Drivers: hv: Add Hyper-V balloon driver
K. Y. Srinivasan [Wed, 14 Nov 2012 09:09:02 +0000 (01:09 -0800)]
Add the basic balloon driver. Windows hosts dynamically manage the guest
memory allocation via a combination memory hot add and ballooning. Memory
hot add is used to grow the guest memory upto the maximum memory that can be
allocatted to the guest. Ballooning is used to both shrink as well as expand
up to the max memory. Supporting hot add needs additional support from the
host. We will support hot add when this support is available. For now,
by setting the VM startup memory to the VM  max memory, we can use
ballooning alone to dynamically manage memory allocation amongst
competing guests on a given host.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/hv/Kconfig
drivers/hv/Makefile
drivers/hv/hv_balloon.c [new file with mode: 0644]

index 70f5dde..b38ef6d 100644 (file)
@@ -13,4 +13,10 @@ config HYPERV_UTILS
        help
          Select this option to enable the Hyper-V Utilities.
 
+config HYPERV_BALLOON
+       tristate "Microsoft Hyper-V Balloon driver"
+       depends on HYPERV
+       help
+         Select this option to enable Hyper-V Balloon driver.
+
 endmenu
index a23938b..e6abfa0 100644 (file)
@@ -1,5 +1,6 @@
 obj-$(CONFIG_HYPERV)           += hv_vmbus.o
 obj-$(CONFIG_HYPERV_UTILS)     += hv_utils.o
+obj-$(CONFIG_HYPERV_BALLOON)   += hv_balloon.o
 
 hv_vmbus-y := vmbus_drv.o \
                 hv.o connection.o channel.o \
diff --git a/drivers/hv/hv_balloon.c b/drivers/hv/hv_balloon.c
new file mode 100644 (file)
index 0000000..bbc4973
--- /dev/null
@@ -0,0 +1,1041 @@
+/*
+ * Copyright (c) 2012, Microsoft Corporation.
+ *
+ * Author:
+ *   K. Y. Srinivasan <kys@microsoft.com>
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published
+ * by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ * NON INFRINGEMENT.  See the GNU General Public License for more
+ * details.
+ *
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/kernel.h>
+#include <linux/mman.h>
+#include <linux/delay.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/kthread.h>
+#include <linux/completion.h>
+#include <linux/memory_hotplug.h>
+#include <linux/memory.h>
+#include <linux/notifier.h>
+#include <linux/mman.h>
+#include <linux/percpu_counter.h>
+
+#include <linux/hyperv.h>
+
+/*
+ * We begin with definitions supporting the Dynamic Memory protocol
+ * with the host.
+ *
+ * Begin protocol definitions.
+ */
+
+
+
+/*
+ * Protocol versions. The low word is the minor version, the high word the major
+ * version.
+ *
+ * History:
+ * Initial version 1.0
+ * Changed to 0.1 on 2009/03/25
+ * Changes to 0.2 on 2009/05/14
+ * Changes to 0.3 on 2009/12/03
+ * Changed to 1.0 on 2011/04/05
+ */
+
+#define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
+#define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
+#define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
+
+enum {
+       DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
+       DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
+
+       DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
+       DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
+
+       DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN8
+};
+
+
+
+/*
+ * Message Types
+ */
+
+enum dm_message_type {
+       /*
+        * Version 0.3
+        */
+       DM_ERROR                        = 0,
+       DM_VERSION_REQUEST              = 1,
+       DM_VERSION_RESPONSE             = 2,
+       DM_CAPABILITIES_REPORT          = 3,
+       DM_CAPABILITIES_RESPONSE        = 4,
+       DM_STATUS_REPORT                = 5,
+       DM_BALLOON_REQUEST              = 6,
+       DM_BALLOON_RESPONSE             = 7,
+       DM_UNBALLOON_REQUEST            = 8,
+       DM_UNBALLOON_RESPONSE           = 9,
+       DM_MEM_HOT_ADD_REQUEST          = 10,
+       DM_MEM_HOT_ADD_RESPONSE         = 11,
+       DM_VERSION_03_MAX               = 11,
+       /*
+        * Version 1.0.
+        */
+       DM_INFO_MESSAGE                 = 12,
+       DM_VERSION_1_MAX                = 12
+};
+
+
+/*
+ * Structures defining the dynamic memory management
+ * protocol.
+ */
+
+union dm_version {
+       struct {
+               __u16 minor_version;
+               __u16 major_version;
+       };
+       __u32 version;
+} __packed;
+
+
+union dm_caps {
+       struct {
+               __u64 balloon:1;
+               __u64 hot_add:1;
+               __u64 reservedz:62;
+       } cap_bits;
+       __u64 caps;
+} __packed;
+
+union dm_mem_page_range {
+       struct  {
+               /*
+                * The PFN number of the first page in the range.
+                * 40 bits is the architectural limit of a PFN
+                * number for AMD64.
+                */
+               __u64 start_page:40;
+               /*
+                * The number of pages in the range.
+                */
+               __u64 page_cnt:24;
+       } finfo;
+       __u64  page_range;
+} __packed;
+
+
+
+/*
+ * The header for all dynamic memory messages:
+ *
+ * type: Type of the message.
+ * size: Size of the message in bytes; including the header.
+ * trans_id: The guest is responsible for manufacturing this ID.
+ */
+
+struct dm_header {
+       __u16 type;
+       __u16 size;
+       __u32 trans_id;
+} __packed;
+
+/*
+ * A generic message format for dynamic memory.
+ * Specific message formats are defined later in the file.
+ */
+
+struct dm_message {
+       struct dm_header hdr;
+       __u8 data[]; /* enclosed message */
+} __packed;
+
+
+/*
+ * Specific message types supporting the dynamic memory protocol.
+ */
+
+/*
+ * Version negotiation message. Sent from the guest to the host.
+ * The guest is free to try different versions until the host
+ * accepts the version.
+ *
+ * dm_version: The protocol version requested.
+ * is_last_attempt: If TRUE, this is the last version guest will request.
+ * reservedz: Reserved field, set to zero.
+ */
+
+struct dm_version_request {
+       struct dm_header hdr;
+       union dm_version version;
+       __u32 is_last_attempt:1;
+       __u32 reservedz:31;
+} __packed;
+
+/*
+ * Version response message; Host to Guest and indicates
+ * if the host has accepted the version sent by the guest.
+ *
+ * is_accepted: If TRUE, host has accepted the version and the guest
+ * should proceed to the next stage of the protocol. FALSE indicates that
+ * guest should re-try with a different version.
+ *
+ * reservedz: Reserved field, set to zero.
+ */
+
+struct dm_version_response {
+       struct dm_header hdr;
+       __u64 is_accepted:1;
+       __u64 reservedz:63;
+} __packed;
+
+/*
+ * Message reporting capabilities. This is sent from the guest to the
+ * host.
+ */
+
+struct dm_capabilities {
+       struct dm_header hdr;
+       union dm_caps caps;
+       __u64 min_page_cnt;
+       __u64 max_page_number;
+} __packed;
+
+/*
+ * Response to the capabilities message. This is sent from the host to the
+ * guest. This message notifies if the host has accepted the guest's
+ * capabilities. If the host has not accepted, the guest must shutdown
+ * the service.
+ *
+ * is_accepted: Indicates if the host has accepted guest's capabilities.
+ * reservedz: Must be 0.
+ */
+
+struct dm_capabilities_resp_msg {
+       struct dm_header hdr;
+       __u64 is_accepted:1;
+       __u64 reservedz:63;
+} __packed;
+
+/*
+ * This message is used to report memory pressure from the guest.
+ * This message is not part of any transaction and there is no
+ * response to this message.
+ *
+ * num_avail: Available memory in pages.
+ * num_committed: Committed memory in pages.
+ * page_file_size: The accumulated size of all page files
+ *                in the system in pages.
+ * zero_free: The nunber of zero and free pages.
+ * page_file_writes: The writes to the page file in pages.
+ * io_diff: An indicator of file cache efficiency or page file activity,
+ *         calculated as File Cache Page Fault Count - Page Read Count.
+ *         This value is in pages.
+ *
+ * Some of these metrics are Windows specific and fortunately
+ * the algorithm on the host side that computes the guest memory
+ * pressure only uses num_committed value.
+ */
+
+struct dm_status {
+       struct dm_header hdr;
+       __u64 num_avail;
+       __u64 num_committed;
+       __u64 page_file_size;
+       __u64 zero_free;
+       __u32 page_file_writes;
+       __u32 io_diff;
+} __packed;
+
+
+/*
+ * Message to ask the guest to allocate memory - balloon up message.
+ * This message is sent from the host to the guest. The guest may not be
+ * able to allocate as much memory as requested.
+ *
+ * num_pages: number of pages to allocate.
+ */
+
+struct dm_balloon {
+       struct dm_header hdr;
+       __u32 num_pages;
+       __u32 reservedz;
+} __packed;
+
+
+/*
+ * Balloon response message; this message is sent from the guest
+ * to the host in response to the balloon message.
+ *
+ * reservedz: Reserved; must be set to zero.
+ * more_pages: If FALSE, this is the last message of the transaction.
+ * if TRUE there will atleast one more message from the guest.
+ *
+ * range_count: The number of ranges in the range array.
+ *
+ * range_array: An array of page ranges returned to the host.
+ *
+ */
+
+struct dm_balloon_response {
+       struct dm_header hdr;
+       __u32 reservedz;
+       __u32 more_pages:1;
+       __u32 range_count:31;
+       union dm_mem_page_range range_array[];
+} __packed;
+
+/*
+ * Un-balloon message; this message is sent from the host
+ * to the guest to give guest more memory.
+ *
+ * more_pages: If FALSE, this is the last message of the transaction.
+ * if TRUE there will atleast one more message from the guest.
+ *
+ * reservedz: Reserved; must be set to zero.
+ *
+ * range_count: The number of ranges in the range array.
+ *
+ * range_array: An array of page ranges returned to the host.
+ *
+ */
+
+struct dm_unballoon_request {
+       struct dm_header hdr;
+       __u32 more_pages:1;
+       __u32 reservedz:31;
+       __u32 range_count;
+       union dm_mem_page_range range_array[];
+} __packed;
+
+/*
+ * Un-balloon response message; this message is sent from the guest
+ * to the host in response to an unballoon request.
+ *
+ */
+
+struct dm_unballoon_response {
+       struct dm_header hdr;
+} __packed;
+
+
+/*
+ * Hot add request message. Message sent from the host to the guest.
+ *
+ * mem_range: Memory range to hot add.
+ *
+ * On Linux we currently don't support this since we cannot hot add
+ * arbitrary granularity of memory.
+ */
+
+struct dm_hot_add {
+       struct dm_header hdr;
+       union dm_mem_page_range range;
+} __packed;
+
+/*
+ * Hot add response message.
+ * This message is sent by the guest to report the status of a hot add request.
+ * If page_count is less than the requested page count, then the host should
+ * assume all further hot add requests will fail, since this indicates that
+ * the guest has hit an upper physical memory barrier.
+ *
+ * Hot adds may also fail due to low resources; in this case, the guest must
+ * not complete this message until the hot add can succeed, and the host must
+ * not send a new hot add request until the response is sent.
+ * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
+ * times it fails the request.
+ *
+ *
+ * page_count: number of pages that were successfully hot added.
+ *
+ * result: result of the operation 1: success, 0: failure.
+ *
+ */
+
+struct dm_hot_add_response {
+       struct dm_header hdr;
+       __u32 page_count;
+       __u32 result;
+} __packed;
+
+/*
+ * Types of information sent from host to the guest.
+ */
+
+enum dm_info_type {
+       INFO_TYPE_MAX_PAGE_CNT = 0,
+       MAX_INFO_TYPE
+};
+
+
+/*
+ * Header for the information message.
+ */
+
+struct dm_info_header {
+       enum dm_info_type type;
+       __u32 data_size;
+} __packed;
+
+/*
+ * This message is sent from the host to the guest to pass
+ * some relevant information (win8 addition).
+ *
+ * reserved: no used.
+ * info_size: size of the information blob.
+ * info: information blob.
+ */
+
+struct dm_info_msg {
+       struct dm_info_header header;
+       __u32 reserved;
+       __u32 info_size;
+       __u8  info[];
+};
+
+/*
+ * End protocol definitions.
+ */
+
+static bool hot_add;
+static bool do_hot_add;
+
+module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
+MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
+
+static atomic_t trans_id = ATOMIC_INIT(0);
+
+static int dm_ring_size = (5 * PAGE_SIZE);
+
+/*
+ * Driver specific state.
+ */
+
+enum hv_dm_state {
+       DM_INITIALIZING = 0,
+       DM_INITIALIZED,
+       DM_BALLOON_UP,
+       DM_BALLOON_DOWN,
+       DM_HOT_ADD,
+       DM_INIT_ERROR
+};
+
+
+static __u8 recv_buffer[PAGE_SIZE];
+static __u8 *send_buffer;
+#define PAGES_IN_2M    512
+
+struct hv_dynmem_device {
+       struct hv_device *dev;
+       enum hv_dm_state state;
+       struct completion host_event;
+       struct completion config_event;
+
+       /*
+        * Number of pages we have currently ballooned out.
+        */
+       unsigned int num_pages_ballooned;
+
+       /*
+        * This thread handles both balloon/hot-add
+        * requests from the host as well as notifying
+        * the host with regards to memory pressure in
+        * the guest.
+        */
+       struct task_struct *thread;
+
+       /*
+        * We start with the highest version we can support
+        * and downgrade based on the host; we save here the
+        * next version to try.
+        */
+       __u32 next_version;
+};
+
+static struct hv_dynmem_device dm_device;
+
+static void hot_add_req(struct hv_dynmem_device *dm, struct dm_hot_add *msg)
+{
+
+       struct dm_hot_add_response resp;
+
+       if (do_hot_add) {
+
+               pr_info("Memory hot add not supported\n");
+
+               /*
+                * Currently we do not support hot add.
+                * Just fail the request.
+                */
+       }
+
+       memset(&resp, 0, sizeof(struct dm_hot_add_response));
+       resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
+       resp.hdr.size = sizeof(struct dm_hot_add_response);
+       resp.hdr.trans_id = atomic_inc_return(&trans_id);
+
+       resp.page_count = 0;
+       resp.result = 0;
+
+       dm->state = DM_INITIALIZED;
+       vmbus_sendpacket(dm->dev->channel, &resp,
+                       sizeof(struct dm_hot_add_response),
+                       (unsigned long)NULL,
+                       VM_PKT_DATA_INBAND, 0);
+
+}
+
+static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
+{
+       switch (msg->header.type) {
+       case INFO_TYPE_MAX_PAGE_CNT:
+               pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
+               pr_info("Data Size is %d\n", msg->header.data_size);
+               break;
+       default:
+               pr_info("Received Unknown type: %d\n", msg->header.type);
+       }
+}
+
+/*
+ * Post our status as it relates memory pressure to the
+ * host. Host expects the guests to post this status
+ * periodically at 1 second intervals.
+ *
+ * The metrics specified in this protocol are very Windows
+ * specific and so we cook up numbers here to convey our memory
+ * pressure.
+ */
+
+static void post_status(struct hv_dynmem_device *dm)
+{
+       struct dm_status status;
+
+
+       memset(&status, 0, sizeof(struct dm_status));
+       status.hdr.type = DM_STATUS_REPORT;
+       status.hdr.size = sizeof(struct dm_status);
+       status.hdr.trans_id = atomic_inc_return(&trans_id);
+
+
+       status.num_committed = vm_memory_committed();
+
+       vmbus_sendpacket(dm->dev->channel, &status,
+                               sizeof(struct dm_status),
+                               (unsigned long)NULL,
+                               VM_PKT_DATA_INBAND, 0);
+
+}
+
+
+
+void free_balloon_pages(struct hv_dynmem_device *dm,
+                        union dm_mem_page_range *range_array)
+{
+       int num_pages = range_array->finfo.page_cnt;
+       __u64 start_frame = range_array->finfo.start_page;
+       struct page *pg;
+       int i;
+
+       for (i = 0; i < num_pages; i++) {
+               pg = pfn_to_page(i + start_frame);
+               __free_page(pg);
+               dm->num_pages_ballooned--;
+       }
+}
+
+
+
+static int  alloc_balloon_pages(struct hv_dynmem_device *dm, int num_pages,
+                        struct dm_balloon_response *bl_resp, int alloc_unit,
+                        bool *alloc_error)
+{
+       int i = 0;
+       struct page *pg;
+
+       if (num_pages < alloc_unit)
+               return 0;
+
+       for (i = 0; (i * alloc_unit) < num_pages; i++) {
+               if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
+                       PAGE_SIZE)
+                       return i * alloc_unit;
+
+               /*
+                * We execute this code in a thread context. Furthermore,
+                * we don't want the kernel to try too hard.
+                */
+               pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
+                               __GFP_NOMEMALLOC | __GFP_NOWARN,
+                               get_order(alloc_unit << PAGE_SHIFT));
+
+               if (!pg) {
+                       *alloc_error = true;
+                       return i * alloc_unit;
+               }
+
+
+               dm->num_pages_ballooned += alloc_unit;
+
+               bl_resp->range_count++;
+               bl_resp->range_array[i].finfo.start_page =
+                       page_to_pfn(pg);
+               bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
+               bl_resp->hdr.size += sizeof(union dm_mem_page_range);
+
+       }
+
+       return num_pages;
+}
+
+
+
+static void balloon_up(struct hv_dynmem_device *dm, struct dm_balloon *req)
+{
+       int num_pages = req->num_pages;
+       int num_ballooned = 0;
+       struct dm_balloon_response *bl_resp;
+       int alloc_unit;
+       int ret;
+       bool alloc_error = false;
+       bool done = false;
+       int i;
+
+
+       /*
+        * Currently, we only support 4k allocations.
+        */
+       alloc_unit = 1;
+
+       while (!done) {
+               bl_resp = (struct dm_balloon_response *)send_buffer;
+               memset(send_buffer, 0, PAGE_SIZE);
+               bl_resp->hdr.type = DM_BALLOON_RESPONSE;
+               bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
+               bl_resp->hdr.size = sizeof(struct dm_balloon_response);
+               bl_resp->more_pages = 1;
+
+
+               num_pages -= num_ballooned;
+               num_ballooned = alloc_balloon_pages(dm, num_pages,
+                                               bl_resp, alloc_unit,
+                                                &alloc_error);
+
+               if ((alloc_error) || (num_ballooned == num_pages)) {
+                       bl_resp->more_pages = 0;
+                       done = true;
+                       dm->state = DM_INITIALIZED;
+               }
+
+               /*
+                * We are pushing a lot of data through the channel;
+                * deal with transient failures caused because of the
+                * lack of space in the ring buffer.
+                */
+
+               do {
+                       ret = vmbus_sendpacket(dm_device.dev->channel,
+                                               bl_resp,
+                                               bl_resp->hdr.size,
+                                               (unsigned long)NULL,
+                                               VM_PKT_DATA_INBAND, 0);
+
+                       if (ret == -EAGAIN)
+                               msleep(20);
+
+               } while (ret == -EAGAIN);
+
+               if (ret) {
+                       /*
+                        * Free up the memory we allocatted.
+                        */
+                       pr_info("Balloon response failed\n");
+
+                       for (i = 0; i < bl_resp->range_count; i++)
+                               free_balloon_pages(dm,
+                                                &bl_resp->range_array[i]);
+
+                       done = true;
+               }
+       }
+
+}
+
+static void balloon_down(struct hv_dynmem_device *dm,
+                       struct dm_unballoon_request *req)
+{
+       union dm_mem_page_range *range_array = req->range_array;
+       int range_count = req->range_count;
+       struct dm_unballoon_response resp;
+       int i;
+
+       for (i = 0; i < range_count; i++)
+               free_balloon_pages(dm, &range_array[i]);
+
+       if (req->more_pages == 1)
+               return;
+
+       memset(&resp, 0, sizeof(struct dm_unballoon_response));
+       resp.hdr.type = DM_UNBALLOON_RESPONSE;
+       resp.hdr.trans_id = atomic_inc_return(&trans_id);
+       resp.hdr.size = sizeof(struct dm_unballoon_response);
+
+       vmbus_sendpacket(dm_device.dev->channel, &resp,
+                               sizeof(struct dm_unballoon_response),
+                               (unsigned long)NULL,
+                               VM_PKT_DATA_INBAND, 0);
+
+       dm->state = DM_INITIALIZED;
+}
+
+static void balloon_onchannelcallback(void *context);
+
+static int dm_thread_func(void *dm_dev)
+{
+       struct hv_dynmem_device *dm = dm_dev;
+       int t;
+       unsigned long  scan_start;
+
+       while (!kthread_should_stop()) {
+               t = wait_for_completion_timeout(&dm_device.config_event, 1*HZ);
+               /*
+                * The host expects us to post information on the memory
+                * pressure every second.
+                */
+
+               if (t == 0)
+                       post_status(dm);
+
+               scan_start = jiffies;
+               switch (dm->state) {
+               case DM_BALLOON_UP:
+                       balloon_up(dm, (struct dm_balloon *)recv_buffer);
+                       break;
+
+               case DM_HOT_ADD:
+                       hot_add_req(dm, (struct dm_hot_add *)recv_buffer);
+                       break;
+               default:
+                       break;
+               }
+
+               if (!time_in_range(jiffies, scan_start, scan_start + HZ))
+                       post_status(dm);
+
+       }
+
+       return 0;
+}
+
+
+static void version_resp(struct hv_dynmem_device *dm,
+                       struct dm_version_response *vresp)
+{
+       struct dm_version_request version_req;
+       int ret;
+
+       if (vresp->is_accepted) {
+               /*
+                * We are done; wakeup the
+                * context waiting for version
+                * negotiation.
+                */
+               complete(&dm->host_event);
+               return;
+       }
+       /*
+        * If there are more versions to try, continue
+        * with negotiations; if not
+        * shutdown the service since we are not able
+        * to negotiate a suitable version number
+        * with the host.
+        */
+       if (dm->next_version == 0)
+               goto version_error;
+
+       dm->next_version = 0;
+       memset(&version_req, 0, sizeof(struct dm_version_request));
+       version_req.hdr.type = DM_VERSION_REQUEST;
+       version_req.hdr.size = sizeof(struct dm_version_request);
+       version_req.hdr.trans_id = atomic_inc_return(&trans_id);
+       version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN7;
+       version_req.is_last_attempt = 1;
+
+       ret = vmbus_sendpacket(dm->dev->channel, &version_req,
+                               sizeof(struct dm_version_request),
+                               (unsigned long)NULL,
+                               VM_PKT_DATA_INBAND, 0);
+
+       if (ret)
+               goto version_error;
+
+       return;
+
+version_error:
+       dm->state = DM_INIT_ERROR;
+       complete(&dm->host_event);
+}
+
+static void cap_resp(struct hv_dynmem_device *dm,
+                       struct dm_capabilities_resp_msg *cap_resp)
+{
+       if (!cap_resp->is_accepted) {
+               pr_info("Capabilities not accepted by host\n");
+               dm->state = DM_INIT_ERROR;
+       }
+       complete(&dm->host_event);
+}
+
+static void balloon_onchannelcallback(void *context)
+{
+       struct hv_device *dev = context;
+       u32 recvlen;
+       u64 requestid;
+       struct dm_message *dm_msg;
+       struct dm_header *dm_hdr;
+       struct hv_dynmem_device *dm = hv_get_drvdata(dev);
+
+       memset(recv_buffer, 0, sizeof(recv_buffer));
+       vmbus_recvpacket(dev->channel, recv_buffer,
+                        PAGE_SIZE, &recvlen, &requestid);
+
+       if (recvlen > 0) {
+               dm_msg = (struct dm_message *)recv_buffer;
+               dm_hdr = &dm_msg->hdr;
+
+               switch (dm_hdr->type) {
+               case DM_VERSION_RESPONSE:
+                       version_resp(dm,
+                                (struct dm_version_response *)dm_msg);
+                       break;
+
+               case DM_CAPABILITIES_RESPONSE:
+                       cap_resp(dm,
+                                (struct dm_capabilities_resp_msg *)dm_msg);
+                       break;
+
+               case DM_BALLOON_REQUEST:
+                       dm->state = DM_BALLOON_UP;
+                       complete(&dm->config_event);
+                       break;
+
+               case DM_UNBALLOON_REQUEST:
+                       dm->state = DM_BALLOON_DOWN;
+                       balloon_down(dm,
+                                (struct dm_unballoon_request *)recv_buffer);
+                       break;
+
+               case DM_MEM_HOT_ADD_REQUEST:
+                       dm->state = DM_HOT_ADD;
+                       complete(&dm->config_event);
+                       break;
+
+               case DM_INFO_MESSAGE:
+                       process_info(dm, (struct dm_info_msg *)dm_msg);
+                       break;
+
+               default:
+                       pr_err("Unhandled message: type: %d\n", dm_hdr->type);
+
+               }
+       }
+
+}
+
+static int balloon_probe(struct hv_device *dev,
+                       const struct hv_vmbus_device_id *dev_id)
+{
+       int ret, t;
+       struct dm_version_request version_req;
+       struct dm_capabilities cap_msg;
+
+       do_hot_add = hot_add;
+
+       /*
+        * First allocate a send buffer.
+        */
+
+       send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
+       if (!send_buffer)
+               return -ENOMEM;
+
+       ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
+                       balloon_onchannelcallback, dev);
+
+       if (ret)
+               return ret;
+
+       dm_device.dev = dev;
+       dm_device.state = DM_INITIALIZING;
+       dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
+       init_completion(&dm_device.host_event);
+       init_completion(&dm_device.config_event);
+
+       dm_device.thread =
+                kthread_run(dm_thread_func, &dm_device, "hv_balloon");
+       if (IS_ERR(dm_device.thread)) {
+               ret = PTR_ERR(dm_device.thread);
+               goto probe_error0;
+       }
+
+       hv_set_drvdata(dev, &dm_device);
+       /*
+        * Initiate the hand shake with the host and negotiate
+        * a version that the host can support. We start with the
+        * highest version number and go down if the host cannot
+        * support it.
+        */
+       memset(&version_req, 0, sizeof(struct dm_version_request));
+       version_req.hdr.type = DM_VERSION_REQUEST;
+       version_req.hdr.size = sizeof(struct dm_version_request);
+       version_req.hdr.trans_id = atomic_inc_return(&trans_id);
+       version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN8;
+       version_req.is_last_attempt = 0;
+
+       ret = vmbus_sendpacket(dev->channel, &version_req,
+                               sizeof(struct dm_version_request),
+                               (unsigned long)NULL,
+                               VM_PKT_DATA_INBAND,
+                               VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
+       if (ret)
+               goto probe_error1;
+
+       t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
+       if (t == 0) {
+               ret = -ETIMEDOUT;
+               goto probe_error1;
+       }
+
+       /*
+        * If we could not negotiate a compatible version with the host
+        * fail the probe function.
+        */
+       if (dm_device.state == DM_INIT_ERROR) {
+               ret = -ETIMEDOUT;
+               goto probe_error1;
+       }
+       /*
+        * Now submit our capabilities to the host.
+        */
+       memset(&cap_msg, 0, sizeof(struct dm_capabilities));
+       cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
+       cap_msg.hdr.size = sizeof(struct dm_capabilities);
+       cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
+
+       cap_msg.caps.cap_bits.balloon = 1;
+       /*
+        * While we currently don't support hot-add,
+        * we still advertise this capability since the
+        * host requires that guests partcipating in the
+        * dynamic memory protocol support hot add.
+        */
+       cap_msg.caps.cap_bits.hot_add = 1;
+
+       /*
+        * Currently the host does not use these
+        * values and we set them to what is done in the
+        * Windows driver.
+        */
+       cap_msg.min_page_cnt = 0;
+       cap_msg.max_page_number = -1;
+
+       ret = vmbus_sendpacket(dev->channel, &cap_msg,
+                               sizeof(struct dm_capabilities),
+                               (unsigned long)NULL,
+                               VM_PKT_DATA_INBAND,
+                               VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
+       if (ret)
+               goto probe_error1;
+
+       t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
+       if (t == 0) {
+               ret = -ETIMEDOUT;
+               goto probe_error1;
+       }
+
+       /*
+        * If the host does not like our capabilities,
+        * fail the probe function.
+        */
+       if (dm_device.state == DM_INIT_ERROR) {
+               ret = -ETIMEDOUT;
+               goto probe_error1;
+       }
+
+       dm_device.state = DM_INITIALIZED;
+
+       return 0;
+
+probe_error1:
+       kthread_stop(dm_device.thread);
+
+probe_error0:
+       vmbus_close(dev->channel);
+       return ret;
+}
+
+static int balloon_remove(struct hv_device *dev)
+{
+       struct hv_dynmem_device *dm = hv_get_drvdata(dev);
+
+       if (dm->num_pages_ballooned != 0)
+               pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
+
+       vmbus_close(dev->channel);
+       kthread_stop(dm->thread);
+
+       return 0;
+}
+
+static const struct hv_vmbus_device_id id_table[] = {
+       /* Dynamic Memory Class ID */
+       /* 525074DC-8985-46e2-8057-A307DC18A502 */
+       { VMBUS_DEVICE(0xdc, 0x74, 0x50, 0X52, 0x85, 0x89, 0xe2, 0x46,
+                      0x80, 0x57, 0xa3, 0x07, 0xdc, 0x18, 0xa5, 0x02)
+       },
+       { },
+};
+
+MODULE_DEVICE_TABLE(vmbus, id_table);
+
+static  struct hv_driver balloon_drv = {
+       .name = "hv_balloon",
+       .id_table = id_table,
+       .probe =  balloon_probe,
+       .remove =  balloon_remove,
+};
+
+static int __init init_balloon_drv(void)
+{
+
+       return vmbus_driver_register(&balloon_drv);
+}
+
+static void exit_balloon_drv(void)
+{
+
+       vmbus_driver_unregister(&balloon_drv);
+}
+
+module_init(init_balloon_drv);
+module_exit(exit_balloon_drv);
+
+MODULE_DESCRIPTION("Hyper-V Balloon");
+MODULE_VERSION(HV_DRV_VERSION);
+MODULE_LICENSE("GPL");