xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision de528723)
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, u8 port_num);
92 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 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 | prandom_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, u8 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, u8 port_num, const struct ib_mad *mad,
221 		      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=%d, "
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, u8 port_num, const struct ib_mad *mad)
384 {
385 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
386 	struct ib_mad_send_buf *send_buf;
387 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
388 	int ret;
389 	unsigned long flags;
390 
391 	if (agent) {
392 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
393 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
394 					      IB_MGMT_BASE_VERSION);
395 		if (IS_ERR(send_buf))
396 			return;
397 		/*
398 		 * We rely here on the fact that MLX QPs don't use the
399 		 * address handle after the send is posted (this is
400 		 * wrong following the IB spec strictly, but we know
401 		 * it's OK for our devices).
402 		 */
403 		spin_lock_irqsave(&dev->sm_lock, flags);
404 		memcpy(send_buf->mad, mad, sizeof *mad);
405 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
406 			ret = ib_post_send_mad(send_buf, NULL);
407 		else
408 			ret = -EINVAL;
409 		spin_unlock_irqrestore(&dev->sm_lock, flags);
410 
411 		if (ret)
412 			ib_free_send_mad(send_buf);
413 	}
414 }
415 
416 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
417 							     struct ib_sa_mad *sa_mad)
418 {
419 	int ret = 0;
420 
421 	/* dispatch to different sa handlers */
422 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
423 	case IB_SA_ATTR_MC_MEMBER_REC:
424 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
425 		break;
426 	default:
427 		break;
428 	}
429 	return ret;
430 }
431 
432 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
433 {
434 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
435 	int i;
436 
437 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
438 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
439 			return i;
440 	}
441 	return -1;
442 }
443 
444 
445 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
446 				   u8 port, u16 pkey, u16 *ix)
447 {
448 	int i, ret;
449 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
450 	u16 slot_pkey;
451 
452 	if (slave == mlx4_master_func_num(dev->dev))
453 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
454 
455 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
456 
457 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
458 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
459 			continue;
460 
461 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
462 
463 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
464 		if (ret)
465 			continue;
466 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
467 			if (slot_pkey & 0x8000) {
468 				*ix = (u16) pkey_ix;
469 				return 0;
470 			} else {
471 				/* take first partial pkey index found */
472 				if (partial_ix == 0xFF)
473 					partial_ix = pkey_ix;
474 			}
475 		}
476 	}
477 
478 	if (partial_ix < 0xFF) {
479 		*ix = (u16) partial_ix;
480 		return 0;
481 	}
482 
483 	return -EINVAL;
484 }
485 
486 static int get_gids_from_l3_hdr(struct ib_grh *grh, union ib_gid *sgid,
487 				union ib_gid *dgid)
488 {
489 	int version = ib_get_rdma_header_version((const union rdma_network_hdr *)grh);
490 	enum rdma_network_type net_type;
491 
492 	if (version == 4)
493 		net_type = RDMA_NETWORK_IPV4;
494 	else if (version == 6)
495 		net_type = RDMA_NETWORK_IPV6;
496 	else
497 		return -EINVAL;
498 
499 	return ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
500 					 sgid, dgid);
501 }
502 
503 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
504 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
505 			  struct ib_grh *grh, struct ib_mad *mad)
506 {
507 	struct ib_sge list;
508 	struct ib_ud_wr wr;
509 	const struct ib_send_wr *bad_wr;
510 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
511 	struct mlx4_ib_demux_pv_qp *tun_qp;
512 	struct mlx4_rcv_tunnel_mad *tun_mad;
513 	struct rdma_ah_attr attr;
514 	struct ib_ah *ah;
515 	struct ib_qp *src_qp = NULL;
516 	unsigned tun_tx_ix = 0;
517 	int dqpn;
518 	int ret = 0;
519 	u16 tun_pkey_ix;
520 	u16 cached_pkey;
521 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
522 
523 	if (dest_qpt > IB_QPT_GSI)
524 		return -EINVAL;
525 
526 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
527 
528 	/* check if proxy qp created */
529 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
530 		return -EAGAIN;
531 
532 	if (!dest_qpt)
533 		tun_qp = &tun_ctx->qp[0];
534 	else
535 		tun_qp = &tun_ctx->qp[1];
536 
537 	/* compute P_Key index to put in tunnel header for slave */
538 	if (dest_qpt) {
539 		u16 pkey_ix;
540 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
541 		if (ret)
542 			return -EINVAL;
543 
544 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
545 		if (ret)
546 			return -EINVAL;
547 		tun_pkey_ix = pkey_ix;
548 	} else
549 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
550 
551 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
552 
553 	/* get tunnel tx data buf for slave */
554 	src_qp = tun_qp->qp;
555 
556 	/* create ah. Just need an empty one with the port num for the post send.
557 	 * The driver will set the force loopback bit in post_send */
558 	memset(&attr, 0, sizeof attr);
559 	attr.type = rdma_ah_find_type(&dev->ib_dev, port);
560 
561 	rdma_ah_set_port_num(&attr, port);
562 	if (is_eth) {
563 		union ib_gid sgid;
564 		union ib_gid dgid;
565 
566 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
567 			return -EINVAL;
568 		rdma_ah_set_grh(&attr, &dgid, 0, 0, 0, 0);
569 	}
570 	ah = rdma_create_ah(tun_ctx->pd, &attr, 0);
571 	if (IS_ERR(ah))
572 		return -ENOMEM;
573 
574 	/* allocate tunnel tx buf after pass failure returns */
575 	spin_lock(&tun_qp->tx_lock);
576 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
577 	    (MLX4_NUM_TUNNEL_BUFS - 1))
578 		ret = -EAGAIN;
579 	else
580 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
581 	spin_unlock(&tun_qp->tx_lock);
582 	if (ret)
583 		goto end;
584 
585 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
586 	if (tun_qp->tx_ring[tun_tx_ix].ah)
587 		rdma_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah, 0);
588 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
589 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
590 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
591 				   sizeof (struct mlx4_rcv_tunnel_mad),
592 				   DMA_TO_DEVICE);
593 
594 	/* copy over to tunnel buffer */
595 	if (grh)
596 		memcpy(&tun_mad->grh, grh, sizeof *grh);
597 	memcpy(&tun_mad->mad, mad, sizeof *mad);
598 
599 	/* adjust tunnel data */
600 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
601 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
602 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
603 
604 	if (is_eth) {
605 		u16 vlan = 0;
606 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
607 						NULL)) {
608 			/* VST mode */
609 			if (vlan != wc->vlan_id)
610 				/* Packet vlan is not the VST-assigned vlan.
611 				 * Drop the packet.
612 				 */
613 				goto out;
614 			 else
615 				/* Remove the vlan tag before forwarding
616 				 * the packet to the VF.
617 				 */
618 				vlan = 0xffff;
619 		} else {
620 			vlan = wc->vlan_id;
621 		}
622 
623 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
624 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
625 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
626 	} else {
627 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
628 		tun_mad->hdr.slid_mac_47_32 = ib_lid_be16(wc->slid);
629 	}
630 
631 	ib_dma_sync_single_for_device(&dev->ib_dev,
632 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
633 				      sizeof (struct mlx4_rcv_tunnel_mad),
634 				      DMA_TO_DEVICE);
635 
636 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
637 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
638 	list.lkey = tun_ctx->pd->local_dma_lkey;
639 
640 	wr.ah = ah;
641 	wr.port_num = port;
642 	wr.remote_qkey = IB_QP_SET_QKEY;
643 	wr.remote_qpn = dqpn;
644 	wr.wr.next = NULL;
645 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
646 	wr.wr.sg_list = &list;
647 	wr.wr.num_sge = 1;
648 	wr.wr.opcode = IB_WR_SEND;
649 	wr.wr.send_flags = IB_SEND_SIGNALED;
650 
651 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
652 	if (!ret)
653 		return 0;
654  out:
655 	spin_lock(&tun_qp->tx_lock);
656 	tun_qp->tx_ix_tail++;
657 	spin_unlock(&tun_qp->tx_lock);
658 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
659 end:
660 	rdma_destroy_ah(ah, 0);
661 	return ret;
662 }
663 
664 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
665 			struct ib_wc *wc, struct ib_grh *grh,
666 			struct ib_mad *mad)
667 {
668 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
669 	int err, other_port;
670 	int slave = -1;
671 	u8 *slave_id;
672 	int is_eth = 0;
673 
674 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
675 		is_eth = 0;
676 	else
677 		is_eth = 1;
678 
679 	if (is_eth) {
680 		union ib_gid dgid;
681 		union ib_gid sgid;
682 
683 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
684 			return -EINVAL;
685 		if (!(wc->wc_flags & IB_WC_GRH)) {
686 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
687 			return -EINVAL;
688 		}
689 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
690 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
691 			return -EINVAL;
692 		}
693 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, dgid.raw, &slave);
694 		if (err && mlx4_is_mf_bonded(dev->dev)) {
695 			other_port = (port == 1) ? 2 : 1;
696 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, dgid.raw, &slave);
697 			if (!err) {
698 				port = other_port;
699 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
700 					 slave, grh->dgid.raw, port, other_port);
701 			}
702 		}
703 		if (err) {
704 			mlx4_ib_warn(ibdev, "failed matching grh\n");
705 			return -ENOENT;
706 		}
707 		if (slave >= dev->dev->caps.sqp_demux) {
708 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
709 				     slave, dev->dev->caps.sqp_demux);
710 			return -ENOENT;
711 		}
712 
713 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
714 			return 0;
715 
716 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
717 		if (err)
718 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
719 				 slave, err);
720 		return 0;
721 	}
722 
723 	/* Initially assume that this mad is for us */
724 	slave = mlx4_master_func_num(dev->dev);
725 
726 	/* See if the slave id is encoded in a response mad */
727 	if (mad->mad_hdr.method & 0x80) {
728 		slave_id = (u8 *) &mad->mad_hdr.tid;
729 		slave = *slave_id;
730 		if (slave != 255) /*255 indicates the dom0*/
731 			*slave_id = 0; /* remap tid */
732 	}
733 
734 	/* If a grh is present, we demux according to it */
735 	if (wc->wc_flags & IB_WC_GRH) {
736 		if (grh->dgid.global.interface_id ==
737 			cpu_to_be64(IB_SA_WELL_KNOWN_GUID) &&
738 		    grh->dgid.global.subnet_prefix == cpu_to_be64(
739 			atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix))) {
740 			slave = 0;
741 		} else {
742 			slave = mlx4_ib_find_real_gid(ibdev, port,
743 						      grh->dgid.global.interface_id);
744 			if (slave < 0) {
745 				mlx4_ib_warn(ibdev, "failed matching grh\n");
746 				return -ENOENT;
747 			}
748 		}
749 	}
750 	/* Class-specific handling */
751 	switch (mad->mad_hdr.mgmt_class) {
752 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
753 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
754 		/* 255 indicates the dom0 */
755 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
756 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
757 				return -EPERM;
758 			/* for a VF. drop unsolicited MADs */
759 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
760 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
761 					     slave, mad->mad_hdr.mgmt_class,
762 					     mad->mad_hdr.method);
763 				return -EINVAL;
764 			}
765 		}
766 		break;
767 	case IB_MGMT_CLASS_SUBN_ADM:
768 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
769 					     (struct ib_sa_mad *) mad))
770 			return 0;
771 		break;
772 	case IB_MGMT_CLASS_CM:
773 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
774 			return 0;
775 		break;
776 	case IB_MGMT_CLASS_DEVICE_MGMT:
777 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
778 			return 0;
779 		break;
780 	default:
781 		/* Drop unsupported classes for slaves in tunnel mode */
782 		if (slave != mlx4_master_func_num(dev->dev)) {
783 			pr_debug("dropping unsupported ingress mad from class:%d "
784 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
785 			return 0;
786 		}
787 	}
788 	/*make sure that no slave==255 was not handled yet.*/
789 	if (slave >= dev->dev->caps.sqp_demux) {
790 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
791 			     slave, dev->dev->caps.sqp_demux);
792 		return -ENOENT;
793 	}
794 
795 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
796 	if (err)
797 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
798 			 slave, err);
799 	return 0;
800 }
801 
802 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
803 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
804 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
805 {
806 	u16 slid, prev_lid = 0;
807 	int err;
808 	struct ib_port_attr pattr;
809 
810 	if (in_wc && in_wc->qp) {
811 		pr_debug("received MAD: port:%d slid:%d sqpn:%d "
812 			 "dlid_bits:%d dqpn:%d wc_flags:0x%x tid:%016llx cls:%x mtd:%x atr:%x\n",
813 			 port_num,
814 			 in_wc->slid, in_wc->src_qp,
815 			 in_wc->dlid_path_bits,
816 			 in_wc->qp->qp_num,
817 			 in_wc->wc_flags,
818 			 be64_to_cpu(in_mad->mad_hdr.tid),
819 			 in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
820 			 be16_to_cpu(in_mad->mad_hdr.attr_id));
821 		if (in_wc->wc_flags & IB_WC_GRH) {
822 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
823 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
824 				 be64_to_cpu(in_grh->sgid.global.interface_id));
825 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
826 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
827 				 be64_to_cpu(in_grh->dgid.global.interface_id));
828 		}
829 	}
830 
831 	slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
832 
833 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
834 		forward_trap(to_mdev(ibdev), port_num, in_mad);
835 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
836 	}
837 
838 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
839 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
840 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
841 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
842 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
843 			return IB_MAD_RESULT_SUCCESS;
844 
845 		/*
846 		 * Don't process SMInfo queries -- the SMA can't handle them.
847 		 */
848 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
849 			return IB_MAD_RESULT_SUCCESS;
850 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
851 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
852 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
853 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
854 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
855 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
856 			return IB_MAD_RESULT_SUCCESS;
857 	} else
858 		return IB_MAD_RESULT_SUCCESS;
859 
860 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
861 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
862 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
863 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
864 	    !ib_query_port(ibdev, port_num, &pattr))
865 		prev_lid = ib_lid_cpu16(pattr.lid);
866 
867 	err = mlx4_MAD_IFC(to_mdev(ibdev),
868 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
869 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
870 			   MLX4_MAD_IFC_NET_VIEW,
871 			   port_num, in_wc, in_grh, in_mad, out_mad);
872 	if (err)
873 		return IB_MAD_RESULT_FAILURE;
874 
875 	if (!out_mad->mad_hdr.status) {
876 		smp_snoop(ibdev, port_num, in_mad, prev_lid);
877 		/* slaves get node desc from FW */
878 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
879 			node_desc_override(ibdev, out_mad);
880 	}
881 
882 	/* set return bit in status of directed route responses */
883 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
884 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
885 
886 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
887 		/* no response for trap repress */
888 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
889 
890 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
891 }
892 
893 static void edit_counter(struct mlx4_counter *cnt, void *counters,
894 			 __be16 attr_id)
895 {
896 	switch (attr_id) {
897 	case IB_PMA_PORT_COUNTERS:
898 	{
899 		struct ib_pma_portcounters *pma_cnt =
900 			(struct ib_pma_portcounters *)counters;
901 
902 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
903 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
904 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
905 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
906 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
907 				     be64_to_cpu(cnt->tx_frames));
908 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
909 				     be64_to_cpu(cnt->rx_frames));
910 		break;
911 	}
912 	case IB_PMA_PORT_COUNTERS_EXT:
913 	{
914 		struct ib_pma_portcounters_ext *pma_cnt_ext =
915 			(struct ib_pma_portcounters_ext *)counters;
916 
917 		pma_cnt_ext->port_xmit_data =
918 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
919 		pma_cnt_ext->port_rcv_data =
920 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
921 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
922 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
923 		break;
924 	}
925 	}
926 }
927 
928 static int iboe_process_mad_port_info(void *out_mad)
929 {
930 	struct ib_class_port_info cpi = {};
931 
932 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
933 	memcpy(out_mad, &cpi, sizeof(cpi));
934 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
935 }
936 
937 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
938 			const struct ib_wc *in_wc, 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 		memset(out_mad->data, 0, sizeof out_mad->data);
970 		switch (counter_stats.counter_mode & 0xf) {
971 		case 0:
972 			edit_counter(&counter_stats,
973 				     (void *)(out_mad->data + 40),
974 				     in_mad->mad_hdr.attr_id);
975 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
976 			break;
977 		default:
978 			err = IB_MAD_RESULT_FAILURE;
979 		}
980 	}
981 
982 	return err;
983 }
984 
985 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
986 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
987 			const struct ib_mad_hdr *in, size_t in_mad_size,
988 			struct ib_mad_hdr *out, size_t *out_mad_size,
989 			u16 *out_mad_pkey_index)
990 {
991 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
992 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
993 	struct ib_mad *out_mad = (struct ib_mad *)out;
994 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
995 
996 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
997 			 *out_mad_size != sizeof(*out_mad)))
998 		return IB_MAD_RESULT_FAILURE;
999 
1000 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
1001 	 * queries, should be called only by VFs and for that specific purpose
1002 	 */
1003 	if (link == IB_LINK_LAYER_INFINIBAND) {
1004 		if (mlx4_is_slave(dev->dev) &&
1005 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
1006 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
1007 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
1008 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
1009 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1010 						in_grh, in_mad, out_mad);
1011 
1012 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
1013 				      in_grh, in_mad, out_mad);
1014 	}
1015 
1016 	if (link == IB_LINK_LAYER_ETHERNET)
1017 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1018 					in_grh, in_mad, out_mad);
1019 
1020 	return -EINVAL;
1021 }
1022 
1023 static void send_handler(struct ib_mad_agent *agent,
1024 			 struct ib_mad_send_wc *mad_send_wc)
1025 {
1026 	if (mad_send_wc->send_buf->context[0])
1027 		rdma_destroy_ah(mad_send_wc->send_buf->context[0], 0);
1028 	ib_free_send_mad(mad_send_wc->send_buf);
1029 }
1030 
1031 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1032 {
1033 	struct ib_mad_agent *agent;
1034 	int p, q;
1035 	int ret;
1036 	enum rdma_link_layer ll;
1037 
1038 	for (p = 0; p < dev->num_ports; ++p) {
1039 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1040 		for (q = 0; q <= 1; ++q) {
1041 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1042 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1043 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1044 							      NULL, 0, send_handler,
1045 							      NULL, NULL, 0);
1046 				if (IS_ERR(agent)) {
1047 					ret = PTR_ERR(agent);
1048 					goto err;
1049 				}
1050 				dev->send_agent[p][q] = agent;
1051 			} else
1052 				dev->send_agent[p][q] = NULL;
1053 		}
1054 	}
1055 
1056 	return 0;
1057 
1058 err:
1059 	for (p = 0; p < dev->num_ports; ++p)
1060 		for (q = 0; q <= 1; ++q)
1061 			if (dev->send_agent[p][q])
1062 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1063 
1064 	return ret;
1065 }
1066 
1067 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1068 {
1069 	struct ib_mad_agent *agent;
1070 	int p, q;
1071 
1072 	for (p = 0; p < dev->num_ports; ++p) {
1073 		for (q = 0; q <= 1; ++q) {
1074 			agent = dev->send_agent[p][q];
1075 			if (agent) {
1076 				dev->send_agent[p][q] = NULL;
1077 				ib_unregister_mad_agent(agent);
1078 			}
1079 		}
1080 
1081 		if (dev->sm_ah[p])
1082 			rdma_destroy_ah(dev->sm_ah[p], 0);
1083 	}
1084 }
1085 
1086 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1087 {
1088 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1089 
1090 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1091 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1092 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1093 }
1094 
1095 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1096 {
1097 	/* re-configure the alias-guid and mcg's */
1098 	if (mlx4_is_master(dev->dev)) {
1099 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1100 
1101 		if (!dev->sriov.is_going_down) {
1102 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1103 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1104 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1105 		}
1106 	}
1107 
1108 	/* Update the sl to vl table from inside client rereg
1109 	 * only if in secure-host mode (snooping is not possible)
1110 	 * and the sl-to-vl change event is not generated by FW.
1111 	 */
1112 	if (!mlx4_is_slave(dev->dev) &&
1113 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1114 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1115 		if (mlx4_is_master(dev->dev))
1116 			/* already in work queue from mlx4_ib_event queueing
1117 			 * mlx4_handle_port_mgmt_change_event, which calls
1118 			 * this procedure. Therefore, call sl2vl_update directly.
1119 			 */
1120 			mlx4_ib_sl2vl_update(dev, port_num);
1121 		else
1122 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
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 		goto out;
1148 
1149 	guid_tbl_blk_num  *= 4;
1150 
1151 	for (i = 0; i < 4; i++) {
1152 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1153 			continue;
1154 		memset(in_mad, 0, sizeof *in_mad);
1155 		memset(out_mad, 0, sizeof *out_mad);
1156 
1157 		in_mad->base_version  = 1;
1158 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1159 		in_mad->class_version = 1;
1160 		in_mad->method        = IB_MGMT_METHOD_GET;
1161 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1162 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1163 
1164 		if (mlx4_MAD_IFC(dev,
1165 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1166 				 port_num, NULL, NULL, in_mad, out_mad)) {
1167 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1168 			goto out;
1169 		}
1170 
1171 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1172 						    port_num,
1173 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1174 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1175 						     port_num,
1176 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1177 	}
1178 
1179 out:
1180 	kfree(in_mad);
1181 	kfree(out_mad);
1182 	return;
1183 }
1184 
1185 void handle_port_mgmt_change_event(struct work_struct *work)
1186 {
1187 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1188 	struct mlx4_ib_dev *dev = ew->ib_dev;
1189 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1190 	u8 port = eqe->event.port_mgmt_change.port;
1191 	u32 changed_attr;
1192 	u32 tbl_block;
1193 	u32 change_bitmap;
1194 
1195 	switch (eqe->subtype) {
1196 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1197 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1198 
1199 		/* Update the SM ah - This should be done before handling
1200 		   the other changed attributes so that MADs can be sent to the SM */
1201 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1202 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1203 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1204 			update_sm_ah(dev, port, lid, sl);
1205 		}
1206 
1207 		/* Check if it is a lid change event */
1208 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1209 			handle_lid_change_event(dev, port);
1210 
1211 		/* Generate GUID changed event */
1212 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1213 			if (mlx4_is_master(dev->dev)) {
1214 				union ib_gid gid;
1215 				int err = 0;
1216 
1217 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1218 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1219 				else
1220 					gid.global.subnet_prefix =
1221 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1222 				if (err) {
1223 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1224 						port, err);
1225 				} else {
1226 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1227 						 port,
1228 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1229 						 be64_to_cpu(gid.global.subnet_prefix));
1230 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1231 						     be64_to_cpu(gid.global.subnet_prefix));
1232 				}
1233 			}
1234 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1235 			/*if master, notify all slaves*/
1236 			if (mlx4_is_master(dev->dev))
1237 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1238 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1239 		}
1240 
1241 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1242 			handle_client_rereg_event(dev, port);
1243 		break;
1244 
1245 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1246 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1247 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1248 			propagate_pkey_ev(dev, port, eqe);
1249 		break;
1250 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1251 		/* paravirtualized master's guid is guid 0 -- does not change */
1252 		if (!mlx4_is_master(dev->dev))
1253 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1254 		/*if master, notify relevant slaves*/
1255 		else if (!dev->sriov.is_going_down) {
1256 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1257 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1258 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1259 		}
1260 		break;
1261 
1262 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1263 		/* cache sl to vl mapping changes for use in
1264 		 * filling QP1 LRH VL field when sending packets
1265 		 */
1266 		if (!mlx4_is_slave(dev->dev)) {
1267 			union sl2vl_tbl_to_u64 sl2vl64;
1268 			int jj;
1269 
1270 			for (jj = 0; jj < 8; jj++) {
1271 				sl2vl64.sl8[jj] =
1272 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1273 				pr_debug("port %u, sl2vl[%d] = %02x\n",
1274 					 port, jj, sl2vl64.sl8[jj]);
1275 			}
1276 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1277 		}
1278 		break;
1279 	default:
1280 		pr_warn("Unsupported subtype 0x%x for "
1281 			"Port Management Change event\n", eqe->subtype);
1282 	}
1283 
1284 	kfree(ew);
1285 }
1286 
1287 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1288 			    enum ib_event_type type)
1289 {
1290 	struct ib_event event;
1291 
1292 	event.device		= &dev->ib_dev;
1293 	event.element.port_num	= port_num;
1294 	event.event		= type;
1295 
1296 	ib_dispatch_event(&event);
1297 }
1298 
1299 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1300 {
1301 	unsigned long flags;
1302 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1303 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1304 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1305 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1306 		queue_work(ctx->wq, &ctx->work);
1307 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1308 }
1309 
1310 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1311 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1312 				  int index)
1313 {
1314 	struct ib_sge sg_list;
1315 	struct ib_recv_wr recv_wr;
1316 	const struct ib_recv_wr *bad_recv_wr;
1317 	int size;
1318 
1319 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1320 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1321 
1322 	sg_list.addr = tun_qp->ring[index].map;
1323 	sg_list.length = size;
1324 	sg_list.lkey = ctx->pd->local_dma_lkey;
1325 
1326 	recv_wr.next = NULL;
1327 	recv_wr.sg_list = &sg_list;
1328 	recv_wr.num_sge = 1;
1329 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1330 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1331 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1332 				      size, DMA_FROM_DEVICE);
1333 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1334 }
1335 
1336 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1337 		int slave, struct ib_sa_mad *sa_mad)
1338 {
1339 	int ret = 0;
1340 
1341 	/* dispatch to different sa handlers */
1342 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1343 	case IB_SA_ATTR_MC_MEMBER_REC:
1344 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1345 		break;
1346 	default:
1347 		break;
1348 	}
1349 	return ret;
1350 }
1351 
1352 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1353 {
1354 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1355 
1356 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1357 }
1358 
1359 
1360 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1361 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1362 			 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
1363 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1364 {
1365 	struct ib_sge list;
1366 	struct ib_ud_wr wr;
1367 	const struct ib_send_wr *bad_wr;
1368 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1369 	struct mlx4_ib_demux_pv_qp *sqp;
1370 	struct mlx4_mad_snd_buf *sqp_mad;
1371 	struct ib_ah *ah;
1372 	struct ib_qp *send_qp = NULL;
1373 	unsigned wire_tx_ix = 0;
1374 	u16 wire_pkey_ix;
1375 	int src_qpnum;
1376 	int ret;
1377 
1378 	sqp_ctx = dev->sriov.sqps[port-1];
1379 
1380 	/* check if proxy qp created */
1381 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1382 		return -EAGAIN;
1383 
1384 	if (dest_qpt == IB_QPT_SMI) {
1385 		src_qpnum = 0;
1386 		sqp = &sqp_ctx->qp[0];
1387 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1388 	} else {
1389 		src_qpnum = 1;
1390 		sqp = &sqp_ctx->qp[1];
1391 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1392 	}
1393 
1394 	send_qp = sqp->qp;
1395 
1396 	ah = rdma_zalloc_drv_obj(sqp_ctx->pd->device, ib_ah);
1397 	if (!ah)
1398 		return -ENOMEM;
1399 
1400 	ah->device = sqp_ctx->pd->device;
1401 	ah->pd = sqp_ctx->pd;
1402 
1403 	/* create ah */
1404 	ret = mlx4_ib_create_ah_slave(ah, attr,
1405 				      rdma_ah_retrieve_grh(attr)->sgid_index,
1406 				      s_mac, vlan_id);
1407 	if (ret)
1408 		goto out;
1409 
1410 	spin_lock(&sqp->tx_lock);
1411 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1412 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1413 		ret = -EAGAIN;
1414 	else
1415 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1416 	spin_unlock(&sqp->tx_lock);
1417 	if (ret)
1418 		goto out;
1419 
1420 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1421 	kfree(sqp->tx_ring[wire_tx_ix].ah);
1422 	sqp->tx_ring[wire_tx_ix].ah = ah;
1423 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1424 				   sqp->tx_ring[wire_tx_ix].buf.map,
1425 				   sizeof (struct mlx4_mad_snd_buf),
1426 				   DMA_TO_DEVICE);
1427 
1428 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1429 
1430 	ib_dma_sync_single_for_device(&dev->ib_dev,
1431 				      sqp->tx_ring[wire_tx_ix].buf.map,
1432 				      sizeof (struct mlx4_mad_snd_buf),
1433 				      DMA_TO_DEVICE);
1434 
1435 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1436 	list.length = sizeof (struct mlx4_mad_snd_buf);
1437 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1438 
1439 	wr.ah = ah;
1440 	wr.port_num = port;
1441 	wr.pkey_index = wire_pkey_ix;
1442 	wr.remote_qkey = qkey;
1443 	wr.remote_qpn = remote_qpn;
1444 	wr.wr.next = NULL;
1445 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1446 	wr.wr.sg_list = &list;
1447 	wr.wr.num_sge = 1;
1448 	wr.wr.opcode = IB_WR_SEND;
1449 	wr.wr.send_flags = IB_SEND_SIGNALED;
1450 
1451 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1452 	if (!ret)
1453 		return 0;
1454 
1455 	spin_lock(&sqp->tx_lock);
1456 	sqp->tx_ix_tail++;
1457 	spin_unlock(&sqp->tx_lock);
1458 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1459 out:
1460 	kfree(ah);
1461 	return ret;
1462 }
1463 
1464 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1465 {
1466 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1467 		return slave;
1468 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1469 }
1470 
1471 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1472 				    struct rdma_ah_attr *ah_attr)
1473 {
1474 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
1475 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1476 		grh->sgid_index = slave;
1477 	else
1478 		grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
1479 }
1480 
1481 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1482 {
1483 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1484 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1485 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1486 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1487 	struct mlx4_ib_ah ah;
1488 	struct rdma_ah_attr ah_attr;
1489 	u8 *slave_id;
1490 	int slave;
1491 	int port;
1492 	u16 vlan_id;
1493 	u8 qos;
1494 	u8 *dmac;
1495 
1496 	/* Get slave that sent this packet */
1497 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1498 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1499 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1500 	    wc->src_qp & 0x4) {
1501 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1502 		return;
1503 	}
1504 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1505 	if (slave != ctx->slave) {
1506 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1507 			     "belongs to another slave\n", wc->src_qp);
1508 		return;
1509 	}
1510 
1511 	/* Map transaction ID */
1512 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1513 				   sizeof (struct mlx4_tunnel_mad),
1514 				   DMA_FROM_DEVICE);
1515 	switch (tunnel->mad.mad_hdr.method) {
1516 	case IB_MGMT_METHOD_SET:
1517 	case IB_MGMT_METHOD_GET:
1518 	case IB_MGMT_METHOD_REPORT:
1519 	case IB_SA_METHOD_GET_TABLE:
1520 	case IB_SA_METHOD_DELETE:
1521 	case IB_SA_METHOD_GET_MULTI:
1522 	case IB_SA_METHOD_GET_TRACE_TBL:
1523 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1524 		if (*slave_id) {
1525 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1526 				     "class:%d slave:%d\n", *slave_id,
1527 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1528 			return;
1529 		} else
1530 			*slave_id = slave;
1531 	default:
1532 		/* nothing */;
1533 	}
1534 
1535 	/* Class-specific handling */
1536 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1537 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1538 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1539 		if (slave != mlx4_master_func_num(dev->dev) &&
1540 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1541 			return;
1542 		break;
1543 	case IB_MGMT_CLASS_SUBN_ADM:
1544 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1545 			      (struct ib_sa_mad *) &tunnel->mad))
1546 			return;
1547 		break;
1548 	case IB_MGMT_CLASS_CM:
1549 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1550 			      (struct ib_mad *) &tunnel->mad))
1551 			return;
1552 		break;
1553 	case IB_MGMT_CLASS_DEVICE_MGMT:
1554 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1555 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1556 			return;
1557 		break;
1558 	default:
1559 		/* Drop unsupported classes for slaves in tunnel mode */
1560 		if (slave != mlx4_master_func_num(dev->dev)) {
1561 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1562 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1563 			return;
1564 		}
1565 	}
1566 
1567 	/* We are using standard ib_core services to send the mad, so generate a
1568 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1569 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1570 	ah.ibah.device = ctx->ib_dev;
1571 
1572 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1573 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1574 	if (port < 0)
1575 		return;
1576 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1577 	ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
1578 
1579 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1580 	if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
1581 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1582 	dmac = rdma_ah_retrieve_dmac(&ah_attr);
1583 	if (dmac)
1584 		memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
1585 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1586 	/* if slave have default vlan use it */
1587 	if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1588 					&vlan_id, &qos))
1589 		rdma_ah_set_sl(&ah_attr, qos);
1590 
1591 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1592 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1593 			     IB_QPT_SMI : IB_QPT_GSI,
1594 			     be16_to_cpu(tunnel->hdr.pkey_index),
1595 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1596 			     be32_to_cpu(tunnel->hdr.qkey),
1597 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1598 }
1599 
1600 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1601 				 enum ib_qp_type qp_type, int is_tun)
1602 {
1603 	int i;
1604 	struct mlx4_ib_demux_pv_qp *tun_qp;
1605 	int rx_buf_size, tx_buf_size;
1606 
1607 	if (qp_type > IB_QPT_GSI)
1608 		return -EINVAL;
1609 
1610 	tun_qp = &ctx->qp[qp_type];
1611 
1612 	tun_qp->ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1613 			       sizeof(struct mlx4_ib_buf),
1614 			       GFP_KERNEL);
1615 	if (!tun_qp->ring)
1616 		return -ENOMEM;
1617 
1618 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1619 				  sizeof (struct mlx4_ib_tun_tx_buf),
1620 				  GFP_KERNEL);
1621 	if (!tun_qp->tx_ring) {
1622 		kfree(tun_qp->ring);
1623 		tun_qp->ring = NULL;
1624 		return -ENOMEM;
1625 	}
1626 
1627 	if (is_tun) {
1628 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1629 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1630 	} else {
1631 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1632 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1633 	}
1634 
1635 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1636 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1637 		if (!tun_qp->ring[i].addr)
1638 			goto err;
1639 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1640 							tun_qp->ring[i].addr,
1641 							rx_buf_size,
1642 							DMA_FROM_DEVICE);
1643 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1644 			kfree(tun_qp->ring[i].addr);
1645 			goto err;
1646 		}
1647 	}
1648 
1649 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1650 		tun_qp->tx_ring[i].buf.addr =
1651 			kmalloc(tx_buf_size, GFP_KERNEL);
1652 		if (!tun_qp->tx_ring[i].buf.addr)
1653 			goto tx_err;
1654 		tun_qp->tx_ring[i].buf.map =
1655 			ib_dma_map_single(ctx->ib_dev,
1656 					  tun_qp->tx_ring[i].buf.addr,
1657 					  tx_buf_size,
1658 					  DMA_TO_DEVICE);
1659 		if (ib_dma_mapping_error(ctx->ib_dev,
1660 					 tun_qp->tx_ring[i].buf.map)) {
1661 			kfree(tun_qp->tx_ring[i].buf.addr);
1662 			goto tx_err;
1663 		}
1664 		tun_qp->tx_ring[i].ah = NULL;
1665 	}
1666 	spin_lock_init(&tun_qp->tx_lock);
1667 	tun_qp->tx_ix_head = 0;
1668 	tun_qp->tx_ix_tail = 0;
1669 	tun_qp->proxy_qpt = qp_type;
1670 
1671 	return 0;
1672 
1673 tx_err:
1674 	while (i > 0) {
1675 		--i;
1676 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1677 				    tx_buf_size, DMA_TO_DEVICE);
1678 		kfree(tun_qp->tx_ring[i].buf.addr);
1679 	}
1680 	i = MLX4_NUM_TUNNEL_BUFS;
1681 err:
1682 	while (i > 0) {
1683 		--i;
1684 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1685 				    rx_buf_size, DMA_FROM_DEVICE);
1686 		kfree(tun_qp->ring[i].addr);
1687 	}
1688 	kfree(tun_qp->tx_ring);
1689 	tun_qp->tx_ring = NULL;
1690 	kfree(tun_qp->ring);
1691 	tun_qp->ring = NULL;
1692 	return -ENOMEM;
1693 }
1694 
1695 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1696 				     enum ib_qp_type qp_type, int is_tun)
1697 {
1698 	int i;
1699 	struct mlx4_ib_demux_pv_qp *tun_qp;
1700 	int rx_buf_size, tx_buf_size;
1701 
1702 	if (qp_type > IB_QPT_GSI)
1703 		return;
1704 
1705 	tun_qp = &ctx->qp[qp_type];
1706 	if (is_tun) {
1707 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1708 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1709 	} else {
1710 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1711 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1712 	}
1713 
1714 
1715 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1716 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1717 				    rx_buf_size, DMA_FROM_DEVICE);
1718 		kfree(tun_qp->ring[i].addr);
1719 	}
1720 
1721 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1722 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1723 				    tx_buf_size, DMA_TO_DEVICE);
1724 		kfree(tun_qp->tx_ring[i].buf.addr);
1725 		if (tun_qp->tx_ring[i].ah)
1726 			rdma_destroy_ah(tun_qp->tx_ring[i].ah, 0);
1727 	}
1728 	kfree(tun_qp->tx_ring);
1729 	kfree(tun_qp->ring);
1730 }
1731 
1732 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1733 {
1734 	struct mlx4_ib_demux_pv_ctx *ctx;
1735 	struct mlx4_ib_demux_pv_qp *tun_qp;
1736 	struct ib_wc wc;
1737 	int ret;
1738 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1739 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1740 
1741 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1742 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1743 		if (wc.status == IB_WC_SUCCESS) {
1744 			switch (wc.opcode) {
1745 			case IB_WC_RECV:
1746 				mlx4_ib_multiplex_mad(ctx, &wc);
1747 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1748 							     wc.wr_id &
1749 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1750 				if (ret)
1751 					pr_err("Failed reposting tunnel "
1752 					       "buf:%lld\n", wc.wr_id);
1753 				break;
1754 			case IB_WC_SEND:
1755 				pr_debug("received tunnel send completion:"
1756 					 "wrid=0x%llx, status=0x%x\n",
1757 					 wc.wr_id, wc.status);
1758 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1759 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1760 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1761 					= NULL;
1762 				spin_lock(&tun_qp->tx_lock);
1763 				tun_qp->tx_ix_tail++;
1764 				spin_unlock(&tun_qp->tx_lock);
1765 
1766 				break;
1767 			default:
1768 				break;
1769 			}
1770 		} else  {
1771 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1772 				 " status = %d, wrid = 0x%llx\n",
1773 				 ctx->slave, wc.status, wc.wr_id);
1774 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1775 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1776 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1777 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1778 					= NULL;
1779 				spin_lock(&tun_qp->tx_lock);
1780 				tun_qp->tx_ix_tail++;
1781 				spin_unlock(&tun_qp->tx_lock);
1782 			}
1783 		}
1784 	}
1785 }
1786 
1787 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1788 {
1789 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1790 
1791 	/* It's worse than that! He's dead, Jim! */
1792 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1793 	       event->event, sqp->port);
1794 }
1795 
1796 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1797 			    enum ib_qp_type qp_type, int create_tun)
1798 {
1799 	int i, ret;
1800 	struct mlx4_ib_demux_pv_qp *tun_qp;
1801 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1802 	struct ib_qp_attr attr;
1803 	int qp_attr_mask_INIT;
1804 
1805 	if (qp_type > IB_QPT_GSI)
1806 		return -EINVAL;
1807 
1808 	tun_qp = &ctx->qp[qp_type];
1809 
1810 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1811 	qp_init_attr.init_attr.send_cq = ctx->cq;
1812 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1813 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1814 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1815 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1816 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1817 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1818 	if (create_tun) {
1819 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1820 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1821 		qp_init_attr.port = ctx->port;
1822 		qp_init_attr.slave = ctx->slave;
1823 		qp_init_attr.proxy_qp_type = qp_type;
1824 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1825 			   IB_QP_QKEY | IB_QP_PORT;
1826 	} else {
1827 		qp_init_attr.init_attr.qp_type = qp_type;
1828 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1829 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1830 	}
1831 	qp_init_attr.init_attr.port_num = ctx->port;
1832 	qp_init_attr.init_attr.qp_context = ctx;
1833 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1834 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1835 	if (IS_ERR(tun_qp->qp)) {
1836 		ret = PTR_ERR(tun_qp->qp);
1837 		tun_qp->qp = NULL;
1838 		pr_err("Couldn't create %s QP (%d)\n",
1839 		       create_tun ? "tunnel" : "special", ret);
1840 		return ret;
1841 	}
1842 
1843 	memset(&attr, 0, sizeof attr);
1844 	attr.qp_state = IB_QPS_INIT;
1845 	ret = 0;
1846 	if (create_tun)
1847 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1848 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1849 					      &attr.pkey_index);
1850 	if (ret || !create_tun)
1851 		attr.pkey_index =
1852 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1853 	attr.qkey = IB_QP1_QKEY;
1854 	attr.port_num = ctx->port;
1855 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1856 	if (ret) {
1857 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1858 		       create_tun ? "tunnel" : "special", ret);
1859 		goto err_qp;
1860 	}
1861 	attr.qp_state = IB_QPS_RTR;
1862 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1863 	if (ret) {
1864 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1865 		       create_tun ? "tunnel" : "special", ret);
1866 		goto err_qp;
1867 	}
1868 	attr.qp_state = IB_QPS_RTS;
1869 	attr.sq_psn = 0;
1870 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1871 	if (ret) {
1872 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1873 		       create_tun ? "tunnel" : "special", ret);
1874 		goto err_qp;
1875 	}
1876 
1877 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1878 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1879 		if (ret) {
1880 			pr_err(" mlx4_ib_post_pv_buf error"
1881 			       " (err = %d, i = %d)\n", ret, i);
1882 			goto err_qp;
1883 		}
1884 	}
1885 	return 0;
1886 
1887 err_qp:
1888 	ib_destroy_qp(tun_qp->qp);
1889 	tun_qp->qp = NULL;
1890 	return ret;
1891 }
1892 
1893 /*
1894  * IB MAD completion callback for real SQPs
1895  */
1896 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1897 {
1898 	struct mlx4_ib_demux_pv_ctx *ctx;
1899 	struct mlx4_ib_demux_pv_qp *sqp;
1900 	struct ib_wc wc;
1901 	struct ib_grh *grh;
1902 	struct ib_mad *mad;
1903 
1904 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1905 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1906 
1907 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1908 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1909 		if (wc.status == IB_WC_SUCCESS) {
1910 			switch (wc.opcode) {
1911 			case IB_WC_SEND:
1912 				kfree(sqp->tx_ring[wc.wr_id &
1913 				      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1914 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1915 					= NULL;
1916 				spin_lock(&sqp->tx_lock);
1917 				sqp->tx_ix_tail++;
1918 				spin_unlock(&sqp->tx_lock);
1919 				break;
1920 			case IB_WC_RECV:
1921 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1922 						(sqp->ring[wc.wr_id &
1923 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1924 				grh = &(((struct mlx4_mad_rcv_buf *)
1925 						(sqp->ring[wc.wr_id &
1926 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1927 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1928 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1929 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1930 					pr_err("Failed reposting SQP "
1931 					       "buf:%lld\n", wc.wr_id);
1932 				break;
1933 			default:
1934 				break;
1935 			}
1936 		} else  {
1937 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1938 				 " status = %d, wrid = 0x%llx\n",
1939 				 ctx->slave, wc.status, wc.wr_id);
1940 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1941 				kfree(sqp->tx_ring[wc.wr_id &
1942 				      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1943 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1944 					= NULL;
1945 				spin_lock(&sqp->tx_lock);
1946 				sqp->tx_ix_tail++;
1947 				spin_unlock(&sqp->tx_lock);
1948 			}
1949 		}
1950 	}
1951 }
1952 
1953 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1954 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1955 {
1956 	struct mlx4_ib_demux_pv_ctx *ctx;
1957 
1958 	*ret_ctx = NULL;
1959 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1960 	if (!ctx)
1961 		return -ENOMEM;
1962 
1963 	ctx->ib_dev = &dev->ib_dev;
1964 	ctx->port = port;
1965 	ctx->slave = slave;
1966 	*ret_ctx = ctx;
1967 	return 0;
1968 }
1969 
1970 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1971 {
1972 	if (dev->sriov.demux[port - 1].tun[slave]) {
1973 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1974 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1975 	}
1976 }
1977 
1978 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1979 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1980 {
1981 	int ret, cq_size;
1982 	struct ib_cq_init_attr cq_attr = {};
1983 
1984 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1985 		return -EEXIST;
1986 
1987 	ctx->state = DEMUX_PV_STATE_STARTING;
1988 	/* have QP0 only if link layer is IB */
1989 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1990 	    IB_LINK_LAYER_INFINIBAND)
1991 		ctx->has_smi = 1;
1992 
1993 	if (ctx->has_smi) {
1994 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1995 		if (ret) {
1996 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1997 			goto err_out;
1998 		}
1999 	}
2000 
2001 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
2002 	if (ret) {
2003 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
2004 		goto err_out_qp0;
2005 	}
2006 
2007 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
2008 	if (ctx->has_smi)
2009 		cq_size *= 2;
2010 
2011 	cq_attr.cqe = cq_size;
2012 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
2013 			       NULL, ctx, &cq_attr);
2014 	if (IS_ERR(ctx->cq)) {
2015 		ret = PTR_ERR(ctx->cq);
2016 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
2017 		goto err_buf;
2018 	}
2019 
2020 	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
2021 	if (IS_ERR(ctx->pd)) {
2022 		ret = PTR_ERR(ctx->pd);
2023 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
2024 		goto err_cq;
2025 	}
2026 
2027 	if (ctx->has_smi) {
2028 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
2029 		if (ret) {
2030 			pr_err("Couldn't create %s QP0 (%d)\n",
2031 			       create_tun ? "tunnel for" : "",  ret);
2032 			goto err_pd;
2033 		}
2034 	}
2035 
2036 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2037 	if (ret) {
2038 		pr_err("Couldn't create %s QP1 (%d)\n",
2039 		       create_tun ? "tunnel for" : "",  ret);
2040 		goto err_qp0;
2041 	}
2042 
2043 	if (create_tun)
2044 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2045 	else
2046 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2047 
2048 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2049 
2050 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2051 	if (ret) {
2052 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2053 		goto err_wq;
2054 	}
2055 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2056 	return 0;
2057 
2058 err_wq:
2059 	ctx->wq = NULL;
2060 	ib_destroy_qp(ctx->qp[1].qp);
2061 	ctx->qp[1].qp = NULL;
2062 
2063 
2064 err_qp0:
2065 	if (ctx->has_smi)
2066 		ib_destroy_qp(ctx->qp[0].qp);
2067 	ctx->qp[0].qp = NULL;
2068 
2069 err_pd:
2070 	ib_dealloc_pd(ctx->pd);
2071 	ctx->pd = NULL;
2072 
2073 err_cq:
2074 	ib_destroy_cq(ctx->cq);
2075 	ctx->cq = NULL;
2076 
2077 err_buf:
2078 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2079 
2080 err_out_qp0:
2081 	if (ctx->has_smi)
2082 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2083 err_out:
2084 	ctx->state = DEMUX_PV_STATE_DOWN;
2085 	return ret;
2086 }
2087 
2088 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2089 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2090 {
2091 	if (!ctx)
2092 		return;
2093 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2094 		ctx->state = DEMUX_PV_STATE_DOWNING;
2095 		if (flush)
2096 			flush_workqueue(ctx->wq);
2097 		if (ctx->has_smi) {
2098 			ib_destroy_qp(ctx->qp[0].qp);
2099 			ctx->qp[0].qp = NULL;
2100 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2101 		}
2102 		ib_destroy_qp(ctx->qp[1].qp);
2103 		ctx->qp[1].qp = NULL;
2104 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2105 		ib_dealloc_pd(ctx->pd);
2106 		ctx->pd = NULL;
2107 		ib_destroy_cq(ctx->cq);
2108 		ctx->cq = NULL;
2109 		ctx->state = DEMUX_PV_STATE_DOWN;
2110 	}
2111 }
2112 
2113 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2114 				  int port, int do_init)
2115 {
2116 	int ret = 0;
2117 
2118 	if (!do_init) {
2119 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2120 		/* for master, destroy real sqp resources */
2121 		if (slave == mlx4_master_func_num(dev->dev))
2122 			destroy_pv_resources(dev, slave, port,
2123 					     dev->sriov.sqps[port - 1], 1);
2124 		/* destroy the tunnel qp resources */
2125 		destroy_pv_resources(dev, slave, port,
2126 				     dev->sriov.demux[port - 1].tun[slave], 1);
2127 		return 0;
2128 	}
2129 
2130 	/* create the tunnel qp resources */
2131 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2132 				  dev->sriov.demux[port - 1].tun[slave]);
2133 
2134 	/* for master, create the real sqp resources */
2135 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2136 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2137 					  dev->sriov.sqps[port - 1]);
2138 	return ret;
2139 }
2140 
2141 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2142 {
2143 	struct mlx4_ib_demux_work *dmxw;
2144 
2145 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2146 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2147 			       dmxw->do_init);
2148 	kfree(dmxw);
2149 	return;
2150 }
2151 
2152 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2153 				       struct mlx4_ib_demux_ctx *ctx,
2154 				       int port)
2155 {
2156 	char name[12];
2157 	int ret = 0;
2158 	int i;
2159 
2160 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2161 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2162 	if (!ctx->tun)
2163 		return -ENOMEM;
2164 
2165 	ctx->dev = dev;
2166 	ctx->port = port;
2167 	ctx->ib_dev = &dev->ib_dev;
2168 
2169 	for (i = 0;
2170 	     i < min(dev->dev->caps.sqp_demux,
2171 	     (u16)(dev->dev->persist->num_vfs + 1));
2172 	     i++) {
2173 		struct mlx4_active_ports actv_ports =
2174 			mlx4_get_active_ports(dev->dev, i);
2175 
2176 		if (!test_bit(port - 1, actv_ports.ports))
2177 			continue;
2178 
2179 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2180 		if (ret) {
2181 			ret = -ENOMEM;
2182 			goto err_mcg;
2183 		}
2184 	}
2185 
2186 	ret = mlx4_ib_mcg_port_init(ctx);
2187 	if (ret) {
2188 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2189 		goto err_mcg;
2190 	}
2191 
2192 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2193 	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2194 	if (!ctx->wq) {
2195 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2196 		ret = -ENOMEM;
2197 		goto err_wq;
2198 	}
2199 
2200 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2201 	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2202 	if (!ctx->ud_wq) {
2203 		pr_err("Failed to create up/down WQ for port %d\n", port);
2204 		ret = -ENOMEM;
2205 		goto err_udwq;
2206 	}
2207 
2208 	return 0;
2209 
2210 err_udwq:
2211 	destroy_workqueue(ctx->wq);
2212 	ctx->wq = NULL;
2213 
2214 err_wq:
2215 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2216 err_mcg:
2217 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2218 		free_pv_object(dev, i, port);
2219 	kfree(ctx->tun);
2220 	ctx->tun = NULL;
2221 	return ret;
2222 }
2223 
2224 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2225 {
2226 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2227 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2228 		flush_workqueue(sqp_ctx->wq);
2229 		if (sqp_ctx->has_smi) {
2230 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2231 			sqp_ctx->qp[0].qp = NULL;
2232 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2233 		}
2234 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2235 		sqp_ctx->qp[1].qp = NULL;
2236 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2237 		ib_dealloc_pd(sqp_ctx->pd);
2238 		sqp_ctx->pd = NULL;
2239 		ib_destroy_cq(sqp_ctx->cq);
2240 		sqp_ctx->cq = NULL;
2241 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2242 	}
2243 }
2244 
2245 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2246 {
2247 	int i;
2248 	if (ctx) {
2249 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2250 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2251 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2252 			if (!ctx->tun[i])
2253 				continue;
2254 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2255 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2256 		}
2257 		flush_workqueue(ctx->wq);
2258 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2259 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2260 			free_pv_object(dev, i, ctx->port);
2261 		}
2262 		kfree(ctx->tun);
2263 		destroy_workqueue(ctx->ud_wq);
2264 		destroy_workqueue(ctx->wq);
2265 	}
2266 }
2267 
2268 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2269 {
2270 	int i;
2271 
2272 	if (!mlx4_is_master(dev->dev))
2273 		return;
2274 	/* initialize or tear down tunnel QPs for the master */
2275 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2276 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2277 	return;
2278 }
2279 
2280 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2281 {
2282 	int i = 0;
2283 	int err;
2284 
2285 	if (!mlx4_is_mfunc(dev->dev))
2286 		return 0;
2287 
2288 	dev->sriov.is_going_down = 0;
2289 	spin_lock_init(&dev->sriov.going_down_lock);
2290 	mlx4_ib_cm_paravirt_init(dev);
2291 
2292 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2293 
2294 	if (mlx4_is_slave(dev->dev)) {
2295 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2296 		return 0;
2297 	}
2298 
2299 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2300 		if (i == mlx4_master_func_num(dev->dev))
2301 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2302 		else
2303 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2304 	}
2305 
2306 	err = mlx4_ib_init_alias_guid_service(dev);
2307 	if (err) {
2308 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2309 		goto paravirt_err;
2310 	}
2311 	err = mlx4_ib_device_register_sysfs(dev);
2312 	if (err) {
2313 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2314 		goto sysfs_err;
2315 	}
2316 
2317 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2318 		     dev->dev->caps.sqp_demux);
2319 	for (i = 0; i < dev->num_ports; i++) {
2320 		union ib_gid gid;
2321 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2322 		if (err)
2323 			goto demux_err;
2324 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2325 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2326 			     be64_to_cpu(gid.global.subnet_prefix));
2327 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2328 				      &dev->sriov.sqps[i]);
2329 		if (err)
2330 			goto demux_err;
2331 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2332 		if (err)
2333 			goto free_pv;
2334 	}
2335 	mlx4_ib_master_tunnels(dev, 1);
2336 	return 0;
2337 
2338 free_pv:
2339 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2340 demux_err:
2341 	while (--i >= 0) {
2342 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2343 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2344 	}
2345 	mlx4_ib_device_unregister_sysfs(dev);
2346 
2347 sysfs_err:
2348 	mlx4_ib_destroy_alias_guid_service(dev);
2349 
2350 paravirt_err:
2351 	mlx4_ib_cm_paravirt_clean(dev, -1);
2352 
2353 	return err;
2354 }
2355 
2356 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2357 {
2358 	int i;
2359 	unsigned long flags;
2360 
2361 	if (!mlx4_is_mfunc(dev->dev))
2362 		return;
2363 
2364 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2365 	dev->sriov.is_going_down = 1;
2366 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2367 	if (mlx4_is_master(dev->dev)) {
2368 		for (i = 0; i < dev->num_ports; i++) {
2369 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2370 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2371 			kfree(dev->sriov.sqps[i]);
2372 			dev->sriov.sqps[i] = NULL;
2373 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2374 		}
2375 
2376 		mlx4_ib_cm_paravirt_clean(dev, -1);
2377 		mlx4_ib_destroy_alias_guid_service(dev);
2378 		mlx4_ib_device_unregister_sysfs(dev);
2379 	}
2380 }
2381