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