f54290baa3dbd84317350a41aaab720649dc1baf
[linux-2.6.git] / drivers / xen / balloon.c
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  *
8  * This program is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public License version 2
10  * as published by the Free Software Foundation; or, when distributed
11  * separately from the Linux kernel or incorporated into other
12  * software packages, subject to the following license:
13  *
14  * Permission is hereby granted, free of charge, to any person obtaining a copy
15  * of this source file (the "Software"), to deal in the Software without
16  * restriction, including without limitation the rights to use, copy, modify,
17  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
18  * and to permit persons to whom the Software is furnished to do so, subject to
19  * the following conditions:
20  *
21  * The above copyright notice and this permission notice shall be included in
22  * all copies or substantial portions of the Software.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
25  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
26  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
27  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
28  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
29  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
30  * IN THE SOFTWARE.
31  */
32
33 #include <linux/kernel.h>
34 #include <linux/sched.h>
35 #include <linux/errno.h>
36 #include <linux/mm.h>
37 #include <linux/bootmem.h>
38 #include <linux/pagemap.h>
39 #include <linux/highmem.h>
40 #include <linux/mutex.h>
41 #include <linux/list.h>
42 #include <linux/gfp.h>
43
44 #include <asm/page.h>
45 #include <asm/pgalloc.h>
46 #include <asm/pgtable.h>
47 #include <asm/tlb.h>
48 #include <asm/e820.h>
49
50 #include <asm/xen/hypervisor.h>
51 #include <asm/xen/hypercall.h>
52
53 #include <xen/xen.h>
54 #include <xen/interface/xen.h>
55 #include <xen/interface/memory.h>
56 #include <xen/balloon.h>
57 #include <xen/features.h>
58 #include <xen/page.h>
59
60 /*
61  * balloon_process() state:
62  *
63  * BP_DONE: done or nothing to do,
64  * BP_EAGAIN: error, go to sleep,
65  * BP_ECANCELED: error, balloon operation canceled.
66  */
67
68 enum bp_state {
69         BP_DONE,
70         BP_EAGAIN,
71         BP_ECANCELED
72 };
73
74
75 static DEFINE_MUTEX(balloon_mutex);
76
77 struct balloon_stats balloon_stats;
78 EXPORT_SYMBOL_GPL(balloon_stats);
79
80 /* We increase/decrease in batches which fit in a page */
81 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
82
83 #ifdef CONFIG_HIGHMEM
84 #define inc_totalhigh_pages() (totalhigh_pages++)
85 #define dec_totalhigh_pages() (totalhigh_pages--)
86 #else
87 #define inc_totalhigh_pages() do {} while(0)
88 #define dec_totalhigh_pages() do {} while(0)
89 #endif
90
91 /* List of ballooned pages, threaded through the mem_map array. */
92 static LIST_HEAD(ballooned_pages);
93
94 /* Main work function, always executed in process context. */
95 static void balloon_process(struct work_struct *work);
96 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
97
98 /* When ballooning out (allocating memory to return to Xen) we don't really
99    want the kernel to try too hard since that can trigger the oom killer. */
100 #define GFP_BALLOON \
101         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
102
103 static void scrub_page(struct page *page)
104 {
105 #ifdef CONFIG_XEN_SCRUB_PAGES
106         clear_highpage(page);
107 #endif
108 }
109
110 /* balloon_append: add the given page to the balloon. */
111 static void __balloon_append(struct page *page)
112 {
113         /* Lowmem is re-populated first, so highmem pages go at list tail. */
114         if (PageHighMem(page)) {
115                 list_add_tail(&page->lru, &ballooned_pages);
116                 balloon_stats.balloon_high++;
117         } else {
118                 list_add(&page->lru, &ballooned_pages);
119                 balloon_stats.balloon_low++;
120         }
121 }
122
123 static void balloon_append(struct page *page)
124 {
125         __balloon_append(page);
126         if (PageHighMem(page))
127                 dec_totalhigh_pages();
128         totalram_pages--;
129 }
130
131 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
132 static struct page *balloon_retrieve(bool prefer_highmem)
133 {
134         struct page *page;
135
136         if (list_empty(&ballooned_pages))
137                 return NULL;
138
139         if (prefer_highmem)
140                 page = list_entry(ballooned_pages.prev, struct page, lru);
141         else
142                 page = list_entry(ballooned_pages.next, struct page, lru);
143         list_del(&page->lru);
144
145         if (PageHighMem(page)) {
146                 balloon_stats.balloon_high--;
147                 inc_totalhigh_pages();
148         }
149         else
150                 balloon_stats.balloon_low--;
151
152         totalram_pages++;
153
154         return page;
155 }
156
157 static struct page *balloon_first_page(void)
158 {
159         if (list_empty(&ballooned_pages))
160                 return NULL;
161         return list_entry(ballooned_pages.next, struct page, lru);
162 }
163
164 static struct page *balloon_next_page(struct page *page)
165 {
166         struct list_head *next = page->lru.next;
167         if (next == &ballooned_pages)
168                 return NULL;
169         return list_entry(next, struct page, lru);
170 }
171
172 static enum bp_state update_schedule(enum bp_state state)
173 {
174         if (state == BP_DONE) {
175                 balloon_stats.schedule_delay = 1;
176                 balloon_stats.retry_count = 1;
177                 return BP_DONE;
178         }
179
180         ++balloon_stats.retry_count;
181
182         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
183                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
184                 balloon_stats.schedule_delay = 1;
185                 balloon_stats.retry_count = 1;
186                 return BP_ECANCELED;
187         }
188
189         balloon_stats.schedule_delay <<= 1;
190
191         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
192                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
193
194         return BP_EAGAIN;
195 }
196
197 static long current_credit(void)
198 {
199         unsigned long target = balloon_stats.target_pages;
200
201         target = min(target,
202                      balloon_stats.current_pages +
203                      balloon_stats.balloon_low +
204                      balloon_stats.balloon_high);
205
206         return target - balloon_stats.current_pages;
207 }
208
209 static enum bp_state increase_reservation(unsigned long nr_pages)
210 {
211         int rc;
212         unsigned long  pfn, i;
213         struct page   *page;
214         struct xen_memory_reservation reservation = {
215                 .address_bits = 0,
216                 .extent_order = 0,
217                 .domid        = DOMID_SELF
218         };
219
220         if (nr_pages > ARRAY_SIZE(frame_list))
221                 nr_pages = ARRAY_SIZE(frame_list);
222
223         page = balloon_first_page();
224         for (i = 0; i < nr_pages; i++) {
225                 if (!page) {
226                         nr_pages = i;
227                         break;
228                 }
229                 frame_list[i] = page_to_pfn(page);
230                 page = balloon_next_page(page);
231         }
232
233         set_xen_guest_handle(reservation.extent_start, frame_list);
234         reservation.nr_extents = nr_pages;
235         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
236         if (rc <= 0)
237                 return BP_EAGAIN;
238
239         for (i = 0; i < rc; i++) {
240                 page = balloon_retrieve(false);
241                 BUG_ON(page == NULL);
242
243                 pfn = page_to_pfn(page);
244                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
245                        phys_to_machine_mapping_valid(pfn));
246
247                 set_phys_to_machine(pfn, frame_list[i]);
248
249                 /* Link back into the page tables if not highmem. */
250                 if (xen_pv_domain() && !PageHighMem(page)) {
251                         int ret;
252                         ret = HYPERVISOR_update_va_mapping(
253                                 (unsigned long)__va(pfn << PAGE_SHIFT),
254                                 mfn_pte(frame_list[i], PAGE_KERNEL),
255                                 0);
256                         BUG_ON(ret);
257                 }
258
259                 /* Relinquish the page back to the allocator. */
260                 ClearPageReserved(page);
261                 init_page_count(page);
262                 __free_page(page);
263         }
264
265         balloon_stats.current_pages += rc;
266
267         return BP_DONE;
268 }
269
270 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
271 {
272         enum bp_state state = BP_DONE;
273         unsigned long  pfn, i;
274         struct page   *page;
275         int ret;
276         struct xen_memory_reservation reservation = {
277                 .address_bits = 0,
278                 .extent_order = 0,
279                 .domid        = DOMID_SELF
280         };
281
282         if (nr_pages > ARRAY_SIZE(frame_list))
283                 nr_pages = ARRAY_SIZE(frame_list);
284
285         for (i = 0; i < nr_pages; i++) {
286                 if ((page = alloc_page(gfp)) == NULL) {
287                         nr_pages = i;
288                         state = BP_EAGAIN;
289                         break;
290                 }
291
292                 pfn = page_to_pfn(page);
293                 frame_list[i] = pfn_to_mfn(pfn);
294
295                 scrub_page(page);
296
297                 if (xen_pv_domain() && !PageHighMem(page)) {
298                         ret = HYPERVISOR_update_va_mapping(
299                                 (unsigned long)__va(pfn << PAGE_SHIFT),
300                                 __pte_ma(0), 0);
301                         BUG_ON(ret);
302                 }
303
304         }
305
306         /* Ensure that ballooned highmem pages don't have kmaps. */
307         kmap_flush_unused();
308         flush_tlb_all();
309
310         /* No more mappings: invalidate P2M and add to balloon. */
311         for (i = 0; i < nr_pages; i++) {
312                 pfn = mfn_to_pfn(frame_list[i]);
313                 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
314                 balloon_append(pfn_to_page(pfn));
315         }
316
317         set_xen_guest_handle(reservation.extent_start, frame_list);
318         reservation.nr_extents   = nr_pages;
319         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
320         BUG_ON(ret != nr_pages);
321
322         balloon_stats.current_pages -= nr_pages;
323
324         return state;
325 }
326
327 /*
328  * We avoid multiple worker processes conflicting via the balloon mutex.
329  * We may of course race updates of the target counts (which are protected
330  * by the balloon lock), or with changes to the Xen hard limit, but we will
331  * recover from these in time.
332  */
333 static void balloon_process(struct work_struct *work)
334 {
335         enum bp_state state = BP_DONE;
336         long credit;
337
338         mutex_lock(&balloon_mutex);
339
340         do {
341                 credit = current_credit();
342
343                 if (credit > 0)
344                         state = increase_reservation(credit);
345
346                 if (credit < 0)
347                         state = decrease_reservation(-credit, GFP_BALLOON);
348
349                 state = update_schedule(state);
350
351 #ifndef CONFIG_PREEMPT
352                 if (need_resched())
353                         schedule();
354 #endif
355         } while (credit && state == BP_DONE);
356
357         /* Schedule more work if there is some still to be done. */
358         if (state == BP_EAGAIN)
359                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
360
361         mutex_unlock(&balloon_mutex);
362 }
363
364 /* Resets the Xen limit, sets new target, and kicks off processing. */
365 void balloon_set_new_target(unsigned long target)
366 {
367         /* No need for lock. Not read-modify-write updates. */
368         balloon_stats.target_pages = target;
369         schedule_delayed_work(&balloon_worker, 0);
370 }
371 EXPORT_SYMBOL_GPL(balloon_set_new_target);
372
373 /**
374  * alloc_xenballooned_pages - get pages that have been ballooned out
375  * @nr_pages: Number of pages to get
376  * @pages: pages returned
377  * @return 0 on success, error otherwise
378  */
379 int alloc_xenballooned_pages(int nr_pages, struct page** pages)
380 {
381         int pgno = 0;
382         struct page* page;
383         mutex_lock(&balloon_mutex);
384         while (pgno < nr_pages) {
385                 page = balloon_retrieve(true);
386                 if (page) {
387                         pages[pgno++] = page;
388                 } else {
389                         enum bp_state st;
390                         st = decrease_reservation(nr_pages - pgno, GFP_HIGHUSER);
391                         if (st != BP_DONE)
392                                 goto out_undo;
393                 }
394         }
395         mutex_unlock(&balloon_mutex);
396         return 0;
397  out_undo:
398         while (pgno)
399                 balloon_append(pages[--pgno]);
400         /* Free the memory back to the kernel soon */
401         schedule_delayed_work(&balloon_worker, 0);
402         mutex_unlock(&balloon_mutex);
403         return -ENOMEM;
404 }
405 EXPORT_SYMBOL(alloc_xenballooned_pages);
406
407 /**
408  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
409  * @nr_pages: Number of pages
410  * @pages: pages to return
411  */
412 void free_xenballooned_pages(int nr_pages, struct page** pages)
413 {
414         int i;
415
416         mutex_lock(&balloon_mutex);
417
418         for (i = 0; i < nr_pages; i++) {
419                 if (pages[i])
420                         balloon_append(pages[i]);
421         }
422
423         /* The balloon may be too large now. Shrink it if needed. */
424         if (current_credit())
425                 schedule_delayed_work(&balloon_worker, 0);
426
427         mutex_unlock(&balloon_mutex);
428 }
429 EXPORT_SYMBOL(free_xenballooned_pages);
430
431 static int __init balloon_init(void)
432 {
433         unsigned long pfn, extra_pfn_end;
434         struct page *page;
435
436         if (!xen_domain())
437                 return -ENODEV;
438
439         pr_info("xen/balloon: Initialising balloon driver.\n");
440
441         balloon_stats.current_pages = xen_pv_domain() ? min(xen_start_info->nr_pages, max_pfn) : max_pfn;
442         balloon_stats.target_pages  = balloon_stats.current_pages;
443         balloon_stats.balloon_low   = 0;
444         balloon_stats.balloon_high  = 0;
445
446         balloon_stats.schedule_delay = 1;
447         balloon_stats.max_schedule_delay = 32;
448         balloon_stats.retry_count = 1;
449         balloon_stats.max_retry_count = RETRY_UNLIMITED;
450
451         /*
452          * Initialise the balloon with excess memory space.  We need
453          * to make sure we don't add memory which doesn't exist or
454          * logically exist.  The E820 map can be trimmed to be smaller
455          * than the amount of physical memory due to the mem= command
456          * line parameter.  And if this is a 32-bit non-HIGHMEM kernel
457          * on a system with memory which requires highmem to access,
458          * don't try to use it.
459          */
460         extra_pfn_end = min(min(max_pfn, e820_end_of_ram_pfn()),
461                             (unsigned long)PFN_DOWN(xen_extra_mem_start + xen_extra_mem_size));
462         for (pfn = PFN_UP(xen_extra_mem_start);
463              pfn < extra_pfn_end;
464              pfn++) {
465                 page = pfn_to_page(pfn);
466                 /* totalram_pages and totalhigh_pages do not include the boot-time
467                    balloon extension, so don't subtract from it. */
468                 __balloon_append(page);
469         }
470
471         return 0;
472 }
473
474 subsys_initcall(balloon_init);
475
476 MODULE_LICENSE("GPL");