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/errno.h>
  45 #include <linux/module.h>
  46 #include <linux/mm.h>
  47 #include <linux/bootmem.h>
  48 #include <linux/pagemap.h>
  49 #include <linux/highmem.h>
  50 #include <linux/mutex.h>
  51 #include <linux/list.h>
  52 #include <linux/gfp.h>
  53 #include <linux/notifier.h>
  54 #include <linux/memory.h>
  55 #include <linux/memory_hotplug.h>
  56 #include <linux/percpu-defs.h>
  57
  58 #include <asm/page.h>
  59 #include <asm/pgalloc.h>
  60 #include <asm/pgtable.h>
  61 #include <asm/tlb.h>
  62
  63 #include <asm/xen/hypervisor.h>
  64 #include <asm/xen/hypercall.h>
  65
  66 #include <xen/xen.h>
  67 #include <xen/interface/xen.h>
  68 #include <xen/interface/memory.h>
  69 #include <xen/balloon.h>
  70 #include <xen/features.h>
  71 #include <xen/page.h>
  72
  73 /*
  74  * balloon_process() state:
  75  *
  76  * BP_DONE: done or nothing to do,
  77  * BP_EAGAIN: error, go to sleep,
  78  * BP_ECANCELED: error, balloon operation canceled.
  79  */
  80
  81 enum bp_state {
  82         BP_DONE,
  83         BP_EAGAIN,
  84         BP_ECANCELED
  85 };
  86
  87
  88 static DEFINE_MUTEX(balloon_mutex);
  89
  90 struct balloon_stats balloon_stats;
  91 EXPORT_SYMBOL_GPL(balloon_stats);
  92
  93 /* We increase/decrease in batches which fit in a page */
  94 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
  95
  96
  97 /* List of ballooned pages, threaded through the mem_map array. */
  98 static LIST_HEAD(ballooned_pages);
  99
 100 /* Main work function, always executed in process context. */
 101 static void balloon_process(struct work_struct *work);
 102 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
 103
 104 /* When ballooning out (allocating memory to return to Xen) we don't really
 105    want the kernel to try too hard since that can trigger the oom killer. */
 106 #define GFP_BALLOON \
 107         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 108
 109 static void scrub_page(struct page *page)
 110 {
 111 #ifdef CONFIG_XEN_SCRUB_PAGES
 112         clear_highpage(page);
 113 #endif
 114 }
 115
 116 /* balloon_append: add the given page to the balloon. */
 117 static void __balloon_append(struct page *page)
 118 {
 119         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 120         if (PageHighMem(page)) {
 121                 list_add_tail(&page->lru, &ballooned_pages);
 122                 balloon_stats.balloon_high++;
 123         } else {
 124                 list_add(&page->lru, &ballooned_pages);
 125                 balloon_stats.balloon_low++;
 126         }
 127 }
 128
 129 static void balloon_append(struct page *page)
 130 {
 131         __balloon_append(page);
 132         adjust_managed_page_count(page, -1);
 133 }
 134
 135 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 136 static struct page *balloon_retrieve(bool prefer_highmem)
 137 {
 138         struct page *page;
 139
 140         if (list_empty(&ballooned_pages))
 141                 return NULL;
 142
 143         if (prefer_highmem)
 144                 page = list_entry(ballooned_pages.prev, struct page, lru);
 145         else
 146                 page = list_entry(ballooned_pages.next, struct page, lru);
 147         list_del(&page->lru);
 148
 149         if (PageHighMem(page))
 150                 balloon_stats.balloon_high--;
 151         else
 152                 balloon_stats.balloon_low--;
 153
 154         adjust_managed_page_count(page, 1);
 155
 156         return page;
 157 }
 158
 159 static struct page *balloon_next_page(struct page *page)
 160 {
 161         struct list_head *next = page->lru.next;
 162         if (next == &ballooned_pages)
 163                 return NULL;
 164         return list_entry(next, struct page, lru);
 165 }
 166
 167 static enum bp_state update_schedule(enum bp_state state)
 168 {
 169         if (state == BP_ECANCELED)
 170                 return BP_ECANCELED;
 171
 172         if (state == BP_DONE) {
 173                 balloon_stats.schedule_delay = 1;
 174                 balloon_stats.retry_count = 1;
 175                 return BP_DONE;
 176         }
 177
 178         ++balloon_stats.retry_count;
 179
 180         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 181                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 182                 balloon_stats.schedule_delay = 1;
 183                 balloon_stats.retry_count = 1;
 184                 return BP_ECANCELED;
 185         }
 186
 187         balloon_stats.schedule_delay <<= 1;
 188
 189         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 190                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 191
 192         return BP_EAGAIN;
 193 }
 194
 195 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 196 static long current_credit(void)
 197 {
 198         return balloon_stats.target_pages - balloon_stats.current_pages -
 199                 balloon_stats.hotplug_pages;
 200 }
 201
 202 static bool balloon_is_inflated(void)
 203 {
 204         if (balloon_stats.balloon_low || balloon_stats.balloon_high ||
 205                         balloon_stats.balloon_hotplug)
 206                 return true;
 207         else
 208                 return false;
 209 }
 210
 211 /*
 212  * reserve_additional_memory() adds memory region of size >= credit above
 213  * max_pfn. New region is section aligned and size is modified to be multiple
 214  * of section size. Those features allow optimal use of address space and
 215  * establish proper alignment when this function is called first time after
 216  * boot (last section not fully populated at boot time contains unused memory
 217  * pages with PG_reserved bit not set; online_pages_range() does not allow page
 218  * onlining in whole range if first onlined page does not have PG_reserved
 219  * bit set). Real size of added memory is established at page onlining stage.
 220  */
 221
 222 static enum bp_state reserve_additional_memory(long credit)
 223 {
 224         int nid, rc;
 225         u64 hotplug_start_paddr;
 226         unsigned long balloon_hotplug = credit;
 227
 228         hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn));
 229         balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION);
 230         nid = memory_add_physaddr_to_nid(hotplug_start_paddr);
 231
 232         rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
 233
 234         if (rc) {
 235                 pr_warn("Cannot add additional memory (%i)\n", rc);
 236                 return BP_ECANCELED;
 237         }
 238
 239         balloon_hotplug -= credit;
 240
 241         balloon_stats.hotplug_pages += credit;
 242         balloon_stats.balloon_hotplug = balloon_hotplug;
 243
 244         return BP_DONE;
 245 }
 246
 247 static void xen_online_page(struct page *page)
 248 {
 249         __online_page_set_limits(page);
 250
 251         mutex_lock(&balloon_mutex);
 252
 253         __balloon_append(page);
 254
 255         if (balloon_stats.hotplug_pages)
 256                 --balloon_stats.hotplug_pages;
 257         else
 258                 --balloon_stats.balloon_hotplug;
 259
 260         mutex_unlock(&balloon_mutex);
 261 }
 262
 263 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 264 {
 265         if (val == MEM_ONLINE)
 266                 schedule_delayed_work(&balloon_worker, 0);
 267
 268         return NOTIFY_OK;
 269 }
 270
 271 static struct notifier_block xen_memory_nb = {
 272         .notifier_call = xen_memory_notifier,
 273         .priority = 0
 274 };
 275 #else
 276 static long current_credit(void)
 277 {
 278         unsigned long target = balloon_stats.target_pages;
 279
 280         target = min(target,
 281                      balloon_stats.current_pages +
 282                      balloon_stats.balloon_low +
 283                      balloon_stats.balloon_high);
 284
 285         return target - balloon_stats.current_pages;
 286 }
 287
 288 static bool balloon_is_inflated(void)
 289 {
 290         if (balloon_stats.balloon_low || balloon_stats.balloon_high)
 291                 return true;
 292         else
 293                 return false;
 294 }
 295
 296 static enum bp_state reserve_additional_memory(long credit)
 297 {
 298         balloon_stats.target_pages = balloon_stats.current_pages;
 299         return BP_DONE;
 300 }
 301 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 302
 303 static enum bp_state increase_reservation(unsigned long nr_pages)
 304 {
 305         int rc;
 306         unsigned long  pfn, i;
 307         struct page   *page;
 308         struct xen_memory_reservation reservation = {
 309                 .address_bits = 0,
 310                 .extent_order = 0,
 311                 .domid        = DOMID_SELF
 312         };
 313
 314 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 315         if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
 316                 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
 317                 balloon_stats.hotplug_pages += nr_pages;
 318                 balloon_stats.balloon_hotplug -= nr_pages;
 319                 return BP_DONE;
 320         }
 321 #endif
 322
 323         if (nr_pages > ARRAY_SIZE(frame_list))
 324                 nr_pages = ARRAY_SIZE(frame_list);
 325
 326         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 327         for (i = 0; i < nr_pages; i++) {
 328                 if (!page) {
 329                         nr_pages = i;
 330                         break;
 331                 }
 332                 frame_list[i] = page_to_pfn(page);
 333                 page = balloon_next_page(page);
 334         }
 335
 336         set_xen_guest_handle(reservation.extent_start, frame_list);
 337         reservation.nr_extents = nr_pages;
 338         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 339         if (rc <= 0)
 340                 return BP_EAGAIN;
 341
 342         for (i = 0; i < rc; i++) {
 343                 page = balloon_retrieve(false);
 344                 BUG_ON(page == NULL);
 345
 346                 pfn = page_to_pfn(page);
 347
 348 #ifdef CONFIG_XEN_HAVE_PVMMU
 349                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 350                         set_phys_to_machine(pfn, frame_list[i]);
 351
 352                         /* Link back into the page tables if not highmem. */
 353                         if (!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                 }
 362 #endif
 363
 364                 /* Relinquish the page back to the allocator. */
 365                 __free_reserved_page(page);
 366         }
 367
 368         balloon_stats.current_pages += rc;
 369
 370         return BP_DONE;
 371 }
 372
 373 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 374 {
 375         enum bp_state state = BP_DONE;
 376         unsigned long  pfn, i;
 377         struct page   *page;
 378         int ret;
 379         struct xen_memory_reservation reservation = {
 380                 .address_bits = 0,
 381                 .extent_order = 0,
 382                 .domid        = DOMID_SELF
 383         };
 384
 385 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 386         if (balloon_stats.hotplug_pages) {
 387                 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
 388                 balloon_stats.hotplug_pages -= nr_pages;
 389                 balloon_stats.balloon_hotplug += nr_pages;
 390                 return BP_DONE;
 391         }
 392 #endif
 393
 394         if (nr_pages > ARRAY_SIZE(frame_list))
 395                 nr_pages = ARRAY_SIZE(frame_list);
 396
 397         for (i = 0; i < nr_pages; i++) {
 398                 page = alloc_page(gfp);
 399                 if (page == NULL) {
 400                         nr_pages = i;
 401                         state = BP_EAGAIN;
 402                         break;
 403                 }
 404                 scrub_page(page);
 405
 406                 frame_list[i] = page_to_pfn(page);
 407         }
 408
 409         /*
 410          * Ensure that ballooned highmem pages don't have kmaps.
 411          *
 412          * Do this before changing the p2m as kmap_flush_unused()
 413          * reads PTEs to obtain pages (and hence needs the original
 414          * p2m entry).
 415          */
 416         kmap_flush_unused();
 417
 418         /* Update direct mapping, invalidate P2M, and add to balloon. */
 419         for (i = 0; i < nr_pages; i++) {
 420                 pfn = frame_list[i];
 421                 frame_list[i] = pfn_to_mfn(pfn);
 422                 page = pfn_to_page(pfn);
 423
 424 #ifdef CONFIG_XEN_HAVE_PVMMU
 425                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 426                         if (!PageHighMem(page)) {
 427                                 ret = HYPERVISOR_update_va_mapping(
 428                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 429                                                 __pte_ma(0), 0);
 430                                 BUG_ON(ret);
 431                         }
 432                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 433                 }
 434 #endif
 435
 436                 balloon_append(page);
 437         }
 438
 439         flush_tlb_all();
 440
 441         set_xen_guest_handle(reservation.extent_start, frame_list);
 442         reservation.nr_extents   = nr_pages;
 443         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 444         BUG_ON(ret != nr_pages);
 445
 446         balloon_stats.current_pages -= nr_pages;
 447
 448         return state;
 449 }
 450
 451 /*
 452  * We avoid multiple worker processes conflicting via the balloon mutex.
 453  * We may of course race updates of the target counts (which are protected
 454  * by the balloon lock), or with changes to the Xen hard limit, but we will
 455  * recover from these in time.
 456  */
 457 static void balloon_process(struct work_struct *work)
 458 {
 459         enum bp_state state = BP_DONE;
 460         long credit;
 461
 462         mutex_lock(&balloon_mutex);
 463
 464         do {
 465                 credit = current_credit();
 466
 467                 if (credit > 0) {
 468                         if (balloon_is_inflated())
 469                                 state = increase_reservation(credit);
 470                         else
 471                                 state = reserve_additional_memory(credit);
 472                 }
 473
 474                 if (credit < 0)
 475                         state = decrease_reservation(-credit, GFP_BALLOON);
 476
 477                 state = update_schedule(state);
 478
 479 #ifndef CONFIG_PREEMPT
 480                 if (need_resched())
 481                         schedule();
 482 #endif
 483         } while (credit && state == BP_DONE);
 484
 485         /* Schedule more work if there is some still to be done. */
 486         if (state == BP_EAGAIN)
 487                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 488
 489         mutex_unlock(&balloon_mutex);
 490 }
 491
 492 /* Resets the Xen limit, sets new target, and kicks off processing. */
 493 void balloon_set_new_target(unsigned long target)
 494 {
 495         /* No need for lock. Not read-modify-write updates. */
 496         balloon_stats.target_pages = target;
 497         schedule_delayed_work(&balloon_worker, 0);
 498 }
 499 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 500
 501 /**
 502  * alloc_xenballooned_pages - get pages that have been ballooned out
 503  * @nr_pages: Number of pages to get
 504  * @pages: pages returned
 505  * @highmem: allow highmem pages
 506  * @return 0 on success, error otherwise
 507  */
 508 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
 509 {
 510         int pgno = 0;
 511         struct page *page;
 512         mutex_lock(&balloon_mutex);
 513         while (pgno < nr_pages) {
 514                 page = balloon_retrieve(highmem);
 515                 if (page && (highmem || !PageHighMem(page))) {
 516                         pages[pgno++] = page;
 517                 } else {
 518                         enum bp_state st;
 519                         if (page)
 520                                 balloon_append(page);
 521                         st = decrease_reservation(nr_pages - pgno,
 522                                         highmem ? GFP_HIGHUSER : GFP_USER);
 523                         if (st != BP_DONE)
 524                                 goto out_undo;
 525                 }
 526         }
 527         mutex_unlock(&balloon_mutex);
 528         return 0;
 529  out_undo:
 530         while (pgno)
 531                 balloon_append(pages[--pgno]);
 532         /* Free the memory back to the kernel soon */
 533         schedule_delayed_work(&balloon_worker, 0);
 534         mutex_unlock(&balloon_mutex);
 535         return -ENOMEM;
 536 }
 537 EXPORT_SYMBOL(alloc_xenballooned_pages);
 538
 539 /**
 540  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 541  * @nr_pages: Number of pages
 542  * @pages: pages to return
 543  */
 544 void free_xenballooned_pages(int nr_pages, struct page **pages)
 545 {
 546         int i;
 547
 548         mutex_lock(&balloon_mutex);
 549
 550         for (i = 0; i < nr_pages; i++) {
 551                 if (pages[i])
 552                         balloon_append(pages[i]);
 553         }
 554
 555         /* The balloon may be too large now. Shrink it if needed. */
 556         if (current_credit())
 557                 schedule_delayed_work(&balloon_worker, 0);
 558
 559         mutex_unlock(&balloon_mutex);
 560 }
 561 EXPORT_SYMBOL(free_xenballooned_pages);
 562
 563 static void __init balloon_add_region(unsigned long start_pfn,
 564                                       unsigned long pages)
 565 {
 566         unsigned long pfn, extra_pfn_end;
 567         struct page *page;
 568
 569         /*
 570          * If the amount of usable memory has been limited (e.g., with
 571          * the 'mem' command line parameter), don't add pages beyond
 572          * this limit.
 573          */
 574         extra_pfn_end = min(max_pfn, start_pfn + pages);
 575
 576         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 577                 page = pfn_to_page(pfn);
 578                 /* totalram_pages and totalhigh_pages do not
 579                    include the boot-time balloon extension, so
 580                    don't subtract from it. */
 581                 __balloon_append(page);
 582         }
 583 }
 584
 585 static int __init balloon_init(void)
 586 {
 587         int i;
 588
 589         if (!xen_domain())
 590                 return -ENODEV;
 591
 592         pr_info("Initialising balloon driver\n");
 593
 594         balloon_stats.current_pages = xen_pv_domain()
 595                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 596                 : get_num_physpages();
 597         balloon_stats.target_pages  = balloon_stats.current_pages;
 598         balloon_stats.balloon_low   = 0;
 599         balloon_stats.balloon_high  = 0;
 600
 601         balloon_stats.schedule_delay = 1;
 602         balloon_stats.max_schedule_delay = 32;
 603         balloon_stats.retry_count = 1;
 604         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 605
 606 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 607         balloon_stats.hotplug_pages = 0;
 608         balloon_stats.balloon_hotplug = 0;
 609
 610         set_online_page_callback(&xen_online_page);
 611         register_memory_notifier(&xen_memory_nb);
 612 #endif
 613
 614         /*
 615          * Initialize the balloon with pages from the extra memory
 616          * regions (see arch/x86/xen/setup.c).
 617          */
 618         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 619                 if (xen_extra_mem[i].size)
 620                         balloon_add_region(PFN_UP(xen_extra_mem[i].start),
 621                                            PFN_DOWN(xen_extra_mem[i].size));
 622
 623         return 0;
 624 }
 625
 626 subsys_initcall(balloon_init);
 627
 628 MODULE_LICENSE("GPL");