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/cred.h>
  45 #include <linux/errno.h>
  46 #include <linux/freezer.h>
  47 #include <linux/kthread.h>
  48 #include <linux/mm.h>
  49 #include <linux/memblock.h>
  50 #include <linux/pagemap.h>
  51 #include <linux/highmem.h>
  52 #include <linux/mutex.h>
  53 #include <linux/list.h>
  54 #include <linux/gfp.h>
  55 #include <linux/notifier.h>
  56 #include <linux/memory.h>
  57 #include <linux/memory_hotplug.h>
  58 #include <linux/percpu-defs.h>
  59 #include <linux/slab.h>
  60 #include <linux/sysctl.h>
  61 #include <linux/moduleparam.h>
  62
  63 #include <asm/page.h>
  64 #include <asm/tlb.h>
  65
  66 #include <asm/xen/hypervisor.h>
  67 #include <asm/xen/hypercall.h>
  68
  69 #include <xen/xen.h>
  70 #include <xen/interface/xen.h>
  71 #include <xen/interface/memory.h>
  72 #include <xen/balloon.h>
  73 #include <xen/features.h>
  74 #include <xen/page.h>
  75 #include <xen/mem-reservation.h>
  76
  77 #undef MODULE_PARAM_PREFIX
  78 #define MODULE_PARAM_PREFIX "xen."
  79
  80 static uint __read_mostly balloon_boot_timeout = 180;
  81 module_param(balloon_boot_timeout, uint, 0444);
  82
  83 static int xen_hotplug_unpopulated;
  84
  85 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
  86
  87 static struct ctl_table balloon_table[] = {
  88         {
  89                 .procname       = "hotplug_unpopulated",
  90                 .data           = &xen_hotplug_unpopulated,
  91                 .maxlen         = sizeof(int),
  92                 .mode           = 0644,
  93                 .proc_handler   = proc_dointvec_minmax,
  94                 .extra1         = SYSCTL_ZERO,
  95                 .extra2         = SYSCTL_ONE,
  96         },
  97         { }
  98 };
  99
 100 static struct ctl_table balloon_root[] = {
 101         {
 102                 .procname       = "balloon",
 103                 .mode           = 0555,
 104                 .child          = balloon_table,
 105         },
 106         { }
 107 };
 108
 109 static struct ctl_table xen_root[] = {
 110         {
 111                 .procname       = "xen",
 112                 .mode           = 0555,
 113                 .child          = balloon_root,
 114         },
 115         { }
 116 };
 117
 118 #endif
 119
 120 /*
 121  * Use one extent per PAGE_SIZE to avoid to break down the page into
 122  * multiple frame.
 123  */
 124 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
 125
 126 /*
 127  * balloon_thread() state:
 128  *
 129  * BP_DONE: done or nothing to do,
 130  * BP_WAIT: wait to be rescheduled,
 131  * BP_EAGAIN: error, go to sleep,
 132  * BP_ECANCELED: error, balloon operation canceled.
 133  */
 134
 135 static enum bp_state {
 136         BP_DONE,
 137         BP_WAIT,
 138         BP_EAGAIN,
 139         BP_ECANCELED
 140 } balloon_state = BP_DONE;
 141
 142 /* Main waiting point for xen-balloon thread. */
 143 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
 144
 145 static DEFINE_MUTEX(balloon_mutex);
 146
 147 struct balloon_stats balloon_stats;
 148 EXPORT_SYMBOL_GPL(balloon_stats);
 149
 150 /* We increase/decrease in batches which fit in a page */
 151 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
 152
 153
 154 /* List of ballooned pages, threaded through the mem_map array. */
 155 static LIST_HEAD(ballooned_pages);
 156 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
 157
 158 /* When ballooning out (allocating memory to return to Xen) we don't really
 159    want the kernel to try too hard since that can trigger the oom killer. */
 160 #define GFP_BALLOON \
 161         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 162
 163 /* balloon_append: add the given page to the balloon. */
 164 static void balloon_append(struct page *page)
 165 {
 166         __SetPageOffline(page);
 167
 168         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 169         if (PageHighMem(page)) {
 170                 list_add_tail(&page->lru, &ballooned_pages);
 171                 balloon_stats.balloon_high++;
 172         } else {
 173                 list_add(&page->lru, &ballooned_pages);
 174                 balloon_stats.balloon_low++;
 175         }
 176         wake_up(&balloon_wq);
 177 }
 178
 179 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 180 static struct page *balloon_retrieve(bool require_lowmem)
 181 {
 182         struct page *page;
 183
 184         if (list_empty(&ballooned_pages))
 185                 return NULL;
 186
 187         page = list_entry(ballooned_pages.next, struct page, lru);
 188         if (require_lowmem && PageHighMem(page))
 189                 return NULL;
 190         list_del(&page->lru);
 191
 192         if (PageHighMem(page))
 193                 balloon_stats.balloon_high--;
 194         else
 195                 balloon_stats.balloon_low--;
 196
 197         __ClearPageOffline(page);
 198         return page;
 199 }
 200
 201 static struct page *balloon_next_page(struct page *page)
 202 {
 203         struct list_head *next = page->lru.next;
 204         if (next == &ballooned_pages)
 205                 return NULL;
 206         return list_entry(next, struct page, lru);
 207 }
 208
 209 static void update_schedule(void)
 210 {
 211         if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
 212                 return;
 213
 214         if (balloon_state == BP_DONE) {
 215                 balloon_stats.schedule_delay = 1;
 216                 balloon_stats.retry_count = 1;
 217                 return;
 218         }
 219
 220         ++balloon_stats.retry_count;
 221
 222         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 223                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 224                 balloon_stats.schedule_delay = 1;
 225                 balloon_stats.retry_count = 1;
 226                 balloon_state = BP_ECANCELED;
 227                 return;
 228         }
 229
 230         balloon_stats.schedule_delay <<= 1;
 231
 232         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 233                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 234
 235         balloon_state = BP_EAGAIN;
 236 }
 237
 238 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 239 static void release_memory_resource(struct resource *resource)
 240 {
 241         if (!resource)
 242                 return;
 243
 244         /*
 245          * No need to reset region to identity mapped since we now
 246          * know that no I/O can be in this region
 247          */
 248         release_resource(resource);
 249         kfree(resource);
 250 }
 251
 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_SYSTEM_RAM | 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         return res;
 274 }
 275
 276 static enum bp_state reserve_additional_memory(void)
 277 {
 278         long credit;
 279         struct resource *resource;
 280         int nid, rc;
 281         unsigned long balloon_hotplug;
 282
 283         credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
 284                 - balloon_stats.total_pages;
 285
 286         /*
 287          * Already hotplugged enough pages?  Wait for them to be
 288          * onlined.
 289          */
 290         if (credit <= 0)
 291                 return BP_WAIT;
 292
 293         balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
 294
 295         resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
 296         if (!resource)
 297                 goto err;
 298
 299         nid = memory_add_physaddr_to_nid(resource->start);
 300
 301 #ifdef CONFIG_XEN_HAVE_PVMMU
 302         /*
 303          * We don't support PV MMU when Linux and Xen is using
 304          * different page granularity.
 305          */
 306         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 307
 308         /*
 309          * add_memory() will build page tables for the new memory so
 310          * the p2m must contain invalid entries so the correct
 311          * non-present PTEs will be written.
 312          *
 313          * If a failure occurs, the original (identity) p2m entries
 314          * are not restored since this region is now known not to
 315          * conflict with any devices.
 316          */
 317         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 318                 unsigned long pfn, i;
 319
 320                 pfn = PFN_DOWN(resource->start);
 321                 for (i = 0; i < balloon_hotplug; i++) {
 322                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
 323                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
 324                                 goto err;
 325                         }
 326                 }
 327         }
 328 #endif
 329
 330         /*
 331          * add_memory_resource() will call online_pages() which in its turn
 332          * will call xen_online_page() callback causing deadlock if we don't
 333          * release balloon_mutex here. Unlocking here is safe because the
 334          * callers drop the mutex before trying again.
 335          */
 336         mutex_unlock(&balloon_mutex);
 337         /* add_memory_resource() requires the device_hotplug lock */
 338         lock_device_hotplug();
 339         rc = add_memory_resource(nid, resource, MEMHP_MERGE_RESOURCE);
 340         unlock_device_hotplug();
 341         mutex_lock(&balloon_mutex);
 342
 343         if (rc) {
 344                 pr_warn("Cannot add additional memory (%i)\n", rc);
 345                 goto err;
 346         }
 347
 348         balloon_stats.total_pages += balloon_hotplug;
 349
 350         return BP_WAIT;
 351   err:
 352         release_memory_resource(resource);
 353         return BP_ECANCELED;
 354 }
 355
 356 static void xen_online_page(struct page *page, unsigned int order)
 357 {
 358         unsigned long i, size = (1 << order);
 359         unsigned long start_pfn = page_to_pfn(page);
 360         struct page *p;
 361
 362         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 363         mutex_lock(&balloon_mutex);
 364         for (i = 0; i < size; i++) {
 365                 p = pfn_to_page(start_pfn + i);
 366                 balloon_append(p);
 367         }
 368         mutex_unlock(&balloon_mutex);
 369 }
 370
 371 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 372 {
 373         if (val == MEM_ONLINE)
 374                 wake_up(&balloon_thread_wq);
 375
 376         return NOTIFY_OK;
 377 }
 378
 379 static struct notifier_block xen_memory_nb = {
 380         .notifier_call = xen_memory_notifier,
 381         .priority = 0
 382 };
 383 #else
 384 static enum bp_state reserve_additional_memory(void)
 385 {
 386         balloon_stats.target_pages = balloon_stats.current_pages +
 387                                      balloon_stats.target_unpopulated;
 388         return BP_ECANCELED;
 389 }
 390 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 391
 392 static long current_credit(void)
 393 {
 394         return balloon_stats.target_pages - balloon_stats.current_pages;
 395 }
 396
 397 static bool balloon_is_inflated(void)
 398 {
 399         return balloon_stats.balloon_low || balloon_stats.balloon_high;
 400 }
 401
 402 static enum bp_state increase_reservation(unsigned long nr_pages)
 403 {
 404         int rc;
 405         unsigned long i;
 406         struct page   *page;
 407
 408         if (nr_pages > ARRAY_SIZE(frame_list))
 409                 nr_pages = ARRAY_SIZE(frame_list);
 410
 411         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 412         for (i = 0; i < nr_pages; i++) {
 413                 if (!page) {
 414                         nr_pages = i;
 415                         break;
 416                 }
 417
 418                 frame_list[i] = page_to_xen_pfn(page);
 419                 page = balloon_next_page(page);
 420         }
 421
 422         rc = xenmem_reservation_increase(nr_pages, frame_list);
 423         if (rc <= 0)
 424                 return BP_EAGAIN;
 425
 426         for (i = 0; i < rc; i++) {
 427                 page = balloon_retrieve(false);
 428                 BUG_ON(page == NULL);
 429
 430                 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
 431
 432                 /* Relinquish the page back to the allocator. */
 433                 free_reserved_page(page);
 434         }
 435
 436         balloon_stats.current_pages += rc;
 437
 438         return BP_DONE;
 439 }
 440
 441 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 442 {
 443         enum bp_state state = BP_DONE;
 444         unsigned long i;
 445         struct page *page, *tmp;
 446         int ret;
 447         LIST_HEAD(pages);
 448
 449         if (nr_pages > ARRAY_SIZE(frame_list))
 450                 nr_pages = ARRAY_SIZE(frame_list);
 451
 452         for (i = 0; i < nr_pages; i++) {
 453                 page = alloc_page(gfp);
 454                 if (page == NULL) {
 455                         nr_pages = i;
 456                         state = BP_EAGAIN;
 457                         break;
 458                 }
 459                 adjust_managed_page_count(page, -1);
 460                 xenmem_reservation_scrub_page(page);
 461                 list_add(&page->lru, &pages);
 462         }
 463
 464         /*
 465          * Ensure that ballooned highmem pages don't have kmaps.
 466          *
 467          * Do this before changing the p2m as kmap_flush_unused()
 468          * reads PTEs to obtain pages (and hence needs the original
 469          * p2m entry).
 470          */
 471         kmap_flush_unused();
 472
 473         /*
 474          * Setup the frame, update direct mapping, invalidate P2M,
 475          * and add to balloon.
 476          */
 477         i = 0;
 478         list_for_each_entry_safe(page, tmp, &pages, lru) {
 479                 frame_list[i++] = xen_page_to_gfn(page);
 480
 481                 xenmem_reservation_va_mapping_reset(1, &page);
 482
 483                 list_del(&page->lru);
 484
 485                 balloon_append(page);
 486         }
 487
 488         flush_tlb_all();
 489
 490         ret = xenmem_reservation_decrease(nr_pages, frame_list);
 491         BUG_ON(ret != nr_pages);
 492
 493         balloon_stats.current_pages -= nr_pages;
 494
 495         return state;
 496 }
 497
 498 /*
 499  * Stop waiting if either state is BP_DONE and ballooning action is
 500  * needed, or if the credit has changed while state is not BP_DONE.
 501  */
 502 static bool balloon_thread_cond(long credit)
 503 {
 504         if (balloon_state == BP_DONE)
 505                 credit = 0;
 506
 507         return current_credit() != credit || kthread_should_stop();
 508 }
 509
 510 /*
 511  * As this is a kthread it is guaranteed to run as a single instance only.
 512  * We may of course race updates of the target counts (which are protected
 513  * by the balloon lock), or with changes to the Xen hard limit, but we will
 514  * recover from these in time.
 515  */
 516 static int balloon_thread(void *unused)
 517 {
 518         long credit;
 519         unsigned long timeout;
 520
 521         set_freezable();
 522         for (;;) {
 523                 switch (balloon_state) {
 524                 case BP_DONE:
 525                 case BP_ECANCELED:
 526                         timeout = 3600 * HZ;
 527                         break;
 528                 case BP_EAGAIN:
 529                         timeout = balloon_stats.schedule_delay * HZ;
 530                         break;
 531                 case BP_WAIT:
 532                         timeout = HZ;
 533                         break;
 534                 }
 535
 536                 credit = current_credit();
 537
 538                 wait_event_freezable_timeout(balloon_thread_wq,
 539                         balloon_thread_cond(credit), timeout);
 540
 541                 if (kthread_should_stop())
 542                         return 0;
 543
 544                 mutex_lock(&balloon_mutex);
 545
 546                 credit = current_credit();
 547
 548                 if (credit > 0) {
 549                         if (balloon_is_inflated())
 550                                 balloon_state = increase_reservation(credit);
 551                         else
 552                                 balloon_state = reserve_additional_memory();
 553                 }
 554
 555                 if (credit < 0) {
 556                         long n_pages;
 557
 558                         n_pages = min(-credit, si_mem_available());
 559                         balloon_state = decrease_reservation(n_pages,
 560                                                              GFP_BALLOON);
 561                         if (balloon_state == BP_DONE && n_pages != -credit &&
 562                             n_pages < totalreserve_pages)
 563                                 balloon_state = BP_EAGAIN;
 564                 }
 565
 566                 update_schedule();
 567
 568                 mutex_unlock(&balloon_mutex);
 569
 570                 cond_resched();
 571         }
 572 }
 573
 574 /* Resets the Xen limit, sets new target, and kicks off processing. */
 575 void balloon_set_new_target(unsigned long target)
 576 {
 577         /* No need for lock. Not read-modify-write updates. */
 578         balloon_stats.target_pages = target;
 579         wake_up(&balloon_thread_wq);
 580 }
 581 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 582
 583 static int add_ballooned_pages(int nr_pages)
 584 {
 585         enum bp_state st;
 586
 587         if (xen_hotplug_unpopulated) {
 588                 st = reserve_additional_memory();
 589                 if (st != BP_ECANCELED) {
 590                         int rc;
 591
 592                         mutex_unlock(&balloon_mutex);
 593                         rc = wait_event_interruptible(balloon_wq,
 594                                    !list_empty(&ballooned_pages));
 595                         mutex_lock(&balloon_mutex);
 596                         return rc ? -ENOMEM : 0;
 597                 }
 598         }
 599
 600         if (si_mem_available() < nr_pages)
 601                 return -ENOMEM;
 602
 603         st = decrease_reservation(nr_pages, GFP_USER);
 604         if (st != BP_DONE)
 605                 return -ENOMEM;
 606
 607         return 0;
 608 }
 609
 610 /**
 611  * alloc_xenballooned_pages - get pages that have been ballooned out
 612  * @nr_pages: Number of pages to get
 613  * @pages: pages returned
 614  * @return 0 on success, error otherwise
 615  */
 616 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
 617 {
 618         int pgno = 0;
 619         struct page *page;
 620         int ret;
 621
 622         mutex_lock(&balloon_mutex);
 623
 624         balloon_stats.target_unpopulated += nr_pages;
 625
 626         while (pgno < nr_pages) {
 627                 page = balloon_retrieve(true);
 628                 if (page) {
 629                         pages[pgno++] = page;
 630 #ifdef CONFIG_XEN_HAVE_PVMMU
 631                         /*
 632                          * We don't support PV MMU when Linux and Xen is using
 633                          * different page granularity.
 634                          */
 635                         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 636
 637                         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 638                                 ret = xen_alloc_p2m_entry(page_to_pfn(page));
 639                                 if (ret < 0)
 640                                         goto out_undo;
 641                         }
 642 #endif
 643                 } else {
 644                         ret = add_ballooned_pages(nr_pages - pgno);
 645                         if (ret < 0)
 646                                 goto out_undo;
 647                 }
 648         }
 649         mutex_unlock(&balloon_mutex);
 650         return 0;
 651  out_undo:
 652         mutex_unlock(&balloon_mutex);
 653         free_xenballooned_pages(pgno, pages);
 654         /*
 655          * NB: free_xenballooned_pages will only subtract pgno pages, but since
 656          * target_unpopulated is incremented with nr_pages at the start we need
 657          * to remove the remaining ones also, or accounting will be screwed.
 658          */
 659         balloon_stats.target_unpopulated -= nr_pages - pgno;
 660         return ret;
 661 }
 662 EXPORT_SYMBOL(alloc_xenballooned_pages);
 663
 664 /**
 665  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 666  * @nr_pages: Number of pages
 667  * @pages: pages to return
 668  */
 669 void free_xenballooned_pages(int nr_pages, struct page **pages)
 670 {
 671         int i;
 672
 673         mutex_lock(&balloon_mutex);
 674
 675         for (i = 0; i < nr_pages; i++) {
 676                 if (pages[i])
 677                         balloon_append(pages[i]);
 678         }
 679
 680         balloon_stats.target_unpopulated -= nr_pages;
 681
 682         /* The balloon may be too large now. Shrink it if needed. */
 683         if (current_credit())
 684                 wake_up(&balloon_thread_wq);
 685
 686         mutex_unlock(&balloon_mutex);
 687 }
 688 EXPORT_SYMBOL(free_xenballooned_pages);
 689
 690 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
 691 static void __init balloon_add_region(unsigned long start_pfn,
 692                                       unsigned long pages)
 693 {
 694         unsigned long pfn, extra_pfn_end;
 695
 696         /*
 697          * If the amount of usable memory has been limited (e.g., with
 698          * the 'mem' command line parameter), don't add pages beyond
 699          * this limit.
 700          */
 701         extra_pfn_end = min(max_pfn, start_pfn + pages);
 702
 703         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 704                 /* totalram_pages and totalhigh_pages do not
 705                    include the boot-time balloon extension, so
 706                    don't subtract from it. */
 707                 balloon_append(pfn_to_page(pfn));
 708         }
 709
 710         balloon_stats.total_pages += extra_pfn_end - start_pfn;
 711 }
 712 #endif
 713
 714 static int __init balloon_init(void)
 715 {
 716         struct task_struct *task;
 717
 718         if (!xen_domain())
 719                 return -ENODEV;
 720
 721         pr_info("Initialising balloon driver\n");
 722
 723 #ifdef CONFIG_XEN_PV
 724         balloon_stats.current_pages = xen_pv_domain()
 725                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 726                 : get_num_physpages();
 727 #else
 728         balloon_stats.current_pages = get_num_physpages();
 729 #endif
 730         balloon_stats.target_pages  = balloon_stats.current_pages;
 731         balloon_stats.balloon_low   = 0;
 732         balloon_stats.balloon_high  = 0;
 733         balloon_stats.total_pages   = balloon_stats.current_pages;
 734
 735         balloon_stats.schedule_delay = 1;
 736         balloon_stats.max_schedule_delay = 32;
 737         balloon_stats.retry_count = 1;
 738         balloon_stats.max_retry_count = 4;
 739
 740 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 741         set_online_page_callback(&xen_online_page);
 742         register_memory_notifier(&xen_memory_nb);
 743         register_sysctl_table(xen_root);
 744 #endif
 745
 746 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
 747         {
 748                 int i;
 749
 750                 /*
 751                  * Initialize the balloon with pages from the extra memory
 752                  * regions (see arch/x86/xen/setup.c).
 753                  */
 754                 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 755                         if (xen_extra_mem[i].n_pfns)
 756                                 balloon_add_region(xen_extra_mem[i].start_pfn,
 757                                                    xen_extra_mem[i].n_pfns);
 758         }
 759 #endif
 760
 761         task = kthread_run(balloon_thread, NULL, "xen-balloon");
 762         if (IS_ERR(task)) {
 763                 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
 764                 return PTR_ERR(task);
 765         }
 766
 767         /* Init the xen-balloon driver. */
 768         xen_balloon_init();
 769
 770         return 0;
 771 }
 772 subsys_initcall(balloon_init);
 773
 774 static int __init balloon_wait_finish(void)
 775 {
 776         long credit, last_credit = 0;
 777         unsigned long last_changed = 0;
 778
 779         if (!xen_domain())
 780                 return -ENODEV;
 781
 782         /* PV guests don't need to wait. */
 783         if (xen_pv_domain() || !current_credit())
 784                 return 0;
 785
 786         pr_notice("Waiting for initial ballooning down having finished.\n");
 787
 788         while ((credit = current_credit()) < 0) {
 789                 if (credit != last_credit) {
 790                         last_changed = jiffies;
 791                         last_credit = credit;
 792                 }
 793                 if (balloon_state == BP_ECANCELED) {
 794                         pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
 795                                      -credit);
 796                         if (jiffies - last_changed >= HZ * balloon_boot_timeout)
 797                                 panic("Initial ballooning failed!\n");
 798                 }
 799
 800                 schedule_timeout_interruptible(HZ / 10);
 801         }
 802
 803         pr_notice("Initial ballooning down finished.\n");
 804
 805         return 0;
 806 }
 807 late_initcall_sync(balloon_wait_finish);