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