2 * Copyright (c) 2012, Microsoft Corporation.
5 * K. Y. Srinivasan <kys@microsoft.com>
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 as published
9 * by the Free Software Foundation.
11 * This program is distributed in the hope that it will be useful, but
12 * WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
14 * NON INFRINGEMENT. See the GNU General Public License for more
19 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
21 #include <linux/kernel.h>
22 #include <linux/mman.h>
23 #include <linux/delay.h>
24 #include <linux/init.h>
25 #include <linux/module.h>
26 #include <linux/slab.h>
27 #include <linux/kthread.h>
28 #include <linux/completion.h>
29 #include <linux/memory_hotplug.h>
30 #include <linux/memory.h>
31 #include <linux/notifier.h>
32 #include <linux/percpu_counter.h>
34 #include <linux/hyperv.h>
37 * We begin with definitions supporting the Dynamic Memory protocol
40 * Begin protocol definitions.
46 * Protocol versions. The low word is the minor version, the high word the major
51 * Changed to 0.1 on 2009/03/25
52 * Changes to 0.2 on 2009/05/14
53 * Changes to 0.3 on 2009/12/03
54 * Changed to 1.0 on 2011/04/05
57 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
58 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
59 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
62 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
63 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
65 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
66 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
68 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN8
77 enum dm_message_type {
82 DM_VERSION_REQUEST = 1,
83 DM_VERSION_RESPONSE = 2,
84 DM_CAPABILITIES_REPORT = 3,
85 DM_CAPABILITIES_RESPONSE = 4,
87 DM_BALLOON_REQUEST = 6,
88 DM_BALLOON_RESPONSE = 7,
89 DM_UNBALLOON_REQUEST = 8,
90 DM_UNBALLOON_RESPONSE = 9,
91 DM_MEM_HOT_ADD_REQUEST = 10,
92 DM_MEM_HOT_ADD_RESPONSE = 11,
93 DM_VERSION_03_MAX = 11,
103 * Structures defining the dynamic memory management
125 union dm_mem_page_range {
128 * The PFN number of the first page in the range.
129 * 40 bits is the architectural limit of a PFN
134 * The number of pages in the range.
144 * The header for all dynamic memory messages:
146 * type: Type of the message.
147 * size: Size of the message in bytes; including the header.
148 * trans_id: The guest is responsible for manufacturing this ID.
158 * A generic message format for dynamic memory.
159 * Specific message formats are defined later in the file.
163 struct dm_header hdr;
164 __u8 data[]; /* enclosed message */
169 * Specific message types supporting the dynamic memory protocol.
173 * Version negotiation message. Sent from the guest to the host.
174 * The guest is free to try different versions until the host
175 * accepts the version.
177 * dm_version: The protocol version requested.
178 * is_last_attempt: If TRUE, this is the last version guest will request.
179 * reservedz: Reserved field, set to zero.
182 struct dm_version_request {
183 struct dm_header hdr;
184 union dm_version version;
185 __u32 is_last_attempt:1;
190 * Version response message; Host to Guest and indicates
191 * if the host has accepted the version sent by the guest.
193 * is_accepted: If TRUE, host has accepted the version and the guest
194 * should proceed to the next stage of the protocol. FALSE indicates that
195 * guest should re-try with a different version.
197 * reservedz: Reserved field, set to zero.
200 struct dm_version_response {
201 struct dm_header hdr;
207 * Message reporting capabilities. This is sent from the guest to the
211 struct dm_capabilities {
212 struct dm_header hdr;
215 __u64 max_page_number;
219 * Response to the capabilities message. This is sent from the host to the
220 * guest. This message notifies if the host has accepted the guest's
221 * capabilities. If the host has not accepted, the guest must shutdown
224 * is_accepted: Indicates if the host has accepted guest's capabilities.
225 * reservedz: Must be 0.
228 struct dm_capabilities_resp_msg {
229 struct dm_header hdr;
235 * This message is used to report memory pressure from the guest.
236 * This message is not part of any transaction and there is no
237 * response to this message.
239 * num_avail: Available memory in pages.
240 * num_committed: Committed memory in pages.
241 * page_file_size: The accumulated size of all page files
242 * in the system in pages.
243 * zero_free: The nunber of zero and free pages.
244 * page_file_writes: The writes to the page file in pages.
245 * io_diff: An indicator of file cache efficiency or page file activity,
246 * calculated as File Cache Page Fault Count - Page Read Count.
247 * This value is in pages.
249 * Some of these metrics are Windows specific and fortunately
250 * the algorithm on the host side that computes the guest memory
251 * pressure only uses num_committed value.
255 struct dm_header hdr;
258 __u64 page_file_size;
260 __u32 page_file_writes;
266 * Message to ask the guest to allocate memory - balloon up message.
267 * This message is sent from the host to the guest. The guest may not be
268 * able to allocate as much memory as requested.
270 * num_pages: number of pages to allocate.
274 struct dm_header hdr;
281 * Balloon response message; this message is sent from the guest
282 * to the host in response to the balloon message.
284 * reservedz: Reserved; must be set to zero.
285 * more_pages: If FALSE, this is the last message of the transaction.
286 * if TRUE there will atleast one more message from the guest.
288 * range_count: The number of ranges in the range array.
290 * range_array: An array of page ranges returned to the host.
294 struct dm_balloon_response {
295 struct dm_header hdr;
298 __u32 range_count:31;
299 union dm_mem_page_range range_array[];
303 * Un-balloon message; this message is sent from the host
304 * to the guest to give guest more memory.
306 * more_pages: If FALSE, this is the last message of the transaction.
307 * if TRUE there will atleast one more message from the guest.
309 * reservedz: Reserved; must be set to zero.
311 * range_count: The number of ranges in the range array.
313 * range_array: An array of page ranges returned to the host.
317 struct dm_unballoon_request {
318 struct dm_header hdr;
322 union dm_mem_page_range range_array[];
326 * Un-balloon response message; this message is sent from the guest
327 * to the host in response to an unballoon request.
331 struct dm_unballoon_response {
332 struct dm_header hdr;
337 * Hot add request message. Message sent from the host to the guest.
339 * mem_range: Memory range to hot add.
341 * On Linux we currently don't support this since we cannot hot add
342 * arbitrary granularity of memory.
346 struct dm_header hdr;
347 union dm_mem_page_range range;
351 * Hot add response message.
352 * This message is sent by the guest to report the status of a hot add request.
353 * If page_count is less than the requested page count, then the host should
354 * assume all further hot add requests will fail, since this indicates that
355 * the guest has hit an upper physical memory barrier.
357 * Hot adds may also fail due to low resources; in this case, the guest must
358 * not complete this message until the hot add can succeed, and the host must
359 * not send a new hot add request until the response is sent.
360 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
361 * times it fails the request.
364 * page_count: number of pages that were successfully hot added.
366 * result: result of the operation 1: success, 0: failure.
370 struct dm_hot_add_response {
371 struct dm_header hdr;
377 * Types of information sent from host to the guest.
381 INFO_TYPE_MAX_PAGE_CNT = 0,
387 * Header for the information message.
390 struct dm_info_header {
391 enum dm_info_type type;
396 * This message is sent from the host to the guest to pass
397 * some relevant information (win8 addition).
400 * info_size: size of the information blob.
401 * info: information blob.
405 struct dm_header hdr;
412 * End protocol definitions.
416 static bool do_hot_add;
418 * Delay reporting memory pressure by
419 * the specified number of seconds.
421 static uint pressure_report_delay = 30;
423 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
424 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
426 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
427 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
428 static atomic_t trans_id = ATOMIC_INIT(0);
430 static int dm_ring_size = (5 * PAGE_SIZE);
433 * Driver specific state.
446 static __u8 recv_buffer[PAGE_SIZE];
447 static __u8 *send_buffer;
448 #define PAGES_IN_2M 512
450 struct hv_dynmem_device {
451 struct hv_device *dev;
452 enum hv_dm_state state;
453 struct completion host_event;
454 struct completion config_event;
457 * Number of pages we have currently ballooned out.
459 unsigned int num_pages_ballooned;
462 * This thread handles both balloon/hot-add
463 * requests from the host as well as notifying
464 * the host with regards to memory pressure in
467 struct task_struct *thread;
470 * We start with the highest version we can support
471 * and downgrade based on the host; we save here the
472 * next version to try.
477 static struct hv_dynmem_device dm_device;
479 static void hot_add_req(struct hv_dynmem_device *dm, struct dm_hot_add *msg)
482 struct dm_hot_add_response resp;
486 pr_info("Memory hot add not supported\n");
489 * Currently we do not support hot add.
490 * Just fail the request.
494 memset(&resp, 0, sizeof(struct dm_hot_add_response));
495 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
496 resp.hdr.size = sizeof(struct dm_hot_add_response);
497 resp.hdr.trans_id = atomic_inc_return(&trans_id);
502 dm->state = DM_INITIALIZED;
503 vmbus_sendpacket(dm->dev->channel, &resp,
504 sizeof(struct dm_hot_add_response),
506 VM_PKT_DATA_INBAND, 0);
510 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
512 struct dm_info_header *info_hdr;
514 info_hdr = (struct dm_info_header *)msg->info;
516 switch (info_hdr->type) {
517 case INFO_TYPE_MAX_PAGE_CNT:
518 pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
519 pr_info("Data Size is %d\n", info_hdr->data_size);
522 pr_info("Received Unknown type: %d\n", info_hdr->type);
526 unsigned long compute_balloon_floor(void)
528 unsigned long min_pages;
529 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
530 /* Simple continuous piecewiese linear function:
531 * max MiB -> min MiB gradient
542 if (totalram_pages < MB2PAGES(128))
543 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
544 else if (totalram_pages < MB2PAGES(512))
545 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
546 else if (totalram_pages < MB2PAGES(2048))
547 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
549 min_pages = MB2PAGES(296) + (totalram_pages >> 5);
555 * Post our status as it relates memory pressure to the
556 * host. Host expects the guests to post this status
557 * periodically at 1 second intervals.
559 * The metrics specified in this protocol are very Windows
560 * specific and so we cook up numbers here to convey our memory
564 static void post_status(struct hv_dynmem_device *dm)
566 struct dm_status status;
569 if (pressure_report_delay > 0) {
570 --pressure_report_delay;
574 memset(&status, 0, sizeof(struct dm_status));
575 status.hdr.type = DM_STATUS_REPORT;
576 status.hdr.size = sizeof(struct dm_status);
577 status.hdr.trans_id = atomic_inc_return(&trans_id);
580 * The host expects the guest to report free memory.
581 * Further, the host expects the pressure information to
582 * include the ballooned out pages.
583 * For a given amount of memory that we are managing, we
584 * need to compute a floor below which we should not balloon.
585 * Compute this and add it to the pressure report.
587 status.num_avail = val.freeram;
588 status.num_committed = vm_memory_committed() +
589 dm->num_pages_ballooned +
590 compute_balloon_floor();
592 vmbus_sendpacket(dm->dev->channel, &status,
593 sizeof(struct dm_status),
595 VM_PKT_DATA_INBAND, 0);
599 static void free_balloon_pages(struct hv_dynmem_device *dm,
600 union dm_mem_page_range *range_array)
602 int num_pages = range_array->finfo.page_cnt;
603 __u64 start_frame = range_array->finfo.start_page;
607 for (i = 0; i < num_pages; i++) {
608 pg = pfn_to_page(i + start_frame);
610 dm->num_pages_ballooned--;
616 static int alloc_balloon_pages(struct hv_dynmem_device *dm, int num_pages,
617 struct dm_balloon_response *bl_resp, int alloc_unit,
623 if (num_pages < alloc_unit)
626 for (i = 0; (i * alloc_unit) < num_pages; i++) {
627 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
629 return i * alloc_unit;
632 * We execute this code in a thread context. Furthermore,
633 * we don't want the kernel to try too hard.
635 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
636 __GFP_NOMEMALLOC | __GFP_NOWARN,
637 get_order(alloc_unit << PAGE_SHIFT));
641 return i * alloc_unit;
645 dm->num_pages_ballooned += alloc_unit;
647 bl_resp->range_count++;
648 bl_resp->range_array[i].finfo.start_page =
650 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
651 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
660 static void balloon_up(struct hv_dynmem_device *dm, struct dm_balloon *req)
662 int num_pages = req->num_pages;
663 int num_ballooned = 0;
664 struct dm_balloon_response *bl_resp;
667 bool alloc_error = false;
673 * Currently, we only support 4k allocations.
678 bl_resp = (struct dm_balloon_response *)send_buffer;
679 memset(send_buffer, 0, PAGE_SIZE);
680 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
681 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
682 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
683 bl_resp->more_pages = 1;
686 num_pages -= num_ballooned;
687 num_ballooned = alloc_balloon_pages(dm, num_pages,
691 if ((alloc_error) || (num_ballooned == num_pages)) {
692 bl_resp->more_pages = 0;
694 dm->state = DM_INITIALIZED;
698 * We are pushing a lot of data through the channel;
699 * deal with transient failures caused because of the
700 * lack of space in the ring buffer.
704 ret = vmbus_sendpacket(dm_device.dev->channel,
708 VM_PKT_DATA_INBAND, 0);
713 } while (ret == -EAGAIN);
717 * Free up the memory we allocatted.
719 pr_info("Balloon response failed\n");
721 for (i = 0; i < bl_resp->range_count; i++)
722 free_balloon_pages(dm,
723 &bl_resp->range_array[i]);
731 static void balloon_down(struct hv_dynmem_device *dm,
732 struct dm_unballoon_request *req)
734 union dm_mem_page_range *range_array = req->range_array;
735 int range_count = req->range_count;
736 struct dm_unballoon_response resp;
739 for (i = 0; i < range_count; i++)
740 free_balloon_pages(dm, &range_array[i]);
742 if (req->more_pages == 1)
745 memset(&resp, 0, sizeof(struct dm_unballoon_response));
746 resp.hdr.type = DM_UNBALLOON_RESPONSE;
747 resp.hdr.trans_id = atomic_inc_return(&trans_id);
748 resp.hdr.size = sizeof(struct dm_unballoon_response);
750 vmbus_sendpacket(dm_device.dev->channel, &resp,
751 sizeof(struct dm_unballoon_response),
753 VM_PKT_DATA_INBAND, 0);
755 dm->state = DM_INITIALIZED;
758 static void balloon_onchannelcallback(void *context);
760 static int dm_thread_func(void *dm_dev)
762 struct hv_dynmem_device *dm = dm_dev;
764 unsigned long scan_start;
766 while (!kthread_should_stop()) {
767 t = wait_for_completion_timeout(&dm_device.config_event, 1*HZ);
769 * The host expects us to post information on the memory
770 * pressure every second.
776 scan_start = jiffies;
779 balloon_up(dm, (struct dm_balloon *)recv_buffer);
783 hot_add_req(dm, (struct dm_hot_add *)recv_buffer);
789 if (!time_in_range(jiffies, scan_start, scan_start + HZ))
798 static void version_resp(struct hv_dynmem_device *dm,
799 struct dm_version_response *vresp)
801 struct dm_version_request version_req;
804 if (vresp->is_accepted) {
806 * We are done; wakeup the
807 * context waiting for version
810 complete(&dm->host_event);
814 * If there are more versions to try, continue
815 * with negotiations; if not
816 * shutdown the service since we are not able
817 * to negotiate a suitable version number
820 if (dm->next_version == 0)
823 dm->next_version = 0;
824 memset(&version_req, 0, sizeof(struct dm_version_request));
825 version_req.hdr.type = DM_VERSION_REQUEST;
826 version_req.hdr.size = sizeof(struct dm_version_request);
827 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
828 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN7;
829 version_req.is_last_attempt = 1;
831 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
832 sizeof(struct dm_version_request),
834 VM_PKT_DATA_INBAND, 0);
842 dm->state = DM_INIT_ERROR;
843 complete(&dm->host_event);
846 static void cap_resp(struct hv_dynmem_device *dm,
847 struct dm_capabilities_resp_msg *cap_resp)
849 if (!cap_resp->is_accepted) {
850 pr_info("Capabilities not accepted by host\n");
851 dm->state = DM_INIT_ERROR;
853 complete(&dm->host_event);
856 static void balloon_onchannelcallback(void *context)
858 struct hv_device *dev = context;
861 struct dm_message *dm_msg;
862 struct dm_header *dm_hdr;
863 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
865 memset(recv_buffer, 0, sizeof(recv_buffer));
866 vmbus_recvpacket(dev->channel, recv_buffer,
867 PAGE_SIZE, &recvlen, &requestid);
870 dm_msg = (struct dm_message *)recv_buffer;
871 dm_hdr = &dm_msg->hdr;
873 switch (dm_hdr->type) {
874 case DM_VERSION_RESPONSE:
876 (struct dm_version_response *)dm_msg);
879 case DM_CAPABILITIES_RESPONSE:
881 (struct dm_capabilities_resp_msg *)dm_msg);
884 case DM_BALLOON_REQUEST:
885 dm->state = DM_BALLOON_UP;
886 complete(&dm->config_event);
889 case DM_UNBALLOON_REQUEST:
890 dm->state = DM_BALLOON_DOWN;
892 (struct dm_unballoon_request *)recv_buffer);
895 case DM_MEM_HOT_ADD_REQUEST:
896 dm->state = DM_HOT_ADD;
897 complete(&dm->config_event);
900 case DM_INFO_MESSAGE:
901 process_info(dm, (struct dm_info_msg *)dm_msg);
905 pr_err("Unhandled message: type: %d\n", dm_hdr->type);
912 static int balloon_probe(struct hv_device *dev,
913 const struct hv_vmbus_device_id *dev_id)
916 struct dm_version_request version_req;
917 struct dm_capabilities cap_msg;
919 do_hot_add = hot_add;
922 * First allocate a send buffer.
925 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
929 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
930 balloon_onchannelcallback, dev);
936 dm_device.state = DM_INITIALIZING;
937 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
938 init_completion(&dm_device.host_event);
939 init_completion(&dm_device.config_event);
942 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
943 if (IS_ERR(dm_device.thread)) {
944 ret = PTR_ERR(dm_device.thread);
948 hv_set_drvdata(dev, &dm_device);
950 * Initiate the hand shake with the host and negotiate
951 * a version that the host can support. We start with the
952 * highest version number and go down if the host cannot
955 memset(&version_req, 0, sizeof(struct dm_version_request));
956 version_req.hdr.type = DM_VERSION_REQUEST;
957 version_req.hdr.size = sizeof(struct dm_version_request);
958 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
959 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN8;
960 version_req.is_last_attempt = 0;
962 ret = vmbus_sendpacket(dev->channel, &version_req,
963 sizeof(struct dm_version_request),
966 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
970 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
977 * If we could not negotiate a compatible version with the host
978 * fail the probe function.
980 if (dm_device.state == DM_INIT_ERROR) {
985 * Now submit our capabilities to the host.
987 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
988 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
989 cap_msg.hdr.size = sizeof(struct dm_capabilities);
990 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
992 cap_msg.caps.cap_bits.balloon = 1;
994 * While we currently don't support hot-add,
995 * we still advertise this capability since the
996 * host requires that guests partcipating in the
997 * dynamic memory protocol support hot add.
999 cap_msg.caps.cap_bits.hot_add = 1;
1002 * Currently the host does not use these
1003 * values and we set them to what is done in the
1006 cap_msg.min_page_cnt = 0;
1007 cap_msg.max_page_number = -1;
1009 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1010 sizeof(struct dm_capabilities),
1011 (unsigned long)NULL,
1013 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1017 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1024 * If the host does not like our capabilities,
1025 * fail the probe function.
1027 if (dm_device.state == DM_INIT_ERROR) {
1032 dm_device.state = DM_INITIALIZED;
1037 kthread_stop(dm_device.thread);
1040 vmbus_close(dev->channel);
1046 static int balloon_remove(struct hv_device *dev)
1048 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1050 if (dm->num_pages_ballooned != 0)
1051 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1053 vmbus_close(dev->channel);
1054 kthread_stop(dm->thread);
1060 static const struct hv_vmbus_device_id id_table[] = {
1061 /* Dynamic Memory Class ID */
1062 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1067 MODULE_DEVICE_TABLE(vmbus, id_table);
1069 static struct hv_driver balloon_drv = {
1070 .name = "hv_balloon",
1071 .id_table = id_table,
1072 .probe = balloon_probe,
1073 .remove = balloon_remove,
1076 static int __init init_balloon_drv(void)
1079 return vmbus_driver_register(&balloon_drv);
1082 static void exit_balloon_drv(void)
1085 vmbus_driver_unregister(&balloon_drv);
1088 module_init(init_balloon_drv);
1089 module_exit(exit_balloon_drv);
1091 MODULE_DESCRIPTION("Hyper-V Balloon");
1092 MODULE_VERSION(HV_DRV_VERSION);
1093 MODULE_LICENSE("GPL");