xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision 82e6fdd6)
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);
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]);
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 	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);
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);
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);
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->qp_num) {
811 		pr_debug("received MAD: slid:%d sqpn:%d "
812 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
813 			in_wc->slid, in_wc->src_qp,
814 			in_wc->dlid_path_bits,
815 			in_wc->qp->qp_num,
816 			in_wc->wc_flags,
817 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
818 			be16_to_cpu(in_mad->mad_hdr.attr_id));
819 		if (in_wc->wc_flags & IB_WC_GRH) {
820 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
821 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
822 				 be64_to_cpu(in_grh->sgid.global.interface_id));
823 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
824 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
825 				 be64_to_cpu(in_grh->dgid.global.interface_id));
826 		}
827 	}
828 
829 	slid = in_wc ? ib_lid_cpu16(in_wc->slid) : be16_to_cpu(IB_LID_PERMISSIVE);
830 
831 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
832 		forward_trap(to_mdev(ibdev), port_num, in_mad);
833 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
834 	}
835 
836 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
837 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
838 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
839 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
840 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
841 			return IB_MAD_RESULT_SUCCESS;
842 
843 		/*
844 		 * Don't process SMInfo queries -- the SMA can't handle them.
845 		 */
846 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
847 			return IB_MAD_RESULT_SUCCESS;
848 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
849 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
850 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
851 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
852 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
853 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
854 			return IB_MAD_RESULT_SUCCESS;
855 	} else
856 		return IB_MAD_RESULT_SUCCESS;
857 
858 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
859 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
860 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
861 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
862 	    !ib_query_port(ibdev, port_num, &pattr))
863 		prev_lid = ib_lid_cpu16(pattr.lid);
864 
865 	err = mlx4_MAD_IFC(to_mdev(ibdev),
866 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
867 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
868 			   MLX4_MAD_IFC_NET_VIEW,
869 			   port_num, in_wc, in_grh, in_mad, out_mad);
870 	if (err)
871 		return IB_MAD_RESULT_FAILURE;
872 
873 	if (!out_mad->mad_hdr.status) {
874 		smp_snoop(ibdev, port_num, in_mad, prev_lid);
875 		/* slaves get node desc from FW */
876 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
877 			node_desc_override(ibdev, out_mad);
878 	}
879 
880 	/* set return bit in status of directed route responses */
881 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
882 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
883 
884 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
885 		/* no response for trap repress */
886 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
887 
888 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
889 }
890 
891 static void edit_counter(struct mlx4_counter *cnt, void *counters,
892 			 __be16 attr_id)
893 {
894 	switch (attr_id) {
895 	case IB_PMA_PORT_COUNTERS:
896 	{
897 		struct ib_pma_portcounters *pma_cnt =
898 			(struct ib_pma_portcounters *)counters;
899 
900 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
901 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
902 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
903 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
904 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
905 				     be64_to_cpu(cnt->tx_frames));
906 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
907 				     be64_to_cpu(cnt->rx_frames));
908 		break;
909 	}
910 	case IB_PMA_PORT_COUNTERS_EXT:
911 	{
912 		struct ib_pma_portcounters_ext *pma_cnt_ext =
913 			(struct ib_pma_portcounters_ext *)counters;
914 
915 		pma_cnt_ext->port_xmit_data =
916 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
917 		pma_cnt_ext->port_rcv_data =
918 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
919 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
920 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
921 		break;
922 	}
923 	}
924 }
925 
926 static int iboe_process_mad_port_info(void *out_mad)
927 {
928 	struct ib_class_port_info cpi = {};
929 
930 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
931 	memcpy(out_mad, &cpi, sizeof(cpi));
932 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
933 }
934 
935 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
936 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
937 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
938 {
939 	struct mlx4_counter counter_stats;
940 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
941 	struct counter_index *tmp_counter;
942 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
943 
944 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
945 		return -EINVAL;
946 
947 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
948 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
949 
950 	memset(&counter_stats, 0, sizeof(counter_stats));
951 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
952 	list_for_each_entry(tmp_counter,
953 			    &dev->counters_table[port_num - 1].counters_list,
954 			    list) {
955 		err = mlx4_get_counter_stats(dev->dev,
956 					     tmp_counter->index,
957 					     &counter_stats, 0);
958 		if (err) {
959 			err = IB_MAD_RESULT_FAILURE;
960 			stats_avail = 0;
961 			break;
962 		}
963 		stats_avail = 1;
964 	}
965 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
966 	if (stats_avail) {
967 		memset(out_mad->data, 0, sizeof out_mad->data);
968 		switch (counter_stats.counter_mode & 0xf) {
969 		case 0:
970 			edit_counter(&counter_stats,
971 				     (void *)(out_mad->data + 40),
972 				     in_mad->mad_hdr.attr_id);
973 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
974 			break;
975 		default:
976 			err = IB_MAD_RESULT_FAILURE;
977 		}
978 	}
979 
980 	return err;
981 }
982 
983 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
984 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
985 			const struct ib_mad_hdr *in, size_t in_mad_size,
986 			struct ib_mad_hdr *out, size_t *out_mad_size,
987 			u16 *out_mad_pkey_index)
988 {
989 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
990 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
991 	struct ib_mad *out_mad = (struct ib_mad *)out;
992 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
993 
994 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
995 			 *out_mad_size != sizeof(*out_mad)))
996 		return IB_MAD_RESULT_FAILURE;
997 
998 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
999 	 * queries, should be called only by VFs and for that specific purpose
1000 	 */
1001 	if (link == IB_LINK_LAYER_INFINIBAND) {
1002 		if (mlx4_is_slave(dev->dev) &&
1003 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
1004 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
1005 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
1006 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
1007 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1008 						in_grh, in_mad, out_mad);
1009 
1010 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
1011 				      in_grh, in_mad, out_mad);
1012 	}
1013 
1014 	if (link == IB_LINK_LAYER_ETHERNET)
1015 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1016 					in_grh, in_mad, out_mad);
1017 
1018 	return -EINVAL;
1019 }
1020 
1021 static void send_handler(struct ib_mad_agent *agent,
1022 			 struct ib_mad_send_wc *mad_send_wc)
1023 {
1024 	if (mad_send_wc->send_buf->context[0])
1025 		rdma_destroy_ah(mad_send_wc->send_buf->context[0]);
1026 	ib_free_send_mad(mad_send_wc->send_buf);
1027 }
1028 
1029 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1030 {
1031 	struct ib_mad_agent *agent;
1032 	int p, q;
1033 	int ret;
1034 	enum rdma_link_layer ll;
1035 
1036 	for (p = 0; p < dev->num_ports; ++p) {
1037 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1038 		for (q = 0; q <= 1; ++q) {
1039 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1040 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1041 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1042 							      NULL, 0, send_handler,
1043 							      NULL, NULL, 0);
1044 				if (IS_ERR(agent)) {
1045 					ret = PTR_ERR(agent);
1046 					goto err;
1047 				}
1048 				dev->send_agent[p][q] = agent;
1049 			} else
1050 				dev->send_agent[p][q] = NULL;
1051 		}
1052 	}
1053 
1054 	return 0;
1055 
1056 err:
1057 	for (p = 0; p < dev->num_ports; ++p)
1058 		for (q = 0; q <= 1; ++q)
1059 			if (dev->send_agent[p][q])
1060 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1061 
1062 	return ret;
1063 }
1064 
1065 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1066 {
1067 	struct ib_mad_agent *agent;
1068 	int p, q;
1069 
1070 	for (p = 0; p < dev->num_ports; ++p) {
1071 		for (q = 0; q <= 1; ++q) {
1072 			agent = dev->send_agent[p][q];
1073 			if (agent) {
1074 				dev->send_agent[p][q] = NULL;
1075 				ib_unregister_mad_agent(agent);
1076 			}
1077 		}
1078 
1079 		if (dev->sm_ah[p])
1080 			rdma_destroy_ah(dev->sm_ah[p]);
1081 	}
1082 }
1083 
1084 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1085 {
1086 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1087 
1088 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1089 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1090 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1091 }
1092 
1093 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1094 {
1095 	/* re-configure the alias-guid and mcg's */
1096 	if (mlx4_is_master(dev->dev)) {
1097 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1098 
1099 		if (!dev->sriov.is_going_down) {
1100 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1101 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1102 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1103 		}
1104 	}
1105 
1106 	/* Update the sl to vl table from inside client rereg
1107 	 * only if in secure-host mode (snooping is not possible)
1108 	 * and the sl-to-vl change event is not generated by FW.
1109 	 */
1110 	if (!mlx4_is_slave(dev->dev) &&
1111 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1112 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1113 		if (mlx4_is_master(dev->dev))
1114 			/* already in work queue from mlx4_ib_event queueing
1115 			 * mlx4_handle_port_mgmt_change_event, which calls
1116 			 * this procedure. Therefore, call sl2vl_update directly.
1117 			 */
1118 			mlx4_ib_sl2vl_update(dev, port_num);
1119 		else
1120 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1121 	}
1122 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1123 }
1124 
1125 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1126 			      struct mlx4_eqe *eqe)
1127 {
1128 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1129 			    GET_MASK_FROM_EQE(eqe));
1130 }
1131 
1132 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1133 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1134 {
1135 	struct ib_smp *in_mad  = NULL;
1136 	struct ib_smp *out_mad  = NULL;
1137 	u16 i;
1138 
1139 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1140 		return;
1141 
1142 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1143 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1144 	if (!in_mad || !out_mad)
1145 		goto out;
1146 
1147 	guid_tbl_blk_num  *= 4;
1148 
1149 	for (i = 0; i < 4; i++) {
1150 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1151 			continue;
1152 		memset(in_mad, 0, sizeof *in_mad);
1153 		memset(out_mad, 0, sizeof *out_mad);
1154 
1155 		in_mad->base_version  = 1;
1156 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1157 		in_mad->class_version = 1;
1158 		in_mad->method        = IB_MGMT_METHOD_GET;
1159 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1160 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1161 
1162 		if (mlx4_MAD_IFC(dev,
1163 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1164 				 port_num, NULL, NULL, in_mad, out_mad)) {
1165 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1166 			goto out;
1167 		}
1168 
1169 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1170 						    port_num,
1171 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1172 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1173 						     port_num,
1174 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1175 	}
1176 
1177 out:
1178 	kfree(in_mad);
1179 	kfree(out_mad);
1180 	return;
1181 }
1182 
1183 void handle_port_mgmt_change_event(struct work_struct *work)
1184 {
1185 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1186 	struct mlx4_ib_dev *dev = ew->ib_dev;
1187 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1188 	u8 port = eqe->event.port_mgmt_change.port;
1189 	u32 changed_attr;
1190 	u32 tbl_block;
1191 	u32 change_bitmap;
1192 
1193 	switch (eqe->subtype) {
1194 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1195 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1196 
1197 		/* Update the SM ah - This should be done before handling
1198 		   the other changed attributes so that MADs can be sent to the SM */
1199 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1200 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1201 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1202 			update_sm_ah(dev, port, lid, sl);
1203 		}
1204 
1205 		/* Check if it is a lid change event */
1206 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1207 			handle_lid_change_event(dev, port);
1208 
1209 		/* Generate GUID changed event */
1210 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1211 			if (mlx4_is_master(dev->dev)) {
1212 				union ib_gid gid;
1213 				int err = 0;
1214 
1215 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1216 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1217 				else
1218 					gid.global.subnet_prefix =
1219 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1220 				if (err) {
1221 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1222 						port, err);
1223 				} else {
1224 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1225 						 port,
1226 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1227 						 be64_to_cpu(gid.global.subnet_prefix));
1228 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1229 						     be64_to_cpu(gid.global.subnet_prefix));
1230 				}
1231 			}
1232 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1233 			/*if master, notify all slaves*/
1234 			if (mlx4_is_master(dev->dev))
1235 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1236 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1237 		}
1238 
1239 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1240 			handle_client_rereg_event(dev, port);
1241 		break;
1242 
1243 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1244 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1245 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1246 			propagate_pkey_ev(dev, port, eqe);
1247 		break;
1248 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1249 		/* paravirtualized master's guid is guid 0 -- does not change */
1250 		if (!mlx4_is_master(dev->dev))
1251 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1252 		/*if master, notify relevant slaves*/
1253 		else if (!dev->sriov.is_going_down) {
1254 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1255 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1256 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1257 		}
1258 		break;
1259 
1260 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1261 		/* cache sl to vl mapping changes for use in
1262 		 * filling QP1 LRH VL field when sending packets
1263 		 */
1264 		if (!mlx4_is_slave(dev->dev)) {
1265 			union sl2vl_tbl_to_u64 sl2vl64;
1266 			int jj;
1267 
1268 			for (jj = 0; jj < 8; jj++) {
1269 				sl2vl64.sl8[jj] =
1270 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1271 				pr_debug("port %u, sl2vl[%d] = %02x\n",
1272 					 port, jj, sl2vl64.sl8[jj]);
1273 			}
1274 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1275 		}
1276 		break;
1277 	default:
1278 		pr_warn("Unsupported subtype 0x%x for "
1279 			"Port Management Change event\n", eqe->subtype);
1280 	}
1281 
1282 	kfree(ew);
1283 }
1284 
1285 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1286 			    enum ib_event_type type)
1287 {
1288 	struct ib_event event;
1289 
1290 	event.device		= &dev->ib_dev;
1291 	event.element.port_num	= port_num;
1292 	event.event		= type;
1293 
1294 	ib_dispatch_event(&event);
1295 }
1296 
1297 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1298 {
1299 	unsigned long flags;
1300 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1301 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1302 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1303 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1304 		queue_work(ctx->wq, &ctx->work);
1305 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1306 }
1307 
1308 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1309 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1310 				  int index)
1311 {
1312 	struct ib_sge sg_list;
1313 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1314 	int size;
1315 
1316 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1317 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1318 
1319 	sg_list.addr = tun_qp->ring[index].map;
1320 	sg_list.length = size;
1321 	sg_list.lkey = ctx->pd->local_dma_lkey;
1322 
1323 	recv_wr.next = NULL;
1324 	recv_wr.sg_list = &sg_list;
1325 	recv_wr.num_sge = 1;
1326 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1327 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1328 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1329 				      size, DMA_FROM_DEVICE);
1330 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1331 }
1332 
1333 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1334 		int slave, struct ib_sa_mad *sa_mad)
1335 {
1336 	int ret = 0;
1337 
1338 	/* dispatch to different sa handlers */
1339 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1340 	case IB_SA_ATTR_MC_MEMBER_REC:
1341 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1342 		break;
1343 	default:
1344 		break;
1345 	}
1346 	return ret;
1347 }
1348 
1349 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1350 {
1351 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1352 
1353 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1354 }
1355 
1356 
1357 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1358 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1359 			 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
1360 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1361 {
1362 	struct ib_sge list;
1363 	struct ib_ud_wr wr;
1364 	struct ib_send_wr *bad_wr;
1365 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1366 	struct mlx4_ib_demux_pv_qp *sqp;
1367 	struct mlx4_mad_snd_buf *sqp_mad;
1368 	struct ib_ah *ah;
1369 	struct ib_qp *send_qp = NULL;
1370 	struct ib_global_route *grh;
1371 	unsigned wire_tx_ix = 0;
1372 	int ret = 0;
1373 	u16 wire_pkey_ix;
1374 	int src_qpnum;
1375 	u8 sgid_index;
1376 
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 	/* create ah */
1397 	grh = rdma_ah_retrieve_grh(attr);
1398 	sgid_index = grh->sgid_index;
1399 	grh->sgid_index = 0;
1400 	ah = rdma_create_ah(sqp_ctx->pd, attr);
1401 	if (IS_ERR(ah))
1402 		return -ENOMEM;
1403 	grh->sgid_index = sgid_index;
1404 	to_mah(ah)->av.ib.gid_index = sgid_index;
1405 	/* get rid of force-loopback bit */
1406 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1407 	spin_lock(&sqp->tx_lock);
1408 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1409 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1410 		ret = -EAGAIN;
1411 	else
1412 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1413 	spin_unlock(&sqp->tx_lock);
1414 	if (ret)
1415 		goto out;
1416 
1417 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1418 	if (sqp->tx_ring[wire_tx_ix].ah)
1419 		rdma_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1420 	sqp->tx_ring[wire_tx_ix].ah = ah;
1421 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1422 				   sqp->tx_ring[wire_tx_ix].buf.map,
1423 				   sizeof (struct mlx4_mad_snd_buf),
1424 				   DMA_TO_DEVICE);
1425 
1426 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1427 
1428 	ib_dma_sync_single_for_device(&dev->ib_dev,
1429 				      sqp->tx_ring[wire_tx_ix].buf.map,
1430 				      sizeof (struct mlx4_mad_snd_buf),
1431 				      DMA_TO_DEVICE);
1432 
1433 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1434 	list.length = sizeof (struct mlx4_mad_snd_buf);
1435 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1436 
1437 	wr.ah = ah;
1438 	wr.port_num = port;
1439 	wr.pkey_index = wire_pkey_ix;
1440 	wr.remote_qkey = qkey;
1441 	wr.remote_qpn = remote_qpn;
1442 	wr.wr.next = NULL;
1443 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1444 	wr.wr.sg_list = &list;
1445 	wr.wr.num_sge = 1;
1446 	wr.wr.opcode = IB_WR_SEND;
1447 	wr.wr.send_flags = IB_SEND_SIGNALED;
1448 	if (s_mac)
1449 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1450 	if (vlan_id < 0x1000)
1451 		vlan_id |= (rdma_ah_get_sl(attr) & 7) << 13;
1452 	to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1453 
1454 
1455 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1456 	if (!ret)
1457 		return 0;
1458 
1459 	spin_lock(&sqp->tx_lock);
1460 	sqp->tx_ix_tail++;
1461 	spin_unlock(&sqp->tx_lock);
1462 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1463 out:
1464 	rdma_destroy_ah(ah);
1465 	return ret;
1466 }
1467 
1468 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1469 {
1470 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1471 		return slave;
1472 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1473 }
1474 
1475 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1476 				    struct rdma_ah_attr *ah_attr)
1477 {
1478 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
1479 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1480 		grh->sgid_index = slave;
1481 	else
1482 		grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
1483 }
1484 
1485 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1486 {
1487 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1488 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1489 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1490 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1491 	struct mlx4_ib_ah ah;
1492 	struct rdma_ah_attr ah_attr;
1493 	u8 *slave_id;
1494 	int slave;
1495 	int port;
1496 	u16 vlan_id;
1497 	u8 qos;
1498 	u8 *dmac;
1499 
1500 	/* Get slave that sent this packet */
1501 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1502 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1503 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1504 	    wc->src_qp & 0x4) {
1505 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1506 		return;
1507 	}
1508 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1509 	if (slave != ctx->slave) {
1510 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1511 			     "belongs to another slave\n", wc->src_qp);
1512 		return;
1513 	}
1514 
1515 	/* Map transaction ID */
1516 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1517 				   sizeof (struct mlx4_tunnel_mad),
1518 				   DMA_FROM_DEVICE);
1519 	switch (tunnel->mad.mad_hdr.method) {
1520 	case IB_MGMT_METHOD_SET:
1521 	case IB_MGMT_METHOD_GET:
1522 	case IB_MGMT_METHOD_REPORT:
1523 	case IB_SA_METHOD_GET_TABLE:
1524 	case IB_SA_METHOD_DELETE:
1525 	case IB_SA_METHOD_GET_MULTI:
1526 	case IB_SA_METHOD_GET_TRACE_TBL:
1527 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1528 		if (*slave_id) {
1529 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1530 				     "class:%d slave:%d\n", *slave_id,
1531 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1532 			return;
1533 		} else
1534 			*slave_id = slave;
1535 	default:
1536 		/* nothing */;
1537 	}
1538 
1539 	/* Class-specific handling */
1540 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1541 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1542 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1543 		if (slave != mlx4_master_func_num(dev->dev) &&
1544 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1545 			return;
1546 		break;
1547 	case IB_MGMT_CLASS_SUBN_ADM:
1548 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1549 			      (struct ib_sa_mad *) &tunnel->mad))
1550 			return;
1551 		break;
1552 	case IB_MGMT_CLASS_CM:
1553 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1554 			      (struct ib_mad *) &tunnel->mad))
1555 			return;
1556 		break;
1557 	case IB_MGMT_CLASS_DEVICE_MGMT:
1558 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1559 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1560 			return;
1561 		break;
1562 	default:
1563 		/* Drop unsupported classes for slaves in tunnel mode */
1564 		if (slave != mlx4_master_func_num(dev->dev)) {
1565 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1566 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1567 			return;
1568 		}
1569 	}
1570 
1571 	/* We are using standard ib_core services to send the mad, so generate a
1572 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1573 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1574 	ah.ibah.device = ctx->ib_dev;
1575 
1576 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1577 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1578 	if (port < 0)
1579 		return;
1580 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1581 	ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
1582 
1583 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1584 	if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
1585 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1586 	dmac = rdma_ah_retrieve_dmac(&ah_attr);
1587 	if (dmac)
1588 		memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
1589 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1590 	/* if slave have default vlan use it */
1591 	if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1592 					&vlan_id, &qos))
1593 		rdma_ah_set_sl(&ah_attr, qos);
1594 
1595 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1596 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1597 			     IB_QPT_SMI : IB_QPT_GSI,
1598 			     be16_to_cpu(tunnel->hdr.pkey_index),
1599 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1600 			     be32_to_cpu(tunnel->hdr.qkey),
1601 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1602 }
1603 
1604 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1605 				 enum ib_qp_type qp_type, int is_tun)
1606 {
1607 	int i;
1608 	struct mlx4_ib_demux_pv_qp *tun_qp;
1609 	int rx_buf_size, tx_buf_size;
1610 
1611 	if (qp_type > IB_QPT_GSI)
1612 		return -EINVAL;
1613 
1614 	tun_qp = &ctx->qp[qp_type];
1615 
1616 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1617 			       GFP_KERNEL);
1618 	if (!tun_qp->ring)
1619 		return -ENOMEM;
1620 
1621 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_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 < MLX4_NUM_TUNNEL_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 < MLX4_NUM_TUNNEL_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 	kfree(tun_qp->tx_ring);
1684 	tun_qp->tx_ring = NULL;
1685 	i = MLX4_NUM_TUNNEL_BUFS;
1686 err:
1687 	while (i > 0) {
1688 		--i;
1689 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1690 				    rx_buf_size, DMA_FROM_DEVICE);
1691 		kfree(tun_qp->ring[i].addr);
1692 	}
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 
1705 	if (qp_type > IB_QPT_GSI)
1706 		return;
1707 
1708 	tun_qp = &ctx->qp[qp_type];
1709 	if (is_tun) {
1710 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1711 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1712 	} else {
1713 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1714 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1715 	}
1716 
1717 
1718 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1719 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1720 				    rx_buf_size, DMA_FROM_DEVICE);
1721 		kfree(tun_qp->ring[i].addr);
1722 	}
1723 
1724 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1725 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1726 				    tx_buf_size, DMA_TO_DEVICE);
1727 		kfree(tun_qp->tx_ring[i].buf.addr);
1728 		if (tun_qp->tx_ring[i].ah)
1729 			rdma_destroy_ah(tun_qp->tx_ring[i].ah);
1730 	}
1731 	kfree(tun_qp->tx_ring);
1732 	kfree(tun_qp->ring);
1733 }
1734 
1735 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1736 {
1737 	struct mlx4_ib_demux_pv_ctx *ctx;
1738 	struct mlx4_ib_demux_pv_qp *tun_qp;
1739 	struct ib_wc wc;
1740 	int ret;
1741 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1742 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1743 
1744 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1745 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1746 		if (wc.status == IB_WC_SUCCESS) {
1747 			switch (wc.opcode) {
1748 			case IB_WC_RECV:
1749 				mlx4_ib_multiplex_mad(ctx, &wc);
1750 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1751 							     wc.wr_id &
1752 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1753 				if (ret)
1754 					pr_err("Failed reposting tunnel "
1755 					       "buf:%lld\n", wc.wr_id);
1756 				break;
1757 			case IB_WC_SEND:
1758 				pr_debug("received tunnel send completion:"
1759 					 "wrid=0x%llx, status=0x%x\n",
1760 					 wc.wr_id, wc.status);
1761 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1762 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1763 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1764 					= NULL;
1765 				spin_lock(&tun_qp->tx_lock);
1766 				tun_qp->tx_ix_tail++;
1767 				spin_unlock(&tun_qp->tx_lock);
1768 
1769 				break;
1770 			default:
1771 				break;
1772 			}
1773 		} else  {
1774 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1775 				 " status = %d, wrid = 0x%llx\n",
1776 				 ctx->slave, wc.status, wc.wr_id);
1777 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1778 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1779 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1780 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1781 					= NULL;
1782 				spin_lock(&tun_qp->tx_lock);
1783 				tun_qp->tx_ix_tail++;
1784 				spin_unlock(&tun_qp->tx_lock);
1785 			}
1786 		}
1787 	}
1788 }
1789 
1790 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1791 {
1792 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1793 
1794 	/* It's worse than that! He's dead, Jim! */
1795 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1796 	       event->event, sqp->port);
1797 }
1798 
1799 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1800 			    enum ib_qp_type qp_type, int create_tun)
1801 {
1802 	int i, ret;
1803 	struct mlx4_ib_demux_pv_qp *tun_qp;
1804 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1805 	struct ib_qp_attr attr;
1806 	int qp_attr_mask_INIT;
1807 
1808 	if (qp_type > IB_QPT_GSI)
1809 		return -EINVAL;
1810 
1811 	tun_qp = &ctx->qp[qp_type];
1812 
1813 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1814 	qp_init_attr.init_attr.send_cq = ctx->cq;
1815 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1816 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1817 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1818 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1819 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1820 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1821 	if (create_tun) {
1822 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1823 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1824 		qp_init_attr.port = ctx->port;
1825 		qp_init_attr.slave = ctx->slave;
1826 		qp_init_attr.proxy_qp_type = qp_type;
1827 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1828 			   IB_QP_QKEY | IB_QP_PORT;
1829 	} else {
1830 		qp_init_attr.init_attr.qp_type = qp_type;
1831 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1832 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1833 	}
1834 	qp_init_attr.init_attr.port_num = ctx->port;
1835 	qp_init_attr.init_attr.qp_context = ctx;
1836 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1837 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1838 	if (IS_ERR(tun_qp->qp)) {
1839 		ret = PTR_ERR(tun_qp->qp);
1840 		tun_qp->qp = NULL;
1841 		pr_err("Couldn't create %s QP (%d)\n",
1842 		       create_tun ? "tunnel" : "special", ret);
1843 		return ret;
1844 	}
1845 
1846 	memset(&attr, 0, sizeof attr);
1847 	attr.qp_state = IB_QPS_INIT;
1848 	ret = 0;
1849 	if (create_tun)
1850 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1851 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1852 					      &attr.pkey_index);
1853 	if (ret || !create_tun)
1854 		attr.pkey_index =
1855 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1856 	attr.qkey = IB_QP1_QKEY;
1857 	attr.port_num = ctx->port;
1858 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1859 	if (ret) {
1860 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1861 		       create_tun ? "tunnel" : "special", ret);
1862 		goto err_qp;
1863 	}
1864 	attr.qp_state = IB_QPS_RTR;
1865 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1866 	if (ret) {
1867 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1868 		       create_tun ? "tunnel" : "special", ret);
1869 		goto err_qp;
1870 	}
1871 	attr.qp_state = IB_QPS_RTS;
1872 	attr.sq_psn = 0;
1873 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1874 	if (ret) {
1875 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1876 		       create_tun ? "tunnel" : "special", ret);
1877 		goto err_qp;
1878 	}
1879 
1880 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1881 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1882 		if (ret) {
1883 			pr_err(" mlx4_ib_post_pv_buf error"
1884 			       " (err = %d, i = %d)\n", ret, i);
1885 			goto err_qp;
1886 		}
1887 	}
1888 	return 0;
1889 
1890 err_qp:
1891 	ib_destroy_qp(tun_qp->qp);
1892 	tun_qp->qp = NULL;
1893 	return ret;
1894 }
1895 
1896 /*
1897  * IB MAD completion callback for real SQPs
1898  */
1899 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1900 {
1901 	struct mlx4_ib_demux_pv_ctx *ctx;
1902 	struct mlx4_ib_demux_pv_qp *sqp;
1903 	struct ib_wc wc;
1904 	struct ib_grh *grh;
1905 	struct ib_mad *mad;
1906 
1907 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1908 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1909 
1910 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1911 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1912 		if (wc.status == IB_WC_SUCCESS) {
1913 			switch (wc.opcode) {
1914 			case IB_WC_SEND:
1915 				rdma_destroy_ah(sqp->tx_ring[wc.wr_id &
1916 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1917 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1918 					= NULL;
1919 				spin_lock(&sqp->tx_lock);
1920 				sqp->tx_ix_tail++;
1921 				spin_unlock(&sqp->tx_lock);
1922 				break;
1923 			case IB_WC_RECV:
1924 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1925 						(sqp->ring[wc.wr_id &
1926 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1927 				grh = &(((struct mlx4_mad_rcv_buf *)
1928 						(sqp->ring[wc.wr_id &
1929 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1930 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1931 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1932 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1933 					pr_err("Failed reposting SQP "
1934 					       "buf:%lld\n", wc.wr_id);
1935 				break;
1936 			default:
1937 				BUG_ON(1);
1938 				break;
1939 			}
1940 		} else  {
1941 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1942 				 " status = %d, wrid = 0x%llx\n",
1943 				 ctx->slave, wc.status, wc.wr_id);
1944 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1945 				rdma_destroy_ah(sqp->tx_ring[wc.wr_id &
1946 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1947 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1948 					= NULL;
1949 				spin_lock(&sqp->tx_lock);
1950 				sqp->tx_ix_tail++;
1951 				spin_unlock(&sqp->tx_lock);
1952 			}
1953 		}
1954 	}
1955 }
1956 
1957 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1958 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1959 {
1960 	struct mlx4_ib_demux_pv_ctx *ctx;
1961 
1962 	*ret_ctx = NULL;
1963 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1964 	if (!ctx)
1965 		return -ENOMEM;
1966 
1967 	ctx->ib_dev = &dev->ib_dev;
1968 	ctx->port = port;
1969 	ctx->slave = slave;
1970 	*ret_ctx = ctx;
1971 	return 0;
1972 }
1973 
1974 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1975 {
1976 	if (dev->sriov.demux[port - 1].tun[slave]) {
1977 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1978 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1979 	}
1980 }
1981 
1982 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1983 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1984 {
1985 	int ret, cq_size;
1986 	struct ib_cq_init_attr cq_attr = {};
1987 
1988 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1989 		return -EEXIST;
1990 
1991 	ctx->state = DEMUX_PV_STATE_STARTING;
1992 	/* have QP0 only if link layer is IB */
1993 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1994 	    IB_LINK_LAYER_INFINIBAND)
1995 		ctx->has_smi = 1;
1996 
1997 	if (ctx->has_smi) {
1998 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1999 		if (ret) {
2000 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
2001 			goto err_out;
2002 		}
2003 	}
2004 
2005 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
2006 	if (ret) {
2007 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
2008 		goto err_out_qp0;
2009 	}
2010 
2011 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
2012 	if (ctx->has_smi)
2013 		cq_size *= 2;
2014 
2015 	cq_attr.cqe = cq_size;
2016 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_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 
2054 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2055 	if (ret) {
2056 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2057 		goto err_wq;
2058 	}
2059 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2060 	return 0;
2061 
2062 err_wq:
2063 	ctx->wq = NULL;
2064 	ib_destroy_qp(ctx->qp[1].qp);
2065 	ctx->qp[1].qp = NULL;
2066 
2067 
2068 err_qp0:
2069 	if (ctx->has_smi)
2070 		ib_destroy_qp(ctx->qp[0].qp);
2071 	ctx->qp[0].qp = NULL;
2072 
2073 err_pd:
2074 	ib_dealloc_pd(ctx->pd);
2075 	ctx->pd = NULL;
2076 
2077 err_cq:
2078 	ib_destroy_cq(ctx->cq);
2079 	ctx->cq = NULL;
2080 
2081 err_buf:
2082 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2083 
2084 err_out_qp0:
2085 	if (ctx->has_smi)
2086 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2087 err_out:
2088 	ctx->state = DEMUX_PV_STATE_DOWN;
2089 	return ret;
2090 }
2091 
2092 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2093 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2094 {
2095 	if (!ctx)
2096 		return;
2097 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2098 		ctx->state = DEMUX_PV_STATE_DOWNING;
2099 		if (flush)
2100 			flush_workqueue(ctx->wq);
2101 		if (ctx->has_smi) {
2102 			ib_destroy_qp(ctx->qp[0].qp);
2103 			ctx->qp[0].qp = NULL;
2104 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2105 		}
2106 		ib_destroy_qp(ctx->qp[1].qp);
2107 		ctx->qp[1].qp = NULL;
2108 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2109 		ib_dealloc_pd(ctx->pd);
2110 		ctx->pd = NULL;
2111 		ib_destroy_cq(ctx->cq);
2112 		ctx->cq = NULL;
2113 		ctx->state = DEMUX_PV_STATE_DOWN;
2114 	}
2115 }
2116 
2117 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2118 				  int port, int do_init)
2119 {
2120 	int ret = 0;
2121 
2122 	if (!do_init) {
2123 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2124 		/* for master, destroy real sqp resources */
2125 		if (slave == mlx4_master_func_num(dev->dev))
2126 			destroy_pv_resources(dev, slave, port,
2127 					     dev->sriov.sqps[port - 1], 1);
2128 		/* destroy the tunnel qp resources */
2129 		destroy_pv_resources(dev, slave, port,
2130 				     dev->sriov.demux[port - 1].tun[slave], 1);
2131 		return 0;
2132 	}
2133 
2134 	/* create the tunnel qp resources */
2135 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2136 				  dev->sriov.demux[port - 1].tun[slave]);
2137 
2138 	/* for master, create the real sqp resources */
2139 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2140 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2141 					  dev->sriov.sqps[port - 1]);
2142 	return ret;
2143 }
2144 
2145 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2146 {
2147 	struct mlx4_ib_demux_work *dmxw;
2148 
2149 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2150 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2151 			       dmxw->do_init);
2152 	kfree(dmxw);
2153 	return;
2154 }
2155 
2156 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2157 				       struct mlx4_ib_demux_ctx *ctx,
2158 				       int port)
2159 {
2160 	char name[12];
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 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2197 	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2198 	if (!ctx->wq) {
2199 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2200 		ret = -ENOMEM;
2201 		goto err_wq;
2202 	}
2203 
2204 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2205 	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2206 	if (!ctx->ud_wq) {
2207 		pr_err("Failed to create up/down WQ for port %d\n", port);
2208 		ret = -ENOMEM;
2209 		goto err_udwq;
2210 	}
2211 
2212 	return 0;
2213 
2214 err_udwq:
2215 	destroy_workqueue(ctx->wq);
2216 	ctx->wq = NULL;
2217 
2218 err_wq:
2219 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2220 err_mcg:
2221 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2222 		free_pv_object(dev, i, port);
2223 	kfree(ctx->tun);
2224 	ctx->tun = NULL;
2225 	return ret;
2226 }
2227 
2228 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2229 {
2230 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2231 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2232 		flush_workqueue(sqp_ctx->wq);
2233 		if (sqp_ctx->has_smi) {
2234 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2235 			sqp_ctx->qp[0].qp = NULL;
2236 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2237 		}
2238 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2239 		sqp_ctx->qp[1].qp = NULL;
2240 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2241 		ib_dealloc_pd(sqp_ctx->pd);
2242 		sqp_ctx->pd = NULL;
2243 		ib_destroy_cq(sqp_ctx->cq);
2244 		sqp_ctx->cq = NULL;
2245 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2246 	}
2247 }
2248 
2249 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2250 {
2251 	int i;
2252 	if (ctx) {
2253 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2254 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2255 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2256 			if (!ctx->tun[i])
2257 				continue;
2258 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2259 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2260 		}
2261 		flush_workqueue(ctx->wq);
2262 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2263 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2264 			free_pv_object(dev, i, ctx->port);
2265 		}
2266 		kfree(ctx->tun);
2267 		destroy_workqueue(ctx->ud_wq);
2268 		destroy_workqueue(ctx->wq);
2269 	}
2270 }
2271 
2272 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2273 {
2274 	int i;
2275 
2276 	if (!mlx4_is_master(dev->dev))
2277 		return;
2278 	/* initialize or tear down tunnel QPs for the master */
2279 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2280 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2281 	return;
2282 }
2283 
2284 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2285 {
2286 	int i = 0;
2287 	int err;
2288 
2289 	if (!mlx4_is_mfunc(dev->dev))
2290 		return 0;
2291 
2292 	dev->sriov.is_going_down = 0;
2293 	spin_lock_init(&dev->sriov.going_down_lock);
2294 	mlx4_ib_cm_paravirt_init(dev);
2295 
2296 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2297 
2298 	if (mlx4_is_slave(dev->dev)) {
2299 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2300 		return 0;
2301 	}
2302 
2303 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2304 		if (i == mlx4_master_func_num(dev->dev))
2305 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2306 		else
2307 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2308 	}
2309 
2310 	err = mlx4_ib_init_alias_guid_service(dev);
2311 	if (err) {
2312 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2313 		goto paravirt_err;
2314 	}
2315 	err = mlx4_ib_device_register_sysfs(dev);
2316 	if (err) {
2317 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2318 		goto sysfs_err;
2319 	}
2320 
2321 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2322 		     dev->dev->caps.sqp_demux);
2323 	for (i = 0; i < dev->num_ports; i++) {
2324 		union ib_gid gid;
2325 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2326 		if (err)
2327 			goto demux_err;
2328 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2329 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2330 			     be64_to_cpu(gid.global.subnet_prefix));
2331 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2332 				      &dev->sriov.sqps[i]);
2333 		if (err)
2334 			goto demux_err;
2335 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2336 		if (err)
2337 			goto free_pv;
2338 	}
2339 	mlx4_ib_master_tunnels(dev, 1);
2340 	return 0;
2341 
2342 free_pv:
2343 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2344 demux_err:
2345 	while (--i >= 0) {
2346 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2347 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2348 	}
2349 	mlx4_ib_device_unregister_sysfs(dev);
2350 
2351 sysfs_err:
2352 	mlx4_ib_destroy_alias_guid_service(dev);
2353 
2354 paravirt_err:
2355 	mlx4_ib_cm_paravirt_clean(dev, -1);
2356 
2357 	return err;
2358 }
2359 
2360 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2361 {
2362 	int i;
2363 	unsigned long flags;
2364 
2365 	if (!mlx4_is_mfunc(dev->dev))
2366 		return;
2367 
2368 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2369 	dev->sriov.is_going_down = 1;
2370 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2371 	if (mlx4_is_master(dev->dev)) {
2372 		for (i = 0; i < dev->num_ports; i++) {
2373 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2374 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2375 			kfree(dev->sriov.sqps[i]);
2376 			dev->sriov.sqps[i] = NULL;
2377 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2378 		}
2379 
2380 		mlx4_ib_cm_paravirt_clean(dev, -1);
2381 		mlx4_ib_destroy_alias_guid_service(dev);
2382 		mlx4_ib_device_unregister_sysfs(dev);
2383 	}
2384 }
2385