev->dev))
1055 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1056 		/*if master, notify relevant slaves*/
1057 		else if (!dev->sriov.is_going_down) {
1058 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1059 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1060 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1061 		}
1062 		break;
1063 	default:
1064 		pr_warn("Unsupported subtype 0x%x for "
1065 			"Port Management Change event\n", eqe->subtype);
1066 	}
1067 
1068 	kfree(ew);
1069 }
1070 
1071 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1072 			    enum ib_event_type type)
1073 {
1074 	struct ib_event event;
1075 
1076 	event.device		= &dev->ib_dev;
1077 	event.element.port_num	= port_num;
1078 	event.event		= type;
1079 
1080 	ib_dispatch_event(&event);
1081 }
1082 
1083 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1084 {
1085 	unsigned long flags;
1086 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1087 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1088 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1089 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1090 		queue_work(ctx->wq, &ctx->work);
1091 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1092 }
1093 
1094 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1095 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1096 				  int index)
1097 {
1098 	struct ib_sge sg_list;
1099 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1100 	int size;
1101 
1102 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1103 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1104 
1105 	sg_list.addr = tun_qp->ring[index].map;
1106 	sg_list.length = size;
1107 	sg_list.lkey = ctx->mr->lkey;
1108 
1109 	recv_wr.next = NULL;
1110 	recv_wr.sg_list = &sg_list;
1111 	recv_wr.num_sge = 1;
1112 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1113 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1114 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1115 				      size, DMA_FROM_DEVICE);
1116 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1117 }
1118 
1119 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1120 		int slave, struct ib_sa_mad *sa_mad)
1121 {
1122 	int ret = 0;
1123 
1124 	/* dispatch to different sa handlers */
1125 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1126 	case IB_SA_ATTR_MC_MEMBER_REC:
1127 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1128 		break;
1129 	default:
1130 		break;
1131 	}
1132 	return ret;
1133 }
1134 
1135 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1136 {
1137 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1138 
1139 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1140 }
1141 
1142 
1143 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1144 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1145 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1146 			 u8 *s_mac, struct ib_mad *mad)
1147 {
1148 	struct ib_sge list;
1149 	struct ib_send_wr wr, *bad_wr;
1150 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1151 	struct mlx4_ib_demux_pv_qp *sqp;
1152 	struct mlx4_mad_snd_buf *sqp_mad;
1153 	struct ib_ah *ah;
1154 	struct ib_qp *send_qp = NULL;
1155 	unsigned wire_tx_ix = 0;
1156 	int ret = 0;
1157 	u16 wire_pkey_ix;
1158 	int src_qpnum;
1159 	u8 sgid_index;
1160 
1161 
1162 	sqp_ctx = dev->sriov.sqps[port-1];
1163 
1164 	/* check if proxy qp created */
1165 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1166 		return -EAGAIN;
1167 
1168 	/* QP0 forwarding only for Dom0 */
1169 	if (dest_qpt == IB_QPT_SMI && (mlx4_master_func_num(dev->dev) != slave))
1170 		return -EINVAL;
1171 
1172 	if (dest_qpt == IB_QPT_SMI) {
1173 		src_qpnum = 0;
1174 		sqp = &sqp_ctx->qp[0];
1175 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1176 	} else {
1177 		src_qpnum = 1;
1178 		sqp = &sqp_ctx->qp[1];
1179 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1180 	}
1181 
1182 	send_qp = sqp->qp;
1183 
1184 	/* create ah */
1185 	sgid_index = attr->grh.sgid_index;
1186 	attr->grh.sgid_index = 0;
1187 	ah = ib_create_ah(sqp_ctx->pd, attr);
1188 	if (IS_ERR(ah))
1189 		return -ENOMEM;
1190 	attr->grh.sgid_index = sgid_index;
1191 	to_mah(ah)->av.ib.gid_index = sgid_index;
1192 	/* get rid of force-loopback bit */
1193 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1194 	spin_lock(&sqp->tx_lock);
1195 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1196 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1197 		ret = -EAGAIN;
1198 	else
1199 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1200 	spin_unlock(&sqp->tx_lock);
1201 	if (ret)
1202 		goto out;
1203 
1204 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1205 	if (sqp->tx_ring[wire_tx_ix].ah)
1206 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1207 	sqp->tx_ring[wire_tx_ix].ah = ah;
1208 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1209 				   sqp->tx_ring[wire_tx_ix].buf.map,
1210 				   sizeof (struct mlx4_mad_snd_buf),
1211 				   DMA_TO_DEVICE);
1212 
1213 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1214 
1215 	ib_dma_sync_single_for_device(&dev->ib_dev,
1216 				      sqp->tx_ring[wire_tx_ix].buf.map,
1217 				      sizeof (struct mlx4_mad_snd_buf),
1218 				      DMA_TO_DEVICE);
1219 
1220 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1221 	list.length = sizeof (struct mlx4_mad_snd_buf);
1222 	list.lkey = sqp_ctx->mr->lkey;
1223 
1224 	wr.wr.ud.ah = ah;
1225 	wr.wr.ud.port_num = port;
1226 	wr.wr.ud.pkey_index = wire_pkey_ix;
1227 	wr.wr.ud.remote_qkey = qkey;
1228 	wr.wr.ud.remote_qpn = remote_qpn;
1229 	wr.next = NULL;
1230 	wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1231 	wr.sg_list = &list;
1232 	wr.num_sge = 1;
1233 	wr.opcode = IB_WR_SEND;
1234 	wr.send_flags = IB_SEND_SIGNALED;
1235 	if (s_mac)
1236 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1237 
1238 
1239 	ret = ib_post_send(send_qp, &wr, &bad_wr);
1240 out:
1241 	if (ret)
1242 		ib_destroy_ah(ah);
1243 	return ret;
1244 }
1245 
1246 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1247 {
1248 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1249 		return slave;
1250 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1251 }
1252 
1253 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1254 				    struct ib_ah_attr *ah_attr)
1255 {
1256 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1257 		ah_attr->grh.sgid_index = slave;
1258 	else
1259 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1260 }
1261 
1262 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1263 {
1264 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1265 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1266 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1267 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1268 	struct mlx4_ib_ah ah;
1269 	struct ib_ah_attr ah_attr;
1270 	u8 *slave_id;
1271 	int slave;
1272 	int port;
1273 
1274 	/* Get slave that sent this packet */
1275 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1276 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1277 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1278 	    wc->src_qp & 0x4) {
1279 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1280 		return;
1281 	}
1282 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1283 	if (slave != ctx->slave) {
1284 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1285 			     "belongs to another slave\n", wc->src_qp);
1286 		return;
1287 	}
1288 	if (slave != mlx4_master_func_num(dev->dev) && !(wc->src_qp & 0x2)) {
1289 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1290 			     "non-master trying to send QP0 packets\n", wc->src_qp);
1291 		return;
1292 	}
1293 
1294 	/* Map transaction ID */
1295 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1296 				   sizeof (struct mlx4_tunnel_mad),
1297 				   DMA_FROM_DEVICE);
1298 	switch (tunnel->mad.mad_hdr.method) {
1299 	case IB_MGMT_METHOD_SET:
1300 	case IB_MGMT_METHOD_GET:
1301 	case IB_MGMT_METHOD_REPORT:
1302 	case IB_SA_METHOD_GET_TABLE:
1303 	case IB_SA_METHOD_DELETE:
1304 	case IB_SA_METHOD_GET_MULTI:
1305 	case IB_SA_METHOD_GET_TRACE_TBL:
1306 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1307 		if (*slave_id) {
1308 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1309 				     "class:%d slave:%d\n", *slave_id,
1310 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1311 			return;
1312 		} else
1313 			*slave_id = slave;
1314 	default:
1315 		/* nothing */;
1316 	}
1317 
1318 	/* Class-specific handling */
1319 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1320 	case IB_MGMT_CLASS_SUBN_ADM:
1321 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1322 			      (struct ib_sa_mad *) &tunnel->mad))
1323 			return;
1324 		break;
1325 	case IB_MGMT_CLASS_CM:
1326 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1327 			      (struct ib_mad *) &tunnel->mad))
1328 			return;
1329 		break;
1330 	case IB_MGMT_CLASS_DEVICE_MGMT:
1331 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1332 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1333 			return;
1334 		break;
1335 	default:
1336 		/* Drop unsupported classes for slaves in tunnel mode */
1337 		if (slave != mlx4_master_func_num(dev->dev)) {
1338 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1339 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1340 			return;
1341 		}
1342 	}
1343 
1344 	/* We are using standard ib_core services to send the mad, so generate a
1345 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1346 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1347 	ah.ibah.device = ctx->ib_dev;
1348 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1349 	if (ah_attr.ah_flags & IB_AH_GRH)
1350 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1351 
1352 	port = mlx4_slave_convert_port(dev->dev, slave, ah_attr.port_num);
1353 	if (port < 0)
1354 		return;
1355 	ah_attr.port_num = port;
1356 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1357 	ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1358 	/* if slave have default vlan use it */
1359 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1360 				    &ah_attr.vlan_id, &ah_attr.sl);
1361 
1362 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1363 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1364 			     IB_QPT_SMI : IB_QPT_GSI,
1365 			     be16_to_cpu(tunnel->hdr.pkey_index),
1366 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1367 			     be32_to_cpu(tunnel->hdr.qkey),
1368 			     &ah_attr, wc->smac, &tunnel->mad);
1369 }
1370 
1371 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1372 				 enum ib_qp_type qp_type, int is_tun)
1373 {
1374 	int i;
1375 	struct mlx4_ib_demux_pv_qp *tun_qp;
1376 	int rx_buf_size, tx_buf_size;
1377 
1378 	if (qp_type > IB_QPT_GSI)
1379 		return -EINVAL;
1380 
1381 	tun_qp = &ctx->qp[qp_type];
1382 
1383 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1384 			       GFP_KERNEL);
1385 	if (!tun_qp->ring)
1386 		return -ENOMEM;
1387 
1388 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1389 				  sizeof (struct mlx4_ib_tun_tx_buf),
1390 				  GFP_KERNEL);
1391 	if (!tun_qp->tx_ring) {
1392 		kfree(tun_qp->ring);
1393 		tun_qp->ring = NULL;
1394 		return -ENOMEM;
1395 	}
1396 
1397 	if (is_tun) {
1398 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1399 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1400 	} else {
1401 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1402 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1403 	}
1404 
1405 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1406 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1407 		if (!tun_qp->ring[i].addr)
1408 			goto err;
1409 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1410 							tun_qp->ring[i].addr,
1411 							rx_buf_size,
1412 							DMA_FROM_DEVICE);
1413 	}
1414 
1415 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1416 		tun_qp->tx_ring[i].buf.addr =
1417 			kmalloc(tx_buf_size, GFP_KERNEL);
1418 		if (!tun_qp->tx_ring[i].buf.addr)
1419 			goto tx_err;
1420 		tun_qp->tx_ring[i].buf.map =
1421 			ib_dma_map_single(ctx->ib_dev,
1422 					  tun_qp->tx_ring[i].buf.addr,
1423 					  tx_buf_size,
1424 					  DMA_TO_DEVICE);
1425 		tun_qp->tx_ring[i].ah = NULL;
1426 	}
1427 	spin_lock_init(&tun_qp->tx_lock);
1428 	tun_qp->tx_ix_head = 0;
1429 	tun_qp->tx_ix_tail = 0;
1430 	tun_qp->proxy_qpt = qp_type;
1431 
1432 	return 0;
1433 
1434 tx_err:
1435 	while (i > 0) {
1436 		--i;
1437 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1438 				    tx_buf_size, DMA_TO_DEVICE);
1439 		kfree(tun_qp->tx_ring[i].buf.addr);
1440 	}
1441 	kfree(tun_qp->tx_ring);
1442 	tun_qp->tx_ring = NULL;
1443 	i = MLX4_NUM_TUNNEL_BUFS;
1444 err:
1445 	while (i > 0) {
1446 		--i;
1447 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1448 				    rx_buf_size, DMA_FROM_DEVICE);
1449 		kfree(tun_qp->ring[i].addr);
1450 	}
1451 	kfree(tun_qp->ring);
1452 	tun_qp->ring = NULL;
1453 	return -ENOMEM;
1454 }
1455 
1456 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1457 				     enum ib_qp_type qp_type, int is_tun)
1458 {
1459 	int i;
1460 	struct mlx4_ib_demux_pv_qp *tun_qp;
1461 	int rx_buf_size, tx_buf_size;
1462 
1463 	if (qp_type > IB_QPT_GSI)
1464 		return;
1465 
1466 	tun_qp = &ctx->qp[qp_type];
1467 	if (is_tun) {
1468 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1469 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1470 	} else {
1471 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1472 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1473 	}
1474 
1475 
1476 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1477 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1478 				    rx_buf_size, DMA_FROM_DEVICE);
1479 		kfree(tun_qp->ring[i].addr);
1480 	}
1481 
1482 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1483 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1484 				    tx_buf_size, DMA_TO_DEVICE);
1485 		kfree(tun_qp->tx_ring[i].buf.addr);
1486 		if (tun_qp->tx_ring[i].ah)
1487 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1488 	}
1489 	kfree(tun_qp->tx_ring);
1490 	kfree(tun_qp->ring);
1491 }
1492 
1493 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1494 {
1495 	struct mlx4_ib_demux_pv_ctx *ctx;
1496 	struct mlx4_ib_demux_pv_qp *tun_qp;
1497 	struct ib_wc wc;
1498 	int ret;
1499 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1500 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1501 
1502 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1503 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1504 		if (wc.status == IB_WC_SUCCESS) {
1505 			switch (wc.opcode) {
1506 			case IB_WC_RECV:
1507 				mlx4_ib_multiplex_mad(ctx, &wc);
1508 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1509 							     wc.wr_id &
1510 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1511 				if (ret)
1512 					pr_err("Failed reposting tunnel "
1513 					       "buf:%lld\n", wc.wr_id);
1514 				break;
1515 			case IB_WC_SEND:
1516 				pr_debug("received tunnel send completion:"
1517 					 "wrid=0x%llx, status=0x%x\n",
1518 					 wc.wr_id, wc.status);
1519 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1520 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1521 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1522 					= NULL;
1523 				spin_lock(&tun_qp->tx_lock);
1524 				tun_qp->tx_ix_tail++;
1525 				spin_unlock(&tun_qp->tx_lock);
1526 
1527 				break;
1528 			default:
1529 				break;
1530 			}
1531 		} else  {
1532 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1533 				 " status = %d, wrid = 0x%llx\n",
1534 				 ctx->slave, wc.status, wc.wr_id);
1535 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1536 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1537 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1538 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1539 					= NULL;
1540 				spin_lock(&tun_qp->tx_lock);
1541 				tun_qp->tx_ix_tail++;
1542 				spin_unlock(&tun_qp->tx_lock);
1543 			}
1544 		}
1545 	}
1546 }
1547 
1548 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1549 {
1550 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1551 
1552 	/* It's worse than that! He's dead, Jim! */
1553 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1554 	       event->event, sqp->port);
1555 }
1556 
1557 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1558 			    enum ib_qp_type qp_type, int create_tun)
1559 {
1560 	int i, ret;
1561 	struct mlx4_ib_demux_pv_qp *tun_qp;
1562 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1563 	struct ib_qp_attr attr;
1564 	int qp_attr_mask_INIT;
1565 
1566 	if (qp_type > IB_QPT_GSI)
1567 		return -EINVAL;
1568 
1569 	tun_qp = &ctx->qp[qp_type];
1570 
1571 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1572 	qp_init_attr.init_attr.send_cq = ctx->cq;
1573 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1574 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1575 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1576 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1577 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1578 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1579 	if (create_tun) {
1580 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1581 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1582 		qp_init_attr.port = ctx->port;
1583 		qp_init_attr.slave = ctx->slave;
1584 		qp_init_attr.proxy_qp_type = qp_type;
1585 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1586 			   IB_QP_QKEY | IB_QP_PORT;
1587 	} else {
1588 		qp_init_attr.init_attr.qp_type = qp_type;
1589 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1590 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1591 	}
1592 	qp_init_attr.init_attr.port_num = ctx->port;
1593 	qp_init_attr.init_attr.qp_context = ctx;
1594 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1595 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1596 	if (IS_ERR(tun_qp->qp)) {
1597 		ret = PTR_ERR(tun_qp->qp);
1598 		tun_qp->qp = NULL;
1599 		pr_err("Couldn't create %s QP (%d)\n",
1600 		       create_tun ? "tunnel" : "special", ret);
1601 		return ret;
1602 	}
1603 
1604 	memset(&attr, 0, sizeof attr);
1605 	attr.qp_state = IB_QPS_INIT;
1606 	ret = 0;
1607 	if (create_tun)
1608 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1609 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1610 					      &attr.pkey_index);
1611 	if (ret || !create_tun)
1612 		attr.pkey_index =
1613 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1614 	attr.qkey = IB_QP1_QKEY;
1615 	attr.port_num = ctx->port;
1616 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1617 	if (ret) {
1618 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1619 		       create_tun ? "tunnel" : "special", ret);
1620 		goto err_qp;
1621 	}
1622 	attr.qp_state = IB_QPS_RTR;
1623 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1624 	if (ret) {
1625 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1626 		       create_tun ? "tunnel" : "special", ret);
1627 		goto err_qp;
1628 	}
1629 	attr.qp_state = IB_QPS_RTS;
1630 	attr.sq_psn = 0;
1631 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1632 	if (ret) {
1633 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1634 		       create_tun ? "tunnel" : "special", ret);
1635 		goto err_qp;
1636 	}
1637 
1638 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1639 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1640 		if (ret) {
1641 			pr_err(" mlx4_ib_post_pv_buf error"
1642 			       " (err = %d, i = %d)\n", ret, i);
1643 			goto err_qp;
1644 		}
1645 	}
1646 	return 0;
1647 
1648 err_qp:
1649 	ib_destroy_qp(tun_qp->qp);
1650 	tun_qp->qp = NULL;
1651 	return ret;
1652 }
1653 
1654 /*
1655  * IB MAD completion callback for real SQPs
1656  */
1657 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1658 {
1659 	struct mlx4_ib_demux_pv_ctx *ctx;
1660 	struct mlx4_ib_demux_pv_qp *sqp;
1661 	struct ib_wc wc;
1662 	struct ib_grh *grh;
1663 	struct ib_mad *mad;
1664 
1665 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1666 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1667 
1668 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1669 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1670 		if (wc.status == IB_WC_SUCCESS) {
1671 			switch (wc.opcode) {
1672 			case IB_WC_SEND:
1673 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1674 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1675 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1676 					= NULL;
1677 				spin_lock(&sqp->tx_lock);
1678 				sqp->tx_ix_tail++;
1679 				spin_unlock(&sqp->tx_lock);
1680 				break;
1681 			case IB_WC_RECV:
1682 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1683 						(sqp->ring[wc.wr_id &
1684 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1685 				grh = &(((struct mlx4_mad_rcv_buf *)
1686 						(sqp->ring[wc.wr_id &
1687 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1688 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1689 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1690 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1691 					pr_err("Failed reposting SQP "
1692 					       "buf:%lld\n", wc.wr_id);
1693 				break;
1694 			default:
1695 				BUG_ON(1);
1696 				break;
1697 			}
1698 		} else  {
1699 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1700 				 " status = %d, wrid = 0x%llx\n",
1701 				 ctx->slave, wc.status, wc.wr_id);
1702 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1703 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1704 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1705 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1706 					= NULL;
1707 				spin_lock(&sqp->tx_lock);
1708 				sqp->tx_ix_tail++;
1709 				spin_unlock(&sqp->tx_lock);
1710 			}
1711 		}
1712 	}
1713 }
1714 
1715 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1716 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1717 {
1718 	struct mlx4_ib_demux_pv_ctx *ctx;
1719 
1720 	*ret_ctx = NULL;
1721 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1722 	if (!ctx) {
1723 		pr_err("failed allocating pv resource context "
1724 		       "for port %d, slave %d\n", port, slave);
1725 		return -ENOMEM;
1726 	}
1727 
1728 	ctx->ib_dev = &dev->ib_dev;
1729 	ctx->port = port;
1730 	ctx->slave = slave;
1731 	*ret_ctx = ctx;
1732 	return 0;
1733 }
1734 
1735 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1736 {
1737 	if (dev->sriov.demux[port - 1].tun[slave]) {
1738 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1739 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1740 	}
1741 }
1742 
1743 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1744 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1745 {
1746 	int ret, cq_size;
1747 
1748 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1749 		return -EEXIST;
1750 
1751 	ctx->state = DEMUX_PV_STATE_STARTING;
1752 	/* have QP0 only on port owner, and only if link layer is IB */
1753 	if (ctx->slave == mlx4_master_func_num(to_mdev(ctx->ib_dev)->dev) &&
1754 	    rdma_port_get_link_layer(ibdev, ctx->port) == IB_LINK_LAYER_INFINIBAND)
1755 		ctx->has_smi = 1;
1756 
1757 	if (ctx->has_smi) {
1758 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1759 		if (ret) {
1760 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1761 			goto err_out;
1762 		}
1763 	}
1764 
1765 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1766 	if (ret) {
1767 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1768 		goto err_out_qp0;
1769 	}
1770 
1771 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1772 	if (ctx->has_smi)
1773 		cq_size *= 2;
1774 
1775 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1776 			       NULL, ctx, cq_size, 0);
1777 	if (IS_ERR(ctx->cq)) {
1778 		ret = PTR_ERR(ctx->cq);
1779 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1780 		goto err_buf;
1781 	}
1782 
1783 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1784 	if (IS_ERR(ctx->pd)) {
1785 		ret = PTR_ERR(ctx->pd);
1786 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1787 		goto err_cq;
1788 	}
1789 
1790 	ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1791 	if (IS_ERR(ctx->mr)) {
1792 		ret = PTR_ERR(ctx->mr);
1793 		pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1794 		goto err_pd;
1795 	}
1796 
1797 	if (ctx->has_smi) {
1798 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1799 		if (ret) {
1800 			pr_err("Couldn't create %s QP0 (%d)\n",
1801 			       create_tun ? "tunnel for" : "",  ret);
1802 			goto err_mr;
1803 		}
1804 	}
1805 
1806 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1807 	if (ret) {
1808 		pr_err("Couldn't create %s QP1 (%d)\n",
1809 		       create_tun ? "tunnel for" : "",  ret);
1810 		goto err_qp0;
1811 	}
1812 
1813 	if (create_tun)
1814 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1815 	else
1816 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1817 
1818 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1819 
1820 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1821 	if (ret) {
1822 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1823 		goto err_wq;
1824 	}
1825 	ctx->state = DEMUX_PV_STATE_ACTIVE;
1826 	return 0;
1827 
1828 err_wq:
1829 	ctx->wq = NULL;
1830 	ib_destroy_qp(ctx->qp[1].qp);
1831 	ctx->qp[1].qp = NULL;
1832 
1833 
1834 err_qp0:
1835 	if (ctx->has_smi)
1836 		ib_destroy_qp(ctx->qp[0].qp);
1837 	ctx->qp[0].qp = NULL;
1838 
1839 err_mr:
1840 	ib_dereg_mr(ctx->mr);
1841 	ctx->mr = NULL;
1842 
1843 err_pd:
1844 	ib_dealloc_pd(ctx->pd);
1845 	ctx->pd = NULL;
1846 
1847 err_cq:
1848 	ib_destroy_cq(ctx->cq);
1849 	ctx->cq = NULL;
1850 
1851 err_buf:
1852 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1853 
1854 err_out_qp0:
1855 	if (ctx->has_smi)
1856 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1857 err_out:
1858 	ctx->state = DEMUX_PV_STATE_DOWN;
1859 	return ret;
1860 }
1861 
1862 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1863 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1864 {
1865 	if (!ctx)
1866 		return;
1867 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
1868 		ctx->state = DEMUX_PV_STATE_DOWNING;
1869 		if (flush)
1870 			flush_workqueue(ctx->wq);
1871 		if (ctx->has_smi) {
1872 			ib_destroy_qp(ctx->qp[0].qp);
1873 			ctx->qp[0].qp = NULL;
1874 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1875 		}
1876 		ib_destroy_qp(ctx->qp[1].qp);
1877 		ctx->qp[1].qp = NULL;
1878 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1879 		ib_dereg_mr(ctx->mr);
1880 		ctx->mr = NULL;
1881 		ib_dealloc_pd(ctx->pd);
1882 		ctx->pd = NULL;
1883 		ib_destroy_cq(ctx->cq);
1884 		ctx->cq = NULL;
1885 		ctx->state = DEMUX_PV_STATE_DOWN;
1886 	}
1887 }
1888 
1889 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1890 				  int port, int do_init)
1891 {
1892 	int ret = 0;
1893 
1894 	if (!do_init) {
1895 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1896 		/* for master, destroy real sqp resources */
1897 		if (slave == mlx4_master_func_num(dev->dev))
1898 			destroy_pv_resources(dev, slave, port,
1899 					     dev->sriov.sqps[port - 1], 1);
1900 		/* destroy the tunnel qp resources */
1901 		destroy_pv_resources(dev, slave, port,
1902 				     dev->sriov.demux[port - 1].tun[slave], 1);
1903 		return 0;
1904 	}
1905 
1906 	/* create the tunnel qp resources */
1907 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1908 				  dev->sriov.demux[port - 1].tun[slave]);
1909 
1910 	/* for master, create the real sqp resources */
1911 	if (!ret && slave == mlx4_master_func_num(dev->dev))
1912 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1913 					  dev->sriov.sqps[port - 1]);
1914 	return ret;
1915 }
1916 
1917 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1918 {
1919 	struct mlx4_ib_demux_work *dmxw;
1920 
1921 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1922 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1923 			       dmxw->do_init);
1924 	kfree(dmxw);
1925 	return;
1926 }
1927 
1928 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1929 				       struct mlx4_ib_demux_ctx *ctx,
1930 				       int port)
1931 {
1932 	char name[12];
1933 	int ret = 0;
1934 	int i;
1935 
1936 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1937 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1938 	if (!ctx->tun)
1939 		return -ENOMEM;
1940 
1941 	ctx->dev = dev;
1942 	ctx->port = port;
1943 	ctx->ib_dev = &dev->ib_dev;
1944 
1945 	for (i = 0;
1946 	     i < min(dev->dev->caps.sqp_demux, (u16)(dev->dev->num_vfs + 1));
1947 	     i++) {
1948 		struct mlx4_active_ports actv_ports =
1949 			mlx4_get_active_ports(dev->dev, i);
1950 
1951 		if (!test_bit(port - 1, actv_ports.ports))
1952 			continue;
1953 
1954 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1955 		if (ret) {
1956 			ret = -ENOMEM;
1957 			goto err_mcg;
1958 		}
1959 	}
1960 
1961 	ret = mlx4_ib_mcg_port_init(ctx);
1962 	if (ret) {
1963 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1964 		goto err_mcg;
1965 	}
1966 
1967 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
1968 	ctx->wq = create_singlethread_workqueue(name);
1969 	if (!ctx->wq) {
1970 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
1971 		ret = -ENOMEM;
1972 		goto err_wq;
1973 	}
1974 
1975 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
1976 	ctx->ud_wq = create_singlethread_workqueue(name);
1977 	if (!ctx->ud_wq) {
1978 		pr_err("Failed to create up/down WQ for port %d\n", port);
1979 		ret = -ENOMEM;
1980 		goto err_udwq;
1981 	}
1982 
1983 	return 0;
1984 
1985 err_udwq:
1986 	destroy_workqueue(ctx->wq);
1987 	ctx->wq = NULL;
1988 
1989 err_wq:
1990 	mlx4_ib_mcg_port_cleanup(ctx, 1);
1991 err_mcg:
1992 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
1993 		free_pv_object(dev, i, port);
1994 	kfree(ctx->tun);
1995 	ctx->tun = NULL;
1996 	return ret;
1997 }
1998 
1999 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2000 {
2001 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2002 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2003 		flush_workqueue(sqp_ctx->wq);
2004 		if (sqp_ctx->has_smi) {
2005 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2006 			sqp_ctx->qp[0].qp = NULL;
2007 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2008 		}
2009 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2010 		sqp_ctx->qp[1].qp = NULL;
2011 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2012 		ib_dereg_mr(sqp_ctx->mr);
2013 		sqp_ctx->mr = NULL;
2014 		ib_dealloc_pd(sqp_ctx->pd);
2015 		sqp_ctx->pd = NULL;
2016 		ib_destroy_cq(sqp_ctx->cq);
2017 		sqp_ctx->cq = NULL;
2018 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2019 	}
2020 }
2021 
2022 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2023 {
2024 	int i;
2025 	if (ctx) {
2026 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2027 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2028 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2029 			if (!ctx->tun[i])
2030 				continue;
2031 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2032 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2033 		}
2034 		flush_workqueue(ctx->wq);
2035 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2036 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2037 			free_pv_object(dev, i, ctx->port);
2038 		}
2039 		kfree(ctx->tun);
2040 		destroy_workqueue(ctx->ud_wq);
2041 		destroy_workqueue(ctx->wq);
2042 	}
2043 }
2044 
2045 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2046 {
2047 	int i;
2048 
2049 	if (!mlx4_is_master(dev->dev))
2050 		return;
2051 	/* initialize or tear down tunnel QPs for the master */
2052 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2053 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2054 	return;
2055 }
2056 
2057 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2058 {
2059 	int i = 0;
2060 	int err;
2061 
2062 	if (!mlx4_is_mfunc(dev->dev))
2063 		return 0;
2064 
2065 	dev->sriov.is_going_down = 0;
2066 	spin_lock_init(&dev->sriov.going_down_lock);
2067 	mlx4_ib_cm_paravirt_init(dev);
2068 
2069 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2070 
2071 	if (mlx4_is_slave(dev->dev)) {
2072 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2073 		return 0;
2074 	}
2075 
2076 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2077 		if (i == mlx4_master_func_num(dev->dev))
2078 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2079 		else
2080 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2081 	}
2082 
2083 	err = mlx4_ib_init_alias_guid_service(dev);
2084 	if (err) {
2085 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2086 		goto paravirt_err;
2087 	}
2088 	err = mlx4_ib_device_register_sysfs(dev);
2089 	if (err) {
2090 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2091 		goto sysfs_err;
2092 	}
2093 
2094 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2095 		     dev->dev->caps.sqp_demux);
2096 	for (i = 0; i < dev->num_ports; i++) {
2097 		union ib_gid gid;
2098 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2099 		if (err)
2100 			goto demux_err;
2101 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2102 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2103 				      &dev->sriov.sqps[i]);
2104 		if (err)
2105 			goto demux_err;
2106 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2107 		if (err)
2108 			goto free_pv;
2109 	}
2110 	mlx4_ib_master_tunnels(dev, 1);
2111 	return 0;
2112 
2113 free_pv:
2114 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2115 demux_err:
2116 	while (--i >= 0) {
2117 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2118 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2119 	}
2120 	mlx4_ib_device_unregister_sysfs(dev);
2121 
2122 sysfs_err:
2123 	mlx4_ib_destroy_alias_guid_service(dev);
2124 
2125 paravirt_err:
2126 	mlx4_ib_cm_paravirt_clean(dev, -1);
2127 
2128 	return err;
2129 }
2130 
2131 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2132 {
2133 	int i;
2134 	unsigned long flags;
2135 
2136 	if (!mlx4_is_mfunc(dev->dev))
2137 		return;
2138 
2139 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2140 	dev->sriov.is_going_down = 1;
2141 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2142 	if (mlx4_is_master(dev->dev)) {
2143 		for (i = 0; i < dev->num_ports; i++) {
2144 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2145 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2146 			kfree(dev->sriov.sqps[i]);
2147 			dev->sriov.sqps[i] = NULL;
2148 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2149 		}
2150 
2151 		mlx4_ib_cm_paravirt_clean(dev, -1);
2152 		mlx4_ib_destroy_alias_guid_service(dev);
2153 		mlx4_ib_device_unregister_sysfs(dev);
2154 	}
2155 }
2156