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>
38 * We begin with definitions supporting the Dynamic Memory protocol
41 * Begin protocol definitions.
47 * Protocol versions. The low word is the minor version, the high word the major
52 * Changed to 0.1 on 2009/03/25
53 * Changes to 0.2 on 2009/05/14
54 * Changes to 0.3 on 2009/12/03
55 * Changed to 1.0 on 2011/04/05
58 #define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
59 #define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
60 #define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
63 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
64 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
65 DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
67 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
68 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
69 DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
71 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
80 enum dm_message_type {
85 DM_VERSION_REQUEST = 1,
86 DM_VERSION_RESPONSE = 2,
87 DM_CAPABILITIES_REPORT = 3,
88 DM_CAPABILITIES_RESPONSE = 4,
90 DM_BALLOON_REQUEST = 6,
91 DM_BALLOON_RESPONSE = 7,
92 DM_UNBALLOON_REQUEST = 8,
93 DM_UNBALLOON_RESPONSE = 9,
94 DM_MEM_HOT_ADD_REQUEST = 10,
95 DM_MEM_HOT_ADD_RESPONSE = 11,
96 DM_VERSION_03_MAX = 11,
100 DM_INFO_MESSAGE = 12,
101 DM_VERSION_1_MAX = 12
106 * Structures defining the dynamic memory management
124 * To support guests that may have alignment
125 * limitations on hot-add, the guest can specify
126 * its alignment requirements; a value of n
127 * represents an alignment of 2^n in mega bytes.
129 __u64 hot_add_alignment:4;
135 union dm_mem_page_range {
138 * The PFN number of the first page in the range.
139 * 40 bits is the architectural limit of a PFN
144 * The number of pages in the range.
154 * The header for all dynamic memory messages:
156 * type: Type of the message.
157 * size: Size of the message in bytes; including the header.
158 * trans_id: The guest is responsible for manufacturing this ID.
168 * A generic message format for dynamic memory.
169 * Specific message formats are defined later in the file.
173 struct dm_header hdr;
174 __u8 data[]; /* enclosed message */
179 * Specific message types supporting the dynamic memory protocol.
183 * Version negotiation message. Sent from the guest to the host.
184 * The guest is free to try different versions until the host
185 * accepts the version.
187 * dm_version: The protocol version requested.
188 * is_last_attempt: If TRUE, this is the last version guest will request.
189 * reservedz: Reserved field, set to zero.
192 struct dm_version_request {
193 struct dm_header hdr;
194 union dm_version version;
195 __u32 is_last_attempt:1;
200 * Version response message; Host to Guest and indicates
201 * if the host has accepted the version sent by the guest.
203 * is_accepted: If TRUE, host has accepted the version and the guest
204 * should proceed to the next stage of the protocol. FALSE indicates that
205 * guest should re-try with a different version.
207 * reservedz: Reserved field, set to zero.
210 struct dm_version_response {
211 struct dm_header hdr;
217 * Message reporting capabilities. This is sent from the guest to the
221 struct dm_capabilities {
222 struct dm_header hdr;
225 __u64 max_page_number;
229 * Response to the capabilities message. This is sent from the host to the
230 * guest. This message notifies if the host has accepted the guest's
231 * capabilities. If the host has not accepted, the guest must shutdown
234 * is_accepted: Indicates if the host has accepted guest's capabilities.
235 * reservedz: Must be 0.
238 struct dm_capabilities_resp_msg {
239 struct dm_header hdr;
245 * This message is used to report memory pressure from the guest.
246 * This message is not part of any transaction and there is no
247 * response to this message.
249 * num_avail: Available memory in pages.
250 * num_committed: Committed memory in pages.
251 * page_file_size: The accumulated size of all page files
252 * in the system in pages.
253 * zero_free: The nunber of zero and free pages.
254 * page_file_writes: The writes to the page file in pages.
255 * io_diff: An indicator of file cache efficiency or page file activity,
256 * calculated as File Cache Page Fault Count - Page Read Count.
257 * This value is in pages.
259 * Some of these metrics are Windows specific and fortunately
260 * the algorithm on the host side that computes the guest memory
261 * pressure only uses num_committed value.
265 struct dm_header hdr;
268 __u64 page_file_size;
270 __u32 page_file_writes;
276 * Message to ask the guest to allocate memory - balloon up message.
277 * This message is sent from the host to the guest. The guest may not be
278 * able to allocate as much memory as requested.
280 * num_pages: number of pages to allocate.
284 struct dm_header hdr;
291 * Balloon response message; this message is sent from the guest
292 * to the host in response to the balloon message.
294 * reservedz: Reserved; must be set to zero.
295 * more_pages: If FALSE, this is the last message of the transaction.
296 * if TRUE there will atleast one more message from the guest.
298 * range_count: The number of ranges in the range array.
300 * range_array: An array of page ranges returned to the host.
304 struct dm_balloon_response {
305 struct dm_header hdr;
308 __u32 range_count:31;
309 union dm_mem_page_range range_array[];
313 * Un-balloon message; this message is sent from the host
314 * to the guest to give guest more memory.
316 * more_pages: If FALSE, this is the last message of the transaction.
317 * if TRUE there will atleast one more message from the guest.
319 * reservedz: Reserved; must be set to zero.
321 * range_count: The number of ranges in the range array.
323 * range_array: An array of page ranges returned to the host.
327 struct dm_unballoon_request {
328 struct dm_header hdr;
332 union dm_mem_page_range range_array[];
336 * Un-balloon response message; this message is sent from the guest
337 * to the host in response to an unballoon request.
341 struct dm_unballoon_response {
342 struct dm_header hdr;
347 * Hot add request message. Message sent from the host to the guest.
349 * mem_range: Memory range to hot add.
351 * On Linux we currently don't support this since we cannot hot add
352 * arbitrary granularity of memory.
356 struct dm_header hdr;
357 union dm_mem_page_range range;
361 * Hot add response message.
362 * This message is sent by the guest to report the status of a hot add request.
363 * If page_count is less than the requested page count, then the host should
364 * assume all further hot add requests will fail, since this indicates that
365 * the guest has hit an upper physical memory barrier.
367 * Hot adds may also fail due to low resources; in this case, the guest must
368 * not complete this message until the hot add can succeed, and the host must
369 * not send a new hot add request until the response is sent.
370 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
371 * times it fails the request.
374 * page_count: number of pages that were successfully hot added.
376 * result: result of the operation 1: success, 0: failure.
380 struct dm_hot_add_response {
381 struct dm_header hdr;
387 * Types of information sent from host to the guest.
391 INFO_TYPE_MAX_PAGE_CNT = 0,
397 * Header for the information message.
400 struct dm_info_header {
401 enum dm_info_type type;
406 * This message is sent from the host to the guest to pass
407 * some relevant information (win8 addition).
410 * info_size: size of the information blob.
411 * info: information blob.
415 struct dm_header hdr;
422 * End protocol definitions.
426 * State to manage hot adding memory into the guest.
427 * The range start_pfn : end_pfn specifies the range
428 * that the host has asked us to hot add. The range
429 * start_pfn : ha_end_pfn specifies the range that we have
430 * currently hot added. We hot add in multiples of 128M
431 * chunks; it is possible that we may not be able to bring
432 * online all the pages in the region. The range
433 * covered_start_pfn:covered_end_pfn defines the pages that can
437 struct hv_hotadd_state {
438 struct list_head list;
439 unsigned long start_pfn;
440 unsigned long covered_start_pfn;
441 unsigned long covered_end_pfn;
442 unsigned long ha_end_pfn;
443 unsigned long end_pfn;
447 struct list_head gap_list;
450 struct hv_hotadd_gap {
451 struct list_head list;
452 unsigned long start_pfn;
453 unsigned long end_pfn;
456 struct balloon_state {
458 struct work_struct wrk;
462 union dm_mem_page_range ha_page_range;
463 union dm_mem_page_range ha_region_range;
464 struct work_struct wrk;
467 static bool hot_add = true;
468 static bool do_hot_add;
470 * Delay reporting memory pressure by
471 * the specified number of seconds.
473 static uint pressure_report_delay = 45;
476 * The last time we posted a pressure report to host.
478 static unsigned long last_post_time;
480 module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
481 MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
483 module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
484 MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
485 static atomic_t trans_id = ATOMIC_INIT(0);
487 static int dm_ring_size = (5 * PAGE_SIZE);
490 * Driver specific state.
503 static __u8 recv_buffer[PAGE_SIZE];
504 static __u8 *send_buffer;
505 #define PAGES_IN_2M 512
506 #define HA_CHUNK (32 * 1024)
508 struct hv_dynmem_device {
509 struct hv_device *dev;
510 enum hv_dm_state state;
511 struct completion host_event;
512 struct completion config_event;
515 * Number of pages we have currently ballooned out.
517 unsigned int num_pages_ballooned;
518 unsigned int num_pages_onlined;
519 unsigned int num_pages_added;
522 * State to manage the ballooning (up) operation.
524 struct balloon_state balloon_wrk;
527 * State to execute the "hot-add" operation.
529 struct hot_add_wrk ha_wrk;
532 * This state tracks if the host has specified a hot-add
535 bool host_specified_ha_region;
538 * State to synchronize hot-add.
540 struct completion ol_waitevent;
543 * This thread handles hot-add
544 * requests from the host as well as notifying
545 * the host with regards to memory pressure in
548 struct task_struct *thread;
551 * Protects ha_region_list, num_pages_onlined counter and individual
552 * regions from ha_region_list.
557 * A list of hot-add regions.
559 struct list_head ha_region_list;
562 * We start with the highest version we can support
563 * and downgrade based on the host; we save here the
564 * next version to try.
569 static struct hv_dynmem_device dm_device;
571 static void post_status(struct hv_dynmem_device *dm);
573 #ifdef CONFIG_MEMORY_HOTPLUG
574 static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
577 struct memory_notify *mem = (struct memory_notify *)v;
582 spin_lock_irqsave(&dm_device.ha_lock, flags);
583 dm_device.num_pages_onlined += mem->nr_pages;
584 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
585 case MEM_CANCEL_ONLINE:
586 if (dm_device.ha_waiting) {
587 dm_device.ha_waiting = false;
588 complete(&dm_device.ol_waitevent);
593 spin_lock_irqsave(&dm_device.ha_lock, flags);
594 dm_device.num_pages_onlined -= mem->nr_pages;
595 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
597 case MEM_GOING_ONLINE:
598 case MEM_GOING_OFFLINE:
599 case MEM_CANCEL_OFFLINE:
605 static struct notifier_block hv_memory_nb = {
606 .notifier_call = hv_memory_notifier,
610 /* Check if the particular page is backed and can be onlined and online it. */
611 static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
613 unsigned long cur_start_pgp;
614 unsigned long cur_end_pgp;
615 struct hv_hotadd_gap *gap;
617 cur_start_pgp = (unsigned long)pfn_to_page(has->covered_start_pfn);
618 cur_end_pgp = (unsigned long)pfn_to_page(has->covered_end_pfn);
620 /* The page is not backed. */
621 if (((unsigned long)pg < cur_start_pgp) ||
622 ((unsigned long)pg >= cur_end_pgp))
625 /* Check for gaps. */
626 list_for_each_entry(gap, &has->gap_list, list) {
627 cur_start_pgp = (unsigned long)
628 pfn_to_page(gap->start_pfn);
629 cur_end_pgp = (unsigned long)
630 pfn_to_page(gap->end_pfn);
631 if (((unsigned long)pg >= cur_start_pgp) &&
632 ((unsigned long)pg < cur_end_pgp)) {
637 /* This frame is currently backed; online the page. */
638 __online_page_set_limits(pg);
639 __online_page_increment_counters(pg);
640 __online_page_free(pg);
643 static void hv_bring_pgs_online(struct hv_hotadd_state *has,
644 unsigned long start_pfn, unsigned long size)
648 for (i = 0; i < size; i++)
649 hv_page_online_one(has, pfn_to_page(start_pfn + i));
652 static void hv_mem_hot_add(unsigned long start, unsigned long size,
653 unsigned long pfn_count,
654 struct hv_hotadd_state *has)
658 unsigned long start_pfn;
659 unsigned long processed_pfn;
660 unsigned long total_pfn = pfn_count;
663 for (i = 0; i < (size/HA_CHUNK); i++) {
664 start_pfn = start + (i * HA_CHUNK);
666 spin_lock_irqsave(&dm_device.ha_lock, flags);
667 has->ha_end_pfn += HA_CHUNK;
669 if (total_pfn > HA_CHUNK) {
670 processed_pfn = HA_CHUNK;
671 total_pfn -= HA_CHUNK;
673 processed_pfn = total_pfn;
677 has->covered_end_pfn += processed_pfn;
678 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
680 init_completion(&dm_device.ol_waitevent);
681 dm_device.ha_waiting = !memhp_auto_online;
683 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
684 ret = add_memory(nid, PFN_PHYS((start_pfn)),
685 (HA_CHUNK << PAGE_SHIFT));
688 pr_info("hot_add memory failed error is %d\n", ret);
689 if (ret == -EEXIST) {
691 * This error indicates that the error
692 * is not a transient failure. This is the
693 * case where the guest's physical address map
694 * precludes hot adding memory. Stop all further
699 spin_lock_irqsave(&dm_device.ha_lock, flags);
700 has->ha_end_pfn -= HA_CHUNK;
701 has->covered_end_pfn -= processed_pfn;
702 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
707 * Wait for the memory block to be onlined when memory onlining
708 * is done outside of kernel (memhp_auto_online). Since the hot
709 * add has succeeded, it is ok to proceed even if the pages in
710 * the hot added region have not been "onlined" within the
713 if (dm_device.ha_waiting)
714 wait_for_completion_timeout(&dm_device.ol_waitevent,
716 post_status(&dm_device);
722 static void hv_online_page(struct page *pg)
724 struct hv_hotadd_state *has;
725 unsigned long cur_start_pgp;
726 unsigned long cur_end_pgp;
729 spin_lock_irqsave(&dm_device.ha_lock, flags);
730 list_for_each_entry(has, &dm_device.ha_region_list, list) {
731 cur_start_pgp = (unsigned long)
732 pfn_to_page(has->start_pfn);
733 cur_end_pgp = (unsigned long)pfn_to_page(has->end_pfn);
735 /* The page belongs to a different HAS. */
736 if (((unsigned long)pg < cur_start_pgp) ||
737 ((unsigned long)pg >= cur_end_pgp))
740 hv_page_online_one(has, pg);
743 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
746 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
748 struct hv_hotadd_state *has;
749 struct hv_hotadd_gap *gap;
750 unsigned long residual, new_inc;
754 spin_lock_irqsave(&dm_device.ha_lock, flags);
755 list_for_each_entry(has, &dm_device.ha_region_list, list) {
757 * If the pfn range we are dealing with is not in the current
758 * "hot add block", move on.
760 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
764 * If the current start pfn is not where the covered_end
765 * is, create a gap and update covered_end_pfn.
767 if (has->covered_end_pfn != start_pfn) {
768 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
774 INIT_LIST_HEAD(&gap->list);
775 gap->start_pfn = has->covered_end_pfn;
776 gap->end_pfn = start_pfn;
777 list_add_tail(&gap->list, &has->gap_list);
779 has->covered_end_pfn = start_pfn;
783 * If the current hot add-request extends beyond
784 * our current limit; extend it.
786 if ((start_pfn + pfn_cnt) > has->end_pfn) {
787 residual = (start_pfn + pfn_cnt - has->end_pfn);
789 * Extend the region by multiples of HA_CHUNK.
791 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
792 if (residual % HA_CHUNK)
795 has->end_pfn += new_inc;
801 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
806 static unsigned long handle_pg_range(unsigned long pg_start,
807 unsigned long pg_count)
809 unsigned long start_pfn = pg_start;
810 unsigned long pfn_cnt = pg_count;
812 struct hv_hotadd_state *has;
813 unsigned long pgs_ol = 0;
814 unsigned long old_covered_state;
815 unsigned long res = 0, flags;
817 spin_lock_irqsave(&dm_device.ha_lock, flags);
818 list_for_each_entry(has, &dm_device.ha_region_list, list) {
820 * If the pfn range we are dealing with is not in the current
821 * "hot add block", move on.
823 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
826 old_covered_state = has->covered_end_pfn;
828 if (start_pfn < has->ha_end_pfn) {
830 * This is the case where we are backing pages
831 * in an already hot added region. Bring
832 * these pages online first.
834 pgs_ol = has->ha_end_pfn - start_pfn;
835 if (pgs_ol > pfn_cnt)
838 has->covered_end_pfn += pgs_ol;
841 * Check if the corresponding memory block is already
842 * online by checking its last previously backed page.
843 * In case it is we need to bring rest (which was not
844 * backed previously) online too.
846 if (start_pfn > has->start_pfn &&
847 !PageReserved(pfn_to_page(start_pfn - 1)))
848 hv_bring_pgs_online(has, start_pfn, pgs_ol);
852 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
854 * We have some residual hot add range
855 * that needs to be hot added; hot add
856 * it now. Hot add a multiple of
857 * of HA_CHUNK that fully covers the pages
860 size = (has->end_pfn - has->ha_end_pfn);
861 if (pfn_cnt <= size) {
862 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
863 if (pfn_cnt % HA_CHUNK)
868 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
869 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
870 spin_lock_irqsave(&dm_device.ha_lock, flags);
873 * If we managed to online any pages that were given to us,
874 * we declare success.
876 res = has->covered_end_pfn - old_covered_state;
879 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
884 static unsigned long process_hot_add(unsigned long pg_start,
885 unsigned long pfn_cnt,
886 unsigned long rg_start,
887 unsigned long rg_size)
889 struct hv_hotadd_state *ha_region = NULL;
896 if (!dm_device.host_specified_ha_region) {
897 covered = pfn_covered(pg_start, pfn_cnt);
906 * If the host has specified a hot-add range; deal with it first.
910 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
914 INIT_LIST_HEAD(&ha_region->list);
915 INIT_LIST_HEAD(&ha_region->gap_list);
917 ha_region->start_pfn = rg_start;
918 ha_region->ha_end_pfn = rg_start;
919 ha_region->covered_start_pfn = pg_start;
920 ha_region->covered_end_pfn = pg_start;
921 ha_region->end_pfn = rg_start + rg_size;
923 spin_lock_irqsave(&dm_device.ha_lock, flags);
924 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
925 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
930 * Process the page range specified; bringing them
931 * online if possible.
933 return handle_pg_range(pg_start, pfn_cnt);
938 static void hot_add_req(struct work_struct *dummy)
940 struct dm_hot_add_response resp;
941 #ifdef CONFIG_MEMORY_HOTPLUG
942 unsigned long pg_start, pfn_cnt;
943 unsigned long rg_start, rg_sz;
945 struct hv_dynmem_device *dm = &dm_device;
947 memset(&resp, 0, sizeof(struct dm_hot_add_response));
948 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
949 resp.hdr.size = sizeof(struct dm_hot_add_response);
951 #ifdef CONFIG_MEMORY_HOTPLUG
952 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
953 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
955 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
956 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
958 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
959 unsigned long region_size;
960 unsigned long region_start;
963 * The host has not specified the hot-add region.
964 * Based on the hot-add page range being specified,
965 * compute a hot-add region that can cover the pages
966 * that need to be hot-added while ensuring the alignment
967 * and size requirements of Linux as it relates to hot-add.
969 region_start = pg_start;
970 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
971 if (pfn_cnt % HA_CHUNK)
972 region_size += HA_CHUNK;
974 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
976 rg_start = region_start;
981 resp.page_count = process_hot_add(pg_start, pfn_cnt,
984 dm->num_pages_added += resp.page_count;
987 * The result field of the response structure has the
988 * following semantics:
990 * 1. If all or some pages hot-added: Guest should return success.
992 * 2. If no pages could be hot-added:
994 * If the guest returns success, then the host
995 * will not attempt any further hot-add operations. This
996 * signifies a permanent failure.
998 * If the guest returns failure, then this failure will be
999 * treated as a transient failure and the host may retry the
1000 * hot-add operation after some delay.
1002 if (resp.page_count > 0)
1004 else if (!do_hot_add)
1009 if (!do_hot_add || (resp.page_count == 0))
1010 pr_info("Memory hot add failed\n");
1012 dm->state = DM_INITIALIZED;
1013 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1014 vmbus_sendpacket(dm->dev->channel, &resp,
1015 sizeof(struct dm_hot_add_response),
1016 (unsigned long)NULL,
1017 VM_PKT_DATA_INBAND, 0);
1020 static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1022 struct dm_info_header *info_hdr;
1024 info_hdr = (struct dm_info_header *)msg->info;
1026 switch (info_hdr->type) {
1027 case INFO_TYPE_MAX_PAGE_CNT:
1028 pr_info("Received INFO_TYPE_MAX_PAGE_CNT\n");
1029 pr_info("Data Size is %d\n", info_hdr->data_size);
1032 pr_info("Received Unknown type: %d\n", info_hdr->type);
1036 static unsigned long compute_balloon_floor(void)
1038 unsigned long min_pages;
1039 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1040 /* Simple continuous piecewiese linear function:
1041 * max MiB -> min MiB gradient
1051 if (totalram_pages < MB2PAGES(128))
1052 min_pages = MB2PAGES(8) + (totalram_pages >> 1);
1053 else if (totalram_pages < MB2PAGES(512))
1054 min_pages = MB2PAGES(40) + (totalram_pages >> 2);
1055 else if (totalram_pages < MB2PAGES(2048))
1056 min_pages = MB2PAGES(104) + (totalram_pages >> 3);
1057 else if (totalram_pages < MB2PAGES(8192))
1058 min_pages = MB2PAGES(232) + (totalram_pages >> 4);
1060 min_pages = MB2PAGES(488) + (totalram_pages >> 5);
1066 * Post our status as it relates memory pressure to the
1067 * host. Host expects the guests to post this status
1068 * periodically at 1 second intervals.
1070 * The metrics specified in this protocol are very Windows
1071 * specific and so we cook up numbers here to convey our memory
1075 static void post_status(struct hv_dynmem_device *dm)
1077 struct dm_status status;
1078 unsigned long now = jiffies;
1079 unsigned long last_post = last_post_time;
1081 if (pressure_report_delay > 0) {
1082 --pressure_report_delay;
1086 if (!time_after(now, (last_post_time + HZ)))
1089 memset(&status, 0, sizeof(struct dm_status));
1090 status.hdr.type = DM_STATUS_REPORT;
1091 status.hdr.size = sizeof(struct dm_status);
1092 status.hdr.trans_id = atomic_inc_return(&trans_id);
1095 * The host expects the guest to report free and committed memory.
1096 * Furthermore, the host expects the pressure information to include
1097 * the ballooned out pages. For a given amount of memory that we are
1098 * managing we need to compute a floor below which we should not
1099 * balloon. Compute this and add it to the pressure report.
1100 * We also need to report all offline pages (num_pages_added -
1101 * num_pages_onlined) as committed to the host, otherwise it can try
1102 * asking us to balloon them out.
1104 status.num_avail = si_mem_available();
1105 status.num_committed = vm_memory_committed() +
1106 dm->num_pages_ballooned +
1107 (dm->num_pages_added > dm->num_pages_onlined ?
1108 dm->num_pages_added - dm->num_pages_onlined : 0) +
1109 compute_balloon_floor();
1112 * If our transaction ID is no longer current, just don't
1113 * send the status. This can happen if we were interrupted
1114 * after we picked our transaction ID.
1116 if (status.hdr.trans_id != atomic_read(&trans_id))
1120 * If the last post time that we sampled has changed,
1121 * we have raced, don't post the status.
1123 if (last_post != last_post_time)
1126 last_post_time = jiffies;
1127 vmbus_sendpacket(dm->dev->channel, &status,
1128 sizeof(struct dm_status),
1129 (unsigned long)NULL,
1130 VM_PKT_DATA_INBAND, 0);
1134 static void free_balloon_pages(struct hv_dynmem_device *dm,
1135 union dm_mem_page_range *range_array)
1137 int num_pages = range_array->finfo.page_cnt;
1138 __u64 start_frame = range_array->finfo.start_page;
1142 for (i = 0; i < num_pages; i++) {
1143 pg = pfn_to_page(i + start_frame);
1145 dm->num_pages_ballooned--;
1151 static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1152 unsigned int num_pages,
1153 struct dm_balloon_response *bl_resp,
1159 if (num_pages < alloc_unit)
1162 for (i = 0; (i * alloc_unit) < num_pages; i++) {
1163 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1165 return i * alloc_unit;
1168 * We execute this code in a thread context. Furthermore,
1169 * we don't want the kernel to try too hard.
1171 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1172 __GFP_NOMEMALLOC | __GFP_NOWARN,
1173 get_order(alloc_unit << PAGE_SHIFT));
1176 return i * alloc_unit;
1178 dm->num_pages_ballooned += alloc_unit;
1181 * If we allocatted 2M pages; split them so we
1182 * can free them in any order we get.
1185 if (alloc_unit != 1)
1186 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1188 bl_resp->range_count++;
1189 bl_resp->range_array[i].finfo.start_page =
1191 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1192 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1201 static void balloon_up(struct work_struct *dummy)
1203 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1204 unsigned int num_ballooned = 0;
1205 struct dm_balloon_response *bl_resp;
1211 unsigned long floor;
1213 /* The host balloons pages in 2M granularity. */
1214 WARN_ON_ONCE(num_pages % PAGES_IN_2M != 0);
1217 * We will attempt 2M allocations. However, if we fail to
1218 * allocate 2M chunks, we will go back to 4k allocations.
1222 avail_pages = si_mem_available();
1223 floor = compute_balloon_floor();
1225 /* Refuse to balloon below the floor, keep the 2M granularity. */
1226 if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1227 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1228 num_pages -= num_pages % PAGES_IN_2M;
1232 bl_resp = (struct dm_balloon_response *)send_buffer;
1233 memset(send_buffer, 0, PAGE_SIZE);
1234 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1235 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1236 bl_resp->more_pages = 1;
1238 num_pages -= num_ballooned;
1239 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1240 bl_resp, alloc_unit);
1242 if (alloc_unit != 1 && num_ballooned == 0) {
1247 if (num_ballooned == 0 || num_ballooned == num_pages) {
1248 bl_resp->more_pages = 0;
1250 dm_device.state = DM_INITIALIZED;
1254 * We are pushing a lot of data through the channel;
1255 * deal with transient failures caused because of the
1256 * lack of space in the ring buffer.
1260 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1261 ret = vmbus_sendpacket(dm_device.dev->channel,
1264 (unsigned long)NULL,
1265 VM_PKT_DATA_INBAND, 0);
1269 post_status(&dm_device);
1270 } while (ret == -EAGAIN);
1274 * Free up the memory we allocatted.
1276 pr_info("Balloon response failed\n");
1278 for (i = 0; i < bl_resp->range_count; i++)
1279 free_balloon_pages(&dm_device,
1280 &bl_resp->range_array[i]);
1288 static void balloon_down(struct hv_dynmem_device *dm,
1289 struct dm_unballoon_request *req)
1291 union dm_mem_page_range *range_array = req->range_array;
1292 int range_count = req->range_count;
1293 struct dm_unballoon_response resp;
1296 for (i = 0; i < range_count; i++) {
1297 free_balloon_pages(dm, &range_array[i]);
1298 complete(&dm_device.config_event);
1301 if (req->more_pages == 1)
1304 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1305 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1306 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1307 resp.hdr.size = sizeof(struct dm_unballoon_response);
1309 vmbus_sendpacket(dm_device.dev->channel, &resp,
1310 sizeof(struct dm_unballoon_response),
1311 (unsigned long)NULL,
1312 VM_PKT_DATA_INBAND, 0);
1314 dm->state = DM_INITIALIZED;
1317 static void balloon_onchannelcallback(void *context);
1319 static int dm_thread_func(void *dm_dev)
1321 struct hv_dynmem_device *dm = dm_dev;
1323 while (!kthread_should_stop()) {
1324 wait_for_completion_interruptible_timeout(
1325 &dm_device.config_event, 1*HZ);
1327 * The host expects us to post information on the memory
1328 * pressure every second.
1330 reinit_completion(&dm_device.config_event);
1338 static void version_resp(struct hv_dynmem_device *dm,
1339 struct dm_version_response *vresp)
1341 struct dm_version_request version_req;
1344 if (vresp->is_accepted) {
1346 * We are done; wakeup the
1347 * context waiting for version
1350 complete(&dm->host_event);
1354 * If there are more versions to try, continue
1355 * with negotiations; if not
1356 * shutdown the service since we are not able
1357 * to negotiate a suitable version number
1360 if (dm->next_version == 0)
1363 memset(&version_req, 0, sizeof(struct dm_version_request));
1364 version_req.hdr.type = DM_VERSION_REQUEST;
1365 version_req.hdr.size = sizeof(struct dm_version_request);
1366 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1367 version_req.version.version = dm->next_version;
1370 * Set the next version to try in case current version fails.
1371 * Win7 protocol ought to be the last one to try.
1373 switch (version_req.version.version) {
1374 case DYNMEM_PROTOCOL_VERSION_WIN8:
1375 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1376 version_req.is_last_attempt = 0;
1379 dm->next_version = 0;
1380 version_req.is_last_attempt = 1;
1383 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1384 sizeof(struct dm_version_request),
1385 (unsigned long)NULL,
1386 VM_PKT_DATA_INBAND, 0);
1394 dm->state = DM_INIT_ERROR;
1395 complete(&dm->host_event);
1398 static void cap_resp(struct hv_dynmem_device *dm,
1399 struct dm_capabilities_resp_msg *cap_resp)
1401 if (!cap_resp->is_accepted) {
1402 pr_info("Capabilities not accepted by host\n");
1403 dm->state = DM_INIT_ERROR;
1405 complete(&dm->host_event);
1408 static void balloon_onchannelcallback(void *context)
1410 struct hv_device *dev = context;
1413 struct dm_message *dm_msg;
1414 struct dm_header *dm_hdr;
1415 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1416 struct dm_balloon *bal_msg;
1417 struct dm_hot_add *ha_msg;
1418 union dm_mem_page_range *ha_pg_range;
1419 union dm_mem_page_range *ha_region;
1421 memset(recv_buffer, 0, sizeof(recv_buffer));
1422 vmbus_recvpacket(dev->channel, recv_buffer,
1423 PAGE_SIZE, &recvlen, &requestid);
1426 dm_msg = (struct dm_message *)recv_buffer;
1427 dm_hdr = &dm_msg->hdr;
1429 switch (dm_hdr->type) {
1430 case DM_VERSION_RESPONSE:
1432 (struct dm_version_response *)dm_msg);
1435 case DM_CAPABILITIES_RESPONSE:
1437 (struct dm_capabilities_resp_msg *)dm_msg);
1440 case DM_BALLOON_REQUEST:
1441 if (dm->state == DM_BALLOON_UP)
1442 pr_warn("Currently ballooning\n");
1443 bal_msg = (struct dm_balloon *)recv_buffer;
1444 dm->state = DM_BALLOON_UP;
1445 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1446 schedule_work(&dm_device.balloon_wrk.wrk);
1449 case DM_UNBALLOON_REQUEST:
1450 dm->state = DM_BALLOON_DOWN;
1452 (struct dm_unballoon_request *)recv_buffer);
1455 case DM_MEM_HOT_ADD_REQUEST:
1456 if (dm->state == DM_HOT_ADD)
1457 pr_warn("Currently hot-adding\n");
1458 dm->state = DM_HOT_ADD;
1459 ha_msg = (struct dm_hot_add *)recv_buffer;
1460 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1462 * This is a normal hot-add request specifying
1465 dm->host_specified_ha_region = false;
1466 ha_pg_range = &ha_msg->range;
1467 dm->ha_wrk.ha_page_range = *ha_pg_range;
1468 dm->ha_wrk.ha_region_range.page_range = 0;
1471 * Host is specifying that we first hot-add
1472 * a region and then partially populate this
1475 dm->host_specified_ha_region = true;
1476 ha_pg_range = &ha_msg->range;
1477 ha_region = &ha_pg_range[1];
1478 dm->ha_wrk.ha_page_range = *ha_pg_range;
1479 dm->ha_wrk.ha_region_range = *ha_region;
1481 schedule_work(&dm_device.ha_wrk.wrk);
1484 case DM_INFO_MESSAGE:
1485 process_info(dm, (struct dm_info_msg *)dm_msg);
1489 pr_err("Unhandled message: type: %d\n", dm_hdr->type);
1496 static int balloon_probe(struct hv_device *dev,
1497 const struct hv_vmbus_device_id *dev_id)
1501 struct dm_version_request version_req;
1502 struct dm_capabilities cap_msg;
1504 do_hot_add = hot_add;
1507 * First allocate a send buffer.
1510 send_buffer = kmalloc(PAGE_SIZE, GFP_KERNEL);
1514 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1515 balloon_onchannelcallback, dev);
1520 dm_device.dev = dev;
1521 dm_device.state = DM_INITIALIZING;
1522 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1523 init_completion(&dm_device.host_event);
1524 init_completion(&dm_device.config_event);
1525 INIT_LIST_HEAD(&dm_device.ha_region_list);
1526 spin_lock_init(&dm_device.ha_lock);
1527 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1528 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1529 dm_device.host_specified_ha_region = false;
1532 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1533 if (IS_ERR(dm_device.thread)) {
1534 ret = PTR_ERR(dm_device.thread);
1538 #ifdef CONFIG_MEMORY_HOTPLUG
1539 set_online_page_callback(&hv_online_page);
1540 register_memory_notifier(&hv_memory_nb);
1543 hv_set_drvdata(dev, &dm_device);
1545 * Initiate the hand shake with the host and negotiate
1546 * a version that the host can support. We start with the
1547 * highest version number and go down if the host cannot
1550 memset(&version_req, 0, sizeof(struct dm_version_request));
1551 version_req.hdr.type = DM_VERSION_REQUEST;
1552 version_req.hdr.size = sizeof(struct dm_version_request);
1553 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1554 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1555 version_req.is_last_attempt = 0;
1557 ret = vmbus_sendpacket(dev->channel, &version_req,
1558 sizeof(struct dm_version_request),
1559 (unsigned long)NULL,
1560 VM_PKT_DATA_INBAND, 0);
1564 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1571 * If we could not negotiate a compatible version with the host
1572 * fail the probe function.
1574 if (dm_device.state == DM_INIT_ERROR) {
1579 * Now submit our capabilities to the host.
1581 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1582 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1583 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1584 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1586 cap_msg.caps.cap_bits.balloon = 1;
1587 cap_msg.caps.cap_bits.hot_add = 1;
1590 * Specify our alignment requirements as it relates
1591 * memory hot-add. Specify 128MB alignment.
1593 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1596 * Currently the host does not use these
1597 * values and we set them to what is done in the
1600 cap_msg.min_page_cnt = 0;
1601 cap_msg.max_page_number = -1;
1603 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1604 sizeof(struct dm_capabilities),
1605 (unsigned long)NULL,
1606 VM_PKT_DATA_INBAND, 0);
1610 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1617 * If the host does not like our capabilities,
1618 * fail the probe function.
1620 if (dm_device.state == DM_INIT_ERROR) {
1625 dm_device.state = DM_INITIALIZED;
1630 #ifdef CONFIG_MEMORY_HOTPLUG
1631 restore_online_page_callback(&hv_online_page);
1633 kthread_stop(dm_device.thread);
1636 vmbus_close(dev->channel);
1642 static int balloon_remove(struct hv_device *dev)
1644 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1645 struct hv_hotadd_state *has, *tmp;
1646 struct hv_hotadd_gap *gap, *tmp_gap;
1647 unsigned long flags;
1649 if (dm->num_pages_ballooned != 0)
1650 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
1652 cancel_work_sync(&dm->balloon_wrk.wrk);
1653 cancel_work_sync(&dm->ha_wrk.wrk);
1655 vmbus_close(dev->channel);
1656 kthread_stop(dm->thread);
1658 #ifdef CONFIG_MEMORY_HOTPLUG
1659 restore_online_page_callback(&hv_online_page);
1660 unregister_memory_notifier(&hv_memory_nb);
1662 spin_lock_irqsave(&dm_device.ha_lock, flags);
1663 list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
1664 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
1665 list_del(&gap->list);
1668 list_del(&has->list);
1671 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
1676 static const struct hv_vmbus_device_id id_table[] = {
1677 /* Dynamic Memory Class ID */
1678 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1683 MODULE_DEVICE_TABLE(vmbus, id_table);
1685 static struct hv_driver balloon_drv = {
1686 .name = "hv_balloon",
1687 .id_table = id_table,
1688 .probe = balloon_probe,
1689 .remove = balloon_remove,
1692 static int __init init_balloon_drv(void)
1695 return vmbus_driver_register(&balloon_drv);
1698 module_init(init_balloon_drv);
1700 MODULE_DESCRIPTION("Hyper-V Balloon");
1701 MODULE_LICENSE("GPL");