drivers/hv/connection.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright (c) 2009, Microsoft Corporation.
   5  *
   6  * Authors:
   7  *   Haiyang Zhang <haiyangz@microsoft.com>
   8  *   Hank Janssen  <hjanssen@microsoft.com>
   9  */
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/kernel.h>
  13 #include <linux/sched.h>
  14 #include <linux/wait.h>
  15 #include <linux/delay.h>
  16 #include <linux/mm.h>
  17 #include <linux/slab.h>
  18 #include <linux/vmalloc.h>
  19 #include <linux/hyperv.h>
  20 #include <linux/export.h>
  21 #include <asm/mshyperv.h>
  22
  23 #include "hyperv_vmbus.h"
  24
  25
  26 struct vmbus_connection vmbus_connection = {
  27         .conn_state             = DISCONNECTED,
  28         .next_gpadl_handle      = ATOMIC_INIT(0xE1E10),
  29
  30         .ready_for_suspend_event= COMPLETION_INITIALIZER(
  31                                   vmbus_connection.ready_for_suspend_event),
  32         .ready_for_resume_event = COMPLETION_INITIALIZER(
  33                                   vmbus_connection.ready_for_resume_event),
  34 };
  35 EXPORT_SYMBOL_GPL(vmbus_connection);
  36
  37 /*
  38  * Negotiated protocol version with the host.
  39  */
  40 __u32 vmbus_proto_version;
  41 EXPORT_SYMBOL_GPL(vmbus_proto_version);
  42
  43 static __u32 vmbus_get_next_version(__u32 current_version)
  44 {
  45         switch (current_version) {
  46         case (VERSION_WIN7):
  47                 return VERSION_WS2008;
  48
  49         case (VERSION_WIN8):
  50                 return VERSION_WIN7;
  51
  52         case (VERSION_WIN8_1):
  53                 return VERSION_WIN8;
  54
  55         case (VERSION_WIN10):
  56                 return VERSION_WIN8_1;
  57
  58         case (VERSION_WIN10_V5):
  59                 return VERSION_WIN10;
  60
  61         case (VERSION_WS2008):
  62         default:
  63                 return VERSION_INVAL;
  64         }
  65 }
  66
  67 int vmbus_negotiate_version(struct vmbus_channel_msginfo *msginfo, u32 version)
  68 {
  69         int ret = 0;
  70         struct vmbus_channel_initiate_contact *msg;
  71         unsigned long flags;
  72
  73         init_completion(&msginfo->waitevent);
  74
  75         msg = (struct vmbus_channel_initiate_contact *)msginfo->msg;
  76
  77         memset(msg, 0, sizeof(*msg));
  78         msg->header.msgtype = CHANNELMSG_INITIATE_CONTACT;
  79         msg->vmbus_version_requested = version;
  80
  81         /*
  82          * VMBus protocol 5.0 (VERSION_WIN10_V5) requires that we must use
  83          * VMBUS_MESSAGE_CONNECTION_ID_4 for the Initiate Contact Message,
  84          * and for subsequent messages, we must use the Message Connection ID
  85          * field in the host-returned Version Response Message. And, with
  86          * VERSION_WIN10_V5, we don't use msg->interrupt_page, but we tell
  87          * the host explicitly that we still use VMBUS_MESSAGE_SINT(2) for
  88          * compatibility.
  89          *
  90          * On old hosts, we should always use VMBUS_MESSAGE_CONNECTION_ID (1).
  91          */
  92         if (version >= VERSION_WIN10_V5) {
  93                 msg->msg_sint = VMBUS_MESSAGE_SINT;
  94                 vmbus_connection.msg_conn_id = VMBUS_MESSAGE_CONNECTION_ID_4;
  95         } else {
  96                 msg->interrupt_page = virt_to_phys(vmbus_connection.int_page);
  97                 vmbus_connection.msg_conn_id = VMBUS_MESSAGE_CONNECTION_ID;
  98         }
  99
 100         msg->monitor_page1 = virt_to_phys(vmbus_connection.monitor_pages[0]);
 101         msg->monitor_page2 = virt_to_phys(vmbus_connection.monitor_pages[1]);
 102         msg->target_vcpu = hv_cpu_number_to_vp_number(VMBUS_CONNECT_CPU);
 103
 104         /*
 105          * Add to list before we send the request since we may
 106          * receive the response before returning from this routine
 107          */
 108         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 109         list_add_tail(&msginfo->msglistentry,
 110                       &vmbus_connection.chn_msg_list);
 111
 112         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 113
 114         ret = vmbus_post_msg(msg,
 115                              sizeof(struct vmbus_channel_initiate_contact),
 116                              true);
 117
 118         trace_vmbus_negotiate_version(msg, ret);
 119
 120         if (ret != 0) {
 121                 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 122                 list_del(&msginfo->msglistentry);
 123                 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock,
 124                                         flags);
 125                 return ret;
 126         }
 127
 128         /* Wait for the connection response */
 129         wait_for_completion(&msginfo->waitevent);
 130
 131         spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
 132         list_del(&msginfo->msglistentry);
 133         spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
 134
 135         /* Check if successful */
 136         if (msginfo->response.version_response.version_supported) {
 137                 vmbus_connection.conn_state = CONNECTED;
 138
 139                 if (version >= VERSION_WIN10_V5)
 140                         vmbus_connection.msg_conn_id =
 141                                 msginfo->response.version_response.msg_conn_id;
 142         } else {
 143                 return -ECONNREFUSED;
 144         }
 145
 146         return ret;
 147 }
 148
 149 /*
 150  * vmbus_connect - Sends a connect request on the partition service connection
 151  */
 152 int vmbus_connect(void)
 153 {
 154         int ret = 0;
 155         struct vmbus_channel_msginfo *msginfo = NULL;
 156         __u32 version;
 157
 158         /* Initialize the vmbus connection */
 159         vmbus_connection.conn_state = CONNECTING;
 160         vmbus_connection.work_queue = create_workqueue("hv_vmbus_con");
 161         if (!vmbus_connection.work_queue) {
 162                 ret = -ENOMEM;
 163                 goto cleanup;
 164         }
 165
 166         vmbus_connection.handle_primary_chan_wq =
 167                 create_workqueue("hv_pri_chan");
 168         if (!vmbus_connection.handle_primary_chan_wq) {
 169                 ret = -ENOMEM;
 170                 goto cleanup;
 171         }
 172
 173         vmbus_connection.handle_sub_chan_wq =
 174                 create_workqueue("hv_sub_chan");
 175         if (!vmbus_connection.handle_sub_chan_wq) {
 176                 ret = -ENOMEM;
 177                 goto cleanup;
 178         }
 179
 180         INIT_LIST_HEAD(&vmbus_connection.chn_msg_list);
 181         spin_lock_init(&vmbus_connection.channelmsg_lock);
 182
 183         INIT_LIST_HEAD(&vmbus_connection.chn_list);
 184         mutex_init(&vmbus_connection.channel_mutex);
 185
 186         /*
 187          * Setup the vmbus event connection for channel interrupt
 188          * abstraction stuff
 189          */
 190         vmbus_connection.int_page =
 191         (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO, 0);
 192         if (vmbus_connection.int_page == NULL) {
 193                 ret = -ENOMEM;
 194                 goto cleanup;
 195         }
 196
 197         vmbus_connection.recv_int_page = vmbus_connection.int_page;
 198         vmbus_connection.send_int_page =
 199                 (void *)((unsigned long)vmbus_connection.int_page +
 200                         (PAGE_SIZE >> 1));
 201
 202         /*
 203          * Setup the monitor notification facility. The 1st page for
 204          * parent->child and the 2nd page for child->parent
 205          */
 206         vmbus_connection.monitor_pages[0] = (void *)__get_free_pages((GFP_KERNEL|__GFP_ZERO), 0);
 207         vmbus_connection.monitor_pages[1] = (void *)__get_free_pages((GFP_KERNEL|__GFP_ZERO), 0);
 208         if ((vmbus_connection.monitor_pages[0] == NULL) ||
 209             (vmbus_connection.monitor_pages[1] == NULL)) {
 210                 ret = -ENOMEM;
 211                 goto cleanup;
 212         }
 213
 214         msginfo = kzalloc(sizeof(*msginfo) +
 215                           sizeof(struct vmbus_channel_initiate_contact),
 216                           GFP_KERNEL);
 217         if (msginfo == NULL) {
 218                 ret = -ENOMEM;
 219                 goto cleanup;
 220         }
 221
 222         /*
 223          * Negotiate a compatible VMBUS version number with the
 224          * host. We start with the highest number we can support
 225          * and work our way down until we negotiate a compatible
 226          * version.
 227          */
 228
 229         version = VERSION_CURRENT;
 230
 231         do {
 232                 ret = vmbus_negotiate_version(msginfo, version);
 233                 if (ret == -ETIMEDOUT)
 234                         goto cleanup;
 235
 236                 if (vmbus_connection.conn_state == CONNECTED)
 237                         break;
 238
 239                 version = vmbus_get_next_version(version);
 240         } while (version != VERSION_INVAL);
 241
 242         if (version == VERSION_INVAL)
 243                 goto cleanup;
 244
 245         vmbus_proto_version = version;
 246         pr_info("Vmbus version:%d.%d\n",
 247                 version >> 16, version & 0xFFFF);
 248
 249         kfree(msginfo);
 250         return 0;
 251
 252 cleanup:
 253         pr_err("Unable to connect to host\n");
 254
 255         vmbus_connection.conn_state = DISCONNECTED;
 256         vmbus_disconnect();
 257
 258         kfree(msginfo);
 259
 260         return ret;
 261 }
 262
 263 void vmbus_disconnect(void)
 264 {
 265         /*
 266          * First send the unload request to the host.
 267          */
 268         vmbus_initiate_unload(false);
 269
 270         if (vmbus_connection.handle_sub_chan_wq)
 271                 destroy_workqueue(vmbus_connection.handle_sub_chan_wq);
 272
 273         if (vmbus_connection.handle_primary_chan_wq)
 274                 destroy_workqueue(vmbus_connection.handle_primary_chan_wq);
 275
 276         if (vmbus_connection.work_queue)
 277                 destroy_workqueue(vmbus_connection.work_queue);
 278
 279         if (vmbus_connection.int_page) {
 280                 free_pages((unsigned long)vmbus_connection.int_page, 0);
 281                 vmbus_connection.int_page = NULL;
 282         }
 283
 284         free_pages((unsigned long)vmbus_connection.monitor_pages[0], 0);
 285         free_pages((unsigned long)vmbus_connection.monitor_pages[1], 0);
 286         vmbus_connection.monitor_pages[0] = NULL;
 287         vmbus_connection.monitor_pages[1] = NULL;
 288 }
 289
 290 /*
 291  * relid2channel - Get the channel object given its
 292  * child relative id (ie channel id)
 293  */
 294 struct vmbus_channel *relid2channel(u32 relid)
 295 {
 296         struct vmbus_channel *channel;
 297         struct vmbus_channel *found_channel  = NULL;
 298         struct list_head *cur, *tmp;
 299         struct vmbus_channel *cur_sc;
 300
 301         BUG_ON(!mutex_is_locked(&vmbus_connection.channel_mutex));
 302
 303         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 304                 if (channel->offermsg.child_relid == relid) {
 305                         found_channel = channel;
 306                         break;
 307                 } else if (!list_empty(&channel->sc_list)) {
 308                         /*
 309                          * Deal with sub-channels.
 310                          */
 311                         list_for_each_safe(cur, tmp, &channel->sc_list) {
 312                                 cur_sc = list_entry(cur, struct vmbus_channel,
 313                                                         sc_list);
 314                                 if (cur_sc->offermsg.child_relid == relid) {
 315                                         found_channel = cur_sc;
 316                                         break;
 317                                 }
 318                         }
 319                 }
 320         }
 321
 322         return found_channel;
 323 }
 324
 325 /*
 326  * vmbus_on_event - Process a channel event notification
 327  *
 328  * For batched channels (default) optimize host to guest signaling
 329  * by ensuring:
 330  * 1. While reading the channel, we disable interrupts from host.
 331  * 2. Ensure that we process all posted messages from the host
 332  *    before returning from this callback.
 333  * 3. Once we return, enable signaling from the host. Once this
 334  *    state is set we check to see if additional packets are
 335  *    available to read. In this case we repeat the process.
 336  *    If this tasklet has been running for a long time
 337  *    then reschedule ourselves.
 338  */
 339 void vmbus_on_event(unsigned long data)
 340 {
 341         struct vmbus_channel *channel = (void *) data;
 342         unsigned long time_limit = jiffies + 2;
 343
 344         trace_vmbus_on_event(channel);
 345
 346         do {
 347                 void (*callback_fn)(void *);
 348
 349                 /* A channel once created is persistent even when
 350                  * there is no driver handling the device. An
 351                  * unloading driver sets the onchannel_callback to NULL.
 352                  */
 353                 callback_fn = READ_ONCE(channel->onchannel_callback);
 354                 if (unlikely(callback_fn == NULL))
 355                         return;
 356
 357                 (*callback_fn)(channel->channel_callback_context);
 358
 359                 if (channel->callback_mode != HV_CALL_BATCHED)
 360                         return;
 361
 362                 if (likely(hv_end_read(&channel->inbound) == 0))
 363                         return;
 364
 365                 hv_begin_read(&channel->inbound);
 366         } while (likely(time_before(jiffies, time_limit)));
 367
 368         /* The time limit (2 jiffies) has been reached */
 369         tasklet_schedule(&channel->callback_event);
 370 }
 371
 372 /*
 373  * vmbus_post_msg - Send a msg on the vmbus's message connection
 374  */
 375 int vmbus_post_msg(void *buffer, size_t buflen, bool can_sleep)
 376 {
 377         struct vmbus_channel_message_header *hdr;
 378         union hv_connection_id conn_id;
 379         int ret = 0;
 380         int retries = 0;
 381         u32 usec = 1;
 382
 383         conn_id.asu32 = 0;
 384         conn_id.u.id = vmbus_connection.msg_conn_id;
 385
 386         /*
 387          * hv_post_message() can have transient failures because of
 388          * insufficient resources. Retry the operation a couple of
 389          * times before giving up.
 390          */
 391         while (retries < 100) {
 392                 ret = hv_post_message(conn_id, 1, buffer, buflen);
 393
 394                 switch (ret) {
 395                 case HV_STATUS_INVALID_CONNECTION_ID:
 396                         /*
 397                          * See vmbus_negotiate_version(): VMBus protocol 5.0
 398                          * requires that we must use
 399                          * VMBUS_MESSAGE_CONNECTION_ID_4 for the Initiate
 400                          * Contact message, but on old hosts that only
 401                          * support VMBus protocol 4.0 or lower, here we get
 402                          * HV_STATUS_INVALID_CONNECTION_ID and we should
 403                          * return an error immediately without retrying.
 404                          */
 405                         hdr = buffer;
 406                         if (hdr->msgtype == CHANNELMSG_INITIATE_CONTACT)
 407                                 return -EINVAL;
 408                         /*
 409                          * We could get this if we send messages too
 410                          * frequently.
 411                          */
 412                         ret = -EAGAIN;
 413                         break;
 414                 case HV_STATUS_INSUFFICIENT_MEMORY:
 415                 case HV_STATUS_INSUFFICIENT_BUFFERS:
 416                         ret = -ENOBUFS;
 417                         break;
 418                 case HV_STATUS_SUCCESS:
 419                         return ret;
 420                 default:
 421                         pr_err("hv_post_msg() failed; error code:%d\n", ret);
 422                         return -EINVAL;
 423                 }
 424
 425                 retries++;
 426                 if (can_sleep && usec > 1000)
 427                         msleep(usec / 1000);
 428                 else if (usec < MAX_UDELAY_MS * 1000)
 429                         udelay(usec);
 430                 else
 431                         mdelay(usec / 1000);
 432
 433                 if (retries < 22)
 434                         usec *= 2;
 435         }
 436         return ret;
 437 }
 438
 439 /*
 440  * vmbus_set_event - Send an event notification to the parent
 441  */
 442 void vmbus_set_event(struct vmbus_channel *channel)
 443 {
 444         u32 child_relid = channel->offermsg.child_relid;
 445
 446         if (!channel->is_dedicated_interrupt)
 447                 vmbus_send_interrupt(child_relid);
 448
 449         ++channel->sig_events;
 450
 451         hv_do_fast_hypercall8(HVCALL_SIGNAL_EVENT, channel->sig_event);
 452 }
 453 EXPORT_SYMBOL_GPL(vmbus_set_event);