xref: /linux/drivers/infiniband/hw/mlx4/mad.c (revision d8ea757b25ec82687c497fc90aa83f9bcea24b5b)
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 
43 #include <linux/mlx4/driver.h>
44 #include "mlx4_ib.h"
45 
46 enum {
47 	MLX4_IB_VENDOR_CLASS1 = 0x9,
48 	MLX4_IB_VENDOR_CLASS2 = 0xa
49 };
50 
51 #define MLX4_TUN_SEND_WRID_SHIFT 34
52 #define MLX4_TUN_QPN_SHIFT 32
53 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
54 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
55 
56 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
57 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
58 
59  /* Port mgmt change event handling */
60 
61 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
62 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
63 #define NUM_IDX_IN_PKEY_TBL_BLK 32
64 #define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
65 #define GUID_TBL_BLK_NUM_ENTRIES 8
66 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
67 
68 struct mlx4_mad_rcv_buf {
69 	struct ib_grh grh;
70 	u8 payload[256];
71 } __packed;
72 
73 struct mlx4_mad_snd_buf {
74 	u8 payload[256];
75 } __packed;
76 
77 struct mlx4_tunnel_mad {
78 	struct ib_grh grh;
79 	struct mlx4_ib_tunnel_header hdr;
80 	struct ib_mad mad;
81 } __packed;
82 
83 struct mlx4_rcv_tunnel_mad {
84 	struct mlx4_rcv_tunnel_hdr hdr;
85 	struct ib_grh grh;
86 	struct ib_mad mad;
87 } __packed;
88 
89 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
90 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
91 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
92 				int block, u32 change_bitmap);
93 
94 __be64 mlx4_ib_gen_node_guid(void)
95 {
96 #define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
97 	return cpu_to_be64(NODE_GUID_HI | prandom_u32());
98 }
99 
100 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
101 {
102 	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
103 		cpu_to_be64(0xff00000000000000LL);
104 }
105 
106 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
107 		 int port, const struct ib_wc *in_wc,
108 		 const struct ib_grh *in_grh,
109 		 const void *in_mad, void *response_mad)
110 {
111 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
112 	void *inbox;
113 	int err;
114 	u32 in_modifier = port;
115 	u8 op_modifier = 0;
116 
117 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
118 	if (IS_ERR(inmailbox))
119 		return PTR_ERR(inmailbox);
120 	inbox = inmailbox->buf;
121 
122 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
123 	if (IS_ERR(outmailbox)) {
124 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
125 		return PTR_ERR(outmailbox);
126 	}
127 
128 	memcpy(inbox, in_mad, 256);
129 
130 	/*
131 	 * Key check traps can't be generated unless we have in_wc to
132 	 * tell us where to send the trap.
133 	 */
134 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
135 		op_modifier |= 0x1;
136 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
137 		op_modifier |= 0x2;
138 	if (mlx4_is_mfunc(dev->dev) &&
139 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
140 		op_modifier |= 0x8;
141 
142 	if (in_wc) {
143 		struct {
144 			__be32		my_qpn;
145 			u32		reserved1;
146 			__be32		rqpn;
147 			u8		sl;
148 			u8		g_path;
149 			u16		reserved2[2];
150 			__be16		pkey;
151 			u32		reserved3[11];
152 			u8		grh[40];
153 		} *ext_info;
154 
155 		memset(inbox + 256, 0, 256);
156 		ext_info = inbox + 256;
157 
158 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
159 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
160 		ext_info->sl     = in_wc->sl << 4;
161 		ext_info->g_path = in_wc->dlid_path_bits |
162 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
163 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
164 
165 		if (in_grh)
166 			memcpy(ext_info->grh, in_grh, 40);
167 
168 		op_modifier |= 0x4;
169 
170 		in_modifier |= in_wc->slid << 16;
171 	}
172 
173 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
174 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
175 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
176 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
177 
178 	if (!err)
179 		memcpy(response_mad, outmailbox->buf, 256);
180 
181 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
182 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
183 
184 	return err;
185 }
186 
187 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
188 {
189 	struct ib_ah *new_ah;
190 	struct ib_ah_attr ah_attr;
191 	unsigned long flags;
192 
193 	if (!dev->send_agent[port_num - 1][0])
194 		return;
195 
196 	memset(&ah_attr, 0, sizeof ah_attr);
197 	ah_attr.dlid     = lid;
198 	ah_attr.sl       = sl;
199 	ah_attr.port_num = port_num;
200 
201 	new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
202 			      &ah_attr);
203 	if (IS_ERR(new_ah))
204 		return;
205 
206 	spin_lock_irqsave(&dev->sm_lock, flags);
207 	if (dev->sm_ah[port_num - 1])
208 		ib_destroy_ah(dev->sm_ah[port_num - 1]);
209 	dev->sm_ah[port_num - 1] = new_ah;
210 	spin_unlock_irqrestore(&dev->sm_lock, flags);
211 }
212 
213 /*
214  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
215  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
216  */
217 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
218 		      u16 prev_lid)
219 {
220 	struct ib_port_info *pinfo;
221 	u16 lid;
222 	__be16 *base;
223 	u32 bn, pkey_change_bitmap;
224 	int i;
225 
226 
227 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
228 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
229 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
230 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
231 		switch (mad->mad_hdr.attr_id) {
232 		case IB_SMP_ATTR_PORT_INFO:
233 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
234 			lid = be16_to_cpu(pinfo->lid);
235 
236 			update_sm_ah(dev, port_num,
237 				     be16_to_cpu(pinfo->sm_lid),
238 				     pinfo->neighbormtu_mastersmsl & 0xf);
239 
240 			if (pinfo->clientrereg_resv_subnetto & 0x80)
241 				handle_client_rereg_event(dev, port_num);
242 
243 			if (prev_lid != lid)
244 				handle_lid_change_event(dev, port_num);
245 			break;
246 
247 		case IB_SMP_ATTR_PKEY_TABLE:
248 			if (!mlx4_is_mfunc(dev->dev)) {
249 				mlx4_ib_dispatch_event(dev, port_num,
250 						       IB_EVENT_PKEY_CHANGE);
251 				break;
252 			}
253 
254 			/* at this point, we are running in the master.
255 			 * Slaves do not receive SMPs.
256 			 */
257 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
258 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
259 			pkey_change_bitmap = 0;
260 			for (i = 0; i < 32; i++) {
261 				pr_debug("PKEY[%d] = x%x\n",
262 					 i + bn*32, be16_to_cpu(base[i]));
263 				if (be16_to_cpu(base[i]) !=
264 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
265 					pkey_change_bitmap |= (1 << i);
266 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
267 						be16_to_cpu(base[i]);
268 				}
269 			}
270 			pr_debug("PKEY Change event: port=%d, "
271 				 "block=0x%x, change_bitmap=0x%x\n",
272 				 port_num, bn, pkey_change_bitmap);
273 
274 			if (pkey_change_bitmap) {
275 				mlx4_ib_dispatch_event(dev, port_num,
276 						       IB_EVENT_PKEY_CHANGE);
277 				if (!dev->sriov.is_going_down)
278 					__propagate_pkey_ev(dev, port_num, bn,
279 							    pkey_change_bitmap);
280 			}
281 			break;
282 
283 		case IB_SMP_ATTR_GUID_INFO:
284 			/* paravirtualized master's guid is guid 0 -- does not change */
285 			if (!mlx4_is_master(dev->dev))
286 				mlx4_ib_dispatch_event(dev, port_num,
287 						       IB_EVENT_GID_CHANGE);
288 			/*if master, notify relevant slaves*/
289 			if (mlx4_is_master(dev->dev) &&
290 			    !dev->sriov.is_going_down) {
291 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
292 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
293 								    (u8 *)(&((struct ib_smp *)mad)->data));
294 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
295 								     (u8 *)(&((struct ib_smp *)mad)->data));
296 			}
297 			break;
298 
299 		default:
300 			break;
301 		}
302 }
303 
304 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
305 				int block, u32 change_bitmap)
306 {
307 	int i, ix, slave, err;
308 	int have_event = 0;
309 
310 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
311 		if (slave == mlx4_master_func_num(dev->dev))
312 			continue;
313 		if (!mlx4_is_slave_active(dev->dev, slave))
314 			continue;
315 
316 		have_event = 0;
317 		for (i = 0; i < 32; i++) {
318 			if (!(change_bitmap & (1 << i)))
319 				continue;
320 			for (ix = 0;
321 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
322 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
323 				    [ix] == i + 32 * block) {
324 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
325 					pr_debug("propagate_pkey_ev: slave %d,"
326 						 " port %d, ix %d (%d)\n",
327 						 slave, port_num, ix, err);
328 					have_event = 1;
329 					break;
330 				}
331 			}
332 			if (have_event)
333 				break;
334 		}
335 	}
336 }
337 
338 static void node_desc_override(struct ib_device *dev,
339 			       struct ib_mad *mad)
340 {
341 	unsigned long flags;
342 
343 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
344 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
345 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
346 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
347 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
348 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
349 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
350 	}
351 }
352 
353 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
354 {
355 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
356 	struct ib_mad_send_buf *send_buf;
357 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
358 	int ret;
359 	unsigned long flags;
360 
361 	if (agent) {
362 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
363 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
364 					      IB_MGMT_BASE_VERSION);
365 		if (IS_ERR(send_buf))
366 			return;
367 		/*
368 		 * We rely here on the fact that MLX QPs don't use the
369 		 * address handle after the send is posted (this is
370 		 * wrong following the IB spec strictly, but we know
371 		 * it's OK for our devices).
372 		 */
373 		spin_lock_irqsave(&dev->sm_lock, flags);
374 		memcpy(send_buf->mad, mad, sizeof *mad);
375 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
376 			ret = ib_post_send_mad(send_buf, NULL);
377 		else
378 			ret = -EINVAL;
379 		spin_unlock_irqrestore(&dev->sm_lock, flags);
380 
381 		if (ret)
382 			ib_free_send_mad(send_buf);
383 	}
384 }
385 
386 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
387 							     struct ib_sa_mad *sa_mad)
388 {
389 	int ret = 0;
390 
391 	/* dispatch to different sa handlers */
392 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
393 	case IB_SA_ATTR_MC_MEMBER_REC:
394 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
395 		break;
396 	default:
397 		break;
398 	}
399 	return ret;
400 }
401 
402 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
403 {
404 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
405 	int i;
406 
407 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
408 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
409 			return i;
410 	}
411 	return -1;
412 }
413 
414 
415 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
416 				   u8 port, u16 pkey, u16 *ix)
417 {
418 	int i, ret;
419 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
420 	u16 slot_pkey;
421 
422 	if (slave == mlx4_master_func_num(dev->dev))
423 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
424 
425 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
426 
427 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
428 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
429 			continue;
430 
431 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
432 
433 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
434 		if (ret)
435 			continue;
436 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
437 			if (slot_pkey & 0x8000) {
438 				*ix = (u16) pkey_ix;
439 				return 0;
440 			} else {
441 				/* take first partial pkey index found */
442 				if (partial_ix == 0xFF)
443 					partial_ix = pkey_ix;
444 			}
445 		}
446 	}
447 
448 	if (partial_ix < 0xFF) {
449 		*ix = (u16) partial_ix;
450 		return 0;
451 	}
452 
453 	return -EINVAL;
454 }
455 
456 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
457 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
458 			  struct ib_grh *grh, struct ib_mad *mad)
459 {
460 	struct ib_sge list;
461 	struct ib_ud_wr wr;
462 	struct ib_send_wr *bad_wr;
463 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
464 	struct mlx4_ib_demux_pv_qp *tun_qp;
465 	struct mlx4_rcv_tunnel_mad *tun_mad;
466 	struct ib_ah_attr attr;
467 	struct ib_ah *ah;
468 	struct ib_qp *src_qp = NULL;
469 	unsigned tun_tx_ix = 0;
470 	int dqpn;
471 	int ret = 0;
472 	u16 tun_pkey_ix;
473 	u16 cached_pkey;
474 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
475 
476 	if (dest_qpt > IB_QPT_GSI)
477 		return -EINVAL;
478 
479 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
480 
481 	/* check if proxy qp created */
482 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
483 		return -EAGAIN;
484 
485 	if (!dest_qpt)
486 		tun_qp = &tun_ctx->qp[0];
487 	else
488 		tun_qp = &tun_ctx->qp[1];
489 
490 	/* compute P_Key index to put in tunnel header for slave */
491 	if (dest_qpt) {
492 		u16 pkey_ix;
493 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
494 		if (ret)
495 			return -EINVAL;
496 
497 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
498 		if (ret)
499 			return -EINVAL;
500 		tun_pkey_ix = pkey_ix;
501 	} else
502 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
503 
504 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
505 
506 	/* get tunnel tx data buf for slave */
507 	src_qp = tun_qp->qp;
508 
509 	/* create ah. Just need an empty one with the port num for the post send.
510 	 * The driver will set the force loopback bit in post_send */
511 	memset(&attr, 0, sizeof attr);
512 	attr.port_num = port;
513 	if (is_eth) {
514 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
515 		attr.ah_flags = IB_AH_GRH;
516 	}
517 	ah = ib_create_ah(tun_ctx->pd, &attr);
518 	if (IS_ERR(ah))
519 		return -ENOMEM;
520 
521 	/* allocate tunnel tx buf after pass failure returns */
522 	spin_lock(&tun_qp->tx_lock);
523 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
524 	    (MLX4_NUM_TUNNEL_BUFS - 1))
525 		ret = -EAGAIN;
526 	else
527 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
528 	spin_unlock(&tun_qp->tx_lock);
529 	if (ret)
530 		goto end;
531 
532 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
533 	if (tun_qp->tx_ring[tun_tx_ix].ah)
534 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
535 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
536 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
537 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
538 				   sizeof (struct mlx4_rcv_tunnel_mad),
539 				   DMA_TO_DEVICE);
540 
541 	/* copy over to tunnel buffer */
542 	if (grh)
543 		memcpy(&tun_mad->grh, grh, sizeof *grh);
544 	memcpy(&tun_mad->mad, mad, sizeof *mad);
545 
546 	/* adjust tunnel data */
547 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
548 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
549 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
550 
551 	if (is_eth) {
552 		u16 vlan = 0;
553 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
554 						NULL)) {
555 			/* VST mode */
556 			if (vlan != wc->vlan_id)
557 				/* Packet vlan is not the VST-assigned vlan.
558 				 * Drop the packet.
559 				 */
560 				goto out;
561 			 else
562 				/* Remove the vlan tag before forwarding
563 				 * the packet to the VF.
564 				 */
565 				vlan = 0xffff;
566 		} else {
567 			vlan = wc->vlan_id;
568 		}
569 
570 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
571 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
572 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
573 	} else {
574 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
575 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
576 	}
577 
578 	ib_dma_sync_single_for_device(&dev->ib_dev,
579 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
580 				      sizeof (struct mlx4_rcv_tunnel_mad),
581 				      DMA_TO_DEVICE);
582 
583 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
584 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
585 	list.lkey = tun_ctx->pd->local_dma_lkey;
586 
587 	wr.ah = ah;
588 	wr.port_num = port;
589 	wr.remote_qkey = IB_QP_SET_QKEY;
590 	wr.remote_qpn = dqpn;
591 	wr.wr.next = NULL;
592 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
593 	wr.wr.sg_list = &list;
594 	wr.wr.num_sge = 1;
595 	wr.wr.opcode = IB_WR_SEND;
596 	wr.wr.send_flags = IB_SEND_SIGNALED;
597 
598 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
599 	if (!ret)
600 		return 0;
601  out:
602 	spin_lock(&tun_qp->tx_lock);
603 	tun_qp->tx_ix_tail++;
604 	spin_unlock(&tun_qp->tx_lock);
605 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
606 end:
607 	ib_destroy_ah(ah);
608 	return ret;
609 }
610 
611 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
612 			struct ib_wc *wc, struct ib_grh *grh,
613 			struct ib_mad *mad)
614 {
615 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
616 	int err, other_port;
617 	int slave = -1;
618 	u8 *slave_id;
619 	int is_eth = 0;
620 
621 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
622 		is_eth = 0;
623 	else
624 		is_eth = 1;
625 
626 	if (is_eth) {
627 		if (!(wc->wc_flags & IB_WC_GRH)) {
628 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
629 			return -EINVAL;
630 		}
631 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
632 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
633 			return -EINVAL;
634 		}
635 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave);
636 		if (err && mlx4_is_mf_bonded(dev->dev)) {
637 			other_port = (port == 1) ? 2 : 1;
638 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, grh->dgid.raw, &slave);
639 			if (!err) {
640 				port = other_port;
641 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
642 					 slave, grh->dgid.raw, port, other_port);
643 			}
644 		}
645 		if (err) {
646 			mlx4_ib_warn(ibdev, "failed matching grh\n");
647 			return -ENOENT;
648 		}
649 		if (slave >= dev->dev->caps.sqp_demux) {
650 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
651 				     slave, dev->dev->caps.sqp_demux);
652 			return -ENOENT;
653 		}
654 
655 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
656 			return 0;
657 
658 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
659 		if (err)
660 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
661 				 slave, err);
662 		return 0;
663 	}
664 
665 	/* Initially assume that this mad is for us */
666 	slave = mlx4_master_func_num(dev->dev);
667 
668 	/* See if the slave id is encoded in a response mad */
669 	if (mad->mad_hdr.method & 0x80) {
670 		slave_id = (u8 *) &mad->mad_hdr.tid;
671 		slave = *slave_id;
672 		if (slave != 255) /*255 indicates the dom0*/
673 			*slave_id = 0; /* remap tid */
674 	}
675 
676 	/* If a grh is present, we demux according to it */
677 	if (wc->wc_flags & IB_WC_GRH) {
678 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
679 		if (slave < 0) {
680 			mlx4_ib_warn(ibdev, "failed matching grh\n");
681 			return -ENOENT;
682 		}
683 	}
684 	/* Class-specific handling */
685 	switch (mad->mad_hdr.mgmt_class) {
686 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
687 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
688 		/* 255 indicates the dom0 */
689 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
690 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
691 				return -EPERM;
692 			/* for a VF. drop unsolicited MADs */
693 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
694 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
695 					     slave, mad->mad_hdr.mgmt_class,
696 					     mad->mad_hdr.method);
697 				return -EINVAL;
698 			}
699 		}
700 		break;
701 	case IB_MGMT_CLASS_SUBN_ADM:
702 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
703 					     (struct ib_sa_mad *) mad))
704 			return 0;
705 		break;
706 	case IB_MGMT_CLASS_CM:
707 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
708 			return 0;
709 		break;
710 	case IB_MGMT_CLASS_DEVICE_MGMT:
711 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
712 			return 0;
713 		break;
714 	default:
715 		/* Drop unsupported classes for slaves in tunnel mode */
716 		if (slave != mlx4_master_func_num(dev->dev)) {
717 			pr_debug("dropping unsupported ingress mad from class:%d "
718 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
719 			return 0;
720 		}
721 	}
722 	/*make sure that no slave==255 was not handled yet.*/
723 	if (slave >= dev->dev->caps.sqp_demux) {
724 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
725 			     slave, dev->dev->caps.sqp_demux);
726 		return -ENOENT;
727 	}
728 
729 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
730 	if (err)
731 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
732 			 slave, err);
733 	return 0;
734 }
735 
736 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
737 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
738 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
739 {
740 	u16 slid, prev_lid = 0;
741 	int err;
742 	struct ib_port_attr pattr;
743 
744 	if (in_wc && in_wc->qp->qp_num) {
745 		pr_debug("received MAD: slid:%d sqpn:%d "
746 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
747 			in_wc->slid, in_wc->src_qp,
748 			in_wc->dlid_path_bits,
749 			in_wc->qp->qp_num,
750 			in_wc->wc_flags,
751 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
752 			be16_to_cpu(in_mad->mad_hdr.attr_id));
753 		if (in_wc->wc_flags & IB_WC_GRH) {
754 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
755 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
756 				 be64_to_cpu(in_grh->sgid.global.interface_id));
757 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
758 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
759 				 be64_to_cpu(in_grh->dgid.global.interface_id));
760 		}
761 	}
762 
763 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
764 
765 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
766 		forward_trap(to_mdev(ibdev), port_num, in_mad);
767 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
768 	}
769 
770 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
771 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
772 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
773 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
774 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
775 			return IB_MAD_RESULT_SUCCESS;
776 
777 		/*
778 		 * Don't process SMInfo queries -- the SMA can't handle them.
779 		 */
780 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
781 			return IB_MAD_RESULT_SUCCESS;
782 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
783 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
784 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
785 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
786 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
787 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
788 			return IB_MAD_RESULT_SUCCESS;
789 	} else
790 		return IB_MAD_RESULT_SUCCESS;
791 
792 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
793 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
794 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
795 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
796 	    !ib_query_port(ibdev, port_num, &pattr))
797 		prev_lid = pattr.lid;
798 
799 	err = mlx4_MAD_IFC(to_mdev(ibdev),
800 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
801 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
802 			   MLX4_MAD_IFC_NET_VIEW,
803 			   port_num, in_wc, in_grh, in_mad, out_mad);
804 	if (err)
805 		return IB_MAD_RESULT_FAILURE;
806 
807 	if (!out_mad->mad_hdr.status) {
808 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
809 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
810 		/* slaves get node desc from FW */
811 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
812 			node_desc_override(ibdev, out_mad);
813 	}
814 
815 	/* set return bit in status of directed route responses */
816 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
817 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
818 
819 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
820 		/* no response for trap repress */
821 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
822 
823 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
824 }
825 
826 static void edit_counter(struct mlx4_counter *cnt, void *counters,
827 			 __be16 attr_id)
828 {
829 	switch (attr_id) {
830 	case IB_PMA_PORT_COUNTERS:
831 	{
832 		struct ib_pma_portcounters *pma_cnt =
833 			(struct ib_pma_portcounters *)counters;
834 
835 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
836 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
837 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
838 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
839 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
840 				     be64_to_cpu(cnt->tx_frames));
841 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
842 				     be64_to_cpu(cnt->rx_frames));
843 		break;
844 	}
845 	case IB_PMA_PORT_COUNTERS_EXT:
846 	{
847 		struct ib_pma_portcounters_ext *pma_cnt_ext =
848 			(struct ib_pma_portcounters_ext *)counters;
849 
850 		pma_cnt_ext->port_xmit_data =
851 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
852 		pma_cnt_ext->port_rcv_data =
853 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
854 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
855 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
856 		break;
857 	}
858 	}
859 }
860 
861 static int iboe_process_mad_port_info(void *out_mad)
862 {
863 	struct ib_class_port_info cpi = {};
864 
865 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
866 	memcpy(out_mad, &cpi, sizeof(cpi));
867 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
868 }
869 
870 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
871 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
872 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
873 {
874 	struct mlx4_counter counter_stats;
875 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
876 	struct counter_index *tmp_counter;
877 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
878 
879 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
880 		return -EINVAL;
881 
882 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
883 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
884 
885 	memset(&counter_stats, 0, sizeof(counter_stats));
886 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
887 	list_for_each_entry(tmp_counter,
888 			    &dev->counters_table[port_num - 1].counters_list,
889 			    list) {
890 		err = mlx4_get_counter_stats(dev->dev,
891 					     tmp_counter->index,
892 					     &counter_stats, 0);
893 		if (err) {
894 			err = IB_MAD_RESULT_FAILURE;
895 			stats_avail = 0;
896 			break;
897 		}
898 		stats_avail = 1;
899 	}
900 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
901 	if (stats_avail) {
902 		memset(out_mad->data, 0, sizeof out_mad->data);
903 		switch (counter_stats.counter_mode & 0xf) {
904 		case 0:
905 			edit_counter(&counter_stats,
906 				     (void *)(out_mad->data + 40),
907 				     in_mad->mad_hdr.attr_id);
908 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
909 			break;
910 		default:
911 			err = IB_MAD_RESULT_FAILURE;
912 		}
913 	}
914 
915 	return err;
916 }
917 
918 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
919 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
920 			const struct ib_mad_hdr *in, size_t in_mad_size,
921 			struct ib_mad_hdr *out, size_t *out_mad_size,
922 			u16 *out_mad_pkey_index)
923 {
924 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
925 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
926 	struct ib_mad *out_mad = (struct ib_mad *)out;
927 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
928 
929 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
930 			 *out_mad_size != sizeof(*out_mad)))
931 		return IB_MAD_RESULT_FAILURE;
932 
933 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
934 	 * queries, should be called only by VFs and for that specific purpose
935 	 */
936 	if (link == IB_LINK_LAYER_INFINIBAND) {
937 		if (mlx4_is_slave(dev->dev) &&
938 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
939 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
940 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
941 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
942 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
943 						in_grh, in_mad, out_mad);
944 
945 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
946 				      in_grh, in_mad, out_mad);
947 	}
948 
949 	if (link == IB_LINK_LAYER_ETHERNET)
950 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
951 					in_grh, in_mad, out_mad);
952 
953 	return -EINVAL;
954 }
955 
956 static void send_handler(struct ib_mad_agent *agent,
957 			 struct ib_mad_send_wc *mad_send_wc)
958 {
959 	if (mad_send_wc->send_buf->context[0])
960 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
961 	ib_free_send_mad(mad_send_wc->send_buf);
962 }
963 
964 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
965 {
966 	struct ib_mad_agent *agent;
967 	int p, q;
968 	int ret;
969 	enum rdma_link_layer ll;
970 
971 	for (p = 0; p < dev->num_ports; ++p) {
972 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
973 		for (q = 0; q <= 1; ++q) {
974 			if (ll == IB_LINK_LAYER_INFINIBAND) {
975 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
976 							      q ? IB_QPT_GSI : IB_QPT_SMI,
977 							      NULL, 0, send_handler,
978 							      NULL, NULL, 0);
979 				if (IS_ERR(agent)) {
980 					ret = PTR_ERR(agent);
981 					goto err;
982 				}
983 				dev->send_agent[p][q] = agent;
984 			} else
985 				dev->send_agent[p][q] = NULL;
986 		}
987 	}
988 
989 	return 0;
990 
991 err:
992 	for (p = 0; p < dev->num_ports; ++p)
993 		for (q = 0; q <= 1; ++q)
994 			if (dev->send_agent[p][q])
995 				ib_unregister_mad_agent(dev->send_agent[p][q]);
996 
997 	return ret;
998 }
999 
1000 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1001 {
1002 	struct ib_mad_agent *agent;
1003 	int p, q;
1004 
1005 	for (p = 0; p < dev->num_ports; ++p) {
1006 		for (q = 0; q <= 1; ++q) {
1007 			agent = dev->send_agent[p][q];
1008 			if (agent) {
1009 				dev->send_agent[p][q] = NULL;
1010 				ib_unregister_mad_agent(agent);
1011 			}
1012 		}
1013 
1014 		if (dev->sm_ah[p])
1015 			ib_destroy_ah(dev->sm_ah[p]);
1016 	}
1017 }
1018 
1019 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1020 {
1021 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1022 
1023 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1024 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1025 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1026 }
1027 
1028 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1029 {
1030 	/* re-configure the alias-guid and mcg's */
1031 	if (mlx4_is_master(dev->dev)) {
1032 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1033 
1034 		if (!dev->sriov.is_going_down) {
1035 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1036 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1037 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1038 		}
1039 	}
1040 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1041 }
1042 
1043 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1044 			      struct mlx4_eqe *eqe)
1045 {
1046 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1047 			    GET_MASK_FROM_EQE(eqe));
1048 }
1049 
1050 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1051 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1052 {
1053 	struct ib_smp *in_mad  = NULL;
1054 	struct ib_smp *out_mad  = NULL;
1055 	u16 i;
1056 
1057 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1058 		return;
1059 
1060 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1061 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1062 	if (!in_mad || !out_mad) {
1063 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1064 		goto out;
1065 	}
1066 
1067 	guid_tbl_blk_num  *= 4;
1068 
1069 	for (i = 0; i < 4; i++) {
1070 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1071 			continue;
1072 		memset(in_mad, 0, sizeof *in_mad);
1073 		memset(out_mad, 0, sizeof *out_mad);
1074 
1075 		in_mad->base_version  = 1;
1076 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1077 		in_mad->class_version = 1;
1078 		in_mad->method        = IB_MGMT_METHOD_GET;
1079 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1080 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1081 
1082 		if (mlx4_MAD_IFC(dev,
1083 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1084 				 port_num, NULL, NULL, in_mad, out_mad)) {
1085 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1086 			goto out;
1087 		}
1088 
1089 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1090 						    port_num,
1091 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1092 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1093 						     port_num,
1094 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1095 	}
1096 
1097 out:
1098 	kfree(in_mad);
1099 	kfree(out_mad);
1100 	return;
1101 }
1102 
1103 void handle_port_mgmt_change_event(struct work_struct *work)
1104 {
1105 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1106 	struct mlx4_ib_dev *dev = ew->ib_dev;
1107 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1108 	u8 port = eqe->event.port_mgmt_change.port;
1109 	u32 changed_attr;
1110 	u32 tbl_block;
1111 	u32 change_bitmap;
1112 
1113 	switch (eqe->subtype) {
1114 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1115 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1116 
1117 		/* Update the SM ah - This should be done before handling
1118 		   the other changed attributes so that MADs can be sent to the SM */
1119 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1120 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1121 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1122 			update_sm_ah(dev, port, lid, sl);
1123 		}
1124 
1125 		/* Check if it is a lid change event */
1126 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1127 			handle_lid_change_event(dev, port);
1128 
1129 		/* Generate GUID changed event */
1130 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1131 			if (mlx4_is_master(dev->dev)) {
1132 				union ib_gid gid;
1133 				int err = 0;
1134 
1135 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1136 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1137 				else
1138 					gid.global.subnet_prefix =
1139 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1140 				if (err) {
1141 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1142 						port, err);
1143 				} else {
1144 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1145 						 port,
1146 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1147 						 be64_to_cpu(gid.global.subnet_prefix));
1148 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1149 						     be64_to_cpu(gid.global.subnet_prefix));
1150 				}
1151 			}
1152 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1153 			/*if master, notify all slaves*/
1154 			if (mlx4_is_master(dev->dev))
1155 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1156 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1157 		}
1158 
1159 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1160 			handle_client_rereg_event(dev, port);
1161 		break;
1162 
1163 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1164 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1165 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1166 			propagate_pkey_ev(dev, port, eqe);
1167 		break;
1168 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1169 		/* paravirtualized master's guid is guid 0 -- does not change */
1170 		if (!mlx4_is_master(dev->dev))
1171 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1172 		/*if master, notify relevant slaves*/
1173 		else if (!dev->sriov.is_going_down) {
1174 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1175 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1176 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1177 		}
1178 		break;
1179 	default:
1180 		pr_warn("Unsupported subtype 0x%x for "
1181 			"Port Management Change event\n", eqe->subtype);
1182 	}
1183 
1184 	kfree(ew);
1185 }
1186 
1187 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1188 			    enum ib_event_type type)
1189 {
1190 	struct ib_event event;
1191 
1192 	event.device		= &dev->ib_dev;
1193 	event.element.port_num	= port_num;
1194 	event.event		= type;
1195 
1196 	ib_dispatch_event(&event);
1197 }
1198 
1199 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1200 {
1201 	unsigned long flags;
1202 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1203 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1204 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1205 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1206 		queue_work(ctx->wq, &ctx->work);
1207 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1208 }
1209 
1210 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1211 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1212 				  int index)
1213 {
1214 	struct ib_sge sg_list;
1215 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1216 	int size;
1217 
1218 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1219 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1220 
1221 	sg_list.addr = tun_qp->ring[index].map;
1222 	sg_list.length = size;
1223 	sg_list.lkey = ctx->pd->local_dma_lkey;
1224 
1225 	recv_wr.next = NULL;
1226 	recv_wr.sg_list = &sg_list;
1227 	recv_wr.num_sge = 1;
1228 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1229 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1230 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1231 				      size, DMA_FROM_DEVICE);
1232 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1233 }
1234 
1235 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1236 		int slave, struct ib_sa_mad *sa_mad)
1237 {
1238 	int ret = 0;
1239 
1240 	/* dispatch to different sa handlers */
1241 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1242 	case IB_SA_ATTR_MC_MEMBER_REC:
1243 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1244 		break;
1245 	default:
1246 		break;
1247 	}
1248 	return ret;
1249 }
1250 
1251 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1252 {
1253 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1254 
1255 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1256 }
1257 
1258 
1259 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1260 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1261 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1262 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1263 {
1264 	struct ib_sge list;
1265 	struct ib_ud_wr wr;
1266 	struct ib_send_wr *bad_wr;
1267 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1268 	struct mlx4_ib_demux_pv_qp *sqp;
1269 	struct mlx4_mad_snd_buf *sqp_mad;
1270 	struct ib_ah *ah;
1271 	struct ib_qp *send_qp = NULL;
1272 	unsigned wire_tx_ix = 0;
1273 	int ret = 0;
1274 	u16 wire_pkey_ix;
1275 	int src_qpnum;
1276 	u8 sgid_index;
1277 
1278 
1279 	sqp_ctx = dev->sriov.sqps[port-1];
1280 
1281 	/* check if proxy qp created */
1282 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1283 		return -EAGAIN;
1284 
1285 	if (dest_qpt == IB_QPT_SMI) {
1286 		src_qpnum = 0;
1287 		sqp = &sqp_ctx->qp[0];
1288 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1289 	} else {
1290 		src_qpnum = 1;
1291 		sqp = &sqp_ctx->qp[1];
1292 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1293 	}
1294 
1295 	send_qp = sqp->qp;
1296 
1297 	/* create ah */
1298 	sgid_index = attr->grh.sgid_index;
1299 	attr->grh.sgid_index = 0;
1300 	ah = ib_create_ah(sqp_ctx->pd, attr);
1301 	if (IS_ERR(ah))
1302 		return -ENOMEM;
1303 	attr->grh.sgid_index = sgid_index;
1304 	to_mah(ah)->av.ib.gid_index = sgid_index;
1305 	/* get rid of force-loopback bit */
1306 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1307 	spin_lock(&sqp->tx_lock);
1308 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1309 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1310 		ret = -EAGAIN;
1311 	else
1312 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1313 	spin_unlock(&sqp->tx_lock);
1314 	if (ret)
1315 		goto out;
1316 
1317 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1318 	if (sqp->tx_ring[wire_tx_ix].ah)
1319 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1320 	sqp->tx_ring[wire_tx_ix].ah = ah;
1321 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1322 				   sqp->tx_ring[wire_tx_ix].buf.map,
1323 				   sizeof (struct mlx4_mad_snd_buf),
1324 				   DMA_TO_DEVICE);
1325 
1326 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1327 
1328 	ib_dma_sync_single_for_device(&dev->ib_dev,
1329 				      sqp->tx_ring[wire_tx_ix].buf.map,
1330 				      sizeof (struct mlx4_mad_snd_buf),
1331 				      DMA_TO_DEVICE);
1332 
1333 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1334 	list.length = sizeof (struct mlx4_mad_snd_buf);
1335 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1336 
1337 	wr.ah = ah;
1338 	wr.port_num = port;
1339 	wr.pkey_index = wire_pkey_ix;
1340 	wr.remote_qkey = qkey;
1341 	wr.remote_qpn = remote_qpn;
1342 	wr.wr.next = NULL;
1343 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1344 	wr.wr.sg_list = &list;
1345 	wr.wr.num_sge = 1;
1346 	wr.wr.opcode = IB_WR_SEND;
1347 	wr.wr.send_flags = IB_SEND_SIGNALED;
1348 	if (s_mac)
1349 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1350 	if (vlan_id < 0x1000)
1351 		vlan_id |= (attr->sl & 7) << 13;
1352 	to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1353 
1354 
1355 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1356 	if (!ret)
1357 		return 0;
1358 
1359 	spin_lock(&sqp->tx_lock);
1360 	sqp->tx_ix_tail++;
1361 	spin_unlock(&sqp->tx_lock);
1362 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1363 out:
1364 	ib_destroy_ah(ah);
1365 	return ret;
1366 }
1367 
1368 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1369 {
1370 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1371 		return slave;
1372 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1373 }
1374 
1375 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1376 				    struct ib_ah_attr *ah_attr)
1377 {
1378 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1379 		ah_attr->grh.sgid_index = slave;
1380 	else
1381 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1382 }
1383 
1384 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1385 {
1386 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1387 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1388 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1389 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1390 	struct mlx4_ib_ah ah;
1391 	struct ib_ah_attr ah_attr;
1392 	u8 *slave_id;
1393 	int slave;
1394 	int port;
1395 	u16 vlan_id;
1396 
1397 	/* Get slave that sent this packet */
1398 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1399 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1400 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1401 	    wc->src_qp & 0x4) {
1402 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1403 		return;
1404 	}
1405 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1406 	if (slave != ctx->slave) {
1407 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1408 			     "belongs to another slave\n", wc->src_qp);
1409 		return;
1410 	}
1411 
1412 	/* Map transaction ID */
1413 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1414 				   sizeof (struct mlx4_tunnel_mad),
1415 				   DMA_FROM_DEVICE);
1416 	switch (tunnel->mad.mad_hdr.method) {
1417 	case IB_MGMT_METHOD_SET:
1418 	case IB_MGMT_METHOD_GET:
1419 	case IB_MGMT_METHOD_REPORT:
1420 	case IB_SA_METHOD_GET_TABLE:
1421 	case IB_SA_METHOD_DELETE:
1422 	case IB_SA_METHOD_GET_MULTI:
1423 	case IB_SA_METHOD_GET_TRACE_TBL:
1424 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1425 		if (*slave_id) {
1426 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1427 				     "class:%d slave:%d\n", *slave_id,
1428 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1429 			return;
1430 		} else
1431 			*slave_id = slave;
1432 	default:
1433 		/* nothing */;
1434 	}
1435 
1436 	/* Class-specific handling */
1437 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1438 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1439 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1440 		if (slave != mlx4_master_func_num(dev->dev) &&
1441 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1442 			return;
1443 		break;
1444 	case IB_MGMT_CLASS_SUBN_ADM:
1445 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1446 			      (struct ib_sa_mad *) &tunnel->mad))
1447 			return;
1448 		break;
1449 	case IB_MGMT_CLASS_CM:
1450 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1451 			      (struct ib_mad *) &tunnel->mad))
1452 			return;
1453 		break;
1454 	case IB_MGMT_CLASS_DEVICE_MGMT:
1455 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1456 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1457 			return;
1458 		break;
1459 	default:
1460 		/* Drop unsupported classes for slaves in tunnel mode */
1461 		if (slave != mlx4_master_func_num(dev->dev)) {
1462 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1463 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1464 			return;
1465 		}
1466 	}
1467 
1468 	/* We are using standard ib_core services to send the mad, so generate a
1469 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1470 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1471 	ah.ibah.device = ctx->ib_dev;
1472 
1473 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1474 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1475 	if (port < 0)
1476 		return;
1477 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1478 
1479 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1480 	if (ah_attr.ah_flags & IB_AH_GRH)
1481 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1482 
1483 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1484 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1485 	/* if slave have default vlan use it */
1486 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1487 				    &vlan_id, &ah_attr.sl);
1488 
1489 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1490 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1491 			     IB_QPT_SMI : IB_QPT_GSI,
1492 			     be16_to_cpu(tunnel->hdr.pkey_index),
1493 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1494 			     be32_to_cpu(tunnel->hdr.qkey),
1495 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1496 }
1497 
1498 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1499 				 enum ib_qp_type qp_type, int is_tun)
1500 {
1501 	int i;
1502 	struct mlx4_ib_demux_pv_qp *tun_qp;
1503 	int rx_buf_size, tx_buf_size;
1504 
1505 	if (qp_type > IB_QPT_GSI)
1506 		return -EINVAL;
1507 
1508 	tun_qp = &ctx->qp[qp_type];
1509 
1510 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1511 			       GFP_KERNEL);
1512 	if (!tun_qp->ring)
1513 		return -ENOMEM;
1514 
1515 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1516 				  sizeof (struct mlx4_ib_tun_tx_buf),
1517 				  GFP_KERNEL);
1518 	if (!tun_qp->tx_ring) {
1519 		kfree(tun_qp->ring);
1520 		tun_qp->ring = NULL;
1521 		return -ENOMEM;
1522 	}
1523 
1524 	if (is_tun) {
1525 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1526 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1527 	} else {
1528 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1529 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1530 	}
1531 
1532 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1533 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1534 		if (!tun_qp->ring[i].addr)
1535 			goto err;
1536 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1537 							tun_qp->ring[i].addr,
1538 							rx_buf_size,
1539 							DMA_FROM_DEVICE);
1540 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1541 			kfree(tun_qp->ring[i].addr);
1542 			goto err;
1543 		}
1544 	}
1545 
1546 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1547 		tun_qp->tx_ring[i].buf.addr =
1548 			kmalloc(tx_buf_size, GFP_KERNEL);
1549 		if (!tun_qp->tx_ring[i].buf.addr)
1550 			goto tx_err;
1551 		tun_qp->tx_ring[i].buf.map =
1552 			ib_dma_map_single(ctx->ib_dev,
1553 					  tun_qp->tx_ring[i].buf.addr,
1554 					  tx_buf_size,
1555 					  DMA_TO_DEVICE);
1556 		if (ib_dma_mapping_error(ctx->ib_dev,
1557 					 tun_qp->tx_ring[i].buf.map)) {
1558 			kfree(tun_qp->tx_ring[i].buf.addr);
1559 			goto tx_err;
1560 		}
1561 		tun_qp->tx_ring[i].ah = NULL;
1562 	}
1563 	spin_lock_init(&tun_qp->tx_lock);
1564 	tun_qp->tx_ix_head = 0;
1565 	tun_qp->tx_ix_tail = 0;
1566 	tun_qp->proxy_qpt = qp_type;
1567 
1568 	return 0;
1569 
1570 tx_err:
1571 	while (i > 0) {
1572 		--i;
1573 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1574 				    tx_buf_size, DMA_TO_DEVICE);
1575 		kfree(tun_qp->tx_ring[i].buf.addr);
1576 	}
1577 	kfree(tun_qp->tx_ring);
1578 	tun_qp->tx_ring = NULL;
1579 	i = MLX4_NUM_TUNNEL_BUFS;
1580 err:
1581 	while (i > 0) {
1582 		--i;
1583 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1584 				    rx_buf_size, DMA_FROM_DEVICE);
1585 		kfree(tun_qp->ring[i].addr);
1586 	}
1587 	kfree(tun_qp->ring);
1588 	tun_qp->ring = NULL;
1589 	return -ENOMEM;
1590 }
1591 
1592 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1593 				     enum ib_qp_type qp_type, int is_tun)
1594 {
1595 	int i;
1596 	struct mlx4_ib_demux_pv_qp *tun_qp;
1597 	int rx_buf_size, tx_buf_size;
1598 
1599 	if (qp_type > IB_QPT_GSI)
1600 		return;
1601 
1602 	tun_qp = &ctx->qp[qp_type];
1603 	if (is_tun) {
1604 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1605 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1606 	} else {
1607 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1608 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1609 	}
1610 
1611 
1612 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1613 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1614 				    rx_buf_size, DMA_FROM_DEVICE);
1615 		kfree(tun_qp->ring[i].addr);
1616 	}
1617 
1618 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1619 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1620 				    tx_buf_size, DMA_TO_DEVICE);
1621 		kfree(tun_qp->tx_ring[i].buf.addr);
1622 		if (tun_qp->tx_ring[i].ah)
1623 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1624 	}
1625 	kfree(tun_qp->tx_ring);
1626 	kfree(tun_qp->ring);
1627 }
1628 
1629 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1630 {
1631 	struct mlx4_ib_demux_pv_ctx *ctx;
1632 	struct mlx4_ib_demux_pv_qp *tun_qp;
1633 	struct ib_wc wc;
1634 	int ret;
1635 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1636 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1637 
1638 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1639 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1640 		if (wc.status == IB_WC_SUCCESS) {
1641 			switch (wc.opcode) {
1642 			case IB_WC_RECV:
1643 				mlx4_ib_multiplex_mad(ctx, &wc);
1644 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1645 							     wc.wr_id &
1646 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1647 				if (ret)
1648 					pr_err("Failed reposting tunnel "
1649 					       "buf:%lld\n", wc.wr_id);
1650 				break;
1651 			case IB_WC_SEND:
1652 				pr_debug("received tunnel send completion:"
1653 					 "wrid=0x%llx, status=0x%x\n",
1654 					 wc.wr_id, wc.status);
1655 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1656 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1657 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1658 					= NULL;
1659 				spin_lock(&tun_qp->tx_lock);
1660 				tun_qp->tx_ix_tail++;
1661 				spin_unlock(&tun_qp->tx_lock);
1662 
1663 				break;
1664 			default:
1665 				break;
1666 			}
1667 		} else  {
1668 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1669 				 " status = %d, wrid = 0x%llx\n",
1670 				 ctx->slave, wc.status, wc.wr_id);
1671 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1672 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1673 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1674 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1675 					= NULL;
1676 				spin_lock(&tun_qp->tx_lock);
1677 				tun_qp->tx_ix_tail++;
1678 				spin_unlock(&tun_qp->tx_lock);
1679 			}
1680 		}
1681 	}
1682 }
1683 
1684 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1685 {
1686 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1687 
1688 	/* It's worse than that! He's dead, Jim! */
1689 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1690 	       event->event, sqp->port);
1691 }
1692 
1693 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1694 			    enum ib_qp_type qp_type, int create_tun)
1695 {
1696 	int i, ret;
1697 	struct mlx4_ib_demux_pv_qp *tun_qp;
1698 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1699 	struct ib_qp_attr attr;
1700 	int qp_attr_mask_INIT;
1701 
1702 	if (qp_type > IB_QPT_GSI)
1703 		return -EINVAL;
1704 
1705 	tun_qp = &ctx->qp[qp_type];
1706 
1707 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1708 	qp_init_attr.init_attr.send_cq = ctx->cq;
1709 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1710 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1711 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1712 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1713 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1714 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1715 	if (create_tun) {
1716 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1717 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1718 		qp_init_attr.port = ctx->port;
1719 		qp_init_attr.slave = ctx->slave;
1720 		qp_init_attr.proxy_qp_type = qp_type;
1721 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1722 			   IB_QP_QKEY | IB_QP_PORT;
1723 	} else {
1724 		qp_init_attr.init_attr.qp_type = qp_type;
1725 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1726 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1727 	}
1728 	qp_init_attr.init_attr.port_num = ctx->port;
1729 	qp_init_attr.init_attr.qp_context = ctx;
1730 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1731 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1732 	if (IS_ERR(tun_qp->qp)) {
1733 		ret = PTR_ERR(tun_qp->qp);
1734 		tun_qp->qp = NULL;
1735 		pr_err("Couldn't create %s QP (%d)\n",
1736 		       create_tun ? "tunnel" : "special", ret);
1737 		return ret;
1738 	}
1739 
1740 	memset(&attr, 0, sizeof attr);
1741 	attr.qp_state = IB_QPS_INIT;
1742 	ret = 0;
1743 	if (create_tun)
1744 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1745 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1746 					      &attr.pkey_index);
1747 	if (ret || !create_tun)
1748 		attr.pkey_index =
1749 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1750 	attr.qkey = IB_QP1_QKEY;
1751 	attr.port_num = ctx->port;
1752 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1753 	if (ret) {
1754 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1755 		       create_tun ? "tunnel" : "special", ret);
1756 		goto err_qp;
1757 	}
1758 	attr.qp_state = IB_QPS_RTR;
1759 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1760 	if (ret) {
1761 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1762 		       create_tun ? "tunnel" : "special", ret);
1763 		goto err_qp;
1764 	}
1765 	attr.qp_state = IB_QPS_RTS;
1766 	attr.sq_psn = 0;
1767 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1768 	if (ret) {
1769 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1770 		       create_tun ? "tunnel" : "special", ret);
1771 		goto err_qp;
1772 	}
1773 
1774 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1775 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1776 		if (ret) {
1777 			pr_err(" mlx4_ib_post_pv_buf error"
1778 			       " (err = %d, i = %d)\n", ret, i);
1779 			goto err_qp;
1780 		}
1781 	}
1782 	return 0;
1783 
1784 err_qp:
1785 	ib_destroy_qp(tun_qp->qp);
1786 	tun_qp->qp = NULL;
1787 	return ret;
1788 }
1789 
1790 /*
1791  * IB MAD completion callback for real SQPs
1792  */
1793 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1794 {
1795 	struct mlx4_ib_demux_pv_ctx *ctx;
1796 	struct mlx4_ib_demux_pv_qp *sqp;
1797 	struct ib_wc wc;
1798 	struct ib_grh *grh;
1799 	struct ib_mad *mad;
1800 
1801 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1802 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1803 
1804 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1805 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1806 		if (wc.status == IB_WC_SUCCESS) {
1807 			switch (wc.opcode) {
1808 			case IB_WC_SEND:
1809 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1810 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1811 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1812 					= NULL;
1813 				spin_lock(&sqp->tx_lock);
1814 				sqp->tx_ix_tail++;
1815 				spin_unlock(&sqp->tx_lock);
1816 				break;
1817 			case IB_WC_RECV:
1818 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1819 						(sqp->ring[wc.wr_id &
1820 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1821 				grh = &(((struct mlx4_mad_rcv_buf *)
1822 						(sqp->ring[wc.wr_id &
1823 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1824 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1825 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1826 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1827 					pr_err("Failed reposting SQP "
1828 					       "buf:%lld\n", wc.wr_id);
1829 				break;
1830 			default:
1831 				BUG_ON(1);
1832 				break;
1833 			}
1834 		} else  {
1835 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1836 				 " status = %d, wrid = 0x%llx\n",
1837 				 ctx->slave, wc.status, wc.wr_id);
1838 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1839 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1840 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1841 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1842 					= NULL;
1843 				spin_lock(&sqp->tx_lock);
1844 				sqp->tx_ix_tail++;
1845 				spin_unlock(&sqp->tx_lock);
1846 			}
1847 		}
1848 	}
1849 }
1850 
1851 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1852 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1853 {
1854 	struct mlx4_ib_demux_pv_ctx *ctx;
1855 
1856 	*ret_ctx = NULL;
1857 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1858 	if (!ctx) {
1859 		pr_err("failed allocating pv resource context "
1860 		       "for port %d, slave %d\n", port, slave);
1861 		return -ENOMEM;
1862 	}
1863 
1864 	ctx->ib_dev = &dev->ib_dev;
1865 	ctx->port = port;
1866 	ctx->slave = slave;
1867 	*ret_ctx = ctx;
1868 	return 0;
1869 }
1870 
1871 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1872 {
1873 	if (dev->sriov.demux[port - 1].tun[slave]) {
1874 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1875 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1876 	}
1877 }
1878 
1879 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1880 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1881 {
1882 	int ret, cq_size;
1883 	struct ib_cq_init_attr cq_attr = {};
1884 
1885 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1886 		return -EEXIST;
1887 
1888 	ctx->state = DEMUX_PV_STATE_STARTING;
1889 	/* have QP0 only if link layer is IB */
1890 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1891 	    IB_LINK_LAYER_INFINIBAND)
1892 		ctx->has_smi = 1;
1893 
1894 	if (ctx->has_smi) {
1895 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1896 		if (ret) {
1897 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1898 			goto err_out;
1899 		}
1900 	}
1901 
1902 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1903 	if (ret) {
1904 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1905 		goto err_out_qp0;
1906 	}
1907 
1908 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1909 	if (ctx->has_smi)
1910 		cq_size *= 2;
1911 
1912 	cq_attr.cqe = cq_size;
1913 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1914 			       NULL, ctx, &cq_attr);
1915 	if (IS_ERR(ctx->cq)) {
1916 		ret = PTR_ERR(ctx->cq);
1917 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1918 		goto err_buf;
1919 	}
1920 
1921 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1922 	if (IS_ERR(ctx->pd)) {
1923 		ret = PTR_ERR(ctx->pd);
1924 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1925 		goto err_cq;
1926 	}
1927 
1928 	if (ctx->has_smi) {
1929 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1930 		if (ret) {
1931 			pr_err("Couldn't create %s QP0 (%d)\n",
1932 			       create_tun ? "tunnel for" : "",  ret);
1933 			goto err_pd;
1934 		}
1935 	}
1936 
1937 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1938 	if (ret) {
1939 		pr_err("Couldn't create %s QP1 (%d)\n",
1940 		       create_tun ? "tunnel for" : "",  ret);
1941 		goto err_qp0;
1942 	}
1943 
1944 	if (create_tun)
1945 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1946 	else
1947 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1948 
1949 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1950 
1951 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1952 	if (ret) {
1953 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1954 		goto err_wq;
1955 	}
1956 	ctx->state = DEMUX_PV_STATE_ACTIVE;
1957 	return 0;
1958 
1959 err_wq:
1960 	ctx->wq = NULL;
1961 	ib_destroy_qp(ctx->qp[1].qp);
1962 	ctx->qp[1].qp = NULL;
1963 
1964 
1965 err_qp0:
1966 	if (ctx->has_smi)
1967 		ib_destroy_qp(ctx->qp[0].qp);
1968 	ctx->qp[0].qp = NULL;
1969 
1970 err_pd:
1971 	ib_dealloc_pd(ctx->pd);
1972 	ctx->pd = NULL;
1973 
1974 err_cq:
1975 	ib_destroy_cq(ctx->cq);
1976 	ctx->cq = NULL;
1977 
1978 err_buf:
1979 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1980 
1981 err_out_qp0:
1982 	if (ctx->has_smi)
1983 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1984 err_out:
1985 	ctx->state = DEMUX_PV_STATE_DOWN;
1986 	return ret;
1987 }
1988 
1989 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1990 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1991 {
1992 	if (!ctx)
1993 		return;
1994 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
1995 		ctx->state = DEMUX_PV_STATE_DOWNING;
1996 		if (flush)
1997 			flush_workqueue(ctx->wq);
1998 		if (ctx->has_smi) {
1999 			ib_destroy_qp(ctx->qp[0].qp);
2000 			ctx->qp[0].qp = NULL;
2001 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2002 		}
2003 		ib_destroy_qp(ctx->qp[1].qp);
2004 		ctx->qp[1].qp = NULL;
2005 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2006 		ib_dealloc_pd(ctx->pd);
2007 		ctx->pd = NULL;
2008 		ib_destroy_cq(ctx->cq);
2009 		ctx->cq = NULL;
2010 		ctx->state = DEMUX_PV_STATE_DOWN;
2011 	}
2012 }
2013 
2014 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2015 				  int port, int do_init)
2016 {
2017 	int ret = 0;
2018 
2019 	if (!do_init) {
2020 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2021 		/* for master, destroy real sqp resources */
2022 		if (slave == mlx4_master_func_num(dev->dev))
2023 			destroy_pv_resources(dev, slave, port,
2024 					     dev->sriov.sqps[port - 1], 1);
2025 		/* destroy the tunnel qp resources */
2026 		destroy_pv_resources(dev, slave, port,
2027 				     dev->sriov.demux[port - 1].tun[slave], 1);
2028 		return 0;
2029 	}
2030 
2031 	/* create the tunnel qp resources */
2032 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2033 				  dev->sriov.demux[port - 1].tun[slave]);
2034 
2035 	/* for master, create the real sqp resources */
2036 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2037 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2038 					  dev->sriov.sqps[port - 1]);
2039 	return ret;
2040 }
2041 
2042 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2043 {
2044 	struct mlx4_ib_demux_work *dmxw;
2045 
2046 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2047 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2048 			       dmxw->do_init);
2049 	kfree(dmxw);
2050 	return;
2051 }
2052 
2053 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2054 				       struct mlx4_ib_demux_ctx *ctx,
2055 				       int port)
2056 {
2057 	char name[12];
2058 	int ret = 0;
2059 	int i;
2060 
2061 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2062 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2063 	if (!ctx->tun)
2064 		return -ENOMEM;
2065 
2066 	ctx->dev = dev;
2067 	ctx->port = port;
2068 	ctx->ib_dev = &dev->ib_dev;
2069 
2070 	for (i = 0;
2071 	     i < min(dev->dev->caps.sqp_demux,
2072 	     (u16)(dev->dev->persist->num_vfs + 1));
2073 	     i++) {
2074 		struct mlx4_active_ports actv_ports =
2075 			mlx4_get_active_ports(dev->dev, i);
2076 
2077 		if (!test_bit(port - 1, actv_ports.ports))
2078 			continue;
2079 
2080 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2081 		if (ret) {
2082 			ret = -ENOMEM;
2083 			goto err_mcg;
2084 		}
2085 	}
2086 
2087 	ret = mlx4_ib_mcg_port_init(ctx);
2088 	if (ret) {
2089 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2090 		goto err_mcg;
2091 	}
2092 
2093 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2094 	ctx->wq = create_singlethread_workqueue(name);
2095 	if (!ctx->wq) {
2096 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2097 		ret = -ENOMEM;
2098 		goto err_wq;
2099 	}
2100 
2101 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2102 	ctx->ud_wq = create_singlethread_workqueue(name);
2103 	if (!ctx->ud_wq) {
2104 		pr_err("Failed to create up/down WQ for port %d\n", port);
2105 		ret = -ENOMEM;
2106 		goto err_udwq;
2107 	}
2108 
2109 	return 0;
2110 
2111 err_udwq:
2112 	destroy_workqueue(ctx->wq);
2113 	ctx->wq = NULL;
2114 
2115 err_wq:
2116 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2117 err_mcg:
2118 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2119 		free_pv_object(dev, i, port);
2120 	kfree(ctx->tun);
2121 	ctx->tun = NULL;
2122 	return ret;
2123 }
2124 
2125 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2126 {
2127 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2128 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2129 		flush_workqueue(sqp_ctx->wq);
2130 		if (sqp_ctx->has_smi) {
2131 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2132 			sqp_ctx->qp[0].qp = NULL;
2133 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2134 		}
2135 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2136 		sqp_ctx->qp[1].qp = NULL;
2137 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2138 		ib_dealloc_pd(sqp_ctx->pd);
2139 		sqp_ctx->pd = NULL;
2140 		ib_destroy_cq(sqp_ctx->cq);
2141 		sqp_ctx->cq = NULL;
2142 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2143 	}
2144 }
2145 
2146 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2147 {
2148 	int i;
2149 	if (ctx) {
2150 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2151 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2152 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2153 			if (!ctx->tun[i])
2154 				continue;
2155 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2156 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2157 		}
2158 		flush_workqueue(ctx->wq);
2159 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2160 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2161 			free_pv_object(dev, i, ctx->port);
2162 		}
2163 		kfree(ctx->tun);
2164 		destroy_workqueue(ctx->ud_wq);
2165 		destroy_workqueue(ctx->wq);
2166 	}
2167 }
2168 
2169 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2170 {
2171 	int i;
2172 
2173 	if (!mlx4_is_master(dev->dev))
2174 		return;
2175 	/* initialize or tear down tunnel QPs for the master */
2176 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2177 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2178 	return;
2179 }
2180 
2181 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2182 {
2183 	int i = 0;
2184 	int err;
2185 
2186 	if (!mlx4_is_mfunc(dev->dev))
2187 		return 0;
2188 
2189 	dev->sriov.is_going_down = 0;
2190 	spin_lock_init(&dev->sriov.going_down_lock);
2191 	mlx4_ib_cm_paravirt_init(dev);
2192 
2193 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2194 
2195 	if (mlx4_is_slave(dev->dev)) {
2196 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2197 		return 0;
2198 	}
2199 
2200 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2201 		if (i == mlx4_master_func_num(dev->dev))
2202 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2203 		else
2204 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2205 	}
2206 
2207 	err = mlx4_ib_init_alias_guid_service(dev);
2208 	if (err) {
2209 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2210 		goto paravirt_err;
2211 	}
2212 	err = mlx4_ib_device_register_sysfs(dev);
2213 	if (err) {
2214 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2215 		goto sysfs_err;
2216 	}
2217 
2218 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2219 		     dev->dev->caps.sqp_demux);
2220 	for (i = 0; i < dev->num_ports; i++) {
2221 		union ib_gid gid;
2222 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2223 		if (err)
2224 			goto demux_err;
2225 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2226 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2227 			     be64_to_cpu(gid.global.subnet_prefix));
2228 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2229 				      &dev->sriov.sqps[i]);
2230 		if (err)
2231 			goto demux_err;
2232 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2233 		if (err)
2234 			goto free_pv;
2235 	}
2236 	mlx4_ib_master_tunnels(dev, 1);
2237 	return 0;
2238 
2239 free_pv:
2240 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2241 demux_err:
2242 	while (--i >= 0) {
2243 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2244 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2245 	}
2246 	mlx4_ib_device_unregister_sysfs(dev);
2247 
2248 sysfs_err:
2249 	mlx4_ib_destroy_alias_guid_service(dev);
2250 
2251 paravirt_err:
2252 	mlx4_ib_cm_paravirt_clean(dev, -1);
2253 
2254 	return err;
2255 }
2256 
2257 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2258 {
2259 	int i;
2260 	unsigned long flags;
2261 
2262 	if (!mlx4_is_mfunc(dev->dev))
2263 		return;
2264 
2265 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2266 	dev->sriov.is_going_down = 1;
2267 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2268 	if (mlx4_is_master(dev->dev)) {
2269 		for (i = 0; i < dev->num_ports; i++) {
2270 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2271 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2272 			kfree(dev->sriov.sqps[i]);
2273 			dev->sriov.sqps[i] = NULL;
2274 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2275 		}
2276 
2277 		mlx4_ib_cm_paravirt_clean(dev, -1);
2278 		mlx4_ib_destroy_alias_guid_service(dev);
2279 		mlx4_ib_device_unregister_sysfs(dev);
2280 	}
2281 }
2282