xref: /linux/drivers/infiniband/hw/mlx4/main.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
1 /*
2  * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
3  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/module.h>
35 #include <linux/init.h>
36 #include <linux/slab.h>
37 #include <linux/errno.h>
38 #include <linux/netdevice.h>
39 #include <linux/inetdevice.h>
40 #include <linux/rtnetlink.h>
41 #include <linux/if_vlan.h>
42 
43 #include <rdma/ib_smi.h>
44 #include <rdma/ib_user_verbs.h>
45 #include <rdma/ib_addr.h>
46 
47 #include <linux/mlx4/driver.h>
48 #include <linux/mlx4/cmd.h>
49 
50 #include "mlx4_ib.h"
51 #include "user.h"
52 
53 #define DRV_NAME	"mlx4_ib"
54 #define DRV_VERSION	"1.0"
55 #define DRV_RELDATE	"April 4, 2008"
56 
57 MODULE_AUTHOR("Roland Dreier");
58 MODULE_DESCRIPTION("Mellanox ConnectX HCA InfiniBand driver");
59 MODULE_LICENSE("Dual BSD/GPL");
60 MODULE_VERSION(DRV_VERSION);
61 
62 static const char mlx4_ib_version[] =
63 	DRV_NAME ": Mellanox ConnectX InfiniBand driver v"
64 	DRV_VERSION " (" DRV_RELDATE ")\n";
65 
66 struct update_gid_work {
67 	struct work_struct	work;
68 	union ib_gid		gids[128];
69 	struct mlx4_ib_dev     *dev;
70 	int			port;
71 };
72 
73 static struct workqueue_struct *wq;
74 
75 static void init_query_mad(struct ib_smp *mad)
76 {
77 	mad->base_version  = 1;
78 	mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
79 	mad->class_version = 1;
80 	mad->method	   = IB_MGMT_METHOD_GET;
81 }
82 
83 static union ib_gid zgid;
84 
85 static int mlx4_ib_query_device(struct ib_device *ibdev,
86 				struct ib_device_attr *props)
87 {
88 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
89 	struct ib_smp *in_mad  = NULL;
90 	struct ib_smp *out_mad = NULL;
91 	int err = -ENOMEM;
92 
93 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
94 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
95 	if (!in_mad || !out_mad)
96 		goto out;
97 
98 	init_query_mad(in_mad);
99 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
100 
101 	err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, 1, NULL, NULL, in_mad, out_mad);
102 	if (err)
103 		goto out;
104 
105 	memset(props, 0, sizeof *props);
106 
107 	props->fw_ver = dev->dev->caps.fw_ver;
108 	props->device_cap_flags    = IB_DEVICE_CHANGE_PHY_PORT |
109 		IB_DEVICE_PORT_ACTIVE_EVENT		|
110 		IB_DEVICE_SYS_IMAGE_GUID		|
111 		IB_DEVICE_RC_RNR_NAK_GEN		|
112 		IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
113 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_PKEY_CNTR)
114 		props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
115 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_QKEY_CNTR)
116 		props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
117 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_APM)
118 		props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
119 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UD_AV_PORT)
120 		props->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
121 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IPOIB_CSUM)
122 		props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
123 	if (dev->dev->caps.max_gso_sz && dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BLH)
124 		props->device_cap_flags |= IB_DEVICE_UD_TSO;
125 	if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_RESERVED_LKEY)
126 		props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
127 	if ((dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_LOCAL_INV) &&
128 	    (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_REMOTE_INV) &&
129 	    (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_FAST_REG_WR))
130 		props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
131 	if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)
132 		props->device_cap_flags |= IB_DEVICE_XRC;
133 
134 	props->vendor_id	   = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
135 		0xffffff;
136 	props->vendor_part_id	   = be16_to_cpup((__be16 *) (out_mad->data + 30));
137 	props->hw_ver		   = be32_to_cpup((__be32 *) (out_mad->data + 32));
138 	memcpy(&props->sys_image_guid, out_mad->data +	4, 8);
139 
140 	props->max_mr_size	   = ~0ull;
141 	props->page_size_cap	   = dev->dev->caps.page_size_cap;
142 	props->max_qp		   = dev->dev->caps.num_qps - dev->dev->caps.reserved_qps;
143 	props->max_qp_wr	   = dev->dev->caps.max_wqes;
144 	props->max_sge		   = min(dev->dev->caps.max_sq_sg,
145 					 dev->dev->caps.max_rq_sg);
146 	props->max_cq		   = dev->dev->caps.num_cqs - dev->dev->caps.reserved_cqs;
147 	props->max_cqe		   = dev->dev->caps.max_cqes;
148 	props->max_mr		   = dev->dev->caps.num_mpts - dev->dev->caps.reserved_mrws;
149 	props->max_pd		   = dev->dev->caps.num_pds - dev->dev->caps.reserved_pds;
150 	props->max_qp_rd_atom	   = dev->dev->caps.max_qp_dest_rdma;
151 	props->max_qp_init_rd_atom = dev->dev->caps.max_qp_init_rdma;
152 	props->max_res_rd_atom	   = props->max_qp_rd_atom * props->max_qp;
153 	props->max_srq		   = dev->dev->caps.num_srqs - dev->dev->caps.reserved_srqs;
154 	props->max_srq_wr	   = dev->dev->caps.max_srq_wqes - 1;
155 	props->max_srq_sge	   = dev->dev->caps.max_srq_sge;
156 	props->max_fast_reg_page_list_len = MLX4_MAX_FAST_REG_PAGES;
157 	props->local_ca_ack_delay  = dev->dev->caps.local_ca_ack_delay;
158 	props->atomic_cap	   = dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_ATOMIC ?
159 		IB_ATOMIC_HCA : IB_ATOMIC_NONE;
160 	props->masked_atomic_cap   = IB_ATOMIC_HCA;
161 	props->max_pkeys	   = dev->dev->caps.pkey_table_len[1];
162 	props->max_mcast_grp	   = dev->dev->caps.num_mgms + dev->dev->caps.num_amgms;
163 	props->max_mcast_qp_attach = dev->dev->caps.num_qp_per_mgm;
164 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
165 					   props->max_mcast_grp;
166 	props->max_map_per_fmr = (1 << (32 - ilog2(dev->dev->caps.num_mpts))) - 1;
167 
168 out:
169 	kfree(in_mad);
170 	kfree(out_mad);
171 
172 	return err;
173 }
174 
175 static enum rdma_link_layer
176 mlx4_ib_port_link_layer(struct ib_device *device, u8 port_num)
177 {
178 	struct mlx4_dev *dev = to_mdev(device)->dev;
179 
180 	return dev->caps.port_mask & (1 << (port_num - 1)) ?
181 		IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
182 }
183 
184 static int ib_link_query_port(struct ib_device *ibdev, u8 port,
185 			      struct ib_port_attr *props,
186 			      struct ib_smp *in_mad,
187 			      struct ib_smp *out_mad)
188 {
189 	int ext_active_speed;
190 	int err;
191 
192 	props->lid		= be16_to_cpup((__be16 *) (out_mad->data + 16));
193 	props->lmc		= out_mad->data[34] & 0x7;
194 	props->sm_lid		= be16_to_cpup((__be16 *) (out_mad->data + 18));
195 	props->sm_sl		= out_mad->data[36] & 0xf;
196 	props->state		= out_mad->data[32] & 0xf;
197 	props->phys_state	= out_mad->data[33] >> 4;
198 	props->port_cap_flags	= be32_to_cpup((__be32 *) (out_mad->data + 20));
199 	props->gid_tbl_len	= to_mdev(ibdev)->dev->caps.gid_table_len[port];
200 	props->max_msg_sz	= to_mdev(ibdev)->dev->caps.max_msg_sz;
201 	props->pkey_tbl_len	= to_mdev(ibdev)->dev->caps.pkey_table_len[port];
202 	props->bad_pkey_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 46));
203 	props->qkey_viol_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 48));
204 	props->active_width	= out_mad->data[31] & 0xf;
205 	props->active_speed	= out_mad->data[35] >> 4;
206 	props->max_mtu		= out_mad->data[41] & 0xf;
207 	props->active_mtu	= out_mad->data[36] >> 4;
208 	props->subnet_timeout	= out_mad->data[51] & 0x1f;
209 	props->max_vl_num	= out_mad->data[37] >> 4;
210 	props->init_type_reply	= out_mad->data[41] >> 4;
211 
212 	/* Check if extended speeds (EDR/FDR/...) are supported */
213 	if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
214 		ext_active_speed = out_mad->data[62] >> 4;
215 
216 		switch (ext_active_speed) {
217 		case 1:
218 			props->active_speed = 16; /* FDR */
219 			break;
220 		case 2:
221 			props->active_speed = 32; /* EDR */
222 			break;
223 		}
224 	}
225 
226 	/* If reported active speed is QDR, check if is FDR-10 */
227 	if (props->active_speed == 4) {
228 		if (to_mdev(ibdev)->dev->caps.ext_port_cap[port] &
229 		    MLX_EXT_PORT_CAP_FLAG_EXTENDED_PORT_INFO) {
230 			init_query_mad(in_mad);
231 			in_mad->attr_id = MLX4_ATTR_EXTENDED_PORT_INFO;
232 			in_mad->attr_mod = cpu_to_be32(port);
233 
234 			err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port,
235 					   NULL, NULL, in_mad, out_mad);
236 			if (err)
237 				return err;
238 
239 			/* Checking LinkSpeedActive for FDR-10 */
240 			if (out_mad->data[15] & 0x1)
241 				props->active_speed = 8;
242 		}
243 	}
244 
245 	return 0;
246 }
247 
248 static u8 state_to_phys_state(enum ib_port_state state)
249 {
250 	return state == IB_PORT_ACTIVE ? 5 : 3;
251 }
252 
253 static int eth_link_query_port(struct ib_device *ibdev, u8 port,
254 			       struct ib_port_attr *props,
255 			       struct ib_smp *out_mad)
256 {
257 	struct mlx4_ib_iboe *iboe = &to_mdev(ibdev)->iboe;
258 	struct net_device *ndev;
259 	enum ib_mtu tmp;
260 
261 	props->active_width	= IB_WIDTH_1X;
262 	props->active_speed	= 4;
263 	props->port_cap_flags	= IB_PORT_CM_SUP;
264 	props->gid_tbl_len	= to_mdev(ibdev)->dev->caps.gid_table_len[port];
265 	props->max_msg_sz	= to_mdev(ibdev)->dev->caps.max_msg_sz;
266 	props->pkey_tbl_len	= 1;
267 	props->bad_pkey_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 46));
268 	props->qkey_viol_cntr	= be16_to_cpup((__be16 *) (out_mad->data + 48));
269 	props->max_mtu		= IB_MTU_4096;
270 	props->subnet_timeout	= 0;
271 	props->max_vl_num	= out_mad->data[37] >> 4;
272 	props->init_type_reply	= 0;
273 	props->state		= IB_PORT_DOWN;
274 	props->phys_state	= state_to_phys_state(props->state);
275 	props->active_mtu	= IB_MTU_256;
276 	spin_lock(&iboe->lock);
277 	ndev = iboe->netdevs[port - 1];
278 	if (!ndev)
279 		goto out;
280 
281 	tmp = iboe_get_mtu(ndev->mtu);
282 	props->active_mtu = tmp ? min(props->max_mtu, tmp) : IB_MTU_256;
283 
284 	props->state		= (netif_running(ndev) && netif_carrier_ok(ndev)) ?
285 					IB_PORT_ACTIVE : IB_PORT_DOWN;
286 	props->phys_state	= state_to_phys_state(props->state);
287 
288 out:
289 	spin_unlock(&iboe->lock);
290 	return 0;
291 }
292 
293 static int mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
294 			      struct ib_port_attr *props)
295 {
296 	struct ib_smp *in_mad  = NULL;
297 	struct ib_smp *out_mad = NULL;
298 	int err = -ENOMEM;
299 
300 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
301 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
302 	if (!in_mad || !out_mad)
303 		goto out;
304 
305 	memset(props, 0, sizeof *props);
306 
307 	init_query_mad(in_mad);
308 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
309 	in_mad->attr_mod = cpu_to_be32(port);
310 
311 	err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
312 	if (err)
313 		goto out;
314 
315 	err = mlx4_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND ?
316 		ib_link_query_port(ibdev, port, props, in_mad, out_mad) :
317 		eth_link_query_port(ibdev, port, props, out_mad);
318 
319 out:
320 	kfree(in_mad);
321 	kfree(out_mad);
322 
323 	return err;
324 }
325 
326 static int __mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
327 			       union ib_gid *gid)
328 {
329 	struct ib_smp *in_mad  = NULL;
330 	struct ib_smp *out_mad = NULL;
331 	int err = -ENOMEM;
332 
333 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
334 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
335 	if (!in_mad || !out_mad)
336 		goto out;
337 
338 	init_query_mad(in_mad);
339 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
340 	in_mad->attr_mod = cpu_to_be32(port);
341 
342 	err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
343 	if (err)
344 		goto out;
345 
346 	memcpy(gid->raw, out_mad->data + 8, 8);
347 
348 	init_query_mad(in_mad);
349 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
350 	in_mad->attr_mod = cpu_to_be32(index / 8);
351 
352 	err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
353 	if (err)
354 		goto out;
355 
356 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
357 
358 out:
359 	kfree(in_mad);
360 	kfree(out_mad);
361 	return err;
362 }
363 
364 static int iboe_query_gid(struct ib_device *ibdev, u8 port, int index,
365 			  union ib_gid *gid)
366 {
367 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
368 
369 	*gid = dev->iboe.gid_table[port - 1][index];
370 
371 	return 0;
372 }
373 
374 static int mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
375 			     union ib_gid *gid)
376 {
377 	if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
378 		return __mlx4_ib_query_gid(ibdev, port, index, gid);
379 	else
380 		return iboe_query_gid(ibdev, port, index, gid);
381 }
382 
383 static int mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
384 			      u16 *pkey)
385 {
386 	struct ib_smp *in_mad  = NULL;
387 	struct ib_smp *out_mad = NULL;
388 	int err = -ENOMEM;
389 
390 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
391 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
392 	if (!in_mad || !out_mad)
393 		goto out;
394 
395 	init_query_mad(in_mad);
396 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
397 	in_mad->attr_mod = cpu_to_be32(index / 32);
398 
399 	err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
400 	if (err)
401 		goto out;
402 
403 	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
404 
405 out:
406 	kfree(in_mad);
407 	kfree(out_mad);
408 	return err;
409 }
410 
411 static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
412 				 struct ib_device_modify *props)
413 {
414 	struct mlx4_cmd_mailbox *mailbox;
415 
416 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
417 		return -EOPNOTSUPP;
418 
419 	if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
420 		return 0;
421 
422 	spin_lock(&to_mdev(ibdev)->sm_lock);
423 	memcpy(ibdev->node_desc, props->node_desc, 64);
424 	spin_unlock(&to_mdev(ibdev)->sm_lock);
425 
426 	/*
427 	 * If possible, pass node desc to FW, so it can generate
428 	 * a 144 trap.  If cmd fails, just ignore.
429 	 */
430 	mailbox = mlx4_alloc_cmd_mailbox(to_mdev(ibdev)->dev);
431 	if (IS_ERR(mailbox))
432 		return 0;
433 
434 	memset(mailbox->buf, 0, 256);
435 	memcpy(mailbox->buf, props->node_desc, 64);
436 	mlx4_cmd(to_mdev(ibdev)->dev, mailbox->dma, 1, 0,
437 		 MLX4_CMD_SET_NODE, MLX4_CMD_TIME_CLASS_A);
438 
439 	mlx4_free_cmd_mailbox(to_mdev(ibdev)->dev, mailbox);
440 
441 	return 0;
442 }
443 
444 static int mlx4_SET_PORT(struct mlx4_ib_dev *dev, u8 port, int reset_qkey_viols,
445 			 u32 cap_mask)
446 {
447 	struct mlx4_cmd_mailbox *mailbox;
448 	int err;
449 	u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
450 
451 	mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
452 	if (IS_ERR(mailbox))
453 		return PTR_ERR(mailbox);
454 
455 	memset(mailbox->buf, 0, 256);
456 
457 	if (dev->dev->flags & MLX4_FLAG_OLD_PORT_CMDS) {
458 		*(u8 *) mailbox->buf	     = !!reset_qkey_viols << 6;
459 		((__be32 *) mailbox->buf)[2] = cpu_to_be32(cap_mask);
460 	} else {
461 		((u8 *) mailbox->buf)[3]     = !!reset_qkey_viols;
462 		((__be32 *) mailbox->buf)[1] = cpu_to_be32(cap_mask);
463 	}
464 
465 	err = mlx4_cmd(dev->dev, mailbox->dma, port, is_eth, MLX4_CMD_SET_PORT,
466 		       MLX4_CMD_TIME_CLASS_B);
467 
468 	mlx4_free_cmd_mailbox(dev->dev, mailbox);
469 	return err;
470 }
471 
472 static int mlx4_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
473 			       struct ib_port_modify *props)
474 {
475 	struct ib_port_attr attr;
476 	u32 cap_mask;
477 	int err;
478 
479 	mutex_lock(&to_mdev(ibdev)->cap_mask_mutex);
480 
481 	err = mlx4_ib_query_port(ibdev, port, &attr);
482 	if (err)
483 		goto out;
484 
485 	cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
486 		~props->clr_port_cap_mask;
487 
488 	err = mlx4_SET_PORT(to_mdev(ibdev), port,
489 			    !!(mask & IB_PORT_RESET_QKEY_CNTR),
490 			    cap_mask);
491 
492 out:
493 	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
494 	return err;
495 }
496 
497 static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
498 						  struct ib_udata *udata)
499 {
500 	struct mlx4_ib_dev *dev = to_mdev(ibdev);
501 	struct mlx4_ib_ucontext *context;
502 	struct mlx4_ib_alloc_ucontext_resp resp;
503 	int err;
504 
505 	if (!dev->ib_active)
506 		return ERR_PTR(-EAGAIN);
507 
508 	resp.qp_tab_size      = dev->dev->caps.num_qps;
509 	resp.bf_reg_size      = dev->dev->caps.bf_reg_size;
510 	resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
511 
512 	context = kmalloc(sizeof *context, GFP_KERNEL);
513 	if (!context)
514 		return ERR_PTR(-ENOMEM);
515 
516 	err = mlx4_uar_alloc(to_mdev(ibdev)->dev, &context->uar);
517 	if (err) {
518 		kfree(context);
519 		return ERR_PTR(err);
520 	}
521 
522 	INIT_LIST_HEAD(&context->db_page_list);
523 	mutex_init(&context->db_page_mutex);
524 
525 	err = ib_copy_to_udata(udata, &resp, sizeof resp);
526 	if (err) {
527 		mlx4_uar_free(to_mdev(ibdev)->dev, &context->uar);
528 		kfree(context);
529 		return ERR_PTR(-EFAULT);
530 	}
531 
532 	return &context->ibucontext;
533 }
534 
535 static int mlx4_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
536 {
537 	struct mlx4_ib_ucontext *context = to_mucontext(ibcontext);
538 
539 	mlx4_uar_free(to_mdev(ibcontext->device)->dev, &context->uar);
540 	kfree(context);
541 
542 	return 0;
543 }
544 
545 static int mlx4_ib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
546 {
547 	struct mlx4_ib_dev *dev = to_mdev(context->device);
548 
549 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
550 		return -EINVAL;
551 
552 	if (vma->vm_pgoff == 0) {
553 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
554 
555 		if (io_remap_pfn_range(vma, vma->vm_start,
556 				       to_mucontext(context)->uar.pfn,
557 				       PAGE_SIZE, vma->vm_page_prot))
558 			return -EAGAIN;
559 	} else if (vma->vm_pgoff == 1 && dev->dev->caps.bf_reg_size != 0) {
560 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
561 
562 		if (io_remap_pfn_range(vma, vma->vm_start,
563 				       to_mucontext(context)->uar.pfn +
564 				       dev->dev->caps.num_uars,
565 				       PAGE_SIZE, vma->vm_page_prot))
566 			return -EAGAIN;
567 	} else
568 		return -EINVAL;
569 
570 	return 0;
571 }
572 
573 static struct ib_pd *mlx4_ib_alloc_pd(struct ib_device *ibdev,
574 				      struct ib_ucontext *context,
575 				      struct ib_udata *udata)
576 {
577 	struct mlx4_ib_pd *pd;
578 	int err;
579 
580 	pd = kmalloc(sizeof *pd, GFP_KERNEL);
581 	if (!pd)
582 		return ERR_PTR(-ENOMEM);
583 
584 	err = mlx4_pd_alloc(to_mdev(ibdev)->dev, &pd->pdn);
585 	if (err) {
586 		kfree(pd);
587 		return ERR_PTR(err);
588 	}
589 
590 	if (context)
591 		if (ib_copy_to_udata(udata, &pd->pdn, sizeof (__u32))) {
592 			mlx4_pd_free(to_mdev(ibdev)->dev, pd->pdn);
593 			kfree(pd);
594 			return ERR_PTR(-EFAULT);
595 		}
596 
597 	return &pd->ibpd;
598 }
599 
600 static int mlx4_ib_dealloc_pd(struct ib_pd *pd)
601 {
602 	mlx4_pd_free(to_mdev(pd->device)->dev, to_mpd(pd)->pdn);
603 	kfree(pd);
604 
605 	return 0;
606 }
607 
608 static struct ib_xrcd *mlx4_ib_alloc_xrcd(struct ib_device *ibdev,
609 					  struct ib_ucontext *context,
610 					  struct ib_udata *udata)
611 {
612 	struct mlx4_ib_xrcd *xrcd;
613 	int err;
614 
615 	if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
616 		return ERR_PTR(-ENOSYS);
617 
618 	xrcd = kmalloc(sizeof *xrcd, GFP_KERNEL);
619 	if (!xrcd)
620 		return ERR_PTR(-ENOMEM);
621 
622 	err = mlx4_xrcd_alloc(to_mdev(ibdev)->dev, &xrcd->xrcdn);
623 	if (err)
624 		goto err1;
625 
626 	xrcd->pd = ib_alloc_pd(ibdev);
627 	if (IS_ERR(xrcd->pd)) {
628 		err = PTR_ERR(xrcd->pd);
629 		goto err2;
630 	}
631 
632 	xrcd->cq = ib_create_cq(ibdev, NULL, NULL, xrcd, 1, 0);
633 	if (IS_ERR(xrcd->cq)) {
634 		err = PTR_ERR(xrcd->cq);
635 		goto err3;
636 	}
637 
638 	return &xrcd->ibxrcd;
639 
640 err3:
641 	ib_dealloc_pd(xrcd->pd);
642 err2:
643 	mlx4_xrcd_free(to_mdev(ibdev)->dev, xrcd->xrcdn);
644 err1:
645 	kfree(xrcd);
646 	return ERR_PTR(err);
647 }
648 
649 static int mlx4_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
650 {
651 	ib_destroy_cq(to_mxrcd(xrcd)->cq);
652 	ib_dealloc_pd(to_mxrcd(xrcd)->pd);
653 	mlx4_xrcd_free(to_mdev(xrcd->device)->dev, to_mxrcd(xrcd)->xrcdn);
654 	kfree(xrcd);
655 
656 	return 0;
657 }
658 
659 static int add_gid_entry(struct ib_qp *ibqp, union ib_gid *gid)
660 {
661 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
662 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
663 	struct mlx4_ib_gid_entry *ge;
664 
665 	ge = kzalloc(sizeof *ge, GFP_KERNEL);
666 	if (!ge)
667 		return -ENOMEM;
668 
669 	ge->gid = *gid;
670 	if (mlx4_ib_add_mc(mdev, mqp, gid)) {
671 		ge->port = mqp->port;
672 		ge->added = 1;
673 	}
674 
675 	mutex_lock(&mqp->mutex);
676 	list_add_tail(&ge->list, &mqp->gid_list);
677 	mutex_unlock(&mqp->mutex);
678 
679 	return 0;
680 }
681 
682 int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
683 		   union ib_gid *gid)
684 {
685 	u8 mac[6];
686 	struct net_device *ndev;
687 	int ret = 0;
688 
689 	if (!mqp->port)
690 		return 0;
691 
692 	spin_lock(&mdev->iboe.lock);
693 	ndev = mdev->iboe.netdevs[mqp->port - 1];
694 	if (ndev)
695 		dev_hold(ndev);
696 	spin_unlock(&mdev->iboe.lock);
697 
698 	if (ndev) {
699 		rdma_get_mcast_mac((struct in6_addr *)gid, mac);
700 		rtnl_lock();
701 		dev_mc_add(mdev->iboe.netdevs[mqp->port - 1], mac);
702 		ret = 1;
703 		rtnl_unlock();
704 		dev_put(ndev);
705 	}
706 
707 	return ret;
708 }
709 
710 static int mlx4_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
711 {
712 	int err;
713 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
714 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
715 
716 	err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw,
717 				    !!(mqp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
718 				    MLX4_PROT_IB_IPV6);
719 	if (err)
720 		return err;
721 
722 	err = add_gid_entry(ibqp, gid);
723 	if (err)
724 		goto err_add;
725 
726 	return 0;
727 
728 err_add:
729 	mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw, MLX4_PROT_IB_IPV6);
730 	return err;
731 }
732 
733 static struct mlx4_ib_gid_entry *find_gid_entry(struct mlx4_ib_qp *qp, u8 *raw)
734 {
735 	struct mlx4_ib_gid_entry *ge;
736 	struct mlx4_ib_gid_entry *tmp;
737 	struct mlx4_ib_gid_entry *ret = NULL;
738 
739 	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
740 		if (!memcmp(raw, ge->gid.raw, 16)) {
741 			ret = ge;
742 			break;
743 		}
744 	}
745 
746 	return ret;
747 }
748 
749 static int mlx4_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
750 {
751 	int err;
752 	struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
753 	struct mlx4_ib_qp *mqp = to_mqp(ibqp);
754 	u8 mac[6];
755 	struct net_device *ndev;
756 	struct mlx4_ib_gid_entry *ge;
757 
758 	err = mlx4_multicast_detach(mdev->dev,
759 				    &mqp->mqp, gid->raw, MLX4_PROT_IB_IPV6);
760 	if (err)
761 		return err;
762 
763 	mutex_lock(&mqp->mutex);
764 	ge = find_gid_entry(mqp, gid->raw);
765 	if (ge) {
766 		spin_lock(&mdev->iboe.lock);
767 		ndev = ge->added ? mdev->iboe.netdevs[ge->port - 1] : NULL;
768 		if (ndev)
769 			dev_hold(ndev);
770 		spin_unlock(&mdev->iboe.lock);
771 		rdma_get_mcast_mac((struct in6_addr *)gid, mac);
772 		if (ndev) {
773 			rtnl_lock();
774 			dev_mc_del(mdev->iboe.netdevs[ge->port - 1], mac);
775 			rtnl_unlock();
776 			dev_put(ndev);
777 		}
778 		list_del(&ge->list);
779 		kfree(ge);
780 	} else
781 		printk(KERN_WARNING "could not find mgid entry\n");
782 
783 	mutex_unlock(&mqp->mutex);
784 
785 	return 0;
786 }
787 
788 static int init_node_data(struct mlx4_ib_dev *dev)
789 {
790 	struct ib_smp *in_mad  = NULL;
791 	struct ib_smp *out_mad = NULL;
792 	int err = -ENOMEM;
793 
794 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
795 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
796 	if (!in_mad || !out_mad)
797 		goto out;
798 
799 	init_query_mad(in_mad);
800 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
801 
802 	err = mlx4_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
803 	if (err)
804 		goto out;
805 
806 	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
807 
808 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
809 
810 	err = mlx4_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
811 	if (err)
812 		goto out;
813 
814 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
815 
816 out:
817 	kfree(in_mad);
818 	kfree(out_mad);
819 	return err;
820 }
821 
822 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
823 			char *buf)
824 {
825 	struct mlx4_ib_dev *dev =
826 		container_of(device, struct mlx4_ib_dev, ib_dev.dev);
827 	return sprintf(buf, "MT%d\n", dev->dev->pdev->device);
828 }
829 
830 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
831 			   char *buf)
832 {
833 	struct mlx4_ib_dev *dev =
834 		container_of(device, struct mlx4_ib_dev, ib_dev.dev);
835 	return sprintf(buf, "%d.%d.%d\n", (int) (dev->dev->caps.fw_ver >> 32),
836 		       (int) (dev->dev->caps.fw_ver >> 16) & 0xffff,
837 		       (int) dev->dev->caps.fw_ver & 0xffff);
838 }
839 
840 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
841 			char *buf)
842 {
843 	struct mlx4_ib_dev *dev =
844 		container_of(device, struct mlx4_ib_dev, ib_dev.dev);
845 	return sprintf(buf, "%x\n", dev->dev->rev_id);
846 }
847 
848 static ssize_t show_board(struct device *device, struct device_attribute *attr,
849 			  char *buf)
850 {
851 	struct mlx4_ib_dev *dev =
852 		container_of(device, struct mlx4_ib_dev, ib_dev.dev);
853 	return sprintf(buf, "%.*s\n", MLX4_BOARD_ID_LEN,
854 		       dev->dev->board_id);
855 }
856 
857 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
858 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
859 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
860 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
861 
862 static struct device_attribute *mlx4_class_attributes[] = {
863 	&dev_attr_hw_rev,
864 	&dev_attr_fw_ver,
865 	&dev_attr_hca_type,
866 	&dev_attr_board_id
867 };
868 
869 static void mlx4_addrconf_ifid_eui48(u8 *eui, u16 vlan_id, struct net_device *dev)
870 {
871 	memcpy(eui, dev->dev_addr, 3);
872 	memcpy(eui + 5, dev->dev_addr + 3, 3);
873 	if (vlan_id < 0x1000) {
874 		eui[3] = vlan_id >> 8;
875 		eui[4] = vlan_id & 0xff;
876 	} else {
877 		eui[3] = 0xff;
878 		eui[4] = 0xfe;
879 	}
880 	eui[0] ^= 2;
881 }
882 
883 static void update_gids_task(struct work_struct *work)
884 {
885 	struct update_gid_work *gw = container_of(work, struct update_gid_work, work);
886 	struct mlx4_cmd_mailbox *mailbox;
887 	union ib_gid *gids;
888 	int err;
889 	struct mlx4_dev	*dev = gw->dev->dev;
890 	struct ib_event event;
891 
892 	mailbox = mlx4_alloc_cmd_mailbox(dev);
893 	if (IS_ERR(mailbox)) {
894 		printk(KERN_WARNING "update gid table failed %ld\n", PTR_ERR(mailbox));
895 		return;
896 	}
897 
898 	gids = mailbox->buf;
899 	memcpy(gids, gw->gids, sizeof gw->gids);
900 
901 	err = mlx4_cmd(dev, mailbox->dma, MLX4_SET_PORT_GID_TABLE << 8 | gw->port,
902 		       1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B);
903 	if (err)
904 		printk(KERN_WARNING "set port command failed\n");
905 	else {
906 		memcpy(gw->dev->iboe.gid_table[gw->port - 1], gw->gids, sizeof gw->gids);
907 		event.device = &gw->dev->ib_dev;
908 		event.element.port_num = gw->port;
909 		event.event    = IB_EVENT_GID_CHANGE;
910 		ib_dispatch_event(&event);
911 	}
912 
913 	mlx4_free_cmd_mailbox(dev, mailbox);
914 	kfree(gw);
915 }
916 
917 static int update_ipv6_gids(struct mlx4_ib_dev *dev, int port, int clear)
918 {
919 	struct net_device *ndev = dev->iboe.netdevs[port - 1];
920 	struct update_gid_work *work;
921 	struct net_device *tmp;
922 	int i;
923 	u8 *hits;
924 	int ret;
925 	union ib_gid gid;
926 	int free;
927 	int found;
928 	int need_update = 0;
929 	u16 vid;
930 
931 	work = kzalloc(sizeof *work, GFP_ATOMIC);
932 	if (!work)
933 		return -ENOMEM;
934 
935 	hits = kzalloc(128, GFP_ATOMIC);
936 	if (!hits) {
937 		ret = -ENOMEM;
938 		goto out;
939 	}
940 
941 	rcu_read_lock();
942 	for_each_netdev_rcu(&init_net, tmp) {
943 		if (ndev && (tmp == ndev || rdma_vlan_dev_real_dev(tmp) == ndev)) {
944 			gid.global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
945 			vid = rdma_vlan_dev_vlan_id(tmp);
946 			mlx4_addrconf_ifid_eui48(&gid.raw[8], vid, ndev);
947 			found = 0;
948 			free = -1;
949 			for (i = 0; i < 128; ++i) {
950 				if (free < 0 &&
951 				    !memcmp(&dev->iboe.gid_table[port - 1][i], &zgid, sizeof zgid))
952 					free = i;
953 				if (!memcmp(&dev->iboe.gid_table[port - 1][i], &gid, sizeof gid)) {
954 					hits[i] = 1;
955 					found = 1;
956 					break;
957 				}
958 			}
959 
960 			if (!found) {
961 				if (tmp == ndev &&
962 				    (memcmp(&dev->iboe.gid_table[port - 1][0],
963 					    &gid, sizeof gid) ||
964 				     !memcmp(&dev->iboe.gid_table[port - 1][0],
965 					     &zgid, sizeof gid))) {
966 					dev->iboe.gid_table[port - 1][0] = gid;
967 					++need_update;
968 					hits[0] = 1;
969 				} else if (free >= 0) {
970 					dev->iboe.gid_table[port - 1][free] = gid;
971 					hits[free] = 1;
972 					++need_update;
973 				}
974 			}
975 		}
976 	}
977 	rcu_read_unlock();
978 
979 	for (i = 0; i < 128; ++i)
980 		if (!hits[i]) {
981 			if (memcmp(&dev->iboe.gid_table[port - 1][i], &zgid, sizeof zgid))
982 				++need_update;
983 			dev->iboe.gid_table[port - 1][i] = zgid;
984 		}
985 
986 	if (need_update) {
987 		memcpy(work->gids, dev->iboe.gid_table[port - 1], sizeof work->gids);
988 		INIT_WORK(&work->work, update_gids_task);
989 		work->port = port;
990 		work->dev = dev;
991 		queue_work(wq, &work->work);
992 	} else
993 		kfree(work);
994 
995 	kfree(hits);
996 	return 0;
997 
998 out:
999 	kfree(work);
1000 	return ret;
1001 }
1002 
1003 static void handle_en_event(struct mlx4_ib_dev *dev, int port, unsigned long event)
1004 {
1005 	switch (event) {
1006 	case NETDEV_UP:
1007 	case NETDEV_CHANGEADDR:
1008 		update_ipv6_gids(dev, port, 0);
1009 		break;
1010 
1011 	case NETDEV_DOWN:
1012 		update_ipv6_gids(dev, port, 1);
1013 		dev->iboe.netdevs[port - 1] = NULL;
1014 	}
1015 }
1016 
1017 static void netdev_added(struct mlx4_ib_dev *dev, int port)
1018 {
1019 	update_ipv6_gids(dev, port, 0);
1020 }
1021 
1022 static void netdev_removed(struct mlx4_ib_dev *dev, int port)
1023 {
1024 	update_ipv6_gids(dev, port, 1);
1025 }
1026 
1027 static int mlx4_ib_netdev_event(struct notifier_block *this, unsigned long event,
1028 				void *ptr)
1029 {
1030 	struct net_device *dev = ptr;
1031 	struct mlx4_ib_dev *ibdev;
1032 	struct net_device *oldnd;
1033 	struct mlx4_ib_iboe *iboe;
1034 	int port;
1035 
1036 	if (!net_eq(dev_net(dev), &init_net))
1037 		return NOTIFY_DONE;
1038 
1039 	ibdev = container_of(this, struct mlx4_ib_dev, iboe.nb);
1040 	iboe = &ibdev->iboe;
1041 
1042 	spin_lock(&iboe->lock);
1043 	mlx4_foreach_ib_transport_port(port, ibdev->dev) {
1044 		oldnd = iboe->netdevs[port - 1];
1045 		iboe->netdevs[port - 1] =
1046 			mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port);
1047 		if (oldnd != iboe->netdevs[port - 1]) {
1048 			if (iboe->netdevs[port - 1])
1049 				netdev_added(ibdev, port);
1050 			else
1051 				netdev_removed(ibdev, port);
1052 		}
1053 	}
1054 
1055 	if (dev == iboe->netdevs[0] ||
1056 	    (iboe->netdevs[0] && rdma_vlan_dev_real_dev(dev) == iboe->netdevs[0]))
1057 		handle_en_event(ibdev, 1, event);
1058 	else if (dev == iboe->netdevs[1]
1059 		 || (iboe->netdevs[1] && rdma_vlan_dev_real_dev(dev) == iboe->netdevs[1]))
1060 		handle_en_event(ibdev, 2, event);
1061 
1062 	spin_unlock(&iboe->lock);
1063 
1064 	return NOTIFY_DONE;
1065 }
1066 
1067 static void *mlx4_ib_add(struct mlx4_dev *dev)
1068 {
1069 	struct mlx4_ib_dev *ibdev;
1070 	int num_ports = 0;
1071 	int i;
1072 	int err;
1073 	struct mlx4_ib_iboe *iboe;
1074 
1075 	printk_once(KERN_INFO "%s", mlx4_ib_version);
1076 
1077 	mlx4_foreach_ib_transport_port(i, dev)
1078 		num_ports++;
1079 
1080 	/* No point in registering a device with no ports... */
1081 	if (num_ports == 0)
1082 		return NULL;
1083 
1084 	ibdev = (struct mlx4_ib_dev *) ib_alloc_device(sizeof *ibdev);
1085 	if (!ibdev) {
1086 		dev_err(&dev->pdev->dev, "Device struct alloc failed\n");
1087 		return NULL;
1088 	}
1089 
1090 	iboe = &ibdev->iboe;
1091 
1092 	if (mlx4_pd_alloc(dev, &ibdev->priv_pdn))
1093 		goto err_dealloc;
1094 
1095 	if (mlx4_uar_alloc(dev, &ibdev->priv_uar))
1096 		goto err_pd;
1097 
1098 	ibdev->uar_map = ioremap((phys_addr_t) ibdev->priv_uar.pfn << PAGE_SHIFT,
1099 				 PAGE_SIZE);
1100 	if (!ibdev->uar_map)
1101 		goto err_uar;
1102 	MLX4_INIT_DOORBELL_LOCK(&ibdev->uar_lock);
1103 
1104 	ibdev->dev = dev;
1105 
1106 	strlcpy(ibdev->ib_dev.name, "mlx4_%d", IB_DEVICE_NAME_MAX);
1107 	ibdev->ib_dev.owner		= THIS_MODULE;
1108 	ibdev->ib_dev.node_type		= RDMA_NODE_IB_CA;
1109 	ibdev->ib_dev.local_dma_lkey	= dev->caps.reserved_lkey;
1110 	ibdev->num_ports		= num_ports;
1111 	ibdev->ib_dev.phys_port_cnt     = ibdev->num_ports;
1112 	ibdev->ib_dev.num_comp_vectors	= dev->caps.num_comp_vectors;
1113 	ibdev->ib_dev.dma_device	= &dev->pdev->dev;
1114 
1115 	ibdev->ib_dev.uverbs_abi_ver	= MLX4_IB_UVERBS_ABI_VERSION;
1116 	ibdev->ib_dev.uverbs_cmd_mask	=
1117 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1118 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1119 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1120 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1121 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1122 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1123 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1124 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1125 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1126 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1127 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1128 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1129 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1130 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1131 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1132 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1133 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST)	|
1134 		(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1135 		(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1136 		(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1137 		(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ)		|
1138 		(1ull << IB_USER_VERBS_CMD_CREATE_XSRQ)		|
1139 		(1ull << IB_USER_VERBS_CMD_OPEN_QP);
1140 
1141 	ibdev->ib_dev.query_device	= mlx4_ib_query_device;
1142 	ibdev->ib_dev.query_port	= mlx4_ib_query_port;
1143 	ibdev->ib_dev.get_link_layer	= mlx4_ib_port_link_layer;
1144 	ibdev->ib_dev.query_gid		= mlx4_ib_query_gid;
1145 	ibdev->ib_dev.query_pkey	= mlx4_ib_query_pkey;
1146 	ibdev->ib_dev.modify_device	= mlx4_ib_modify_device;
1147 	ibdev->ib_dev.modify_port	= mlx4_ib_modify_port;
1148 	ibdev->ib_dev.alloc_ucontext	= mlx4_ib_alloc_ucontext;
1149 	ibdev->ib_dev.dealloc_ucontext	= mlx4_ib_dealloc_ucontext;
1150 	ibdev->ib_dev.mmap		= mlx4_ib_mmap;
1151 	ibdev->ib_dev.alloc_pd		= mlx4_ib_alloc_pd;
1152 	ibdev->ib_dev.dealloc_pd	= mlx4_ib_dealloc_pd;
1153 	ibdev->ib_dev.create_ah		= mlx4_ib_create_ah;
1154 	ibdev->ib_dev.query_ah		= mlx4_ib_query_ah;
1155 	ibdev->ib_dev.destroy_ah	= mlx4_ib_destroy_ah;
1156 	ibdev->ib_dev.create_srq	= mlx4_ib_create_srq;
1157 	ibdev->ib_dev.modify_srq	= mlx4_ib_modify_srq;
1158 	ibdev->ib_dev.query_srq		= mlx4_ib_query_srq;
1159 	ibdev->ib_dev.destroy_srq	= mlx4_ib_destroy_srq;
1160 	ibdev->ib_dev.post_srq_recv	= mlx4_ib_post_srq_recv;
1161 	ibdev->ib_dev.create_qp		= mlx4_ib_create_qp;
1162 	ibdev->ib_dev.modify_qp		= mlx4_ib_modify_qp;
1163 	ibdev->ib_dev.query_qp		= mlx4_ib_query_qp;
1164 	ibdev->ib_dev.destroy_qp	= mlx4_ib_destroy_qp;
1165 	ibdev->ib_dev.post_send		= mlx4_ib_post_send;
1166 	ibdev->ib_dev.post_recv		= mlx4_ib_post_recv;
1167 	ibdev->ib_dev.create_cq		= mlx4_ib_create_cq;
1168 	ibdev->ib_dev.modify_cq		= mlx4_ib_modify_cq;
1169 	ibdev->ib_dev.resize_cq		= mlx4_ib_resize_cq;
1170 	ibdev->ib_dev.destroy_cq	= mlx4_ib_destroy_cq;
1171 	ibdev->ib_dev.poll_cq		= mlx4_ib_poll_cq;
1172 	ibdev->ib_dev.req_notify_cq	= mlx4_ib_arm_cq;
1173 	ibdev->ib_dev.get_dma_mr	= mlx4_ib_get_dma_mr;
1174 	ibdev->ib_dev.reg_user_mr	= mlx4_ib_reg_user_mr;
1175 	ibdev->ib_dev.dereg_mr		= mlx4_ib_dereg_mr;
1176 	ibdev->ib_dev.alloc_fast_reg_mr = mlx4_ib_alloc_fast_reg_mr;
1177 	ibdev->ib_dev.alloc_fast_reg_page_list = mlx4_ib_alloc_fast_reg_page_list;
1178 	ibdev->ib_dev.free_fast_reg_page_list  = mlx4_ib_free_fast_reg_page_list;
1179 	ibdev->ib_dev.attach_mcast	= mlx4_ib_mcg_attach;
1180 	ibdev->ib_dev.detach_mcast	= mlx4_ib_mcg_detach;
1181 	ibdev->ib_dev.process_mad	= mlx4_ib_process_mad;
1182 
1183 	ibdev->ib_dev.alloc_fmr		= mlx4_ib_fmr_alloc;
1184 	ibdev->ib_dev.map_phys_fmr	= mlx4_ib_map_phys_fmr;
1185 	ibdev->ib_dev.unmap_fmr		= mlx4_ib_unmap_fmr;
1186 	ibdev->ib_dev.dealloc_fmr	= mlx4_ib_fmr_dealloc;
1187 
1188 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) {
1189 		ibdev->ib_dev.alloc_xrcd = mlx4_ib_alloc_xrcd;
1190 		ibdev->ib_dev.dealloc_xrcd = mlx4_ib_dealloc_xrcd;
1191 		ibdev->ib_dev.uverbs_cmd_mask |=
1192 			(1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
1193 			(1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
1194 	}
1195 
1196 	spin_lock_init(&iboe->lock);
1197 
1198 	if (init_node_data(ibdev))
1199 		goto err_map;
1200 
1201 	for (i = 0; i < ibdev->num_ports; ++i) {
1202 		if (mlx4_ib_port_link_layer(&ibdev->ib_dev, i + 1) ==
1203 						IB_LINK_LAYER_ETHERNET) {
1204 			err = mlx4_counter_alloc(ibdev->dev, &ibdev->counters[i]);
1205 			if (err)
1206 				ibdev->counters[i] = -1;
1207 		} else
1208 				ibdev->counters[i] = -1;
1209 	}
1210 
1211 	spin_lock_init(&ibdev->sm_lock);
1212 	mutex_init(&ibdev->cap_mask_mutex);
1213 
1214 	if (ib_register_device(&ibdev->ib_dev, NULL))
1215 		goto err_counter;
1216 
1217 	if (mlx4_ib_mad_init(ibdev))
1218 		goto err_reg;
1219 
1220 	if (dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE && !iboe->nb.notifier_call) {
1221 		iboe->nb.notifier_call = mlx4_ib_netdev_event;
1222 		err = register_netdevice_notifier(&iboe->nb);
1223 		if (err)
1224 			goto err_reg;
1225 	}
1226 
1227 	for (i = 0; i < ARRAY_SIZE(mlx4_class_attributes); ++i) {
1228 		if (device_create_file(&ibdev->ib_dev.dev,
1229 				       mlx4_class_attributes[i]))
1230 			goto err_notif;
1231 	}
1232 
1233 	ibdev->ib_active = true;
1234 
1235 	return ibdev;
1236 
1237 err_notif:
1238 	if (unregister_netdevice_notifier(&ibdev->iboe.nb))
1239 		printk(KERN_WARNING "failure unregistering notifier\n");
1240 	flush_workqueue(wq);
1241 
1242 err_reg:
1243 	ib_unregister_device(&ibdev->ib_dev);
1244 
1245 err_counter:
1246 	for (; i; --i)
1247 		mlx4_counter_free(ibdev->dev, ibdev->counters[i - 1]);
1248 
1249 err_map:
1250 	iounmap(ibdev->uar_map);
1251 
1252 err_uar:
1253 	mlx4_uar_free(dev, &ibdev->priv_uar);
1254 
1255 err_pd:
1256 	mlx4_pd_free(dev, ibdev->priv_pdn);
1257 
1258 err_dealloc:
1259 	ib_dealloc_device(&ibdev->ib_dev);
1260 
1261 	return NULL;
1262 }
1263 
1264 static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
1265 {
1266 	struct mlx4_ib_dev *ibdev = ibdev_ptr;
1267 	int p;
1268 
1269 	mlx4_ib_mad_cleanup(ibdev);
1270 	ib_unregister_device(&ibdev->ib_dev);
1271 	if (ibdev->iboe.nb.notifier_call) {
1272 		if (unregister_netdevice_notifier(&ibdev->iboe.nb))
1273 			printk(KERN_WARNING "failure unregistering notifier\n");
1274 		ibdev->iboe.nb.notifier_call = NULL;
1275 	}
1276 	iounmap(ibdev->uar_map);
1277 	for (p = 0; p < ibdev->num_ports; ++p)
1278 		mlx4_counter_free(ibdev->dev, ibdev->counters[p]);
1279 	mlx4_foreach_port(p, dev, MLX4_PORT_TYPE_IB)
1280 		mlx4_CLOSE_PORT(dev, p);
1281 
1282 	mlx4_uar_free(dev, &ibdev->priv_uar);
1283 	mlx4_pd_free(dev, ibdev->priv_pdn);
1284 	ib_dealloc_device(&ibdev->ib_dev);
1285 }
1286 
1287 static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
1288 			  enum mlx4_dev_event event, int port)
1289 {
1290 	struct ib_event ibev;
1291 	struct mlx4_ib_dev *ibdev = to_mdev((struct ib_device *) ibdev_ptr);
1292 
1293 	if (port > ibdev->num_ports)
1294 		return;
1295 
1296 	switch (event) {
1297 	case MLX4_DEV_EVENT_PORT_UP:
1298 		ibev.event = IB_EVENT_PORT_ACTIVE;
1299 		break;
1300 
1301 	case MLX4_DEV_EVENT_PORT_DOWN:
1302 		ibev.event = IB_EVENT_PORT_ERR;
1303 		break;
1304 
1305 	case MLX4_DEV_EVENT_CATASTROPHIC_ERROR:
1306 		ibdev->ib_active = false;
1307 		ibev.event = IB_EVENT_DEVICE_FATAL;
1308 		break;
1309 
1310 	default:
1311 		return;
1312 	}
1313 
1314 	ibev.device	      = ibdev_ptr;
1315 	ibev.element.port_num = port;
1316 
1317 	ib_dispatch_event(&ibev);
1318 }
1319 
1320 static struct mlx4_interface mlx4_ib_interface = {
1321 	.add		= mlx4_ib_add,
1322 	.remove		= mlx4_ib_remove,
1323 	.event		= mlx4_ib_event,
1324 	.protocol	= MLX4_PROT_IB_IPV6
1325 };
1326 
1327 static int __init mlx4_ib_init(void)
1328 {
1329 	int err;
1330 
1331 	wq = create_singlethread_workqueue("mlx4_ib");
1332 	if (!wq)
1333 		return -ENOMEM;
1334 
1335 	err = mlx4_register_interface(&mlx4_ib_interface);
1336 	if (err) {
1337 		destroy_workqueue(wq);
1338 		return err;
1339 	}
1340 
1341 	return 0;
1342 }
1343 
1344 static void __exit mlx4_ib_cleanup(void)
1345 {
1346 	mlx4_unregister_interface(&mlx4_ib_interface);
1347 	destroy_workqueue(wq);
1348 }
1349 
1350 module_init(mlx4_ib_init);
1351 module_exit(mlx4_ib_cleanup);
1352