drivers/infiniband/hw/mlx4/mad.c

   1 /*
   2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <rdma/ib_mad.h>
  34 #include <rdma/ib_smi.h>
  35
  36 #include <linux/mlx4/cmd.h>
  37 #include <linux/gfp.h>
  38 #include <rdma/ib_pma.h>
  39
  40 #include "mlx4_ib.h"
  41
  42 enum {
  43         MLX4_IB_VENDOR_CLASS1 = 0x9,
  44         MLX4_IB_VENDOR_CLASS2 = 0xa
  45 };
  46
  47 /* Counters should be saturate once they reach their maximum value */
  48 #define ASSIGN_32BIT_COUNTER(counter, value) do {\
  49         if ((value) > (u32)~0U)                  \
  50                 counter = cpu_to_be32((u32)~0U); \
  51         else                                     \
  52                 counter = cpu_to_be32(value);    \
  53 } while (0)
  54
  55 int mlx4_MAD_IFC(struct mlx4_ib_dev *dev, int ignore_mkey, int ignore_bkey,
  56                  int port, struct ib_wc *in_wc, struct ib_grh *in_grh,
  57                  void *in_mad, void *response_mad)
  58 {
  59         struct mlx4_cmd_mailbox *inmailbox, *outmailbox;
  60         void *inbox;
  61         int err;
  62         u32 in_modifier = port;
  63         u8 op_modifier = 0;
  64
  65         inmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
  66         if (IS_ERR(inmailbox))
  67                 return PTR_ERR(inmailbox);
  68         inbox = inmailbox->buf;
  69
  70         outmailbox = mlx4_alloc_cmd_mailbox(dev->dev);
  71         if (IS_ERR(outmailbox)) {
  72                 mlx4_free_cmd_mailbox(dev->dev, inmailbox);
  73                 return PTR_ERR(outmailbox);
  74         }
  75
  76         memcpy(inbox, in_mad, 256);
  77
  78         /*
  79          * Key check traps can't be generated unless we have in_wc to
  80          * tell us where to send the trap.
  81          */
  82         if (ignore_mkey || !in_wc)
  83                 op_modifier |= 0x1;
  84         if (ignore_bkey || !in_wc)
  85                 op_modifier |= 0x2;
  86
  87         if (in_wc) {
  88                 struct {
  89                         __be32          my_qpn;
  90                         u32             reserved1;
  91                         __be32          rqpn;
  92                         u8              sl;
  93                         u8              g_path;
  94                         u16             reserved2[2];
  95                         __be16          pkey;
  96                         u32             reserved3[11];
  97                         u8              grh[40];
  98                 } *ext_info;
  99
 100                 memset(inbox + 256, 0, 256);
 101                 ext_info = inbox + 256;
 102
 103                 ext_info->my_qpn = cpu_to_be32(in_wc->qp->qp_num);
 104                 ext_info->rqpn   = cpu_to_be32(in_wc->src_qp);
 105                 ext_info->sl     = in_wc->sl << 4;
 106                 ext_info->g_path = in_wc->dlid_path_bits |
 107                         (in_wc->wc_flags & IB_WC_GRH ? 0x80 : 0);
 108                 ext_info->pkey   = cpu_to_be16(in_wc->pkey_index);
 109
 110                 if (in_grh)
 111                         memcpy(ext_info->grh, in_grh, 40);
 112
 113                 op_modifier |= 0x4;
 114
 115                 in_modifier |= in_wc->slid << 16;
 116         }
 117
 118         err = mlx4_cmd_box(dev->dev, inmailbox->dma, outmailbox->dma,
 119                            in_modifier, op_modifier,
 120                            MLX4_CMD_MAD_IFC, MLX4_CMD_TIME_CLASS_C);
 121
 122         if (!err)
 123                 memcpy(response_mad, outmailbox->buf, 256);
 124
 125         mlx4_free_cmd_mailbox(dev->dev, inmailbox);
 126         mlx4_free_cmd_mailbox(dev->dev, outmailbox);
 127
 128         return err;
 129 }
 130
 131 static void update_sm_ah(struct mlx4_ib_dev *dev, u8 port_num, u16 lid, u8 sl)
 132 {
 133         struct ib_ah *new_ah;
 134         struct ib_ah_attr ah_attr;
 135
 136         if (!dev->send_agent[port_num - 1][0])
 137                 return;
 138
 139         memset(&ah_attr, 0, sizeof ah_attr);
 140         ah_attr.dlid     = lid;
 141         ah_attr.sl       = sl;
 142         ah_attr.port_num = port_num;
 143
 144         new_ah = ib_create_ah(dev->send_agent[port_num - 1][0]->qp->pd,
 145                               &ah_attr);
 146         if (IS_ERR(new_ah))
 147                 return;
 148
 149         spin_lock(&dev->sm_lock);
 150         if (dev->sm_ah[port_num - 1])
 151                 ib_destroy_ah(dev->sm_ah[port_num - 1]);
 152         dev->sm_ah[port_num - 1] = new_ah;
 153         spin_unlock(&dev->sm_lock);
 154 }
 155
 156 /*
 157  * Snoop SM MADs for port info and P_Key table sets, so we can
 158  * synthesize LID change and P_Key change events.
 159  */
 160 static void smp_snoop(struct ib_device *ibdev, u8 port_num, struct ib_mad *mad,
 161                                 u16 prev_lid)
 162 {
 163         struct ib_event event;
 164
 165         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 166              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 167             mad->mad_hdr.method == IB_MGMT_METHOD_SET) {
 168                 if (mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO) {
 169                         struct ib_port_info *pinfo =
 170                                 (struct ib_port_info *) ((struct ib_smp *) mad)->data;
 171                         u16 lid = be16_to_cpu(pinfo->lid);
 172
 173                         update_sm_ah(to_mdev(ibdev), port_num,
 174                                      be16_to_cpu(pinfo->sm_lid),
 175                                      pinfo->neighbormtu_mastersmsl & 0xf);
 176
 177                         event.device           = ibdev;
 178                         event.element.port_num = port_num;
 179
 180                         if (pinfo->clientrereg_resv_subnetto & 0x80) {
 181                                 event.event    = IB_EVENT_CLIENT_REREGISTER;
 182                                 ib_dispatch_event(&event);
 183                         }
 184
 185                         if (prev_lid != lid) {
 186                                 event.event    = IB_EVENT_LID_CHANGE;
 187                                 ib_dispatch_event(&event);
 188                         }
 189                 }
 190
 191                 if (mad->mad_hdr.attr_id == IB_SMP_ATTR_PKEY_TABLE) {
 192                         event.device           = ibdev;
 193                         event.event            = IB_EVENT_PKEY_CHANGE;
 194                         event.element.port_num = port_num;
 195                         ib_dispatch_event(&event);
 196                 }
 197         }
 198 }
 199
 200 static void node_desc_override(struct ib_device *dev,
 201                                struct ib_mad *mad)
 202 {
 203         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 204              mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 205             mad->mad_hdr.method == IB_MGMT_METHOD_GET_RESP &&
 206             mad->mad_hdr.attr_id == IB_SMP_ATTR_NODE_DESC) {
 207                 spin_lock(&to_mdev(dev)->sm_lock);
 208                 memcpy(((struct ib_smp *) mad)->data, dev->node_desc, 64);
 209                 spin_unlock(&to_mdev(dev)->sm_lock);
 210         }
 211 }
 212
 213 static void forward_trap(struct mlx4_ib_dev *dev, u8 port_num, struct ib_mad *mad)
 214 {
 215         int qpn = mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED;
 216         struct ib_mad_send_buf *send_buf;
 217         struct ib_mad_agent *agent = dev->send_agent[port_num - 1][qpn];
 218         int ret;
 219
 220         if (agent) {
 221                 send_buf = ib_create_send_mad(agent, qpn, 0, 0, IB_MGMT_MAD_HDR,
 222                                               IB_MGMT_MAD_DATA, GFP_ATOMIC);
 223                 if (IS_ERR(send_buf))
 224                         return;
 225                 /*
 226                  * We rely here on the fact that MLX QPs don't use the
 227                  * address handle after the send is posted (this is
 228                  * wrong following the IB spec strictly, but we know
 229                  * it's OK for our devices).
 230                  */
 231                 spin_lock(&dev->sm_lock);
 232                 memcpy(send_buf->mad, mad, sizeof *mad);
 233                 if ((send_buf->ah = dev->sm_ah[port_num - 1]))
 234                         ret = ib_post_send_mad(send_buf, NULL);
 235                 else
 236                         ret = -EINVAL;
 237                 spin_unlock(&dev->sm_lock);
 238
 239                 if (ret)
 240                         ib_free_send_mad(send_buf);
 241         }
 242 }
 243
 244 static int ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 245                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 246                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 247 {
 248         u16 slid, prev_lid = 0;
 249         int err;
 250         struct ib_port_attr pattr;
 251
 252         slid = in_wc ? in_wc->slid : be16_to_cpu(IB_LID_PERMISSIVE);
 253
 254         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP && slid == 0) {
 255                 forward_trap(to_mdev(ibdev), port_num, in_mad);
 256                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 257         }
 258
 259         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 260             in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 261                 if (in_mad->mad_hdr.method   != IB_MGMT_METHOD_GET &&
 262                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_SET &&
 263                     in_mad->mad_hdr.method   != IB_MGMT_METHOD_TRAP_REPRESS)
 264                         return IB_MAD_RESULT_SUCCESS;
 265
 266                 /*
 267                  * Don't process SMInfo queries -- the SMA can't handle them.
 268                  */
 269                 if (in_mad->mad_hdr.attr_id == IB_SMP_ATTR_SM_INFO)
 270                         return IB_MAD_RESULT_SUCCESS;
 271         } else if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_PERF_MGMT ||
 272                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS1   ||
 273                    in_mad->mad_hdr.mgmt_class == MLX4_IB_VENDOR_CLASS2   ||
 274                    in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CONG_MGMT) {
 275                 if (in_mad->mad_hdr.method  != IB_MGMT_METHOD_GET &&
 276                     in_mad->mad_hdr.method  != IB_MGMT_METHOD_SET)
 277                         return IB_MAD_RESULT_SUCCESS;
 278         } else
 279                 return IB_MAD_RESULT_SUCCESS;
 280
 281         if ((in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED ||
 282              in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) &&
 283             in_mad->mad_hdr.method == IB_MGMT_METHOD_SET &&
 284             in_mad->mad_hdr.attr_id == IB_SMP_ATTR_PORT_INFO &&
 285             !ib_query_port(ibdev, port_num, &pattr))
 286                 prev_lid = pattr.lid;
 287
 288         err = mlx4_MAD_IFC(to_mdev(ibdev),
 289                            mad_flags & IB_MAD_IGNORE_MKEY,
 290                            mad_flags & IB_MAD_IGNORE_BKEY,
 291                            port_num, in_wc, in_grh, in_mad, out_mad);
 292         if (err)
 293                 return IB_MAD_RESULT_FAILURE;
 294
 295         if (!out_mad->mad_hdr.status) {
 296                 smp_snoop(ibdev, port_num, in_mad, prev_lid);
 297                 node_desc_override(ibdev, out_mad);
 298         }
 299
 300         /* set return bit in status of directed route responses */
 301         if (in_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 302                 out_mad->mad_hdr.status |= cpu_to_be16(1 << 15);
 303
 304         if (in_mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS)
 305                 /* no response for trap repress */
 306                 return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
 307
 308         return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 309 }
 310
 311 static void edit_counter(struct mlx4_counter *cnt,
 312                                         struct ib_pma_portcounters *pma_cnt)
 313 {
 314         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_data,
 315                              (be64_to_cpu(cnt->tx_bytes) >> 2));
 316         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_data,
 317                              (be64_to_cpu(cnt->rx_bytes) >> 2));
 318         ASSIGN_32BIT_COUNTER(pma_cnt->port_xmit_packets,
 319                              be64_to_cpu(cnt->tx_frames));
 320         ASSIGN_32BIT_COUNTER(pma_cnt->port_rcv_packets,
 321                              be64_to_cpu(cnt->rx_frames));
 322 }
 323
 324 static int iboe_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 325                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 326                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 327 {
 328         struct mlx4_cmd_mailbox *mailbox;
 329         struct mlx4_ib_dev *dev = to_mdev(ibdev);
 330         int err;
 331         u32 inmod = dev->counters[port_num - 1] & 0xffff;
 332         u8 mode;
 333
 334         if (in_mad->mad_hdr.mgmt_class != IB_MGMT_CLASS_PERF_MGMT)
 335                 return -EINVAL;
 336
 337         mailbox = mlx4_alloc_cmd_mailbox(dev->dev);
 338         if (IS_ERR(mailbox))
 339                 return IB_MAD_RESULT_FAILURE;
 340
 341         err = mlx4_cmd_box(dev->dev, 0, mailbox->dma, inmod, 0,
 342                            MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C);
 343         if (err)
 344                 err = IB_MAD_RESULT_FAILURE;
 345         else {
 346                 memset(out_mad->data, 0, sizeof out_mad->data);
 347                 mode = ((struct mlx4_counter *)mailbox->buf)->counter_mode;
 348                 switch (mode & 0xf) {
 349                 case 0:
 350                         edit_counter(mailbox->buf,
 351                                                 (void *)(out_mad->data + 40));
 352                         err = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
 353                         break;
 354                 default:
 355                         err = IB_MAD_RESULT_FAILURE;
 356                 }
 357         }
 358
 359         mlx4_free_cmd_mailbox(dev->dev, mailbox);
 360
 361         return err;
 362 }
 363
 364 int mlx4_ib_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
 365                         struct ib_wc *in_wc, struct ib_grh *in_grh,
 366                         struct ib_mad *in_mad, struct ib_mad *out_mad)
 367 {
 368         switch (rdma_port_get_link_layer(ibdev, port_num)) {
 369         case IB_LINK_LAYER_INFINIBAND:
 370                 return ib_process_mad(ibdev, mad_flags, port_num, in_wc,
 371                                       in_grh, in_mad, out_mad);
 372         case IB_LINK_LAYER_ETHERNET:
 373                 return iboe_process_mad(ibdev, mad_flags, port_num, in_wc,
 374                                           in_grh, in_mad, out_mad);
 375         default:
 376                 return -EINVAL;
 377         }
 378 }
 379
 380 static void send_handler(struct ib_mad_agent *agent,
 381                          struct ib_mad_send_wc *mad_send_wc)
 382 {
 383         ib_free_send_mad(mad_send_wc->send_buf);
 384 }
 385
 386 int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 387 {
 388         struct ib_mad_agent *agent;
 389         int p, q;
 390         int ret;
 391         enum rdma_link_layer ll;
 392
 393         for (p = 0; p < dev->num_ports; ++p) {
 394                 ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 395                 for (q = 0; q <= 1; ++q) {
 396                         if (ll == IB_LINK_LAYER_INFINIBAND) {
 397                                 agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
 398                                                               q ? IB_QPT_GSI : IB_QPT_SMI,
 399                                                               NULL, 0, send_handler,
 400                                                               NULL, NULL);
 401                                 if (IS_ERR(agent)) {
 402                                         ret = PTR_ERR(agent);
 403                                         goto err;
 404                                 }
 405                                 dev->send_agent[p][q] = agent;
 406                         } else
 407                                 dev->send_agent[p][q] = NULL;
 408                 }
 409         }
 410
 411         return 0;
 412
 413 err:
 414         for (p = 0; p < dev->num_ports; ++p)
 415                 for (q = 0; q <= 1; ++q)
 416                         if (dev->send_agent[p][q])
 417                                 ib_unregister_mad_agent(dev->send_agent[p][q]);
 418
 419         return ret;
 420 }
 421
 422 void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 423 {
 424         struct ib_mad_agent *agent;
 425         int p, q;
 426
 427         for (p = 0; p < dev->num_ports; ++p) {
 428                 for (q = 0; q <= 1; ++q) {
 429                         agent = dev->send_agent[p][q];
 430                         if (agent) {
 431                                 dev->send_agent[p][q] = NULL;
 432                                 ib_unregister_mad_agent(agent);
 433                         }
 434                 }
 435
 436                 if (dev->sm_ah[p])
 437                         ib_destroy_ah(dev->sm_ah[p]);
 438         }
 439 }