Loading...
1/******************************************************************************
2 * Xen balloon driver - enables returning/claiming memory to/from Xen.
3 *
4 * Copyright (c) 2003, B Dragovic
5 * Copyright (c) 2003-2004, M Williamson, K Fraser
6 * Copyright (c) 2005 Dan M. Smith, IBM Corporation
7 * Copyright (c) 2010 Daniel Kiper
8 *
9 * Memory hotplug support was written by Daniel Kiper. Work on
10 * it was sponsored by Google under Google Summer of Code 2010
11 * program. Jeremy Fitzhardinge from Citrix was the mentor for
12 * this project.
13 *
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public License version 2
16 * as published by the Free Software Foundation; or, when distributed
17 * separately from the Linux kernel or incorporated into other
18 * software packages, subject to the following license:
19 *
20 * Permission is hereby granted, free of charge, to any person obtaining a copy
21 * of this source file (the "Software"), to deal in the Software without
22 * restriction, including without limitation the rights to use, copy, modify,
23 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
24 * and to permit persons to whom the Software is furnished to do so, subject to
25 * the following conditions:
26 *
27 * The above copyright notice and this permission notice shall be included in
28 * all copies or substantial portions of the Software.
29 *
30 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
31 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
32 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
33 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
34 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
35 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
36 * IN THE SOFTWARE.
37 */
38
39#include <linux/kernel.h>
40#include <linux/sched.h>
41#include <linux/errno.h>
42#include <linux/module.h>
43#include <linux/mm.h>
44#include <linux/bootmem.h>
45#include <linux/pagemap.h>
46#include <linux/highmem.h>
47#include <linux/mutex.h>
48#include <linux/list.h>
49#include <linux/gfp.h>
50#include <linux/notifier.h>
51#include <linux/memory.h>
52#include <linux/memory_hotplug.h>
53
54#include <asm/page.h>
55#include <asm/pgalloc.h>
56#include <asm/pgtable.h>
57#include <asm/tlb.h>
58#include <asm/e820.h>
59
60#include <asm/xen/hypervisor.h>
61#include <asm/xen/hypercall.h>
62
63#include <xen/xen.h>
64#include <xen/interface/xen.h>
65#include <xen/interface/memory.h>
66#include <xen/balloon.h>
67#include <xen/features.h>
68#include <xen/page.h>
69
70/*
71 * balloon_process() state:
72 *
73 * BP_DONE: done or nothing to do,
74 * BP_EAGAIN: error, go to sleep,
75 * BP_ECANCELED: error, balloon operation canceled.
76 */
77
78enum bp_state {
79 BP_DONE,
80 BP_EAGAIN,
81 BP_ECANCELED
82};
83
84
85static DEFINE_MUTEX(balloon_mutex);
86
87struct balloon_stats balloon_stats;
88EXPORT_SYMBOL_GPL(balloon_stats);
89
90/* We increase/decrease in batches which fit in a page */
91static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
92
93#ifdef CONFIG_HIGHMEM
94#define inc_totalhigh_pages() (totalhigh_pages++)
95#define dec_totalhigh_pages() (totalhigh_pages--)
96#else
97#define inc_totalhigh_pages() do {} while (0)
98#define dec_totalhigh_pages() do {} while (0)
99#endif
100
101/* List of ballooned pages, threaded through the mem_map array. */
102static LIST_HEAD(ballooned_pages);
103
104/* Main work function, always executed in process context. */
105static void balloon_process(struct work_struct *work);
106static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
107
108/* When ballooning out (allocating memory to return to Xen) we don't really
109 want the kernel to try too hard since that can trigger the oom killer. */
110#define GFP_BALLOON \
111 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
112
113static void scrub_page(struct page *page)
114{
115#ifdef CONFIG_XEN_SCRUB_PAGES
116 clear_highpage(page);
117#endif
118}
119
120/* balloon_append: add the given page to the balloon. */
121static void __balloon_append(struct page *page)
122{
123 /* Lowmem is re-populated first, so highmem pages go at list tail. */
124 if (PageHighMem(page)) {
125 list_add_tail(&page->lru, &ballooned_pages);
126 balloon_stats.balloon_high++;
127 } else {
128 list_add(&page->lru, &ballooned_pages);
129 balloon_stats.balloon_low++;
130 }
131}
132
133static void balloon_append(struct page *page)
134{
135 __balloon_append(page);
136 if (PageHighMem(page))
137 dec_totalhigh_pages();
138 totalram_pages--;
139}
140
141/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
142static struct page *balloon_retrieve(bool prefer_highmem)
143{
144 struct page *page;
145
146 if (list_empty(&ballooned_pages))
147 return NULL;
148
149 if (prefer_highmem)
150 page = list_entry(ballooned_pages.prev, struct page, lru);
151 else
152 page = list_entry(ballooned_pages.next, struct page, lru);
153 list_del(&page->lru);
154
155 if (PageHighMem(page)) {
156 balloon_stats.balloon_high--;
157 inc_totalhigh_pages();
158 } else
159 balloon_stats.balloon_low--;
160
161 totalram_pages++;
162
163 return page;
164}
165
166static struct page *balloon_first_page(void)
167{
168 if (list_empty(&ballooned_pages))
169 return NULL;
170 return list_entry(ballooned_pages.next, struct page, lru);
171}
172
173static struct page *balloon_next_page(struct page *page)
174{
175 struct list_head *next = page->lru.next;
176 if (next == &ballooned_pages)
177 return NULL;
178 return list_entry(next, struct page, lru);
179}
180
181static enum bp_state update_schedule(enum bp_state state)
182{
183 if (state == BP_DONE) {
184 balloon_stats.schedule_delay = 1;
185 balloon_stats.retry_count = 1;
186 return BP_DONE;
187 }
188
189 ++balloon_stats.retry_count;
190
191 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
192 balloon_stats.retry_count > balloon_stats.max_retry_count) {
193 balloon_stats.schedule_delay = 1;
194 balloon_stats.retry_count = 1;
195 return BP_ECANCELED;
196 }
197
198 balloon_stats.schedule_delay <<= 1;
199
200 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
201 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
202
203 return BP_EAGAIN;
204}
205
206#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
207static long current_credit(void)
208{
209 return balloon_stats.target_pages - balloon_stats.current_pages -
210 balloon_stats.hotplug_pages;
211}
212
213static bool balloon_is_inflated(void)
214{
215 if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
216 balloon_stats.balloon_hotplug)
217 return true;
218 else
219 return false;
220}
221
222/*
223 * reserve_additional_memory() adds memory region of size >= credit above
224 * max_pfn. New region is section aligned and size is modified to be multiple
225 * of section size. Those features allow optimal use of address space and
226 * establish proper alignment when this function is called first time after
227 * boot (last section not fully populated at boot time contains unused memory
228 * pages with PG_reserved bit not set; online_pages_range() does not allow page
229 * onlining in whole range if first onlined page does not have PG_reserved
230 * bit set). Real size of added memory is established at page onlining stage.
231 */
232
233static enum bp_state reserve_additional_memory(long credit)
234{
235 int nid, rc;
236 u64 hotplug_start_paddr;
237 unsigned long balloon_hotplug = credit;
238
239 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
240 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
241 nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
242
243 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
244
245 if (rc) {
246 pr_info("xen_balloon: %s: add_memory() failed: %i\n", __func__, rc);
247 return BP_EAGAIN;
248 }
249
250 balloon_hotplug -= credit;
251
252 balloon_stats.hotplug_pages += credit;
253 balloon_stats.balloon_hotplug = balloon_hotplug;
254
255 return BP_DONE;
256}
257
258static void xen_online_page(struct page *page)
259{
260 __online_page_set_limits(page);
261
262 mutex_lock(&balloon_mutex);
263
264 __balloon_append(page);
265
266 if (balloon_stats.hotplug_pages)
267 --balloon_stats.hotplug_pages;
268 else
269 --balloon_stats.balloon_hotplug;
270
271 mutex_unlock(&balloon_mutex);
272}
273
274static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
275{
276 if (val == MEM_ONLINE)
277 schedule_delayed_work(&balloon_worker, 0);
278
279 return NOTIFY_OK;
280}
281
282static struct notifier_block xen_memory_nb = {
283 .notifier_call = xen_memory_notifier,
284 .priority = 0
285};
286#else
287static long current_credit(void)
288{
289 unsigned long target = balloon_stats.target_pages;
290
291 target = min(target,
292 balloon_stats.current_pages +
293 balloon_stats.balloon_low +
294 balloon_stats.balloon_high);
295
296 return target - balloon_stats.current_pages;
297}
298
299static bool balloon_is_inflated(void)
300{
301 if (balloon_stats.balloon_low || balloon_stats.balloon_high)
302 return true;
303 else
304 return false;
305}
306
307static enum bp_state reserve_additional_memory(long credit)
308{
309 balloon_stats.target_pages = balloon_stats.current_pages;
310 return BP_DONE;
311}
312#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
313
314static enum bp_state increase_reservation(unsigned long nr_pages)
315{
316 int rc;
317 unsigned long pfn, i;
318 struct page *page;
319 struct xen_memory_reservation reservation = {
320 .address_bits = 0,
321 .extent_order = 0,
322 .domid = DOMID_SELF
323 };
324
325#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
326 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
327 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
328 balloon_stats.hotplug_pages += nr_pages;
329 balloon_stats.balloon_hotplug -= nr_pages;
330 return BP_DONE;
331 }
332#endif
333
334 if (nr_pages > ARRAY_SIZE(frame_list))
335 nr_pages = ARRAY_SIZE(frame_list);
336
337 page = balloon_first_page();
338 for (i = 0; i < nr_pages; i++) {
339 if (!page) {
340 nr_pages = i;
341 break;
342 }
343 frame_list[i] = page_to_pfn(page);
344 page = balloon_next_page(page);
345 }
346
347 set_xen_guest_handle(reservation.extent_start, frame_list);
348 reservation.nr_extents = nr_pages;
349 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
350 if (rc <= 0)
351 return BP_EAGAIN;
352
353 for (i = 0; i < rc; i++) {
354 page = balloon_retrieve(false);
355 BUG_ON(page == NULL);
356
357 pfn = page_to_pfn(page);
358 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
359 phys_to_machine_mapping_valid(pfn));
360
361 set_phys_to_machine(pfn, frame_list[i]);
362
363 /* Link back into the page tables if not highmem. */
364 if (xen_pv_domain() && !PageHighMem(page)) {
365 int ret;
366 ret = HYPERVISOR_update_va_mapping(
367 (unsigned long)__va(pfn << PAGE_SHIFT),
368 mfn_pte(frame_list[i], PAGE_KERNEL),
369 0);
370 BUG_ON(ret);
371 }
372
373 /* Relinquish the page back to the allocator. */
374 ClearPageReserved(page);
375 init_page_count(page);
376 __free_page(page);
377 }
378
379 balloon_stats.current_pages += rc;
380
381 return BP_DONE;
382}
383
384static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
385{
386 enum bp_state state = BP_DONE;
387 unsigned long pfn, i;
388 struct page *page;
389 int ret;
390 struct xen_memory_reservation reservation = {
391 .address_bits = 0,
392 .extent_order = 0,
393 .domid = DOMID_SELF
394 };
395
396#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
397 if (balloon_stats.hotplug_pages) {
398 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
399 balloon_stats.hotplug_pages -= nr_pages;
400 balloon_stats.balloon_hotplug += nr_pages;
401 return BP_DONE;
402 }
403#endif
404
405 if (nr_pages > ARRAY_SIZE(frame_list))
406 nr_pages = ARRAY_SIZE(frame_list);
407
408 for (i = 0; i < nr_pages; i++) {
409 if ((page = alloc_page(gfp)) == NULL) {
410 nr_pages = i;
411 state = BP_EAGAIN;
412 break;
413 }
414
415 pfn = page_to_pfn(page);
416 frame_list[i] = pfn_to_mfn(pfn);
417
418 scrub_page(page);
419
420 if (xen_pv_domain() && !PageHighMem(page)) {
421 ret = HYPERVISOR_update_va_mapping(
422 (unsigned long)__va(pfn << PAGE_SHIFT),
423 __pte_ma(0), 0);
424 BUG_ON(ret);
425 }
426
427 }
428
429 /* Ensure that ballooned highmem pages don't have kmaps. */
430 kmap_flush_unused();
431 flush_tlb_all();
432
433 /* No more mappings: invalidate P2M and add to balloon. */
434 for (i = 0; i < nr_pages; i++) {
435 pfn = mfn_to_pfn(frame_list[i]);
436 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
437 balloon_append(pfn_to_page(pfn));
438 }
439
440 set_xen_guest_handle(reservation.extent_start, frame_list);
441 reservation.nr_extents = nr_pages;
442 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
443 BUG_ON(ret != nr_pages);
444
445 balloon_stats.current_pages -= nr_pages;
446
447 return state;
448}
449
450/*
451 * We avoid multiple worker processes conflicting via the balloon mutex.
452 * We may of course race updates of the target counts (which are protected
453 * by the balloon lock), or with changes to the Xen hard limit, but we will
454 * recover from these in time.
455 */
456static void balloon_process(struct work_struct *work)
457{
458 enum bp_state state = BP_DONE;
459 long credit;
460
461 mutex_lock(&balloon_mutex);
462
463 do {
464 credit = current_credit();
465
466 if (credit > 0) {
467 if (balloon_is_inflated())
468 state = increase_reservation(credit);
469 else
470 state = reserve_additional_memory(credit);
471 }
472
473 if (credit < 0)
474 state = decrease_reservation(-credit, GFP_BALLOON);
475
476 state = update_schedule(state);
477
478#ifndef CONFIG_PREEMPT
479 if (need_resched())
480 schedule();
481#endif
482 } while (credit && state == BP_DONE);
483
484 /* Schedule more work if there is some still to be done. */
485 if (state == BP_EAGAIN)
486 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
487
488 mutex_unlock(&balloon_mutex);
489}
490
491/* Resets the Xen limit, sets new target, and kicks off processing. */
492void balloon_set_new_target(unsigned long target)
493{
494 /* No need for lock. Not read-modify-write updates. */
495 balloon_stats.target_pages = target;
496 schedule_delayed_work(&balloon_worker, 0);
497}
498EXPORT_SYMBOL_GPL(balloon_set_new_target);
499
500/**
501 * alloc_xenballooned_pages - get pages that have been ballooned out
502 * @nr_pages: Number of pages to get
503 * @pages: pages returned
504 * @highmem: allow highmem pages
505 * @return 0 on success, error otherwise
506 */
507int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
508{
509 int pgno = 0;
510 struct page *page;
511 mutex_lock(&balloon_mutex);
512 while (pgno < nr_pages) {
513 page = balloon_retrieve(highmem);
514 if (page && (highmem || !PageHighMem(page))) {
515 pages[pgno++] = page;
516 } else {
517 enum bp_state st;
518 if (page)
519 balloon_append(page);
520 st = decrease_reservation(nr_pages - pgno,
521 highmem ? GFP_HIGHUSER : GFP_USER);
522 if (st != BP_DONE)
523 goto out_undo;
524 }
525 }
526 mutex_unlock(&balloon_mutex);
527 return 0;
528 out_undo:
529 while (pgno)
530 balloon_append(pages[--pgno]);
531 /* Free the memory back to the kernel soon */
532 schedule_delayed_work(&balloon_worker, 0);
533 mutex_unlock(&balloon_mutex);
534 return -ENOMEM;
535}
536EXPORT_SYMBOL(alloc_xenballooned_pages);
537
538/**
539 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
540 * @nr_pages: Number of pages
541 * @pages: pages to return
542 */
543void free_xenballooned_pages(int nr_pages, struct page **pages)
544{
545 int i;
546
547 mutex_lock(&balloon_mutex);
548
549 for (i = 0; i < nr_pages; i++) {
550 if (pages[i])
551 balloon_append(pages[i]);
552 }
553
554 /* The balloon may be too large now. Shrink it if needed. */
555 if (current_credit())
556 schedule_delayed_work(&balloon_worker, 0);
557
558 mutex_unlock(&balloon_mutex);
559}
560EXPORT_SYMBOL(free_xenballooned_pages);
561
562static void __init balloon_add_region(unsigned long start_pfn,
563 unsigned long pages)
564{
565 unsigned long pfn, extra_pfn_end;
566 struct page *page;
567
568 /*
569 * If the amount of usable memory has been limited (e.g., with
570 * the 'mem' command line parameter), don't add pages beyond
571 * this limit.
572 */
573 extra_pfn_end = min(max_pfn, start_pfn + pages);
574
575 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
576 page = pfn_to_page(pfn);
577 /* totalram_pages and totalhigh_pages do not
578 include the boot-time balloon extension, so
579 don't subtract from it. */
580 __balloon_append(page);
581 }
582}
583
584static int __init balloon_init(void)
585{
586 int i;
587
588 if (!xen_domain())
589 return -ENODEV;
590
591 pr_info("xen/balloon: Initialising balloon driver.\n");
592
593 balloon_stats.current_pages = xen_pv_domain()
594 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
595 : max_pfn;
596 balloon_stats.target_pages = balloon_stats.current_pages;
597 balloon_stats.balloon_low = 0;
598 balloon_stats.balloon_high = 0;
599
600 balloon_stats.schedule_delay = 1;
601 balloon_stats.max_schedule_delay = 32;
602 balloon_stats.retry_count = 1;
603 balloon_stats.max_retry_count = RETRY_UNLIMITED;
604
605#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
606 balloon_stats.hotplug_pages = 0;
607 balloon_stats.balloon_hotplug = 0;
608
609 set_online_page_callback(&xen_online_page);
610 register_memory_notifier(&xen_memory_nb);
611#endif
612
613 /*
614 * Initialize the balloon with pages from the extra memory
615 * regions (see arch/x86/xen/setup.c).
616 */
617 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
618 if (xen_extra_mem[i].size)
619 balloon_add_region(PFN_UP(xen_extra_mem[i].start),
620 PFN_DOWN(xen_extra_mem[i].size));
621
622 return 0;
623}
624
625subsys_initcall(balloon_init);
626
627MODULE_LICENSE("GPL");
1/******************************************************************************
2 * Xen balloon driver - enables returning/claiming memory to/from Xen.
3 *
4 * Copyright (c) 2003, B Dragovic
5 * Copyright (c) 2003-2004, M Williamson, K Fraser
6 * Copyright (c) 2005 Dan M. Smith, IBM Corporation
7 * Copyright (c) 2010 Daniel Kiper
8 *
9 * Memory hotplug support was written by Daniel Kiper. Work on
10 * it was sponsored by Google under Google Summer of Code 2010
11 * program. Jeremy Fitzhardinge from Citrix was the mentor for
12 * this project.
13 *
14 * This program is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU General Public License version 2
16 * as published by the Free Software Foundation; or, when distributed
17 * separately from the Linux kernel or incorporated into other
18 * software packages, subject to the following license:
19 *
20 * Permission is hereby granted, free of charge, to any person obtaining a copy
21 * of this source file (the "Software"), to deal in the Software without
22 * restriction, including without limitation the rights to use, copy, modify,
23 * merge, publish, distribute, sublicense, and/or sell copies of the Software,
24 * and to permit persons to whom the Software is furnished to do so, subject to
25 * the following conditions:
26 *
27 * The above copyright notice and this permission notice shall be included in
28 * all copies or substantial portions of the Software.
29 *
30 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
31 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
32 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
33 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
34 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
35 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
36 * IN THE SOFTWARE.
37 */
38
39#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
40
41#include <linux/cpu.h>
42#include <linux/kernel.h>
43#include <linux/sched.h>
44#include <linux/cred.h>
45#include <linux/errno.h>
46#include <linux/mm.h>
47#include <linux/memblock.h>
48#include <linux/pagemap.h>
49#include <linux/highmem.h>
50#include <linux/mutex.h>
51#include <linux/list.h>
52#include <linux/gfp.h>
53#include <linux/notifier.h>
54#include <linux/memory.h>
55#include <linux/memory_hotplug.h>
56#include <linux/percpu-defs.h>
57#include <linux/slab.h>
58#include <linux/sysctl.h>
59
60#include <asm/page.h>
61#include <asm/pgalloc.h>
62#include <asm/pgtable.h>
63#include <asm/tlb.h>
64
65#include <asm/xen/hypervisor.h>
66#include <asm/xen/hypercall.h>
67
68#include <xen/xen.h>
69#include <xen/interface/xen.h>
70#include <xen/interface/memory.h>
71#include <xen/balloon.h>
72#include <xen/features.h>
73#include <xen/page.h>
74#include <xen/mem-reservation.h>
75
76static int xen_hotplug_unpopulated;
77
78#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
79
80static struct ctl_table balloon_table[] = {
81 {
82 .procname = "hotplug_unpopulated",
83 .data = &xen_hotplug_unpopulated,
84 .maxlen = sizeof(int),
85 .mode = 0644,
86 .proc_handler = proc_dointvec_minmax,
87 .extra1 = SYSCTL_ZERO,
88 .extra2 = SYSCTL_ONE,
89 },
90 { }
91};
92
93static struct ctl_table balloon_root[] = {
94 {
95 .procname = "balloon",
96 .mode = 0555,
97 .child = balloon_table,
98 },
99 { }
100};
101
102static struct ctl_table xen_root[] = {
103 {
104 .procname = "xen",
105 .mode = 0555,
106 .child = balloon_root,
107 },
108 { }
109};
110
111#endif
112
113/*
114 * Use one extent per PAGE_SIZE to avoid to break down the page into
115 * multiple frame.
116 */
117#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
118
119/*
120 * balloon_process() state:
121 *
122 * BP_DONE: done or nothing to do,
123 * BP_WAIT: wait to be rescheduled,
124 * BP_EAGAIN: error, go to sleep,
125 * BP_ECANCELED: error, balloon operation canceled.
126 */
127
128enum bp_state {
129 BP_DONE,
130 BP_WAIT,
131 BP_EAGAIN,
132 BP_ECANCELED
133};
134
135
136static DEFINE_MUTEX(balloon_mutex);
137
138struct balloon_stats balloon_stats;
139EXPORT_SYMBOL_GPL(balloon_stats);
140
141/* We increase/decrease in batches which fit in a page */
142static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
143
144
145/* List of ballooned pages, threaded through the mem_map array. */
146static LIST_HEAD(ballooned_pages);
147static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
148
149/* Main work function, always executed in process context. */
150static void balloon_process(struct work_struct *work);
151static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
152
153/* When ballooning out (allocating memory to return to Xen) we don't really
154 want the kernel to try too hard since that can trigger the oom killer. */
155#define GFP_BALLOON \
156 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
157
158/* balloon_append: add the given page to the balloon. */
159static void balloon_append(struct page *page)
160{
161 __SetPageOffline(page);
162
163 /* Lowmem is re-populated first, so highmem pages go at list tail. */
164 if (PageHighMem(page)) {
165 list_add_tail(&page->lru, &ballooned_pages);
166 balloon_stats.balloon_high++;
167 } else {
168 list_add(&page->lru, &ballooned_pages);
169 balloon_stats.balloon_low++;
170 }
171 wake_up(&balloon_wq);
172}
173
174/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
175static struct page *balloon_retrieve(bool require_lowmem)
176{
177 struct page *page;
178
179 if (list_empty(&ballooned_pages))
180 return NULL;
181
182 page = list_entry(ballooned_pages.next, struct page, lru);
183 if (require_lowmem && PageHighMem(page))
184 return NULL;
185 list_del(&page->lru);
186
187 if (PageHighMem(page))
188 balloon_stats.balloon_high--;
189 else
190 balloon_stats.balloon_low--;
191
192 __ClearPageOffline(page);
193 return page;
194}
195
196static struct page *balloon_next_page(struct page *page)
197{
198 struct list_head *next = page->lru.next;
199 if (next == &ballooned_pages)
200 return NULL;
201 return list_entry(next, struct page, lru);
202}
203
204static enum bp_state update_schedule(enum bp_state state)
205{
206 if (state == BP_WAIT)
207 return BP_WAIT;
208
209 if (state == BP_ECANCELED)
210 return BP_ECANCELED;
211
212 if (state == BP_DONE) {
213 balloon_stats.schedule_delay = 1;
214 balloon_stats.retry_count = 1;
215 return BP_DONE;
216 }
217
218 ++balloon_stats.retry_count;
219
220 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
221 balloon_stats.retry_count > balloon_stats.max_retry_count) {
222 balloon_stats.schedule_delay = 1;
223 balloon_stats.retry_count = 1;
224 return BP_ECANCELED;
225 }
226
227 balloon_stats.schedule_delay <<= 1;
228
229 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
230 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
231
232 return BP_EAGAIN;
233}
234
235#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
236static void release_memory_resource(struct resource *resource)
237{
238 if (!resource)
239 return;
240
241 /*
242 * No need to reset region to identity mapped since we now
243 * know that no I/O can be in this region
244 */
245 release_resource(resource);
246 kfree(resource);
247}
248
249static struct resource *additional_memory_resource(phys_addr_t size)
250{
251 struct resource *res;
252 int ret;
253
254 res = kzalloc(sizeof(*res), GFP_KERNEL);
255 if (!res)
256 return NULL;
257
258 res->name = "System RAM";
259 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
260
261 ret = allocate_resource(&iomem_resource, res,
262 size, 0, -1,
263 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
264 if (ret < 0) {
265 pr_err("Cannot allocate new System RAM resource\n");
266 kfree(res);
267 return NULL;
268 }
269
270#ifdef CONFIG_SPARSEMEM
271 {
272 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
273 unsigned long pfn = res->start >> PAGE_SHIFT;
274
275 if (pfn > limit) {
276 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
277 pfn, limit);
278 release_memory_resource(res);
279 return NULL;
280 }
281 }
282#endif
283
284 return res;
285}
286
287static enum bp_state reserve_additional_memory(void)
288{
289 long credit;
290 struct resource *resource;
291 int nid, rc;
292 unsigned long balloon_hotplug;
293
294 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
295 - balloon_stats.total_pages;
296
297 /*
298 * Already hotplugged enough pages? Wait for them to be
299 * onlined.
300 */
301 if (credit <= 0)
302 return BP_WAIT;
303
304 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
305
306 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
307 if (!resource)
308 goto err;
309
310 nid = memory_add_physaddr_to_nid(resource->start);
311
312#ifdef CONFIG_XEN_HAVE_PVMMU
313 /*
314 * We don't support PV MMU when Linux and Xen is using
315 * different page granularity.
316 */
317 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
318
319 /*
320 * add_memory() will build page tables for the new memory so
321 * the p2m must contain invalid entries so the correct
322 * non-present PTEs will be written.
323 *
324 * If a failure occurs, the original (identity) p2m entries
325 * are not restored since this region is now known not to
326 * conflict with any devices.
327 */
328 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
329 unsigned long pfn, i;
330
331 pfn = PFN_DOWN(resource->start);
332 for (i = 0; i < balloon_hotplug; i++) {
333 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
334 pr_warn("set_phys_to_machine() failed, no memory added\n");
335 goto err;
336 }
337 }
338 }
339#endif
340
341 /*
342 * add_memory_resource() will call online_pages() which in its turn
343 * will call xen_online_page() callback causing deadlock if we don't
344 * release balloon_mutex here. Unlocking here is safe because the
345 * callers drop the mutex before trying again.
346 */
347 mutex_unlock(&balloon_mutex);
348 /* add_memory_resource() requires the device_hotplug lock */
349 lock_device_hotplug();
350 rc = add_memory_resource(nid, resource);
351 unlock_device_hotplug();
352 mutex_lock(&balloon_mutex);
353
354 if (rc) {
355 pr_warn("Cannot add additional memory (%i)\n", rc);
356 goto err;
357 }
358
359 balloon_stats.total_pages += balloon_hotplug;
360
361 return BP_WAIT;
362 err:
363 release_memory_resource(resource);
364 return BP_ECANCELED;
365}
366
367static void xen_online_page(struct page *page, unsigned int order)
368{
369 unsigned long i, size = (1 << order);
370 unsigned long start_pfn = page_to_pfn(page);
371 struct page *p;
372
373 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
374 mutex_lock(&balloon_mutex);
375 for (i = 0; i < size; i++) {
376 p = pfn_to_page(start_pfn + i);
377 __online_page_set_limits(p);
378 balloon_append(p);
379 }
380 mutex_unlock(&balloon_mutex);
381}
382
383static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
384{
385 if (val == MEM_ONLINE)
386 schedule_delayed_work(&balloon_worker, 0);
387
388 return NOTIFY_OK;
389}
390
391static struct notifier_block xen_memory_nb = {
392 .notifier_call = xen_memory_notifier,
393 .priority = 0
394};
395#else
396static enum bp_state reserve_additional_memory(void)
397{
398 balloon_stats.target_pages = balloon_stats.current_pages;
399 return BP_ECANCELED;
400}
401#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
402
403static long current_credit(void)
404{
405 return balloon_stats.target_pages - balloon_stats.current_pages;
406}
407
408static bool balloon_is_inflated(void)
409{
410 return balloon_stats.balloon_low || balloon_stats.balloon_high;
411}
412
413static enum bp_state increase_reservation(unsigned long nr_pages)
414{
415 int rc;
416 unsigned long i;
417 struct page *page;
418
419 if (nr_pages > ARRAY_SIZE(frame_list))
420 nr_pages = ARRAY_SIZE(frame_list);
421
422 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
423 for (i = 0; i < nr_pages; i++) {
424 if (!page) {
425 nr_pages = i;
426 break;
427 }
428
429 frame_list[i] = page_to_xen_pfn(page);
430 page = balloon_next_page(page);
431 }
432
433 rc = xenmem_reservation_increase(nr_pages, frame_list);
434 if (rc <= 0)
435 return BP_EAGAIN;
436
437 for (i = 0; i < rc; i++) {
438 page = balloon_retrieve(false);
439 BUG_ON(page == NULL);
440
441 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
442
443 /* Relinquish the page back to the allocator. */
444 free_reserved_page(page);
445 }
446
447 balloon_stats.current_pages += rc;
448
449 return BP_DONE;
450}
451
452static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
453{
454 enum bp_state state = BP_DONE;
455 unsigned long i;
456 struct page *page, *tmp;
457 int ret;
458 LIST_HEAD(pages);
459
460 if (nr_pages > ARRAY_SIZE(frame_list))
461 nr_pages = ARRAY_SIZE(frame_list);
462
463 for (i = 0; i < nr_pages; i++) {
464 page = alloc_page(gfp);
465 if (page == NULL) {
466 nr_pages = i;
467 state = BP_EAGAIN;
468 break;
469 }
470 adjust_managed_page_count(page, -1);
471 xenmem_reservation_scrub_page(page);
472 list_add(&page->lru, &pages);
473 }
474
475 /*
476 * Ensure that ballooned highmem pages don't have kmaps.
477 *
478 * Do this before changing the p2m as kmap_flush_unused()
479 * reads PTEs to obtain pages (and hence needs the original
480 * p2m entry).
481 */
482 kmap_flush_unused();
483
484 /*
485 * Setup the frame, update direct mapping, invalidate P2M,
486 * and add to balloon.
487 */
488 i = 0;
489 list_for_each_entry_safe(page, tmp, &pages, lru) {
490 frame_list[i++] = xen_page_to_gfn(page);
491
492 xenmem_reservation_va_mapping_reset(1, &page);
493
494 list_del(&page->lru);
495
496 balloon_append(page);
497 }
498
499 flush_tlb_all();
500
501 ret = xenmem_reservation_decrease(nr_pages, frame_list);
502 BUG_ON(ret != nr_pages);
503
504 balloon_stats.current_pages -= nr_pages;
505
506 return state;
507}
508
509/*
510 * As this is a work item it is guaranteed to run as a single instance only.
511 * We may of course race updates of the target counts (which are protected
512 * by the balloon lock), or with changes to the Xen hard limit, but we will
513 * recover from these in time.
514 */
515static void balloon_process(struct work_struct *work)
516{
517 enum bp_state state = BP_DONE;
518 long credit;
519
520
521 do {
522 mutex_lock(&balloon_mutex);
523
524 credit = current_credit();
525
526 if (credit > 0) {
527 if (balloon_is_inflated())
528 state = increase_reservation(credit);
529 else
530 state = reserve_additional_memory();
531 }
532
533 if (credit < 0) {
534 long n_pages;
535
536 n_pages = min(-credit, si_mem_available());
537 state = decrease_reservation(n_pages, GFP_BALLOON);
538 if (state == BP_DONE && n_pages != -credit &&
539 n_pages < totalreserve_pages)
540 state = BP_EAGAIN;
541 }
542
543 state = update_schedule(state);
544
545 mutex_unlock(&balloon_mutex);
546
547 cond_resched();
548
549 } while (credit && state == BP_DONE);
550
551 /* Schedule more work if there is some still to be done. */
552 if (state == BP_EAGAIN)
553 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
554}
555
556/* Resets the Xen limit, sets new target, and kicks off processing. */
557void balloon_set_new_target(unsigned long target)
558{
559 /* No need for lock. Not read-modify-write updates. */
560 balloon_stats.target_pages = target;
561 schedule_delayed_work(&balloon_worker, 0);
562}
563EXPORT_SYMBOL_GPL(balloon_set_new_target);
564
565static int add_ballooned_pages(int nr_pages)
566{
567 enum bp_state st;
568
569 if (xen_hotplug_unpopulated) {
570 st = reserve_additional_memory();
571 if (st != BP_ECANCELED) {
572 mutex_unlock(&balloon_mutex);
573 wait_event(balloon_wq,
574 !list_empty(&ballooned_pages));
575 mutex_lock(&balloon_mutex);
576 return 0;
577 }
578 }
579
580 if (si_mem_available() < nr_pages)
581 return -ENOMEM;
582
583 st = decrease_reservation(nr_pages, GFP_USER);
584 if (st != BP_DONE)
585 return -ENOMEM;
586
587 return 0;
588}
589
590/**
591 * alloc_xenballooned_pages - get pages that have been ballooned out
592 * @nr_pages: Number of pages to get
593 * @pages: pages returned
594 * @return 0 on success, error otherwise
595 */
596int alloc_xenballooned_pages(int nr_pages, struct page **pages)
597{
598 int pgno = 0;
599 struct page *page;
600 int ret;
601
602 mutex_lock(&balloon_mutex);
603
604 balloon_stats.target_unpopulated += nr_pages;
605
606 while (pgno < nr_pages) {
607 page = balloon_retrieve(true);
608 if (page) {
609 pages[pgno++] = page;
610#ifdef CONFIG_XEN_HAVE_PVMMU
611 /*
612 * We don't support PV MMU when Linux and Xen is using
613 * different page granularity.
614 */
615 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
616
617 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
618 ret = xen_alloc_p2m_entry(page_to_pfn(page));
619 if (ret < 0)
620 goto out_undo;
621 }
622#endif
623 } else {
624 ret = add_ballooned_pages(nr_pages - pgno);
625 if (ret < 0)
626 goto out_undo;
627 }
628 }
629 mutex_unlock(&balloon_mutex);
630 return 0;
631 out_undo:
632 mutex_unlock(&balloon_mutex);
633 free_xenballooned_pages(pgno, pages);
634 return ret;
635}
636EXPORT_SYMBOL(alloc_xenballooned_pages);
637
638/**
639 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
640 * @nr_pages: Number of pages
641 * @pages: pages to return
642 */
643void free_xenballooned_pages(int nr_pages, struct page **pages)
644{
645 int i;
646
647 mutex_lock(&balloon_mutex);
648
649 for (i = 0; i < nr_pages; i++) {
650 if (pages[i])
651 balloon_append(pages[i]);
652 }
653
654 balloon_stats.target_unpopulated -= nr_pages;
655
656 /* The balloon may be too large now. Shrink it if needed. */
657 if (current_credit())
658 schedule_delayed_work(&balloon_worker, 0);
659
660 mutex_unlock(&balloon_mutex);
661}
662EXPORT_SYMBOL(free_xenballooned_pages);
663
664#ifdef CONFIG_XEN_PV
665static void __init balloon_add_region(unsigned long start_pfn,
666 unsigned long pages)
667{
668 unsigned long pfn, extra_pfn_end;
669
670 /*
671 * If the amount of usable memory has been limited (e.g., with
672 * the 'mem' command line parameter), don't add pages beyond
673 * this limit.
674 */
675 extra_pfn_end = min(max_pfn, start_pfn + pages);
676
677 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
678 /* totalram_pages and totalhigh_pages do not
679 include the boot-time balloon extension, so
680 don't subtract from it. */
681 balloon_append(pfn_to_page(pfn));
682 }
683
684 balloon_stats.total_pages += extra_pfn_end - start_pfn;
685}
686#endif
687
688static int __init balloon_init(void)
689{
690 if (!xen_domain())
691 return -ENODEV;
692
693 pr_info("Initialising balloon driver\n");
694
695#ifdef CONFIG_XEN_PV
696 balloon_stats.current_pages = xen_pv_domain()
697 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
698 : get_num_physpages();
699#else
700 balloon_stats.current_pages = get_num_physpages();
701#endif
702 balloon_stats.target_pages = balloon_stats.current_pages;
703 balloon_stats.balloon_low = 0;
704 balloon_stats.balloon_high = 0;
705 balloon_stats.total_pages = balloon_stats.current_pages;
706
707 balloon_stats.schedule_delay = 1;
708 balloon_stats.max_schedule_delay = 32;
709 balloon_stats.retry_count = 1;
710 balloon_stats.max_retry_count = 4;
711
712#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
713 set_online_page_callback(&xen_online_page);
714 register_memory_notifier(&xen_memory_nb);
715 register_sysctl_table(xen_root);
716#endif
717
718#ifdef CONFIG_XEN_PV
719 {
720 int i;
721
722 /*
723 * Initialize the balloon with pages from the extra memory
724 * regions (see arch/x86/xen/setup.c).
725 */
726 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
727 if (xen_extra_mem[i].n_pfns)
728 balloon_add_region(xen_extra_mem[i].start_pfn,
729 xen_extra_mem[i].n_pfns);
730 }
731#endif
732
733 /* Init the xen-balloon driver. */
734 xen_balloon_init();
735
736 return 0;
737}
738subsys_initcall(balloon_init);