sys/dev/hyperv/vmbus/hv_channel.c

   1 /*-
   2  * Copyright (c) 2009-2012 Microsoft Corp.
   3  * Copyright (c) 2012 NetApp Inc.
   4  * Copyright (c) 2012 Citrix Inc.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice unmodified, this list of conditions, and the following
  12  *    disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27  */
  28
  29 #include <sys/cdefs.h>
  30 __FBSDID("$FreeBSD$");
  31
  32 #include <sys/param.h>
  33 #include <sys/kernel.h>
  34 #include <sys/malloc.h>
  35 #include <sys/systm.h>
  36 #include <sys/mbuf.h>
  37 #include <sys/lock.h>
  38 #include <sys/mutex.h>
  39 #include <sys/sysctl.h>
  40 #include <machine/bus.h>
  41 #include <vm/vm.h>
  42 #include <vm/vm_param.h>
  43 #include <vm/pmap.h>
  44
  45 #include "hv_vmbus_priv.h"
  46
  47 static int      vmbus_channel_create_gpadl_header(
  48                         /* must be phys and virt contiguous*/
  49                         void*                           contig_buffer,
  50                         /* page-size multiple */
  51                         uint32_t                        size,
  52                         hv_vmbus_channel_msg_info**     msg_info,
  53                         uint32_t*                       message_count);
  54
  55 static void     vmbus_channel_set_event(hv_vmbus_channel* channel);
  56 static void     VmbusProcessChannelEvent(void* channel, int pending);
  57
  58 /**
  59  *  @brief Trigger an event notification on the specified channel
  60  */
  61 static void
  62 vmbus_channel_set_event(hv_vmbus_channel *channel)
  63 {
  64         hv_vmbus_monitor_page *monitor_page;
  65
  66         if (channel->offer_msg.monitor_allocated) {
  67                 /* Each uint32_t represents 32 channels */
  68                 synch_set_bit((channel->offer_msg.child_rel_id & 31),
  69                         ((uint32_t *)hv_vmbus_g_connection.send_interrupt_page
  70                                 + ((channel->offer_msg.child_rel_id >> 5))));
  71
  72                 monitor_page = (hv_vmbus_monitor_page *)
  73                         hv_vmbus_g_connection.monitor_page_2;
  74
  75                 synch_set_bit(channel->monitor_bit,
  76                         (uint32_t *)&monitor_page->
  77                                 trigger_group[channel->monitor_group].u.pending);
  78         } else {
  79                 hv_vmbus_set_event(channel);
  80         }
  81
  82 }
  83
  84 static int
  85 vmbus_channel_sysctl_monalloc(SYSCTL_HANDLER_ARGS)
  86 {
  87         struct hv_vmbus_channel *chan = arg1;
  88         int alloc = 0;
  89
  90         if (chan->offer_msg.monitor_allocated)
  91                 alloc = 1;
  92         return sysctl_handle_int(oidp, &alloc, 0, req);
  93 }
  94
  95 static void
  96 vmbus_channel_sysctl_create(hv_vmbus_channel* channel)
  97 {
  98         device_t dev;
  99         struct sysctl_oid *devch_sysctl;
 100         struct sysctl_oid *devch_id_sysctl, *devch_sub_sysctl;
 101         struct sysctl_oid *devch_id_in_sysctl, *devch_id_out_sysctl;
 102         struct sysctl_ctx_list *ctx;
 103         uint32_t ch_id;
 104         uint16_t sub_ch_id;
 105         char name[16];
 106
 107         hv_vmbus_channel* primary_ch = channel->primary_channel;
 108
 109         if (primary_ch == NULL) {
 110                 dev = channel->device->device;
 111                 ch_id = channel->offer_msg.child_rel_id;
 112         } else {
 113                 dev = primary_ch->device->device;
 114                 ch_id = primary_ch->offer_msg.child_rel_id;
 115                 sub_ch_id = channel->offer_msg.offer.sub_channel_index;
 116         }
 117         ctx = device_get_sysctl_ctx(dev);
 118         /* This creates dev.DEVNAME.DEVUNIT.channel tree */
 119         devch_sysctl = SYSCTL_ADD_NODE(ctx,
 120                     SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 121                     OID_AUTO, "channel", CTLFLAG_RD, 0, "");
 122         /* This creates dev.DEVNAME.DEVUNIT.channel.CHANID tree */
 123         snprintf(name, sizeof(name), "%d", ch_id);
 124         devch_id_sysctl = SYSCTL_ADD_NODE(ctx,
 125                     SYSCTL_CHILDREN(devch_sysctl),
 126                     OID_AUTO, name, CTLFLAG_RD, 0, "");
 127
 128         if (primary_ch != NULL) {
 129                 devch_sub_sysctl = SYSCTL_ADD_NODE(ctx,
 130                         SYSCTL_CHILDREN(devch_id_sysctl),
 131                         OID_AUTO, "sub", CTLFLAG_RD, 0, "");
 132                 snprintf(name, sizeof(name), "%d", sub_ch_id);
 133                 devch_id_sysctl = SYSCTL_ADD_NODE(ctx,
 134                         SYSCTL_CHILDREN(devch_sub_sysctl),
 135                         OID_AUTO, name, CTLFLAG_RD, 0, "");
 136
 137                 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(devch_id_sysctl),
 138                     OID_AUTO, "chanid", CTLFLAG_RD,
 139                     &channel->offer_msg.child_rel_id, 0, "channel id");
 140         }
 141         SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(devch_id_sysctl), OID_AUTO,
 142             "cpu", CTLFLAG_RD, &channel->target_cpu, 0, "owner CPU id");
 143         SYSCTL_ADD_PROC(ctx, SYSCTL_CHILDREN(devch_id_sysctl), OID_AUTO,
 144             "monitor_allocated", CTLTYPE_INT | CTLFLAG_RD, channel, 0,
 145             vmbus_channel_sysctl_monalloc, "I",
 146             "is monitor allocated to this channel");
 147
 148         devch_id_in_sysctl = SYSCTL_ADD_NODE(ctx,
 149                     SYSCTL_CHILDREN(devch_id_sysctl),
 150                     OID_AUTO,
 151                     "in",
 152                     CTLFLAG_RD, 0, "");
 153         devch_id_out_sysctl = SYSCTL_ADD_NODE(ctx,
 154                     SYSCTL_CHILDREN(devch_id_sysctl),
 155                     OID_AUTO,
 156                     "out",
 157                     CTLFLAG_RD, 0, "");
 158         hv_ring_buffer_stat(ctx,
 159                 SYSCTL_CHILDREN(devch_id_in_sysctl),
 160                 &(channel->inbound),
 161                 "inbound ring buffer stats");
 162         hv_ring_buffer_stat(ctx,
 163                 SYSCTL_CHILDREN(devch_id_out_sysctl),
 164                 &(channel->outbound),
 165                 "outbound ring buffer stats");
 166 }
 167
 168 /**
 169  * @brief Open the specified channel
 170  */
 171 int
 172 hv_vmbus_channel_open(
 173         hv_vmbus_channel*               new_channel,
 174         uint32_t                        send_ring_buffer_size,
 175         uint32_t                        recv_ring_buffer_size,
 176         void*                           user_data,
 177         uint32_t                        user_data_len,
 178         hv_vmbus_pfn_channel_callback   pfn_on_channel_callback,
 179         void*                           context)
 180 {
 181
 182         int ret = 0;
 183         void *in, *out;
 184         hv_vmbus_channel_open_channel*  open_msg;
 185         hv_vmbus_channel_msg_info*      open_info;
 186
 187         mtx_lock(&new_channel->sc_lock);
 188         if (new_channel->state == HV_CHANNEL_OPEN_STATE) {
 189             new_channel->state = HV_CHANNEL_OPENING_STATE;
 190         } else {
 191             mtx_unlock(&new_channel->sc_lock);
 192             if(bootverbose)
 193                 printf("VMBUS: Trying to open channel <%p> which in "
 194                     "%d state.\n", new_channel, new_channel->state);
 195             return (EINVAL);
 196         }
 197         mtx_unlock(&new_channel->sc_lock);
 198
 199         new_channel->on_channel_callback = pfn_on_channel_callback;
 200         new_channel->channel_callback_context = context;
 201
 202         new_channel->rxq = hv_vmbus_g_context.hv_event_queue[new_channel->target_cpu];
 203         TASK_INIT(&new_channel->channel_task, 0, VmbusProcessChannelEvent, new_channel);
 204
 205         /* Allocate the ring buffer */
 206         out = contigmalloc((send_ring_buffer_size + recv_ring_buffer_size),
 207             M_DEVBUF, M_ZERO, 0UL, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
 208         KASSERT(out != NULL,
 209             ("Error VMBUS: contigmalloc failed to allocate Ring Buffer!"));
 210         if (out == NULL)
 211                 return (ENOMEM);
 212
 213         in = ((uint8_t *) out + send_ring_buffer_size);
 214
 215         new_channel->ring_buffer_pages = out;
 216         new_channel->ring_buffer_page_count = (send_ring_buffer_size +
 217             recv_ring_buffer_size) >> PAGE_SHIFT;
 218         new_channel->ring_buffer_size = send_ring_buffer_size +
 219             recv_ring_buffer_size;
 220
 221         hv_vmbus_ring_buffer_init(
 222                 &new_channel->outbound,
 223                 out,
 224                 send_ring_buffer_size);
 225
 226         hv_vmbus_ring_buffer_init(
 227                 &new_channel->inbound,
 228                 in,
 229                 recv_ring_buffer_size);
 230
 231         /* Create sysctl tree for this channel */
 232         vmbus_channel_sysctl_create(new_channel);
 233
 234         /**
 235          * Establish the gpadl for the ring buffer
 236          */
 237         new_channel->ring_buffer_gpadl_handle = 0;
 238
 239         ret = hv_vmbus_channel_establish_gpadl(new_channel,
 240                 new_channel->outbound.ring_buffer,
 241                 send_ring_buffer_size + recv_ring_buffer_size,
 242                 &new_channel->ring_buffer_gpadl_handle);
 243
 244         /**
 245          * Create and init the channel open message
 246          */
 247         open_info = (hv_vmbus_channel_msg_info*) malloc(
 248                 sizeof(hv_vmbus_channel_msg_info) +
 249                         sizeof(hv_vmbus_channel_open_channel),
 250                 M_DEVBUF,
 251                 M_NOWAIT);
 252         KASSERT(open_info != NULL,
 253             ("Error VMBUS: malloc failed to allocate Open Channel message!"));
 254
 255         if (open_info == NULL)
 256                 return (ENOMEM);
 257
 258         sema_init(&open_info->wait_sema, 0, "Open Info Sema");
 259
 260         open_msg = (hv_vmbus_channel_open_channel*) open_info->msg;
 261         open_msg->header.message_type = HV_CHANNEL_MESSAGE_OPEN_CHANNEL;
 262         open_msg->open_id = new_channel->offer_msg.child_rel_id;
 263         open_msg->child_rel_id = new_channel->offer_msg.child_rel_id;
 264         open_msg->ring_buffer_gpadl_handle =
 265                 new_channel->ring_buffer_gpadl_handle;
 266         open_msg->downstream_ring_buffer_page_offset = send_ring_buffer_size
 267                 >> PAGE_SHIFT;
 268         open_msg->target_vcpu = new_channel->target_vcpu;
 269
 270         if (user_data_len)
 271                 memcpy(open_msg->user_data, user_data, user_data_len);
 272
 273         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 274         TAILQ_INSERT_TAIL(
 275                 &hv_vmbus_g_connection.channel_msg_anchor,
 276                 open_info,
 277                 msg_list_entry);
 278         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 279
 280         ret = hv_vmbus_post_message(
 281                 open_msg, sizeof(hv_vmbus_channel_open_channel));
 282
 283         if (ret != 0)
 284             goto cleanup;
 285
 286         ret = sema_timedwait(&open_info->wait_sema, 5 * hz); /* KYS 5 seconds */
 287
 288         if (ret) {
 289             if(bootverbose)
 290                 printf("VMBUS: channel <%p> open timeout.\n", new_channel);
 291             goto cleanup;
 292         }
 293
 294         if (open_info->response.open_result.status == 0) {
 295             new_channel->state = HV_CHANNEL_OPENED_STATE;
 296             if(bootverbose)
 297                 printf("VMBUS: channel <%p> open success.\n", new_channel);
 298         } else {
 299             if(bootverbose)
 300                 printf("Error VMBUS: channel <%p> open failed - %d!\n",
 301                         new_channel, open_info->response.open_result.status);
 302         }
 303
 304         cleanup:
 305         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 306         TAILQ_REMOVE(
 307                 &hv_vmbus_g_connection.channel_msg_anchor,
 308                 open_info,
 309                 msg_list_entry);
 310         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 311         sema_destroy(&open_info->wait_sema);
 312         free(open_info, M_DEVBUF);
 313
 314         return (ret);
 315 }
 316
 317 /**
 318  * @brief Create a gpadl for the specified buffer
 319  */
 320 static int
 321 vmbus_channel_create_gpadl_header(
 322         void*                           contig_buffer,
 323         uint32_t                        size,   /* page-size multiple */
 324         hv_vmbus_channel_msg_info**     msg_info,
 325         uint32_t*                       message_count)
 326 {
 327         int                             i;
 328         int                             page_count;
 329         unsigned long long              pfn;
 330         uint32_t                        msg_size;
 331         hv_vmbus_channel_gpadl_header*  gpa_header;
 332         hv_vmbus_channel_gpadl_body*    gpadl_body;
 333         hv_vmbus_channel_msg_info*      msg_header;
 334         hv_vmbus_channel_msg_info*      msg_body;
 335
 336         int pfnSum, pfnCount, pfnLeft, pfnCurr, pfnSize;
 337
 338         page_count = size >> PAGE_SHIFT;
 339         pfn = hv_get_phys_addr(contig_buffer) >> PAGE_SHIFT;
 340
 341         /*do we need a gpadl body msg */
 342         pfnSize = HV_MAX_SIZE_CHANNEL_MESSAGE
 343             - sizeof(hv_vmbus_channel_gpadl_header)
 344             - sizeof(hv_gpa_range);
 345         pfnCount = pfnSize / sizeof(uint64_t);
 346
 347         if (page_count > pfnCount) { /* if(we need a gpadl body)        */
 348             /* fill in the header               */
 349             msg_size = sizeof(hv_vmbus_channel_msg_info)
 350                 + sizeof(hv_vmbus_channel_gpadl_header)
 351                 + sizeof(hv_gpa_range)
 352                 + pfnCount * sizeof(uint64_t);
 353             msg_header = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 354             KASSERT(
 355                 msg_header != NULL,
 356                 ("Error VMBUS: malloc failed to allocate Gpadl Message!"));
 357             if (msg_header == NULL)
 358                 return (ENOMEM);
 359
 360             TAILQ_INIT(&msg_header->sub_msg_list_anchor);
 361             msg_header->message_size = msg_size;
 362
 363             gpa_header = (hv_vmbus_channel_gpadl_header*) msg_header->msg;
 364             gpa_header->range_count = 1;
 365             gpa_header->range_buf_len = sizeof(hv_gpa_range)
 366                 + page_count * sizeof(uint64_t);
 367             gpa_header->range[0].byte_offset = 0;
 368             gpa_header->range[0].byte_count = size;
 369             for (i = 0; i < pfnCount; i++) {
 370                 gpa_header->range[0].pfn_array[i] = pfn + i;
 371             }
 372             *msg_info = msg_header;
 373             *message_count = 1;
 374
 375             pfnSum = pfnCount;
 376             pfnLeft = page_count - pfnCount;
 377
 378             /*
 379              *  figure out how many pfns we can fit
 380              */
 381             pfnSize = HV_MAX_SIZE_CHANNEL_MESSAGE
 382                 - sizeof(hv_vmbus_channel_gpadl_body);
 383             pfnCount = pfnSize / sizeof(uint64_t);
 384
 385             /*
 386              * fill in the body
 387              */
 388             while (pfnLeft) {
 389                 if (pfnLeft > pfnCount) {
 390                     pfnCurr = pfnCount;
 391                 } else {
 392                     pfnCurr = pfnLeft;
 393                 }
 394
 395                 msg_size = sizeof(hv_vmbus_channel_msg_info) +
 396                     sizeof(hv_vmbus_channel_gpadl_body) +
 397                     pfnCurr * sizeof(uint64_t);
 398                 msg_body = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 399                 KASSERT(
 400                     msg_body != NULL,
 401                     ("Error VMBUS: malloc failed to allocate Gpadl msg_body!"));
 402                 if (msg_body == NULL)
 403                     return (ENOMEM);
 404
 405                 msg_body->message_size = msg_size;
 406                 (*message_count)++;
 407                 gpadl_body =
 408                     (hv_vmbus_channel_gpadl_body*) msg_body->msg;
 409                 /*
 410                  * gpadl_body->gpadl = kbuffer;
 411                  */
 412                 for (i = 0; i < pfnCurr; i++) {
 413                     gpadl_body->pfn[i] = pfn + pfnSum + i;
 414                 }
 415
 416                 TAILQ_INSERT_TAIL(
 417                     &msg_header->sub_msg_list_anchor,
 418                     msg_body,
 419                     msg_list_entry);
 420                 pfnSum += pfnCurr;
 421                 pfnLeft -= pfnCurr;
 422             }
 423         } else { /* else everything fits in a header */
 424
 425             msg_size = sizeof(hv_vmbus_channel_msg_info) +
 426                 sizeof(hv_vmbus_channel_gpadl_header) +
 427                 sizeof(hv_gpa_range) +
 428                 page_count * sizeof(uint64_t);
 429             msg_header = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 430             KASSERT(
 431                 msg_header != NULL,
 432                 ("Error VMBUS: malloc failed to allocate Gpadl Message!"));
 433             if (msg_header == NULL)
 434                 return (ENOMEM);
 435
 436             msg_header->message_size = msg_size;
 437
 438             gpa_header = (hv_vmbus_channel_gpadl_header*) msg_header->msg;
 439             gpa_header->range_count = 1;
 440             gpa_header->range_buf_len = sizeof(hv_gpa_range) +
 441                 page_count * sizeof(uint64_t);
 442             gpa_header->range[0].byte_offset = 0;
 443             gpa_header->range[0].byte_count = size;
 444             for (i = 0; i < page_count; i++) {
 445                 gpa_header->range[0].pfn_array[i] = pfn + i;
 446             }
 447
 448             *msg_info = msg_header;
 449             *message_count = 1;
 450         }
 451
 452         return (0);
 453 }
 454
 455 /**
 456  * @brief Establish a GPADL for the specified buffer
 457  */
 458 int
 459 hv_vmbus_channel_establish_gpadl(
 460         hv_vmbus_channel*       channel,
 461         void*                   contig_buffer,
 462         uint32_t                size, /* page-size multiple */
 463         uint32_t*               gpadl_handle)
 464
 465 {
 466         int ret = 0;
 467         hv_vmbus_channel_gpadl_header*  gpadl_msg;
 468         hv_vmbus_channel_gpadl_body*    gpadl_body;
 469         hv_vmbus_channel_msg_info*      msg_info;
 470         hv_vmbus_channel_msg_info*      sub_msg_info;
 471         uint32_t                        msg_count;
 472         hv_vmbus_channel_msg_info*      curr;
 473         uint32_t                        next_gpadl_handle;
 474
 475         next_gpadl_handle = atomic_fetchadd_int(
 476             &hv_vmbus_g_connection.next_gpadl_handle, 1);
 477
 478         ret = vmbus_channel_create_gpadl_header(
 479                 contig_buffer, size, &msg_info, &msg_count);
 480
 481         if(ret != 0) {
 482                 /*
 483                  * XXX
 484                  * We can _not_ even revert the above incremental,
 485                  * if multiple GPADL establishments are running
 486                  * parallelly, decrement the global next_gpadl_handle
 487                  * is calling for _big_ trouble.  A better solution
 488                  * is to have a 0-based GPADL id bitmap ...
 489                  */
 490                 return ret;
 491         }
 492
 493         sema_init(&msg_info->wait_sema, 0, "Open Info Sema");
 494         gpadl_msg = (hv_vmbus_channel_gpadl_header*) msg_info->msg;
 495         gpadl_msg->header.message_type = HV_CHANNEL_MESSAGEL_GPADL_HEADER;
 496         gpadl_msg->child_rel_id = channel->offer_msg.child_rel_id;
 497         gpadl_msg->gpadl = next_gpadl_handle;
 498
 499         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 500         TAILQ_INSERT_TAIL(
 501                 &hv_vmbus_g_connection.channel_msg_anchor,
 502                 msg_info,
 503                 msg_list_entry);
 504
 505         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 506
 507         ret = hv_vmbus_post_message(
 508                 gpadl_msg,
 509                 msg_info->message_size -
 510                     (uint32_t) sizeof(hv_vmbus_channel_msg_info));
 511
 512         if (ret != 0)
 513             goto cleanup;
 514
 515         if (msg_count > 1) {
 516             TAILQ_FOREACH(curr,
 517                     &msg_info->sub_msg_list_anchor, msg_list_entry) {
 518                 sub_msg_info = curr;
 519                 gpadl_body =
 520                     (hv_vmbus_channel_gpadl_body*) sub_msg_info->msg;
 521
 522                 gpadl_body->header.message_type =
 523                     HV_CHANNEL_MESSAGE_GPADL_BODY;
 524                 gpadl_body->gpadl = next_gpadl_handle;
 525
 526                 ret = hv_vmbus_post_message(
 527                         gpadl_body,
 528                         sub_msg_info->message_size
 529                             - (uint32_t) sizeof(hv_vmbus_channel_msg_info));
 530                  /* if (the post message failed) give up and clean up */
 531                 if(ret != 0)
 532                     goto cleanup;
 533             }
 534         }
 535
 536         ret = sema_timedwait(&msg_info->wait_sema, 5 * hz); /* KYS 5 seconds*/
 537         if (ret != 0)
 538             goto cleanup;
 539
 540         *gpadl_handle = gpadl_msg->gpadl;
 541
 542 cleanup:
 543
 544         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 545         TAILQ_REMOVE(&hv_vmbus_g_connection.channel_msg_anchor,
 546                 msg_info, msg_list_entry);
 547         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 548
 549         sema_destroy(&msg_info->wait_sema);
 550         free(msg_info, M_DEVBUF);
 551
 552         return (ret);
 553 }
 554
 555 /**
 556  * @brief Teardown the specified GPADL handle
 557  */
 558 int
 559 hv_vmbus_channel_teardown_gpdal(
 560         hv_vmbus_channel*       channel,
 561         uint32_t                gpadl_handle)
 562 {
 563         int                                     ret = 0;
 564         hv_vmbus_channel_gpadl_teardown*        msg;
 565         hv_vmbus_channel_msg_info*              info;
 566
 567         info = (hv_vmbus_channel_msg_info *)
 568                 malloc( sizeof(hv_vmbus_channel_msg_info) +
 569                         sizeof(hv_vmbus_channel_gpadl_teardown),
 570                                 M_DEVBUF, M_NOWAIT);
 571         KASSERT(info != NULL,
 572             ("Error VMBUS: malloc failed to allocate Gpadl Teardown Msg!"));
 573         if (info == NULL) {
 574             ret = ENOMEM;
 575             goto cleanup;
 576         }
 577
 578         sema_init(&info->wait_sema, 0, "Open Info Sema");
 579
 580         msg = (hv_vmbus_channel_gpadl_teardown*) info->msg;
 581
 582         msg->header.message_type = HV_CHANNEL_MESSAGE_GPADL_TEARDOWN;
 583         msg->child_rel_id = channel->offer_msg.child_rel_id;
 584         msg->gpadl = gpadl_handle;
 585
 586         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 587         TAILQ_INSERT_TAIL(&hv_vmbus_g_connection.channel_msg_anchor,
 588                         info, msg_list_entry);
 589         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 590
 591         ret = hv_vmbus_post_message(msg,
 592                         sizeof(hv_vmbus_channel_gpadl_teardown));
 593         if (ret != 0)
 594             goto cleanup;
 595
 596         ret = sema_timedwait(&info->wait_sema, 5 * hz); /* KYS 5 seconds */
 597
 598 cleanup:
 599         /*
 600          * Received a torndown response
 601          */
 602         mtx_lock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 603         TAILQ_REMOVE(&hv_vmbus_g_connection.channel_msg_anchor,
 604                         info, msg_list_entry);
 605         mtx_unlock_spin(&hv_vmbus_g_connection.channel_msg_lock);
 606         sema_destroy(&info->wait_sema);
 607         free(info, M_DEVBUF);
 608
 609         return (ret);
 610 }
 611
 612 static void
 613 hv_vmbus_channel_close_internal(hv_vmbus_channel *channel)
 614 {
 615         int ret = 0;
 616         struct taskqueue *rxq = channel->rxq;
 617         hv_vmbus_channel_close_channel* msg;
 618         hv_vmbus_channel_msg_info* info;
 619
 620         channel->state = HV_CHANNEL_OPEN_STATE;
 621         channel->sc_creation_callback = NULL;
 622
 623         /*
 624          * set rxq to NULL to avoid more requests be scheduled
 625          */
 626         channel->rxq = NULL;
 627         taskqueue_drain(rxq, &channel->channel_task);
 628         channel->on_channel_callback = NULL;
 629
 630         /**
 631          * Send a closing message
 632          */
 633         info = (hv_vmbus_channel_msg_info *)
 634                 malloc( sizeof(hv_vmbus_channel_msg_info) +
 635                         sizeof(hv_vmbus_channel_close_channel),
 636                                 M_DEVBUF, M_NOWAIT);
 637         KASSERT(info != NULL, ("VMBUS: malloc failed hv_vmbus_channel_close!"));
 638         if(info == NULL)
 639             return;
 640
 641         msg = (hv_vmbus_channel_close_channel*) info->msg;
 642         msg->header.message_type = HV_CHANNEL_MESSAGE_CLOSE_CHANNEL;
 643         msg->child_rel_id = channel->offer_msg.child_rel_id;
 644
 645         ret = hv_vmbus_post_message(
 646                 msg, sizeof(hv_vmbus_channel_close_channel));
 647
 648         /* Tear down the gpadl for the channel's ring buffer */
 649         if (channel->ring_buffer_gpadl_handle) {
 650                 hv_vmbus_channel_teardown_gpdal(channel,
 651                         channel->ring_buffer_gpadl_handle);
 652         }
 653
 654         /* TODO: Send a msg to release the childRelId */
 655
 656         /* cleanup the ring buffers for this channel */
 657         hv_ring_buffer_cleanup(&channel->outbound);
 658         hv_ring_buffer_cleanup(&channel->inbound);
 659
 660         contigfree(channel->ring_buffer_pages, channel->ring_buffer_size,
 661             M_DEVBUF);
 662
 663         free(info, M_DEVBUF);
 664 }
 665
 666 /**
 667  * @brief Close the specified channel
 668  */
 669 void
 670 hv_vmbus_channel_close(hv_vmbus_channel *channel)
 671 {
 672         hv_vmbus_channel*       sub_channel;
 673
 674         if (channel->primary_channel != NULL) {
 675                 /*
 676                  * We only close multi-channels when the primary is
 677                  * closed.
 678                  */
 679                 return;
 680         }
 681
 682         /*
 683          * Close all multi-channels first.
 684          */
 685         TAILQ_FOREACH(sub_channel, &channel->sc_list_anchor,
 686             sc_list_entry) {
 687                 if (sub_channel->state != HV_CHANNEL_OPENED_STATE)
 688                         continue;
 689                 hv_vmbus_channel_close_internal(sub_channel);
 690         }
 691         /*
 692          * Then close the primary channel.
 693          */
 694         hv_vmbus_channel_close_internal(channel);
 695 }
 696
 697 /**
 698  * @brief Send the specified buffer on the given channel
 699  */
 700 int
 701 hv_vmbus_channel_send_packet(
 702         hv_vmbus_channel*       channel,
 703         void*                   buffer,
 704         uint32_t                buffer_len,
 705         uint64_t                request_id,
 706         hv_vmbus_packet_type    type,
 707         uint32_t                flags)
 708 {
 709         int                     ret = 0;
 710         hv_vm_packet_descriptor desc;
 711         uint32_t                packet_len;
 712         uint64_t                aligned_data;
 713         uint32_t                packet_len_aligned;
 714         boolean_t               need_sig;
 715         hv_vmbus_sg_buffer_list buffer_list[3];
 716
 717         packet_len = sizeof(hv_vm_packet_descriptor) + buffer_len;
 718         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 719         aligned_data = 0;
 720
 721         /* Setup the descriptor */
 722         desc.type = type;   /* HV_VMBUS_PACKET_TYPE_DATA_IN_BAND;             */
 723         desc.flags = flags; /* HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED */
 724                             /* in 8-bytes granularity */
 725         desc.data_offset8 = sizeof(hv_vm_packet_descriptor) >> 3;
 726         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 727         desc.transaction_id = request_id;
 728
 729         buffer_list[0].data = &desc;
 730         buffer_list[0].length = sizeof(hv_vm_packet_descriptor);
 731
 732         buffer_list[1].data = buffer;
 733         buffer_list[1].length = buffer_len;
 734
 735         buffer_list[2].data = &aligned_data;
 736         buffer_list[2].length = packet_len_aligned - packet_len;
 737
 738         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 739             &need_sig);
 740
 741         /* TODO: We should determine if this is optional */
 742         if (ret == 0 && need_sig) {
 743                 vmbus_channel_set_event(channel);
 744         }
 745
 746         return (ret);
 747 }
 748
 749 /**
 750  * @brief Send a range of single-page buffer packets using
 751  * a GPADL Direct packet type
 752  */
 753 int
 754 hv_vmbus_channel_send_packet_pagebuffer(
 755         hv_vmbus_channel*       channel,
 756         hv_vmbus_page_buffer    page_buffers[],
 757         uint32_t                page_count,
 758         void*                   buffer,
 759         uint32_t                buffer_len,
 760         uint64_t                request_id)
 761 {
 762
 763         int                                     ret = 0;
 764         boolean_t                               need_sig;
 765         uint32_t                                packet_len;
 766         uint32_t                                page_buflen;
 767         uint32_t                                packetLen_aligned;
 768         hv_vmbus_sg_buffer_list                 buffer_list[4];
 769         hv_vmbus_channel_packet_page_buffer     desc;
 770         uint32_t                                descSize;
 771         uint64_t                                alignedData = 0;
 772
 773         if (page_count > HV_MAX_PAGE_BUFFER_COUNT)
 774                 return (EINVAL);
 775
 776         /*
 777          * Adjust the size down since hv_vmbus_channel_packet_page_buffer
 778          *  is the largest size we support
 779          */
 780         descSize = __offsetof(hv_vmbus_channel_packet_page_buffer, range);
 781         page_buflen = sizeof(hv_vmbus_page_buffer) * page_count;
 782         packet_len = descSize + page_buflen + buffer_len;
 783         packetLen_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 784
 785         /* Setup the descriptor */
 786         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 787         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 788         /* in 8-bytes granularity */
 789         desc.data_offset8 = (descSize + page_buflen) >> 3;
 790         desc.length8 = (uint16_t) (packetLen_aligned >> 3);
 791         desc.transaction_id = request_id;
 792         desc.range_count = page_count;
 793
 794         buffer_list[0].data = &desc;
 795         buffer_list[0].length = descSize;
 796
 797         buffer_list[1].data = page_buffers;
 798         buffer_list[1].length = page_buflen;
 799
 800         buffer_list[2].data = buffer;
 801         buffer_list[2].length = buffer_len;
 802
 803         buffer_list[3].data = &alignedData;
 804         buffer_list[3].length = packetLen_aligned - packet_len;
 805
 806         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 4,
 807             &need_sig);
 808
 809         /* TODO: We should determine if this is optional */
 810         if (ret == 0 && need_sig) {
 811                 vmbus_channel_set_event(channel);
 812         }
 813
 814         return (ret);
 815 }
 816
 817 /**
 818  * @brief Send a multi-page buffer packet using a GPADL Direct packet type
 819  */
 820 int
 821 hv_vmbus_channel_send_packet_multipagebuffer(
 822         hv_vmbus_channel*               channel,
 823         hv_vmbus_multipage_buffer*      multi_page_buffer,
 824         void*                           buffer,
 825         uint32_t                        buffer_len,
 826         uint64_t                        request_id)
 827 {
 828
 829         int                     ret = 0;
 830         uint32_t                desc_size;
 831         boolean_t               need_sig;
 832         uint32_t                packet_len;
 833         uint32_t                packet_len_aligned;
 834         uint32_t                pfn_count;
 835         uint64_t                aligned_data = 0;
 836         hv_vmbus_sg_buffer_list buffer_list[3];
 837         hv_vmbus_channel_packet_multipage_buffer desc;
 838
 839         pfn_count =
 840             HV_NUM_PAGES_SPANNED(
 841                     multi_page_buffer->offset,
 842                     multi_page_buffer->length);
 843
 844         if ((pfn_count == 0) || (pfn_count > HV_MAX_MULTIPAGE_BUFFER_COUNT))
 845             return (EINVAL);
 846         /*
 847          * Adjust the size down since hv_vmbus_channel_packet_multipage_buffer
 848          * is the largest size we support
 849          */
 850         desc_size =
 851             sizeof(hv_vmbus_channel_packet_multipage_buffer) -
 852                     ((HV_MAX_MULTIPAGE_BUFFER_COUNT - pfn_count) *
 853                         sizeof(uint64_t));
 854         packet_len = desc_size + buffer_len;
 855         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 856
 857         /*
 858          * Setup the descriptor
 859          */
 860         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 861         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 862         desc.data_offset8 = desc_size >> 3; /* in 8-bytes granularity */
 863         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 864         desc.transaction_id = request_id;
 865         desc.range_count = 1;
 866
 867         desc.range.length = multi_page_buffer->length;
 868         desc.range.offset = multi_page_buffer->offset;
 869
 870         memcpy(desc.range.pfn_array, multi_page_buffer->pfn_array,
 871                 pfn_count * sizeof(uint64_t));
 872
 873         buffer_list[0].data = &desc;
 874         buffer_list[0].length = desc_size;
 875
 876         buffer_list[1].data = buffer;
 877         buffer_list[1].length = buffer_len;
 878
 879         buffer_list[2].data = &aligned_data;
 880         buffer_list[2].length = packet_len_aligned - packet_len;
 881
 882         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 883             &need_sig);
 884
 885         /* TODO: We should determine if this is optional */
 886         if (ret == 0 && need_sig) {
 887             vmbus_channel_set_event(channel);
 888         }
 889
 890         return (ret);
 891 }
 892
 893 /**
 894  * @brief Retrieve the user packet on the specified channel
 895  */
 896 int
 897 hv_vmbus_channel_recv_packet(
 898         hv_vmbus_channel*       channel,
 899         void*                   Buffer,
 900         uint32_t                buffer_len,
 901         uint32_t*               buffer_actual_len,
 902         uint64_t*               request_id)
 903 {
 904         int                     ret;
 905         uint32_t                user_len;
 906         uint32_t                packet_len;
 907         hv_vm_packet_descriptor desc;
 908
 909         *buffer_actual_len = 0;
 910         *request_id = 0;
 911
 912         ret = hv_ring_buffer_peek(&channel->inbound, &desc,
 913                 sizeof(hv_vm_packet_descriptor));
 914         if (ret != 0)
 915                 return (0);
 916
 917         packet_len = desc.length8 << 3;
 918         user_len = packet_len - (desc.data_offset8 << 3);
 919
 920         *buffer_actual_len = user_len;
 921
 922         if (user_len > buffer_len)
 923                 return (EINVAL);
 924
 925         *request_id = desc.transaction_id;
 926
 927         /* Copy over the packet to the user buffer */
 928         ret = hv_ring_buffer_read(&channel->inbound, Buffer, user_len,
 929                 (desc.data_offset8 << 3));
 930
 931         return (0);
 932 }
 933
 934 /**
 935  * @brief Retrieve the raw packet on the specified channel
 936  */
 937 int
 938 hv_vmbus_channel_recv_packet_raw(
 939         hv_vmbus_channel*       channel,
 940         void*                   buffer,
 941         uint32_t                buffer_len,
 942         uint32_t*               buffer_actual_len,
 943         uint64_t*               request_id)
 944 {
 945         int             ret;
 946         uint32_t        packetLen;
 947         hv_vm_packet_descriptor desc;
 948
 949         *buffer_actual_len = 0;
 950         *request_id = 0;
 951
 952         ret = hv_ring_buffer_peek(
 953                 &channel->inbound, &desc,
 954                 sizeof(hv_vm_packet_descriptor));
 955
 956         if (ret != 0)
 957             return (0);
 958
 959         packetLen = desc.length8 << 3;
 960         *buffer_actual_len = packetLen;
 961
 962         if (packetLen > buffer_len)
 963             return (ENOBUFS);
 964
 965         *request_id = desc.transaction_id;
 966
 967         /* Copy over the entire packet to the user buffer */
 968         ret = hv_ring_buffer_read(&channel->inbound, buffer, packetLen, 0);
 969
 970         return (0);
 971 }
 972
 973
 974 /**
 975  * Process a channel event notification
 976  */
 977 static void
 978 VmbusProcessChannelEvent(void* context, int pending)
 979 {
 980         void* arg;
 981         uint32_t bytes_to_read;
 982         hv_vmbus_channel* channel = (hv_vmbus_channel*)context;
 983         boolean_t is_batched_reading;
 984
 985         /**
 986          * Find the channel based on this relid and invokes
 987          * the channel callback to process the event
 988          */
 989
 990         if (channel == NULL) {
 991                 return;
 992         }
 993         /**
 994          * To deal with the race condition where we might
 995          * receive a packet while the relevant driver is
 996          * being unloaded, dispatch the callback while
 997          * holding the channel lock. The unloading driver
 998          * will acquire the same channel lock to set the
 999          * callback to NULL. This closes the window.
1000          */
1001
1002         if (channel->on_channel_callback != NULL) {
1003                 arg = channel->channel_callback_context;
1004                 is_batched_reading = channel->batched_reading;
1005                 /*
1006                  * Optimize host to guest signaling by ensuring:
1007                  * 1. While reading the channel, we disable interrupts from
1008                  *    host.
1009                  * 2. Ensure that we process all posted messages from the host
1010                  *    before returning from this callback.
1011                  * 3. Once we return, enable signaling from the host. Once this
1012                  *    state is set we check to see if additional packets are
1013                  *    available to read. In this case we repeat the process.
1014                  */
1015                 do {
1016                         if (is_batched_reading)
1017                                 hv_ring_buffer_read_begin(&channel->inbound);
1018
1019                         channel->on_channel_callback(arg);
1020
1021                         if (is_batched_reading)
1022                                 bytes_to_read =
1023                                     hv_ring_buffer_read_end(&channel->inbound);
1024                         else
1025                                 bytes_to_read = 0;
1026                 } while (is_batched_reading && (bytes_to_read != 0));
1027         }
1028 }