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/bootmem.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
  62 #include <asm/page.h>
  63 #include <asm/pgalloc.h>
  64 #include <asm/pgtable.h>
  65 #include <asm/tlb.h>
  66
  67 #include <asm/xen/hypervisor.h>
  68 #include <asm/xen/hypercall.h>
  69
  70 #include <xen/xen.h>
  71 #include <xen/interface/xen.h>
  72 #include <xen/interface/memory.h>
  73 #include <xen/balloon.h>
  74 #include <xen/features.h>
  75 #include <xen/page.h>
  76
  77 static int xen_hotplug_unpopulated;
  78
  79 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
  80
  81 static int zero;
  82 static int one = 1;
  83
  84 static struct ctl_table balloon_table[] = {
  85         {
  86                 .procname       = "hotplug_unpopulated",
  87                 .data           = &xen_hotplug_unpopulated,
  88                 .maxlen         = sizeof(int),
  89                 .mode           = 0644,
  90                 .proc_handler   = proc_dointvec_minmax,
  91                 .extra1         = &zero,
  92                 .extra2         = &one,
  93         },
  94         { }
  95 };
  96
  97 static struct ctl_table balloon_root[] = {
  98         {
  99                 .procname       = "balloon",
 100                 .mode           = 0555,
 101                 .child          = balloon_table,
 102         },
 103         { }
 104 };
 105
 106 static struct ctl_table xen_root[] = {
 107         {
 108                 .procname       = "xen",
 109                 .mode           = 0555,
 110                 .child          = balloon_root,
 111         },
 112         { }
 113 };
 114
 115 #endif
 116
 117 /*
 118  * Use one extent per PAGE_SIZE to avoid to break down the page into
 119  * multiple frame.
 120  */
 121 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)
 122
 123 /*
 124  * balloon_thread() state:
 125  *
 126  * BP_DONE: done or nothing to do,
 127  * BP_WAIT: wait to be rescheduled,
 128  * BP_EAGAIN: error, go to sleep,
 129  * BP_ECANCELED: error, balloon operation canceled.
 130  */
 131
 132 enum bp_state {
 133         BP_DONE,
 134         BP_WAIT,
 135         BP_EAGAIN,
 136         BP_ECANCELED
 137 };
 138
 139 /* Main waiting point for xen-balloon thread. */
 140 static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);
 141
 142 static DEFINE_MUTEX(balloon_mutex);
 143
 144 struct balloon_stats balloon_stats;
 145 EXPORT_SYMBOL_GPL(balloon_stats);
 146
 147 /* We increase/decrease in batches which fit in a page */
 148 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];
 149
 150
 151 /* List of ballooned pages, threaded through the mem_map array. */
 152 static LIST_HEAD(ballooned_pages);
 153 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);
 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 }
 185
 186 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 187 static struct page *balloon_retrieve(bool require_lowmem)
 188 {
 189         struct page *page;
 190
 191         if (list_empty(&ballooned_pages))
 192                 return NULL;
 193
 194         page = list_entry(ballooned_pages.next, struct page, lru);
 195         if (require_lowmem && PageHighMem(page))
 196                 return NULL;
 197         list_del(&page->lru);
 198
 199         if (PageHighMem(page))
 200                 balloon_stats.balloon_high--;
 201         else
 202                 balloon_stats.balloon_low--;
 203
 204         return page;
 205 }
 206
 207 static struct page *balloon_next_page(struct page *page)
 208 {
 209         struct list_head *next = page->lru.next;
 210         if (next == &ballooned_pages)
 211                 return NULL;
 212         return list_entry(next, struct page, lru);
 213 }
 214
 215 static enum bp_state update_schedule(enum bp_state state)
 216 {
 217         if (state == BP_WAIT)
 218                 return BP_WAIT;
 219
 220         if (state == BP_ECANCELED)
 221                 return BP_ECANCELED;
 222
 223         if (state == BP_DONE) {
 224                 balloon_stats.schedule_delay = 1;
 225                 balloon_stats.retry_count = 1;
 226                 return BP_DONE;
 227         }
 228
 229         ++balloon_stats.retry_count;
 230
 231         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 232                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 233                 balloon_stats.schedule_delay = 1;
 234                 balloon_stats.retry_count = 1;
 235                 return BP_ECANCELED;
 236         }
 237
 238         balloon_stats.schedule_delay <<= 1;
 239
 240         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 241                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 242
 243         return BP_EAGAIN;
 244 }
 245
 246 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 247 static void release_memory_resource(struct resource *resource)
 248 {
 249         if (!resource)
 250                 return;
 251
 252         /*
 253          * No need to reset region to identity mapped since we now
 254          * know that no I/O can be in this region
 255          */
 256         release_resource(resource);
 257         kfree(resource);
 258 }
 259
 260 static struct resource *additional_memory_resource(phys_addr_t size)
 261 {
 262         struct resource *res;
 263         int ret;
 264
 265         res = kzalloc(sizeof(*res), GFP_KERNEL);
 266         if (!res)
 267                 return NULL;
 268
 269         res->name = "System RAM";
 270         res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
 271
 272         ret = allocate_resource(&iomem_resource, res,
 273                                 size, 0, -1,
 274                                 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
 275         if (ret < 0) {
 276                 pr_err("Cannot allocate new System RAM resource\n");
 277                 kfree(res);
 278                 return NULL;
 279         }
 280
 281 #ifdef CONFIG_SPARSEMEM
 282         {
 283                 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT);
 284                 unsigned long pfn = res->start >> PAGE_SHIFT;
 285
 286                 if (pfn > limit) {
 287                         pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n",
 288                                pfn, limit);
 289                         release_memory_resource(res);
 290                         return NULL;
 291                 }
 292         }
 293 #endif
 294
 295         return res;
 296 }
 297
 298 static enum bp_state reserve_additional_memory(void)
 299 {
 300         long credit;
 301         struct resource *resource;
 302         int nid, rc;
 303         unsigned long balloon_hotplug;
 304
 305         credit = balloon_stats.target_pages + balloon_stats.target_unpopulated
 306                 - balloon_stats.total_pages;
 307
 308         /*
 309          * Already hotplugged enough pages?  Wait for them to be
 310          * onlined.
 311          */
 312         if (credit <= 0)
 313                 return BP_WAIT;
 314
 315         balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
 316
 317         resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
 318         if (!resource)
 319                 goto err;
 320
 321         nid = memory_add_physaddr_to_nid(resource->start);
 322
 323 #ifdef CONFIG_XEN_HAVE_PVMMU
 324         /*
 325          * We don't support PV MMU when Linux and Xen is using
 326          * different page granularity.
 327          */
 328         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 329
 330         /*
 331          * add_memory() will build page tables for the new memory so
 332          * the p2m must contain invalid entries so the correct
 333          * non-present PTEs will be written.
 334          *
 335          * If a failure occurs, the original (identity) p2m entries
 336          * are not restored since this region is now known not to
 337          * conflict with any devices.
 338          */
 339         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 340                 unsigned long pfn, i;
 341
 342                 pfn = PFN_DOWN(resource->start);
 343                 for (i = 0; i < balloon_hotplug; i++) {
 344                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
 345                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
 346                                 goto err;
 347                         }
 348                 }
 349         }
 350 #endif
 351
 352         /*
 353          * add_memory_resource() will call online_pages() which in its turn
 354          * will call xen_online_page() callback causing deadlock if we don't
 355          * release balloon_mutex here. Unlocking here is safe because the
 356          * callers drop the mutex before trying again.
 357          */
 358         mutex_unlock(&balloon_mutex);
 359         /* add_memory_resource() requires the device_hotplug lock */
 360         lock_device_hotplug();
 361         rc = add_memory_resource(nid, resource, memhp_auto_online);
 362         unlock_device_hotplug();
 363         mutex_lock(&balloon_mutex);
 364
 365         if (rc) {
 366                 pr_warn("Cannot add additional memory (%i)\n", rc);
 367                 goto err;
 368         }
 369
 370         balloon_stats.total_pages += balloon_hotplug;
 371
 372         return BP_WAIT;
 373   err:
 374         release_memory_resource(resource);
 375         return BP_ECANCELED;
 376 }
 377
 378 static void xen_online_page(struct page *page)
 379 {
 380         __online_page_set_limits(page);
 381
 382         mutex_lock(&balloon_mutex);
 383
 384         __balloon_append(page);
 385
 386         mutex_unlock(&balloon_mutex);
 387 }
 388
 389 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 390 {
 391         if (val == MEM_ONLINE)
 392                 wake_up(&balloon_thread_wq);
 393
 394         return NOTIFY_OK;
 395 }
 396
 397 static struct notifier_block xen_memory_nb = {
 398         .notifier_call = xen_memory_notifier,
 399         .priority = 0
 400 };
 401 #else
 402 static enum bp_state reserve_additional_memory(void)
 403 {
 404         balloon_stats.target_pages = balloon_stats.current_pages +
 405                                      balloon_stats.target_unpopulated;
 406         return BP_ECANCELED;
 407 }
 408 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 409
 410 static long current_credit(void)
 411 {
 412         return balloon_stats.target_pages - balloon_stats.current_pages;
 413 }
 414
 415 static bool balloon_is_inflated(void)
 416 {
 417         return balloon_stats.balloon_low || balloon_stats.balloon_high;
 418 }
 419
 420 static enum bp_state increase_reservation(unsigned long nr_pages)
 421 {
 422         int rc;
 423         unsigned long i;
 424         struct page   *page;
 425         struct xen_memory_reservation reservation = {
 426                 .address_bits = 0,
 427                 .extent_order = EXTENT_ORDER,
 428                 .domid        = DOMID_SELF
 429         };
 430
 431         if (nr_pages > ARRAY_SIZE(frame_list))
 432                 nr_pages = ARRAY_SIZE(frame_list);
 433
 434         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 435         for (i = 0; i < nr_pages; i++) {
 436                 if (!page) {
 437                         nr_pages = i;
 438                         break;
 439                 }
 440
 441                 /* XENMEM_populate_physmap requires a PFN based on Xen
 442                  * granularity.
 443                  */
 444                 frame_list[i] = page_to_xen_pfn(page);
 445                 page = balloon_next_page(page);
 446         }
 447
 448         set_xen_guest_handle(reservation.extent_start, frame_list);
 449         reservation.nr_extents = nr_pages;
 450         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 451         if (rc <= 0)
 452                 return BP_EAGAIN;
 453
 454         for (i = 0; i < rc; i++) {
 455                 page = balloon_retrieve(false);
 456                 BUG_ON(page == NULL);
 457
 458 #ifdef CONFIG_XEN_HAVE_PVMMU
 459                 /*
 460                  * We don't support PV MMU when Linux and Xen is using
 461                  * different page granularity.
 462                  */
 463                 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 464
 465                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 466                         unsigned long pfn = page_to_pfn(page);
 467
 468                         set_phys_to_machine(pfn, frame_list[i]);
 469
 470                         /* Link back into the page tables if not highmem. */
 471                         if (!PageHighMem(page)) {
 472                                 int ret;
 473                                 ret = HYPERVISOR_update_va_mapping(
 474                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 475                                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 476                                                 0);
 477                                 BUG_ON(ret);
 478                         }
 479                 }
 480 #endif
 481
 482                 /* Relinquish the page back to the allocator. */
 483                 free_reserved_page(page);
 484         }
 485
 486         balloon_stats.current_pages += rc;
 487
 488         return BP_DONE;
 489 }
 490
 491 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 492 {
 493         enum bp_state state = BP_DONE;
 494         unsigned long i;
 495         struct page *page, *tmp;
 496         int ret;
 497         struct xen_memory_reservation reservation = {
 498                 .address_bits = 0,
 499                 .extent_order = EXTENT_ORDER,
 500                 .domid        = DOMID_SELF
 501         };
 502         LIST_HEAD(pages);
 503
 504         if (nr_pages > ARRAY_SIZE(frame_list))
 505                 nr_pages = ARRAY_SIZE(frame_list);
 506
 507         for (i = 0; i < nr_pages; i++) {
 508                 page = alloc_page(gfp);
 509                 if (page == NULL) {
 510                         nr_pages = i;
 511                         state = BP_EAGAIN;
 512                         break;
 513                 }
 514                 adjust_managed_page_count(page, -1);
 515                 scrub_page(page);
 516                 list_add(&page->lru, &pages);
 517         }
 518
 519         /*
 520          * Ensure that ballooned highmem pages don't have kmaps.
 521          *
 522          * Do this before changing the p2m as kmap_flush_unused()
 523          * reads PTEs to obtain pages (and hence needs the original
 524          * p2m entry).
 525          */
 526         kmap_flush_unused();
 527
 528         /*
 529          * Setup the frame, update direct mapping, invalidate P2M,
 530          * and add to balloon.
 531          */
 532         i = 0;
 533         list_for_each_entry_safe(page, tmp, &pages, lru) {
 534                 /* XENMEM_decrease_reservation requires a GFN */
 535                 frame_list[i++] = xen_page_to_gfn(page);
 536
 537 #ifdef CONFIG_XEN_HAVE_PVMMU
 538                 /*
 539                  * We don't support PV MMU when Linux and Xen is using
 540                  * different page granularity.
 541                  */
 542                 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 543
 544                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 545                         unsigned long pfn = page_to_pfn(page);
 546
 547                         if (!PageHighMem(page)) {
 548                                 ret = HYPERVISOR_update_va_mapping(
 549                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 550                                                 __pte_ma(0), 0);
 551                                 BUG_ON(ret);
 552                         }
 553                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 554                 }
 555 #endif
 556                 list_del(&page->lru);
 557
 558                 balloon_append(page);
 559         }
 560
 561         flush_tlb_all();
 562
 563         set_xen_guest_handle(reservation.extent_start, frame_list);
 564         reservation.nr_extents   = nr_pages;
 565         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 566         BUG_ON(ret != nr_pages);
 567
 568         balloon_stats.current_pages -= nr_pages;
 569
 570         return state;
 571 }
 572
 573 /*
 574  * Stop waiting if either state is BP_DONE and ballooning action is
 575  * needed, or if the credit has changed while state is not BP_DONE.
 576  */
 577 static bool balloon_thread_cond(enum bp_state state, long credit)
 578 {
 579         if (state == BP_DONE)
 580                 credit = 0;
 581
 582         return current_credit() != credit || kthread_should_stop();
 583 }
 584
 585 /*
 586  * As this is a kthread it is guaranteed to run as a single instance only.
 587  * We may of course race updates of the target counts (which are protected
 588  * by the balloon lock), or with changes to the Xen hard limit, but we will
 589  * recover from these in time.
 590  */
 591 static int balloon_thread(void *unused)
 592 {
 593         enum bp_state state = BP_DONE;
 594         long credit;
 595         unsigned long timeout;
 596
 597         set_freezable();
 598         for (;;) {
 599                 switch (state) {
 600                 case BP_DONE:
 601                 case BP_ECANCELED:
 602                         timeout = 3600 * HZ;
 603                         break;
 604                 case BP_EAGAIN:
 605                         timeout = balloon_stats.schedule_delay * HZ;
 606                         break;
 607                 case BP_WAIT:
 608                         timeout = HZ;
 609                         break;
 610                 }
 611
 612                 credit = current_credit();
 613
 614                 wait_event_freezable_timeout(balloon_thread_wq,
 615                         balloon_thread_cond(state, credit), timeout);
 616
 617                 if (kthread_should_stop())
 618                         return 0;
 619
 620                 mutex_lock(&balloon_mutex);
 621
 622                 credit = current_credit();
 623
 624                 if (credit > 0) {
 625                         if (balloon_is_inflated())
 626                                 state = increase_reservation(credit);
 627                         else
 628                                 state = reserve_additional_memory();
 629                 }
 630
 631                 if (credit < 0) {
 632                         long n_pages;
 633
 634                         n_pages = min(-credit, si_mem_available());
 635                         state = decrease_reservation(n_pages, GFP_BALLOON);
 636                         if (state == BP_DONE && n_pages != -credit &&
 637                             n_pages < totalreserve_pages)
 638                                 state = BP_EAGAIN;
 639                 }
 640
 641                 state = update_schedule(state);
 642
 643                 mutex_unlock(&balloon_mutex);
 644
 645                 cond_resched();
 646         }
 647 }
 648
 649 /* Resets the Xen limit, sets new target, and kicks off processing. */
 650 void balloon_set_new_target(unsigned long target)
 651 {
 652         /* No need for lock. Not read-modify-write updates. */
 653         balloon_stats.target_pages = target;
 654         wake_up(&balloon_thread_wq);
 655 }
 656 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 657
 658 static int add_ballooned_pages(int nr_pages)
 659 {
 660         enum bp_state st;
 661
 662         if (xen_hotplug_unpopulated) {
 663                 st = reserve_additional_memory();
 664                 if (st != BP_ECANCELED) {
 665                         int rc;
 666
 667                         mutex_unlock(&balloon_mutex);
 668                         rc = wait_event_interruptible(balloon_wq,
 669                                    !list_empty(&ballooned_pages));
 670                         mutex_lock(&balloon_mutex);
 671                         return rc ? -ENOMEM : 0;
 672                 }
 673         }
 674
 675         if (si_mem_available() < nr_pages)
 676                 return -ENOMEM;
 677
 678         st = decrease_reservation(nr_pages, GFP_USER);
 679         if (st != BP_DONE)
 680                 return -ENOMEM;
 681
 682         return 0;
 683 }
 684
 685 /**
 686  * alloc_xenballooned_pages - get pages that have been ballooned out
 687  * @nr_pages: Number of pages to get
 688  * @pages: pages returned
 689  * @return 0 on success, error otherwise
 690  */
 691 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
 692 {
 693         int pgno = 0;
 694         struct page *page;
 695         int ret;
 696
 697         mutex_lock(&balloon_mutex);
 698
 699         balloon_stats.target_unpopulated += nr_pages;
 700
 701         while (pgno < nr_pages) {
 702                 page = balloon_retrieve(true);
 703                 if (page) {
 704                         pages[pgno++] = page;
 705 #ifdef CONFIG_XEN_HAVE_PVMMU
 706                         /*
 707                          * We don't support PV MMU when Linux and Xen is using
 708                          * different page granularity.
 709                          */
 710                         BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);
 711
 712                         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 713                                 ret = xen_alloc_p2m_entry(page_to_pfn(page));
 714                                 if (ret < 0)
 715                                         goto out_undo;
 716                         }
 717 #endif
 718                 } else {
 719                         ret = add_ballooned_pages(nr_pages - pgno);
 720                         if (ret < 0)
 721                                 goto out_undo;
 722                 }
 723         }
 724         mutex_unlock(&balloon_mutex);
 725         return 0;
 726  out_undo:
 727         mutex_unlock(&balloon_mutex);
 728         free_xenballooned_pages(pgno, pages);
 729         /*
 730          * NB: free_xenballooned_pages will only subtract pgno pages, but since
 731          * target_unpopulated is incremented with nr_pages at the start we need
 732          * to remove the remaining ones also, or accounting will be screwed.
 733          */
 734         balloon_stats.target_unpopulated -= nr_pages - pgno;
 735         return ret;
 736 }
 737 EXPORT_SYMBOL(alloc_xenballooned_pages);
 738
 739 /**
 740  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 741  * @nr_pages: Number of pages
 742  * @pages: pages to return
 743  */
 744 void free_xenballooned_pages(int nr_pages, struct page **pages)
 745 {
 746         int i;
 747
 748         mutex_lock(&balloon_mutex);
 749
 750         for (i = 0; i < nr_pages; i++) {
 751                 if (pages[i])
 752                         balloon_append(pages[i]);
 753         }
 754
 755         balloon_stats.target_unpopulated -= nr_pages;
 756
 757         /* The balloon may be too large now. Shrink it if needed. */
 758         if (current_credit())
 759                 wake_up(&balloon_thread_wq);
 760
 761         mutex_unlock(&balloon_mutex);
 762 }
 763 EXPORT_SYMBOL(free_xenballooned_pages);
 764
 765 #ifdef CONFIG_XEN_PV
 766 static void __init balloon_add_region(unsigned long start_pfn,
 767                                       unsigned long pages)
 768 {
 769         unsigned long pfn, extra_pfn_end;
 770         struct page *page;
 771
 772         /*
 773          * If the amount of usable memory has been limited (e.g., with
 774          * the 'mem' command line parameter), don't add pages beyond
 775          * this limit.
 776          */
 777         extra_pfn_end = min(max_pfn, start_pfn + pages);
 778
 779         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 780                 page = pfn_to_page(pfn);
 781                 /* totalram_pages and totalhigh_pages do not
 782                    include the boot-time balloon extension, so
 783                    don't subtract from it. */
 784                 __balloon_append(page);
 785         }
 786
 787         balloon_stats.total_pages += extra_pfn_end - start_pfn;
 788 }
 789 #endif
 790
 791 static int __init balloon_init(void)
 792 {
 793         struct task_struct *task;
 794
 795         if (!xen_domain())
 796                 return -ENODEV;
 797
 798         pr_info("Initialising balloon driver\n");
 799
 800 #ifdef CONFIG_XEN_PV
 801         balloon_stats.current_pages = xen_pv_domain()
 802                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 803                 : get_num_physpages();
 804 #else
 805         balloon_stats.current_pages = get_num_physpages();
 806 #endif
 807         balloon_stats.target_pages  = balloon_stats.current_pages;
 808         balloon_stats.balloon_low   = 0;
 809         balloon_stats.balloon_high  = 0;
 810         balloon_stats.total_pages   = balloon_stats.current_pages;
 811
 812         balloon_stats.schedule_delay = 1;
 813         balloon_stats.max_schedule_delay = 32;
 814         balloon_stats.retry_count = 1;
 815         balloon_stats.max_retry_count = 4;
 816
 817 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 818         set_online_page_callback(&xen_online_page);
 819         register_memory_notifier(&xen_memory_nb);
 820         register_sysctl_table(xen_root);
 821 #endif
 822
 823 #ifdef CONFIG_XEN_PV
 824         {
 825                 int i;
 826
 827                 /*
 828                  * Initialize the balloon with pages from the extra memory
 829                  * regions (see arch/x86/xen/setup.c).
 830                  */
 831                 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 832                         if (xen_extra_mem[i].n_pfns)
 833                                 balloon_add_region(xen_extra_mem[i].start_pfn,
 834                                                    xen_extra_mem[i].n_pfns);
 835         }
 836 #endif
 837
 838         task = kthread_run(balloon_thread, NULL, "xen-balloon");
 839         if (IS_ERR(task)) {
 840                 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");
 841                 return PTR_ERR(task);
 842         }
 843
 844         /* Init the xen-balloon driver. */
 845         xen_balloon_init();
 846
 847         return 0;
 848 }
 849 subsys_initcall(balloon_init);