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