xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision bc5aa3a0)
1 /*
2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <rdma/ib_mad.h>
34 #include <rdma/ib_smi.h>
35 #include <rdma/ib_sa.h>
36 #include <rdma/ib_cache.h>
37 
38 #include <linux/random.h>
39 #include <linux/mlx4/cmd.h>
40 #include <linux/gfp.h>
41 #include <rdma/ib_pma.h>
42 
43 #include <linux/mlx4/driver.h>
44 #include "mlx4_ib.h"
45 
46 enum {
47 	MLX4_IB_VENDOR_CLASS1 = 0x9,
48 	MLX4_IB_VENDOR_CLASS2 = 0xa
49 };
50 
51 #define MLX4_TUN_SEND_WRID_SHIFT 34
52 #define MLX4_TUN_QPN_SHIFT 32
53 #define MLX4_TUN_WRID_RECV (((u64) 1) << MLX4_TUN_SEND_WRID_SHIFT)
54 #define MLX4_TUN_SET_WRID_QPN(a) (((u64) ((a) & 0x3)) << MLX4_TUN_QPN_SHIFT)
55 
56 #define MLX4_TUN_IS_RECV(a)  (((a) >>  MLX4_TUN_SEND_WRID_SHIFT) & 0x1)
57 #define MLX4_TUN_WRID_QPN(a) (((a) >> MLX4_TUN_QPN_SHIFT) & 0x3)
58 
59  /* Port mgmt change event handling */
60 
61 #define GET_BLK_PTR_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.block_ptr)
62 #define GET_MASK_FROM_EQE(eqe) be32_to_cpu(eqe->event.port_mgmt_change.params.tbl_change_info.tbl_entries_mask)
63 #define NUM_IDX_IN_PKEY_TBL_BLK 32
64 #define GUID_TBL_ENTRY_SIZE 8	   /* size in bytes */
65 #define GUID_TBL_BLK_NUM_ENTRIES 8
66 #define GUID_TBL_BLK_SIZE (GUID_TBL_ENTRY_SIZE * GUID_TBL_BLK_NUM_ENTRIES)
67 
68 struct mlx4_mad_rcv_buf {
69 	struct ib_grh grh;
70 	u8 payload[256];
71 } __packed;
72 
73 struct mlx4_mad_snd_buf {
74 	u8 payload[256];
75 } __packed;
76 
77 struct mlx4_tunnel_mad {
78 	struct ib_grh grh;
79 	struct mlx4_ib_tunnel_header hdr;
80 	struct ib_mad mad;
81 } __packed;
82 
83 struct mlx4_rcv_tunnel_mad {
84 	struct mlx4_rcv_tunnel_hdr hdr;
85 	struct ib_grh grh;
86 	struct ib_mad mad;
87 } __packed;
88 
89 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num);
90 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num);
91 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
92 				int block, u32 change_bitmap);
93 
94 __be64 mlx4_ib_gen_node_guid(void)
95 {
96 #define NODE_GUID_HI	((u64) (((u64)IB_OPENIB_OUI) << 40))
97 	return cpu_to_be64(NODE_GUID_HI | prandom_u32());
98 }
99 
100 __be64 mlx4_ib_get_new_demux_tid(struct mlx4_ib_demux_ctx *ctx)
101 {
102 	return cpu_to_be64(atomic_inc_return(&ctx->tid)) |
103 		cpu_to_be64(0xff00000000000000LL);
104 }
105 
106 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int mad_ifc_flags,
107 		 int port, const struct ib_wc *in_wc,
108 		 const struct ib_grh *in_grh,
109 		 const void *in_mad, void *response_mad)
110 {
111 	struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
112 	void *inbox;
113 	int err;
114 	u32 in_modifier = port;
115 	u8 op_modifier = 0;
116 
117 	inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
118 	if (IS_ERR(inmailbox))
119 		return PTR_ERR(inmailbox);
120 	inbox = inmailbox->buf;
121 
122 	outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
123 	if (IS_ERR(outmailbox)) {
124 		mlx4_free_cmd_mailbox(dev->dev, inmailbox);
125 		return PTR_ERR(outmailbox);
126 	}
127 
128 	memcpy(inbox, in_mad, 256);
129 
130 	/*
131 	 * Key check traps can't be generated unless we have in_wc to
132 	 * tell us where to send the trap.
133 	 */
134 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_MKEY) || !in_wc)
135 		op_modifier |= 0x1;
136 	if ((mad_ifc_flags & MLX4_MAD_IFC_IGNORE_BKEY) || !in_wc)
137 		op_modifier |= 0x2;
138 	if (mlx4_is_mfunc(dev->dev) &&
139 	    (mad_ifc_flags & MLX4_MAD_IFC_NET_VIEW || in_wc))
140 		op_modifier |= 0x8;
141 
142 	if (in_wc) {
143 		struct {
144 			__be32		my_qpn;
145 			u32		reserved1;
146 			__be32		rqpn;
147 			u8		sl;
148 			u8		g_path;
149 			u16		reserved2[2];
150 			__be16		pkey;
151 			u32		reserved3[11];
152 			u8		grh[40];
153 		} *ext_info;
154 
155 		memset(inbox + 256, 0, 256);
156 		ext_info = inbox + 256;
157 
158 		ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
159 		ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
160 		ext_info->sl     = in_wc->sl << 4;
161 		ext_info->g_path = in_wc->dlid_path_bits |
162 			(in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
163 		ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
164 
165 		if (in_grh)
166 			memcpy(ext_info->grh, in_grh, 40);
167 
168 		op_modifier |= 0x4;
169 
170 		in_modifier |= in_wc->slid << 16;
171 	}
172 
173 	err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma, in_modifier,
174 			   mlx4_is_master(dev->dev) ? (op_modifier & ~0x8) : op_modifier,
175 			   MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C,
176 			   (op_modifier & 0x8) ? MLX4_CMD_NATIVE : MLX4_CMD_WRAPPED);
177 
178 	if (!err)
179 		memcpy(response_mad, outmailbox->buf, 256);
180 
181 	mlx4_free_cmd_mailbox(dev->dev, inmailbox);
182 	mlx4_free_cmd_mailbox(dev->dev, outmailbox);
183 
184 	return err;
185 }
186 
187 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
188 {
189 	struct ib_ah *new_ah;
190 	struct ib_ah_attr ah_attr;
191 	unsigned long flags;
192 
193 	if (!dev->send_agent[port_num - 1][0])
194 		return;
195 
196 	memset(&ah_attr, 0, sizeof ah_attr);
197 	ah_attr.dlid     = lid;
198 	ah_attr.sl       = sl;
199 	ah_attr.port_num = port_num;
200 
201 	new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
202 			      &ah_attr);
203 	if (IS_ERR(new_ah))
204 		return;
205 
206 	spin_lock_irqsave(&dev->sm_lock, flags);
207 	if (dev->sm_ah[port_num - 1])
208 		ib_destroy_ah(dev->sm_ah[port_num - 1]);
209 	dev->sm_ah[port_num - 1] = new_ah;
210 	spin_unlock_irqrestore(&dev->sm_lock, flags);
211 }
212 
213 /*
214  * Snoop SM MADs for port info, GUID info, and  P_Key table sets, so we can
215  * synthesize LID change, Client-Rereg, GID change, and P_Key change events.
216  */
217 static void smp_snoop(struct ib_device *ibdev, u8 port_num, const struct ib_mad *mad,
218 		      u16 prev_lid)
219 {
220 	struct ib_port_info *pinfo;
221 	u16 lid;
222 	__be16 *base;
223 	u32 bn, pkey_change_bitmap;
224 	int i;
225 
226 
227 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
228 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
229 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
230 	    mad->mad_hdr.method == IB_MGMT_METHOD_SET)
231 		switch (mad->mad_hdr.attr_id) {
232 		case IB_SMP_ATTR_PORT_INFO:
233 			pinfo = (struct ib_port_info *) ((struct ib_smp *) mad)->data;
234 			lid = be16_to_cpu(pinfo->lid);
235 
236 			update_sm_ah(dev, port_num,
237 				     be16_to_cpu(pinfo->sm_lid),
238 				     pinfo->neighbormtu_mastersmsl & 0xf);
239 
240 			if (pinfo->clientrereg_resv_subnetto & 0x80)
241 				handle_client_rereg_event(dev, port_num);
242 
243 			if (prev_lid != lid)
244 				handle_lid_change_event(dev, port_num);
245 			break;
246 
247 		case IB_SMP_ATTR_PKEY_TABLE:
248 			if (!mlx4_is_mfunc(dev->dev)) {
249 				mlx4_ib_dispatch_event(dev, port_num,
250 						       IB_EVENT_PKEY_CHANGE);
251 				break;
252 			}
253 
254 			/* at this point, we are running in the master.
255 			 * Slaves do not receive SMPs.
256 			 */
257 			bn  = be32_to_cpu(((struct ib_smp *)mad)->attr_mod) & 0xFFFF;
258 			base = (__be16 *) &(((struct ib_smp *)mad)->data[0]);
259 			pkey_change_bitmap = 0;
260 			for (i = 0; i < 32; i++) {
261 				pr_debug("PKEY[%d] = x%x\n",
262 					 i + bn*32, be16_to_cpu(base[i]));
263 				if (be16_to_cpu(base[i]) !=
264 				    dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32]) {
265 					pkey_change_bitmap |= (1 << i);
266 					dev->pkeys.phys_pkey_cache[port_num - 1][i + bn*32] =
267 						be16_to_cpu(base[i]);
268 				}
269 			}
270 			pr_debug("PKEY Change event: port=%d, "
271 				 "block=0x%x, change_bitmap=0x%x\n",
272 				 port_num, bn, pkey_change_bitmap);
273 
274 			if (pkey_change_bitmap) {
275 				mlx4_ib_dispatch_event(dev, port_num,
276 						       IB_EVENT_PKEY_CHANGE);
277 				if (!dev->sriov.is_going_down)
278 					__propagate_pkey_ev(dev, port_num, bn,
279 							    pkey_change_bitmap);
280 			}
281 			break;
282 
283 		case IB_SMP_ATTR_GUID_INFO:
284 			/* paravirtualized master's guid is guid 0 -- does not change */
285 			if (!mlx4_is_master(dev->dev))
286 				mlx4_ib_dispatch_event(dev, port_num,
287 						       IB_EVENT_GID_CHANGE);
288 			/*if master, notify relevant slaves*/
289 			if (mlx4_is_master(dev->dev) &&
290 			    !dev->sriov.is_going_down) {
291 				bn = be32_to_cpu(((struct ib_smp *)mad)->attr_mod);
292 				mlx4_ib_update_cache_on_guid_change(dev, bn, port_num,
293 								    (u8 *)(&((struct ib_smp *)mad)->data));
294 				mlx4_ib_notify_slaves_on_guid_change(dev, bn, port_num,
295 								     (u8 *)(&((struct ib_smp *)mad)->data));
296 			}
297 			break;
298 
299 		default:
300 			break;
301 		}
302 }
303 
304 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
305 				int block, u32 change_bitmap)
306 {
307 	int i, ix, slave, err;
308 	int have_event = 0;
309 
310 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
311 		if (slave == mlx4_master_func_num(dev->dev))
312 			continue;
313 		if (!mlx4_is_slave_active(dev->dev, slave))
314 			continue;
315 
316 		have_event = 0;
317 		for (i = 0; i < 32; i++) {
318 			if (!(change_bitmap & (1 << i)))
319 				continue;
320 			for (ix = 0;
321 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
322 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
323 				    [ix] == i + 32 * block) {
324 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
325 					pr_debug("propagate_pkey_ev: slave %d,"
326 						 " port %d, ix %d (%d)\n",
327 						 slave, port_num, ix, err);
328 					have_event = 1;
329 					break;
330 				}
331 			}
332 			if (have_event)
333 				break;
334 		}
335 	}
336 }
337 
338 static void node_desc_override(struct ib_device *dev,
339 			       struct ib_mad *mad)
340 {
341 	unsigned long flags;
342 
343 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
344 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
345 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
346 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
347 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
348 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
349 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
350 	}
351 }
352 
353 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
354 {
355 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
356 	struct ib_mad_send_buf *send_buf;
357 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
358 	int ret;
359 	unsigned long flags;
360 
361 	if (agent) {
362 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
363 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
364 					      IB_MGMT_BASE_VERSION);
365 		if (IS_ERR(send_buf))
366 			return;
367 		/*
368 		 * We rely here on the fact that MLX QPs don't use the
369 		 * address handle after the send is posted (this is
370 		 * wrong following the IB spec strictly, but we know
371 		 * it's OK for our devices).
372 		 */
373 		spin_lock_irqsave(&dev->sm_lock, flags);
374 		memcpy(send_buf->mad, mad, sizeof *mad);
375 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
376 			ret = ib_post_send_mad(send_buf, NULL);
377 		else
378 			ret = -EINVAL;
379 		spin_unlock_irqrestore(&dev->sm_lock, flags);
380 
381 		if (ret)
382 			ib_free_send_mad(send_buf);
383 	}
384 }
385 
386 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
387 							     struct ib_sa_mad *sa_mad)
388 {
389 	int ret = 0;
390 
391 	/* dispatch to different sa handlers */
392 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
393 	case IB_SA_ATTR_MC_MEMBER_REC:
394 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
395 		break;
396 	default:
397 		break;
398 	}
399 	return ret;
400 }
401 
402 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
403 {
404 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
405 	int i;
406 
407 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
408 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
409 			return i;
410 	}
411 	return -1;
412 }
413 
414 
415 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
416 				   u8 port, u16 pkey, u16 *ix)
417 {
418 	int i, ret;
419 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
420 	u16 slot_pkey;
421 
422 	if (slave == mlx4_master_func_num(dev->dev))
423 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
424 
425 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
426 
427 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
428 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
429 			continue;
430 
431 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
432 
433 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
434 		if (ret)
435 			continue;
436 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
437 			if (slot_pkey & 0x8000) {
438 				*ix = (u16) pkey_ix;
439 				return 0;
440 			} else {
441 				/* take first partial pkey index found */
442 				if (partial_ix == 0xFF)
443 					partial_ix = pkey_ix;
444 			}
445 		}
446 	}
447 
448 	if (partial_ix < 0xFF) {
449 		*ix = (u16) partial_ix;
450 		return 0;
451 	}
452 
453 	return -EINVAL;
454 }
455 
456 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
457 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
458 			  struct ib_grh *grh, struct ib_mad *mad)
459 {
460 	struct ib_sge list;
461 	struct ib_ud_wr wr;
462 	struct ib_send_wr *bad_wr;
463 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
464 	struct mlx4_ib_demux_pv_qp *tun_qp;
465 	struct mlx4_rcv_tunnel_mad *tun_mad;
466 	struct ib_ah_attr attr;
467 	struct ib_ah *ah;
468 	struct ib_qp *src_qp = NULL;
469 	unsigned tun_tx_ix = 0;
470 	int dqpn;
471 	int ret = 0;
472 	u16 tun_pkey_ix;
473 	u16 cached_pkey;
474 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
475 
476 	if (dest_qpt > IB_QPT_GSI)
477 		return -EINVAL;
478 
479 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
480 
481 	/* check if proxy qp created */
482 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
483 		return -EAGAIN;
484 
485 	if (!dest_qpt)
486 		tun_qp = &tun_ctx->qp[0];
487 	else
488 		tun_qp = &tun_ctx->qp[1];
489 
490 	/* compute P_Key index to put in tunnel header for slave */
491 	if (dest_qpt) {
492 		u16 pkey_ix;
493 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
494 		if (ret)
495 			return -EINVAL;
496 
497 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
498 		if (ret)
499 			return -EINVAL;
500 		tun_pkey_ix = pkey_ix;
501 	} else
502 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
503 
504 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
505 
506 	/* get tunnel tx data buf for slave */
507 	src_qp = tun_qp->qp;
508 
509 	/* create ah. Just need an empty one with the port num for the post send.
510 	 * The driver will set the force loopback bit in post_send */
511 	memset(&attr, 0, sizeof attr);
512 	attr.port_num = port;
513 	if (is_eth) {
514 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
515 		attr.ah_flags = IB_AH_GRH;
516 	}
517 	ah = ib_create_ah(tun_ctx->pd, &attr);
518 	if (IS_ERR(ah))
519 		return -ENOMEM;
520 
521 	/* allocate tunnel tx buf after pass failure returns */
522 	spin_lock(&tun_qp->tx_lock);
523 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
524 	    (MLX4_NUM_TUNNEL_BUFS - 1))
525 		ret = -EAGAIN;
526 	else
527 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
528 	spin_unlock(&tun_qp->tx_lock);
529 	if (ret)
530 		goto end;
531 
532 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
533 	if (tun_qp->tx_ring[tun_tx_ix].ah)
534 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
535 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
536 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
537 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
538 				   sizeof (struct mlx4_rcv_tunnel_mad),
539 				   DMA_TO_DEVICE);
540 
541 	/* copy over to tunnel buffer */
542 	if (grh)
543 		memcpy(&tun_mad->grh, grh, sizeof *grh);
544 	memcpy(&tun_mad->mad, mad, sizeof *mad);
545 
546 	/* adjust tunnel data */
547 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
548 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
549 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
550 
551 	if (is_eth) {
552 		u16 vlan = 0;
553 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
554 						NULL)) {
555 			/* VST mode */
556 			if (vlan != wc->vlan_id)
557 				/* Packet vlan is not the VST-assigned vlan.
558 				 * Drop the packet.
559 				 */
560 				goto out;
561 			 else
562 				/* Remove the vlan tag before forwarding
563 				 * the packet to the VF.
564 				 */
565 				vlan = 0xffff;
566 		} else {
567 			vlan = wc->vlan_id;
568 		}
569 
570 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
571 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
572 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
573 	} else {
574 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
575 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
576 	}
577 
578 	ib_dma_sync_single_for_device(&dev->ib_dev,
579 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
580 				      sizeof (struct mlx4_rcv_tunnel_mad),
581 				      DMA_TO_DEVICE);
582 
583 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
584 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
585 	list.lkey = tun_ctx->pd->local_dma_lkey;
586 
587 	wr.ah = ah;
588 	wr.port_num = port;
589 	wr.remote_qkey = IB_QP_SET_QKEY;
590 	wr.remote_qpn = dqpn;
591 	wr.wr.next = NULL;
592 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
593 	wr.wr.sg_list = &list;
594 	wr.wr.num_sge = 1;
595 	wr.wr.opcode = IB_WR_SEND;
596 	wr.wr.send_flags = IB_SEND_SIGNALED;
597 
598 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
599 	if (!ret)
600 		return 0;
601  out:
602 	spin_lock(&tun_qp->tx_lock);
603 	tun_qp->tx_ix_tail++;
604 	spin_unlock(&tun_qp->tx_lock);
605 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
606 end:
607 	ib_destroy_ah(ah);
608 	return ret;
609 }
610 
611 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
612 			struct ib_wc *wc, struct ib_grh *grh,
613 			struct ib_mad *mad)
614 {
615 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
616 	int err, other_port;
617 	int slave = -1;
618 	u8 *slave_id;
619 	int is_eth = 0;
620 
621 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
622 		is_eth = 0;
623 	else
624 		is_eth = 1;
625 
626 	if (is_eth) {
627 		if (!(wc->wc_flags & IB_WC_GRH)) {
628 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
629 			return -EINVAL;
630 		}
631 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
632 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
633 			return -EINVAL;
634 		}
635 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave);
636 		if (err && mlx4_is_mf_bonded(dev->dev)) {
637 			other_port = (port == 1) ? 2 : 1;
638 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, grh->dgid.raw, &slave);
639 			if (!err) {
640 				port = other_port;
641 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
642 					 slave, grh->dgid.raw, port, other_port);
643 			}
644 		}
645 		if (err) {
646 			mlx4_ib_warn(ibdev, "failed matching grh\n");
647 			return -ENOENT;
648 		}
649 		if (slave >= dev->dev->caps.sqp_demux) {
650 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
651 				     slave, dev->dev->caps.sqp_demux);
652 			return -ENOENT;
653 		}
654 
655 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
656 			return 0;
657 
658 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
659 		if (err)
660 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
661 				 slave, err);
662 		return 0;
663 	}
664 
665 	/* Initially assume that this mad is for us */
666 	slave = mlx4_master_func_num(dev->dev);
667 
668 	/* See if the slave id is encoded in a response mad */
669 	if (mad->mad_hdr.method & 0x80) {
670 		slave_id = (u8 *) &mad->mad_hdr.tid;
671 		slave = *slave_id;
672 		if (slave != 255) /*255 indicates the dom0*/
673 			*slave_id = 0; /* remap tid */
674 	}
675 
676 	/* If a grh is present, we demux according to it */
677 	if (wc->wc_flags & IB_WC_GRH) {
678 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
679 		if (slave < 0) {
680 			mlx4_ib_warn(ibdev, "failed matching grh\n");
681 			return -ENOENT;
682 		}
683 	}
684 	/* Class-specific handling */
685 	switch (mad->mad_hdr.mgmt_class) {
686 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
687 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
688 		/* 255 indicates the dom0 */
689 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
690 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
691 				return -EPERM;
692 			/* for a VF. drop unsolicited MADs */
693 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
694 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
695 					     slave, mad->mad_hdr.mgmt_class,
696 					     mad->mad_hdr.method);
697 				return -EINVAL;
698 			}
699 		}
700 		break;
701 	case IB_MGMT_CLASS_SUBN_ADM:
702 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
703 					     (struct ib_sa_mad *) mad))
704 			return 0;
705 		break;
706 	case IB_MGMT_CLASS_CM:
707 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
708 			return 0;
709 		break;
710 	case IB_MGMT_CLASS_DEVICE_MGMT:
711 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
712 			return 0;
713 		break;
714 	default:
715 		/* Drop unsupported classes for slaves in tunnel mode */
716 		if (slave != mlx4_master_func_num(dev->dev)) {
717 			pr_debug("dropping unsupported ingress mad from class:%d "
718 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
719 			return 0;
720 		}
721 	}
722 	/*make sure that no slave==255 was not handled yet.*/
723 	if (slave >= dev->dev->caps.sqp_demux) {
724 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
725 			     slave, dev->dev->caps.sqp_demux);
726 		return -ENOENT;
727 	}
728 
729 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
730 	if (err)
731 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
732 			 slave, err);
733 	return 0;
734 }
735 
736 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
737 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
738 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
739 {
740 	u16 slid, prev_lid = 0;
741 	int err;
742 	struct ib_port_attr pattr;
743 
744 	if (in_wc && in_wc->qp->qp_num) {
745 		pr_debug("received MAD: slid:%d sqpn:%d "
746 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
747 			in_wc->slid, in_wc->src_qp,
748 			in_wc->dlid_path_bits,
749 			in_wc->qp->qp_num,
750 			in_wc->wc_flags,
751 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
752 			be16_to_cpu(in_mad->mad_hdr.attr_id));
753 		if (in_wc->wc_flags & IB_WC_GRH) {
754 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
755 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
756 				 be64_to_cpu(in_grh->sgid.global.interface_id));
757 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
758 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
759 				 be64_to_cpu(in_grh->dgid.global.interface_id));
760 		}
761 	}
762 
763 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
764 
765 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
766 		forward_trap(to_mdev(ibdev), port_num, in_mad);
767 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
768 	}
769 
770 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
771 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
772 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
773 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
774 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
775 			return IB_MAD_RESULT_SUCCESS;
776 
777 		/*
778 		 * Don't process SMInfo queries -- the SMA can't handle them.
779 		 */
780 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
781 			return IB_MAD_RESULT_SUCCESS;
782 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
783 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
784 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
785 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
786 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
787 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
788 			return IB_MAD_RESULT_SUCCESS;
789 	} else
790 		return IB_MAD_RESULT_SUCCESS;
791 
792 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
793 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
794 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
795 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
796 	    !ib_query_port(ibdev, port_num, &pattr))
797 		prev_lid = pattr.lid;
798 
799 	err = mlx4_MAD_IFC(to_mdev(ibdev),
800 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
801 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
802 			   MLX4_MAD_IFC_NET_VIEW,
803 			   port_num, in_wc, in_grh, in_mad, out_mad);
804 	if (err)
805 		return IB_MAD_RESULT_FAILURE;
806 
807 	if (!out_mad->mad_hdr.status) {
808 		if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_PORT_MNG_CHG_EV))
809 			smp_snoop(ibdev, port_num, in_mad, prev_lid);
810 		/* slaves get node desc from FW */
811 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
812 			node_desc_override(ibdev, out_mad);
813 	}
814 
815 	/* set return bit in status of directed route responses */
816 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
817 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
818 
819 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
820 		/* no response for trap repress */
821 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
822 
823 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
824 }
825 
826 static void edit_counter(struct mlx4_counter *cnt, void *counters,
827 			 __be16 attr_id)
828 {
829 	switch (attr_id) {
830 	case IB_PMA_PORT_COUNTERS:
831 	{
832 		struct ib_pma_portcounters *pma_cnt =
833 			(struct ib_pma_portcounters *)counters;
834 
835 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
836 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
837 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
838 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
839 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
840 				     be64_to_cpu(cnt->tx_frames));
841 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
842 				     be64_to_cpu(cnt->rx_frames));
843 		break;
844 	}
845 	case IB_PMA_PORT_COUNTERS_EXT:
846 	{
847 		struct ib_pma_portcounters_ext *pma_cnt_ext =
848 			(struct ib_pma_portcounters_ext *)counters;
849 
850 		pma_cnt_ext->port_xmit_data =
851 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
852 		pma_cnt_ext->port_rcv_data =
853 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
854 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
855 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
856 		break;
857 	}
858 	}
859 }
860 
861 static int iboe_process_mad_port_info(void *out_mad)
862 {
863 	struct ib_class_port_info cpi = {};
864 
865 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
866 	memcpy(out_mad, &cpi, sizeof(cpi));
867 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
868 }
869 
870 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
871 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
872 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
873 {
874 	struct mlx4_counter counter_stats;
875 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
876 	struct counter_index *tmp_counter;
877 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
878 
879 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
880 		return -EINVAL;
881 
882 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
883 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
884 
885 	memset(&counter_stats, 0, sizeof(counter_stats));
886 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
887 	list_for_each_entry(tmp_counter,
888 			    &dev->counters_table[port_num - 1].counters_list,
889 			    list) {
890 		err = mlx4_get_counter_stats(dev->dev,
891 					     tmp_counter->index,
892 					     &counter_stats, 0);
893 		if (err) {
894 			err = IB_MAD_RESULT_FAILURE;
895 			stats_avail = 0;
896 			break;
897 		}
898 		stats_avail = 1;
899 	}
900 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
901 	if (stats_avail) {
902 		memset(out_mad->data, 0, sizeof out_mad->data);
903 		switch (counter_stats.counter_mode & 0xf) {
904 		case 0:
905 			edit_counter(&counter_stats,
906 				     (void *)(out_mad->data + 40),
907 				     in_mad->mad_hdr.attr_id);
908 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
909 			break;
910 		default:
911 			err = IB_MAD_RESULT_FAILURE;
912 		}
913 	}
914 
915 	return err;
916 }
917 
918 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
919 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
920 			const struct ib_mad_hdr *in, size_t in_mad_size,
921 			struct ib_mad_hdr *out, size_t *out_mad_size,
922 			u16 *out_mad_pkey_index)
923 {
924 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
925 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
926 	struct ib_mad *out_mad = (struct ib_mad *)out;
927 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
928 
929 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
930 			 *out_mad_size != sizeof(*out_mad)))
931 		return IB_MAD_RESULT_FAILURE;
932 
933 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
934 	 * queries, should be called only by VFs and for that specific purpose
935 	 */
936 	if (link == IB_LINK_LAYER_INFINIBAND) {
937 		if (mlx4_is_slave(dev->dev) &&
938 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
939 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
940 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
941 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
942 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
943 						in_grh, in_mad, out_mad);
944 
945 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
946 				      in_grh, in_mad, out_mad);
947 	}
948 
949 	if (link == IB_LINK_LAYER_ETHERNET)
950 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
951 					in_grh, in_mad, out_mad);
952 
953 	return -EINVAL;
954 }
955 
956 static void send_handler(struct ib_mad_agent *agent,
957 			 struct ib_mad_send_wc *mad_send_wc)
958 {
959 	if (mad_send_wc->send_buf->context[0])
960 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
961 	ib_free_send_mad(mad_send_wc->send_buf);
962 }
963 
964 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
965 {
966 	struct ib_mad_agent *agent;
967 	int p, q;
968 	int ret;
969 	enum rdma_link_layer ll;
970 
971 	for (p = 0; p < dev->num_ports; ++p) {
972 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
973 		for (q = 0; q <= 1; ++q) {
974 			if (ll == IB_LINK_LAYER_INFINIBAND) {
975 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
976 							      q ? IB_QPT_GSI : IB_QPT_SMI,
977 							      NULL, 0, send_handler,
978 							      NULL, NULL, 0);
979 				if (IS_ERR(agent)) {
980 					ret = PTR_ERR(agent);
981 					goto err;
982 				}
983 				dev->send_agent[p][q] = agent;
984 			} else
985 				dev->send_agent[p][q] = NULL;
986 		}
987 	}
988 
989 	return 0;
990 
991 err:
992 	for (p = 0; p < dev->num_ports; ++p)
993 		for (q = 0; q <= 1; ++q)
994 			if (dev->send_agent[p][q])
995 				ib_unregister_mad_agent(dev->send_agent[p][q]);
996 
997 	return ret;
998 }
999 
1000 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1001 {
1002 	struct ib_mad_agent *agent;
1003 	int p, q;
1004 
1005 	for (p = 0; p < dev->num_ports; ++p) {
1006 		for (q = 0; q <= 1; ++q) {
1007 			agent = dev->send_agent[p][q];
1008 			if (agent) {
1009 				dev->send_agent[p][q] = NULL;
1010 				ib_unregister_mad_agent(agent);
1011 			}
1012 		}
1013 
1014 		if (dev->sm_ah[p])
1015 			ib_destroy_ah(dev->sm_ah[p]);
1016 	}
1017 }
1018 
1019 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1020 {
1021 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1022 
1023 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1024 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1025 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1026 }
1027 
1028 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1029 {
1030 	/* re-configure the alias-guid and mcg's */
1031 	if (mlx4_is_master(dev->dev)) {
1032 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1033 
1034 		if (!dev->sriov.is_going_down) {
1035 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1036 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1037 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1038 		}
1039 	}
1040 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1041 }
1042 
1043 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1044 			      struct mlx4_eqe *eqe)
1045 {
1046 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1047 			    GET_MASK_FROM_EQE(eqe));
1048 }
1049 
1050 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1051 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1052 {
1053 	struct ib_smp *in_mad  = NULL;
1054 	struct ib_smp *out_mad  = NULL;
1055 	u16 i;
1056 
1057 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1058 		return;
1059 
1060 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1061 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1062 	if (!in_mad || !out_mad) {
1063 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1064 		goto out;
1065 	}
1066 
1067 	guid_tbl_blk_num  *= 4;
1068 
1069 	for (i = 0; i < 4; i++) {
1070 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1071 			continue;
1072 		memset(in_mad, 0, sizeof *in_mad);
1073 		memset(out_mad, 0, sizeof *out_mad);
1074 
1075 		in_mad->base_version  = 1;
1076 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1077 		in_mad->class_version = 1;
1078 		in_mad->method        = IB_MGMT_METHOD_GET;
1079 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1080 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1081 
1082 		if (mlx4_MAD_IFC(dev,
1083 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1084 				 port_num, NULL, NULL, in_mad, out_mad)) {
1085 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1086 			goto out;
1087 		}
1088 
1089 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1090 						    port_num,
1091 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1092 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1093 						     port_num,
1094 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1095 	}
1096 
1097 out:
1098 	kfree(in_mad);
1099 	kfree(out_mad);
1100 	return;
1101 }
1102 
1103 void handle_port_mgmt_change_event(struct work_struct *work)
1104 {
1105 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1106 	struct mlx4_ib_dev *dev = ew->ib_dev;
1107 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1108 	u8 port = eqe->event.port_mgmt_change.port;
1109 	u32 changed_attr;
1110 	u32 tbl_block;
1111 	u32 change_bitmap;
1112 
1113 	switch (eqe->subtype) {
1114 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1115 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1116 
1117 		/* Update the SM ah - This should be done before handling
1118 		   the other changed attributes so that MADs can be sent to the SM */
1119 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1120 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1121 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1122 			update_sm_ah(dev, port, lid, sl);
1123 		}
1124 
1125 		/* Check if it is a lid change event */
1126 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1127 			handle_lid_change_event(dev, port);
1128 
1129 		/* Generate GUID changed event */
1130 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1131 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1132 			/*if master, notify all slaves*/
1133 			if (mlx4_is_master(dev->dev))
1134 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1135 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1136 		}
1137 
1138 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1139 			handle_client_rereg_event(dev, port);
1140 		break;
1141 
1142 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1143 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1144 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1145 			propagate_pkey_ev(dev, port, eqe);
1146 		break;
1147 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1148 		/* paravirtualized master's guid is guid 0 -- does not change */
1149 		if (!mlx4_is_master(dev->dev))
1150 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1151 		/*if master, notify relevant slaves*/
1152 		else if (!dev->sriov.is_going_down) {
1153 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1154 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1155 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1156 		}
1157 		break;
1158 	default:
1159 		pr_warn("Unsupported subtype 0x%x for "
1160 			"Port Management Change event\n", eqe->subtype);
1161 	}
1162 
1163 	kfree(ew);
1164 }
1165 
1166 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1167 			    enum ib_event_type type)
1168 {
1169 	struct ib_event event;
1170 
1171 	event.device		= &dev->ib_dev;
1172 	event.element.port_num	= port_num;
1173 	event.event		= type;
1174 
1175 	ib_dispatch_event(&event);
1176 }
1177 
1178 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1179 {
1180 	unsigned long flags;
1181 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1182 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1183 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1184 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1185 		queue_work(ctx->wq, &ctx->work);
1186 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1187 }
1188 
1189 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1190 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1191 				  int index)
1192 {
1193 	struct ib_sge sg_list;
1194 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1195 	int size;
1196 
1197 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1198 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1199 
1200 	sg_list.addr = tun_qp->ring[index].map;
1201 	sg_list.length = size;
1202 	sg_list.lkey = ctx->pd->local_dma_lkey;
1203 
1204 	recv_wr.next = NULL;
1205 	recv_wr.sg_list = &sg_list;
1206 	recv_wr.num_sge = 1;
1207 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1208 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1209 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1210 				      size, DMA_FROM_DEVICE);
1211 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1212 }
1213 
1214 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1215 		int slave, struct ib_sa_mad *sa_mad)
1216 {
1217 	int ret = 0;
1218 
1219 	/* dispatch to different sa handlers */
1220 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1221 	case IB_SA_ATTR_MC_MEMBER_REC:
1222 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1223 		break;
1224 	default:
1225 		break;
1226 	}
1227 	return ret;
1228 }
1229 
1230 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1231 {
1232 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1233 
1234 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1235 }
1236 
1237 
1238 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1239 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1240 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1241 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1242 {
1243 	struct ib_sge list;
1244 	struct ib_ud_wr wr;
1245 	struct ib_send_wr *bad_wr;
1246 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1247 	struct mlx4_ib_demux_pv_qp *sqp;
1248 	struct mlx4_mad_snd_buf *sqp_mad;
1249 	struct ib_ah *ah;
1250 	struct ib_qp *send_qp = NULL;
1251 	unsigned wire_tx_ix = 0;
1252 	int ret = 0;
1253 	u16 wire_pkey_ix;
1254 	int src_qpnum;
1255 	u8 sgid_index;
1256 
1257 
1258 	sqp_ctx = dev->sriov.sqps[port-1];
1259 
1260 	/* check if proxy qp created */
1261 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1262 		return -EAGAIN;
1263 
1264 	if (dest_qpt == IB_QPT_SMI) {
1265 		src_qpnum = 0;
1266 		sqp = &sqp_ctx->qp[0];
1267 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1268 	} else {
1269 		src_qpnum = 1;
1270 		sqp = &sqp_ctx->qp[1];
1271 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1272 	}
1273 
1274 	send_qp = sqp->qp;
1275 
1276 	/* create ah */
1277 	sgid_index = attr->grh.sgid_index;
1278 	attr->grh.sgid_index = 0;
1279 	ah = ib_create_ah(sqp_ctx->pd, attr);
1280 	if (IS_ERR(ah))
1281 		return -ENOMEM;
1282 	attr->grh.sgid_index = sgid_index;
1283 	to_mah(ah)->av.ib.gid_index = sgid_index;
1284 	/* get rid of force-loopback bit */
1285 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1286 	spin_lock(&sqp->tx_lock);
1287 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1288 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1289 		ret = -EAGAIN;
1290 	else
1291 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1292 	spin_unlock(&sqp->tx_lock);
1293 	if (ret)
1294 		goto out;
1295 
1296 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1297 	if (sqp->tx_ring[wire_tx_ix].ah)
1298 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1299 	sqp->tx_ring[wire_tx_ix].ah = ah;
1300 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1301 				   sqp->tx_ring[wire_tx_ix].buf.map,
1302 				   sizeof (struct mlx4_mad_snd_buf),
1303 				   DMA_TO_DEVICE);
1304 
1305 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1306 
1307 	ib_dma_sync_single_for_device(&dev->ib_dev,
1308 				      sqp->tx_ring[wire_tx_ix].buf.map,
1309 				      sizeof (struct mlx4_mad_snd_buf),
1310 				      DMA_TO_DEVICE);
1311 
1312 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1313 	list.length = sizeof (struct mlx4_mad_snd_buf);
1314 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1315 
1316 	wr.ah = ah;
1317 	wr.port_num = port;
1318 	wr.pkey_index = wire_pkey_ix;
1319 	wr.remote_qkey = qkey;
1320 	wr.remote_qpn = remote_qpn;
1321 	wr.wr.next = NULL;
1322 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1323 	wr.wr.sg_list = &list;
1324 	wr.wr.num_sge = 1;
1325 	wr.wr.opcode = IB_WR_SEND;
1326 	wr.wr.send_flags = IB_SEND_SIGNALED;
1327 	if (s_mac)
1328 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1329 	if (vlan_id < 0x1000)
1330 		vlan_id |= (attr->sl & 7) << 13;
1331 	to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1332 
1333 
1334 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1335 	if (!ret)
1336 		return 0;
1337 
1338 	spin_lock(&sqp->tx_lock);
1339 	sqp->tx_ix_tail++;
1340 	spin_unlock(&sqp->tx_lock);
1341 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1342 out:
1343 	ib_destroy_ah(ah);
1344 	return ret;
1345 }
1346 
1347 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1348 {
1349 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1350 		return slave;
1351 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1352 }
1353 
1354 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1355 				    struct ib_ah_attr *ah_attr)
1356 {
1357 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1358 		ah_attr->grh.sgid_index = slave;
1359 	else
1360 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1361 }
1362 
1363 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1364 {
1365 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1366 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1367 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1368 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1369 	struct mlx4_ib_ah ah;
1370 	struct ib_ah_attr ah_attr;
1371 	u8 *slave_id;
1372 	int slave;
1373 	int port;
1374 	u16 vlan_id;
1375 
1376 	/* Get slave that sent this packet */
1377 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1378 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1379 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1380 	    wc->src_qp & 0x4) {
1381 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1382 		return;
1383 	}
1384 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1385 	if (slave != ctx->slave) {
1386 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1387 			     "belongs to another slave\n", wc->src_qp);
1388 		return;
1389 	}
1390 
1391 	/* Map transaction ID */
1392 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1393 				   sizeof (struct mlx4_tunnel_mad),
1394 				   DMA_FROM_DEVICE);
1395 	switch (tunnel->mad.mad_hdr.method) {
1396 	case IB_MGMT_METHOD_SET:
1397 	case IB_MGMT_METHOD_GET:
1398 	case IB_MGMT_METHOD_REPORT:
1399 	case IB_SA_METHOD_GET_TABLE:
1400 	case IB_SA_METHOD_DELETE:
1401 	case IB_SA_METHOD_GET_MULTI:
1402 	case IB_SA_METHOD_GET_TRACE_TBL:
1403 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1404 		if (*slave_id) {
1405 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1406 				     "class:%d slave:%d\n", *slave_id,
1407 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1408 			return;
1409 		} else
1410 			*slave_id = slave;
1411 	default:
1412 		/* nothing */;
1413 	}
1414 
1415 	/* Class-specific handling */
1416 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1417 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1418 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1419 		if (slave != mlx4_master_func_num(dev->dev) &&
1420 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1421 			return;
1422 		break;
1423 	case IB_MGMT_CLASS_SUBN_ADM:
1424 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1425 			      (struct ib_sa_mad *) &tunnel->mad))
1426 			return;
1427 		break;
1428 	case IB_MGMT_CLASS_CM:
1429 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1430 			      (struct ib_mad *) &tunnel->mad))
1431 			return;
1432 		break;
1433 	case IB_MGMT_CLASS_DEVICE_MGMT:
1434 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1435 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1436 			return;
1437 		break;
1438 	default:
1439 		/* Drop unsupported classes for slaves in tunnel mode */
1440 		if (slave != mlx4_master_func_num(dev->dev)) {
1441 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1442 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1443 			return;
1444 		}
1445 	}
1446 
1447 	/* We are using standard ib_core services to send the mad, so generate a
1448 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1449 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1450 	ah.ibah.device = ctx->ib_dev;
1451 
1452 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1453 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1454 	if (port < 0)
1455 		return;
1456 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1457 
1458 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1459 	if (ah_attr.ah_flags & IB_AH_GRH)
1460 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1461 
1462 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1463 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1464 	/* if slave have default vlan use it */
1465 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1466 				    &vlan_id, &ah_attr.sl);
1467 
1468 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1469 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1470 			     IB_QPT_SMI : IB_QPT_GSI,
1471 			     be16_to_cpu(tunnel->hdr.pkey_index),
1472 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1473 			     be32_to_cpu(tunnel->hdr.qkey),
1474 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1475 }
1476 
1477 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1478 				 enum ib_qp_type qp_type, int is_tun)
1479 {
1480 	int i;
1481 	struct mlx4_ib_demux_pv_qp *tun_qp;
1482 	int rx_buf_size, tx_buf_size;
1483 
1484 	if (qp_type > IB_QPT_GSI)
1485 		return -EINVAL;
1486 
1487 	tun_qp = &ctx->qp[qp_type];
1488 
1489 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1490 			       GFP_KERNEL);
1491 	if (!tun_qp->ring)
1492 		return -ENOMEM;
1493 
1494 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1495 				  sizeof (struct mlx4_ib_tun_tx_buf),
1496 				  GFP_KERNEL);
1497 	if (!tun_qp->tx_ring) {
1498 		kfree(tun_qp->ring);
1499 		tun_qp->ring = NULL;
1500 		return -ENOMEM;
1501 	}
1502 
1503 	if (is_tun) {
1504 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1505 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1506 	} else {
1507 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1508 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1509 	}
1510 
1511 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1512 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1513 		if (!tun_qp->ring[i].addr)
1514 			goto err;
1515 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1516 							tun_qp->ring[i].addr,
1517 							rx_buf_size,
1518 							DMA_FROM_DEVICE);
1519 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1520 			kfree(tun_qp->ring[i].addr);
1521 			goto err;
1522 		}
1523 	}
1524 
1525 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1526 		tun_qp->tx_ring[i].buf.addr =
1527 			kmalloc(tx_buf_size, GFP_KERNEL);
1528 		if (!tun_qp->tx_ring[i].buf.addr)
1529 			goto tx_err;
1530 		tun_qp->tx_ring[i].buf.map =
1531 			ib_dma_map_single(ctx->ib_dev,
1532 					  tun_qp->tx_ring[i].buf.addr,
1533 					  tx_buf_size,
1534 					  DMA_TO_DEVICE);
1535 		if (ib_dma_mapping_error(ctx->ib_dev,
1536 					 tun_qp->tx_ring[i].buf.map)) {
1537 			kfree(tun_qp->tx_ring[i].buf.addr);
1538 			goto tx_err;
1539 		}
1540 		tun_qp->tx_ring[i].ah = NULL;
1541 	}
1542 	spin_lock_init(&tun_qp->tx_lock);
1543 	tun_qp->tx_ix_head = 0;
1544 	tun_qp->tx_ix_tail = 0;
1545 	tun_qp->proxy_qpt = qp_type;
1546 
1547 	return 0;
1548 
1549 tx_err:
1550 	while (i > 0) {
1551 		--i;
1552 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1553 				    tx_buf_size, DMA_TO_DEVICE);
1554 		kfree(tun_qp->tx_ring[i].buf.addr);
1555 	}
1556 	kfree(tun_qp->tx_ring);
1557 	tun_qp->tx_ring = NULL;
1558 	i = MLX4_NUM_TUNNEL_BUFS;
1559 err:
1560 	while (i > 0) {
1561 		--i;
1562 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1563 				    rx_buf_size, DMA_FROM_DEVICE);
1564 		kfree(tun_qp->ring[i].addr);
1565 	}
1566 	kfree(tun_qp->ring);
1567 	tun_qp->ring = NULL;
1568 	return -ENOMEM;
1569 }
1570 
1571 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1572 				     enum ib_qp_type qp_type, int is_tun)
1573 {
1574 	int i;
1575 	struct mlx4_ib_demux_pv_qp *tun_qp;
1576 	int rx_buf_size, tx_buf_size;
1577 
1578 	if (qp_type > IB_QPT_GSI)
1579 		return;
1580 
1581 	tun_qp = &ctx->qp[qp_type];
1582 	if (is_tun) {
1583 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1584 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1585 	} else {
1586 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1587 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1588 	}
1589 
1590 
1591 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1592 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1593 				    rx_buf_size, DMA_FROM_DEVICE);
1594 		kfree(tun_qp->ring[i].addr);
1595 	}
1596 
1597 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1598 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1599 				    tx_buf_size, DMA_TO_DEVICE);
1600 		kfree(tun_qp->tx_ring[i].buf.addr);
1601 		if (tun_qp->tx_ring[i].ah)
1602 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1603 	}
1604 	kfree(tun_qp->tx_ring);
1605 	kfree(tun_qp->ring);
1606 }
1607 
1608 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1609 {
1610 	struct mlx4_ib_demux_pv_ctx *ctx;
1611 	struct mlx4_ib_demux_pv_qp *tun_qp;
1612 	struct ib_wc wc;
1613 	int ret;
1614 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1615 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1616 
1617 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1618 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1619 		if (wc.status == IB_WC_SUCCESS) {
1620 			switch (wc.opcode) {
1621 			case IB_WC_RECV:
1622 				mlx4_ib_multiplex_mad(ctx, &wc);
1623 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1624 							     wc.wr_id &
1625 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1626 				if (ret)
1627 					pr_err("Failed reposting tunnel "
1628 					       "buf:%lld\n", wc.wr_id);
1629 				break;
1630 			case IB_WC_SEND:
1631 				pr_debug("received tunnel send completion:"
1632 					 "wrid=0x%llx, status=0x%x\n",
1633 					 wc.wr_id, wc.status);
1634 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1635 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1636 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1637 					= NULL;
1638 				spin_lock(&tun_qp->tx_lock);
1639 				tun_qp->tx_ix_tail++;
1640 				spin_unlock(&tun_qp->tx_lock);
1641 
1642 				break;
1643 			default:
1644 				break;
1645 			}
1646 		} else  {
1647 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1648 				 " status = %d, wrid = 0x%llx\n",
1649 				 ctx->slave, wc.status, wc.wr_id);
1650 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1651 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1652 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1653 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1654 					= NULL;
1655 				spin_lock(&tun_qp->tx_lock);
1656 				tun_qp->tx_ix_tail++;
1657 				spin_unlock(&tun_qp->tx_lock);
1658 			}
1659 		}
1660 	}
1661 }
1662 
1663 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1664 {
1665 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1666 
1667 	/* It's worse than that! He's dead, Jim! */
1668 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1669 	       event->event, sqp->port);
1670 }
1671 
1672 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1673 			    enum ib_qp_type qp_type, int create_tun)
1674 {
1675 	int i, ret;
1676 	struct mlx4_ib_demux_pv_qp *tun_qp;
1677 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1678 	struct ib_qp_attr attr;
1679 	int qp_attr_mask_INIT;
1680 
1681 	if (qp_type > IB_QPT_GSI)
1682 		return -EINVAL;
1683 
1684 	tun_qp = &ctx->qp[qp_type];
1685 
1686 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1687 	qp_init_attr.init_attr.send_cq = ctx->cq;
1688 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1689 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1690 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1691 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1692 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1693 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1694 	if (create_tun) {
1695 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1696 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1697 		qp_init_attr.port = ctx->port;
1698 		qp_init_attr.slave = ctx->slave;
1699 		qp_init_attr.proxy_qp_type = qp_type;
1700 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1701 			   IB_QP_QKEY | IB_QP_PORT;
1702 	} else {
1703 		qp_init_attr.init_attr.qp_type = qp_type;
1704 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1705 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1706 	}
1707 	qp_init_attr.init_attr.port_num = ctx->port;
1708 	qp_init_attr.init_attr.qp_context = ctx;
1709 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1710 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1711 	if (IS_ERR(tun_qp->qp)) {
1712 		ret = PTR_ERR(tun_qp->qp);
1713 		tun_qp->qp = NULL;
1714 		pr_err("Couldn't create %s QP (%d)\n",
1715 		       create_tun ? "tunnel" : "special", ret);
1716 		return ret;
1717 	}
1718 
1719 	memset(&attr, 0, sizeof attr);
1720 	attr.qp_state = IB_QPS_INIT;
1721 	ret = 0;
1722 	if (create_tun)
1723 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1724 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1725 					      &attr.pkey_index);
1726 	if (ret || !create_tun)
1727 		attr.pkey_index =
1728 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1729 	attr.qkey = IB_QP1_QKEY;
1730 	attr.port_num = ctx->port;
1731 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1732 	if (ret) {
1733 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1734 		       create_tun ? "tunnel" : "special", ret);
1735 		goto err_qp;
1736 	}
1737 	attr.qp_state = IB_QPS_RTR;
1738 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1739 	if (ret) {
1740 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1741 		       create_tun ? "tunnel" : "special", ret);
1742 		goto err_qp;
1743 	}
1744 	attr.qp_state = IB_QPS_RTS;
1745 	attr.sq_psn = 0;
1746 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1747 	if (ret) {
1748 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1749 		       create_tun ? "tunnel" : "special", ret);
1750 		goto err_qp;
1751 	}
1752 
1753 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1754 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1755 		if (ret) {
1756 			pr_err(" mlx4_ib_post_pv_buf error"
1757 			       " (err = %d, i = %d)\n", ret, i);
1758 			goto err_qp;
1759 		}
1760 	}
1761 	return 0;
1762 
1763 err_qp:
1764 	ib_destroy_qp(tun_qp->qp);
1765 	tun_qp->qp = NULL;
1766 	return ret;
1767 }
1768 
1769 /*
1770  * IB MAD completion callback for real SQPs
1771  */
1772 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1773 {
1774 	struct mlx4_ib_demux_pv_ctx *ctx;
1775 	struct mlx4_ib_demux_pv_qp *sqp;
1776 	struct ib_wc wc;
1777 	struct ib_grh *grh;
1778 	struct ib_mad *mad;
1779 
1780 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1781 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1782 
1783 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1784 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1785 		if (wc.status == IB_WC_SUCCESS) {
1786 			switch (wc.opcode) {
1787 			case IB_WC_SEND:
1788 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1789 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1790 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1791 					= NULL;
1792 				spin_lock(&sqp->tx_lock);
1793 				sqp->tx_ix_tail++;
1794 				spin_unlock(&sqp->tx_lock);
1795 				break;
1796 			case IB_WC_RECV:
1797 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1798 						(sqp->ring[wc.wr_id &
1799 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1800 				grh = &(((struct mlx4_mad_rcv_buf *)
1801 						(sqp->ring[wc.wr_id &
1802 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1803 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1804 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1805 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1806 					pr_err("Failed reposting SQP "
1807 					       "buf:%lld\n", wc.wr_id);
1808 				break;
1809 			default:
1810 				BUG_ON(1);
1811 				break;
1812 			}
1813 		} else  {
1814 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1815 				 " status = %d, wrid = 0x%llx\n",
1816 				 ctx->slave, wc.status, wc.wr_id);
1817 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1818 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1819 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1820 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1821 					= NULL;
1822 				spin_lock(&sqp->tx_lock);
1823 				sqp->tx_ix_tail++;
1824 				spin_unlock(&sqp->tx_lock);
1825 			}
1826 		}
1827 	}
1828 }
1829 
1830 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1831 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1832 {
1833 	struct mlx4_ib_demux_pv_ctx *ctx;
1834 
1835 	*ret_ctx = NULL;
1836 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1837 	if (!ctx) {
1838 		pr_err("failed allocating pv resource context "
1839 		       "for port %d, slave %d\n", port, slave);
1840 		return -ENOMEM;
1841 	}
1842 
1843 	ctx->ib_dev = &dev->ib_dev;
1844 	ctx->port = port;
1845 	ctx->slave = slave;
1846 	*ret_ctx = ctx;
1847 	return 0;
1848 }
1849 
1850 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1851 {
1852 	if (dev->sriov.demux[port - 1].tun[slave]) {
1853 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1854 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1855 	}
1856 }
1857 
1858 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1859 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1860 {
1861 	int ret, cq_size;
1862 	struct ib_cq_init_attr cq_attr = {};
1863 
1864 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1865 		return -EEXIST;
1866 
1867 	ctx->state = DEMUX_PV_STATE_STARTING;
1868 	/* have QP0 only if link layer is IB */
1869 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1870 	    IB_LINK_LAYER_INFINIBAND)
1871 		ctx->has_smi = 1;
1872 
1873 	if (ctx->has_smi) {
1874 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1875 		if (ret) {
1876 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1877 			goto err_out;
1878 		}
1879 	}
1880 
1881 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1882 	if (ret) {
1883 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1884 		goto err_out_qp0;
1885 	}
1886 
1887 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1888 	if (ctx->has_smi)
1889 		cq_size *= 2;
1890 
1891 	cq_attr.cqe = cq_size;
1892 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1893 			       NULL, ctx, &cq_attr);
1894 	if (IS_ERR(ctx->cq)) {
1895 		ret = PTR_ERR(ctx->cq);
1896 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1897 		goto err_buf;
1898 	}
1899 
1900 	ctx->pd = ib_alloc_pd(ctx->ib_dev);
1901 	if (IS_ERR(ctx->pd)) {
1902 		ret = PTR_ERR(ctx->pd);
1903 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1904 		goto err_cq;
1905 	}
1906 
1907 	if (ctx->has_smi) {
1908 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1909 		if (ret) {
1910 			pr_err("Couldn't create %s QP0 (%d)\n",
1911 			       create_tun ? "tunnel for" : "",  ret);
1912 			goto err_pd;
1913 		}
1914 	}
1915 
1916 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1917 	if (ret) {
1918 		pr_err("Couldn't create %s QP1 (%d)\n",
1919 		       create_tun ? "tunnel for" : "",  ret);
1920 		goto err_qp0;
1921 	}
1922 
1923 	if (create_tun)
1924 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
1925 	else
1926 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
1927 
1928 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
1929 
1930 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1931 	if (ret) {
1932 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
1933 		goto err_wq;
1934 	}
1935 	ctx->state = DEMUX_PV_STATE_ACTIVE;
1936 	return 0;
1937 
1938 err_wq:
1939 	ctx->wq = NULL;
1940 	ib_destroy_qp(ctx->qp[1].qp);
1941 	ctx->qp[1].qp = NULL;
1942 
1943 
1944 err_qp0:
1945 	if (ctx->has_smi)
1946 		ib_destroy_qp(ctx->qp[0].qp);
1947 	ctx->qp[0].qp = NULL;
1948 
1949 err_pd:
1950 	ib_dealloc_pd(ctx->pd);
1951 	ctx->pd = NULL;
1952 
1953 err_cq:
1954 	ib_destroy_cq(ctx->cq);
1955 	ctx->cq = NULL;
1956 
1957 err_buf:
1958 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
1959 
1960 err_out_qp0:
1961 	if (ctx->has_smi)
1962 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
1963 err_out:
1964 	ctx->state = DEMUX_PV_STATE_DOWN;
1965 	return ret;
1966 }
1967 
1968 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
1969 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
1970 {
1971 	if (!ctx)
1972 		return;
1973 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
1974 		ctx->state = DEMUX_PV_STATE_DOWNING;
1975 		if (flush)
1976 			flush_workqueue(ctx->wq);
1977 		if (ctx->has_smi) {
1978 			ib_destroy_qp(ctx->qp[0].qp);
1979 			ctx->qp[0].qp = NULL;
1980 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
1981 		}
1982 		ib_destroy_qp(ctx->qp[1].qp);
1983 		ctx->qp[1].qp = NULL;
1984 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
1985 		ib_dealloc_pd(ctx->pd);
1986 		ctx->pd = NULL;
1987 		ib_destroy_cq(ctx->cq);
1988 		ctx->cq = NULL;
1989 		ctx->state = DEMUX_PV_STATE_DOWN;
1990 	}
1991 }
1992 
1993 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
1994 				  int port, int do_init)
1995 {
1996 	int ret = 0;
1997 
1998 	if (!do_init) {
1999 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2000 		/* for master, destroy real sqp resources */
2001 		if (slave == mlx4_master_func_num(dev->dev))
2002 			destroy_pv_resources(dev, slave, port,
2003 					     dev->sriov.sqps[port - 1], 1);
2004 		/* destroy the tunnel qp resources */
2005 		destroy_pv_resources(dev, slave, port,
2006 				     dev->sriov.demux[port - 1].tun[slave], 1);
2007 		return 0;
2008 	}
2009 
2010 	/* create the tunnel qp resources */
2011 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2012 				  dev->sriov.demux[port - 1].tun[slave]);
2013 
2014 	/* for master, create the real sqp resources */
2015 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2016 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2017 					  dev->sriov.sqps[port - 1]);
2018 	return ret;
2019 }
2020 
2021 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2022 {
2023 	struct mlx4_ib_demux_work *dmxw;
2024 
2025 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2026 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2027 			       dmxw->do_init);
2028 	kfree(dmxw);
2029 	return;
2030 }
2031 
2032 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2033 				       struct mlx4_ib_demux_ctx *ctx,
2034 				       int port)
2035 {
2036 	char name[12];
2037 	int ret = 0;
2038 	int i;
2039 
2040 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2041 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2042 	if (!ctx->tun)
2043 		return -ENOMEM;
2044 
2045 	ctx->dev = dev;
2046 	ctx->port = port;
2047 	ctx->ib_dev = &dev->ib_dev;
2048 
2049 	for (i = 0;
2050 	     i < min(dev->dev->caps.sqp_demux,
2051 	     (u16)(dev->dev->persist->num_vfs + 1));
2052 	     i++) {
2053 		struct mlx4_active_ports actv_ports =
2054 			mlx4_get_active_ports(dev->dev, i);
2055 
2056 		if (!test_bit(port - 1, actv_ports.ports))
2057 			continue;
2058 
2059 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2060 		if (ret) {
2061 			ret = -ENOMEM;
2062 			goto err_mcg;
2063 		}
2064 	}
2065 
2066 	ret = mlx4_ib_mcg_port_init(ctx);
2067 	if (ret) {
2068 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2069 		goto err_mcg;
2070 	}
2071 
2072 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2073 	ctx->wq = create_singlethread_workqueue(name);
2074 	if (!ctx->wq) {
2075 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2076 		ret = -ENOMEM;
2077 		goto err_wq;
2078 	}
2079 
2080 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2081 	ctx->ud_wq = create_singlethread_workqueue(name);
2082 	if (!ctx->ud_wq) {
2083 		pr_err("Failed to create up/down WQ for port %d\n", port);
2084 		ret = -ENOMEM;
2085 		goto err_udwq;
2086 	}
2087 
2088 	return 0;
2089 
2090 err_udwq:
2091 	destroy_workqueue(ctx->wq);
2092 	ctx->wq = NULL;
2093 
2094 err_wq:
2095 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2096 err_mcg:
2097 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2098 		free_pv_object(dev, i, port);
2099 	kfree(ctx->tun);
2100 	ctx->tun = NULL;
2101 	return ret;
2102 }
2103 
2104 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2105 {
2106 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2107 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2108 		flush_workqueue(sqp_ctx->wq);
2109 		if (sqp_ctx->has_smi) {
2110 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2111 			sqp_ctx->qp[0].qp = NULL;
2112 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2113 		}
2114 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2115 		sqp_ctx->qp[1].qp = NULL;
2116 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2117 		ib_dealloc_pd(sqp_ctx->pd);
2118 		sqp_ctx->pd = NULL;
2119 		ib_destroy_cq(sqp_ctx->cq);
2120 		sqp_ctx->cq = NULL;
2121 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2122 	}
2123 }
2124 
2125 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2126 {
2127 	int i;
2128 	if (ctx) {
2129 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2130 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2131 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2132 			if (!ctx->tun[i])
2133 				continue;
2134 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2135 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2136 		}
2137 		flush_workqueue(ctx->wq);
2138 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2139 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2140 			free_pv_object(dev, i, ctx->port);
2141 		}
2142 		kfree(ctx->tun);
2143 		destroy_workqueue(ctx->ud_wq);
2144 		destroy_workqueue(ctx->wq);
2145 	}
2146 }
2147 
2148 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2149 {
2150 	int i;
2151 
2152 	if (!mlx4_is_master(dev->dev))
2153 		return;
2154 	/* initialize or tear down tunnel QPs for the master */
2155 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2156 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2157 	return;
2158 }
2159 
2160 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2161 {
2162 	int i = 0;
2163 	int err;
2164 
2165 	if (!mlx4_is_mfunc(dev->dev))
2166 		return 0;
2167 
2168 	dev->sriov.is_going_down = 0;
2169 	spin_lock_init(&dev->sriov.going_down_lock);
2170 	mlx4_ib_cm_paravirt_init(dev);
2171 
2172 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2173 
2174 	if (mlx4_is_slave(dev->dev)) {
2175 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2176 		return 0;
2177 	}
2178 
2179 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2180 		if (i == mlx4_master_func_num(dev->dev))
2181 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2182 		else
2183 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2184 	}
2185 
2186 	err = mlx4_ib_init_alias_guid_service(dev);
2187 	if (err) {
2188 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2189 		goto paravirt_err;
2190 	}
2191 	err = mlx4_ib_device_register_sysfs(dev);
2192 	if (err) {
2193 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2194 		goto sysfs_err;
2195 	}
2196 
2197 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2198 		     dev->dev->caps.sqp_demux);
2199 	for (i = 0; i < dev->num_ports; i++) {
2200 		union ib_gid gid;
2201 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2202 		if (err)
2203 			goto demux_err;
2204 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2205 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2206 				      &dev->sriov.sqps[i]);
2207 		if (err)
2208 			goto demux_err;
2209 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2210 		if (err)
2211 			goto free_pv;
2212 	}
2213 	mlx4_ib_master_tunnels(dev, 1);
2214 	return 0;
2215 
2216 free_pv:
2217 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2218 demux_err:
2219 	while (--i >= 0) {
2220 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2221 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2222 	}
2223 	mlx4_ib_device_unregister_sysfs(dev);
2224 
2225 sysfs_err:
2226 	mlx4_ib_destroy_alias_guid_service(dev);
2227 
2228 paravirt_err:
2229 	mlx4_ib_cm_paravirt_clean(dev, -1);
2230 
2231 	return err;
2232 }
2233 
2234 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2235 {
2236 	int i;
2237 	unsigned long flags;
2238 
2239 	if (!mlx4_is_mfunc(dev->dev))
2240 		return;
2241 
2242 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2243 	dev->sriov.is_going_down = 1;
2244 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2245 	if (mlx4_is_master(dev->dev)) {
2246 		for (i = 0; i < dev->num_ports; i++) {
2247 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2248 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2249 			kfree(dev->sriov.sqps[i]);
2250 			dev->sriov.sqps[i] = NULL;
2251 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2252 		}
2253 
2254 		mlx4_ib_cm_paravirt_clean(dev, -1);
2255 		mlx4_ib_destroy_alias_guid_service(dev);
2256 		mlx4_ib_device_unregister_sysfs(dev);
2257 	}
2258 }
2259