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