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/jiffies.h>
23 #include <linux/mman.h>
24 #include <linux/delay.h>
25 #include <linux/init.h>
26 #include <linux/module.h>
27 #include <linux/slab.h>
28 #include <linux/kthread.h>
29 #include <linux/completion.h>
30 #include <linux/memory_hotplug.h>
31 #include <linux/memory.h>
32 #include <linux/notifier.h>
33 #include <linux/percpu_counter.h>
35 #include <linux/hyperv.h>
37 #define CREATE_TRACE_POINTS
38 #include "hv_trace_balloon.h"
41 * We begin with definitions supporting the Dynamic Memory protocol
44 * Begin protocol definitions.
50 * Protocol versions. The low word is the minor version, the high word the major
55 * Changed to 0.1 on 2009/03/25
56 * Changes to 0.2 on 2009/05/14
57 * Changes to 0.3 on 2009/12/03
58 * Changed to 1.0 on 2011/04/05
61 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
62 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
63 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
66 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
67 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
68 DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
70 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
71 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
72 DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
74 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
83 enum dm_message_type {
88 DM_VERSION_REQUEST = 1,
89 DM_VERSION_RESPONSE = 2,
90 DM_CAPABILITIES_REPORT = 3,
91 DM_CAPABILITIES_RESPONSE = 4,
93 DM_BALLOON_REQUEST = 6,
94 DM_BALLOON_RESPONSE = 7,
95 DM_UNBALLOON_REQUEST = 8,
96 DM_UNBALLOON_RESPONSE = 9,
97 DM_MEM_HOT_ADD_REQUEST = 10,
98 DM_MEM_HOT_ADD_RESPONSE = 11,
99 DM_VERSION_03_MAX = 11,
103 DM_INFO_MESSAGE = 12,
104 DM_VERSION_1_MAX = 12
109 * Structures defining the dynamic memory management
127 * To support guests that may have alignment
128 * limitations on hot-add, the guest can specify
129 * its alignment requirements; a value of n
130 * represents an alignment of 2^n in mega bytes.
132 __u64 hot_add_alignment:4;
138 union dm_mem_page_range {
141 * The PFN number of the first page in the range.
142 * 40 bits is the architectural limit of a PFN
147 * The number of pages in the range.
157 * The header for all dynamic memory messages:
159 * type: Type of the message.
160 * size: Size of the message in bytes; including the header.
161 * trans_id: The guest is responsible for manufacturing this ID.
171 * A generic message format for dynamic memory.
172 * Specific message formats are defined later in the file.
176 struct dm_header hdr;
177 __u8 data[]; /* enclosed message */
182 * Specific message types supporting the dynamic memory protocol.
186 * Version negotiation message. Sent from the guest to the host.
187 * The guest is free to try different versions until the host
188 * accepts the version.
190 * dm_version: The protocol version requested.
191 * is_last_attempt: If TRUE, this is the last version guest will request.
192 * reservedz: Reserved field, set to zero.
195 struct dm_version_request {
196 struct dm_header hdr;
197 union dm_version version;
198 __u32 is_last_attempt:1;
203 * Version response message; Host to Guest and indicates
204 * if the host has accepted the version sent by the guest.
206 * is_accepted: If TRUE, host has accepted the version and the guest
207 * should proceed to the next stage of the protocol. FALSE indicates that
208 * guest should re-try with a different version.
210 * reservedz: Reserved field, set to zero.
213 struct dm_version_response {
214 struct dm_header hdr;
220 * Message reporting capabilities. This is sent from the guest to the
224 struct dm_capabilities {
225 struct dm_header hdr;
228 __u64 max_page_number;
232 * Response to the capabilities message. This is sent from the host to the
233 * guest. This message notifies if the host has accepted the guest's
234 * capabilities. If the host has not accepted, the guest must shutdown
237 * is_accepted: Indicates if the host has accepted guest's capabilities.
238 * reservedz: Must be 0.
241 struct dm_capabilities_resp_msg {
242 struct dm_header hdr;
248 * This message is used to report memory pressure from the guest.
249 * This message is not part of any transaction and there is no
250 * response to this message.
252 * num_avail: Available memory in pages.
253 * num_committed: Committed memory in pages.
254 * page_file_size: The accumulated size of all page files
255 * in the system in pages.
256 * zero_free: The nunber of zero and free pages.
257 * page_file_writes: The writes to the page file in pages.
258 * io_diff: An indicator of file cache efficiency or page file activity,
259 * calculated as File Cache Page Fault Count - Page Read Count.
260 * This value is in pages.
262 * Some of these metrics are Windows specific and fortunately
263 * the algorithm on the host side that computes the guest memory
264 * pressure only uses num_committed value.
268 struct dm_header hdr;
271 __u64 page_file_size;
273 __u32 page_file_writes;
279 * Message to ask the guest to allocate memory - balloon up message.
280 * This message is sent from the host to the guest. The guest may not be
281 * able to allocate as much memory as requested.
283 * num_pages: number of pages to allocate.
287 struct dm_header hdr;
294 * Balloon response message; this message is sent from the guest
295 * to the host in response to the balloon message.
297 * reservedz: Reserved; must be set to zero.
298 * more_pages: If FALSE, this is the last message of the transaction.
299 * if TRUE there will atleast one more message from the guest.
301 * range_count: The number of ranges in the range array.
303 * range_array: An array of page ranges returned to the host.
307 struct dm_balloon_response {
308 struct dm_header hdr;
311 __u32 range_count:31;
312 union dm_mem_page_range range_array[];
316 * Un-balloon message; this message is sent from the host
317 * to the guest to give guest more memory.
319 * more_pages: If FALSE, this is the last message of the transaction.
320 * if TRUE there will atleast one more message from the guest.
322 * reservedz: Reserved; must be set to zero.
324 * range_count: The number of ranges in the range array.
326 * range_array: An array of page ranges returned to the host.
330 struct dm_unballoon_request {
331 struct dm_header hdr;
335 union dm_mem_page_range range_array[];
339 * Un-balloon response message; this message is sent from the guest
340 * to the host in response to an unballoon request.
344 struct dm_unballoon_response {
345 struct dm_header hdr;
350 * Hot add request message. Message sent from the host to the guest.
352 * mem_range: Memory range to hot add.
354 * On Linux we currently don't support this since we cannot hot add
355 * arbitrary granularity of memory.
359 struct dm_header hdr;
360 union dm_mem_page_range range;
364 * Hot add response message.
365 * This message is sent by the guest to report the status of a hot add request.
366 * If page_count is less than the requested page count, then the host should
367 * assume all further hot add requests will fail, since this indicates that
368 * the guest has hit an upper physical memory barrier.
370 * Hot adds may also fail due to low resources; in this case, the guest must
371 * not complete this message until the hot add can succeed, and the host must
372 * not send a new hot add request until the response is sent.
373 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
374 * times it fails the request.
377 * page_count: number of pages that were successfully hot added.
379 * result: result of the operation 1: success, 0: failure.
383 struct dm_hot_add_response {
384 struct dm_header hdr;
390 * Types of information sent from host to the guest.
394 INFO_TYPE_MAX_PAGE_CNT = 0,
400 * Header for the information message.
403 struct dm_info_header {
404 enum dm_info_type type;
409 * This message is sent from the host to the guest to pass
410 * some relevant information (win8 addition).
413 * info_size: size of the information blob.
414 * info: information blob.
418 struct dm_header hdr;
425 * End protocol definitions.
429 * State to manage hot adding memory into the guest.
430 * The range start_pfn : end_pfn specifies the range
431 * that the host has asked us to hot add. The range
432 * start_pfn : ha_end_pfn specifies the range that we have
433 * currently hot added. We hot add in multiples of 128M
434 * chunks; it is possible that we may not be able to bring
435 * online all the pages in the region. The range
436 * covered_start_pfn:covered_end_pfn defines the pages that can
440 struct hv_hotadd_state {
441 struct list_head list;
442 unsigned long start_pfn;
443 unsigned long covered_start_pfn;
444 unsigned long covered_end_pfn;
445 unsigned long ha_end_pfn;
446 unsigned long end_pfn;
450 struct list_head gap_list;
453 struct hv_hotadd_gap {
454 struct list_head list;
455 unsigned long start_pfn;
456 unsigned long end_pfn;
459 struct balloon_state {
461 struct work_struct wrk;
465 union dm_mem_page_range ha_page_range;
466 union dm_mem_page_range ha_region_range;
467 struct work_struct wrk;
470 static bool hot_add = true;
471 static bool do_hot_add;
473 * Delay reporting memory pressure by
474 * the specified number of seconds.
476 static uint pressure_report_delay = 45;
479 * The last time we posted a pressure report to host.
481 static unsigned long last_post_time;
483 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
484 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
486 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
487 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
488 static atomic_t trans_id = ATOMIC_INIT(0);
490 static int dm_ring_size = (5 * PAGE_SIZE);
493 * Driver specific state.
506 static __u8 recv_buffer[PAGE_SIZE];
507 static __u8 *send_buffer;
508 #define PAGES_IN_2M 512
509 #define HA_CHUNK (32 * 1024)
511 struct hv_dynmem_device {
512 struct hv_device *dev;
513 enum hv_dm_state state;
514 struct completion host_event;
515 struct completion config_event;
518 * Number of pages we have currently ballooned out.
520 unsigned int num_pages_ballooned;
521 unsigned int num_pages_onlined;
522 unsigned int num_pages_added;
525 * State to manage the ballooning (up) operation.
527 struct balloon_state balloon_wrk;
530 * State to execute the "hot-add" operation.
532 struct hot_add_wrk ha_wrk;
535 * This state tracks if the host has specified a hot-add
538 bool host_specified_ha_region;
541 * State to synchronize hot-add.
543 struct completion ol_waitevent;
546 * This thread handles hot-add
547 * requests from the host as well as notifying
548 * the host with regards to memory pressure in
551 struct task_struct *thread;
554 * Protects ha_region_list, num_pages_onlined counter and individual
555 * regions from ha_region_list.
560 * A list of hot-add regions.
562 struct list_head ha_region_list;
565 * We start with the highest version we can support
566 * and downgrade based on the host; we save here the
567 * next version to try.
572 * The negotiated version agreed by host.
577 static struct hv_dynmem_device dm_device;
579 static void post_status(struct hv_dynmem_device *dm);
581 #ifdef CONFIG_MEMORY_HOTPLUG
582 static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
585 struct hv_hotadd_gap *gap;
587 /* The page is not backed. */
588 if ((pfn < has->covered_start_pfn) || (pfn >= has->covered_end_pfn))
591 /* Check for gaps. */
592 list_for_each_entry(gap, &has->gap_list, list) {
593 if ((pfn >= gap->start_pfn) && (pfn < gap->end_pfn))
600 static unsigned long hv_page_offline_check(unsigned long start_pfn,
601 unsigned long nr_pages)
603 unsigned long pfn = start_pfn, count = 0;
604 struct hv_hotadd_state *has;
607 while (pfn < start_pfn + nr_pages) {
609 * Search for HAS which covers the pfn and when we find one
610 * count how many consequitive PFNs are covered.
613 list_for_each_entry(has, &dm_device.ha_region_list, list) {
614 while ((pfn >= has->start_pfn) &&
615 (pfn < has->end_pfn) &&
616 (pfn < start_pfn + nr_pages)) {
618 if (has_pfn_is_backed(has, pfn))
625 * This PFN is not in any HAS (e.g. we're offlining a region
626 * which was present at boot), no need to account for it. Go
636 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
639 struct memory_notify *mem = (struct memory_notify *)v;
640 unsigned long flags, pfn_count;
644 case MEM_CANCEL_ONLINE:
645 if (dm_device.ha_waiting) {
646 dm_device.ha_waiting = false;
647 complete(&dm_device.ol_waitevent);
652 spin_lock_irqsave(&dm_device.ha_lock, flags);
653 pfn_count = hv_page_offline_check(mem->start_pfn,
655 if (pfn_count <= dm_device.num_pages_onlined) {
656 dm_device.num_pages_onlined -= pfn_count;
659 * We're offlining more pages than we managed to online.
660 * This is unexpected. In any case don't let
661 * num_pages_onlined wrap around zero.
664 dm_device.num_pages_onlined = 0;
666 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
668 case MEM_GOING_ONLINE:
669 case MEM_GOING_OFFLINE:
670 case MEM_CANCEL_OFFLINE:
676 static struct notifier_block hv_memory_nb = {
677 .notifier_call = hv_memory_notifier,
681 /* Check if the particular page is backed and can be onlined and online it. */
682 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
684 if (!has_pfn_is_backed(has, page_to_pfn(pg)))
687 /* This frame is currently backed; online the page. */
688 __online_page_set_limits(pg);
689 __online_page_increment_counters(pg);
690 __online_page_free(pg);
692 lockdep_assert_held(&dm_device.ha_lock);
693 dm_device.num_pages_onlined++;
696 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
697 unsigned long start_pfn, unsigned long size)
701 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
702 for (i = 0; i < size; i++)
703 hv_page_online_one(has, pfn_to_page(start_pfn + i));
706 static void hv_mem_hot_add(unsigned long start, unsigned long size,
707 unsigned long pfn_count,
708 struct hv_hotadd_state *has)
712 unsigned long start_pfn;
713 unsigned long processed_pfn;
714 unsigned long total_pfn = pfn_count;
717 for (i = 0; i < (size/HA_CHUNK); i++) {
718 start_pfn = start + (i * HA_CHUNK);
720 spin_lock_irqsave(&dm_device.ha_lock, flags);
721 has->ha_end_pfn += HA_CHUNK;
723 if (total_pfn > HA_CHUNK) {
724 processed_pfn = HA_CHUNK;
725 total_pfn -= HA_CHUNK;
727 processed_pfn = total_pfn;
731 has->covered_end_pfn += processed_pfn;
732 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
734 init_completion(&dm_device.ol_waitevent);
735 dm_device.ha_waiting = !memhp_auto_online;
737 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
738 ret = add_memory(nid, PFN_PHYS((start_pfn)),
739 (HA_CHUNK << PAGE_SHIFT));
742 pr_err("hot_add memory failed error is %d\n", ret);
743 if (ret == -EEXIST) {
745 * This error indicates that the error
746 * is not a transient failure. This is the
747 * case where the guest's physical address map
748 * precludes hot adding memory. Stop all further
753 spin_lock_irqsave(&dm_device.ha_lock, flags);
754 has->ha_end_pfn -= HA_CHUNK;
755 has->covered_end_pfn -= processed_pfn;
756 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
761 * Wait for the memory block to be onlined when memory onlining
762 * is done outside of kernel (memhp_auto_online). Since the hot
763 * add has succeeded, it is ok to proceed even if the pages in
764 * the hot added region have not been "onlined" within the
767 if (dm_device.ha_waiting)
768 wait_for_completion_timeout(&dm_device.ol_waitevent,
770 post_status(&dm_device);
774 static void hv_online_page(struct page *pg)
776 struct hv_hotadd_state *has;
778 unsigned long pfn = page_to_pfn(pg);
780 spin_lock_irqsave(&dm_device.ha_lock, flags);
781 list_for_each_entry(has, &dm_device.ha_region_list, list) {
782 /* The page belongs to a different HAS. */
783 if ((pfn < has->start_pfn) || (pfn >= has->end_pfn))
786 hv_page_online_one(has, pg);
789 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
792 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
794 struct hv_hotadd_state *has;
795 struct hv_hotadd_gap *gap;
796 unsigned long residual, new_inc;
800 spin_lock_irqsave(&dm_device.ha_lock, flags);
801 list_for_each_entry(has, &dm_device.ha_region_list, list) {
803 * If the pfn range we are dealing with is not in the current
804 * "hot add block", move on.
806 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
810 * If the current start pfn is not where the covered_end
811 * is, create a gap and update covered_end_pfn.
813 if (has->covered_end_pfn != start_pfn) {
814 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
820 INIT_LIST_HEAD(&gap->list);
821 gap->start_pfn = has->covered_end_pfn;
822 gap->end_pfn = start_pfn;
823 list_add_tail(&gap->list, &has->gap_list);
825 has->covered_end_pfn = start_pfn;
829 * If the current hot add-request extends beyond
830 * our current limit; extend it.
832 if ((start_pfn + pfn_cnt) > has->end_pfn) {
833 residual = (start_pfn + pfn_cnt - has->end_pfn);
835 * Extend the region by multiples of HA_CHUNK.
837 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
838 if (residual % HA_CHUNK)
841 has->end_pfn += new_inc;
847 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
852 static unsigned long handle_pg_range(unsigned long pg_start,
853 unsigned long pg_count)
855 unsigned long start_pfn = pg_start;
856 unsigned long pfn_cnt = pg_count;
858 struct hv_hotadd_state *has;
859 unsigned long pgs_ol = 0;
860 unsigned long old_covered_state;
861 unsigned long res = 0, flags;
863 pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
866 spin_lock_irqsave(&dm_device.ha_lock, flags);
867 list_for_each_entry(has, &dm_device.ha_region_list, list) {
869 * If the pfn range we are dealing with is not in the current
870 * "hot add block", move on.
872 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
875 old_covered_state = has->covered_end_pfn;
877 if (start_pfn < has->ha_end_pfn) {
879 * This is the case where we are backing pages
880 * in an already hot added region. Bring
881 * these pages online first.
883 pgs_ol = has->ha_end_pfn - start_pfn;
884 if (pgs_ol > pfn_cnt)
887 has->covered_end_pfn += pgs_ol;
890 * Check if the corresponding memory block is already
891 * online by checking its last previously backed page.
892 * In case it is we need to bring rest (which was not
893 * backed previously) online too.
895 if (start_pfn > has->start_pfn &&
896 !PageReserved(pfn_to_page(start_pfn - 1)))
897 hv_bring_pgs_online(has, start_pfn, pgs_ol);
901 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
903 * We have some residual hot add range
904 * that needs to be hot added; hot add
905 * it now. Hot add a multiple of
906 * of HA_CHUNK that fully covers the pages
909 size = (has->end_pfn - has->ha_end_pfn);
910 if (pfn_cnt <= size) {
911 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
912 if (pfn_cnt % HA_CHUNK)
917 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
918 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
919 spin_lock_irqsave(&dm_device.ha_lock, flags);
922 * If we managed to online any pages that were given to us,
923 * we declare success.
925 res = has->covered_end_pfn - old_covered_state;
928 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
933 static unsigned long process_hot_add(unsigned long pg_start,
934 unsigned long pfn_cnt,
935 unsigned long rg_start,
936 unsigned long rg_size)
938 struct hv_hotadd_state *ha_region = NULL;
945 if (!dm_device.host_specified_ha_region) {
946 covered = pfn_covered(pg_start, pfn_cnt);
955 * If the host has specified a hot-add range; deal with it first.
959 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
963 INIT_LIST_HEAD(&ha_region->list);
964 INIT_LIST_HEAD(&ha_region->gap_list);
966 ha_region->start_pfn = rg_start;
967 ha_region->ha_end_pfn = rg_start;
968 ha_region->covered_start_pfn = pg_start;
969 ha_region->covered_end_pfn = pg_start;
970 ha_region->end_pfn = rg_start + rg_size;
972 spin_lock_irqsave(&dm_device.ha_lock, flags);
973 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
974 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
979 * Process the page range specified; bringing them
980 * online if possible.
982 return handle_pg_range(pg_start, pfn_cnt);
987 static void hot_add_req(struct work_struct *dummy)
989 struct dm_hot_add_response resp;
990 #ifdef CONFIG_MEMORY_HOTPLUG
991 unsigned long pg_start, pfn_cnt;
992 unsigned long rg_start, rg_sz;
994 struct hv_dynmem_device *dm = &dm_device;
996 memset(&resp, 0, sizeof(struct dm_hot_add_response));
997 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
998 resp.hdr.size = sizeof(struct dm_hot_add_response);
1000 #ifdef CONFIG_MEMORY_HOTPLUG
1001 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1002 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1004 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1005 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1007 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1008 unsigned long region_size;
1009 unsigned long region_start;
1012 * The host has not specified the hot-add region.
1013 * Based on the hot-add page range being specified,
1014 * compute a hot-add region that can cover the pages
1015 * that need to be hot-added while ensuring the alignment
1016 * and size requirements of Linux as it relates to hot-add.
1018 region_start = pg_start;
1019 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1020 if (pfn_cnt % HA_CHUNK)
1021 region_size += HA_CHUNK;
1023 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1025 rg_start = region_start;
1026 rg_sz = region_size;
1030 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1033 dm->num_pages_added += resp.page_count;
1036 * The result field of the response structure has the
1037 * following semantics:
1039 * 1. If all or some pages hot-added: Guest should return success.
1041 * 2. If no pages could be hot-added:
1043 * If the guest returns success, then the host
1044 * will not attempt any further hot-add operations. This
1045 * signifies a permanent failure.
1047 * If the guest returns failure, then this failure will be
1048 * treated as a transient failure and the host may retry the
1049 * hot-add operation after some delay.
1051 if (resp.page_count > 0)
1053 else if (!do_hot_add)
1058 if (!do_hot_add || (resp.page_count == 0))
1059 pr_err("Memory hot add failed\n");
1061 dm->state = DM_INITIALIZED;
1062 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1063 vmbus_sendpacket(dm->dev->channel, &resp,
1064 sizeof(struct dm_hot_add_response),
1065 (unsigned long)NULL,
1066 VM_PKT_DATA_INBAND, 0);
1069 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1071 struct dm_info_header *info_hdr;
1073 info_hdr = (struct dm_info_header *)msg->info;
1075 switch (info_hdr->type) {
1076 case INFO_TYPE_MAX_PAGE_CNT:
1077 if (info_hdr->data_size == sizeof(__u64)) {
1078 __u64 *max_page_count = (__u64 *)&info_hdr[1];
1080 pr_info("Max. dynamic memory size: %llu MB\n",
1081 (*max_page_count) >> (20 - PAGE_SHIFT));
1086 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1090 static unsigned long compute_balloon_floor(void)
1092 unsigned long min_pages;
1093 unsigned long nr_pages = totalram_pages();
1094 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1095 /* Simple continuous piecewiese linear function:
1096 * max MiB -> min MiB gradient
1106 if (nr_pages < MB2PAGES(128))
1107 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1108 else if (nr_pages < MB2PAGES(512))
1109 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1110 else if (nr_pages < MB2PAGES(2048))
1111 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1112 else if (nr_pages < MB2PAGES(8192))
1113 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1115 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1121 * Post our status as it relates memory pressure to the
1122 * host. Host expects the guests to post this status
1123 * periodically at 1 second intervals.
1125 * The metrics specified in this protocol are very Windows
1126 * specific and so we cook up numbers here to convey our memory
1130 static void post_status(struct hv_dynmem_device *dm)
1132 struct dm_status status;
1133 unsigned long now = jiffies;
1134 unsigned long last_post = last_post_time;
1136 if (pressure_report_delay > 0) {
1137 --pressure_report_delay;
1141 if (!time_after(now, (last_post_time + HZ)))
1144 memset(&status, 0, sizeof(struct dm_status));
1145 status.hdr.type = DM_STATUS_REPORT;
1146 status.hdr.size = sizeof(struct dm_status);
1147 status.hdr.trans_id = atomic_inc_return(&trans_id);
1150 * The host expects the guest to report free and committed memory.
1151 * Furthermore, the host expects the pressure information to include
1152 * the ballooned out pages. For a given amount of memory that we are
1153 * managing we need to compute a floor below which we should not
1154 * balloon. Compute this and add it to the pressure report.
1155 * We also need to report all offline pages (num_pages_added -
1156 * num_pages_onlined) as committed to the host, otherwise it can try
1157 * asking us to balloon them out.
1159 status.num_avail = si_mem_available();
1160 status.num_committed = vm_memory_committed() +
1161 dm->num_pages_ballooned +
1162 (dm->num_pages_added > dm->num_pages_onlined ?
1163 dm->num_pages_added - dm->num_pages_onlined : 0) +
1164 compute_balloon_floor();
1166 trace_balloon_status(status.num_avail, status.num_committed,
1167 vm_memory_committed(), dm->num_pages_ballooned,
1168 dm->num_pages_added, dm->num_pages_onlined);
1170 * If our transaction ID is no longer current, just don't
1171 * send the status. This can happen if we were interrupted
1172 * after we picked our transaction ID.
1174 if (status.hdr.trans_id != atomic_read(&trans_id))
1178 * If the last post time that we sampled has changed,
1179 * we have raced, don't post the status.
1181 if (last_post != last_post_time)
1184 last_post_time = jiffies;
1185 vmbus_sendpacket(dm->dev->channel, &status,
1186 sizeof(struct dm_status),
1187 (unsigned long)NULL,
1188 VM_PKT_DATA_INBAND, 0);
1192 static void free_balloon_pages(struct hv_dynmem_device *dm,
1193 union dm_mem_page_range *range_array)
1195 int num_pages = range_array->finfo.page_cnt;
1196 __u64 start_frame = range_array->finfo.start_page;
1200 for (i = 0; i < num_pages; i++) {
1201 pg = pfn_to_page(i + start_frame);
1203 dm->num_pages_ballooned--;
1209 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1210 unsigned int num_pages,
1211 struct dm_balloon_response *bl_resp,
1217 if (num_pages < alloc_unit)
1220 for (i = 0; (i * alloc_unit) < num_pages; i++) {
1221 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1223 return i * alloc_unit;
1226 * We execute this code in a thread context. Furthermore,
1227 * we don't want the kernel to try too hard.
1229 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1230 __GFP_NOMEMALLOC | __GFP_NOWARN,
1231 get_order(alloc_unit << PAGE_SHIFT));
1234 return i * alloc_unit;
1236 dm->num_pages_ballooned += alloc_unit;
1239 * If we allocatted 2M pages; split them so we
1240 * can free them in any order we get.
1243 if (alloc_unit != 1)
1244 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1246 bl_resp->range_count++;
1247 bl_resp->range_array[i].finfo.start_page =
1249 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1250 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1257 static void balloon_up(struct work_struct *dummy)
1259 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1260 unsigned int num_ballooned = 0;
1261 struct dm_balloon_response *bl_resp;
1267 unsigned long floor;
1269 /* The host balloons pages in 2M granularity. */
1270 WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1273 * We will attempt 2M allocations. However, if we fail to
1274 * allocate 2M chunks, we will go back to 4k allocations.
1278 avail_pages = si_mem_available();
1279 floor = compute_balloon_floor();
1281 /* Refuse to balloon below the floor, keep the 2M granularity. */
1282 if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1283 pr_warn("Balloon request will be partially fulfilled. %s\n",
1284 avail_pages < num_pages ? "Not enough memory." :
1285 "Balloon floor reached.");
1287 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1288 num_pages -= num_pages % PAGES_IN_2M;
1292 bl_resp = (struct dm_balloon_response *)send_buffer;
1293 memset(send_buffer, 0, PAGE_SIZE);
1294 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1295 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1296 bl_resp->more_pages = 1;
1298 num_pages -= num_ballooned;
1299 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1300 bl_resp, alloc_unit);
1302 if (alloc_unit != 1 && num_ballooned == 0) {
1307 if (num_ballooned == 0 || num_ballooned == num_pages) {
1308 pr_debug("Ballooned %u out of %u requested pages.\n",
1309 num_pages, dm_device.balloon_wrk.num_pages);
1311 bl_resp->more_pages = 0;
1313 dm_device.state = DM_INITIALIZED;
1317 * We are pushing a lot of data through the channel;
1318 * deal with transient failures caused because of the
1319 * lack of space in the ring buffer.
1323 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1324 ret = vmbus_sendpacket(dm_device.dev->channel,
1327 (unsigned long)NULL,
1328 VM_PKT_DATA_INBAND, 0);
1332 post_status(&dm_device);
1333 } while (ret == -EAGAIN);
1337 * Free up the memory we allocatted.
1339 pr_err("Balloon response failed\n");
1341 for (i = 0; i < bl_resp->range_count; i++)
1342 free_balloon_pages(&dm_device,
1343 &bl_resp->range_array[i]);
1351 static void balloon_down(struct hv_dynmem_device *dm,
1352 struct dm_unballoon_request *req)
1354 union dm_mem_page_range *range_array = req->range_array;
1355 int range_count = req->range_count;
1356 struct dm_unballoon_response resp;
1358 unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1360 for (i = 0; i < range_count; i++) {
1361 free_balloon_pages(dm, &range_array[i]);
1362 complete(&dm_device.config_event);
1365 pr_debug("Freed %u ballooned pages.\n",
1366 prev_pages_ballooned - dm->num_pages_ballooned);
1368 if (req->more_pages == 1)
1371 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1372 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1373 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1374 resp.hdr.size = sizeof(struct dm_unballoon_response);
1376 vmbus_sendpacket(dm_device.dev->channel, &resp,
1377 sizeof(struct dm_unballoon_response),
1378 (unsigned long)NULL,
1379 VM_PKT_DATA_INBAND, 0);
1381 dm->state = DM_INITIALIZED;
1384 static void balloon_onchannelcallback(void *context);
1386 static int dm_thread_func(void *dm_dev)
1388 struct hv_dynmem_device *dm = dm_dev;
1390 while (!kthread_should_stop()) {
1391 wait_for_completion_interruptible_timeout(
1392 &dm_device.config_event, 1*HZ);
1394 * The host expects us to post information on the memory
1395 * pressure every second.
1397 reinit_completion(&dm_device.config_event);
1405 static void version_resp(struct hv_dynmem_device *dm,
1406 struct dm_version_response *vresp)
1408 struct dm_version_request version_req;
1411 if (vresp->is_accepted) {
1413 * We are done; wakeup the
1414 * context waiting for version
1417 complete(&dm->host_event);
1421 * If there are more versions to try, continue
1422 * with negotiations; if not
1423 * shutdown the service since we are not able
1424 * to negotiate a suitable version number
1427 if (dm->next_version == 0)
1430 memset(&version_req, 0, sizeof(struct dm_version_request));
1431 version_req.hdr.type = DM_VERSION_REQUEST;
1432 version_req.hdr.size = sizeof(struct dm_version_request);
1433 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1434 version_req.version.version = dm->next_version;
1435 dm->version = version_req.version.version;
1438 * Set the next version to try in case current version fails.
1439 * Win7 protocol ought to be the last one to try.
1441 switch (version_req.version.version) {
1442 case DYNMEM_PROTOCOL_VERSION_WIN8:
1443 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1444 version_req.is_last_attempt = 0;
1447 dm->next_version = 0;
1448 version_req.is_last_attempt = 1;
1451 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1452 sizeof(struct dm_version_request),
1453 (unsigned long)NULL,
1454 VM_PKT_DATA_INBAND, 0);
1462 dm->state = DM_INIT_ERROR;
1463 complete(&dm->host_event);
1466 static void cap_resp(struct hv_dynmem_device *dm,
1467 struct dm_capabilities_resp_msg *cap_resp)
1469 if (!cap_resp->is_accepted) {
1470 pr_err("Capabilities not accepted by host\n");
1471 dm->state = DM_INIT_ERROR;
1473 complete(&dm->host_event);
1476 static void balloon_onchannelcallback(void *context)
1478 struct hv_device *dev = context;
1481 struct dm_message *dm_msg;
1482 struct dm_header *dm_hdr;
1483 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1484 struct dm_balloon *bal_msg;
1485 struct dm_hot_add *ha_msg;
1486 union dm_mem_page_range *ha_pg_range;
1487 union dm_mem_page_range *ha_region;
1489 memset(recv_buffer, 0, sizeof(recv_buffer));
1490 vmbus_recvpacket(dev->channel, recv_buffer,
1491 PAGE_SIZE, &recvlen, &requestid);
1494 dm_msg = (struct dm_message *)recv_buffer;
1495 dm_hdr = &dm_msg->hdr;
1497 switch (dm_hdr->type) {
1498 case DM_VERSION_RESPONSE:
1500 (struct dm_version_response *)dm_msg);
1503 case DM_CAPABILITIES_RESPONSE:
1505 (struct dm_capabilities_resp_msg *)dm_msg);
1508 case DM_BALLOON_REQUEST:
1509 if (dm->state == DM_BALLOON_UP)
1510 pr_warn("Currently ballooning\n");
1511 bal_msg = (struct dm_balloon *)recv_buffer;
1512 dm->state = DM_BALLOON_UP;
1513 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1514 schedule_work(&dm_device.balloon_wrk.wrk);
1517 case DM_UNBALLOON_REQUEST:
1518 dm->state = DM_BALLOON_DOWN;
1520 (struct dm_unballoon_request *)recv_buffer);
1523 case DM_MEM_HOT_ADD_REQUEST:
1524 if (dm->state == DM_HOT_ADD)
1525 pr_warn("Currently hot-adding\n");
1526 dm->state = DM_HOT_ADD;
1527 ha_msg = (struct dm_hot_add *)recv_buffer;
1528 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1530 * This is a normal hot-add request specifying
1533 dm->host_specified_ha_region = false;
1534 ha_pg_range = &ha_msg->range;
1535 dm->ha_wrk.ha_page_range = *ha_pg_range;
1536 dm->ha_wrk.ha_region_range.page_range = 0;
1539 * Host is specifying that we first hot-add
1540 * a region and then partially populate this
1543 dm->host_specified_ha_region = true;
1544 ha_pg_range = &ha_msg->range;
1545 ha_region = &ha_pg_range[1];
1546 dm->ha_wrk.ha_page_range = *ha_pg_range;
1547 dm->ha_wrk.ha_region_range = *ha_region;
1549 schedule_work(&dm_device.ha_wrk.wrk);
1552 case DM_INFO_MESSAGE:
1553 process_info(dm, (struct dm_info_msg *)dm_msg);
1557 pr_warn("Unhandled message: type: %d\n", dm_hdr->type);
1564 static int balloon_probe(struct hv_device *dev,
1565 const struct hv_vmbus_device_id *dev_id)
1569 struct dm_version_request version_req;
1570 struct dm_capabilities cap_msg;
1572 #ifdef CONFIG_MEMORY_HOTPLUG
1573 do_hot_add = hot_add;
1579 * First allocate a send buffer.
1582 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1586 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1587 balloon_onchannelcallback, dev);
1592 dm_device.dev = dev;
1593 dm_device.state = DM_INITIALIZING;
1594 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1595 init_completion(&dm_device.host_event);
1596 init_completion(&dm_device.config_event);
1597 INIT_LIST_HEAD(&dm_device.ha_region_list);
1598 spin_lock_init(&dm_device.ha_lock);
1599 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1600 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1601 dm_device.host_specified_ha_region = false;
1604 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1605 if (IS_ERR(dm_device.thread)) {
1606 ret = PTR_ERR(dm_device.thread);
1610 #ifdef CONFIG_MEMORY_HOTPLUG
1611 set_online_page_callback(&hv_online_page);
1612 register_memory_notifier(&hv_memory_nb);
1615 hv_set_drvdata(dev, &dm_device);
1617 * Initiate the hand shake with the host and negotiate
1618 * a version that the host can support. We start with the
1619 * highest version number and go down if the host cannot
1622 memset(&version_req, 0, sizeof(struct dm_version_request));
1623 version_req.hdr.type = DM_VERSION_REQUEST;
1624 version_req.hdr.size = sizeof(struct dm_version_request);
1625 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1626 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1627 version_req.is_last_attempt = 0;
1628 dm_device.version = version_req.version.version;
1630 ret = vmbus_sendpacket(dev->channel, &version_req,
1631 sizeof(struct dm_version_request),
1632 (unsigned long)NULL,
1633 VM_PKT_DATA_INBAND, 0);
1637 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1644 * If we could not negotiate a compatible version with the host
1645 * fail the probe function.
1647 if (dm_device.state == DM_INIT_ERROR) {
1652 pr_info("Using Dynamic Memory protocol version %u.%u\n",
1653 DYNMEM_MAJOR_VERSION(dm_device.version),
1654 DYNMEM_MINOR_VERSION(dm_device.version));
1657 * Now submit our capabilities to the host.
1659 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1660 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1661 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1662 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1664 cap_msg.caps.cap_bits.balloon = 1;
1665 cap_msg.caps.cap_bits.hot_add = 1;
1668 * Specify our alignment requirements as it relates
1669 * memory hot-add. Specify 128MB alignment.
1671 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1674 * Currently the host does not use these
1675 * values and we set them to what is done in the
1678 cap_msg.min_page_cnt = 0;
1679 cap_msg.max_page_number = -1;
1681 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1682 sizeof(struct dm_capabilities),
1683 (unsigned long)NULL,
1684 VM_PKT_DATA_INBAND, 0);
1688 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1695 * If the host does not like our capabilities,
1696 * fail the probe function.
1698 if (dm_device.state == DM_INIT_ERROR) {
1703 dm_device.state = DM_INITIALIZED;
1704 last_post_time = jiffies;
1709 #ifdef CONFIG_MEMORY_HOTPLUG
1710 restore_online_page_callback(&hv_online_page);
1712 kthread_stop(dm_device.thread);
1715 vmbus_close(dev->channel);
1721 static int balloon_remove(struct hv_device *dev)
1723 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1724 struct hv_hotadd_state *has, *tmp;
1725 struct hv_hotadd_gap *gap, *tmp_gap;
1726 unsigned long flags;
1728 if (dm->num_pages_ballooned != 0)
1729 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1731 cancel_work_sync(&dm->balloon_wrk.wrk);
1732 cancel_work_sync(&dm->ha_wrk.wrk);
1734 vmbus_close(dev->channel);
1735 kthread_stop(dm->thread);
1737 #ifdef CONFIG_MEMORY_HOTPLUG
1738 restore_online_page_callback(&hv_online_page);
1739 unregister_memory_notifier(&hv_memory_nb);
1741 spin_lock_irqsave(&dm_device.ha_lock, flags);
1742 list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1743 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1744 list_del(&gap->list);
1747 list_del(&has->list);
1750 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1755 static const struct hv_vmbus_device_id id_table[] = {
1756 /* Dynamic Memory Class ID */
1757 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1762 MODULE_DEVICE_TABLE(vmbus, id_table);
1764 static struct hv_driver balloon_drv = {
1765 .name = "hv_balloon",
1766 .id_table = id_table,
1767 .probe = balloon_probe,
1768 .remove = balloon_remove,
1770 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
1774 static int __init init_balloon_drv(void)
1777 return vmbus_driver_register(&balloon_drv);
1780 module_init(init_balloon_drv);
1782 MODULE_DESCRIPTION("Hyper-V Balloon");
1783 MODULE_LICENSE("GPL");