2 * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
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>
43 #include <rdma/ib_smi.h>
44 #include <rdma/ib_user_verbs.h>
45 #include <rdma/ib_addr.h>
47 #include <linux/mlx4/driver.h>
48 #include <linux/mlx4/cmd.h>
53 #define DRV_NAME "mlx4_ib"
54 #define DRV_VERSION "1.0"
55 #define DRV_RELDATE "April 4, 2008"
57 MODULE_AUTHOR("Roland Dreier");
58 MODULE_DESCRIPTION("Mellanox ConnectX HCA InfiniBand driver");
59 MODULE_LICENSE("Dual BSD/GPL");
60 MODULE_VERSION(DRV_VERSION);
62 static const char mlx4_ib_version[] =
63 DRV_NAME ": Mellanox ConnectX InfiniBand driver v"
64 DRV_VERSION " (" DRV_RELDATE ")\n";
66 struct update_gid_work {
67 struct work_struct work;
68 union ib_gid gids[128];
69 struct mlx4_ib_dev *dev;
73 static struct workqueue_struct *wq;
75 static void init_query_mad(struct ib_smp *mad)
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;
83 static union ib_gid zgid;
85 static int mlx4_ib_query_device(struct ib_device *ibdev,
86 struct ib_device_attr *props)
88 struct mlx4_ib_dev *dev = to_mdev(ibdev);
89 struct ib_smp *in_mad = NULL;
90 struct ib_smp *out_mad = NULL;
93 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
94 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
95 if (!in_mad || !out_mad)
98 init_query_mad(in_mad);
99 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
101 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, 1, NULL, NULL, in_mad, out_mad);
105 memset(props, 0, sizeof *props);
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;
134 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
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);
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 props->max_ah = INT_MAX;
176 static enum rdma_link_layer
177 mlx4_ib_port_link_layer(struct ib_device *device, u8 port_num)
179 struct mlx4_dev *dev = to_mdev(device)->dev;
181 return dev->caps.port_mask & (1 << (port_num - 1)) ?
182 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
185 static int ib_link_query_port(struct ib_device *ibdev, u8 port,
186 struct ib_port_attr *props,
187 struct ib_smp *in_mad,
188 struct ib_smp *out_mad)
190 int ext_active_speed;
193 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16));
194 props->lmc = out_mad->data[34] & 0x7;
195 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18));
196 props->sm_sl = out_mad->data[36] & 0xf;
197 props->state = out_mad->data[32] & 0xf;
198 props->phys_state = out_mad->data[33] >> 4;
199 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20));
200 props->gid_tbl_len = to_mdev(ibdev)->dev->caps.gid_table_len[port];
201 props->max_msg_sz = to_mdev(ibdev)->dev->caps.max_msg_sz;
202 props->pkey_tbl_len = to_mdev(ibdev)->dev->caps.pkey_table_len[port];
203 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
204 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
205 props->active_width = out_mad->data[31] & 0xf;
206 props->active_speed = out_mad->data[35] >> 4;
207 props->max_mtu = out_mad->data[41] & 0xf;
208 props->active_mtu = out_mad->data[36] >> 4;
209 props->subnet_timeout = out_mad->data[51] & 0x1f;
210 props->max_vl_num = out_mad->data[37] >> 4;
211 props->init_type_reply = out_mad->data[41] >> 4;
213 /* Check if extended speeds (EDR/FDR/...) are supported */
214 if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
215 ext_active_speed = out_mad->data[62] >> 4;
217 switch (ext_active_speed) {
219 props->active_speed = 16; /* FDR */
222 props->active_speed = 32; /* EDR */
227 /* If reported active speed is QDR, check if is FDR-10 */
228 if (props->active_speed == 4) {
229 if (to_mdev(ibdev)->dev->caps.ext_port_cap[port] &
230 MLX_EXT_PORT_CAP_FLAG_EXTENDED_PORT_INFO) {
231 init_query_mad(in_mad);
232 in_mad->attr_id = MLX4_ATTR_EXTENDED_PORT_INFO;
233 in_mad->attr_mod = cpu_to_be32(port);
235 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port,
236 NULL, NULL, in_mad, out_mad);
240 /* Checking LinkSpeedActive for FDR-10 */
241 if (out_mad->data[15] & 0x1)
242 props->active_speed = 8;
249 static u8 state_to_phys_state(enum ib_port_state state)
251 return state == IB_PORT_ACTIVE ? 5 : 3;
254 static int eth_link_query_port(struct ib_device *ibdev, u8 port,
255 struct ib_port_attr *props,
256 struct ib_smp *out_mad)
258 struct mlx4_ib_iboe *iboe = &to_mdev(ibdev)->iboe;
259 struct net_device *ndev;
262 props->active_width = IB_WIDTH_1X;
263 props->active_speed = 4;
264 props->port_cap_flags = IB_PORT_CM_SUP;
265 props->gid_tbl_len = to_mdev(ibdev)->dev->caps.gid_table_len[port];
266 props->max_msg_sz = to_mdev(ibdev)->dev->caps.max_msg_sz;
267 props->pkey_tbl_len = 1;
268 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
269 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
270 props->max_mtu = IB_MTU_4096;
271 props->subnet_timeout = 0;
272 props->max_vl_num = out_mad->data[37] >> 4;
273 props->init_type_reply = 0;
274 props->state = IB_PORT_DOWN;
275 props->phys_state = state_to_phys_state(props->state);
276 props->active_mtu = IB_MTU_256;
277 spin_lock(&iboe->lock);
278 ndev = iboe->netdevs[port - 1];
282 tmp = iboe_get_mtu(ndev->mtu);
283 props->active_mtu = tmp ? min(props->max_mtu, tmp) : IB_MTU_256;
285 props->state = (netif_running(ndev) && netif_carrier_ok(ndev)) ?
286 IB_PORT_ACTIVE : IB_PORT_DOWN;
287 props->phys_state = state_to_phys_state(props->state);
290 spin_unlock(&iboe->lock);
294 static int mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
295 struct ib_port_attr *props)
297 struct ib_smp *in_mad = NULL;
298 struct ib_smp *out_mad = NULL;
301 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
302 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
303 if (!in_mad || !out_mad)
306 memset(props, 0, sizeof *props);
308 init_query_mad(in_mad);
309 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
310 in_mad->attr_mod = cpu_to_be32(port);
312 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
316 err = mlx4_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND ?
317 ib_link_query_port(ibdev, port, props, in_mad, out_mad) :
318 eth_link_query_port(ibdev, port, props, out_mad);
327 static int __mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
330 struct ib_smp *in_mad = NULL;
331 struct ib_smp *out_mad = NULL;
334 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
335 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
336 if (!in_mad || !out_mad)
339 init_query_mad(in_mad);
340 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
341 in_mad->attr_mod = cpu_to_be32(port);
343 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
347 memcpy(gid->raw, out_mad->data + 8, 8);
349 init_query_mad(in_mad);
350 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
351 in_mad->attr_mod = cpu_to_be32(index / 8);
353 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
357 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
365 static int iboe_query_gid(struct ib_device *ibdev, u8 port, int index,
368 struct mlx4_ib_dev *dev = to_mdev(ibdev);
370 *gid = dev->iboe.gid_table[port - 1][index];
375 static int mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
378 if (rdma_port_get_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND)
379 return __mlx4_ib_query_gid(ibdev, port, index, gid);
381 return iboe_query_gid(ibdev, port, index, gid);
384 static int mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
387 struct ib_smp *in_mad = NULL;
388 struct ib_smp *out_mad = NULL;
391 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
392 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
393 if (!in_mad || !out_mad)
396 init_query_mad(in_mad);
397 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE;
398 in_mad->attr_mod = cpu_to_be32(index / 32);
400 err = mlx4_MAD_IFC(to_mdev(ibdev), 1, 1, port, NULL, NULL, in_mad, out_mad);
404 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
412 static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
413 struct ib_device_modify *props)
415 struct mlx4_cmd_mailbox *mailbox;
417 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
420 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
423 spin_lock(&to_mdev(ibdev)->sm_lock);
424 memcpy(ibdev->node_desc, props->node_desc, 64);
425 spin_unlock(&to_mdev(ibdev)->sm_lock);
428 * If possible, pass node desc to FW, so it can generate
429 * a 144 trap. If cmd fails, just ignore.
431 mailbox = mlx4_alloc_cmd_mailbox(to_mdev(ibdev)->dev);
435 memset(mailbox->buf, 0, 256);
436 memcpy(mailbox->buf, props->node_desc, 64);
437 mlx4_cmd(to_mdev(ibdev)->dev, mailbox->dma, 1, 0,
438 MLX4_CMD_SET_NODE, MLX4_CMD_TIME_CLASS_A);
440 mlx4_free_cmd_mailbox(to_mdev(ibdev)->dev, mailbox);
445 static int mlx4_SET_PORT(struct mlx4_ib_dev *dev, u8 port, int reset_qkey_viols,
448 struct mlx4_cmd_mailbox *mailbox;
450 u8 is_eth = dev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
452 mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
454 return PTR_ERR(mailbox);
456 memset(mailbox->buf, 0, 256);
458 if (dev->dev->flags & MLX4_FLAG_OLD_PORT_CMDS) {
459 *(u8 *) mailbox->buf = !!reset_qkey_viols << 6;
460 ((__be32 *) mailbox->buf)[2] = cpu_to_be32(cap_mask);
462 ((u8 *) mailbox->buf)[3] = !!reset_qkey_viols;
463 ((__be32 *) mailbox->buf)[1] = cpu_to_be32(cap_mask);
466 err = mlx4_cmd(dev->dev, mailbox->dma, port, is_eth, MLX4_CMD_SET_PORT,
467 MLX4_CMD_TIME_CLASS_B);
469 mlx4_free_cmd_mailbox(dev->dev, mailbox);
473 static int mlx4_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
474 struct ib_port_modify *props)
476 struct ib_port_attr attr;
480 mutex_lock(&to_mdev(ibdev)->cap_mask_mutex);
482 err = mlx4_ib_query_port(ibdev, port, &attr);
486 cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
487 ~props->clr_port_cap_mask;
489 err = mlx4_SET_PORT(to_mdev(ibdev), port,
490 !!(mask & IB_PORT_RESET_QKEY_CNTR),
494 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
498 static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
499 struct ib_udata *udata)
501 struct mlx4_ib_dev *dev = to_mdev(ibdev);
502 struct mlx4_ib_ucontext *context;
503 struct mlx4_ib_alloc_ucontext_resp resp;
507 return ERR_PTR(-EAGAIN);
509 resp.qp_tab_size = dev->dev->caps.num_qps;
510 resp.bf_reg_size = dev->dev->caps.bf_reg_size;
511 resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
513 context = kmalloc(sizeof *context, GFP_KERNEL);
515 return ERR_PTR(-ENOMEM);
517 err = mlx4_uar_alloc(to_mdev(ibdev)->dev, &context->uar);
523 INIT_LIST_HEAD(&context->db_page_list);
524 mutex_init(&context->db_page_mutex);
526 err = ib_copy_to_udata(udata, &resp, sizeof resp);
528 mlx4_uar_free(to_mdev(ibdev)->dev, &context->uar);
530 return ERR_PTR(-EFAULT);
533 return &context->ibucontext;
536 static int mlx4_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
538 struct mlx4_ib_ucontext *context = to_mucontext(ibcontext);
540 mlx4_uar_free(to_mdev(ibcontext->device)->dev, &context->uar);
546 static int mlx4_ib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
548 struct mlx4_ib_dev *dev = to_mdev(context->device);
550 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
553 if (vma->vm_pgoff == 0) {
554 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
556 if (io_remap_pfn_range(vma, vma->vm_start,
557 to_mucontext(context)->uar.pfn,
558 PAGE_SIZE, vma->vm_page_prot))
560 } else if (vma->vm_pgoff == 1 && dev->dev->caps.bf_reg_size != 0) {
561 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
563 if (io_remap_pfn_range(vma, vma->vm_start,
564 to_mucontext(context)->uar.pfn +
565 dev->dev->caps.num_uars,
566 PAGE_SIZE, vma->vm_page_prot))
574 static struct ib_pd *mlx4_ib_alloc_pd(struct ib_device *ibdev,
575 struct ib_ucontext *context,
576 struct ib_udata *udata)
578 struct mlx4_ib_pd *pd;
581 pd = kmalloc(sizeof *pd, GFP_KERNEL);
583 return ERR_PTR(-ENOMEM);
585 err = mlx4_pd_alloc(to_mdev(ibdev)->dev, &pd->pdn);
592 if (ib_copy_to_udata(udata, &pd->pdn, sizeof (__u32))) {
593 mlx4_pd_free(to_mdev(ibdev)->dev, pd->pdn);
595 return ERR_PTR(-EFAULT);
601 static int mlx4_ib_dealloc_pd(struct ib_pd *pd)
603 mlx4_pd_free(to_mdev(pd->device)->dev, to_mpd(pd)->pdn);
609 static struct ib_xrcd *mlx4_ib_alloc_xrcd(struct ib_device *ibdev,
610 struct ib_ucontext *context,
611 struct ib_udata *udata)
613 struct mlx4_ib_xrcd *xrcd;
616 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
617 return ERR_PTR(-ENOSYS);
619 xrcd = kmalloc(sizeof *xrcd, GFP_KERNEL);
621 return ERR_PTR(-ENOMEM);
623 err = mlx4_xrcd_alloc(to_mdev(ibdev)->dev, &xrcd->xrcdn);
627 xrcd->pd = ib_alloc_pd(ibdev);
628 if (IS_ERR(xrcd->pd)) {
629 err = PTR_ERR(xrcd->pd);
633 xrcd->cq = ib_create_cq(ibdev, NULL, NULL, xrcd, 1, 0);
634 if (IS_ERR(xrcd->cq)) {
635 err = PTR_ERR(xrcd->cq);
639 return &xrcd->ibxrcd;
642 ib_dealloc_pd(xrcd->pd);
644 mlx4_xrcd_free(to_mdev(ibdev)->dev, xrcd->xrcdn);
650 static int mlx4_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
652 ib_destroy_cq(to_mxrcd(xrcd)->cq);
653 ib_dealloc_pd(to_mxrcd(xrcd)->pd);
654 mlx4_xrcd_free(to_mdev(xrcd->device)->dev, to_mxrcd(xrcd)->xrcdn);
660 static int add_gid_entry(struct ib_qp *ibqp, union ib_gid *gid)
662 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
663 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
664 struct mlx4_ib_gid_entry *ge;
666 ge = kzalloc(sizeof *ge, GFP_KERNEL);
671 if (mlx4_ib_add_mc(mdev, mqp, gid)) {
672 ge->port = mqp->port;
676 mutex_lock(&mqp->mutex);
677 list_add_tail(&ge->list, &mqp->gid_list);
678 mutex_unlock(&mqp->mutex);
683 int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
687 struct net_device *ndev;
693 spin_lock(&mdev->iboe.lock);
694 ndev = mdev->iboe.netdevs[mqp->port - 1];
697 spin_unlock(&mdev->iboe.lock);
700 rdma_get_mcast_mac((struct in6_addr *)gid, mac);
702 dev_mc_add(mdev->iboe.netdevs[mqp->port - 1], mac);
711 static int mlx4_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
714 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
715 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
717 err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw,
718 !!(mqp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
723 err = add_gid_entry(ibqp, gid);
730 mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw, MLX4_PROT_IB_IPV6);
734 static struct mlx4_ib_gid_entry *find_gid_entry(struct mlx4_ib_qp *qp, u8 *raw)
736 struct mlx4_ib_gid_entry *ge;
737 struct mlx4_ib_gid_entry *tmp;
738 struct mlx4_ib_gid_entry *ret = NULL;
740 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
741 if (!memcmp(raw, ge->gid.raw, 16)) {
750 static int mlx4_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
753 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
754 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
756 struct net_device *ndev;
757 struct mlx4_ib_gid_entry *ge;
759 err = mlx4_multicast_detach(mdev->dev,
760 &mqp->mqp, gid->raw, MLX4_PROT_IB_IPV6);
764 mutex_lock(&mqp->mutex);
765 ge = find_gid_entry(mqp, gid->raw);
767 spin_lock(&mdev->iboe.lock);
768 ndev = ge->added ? mdev->iboe.netdevs[ge->port - 1] : NULL;
771 spin_unlock(&mdev->iboe.lock);
772 rdma_get_mcast_mac((struct in6_addr *)gid, mac);
775 dev_mc_del(mdev->iboe.netdevs[ge->port - 1], mac);
782 printk(KERN_WARNING "could not find mgid entry\n");
784 mutex_unlock(&mqp->mutex);
789 static int init_node_data(struct mlx4_ib_dev *dev)
791 struct ib_smp *in_mad = NULL;
792 struct ib_smp *out_mad = NULL;
795 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
796 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
797 if (!in_mad || !out_mad)
800 init_query_mad(in_mad);
801 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
803 err = mlx4_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
807 memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
809 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
811 err = mlx4_MAD_IFC(dev, 1, 1, 1, NULL, NULL, in_mad, out_mad);
815 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
823 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
826 struct mlx4_ib_dev *dev =
827 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
828 return sprintf(buf, "MT%d\n", dev->dev->pdev->device);
831 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
834 struct mlx4_ib_dev *dev =
835 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
836 return sprintf(buf, "%d.%d.%d\n", (int) (dev->dev->caps.fw_ver >> 32),
837 (int) (dev->dev->caps.fw_ver >> 16) & 0xffff,
838 (int) dev->dev->caps.fw_ver & 0xffff);
841 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
844 struct mlx4_ib_dev *dev =
845 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
846 return sprintf(buf, "%x\n", dev->dev->rev_id);
849 static ssize_t show_board(struct device *device, struct device_attribute *attr,
852 struct mlx4_ib_dev *dev =
853 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
854 return sprintf(buf, "%.*s\n", MLX4_BOARD_ID_LEN,
858 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
859 static DEVICE_ATTR(fw_ver, S_IRUGO, show_fw_ver, NULL);
860 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
861 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
863 static struct device_attribute *mlx4_class_attributes[] = {
870 static void mlx4_addrconf_ifid_eui48(u8 *eui, u16 vlan_id, struct net_device *dev)
872 memcpy(eui, dev->dev_addr, 3);
873 memcpy(eui + 5, dev->dev_addr + 3, 3);
874 if (vlan_id < 0x1000) {
875 eui[3] = vlan_id >> 8;
876 eui[4] = vlan_id & 0xff;
884 static void update_gids_task(struct work_struct *work)
886 struct update_gid_work *gw = container_of(work, struct update_gid_work, work);
887 struct mlx4_cmd_mailbox *mailbox;
890 struct mlx4_dev *dev = gw->dev->dev;
891 struct ib_event event;
893 mailbox = mlx4_alloc_cmd_mailbox(dev);
894 if (IS_ERR(mailbox)) {
895 printk(KERN_WARNING "update gid table failed %ld\n", PTR_ERR(mailbox));
900 memcpy(gids, gw->gids, sizeof gw->gids);
902 err = mlx4_cmd(dev, mailbox->dma, MLX4_SET_PORT_GID_TABLE << 8 | gw->port,
903 1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B);
905 printk(KERN_WARNING "set port command failed\n");
907 memcpy(gw->dev->iboe.gid_table[gw->port - 1], gw->gids, sizeof gw->gids);
908 event.device = &gw->dev->ib_dev;
909 event.element.port_num = gw->port;
910 event.event = IB_EVENT_GID_CHANGE;
911 ib_dispatch_event(&event);
914 mlx4_free_cmd_mailbox(dev, mailbox);
918 static int update_ipv6_gids(struct mlx4_ib_dev *dev, int port, int clear)
920 struct net_device *ndev = dev->iboe.netdevs[port - 1];
921 struct update_gid_work *work;
922 struct net_device *tmp;
932 work = kzalloc(sizeof *work, GFP_ATOMIC);
936 hits = kzalloc(128, GFP_ATOMIC);
943 for_each_netdev_rcu(&init_net, tmp) {
944 if (ndev && (tmp == ndev || rdma_vlan_dev_real_dev(tmp) == ndev)) {
945 gid.global.subnet_prefix = cpu_to_be64(0xfe80000000000000LL);
946 vid = rdma_vlan_dev_vlan_id(tmp);
947 mlx4_addrconf_ifid_eui48(&gid.raw[8], vid, ndev);
950 for (i = 0; i < 128; ++i) {
952 !memcmp(&dev->iboe.gid_table[port - 1][i], &zgid, sizeof zgid))
954 if (!memcmp(&dev->iboe.gid_table[port - 1][i], &gid, sizeof gid)) {
963 (memcmp(&dev->iboe.gid_table[port - 1][0],
965 !memcmp(&dev->iboe.gid_table[port - 1][0],
966 &zgid, sizeof gid))) {
967 dev->iboe.gid_table[port - 1][0] = gid;
970 } else if (free >= 0) {
971 dev->iboe.gid_table[port - 1][free] = gid;
980 for (i = 0; i < 128; ++i)
982 if (memcmp(&dev->iboe.gid_table[port - 1][i], &zgid, sizeof zgid))
984 dev->iboe.gid_table[port - 1][i] = zgid;
988 memcpy(work->gids, dev->iboe.gid_table[port - 1], sizeof work->gids);
989 INIT_WORK(&work->work, update_gids_task);
992 queue_work(wq, &work->work);
1004 static void handle_en_event(struct mlx4_ib_dev *dev, int port, unsigned long event)
1008 case NETDEV_CHANGEADDR:
1009 update_ipv6_gids(dev, port, 0);
1013 update_ipv6_gids(dev, port, 1);
1014 dev->iboe.netdevs[port - 1] = NULL;
1018 static void netdev_added(struct mlx4_ib_dev *dev, int port)
1020 update_ipv6_gids(dev, port, 0);
1023 static void netdev_removed(struct mlx4_ib_dev *dev, int port)
1025 update_ipv6_gids(dev, port, 1);
1028 static int mlx4_ib_netdev_event(struct notifier_block *this, unsigned long event,
1031 struct net_device *dev = ptr;
1032 struct mlx4_ib_dev *ibdev;
1033 struct net_device *oldnd;
1034 struct mlx4_ib_iboe *iboe;
1037 if (!net_eq(dev_net(dev), &init_net))
1040 ibdev = container_of(this, struct mlx4_ib_dev, iboe.nb);
1041 iboe = &ibdev->iboe;
1043 spin_lock(&iboe->lock);
1044 mlx4_foreach_ib_transport_port(port, ibdev->dev) {
1045 oldnd = iboe->netdevs[port - 1];
1046 iboe->netdevs[port - 1] =
1047 mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port);
1048 if (oldnd != iboe->netdevs[port - 1]) {
1049 if (iboe->netdevs[port - 1])
1050 netdev_added(ibdev, port);
1052 netdev_removed(ibdev, port);
1056 if (dev == iboe->netdevs[0] ||
1057 (iboe->netdevs[0] && rdma_vlan_dev_real_dev(dev) == iboe->netdevs[0]))
1058 handle_en_event(ibdev, 1, event);
1059 else if (dev == iboe->netdevs[1]
1060 || (iboe->netdevs[1] && rdma_vlan_dev_real_dev(dev) == iboe->netdevs[1]))
1061 handle_en_event(ibdev, 2, event);
1063 spin_unlock(&iboe->lock);
1068 static void *mlx4_ib_add(struct mlx4_dev *dev)
1070 struct mlx4_ib_dev *ibdev;
1074 struct mlx4_ib_iboe *iboe;
1076 printk_once(KERN_INFO "%s", mlx4_ib_version);
1078 mlx4_foreach_ib_transport_port(i, dev)
1081 /* No point in registering a device with no ports... */
1085 ibdev = (struct mlx4_ib_dev *) ib_alloc_device(sizeof *ibdev);
1087 dev_err(&dev->pdev->dev, "Device struct alloc failed\n");
1091 iboe = &ibdev->iboe;
1093 if (mlx4_pd_alloc(dev, &ibdev->priv_pdn))
1096 if (mlx4_uar_alloc(dev, &ibdev->priv_uar))
1099 ibdev->uar_map = ioremap((phys_addr_t) ibdev->priv_uar.pfn << PAGE_SHIFT,
1101 if (!ibdev->uar_map)
1103 MLX4_INIT_DOORBELL_LOCK(&ibdev->uar_lock);
1107 strlcpy(ibdev->ib_dev.name, "mlx4_%d", IB_DEVICE_NAME_MAX);
1108 ibdev->ib_dev.owner = THIS_MODULE;
1109 ibdev->ib_dev.node_type = RDMA_NODE_IB_CA;
1110 ibdev->ib_dev.local_dma_lkey = dev->caps.reserved_lkey;
1111 ibdev->num_ports = num_ports;
1112 ibdev->ib_dev.phys_port_cnt = ibdev->num_ports;
1113 ibdev->ib_dev.num_comp_vectors = dev->caps.num_comp_vectors;
1114 ibdev->ib_dev.dma_device = &dev->pdev->dev;
1116 ibdev->ib_dev.uverbs_abi_ver = MLX4_IB_UVERBS_ABI_VERSION;
1117 ibdev->ib_dev.uverbs_cmd_mask =
1118 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
1119 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
1120 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
1121 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
1122 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
1123 (1ull << IB_USER_VERBS_CMD_REG_MR) |
1124 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
1125 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1126 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
1127 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
1128 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
1129 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
1130 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
1131 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
1132 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
1133 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
1134 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
1135 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
1136 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
1137 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
1138 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
1139 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
1140 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
1142 ibdev->ib_dev.query_device = mlx4_ib_query_device;
1143 ibdev->ib_dev.query_port = mlx4_ib_query_port;
1144 ibdev->ib_dev.get_link_layer = mlx4_ib_port_link_layer;
1145 ibdev->ib_dev.query_gid = mlx4_ib_query_gid;
1146 ibdev->ib_dev.query_pkey = mlx4_ib_query_pkey;
1147 ibdev->ib_dev.modify_device = mlx4_ib_modify_device;
1148 ibdev->ib_dev.modify_port = mlx4_ib_modify_port;
1149 ibdev->ib_dev.alloc_ucontext = mlx4_ib_alloc_ucontext;
1150 ibdev->ib_dev.dealloc_ucontext = mlx4_ib_dealloc_ucontext;
1151 ibdev->ib_dev.mmap = mlx4_ib_mmap;
1152 ibdev->ib_dev.alloc_pd = mlx4_ib_alloc_pd;
1153 ibdev->ib_dev.dealloc_pd = mlx4_ib_dealloc_pd;
1154 ibdev->ib_dev.create_ah = mlx4_ib_create_ah;
1155 ibdev->ib_dev.query_ah = mlx4_ib_query_ah;
1156 ibdev->ib_dev.destroy_ah = mlx4_ib_destroy_ah;
1157 ibdev->ib_dev.create_srq = mlx4_ib_create_srq;
1158 ibdev->ib_dev.modify_srq = mlx4_ib_modify_srq;
1159 ibdev->ib_dev.query_srq = mlx4_ib_query_srq;
1160 ibdev->ib_dev.destroy_srq = mlx4_ib_destroy_srq;
1161 ibdev->ib_dev.post_srq_recv = mlx4_ib_post_srq_recv;
1162 ibdev->ib_dev.create_qp = mlx4_ib_create_qp;
1163 ibdev->ib_dev.modify_qp = mlx4_ib_modify_qp;
1164 ibdev->ib_dev.query_qp = mlx4_ib_query_qp;
1165 ibdev->ib_dev.destroy_qp = mlx4_ib_destroy_qp;
1166 ibdev->ib_dev.post_send = mlx4_ib_post_send;
1167 ibdev->ib_dev.post_recv = mlx4_ib_post_recv;
1168 ibdev->ib_dev.create_cq = mlx4_ib_create_cq;
1169 ibdev->ib_dev.modify_cq = mlx4_ib_modify_cq;
1170 ibdev->ib_dev.resize_cq = mlx4_ib_resize_cq;
1171 ibdev->ib_dev.destroy_cq = mlx4_ib_destroy_cq;
1172 ibdev->ib_dev.poll_cq = mlx4_ib_poll_cq;
1173 ibdev->ib_dev.req_notify_cq = mlx4_ib_arm_cq;
1174 ibdev->ib_dev.get_dma_mr = mlx4_ib_get_dma_mr;
1175 ibdev->ib_dev.reg_user_mr = mlx4_ib_reg_user_mr;
1176 ibdev->ib_dev.dereg_mr = mlx4_ib_dereg_mr;
1177 ibdev->ib_dev.alloc_fast_reg_mr = mlx4_ib_alloc_fast_reg_mr;
1178 ibdev->ib_dev.alloc_fast_reg_page_list = mlx4_ib_alloc_fast_reg_page_list;
1179 ibdev->ib_dev.free_fast_reg_page_list = mlx4_ib_free_fast_reg_page_list;
1180 ibdev->ib_dev.attach_mcast = mlx4_ib_mcg_attach;
1181 ibdev->ib_dev.detach_mcast = mlx4_ib_mcg_detach;
1182 ibdev->ib_dev.process_mad = mlx4_ib_process_mad;
1184 ibdev->ib_dev.alloc_fmr = mlx4_ib_fmr_alloc;
1185 ibdev->ib_dev.map_phys_fmr = mlx4_ib_map_phys_fmr;
1186 ibdev->ib_dev.unmap_fmr = mlx4_ib_unmap_fmr;
1187 ibdev->ib_dev.dealloc_fmr = mlx4_ib_fmr_dealloc;
1189 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) {
1190 ibdev->ib_dev.alloc_xrcd = mlx4_ib_alloc_xrcd;
1191 ibdev->ib_dev.dealloc_xrcd = mlx4_ib_dealloc_xrcd;
1192 ibdev->ib_dev.uverbs_cmd_mask |=
1193 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
1194 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
1197 spin_lock_init(&iboe->lock);
1199 if (init_node_data(ibdev))
1202 for (i = 0; i < ibdev->num_ports; ++i) {
1203 if (mlx4_ib_port_link_layer(&ibdev->ib_dev, i + 1) ==
1204 IB_LINK_LAYER_ETHERNET) {
1205 err = mlx4_counter_alloc(ibdev->dev, &ibdev->counters[i]);
1207 ibdev->counters[i] = -1;
1209 ibdev->counters[i] = -1;
1212 spin_lock_init(&ibdev->sm_lock);
1213 mutex_init(&ibdev->cap_mask_mutex);
1215 if (ib_register_device(&ibdev->ib_dev, NULL))
1218 if (mlx4_ib_mad_init(ibdev))
1221 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE && !iboe->nb.notifier_call) {
1222 iboe->nb.notifier_call = mlx4_ib_netdev_event;
1223 err = register_netdevice_notifier(&iboe->nb);
1228 for (i = 0; i < ARRAY_SIZE(mlx4_class_attributes); ++i) {
1229 if (device_create_file(&ibdev->ib_dev.dev,
1230 mlx4_class_attributes[i]))
1234 ibdev->ib_active = true;
1239 if (unregister_netdevice_notifier(&ibdev->iboe.nb))
1240 printk(KERN_WARNING "failure unregistering notifier\n");
1241 flush_workqueue(wq);
1244 ib_unregister_device(&ibdev->ib_dev);
1248 if (ibdev->counters[i - 1] != -1)
1249 mlx4_counter_free(ibdev->dev, ibdev->counters[i - 1]);
1252 iounmap(ibdev->uar_map);
1255 mlx4_uar_free(dev, &ibdev->priv_uar);
1258 mlx4_pd_free(dev, ibdev->priv_pdn);
1261 ib_dealloc_device(&ibdev->ib_dev);
1266 static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
1268 struct mlx4_ib_dev *ibdev = ibdev_ptr;
1271 mlx4_ib_mad_cleanup(ibdev);
1272 ib_unregister_device(&ibdev->ib_dev);
1273 if (ibdev->iboe.nb.notifier_call) {
1274 if (unregister_netdevice_notifier(&ibdev->iboe.nb))
1275 printk(KERN_WARNING "failure unregistering notifier\n");
1276 ibdev->iboe.nb.notifier_call = NULL;
1278 iounmap(ibdev->uar_map);
1279 for (p = 0; p < ibdev->num_ports; ++p)
1280 if (ibdev->counters[p] != -1)
1281 mlx4_counter_free(ibdev->dev, ibdev->counters[p]);
1282 mlx4_foreach_port(p, dev, MLX4_PORT_TYPE_IB)
1283 mlx4_CLOSE_PORT(dev, p);
1285 mlx4_uar_free(dev, &ibdev->priv_uar);
1286 mlx4_pd_free(dev, ibdev->priv_pdn);
1287 ib_dealloc_device(&ibdev->ib_dev);
1290 static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
1291 enum mlx4_dev_event event, int port)
1293 struct ib_event ibev;
1294 struct mlx4_ib_dev *ibdev = to_mdev((struct ib_device *) ibdev_ptr);
1296 if (port > ibdev->num_ports)
1300 case MLX4_DEV_EVENT_PORT_UP:
1301 ibev.event = IB_EVENT_PORT_ACTIVE;
1304 case MLX4_DEV_EVENT_PORT_DOWN:
1305 ibev.event = IB_EVENT_PORT_ERR;
1308 case MLX4_DEV_EVENT_CATASTROPHIC_ERROR:
1309 ibdev->ib_active = false;
1310 ibev.event = IB_EVENT_DEVICE_FATAL;
1317 ibev.device = ibdev_ptr;
1318 ibev.element.port_num = port;
1320 ib_dispatch_event(&ibev);
1323 static struct mlx4_interface mlx4_ib_interface = {
1325 .remove = mlx4_ib_remove,
1326 .event = mlx4_ib_event,
1327 .protocol = MLX4_PROT_IB_IPV6
1330 static int __init mlx4_ib_init(void)
1334 wq = create_singlethread_workqueue("mlx4_ib");
1338 err = mlx4_register_interface(&mlx4_ib_interface);
1340 destroy_workqueue(wq);
1347 static void __exit mlx4_ib_cleanup(void)
1349 mlx4_unregister_interface(&mlx4_ib_interface);
1350 destroy_workqueue(wq);
1353 module_init(mlx4_ib_init);
1354 module_exit(mlx4_ib_cleanup);