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/errno.h>
45#include <linux/module.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
58#include <asm/page.h>
59#include <asm/pgalloc.h>
60#include <asm/pgtable.h>
61#include <asm/tlb.h>
62
63#include <asm/xen/hypervisor.h>
64#include <asm/xen/hypercall.h>
65
66#include <xen/xen.h>
67#include <xen/interface/xen.h>
68#include <xen/interface/memory.h>
69#include <xen/balloon.h>
70#include <xen/features.h>
71#include <xen/page.h>
72
73/*
74 * balloon_process() state:
75 *
76 * BP_DONE: done or nothing to do,
77 * BP_EAGAIN: error, go to sleep,
78 * BP_ECANCELED: error, balloon operation canceled.
79 */
80
81enum bp_state {
82 BP_DONE,
83 BP_EAGAIN,
84 BP_ECANCELED
85};
86
87
88static DEFINE_MUTEX(balloon_mutex);
89
90struct balloon_stats balloon_stats;
91EXPORT_SYMBOL_GPL(balloon_stats);
92
93/* We increase/decrease in batches which fit in a page */
94static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
95static DEFINE_PER_CPU(struct page *, balloon_scratch_page);
96
97
98/* List of ballooned pages, threaded through the mem_map array. */
99static LIST_HEAD(ballooned_pages);
100
101/* Main work function, always executed in process context. */
102static void balloon_process(struct work_struct *work);
103static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
104
105/* When ballooning out (allocating memory to return to Xen) we don't really
106 want the kernel to try too hard since that can trigger the oom killer. */
107#define GFP_BALLOON \
108 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
109
110static void scrub_page(struct page *page)
111{
112#ifdef CONFIG_XEN_SCRUB_PAGES
113 clear_highpage(page);
114#endif
115}
116
117/* balloon_append: add the given page to the balloon. */
118static void __balloon_append(struct page *page)
119{
120 /* Lowmem is re-populated first, so highmem pages go at list tail. */
121 if (PageHighMem(page)) {
122 list_add_tail(&page->lru, &ballooned_pages);
123 balloon_stats.balloon_high++;
124 } else {
125 list_add(&page->lru, &ballooned_pages);
126 balloon_stats.balloon_low++;
127 }
128}
129
130static void balloon_append(struct page *page)
131{
132 __balloon_append(page);
133 adjust_managed_page_count(page, -1);
134}
135
136/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
137static struct page *balloon_retrieve(bool prefer_highmem)
138{
139 struct page *page;
140
141 if (list_empty(&ballooned_pages))
142 return NULL;
143
144 if (prefer_highmem)
145 page = list_entry(ballooned_pages.prev, struct page, lru);
146 else
147 page = list_entry(ballooned_pages.next, struct page, lru);
148 list_del(&page->lru);
149
150 if (PageHighMem(page))
151 balloon_stats.balloon_high--;
152 else
153 balloon_stats.balloon_low--;
154
155 adjust_managed_page_count(page, 1);
156
157 return page;
158}
159
160static struct page *balloon_next_page(struct page *page)
161{
162 struct list_head *next = page->lru.next;
163 if (next == &ballooned_pages)
164 return NULL;
165 return list_entry(next, struct page, lru);
166}
167
168static enum bp_state update_schedule(enum bp_state state)
169{
170 if (state == BP_DONE) {
171 balloon_stats.schedule_delay = 1;
172 balloon_stats.retry_count = 1;
173 return BP_DONE;
174 }
175
176 ++balloon_stats.retry_count;
177
178 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
179 balloon_stats.retry_count > balloon_stats.max_retry_count) {
180 balloon_stats.schedule_delay = 1;
181 balloon_stats.retry_count = 1;
182 return BP_ECANCELED;
183 }
184
185 balloon_stats.schedule_delay <<= 1;
186
187 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
188 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
189
190 return BP_EAGAIN;
191}
192
193#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
194static long current_credit(void)
195{
196 return balloon_stats.target_pages - balloon_stats.current_pages -
197 balloon_stats.hotplug_pages;
198}
199
200static bool balloon_is_inflated(void)
201{
202 if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
203 balloon_stats.balloon_hotplug)
204 return true;
205 else
206 return false;
207}
208
209/*
210 * reserve_additional_memory() adds memory region of size >= credit above
211 * max_pfn. New region is section aligned and size is modified to be multiple
212 * of section size. Those features allow optimal use of address space and
213 * establish proper alignment when this function is called first time after
214 * boot (last section not fully populated at boot time contains unused memory
215 * pages with PG_reserved bit not set; online_pages_range() does not allow page
216 * onlining in whole range if first onlined page does not have PG_reserved
217 * bit set). Real size of added memory is established at page onlining stage.
218 */
219
220static enum bp_state reserve_additional_memory(long credit)
221{
222 int nid, rc;
223 u64 hotplug_start_paddr;
224 unsigned long balloon_hotplug = credit;
225
226 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
227 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
228 nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
229
230 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
231
232 if (rc) {
233 pr_info("%s: add_memory() failed: %i\n", __func__, rc);
234 return BP_EAGAIN;
235 }
236
237 balloon_hotplug -= credit;
238
239 balloon_stats.hotplug_pages += credit;
240 balloon_stats.balloon_hotplug = balloon_hotplug;
241
242 return BP_DONE;
243}
244
245static void xen_online_page(struct page *page)
246{
247 __online_page_set_limits(page);
248
249 mutex_lock(&balloon_mutex);
250
251 __balloon_append(page);
252
253 if (balloon_stats.hotplug_pages)
254 --balloon_stats.hotplug_pages;
255 else
256 --balloon_stats.balloon_hotplug;
257
258 mutex_unlock(&balloon_mutex);
259}
260
261static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
262{
263 if (val == MEM_ONLINE)
264 schedule_delayed_work(&balloon_worker, 0);
265
266 return NOTIFY_OK;
267}
268
269static struct notifier_block xen_memory_nb = {
270 .notifier_call = xen_memory_notifier,
271 .priority = 0
272};
273#else
274static long current_credit(void)
275{
276 unsigned long target = balloon_stats.target_pages;
277
278 target = min(target,
279 balloon_stats.current_pages +
280 balloon_stats.balloon_low +
281 balloon_stats.balloon_high);
282
283 return target - balloon_stats.current_pages;
284}
285
286static bool balloon_is_inflated(void)
287{
288 if (balloon_stats.balloon_low || balloon_stats.balloon_high)
289 return true;
290 else
291 return false;
292}
293
294static enum bp_state reserve_additional_memory(long credit)
295{
296 balloon_stats.target_pages = balloon_stats.current_pages;
297 return BP_DONE;
298}
299#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
300
301static enum bp_state increase_reservation(unsigned long nr_pages)
302{
303 int rc;
304 unsigned long pfn, i;
305 struct page *page;
306 struct xen_memory_reservation reservation = {
307 .address_bits = 0,
308 .extent_order = 0,
309 .domid = DOMID_SELF
310 };
311
312#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
313 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
314 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
315 balloon_stats.hotplug_pages += nr_pages;
316 balloon_stats.balloon_hotplug -= nr_pages;
317 return BP_DONE;
318 }
319#endif
320
321 if (nr_pages > ARRAY_SIZE(frame_list))
322 nr_pages = ARRAY_SIZE(frame_list);
323
324 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
325 for (i = 0; i < nr_pages; i++) {
326 if (!page) {
327 nr_pages = i;
328 break;
329 }
330 frame_list[i] = page_to_pfn(page);
331 page = balloon_next_page(page);
332 }
333
334 set_xen_guest_handle(reservation.extent_start, frame_list);
335 reservation.nr_extents = nr_pages;
336 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
337 if (rc <= 0)
338 return BP_EAGAIN;
339
340 for (i = 0; i < rc; i++) {
341 page = balloon_retrieve(false);
342 BUG_ON(page == NULL);
343
344 pfn = page_to_pfn(page);
345
346#ifdef CONFIG_XEN_HAVE_PVMMU
347 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
348 set_phys_to_machine(pfn, frame_list[i]);
349
350 /* Link back into the page tables if not highmem. */
351 if (!PageHighMem(page)) {
352 int ret;
353 ret = HYPERVISOR_update_va_mapping(
354 (unsigned long)__va(pfn << PAGE_SHIFT),
355 mfn_pte(frame_list[i], PAGE_KERNEL),
356 0);
357 BUG_ON(ret);
358 }
359 }
360#endif
361
362 /* Relinquish the page back to the allocator. */
363 __free_reserved_page(page);
364 }
365
366 balloon_stats.current_pages += rc;
367
368 return BP_DONE;
369}
370
371static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
372{
373 enum bp_state state = BP_DONE;
374 unsigned long pfn, i;
375 struct page *page;
376 int ret;
377 struct xen_memory_reservation reservation = {
378 .address_bits = 0,
379 .extent_order = 0,
380 .domid = DOMID_SELF
381 };
382
383#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
384 if (balloon_stats.hotplug_pages) {
385 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
386 balloon_stats.hotplug_pages -= nr_pages;
387 balloon_stats.balloon_hotplug += nr_pages;
388 return BP_DONE;
389 }
390#endif
391
392 if (nr_pages > ARRAY_SIZE(frame_list))
393 nr_pages = ARRAY_SIZE(frame_list);
394
395 for (i = 0; i < nr_pages; i++) {
396 page = alloc_page(gfp);
397 if (page == NULL) {
398 nr_pages = i;
399 state = BP_EAGAIN;
400 break;
401 }
402 scrub_page(page);
403
404 frame_list[i] = page_to_pfn(page);
405 }
406
407 /*
408 * Ensure that ballooned highmem pages don't have kmaps.
409 *
410 * Do this before changing the p2m as kmap_flush_unused()
411 * reads PTEs to obtain pages (and hence needs the original
412 * p2m entry).
413 */
414 kmap_flush_unused();
415
416 /* Update direct mapping, invalidate P2M, and add to balloon. */
417 for (i = 0; i < nr_pages; i++) {
418 pfn = frame_list[i];
419 frame_list[i] = pfn_to_mfn(pfn);
420 page = pfn_to_page(pfn);
421
422#ifdef CONFIG_XEN_HAVE_PVMMU
423 /*
424 * Ballooned out frames are effectively replaced with
425 * a scratch frame. Ensure direct mappings and the
426 * p2m are consistent.
427 */
428 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
429 unsigned long p;
430 struct page *scratch_page = get_balloon_scratch_page();
431
432 if (!PageHighMem(page)) {
433 ret = HYPERVISOR_update_va_mapping(
434 (unsigned long)__va(pfn << PAGE_SHIFT),
435 pfn_pte(page_to_pfn(scratch_page),
436 PAGE_KERNEL_RO), 0);
437 BUG_ON(ret);
438 }
439 p = page_to_pfn(scratch_page);
440 __set_phys_to_machine(pfn, pfn_to_mfn(p));
441
442 put_balloon_scratch_page();
443 }
444#endif
445
446 balloon_append(page);
447 }
448
449 flush_tlb_all();
450
451 set_xen_guest_handle(reservation.extent_start, frame_list);
452 reservation.nr_extents = nr_pages;
453 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
454 BUG_ON(ret != nr_pages);
455
456 balloon_stats.current_pages -= nr_pages;
457
458 return state;
459}
460
461/*
462 * We avoid multiple worker processes conflicting via the balloon mutex.
463 * We may of course race updates of the target counts (which are protected
464 * by the balloon lock), or with changes to the Xen hard limit, but we will
465 * recover from these in time.
466 */
467static void balloon_process(struct work_struct *work)
468{
469 enum bp_state state = BP_DONE;
470 long credit;
471
472 mutex_lock(&balloon_mutex);
473
474 do {
475 credit = current_credit();
476
477 if (credit > 0) {
478 if (balloon_is_inflated())
479 state = increase_reservation(credit);
480 else
481 state = reserve_additional_memory(credit);
482 }
483
484 if (credit < 0)
485 state = decrease_reservation(-credit, GFP_BALLOON);
486
487 state = update_schedule(state);
488
489#ifndef CONFIG_PREEMPT
490 if (need_resched())
491 schedule();
492#endif
493 } while (credit && state == BP_DONE);
494
495 /* Schedule more work if there is some still to be done. */
496 if (state == BP_EAGAIN)
497 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
498
499 mutex_unlock(&balloon_mutex);
500}
501
502struct page *get_balloon_scratch_page(void)
503{
504 struct page *ret = get_cpu_var(balloon_scratch_page);
505 BUG_ON(ret == NULL);
506 return ret;
507}
508
509void put_balloon_scratch_page(void)
510{
511 put_cpu_var(balloon_scratch_page);
512}
513
514/* Resets the Xen limit, sets new target, and kicks off processing. */
515void balloon_set_new_target(unsigned long target)
516{
517 /* No need for lock. Not read-modify-write updates. */
518 balloon_stats.target_pages = target;
519 schedule_delayed_work(&balloon_worker, 0);
520}
521EXPORT_SYMBOL_GPL(balloon_set_new_target);
522
523/**
524 * alloc_xenballooned_pages - get pages that have been ballooned out
525 * @nr_pages: Number of pages to get
526 * @pages: pages returned
527 * @highmem: allow highmem pages
528 * @return 0 on success, error otherwise
529 */
530int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
531{
532 int pgno = 0;
533 struct page *page;
534 mutex_lock(&balloon_mutex);
535 while (pgno < nr_pages) {
536 page = balloon_retrieve(highmem);
537 if (page && (highmem || !PageHighMem(page))) {
538 pages[pgno++] = page;
539 } else {
540 enum bp_state st;
541 if (page)
542 balloon_append(page);
543 st = decrease_reservation(nr_pages - pgno,
544 highmem ? GFP_HIGHUSER : GFP_USER);
545 if (st != BP_DONE)
546 goto out_undo;
547 }
548 }
549 mutex_unlock(&balloon_mutex);
550 return 0;
551 out_undo:
552 while (pgno)
553 balloon_append(pages[--pgno]);
554 /* Free the memory back to the kernel soon */
555 schedule_delayed_work(&balloon_worker, 0);
556 mutex_unlock(&balloon_mutex);
557 return -ENOMEM;
558}
559EXPORT_SYMBOL(alloc_xenballooned_pages);
560
561/**
562 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
563 * @nr_pages: Number of pages
564 * @pages: pages to return
565 */
566void free_xenballooned_pages(int nr_pages, struct page **pages)
567{
568 int i;
569
570 mutex_lock(&balloon_mutex);
571
572 for (i = 0; i < nr_pages; i++) {
573 if (pages[i])
574 balloon_append(pages[i]);
575 }
576
577 /* The balloon may be too large now. Shrink it if needed. */
578 if (current_credit())
579 schedule_delayed_work(&balloon_worker, 0);
580
581 mutex_unlock(&balloon_mutex);
582}
583EXPORT_SYMBOL(free_xenballooned_pages);
584
585static void __init balloon_add_region(unsigned long start_pfn,
586 unsigned long pages)
587{
588 unsigned long pfn, extra_pfn_end;
589 struct page *page;
590
591 /*
592 * If the amount of usable memory has been limited (e.g., with
593 * the 'mem' command line parameter), don't add pages beyond
594 * this limit.
595 */
596 extra_pfn_end = min(max_pfn, start_pfn + pages);
597
598 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
599 page = pfn_to_page(pfn);
600 /* totalram_pages and totalhigh_pages do not
601 include the boot-time balloon extension, so
602 don't subtract from it. */
603 __balloon_append(page);
604 }
605}
606
607static int alloc_balloon_scratch_page(int cpu)
608{
609 if (per_cpu(balloon_scratch_page, cpu) != NULL)
610 return 0;
611
612 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL);
613 if (per_cpu(balloon_scratch_page, cpu) == NULL) {
614 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu);
615 return -ENOMEM;
616 }
617
618 return 0;
619}
620
621
622static int balloon_cpu_notify(struct notifier_block *self,
623 unsigned long action, void *hcpu)
624{
625 int cpu = (long)hcpu;
626 switch (action) {
627 case CPU_UP_PREPARE:
628 if (alloc_balloon_scratch_page(cpu))
629 return NOTIFY_BAD;
630 break;
631 default:
632 break;
633 }
634 return NOTIFY_OK;
635}
636
637static struct notifier_block balloon_cpu_notifier = {
638 .notifier_call = balloon_cpu_notify,
639};
640
641static int __init balloon_init(void)
642{
643 int i, cpu;
644
645 if (!xen_domain())
646 return -ENODEV;
647
648 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
649 register_cpu_notifier(&balloon_cpu_notifier);
650
651 get_online_cpus();
652 for_each_online_cpu(cpu) {
653 if (alloc_balloon_scratch_page(cpu)) {
654 put_online_cpus();
655 unregister_cpu_notifier(&balloon_cpu_notifier);
656 return -ENOMEM;
657 }
658 }
659 put_online_cpus();
660 }
661
662 pr_info("Initialising balloon driver\n");
663
664 balloon_stats.current_pages = xen_pv_domain()
665 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
666 : get_num_physpages();
667 balloon_stats.target_pages = balloon_stats.current_pages;
668 balloon_stats.balloon_low = 0;
669 balloon_stats.balloon_high = 0;
670
671 balloon_stats.schedule_delay = 1;
672 balloon_stats.max_schedule_delay = 32;
673 balloon_stats.retry_count = 1;
674 balloon_stats.max_retry_count = RETRY_UNLIMITED;
675
676#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
677 balloon_stats.hotplug_pages = 0;
678 balloon_stats.balloon_hotplug = 0;
679
680 set_online_page_callback(&xen_online_page);
681 register_memory_notifier(&xen_memory_nb);
682#endif
683
684 /*
685 * Initialize the balloon with pages from the extra memory
686 * regions (see arch/x86/xen/setup.c).
687 */
688 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
689 if (xen_extra_mem[i].size)
690 balloon_add_region(PFN_UP(xen_extra_mem[i].start),
691 PFN_DOWN(xen_extra_mem[i].size));
692
693 return 0;
694}
695
696subsys_initcall(balloon_init);
697
698static int __init balloon_clear(void)
699{
700 int cpu;
701
702 for_each_possible_cpu(cpu)
703 per_cpu(balloon_scratch_page, cpu) = NULL;
704
705 return 0;
706}
707early_initcall(balloon_clear);
708
709MODULE_LICENSE("GPL");