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