xref: /linux/drivers/infiniband/hw/mlx4/mad.c (revision 52c996d3f40b40f87ef9dc80596903309682acc3)
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 #include <linux/ip.h>
43 #include <net/ipv6.h>
44 
45 #include <linux/mlx4/driver.h>
46 #include "mlx4_ib.h"
47 
48 enum {
49 	MLX4_IB_VENDOR_CLASS1 = 0x9,
50 	MLX4_IB_VENDOR_CLASS2 = 0xa
51 };
52 
53 #define MLX4_TUN_SEND_WRID_SHIFT 34
54 #define MLX4_TUN_QPN_SHIFT 32
55 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
56 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
57 
58 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
59 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
60 
61  /* Port mgmt change event handling */
62 
63 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
64 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
65 #define NUM_IDX_IN_PKEY_TBL_BLK 32
66 #define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
67 #define GUID_TBL_BLK_NUM_ENTRIES 8
68 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
69 
70 struct mlx4_mad_rcv_buf {
71 	struct ib_grh grh;
72 	u8 payload[256];
73 } __packed;
74 
75 struct mlx4_mad_snd_buf {
76 	u8 payload[256];
77 } __packed;
78 
79 struct mlx4_tunnel_mad {
80 	struct ib_grh grh;
81 	struct mlx4_ib_tunnel_header hdr;
82 	struct ib_mad mad;
83 } __packed;
84 
85 struct mlx4_rcv_tunnel_mad {
86 	struct mlx4_rcv_tunnel_hdr hdr;
87 	struct ib_grh grh;
88 	struct ib_mad mad;
89 } __packed;
90 
91 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u32 port_num);
92 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u32 port_num);
93 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
94 				int block, u32 change_bitmap);
95 
96 __be64 mlx4_ib_gen_node_guid(void)
97 {
98 #define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
99 	return cpu_to_be64(NODE_GUID_HI | get_random_u32());
100 }
101 
102 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
103 {
104 	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
105 		cpu_to_be64(0xff00000000000000LL);
106 }
107 
108 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
109 		 int port, const struct ib_wc *in_wc,
110 		 const struct ib_grh *in_grh,
111 		 const void *in_mad, void *response_mad)
112 {
113 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
114 	void *inbox;
115 	int err;
116 	u32 in_modifier = port;
117 	u8 op_modifier = 0;
118 
119 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
120 	if (IS_ERR(inmailbox))
121 		return PTR_ERR(inmailbox);
122 	inbox = inmailbox->buf;
123 
124 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
125 	if (IS_ERR(outmailbox)) {
126 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
127 		return PTR_ERR(outmailbox);
128 	}
129 
130 	memcpy(inbox, in_mad, 256);
131 
132 	/*
133 	 * Key check traps can't be generated unless we have in_wc to
134 	 * tell us where to send the trap.
135 	 */
136 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
137 		op_modifier |= 0x1;
138 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
139 		op_modifier |= 0x2;
140 	if (mlx4_is_mfunc(dev->dev) &&
141 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
142 		op_modifier |= 0x8;
143 
144 	if (in_wc) {
145 		struct {
146 			__be32		my_qpn;
147 			u32		reserved1;
148 			__be32		rqpn;
149 			u8		sl;
150 			u8		g_path;
151 			u16		reserved2[2];
152 			__be16		pkey;
153 			u32		reserved3[11];
154 			u8		grh[40];
155 		} *ext_info;
156 
157 		memset(inbox + 256, 0, 256);
158 		ext_info = inbox + 256;
159 
160 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
161 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
162 		ext_info->sl     = in_wc->sl << 4;
163 		ext_info->g_path = in_wc->dlid_path_bits |
164 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
165 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
166 
167 		if (in_grh)
168 			memcpy(ext_info->grh, in_grh, 40);
169 
170 		op_modifier |= 0x4;
171 
172 		in_modifier |= ib_lid_cpu16(in_wc->slid) << 16;
173 	}
174 
175 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
176 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
177 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
178 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
179 
180 	if (!err)
181 		memcpy(response_mad, outmailbox->buf, 256);
182 
183 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
184 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
185 
186 	return err;
187 }
188 
189 static void update_sm_ah(struct mlx4_ib_dev *dev, u32 port_num, u16 lid, u8 sl)
190 {
191 	struct ib_ah *new_ah;
192 	struct rdma_ah_attr ah_attr;
193 	unsigned long flags;
194 
195 	if (!dev->send_agent[port_num - 1][0])
196 		return;
197 
198 	memset(&ah_attr, 0, sizeof ah_attr);
199 	ah_attr.type = rdma_ah_find_type(&dev->ib_dev, port_num);
200 	rdma_ah_set_dlid(&ah_attr, lid);
201 	rdma_ah_set_sl(&ah_attr, sl);
202 	rdma_ah_set_port_num(&ah_attr, port_num);
203 
204 	new_ah = rdma_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
205 				&ah_attr, 0);
206 	if (IS_ERR(new_ah))
207 		return;
208 
209 	spin_lock_irqsave(&dev->sm_lock, flags);
210 	if (dev->sm_ah[port_num - 1])
211 		rdma_destroy_ah(dev->sm_ah[port_num - 1], 0);
212 	dev->sm_ah[port_num - 1] = new_ah;
213 	spin_unlock_irqrestore(&dev->sm_lock, flags);
214 }
215 
216 /*
217  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
218  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
219  */
220 static void smp_snoop(struct ib_device *ibdev, u32 port_num,
221 		      const struct ib_mad *mad, u16 prev_lid)
222 {
223 	struct ib_port_info *pinfo;
224 	u16 lid;
225 	__be16 *base;
226 	u32 bn, pkey_change_bitmap;
227 	int i;
228 
229 
230 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
231 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
232 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
233 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
234 		switch (mad->mad_hdr.attr_id) {
235 		case IB_SMP_ATTR_PORT_INFO:
236 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
237 				return;
238 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
239 			lid = be16_to_cpu(pinfo->lid);
240 
241 			update_sm_ah(dev, port_num,
242 				     be16_to_cpu(pinfo->sm_lid),
243 				     pinfo->neighbormtu_mastersmsl & 0xf);
244 
245 			if (pinfo->clientrereg_resv_subnetto & 0x80)
246 				handle_client_rereg_event(dev, port_num);
247 
248 			if (prev_lid != lid)
249 				handle_lid_change_event(dev, port_num);
250 			break;
251 
252 		case IB_SMP_ATTR_PKEY_TABLE:
253 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
254 				return;
255 			if (!mlx4_is_mfunc(dev->dev)) {
256 				mlx4_ib_dispatch_event(dev, port_num,
257 						       IB_EVENT_PKEY_CHANGE);
258 				break;
259 			}
260 
261 			/* at this point, we are running in the master.
262 			 * Slaves do not receive SMPs.
263 			 */
264 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
265 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
266 			pkey_change_bitmap = 0;
267 			for (i = 0; i < 32; i++) {
268 				pr_debug("PKEY[%d] = x%x\n",
269 					 i + bn*32, be16_to_cpu(base[i]));
270 				if (be16_to_cpu(base[i]) !=
271 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
272 					pkey_change_bitmap |= (1 << i);
273 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
274 						be16_to_cpu(base[i]);
275 				}
276 			}
277 			pr_debug("PKEY Change event: port=%u, "
278 				 "block=0x%x, change_bitmap=0x%x\n",
279 				 port_num, bn, pkey_change_bitmap);
280 
281 			if (pkey_change_bitmap) {
282 				mlx4_ib_dispatch_event(dev, port_num,
283 						       IB_EVENT_PKEY_CHANGE);
284 				if (!dev->sriov.is_going_down)
285 					__propagate_pkey_ev(dev, port_num, bn,
286 							    pkey_change_bitmap);
287 			}
288 			break;
289 
290 		case IB_SMP_ATTR_GUID_INFO:
291 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
292 				return;
293 			/* paravirtualized master's guid is guid 0 -- does not change */
294 			if (!mlx4_is_master(dev->dev))
295 				mlx4_ib_dispatch_event(dev, port_num,
296 						       IB_EVENT_GID_CHANGE);
297 			/*if master, notify relevant slaves*/
298 			if (mlx4_is_master(dev->dev) &&
299 			    !dev->sriov.is_going_down) {
300 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
301 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
302 								    (u8 *)(&((struct ib_smp *)mad)->data));
303 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
304 								     (u8 *)(&((struct ib_smp *)mad)->data));
305 			}
306 			break;
307 
308 		case IB_SMP_ATTR_SL_TO_VL_TABLE:
309 			/* cache sl to vl mapping changes for use in
310 			 * filling QP1 LRH VL field when sending packets
311 			 */
312 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV &&
313 			    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)
314 				return;
315 			if (!mlx4_is_slave(dev->dev)) {
316 				union sl2vl_tbl_to_u64 sl2vl64;
317 				int jj;
318 
319 				for (jj = 0; jj < 8; jj++) {
320 					sl2vl64.sl8[jj] = ((struct ib_smp *)mad)->data[jj];
321 					pr_debug("port %u, sl2vl[%d] = %02x\n",
322 						 port_num, jj, sl2vl64.sl8[jj]);
323 				}
324 				atomic64_set(&dev->sl2vl[port_num - 1], sl2vl64.sl64);
325 			}
326 			break;
327 
328 		default:
329 			break;
330 		}
331 }
332 
333 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
334 				int block, u32 change_bitmap)
335 {
336 	int i, ix, slave, err;
337 	int have_event = 0;
338 
339 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
340 		if (slave == mlx4_master_func_num(dev->dev))
341 			continue;
342 		if (!mlx4_is_slave_active(dev->dev, slave))
343 			continue;
344 
345 		have_event = 0;
346 		for (i = 0; i < 32; i++) {
347 			if (!(change_bitmap & (1 << i)))
348 				continue;
349 			for (ix = 0;
350 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
351 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
352 				    [ix] == i + 32 * block) {
353 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
354 					pr_debug("propagate_pkey_ev: slave %d,"
355 						 " port %d, ix %d (%d)\n",
356 						 slave, port_num, ix, err);
357 					have_event = 1;
358 					break;
359 				}
360 			}
361 			if (have_event)
362 				break;
363 		}
364 	}
365 }
366 
367 static void node_desc_override(struct ib_device *dev,
368 			       struct ib_mad *mad)
369 {
370 	unsigned long flags;
371 
372 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
373 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
374 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
375 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
376 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
377 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
378 		       IB_DEVICE_NODE_DESC_MAX);
379 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
380 	}
381 }
382 
383 static void forward_trap(struct mlx4_ib_dev *dev, u32 port_num,
384 			 const struct ib_mad *mad)
385 {
386 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
387 	struct ib_mad_send_buf *send_buf;
388 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
389 	int ret;
390 	unsigned long flags;
391 
392 	if (agent) {
393 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
394 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
395 					      IB_MGMT_BASE_VERSION);
396 		if (IS_ERR(send_buf))
397 			return;
398 		/*
399 		 * We rely here on the fact that MLX QPs don't use the
400 		 * address handle after the send is posted (this is
401 		 * wrong following the IB spec strictly, but we know
402 		 * it's OK for our devices).
403 		 */
404 		spin_lock_irqsave(&dev->sm_lock, flags);
405 		memcpy(send_buf->mad, mad, sizeof *mad);
406 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
407 			ret = ib_post_send_mad(send_buf, NULL);
408 		else
409 			ret = -EINVAL;
410 		spin_unlock_irqrestore(&dev->sm_lock, flags);
411 
412 		if (ret)
413 			ib_free_send_mad(send_buf);
414 	}
415 }
416 
417 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
418 							     struct ib_sa_mad *sa_mad)
419 {
420 	int ret = 0;
421 
422 	/* dispatch to different sa handlers */
423 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
424 	case IB_SA_ATTR_MC_MEMBER_REC:
425 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
426 		break;
427 	default:
428 		break;
429 	}
430 	return ret;
431 }
432 
433 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u32 port, __be64 guid)
434 {
435 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
436 	int i;
437 
438 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
439 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
440 			return i;
441 	}
442 	return -1;
443 }
444 
445 
446 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
447 				   u32 port, u16 pkey, u16 *ix)
448 {
449 	int i, ret;
450 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
451 	u16 slot_pkey;
452 
453 	if (slave == mlx4_master_func_num(dev->dev))
454 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
455 
456 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
457 
458 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
459 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
460 			continue;
461 
462 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
463 
464 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
465 		if (ret)
466 			continue;
467 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
468 			if (slot_pkey & 0x8000) {
469 				*ix = (u16) pkey_ix;
470 				return 0;
471 			} else {
472 				/* take first partial pkey index found */
473 				if (partial_ix == 0xFF)
474 					partial_ix = pkey_ix;
475 			}
476 		}
477 	}
478 
479 	if (partial_ix < 0xFF) {
480 		*ix = (u16) partial_ix;
481 		return 0;
482 	}
483 
484 	return -EINVAL;
485 }
486 
487 static int get_gids_from_l3_hdr(struct ib_grh *grh, union ib_gid *sgid,
488 				union ib_gid *dgid)
489 {
490 	int version = ib_get_rdma_header_version((const union rdma_network_hdr *)grh);
491 	enum rdma_network_type net_type;
492 
493 	if (version == 4)
494 		net_type = RDMA_NETWORK_IPV4;
495 	else if (version == 6)
496 		net_type = RDMA_NETWORK_IPV6;
497 	else
498 		return -EINVAL;
499 
500 	return ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
501 					 sgid, dgid);
502 }
503 
504 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
505 {
506 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
507 
508 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
509 }
510 
511 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u32 port,
512 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
513 			  struct ib_grh *grh, struct ib_mad *mad)
514 {
515 	struct ib_sge list;
516 	struct ib_ud_wr wr;
517 	const struct ib_send_wr *bad_wr;
518 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
519 	struct mlx4_ib_demux_pv_qp *tun_qp;
520 	struct mlx4_rcv_tunnel_mad *tun_mad;
521 	struct rdma_ah_attr attr;
522 	struct ib_ah *ah;
523 	struct ib_qp *src_qp = NULL;
524 	unsigned tun_tx_ix = 0;
525 	int dqpn;
526 	int ret = 0;
527 	u16 tun_pkey_ix;
528 	u16 cached_pkey;
529 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
530 
531 	if (dest_qpt > IB_QPT_GSI) {
532 		pr_debug("dest_qpt (%d) > IB_QPT_GSI\n", dest_qpt);
533 		return -EINVAL;
534 	}
535 
536 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
537 
538 	/* check if proxy qp created */
539 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
540 		return -EAGAIN;
541 
542 	if (!dest_qpt)
543 		tun_qp = &tun_ctx->qp[0];
544 	else
545 		tun_qp = &tun_ctx->qp[1];
546 
547 	/* compute P_Key index to put in tunnel header for slave */
548 	if (dest_qpt) {
549 		u16 pkey_ix;
550 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
551 		if (ret) {
552 			pr_debug("unable to get %s cached pkey for index %d, ret %d\n",
553 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
554 				 wc->pkey_index, ret);
555 			return -EINVAL;
556 		}
557 
558 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
559 		if (ret) {
560 			pr_debug("unable to get %s pkey ix for pkey 0x%x, ret %d\n",
561 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
562 				 cached_pkey, ret);
563 			return -EINVAL;
564 		}
565 		tun_pkey_ix = pkey_ix;
566 	} else
567 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
568 
569 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
570 
571 	/* get tunnel tx data buf for slave */
572 	src_qp = tun_qp->qp;
573 
574 	/* create ah. Just need an empty one with the port num for the post send.
575 	 * The driver will set the force loopback bit in post_send */
576 	memset(&attr, 0, sizeof attr);
577 	attr.type = rdma_ah_find_type(&dev->ib_dev, port);
578 
579 	rdma_ah_set_port_num(&attr, port);
580 	if (is_eth) {
581 		union ib_gid sgid;
582 		union ib_gid dgid;
583 
584 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
585 			return -EINVAL;
586 		rdma_ah_set_grh(&attr, &dgid, 0, 0, 0, 0);
587 	}
588 	ah = rdma_create_ah(tun_ctx->pd, &attr, 0);
589 	if (IS_ERR(ah))
590 		return -ENOMEM;
591 
592 	/* allocate tunnel tx buf after pass failure returns */
593 	spin_lock(&tun_qp->tx_lock);
594 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
595 	    (MLX4_NUM_TUNNEL_BUFS - 1))
596 		ret = -EAGAIN;
597 	else
598 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
599 	spin_unlock(&tun_qp->tx_lock);
600 	if (ret)
601 		goto end;
602 
603 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
604 	if (tun_qp->tx_ring[tun_tx_ix].ah)
605 		rdma_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah, 0);
606 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
607 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
608 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
609 				   sizeof (struct mlx4_rcv_tunnel_mad),
610 				   DMA_TO_DEVICE);
611 
612 	/* copy over to tunnel buffer */
613 	if (grh)
614 		memcpy(&tun_mad->grh, grh, sizeof *grh);
615 	memcpy(&tun_mad->mad, mad, sizeof *mad);
616 
617 	/* adjust tunnel data */
618 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
619 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
620 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
621 
622 	if (is_eth) {
623 		u16 vlan = 0;
624 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
625 						NULL)) {
626 			/* VST mode */
627 			if (vlan != wc->vlan_id)
628 				/* Packet vlan is not the VST-assigned vlan.
629 				 * Drop the packet.
630 				 */
631 				goto out;
632 			 else
633 				/* Remove the vlan tag before forwarding
634 				 * the packet to the VF.
635 				 */
636 				vlan = 0xffff;
637 		} else {
638 			vlan = wc->vlan_id;
639 		}
640 
641 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
642 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
643 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
644 	} else {
645 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
646 		tun_mad->hdr.slid_mac_47_32 = ib_lid_be16(wc->slid);
647 	}
648 
649 	ib_dma_sync_single_for_device(&dev->ib_dev,
650 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
651 				      sizeof (struct mlx4_rcv_tunnel_mad),
652 				      DMA_TO_DEVICE);
653 
654 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
655 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
656 	list.lkey = tun_ctx->pd->local_dma_lkey;
657 
658 	wr.ah = ah;
659 	wr.port_num = port;
660 	wr.remote_qkey = IB_QP_SET_QKEY;
661 	wr.remote_qpn = dqpn;
662 	wr.wr.next = NULL;
663 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
664 	wr.wr.sg_list = &list;
665 	wr.wr.num_sge = 1;
666 	wr.wr.opcode = IB_WR_SEND;
667 	wr.wr.send_flags = IB_SEND_SIGNALED;
668 
669 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
670 	if (!ret)
671 		return 0;
672  out:
673 	spin_lock(&tun_qp->tx_lock);
674 	tun_qp->tx_ix_tail++;
675 	spin_unlock(&tun_qp->tx_lock);
676 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
677 end:
678 	rdma_destroy_ah(ah, 0);
679 	return ret;
680 }
681 
682 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u32 port,
683 			struct ib_wc *wc, struct ib_grh *grh,
684 			struct ib_mad *mad)
685 {
686 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
687 	int err, other_port;
688 	int slave = -1;
689 	u8 *slave_id;
690 	int is_eth = 0;
691 
692 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
693 		is_eth = 0;
694 	else
695 		is_eth = 1;
696 
697 	if (is_eth) {
698 		union ib_gid dgid;
699 		union ib_gid sgid;
700 
701 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
702 			return -EINVAL;
703 		if (!(wc->wc_flags & IB_WC_GRH)) {
704 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
705 			return -EINVAL;
706 		}
707 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
708 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
709 			return -EINVAL;
710 		}
711 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, dgid.raw, &slave);
712 		if (err && mlx4_is_mf_bonded(dev->dev)) {
713 			other_port = (port == 1) ? 2 : 1;
714 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, dgid.raw, &slave);
715 			if (!err) {
716 				port = other_port;
717 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
718 					 slave, grh->dgid.raw, port, other_port);
719 			}
720 		}
721 		if (err) {
722 			mlx4_ib_warn(ibdev, "failed matching grh\n");
723 			return -ENOENT;
724 		}
725 		if (slave >= dev->dev->caps.sqp_demux) {
726 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
727 				     slave, dev->dev->caps.sqp_demux);
728 			return -ENOENT;
729 		}
730 
731 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
732 			return 0;
733 
734 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
735 		if (err)
736 			pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
737 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
738 				 slave, err);
739 		return 0;
740 	}
741 
742 	/* Initially assume that this mad is for us */
743 	slave = mlx4_master_func_num(dev->dev);
744 
745 	/* See if the slave id is encoded in a response mad */
746 	if (mad->mad_hdr.method & 0x80) {
747 		slave_id = (u8 *) &mad->mad_hdr.tid;
748 		slave = *slave_id;
749 		if (slave != 255) /*255 indicates the dom0*/
750 			*slave_id = 0; /* remap tid */
751 	}
752 
753 	/* If a grh is present, we demux according to it */
754 	if (wc->wc_flags & IB_WC_GRH) {
755 		if (grh->dgid.global.interface_id ==
756 			cpu_to_be64(IB_SA_WELL_KNOWN_GUID) &&
757 		    grh->dgid.global.subnet_prefix == cpu_to_be64(
758 			atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix))) {
759 			slave = 0;
760 		} else {
761 			slave = mlx4_ib_find_real_gid(ibdev, port,
762 						      grh->dgid.global.interface_id);
763 			if (slave < 0) {
764 				mlx4_ib_warn(ibdev, "failed matching grh\n");
765 				return -ENOENT;
766 			}
767 		}
768 	}
769 	/* Class-specific handling */
770 	switch (mad->mad_hdr.mgmt_class) {
771 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
772 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
773 		/* 255 indicates the dom0 */
774 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
775 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
776 				return -EPERM;
777 			/* for a VF. drop unsolicited MADs */
778 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
779 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
780 					     slave, mad->mad_hdr.mgmt_class,
781 					     mad->mad_hdr.method);
782 				return -EINVAL;
783 			}
784 		}
785 		break;
786 	case IB_MGMT_CLASS_SUBN_ADM:
787 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
788 					     (struct ib_sa_mad *) mad))
789 			return 0;
790 		break;
791 	case IB_MGMT_CLASS_CM:
792 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
793 			return 0;
794 		break;
795 	case IB_MGMT_CLASS_DEVICE_MGMT:
796 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
797 			return 0;
798 		break;
799 	default:
800 		/* Drop unsupported classes for slaves in tunnel mode */
801 		if (slave != mlx4_master_func_num(dev->dev)) {
802 			pr_debug("dropping unsupported ingress mad from class:%d "
803 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
804 			return 0;
805 		}
806 	}
807 	/*make sure that no slave==255 was not handled yet.*/
808 	if (slave >= dev->dev->caps.sqp_demux) {
809 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
810 			     slave, dev->dev->caps.sqp_demux);
811 		return -ENOENT;
812 	}
813 
814 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
815 	if (err)
816 		pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
817 			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
818 			 slave, err);
819 	return 0;
820 }
821 
822 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
823 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
824 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
825 {
826 	u16 slid, prev_lid = 0;
827 	int err;
828 	struct ib_port_attr pattr;
829 
830 	slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
831 
832 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
833 		forward_trap(to_mdev(ibdev), port_num, in_mad);
834 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
835 	}
836 
837 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
838 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
839 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
840 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
841 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
842 			return IB_MAD_RESULT_SUCCESS;
843 
844 		/*
845 		 * Don't process SMInfo queries -- the SMA can't handle them.
846 		 */
847 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
848 			return IB_MAD_RESULT_SUCCESS;
849 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
850 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
851 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
852 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
853 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
854 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
855 			return IB_MAD_RESULT_SUCCESS;
856 	} else
857 		return IB_MAD_RESULT_SUCCESS;
858 
859 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
860 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
861 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
862 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
863 	    !ib_query_port(ibdev, port_num, &pattr))
864 		prev_lid = ib_lid_cpu16(pattr.lid);
865 
866 	err = mlx4_MAD_IFC(to_mdev(ibdev),
867 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
868 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
869 			   MLX4_MAD_IFC_NET_VIEW,
870 			   port_num, in_wc, in_grh, in_mad, out_mad);
871 	if (err)
872 		return IB_MAD_RESULT_FAILURE;
873 
874 	if (!out_mad->mad_hdr.status) {
875 		smp_snoop(ibdev, port_num, in_mad, prev_lid);
876 		/* slaves get node desc from FW */
877 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
878 			node_desc_override(ibdev, out_mad);
879 	}
880 
881 	/* set return bit in status of directed route responses */
882 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
883 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
884 
885 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
886 		/* no response for trap repress */
887 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
888 
889 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
890 }
891 
892 static void edit_counter(struct mlx4_counter *cnt, void *counters,
893 			 __be16 attr_id)
894 {
895 	switch (attr_id) {
896 	case IB_PMA_PORT_COUNTERS:
897 	{
898 		struct ib_pma_portcounters *pma_cnt =
899 			(struct ib_pma_portcounters *)counters;
900 
901 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
902 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
903 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
904 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
905 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
906 				     be64_to_cpu(cnt->tx_frames));
907 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
908 				     be64_to_cpu(cnt->rx_frames));
909 		break;
910 	}
911 	case IB_PMA_PORT_COUNTERS_EXT:
912 	{
913 		struct ib_pma_portcounters_ext *pma_cnt_ext =
914 			(struct ib_pma_portcounters_ext *)counters;
915 
916 		pma_cnt_ext->port_xmit_data =
917 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
918 		pma_cnt_ext->port_rcv_data =
919 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
920 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
921 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
922 		break;
923 	}
924 	}
925 }
926 
927 static int iboe_process_mad_port_info(void *out_mad)
928 {
929 	struct ib_class_port_info cpi = {};
930 
931 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
932 	memcpy(out_mad, &cpi, sizeof(cpi));
933 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
934 }
935 
936 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags,
937 			    u32 port_num, const struct ib_wc *in_wc,
938 			    const struct ib_grh *in_grh,
939 			    const struct ib_mad *in_mad, struct ib_mad *out_mad)
940 {
941 	struct mlx4_counter counter_stats;
942 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
943 	struct counter_index *tmp_counter;
944 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
945 
946 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
947 		return -EINVAL;
948 
949 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
950 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
951 
952 	memset(&counter_stats, 0, sizeof(counter_stats));
953 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
954 	list_for_each_entry(tmp_counter,
955 			    &dev->counters_table[port_num - 1].counters_list,
956 			    list) {
957 		err = mlx4_get_counter_stats(dev->dev,
958 					     tmp_counter->index,
959 					     &counter_stats, 0);
960 		if (err) {
961 			err = IB_MAD_RESULT_FAILURE;
962 			stats_avail = 0;
963 			break;
964 		}
965 		stats_avail = 1;
966 	}
967 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
968 	if (stats_avail) {
969 		switch (counter_stats.counter_mode & 0xf) {
970 		case 0:
971 			edit_counter(&counter_stats,
972 				     (void *)(out_mad->data + 40),
973 				     in_mad->mad_hdr.attr_id);
974 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
975 			break;
976 		default:
977 			err = IB_MAD_RESULT_FAILURE;
978 		}
979 	}
980 
981 	return err;
982 }
983 
984 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u32 port_num,
985 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
986 			const struct ib_mad *in, struct ib_mad *out,
987 			size_t *out_mad_size, u16 *out_mad_pkey_index)
988 {
989 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
990 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
991 
992 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
993 	 * queries, should be called only by VFs and for that specific purpose
994 	 */
995 	if (link == IB_LINK_LAYER_INFINIBAND) {
996 		if (mlx4_is_slave(dev->dev) &&
997 		    (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
998 		     (in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
999 		      in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
1000 		      in->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
1001 			return iboe_process_mad(ibdev, mad_flags, port_num,
1002 						in_wc, in_grh, in, out);
1003 
1004 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc, in_grh,
1005 				      in, out);
1006 	}
1007 
1008 	if (link == IB_LINK_LAYER_ETHERNET)
1009 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1010 					in_grh, in, out);
1011 
1012 	return -EINVAL;
1013 }
1014 
1015 static void send_handler(struct ib_mad_agent *agent,
1016 			 struct ib_mad_send_wc *mad_send_wc)
1017 {
1018 	if (mad_send_wc->send_buf->context[0])
1019 		rdma_destroy_ah(mad_send_wc->send_buf->context[0], 0);
1020 	ib_free_send_mad(mad_send_wc->send_buf);
1021 }
1022 
1023 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1024 {
1025 	struct ib_mad_agent *agent;
1026 	int p, q;
1027 	int ret;
1028 	enum rdma_link_layer ll;
1029 
1030 	for (p = 0; p < dev->num_ports; ++p) {
1031 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1032 		for (q = 0; q <= 1; ++q) {
1033 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1034 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1035 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1036 							      NULL, 0, send_handler,
1037 							      NULL, NULL, 0);
1038 				if (IS_ERR(agent)) {
1039 					ret = PTR_ERR(agent);
1040 					goto err;
1041 				}
1042 				dev->send_agent[p][q] = agent;
1043 			} else
1044 				dev->send_agent[p][q] = NULL;
1045 		}
1046 	}
1047 
1048 	return 0;
1049 
1050 err:
1051 	for (p = 0; p < dev->num_ports; ++p)
1052 		for (q = 0; q <= 1; ++q)
1053 			if (dev->send_agent[p][q])
1054 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1055 
1056 	return ret;
1057 }
1058 
1059 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1060 {
1061 	struct ib_mad_agent *agent;
1062 	int p, q;
1063 
1064 	for (p = 0; p < dev->num_ports; ++p) {
1065 		for (q = 0; q <= 1; ++q) {
1066 			agent = dev->send_agent[p][q];
1067 			if (agent) {
1068 				dev->send_agent[p][q] = NULL;
1069 				ib_unregister_mad_agent(agent);
1070 			}
1071 		}
1072 
1073 		if (dev->sm_ah[p])
1074 			rdma_destroy_ah(dev->sm_ah[p], 0);
1075 	}
1076 }
1077 
1078 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u32 port_num)
1079 {
1080 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1081 
1082 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1083 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1084 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1085 }
1086 
1087 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u32 port_num)
1088 {
1089 	/* re-configure the alias-guid and mcg's */
1090 	if (mlx4_is_master(dev->dev)) {
1091 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1092 
1093 		if (!dev->sriov.is_going_down) {
1094 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1095 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1096 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1097 		}
1098 	}
1099 
1100 	/* Update the sl to vl table from inside client rereg
1101 	 * only if in secure-host mode (snooping is not possible)
1102 	 * and the sl-to-vl change event is not generated by FW.
1103 	 */
1104 	if (!mlx4_is_slave(dev->dev) &&
1105 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1106 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1107 		if (mlx4_is_master(dev->dev))
1108 			/* already in work queue from mlx4_ib_event queueing
1109 			 * mlx4_handle_port_mgmt_change_event, which calls
1110 			 * this procedure. Therefore, call sl2vl_update directly.
1111 			 */
1112 			mlx4_ib_sl2vl_update(dev, port_num);
1113 		else
1114 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1115 	}
1116 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1117 }
1118 
1119 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1120 			      struct mlx4_eqe *eqe)
1121 {
1122 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1123 			    GET_MASK_FROM_EQE(eqe));
1124 }
1125 
1126 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u32 port_num,
1127 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1128 {
1129 	struct ib_smp *in_mad  = NULL;
1130 	struct ib_smp *out_mad  = NULL;
1131 	u16 i;
1132 
1133 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1134 		return;
1135 
1136 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1137 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1138 	if (!in_mad || !out_mad)
1139 		goto out;
1140 
1141 	guid_tbl_blk_num  *= 4;
1142 
1143 	for (i = 0; i < 4; i++) {
1144 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1145 			continue;
1146 		memset(in_mad, 0, sizeof *in_mad);
1147 		memset(out_mad, 0, sizeof *out_mad);
1148 
1149 		in_mad->base_version  = 1;
1150 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1151 		in_mad->class_version = 1;
1152 		in_mad->method        = IB_MGMT_METHOD_GET;
1153 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1154 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1155 
1156 		if (mlx4_MAD_IFC(dev,
1157 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1158 				 port_num, NULL, NULL, in_mad, out_mad)) {
1159 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1160 			goto out;
1161 		}
1162 
1163 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1164 						    port_num,
1165 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1166 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1167 						     port_num,
1168 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1169 	}
1170 
1171 out:
1172 	kfree(in_mad);
1173 	kfree(out_mad);
1174 	return;
1175 }
1176 
1177 void handle_port_mgmt_change_event(struct work_struct *work)
1178 {
1179 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1180 	struct mlx4_ib_dev *dev = ew->ib_dev;
1181 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1182 	u32 port = eqe->event.port_mgmt_change.port;
1183 	u32 changed_attr;
1184 	u32 tbl_block;
1185 	u32 change_bitmap;
1186 
1187 	switch (eqe->subtype) {
1188 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1189 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1190 
1191 		/* Update the SM ah - This should be done before handling
1192 		   the other changed attributes so that MADs can be sent to the SM */
1193 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1194 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1195 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1196 			update_sm_ah(dev, port, lid, sl);
1197 		}
1198 
1199 		/* Check if it is a lid change event */
1200 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1201 			handle_lid_change_event(dev, port);
1202 
1203 		/* Generate GUID changed event */
1204 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1205 			if (mlx4_is_master(dev->dev)) {
1206 				union ib_gid gid;
1207 				int err = 0;
1208 
1209 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1210 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1211 				else
1212 					gid.global.subnet_prefix =
1213 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1214 				if (err) {
1215 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1216 						port, err);
1217 				} else {
1218 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1219 						 port,
1220 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1221 						 be64_to_cpu(gid.global.subnet_prefix));
1222 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1223 						     be64_to_cpu(gid.global.subnet_prefix));
1224 				}
1225 			}
1226 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1227 			/*if master, notify all slaves*/
1228 			if (mlx4_is_master(dev->dev))
1229 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1230 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1231 		}
1232 
1233 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1234 			handle_client_rereg_event(dev, port);
1235 		break;
1236 
1237 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1238 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1239 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1240 			propagate_pkey_ev(dev, port, eqe);
1241 		break;
1242 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1243 		/* paravirtualized master's guid is guid 0 -- does not change */
1244 		if (!mlx4_is_master(dev->dev))
1245 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1246 		/*if master, notify relevant slaves*/
1247 		else if (!dev->sriov.is_going_down) {
1248 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1249 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1250 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1251 		}
1252 		break;
1253 
1254 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1255 		/* cache sl to vl mapping changes for use in
1256 		 * filling QP1 LRH VL field when sending packets
1257 		 */
1258 		if (!mlx4_is_slave(dev->dev)) {
1259 			union sl2vl_tbl_to_u64 sl2vl64;
1260 			int jj;
1261 
1262 			for (jj = 0; jj < 8; jj++) {
1263 				sl2vl64.sl8[jj] =
1264 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1265 				pr_debug("port %u, sl2vl[%d] = %02x\n",
1266 					 port, jj, sl2vl64.sl8[jj]);
1267 			}
1268 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1269 		}
1270 		break;
1271 	default:
1272 		pr_warn("Unsupported subtype 0x%x for "
1273 			"Port Management Change event\n", eqe->subtype);
1274 	}
1275 
1276 	kfree(ew);
1277 }
1278 
1279 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u32 port_num,
1280 			    enum ib_event_type type)
1281 {
1282 	struct ib_event event;
1283 
1284 	event.device		= &dev->ib_dev;
1285 	event.element.port_num	= port_num;
1286 	event.event		= type;
1287 
1288 	ib_dispatch_event(&event);
1289 }
1290 
1291 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1292 {
1293 	unsigned long flags;
1294 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1295 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1296 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1297 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1298 		queue_work(ctx->wq, &ctx->work);
1299 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1300 }
1301 
1302 static void mlx4_ib_wire_comp_handler(struct ib_cq *cq, void *arg)
1303 {
1304 	unsigned long flags;
1305 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1306 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1307 
1308 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1309 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1310 		queue_work(ctx->wi_wq, &ctx->work);
1311 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1312 }
1313 
1314 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1315 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1316 				  int index)
1317 {
1318 	struct ib_sge sg_list;
1319 	struct ib_recv_wr recv_wr;
1320 	const struct ib_recv_wr *bad_recv_wr;
1321 	int size;
1322 
1323 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1324 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1325 
1326 	sg_list.addr = tun_qp->ring[index].map;
1327 	sg_list.length = size;
1328 	sg_list.lkey = ctx->pd->local_dma_lkey;
1329 
1330 	recv_wr.next = NULL;
1331 	recv_wr.sg_list = &sg_list;
1332 	recv_wr.num_sge = 1;
1333 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1334 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1335 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1336 				      size, DMA_FROM_DEVICE);
1337 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1338 }
1339 
1340 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1341 		int slave, struct ib_sa_mad *sa_mad)
1342 {
1343 	int ret = 0;
1344 
1345 	/* dispatch to different sa handlers */
1346 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1347 	case IB_SA_ATTR_MC_MEMBER_REC:
1348 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1349 		break;
1350 	default:
1351 		break;
1352 	}
1353 	return ret;
1354 }
1355 
1356 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u32 port,
1357 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1358 			 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
1359 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1360 {
1361 	struct ib_sge list;
1362 	struct ib_ud_wr wr;
1363 	const struct ib_send_wr *bad_wr;
1364 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1365 	struct mlx4_ib_demux_pv_qp *sqp;
1366 	struct mlx4_mad_snd_buf *sqp_mad;
1367 	struct ib_ah *ah;
1368 	struct ib_qp *send_qp = NULL;
1369 	unsigned wire_tx_ix = 0;
1370 	u16 wire_pkey_ix;
1371 	int src_qpnum;
1372 	int ret;
1373 
1374 	sqp_ctx = dev->sriov.sqps[port-1];
1375 
1376 	/* check if proxy qp created */
1377 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1378 		return -EAGAIN;
1379 
1380 	if (dest_qpt == IB_QPT_SMI) {
1381 		src_qpnum = 0;
1382 		sqp = &sqp_ctx->qp[0];
1383 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1384 	} else {
1385 		src_qpnum = 1;
1386 		sqp = &sqp_ctx->qp[1];
1387 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1388 	}
1389 
1390 	send_qp = sqp->qp;
1391 
1392 	ah = rdma_zalloc_drv_obj(sqp_ctx->pd->device, ib_ah);
1393 	if (!ah)
1394 		return -ENOMEM;
1395 
1396 	ah->device = sqp_ctx->pd->device;
1397 	ah->pd = sqp_ctx->pd;
1398 
1399 	/* create ah */
1400 	ret = mlx4_ib_create_ah_slave(ah, attr,
1401 				      rdma_ah_retrieve_grh(attr)->sgid_index,
1402 				      s_mac, vlan_id);
1403 	if (ret)
1404 		goto out;
1405 
1406 	spin_lock(&sqp->tx_lock);
1407 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1408 	    (MLX4_NUM_WIRE_BUFS - 1))
1409 		ret = -EAGAIN;
1410 	else
1411 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_WIRE_BUFS - 1);
1412 	spin_unlock(&sqp->tx_lock);
1413 	if (ret)
1414 		goto out;
1415 
1416 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1417 	kfree(sqp->tx_ring[wire_tx_ix].ah);
1418 	sqp->tx_ring[wire_tx_ix].ah = ah;
1419 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1420 				   sqp->tx_ring[wire_tx_ix].buf.map,
1421 				   sizeof (struct mlx4_mad_snd_buf),
1422 				   DMA_TO_DEVICE);
1423 
1424 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1425 
1426 	ib_dma_sync_single_for_device(&dev->ib_dev,
1427 				      sqp->tx_ring[wire_tx_ix].buf.map,
1428 				      sizeof (struct mlx4_mad_snd_buf),
1429 				      DMA_TO_DEVICE);
1430 
1431 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1432 	list.length = sizeof (struct mlx4_mad_snd_buf);
1433 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1434 
1435 	wr.ah = ah;
1436 	wr.port_num = port;
1437 	wr.pkey_index = wire_pkey_ix;
1438 	wr.remote_qkey = qkey;
1439 	wr.remote_qpn = remote_qpn;
1440 	wr.wr.next = NULL;
1441 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1442 	wr.wr.sg_list = &list;
1443 	wr.wr.num_sge = 1;
1444 	wr.wr.opcode = IB_WR_SEND;
1445 	wr.wr.send_flags = IB_SEND_SIGNALED;
1446 
1447 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1448 	if (!ret)
1449 		return 0;
1450 
1451 	spin_lock(&sqp->tx_lock);
1452 	sqp->tx_ix_tail++;
1453 	spin_unlock(&sqp->tx_lock);
1454 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1455 out:
1456 	kfree(ah);
1457 	return ret;
1458 }
1459 
1460 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1461 {
1462 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1463 		return slave;
1464 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1465 }
1466 
1467 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1468 				    struct rdma_ah_attr *ah_attr)
1469 {
1470 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
1471 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1472 		grh->sgid_index = slave;
1473 	else
1474 		grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
1475 }
1476 
1477 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1478 {
1479 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1480 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1481 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1482 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1483 	struct mlx4_ib_ah ah;
1484 	struct rdma_ah_attr ah_attr;
1485 	u8 *slave_id;
1486 	int slave;
1487 	int port;
1488 	u16 vlan_id;
1489 	u8 qos;
1490 	u8 *dmac;
1491 	int sts;
1492 
1493 	/* Get slave that sent this packet */
1494 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1495 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1496 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1497 	    wc->src_qp & 0x4) {
1498 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1499 		return;
1500 	}
1501 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1502 	if (slave != ctx->slave) {
1503 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1504 			     "belongs to another slave\n", wc->src_qp);
1505 		return;
1506 	}
1507 
1508 	/* Map transaction ID */
1509 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1510 				   sizeof (struct mlx4_tunnel_mad),
1511 				   DMA_FROM_DEVICE);
1512 	switch (tunnel->mad.mad_hdr.method) {
1513 	case IB_MGMT_METHOD_SET:
1514 	case IB_MGMT_METHOD_GET:
1515 	case IB_MGMT_METHOD_REPORT:
1516 	case IB_SA_METHOD_GET_TABLE:
1517 	case IB_SA_METHOD_DELETE:
1518 	case IB_SA_METHOD_GET_MULTI:
1519 	case IB_SA_METHOD_GET_TRACE_TBL:
1520 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1521 		if (*slave_id) {
1522 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1523 				     "class:%d slave:%d\n", *slave_id,
1524 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1525 			return;
1526 		} else
1527 			*slave_id = slave;
1528 		break;
1529 	default:
1530 		/* nothing */;
1531 	}
1532 
1533 	/* Class-specific handling */
1534 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1535 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1536 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1537 		if (slave != mlx4_master_func_num(dev->dev) &&
1538 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1539 			return;
1540 		break;
1541 	case IB_MGMT_CLASS_SUBN_ADM:
1542 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1543 			      (struct ib_sa_mad *) &tunnel->mad))
1544 			return;
1545 		break;
1546 	case IB_MGMT_CLASS_CM:
1547 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1548 			      (struct ib_mad *) &tunnel->mad))
1549 			return;
1550 		break;
1551 	case IB_MGMT_CLASS_DEVICE_MGMT:
1552 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1553 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1554 			return;
1555 		break;
1556 	default:
1557 		/* Drop unsupported classes for slaves in tunnel mode */
1558 		if (slave != mlx4_master_func_num(dev->dev)) {
1559 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1560 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1561 			return;
1562 		}
1563 	}
1564 
1565 	/* We are using standard ib_core services to send the mad, so generate a
1566 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1567 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1568 	ah.ibah.device = ctx->ib_dev;
1569 
1570 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1571 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1572 	if (port < 0)
1573 		return;
1574 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1575 	ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
1576 
1577 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1578 	if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
1579 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1580 	dmac = rdma_ah_retrieve_dmac(&ah_attr);
1581 	if (dmac)
1582 		memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
1583 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1584 	/* if slave have default vlan use it */
1585 	if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1586 					&vlan_id, &qos))
1587 		rdma_ah_set_sl(&ah_attr, qos);
1588 
1589 	sts = mlx4_ib_send_to_wire(dev, slave, ctx->port,
1590 				   is_proxy_qp0(dev, wc->src_qp, slave) ?
1591 				   IB_QPT_SMI : IB_QPT_GSI,
1592 				   be16_to_cpu(tunnel->hdr.pkey_index),
1593 				   be32_to_cpu(tunnel->hdr.remote_qpn),
1594 				   be32_to_cpu(tunnel->hdr.qkey),
1595 				   &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1596 	if (sts)
1597 		pr_debug("failed sending %s to wire on behalf of slave %d (%d)\n",
1598 			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
1599 			 slave, sts);
1600 }
1601 
1602 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1603 				 enum ib_qp_type qp_type, int is_tun)
1604 {
1605 	int i;
1606 	struct mlx4_ib_demux_pv_qp *tun_qp;
1607 	int rx_buf_size, tx_buf_size;
1608 	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1609 
1610 	if (qp_type > IB_QPT_GSI)
1611 		return -EINVAL;
1612 
1613 	tun_qp = &ctx->qp[qp_type];
1614 
1615 	tun_qp->ring = kcalloc(nmbr_bufs,
1616 			       sizeof(struct mlx4_ib_buf),
1617 			       GFP_KERNEL);
1618 	if (!tun_qp->ring)
1619 		return -ENOMEM;
1620 
1621 	tun_qp->tx_ring = kcalloc(nmbr_bufs,
1622 				  sizeof (struct mlx4_ib_tun_tx_buf),
1623 				  GFP_KERNEL);
1624 	if (!tun_qp->tx_ring) {
1625 		kfree(tun_qp->ring);
1626 		tun_qp->ring = NULL;
1627 		return -ENOMEM;
1628 	}
1629 
1630 	if (is_tun) {
1631 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1632 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1633 	} else {
1634 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1635 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1636 	}
1637 
1638 	for (i = 0; i < nmbr_bufs; i++) {
1639 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1640 		if (!tun_qp->ring[i].addr)
1641 			goto err;
1642 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1643 							tun_qp->ring[i].addr,
1644 							rx_buf_size,
1645 							DMA_FROM_DEVICE);
1646 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1647 			kfree(tun_qp->ring[i].addr);
1648 			goto err;
1649 		}
1650 	}
1651 
1652 	for (i = 0; i < nmbr_bufs; i++) {
1653 		tun_qp->tx_ring[i].buf.addr =
1654 			kmalloc(tx_buf_size, GFP_KERNEL);
1655 		if (!tun_qp->tx_ring[i].buf.addr)
1656 			goto tx_err;
1657 		tun_qp->tx_ring[i].buf.map =
1658 			ib_dma_map_single(ctx->ib_dev,
1659 					  tun_qp->tx_ring[i].buf.addr,
1660 					  tx_buf_size,
1661 					  DMA_TO_DEVICE);
1662 		if (ib_dma_mapping_error(ctx->ib_dev,
1663 					 tun_qp->tx_ring[i].buf.map)) {
1664 			kfree(tun_qp->tx_ring[i].buf.addr);
1665 			goto tx_err;
1666 		}
1667 		tun_qp->tx_ring[i].ah = NULL;
1668 	}
1669 	spin_lock_init(&tun_qp->tx_lock);
1670 	tun_qp->tx_ix_head = 0;
1671 	tun_qp->tx_ix_tail = 0;
1672 	tun_qp->proxy_qpt = qp_type;
1673 
1674 	return 0;
1675 
1676 tx_err:
1677 	while (i > 0) {
1678 		--i;
1679 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1680 				    tx_buf_size, DMA_TO_DEVICE);
1681 		kfree(tun_qp->tx_ring[i].buf.addr);
1682 	}
1683 	i = nmbr_bufs;
1684 err:
1685 	while (i > 0) {
1686 		--i;
1687 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1688 				    rx_buf_size, DMA_FROM_DEVICE);
1689 		kfree(tun_qp->ring[i].addr);
1690 	}
1691 	kfree(tun_qp->tx_ring);
1692 	tun_qp->tx_ring = NULL;
1693 	kfree(tun_qp->ring);
1694 	tun_qp->ring = NULL;
1695 	return -ENOMEM;
1696 }
1697 
1698 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1699 				     enum ib_qp_type qp_type, int is_tun)
1700 {
1701 	int i;
1702 	struct mlx4_ib_demux_pv_qp *tun_qp;
1703 	int rx_buf_size, tx_buf_size;
1704 	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1705 
1706 	if (qp_type > IB_QPT_GSI)
1707 		return;
1708 
1709 	tun_qp = &ctx->qp[qp_type];
1710 	if (is_tun) {
1711 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1712 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1713 	} else {
1714 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1715 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1716 	}
1717 
1718 
1719 	for (i = 0; i < nmbr_bufs; i++) {
1720 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1721 				    rx_buf_size, DMA_FROM_DEVICE);
1722 		kfree(tun_qp->ring[i].addr);
1723 	}
1724 
1725 	for (i = 0; i < nmbr_bufs; i++) {
1726 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1727 				    tx_buf_size, DMA_TO_DEVICE);
1728 		kfree(tun_qp->tx_ring[i].buf.addr);
1729 		if (tun_qp->tx_ring[i].ah)
1730 			rdma_destroy_ah(tun_qp->tx_ring[i].ah, 0);
1731 	}
1732 	kfree(tun_qp->tx_ring);
1733 	kfree(tun_qp->ring);
1734 }
1735 
1736 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1737 {
1738 	struct mlx4_ib_demux_pv_ctx *ctx;
1739 	struct mlx4_ib_demux_pv_qp *tun_qp;
1740 	struct ib_wc wc;
1741 	int ret;
1742 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1743 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1744 
1745 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1746 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1747 		if (wc.status == IB_WC_SUCCESS) {
1748 			switch (wc.opcode) {
1749 			case IB_WC_RECV:
1750 				mlx4_ib_multiplex_mad(ctx, &wc);
1751 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1752 							     wc.wr_id &
1753 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1754 				if (ret)
1755 					pr_err("Failed reposting tunnel "
1756 					       "buf:%lld\n", wc.wr_id);
1757 				break;
1758 			case IB_WC_SEND:
1759 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1760 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1761 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1762 					= NULL;
1763 				spin_lock(&tun_qp->tx_lock);
1764 				tun_qp->tx_ix_tail++;
1765 				spin_unlock(&tun_qp->tx_lock);
1766 
1767 				break;
1768 			default:
1769 				break;
1770 			}
1771 		} else  {
1772 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1773 				 " status = %d, wrid = 0x%llx\n",
1774 				 ctx->slave, wc.status, wc.wr_id);
1775 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1776 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1777 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1778 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1779 					= NULL;
1780 				spin_lock(&tun_qp->tx_lock);
1781 				tun_qp->tx_ix_tail++;
1782 				spin_unlock(&tun_qp->tx_lock);
1783 			}
1784 		}
1785 	}
1786 }
1787 
1788 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1789 {
1790 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1791 
1792 	/* It's worse than that! He's dead, Jim! */
1793 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1794 	       event->event, sqp->port);
1795 }
1796 
1797 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1798 			    enum ib_qp_type qp_type, int create_tun)
1799 {
1800 	int i, ret;
1801 	struct mlx4_ib_demux_pv_qp *tun_qp;
1802 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1803 	struct ib_qp_attr attr;
1804 	int qp_attr_mask_INIT;
1805 	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1806 
1807 	if (qp_type > IB_QPT_GSI)
1808 		return -EINVAL;
1809 
1810 	tun_qp = &ctx->qp[qp_type];
1811 
1812 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1813 	qp_init_attr.init_attr.send_cq = ctx->cq;
1814 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1815 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1816 	qp_init_attr.init_attr.cap.max_send_wr = nmbr_bufs;
1817 	qp_init_attr.init_attr.cap.max_recv_wr = nmbr_bufs;
1818 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1819 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1820 	if (create_tun) {
1821 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1822 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1823 		qp_init_attr.port = ctx->port;
1824 		qp_init_attr.slave = ctx->slave;
1825 		qp_init_attr.proxy_qp_type = qp_type;
1826 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1827 			   IB_QP_QKEY | IB_QP_PORT;
1828 	} else {
1829 		qp_init_attr.init_attr.qp_type = qp_type;
1830 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1831 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1832 	}
1833 	qp_init_attr.init_attr.port_num = ctx->port;
1834 	qp_init_attr.init_attr.qp_context = ctx;
1835 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1836 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1837 	if (IS_ERR(tun_qp->qp)) {
1838 		ret = PTR_ERR(tun_qp->qp);
1839 		tun_qp->qp = NULL;
1840 		pr_err("Couldn't create %s QP (%d)\n",
1841 		       create_tun ? "tunnel" : "special", ret);
1842 		return ret;
1843 	}
1844 
1845 	memset(&attr, 0, sizeof attr);
1846 	attr.qp_state = IB_QPS_INIT;
1847 	ret = 0;
1848 	if (create_tun)
1849 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1850 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1851 					      &attr.pkey_index);
1852 	if (ret || !create_tun)
1853 		attr.pkey_index =
1854 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1855 	attr.qkey = IB_QP1_QKEY;
1856 	attr.port_num = ctx->port;
1857 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1858 	if (ret) {
1859 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1860 		       create_tun ? "tunnel" : "special", ret);
1861 		goto err_qp;
1862 	}
1863 	attr.qp_state = IB_QPS_RTR;
1864 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1865 	if (ret) {
1866 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1867 		       create_tun ? "tunnel" : "special", ret);
1868 		goto err_qp;
1869 	}
1870 	attr.qp_state = IB_QPS_RTS;
1871 	attr.sq_psn = 0;
1872 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1873 	if (ret) {
1874 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1875 		       create_tun ? "tunnel" : "special", ret);
1876 		goto err_qp;
1877 	}
1878 
1879 	for (i = 0; i < nmbr_bufs; i++) {
1880 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1881 		if (ret) {
1882 			pr_err(" mlx4_ib_post_pv_buf error"
1883 			       " (err = %d, i = %d)\n", ret, i);
1884 			goto err_qp;
1885 		}
1886 	}
1887 	return 0;
1888 
1889 err_qp:
1890 	ib_destroy_qp(tun_qp->qp);
1891 	tun_qp->qp = NULL;
1892 	return ret;
1893 }
1894 
1895 /*
1896  * IB MAD completion callback for real SQPs
1897  */
1898 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1899 {
1900 	struct mlx4_ib_demux_pv_ctx *ctx;
1901 	struct mlx4_ib_demux_pv_qp *sqp;
1902 	struct ib_wc wc;
1903 	struct ib_grh *grh;
1904 	struct ib_mad *mad;
1905 
1906 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1907 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1908 
1909 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1910 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1911 		if (wc.status == IB_WC_SUCCESS) {
1912 			switch (wc.opcode) {
1913 			case IB_WC_SEND:
1914 				kfree(sqp->tx_ring[wc.wr_id &
1915 				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
1916 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
1917 					= NULL;
1918 				spin_lock(&sqp->tx_lock);
1919 				sqp->tx_ix_tail++;
1920 				spin_unlock(&sqp->tx_lock);
1921 				break;
1922 			case IB_WC_RECV:
1923 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1924 						(sqp->ring[wc.wr_id &
1925 						(MLX4_NUM_WIRE_BUFS - 1)].addr))->payload);
1926 				grh = &(((struct mlx4_mad_rcv_buf *)
1927 						(sqp->ring[wc.wr_id &
1928 						(MLX4_NUM_WIRE_BUFS - 1)].addr))->grh);
1929 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1930 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1931 							   (MLX4_NUM_WIRE_BUFS - 1)))
1932 					pr_err("Failed reposting SQP "
1933 					       "buf:%lld\n", wc.wr_id);
1934 				break;
1935 			default:
1936 				break;
1937 			}
1938 		} else  {
1939 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1940 				 " status = %d, wrid = 0x%llx\n",
1941 				 ctx->slave, wc.status, wc.wr_id);
1942 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1943 				kfree(sqp->tx_ring[wc.wr_id &
1944 				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
1945 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
1946 					= NULL;
1947 				spin_lock(&sqp->tx_lock);
1948 				sqp->tx_ix_tail++;
1949 				spin_unlock(&sqp->tx_lock);
1950 			}
1951 		}
1952 	}
1953 }
1954 
1955 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1956 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1957 {
1958 	struct mlx4_ib_demux_pv_ctx *ctx;
1959 
1960 	*ret_ctx = NULL;
1961 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1962 	if (!ctx)
1963 		return -ENOMEM;
1964 
1965 	ctx->ib_dev = &dev->ib_dev;
1966 	ctx->port = port;
1967 	ctx->slave = slave;
1968 	*ret_ctx = ctx;
1969 	return 0;
1970 }
1971 
1972 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1973 {
1974 	if (dev->sriov.demux[port - 1].tun[slave]) {
1975 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1976 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1977 	}
1978 }
1979 
1980 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1981 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1982 {
1983 	int ret, cq_size;
1984 	struct ib_cq_init_attr cq_attr = {};
1985 	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1986 
1987 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1988 		return -EEXIST;
1989 
1990 	ctx->state = DEMUX_PV_STATE_STARTING;
1991 	/* have QP0 only if link layer is IB */
1992 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1993 	    IB_LINK_LAYER_INFINIBAND)
1994 		ctx->has_smi = 1;
1995 
1996 	if (ctx->has_smi) {
1997 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1998 		if (ret) {
1999 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
2000 			goto err_out;
2001 		}
2002 	}
2003 
2004 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
2005 	if (ret) {
2006 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
2007 		goto err_out_qp0;
2008 	}
2009 
2010 	cq_size = 2 * nmbr_bufs;
2011 	if (ctx->has_smi)
2012 		cq_size *= 2;
2013 
2014 	cq_attr.cqe = cq_size;
2015 	ctx->cq = ib_create_cq(ctx->ib_dev,
2016 			       create_tun ? mlx4_ib_tunnel_comp_handler : mlx4_ib_wire_comp_handler,
2017 			       NULL, ctx, &cq_attr);
2018 	if (IS_ERR(ctx->cq)) {
2019 		ret = PTR_ERR(ctx->cq);
2020 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
2021 		goto err_buf;
2022 	}
2023 
2024 	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
2025 	if (IS_ERR(ctx->pd)) {
2026 		ret = PTR_ERR(ctx->pd);
2027 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
2028 		goto err_cq;
2029 	}
2030 
2031 	if (ctx->has_smi) {
2032 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
2033 		if (ret) {
2034 			pr_err("Couldn't create %s QP0 (%d)\n",
2035 			       create_tun ? "tunnel for" : "",  ret);
2036 			goto err_pd;
2037 		}
2038 	}
2039 
2040 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2041 	if (ret) {
2042 		pr_err("Couldn't create %s QP1 (%d)\n",
2043 		       create_tun ? "tunnel for" : "",  ret);
2044 		goto err_qp0;
2045 	}
2046 
2047 	if (create_tun)
2048 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2049 	else
2050 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2051 
2052 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2053 	ctx->wi_wq = to_mdev(ibdev)->sriov.demux[port - 1].wi_wq;
2054 
2055 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2056 	if (ret) {
2057 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2058 		goto err_wq;
2059 	}
2060 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2061 	return 0;
2062 
2063 err_wq:
2064 	ctx->wq = NULL;
2065 	ib_destroy_qp(ctx->qp[1].qp);
2066 	ctx->qp[1].qp = NULL;
2067 
2068 
2069 err_qp0:
2070 	if (ctx->has_smi)
2071 		ib_destroy_qp(ctx->qp[0].qp);
2072 	ctx->qp[0].qp = NULL;
2073 
2074 err_pd:
2075 	ib_dealloc_pd(ctx->pd);
2076 	ctx->pd = NULL;
2077 
2078 err_cq:
2079 	ib_destroy_cq(ctx->cq);
2080 	ctx->cq = NULL;
2081 
2082 err_buf:
2083 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2084 
2085 err_out_qp0:
2086 	if (ctx->has_smi)
2087 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2088 err_out:
2089 	ctx->state = DEMUX_PV_STATE_DOWN;
2090 	return ret;
2091 }
2092 
2093 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2094 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2095 {
2096 	if (!ctx)
2097 		return;
2098 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2099 		ctx->state = DEMUX_PV_STATE_DOWNING;
2100 		if (flush)
2101 			flush_workqueue(ctx->wq);
2102 		if (ctx->has_smi) {
2103 			ib_destroy_qp(ctx->qp[0].qp);
2104 			ctx->qp[0].qp = NULL;
2105 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2106 		}
2107 		ib_destroy_qp(ctx->qp[1].qp);
2108 		ctx->qp[1].qp = NULL;
2109 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2110 		ib_dealloc_pd(ctx->pd);
2111 		ctx->pd = NULL;
2112 		ib_destroy_cq(ctx->cq);
2113 		ctx->cq = NULL;
2114 		ctx->state = DEMUX_PV_STATE_DOWN;
2115 	}
2116 }
2117 
2118 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2119 				  int port, int do_init)
2120 {
2121 	int ret = 0;
2122 
2123 	if (!do_init) {
2124 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2125 		/* for master, destroy real sqp resources */
2126 		if (slave == mlx4_master_func_num(dev->dev))
2127 			destroy_pv_resources(dev, slave, port,
2128 					     dev->sriov.sqps[port - 1], 1);
2129 		/* destroy the tunnel qp resources */
2130 		destroy_pv_resources(dev, slave, port,
2131 				     dev->sriov.demux[port - 1].tun[slave], 1);
2132 		return 0;
2133 	}
2134 
2135 	/* create the tunnel qp resources */
2136 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2137 				  dev->sriov.demux[port - 1].tun[slave]);
2138 
2139 	/* for master, create the real sqp resources */
2140 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2141 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2142 					  dev->sriov.sqps[port - 1]);
2143 	return ret;
2144 }
2145 
2146 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2147 {
2148 	struct mlx4_ib_demux_work *dmxw;
2149 
2150 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2151 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2152 			       dmxw->do_init);
2153 	kfree(dmxw);
2154 	return;
2155 }
2156 
2157 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2158 				       struct mlx4_ib_demux_ctx *ctx,
2159 				       int port)
2160 {
2161 	int ret = 0;
2162 	int i;
2163 
2164 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2165 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2166 	if (!ctx->tun)
2167 		return -ENOMEM;
2168 
2169 	ctx->dev = dev;
2170 	ctx->port = port;
2171 	ctx->ib_dev = &dev->ib_dev;
2172 
2173 	for (i = 0;
2174 	     i < min(dev->dev->caps.sqp_demux,
2175 	     (u16)(dev->dev->persist->num_vfs + 1));
2176 	     i++) {
2177 		struct mlx4_active_ports actv_ports =
2178 			mlx4_get_active_ports(dev->dev, i);
2179 
2180 		if (!test_bit(port - 1, actv_ports.ports))
2181 			continue;
2182 
2183 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2184 		if (ret) {
2185 			ret = -ENOMEM;
2186 			goto err_mcg;
2187 		}
2188 	}
2189 
2190 	ret = mlx4_ib_mcg_port_init(ctx);
2191 	if (ret) {
2192 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2193 		goto err_mcg;
2194 	}
2195 
2196 	ctx->wq = alloc_ordered_workqueue("mlx4_ibt%d", WQ_MEM_RECLAIM, port);
2197 	if (!ctx->wq) {
2198 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2199 		ret = -ENOMEM;
2200 		goto err_wq;
2201 	}
2202 
2203 	ctx->wi_wq = alloc_ordered_workqueue("mlx4_ibwi%d", WQ_MEM_RECLAIM, port);
2204 	if (!ctx->wi_wq) {
2205 		pr_err("Failed to create wire WQ for port %d\n", port);
2206 		ret = -ENOMEM;
2207 		goto err_wiwq;
2208 	}
2209 
2210 	ctx->ud_wq = alloc_ordered_workqueue("mlx4_ibud%d", WQ_MEM_RECLAIM, port);
2211 	if (!ctx->ud_wq) {
2212 		pr_err("Failed to create up/down WQ for port %d\n", port);
2213 		ret = -ENOMEM;
2214 		goto err_udwq;
2215 	}
2216 
2217 	return 0;
2218 
2219 err_udwq:
2220 	destroy_workqueue(ctx->wi_wq);
2221 	ctx->wi_wq = NULL;
2222 
2223 err_wiwq:
2224 	destroy_workqueue(ctx->wq);
2225 	ctx->wq = NULL;
2226 
2227 err_wq:
2228 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2229 err_mcg:
2230 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2231 		free_pv_object(dev, i, port);
2232 	kfree(ctx->tun);
2233 	ctx->tun = NULL;
2234 	return ret;
2235 }
2236 
2237 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2238 {
2239 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2240 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2241 		flush_workqueue(sqp_ctx->wq);
2242 		if (sqp_ctx->has_smi) {
2243 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2244 			sqp_ctx->qp[0].qp = NULL;
2245 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2246 		}
2247 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2248 		sqp_ctx->qp[1].qp = NULL;
2249 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2250 		ib_dealloc_pd(sqp_ctx->pd);
2251 		sqp_ctx->pd = NULL;
2252 		ib_destroy_cq(sqp_ctx->cq);
2253 		sqp_ctx->cq = NULL;
2254 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2255 	}
2256 }
2257 
2258 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2259 {
2260 	int i;
2261 	if (ctx) {
2262 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2263 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2264 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2265 			if (!ctx->tun[i])
2266 				continue;
2267 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2268 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2269 		}
2270 		flush_workqueue(ctx->wq);
2271 		flush_workqueue(ctx->wi_wq);
2272 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2273 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2274 			free_pv_object(dev, i, ctx->port);
2275 		}
2276 		kfree(ctx->tun);
2277 		destroy_workqueue(ctx->ud_wq);
2278 		destroy_workqueue(ctx->wi_wq);
2279 		destroy_workqueue(ctx->wq);
2280 	}
2281 }
2282 
2283 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2284 {
2285 	int i;
2286 
2287 	if (!mlx4_is_master(dev->dev))
2288 		return;
2289 	/* initialize or tear down tunnel QPs for the master */
2290 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2291 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2292 	return;
2293 }
2294 
2295 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2296 {
2297 	int i = 0;
2298 	int err;
2299 
2300 	if (!mlx4_is_mfunc(dev->dev))
2301 		return 0;
2302 
2303 	dev->sriov.is_going_down = 0;
2304 	spin_lock_init(&dev->sriov.going_down_lock);
2305 	mlx4_ib_cm_paravirt_init(dev);
2306 
2307 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2308 
2309 	if (mlx4_is_slave(dev->dev)) {
2310 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2311 		return 0;
2312 	}
2313 
2314 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2315 		if (i == mlx4_master_func_num(dev->dev))
2316 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2317 		else
2318 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2319 	}
2320 
2321 	err = mlx4_ib_init_alias_guid_service(dev);
2322 	if (err) {
2323 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2324 		goto paravirt_err;
2325 	}
2326 	err = mlx4_ib_device_register_sysfs(dev);
2327 	if (err) {
2328 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2329 		goto sysfs_err;
2330 	}
2331 
2332 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2333 		     dev->dev->caps.sqp_demux);
2334 	for (i = 0; i < dev->num_ports; i++) {
2335 		union ib_gid gid;
2336 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2337 		if (err)
2338 			goto demux_err;
2339 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2340 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2341 			     be64_to_cpu(gid.global.subnet_prefix));
2342 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2343 				      &dev->sriov.sqps[i]);
2344 		if (err)
2345 			goto demux_err;
2346 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2347 		if (err)
2348 			goto free_pv;
2349 	}
2350 	mlx4_ib_master_tunnels(dev, 1);
2351 	return 0;
2352 
2353 free_pv:
2354 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2355 demux_err:
2356 	while (--i >= 0) {
2357 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2358 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2359 	}
2360 	mlx4_ib_device_unregister_sysfs(dev);
2361 
2362 sysfs_err:
2363 	mlx4_ib_destroy_alias_guid_service(dev);
2364 
2365 paravirt_err:
2366 	mlx4_ib_cm_paravirt_clean(dev, -1);
2367 
2368 	return err;
2369 }
2370 
2371 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2372 {
2373 	int i;
2374 	unsigned long flags;
2375 
2376 	if (!mlx4_is_mfunc(dev->dev))
2377 		return;
2378 
2379 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2380 	dev->sriov.is_going_down = 1;
2381 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2382 	if (mlx4_is_master(dev->dev)) {
2383 		for (i = 0; i < dev->num_ports; i++) {
2384 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2385 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2386 			kfree(dev->sriov.sqps[i]);
2387 			dev->sriov.sqps[i] = NULL;
2388 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2389 		}
2390 
2391 		mlx4_ib_cm_paravirt_clean(dev, -1);
2392 		mlx4_ib_destroy_alias_guid_service(dev);
2393 		mlx4_ib_device_unregister_sysfs(dev);
2394 	}
2395 }
2396