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
  41 #include <machine/atomic.h>
  42 #include <machine/bus.h>
  43
  44 #include <vm/vm.h>
  45 #include <vm/vm_param.h>
  46 #include <vm/pmap.h>
  47
  48 #include <dev/hyperv/vmbus/hv_vmbus_priv.h>
  49 #include <dev/hyperv/vmbus/hyperv_var.h>
  50 #include <dev/hyperv/vmbus/vmbus_reg.h>
  51 #include <dev/hyperv/vmbus/vmbus_var.h>
  52
  53 static void     vmbus_channel_set_event(hv_vmbus_channel* channel);
  54 static void     VmbusProcessChannelEvent(void* channel, int pending);
  55 static void     vmbus_chan_update_evtflagcnt(struct vmbus_softc *,
  56                     const struct hv_vmbus_channel *);
  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         struct vmbus_softc *sc = channel->vmbus_sc;
  65         uint32_t chanid = channel->ch_id;
  66
  67         atomic_set_long(&sc->vmbus_tx_evtflags[chanid >> VMBUS_EVTFLAG_SHIFT],
  68             1UL << (chanid & VMBUS_EVTFLAG_MASK));
  69
  70         if (channel->ch_flags & VMBUS_CHAN_FLAG_HASMNF) {
  71                 hv_vmbus_monitor_page *monitor_page;
  72
  73                 monitor_page = sc->vmbus_mnf2;
  74                 synch_set_bit(channel->monitor_bit,
  75                         (uint32_t *)&monitor_page->
  76                                 trigger_group[channel->monitor_group].u.pending);
  77         } else {
  78                 hypercall_signal_event(channel->ch_sigevt_dma.hv_paddr);
  79         }
  80
  81 }
  82
  83 static int
  84 vmbus_channel_sysctl_monalloc(SYSCTL_HANDLER_ARGS)
  85 {
  86         struct hv_vmbus_channel *chan = arg1;
  87         int alloc = 0;
  88
  89         if (chan->ch_flags & VMBUS_CHAN_FLAG_HASMNF)
  90                 alloc = 1;
  91         return sysctl_handle_int(oidp, &alloc, 0, req);
  92 }
  93
  94 static void
  95 vmbus_channel_sysctl_create(hv_vmbus_channel* channel)
  96 {
  97         device_t dev;
  98         struct sysctl_oid *devch_sysctl;
  99         struct sysctl_oid *devch_id_sysctl, *devch_sub_sysctl;
 100         struct sysctl_oid *devch_id_in_sysctl, *devch_id_out_sysctl;
 101         struct sysctl_ctx_list *ctx;
 102         uint32_t ch_id;
 103         uint16_t sub_ch_id;
 104         char name[16];
 105
 106         hv_vmbus_channel* primary_ch = channel->primary_channel;
 107
 108         if (primary_ch == NULL) {
 109                 dev = channel->ch_dev;
 110                 ch_id = channel->ch_id;
 111         } else {
 112                 dev = primary_ch->ch_dev;
 113                 ch_id = primary_ch->ch_id;
 114                 sub_ch_id = channel->ch_subidx;
 115         }
 116         ctx = &channel->ch_sysctl_ctx;
 117         sysctl_ctx_init(ctx);
 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 | CTLFLAG_MPSAFE, 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 | CTLFLAG_MPSAFE, 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 | CTLFLAG_MPSAFE, 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 | CTLFLAG_MPSAFE, 0, "");
 136
 137                 SYSCTL_ADD_UINT(ctx, SYSCTL_CHILDREN(devch_id_sysctl),
 138                     OID_AUTO, "chanid", CTLFLAG_RD,
 139                     &channel->ch_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 | CTLFLAG_MPSAFE,
 145             channel, 0, 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 | CTLFLAG_MPSAFE, 0, "");
 153         devch_id_out_sysctl = SYSCTL_ADD_NODE(ctx,
 154                     SYSCTL_CHILDREN(devch_id_sysctl),
 155                     OID_AUTO,
 156                     "out",
 157                     CTLFLAG_RD | CTLFLAG_MPSAFE, 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         struct vmbus_softc *sc = new_channel->vmbus_sc;
 182         const struct vmbus_chanmsg_chopen_resp *resp;
 183         const struct vmbus_message *msg;
 184         struct vmbus_chanmsg_chopen *req;
 185         struct vmbus_msghc *mh;
 186         uint32_t status;
 187         int ret = 0;
 188         void *in, *out;
 189
 190         if (user_data_len > VMBUS_CHANMSG_CHOPEN_UDATA_SIZE) {
 191                 device_printf(sc->vmbus_dev,
 192                     "invalid udata len %u for chan%u\n",
 193                     user_data_len, new_channel->ch_id);
 194                 return EINVAL;
 195         }
 196
 197         mtx_lock(&new_channel->sc_lock);
 198         if (new_channel->state == HV_CHANNEL_OPEN_STATE) {
 199             new_channel->state = HV_CHANNEL_OPENING_STATE;
 200         } else {
 201             mtx_unlock(&new_channel->sc_lock);
 202             if(bootverbose)
 203                 printf("VMBUS: Trying to open channel <%p> which in "
 204                     "%d state.\n", new_channel, new_channel->state);
 205             return (EINVAL);
 206         }
 207         mtx_unlock(&new_channel->sc_lock);
 208
 209         new_channel->on_channel_callback = pfn_on_channel_callback;
 210         new_channel->channel_callback_context = context;
 211
 212         vmbus_chan_update_evtflagcnt(sc, new_channel);
 213
 214         new_channel->rxq = VMBUS_PCPU_GET(new_channel->vmbus_sc, event_tq,
 215             new_channel->target_cpu);
 216         TASK_INIT(&new_channel->channel_task, 0, VmbusProcessChannelEvent, new_channel);
 217
 218         /* Allocate the ring buffer */
 219         out = contigmalloc((send_ring_buffer_size + recv_ring_buffer_size),
 220             M_DEVBUF, M_ZERO, 0UL, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
 221         KASSERT(out != NULL,
 222             ("Error VMBUS: contigmalloc failed to allocate Ring Buffer!"));
 223         if (out == NULL)
 224                 return (ENOMEM);
 225
 226         in = ((uint8_t *) out + send_ring_buffer_size);
 227
 228         new_channel->ring_buffer_pages = out;
 229         new_channel->ring_buffer_page_count = (send_ring_buffer_size +
 230             recv_ring_buffer_size) >> PAGE_SHIFT;
 231         new_channel->ring_buffer_size = send_ring_buffer_size +
 232             recv_ring_buffer_size;
 233
 234         hv_vmbus_ring_buffer_init(
 235                 &new_channel->outbound,
 236                 out,
 237                 send_ring_buffer_size);
 238
 239         hv_vmbus_ring_buffer_init(
 240                 &new_channel->inbound,
 241                 in,
 242                 recv_ring_buffer_size);
 243
 244         /* Create sysctl tree for this channel */
 245         vmbus_channel_sysctl_create(new_channel);
 246
 247         /**
 248          * Establish the gpadl for the ring buffer
 249          */
 250         new_channel->ring_buffer_gpadl_handle = 0;
 251
 252         ret = hv_vmbus_channel_establish_gpadl(new_channel,
 253                 new_channel->outbound.ring_buffer,
 254                 send_ring_buffer_size + recv_ring_buffer_size,
 255                 &new_channel->ring_buffer_gpadl_handle);
 256
 257         /*
 258          * Open channel w/ the bufring GPADL on the target CPU.
 259          */
 260         mh = vmbus_msghc_get(sc, sizeof(*req));
 261         if (mh == NULL) {
 262                 device_printf(sc->vmbus_dev,
 263                     "can not get msg hypercall for chopen(chan%u)\n",
 264                     new_channel->ch_id);
 265                 return ENXIO;
 266         }
 267
 268         req = vmbus_msghc_dataptr(mh);
 269         req->chm_hdr.chm_type = VMBUS_CHANMSG_TYPE_CHOPEN;
 270         req->chm_chanid = new_channel->ch_id;
 271         req->chm_openid = new_channel->ch_id;
 272         req->chm_gpadl = new_channel->ring_buffer_gpadl_handle;
 273         req->chm_vcpuid = new_channel->target_vcpu;
 274         req->chm_rxbr_pgofs = send_ring_buffer_size >> PAGE_SHIFT;
 275         if (user_data_len)
 276                 memcpy(req->chm_udata, user_data, user_data_len);
 277
 278         ret = vmbus_msghc_exec(sc, mh);
 279         if (ret != 0) {
 280                 device_printf(sc->vmbus_dev,
 281                     "chopen(chan%u) msg hypercall exec failed: %d\n",
 282                     new_channel->ch_id, ret);
 283                 vmbus_msghc_put(sc, mh);
 284                 return ret;
 285         }
 286
 287         msg = vmbus_msghc_wait_result(sc, mh);
 288         resp = (const struct vmbus_chanmsg_chopen_resp *)msg->msg_data;
 289         status = resp->chm_status;
 290
 291         vmbus_msghc_put(sc, mh);
 292
 293         if (status == 0) {
 294                 new_channel->state = HV_CHANNEL_OPENED_STATE;
 295                 if (bootverbose) {
 296                         device_printf(sc->vmbus_dev, "chan%u opened\n",
 297                             new_channel->ch_id);
 298                 }
 299         } else {
 300                 device_printf(sc->vmbus_dev, "failed to open chan%u\n",
 301                     new_channel->ch_id);
 302                 ret = ENXIO;
 303         }
 304         return (ret);
 305 }
 306
 307 /**
 308  * @brief Establish a GPADL for the specified buffer
 309  */
 310 int
 311 hv_vmbus_channel_establish_gpadl(struct hv_vmbus_channel *channel,
 312     void *contig_buffer, uint32_t size, uint32_t *gpadl0)
 313 {
 314         struct vmbus_softc *sc = channel->vmbus_sc;
 315         struct vmbus_msghc *mh;
 316         struct vmbus_chanmsg_gpadl_conn *req;
 317         const struct vmbus_message *msg;
 318         size_t reqsz;
 319         uint32_t gpadl, status;
 320         int page_count, range_len, i, cnt, error;
 321         uint64_t page_id, paddr;
 322
 323         /*
 324          * Preliminary checks.
 325          */
 326
 327         KASSERT((size & PAGE_MASK) == 0,
 328             ("invalid GPA size %u, not multiple page size", size));
 329         page_count = size >> PAGE_SHIFT;
 330
 331         paddr = hv_get_phys_addr(contig_buffer);
 332         KASSERT((paddr & PAGE_MASK) == 0,
 333             ("GPA is not page aligned %jx", (uintmax_t)paddr));
 334         page_id = paddr >> PAGE_SHIFT;
 335
 336         range_len = __offsetof(struct vmbus_gpa_range, gpa_page[page_count]);
 337         /*
 338          * We don't support multiple GPA ranges.
 339          */
 340         if (range_len > UINT16_MAX) {
 341                 device_printf(sc->vmbus_dev, "GPA too large, %d pages\n",
 342                     page_count);
 343                 return EOPNOTSUPP;
 344         }
 345
 346         /*
 347          * Allocate GPADL id.
 348          */
 349         gpadl = vmbus_gpadl_alloc(sc);
 350         *gpadl0 = gpadl;
 351
 352         /*
 353          * Connect this GPADL to the target channel.
 354          *
 355          * NOTE:
 356          * Since each message can only hold small set of page
 357          * addresses, several messages may be required to
 358          * complete the connection.
 359          */
 360         if (page_count > VMBUS_CHANMSG_GPADL_CONN_PGMAX)
 361                 cnt = VMBUS_CHANMSG_GPADL_CONN_PGMAX;
 362         else
 363                 cnt = page_count;
 364         page_count -= cnt;
 365
 366         reqsz = __offsetof(struct vmbus_chanmsg_gpadl_conn,
 367             chm_range.gpa_page[cnt]);
 368         mh = vmbus_msghc_get(sc, reqsz);
 369         if (mh == NULL) {
 370                 device_printf(sc->vmbus_dev,
 371                     "can not get msg hypercall for gpadl->chan%u\n",
 372                     channel->ch_id);
 373                 return EIO;
 374         }
 375
 376         req = vmbus_msghc_dataptr(mh);
 377         req->chm_hdr.chm_type = VMBUS_CHANMSG_TYPE_GPADL_CONN;
 378         req->chm_chanid = channel->ch_id;
 379         req->chm_gpadl = gpadl;
 380         req->chm_range_len = range_len;
 381         req->chm_range_cnt = 1;
 382         req->chm_range.gpa_len = size;
 383         req->chm_range.gpa_ofs = 0;
 384         for (i = 0; i < cnt; ++i)
 385                 req->chm_range.gpa_page[i] = page_id++;
 386
 387         error = vmbus_msghc_exec(sc, mh);
 388         if (error) {
 389                 device_printf(sc->vmbus_dev,
 390                     "gpadl->chan%u msg hypercall exec failed: %d\n",
 391                     channel->ch_id, error);
 392                 vmbus_msghc_put(sc, mh);
 393                 return error;
 394         }
 395
 396         while (page_count > 0) {
 397                 struct vmbus_chanmsg_gpadl_subconn *subreq;
 398
 399                 if (page_count > VMBUS_CHANMSG_GPADL_SUBCONN_PGMAX)
 400                         cnt = VMBUS_CHANMSG_GPADL_SUBCONN_PGMAX;
 401                 else
 402                         cnt = page_count;
 403                 page_count -= cnt;
 404
 405                 reqsz = __offsetof(struct vmbus_chanmsg_gpadl_subconn,
 406                     chm_gpa_page[cnt]);
 407                 vmbus_msghc_reset(mh, reqsz);
 408
 409                 subreq = vmbus_msghc_dataptr(mh);
 410                 subreq->chm_hdr.chm_type = VMBUS_CHANMSG_TYPE_GPADL_SUBCONN;
 411                 subreq->chm_gpadl = gpadl;
 412                 for (i = 0; i < cnt; ++i)
 413                         subreq->chm_gpa_page[i] = page_id++;
 414
 415                 vmbus_msghc_exec_noresult(mh);
 416         }
 417         KASSERT(page_count == 0, ("invalid page count %d", page_count));
 418
 419         msg = vmbus_msghc_wait_result(sc, mh);
 420         status = ((const struct vmbus_chanmsg_gpadl_connresp *)
 421             msg->msg_data)->chm_status;
 422
 423         vmbus_msghc_put(sc, mh);
 424
 425         if (status != 0) {
 426                 device_printf(sc->vmbus_dev, "gpadl->chan%u failed: "
 427                     "status %u\n", channel->ch_id, status);
 428                 return EIO;
 429         } else {
 430                 if (bootverbose) {
 431                         device_printf(sc->vmbus_dev, "gpadl->chan%u "
 432                             "succeeded\n", channel->ch_id);
 433                 }
 434         }
 435         return 0;
 436 }
 437
 438 /*
 439  * Disconnect the GPA from the target channel
 440  */
 441 int
 442 hv_vmbus_channel_teardown_gpdal(struct hv_vmbus_channel *chan, uint32_t gpadl)
 443 {
 444         struct vmbus_softc *sc = chan->vmbus_sc;
 445         struct vmbus_msghc *mh;
 446         struct vmbus_chanmsg_gpadl_disconn *req;
 447         int error;
 448
 449         mh = vmbus_msghc_get(sc, sizeof(*req));
 450         if (mh == NULL) {
 451                 device_printf(sc->vmbus_dev,
 452                     "can not get msg hypercall for gpa x->chan%u\n",
 453                     chan->ch_id);
 454                 return EBUSY;
 455         }
 456
 457         req = vmbus_msghc_dataptr(mh);
 458         req->chm_hdr.chm_type = VMBUS_CHANMSG_TYPE_GPADL_DISCONN;
 459         req->chm_chanid = chan->ch_id;
 460         req->chm_gpadl = gpadl;
 461
 462         error = vmbus_msghc_exec(sc, mh);
 463         if (error) {
 464                 device_printf(sc->vmbus_dev,
 465                     "gpa x->chan%u msg hypercall exec failed: %d\n",
 466                     chan->ch_id, error);
 467                 vmbus_msghc_put(sc, mh);
 468                 return error;
 469         }
 470
 471         vmbus_msghc_wait_result(sc, mh);
 472         /* Discard result; no useful information */
 473         vmbus_msghc_put(sc, mh);
 474
 475         return 0;
 476 }
 477
 478 static void
 479 hv_vmbus_channel_close_internal(hv_vmbus_channel *channel)
 480 {
 481         struct vmbus_softc *sc = channel->vmbus_sc;
 482         struct vmbus_msghc *mh;
 483         struct vmbus_chanmsg_chclose *req;
 484         struct taskqueue *rxq = channel->rxq;
 485         int error;
 486
 487         channel->state = HV_CHANNEL_OPEN_STATE;
 488         sysctl_ctx_free(&channel->ch_sysctl_ctx);
 489
 490         /*
 491          * set rxq to NULL to avoid more requests be scheduled
 492          */
 493         channel->rxq = NULL;
 494         taskqueue_drain(rxq, &channel->channel_task);
 495         channel->on_channel_callback = NULL;
 496
 497         /**
 498          * Send a closing message
 499          */
 500
 501         mh = vmbus_msghc_get(sc, sizeof(*req));
 502         if (mh == NULL) {
 503                 device_printf(sc->vmbus_dev,
 504                     "can not get msg hypercall for chclose(chan%u)\n",
 505                     channel->ch_id);
 506                 return;
 507         }
 508
 509         req = vmbus_msghc_dataptr(mh);
 510         req->chm_hdr.chm_type = VMBUS_CHANMSG_TYPE_CHCLOSE;
 511         req->chm_chanid = channel->ch_id;
 512
 513         error = vmbus_msghc_exec_noresult(mh);
 514         vmbus_msghc_put(sc, mh);
 515
 516         if (error) {
 517                 device_printf(sc->vmbus_dev,
 518                     "chclose(chan%u) msg hypercall exec failed: %d\n",
 519                     channel->ch_id, error);
 520                 return;
 521         } else if (bootverbose) {
 522                 device_printf(sc->vmbus_dev, "close chan%u\n",
 523                     channel->ch_id);
 524         }
 525
 526         /* Tear down the gpadl for the channel's ring buffer */
 527         if (channel->ring_buffer_gpadl_handle) {
 528                 hv_vmbus_channel_teardown_gpdal(channel,
 529                         channel->ring_buffer_gpadl_handle);
 530         }
 531
 532         /* TODO: Send a msg to release the childRelId */
 533
 534         /* cleanup the ring buffers for this channel */
 535         hv_ring_buffer_cleanup(&channel->outbound);
 536         hv_ring_buffer_cleanup(&channel->inbound);
 537
 538         contigfree(channel->ring_buffer_pages, channel->ring_buffer_size,
 539             M_DEVBUF);
 540 }
 541
 542 /**
 543  * @brief Close the specified channel
 544  */
 545 void
 546 hv_vmbus_channel_close(hv_vmbus_channel *channel)
 547 {
 548         hv_vmbus_channel*       sub_channel;
 549
 550         if (channel->primary_channel != NULL) {
 551                 /*
 552                  * We only close multi-channels when the primary is
 553                  * closed.
 554                  */
 555                 return;
 556         }
 557
 558         /*
 559          * Close all multi-channels first.
 560          */
 561         TAILQ_FOREACH(sub_channel, &channel->sc_list_anchor,
 562             sc_list_entry) {
 563                 if (sub_channel->state != HV_CHANNEL_OPENED_STATE)
 564                         continue;
 565                 hv_vmbus_channel_close_internal(sub_channel);
 566         }
 567         /*
 568          * Then close the primary channel.
 569          */
 570         hv_vmbus_channel_close_internal(channel);
 571 }
 572
 573 /**
 574  * @brief Send the specified buffer on the given channel
 575  */
 576 int
 577 hv_vmbus_channel_send_packet(
 578         hv_vmbus_channel*       channel,
 579         void*                   buffer,
 580         uint32_t                buffer_len,
 581         uint64_t                request_id,
 582         hv_vmbus_packet_type    type,
 583         uint32_t                flags)
 584 {
 585         int                     ret = 0;
 586         hv_vm_packet_descriptor desc;
 587         uint32_t                packet_len;
 588         uint64_t                aligned_data;
 589         uint32_t                packet_len_aligned;
 590         boolean_t               need_sig;
 591         hv_vmbus_sg_buffer_list buffer_list[3];
 592
 593         packet_len = sizeof(hv_vm_packet_descriptor) + buffer_len;
 594         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 595         aligned_data = 0;
 596
 597         /* Setup the descriptor */
 598         desc.type = type;   /* HV_VMBUS_PACKET_TYPE_DATA_IN_BAND;             */
 599         desc.flags = flags; /* HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED */
 600                             /* in 8-bytes granularity */
 601         desc.data_offset8 = sizeof(hv_vm_packet_descriptor) >> 3;
 602         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 603         desc.transaction_id = request_id;
 604
 605         buffer_list[0].data = &desc;
 606         buffer_list[0].length = sizeof(hv_vm_packet_descriptor);
 607
 608         buffer_list[1].data = buffer;
 609         buffer_list[1].length = buffer_len;
 610
 611         buffer_list[2].data = &aligned_data;
 612         buffer_list[2].length = packet_len_aligned - packet_len;
 613
 614         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 615             &need_sig);
 616
 617         /* TODO: We should determine if this is optional */
 618         if (ret == 0 && need_sig) {
 619                 vmbus_channel_set_event(channel);
 620         }
 621
 622         return (ret);
 623 }
 624
 625 /**
 626  * @brief Send a range of single-page buffer packets using
 627  * a GPADL Direct packet type
 628  */
 629 int
 630 hv_vmbus_channel_send_packet_pagebuffer(
 631         hv_vmbus_channel*       channel,
 632         hv_vmbus_page_buffer    page_buffers[],
 633         uint32_t                page_count,
 634         void*                   buffer,
 635         uint32_t                buffer_len,
 636         uint64_t                request_id)
 637 {
 638
 639         int                                     ret = 0;
 640         boolean_t                               need_sig;
 641         uint32_t                                packet_len;
 642         uint32_t                                page_buflen;
 643         uint32_t                                packetLen_aligned;
 644         hv_vmbus_sg_buffer_list                 buffer_list[4];
 645         hv_vmbus_channel_packet_page_buffer     desc;
 646         uint32_t                                descSize;
 647         uint64_t                                alignedData = 0;
 648
 649         if (page_count > HV_MAX_PAGE_BUFFER_COUNT)
 650                 return (EINVAL);
 651
 652         /*
 653          * Adjust the size down since hv_vmbus_channel_packet_page_buffer
 654          *  is the largest size we support
 655          */
 656         descSize = __offsetof(hv_vmbus_channel_packet_page_buffer, range);
 657         page_buflen = sizeof(hv_vmbus_page_buffer) * page_count;
 658         packet_len = descSize + page_buflen + buffer_len;
 659         packetLen_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 660
 661         /* Setup the descriptor */
 662         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 663         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 664         /* in 8-bytes granularity */
 665         desc.data_offset8 = (descSize + page_buflen) >> 3;
 666         desc.length8 = (uint16_t) (packetLen_aligned >> 3);
 667         desc.transaction_id = request_id;
 668         desc.range_count = page_count;
 669
 670         buffer_list[0].data = &desc;
 671         buffer_list[0].length = descSize;
 672
 673         buffer_list[1].data = page_buffers;
 674         buffer_list[1].length = page_buflen;
 675
 676         buffer_list[2].data = buffer;
 677         buffer_list[2].length = buffer_len;
 678
 679         buffer_list[3].data = &alignedData;
 680         buffer_list[3].length = packetLen_aligned - packet_len;
 681
 682         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 4,
 683             &need_sig);
 684
 685         /* TODO: We should determine if this is optional */
 686         if (ret == 0 && need_sig) {
 687                 vmbus_channel_set_event(channel);
 688         }
 689
 690         return (ret);
 691 }
 692
 693 /**
 694  * @brief Send a multi-page buffer packet using a GPADL Direct packet type
 695  */
 696 int
 697 hv_vmbus_channel_send_packet_multipagebuffer(
 698         hv_vmbus_channel*               channel,
 699         hv_vmbus_multipage_buffer*      multi_page_buffer,
 700         void*                           buffer,
 701         uint32_t                        buffer_len,
 702         uint64_t                        request_id)
 703 {
 704
 705         int                     ret = 0;
 706         uint32_t                desc_size;
 707         boolean_t               need_sig;
 708         uint32_t                packet_len;
 709         uint32_t                packet_len_aligned;
 710         uint32_t                pfn_count;
 711         uint64_t                aligned_data = 0;
 712         hv_vmbus_sg_buffer_list buffer_list[3];
 713         hv_vmbus_channel_packet_multipage_buffer desc;
 714
 715         pfn_count =
 716             HV_NUM_PAGES_SPANNED(
 717                     multi_page_buffer->offset,
 718                     multi_page_buffer->length);
 719
 720         if ((pfn_count == 0) || (pfn_count > HV_MAX_MULTIPAGE_BUFFER_COUNT))
 721             return (EINVAL);
 722         /*
 723          * Adjust the size down since hv_vmbus_channel_packet_multipage_buffer
 724          * is the largest size we support
 725          */
 726         desc_size =
 727             sizeof(hv_vmbus_channel_packet_multipage_buffer) -
 728                     ((HV_MAX_MULTIPAGE_BUFFER_COUNT - pfn_count) *
 729                         sizeof(uint64_t));
 730         packet_len = desc_size + buffer_len;
 731         packet_len_aligned = HV_ALIGN_UP(packet_len, sizeof(uint64_t));
 732
 733         /*
 734          * Setup the descriptor
 735          */
 736         desc.type = HV_VMBUS_PACKET_TYPE_DATA_USING_GPA_DIRECT;
 737         desc.flags = HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED;
 738         desc.data_offset8 = desc_size >> 3; /* in 8-bytes granularity */
 739         desc.length8 = (uint16_t) (packet_len_aligned >> 3);
 740         desc.transaction_id = request_id;
 741         desc.range_count = 1;
 742
 743         desc.range.length = multi_page_buffer->length;
 744         desc.range.offset = multi_page_buffer->offset;
 745
 746         memcpy(desc.range.pfn_array, multi_page_buffer->pfn_array,
 747                 pfn_count * sizeof(uint64_t));
 748
 749         buffer_list[0].data = &desc;
 750         buffer_list[0].length = desc_size;
 751
 752         buffer_list[1].data = buffer;
 753         buffer_list[1].length = buffer_len;
 754
 755         buffer_list[2].data = &aligned_data;
 756         buffer_list[2].length = packet_len_aligned - packet_len;
 757
 758         ret = hv_ring_buffer_write(&channel->outbound, buffer_list, 3,
 759             &need_sig);
 760
 761         /* TODO: We should determine if this is optional */
 762         if (ret == 0 && need_sig) {
 763             vmbus_channel_set_event(channel);
 764         }
 765
 766         return (ret);
 767 }
 768
 769 /**
 770  * @brief Retrieve the user packet on the specified channel
 771  */
 772 int
 773 hv_vmbus_channel_recv_packet(
 774         hv_vmbus_channel*       channel,
 775         void*                   Buffer,
 776         uint32_t                buffer_len,
 777         uint32_t*               buffer_actual_len,
 778         uint64_t*               request_id)
 779 {
 780         int                     ret;
 781         uint32_t                user_len;
 782         uint32_t                packet_len;
 783         hv_vm_packet_descriptor desc;
 784
 785         *buffer_actual_len = 0;
 786         *request_id = 0;
 787
 788         ret = hv_ring_buffer_peek(&channel->inbound, &desc,
 789                 sizeof(hv_vm_packet_descriptor));
 790         if (ret != 0)
 791                 return (0);
 792
 793         packet_len = desc.length8 << 3;
 794         user_len = packet_len - (desc.data_offset8 << 3);
 795
 796         *buffer_actual_len = user_len;
 797
 798         if (user_len > buffer_len)
 799                 return (EINVAL);
 800
 801         *request_id = desc.transaction_id;
 802
 803         /* Copy over the packet to the user buffer */
 804         ret = hv_ring_buffer_read(&channel->inbound, Buffer, user_len,
 805                 (desc.data_offset8 << 3));
 806
 807         return (0);
 808 }
 809
 810 /**
 811  * @brief Retrieve the raw packet on the specified channel
 812  */
 813 int
 814 hv_vmbus_channel_recv_packet_raw(
 815         hv_vmbus_channel*       channel,
 816         void*                   buffer,
 817         uint32_t                buffer_len,
 818         uint32_t*               buffer_actual_len,
 819         uint64_t*               request_id)
 820 {
 821         int             ret;
 822         uint32_t        packetLen;
 823         hv_vm_packet_descriptor desc;
 824
 825         *buffer_actual_len = 0;
 826         *request_id = 0;
 827
 828         ret = hv_ring_buffer_peek(
 829                 &channel->inbound, &desc,
 830                 sizeof(hv_vm_packet_descriptor));
 831
 832         if (ret != 0)
 833             return (0);
 834
 835         packetLen = desc.length8 << 3;
 836         *buffer_actual_len = packetLen;
 837
 838         if (packetLen > buffer_len)
 839             return (ENOBUFS);
 840
 841         *request_id = desc.transaction_id;
 842
 843         /* Copy over the entire packet to the user buffer */
 844         ret = hv_ring_buffer_read(&channel->inbound, buffer, packetLen, 0);
 845
 846         return (0);
 847 }
 848
 849
 850 /**
 851  * Process a channel event notification
 852  */
 853 static void
 854 VmbusProcessChannelEvent(void* context, int pending)
 855 {
 856         void* arg;
 857         uint32_t bytes_to_read;
 858         hv_vmbus_channel* channel = (hv_vmbus_channel*)context;
 859         boolean_t is_batched_reading;
 860
 861         if (channel->on_channel_callback != NULL) {
 862                 arg = channel->channel_callback_context;
 863                 is_batched_reading = channel->batched_reading;
 864                 /*
 865                  * Optimize host to guest signaling by ensuring:
 866                  * 1. While reading the channel, we disable interrupts from
 867                  *    host.
 868                  * 2. Ensure that we process all posted messages from the host
 869                  *    before returning from this callback.
 870                  * 3. Once we return, enable signaling from the host. Once this
 871                  *    state is set we check to see if additional packets are
 872                  *    available to read. In this case we repeat the process.
 873                  */
 874                 do {
 875                         if (is_batched_reading)
 876                                 hv_ring_buffer_read_begin(&channel->inbound);
 877
 878                         channel->on_channel_callback(arg);
 879
 880                         if (is_batched_reading)
 881                                 bytes_to_read =
 882                                     hv_ring_buffer_read_end(&channel->inbound);
 883                         else
 884                                 bytes_to_read = 0;
 885                 } while (is_batched_reading && (bytes_to_read != 0));
 886         }
 887 }
 888
 889 static __inline void
 890 vmbus_event_flags_proc(struct vmbus_softc *sc, volatile u_long *event_flags,
 891     int flag_cnt)
 892 {
 893         int f;
 894
 895         for (f = 0; f < flag_cnt; ++f) {
 896                 uint32_t rel_id_base;
 897                 u_long flags;
 898                 int bit;
 899
 900                 if (event_flags[f] == 0)
 901                         continue;
 902
 903                 flags = atomic_swap_long(&event_flags[f], 0);
 904                 rel_id_base = f << VMBUS_EVTFLAG_SHIFT;
 905
 906                 while ((bit = ffsl(flags)) != 0) {
 907                         struct hv_vmbus_channel *channel;
 908                         uint32_t rel_id;
 909
 910                         --bit;  /* NOTE: ffsl is 1-based */
 911                         flags &= ~(1UL << bit);
 912
 913                         rel_id = rel_id_base + bit;
 914                         channel = sc->vmbus_chmap[rel_id];
 915
 916                         /* if channel is closed or closing */
 917                         if (channel == NULL || channel->rxq == NULL)
 918                                 continue;
 919
 920                         if (channel->batched_reading)
 921                                 hv_ring_buffer_read_begin(&channel->inbound);
 922                         taskqueue_enqueue(channel->rxq, &channel->channel_task);
 923                 }
 924         }
 925 }
 926
 927 void
 928 vmbus_event_proc(struct vmbus_softc *sc, int cpu)
 929 {
 930         struct vmbus_evtflags *eventf;
 931
 932         /*
 933          * On Host with Win8 or above, the event page can be checked directly
 934          * to get the id of the channel that has the pending interrupt.
 935          */
 936         eventf = VMBUS_PCPU_GET(sc, event_flags, cpu) + VMBUS_SINT_MESSAGE;
 937         vmbus_event_flags_proc(sc, eventf->evt_flags,
 938             VMBUS_PCPU_GET(sc, event_flags_cnt, cpu));
 939 }
 940
 941 void
 942 vmbus_event_proc_compat(struct vmbus_softc *sc, int cpu)
 943 {
 944         struct vmbus_evtflags *eventf;
 945
 946         eventf = VMBUS_PCPU_GET(sc, event_flags, cpu) + VMBUS_SINT_MESSAGE;
 947         if (atomic_testandclear_long(&eventf->evt_flags[0], 0)) {
 948                 vmbus_event_flags_proc(sc, sc->vmbus_rx_evtflags,
 949                     VMBUS_CHAN_MAX_COMPAT >> VMBUS_EVTFLAG_SHIFT);
 950         }
 951 }
 952
 953 static void
 954 vmbus_chan_update_evtflagcnt(struct vmbus_softc *sc,
 955     const struct hv_vmbus_channel *chan)
 956 {
 957         volatile int *flag_cnt_ptr;
 958         int flag_cnt;
 959
 960         flag_cnt = (chan->ch_id / VMBUS_EVTFLAG_LEN) + 1;
 961         flag_cnt_ptr = VMBUS_PCPU_PTR(sc, event_flags_cnt, chan->target_cpu);
 962
 963         for (;;) {
 964                 int old_flag_cnt;
 965
 966                 old_flag_cnt = *flag_cnt_ptr;
 967                 if (old_flag_cnt >= flag_cnt)
 968                         break;
 969                 if (atomic_cmpset_int(flag_cnt_ptr, old_flag_cnt, flag_cnt)) {
 970                         if (bootverbose) {
 971                                 device_printf(sc->vmbus_dev,
 972                                     "channel%u update cpu%d flag_cnt to %d\n",
 973                                     chan->ch_id,
 974                                     chan->target_cpu, flag_cnt);
 975                         }
 976                         break;
 977                 }
 978         }
 979 }