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/mm.h>
43#include <linux/bootmem.h>
44#include <linux/pagemap.h>
45#include <linux/highmem.h>
46#include <linux/mutex.h>
47#include <linux/list.h>
48#include <linux/gfp.h>
49#include <linux/notifier.h>
50#include <linux/memory.h>
51#include <linux/memory_hotplug.h>
52
53#include <asm/page.h>
54#include <asm/pgalloc.h>
55#include <asm/pgtable.h>
56#include <asm/tlb.h>
57#include <asm/e820.h>
58
59#include <asm/xen/hypervisor.h>
60#include <asm/xen/hypercall.h>
61
62#include <xen/xen.h>
63#include <xen/interface/xen.h>
64#include <xen/interface/memory.h>
65#include <xen/balloon.h>
66#include <xen/features.h>
67#include <xen/page.h>
68
69/*
70 * balloon_process() state:
71 *
72 * BP_DONE: done or nothing to do,
73 * BP_EAGAIN: error, go to sleep,
74 * BP_ECANCELED: error, balloon operation canceled.
75 */
76
77enum bp_state {
78 BP_DONE,
79 BP_EAGAIN,
80 BP_ECANCELED
81};
82
83
84static DEFINE_MUTEX(balloon_mutex);
85
86struct balloon_stats balloon_stats;
87EXPORT_SYMBOL_GPL(balloon_stats);
88
89/* We increase/decrease in batches which fit in a page */
90static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
91
92#ifdef CONFIG_HIGHMEM
93#define inc_totalhigh_pages() (totalhigh_pages++)
94#define dec_totalhigh_pages() (totalhigh_pages--)
95#else
96#define inc_totalhigh_pages() do {} while(0)
97#define dec_totalhigh_pages() do {} while(0)
98#endif
99
100/* List of ballooned pages, threaded through the mem_map array. */
101static LIST_HEAD(ballooned_pages);
102
103/* Main work function, always executed in process context. */
104static void balloon_process(struct work_struct *work);
105static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
106
107/* When ballooning out (allocating memory to return to Xen) we don't really
108 want the kernel to try too hard since that can trigger the oom killer. */
109#define GFP_BALLOON \
110 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
111
112static void scrub_page(struct page *page)
113{
114#ifdef CONFIG_XEN_SCRUB_PAGES
115 clear_highpage(page);
116#endif
117}
118
119/* balloon_append: add the given page to the balloon. */
120static void __balloon_append(struct page *page)
121{
122 /* Lowmem is re-populated first, so highmem pages go at list tail. */
123 if (PageHighMem(page)) {
124 list_add_tail(&page->lru, &ballooned_pages);
125 balloon_stats.balloon_high++;
126 } else {
127 list_add(&page->lru, &ballooned_pages);
128 balloon_stats.balloon_low++;
129 }
130}
131
132static void balloon_append(struct page *page)
133{
134 __balloon_append(page);
135 if (PageHighMem(page))
136 dec_totalhigh_pages();
137 totalram_pages--;
138}
139
140/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
141static struct page *balloon_retrieve(bool prefer_highmem)
142{
143 struct page *page;
144
145 if (list_empty(&ballooned_pages))
146 return NULL;
147
148 if (prefer_highmem)
149 page = list_entry(ballooned_pages.prev, struct page, lru);
150 else
151 page = list_entry(ballooned_pages.next, struct page, lru);
152 list_del(&page->lru);
153
154 if (PageHighMem(page)) {
155 balloon_stats.balloon_high--;
156 inc_totalhigh_pages();
157 }
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 * @return 0 on success, error otherwise
505 */
506int alloc_xenballooned_pages(int nr_pages, struct page** pages)
507{
508 int pgno = 0;
509 struct page* page;
510 mutex_lock(&balloon_mutex);
511 while (pgno < nr_pages) {
512 page = balloon_retrieve(true);
513 if (page) {
514 pages[pgno++] = page;
515 } else {
516 enum bp_state st;
517 st = decrease_reservation(nr_pages - pgno, GFP_HIGHUSER);
518 if (st != BP_DONE)
519 goto out_undo;
520 }
521 }
522 mutex_unlock(&balloon_mutex);
523 return 0;
524 out_undo:
525 while (pgno)
526 balloon_append(pages[--pgno]);
527 /* Free the memory back to the kernel soon */
528 schedule_delayed_work(&balloon_worker, 0);
529 mutex_unlock(&balloon_mutex);
530 return -ENOMEM;
531}
532EXPORT_SYMBOL(alloc_xenballooned_pages);
533
534/**
535 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
536 * @nr_pages: Number of pages
537 * @pages: pages to return
538 */
539void free_xenballooned_pages(int nr_pages, struct page** pages)
540{
541 int i;
542
543 mutex_lock(&balloon_mutex);
544
545 for (i = 0; i < nr_pages; i++) {
546 if (pages[i])
547 balloon_append(pages[i]);
548 }
549
550 /* The balloon may be too large now. Shrink it if needed. */
551 if (current_credit())
552 schedule_delayed_work(&balloon_worker, 0);
553
554 mutex_unlock(&balloon_mutex);
555}
556EXPORT_SYMBOL(free_xenballooned_pages);
557
558static int __init balloon_init(void)
559{
560 unsigned long pfn, extra_pfn_end;
561 struct page *page;
562
563 if (!xen_domain())
564 return -ENODEV;
565
566 pr_info("xen/balloon: Initialising balloon driver.\n");
567
568 balloon_stats.current_pages = xen_pv_domain() ? min(xen_start_info->nr_pages, max_pfn) : max_pfn;
569 balloon_stats.target_pages = balloon_stats.current_pages;
570 balloon_stats.balloon_low = 0;
571 balloon_stats.balloon_high = 0;
572
573 balloon_stats.schedule_delay = 1;
574 balloon_stats.max_schedule_delay = 32;
575 balloon_stats.retry_count = 1;
576 balloon_stats.max_retry_count = RETRY_UNLIMITED;
577
578#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
579 balloon_stats.hotplug_pages = 0;
580 balloon_stats.balloon_hotplug = 0;
581
582 set_online_page_callback(&xen_online_page);
583 register_memory_notifier(&xen_memory_nb);
584#endif
585
586 /*
587 * Initialise the balloon with excess memory space. We need
588 * to make sure we don't add memory which doesn't exist or
589 * logically exist. The E820 map can be trimmed to be smaller
590 * than the amount of physical memory due to the mem= command
591 * line parameter. And if this is a 32-bit non-HIGHMEM kernel
592 * on a system with memory which requires highmem to access,
593 * don't try to use it.
594 */
595 extra_pfn_end = min(min(max_pfn, e820_end_of_ram_pfn()),
596 (unsigned long)PFN_DOWN(xen_extra_mem_start + xen_extra_mem_size));
597 for (pfn = PFN_UP(xen_extra_mem_start);
598 pfn < extra_pfn_end;
599 pfn++) {
600 page = pfn_to_page(pfn);
601 /* totalram_pages and totalhigh_pages do not include the boot-time
602 balloon extension, so don't subtract from it. */
603 __balloon_append(page);
604 }
605
606 return 0;
607}
608
609subsys_initcall(balloon_init);
610
611MODULE_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/bootmem.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
75static int xen_hotplug_unpopulated;
76
77#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
78
79static int zero;
80static int one = 1;
81
82static struct ctl_table balloon_table[] = {
83 {
84 .procname = "hotplug_unpopulated",
85 .data = &xen_hotplug_unpopulated,
86 .maxlen = sizeof(int),
87 .mode = 0644,
88 .proc_handler = proc_dointvec_minmax,
89 .extra1 = &zero,
90 .extra2 = &one,
91 },
92 { }
93};
94
95static struct ctl_table balloon_root[] = {
96 {
97 .procname = "balloon",
98 .mode = 0555,
99 .child = balloon_table,
100 },
101 { }
102};
103
104static struct ctl_table xen_root[] = {
105 {
106 .procname = "xen",
107 .mode = 0555,
108 .child = balloon_root,
109 },
110 { }
111};
112
113#endif
114
115/*
116 * Use one extent per PAGE_SIZE to avoid to break down the page into
117 * multiple frame.
118 */
119#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
120
121/*
122 * balloon_process() state:
123 *
124 * BP_DONE: done or nothing to do,
125 * BP_WAIT: wait to be rescheduled,
126 * BP_EAGAIN: error, go to sleep,
127 * BP_ECANCELED: error, balloon operation canceled.
128 */
129
130enum bp_state {
131 BP_DONE,
132 BP_WAIT,
133 BP_EAGAIN,
134 BP_ECANCELED
135};
136
137
138static DEFINE_MUTEX(balloon_mutex);
139
140struct balloon_stats balloon_stats;
141EXPORT_SYMBOL_GPL(balloon_stats);
142
143/* We increase/decrease in batches which fit in a page */
144static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
145
146
147/* List of ballooned pages, threaded through the mem_map array. */
148static LIST_HEAD(ballooned_pages);
149static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
150
151/* Main work function, always executed in process context. */
152static void balloon_process(struct work_struct *work);
153static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
154
155/* When ballooning out (allocating memory to return to Xen) we don't really
156 want the kernel to try too hard since that can trigger the oom killer. */
157#define GFP_BALLOON \
158 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
159
160static void scrub_page(struct page *page)
161{
162#ifdef CONFIG_XEN_SCRUB_PAGES
163 clear_highpage(page);
164#endif
165}
166
167/* balloon_append: add the given page to the balloon. */
168static void __balloon_append(struct page *page)
169{
170 /* Lowmem is re-populated first, so highmem pages go at list tail. */
171 if (PageHighMem(page)) {
172 list_add_tail(&page->lru, &ballooned_pages);
173 balloon_stats.balloon_high++;
174 } else {
175 list_add(&page->lru, &ballooned_pages);
176 balloon_stats.balloon_low++;
177 }
178 wake_up(&balloon_wq);
179}
180
181static void balloon_append(struct page *page)
182{
183 __balloon_append(page);
184}
185
186/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
187static struct page *balloon_retrieve(bool require_lowmem)
188{
189 struct page *page;
190
191 if (list_empty(&ballooned_pages))
192 return NULL;
193
194 page = list_entry(ballooned_pages.next, struct page, lru);
195 if (require_lowmem && PageHighMem(page))
196 return NULL;
197 list_del(&page->lru);
198
199 if (PageHighMem(page))
200 balloon_stats.balloon_high--;
201 else
202 balloon_stats.balloon_low--;
203
204 return page;
205}
206
207static struct page *balloon_next_page(struct page *page)
208{
209 struct list_head *next = page->lru.next;
210 if (next == &ballooned_pages)
211 return NULL;
212 return list_entry(next, struct page, lru);
213}
214
215static enum bp_state update_schedule(enum bp_state state)
216{
217 if (state == BP_WAIT)
218 return BP_WAIT;
219
220 if (state == BP_ECANCELED)
221 return BP_ECANCELED;
222
223 if (state == BP_DONE) {
224 balloon_stats.schedule_delay = 1;
225 balloon_stats.retry_count = 1;
226 return BP_DONE;
227 }
228
229 ++balloon_stats.retry_count;
230
231 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
232 balloon_stats.retry_count > balloon_stats.max_retry_count) {
233 balloon_stats.schedule_delay = 1;
234 balloon_stats.retry_count = 1;
235 return BP_ECANCELED;
236 }
237
238 balloon_stats.schedule_delay <<= 1;
239
240 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
241 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
242
243 return BP_EAGAIN;
244}
245
246#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
247static void release_memory_resource(struct resource *resource)
248{
249 if (!resource)
250 return;
251
252 /*
253 * No need to reset region to identity mapped since we now
254 * know that no I/O can be in this region
255 */
256 release_resource(resource);
257 kfree(resource);
258}
259
260/*
261 * Host memory not allocated to dom0. We can use this range for hotplug-based
262 * ballooning.
263 *
264 * It's a type-less resource. Setting IORESOURCE_MEM will make resource
265 * management algorithms (arch_remove_reservations()) look into guest e820,
266 * which we don't want.
267 */
268static struct resource hostmem_resource = {
269 .name = "Host RAM",
270};
271
272void __attribute__((weak)) __init arch_xen_balloon_init(struct resource *res)
273{}
274
275static struct resource *additional_memory_resource(phys_addr_t size)
276{
277 struct resource *res, *res_hostmem;
278 int ret = -ENOMEM;
279
280 res = kzalloc(sizeof(*res), GFP_KERNEL);
281 if (!res)
282 return NULL;
283
284 res->name = "System RAM";
285 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
286
287 res_hostmem = kzalloc(sizeof(*res), GFP_KERNEL);
288 if (res_hostmem) {
289 /* Try to grab a range from hostmem */
290 res_hostmem->name = "Host memory";
291 ret = allocate_resource(&hostmem_resource, res_hostmem,
292 size, 0, -1,
293 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
294 }
295
296 if (!ret) {
297 /*
298 * Insert this resource into iomem. Because hostmem_resource
299 * tracks portion of guest e820 marked as UNUSABLE noone else
300 * should try to use it.
301 */
302 res->start = res_hostmem->start;
303 res->end = res_hostmem->end;
304 ret = insert_resource(&iomem_resource, res);
305 if (ret < 0) {
306 pr_err("Can't insert iomem_resource [%llx - %llx]\n",
307 res->start, res->end);
308 release_memory_resource(res_hostmem);
309 res_hostmem = NULL;
310 res->start = res->end = 0;
311 }
312 }
313
314 if (ret) {
315 ret = allocate_resource(&iomem_resource, res,
316 size, 0, -1,
317 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
318 if (ret < 0) {
319 pr_err("Cannot allocate new System RAM resource\n");
320 kfree(res);
321 return NULL;
322 }
323 }
324
325#ifdef CONFIG_SPARSEMEM
326 {
327 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
328 unsigned long pfn = res->start >> PAGE_SHIFT;
329
330 if (pfn > limit) {
331 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
332 pfn, limit);
333 release_memory_resource(res);
334 release_memory_resource(res_hostmem);
335 return NULL;
336 }
337 }
338#endif
339
340 return res;
341}
342
343static enum bp_state reserve_additional_memory(void)
344{
345 long credit;
346 struct resource *resource;
347 int nid, rc;
348 unsigned long balloon_hotplug;
349
350 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
351 - balloon_stats.total_pages;
352
353 /*
354 * Already hotplugged enough pages? Wait for them to be
355 * onlined.
356 */
357 if (credit <= 0)
358 return BP_WAIT;
359
360 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
361
362 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
363 if (!resource)
364 goto err;
365
366 nid = memory_add_physaddr_to_nid(resource->start);
367
368#ifdef CONFIG_XEN_HAVE_PVMMU
369 /*
370 * We don't support PV MMU when Linux and Xen is using
371 * different page granularity.
372 */
373 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
374
375 /*
376 * add_memory() will build page tables for the new memory so
377 * the p2m must contain invalid entries so the correct
378 * non-present PTEs will be written.
379 *
380 * If a failure occurs, the original (identity) p2m entries
381 * are not restored since this region is now known not to
382 * conflict with any devices.
383 */
384 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
385 unsigned long pfn, i;
386
387 pfn = PFN_DOWN(resource->start);
388 for (i = 0; i < balloon_hotplug; i++) {
389 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
390 pr_warn("set_phys_to_machine() failed, no memory added\n");
391 goto err;
392 }
393 }
394 }
395#endif
396
397 /*
398 * add_memory_resource() will call online_pages() which in its turn
399 * will call xen_online_page() callback causing deadlock if we don't
400 * release balloon_mutex here. Unlocking here is safe because the
401 * callers drop the mutex before trying again.
402 */
403 mutex_unlock(&balloon_mutex);
404 rc = add_memory_resource(nid, resource, memhp_auto_online);
405 mutex_lock(&balloon_mutex);
406
407 if (rc) {
408 pr_warn("Cannot add additional memory (%i)\n", rc);
409 goto err;
410 }
411
412 balloon_stats.total_pages += balloon_hotplug;
413
414 return BP_WAIT;
415 err:
416 release_memory_resource(resource);
417 return BP_ECANCELED;
418}
419
420static void xen_online_page(struct page *page)
421{
422 __online_page_set_limits(page);
423
424 mutex_lock(&balloon_mutex);
425
426 __balloon_append(page);
427
428 mutex_unlock(&balloon_mutex);
429}
430
431static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
432{
433 if (val == MEM_ONLINE)
434 schedule_delayed_work(&balloon_worker, 0);
435
436 return NOTIFY_OK;
437}
438
439static struct notifier_block xen_memory_nb = {
440 .notifier_call = xen_memory_notifier,
441 .priority = 0
442};
443#else
444static enum bp_state reserve_additional_memory(void)
445{
446 balloon_stats.target_pages = balloon_stats.current_pages;
447 return BP_ECANCELED;
448}
449#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
450
451static long current_credit(void)
452{
453 return balloon_stats.target_pages - balloon_stats.current_pages;
454}
455
456static bool balloon_is_inflated(void)
457{
458 return balloon_stats.balloon_low || balloon_stats.balloon_high;
459}
460
461static enum bp_state increase_reservation(unsigned long nr_pages)
462{
463 int rc;
464 unsigned long i;
465 struct page *page;
466 struct xen_memory_reservation reservation = {
467 .address_bits = 0,
468 .extent_order = EXTENT_ORDER,
469 .domid = DOMID_SELF
470 };
471
472 if (nr_pages > ARRAY_SIZE(frame_list))
473 nr_pages = ARRAY_SIZE(frame_list);
474
475 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
476 for (i = 0; i < nr_pages; i++) {
477 if (!page) {
478 nr_pages = i;
479 break;
480 }
481
482 /* XENMEM_populate_physmap requires a PFN based on Xen
483 * granularity.
484 */
485 frame_list[i] = page_to_xen_pfn(page);
486 page = balloon_next_page(page);
487 }
488
489 set_xen_guest_handle(reservation.extent_start, frame_list);
490 reservation.nr_extents = nr_pages;
491 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
492 if (rc <= 0)
493 return BP_EAGAIN;
494
495 for (i = 0; i < rc; i++) {
496 page = balloon_retrieve(false);
497 BUG_ON(page == NULL);
498
499#ifdef CONFIG_XEN_HAVE_PVMMU
500 /*
501 * We don't support PV MMU when Linux and Xen is using
502 * different page granularity.
503 */
504 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
505
506 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
507 unsigned long pfn = page_to_pfn(page);
508
509 set_phys_to_machine(pfn, frame_list[i]);
510
511 /* Link back into the page tables if not highmem. */
512 if (!PageHighMem(page)) {
513 int ret;
514 ret = HYPERVISOR_update_va_mapping(
515 (unsigned long)__va(pfn << PAGE_SHIFT),
516 mfn_pte(frame_list[i], PAGE_KERNEL),
517 0);
518 BUG_ON(ret);
519 }
520 }
521#endif
522
523 /* Relinquish the page back to the allocator. */
524 free_reserved_page(page);
525 }
526
527 balloon_stats.current_pages += rc;
528
529 return BP_DONE;
530}
531
532static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
533{
534 enum bp_state state = BP_DONE;
535 unsigned long i;
536 struct page *page, *tmp;
537 int ret;
538 struct xen_memory_reservation reservation = {
539 .address_bits = 0,
540 .extent_order = EXTENT_ORDER,
541 .domid = DOMID_SELF
542 };
543 LIST_HEAD(pages);
544
545 if (nr_pages > ARRAY_SIZE(frame_list))
546 nr_pages = ARRAY_SIZE(frame_list);
547
548 for (i = 0; i < nr_pages; i++) {
549 page = alloc_page(gfp);
550 if (page == NULL) {
551 nr_pages = i;
552 state = BP_EAGAIN;
553 break;
554 }
555 adjust_managed_page_count(page, -1);
556 scrub_page(page);
557 list_add(&page->lru, &pages);
558 }
559
560 /*
561 * Ensure that ballooned highmem pages don't have kmaps.
562 *
563 * Do this before changing the p2m as kmap_flush_unused()
564 * reads PTEs to obtain pages (and hence needs the original
565 * p2m entry).
566 */
567 kmap_flush_unused();
568
569 /*
570 * Setup the frame, update direct mapping, invalidate P2M,
571 * and add to balloon.
572 */
573 i = 0;
574 list_for_each_entry_safe(page, tmp, &pages, lru) {
575 /* XENMEM_decrease_reservation requires a GFN */
576 frame_list[i++] = xen_page_to_gfn(page);
577
578#ifdef CONFIG_XEN_HAVE_PVMMU
579 /*
580 * We don't support PV MMU when Linux and Xen is using
581 * different page granularity.
582 */
583 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
584
585 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
586 unsigned long pfn = page_to_pfn(page);
587
588 if (!PageHighMem(page)) {
589 ret = HYPERVISOR_update_va_mapping(
590 (unsigned long)__va(pfn << PAGE_SHIFT),
591 __pte_ma(0), 0);
592 BUG_ON(ret);
593 }
594 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
595 }
596#endif
597 list_del(&page->lru);
598
599 balloon_append(page);
600 }
601
602 flush_tlb_all();
603
604 set_xen_guest_handle(reservation.extent_start, frame_list);
605 reservation.nr_extents = nr_pages;
606 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
607 BUG_ON(ret != nr_pages);
608
609 balloon_stats.current_pages -= nr_pages;
610
611 return state;
612}
613
614/*
615 * As this is a work item it is guaranteed to run as a single instance only.
616 * We may of course race updates of the target counts (which are protected
617 * by the balloon lock), or with changes to the Xen hard limit, but we will
618 * recover from these in time.
619 */
620static void balloon_process(struct work_struct *work)
621{
622 enum bp_state state = BP_DONE;
623 long credit;
624
625
626 do {
627 mutex_lock(&balloon_mutex);
628
629 credit = current_credit();
630
631 if (credit > 0) {
632 if (balloon_is_inflated())
633 state = increase_reservation(credit);
634 else
635 state = reserve_additional_memory();
636 }
637
638 if (credit < 0)
639 state = decrease_reservation(-credit, GFP_BALLOON);
640
641 state = update_schedule(state);
642
643 mutex_unlock(&balloon_mutex);
644
645 cond_resched();
646
647 } while (credit && state == BP_DONE);
648
649 /* Schedule more work if there is some still to be done. */
650 if (state == BP_EAGAIN)
651 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
652}
653
654/* Resets the Xen limit, sets new target, and kicks off processing. */
655void balloon_set_new_target(unsigned long target)
656{
657 /* No need for lock. Not read-modify-write updates. */
658 balloon_stats.target_pages = target;
659 schedule_delayed_work(&balloon_worker, 0);
660}
661EXPORT_SYMBOL_GPL(balloon_set_new_target);
662
663static int add_ballooned_pages(int nr_pages)
664{
665 enum bp_state st;
666
667 if (xen_hotplug_unpopulated) {
668 st = reserve_additional_memory();
669 if (st != BP_ECANCELED) {
670 mutex_unlock(&balloon_mutex);
671 wait_event(balloon_wq,
672 !list_empty(&ballooned_pages));
673 mutex_lock(&balloon_mutex);
674 return 0;
675 }
676 }
677
678 st = decrease_reservation(nr_pages, GFP_USER);
679 if (st != BP_DONE)
680 return -ENOMEM;
681
682 return 0;
683}
684
685/**
686 * alloc_xenballooned_pages - get pages that have been ballooned out
687 * @nr_pages: Number of pages to get
688 * @pages: pages returned
689 * @return 0 on success, error otherwise
690 */
691int alloc_xenballooned_pages(int nr_pages, struct page **pages)
692{
693 int pgno = 0;
694 struct page *page;
695 int ret;
696
697 mutex_lock(&balloon_mutex);
698
699 balloon_stats.target_unpopulated += nr_pages;
700
701 while (pgno < nr_pages) {
702 page = balloon_retrieve(true);
703 if (page) {
704 pages[pgno++] = page;
705#ifdef CONFIG_XEN_HAVE_PVMMU
706 /*
707 * We don't support PV MMU when Linux and Xen is using
708 * different page granularity.
709 */
710 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
711
712 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
713 ret = xen_alloc_p2m_entry(page_to_pfn(page));
714 if (ret < 0)
715 goto out_undo;
716 }
717#endif
718 } else {
719 ret = add_ballooned_pages(nr_pages - pgno);
720 if (ret < 0)
721 goto out_undo;
722 }
723 }
724 mutex_unlock(&balloon_mutex);
725 return 0;
726 out_undo:
727 mutex_unlock(&balloon_mutex);
728 free_xenballooned_pages(pgno, pages);
729 return ret;
730}
731EXPORT_SYMBOL(alloc_xenballooned_pages);
732
733/**
734 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
735 * @nr_pages: Number of pages
736 * @pages: pages to return
737 */
738void free_xenballooned_pages(int nr_pages, struct page **pages)
739{
740 int i;
741
742 mutex_lock(&balloon_mutex);
743
744 for (i = 0; i < nr_pages; i++) {
745 if (pages[i])
746 balloon_append(pages[i]);
747 }
748
749 balloon_stats.target_unpopulated -= nr_pages;
750
751 /* The balloon may be too large now. Shrink it if needed. */
752 if (current_credit())
753 schedule_delayed_work(&balloon_worker, 0);
754
755 mutex_unlock(&balloon_mutex);
756}
757EXPORT_SYMBOL(free_xenballooned_pages);
758
759#ifdef CONFIG_XEN_PV
760static void __init balloon_add_region(unsigned long start_pfn,
761 unsigned long pages)
762{
763 unsigned long pfn, extra_pfn_end;
764 struct page *page;
765
766 /*
767 * If the amount of usable memory has been limited (e.g., with
768 * the 'mem' command line parameter), don't add pages beyond
769 * this limit.
770 */
771 extra_pfn_end = min(max_pfn, start_pfn + pages);
772
773 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
774 page = pfn_to_page(pfn);
775 /* totalram_pages and totalhigh_pages do not
776 include the boot-time balloon extension, so
777 don't subtract from it. */
778 __balloon_append(page);
779 }
780
781 balloon_stats.total_pages += extra_pfn_end - start_pfn;
782}
783#endif
784
785static int __init balloon_init(void)
786{
787 if (!xen_domain())
788 return -ENODEV;
789
790 pr_info("Initialising balloon driver\n");
791
792#ifdef CONFIG_XEN_PV
793 balloon_stats.current_pages = xen_pv_domain()
794 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
795 : get_num_physpages();
796#else
797 balloon_stats.current_pages = get_num_physpages();
798#endif
799 balloon_stats.target_pages = balloon_stats.current_pages;
800 balloon_stats.balloon_low = 0;
801 balloon_stats.balloon_high = 0;
802 balloon_stats.total_pages = balloon_stats.current_pages;
803
804 balloon_stats.schedule_delay = 1;
805 balloon_stats.max_schedule_delay = 32;
806 balloon_stats.retry_count = 1;
807 balloon_stats.max_retry_count = RETRY_UNLIMITED;
808
809#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
810 set_online_page_callback(&xen_online_page);
811 register_memory_notifier(&xen_memory_nb);
812 register_sysctl_table(xen_root);
813
814 arch_xen_balloon_init(&hostmem_resource);
815#endif
816
817#ifdef CONFIG_XEN_PV
818 {
819 int i;
820
821 /*
822 * Initialize the balloon with pages from the extra memory
823 * regions (see arch/x86/xen/setup.c).
824 */
825 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
826 if (xen_extra_mem[i].n_pfns)
827 balloon_add_region(xen_extra_mem[i].start_pfn,
828 xen_extra_mem[i].n_pfns);
829 }
830#endif
831
832 /* Init the xen-balloon driver. */
833 xen_balloon_init();
834
835 return 0;
836}
837subsys_initcall(balloon_init);