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 #include <linux/kernel.h>
  40 #include <linux/sched.h>
  41 #include <linux/errno.h>
  42 #include <linux/module.h>
  43 #include <linux/mm.h>
  44 #include <linux/bootmem.h>
  45 #include <linux/pagemap.h>
  46 #include <linux/highmem.h>
  47 #include <linux/mutex.h>
  48 #include <linux/list.h>
  49 #include <linux/gfp.h>
  50 #include <linux/notifier.h>
  51 #include <linux/memory.h>
  52 #include <linux/memory_hotplug.h>
  53
  54 #include <asm/page.h>
  55 #include <asm/pgalloc.h>
  56 #include <asm/pgtable.h>
  57 #include <asm/tlb.h>
  58
  59 #include <asm/xen/hypervisor.h>
  60 #include <asm/xen/hypercall.h>
  61
  62 #include <xen/xen.h>
  63 #include <xen/interface/xen.h>
  64 #include <xen/interface/memory.h>
  65 #include <xen/balloon.h>
  66 #include <xen/features.h>
  67 #include <xen/page.h>
  68
  69 /*
  70  * balloon_process() state:
  71  *
  72  * BP_DONE: done or nothing to do,
  73  * BP_EAGAIN: error, go to sleep,
  74  * BP_ECANCELED: error, balloon operation canceled.
  75  */
  76
  77 enum bp_state {
  78         BP_DONE,
  79         BP_EAGAIN,
  80         BP_ECANCELED
  81 };
  82
  83
  84 static DEFINE_MUTEX(balloon_mutex);
  85
  86 struct balloon_stats balloon_stats;
  87 EXPORT_SYMBOL_GPL(balloon_stats);
  88
  89 /* We increase/decrease in batches which fit in a page */
  90 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
  91
  92 #ifdef CONFIG_HIGHMEM
  93 #define inc_totalhigh_pages() (totalhigh_pages++)
  94 #define dec_totalhigh_pages() (totalhigh_pages--)
  95 #else
  96 #define inc_totalhigh_pages() do {} while (0)
  97 #define dec_totalhigh_pages() do {} while (0)
  98 #endif
  99
 100 /* List of ballooned pages, threaded through the mem_map array. */
 101 static LIST_HEAD(ballooned_pages);
 102
 103 /* Main work function, always executed in process context. */
 104 static void balloon_process(struct work_struct *work);
 105 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
 106
 107 /* When ballooning out (allocating memory to return to Xen) we don't really
 108    want the kernel to try too hard since that can trigger the oom killer. */
 109 #define GFP_BALLOON \
 110         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 111
 112 static void scrub_page(struct page *page)
 113 {
 114 #ifdef CONFIG_XEN_SCRUB_PAGES
 115         clear_highpage(page);
 116 #endif
 117 }
 118
 119 /* balloon_append: add the given page to the balloon. */
 120 static void __balloon_append(struct page *page)
 121 {
 122         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 123         if (PageHighMem(page)) {
 124                 list_add_tail(&page->lru, &ballooned_pages);
 125                 balloon_stats.balloon_high++;
 126         } else {
 127                 list_add(&page->lru, &ballooned_pages);
 128                 balloon_stats.balloon_low++;
 129         }
 130 }
 131
 132 static void balloon_append(struct page *page)
 133 {
 134         __balloon_append(page);
 135         if (PageHighMem(page))
 136                 dec_totalhigh_pages();
 137         totalram_pages--;
 138 }
 139
 140 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 141 static struct page *balloon_retrieve(bool prefer_highmem)
 142 {
 143         struct page *page;
 144
 145         if (list_empty(&ballooned_pages))
 146                 return NULL;
 147
 148         if (prefer_highmem)
 149                 page = list_entry(ballooned_pages.prev, struct page, lru);
 150         else
 151                 page = list_entry(ballooned_pages.next, struct page, lru);
 152         list_del(&page->lru);
 153
 154         if (PageHighMem(page)) {
 155                 balloon_stats.balloon_high--;
 156                 inc_totalhigh_pages();
 157         } else
 158                 balloon_stats.balloon_low--;
 159
 160         totalram_pages++;
 161
 162         return page;
 163 }
 164
 165 static struct page *balloon_first_page(void)
 166 {
 167         if (list_empty(&ballooned_pages))
 168                 return NULL;
 169         return list_entry(ballooned_pages.next, struct page, lru);
 170 }
 171
 172 static struct page *balloon_next_page(struct page *page)
 173 {
 174         struct list_head *next = page->lru.next;
 175         if (next == &ballooned_pages)
 176                 return NULL;
 177         return list_entry(next, struct page, lru);
 178 }
 179
 180 static enum bp_state update_schedule(enum bp_state state)
 181 {
 182         if (state == BP_DONE) {
 183                 balloon_stats.schedule_delay = 1;
 184                 balloon_stats.retry_count = 1;
 185                 return BP_DONE;
 186         }
 187
 188         ++balloon_stats.retry_count;
 189
 190         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 191                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 192                 balloon_stats.schedule_delay = 1;
 193                 balloon_stats.retry_count = 1;
 194                 return BP_ECANCELED;
 195         }
 196
 197         balloon_stats.schedule_delay <<= 1;
 198
 199         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 200                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 201
 202         return BP_EAGAIN;
 203 }
 204
 205 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 206 static long current_credit(void)
 207 {
 208         return balloon_stats.target_pages - balloon_stats.current_pages -
 209                 balloon_stats.hotplug_pages;
 210 }
 211
 212 static bool balloon_is_inflated(void)
 213 {
 214         if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
 215                         balloon_stats.balloon_hotplug)
 216                 return true;
 217         else
 218                 return false;
 219 }
 220
 221 /*
 222  * reserve_additional_memory() adds memory region of size >= credit above
 223  * max_pfn. New region is section aligned and size is modified to be multiple
 224  * of section size. Those features allow optimal use of address space and
 225  * establish proper alignment when this function is called first time after
 226  * boot (last section not fully populated at boot time contains unused memory
 227  * pages with PG_reserved bit not set; online_pages_range() does not allow page
 228  * onlining in whole range if first onlined page does not have PG_reserved
 229  * bit set). Real size of added memory is established at page onlining stage.
 230  */
 231
 232 static enum bp_state reserve_additional_memory(long credit)
 233 {
 234         int nid, rc;
 235         u64 hotplug_start_paddr;
 236         unsigned long balloon_hotplug = credit;
 237
 238         hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
 239         balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
 240         nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
 241
 242         rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
 243
 244         if (rc) {
 245                 pr_info("xen_balloon: %s: add_memory() failed: %i\n", __func__, rc);
 246                 return BP_EAGAIN;
 247         }
 248
 249         balloon_hotplug -= credit;
 250
 251         balloon_stats.hotplug_pages += credit;
 252         balloon_stats.balloon_hotplug = balloon_hotplug;
 253
 254         return BP_DONE;
 255 }
 256
 257 static void xen_online_page(struct page *page)
 258 {
 259         __online_page_set_limits(page);
 260
 261         mutex_lock(&balloon_mutex);
 262
 263         __balloon_append(page);
 264
 265         if (balloon_stats.hotplug_pages)
 266                 --balloon_stats.hotplug_pages;
 267         else
 268                 --balloon_stats.balloon_hotplug;
 269
 270         mutex_unlock(&balloon_mutex);
 271 }
 272
 273 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 274 {
 275         if (val == MEM_ONLINE)
 276                 schedule_delayed_work(&balloon_worker, 0);
 277
 278         return NOTIFY_OK;
 279 }
 280
 281 static struct notifier_block xen_memory_nb = {
 282         .notifier_call = xen_memory_notifier,
 283         .priority = 0
 284 };
 285 #else
 286 static long current_credit(void)
 287 {
 288         unsigned long target = balloon_stats.target_pages;
 289
 290         target = min(target,
 291                      balloon_stats.current_pages +
 292                      balloon_stats.balloon_low +
 293                      balloon_stats.balloon_high);
 294
 295         return target - balloon_stats.current_pages;
 296 }
 297
 298 static bool balloon_is_inflated(void)
 299 {
 300         if (balloon_stats.balloon_low || balloon_stats.balloon_high)
 301                 return true;
 302         else
 303                 return false;
 304 }
 305
 306 static enum bp_state reserve_additional_memory(long credit)
 307 {
 308         balloon_stats.target_pages = balloon_stats.current_pages;
 309         return BP_DONE;
 310 }
 311 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 312
 313 static enum bp_state increase_reservation(unsigned long nr_pages)
 314 {
 315         int rc;
 316         unsigned long  pfn, i;
 317         struct page   *page;
 318         struct xen_memory_reservation reservation = {
 319                 .address_bits = 0,
 320                 .extent_order = 0,
 321                 .domid        = DOMID_SELF
 322         };
 323
 324 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 325         if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
 326                 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
 327                 balloon_stats.hotplug_pages += nr_pages;
 328                 balloon_stats.balloon_hotplug -= nr_pages;
 329                 return BP_DONE;
 330         }
 331 #endif
 332
 333         if (nr_pages > ARRAY_SIZE(frame_list))
 334                 nr_pages = ARRAY_SIZE(frame_list);
 335
 336         page = balloon_first_page();
 337         for (i = 0; i < nr_pages; i++) {
 338                 if (!page) {
 339                         nr_pages = i;
 340                         break;
 341                 }
 342                 frame_list[i] = page_to_pfn(page);
 343                 page = balloon_next_page(page);
 344         }
 345
 346         set_xen_guest_handle(reservation.extent_start, frame_list);
 347         reservation.nr_extents = nr_pages;
 348         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 349         if (rc <= 0)
 350                 return BP_EAGAIN;
 351
 352         for (i = 0; i < rc; i++) {
 353                 page = balloon_retrieve(false);
 354                 BUG_ON(page == NULL);
 355
 356                 pfn = page_to_pfn(page);
 357                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
 358                        phys_to_machine_mapping_valid(pfn));
 359
 360                 set_phys_to_machine(pfn, frame_list[i]);
 361
 362 #ifdef CONFIG_XEN_HAVE_PVMMU
 363                 /* Link back into the page tables if not highmem. */
 364                 if (xen_pv_domain() && !PageHighMem(page)) {
 365                         int ret;
 366                         ret = HYPERVISOR_update_va_mapping(
 367                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 368                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 369                                 0);
 370                         BUG_ON(ret);
 371                 }
 372 #endif
 373
 374                 /* Relinquish the page back to the allocator. */
 375                 ClearPageReserved(page);
 376                 init_page_count(page);
 377                 __free_page(page);
 378         }
 379
 380         balloon_stats.current_pages += rc;
 381
 382         return BP_DONE;
 383 }
 384
 385 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 386 {
 387         enum bp_state state = BP_DONE;
 388         unsigned long  pfn, i;
 389         struct page   *page;
 390         int ret;
 391         struct xen_memory_reservation reservation = {
 392                 .address_bits = 0,
 393                 .extent_order = 0,
 394                 .domid        = DOMID_SELF
 395         };
 396
 397 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 398         if (balloon_stats.hotplug_pages) {
 399                 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
 400                 balloon_stats.hotplug_pages -= nr_pages;
 401                 balloon_stats.balloon_hotplug += nr_pages;
 402                 return BP_DONE;
 403         }
 404 #endif
 405
 406         if (nr_pages > ARRAY_SIZE(frame_list))
 407                 nr_pages = ARRAY_SIZE(frame_list);
 408
 409         for (i = 0; i < nr_pages; i++) {
 410                 if ((page = alloc_page(gfp)) == NULL) {
 411                         nr_pages = i;
 412                         state = BP_EAGAIN;
 413                         break;
 414                 }
 415
 416                 pfn = page_to_pfn(page);
 417                 frame_list[i] = pfn_to_mfn(pfn);
 418
 419                 scrub_page(page);
 420
 421 #ifdef CONFIG_XEN_HAVE_PVMMU
 422                 if (xen_pv_domain() && !PageHighMem(page)) {
 423                         ret = HYPERVISOR_update_va_mapping(
 424                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 425                                 __pte_ma(0), 0);
 426                         BUG_ON(ret);
 427                 }
 428 #endif
 429         }
 430
 431         /* Ensure that ballooned highmem pages don't have kmaps. */
 432         kmap_flush_unused();
 433         flush_tlb_all();
 434
 435         /* No more mappings: invalidate P2M and add to balloon. */
 436         for (i = 0; i < nr_pages; i++) {
 437                 pfn = mfn_to_pfn(frame_list[i]);
 438                 __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 439                 balloon_append(pfn_to_page(pfn));
 440         }
 441
 442         set_xen_guest_handle(reservation.extent_start, frame_list);
 443         reservation.nr_extents   = nr_pages;
 444         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 445         BUG_ON(ret != nr_pages);
 446
 447         balloon_stats.current_pages -= nr_pages;
 448
 449         return state;
 450 }
 451
 452 /*
 453  * We avoid multiple worker processes conflicting via the balloon mutex.
 454  * We may of course race updates of the target counts (which are protected
 455  * by the balloon lock), or with changes to the Xen hard limit, but we will
 456  * recover from these in time.
 457  */
 458 static void balloon_process(struct work_struct *work)
 459 {
 460         enum bp_state state = BP_DONE;
 461         long credit;
 462
 463         mutex_lock(&balloon_mutex);
 464
 465         do {
 466                 credit = current_credit();
 467
 468                 if (credit > 0) {
 469                         if (balloon_is_inflated())
 470                                 state = increase_reservation(credit);
 471                         else
 472                                 state = reserve_additional_memory(credit);
 473                 }
 474
 475                 if (credit < 0)
 476                         state = decrease_reservation(-credit, GFP_BALLOON);
 477
 478                 state = update_schedule(state);
 479
 480 #ifndef CONFIG_PREEMPT
 481                 if (need_resched())
 482                         schedule();
 483 #endif
 484         } while (credit && state == BP_DONE);
 485
 486         /* Schedule more work if there is some still to be done. */
 487         if (state == BP_EAGAIN)
 488                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 489
 490         mutex_unlock(&balloon_mutex);
 491 }
 492
 493 /* Resets the Xen limit, sets new target, and kicks off processing. */
 494 void balloon_set_new_target(unsigned long target)
 495 {
 496         /* No need for lock. Not read-modify-write updates. */
 497         balloon_stats.target_pages = target;
 498         schedule_delayed_work(&balloon_worker, 0);
 499 }
 500 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 501
 502 /**
 503  * alloc_xenballooned_pages - get pages that have been ballooned out
 504  * @nr_pages: Number of pages to get
 505  * @pages: pages returned
 506  * @highmem: allow highmem pages
 507  * @return 0 on success, error otherwise
 508  */
 509 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
 510 {
 511         int pgno = 0;
 512         struct page *page;
 513         mutex_lock(&balloon_mutex);
 514         while (pgno < nr_pages) {
 515                 page = balloon_retrieve(highmem);
 516                 if (page && (highmem || !PageHighMem(page))) {
 517                         pages[pgno++] = page;
 518                 } else {
 519                         enum bp_state st;
 520                         if (page)
 521                                 balloon_append(page);
 522                         st = decrease_reservation(nr_pages - pgno,
 523                                         highmem ? GFP_HIGHUSER : GFP_USER);
 524                         if (st != BP_DONE)
 525                                 goto out_undo;
 526                 }
 527         }
 528         mutex_unlock(&balloon_mutex);
 529         return 0;
 530  out_undo:
 531         while (pgno)
 532                 balloon_append(pages[--pgno]);
 533         /* Free the memory back to the kernel soon */
 534         schedule_delayed_work(&balloon_worker, 0);
 535         mutex_unlock(&balloon_mutex);
 536         return -ENOMEM;
 537 }
 538 EXPORT_SYMBOL(alloc_xenballooned_pages);
 539
 540 /**
 541  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 542  * @nr_pages: Number of pages
 543  * @pages: pages to return
 544  */
 545 void free_xenballooned_pages(int nr_pages, struct page **pages)
 546 {
 547         int i;
 548
 549         mutex_lock(&balloon_mutex);
 550
 551         for (i = 0; i < nr_pages; i++) {
 552                 if (pages[i])
 553                         balloon_append(pages[i]);
 554         }
 555
 556         /* The balloon may be too large now. Shrink it if needed. */
 557         if (current_credit())
 558                 schedule_delayed_work(&balloon_worker, 0);
 559
 560         mutex_unlock(&balloon_mutex);
 561 }
 562 EXPORT_SYMBOL(free_xenballooned_pages);
 563
 564 static void __init balloon_add_region(unsigned long start_pfn,
 565                                       unsigned long pages)
 566 {
 567         unsigned long pfn, extra_pfn_end;
 568         struct page *page;
 569
 570         /*
 571          * If the amount of usable memory has been limited (e.g., with
 572          * the 'mem' command line parameter), don't add pages beyond
 573          * this limit.
 574          */
 575         extra_pfn_end = min(max_pfn, start_pfn + pages);
 576
 577         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 578                 page = pfn_to_page(pfn);
 579                 /* totalram_pages and totalhigh_pages do not
 580                    include the boot-time balloon extension, so
 581                    don't subtract from it. */
 582                 __balloon_append(page);
 583         }
 584 }
 585
 586 static int __init balloon_init(void)
 587 {
 588         int i;
 589
 590         if (!xen_domain())
 591                 return -ENODEV;
 592
 593         pr_info("xen/balloon: Initialising balloon driver.\n");
 594
 595         balloon_stats.current_pages = xen_pv_domain()
 596                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 597                 : max_pfn;
 598         balloon_stats.target_pages  = balloon_stats.current_pages;
 599         balloon_stats.balloon_low   = 0;
 600         balloon_stats.balloon_high  = 0;
 601
 602         balloon_stats.schedule_delay = 1;
 603         balloon_stats.max_schedule_delay = 32;
 604         balloon_stats.retry_count = 1;
 605         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 606
 607 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 608         balloon_stats.hotplug_pages = 0;
 609         balloon_stats.balloon_hotplug = 0;
 610
 611         set_online_page_callback(&xen_online_page);
 612         register_memory_notifier(&xen_memory_nb);
 613 #endif
 614
 615         /*
 616          * Initialize the balloon with pages from the extra memory
 617          * regions (see arch/x86/xen/setup.c).
 618          */
 619         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 620                 if (xen_extra_mem[i].size)
 621                         balloon_add_region(PFN_UP(xen_extra_mem[i].start),
 622                                            PFN_DOWN(xen_extra_mem[i].size));
 623
 624         return 0;
 625 }
 626
 627 subsys_initcall(balloon_init);
 628
 629 MODULE_LICENSE("GPL");