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>
42 #include <linux/sched/mm.h>
43 #include <linux/sched/task.h>
46 #include <net/addrconf.h>
47 #include <net/devlink.h>
49 #include <rdma/ib_smi.h>
50 #include <rdma/ib_user_verbs.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/ib_cache.h>
54 #include <net/bonding.h>
56 #include <linux/mlx4/driver.h>
57 #include <linux/mlx4/cmd.h>
58 #include <linux/mlx4/qp.h>
61 #include <rdma/mlx4-abi.h>
63 #define DRV_NAME MLX4_IB_DRV_NAME
64 #define DRV_VERSION "4.0-0"
66 #define MLX4_IB_FLOW_MAX_PRIO 0xFFF
67 #define MLX4_IB_FLOW_QPN_MASK 0xFFFFFF
68 #define MLX4_IB_CARD_REV_A0 0xA0
70 MODULE_AUTHOR("Roland Dreier");
71 MODULE_DESCRIPTION("Mellanox ConnectX HCA InfiniBand driver");
72 MODULE_LICENSE("Dual BSD/GPL");
73 MODULE_VERSION(DRV_VERSION);
75 int mlx4_ib_sm_guid_assign = 0;
76 module_param_named(sm_guid_assign, mlx4_ib_sm_guid_assign, int, 0444);
77 MODULE_PARM_DESC(sm_guid_assign, "Enable SM alias_GUID assignment if sm_guid_assign > 0 (Default: 0)");
79 static const char mlx4_ib_version[] =
80 DRV_NAME ": Mellanox ConnectX InfiniBand driver v"
83 static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init);
85 static struct workqueue_struct *wq;
87 static void init_query_mad(struct ib_smp *mad)
89 mad->base_version = 1;
90 mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
91 mad->class_version = 1;
92 mad->method = IB_MGMT_METHOD_GET;
95 static int check_flow_steering_support(struct mlx4_dev *dev)
97 int eth_num_ports = 0;
100 int dmfs = dev->caps.steering_mode == MLX4_STEERING_MODE_DEVICE_MANAGED;
104 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH)
106 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
108 dmfs &= (!ib_num_ports ||
109 (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_IPOIB)) &&
111 (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_FS_EN));
112 if (ib_num_ports && mlx4_is_mfunc(dev)) {
113 pr_warn("Device managed flow steering is unavailable for IB port in multifunction env.\n");
120 static int num_ib_ports(struct mlx4_dev *dev)
125 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
131 static struct net_device *mlx4_ib_get_netdev(struct ib_device *device, u8 port_num)
133 struct mlx4_ib_dev *ibdev = to_mdev(device);
134 struct net_device *dev;
137 dev = mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port_num);
140 if (mlx4_is_bonded(ibdev->dev)) {
141 struct net_device *upper = NULL;
143 upper = netdev_master_upper_dev_get_rcu(dev);
145 struct net_device *active;
147 active = bond_option_active_slave_get_rcu(netdev_priv(upper));
160 static int mlx4_ib_update_gids_v1(struct gid_entry *gids,
161 struct mlx4_ib_dev *ibdev,
164 struct mlx4_cmd_mailbox *mailbox;
166 struct mlx4_dev *dev = ibdev->dev;
168 union ib_gid *gid_tbl;
170 mailbox = mlx4_alloc_cmd_mailbox(dev);
174 gid_tbl = mailbox->buf;
176 for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i)
177 memcpy(&gid_tbl[i], &gids[i].gid, sizeof(union ib_gid));
179 err = mlx4_cmd(dev, mailbox->dma,
180 MLX4_SET_PORT_GID_TABLE << 8 | port_num,
181 1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
183 if (mlx4_is_bonded(dev))
184 err += mlx4_cmd(dev, mailbox->dma,
185 MLX4_SET_PORT_GID_TABLE << 8 | 2,
186 1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
189 mlx4_free_cmd_mailbox(dev, mailbox);
193 static int mlx4_ib_update_gids_v1_v2(struct gid_entry *gids,
194 struct mlx4_ib_dev *ibdev,
197 struct mlx4_cmd_mailbox *mailbox;
199 struct mlx4_dev *dev = ibdev->dev;
210 mailbox = mlx4_alloc_cmd_mailbox(dev);
214 gid_tbl = mailbox->buf;
215 for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i) {
216 memcpy(&gid_tbl[i].gid, &gids[i].gid, sizeof(union ib_gid));
217 if (gids[i].gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
218 gid_tbl[i].version = 2;
219 if (!ipv6_addr_v4mapped((struct in6_addr *)&gids[i].gid))
222 memset(&gid_tbl[i].gid, 0, 12);
226 err = mlx4_cmd(dev, mailbox->dma,
227 MLX4_SET_PORT_ROCE_ADDR << 8 | port_num,
228 1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
230 if (mlx4_is_bonded(dev))
231 err += mlx4_cmd(dev, mailbox->dma,
232 MLX4_SET_PORT_ROCE_ADDR << 8 | 2,
233 1, MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
236 mlx4_free_cmd_mailbox(dev, mailbox);
240 static int mlx4_ib_update_gids(struct gid_entry *gids,
241 struct mlx4_ib_dev *ibdev,
244 if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2)
245 return mlx4_ib_update_gids_v1_v2(gids, ibdev, port_num);
247 return mlx4_ib_update_gids_v1(gids, ibdev, port_num);
250 static int mlx4_ib_add_gid(struct ib_device *device,
253 const union ib_gid *gid,
254 const struct ib_gid_attr *attr,
257 struct mlx4_ib_dev *ibdev = to_mdev(device);
258 struct mlx4_ib_iboe *iboe = &ibdev->iboe;
259 struct mlx4_port_gid_table *port_gid_table;
260 int free = -1, found = -1;
264 struct gid_entry *gids = NULL;
266 if (!rdma_cap_roce_gid_table(device, port_num))
269 if (port_num > MLX4_MAX_PORTS)
275 port_gid_table = &iboe->gids[port_num - 1];
276 spin_lock_bh(&iboe->lock);
277 for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i) {
278 if (!memcmp(&port_gid_table->gids[i].gid, gid, sizeof(*gid)) &&
279 (port_gid_table->gids[i].gid_type == attr->gid_type)) {
283 if (free < 0 && !memcmp(&port_gid_table->gids[i].gid, &zgid, sizeof(*gid)))
284 free = i; /* HW has space */
291 port_gid_table->gids[free].ctx = kmalloc(sizeof(*port_gid_table->gids[free].ctx), GFP_ATOMIC);
292 if (!port_gid_table->gids[free].ctx) {
295 *context = port_gid_table->gids[free].ctx;
296 memcpy(&port_gid_table->gids[free].gid, gid, sizeof(*gid));
297 port_gid_table->gids[free].gid_type = attr->gid_type;
298 port_gid_table->gids[free].ctx->real_index = free;
299 port_gid_table->gids[free].ctx->refcount = 1;
304 struct gid_cache_context *ctx = port_gid_table->gids[found].ctx;
308 if (!ret && hw_update) {
309 gids = kmalloc(sizeof(*gids) * MLX4_MAX_PORT_GIDS, GFP_ATOMIC);
313 for (i = 0; i < MLX4_MAX_PORT_GIDS; i++) {
314 memcpy(&gids[i].gid, &port_gid_table->gids[i].gid, sizeof(union ib_gid));
315 gids[i].gid_type = port_gid_table->gids[i].gid_type;
319 spin_unlock_bh(&iboe->lock);
321 if (!ret && hw_update) {
322 ret = mlx4_ib_update_gids(gids, ibdev, port_num);
329 static int mlx4_ib_del_gid(struct ib_device *device,
334 struct gid_cache_context *ctx = *context;
335 struct mlx4_ib_dev *ibdev = to_mdev(device);
336 struct mlx4_ib_iboe *iboe = &ibdev->iboe;
337 struct mlx4_port_gid_table *port_gid_table;
340 struct gid_entry *gids = NULL;
342 if (!rdma_cap_roce_gid_table(device, port_num))
345 if (port_num > MLX4_MAX_PORTS)
348 port_gid_table = &iboe->gids[port_num - 1];
349 spin_lock_bh(&iboe->lock);
352 if (!ctx->refcount) {
353 unsigned int real_index = ctx->real_index;
355 memcpy(&port_gid_table->gids[real_index].gid, &zgid, sizeof(zgid));
356 kfree(port_gid_table->gids[real_index].ctx);
357 port_gid_table->gids[real_index].ctx = NULL;
361 if (!ret && hw_update) {
364 gids = kmalloc(sizeof(*gids) * MLX4_MAX_PORT_GIDS, GFP_ATOMIC);
368 for (i = 0; i < MLX4_MAX_PORT_GIDS; i++)
369 memcpy(&gids[i].gid, &port_gid_table->gids[i].gid, sizeof(union ib_gid));
372 spin_unlock_bh(&iboe->lock);
374 if (!ret && hw_update) {
375 ret = mlx4_ib_update_gids(gids, ibdev, port_num);
381 int mlx4_ib_gid_index_to_real_index(struct mlx4_ib_dev *ibdev,
382 u8 port_num, int index)
384 struct mlx4_ib_iboe *iboe = &ibdev->iboe;
385 struct gid_cache_context *ctx = NULL;
387 struct mlx4_port_gid_table *port_gid_table;
388 int real_index = -EINVAL;
392 struct ib_gid_attr attr;
394 if (port_num > MLX4_MAX_PORTS)
397 if (mlx4_is_bonded(ibdev->dev))
400 if (!rdma_cap_roce_gid_table(&ibdev->ib_dev, port_num))
403 ret = ib_get_cached_gid(&ibdev->ib_dev, port_num, index, &gid, &attr);
410 if (!memcmp(&gid, &zgid, sizeof(gid)))
413 spin_lock_irqsave(&iboe->lock, flags);
414 port_gid_table = &iboe->gids[port_num - 1];
416 for (i = 0; i < MLX4_MAX_PORT_GIDS; ++i)
417 if (!memcmp(&port_gid_table->gids[i].gid, &gid, sizeof(gid)) &&
418 attr.gid_type == port_gid_table->gids[i].gid_type) {
419 ctx = port_gid_table->gids[i].ctx;
423 real_index = ctx->real_index;
424 spin_unlock_irqrestore(&iboe->lock, flags);
428 static int mlx4_ib_query_device(struct ib_device *ibdev,
429 struct ib_device_attr *props,
430 struct ib_udata *uhw)
432 struct mlx4_ib_dev *dev = to_mdev(ibdev);
433 struct ib_smp *in_mad = NULL;
434 struct ib_smp *out_mad = NULL;
437 struct mlx4_uverbs_ex_query_device cmd;
438 struct mlx4_uverbs_ex_query_device_resp resp = {.comp_mask = 0};
439 struct mlx4_clock_params clock_params;
442 if (uhw->inlen < sizeof(cmd))
445 err = ib_copy_from_udata(&cmd, uhw, sizeof(cmd));
456 resp.response_length = offsetof(typeof(resp), response_length) +
457 sizeof(resp.response_length);
458 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
459 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
461 if (!in_mad || !out_mad)
464 init_query_mad(in_mad);
465 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
467 err = mlx4_MAD_IFC(to_mdev(ibdev), MLX4_MAD_IFC_IGNORE_KEYS,
468 1, NULL, NULL, in_mad, out_mad);
472 memset(props, 0, sizeof *props);
474 have_ib_ports = num_ib_ports(dev->dev);
476 props->fw_ver = dev->dev->caps.fw_ver;
477 props->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
478 IB_DEVICE_PORT_ACTIVE_EVENT |
479 IB_DEVICE_SYS_IMAGE_GUID |
480 IB_DEVICE_RC_RNR_NAK_GEN |
481 IB_DEVICE_BLOCK_MULTICAST_LOOPBACK;
482 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_PKEY_CNTR)
483 props->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
484 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BAD_QKEY_CNTR)
485 props->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
486 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_APM && have_ib_ports)
487 props->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
488 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_UD_AV_PORT)
489 props->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
490 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IPOIB_CSUM)
491 props->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
492 if (dev->dev->caps.max_gso_sz &&
493 (dev->dev->rev_id != MLX4_IB_CARD_REV_A0) &&
494 (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_BLH))
495 props->device_cap_flags |= IB_DEVICE_UD_TSO;
496 if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_RESERVED_LKEY)
497 props->device_cap_flags |= IB_DEVICE_LOCAL_DMA_LKEY;
498 if ((dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_LOCAL_INV) &&
499 (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_REMOTE_INV) &&
500 (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_FAST_REG_WR))
501 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
502 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC)
503 props->device_cap_flags |= IB_DEVICE_XRC;
504 if (dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW)
505 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW;
506 if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN) {
507 if (dev->dev->caps.bmme_flags & MLX4_BMME_FLAG_WIN_TYPE_2B)
508 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2B;
510 props->device_cap_flags |= IB_DEVICE_MEM_WINDOW_TYPE_2A;
512 if (dev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED)
513 props->device_cap_flags |= IB_DEVICE_MANAGED_FLOW_STEERING;
515 props->device_cap_flags |= IB_DEVICE_RAW_IP_CSUM;
517 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
519 props->vendor_part_id = dev->dev->persist->pdev->device;
520 props->hw_ver = be32_to_cpup((__be32 *) (out_mad->data + 32));
521 memcpy(&props->sys_image_guid, out_mad->data + 4, 8);
523 props->max_mr_size = ~0ull;
524 props->page_size_cap = dev->dev->caps.page_size_cap;
525 props->max_qp = dev->dev->quotas.qp;
526 props->max_qp_wr = dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE;
527 props->max_sge = min(dev->dev->caps.max_sq_sg,
528 dev->dev->caps.max_rq_sg);
529 props->max_sge_rd = MLX4_MAX_SGE_RD;
530 props->max_cq = dev->dev->quotas.cq;
531 props->max_cqe = dev->dev->caps.max_cqes;
532 props->max_mr = dev->dev->quotas.mpt;
533 props->max_pd = dev->dev->caps.num_pds - dev->dev->caps.reserved_pds;
534 props->max_qp_rd_atom = dev->dev->caps.max_qp_dest_rdma;
535 props->max_qp_init_rd_atom = dev->dev->caps.max_qp_init_rdma;
536 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp;
537 props->max_srq = dev->dev->quotas.srq;
538 props->max_srq_wr = dev->dev->caps.max_srq_wqes - 1;
539 props->max_srq_sge = dev->dev->caps.max_srq_sge;
540 props->max_fast_reg_page_list_len = MLX4_MAX_FAST_REG_PAGES;
541 props->local_ca_ack_delay = dev->dev->caps.local_ca_ack_delay;
542 props->atomic_cap = dev->dev->caps.flags & MLX4_DEV_CAP_FLAG_ATOMIC ?
543 IB_ATOMIC_HCA : IB_ATOMIC_NONE;
544 props->masked_atomic_cap = props->atomic_cap;
545 props->max_pkeys = dev->dev->caps.pkey_table_len[1];
546 props->max_mcast_grp = dev->dev->caps.num_mgms + dev->dev->caps.num_amgms;
547 props->max_mcast_qp_attach = dev->dev->caps.num_qp_per_mgm;
548 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
549 props->max_mcast_grp;
550 props->max_map_per_fmr = dev->dev->caps.max_fmr_maps;
551 props->hca_core_clock = dev->dev->caps.hca_core_clock * 1000UL;
552 props->timestamp_mask = 0xFFFFFFFFFFFFULL;
553 props->max_ah = INT_MAX;
555 if (!mlx4_is_slave(dev->dev))
556 err = mlx4_get_internal_clock_params(dev->dev, &clock_params);
558 if (uhw->outlen >= resp.response_length + sizeof(resp.hca_core_clock_offset)) {
559 resp.response_length += sizeof(resp.hca_core_clock_offset);
560 if (!err && !mlx4_is_slave(dev->dev)) {
561 resp.comp_mask |= QUERY_DEVICE_RESP_MASK_TIMESTAMP;
562 resp.hca_core_clock_offset = clock_params.offset % PAGE_SIZE;
567 err = ib_copy_to_udata(uhw, &resp, resp.response_length);
578 static enum rdma_link_layer
579 mlx4_ib_port_link_layer(struct ib_device *device, u8 port_num)
581 struct mlx4_dev *dev = to_mdev(device)->dev;
583 return dev->caps.port_mask[port_num] == MLX4_PORT_TYPE_IB ?
584 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
587 static int ib_link_query_port(struct ib_device *ibdev, u8 port,
588 struct ib_port_attr *props, int netw_view)
590 struct ib_smp *in_mad = NULL;
591 struct ib_smp *out_mad = NULL;
592 int ext_active_speed;
593 int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
596 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
597 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
598 if (!in_mad || !out_mad)
601 init_query_mad(in_mad);
602 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
603 in_mad->attr_mod = cpu_to_be32(port);
605 if (mlx4_is_mfunc(to_mdev(ibdev)->dev) && netw_view)
606 mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
608 err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
614 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16));
615 props->lmc = out_mad->data[34] & 0x7;
616 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18));
617 props->sm_sl = out_mad->data[36] & 0xf;
618 props->state = out_mad->data[32] & 0xf;
619 props->phys_state = out_mad->data[33] >> 4;
620 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20));
622 props->gid_tbl_len = out_mad->data[50];
624 props->gid_tbl_len = to_mdev(ibdev)->dev->caps.gid_table_len[port];
625 props->max_msg_sz = to_mdev(ibdev)->dev->caps.max_msg_sz;
626 props->pkey_tbl_len = to_mdev(ibdev)->dev->caps.pkey_table_len[port];
627 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46));
628 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48));
629 props->active_width = out_mad->data[31] & 0xf;
630 props->active_speed = out_mad->data[35] >> 4;
631 props->max_mtu = out_mad->data[41] & 0xf;
632 props->active_mtu = out_mad->data[36] >> 4;
633 props->subnet_timeout = out_mad->data[51] & 0x1f;
634 props->max_vl_num = out_mad->data[37] >> 4;
635 props->init_type_reply = out_mad->data[41] >> 4;
637 /* Check if extended speeds (EDR/FDR/...) are supported */
638 if (props->port_cap_flags & IB_PORT_EXTENDED_SPEEDS_SUP) {
639 ext_active_speed = out_mad->data[62] >> 4;
641 switch (ext_active_speed) {
643 props->active_speed = IB_SPEED_FDR;
646 props->active_speed = IB_SPEED_EDR;
651 /* If reported active speed is QDR, check if is FDR-10 */
652 if (props->active_speed == IB_SPEED_QDR) {
653 init_query_mad(in_mad);
654 in_mad->attr_id = MLX4_ATTR_EXTENDED_PORT_INFO;
655 in_mad->attr_mod = cpu_to_be32(port);
657 err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port,
658 NULL, NULL, in_mad, out_mad);
662 /* Checking LinkSpeedActive for FDR-10 */
663 if (out_mad->data[15] & 0x1)
664 props->active_speed = IB_SPEED_FDR10;
667 /* Avoid wrong speed value returned by FW if the IB link is down. */
668 if (props->state == IB_PORT_DOWN)
669 props->active_speed = IB_SPEED_SDR;
677 static u8 state_to_phys_state(enum ib_port_state state)
679 return state == IB_PORT_ACTIVE ? 5 : 3;
682 static int eth_link_query_port(struct ib_device *ibdev, u8 port,
683 struct ib_port_attr *props)
686 struct mlx4_ib_dev *mdev = to_mdev(ibdev);
687 struct mlx4_ib_iboe *iboe = &mdev->iboe;
688 struct net_device *ndev;
690 struct mlx4_cmd_mailbox *mailbox;
692 int is_bonded = mlx4_is_bonded(mdev->dev);
694 mailbox = mlx4_alloc_cmd_mailbox(mdev->dev);
696 return PTR_ERR(mailbox);
698 err = mlx4_cmd_box(mdev->dev, 0, mailbox->dma, port, 0,
699 MLX4_CMD_QUERY_PORT, MLX4_CMD_TIME_CLASS_B,
704 props->active_width = (((u8 *)mailbox->buf)[5] == 0x40) ||
705 (((u8 *)mailbox->buf)[5] == 0x20 /*56Gb*/) ?
706 IB_WIDTH_4X : IB_WIDTH_1X;
707 props->active_speed = (((u8 *)mailbox->buf)[5] == 0x20 /*56Gb*/) ?
708 IB_SPEED_FDR : IB_SPEED_QDR;
709 props->port_cap_flags = IB_PORT_CM_SUP | IB_PORT_IP_BASED_GIDS;
710 props->gid_tbl_len = mdev->dev->caps.gid_table_len[port];
711 props->max_msg_sz = mdev->dev->caps.max_msg_sz;
712 props->pkey_tbl_len = 1;
713 props->max_mtu = IB_MTU_4096;
714 props->max_vl_num = 2;
715 props->state = IB_PORT_DOWN;
716 props->phys_state = state_to_phys_state(props->state);
717 props->active_mtu = IB_MTU_256;
718 spin_lock_bh(&iboe->lock);
719 ndev = iboe->netdevs[port - 1];
720 if (ndev && is_bonded) {
721 rcu_read_lock(); /* required to get upper dev */
722 ndev = netdev_master_upper_dev_get_rcu(ndev);
728 tmp = iboe_get_mtu(ndev->mtu);
729 props->active_mtu = tmp ? min(props->max_mtu, tmp) : IB_MTU_256;
731 props->state = (netif_running(ndev) && netif_carrier_ok(ndev)) ?
732 IB_PORT_ACTIVE : IB_PORT_DOWN;
733 props->phys_state = state_to_phys_state(props->state);
735 spin_unlock_bh(&iboe->lock);
737 mlx4_free_cmd_mailbox(mdev->dev, mailbox);
741 int __mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
742 struct ib_port_attr *props, int netw_view)
746 /* props being zeroed by the caller, avoid zeroing it here */
748 err = mlx4_ib_port_link_layer(ibdev, port) == IB_LINK_LAYER_INFINIBAND ?
749 ib_link_query_port(ibdev, port, props, netw_view) :
750 eth_link_query_port(ibdev, port, props);
755 static int mlx4_ib_query_port(struct ib_device *ibdev, u8 port,
756 struct ib_port_attr *props)
758 /* returns host view */
759 return __mlx4_ib_query_port(ibdev, port, props, 0);
762 int __mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
763 union ib_gid *gid, int netw_view)
765 struct ib_smp *in_mad = NULL;
766 struct ib_smp *out_mad = NULL;
768 struct mlx4_ib_dev *dev = to_mdev(ibdev);
770 int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
772 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
773 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
774 if (!in_mad || !out_mad)
777 init_query_mad(in_mad);
778 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO;
779 in_mad->attr_mod = cpu_to_be32(port);
781 if (mlx4_is_mfunc(dev->dev) && netw_view)
782 mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
784 err = mlx4_MAD_IFC(dev, mad_ifc_flags, port, NULL, NULL, in_mad, out_mad);
788 memcpy(gid->raw, out_mad->data + 8, 8);
790 if (mlx4_is_mfunc(dev->dev) && !netw_view) {
792 /* For any index > 0, return the null guid */
799 init_query_mad(in_mad);
800 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO;
801 in_mad->attr_mod = cpu_to_be32(index / 8);
803 err = mlx4_MAD_IFC(dev, mad_ifc_flags, port,
804 NULL, NULL, in_mad, out_mad);
808 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
812 memset(gid->raw + 8, 0, 8);
818 static int mlx4_ib_query_gid(struct ib_device *ibdev, u8 port, int index,
823 if (rdma_protocol_ib(ibdev, port))
824 return __mlx4_ib_query_gid(ibdev, port, index, gid, 0);
826 if (!rdma_protocol_roce(ibdev, port))
829 if (!rdma_cap_roce_gid_table(ibdev, port))
832 ret = ib_get_cached_gid(ibdev, port, index, gid, NULL);
833 if (ret == -EAGAIN) {
834 memcpy(gid, &zgid, sizeof(*gid));
841 static int mlx4_ib_query_sl2vl(struct ib_device *ibdev, u8 port, u64 *sl2vl_tbl)
843 union sl2vl_tbl_to_u64 sl2vl64;
844 struct ib_smp *in_mad = NULL;
845 struct ib_smp *out_mad = NULL;
846 int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
850 if (mlx4_is_slave(to_mdev(ibdev)->dev)) {
855 in_mad = kzalloc(sizeof(*in_mad), GFP_KERNEL);
856 out_mad = kmalloc(sizeof(*out_mad), GFP_KERNEL);
857 if (!in_mad || !out_mad)
860 init_query_mad(in_mad);
861 in_mad->attr_id = IB_SMP_ATTR_SL_TO_VL_TABLE;
862 in_mad->attr_mod = 0;
864 if (mlx4_is_mfunc(to_mdev(ibdev)->dev))
865 mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
867 err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
872 for (jj = 0; jj < 8; jj++)
873 sl2vl64.sl8[jj] = ((struct ib_smp *)out_mad)->data[jj];
874 *sl2vl_tbl = sl2vl64.sl64;
882 static void mlx4_init_sl2vl_tbl(struct mlx4_ib_dev *mdev)
888 for (i = 1; i <= mdev->dev->caps.num_ports; i++) {
889 if (mdev->dev->caps.port_type[i] == MLX4_PORT_TYPE_ETH)
891 err = mlx4_ib_query_sl2vl(&mdev->ib_dev, i, &sl2vl);
893 pr_err("Unable to get default sl to vl mapping for port %d. Using all zeroes (%d)\n",
897 atomic64_set(&mdev->sl2vl[i - 1], sl2vl);
901 int __mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
902 u16 *pkey, int netw_view)
904 struct ib_smp *in_mad = NULL;
905 struct ib_smp *out_mad = NULL;
906 int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
909 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
910 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
911 if (!in_mad || !out_mad)
914 init_query_mad(in_mad);
915 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE;
916 in_mad->attr_mod = cpu_to_be32(index / 32);
918 if (mlx4_is_mfunc(to_mdev(ibdev)->dev) && netw_view)
919 mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
921 err = mlx4_MAD_IFC(to_mdev(ibdev), mad_ifc_flags, port, NULL, NULL,
926 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
934 static int mlx4_ib_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
936 return __mlx4_ib_query_pkey(ibdev, port, index, pkey, 0);
939 static int mlx4_ib_modify_device(struct ib_device *ibdev, int mask,
940 struct ib_device_modify *props)
942 struct mlx4_cmd_mailbox *mailbox;
945 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
948 if (!(mask & IB_DEVICE_MODIFY_NODE_DESC))
951 if (mlx4_is_slave(to_mdev(ibdev)->dev))
954 spin_lock_irqsave(&to_mdev(ibdev)->sm_lock, flags);
955 memcpy(ibdev->node_desc, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
956 spin_unlock_irqrestore(&to_mdev(ibdev)->sm_lock, flags);
959 * If possible, pass node desc to FW, so it can generate
960 * a 144 trap. If cmd fails, just ignore.
962 mailbox = mlx4_alloc_cmd_mailbox(to_mdev(ibdev)->dev);
966 memcpy(mailbox->buf, props->node_desc, IB_DEVICE_NODE_DESC_MAX);
967 mlx4_cmd(to_mdev(ibdev)->dev, mailbox->dma, 1, 0,
968 MLX4_CMD_SET_NODE, MLX4_CMD_TIME_CLASS_A, MLX4_CMD_NATIVE);
970 mlx4_free_cmd_mailbox(to_mdev(ibdev)->dev, mailbox);
975 static int mlx4_ib_SET_PORT(struct mlx4_ib_dev *dev, u8 port, int reset_qkey_viols,
978 struct mlx4_cmd_mailbox *mailbox;
981 mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
983 return PTR_ERR(mailbox);
985 if (dev->dev->flags & MLX4_FLAG_OLD_PORT_CMDS) {
986 *(u8 *) mailbox->buf = !!reset_qkey_viols << 6;
987 ((__be32 *) mailbox->buf)[2] = cpu_to_be32(cap_mask);
989 ((u8 *) mailbox->buf)[3] = !!reset_qkey_viols;
990 ((__be32 *) mailbox->buf)[1] = cpu_to_be32(cap_mask);
993 err = mlx4_cmd(dev->dev, mailbox->dma, port, MLX4_SET_PORT_IB_OPCODE,
994 MLX4_CMD_SET_PORT, MLX4_CMD_TIME_CLASS_B,
997 mlx4_free_cmd_mailbox(dev->dev, mailbox);
1001 static int mlx4_ib_modify_port(struct ib_device *ibdev, u8 port, int mask,
1002 struct ib_port_modify *props)
1004 struct mlx4_ib_dev *mdev = to_mdev(ibdev);
1005 u8 is_eth = mdev->dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH;
1006 struct ib_port_attr attr;
1010 /* return OK if this is RoCE. CM calls ib_modify_port() regardless
1011 * of whether port link layer is ETH or IB. For ETH ports, qkey
1012 * violations and port capabilities are not meaningful.
1017 mutex_lock(&mdev->cap_mask_mutex);
1019 err = ib_query_port(ibdev, port, &attr);
1023 cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
1024 ~props->clr_port_cap_mask;
1026 err = mlx4_ib_SET_PORT(mdev, port,
1027 !!(mask & IB_PORT_RESET_QKEY_CNTR),
1031 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
1035 static struct ib_ucontext *mlx4_ib_alloc_ucontext(struct ib_device *ibdev,
1036 struct ib_udata *udata)
1038 struct mlx4_ib_dev *dev = to_mdev(ibdev);
1039 struct mlx4_ib_ucontext *context;
1040 struct mlx4_ib_alloc_ucontext_resp_v3 resp_v3;
1041 struct mlx4_ib_alloc_ucontext_resp resp;
1044 if (!dev->ib_active)
1045 return ERR_PTR(-EAGAIN);
1047 if (ibdev->uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION) {
1048 resp_v3.qp_tab_size = dev->dev->caps.num_qps;
1049 resp_v3.bf_reg_size = dev->dev->caps.bf_reg_size;
1050 resp_v3.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
1052 resp.dev_caps = dev->dev->caps.userspace_caps;
1053 resp.qp_tab_size = dev->dev->caps.num_qps;
1054 resp.bf_reg_size = dev->dev->caps.bf_reg_size;
1055 resp.bf_regs_per_page = dev->dev->caps.bf_regs_per_page;
1056 resp.cqe_size = dev->dev->caps.cqe_size;
1059 context = kzalloc(sizeof(*context), GFP_KERNEL);
1061 return ERR_PTR(-ENOMEM);
1063 err = mlx4_uar_alloc(to_mdev(ibdev)->dev, &context->uar);
1066 return ERR_PTR(err);
1069 INIT_LIST_HEAD(&context->db_page_list);
1070 mutex_init(&context->db_page_mutex);
1072 if (ibdev->uverbs_abi_ver == MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION)
1073 err = ib_copy_to_udata(udata, &resp_v3, sizeof(resp_v3));
1075 err = ib_copy_to_udata(udata, &resp, sizeof(resp));
1078 mlx4_uar_free(to_mdev(ibdev)->dev, &context->uar);
1080 return ERR_PTR(-EFAULT);
1083 return &context->ibucontext;
1086 static int mlx4_ib_dealloc_ucontext(struct ib_ucontext *ibcontext)
1088 struct mlx4_ib_ucontext *context = to_mucontext(ibcontext);
1090 mlx4_uar_free(to_mdev(ibcontext->device)->dev, &context->uar);
1096 static void mlx4_ib_vma_open(struct vm_area_struct *area)
1098 /* vma_open is called when a new VMA is created on top of our VMA.
1099 * This is done through either mremap flow or split_vma (usually due
1100 * to mlock, madvise, munmap, etc.). We do not support a clone of the
1101 * vma, as this VMA is strongly hardware related. Therefore we set the
1102 * vm_ops of the newly created/cloned VMA to NULL, to prevent it from
1103 * calling us again and trying to do incorrect actions. We assume that
1104 * the original vma size is exactly a single page that there will be no
1105 * "splitting" operations on.
1107 area->vm_ops = NULL;
1110 static void mlx4_ib_vma_close(struct vm_area_struct *area)
1112 struct mlx4_ib_vma_private_data *mlx4_ib_vma_priv_data;
1114 /* It's guaranteed that all VMAs opened on a FD are closed before the
1115 * file itself is closed, therefore no sync is needed with the regular
1116 * closing flow. (e.g. mlx4_ib_dealloc_ucontext) However need a sync
1117 * with accessing the vma as part of mlx4_ib_disassociate_ucontext.
1118 * The close operation is usually called under mm->mmap_sem except when
1119 * process is exiting. The exiting case is handled explicitly as part
1120 * of mlx4_ib_disassociate_ucontext.
1122 mlx4_ib_vma_priv_data = (struct mlx4_ib_vma_private_data *)
1123 area->vm_private_data;
1125 /* set the vma context pointer to null in the mlx4_ib driver's private
1126 * data to protect against a race condition in mlx4_ib_dissassociate_ucontext().
1128 mlx4_ib_vma_priv_data->vma = NULL;
1131 static const struct vm_operations_struct mlx4_ib_vm_ops = {
1132 .open = mlx4_ib_vma_open,
1133 .close = mlx4_ib_vma_close
1136 static void mlx4_ib_disassociate_ucontext(struct ib_ucontext *ibcontext)
1140 struct vm_area_struct *vma;
1141 struct mlx4_ib_ucontext *context = to_mucontext(ibcontext);
1142 struct task_struct *owning_process = NULL;
1143 struct mm_struct *owning_mm = NULL;
1145 owning_process = get_pid_task(ibcontext->tgid, PIDTYPE_PID);
1146 if (!owning_process)
1149 owning_mm = get_task_mm(owning_process);
1151 pr_info("no mm, disassociate ucontext is pending task termination\n");
1153 /* make sure that task is dead before returning, it may
1154 * prevent a rare case of module down in parallel to a
1155 * call to mlx4_ib_vma_close.
1157 put_task_struct(owning_process);
1159 owning_process = get_pid_task(ibcontext->tgid,
1161 if (!owning_process ||
1162 owning_process->state == TASK_DEAD) {
1163 pr_info("disassociate ucontext done, task was terminated\n");
1164 /* in case task was dead need to release the task struct */
1166 put_task_struct(owning_process);
1172 /* need to protect from a race on closing the vma as part of
1173 * mlx4_ib_vma_close().
1175 down_write(&owning_mm->mmap_sem);
1176 for (i = 0; i < HW_BAR_COUNT; i++) {
1177 vma = context->hw_bar_info[i].vma;
1181 ret = zap_vma_ptes(context->hw_bar_info[i].vma,
1182 context->hw_bar_info[i].vma->vm_start,
1185 pr_err("Error: zap_vma_ptes failed for index=%d, ret=%d\n", i, ret);
1189 context->hw_bar_info[i].vma->vm_flags &=
1190 ~(VM_SHARED | VM_MAYSHARE);
1191 /* context going to be destroyed, should not access ops any more */
1192 context->hw_bar_info[i].vma->vm_ops = NULL;
1195 up_write(&owning_mm->mmap_sem);
1197 put_task_struct(owning_process);
1200 static void mlx4_ib_set_vma_data(struct vm_area_struct *vma,
1201 struct mlx4_ib_vma_private_data *vma_private_data)
1203 vma_private_data->vma = vma;
1204 vma->vm_private_data = vma_private_data;
1205 vma->vm_ops = &mlx4_ib_vm_ops;
1208 static int mlx4_ib_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
1210 struct mlx4_ib_dev *dev = to_mdev(context->device);
1211 struct mlx4_ib_ucontext *mucontext = to_mucontext(context);
1213 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1216 if (vma->vm_pgoff == 0) {
1217 /* We prevent double mmaping on same context */
1218 if (mucontext->hw_bar_info[HW_BAR_DB].vma)
1221 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1223 if (io_remap_pfn_range(vma, vma->vm_start,
1224 to_mucontext(context)->uar.pfn,
1225 PAGE_SIZE, vma->vm_page_prot))
1228 mlx4_ib_set_vma_data(vma, &mucontext->hw_bar_info[HW_BAR_DB]);
1230 } else if (vma->vm_pgoff == 1 && dev->dev->caps.bf_reg_size != 0) {
1231 /* We prevent double mmaping on same context */
1232 if (mucontext->hw_bar_info[HW_BAR_BF].vma)
1235 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
1237 if (io_remap_pfn_range(vma, vma->vm_start,
1238 to_mucontext(context)->uar.pfn +
1239 dev->dev->caps.num_uars,
1240 PAGE_SIZE, vma->vm_page_prot))
1243 mlx4_ib_set_vma_data(vma, &mucontext->hw_bar_info[HW_BAR_BF]);
1245 } else if (vma->vm_pgoff == 3) {
1246 struct mlx4_clock_params params;
1249 /* We prevent double mmaping on same context */
1250 if (mucontext->hw_bar_info[HW_BAR_CLOCK].vma)
1253 ret = mlx4_get_internal_clock_params(dev->dev, ¶ms);
1258 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1259 if (io_remap_pfn_range(vma, vma->vm_start,
1260 (pci_resource_start(dev->dev->persist->pdev,
1264 PAGE_SIZE, vma->vm_page_prot))
1267 mlx4_ib_set_vma_data(vma,
1268 &mucontext->hw_bar_info[HW_BAR_CLOCK]);
1276 static struct ib_pd *mlx4_ib_alloc_pd(struct ib_device *ibdev,
1277 struct ib_ucontext *context,
1278 struct ib_udata *udata)
1280 struct mlx4_ib_pd *pd;
1283 pd = kmalloc(sizeof *pd, GFP_KERNEL);
1285 return ERR_PTR(-ENOMEM);
1287 err = mlx4_pd_alloc(to_mdev(ibdev)->dev, &pd->pdn);
1290 return ERR_PTR(err);
1294 if (ib_copy_to_udata(udata, &pd->pdn, sizeof (__u32))) {
1295 mlx4_pd_free(to_mdev(ibdev)->dev, pd->pdn);
1297 return ERR_PTR(-EFAULT);
1303 static int mlx4_ib_dealloc_pd(struct ib_pd *pd)
1305 mlx4_pd_free(to_mdev(pd->device)->dev, to_mpd(pd)->pdn);
1311 static struct ib_xrcd *mlx4_ib_alloc_xrcd(struct ib_device *ibdev,
1312 struct ib_ucontext *context,
1313 struct ib_udata *udata)
1315 struct mlx4_ib_xrcd *xrcd;
1316 struct ib_cq_init_attr cq_attr = {};
1319 if (!(to_mdev(ibdev)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1320 return ERR_PTR(-ENOSYS);
1322 xrcd = kmalloc(sizeof *xrcd, GFP_KERNEL);
1324 return ERR_PTR(-ENOMEM);
1326 err = mlx4_xrcd_alloc(to_mdev(ibdev)->dev, &xrcd->xrcdn);
1330 xrcd->pd = ib_alloc_pd(ibdev, 0);
1331 if (IS_ERR(xrcd->pd)) {
1332 err = PTR_ERR(xrcd->pd);
1337 xrcd->cq = ib_create_cq(ibdev, NULL, NULL, xrcd, &cq_attr);
1338 if (IS_ERR(xrcd->cq)) {
1339 err = PTR_ERR(xrcd->cq);
1343 return &xrcd->ibxrcd;
1346 ib_dealloc_pd(xrcd->pd);
1348 mlx4_xrcd_free(to_mdev(ibdev)->dev, xrcd->xrcdn);
1351 return ERR_PTR(err);
1354 static int mlx4_ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1356 ib_destroy_cq(to_mxrcd(xrcd)->cq);
1357 ib_dealloc_pd(to_mxrcd(xrcd)->pd);
1358 mlx4_xrcd_free(to_mdev(xrcd->device)->dev, to_mxrcd(xrcd)->xrcdn);
1364 static int add_gid_entry(struct ib_qp *ibqp, union ib_gid *gid)
1366 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
1367 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
1368 struct mlx4_ib_gid_entry *ge;
1370 ge = kzalloc(sizeof *ge, GFP_KERNEL);
1375 if (mlx4_ib_add_mc(mdev, mqp, gid)) {
1376 ge->port = mqp->port;
1380 mutex_lock(&mqp->mutex);
1381 list_add_tail(&ge->list, &mqp->gid_list);
1382 mutex_unlock(&mqp->mutex);
1387 static void mlx4_ib_delete_counters_table(struct mlx4_ib_dev *ibdev,
1388 struct mlx4_ib_counters *ctr_table)
1390 struct counter_index *counter, *tmp_count;
1392 mutex_lock(&ctr_table->mutex);
1393 list_for_each_entry_safe(counter, tmp_count, &ctr_table->counters_list,
1395 if (counter->allocated)
1396 mlx4_counter_free(ibdev->dev, counter->index);
1397 list_del(&counter->list);
1400 mutex_unlock(&ctr_table->mutex);
1403 int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
1406 struct net_device *ndev;
1412 spin_lock_bh(&mdev->iboe.lock);
1413 ndev = mdev->iboe.netdevs[mqp->port - 1];
1416 spin_unlock_bh(&mdev->iboe.lock);
1426 struct mlx4_ib_steering {
1427 struct list_head list;
1428 struct mlx4_flow_reg_id reg_id;
1432 #define LAST_ETH_FIELD vlan_tag
1433 #define LAST_IB_FIELD sl
1434 #define LAST_IPV4_FIELD dst_ip
1435 #define LAST_TCP_UDP_FIELD src_port
1437 /* Field is the last supported field */
1438 #define FIELDS_NOT_SUPPORTED(filter, field)\
1439 memchr_inv((void *)&filter.field +\
1440 sizeof(filter.field), 0,\
1442 offsetof(typeof(filter), field) -\
1443 sizeof(filter.field))
1445 static int parse_flow_attr(struct mlx4_dev *dev,
1447 union ib_flow_spec *ib_spec,
1448 struct _rule_hw *mlx4_spec)
1450 enum mlx4_net_trans_rule_id type;
1452 switch (ib_spec->type) {
1453 case IB_FLOW_SPEC_ETH:
1454 if (FIELDS_NOT_SUPPORTED(ib_spec->eth.mask, LAST_ETH_FIELD))
1457 type = MLX4_NET_TRANS_RULE_ID_ETH;
1458 memcpy(mlx4_spec->eth.dst_mac, ib_spec->eth.val.dst_mac,
1460 memcpy(mlx4_spec->eth.dst_mac_msk, ib_spec->eth.mask.dst_mac,
1462 mlx4_spec->eth.vlan_tag = ib_spec->eth.val.vlan_tag;
1463 mlx4_spec->eth.vlan_tag_msk = ib_spec->eth.mask.vlan_tag;
1465 case IB_FLOW_SPEC_IB:
1466 if (FIELDS_NOT_SUPPORTED(ib_spec->ib.mask, LAST_IB_FIELD))
1469 type = MLX4_NET_TRANS_RULE_ID_IB;
1470 mlx4_spec->ib.l3_qpn =
1471 cpu_to_be32(qp_num);
1472 mlx4_spec->ib.qpn_mask =
1473 cpu_to_be32(MLX4_IB_FLOW_QPN_MASK);
1477 case IB_FLOW_SPEC_IPV4:
1478 if (FIELDS_NOT_SUPPORTED(ib_spec->ipv4.mask, LAST_IPV4_FIELD))
1481 type = MLX4_NET_TRANS_RULE_ID_IPV4;
1482 mlx4_spec->ipv4.src_ip = ib_spec->ipv4.val.src_ip;
1483 mlx4_spec->ipv4.src_ip_msk = ib_spec->ipv4.mask.src_ip;
1484 mlx4_spec->ipv4.dst_ip = ib_spec->ipv4.val.dst_ip;
1485 mlx4_spec->ipv4.dst_ip_msk = ib_spec->ipv4.mask.dst_ip;
1488 case IB_FLOW_SPEC_TCP:
1489 case IB_FLOW_SPEC_UDP:
1490 if (FIELDS_NOT_SUPPORTED(ib_spec->tcp_udp.mask, LAST_TCP_UDP_FIELD))
1493 type = ib_spec->type == IB_FLOW_SPEC_TCP ?
1494 MLX4_NET_TRANS_RULE_ID_TCP :
1495 MLX4_NET_TRANS_RULE_ID_UDP;
1496 mlx4_spec->tcp_udp.dst_port = ib_spec->tcp_udp.val.dst_port;
1497 mlx4_spec->tcp_udp.dst_port_msk = ib_spec->tcp_udp.mask.dst_port;
1498 mlx4_spec->tcp_udp.src_port = ib_spec->tcp_udp.val.src_port;
1499 mlx4_spec->tcp_udp.src_port_msk = ib_spec->tcp_udp.mask.src_port;
1505 if (mlx4_map_sw_to_hw_steering_id(dev, type) < 0 ||
1506 mlx4_hw_rule_sz(dev, type) < 0)
1508 mlx4_spec->id = cpu_to_be16(mlx4_map_sw_to_hw_steering_id(dev, type));
1509 mlx4_spec->size = mlx4_hw_rule_sz(dev, type) >> 2;
1510 return mlx4_hw_rule_sz(dev, type);
1513 struct default_rules {
1514 __u32 mandatory_fields[IB_FLOW_SPEC_SUPPORT_LAYERS];
1515 __u32 mandatory_not_fields[IB_FLOW_SPEC_SUPPORT_LAYERS];
1516 __u32 rules_create_list[IB_FLOW_SPEC_SUPPORT_LAYERS];
1519 static const struct default_rules default_table[] = {
1521 .mandatory_fields = {IB_FLOW_SPEC_IPV4},
1522 .mandatory_not_fields = {IB_FLOW_SPEC_ETH},
1523 .rules_create_list = {IB_FLOW_SPEC_IB},
1524 .link_layer = IB_LINK_LAYER_INFINIBAND
1528 static int __mlx4_ib_default_rules_match(struct ib_qp *qp,
1529 struct ib_flow_attr *flow_attr)
1533 const struct default_rules *pdefault_rules = default_table;
1534 u8 link_layer = rdma_port_get_link_layer(qp->device, flow_attr->port);
1536 for (i = 0; i < ARRAY_SIZE(default_table); i++, pdefault_rules++) {
1537 __u32 field_types[IB_FLOW_SPEC_SUPPORT_LAYERS];
1538 memset(&field_types, 0, sizeof(field_types));
1540 if (link_layer != pdefault_rules->link_layer)
1543 ib_flow = flow_attr + 1;
1544 /* we assume the specs are sorted */
1545 for (j = 0, k = 0; k < IB_FLOW_SPEC_SUPPORT_LAYERS &&
1546 j < flow_attr->num_of_specs; k++) {
1547 union ib_flow_spec *current_flow =
1548 (union ib_flow_spec *)ib_flow;
1550 /* same layer but different type */
1551 if (((current_flow->type & IB_FLOW_SPEC_LAYER_MASK) ==
1552 (pdefault_rules->mandatory_fields[k] &
1553 IB_FLOW_SPEC_LAYER_MASK)) &&
1554 (current_flow->type !=
1555 pdefault_rules->mandatory_fields[k]))
1558 /* same layer, try match next one */
1559 if (current_flow->type ==
1560 pdefault_rules->mandatory_fields[k]) {
1563 ((union ib_flow_spec *)ib_flow)->size;
1567 ib_flow = flow_attr + 1;
1568 for (j = 0; j < flow_attr->num_of_specs;
1569 j++, ib_flow += ((union ib_flow_spec *)ib_flow)->size)
1570 for (k = 0; k < IB_FLOW_SPEC_SUPPORT_LAYERS; k++)
1571 /* same layer and same type */
1572 if (((union ib_flow_spec *)ib_flow)->type ==
1573 pdefault_rules->mandatory_not_fields[k])
1582 static int __mlx4_ib_create_default_rules(
1583 struct mlx4_ib_dev *mdev,
1585 const struct default_rules *pdefault_rules,
1586 struct _rule_hw *mlx4_spec) {
1590 for (i = 0; i < ARRAY_SIZE(pdefault_rules->rules_create_list); i++) {
1592 union ib_flow_spec ib_spec;
1593 switch (pdefault_rules->rules_create_list[i]) {
1597 case IB_FLOW_SPEC_IB:
1598 ib_spec.type = IB_FLOW_SPEC_IB;
1599 ib_spec.size = sizeof(struct ib_flow_spec_ib);
1606 /* We must put empty rule, qpn is being ignored */
1607 ret = parse_flow_attr(mdev->dev, 0, &ib_spec,
1610 pr_info("invalid parsing\n");
1614 mlx4_spec = (void *)mlx4_spec + ret;
1620 static int __mlx4_ib_create_flow(struct ib_qp *qp, struct ib_flow_attr *flow_attr,
1622 enum mlx4_net_trans_promisc_mode flow_type,
1628 struct mlx4_ib_dev *mdev = to_mdev(qp->device);
1629 struct mlx4_cmd_mailbox *mailbox;
1630 struct mlx4_net_trans_rule_hw_ctrl *ctrl;
1633 static const u16 __mlx4_domain[] = {
1634 [IB_FLOW_DOMAIN_USER] = MLX4_DOMAIN_UVERBS,
1635 [IB_FLOW_DOMAIN_ETHTOOL] = MLX4_DOMAIN_ETHTOOL,
1636 [IB_FLOW_DOMAIN_RFS] = MLX4_DOMAIN_RFS,
1637 [IB_FLOW_DOMAIN_NIC] = MLX4_DOMAIN_NIC,
1640 if (flow_attr->priority > MLX4_IB_FLOW_MAX_PRIO) {
1641 pr_err("Invalid priority value %d\n", flow_attr->priority);
1645 if (domain >= IB_FLOW_DOMAIN_NUM) {
1646 pr_err("Invalid domain value %d\n", domain);
1650 if (mlx4_map_sw_to_hw_steering_mode(mdev->dev, flow_type) < 0)
1653 mailbox = mlx4_alloc_cmd_mailbox(mdev->dev);
1654 if (IS_ERR(mailbox))
1655 return PTR_ERR(mailbox);
1656 ctrl = mailbox->buf;
1658 ctrl->prio = cpu_to_be16(__mlx4_domain[domain] |
1659 flow_attr->priority);
1660 ctrl->type = mlx4_map_sw_to_hw_steering_mode(mdev->dev, flow_type);
1661 ctrl->port = flow_attr->port;
1662 ctrl->qpn = cpu_to_be32(qp->qp_num);
1664 ib_flow = flow_attr + 1;
1665 size += sizeof(struct mlx4_net_trans_rule_hw_ctrl);
1666 /* Add default flows */
1667 default_flow = __mlx4_ib_default_rules_match(qp, flow_attr);
1668 if (default_flow >= 0) {
1669 ret = __mlx4_ib_create_default_rules(
1670 mdev, qp, default_table + default_flow,
1671 mailbox->buf + size);
1673 mlx4_free_cmd_mailbox(mdev->dev, mailbox);
1678 for (i = 0; i < flow_attr->num_of_specs; i++) {
1679 ret = parse_flow_attr(mdev->dev, qp->qp_num, ib_flow,
1680 mailbox->buf + size);
1682 mlx4_free_cmd_mailbox(mdev->dev, mailbox);
1685 ib_flow += ((union ib_flow_spec *) ib_flow)->size;
1689 if (mlx4_is_master(mdev->dev) && flow_type == MLX4_FS_REGULAR &&
1690 flow_attr->num_of_specs == 1) {
1691 struct _rule_hw *rule_header = (struct _rule_hw *)(ctrl + 1);
1692 enum ib_flow_spec_type header_spec =
1693 ((union ib_flow_spec *)(flow_attr + 1))->type;
1695 if (header_spec == IB_FLOW_SPEC_ETH)
1696 mlx4_handle_eth_header_mcast_prio(ctrl, rule_header);
1699 ret = mlx4_cmd_imm(mdev->dev, mailbox->dma, reg_id, size >> 2, 0,
1700 MLX4_QP_FLOW_STEERING_ATTACH, MLX4_CMD_TIME_CLASS_A,
1703 pr_err("mcg table is full. Fail to register network rule.\n");
1704 else if (ret == -ENXIO)
1705 pr_err("Device managed flow steering is disabled. Fail to register network rule.\n");
1707 pr_err("Invalid argument. Fail to register network rule.\n");
1709 mlx4_free_cmd_mailbox(mdev->dev, mailbox);
1713 static int __mlx4_ib_destroy_flow(struct mlx4_dev *dev, u64 reg_id)
1716 err = mlx4_cmd(dev, reg_id, 0, 0,
1717 MLX4_QP_FLOW_STEERING_DETACH, MLX4_CMD_TIME_CLASS_A,
1720 pr_err("Fail to detach network rule. registration id = 0x%llx\n",
1725 static int mlx4_ib_tunnel_steer_add(struct ib_qp *qp, struct ib_flow_attr *flow_attr,
1729 union ib_flow_spec *ib_spec;
1730 struct mlx4_dev *dev = to_mdev(qp->device)->dev;
1733 if (dev->caps.tunnel_offload_mode != MLX4_TUNNEL_OFFLOAD_MODE_VXLAN ||
1734 dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_STATIC)
1735 return 0; /* do nothing */
1737 ib_flow = flow_attr + 1;
1738 ib_spec = (union ib_flow_spec *)ib_flow;
1740 if (ib_spec->type != IB_FLOW_SPEC_ETH || flow_attr->num_of_specs != 1)
1741 return 0; /* do nothing */
1743 err = mlx4_tunnel_steer_add(to_mdev(qp->device)->dev, ib_spec->eth.val.dst_mac,
1744 flow_attr->port, qp->qp_num,
1745 MLX4_DOMAIN_UVERBS | (flow_attr->priority & 0xff),
1750 static int mlx4_ib_add_dont_trap_rule(struct mlx4_dev *dev,
1751 struct ib_flow_attr *flow_attr,
1752 enum mlx4_net_trans_promisc_mode *type)
1756 if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_UC_MC_SNIFFER) ||
1757 (dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_STATIC) ||
1758 (flow_attr->num_of_specs > 1) || (flow_attr->priority != 0)) {
1762 if (flow_attr->num_of_specs == 0) {
1763 type[0] = MLX4_FS_MC_SNIFFER;
1764 type[1] = MLX4_FS_UC_SNIFFER;
1766 union ib_flow_spec *ib_spec;
1768 ib_spec = (union ib_flow_spec *)(flow_attr + 1);
1769 if (ib_spec->type != IB_FLOW_SPEC_ETH)
1772 /* if all is zero than MC and UC */
1773 if (is_zero_ether_addr(ib_spec->eth.mask.dst_mac)) {
1774 type[0] = MLX4_FS_MC_SNIFFER;
1775 type[1] = MLX4_FS_UC_SNIFFER;
1777 u8 mac[ETH_ALEN] = {ib_spec->eth.mask.dst_mac[0] ^ 0x01,
1778 ib_spec->eth.mask.dst_mac[1],
1779 ib_spec->eth.mask.dst_mac[2],
1780 ib_spec->eth.mask.dst_mac[3],
1781 ib_spec->eth.mask.dst_mac[4],
1782 ib_spec->eth.mask.dst_mac[5]};
1784 /* Above xor was only on MC bit, non empty mask is valid
1785 * only if this bit is set and rest are zero.
1787 if (!is_zero_ether_addr(&mac[0]))
1790 if (is_multicast_ether_addr(ib_spec->eth.val.dst_mac))
1791 type[0] = MLX4_FS_MC_SNIFFER;
1793 type[0] = MLX4_FS_UC_SNIFFER;
1800 static struct ib_flow *mlx4_ib_create_flow(struct ib_qp *qp,
1801 struct ib_flow_attr *flow_attr,
1804 int err = 0, i = 0, j = 0;
1805 struct mlx4_ib_flow *mflow;
1806 enum mlx4_net_trans_promisc_mode type[2];
1807 struct mlx4_dev *dev = (to_mdev(qp->device))->dev;
1808 int is_bonded = mlx4_is_bonded(dev);
1810 if (flow_attr->port < 1 || flow_attr->port > qp->device->phys_port_cnt)
1811 return ERR_PTR(-EINVAL);
1813 if ((flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP) &&
1814 (flow_attr->type != IB_FLOW_ATTR_NORMAL))
1815 return ERR_PTR(-EOPNOTSUPP);
1817 memset(type, 0, sizeof(type));
1819 mflow = kzalloc(sizeof(*mflow), GFP_KERNEL);
1825 switch (flow_attr->type) {
1826 case IB_FLOW_ATTR_NORMAL:
1827 /* If dont trap flag (continue match) is set, under specific
1828 * condition traffic be replicated to given qp,
1829 * without stealing it
1831 if (unlikely(flow_attr->flags & IB_FLOW_ATTR_FLAGS_DONT_TRAP)) {
1832 err = mlx4_ib_add_dont_trap_rule(dev,
1838 type[0] = MLX4_FS_REGULAR;
1842 case IB_FLOW_ATTR_ALL_DEFAULT:
1843 type[0] = MLX4_FS_ALL_DEFAULT;
1846 case IB_FLOW_ATTR_MC_DEFAULT:
1847 type[0] = MLX4_FS_MC_DEFAULT;
1850 case IB_FLOW_ATTR_SNIFFER:
1851 type[0] = MLX4_FS_MIRROR_RX_PORT;
1852 type[1] = MLX4_FS_MIRROR_SX_PORT;
1860 while (i < ARRAY_SIZE(type) && type[i]) {
1861 err = __mlx4_ib_create_flow(qp, flow_attr, domain, type[i],
1862 &mflow->reg_id[i].id);
1864 goto err_create_flow;
1866 /* Application always sees one port so the mirror rule
1867 * must be on port #2
1869 flow_attr->port = 2;
1870 err = __mlx4_ib_create_flow(qp, flow_attr,
1872 &mflow->reg_id[j].mirror);
1873 flow_attr->port = 1;
1875 goto err_create_flow;
1882 if (i < ARRAY_SIZE(type) && flow_attr->type == IB_FLOW_ATTR_NORMAL) {
1883 err = mlx4_ib_tunnel_steer_add(qp, flow_attr,
1884 &mflow->reg_id[i].id);
1886 goto err_create_flow;
1889 flow_attr->port = 2;
1890 err = mlx4_ib_tunnel_steer_add(qp, flow_attr,
1891 &mflow->reg_id[j].mirror);
1892 flow_attr->port = 1;
1894 goto err_create_flow;
1897 /* function to create mirror rule */
1901 return &mflow->ibflow;
1905 (void)__mlx4_ib_destroy_flow(to_mdev(qp->device)->dev,
1906 mflow->reg_id[i].id);
1911 (void)__mlx4_ib_destroy_flow(to_mdev(qp->device)->dev,
1912 mflow->reg_id[j].mirror);
1917 return ERR_PTR(err);
1920 static int mlx4_ib_destroy_flow(struct ib_flow *flow_id)
1924 struct mlx4_ib_dev *mdev = to_mdev(flow_id->qp->device);
1925 struct mlx4_ib_flow *mflow = to_mflow(flow_id);
1927 while (i < ARRAY_SIZE(mflow->reg_id) && mflow->reg_id[i].id) {
1928 err = __mlx4_ib_destroy_flow(mdev->dev, mflow->reg_id[i].id);
1931 if (mflow->reg_id[i].mirror) {
1932 err = __mlx4_ib_destroy_flow(mdev->dev,
1933 mflow->reg_id[i].mirror);
1944 static int mlx4_ib_mcg_attach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
1947 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
1948 struct mlx4_dev *dev = mdev->dev;
1949 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
1950 struct mlx4_ib_steering *ib_steering = NULL;
1951 enum mlx4_protocol prot = MLX4_PROT_IB_IPV6;
1952 struct mlx4_flow_reg_id reg_id;
1954 if (mdev->dev->caps.steering_mode ==
1955 MLX4_STEERING_MODE_DEVICE_MANAGED) {
1956 ib_steering = kmalloc(sizeof(*ib_steering), GFP_KERNEL);
1961 err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw, mqp->port,
1963 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
1966 pr_err("multicast attach op failed, err %d\n", err);
1971 if (mlx4_is_bonded(dev)) {
1972 err = mlx4_multicast_attach(mdev->dev, &mqp->mqp, gid->raw,
1973 (mqp->port == 1) ? 2 : 1,
1975 MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK),
1976 prot, ®_id.mirror);
1981 err = add_gid_entry(ibqp, gid);
1986 memcpy(ib_steering->gid.raw, gid->raw, 16);
1987 ib_steering->reg_id = reg_id;
1988 mutex_lock(&mqp->mutex);
1989 list_add(&ib_steering->list, &mqp->steering_rules);
1990 mutex_unlock(&mqp->mutex);
1995 mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
1998 mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
1999 prot, reg_id.mirror);
2006 static struct mlx4_ib_gid_entry *find_gid_entry(struct mlx4_ib_qp *qp, u8 *raw)
2008 struct mlx4_ib_gid_entry *ge;
2009 struct mlx4_ib_gid_entry *tmp;
2010 struct mlx4_ib_gid_entry *ret = NULL;
2012 list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
2013 if (!memcmp(raw, ge->gid.raw, 16)) {
2022 static int mlx4_ib_mcg_detach(struct ib_qp *ibqp, union ib_gid *gid, u16 lid)
2025 struct mlx4_ib_dev *mdev = to_mdev(ibqp->device);
2026 struct mlx4_dev *dev = mdev->dev;
2027 struct mlx4_ib_qp *mqp = to_mqp(ibqp);
2028 struct net_device *ndev;
2029 struct mlx4_ib_gid_entry *ge;
2030 struct mlx4_flow_reg_id reg_id = {0, 0};
2031 enum mlx4_protocol prot = MLX4_PROT_IB_IPV6;
2033 if (mdev->dev->caps.steering_mode ==
2034 MLX4_STEERING_MODE_DEVICE_MANAGED) {
2035 struct mlx4_ib_steering *ib_steering;
2037 mutex_lock(&mqp->mutex);
2038 list_for_each_entry(ib_steering, &mqp->steering_rules, list) {
2039 if (!memcmp(ib_steering->gid.raw, gid->raw, 16)) {
2040 list_del(&ib_steering->list);
2044 mutex_unlock(&mqp->mutex);
2045 if (&ib_steering->list == &mqp->steering_rules) {
2046 pr_err("Couldn't find reg_id for mgid. Steering rule is left attached\n");
2049 reg_id = ib_steering->reg_id;
2053 err = mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
2058 if (mlx4_is_bonded(dev)) {
2059 err = mlx4_multicast_detach(mdev->dev, &mqp->mqp, gid->raw,
2060 prot, reg_id.mirror);
2065 mutex_lock(&mqp->mutex);
2066 ge = find_gid_entry(mqp, gid->raw);
2068 spin_lock_bh(&mdev->iboe.lock);
2069 ndev = ge->added ? mdev->iboe.netdevs[ge->port - 1] : NULL;
2072 spin_unlock_bh(&mdev->iboe.lock);
2075 list_del(&ge->list);
2078 pr_warn("could not find mgid entry\n");
2080 mutex_unlock(&mqp->mutex);
2085 static int init_node_data(struct mlx4_ib_dev *dev)
2087 struct ib_smp *in_mad = NULL;
2088 struct ib_smp *out_mad = NULL;
2089 int mad_ifc_flags = MLX4_MAD_IFC_IGNORE_KEYS;
2092 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL);
2093 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
2094 if (!in_mad || !out_mad)
2097 init_query_mad(in_mad);
2098 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
2099 if (mlx4_is_master(dev->dev))
2100 mad_ifc_flags |= MLX4_MAD_IFC_NET_VIEW;
2102 err = mlx4_MAD_IFC(dev, mad_ifc_flags, 1, NULL, NULL, in_mad, out_mad);
2106 memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
2108 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
2110 err = mlx4_MAD_IFC(dev, mad_ifc_flags, 1, NULL, NULL, in_mad, out_mad);
2114 dev->dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
2115 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
2123 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
2126 struct mlx4_ib_dev *dev =
2127 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
2128 return sprintf(buf, "MT%d\n", dev->dev->persist->pdev->device);
2131 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
2134 struct mlx4_ib_dev *dev =
2135 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
2136 return sprintf(buf, "%x\n", dev->dev->rev_id);
2139 static ssize_t show_board(struct device *device, struct device_attribute *attr,
2142 struct mlx4_ib_dev *dev =
2143 container_of(device, struct mlx4_ib_dev, ib_dev.dev);
2144 return sprintf(buf, "%.*s\n", MLX4_BOARD_ID_LEN,
2145 dev->dev->board_id);
2148 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
2149 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca, NULL);
2150 static DEVICE_ATTR(board_id, S_IRUGO, show_board, NULL);
2152 static struct device_attribute *mlx4_class_attributes[] = {
2158 struct diag_counter {
2163 #define DIAG_COUNTER(_name, _offset) \
2164 { .name = #_name, .offset = _offset }
2166 static const struct diag_counter diag_basic[] = {
2167 DIAG_COUNTER(rq_num_lle, 0x00),
2168 DIAG_COUNTER(sq_num_lle, 0x04),
2169 DIAG_COUNTER(rq_num_lqpoe, 0x08),
2170 DIAG_COUNTER(sq_num_lqpoe, 0x0C),
2171 DIAG_COUNTER(rq_num_lpe, 0x18),
2172 DIAG_COUNTER(sq_num_lpe, 0x1C),
2173 DIAG_COUNTER(rq_num_wrfe, 0x20),
2174 DIAG_COUNTER(sq_num_wrfe, 0x24),
2175 DIAG_COUNTER(sq_num_mwbe, 0x2C),
2176 DIAG_COUNTER(sq_num_bre, 0x34),
2177 DIAG_COUNTER(sq_num_rire, 0x44),
2178 DIAG_COUNTER(rq_num_rire, 0x48),
2179 DIAG_COUNTER(sq_num_rae, 0x4C),
2180 DIAG_COUNTER(rq_num_rae, 0x50),
2181 DIAG_COUNTER(sq_num_roe, 0x54),
2182 DIAG_COUNTER(sq_num_tree, 0x5C),
2183 DIAG_COUNTER(sq_num_rree, 0x64),
2184 DIAG_COUNTER(rq_num_rnr, 0x68),
2185 DIAG_COUNTER(sq_num_rnr, 0x6C),
2186 DIAG_COUNTER(rq_num_oos, 0x100),
2187 DIAG_COUNTER(sq_num_oos, 0x104),
2190 static const struct diag_counter diag_ext[] = {
2191 DIAG_COUNTER(rq_num_dup, 0x130),
2192 DIAG_COUNTER(sq_num_to, 0x134),
2195 static const struct diag_counter diag_device_only[] = {
2196 DIAG_COUNTER(num_cqovf, 0x1A0),
2197 DIAG_COUNTER(rq_num_udsdprd, 0x118),
2200 static struct rdma_hw_stats *mlx4_ib_alloc_hw_stats(struct ib_device *ibdev,
2203 struct mlx4_ib_dev *dev = to_mdev(ibdev);
2204 struct mlx4_ib_diag_counters *diag = dev->diag_counters;
2206 if (!diag[!!port_num].name)
2209 return rdma_alloc_hw_stats_struct(diag[!!port_num].name,
2210 diag[!!port_num].num_counters,
2211 RDMA_HW_STATS_DEFAULT_LIFESPAN);
2214 static int mlx4_ib_get_hw_stats(struct ib_device *ibdev,
2215 struct rdma_hw_stats *stats,
2218 struct mlx4_ib_dev *dev = to_mdev(ibdev);
2219 struct mlx4_ib_diag_counters *diag = dev->diag_counters;
2220 u32 hw_value[ARRAY_SIZE(diag_device_only) +
2221 ARRAY_SIZE(diag_ext) + ARRAY_SIZE(diag_basic)] = {};
2225 ret = mlx4_query_diag_counters(dev->dev,
2226 MLX4_OP_MOD_QUERY_TRANSPORT_CI_ERRORS,
2227 diag[!!port].offset, hw_value,
2228 diag[!!port].num_counters, port);
2233 for (i = 0; i < diag[!!port].num_counters; i++)
2234 stats->value[i] = hw_value[i];
2236 return diag[!!port].num_counters;
2239 static int __mlx4_ib_alloc_diag_counters(struct mlx4_ib_dev *ibdev,
2247 num_counters = ARRAY_SIZE(diag_basic);
2249 if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT)
2250 num_counters += ARRAY_SIZE(diag_ext);
2253 num_counters += ARRAY_SIZE(diag_device_only);
2255 *name = kcalloc(num_counters, sizeof(**name), GFP_KERNEL);
2259 *offset = kcalloc(num_counters, sizeof(**offset), GFP_KERNEL);
2263 *num = num_counters;
2272 static void mlx4_ib_fill_diag_counters(struct mlx4_ib_dev *ibdev,
2280 for (i = 0, j = 0; i < ARRAY_SIZE(diag_basic); i++, j++) {
2281 name[i] = diag_basic[i].name;
2282 offset[i] = diag_basic[i].offset;
2285 if (ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT) {
2286 for (i = 0; i < ARRAY_SIZE(diag_ext); i++, j++) {
2287 name[j] = diag_ext[i].name;
2288 offset[j] = diag_ext[i].offset;
2293 for (i = 0; i < ARRAY_SIZE(diag_device_only); i++, j++) {
2294 name[j] = diag_device_only[i].name;
2295 offset[j] = diag_device_only[i].offset;
2300 static int mlx4_ib_alloc_diag_counters(struct mlx4_ib_dev *ibdev)
2302 struct mlx4_ib_diag_counters *diag = ibdev->diag_counters;
2305 bool per_port = !!(ibdev->dev->caps.flags2 &
2306 MLX4_DEV_CAP_FLAG2_DIAG_PER_PORT);
2308 if (mlx4_is_slave(ibdev->dev))
2311 for (i = 0; i < MLX4_DIAG_COUNTERS_TYPES; i++) {
2312 /* i == 1 means we are building port counters */
2316 ret = __mlx4_ib_alloc_diag_counters(ibdev, &diag[i].name,
2318 &diag[i].num_counters, i);
2322 mlx4_ib_fill_diag_counters(ibdev, diag[i].name,
2326 ibdev->ib_dev.get_hw_stats = mlx4_ib_get_hw_stats;
2327 ibdev->ib_dev.alloc_hw_stats = mlx4_ib_alloc_hw_stats;
2333 kfree(diag[i - 1].name);
2334 kfree(diag[i - 1].offset);
2340 static void mlx4_ib_diag_cleanup(struct mlx4_ib_dev *ibdev)
2344 for (i = 0; i < MLX4_DIAG_COUNTERS_TYPES; i++) {
2345 kfree(ibdev->diag_counters[i].offset);
2346 kfree(ibdev->diag_counters[i].name);
2350 #define MLX4_IB_INVALID_MAC ((u64)-1)
2351 static void mlx4_ib_update_qps(struct mlx4_ib_dev *ibdev,
2352 struct net_device *dev,
2356 u64 release_mac = MLX4_IB_INVALID_MAC;
2357 struct mlx4_ib_qp *qp;
2359 read_lock(&dev_base_lock);
2360 new_smac = mlx4_mac_to_u64(dev->dev_addr);
2361 read_unlock(&dev_base_lock);
2363 atomic64_set(&ibdev->iboe.mac[port - 1], new_smac);
2365 /* no need for update QP1 and mac registration in non-SRIOV */
2366 if (!mlx4_is_mfunc(ibdev->dev))
2369 mutex_lock(&ibdev->qp1_proxy_lock[port - 1]);
2370 qp = ibdev->qp1_proxy[port - 1];
2374 struct mlx4_update_qp_params update_params;
2376 mutex_lock(&qp->mutex);
2377 old_smac = qp->pri.smac;
2378 if (new_smac == old_smac)
2381 new_smac_index = mlx4_register_mac(ibdev->dev, port, new_smac);
2383 if (new_smac_index < 0)
2386 update_params.smac_index = new_smac_index;
2387 if (mlx4_update_qp(ibdev->dev, qp->mqp.qpn, MLX4_UPDATE_QP_SMAC,
2389 release_mac = new_smac;
2392 /* if old port was zero, no mac was yet registered for this QP */
2393 if (qp->pri.smac_port)
2394 release_mac = old_smac;
2395 qp->pri.smac = new_smac;
2396 qp->pri.smac_port = port;
2397 qp->pri.smac_index = new_smac_index;
2401 if (release_mac != MLX4_IB_INVALID_MAC)
2402 mlx4_unregister_mac(ibdev->dev, port, release_mac);
2404 mutex_unlock(&qp->mutex);
2405 mutex_unlock(&ibdev->qp1_proxy_lock[port - 1]);
2408 static void mlx4_ib_scan_netdevs(struct mlx4_ib_dev *ibdev,
2409 struct net_device *dev,
2410 unsigned long event)
2413 struct mlx4_ib_iboe *iboe;
2414 int update_qps_port = -1;
2419 iboe = &ibdev->iboe;
2421 spin_lock_bh(&iboe->lock);
2422 mlx4_foreach_ib_transport_port(port, ibdev->dev) {
2424 iboe->netdevs[port - 1] =
2425 mlx4_get_protocol_dev(ibdev->dev, MLX4_PROT_ETH, port);
2427 if (dev == iboe->netdevs[port - 1] &&
2428 (event == NETDEV_CHANGEADDR || event == NETDEV_REGISTER ||
2429 event == NETDEV_UP || event == NETDEV_CHANGE))
2430 update_qps_port = port;
2433 spin_unlock_bh(&iboe->lock);
2435 if (update_qps_port > 0)
2436 mlx4_ib_update_qps(ibdev, dev, update_qps_port);
2439 static int mlx4_ib_netdev_event(struct notifier_block *this,
2440 unsigned long event, void *ptr)
2442 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
2443 struct mlx4_ib_dev *ibdev;
2445 if (!net_eq(dev_net(dev), &init_net))
2448 ibdev = container_of(this, struct mlx4_ib_dev, iboe.nb);
2449 mlx4_ib_scan_netdevs(ibdev, dev, event);
2454 static void init_pkeys(struct mlx4_ib_dev *ibdev)
2460 if (mlx4_is_master(ibdev->dev)) {
2461 for (slave = 0; slave <= ibdev->dev->persist->num_vfs;
2463 for (port = 1; port <= ibdev->dev->caps.num_ports; ++port) {
2465 i < ibdev->dev->phys_caps.pkey_phys_table_len[port];
2467 ibdev->pkeys.virt2phys_pkey[slave][port - 1][i] =
2468 /* master has the identity virt2phys pkey mapping */
2469 (slave == mlx4_master_func_num(ibdev->dev) || !i) ? i :
2470 ibdev->dev->phys_caps.pkey_phys_table_len[port] - 1;
2471 mlx4_sync_pkey_table(ibdev->dev, slave, port, i,
2472 ibdev->pkeys.virt2phys_pkey[slave][port - 1][i]);
2476 /* initialize pkey cache */
2477 for (port = 1; port <= ibdev->dev->caps.num_ports; ++port) {
2479 i < ibdev->dev->phys_caps.pkey_phys_table_len[port];
2481 ibdev->pkeys.phys_pkey_cache[port-1][i] =
2487 static void mlx4_ib_alloc_eqs(struct mlx4_dev *dev, struct mlx4_ib_dev *ibdev)
2489 int i, j, eq = 0, total_eqs = 0;
2491 ibdev->eq_table = kcalloc(dev->caps.num_comp_vectors,
2492 sizeof(ibdev->eq_table[0]), GFP_KERNEL);
2493 if (!ibdev->eq_table)
2496 for (i = 1; i <= dev->caps.num_ports; i++) {
2497 for (j = 0; j < mlx4_get_eqs_per_port(dev, i);
2499 if (i > 1 && mlx4_is_eq_shared(dev, total_eqs))
2501 ibdev->eq_table[eq] = total_eqs;
2502 if (!mlx4_assign_eq(dev, i,
2503 &ibdev->eq_table[eq]))
2506 ibdev->eq_table[eq] = -1;
2510 for (i = eq; i < dev->caps.num_comp_vectors;
2511 ibdev->eq_table[i++] = -1)
2514 /* Advertise the new number of EQs to clients */
2515 ibdev->ib_dev.num_comp_vectors = eq;
2518 static void mlx4_ib_free_eqs(struct mlx4_dev *dev, struct mlx4_ib_dev *ibdev)
2521 int total_eqs = ibdev->ib_dev.num_comp_vectors;
2523 /* no eqs were allocated */
2524 if (!ibdev->eq_table)
2527 /* Reset the advertised EQ number */
2528 ibdev->ib_dev.num_comp_vectors = 0;
2530 for (i = 0; i < total_eqs; i++)
2531 mlx4_release_eq(dev, ibdev->eq_table[i]);
2533 kfree(ibdev->eq_table);
2534 ibdev->eq_table = NULL;
2537 static int mlx4_port_immutable(struct ib_device *ibdev, u8 port_num,
2538 struct ib_port_immutable *immutable)
2540 struct ib_port_attr attr;
2541 struct mlx4_ib_dev *mdev = to_mdev(ibdev);
2544 if (mlx4_ib_port_link_layer(ibdev, port_num) == IB_LINK_LAYER_INFINIBAND) {
2545 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
2546 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2548 if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE)
2549 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
2550 if (mdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2)
2551 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE |
2552 RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
2553 immutable->core_cap_flags |= RDMA_CORE_PORT_RAW_PACKET;
2554 if (immutable->core_cap_flags & (RDMA_CORE_PORT_IBA_ROCE |
2555 RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP))
2556 immutable->max_mad_size = IB_MGMT_MAD_SIZE;
2559 err = ib_query_port(ibdev, port_num, &attr);
2563 immutable->pkey_tbl_len = attr.pkey_tbl_len;
2564 immutable->gid_tbl_len = attr.gid_tbl_len;
2569 static void get_fw_ver_str(struct ib_device *device, char *str,
2572 struct mlx4_ib_dev *dev =
2573 container_of(device, struct mlx4_ib_dev, ib_dev);
2574 snprintf(str, str_len, "%d.%d.%d",
2575 (int) (dev->dev->caps.fw_ver >> 32),
2576 (int) (dev->dev->caps.fw_ver >> 16) & 0xffff,
2577 (int) dev->dev->caps.fw_ver & 0xffff);
2580 static void *mlx4_ib_add(struct mlx4_dev *dev)
2582 struct mlx4_ib_dev *ibdev;
2586 struct mlx4_ib_iboe *iboe;
2587 int ib_num_ports = 0;
2588 int num_req_counters;
2591 struct counter_index *new_counter_index = NULL;
2593 pr_info_once("%s", mlx4_ib_version);
2596 mlx4_foreach_ib_transport_port(i, dev)
2599 /* No point in registering a device with no ports... */
2603 ibdev = (struct mlx4_ib_dev *) ib_alloc_device(sizeof *ibdev);
2605 dev_err(&dev->persist->pdev->dev,
2606 "Device struct alloc failed\n");
2610 iboe = &ibdev->iboe;
2612 if (mlx4_pd_alloc(dev, &ibdev->priv_pdn))
2615 if (mlx4_uar_alloc(dev, &ibdev->priv_uar))
2618 ibdev->uar_map = ioremap((phys_addr_t) ibdev->priv_uar.pfn << PAGE_SHIFT,
2620 if (!ibdev->uar_map)
2622 MLX4_INIT_DOORBELL_LOCK(&ibdev->uar_lock);
2625 ibdev->bond_next_port = 0;
2627 strlcpy(ibdev->ib_dev.name, "mlx4_%d", IB_DEVICE_NAME_MAX);
2628 ibdev->ib_dev.owner = THIS_MODULE;
2629 ibdev->ib_dev.node_type = RDMA_NODE_IB_CA;
2630 ibdev->ib_dev.local_dma_lkey = dev->caps.reserved_lkey;
2631 ibdev->num_ports = num_ports;
2632 ibdev->ib_dev.phys_port_cnt = mlx4_is_bonded(dev) ?
2633 1 : ibdev->num_ports;
2634 ibdev->ib_dev.num_comp_vectors = dev->caps.num_comp_vectors;
2635 ibdev->ib_dev.dev.parent = &dev->persist->pdev->dev;
2636 ibdev->ib_dev.get_netdev = mlx4_ib_get_netdev;
2637 ibdev->ib_dev.add_gid = mlx4_ib_add_gid;
2638 ibdev->ib_dev.del_gid = mlx4_ib_del_gid;
2640 if (dev->caps.userspace_caps)
2641 ibdev->ib_dev.uverbs_abi_ver = MLX4_IB_UVERBS_ABI_VERSION;
2643 ibdev->ib_dev.uverbs_abi_ver = MLX4_IB_UVERBS_NO_DEV_CAPS_ABI_VERSION;
2645 ibdev->ib_dev.uverbs_cmd_mask =
2646 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) |
2647 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) |
2648 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) |
2649 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) |
2650 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) |
2651 (1ull << IB_USER_VERBS_CMD_REG_MR) |
2652 (1ull << IB_USER_VERBS_CMD_REREG_MR) |
2653 (1ull << IB_USER_VERBS_CMD_DEREG_MR) |
2654 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
2655 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) |
2656 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) |
2657 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) |
2658 (1ull << IB_USER_VERBS_CMD_CREATE_QP) |
2659 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) |
2660 (1ull << IB_USER_VERBS_CMD_QUERY_QP) |
2661 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) |
2662 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) |
2663 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST) |
2664 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) |
2665 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) |
2666 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) |
2667 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) |
2668 (1ull << IB_USER_VERBS_CMD_CREATE_XSRQ) |
2669 (1ull << IB_USER_VERBS_CMD_OPEN_QP);
2671 ibdev->ib_dev.query_device = mlx4_ib_query_device;
2672 ibdev->ib_dev.query_port = mlx4_ib_query_port;
2673 ibdev->ib_dev.get_link_layer = mlx4_ib_port_link_layer;
2674 ibdev->ib_dev.query_gid = mlx4_ib_query_gid;
2675 ibdev->ib_dev.query_pkey = mlx4_ib_query_pkey;
2676 ibdev->ib_dev.modify_device = mlx4_ib_modify_device;
2677 ibdev->ib_dev.modify_port = mlx4_ib_modify_port;
2678 ibdev->ib_dev.alloc_ucontext = mlx4_ib_alloc_ucontext;
2679 ibdev->ib_dev.dealloc_ucontext = mlx4_ib_dealloc_ucontext;
2680 ibdev->ib_dev.mmap = mlx4_ib_mmap;
2681 ibdev->ib_dev.alloc_pd = mlx4_ib_alloc_pd;
2682 ibdev->ib_dev.dealloc_pd = mlx4_ib_dealloc_pd;
2683 ibdev->ib_dev.create_ah = mlx4_ib_create_ah;
2684 ibdev->ib_dev.query_ah = mlx4_ib_query_ah;
2685 ibdev->ib_dev.destroy_ah = mlx4_ib_destroy_ah;
2686 ibdev->ib_dev.create_srq = mlx4_ib_create_srq;
2687 ibdev->ib_dev.modify_srq = mlx4_ib_modify_srq;
2688 ibdev->ib_dev.query_srq = mlx4_ib_query_srq;
2689 ibdev->ib_dev.destroy_srq = mlx4_ib_destroy_srq;
2690 ibdev->ib_dev.post_srq_recv = mlx4_ib_post_srq_recv;
2691 ibdev->ib_dev.create_qp = mlx4_ib_create_qp;
2692 ibdev->ib_dev.modify_qp = mlx4_ib_modify_qp;
2693 ibdev->ib_dev.query_qp = mlx4_ib_query_qp;
2694 ibdev->ib_dev.destroy_qp = mlx4_ib_destroy_qp;
2695 ibdev->ib_dev.post_send = mlx4_ib_post_send;
2696 ibdev->ib_dev.post_recv = mlx4_ib_post_recv;
2697 ibdev->ib_dev.create_cq = mlx4_ib_create_cq;
2698 ibdev->ib_dev.modify_cq = mlx4_ib_modify_cq;
2699 ibdev->ib_dev.resize_cq = mlx4_ib_resize_cq;
2700 ibdev->ib_dev.destroy_cq = mlx4_ib_destroy_cq;
2701 ibdev->ib_dev.poll_cq = mlx4_ib_poll_cq;
2702 ibdev->ib_dev.req_notify_cq = mlx4_ib_arm_cq;
2703 ibdev->ib_dev.get_dma_mr = mlx4_ib_get_dma_mr;
2704 ibdev->ib_dev.reg_user_mr = mlx4_ib_reg_user_mr;
2705 ibdev->ib_dev.rereg_user_mr = mlx4_ib_rereg_user_mr;
2706 ibdev->ib_dev.dereg_mr = mlx4_ib_dereg_mr;
2707 ibdev->ib_dev.alloc_mr = mlx4_ib_alloc_mr;
2708 ibdev->ib_dev.map_mr_sg = mlx4_ib_map_mr_sg;
2709 ibdev->ib_dev.attach_mcast = mlx4_ib_mcg_attach;
2710 ibdev->ib_dev.detach_mcast = mlx4_ib_mcg_detach;
2711 ibdev->ib_dev.process_mad = mlx4_ib_process_mad;
2712 ibdev->ib_dev.get_port_immutable = mlx4_port_immutable;
2713 ibdev->ib_dev.get_dev_fw_str = get_fw_ver_str;
2714 ibdev->ib_dev.disassociate_ucontext = mlx4_ib_disassociate_ucontext;
2716 if (!mlx4_is_slave(ibdev->dev)) {
2717 ibdev->ib_dev.alloc_fmr = mlx4_ib_fmr_alloc;
2718 ibdev->ib_dev.map_phys_fmr = mlx4_ib_map_phys_fmr;
2719 ibdev->ib_dev.unmap_fmr = mlx4_ib_unmap_fmr;
2720 ibdev->ib_dev.dealloc_fmr = mlx4_ib_fmr_dealloc;
2723 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW ||
2724 dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN) {
2725 ibdev->ib_dev.alloc_mw = mlx4_ib_alloc_mw;
2726 ibdev->ib_dev.dealloc_mw = mlx4_ib_dealloc_mw;
2728 ibdev->ib_dev.uverbs_cmd_mask |=
2729 (1ull << IB_USER_VERBS_CMD_ALLOC_MW) |
2730 (1ull << IB_USER_VERBS_CMD_DEALLOC_MW);
2733 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) {
2734 ibdev->ib_dev.alloc_xrcd = mlx4_ib_alloc_xrcd;
2735 ibdev->ib_dev.dealloc_xrcd = mlx4_ib_dealloc_xrcd;
2736 ibdev->ib_dev.uverbs_cmd_mask |=
2737 (1ull << IB_USER_VERBS_CMD_OPEN_XRCD) |
2738 (1ull << IB_USER_VERBS_CMD_CLOSE_XRCD);
2741 if (check_flow_steering_support(dev)) {
2742 ibdev->steering_support = MLX4_STEERING_MODE_DEVICE_MANAGED;
2743 ibdev->ib_dev.create_flow = mlx4_ib_create_flow;
2744 ibdev->ib_dev.destroy_flow = mlx4_ib_destroy_flow;
2746 ibdev->ib_dev.uverbs_ex_cmd_mask |=
2747 (1ull << IB_USER_VERBS_EX_CMD_CREATE_FLOW) |
2748 (1ull << IB_USER_VERBS_EX_CMD_DESTROY_FLOW);
2751 ibdev->ib_dev.uverbs_ex_cmd_mask |=
2752 (1ull << IB_USER_VERBS_EX_CMD_QUERY_DEVICE) |
2753 (1ull << IB_USER_VERBS_EX_CMD_CREATE_CQ) |
2754 (1ull << IB_USER_VERBS_EX_CMD_CREATE_QP);
2756 mlx4_ib_alloc_eqs(dev, ibdev);
2758 spin_lock_init(&iboe->lock);
2760 if (init_node_data(ibdev))
2762 mlx4_init_sl2vl_tbl(ibdev);
2764 for (i = 0; i < ibdev->num_ports; ++i) {
2765 mutex_init(&ibdev->counters_table[i].mutex);
2766 INIT_LIST_HEAD(&ibdev->counters_table[i].counters_list);
2769 num_req_counters = mlx4_is_bonded(dev) ? 1 : ibdev->num_ports;
2770 for (i = 0; i < num_req_counters; ++i) {
2771 mutex_init(&ibdev->qp1_proxy_lock[i]);
2773 if (mlx4_ib_port_link_layer(&ibdev->ib_dev, i + 1) ==
2774 IB_LINK_LAYER_ETHERNET) {
2775 err = mlx4_counter_alloc(ibdev->dev, &counter_index);
2776 /* if failed to allocate a new counter, use default */
2779 mlx4_get_default_counter_index(dev,
2783 } else { /* IB_LINK_LAYER_INFINIBAND use the default counter */
2784 counter_index = mlx4_get_default_counter_index(dev,
2787 new_counter_index = kmalloc(sizeof(*new_counter_index),
2789 if (!new_counter_index) {
2791 mlx4_counter_free(ibdev->dev, counter_index);
2794 new_counter_index->index = counter_index;
2795 new_counter_index->allocated = allocated;
2796 list_add_tail(&new_counter_index->list,
2797 &ibdev->counters_table[i].counters_list);
2798 ibdev->counters_table[i].default_counter = counter_index;
2799 pr_info("counter index %d for port %d allocated %d\n",
2800 counter_index, i + 1, allocated);
2802 if (mlx4_is_bonded(dev))
2803 for (i = 1; i < ibdev->num_ports ; ++i) {
2805 kmalloc(sizeof(struct counter_index),
2807 if (!new_counter_index)
2809 new_counter_index->index = counter_index;
2810 new_counter_index->allocated = 0;
2811 list_add_tail(&new_counter_index->list,
2812 &ibdev->counters_table[i].counters_list);
2813 ibdev->counters_table[i].default_counter =
2817 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
2820 spin_lock_init(&ibdev->sm_lock);
2821 mutex_init(&ibdev->cap_mask_mutex);
2822 INIT_LIST_HEAD(&ibdev->qp_list);
2823 spin_lock_init(&ibdev->reset_flow_resource_lock);
2825 if (ibdev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED &&
2827 ibdev->steer_qpn_count = MLX4_IB_UC_MAX_NUM_QPS;
2828 err = mlx4_qp_reserve_range(dev, ibdev->steer_qpn_count,
2829 MLX4_IB_UC_STEER_QPN_ALIGN,
2830 &ibdev->steer_qpn_base, 0);
2834 ibdev->ib_uc_qpns_bitmap =
2835 kmalloc(BITS_TO_LONGS(ibdev->steer_qpn_count) *
2838 if (!ibdev->ib_uc_qpns_bitmap)
2839 goto err_steer_qp_release;
2841 if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_DMFS_IPOIB) {
2842 bitmap_zero(ibdev->ib_uc_qpns_bitmap,
2843 ibdev->steer_qpn_count);
2844 err = mlx4_FLOW_STEERING_IB_UC_QP_RANGE(
2845 dev, ibdev->steer_qpn_base,
2846 ibdev->steer_qpn_base +
2847 ibdev->steer_qpn_count - 1);
2849 goto err_steer_free_bitmap;
2851 bitmap_fill(ibdev->ib_uc_qpns_bitmap,
2852 ibdev->steer_qpn_count);
2856 for (j = 1; j <= ibdev->dev->caps.num_ports; j++)
2857 atomic64_set(&iboe->mac[j - 1], ibdev->dev->caps.def_mac[j]);
2859 if (mlx4_ib_alloc_diag_counters(ibdev))
2860 goto err_steer_free_bitmap;
2862 if (ib_register_device(&ibdev->ib_dev, NULL))
2863 goto err_diag_counters;
2865 if (mlx4_ib_mad_init(ibdev))
2868 if (mlx4_ib_init_sriov(ibdev))
2871 if (!iboe->nb.notifier_call) {
2872 iboe->nb.notifier_call = mlx4_ib_netdev_event;
2873 err = register_netdevice_notifier(&iboe->nb);
2875 iboe->nb.notifier_call = NULL;
2879 if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ROCE_V1_V2) {
2880 err = mlx4_config_roce_v2_port(dev, ROCE_V2_UDP_DPORT);
2885 for (j = 0; j < ARRAY_SIZE(mlx4_class_attributes); ++j) {
2886 if (device_create_file(&ibdev->ib_dev.dev,
2887 mlx4_class_attributes[j]))
2891 ibdev->ib_active = true;
2892 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
2893 devlink_port_type_ib_set(mlx4_get_devlink_port(dev, i),
2896 if (mlx4_is_mfunc(ibdev->dev))
2899 /* create paravirt contexts for any VFs which are active */
2900 if (mlx4_is_master(ibdev->dev)) {
2901 for (j = 0; j < MLX4_MFUNC_MAX; j++) {
2902 if (j == mlx4_master_func_num(ibdev->dev))
2904 if (mlx4_is_slave_active(ibdev->dev, j))
2905 do_slave_init(ibdev, j, 1);
2911 if (ibdev->iboe.nb.notifier_call) {
2912 if (unregister_netdevice_notifier(&ibdev->iboe.nb))
2913 pr_warn("failure unregistering notifier\n");
2914 ibdev->iboe.nb.notifier_call = NULL;
2916 flush_workqueue(wq);
2918 mlx4_ib_close_sriov(ibdev);
2921 mlx4_ib_mad_cleanup(ibdev);
2924 ib_unregister_device(&ibdev->ib_dev);
2927 mlx4_ib_diag_cleanup(ibdev);
2929 err_steer_free_bitmap:
2930 kfree(ibdev->ib_uc_qpns_bitmap);
2932 err_steer_qp_release:
2933 if (ibdev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED)
2934 mlx4_qp_release_range(dev, ibdev->steer_qpn_base,
2935 ibdev->steer_qpn_count);
2937 for (i = 0; i < ibdev->num_ports; ++i)
2938 mlx4_ib_delete_counters_table(ibdev, &ibdev->counters_table[i]);
2941 mlx4_ib_free_eqs(dev, ibdev);
2942 iounmap(ibdev->uar_map);
2945 mlx4_uar_free(dev, &ibdev->priv_uar);
2948 mlx4_pd_free(dev, ibdev->priv_pdn);
2951 ib_dealloc_device(&ibdev->ib_dev);
2956 int mlx4_ib_steer_qp_alloc(struct mlx4_ib_dev *dev, int count, int *qpn)
2960 WARN_ON(!dev->ib_uc_qpns_bitmap);
2962 offset = bitmap_find_free_region(dev->ib_uc_qpns_bitmap,
2963 dev->steer_qpn_count,
2964 get_count_order(count));
2968 *qpn = dev->steer_qpn_base + offset;
2972 void mlx4_ib_steer_qp_free(struct mlx4_ib_dev *dev, u32 qpn, int count)
2975 dev->steering_support != MLX4_STEERING_MODE_DEVICE_MANAGED)
2978 BUG_ON(qpn < dev->steer_qpn_base);
2980 bitmap_release_region(dev->ib_uc_qpns_bitmap,
2981 qpn - dev->steer_qpn_base,
2982 get_count_order(count));
2985 int mlx4_ib_steer_qp_reg(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
2990 struct ib_flow_attr *flow = NULL;
2991 struct ib_flow_spec_ib *ib_spec;
2994 flow_size = sizeof(struct ib_flow_attr) +
2995 sizeof(struct ib_flow_spec_ib);
2996 flow = kzalloc(flow_size, GFP_KERNEL);
2999 flow->port = mqp->port;
3000 flow->num_of_specs = 1;
3001 flow->size = flow_size;
3002 ib_spec = (struct ib_flow_spec_ib *)(flow + 1);
3003 ib_spec->type = IB_FLOW_SPEC_IB;
3004 ib_spec->size = sizeof(struct ib_flow_spec_ib);
3005 /* Add an empty rule for IB L2 */
3006 memset(&ib_spec->mask, 0, sizeof(ib_spec->mask));
3008 err = __mlx4_ib_create_flow(&mqp->ibqp, flow,
3013 err = __mlx4_ib_destroy_flow(mdev->dev, mqp->reg_id);
3019 static void mlx4_ib_remove(struct mlx4_dev *dev, void *ibdev_ptr)
3021 struct mlx4_ib_dev *ibdev = ibdev_ptr;
3025 mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_IB)
3026 devlink_port_type_clear(mlx4_get_devlink_port(dev, i));
3027 ibdev->ib_active = false;
3028 flush_workqueue(wq);
3030 mlx4_ib_close_sriov(ibdev);
3031 mlx4_ib_mad_cleanup(ibdev);
3032 ib_unregister_device(&ibdev->ib_dev);
3033 mlx4_ib_diag_cleanup(ibdev);
3034 if (ibdev->iboe.nb.notifier_call) {
3035 if (unregister_netdevice_notifier(&ibdev->iboe.nb))
3036 pr_warn("failure unregistering notifier\n");
3037 ibdev->iboe.nb.notifier_call = NULL;
3040 if (ibdev->steering_support == MLX4_STEERING_MODE_DEVICE_MANAGED) {
3041 mlx4_qp_release_range(dev, ibdev->steer_qpn_base,
3042 ibdev->steer_qpn_count);
3043 kfree(ibdev->ib_uc_qpns_bitmap);
3046 iounmap(ibdev->uar_map);
3047 for (p = 0; p < ibdev->num_ports; ++p)
3048 mlx4_ib_delete_counters_table(ibdev, &ibdev->counters_table[p]);
3050 mlx4_foreach_port(p, dev, MLX4_PORT_TYPE_IB)
3051 mlx4_CLOSE_PORT(dev, p);
3053 mlx4_ib_free_eqs(dev, ibdev);
3055 mlx4_uar_free(dev, &ibdev->priv_uar);
3056 mlx4_pd_free(dev, ibdev->priv_pdn);
3057 ib_dealloc_device(&ibdev->ib_dev);
3060 static void do_slave_init(struct mlx4_ib_dev *ibdev, int slave, int do_init)
3062 struct mlx4_ib_demux_work **dm = NULL;
3063 struct mlx4_dev *dev = ibdev->dev;
3065 unsigned long flags;
3066 struct mlx4_active_ports actv_ports;
3068 unsigned int first_port;
3070 if (!mlx4_is_master(dev))
3073 actv_ports = mlx4_get_active_ports(dev, slave);
3074 ports = bitmap_weight(actv_ports.ports, dev->caps.num_ports);
3075 first_port = find_first_bit(actv_ports.ports, dev->caps.num_ports);
3077 dm = kcalloc(ports, sizeof(*dm), GFP_ATOMIC);
3081 for (i = 0; i < ports; i++) {
3082 dm[i] = kmalloc(sizeof (struct mlx4_ib_demux_work), GFP_ATOMIC);
3088 INIT_WORK(&dm[i]->work, mlx4_ib_tunnels_update_work);
3089 dm[i]->port = first_port + i + 1;
3090 dm[i]->slave = slave;
3091 dm[i]->do_init = do_init;
3094 /* initialize or tear down tunnel QPs for the slave */
3095 spin_lock_irqsave(&ibdev->sriov.going_down_lock, flags);
3096 if (!ibdev->sriov.is_going_down) {
3097 for (i = 0; i < ports; i++)
3098 queue_work(ibdev->sriov.demux[i].ud_wq, &dm[i]->work);
3099 spin_unlock_irqrestore(&ibdev->sriov.going_down_lock, flags);
3101 spin_unlock_irqrestore(&ibdev->sriov.going_down_lock, flags);
3102 for (i = 0; i < ports; i++)
3110 static void mlx4_ib_handle_catas_error(struct mlx4_ib_dev *ibdev)
3112 struct mlx4_ib_qp *mqp;
3113 unsigned long flags_qp;
3114 unsigned long flags_cq;
3115 struct mlx4_ib_cq *send_mcq, *recv_mcq;
3116 struct list_head cq_notify_list;
3117 struct mlx4_cq *mcq;
3118 unsigned long flags;
3120 pr_warn("mlx4_ib_handle_catas_error was started\n");
3121 INIT_LIST_HEAD(&cq_notify_list);
3123 /* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
3124 spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
3126 list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
3127 spin_lock_irqsave(&mqp->sq.lock, flags_qp);
3128 if (mqp->sq.tail != mqp->sq.head) {
3129 send_mcq = to_mcq(mqp->ibqp.send_cq);
3130 spin_lock_irqsave(&send_mcq->lock, flags_cq);
3131 if (send_mcq->mcq.comp &&
3132 mqp->ibqp.send_cq->comp_handler) {
3133 if (!send_mcq->mcq.reset_notify_added) {
3134 send_mcq->mcq.reset_notify_added = 1;
3135 list_add_tail(&send_mcq->mcq.reset_notify,
3139 spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
3141 spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
3142 /* Now, handle the QP's receive queue */
3143 spin_lock_irqsave(&mqp->rq.lock, flags_qp);
3144 /* no handling is needed for SRQ */
3145 if (!mqp->ibqp.srq) {
3146 if (mqp->rq.tail != mqp->rq.head) {
3147 recv_mcq = to_mcq(mqp->ibqp.recv_cq);
3148 spin_lock_irqsave(&recv_mcq->lock, flags_cq);
3149 if (recv_mcq->mcq.comp &&
3150 mqp->ibqp.recv_cq->comp_handler) {
3151 if (!recv_mcq->mcq.reset_notify_added) {
3152 recv_mcq->mcq.reset_notify_added = 1;
3153 list_add_tail(&recv_mcq->mcq.reset_notify,
3157 spin_unlock_irqrestore(&recv_mcq->lock,
3161 spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
3164 list_for_each_entry(mcq, &cq_notify_list, reset_notify) {
3167 spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
3168 pr_warn("mlx4_ib_handle_catas_error ended\n");
3171 static void handle_bonded_port_state_event(struct work_struct *work)
3173 struct ib_event_work *ew =
3174 container_of(work, struct ib_event_work, work);
3175 struct mlx4_ib_dev *ibdev = ew->ib_dev;
3176 enum ib_port_state bonded_port_state = IB_PORT_NOP;
3178 struct ib_event ibev;
3181 spin_lock_bh(&ibdev->iboe.lock);
3182 for (i = 0; i < MLX4_MAX_PORTS; ++i) {
3183 struct net_device *curr_netdev = ibdev->iboe.netdevs[i];
3184 enum ib_port_state curr_port_state;
3190 (netif_running(curr_netdev) &&
3191 netif_carrier_ok(curr_netdev)) ?
3192 IB_PORT_ACTIVE : IB_PORT_DOWN;
3194 bonded_port_state = (bonded_port_state != IB_PORT_ACTIVE) ?
3195 curr_port_state : IB_PORT_ACTIVE;
3197 spin_unlock_bh(&ibdev->iboe.lock);
3199 ibev.device = &ibdev->ib_dev;
3200 ibev.element.port_num = 1;
3201 ibev.event = (bonded_port_state == IB_PORT_ACTIVE) ?
3202 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
3204 ib_dispatch_event(&ibev);
3207 void mlx4_ib_sl2vl_update(struct mlx4_ib_dev *mdev, int port)
3212 err = mlx4_ib_query_sl2vl(&mdev->ib_dev, port, &sl2vl);
3214 pr_err("Unable to get current sl to vl mapping for port %d. Using all zeroes (%d)\n",
3218 atomic64_set(&mdev->sl2vl[port - 1], sl2vl);
3221 static void ib_sl2vl_update_work(struct work_struct *work)
3223 struct ib_event_work *ew = container_of(work, struct ib_event_work, work);
3224 struct mlx4_ib_dev *mdev = ew->ib_dev;
3225 int port = ew->port;
3227 mlx4_ib_sl2vl_update(mdev, port);
3232 void mlx4_sched_ib_sl2vl_update_work(struct mlx4_ib_dev *ibdev,
3235 struct ib_event_work *ew;
3237 ew = kmalloc(sizeof(*ew), GFP_ATOMIC);
3239 INIT_WORK(&ew->work, ib_sl2vl_update_work);
3242 queue_work(wq, &ew->work);
3246 static void mlx4_ib_event(struct mlx4_dev *dev, void *ibdev_ptr,
3247 enum mlx4_dev_event event, unsigned long param)
3249 struct ib_event ibev;
3250 struct mlx4_ib_dev *ibdev = to_mdev((struct ib_device *) ibdev_ptr);
3251 struct mlx4_eqe *eqe = NULL;
3252 struct ib_event_work *ew;
3255 if (mlx4_is_bonded(dev) &&
3256 ((event == MLX4_DEV_EVENT_PORT_UP) ||
3257 (event == MLX4_DEV_EVENT_PORT_DOWN))) {
3258 ew = kmalloc(sizeof(*ew), GFP_ATOMIC);
3261 INIT_WORK(&ew->work, handle_bonded_port_state_event);
3263 queue_work(wq, &ew->work);
3267 if (event == MLX4_DEV_EVENT_PORT_MGMT_CHANGE)
3268 eqe = (struct mlx4_eqe *)param;
3273 case MLX4_DEV_EVENT_PORT_UP:
3274 if (p > ibdev->num_ports)
3276 if (!mlx4_is_slave(dev) &&
3277 rdma_port_get_link_layer(&ibdev->ib_dev, p) ==
3278 IB_LINK_LAYER_INFINIBAND) {
3279 if (mlx4_is_master(dev))
3280 mlx4_ib_invalidate_all_guid_record(ibdev, p);
3281 if (ibdev->dev->flags & MLX4_FLAG_SECURE_HOST &&
3282 !(ibdev->dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SL_TO_VL_CHANGE_EVENT))
3283 mlx4_sched_ib_sl2vl_update_work(ibdev, p);
3285 ibev.event = IB_EVENT_PORT_ACTIVE;
3288 case MLX4_DEV_EVENT_PORT_DOWN:
3289 if (p > ibdev->num_ports)
3291 ibev.event = IB_EVENT_PORT_ERR;
3294 case MLX4_DEV_EVENT_CATASTROPHIC_ERROR:
3295 ibdev->ib_active = false;
3296 ibev.event = IB_EVENT_DEVICE_FATAL;
3297 mlx4_ib_handle_catas_error(ibdev);
3300 case MLX4_DEV_EVENT_PORT_MGMT_CHANGE:
3301 ew = kmalloc(sizeof *ew, GFP_ATOMIC);
3305 INIT_WORK(&ew->work, handle_port_mgmt_change_event);
3306 memcpy(&ew->ib_eqe, eqe, sizeof *eqe);
3308 /* need to queue only for port owner, which uses GEN_EQE */
3309 if (mlx4_is_master(dev))
3310 queue_work(wq, &ew->work);
3312 handle_port_mgmt_change_event(&ew->work);
3315 case MLX4_DEV_EVENT_SLAVE_INIT:
3316 /* here, p is the slave id */
3317 do_slave_init(ibdev, p, 1);
3318 if (mlx4_is_master(dev)) {
3321 for (i = 1; i <= ibdev->num_ports; i++) {
3322 if (rdma_port_get_link_layer(&ibdev->ib_dev, i)
3323 == IB_LINK_LAYER_INFINIBAND)
3324 mlx4_ib_slave_alias_guid_event(ibdev,
3331 case MLX4_DEV_EVENT_SLAVE_SHUTDOWN:
3332 if (mlx4_is_master(dev)) {
3335 for (i = 1; i <= ibdev->num_ports; i++) {
3336 if (rdma_port_get_link_layer(&ibdev->ib_dev, i)
3337 == IB_LINK_LAYER_INFINIBAND)
3338 mlx4_ib_slave_alias_guid_event(ibdev,
3343 /* here, p is the slave id */
3344 do_slave_init(ibdev, p, 0);
3351 ibev.device = ibdev_ptr;
3352 ibev.element.port_num = mlx4_is_bonded(ibdev->dev) ? 1 : (u8)p;
3354 ib_dispatch_event(&ibev);
3357 static struct mlx4_interface mlx4_ib_interface = {
3359 .remove = mlx4_ib_remove,
3360 .event = mlx4_ib_event,
3361 .protocol = MLX4_PROT_IB_IPV6,
3362 .flags = MLX4_INTFF_BONDING
3365 static int __init mlx4_ib_init(void)
3369 wq = alloc_ordered_workqueue("mlx4_ib", WQ_MEM_RECLAIM);
3373 err = mlx4_ib_mcg_init();
3377 err = mlx4_register_interface(&mlx4_ib_interface);
3384 mlx4_ib_mcg_destroy();
3387 destroy_workqueue(wq);
3391 static void __exit mlx4_ib_cleanup(void)
3393 mlx4_unregister_interface(&mlx4_ib_interface);
3394 mlx4_ib_mcg_destroy();
3395 destroy_workqueue(wq);
3398 module_init(mlx4_ib_init);
3399 module_exit(mlx4_ib_cleanup);