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