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#include <linux/sizes.h>
29
30#include <linux/hyperv.h>
31#include <asm/hyperv-tlfs.h>
32
33#include <asm/mshyperv.h>
34
35#define CREATE_TRACE_POINTS
36#include "hv_trace_balloon.h"
37
38/*
39 * We begin with definitions supporting the Dynamic Memory protocol
40 * with the host.
41 *
42 * Begin protocol definitions.
43 */
44
45/*
46 * Protocol versions. The low word is the minor version, the high word the major
47 * version.
48 *
49 * History:
50 * Initial version 1.0
51 * Changed to 0.1 on 2009/03/25
52 * Changes to 0.2 on 2009/05/14
53 * Changes to 0.3 on 2009/12/03
54 * Changed to 1.0 on 2011/04/05
55 */
56
57#define DYNMEM_MAKE_VERSION(Major, Minor) ((__u32)(((Major) << 16) | (Minor)))
58#define DYNMEM_MAJOR_VERSION(Version) ((__u32)(Version) >> 16)
59#define DYNMEM_MINOR_VERSION(Version) ((__u32)(Version) & 0xff)
60
61enum {
62 DYNMEM_PROTOCOL_VERSION_1 = DYNMEM_MAKE_VERSION(0, 3),
63 DYNMEM_PROTOCOL_VERSION_2 = DYNMEM_MAKE_VERSION(1, 0),
64 DYNMEM_PROTOCOL_VERSION_3 = DYNMEM_MAKE_VERSION(2, 0),
65
66 DYNMEM_PROTOCOL_VERSION_WIN7 = DYNMEM_PROTOCOL_VERSION_1,
67 DYNMEM_PROTOCOL_VERSION_WIN8 = DYNMEM_PROTOCOL_VERSION_2,
68 DYNMEM_PROTOCOL_VERSION_WIN10 = DYNMEM_PROTOCOL_VERSION_3,
69
70 DYNMEM_PROTOCOL_VERSION_CURRENT = DYNMEM_PROTOCOL_VERSION_WIN10
71};
72
73/*
74 * Message Types
75 */
76
77enum dm_message_type {
78 /*
79 * Version 0.3
80 */
81 DM_ERROR = 0,
82 DM_VERSION_REQUEST = 1,
83 DM_VERSION_RESPONSE = 2,
84 DM_CAPABILITIES_REPORT = 3,
85 DM_CAPABILITIES_RESPONSE = 4,
86 DM_STATUS_REPORT = 5,
87 DM_BALLOON_REQUEST = 6,
88 DM_BALLOON_RESPONSE = 7,
89 DM_UNBALLOON_REQUEST = 8,
90 DM_UNBALLOON_RESPONSE = 9,
91 DM_MEM_HOT_ADD_REQUEST = 10,
92 DM_MEM_HOT_ADD_RESPONSE = 11,
93 DM_VERSION_03_MAX = 11,
94 /*
95 * Version 1.0.
96 */
97 DM_INFO_MESSAGE = 12,
98 DM_VERSION_1_MAX = 12
99};
100
101/*
102 * Structures defining the dynamic memory management
103 * protocol.
104 */
105
106union dm_version {
107 struct {
108 __u16 minor_version;
109 __u16 major_version;
110 };
111 __u32 version;
112} __packed;
113
114union dm_caps {
115 struct {
116 __u64 balloon:1;
117 __u64 hot_add:1;
118 /*
119 * To support guests that may have alignment
120 * limitations on hot-add, the guest can specify
121 * its alignment requirements; a value of n
122 * represents an alignment of 2^n in mega bytes.
123 */
124 __u64 hot_add_alignment:4;
125 __u64 reservedz:58;
126 } cap_bits;
127 __u64 caps;
128} __packed;
129
130union dm_mem_page_range {
131 struct {
132 /*
133 * The PFN number of the first page in the range.
134 * 40 bits is the architectural limit of a PFN
135 * number for AMD64.
136 */
137 __u64 start_page:40;
138 /*
139 * The number of pages in the range.
140 */
141 __u64 page_cnt:24;
142 } finfo;
143 __u64 page_range;
144} __packed;
145
146/*
147 * The header for all dynamic memory messages:
148 *
149 * type: Type of the message.
150 * size: Size of the message in bytes; including the header.
151 * trans_id: The guest is responsible for manufacturing this ID.
152 */
153
154struct dm_header {
155 __u16 type;
156 __u16 size;
157 __u32 trans_id;
158} __packed;
159
160/*
161 * A generic message format for dynamic memory.
162 * Specific message formats are defined later in the file.
163 */
164
165struct dm_message {
166 struct dm_header hdr;
167 __u8 data[]; /* enclosed message */
168} __packed;
169
170/*
171 * Specific message types supporting the dynamic memory protocol.
172 */
173
174/*
175 * Version negotiation message. Sent from the guest to the host.
176 * The guest is free to try different versions until the host
177 * accepts the version.
178 *
179 * dm_version: The protocol version requested.
180 * is_last_attempt: If TRUE, this is the last version guest will request.
181 * reservedz: Reserved field, set to zero.
182 */
183
184struct dm_version_request {
185 struct dm_header hdr;
186 union dm_version version;
187 __u32 is_last_attempt:1;
188 __u32 reservedz:31;
189} __packed;
190
191/*
192 * Version response message; Host to Guest and indicates
193 * if the host has accepted the version sent by the guest.
194 *
195 * is_accepted: If TRUE, host has accepted the version and the guest
196 * should proceed to the next stage of the protocol. FALSE indicates that
197 * guest should re-try with a different version.
198 *
199 * reservedz: Reserved field, set to zero.
200 */
201
202struct dm_version_response {
203 struct dm_header hdr;
204 __u64 is_accepted:1;
205 __u64 reservedz:63;
206} __packed;
207
208/*
209 * Message reporting capabilities. This is sent from the guest to the
210 * host.
211 */
212
213struct dm_capabilities {
214 struct dm_header hdr;
215 union dm_caps caps;
216 __u64 min_page_cnt;
217 __u64 max_page_number;
218} __packed;
219
220/*
221 * Response to the capabilities message. This is sent from the host to the
222 * guest. This message notifies if the host has accepted the guest's
223 * capabilities. If the host has not accepted, the guest must shutdown
224 * the service.
225 *
226 * is_accepted: Indicates if the host has accepted guest's capabilities.
227 * reservedz: Must be 0.
228 */
229
230struct dm_capabilities_resp_msg {
231 struct dm_header hdr;
232 __u64 is_accepted:1;
233 __u64 reservedz:63;
234} __packed;
235
236/*
237 * This message is used to report memory pressure from the guest.
238 * This message is not part of any transaction and there is no
239 * response to this message.
240 *
241 * num_avail: Available memory in pages.
242 * num_committed: Committed memory in pages.
243 * page_file_size: The accumulated size of all page files
244 * in the system in pages.
245 * zero_free: The number of zero and free pages.
246 * page_file_writes: The writes to the page file in pages.
247 * io_diff: An indicator of file cache efficiency or page file activity,
248 * calculated as File Cache Page Fault Count - Page Read Count.
249 * This value is in pages.
250 *
251 * Some of these metrics are Windows specific and fortunately
252 * the algorithm on the host side that computes the guest memory
253 * pressure only uses num_committed value.
254 */
255
256struct dm_status {
257 struct dm_header hdr;
258 __u64 num_avail;
259 __u64 num_committed;
260 __u64 page_file_size;
261 __u64 zero_free;
262 __u32 page_file_writes;
263 __u32 io_diff;
264} __packed;
265
266/*
267 * Message to ask the guest to allocate memory - balloon up message.
268 * This message is sent from the host to the guest. The guest may not be
269 * able to allocate as much memory as requested.
270 *
271 * num_pages: number of pages to allocate.
272 */
273
274struct dm_balloon {
275 struct dm_header hdr;
276 __u32 num_pages;
277 __u32 reservedz;
278} __packed;
279
280/*
281 * Balloon response message; this message is sent from the guest
282 * to the host in response to the balloon message.
283 *
284 * reservedz: Reserved; must be set to zero.
285 * more_pages: If FALSE, this is the last message of the transaction.
286 * if TRUE there will be at least one more message from the guest.
287 *
288 * range_count: The number of ranges in the range array.
289 *
290 * range_array: An array of page ranges returned to the host.
291 *
292 */
293
294struct dm_balloon_response {
295 struct dm_header hdr;
296 __u32 reservedz;
297 __u32 more_pages:1;
298 __u32 range_count:31;
299 union dm_mem_page_range range_array[];
300} __packed;
301
302/*
303 * Un-balloon message; this message is sent from the host
304 * to the guest to give guest more memory.
305 *
306 * more_pages: If FALSE, this is the last message of the transaction.
307 * if TRUE there will be at least one more message from the guest.
308 *
309 * reservedz: Reserved; must be set to zero.
310 *
311 * range_count: The number of ranges in the range array.
312 *
313 * range_array: An array of page ranges returned to the host.
314 *
315 */
316
317struct dm_unballoon_request {
318 struct dm_header hdr;
319 __u32 more_pages:1;
320 __u32 reservedz:31;
321 __u32 range_count;
322 union dm_mem_page_range range_array[];
323} __packed;
324
325/*
326 * Un-balloon response message; this message is sent from the guest
327 * to the host in response to an unballoon request.
328 *
329 */
330
331struct dm_unballoon_response {
332 struct dm_header hdr;
333} __packed;
334
335/*
336 * Hot add request message. Message sent from the host to the guest.
337 *
338 * mem_range: Memory range to hot add.
339 *
340 */
341
342struct dm_hot_add {
343 struct dm_header hdr;
344 union dm_mem_page_range range;
345} __packed;
346
347/*
348 * Hot add response message.
349 * This message is sent by the guest to report the status of a hot add request.
350 * If page_count is less than the requested page count, then the host should
351 * assume all further hot add requests will fail, since this indicates that
352 * the guest has hit an upper physical memory barrier.
353 *
354 * Hot adds may also fail due to low resources; in this case, the guest must
355 * not complete this message until the hot add can succeed, and the host must
356 * not send a new hot add request until the response is sent.
357 * If VSC fails to hot add memory DYNMEM_NUMBER_OF_UNSUCCESSFUL_HOTADD_ATTEMPTS
358 * times it fails the request.
359 *
360 *
361 * page_count: number of pages that were successfully hot added.
362 *
363 * result: result of the operation 1: success, 0: failure.
364 *
365 */
366
367struct dm_hot_add_response {
368 struct dm_header hdr;
369 __u32 page_count;
370 __u32 result;
371} __packed;
372
373/*
374 * Types of information sent from host to the guest.
375 */
376
377enum dm_info_type {
378 INFO_TYPE_MAX_PAGE_CNT = 0,
379 MAX_INFO_TYPE
380};
381
382/*
383 * Header for the information message.
384 */
385
386struct dm_info_header {
387 enum dm_info_type type;
388 __u32 data_size;
389} __packed;
390
391/*
392 * This message is sent from the host to the guest to pass
393 * some relevant information (win8 addition).
394 *
395 * reserved: no used.
396 * info_size: size of the information blob.
397 * info: information blob.
398 */
399
400struct dm_info_msg {
401 struct dm_header hdr;
402 __u32 reserved;
403 __u32 info_size;
404 __u8 info[];
405};
406
407/*
408 * End protocol definitions.
409 */
410
411/*
412 * State to manage hot adding memory into the guest.
413 * The range start_pfn : end_pfn specifies the range
414 * that the host has asked us to hot add. The range
415 * start_pfn : ha_end_pfn specifies the range that we have
416 * currently hot added. We hot add in chunks equal to the
417 * memory block size; it is possible that we may not be able
418 * to bring online all the pages in the region. The range
419 * covered_start_pfn:covered_end_pfn defines the pages that can
420 * be brought online.
421 */
422
423struct hv_hotadd_state {
424 struct list_head list;
425 unsigned long start_pfn;
426 unsigned long covered_start_pfn;
427 unsigned long covered_end_pfn;
428 unsigned long ha_end_pfn;
429 unsigned long end_pfn;
430 /*
431 * A list of gaps.
432 */
433 struct list_head gap_list;
434};
435
436struct hv_hotadd_gap {
437 struct list_head list;
438 unsigned long start_pfn;
439 unsigned long end_pfn;
440};
441
442struct balloon_state {
443 __u32 num_pages;
444 struct work_struct wrk;
445};
446
447struct hot_add_wrk {
448 union dm_mem_page_range ha_page_range;
449 union dm_mem_page_range ha_region_range;
450 struct work_struct wrk;
451};
452
453static bool allow_hibernation;
454static bool hot_add = true;
455static bool do_hot_add;
456/*
457 * Delay reporting memory pressure by
458 * the specified number of seconds.
459 */
460static uint pressure_report_delay = 45;
461extern unsigned int page_reporting_order;
462#define HV_MAX_FAILURES 2
463
464/*
465 * The last time we posted a pressure report to host.
466 */
467static unsigned long last_post_time;
468
469static int hv_hypercall_multi_failure;
470
471module_param(hot_add, bool, 0644);
472MODULE_PARM_DESC(hot_add, "If set attempt memory hot_add");
473
474module_param(pressure_report_delay, uint, 0644);
475MODULE_PARM_DESC(pressure_report_delay, "Delay in secs in reporting pressure");
476static atomic_t trans_id = ATOMIC_INIT(0);
477
478static int dm_ring_size = VMBUS_RING_SIZE(16 * 1024);
479
480/*
481 * Driver specific state.
482 */
483
484enum hv_dm_state {
485 DM_INITIALIZING = 0,
486 DM_INITIALIZED,
487 DM_BALLOON_UP,
488 DM_BALLOON_DOWN,
489 DM_HOT_ADD,
490 DM_INIT_ERROR
491};
492
493static __u8 recv_buffer[HV_HYP_PAGE_SIZE];
494static __u8 balloon_up_send_buffer[HV_HYP_PAGE_SIZE];
495
496static unsigned long ha_pages_in_chunk;
497#define HA_BYTES_IN_CHUNK (ha_pages_in_chunk << PAGE_SHIFT)
498
499#define PAGES_IN_2M (2 * 1024 * 1024 / PAGE_SIZE)
500
501struct hv_dynmem_device {
502 struct hv_device *dev;
503 enum hv_dm_state state;
504 struct completion host_event;
505 struct completion config_event;
506
507 /*
508 * Number of pages we have currently ballooned out.
509 */
510 unsigned int num_pages_ballooned;
511 unsigned int num_pages_onlined;
512 unsigned int num_pages_added;
513
514 /*
515 * State to manage the ballooning (up) operation.
516 */
517 struct balloon_state balloon_wrk;
518
519 /*
520 * State to execute the "hot-add" operation.
521 */
522 struct hot_add_wrk ha_wrk;
523
524 /*
525 * This state tracks if the host has specified a hot-add
526 * region.
527 */
528 bool host_specified_ha_region;
529
530 /*
531 * State to synchronize hot-add.
532 */
533 struct completion ol_waitevent;
534 /*
535 * This thread handles hot-add
536 * requests from the host as well as notifying
537 * the host with regards to memory pressure in
538 * the guest.
539 */
540 struct task_struct *thread;
541
542 /*
543 * Protects ha_region_list, num_pages_onlined counter and individual
544 * regions from ha_region_list.
545 */
546 spinlock_t ha_lock;
547
548 /*
549 * A list of hot-add regions.
550 */
551 struct list_head ha_region_list;
552
553 /*
554 * We start with the highest version we can support
555 * and downgrade based on the host; we save here the
556 * next version to try.
557 */
558 __u32 next_version;
559
560 /*
561 * The negotiated version agreed by host.
562 */
563 __u32 version;
564
565 struct page_reporting_dev_info pr_dev_info;
566
567 /*
568 * Maximum number of pages that can be hot_add-ed
569 */
570 __u64 max_dynamic_page_count;
571};
572
573static struct hv_dynmem_device dm_device;
574
575static void post_status(struct hv_dynmem_device *dm);
576
577static void enable_page_reporting(void);
578
579static void disable_page_reporting(void);
580
581#ifdef CONFIG_MEMORY_HOTPLUG
582static inline bool has_pfn_is_backed(struct hv_hotadd_state *has,
583 unsigned long pfn)
584{
585 struct hv_hotadd_gap *gap;
586
587 /* The page is not backed. */
588 if (pfn < has->covered_start_pfn || pfn >= has->covered_end_pfn)
589 return false;
590
591 /* Check for gaps. */
592 list_for_each_entry(gap, &has->gap_list, list) {
593 if (pfn >= gap->start_pfn && pfn < gap->end_pfn)
594 return false;
595 }
596
597 return true;
598}
599
600static unsigned long hv_page_offline_check(unsigned long start_pfn,
601 unsigned long nr_pages)
602{
603 unsigned long pfn = start_pfn, count = 0;
604 struct hv_hotadd_state *has;
605 bool found;
606
607 while (pfn < start_pfn + nr_pages) {
608 /*
609 * Search for HAS which covers the pfn and when we find one
610 * count how many consequitive PFNs are covered.
611 */
612 found = false;
613 list_for_each_entry(has, &dm_device.ha_region_list, list) {
614 while ((pfn >= has->start_pfn) &&
615 (pfn < has->end_pfn) &&
616 (pfn < start_pfn + nr_pages)) {
617 found = true;
618 if (has_pfn_is_backed(has, pfn))
619 count++;
620 pfn++;
621 }
622 }
623
624 /*
625 * This PFN is not in any HAS (e.g. we're offlining a region
626 * which was present at boot), no need to account for it. Go
627 * to the next one.
628 */
629 if (!found)
630 pfn++;
631 }
632
633 return count;
634}
635
636static int hv_memory_notifier(struct notifier_block *nb, unsigned long val,
637 void *v)
638{
639 struct memory_notify *mem = (struct memory_notify *)v;
640 unsigned long pfn_count;
641
642 switch (val) {
643 case MEM_ONLINE:
644 case MEM_CANCEL_ONLINE:
645 complete(&dm_device.ol_waitevent);
646 break;
647
648 case MEM_OFFLINE:
649 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
650 pfn_count = hv_page_offline_check(mem->start_pfn,
651 mem->nr_pages);
652 if (pfn_count <= dm_device.num_pages_onlined) {
653 dm_device.num_pages_onlined -= pfn_count;
654 } else {
655 /*
656 * We're offlining more pages than we
657 * managed to online. This is
658 * unexpected. In any case don't let
659 * num_pages_onlined wrap around zero.
660 */
661 WARN_ON_ONCE(1);
662 dm_device.num_pages_onlined = 0;
663 }
664 }
665 break;
666 case MEM_GOING_ONLINE:
667 case MEM_GOING_OFFLINE:
668 case MEM_CANCEL_OFFLINE:
669 break;
670 }
671 return NOTIFY_OK;
672}
673
674static struct notifier_block hv_memory_nb = {
675 .notifier_call = hv_memory_notifier,
676 .priority = 0
677};
678
679/* Check if the particular page is backed and can be onlined and online it. */
680static void hv_page_online_one(struct hv_hotadd_state *has, struct page *pg)
681{
682 if (!has_pfn_is_backed(has, page_to_pfn(pg))) {
683 if (!PageOffline(pg))
684 __SetPageOffline(pg);
685 return;
686 } else if (!PageOffline(pg))
687 return;
688
689 /* This frame is currently backed; online the page. */
690 generic_online_page(pg, 0);
691
692 lockdep_assert_held(&dm_device.ha_lock);
693 dm_device.num_pages_onlined++;
694}
695
696static void hv_bring_pgs_online(struct hv_hotadd_state *has,
697 unsigned long start_pfn, unsigned long size)
698{
699 int i;
700
701 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
702 for (i = 0; i < size; i++)
703 hv_page_online_one(has, pfn_to_page(start_pfn + i));
704}
705
706static void hv_mem_hot_add(unsigned long start, unsigned long size,
707 unsigned long pfn_count,
708 struct hv_hotadd_state *has)
709{
710 int ret = 0;
711 int i, nid;
712 unsigned long start_pfn;
713 unsigned long processed_pfn;
714 unsigned long total_pfn = pfn_count;
715
716 for (i = 0; i < (size/ha_pages_in_chunk); i++) {
717 start_pfn = start + (i * ha_pages_in_chunk);
718
719 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
720 has->ha_end_pfn += ha_pages_in_chunk;
721 processed_pfn = umin(total_pfn, ha_pages_in_chunk);
722 total_pfn -= processed_pfn;
723 has->covered_end_pfn += processed_pfn;
724 }
725
726 reinit_completion(&dm_device.ol_waitevent);
727
728 nid = memory_add_physaddr_to_nid(PFN_PHYS(start_pfn));
729 ret = add_memory(nid, PFN_PHYS((start_pfn)),
730 HA_BYTES_IN_CHUNK, MHP_MERGE_RESOURCE);
731
732 if (ret) {
733 pr_err("hot_add memory failed error is %d\n", ret);
734 if (ret == -EEXIST) {
735 /*
736 * This error indicates that the error
737 * is not a transient failure. This is the
738 * case where the guest's physical address map
739 * precludes hot adding memory. Stop all further
740 * memory hot-add.
741 */
742 do_hot_add = false;
743 }
744 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
745 has->ha_end_pfn -= ha_pages_in_chunk;
746 has->covered_end_pfn -= processed_pfn;
747 }
748 break;
749 }
750
751 /*
752 * Wait for memory to get onlined. If the kernel onlined the
753 * memory when adding it, this will return directly. Otherwise,
754 * it will wait for user space to online the memory. This helps
755 * to avoid adding memory faster than it is getting onlined. As
756 * adding succeeded, it is ok to proceed even if the memory was
757 * not onlined in time.
758 */
759 wait_for_completion_timeout(&dm_device.ol_waitevent, secs_to_jiffies(5));
760 post_status(&dm_device);
761 }
762}
763
764static void hv_online_page(struct page *pg, unsigned int order)
765{
766 struct hv_hotadd_state *has;
767 unsigned long pfn = page_to_pfn(pg);
768
769 guard(spinlock_irqsave)(&dm_device.ha_lock);
770 list_for_each_entry(has, &dm_device.ha_region_list, list) {
771 /* The page belongs to a different HAS. */
772 if (pfn < has->start_pfn ||
773 (pfn + (1UL << order) > has->end_pfn))
774 continue;
775
776 hv_bring_pgs_online(has, pfn, 1UL << order);
777 break;
778 }
779}
780
781static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
782{
783 struct hv_hotadd_state *has;
784 struct hv_hotadd_gap *gap;
785 unsigned long residual;
786 int ret = 0;
787
788 guard(spinlock_irqsave)(&dm_device.ha_lock);
789 list_for_each_entry(has, &dm_device.ha_region_list, list) {
790 /*
791 * If the pfn range we are dealing with is not in the current
792 * "hot add block", move on.
793 */
794 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
795 continue;
796
797 /*
798 * If the current start pfn is not where the covered_end
799 * is, create a gap and update covered_end_pfn.
800 */
801 if (has->covered_end_pfn != start_pfn) {
802 gap = kzalloc(sizeof(struct hv_hotadd_gap), GFP_ATOMIC);
803 if (!gap) {
804 ret = -ENOMEM;
805 break;
806 }
807
808 INIT_LIST_HEAD(&gap->list);
809 gap->start_pfn = has->covered_end_pfn;
810 gap->end_pfn = start_pfn;
811 list_add_tail(&gap->list, &has->gap_list);
812
813 has->covered_end_pfn = start_pfn;
814 }
815
816 /*
817 * If the current hot add-request extends beyond
818 * our current limit; extend it.
819 */
820 if ((start_pfn + pfn_cnt) > has->end_pfn) {
821 /* Extend the region by multiples of ha_pages_in_chunk */
822 residual = (start_pfn + pfn_cnt - has->end_pfn);
823 has->end_pfn += ALIGN(residual, ha_pages_in_chunk);
824 }
825
826 ret = 1;
827 break;
828 }
829
830 return ret;
831}
832
833static unsigned long handle_pg_range(unsigned long pg_start,
834 unsigned long pg_count)
835{
836 unsigned long start_pfn = pg_start;
837 unsigned long pfn_cnt = pg_count;
838 unsigned long size;
839 struct hv_hotadd_state *has;
840 unsigned long pgs_ol = 0;
841 unsigned long old_covered_state;
842 unsigned long res = 0, flags;
843
844 pr_debug("Hot adding %lu pages starting at pfn 0x%lx.\n", pg_count,
845 pg_start);
846
847 spin_lock_irqsave(&dm_device.ha_lock, flags);
848 list_for_each_entry(has, &dm_device.ha_region_list, list) {
849 /*
850 * If the pfn range we are dealing with is not in the current
851 * "hot add block", move on.
852 */
853 if (start_pfn < has->start_pfn || start_pfn >= has->end_pfn)
854 continue;
855
856 old_covered_state = has->covered_end_pfn;
857
858 if (start_pfn < has->ha_end_pfn) {
859 /*
860 * This is the case where we are backing pages
861 * in an already hot added region. Bring
862 * these pages online first.
863 */
864 pgs_ol = has->ha_end_pfn - start_pfn;
865 if (pgs_ol > pfn_cnt)
866 pgs_ol = pfn_cnt;
867
868 has->covered_end_pfn += pgs_ol;
869 pfn_cnt -= pgs_ol;
870 /*
871 * Check if the corresponding memory block is already
872 * online. It is possible to observe struct pages still
873 * being uninitialized here so check section instead.
874 * In case the section is online we need to bring the
875 * rest of pfns (which were not backed previously)
876 * online too.
877 */
878 if (start_pfn > has->start_pfn &&
879 online_section_nr(pfn_to_section_nr(start_pfn)))
880 hv_bring_pgs_online(has, start_pfn, pgs_ol);
881 }
882
883 if (has->ha_end_pfn < has->end_pfn && pfn_cnt > 0) {
884 /*
885 * We have some residual hot add range
886 * that needs to be hot added; hot add
887 * it now. Hot add a multiple of
888 * ha_pages_in_chunk that fully covers the pages
889 * we have.
890 */
891 size = (has->end_pfn - has->ha_end_pfn);
892 if (pfn_cnt <= size) {
893 size = ALIGN(pfn_cnt, ha_pages_in_chunk);
894 } else {
895 pfn_cnt = size;
896 }
897 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
898 hv_mem_hot_add(has->ha_end_pfn, size, pfn_cnt, has);
899 spin_lock_irqsave(&dm_device.ha_lock, flags);
900 }
901 /*
902 * If we managed to online any pages that were given to us,
903 * we declare success.
904 */
905 res = has->covered_end_pfn - old_covered_state;
906 break;
907 }
908 spin_unlock_irqrestore(&dm_device.ha_lock, flags);
909
910 return res;
911}
912
913static unsigned long process_hot_add(unsigned long pg_start,
914 unsigned long pfn_cnt,
915 unsigned long rg_start,
916 unsigned long rg_size)
917{
918 struct hv_hotadd_state *ha_region = NULL;
919 int covered;
920
921 if (pfn_cnt == 0)
922 return 0;
923
924 if (!dm_device.host_specified_ha_region) {
925 covered = pfn_covered(pg_start, pfn_cnt);
926 if (covered < 0)
927 return 0;
928
929 if (covered)
930 goto do_pg_range;
931 }
932
933 /*
934 * If the host has specified a hot-add range; deal with it first.
935 */
936
937 if (rg_size != 0) {
938 ha_region = kzalloc(sizeof(struct hv_hotadd_state), GFP_KERNEL);
939 if (!ha_region)
940 return 0;
941
942 INIT_LIST_HEAD(&ha_region->list);
943 INIT_LIST_HEAD(&ha_region->gap_list);
944
945 ha_region->start_pfn = rg_start;
946 ha_region->ha_end_pfn = rg_start;
947 ha_region->covered_start_pfn = pg_start;
948 ha_region->covered_end_pfn = pg_start;
949 ha_region->end_pfn = rg_start + rg_size;
950
951 scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
952 list_add_tail(&ha_region->list, &dm_device.ha_region_list);
953 }
954 }
955
956do_pg_range:
957 /*
958 * Process the page range specified; bringing them
959 * online if possible.
960 */
961 return handle_pg_range(pg_start, pfn_cnt);
962}
963
964#endif
965
966static void hot_add_req(struct work_struct *dummy)
967{
968 struct dm_hot_add_response resp;
969#ifdef CONFIG_MEMORY_HOTPLUG
970 unsigned long pg_start, pfn_cnt;
971 unsigned long rg_start, rg_sz;
972#endif
973 struct hv_dynmem_device *dm = &dm_device;
974
975 memset(&resp, 0, sizeof(struct dm_hot_add_response));
976 resp.hdr.type = DM_MEM_HOT_ADD_RESPONSE;
977 resp.hdr.size = sizeof(struct dm_hot_add_response);
978
979#ifdef CONFIG_MEMORY_HOTPLUG
980 pg_start = dm->ha_wrk.ha_page_range.finfo.start_page;
981 pfn_cnt = dm->ha_wrk.ha_page_range.finfo.page_cnt;
982
983 rg_start = dm->ha_wrk.ha_region_range.finfo.start_page;
984 rg_sz = dm->ha_wrk.ha_region_range.finfo.page_cnt;
985
986 if (rg_start == 0 && !dm->host_specified_ha_region) {
987 /*
988 * The host has not specified the hot-add region.
989 * Based on the hot-add page range being specified,
990 * compute a hot-add region that can cover the pages
991 * that need to be hot-added while ensuring the alignment
992 * and size requirements of Linux as it relates to hot-add.
993 */
994 rg_start = ALIGN_DOWN(pg_start, ha_pages_in_chunk);
995 rg_sz = ALIGN(pfn_cnt, ha_pages_in_chunk);
996 }
997
998 if (do_hot_add)
999 resp.page_count = process_hot_add(pg_start, pfn_cnt,
1000 rg_start, rg_sz);
1001
1002 dm->num_pages_added += resp.page_count;
1003#endif
1004 /*
1005 * The result field of the response structure has the
1006 * following semantics:
1007 *
1008 * 1. If all or some pages hot-added: Guest should return success.
1009 *
1010 * 2. If no pages could be hot-added:
1011 *
1012 * If the guest returns success, then the host
1013 * will not attempt any further hot-add operations. This
1014 * signifies a permanent failure.
1015 *
1016 * If the guest returns failure, then this failure will be
1017 * treated as a transient failure and the host may retry the
1018 * hot-add operation after some delay.
1019 */
1020 if (resp.page_count > 0)
1021 resp.result = 1;
1022 else if (!do_hot_add)
1023 resp.result = 1;
1024 else
1025 resp.result = 0;
1026
1027 if (!do_hot_add || resp.page_count == 0) {
1028 if (!allow_hibernation)
1029 pr_err("Memory hot add failed\n");
1030 else
1031 pr_info("Ignore hot-add request!\n");
1032 }
1033
1034 dm->state = DM_INITIALIZED;
1035 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1036 vmbus_sendpacket(dm->dev->channel, &resp,
1037 sizeof(struct dm_hot_add_response),
1038 (unsigned long)NULL,
1039 VM_PKT_DATA_INBAND, 0);
1040}
1041
1042static void process_info(struct hv_dynmem_device *dm, struct dm_info_msg *msg)
1043{
1044 struct dm_info_header *info_hdr;
1045
1046 info_hdr = (struct dm_info_header *)msg->info;
1047
1048 switch (info_hdr->type) {
1049 case INFO_TYPE_MAX_PAGE_CNT:
1050 if (info_hdr->data_size == sizeof(__u64)) {
1051 __u64 *max_page_count = (__u64 *)&info_hdr[1];
1052
1053 pr_info("Max. dynamic memory size: %llu MB\n",
1054 (*max_page_count) >> (20 - HV_HYP_PAGE_SHIFT));
1055 dm->max_dynamic_page_count = *max_page_count;
1056 }
1057
1058 break;
1059 default:
1060 pr_warn("Received Unknown type: %d\n", info_hdr->type);
1061 }
1062}
1063
1064static unsigned long compute_balloon_floor(void)
1065{
1066 unsigned long min_pages;
1067 unsigned long nr_pages = totalram_pages();
1068#define MB2PAGES(mb) ((mb) << (20 - PAGE_SHIFT))
1069 /* Simple continuous piecewiese linear function:
1070 * max MiB -> min MiB gradient
1071 * 0 0
1072 * 16 16
1073 * 32 24
1074 * 128 72 (1/2)
1075 * 512 168 (1/4)
1076 * 2048 360 (1/8)
1077 * 8192 744 (1/16)
1078 * 32768 1512 (1/32)
1079 */
1080 if (nr_pages < MB2PAGES(128))
1081 min_pages = MB2PAGES(8) + (nr_pages >> 1);
1082 else if (nr_pages < MB2PAGES(512))
1083 min_pages = MB2PAGES(40) + (nr_pages >> 2);
1084 else if (nr_pages < MB2PAGES(2048))
1085 min_pages = MB2PAGES(104) + (nr_pages >> 3);
1086 else if (nr_pages < MB2PAGES(8192))
1087 min_pages = MB2PAGES(232) + (nr_pages >> 4);
1088 else
1089 min_pages = MB2PAGES(488) + (nr_pages >> 5);
1090#undef MB2PAGES
1091 return min_pages;
1092}
1093
1094/*
1095 * Compute total committed memory pages
1096 */
1097
1098static unsigned long get_pages_committed(struct hv_dynmem_device *dm)
1099{
1100 return vm_memory_committed() +
1101 dm->num_pages_ballooned +
1102 (dm->num_pages_added > dm->num_pages_onlined ?
1103 dm->num_pages_added - dm->num_pages_onlined : 0) +
1104 compute_balloon_floor();
1105}
1106
1107/*
1108 * Post our status as it relates memory pressure to the
1109 * host. Host expects the guests to post this status
1110 * periodically at 1 second intervals.
1111 *
1112 * The metrics specified in this protocol are very Windows
1113 * specific and so we cook up numbers here to convey our memory
1114 * pressure.
1115 */
1116
1117static void post_status(struct hv_dynmem_device *dm)
1118{
1119 struct dm_status status;
1120 unsigned long now = jiffies;
1121 unsigned long last_post = last_post_time;
1122 unsigned long num_pages_avail, num_pages_committed;
1123
1124 if (pressure_report_delay > 0) {
1125 --pressure_report_delay;
1126 return;
1127 }
1128
1129 if (!time_after(now, (last_post_time + HZ)))
1130 return;
1131
1132 memset(&status, 0, sizeof(struct dm_status));
1133 status.hdr.type = DM_STATUS_REPORT;
1134 status.hdr.size = sizeof(struct dm_status);
1135 status.hdr.trans_id = atomic_inc_return(&trans_id);
1136
1137 /*
1138 * The host expects the guest to report free and committed memory.
1139 * Furthermore, the host expects the pressure information to include
1140 * the ballooned out pages. For a given amount of memory that we are
1141 * managing we need to compute a floor below which we should not
1142 * balloon. Compute this and add it to the pressure report.
1143 * We also need to report all offline pages (num_pages_added -
1144 * num_pages_onlined) as committed to the host, otherwise it can try
1145 * asking us to balloon them out.
1146 */
1147 num_pages_avail = si_mem_available();
1148 num_pages_committed = get_pages_committed(dm);
1149
1150 trace_balloon_status(num_pages_avail, num_pages_committed,
1151 vm_memory_committed(), dm->num_pages_ballooned,
1152 dm->num_pages_added, dm->num_pages_onlined);
1153
1154 /* Convert numbers of pages into numbers of HV_HYP_PAGEs. */
1155 status.num_avail = num_pages_avail * NR_HV_HYP_PAGES_IN_PAGE;
1156 status.num_committed = num_pages_committed * NR_HV_HYP_PAGES_IN_PAGE;
1157
1158 /*
1159 * If our transaction ID is no longer current, just don't
1160 * send the status. This can happen if we were interrupted
1161 * after we picked our transaction ID.
1162 */
1163 if (status.hdr.trans_id != atomic_read(&trans_id))
1164 return;
1165
1166 /*
1167 * If the last post time that we sampled has changed,
1168 * we have raced, don't post the status.
1169 */
1170 if (last_post != last_post_time)
1171 return;
1172
1173 last_post_time = jiffies;
1174 vmbus_sendpacket(dm->dev->channel, &status,
1175 sizeof(struct dm_status),
1176 (unsigned long)NULL,
1177 VM_PKT_DATA_INBAND, 0);
1178}
1179
1180static void free_balloon_pages(struct hv_dynmem_device *dm,
1181 union dm_mem_page_range *range_array)
1182{
1183 int num_pages = range_array->finfo.page_cnt;
1184 __u64 start_frame = range_array->finfo.start_page;
1185 struct page *pg;
1186 int i;
1187
1188 for (i = 0; i < num_pages; i++) {
1189 pg = pfn_to_page(i + start_frame);
1190 __ClearPageOffline(pg);
1191 __free_page(pg);
1192 dm->num_pages_ballooned--;
1193 adjust_managed_page_count(pg, 1);
1194 }
1195}
1196
1197static unsigned int alloc_balloon_pages(struct hv_dynmem_device *dm,
1198 unsigned int num_pages,
1199 struct dm_balloon_response *bl_resp,
1200 int alloc_unit)
1201{
1202 unsigned int i, j;
1203 struct page *pg;
1204
1205 for (i = 0; i < num_pages / alloc_unit; i++) {
1206 if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
1207 HV_HYP_PAGE_SIZE)
1208 return i * alloc_unit;
1209
1210 /*
1211 * We execute this code in a thread context. Furthermore,
1212 * we don't want the kernel to try too hard.
1213 */
1214 pg = alloc_pages(GFP_HIGHUSER | __GFP_NORETRY |
1215 __GFP_NOMEMALLOC | __GFP_NOWARN,
1216 get_order(alloc_unit << PAGE_SHIFT));
1217
1218 if (!pg)
1219 return i * alloc_unit;
1220
1221 dm->num_pages_ballooned += alloc_unit;
1222
1223 /*
1224 * If we allocatted 2M pages; split them so we
1225 * can free them in any order we get.
1226 */
1227
1228 if (alloc_unit != 1)
1229 split_page(pg, get_order(alloc_unit << PAGE_SHIFT));
1230
1231 /* mark all pages offline */
1232 for (j = 0; j < alloc_unit; j++) {
1233 __SetPageOffline(pg + j);
1234 adjust_managed_page_count(pg + j, -1);
1235 }
1236
1237 bl_resp->range_count++;
1238 bl_resp->range_array[i].finfo.start_page =
1239 page_to_pfn(pg);
1240 bl_resp->range_array[i].finfo.page_cnt = alloc_unit;
1241 bl_resp->hdr.size += sizeof(union dm_mem_page_range);
1242 }
1243
1244 return i * alloc_unit;
1245}
1246
1247static void balloon_up(struct work_struct *dummy)
1248{
1249 unsigned int num_pages = dm_device.balloon_wrk.num_pages;
1250 unsigned int num_ballooned = 0;
1251 struct dm_balloon_response *bl_resp;
1252 int alloc_unit;
1253 int ret;
1254 bool done = false;
1255 int i;
1256 long avail_pages;
1257 unsigned long floor;
1258
1259 /*
1260 * We will attempt 2M allocations. However, if we fail to
1261 * allocate 2M chunks, we will go back to PAGE_SIZE allocations.
1262 */
1263 alloc_unit = PAGES_IN_2M;
1264
1265 avail_pages = si_mem_available();
1266 floor = compute_balloon_floor();
1267
1268 /* Refuse to balloon below the floor. */
1269 if (avail_pages < num_pages || avail_pages - num_pages < floor) {
1270 pr_info("Balloon request will be partially fulfilled. %s\n",
1271 avail_pages < num_pages ? "Not enough memory." :
1272 "Balloon floor reached.");
1273
1274 num_pages = avail_pages > floor ? (avail_pages - floor) : 0;
1275 }
1276
1277 while (!done) {
1278 memset(balloon_up_send_buffer, 0, HV_HYP_PAGE_SIZE);
1279 bl_resp = (struct dm_balloon_response *)balloon_up_send_buffer;
1280 bl_resp->hdr.type = DM_BALLOON_RESPONSE;
1281 bl_resp->hdr.size = sizeof(struct dm_balloon_response);
1282 bl_resp->more_pages = 1;
1283
1284 num_pages -= num_ballooned;
1285 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
1286 bl_resp, alloc_unit);
1287
1288 if (alloc_unit != 1 && num_ballooned == 0) {
1289 alloc_unit = 1;
1290 continue;
1291 }
1292
1293 if (num_ballooned == 0 || num_ballooned == num_pages) {
1294 pr_debug("Ballooned %u out of %u requested pages.\n",
1295 num_pages, dm_device.balloon_wrk.num_pages);
1296
1297 bl_resp->more_pages = 0;
1298 done = true;
1299 dm_device.state = DM_INITIALIZED;
1300 }
1301
1302 /*
1303 * We are pushing a lot of data through the channel;
1304 * deal with transient failures caused because of the
1305 * lack of space in the ring buffer.
1306 */
1307
1308 do {
1309 bl_resp->hdr.trans_id = atomic_inc_return(&trans_id);
1310 ret = vmbus_sendpacket(dm_device.dev->channel,
1311 bl_resp,
1312 bl_resp->hdr.size,
1313 (unsigned long)NULL,
1314 VM_PKT_DATA_INBAND, 0);
1315
1316 if (ret == -EAGAIN)
1317 msleep(20);
1318 post_status(&dm_device);
1319 } while (ret == -EAGAIN);
1320
1321 if (ret) {
1322 /*
1323 * Free up the memory we allocatted.
1324 */
1325 pr_err("Balloon response failed\n");
1326
1327 for (i = 0; i < bl_resp->range_count; i++)
1328 free_balloon_pages(&dm_device,
1329 &bl_resp->range_array[i]);
1330
1331 done = true;
1332 }
1333 }
1334}
1335
1336static void balloon_down(struct hv_dynmem_device *dm,
1337 struct dm_unballoon_request *req)
1338{
1339 union dm_mem_page_range *range_array = req->range_array;
1340 int range_count = req->range_count;
1341 struct dm_unballoon_response resp;
1342 int i;
1343 unsigned int prev_pages_ballooned = dm->num_pages_ballooned;
1344
1345 for (i = 0; i < range_count; i++) {
1346 free_balloon_pages(dm, &range_array[i]);
1347 complete(&dm_device.config_event);
1348 }
1349
1350 pr_debug("Freed %u ballooned pages.\n",
1351 prev_pages_ballooned - dm->num_pages_ballooned);
1352
1353 if (req->more_pages == 1)
1354 return;
1355
1356 memset(&resp, 0, sizeof(struct dm_unballoon_response));
1357 resp.hdr.type = DM_UNBALLOON_RESPONSE;
1358 resp.hdr.trans_id = atomic_inc_return(&trans_id);
1359 resp.hdr.size = sizeof(struct dm_unballoon_response);
1360
1361 vmbus_sendpacket(dm_device.dev->channel, &resp,
1362 sizeof(struct dm_unballoon_response),
1363 (unsigned long)NULL,
1364 VM_PKT_DATA_INBAND, 0);
1365
1366 dm->state = DM_INITIALIZED;
1367}
1368
1369static void balloon_onchannelcallback(void *context);
1370
1371static int dm_thread_func(void *dm_dev)
1372{
1373 struct hv_dynmem_device *dm = dm_dev;
1374
1375 while (!kthread_should_stop()) {
1376 wait_for_completion_interruptible_timeout(&dm_device.config_event,
1377 secs_to_jiffies(1));
1378 /*
1379 * The host expects us to post information on the memory
1380 * pressure every second.
1381 */
1382 reinit_completion(&dm_device.config_event);
1383 post_status(dm);
1384 /*
1385 * disable free page reporting if multiple hypercall
1386 * failure flag set. It is not done in the page_reporting
1387 * callback context as that causes a deadlock between
1388 * page_reporting_process() and page_reporting_unregister()
1389 */
1390 if (hv_hypercall_multi_failure >= HV_MAX_FAILURES) {
1391 pr_err("Multiple failures in cold memory discard hypercall, disabling page reporting\n");
1392 disable_page_reporting();
1393 /* Reset the flag after disabling reporting */
1394 hv_hypercall_multi_failure = 0;
1395 }
1396 }
1397
1398 return 0;
1399}
1400
1401static void version_resp(struct hv_dynmem_device *dm,
1402 struct dm_version_response *vresp)
1403{
1404 struct dm_version_request version_req;
1405 int ret;
1406
1407 if (vresp->is_accepted) {
1408 /*
1409 * We are done; wakeup the
1410 * context waiting for version
1411 * negotiation.
1412 */
1413 complete(&dm->host_event);
1414 return;
1415 }
1416 /*
1417 * If there are more versions to try, continue
1418 * with negotiations; if not
1419 * shutdown the service since we are not able
1420 * to negotiate a suitable version number
1421 * with the host.
1422 */
1423 if (dm->next_version == 0)
1424 goto version_error;
1425
1426 memset(&version_req, 0, sizeof(struct dm_version_request));
1427 version_req.hdr.type = DM_VERSION_REQUEST;
1428 version_req.hdr.size = sizeof(struct dm_version_request);
1429 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1430 version_req.version.version = dm->next_version;
1431 dm->version = version_req.version.version;
1432
1433 /*
1434 * Set the next version to try in case current version fails.
1435 * Win7 protocol ought to be the last one to try.
1436 */
1437 switch (version_req.version.version) {
1438 case DYNMEM_PROTOCOL_VERSION_WIN8:
1439 dm->next_version = DYNMEM_PROTOCOL_VERSION_WIN7;
1440 version_req.is_last_attempt = 0;
1441 break;
1442 default:
1443 dm->next_version = 0;
1444 version_req.is_last_attempt = 1;
1445 }
1446
1447 ret = vmbus_sendpacket(dm->dev->channel, &version_req,
1448 sizeof(struct dm_version_request),
1449 (unsigned long)NULL,
1450 VM_PKT_DATA_INBAND, 0);
1451
1452 if (ret)
1453 goto version_error;
1454
1455 return;
1456
1457version_error:
1458 dm->state = DM_INIT_ERROR;
1459 complete(&dm->host_event);
1460}
1461
1462static void cap_resp(struct hv_dynmem_device *dm,
1463 struct dm_capabilities_resp_msg *cap_resp)
1464{
1465 if (!cap_resp->is_accepted) {
1466 pr_err("Capabilities not accepted by host\n");
1467 dm->state = DM_INIT_ERROR;
1468 }
1469 complete(&dm->host_event);
1470}
1471
1472static void balloon_onchannelcallback(void *context)
1473{
1474 struct hv_device *dev = context;
1475 u32 recvlen;
1476 u64 requestid;
1477 struct dm_message *dm_msg;
1478 struct dm_header *dm_hdr;
1479 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1480 struct dm_balloon *bal_msg;
1481 struct dm_hot_add *ha_msg;
1482 union dm_mem_page_range *ha_pg_range;
1483 union dm_mem_page_range *ha_region;
1484
1485 memset(recv_buffer, 0, sizeof(recv_buffer));
1486 vmbus_recvpacket(dev->channel, recv_buffer,
1487 HV_HYP_PAGE_SIZE, &recvlen, &requestid);
1488
1489 if (recvlen > 0) {
1490 dm_msg = (struct dm_message *)recv_buffer;
1491 dm_hdr = &dm_msg->hdr;
1492
1493 switch (dm_hdr->type) {
1494 case DM_VERSION_RESPONSE:
1495 version_resp(dm,
1496 (struct dm_version_response *)dm_msg);
1497 break;
1498
1499 case DM_CAPABILITIES_RESPONSE:
1500 cap_resp(dm,
1501 (struct dm_capabilities_resp_msg *)dm_msg);
1502 break;
1503
1504 case DM_BALLOON_REQUEST:
1505 if (allow_hibernation) {
1506 pr_info("Ignore balloon-up request!\n");
1507 break;
1508 }
1509
1510 if (dm->state == DM_BALLOON_UP)
1511 pr_warn("Currently ballooning\n");
1512 bal_msg = (struct dm_balloon *)recv_buffer;
1513 dm->state = DM_BALLOON_UP;
1514 dm_device.balloon_wrk.num_pages = bal_msg->num_pages;
1515 schedule_work(&dm_device.balloon_wrk.wrk);
1516 break;
1517
1518 case DM_UNBALLOON_REQUEST:
1519 if (allow_hibernation) {
1520 pr_info("Ignore balloon-down request!\n");
1521 break;
1522 }
1523
1524 dm->state = DM_BALLOON_DOWN;
1525 balloon_down(dm,
1526 (struct dm_unballoon_request *)recv_buffer);
1527 break;
1528
1529 case DM_MEM_HOT_ADD_REQUEST:
1530 if (dm->state == DM_HOT_ADD)
1531 pr_warn("Currently hot-adding\n");
1532 dm->state = DM_HOT_ADD;
1533 ha_msg = (struct dm_hot_add *)recv_buffer;
1534 if (ha_msg->hdr.size == sizeof(struct dm_hot_add)) {
1535 /*
1536 * This is a normal hot-add request specifying
1537 * hot-add memory.
1538 */
1539 dm->host_specified_ha_region = false;
1540 ha_pg_range = &ha_msg->range;
1541 dm->ha_wrk.ha_page_range = *ha_pg_range;
1542 dm->ha_wrk.ha_region_range.page_range = 0;
1543 } else {
1544 /*
1545 * Host is specifying that we first hot-add
1546 * a region and then partially populate this
1547 * region.
1548 */
1549 dm->host_specified_ha_region = true;
1550 ha_pg_range = &ha_msg->range;
1551 ha_region = &ha_pg_range[1];
1552 dm->ha_wrk.ha_page_range = *ha_pg_range;
1553 dm->ha_wrk.ha_region_range = *ha_region;
1554 }
1555 schedule_work(&dm_device.ha_wrk.wrk);
1556 break;
1557
1558 case DM_INFO_MESSAGE:
1559 process_info(dm, (struct dm_info_msg *)dm_msg);
1560 break;
1561
1562 default:
1563 pr_warn_ratelimited("Unhandled message: type: %d\n", dm_hdr->type);
1564 }
1565 }
1566}
1567
1568#define HV_LARGE_REPORTING_ORDER 9
1569#define HV_LARGE_REPORTING_LEN (HV_HYP_PAGE_SIZE << \
1570 HV_LARGE_REPORTING_ORDER)
1571static int hv_free_page_report(struct page_reporting_dev_info *pr_dev_info,
1572 struct scatterlist *sgl, unsigned int nents)
1573{
1574 unsigned long flags;
1575 struct hv_memory_hint *hint;
1576 int i, order;
1577 u64 status;
1578 struct scatterlist *sg;
1579
1580 WARN_ON_ONCE(nents > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1581 WARN_ON_ONCE(sgl->length < (HV_HYP_PAGE_SIZE << page_reporting_order));
1582 local_irq_save(flags);
1583 hint = *this_cpu_ptr(hyperv_pcpu_input_arg);
1584 if (!hint) {
1585 local_irq_restore(flags);
1586 return -ENOSPC;
1587 }
1588
1589 hint->type = HV_EXT_MEMORY_HEAT_HINT_TYPE_COLD_DISCARD;
1590 hint->reserved = 0;
1591 for_each_sg(sgl, sg, nents, i) {
1592 union hv_gpa_page_range *range;
1593
1594 range = &hint->ranges[i];
1595 range->address_space = 0;
1596 order = get_order(sg->length);
1597 /*
1598 * Hyper-V expects the additional_pages field in the units
1599 * of one of these 3 sizes, 4Kbytes, 2Mbytes or 1Gbytes.
1600 * This is dictated by the values of the fields page.largesize
1601 * and page_size.
1602 * This code however, only uses 4Kbytes and 2Mbytes units
1603 * and not 1Gbytes unit.
1604 */
1605
1606 /* page reporting for pages 2MB or higher */
1607 if (order >= HV_LARGE_REPORTING_ORDER) {
1608 range->page.largepage = 1;
1609 range->page_size = HV_GPA_PAGE_RANGE_PAGE_SIZE_2MB;
1610 range->base_large_pfn = page_to_hvpfn(
1611 sg_page(sg)) >> HV_LARGE_REPORTING_ORDER;
1612 range->page.additional_pages =
1613 (sg->length / HV_LARGE_REPORTING_LEN) - 1;
1614 } else {
1615 /* Page reporting for pages below 2MB */
1616 range->page.basepfn = page_to_hvpfn(sg_page(sg));
1617 range->page.largepage = false;
1618 range->page.additional_pages =
1619 (sg->length / HV_HYP_PAGE_SIZE) - 1;
1620 }
1621 }
1622
1623 status = hv_do_rep_hypercall(HV_EXT_CALL_MEMORY_HEAT_HINT, nents, 0,
1624 hint, NULL);
1625 local_irq_restore(flags);
1626 if (!hv_result_success(status)) {
1627 pr_err("Cold memory discard hypercall failed with status %llx\n",
1628 status);
1629 if (hv_hypercall_multi_failure > 0)
1630 hv_hypercall_multi_failure++;
1631
1632 if (hv_result(status) == HV_STATUS_INVALID_PARAMETER) {
1633 pr_err("Underlying Hyper-V does not support order less than 9. Hypercall failed\n");
1634 pr_err("Defaulting to page_reporting_order %d\n",
1635 pageblock_order);
1636 page_reporting_order = pageblock_order;
1637 hv_hypercall_multi_failure++;
1638 return -EINVAL;
1639 }
1640
1641 return -EINVAL;
1642 }
1643
1644 return 0;
1645}
1646
1647static void enable_page_reporting(void)
1648{
1649 int ret;
1650
1651 if (!hv_query_ext_cap(HV_EXT_CAPABILITY_MEMORY_COLD_DISCARD_HINT)) {
1652 pr_debug("Cold memory discard hint not supported by Hyper-V\n");
1653 return;
1654 }
1655
1656 BUILD_BUG_ON(PAGE_REPORTING_CAPACITY > HV_MEMORY_HINT_MAX_GPA_PAGE_RANGES);
1657 dm_device.pr_dev_info.report = hv_free_page_report;
1658 /*
1659 * We let the page_reporting_order parameter decide the order
1660 * in the page_reporting code
1661 */
1662 dm_device.pr_dev_info.order = 0;
1663 ret = page_reporting_register(&dm_device.pr_dev_info);
1664 if (ret < 0) {
1665 dm_device.pr_dev_info.report = NULL;
1666 pr_err("Failed to enable cold memory discard: %d\n", ret);
1667 } else {
1668 pr_info("Cold memory discard hint enabled with order %d\n",
1669 page_reporting_order);
1670 }
1671}
1672
1673static void disable_page_reporting(void)
1674{
1675 if (dm_device.pr_dev_info.report) {
1676 page_reporting_unregister(&dm_device.pr_dev_info);
1677 dm_device.pr_dev_info.report = NULL;
1678 }
1679}
1680
1681static int ballooning_enabled(void)
1682{
1683 /*
1684 * Disable ballooning if the page size is not 4k (HV_HYP_PAGE_SIZE),
1685 * since currently it's unclear to us whether an unballoon request can
1686 * make sure all page ranges are guest page size aligned.
1687 */
1688 if (PAGE_SIZE != HV_HYP_PAGE_SIZE) {
1689 pr_info("Ballooning disabled because page size is not 4096 bytes\n");
1690 return 0;
1691 }
1692
1693 return 1;
1694}
1695
1696static int hot_add_enabled(void)
1697{
1698 /*
1699 * Disable hot add on ARM64, because we currently rely on
1700 * memory_add_physaddr_to_nid() to get a node id of a hot add range,
1701 * however ARM64's memory_add_physaddr_to_nid() always return 0 and
1702 * DM_MEM_HOT_ADD_REQUEST doesn't have the NUMA node information for
1703 * add_memory().
1704 */
1705 if (IS_ENABLED(CONFIG_ARM64)) {
1706 pr_info("Memory hot add disabled on ARM64\n");
1707 return 0;
1708 }
1709
1710 return 1;
1711}
1712
1713static int balloon_connect_vsp(struct hv_device *dev)
1714{
1715 struct dm_version_request version_req;
1716 struct dm_capabilities cap_msg;
1717 unsigned long t;
1718 int ret;
1719
1720 /*
1721 * max_pkt_size should be large enough for one vmbus packet header plus
1722 * our receive buffer size. Hyper-V sends messages up to
1723 * HV_HYP_PAGE_SIZE bytes long on balloon channel.
1724 */
1725 dev->channel->max_pkt_size = HV_HYP_PAGE_SIZE * 2;
1726
1727 ret = vmbus_open(dev->channel, dm_ring_size, dm_ring_size, NULL, 0,
1728 balloon_onchannelcallback, dev);
1729 if (ret)
1730 return ret;
1731
1732 /*
1733 * Initiate the hand shake with the host and negotiate
1734 * a version that the host can support. We start with the
1735 * highest version number and go down if the host cannot
1736 * support it.
1737 */
1738 memset(&version_req, 0, sizeof(struct dm_version_request));
1739 version_req.hdr.type = DM_VERSION_REQUEST;
1740 version_req.hdr.size = sizeof(struct dm_version_request);
1741 version_req.hdr.trans_id = atomic_inc_return(&trans_id);
1742 version_req.version.version = DYNMEM_PROTOCOL_VERSION_WIN10;
1743 version_req.is_last_attempt = 0;
1744 dm_device.version = version_req.version.version;
1745
1746 ret = vmbus_sendpacket(dev->channel, &version_req,
1747 sizeof(struct dm_version_request),
1748 (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1749 if (ret)
1750 goto out;
1751
1752 t = wait_for_completion_timeout(&dm_device.host_event, secs_to_jiffies(5));
1753 if (t == 0) {
1754 ret = -ETIMEDOUT;
1755 goto out;
1756 }
1757
1758 /*
1759 * If we could not negotiate a compatible version with the host
1760 * fail the probe function.
1761 */
1762 if (dm_device.state == DM_INIT_ERROR) {
1763 ret = -EPROTO;
1764 goto out;
1765 }
1766
1767 pr_info("Using Dynamic Memory protocol version %u.%u\n",
1768 DYNMEM_MAJOR_VERSION(dm_device.version),
1769 DYNMEM_MINOR_VERSION(dm_device.version));
1770
1771 /*
1772 * Now submit our capabilities to the host.
1773 */
1774 memset(&cap_msg, 0, sizeof(struct dm_capabilities));
1775 cap_msg.hdr.type = DM_CAPABILITIES_REPORT;
1776 cap_msg.hdr.size = sizeof(struct dm_capabilities);
1777 cap_msg.hdr.trans_id = atomic_inc_return(&trans_id);
1778
1779 /*
1780 * When hibernation (i.e. virtual ACPI S4 state) is enabled, the host
1781 * currently still requires the bits to be set, so we have to add code
1782 * to fail the host's hot-add and balloon up/down requests, if any.
1783 */
1784 cap_msg.caps.cap_bits.balloon = ballooning_enabled();
1785 cap_msg.caps.cap_bits.hot_add = hot_add_enabled();
1786
1787 /*
1788 * Specify our alignment requirements for memory hot-add. The value is
1789 * the log base 2 of the number of megabytes in a chunk. For example,
1790 * with 256 MiB chunks, the value is 8. The number of MiB in a chunk
1791 * must be a power of 2.
1792 */
1793 cap_msg.caps.cap_bits.hot_add_alignment =
1794 ilog2(HA_BYTES_IN_CHUNK / SZ_1M);
1795
1796 /*
1797 * Currently the host does not use these
1798 * values and we set them to what is done in the
1799 * Windows driver.
1800 */
1801 cap_msg.min_page_cnt = 0;
1802 cap_msg.max_page_number = -1;
1803
1804 ret = vmbus_sendpacket(dev->channel, &cap_msg,
1805 sizeof(struct dm_capabilities),
1806 (unsigned long)NULL, VM_PKT_DATA_INBAND, 0);
1807 if (ret)
1808 goto out;
1809
1810 t = wait_for_completion_timeout(&dm_device.host_event, secs_to_jiffies(5));
1811 if (t == 0) {
1812 ret = -ETIMEDOUT;
1813 goto out;
1814 }
1815
1816 /*
1817 * If the host does not like our capabilities,
1818 * fail the probe function.
1819 */
1820 if (dm_device.state == DM_INIT_ERROR) {
1821 ret = -EPROTO;
1822 goto out;
1823 }
1824
1825 return 0;
1826out:
1827 vmbus_close(dev->channel);
1828 return ret;
1829}
1830
1831/*
1832 * DEBUGFS Interface
1833 */
1834#ifdef CONFIG_DEBUG_FS
1835
1836/**
1837 * hv_balloon_debug_show - shows statistics of balloon operations.
1838 * @f: pointer to the &struct seq_file.
1839 * @offset: ignored.
1840 *
1841 * Provides the statistics that can be accessed in hv-balloon in the debugfs.
1842 *
1843 * Return: zero on success or an error code.
1844 */
1845static int hv_balloon_debug_show(struct seq_file *f, void *offset)
1846{
1847 struct hv_dynmem_device *dm = f->private;
1848 char *sname;
1849
1850 seq_printf(f, "%-22s: %u.%u\n", "host_version",
1851 DYNMEM_MAJOR_VERSION(dm->version),
1852 DYNMEM_MINOR_VERSION(dm->version));
1853
1854 seq_printf(f, "%-22s:", "capabilities");
1855 if (ballooning_enabled())
1856 seq_puts(f, " enabled");
1857
1858 if (hot_add_enabled())
1859 seq_puts(f, " hot_add");
1860
1861 seq_puts(f, "\n");
1862
1863 seq_printf(f, "%-22s: %u", "state", dm->state);
1864 switch (dm->state) {
1865 case DM_INITIALIZING:
1866 sname = "Initializing";
1867 break;
1868 case DM_INITIALIZED:
1869 sname = "Initialized";
1870 break;
1871 case DM_BALLOON_UP:
1872 sname = "Balloon Up";
1873 break;
1874 case DM_BALLOON_DOWN:
1875 sname = "Balloon Down";
1876 break;
1877 case DM_HOT_ADD:
1878 sname = "Hot Add";
1879 break;
1880 case DM_INIT_ERROR:
1881 sname = "Error";
1882 break;
1883 default:
1884 sname = "Unknown";
1885 }
1886 seq_printf(f, " (%s)\n", sname);
1887
1888 /* HV Page Size */
1889 seq_printf(f, "%-22s: %ld\n", "page_size", HV_HYP_PAGE_SIZE);
1890
1891 /* Pages added with hot_add */
1892 seq_printf(f, "%-22s: %u\n", "pages_added", dm->num_pages_added);
1893
1894 /* pages that are "onlined"/used from pages_added */
1895 seq_printf(f, "%-22s: %u\n", "pages_onlined", dm->num_pages_onlined);
1896
1897 /* pages we have given back to host */
1898 seq_printf(f, "%-22s: %u\n", "pages_ballooned", dm->num_pages_ballooned);
1899
1900 seq_printf(f, "%-22s: %lu\n", "total_pages_committed",
1901 get_pages_committed(dm));
1902
1903 seq_printf(f, "%-22s: %llu\n", "max_dynamic_page_count",
1904 dm->max_dynamic_page_count);
1905
1906 return 0;
1907}
1908
1909DEFINE_SHOW_ATTRIBUTE(hv_balloon_debug);
1910
1911static void hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1912{
1913 debugfs_create_file("hv-balloon", 0444, NULL, b,
1914 &hv_balloon_debug_fops);
1915}
1916
1917static void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1918{
1919 debugfs_lookup_and_remove("hv-balloon", NULL);
1920}
1921
1922#else
1923
1924static inline void hv_balloon_debugfs_init(struct hv_dynmem_device *b)
1925{
1926}
1927
1928static inline void hv_balloon_debugfs_exit(struct hv_dynmem_device *b)
1929{
1930}
1931
1932#endif /* CONFIG_DEBUG_FS */
1933
1934static int balloon_probe(struct hv_device *dev,
1935 const struct hv_vmbus_device_id *dev_id)
1936{
1937 int ret;
1938
1939 allow_hibernation = hv_is_hibernation_supported();
1940 if (allow_hibernation)
1941 hot_add = false;
1942
1943#ifdef CONFIG_MEMORY_HOTPLUG
1944 /*
1945 * Hot-add must operate in chunks that are of size equal to the
1946 * memory block size because that's what the core add_memory()
1947 * interface requires. The Hyper-V interface requires that the memory
1948 * block size be a power of 2, which is guaranteed by the check in
1949 * memory_dev_init().
1950 */
1951 ha_pages_in_chunk = memory_block_size_bytes() / PAGE_SIZE;
1952 do_hot_add = hot_add;
1953#else
1954 /*
1955 * Without MEMORY_HOTPLUG, the guest returns a failure status for all
1956 * hot add requests from Hyper-V, and the chunk size is used only to
1957 * specify alignment to Hyper-V as required by the host/guest protocol.
1958 * Somewhat arbitrarily, use 128 MiB.
1959 */
1960 ha_pages_in_chunk = SZ_128M / PAGE_SIZE;
1961 do_hot_add = false;
1962#endif
1963 dm_device.dev = dev;
1964 dm_device.state = DM_INITIALIZING;
1965 dm_device.next_version = DYNMEM_PROTOCOL_VERSION_WIN8;
1966 init_completion(&dm_device.host_event);
1967 init_completion(&dm_device.config_event);
1968 INIT_LIST_HEAD(&dm_device.ha_region_list);
1969 spin_lock_init(&dm_device.ha_lock);
1970 INIT_WORK(&dm_device.balloon_wrk.wrk, balloon_up);
1971 INIT_WORK(&dm_device.ha_wrk.wrk, hot_add_req);
1972 dm_device.host_specified_ha_region = false;
1973
1974#ifdef CONFIG_MEMORY_HOTPLUG
1975 set_online_page_callback(&hv_online_page);
1976 init_completion(&dm_device.ol_waitevent);
1977 register_memory_notifier(&hv_memory_nb);
1978#endif
1979
1980 hv_set_drvdata(dev, &dm_device);
1981
1982 ret = balloon_connect_vsp(dev);
1983 if (ret != 0)
1984 goto connect_error;
1985
1986 enable_page_reporting();
1987 dm_device.state = DM_INITIALIZED;
1988
1989 dm_device.thread =
1990 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
1991 if (IS_ERR(dm_device.thread)) {
1992 ret = PTR_ERR(dm_device.thread);
1993 goto probe_error;
1994 }
1995
1996 hv_balloon_debugfs_init(&dm_device);
1997
1998 return 0;
1999
2000probe_error:
2001 dm_device.state = DM_INIT_ERROR;
2002 dm_device.thread = NULL;
2003 disable_page_reporting();
2004 vmbus_close(dev->channel);
2005connect_error:
2006#ifdef CONFIG_MEMORY_HOTPLUG
2007 unregister_memory_notifier(&hv_memory_nb);
2008 restore_online_page_callback(&hv_online_page);
2009#endif
2010 return ret;
2011}
2012
2013static void balloon_remove(struct hv_device *dev)
2014{
2015 struct hv_dynmem_device *dm = hv_get_drvdata(dev);
2016 struct hv_hotadd_state *has, *tmp;
2017 struct hv_hotadd_gap *gap, *tmp_gap;
2018
2019 if (dm->num_pages_ballooned != 0)
2020 pr_warn("Ballooned pages: %d\n", dm->num_pages_ballooned);
2021
2022 hv_balloon_debugfs_exit(dm);
2023
2024 cancel_work_sync(&dm->balloon_wrk.wrk);
2025 cancel_work_sync(&dm->ha_wrk.wrk);
2026
2027 kthread_stop(dm->thread);
2028
2029 /*
2030 * This is to handle the case when balloon_resume()
2031 * call has failed and some cleanup has been done as
2032 * a part of the error handling.
2033 */
2034 if (dm_device.state != DM_INIT_ERROR) {
2035 disable_page_reporting();
2036 vmbus_close(dev->channel);
2037#ifdef CONFIG_MEMORY_HOTPLUG
2038 unregister_memory_notifier(&hv_memory_nb);
2039 restore_online_page_callback(&hv_online_page);
2040#endif
2041 }
2042
2043 guard(spinlock_irqsave)(&dm_device.ha_lock);
2044 list_for_each_entry_safe(has, tmp, &dm->ha_region_list, list) {
2045 list_for_each_entry_safe(gap, tmp_gap, &has->gap_list, list) {
2046 list_del(&gap->list);
2047 kfree(gap);
2048 }
2049 list_del(&has->list);
2050 kfree(has);
2051 }
2052}
2053
2054static int balloon_suspend(struct hv_device *hv_dev)
2055{
2056 struct hv_dynmem_device *dm = hv_get_drvdata(hv_dev);
2057
2058 tasklet_disable(&hv_dev->channel->callback_event);
2059
2060 cancel_work_sync(&dm->balloon_wrk.wrk);
2061 cancel_work_sync(&dm->ha_wrk.wrk);
2062
2063 if (dm->thread) {
2064 kthread_stop(dm->thread);
2065 dm->thread = NULL;
2066 vmbus_close(hv_dev->channel);
2067 }
2068
2069 tasklet_enable(&hv_dev->channel->callback_event);
2070
2071 return 0;
2072}
2073
2074static int balloon_resume(struct hv_device *dev)
2075{
2076 int ret;
2077
2078 dm_device.state = DM_INITIALIZING;
2079
2080 ret = balloon_connect_vsp(dev);
2081
2082 if (ret != 0)
2083 goto out;
2084
2085 dm_device.thread =
2086 kthread_run(dm_thread_func, &dm_device, "hv_balloon");
2087 if (IS_ERR(dm_device.thread)) {
2088 ret = PTR_ERR(dm_device.thread);
2089 dm_device.thread = NULL;
2090 goto close_channel;
2091 }
2092
2093 dm_device.state = DM_INITIALIZED;
2094 return 0;
2095close_channel:
2096 vmbus_close(dev->channel);
2097out:
2098 dm_device.state = DM_INIT_ERROR;
2099 disable_page_reporting();
2100#ifdef CONFIG_MEMORY_HOTPLUG
2101 unregister_memory_notifier(&hv_memory_nb);
2102 restore_online_page_callback(&hv_online_page);
2103#endif
2104 return ret;
2105}
2106
2107static const struct hv_vmbus_device_id id_table[] = {
2108 /* Dynamic Memory Class ID */
2109 /* 525074DC-8985-46e2-8057-A307DC18A502 */
2110 { HV_DM_GUID, },
2111 { },
2112};
2113
2114MODULE_DEVICE_TABLE(vmbus, id_table);
2115
2116static struct hv_driver balloon_drv = {
2117 .name = "hv_balloon",
2118 .id_table = id_table,
2119 .probe = balloon_probe,
2120 .remove = balloon_remove,
2121 .suspend = balloon_suspend,
2122 .resume = balloon_resume,
2123 .driver = {
2124 .probe_type = PROBE_PREFER_ASYNCHRONOUS,
2125 },
2126};
2127
2128static int __init init_balloon_drv(void)
2129{
2130 return vmbus_driver_register(&balloon_drv);
2131}
2132
2133module_init(init_balloon_drv);
2134
2135MODULE_DESCRIPTION("Hyper-V Balloon");
2136MODULE_LICENSE("GPL");