]> Git Repo - linux.git/commitdiff
Drivers: hv: vmbus: Resume after fixing up old primary channels
authorDexuan Cui <[email protected]>
Thu, 5 Sep 2019 23:01:22 +0000 (23:01 +0000)
committerSasha Levin <[email protected]>
Fri, 6 Sep 2019 18:52:44 +0000 (14:52 -0400)
When the host re-offers the primary channels upon resume, the host only
guarantees the Instance GUID  doesn't change, so vmbus_bus_suspend()
should invalidate channel->offermsg.child_relid and figure out the
number of primary channels that need to be fixed up upon resume.

Upon resume, vmbus_onoffer() finds the old channel structs, and maps
the new offers to the old channels, and fixes up the old structs,
and finally the resume callbacks of the VSC drivers will re-open
the channels.

Signed-off-by: Dexuan Cui <[email protected]>
Reviewed-by: Michael Kelley <[email protected]>
Signed-off-by: Sasha Levin <[email protected]>
drivers/hv/channel_mgmt.c
drivers/hv/connection.c
drivers/hv/hyperv_vmbus.h
drivers/hv/vmbus_drv.c
include/linux/hyperv.h

index 5518d031f62a818e195be7b44c58158e5569fe66..8eb167540b4f66e658e7977858635f6dc8139700 100644 (file)
@@ -407,7 +407,15 @@ void hv_process_channel_removal(struct vmbus_channel *channel)
                cpumask_clear_cpu(channel->target_cpu,
                                  &primary_channel->alloced_cpus_in_node);
 
-       vmbus_release_relid(channel->offermsg.child_relid);
+       /*
+        * Upon suspend, an in-use hv_sock channel is marked as "rescinded" and
+        * the relid is invalidated; after hibernation, when the user-space app
+        * destroys the channel, the relid is INVALID_RELID, and in this case
+        * it's unnecessary and unsafe to release the old relid, since the same
+        * relid can refer to a completely different channel now.
+        */
+       if (channel->offermsg.child_relid != INVALID_RELID)
+               vmbus_release_relid(channel->offermsg.child_relid);
 
        free_channel(channel);
 }
@@ -851,6 +859,36 @@ void vmbus_initiate_unload(bool crash)
                vmbus_wait_for_unload();
 }
 
+static void check_ready_for_resume_event(void)
+{
+       /*
+        * If all the old primary channels have been fixed up, then it's safe
+        * to resume.
+        */
+       if (atomic_dec_and_test(&vmbus_connection.nr_chan_fixup_on_resume))
+               complete(&vmbus_connection.ready_for_resume_event);
+}
+
+static void vmbus_setup_channel_state(struct vmbus_channel *channel,
+                                     struct vmbus_channel_offer_channel *offer)
+{
+       /*
+        * Setup state for signalling the host.
+        */
+       channel->sig_event = VMBUS_EVENT_CONNECTION_ID;
+
+       if (vmbus_proto_version != VERSION_WS2008) {
+               channel->is_dedicated_interrupt =
+                               (offer->is_dedicated_interrupt != 0);
+               channel->sig_event = offer->connection_id;
+       }
+
+       memcpy(&channel->offermsg, offer,
+              sizeof(struct vmbus_channel_offer_channel));
+       channel->monitor_grp = (u8)offer->monitorid / 32;
+       channel->monitor_bit = (u8)offer->monitorid % 32;
+}
+
 /*
  * find_primary_channel_by_offer - Get the channel object given the new offer.
  * This is only used in the resume path of hibernation.
@@ -902,14 +940,29 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
                atomic_dec(&vmbus_connection.offer_in_progress);
 
                /*
-                * We're resuming from hibernation: we expect the host to send
-                * exactly the same offers that we had before the hibernation.
+                * We're resuming from hibernation: all the sub-channel and
+                * hv_sock channels we had before the hibernation should have
+                * been cleaned up, and now we must be seeing a re-offered
+                * primary channel that we had before the hibernation.
                 */
+
+               WARN_ON(oldchannel->offermsg.child_relid != INVALID_RELID);
+               /* Fix up the relid. */
+               oldchannel->offermsg.child_relid = offer->child_relid;
+
                offer_sz = sizeof(*offer);
-               if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0)
+               if (memcmp(offer, &oldchannel->offermsg, offer_sz) == 0) {
+                       check_ready_for_resume_event();
                        return;
+               }
 
-               pr_debug("Mismatched offer from the host (relid=%d)\n",
+               /*
+                * This is not an error, since the host can also change the
+                * other field(s) of the offer, e.g. on WS RS5 (Build 17763),
+                * the offer->connection_id of the Mellanox VF vmbus device
+                * can change when the host reoffers the device upon resume.
+                */
+               pr_debug("vmbus offer changed: relid=%d\n",
                         offer->child_relid);
 
                print_hex_dump_debug("Old vmbus offer: ", DUMP_PREFIX_OFFSET,
@@ -917,6 +970,12 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
                                     false);
                print_hex_dump_debug("New vmbus offer: ", DUMP_PREFIX_OFFSET,
                                     16, 4, offer, offer_sz, false);
+
+               /* Fix up the old channel. */
+               vmbus_setup_channel_state(oldchannel, offer);
+
+               check_ready_for_resume_event();
+
                return;
        }
 
@@ -929,21 +988,7 @@ static void vmbus_onoffer(struct vmbus_channel_message_header *hdr)
                return;
        }
 
-       /*
-        * Setup state for signalling the host.
-        */
-       newchannel->sig_event = VMBUS_EVENT_CONNECTION_ID;
-
-       if (vmbus_proto_version != VERSION_WS2008) {
-               newchannel->is_dedicated_interrupt =
-                               (offer->is_dedicated_interrupt != 0);
-               newchannel->sig_event = offer->connection_id;
-       }
-
-       memcpy(&newchannel->offermsg, offer,
-              sizeof(struct vmbus_channel_offer_channel));
-       newchannel->monitor_grp = (u8)offer->monitorid / 32;
-       newchannel->monitor_bit = (u8)offer->monitorid % 32;
+       vmbus_setup_channel_state(newchannel, offer);
 
        vmbus_process_offer(newchannel);
 }
index 99851ea682eb26515a826d469a8c64893bd9e1b3..6e4c015783ffca5ec063c59ff5a478d2dc93de46 100644 (file)
@@ -29,6 +29,8 @@ struct vmbus_connection vmbus_connection = {
 
        .ready_for_suspend_event= COMPLETION_INITIALIZER(
                                  vmbus_connection.ready_for_suspend_event),
+       .ready_for_resume_event = COMPLETION_INITIALIZER(
+                                 vmbus_connection.ready_for_resume_event),
 };
 EXPORT_SYMBOL_GPL(vmbus_connection);
 
index 974b747ca1fc03c1d3bf989fa5acc7c7e5c46418..f7a5f5615f34cb47b116fbcc2e5f82b40b40d8bd 100644 (file)
@@ -272,6 +272,20 @@ struct vmbus_connection {
         * drop to zero.
         */
        struct completion ready_for_suspend_event;
+
+       /*
+        * The number of primary channels that should be "fixed up"
+        * upon resume: these channels are re-offered upon resume, and some
+        * fields of the channel offers (i.e. child_relid and connection_id)
+        * can change, so the old offermsg must be fixed up, before the resume
+        * callbacks of the VSC drivers start to further touch the channels.
+        */
+       atomic_t nr_chan_fixup_on_resume;
+       /*
+        * vmbus_bus_resume() waits for "nr_chan_fixup_on_resume" to
+        * drop to zero.
+        */
+       struct completion ready_for_resume_event;
 };
 
 
index 32ec951d334f61bfd193451eb56a35ab152e79ce..391f0b225c9ae4698c3cc3d9958b10873cf861a0 100644 (file)
@@ -2164,9 +2164,17 @@ static int vmbus_bus_suspend(struct device *dev)
        if (atomic_read(&vmbus_connection.nr_chan_close_on_suspend) > 0)
                wait_for_completion(&vmbus_connection.ready_for_suspend_event);
 
+       WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) != 0);
+
        mutex_lock(&vmbus_connection.channel_mutex);
 
        list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
+               /*
+                * Invalidate the field. Upon resume, vmbus_onoffer() will fix
+                * up the field, and the other fields (if necessary).
+                */
+               channel->offermsg.child_relid = INVALID_RELID;
+
                if (is_hvsock_channel(channel)) {
                        if (!channel->rescind) {
                                pr_err("hv_sock channel not rescinded!\n");
@@ -2181,6 +2189,8 @@ static int vmbus_bus_suspend(struct device *dev)
                        WARN_ON_ONCE(1);
                }
                spin_unlock_irqrestore(&channel->lock, flags);
+
+               atomic_inc(&vmbus_connection.nr_chan_fixup_on_resume);
        }
 
        mutex_unlock(&vmbus_connection.channel_mutex);
@@ -2189,6 +2199,9 @@ static int vmbus_bus_suspend(struct device *dev)
 
        vmbus_connection.conn_state = DISCONNECTED;
 
+       /* Reset the event for the next resume. */
+       reinit_completion(&vmbus_connection.ready_for_resume_event);
+
        return 0;
 }
 
@@ -2223,8 +2236,12 @@ static int vmbus_bus_resume(struct device *dev)
        if (ret != 0)
                return ret;
 
+       WARN_ON(atomic_read(&vmbus_connection.nr_chan_fixup_on_resume) == 0);
+
        vmbus_request_offers();
 
+       wait_for_completion(&vmbus_connection.ready_for_resume_event);
+
        /* Reset the event for the next suspend. */
        reinit_completion(&vmbus_connection.ready_for_suspend_event);
 
index 8a60e7766037d669db5ce65475af076af0fbf190..a3aa9e9ef6f23d6c491b169b8f6a4fc445ebbe7f 100644 (file)
@@ -426,6 +426,9 @@ enum vmbus_channel_message_type {
        CHANNELMSG_COUNT
 };
 
+/* Hyper-V supports about 2048 channels, and the RELIDs start with 1. */
+#define INVALID_RELID  U32_MAX
+
 struct vmbus_channel_message_header {
        enum vmbus_channel_message_type msgtype;
        u32 padding;
This page took 0.064744 seconds and 4 git commands to generate.