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 struct resource *additional_memory_resource(phys_addr_t size)
 251 {
 252         struct resource *res;
 253         int ret;
 254
 255         res = kzalloc(sizeof(*res), GFP_KERNEL);
 256         if (!res)
 257                 return NULL;
 258
 259         res->name = "System RAM";
 260         res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 261
 262         ret = allocate_resource(&iomem_resource, res,
 263                                 size, 0, -1,
 264                                 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
 265         if (ret < 0) {
 266                 pr_err("Cannot allocate new System RAM resource\n");
 267                 kfree(res);
 268                 return NULL;
 269         }
 270
 271         return res;
 272 }
 273
 274 static void release_memory_resource(struct resource *resource)
 275 {
 276         if (!resource)
 277                 return;
 278
 279         /*
 280          * No need to reset region to identity mapped since we now
 281          * know that no I/O can be in this region
 282          */
 283         release_resource(resource);
 284         kfree(resource);
 285 }
 286
 287 static 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         rc = add_memory_resource(nid, resource, memhp_auto_online);
 349         mutex_lock(&balloon_mutex);
 350
 351         if (rc) {
 352                 pr_warn("Cannot add additional memory (%i)\n", rc);
 353                 goto err;
 354         }
 355
 356         balloon_stats.total_pages += balloon_hotplug;
 357
 358         return BP_WAIT;
 359   err:
 360         release_memory_resource(resource);
 361         return BP_ECANCELED;
 362 }
 363
 364 static void xen_online_page(struct page *page)
 365 {
 366         __online_page_set_limits(page);
 367
 368         mutex_lock(&balloon_mutex);
 369
 370         __balloon_append(page);
 371
 372         mutex_unlock(&balloon_mutex);
 373 }
 374
 375 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 376 {
 377         if (val == MEM_ONLINE)
 378                 schedule_delayed_work(&balloon_worker, 0);
 379
 380         return NOTIFY_OK;
 381 }
 382
 383 static struct notifier_block xen_memory_nb = {
 384         .notifier_call = xen_memory_notifier,
 385         .priority = 0
 386 };
 387 #else
 388 static enum bp_state reserve_additional_memory(void)
 389 {
 390         balloon_stats.target_pages = balloon_stats.current_pages;
 391         return BP_ECANCELED;
 392 }
 393 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 394
 395 static long current_credit(void)
 396 {
 397         return balloon_stats.target_pages - balloon_stats.current_pages;
 398 }
 399
 400 static bool balloon_is_inflated(void)
 401 {
 402         return balloon_stats.balloon_low || balloon_stats.balloon_high;
 403 }
 404
 405 static enum bp_state increase_reservation(unsigned long nr_pages)
 406 {
 407         int rc;
 408         unsigned long i;
 409         struct page   *page;
 410         struct xen_memory_reservation reservation = {
 411                 .address_bits = 0,
 412                 .extent_order = EXTENT_ORDER,
 413                 .domid        = DOMID_SELF
 414         };
 415
 416         if (nr_pages > ARRAY_SIZE(frame_list))
 417                 nr_pages = ARRAY_SIZE(frame_list);
 418
 419         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 420         for (i = 0; i < nr_pages; i++) {
 421                 if (!page) {
 422                         nr_pages = i;
 423                         break;
 424                 }
 425
 426                 /* XENMEM_populate_physmap requires a PFN based on Xen
 427                  * granularity.
 428                  */
 429                 frame_list[i] = page_to_xen_pfn(page);
 430                 page = balloon_next_page(page);
 431         }
 432
 433         set_xen_guest_handle(reservation.extent_start, frame_list);
 434         reservation.nr_extents = nr_pages;
 435         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 436         if (rc <= 0)
 437                 return BP_EAGAIN;
 438
 439         for (i = 0; i < rc; i++) {
 440                 page = balloon_retrieve(false);
 441                 BUG_ON(page == NULL);
 442
 443 #ifdef CONFIG_XEN_HAVE_PVMMU
 444                 /*
 445                  * We don't support PV MMU when Linux and Xen is using
 446                  * different page granularity.
 447                  */
 448                 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 449
 450                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 451                         unsigned long pfn = page_to_pfn(page);
 452
 453                         set_phys_to_machine(pfn, frame_list[i]);
 454
 455                         /* Link back into the page tables if not highmem. */
 456                         if (!PageHighMem(page)) {
 457                                 int ret;
 458                                 ret = HYPERVISOR_update_va_mapping(
 459                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 460                                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 461                                                 0);
 462                                 BUG_ON(ret);
 463                         }
 464                 }
 465 #endif
 466
 467                 /* Relinquish the page back to the allocator. */
 468                 __free_reserved_page(page);
 469         }
 470
 471         balloon_stats.current_pages += rc;
 472
 473         return BP_DONE;
 474 }
 475
 476 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 477 {
 478         enum bp_state state = BP_DONE;
 479         unsigned long i;
 480         struct page *page, *tmp;
 481         int ret;
 482         struct xen_memory_reservation reservation = {
 483                 .address_bits = 0,
 484                 .extent_order = EXTENT_ORDER,
 485                 .domid        = DOMID_SELF
 486         };
 487         LIST_HEAD(pages);
 488
 489         if (nr_pages > ARRAY_SIZE(frame_list))
 490                 nr_pages = ARRAY_SIZE(frame_list);
 491
 492         for (i = 0; i < nr_pages; i++) {
 493                 page = alloc_page(gfp);
 494                 if (page == NULL) {
 495                         nr_pages = i;
 496                         state = BP_EAGAIN;
 497                         break;
 498                 }
 499                 scrub_page(page);
 500                 list_add(&page->lru, &pages);
 501         }
 502
 503         /*
 504          * Ensure that ballooned highmem pages don't have kmaps.
 505          *
 506          * Do this before changing the p2m as kmap_flush_unused()
 507          * reads PTEs to obtain pages (and hence needs the original
 508          * p2m entry).
 509          */
 510         kmap_flush_unused();
 511
 512         /*
 513          * Setup the frame, update direct mapping, invalidate P2M,
 514          * and add to balloon.
 515          */
 516         i = 0;
 517         list_for_each_entry_safe(page, tmp, &pages, lru) {
 518                 /* XENMEM_decrease_reservation requires a GFN */
 519                 frame_list[i++] = xen_page_to_gfn(page);
 520
 521 #ifdef CONFIG_XEN_HAVE_PVMMU
 522                 /*
 523                  * We don't support PV MMU when Linux and Xen is using
 524                  * different page granularity.
 525                  */
 526                 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 527
 528                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 529                         unsigned long pfn = page_to_pfn(page);
 530
 531                         if (!PageHighMem(page)) {
 532                                 ret = HYPERVISOR_update_va_mapping(
 533                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 534                                                 __pte_ma(0), 0);
 535                                 BUG_ON(ret);
 536                         }
 537                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 538                 }
 539 #endif
 540                 list_del(&page->lru);
 541
 542                 balloon_append(page);
 543         }
 544
 545         flush_tlb_all();
 546
 547         set_xen_guest_handle(reservation.extent_start, frame_list);
 548         reservation.nr_extents   = nr_pages;
 549         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 550         BUG_ON(ret != nr_pages);
 551
 552         balloon_stats.current_pages -= nr_pages;
 553
 554         return state;
 555 }
 556
 557 /*
 558  * As this is a work item it is guaranteed to run as a single instance only.
 559  * We may of course race updates of the target counts (which are protected
 560  * by the balloon lock), or with changes to the Xen hard limit, but we will
 561  * recover from these in time.
 562  */
 563 static void balloon_process(struct work_struct *work)
 564 {
 565         enum bp_state state = BP_DONE;
 566         long credit;
 567
 568
 569         do {
 570                 mutex_lock(&balloon_mutex);
 571
 572                 credit = current_credit();
 573
 574                 if (credit > 0) {
 575                         if (balloon_is_inflated())
 576                                 state = increase_reservation(credit);
 577                         else
 578                                 state = reserve_additional_memory();
 579                 }
 580
 581                 if (credit < 0)
 582                         state = decrease_reservation(-credit, GFP_BALLOON);
 583
 584                 state = update_schedule(state);
 585
 586                 mutex_unlock(&balloon_mutex);
 587
 588                 cond_resched();
 589
 590         } while (credit && state == BP_DONE);
 591
 592         /* Schedule more work if there is some still to be done. */
 593         if (state == BP_EAGAIN)
 594                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 595 }
 596
 597 /* Resets the Xen limit, sets new target, and kicks off processing. */
 598 void balloon_set_new_target(unsigned long target)
 599 {
 600         /* No need for lock. Not read-modify-write updates. */
 601         balloon_stats.target_pages = target;
 602         schedule_delayed_work(&balloon_worker, 0);
 603 }
 604 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 605
 606 static int add_ballooned_pages(int nr_pages)
 607 {
 608         enum bp_state st;
 609
 610         if (xen_hotplug_unpopulated) {
 611                 st = reserve_additional_memory();
 612                 if (st != BP_ECANCELED) {
 613                         mutex_unlock(&balloon_mutex);
 614                         wait_event(balloon_wq,
 615                                    !list_empty(&ballooned_pages));
 616                         mutex_lock(&balloon_mutex);
 617                         return 0;
 618                 }
 619         }
 620
 621         st = decrease_reservation(nr_pages, GFP_USER);
 622         if (st != BP_DONE)
 623                 return -ENOMEM;
 624
 625         return 0;
 626 }
 627
 628 /**
 629  * alloc_xenballooned_pages - get pages that have been ballooned out
 630  * @nr_pages: Number of pages to get
 631  * @pages: pages returned
 632  * @return 0 on success, error otherwise
 633  */
 634 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
 635 {
 636         int pgno = 0;
 637         struct page *page;
 638         int ret;
 639
 640         mutex_lock(&balloon_mutex);
 641
 642         balloon_stats.target_unpopulated += nr_pages;
 643
 644         while (pgno < nr_pages) {
 645                 page = balloon_retrieve(true);
 646                 if (page) {
 647                         pages[pgno++] = page;
 648 #ifdef CONFIG_XEN_HAVE_PVMMU
 649                         /*
 650                          * We don't support PV MMU when Linux and Xen is using
 651                          * different page granularity.
 652                          */
 653                         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 654
 655                         ret = xen_alloc_p2m_entry(page_to_pfn(page));
 656                         if (ret < 0)
 657                                 goto out_undo;
 658 #endif
 659                 } else {
 660                         ret = add_ballooned_pages(nr_pages - pgno);
 661                         if (ret < 0)
 662                                 goto out_undo;
 663                 }
 664         }
 665         mutex_unlock(&balloon_mutex);
 666         return 0;
 667  out_undo:
 668         mutex_unlock(&balloon_mutex);
 669         free_xenballooned_pages(pgno, pages);
 670         return ret;
 671 }
 672 EXPORT_SYMBOL(alloc_xenballooned_pages);
 673
 674 /**
 675  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 676  * @nr_pages: Number of pages
 677  * @pages: pages to return
 678  */
 679 void free_xenballooned_pages(int nr_pages, struct page **pages)
 680 {
 681         int i;
 682
 683         mutex_lock(&balloon_mutex);
 684
 685         for (i = 0; i < nr_pages; i++) {
 686                 if (pages[i])
 687                         balloon_append(pages[i]);
 688         }
 689
 690         balloon_stats.target_unpopulated -= nr_pages;
 691
 692         /* The balloon may be too large now. Shrink it if needed. */
 693         if (current_credit())
 694                 schedule_delayed_work(&balloon_worker, 0);
 695
 696         mutex_unlock(&balloon_mutex);
 697 }
 698 EXPORT_SYMBOL(free_xenballooned_pages);
 699
 700 static void __init balloon_add_region(unsigned long start_pfn,
 701                                       unsigned long pages)
 702 {
 703         unsigned long pfn, extra_pfn_end;
 704         struct page *page;
 705
 706         /*
 707          * If the amount of usable memory has been limited (e.g., with
 708          * the 'mem' command line parameter), don't add pages beyond
 709          * this limit.
 710          */
 711         extra_pfn_end = min(max_pfn, start_pfn + pages);
 712
 713         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 714                 page = pfn_to_page(pfn);
 715                 /* totalram_pages and totalhigh_pages do not
 716                    include the boot-time balloon extension, so
 717                    don't subtract from it. */
 718                 __balloon_append(page);
 719         }
 720
 721         balloon_stats.total_pages += extra_pfn_end - start_pfn;
 722 }
 723
 724 static int __init balloon_init(void)
 725 {
 726         int i;
 727
 728         if (!xen_domain())
 729                 return -ENODEV;
 730
 731         pr_info("Initialising balloon driver\n");
 732
 733         balloon_stats.current_pages = xen_pv_domain()
 734                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 735                 : get_num_physpages();
 736         balloon_stats.target_pages  = balloon_stats.current_pages;
 737         balloon_stats.balloon_low   = 0;
 738         balloon_stats.balloon_high  = 0;
 739         balloon_stats.total_pages   = balloon_stats.current_pages;
 740
 741         balloon_stats.schedule_delay = 1;
 742         balloon_stats.max_schedule_delay = 32;
 743         balloon_stats.retry_count = 1;
 744         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 745
 746 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 747         set_online_page_callback(&xen_online_page);
 748         register_memory_notifier(&xen_memory_nb);
 749         register_sysctl_table(xen_root);
 750 #endif
 751
 752         /*
 753          * Initialize the balloon with pages from the extra memory
 754          * regions (see arch/x86/xen/setup.c).
 755          */
 756         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 757                 if (xen_extra_mem[i].n_pfns)
 758                         balloon_add_region(xen_extra_mem[i].start_pfn,
 759                                            xen_extra_mem[i].n_pfns);
 760
 761         return 0;
 762 }
 763
 764 subsys_initcall(balloon_init);
 765
 766 MODULE_LICENSE("GPL");