xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_mad.c (revision 8657387683946d0c03e09fe77029edfe309eeb20)
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 <dev/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 | random());
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 			if (vlan != wc->vlan_id)
556 				/* VST and default vlan is not the packet vlan drop the
557 				 * packet*/
558 				goto out;
559 			 else
560 				/* VST , remove hide the vlan from the VF */
561 				vlan = 0;
562 		} else {
563 			vlan = wc->vlan_id;
564 		}
565 
566 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
567 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
568 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
569 	} else {
570 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
571 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
572 	}
573 
574 	ib_dma_sync_single_for_device(&dev->ib_dev,
575 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
576 				      sizeof (struct mlx4_rcv_tunnel_mad),
577 				      DMA_TO_DEVICE);
578 
579 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
580 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
581 	list.lkey = tun_ctx->mr->lkey;
582 
583 	wr.wr.ud.ah = ah;
584 	wr.wr.ud.port_num = port;
585 	wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
586 	wr.wr.ud.remote_qpn = dqpn;
587 	wr.next = NULL;
588 	wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
589 	wr.sg_list = &list;
590 	wr.num_sge = 1;
591 	wr.opcode = IB_WR_SEND;
592 	wr.send_flags = IB_SEND_SIGNALED;
593 
594 	ret = ib_post_send(src_qp, &wr, &bad_wr);
595 out:
596 	if (ret)
597 		ib_destroy_ah(ah);
598 	return ret;
599 }
600 
601 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
602 			struct ib_wc *wc, struct ib_grh *grh,
603 			struct ib_mad *mad)
604 {
605 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
606 	int err;
607 	int slave;
608 	u8 *slave_id;
609 	int is_eth = 0;
610 
611 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
612 		is_eth = 0;
613 	else
614 		is_eth = 1;
615 
616 	if (is_eth) {
617 		if (!(wc->wc_flags & IB_WC_GRH)) {
618 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
619 			return -EINVAL;
620 		}
621 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
622 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
623 			return -EINVAL;
624 		}
625 		if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
626 			mlx4_ib_warn(ibdev, "failed matching grh\n");
627 			return -ENOENT;
628 		}
629 		if (slave >= dev->dev->caps.sqp_demux) {
630 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
631 				     slave, dev->dev->caps.sqp_demux);
632 			return -ENOENT;
633 		}
634 
635 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad, is_eth))
636 			return 0;
637 
638 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
639 		if (err)
640 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
641 				 slave, err);
642 		return 0;
643 	}
644 
645 	/* Initially assume that this mad is for us */
646 	slave = mlx4_master_func_num(dev->dev);
647 
648 	/* See if the slave id is encoded in a response mad */
649 	if (mad->mad_hdr.method & 0x80) {
650 		slave_id = (u8 *) &mad->mad_hdr.tid;
651 		slave = *slave_id;
652 		if (slave != 255) /*255 indicates the dom0*/
653 			*slave_id = 0; /* remap tid */
654 	}
655 
656 	/* If a grh is present, we demux according to it */
657 	if (wc->wc_flags & IB_WC_GRH) {
658 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
659 		if (slave < 0) {
660 			mlx4_ib_warn(ibdev, "failed matching grh\n");
661 			return -ENOENT;
662 		}
663 	}
664 	/* Class-specific handling */
665 	switch (mad->mad_hdr.mgmt_class) {
666 	case IB_MGMT_CLASS_SUBN_ADM:
667 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
668 					     (struct ib_sa_mad *) mad))
669 			return 0;
670 		break;
671 	case IB_MGMT_CLASS_CM:
672 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad, is_eth))
673 			return 0;
674 		break;
675 	case IB_MGMT_CLASS_DEVICE_MGMT:
676 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
677 			return 0;
678 		break;
679 	default:
680 		/* Drop unsupported classes for slaves in tunnel mode */
681 		if (slave != mlx4_master_func_num(dev->dev)) {
682 			pr_debug("dropping unsupported ingress mad from class:%d "
683 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
684 			return 0;
685 		}
686 	}
687 	/*make sure that no slave==255 was not handled yet.*/
688 	if (slave >= dev->dev->caps.sqp_demux) {
689 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
690 			     slave, dev->dev->caps.sqp_demux);
691 		return -ENOENT;
692 	}
693 
694 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
695 	if (err)
696 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
697 			 slave, err);
698 	return 0;
699 }
700 
701 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
702 			struct ib_wc *in_wc, struct ib_grh *in_grh,
703 			struct ib_mad *in_mad, struct ib_mad *out_mad)
704 {
705 	u16 slid, prev_lid = 0;
706 	int err;
707 	struct ib_port_attr pattr;
708 
709 	if (in_wc && in_wc->qp->qp_num) {
710 		pr_debug("received MAD: slid:%d sqpn:%d "
711 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
712 			in_wc->slid, in_wc->src_qp,
713 			in_wc->dlid_path_bits,
714 			in_wc->qp->qp_num,
715 			in_wc->wc_flags,
716 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
717 			be16_to_cpu(in_mad->mad_hdr.attr_id));
718 		if (in_wc->wc_flags & IB_WC_GRH) {
719 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
720 				 (unsigned long long)be64_to_cpu(in_grh->sgid.global.subnet_prefix),
721 				 (unsigned long long)be64_to_cpu(in_grh->sgid.global.interface_id));
722 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
723 				 (unsigned long long)be64_to_cpu(in_grh->dgid.global.subnet_prefix),
724 				 (unsigned long long)be64_to_cpu(in_grh->dgid.global.interface_id));
725 		}
726 	}
727 
728 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
729 
730 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
731 		forward_trap(to_mdev(ibdev), port_num, in_mad);
732 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
733 	}
734 
735 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
736 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
737 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
738 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
739 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
740 			return IB_MAD_RESULT_SUCCESS;
741 
742 		/*
743 		 * Don't process SMInfo queries -- the SMA can't handle them.
744 		 */
745 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
746 			return IB_MAD_RESULT_SUCCESS;
747 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
748 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
749 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
750 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
751 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
752 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
753 			return IB_MAD_RESULT_SUCCESS;
754 	} else
755 		return IB_MAD_RESULT_SUCCESS;
756 
757 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
758 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
759 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
760 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
761 	    !ib_query_port(ibdev, port_num, &pattr))
762 		prev_lid = pattr.lid;
763 
764 	err = mlx4_MAD_IFC(to_mdev(ibdev),
765 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
766 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
767 			   MLX4_MAD_IFC_NET_VIEW,
768 			   port_num, in_wc, in_grh, in_mad, out_mad);
769 	if (err)
770 		return IB_MAD_RESULT_FAILURE;
771 
772 	if (!out_mad->mad_hdr.status) {
773 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
774 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
775 		/* slaves get node desc from FW */
776 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
777 			node_desc_override(ibdev, out_mad);
778 	}
779 
780 	/* set return bit in status of directed route responses */
781 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
782 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
783 
784 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
785 		/* no response for trap repress */
786 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
787 
788 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
789 }
790 
791 static void edit_counter_ext(struct mlx4_if_stat_extended *cnt, void *counters,
792 			     __be16 attr_id)
793 {
794 	switch (attr_id) {
795 	case IB_PMA_PORT_COUNTERS:
796 	{
797 		struct ib_pma_portcounters *pma_cnt =
798 				(struct ib_pma_portcounters *)counters;
799 		pma_cnt->port_xmit_data =
800 			cpu_to_be32((be64_to_cpu(cnt->counters[0].
801 						 IfTxUnicastOctets) +
802 				     be64_to_cpu(cnt->counters[0].
803 						 IfTxMulticastOctets) +
804 				     be64_to_cpu(cnt->counters[0].
805 						 IfTxBroadcastOctets) +
806 				     be64_to_cpu(cnt->counters[0].
807 						 IfTxDroppedOctets)) >> 2);
808 		pma_cnt->port_rcv_data  =
809 			cpu_to_be32((be64_to_cpu(cnt->counters[0].
810 						 IfRxUnicastOctets) +
811 				     be64_to_cpu(cnt->counters[0].
812 						 IfRxMulticastOctets) +
813 				     be64_to_cpu(cnt->counters[0].
814 						 IfRxBroadcastOctets) +
815 				     be64_to_cpu(cnt->counters[0].
816 						 IfRxNoBufferOctets) +
817 				     be64_to_cpu(cnt->counters[0].
818 						 IfRxErrorOctets)) >> 2);
819 		pma_cnt->port_xmit_packets =
820 			cpu_to_be32(be64_to_cpu(cnt->counters[0].
821 						IfTxUnicastFrames) +
822 				    be64_to_cpu(cnt->counters[0].
823 						IfTxMulticastFrames) +
824 				    be64_to_cpu(cnt->counters[0].
825 						IfTxBroadcastFrames) +
826 				    be64_to_cpu(cnt->counters[0].
827 						IfTxDroppedFrames));
828 		pma_cnt->port_rcv_packets  =
829 			cpu_to_be32(be64_to_cpu(cnt->counters[0].
830 						IfRxUnicastFrames) +
831 				    be64_to_cpu(cnt->counters[0].
832 						IfRxMulticastFrames) +
833 				    be64_to_cpu(cnt->counters[0].
834 						IfRxBroadcastFrames) +
835 				    be64_to_cpu(cnt->counters[0].
836 						IfRxNoBufferFrames) +
837 				    be64_to_cpu(cnt->counters[0].
838 						IfRxErrorFrames));
839 		pma_cnt->port_rcv_errors = cpu_to_be32(be64_to_cpu(cnt->
840 						       counters[0].
841 						       IfRxErrorFrames));
842 		break;
843 	}
844 
845 	case IB_PMA_PORT_COUNTERS_EXT:
846 	{
847 		struct ib_pma_portcounters_ext *pma_cnt_ext =
848 				(struct ib_pma_portcounters_ext *)counters;
849 
850 		pma_cnt_ext->port_xmit_data =
851 			cpu_to_be64((be64_to_cpu(cnt->counters[0].
852 						 IfTxUnicastOctets) +
853 				     be64_to_cpu(cnt->counters[0].
854 						 IfTxMulticastOctets) +
855 				     be64_to_cpu(cnt->counters[0].
856 						 IfTxBroadcastOctets) +
857 				     be64_to_cpu(cnt->counters[0].
858 						 IfTxDroppedOctets)) >> 2);
859 		pma_cnt_ext->port_rcv_data  =
860 			cpu_to_be64((be64_to_cpu(cnt->counters[0].
861 						 IfRxUnicastOctets) +
862 				     be64_to_cpu(cnt->counters[0].
863 						 IfRxMulticastOctets) +
864 				     be64_to_cpu(cnt->counters[0].
865 						 IfRxBroadcastOctets) +
866 				     be64_to_cpu(cnt->counters[0].
867 						 IfRxNoBufferOctets) +
868 				     be64_to_cpu(cnt->counters[0].
869 						 IfRxErrorOctets)) >> 2);
870 		pma_cnt_ext->port_xmit_packets =
871 			cpu_to_be64(be64_to_cpu(cnt->counters[0].
872 						IfTxUnicastFrames) +
873 				    be64_to_cpu(cnt->counters[0].
874 						IfTxMulticastFrames) +
875 				    be64_to_cpu(cnt->counters[0].
876 						IfTxBroadcastFrames) +
877 				    be64_to_cpu(cnt->counters[0].
878 						IfTxDroppedFrames));
879 		pma_cnt_ext->port_rcv_packets  =
880 			cpu_to_be64(be64_to_cpu(cnt->counters[0].
881 						IfRxUnicastFrames) +
882 				    be64_to_cpu(cnt->counters[0].
883 						IfRxMulticastFrames) +
884 				    be64_to_cpu(cnt->counters[0].
885 						IfRxBroadcastFrames) +
886 				    be64_to_cpu(cnt->counters[0].
887 						IfRxNoBufferFrames) +
888 				    be64_to_cpu(cnt->counters[0].
889 						IfRxErrorFrames));
890 		pma_cnt_ext->port_unicast_xmit_packets = cnt->counters[0].
891 						IfTxUnicastFrames;
892 		pma_cnt_ext->port_unicast_rcv_packets = cnt->counters[0].
893 						IfRxUnicastFrames;
894 		pma_cnt_ext->port_multicast_xmit_packets =
895 			cpu_to_be64(be64_to_cpu(cnt->counters[0].
896 						IfTxMulticastFrames) +
897 				    be64_to_cpu(cnt->counters[0].
898 						IfTxBroadcastFrames));
899 		pma_cnt_ext->port_multicast_rcv_packets =
900 			cpu_to_be64(be64_to_cpu(cnt->counters[0].
901 						IfTxMulticastFrames) +
902 				    be64_to_cpu(cnt->counters[0].
903 						IfTxBroadcastFrames));
904 
905 		break;
906 	}
907 
908 	default:
909 		pr_warn("Unsupported attr_id 0x%x\n", attr_id);
910 		break;
911 	}
912 
913 }
914 
915 static void edit_counter(struct mlx4_if_stat_basic *cnt, void *counters,
916 			 __be16	attr_id)
917 {
918 	switch (attr_id) {
919 	case IB_PMA_PORT_COUNTERS:
920 	{
921 		struct ib_pma_portcounters *pma_cnt =
922 				(struct ib_pma_portcounters *) counters;
923 		pma_cnt->port_xmit_data =
924 			cpu_to_be32(be64_to_cpu(
925 				    cnt->counters[0].IfTxOctets) >> 2);
926 		pma_cnt->port_rcv_data  =
927 			cpu_to_be32(be64_to_cpu(
928 				    cnt->counters[0].IfRxOctets) >> 2);
929 		pma_cnt->port_xmit_packets =
930 			cpu_to_be32(be64_to_cpu(cnt->counters[0].IfTxFrames));
931 		pma_cnt->port_rcv_packets  =
932 			cpu_to_be32(be64_to_cpu(cnt->counters[0].IfRxFrames));
933 		break;
934 	}
935 	case IB_PMA_PORT_COUNTERS_EXT:
936 	{
937 		struct ib_pma_portcounters_ext *pma_cnt_ext =
938 				(struct ib_pma_portcounters_ext *) counters;
939 
940 		pma_cnt_ext->port_xmit_data =
941 			cpu_to_be64((be64_to_cpu(cnt->counters[0].
942 						 IfTxOctets) >> 2));
943 		pma_cnt_ext->port_rcv_data  =
944 			cpu_to_be64((be64_to_cpu(cnt->counters[0].
945 						 IfRxOctets) >> 2));
946 		pma_cnt_ext->port_xmit_packets = cnt->counters[0].IfTxFrames;
947 		pma_cnt_ext->port_rcv_packets  = cnt->counters[0].IfRxFrames;
948 		break;
949 	}
950 	default:
951 		pr_warn("Unsupported attr_id 0x%x\n", attr_id);
952 		break;
953 	}
954 }
955 
956 int mlx4_ib_query_if_stat(struct mlx4_ib_dev *dev, u32 counter_index,
957 		       union mlx4_counter *counter, u8 clear)
958 {
959 	struct mlx4_cmd_mailbox *mailbox;
960 	int err;
961 	u32 inmod = counter_index | ((clear & 1) << 31);
962 
963 	mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
964 	if (IS_ERR(mailbox))
965 		return IB_MAD_RESULT_FAILURE;
966 
967 	err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
968 			   MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
969 			   MLX4_CMD_NATIVE);
970 	if (!err)
971 		memcpy(counter, mailbox->buf, MLX4_IF_STAT_SZ(1));
972 
973 	mlx4_free_cmd_mailbox(dev->dev, mailbox);
974 
975 	return err;
976 }
977 
978 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
979 			struct ib_wc *in_wc, struct ib_grh *in_grh,
980 			struct ib_mad *in_mad, struct ib_mad *out_mad)
981 {
982 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
983 	int err;
984 	u32 counter_index = dev->counters[port_num - 1].counter_index & 0xffff;
985 	u8 mode;
986 	char				counter_buf[MLX4_IF_STAT_SZ(1)] __aligned(8);
987 	union  mlx4_counter		*counter = (union mlx4_counter *)
988 						   counter_buf;
989 
990 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
991 		return -EINVAL;
992 
993 	/* in case of default counter IB shares the counter with ETH */
994 	/* the state could be -EEXIST or -ENOSPC */
995 	if (dev->counters[port_num - 1].status) {
996 		memset(out_mad->data, 0, sizeof out_mad->data);
997 		err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
998 	} else {
999 		if (mlx4_ib_query_if_stat(dev, counter_index, counter, 0))
1000 			return IB_MAD_RESULT_FAILURE;
1001 
1002 		memset(out_mad->data, 0, sizeof(out_mad->data));
1003 		mode = counter->control.cnt_mode & 0xFF;
1004 		err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
1005 		switch (mode & 0xf) {
1006 		case 0:
1007 			edit_counter((void *)counter,
1008 				     (void *)(out_mad->data + 40),
1009 				     in_mad->mad_hdr.attr_id);
1010 			break;
1011 		case 1:
1012 			edit_counter_ext((void *)counter,
1013 					 (void *)(out_mad->data + 40),
1014 					 in_mad->mad_hdr.attr_id);
1015 			break;
1016 		default:
1017 			err = IB_MAD_RESULT_FAILURE;
1018 		}
1019 	}
1020 
1021 	return err;
1022 }
1023 
1024 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
1025 			struct ib_wc *in_wc, struct ib_grh *in_grh,
1026 			struct ib_mad *in_mad, struct ib_mad *out_mad)
1027 {
1028 	switch (rdma_port_get_link_layer(ibdev, port_num)) {
1029 	case IB_LINK_LAYER_INFINIBAND:
1030 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
1031 				      in_grh, in_mad, out_mad);
1032 	case IB_LINK_LAYER_ETHERNET:
1033 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1034 					  in_grh, in_mad, out_mad);
1035 	default:
1036 		return -EINVAL;
1037 	}
1038 }
1039 
1040 static void send_handler(struct ib_mad_agent *agent,
1041 			 struct ib_mad_send_wc *mad_send_wc)
1042 {
1043 	if (mad_send_wc->send_buf->context[0])
1044 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
1045 	ib_free_send_mad(mad_send_wc->send_buf);
1046 }
1047 
1048 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1049 {
1050 	struct ib_mad_agent *agent;
1051 	int p, q;
1052 	int ret;
1053 	enum rdma_link_layer ll;
1054 
1055 	for (p = 0; p < dev->num_ports; ++p) {
1056 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1057 		for (q = 0; q <= 1; ++q) {
1058 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1059 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1060 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1061 							      NULL, 0, send_handler,
1062 							      NULL, NULL);
1063 				if (IS_ERR(agent)) {
1064 					ret = PTR_ERR(agent);
1065 					goto err;
1066 				}
1067 				dev->send_agent[p][q] = agent;
1068 			} else
1069 				dev->send_agent[p][q] = NULL;
1070 		}
1071 	}
1072 
1073 	return 0;
1074 
1075 err:
1076 	for (p = 0; p < dev->num_ports; ++p)
1077 		for (q = 0; q <= 1; ++q)
1078 			if (dev->send_agent[p][q])
1079 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1080 
1081 	return ret;
1082 }
1083 
1084 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1085 {
1086 	struct ib_mad_agent *agent;
1087 	int p, q;
1088 
1089 	for (p = 0; p < dev->num_ports; ++p) {
1090 		for (q = 0; q <= 1; ++q) {
1091 			agent = dev->send_agent[p][q];
1092 			if (agent) {
1093 				dev->send_agent[p][q] = NULL;
1094 				ib_unregister_mad_agent(agent);
1095 			}
1096 		}
1097 
1098 		if (dev->sm_ah[p])
1099 			ib_destroy_ah(dev->sm_ah[p]);
1100 	}
1101 }
1102 
1103 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1104 {
1105 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1106 
1107 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1108 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1109 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK, 0, 0);
1110 }
1111 
1112 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1113 {
1114 	/* re-configure the alias-guid and mcg's */
1115 	if (mlx4_is_master(dev->dev)) {
1116 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1117 
1118 		if (!dev->sriov.is_going_down) {
1119 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1120 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1121 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK, 0, 0);
1122 		}
1123 	}
1124 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1125 }
1126 
1127 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1128 			      struct mlx4_eqe *eqe)
1129 {
1130 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1131 			    GET_MASK_FROM_EQE(eqe));
1132 }
1133 
1134 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1135 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1136 {
1137 	struct ib_smp *in_mad  = NULL;
1138 	struct ib_smp *out_mad  = NULL;
1139 	u16 i;
1140 
1141 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1142 		return;
1143 
1144 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1145 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1146 	if (!in_mad || !out_mad) {
1147 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1148 		goto out;
1149 	}
1150 
1151 	guid_tbl_blk_num  *= 4;
1152 
1153 	for (i = 0; i < 4; i++) {
1154 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1155 			continue;
1156 		memset(in_mad, 0, sizeof *in_mad);
1157 		memset(out_mad, 0, sizeof *out_mad);
1158 
1159 		in_mad->base_version  = 1;
1160 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1161 		in_mad->class_version = 1;
1162 		in_mad->method        = IB_MGMT_METHOD_GET;
1163 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1164 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1165 
1166 		if (mlx4_MAD_IFC(dev,
1167 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1168 				 port_num, NULL, NULL, in_mad, out_mad)) {
1169 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1170 			goto out;
1171 		}
1172 
1173 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1174 						    port_num,
1175 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1176 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1177 						     port_num,
1178 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1179 	}
1180 
1181 out:
1182 	kfree(in_mad);
1183 	kfree(out_mad);
1184 	return;
1185 }
1186 
1187 void handle_port_mgmt_change_event(struct work_struct *work)
1188 {
1189 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1190 	struct mlx4_ib_dev *dev = ew->ib_dev;
1191 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1192 	u8 port = eqe->event.port_mgmt_change.port;
1193 	u32 changed_attr;
1194 	u32 tbl_block;
1195 	u32 change_bitmap;
1196 
1197 	switch (eqe->subtype) {
1198 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1199 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1200 
1201 		/* Update the SM ah - This should be done before handling
1202 		   the other changed attributes so that MADs can be sent to the SM */
1203 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1204 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1205 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1206 			update_sm_ah(dev, port, lid, sl);
1207 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_SM_CHANGE);
1208 			if (mlx4_is_master(dev->dev))
1209 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1210 							    changed_attr & MSTR_SM_CHANGE_MASK,
1211 							    lid, sl);
1212 		}
1213 
1214 		/* Check if it is a lid change event */
1215 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1216 			handle_lid_change_event(dev, port);
1217 
1218 		/* Generate GUID changed event */
1219 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1220 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1221 			/*if master, notify all slaves*/
1222 			if (mlx4_is_master(dev->dev))
1223 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1224 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK, 0, 0);
1225 		}
1226 
1227 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1228 			handle_client_rereg_event(dev, port);
1229 		break;
1230 
1231 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1232 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1233 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1234 			propagate_pkey_ev(dev, port, eqe);
1235 		break;
1236 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1237 		/* paravirtualized master's guid is guid 0 -- does not change */
1238 		if (!mlx4_is_master(dev->dev))
1239 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1240 		/*if master, notify relevant slaves*/
1241 		else if (!dev->sriov.is_going_down) {
1242 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1243 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1244 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1245 		}
1246 		break;
1247 	default:
1248 		pr_warn("Unsupported subtype 0x%x for "
1249 			"Port Management Change event\n", eqe->subtype);
1250 	}
1251 
1252 	kfree(ew);
1253 }
1254 
1255 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1256 			    enum ib_event_type type)
1257 {
1258 	struct ib_event event;
1259 
1260 	event.device		= &dev->ib_dev;
1261 	event.element.port_num	= port_num;
1262 	event.event		= type;
1263 
1264 	ib_dispatch_event(&event);
1265 }
1266 
1267 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1268 {
1269 	unsigned long flags;
1270 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1271 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1272 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1273 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1274 		queue_work(ctx->wq, &ctx->work);
1275 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1276 }
1277 
1278 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1279 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1280 				  int index)
1281 {
1282 	struct ib_sge sg_list;
1283 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1284 	int size;
1285 
1286 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1287 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1288 
1289 	sg_list.addr = tun_qp->ring[index].map;
1290 	sg_list.length = size;
1291 	sg_list.lkey = ctx->mr->lkey;
1292 
1293 	recv_wr.next = NULL;
1294 	recv_wr.sg_list = &sg_list;
1295 	recv_wr.num_sge = 1;
1296 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1297 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1298 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1299 				      size, DMA_FROM_DEVICE);
1300 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1301 }
1302 
1303 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1304 		int slave, struct ib_sa_mad *sa_mad)
1305 {
1306 	int ret = 0;
1307 
1308 	/* dispatch to different sa handlers */
1309 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1310 	case IB_SA_ATTR_MC_MEMBER_REC:
1311 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1312 		break;
1313 	default:
1314 		break;
1315 	}
1316 	return ret;
1317 }
1318 
1319 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1320 {
1321 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1322 
1323 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1324 }
1325 
1326 
1327 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1328 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1329 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1330 			 u8 *s_mac, struct ib_mad *mad)
1331 {
1332 	struct ib_sge list;
1333 	struct ib_send_wr wr, *bad_wr;
1334 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1335 	struct mlx4_ib_demux_pv_qp *sqp;
1336 	struct mlx4_mad_snd_buf *sqp_mad;
1337 	struct ib_ah *ah;
1338 	struct ib_qp *send_qp = NULL;
1339 	unsigned wire_tx_ix = 0;
1340 	int ret = 0;
1341 	u16 wire_pkey_ix;
1342 	int src_qpnum;
1343 	u8 sgid_index;
1344 
1345 
1346 	sqp_ctx = dev->sriov.sqps[port-1];
1347 
1348 	/* check if proxy qp created */
1349 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1350 		return -EAGAIN;
1351 
1352 	/* QP0 forwarding only for Dom0 */
1353 	if (dest_qpt == IB_QPT_SMI && (mlx4_master_func_num(dev->dev) != slave))
1354 		return -EINVAL;
1355 
1356 	if (dest_qpt == IB_QPT_SMI) {
1357 		src_qpnum = 0;
1358 		sqp = &sqp_ctx->qp[0];
1359 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1360 	} else {
1361 		src_qpnum = 1;
1362 		sqp = &sqp_ctx->qp[1];
1363 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1364 	}
1365 
1366 	send_qp = sqp->qp;
1367 
1368 	/* create ah */
1369 	sgid_index = attr->grh.sgid_index;
1370 	attr->grh.sgid_index = 0;
1371 	ah = ib_create_ah(sqp_ctx->pd, attr);
1372 	if (IS_ERR(ah))
1373 		return -ENOMEM;
1374 	attr->grh.sgid_index = sgid_index;
1375 	to_mah(ah)->av.ib.gid_index = sgid_index;
1376 	/* get rid of force-loopback bit */
1377 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1378 	spin_lock(&sqp->tx_lock);
1379 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1380 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1381 		ret = -EAGAIN;
1382 	else
1383 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1384 	spin_unlock(&sqp->tx_lock);
1385 	if (ret)
1386 		goto out;
1387 
1388 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1389 	if (sqp->tx_ring[wire_tx_ix].ah)
1390 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1391 	sqp->tx_ring[wire_tx_ix].ah = ah;
1392 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1393 				   sqp->tx_ring[wire_tx_ix].buf.map,
1394 				   sizeof (struct mlx4_mad_snd_buf),
1395 				   DMA_TO_DEVICE);
1396 
1397 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1398 
1399 	ib_dma_sync_single_for_device(&dev->ib_dev,
1400 				      sqp->tx_ring[wire_tx_ix].buf.map,
1401 				      sizeof (struct mlx4_mad_snd_buf),
1402 				      DMA_TO_DEVICE);
1403 
1404 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1405 	list.length = sizeof (struct mlx4_mad_snd_buf);
1406 	list.lkey = sqp_ctx->mr->lkey;
1407 
1408 	wr.wr.ud.ah = ah;
1409 	wr.wr.ud.port_num = port;
1410 	wr.wr.ud.pkey_index = wire_pkey_ix;
1411 	wr.wr.ud.remote_qkey = qkey;
1412 	wr.wr.ud.remote_qpn = remote_qpn;
1413 	wr.next = NULL;
1414 	wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1415 	wr.sg_list = &list;
1416 	wr.num_sge = 1;
1417 	wr.opcode = IB_WR_SEND;
1418 	wr.send_flags = IB_SEND_SIGNALED;
1419 	if (s_mac)
1420 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1421 
1422 
1423 	ret = ib_post_send(send_qp, &wr, &bad_wr);
1424 out:
1425 	if (ret)
1426 		ib_destroy_ah(ah);
1427 	return ret;
1428 }
1429 
1430 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1431 {
1432 	int gids;
1433 	int vfs;
1434 
1435 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1436 		return slave;
1437 
1438 	gids = MLX4_ROCE_MAX_GIDS - MLX4_ROCE_PF_GIDS;
1439 	vfs = dev->dev->num_vfs;
1440 
1441 	if (slave == 0)
1442 		return 0;
1443 	if (slave <= gids % vfs)
1444 		return MLX4_ROCE_PF_GIDS + ((gids / vfs) + 1) * (slave - 1);
1445 
1446 	return MLX4_ROCE_PF_GIDS + (gids % vfs) + ((gids / vfs) * (slave - 1));
1447 }
1448 
1449 static int get_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1450 			       struct ib_ah_attr *ah_attr)
1451 {
1452 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND) {
1453 		ah_attr->grh.sgid_index = slave;
1454 		return 0;
1455 	}
1456 	ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1457 	return 0;
1458 }
1459 
1460 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1461 {
1462 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1463 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1464 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1465 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1466 	struct mlx4_ib_ah ah;
1467 	struct ib_ah_attr ah_attr;
1468 	u8 *slave_id;
1469 	int slave;
1470 
1471 	/* Get slave that sent this packet */
1472 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1473 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1474 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1475 	    wc->src_qp & 0x4) {
1476 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1477 		return;
1478 	}
1479 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1480 	if (slave != ctx->slave) {
1481 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1482 			     "belongs to another slave\n", wc->src_qp);
1483 		return;
1484 	}
1485 	if (slave != mlx4_master_func_num(dev->dev) && !(wc->src_qp & 0x2)) {
1486 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1487 			     "non-master trying to send QP0 packets\n", wc->src_qp);
1488 		return;
1489 	}
1490 
1491 	/* Map transaction ID */
1492 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1493 				   sizeof (struct mlx4_tunnel_mad),
1494 				   DMA_FROM_DEVICE);
1495 	switch (tunnel->mad.mad_hdr.method) {
1496 	case IB_MGMT_METHOD_SET:
1497 	case IB_MGMT_METHOD_GET:
1498 	case IB_MGMT_METHOD_REPORT:
1499 	case IB_SA_METHOD_GET_TABLE:
1500 	case IB_SA_METHOD_DELETE:
1501 	case IB_SA_METHOD_GET_MULTI:
1502 	case IB_SA_METHOD_GET_TRACE_TBL:
1503 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1504 		if (*slave_id) {
1505 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1506 				     "class:%d slave:%d\n", *slave_id,
1507 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1508 			return;
1509 		} else
1510 			*slave_id = slave;
1511 	default:
1512 		/* nothing */;
1513 	}
1514 
1515 	/* Class-specific handling */
1516 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1517 	case IB_MGMT_CLASS_SUBN_ADM:
1518 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1519 			      (struct ib_sa_mad *) &tunnel->mad))
1520 			return;
1521 		break;
1522 	case IB_MGMT_CLASS_CM:
1523 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1524 			      (struct ib_mad *) &tunnel->mad))
1525 			return;
1526 		break;
1527 	case IB_MGMT_CLASS_DEVICE_MGMT:
1528 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1529 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1530 			return;
1531 		break;
1532 	default:
1533 		/* Drop unsupported classes for slaves in tunnel mode */
1534 		if (slave != mlx4_master_func_num(dev->dev)) {
1535 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1536 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1537 			return;
1538 		}
1539 	}
1540 
1541 	/* We are using standard ib_core services to send the mad, so generate a
1542 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1543 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1544 	ah.ibah.device = ctx->ib_dev;
1545 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1546 	if (ah_attr.ah_flags & IB_AH_GRH)
1547 		if (get_real_sgid_index(dev, slave, ctx->port, &ah_attr))
1548 			return;
1549 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1550 	ah_attr.vlan_id = tunnel->hdr.vlan;
1551 	/* if slave have default vlan use it */
1552 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1553 				    &ah_attr.vlan_id, &ah_attr.sl);
1554 
1555 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1556 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1557 			     IB_QPT_SMI : IB_QPT_GSI,
1558 			     be16_to_cpu(tunnel->hdr.pkey_index),
1559 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1560 			     be32_to_cpu(tunnel->hdr.qkey),
1561 			     &ah_attr, wc->smac, &tunnel->mad);
1562 }
1563 
1564 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1565 				 enum ib_qp_type qp_type, int is_tun)
1566 {
1567 	int i;
1568 	struct mlx4_ib_demux_pv_qp *tun_qp;
1569 	int rx_buf_size, tx_buf_size;
1570 
1571 	if (qp_type > IB_QPT_GSI)
1572 		return -EINVAL;
1573 
1574 	tun_qp = &ctx->qp[qp_type];
1575 
1576 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1577 			       GFP_KERNEL);
1578 	if (!tun_qp->ring)
1579 		return -ENOMEM;
1580 
1581 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1582 				  sizeof (struct mlx4_ib_tun_tx_buf),
1583 				  GFP_KERNEL);
1584 	if (!tun_qp->tx_ring) {
1585 		kfree(tun_qp->ring);
1586 		tun_qp->ring = NULL;
1587 		return -ENOMEM;
1588 	}
1589 
1590 	if (is_tun) {
1591 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1592 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1593 	} else {
1594 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1595 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1596 	}
1597 
1598 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1599 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1600 		if (!tun_qp->ring[i].addr)
1601 			goto err;
1602 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1603 							tun_qp->ring[i].addr,
1604 							rx_buf_size,
1605 							DMA_FROM_DEVICE);
1606 		if (unlikely(ib_dma_mapping_error(ctx->ib_dev,
1607 						  tun_qp->ring[i].map))) {
1608 			mlx4_ib_warn(ctx->ib_dev, "ib_dma_map_single failed\n");
1609 			kfree(tun_qp->ring[i].addr);
1610 			goto err;
1611 		}
1612 	}
1613 
1614 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1615 		tun_qp->tx_ring[i].buf.addr =
1616 			kmalloc(tx_buf_size, GFP_KERNEL);
1617 		if (!tun_qp->tx_ring[i].buf.addr)
1618 			goto tx_err;
1619 		tun_qp->tx_ring[i].buf.map =
1620 			ib_dma_map_single(ctx->ib_dev,
1621 					  tun_qp->tx_ring[i].buf.addr,
1622 					  tx_buf_size,
1623 					  DMA_TO_DEVICE);
1624 		if (unlikely(ib_dma_mapping_error(ctx->ib_dev,
1625 						  tun_qp->tx_ring[i].buf.map))) {
1626 			mlx4_ib_warn(ctx->ib_dev, "ib_dma_map_single failed\n");
1627 			kfree(tun_qp->tx_ring[i].buf.addr);
1628 			goto tx_err;
1629 		}
1630 		tun_qp->tx_ring[i].ah = NULL;
1631 	}
1632 	spin_lock_init(&tun_qp->tx_lock);
1633 	tun_qp->tx_ix_head = 0;
1634 	tun_qp->tx_ix_tail = 0;
1635 	tun_qp->proxy_qpt = qp_type;
1636 
1637 	return 0;
1638 
1639 tx_err:
1640 	while (i > 0) {
1641 		--i;
1642 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1643 				    tx_buf_size, DMA_TO_DEVICE);
1644 		kfree(tun_qp->tx_ring[i].buf.addr);
1645 	}
1646 	kfree(tun_qp->tx_ring);
1647 	tun_qp->tx_ring = NULL;
1648 	i = MLX4_NUM_TUNNEL_BUFS;
1649 err:
1650 	while (i > 0) {
1651 		--i;
1652 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1653 				    rx_buf_size, DMA_FROM_DEVICE);
1654 		kfree(tun_qp->ring[i].addr);
1655 	}
1656 	kfree(tun_qp->ring);
1657 	tun_qp->ring = NULL;
1658 	return -ENOMEM;
1659 }
1660 
1661 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1662 				     enum ib_qp_type qp_type, int is_tun)
1663 {
1664 	int i;
1665 	struct mlx4_ib_demux_pv_qp *tun_qp;
1666 	int rx_buf_size, tx_buf_size;
1667 
1668 	if (qp_type > IB_QPT_GSI)
1669 		return;
1670 
1671 	tun_qp = &ctx->qp[qp_type];
1672 	if (is_tun) {
1673 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1674 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1675 	} else {
1676 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1677 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1678 	}
1679 
1680 
1681 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1682 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1683 				    rx_buf_size, DMA_FROM_DEVICE);
1684 		kfree(tun_qp->ring[i].addr);
1685 	}
1686 
1687 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1688 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1689 				    tx_buf_size, DMA_TO_DEVICE);
1690 		kfree(tun_qp->tx_ring[i].buf.addr);
1691 		if (tun_qp->tx_ring[i].ah)
1692 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1693 	}
1694 	kfree(tun_qp->tx_ring);
1695 	kfree(tun_qp->ring);
1696 }
1697 
1698 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1699 {
1700 	struct mlx4_ib_demux_pv_ctx *ctx;
1701 	struct mlx4_ib_demux_pv_qp *tun_qp;
1702 	struct ib_wc wc;
1703 	int ret;
1704 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1705 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1706 
1707 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1708 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1709 		if (wc.status == IB_WC_SUCCESS) {
1710 			switch (wc.opcode) {
1711 			case IB_WC_RECV:
1712 				mlx4_ib_multiplex_mad(ctx, &wc);
1713 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1714 							     wc.wr_id &
1715 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1716 				if (ret)
1717 					pr_err("Failed reposting tunnel "
1718 					       "buf:%lld\n", (unsigned long long)wc.wr_id);
1719 				break;
1720 			case IB_WC_SEND:
1721 				pr_debug("received tunnel send completion:"
1722 					 "wrid=0x%llx, status=0x%x\n",
1723 					 (unsigned long long)wc.wr_id, wc.status);
1724 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1725 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1726 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1727 					= NULL;
1728 				spin_lock(&tun_qp->tx_lock);
1729 				tun_qp->tx_ix_tail++;
1730 				spin_unlock(&tun_qp->tx_lock);
1731 
1732 				break;
1733 			default:
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, (unsigned long long)wc.wr_id);
1740 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1741 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1742 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1743 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1744 					= NULL;
1745 				spin_lock(&tun_qp->tx_lock);
1746 				tun_qp->tx_ix_tail++;
1747 				spin_unlock(&tun_qp->tx_lock);
1748 			}
1749 		}
1750 	}
1751 }
1752 
1753 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1754 {
1755 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1756 
1757 	/* It's worse than that! He's dead, Jim! */
1758 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1759 	       event->event, sqp->port);
1760 }
1761 
1762 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1763 			    enum ib_qp_type qp_type, int create_tun)
1764 {
1765 	int i, ret;
1766 	struct mlx4_ib_demux_pv_qp *tun_qp;
1767 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1768 	struct ib_qp_attr attr;
1769 	int qp_attr_mask_INIT;
1770 
1771 	if (qp_type > IB_QPT_GSI)
1772 		return -EINVAL;
1773 
1774 	tun_qp = &ctx->qp[qp_type];
1775 
1776 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1777 	qp_init_attr.init_attr.send_cq = ctx->cq;
1778 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1779 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1780 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1781 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1782 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1783 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1784 	if (create_tun) {
1785 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1786 		qp_init_attr.init_attr.create_flags = (enum ib_qp_create_flags)MLX4_IB_SRIOV_TUNNEL_QP;
1787 		qp_init_attr.port = ctx->port;
1788 		qp_init_attr.slave = ctx->slave;
1789 		qp_init_attr.proxy_qp_type = qp_type;
1790 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1791 			   IB_QP_QKEY | IB_QP_PORT;
1792 	} else {
1793 		qp_init_attr.init_attr.qp_type = qp_type;
1794 		qp_init_attr.init_attr.create_flags = (enum ib_qp_create_flags)MLX4_IB_SRIOV_SQP;
1795 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1796 	}
1797 	qp_init_attr.init_attr.port_num = ctx->port;
1798 	qp_init_attr.init_attr.qp_context = ctx;
1799 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1800 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1801 	if (IS_ERR(tun_qp->qp)) {
1802 		ret = PTR_ERR(tun_qp->qp);
1803 		tun_qp->qp = NULL;
1804 		pr_err("Couldn't create %s QP (%d)\n",
1805 		       create_tun ? "tunnel" : "special", ret);
1806 		return ret;
1807 	}
1808 
1809 	memset(&attr, 0, sizeof attr);
1810 	attr.qp_state = IB_QPS_INIT;
1811 	ret = 0;
1812 	if (create_tun)
1813 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1814 					      ctx->port, 0xFFFF, &attr.pkey_index);
1815 	if (ret || !create_tun)
1816 		attr.pkey_index =
1817 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1818 	attr.qkey = IB_QP1_QKEY;
1819 	attr.port_num = ctx->port;
1820 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1821 	if (ret) {
1822 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1823 		       create_tun ? "tunnel" : "special", ret);
1824 		goto err_qp;
1825 	}
1826 	attr.qp_state = IB_QPS_RTR;
1827 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1828 	if (ret) {
1829 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1830 		       create_tun ? "tunnel" : "special", ret);
1831 		goto err_qp;
1832 	}
1833 	attr.qp_state = IB_QPS_RTS;
1834 	attr.sq_psn = 0;
1835 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1836 	if (ret) {
1837 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1838 		       create_tun ? "tunnel" : "special", ret);
1839 		goto err_qp;
1840 	}
1841 
1842 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1843 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1844 		if (ret) {
1845 			pr_err(" mlx4_ib_post_pv_buf error"
1846 			       " (err = %d, i = %d)\n", ret, i);
1847 			goto err_qp;
1848 		}
1849 	}
1850 	return 0;
1851 
1852 err_qp:
1853 	ib_destroy_qp(tun_qp->qp);
1854 	tun_qp->qp = NULL;
1855 	return ret;
1856 }
1857 
1858 /*
1859  * IB MAD completion callback for real SQPs
1860  */
1861 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1862 {
1863 	struct mlx4_ib_demux_pv_ctx *ctx;
1864 	struct mlx4_ib_demux_pv_qp *sqp;
1865 	struct ib_wc wc;
1866 	struct ib_grh *grh;
1867 	struct ib_mad *mad;
1868 
1869 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1870 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1871 
1872 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1873 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1874 		if (wc.status == IB_WC_SUCCESS) {
1875 			switch (wc.opcode) {
1876 			case IB_WC_SEND:
1877 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1878 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1879 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1880 					= NULL;
1881 				spin_lock(&sqp->tx_lock);
1882 				sqp->tx_ix_tail++;
1883 				spin_unlock(&sqp->tx_lock);
1884 				break;
1885 			case IB_WC_RECV:
1886 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1887 						(sqp->ring[wc.wr_id &
1888 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1889 				grh = &(((struct mlx4_mad_rcv_buf *)
1890 						(sqp->ring[wc.wr_id &
1891 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1892 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1893 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1894 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1895 					pr_err("Failed reposting SQP "
1896 					       "buf:%lld\n", (unsigned long long)wc.wr_id);
1897 				break;
1898 			default:
1899 				BUG_ON(1);
1900 				break;
1901 			}
1902 		} else  {
1903 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1904 				 " status = %d, wrid = 0x%llx\n",
1905 				 ctx->slave, wc.status, (unsigned long long)wc.wr_id);
1906 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1907 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1908 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1909 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1910 					= NULL;
1911 				spin_lock(&sqp->tx_lock);
1912 				sqp->tx_ix_tail++;
1913 				spin_unlock(&sqp->tx_lock);
1914 			}
1915 		}
1916 	}
1917 }
1918 
1919 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1920 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1921 {
1922 	struct mlx4_ib_demux_pv_ctx *ctx;
1923 
1924 	*ret_ctx = NULL;
1925 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1926 	if (!ctx) {
1927 		pr_err("failed allocating pv resource context "
1928 		       "for port %d, slave %d\n", port, slave);
1929 		return -ENOMEM;
1930 	}
1931 
1932 	ctx->ib_dev = &dev->ib_dev;
1933 	ctx->port = port;
1934 	ctx->slave = slave;
1935 	*ret_ctx = ctx;
1936 	return 0;
1937 }
1938 
1939 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1940 {
1941 	if (dev->sriov.demux[port - 1].tun[slave]) {
1942 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1943 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1944 	}
1945 }
1946 
1947 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1948 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1949 {
1950 	int ret, cq_size;
1951 
1952 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1953 		return -EEXIST;
1954 
1955 	ctx->state = DEMUX_PV_STATE_STARTING;
1956 	/* have QP0 only on port owner, and only if link layer is IB */
1957 	if (ctx->slave == mlx4_master_func_num(to_mdev(ctx->ib_dev)->dev) &&
1958 	    rdma_port_get_link_layer(ibdev, ctx->port) == IB_LINK_LAYER_INFINIBAND)
1959 		ctx->has_smi = 1;
1960 
1961 	if (ctx->has_smi) {
1962 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1963 		if (ret) {
1964 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1965 			goto err_out;
1966 		}
1967 	}
1968 
1969 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1970 	if (ret) {
1971 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1972 		goto err_out_qp0;
1973 	}
1974 
1975 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1976 	if (ctx->has_smi)
1977 		cq_size *= 2;
1978 
1979 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1980 			       NULL, ctx, cq_size, 0);
1981 	if (IS_ERR(ctx->cq)) {
1982 		ret = PTR_ERR(ctx->cq);
1983 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1984 		goto err_buf;
1985 	}
1986 
1987 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1988 	if (IS_ERR(ctx->pd)) {
1989 		ret = PTR_ERR(ctx->pd);
1990 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1991 		goto err_cq;
1992 	}
1993 
1994 	ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1995 	if (IS_ERR(ctx->mr)) {
1996 		ret = PTR_ERR(ctx->mr);
1997 		pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1998 		goto err_pd;
1999 	}
2000 
2001 	if (ctx->has_smi) {
2002 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
2003 		if (ret) {
2004 			pr_err("Couldn't create %s QP0 (%d)\n",
2005 			       create_tun ? "tunnel for" : "",  ret);
2006 			goto err_mr;
2007 		}
2008 	}
2009 
2010 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2011 	if (ret) {
2012 		pr_err("Couldn't create %s QP1 (%d)\n",
2013 		       create_tun ? "tunnel for" : "",  ret);
2014 		goto err_qp0;
2015 	}
2016 
2017 	if (create_tun)
2018 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2019 	else
2020 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2021 
2022 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2023 
2024 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2025 	if (ret) {
2026 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2027 		goto err_wq;
2028 	}
2029 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2030 	return 0;
2031 
2032 err_wq:
2033 	ctx->wq = NULL;
2034 	ib_destroy_qp(ctx->qp[1].qp);
2035 	ctx->qp[1].qp = NULL;
2036 
2037 
2038 err_qp0:
2039 	if (ctx->has_smi)
2040 		ib_destroy_qp(ctx->qp[0].qp);
2041 	ctx->qp[0].qp = NULL;
2042 
2043 err_mr:
2044 	ib_dereg_mr(ctx->mr);
2045 	ctx->mr = NULL;
2046 
2047 err_pd:
2048 	ib_dealloc_pd(ctx->pd);
2049 	ctx->pd = NULL;
2050 
2051 err_cq:
2052 	ib_destroy_cq(ctx->cq);
2053 	ctx->cq = NULL;
2054 
2055 err_buf:
2056 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2057 
2058 err_out_qp0:
2059 	if (ctx->has_smi)
2060 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2061 err_out:
2062 	ctx->state = DEMUX_PV_STATE_DOWN;
2063 	return ret;
2064 }
2065 
2066 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2067 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2068 {
2069 	if (!ctx)
2070 		return;
2071 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2072 		ctx->state = DEMUX_PV_STATE_DOWNING;
2073 		if (flush)
2074 			flush_workqueue(ctx->wq);
2075 		if (ctx->has_smi) {
2076 			ib_destroy_qp(ctx->qp[0].qp);
2077 			ctx->qp[0].qp = NULL;
2078 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2079 		}
2080 		ib_destroy_qp(ctx->qp[1].qp);
2081 		ctx->qp[1].qp = NULL;
2082 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2083 		ib_dereg_mr(ctx->mr);
2084 		ctx->mr = NULL;
2085 		ib_dealloc_pd(ctx->pd);
2086 		ctx->pd = NULL;
2087 		ib_destroy_cq(ctx->cq);
2088 		ctx->cq = NULL;
2089 		ctx->state = DEMUX_PV_STATE_DOWN;
2090 	}
2091 }
2092 
2093 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2094 				  int port, int do_init)
2095 {
2096 	int ret = 0;
2097 
2098 	if (!do_init) {
2099 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2100 		/* for master, destroy real sqp resources */
2101 		if (slave == mlx4_master_func_num(dev->dev))
2102 			destroy_pv_resources(dev, slave, port,
2103 					     dev->sriov.sqps[port - 1], 1);
2104 		/* destroy the tunnel qp resources */
2105 		destroy_pv_resources(dev, slave, port,
2106 				     dev->sriov.demux[port - 1].tun[slave], 1);
2107 		return 0;
2108 	}
2109 
2110 	/* create the tunnel qp resources */
2111 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2112 				  dev->sriov.demux[port - 1].tun[slave]);
2113 
2114 	/* for master, create the real sqp resources */
2115 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2116 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2117 					  dev->sriov.sqps[port - 1]);
2118 	return ret;
2119 }
2120 
2121 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2122 {
2123 	struct mlx4_ib_demux_work *dmxw;
2124 
2125 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2126 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2127 			       dmxw->do_init);
2128 	kfree(dmxw);
2129 	return;
2130 }
2131 
2132 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2133 				       struct mlx4_ib_demux_ctx *ctx,
2134 				       int port)
2135 {
2136 	char name[12];
2137 	int ret = 0;
2138 	int i;
2139 
2140 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2141 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2142 	if (!ctx->tun)
2143 		return -ENOMEM;
2144 
2145 	ctx->dev = dev;
2146 	ctx->port = port;
2147 	ctx->ib_dev = &dev->ib_dev;
2148 
2149 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2150 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2151 		if (ret) {
2152 			ret = -ENOMEM;
2153 			goto err_mcg;
2154 		}
2155 	}
2156 
2157 	ret = mlx4_ib_mcg_port_init(ctx);
2158 	if (ret) {
2159 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2160 		goto err_mcg;
2161 	}
2162 
2163 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2164 	ctx->wq = create_singlethread_workqueue(name);
2165 	if (!ctx->wq) {
2166 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2167 		ret = -ENOMEM;
2168 		goto err_wq;
2169 	}
2170 
2171 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2172 	ctx->ud_wq = create_singlethread_workqueue(name);
2173 	if (!ctx->ud_wq) {
2174 		pr_err("Failed to create up/down WQ for port %d\n", port);
2175 		ret = -ENOMEM;
2176 		goto err_udwq;
2177 	}
2178 
2179 	return 0;
2180 
2181 err_udwq:
2182 	destroy_workqueue(ctx->wq);
2183 	ctx->wq = NULL;
2184 
2185 err_wq:
2186 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2187 err_mcg:
2188 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2189 		free_pv_object(dev, i, port);
2190 	kfree(ctx->tun);
2191 	ctx->tun = NULL;
2192 	return ret;
2193 }
2194 
2195 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2196 {
2197 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2198 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2199 		flush_workqueue(sqp_ctx->wq);
2200 		if (sqp_ctx->has_smi) {
2201 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2202 			sqp_ctx->qp[0].qp = NULL;
2203 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2204 		}
2205 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2206 		sqp_ctx->qp[1].qp = NULL;
2207 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2208 		ib_dereg_mr(sqp_ctx->mr);
2209 		sqp_ctx->mr = NULL;
2210 		ib_dealloc_pd(sqp_ctx->pd);
2211 		sqp_ctx->pd = NULL;
2212 		ib_destroy_cq(sqp_ctx->cq);
2213 		sqp_ctx->cq = NULL;
2214 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2215 	}
2216 }
2217 
2218 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2219 {
2220 	int i;
2221 	if (ctx) {
2222 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2223 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2224 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2225 			if (!ctx->tun[i])
2226 				continue;
2227 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2228 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2229 		}
2230 		flush_workqueue(ctx->wq);
2231 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2232 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2233 			free_pv_object(dev, i, ctx->port);
2234 		}
2235 		kfree(ctx->tun);
2236 		destroy_workqueue(ctx->ud_wq);
2237 		destroy_workqueue(ctx->wq);
2238 	}
2239 }
2240 
2241 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2242 {
2243 	int i;
2244 
2245 	if (!mlx4_is_master(dev->dev))
2246 		return;
2247 	/* initialize or tear down tunnel QPs for the master */
2248 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2249 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2250 	return;
2251 }
2252 
2253 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2254 {
2255 	int i = 0;
2256 	int err;
2257 
2258 	if (!mlx4_is_mfunc(dev->dev))
2259 		return 0;
2260 
2261 	dev->sriov.is_going_down = 0;
2262 	spin_lock_init(&dev->sriov.going_down_lock);
2263 	mlx4_ib_cm_paravirt_init(dev);
2264 
2265 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2266 
2267 	if (mlx4_is_slave(dev->dev)) {
2268 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2269 		return 0;
2270 	}
2271 
2272 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2273 		if (i == mlx4_master_func_num(dev->dev))
2274 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2275 		else
2276 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2277 	}
2278 
2279 	err = mlx4_ib_init_alias_guid_service(dev);
2280 	if (err) {
2281 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2282 		goto paravirt_err;
2283 	}
2284 	err = mlx4_ib_device_register_sysfs(dev);
2285 	if (err) {
2286 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2287 		goto sysfs_err;
2288 	}
2289 
2290 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2291 		     dev->dev->caps.sqp_demux);
2292 	for (i = 0; i < dev->num_ports; i++) {
2293 		union ib_gid gid;
2294 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2295 		if (err)
2296 			goto demux_err;
2297 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2298 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2299 				      &dev->sriov.sqps[i]);
2300 		if (err)
2301 			goto demux_err;
2302 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2303 		if (err)
2304 			goto demux_err;
2305 	}
2306 	mlx4_ib_master_tunnels(dev, 1);
2307 	return 0;
2308 
2309 demux_err:
2310 	while (i > 0) {
2311 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2312 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2313 		--i;
2314 	}
2315 	mlx4_ib_device_unregister_sysfs(dev);
2316 
2317 sysfs_err:
2318 	mlx4_ib_destroy_alias_guid_service(dev);
2319 
2320 paravirt_err:
2321 	mlx4_ib_cm_paravirt_clean(dev, -1);
2322 
2323 	return err;
2324 }
2325 
2326 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2327 {
2328 	int i;
2329 	unsigned long flags;
2330 
2331 	if (!mlx4_is_mfunc(dev->dev))
2332 		return;
2333 
2334 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2335 	dev->sriov.is_going_down = 1;
2336 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2337 	if (mlx4_is_master(dev->dev)) {
2338 		for (i = 0; i < dev->num_ports; i++) {
2339 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2340 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2341 			kfree(dev->sriov.sqps[i]);
2342 			dev->sriov.sqps[i] = NULL;
2343 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2344 		}
2345 
2346 		mlx4_ib_cm_paravirt_clean(dev, -1);
2347 		mlx4_ib_destroy_alias_guid_service(dev);
2348 		mlx4_ib_device_unregister_sysfs(dev);
2349 	}
2350 }
2351