xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision 13dd8710)
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 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
504 {
505 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
506 
507 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
508 }
509 
510 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
511 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
512 			  struct ib_grh *grh, struct ib_mad *mad)
513 {
514 	struct ib_sge list;
515 	struct ib_ud_wr wr;
516 	const struct ib_send_wr *bad_wr;
517 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
518 	struct mlx4_ib_demux_pv_qp *tun_qp;
519 	struct mlx4_rcv_tunnel_mad *tun_mad;
520 	struct rdma_ah_attr attr;
521 	struct ib_ah *ah;
522 	struct ib_qp *src_qp = NULL;
523 	unsigned tun_tx_ix = 0;
524 	int dqpn;
525 	int ret = 0;
526 	u16 tun_pkey_ix;
527 	u16 cached_pkey;
528 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
529 
530 	if (dest_qpt > IB_QPT_GSI) {
531 		pr_debug("dest_qpt (%d) > IB_QPT_GSI\n", dest_qpt);
532 		return -EINVAL;
533 	}
534 
535 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
536 
537 	/* check if proxy qp created */
538 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
539 		return -EAGAIN;
540 
541 	if (!dest_qpt)
542 		tun_qp = &tun_ctx->qp[0];
543 	else
544 		tun_qp = &tun_ctx->qp[1];
545 
546 	/* compute P_Key index to put in tunnel header for slave */
547 	if (dest_qpt) {
548 		u16 pkey_ix;
549 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
550 		if (ret) {
551 			pr_debug("unable to get %s cached pkey for index %d, ret %d\n",
552 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
553 				 wc->pkey_index, ret);
554 			return -EINVAL;
555 		}
556 
557 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
558 		if (ret) {
559 			pr_debug("unable to get %s pkey ix for pkey 0x%x, ret %d\n",
560 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
561 				 cached_pkey, ret);
562 			return -EINVAL;
563 		}
564 		tun_pkey_ix = pkey_ix;
565 	} else
566 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
567 
568 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
569 
570 	/* get tunnel tx data buf for slave */
571 	src_qp = tun_qp->qp;
572 
573 	/* create ah. Just need an empty one with the port num for the post send.
574 	 * The driver will set the force loopback bit in post_send */
575 	memset(&attr, 0, sizeof attr);
576 	attr.type = rdma_ah_find_type(&dev->ib_dev, port);
577 
578 	rdma_ah_set_port_num(&attr, port);
579 	if (is_eth) {
580 		union ib_gid sgid;
581 		union ib_gid dgid;
582 
583 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
584 			return -EINVAL;
585 		rdma_ah_set_grh(&attr, &dgid, 0, 0, 0, 0);
586 	}
587 	ah = rdma_create_ah(tun_ctx->pd, &attr, 0);
588 	if (IS_ERR(ah))
589 		return -ENOMEM;
590 
591 	/* allocate tunnel tx buf after pass failure returns */
592 	spin_lock(&tun_qp->tx_lock);
593 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
594 	    (MLX4_NUM_TUNNEL_BUFS - 1))
595 		ret = -EAGAIN;
596 	else
597 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
598 	spin_unlock(&tun_qp->tx_lock);
599 	if (ret)
600 		goto end;
601 
602 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
603 	if (tun_qp->tx_ring[tun_tx_ix].ah)
604 		rdma_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah, 0);
605 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
606 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
607 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
608 				   sizeof (struct mlx4_rcv_tunnel_mad),
609 				   DMA_TO_DEVICE);
610 
611 	/* copy over to tunnel buffer */
612 	if (grh)
613 		memcpy(&tun_mad->grh, grh, sizeof *grh);
614 	memcpy(&tun_mad->mad, mad, sizeof *mad);
615 
616 	/* adjust tunnel data */
617 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
618 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
619 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
620 
621 	if (is_eth) {
622 		u16 vlan = 0;
623 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
624 						NULL)) {
625 			/* VST mode */
626 			if (vlan != wc->vlan_id)
627 				/* Packet vlan is not the VST-assigned vlan.
628 				 * Drop the packet.
629 				 */
630 				goto out;
631 			 else
632 				/* Remove the vlan tag before forwarding
633 				 * the packet to the VF.
634 				 */
635 				vlan = 0xffff;
636 		} else {
637 			vlan = wc->vlan_id;
638 		}
639 
640 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
641 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
642 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
643 	} else {
644 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
645 		tun_mad->hdr.slid_mac_47_32 = ib_lid_be16(wc->slid);
646 	}
647 
648 	ib_dma_sync_single_for_device(&dev->ib_dev,
649 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
650 				      sizeof (struct mlx4_rcv_tunnel_mad),
651 				      DMA_TO_DEVICE);
652 
653 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
654 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
655 	list.lkey = tun_ctx->pd->local_dma_lkey;
656 
657 	wr.ah = ah;
658 	wr.port_num = port;
659 	wr.remote_qkey = IB_QP_SET_QKEY;
660 	wr.remote_qpn = dqpn;
661 	wr.wr.next = NULL;
662 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
663 	wr.wr.sg_list = &list;
664 	wr.wr.num_sge = 1;
665 	wr.wr.opcode = IB_WR_SEND;
666 	wr.wr.send_flags = IB_SEND_SIGNALED;
667 
668 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
669 	if (!ret)
670 		return 0;
671  out:
672 	spin_lock(&tun_qp->tx_lock);
673 	tun_qp->tx_ix_tail++;
674 	spin_unlock(&tun_qp->tx_lock);
675 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
676 end:
677 	rdma_destroy_ah(ah, 0);
678 	return ret;
679 }
680 
681 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
682 			struct ib_wc *wc, struct ib_grh *grh,
683 			struct ib_mad *mad)
684 {
685 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
686 	int err, other_port;
687 	int slave = -1;
688 	u8 *slave_id;
689 	int is_eth = 0;
690 
691 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
692 		is_eth = 0;
693 	else
694 		is_eth = 1;
695 
696 	if (is_eth) {
697 		union ib_gid dgid;
698 		union ib_gid sgid;
699 
700 		if (get_gids_from_l3_hdr(grh, &sgid, &dgid))
701 			return -EINVAL;
702 		if (!(wc->wc_flags & IB_WC_GRH)) {
703 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
704 			return -EINVAL;
705 		}
706 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
707 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
708 			return -EINVAL;
709 		}
710 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, dgid.raw, &slave);
711 		if (err && mlx4_is_mf_bonded(dev->dev)) {
712 			other_port = (port == 1) ? 2 : 1;
713 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, dgid.raw, &slave);
714 			if (!err) {
715 				port = other_port;
716 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
717 					 slave, grh->dgid.raw, port, other_port);
718 			}
719 		}
720 		if (err) {
721 			mlx4_ib_warn(ibdev, "failed matching grh\n");
722 			return -ENOENT;
723 		}
724 		if (slave >= dev->dev->caps.sqp_demux) {
725 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
726 				     slave, dev->dev->caps.sqp_demux);
727 			return -ENOENT;
728 		}
729 
730 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
731 			return 0;
732 
733 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
734 		if (err)
735 			pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
736 				 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
737 				 slave, err);
738 		return 0;
739 	}
740 
741 	/* Initially assume that this mad is for us */
742 	slave = mlx4_master_func_num(dev->dev);
743 
744 	/* See if the slave id is encoded in a response mad */
745 	if (mad->mad_hdr.method & 0x80) {
746 		slave_id = (u8 *) &mad->mad_hdr.tid;
747 		slave = *slave_id;
748 		if (slave != 255) /*255 indicates the dom0*/
749 			*slave_id = 0; /* remap tid */
750 	}
751 
752 	/* If a grh is present, we demux according to it */
753 	if (wc->wc_flags & IB_WC_GRH) {
754 		if (grh->dgid.global.interface_id ==
755 			cpu_to_be64(IB_SA_WELL_KNOWN_GUID) &&
756 		    grh->dgid.global.subnet_prefix == cpu_to_be64(
757 			atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix))) {
758 			slave = 0;
759 		} else {
760 			slave = mlx4_ib_find_real_gid(ibdev, port,
761 						      grh->dgid.global.interface_id);
762 			if (slave < 0) {
763 				mlx4_ib_warn(ibdev, "failed matching grh\n");
764 				return -ENOENT;
765 			}
766 		}
767 	}
768 	/* Class-specific handling */
769 	switch (mad->mad_hdr.mgmt_class) {
770 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
771 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
772 		/* 255 indicates the dom0 */
773 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
774 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
775 				return -EPERM;
776 			/* for a VF. drop unsolicited MADs */
777 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
778 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
779 					     slave, mad->mad_hdr.mgmt_class,
780 					     mad->mad_hdr.method);
781 				return -EINVAL;
782 			}
783 		}
784 		break;
785 	case IB_MGMT_CLASS_SUBN_ADM:
786 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
787 					     (struct ib_sa_mad *) mad))
788 			return 0;
789 		break;
790 	case IB_MGMT_CLASS_CM:
791 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
792 			return 0;
793 		break;
794 	case IB_MGMT_CLASS_DEVICE_MGMT:
795 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
796 			return 0;
797 		break;
798 	default:
799 		/* Drop unsupported classes for slaves in tunnel mode */
800 		if (slave != mlx4_master_func_num(dev->dev)) {
801 			pr_debug("dropping unsupported ingress mad from class:%d "
802 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
803 			return 0;
804 		}
805 	}
806 	/*make sure that no slave==255 was not handled yet.*/
807 	if (slave >= dev->dev->caps.sqp_demux) {
808 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
809 			     slave, dev->dev->caps.sqp_demux);
810 		return -ENOENT;
811 	}
812 
813 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
814 	if (err)
815 		pr_debug("failed sending %s to slave %d via tunnel qp (%d)\n",
816 			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
817 			 slave, err);
818 	return 0;
819 }
820 
821 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
822 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
823 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
824 {
825 	u16 slid, prev_lid = 0;
826 	int err;
827 	struct ib_port_attr pattr;
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 		switch (counter_stats.counter_mode & 0xf) {
968 		case 0:
969 			edit_counter(&counter_stats,
970 				     (void *)(out_mad->data + 40),
971 				     in_mad->mad_hdr.attr_id);
972 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
973 			break;
974 		default:
975 			err = IB_MAD_RESULT_FAILURE;
976 		}
977 	}
978 
979 	return err;
980 }
981 
982 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
983 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
984 			const struct ib_mad *in, struct ib_mad *out,
985 			size_t *out_mad_size, u16 *out_mad_pkey_index)
986 {
987 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
988 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
989 
990 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
991 	 * queries, should be called only by VFs and for that specific purpose
992 	 */
993 	if (link == IB_LINK_LAYER_INFINIBAND) {
994 		if (mlx4_is_slave(dev->dev) &&
995 		    (in->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
996 		     (in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
997 		      in->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
998 		      in->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
999 			return iboe_process_mad(ibdev, mad_flags, port_num,
1000 						in_wc, in_grh, in, out);
1001 
1002 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc, in_grh,
1003 				      in, out);
1004 	}
1005 
1006 	if (link == IB_LINK_LAYER_ETHERNET)
1007 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
1008 					in_grh, in, out);
1009 
1010 	return -EINVAL;
1011 }
1012 
1013 static void send_handler(struct ib_mad_agent *agent,
1014 			 struct ib_mad_send_wc *mad_send_wc)
1015 {
1016 	if (mad_send_wc->send_buf->context[0])
1017 		rdma_destroy_ah(mad_send_wc->send_buf->context[0], 0);
1018 	ib_free_send_mad(mad_send_wc->send_buf);
1019 }
1020 
1021 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
1022 {
1023 	struct ib_mad_agent *agent;
1024 	int p, q;
1025 	int ret;
1026 	enum rdma_link_layer ll;
1027 
1028 	for (p = 0; p < dev->num_ports; ++p) {
1029 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1030 		for (q = 0; q <= 1; ++q) {
1031 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1032 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1033 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1034 							      NULL, 0, send_handler,
1035 							      NULL, NULL, 0);
1036 				if (IS_ERR(agent)) {
1037 					ret = PTR_ERR(agent);
1038 					goto err;
1039 				}
1040 				dev->send_agent[p][q] = agent;
1041 			} else
1042 				dev->send_agent[p][q] = NULL;
1043 		}
1044 	}
1045 
1046 	return 0;
1047 
1048 err:
1049 	for (p = 0; p < dev->num_ports; ++p)
1050 		for (q = 0; q <= 1; ++q)
1051 			if (dev->send_agent[p][q])
1052 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1053 
1054 	return ret;
1055 }
1056 
1057 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1058 {
1059 	struct ib_mad_agent *agent;
1060 	int p, q;
1061 
1062 	for (p = 0; p < dev->num_ports; ++p) {
1063 		for (q = 0; q <= 1; ++q) {
1064 			agent = dev->send_agent[p][q];
1065 			if (agent) {
1066 				dev->send_agent[p][q] = NULL;
1067 				ib_unregister_mad_agent(agent);
1068 			}
1069 		}
1070 
1071 		if (dev->sm_ah[p])
1072 			rdma_destroy_ah(dev->sm_ah[p], 0);
1073 	}
1074 }
1075 
1076 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1077 {
1078 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1079 
1080 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1081 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1082 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1083 }
1084 
1085 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1086 {
1087 	/* re-configure the alias-guid and mcg's */
1088 	if (mlx4_is_master(dev->dev)) {
1089 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1090 
1091 		if (!dev->sriov.is_going_down) {
1092 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1093 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1094 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1095 		}
1096 	}
1097 
1098 	/* Update the sl to vl table from inside client rereg
1099 	 * only if in secure-host mode (snooping is not possible)
1100 	 * and the sl-to-vl change event is not generated by FW.
1101 	 */
1102 	if (!mlx4_is_slave(dev->dev) &&
1103 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1104 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1105 		if (mlx4_is_master(dev->dev))
1106 			/* already in work queue from mlx4_ib_event queueing
1107 			 * mlx4_handle_port_mgmt_change_event, which calls
1108 			 * this procedure. Therefore, call sl2vl_update directly.
1109 			 */
1110 			mlx4_ib_sl2vl_update(dev, port_num);
1111 		else
1112 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1113 	}
1114 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1115 }
1116 
1117 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1118 			      struct mlx4_eqe *eqe)
1119 {
1120 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1121 			    GET_MASK_FROM_EQE(eqe));
1122 }
1123 
1124 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1125 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1126 {
1127 	struct ib_smp *in_mad  = NULL;
1128 	struct ib_smp *out_mad  = NULL;
1129 	u16 i;
1130 
1131 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1132 		return;
1133 
1134 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1135 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1136 	if (!in_mad || !out_mad)
1137 		goto out;
1138 
1139 	guid_tbl_blk_num  *= 4;
1140 
1141 	for (i = 0; i < 4; i++) {
1142 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1143 			continue;
1144 		memset(in_mad, 0, sizeof *in_mad);
1145 		memset(out_mad, 0, sizeof *out_mad);
1146 
1147 		in_mad->base_version  = 1;
1148 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1149 		in_mad->class_version = 1;
1150 		in_mad->method        = IB_MGMT_METHOD_GET;
1151 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1152 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1153 
1154 		if (mlx4_MAD_IFC(dev,
1155 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1156 				 port_num, NULL, NULL, in_mad, out_mad)) {
1157 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1158 			goto out;
1159 		}
1160 
1161 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1162 						    port_num,
1163 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1164 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1165 						     port_num,
1166 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1167 	}
1168 
1169 out:
1170 	kfree(in_mad);
1171 	kfree(out_mad);
1172 	return;
1173 }
1174 
1175 void handle_port_mgmt_change_event(struct work_struct *work)
1176 {
1177 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1178 	struct mlx4_ib_dev *dev = ew->ib_dev;
1179 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1180 	u8 port = eqe->event.port_mgmt_change.port;
1181 	u32 changed_attr;
1182 	u32 tbl_block;
1183 	u32 change_bitmap;
1184 
1185 	switch (eqe->subtype) {
1186 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1187 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1188 
1189 		/* Update the SM ah - This should be done before handling
1190 		   the other changed attributes so that MADs can be sent to the SM */
1191 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1192 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1193 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1194 			update_sm_ah(dev, port, lid, sl);
1195 		}
1196 
1197 		/* Check if it is a lid change event */
1198 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1199 			handle_lid_change_event(dev, port);
1200 
1201 		/* Generate GUID changed event */
1202 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1203 			if (mlx4_is_master(dev->dev)) {
1204 				union ib_gid gid;
1205 				int err = 0;
1206 
1207 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1208 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1209 				else
1210 					gid.global.subnet_prefix =
1211 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1212 				if (err) {
1213 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1214 						port, err);
1215 				} else {
1216 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1217 						 port,
1218 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1219 						 be64_to_cpu(gid.global.subnet_prefix));
1220 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1221 						     be64_to_cpu(gid.global.subnet_prefix));
1222 				}
1223 			}
1224 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1225 			/*if master, notify all slaves*/
1226 			if (mlx4_is_master(dev->dev))
1227 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1228 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1229 		}
1230 
1231 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1232 			handle_client_rereg_event(dev, port);
1233 		break;
1234 
1235 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1236 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1237 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1238 			propagate_pkey_ev(dev, port, eqe);
1239 		break;
1240 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1241 		/* paravirtualized master's guid is guid 0 -- does not change */
1242 		if (!mlx4_is_master(dev->dev))
1243 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1244 		/*if master, notify relevant slaves*/
1245 		else if (!dev->sriov.is_going_down) {
1246 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1247 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1248 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1249 		}
1250 		break;
1251 
1252 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1253 		/* cache sl to vl mapping changes for use in
1254 		 * filling QP1 LRH VL field when sending packets
1255 		 */
1256 		if (!mlx4_is_slave(dev->dev)) {
1257 			union sl2vl_tbl_to_u64 sl2vl64;
1258 			int jj;
1259 
1260 			for (jj = 0; jj < 8; jj++) {
1261 				sl2vl64.sl8[jj] =
1262 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1263 				pr_debug("port %u, sl2vl[%d] = %02x\n",
1264 					 port, jj, sl2vl64.sl8[jj]);
1265 			}
1266 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1267 		}
1268 		break;
1269 	default:
1270 		pr_warn("Unsupported subtype 0x%x for "
1271 			"Port Management Change event\n", eqe->subtype);
1272 	}
1273 
1274 	kfree(ew);
1275 }
1276 
1277 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1278 			    enum ib_event_type type)
1279 {
1280 	struct ib_event event;
1281 
1282 	event.device		= &dev->ib_dev;
1283 	event.element.port_num	= port_num;
1284 	event.event		= type;
1285 
1286 	ib_dispatch_event(&event);
1287 }
1288 
1289 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1290 {
1291 	unsigned long flags;
1292 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1293 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1294 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1295 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1296 		queue_work(ctx->wq, &ctx->work);
1297 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1298 }
1299 
1300 static void mlx4_ib_wire_comp_handler(struct ib_cq *cq, void *arg)
1301 {
1302 	unsigned long flags;
1303 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1304 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1305 
1306 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1307 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1308 		queue_work(ctx->wi_wq, &ctx->work);
1309 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1310 }
1311 
1312 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1313 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1314 				  int index)
1315 {
1316 	struct ib_sge sg_list;
1317 	struct ib_recv_wr recv_wr;
1318 	const struct ib_recv_wr *bad_recv_wr;
1319 	int size;
1320 
1321 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1322 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1323 
1324 	sg_list.addr = tun_qp->ring[index].map;
1325 	sg_list.length = size;
1326 	sg_list.lkey = ctx->pd->local_dma_lkey;
1327 
1328 	recv_wr.next = NULL;
1329 	recv_wr.sg_list = &sg_list;
1330 	recv_wr.num_sge = 1;
1331 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1332 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1333 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1334 				      size, DMA_FROM_DEVICE);
1335 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1336 }
1337 
1338 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1339 		int slave, struct ib_sa_mad *sa_mad)
1340 {
1341 	int ret = 0;
1342 
1343 	/* dispatch to different sa handlers */
1344 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1345 	case IB_SA_ATTR_MC_MEMBER_REC:
1346 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1347 		break;
1348 	default:
1349 		break;
1350 	}
1351 	return ret;
1352 }
1353 
1354 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1355 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1356 			 u32 remote_qpn, u32 qkey, struct rdma_ah_attr *attr,
1357 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1358 {
1359 	struct ib_sge list;
1360 	struct ib_ud_wr wr;
1361 	const struct ib_send_wr *bad_wr;
1362 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1363 	struct mlx4_ib_demux_pv_qp *sqp;
1364 	struct mlx4_mad_snd_buf *sqp_mad;
1365 	struct ib_ah *ah;
1366 	struct ib_qp *send_qp = NULL;
1367 	unsigned wire_tx_ix = 0;
1368 	u16 wire_pkey_ix;
1369 	int src_qpnum;
1370 	int ret;
1371 
1372 	sqp_ctx = dev->sriov.sqps[port-1];
1373 
1374 	/* check if proxy qp created */
1375 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1376 		return -EAGAIN;
1377 
1378 	if (dest_qpt == IB_QPT_SMI) {
1379 		src_qpnum = 0;
1380 		sqp = &sqp_ctx->qp[0];
1381 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1382 	} else {
1383 		src_qpnum = 1;
1384 		sqp = &sqp_ctx->qp[1];
1385 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1386 	}
1387 
1388 	send_qp = sqp->qp;
1389 
1390 	ah = rdma_zalloc_drv_obj(sqp_ctx->pd->device, ib_ah);
1391 	if (!ah)
1392 		return -ENOMEM;
1393 
1394 	ah->device = sqp_ctx->pd->device;
1395 	ah->pd = sqp_ctx->pd;
1396 
1397 	/* create ah */
1398 	ret = mlx4_ib_create_ah_slave(ah, attr,
1399 				      rdma_ah_retrieve_grh(attr)->sgid_index,
1400 				      s_mac, vlan_id);
1401 	if (ret)
1402 		goto out;
1403 
1404 	spin_lock(&sqp->tx_lock);
1405 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1406 	    (MLX4_NUM_WIRE_BUFS - 1))
1407 		ret = -EAGAIN;
1408 	else
1409 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_WIRE_BUFS - 1);
1410 	spin_unlock(&sqp->tx_lock);
1411 	if (ret)
1412 		goto out;
1413 
1414 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1415 	kfree(sqp->tx_ring[wire_tx_ix].ah);
1416 	sqp->tx_ring[wire_tx_ix].ah = ah;
1417 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1418 				   sqp->tx_ring[wire_tx_ix].buf.map,
1419 				   sizeof (struct mlx4_mad_snd_buf),
1420 				   DMA_TO_DEVICE);
1421 
1422 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1423 
1424 	ib_dma_sync_single_for_device(&dev->ib_dev,
1425 				      sqp->tx_ring[wire_tx_ix].buf.map,
1426 				      sizeof (struct mlx4_mad_snd_buf),
1427 				      DMA_TO_DEVICE);
1428 
1429 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1430 	list.length = sizeof (struct mlx4_mad_snd_buf);
1431 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1432 
1433 	wr.ah = ah;
1434 	wr.port_num = port;
1435 	wr.pkey_index = wire_pkey_ix;
1436 	wr.remote_qkey = qkey;
1437 	wr.remote_qpn = remote_qpn;
1438 	wr.wr.next = NULL;
1439 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1440 	wr.wr.sg_list = &list;
1441 	wr.wr.num_sge = 1;
1442 	wr.wr.opcode = IB_WR_SEND;
1443 	wr.wr.send_flags = IB_SEND_SIGNALED;
1444 
1445 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1446 	if (!ret)
1447 		return 0;
1448 
1449 	spin_lock(&sqp->tx_lock);
1450 	sqp->tx_ix_tail++;
1451 	spin_unlock(&sqp->tx_lock);
1452 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1453 out:
1454 	kfree(ah);
1455 	return ret;
1456 }
1457 
1458 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1459 {
1460 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1461 		return slave;
1462 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1463 }
1464 
1465 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1466 				    struct rdma_ah_attr *ah_attr)
1467 {
1468 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
1469 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1470 		grh->sgid_index = slave;
1471 	else
1472 		grh->sgid_index += get_slave_base_gid_ix(dev, slave, port);
1473 }
1474 
1475 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1476 {
1477 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1478 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1479 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1480 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1481 	struct mlx4_ib_ah ah;
1482 	struct rdma_ah_attr ah_attr;
1483 	u8 *slave_id;
1484 	int slave;
1485 	int port;
1486 	u16 vlan_id;
1487 	u8 qos;
1488 	u8 *dmac;
1489 	int sts;
1490 
1491 	/* Get slave that sent this packet */
1492 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1493 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1494 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1495 	    wc->src_qp & 0x4) {
1496 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1497 		return;
1498 	}
1499 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1500 	if (slave != ctx->slave) {
1501 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1502 			     "belongs to another slave\n", wc->src_qp);
1503 		return;
1504 	}
1505 
1506 	/* Map transaction ID */
1507 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1508 				   sizeof (struct mlx4_tunnel_mad),
1509 				   DMA_FROM_DEVICE);
1510 	switch (tunnel->mad.mad_hdr.method) {
1511 	case IB_MGMT_METHOD_SET:
1512 	case IB_MGMT_METHOD_GET:
1513 	case IB_MGMT_METHOD_REPORT:
1514 	case IB_SA_METHOD_GET_TABLE:
1515 	case IB_SA_METHOD_DELETE:
1516 	case IB_SA_METHOD_GET_MULTI:
1517 	case IB_SA_METHOD_GET_TRACE_TBL:
1518 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1519 		if (*slave_id) {
1520 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1521 				     "class:%d slave:%d\n", *slave_id,
1522 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1523 			return;
1524 		} else
1525 			*slave_id = slave;
1526 	default:
1527 		/* nothing */;
1528 	}
1529 
1530 	/* Class-specific handling */
1531 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1532 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1533 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1534 		if (slave != mlx4_master_func_num(dev->dev) &&
1535 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1536 			return;
1537 		break;
1538 	case IB_MGMT_CLASS_SUBN_ADM:
1539 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1540 			      (struct ib_sa_mad *) &tunnel->mad))
1541 			return;
1542 		break;
1543 	case IB_MGMT_CLASS_CM:
1544 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1545 			      (struct ib_mad *) &tunnel->mad))
1546 			return;
1547 		break;
1548 	case IB_MGMT_CLASS_DEVICE_MGMT:
1549 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1550 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1551 			return;
1552 		break;
1553 	default:
1554 		/* Drop unsupported classes for slaves in tunnel mode */
1555 		if (slave != mlx4_master_func_num(dev->dev)) {
1556 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1557 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1558 			return;
1559 		}
1560 	}
1561 
1562 	/* We are using standard ib_core services to send the mad, so generate a
1563 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1564 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1565 	ah.ibah.device = ctx->ib_dev;
1566 
1567 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1568 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1569 	if (port < 0)
1570 		return;
1571 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1572 	ah.ibah.type = rdma_ah_find_type(&dev->ib_dev, port);
1573 
1574 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1575 	if (rdma_ah_get_ah_flags(&ah_attr) & IB_AH_GRH)
1576 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1577 	dmac = rdma_ah_retrieve_dmac(&ah_attr);
1578 	if (dmac)
1579 		memcpy(dmac, tunnel->hdr.mac, ETH_ALEN);
1580 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1581 	/* if slave have default vlan use it */
1582 	if (mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1583 					&vlan_id, &qos))
1584 		rdma_ah_set_sl(&ah_attr, qos);
1585 
1586 	sts = mlx4_ib_send_to_wire(dev, slave, ctx->port,
1587 				   is_proxy_qp0(dev, wc->src_qp, slave) ?
1588 				   IB_QPT_SMI : IB_QPT_GSI,
1589 				   be16_to_cpu(tunnel->hdr.pkey_index),
1590 				   be32_to_cpu(tunnel->hdr.remote_qpn),
1591 				   be32_to_cpu(tunnel->hdr.qkey),
1592 				   &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1593 	if (sts)
1594 		pr_debug("failed sending %s to wire on behalf of slave %d (%d)\n",
1595 			 is_proxy_qp0(dev, wc->src_qp, slave) ? "SMI" : "GSI",
1596 			 slave, sts);
1597 }
1598 
1599 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1600 				 enum ib_qp_type qp_type, int is_tun)
1601 {
1602 	int i;
1603 	struct mlx4_ib_demux_pv_qp *tun_qp;
1604 	int rx_buf_size, tx_buf_size;
1605 	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
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(nmbr_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(nmbr_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 < nmbr_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 < nmbr_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 = nmbr_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 	const int nmbr_bufs = is_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1702 
1703 	if (qp_type > IB_QPT_GSI)
1704 		return;
1705 
1706 	tun_qp = &ctx->qp[qp_type];
1707 	if (is_tun) {
1708 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1709 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1710 	} else {
1711 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1712 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1713 	}
1714 
1715 
1716 	for (i = 0; i < nmbr_bufs; i++) {
1717 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1718 				    rx_buf_size, DMA_FROM_DEVICE);
1719 		kfree(tun_qp->ring[i].addr);
1720 	}
1721 
1722 	for (i = 0; i < nmbr_bufs; i++) {
1723 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1724 				    tx_buf_size, DMA_TO_DEVICE);
1725 		kfree(tun_qp->tx_ring[i].buf.addr);
1726 		if (tun_qp->tx_ring[i].ah)
1727 			rdma_destroy_ah(tun_qp->tx_ring[i].ah, 0);
1728 	}
1729 	kfree(tun_qp->tx_ring);
1730 	kfree(tun_qp->ring);
1731 }
1732 
1733 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1734 {
1735 	struct mlx4_ib_demux_pv_ctx *ctx;
1736 	struct mlx4_ib_demux_pv_qp *tun_qp;
1737 	struct ib_wc wc;
1738 	int ret;
1739 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1740 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1741 
1742 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1743 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1744 		if (wc.status == IB_WC_SUCCESS) {
1745 			switch (wc.opcode) {
1746 			case IB_WC_RECV:
1747 				mlx4_ib_multiplex_mad(ctx, &wc);
1748 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1749 							     wc.wr_id &
1750 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1751 				if (ret)
1752 					pr_err("Failed reposting tunnel "
1753 					       "buf:%lld\n", wc.wr_id);
1754 				break;
1755 			case IB_WC_SEND:
1756 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1757 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1758 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1759 					= NULL;
1760 				spin_lock(&tun_qp->tx_lock);
1761 				tun_qp->tx_ix_tail++;
1762 				spin_unlock(&tun_qp->tx_lock);
1763 
1764 				break;
1765 			default:
1766 				break;
1767 			}
1768 		} else  {
1769 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1770 				 " status = %d, wrid = 0x%llx\n",
1771 				 ctx->slave, wc.status, wc.wr_id);
1772 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1773 				rdma_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1774 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah, 0);
1775 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1776 					= NULL;
1777 				spin_lock(&tun_qp->tx_lock);
1778 				tun_qp->tx_ix_tail++;
1779 				spin_unlock(&tun_qp->tx_lock);
1780 			}
1781 		}
1782 	}
1783 }
1784 
1785 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1786 {
1787 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1788 
1789 	/* It's worse than that! He's dead, Jim! */
1790 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1791 	       event->event, sqp->port);
1792 }
1793 
1794 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1795 			    enum ib_qp_type qp_type, int create_tun)
1796 {
1797 	int i, ret;
1798 	struct mlx4_ib_demux_pv_qp *tun_qp;
1799 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1800 	struct ib_qp_attr attr;
1801 	int qp_attr_mask_INIT;
1802 	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
1803 
1804 	if (qp_type > IB_QPT_GSI)
1805 		return -EINVAL;
1806 
1807 	tun_qp = &ctx->qp[qp_type];
1808 
1809 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1810 	qp_init_attr.init_attr.send_cq = ctx->cq;
1811 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1812 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1813 	qp_init_attr.init_attr.cap.max_send_wr = nmbr_bufs;
1814 	qp_init_attr.init_attr.cap.max_recv_wr = nmbr_bufs;
1815 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1816 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1817 	if (create_tun) {
1818 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1819 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1820 		qp_init_attr.port = ctx->port;
1821 		qp_init_attr.slave = ctx->slave;
1822 		qp_init_attr.proxy_qp_type = qp_type;
1823 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1824 			   IB_QP_QKEY | IB_QP_PORT;
1825 	} else {
1826 		qp_init_attr.init_attr.qp_type = qp_type;
1827 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1828 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1829 	}
1830 	qp_init_attr.init_attr.port_num = ctx->port;
1831 	qp_init_attr.init_attr.qp_context = ctx;
1832 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1833 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1834 	if (IS_ERR(tun_qp->qp)) {
1835 		ret = PTR_ERR(tun_qp->qp);
1836 		tun_qp->qp = NULL;
1837 		pr_err("Couldn't create %s QP (%d)\n",
1838 		       create_tun ? "tunnel" : "special", ret);
1839 		return ret;
1840 	}
1841 
1842 	memset(&attr, 0, sizeof attr);
1843 	attr.qp_state = IB_QPS_INIT;
1844 	ret = 0;
1845 	if (create_tun)
1846 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1847 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1848 					      &attr.pkey_index);
1849 	if (ret || !create_tun)
1850 		attr.pkey_index =
1851 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1852 	attr.qkey = IB_QP1_QKEY;
1853 	attr.port_num = ctx->port;
1854 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1855 	if (ret) {
1856 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1857 		       create_tun ? "tunnel" : "special", ret);
1858 		goto err_qp;
1859 	}
1860 	attr.qp_state = IB_QPS_RTR;
1861 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1862 	if (ret) {
1863 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1864 		       create_tun ? "tunnel" : "special", ret);
1865 		goto err_qp;
1866 	}
1867 	attr.qp_state = IB_QPS_RTS;
1868 	attr.sq_psn = 0;
1869 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1870 	if (ret) {
1871 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1872 		       create_tun ? "tunnel" : "special", ret);
1873 		goto err_qp;
1874 	}
1875 
1876 	for (i = 0; i < nmbr_bufs; i++) {
1877 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1878 		if (ret) {
1879 			pr_err(" mlx4_ib_post_pv_buf error"
1880 			       " (err = %d, i = %d)\n", ret, i);
1881 			goto err_qp;
1882 		}
1883 	}
1884 	return 0;
1885 
1886 err_qp:
1887 	ib_destroy_qp(tun_qp->qp);
1888 	tun_qp->qp = NULL;
1889 	return ret;
1890 }
1891 
1892 /*
1893  * IB MAD completion callback for real SQPs
1894  */
1895 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1896 {
1897 	struct mlx4_ib_demux_pv_ctx *ctx;
1898 	struct mlx4_ib_demux_pv_qp *sqp;
1899 	struct ib_wc wc;
1900 	struct ib_grh *grh;
1901 	struct ib_mad *mad;
1902 
1903 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1904 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1905 
1906 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1907 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1908 		if (wc.status == IB_WC_SUCCESS) {
1909 			switch (wc.opcode) {
1910 			case IB_WC_SEND:
1911 				kfree(sqp->tx_ring[wc.wr_id &
1912 				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
1913 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
1914 					= NULL;
1915 				spin_lock(&sqp->tx_lock);
1916 				sqp->tx_ix_tail++;
1917 				spin_unlock(&sqp->tx_lock);
1918 				break;
1919 			case IB_WC_RECV:
1920 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1921 						(sqp->ring[wc.wr_id &
1922 						(MLX4_NUM_WIRE_BUFS - 1)].addr))->payload);
1923 				grh = &(((struct mlx4_mad_rcv_buf *)
1924 						(sqp->ring[wc.wr_id &
1925 						(MLX4_NUM_WIRE_BUFS - 1)].addr))->grh);
1926 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1927 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1928 							   (MLX4_NUM_WIRE_BUFS - 1)))
1929 					pr_err("Failed reposting SQP "
1930 					       "buf:%lld\n", wc.wr_id);
1931 				break;
1932 			default:
1933 				break;
1934 			}
1935 		} else  {
1936 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1937 				 " status = %d, wrid = 0x%llx\n",
1938 				 ctx->slave, wc.status, wc.wr_id);
1939 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1940 				kfree(sqp->tx_ring[wc.wr_id &
1941 				      (MLX4_NUM_WIRE_BUFS - 1)].ah);
1942 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_WIRE_BUFS - 1)].ah
1943 					= NULL;
1944 				spin_lock(&sqp->tx_lock);
1945 				sqp->tx_ix_tail++;
1946 				spin_unlock(&sqp->tx_lock);
1947 			}
1948 		}
1949 	}
1950 }
1951 
1952 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1953 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1954 {
1955 	struct mlx4_ib_demux_pv_ctx *ctx;
1956 
1957 	*ret_ctx = NULL;
1958 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1959 	if (!ctx)
1960 		return -ENOMEM;
1961 
1962 	ctx->ib_dev = &dev->ib_dev;
1963 	ctx->port = port;
1964 	ctx->slave = slave;
1965 	*ret_ctx = ctx;
1966 	return 0;
1967 }
1968 
1969 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1970 {
1971 	if (dev->sriov.demux[port - 1].tun[slave]) {
1972 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1973 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1974 	}
1975 }
1976 
1977 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1978 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1979 {
1980 	int ret, cq_size;
1981 	struct ib_cq_init_attr cq_attr = {};
1982 	const int nmbr_bufs = create_tun ? MLX4_NUM_TUNNEL_BUFS : MLX4_NUM_WIRE_BUFS;
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 * nmbr_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,
2013 			       create_tun ? mlx4_ib_tunnel_comp_handler : mlx4_ib_wire_comp_handler,
2014 			       NULL, ctx, &cq_attr);
2015 	if (IS_ERR(ctx->cq)) {
2016 		ret = PTR_ERR(ctx->cq);
2017 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
2018 		goto err_buf;
2019 	}
2020 
2021 	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
2022 	if (IS_ERR(ctx->pd)) {
2023 		ret = PTR_ERR(ctx->pd);
2024 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
2025 		goto err_cq;
2026 	}
2027 
2028 	if (ctx->has_smi) {
2029 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
2030 		if (ret) {
2031 			pr_err("Couldn't create %s QP0 (%d)\n",
2032 			       create_tun ? "tunnel for" : "",  ret);
2033 			goto err_pd;
2034 		}
2035 	}
2036 
2037 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2038 	if (ret) {
2039 		pr_err("Couldn't create %s QP1 (%d)\n",
2040 		       create_tun ? "tunnel for" : "",  ret);
2041 		goto err_qp0;
2042 	}
2043 
2044 	if (create_tun)
2045 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2046 	else
2047 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2048 
2049 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2050 	ctx->wi_wq = to_mdev(ibdev)->sriov.demux[port - 1].wi_wq;
2051 
2052 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2053 	if (ret) {
2054 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2055 		goto err_wq;
2056 	}
2057 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2058 	return 0;
2059 
2060 err_wq:
2061 	ctx->wq = NULL;
2062 	ib_destroy_qp(ctx->qp[1].qp);
2063 	ctx->qp[1].qp = NULL;
2064 
2065 
2066 err_qp0:
2067 	if (ctx->has_smi)
2068 		ib_destroy_qp(ctx->qp[0].qp);
2069 	ctx->qp[0].qp = NULL;
2070 
2071 err_pd:
2072 	ib_dealloc_pd(ctx->pd);
2073 	ctx->pd = NULL;
2074 
2075 err_cq:
2076 	ib_destroy_cq(ctx->cq);
2077 	ctx->cq = NULL;
2078 
2079 err_buf:
2080 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2081 
2082 err_out_qp0:
2083 	if (ctx->has_smi)
2084 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2085 err_out:
2086 	ctx->state = DEMUX_PV_STATE_DOWN;
2087 	return ret;
2088 }
2089 
2090 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2091 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2092 {
2093 	if (!ctx)
2094 		return;
2095 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2096 		ctx->state = DEMUX_PV_STATE_DOWNING;
2097 		if (flush)
2098 			flush_workqueue(ctx->wq);
2099 		if (ctx->has_smi) {
2100 			ib_destroy_qp(ctx->qp[0].qp);
2101 			ctx->qp[0].qp = NULL;
2102 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2103 		}
2104 		ib_destroy_qp(ctx->qp[1].qp);
2105 		ctx->qp[1].qp = NULL;
2106 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2107 		ib_dealloc_pd(ctx->pd);
2108 		ctx->pd = NULL;
2109 		ib_destroy_cq(ctx->cq);
2110 		ctx->cq = NULL;
2111 		ctx->state = DEMUX_PV_STATE_DOWN;
2112 	}
2113 }
2114 
2115 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2116 				  int port, int do_init)
2117 {
2118 	int ret = 0;
2119 
2120 	if (!do_init) {
2121 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2122 		/* for master, destroy real sqp resources */
2123 		if (slave == mlx4_master_func_num(dev->dev))
2124 			destroy_pv_resources(dev, slave, port,
2125 					     dev->sriov.sqps[port - 1], 1);
2126 		/* destroy the tunnel qp resources */
2127 		destroy_pv_resources(dev, slave, port,
2128 				     dev->sriov.demux[port - 1].tun[slave], 1);
2129 		return 0;
2130 	}
2131 
2132 	/* create the tunnel qp resources */
2133 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2134 				  dev->sriov.demux[port - 1].tun[slave]);
2135 
2136 	/* for master, create the real sqp resources */
2137 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2138 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2139 					  dev->sriov.sqps[port - 1]);
2140 	return ret;
2141 }
2142 
2143 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2144 {
2145 	struct mlx4_ib_demux_work *dmxw;
2146 
2147 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2148 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2149 			       dmxw->do_init);
2150 	kfree(dmxw);
2151 	return;
2152 }
2153 
2154 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2155 				       struct mlx4_ib_demux_ctx *ctx,
2156 				       int port)
2157 {
2158 	char name[12];
2159 	int ret = 0;
2160 	int i;
2161 
2162 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2163 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2164 	if (!ctx->tun)
2165 		return -ENOMEM;
2166 
2167 	ctx->dev = dev;
2168 	ctx->port = port;
2169 	ctx->ib_dev = &dev->ib_dev;
2170 
2171 	for (i = 0;
2172 	     i < min(dev->dev->caps.sqp_demux,
2173 	     (u16)(dev->dev->persist->num_vfs + 1));
2174 	     i++) {
2175 		struct mlx4_active_ports actv_ports =
2176 			mlx4_get_active_ports(dev->dev, i);
2177 
2178 		if (!test_bit(port - 1, actv_ports.ports))
2179 			continue;
2180 
2181 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2182 		if (ret) {
2183 			ret = -ENOMEM;
2184 			goto err_mcg;
2185 		}
2186 	}
2187 
2188 	ret = mlx4_ib_mcg_port_init(ctx);
2189 	if (ret) {
2190 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2191 		goto err_mcg;
2192 	}
2193 
2194 	snprintf(name, sizeof(name), "mlx4_ibt%d", port);
2195 	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2196 	if (!ctx->wq) {
2197 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2198 		ret = -ENOMEM;
2199 		goto err_wq;
2200 	}
2201 
2202 	snprintf(name, sizeof(name), "mlx4_ibwi%d", port);
2203 	ctx->wi_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2204 	if (!ctx->wi_wq) {
2205 		pr_err("Failed to create wire WQ for port %d\n", port);
2206 		ret = -ENOMEM;
2207 		goto err_wiwq;
2208 	}
2209 
2210 	snprintf(name, sizeof(name), "mlx4_ibud%d", port);
2211 	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2212 	if (!ctx->ud_wq) {
2213 		pr_err("Failed to create up/down WQ for port %d\n", port);
2214 		ret = -ENOMEM;
2215 		goto err_udwq;
2216 	}
2217 
2218 	return 0;
2219 
2220 err_udwq:
2221 	destroy_workqueue(ctx->wi_wq);
2222 	ctx->wi_wq = NULL;
2223 
2224 err_wiwq:
2225 	destroy_workqueue(ctx->wq);
2226 	ctx->wq = NULL;
2227 
2228 err_wq:
2229 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2230 err_mcg:
2231 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2232 		free_pv_object(dev, i, port);
2233 	kfree(ctx->tun);
2234 	ctx->tun = NULL;
2235 	return ret;
2236 }
2237 
2238 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2239 {
2240 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2241 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2242 		flush_workqueue(sqp_ctx->wq);
2243 		if (sqp_ctx->has_smi) {
2244 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2245 			sqp_ctx->qp[0].qp = NULL;
2246 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2247 		}
2248 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2249 		sqp_ctx->qp[1].qp = NULL;
2250 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2251 		ib_dealloc_pd(sqp_ctx->pd);
2252 		sqp_ctx->pd = NULL;
2253 		ib_destroy_cq(sqp_ctx->cq);
2254 		sqp_ctx->cq = NULL;
2255 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2256 	}
2257 }
2258 
2259 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2260 {
2261 	int i;
2262 	if (ctx) {
2263 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2264 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2265 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2266 			if (!ctx->tun[i])
2267 				continue;
2268 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2269 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2270 		}
2271 		flush_workqueue(ctx->wq);
2272 		flush_workqueue(ctx->wi_wq);
2273 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2274 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2275 			free_pv_object(dev, i, ctx->port);
2276 		}
2277 		kfree(ctx->tun);
2278 		destroy_workqueue(ctx->ud_wq);
2279 		destroy_workqueue(ctx->wi_wq);
2280 		destroy_workqueue(ctx->wq);
2281 	}
2282 }
2283 
2284 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2285 {
2286 	int i;
2287 
2288 	if (!mlx4_is_master(dev->dev))
2289 		return;
2290 	/* initialize or tear down tunnel QPs for the master */
2291 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2292 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2293 	return;
2294 }
2295 
2296 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2297 {
2298 	int i = 0;
2299 	int err;
2300 
2301 	if (!mlx4_is_mfunc(dev->dev))
2302 		return 0;
2303 
2304 	dev->sriov.is_going_down = 0;
2305 	spin_lock_init(&dev->sriov.going_down_lock);
2306 	mlx4_ib_cm_paravirt_init(dev);
2307 
2308 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2309 
2310 	if (mlx4_is_slave(dev->dev)) {
2311 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2312 		return 0;
2313 	}
2314 
2315 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2316 		if (i == mlx4_master_func_num(dev->dev))
2317 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2318 		else
2319 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2320 	}
2321 
2322 	err = mlx4_ib_init_alias_guid_service(dev);
2323 	if (err) {
2324 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2325 		goto paravirt_err;
2326 	}
2327 	err = mlx4_ib_device_register_sysfs(dev);
2328 	if (err) {
2329 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2330 		goto sysfs_err;
2331 	}
2332 
2333 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2334 		     dev->dev->caps.sqp_demux);
2335 	for (i = 0; i < dev->num_ports; i++) {
2336 		union ib_gid gid;
2337 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2338 		if (err)
2339 			goto demux_err;
2340 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2341 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2342 			     be64_to_cpu(gid.global.subnet_prefix));
2343 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2344 				      &dev->sriov.sqps[i]);
2345 		if (err)
2346 			goto demux_err;
2347 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2348 		if (err)
2349 			goto free_pv;
2350 	}
2351 	mlx4_ib_master_tunnels(dev, 1);
2352 	return 0;
2353 
2354 free_pv:
2355 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2356 demux_err:
2357 	while (--i >= 0) {
2358 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2359 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2360 	}
2361 	mlx4_ib_device_unregister_sysfs(dev);
2362 
2363 sysfs_err:
2364 	mlx4_ib_destroy_alias_guid_service(dev);
2365 
2366 paravirt_err:
2367 	mlx4_ib_cm_paravirt_clean(dev, -1);
2368 
2369 	return err;
2370 }
2371 
2372 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2373 {
2374 	int i;
2375 	unsigned long flags;
2376 
2377 	if (!mlx4_is_mfunc(dev->dev))
2378 		return;
2379 
2380 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2381 	dev->sriov.is_going_down = 1;
2382 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2383 	if (mlx4_is_master(dev->dev)) {
2384 		for (i = 0; i < dev->num_ports; i++) {
2385 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2386 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2387 			kfree(dev->sriov.sqps[i]);
2388 			dev->sriov.sqps[i] = NULL;
2389 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2390 		}
2391 
2392 		mlx4_ib_cm_paravirt_clean(dev, -1);
2393 		mlx4_ib_destroy_alias_guid_service(dev);
2394 		mlx4_ib_device_unregister_sysfs(dev);
2395 	}
2396 }
2397