xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision 179dd8c0348af75b02c7d72eaaf1cb179f1721ef)
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 "mlx4_ib.h"
44 
45 enum {
46 	MLX4_IB_VENDOR_CLASS1 = 0x9,
47 	MLX4_IB_VENDOR_CLASS2 = 0xa
48 };
49 
50 #define MLX4_TUN_SEND_WRID_SHIFT 34
51 #define MLX4_TUN_QPN_SHIFT 32
52 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
53 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
54 
55 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
56 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
57 
58  /* Port mgmt change event handling */
59 
60 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
61 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
62 #define NUM_IDX_IN_PKEY_TBL_BLK 32
63 #define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
64 #define GUID_TBL_BLK_NUM_ENTRIES 8
65 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
66 
67 struct mlx4_mad_rcv_buf {
68 	struct ib_grh grh;
69 	u8 payload[256];
70 } __packed;
71 
72 struct mlx4_mad_snd_buf {
73 	u8 payload[256];
74 } __packed;
75 
76 struct mlx4_tunnel_mad {
77 	struct ib_grh grh;
78 	struct mlx4_ib_tunnel_header hdr;
79 	struct ib_mad mad;
80 } __packed;
81 
82 struct mlx4_rcv_tunnel_mad {
83 	struct mlx4_rcv_tunnel_hdr hdr;
84 	struct ib_grh grh;
85 	struct ib_mad mad;
86 } __packed;
87 
88 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
89 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
90 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
91 				int block, u32 change_bitmap);
92 
93 __be64 mlx4_ib_gen_node_guid(void)
94 {
95 #define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
96 	return cpu_to_be64(NODE_GUID_HI | prandom_u32());
97 }
98 
99 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
100 {
101 	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
102 		cpu_to_be64(0xff00000000000000LL);
103 }
104 
105 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
106 		 int port, const struct ib_wc *in_wc,
107 		 const struct ib_grh *in_grh,
108 		 const void *in_mad, void *response_mad)
109 {
110 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
111 	void *inbox;
112 	int err;
113 	u32 in_modifier = port;
114 	u8 op_modifier = 0;
115 
116 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
117 	if (IS_ERR(inmailbox))
118 		return PTR_ERR(inmailbox);
119 	inbox = inmailbox->buf;
120 
121 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
122 	if (IS_ERR(outmailbox)) {
123 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
124 		return PTR_ERR(outmailbox);
125 	}
126 
127 	memcpy(inbox, in_mad, 256);
128 
129 	/*
130 	 * Key check traps can't be generated unless we have in_wc to
131 	 * tell us where to send the trap.
132 	 */
133 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
134 		op_modifier |= 0x1;
135 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
136 		op_modifier |= 0x2;
137 	if (mlx4_is_mfunc(dev->dev) &&
138 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
139 		op_modifier |= 0x8;
140 
141 	if (in_wc) {
142 		struct {
143 			__be32		my_qpn;
144 			u32		reserved1;
145 			__be32		rqpn;
146 			u8		sl;
147 			u8		g_path;
148 			u16		reserved2[2];
149 			__be16		pkey;
150 			u32		reserved3[11];
151 			u8		grh[40];
152 		} *ext_info;
153 
154 		memset(inbox + 256, 0, 256);
155 		ext_info = inbox + 256;
156 
157 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
158 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
159 		ext_info->sl     = in_wc->sl << 4;
160 		ext_info->g_path = in_wc->dlid_path_bits |
161 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
162 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
163 
164 		if (in_grh)
165 			memcpy(ext_info->grh, in_grh, 40);
166 
167 		op_modifier |= 0x4;
168 
169 		in_modifier |= in_wc->slid << 16;
170 	}
171 
172 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
173 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
174 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
175 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
176 
177 	if (!err)
178 		memcpy(response_mad, outmailbox->buf, 256);
179 
180 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
181 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
182 
183 	return err;
184 }
185 
186 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
187 {
188 	struct ib_ah *new_ah;
189 	struct ib_ah_attr ah_attr;
190 	unsigned long flags;
191 
192 	if (!dev->send_agent[port_num - 1][0])
193 		return;
194 
195 	memset(&ah_attr, 0, sizeof ah_attr);
196 	ah_attr.dlid     = lid;
197 	ah_attr.sl       = sl;
198 	ah_attr.port_num = port_num;
199 
200 	new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
201 			      &ah_attr);
202 	if (IS_ERR(new_ah))
203 		return;
204 
205 	spin_lock_irqsave(&dev->sm_lock, flags);
206 	if (dev->sm_ah[port_num - 1])
207 		ib_destroy_ah(dev->sm_ah[port_num - 1]);
208 	dev->sm_ah[port_num - 1] = new_ah;
209 	spin_unlock_irqrestore(&dev->sm_lock, flags);
210 }
211 
212 /*
213  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
214  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
215  */
216 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
217 		      u16 prev_lid)
218 {
219 	struct ib_port_info *pinfo;
220 	u16 lid;
221 	__be16 *base;
222 	u32 bn, pkey_change_bitmap;
223 	int i;
224 
225 
226 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
227 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
228 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
229 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
230 		switch (mad->mad_hdr.attr_id) {
231 		case IB_SMP_ATTR_PORT_INFO:
232 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
233 			lid = be16_to_cpu(pinfo->lid);
234 
235 			update_sm_ah(dev, port_num,
236 				     be16_to_cpu(pinfo->sm_lid),
237 				     pinfo->neighbormtu_mastersmsl & 0xf);
238 
239 			if (pinfo->clientrereg_resv_subnetto & 0x80)
240 				handle_client_rereg_event(dev, port_num);
241 
242 			if (prev_lid != lid)
243 				handle_lid_change_event(dev, port_num);
244 			break;
245 
246 		case IB_SMP_ATTR_PKEY_TABLE:
247 			if (!mlx4_is_mfunc(dev->dev)) {
248 				mlx4_ib_dispatch_event(dev, port_num,
249 						       IB_EVENT_PKEY_CHANGE);
250 				break;
251 			}
252 
253 			/* at this point, we are running in the master.
254 			 * Slaves do not receive SMPs.
255 			 */
256 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
257 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
258 			pkey_change_bitmap = 0;
259 			for (i = 0; i < 32; i++) {
260 				pr_debug("PKEY[%d] = x%x\n",
261 					 i + bn*32, be16_to_cpu(base[i]));
262 				if (be16_to_cpu(base[i]) !=
263 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
264 					pkey_change_bitmap |= (1 << i);
265 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
266 						be16_to_cpu(base[i]);
267 				}
268 			}
269 			pr_debug("PKEY Change event: port=%d, "
270 				 "block=0x%x, change_bitmap=0x%x\n",
271 				 port_num, bn, pkey_change_bitmap);
272 
273 			if (pkey_change_bitmap) {
274 				mlx4_ib_dispatch_event(dev, port_num,
275 						       IB_EVENT_PKEY_CHANGE);
276 				if (!dev->sriov.is_going_down)
277 					__propagate_pkey_ev(dev, port_num, bn,
278 							    pkey_change_bitmap);
279 			}
280 			break;
281 
282 		case IB_SMP_ATTR_GUID_INFO:
283 			/* paravirtualized master's guid is guid 0 -- does not change */
284 			if (!mlx4_is_master(dev->dev))
285 				mlx4_ib_dispatch_event(dev, port_num,
286 						       IB_EVENT_GID_CHANGE);
287 			/*if master, notify relevant slaves*/
288 			if (mlx4_is_master(dev->dev) &&
289 			    !dev->sriov.is_going_down) {
290 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
291 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
292 								    (u8 *)(&((struct ib_smp *)mad)->data));
293 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
294 								     (u8 *)(&((struct ib_smp *)mad)->data));
295 			}
296 			break;
297 
298 		default:
299 			break;
300 		}
301 }
302 
303 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
304 				int block, u32 change_bitmap)
305 {
306 	int i, ix, slave, err;
307 	int have_event = 0;
308 
309 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
310 		if (slave == mlx4_master_func_num(dev->dev))
311 			continue;
312 		if (!mlx4_is_slave_active(dev->dev, slave))
313 			continue;
314 
315 		have_event = 0;
316 		for (i = 0; i < 32; i++) {
317 			if (!(change_bitmap & (1 << i)))
318 				continue;
319 			for (ix = 0;
320 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
321 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
322 				    [ix] == i + 32 * block) {
323 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
324 					pr_debug("propagate_pkey_ev: slave %d,"
325 						 " port %d, ix %d (%d)\n",
326 						 slave, port_num, ix, err);
327 					have_event = 1;
328 					break;
329 				}
330 			}
331 			if (have_event)
332 				break;
333 		}
334 	}
335 }
336 
337 static void node_desc_override(struct ib_device *dev,
338 			       struct ib_mad *mad)
339 {
340 	unsigned long flags;
341 
342 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
343 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
344 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
345 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
346 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
347 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
348 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
349 	}
350 }
351 
352 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
353 {
354 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
355 	struct ib_mad_send_buf *send_buf;
356 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
357 	int ret;
358 	unsigned long flags;
359 
360 	if (agent) {
361 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
362 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
363 					      IB_MGMT_BASE_VERSION);
364 		if (IS_ERR(send_buf))
365 			return;
366 		/*
367 		 * We rely here on the fact that MLX QPs don't use the
368 		 * address handle after the send is posted (this is
369 		 * wrong following the IB spec strictly, but we know
370 		 * it's OK for our devices).
371 		 */
372 		spin_lock_irqsave(&dev->sm_lock, flags);
373 		memcpy(send_buf->mad, mad, sizeof *mad);
374 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
375 			ret = ib_post_send_mad(send_buf, NULL);
376 		else
377 			ret = -EINVAL;
378 		spin_unlock_irqrestore(&dev->sm_lock, flags);
379 
380 		if (ret)
381 			ib_free_send_mad(send_buf);
382 	}
383 }
384 
385 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
386 							     struct ib_sa_mad *sa_mad)
387 {
388 	int ret = 0;
389 
390 	/* dispatch to different sa handlers */
391 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
392 	case IB_SA_ATTR_MC_MEMBER_REC:
393 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
394 		break;
395 	default:
396 		break;
397 	}
398 	return ret;
399 }
400 
401 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
402 {
403 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
404 	int i;
405 
406 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
407 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
408 			return i;
409 	}
410 	return -1;
411 }
412 
413 
414 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
415 				   u8 port, u16 pkey, u16 *ix)
416 {
417 	int i, ret;
418 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
419 	u16 slot_pkey;
420 
421 	if (slave == mlx4_master_func_num(dev->dev))
422 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
423 
424 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
425 
426 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
427 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
428 			continue;
429 
430 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
431 
432 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
433 		if (ret)
434 			continue;
435 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
436 			if (slot_pkey & 0x8000) {
437 				*ix = (u16) pkey_ix;
438 				return 0;
439 			} else {
440 				/* take first partial pkey index found */
441 				if (partial_ix == 0xFF)
442 					partial_ix = pkey_ix;
443 			}
444 		}
445 	}
446 
447 	if (partial_ix < 0xFF) {
448 		*ix = (u16) partial_ix;
449 		return 0;
450 	}
451 
452 	return -EINVAL;
453 }
454 
455 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
456 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
457 			  struct ib_grh *grh, struct ib_mad *mad)
458 {
459 	struct ib_sge list;
460 	struct ib_send_wr wr, *bad_wr;
461 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
462 	struct mlx4_ib_demux_pv_qp *tun_qp;
463 	struct mlx4_rcv_tunnel_mad *tun_mad;
464 	struct ib_ah_attr attr;
465 	struct ib_ah *ah;
466 	struct ib_qp *src_qp = NULL;
467 	unsigned tun_tx_ix = 0;
468 	int dqpn;
469 	int ret = 0;
470 	u16 tun_pkey_ix;
471 	u16 cached_pkey;
472 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
473 
474 	if (dest_qpt > IB_QPT_GSI)
475 		return -EINVAL;
476 
477 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
478 
479 	/* check if proxy qp created */
480 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
481 		return -EAGAIN;
482 
483 	if (!dest_qpt)
484 		tun_qp = &tun_ctx->qp[0];
485 	else
486 		tun_qp = &tun_ctx->qp[1];
487 
488 	/* compute P_Key index to put in tunnel header for slave */
489 	if (dest_qpt) {
490 		u16 pkey_ix;
491 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
492 		if (ret)
493 			return -EINVAL;
494 
495 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
496 		if (ret)
497 			return -EINVAL;
498 		tun_pkey_ix = pkey_ix;
499 	} else
500 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
501 
502 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
503 
504 	/* get tunnel tx data buf for slave */
505 	src_qp = tun_qp->qp;
506 
507 	/* create ah. Just need an empty one with the port num for the post send.
508 	 * The driver will set the force loopback bit in post_send */
509 	memset(&attr, 0, sizeof attr);
510 	attr.port_num = port;
511 	if (is_eth) {
512 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
513 		attr.ah_flags = IB_AH_GRH;
514 	}
515 	ah = ib_create_ah(tun_ctx->pd, &attr);
516 	if (IS_ERR(ah))
517 		return -ENOMEM;
518 
519 	/* allocate tunnel tx buf after pass failure returns */
520 	spin_lock(&tun_qp->tx_lock);
521 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
522 	    (MLX4_NUM_TUNNEL_BUFS - 1))
523 		ret = -EAGAIN;
524 	else
525 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
526 	spin_unlock(&tun_qp->tx_lock);
527 	if (ret)
528 		goto out;
529 
530 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
531 	if (tun_qp->tx_ring[tun_tx_ix].ah)
532 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
533 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
534 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
535 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
536 				   sizeof (struct mlx4_rcv_tunnel_mad),
537 				   DMA_TO_DEVICE);
538 
539 	/* copy over to tunnel buffer */
540 	if (grh)
541 		memcpy(&tun_mad->grh, grh, sizeof *grh);
542 	memcpy(&tun_mad->mad, mad, sizeof *mad);
543 
544 	/* adjust tunnel data */
545 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
546 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
547 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
548 
549 	if (is_eth) {
550 		u16 vlan = 0;
551 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
552 						NULL)) {
553 			/* VST mode */
554 			if (vlan != wc->vlan_id)
555 				/* Packet vlan is not the VST-assigned vlan.
556 				 * Drop the packet.
557 				 */
558 				goto out;
559 			 else
560 				/* Remove the vlan tag before forwarding
561 				 * the packet to the VF.
562 				 */
563 				vlan = 0xffff;
564 		} else {
565 			vlan = wc->vlan_id;
566 		}
567 
568 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
569 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
570 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
571 	} else {
572 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
573 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
574 	}
575 
576 	ib_dma_sync_single_for_device(&dev->ib_dev,
577 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
578 				      sizeof (struct mlx4_rcv_tunnel_mad),
579 				      DMA_TO_DEVICE);
580 
581 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
582 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
583 	list.lkey = tun_ctx->mr->lkey;
584 
585 	wr.wr.ud.ah = ah;
586 	wr.wr.ud.port_num = port;
587 	wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
588 	wr.wr.ud.remote_qpn = dqpn;
589 	wr.next = NULL;
590 	wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
591 	wr.sg_list = &list;
592 	wr.num_sge = 1;
593 	wr.opcode = IB_WR_SEND;
594 	wr.send_flags = IB_SEND_SIGNALED;
595 
596 	ret = ib_post_send(src_qp, &wr, &bad_wr);
597 out:
598 	if (ret)
599 		ib_destroy_ah(ah);
600 	return ret;
601 }
602 
603 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
604 			struct ib_wc *wc, struct ib_grh *grh,
605 			struct ib_mad *mad)
606 {
607 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
608 	int err;
609 	int slave;
610 	u8 *slave_id;
611 	int is_eth = 0;
612 
613 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
614 		is_eth = 0;
615 	else
616 		is_eth = 1;
617 
618 	if (is_eth) {
619 		if (!(wc->wc_flags & IB_WC_GRH)) {
620 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
621 			return -EINVAL;
622 		}
623 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
624 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
625 			return -EINVAL;
626 		}
627 		if (mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave)) {
628 			mlx4_ib_warn(ibdev, "failed matching grh\n");
629 			return -ENOENT;
630 		}
631 		if (slave >= dev->dev->caps.sqp_demux) {
632 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
633 				     slave, dev->dev->caps.sqp_demux);
634 			return -ENOENT;
635 		}
636 
637 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
638 			return 0;
639 
640 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
641 		if (err)
642 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
643 				 slave, err);
644 		return 0;
645 	}
646 
647 	/* Initially assume that this mad is for us */
648 	slave = mlx4_master_func_num(dev->dev);
649 
650 	/* See if the slave id is encoded in a response mad */
651 	if (mad->mad_hdr.method & 0x80) {
652 		slave_id = (u8 *) &mad->mad_hdr.tid;
653 		slave = *slave_id;
654 		if (slave != 255) /*255 indicates the dom0*/
655 			*slave_id = 0; /* remap tid */
656 	}
657 
658 	/* If a grh is present, we demux according to it */
659 	if (wc->wc_flags & IB_WC_GRH) {
660 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
661 		if (slave < 0) {
662 			mlx4_ib_warn(ibdev, "failed matching grh\n");
663 			return -ENOENT;
664 		}
665 	}
666 	/* Class-specific handling */
667 	switch (mad->mad_hdr.mgmt_class) {
668 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
669 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
670 		/* 255 indicates the dom0 */
671 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
672 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
673 				return -EPERM;
674 			/* for a VF. drop unsolicited MADs */
675 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
676 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
677 					     slave, mad->mad_hdr.mgmt_class,
678 					     mad->mad_hdr.method);
679 				return -EINVAL;
680 			}
681 		}
682 		break;
683 	case IB_MGMT_CLASS_SUBN_ADM:
684 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
685 					     (struct ib_sa_mad *) mad))
686 			return 0;
687 		break;
688 	case IB_MGMT_CLASS_CM:
689 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
690 			return 0;
691 		break;
692 	case IB_MGMT_CLASS_DEVICE_MGMT:
693 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
694 			return 0;
695 		break;
696 	default:
697 		/* Drop unsupported classes for slaves in tunnel mode */
698 		if (slave != mlx4_master_func_num(dev->dev)) {
699 			pr_debug("dropping unsupported ingress mad from class:%d "
700 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
701 			return 0;
702 		}
703 	}
704 	/*make sure that no slave==255 was not handled yet.*/
705 	if (slave >= dev->dev->caps.sqp_demux) {
706 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
707 			     slave, dev->dev->caps.sqp_demux);
708 		return -ENOENT;
709 	}
710 
711 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
712 	if (err)
713 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
714 			 slave, err);
715 	return 0;
716 }
717 
718 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
719 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
720 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
721 {
722 	u16 slid, prev_lid = 0;
723 	int err;
724 	struct ib_port_attr pattr;
725 
726 	if (in_wc && in_wc->qp->qp_num) {
727 		pr_debug("received MAD: slid:%d sqpn:%d "
728 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
729 			in_wc->slid, in_wc->src_qp,
730 			in_wc->dlid_path_bits,
731 			in_wc->qp->qp_num,
732 			in_wc->wc_flags,
733 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
734 			be16_to_cpu(in_mad->mad_hdr.attr_id));
735 		if (in_wc->wc_flags & IB_WC_GRH) {
736 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
737 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
738 				 be64_to_cpu(in_grh->sgid.global.interface_id));
739 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
740 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
741 				 be64_to_cpu(in_grh->dgid.global.interface_id));
742 		}
743 	}
744 
745 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
746 
747 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
748 		forward_trap(to_mdev(ibdev), port_num, in_mad);
749 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
750 	}
751 
752 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
753 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
754 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
755 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
756 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
757 			return IB_MAD_RESULT_SUCCESS;
758 
759 		/*
760 		 * Don't process SMInfo queries -- the SMA can't handle them.
761 		 */
762 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
763 			return IB_MAD_RESULT_SUCCESS;
764 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
765 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
766 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
767 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
768 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
769 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
770 			return IB_MAD_RESULT_SUCCESS;
771 	} else
772 		return IB_MAD_RESULT_SUCCESS;
773 
774 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
775 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
776 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
777 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
778 	    !ib_query_port(ibdev, port_num, &pattr))
779 		prev_lid = pattr.lid;
780 
781 	err = mlx4_MAD_IFC(to_mdev(ibdev),
782 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
783 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
784 			   MLX4_MAD_IFC_NET_VIEW,
785 			   port_num, in_wc, in_grh, in_mad, out_mad);
786 	if (err)
787 		return IB_MAD_RESULT_FAILURE;
788 
789 	if (!out_mad->mad_hdr.status) {
790 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
791 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
792 		/* slaves get node desc from FW */
793 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
794 			node_desc_override(ibdev, out_mad);
795 	}
796 
797 	/* set return bit in status of directed route responses */
798 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
799 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
800 
801 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
802 		/* no response for trap repress */
803 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
804 
805 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
806 }
807 
808 static void edit_counter(struct mlx4_counter *cnt,
809 					struct ib_pma_portcounters *pma_cnt)
810 {
811 	ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
812 			     (be64_to_cpu(cnt->tx_bytes) >> 2));
813 	ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
814 			     (be64_to_cpu(cnt->rx_bytes) >> 2));
815 	ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
816 			     be64_to_cpu(cnt->tx_frames));
817 	ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
818 			     be64_to_cpu(cnt->rx_frames));
819 }
820 
821 static int iboe_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 	struct mlx4_counter counter_stats;
826 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
827 	int err;
828 
829 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
830 		return -EINVAL;
831 
832 	memset(&counter_stats, 0, sizeof(counter_stats));
833 	err = mlx4_get_counter_stats(dev->dev,
834 				     dev->counters[port_num - 1].index,
835 				     &counter_stats, 0);
836 	if (err)
837 		err = IB_MAD_RESULT_FAILURE;
838 	else {
839 		memset(out_mad->data, 0, sizeof out_mad->data);
840 		switch (counter_stats.counter_mode & 0xf) {
841 		case 0:
842 			edit_counter(&counter_stats,
843 				     (void *)(out_mad->data + 40));
844 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
845 			break;
846 		default:
847 			err = IB_MAD_RESULT_FAILURE;
848 		}
849 	}
850 
851 	return err;
852 }
853 
854 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
855 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
856 			const struct ib_mad_hdr *in, size_t in_mad_size,
857 			struct ib_mad_hdr *out, size_t *out_mad_size,
858 			u16 *out_mad_pkey_index)
859 {
860 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
861 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
862 	struct ib_mad *out_mad = (struct ib_mad *)out;
863 
864 	BUG_ON(in_mad_size != sizeof(*in_mad) ||
865 	       *out_mad_size != sizeof(*out_mad));
866 
867 	switch (rdma_port_get_link_layer(ibdev, port_num)) {
868 	case IB_LINK_LAYER_INFINIBAND:
869 		if (!mlx4_is_slave(dev->dev))
870 			return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
871 					      in_grh, in_mad, out_mad);
872 	case IB_LINK_LAYER_ETHERNET:
873 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
874 					  in_grh, in_mad, out_mad);
875 	default:
876 		return -EINVAL;
877 	}
878 }
879 
880 static void send_handler(struct ib_mad_agent *agent,
881 			 struct ib_mad_send_wc *mad_send_wc)
882 {
883 	if (mad_send_wc->send_buf->context[0])
884 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
885 	ib_free_send_mad(mad_send_wc->send_buf);
886 }
887 
888 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
889 {
890 	struct ib_mad_agent *agent;
891 	int p, q;
892 	int ret;
893 	enum rdma_link_layer ll;
894 
895 	for (p = 0; p < dev->num_ports; ++p) {
896 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
897 		for (q = 0; q <= 1; ++q) {
898 			if (ll == IB_LINK_LAYER_INFINIBAND) {
899 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
900 							      q ? IB_QPT_GSI : IB_QPT_SMI,
901 							      NULL, 0, send_handler,
902 							      NULL, NULL, 0);
903 				if (IS_ERR(agent)) {
904 					ret = PTR_ERR(agent);
905 					goto err;
906 				}
907 				dev->send_agent[p][q] = agent;
908 			} else
909 				dev->send_agent[p][q] = NULL;
910 		}
911 	}
912 
913 	return 0;
914 
915 err:
916 	for (p = 0; p < dev->num_ports; ++p)
917 		for (q = 0; q <= 1; ++q)
918 			if (dev->send_agent[p][q])
919 				ib_unregister_mad_agent(dev->send_agent[p][q]);
920 
921 	return ret;
922 }
923 
924 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
925 {
926 	struct ib_mad_agent *agent;
927 	int p, q;
928 
929 	for (p = 0; p < dev->num_ports; ++p) {
930 		for (q = 0; q <= 1; ++q) {
931 			agent = dev->send_agent[p][q];
932 			if (agent) {
933 				dev->send_agent[p][q] = NULL;
934 				ib_unregister_mad_agent(agent);
935 			}
936 		}
937 
938 		if (dev->sm_ah[p])
939 			ib_destroy_ah(dev->sm_ah[p]);
940 	}
941 }
942 
943 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
944 {
945 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
946 
947 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
948 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
949 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
950 }
951 
952 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
953 {
954 	/* re-configure the alias-guid and mcg's */
955 	if (mlx4_is_master(dev->dev)) {
956 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
957 
958 		if (!dev->sriov.is_going_down) {
959 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
960 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
961 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
962 		}
963 	}
964 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
965 }
966 
967 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
968 			      struct mlx4_eqe *eqe)
969 {
970 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
971 			    GET_MASK_FROM_EQE(eqe));
972 }
973 
974 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
975 				      u32 guid_tbl_blk_num, u32 change_bitmap)
976 {
977 	struct ib_smp *in_mad  = NULL;
978 	struct ib_smp *out_mad  = NULL;
979 	u16 i;
980 
981 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
982 		return;
983 
984 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
985 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
986 	if (!in_mad || !out_mad) {
987 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
988 		goto out;
989 	}
990 
991 	guid_tbl_blk_num  *= 4;
992 
993 	for (i = 0; i < 4; i++) {
994 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
995 			continue;
996 		memset(in_mad, 0, sizeof *in_mad);
997 		memset(out_mad, 0, sizeof *out_mad);
998 
999 		in_mad->base_version  = 1;
1000 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1001 		in_mad->class_version = 1;
1002 		in_mad->method        = IB_MGMT_METHOD_GET;
1003 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1004 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1005 
1006 		if (mlx4_MAD_IFC(dev,
1007 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1008 				 port_num, NULL, NULL, in_mad, out_mad)) {
1009 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1010 			goto out;
1011 		}
1012 
1013 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1014 						    port_num,
1015 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1016 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1017 						     port_num,
1018 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1019 	}
1020 
1021 out:
1022 	kfree(in_mad);
1023 	kfree(out_mad);
1024 	return;
1025 }
1026 
1027 void handle_port_mgmt_change_event(struct work_struct *work)
1028 {
1029 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1030 	struct mlx4_ib_dev *dev = ew->ib_dev;
1031 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1032 	u8 port = eqe->event.port_mgmt_change.port;
1033 	u32 changed_attr;
1034 	u32 tbl_block;
1035 	u32 change_bitmap;
1036 
1037 	switch (eqe->subtype) {
1038 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1039 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1040 
1041 		/* Update the SM ah - This should be done before handling
1042 		   the other changed attributes so that MADs can be sent to the SM */
1043 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1044 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1045 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1046 			update_sm_ah(dev, port, lid, sl);
1047 		}
1048 
1049 		/* Check if it is a lid change event */
1050 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1051 			handle_lid_change_event(dev, port);
1052 
1053 		/* Generate GUID changed event */
1054 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1055 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1056 			/*if master, notify all slaves*/
1057 			if (mlx4_is_master(dev->dev))
1058 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1059 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1060 		}
1061 
1062 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1063 			handle_client_rereg_event(dev, port);
1064 		break;
1065 
1066 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1067 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1068 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1069 			propagate_pkey_ev(dev, port, eqe);
1070 		break;
1071 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1072 		/* paravirtualized master's guid is guid 0 -- does not change */
1073 		if (!mlx4_is_master(dev->dev))
1074 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1075 		/*if master, notify relevant slaves*/
1076 		else if (!dev->sriov.is_going_down) {
1077 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1078 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1079 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1080 		}
1081 		break;
1082 	default:
1083 		pr_warn("Unsupported subtype 0x%x for "
1084 			"Port Management Change event\n", eqe->subtype);
1085 	}
1086 
1087 	kfree(ew);
1088 }
1089 
1090 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1091 			    enum ib_event_type type)
1092 {
1093 	struct ib_event event;
1094 
1095 	event.device		= &dev->ib_dev;
1096 	event.element.port_num	= port_num;
1097 	event.event		= type;
1098 
1099 	ib_dispatch_event(&event);
1100 }
1101 
1102 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1103 {
1104 	unsigned long flags;
1105 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1106 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1107 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1108 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1109 		queue_work(ctx->wq, &ctx->work);
1110 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1111 }
1112 
1113 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1114 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1115 				  int index)
1116 {
1117 	struct ib_sge sg_list;
1118 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1119 	int size;
1120 
1121 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1122 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1123 
1124 	sg_list.addr = tun_qp->ring[index].map;
1125 	sg_list.length = size;
1126 	sg_list.lkey = ctx->mr->lkey;
1127 
1128 	recv_wr.next = NULL;
1129 	recv_wr.sg_list = &sg_list;
1130 	recv_wr.num_sge = 1;
1131 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1132 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1133 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1134 				      size, DMA_FROM_DEVICE);
1135 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1136 }
1137 
1138 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1139 		int slave, struct ib_sa_mad *sa_mad)
1140 {
1141 	int ret = 0;
1142 
1143 	/* dispatch to different sa handlers */
1144 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1145 	case IB_SA_ATTR_MC_MEMBER_REC:
1146 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1147 		break;
1148 	default:
1149 		break;
1150 	}
1151 	return ret;
1152 }
1153 
1154 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1155 {
1156 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1157 
1158 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1159 }
1160 
1161 
1162 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1163 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1164 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1165 			 u8 *s_mac, struct ib_mad *mad)
1166 {
1167 	struct ib_sge list;
1168 	struct ib_send_wr wr, *bad_wr;
1169 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1170 	struct mlx4_ib_demux_pv_qp *sqp;
1171 	struct mlx4_mad_snd_buf *sqp_mad;
1172 	struct ib_ah *ah;
1173 	struct ib_qp *send_qp = NULL;
1174 	unsigned wire_tx_ix = 0;
1175 	int ret = 0;
1176 	u16 wire_pkey_ix;
1177 	int src_qpnum;
1178 	u8 sgid_index;
1179 
1180 
1181 	sqp_ctx = dev->sriov.sqps[port-1];
1182 
1183 	/* check if proxy qp created */
1184 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1185 		return -EAGAIN;
1186 
1187 	if (dest_qpt == IB_QPT_SMI) {
1188 		src_qpnum = 0;
1189 		sqp = &sqp_ctx->qp[0];
1190 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1191 	} else {
1192 		src_qpnum = 1;
1193 		sqp = &sqp_ctx->qp[1];
1194 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1195 	}
1196 
1197 	send_qp = sqp->qp;
1198 
1199 	/* create ah */
1200 	sgid_index = attr->grh.sgid_index;
1201 	attr->grh.sgid_index = 0;
1202 	ah = ib_create_ah(sqp_ctx->pd, attr);
1203 	if (IS_ERR(ah))
1204 		return -ENOMEM;
1205 	attr->grh.sgid_index = sgid_index;
1206 	to_mah(ah)->av.ib.gid_index = sgid_index;
1207 	/* get rid of force-loopback bit */
1208 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1209 	spin_lock(&sqp->tx_lock);
1210 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1211 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1212 		ret = -EAGAIN;
1213 	else
1214 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1215 	spin_unlock(&sqp->tx_lock);
1216 	if (ret)
1217 		goto out;
1218 
1219 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1220 	if (sqp->tx_ring[wire_tx_ix].ah)
1221 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1222 	sqp->tx_ring[wire_tx_ix].ah = ah;
1223 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1224 				   sqp->tx_ring[wire_tx_ix].buf.map,
1225 				   sizeof (struct mlx4_mad_snd_buf),
1226 				   DMA_TO_DEVICE);
1227 
1228 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1229 
1230 	ib_dma_sync_single_for_device(&dev->ib_dev,
1231 				      sqp->tx_ring[wire_tx_ix].buf.map,
1232 				      sizeof (struct mlx4_mad_snd_buf),
1233 				      DMA_TO_DEVICE);
1234 
1235 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1236 	list.length = sizeof (struct mlx4_mad_snd_buf);
1237 	list.lkey = sqp_ctx->mr->lkey;
1238 
1239 	wr.wr.ud.ah = ah;
1240 	wr.wr.ud.port_num = port;
1241 	wr.wr.ud.pkey_index = wire_pkey_ix;
1242 	wr.wr.ud.remote_qkey = qkey;
1243 	wr.wr.ud.remote_qpn = remote_qpn;
1244 	wr.next = NULL;
1245 	wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1246 	wr.sg_list = &list;
1247 	wr.num_sge = 1;
1248 	wr.opcode = IB_WR_SEND;
1249 	wr.send_flags = IB_SEND_SIGNALED;
1250 	if (s_mac)
1251 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1252 
1253 
1254 	ret = ib_post_send(send_qp, &wr, &bad_wr);
1255 out:
1256 	if (ret)
1257 		ib_destroy_ah(ah);
1258 	return ret;
1259 }
1260 
1261 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1262 {
1263 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1264 		return slave;
1265 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1266 }
1267 
1268 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1269 				    struct ib_ah_attr *ah_attr)
1270 {
1271 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1272 		ah_attr->grh.sgid_index = slave;
1273 	else
1274 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1275 }
1276 
1277 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1278 {
1279 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1280 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1281 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1282 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1283 	struct mlx4_ib_ah ah;
1284 	struct ib_ah_attr ah_attr;
1285 	u8 *slave_id;
1286 	int slave;
1287 	int port;
1288 
1289 	/* Get slave that sent this packet */
1290 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1291 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1292 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1293 	    wc->src_qp & 0x4) {
1294 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1295 		return;
1296 	}
1297 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1298 	if (slave != ctx->slave) {
1299 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1300 			     "belongs to another slave\n", wc->src_qp);
1301 		return;
1302 	}
1303 
1304 	/* Map transaction ID */
1305 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1306 				   sizeof (struct mlx4_tunnel_mad),
1307 				   DMA_FROM_DEVICE);
1308 	switch (tunnel->mad.mad_hdr.method) {
1309 	case IB_MGMT_METHOD_SET:
1310 	case IB_MGMT_METHOD_GET:
1311 	case IB_MGMT_METHOD_REPORT:
1312 	case IB_SA_METHOD_GET_TABLE:
1313 	case IB_SA_METHOD_DELETE:
1314 	case IB_SA_METHOD_GET_MULTI:
1315 	case IB_SA_METHOD_GET_TRACE_TBL:
1316 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1317 		if (*slave_id) {
1318 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1319 				     "class:%d slave:%d\n", *slave_id,
1320 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1321 			return;
1322 		} else
1323 			*slave_id = slave;
1324 	default:
1325 		/* nothing */;
1326 	}
1327 
1328 	/* Class-specific handling */
1329 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1330 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1331 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1332 		if (slave != mlx4_master_func_num(dev->dev) &&
1333 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1334 			return;
1335 		break;
1336 	case IB_MGMT_CLASS_SUBN_ADM:
1337 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1338 			      (struct ib_sa_mad *) &tunnel->mad))
1339 			return;
1340 		break;
1341 	case IB_MGMT_CLASS_CM:
1342 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1343 			      (struct ib_mad *) &tunnel->mad))
1344 			return;
1345 		break;
1346 	case IB_MGMT_CLASS_DEVICE_MGMT:
1347 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1348 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1349 			return;
1350 		break;
1351 	default:
1352 		/* Drop unsupported classes for slaves in tunnel mode */
1353 		if (slave != mlx4_master_func_num(dev->dev)) {
1354 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1355 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1356 			return;
1357 		}
1358 	}
1359 
1360 	/* We are using standard ib_core services to send the mad, so generate a
1361 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1362 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1363 	ah.ibah.device = ctx->ib_dev;
1364 
1365 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1366 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1367 	if (port < 0)
1368 		return;
1369 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1370 
1371 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1372 	if (ah_attr.ah_flags & IB_AH_GRH)
1373 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1374 
1375 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1376 	ah_attr.vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1377 	/* if slave have default vlan use it */
1378 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1379 				    &ah_attr.vlan_id, &ah_attr.sl);
1380 
1381 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1382 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1383 			     IB_QPT_SMI : IB_QPT_GSI,
1384 			     be16_to_cpu(tunnel->hdr.pkey_index),
1385 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1386 			     be32_to_cpu(tunnel->hdr.qkey),
1387 			     &ah_attr, wc->smac, &tunnel->mad);
1388 }
1389 
1390 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1391 				 enum ib_qp_type qp_type, int is_tun)
1392 {
1393 	int i;
1394 	struct mlx4_ib_demux_pv_qp *tun_qp;
1395 	int rx_buf_size, tx_buf_size;
1396 
1397 	if (qp_type > IB_QPT_GSI)
1398 		return -EINVAL;
1399 
1400 	tun_qp = &ctx->qp[qp_type];
1401 
1402 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1403 			       GFP_KERNEL);
1404 	if (!tun_qp->ring)
1405 		return -ENOMEM;
1406 
1407 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1408 				  sizeof (struct mlx4_ib_tun_tx_buf),
1409 				  GFP_KERNEL);
1410 	if (!tun_qp->tx_ring) {
1411 		kfree(tun_qp->ring);
1412 		tun_qp->ring = NULL;
1413 		return -ENOMEM;
1414 	}
1415 
1416 	if (is_tun) {
1417 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1418 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1419 	} else {
1420 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1421 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1422 	}
1423 
1424 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1425 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1426 		if (!tun_qp->ring[i].addr)
1427 			goto err;
1428 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1429 							tun_qp->ring[i].addr,
1430 							rx_buf_size,
1431 							DMA_FROM_DEVICE);
1432 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1433 			kfree(tun_qp->ring[i].addr);
1434 			goto err;
1435 		}
1436 	}
1437 
1438 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1439 		tun_qp->tx_ring[i].buf.addr =
1440 			kmalloc(tx_buf_size, GFP_KERNEL);
1441 		if (!tun_qp->tx_ring[i].buf.addr)
1442 			goto tx_err;
1443 		tun_qp->tx_ring[i].buf.map =
1444 			ib_dma_map_single(ctx->ib_dev,
1445 					  tun_qp->tx_ring[i].buf.addr,
1446 					  tx_buf_size,
1447 					  DMA_TO_DEVICE);
1448 		if (ib_dma_mapping_error(ctx->ib_dev,
1449 					 tun_qp->tx_ring[i].buf.map)) {
1450 			kfree(tun_qp->tx_ring[i].buf.addr);
1451 			goto tx_err;
1452 		}
1453 		tun_qp->tx_ring[i].ah = NULL;
1454 	}
1455 	spin_lock_init(&tun_qp->tx_lock);
1456 	tun_qp->tx_ix_head = 0;
1457 	tun_qp->tx_ix_tail = 0;
1458 	tun_qp->proxy_qpt = qp_type;
1459 
1460 	return 0;
1461 
1462 tx_err:
1463 	while (i > 0) {
1464 		--i;
1465 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1466 				    tx_buf_size, DMA_TO_DEVICE);
1467 		kfree(tun_qp->tx_ring[i].buf.addr);
1468 	}
1469 	kfree(tun_qp->tx_ring);
1470 	tun_qp->tx_ring = NULL;
1471 	i = MLX4_NUM_TUNNEL_BUFS;
1472 err:
1473 	while (i > 0) {
1474 		--i;
1475 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1476 				    rx_buf_size, DMA_FROM_DEVICE);
1477 		kfree(tun_qp->ring[i].addr);
1478 	}
1479 	kfree(tun_qp->ring);
1480 	tun_qp->ring = NULL;
1481 	return -ENOMEM;
1482 }
1483 
1484 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1485 				     enum ib_qp_type qp_type, int is_tun)
1486 {
1487 	int i;
1488 	struct mlx4_ib_demux_pv_qp *tun_qp;
1489 	int rx_buf_size, tx_buf_size;
1490 
1491 	if (qp_type > IB_QPT_GSI)
1492 		return;
1493 
1494 	tun_qp = &ctx->qp[qp_type];
1495 	if (is_tun) {
1496 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1497 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1498 	} else {
1499 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1500 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1501 	}
1502 
1503 
1504 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1505 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1506 				    rx_buf_size, DMA_FROM_DEVICE);
1507 		kfree(tun_qp->ring[i].addr);
1508 	}
1509 
1510 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1511 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1512 				    tx_buf_size, DMA_TO_DEVICE);
1513 		kfree(tun_qp->tx_ring[i].buf.addr);
1514 		if (tun_qp->tx_ring[i].ah)
1515 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1516 	}
1517 	kfree(tun_qp->tx_ring);
1518 	kfree(tun_qp->ring);
1519 }
1520 
1521 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1522 {
1523 	struct mlx4_ib_demux_pv_ctx *ctx;
1524 	struct mlx4_ib_demux_pv_qp *tun_qp;
1525 	struct ib_wc wc;
1526 	int ret;
1527 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1528 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1529 
1530 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1531 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1532 		if (wc.status == IB_WC_SUCCESS) {
1533 			switch (wc.opcode) {
1534 			case IB_WC_RECV:
1535 				mlx4_ib_multiplex_mad(ctx, &wc);
1536 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1537 							     wc.wr_id &
1538 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1539 				if (ret)
1540 					pr_err("Failed reposting tunnel "
1541 					       "buf:%lld\n", wc.wr_id);
1542 				break;
1543 			case IB_WC_SEND:
1544 				pr_debug("received tunnel send completion:"
1545 					 "wrid=0x%llx, status=0x%x\n",
1546 					 wc.wr_id, wc.status);
1547 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1548 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1549 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1550 					= NULL;
1551 				spin_lock(&tun_qp->tx_lock);
1552 				tun_qp->tx_ix_tail++;
1553 				spin_unlock(&tun_qp->tx_lock);
1554 
1555 				break;
1556 			default:
1557 				break;
1558 			}
1559 		} else  {
1560 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1561 				 " status = %d, wrid = 0x%llx\n",
1562 				 ctx->slave, wc.status, wc.wr_id);
1563 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1564 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1565 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1566 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1567 					= NULL;
1568 				spin_lock(&tun_qp->tx_lock);
1569 				tun_qp->tx_ix_tail++;
1570 				spin_unlock(&tun_qp->tx_lock);
1571 			}
1572 		}
1573 	}
1574 }
1575 
1576 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1577 {
1578 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1579 
1580 	/* It's worse than that! He's dead, Jim! */
1581 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1582 	       event->event, sqp->port);
1583 }
1584 
1585 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1586 			    enum ib_qp_type qp_type, int create_tun)
1587 {
1588 	int i, ret;
1589 	struct mlx4_ib_demux_pv_qp *tun_qp;
1590 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1591 	struct ib_qp_attr attr;
1592 	int qp_attr_mask_INIT;
1593 
1594 	if (qp_type > IB_QPT_GSI)
1595 		return -EINVAL;
1596 
1597 	tun_qp = &ctx->qp[qp_type];
1598 
1599 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1600 	qp_init_attr.init_attr.send_cq = ctx->cq;
1601 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1602 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1603 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1604 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1605 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1606 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1607 	if (create_tun) {
1608 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1609 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1610 		qp_init_attr.port = ctx->port;
1611 		qp_init_attr.slave = ctx->slave;
1612 		qp_init_attr.proxy_qp_type = qp_type;
1613 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1614 			   IB_QP_QKEY | IB_QP_PORT;
1615 	} else {
1616 		qp_init_attr.init_attr.qp_type = qp_type;
1617 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1618 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1619 	}
1620 	qp_init_attr.init_attr.port_num = ctx->port;
1621 	qp_init_attr.init_attr.qp_context = ctx;
1622 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1623 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1624 	if (IS_ERR(tun_qp->qp)) {
1625 		ret = PTR_ERR(tun_qp->qp);
1626 		tun_qp->qp = NULL;
1627 		pr_err("Couldn't create %s QP (%d)\n",
1628 		       create_tun ? "tunnel" : "special", ret);
1629 		return ret;
1630 	}
1631 
1632 	memset(&attr, 0, sizeof attr);
1633 	attr.qp_state = IB_QPS_INIT;
1634 	ret = 0;
1635 	if (create_tun)
1636 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1637 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1638 					      &attr.pkey_index);
1639 	if (ret || !create_tun)
1640 		attr.pkey_index =
1641 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1642 	attr.qkey = IB_QP1_QKEY;
1643 	attr.port_num = ctx->port;
1644 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1645 	if (ret) {
1646 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1647 		       create_tun ? "tunnel" : "special", ret);
1648 		goto err_qp;
1649 	}
1650 	attr.qp_state = IB_QPS_RTR;
1651 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1652 	if (ret) {
1653 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1654 		       create_tun ? "tunnel" : "special", ret);
1655 		goto err_qp;
1656 	}
1657 	attr.qp_state = IB_QPS_RTS;
1658 	attr.sq_psn = 0;
1659 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1660 	if (ret) {
1661 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1662 		       create_tun ? "tunnel" : "special", ret);
1663 		goto err_qp;
1664 	}
1665 
1666 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1667 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1668 		if (ret) {
1669 			pr_err(" mlx4_ib_post_pv_buf error"
1670 			       " (err = %d, i = %d)\n", ret, i);
1671 			goto err_qp;
1672 		}
1673 	}
1674 	return 0;
1675 
1676 err_qp:
1677 	ib_destroy_qp(tun_qp->qp);
1678 	tun_qp->qp = NULL;
1679 	return ret;
1680 }
1681 
1682 /*
1683  * IB MAD completion callback for real SQPs
1684  */
1685 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1686 {
1687 	struct mlx4_ib_demux_pv_ctx *ctx;
1688 	struct mlx4_ib_demux_pv_qp *sqp;
1689 	struct ib_wc wc;
1690 	struct ib_grh *grh;
1691 	struct ib_mad *mad;
1692 
1693 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1694 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1695 
1696 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1697 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1698 		if (wc.status == IB_WC_SUCCESS) {
1699 			switch (wc.opcode) {
1700 			case IB_WC_SEND:
1701 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1702 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1703 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1704 					= NULL;
1705 				spin_lock(&sqp->tx_lock);
1706 				sqp->tx_ix_tail++;
1707 				spin_unlock(&sqp->tx_lock);
1708 				break;
1709 			case IB_WC_RECV:
1710 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1711 						(sqp->ring[wc.wr_id &
1712 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1713 				grh = &(((struct mlx4_mad_rcv_buf *)
1714 						(sqp->ring[wc.wr_id &
1715 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1716 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1717 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1718 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1719 					pr_err("Failed reposting SQP "
1720 					       "buf:%lld\n", wc.wr_id);
1721 				break;
1722 			default:
1723 				BUG_ON(1);
1724 				break;
1725 			}
1726 		} else  {
1727 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1728 				 " status = %d, wrid = 0x%llx\n",
1729 				 ctx->slave, wc.status, wc.wr_id);
1730 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1731 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1732 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1733 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1734 					= NULL;
1735 				spin_lock(&sqp->tx_lock);
1736 				sqp->tx_ix_tail++;
1737 				spin_unlock(&sqp->tx_lock);
1738 			}
1739 		}
1740 	}
1741 }
1742 
1743 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1744 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1745 {
1746 	struct mlx4_ib_demux_pv_ctx *ctx;
1747 
1748 	*ret_ctx = NULL;
1749 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1750 	if (!ctx) {
1751 		pr_err("failed allocating pv resource context "
1752 		       "for port %d, slave %d\n", port, slave);
1753 		return -ENOMEM;
1754 	}
1755 
1756 	ctx->ib_dev = &dev->ib_dev;
1757 	ctx->port = port;
1758 	ctx->slave = slave;
1759 	*ret_ctx = ctx;
1760 	return 0;
1761 }
1762 
1763 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1764 {
1765 	if (dev->sriov.demux[port - 1].tun[slave]) {
1766 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1767 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1768 	}
1769 }
1770 
1771 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1772 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1773 {
1774 	int ret, cq_size;
1775 	struct ib_cq_init_attr cq_attr = {};
1776 
1777 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1778 		return -EEXIST;
1779 
1780 	ctx->state = DEMUX_PV_STATE_STARTING;
1781 	/* have QP0 only if link layer is IB */
1782 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1783 	    IB_LINK_LAYER_INFINIBAND)
1784 		ctx->has_smi = 1;
1785 
1786 	if (ctx->has_smi) {
1787 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1788 		if (ret) {
1789 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1790 			goto err_out;
1791 		}
1792 	}
1793 
1794 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1795 	if (ret) {
1796 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1797 		goto err_out_qp0;
1798 	}
1799 
1800 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1801 	if (ctx->has_smi)
1802 		cq_size *= 2;
1803 
1804 	cq_attr.cqe = cq_size;
1805 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1806 			       NULL, ctx, &cq_attr);
1807 	if (IS_ERR(ctx->cq)) {
1808 		ret = PTR_ERR(ctx->cq);
1809 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1810 		goto err_buf;
1811 	}
1812 
1813 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1814 	if (IS_ERR(ctx->pd)) {
1815 		ret = PTR_ERR(ctx->pd);
1816 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1817 		goto err_cq;
1818 	}
1819 
1820 	ctx->mr = ib_get_dma_mr(ctx->pd, IB_ACCESS_LOCAL_WRITE);
1821 	if (IS_ERR(ctx->mr)) {
1822 		ret = PTR_ERR(ctx->mr);
1823 		pr_err("Couldn't get tunnel DMA MR (%d)\n", ret);
1824 		goto err_pd;
1825 	}
1826 
1827 	if (ctx->has_smi) {
1828 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1829 		if (ret) {
1830 			pr_err("Couldn't create %s QP0 (%d)\n",
1831 			       create_tun ? "tunnel for" : "",  ret);
1832 			goto err_mr;
1833 		}
1834 	}
1835 
1836 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1837 	if (ret) {
1838 		pr_err("Couldn't create %s QP1 (%d)\n",
1839 		       create_tun ? "tunnel for" : "",  ret);
1840 		goto err_qp0;
1841 	}
1842 
1843 	if (create_tun)
1844 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1845 	else
1846 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1847 
1848 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1849 
1850 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1851 	if (ret) {
1852 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1853 		goto err_wq;
1854 	}
1855 	ctx->state = DEMUX_PV_STATE_ACTIVE;
1856 	return 0;
1857 
1858 err_wq:
1859 	ctx->wq = NULL;
1860 	ib_destroy_qp(ctx->qp[1].qp);
1861 	ctx->qp[1].qp = NULL;
1862 
1863 
1864 err_qp0:
1865 	if (ctx->has_smi)
1866 		ib_destroy_qp(ctx->qp[0].qp);
1867 	ctx->qp[0].qp = NULL;
1868 
1869 err_mr:
1870 	ib_dereg_mr(ctx->mr);
1871 	ctx->mr = NULL;
1872 
1873 err_pd:
1874 	ib_dealloc_pd(ctx->pd);
1875 	ctx->pd = NULL;
1876 
1877 err_cq:
1878 	ib_destroy_cq(ctx->cq);
1879 	ctx->cq = NULL;
1880 
1881 err_buf:
1882 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1883 
1884 err_out_qp0:
1885 	if (ctx->has_smi)
1886 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1887 err_out:
1888 	ctx->state = DEMUX_PV_STATE_DOWN;
1889 	return ret;
1890 }
1891 
1892 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1893 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1894 {
1895 	if (!ctx)
1896 		return;
1897 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
1898 		ctx->state = DEMUX_PV_STATE_DOWNING;
1899 		if (flush)
1900 			flush_workqueue(ctx->wq);
1901 		if (ctx->has_smi) {
1902 			ib_destroy_qp(ctx->qp[0].qp);
1903 			ctx->qp[0].qp = NULL;
1904 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1905 		}
1906 		ib_destroy_qp(ctx->qp[1].qp);
1907 		ctx->qp[1].qp = NULL;
1908 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1909 		ib_dereg_mr(ctx->mr);
1910 		ctx->mr = NULL;
1911 		ib_dealloc_pd(ctx->pd);
1912 		ctx->pd = NULL;
1913 		ib_destroy_cq(ctx->cq);
1914 		ctx->cq = NULL;
1915 		ctx->state = DEMUX_PV_STATE_DOWN;
1916 	}
1917 }
1918 
1919 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1920 				  int port, int do_init)
1921 {
1922 	int ret = 0;
1923 
1924 	if (!do_init) {
1925 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
1926 		/* for master, destroy real sqp resources */
1927 		if (slave == mlx4_master_func_num(dev->dev))
1928 			destroy_pv_resources(dev, slave, port,
1929 					     dev->sriov.sqps[port - 1], 1);
1930 		/* destroy the tunnel qp resources */
1931 		destroy_pv_resources(dev, slave, port,
1932 				     dev->sriov.demux[port - 1].tun[slave], 1);
1933 		return 0;
1934 	}
1935 
1936 	/* create the tunnel qp resources */
1937 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
1938 				  dev->sriov.demux[port - 1].tun[slave]);
1939 
1940 	/* for master, create the real sqp resources */
1941 	if (!ret && slave == mlx4_master_func_num(dev->dev))
1942 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
1943 					  dev->sriov.sqps[port - 1]);
1944 	return ret;
1945 }
1946 
1947 void mlx4_ib_tunnels_update_work(struct work_struct *work)
1948 {
1949 	struct mlx4_ib_demux_work *dmxw;
1950 
1951 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
1952 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
1953 			       dmxw->do_init);
1954 	kfree(dmxw);
1955 	return;
1956 }
1957 
1958 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
1959 				       struct mlx4_ib_demux_ctx *ctx,
1960 				       int port)
1961 {
1962 	char name[12];
1963 	int ret = 0;
1964 	int i;
1965 
1966 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
1967 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
1968 	if (!ctx->tun)
1969 		return -ENOMEM;
1970 
1971 	ctx->dev = dev;
1972 	ctx->port = port;
1973 	ctx->ib_dev = &dev->ib_dev;
1974 
1975 	for (i = 0;
1976 	     i < min(dev->dev->caps.sqp_demux,
1977 	     (u16)(dev->dev->persist->num_vfs + 1));
1978 	     i++) {
1979 		struct mlx4_active_ports actv_ports =
1980 			mlx4_get_active_ports(dev->dev, i);
1981 
1982 		if (!test_bit(port - 1, actv_ports.ports))
1983 			continue;
1984 
1985 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
1986 		if (ret) {
1987 			ret = -ENOMEM;
1988 			goto err_mcg;
1989 		}
1990 	}
1991 
1992 	ret = mlx4_ib_mcg_port_init(ctx);
1993 	if (ret) {
1994 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
1995 		goto err_mcg;
1996 	}
1997 
1998 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
1999 	ctx->wq = create_singlethread_workqueue(name);
2000 	if (!ctx->wq) {
2001 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2002 		ret = -ENOMEM;
2003 		goto err_wq;
2004 	}
2005 
2006 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2007 	ctx->ud_wq = create_singlethread_workqueue(name);
2008 	if (!ctx->ud_wq) {
2009 		pr_err("Failed to create up/down WQ for port %d\n", port);
2010 		ret = -ENOMEM;
2011 		goto err_udwq;
2012 	}
2013 
2014 	return 0;
2015 
2016 err_udwq:
2017 	destroy_workqueue(ctx->wq);
2018 	ctx->wq = NULL;
2019 
2020 err_wq:
2021 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2022 err_mcg:
2023 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2024 		free_pv_object(dev, i, port);
2025 	kfree(ctx->tun);
2026 	ctx->tun = NULL;
2027 	return ret;
2028 }
2029 
2030 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2031 {
2032 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2033 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2034 		flush_workqueue(sqp_ctx->wq);
2035 		if (sqp_ctx->has_smi) {
2036 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2037 			sqp_ctx->qp[0].qp = NULL;
2038 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2039 		}
2040 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2041 		sqp_ctx->qp[1].qp = NULL;
2042 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2043 		ib_dereg_mr(sqp_ctx->mr);
2044 		sqp_ctx->mr = NULL;
2045 		ib_dealloc_pd(sqp_ctx->pd);
2046 		sqp_ctx->pd = NULL;
2047 		ib_destroy_cq(sqp_ctx->cq);
2048 		sqp_ctx->cq = NULL;
2049 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2050 	}
2051 }
2052 
2053 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2054 {
2055 	int i;
2056 	if (ctx) {
2057 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2058 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2059 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2060 			if (!ctx->tun[i])
2061 				continue;
2062 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2063 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2064 		}
2065 		flush_workqueue(ctx->wq);
2066 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2067 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2068 			free_pv_object(dev, i, ctx->port);
2069 		}
2070 		kfree(ctx->tun);
2071 		destroy_workqueue(ctx->ud_wq);
2072 		destroy_workqueue(ctx->wq);
2073 	}
2074 }
2075 
2076 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2077 {
2078 	int i;
2079 
2080 	if (!mlx4_is_master(dev->dev))
2081 		return;
2082 	/* initialize or tear down tunnel QPs for the master */
2083 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2084 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2085 	return;
2086 }
2087 
2088 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2089 {
2090 	int i = 0;
2091 	int err;
2092 
2093 	if (!mlx4_is_mfunc(dev->dev))
2094 		return 0;
2095 
2096 	dev->sriov.is_going_down = 0;
2097 	spin_lock_init(&dev->sriov.going_down_lock);
2098 	mlx4_ib_cm_paravirt_init(dev);
2099 
2100 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2101 
2102 	if (mlx4_is_slave(dev->dev)) {
2103 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2104 		return 0;
2105 	}
2106 
2107 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2108 		if (i == mlx4_master_func_num(dev->dev))
2109 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2110 		else
2111 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2112 	}
2113 
2114 	err = mlx4_ib_init_alias_guid_service(dev);
2115 	if (err) {
2116 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2117 		goto paravirt_err;
2118 	}
2119 	err = mlx4_ib_device_register_sysfs(dev);
2120 	if (err) {
2121 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2122 		goto sysfs_err;
2123 	}
2124 
2125 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2126 		     dev->dev->caps.sqp_demux);
2127 	for (i = 0; i < dev->num_ports; i++) {
2128 		union ib_gid gid;
2129 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2130 		if (err)
2131 			goto demux_err;
2132 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2133 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2134 				      &dev->sriov.sqps[i]);
2135 		if (err)
2136 			goto demux_err;
2137 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2138 		if (err)
2139 			goto free_pv;
2140 	}
2141 	mlx4_ib_master_tunnels(dev, 1);
2142 	return 0;
2143 
2144 free_pv:
2145 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2146 demux_err:
2147 	while (--i >= 0) {
2148 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2149 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2150 	}
2151 	mlx4_ib_device_unregister_sysfs(dev);
2152 
2153 sysfs_err:
2154 	mlx4_ib_destroy_alias_guid_service(dev);
2155 
2156 paravirt_err:
2157 	mlx4_ib_cm_paravirt_clean(dev, -1);
2158 
2159 	return err;
2160 }
2161 
2162 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2163 {
2164 	int i;
2165 	unsigned long flags;
2166 
2167 	if (!mlx4_is_mfunc(dev->dev))
2168 		return;
2169 
2170 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2171 	dev->sriov.is_going_down = 1;
2172 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2173 	if (mlx4_is_master(dev->dev)) {
2174 		for (i = 0; i < dev->num_ports; i++) {
2175 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2176 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2177 			kfree(dev->sriov.sqps[i]);
2178 			dev->sriov.sqps[i] = NULL;
2179 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2180 		}
2181 
2182 		mlx4_ib_cm_paravirt_clean(dev, -1);
2183 		mlx4_ib_destroy_alias_guid_service(dev);
2184 		mlx4_ib_device_unregister_sysfs(dev);
2185 	}
2186 }
2187