git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/core/user_mad.c

   1 /*
   2  * Copyright (c) 2004 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
   4  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: user_mad.c 2814 2005-07-06 19:14:09Z halr $
  35  */
  36
  37 #include <linux/module.h>
  38 #include <linux/init.h>
  39 #include <linux/device.h>
  40 #include <linux/err.h>
  41 #include <linux/fs.h>
  42 #include <linux/cdev.h>
  43 #include <linux/pci.h>
  44 #include <linux/dma-mapping.h>
  45 #include <linux/poll.h>
  46 #include <linux/rwsem.h>
  47 #include <linux/kref.h>
  48
  49 #include <asm/uaccess.h>
  50 #include <asm/semaphore.h>
  51
  52 #include <rdma/ib_mad.h>
  53 #include <rdma/ib_user_mad.h>
  54
  55 MODULE_AUTHOR("Roland Dreier");
  56 MODULE_DESCRIPTION("InfiniBand userspace MAD packet access");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 enum {
  60         IB_UMAD_MAX_PORTS  = 64,
  61         IB_UMAD_MAX_AGENTS = 32,
  62
  63         IB_UMAD_MAJOR      = 231,
  64         IB_UMAD_MINOR_BASE = 0
  65 };
  66
  67 /*
  68  * Our lifetime rules for these structs are the following: each time a
  69  * device special file is opened, we look up the corresponding struct
  70  * ib_umad_port by minor in the umad_port[] table while holding the
  71  * port_lock.  If this lookup succeeds, we take a reference on the
  72  * ib_umad_port's struct ib_umad_device while still holding the
  73  * port_lock; if the lookup fails, we fail the open().  We drop these
  74  * references in the corresponding close().
  75  *
  76  * In addition to references coming from open character devices, there
  77  * is one more reference to each ib_umad_device representing the
  78  * module's reference taken when allocating the ib_umad_device in
  79  * ib_umad_add_one().
  80  *
  81  * When destroying an ib_umad_device, we clear all of its
  82  * ib_umad_ports from umad_port[] while holding port_lock before
  83  * dropping the module's reference to the ib_umad_device.  This is
  84  * always safe because any open() calls will either succeed and obtain
  85  * a reference before we clear the umad_port[] entries, or fail after
  86  * we clear the umad_port[] entries.
  87  */
  88
  89 struct ib_umad_port {
  90         struct cdev           *dev;
  91         struct class_device   *class_dev;
  92
  93         struct cdev           *sm_dev;
  94         struct class_device   *sm_class_dev;
  95         struct semaphore       sm_sem;
  96
  97         struct ib_device      *ib_dev;
  98         struct ib_umad_device *umad_dev;
  99         int                    dev_num;
 100         u8                     port_num;
 101 };
 102
 103 struct ib_umad_device {
 104         int                  start_port, end_port;
 105         struct kref          ref;
 106         struct ib_umad_port  port[0];
 107 };
 108
 109 struct ib_umad_file {
 110         struct ib_umad_port *port;
 111         spinlock_t           recv_lock;
 112         struct list_head     recv_list;
 113         wait_queue_head_t    recv_wait;
 114         struct rw_semaphore  agent_mutex;
 115         struct ib_mad_agent *agent[IB_UMAD_MAX_AGENTS];
 116         struct ib_mr        *mr[IB_UMAD_MAX_AGENTS];
 117 };
 118
 119 struct ib_umad_packet {
 120         struct ib_mad_send_buf *msg;
 121         struct list_head   list;
 122         int                length;
 123         struct ib_user_mad mad;
 124 };
 125
 126 static struct class *umad_class;
 127
 128 static const dev_t base_dev = MKDEV(IB_UMAD_MAJOR, IB_UMAD_MINOR_BASE);
 129
 130 static DEFINE_SPINLOCK(port_lock);
 131 static struct ib_umad_port *umad_port[IB_UMAD_MAX_PORTS];
 132 static DECLARE_BITMAP(dev_map, IB_UMAD_MAX_PORTS * 2);
 133
 134 static void ib_umad_add_one(struct ib_device *device);
 135 static void ib_umad_remove_one(struct ib_device *device);
 136
 137 static void ib_umad_release_dev(struct kref *ref)
 138 {
 139         struct ib_umad_device *dev =
 140                 container_of(ref, struct ib_umad_device, ref);
 141
 142         kfree(dev);
 143 }
 144
 145 static int queue_packet(struct ib_umad_file *file,
 146                         struct ib_mad_agent *agent,
 147                         struct ib_umad_packet *packet)
 148 {
 149         int ret = 1;
 150
 151         down_read(&file->agent_mutex);
 152         for (packet->mad.hdr.id = 0;
 153              packet->mad.hdr.id < IB_UMAD_MAX_AGENTS;
 154              packet->mad.hdr.id++)
 155                 if (agent == file->agent[packet->mad.hdr.id]) {
 156                         spin_lock_irq(&file->recv_lock);
 157                         list_add_tail(&packet->list, &file->recv_list);
 158                         spin_unlock_irq(&file->recv_lock);
 159                         wake_up_interruptible(&file->recv_wait);
 160                         ret = 0;
 161                         break;
 162                 }
 163
 164         up_read(&file->agent_mutex);
 165
 166         return ret;
 167 }
 168
 169 static void send_handler(struct ib_mad_agent *agent,
 170                          struct ib_mad_send_wc *send_wc)
 171 {
 172         struct ib_umad_file *file = agent->context;
 173         struct ib_umad_packet *timeout;
 174         struct ib_umad_packet *packet = send_wc->send_buf->context[0];
 175
 176         ib_destroy_ah(packet->msg->ah);
 177         ib_free_send_mad(packet->msg);
 178
 179         if (send_wc->status == IB_WC_RESP_TIMEOUT_ERR) {
 180                 timeout = kzalloc(sizeof *timeout + IB_MGMT_MAD_HDR, GFP_KERNEL);
 181                 if (!timeout)
 182                         goto out;
 183
 184                 timeout->length         = IB_MGMT_MAD_HDR;
 185                 timeout->mad.hdr.id     = packet->mad.hdr.id;
 186                 timeout->mad.hdr.status = ETIMEDOUT;
 187                 memcpy(timeout->mad.data, packet->mad.data,
 188                        sizeof (struct ib_mad_hdr));
 189
 190                 if (!queue_packet(file, agent, timeout))
 191                                 return;
 192         }
 193 out:
 194         kfree(packet);
 195 }
 196
 197 static void recv_handler(struct ib_mad_agent *agent,
 198                          struct ib_mad_recv_wc *mad_recv_wc)
 199 {
 200         struct ib_umad_file *file = agent->context;
 201         struct ib_umad_packet *packet;
 202         int length;
 203
 204         if (mad_recv_wc->wc->status != IB_WC_SUCCESS)
 205                 goto out;
 206
 207         length = mad_recv_wc->mad_len;
 208         packet = kzalloc(sizeof *packet + length, GFP_KERNEL);
 209         if (!packet)
 210                 goto out;
 211
 212         packet->length = length;
 213
 214         ib_coalesce_recv_mad(mad_recv_wc, packet->mad.data);
 215
 216         packet->mad.hdr.status    = 0;
 217         packet->mad.hdr.length    = length + sizeof (struct ib_user_mad);
 218         packet->mad.hdr.qpn       = cpu_to_be32(mad_recv_wc->wc->src_qp);
 219         packet->mad.hdr.lid       = cpu_to_be16(mad_recv_wc->wc->slid);
 220         packet->mad.hdr.sl        = mad_recv_wc->wc->sl;
 221         packet->mad.hdr.path_bits = mad_recv_wc->wc->dlid_path_bits;
 222         packet->mad.hdr.grh_present = !!(mad_recv_wc->wc->wc_flags & IB_WC_GRH);
 223         if (packet->mad.hdr.grh_present) {
 224                 /* XXX parse GRH */
 225                 packet->mad.hdr.gid_index       = 0;
 226                 packet->mad.hdr.hop_limit       = 0;
 227                 packet->mad.hdr.traffic_class   = 0;
 228                 memset(packet->mad.hdr.gid, 0, 16);
 229                 packet->mad.hdr.flow_label      = 0;
 230         }
 231
 232         if (queue_packet(file, agent, packet))
 233                 kfree(packet);
 234
 235 out:
 236         ib_free_recv_mad(mad_recv_wc);
 237 }
 238
 239 static ssize_t ib_umad_read(struct file *filp, char __user *buf,
 240                             size_t count, loff_t *pos)
 241 {
 242         struct ib_umad_file *file = filp->private_data;
 243         struct ib_umad_packet *packet;
 244         ssize_t ret;
 245
 246         if (count < sizeof (struct ib_user_mad) + sizeof (struct ib_mad))
 247                 return -EINVAL;
 248
 249         spin_lock_irq(&file->recv_lock);
 250
 251         while (list_empty(&file->recv_list)) {
 252                 spin_unlock_irq(&file->recv_lock);
 253
 254                 if (filp->f_flags & O_NONBLOCK)
 255                         return -EAGAIN;
 256
 257                 if (wait_event_interruptible(file->recv_wait,
 258                                              !list_empty(&file->recv_list)))
 259                         return -ERESTARTSYS;
 260
 261                 spin_lock_irq(&file->recv_lock);
 262         }
 263
 264         packet = list_entry(file->recv_list.next, struct ib_umad_packet, list);
 265         list_del(&packet->list);
 266
 267         spin_unlock_irq(&file->recv_lock);
 268
 269         if (count < packet->length + sizeof (struct ib_user_mad)) {
 270                 /* Return length needed (and first RMPP segment) if too small */
 271                 if (copy_to_user(buf, &packet->mad,
 272                                  sizeof (struct ib_user_mad) + sizeof (struct ib_mad)))
 273                         ret = -EFAULT;
 274                 else
 275                         ret = -ENOSPC;
 276         } else if (copy_to_user(buf, &packet->mad,
 277                                 packet->length + sizeof (struct ib_user_mad)))
 278                 ret = -EFAULT;
 279         else
 280                 ret = packet->length + sizeof (struct ib_user_mad);
 281         if (ret < 0) {
 282                 /* Requeue packet */
 283                 spin_lock_irq(&file->recv_lock);
 284                 list_add(&packet->list, &file->recv_list);
 285                 spin_unlock_irq(&file->recv_lock);
 286         } else
 287                 kfree(packet);
 288         return ret;
 289 }
 290
 291 static ssize_t ib_umad_write(struct file *filp, const char __user *buf,
 292                              size_t count, loff_t *pos)
 293 {
 294         struct ib_umad_file *file = filp->private_data;
 295         struct ib_umad_packet *packet;
 296         struct ib_mad_agent *agent;
 297         struct ib_ah_attr ah_attr;
 298         struct ib_ah *ah;
 299         struct ib_rmpp_mad *rmpp_mad;
 300         u8 method;
 301         __be64 *tid;
 302         int ret, length, hdr_len, copy_offset;
 303         int rmpp_active = 0;
 304
 305         if (count < sizeof (struct ib_user_mad))
 306                 return -EINVAL;
 307
 308         length = count - sizeof (struct ib_user_mad);
 309         packet = kmalloc(sizeof *packet + IB_MGMT_RMPP_HDR, GFP_KERNEL);
 310         if (!packet)
 311                 return -ENOMEM;
 312
 313         if (copy_from_user(&packet->mad, buf,
 314                             sizeof (struct ib_user_mad) + IB_MGMT_RMPP_HDR)) {
 315                 ret = -EFAULT;
 316                 goto err;
 317         }
 318
 319         if (packet->mad.hdr.id < 0 ||
 320             packet->mad.hdr.id >= IB_UMAD_MAX_AGENTS) {
 321                 ret = -EINVAL;
 322                 goto err;
 323         }
 324
 325         down_read(&file->agent_mutex);
 326
 327         agent = file->agent[packet->mad.hdr.id];
 328         if (!agent) {
 329                 ret = -EINVAL;
 330                 goto err_up;
 331         }
 332
 333         memset(&ah_attr, 0, sizeof ah_attr);
 334         ah_attr.dlid          = be16_to_cpu(packet->mad.hdr.lid);
 335         ah_attr.sl            = packet->mad.hdr.sl;
 336         ah_attr.src_path_bits = packet->mad.hdr.path_bits;
 337         ah_attr.port_num      = file->port->port_num;
 338         if (packet->mad.hdr.grh_present) {
 339                 ah_attr.ah_flags = IB_AH_GRH;
 340                 memcpy(ah_attr.grh.dgid.raw, packet->mad.hdr.gid, 16);
 341                 ah_attr.grh.flow_label     = be32_to_cpu(packet->mad.hdr.flow_label);
 342                 ah_attr.grh.hop_limit      = packet->mad.hdr.hop_limit;
 343                 ah_attr.grh.traffic_class  = packet->mad.hdr.traffic_class;
 344         }
 345
 346         ah = ib_create_ah(agent->qp->pd, &ah_attr);
 347         if (IS_ERR(ah)) {
 348                 ret = PTR_ERR(ah);
 349                 goto err_up;
 350         }
 351
 352         rmpp_mad = (struct ib_rmpp_mad *) packet->mad.data;
 353         if (ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) & IB_MGMT_RMPP_FLAG_ACTIVE) {
 354                 /* RMPP active */
 355                 if (!agent->rmpp_version) {
 356                         ret = -EINVAL;
 357                         goto err_ah;
 358                 }
 359
 360                 /* Validate that the management class can support RMPP */
 361                 if (rmpp_mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_ADM) {
 362                         hdr_len = IB_MGMT_SA_HDR;
 363                 } else if ((rmpp_mad->mad_hdr.mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 364                             (rmpp_mad->mad_hdr.mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)) {
 365                                 hdr_len = IB_MGMT_VENDOR_HDR;
 366                 } else {
 367                         ret = -EINVAL;
 368                         goto err_ah;
 369                 }
 370                 rmpp_active = 1;
 371                 copy_offset = IB_MGMT_RMPP_HDR;
 372         } else {
 373                 hdr_len = IB_MGMT_MAD_HDR;
 374                 copy_offset = IB_MGMT_MAD_HDR;
 375         }
 376
 377         packet->msg = ib_create_send_mad(agent,
 378                                          be32_to_cpu(packet->mad.hdr.qpn),
 379                                          0, rmpp_active,
 380                                          hdr_len, length - hdr_len,
 381                                          GFP_KERNEL);
 382         if (IS_ERR(packet->msg)) {
 383                 ret = PTR_ERR(packet->msg);
 384                 goto err_ah;
 385         }
 386
 387         packet->msg->ah         = ah;
 388         packet->msg->timeout_ms = packet->mad.hdr.timeout_ms;
 389         packet->msg->retries    = packet->mad.hdr.retries;
 390         packet->msg->context[0] = packet;
 391
 392         /* Copy MAD headers (RMPP header in place) */
 393         memcpy(packet->msg->mad, packet->mad.data, IB_MGMT_MAD_HDR);
 394         /* Now, copy rest of message from user into send buffer */
 395         if (copy_from_user(packet->msg->mad + copy_offset,
 396                            buf + sizeof (struct ib_user_mad) + copy_offset,
 397                            length - copy_offset)) {
 398                 ret = -EFAULT;
 399                 goto err_msg;
 400         }
 401
 402         /*
 403          * If userspace is generating a request that will generate a
 404          * response, we need to make sure the high-order part of the
 405          * transaction ID matches the agent being used to send the
 406          * MAD.
 407          */
 408         method = ((struct ib_mad_hdr *) packet->msg->mad)->method;
 409
 410         if (!(method & IB_MGMT_METHOD_RESP)       &&
 411             method != IB_MGMT_METHOD_TRAP_REPRESS &&
 412             method != IB_MGMT_METHOD_SEND) {
 413                 tid = &((struct ib_mad_hdr *) packet->msg->mad)->tid;
 414                 *tid = cpu_to_be64(((u64) agent->hi_tid) << 32 |
 415                                    (be64_to_cpup(tid) & 0xffffffff));
 416         }
 417
 418         ret = ib_post_send_mad(packet->msg, NULL);
 419         if (ret)
 420                 goto err_msg;
 421
 422         up_read(&file->agent_mutex);
 423
 424         return count;
 425
 426 err_msg:
 427         ib_free_send_mad(packet->msg);
 428
 429 err_ah:
 430         ib_destroy_ah(ah);
 431
 432 err_up:
 433         up_read(&file->agent_mutex);
 434
 435 err:
 436         kfree(packet);
 437         return ret;
 438 }
 439
 440 static unsigned int ib_umad_poll(struct file *filp, struct poll_table_struct *wait)
 441 {
 442         struct ib_umad_file *file = filp->private_data;
 443
 444         /* we will always be able to post a MAD send */
 445         unsigned int mask = POLLOUT | POLLWRNORM;
 446
 447         poll_wait(filp, &file->recv_wait, wait);
 448
 449         if (!list_empty(&file->recv_list))
 450                 mask |= POLLIN | POLLRDNORM;
 451
 452         return mask;
 453 }
 454
 455 static int ib_umad_reg_agent(struct ib_umad_file *file, unsigned long arg)
 456 {
 457         struct ib_user_mad_reg_req ureq;
 458         struct ib_mad_reg_req req;
 459         struct ib_mad_agent *agent;
 460         int agent_id;
 461         int ret;
 462
 463         down_write(&file->agent_mutex);
 464
 465         if (copy_from_user(&ureq, (void __user *) arg, sizeof ureq)) {
 466                 ret = -EFAULT;
 467                 goto out;
 468         }
 469
 470         if (ureq.qpn != 0 && ureq.qpn != 1) {
 471                 ret = -EINVAL;
 472                 goto out;
 473         }
 474
 475         for (agent_id = 0; agent_id < IB_UMAD_MAX_AGENTS; ++agent_id)
 476                 if (!file->agent[agent_id])
 477                         goto found;
 478
 479         ret = -ENOMEM;
 480         goto out;
 481
 482 found:
 483         if (ureq.mgmt_class) {
 484                 req.mgmt_class         = ureq.mgmt_class;
 485                 req.mgmt_class_version = ureq.mgmt_class_version;
 486                 memcpy(req.method_mask, ureq.method_mask, sizeof req.method_mask);
 487                 memcpy(req.oui,         ureq.oui,         sizeof req.oui);
 488         }
 489
 490         agent = ib_register_mad_agent(file->port->ib_dev, file->port->port_num,
 491                                       ureq.qpn ? IB_QPT_GSI : IB_QPT_SMI,
 492                                       ureq.mgmt_class ? &req : NULL,
 493                                       ureq.rmpp_version,
 494                                       send_handler, recv_handler, file);
 495         if (IS_ERR(agent)) {
 496                 ret = PTR_ERR(agent);
 497                 goto out;
 498         }
 499
 500         file->agent[agent_id] = agent;
 501
 502         file->mr[agent_id] = ib_get_dma_mr(agent->qp->pd, IB_ACCESS_LOCAL_WRITE);
 503         if (IS_ERR(file->mr[agent_id])) {
 504                 ret = -ENOMEM;
 505                 goto err;
 506         }
 507
 508         if (put_user(agent_id,
 509                      (u32 __user *) (arg + offsetof(struct ib_user_mad_reg_req, id)))) {
 510                 ret = -EFAULT;
 511                 goto err_mr;
 512         }
 513
 514         ret = 0;
 515         goto out;
 516
 517 err_mr:
 518         ib_dereg_mr(file->mr[agent_id]);
 519
 520 err:
 521         file->agent[agent_id] = NULL;
 522         ib_unregister_mad_agent(agent);
 523
 524 out:
 525         up_write(&file->agent_mutex);
 526         return ret;
 527 }
 528
 529 static int ib_umad_unreg_agent(struct ib_umad_file *file, unsigned long arg)
 530 {
 531         u32 id;
 532         int ret = 0;
 533
 534         down_write(&file->agent_mutex);
 535
 536         if (get_user(id, (u32 __user *) arg)) {
 537                 ret = -EFAULT;
 538                 goto out;
 539         }
 540
 541         if (id < 0 || id >= IB_UMAD_MAX_AGENTS || !file->agent[id]) {
 542                 ret = -EINVAL;
 543                 goto out;
 544         }
 545
 546         ib_dereg_mr(file->mr[id]);
 547         ib_unregister_mad_agent(file->agent[id]);
 548         file->agent[id] = NULL;
 549
 550 out:
 551         up_write(&file->agent_mutex);
 552         return ret;
 553 }
 554
 555 static long ib_umad_ioctl(struct file *filp, unsigned int cmd,
 556                           unsigned long arg)
 557 {
 558         switch (cmd) {
 559         case IB_USER_MAD_REGISTER_AGENT:
 560                 return ib_umad_reg_agent(filp->private_data, arg);
 561         case IB_USER_MAD_UNREGISTER_AGENT:
 562                 return ib_umad_unreg_agent(filp->private_data, arg);
 563         default:
 564                 return -ENOIOCTLCMD;
 565         }
 566 }
 567
 568 static int ib_umad_open(struct inode *inode, struct file *filp)
 569 {
 570         struct ib_umad_port *port;
 571         struct ib_umad_file *file;
 572
 573         spin_lock(&port_lock);
 574         port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE];
 575         if (port)
 576                 kref_get(&port->umad_dev->ref);
 577         spin_unlock(&port_lock);
 578
 579         if (!port)
 580                 return -ENXIO;
 581
 582         file = kzalloc(sizeof *file, GFP_KERNEL);
 583         if (!file) {
 584                 kref_put(&port->umad_dev->ref, ib_umad_release_dev);
 585                 return -ENOMEM;
 586         }
 587
 588         spin_lock_init(&file->recv_lock);
 589         init_rwsem(&file->agent_mutex);
 590         INIT_LIST_HEAD(&file->recv_list);
 591         init_waitqueue_head(&file->recv_wait);
 592
 593         file->port = port;
 594         filp->private_data = file;
 595
 596         return 0;
 597 }
 598
 599 static int ib_umad_close(struct inode *inode, struct file *filp)
 600 {
 601         struct ib_umad_file *file = filp->private_data;
 602         struct ib_umad_device *dev = file->port->umad_dev;
 603         struct ib_umad_packet *packet, *tmp;
 604         int i;
 605
 606         for (i = 0; i < IB_UMAD_MAX_AGENTS; ++i)
 607                 if (file->agent[i]) {
 608                         ib_dereg_mr(file->mr[i]);
 609                         ib_unregister_mad_agent(file->agent[i]);
 610                 }
 611
 612         list_for_each_entry_safe(packet, tmp, &file->recv_list, list)
 613                 kfree(packet);
 614
 615         kfree(file);
 616
 617         kref_put(&dev->ref, ib_umad_release_dev);
 618
 619         return 0;
 620 }
 621
 622 static struct file_operations umad_fops = {
 623         .owner          = THIS_MODULE,
 624         .read           = ib_umad_read,
 625         .write          = ib_umad_write,
 626         .poll           = ib_umad_poll,
 627         .unlocked_ioctl = ib_umad_ioctl,
 628         .compat_ioctl   = ib_umad_ioctl,
 629         .open           = ib_umad_open,
 630         .release        = ib_umad_close
 631 };
 632
 633 static int ib_umad_sm_open(struct inode *inode, struct file *filp)
 634 {
 635         struct ib_umad_port *port;
 636         struct ib_port_modify props = {
 637                 .set_port_cap_mask = IB_PORT_SM
 638         };
 639         int ret;
 640
 641         spin_lock(&port_lock);
 642         port = umad_port[iminor(inode) - IB_UMAD_MINOR_BASE - IB_UMAD_MAX_PORTS];
 643         if (port)
 644                 kref_get(&port->umad_dev->ref);
 645         spin_unlock(&port_lock);
 646
 647         if (!port)
 648                 return -ENXIO;
 649
 650         if (filp->f_flags & O_NONBLOCK) {
 651                 if (down_trylock(&port->sm_sem)) {
 652                         ret = -EAGAIN;
 653                         goto fail;
 654                 }
 655         } else {
 656                 if (down_interruptible(&port->sm_sem)) {
 657                         ret = -ERESTARTSYS;
 658                         goto fail;
 659                 }
 660         }
 661
 662         ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
 663         if (ret) {
 664                 up(&port->sm_sem);
 665                 goto fail;
 666         }
 667
 668         filp->private_data = port;
 669
 670         return 0;
 671
 672 fail:
 673         kref_put(&port->umad_dev->ref, ib_umad_release_dev);
 674         return ret;
 675 }
 676
 677 static int ib_umad_sm_close(struct inode *inode, struct file *filp)
 678 {
 679         struct ib_umad_port *port = filp->private_data;
 680         struct ib_port_modify props = {
 681                 .clr_port_cap_mask = IB_PORT_SM
 682         };
 683         int ret;
 684
 685         ret = ib_modify_port(port->ib_dev, port->port_num, 0, &props);
 686         up(&port->sm_sem);
 687
 688         kref_put(&port->umad_dev->ref, ib_umad_release_dev);
 689
 690         return ret;
 691 }
 692
 693 static struct file_operations umad_sm_fops = {
 694         .owner   = THIS_MODULE,
 695         .open    = ib_umad_sm_open,
 696         .release = ib_umad_sm_close
 697 };
 698
 699 static struct ib_client umad_client = {
 700         .name   = "umad",
 701         .add    = ib_umad_add_one,
 702         .remove = ib_umad_remove_one
 703 };
 704
 705 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
 706 {
 707         struct ib_umad_port *port = class_get_devdata(class_dev);
 708
 709         if (!port)
 710                 return -ENODEV;
 711
 712         return sprintf(buf, "%s\n", port->ib_dev->name);
 713 }
 714 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
 715
 716 static ssize_t show_port(struct class_device *class_dev, char *buf)
 717 {
 718         struct ib_umad_port *port = class_get_devdata(class_dev);
 719
 720         if (!port)
 721                 return -ENODEV;
 722
 723         return sprintf(buf, "%d\n", port->port_num);
 724 }
 725 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
 726
 727 static ssize_t show_abi_version(struct class *class, char *buf)
 728 {
 729         return sprintf(buf, "%d\n", IB_USER_MAD_ABI_VERSION);
 730 }
 731 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
 732
 733 static int ib_umad_init_port(struct ib_device *device, int port_num,
 734                              struct ib_umad_port *port)
 735 {
 736         spin_lock(&port_lock);
 737         port->dev_num = find_first_zero_bit(dev_map, IB_UMAD_MAX_PORTS);
 738         if (port->dev_num >= IB_UMAD_MAX_PORTS) {
 739                 spin_unlock(&port_lock);
 740                 return -1;
 741         }
 742         set_bit(port->dev_num, dev_map);
 743         spin_unlock(&port_lock);
 744
 745         port->ib_dev   = device;
 746         port->port_num = port_num;
 747         init_MUTEX(&port->sm_sem);
 748
 749         port->dev = cdev_alloc();
 750         if (!port->dev)
 751                 return -1;
 752         port->dev->owner = THIS_MODULE;
 753         port->dev->ops   = &umad_fops;
 754         kobject_set_name(&port->dev->kobj, "umad%d", port->dev_num);
 755         if (cdev_add(port->dev, base_dev + port->dev_num, 1))
 756                 goto err_cdev;
 757
 758         port->class_dev = class_device_create(umad_class, NULL, port->dev->dev,
 759                                               device->dma_device,
 760                                               "umad%d", port->dev_num);
 761         if (IS_ERR(port->class_dev))
 762                 goto err_cdev;
 763
 764         if (class_device_create_file(port->class_dev, &class_device_attr_ibdev))
 765                 goto err_class;
 766         if (class_device_create_file(port->class_dev, &class_device_attr_port))
 767                 goto err_class;
 768
 769         port->sm_dev = cdev_alloc();
 770         if (!port->sm_dev)
 771                 goto err_class;
 772         port->sm_dev->owner = THIS_MODULE;
 773         port->sm_dev->ops   = &umad_sm_fops;
 774         kobject_set_name(&port->dev->kobj, "issm%d", port->dev_num);
 775         if (cdev_add(port->sm_dev, base_dev + port->dev_num + IB_UMAD_MAX_PORTS, 1))
 776                 goto err_sm_cdev;
 777
 778         port->sm_class_dev = class_device_create(umad_class, NULL, port->sm_dev->dev,
 779                                                  device->dma_device,
 780                                                  "issm%d", port->dev_num);
 781         if (IS_ERR(port->sm_class_dev))
 782                 goto err_sm_cdev;
 783
 784         class_set_devdata(port->class_dev,    port);
 785         class_set_devdata(port->sm_class_dev, port);
 786
 787         if (class_device_create_file(port->sm_class_dev, &class_device_attr_ibdev))
 788                 goto err_sm_class;
 789         if (class_device_create_file(port->sm_class_dev, &class_device_attr_port))
 790                 goto err_sm_class;
 791
 792         spin_lock(&port_lock);
 793         umad_port[port->dev_num] = port;
 794         spin_unlock(&port_lock);
 795
 796         return 0;
 797
 798 err_sm_class:
 799         class_device_destroy(umad_class, port->sm_dev->dev);
 800
 801 err_sm_cdev:
 802         cdev_del(port->sm_dev);
 803
 804 err_class:
 805         class_device_destroy(umad_class, port->dev->dev);
 806
 807 err_cdev:
 808         cdev_del(port->dev);
 809         clear_bit(port->dev_num, dev_map);
 810
 811         return -1;
 812 }
 813
 814 static void ib_umad_kill_port(struct ib_umad_port *port)
 815 {
 816         class_set_devdata(port->class_dev,    NULL);
 817         class_set_devdata(port->sm_class_dev, NULL);
 818
 819         class_device_destroy(umad_class, port->dev->dev);
 820         class_device_destroy(umad_class, port->sm_dev->dev);
 821
 822         cdev_del(port->dev);
 823         cdev_del(port->sm_dev);
 824
 825         spin_lock(&port_lock);
 826         umad_port[port->dev_num] = NULL;
 827         spin_unlock(&port_lock);
 828
 829         clear_bit(port->dev_num, dev_map);
 830 }
 831
 832 static void ib_umad_add_one(struct ib_device *device)
 833 {
 834         struct ib_umad_device *umad_dev;
 835         int s, e, i;
 836
 837         if (device->node_type == IB_NODE_SWITCH)
 838                 s = e = 0;
 839         else {
 840                 s = 1;
 841                 e = device->phys_port_cnt;
 842         }
 843
 844         umad_dev = kzalloc(sizeof *umad_dev +
 845                            (e - s + 1) * sizeof (struct ib_umad_port),
 846                            GFP_KERNEL);
 847         if (!umad_dev)
 848                 return;
 849
 850         kref_init(&umad_dev->ref);
 851
 852         umad_dev->start_port = s;
 853         umad_dev->end_port   = e;
 854
 855         for (i = s; i <= e; ++i) {
 856                 umad_dev->port[i - s].umad_dev = umad_dev;
 857
 858                 if (ib_umad_init_port(device, i, &umad_dev->port[i - s]))
 859                         goto err;
 860         }
 861
 862         ib_set_client_data(device, &umad_client, umad_dev);
 863
 864         return;
 865
 866 err:
 867         while (--i >= s)
 868                 ib_umad_kill_port(&umad_dev->port[i]);
 869
 870         kref_put(&umad_dev->ref, ib_umad_release_dev);
 871 }
 872
 873 static void ib_umad_remove_one(struct ib_device *device)
 874 {
 875         struct ib_umad_device *umad_dev = ib_get_client_data(device, &umad_client);
 876         int i;
 877
 878         if (!umad_dev)
 879                 return;
 880
 881         for (i = 0; i <= umad_dev->end_port - umad_dev->start_port; ++i)
 882                 ib_umad_kill_port(&umad_dev->port[i]);
 883
 884         kref_put(&umad_dev->ref, ib_umad_release_dev);
 885 }
 886
 887 static int __init ib_umad_init(void)
 888 {
 889         int ret;
 890
 891         ret = register_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2,
 892                                      "infiniband_mad");
 893         if (ret) {
 894                 printk(KERN_ERR "user_mad: couldn't register device number\n");
 895                 goto out;
 896         }
 897
 898         umad_class = class_create(THIS_MODULE, "infiniband_mad");
 899         if (IS_ERR(umad_class)) {
 900                 ret = PTR_ERR(umad_class);
 901                 printk(KERN_ERR "user_mad: couldn't create class infiniband_mad\n");
 902                 goto out_chrdev;
 903         }
 904
 905         ret = class_create_file(umad_class, &class_attr_abi_version);
 906         if (ret) {
 907                 printk(KERN_ERR "user_mad: couldn't create abi_version attribute\n");
 908                 goto out_class;
 909         }
 910
 911         ret = ib_register_client(&umad_client);
 912         if (ret) {
 913                 printk(KERN_ERR "user_mad: couldn't register ib_umad client\n");
 914                 goto out_class;
 915         }
 916
 917         return 0;
 918
 919 out_class:
 920         class_destroy(umad_class);
 921
 922 out_chrdev:
 923         unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
 924
 925 out:
 926         return ret;
 927 }
 928
 929 static void __exit ib_umad_cleanup(void)
 930 {
 931         ib_unregister_client(&umad_client);
 932         class_destroy(umad_class);
 933         unregister_chrdev_region(base_dev, IB_UMAD_MAX_PORTS * 2);
 934 }
 935
 936 module_init(ib_umad_init);
 937 module_exit(ib_umad_cleanup);