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 * The negotiated version agreed by host.
574 static struct hv_dynmem_device dm_device
;
576 static void post_status(struct hv_dynmem_device
*dm
);
578 #ifdef CONFIG_MEMORY_HOTPLUG
579 static int hv_memory_notifier(struct notifier_block
*nb
, unsigned long val
,
582 struct memory_notify
*mem
= (struct memory_notify
*)v
;
587 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
588 dm_device
.num_pages_onlined
+= mem
->nr_pages
;
589 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
591 case MEM_CANCEL_ONLINE
:
592 if (dm_device
.ha_waiting
) {
593 dm_device
.ha_waiting
= false;
594 complete(&dm_device
.ol_waitevent
);
599 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
600 dm_device
.num_pages_onlined
-= mem
->nr_pages
;
601 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
603 case MEM_GOING_ONLINE
:
604 case MEM_GOING_OFFLINE
:
605 case MEM_CANCEL_OFFLINE
:
611 static struct notifier_block hv_memory_nb
= {
612 .notifier_call
= hv_memory_notifier
,
616 /* Check if the particular page is backed and can be onlined and online it. */
617 static void hv_page_online_one(struct hv_hotadd_state
*has
, struct page
*pg
)
619 unsigned long cur_start_pgp
;
620 unsigned long cur_end_pgp
;
621 struct hv_hotadd_gap
*gap
;
623 cur_start_pgp
= (unsigned long)pfn_to_page(has
->covered_start_pfn
);
624 cur_end_pgp
= (unsigned long)pfn_to_page(has
->covered_end_pfn
);
626 /* The page is not backed. */
627 if (((unsigned long)pg
< cur_start_pgp
) ||
628 ((unsigned long)pg
>= cur_end_pgp
))
631 /* Check for gaps. */
632 list_for_each_entry(gap
, &has
->gap_list
, list
) {
633 cur_start_pgp
= (unsigned long)
634 pfn_to_page(gap
->start_pfn
);
635 cur_end_pgp
= (unsigned long)
636 pfn_to_page(gap
->end_pfn
);
637 if (((unsigned long)pg
>= cur_start_pgp
) &&
638 ((unsigned long)pg
< cur_end_pgp
)) {
643 /* This frame is currently backed; online the page. */
644 __online_page_set_limits(pg
);
645 __online_page_increment_counters(pg
);
646 __online_page_free(pg
);
649 static void hv_bring_pgs_online(struct hv_hotadd_state
*has
,
650 unsigned long start_pfn
, unsigned long size
)
654 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size
, start_pfn
);
655 for (i
= 0; i
< size
; i
++)
656 hv_page_online_one(has
, pfn_to_page(start_pfn
+ i
));
659 static void hv_mem_hot_add(unsigned long start
, unsigned long size
,
660 unsigned long pfn_count
,
661 struct hv_hotadd_state
*has
)
665 unsigned long start_pfn
;
666 unsigned long processed_pfn
;
667 unsigned long total_pfn
= pfn_count
;
670 for (i
= 0; i
< (size
/HA_CHUNK
); i
++) {
671 start_pfn
= start
+ (i
* HA_CHUNK
);
673 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
674 has
->ha_end_pfn
+= HA_CHUNK
;
676 if (total_pfn
> HA_CHUNK
) {
677 processed_pfn
= HA_CHUNK
;
678 total_pfn
-= HA_CHUNK
;
680 processed_pfn
= total_pfn
;
684 has
->covered_end_pfn
+= processed_pfn
;
685 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
687 init_completion(&dm_device
.ol_waitevent
);
688 dm_device
.ha_waiting
= !memhp_auto_online
;
690 nid
= memory_add_physaddr_to_nid(PFN_PHYS(start_pfn
));
691 ret
= add_memory(nid
, PFN_PHYS((start_pfn
)),
692 (HA_CHUNK
<< PAGE_SHIFT
));
695 pr_warn("hot_add memory failed error is %d\n", ret
);
696 if (ret
== -EEXIST
) {
698 * This error indicates that the error
699 * is not a transient failure. This is the
700 * case where the guest's physical address map
701 * precludes hot adding memory. Stop all further
706 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
707 has
->ha_end_pfn
-= HA_CHUNK
;
708 has
->covered_end_pfn
-= processed_pfn
;
709 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
714 * Wait for the memory block to be onlined when memory onlining
715 * is done outside of kernel (memhp_auto_online). Since the hot
716 * add has succeeded, it is ok to proceed even if the pages in
717 * the hot added region have not been "onlined" within the
720 if (dm_device
.ha_waiting
)
721 wait_for_completion_timeout(&dm_device
.ol_waitevent
,
723 post_status(&dm_device
);
729 static void hv_online_page(struct page
*pg
)
731 struct hv_hotadd_state
*has
;
732 unsigned long cur_start_pgp
;
733 unsigned long cur_end_pgp
;
736 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
737 list_for_each_entry(has
, &dm_device
.ha_region_list
, list
) {
738 cur_start_pgp
= (unsigned long)
739 pfn_to_page(has
->start_pfn
);
740 cur_end_pgp
= (unsigned long)pfn_to_page(has
->end_pfn
);
742 /* The page belongs to a different HAS. */
743 if (((unsigned long)pg
< cur_start_pgp
) ||
744 ((unsigned long)pg
>= cur_end_pgp
))
747 hv_page_online_one(has
, pg
);
750 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
753 static int pfn_covered(unsigned long start_pfn
, unsigned long pfn_cnt
)
755 struct hv_hotadd_state
*has
;
756 struct hv_hotadd_gap
*gap
;
757 unsigned long residual
, new_inc
;
761 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
762 list_for_each_entry(has
, &dm_device
.ha_region_list
, list
) {
764 * If the pfn range we are dealing with is not in the current
765 * "hot add block", move on.
767 if (start_pfn
< has
->start_pfn
|| start_pfn
>= has
->end_pfn
)
771 * If the current start pfn is not where the covered_end
772 * is, create a gap and update covered_end_pfn.
774 if (has
->covered_end_pfn
!= start_pfn
) {
775 gap
= kzalloc(sizeof(struct hv_hotadd_gap
), GFP_ATOMIC
);
781 INIT_LIST_HEAD(&gap
->list
);
782 gap
->start_pfn
= has
->covered_end_pfn
;
783 gap
->end_pfn
= start_pfn
;
784 list_add_tail(&gap
->list
, &has
->gap_list
);
786 has
->covered_end_pfn
= start_pfn
;
790 * If the current hot add-request extends beyond
791 * our current limit; extend it.
793 if ((start_pfn
+ pfn_cnt
) > has
->end_pfn
) {
794 residual
= (start_pfn
+ pfn_cnt
- has
->end_pfn
);
796 * Extend the region by multiples of HA_CHUNK.
798 new_inc
= (residual
/ HA_CHUNK
) * HA_CHUNK
;
799 if (residual
% HA_CHUNK
)
802 has
->end_pfn
+= new_inc
;
808 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
813 static unsigned long handle_pg_range(unsigned long pg_start
,
814 unsigned long pg_count
)
816 unsigned long start_pfn
= pg_start
;
817 unsigned long pfn_cnt
= pg_count
;
819 struct hv_hotadd_state
*has
;
820 unsigned long pgs_ol
= 0;
821 unsigned long old_covered_state
;
822 unsigned long res
= 0, flags
;
824 pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count
,
827 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
828 list_for_each_entry(has
, &dm_device
.ha_region_list
, list
) {
830 * If the pfn range we are dealing with is not in the current
831 * "hot add block", move on.
833 if (start_pfn
< has
->start_pfn
|| start_pfn
>= has
->end_pfn
)
836 old_covered_state
= has
->covered_end_pfn
;
838 if (start_pfn
< has
->ha_end_pfn
) {
840 * This is the case where we are backing pages
841 * in an already hot added region. Bring
842 * these pages online first.
844 pgs_ol
= has
->ha_end_pfn
- start_pfn
;
845 if (pgs_ol
> pfn_cnt
)
848 has
->covered_end_pfn
+= pgs_ol
;
851 * Check if the corresponding memory block is already
852 * online by checking its last previously backed page.
853 * In case it is we need to bring rest (which was not
854 * backed previously) online too.
856 if (start_pfn
> has
->start_pfn
&&
857 !PageReserved(pfn_to_page(start_pfn
- 1)))
858 hv_bring_pgs_online(has
, start_pfn
, pgs_ol
);
862 if ((has
->ha_end_pfn
< has
->end_pfn
) && (pfn_cnt
> 0)) {
864 * We have some residual hot add range
865 * that needs to be hot added; hot add
866 * it now. Hot add a multiple of
867 * of HA_CHUNK that fully covers the pages
870 size
= (has
->end_pfn
- has
->ha_end_pfn
);
871 if (pfn_cnt
<= size
) {
872 size
= ((pfn_cnt
/ HA_CHUNK
) * HA_CHUNK
);
873 if (pfn_cnt
% HA_CHUNK
)
878 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
879 hv_mem_hot_add(has
->ha_end_pfn
, size
, pfn_cnt
, has
);
880 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
883 * If we managed to online any pages that were given to us,
884 * we declare success.
886 res
= has
->covered_end_pfn
- old_covered_state
;
889 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
894 static unsigned long process_hot_add(unsigned long pg_start
,
895 unsigned long pfn_cnt
,
896 unsigned long rg_start
,
897 unsigned long rg_size
)
899 struct hv_hotadd_state
*ha_region
= NULL
;
906 if (!dm_device
.host_specified_ha_region
) {
907 covered
= pfn_covered(pg_start
, pfn_cnt
);
916 * If the host has specified a hot-add range; deal with it first.
920 ha_region
= kzalloc(sizeof(struct hv_hotadd_state
), GFP_KERNEL
);
924 INIT_LIST_HEAD(&ha_region
->list
);
925 INIT_LIST_HEAD(&ha_region
->gap_list
);
927 ha_region
->start_pfn
= rg_start
;
928 ha_region
->ha_end_pfn
= rg_start
;
929 ha_region
->covered_start_pfn
= pg_start
;
930 ha_region
->covered_end_pfn
= pg_start
;
931 ha_region
->end_pfn
= rg_start
+ rg_size
;
933 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
934 list_add_tail(&ha_region
->list
, &dm_device
.ha_region_list
);
935 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
940 * Process the page range specified; bringing them
941 * online if possible.
943 return handle_pg_range(pg_start
, pfn_cnt
);
948 static void hot_add_req(struct work_struct
*dummy
)
950 struct dm_hot_add_response resp
;
951 #ifdef CONFIG_MEMORY_HOTPLUG
952 unsigned long pg_start
, pfn_cnt
;
953 unsigned long rg_start
, rg_sz
;
955 struct hv_dynmem_device
*dm
= &dm_device
;
957 memset(&resp
, 0, sizeof(struct dm_hot_add_response
));
958 resp
.hdr
.type
= DM_MEM_HOT_ADD_RESPONSE
;
959 resp
.hdr
.size
= sizeof(struct dm_hot_add_response
);
961 #ifdef CONFIG_MEMORY_HOTPLUG
962 pg_start
= dm
->ha_wrk
.ha_page_range
.finfo
.start_page
;
963 pfn_cnt
= dm
->ha_wrk
.ha_page_range
.finfo
.page_cnt
;
965 rg_start
= dm
->ha_wrk
.ha_region_range
.finfo
.start_page
;
966 rg_sz
= dm
->ha_wrk
.ha_region_range
.finfo
.page_cnt
;
968 if ((rg_start
== 0) && (!dm
->host_specified_ha_region
)) {
969 unsigned long region_size
;
970 unsigned long region_start
;
973 * The host has not specified the hot-add region.
974 * Based on the hot-add page range being specified,
975 * compute a hot-add region that can cover the pages
976 * that need to be hot-added while ensuring the alignment
977 * and size requirements of Linux as it relates to hot-add.
979 region_start
= pg_start
;
980 region_size
= (pfn_cnt
/ HA_CHUNK
) * HA_CHUNK
;
981 if (pfn_cnt
% HA_CHUNK
)
982 region_size
+= HA_CHUNK
;
984 region_start
= (pg_start
/ HA_CHUNK
) * HA_CHUNK
;
986 rg_start
= region_start
;
991 resp
.page_count
= process_hot_add(pg_start
, pfn_cnt
,
994 dm
->num_pages_added
+= resp
.page_count
;
997 * The result field of the response structure has the
998 * following semantics:
1000 * 1. If all or some pages hot-added: Guest should return success.
1002 * 2. If no pages could be hot-added:
1004 * If the guest returns success, then the host
1005 * will not attempt any further hot-add operations. This
1006 * signifies a permanent failure.
1008 * If the guest returns failure, then this failure will be
1009 * treated as a transient failure and the host may retry the
1010 * hot-add operation after some delay.
1012 if (resp
.page_count
> 0)
1014 else if (!do_hot_add
)
1019 if (!do_hot_add
|| (resp
.page_count
== 0))
1020 pr_info("Memory hot add failed\n");
1022 dm
->state
= DM_INITIALIZED
;
1023 resp
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1024 vmbus_sendpacket(dm
->dev
->channel
, &resp
,
1025 sizeof(struct dm_hot_add_response
),
1026 (unsigned long)NULL
,
1027 VM_PKT_DATA_INBAND
, 0);
1030 static void process_info(struct hv_dynmem_device
*dm
, struct dm_info_msg
*msg
)
1032 struct dm_info_header
*info_hdr
;
1034 info_hdr
= (struct dm_info_header
*)msg
->info
;
1036 switch (info_hdr
->type
) {
1037 case INFO_TYPE_MAX_PAGE_CNT
:
1038 if (info_hdr
->data_size
== sizeof(__u64
)) {
1039 __u64
*max_page_count
= (__u64
*)&info_hdr
[1];
1041 pr_info("INFO_TYPE_MAX_PAGE_CNT = %llu\n",
1047 pr_info("Received Unknown type: %d\n", info_hdr
->type
);
1051 static unsigned long compute_balloon_floor(void)
1053 unsigned long min_pages
;
1054 #define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1055 /* Simple continuous piecewiese linear function:
1056 * max MiB -> min MiB gradient
1066 if (totalram_pages
< MB2PAGES(128))
1067 min_pages
= MB2PAGES(8) + (totalram_pages
>> 1);
1068 else if (totalram_pages
< MB2PAGES(512))
1069 min_pages
= MB2PAGES(40) + (totalram_pages
>> 2);
1070 else if (totalram_pages
< MB2PAGES(2048))
1071 min_pages
= MB2PAGES(104) + (totalram_pages
>> 3);
1072 else if (totalram_pages
< MB2PAGES(8192))
1073 min_pages
= MB2PAGES(232) + (totalram_pages
>> 4);
1075 min_pages
= MB2PAGES(488) + (totalram_pages
>> 5);
1081 * Post our status as it relates memory pressure to the
1082 * host. Host expects the guests to post this status
1083 * periodically at 1 second intervals.
1085 * The metrics specified in this protocol are very Windows
1086 * specific and so we cook up numbers here to convey our memory
1090 static void post_status(struct hv_dynmem_device
*dm
)
1092 struct dm_status status
;
1093 unsigned long now
= jiffies
;
1094 unsigned long last_post
= last_post_time
;
1096 if (pressure_report_delay
> 0) {
1097 --pressure_report_delay
;
1101 if (!time_after(now
, (last_post_time
+ HZ
)))
1104 memset(&status
, 0, sizeof(struct dm_status
));
1105 status
.hdr
.type
= DM_STATUS_REPORT
;
1106 status
.hdr
.size
= sizeof(struct dm_status
);
1107 status
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1110 * The host expects the guest to report free and committed memory.
1111 * Furthermore, the host expects the pressure information to include
1112 * the ballooned out pages. For a given amount of memory that we are
1113 * managing we need to compute a floor below which we should not
1114 * balloon. Compute this and add it to the pressure report.
1115 * We also need to report all offline pages (num_pages_added -
1116 * num_pages_onlined) as committed to the host, otherwise it can try
1117 * asking us to balloon them out.
1119 status
.num_avail
= si_mem_available();
1120 status
.num_committed
= vm_memory_committed() +
1121 dm
->num_pages_ballooned
+
1122 (dm
->num_pages_added
> dm
->num_pages_onlined
?
1123 dm
->num_pages_added
- dm
->num_pages_onlined
: 0) +
1124 compute_balloon_floor();
1127 * If our transaction ID is no longer current, just don't
1128 * send the status. This can happen if we were interrupted
1129 * after we picked our transaction ID.
1131 if (status
.hdr
.trans_id
!= atomic_read(&trans_id
))
1135 * If the last post time that we sampled has changed,
1136 * we have raced, don't post the status.
1138 if (last_post
!= last_post_time
)
1141 last_post_time
= jiffies
;
1142 vmbus_sendpacket(dm
->dev
->channel
, &status
,
1143 sizeof(struct dm_status
),
1144 (unsigned long)NULL
,
1145 VM_PKT_DATA_INBAND
, 0);
1149 static void free_balloon_pages(struct hv_dynmem_device
*dm
,
1150 union dm_mem_page_range
*range_array
)
1152 int num_pages
= range_array
->finfo
.page_cnt
;
1153 __u64 start_frame
= range_array
->finfo
.start_page
;
1157 for (i
= 0; i
< num_pages
; i
++) {
1158 pg
= pfn_to_page(i
+ start_frame
);
1160 dm
->num_pages_ballooned
--;
1166 static unsigned int alloc_balloon_pages(struct hv_dynmem_device
*dm
,
1167 unsigned int num_pages
,
1168 struct dm_balloon_response
*bl_resp
,
1174 if (num_pages
< alloc_unit
)
1177 for (i
= 0; (i
* alloc_unit
) < num_pages
; i
++) {
1178 if (bl_resp
->hdr
.size
+ sizeof(union dm_mem_page_range
) >
1180 return i
* alloc_unit
;
1183 * We execute this code in a thread context. Furthermore,
1184 * we don't want the kernel to try too hard.
1186 pg
= alloc_pages(GFP_HIGHUSER
| __GFP_NORETRY
|
1187 __GFP_NOMEMALLOC
| __GFP_NOWARN
,
1188 get_order(alloc_unit
<< PAGE_SHIFT
));
1191 return i
* alloc_unit
;
1193 dm
->num_pages_ballooned
+= alloc_unit
;
1196 * If we allocatted 2M pages; split them so we
1197 * can free them in any order we get.
1200 if (alloc_unit
!= 1)
1201 split_page(pg
, get_order(alloc_unit
<< PAGE_SHIFT
));
1203 bl_resp
->range_count
++;
1204 bl_resp
->range_array
[i
].finfo
.start_page
=
1206 bl_resp
->range_array
[i
].finfo
.page_cnt
= alloc_unit
;
1207 bl_resp
->hdr
.size
+= sizeof(union dm_mem_page_range
);
1214 static void balloon_up(struct work_struct
*dummy
)
1216 unsigned int num_pages
= dm_device
.balloon_wrk
.num_pages
;
1217 unsigned int num_ballooned
= 0;
1218 struct dm_balloon_response
*bl_resp
;
1224 unsigned long floor
;
1226 /* The host balloons pages in 2M granularity. */
1227 WARN_ON_ONCE(num_pages
% PAGES_IN_2M
!= 0);
1230 * We will attempt 2M allocations. However, if we fail to
1231 * allocate 2M chunks, we will go back to 4k allocations.
1235 avail_pages
= si_mem_available();
1236 floor
= compute_balloon_floor();
1238 /* Refuse to balloon below the floor, keep the 2M granularity. */
1239 if (avail_pages
< num_pages
|| avail_pages
- num_pages
< floor
) {
1240 pr_warn("Balloon request will be partially fulfilled. %s\n",
1241 avail_pages
< num_pages
? "Not enough memory." :
1242 "Balloon floor reached.");
1244 num_pages
= avail_pages
> floor
? (avail_pages
- floor
) : 0;
1245 num_pages
-= num_pages
% PAGES_IN_2M
;
1249 bl_resp
= (struct dm_balloon_response
*)send_buffer
;
1250 memset(send_buffer
, 0, PAGE_SIZE
);
1251 bl_resp
->hdr
.type
= DM_BALLOON_RESPONSE
;
1252 bl_resp
->hdr
.size
= sizeof(struct dm_balloon_response
);
1253 bl_resp
->more_pages
= 1;
1255 num_pages
-= num_ballooned
;
1256 num_ballooned
= alloc_balloon_pages(&dm_device
, num_pages
,
1257 bl_resp
, alloc_unit
);
1259 if (alloc_unit
!= 1 && num_ballooned
== 0) {
1264 if (num_ballooned
== 0 || num_ballooned
== num_pages
) {
1265 pr_debug("Ballooned %u out of %u requested pages.\n",
1266 num_pages
, dm_device
.balloon_wrk
.num_pages
);
1268 bl_resp
->more_pages
= 0;
1270 dm_device
.state
= DM_INITIALIZED
;
1274 * We are pushing a lot of data through the channel;
1275 * deal with transient failures caused because of the
1276 * lack of space in the ring buffer.
1280 bl_resp
->hdr
.trans_id
= atomic_inc_return(&trans_id
);
1281 ret
= vmbus_sendpacket(dm_device
.dev
->channel
,
1284 (unsigned long)NULL
,
1285 VM_PKT_DATA_INBAND
, 0);
1289 post_status(&dm_device
);
1290 } while (ret
== -EAGAIN
);
1294 * Free up the memory we allocatted.
1296 pr_info("Balloon response failed\n");
1298 for (i
= 0; i
< bl_resp
->range_count
; i
++)
1299 free_balloon_pages(&dm_device
,
1300 &bl_resp
->range_array
[i
]);
1308 static void balloon_down(struct hv_dynmem_device
*dm
,
1309 struct dm_unballoon_request
*req
)
1311 union dm_mem_page_range
*range_array
= req
->range_array
;
1312 int range_count
= req
->range_count
;
1313 struct dm_unballoon_response resp
;
1315 unsigned int prev_pages_ballooned
= dm
->num_pages_ballooned
;
1317 for (i
= 0; i
< range_count
; i
++) {
1318 free_balloon_pages(dm
, &range_array
[i
]);
1319 complete(&dm_device
.config_event
);
1322 pr_debug("Freed %u ballooned pages.\n",
1323 prev_pages_ballooned
- dm
->num_pages_ballooned
);
1325 if (req
->more_pages
== 1)
1328 memset(&resp
, 0, sizeof(struct dm_unballoon_response
));
1329 resp
.hdr
.type
= DM_UNBALLOON_RESPONSE
;
1330 resp
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1331 resp
.hdr
.size
= sizeof(struct dm_unballoon_response
);
1333 vmbus_sendpacket(dm_device
.dev
->channel
, &resp
,
1334 sizeof(struct dm_unballoon_response
),
1335 (unsigned long)NULL
,
1336 VM_PKT_DATA_INBAND
, 0);
1338 dm
->state
= DM_INITIALIZED
;
1341 static void balloon_onchannelcallback(void *context
);
1343 static int dm_thread_func(void *dm_dev
)
1345 struct hv_dynmem_device
*dm
= dm_dev
;
1347 while (!kthread_should_stop()) {
1348 wait_for_completion_interruptible_timeout(
1349 &dm_device
.config_event
, 1*HZ
);
1351 * The host expects us to post information on the memory
1352 * pressure every second.
1354 reinit_completion(&dm_device
.config_event
);
1362 static void version_resp(struct hv_dynmem_device
*dm
,
1363 struct dm_version_response
*vresp
)
1365 struct dm_version_request version_req
;
1368 if (vresp
->is_accepted
) {
1370 * We are done; wakeup the
1371 * context waiting for version
1374 complete(&dm
->host_event
);
1378 * If there are more versions to try, continue
1379 * with negotiations; if not
1380 * shutdown the service since we are not able
1381 * to negotiate a suitable version number
1384 if (dm
->next_version
== 0)
1387 memset(&version_req
, 0, sizeof(struct dm_version_request
));
1388 version_req
.hdr
.type
= DM_VERSION_REQUEST
;
1389 version_req
.hdr
.size
= sizeof(struct dm_version_request
);
1390 version_req
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1391 version_req
.version
.version
= dm
->next_version
;
1392 dm
->version
= version_req
.version
.version
;
1395 * Set the next version to try in case current version fails.
1396 * Win7 protocol ought to be the last one to try.
1398 switch (version_req
.version
.version
) {
1399 case DYNMEM_PROTOCOL_VERSION_WIN8
:
1400 dm
->next_version
= DYNMEM_PROTOCOL_VERSION_WIN7
;
1401 version_req
.is_last_attempt
= 0;
1404 dm
->next_version
= 0;
1405 version_req
.is_last_attempt
= 1;
1408 ret
= vmbus_sendpacket(dm
->dev
->channel
, &version_req
,
1409 sizeof(struct dm_version_request
),
1410 (unsigned long)NULL
,
1411 VM_PKT_DATA_INBAND
, 0);
1419 dm
->state
= DM_INIT_ERROR
;
1420 complete(&dm
->host_event
);
1423 static void cap_resp(struct hv_dynmem_device
*dm
,
1424 struct dm_capabilities_resp_msg
*cap_resp
)
1426 if (!cap_resp
->is_accepted
) {
1427 pr_info("Capabilities not accepted by host\n");
1428 dm
->state
= DM_INIT_ERROR
;
1430 complete(&dm
->host_event
);
1433 static void balloon_onchannelcallback(void *context
)
1435 struct hv_device
*dev
= context
;
1438 struct dm_message
*dm_msg
;
1439 struct dm_header
*dm_hdr
;
1440 struct hv_dynmem_device
*dm
= hv_get_drvdata(dev
);
1441 struct dm_balloon
*bal_msg
;
1442 struct dm_hot_add
*ha_msg
;
1443 union dm_mem_page_range
*ha_pg_range
;
1444 union dm_mem_page_range
*ha_region
;
1446 memset(recv_buffer
, 0, sizeof(recv_buffer
));
1447 vmbus_recvpacket(dev
->channel
, recv_buffer
,
1448 PAGE_SIZE
, &recvlen
, &requestid
);
1451 dm_msg
= (struct dm_message
*)recv_buffer
;
1452 dm_hdr
= &dm_msg
->hdr
;
1454 switch (dm_hdr
->type
) {
1455 case DM_VERSION_RESPONSE
:
1457 (struct dm_version_response
*)dm_msg
);
1460 case DM_CAPABILITIES_RESPONSE
:
1462 (struct dm_capabilities_resp_msg
*)dm_msg
);
1465 case DM_BALLOON_REQUEST
:
1466 if (dm
->state
== DM_BALLOON_UP
)
1467 pr_warn("Currently ballooning\n");
1468 bal_msg
= (struct dm_balloon
*)recv_buffer
;
1469 dm
->state
= DM_BALLOON_UP
;
1470 dm_device
.balloon_wrk
.num_pages
= bal_msg
->num_pages
;
1471 schedule_work(&dm_device
.balloon_wrk
.wrk
);
1474 case DM_UNBALLOON_REQUEST
:
1475 dm
->state
= DM_BALLOON_DOWN
;
1477 (struct dm_unballoon_request
*)recv_buffer
);
1480 case DM_MEM_HOT_ADD_REQUEST
:
1481 if (dm
->state
== DM_HOT_ADD
)
1482 pr_warn("Currently hot-adding\n");
1483 dm
->state
= DM_HOT_ADD
;
1484 ha_msg
= (struct dm_hot_add
*)recv_buffer
;
1485 if (ha_msg
->hdr
.size
== sizeof(struct dm_hot_add
)) {
1487 * This is a normal hot-add request specifying
1490 dm
->host_specified_ha_region
= false;
1491 ha_pg_range
= &ha_msg
->range
;
1492 dm
->ha_wrk
.ha_page_range
= *ha_pg_range
;
1493 dm
->ha_wrk
.ha_region_range
.page_range
= 0;
1496 * Host is specifying that we first hot-add
1497 * a region and then partially populate this
1500 dm
->host_specified_ha_region
= true;
1501 ha_pg_range
= &ha_msg
->range
;
1502 ha_region
= &ha_pg_range
[1];
1503 dm
->ha_wrk
.ha_page_range
= *ha_pg_range
;
1504 dm
->ha_wrk
.ha_region_range
= *ha_region
;
1506 schedule_work(&dm_device
.ha_wrk
.wrk
);
1509 case DM_INFO_MESSAGE
:
1510 process_info(dm
, (struct dm_info_msg
*)dm_msg
);
1514 pr_err("Unhandled message: type: %d\n", dm_hdr
->type
);
1521 static int balloon_probe(struct hv_device
*dev
,
1522 const struct hv_vmbus_device_id
*dev_id
)
1526 struct dm_version_request version_req
;
1527 struct dm_capabilities cap_msg
;
1529 #ifdef CONFIG_MEMORY_HOTPLUG
1530 do_hot_add
= hot_add
;
1536 * First allocate a send buffer.
1539 send_buffer
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
1543 ret
= vmbus_open(dev
->channel
, dm_ring_size
, dm_ring_size
, NULL
, 0,
1544 balloon_onchannelcallback
, dev
);
1549 dm_device
.dev
= dev
;
1550 dm_device
.state
= DM_INITIALIZING
;
1551 dm_device
.next_version
= DYNMEM_PROTOCOL_VERSION_WIN8
;
1552 init_completion(&dm_device
.host_event
);
1553 init_completion(&dm_device
.config_event
);
1554 INIT_LIST_HEAD(&dm_device
.ha_region_list
);
1555 spin_lock_init(&dm_device
.ha_lock
);
1556 INIT_WORK(&dm_device
.balloon_wrk
.wrk
, balloon_up
);
1557 INIT_WORK(&dm_device
.ha_wrk
.wrk
, hot_add_req
);
1558 dm_device
.host_specified_ha_region
= false;
1561 kthread_run(dm_thread_func
, &dm_device
, "hv_balloon");
1562 if (IS_ERR(dm_device
.thread
)) {
1563 ret
= PTR_ERR(dm_device
.thread
);
1567 #ifdef CONFIG_MEMORY_HOTPLUG
1568 set_online_page_callback(&hv_online_page
);
1569 register_memory_notifier(&hv_memory_nb
);
1572 hv_set_drvdata(dev
, &dm_device
);
1574 * Initiate the hand shake with the host and negotiate
1575 * a version that the host can support. We start with the
1576 * highest version number and go down if the host cannot
1579 memset(&version_req
, 0, sizeof(struct dm_version_request
));
1580 version_req
.hdr
.type
= DM_VERSION_REQUEST
;
1581 version_req
.hdr
.size
= sizeof(struct dm_version_request
);
1582 version_req
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1583 version_req
.version
.version
= DYNMEM_PROTOCOL_VERSION_WIN10
;
1584 version_req
.is_last_attempt
= 0;
1585 dm_device
.version
= version_req
.version
.version
;
1587 ret
= vmbus_sendpacket(dev
->channel
, &version_req
,
1588 sizeof(struct dm_version_request
),
1589 (unsigned long)NULL
,
1590 VM_PKT_DATA_INBAND
, 0);
1594 t
= wait_for_completion_timeout(&dm_device
.host_event
, 5*HZ
);
1601 * If we could not negotiate a compatible version with the host
1602 * fail the probe function.
1604 if (dm_device
.state
== DM_INIT_ERROR
) {
1609 pr_info("Using Dynamic Memory protocol version %u.%u\n",
1610 DYNMEM_MAJOR_VERSION(dm_device
.version
),
1611 DYNMEM_MINOR_VERSION(dm_device
.version
));
1614 * Now submit our capabilities to the host.
1616 memset(&cap_msg
, 0, sizeof(struct dm_capabilities
));
1617 cap_msg
.hdr
.type
= DM_CAPABILITIES_REPORT
;
1618 cap_msg
.hdr
.size
= sizeof(struct dm_capabilities
);
1619 cap_msg
.hdr
.trans_id
= atomic_inc_return(&trans_id
);
1621 cap_msg
.caps
.cap_bits
.balloon
= 1;
1622 cap_msg
.caps
.cap_bits
.hot_add
= 1;
1625 * Specify our alignment requirements as it relates
1626 * memory hot-add. Specify 128MB alignment.
1628 cap_msg
.caps
.cap_bits
.hot_add_alignment
= 7;
1631 * Currently the host does not use these
1632 * values and we set them to what is done in the
1635 cap_msg
.min_page_cnt
= 0;
1636 cap_msg
.max_page_number
= -1;
1638 ret
= vmbus_sendpacket(dev
->channel
, &cap_msg
,
1639 sizeof(struct dm_capabilities
),
1640 (unsigned long)NULL
,
1641 VM_PKT_DATA_INBAND
, 0);
1645 t
= wait_for_completion_timeout(&dm_device
.host_event
, 5*HZ
);
1652 * If the host does not like our capabilities,
1653 * fail the probe function.
1655 if (dm_device
.state
== DM_INIT_ERROR
) {
1660 dm_device
.state
= DM_INITIALIZED
;
1665 #ifdef CONFIG_MEMORY_HOTPLUG
1666 restore_online_page_callback(&hv_online_page
);
1668 kthread_stop(dm_device
.thread
);
1671 vmbus_close(dev
->channel
);
1677 static int balloon_remove(struct hv_device
*dev
)
1679 struct hv_dynmem_device
*dm
= hv_get_drvdata(dev
);
1680 struct hv_hotadd_state
*has
, *tmp
;
1681 struct hv_hotadd_gap
*gap
, *tmp_gap
;
1682 unsigned long flags
;
1684 if (dm
->num_pages_ballooned
!= 0)
1685 pr_warn("Ballooned pages: %d\n", dm
->num_pages_ballooned
);
1687 cancel_work_sync(&dm
->balloon_wrk
.wrk
);
1688 cancel_work_sync(&dm
->ha_wrk
.wrk
);
1690 vmbus_close(dev
->channel
);
1691 kthread_stop(dm
->thread
);
1693 #ifdef CONFIG_MEMORY_HOTPLUG
1694 restore_online_page_callback(&hv_online_page
);
1695 unregister_memory_notifier(&hv_memory_nb
);
1697 spin_lock_irqsave(&dm_device
.ha_lock
, flags
);
1698 list_for_each_entry_safe(has
, tmp
, &dm
->ha_region_list
, list
) {
1699 list_for_each_entry_safe(gap
, tmp_gap
, &has
->gap_list
, list
) {
1700 list_del(&gap
->list
);
1703 list_del(&has
->list
);
1706 spin_unlock_irqrestore(&dm_device
.ha_lock
, flags
);
1711 static const struct hv_vmbus_device_id id_table
[] = {
1712 /* Dynamic Memory Class ID */
1713 /* 525074DC-8985-46e2-8057-A307DC18A502 */
1718 MODULE_DEVICE_TABLE(vmbus
, id_table
);
1720 static struct hv_driver balloon_drv
= {
1721 .name
= "hv_balloon",
1722 .id_table
= id_table
,
1723 .probe
= balloon_probe
,
1724 .remove
= balloon_remove
,
1727 static int __init
init_balloon_drv(void)
1730 return vmbus_driver_register(&balloon_drv
);
1733 module_init(init_balloon_drv
);
1735 MODULE_DESCRIPTION("Hyper-V Balloon");
1736 MODULE_LICENSE("GPL");