xref: /openbmc/linux/drivers/infiniband/hw/mlx4/mad.c (revision de2bdb3d)
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 			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("port %u, sl2vl[%d] = %02x\n",
319 						 port_num, jj, sl2vl64.sl8[jj]);
320 				}
321 				atomic64_set(&dev->sl2vl[port_num - 1], sl2vl64.sl64);
322 			}
323 			break;
324 
325 		default:
326 			break;
327 		}
328 }
329 
330 static void __propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
331 				int block, u32 change_bitmap)
332 {
333 	int i, ix, slave, err;
334 	int have_event = 0;
335 
336 	for (slave = 0; slave < dev->dev->caps.sqp_demux; slave++) {
337 		if (slave == mlx4_master_func_num(dev->dev))
338 			continue;
339 		if (!mlx4_is_slave_active(dev->dev, slave))
340 			continue;
341 
342 		have_event = 0;
343 		for (i = 0; i < 32; i++) {
344 			if (!(change_bitmap & (1 << i)))
345 				continue;
346 			for (ix = 0;
347 			     ix < dev->dev->caps.pkey_table_len[port_num]; ix++) {
348 				if (dev->pkeys.virt2phys_pkey[slave][port_num - 1]
349 				    [ix] == i + 32 * block) {
350 					err = mlx4_gen_pkey_eqe(dev->dev, slave, port_num);
351 					pr_debug("propagate_pkey_ev: slave %d,"
352 						 " port %d, ix %d (%d)\n",
353 						 slave, port_num, ix, err);
354 					have_event = 1;
355 					break;
356 				}
357 			}
358 			if (have_event)
359 				break;
360 		}
361 	}
362 }
363 
364 static void node_desc_override(struct ib_device *dev,
365 			       struct ib_mad *mad)
366 {
367 	unsigned long flags;
368 
369 	if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
370 	     mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
371 	    mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
372 	    mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
373 		spin_lock_irqsave(&to_mdev(dev)->sm_lock, flags);
374 		memcpy(((struct ib_smp *) mad)->data, dev->node_desc,
375 		       IB_DEVICE_NODE_DESC_MAX);
376 		spin_unlock_irqrestore(&to_mdev(dev)->sm_lock, flags);
377 	}
378 }
379 
380 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, const struct ib_mad *mad)
381 {
382 	int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
383 	struct ib_mad_send_buf *send_buf;
384 	struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
385 	int ret;
386 	unsigned long flags;
387 
388 	if (agent) {
389 		send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
390 					      IB_MGMT_MAD_DATA, GFP_ATOMIC,
391 					      IB_MGMT_BASE_VERSION);
392 		if (IS_ERR(send_buf))
393 			return;
394 		/*
395 		 * We rely here on the fact that MLX QPs don't use the
396 		 * address handle after the send is posted (this is
397 		 * wrong following the IB spec strictly, but we know
398 		 * it's OK for our devices).
399 		 */
400 		spin_lock_irqsave(&dev->sm_lock, flags);
401 		memcpy(send_buf->mad, mad, sizeof *mad);
402 		if ((send_buf->ah = dev->sm_ah[port_num - 1]))
403 			ret = ib_post_send_mad(send_buf, NULL);
404 		else
405 			ret = -EINVAL;
406 		spin_unlock_irqrestore(&dev->sm_lock, flags);
407 
408 		if (ret)
409 			ib_free_send_mad(send_buf);
410 	}
411 }
412 
413 static int mlx4_ib_demux_sa_handler(struct ib_device *ibdev, int port, int slave,
414 							     struct ib_sa_mad *sa_mad)
415 {
416 	int ret = 0;
417 
418 	/* dispatch to different sa handlers */
419 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
420 	case IB_SA_ATTR_MC_MEMBER_REC:
421 		ret = mlx4_ib_mcg_demux_handler(ibdev, port, slave, sa_mad);
422 		break;
423 	default:
424 		break;
425 	}
426 	return ret;
427 }
428 
429 int mlx4_ib_find_real_gid(struct ib_device *ibdev, u8 port, __be64 guid)
430 {
431 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
432 	int i;
433 
434 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
435 		if (dev->sriov.demux[port - 1].guid_cache[i] == guid)
436 			return i;
437 	}
438 	return -1;
439 }
440 
441 
442 static int find_slave_port_pkey_ix(struct mlx4_ib_dev *dev, int slave,
443 				   u8 port, u16 pkey, u16 *ix)
444 {
445 	int i, ret;
446 	u8 unassigned_pkey_ix, pkey_ix, partial_ix = 0xFF;
447 	u16 slot_pkey;
448 
449 	if (slave == mlx4_master_func_num(dev->dev))
450 		return ib_find_cached_pkey(&dev->ib_dev, port, pkey, ix);
451 
452 	unassigned_pkey_ix = dev->dev->phys_caps.pkey_phys_table_len[port] - 1;
453 
454 	for (i = 0; i < dev->dev->caps.pkey_table_len[port]; i++) {
455 		if (dev->pkeys.virt2phys_pkey[slave][port - 1][i] == unassigned_pkey_ix)
456 			continue;
457 
458 		pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][i];
459 
460 		ret = ib_get_cached_pkey(&dev->ib_dev, port, pkey_ix, &slot_pkey);
461 		if (ret)
462 			continue;
463 		if ((slot_pkey & 0x7FFF) == (pkey & 0x7FFF)) {
464 			if (slot_pkey & 0x8000) {
465 				*ix = (u16) pkey_ix;
466 				return 0;
467 			} else {
468 				/* take first partial pkey index found */
469 				if (partial_ix == 0xFF)
470 					partial_ix = pkey_ix;
471 			}
472 		}
473 	}
474 
475 	if (partial_ix < 0xFF) {
476 		*ix = (u16) partial_ix;
477 		return 0;
478 	}
479 
480 	return -EINVAL;
481 }
482 
483 int mlx4_ib_send_to_slave(struct mlx4_ib_dev *dev, int slave, u8 port,
484 			  enum ib_qp_type dest_qpt, struct ib_wc *wc,
485 			  struct ib_grh *grh, struct ib_mad *mad)
486 {
487 	struct ib_sge list;
488 	struct ib_ud_wr wr;
489 	struct ib_send_wr *bad_wr;
490 	struct mlx4_ib_demux_pv_ctx *tun_ctx;
491 	struct mlx4_ib_demux_pv_qp *tun_qp;
492 	struct mlx4_rcv_tunnel_mad *tun_mad;
493 	struct ib_ah_attr attr;
494 	struct ib_ah *ah;
495 	struct ib_qp *src_qp = NULL;
496 	unsigned tun_tx_ix = 0;
497 	int dqpn;
498 	int ret = 0;
499 	u16 tun_pkey_ix;
500 	u16 cached_pkey;
501 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
502 
503 	if (dest_qpt > IB_QPT_GSI)
504 		return -EINVAL;
505 
506 	tun_ctx = dev->sriov.demux[port-1].tun[slave];
507 
508 	/* check if proxy qp created */
509 	if (!tun_ctx || tun_ctx->state != DEMUX_PV_STATE_ACTIVE)
510 		return -EAGAIN;
511 
512 	if (!dest_qpt)
513 		tun_qp = &tun_ctx->qp[0];
514 	else
515 		tun_qp = &tun_ctx->qp[1];
516 
517 	/* compute P_Key index to put in tunnel header for slave */
518 	if (dest_qpt) {
519 		u16 pkey_ix;
520 		ret = ib_get_cached_pkey(&dev->ib_dev, port, wc->pkey_index, &cached_pkey);
521 		if (ret)
522 			return -EINVAL;
523 
524 		ret = find_slave_port_pkey_ix(dev, slave, port, cached_pkey, &pkey_ix);
525 		if (ret)
526 			return -EINVAL;
527 		tun_pkey_ix = pkey_ix;
528 	} else
529 		tun_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
530 
531 	dqpn = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave + port + (dest_qpt * 2) - 1;
532 
533 	/* get tunnel tx data buf for slave */
534 	src_qp = tun_qp->qp;
535 
536 	/* create ah. Just need an empty one with the port num for the post send.
537 	 * The driver will set the force loopback bit in post_send */
538 	memset(&attr, 0, sizeof attr);
539 	attr.port_num = port;
540 	if (is_eth) {
541 		memcpy(&attr.grh.dgid.raw[0], &grh->dgid.raw[0], 16);
542 		attr.ah_flags = IB_AH_GRH;
543 	}
544 	ah = ib_create_ah(tun_ctx->pd, &attr);
545 	if (IS_ERR(ah))
546 		return -ENOMEM;
547 
548 	/* allocate tunnel tx buf after pass failure returns */
549 	spin_lock(&tun_qp->tx_lock);
550 	if (tun_qp->tx_ix_head - tun_qp->tx_ix_tail >=
551 	    (MLX4_NUM_TUNNEL_BUFS - 1))
552 		ret = -EAGAIN;
553 	else
554 		tun_tx_ix = (++tun_qp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
555 	spin_unlock(&tun_qp->tx_lock);
556 	if (ret)
557 		goto end;
558 
559 	tun_mad = (struct mlx4_rcv_tunnel_mad *) (tun_qp->tx_ring[tun_tx_ix].buf.addr);
560 	if (tun_qp->tx_ring[tun_tx_ix].ah)
561 		ib_destroy_ah(tun_qp->tx_ring[tun_tx_ix].ah);
562 	tun_qp->tx_ring[tun_tx_ix].ah = ah;
563 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
564 				   tun_qp->tx_ring[tun_tx_ix].buf.map,
565 				   sizeof (struct mlx4_rcv_tunnel_mad),
566 				   DMA_TO_DEVICE);
567 
568 	/* copy over to tunnel buffer */
569 	if (grh)
570 		memcpy(&tun_mad->grh, grh, sizeof *grh);
571 	memcpy(&tun_mad->mad, mad, sizeof *mad);
572 
573 	/* adjust tunnel data */
574 	tun_mad->hdr.pkey_index = cpu_to_be16(tun_pkey_ix);
575 	tun_mad->hdr.flags_src_qp = cpu_to_be32(wc->src_qp & 0xFFFFFF);
576 	tun_mad->hdr.g_ml_path = (grh && (wc->wc_flags & IB_WC_GRH)) ? 0x80 : 0;
577 
578 	if (is_eth) {
579 		u16 vlan = 0;
580 		if (mlx4_get_slave_default_vlan(dev->dev, port, slave, &vlan,
581 						NULL)) {
582 			/* VST mode */
583 			if (vlan != wc->vlan_id)
584 				/* Packet vlan is not the VST-assigned vlan.
585 				 * Drop the packet.
586 				 */
587 				goto out;
588 			 else
589 				/* Remove the vlan tag before forwarding
590 				 * the packet to the VF.
591 				 */
592 				vlan = 0xffff;
593 		} else {
594 			vlan = wc->vlan_id;
595 		}
596 
597 		tun_mad->hdr.sl_vid = cpu_to_be16(vlan);
598 		memcpy((char *)&tun_mad->hdr.mac_31_0, &(wc->smac[0]), 4);
599 		memcpy((char *)&tun_mad->hdr.slid_mac_47_32, &(wc->smac[4]), 2);
600 	} else {
601 		tun_mad->hdr.sl_vid = cpu_to_be16(((u16)(wc->sl)) << 12);
602 		tun_mad->hdr.slid_mac_47_32 = cpu_to_be16(wc->slid);
603 	}
604 
605 	ib_dma_sync_single_for_device(&dev->ib_dev,
606 				      tun_qp->tx_ring[tun_tx_ix].buf.map,
607 				      sizeof (struct mlx4_rcv_tunnel_mad),
608 				      DMA_TO_DEVICE);
609 
610 	list.addr = tun_qp->tx_ring[tun_tx_ix].buf.map;
611 	list.length = sizeof (struct mlx4_rcv_tunnel_mad);
612 	list.lkey = tun_ctx->pd->local_dma_lkey;
613 
614 	wr.ah = ah;
615 	wr.port_num = port;
616 	wr.remote_qkey = IB_QP_SET_QKEY;
617 	wr.remote_qpn = dqpn;
618 	wr.wr.next = NULL;
619 	wr.wr.wr_id = ((u64) tun_tx_ix) | MLX4_TUN_SET_WRID_QPN(dest_qpt);
620 	wr.wr.sg_list = &list;
621 	wr.wr.num_sge = 1;
622 	wr.wr.opcode = IB_WR_SEND;
623 	wr.wr.send_flags = IB_SEND_SIGNALED;
624 
625 	ret = ib_post_send(src_qp, &wr.wr, &bad_wr);
626 	if (!ret)
627 		return 0;
628  out:
629 	spin_lock(&tun_qp->tx_lock);
630 	tun_qp->tx_ix_tail++;
631 	spin_unlock(&tun_qp->tx_lock);
632 	tun_qp->tx_ring[tun_tx_ix].ah = NULL;
633 end:
634 	ib_destroy_ah(ah);
635 	return ret;
636 }
637 
638 static int mlx4_ib_demux_mad(struct ib_device *ibdev, u8 port,
639 			struct ib_wc *wc, struct ib_grh *grh,
640 			struct ib_mad *mad)
641 {
642 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
643 	int err, other_port;
644 	int slave = -1;
645 	u8 *slave_id;
646 	int is_eth = 0;
647 
648 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
649 		is_eth = 0;
650 	else
651 		is_eth = 1;
652 
653 	if (is_eth) {
654 		if (!(wc->wc_flags & IB_WC_GRH)) {
655 			mlx4_ib_warn(ibdev, "RoCE grh not present.\n");
656 			return -EINVAL;
657 		}
658 		if (mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_CM) {
659 			mlx4_ib_warn(ibdev, "RoCE mgmt class is not CM\n");
660 			return -EINVAL;
661 		}
662 		err = mlx4_get_slave_from_roce_gid(dev->dev, port, grh->dgid.raw, &slave);
663 		if (err && mlx4_is_mf_bonded(dev->dev)) {
664 			other_port = (port == 1) ? 2 : 1;
665 			err = mlx4_get_slave_from_roce_gid(dev->dev, other_port, grh->dgid.raw, &slave);
666 			if (!err) {
667 				port = other_port;
668 				pr_debug("resolved slave %d from gid %pI6 wire port %d other %d\n",
669 					 slave, grh->dgid.raw, port, other_port);
670 			}
671 		}
672 		if (err) {
673 			mlx4_ib_warn(ibdev, "failed matching grh\n");
674 			return -ENOENT;
675 		}
676 		if (slave >= dev->dev->caps.sqp_demux) {
677 			mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
678 				     slave, dev->dev->caps.sqp_demux);
679 			return -ENOENT;
680 		}
681 
682 		if (mlx4_ib_demux_cm_handler(ibdev, port, NULL, mad))
683 			return 0;
684 
685 		err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
686 		if (err)
687 			pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
688 				 slave, err);
689 		return 0;
690 	}
691 
692 	/* Initially assume that this mad is for us */
693 	slave = mlx4_master_func_num(dev->dev);
694 
695 	/* See if the slave id is encoded in a response mad */
696 	if (mad->mad_hdr.method & 0x80) {
697 		slave_id = (u8 *) &mad->mad_hdr.tid;
698 		slave = *slave_id;
699 		if (slave != 255) /*255 indicates the dom0*/
700 			*slave_id = 0; /* remap tid */
701 	}
702 
703 	/* If a grh is present, we demux according to it */
704 	if (wc->wc_flags & IB_WC_GRH) {
705 		slave = mlx4_ib_find_real_gid(ibdev, port, grh->dgid.global.interface_id);
706 		if (slave < 0) {
707 			mlx4_ib_warn(ibdev, "failed matching grh\n");
708 			return -ENOENT;
709 		}
710 	}
711 	/* Class-specific handling */
712 	switch (mad->mad_hdr.mgmt_class) {
713 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
714 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
715 		/* 255 indicates the dom0 */
716 		if (slave != 255 && slave != mlx4_master_func_num(dev->dev)) {
717 			if (!mlx4_vf_smi_enabled(dev->dev, slave, port))
718 				return -EPERM;
719 			/* for a VF. drop unsolicited MADs */
720 			if (!(mad->mad_hdr.method & IB_MGMT_METHOD_RESP)) {
721 				mlx4_ib_warn(ibdev, "demux QP0. rejecting unsolicited mad for slave %d class 0x%x, method 0x%x\n",
722 					     slave, mad->mad_hdr.mgmt_class,
723 					     mad->mad_hdr.method);
724 				return -EINVAL;
725 			}
726 		}
727 		break;
728 	case IB_MGMT_CLASS_SUBN_ADM:
729 		if (mlx4_ib_demux_sa_handler(ibdev, port, slave,
730 					     (struct ib_sa_mad *) mad))
731 			return 0;
732 		break;
733 	case IB_MGMT_CLASS_CM:
734 		if (mlx4_ib_demux_cm_handler(ibdev, port, &slave, mad))
735 			return 0;
736 		break;
737 	case IB_MGMT_CLASS_DEVICE_MGMT:
738 		if (mad->mad_hdr.method != IB_MGMT_METHOD_GET_RESP)
739 			return 0;
740 		break;
741 	default:
742 		/* Drop unsupported classes for slaves in tunnel mode */
743 		if (slave != mlx4_master_func_num(dev->dev)) {
744 			pr_debug("dropping unsupported ingress mad from class:%d "
745 				 "for slave:%d\n", mad->mad_hdr.mgmt_class, slave);
746 			return 0;
747 		}
748 	}
749 	/*make sure that no slave==255 was not handled yet.*/
750 	if (slave >= dev->dev->caps.sqp_demux) {
751 		mlx4_ib_warn(ibdev, "slave id: %d is bigger than allowed:%d\n",
752 			     slave, dev->dev->caps.sqp_demux);
753 		return -ENOENT;
754 	}
755 
756 	err = mlx4_ib_send_to_slave(dev, slave, port, wc->qp->qp_type, wc, grh, mad);
757 	if (err)
758 		pr_debug("failed sending to slave %d via tunnel qp (%d)\n",
759 			 slave, err);
760 	return 0;
761 }
762 
763 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
764 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
765 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
766 {
767 	u16 slid, prev_lid = 0;
768 	int err;
769 	struct ib_port_attr pattr;
770 
771 	if (in_wc && in_wc->qp->qp_num) {
772 		pr_debug("received MAD: slid:%d sqpn:%d "
773 			"dlid_bits:%d dqpn:%d wc_flags:0x%x, cls %x, mtd %x, atr %x\n",
774 			in_wc->slid, in_wc->src_qp,
775 			in_wc->dlid_path_bits,
776 			in_wc->qp->qp_num,
777 			in_wc->wc_flags,
778 			in_mad->mad_hdr.mgmt_class, in_mad->mad_hdr.method,
779 			be16_to_cpu(in_mad->mad_hdr.attr_id));
780 		if (in_wc->wc_flags & IB_WC_GRH) {
781 			pr_debug("sgid_hi:0x%016llx sgid_lo:0x%016llx\n",
782 				 be64_to_cpu(in_grh->sgid.global.subnet_prefix),
783 				 be64_to_cpu(in_grh->sgid.global.interface_id));
784 			pr_debug("dgid_hi:0x%016llx dgid_lo:0x%016llx\n",
785 				 be64_to_cpu(in_grh->dgid.global.subnet_prefix),
786 				 be64_to_cpu(in_grh->dgid.global.interface_id));
787 		}
788 	}
789 
790 	slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
791 
792 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
793 		forward_trap(to_mdev(ibdev), port_num, in_mad);
794 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
795 	}
796 
797 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
798 	    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
799 		if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
800 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
801 		    in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
802 			return IB_MAD_RESULT_SUCCESS;
803 
804 		/*
805 		 * Don't process SMInfo queries -- the SMA can't handle them.
806 		 */
807 		if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
808 			return IB_MAD_RESULT_SUCCESS;
809 	} else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
810 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
811 		   in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
812 		   in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
813 		if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
814 		    in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
815 			return IB_MAD_RESULT_SUCCESS;
816 	} else
817 		return IB_MAD_RESULT_SUCCESS;
818 
819 	if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
820 	     in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
821 	    in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
822 	    in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
823 	    !ib_query_port(ibdev, port_num, &pattr))
824 		prev_lid = pattr.lid;
825 
826 	err = mlx4_MAD_IFC(to_mdev(ibdev),
827 			   (mad_flags & IB_MAD_IGNORE_MKEY ? MLX4_MAD_IFC_IGNORE_MKEY : 0) |
828 			   (mad_flags & IB_MAD_IGNORE_BKEY ? MLX4_MAD_IFC_IGNORE_BKEY : 0) |
829 			   MLX4_MAD_IFC_NET_VIEW,
830 			   port_num, in_wc, in_grh, in_mad, out_mad);
831 	if (err)
832 		return IB_MAD_RESULT_FAILURE;
833 
834 	if (!out_mad->mad_hdr.status) {
835 		smp_snoop(ibdev, port_num, in_mad, prev_lid);
836 		/* slaves get node desc from FW */
837 		if (!mlx4_is_slave(to_mdev(ibdev)->dev))
838 			node_desc_override(ibdev, out_mad);
839 	}
840 
841 	/* set return bit in status of directed route responses */
842 	if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
843 		out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
844 
845 	if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
846 		/* no response for trap repress */
847 		return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
848 
849 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
850 }
851 
852 static void edit_counter(struct mlx4_counter *cnt, void *counters,
853 			 __be16 attr_id)
854 {
855 	switch (attr_id) {
856 	case IB_PMA_PORT_COUNTERS:
857 	{
858 		struct ib_pma_portcounters *pma_cnt =
859 			(struct ib_pma_portcounters *)counters;
860 
861 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
862 				     (be64_to_cpu(cnt->tx_bytes) >> 2));
863 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
864 				     (be64_to_cpu(cnt->rx_bytes) >> 2));
865 		ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
866 				     be64_to_cpu(cnt->tx_frames));
867 		ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
868 				     be64_to_cpu(cnt->rx_frames));
869 		break;
870 	}
871 	case IB_PMA_PORT_COUNTERS_EXT:
872 	{
873 		struct ib_pma_portcounters_ext *pma_cnt_ext =
874 			(struct ib_pma_portcounters_ext *)counters;
875 
876 		pma_cnt_ext->port_xmit_data =
877 			cpu_to_be64(be64_to_cpu(cnt->tx_bytes) >> 2);
878 		pma_cnt_ext->port_rcv_data =
879 			cpu_to_be64(be64_to_cpu(cnt->rx_bytes) >> 2);
880 		pma_cnt_ext->port_xmit_packets = cnt->tx_frames;
881 		pma_cnt_ext->port_rcv_packets = cnt->rx_frames;
882 		break;
883 	}
884 	}
885 }
886 
887 static int iboe_process_mad_port_info(void *out_mad)
888 {
889 	struct ib_class_port_info cpi = {};
890 
891 	cpi.capability_mask = IB_PMA_CLASS_CAP_EXT_WIDTH;
892 	memcpy(out_mad, &cpi, sizeof(cpi));
893 	return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
894 }
895 
896 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
897 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
898 			const struct ib_mad *in_mad, struct ib_mad *out_mad)
899 {
900 	struct mlx4_counter counter_stats;
901 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
902 	struct counter_index *tmp_counter;
903 	int err = IB_MAD_RESULT_FAILURE, stats_avail = 0;
904 
905 	if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
906 		return -EINVAL;
907 
908 	if (in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)
909 		return iboe_process_mad_port_info((void *)(out_mad->data + 40));
910 
911 	memset(&counter_stats, 0, sizeof(counter_stats));
912 	mutex_lock(&dev->counters_table[port_num - 1].mutex);
913 	list_for_each_entry(tmp_counter,
914 			    &dev->counters_table[port_num - 1].counters_list,
915 			    list) {
916 		err = mlx4_get_counter_stats(dev->dev,
917 					     tmp_counter->index,
918 					     &counter_stats, 0);
919 		if (err) {
920 			err = IB_MAD_RESULT_FAILURE;
921 			stats_avail = 0;
922 			break;
923 		}
924 		stats_avail = 1;
925 	}
926 	mutex_unlock(&dev->counters_table[port_num - 1].mutex);
927 	if (stats_avail) {
928 		memset(out_mad->data, 0, sizeof out_mad->data);
929 		switch (counter_stats.counter_mode & 0xf) {
930 		case 0:
931 			edit_counter(&counter_stats,
932 				     (void *)(out_mad->data + 40),
933 				     in_mad->mad_hdr.attr_id);
934 			err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
935 			break;
936 		default:
937 			err = IB_MAD_RESULT_FAILURE;
938 		}
939 	}
940 
941 	return err;
942 }
943 
944 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
945 			const struct ib_wc *in_wc, const struct ib_grh *in_grh,
946 			const struct ib_mad_hdr *in, size_t in_mad_size,
947 			struct ib_mad_hdr *out, size_t *out_mad_size,
948 			u16 *out_mad_pkey_index)
949 {
950 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
951 	const struct ib_mad *in_mad = (const struct ib_mad *)in;
952 	struct ib_mad *out_mad = (struct ib_mad *)out;
953 	enum rdma_link_layer link = rdma_port_get_link_layer(ibdev, port_num);
954 
955 	if (WARN_ON_ONCE(in_mad_size != sizeof(*in_mad) ||
956 			 *out_mad_size != sizeof(*out_mad)))
957 		return IB_MAD_RESULT_FAILURE;
958 
959 	/* iboe_process_mad() which uses the HCA flow-counters to implement IB PMA
960 	 * queries, should be called only by VFs and for that specific purpose
961 	 */
962 	if (link == IB_LINK_LAYER_INFINIBAND) {
963 		if (mlx4_is_slave(dev->dev) &&
964 		    (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT &&
965 		     (in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS ||
966 		      in_mad->mad_hdr.attr_id == IB_PMA_PORT_COUNTERS_EXT ||
967 		      in_mad->mad_hdr.attr_id == IB_PMA_CLASS_PORT_INFO)))
968 			return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
969 						in_grh, in_mad, out_mad);
970 
971 		return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
972 				      in_grh, in_mad, out_mad);
973 	}
974 
975 	if (link == IB_LINK_LAYER_ETHERNET)
976 		return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
977 					in_grh, in_mad, out_mad);
978 
979 	return -EINVAL;
980 }
981 
982 static void send_handler(struct ib_mad_agent *agent,
983 			 struct ib_mad_send_wc *mad_send_wc)
984 {
985 	if (mad_send_wc->send_buf->context[0])
986 		ib_destroy_ah(mad_send_wc->send_buf->context[0]);
987 	ib_free_send_mad(mad_send_wc->send_buf);
988 }
989 
990 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
991 {
992 	struct ib_mad_agent *agent;
993 	int p, q;
994 	int ret;
995 	enum rdma_link_layer ll;
996 
997 	for (p = 0; p < dev->num_ports; ++p) {
998 		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
999 		for (q = 0; q <= 1; ++q) {
1000 			if (ll == IB_LINK_LAYER_INFINIBAND) {
1001 				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
1002 							      q ? IB_QPT_GSI : IB_QPT_SMI,
1003 							      NULL, 0, send_handler,
1004 							      NULL, NULL, 0);
1005 				if (IS_ERR(agent)) {
1006 					ret = PTR_ERR(agent);
1007 					goto err;
1008 				}
1009 				dev->send_agent[p][q] = agent;
1010 			} else
1011 				dev->send_agent[p][q] = NULL;
1012 		}
1013 	}
1014 
1015 	return 0;
1016 
1017 err:
1018 	for (p = 0; p < dev->num_ports; ++p)
1019 		for (q = 0; q <= 1; ++q)
1020 			if (dev->send_agent[p][q])
1021 				ib_unregister_mad_agent(dev->send_agent[p][q]);
1022 
1023 	return ret;
1024 }
1025 
1026 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
1027 {
1028 	struct ib_mad_agent *agent;
1029 	int p, q;
1030 
1031 	for (p = 0; p < dev->num_ports; ++p) {
1032 		for (q = 0; q <= 1; ++q) {
1033 			agent = dev->send_agent[p][q];
1034 			if (agent) {
1035 				dev->send_agent[p][q] = NULL;
1036 				ib_unregister_mad_agent(agent);
1037 			}
1038 		}
1039 
1040 		if (dev->sm_ah[p])
1041 			ib_destroy_ah(dev->sm_ah[p]);
1042 	}
1043 }
1044 
1045 static void handle_lid_change_event(struct mlx4_ib_dev *dev, u8 port_num)
1046 {
1047 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_LID_CHANGE);
1048 
1049 	if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1050 		mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1051 					    MLX4_EQ_PORT_INFO_LID_CHANGE_MASK);
1052 }
1053 
1054 static void handle_client_rereg_event(struct mlx4_ib_dev *dev, u8 port_num)
1055 {
1056 	/* re-configure the alias-guid and mcg's */
1057 	if (mlx4_is_master(dev->dev)) {
1058 		mlx4_ib_invalidate_all_guid_record(dev, port_num);
1059 
1060 		if (!dev->sriov.is_going_down) {
1061 			mlx4_ib_mcg_port_cleanup(&dev->sriov.demux[port_num - 1], 0);
1062 			mlx4_gen_slaves_port_mgt_ev(dev->dev, port_num,
1063 						    MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK);
1064 		}
1065 	}
1066 
1067 	/* Update the sl to vl table from inside client rereg
1068 	 * only if in secure-host mode (snooping is not possible)
1069 	 * and the sl-to-vl change event is not generated by FW.
1070 	 */
1071 	if (!mlx4_is_slave(dev->dev) &&
1072 	    dev->dev->flags & MLX4_FLAG_SECURE_HOST &&
1073 	    !(dev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT)) {
1074 		if (mlx4_is_master(dev->dev))
1075 			/* already in work queue from mlx4_ib_event queueing
1076 			 * mlx4_handle_port_mgmt_change_event, which calls
1077 			 * this procedure. Therefore, call sl2vl_update directly.
1078 			 */
1079 			mlx4_ib_sl2vl_update(dev, port_num);
1080 		else
1081 			mlx4_sched_ib_sl2vl_update_work(dev, port_num);
1082 	}
1083 	mlx4_ib_dispatch_event(dev, port_num, IB_EVENT_CLIENT_REREGISTER);
1084 }
1085 
1086 static void propagate_pkey_ev(struct mlx4_ib_dev *dev, int port_num,
1087 			      struct mlx4_eqe *eqe)
1088 {
1089 	__propagate_pkey_ev(dev, port_num, GET_BLK_PTR_FROM_EQE(eqe),
1090 			    GET_MASK_FROM_EQE(eqe));
1091 }
1092 
1093 static void handle_slaves_guid_change(struct mlx4_ib_dev *dev, u8 port_num,
1094 				      u32 guid_tbl_blk_num, u32 change_bitmap)
1095 {
1096 	struct ib_smp *in_mad  = NULL;
1097 	struct ib_smp *out_mad  = NULL;
1098 	u16 i;
1099 
1100 	if (!mlx4_is_mfunc(dev->dev) || !mlx4_is_master(dev->dev))
1101 		return;
1102 
1103 	in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
1104 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1105 	if (!in_mad || !out_mad) {
1106 		mlx4_ib_warn(&dev->ib_dev, "failed to allocate memory for guid info mads\n");
1107 		goto out;
1108 	}
1109 
1110 	guid_tbl_blk_num  *= 4;
1111 
1112 	for (i = 0; i < 4; i++) {
1113 		if (change_bitmap && (!((change_bitmap >> (8 * i)) & 0xff)))
1114 			continue;
1115 		memset(in_mad, 0, sizeof *in_mad);
1116 		memset(out_mad, 0, sizeof *out_mad);
1117 
1118 		in_mad->base_version  = 1;
1119 		in_mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
1120 		in_mad->class_version = 1;
1121 		in_mad->method        = IB_MGMT_METHOD_GET;
1122 		in_mad->attr_id       = IB_SMP_ATTR_GUID_INFO;
1123 		in_mad->attr_mod      = cpu_to_be32(guid_tbl_blk_num + i);
1124 
1125 		if (mlx4_MAD_IFC(dev,
1126 				 MLX4_MAD_IFC_IGNORE_KEYS | MLX4_MAD_IFC_NET_VIEW,
1127 				 port_num, NULL, NULL, in_mad, out_mad)) {
1128 			mlx4_ib_warn(&dev->ib_dev, "Failed in get GUID INFO MAD_IFC\n");
1129 			goto out;
1130 		}
1131 
1132 		mlx4_ib_update_cache_on_guid_change(dev, guid_tbl_blk_num + i,
1133 						    port_num,
1134 						    (u8 *)(&((struct ib_smp *)out_mad)->data));
1135 		mlx4_ib_notify_slaves_on_guid_change(dev, guid_tbl_blk_num + i,
1136 						     port_num,
1137 						     (u8 *)(&((struct ib_smp *)out_mad)->data));
1138 	}
1139 
1140 out:
1141 	kfree(in_mad);
1142 	kfree(out_mad);
1143 	return;
1144 }
1145 
1146 void handle_port_mgmt_change_event(struct work_struct *work)
1147 {
1148 	struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
1149 	struct mlx4_ib_dev *dev = ew->ib_dev;
1150 	struct mlx4_eqe *eqe = &(ew->ib_eqe);
1151 	u8 port = eqe->event.port_mgmt_change.port;
1152 	u32 changed_attr;
1153 	u32 tbl_block;
1154 	u32 change_bitmap;
1155 
1156 	switch (eqe->subtype) {
1157 	case MLX4_DEV_PMC_SUBTYPE_PORT_INFO:
1158 		changed_attr = be32_to_cpu(eqe->event.port_mgmt_change.params.port_info.changed_attr);
1159 
1160 		/* Update the SM ah - This should be done before handling
1161 		   the other changed attributes so that MADs can be sent to the SM */
1162 		if (changed_attr & MSTR_SM_CHANGE_MASK) {
1163 			u16 lid = be16_to_cpu(eqe->event.port_mgmt_change.params.port_info.mstr_sm_lid);
1164 			u8 sl = eqe->event.port_mgmt_change.params.port_info.mstr_sm_sl & 0xf;
1165 			update_sm_ah(dev, port, lid, sl);
1166 		}
1167 
1168 		/* Check if it is a lid change event */
1169 		if (changed_attr & MLX4_EQ_PORT_INFO_LID_CHANGE_MASK)
1170 			handle_lid_change_event(dev, port);
1171 
1172 		/* Generate GUID changed event */
1173 		if (changed_attr & MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK) {
1174 			if (mlx4_is_master(dev->dev)) {
1175 				union ib_gid gid;
1176 				int err = 0;
1177 
1178 				if (!eqe->event.port_mgmt_change.params.port_info.gid_prefix)
1179 					err = __mlx4_ib_query_gid(&dev->ib_dev, port, 0, &gid, 1);
1180 				else
1181 					gid.global.subnet_prefix =
1182 						eqe->event.port_mgmt_change.params.port_info.gid_prefix;
1183 				if (err) {
1184 					pr_warn("Could not change QP1 subnet prefix for port %d: query_gid error (%d)\n",
1185 						port, err);
1186 				} else {
1187 					pr_debug("Changing QP1 subnet prefix for port %d. old=0x%llx. new=0x%llx\n",
1188 						 port,
1189 						 (u64)atomic64_read(&dev->sriov.demux[port - 1].subnet_prefix),
1190 						 be64_to_cpu(gid.global.subnet_prefix));
1191 					atomic64_set(&dev->sriov.demux[port - 1].subnet_prefix,
1192 						     be64_to_cpu(gid.global.subnet_prefix));
1193 				}
1194 			}
1195 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1196 			/*if master, notify all slaves*/
1197 			if (mlx4_is_master(dev->dev))
1198 				mlx4_gen_slaves_port_mgt_ev(dev->dev, port,
1199 							    MLX4_EQ_PORT_INFO_GID_PFX_CHANGE_MASK);
1200 		}
1201 
1202 		if (changed_attr & MLX4_EQ_PORT_INFO_CLIENT_REREG_MASK)
1203 			handle_client_rereg_event(dev, port);
1204 		break;
1205 
1206 	case MLX4_DEV_PMC_SUBTYPE_PKEY_TABLE:
1207 		mlx4_ib_dispatch_event(dev, port, IB_EVENT_PKEY_CHANGE);
1208 		if (mlx4_is_master(dev->dev) && !dev->sriov.is_going_down)
1209 			propagate_pkey_ev(dev, port, eqe);
1210 		break;
1211 	case MLX4_DEV_PMC_SUBTYPE_GUID_INFO:
1212 		/* paravirtualized master's guid is guid 0 -- does not change */
1213 		if (!mlx4_is_master(dev->dev))
1214 			mlx4_ib_dispatch_event(dev, port, IB_EVENT_GID_CHANGE);
1215 		/*if master, notify relevant slaves*/
1216 		else if (!dev->sriov.is_going_down) {
1217 			tbl_block = GET_BLK_PTR_FROM_EQE(eqe);
1218 			change_bitmap = GET_MASK_FROM_EQE(eqe);
1219 			handle_slaves_guid_change(dev, port, tbl_block, change_bitmap);
1220 		}
1221 		break;
1222 
1223 	case MLX4_DEV_PMC_SUBTYPE_SL_TO_VL_MAP:
1224 		/* cache sl to vl mapping changes for use in
1225 		 * filling QP1 LRH VL field when sending packets
1226 		 */
1227 		if (!mlx4_is_slave(dev->dev)) {
1228 			union sl2vl_tbl_to_u64 sl2vl64;
1229 			int jj;
1230 
1231 			for (jj = 0; jj < 8; jj++) {
1232 				sl2vl64.sl8[jj] =
1233 					eqe->event.port_mgmt_change.params.sl2vl_tbl_change_info.sl2vl_table[jj];
1234 				pr_debug("port %u, sl2vl[%d] = %02x\n",
1235 					 port, jj, sl2vl64.sl8[jj]);
1236 			}
1237 			atomic64_set(&dev->sl2vl[port - 1], sl2vl64.sl64);
1238 		}
1239 		break;
1240 	default:
1241 		pr_warn("Unsupported subtype 0x%x for "
1242 			"Port Management Change event\n", eqe->subtype);
1243 	}
1244 
1245 	kfree(ew);
1246 }
1247 
1248 void mlx4_ib_dispatch_event(struct mlx4_ib_dev *dev, u8 port_num,
1249 			    enum ib_event_type type)
1250 {
1251 	struct ib_event event;
1252 
1253 	event.device		= &dev->ib_dev;
1254 	event.element.port_num	= port_num;
1255 	event.event		= type;
1256 
1257 	ib_dispatch_event(&event);
1258 }
1259 
1260 static void mlx4_ib_tunnel_comp_handler(struct ib_cq *cq, void *arg)
1261 {
1262 	unsigned long flags;
1263 	struct mlx4_ib_demux_pv_ctx *ctx = cq->cq_context;
1264 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1265 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
1266 	if (!dev->sriov.is_going_down && ctx->state == DEMUX_PV_STATE_ACTIVE)
1267 		queue_work(ctx->wq, &ctx->work);
1268 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
1269 }
1270 
1271 static int mlx4_ib_post_pv_qp_buf(struct mlx4_ib_demux_pv_ctx *ctx,
1272 				  struct mlx4_ib_demux_pv_qp *tun_qp,
1273 				  int index)
1274 {
1275 	struct ib_sge sg_list;
1276 	struct ib_recv_wr recv_wr, *bad_recv_wr;
1277 	int size;
1278 
1279 	size = (tun_qp->qp->qp_type == IB_QPT_UD) ?
1280 		sizeof (struct mlx4_tunnel_mad) : sizeof (struct mlx4_mad_rcv_buf);
1281 
1282 	sg_list.addr = tun_qp->ring[index].map;
1283 	sg_list.length = size;
1284 	sg_list.lkey = ctx->pd->local_dma_lkey;
1285 
1286 	recv_wr.next = NULL;
1287 	recv_wr.sg_list = &sg_list;
1288 	recv_wr.num_sge = 1;
1289 	recv_wr.wr_id = (u64) index | MLX4_TUN_WRID_RECV |
1290 		MLX4_TUN_SET_WRID_QPN(tun_qp->proxy_qpt);
1291 	ib_dma_sync_single_for_device(ctx->ib_dev, tun_qp->ring[index].map,
1292 				      size, DMA_FROM_DEVICE);
1293 	return ib_post_recv(tun_qp->qp, &recv_wr, &bad_recv_wr);
1294 }
1295 
1296 static int mlx4_ib_multiplex_sa_handler(struct ib_device *ibdev, int port,
1297 		int slave, struct ib_sa_mad *sa_mad)
1298 {
1299 	int ret = 0;
1300 
1301 	/* dispatch to different sa handlers */
1302 	switch (be16_to_cpu(sa_mad->mad_hdr.attr_id)) {
1303 	case IB_SA_ATTR_MC_MEMBER_REC:
1304 		ret = mlx4_ib_mcg_multiplex_handler(ibdev, port, slave, sa_mad);
1305 		break;
1306 	default:
1307 		break;
1308 	}
1309 	return ret;
1310 }
1311 
1312 static int is_proxy_qp0(struct mlx4_ib_dev *dev, int qpn, int slave)
1313 {
1314 	int proxy_start = dev->dev->phys_caps.base_proxy_sqpn + 8 * slave;
1315 
1316 	return (qpn >= proxy_start && qpn <= proxy_start + 1);
1317 }
1318 
1319 
1320 int mlx4_ib_send_to_wire(struct mlx4_ib_dev *dev, int slave, u8 port,
1321 			 enum ib_qp_type dest_qpt, u16 pkey_index,
1322 			 u32 remote_qpn, u32 qkey, struct ib_ah_attr *attr,
1323 			 u8 *s_mac, u16 vlan_id, struct ib_mad *mad)
1324 {
1325 	struct ib_sge list;
1326 	struct ib_ud_wr wr;
1327 	struct ib_send_wr *bad_wr;
1328 	struct mlx4_ib_demux_pv_ctx *sqp_ctx;
1329 	struct mlx4_ib_demux_pv_qp *sqp;
1330 	struct mlx4_mad_snd_buf *sqp_mad;
1331 	struct ib_ah *ah;
1332 	struct ib_qp *send_qp = NULL;
1333 	unsigned wire_tx_ix = 0;
1334 	int ret = 0;
1335 	u16 wire_pkey_ix;
1336 	int src_qpnum;
1337 	u8 sgid_index;
1338 
1339 
1340 	sqp_ctx = dev->sriov.sqps[port-1];
1341 
1342 	/* check if proxy qp created */
1343 	if (!sqp_ctx || sqp_ctx->state != DEMUX_PV_STATE_ACTIVE)
1344 		return -EAGAIN;
1345 
1346 	if (dest_qpt == IB_QPT_SMI) {
1347 		src_qpnum = 0;
1348 		sqp = &sqp_ctx->qp[0];
1349 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][0];
1350 	} else {
1351 		src_qpnum = 1;
1352 		sqp = &sqp_ctx->qp[1];
1353 		wire_pkey_ix = dev->pkeys.virt2phys_pkey[slave][port - 1][pkey_index];
1354 	}
1355 
1356 	send_qp = sqp->qp;
1357 
1358 	/* create ah */
1359 	sgid_index = attr->grh.sgid_index;
1360 	attr->grh.sgid_index = 0;
1361 	ah = ib_create_ah(sqp_ctx->pd, attr);
1362 	if (IS_ERR(ah))
1363 		return -ENOMEM;
1364 	attr->grh.sgid_index = sgid_index;
1365 	to_mah(ah)->av.ib.gid_index = sgid_index;
1366 	/* get rid of force-loopback bit */
1367 	to_mah(ah)->av.ib.port_pd &= cpu_to_be32(0x7FFFFFFF);
1368 	spin_lock(&sqp->tx_lock);
1369 	if (sqp->tx_ix_head - sqp->tx_ix_tail >=
1370 	    (MLX4_NUM_TUNNEL_BUFS - 1))
1371 		ret = -EAGAIN;
1372 	else
1373 		wire_tx_ix = (++sqp->tx_ix_head) & (MLX4_NUM_TUNNEL_BUFS - 1);
1374 	spin_unlock(&sqp->tx_lock);
1375 	if (ret)
1376 		goto out;
1377 
1378 	sqp_mad = (struct mlx4_mad_snd_buf *) (sqp->tx_ring[wire_tx_ix].buf.addr);
1379 	if (sqp->tx_ring[wire_tx_ix].ah)
1380 		ib_destroy_ah(sqp->tx_ring[wire_tx_ix].ah);
1381 	sqp->tx_ring[wire_tx_ix].ah = ah;
1382 	ib_dma_sync_single_for_cpu(&dev->ib_dev,
1383 				   sqp->tx_ring[wire_tx_ix].buf.map,
1384 				   sizeof (struct mlx4_mad_snd_buf),
1385 				   DMA_TO_DEVICE);
1386 
1387 	memcpy(&sqp_mad->payload, mad, sizeof *mad);
1388 
1389 	ib_dma_sync_single_for_device(&dev->ib_dev,
1390 				      sqp->tx_ring[wire_tx_ix].buf.map,
1391 				      sizeof (struct mlx4_mad_snd_buf),
1392 				      DMA_TO_DEVICE);
1393 
1394 	list.addr = sqp->tx_ring[wire_tx_ix].buf.map;
1395 	list.length = sizeof (struct mlx4_mad_snd_buf);
1396 	list.lkey = sqp_ctx->pd->local_dma_lkey;
1397 
1398 	wr.ah = ah;
1399 	wr.port_num = port;
1400 	wr.pkey_index = wire_pkey_ix;
1401 	wr.remote_qkey = qkey;
1402 	wr.remote_qpn = remote_qpn;
1403 	wr.wr.next = NULL;
1404 	wr.wr.wr_id = ((u64) wire_tx_ix) | MLX4_TUN_SET_WRID_QPN(src_qpnum);
1405 	wr.wr.sg_list = &list;
1406 	wr.wr.num_sge = 1;
1407 	wr.wr.opcode = IB_WR_SEND;
1408 	wr.wr.send_flags = IB_SEND_SIGNALED;
1409 	if (s_mac)
1410 		memcpy(to_mah(ah)->av.eth.s_mac, s_mac, 6);
1411 	if (vlan_id < 0x1000)
1412 		vlan_id |= (attr->sl & 7) << 13;
1413 	to_mah(ah)->av.eth.vlan = cpu_to_be16(vlan_id);
1414 
1415 
1416 	ret = ib_post_send(send_qp, &wr.wr, &bad_wr);
1417 	if (!ret)
1418 		return 0;
1419 
1420 	spin_lock(&sqp->tx_lock);
1421 	sqp->tx_ix_tail++;
1422 	spin_unlock(&sqp->tx_lock);
1423 	sqp->tx_ring[wire_tx_ix].ah = NULL;
1424 out:
1425 	ib_destroy_ah(ah);
1426 	return ret;
1427 }
1428 
1429 static int get_slave_base_gid_ix(struct mlx4_ib_dev *dev, int slave, int port)
1430 {
1431 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1432 		return slave;
1433 	return mlx4_get_base_gid_ix(dev->dev, slave, port);
1434 }
1435 
1436 static void fill_in_real_sgid_index(struct mlx4_ib_dev *dev, int slave, int port,
1437 				    struct ib_ah_attr *ah_attr)
1438 {
1439 	if (rdma_port_get_link_layer(&dev->ib_dev, port) == IB_LINK_LAYER_INFINIBAND)
1440 		ah_attr->grh.sgid_index = slave;
1441 	else
1442 		ah_attr->grh.sgid_index += get_slave_base_gid_ix(dev, slave, port);
1443 }
1444 
1445 static void mlx4_ib_multiplex_mad(struct mlx4_ib_demux_pv_ctx *ctx, struct ib_wc *wc)
1446 {
1447 	struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
1448 	struct mlx4_ib_demux_pv_qp *tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc->wr_id)];
1449 	int wr_ix = wc->wr_id & (MLX4_NUM_TUNNEL_BUFS - 1);
1450 	struct mlx4_tunnel_mad *tunnel = tun_qp->ring[wr_ix].addr;
1451 	struct mlx4_ib_ah ah;
1452 	struct ib_ah_attr ah_attr;
1453 	u8 *slave_id;
1454 	int slave;
1455 	int port;
1456 	u16 vlan_id;
1457 
1458 	/* Get slave that sent this packet */
1459 	if (wc->src_qp < dev->dev->phys_caps.base_proxy_sqpn ||
1460 	    wc->src_qp >= dev->dev->phys_caps.base_proxy_sqpn + 8 * MLX4_MFUNC_MAX ||
1461 	    (wc->src_qp & 0x1) != ctx->port - 1 ||
1462 	    wc->src_qp & 0x4) {
1463 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d\n", wc->src_qp);
1464 		return;
1465 	}
1466 	slave = ((wc->src_qp & ~0x7) - dev->dev->phys_caps.base_proxy_sqpn) / 8;
1467 	if (slave != ctx->slave) {
1468 		mlx4_ib_warn(ctx->ib_dev, "can't multiplex bad sqp:%d: "
1469 			     "belongs to another slave\n", wc->src_qp);
1470 		return;
1471 	}
1472 
1473 	/* Map transaction ID */
1474 	ib_dma_sync_single_for_cpu(ctx->ib_dev, tun_qp->ring[wr_ix].map,
1475 				   sizeof (struct mlx4_tunnel_mad),
1476 				   DMA_FROM_DEVICE);
1477 	switch (tunnel->mad.mad_hdr.method) {
1478 	case IB_MGMT_METHOD_SET:
1479 	case IB_MGMT_METHOD_GET:
1480 	case IB_MGMT_METHOD_REPORT:
1481 	case IB_SA_METHOD_GET_TABLE:
1482 	case IB_SA_METHOD_DELETE:
1483 	case IB_SA_METHOD_GET_MULTI:
1484 	case IB_SA_METHOD_GET_TRACE_TBL:
1485 		slave_id = (u8 *) &tunnel->mad.mad_hdr.tid;
1486 		if (*slave_id) {
1487 			mlx4_ib_warn(ctx->ib_dev, "egress mad has non-null tid msb:%d "
1488 				     "class:%d slave:%d\n", *slave_id,
1489 				     tunnel->mad.mad_hdr.mgmt_class, slave);
1490 			return;
1491 		} else
1492 			*slave_id = slave;
1493 	default:
1494 		/* nothing */;
1495 	}
1496 
1497 	/* Class-specific handling */
1498 	switch (tunnel->mad.mad_hdr.mgmt_class) {
1499 	case IB_MGMT_CLASS_SUBN_LID_ROUTED:
1500 	case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
1501 		if (slave != mlx4_master_func_num(dev->dev) &&
1502 		    !mlx4_vf_smi_enabled(dev->dev, slave, ctx->port))
1503 			return;
1504 		break;
1505 	case IB_MGMT_CLASS_SUBN_ADM:
1506 		if (mlx4_ib_multiplex_sa_handler(ctx->ib_dev, ctx->port, slave,
1507 			      (struct ib_sa_mad *) &tunnel->mad))
1508 			return;
1509 		break;
1510 	case IB_MGMT_CLASS_CM:
1511 		if (mlx4_ib_multiplex_cm_handler(ctx->ib_dev, ctx->port, slave,
1512 			      (struct ib_mad *) &tunnel->mad))
1513 			return;
1514 		break;
1515 	case IB_MGMT_CLASS_DEVICE_MGMT:
1516 		if (tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_GET &&
1517 		    tunnel->mad.mad_hdr.method != IB_MGMT_METHOD_SET)
1518 			return;
1519 		break;
1520 	default:
1521 		/* Drop unsupported classes for slaves in tunnel mode */
1522 		if (slave != mlx4_master_func_num(dev->dev)) {
1523 			mlx4_ib_warn(ctx->ib_dev, "dropping unsupported egress mad from class:%d "
1524 				     "for slave:%d\n", tunnel->mad.mad_hdr.mgmt_class, slave);
1525 			return;
1526 		}
1527 	}
1528 
1529 	/* We are using standard ib_core services to send the mad, so generate a
1530 	 * stadard address handle by decoding the tunnelled mlx4_ah fields */
1531 	memcpy(&ah.av, &tunnel->hdr.av, sizeof (struct mlx4_av));
1532 	ah.ibah.device = ctx->ib_dev;
1533 
1534 	port = be32_to_cpu(ah.av.ib.port_pd) >> 24;
1535 	port = mlx4_slave_convert_port(dev->dev, slave, port);
1536 	if (port < 0)
1537 		return;
1538 	ah.av.ib.port_pd = cpu_to_be32(port << 24 | (be32_to_cpu(ah.av.ib.port_pd) & 0xffffff));
1539 
1540 	mlx4_ib_query_ah(&ah.ibah, &ah_attr);
1541 	if (ah_attr.ah_flags & IB_AH_GRH)
1542 		fill_in_real_sgid_index(dev, slave, ctx->port, &ah_attr);
1543 
1544 	memcpy(ah_attr.dmac, tunnel->hdr.mac, 6);
1545 	vlan_id = be16_to_cpu(tunnel->hdr.vlan);
1546 	/* if slave have default vlan use it */
1547 	mlx4_get_slave_default_vlan(dev->dev, ctx->port, slave,
1548 				    &vlan_id, &ah_attr.sl);
1549 
1550 	mlx4_ib_send_to_wire(dev, slave, ctx->port,
1551 			     is_proxy_qp0(dev, wc->src_qp, slave) ?
1552 			     IB_QPT_SMI : IB_QPT_GSI,
1553 			     be16_to_cpu(tunnel->hdr.pkey_index),
1554 			     be32_to_cpu(tunnel->hdr.remote_qpn),
1555 			     be32_to_cpu(tunnel->hdr.qkey),
1556 			     &ah_attr, wc->smac, vlan_id, &tunnel->mad);
1557 }
1558 
1559 static int mlx4_ib_alloc_pv_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1560 				 enum ib_qp_type qp_type, int is_tun)
1561 {
1562 	int i;
1563 	struct mlx4_ib_demux_pv_qp *tun_qp;
1564 	int rx_buf_size, tx_buf_size;
1565 
1566 	if (qp_type > IB_QPT_GSI)
1567 		return -EINVAL;
1568 
1569 	tun_qp = &ctx->qp[qp_type];
1570 
1571 	tun_qp->ring = kzalloc(sizeof (struct mlx4_ib_buf) * MLX4_NUM_TUNNEL_BUFS,
1572 			       GFP_KERNEL);
1573 	if (!tun_qp->ring)
1574 		return -ENOMEM;
1575 
1576 	tun_qp->tx_ring = kcalloc(MLX4_NUM_TUNNEL_BUFS,
1577 				  sizeof (struct mlx4_ib_tun_tx_buf),
1578 				  GFP_KERNEL);
1579 	if (!tun_qp->tx_ring) {
1580 		kfree(tun_qp->ring);
1581 		tun_qp->ring = NULL;
1582 		return -ENOMEM;
1583 	}
1584 
1585 	if (is_tun) {
1586 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1587 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1588 	} else {
1589 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1590 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1591 	}
1592 
1593 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1594 		tun_qp->ring[i].addr = kmalloc(rx_buf_size, GFP_KERNEL);
1595 		if (!tun_qp->ring[i].addr)
1596 			goto err;
1597 		tun_qp->ring[i].map = ib_dma_map_single(ctx->ib_dev,
1598 							tun_qp->ring[i].addr,
1599 							rx_buf_size,
1600 							DMA_FROM_DEVICE);
1601 		if (ib_dma_mapping_error(ctx->ib_dev, tun_qp->ring[i].map)) {
1602 			kfree(tun_qp->ring[i].addr);
1603 			goto err;
1604 		}
1605 	}
1606 
1607 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1608 		tun_qp->tx_ring[i].buf.addr =
1609 			kmalloc(tx_buf_size, GFP_KERNEL);
1610 		if (!tun_qp->tx_ring[i].buf.addr)
1611 			goto tx_err;
1612 		tun_qp->tx_ring[i].buf.map =
1613 			ib_dma_map_single(ctx->ib_dev,
1614 					  tun_qp->tx_ring[i].buf.addr,
1615 					  tx_buf_size,
1616 					  DMA_TO_DEVICE);
1617 		if (ib_dma_mapping_error(ctx->ib_dev,
1618 					 tun_qp->tx_ring[i].buf.map)) {
1619 			kfree(tun_qp->tx_ring[i].buf.addr);
1620 			goto tx_err;
1621 		}
1622 		tun_qp->tx_ring[i].ah = NULL;
1623 	}
1624 	spin_lock_init(&tun_qp->tx_lock);
1625 	tun_qp->tx_ix_head = 0;
1626 	tun_qp->tx_ix_tail = 0;
1627 	tun_qp->proxy_qpt = qp_type;
1628 
1629 	return 0;
1630 
1631 tx_err:
1632 	while (i > 0) {
1633 		--i;
1634 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1635 				    tx_buf_size, DMA_TO_DEVICE);
1636 		kfree(tun_qp->tx_ring[i].buf.addr);
1637 	}
1638 	kfree(tun_qp->tx_ring);
1639 	tun_qp->tx_ring = NULL;
1640 	i = MLX4_NUM_TUNNEL_BUFS;
1641 err:
1642 	while (i > 0) {
1643 		--i;
1644 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1645 				    rx_buf_size, DMA_FROM_DEVICE);
1646 		kfree(tun_qp->ring[i].addr);
1647 	}
1648 	kfree(tun_qp->ring);
1649 	tun_qp->ring = NULL;
1650 	return -ENOMEM;
1651 }
1652 
1653 static void mlx4_ib_free_pv_qp_bufs(struct mlx4_ib_demux_pv_ctx *ctx,
1654 				     enum ib_qp_type qp_type, int is_tun)
1655 {
1656 	int i;
1657 	struct mlx4_ib_demux_pv_qp *tun_qp;
1658 	int rx_buf_size, tx_buf_size;
1659 
1660 	if (qp_type > IB_QPT_GSI)
1661 		return;
1662 
1663 	tun_qp = &ctx->qp[qp_type];
1664 	if (is_tun) {
1665 		rx_buf_size = sizeof (struct mlx4_tunnel_mad);
1666 		tx_buf_size = sizeof (struct mlx4_rcv_tunnel_mad);
1667 	} else {
1668 		rx_buf_size = sizeof (struct mlx4_mad_rcv_buf);
1669 		tx_buf_size = sizeof (struct mlx4_mad_snd_buf);
1670 	}
1671 
1672 
1673 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1674 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->ring[i].map,
1675 				    rx_buf_size, DMA_FROM_DEVICE);
1676 		kfree(tun_qp->ring[i].addr);
1677 	}
1678 
1679 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1680 		ib_dma_unmap_single(ctx->ib_dev, tun_qp->tx_ring[i].buf.map,
1681 				    tx_buf_size, DMA_TO_DEVICE);
1682 		kfree(tun_qp->tx_ring[i].buf.addr);
1683 		if (tun_qp->tx_ring[i].ah)
1684 			ib_destroy_ah(tun_qp->tx_ring[i].ah);
1685 	}
1686 	kfree(tun_qp->tx_ring);
1687 	kfree(tun_qp->ring);
1688 }
1689 
1690 static void mlx4_ib_tunnel_comp_worker(struct work_struct *work)
1691 {
1692 	struct mlx4_ib_demux_pv_ctx *ctx;
1693 	struct mlx4_ib_demux_pv_qp *tun_qp;
1694 	struct ib_wc wc;
1695 	int ret;
1696 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1697 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1698 
1699 	while (ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1700 		tun_qp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1701 		if (wc.status == IB_WC_SUCCESS) {
1702 			switch (wc.opcode) {
1703 			case IB_WC_RECV:
1704 				mlx4_ib_multiplex_mad(ctx, &wc);
1705 				ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp,
1706 							     wc.wr_id &
1707 							     (MLX4_NUM_TUNNEL_BUFS - 1));
1708 				if (ret)
1709 					pr_err("Failed reposting tunnel "
1710 					       "buf:%lld\n", wc.wr_id);
1711 				break;
1712 			case IB_WC_SEND:
1713 				pr_debug("received tunnel send completion:"
1714 					 "wrid=0x%llx, status=0x%x\n",
1715 					 wc.wr_id, wc.status);
1716 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1717 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1718 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1719 					= NULL;
1720 				spin_lock(&tun_qp->tx_lock);
1721 				tun_qp->tx_ix_tail++;
1722 				spin_unlock(&tun_qp->tx_lock);
1723 
1724 				break;
1725 			default:
1726 				break;
1727 			}
1728 		} else  {
1729 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1730 				 " status = %d, wrid = 0x%llx\n",
1731 				 ctx->slave, wc.status, wc.wr_id);
1732 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1733 				ib_destroy_ah(tun_qp->tx_ring[wc.wr_id &
1734 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1735 				tun_qp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1736 					= NULL;
1737 				spin_lock(&tun_qp->tx_lock);
1738 				tun_qp->tx_ix_tail++;
1739 				spin_unlock(&tun_qp->tx_lock);
1740 			}
1741 		}
1742 	}
1743 }
1744 
1745 static void pv_qp_event_handler(struct ib_event *event, void *qp_context)
1746 {
1747 	struct mlx4_ib_demux_pv_ctx *sqp = qp_context;
1748 
1749 	/* It's worse than that! He's dead, Jim! */
1750 	pr_err("Fatal error (%d) on a MAD QP on port %d\n",
1751 	       event->event, sqp->port);
1752 }
1753 
1754 static int create_pv_sqp(struct mlx4_ib_demux_pv_ctx *ctx,
1755 			    enum ib_qp_type qp_type, int create_tun)
1756 {
1757 	int i, ret;
1758 	struct mlx4_ib_demux_pv_qp *tun_qp;
1759 	struct mlx4_ib_qp_tunnel_init_attr qp_init_attr;
1760 	struct ib_qp_attr attr;
1761 	int qp_attr_mask_INIT;
1762 
1763 	if (qp_type > IB_QPT_GSI)
1764 		return -EINVAL;
1765 
1766 	tun_qp = &ctx->qp[qp_type];
1767 
1768 	memset(&qp_init_attr, 0, sizeof qp_init_attr);
1769 	qp_init_attr.init_attr.send_cq = ctx->cq;
1770 	qp_init_attr.init_attr.recv_cq = ctx->cq;
1771 	qp_init_attr.init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
1772 	qp_init_attr.init_attr.cap.max_send_wr = MLX4_NUM_TUNNEL_BUFS;
1773 	qp_init_attr.init_attr.cap.max_recv_wr = MLX4_NUM_TUNNEL_BUFS;
1774 	qp_init_attr.init_attr.cap.max_send_sge = 1;
1775 	qp_init_attr.init_attr.cap.max_recv_sge = 1;
1776 	if (create_tun) {
1777 		qp_init_attr.init_attr.qp_type = IB_QPT_UD;
1778 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_TUNNEL_QP;
1779 		qp_init_attr.port = ctx->port;
1780 		qp_init_attr.slave = ctx->slave;
1781 		qp_init_attr.proxy_qp_type = qp_type;
1782 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX |
1783 			   IB_QP_QKEY | IB_QP_PORT;
1784 	} else {
1785 		qp_init_attr.init_attr.qp_type = qp_type;
1786 		qp_init_attr.init_attr.create_flags = MLX4_IB_SRIOV_SQP;
1787 		qp_attr_mask_INIT = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY;
1788 	}
1789 	qp_init_attr.init_attr.port_num = ctx->port;
1790 	qp_init_attr.init_attr.qp_context = ctx;
1791 	qp_init_attr.init_attr.event_handler = pv_qp_event_handler;
1792 	tun_qp->qp = ib_create_qp(ctx->pd, &qp_init_attr.init_attr);
1793 	if (IS_ERR(tun_qp->qp)) {
1794 		ret = PTR_ERR(tun_qp->qp);
1795 		tun_qp->qp = NULL;
1796 		pr_err("Couldn't create %s QP (%d)\n",
1797 		       create_tun ? "tunnel" : "special", ret);
1798 		return ret;
1799 	}
1800 
1801 	memset(&attr, 0, sizeof attr);
1802 	attr.qp_state = IB_QPS_INIT;
1803 	ret = 0;
1804 	if (create_tun)
1805 		ret = find_slave_port_pkey_ix(to_mdev(ctx->ib_dev), ctx->slave,
1806 					      ctx->port, IB_DEFAULT_PKEY_FULL,
1807 					      &attr.pkey_index);
1808 	if (ret || !create_tun)
1809 		attr.pkey_index =
1810 			to_mdev(ctx->ib_dev)->pkeys.virt2phys_pkey[ctx->slave][ctx->port - 1][0];
1811 	attr.qkey = IB_QP1_QKEY;
1812 	attr.port_num = ctx->port;
1813 	ret = ib_modify_qp(tun_qp->qp, &attr, qp_attr_mask_INIT);
1814 	if (ret) {
1815 		pr_err("Couldn't change %s qp state to INIT (%d)\n",
1816 		       create_tun ? "tunnel" : "special", ret);
1817 		goto err_qp;
1818 	}
1819 	attr.qp_state = IB_QPS_RTR;
1820 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE);
1821 	if (ret) {
1822 		pr_err("Couldn't change %s qp state to RTR (%d)\n",
1823 		       create_tun ? "tunnel" : "special", ret);
1824 		goto err_qp;
1825 	}
1826 	attr.qp_state = IB_QPS_RTS;
1827 	attr.sq_psn = 0;
1828 	ret = ib_modify_qp(tun_qp->qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
1829 	if (ret) {
1830 		pr_err("Couldn't change %s qp state to RTS (%d)\n",
1831 		       create_tun ? "tunnel" : "special", ret);
1832 		goto err_qp;
1833 	}
1834 
1835 	for (i = 0; i < MLX4_NUM_TUNNEL_BUFS; i++) {
1836 		ret = mlx4_ib_post_pv_qp_buf(ctx, tun_qp, i);
1837 		if (ret) {
1838 			pr_err(" mlx4_ib_post_pv_buf error"
1839 			       " (err = %d, i = %d)\n", ret, i);
1840 			goto err_qp;
1841 		}
1842 	}
1843 	return 0;
1844 
1845 err_qp:
1846 	ib_destroy_qp(tun_qp->qp);
1847 	tun_qp->qp = NULL;
1848 	return ret;
1849 }
1850 
1851 /*
1852  * IB MAD completion callback for real SQPs
1853  */
1854 static void mlx4_ib_sqp_comp_worker(struct work_struct *work)
1855 {
1856 	struct mlx4_ib_demux_pv_ctx *ctx;
1857 	struct mlx4_ib_demux_pv_qp *sqp;
1858 	struct ib_wc wc;
1859 	struct ib_grh *grh;
1860 	struct ib_mad *mad;
1861 
1862 	ctx = container_of(work, struct mlx4_ib_demux_pv_ctx, work);
1863 	ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
1864 
1865 	while (mlx4_ib_poll_cq(ctx->cq, 1, &wc) == 1) {
1866 		sqp = &ctx->qp[MLX4_TUN_WRID_QPN(wc.wr_id)];
1867 		if (wc.status == IB_WC_SUCCESS) {
1868 			switch (wc.opcode) {
1869 			case IB_WC_SEND:
1870 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1871 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1872 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1873 					= NULL;
1874 				spin_lock(&sqp->tx_lock);
1875 				sqp->tx_ix_tail++;
1876 				spin_unlock(&sqp->tx_lock);
1877 				break;
1878 			case IB_WC_RECV:
1879 				mad = (struct ib_mad *) &(((struct mlx4_mad_rcv_buf *)
1880 						(sqp->ring[wc.wr_id &
1881 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->payload);
1882 				grh = &(((struct mlx4_mad_rcv_buf *)
1883 						(sqp->ring[wc.wr_id &
1884 						(MLX4_NUM_TUNNEL_BUFS - 1)].addr))->grh);
1885 				mlx4_ib_demux_mad(ctx->ib_dev, ctx->port, &wc, grh, mad);
1886 				if (mlx4_ib_post_pv_qp_buf(ctx, sqp, wc.wr_id &
1887 							   (MLX4_NUM_TUNNEL_BUFS - 1)))
1888 					pr_err("Failed reposting SQP "
1889 					       "buf:%lld\n", wc.wr_id);
1890 				break;
1891 			default:
1892 				BUG_ON(1);
1893 				break;
1894 			}
1895 		} else  {
1896 			pr_debug("mlx4_ib: completion error in tunnel: %d."
1897 				 " status = %d, wrid = 0x%llx\n",
1898 				 ctx->slave, wc.status, wc.wr_id);
1899 			if (!MLX4_TUN_IS_RECV(wc.wr_id)) {
1900 				ib_destroy_ah(sqp->tx_ring[wc.wr_id &
1901 					      (MLX4_NUM_TUNNEL_BUFS - 1)].ah);
1902 				sqp->tx_ring[wc.wr_id & (MLX4_NUM_TUNNEL_BUFS - 1)].ah
1903 					= NULL;
1904 				spin_lock(&sqp->tx_lock);
1905 				sqp->tx_ix_tail++;
1906 				spin_unlock(&sqp->tx_lock);
1907 			}
1908 		}
1909 	}
1910 }
1911 
1912 static int alloc_pv_object(struct mlx4_ib_dev *dev, int slave, int port,
1913 			       struct mlx4_ib_demux_pv_ctx **ret_ctx)
1914 {
1915 	struct mlx4_ib_demux_pv_ctx *ctx;
1916 
1917 	*ret_ctx = NULL;
1918 	ctx = kzalloc(sizeof (struct mlx4_ib_demux_pv_ctx), GFP_KERNEL);
1919 	if (!ctx) {
1920 		pr_err("failed allocating pv resource context "
1921 		       "for port %d, slave %d\n", port, slave);
1922 		return -ENOMEM;
1923 	}
1924 
1925 	ctx->ib_dev = &dev->ib_dev;
1926 	ctx->port = port;
1927 	ctx->slave = slave;
1928 	*ret_ctx = ctx;
1929 	return 0;
1930 }
1931 
1932 static void free_pv_object(struct mlx4_ib_dev *dev, int slave, int port)
1933 {
1934 	if (dev->sriov.demux[port - 1].tun[slave]) {
1935 		kfree(dev->sriov.demux[port - 1].tun[slave]);
1936 		dev->sriov.demux[port - 1].tun[slave] = NULL;
1937 	}
1938 }
1939 
1940 static int create_pv_resources(struct ib_device *ibdev, int slave, int port,
1941 			       int create_tun, struct mlx4_ib_demux_pv_ctx *ctx)
1942 {
1943 	int ret, cq_size;
1944 	struct ib_cq_init_attr cq_attr = {};
1945 
1946 	if (ctx->state != DEMUX_PV_STATE_DOWN)
1947 		return -EEXIST;
1948 
1949 	ctx->state = DEMUX_PV_STATE_STARTING;
1950 	/* have QP0 only if link layer is IB */
1951 	if (rdma_port_get_link_layer(ibdev, ctx->port) ==
1952 	    IB_LINK_LAYER_INFINIBAND)
1953 		ctx->has_smi = 1;
1954 
1955 	if (ctx->has_smi) {
1956 		ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_SMI, create_tun);
1957 		if (ret) {
1958 			pr_err("Failed allocating qp0 tunnel bufs (%d)\n", ret);
1959 			goto err_out;
1960 		}
1961 	}
1962 
1963 	ret = mlx4_ib_alloc_pv_bufs(ctx, IB_QPT_GSI, create_tun);
1964 	if (ret) {
1965 		pr_err("Failed allocating qp1 tunnel bufs (%d)\n", ret);
1966 		goto err_out_qp0;
1967 	}
1968 
1969 	cq_size = 2 * MLX4_NUM_TUNNEL_BUFS;
1970 	if (ctx->has_smi)
1971 		cq_size *= 2;
1972 
1973 	cq_attr.cqe = cq_size;
1974 	ctx->cq = ib_create_cq(ctx->ib_dev, mlx4_ib_tunnel_comp_handler,
1975 			       NULL, ctx, &cq_attr);
1976 	if (IS_ERR(ctx->cq)) {
1977 		ret = PTR_ERR(ctx->cq);
1978 		pr_err("Couldn't create tunnel CQ (%d)\n", ret);
1979 		goto err_buf;
1980 	}
1981 
1982 	ctx->pd = ib_alloc_pd(ctx->ib_dev, 0);
1983 	if (IS_ERR(ctx->pd)) {
1984 		ret = PTR_ERR(ctx->pd);
1985 		pr_err("Couldn't create tunnel PD (%d)\n", ret);
1986 		goto err_cq;
1987 	}
1988 
1989 	if (ctx->has_smi) {
1990 		ret = create_pv_sqp(ctx, IB_QPT_SMI, create_tun);
1991 		if (ret) {
1992 			pr_err("Couldn't create %s QP0 (%d)\n",
1993 			       create_tun ? "tunnel for" : "",  ret);
1994 			goto err_pd;
1995 		}
1996 	}
1997 
1998 	ret = create_pv_sqp(ctx, IB_QPT_GSI, create_tun);
1999 	if (ret) {
2000 		pr_err("Couldn't create %s QP1 (%d)\n",
2001 		       create_tun ? "tunnel for" : "",  ret);
2002 		goto err_qp0;
2003 	}
2004 
2005 	if (create_tun)
2006 		INIT_WORK(&ctx->work, mlx4_ib_tunnel_comp_worker);
2007 	else
2008 		INIT_WORK(&ctx->work, mlx4_ib_sqp_comp_worker);
2009 
2010 	ctx->wq = to_mdev(ibdev)->sriov.demux[port - 1].wq;
2011 
2012 	ret = ib_req_notify_cq(ctx->cq, IB_CQ_NEXT_COMP);
2013 	if (ret) {
2014 		pr_err("Couldn't arm tunnel cq (%d)\n", ret);
2015 		goto err_wq;
2016 	}
2017 	ctx->state = DEMUX_PV_STATE_ACTIVE;
2018 	return 0;
2019 
2020 err_wq:
2021 	ctx->wq = NULL;
2022 	ib_destroy_qp(ctx->qp[1].qp);
2023 	ctx->qp[1].qp = NULL;
2024 
2025 
2026 err_qp0:
2027 	if (ctx->has_smi)
2028 		ib_destroy_qp(ctx->qp[0].qp);
2029 	ctx->qp[0].qp = NULL;
2030 
2031 err_pd:
2032 	ib_dealloc_pd(ctx->pd);
2033 	ctx->pd = NULL;
2034 
2035 err_cq:
2036 	ib_destroy_cq(ctx->cq);
2037 	ctx->cq = NULL;
2038 
2039 err_buf:
2040 	mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, create_tun);
2041 
2042 err_out_qp0:
2043 	if (ctx->has_smi)
2044 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, create_tun);
2045 err_out:
2046 	ctx->state = DEMUX_PV_STATE_DOWN;
2047 	return ret;
2048 }
2049 
2050 static void destroy_pv_resources(struct mlx4_ib_dev *dev, int slave, int port,
2051 				 struct mlx4_ib_demux_pv_ctx *ctx, int flush)
2052 {
2053 	if (!ctx)
2054 		return;
2055 	if (ctx->state > DEMUX_PV_STATE_DOWN) {
2056 		ctx->state = DEMUX_PV_STATE_DOWNING;
2057 		if (flush)
2058 			flush_workqueue(ctx->wq);
2059 		if (ctx->has_smi) {
2060 			ib_destroy_qp(ctx->qp[0].qp);
2061 			ctx->qp[0].qp = NULL;
2062 			mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_SMI, 1);
2063 		}
2064 		ib_destroy_qp(ctx->qp[1].qp);
2065 		ctx->qp[1].qp = NULL;
2066 		mlx4_ib_free_pv_qp_bufs(ctx, IB_QPT_GSI, 1);
2067 		ib_dealloc_pd(ctx->pd);
2068 		ctx->pd = NULL;
2069 		ib_destroy_cq(ctx->cq);
2070 		ctx->cq = NULL;
2071 		ctx->state = DEMUX_PV_STATE_DOWN;
2072 	}
2073 }
2074 
2075 static int mlx4_ib_tunnels_update(struct mlx4_ib_dev *dev, int slave,
2076 				  int port, int do_init)
2077 {
2078 	int ret = 0;
2079 
2080 	if (!do_init) {
2081 		clean_vf_mcast(&dev->sriov.demux[port - 1], slave);
2082 		/* for master, destroy real sqp resources */
2083 		if (slave == mlx4_master_func_num(dev->dev))
2084 			destroy_pv_resources(dev, slave, port,
2085 					     dev->sriov.sqps[port - 1], 1);
2086 		/* destroy the tunnel qp resources */
2087 		destroy_pv_resources(dev, slave, port,
2088 				     dev->sriov.demux[port - 1].tun[slave], 1);
2089 		return 0;
2090 	}
2091 
2092 	/* create the tunnel qp resources */
2093 	ret = create_pv_resources(&dev->ib_dev, slave, port, 1,
2094 				  dev->sriov.demux[port - 1].tun[slave]);
2095 
2096 	/* for master, create the real sqp resources */
2097 	if (!ret && slave == mlx4_master_func_num(dev->dev))
2098 		ret = create_pv_resources(&dev->ib_dev, slave, port, 0,
2099 					  dev->sriov.sqps[port - 1]);
2100 	return ret;
2101 }
2102 
2103 void mlx4_ib_tunnels_update_work(struct work_struct *work)
2104 {
2105 	struct mlx4_ib_demux_work *dmxw;
2106 
2107 	dmxw = container_of(work, struct mlx4_ib_demux_work, work);
2108 	mlx4_ib_tunnels_update(dmxw->dev, dmxw->slave, (int) dmxw->port,
2109 			       dmxw->do_init);
2110 	kfree(dmxw);
2111 	return;
2112 }
2113 
2114 static int mlx4_ib_alloc_demux_ctx(struct mlx4_ib_dev *dev,
2115 				       struct mlx4_ib_demux_ctx *ctx,
2116 				       int port)
2117 {
2118 	char name[12];
2119 	int ret = 0;
2120 	int i;
2121 
2122 	ctx->tun = kcalloc(dev->dev->caps.sqp_demux,
2123 			   sizeof (struct mlx4_ib_demux_pv_ctx *), GFP_KERNEL);
2124 	if (!ctx->tun)
2125 		return -ENOMEM;
2126 
2127 	ctx->dev = dev;
2128 	ctx->port = port;
2129 	ctx->ib_dev = &dev->ib_dev;
2130 
2131 	for (i = 0;
2132 	     i < min(dev->dev->caps.sqp_demux,
2133 	     (u16)(dev->dev->persist->num_vfs + 1));
2134 	     i++) {
2135 		struct mlx4_active_ports actv_ports =
2136 			mlx4_get_active_ports(dev->dev, i);
2137 
2138 		if (!test_bit(port - 1, actv_ports.ports))
2139 			continue;
2140 
2141 		ret = alloc_pv_object(dev, i, port, &ctx->tun[i]);
2142 		if (ret) {
2143 			ret = -ENOMEM;
2144 			goto err_mcg;
2145 		}
2146 	}
2147 
2148 	ret = mlx4_ib_mcg_port_init(ctx);
2149 	if (ret) {
2150 		pr_err("Failed initializing mcg para-virt (%d)\n", ret);
2151 		goto err_mcg;
2152 	}
2153 
2154 	snprintf(name, sizeof name, "mlx4_ibt%d", port);
2155 	ctx->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2156 	if (!ctx->wq) {
2157 		pr_err("Failed to create tunnelling WQ for port %d\n", port);
2158 		ret = -ENOMEM;
2159 		goto err_wq;
2160 	}
2161 
2162 	snprintf(name, sizeof name, "mlx4_ibud%d", port);
2163 	ctx->ud_wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
2164 	if (!ctx->ud_wq) {
2165 		pr_err("Failed to create up/down WQ for port %d\n", port);
2166 		ret = -ENOMEM;
2167 		goto err_udwq;
2168 	}
2169 
2170 	return 0;
2171 
2172 err_udwq:
2173 	destroy_workqueue(ctx->wq);
2174 	ctx->wq = NULL;
2175 
2176 err_wq:
2177 	mlx4_ib_mcg_port_cleanup(ctx, 1);
2178 err_mcg:
2179 	for (i = 0; i < dev->dev->caps.sqp_demux; i++)
2180 		free_pv_object(dev, i, port);
2181 	kfree(ctx->tun);
2182 	ctx->tun = NULL;
2183 	return ret;
2184 }
2185 
2186 static void mlx4_ib_free_sqp_ctx(struct mlx4_ib_demux_pv_ctx *sqp_ctx)
2187 {
2188 	if (sqp_ctx->state > DEMUX_PV_STATE_DOWN) {
2189 		sqp_ctx->state = DEMUX_PV_STATE_DOWNING;
2190 		flush_workqueue(sqp_ctx->wq);
2191 		if (sqp_ctx->has_smi) {
2192 			ib_destroy_qp(sqp_ctx->qp[0].qp);
2193 			sqp_ctx->qp[0].qp = NULL;
2194 			mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_SMI, 0);
2195 		}
2196 		ib_destroy_qp(sqp_ctx->qp[1].qp);
2197 		sqp_ctx->qp[1].qp = NULL;
2198 		mlx4_ib_free_pv_qp_bufs(sqp_ctx, IB_QPT_GSI, 0);
2199 		ib_dealloc_pd(sqp_ctx->pd);
2200 		sqp_ctx->pd = NULL;
2201 		ib_destroy_cq(sqp_ctx->cq);
2202 		sqp_ctx->cq = NULL;
2203 		sqp_ctx->state = DEMUX_PV_STATE_DOWN;
2204 	}
2205 }
2206 
2207 static void mlx4_ib_free_demux_ctx(struct mlx4_ib_demux_ctx *ctx)
2208 {
2209 	int i;
2210 	if (ctx) {
2211 		struct mlx4_ib_dev *dev = to_mdev(ctx->ib_dev);
2212 		mlx4_ib_mcg_port_cleanup(ctx, 1);
2213 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2214 			if (!ctx->tun[i])
2215 				continue;
2216 			if (ctx->tun[i]->state > DEMUX_PV_STATE_DOWN)
2217 				ctx->tun[i]->state = DEMUX_PV_STATE_DOWNING;
2218 		}
2219 		flush_workqueue(ctx->wq);
2220 		for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2221 			destroy_pv_resources(dev, i, ctx->port, ctx->tun[i], 0);
2222 			free_pv_object(dev, i, ctx->port);
2223 		}
2224 		kfree(ctx->tun);
2225 		destroy_workqueue(ctx->ud_wq);
2226 		destroy_workqueue(ctx->wq);
2227 	}
2228 }
2229 
2230 static void mlx4_ib_master_tunnels(struct mlx4_ib_dev *dev, int do_init)
2231 {
2232 	int i;
2233 
2234 	if (!mlx4_is_master(dev->dev))
2235 		return;
2236 	/* initialize or tear down tunnel QPs for the master */
2237 	for (i = 0; i < dev->dev->caps.num_ports; i++)
2238 		mlx4_ib_tunnels_update(dev, mlx4_master_func_num(dev->dev), i + 1, do_init);
2239 	return;
2240 }
2241 
2242 int mlx4_ib_init_sriov(struct mlx4_ib_dev *dev)
2243 {
2244 	int i = 0;
2245 	int err;
2246 
2247 	if (!mlx4_is_mfunc(dev->dev))
2248 		return 0;
2249 
2250 	dev->sriov.is_going_down = 0;
2251 	spin_lock_init(&dev->sriov.going_down_lock);
2252 	mlx4_ib_cm_paravirt_init(dev);
2253 
2254 	mlx4_ib_warn(&dev->ib_dev, "multi-function enabled\n");
2255 
2256 	if (mlx4_is_slave(dev->dev)) {
2257 		mlx4_ib_warn(&dev->ib_dev, "operating in qp1 tunnel mode\n");
2258 		return 0;
2259 	}
2260 
2261 	for (i = 0; i < dev->dev->caps.sqp_demux; i++) {
2262 		if (i == mlx4_master_func_num(dev->dev))
2263 			mlx4_put_slave_node_guid(dev->dev, i, dev->ib_dev.node_guid);
2264 		else
2265 			mlx4_put_slave_node_guid(dev->dev, i, mlx4_ib_gen_node_guid());
2266 	}
2267 
2268 	err = mlx4_ib_init_alias_guid_service(dev);
2269 	if (err) {
2270 		mlx4_ib_warn(&dev->ib_dev, "Failed init alias guid process.\n");
2271 		goto paravirt_err;
2272 	}
2273 	err = mlx4_ib_device_register_sysfs(dev);
2274 	if (err) {
2275 		mlx4_ib_warn(&dev->ib_dev, "Failed to register sysfs\n");
2276 		goto sysfs_err;
2277 	}
2278 
2279 	mlx4_ib_warn(&dev->ib_dev, "initializing demux service for %d qp1 clients\n",
2280 		     dev->dev->caps.sqp_demux);
2281 	for (i = 0; i < dev->num_ports; i++) {
2282 		union ib_gid gid;
2283 		err = __mlx4_ib_query_gid(&dev->ib_dev, i + 1, 0, &gid, 1);
2284 		if (err)
2285 			goto demux_err;
2286 		dev->sriov.demux[i].guid_cache[0] = gid.global.interface_id;
2287 		atomic64_set(&dev->sriov.demux[i].subnet_prefix,
2288 			     be64_to_cpu(gid.global.subnet_prefix));
2289 		err = alloc_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1,
2290 				      &dev->sriov.sqps[i]);
2291 		if (err)
2292 			goto demux_err;
2293 		err = mlx4_ib_alloc_demux_ctx(dev, &dev->sriov.demux[i], i + 1);
2294 		if (err)
2295 			goto free_pv;
2296 	}
2297 	mlx4_ib_master_tunnels(dev, 1);
2298 	return 0;
2299 
2300 free_pv:
2301 	free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2302 demux_err:
2303 	while (--i >= 0) {
2304 		free_pv_object(dev, mlx4_master_func_num(dev->dev), i + 1);
2305 		mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2306 	}
2307 	mlx4_ib_device_unregister_sysfs(dev);
2308 
2309 sysfs_err:
2310 	mlx4_ib_destroy_alias_guid_service(dev);
2311 
2312 paravirt_err:
2313 	mlx4_ib_cm_paravirt_clean(dev, -1);
2314 
2315 	return err;
2316 }
2317 
2318 void mlx4_ib_close_sriov(struct mlx4_ib_dev *dev)
2319 {
2320 	int i;
2321 	unsigned long flags;
2322 
2323 	if (!mlx4_is_mfunc(dev->dev))
2324 		return;
2325 
2326 	spin_lock_irqsave(&dev->sriov.going_down_lock, flags);
2327 	dev->sriov.is_going_down = 1;
2328 	spin_unlock_irqrestore(&dev->sriov.going_down_lock, flags);
2329 	if (mlx4_is_master(dev->dev)) {
2330 		for (i = 0; i < dev->num_ports; i++) {
2331 			flush_workqueue(dev->sriov.demux[i].ud_wq);
2332 			mlx4_ib_free_sqp_ctx(dev->sriov.sqps[i]);
2333 			kfree(dev->sriov.sqps[i]);
2334 			dev->sriov.sqps[i] = NULL;
2335 			mlx4_ib_free_demux_ctx(&dev->sriov.demux[i]);
2336 		}
2337 
2338 		mlx4_ib_cm_paravirt_clean(dev, -1);
2339 		mlx4_ib_destroy_alias_guid_service(dev);
2340 		mlx4_ib_device_unregister_sysfs(dev);
2341 	}
2342 }
2343