xref: /freebsd/sys/dev/mlx4/mlx4_ib/mlx4_ib_mad.c (revision 62cfcf62f627e5093fb37026a6d8c98e4d2ef04c)
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 <dev/mlx4/cmd.h>
40 #include <dev/mlx4/driver.h>
41 #include <linux/gfp.h>
42 #include <rdma/ib_pma.h>
43 
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 | random());
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 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
234 				return;
235 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
236 			lid = be16_to_cpu(pinfo->lid);
237 
238 			update_sm_ah(dev, port_num,
239 				     be16_to_cpu(pinfo->sm_lid),
240 				     pinfo->neighbormtu_mastersmsl & 0xf);
241 
242 			if (pinfo->clientrereg_resv_subnetto & 0x80)
243 				handle_client_rereg_event(dev, port_num);
244 
245 			if (prev_lid != lid)
246 				handle_lid_change_event(dev, port_num);
247 			break;
248 
249 		case IB_SMP_ATTR_PKEY_TABLE:
250 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
251 				return;
252 			if (!mlx4_is_mfunc(dev->dev)) {
253 				mlx4_ib_dispatch_event(dev, port_num,
254 						       IB_EVENT_PKEY_CHANGE);
255 				break;
256 			}
257 
258 			/* at this point, we are running in the master.
259 			 * Slaves do not receive SMPs.
260 			 */
261 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
262 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
263 			pkey_change_bitmap = 0;
264 			for (i = 0; i < 32; i++) {
265 				pr_debug("PKEY[%d] = x%x\n",
266 					 i + bn*32, be16_to_cpu(base[i]));
267 				if (be16_to_cpu(base[i]) !=
268 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
269 					pkey_change_bitmap |= (1 << i);
270 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
271 						be16_to_cpu(base[i]);
272 				}
273 			}
274 			pr_debug("PKEY Change event: port=%d, "
275 				 "block=0x%x, change_bitmap=0x%x\n",
276 				 port_num, bn, pkey_change_bitmap);
277 
278 			if (pkey_change_bitmap) {
279 				mlx4_ib_dispatch_event(dev, port_num,
280 						       IB_EVENT_PKEY_CHANGE);
281 				if (!dev->sriov.is_going_down)
282 					__propagate_pkey_ev(dev, port_num, bn,
283 							    pkey_change_bitmap);
284 			}
285 			break;
286 
287 		case IB_SMP_ATTR_GUID_INFO:
288 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV)
289 				return;
290 			/* paravirtualized master's guid is guid 0 -- does not change */
291 			if (!mlx4_is_master(dev->dev))
292 				mlx4_ib_dispatch_event(dev, port_num,
293 						       IB_EVENT_GID_CHANGE);
294 			/*if master, notify relevant slaves*/
295 			if (mlx4_is_master(dev->dev) &&
296 			    !dev->sriov.is_going_down) {
297 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
298 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
299 								    (u8 *)(&((struct ib_smp *)mad)->data));
300 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
301 								     (u8 *)(&((struct ib_smp *)mad)->data));
302 			}
303 			break;
304 
305 		case IB_SMP_ATTR_SL_TO_VL_TABLE:
306 			/* cache sl to vl mapping changes for use in
307 			 * filling QP1 LRH VL field when sending packets
308 			 */
309 			if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV &&
310 			    dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)
311 				return;
312 			if (!mlx4_is_slave(dev->dev)) {
313 				union sl2vl_tbl_to_u64 sl2vl64;
314 				int jj;
315 
316 				for (jj = 0; jj < 8; jj++) {
317 					sl2vl64.sl8[jj] = ((struct ib_smp *)mad)->data[jj];
318 					pr_debug("sl2vl[%d] = %02x\n", jj, sl2vl64.sl8[jj]);
319 				}
320 				atomic64_set(&dev->sl2vl[port_num - 1], sl2vl64.sl64);
321 			}
322 			break;
323 
324 		default:
325 			break;
326 		}
327 }
328 
329 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
330 				int block, u32 change_bitmap)
331 {
332 	int i, ix, slave, err;
333 	int have_event = 0;
334 
335 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
336 		if (slave == mlx4_master_func_num(dev->dev))
337 			continue;
338 		if (!mlx4_is_slave_active(dev->dev, slave))
339 			continue;
340 
341 		have_event = 0;
342 		for (i = 0; i < 32; i++) {
343 			if (!(change_bitmap & (1 << i)))
344 				continue;
345 			for (ix = 0;
346 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
347 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
348 				    [ix] == i + 32 * block) {
349 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
350 					pr_debug("propagate_pkey_ev: slave %d,"
351 						 " port %d, ix %d (%d)\n",
352 						 slave, port_num, ix, err);
353 					have_event = 1;
354 					break;
355 				}
356 			}
357 			if (have_event)
358 				break;
359 		}
360 	}
361 }
362 
363 static void node_desc_override(struct ib_device *dev,
364 			       struct ib_mad *mad)
365 {
366 	unsigned long flags;
367 
368 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
369 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
370 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
371 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
372 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
373 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
374 		       IB_DEVICE_NODE_DESC_MAX);
375 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
376 	}
377 }
378 
379 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
380 {
381 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
382 	struct ib_mad_send_buf *send_buf;
383 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
384 	int ret;
385 	unsigned long flags;
386 
387 	if (agent) {
388 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
389 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
390 					      IB_MGMT_BASE_VERSION);
391 		if (IS_ERR(send_buf))
392 			return;
393 		/*
394 		 * We rely here on the fact that MLX QPs don't use the
395 		 * address handle after the send is posted (this is
396 		 * wrong following the IB spec strictly, but we know
397 		 * it's OK for our devices).
398 		 */
399 		spin_lock_irqsave(&dev->sm_lock, flags);
400 		memcpy(send_buf->mad, mad, sizeof *mad);
401 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
402 			ret = ib_post_send_mad(send_buf, NULL);
403 		else
404 			ret = -EINVAL;
405 		spin_unlock_irqrestore(&dev->sm_lock, flags);
406 
407 		if (ret)
408 			ib_free_send_mad(send_buf);
409 	}
410 }
411 
412 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
413 							     struct ib_sa_mad *sa_mad)
414 {
415 	int ret = 0;
416 
417 	/* dispatch to different sa handlers */
418 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
419 	case IB_SA_ATTR_MC_MEMBER_REC:
420 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
421 		break;
422 	default:
423 		break;
424 	}
425 	return ret;
426 }
427 
428 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
429 {
430 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
431 	int i;
432 
433 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
434 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
435 			return i;
436 	}
437 	return -1;
438 }
439 
440 
441 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
442 				   u8 port, u16 pkey, u16 *ix)
443 {
444 	int i, ret;
445 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
446 	u16 slot_pkey;
447 
448 	if (slave == mlx4_master_func_num(dev->dev))
449 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
450 
451 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
452 
453 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
454 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
455 			continue;
456 
457 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
458 
459 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
460 		if (ret)
461 			continue;
462 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
463 			if (slot_pkey & 0x8000) {
464 				*ix = (u16) pkey_ix;
465 				return 0;
466 			} else {
467 				/* take first partial pkey index found */
468 				if (partial_ix == 0xFF)
469 					partial_ix = pkey_ix;
470 			}
471 		}
472 	}
473 
474 	if (partial_ix < 0xFF) {
475 		*ix = (u16) partial_ix;
476 		return 0;
477 	}
478 
479 	return -EINVAL;
480 }
481 
482 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
483 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
484 			  struct ib_grh *grh, struct ib_mad *mad)
485 {
486 	struct ib_sge list;
487 	struct ib_ud_wr wr;
488 	struct ib_send_wr *bad_wr;
489 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
490 	struct mlx4_ib_demux_pv_qp *tun_qp;
491 	struct mlx4_rcv_tunnel_mad *tun_mad;
492 	struct ib_ah_attr attr;
493 	struct ib_ah *ah;
494 	struct ib_qp *src_qp = NULL;
495 	unsigned tun_tx_ix = 0;
496 	int dqpn;
497 	int ret = 0;
498 	u16 tun_pkey_ix;
499 	u16 cached_pkey;
500 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
501 
502 	if (dest_qpt > IB_QPT_GSI)
503 		return -EINVAL;
504 
505 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
506 
507 	/* check if proxy qp created */
508 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
509 		return -EAGAIN;
510 
511 	if (!dest_qpt)
512 		tun_qp = &tun_ctx->qp[0];
513 	else
514 		tun_qp = &tun_ctx->qp[1];
515 
516 	/* compute P_Key index to put in tunnel header for slave */
517 	if (dest_qpt) {
518 		u16 pkey_ix;
519 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
520 		if (ret)
521 			return -EINVAL;
522 
523 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
524 		if (ret)
525 			return -EINVAL;
526 		tun_pkey_ix = pkey_ix;
527 	} else
528 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
529 
530 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
531 
532 	/* get tunnel tx data buf for slave */
533 	src_qp = tun_qp->qp;
534 
535 	/* create ah. Just need an empty one with the port num for the post send.
536 	 * The driver will set the force loopback bit in post_send */
537 	memset(&attr, 0, sizeof attr);
538 	attr.port_num = port;
539 	if (is_eth) {
540 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
541 		attr.ah_flags = IB_AH_GRH;
542 	}
543 	ah = ib_create_ah(tun_ctx->pd, &attr);
544 	if (IS_ERR(ah))
545 		return -ENOMEM;
546 
547 	/* allocate tunnel tx buf after pass failure returns */
548 	spin_lock(&tun_qp->tx_lock);
549 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
550 	    (MLX4_NUM_TUNNEL_BUFS - 1))
551 		ret = -EAGAIN;
552 	else
553 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
554 	spin_unlock(&tun_qp->tx_lock);
555 	if (ret)
556 		goto end;
557 
558 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
559 	if (tun_qp->tx_ring[tun_tx_ix].ah)
560 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
561 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
562 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
563 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
564 				   sizeof (struct mlx4_rcv_tunnel_mad),
565 				   DMA_TO_DEVICE);
566 
567 	/* copy over to tunnel buffer */
568 	if (grh)
569 		memcpy(&tun_mad->grh, grh, sizeof *grh);
570 	memcpy(&tun_mad->mad, mad, sizeof *mad);
571 
572 	/* adjust tunnel data */
573 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
574 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
575 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
576 
577 	if (is_eth) {
578 		u16 vlan = 0;
579 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
580 						NULL)) {
581 			/* VST mode */
582 			if (vlan != wc->vlan_id) {
583 				/* Packet vlan is not the VST-assigned vlan.
584 				 * Drop the packet.
585 				 */
586 				ret = -EPERM;
587 				goto out;
588 			} else {
589 				/* Remove the vlan tag before forwarding
590 				 * the packet to the VF.
591 				 */
592 				vlan = 0xffff;
593 			}
594 		} else {
595 			vlan = wc->vlan_id;
596 		}
597 
598 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
599 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
600 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
601 	} else {
602 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
603 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
604 	}
605 
606 	ib_dma_sync_single_for_device(&dev->ib_dev,
607 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
608 				      sizeof (struct mlx4_rcv_tunnel_mad),
609 				      DMA_TO_DEVICE);
610 
611 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
612 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
613 	list.lkey = tun_ctx->pd->local_dma_lkey;
614 
615 	wr.ah = ah;
616 	wr.port_num = port;
617 	wr.remote_qkey = IB_QP_SET_QKEY;
618 	wr.remote_qpn = dqpn;
619 	wr.wr.next = NULL;
620 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
621 	wr.wr.sg_list = &list;
622 	wr.wr.num_sge = 1;
623 	wr.wr.opcode = IB_WR_SEND;
624 	wr.wr.send_flags = IB_SEND_SIGNALED;
625 
626 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
627 	if (!ret)
628 		return 0;
629  out:
630 	spin_lock(&tun_qp->tx_lock);
631 	tun_qp->tx_ix_tail++;
632 	spin_unlock(&tun_qp->tx_lock);
633 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
634 end:
635 	ib_destroy_ah(ah);
636 	return ret;
637 }
638 
639 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
640 			struct ib_wc *wc, struct ib_grh *grh,
641 			struct ib_mad *mad)
642 {
643 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
644 	int err, other_port;
645 	int slave = -1;
646 	u8 *slave_id;
647 	int is_eth = 0;
648 
649 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
650 		is_eth = 0;
651 	else
652 		is_eth = 1;
653 
654 	if (is_eth) {
655 		if (!(wc->wc_flags & IB_WC_GRH)) {
656 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
657 			return -EINVAL;
658 		}
659 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
660 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
661 			return -EINVAL;
662 		}
663 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave);
664 		if (err && mlx4_is_mf_bonded(dev->dev)) {
665 			other_port = (port == 1) ? 2 : 1;
666 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, grh->dgid.raw, &slave);
667 			if (!err) {
668 				port = other_port;
669 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
670 					 slave, grh->dgid.raw, port, other_port);
671 			}
672 		}
673 		if (err) {
674 			mlx4_ib_warn(ibdev, "failed matching grh\n");
675 			return -ENOENT;
676 		}
677 		if (slave >= dev->dev->caps.sqp_demux) {
678 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
679 				     slave, dev->dev->caps.sqp_demux);
680 			return -ENOENT;
681 		}
682 
683 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
684 			return 0;
685 
686 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
687 		if (err)
688 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
689 				 slave, err);
690 		return 0;
691 	}
692 
693 	/* Initially assume that this mad is for us */
694 	slave = mlx4_master_func_num(dev->dev);
695 
696 	/* See if the slave id is encoded in a response mad */
697 	if (mad->mad_hdr.method & 0x80) {
698 		slave_id = (u8 *) &mad->mad_hdr.tid;
699 		slave = *slave_id;
700 		if (slave != 255) /*255 indicates the dom0*/
701 			*slave_id = 0; /* remap tid */
702 	}
703 
704 	/* If a grh is present, we demux according to it */
705 	if (wc->wc_flags & IB_WC_GRH) {
706 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
707 		if (slave < 0) {
708 			mlx4_ib_warn(ibdev, "failed matching grh\n");
709 			return -ENOENT;
710 		}
711 	}
712 	/* Class-specific handling */
713 	switch (mad->mad_hdr.mgmt_class) {
714 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
715 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
716 		/* 255 indicates the dom0 */
717 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
718 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
719 				return -EPERM;
720 			/* for a VF. drop unsolicited MADs */
721 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
722 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
723 					     slave, mad->mad_hdr.mgmt_class,
724 					     mad->mad_hdr.method);
725 				return -EINVAL;
726 			}
727 		}
728 		break;
729 	case IB_MGMT_CLASS_SUBN_ADM:
730 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
731 					     (struct ib_sa_mad *) mad))
732 			return 0;
733 		break;
734 	case IB_MGMT_CLASS_CM:
735 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
736 			return 0;
737 		break;
738 	case IB_MGMT_CLASS_DEVICE_MGMT:
739 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
740 			return 0;
741 		break;
742 	default:
743 		/* Drop unsupported classes for slaves in tunnel mode */
744 		if (slave != mlx4_master_func_num(dev->dev)) {
745 			pr_debug("dropping unsupported ingress mad from class:%d "
746 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
747 			return 0;
748 		}
749 	}
750 	/*make sure that no slave==255 was not handled yet.*/
751 	if (slave >= dev->dev->caps.sqp_demux) {
752 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
753 			     slave, dev->dev->caps.sqp_demux);
754 		return -ENOENT;
755 	}
756 
757 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
758 	if (err)
759 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
760 			 slave, err);
761 	return 0;
762 }
763 
764 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
765 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
766 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
767 {
768 	u16 slid, prev_lid = 0;
769 	int err;
770 	struct ib_port_attr pattr;
771 
772 	if (in_wc && in_wc->qp->qp_num) {
773 		pr_debug("received MAD: slid:%d sqpn:%d "
774 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
775 			in_wc->slid, in_wc->src_qp,
776 			in_wc->dlid_path_bits,
777 			in_wc->qp->qp_num,
778 			in_wc->wc_flags,
779 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
780 			be16_to_cpu(in_mad->mad_hdr.attr_id));
781 		if (in_wc->wc_flags & IB_WC_GRH) {
782 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
783 				 (unsigned long long)be64_to_cpu(in_grh->sgid.global.subnet_prefix),
784 				 (unsigned long long)be64_to_cpu(in_grh->sgid.global.interface_id));
785 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
786 				 (unsigned long long)be64_to_cpu(in_grh->dgid.global.subnet_prefix),
787 				 (unsigned long long)be64_to_cpu(in_grh->dgid.global.interface_id));
788 		}
789 	}
790 
791 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
792 
793 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
794 		forward_trap(to_mdev(ibdev), port_num, in_mad);
795 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
796 	}
797 
798 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
799 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
800 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
801 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
802 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
803 			return IB_MAD_RESULT_SUCCESS;
804 
805 		/*
806 		 * Don't process SMInfo queries -- the SMA can't handle them.
807 		 */
808 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
809 			return IB_MAD_RESULT_SUCCESS;
810 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
811 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
812 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
813 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
814 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
815 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
816 			return IB_MAD_RESULT_SUCCESS;
817 	} else
818 		return IB_MAD_RESULT_SUCCESS;
819 
820 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
821 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
822 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
823 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
824 	    !ib_query_port(ibdev, port_num, &pattr))
825 		prev_lid = pattr.lid;
826 
827 	err = mlx4_MAD_IFC(to_mdev(ibdev),
828 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
829 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
830 			   MLX4_MAD_IFC_NET_VIEW,
831 			   port_num, in_wc, in_grh, in_mad, out_mad);
832 	if (err)
833 		return IB_MAD_RESULT_FAILURE;
834 
835 	if (!out_mad->mad_hdr.status) {
836 		smp_snoop(ibdev, port_num, in_mad, prev_lid);
837 		/* slaves get node desc from FW */
838 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
839 			node_desc_override(ibdev, out_mad);
840 	}
841 
842 	/* set return bit in status of directed route responses */
843 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
844 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
845 
846 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
847 		/* no response for trap repress */
848 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
849 
850 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
851 }
852 
853 static void edit_counter(struct mlx4_counter *cnt, void *counters,
854 			 __be16 attr_id)
855 {
856 	switch (attr_id) {
857 	case IB_PMA_PORT_COUNTERS:
858 	{
859 		struct ib_pma_portcounters *pma_cnt =
860 			(struct ib_pma_portcounters *)counters;
861 
862 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
863 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
864 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
865 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
866 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
867 				     be64_to_cpu(cnt->tx_frames));
868 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
869 				     be64_to_cpu(cnt->rx_frames));
870 		break;
871 	}
872 	case IB_PMA_PORT_COUNTERS_EXT:
873 	{
874 		struct ib_pma_portcounters_ext *pma_cnt_ext =
875 			(struct ib_pma_portcounters_ext *)counters;
876 
877 		pma_cnt_ext->port_xmit_data =
878 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
879 		pma_cnt_ext->port_rcv_data =
880 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
881 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
882 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
883 		break;
884 	}
885 	default:
886 		break;
887 	}
888 }
889 
890 static int iboe_process_mad_port_info(void *out_mad)
891 {
892 	struct ib_class_port_info cpi = {};
893 
894 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
895 	memcpy(out_mad, &cpi, sizeof(cpi));
896 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
897 }
898 
899 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
900 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
901 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
902 {
903 	struct mlx4_counter counter_stats;
904 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
905 	struct counter_index *tmp_counter;
906 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
907 
908 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
909 		return -EINVAL;
910 
911 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
912 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
913 
914 	memset(&counter_stats, 0, sizeof(counter_stats));
915 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
916 	list_for_each_entry(tmp_counter,
917 			    &dev->counters_table[port_num - 1].counters_list,
918 			    list) {
919 		err = mlx4_get_counter_stats(dev->dev,
920 					     tmp_counter->index,
921 					     &counter_stats, 0);
922 		if (err) {
923 			err = IB_MAD_RESULT_FAILURE;
924 			stats_avail = 0;
925 			break;
926 		}
927 		stats_avail = 1;
928 	}
929 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
930 	if (stats_avail) {
931 		memset(out_mad->data, 0, sizeof out_mad->data);
932 		switch (counter_stats.counter_mode & 0xf) {
933 		case 0:
934 			edit_counter(&counter_stats,
935 				     (void *)(out_mad->data + 40),
936 				     in_mad->mad_hdr.attr_id);
937 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
938 			break;
939 		default:
940 			err = IB_MAD_RESULT_FAILURE;
941 		}
942 	}
943 
944 	return err;
945 }
946 
947 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
948 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
949 			const struct ib_mad_hdr *in, size_t in_mad_size,
950 			struct ib_mad_hdr *out, size_t *out_mad_size,
951 			u16 *out_mad_pkey_index)
952 {
953 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
954 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
955 	struct ib_mad *out_mad = (struct ib_mad *)out;
956 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
957 
958 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
959 			 *out_mad_size != sizeof(*out_mad)))
960 		return IB_MAD_RESULT_FAILURE;
961 
962 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
963 	 * queries, should be called only by VFs and for that specific purpose
964 	 */
965 	if (link == IB_LINK_LAYER_INFINIBAND) {
966 		if (mlx4_is_slave(dev->dev) &&
967 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
968 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
969 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
970 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
971 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
972 						in_grh, in_mad, out_mad);
973 
974 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
975 				      in_grh, in_mad, out_mad);
976 	}
977 
978 	if (link == IB_LINK_LAYER_ETHERNET)
979 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
980 					in_grh, in_mad, out_mad);
981 
982 	return -EINVAL;
983 }
984 
985 static void send_handler(struct ib_mad_agent *agent,
986 			 struct ib_mad_send_wc *mad_send_wc)
987 {
988 	if (mad_send_wc->send_buf->context[0])
989 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
990 	ib_free_send_mad(mad_send_wc->send_buf);
991 }
992 
993 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
994 {
995 	struct ib_mad_agent *agent;
996 	int p, q;
997 	int ret;
998 	enum rdma_link_layer ll;
999 
1000 	for (p = 0; p < dev->num_ports; ++p) {
1001 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
1002 		for (q = 0; q <= 1; ++q) {
1003 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1004 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1005 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1006 							      NULL, 0, send_handler,
1007 							      NULL, NULL, 0);
1008 				if (IS_ERR(agent)) {
1009 					ret = PTR_ERR(agent);
1010 					goto err;
1011 				}
1012 				dev->send_agent[p][q] = agent;
1013 			} else
1014 				dev->send_agent[p][q] = NULL;
1015 		}
1016 	}
1017 
1018 	return 0;
1019 
1020 err:
1021 	for (p = 0; p < dev->num_ports; ++p)
1022 		for (q = 0; q <= 1; ++q)
1023 			if (dev->send_agent[p][q])
1024 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1025 
1026 	return ret;
1027 }
1028 
1029 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1030 {
1031 	struct ib_mad_agent *agent;
1032 	int p, q;
1033 
1034 	for (p = 0; p < dev->num_ports; ++p) {
1035 		for (q = 0; q <= 1; ++q) {
1036 			agent = dev->send_agent[p][q];
1037 			if (agent) {
1038 				dev->send_agent[p][q] = NULL;
1039 				ib_unregister_mad_agent(agent);
1040 			}
1041 		}
1042 
1043 		if (dev->sm_ah[p])
1044 			ib_destroy_ah(dev->sm_ah[p]);
1045 	}
1046 }
1047 
1048 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1049 {
1050 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1051 
1052 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1053 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1054 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1055 }
1056 
1057 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1058 {
1059 	/* re-configure the alias-guid and mcg's */
1060 	if (mlx4_is_master(dev->dev)) {
1061 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1062 
1063 		if (!dev->sriov.is_going_down) {
1064 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1065 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1066 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1067 		}
1068 	}
1069 
1070 	/* Update the sl to vl table from inside client rereg
1071 	 * only if in secure-host mode (snooping is not possible)
1072 	 * and the sl-to-vl change event is not generated by FW.
1073 	 */
1074 	if (!mlx4_is_slave(dev->dev) &&
1075 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1076 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1077 		if (mlx4_is_master(dev->dev))
1078 			/* already in work queue from mlx4_ib_event queueing
1079 			 * mlx4_handle_port_mgmt_change_event, which calls
1080 			 * this procedure. Therefore, call sl2vl_update directly.
1081 			 */
1082 			mlx4_ib_sl2vl_update(dev, port_num);
1083 		else
1084 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1085 	}
1086 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1087 }
1088 
1089 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1090 			      struct mlx4_eqe *eqe)
1091 {
1092 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1093 			    GET_MASK_FROM_EQE(eqe));
1094 }
1095 
1096 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1097 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1098 {
1099 	struct ib_smp *in_mad  = NULL;
1100 	struct ib_smp *out_mad  = NULL;
1101 	u16 i;
1102 
1103 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1104 		return;
1105 
1106 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1107 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1108 	if (!in_mad || !out_mad) {
1109 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1110 		goto out;
1111 	}
1112 
1113 	guid_tbl_blk_num  *= 4;
1114 
1115 	for (i = 0; i < 4; i++) {
1116 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1117 			continue;
1118 		memset(in_mad, 0, sizeof *in_mad);
1119 		memset(out_mad, 0, sizeof *out_mad);
1120 
1121 		in_mad->base_version  = 1;
1122 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1123 		in_mad->class_version = 1;
1124 		in_mad->method        = IB_MGMT_METHOD_GET;
1125 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1126 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1127 
1128 		if (mlx4_MAD_IFC(dev,
1129 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1130 				 port_num, NULL, NULL, in_mad, out_mad)) {
1131 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1132 			goto out;
1133 		}
1134 
1135 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1136 						    port_num,
1137 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1138 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1139 						     port_num,
1140 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1141 	}
1142 
1143 out:
1144 	kfree(in_mad);
1145 	kfree(out_mad);
1146 	return;
1147 }
1148 
1149 void handle_port_mgmt_change_event(struct work_struct *work)
1150 {
1151 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1152 	struct mlx4_ib_dev *dev = ew->ib_dev;
1153 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1154 	u8 port = eqe->event.port_mgmt_change.port;
1155 	u32 changed_attr;
1156 	u32 tbl_block;
1157 	u32 change_bitmap;
1158 
1159 	switch (eqe->subtype) {
1160 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1161 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1162 
1163 		/* Update the SM ah - This should be done before handling
1164 		   the other changed attributes so that MADs can be sent to the SM */
1165 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1166 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1167 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1168 			update_sm_ah(dev, port, lid, sl);
1169 		}
1170 
1171 		/* Check if it is a lid change event */
1172 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1173 			handle_lid_change_event(dev, port);
1174 
1175 		/* Generate GUID changed event */
1176 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1177 			if (mlx4_is_master(dev->dev)) {
1178 				union ib_gid gid;
1179 				int err = 0;
1180 
1181 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1182 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1183 				else
1184 					gid.global.subnet_prefix =
1185 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1186 				if (err) {
1187 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1188 						port, err);
1189 				} else {
1190 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1191 						 port,
1192 						 (long long)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1193 						 (long long)be64_to_cpu(gid.global.subnet_prefix));
1194 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1195 						     be64_to_cpu(gid.global.subnet_prefix));
1196 				}
1197 			}
1198 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1199 			/*if master, notify all slaves*/
1200 			if (mlx4_is_master(dev->dev))
1201 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1202 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1203 		}
1204 
1205 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1206 			handle_client_rereg_event(dev, port);
1207 		break;
1208 
1209 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1210 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1211 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1212 			propagate_pkey_ev(dev, port, eqe);
1213 		break;
1214 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1215 		/* paravirtualized master's guid is guid 0 -- does not change */
1216 		if (!mlx4_is_master(dev->dev))
1217 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1218 		/*if master, notify relevant slaves*/
1219 		else if (!dev->sriov.is_going_down) {
1220 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1221 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1222 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1223 		}
1224 		break;
1225 
1226 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1227 		/* cache sl to vl mapping changes for use in
1228 		 * filling QP1 LRH VL field when sending packets
1229 		 */
1230 		if (!mlx4_is_slave(dev->dev)) {
1231 			union sl2vl_tbl_to_u64 sl2vl64;
1232 			int jj;
1233 
1234 			for (jj = 0; jj < 8; jj++) {
1235 				sl2vl64.sl8[jj] =
1236 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1237 				pr_debug("sl2vl[%d] = %02x\n", jj, sl2vl64.sl8[jj]);
1238 			}
1239 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1240 		}
1241 		break;
1242 	default:
1243 		pr_warn("Unsupported subtype 0x%x for "
1244 			"Port Management Change event\n", eqe->subtype);
1245 	}
1246 
1247 	kfree(ew);
1248 }
1249 
1250 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1251 			    enum ib_event_type type)
1252 {
1253 	struct ib_event event;
1254 
1255 	event.device		= &dev->ib_dev;
1256 	event.element.port_num	= port_num;
1257 	event.event		= type;
1258 
1259 	ib_dispatch_event(&event);
1260 }
1261 
1262 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1263 {
1264 	unsigned long flags;
1265 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1266 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1267 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1268 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1269 		queue_work(ctx->wq, &ctx->work);
1270 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1271 }
1272 
1273 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1274 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1275 				  int index)
1276 {
1277 	struct ib_sge sg_list;
1278 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1279 	int size;
1280 
1281 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1282 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1283 
1284 	sg_list.addr = tun_qp->ring[index].map;
1285 	sg_list.length = size;
1286 	sg_list.lkey = ctx->pd->local_dma_lkey;
1287 
1288 	recv_wr.next = NULL;
1289 	recv_wr.sg_list = &sg_list;
1290 	recv_wr.num_sge = 1;
1291 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1292 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1293 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1294 				      size, DMA_FROM_DEVICE);
1295 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1296 }
1297 
1298 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1299 		int slave, struct ib_sa_mad *sa_mad)
1300 {
1301 	int ret = 0;
1302 
1303 	/* dispatch to different sa handlers */
1304 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1305 	case IB_SA_ATTR_MC_MEMBER_REC:
1306 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1307 		break;
1308 	default:
1309 		break;
1310 	}
1311 	return ret;
1312 }
1313 
1314 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1315 {
1316 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1317 
1318 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1319 }
1320 
1321 
1322 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1323 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1324 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1325 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1326 {
1327 	struct ib_sge list;
1328 	struct ib_ud_wr wr;
1329 	struct ib_send_wr *bad_wr;
1330 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1331 	struct mlx4_ib_demux_pv_qp *sqp;
1332 	struct mlx4_mad_snd_buf *sqp_mad;
1333 	struct ib_ah *ah;
1334 	struct ib_qp *send_qp = NULL;
1335 	unsigned wire_tx_ix = 0;
1336 	int ret = 0;
1337 	u16 wire_pkey_ix;
1338 	int src_qpnum;
1339 	u8 sgid_index;
1340 
1341 
1342 	sqp_ctx = dev->sriov.sqps[port-1];
1343 
1344 	/* check if proxy qp created */
1345 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1346 		return -EAGAIN;
1347 
1348 	if (dest_qpt == IB_QPT_SMI) {
1349 		src_qpnum = 0;
1350 		sqp = &sqp_ctx->qp[0];
1351 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1352 	} else {
1353 		src_qpnum = 1;
1354 		sqp = &sqp_ctx->qp[1];
1355 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1356 	}
1357 
1358 	send_qp = sqp->qp;
1359 
1360 	/* create ah */
1361 	sgid_index = attr->grh.sgid_index;
1362 	attr->grh.sgid_index = 0;
1363 	ah = ib_create_ah(sqp_ctx->pd, attr);
1364 	if (IS_ERR(ah))
1365 		return -ENOMEM;
1366 	attr->grh.sgid_index = sgid_index;
1367 	to_mah(ah)->av.ib.gid_index = sgid_index;
1368 	/* get rid of force-loopback bit */
1369 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1370 	spin_lock(&sqp->tx_lock);
1371 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1372 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1373 		ret = -EAGAIN;
1374 	else
1375 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1376 	spin_unlock(&sqp->tx_lock);
1377 	if (ret)
1378 		goto out;
1379 
1380 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1381 	if (sqp->tx_ring[wire_tx_ix].ah)
1382 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1383 	sqp->tx_ring[wire_tx_ix].ah = ah;
1384 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1385 				   sqp->tx_ring[wire_tx_ix].buf.map,
1386 				   sizeof (struct mlx4_mad_snd_buf),
1387 				   DMA_TO_DEVICE);
1388 
1389 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1390 
1391 	ib_dma_sync_single_for_device(&dev->ib_dev,
1392 				      sqp->tx_ring[wire_tx_ix].buf.map,
1393 				      sizeof (struct mlx4_mad_snd_buf),
1394 				      DMA_TO_DEVICE);
1395 
1396 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1397 	list.length = sizeof (struct mlx4_mad_snd_buf);
1398 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1399 
1400 	wr.ah = ah;
1401 	wr.port_num = port;
1402 	wr.pkey_index = wire_pkey_ix;
1403 	wr.remote_qkey = qkey;
1404 	wr.remote_qpn = remote_qpn;
1405 	wr.wr.next = NULL;
1406 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1407 	wr.wr.sg_list = &list;
1408 	wr.wr.num_sge = 1;
1409 	wr.wr.opcode = IB_WR_SEND;
1410 	wr.wr.send_flags = IB_SEND_SIGNALED;
1411 	if (s_mac)
1412 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1413 	if (vlan_id < 0x1000)
1414 		vlan_id |= (attr->sl & 7) << 13;
1415 	to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1416 
1417 
1418 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1419 	if (!ret)
1420 		return 0;
1421 
1422 	spin_lock(&sqp->tx_lock);
1423 	sqp->tx_ix_tail++;
1424 	spin_unlock(&sqp->tx_lock);
1425 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1426 out:
1427 	ib_destroy_ah(ah);
1428 	return ret;
1429 }
1430 
1431 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1432 {
1433 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1434 		return slave;
1435 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1436 }
1437 
1438 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1439 				    struct ib_ah_attr *ah_attr)
1440 {
1441 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1442 		ah_attr->grh.sgid_index = slave;
1443 	else
1444 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1445 }
1446 
1447 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1448 {
1449 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1450 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1451 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1452 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1453 	struct mlx4_ib_ah ah;
1454 	struct ib_ah_attr ah_attr;
1455 	u8 *slave_id;
1456 	int slave;
1457 	int port;
1458 	u16 vlan_id;
1459 
1460 	/* Get slave that sent this packet */
1461 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1462 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1463 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1464 	    wc->src_qp & 0x4) {
1465 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1466 		return;
1467 	}
1468 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1469 	if (slave != ctx->slave) {
1470 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1471 			     "belongs to another slave\n", wc->src_qp);
1472 		return;
1473 	}
1474 
1475 	/* Map transaction ID */
1476 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1477 				   sizeof (struct mlx4_tunnel_mad),
1478 				   DMA_FROM_DEVICE);
1479 	switch (tunnel->mad.mad_hdr.method) {
1480 	case IB_MGMT_METHOD_SET:
1481 	case IB_MGMT_METHOD_GET:
1482 	case IB_MGMT_METHOD_REPORT:
1483 	case IB_SA_METHOD_GET_TABLE:
1484 	case IB_SA_METHOD_DELETE:
1485 	case IB_SA_METHOD_GET_MULTI:
1486 	case IB_SA_METHOD_GET_TRACE_TBL:
1487 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1488 		if (*slave_id) {
1489 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1490 				     "class:%d slave:%d\n", *slave_id,
1491 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1492 			return;
1493 		} else
1494 			*slave_id = slave;
1495 	default:
1496 		/* nothing */;
1497 	}
1498 
1499 	/* Class-specific handling */
1500 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1501 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1502 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1503 		if (slave != mlx4_master_func_num(dev->dev) &&
1504 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1505 			return;
1506 		break;
1507 	case IB_MGMT_CLASS_SUBN_ADM:
1508 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1509 			      (struct ib_sa_mad *) &tunnel->mad))
1510 			return;
1511 		break;
1512 	case IB_MGMT_CLASS_CM:
1513 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1514 			      (struct ib_mad *) &tunnel->mad))
1515 			return;
1516 		break;
1517 	case IB_MGMT_CLASS_DEVICE_MGMT:
1518 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1519 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1520 			return;
1521 		break;
1522 	default:
1523 		/* Drop unsupported classes for slaves in tunnel mode */
1524 		if (slave != mlx4_master_func_num(dev->dev)) {
1525 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1526 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1527 			return;
1528 		}
1529 	}
1530 
1531 	/* We are using standard ib_core services to send the mad, so generate a
1532 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1533 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1534 	ah.ibah.device = ctx->ib_dev;
1535 
1536 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1537 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1538 	if (port < 0)
1539 		return;
1540 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1541 
1542 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1543 	if (ah_attr.ah_flags & IB_AH_GRH)
1544 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1545 
1546 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1547 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1548 	/* if slave have default vlan use it */
1549 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1550 				    &vlan_id, &ah_attr.sl);
1551 
1552 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1553 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1554 			     IB_QPT_SMI : IB_QPT_GSI,
1555 			     be16_to_cpu(tunnel->hdr.pkey_index),
1556 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1557 			     be32_to_cpu(tunnel->hdr.qkey),
1558 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1559 }
1560 
1561 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1562 				 enum ib_qp_type qp_type, int is_tun)
1563 {
1564 	int i;
1565 	struct mlx4_ib_demux_pv_qp *tun_qp;
1566 	int rx_buf_size, tx_buf_size;
1567 
1568 	if (qp_type > IB_QPT_GSI)
1569 		return -EINVAL;
1570 
1571 	tun_qp = &ctx->qp[qp_type];
1572 
1573 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1574 			       GFP_KERNEL);
1575 	if (!tun_qp->ring)
1576 		return -ENOMEM;
1577 
1578 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1579 				  sizeof (struct mlx4_ib_tun_tx_buf),
1580 				  GFP_KERNEL);
1581 	if (!tun_qp->tx_ring) {
1582 		kfree(tun_qp->ring);
1583 		tun_qp->ring = NULL;
1584 		return -ENOMEM;
1585 	}
1586 
1587 	if (is_tun) {
1588 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1589 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1590 	} else {
1591 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1592 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1593 	}
1594 
1595 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1596 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1597 		if (!tun_qp->ring[i].addr)
1598 			goto err;
1599 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1600 							tun_qp->ring[i].addr,
1601 							rx_buf_size,
1602 							DMA_FROM_DEVICE);
1603 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1604 			kfree(tun_qp->ring[i].addr);
1605 			goto err;
1606 		}
1607 	}
1608 
1609 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1610 		tun_qp->tx_ring[i].buf.addr =
1611 			kmalloc(tx_buf_size, GFP_KERNEL);
1612 		if (!tun_qp->tx_ring[i].buf.addr)
1613 			goto tx_err;
1614 		tun_qp->tx_ring[i].buf.map =
1615 			ib_dma_map_single(ctx->ib_dev,
1616 					  tun_qp->tx_ring[i].buf.addr,
1617 					  tx_buf_size,
1618 					  DMA_TO_DEVICE);
1619 		if (ib_dma_mapping_error(ctx->ib_dev,
1620 					 tun_qp->tx_ring[i].buf.map)) {
1621 			kfree(tun_qp->tx_ring[i].buf.addr);
1622 			goto tx_err;
1623 		}
1624 		tun_qp->tx_ring[i].ah = NULL;
1625 	}
1626 	spin_lock_init(&tun_qp->tx_lock);
1627 	tun_qp->tx_ix_head = 0;
1628 	tun_qp->tx_ix_tail = 0;
1629 	tun_qp->proxy_qpt = qp_type;
1630 
1631 	return 0;
1632 
1633 tx_err:
1634 	while (i > 0) {
1635 		--i;
1636 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1637 				    tx_buf_size, DMA_TO_DEVICE);
1638 		kfree(tun_qp->tx_ring[i].buf.addr);
1639 	}
1640 	kfree(tun_qp->tx_ring);
1641 	tun_qp->tx_ring = NULL;
1642 	i = MLX4_NUM_TUNNEL_BUFS;
1643 err:
1644 	while (i > 0) {
1645 		--i;
1646 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1647 				    rx_buf_size, DMA_FROM_DEVICE);
1648 		kfree(tun_qp->ring[i].addr);
1649 	}
1650 	kfree(tun_qp->ring);
1651 	tun_qp->ring = NULL;
1652 	return -ENOMEM;
1653 }
1654 
1655 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1656 				     enum ib_qp_type qp_type, int is_tun)
1657 {
1658 	int i;
1659 	struct mlx4_ib_demux_pv_qp *tun_qp;
1660 	int rx_buf_size, tx_buf_size;
1661 
1662 	if (qp_type > IB_QPT_GSI)
1663 		return;
1664 
1665 	tun_qp = &ctx->qp[qp_type];
1666 	if (is_tun) {
1667 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1668 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1669 	} else {
1670 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1671 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1672 	}
1673 
1674 
1675 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1676 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1677 				    rx_buf_size, DMA_FROM_DEVICE);
1678 		kfree(tun_qp->ring[i].addr);
1679 	}
1680 
1681 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1682 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1683 				    tx_buf_size, DMA_TO_DEVICE);
1684 		kfree(tun_qp->tx_ring[i].buf.addr);
1685 		if (tun_qp->tx_ring[i].ah)
1686 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1687 	}
1688 	kfree(tun_qp->tx_ring);
1689 	kfree(tun_qp->ring);
1690 }
1691 
1692 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1693 {
1694 	struct mlx4_ib_demux_pv_ctx *ctx;
1695 	struct mlx4_ib_demux_pv_qp *tun_qp;
1696 	struct ib_wc wc;
1697 	int ret;
1698 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1699 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1700 
1701 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1702 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1703 		if (wc.status == IB_WC_SUCCESS) {
1704 			switch (wc.opcode) {
1705 			case IB_WC_RECV:
1706 				mlx4_ib_multiplex_mad(ctx, &wc);
1707 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1708 							     wc.wr_id &
1709 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1710 				if (ret)
1711 					pr_err("Failed reposting tunnel "
1712 					       "buf:%lld\n", (unsigned long long)wc.wr_id);
1713 				break;
1714 			case IB_WC_SEND:
1715 				pr_debug("received tunnel send completion:"
1716 					 "wrid=0x%llx, status=0x%x\n",
1717 					 (unsigned long long)wc.wr_id, wc.status);
1718 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1719 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1720 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1721 					= NULL;
1722 				spin_lock(&tun_qp->tx_lock);
1723 				tun_qp->tx_ix_tail++;
1724 				spin_unlock(&tun_qp->tx_lock);
1725 
1726 				break;
1727 			default:
1728 				break;
1729 			}
1730 		} else  {
1731 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1732 				 " status = %d, wrid = 0x%llx\n",
1733 				 ctx->slave, wc.status, (unsigned long long)wc.wr_id);
1734 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1735 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1736 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1737 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1738 					= NULL;
1739 				spin_lock(&tun_qp->tx_lock);
1740 				tun_qp->tx_ix_tail++;
1741 				spin_unlock(&tun_qp->tx_lock);
1742 			}
1743 		}
1744 	}
1745 }
1746 
1747 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1748 {
1749 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1750 
1751 	/* It's worse than that! He's dead, Jim! */
1752 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1753 	       event->event, sqp->port);
1754 }
1755 
1756 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1757 			    enum ib_qp_type qp_type, int create_tun)
1758 {
1759 	int i, ret;
1760 	struct mlx4_ib_demux_pv_qp *tun_qp;
1761 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1762 	struct ib_qp_attr attr;
1763 	int qp_attr_mask_INIT;
1764 
1765 	if (qp_type > IB_QPT_GSI)
1766 		return -EINVAL;
1767 
1768 	tun_qp = &ctx->qp[qp_type];
1769 
1770 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1771 	qp_init_attr.init_attr.send_cq = ctx->cq;
1772 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1773 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1774 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1775 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1776 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1777 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1778 	if (create_tun) {
1779 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1780 		qp_init_attr.init_attr.create_flags =
1781 		    (enum ib_qp_create_flags)MLX4_IB_SRIOV_TUNNEL_QP;
1782 		qp_init_attr.port = ctx->port;
1783 		qp_init_attr.slave = ctx->slave;
1784 		qp_init_attr.proxy_qp_type = qp_type;
1785 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1786 			   IB_QP_QKEY | IB_QP_PORT;
1787 	} else {
1788 		qp_init_attr.init_attr.qp_type = qp_type;
1789 		qp_init_attr.init_attr.create_flags =
1790 		    (enum ib_qp_create_flags)MLX4_IB_SRIOV_SQP;
1791 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1792 	}
1793 	qp_init_attr.init_attr.port_num = ctx->port;
1794 	qp_init_attr.init_attr.qp_context = ctx;
1795 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1796 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1797 	if (IS_ERR(tun_qp->qp)) {
1798 		ret = PTR_ERR(tun_qp->qp);
1799 		tun_qp->qp = NULL;
1800 		pr_err("Couldn't create %s QP (%d)\n",
1801 		       create_tun ? "tunnel" : "special", ret);
1802 		return ret;
1803 	}
1804 
1805 	memset(&attr, 0, sizeof attr);
1806 	attr.qp_state = IB_QPS_INIT;
1807 	ret = 0;
1808 	if (create_tun)
1809 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1810 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1811 					      &attr.pkey_index);
1812 	if (ret || !create_tun)
1813 		attr.pkey_index =
1814 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1815 	attr.qkey = IB_QP1_QKEY;
1816 	attr.port_num = ctx->port;
1817 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1818 	if (ret) {
1819 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1820 		       create_tun ? "tunnel" : "special", ret);
1821 		goto err_qp;
1822 	}
1823 	attr.qp_state = IB_QPS_RTR;
1824 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1825 	if (ret) {
1826 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1827 		       create_tun ? "tunnel" : "special", ret);
1828 		goto err_qp;
1829 	}
1830 	attr.qp_state = IB_QPS_RTS;
1831 	attr.sq_psn = 0;
1832 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1833 	if (ret) {
1834 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1835 		       create_tun ? "tunnel" : "special", ret);
1836 		goto err_qp;
1837 	}
1838 
1839 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1840 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1841 		if (ret) {
1842 			pr_err(" mlx4_ib_post_pv_buf error"
1843 			       " (err = %d, i = %d)\n", ret, i);
1844 			goto err_qp;
1845 		}
1846 	}
1847 	return 0;
1848 
1849 err_qp:
1850 	ib_destroy_qp(tun_qp->qp);
1851 	tun_qp->qp = NULL;
1852 	return ret;
1853 }
1854 
1855 /*
1856  * IB MAD completion callback for real SQPs
1857  */
1858 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1859 {
1860 	struct mlx4_ib_demux_pv_ctx *ctx;
1861 	struct mlx4_ib_demux_pv_qp *sqp;
1862 	struct ib_wc wc;
1863 	struct ib_grh *grh;
1864 	struct ib_mad *mad;
1865 
1866 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1867 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1868 
1869 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1870 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1871 		if (wc.status == IB_WC_SUCCESS) {
1872 			switch (wc.opcode) {
1873 			case IB_WC_SEND:
1874 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1875 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1876 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1877 					= NULL;
1878 				spin_lock(&sqp->tx_lock);
1879 				sqp->tx_ix_tail++;
1880 				spin_unlock(&sqp->tx_lock);
1881 				break;
1882 			case IB_WC_RECV:
1883 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1884 						(sqp->ring[wc.wr_id &
1885 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1886 				grh = &(((struct mlx4_mad_rcv_buf *)
1887 						(sqp->ring[wc.wr_id &
1888 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1889 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1890 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1891 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1892 					pr_err("Failed reposting SQP "
1893 					       "buf:%lld\n", (unsigned long long)wc.wr_id);
1894 				break;
1895 			default:
1896 				BUG_ON(1);
1897 				break;
1898 			}
1899 		} else  {
1900 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1901 				 " status = %d, wrid = 0x%llx\n",
1902 				 ctx->slave, wc.status, (unsigned long long)wc.wr_id);
1903 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1904 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1905 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1906 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1907 					= NULL;
1908 				spin_lock(&sqp->tx_lock);
1909 				sqp->tx_ix_tail++;
1910 				spin_unlock(&sqp->tx_lock);
1911 			}
1912 		}
1913 	}
1914 }
1915 
1916 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1917 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1918 {
1919 	struct mlx4_ib_demux_pv_ctx *ctx;
1920 
1921 	*ret_ctx = NULL;
1922 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1923 	if (!ctx) {
1924 		pr_err("failed allocating pv resource context "
1925 		       "for port %d, slave %d\n", port, slave);
1926 		return -ENOMEM;
1927 	}
1928 
1929 	ctx->ib_dev = &dev->ib_dev;
1930 	ctx->port = port;
1931 	ctx->slave = slave;
1932 	*ret_ctx = ctx;
1933 	return 0;
1934 }
1935 
1936 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1937 {
1938 	if (dev->sriov.demux[port - 1].tun[slave]) {
1939 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1940 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1941 	}
1942 }
1943 
1944 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1945 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1946 {
1947 	int ret, cq_size;
1948 	struct ib_cq_init_attr cq_attr = {};
1949 
1950 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1951 		return -EEXIST;
1952 
1953 	ctx->state = DEMUX_PV_STATE_STARTING;
1954 	/* have QP0 only if link layer is IB */
1955 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1956 	    IB_LINK_LAYER_INFINIBAND)
1957 		ctx->has_smi = 1;
1958 
1959 	if (ctx->has_smi) {
1960 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1961 		if (ret) {
1962 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1963 			goto err_out;
1964 		}
1965 	}
1966 
1967 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1968 	if (ret) {
1969 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1970 		goto err_out_qp0;
1971 	}
1972 
1973 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1974 	if (ctx->has_smi)
1975 		cq_size *= 2;
1976 
1977 	cq_attr.cqe = cq_size;
1978 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1979 			       NULL, ctx, &cq_attr);
1980 	if (IS_ERR(ctx->cq)) {
1981 		ret = PTR_ERR(ctx->cq);
1982 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1983 		goto err_buf;
1984 	}
1985 
1986 	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
1987 	if (IS_ERR(ctx->pd)) {
1988 		ret = PTR_ERR(ctx->pd);
1989 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1990 		goto err_cq;
1991 	}
1992 
1993 	if (ctx->has_smi) {
1994 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1995 		if (ret) {
1996 			pr_err("Couldn't create %s QP0 (%d)\n",
1997 			       create_tun ? "tunnel for" : "",  ret);
1998 			goto err_pd;
1999 		}
2000 	}
2001 
2002 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
2003 	if (ret) {
2004 		pr_err("Couldn't create %s QP1 (%d)\n",
2005 		       create_tun ? "tunnel for" : "",  ret);
2006 		goto err_qp0;
2007 	}
2008 
2009 	if (create_tun)
2010 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2011 	else
2012 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2013 
2014 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2015 
2016 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2017 	if (ret) {
2018 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2019 		goto err_wq;
2020 	}
2021 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2022 	return 0;
2023 
2024 err_wq:
2025 	ctx->wq = NULL;
2026 	ib_destroy_qp(ctx->qp[1].qp);
2027 	ctx->qp[1].qp = NULL;
2028 
2029 
2030 err_qp0:
2031 	if (ctx->has_smi)
2032 		ib_destroy_qp(ctx->qp[0].qp);
2033 	ctx->qp[0].qp = NULL;
2034 
2035 err_pd:
2036 	ib_dealloc_pd(ctx->pd);
2037 	ctx->pd = NULL;
2038 
2039 err_cq:
2040 	ib_destroy_cq(ctx->cq);
2041 	ctx->cq = NULL;
2042 
2043 err_buf:
2044 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2045 
2046 err_out_qp0:
2047 	if (ctx->has_smi)
2048 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2049 err_out:
2050 	ctx->state = DEMUX_PV_STATE_DOWN;
2051 	return ret;
2052 }
2053 
2054 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2055 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2056 {
2057 	if (!ctx)
2058 		return;
2059 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2060 		ctx->state = DEMUX_PV_STATE_DOWNING;
2061 		if (flush)
2062 			flush_workqueue(ctx->wq);
2063 		if (ctx->has_smi) {
2064 			ib_destroy_qp(ctx->qp[0].qp);
2065 			ctx->qp[0].qp = NULL;
2066 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2067 		}
2068 		ib_destroy_qp(ctx->qp[1].qp);
2069 		ctx->qp[1].qp = NULL;
2070 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2071 		ib_dealloc_pd(ctx->pd);
2072 		ctx->pd = NULL;
2073 		ib_destroy_cq(ctx->cq);
2074 		ctx->cq = NULL;
2075 		ctx->state = DEMUX_PV_STATE_DOWN;
2076 	}
2077 }
2078 
2079 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2080 				  int port, int do_init)
2081 {
2082 	int ret = 0;
2083 
2084 	if (!do_init) {
2085 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2086 		/* for master, destroy real sqp resources */
2087 		if (slave == mlx4_master_func_num(dev->dev))
2088 			destroy_pv_resources(dev, slave, port,
2089 					     dev->sriov.sqps[port - 1], 1);
2090 		/* destroy the tunnel qp resources */
2091 		destroy_pv_resources(dev, slave, port,
2092 				     dev->sriov.demux[port - 1].tun[slave], 1);
2093 		return 0;
2094 	}
2095 
2096 	/* create the tunnel qp resources */
2097 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2098 				  dev->sriov.demux[port - 1].tun[slave]);
2099 
2100 	/* for master, create the real sqp resources */
2101 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2102 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2103 					  dev->sriov.sqps[port - 1]);
2104 	return ret;
2105 }
2106 
2107 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2108 {
2109 	struct mlx4_ib_demux_work *dmxw;
2110 
2111 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2112 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2113 			       dmxw->do_init);
2114 	kfree(dmxw);
2115 	return;
2116 }
2117 
2118 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2119 				       struct mlx4_ib_demux_ctx *ctx,
2120 				       int port)
2121 {
2122 	char name[12];
2123 	int ret = 0;
2124 	int i;
2125 
2126 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2127 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2128 	if (!ctx->tun)
2129 		return -ENOMEM;
2130 
2131 	ctx->dev = dev;
2132 	ctx->port = port;
2133 	ctx->ib_dev = &dev->ib_dev;
2134 
2135 	for (i = 0;
2136 	     i < min(dev->dev->caps.sqp_demux,
2137 	     (u16)(dev->dev->persist->num_vfs + 1));
2138 	     i++) {
2139 		struct mlx4_active_ports actv_ports =
2140 			mlx4_get_active_ports(dev->dev, i);
2141 
2142 		if (!test_bit(port - 1, actv_ports.ports))
2143 			continue;
2144 
2145 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2146 		if (ret) {
2147 			ret = -ENOMEM;
2148 			goto err_mcg;
2149 		}
2150 	}
2151 
2152 	ret = mlx4_ib_mcg_port_init(ctx);
2153 	if (ret) {
2154 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2155 		goto err_mcg;
2156 	}
2157 
2158 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2159 	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2160 	if (!ctx->wq) {
2161 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2162 		ret = -ENOMEM;
2163 		goto err_wq;
2164 	}
2165 
2166 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2167 	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2168 	if (!ctx->ud_wq) {
2169 		pr_err("Failed to create up/down WQ for port %d\n", port);
2170 		ret = -ENOMEM;
2171 		goto err_udwq;
2172 	}
2173 
2174 	return 0;
2175 
2176 err_udwq:
2177 	destroy_workqueue(ctx->wq);
2178 	ctx->wq = NULL;
2179 
2180 err_wq:
2181 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2182 err_mcg:
2183 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2184 		free_pv_object(dev, i, port);
2185 	kfree(ctx->tun);
2186 	ctx->tun = NULL;
2187 	return ret;
2188 }
2189 
2190 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2191 {
2192 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2193 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2194 		flush_workqueue(sqp_ctx->wq);
2195 		if (sqp_ctx->has_smi) {
2196 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2197 			sqp_ctx->qp[0].qp = NULL;
2198 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2199 		}
2200 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2201 		sqp_ctx->qp[1].qp = NULL;
2202 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2203 		ib_dealloc_pd(sqp_ctx->pd);
2204 		sqp_ctx->pd = NULL;
2205 		ib_destroy_cq(sqp_ctx->cq);
2206 		sqp_ctx->cq = NULL;
2207 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2208 	}
2209 }
2210 
2211 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2212 {
2213 	int i;
2214 	if (ctx) {
2215 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2216 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2217 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2218 			if (!ctx->tun[i])
2219 				continue;
2220 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2221 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2222 		}
2223 		flush_workqueue(ctx->wq);
2224 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2225 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2226 			free_pv_object(dev, i, ctx->port);
2227 		}
2228 		kfree(ctx->tun);
2229 		destroy_workqueue(ctx->ud_wq);
2230 		destroy_workqueue(ctx->wq);
2231 	}
2232 }
2233 
2234 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2235 {
2236 	int i;
2237 
2238 	if (!mlx4_is_master(dev->dev))
2239 		return;
2240 	/* initialize or tear down tunnel QPs for the master */
2241 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2242 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2243 	return;
2244 }
2245 
2246 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2247 {
2248 	int i = 0;
2249 	int err;
2250 
2251 	if (!mlx4_is_mfunc(dev->dev))
2252 		return 0;
2253 
2254 	dev->sriov.is_going_down = 0;
2255 	spin_lock_init(&dev->sriov.going_down_lock);
2256 	mlx4_ib_cm_paravirt_init(dev);
2257 
2258 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2259 
2260 	if (mlx4_is_slave(dev->dev)) {
2261 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2262 		return 0;
2263 	}
2264 
2265 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2266 		if (i == mlx4_master_func_num(dev->dev))
2267 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2268 		else
2269 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2270 	}
2271 
2272 	err = mlx4_ib_init_alias_guid_service(dev);
2273 	if (err) {
2274 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2275 		goto paravirt_err;
2276 	}
2277 	err = mlx4_ib_device_register_sysfs(dev);
2278 	if (err) {
2279 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2280 		goto sysfs_err;
2281 	}
2282 
2283 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2284 		     dev->dev->caps.sqp_demux);
2285 	for (i = 0; i < dev->num_ports; i++) {
2286 		union ib_gid gid;
2287 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2288 		if (err)
2289 			goto demux_err;
2290 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2291 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2292 			     be64_to_cpu(gid.global.subnet_prefix));
2293 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2294 				      &dev->sriov.sqps[i]);
2295 		if (err)
2296 			goto demux_err;
2297 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2298 		if (err)
2299 			goto free_pv;
2300 	}
2301 	mlx4_ib_master_tunnels(dev, 1);
2302 	return 0;
2303 
2304 free_pv:
2305 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2306 demux_err:
2307 	while (--i >= 0) {
2308 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2309 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2310 	}
2311 	mlx4_ib_device_unregister_sysfs(dev);
2312 
2313 sysfs_err:
2314 	mlx4_ib_destroy_alias_guid_service(dev);
2315 
2316 paravirt_err:
2317 	mlx4_ib_cm_paravirt_clean(dev, -1);
2318 
2319 	return err;
2320 }
2321 
2322 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2323 {
2324 	int i;
2325 	unsigned long flags;
2326 
2327 	if (!mlx4_is_mfunc(dev->dev))
2328 		return;
2329 
2330 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2331 	dev->sriov.is_going_down = 1;
2332 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2333 	if (mlx4_is_master(dev->dev)) {
2334 		for (i = 0; i < dev->num_ports; i++) {
2335 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2336 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2337 			kfree(dev->sriov.sqps[i]);
2338 			dev->sriov.sqps[i] = NULL;
2339 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2340 		}
2341 
2342 		mlx4_ib_cm_paravirt_clean(dev, -1);
2343 		mlx4_ib_destroy_alias_guid_service(dev);
2344 		mlx4_ib_device_unregister_sysfs(dev);
2345 	}
2346 }
2347