Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (c) 2012, Microsoft Corporation.
4 *
5 * Author:
6 * K. Y. Srinivasan <kys@microsoft.com>
7 */
8
9#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11#include <linux/cleanup.h>
12#include <linux/kernel.h>
13#include <linux/jiffies.h>
14#include <linux/mman.h>
15#include <linux/debugfs.h>
16#include <linux/delay.h>
17#include <linux/init.h>
18#include <linux/module.h>
19#include <linux/slab.h>
20#include <linux/kthread.h>
21#include <linux/completion.h>
22#include <linux/count_zeros.h>
23#include <linux/memory_hotplug.h>
24#include <linux/memory.h>
25#include <linux/notifier.h>
26#include <linux/percpu_counter.h>
27#include <linux/page_reporting.h>
28
29#include <linux/hyperv.h>
30#include <asm/hyperv-tlfs.h>
31
32#include <asm/mshyperv.h>
33
34#define CREATE_TRACE_POINTS
35#include "hv_trace_balloon.h"
36
37/*
38 * We begin with definitions supporting the Dynamic Memory protocol
39 * with the host.
40 *
41 * Begin protocol definitions.
42 */
43
44
45
46/*
47 * Protocol versions. The low word is the minor version, the high word the major
48 * version.
49 *
50 * History:
51 * Initial version 1.0
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
56 */
57
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)
61
62enum {
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),
66
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,
70
71 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
72};
73
74
75
76/*
77 * Message Types
78 */
79
80enum dm_message_type {
81 /*
82 * Version 0.3
83 */
84 DM_ERROR = 0,
85 DM_VERSION_REQUEST = 1,
86 DM_VERSION_RESPONSE = 2,
87 DM_CAPABILITIES_REPORT = 3,
88 DM_CAPABILITIES_RESPONSE = 4,
89 DM_STATUS_REPORT = 5,
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,
97 /*
98 * Version 1.0.
99 */
100 DM_INFO_MESSAGE = 12,
101 DM_VERSION_1_MAX = 12
102};
103
104
105/*
106 * Structures defining the dynamic memory management
107 * protocol.
108 */
109
110union dm_version {
111 struct {
112 __u16 minor_version;
113 __u16 major_version;
114 };
115 __u32 version;
116} __packed;
117
118
119union dm_caps {
120 struct {
121 __u64 balloon:1;
122 __u64 hot_add:1;
123 /*
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.
128 */
129 __u64 hot_add_alignment:4;
130 __u64 reservedz:58;
131 } cap_bits;
132 __u64 caps;
133} __packed;
134
135union dm_mem_page_range {
136 struct {
137 /*
138 * The PFN number of the first page in the range.
139 * 40 bits is the architectural limit of a PFN
140 * number for AMD64.
141 */
142 __u64 start_page:40;
143 /*
144 * The number of pages in the range.
145 */
146 __u64 page_cnt:24;
147 } finfo;
148 __u64 page_range;
149} __packed;
150
151
152
153/*
154 * The header for all dynamic memory messages:
155 *
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.
159 */
160
161struct dm_header {
162 __u16 type;
163 __u16 size;
164 __u32 trans_id;
165} __packed;
166
167/*
168 * A generic message format for dynamic memory.
169 * Specific message formats are defined later in the file.
170 */
171
172struct dm_message {
173 struct dm_header hdr;
174 __u8 data[]; /* enclosed message */
175} __packed;
176
177
178/*
179 * Specific message types supporting the dynamic memory protocol.
180 */
181
182/*
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.
186 *
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.
190 */
191
192struct dm_version_request {
193 struct dm_header hdr;
194 union dm_version version;
195 __u32 is_last_attempt:1;
196 __u32 reservedz:31;
197} __packed;
198
199/*
200 * Version response message; Host to Guest and indicates
201 * if the host has accepted the version sent by the guest.
202 *
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.
206 *
207 * reservedz: Reserved field, set to zero.
208 */
209
210struct dm_version_response {
211 struct dm_header hdr;
212 __u64 is_accepted:1;
213 __u64 reservedz:63;
214} __packed;
215
216/*
217 * Message reporting capabilities. This is sent from the guest to the
218 * host.
219 */
220
221struct dm_capabilities {
222 struct dm_header hdr;
223 union dm_caps caps;
224 __u64 min_page_cnt;
225 __u64 max_page_number;
226} __packed;
227
228/*
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
232 * the service.
233 *
234 * is_accepted: Indicates if the host has accepted guest's capabilities.
235 * reservedz: Must be 0.
236 */
237
238struct dm_capabilities_resp_msg {
239 struct dm_header hdr;
240 __u64 is_accepted:1;
241 __u64 reservedz:63;
242} __packed;
243
244/*
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.
248 *
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 number 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.
258 *
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.
262 */
263
264struct dm_status {
265 struct dm_header hdr;
266 __u64 num_avail;
267 __u64 num_committed;
268 __u64 page_file_size;
269 __u64 zero_free;
270 __u32 page_file_writes;
271 __u32 io_diff;
272} __packed;
273
274
275/*
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.
279 *
280 * num_pages: number of pages to allocate.
281 */
282
283struct dm_balloon {
284 struct dm_header hdr;
285 __u32 num_pages;
286 __u32 reservedz;
287} __packed;
288
289
290/*
291 * Balloon response message; this message is sent from the guest
292 * to the host in response to the balloon message.
293 *
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.
297 *
298 * range_count: The number of ranges in the range array.
299 *
300 * range_array: An array of page ranges returned to the host.
301 *
302 */
303
304struct dm_balloon_response {
305 struct dm_header hdr;
306 __u32 reservedz;
307 __u32 more_pages:1;
308 __u32 range_count:31;
309 union dm_mem_page_range range_array[];
310} __packed;
311
312/*
313 * Un-balloon message; this message is sent from the host
314 * to the guest to give guest more memory.
315 *
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.
318 *
319 * reservedz: Reserved; must be set to zero.
320 *
321 * range_count: The number of ranges in the range array.
322 *
323 * range_array: An array of page ranges returned to the host.
324 *
325 */
326
327struct dm_unballoon_request {
328 struct dm_header hdr;
329 __u32 more_pages:1;
330 __u32 reservedz:31;
331 __u32 range_count;
332 union dm_mem_page_range range_array[];
333} __packed;
334
335/*
336 * Un-balloon response message; this message is sent from the guest
337 * to the host in response to an unballoon request.
338 *
339 */
340
341struct dm_unballoon_response {
342 struct dm_header hdr;
343} __packed;
344
345
346/*
347 * Hot add request message. Message sent from the host to the guest.
348 *
349 * mem_range: Memory range to hot add.
350 *
351 */
352
353struct dm_hot_add {
354 struct dm_header hdr;
355 union dm_mem_page_range range;
356} __packed;
357
358/*
359 * Hot add response message.
360 * This message is sent by the guest to report the status of a hot add request.
361 * If page_count is less than the requested page count, then the host should
362 * assume all further hot add requests will fail, since this indicates that
363 * the guest has hit an upper physical memory barrier.
364 *
365 * Hot adds may also fail due to low resources; in this case, the guest must
366 * not complete this message until the hot add can succeed, and the host must
367 * not send a new hot add request until the response is sent.
368 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
369 * times it fails the request.
370 *
371 *
372 * page_count: number of pages that were successfully hot added.
373 *
374 * result: result of the operation 1: success, 0: failure.
375 *
376 */
377
378struct dm_hot_add_response {
379 struct dm_header hdr;
380 __u32 page_count;
381 __u32 result;
382} __packed;
383
384/*
385 * Types of information sent from host to the guest.
386 */
387
388enum dm_info_type {
389 INFO_TYPE_MAX_PAGE_CNT = 0,
390 MAX_INFO_TYPE
391};
392
393
394/*
395 * Header for the information message.
396 */
397
398struct dm_info_header {
399 enum dm_info_type type;
400 __u32 data_size;
401} __packed;
402
403/*
404 * This message is sent from the host to the guest to pass
405 * some relevant information (win8 addition).
406 *
407 * reserved: no used.
408 * info_size: size of the information blob.
409 * info: information blob.
410 */
411
412struct dm_info_msg {
413 struct dm_header hdr;
414 __u32 reserved;
415 __u32 info_size;
416 __u8 info[];
417};
418
419/*
420 * End protocol definitions.
421 */
422
423/*
424 * State to manage hot adding memory into the guest.
425 * The range start_pfn : end_pfn specifies the range
426 * that the host has asked us to hot add. The range
427 * start_pfn : ha_end_pfn specifies the range that we have
428 * currently hot added. We hot add in multiples of 128M
429 * chunks; it is possible that we may not be able to bring
430 * online all the pages in the region. The range
431 * covered_start_pfn:covered_end_pfn defines the pages that can
432 * be brough online.
433 */
434
435struct hv_hotadd_state {
436 struct list_head list;
437 unsigned long start_pfn;
438 unsigned long covered_start_pfn;
439 unsigned long covered_end_pfn;
440 unsigned long ha_end_pfn;
441 unsigned long end_pfn;
442 /*
443 * A list of gaps.
444 */
445 struct list_head gap_list;
446};
447
448struct hv_hotadd_gap {
449 struct list_head list;
450 unsigned long start_pfn;
451 unsigned long end_pfn;
452};
453
454struct balloon_state {
455 __u32 num_pages;
456 struct work_struct wrk;
457};
458
459struct hot_add_wrk {
460 union dm_mem_page_range ha_page_range;
461 union dm_mem_page_range ha_region_range;
462 struct work_struct wrk;
463};
464
465static bool allow_hibernation;
466static bool hot_add = true;
467static bool do_hot_add;
468/*
469 * Delay reporting memory pressure by
470 * the specified number of seconds.
471 */
472static uint pressure_report_delay = 45;
473extern unsigned int page_reporting_order;
474#define HV_MAX_FAILURES 2
475
476/*
477 * The last time we posted a pressure report to host.
478 */
479static unsigned long last_post_time;
480
481static int hv_hypercall_multi_failure;
482
483module_param(hot_add, bool, (S_IRUGO | S_IWUSR));
484MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
485
486module_param(pressure_report_delay, uint, (S_IRUGO | S_IWUSR));
487MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
488static atomic_t trans_id = ATOMIC_INIT(0);
489
490static int dm_ring_size = VMBUS_RING_SIZE(16 * 1024);
491
492/*
493 * Driver specific state.
494 */
495
496enum hv_dm_state {
497 DM_INITIALIZING = 0,
498 DM_INITIALIZED,
499 DM_BALLOON_UP,
500 DM_BALLOON_DOWN,
501 DM_HOT_ADD,
502 DM_INIT_ERROR
503};
504
505
506static __u8 recv_buffer[HV_HYP_PAGE_SIZE];
507static __u8 balloon_up_send_buffer[HV_HYP_PAGE_SIZE];
508#define PAGES_IN_2M (2 * 1024 * 1024 / PAGE_SIZE)
509#define HA_CHUNK (128 * 1024 * 1024 / PAGE_SIZE)
510
511struct hv_dynmem_device {
512 struct hv_device *dev;
513 enum hv_dm_state state;
514 struct completion host_event;
515 struct completion config_event;
516
517 /*
518 * Number of pages we have currently ballooned out.
519 */
520 unsigned int num_pages_ballooned;
521 unsigned int num_pages_onlined;
522 unsigned int num_pages_added;
523
524 /*
525 * State to manage the ballooning (up) operation.
526 */
527 struct balloon_state balloon_wrk;
528
529 /*
530 * State to execute the "hot-add" operation.
531 */
532 struct hot_add_wrk ha_wrk;
533
534 /*
535 * This state tracks if the host has specified a hot-add
536 * region.
537 */
538 bool host_specified_ha_region;
539
540 /*
541 * State to synchronize hot-add.
542 */
543 struct completion ol_waitevent;
544 /*
545 * This thread handles hot-add
546 * requests from the host as well as notifying
547 * the host with regards to memory pressure in
548 * the guest.
549 */
550 struct task_struct *thread;
551
552 /*
553 * Protects ha_region_list, num_pages_onlined counter and individual
554 * regions from ha_region_list.
555 */
556 spinlock_t ha_lock;
557
558 /*
559 * A list of hot-add regions.
560 */
561 struct list_head ha_region_list;
562
563 /*
564 * We start with the highest version we can support
565 * and downgrade based on the host; we save here the
566 * next version to try.
567 */
568 __u32 next_version;
569
570 /*
571 * The negotiated version agreed by host.
572 */
573 __u32 version;
574
575 struct page_reporting_dev_info pr_dev_info;
576
577 /*
578 * Maximum number of pages that can be hot_add-ed
579 */
580 __u64 max_dynamic_page_count;
581};
582
583static struct hv_dynmem_device dm_device;
584
585static void post_status(struct hv_dynmem_device *dm);
586
587static void enable_page_reporting(void);
588
589static void disable_page_reporting(void);
590
591#ifdef CONFIG_MEMORY_HOTPLUG
592static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
593 unsigned long pfn)
594{
595 struct hv_hotadd_gap *gap;
596
597 /* The page is not backed. */
598 if ((pfn < has->covered_start_pfn) || (pfn >= has->covered_end_pfn))
599 return false;
600
601 /* Check for gaps. */
602 list_for_each_entry(gap, &has->gap_list, list) {
603 if ((pfn >= gap->start_pfn) && (pfn < gap->end_pfn))
604 return false;
605 }
606
607 return true;
608}
609
610static unsigned long hv_page_offline_check(unsigned long start_pfn,
611 unsigned long nr_pages)
612{
613 unsigned long pfn = start_pfn, count = 0;
614 struct hv_hotadd_state *has;
615 bool found;
616
617 while (pfn < start_pfn + nr_pages) {
618 /*
619 * Search for HAS which covers the pfn and when we find one
620 * count how many consequitive PFNs are covered.
621 */
622 found = false;
623 list_for_each_entry(has, &dm_device.ha_region_list, list) {
624 while ((pfn >= has->start_pfn) &&
625 (pfn < has->end_pfn) &&
626 (pfn < start_pfn + nr_pages)) {
627 found = true;
628 if (has_pfn_is_backed(has, pfn))
629 count++;
630 pfn++;
631 }
632 }
633
634 /*
635 * This PFN is not in any HAS (e.g. we're offlining a region
636 * which was present at boot), no need to account for it. Go
637 * to the next one.
638 */
639 if (!found)
640 pfn++;
641 }
642
643 return count;
644}
645
646static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
647 void *v)
648{
649 struct memory_notify *mem = (struct memory_notify *)v;
650 unsigned long pfn_count;
651
652 switch (val) {
653 case MEM_ONLINE:
654 case MEM_CANCEL_ONLINE:
655 complete(&dm_device.ol_waitevent);
656 break;
657
658 case MEM_OFFLINE:
659 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
660 pfn_count = hv_page_offline_check(mem->start_pfn,
661 mem->nr_pages);
662 if (pfn_count <= dm_device.num_pages_onlined) {
663 dm_device.num_pages_onlined -= pfn_count;
664 } else {
665 /*
666 * We're offlining more pages than we
667 * managed to online. This is
668 * unexpected. In any case don't let
669 * num_pages_onlined wrap around zero.
670 */
671 WARN_ON_ONCE(1);
672 dm_device.num_pages_onlined = 0;
673 }
674 }
675 break;
676 case MEM_GOING_ONLINE:
677 case MEM_GOING_OFFLINE:
678 case MEM_CANCEL_OFFLINE:
679 break;
680 }
681 return NOTIFY_OK;
682}
683
684static struct notifier_block hv_memory_nb = {
685 .notifier_call = hv_memory_notifier,
686 .priority = 0
687};
688
689/* Check if the particular page is backed and can be onlined and online it. */
690static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
691{
692 if (!has_pfn_is_backed(has, page_to_pfn(pg))) {
693 if (!PageOffline(pg))
694 __SetPageOffline(pg);
695 return;
696 }
697 if (PageOffline(pg))
698 __ClearPageOffline(pg);
699
700 /* This frame is currently backed; online the page. */
701 generic_online_page(pg, 0);
702
703 lockdep_assert_held(&dm_device.ha_lock);
704 dm_device.num_pages_onlined++;
705}
706
707static void hv_bring_pgs_online(struct hv_hotadd_state *has,
708 unsigned long start_pfn, unsigned long size)
709{
710 int i;
711
712 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
713 for (i = 0; i < size; i++)
714 hv_page_online_one(has, pfn_to_page(start_pfn + i));
715}
716
717static void hv_mem_hot_add(unsigned long start, unsigned long size,
718 unsigned long pfn_count,
719 struct hv_hotadd_state *has)
720{
721 int ret = 0;
722 int i, nid;
723 unsigned long start_pfn;
724 unsigned long processed_pfn;
725 unsigned long total_pfn = pfn_count;
726
727 for (i = 0; i < (size/HA_CHUNK); i++) {
728 start_pfn = start + (i * HA_CHUNK);
729
730 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
731 has->ha_end_pfn += HA_CHUNK;
732
733 if (total_pfn > HA_CHUNK) {
734 processed_pfn = HA_CHUNK;
735 total_pfn -= HA_CHUNK;
736 } else {
737 processed_pfn = total_pfn;
738 total_pfn = 0;
739 }
740
741 has->covered_end_pfn += processed_pfn;
742 }
743
744 reinit_completion(&dm_device.ol_waitevent);
745
746 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
747 ret = add_memory(nid, PFN_PHYS((start_pfn)),
748 (HA_CHUNK << PAGE_SHIFT), MHP_MERGE_RESOURCE);
749
750 if (ret) {
751 pr_err("hot_add memory failed error is %d\n", ret);
752 if (ret == -EEXIST) {
753 /*
754 * This error indicates that the error
755 * is not a transient failure. This is the
756 * case where the guest's physical address map
757 * precludes hot adding memory. Stop all further
758 * memory hot-add.
759 */
760 do_hot_add = false;
761 }
762 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
763 has->ha_end_pfn -= HA_CHUNK;
764 has->covered_end_pfn -= processed_pfn;
765 }
766 break;
767 }
768
769 /*
770 * Wait for memory to get onlined. If the kernel onlined the
771 * memory when adding it, this will return directly. Otherwise,
772 * it will wait for user space to online the memory. This helps
773 * to avoid adding memory faster than it is getting onlined. As
774 * adding succeeded, it is ok to proceed even if the memory was
775 * not onlined in time.
776 */
777 wait_for_completion_timeout(&dm_device.ol_waitevent, 5 * HZ);
778 post_status(&dm_device);
779 }
780}
781
782static void hv_online_page(struct page *pg, unsigned int order)
783{
784 struct hv_hotadd_state *has;
785 unsigned long pfn = page_to_pfn(pg);
786
787 guard(spinlock_irqsave)(&dm_device.ha_lock);
788 list_for_each_entry(has, &dm_device.ha_region_list, list) {
789 /* The page belongs to a different HAS. */
790 if ((pfn < has->start_pfn) ||
791 (pfn + (1UL << order) > has->end_pfn))
792 continue;
793
794 hv_bring_pgs_online(has, pfn, 1UL << order);
795 break;
796 }
797}
798
799static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
800{
801 struct hv_hotadd_state *has;
802 struct hv_hotadd_gap *gap;
803 unsigned long residual, new_inc;
804 int ret = 0;
805
806 guard(spinlock_irqsave)(&dm_device.ha_lock);
807 list_for_each_entry(has, &dm_device.ha_region_list, list) {
808 /*
809 * If the pfn range we are dealing with is not in the current
810 * "hot add block", move on.
811 */
812 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
813 continue;
814
815 /*
816 * If the current start pfn is not where the covered_end
817 * is, create a gap and update covered_end_pfn.
818 */
819 if (has->covered_end_pfn != start_pfn) {
820 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
821 if (!gap) {
822 ret = -ENOMEM;
823 break;
824 }
825
826 INIT_LIST_HEAD(&gap->list);
827 gap->start_pfn = has->covered_end_pfn;
828 gap->end_pfn = start_pfn;
829 list_add_tail(&gap->list, &has->gap_list);
830
831 has->covered_end_pfn = start_pfn;
832 }
833
834 /*
835 * If the current hot add-request extends beyond
836 * our current limit; extend it.
837 */
838 if ((start_pfn + pfn_cnt) > has->end_pfn) {
839 residual = (start_pfn + pfn_cnt - has->end_pfn);
840 /*
841 * Extend the region by multiples of HA_CHUNK.
842 */
843 new_inc = (residual / HA_CHUNK) * HA_CHUNK;
844 if (residual % HA_CHUNK)
845 new_inc += HA_CHUNK;
846
847 has->end_pfn += new_inc;
848 }
849
850 ret = 1;
851 break;
852 }
853
854 return ret;
855}
856
857static unsigned long handle_pg_range(unsigned long pg_start,
858 unsigned long pg_count)
859{
860 unsigned long start_pfn = pg_start;
861 unsigned long pfn_cnt = pg_count;
862 unsigned long size;
863 struct hv_hotadd_state *has;
864 unsigned long pgs_ol = 0;
865 unsigned long old_covered_state;
866 unsigned long res = 0, flags;
867
868 pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
869 pg_start);
870
871 spin_lock_irqsave(&dm_device.ha_lock, flags);
872 list_for_each_entry(has, &dm_device.ha_region_list, list) {
873 /*
874 * If the pfn range we are dealing with is not in the current
875 * "hot add block", move on.
876 */
877 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
878 continue;
879
880 old_covered_state = has->covered_end_pfn;
881
882 if (start_pfn < has->ha_end_pfn) {
883 /*
884 * This is the case where we are backing pages
885 * in an already hot added region. Bring
886 * these pages online first.
887 */
888 pgs_ol = has->ha_end_pfn - start_pfn;
889 if (pgs_ol > pfn_cnt)
890 pgs_ol = pfn_cnt;
891
892 has->covered_end_pfn += pgs_ol;
893 pfn_cnt -= pgs_ol;
894 /*
895 * Check if the corresponding memory block is already
896 * online. It is possible to observe struct pages still
897 * being uninitialized here so check section instead.
898 * In case the section is online we need to bring the
899 * rest of pfns (which were not backed previously)
900 * online too.
901 */
902 if (start_pfn > has->start_pfn &&
903 online_section_nr(pfn_to_section_nr(start_pfn)))
904 hv_bring_pgs_online(has, start_pfn, pgs_ol);
905
906 }
907
908 if ((has->ha_end_pfn < has->end_pfn) && (pfn_cnt > 0)) {
909 /*
910 * We have some residual hot add range
911 * that needs to be hot added; hot add
912 * it now. Hot add a multiple of
913 * HA_CHUNK that fully covers the pages
914 * we have.
915 */
916 size = (has->end_pfn - has->ha_end_pfn);
917 if (pfn_cnt <= size) {
918 size = ((pfn_cnt / HA_CHUNK) * HA_CHUNK);
919 if (pfn_cnt % HA_CHUNK)
920 size += HA_CHUNK;
921 } else {
922 pfn_cnt = size;
923 }
924 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
925 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
926 spin_lock_irqsave(&dm_device.ha_lock, flags);
927 }
928 /*
929 * If we managed to online any pages that were given to us,
930 * we declare success.
931 */
932 res = has->covered_end_pfn - old_covered_state;
933 break;
934 }
935 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
936
937 return res;
938}
939
940static unsigned long process_hot_add(unsigned long pg_start,
941 unsigned long pfn_cnt,
942 unsigned long rg_start,
943 unsigned long rg_size)
944{
945 struct hv_hotadd_state *ha_region = NULL;
946 int covered;
947
948 if (pfn_cnt == 0)
949 return 0;
950
951 if (!dm_device.host_specified_ha_region) {
952 covered = pfn_covered(pg_start, pfn_cnt);
953 if (covered < 0)
954 return 0;
955
956 if (covered)
957 goto do_pg_range;
958 }
959
960 /*
961 * If the host has specified a hot-add range; deal with it first.
962 */
963
964 if (rg_size != 0) {
965 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
966 if (!ha_region)
967 return 0;
968
969 INIT_LIST_HEAD(&ha_region->list);
970 INIT_LIST_HEAD(&ha_region->gap_list);
971
972 ha_region->start_pfn = rg_start;
973 ha_region->ha_end_pfn = rg_start;
974 ha_region->covered_start_pfn = pg_start;
975 ha_region->covered_end_pfn = pg_start;
976 ha_region->end_pfn = rg_start + rg_size;
977
978 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
979 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
980 }
981 }
982
983do_pg_range:
984 /*
985 * Process the page range specified; bringing them
986 * online if possible.
987 */
988 return handle_pg_range(pg_start, pfn_cnt);
989}
990
991#endif
992
993static void hot_add_req(struct work_struct *dummy)
994{
995 struct dm_hot_add_response resp;
996#ifdef CONFIG_MEMORY_HOTPLUG
997 unsigned long pg_start, pfn_cnt;
998 unsigned long rg_start, rg_sz;
999#endif
1000 struct hv_dynmem_device *dm = &dm_device;
1001
1002 memset(&resp, 0, sizeof(struct dm_hot_add_response));
1003 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
1004 resp.hdr.size = sizeof(struct dm_hot_add_response);
1005
1006#ifdef CONFIG_MEMORY_HOTPLUG
1007 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
1008 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
1009
1010 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
1011 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
1012
1013 if ((rg_start == 0) && (!dm->host_specified_ha_region)) {
1014 unsigned long region_size;
1015 unsigned long region_start;
1016
1017 /*
1018 * The host has not specified the hot-add region.
1019 * Based on the hot-add page range being specified,
1020 * compute a hot-add region that can cover the pages
1021 * that need to be hot-added while ensuring the alignment
1022 * and size requirements of Linux as it relates to hot-add.
1023 */
1024 region_size = (pfn_cnt / HA_CHUNK) * HA_CHUNK;
1025 if (pfn_cnt % HA_CHUNK)
1026 region_size += HA_CHUNK;
1027
1028 region_start = (pg_start / HA_CHUNK) * HA_CHUNK;
1029
1030 rg_start = region_start;
1031 rg_sz = region_size;
1032 }
1033
1034 if (do_hot_add)
1035 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1036 rg_start, rg_sz);
1037
1038 dm->num_pages_added += resp.page_count;
1039#endif
1040 /*
1041 * The result field of the response structure has the
1042 * following semantics:
1043 *
1044 * 1. If all or some pages hot-added: Guest should return success.
1045 *
1046 * 2. If no pages could be hot-added:
1047 *
1048 * If the guest returns success, then the host
1049 * will not attempt any further hot-add operations. This
1050 * signifies a permanent failure.
1051 *
1052 * If the guest returns failure, then this failure will be
1053 * treated as a transient failure and the host may retry the
1054 * hot-add operation after some delay.
1055 */
1056 if (resp.page_count > 0)
1057 resp.result = 1;
1058 else if (!do_hot_add)
1059 resp.result = 1;
1060 else
1061 resp.result = 0;
1062
1063 if (!do_hot_add || resp.page_count == 0) {
1064 if (!allow_hibernation)
1065 pr_err("Memory hot add failed\n");
1066 else
1067 pr_info("Ignore hot-add request!\n");
1068 }
1069
1070 dm->state = DM_INITIALIZED;
1071 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1072 vmbus_sendpacket(dm->dev->channel, &resp,
1073 sizeof(struct dm_hot_add_response),
1074 (unsigned long)NULL,
1075 VM_PKT_DATA_INBAND, 0);
1076}
1077
1078static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1079{
1080 struct dm_info_header *info_hdr;
1081
1082 info_hdr = (struct dm_info_header *)msg->info;
1083
1084 switch (info_hdr->type) {
1085 case INFO_TYPE_MAX_PAGE_CNT:
1086 if (info_hdr->data_size == sizeof(__u64)) {
1087 __u64 *max_page_count = (__u64 *)&info_hdr[1];
1088
1089 pr_info("Max. dynamic memory size: %llu MB\n",
1090 (*max_page_count) >> (20 - HV_HYP_PAGE_SHIFT));
1091 dm->max_dynamic_page_count = *max_page_count;
1092 }
1093
1094 break;
1095 default:
1096 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1097 }
1098}
1099
1100static unsigned long compute_balloon_floor(void)
1101{
1102 unsigned long min_pages;
1103 unsigned long nr_pages = totalram_pages();
1104#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1105 /* Simple continuous piecewiese linear function:
1106 * max MiB -> min MiB gradient
1107 * 0 0
1108 * 16 16
1109 * 32 24
1110 * 128 72 (1/2)
1111 * 512 168 (1/4)
1112 * 2048 360 (1/8)
1113 * 8192 744 (1/16)
1114 * 32768 1512 (1/32)
1115 */
1116 if (nr_pages < MB2PAGES(128))
1117 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1118 else if (nr_pages < MB2PAGES(512))
1119 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1120 else if (nr_pages < MB2PAGES(2048))
1121 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1122 else if (nr_pages < MB2PAGES(8192))
1123 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1124 else
1125 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1126#undef MB2PAGES
1127 return min_pages;
1128}
1129
1130/*
1131 * Compute total committed memory pages
1132 */
1133
1134static unsigned long get_pages_committed(struct hv_dynmem_device *dm)
1135{
1136 return vm_memory_committed() +
1137 dm->num_pages_ballooned +
1138 (dm->num_pages_added > dm->num_pages_onlined ?
1139 dm->num_pages_added - dm->num_pages_onlined : 0) +
1140 compute_balloon_floor();
1141}
1142
1143/*
1144 * Post our status as it relates memory pressure to the
1145 * host. Host expects the guests to post this status
1146 * periodically at 1 second intervals.
1147 *
1148 * The metrics specified in this protocol are very Windows
1149 * specific and so we cook up numbers here to convey our memory
1150 * pressure.
1151 */
1152
1153static void post_status(struct hv_dynmem_device *dm)
1154{
1155 struct dm_status status;
1156 unsigned long now = jiffies;
1157 unsigned long last_post = last_post_time;
1158 unsigned long num_pages_avail, num_pages_committed;
1159
1160 if (pressure_report_delay > 0) {
1161 --pressure_report_delay;
1162 return;
1163 }
1164
1165 if (!time_after(now, (last_post_time + HZ)))
1166 return;
1167
1168 memset(&status, 0, sizeof(struct dm_status));
1169 status.hdr.type = DM_STATUS_REPORT;
1170 status.hdr.size = sizeof(struct dm_status);
1171 status.hdr.trans_id = atomic_inc_return(&trans_id);
1172
1173 /*
1174 * The host expects the guest to report free and committed memory.
1175 * Furthermore, the host expects the pressure information to include
1176 * the ballooned out pages. For a given amount of memory that we are
1177 * managing we need to compute a floor below which we should not
1178 * balloon. Compute this and add it to the pressure report.
1179 * We also need to report all offline pages (num_pages_added -
1180 * num_pages_onlined) as committed to the host, otherwise it can try
1181 * asking us to balloon them out.
1182 */
1183 num_pages_avail = si_mem_available();
1184 num_pages_committed = get_pages_committed(dm);
1185
1186 trace_balloon_status(num_pages_avail, num_pages_committed,
1187 vm_memory_committed(), dm->num_pages_ballooned,
1188 dm->num_pages_added, dm->num_pages_onlined);
1189
1190 /* Convert numbers of pages into numbers of HV_HYP_PAGEs. */
1191 status.num_avail = num_pages_avail * NR_HV_HYP_PAGES_IN_PAGE;
1192 status.num_committed = num_pages_committed * NR_HV_HYP_PAGES_IN_PAGE;
1193
1194 /*
1195 * If our transaction ID is no longer current, just don't
1196 * send the status. This can happen if we were interrupted
1197 * after we picked our transaction ID.
1198 */
1199 if (status.hdr.trans_id != atomic_read(&trans_id))
1200 return;
1201
1202 /*
1203 * If the last post time that we sampled has changed,
1204 * we have raced, don't post the status.
1205 */
1206 if (last_post != last_post_time)
1207 return;
1208
1209 last_post_time = jiffies;
1210 vmbus_sendpacket(dm->dev->channel, &status,
1211 sizeof(struct dm_status),
1212 (unsigned long)NULL,
1213 VM_PKT_DATA_INBAND, 0);
1214
1215}
1216
1217static void free_balloon_pages(struct hv_dynmem_device *dm,
1218 union dm_mem_page_range *range_array)
1219{
1220 int num_pages = range_array->finfo.page_cnt;
1221 __u64 start_frame = range_array->finfo.start_page;
1222 struct page *pg;
1223 int i;
1224
1225 for (i = 0; i < num_pages; i++) {
1226 pg = pfn_to_page(i + start_frame);
1227 __ClearPageOffline(pg);
1228 __free_page(pg);
1229 dm->num_pages_ballooned--;
1230 adjust_managed_page_count(pg, 1);
1231 }
1232}
1233
1234
1235
1236static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1237 unsigned int num_pages,
1238 struct dm_balloon_response *bl_resp,
1239 int alloc_unit)
1240{
1241 unsigned int i, j;
1242 struct page *pg;
1243
1244 for (i = 0; i < num_pages / alloc_unit; i++) {
1245 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1246 HV_HYP_PAGE_SIZE)
1247 return i * alloc_unit;
1248
1249 /*
1250 * We execute this code in a thread context. Furthermore,
1251 * we don't want the kernel to try too hard.
1252 */
1253 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1254 __GFP_NOMEMALLOC | __GFP_NOWARN,
1255 get_order(alloc_unit << PAGE_SHIFT));
1256
1257 if (!pg)
1258 return i * alloc_unit;
1259
1260 dm->num_pages_ballooned += alloc_unit;
1261
1262 /*
1263 * If we allocatted 2M pages; split them so we
1264 * can free them in any order we get.
1265 */
1266
1267 if (alloc_unit != 1)
1268 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1269
1270 /* mark all pages offline */
1271 for (j = 0; j < alloc_unit; j++) {
1272 __SetPageOffline(pg + j);
1273 adjust_managed_page_count(pg + j, -1);
1274 }
1275
1276 bl_resp->range_count++;
1277 bl_resp->range_array[i].finfo.start_page =
1278 page_to_pfn(pg);
1279 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1280 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1281
1282 }
1283
1284 return i * alloc_unit;
1285}
1286
1287static void balloon_up(struct work_struct *dummy)
1288{
1289 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1290 unsigned int num_ballooned = 0;
1291 struct dm_balloon_response *bl_resp;
1292 int alloc_unit;
1293 int ret;
1294 bool done = false;
1295 int i;
1296 long avail_pages;
1297 unsigned long floor;
1298
1299 /*
1300 * We will attempt 2M allocations. However, if we fail to
1301 * allocate 2M chunks, we will go back to PAGE_SIZE allocations.
1302 */
1303 alloc_unit = PAGES_IN_2M;
1304
1305 avail_pages = si_mem_available();
1306 floor = compute_balloon_floor();
1307
1308 /* Refuse to balloon below the floor. */
1309 if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1310 pr_info("Balloon request will be partially fulfilled. %s\n",
1311 avail_pages < num_pages ? "Not enough memory." :
1312 "Balloon floor reached.");
1313
1314 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1315 }
1316
1317 while (!done) {
1318 memset(balloon_up_send_buffer, 0, HV_HYP_PAGE_SIZE);
1319 bl_resp = (struct dm_balloon_response *)balloon_up_send_buffer;
1320 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1321 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1322 bl_resp->more_pages = 1;
1323
1324 num_pages -= num_ballooned;
1325 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1326 bl_resp, alloc_unit);
1327
1328 if (alloc_unit != 1 && num_ballooned == 0) {
1329 alloc_unit = 1;
1330 continue;
1331 }
1332
1333 if (num_ballooned == 0 || num_ballooned == num_pages) {
1334 pr_debug("Ballooned %u out of %u requested pages.\n",
1335 num_pages, dm_device.balloon_wrk.num_pages);
1336
1337 bl_resp->more_pages = 0;
1338 done = true;
1339 dm_device.state = DM_INITIALIZED;
1340 }
1341
1342 /*
1343 * We are pushing a lot of data through the channel;
1344 * deal with transient failures caused because of the
1345 * lack of space in the ring buffer.
1346 */
1347
1348 do {
1349 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1350 ret = vmbus_sendpacket(dm_device.dev->channel,
1351 bl_resp,
1352 bl_resp->hdr.size,
1353 (unsigned long)NULL,
1354 VM_PKT_DATA_INBAND, 0);
1355
1356 if (ret == -EAGAIN)
1357 msleep(20);
1358 post_status(&dm_device);
1359 } while (ret == -EAGAIN);
1360
1361 if (ret) {
1362 /*
1363 * Free up the memory we allocatted.
1364 */
1365 pr_err("Balloon response failed\n");
1366
1367 for (i = 0; i < bl_resp->range_count; i++)
1368 free_balloon_pages(&dm_device,
1369 &bl_resp->range_array[i]);
1370
1371 done = true;
1372 }
1373 }
1374
1375}
1376
1377static void balloon_down(struct hv_dynmem_device *dm,
1378 struct dm_unballoon_request *req)
1379{
1380 union dm_mem_page_range *range_array = req->range_array;
1381 int range_count = req->range_count;
1382 struct dm_unballoon_response resp;
1383 int i;
1384 unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1385
1386 for (i = 0; i < range_count; i++) {
1387 free_balloon_pages(dm, &range_array[i]);
1388 complete(&dm_device.config_event);
1389 }
1390
1391 pr_debug("Freed %u ballooned pages.\n",
1392 prev_pages_ballooned - dm->num_pages_ballooned);
1393
1394 if (req->more_pages == 1)
1395 return;
1396
1397 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1398 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1399 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1400 resp.hdr.size = sizeof(struct dm_unballoon_response);
1401
1402 vmbus_sendpacket(dm_device.dev->channel, &resp,
1403 sizeof(struct dm_unballoon_response),
1404 (unsigned long)NULL,
1405 VM_PKT_DATA_INBAND, 0);
1406
1407 dm->state = DM_INITIALIZED;
1408}
1409
1410static void balloon_onchannelcallback(void *context);
1411
1412static int dm_thread_func(void *dm_dev)
1413{
1414 struct hv_dynmem_device *dm = dm_dev;
1415
1416 while (!kthread_should_stop()) {
1417 wait_for_completion_interruptible_timeout(
1418 &dm_device.config_event, 1*HZ);
1419 /*
1420 * The host expects us to post information on the memory
1421 * pressure every second.
1422 */
1423 reinit_completion(&dm_device.config_event);
1424 post_status(dm);
1425 /*
1426 * disable free page reporting if multiple hypercall
1427 * failure flag set. It is not done in the page_reporting
1428 * callback context as that causes a deadlock between
1429 * page_reporting_process() and page_reporting_unregister()
1430 */
1431 if (hv_hypercall_multi_failure >= HV_MAX_FAILURES) {
1432 pr_err("Multiple failures in cold memory discard hypercall, disabling page reporting\n");
1433 disable_page_reporting();
1434 /* Reset the flag after disabling reporting */
1435 hv_hypercall_multi_failure = 0;
1436 }
1437 }
1438
1439 return 0;
1440}
1441
1442
1443static void version_resp(struct hv_dynmem_device *dm,
1444 struct dm_version_response *vresp)
1445{
1446 struct dm_version_request version_req;
1447 int ret;
1448
1449 if (vresp->is_accepted) {
1450 /*
1451 * We are done; wakeup the
1452 * context waiting for version
1453 * negotiation.
1454 */
1455 complete(&dm->host_event);
1456 return;
1457 }
1458 /*
1459 * If there are more versions to try, continue
1460 * with negotiations; if not
1461 * shutdown the service since we are not able
1462 * to negotiate a suitable version number
1463 * with the host.
1464 */
1465 if (dm->next_version == 0)
1466 goto version_error;
1467
1468 memset(&version_req, 0, sizeof(struct dm_version_request));
1469 version_req.hdr.type = DM_VERSION_REQUEST;
1470 version_req.hdr.size = sizeof(struct dm_version_request);
1471 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1472 version_req.version.version = dm->next_version;
1473 dm->version = version_req.version.version;
1474
1475 /*
1476 * Set the next version to try in case current version fails.
1477 * Win7 protocol ought to be the last one to try.
1478 */
1479 switch (version_req.version.version) {
1480 case DYNMEM_PROTOCOL_VERSION_WIN8:
1481 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1482 version_req.is_last_attempt = 0;
1483 break;
1484 default:
1485 dm->next_version = 0;
1486 version_req.is_last_attempt = 1;
1487 }
1488
1489 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1490 sizeof(struct dm_version_request),
1491 (unsigned long)NULL,
1492 VM_PKT_DATA_INBAND, 0);
1493
1494 if (ret)
1495 goto version_error;
1496
1497 return;
1498
1499version_error:
1500 dm->state = DM_INIT_ERROR;
1501 complete(&dm->host_event);
1502}
1503
1504static void cap_resp(struct hv_dynmem_device *dm,
1505 struct dm_capabilities_resp_msg *cap_resp)
1506{
1507 if (!cap_resp->is_accepted) {
1508 pr_err("Capabilities not accepted by host\n");
1509 dm->state = DM_INIT_ERROR;
1510 }
1511 complete(&dm->host_event);
1512}
1513
1514static void balloon_onchannelcallback(void *context)
1515{
1516 struct hv_device *dev = context;
1517 u32 recvlen;
1518 u64 requestid;
1519 struct dm_message *dm_msg;
1520 struct dm_header *dm_hdr;
1521 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1522 struct dm_balloon *bal_msg;
1523 struct dm_hot_add *ha_msg;
1524 union dm_mem_page_range *ha_pg_range;
1525 union dm_mem_page_range *ha_region;
1526
1527 memset(recv_buffer, 0, sizeof(recv_buffer));
1528 vmbus_recvpacket(dev->channel, recv_buffer,
1529 HV_HYP_PAGE_SIZE, &recvlen, &requestid);
1530
1531 if (recvlen > 0) {
1532 dm_msg = (struct dm_message *)recv_buffer;
1533 dm_hdr = &dm_msg->hdr;
1534
1535 switch (dm_hdr->type) {
1536 case DM_VERSION_RESPONSE:
1537 version_resp(dm,
1538 (struct dm_version_response *)dm_msg);
1539 break;
1540
1541 case DM_CAPABILITIES_RESPONSE:
1542 cap_resp(dm,
1543 (struct dm_capabilities_resp_msg *)dm_msg);
1544 break;
1545
1546 case DM_BALLOON_REQUEST:
1547 if (allow_hibernation) {
1548 pr_info("Ignore balloon-up request!\n");
1549 break;
1550 }
1551
1552 if (dm->state == DM_BALLOON_UP)
1553 pr_warn("Currently ballooning\n");
1554 bal_msg = (struct dm_balloon *)recv_buffer;
1555 dm->state = DM_BALLOON_UP;
1556 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1557 schedule_work(&dm_device.balloon_wrk.wrk);
1558 break;
1559
1560 case DM_UNBALLOON_REQUEST:
1561 if (allow_hibernation) {
1562 pr_info("Ignore balloon-down request!\n");
1563 break;
1564 }
1565
1566 dm->state = DM_BALLOON_DOWN;
1567 balloon_down(dm,
1568 (struct dm_unballoon_request *)recv_buffer);
1569 break;
1570
1571 case DM_MEM_HOT_ADD_REQUEST:
1572 if (dm->state == DM_HOT_ADD)
1573 pr_warn("Currently hot-adding\n");
1574 dm->state = DM_HOT_ADD;
1575 ha_msg = (struct dm_hot_add *)recv_buffer;
1576 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1577 /*
1578 * This is a normal hot-add request specifying
1579 * hot-add memory.
1580 */
1581 dm->host_specified_ha_region = false;
1582 ha_pg_range = &ha_msg->range;
1583 dm->ha_wrk.ha_page_range = *ha_pg_range;
1584 dm->ha_wrk.ha_region_range.page_range = 0;
1585 } else {
1586 /*
1587 * Host is specifying that we first hot-add
1588 * a region and then partially populate this
1589 * region.
1590 */
1591 dm->host_specified_ha_region = true;
1592 ha_pg_range = &ha_msg->range;
1593 ha_region = &ha_pg_range[1];
1594 dm->ha_wrk.ha_page_range = *ha_pg_range;
1595 dm->ha_wrk.ha_region_range = *ha_region;
1596 }
1597 schedule_work(&dm_device.ha_wrk.wrk);
1598 break;
1599
1600 case DM_INFO_MESSAGE:
1601 process_info(dm, (struct dm_info_msg *)dm_msg);
1602 break;
1603
1604 default:
1605 pr_warn_ratelimited("Unhandled message: type: %d\n", dm_hdr->type);
1606
1607 }
1608 }
1609
1610}
1611
1612#define HV_LARGE_REPORTING_ORDER 9
1613#define HV_LARGE_REPORTING_LEN (HV_HYP_PAGE_SIZE << \
1614 HV_LARGE_REPORTING_ORDER)
1615static int hv_free_page_report(struct page_reporting_dev_info *pr_dev_info,
1616 struct scatterlist *sgl, unsigned int nents)
1617{
1618 unsigned long flags;
1619 struct hv_memory_hint *hint;
1620 int i, order;
1621 u64 status;
1622 struct scatterlist *sg;
1623
1624 WARN_ON_ONCE(nents > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1625 WARN_ON_ONCE(sgl->length < (HV_HYP_PAGE_SIZE << page_reporting_order));
1626 local_irq_save(flags);
1627 hint = *this_cpu_ptr(hyperv_pcpu_input_arg);
1628 if (!hint) {
1629 local_irq_restore(flags);
1630 return -ENOSPC;
1631 }
1632
1633 hint->type = HV_EXT_MEMORY_HEAT_HINT_TYPE_COLD_DISCARD;
1634 hint->reserved = 0;
1635 for_each_sg(sgl, sg, nents, i) {
1636 union hv_gpa_page_range *range;
1637
1638 range = &hint->ranges[i];
1639 range->address_space = 0;
1640 order = get_order(sg->length);
1641 /*
1642 * Hyper-V expects the additional_pages field in the units
1643 * of one of these 3 sizes, 4Kbytes, 2Mbytes or 1Gbytes.
1644 * This is dictated by the values of the fields page.largesize
1645 * and page_size.
1646 * This code however, only uses 4Kbytes and 2Mbytes units
1647 * and not 1Gbytes unit.
1648 */
1649
1650 /* page reporting for pages 2MB or higher */
1651 if (order >= HV_LARGE_REPORTING_ORDER ) {
1652 range->page.largepage = 1;
1653 range->page_size = HV_GPA_PAGE_RANGE_PAGE_SIZE_2MB;
1654 range->base_large_pfn = page_to_hvpfn(
1655 sg_page(sg)) >> HV_LARGE_REPORTING_ORDER;
1656 range->page.additional_pages =
1657 (sg->length / HV_LARGE_REPORTING_LEN) - 1;
1658 } else {
1659 /* Page reporting for pages below 2MB */
1660 range->page.basepfn = page_to_hvpfn(sg_page(sg));
1661 range->page.largepage = false;
1662 range->page.additional_pages =
1663 (sg->length / HV_HYP_PAGE_SIZE) - 1;
1664 }
1665
1666 }
1667
1668 status = hv_do_rep_hypercall(HV_EXT_CALL_MEMORY_HEAT_HINT, nents, 0,
1669 hint, NULL);
1670 local_irq_restore(flags);
1671 if (!hv_result_success(status)) {
1672
1673 pr_err("Cold memory discard hypercall failed with status %llx\n",
1674 status);
1675 if (hv_hypercall_multi_failure > 0)
1676 hv_hypercall_multi_failure++;
1677
1678 if (hv_result(status) == HV_STATUS_INVALID_PARAMETER) {
1679 pr_err("Underlying Hyper-V does not support order less than 9. Hypercall failed\n");
1680 pr_err("Defaulting to page_reporting_order %d\n",
1681 pageblock_order);
1682 page_reporting_order = pageblock_order;
1683 hv_hypercall_multi_failure++;
1684 return -EINVAL;
1685 }
1686
1687 return -EINVAL;
1688 }
1689
1690 return 0;
1691}
1692
1693static void enable_page_reporting(void)
1694{
1695 int ret;
1696
1697 if (!hv_query_ext_cap(HV_EXT_CAPABILITY_MEMORY_COLD_DISCARD_HINT)) {
1698 pr_debug("Cold memory discard hint not supported by Hyper-V\n");
1699 return;
1700 }
1701
1702 BUILD_BUG_ON(PAGE_REPORTING_CAPACITY > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1703 dm_device.pr_dev_info.report = hv_free_page_report;
1704 /*
1705 * We let the page_reporting_order parameter decide the order
1706 * in the page_reporting code
1707 */
1708 dm_device.pr_dev_info.order = 0;
1709 ret = page_reporting_register(&dm_device.pr_dev_info);
1710 if (ret < 0) {
1711 dm_device.pr_dev_info.report = NULL;
1712 pr_err("Failed to enable cold memory discard: %d\n", ret);
1713 } else {
1714 pr_info("Cold memory discard hint enabled with order %d\n",
1715 page_reporting_order);
1716 }
1717}
1718
1719static void disable_page_reporting(void)
1720{
1721 if (dm_device.pr_dev_info.report) {
1722 page_reporting_unregister(&dm_device.pr_dev_info);
1723 dm_device.pr_dev_info.report = NULL;
1724 }
1725}
1726
1727static int ballooning_enabled(void)
1728{
1729 /*
1730 * Disable ballooning if the page size is not 4k (HV_HYP_PAGE_SIZE),
1731 * since currently it's unclear to us whether an unballoon request can
1732 * make sure all page ranges are guest page size aligned.
1733 */
1734 if (PAGE_SIZE != HV_HYP_PAGE_SIZE) {
1735 pr_info("Ballooning disabled because page size is not 4096 bytes\n");
1736 return 0;
1737 }
1738
1739 return 1;
1740}
1741
1742static int hot_add_enabled(void)
1743{
1744 /*
1745 * Disable hot add on ARM64, because we currently rely on
1746 * memory_add_physaddr_to_nid() to get a node id of a hot add range,
1747 * however ARM64's memory_add_physaddr_to_nid() always return 0 and
1748 * DM_MEM_HOT_ADD_REQUEST doesn't have the NUMA node information for
1749 * add_memory().
1750 */
1751 if (IS_ENABLED(CONFIG_ARM64)) {
1752 pr_info("Memory hot add disabled on ARM64\n");
1753 return 0;
1754 }
1755
1756 return 1;
1757}
1758
1759static int balloon_connect_vsp(struct hv_device *dev)
1760{
1761 struct dm_version_request version_req;
1762 struct dm_capabilities cap_msg;
1763 unsigned long t;
1764 int ret;
1765
1766 /*
1767 * max_pkt_size should be large enough for one vmbus packet header plus
1768 * our receive buffer size. Hyper-V sends messages up to
1769 * HV_HYP_PAGE_SIZE bytes long on balloon channel.
1770 */
1771 dev->channel->max_pkt_size = HV_HYP_PAGE_SIZE * 2;
1772
1773 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1774 balloon_onchannelcallback, dev);
1775 if (ret)
1776 return ret;
1777
1778 /*
1779 * Initiate the hand shake with the host and negotiate
1780 * a version that the host can support. We start with the
1781 * highest version number and go down if the host cannot
1782 * support it.
1783 */
1784 memset(&version_req, 0, sizeof(struct dm_version_request));
1785 version_req.hdr.type = DM_VERSION_REQUEST;
1786 version_req.hdr.size = sizeof(struct dm_version_request);
1787 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1788 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1789 version_req.is_last_attempt = 0;
1790 dm_device.version = version_req.version.version;
1791
1792 ret = vmbus_sendpacket(dev->channel, &version_req,
1793 sizeof(struct dm_version_request),
1794 (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1795 if (ret)
1796 goto out;
1797
1798 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1799 if (t == 0) {
1800 ret = -ETIMEDOUT;
1801 goto out;
1802 }
1803
1804 /*
1805 * If we could not negotiate a compatible version with the host
1806 * fail the probe function.
1807 */
1808 if (dm_device.state == DM_INIT_ERROR) {
1809 ret = -EPROTO;
1810 goto out;
1811 }
1812
1813 pr_info("Using Dynamic Memory protocol version %u.%u\n",
1814 DYNMEM_MAJOR_VERSION(dm_device.version),
1815 DYNMEM_MINOR_VERSION(dm_device.version));
1816
1817 /*
1818 * Now submit our capabilities to the host.
1819 */
1820 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1821 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1822 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1823 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1824
1825 /*
1826 * When hibernation (i.e. virtual ACPI S4 state) is enabled, the host
1827 * currently still requires the bits to be set, so we have to add code
1828 * to fail the host's hot-add and balloon up/down requests, if any.
1829 */
1830 cap_msg.caps.cap_bits.balloon = ballooning_enabled();
1831 cap_msg.caps.cap_bits.hot_add = hot_add_enabled();
1832
1833 /*
1834 * Specify our alignment requirements as it relates
1835 * memory hot-add. Specify 128MB alignment.
1836 */
1837 cap_msg.caps.cap_bits.hot_add_alignment = 7;
1838
1839 /*
1840 * Currently the host does not use these
1841 * values and we set them to what is done in the
1842 * Windows driver.
1843 */
1844 cap_msg.min_page_cnt = 0;
1845 cap_msg.max_page_number = -1;
1846
1847 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1848 sizeof(struct dm_capabilities),
1849 (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1850 if (ret)
1851 goto out;
1852
1853 t = wait_for_completion_timeout(&dm_device.host_event, 5*HZ);
1854 if (t == 0) {
1855 ret = -ETIMEDOUT;
1856 goto out;
1857 }
1858
1859 /*
1860 * If the host does not like our capabilities,
1861 * fail the probe function.
1862 */
1863 if (dm_device.state == DM_INIT_ERROR) {
1864 ret = -EPROTO;
1865 goto out;
1866 }
1867
1868 return 0;
1869out:
1870 vmbus_close(dev->channel);
1871 return ret;
1872}
1873
1874/*
1875 * DEBUGFS Interface
1876 */
1877#ifdef CONFIG_DEBUG_FS
1878
1879/**
1880 * hv_balloon_debug_show - shows statistics of balloon operations.
1881 * @f: pointer to the &struct seq_file.
1882 * @offset: ignored.
1883 *
1884 * Provides the statistics that can be accessed in hv-balloon in the debugfs.
1885 *
1886 * Return: zero on success or an error code.
1887 */
1888static int hv_balloon_debug_show(struct seq_file *f, void *offset)
1889{
1890 struct hv_dynmem_device *dm = f->private;
1891 char *sname;
1892
1893 seq_printf(f, "%-22s: %u.%u\n", "host_version",
1894 DYNMEM_MAJOR_VERSION(dm->version),
1895 DYNMEM_MINOR_VERSION(dm->version));
1896
1897 seq_printf(f, "%-22s:", "capabilities");
1898 if (ballooning_enabled())
1899 seq_puts(f, " enabled");
1900
1901 if (hot_add_enabled())
1902 seq_puts(f, " hot_add");
1903
1904 seq_puts(f, "\n");
1905
1906 seq_printf(f, "%-22s: %u", "state", dm->state);
1907 switch (dm->state) {
1908 case DM_INITIALIZING:
1909 sname = "Initializing";
1910 break;
1911 case DM_INITIALIZED:
1912 sname = "Initialized";
1913 break;
1914 case DM_BALLOON_UP:
1915 sname = "Balloon Up";
1916 break;
1917 case DM_BALLOON_DOWN:
1918 sname = "Balloon Down";
1919 break;
1920 case DM_HOT_ADD:
1921 sname = "Hot Add";
1922 break;
1923 case DM_INIT_ERROR:
1924 sname = "Error";
1925 break;
1926 default:
1927 sname = "Unknown";
1928 }
1929 seq_printf(f, " (%s)\n", sname);
1930
1931 /* HV Page Size */
1932 seq_printf(f, "%-22s: %ld\n", "page_size", HV_HYP_PAGE_SIZE);
1933
1934 /* Pages added with hot_add */
1935 seq_printf(f, "%-22s: %u\n", "pages_added", dm->num_pages_added);
1936
1937 /* pages that are "onlined"/used from pages_added */
1938 seq_printf(f, "%-22s: %u\n", "pages_onlined", dm->num_pages_onlined);
1939
1940 /* pages we have given back to host */
1941 seq_printf(f, "%-22s: %u\n", "pages_ballooned", dm->num_pages_ballooned);
1942
1943 seq_printf(f, "%-22s: %lu\n", "total_pages_committed",
1944 get_pages_committed(dm));
1945
1946 seq_printf(f, "%-22s: %llu\n", "max_dynamic_page_count",
1947 dm->max_dynamic_page_count);
1948
1949 return 0;
1950}
1951
1952DEFINE_SHOW_ATTRIBUTE(hv_balloon_debug);
1953
1954static void hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1955{
1956 debugfs_create_file("hv-balloon", 0444, NULL, b,
1957 &hv_balloon_debug_fops);
1958}
1959
1960static void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1961{
1962 debugfs_lookup_and_remove("hv-balloon", NULL);
1963}
1964
1965#else
1966
1967static inline void hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1968{
1969}
1970
1971static inline void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1972{
1973}
1974
1975#endif /* CONFIG_DEBUG_FS */
1976
1977static int balloon_probe(struct hv_device *dev,
1978 const struct hv_vmbus_device_id *dev_id)
1979{
1980 int ret;
1981
1982 allow_hibernation = hv_is_hibernation_supported();
1983 if (allow_hibernation)
1984 hot_add = false;
1985
1986#ifdef CONFIG_MEMORY_HOTPLUG
1987 do_hot_add = hot_add;
1988#else
1989 do_hot_add = false;
1990#endif
1991 dm_device.dev = dev;
1992 dm_device.state = DM_INITIALIZING;
1993 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1994 init_completion(&dm_device.host_event);
1995 init_completion(&dm_device.config_event);
1996 INIT_LIST_HEAD(&dm_device.ha_region_list);
1997 spin_lock_init(&dm_device.ha_lock);
1998 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1999 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
2000 dm_device.host_specified_ha_region = false;
2001
2002#ifdef CONFIG_MEMORY_HOTPLUG
2003 set_online_page_callback(&hv_online_page);
2004 init_completion(&dm_device.ol_waitevent);
2005 register_memory_notifier(&hv_memory_nb);
2006#endif
2007
2008 hv_set_drvdata(dev, &dm_device);
2009
2010 ret = balloon_connect_vsp(dev);
2011 if (ret != 0)
2012 goto connect_error;
2013
2014 enable_page_reporting();
2015 dm_device.state = DM_INITIALIZED;
2016
2017 dm_device.thread =
2018 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2019 if (IS_ERR(dm_device.thread)) {
2020 ret = PTR_ERR(dm_device.thread);
2021 goto probe_error;
2022 }
2023
2024 hv_balloon_debugfs_init(&dm_device);
2025
2026 return 0;
2027
2028probe_error:
2029 dm_device.state = DM_INIT_ERROR;
2030 dm_device.thread = NULL;
2031 disable_page_reporting();
2032 vmbus_close(dev->channel);
2033connect_error:
2034#ifdef CONFIG_MEMORY_HOTPLUG
2035 unregister_memory_notifier(&hv_memory_nb);
2036 restore_online_page_callback(&hv_online_page);
2037#endif
2038 return ret;
2039}
2040
2041static void balloon_remove(struct hv_device *dev)
2042{
2043 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
2044 struct hv_hotadd_state *has, *tmp;
2045 struct hv_hotadd_gap *gap, *tmp_gap;
2046
2047 if (dm->num_pages_ballooned != 0)
2048 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
2049
2050 hv_balloon_debugfs_exit(dm);
2051
2052 cancel_work_sync(&dm->balloon_wrk.wrk);
2053 cancel_work_sync(&dm->ha_wrk.wrk);
2054
2055 kthread_stop(dm->thread);
2056
2057 /*
2058 * This is to handle the case when balloon_resume()
2059 * call has failed and some cleanup has been done as
2060 * a part of the error handling.
2061 */
2062 if (dm_device.state != DM_INIT_ERROR) {
2063 disable_page_reporting();
2064 vmbus_close(dev->channel);
2065#ifdef CONFIG_MEMORY_HOTPLUG
2066 unregister_memory_notifier(&hv_memory_nb);
2067 restore_online_page_callback(&hv_online_page);
2068#endif
2069 }
2070
2071 guard(spinlock_irqsave)(&dm_device.ha_lock);
2072 list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
2073 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
2074 list_del(&gap->list);
2075 kfree(gap);
2076 }
2077 list_del(&has->list);
2078 kfree(has);
2079 }
2080}
2081
2082static int balloon_suspend(struct hv_device *hv_dev)
2083{
2084 struct hv_dynmem_device *dm = hv_get_drvdata(hv_dev);
2085
2086 tasklet_disable(&hv_dev->channel->callback_event);
2087
2088 cancel_work_sync(&dm->balloon_wrk.wrk);
2089 cancel_work_sync(&dm->ha_wrk.wrk);
2090
2091 if (dm->thread) {
2092 kthread_stop(dm->thread);
2093 dm->thread = NULL;
2094 vmbus_close(hv_dev->channel);
2095 }
2096
2097 tasklet_enable(&hv_dev->channel->callback_event);
2098
2099 return 0;
2100
2101}
2102
2103static int balloon_resume(struct hv_device *dev)
2104{
2105 int ret;
2106
2107 dm_device.state = DM_INITIALIZING;
2108
2109 ret = balloon_connect_vsp(dev);
2110
2111 if (ret != 0)
2112 goto out;
2113
2114 dm_device.thread =
2115 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2116 if (IS_ERR(dm_device.thread)) {
2117 ret = PTR_ERR(dm_device.thread);
2118 dm_device.thread = NULL;
2119 goto close_channel;
2120 }
2121
2122 dm_device.state = DM_INITIALIZED;
2123 return 0;
2124close_channel:
2125 vmbus_close(dev->channel);
2126out:
2127 dm_device.state = DM_INIT_ERROR;
2128 disable_page_reporting();
2129#ifdef CONFIG_MEMORY_HOTPLUG
2130 unregister_memory_notifier(&hv_memory_nb);
2131 restore_online_page_callback(&hv_online_page);
2132#endif
2133 return ret;
2134}
2135
2136static const struct hv_vmbus_device_id id_table[] = {
2137 /* Dynamic Memory Class ID */
2138 /* 525074DC-8985-46e2-8057-A307DC18A502 */
2139 { HV_DM_GUID, },
2140 { },
2141};
2142
2143MODULE_DEVICE_TABLE(vmbus, id_table);
2144
2145static struct hv_driver balloon_drv = {
2146 .name = "hv_balloon",
2147 .id_table = id_table,
2148 .probe = balloon_probe,
2149 .remove = balloon_remove,
2150 .suspend = balloon_suspend,
2151 .resume = balloon_resume,
2152 .driver = {
2153 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
2154 },
2155};
2156
2157static int __init init_balloon_drv(void)
2158{
2159
2160 return vmbus_driver_register(&balloon_drv);
2161}
2162
2163module_init(init_balloon_drv);
2164
2165MODULE_DESCRIPTION("Hyper-V Balloon");
2166MODULE_LICENSE("GPL");