drivers/infiniband/hw/ocrdma/ocrdma_main.c

   1 /* This file is part of the Emulex RoCE Device Driver for
   2  * RoCE (RDMA over Converged Ethernet) adapters.
   3  * Copyright (C) 2012-2015 Emulex. All rights reserved.
   4  * EMULEX and SLI are trademarks of Emulex.
   5  * www.emulex.com
   6  *
   7  * This software is available to you under a choice of one of two licenses.
   8  * You may choose to be licensed under the terms of the GNU General Public
   9  * License (GPL) Version 2, available from the file COPYING in the main
  10  * directory of this source tree, or the BSD license below:
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  *
  16  * - Redistributions of source code must retain the above copyright notice,
  17  *   this list of conditions and the following disclaimer.
  18  *
  19  * - Redistributions in binary form must reproduce the above copyright
  20  *   notice, this list of conditions and the following disclaimer in
  21  *   the documentation and/or other materials provided with the distribution.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  24  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,THE
  25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  26  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
  27  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  28  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  29  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  30  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  31  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  32  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
  33  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  34  *
  35  * Contact Information:
  36  * linux-drivers@emulex.com
  37  *
  38  * Emulex
  39  * 3333 Susan Street
  40  * Costa Mesa, CA 92626
  41  */
  42
  43 #include <linux/module.h>
  44 #include <linux/idr.h>
  45 #include <rdma/ib_verbs.h>
  46 #include <rdma/ib_user_verbs.h>
  47 #include <rdma/ib_addr.h>
  48 #include <rdma/ib_mad.h>
  49
  50 #include <linux/netdevice.h>
  51 #include <net/addrconf.h>
  52
  53 #include "ocrdma.h"
  54 #include "ocrdma_verbs.h"
  55 #include "ocrdma_ah.h"
  56 #include "be_roce.h"
  57 #include "ocrdma_hw.h"
  58 #include "ocrdma_stats.h"
  59 #include <rdma/ocrdma-abi.h>
  60
  61 MODULE_VERSION(OCRDMA_ROCE_DRV_VERSION);
  62 MODULE_DESCRIPTION(OCRDMA_ROCE_DRV_DESC " " OCRDMA_ROCE_DRV_VERSION);
  63 MODULE_AUTHOR("Emulex Corporation");
  64 MODULE_LICENSE("Dual BSD/GPL");
  65
  66 static DEFINE_IDR(ocrdma_dev_id);
  67
  68 void ocrdma_get_guid(struct ocrdma_dev *dev, u8 *guid)
  69 {
  70         u8 mac_addr[6];
  71
  72         memcpy(&mac_addr[0], &dev->nic_info.mac_addr[0], ETH_ALEN);
  73         guid[0] = mac_addr[0] ^ 2;
  74         guid[1] = mac_addr[1];
  75         guid[2] = mac_addr[2];
  76         guid[3] = 0xff;
  77         guid[4] = 0xfe;
  78         guid[5] = mac_addr[3];
  79         guid[6] = mac_addr[4];
  80         guid[7] = mac_addr[5];
  81 }
  82 static enum rdma_link_layer ocrdma_link_layer(struct ib_device *device,
  83                                               u8 port_num)
  84 {
  85         return IB_LINK_LAYER_ETHERNET;
  86 }
  87
  88 static int ocrdma_port_immutable(struct ib_device *ibdev, u8 port_num,
  89                                  struct ib_port_immutable *immutable)
  90 {
  91         struct ib_port_attr attr;
  92         struct ocrdma_dev *dev;
  93         int err;
  94
  95         dev = get_ocrdma_dev(ibdev);
  96         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE;
  97         if (ocrdma_is_udp_encap_supported(dev))
  98                 immutable->core_cap_flags |= RDMA_CORE_CAP_PROT_ROCE_UDP_ENCAP;
  99
 100         err = ib_query_port(ibdev, port_num, &attr);
 101         if (err)
 102                 return err;
 103
 104         immutable->pkey_tbl_len = attr.pkey_tbl_len;
 105         immutable->gid_tbl_len = attr.gid_tbl_len;
 106         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
 107
 108         return 0;
 109 }
 110
 111 static void get_dev_fw_str(struct ib_device *device, char *str,
 112                            size_t str_len)
 113 {
 114         struct ocrdma_dev *dev = get_ocrdma_dev(device);
 115
 116         snprintf(str, str_len, "%s", &dev->attr.fw_ver[0]);
 117 }
 118
 119 static int ocrdma_register_device(struct ocrdma_dev *dev)
 120 {
 121         strlcpy(dev->ibdev.name, "ocrdma%d", IB_DEVICE_NAME_MAX);
 122         ocrdma_get_guid(dev, (u8 *)&dev->ibdev.node_guid);
 123         BUILD_BUG_ON(sizeof(OCRDMA_NODE_DESC) > IB_DEVICE_NODE_DESC_MAX);
 124         memcpy(dev->ibdev.node_desc, OCRDMA_NODE_DESC,
 125                sizeof(OCRDMA_NODE_DESC));
 126         dev->ibdev.owner = THIS_MODULE;
 127         dev->ibdev.uverbs_abi_ver = OCRDMA_ABI_VERSION;
 128         dev->ibdev.uverbs_cmd_mask =
 129             OCRDMA_UVERBS(GET_CONTEXT) |
 130             OCRDMA_UVERBS(QUERY_DEVICE) |
 131             OCRDMA_UVERBS(QUERY_PORT) |
 132             OCRDMA_UVERBS(ALLOC_PD) |
 133             OCRDMA_UVERBS(DEALLOC_PD) |
 134             OCRDMA_UVERBS(REG_MR) |
 135             OCRDMA_UVERBS(DEREG_MR) |
 136             OCRDMA_UVERBS(CREATE_COMP_CHANNEL) |
 137             OCRDMA_UVERBS(CREATE_CQ) |
 138             OCRDMA_UVERBS(RESIZE_CQ) |
 139             OCRDMA_UVERBS(DESTROY_CQ) |
 140             OCRDMA_UVERBS(REQ_NOTIFY_CQ) |
 141             OCRDMA_UVERBS(CREATE_QP) |
 142             OCRDMA_UVERBS(MODIFY_QP) |
 143             OCRDMA_UVERBS(QUERY_QP) |
 144             OCRDMA_UVERBS(DESTROY_QP) |
 145             OCRDMA_UVERBS(POLL_CQ) |
 146             OCRDMA_UVERBS(POST_SEND) |
 147             OCRDMA_UVERBS(POST_RECV);
 148
 149         dev->ibdev.uverbs_cmd_mask |=
 150             OCRDMA_UVERBS(CREATE_AH) |
 151              OCRDMA_UVERBS(MODIFY_AH) |
 152              OCRDMA_UVERBS(QUERY_AH) |
 153              OCRDMA_UVERBS(DESTROY_AH);
 154
 155         dev->ibdev.node_type = RDMA_NODE_IB_CA;
 156         dev->ibdev.phys_port_cnt = 1;
 157         dev->ibdev.num_comp_vectors = dev->eq_cnt;
 158
 159         /* mandatory verbs. */
 160         dev->ibdev.query_device = ocrdma_query_device;
 161         dev->ibdev.query_port = ocrdma_query_port;
 162         dev->ibdev.modify_port = ocrdma_modify_port;
 163         dev->ibdev.query_gid = ocrdma_query_gid;
 164         dev->ibdev.get_netdev = ocrdma_get_netdev;
 165         dev->ibdev.add_gid = ocrdma_add_gid;
 166         dev->ibdev.del_gid = ocrdma_del_gid;
 167         dev->ibdev.get_link_layer = ocrdma_link_layer;
 168         dev->ibdev.alloc_pd = ocrdma_alloc_pd;
 169         dev->ibdev.dealloc_pd = ocrdma_dealloc_pd;
 170
 171         dev->ibdev.create_cq = ocrdma_create_cq;
 172         dev->ibdev.destroy_cq = ocrdma_destroy_cq;
 173         dev->ibdev.resize_cq = ocrdma_resize_cq;
 174
 175         dev->ibdev.create_qp = ocrdma_create_qp;
 176         dev->ibdev.modify_qp = ocrdma_modify_qp;
 177         dev->ibdev.query_qp = ocrdma_query_qp;
 178         dev->ibdev.destroy_qp = ocrdma_destroy_qp;
 179
 180         dev->ibdev.query_pkey = ocrdma_query_pkey;
 181         dev->ibdev.create_ah = ocrdma_create_ah;
 182         dev->ibdev.destroy_ah = ocrdma_destroy_ah;
 183         dev->ibdev.query_ah = ocrdma_query_ah;
 184         dev->ibdev.modify_ah = ocrdma_modify_ah;
 185
 186         dev->ibdev.poll_cq = ocrdma_poll_cq;
 187         dev->ibdev.post_send = ocrdma_post_send;
 188         dev->ibdev.post_recv = ocrdma_post_recv;
 189         dev->ibdev.req_notify_cq = ocrdma_arm_cq;
 190
 191         dev->ibdev.get_dma_mr = ocrdma_get_dma_mr;
 192         dev->ibdev.dereg_mr = ocrdma_dereg_mr;
 193         dev->ibdev.reg_user_mr = ocrdma_reg_user_mr;
 194
 195         dev->ibdev.alloc_mr = ocrdma_alloc_mr;
 196         dev->ibdev.map_mr_sg = ocrdma_map_mr_sg;
 197
 198         /* mandatory to support user space verbs consumer. */
 199         dev->ibdev.alloc_ucontext = ocrdma_alloc_ucontext;
 200         dev->ibdev.dealloc_ucontext = ocrdma_dealloc_ucontext;
 201         dev->ibdev.mmap = ocrdma_mmap;
 202         dev->ibdev.dev.parent = &dev->nic_info.pdev->dev;
 203
 204         dev->ibdev.process_mad = ocrdma_process_mad;
 205         dev->ibdev.get_port_immutable = ocrdma_port_immutable;
 206         dev->ibdev.get_dev_fw_str     = get_dev_fw_str;
 207
 208         if (ocrdma_get_asic_type(dev) == OCRDMA_ASIC_GEN_SKH_R) {
 209                 dev->ibdev.uverbs_cmd_mask |=
 210                      OCRDMA_UVERBS(CREATE_SRQ) |
 211                      OCRDMA_UVERBS(MODIFY_SRQ) |
 212                      OCRDMA_UVERBS(QUERY_SRQ) |
 213                      OCRDMA_UVERBS(DESTROY_SRQ) |
 214                      OCRDMA_UVERBS(POST_SRQ_RECV);
 215
 216                 dev->ibdev.create_srq = ocrdma_create_srq;
 217                 dev->ibdev.modify_srq = ocrdma_modify_srq;
 218                 dev->ibdev.query_srq = ocrdma_query_srq;
 219                 dev->ibdev.destroy_srq = ocrdma_destroy_srq;
 220                 dev->ibdev.post_srq_recv = ocrdma_post_srq_recv;
 221         }
 222         return ib_register_device(&dev->ibdev, NULL);
 223 }
 224
 225 static int ocrdma_alloc_resources(struct ocrdma_dev *dev)
 226 {
 227         mutex_init(&dev->dev_lock);
 228         dev->cq_tbl = kzalloc(sizeof(struct ocrdma_cq *) *
 229                               OCRDMA_MAX_CQ, GFP_KERNEL);
 230         if (!dev->cq_tbl)
 231                 goto alloc_err;
 232
 233         if (dev->attr.max_qp) {
 234                 dev->qp_tbl = kzalloc(sizeof(struct ocrdma_qp *) *
 235                                       OCRDMA_MAX_QP, GFP_KERNEL);
 236                 if (!dev->qp_tbl)
 237                         goto alloc_err;
 238         }
 239
 240         dev->stag_arr = kzalloc(sizeof(u64) * OCRDMA_MAX_STAG, GFP_KERNEL);
 241         if (dev->stag_arr == NULL)
 242                 goto alloc_err;
 243
 244         ocrdma_alloc_pd_pool(dev);
 245
 246         if (!ocrdma_alloc_stats_resources(dev)) {
 247                 pr_err("%s: stats resource allocation failed\n", __func__);
 248                 goto alloc_err;
 249         }
 250
 251         spin_lock_init(&dev->av_tbl.lock);
 252         spin_lock_init(&dev->flush_q_lock);
 253         return 0;
 254 alloc_err:
 255         pr_err("%s(%d) error.\n", __func__, dev->id);
 256         return -ENOMEM;
 257 }
 258
 259 static void ocrdma_free_resources(struct ocrdma_dev *dev)
 260 {
 261         ocrdma_release_stats_resources(dev);
 262         kfree(dev->stag_arr);
 263         kfree(dev->qp_tbl);
 264         kfree(dev->cq_tbl);
 265 }
 266
 267 /* OCRDMA sysfs interface */
 268 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
 269                         char *buf)
 270 {
 271         struct ocrdma_dev *dev = dev_get_drvdata(device);
 272
 273         return scnprintf(buf, PAGE_SIZE, "0x%x\n", dev->nic_info.pdev->vendor);
 274 }
 275
 276 static ssize_t show_hca_type(struct device *device,
 277                              struct device_attribute *attr, char *buf)
 278 {
 279         struct ocrdma_dev *dev = dev_get_drvdata(device);
 280
 281         return scnprintf(buf, PAGE_SIZE, "%s\n", &dev->model_number[0]);
 282 }
 283
 284 static DEVICE_ATTR(hw_rev, S_IRUGO, show_rev, NULL);
 285 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca_type, NULL);
 286
 287 static struct device_attribute *ocrdma_attributes[] = {
 288         &dev_attr_hw_rev,
 289         &dev_attr_hca_type
 290 };
 291
 292 static void ocrdma_remove_sysfiles(struct ocrdma_dev *dev)
 293 {
 294         int i;
 295
 296         for (i = 0; i < ARRAY_SIZE(ocrdma_attributes); i++)
 297                 device_remove_file(&dev->ibdev.dev, ocrdma_attributes[i]);
 298 }
 299
 300 static struct ocrdma_dev *ocrdma_add(struct be_dev_info *dev_info)
 301 {
 302         int status = 0, i;
 303         u8 lstate = 0;
 304         struct ocrdma_dev *dev;
 305
 306         dev = (struct ocrdma_dev *)ib_alloc_device(sizeof(struct ocrdma_dev));
 307         if (!dev) {
 308                 pr_err("Unable to allocate ib device\n");
 309                 return NULL;
 310         }
 311         dev->mbx_cmd = kzalloc(sizeof(struct ocrdma_mqe_emb_cmd), GFP_KERNEL);
 312         if (!dev->mbx_cmd)
 313                 goto idr_err;
 314
 315         memcpy(&dev->nic_info, dev_info, sizeof(*dev_info));
 316         dev->id = idr_alloc(&ocrdma_dev_id, NULL, 0, 0, GFP_KERNEL);
 317         if (dev->id < 0)
 318                 goto idr_err;
 319
 320         status = ocrdma_init_hw(dev);
 321         if (status)
 322                 goto init_err;
 323
 324         status = ocrdma_alloc_resources(dev);
 325         if (status)
 326                 goto alloc_err;
 327
 328         ocrdma_init_service_level(dev);
 329         status = ocrdma_register_device(dev);
 330         if (status)
 331                 goto alloc_err;
 332
 333         /* Query Link state and update */
 334         status = ocrdma_mbx_get_link_speed(dev, NULL, &lstate);
 335         if (!status)
 336                 ocrdma_update_link_state(dev, lstate);
 337
 338         for (i = 0; i < ARRAY_SIZE(ocrdma_attributes); i++)
 339                 if (device_create_file(&dev->ibdev.dev, ocrdma_attributes[i]))
 340                         goto sysfs_err;
 341         /* Init stats */
 342         ocrdma_add_port_stats(dev);
 343         /* Interrupt Moderation */
 344         INIT_DELAYED_WORK(&dev->eqd_work, ocrdma_eqd_set_task);
 345         schedule_delayed_work(&dev->eqd_work, msecs_to_jiffies(1000));
 346
 347         pr_info("%s %s: %s \"%s\" port %d\n",
 348                 dev_name(&dev->nic_info.pdev->dev), hca_name(dev),
 349                 port_speed_string(dev), dev->model_number,
 350                 dev->hba_port_num);
 351         pr_info("%s ocrdma%d driver loaded successfully\n",
 352                 dev_name(&dev->nic_info.pdev->dev), dev->id);
 353         return dev;
 354
 355 sysfs_err:
 356         ocrdma_remove_sysfiles(dev);
 357 alloc_err:
 358         ocrdma_free_resources(dev);
 359         ocrdma_cleanup_hw(dev);
 360 init_err:
 361         idr_remove(&ocrdma_dev_id, dev->id);
 362 idr_err:
 363         kfree(dev->mbx_cmd);
 364         ib_dealloc_device(&dev->ibdev);
 365         pr_err("%s() leaving. ret=%d\n", __func__, status);
 366         return NULL;
 367 }
 368
 369 static void ocrdma_remove_free(struct ocrdma_dev *dev)
 370 {
 371
 372         idr_remove(&ocrdma_dev_id, dev->id);
 373         kfree(dev->mbx_cmd);
 374         ib_dealloc_device(&dev->ibdev);
 375 }
 376
 377 static void ocrdma_remove(struct ocrdma_dev *dev)
 378 {
 379         /* first unregister with stack to stop all the active traffic
 380          * of the registered clients.
 381          */
 382         cancel_delayed_work_sync(&dev->eqd_work);
 383         ocrdma_remove_sysfiles(dev);
 384         ib_unregister_device(&dev->ibdev);
 385
 386         ocrdma_rem_port_stats(dev);
 387         ocrdma_free_resources(dev);
 388         ocrdma_cleanup_hw(dev);
 389         ocrdma_remove_free(dev);
 390 }
 391
 392 static int ocrdma_dispatch_port_active(struct ocrdma_dev *dev)
 393 {
 394         struct ib_event port_event;
 395
 396         port_event.event = IB_EVENT_PORT_ACTIVE;
 397         port_event.element.port_num = 1;
 398         port_event.device = &dev->ibdev;
 399         ib_dispatch_event(&port_event);
 400         return 0;
 401 }
 402
 403 static int ocrdma_dispatch_port_error(struct ocrdma_dev *dev)
 404 {
 405         struct ib_event err_event;
 406
 407         err_event.event = IB_EVENT_PORT_ERR;
 408         err_event.element.port_num = 1;
 409         err_event.device = &dev->ibdev;
 410         ib_dispatch_event(&err_event);
 411         return 0;
 412 }
 413
 414 static void ocrdma_shutdown(struct ocrdma_dev *dev)
 415 {
 416         ocrdma_dispatch_port_error(dev);
 417         ocrdma_remove(dev);
 418 }
 419
 420 /* event handling via NIC driver ensures that all the NIC specific
 421  * initialization done before RoCE driver notifies
 422  * event to stack.
 423  */
 424 static void ocrdma_event_handler(struct ocrdma_dev *dev, u32 event)
 425 {
 426         switch (event) {
 427         case BE_DEV_SHUTDOWN:
 428                 ocrdma_shutdown(dev);
 429                 break;
 430         default:
 431                 break;
 432         }
 433 }
 434
 435 void ocrdma_update_link_state(struct ocrdma_dev *dev, u8 lstate)
 436 {
 437         if (!(dev->flags & OCRDMA_FLAGS_LINK_STATUS_INIT)) {
 438                 dev->flags |= OCRDMA_FLAGS_LINK_STATUS_INIT;
 439                 if (!lstate)
 440                         return;
 441         }
 442
 443         if (!lstate)
 444                 ocrdma_dispatch_port_error(dev);
 445         else
 446                 ocrdma_dispatch_port_active(dev);
 447 }
 448
 449 static struct ocrdma_driver ocrdma_drv = {
 450         .name                   = "ocrdma_driver",
 451         .add                    = ocrdma_add,
 452         .remove                 = ocrdma_remove,
 453         .state_change_handler   = ocrdma_event_handler,
 454         .be_abi_version         = OCRDMA_BE_ROCE_ABI_VERSION,
 455 };
 456
 457 static int __init ocrdma_init_module(void)
 458 {
 459         int status;
 460
 461         ocrdma_init_debugfs();
 462
 463         status = be_roce_register_driver(&ocrdma_drv);
 464         if (status)
 465                 goto err_be_reg;
 466
 467         return 0;
 468
 469 err_be_reg:
 470
 471         return status;
 472 }
 473
 474 static void __exit ocrdma_exit_module(void)
 475 {
 476         be_roce_unregister_driver(&ocrdma_drv);
 477         ocrdma_rem_debugfs();
 478         idr_destroy(&ocrdma_dev_id);
 479 }
 480
 481 module_init(ocrdma_init_module);
 482 module_exit(ocrdma_exit_module);