sys/dev/hyperv/vmbus/hv_channel.c

   1 /*-
   2  * Copyright (c) 2009-2012,2016 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 <dev/hyperv/vmbus/hv_vmbus_priv.h>
  46 #include <dev/hyperv/vmbus/vmbus_var.h>
  47
  48 static int      vmbus_channel_create_gpadl_header(
  49                         /* must be phys and virt contiguous*/
  50                         void*                           contig_buffer,
  51                         /* page-size multiple */
  52                         uint32_t                        size,
  53                         hv_vmbus_channel_msg_info**     msg_info,
  54                         uint32_t*                       message_count);
  55
  56 static void     vmbus_channel_set_event(hv_vmbus_channel* channel);
  57 static void     VmbusProcessChannelEvent(void* channel, int pending);
  58
  59 /**
  60  *  @brief Trigger an event notification on the specified channel
  61  */
  62 static void
  63 vmbus_channel_set_event(hv_vmbus_channel *channel)
  64 {
  65         hv_vmbus_monitor_page *monitor_page;
  66
  67         if (channel->offer_msg.monitor_allocated) {
  68                 /* Each uint32_t represents 32 channels */
  69                 synch_set_bit((channel->offer_msg.child_rel_id & 31),
  70                         ((uint32_t *)hv_vmbus_g_connection.send_interrupt_page
  71                                 + ((channel->offer_msg.child_rel_id >> 5))));
  72
  73                 monitor_page = (hv_vmbus_monitor_page *)
  74                         hv_vmbus_g_connection.monitor_page_2;
  75
  76                 synch_set_bit(channel->monitor_bit,
  77                         (uint32_t *)&monitor_page->
  78                                 trigger_group[channel->monitor_group].u.pending);
  79         } else {
  80                 hv_vmbus_set_event(channel);
  81         }
  82
  83 }
  84
  85 static int
  86 vmbus_channel_sysctl_monalloc(SYSCTL_HANDLER_ARGS)
  87 {
  88         struct hv_vmbus_channel *chan = arg1;
  89         int alloc = 0;
  90
  91         if (chan->offer_msg.monitor_allocated)
  92                 alloc = 1;
  93         return sysctl_handle_int(oidp, &alloc, 0, req);
  94 }
  95
  96 static void
  97 vmbus_channel_sysctl_create(hv_vmbus_channel* channel)
  98 {
  99         device_t dev;
 100         struct sysctl_oid *devch_sysctl;
 101         struct sysctl_oid *devch_id_sysctl, *devch_sub_sysctl;
 102         struct sysctl_oid *devch_id_in_sysctl, *devch_id_out_sysctl;
 103         struct sysctl_ctx_list *ctx;
 104         uint32_t ch_id;
 105         uint16_t sub_ch_id;
 106         char name[16];
 107
 108         hv_vmbus_channel* primary_ch = channel->primary_channel;
 109
 110         if (primary_ch == NULL) {
 111                 dev = channel->device->device;
 112                 ch_id = channel->offer_msg.child_rel_id;
 113         } else {
 114                 dev = primary_ch->device->device;
 115                 ch_id = primary_ch->offer_msg.child_rel_id;
 116                 sub_ch_id = channel->offer_msg.offer.sub_channel_index;
 117         }
 118         ctx = device_get_sysctl_ctx(dev);
 119         /* This creates dev.DEVNAME.DEVUNIT.channel tree */
 120         devch_sysctl = SYSCTL_ADD_NODE(ctx,
 121                     SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 122                     OID_AUTO, "channel", CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 123         /* This creates dev.DEVNAME.DEVUNIT.channel.CHANID tree */
 124         snprintf(name, sizeof(name), "%d", ch_id);
 125         devch_id_sysctl = SYSCTL_ADD_NODE(ctx,
 126                     SYSCTL_CHILDREN(devch_sysctl),
 127                     OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 128
 129         if (primary_ch != NULL) {
 130                 devch_sub_sysctl = SYSCTL_ADD_NODE(ctx,
 131                         SYSCTL_CHILDREN(devch_id_sysctl),
 132                         OID_AUTO, "sub", CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 133                 snprintf(name, sizeof(name), "%d", sub_ch_id);
 134                 devch_id_sysctl = SYSCTL_ADD_NODE(ctx,
 135                         SYSCTL_CHILDREN(devch_sub_sysctl),
 136                         OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 137
 138                 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(devch_id_sysctl),
 139                     OID_AUTO, "chanid", CTLFLAG_RD,
 140                     &channel->offer_msg.child_rel_id, 0, "channel id");
 141         }
 142         SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(devch_id_sysctl), OID_AUTO,
 143             "cpu", CTLFLAG_RD, &channel->target_cpu, 0, "owner CPU id");
 144         SYSCTL_ADD_PROC(ctx, SYSCTL_CHILDREN(devch_id_sysctl), OID_AUTO,
 145             "monitor_allocated", CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE,
 146             channel, 0, vmbus_channel_sysctl_monalloc, "I",
 147             "is monitor allocated to this channel");
 148
 149         devch_id_in_sysctl = SYSCTL_ADD_NODE(ctx,
 150                     SYSCTL_CHILDREN(devch_id_sysctl),
 151                     OID_AUTO,
 152                     "in",
 153                     CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 154         devch_id_out_sysctl = SYSCTL_ADD_NODE(ctx,
 155                     SYSCTL_CHILDREN(devch_id_sysctl),
 156                     OID_AUTO,
 157                     "out",
 158                     CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "");
 159         hv_ring_buffer_stat(ctx,
 160                 SYSCTL_CHILDREN(devch_id_in_sysctl),
 161                 &(channel->inbound),
 162                 "inbound ring buffer stats");
 163         hv_ring_buffer_stat(ctx,
 164                 SYSCTL_CHILDREN(devch_id_out_sysctl),
 165                 &(channel->outbound),
 166                 "outbound ring buffer stats");
 167 }
 168
 169 /**
 170  * @brief Open the specified channel
 171  */
 172 int
 173 hv_vmbus_channel_open(
 174         hv_vmbus_channel*               new_channel,
 175         uint32_t                        send_ring_buffer_size,
 176         uint32_t                        recv_ring_buffer_size,
 177         void*                           user_data,
 178         uint32_t                        user_data_len,
 179         hv_vmbus_pfn_channel_callback   pfn_on_channel_callback,
 180         void*                           context)
 181 {
 182
 183         int ret = 0;
 184         void *in, *out;
 185         hv_vmbus_channel_open_channel*  open_msg;
 186         hv_vmbus_channel_msg_info*      open_info;
 187
 188         mtx_lock(&new_channel->sc_lock);
 189         if (new_channel->state == HV_CHANNEL_OPEN_STATE) {
 190             new_channel->state = HV_CHANNEL_OPENING_STATE;
 191         } else {
 192             mtx_unlock(&new_channel->sc_lock);
 193             if(bootverbose)
 194                 printf("VMBUS: Trying to open channel <%p> which in "
 195                     "%d state.\n", new_channel, new_channel->state);
 196             return (EINVAL);
 197         }
 198         mtx_unlock(&new_channel->sc_lock);
 199
 200         new_channel->on_channel_callback = pfn_on_channel_callback;
 201         new_channel->channel_callback_context = context;
 202
 203         vmbus_on_channel_open(new_channel);
 204
 205         new_channel->rxq = VMBUS_PCPU_GET(vmbus_get_softc(), event_tq,
 206             new_channel->target_cpu);
 207         TASK_INIT(&new_channel->channel_task, 0, VmbusProcessChannelEvent, new_channel);
 208
 209         /* Allocate the ring buffer */
 210         out = contigmalloc((send_ring_buffer_size + recv_ring_buffer_size),
 211             M_DEVBUF, M_ZERO, 0UL, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
 212         KASSERT(out != NULL,
 213             ("Error VMBUS: contigmalloc failed to allocate Ring Buffer!"));
 214         if (out == NULL)
 215                 return (ENOMEM);
 216
 217         in = ((uint8_t *) out + send_ring_buffer_size);
 218
 219         new_channel->ring_buffer_pages = out;
 220         new_channel->ring_buffer_page_count = (send_ring_buffer_size +
 221             recv_ring_buffer_size) >> PAGE_SHIFT;
 222         new_channel->ring_buffer_size = send_ring_buffer_size +
 223             recv_ring_buffer_size;
 224
 225         hv_vmbus_ring_buffer_init(
 226                 &new_channel->outbound,
 227                 out,
 228                 send_ring_buffer_size);
 229
 230         hv_vmbus_ring_buffer_init(
 231                 &new_channel->inbound,
 232                 in,
 233                 recv_ring_buffer_size);
 234
 235         /* Create sysctl tree for this channel */
 236         vmbus_channel_sysctl_create(new_channel);
 237
 238         /**
 239          * Establish the gpadl for the ring buffer
 240          */
 241         new_channel->ring_buffer_gpadl_handle = 0;
 242
 243         ret = hv_vmbus_channel_establish_gpadl(new_channel,
 244                 new_channel->outbound.ring_buffer,
 245                 send_ring_buffer_size + recv_ring_buffer_size,
 246                 &new_channel->ring_buffer_gpadl_handle);
 247
 248         /**
 249          * Create and init the channel open message
 250          */
 251         open_info = (hv_vmbus_channel_msg_info*) malloc(
 252                 sizeof(hv_vmbus_channel_msg_info) +
 253                         sizeof(hv_vmbus_channel_open_channel),
 254                 M_DEVBUF,
 255                 M_NOWAIT);
 256         KASSERT(open_info != NULL,
 257             ("Error VMBUS: malloc failed to allocate Open Channel message!"));
 258
 259         if (open_info == NULL)
 260                 return (ENOMEM);
 261
 262         sema_init(&open_info->wait_sema, 0, "Open Info Sema");
 263
 264         open_msg = (hv_vmbus_channel_open_channel*) open_info->msg;
 265         open_msg->header.message_type = HV_CHANNEL_MESSAGE_OPEN_CHANNEL;
 266         open_msg->open_id = new_channel->offer_msg.child_rel_id;
 267         open_msg->child_rel_id = new_channel->offer_msg.child_rel_id;
 268         open_msg->ring_buffer_gpadl_handle =
 269                 new_channel->ring_buffer_gpadl_handle;
 270         open_msg->downstream_ring_buffer_page_offset = send_ring_buffer_size
 271                 >> PAGE_SHIFT;
 272         open_msg->target_vcpu = new_channel->target_vcpu;
 273
 274         if (user_data_len)
 275                 memcpy(open_msg->user_data, user_data, user_data_len);
 276
 277         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 278         TAILQ_INSERT_TAIL(
 279                 &hv_vmbus_g_connection.channel_msg_anchor,
 280                 open_info,
 281                 msg_list_entry);
 282         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 283
 284         ret = hv_vmbus_post_message(
 285                 open_msg, sizeof(hv_vmbus_channel_open_channel));
 286
 287         if (ret != 0)
 288             goto cleanup;
 289
 290         ret = sema_timedwait(&open_info->wait_sema, 5 * hz); /* KYS 5 seconds */
 291
 292         if (ret) {
 293             if(bootverbose)
 294                 printf("VMBUS: channel <%p> open timeout.\n", new_channel);
 295             goto cleanup;
 296         }
 297
 298         if (open_info->response.open_result.status == 0) {
 299             new_channel->state = HV_CHANNEL_OPENED_STATE;
 300             if(bootverbose)
 301                 printf("VMBUS: channel <%p> open success.\n", new_channel);
 302         } else {
 303             if(bootverbose)
 304                 printf("Error VMBUS: channel <%p> open failed - %d!\n",
 305                         new_channel, open_info->response.open_result.status);
 306         }
 307
 308         cleanup:
 309         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 310         TAILQ_REMOVE(
 311                 &hv_vmbus_g_connection.channel_msg_anchor,
 312                 open_info,
 313                 msg_list_entry);
 314         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 315         sema_destroy(&open_info->wait_sema);
 316         free(open_info, M_DEVBUF);
 317
 318         return (ret);
 319 }
 320
 321 /**
 322  * @brief Create a gpadl for the specified buffer
 323  */
 324 static int
 325 vmbus_channel_create_gpadl_header(
 326         void*                           contig_buffer,
 327         uint32_t                        size,   /* page-size multiple */
 328         hv_vmbus_channel_msg_info**     msg_info,
 329         uint32_t*                       message_count)
 330 {
 331         int                             i;
 332         int                             page_count;
 333         unsigned long long              pfn;
 334         uint32_t                        msg_size;
 335         hv_vmbus_channel_gpadl_header*  gpa_header;
 336         hv_vmbus_channel_gpadl_body*    gpadl_body;
 337         hv_vmbus_channel_msg_info*      msg_header;
 338         hv_vmbus_channel_msg_info*      msg_body;
 339
 340         int pfnSum, pfnCount, pfnLeft, pfnCurr, pfnSize;
 341
 342         page_count = size >> PAGE_SHIFT;
 343         pfn = hv_get_phys_addr(contig_buffer) >> PAGE_SHIFT;
 344
 345         /*do we need a gpadl body msg */
 346         pfnSize = HV_MAX_SIZE_CHANNEL_MESSAGE
 347             - sizeof(hv_vmbus_channel_gpadl_header)
 348             - sizeof(hv_gpa_range);
 349         pfnCount = pfnSize / sizeof(uint64_t);
 350
 351         if (page_count > pfnCount) { /* if(we need a gpadl body)        */
 352             /* fill in the header               */
 353             msg_size = sizeof(hv_vmbus_channel_msg_info)
 354                 + sizeof(hv_vmbus_channel_gpadl_header)
 355                 + sizeof(hv_gpa_range)
 356                 + pfnCount * sizeof(uint64_t);
 357             msg_header = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 358             KASSERT(
 359                 msg_header != NULL,
 360                 ("Error VMBUS: malloc failed to allocate Gpadl Message!"));
 361             if (msg_header == NULL)
 362                 return (ENOMEM);
 363
 364             TAILQ_INIT(&msg_header->sub_msg_list_anchor);
 365             msg_header->message_size = msg_size;
 366
 367             gpa_header = (hv_vmbus_channel_gpadl_header*) msg_header->msg;
 368             gpa_header->range_count = 1;
 369             gpa_header->range_buf_len = sizeof(hv_gpa_range)
 370                 + page_count * sizeof(uint64_t);
 371             gpa_header->range[0].byte_offset = 0;
 372             gpa_header->range[0].byte_count = size;
 373             for (i = 0; i < pfnCount; i++) {
 374                 gpa_header->range[0].pfn_array[i] = pfn + i;
 375             }
 376             *msg_info = msg_header;
 377             *message_count = 1;
 378
 379             pfnSum = pfnCount;
 380             pfnLeft = page_count - pfnCount;
 381
 382             /*
 383              *  figure out how many pfns we can fit
 384              */
 385             pfnSize = HV_MAX_SIZE_CHANNEL_MESSAGE
 386                 - sizeof(hv_vmbus_channel_gpadl_body);
 387             pfnCount = pfnSize / sizeof(uint64_t);
 388
 389             /*
 390              * fill in the body
 391              */
 392             while (pfnLeft) {
 393                 if (pfnLeft > pfnCount) {
 394                     pfnCurr = pfnCount;
 395                 } else {
 396                     pfnCurr = pfnLeft;
 397                 }
 398
 399                 msg_size = sizeof(hv_vmbus_channel_msg_info) +
 400                     sizeof(hv_vmbus_channel_gpadl_body) +
 401                     pfnCurr * sizeof(uint64_t);
 402                 msg_body = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 403                 KASSERT(
 404                     msg_body != NULL,
 405                     ("Error VMBUS: malloc failed to allocate Gpadl msg_body!"));
 406                 if (msg_body == NULL)
 407                     return (ENOMEM);
 408
 409                 msg_body->message_size = msg_size;
 410                 (*message_count)++;
 411                 gpadl_body =
 412                     (hv_vmbus_channel_gpadl_body*) msg_body->msg;
 413                 /*
 414                  * gpadl_body->gpadl = kbuffer;
 415                  */
 416                 for (i = 0; i < pfnCurr; i++) {
 417                     gpadl_body->pfn[i] = pfn + pfnSum + i;
 418                 }
 419
 420                 TAILQ_INSERT_TAIL(
 421                     &msg_header->sub_msg_list_anchor,
 422                     msg_body,
 423                     msg_list_entry);
 424                 pfnSum += pfnCurr;
 425                 pfnLeft -= pfnCurr;
 426             }
 427         } else { /* else everything fits in a header */
 428
 429             msg_size = sizeof(hv_vmbus_channel_msg_info) +
 430                 sizeof(hv_vmbus_channel_gpadl_header) +
 431                 sizeof(hv_gpa_range) +
 432                 page_count * sizeof(uint64_t);
 433             msg_header = malloc(msg_size, M_DEVBUF, M_NOWAIT | M_ZERO);
 434             KASSERT(
 435                 msg_header != NULL,
 436                 ("Error VMBUS: malloc failed to allocate Gpadl Message!"));
 437             if (msg_header == NULL)
 438                 return (ENOMEM);
 439
 440             msg_header->message_size = msg_size;
 441
 442             gpa_header = (hv_vmbus_channel_gpadl_header*) msg_header->msg;
 443             gpa_header->range_count = 1;
 444             gpa_header->range_buf_len = sizeof(hv_gpa_range) +
 445                 page_count * sizeof(uint64_t);
 446             gpa_header->range[0].byte_offset = 0;
 447             gpa_header->range[0].byte_count = size;
 448             for (i = 0; i < page_count; i++) {
 449                 gpa_header->range[0].pfn_array[i] = pfn + i;
 450             }
 451
 452             *msg_info = msg_header;
 453             *message_count = 1;
 454         }
 455
 456         return (0);
 457 }
 458
 459 /**
 460  * @brief Establish a GPADL for the specified buffer
 461  */
 462 int
 463 hv_vmbus_channel_establish_gpadl(
 464         hv_vmbus_channel*       channel,
 465         void*                   contig_buffer,
 466         uint32_t                size, /* page-size multiple */
 467         uint32_t*               gpadl_handle)
 468
 469 {
 470         int ret = 0;
 471         hv_vmbus_channel_gpadl_header*  gpadl_msg;
 472         hv_vmbus_channel_gpadl_body*    gpadl_body;
 473         hv_vmbus_channel_msg_info*      msg_info;
 474         hv_vmbus_channel_msg_info*      sub_msg_info;
 475         uint32_t                        msg_count;
 476         hv_vmbus_channel_msg_info*      curr;
 477         uint32_t                        next_gpadl_handle;
 478
 479         next_gpadl_handle = atomic_fetchadd_int(
 480             &hv_vmbus_g_connection.next_gpadl_handle, 1);
 481
 482         ret = vmbus_channel_create_gpadl_header(
 483                 contig_buffer, size, &msg_info, &msg_count);
 484
 485         if(ret != 0) {
 486                 /*
 487                  * XXX
 488                  * We can _not_ even revert the above incremental,
 489                  * if multiple GPADL establishments are running
 490                  * parallelly, decrement the global next_gpadl_handle
 491                  * is calling for _big_ trouble.  A better solution
 492                  * is to have a 0-based GPADL id bitmap ...
 493                  */
 494                 return ret;
 495         }
 496
 497         sema_init(&msg_info->wait_sema, 0, "Open Info Sema");
 498         gpadl_msg = (hv_vmbus_channel_gpadl_header*) msg_info->msg;
 499         gpadl_msg->header.message_type = HV_CHANNEL_MESSAGEL_GPADL_HEADER;
 500         gpadl_msg->child_rel_id = channel->offer_msg.child_rel_id;
 501         gpadl_msg->gpadl = next_gpadl_handle;
 502
 503         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 504         TAILQ_INSERT_TAIL(
 505                 &hv_vmbus_g_connection.channel_msg_anchor,
 506                 msg_info,
 507                 msg_list_entry);
 508
 509         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 510
 511         ret = hv_vmbus_post_message(
 512                 gpadl_msg,
 513                 msg_info->message_size -
 514                     (uint32_t) sizeof(hv_vmbus_channel_msg_info));
 515
 516         if (ret != 0)
 517             goto cleanup;
 518
 519         if (msg_count > 1) {
 520             TAILQ_FOREACH(curr,
 521                     &msg_info->sub_msg_list_anchor, msg_list_entry) {
 522                 sub_msg_info = curr;
 523                 gpadl_body =
 524                     (hv_vmbus_channel_gpadl_body*) sub_msg_info->msg;
 525
 526                 gpadl_body->header.message_type =
 527                     HV_CHANNEL_MESSAGE_GPADL_BODY;
 528                 gpadl_body->gpadl = next_gpadl_handle;
 529
 530                 ret = hv_vmbus_post_message(
 531                         gpadl_body,
 532                         sub_msg_info->message_size
 533                             - (uint32_t) sizeof(hv_vmbus_channel_msg_info));
 534                  /* if (the post message failed) give up and clean up */
 535                 if(ret != 0)
 536                     goto cleanup;
 537             }
 538         }
 539
 540         ret = sema_timedwait(&msg_info->wait_sema, 5 * hz); /* KYS 5 seconds*/
 541         if (ret != 0)
 542             goto cleanup;
 543
 544         *gpadl_handle = gpadl_msg->gpadl;
 545
 546 cleanup:
 547
 548         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 549         TAILQ_REMOVE(&hv_vmbus_g_connection.channel_msg_anchor,
 550                 msg_info, msg_list_entry);
 551         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 552
 553         sema_destroy(&msg_info->wait_sema);
 554         free(msg_info, M_DEVBUF);
 555
 556         return (ret);
 557 }
 558
 559 /**
 560  * @brief Teardown the specified GPADL handle
 561  */
 562 int
 563 hv_vmbus_channel_teardown_gpdal(
 564         hv_vmbus_channel*       channel,
 565         uint32_t                gpadl_handle)
 566 {
 567         int                                     ret = 0;
 568         hv_vmbus_channel_gpadl_teardown*        msg;
 569         hv_vmbus_channel_msg_info*              info;
 570
 571         info = (hv_vmbus_channel_msg_info *)
 572                 malloc( sizeof(hv_vmbus_channel_msg_info) +
 573                         sizeof(hv_vmbus_channel_gpadl_teardown),
 574                                 M_DEVBUF, M_NOWAIT);
 575         KASSERT(info != NULL,
 576             ("Error VMBUS: malloc failed to allocate Gpadl Teardown Msg!"));
 577         if (info == NULL) {
 578             ret = ENOMEM;
 579             goto cleanup;
 580         }
 581
 582         sema_init(&info->wait_sema, 0, "Open Info Sema");
 583
 584         msg = (hv_vmbus_channel_gpadl_teardown*) info->msg;
 585
 586         msg->header.message_type = HV_CHANNEL_MESSAGE_GPADL_TEARDOWN;
 587         msg->child_rel_id = channel->offer_msg.child_rel_id;
 588         msg->gpadl = gpadl_handle;
 589
 590         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 591         TAILQ_INSERT_TAIL(&hv_vmbus_g_connection.channel_msg_anchor,
 592                         info, msg_list_entry);
 593         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 594
 595         ret = hv_vmbus_post_message(msg,
 596                         sizeof(hv_vmbus_channel_gpadl_teardown));
 597         if (ret != 0)
 598             goto cleanup;
 599
 600         ret = sema_timedwait(&info->wait_sema, 5 * hz); /* KYS 5 seconds */
 601
 602 cleanup:
 603         /*
 604          * Received a torndown response
 605          */
 606         mtx_lock(&hv_vmbus_g_connection.channel_msg_lock);
 607         TAILQ_REMOVE(&hv_vmbus_g_connection.channel_msg_anchor,
 608                         info, msg_list_entry);
 609         mtx_unlock(&hv_vmbus_g_connection.channel_msg_lock);
 610         sema_destroy(&info->wait_sema);
 611         free(info, M_DEVBUF);
 612
 613         return (ret);
 614 }
 615
 616 static void
 617 hv_vmbus_channel_close_internal(hv_vmbus_channel *channel)
 618 {
 619         int ret = 0;
 620         struct taskqueue *rxq = channel->rxq;
 621         hv_vmbus_channel_close_channel* msg;
 622         hv_vmbus_channel_msg_info* info;
 623
 624         channel->state = HV_CHANNEL_OPEN_STATE;
 625
 626         /*
 627          * set rxq to NULL to avoid more requests be scheduled
 628          */
 629         channel->rxq = NULL;
 630         taskqueue_drain(rxq, &channel->channel_task);
 631         channel->on_channel_callback = NULL;
 632
 633         /**
 634          * Send a closing message
 635          */
 636         info = (hv_vmbus_channel_msg_info *)
 637                 malloc( sizeof(hv_vmbus_channel_msg_info) +
 638                         sizeof(hv_vmbus_channel_close_channel),
 639                                 M_DEVBUF, M_NOWAIT);
 640         KASSERT(info != NULL, ("VMBUS: malloc failed hv_vmbus_channel_close!"));
 641         if(info == NULL)
 642             return;
 643
 644         msg = (hv_vmbus_channel_close_channel*) info->msg;
 645         msg->header.message_type = HV_CHANNEL_MESSAGE_CLOSE_CHANNEL;
 646         msg->child_rel_id = channel->offer_msg.child_rel_id;
 647
 648         ret = hv_vmbus_post_message(
 649                 msg, sizeof(hv_vmbus_channel_close_channel));
 650
 651         /* Tear down the gpadl for the channel's ring buffer */
 652         if (channel->ring_buffer_gpadl_handle) {
 653                 hv_vmbus_channel_teardown_gpdal(channel,
 654                         channel->ring_buffer_gpadl_handle);
 655         }
 656
 657         /* TODO: Send a msg to release the childRelId */
 658
 659         /* cleanup the ring buffers for this channel */
 660         hv_ring_buffer_cleanup(&channel->outbound);
 661         hv_ring_buffer_cleanup(&channel->inbound);
 662
 663         contigfree(channel->ring_buffer_pages, channel->ring_buffer_size,
 664             M_DEVBUF);
 665
 666         free(info, M_DEVBUF);
 667 }
 668
 669 /**
 670  * @brief Close the specified channel
 671  */
 672 void
 673 hv_vmbus_channel_close(hv_vmbus_channel *channel)
 674 {
 675         hv_vmbus_channel*       sub_channel;
 676
 677         if (channel->primary_channel != NULL) {
 678                 /*
 679                  * We only close multi-channels when the primary is
 680                  * closed.
 681                  */
 682                 return;
 683         }
 684
 685         /*
 686          * Close all multi-channels first.
 687          */
 688         TAILQ_FOREACH(sub_channel, &channel->sc_list_anchor,
 689             sc_list_entry) {
 690                 if (sub_channel->state != HV_CHANNEL_OPENED_STATE)
 691                         continue;
 692                 hv_vmbus_channel_close_internal(sub_channel);
 693         }
 694         /*
 695          * Then close the primary channel.
 696          */
 697         hv_vmbus_channel_close_internal(channel);
 698 }
 699
 700 /**
 701  * @brief Send the specified buffer on the given channel
 702  */
 703 int
 704 hv_vmbus_channel_send_packet(
 705         hv_vmbus_channel*       channel,
 706         void*                   buffer,
 707         uint32_t                buffer_len,
 708         uint64_t                request_id,
 709         hv_vmbus_packet_type    type,
 710         uint32_t                flags)
 711 {
 712         int                     ret = 0;
 713         hv_vm_packet_descriptor desc;
 714         uint32_t                packet_len;
 715         uint64_t                aligned_data;
 716         uint32_t                packet_len_aligned;
 717         boolean_t               need_sig;
 718         hv_vmbus_sg_buffer_list buffer_list[3];
 719
 720         packet_len = sizeof(hv_vm_packet_descriptor) + buffer_len;
 721         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 722         aligned_data = 0;
 723
 724         /* Setup the descriptor */
 725         desc.type = type;   /* HV_VMBUS_PACKET_TYPE_DATA_IN_BAND;             */
 726         desc.flags = flags; /* HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED */
 727                             /* in 8-bytes granularity */
 728         desc.data_offset8 = sizeof(hv_vm_packet_descriptor) >> 3;
 729         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 730         desc.transaction_id = request_id;
 731
 732         buffer_list[0].data = &desc;
 733         buffer_list[0].length = sizeof(hv_vm_packet_descriptor);
 734
 735         buffer_list[1].data = buffer;
 736         buffer_list[1].length = buffer_len;
 737
 738         buffer_list[2].data = &aligned_data;
 739         buffer_list[2].length = packet_len_aligned - packet_len;
 740
 741         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 742             &need_sig);
 743
 744         /* TODO: We should determine if this is optional */
 745         if (ret == 0 && need_sig) {
 746                 vmbus_channel_set_event(channel);
 747         }
 748
 749         return (ret);
 750 }
 751
 752 /**
 753  * @brief Send a range of single-page buffer packets using
 754  * a GPADL Direct packet type
 755  */
 756 int
 757 hv_vmbus_channel_send_packet_pagebuffer(
 758         hv_vmbus_channel*       channel,
 759         hv_vmbus_page_buffer    page_buffers[],
 760         uint32_t                page_count,
 761         void*                   buffer,
 762         uint32_t                buffer_len,
 763         uint64_t                request_id)
 764 {
 765
 766         int                                     ret = 0;
 767         boolean_t                               need_sig;
 768         uint32_t                                packet_len;
 769         uint32_t                                page_buflen;
 770         uint32_t                                packetLen_aligned;
 771         hv_vmbus_sg_buffer_list                 buffer_list[4];
 772         hv_vmbus_channel_packet_page_buffer     desc;
 773         uint32_t                                descSize;
 774         uint64_t                                alignedData = 0;
 775
 776         if (page_count > HV_MAX_PAGE_BUFFER_COUNT)
 777                 return (EINVAL);
 778
 779         /*
 780          * Adjust the size down since hv_vmbus_channel_packet_page_buffer
 781          *  is the largest size we support
 782          */
 783         descSize = __offsetof(hv_vmbus_channel_packet_page_buffer, range);
 784         page_buflen = sizeof(hv_vmbus_page_buffer) * page_count;
 785         packet_len = descSize + page_buflen + buffer_len;
 786         packetLen_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 787
 788         /* Setup the descriptor */
 789         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 790         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 791         /* in 8-bytes granularity */
 792         desc.data_offset8 = (descSize + page_buflen) >> 3;
 793         desc.length8 = (uint16_t) (packetLen_aligned >> 3);
 794         desc.transaction_id = request_id;
 795         desc.range_count = page_count;
 796
 797         buffer_list[0].data = &desc;
 798         buffer_list[0].length = descSize;
 799
 800         buffer_list[1].data = page_buffers;
 801         buffer_list[1].length = page_buflen;
 802
 803         buffer_list[2].data = buffer;
 804         buffer_list[2].length = buffer_len;
 805
 806         buffer_list[3].data = &alignedData;
 807         buffer_list[3].length = packetLen_aligned - packet_len;
 808
 809         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 4,
 810             &need_sig);
 811
 812         /* TODO: We should determine if this is optional */
 813         if (ret == 0 && need_sig) {
 814                 vmbus_channel_set_event(channel);
 815         }
 816
 817         return (ret);
 818 }
 819
 820 /**
 821  * @brief Send a multi-page buffer packet using a GPADL Direct packet type
 822  */
 823 int
 824 hv_vmbus_channel_send_packet_multipagebuffer(
 825         hv_vmbus_channel*               channel,
 826         hv_vmbus_multipage_buffer*      multi_page_buffer,
 827         void*                           buffer,
 828         uint32_t                        buffer_len,
 829         uint64_t                        request_id)
 830 {
 831
 832         int                     ret = 0;
 833         uint32_t                desc_size;
 834         boolean_t               need_sig;
 835         uint32_t                packet_len;
 836         uint32_t                packet_len_aligned;
 837         uint32_t                pfn_count;
 838         uint64_t                aligned_data = 0;
 839         hv_vmbus_sg_buffer_list buffer_list[3];
 840         hv_vmbus_channel_packet_multipage_buffer desc;
 841
 842         pfn_count =
 843             HV_NUM_PAGES_SPANNED(
 844                     multi_page_buffer->offset,
 845                     multi_page_buffer->length);
 846
 847         if ((pfn_count == 0) || (pfn_count > HV_MAX_MULTIPAGE_BUFFER_COUNT))
 848             return (EINVAL);
 849         /*
 850          * Adjust the size down since hv_vmbus_channel_packet_multipage_buffer
 851          * is the largest size we support
 852          */
 853         desc_size =
 854             sizeof(hv_vmbus_channel_packet_multipage_buffer) -
 855                     ((HV_MAX_MULTIPAGE_BUFFER_COUNT - pfn_count) *
 856                         sizeof(uint64_t));
 857         packet_len = desc_size + buffer_len;
 858         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 859
 860         /*
 861          * Setup the descriptor
 862          */
 863         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 864         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 865         desc.data_offset8 = desc_size >> 3; /* in 8-bytes granularity */
 866         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 867         desc.transaction_id = request_id;
 868         desc.range_count = 1;
 869
 870         desc.range.length = multi_page_buffer->length;
 871         desc.range.offset = multi_page_buffer->offset;
 872
 873         memcpy(desc.range.pfn_array, multi_page_buffer->pfn_array,
 874                 pfn_count * sizeof(uint64_t));
 875
 876         buffer_list[0].data = &desc;
 877         buffer_list[0].length = desc_size;
 878
 879         buffer_list[1].data = buffer;
 880         buffer_list[1].length = buffer_len;
 881
 882         buffer_list[2].data = &aligned_data;
 883         buffer_list[2].length = packet_len_aligned - packet_len;
 884
 885         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 886             &need_sig);
 887
 888         /* TODO: We should determine if this is optional */
 889         if (ret == 0 && need_sig) {
 890             vmbus_channel_set_event(channel);
 891         }
 892
 893         return (ret);
 894 }
 895
 896 /**
 897  * @brief Retrieve the user packet on the specified channel
 898  */
 899 int
 900 hv_vmbus_channel_recv_packet(
 901         hv_vmbus_channel*       channel,
 902         void*                   Buffer,
 903         uint32_t                buffer_len,
 904         uint32_t*               buffer_actual_len,
 905         uint64_t*               request_id)
 906 {
 907         int                     ret;
 908         uint32_t                user_len;
 909         uint32_t                packet_len;
 910         hv_vm_packet_descriptor desc;
 911
 912         *buffer_actual_len = 0;
 913         *request_id = 0;
 914
 915         ret = hv_ring_buffer_peek(&channel->inbound, &desc,
 916                 sizeof(hv_vm_packet_descriptor));
 917         if (ret != 0)
 918                 return (0);
 919
 920         packet_len = desc.length8 << 3;
 921         user_len = packet_len - (desc.data_offset8 << 3);
 922
 923         *buffer_actual_len = user_len;
 924
 925         if (user_len > buffer_len)
 926                 return (EINVAL);
 927
 928         *request_id = desc.transaction_id;
 929
 930         /* Copy over the packet to the user buffer */
 931         ret = hv_ring_buffer_read(&channel->inbound, Buffer, user_len,
 932                 (desc.data_offset8 << 3));
 933
 934         return (0);
 935 }
 936
 937 /**
 938  * @brief Retrieve the raw packet on the specified channel
 939  */
 940 int
 941 hv_vmbus_channel_recv_packet_raw(
 942         hv_vmbus_channel*       channel,
 943         void*                   buffer,
 944         uint32_t                buffer_len,
 945         uint32_t*               buffer_actual_len,
 946         uint64_t*               request_id)
 947 {
 948         int             ret;
 949         uint32_t        packetLen;
 950         hv_vm_packet_descriptor desc;
 951
 952         *buffer_actual_len = 0;
 953         *request_id = 0;
 954
 955         ret = hv_ring_buffer_peek(
 956                 &channel->inbound, &desc,
 957                 sizeof(hv_vm_packet_descriptor));
 958
 959         if (ret != 0)
 960             return (0);
 961
 962         packetLen = desc.length8 << 3;
 963         *buffer_actual_len = packetLen;
 964
 965         if (packetLen > buffer_len)
 966             return (ENOBUFS);
 967
 968         *request_id = desc.transaction_id;
 969
 970         /* Copy over the entire packet to the user buffer */
 971         ret = hv_ring_buffer_read(&channel->inbound, buffer, packetLen, 0);
 972
 973         return (0);
 974 }
 975
 976
 977 /**
 978  * Process a channel event notification
 979  */
 980 static void
 981 VmbusProcessChannelEvent(void* context, int pending)
 982 {
 983         void* arg;
 984         uint32_t bytes_to_read;
 985         hv_vmbus_channel* channel = (hv_vmbus_channel*)context;
 986         boolean_t is_batched_reading;
 987
 988         /**
 989          * Find the channel based on this relid and invokes
 990          * the channel callback to process the event
 991          */
 992
 993         if (channel == NULL) {
 994                 return;
 995         }
 996         /**
 997          * To deal with the race condition where we might
 998          * receive a packet while the relevant driver is
 999          * being unloaded, dispatch the callback while
1000          * holding the channel lock. The unloading driver
1001          * will acquire the same channel lock to set the
1002          * callback to NULL. This closes the window.
1003          */
1004
1005         if (channel->on_channel_callback != NULL) {
1006                 arg = channel->channel_callback_context;
1007                 is_batched_reading = channel->batched_reading;
1008                 /*
1009                  * Optimize host to guest signaling by ensuring:
1010                  * 1. While reading the channel, we disable interrupts from
1011                  *    host.
1012                  * 2. Ensure that we process all posted messages from the host
1013                  *    before returning from this callback.
1014                  * 3. Once we return, enable signaling from the host. Once this
1015                  *    state is set we check to see if additional packets are
1016                  *    available to read. In this case we repeat the process.
1017                  */
1018                 do {
1019                         if (is_batched_reading)
1020                                 hv_ring_buffer_read_begin(&channel->inbound);
1021
1022                         channel->on_channel_callback(arg);
1023
1024                         if (is_batched_reading)
1025                                 bytes_to_read =
1026                                     hv_ring_buffer_read_end(&channel->inbound);
1027                         else
1028                                 bytes_to_read = 0;
1029                 } while (is_batched_reading && (bytes_to_read != 0));
1030         }
1031 }