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 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
  84 static int xen_hotplug_unpopulated;
  85
  86 static struct ctl_table balloon_table[] = {
  87         {
  88                 .procname       = "hotplug_unpopulated",
  89                 .data           = &xen_hotplug_unpopulated,
  90                 .maxlen         = sizeof(int),
  91                 .mode           = 0644,
  92                 .proc_handler   = proc_dointvec_minmax,
  93                 .extra1         = SYSCTL_ZERO,
  94                 .extra2         = SYSCTL_ONE,
  95         },
  96         { }
  97 };
  98
  99 static struct ctl_table balloon_root[] = {
 100         {
 101                 .procname       = "balloon",
 102                 .mode           = 0555,
 103                 .child          = balloon_table,
 104         },
 105         { }
 106 };
 107
 108 static struct ctl_table xen_root[] = {
 109         {
 110                 .procname       = "xen",
 111                 .mode           = 0555,
 112                 .child          = balloon_root,
 113         },
 114         { }
 115 };
 116
 117 #else
 118 #define xen_hotplug_unpopulated 0
 119 #endif
 120
 121 /*
 122  * Use one extent per PAGE_SIZE to avoid to break down the page into
 123  * multiple frame.
 124  */
 125 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
 126
 127 /*
 128  * balloon_thread() state:
 129  *
 130  * BP_DONE: done or nothing to do,
 131  * BP_WAIT: wait to be rescheduled,
 132  * BP_EAGAIN: error, go to sleep,
 133  * BP_ECANCELED: error, balloon operation canceled.
 134  */
 135
 136 static enum bp_state {
 137         BP_DONE,
 138         BP_WAIT,
 139         BP_EAGAIN,
 140         BP_ECANCELED
 141 } balloon_state = BP_DONE;
 142
 143 /* Main waiting point for xen-balloon thread. */
 144 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
 145
 146 static DEFINE_MUTEX(balloon_mutex);
 147
 148 struct balloon_stats balloon_stats;
 149 EXPORT_SYMBOL_GPL(balloon_stats);
 150
 151 /* We increase/decrease in batches which fit in a page */
 152 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
 153
 154
 155 /* List of ballooned pages, threaded through the mem_map array. */
 156 static LIST_HEAD(ballooned_pages);
 157 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
 158
 159 /* When ballooning out (allocating memory to return to Xen) we don't really
 160    want the kernel to try too hard since that can trigger the oom killer. */
 161 #define GFP_BALLOON \
 162         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 163
 164 /* balloon_append: add the given page to the balloon. */
 165 static void balloon_append(struct page *page)
 166 {
 167         __SetPageOffline(page);
 168
 169         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 170         if (PageHighMem(page)) {
 171                 list_add_tail(&page->lru, &ballooned_pages);
 172                 balloon_stats.balloon_high++;
 173         } else {
 174                 list_add(&page->lru, &ballooned_pages);
 175                 balloon_stats.balloon_low++;
 176         }
 177         wake_up(&balloon_wq);
 178 }
 179
 180 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 181 static struct page *balloon_retrieve(bool require_lowmem)
 182 {
 183         struct page *page;
 184
 185         if (list_empty(&ballooned_pages))
 186                 return NULL;
 187
 188         page = list_entry(ballooned_pages.next, struct page, lru);
 189         if (require_lowmem && PageHighMem(page))
 190                 return NULL;
 191         list_del(&page->lru);
 192
 193         if (PageHighMem(page))
 194                 balloon_stats.balloon_high--;
 195         else
 196                 balloon_stats.balloon_low--;
 197
 198         __ClearPageOffline(page);
 199         return page;
 200 }
 201
 202 static struct page *balloon_next_page(struct page *page)
 203 {
 204         struct list_head *next = page->lru.next;
 205         if (next == &ballooned_pages)
 206                 return NULL;
 207         return list_entry(next, struct page, lru);
 208 }
 209
 210 static void update_schedule(void)
 211 {
 212         if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)
 213                 return;
 214
 215         if (balloon_state == BP_DONE) {
 216                 balloon_stats.schedule_delay = 1;
 217                 balloon_stats.retry_count = 1;
 218                 return;
 219         }
 220
 221         ++balloon_stats.retry_count;
 222
 223         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 224                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 225                 balloon_stats.schedule_delay = 1;
 226                 balloon_stats.retry_count = 1;
 227                 balloon_state = BP_ECANCELED;
 228                 return;
 229         }
 230
 231         balloon_stats.schedule_delay <<= 1;
 232
 233         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 234                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 235
 236         balloon_state = BP_EAGAIN;
 237 }
 238
 239 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 240 static void release_memory_resource(struct resource *resource)
 241 {
 242         if (!resource)
 243                 return;
 244
 245         /*
 246          * No need to reset region to identity mapped since we now
 247          * know that no I/O can be in this region
 248          */
 249         release_resource(resource);
 250         kfree(resource);
 251 }
 252
 253 static struct resource *additional_memory_resource(phys_addr_t size)
 254 {
 255         struct resource *res;
 256         int ret;
 257
 258         res = kzalloc(sizeof(*res), GFP_KERNEL);
 259         if (!res)
 260                 return NULL;
 261
 262         res->name = "System RAM";
 263         res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 264
 265         ret = allocate_resource(&iomem_resource, res,
 266                                 size, 0, -1,
 267                                 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
 268         if (ret < 0) {
 269                 pr_err("Cannot allocate new System RAM resource\n");
 270                 kfree(res);
 271                 return NULL;
 272         }
 273
 274         return res;
 275 }
 276
 277 static enum bp_state reserve_additional_memory(void)
 278 {
 279         long credit;
 280         struct resource *resource;
 281         int nid, rc;
 282         unsigned long balloon_hotplug;
 283
 284         credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
 285                 - balloon_stats.total_pages;
 286
 287         /*
 288          * Already hotplugged enough pages?  Wait for them to be
 289          * onlined.
 290          */
 291         if (credit <= 0)
 292                 return BP_WAIT;
 293
 294         balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
 295
 296         resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
 297         if (!resource)
 298                 goto err;
 299
 300         nid = memory_add_physaddr_to_nid(resource->start);
 301
 302 #ifdef CONFIG_XEN_HAVE_PVMMU
 303         /*
 304          * We don't support PV MMU when Linux and Xen is using
 305          * different page granularity.
 306          */
 307         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 308
 309         /*
 310          * add_memory() will build page tables for the new memory so
 311          * the p2m must contain invalid entries so the correct
 312          * non-present PTEs will be written.
 313          *
 314          * If a failure occurs, the original (identity) p2m entries
 315          * are not restored since this region is now known not to
 316          * conflict with any devices.
 317          */
 318         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 319                 unsigned long pfn, i;
 320
 321                 pfn = PFN_DOWN(resource->start);
 322                 for (i = 0; i < balloon_hotplug; i++) {
 323                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
 324                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
 325                                 goto err;
 326                         }
 327                 }
 328         }
 329 #endif
 330
 331         /*
 332          * add_memory_resource() will call online_pages() which in its turn
 333          * will call xen_online_page() callback causing deadlock if we don't
 334          * release balloon_mutex here. Unlocking here is safe because the
 335          * callers drop the mutex before trying again.
 336          */
 337         mutex_unlock(&balloon_mutex);
 338         /* add_memory_resource() requires the device_hotplug lock */
 339         lock_device_hotplug();
 340         rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);
 341         unlock_device_hotplug();
 342         mutex_lock(&balloon_mutex);
 343
 344         if (rc) {
 345                 pr_warn("Cannot add additional memory (%i)\n", rc);
 346                 goto err;
 347         }
 348
 349         balloon_stats.total_pages += balloon_hotplug;
 350
 351         return BP_WAIT;
 352   err:
 353         release_memory_resource(resource);
 354         return BP_ECANCELED;
 355 }
 356
 357 static void xen_online_page(struct page *page, unsigned int order)
 358 {
 359         unsigned long i, size = (1 << order);
 360         unsigned long start_pfn = page_to_pfn(page);
 361         struct page *p;
 362
 363         pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);
 364         mutex_lock(&balloon_mutex);
 365         for (i = 0; i < size; i++) {
 366                 p = pfn_to_page(start_pfn + i);
 367                 balloon_append(p);
 368         }
 369         mutex_unlock(&balloon_mutex);
 370 }
 371
 372 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 373 {
 374         if (val == MEM_ONLINE)
 375                 wake_up(&balloon_thread_wq);
 376
 377         return NOTIFY_OK;
 378 }
 379
 380 static struct notifier_block xen_memory_nb = {
 381         .notifier_call = xen_memory_notifier,
 382         .priority = 0
 383 };
 384 #else
 385 static enum bp_state reserve_additional_memory(void)
 386 {
 387         balloon_stats.target_pages = balloon_stats.current_pages +
 388                                      balloon_stats.target_unpopulated;
 389         return BP_ECANCELED;
 390 }
 391 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 392
 393 static long current_credit(void)
 394 {
 395         return balloon_stats.target_pages - balloon_stats.current_pages;
 396 }
 397
 398 static bool balloon_is_inflated(void)
 399 {
 400         return balloon_stats.balloon_low || balloon_stats.balloon_high;
 401 }
 402
 403 static enum bp_state increase_reservation(unsigned long nr_pages)
 404 {
 405         int rc;
 406         unsigned long i;
 407         struct page   *page;
 408
 409         if (nr_pages > ARRAY_SIZE(frame_list))
 410                 nr_pages = ARRAY_SIZE(frame_list);
 411
 412         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 413         for (i = 0; i < nr_pages; i++) {
 414                 if (!page) {
 415                         nr_pages = i;
 416                         break;
 417                 }
 418
 419                 frame_list[i] = page_to_xen_pfn(page);
 420                 page = balloon_next_page(page);
 421         }
 422
 423         rc = xenmem_reservation_increase(nr_pages, frame_list);
 424         if (rc <= 0)
 425                 return BP_EAGAIN;
 426
 427         for (i = 0; i < rc; i++) {
 428                 page = balloon_retrieve(false);
 429                 BUG_ON(page == NULL);
 430
 431                 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);
 432
 433                 /* Relinquish the page back to the allocator. */
 434                 free_reserved_page(page);
 435         }
 436
 437         balloon_stats.current_pages += rc;
 438
 439         return BP_DONE;
 440 }
 441
 442 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 443 {
 444         enum bp_state state = BP_DONE;
 445         unsigned long i;
 446         struct page *page, *tmp;
 447         int ret;
 448         LIST_HEAD(pages);
 449
 450         if (nr_pages > ARRAY_SIZE(frame_list))
 451                 nr_pages = ARRAY_SIZE(frame_list);
 452
 453         for (i = 0; i < nr_pages; i++) {
 454                 page = alloc_page(gfp);
 455                 if (page == NULL) {
 456                         nr_pages = i;
 457                         state = BP_EAGAIN;
 458                         break;
 459                 }
 460                 adjust_managed_page_count(page, -1);
 461                 xenmem_reservation_scrub_page(page);
 462                 list_add(&page->lru, &pages);
 463         }
 464
 465         /*
 466          * Ensure that ballooned highmem pages don't have kmaps.
 467          *
 468          * Do this before changing the p2m as kmap_flush_unused()
 469          * reads PTEs to obtain pages (and hence needs the original
 470          * p2m entry).
 471          */
 472         kmap_flush_unused();
 473
 474         /*
 475          * Setup the frame, update direct mapping, invalidate P2M,
 476          * and add to balloon.
 477          */
 478         i = 0;
 479         list_for_each_entry_safe(page, tmp, &pages, lru) {
 480                 frame_list[i++] = xen_page_to_gfn(page);
 481
 482                 xenmem_reservation_va_mapping_reset(1, &page);
 483
 484                 list_del(&page->lru);
 485
 486                 balloon_append(page);
 487         }
 488
 489         flush_tlb_all();
 490
 491         ret = xenmem_reservation_decrease(nr_pages, frame_list);
 492         BUG_ON(ret != nr_pages);
 493
 494         balloon_stats.current_pages -= nr_pages;
 495
 496         return state;
 497 }
 498
 499 /*
 500  * Stop waiting if either state is BP_DONE and ballooning action is
 501  * needed, or if the credit has changed while state is not BP_DONE.
 502  */
 503 static bool balloon_thread_cond(long credit)
 504 {
 505         if (balloon_state == BP_DONE)
 506                 credit = 0;
 507
 508         return current_credit() != credit || kthread_should_stop();
 509 }
 510
 511 /*
 512  * As this is a kthread it is guaranteed to run as a single instance only.
 513  * We may of course race updates of the target counts (which are protected
 514  * by the balloon lock), or with changes to the Xen hard limit, but we will
 515  * recover from these in time.
 516  */
 517 static int balloon_thread(void *unused)
 518 {
 519         long credit;
 520         unsigned long timeout;
 521
 522         set_freezable();
 523         for (;;) {
 524                 switch (balloon_state) {
 525                 case BP_DONE:
 526                 case BP_ECANCELED:
 527                         timeout = 3600 * HZ;
 528                         break;
 529                 case BP_EAGAIN:
 530                         timeout = balloon_stats.schedule_delay * HZ;
 531                         break;
 532                 case BP_WAIT:
 533                         timeout = HZ;
 534                         break;
 535                 }
 536
 537                 credit = current_credit();
 538
 539                 wait_event_freezable_timeout(balloon_thread_wq,
 540                         balloon_thread_cond(credit), timeout);
 541
 542                 if (kthread_should_stop())
 543                         return 0;
 544
 545                 mutex_lock(&balloon_mutex);
 546
 547                 credit = current_credit();
 548
 549                 if (credit > 0) {
 550                         if (balloon_is_inflated())
 551                                 balloon_state = increase_reservation(credit);
 552                         else
 553                                 balloon_state = reserve_additional_memory();
 554                 }
 555
 556                 if (credit < 0) {
 557                         long n_pages;
 558
 559                         n_pages = min(-credit, si_mem_available());
 560                         balloon_state = decrease_reservation(n_pages,
 561                                                              GFP_BALLOON);
 562                         if (balloon_state == BP_DONE && n_pages != -credit &&
 563                             n_pages < totalreserve_pages)
 564                                 balloon_state = BP_EAGAIN;
 565                 }
 566
 567                 update_schedule();
 568
 569                 mutex_unlock(&balloon_mutex);
 570
 571                 cond_resched();
 572         }
 573 }
 574
 575 /* Resets the Xen limit, sets new target, and kicks off processing. */
 576 void balloon_set_new_target(unsigned long target)
 577 {
 578         /* No need for lock. Not read-modify-write updates. */
 579         balloon_stats.target_pages = target;
 580         wake_up(&balloon_thread_wq);
 581 }
 582 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 583
 584 #ifndef CONFIG_XEN_UNPOPULATED_ALLOC
 585 static int add_ballooned_pages(unsigned int nr_pages)
 586 {
 587         enum bp_state st;
 588
 589         if (xen_hotplug_unpopulated) {
 590                 st = reserve_additional_memory();
 591                 if (st != BP_ECANCELED) {
 592                         int rc;
 593
 594                         mutex_unlock(&balloon_mutex);
 595                         rc = wait_event_interruptible(balloon_wq,
 596                                    !list_empty(&ballooned_pages));
 597                         mutex_lock(&balloon_mutex);
 598                         return rc ? -ENOMEM : 0;
 599                 }
 600         }
 601
 602         if (si_mem_available() < nr_pages)
 603                 return -ENOMEM;
 604
 605         st = decrease_reservation(nr_pages, GFP_USER);
 606         if (st != BP_DONE)
 607                 return -ENOMEM;
 608
 609         return 0;
 610 }
 611
 612 /**
 613  * xen_alloc_unpopulated_pages - get pages that have been ballooned out
 614  * @nr_pages: Number of pages to get
 615  * @pages: pages returned
 616  * @return 0 on success, error otherwise
 617  */
 618 int xen_alloc_unpopulated_pages(unsigned int nr_pages, struct page **pages)
 619 {
 620         unsigned int pgno = 0;
 621         struct page *page;
 622         int ret;
 623
 624         mutex_lock(&balloon_mutex);
 625
 626         balloon_stats.target_unpopulated += nr_pages;
 627
 628         while (pgno < nr_pages) {
 629                 page = balloon_retrieve(true);
 630                 if (page) {
 631                         pages[pgno++] = page;
 632 #ifdef CONFIG_XEN_HAVE_PVMMU
 633                         /*
 634                          * We don't support PV MMU when Linux and Xen is using
 635                          * different page granularity.
 636                          */
 637                         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 638
 639                         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 640                                 ret = xen_alloc_p2m_entry(page_to_pfn(page));
 641                                 if (ret < 0)
 642                                         goto out_undo;
 643                         }
 644 #endif
 645                 } else {
 646                         ret = add_ballooned_pages(nr_pages - pgno);
 647                         if (ret < 0)
 648                                 goto out_undo;
 649                 }
 650         }
 651         mutex_unlock(&balloon_mutex);
 652         return 0;
 653  out_undo:
 654         mutex_unlock(&balloon_mutex);
 655         xen_free_unpopulated_pages(pgno, pages);
 656         /*
 657          * NB: free_xenballooned_pages will only subtract pgno pages, but since
 658          * target_unpopulated is incremented with nr_pages at the start we need
 659          * to remove the remaining ones also, or accounting will be screwed.
 660          */
 661         balloon_stats.target_unpopulated -= nr_pages - pgno;
 662         return ret;
 663 }
 664 EXPORT_SYMBOL(xen_alloc_unpopulated_pages);
 665
 666 /**
 667  * xen_free_unpopulated_pages - return pages retrieved with get_ballooned_pages
 668  * @nr_pages: Number of pages
 669  * @pages: pages to return
 670  */
 671 void xen_free_unpopulated_pages(unsigned int nr_pages, struct page **pages)
 672 {
 673         unsigned int i;
 674
 675         mutex_lock(&balloon_mutex);
 676
 677         for (i = 0; i < nr_pages; i++) {
 678                 if (pages[i])
 679                         balloon_append(pages[i]);
 680         }
 681
 682         balloon_stats.target_unpopulated -= nr_pages;
 683
 684         /* The balloon may be too large now. Shrink it if needed. */
 685         if (current_credit())
 686                 wake_up(&balloon_thread_wq);
 687
 688         mutex_unlock(&balloon_mutex);
 689 }
 690 EXPORT_SYMBOL(xen_free_unpopulated_pages);
 691
 692 #if defined(CONFIG_XEN_PV)
 693 static void __init balloon_add_region(unsigned long start_pfn,
 694                                       unsigned long pages)
 695 {
 696         unsigned long pfn, extra_pfn_end;
 697
 698         /*
 699          * If the amount of usable memory has been limited (e.g., with
 700          * the 'mem' command line parameter), don't add pages beyond
 701          * this limit.
 702          */
 703         extra_pfn_end = min(max_pfn, start_pfn + pages);
 704
 705         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 706                 /* totalram_pages and totalhigh_pages do not
 707                    include the boot-time balloon extension, so
 708                    don't subtract from it. */
 709                 balloon_append(pfn_to_page(pfn));
 710         }
 711
 712         balloon_stats.total_pages += extra_pfn_end - start_pfn;
 713 }
 714 #endif
 715 #endif
 716
 717 static int __init balloon_init(void)
 718 {
 719         struct task_struct *task;
 720
 721         if (!xen_domain())
 722                 return -ENODEV;
 723
 724         pr_info("Initialising balloon driver\n");
 725
 726 #ifdef CONFIG_XEN_PV
 727         balloon_stats.current_pages = xen_pv_domain()
 728                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 729                 : get_num_physpages();
 730 #else
 731         balloon_stats.current_pages = get_num_physpages();
 732 #endif
 733         balloon_stats.target_pages  = balloon_stats.current_pages;
 734         balloon_stats.balloon_low   = 0;
 735         balloon_stats.balloon_high  = 0;
 736         balloon_stats.total_pages   = balloon_stats.current_pages;
 737
 738         balloon_stats.schedule_delay = 1;
 739         balloon_stats.max_schedule_delay = 32;
 740         balloon_stats.retry_count = 1;
 741         balloon_stats.max_retry_count = 4;
 742
 743 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 744         set_online_page_callback(&xen_online_page);
 745         register_memory_notifier(&xen_memory_nb);
 746         register_sysctl_table(xen_root);
 747 #endif
 748
 749 #if defined(CONFIG_XEN_PV) && !defined(CONFIG_XEN_UNPOPULATED_ALLOC)
 750         {
 751                 int i;
 752
 753                 /*
 754                  * Initialize the balloon with pages from the extra memory
 755                  * regions (see arch/x86/xen/setup.c).
 756                  */
 757                 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 758                         if (xen_extra_mem[i].n_pfns)
 759                                 balloon_add_region(xen_extra_mem[i].start_pfn,
 760                                                    xen_extra_mem[i].n_pfns);
 761         }
 762 #endif
 763
 764         task = kthread_run(balloon_thread, NULL, "xen-balloon");
 765         if (IS_ERR(task)) {
 766                 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
 767                 return PTR_ERR(task);
 768         }
 769
 770         /* Init the xen-balloon driver. */
 771         xen_balloon_init();
 772
 773         return 0;
 774 }
 775 subsys_initcall(balloon_init);
 776
 777 static int __init balloon_wait_finish(void)
 778 {
 779         long credit, last_credit = 0;
 780         unsigned long last_changed = 0;
 781
 782         if (!xen_domain())
 783                 return -ENODEV;
 784
 785         /* PV guests don't need to wait. */
 786         if (xen_pv_domain() || !current_credit())
 787                 return 0;
 788
 789         pr_notice("Waiting for initial ballooning down having finished.\n");
 790
 791         while ((credit = current_credit()) < 0) {
 792                 if (credit != last_credit) {
 793                         last_changed = jiffies;
 794                         last_credit = credit;
 795                 }
 796                 if (balloon_state == BP_ECANCELED) {
 797                         pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",
 798                                      -credit);
 799                         if (jiffies - last_changed >= HZ * balloon_boot_timeout)
 800                                 panic("Initial ballooning failed!\n");
 801                 }
 802
 803                 schedule_timeout_interruptible(HZ / 10);
 804         }
 805
 806         pr_notice("Initial ballooning down finished.\n");
 807
 808         return 0;
 809 }
 810 late_initcall_sync(balloon_wait_finish);