1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2009, Microsoft Corporation.
6 * Haiyang Zhang <haiyangz@microsoft.com>
7 * Hank Janssen <hjanssen@microsoft.com>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/interrupt.h>
13 #include <linux/sched.h>
14 #include <linux/wait.h>
16 #include <linux/slab.h>
17 #include <linux/list.h>
18 #include <linux/module.h>
19 #include <linux/completion.h>
20 #include <linux/delay.h>
21 #include <linux/cpu.h>
22 #include <linux/hyperv.h>
23 #include <asm/mshyperv.h>
25 #include "hyperv_vmbus.h"
27 static void init_vp_index(struct vmbus_channel *channel);
29 const struct vmbus_device vmbus_devs[] = {
34 .allowed_in_isolated = false,
38 { .dev_type = HV_SCSI,
41 .allowed_in_isolated = true,
48 .allowed_in_isolated = false,
55 .allowed_in_isolated = true,
62 .allowed_in_isolated = false,
66 { .dev_type = HV_PCIE,
69 .allowed_in_isolated = false,
72 /* Synthetic Frame Buffer */
76 .allowed_in_isolated = false,
79 /* Synthetic Keyboard */
83 .allowed_in_isolated = false,
87 { .dev_type = HV_MOUSE,
90 .allowed_in_isolated = false,
97 .allowed_in_isolated = false,
103 .perf_device = false,
104 .allowed_in_isolated = true,
110 .perf_device = false,
111 .allowed_in_isolated = true,
115 { .dev_type = HV_SHUTDOWN,
117 .perf_device = false,
118 .allowed_in_isolated = true,
122 { .dev_type = HV_FCOPY,
124 .perf_device = false,
125 .allowed_in_isolated = false,
129 { .dev_type = HV_BACKUP,
131 .perf_device = false,
132 .allowed_in_isolated = false,
138 .perf_device = false,
139 .allowed_in_isolated = false,
143 { .dev_type = HV_UNKNOWN,
144 .perf_device = false,
145 .allowed_in_isolated = false,
149 static const struct {
151 } vmbus_unsupported_devs[] = {
159 * The rescinded channel may be blocked waiting for a response from the host;
162 static void vmbus_rescind_cleanup(struct vmbus_channel *channel)
164 struct vmbus_channel_msginfo *msginfo;
168 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
169 channel->rescind = true;
170 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
173 if (msginfo->waiting_channel == channel) {
174 complete(&msginfo->waitevent);
178 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
181 static bool is_unsupported_vmbus_devs(const guid_t *guid)
185 for (i = 0; i < ARRAY_SIZE(vmbus_unsupported_devs); i++)
186 if (guid_equal(guid, &vmbus_unsupported_devs[i].guid))
191 static u16 hv_get_dev_type(const struct vmbus_channel *channel)
193 const guid_t *guid = &channel->offermsg.offer.if_type;
196 if (is_hvsock_channel(channel) || is_unsupported_vmbus_devs(guid))
199 for (i = HV_IDE; i < HV_UNKNOWN; i++) {
200 if (guid_equal(guid, &vmbus_devs[i].guid))
203 pr_info("Unknown GUID: %pUl\n", guid);
208 * vmbus_prep_negotiate_resp() - Create default response for Negotiate message
209 * @icmsghdrp: Pointer to msg header structure
210 * @buf: Raw buffer channel data
211 * @buflen: Length of the raw buffer channel data.
212 * @fw_version: The framework versions we can support.
213 * @fw_vercnt: The size of @fw_version.
214 * @srv_version: The service versions we can support.
215 * @srv_vercnt: The size of @srv_version.
216 * @nego_fw_version: The selected framework version.
217 * @nego_srv_version: The selected service version.
219 * Note: Versions are given in decreasing order.
221 * Set up and fill in default negotiate response message.
222 * Mainly used by Hyper-V drivers.
224 bool vmbus_prep_negotiate_resp(struct icmsg_hdr *icmsghdrp, u8 *buf,
225 u32 buflen, const int *fw_version, int fw_vercnt,
226 const int *srv_version, int srv_vercnt,
227 int *nego_fw_version, int *nego_srv_version)
229 int icframe_major, icframe_minor;
230 int icmsg_major, icmsg_minor;
231 int fw_major, fw_minor;
232 int srv_major, srv_minor;
234 bool found_match = false;
235 struct icmsg_negotiate *negop;
237 /* Check that there's enough space for icframe_vercnt, icmsg_vercnt */
238 if (buflen < ICMSG_HDR + offsetof(struct icmsg_negotiate, reserved)) {
239 pr_err_ratelimited("Invalid icmsg negotiate\n");
243 icmsghdrp->icmsgsize = 0x10;
244 negop = (struct icmsg_negotiate *)&buf[ICMSG_HDR];
246 icframe_major = negop->icframe_vercnt;
249 icmsg_major = negop->icmsg_vercnt;
252 /* Validate negop packet */
253 if (icframe_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
254 icmsg_major > IC_VERSION_NEGOTIATION_MAX_VER_COUNT ||
255 ICMSG_NEGOTIATE_PKT_SIZE(icframe_major, icmsg_major) > buflen) {
256 pr_err_ratelimited("Invalid icmsg negotiate - icframe_major: %u, icmsg_major: %u\n",
257 icframe_major, icmsg_major);
262 * Select the framework version number we will
266 for (i = 0; i < fw_vercnt; i++) {
267 fw_major = (fw_version[i] >> 16);
268 fw_minor = (fw_version[i] & 0xFFFF);
270 for (j = 0; j < negop->icframe_vercnt; j++) {
271 if ((negop->icversion_data[j].major == fw_major) &&
272 (negop->icversion_data[j].minor == fw_minor)) {
273 icframe_major = negop->icversion_data[j].major;
274 icframe_minor = negop->icversion_data[j].minor;
289 for (i = 0; i < srv_vercnt; i++) {
290 srv_major = (srv_version[i] >> 16);
291 srv_minor = (srv_version[i] & 0xFFFF);
293 for (j = negop->icframe_vercnt;
294 (j < negop->icframe_vercnt + negop->icmsg_vercnt);
297 if ((negop->icversion_data[j].major == srv_major) &&
298 (negop->icversion_data[j].minor == srv_minor)) {
300 icmsg_major = negop->icversion_data[j].major;
301 icmsg_minor = negop->icversion_data[j].minor;
312 * Respond with the framework and service
313 * version numbers we can support.
318 negop->icframe_vercnt = 0;
319 negop->icmsg_vercnt = 0;
321 negop->icframe_vercnt = 1;
322 negop->icmsg_vercnt = 1;
326 *nego_fw_version = (icframe_major << 16) | icframe_minor;
328 if (nego_srv_version)
329 *nego_srv_version = (icmsg_major << 16) | icmsg_minor;
331 negop->icversion_data[0].major = icframe_major;
332 negop->icversion_data[0].minor = icframe_minor;
333 negop->icversion_data[1].major = icmsg_major;
334 negop->icversion_data[1].minor = icmsg_minor;
337 EXPORT_SYMBOL_GPL(vmbus_prep_negotiate_resp);
340 * alloc_channel - Allocate and initialize a vmbus channel object
342 static struct vmbus_channel *alloc_channel(void)
344 struct vmbus_channel *channel;
346 channel = kzalloc(sizeof(*channel), GFP_ATOMIC);
350 spin_lock_init(&channel->sched_lock);
351 init_completion(&channel->rescind_event);
353 INIT_LIST_HEAD(&channel->sc_list);
355 tasklet_init(&channel->callback_event,
356 vmbus_on_event, (unsigned long)channel);
358 hv_ringbuffer_pre_init(channel);
364 * free_channel - Release the resources used by the vmbus channel object
366 static void free_channel(struct vmbus_channel *channel)
368 tasklet_kill(&channel->callback_event);
369 vmbus_remove_channel_attr_group(channel);
371 kobject_put(&channel->kobj);
374 void vmbus_channel_map_relid(struct vmbus_channel *channel)
376 if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
379 * The mapping of the channel's relid is visible from the CPUs that
380 * execute vmbus_chan_sched() by the time that vmbus_chan_sched() will
383 * (a) In the "normal (i.e., not resuming from hibernation)" path,
384 * the full barrier in virt_store_mb() guarantees that the store
385 * is propagated to all CPUs before the add_channel_work work
386 * is queued. In turn, add_channel_work is queued before the
387 * channel's ring buffer is allocated/initialized and the
388 * OPENCHANNEL message for the channel is sent in vmbus_open().
389 * Hyper-V won't start sending the interrupts for the channel
390 * before the OPENCHANNEL message is acked. The memory barrier
391 * in vmbus_chan_sched() -> sync_test_and_clear_bit() ensures
392 * that vmbus_chan_sched() must find the channel's relid in
393 * recv_int_page before retrieving the channel pointer from the
396 * (b) In the "resuming from hibernation" path, the virt_store_mb()
397 * guarantees that the store is propagated to all CPUs before
398 * the VMBus connection is marked as ready for the resume event
399 * (cf. check_ready_for_resume_event()). The interrupt handler
400 * of the VMBus driver and vmbus_chan_sched() can not run before
401 * vmbus_bus_resume() has completed execution (cf. resume_noirq).
404 vmbus_connection.channels[channel->offermsg.child_relid],
408 void vmbus_channel_unmap_relid(struct vmbus_channel *channel)
410 if (WARN_ON(channel->offermsg.child_relid >= MAX_CHANNEL_RELIDS))
413 vmbus_connection.channels[channel->offermsg.child_relid],
417 static void vmbus_release_relid(u32 relid)
419 struct vmbus_channel_relid_released msg;
422 memset(&msg, 0, sizeof(struct vmbus_channel_relid_released));
423 msg.child_relid = relid;
424 msg.header.msgtype = CHANNELMSG_RELID_RELEASED;
425 ret = vmbus_post_msg(&msg, sizeof(struct vmbus_channel_relid_released),
428 trace_vmbus_release_relid(&msg, ret);
431 void hv_process_channel_removal(struct vmbus_channel *channel)
433 lockdep_assert_held(&vmbus_connection.channel_mutex);
434 BUG_ON(!channel->rescind);
437 * hv_process_channel_removal() could find INVALID_RELID only for
438 * hv_sock channels. See the inline comments in vmbus_onoffer().
440 WARN_ON(channel->offermsg.child_relid == INVALID_RELID &&
441 !is_hvsock_channel(channel));
444 * Upon suspend, an in-use hv_sock channel is removed from the array of
445 * channels and the relid is invalidated. After hibernation, when the
446 * user-space application destroys the channel, it's unnecessary and
447 * unsafe to remove the channel from the array of channels. See also
448 * the inline comments before the call of vmbus_release_relid() below.
450 if (channel->offermsg.child_relid != INVALID_RELID)
451 vmbus_channel_unmap_relid(channel);
453 if (channel->primary_channel == NULL)
454 list_del(&channel->listentry);
456 list_del(&channel->sc_list);
459 * If this is a "perf" channel, updates the hv_numa_map[] masks so that
460 * init_vp_index() can (re-)use the CPU.
462 if (hv_is_perf_channel(channel))
463 hv_clear_allocated_cpu(channel->target_cpu);
466 * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
467 * the relid is invalidated; after hibernation, when the user-space app
468 * destroys the channel, the relid is INVALID_RELID, and in this case
469 * it's unnecessary and unsafe to release the old relid, since the same
470 * relid can refer to a completely different channel now.
472 if (channel->offermsg.child_relid != INVALID_RELID)
473 vmbus_release_relid(channel->offermsg.child_relid);
475 free_channel(channel);
478 void vmbus_free_channels(void)
480 struct vmbus_channel *channel, *tmp;
482 list_for_each_entry_safe(channel, tmp, &vmbus_connection.chn_list,
484 /* hv_process_channel_removal() needs this */
485 channel->rescind = true;
487 vmbus_device_unregister(channel->device_obj);
491 /* Note: the function can run concurrently for primary/sub channels. */
492 static void vmbus_add_channel_work(struct work_struct *work)
494 struct vmbus_channel *newchannel =
495 container_of(work, struct vmbus_channel, add_channel_work);
496 struct vmbus_channel *primary_channel = newchannel->primary_channel;
500 * This state is used to indicate a successful open
501 * so that when we do close the channel normally, we
502 * can cleanup properly.
504 newchannel->state = CHANNEL_OPEN_STATE;
506 if (primary_channel != NULL) {
507 /* newchannel is a sub-channel. */
508 struct hv_device *dev = primary_channel->device_obj;
510 if (vmbus_add_channel_kobj(dev, newchannel))
513 if (primary_channel->sc_creation_callback != NULL)
514 primary_channel->sc_creation_callback(newchannel);
516 newchannel->probe_done = true;
521 * Start the process of binding the primary channel to the driver
523 newchannel->device_obj = vmbus_device_create(
524 &newchannel->offermsg.offer.if_type,
525 &newchannel->offermsg.offer.if_instance,
527 if (!newchannel->device_obj)
530 newchannel->device_obj->device_id = newchannel->device_id;
532 * Add the new device to the bus. This will kick off device-driver
533 * binding which eventually invokes the device driver's AddDevice()
536 ret = vmbus_device_register(newchannel->device_obj);
539 pr_err("unable to add child device object (relid %d)\n",
540 newchannel->offermsg.child_relid);
541 kfree(newchannel->device_obj);
545 newchannel->probe_done = true;
549 mutex_lock(&vmbus_connection.channel_mutex);
552 * We need to set the flag, otherwise
553 * vmbus_onoffer_rescind() can be blocked.
555 newchannel->probe_done = true;
557 if (primary_channel == NULL)
558 list_del(&newchannel->listentry);
560 list_del(&newchannel->sc_list);
562 /* vmbus_process_offer() has mapped the channel. */
563 vmbus_channel_unmap_relid(newchannel);
565 mutex_unlock(&vmbus_connection.channel_mutex);
567 vmbus_release_relid(newchannel->offermsg.child_relid);
569 free_channel(newchannel);
573 * vmbus_process_offer - Process the offer by creating a channel/device
574 * associated with this offer
576 static void vmbus_process_offer(struct vmbus_channel *newchannel)
578 struct vmbus_channel *channel;
579 struct workqueue_struct *wq;
583 * Synchronize vmbus_process_offer() and CPU hotplugging:
587 * [vmbus_process_offer()] [Hot removal of the CPU]
589 * CPU_READ_LOCK CPUS_WRITE_LOCK
590 * LOAD cpu_online_mask SEARCH chn_list
591 * STORE target_cpu LOAD target_cpu
592 * INSERT chn_list STORE cpu_online_mask
593 * CPUS_READ_UNLOCK CPUS_WRITE_UNLOCK
595 * Forbids: CPU1's LOAD from *not* seing CPU2's STORE &&
596 * CPU2's SEARCH from *not* seeing CPU1's INSERT
598 * Forbids: CPU2's SEARCH from seeing CPU1's INSERT &&
599 * CPU2's LOAD from *not* seing CPU1's STORE
604 * Serializes the modifications of the chn_list list as well as
605 * the accesses to next_numa_node_id in init_vp_index().
607 mutex_lock(&vmbus_connection.channel_mutex);
609 list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
610 if (guid_equal(&channel->offermsg.offer.if_type,
611 &newchannel->offermsg.offer.if_type) &&
612 guid_equal(&channel->offermsg.offer.if_instance,
613 &newchannel->offermsg.offer.if_instance)) {
615 newchannel->primary_channel = channel;
620 init_vp_index(newchannel);
622 /* Remember the channels that should be cleaned up upon suspend. */
623 if (is_hvsock_channel(newchannel) || is_sub_channel(newchannel))
624 atomic_inc(&vmbus_connection.nr_chan_close_on_suspend);
627 * Now that we have acquired the channel_mutex,
628 * we can release the potentially racing rescind thread.
630 atomic_dec(&vmbus_connection.offer_in_progress);
633 list_add_tail(&newchannel->listentry,
634 &vmbus_connection.chn_list);
637 * Check to see if this is a valid sub-channel.
639 if (newchannel->offermsg.offer.sub_channel_index == 0) {
640 mutex_unlock(&vmbus_connection.channel_mutex);
642 * Don't call free_channel(), because newchannel->kobj
643 * is not initialized yet.
650 * Process the sub-channel.
652 list_add_tail(&newchannel->sc_list, &channel->sc_list);
655 vmbus_channel_map_relid(newchannel);
657 mutex_unlock(&vmbus_connection.channel_mutex);
661 * vmbus_process_offer() mustn't call channel->sc_creation_callback()
662 * directly for sub-channels, because sc_creation_callback() ->
663 * vmbus_open() may never get the host's response to the
664 * OPEN_CHANNEL message (the host may rescind a channel at any time,
665 * e.g. in the case of hot removing a NIC), and vmbus_onoffer_rescind()
666 * may not wake up the vmbus_open() as it's blocked due to a non-zero
667 * vmbus_connection.offer_in_progress, and finally we have a deadlock.
669 * The above is also true for primary channels, if the related device
670 * drivers use sync probing mode by default.
672 * And, usually the handling of primary channels and sub-channels can
673 * depend on each other, so we should offload them to different
674 * workqueues to avoid possible deadlock, e.g. in sync-probing mode,
675 * NIC1's netvsc_subchan_work() can race with NIC2's netvsc_probe() ->
676 * rtnl_lock(), and causes deadlock: the former gets the rtnl_lock
677 * and waits for all the sub-channels to appear, but the latter
678 * can't get the rtnl_lock and this blocks the handling of
681 INIT_WORK(&newchannel->add_channel_work, vmbus_add_channel_work);
682 wq = fnew ? vmbus_connection.handle_primary_chan_wq :
683 vmbus_connection.handle_sub_chan_wq;
684 queue_work(wq, &newchannel->add_channel_work);
688 * Check if CPUs used by other channels of the same device.
689 * It should only be called by init_vp_index().
691 static bool hv_cpuself_used(u32 cpu, struct vmbus_channel *chn)
693 struct vmbus_channel *primary = chn->primary_channel;
694 struct vmbus_channel *sc;
696 lockdep_assert_held(&vmbus_connection.channel_mutex);
701 if (primary->target_cpu == cpu)
704 list_for_each_entry(sc, &primary->sc_list, sc_list)
705 if (sc != chn && sc->target_cpu == cpu)
712 * We use this state to statically distribute the channel interrupt load.
714 static int next_numa_node_id;
717 * We can statically distribute the incoming channel interrupt load
718 * by binding a channel to VCPU.
720 * For non-performance critical channels we assign the VMBUS_CONNECT_CPU.
721 * Performance critical channels will be distributed evenly among all
722 * the available NUMA nodes. Once the node is assigned, we will assign
723 * the CPU based on a simple round robin scheme.
725 static void init_vp_index(struct vmbus_channel *channel)
727 bool perf_chn = hv_is_perf_channel(channel);
728 u32 i, ncpu = num_online_cpus();
729 cpumask_var_t available_mask;
730 struct cpumask *allocated_mask;
735 !alloc_cpumask_var(&available_mask, GFP_KERNEL)) {
737 * If the channel is not a performance critical
738 * channel, bind it to VMBUS_CONNECT_CPU.
739 * In case alloc_cpumask_var() fails, bind it to
742 channel->target_cpu = VMBUS_CONNECT_CPU;
744 hv_set_allocated_cpu(VMBUS_CONNECT_CPU);
748 for (i = 1; i <= ncpu + 1; i++) {
750 numa_node = next_numa_node_id++;
751 if (numa_node == nr_node_ids) {
752 next_numa_node_id = 0;
755 if (cpumask_empty(cpumask_of_node(numa_node)))
759 allocated_mask = &hv_context.hv_numa_map[numa_node];
761 if (cpumask_equal(allocated_mask, cpumask_of_node(numa_node))) {
763 * We have cycled through all the CPUs in the node;
764 * reset the allocated map.
766 cpumask_clear(allocated_mask);
769 cpumask_xor(available_mask, allocated_mask,
770 cpumask_of_node(numa_node));
772 target_cpu = cpumask_first(available_mask);
773 cpumask_set_cpu(target_cpu, allocated_mask);
775 if (channel->offermsg.offer.sub_channel_index >= ncpu ||
776 i > ncpu || !hv_cpuself_used(target_cpu, channel))
780 channel->target_cpu = target_cpu;
782 free_cpumask_var(available_mask);
785 #define UNLOAD_DELAY_UNIT_MS 10 /* 10 milliseconds */
786 #define UNLOAD_WAIT_MS (100*1000) /* 100 seconds */
787 #define UNLOAD_WAIT_LOOPS (UNLOAD_WAIT_MS/UNLOAD_DELAY_UNIT_MS)
788 #define UNLOAD_MSG_MS (5*1000) /* Every 5 seconds */
789 #define UNLOAD_MSG_LOOPS (UNLOAD_MSG_MS/UNLOAD_DELAY_UNIT_MS)
791 static void vmbus_wait_for_unload(void)
795 struct hv_message *msg;
796 struct vmbus_channel_message_header *hdr;
800 * CHANNELMSG_UNLOAD_RESPONSE is always delivered to the CPU which was
801 * used for initial contact or to CPU0 depending on host version. When
802 * we're crashing on a different CPU let's hope that IRQ handler on
803 * the cpu which receives CHANNELMSG_UNLOAD_RESPONSE is still
804 * functional and vmbus_unload_response() will complete
805 * vmbus_connection.unload_event. If not, the last thing we can do is
806 * read message pages for all CPUs directly.
808 * Wait up to 100 seconds since an Azure host must writeback any dirty
809 * data in its disk cache before the VMbus UNLOAD request will
810 * complete. This flushing has been empirically observed to take up
811 * to 50 seconds in cases with a lot of dirty data, so allow additional
812 * leeway and for inaccuracies in mdelay(). But eventually time out so
813 * that the panic path can't get hung forever in case the response
814 * message isn't seen.
816 for (i = 1; i <= UNLOAD_WAIT_LOOPS; i++) {
817 if (completion_done(&vmbus_connection.unload_event))
820 for_each_online_cpu(cpu) {
821 struct hv_per_cpu_context *hv_cpu
822 = per_cpu_ptr(hv_context.cpu_context, cpu);
824 page_addr = hv_cpu->synic_message_page;
825 msg = (struct hv_message *)page_addr
826 + VMBUS_MESSAGE_SINT;
828 message_type = READ_ONCE(msg->header.message_type);
829 if (message_type == HVMSG_NONE)
832 hdr = (struct vmbus_channel_message_header *)
835 if (hdr->msgtype == CHANNELMSG_UNLOAD_RESPONSE)
836 complete(&vmbus_connection.unload_event);
838 vmbus_signal_eom(msg, message_type);
842 * Give a notice periodically so someone watching the
843 * serial output won't think it is completely hung.
845 if (!(i % UNLOAD_MSG_LOOPS))
846 pr_notice("Waiting for VMBus UNLOAD to complete\n");
848 mdelay(UNLOAD_DELAY_UNIT_MS);
850 pr_err("Continuing even though VMBus UNLOAD did not complete\n");
854 * We're crashing and already got the UNLOAD_RESPONSE, cleanup all
855 * maybe-pending messages on all CPUs to be able to receive new
856 * messages after we reconnect.
858 for_each_online_cpu(cpu) {
859 struct hv_per_cpu_context *hv_cpu
860 = per_cpu_ptr(hv_context.cpu_context, cpu);
862 page_addr = hv_cpu->synic_message_page;
863 msg = (struct hv_message *)page_addr + VMBUS_MESSAGE_SINT;
864 msg->header.message_type = HVMSG_NONE;
869 * vmbus_unload_response - Handler for the unload response.
871 static void vmbus_unload_response(struct vmbus_channel_message_header *hdr)
874 * This is a global event; just wakeup the waiting thread.
875 * Once we successfully unload, we can cleanup the monitor state.
877 * NB. A malicious or compromised Hyper-V could send a spurious
878 * message of type CHANNELMSG_UNLOAD_RESPONSE, and trigger a call
879 * of the complete() below. Make sure that unload_event has been
880 * initialized by the time this complete() is executed.
882 complete(&vmbus_connection.unload_event);
885 void vmbus_initiate_unload(bool crash)
887 struct vmbus_channel_message_header hdr;
889 if (xchg(&vmbus_connection.conn_state, DISCONNECTED) == DISCONNECTED)
892 /* Pre-Win2012R2 hosts don't support reconnect */
893 if (vmbus_proto_version < VERSION_WIN8_1)
896 reinit_completion(&vmbus_connection.unload_event);
897 memset(&hdr, 0, sizeof(struct vmbus_channel_message_header));
898 hdr.msgtype = CHANNELMSG_UNLOAD;
899 vmbus_post_msg(&hdr, sizeof(struct vmbus_channel_message_header),
903 * vmbus_initiate_unload() is also called on crash and the crash can be
904 * happening in an interrupt context, where scheduling is impossible.
907 wait_for_completion(&vmbus_connection.unload_event);
909 vmbus_wait_for_unload();
912 static void check_ready_for_resume_event(void)
915 * If all the old primary channels have been fixed up, then it's safe
918 if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
919 complete(&vmbus_connection.ready_for_resume_event);
922 static void vmbus_setup_channel_state(struct vmbus_channel *channel,
923 struct vmbus_channel_offer_channel *offer)
926 * Setup state for signalling the host.
928 channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
930 channel->is_dedicated_interrupt =
931 (offer->is_dedicated_interrupt != 0);
932 channel->sig_event = offer->connection_id;
934 memcpy(&channel->offermsg, offer,
935 sizeof(struct vmbus_channel_offer_channel));
936 channel->monitor_grp = (u8)offer->monitorid / 32;
937 channel->monitor_bit = (u8)offer->monitorid % 32;
938 channel->device_id = hv_get_dev_type(channel);
942 * find_primary_channel_by_offer - Get the channel object given the new offer.
943 * This is only used in the resume path of hibernation.
945 static struct vmbus_channel *
946 find_primary_channel_by_offer(const struct vmbus_channel_offer_channel *offer)
948 struct vmbus_channel *channel = NULL, *iter;
949 const guid_t *inst1, *inst2;
951 /* Ignore sub-channel offers. */
952 if (offer->offer.sub_channel_index != 0)
955 mutex_lock(&vmbus_connection.channel_mutex);
957 list_for_each_entry(iter, &vmbus_connection.chn_list, listentry) {
958 inst1 = &iter->offermsg.offer.if_instance;
959 inst2 = &offer->offer.if_instance;
961 if (guid_equal(inst1, inst2)) {
967 mutex_unlock(&vmbus_connection.channel_mutex);
972 static bool vmbus_is_valid_offer(const struct vmbus_channel_offer_channel *offer)
974 const guid_t *guid = &offer->offer.if_type;
977 if (!hv_is_isolation_supported())
980 if (is_hvsock_offer(offer))
983 for (i = 0; i < ARRAY_SIZE(vmbus_devs); i++) {
984 if (guid_equal(guid, &vmbus_devs[i].guid))
985 return vmbus_devs[i].allowed_in_isolated;
991 * vmbus_onoffer - Handler for channel offers from vmbus in parent partition.
994 static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
996 struct vmbus_channel_offer_channel *offer;
997 struct vmbus_channel *oldchannel, *newchannel;
1000 offer = (struct vmbus_channel_offer_channel *)hdr;
1002 trace_vmbus_onoffer(offer);
1004 if (!vmbus_is_valid_offer(offer)) {
1005 pr_err_ratelimited("Invalid offer %d from the host supporting isolation\n",
1006 offer->child_relid);
1007 atomic_dec(&vmbus_connection.offer_in_progress);
1011 oldchannel = find_primary_channel_by_offer(offer);
1013 if (oldchannel != NULL) {
1015 * We're resuming from hibernation: all the sub-channel and
1016 * hv_sock channels we had before the hibernation should have
1017 * been cleaned up, and now we must be seeing a re-offered
1018 * primary channel that we had before the hibernation.
1022 * { Initially: channel relid = INVALID_RELID,
1023 * channels[valid_relid] = NULL }
1027 * [vmbus_onoffer()] [vmbus_device_release()]
1029 * LOCK channel_mutex LOCK channel_mutex
1030 * STORE channel relid = valid_relid LOAD r1 = channel relid
1031 * MAP_RELID channel if (r1 != INVALID_RELID)
1032 * UNLOCK channel_mutex UNMAP_RELID channel
1033 * UNLOCK channel_mutex
1035 * Forbids: r1 == valid_relid &&
1036 * channels[valid_relid] == channel
1038 * Note. r1 can be INVALID_RELID only for an hv_sock channel.
1039 * None of the hv_sock channels which were present before the
1040 * suspend are re-offered upon the resume. See the WARN_ON()
1041 * in hv_process_channel_removal().
1043 mutex_lock(&vmbus_connection.channel_mutex);
1045 atomic_dec(&vmbus_connection.offer_in_progress);
1047 WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
1048 /* Fix up the relid. */
1049 oldchannel->offermsg.child_relid = offer->child_relid;
1051 offer_sz = sizeof(*offer);
1052 if (memcmp(offer, &oldchannel->offermsg, offer_sz) != 0) {
1054 * This is not an error, since the host can also change
1055 * the other field(s) of the offer, e.g. on WS RS5
1056 * (Build 17763), the offer->connection_id of the
1057 * Mellanox VF vmbus device can change when the host
1058 * reoffers the device upon resume.
1060 pr_debug("vmbus offer changed: relid=%d\n",
1061 offer->child_relid);
1063 print_hex_dump_debug("Old vmbus offer: ",
1064 DUMP_PREFIX_OFFSET, 16, 4,
1065 &oldchannel->offermsg, offer_sz,
1067 print_hex_dump_debug("New vmbus offer: ",
1068 DUMP_PREFIX_OFFSET, 16, 4,
1069 offer, offer_sz, false);
1071 /* Fix up the old channel. */
1072 vmbus_setup_channel_state(oldchannel, offer);
1075 /* Add the channel back to the array of channels. */
1076 vmbus_channel_map_relid(oldchannel);
1077 check_ready_for_resume_event();
1079 mutex_unlock(&vmbus_connection.channel_mutex);
1083 /* Allocate the channel object and save this offer. */
1084 newchannel = alloc_channel();
1086 vmbus_release_relid(offer->child_relid);
1087 atomic_dec(&vmbus_connection.offer_in_progress);
1088 pr_err("Unable to allocate channel object\n");
1092 vmbus_setup_channel_state(newchannel, offer);
1094 vmbus_process_offer(newchannel);
1097 static void check_ready_for_suspend_event(void)
1100 * If all the sub-channels or hv_sock channels have been cleaned up,
1101 * then it's safe to suspend.
1103 if (atomic_dec_and_test(&vmbus_connection.nr_chan_close_on_suspend))
1104 complete(&vmbus_connection.ready_for_suspend_event);
1108 * vmbus_onoffer_rescind - Rescind offer handler.
1110 * We queue a work item to process this offer synchronously
1112 static void vmbus_onoffer_rescind(struct vmbus_channel_message_header *hdr)
1114 struct vmbus_channel_rescind_offer *rescind;
1115 struct vmbus_channel *channel;
1117 bool clean_up_chan_for_suspend;
1119 rescind = (struct vmbus_channel_rescind_offer *)hdr;
1121 trace_vmbus_onoffer_rescind(rescind);
1124 * The offer msg and the corresponding rescind msg
1125 * from the host are guranteed to be ordered -
1126 * offer comes in first and then the rescind.
1127 * Since we process these events in work elements,
1128 * and with preemption, we may end up processing
1129 * the events out of order. We rely on the synchronization
1130 * provided by offer_in_progress and by channel_mutex for
1131 * ordering these events:
1133 * { Initially: offer_in_progress = 1 }
1137 * [vmbus_onoffer()] [vmbus_onoffer_rescind()]
1139 * LOCK channel_mutex WAIT_ON offer_in_progress == 0
1140 * DECREMENT offer_in_progress LOCK channel_mutex
1141 * STORE channels[] LOAD channels[]
1142 * UNLOCK channel_mutex UNLOCK channel_mutex
1144 * Forbids: CPU2's LOAD from *not* seeing CPU1's STORE
1147 while (atomic_read(&vmbus_connection.offer_in_progress) != 0) {
1149 * We wait here until any channel offer is currently
1155 mutex_lock(&vmbus_connection.channel_mutex);
1156 channel = relid2channel(rescind->child_relid);
1157 if (channel != NULL) {
1159 * Guarantee that no other instance of vmbus_onoffer_rescind()
1160 * has got a reference to the channel object. Synchronize on
1161 * &vmbus_connection.channel_mutex.
1163 if (channel->rescind_ref) {
1164 mutex_unlock(&vmbus_connection.channel_mutex);
1167 channel->rescind_ref = true;
1169 mutex_unlock(&vmbus_connection.channel_mutex);
1171 if (channel == NULL) {
1173 * We failed in processing the offer message;
1174 * we would have cleaned up the relid in that
1180 clean_up_chan_for_suspend = is_hvsock_channel(channel) ||
1181 is_sub_channel(channel);
1183 * Before setting channel->rescind in vmbus_rescind_cleanup(), we
1184 * should make sure the channel callback is not running any more.
1186 vmbus_reset_channel_cb(channel);
1189 * Now wait for offer handling to complete.
1191 vmbus_rescind_cleanup(channel);
1192 while (READ_ONCE(channel->probe_done) == false) {
1194 * We wait here until any channel offer is currently
1201 * At this point, the rescind handling can proceed safely.
1204 if (channel->device_obj) {
1205 if (channel->chn_rescind_callback) {
1206 channel->chn_rescind_callback(channel);
1208 if (clean_up_chan_for_suspend)
1209 check_ready_for_suspend_event();
1214 * We will have to unregister this device from the
1217 dev = get_device(&channel->device_obj->device);
1219 vmbus_device_unregister(channel->device_obj);
1222 } else if (channel->primary_channel != NULL) {
1224 * Sub-channel is being rescinded. Following is the channel
1225 * close sequence when initiated from the driveri (refer to
1226 * vmbus_close() for details):
1227 * 1. Close all sub-channels first
1228 * 2. Then close the primary channel.
1230 mutex_lock(&vmbus_connection.channel_mutex);
1231 if (channel->state == CHANNEL_OPEN_STATE) {
1233 * The channel is currently not open;
1234 * it is safe for us to cleanup the channel.
1236 hv_process_channel_removal(channel);
1238 complete(&channel->rescind_event);
1240 mutex_unlock(&vmbus_connection.channel_mutex);
1243 /* The "channel" may have been freed. Do not access it any longer. */
1245 if (clean_up_chan_for_suspend)
1246 check_ready_for_suspend_event();
1249 void vmbus_hvsock_device_unregister(struct vmbus_channel *channel)
1251 BUG_ON(!is_hvsock_channel(channel));
1253 /* We always get a rescind msg when a connection is closed. */
1254 while (!READ_ONCE(channel->probe_done) || !READ_ONCE(channel->rescind))
1257 vmbus_device_unregister(channel->device_obj);
1259 EXPORT_SYMBOL_GPL(vmbus_hvsock_device_unregister);
1263 * vmbus_onoffers_delivered -
1264 * This is invoked when all offers have been delivered.
1266 * Nothing to do here.
1268 static void vmbus_onoffers_delivered(
1269 struct vmbus_channel_message_header *hdr)
1274 * vmbus_onopen_result - Open result handler.
1276 * This is invoked when we received a response to our channel open request.
1277 * Find the matching request, copy the response and signal the requesting
1280 static void vmbus_onopen_result(struct vmbus_channel_message_header *hdr)
1282 struct vmbus_channel_open_result *result;
1283 struct vmbus_channel_msginfo *msginfo;
1284 struct vmbus_channel_message_header *requestheader;
1285 struct vmbus_channel_open_channel *openmsg;
1286 unsigned long flags;
1288 result = (struct vmbus_channel_open_result *)hdr;
1290 trace_vmbus_onopen_result(result);
1293 * Find the open msg, copy the result and signal/unblock the wait event
1295 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1297 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1300 (struct vmbus_channel_message_header *)msginfo->msg;
1302 if (requestheader->msgtype == CHANNELMSG_OPENCHANNEL) {
1304 (struct vmbus_channel_open_channel *)msginfo->msg;
1305 if (openmsg->child_relid == result->child_relid &&
1306 openmsg->openid == result->openid) {
1307 memcpy(&msginfo->response.open_result,
1310 struct vmbus_channel_open_result));
1311 complete(&msginfo->waitevent);
1316 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1320 * vmbus_ongpadl_created - GPADL created handler.
1322 * This is invoked when we received a response to our gpadl create request.
1323 * Find the matching request, copy the response and signal the requesting
1326 static void vmbus_ongpadl_created(struct vmbus_channel_message_header *hdr)
1328 struct vmbus_channel_gpadl_created *gpadlcreated;
1329 struct vmbus_channel_msginfo *msginfo;
1330 struct vmbus_channel_message_header *requestheader;
1331 struct vmbus_channel_gpadl_header *gpadlheader;
1332 unsigned long flags;
1334 gpadlcreated = (struct vmbus_channel_gpadl_created *)hdr;
1336 trace_vmbus_ongpadl_created(gpadlcreated);
1339 * Find the establish msg, copy the result and signal/unblock the wait
1342 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1344 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1347 (struct vmbus_channel_message_header *)msginfo->msg;
1349 if (requestheader->msgtype == CHANNELMSG_GPADL_HEADER) {
1351 (struct vmbus_channel_gpadl_header *)requestheader;
1353 if ((gpadlcreated->child_relid ==
1354 gpadlheader->child_relid) &&
1355 (gpadlcreated->gpadl == gpadlheader->gpadl)) {
1356 memcpy(&msginfo->response.gpadl_created,
1359 struct vmbus_channel_gpadl_created));
1360 complete(&msginfo->waitevent);
1365 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1369 * vmbus_onmodifychannel_response - Modify Channel response handler.
1371 * This is invoked when we received a response to our channel modify request.
1372 * Find the matching request, copy the response and signal the requesting thread.
1374 static void vmbus_onmodifychannel_response(struct vmbus_channel_message_header *hdr)
1376 struct vmbus_channel_modifychannel_response *response;
1377 struct vmbus_channel_msginfo *msginfo;
1378 unsigned long flags;
1380 response = (struct vmbus_channel_modifychannel_response *)hdr;
1382 trace_vmbus_onmodifychannel_response(response);
1385 * Find the modify msg, copy the response and signal/unblock the wait event.
1387 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1389 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list, msglistentry) {
1390 struct vmbus_channel_message_header *responseheader =
1391 (struct vmbus_channel_message_header *)msginfo->msg;
1393 if (responseheader->msgtype == CHANNELMSG_MODIFYCHANNEL) {
1394 struct vmbus_channel_modifychannel *modifymsg;
1396 modifymsg = (struct vmbus_channel_modifychannel *)msginfo->msg;
1397 if (modifymsg->child_relid == response->child_relid) {
1398 memcpy(&msginfo->response.modify_response, response,
1400 complete(&msginfo->waitevent);
1405 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1409 * vmbus_ongpadl_torndown - GPADL torndown handler.
1411 * This is invoked when we received a response to our gpadl teardown request.
1412 * Find the matching request, copy the response and signal the requesting
1415 static void vmbus_ongpadl_torndown(
1416 struct vmbus_channel_message_header *hdr)
1418 struct vmbus_channel_gpadl_torndown *gpadl_torndown;
1419 struct vmbus_channel_msginfo *msginfo;
1420 struct vmbus_channel_message_header *requestheader;
1421 struct vmbus_channel_gpadl_teardown *gpadl_teardown;
1422 unsigned long flags;
1424 gpadl_torndown = (struct vmbus_channel_gpadl_torndown *)hdr;
1426 trace_vmbus_ongpadl_torndown(gpadl_torndown);
1429 * Find the open msg, copy the result and signal/unblock the wait event
1431 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1433 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1436 (struct vmbus_channel_message_header *)msginfo->msg;
1438 if (requestheader->msgtype == CHANNELMSG_GPADL_TEARDOWN) {
1440 (struct vmbus_channel_gpadl_teardown *)requestheader;
1442 if (gpadl_torndown->gpadl == gpadl_teardown->gpadl) {
1443 memcpy(&msginfo->response.gpadl_torndown,
1446 struct vmbus_channel_gpadl_torndown));
1447 complete(&msginfo->waitevent);
1452 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1456 * vmbus_onversion_response - Version response handler
1458 * This is invoked when we received a response to our initiate contact request.
1459 * Find the matching request, copy the response and signal the requesting
1462 static void vmbus_onversion_response(
1463 struct vmbus_channel_message_header *hdr)
1465 struct vmbus_channel_msginfo *msginfo;
1466 struct vmbus_channel_message_header *requestheader;
1467 struct vmbus_channel_version_response *version_response;
1468 unsigned long flags;
1470 version_response = (struct vmbus_channel_version_response *)hdr;
1472 trace_vmbus_onversion_response(version_response);
1474 spin_lock_irqsave(&vmbus_connection.channelmsg_lock, flags);
1476 list_for_each_entry(msginfo, &vmbus_connection.chn_msg_list,
1479 (struct vmbus_channel_message_header *)msginfo->msg;
1481 if (requestheader->msgtype ==
1482 CHANNELMSG_INITIATE_CONTACT) {
1483 memcpy(&msginfo->response.version_response,
1485 sizeof(struct vmbus_channel_version_response));
1486 complete(&msginfo->waitevent);
1489 spin_unlock_irqrestore(&vmbus_connection.channelmsg_lock, flags);
1492 /* Channel message dispatch table */
1493 const struct vmbus_channel_message_table_entry
1494 channel_message_table[CHANNELMSG_COUNT] = {
1495 { CHANNELMSG_INVALID, 0, NULL, 0},
1496 { CHANNELMSG_OFFERCHANNEL, 0, vmbus_onoffer,
1497 sizeof(struct vmbus_channel_offer_channel)},
1498 { CHANNELMSG_RESCIND_CHANNELOFFER, 0, vmbus_onoffer_rescind,
1499 sizeof(struct vmbus_channel_rescind_offer) },
1500 { CHANNELMSG_REQUESTOFFERS, 0, NULL, 0},
1501 { CHANNELMSG_ALLOFFERS_DELIVERED, 1, vmbus_onoffers_delivered, 0},
1502 { CHANNELMSG_OPENCHANNEL, 0, NULL, 0},
1503 { CHANNELMSG_OPENCHANNEL_RESULT, 1, vmbus_onopen_result,
1504 sizeof(struct vmbus_channel_open_result)},
1505 { CHANNELMSG_CLOSECHANNEL, 0, NULL, 0},
1506 { CHANNELMSG_GPADL_HEADER, 0, NULL, 0},
1507 { CHANNELMSG_GPADL_BODY, 0, NULL, 0},
1508 { CHANNELMSG_GPADL_CREATED, 1, vmbus_ongpadl_created,
1509 sizeof(struct vmbus_channel_gpadl_created)},
1510 { CHANNELMSG_GPADL_TEARDOWN, 0, NULL, 0},
1511 { CHANNELMSG_GPADL_TORNDOWN, 1, vmbus_ongpadl_torndown,
1512 sizeof(struct vmbus_channel_gpadl_torndown) },
1513 { CHANNELMSG_RELID_RELEASED, 0, NULL, 0},
1514 { CHANNELMSG_INITIATE_CONTACT, 0, NULL, 0},
1515 { CHANNELMSG_VERSION_RESPONSE, 1, vmbus_onversion_response,
1516 sizeof(struct vmbus_channel_version_response)},
1517 { CHANNELMSG_UNLOAD, 0, NULL, 0},
1518 { CHANNELMSG_UNLOAD_RESPONSE, 1, vmbus_unload_response, 0},
1519 { CHANNELMSG_18, 0, NULL, 0},
1520 { CHANNELMSG_19, 0, NULL, 0},
1521 { CHANNELMSG_20, 0, NULL, 0},
1522 { CHANNELMSG_TL_CONNECT_REQUEST, 0, NULL, 0},
1523 { CHANNELMSG_MODIFYCHANNEL, 0, NULL, 0},
1524 { CHANNELMSG_TL_CONNECT_RESULT, 0, NULL, 0},
1525 { CHANNELMSG_MODIFYCHANNEL_RESPONSE, 1, vmbus_onmodifychannel_response,
1526 sizeof(struct vmbus_channel_modifychannel_response)},
1530 * vmbus_onmessage - Handler for channel protocol messages.
1532 * This is invoked in the vmbus worker thread context.
1534 void vmbus_onmessage(struct vmbus_channel_message_header *hdr)
1536 trace_vmbus_on_message(hdr);
1539 * vmbus_on_msg_dpc() makes sure the hdr->msgtype here can not go
1540 * out of bound and the message_handler pointer can not be NULL.
1542 channel_message_table[hdr->msgtype].message_handler(hdr);
1546 * vmbus_request_offers - Send a request to get all our pending offers.
1548 int vmbus_request_offers(void)
1550 struct vmbus_channel_message_header *msg;
1551 struct vmbus_channel_msginfo *msginfo;
1554 msginfo = kzalloc(sizeof(*msginfo) +
1555 sizeof(struct vmbus_channel_message_header),
1560 msg = (struct vmbus_channel_message_header *)msginfo->msg;
1562 msg->msgtype = CHANNELMSG_REQUESTOFFERS;
1564 ret = vmbus_post_msg(msg, sizeof(struct vmbus_channel_message_header),
1567 trace_vmbus_request_offers(ret);
1570 pr_err("Unable to request offers - %d\n", ret);
1581 void vmbus_set_sc_create_callback(struct vmbus_channel *primary_channel,
1582 void (*sc_cr_cb)(struct vmbus_channel *new_sc))
1584 primary_channel->sc_creation_callback = sc_cr_cb;
1586 EXPORT_SYMBOL_GPL(vmbus_set_sc_create_callback);
1588 void vmbus_set_chn_rescind_callback(struct vmbus_channel *channel,
1589 void (*chn_rescind_cb)(struct vmbus_channel *))
1591 channel->chn_rescind_callback = chn_rescind_cb;
1593 EXPORT_SYMBOL_GPL(vmbus_set_chn_rescind_callback);