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 #ifdef CONFIG_XEN_HAVE_PVMMU
 233         /*
 234          * add_memory() will build page tables for the new memory so
 235          * the p2m must contain invalid entries so the correct
 236          * non-present PTEs will be written.
 237          *
 238          * If a failure occurs, the original (identity) p2m entries
 239          * are not restored since this region is now known not to
 240          * conflict with any devices.
 241          */
 242         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 243                 unsigned long pfn, i;
 244
 245                 pfn = PFN_DOWN(hotplug_start_paddr);
 246                 for (i = 0; i < balloon_hotplug; i++) {
 247                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
 248                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
 249                                 return BP_ECANCELED;
 250                         }
 251                 }
 252         }
 253 #endif
 254
 255         rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT);
 256
 257         if (rc) {
 258                 pr_warn("Cannot add additional memory (%i)\n", rc);
 259                 return BP_ECANCELED;
 260         }
 261
 262         balloon_hotplug -= credit;
 263
 264         balloon_stats.hotplug_pages += credit;
 265         balloon_stats.balloon_hotplug = balloon_hotplug;
 266
 267         return BP_DONE;
 268 }
 269
 270 static void xen_online_page(struct page *page)
 271 {
 272         __online_page_set_limits(page);
 273
 274         mutex_lock(&balloon_mutex);
 275
 276         __balloon_append(page);
 277
 278         if (balloon_stats.hotplug_pages)
 279                 --balloon_stats.hotplug_pages;
 280         else
 281                 --balloon_stats.balloon_hotplug;
 282
 283         mutex_unlock(&balloon_mutex);
 284 }
 285
 286 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 287 {
 288         if (val == MEM_ONLINE)
 289                 schedule_delayed_work(&balloon_worker, 0);
 290
 291         return NOTIFY_OK;
 292 }
 293
 294 static struct notifier_block xen_memory_nb = {
 295         .notifier_call = xen_memory_notifier,
 296         .priority = 0
 297 };
 298 #else
 299 static long current_credit(void)
 300 {
 301         unsigned long target = balloon_stats.target_pages;
 302
 303         target = min(target,
 304                      balloon_stats.current_pages +
 305                      balloon_stats.balloon_low +
 306                      balloon_stats.balloon_high);
 307
 308         return target - balloon_stats.current_pages;
 309 }
 310
 311 static bool balloon_is_inflated(void)
 312 {
 313         if (balloon_stats.balloon_low || balloon_stats.balloon_high)
 314                 return true;
 315         else
 316                 return false;
 317 }
 318
 319 static enum bp_state reserve_additional_memory(long credit)
 320 {
 321         balloon_stats.target_pages = balloon_stats.current_pages;
 322         return BP_DONE;
 323 }
 324 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 325
 326 static enum bp_state increase_reservation(unsigned long nr_pages)
 327 {
 328         int rc;
 329         unsigned long  pfn, i;
 330         struct page   *page;
 331         struct xen_memory_reservation reservation = {
 332                 .address_bits = 0,
 333                 .extent_order = 0,
 334                 .domid        = DOMID_SELF
 335         };
 336
 337 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 338         if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) {
 339                 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug);
 340                 balloon_stats.hotplug_pages += nr_pages;
 341                 balloon_stats.balloon_hotplug -= nr_pages;
 342                 return BP_DONE;
 343         }
 344 #endif
 345
 346         if (nr_pages > ARRAY_SIZE(frame_list))
 347                 nr_pages = ARRAY_SIZE(frame_list);
 348
 349         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 350         for (i = 0; i < nr_pages; i++) {
 351                 if (!page) {
 352                         nr_pages = i;
 353                         break;
 354                 }
 355                 frame_list[i] = page_to_pfn(page);
 356                 page = balloon_next_page(page);
 357         }
 358
 359         set_xen_guest_handle(reservation.extent_start, frame_list);
 360         reservation.nr_extents = nr_pages;
 361         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 362         if (rc <= 0)
 363                 return BP_EAGAIN;
 364
 365         for (i = 0; i < rc; i++) {
 366                 page = balloon_retrieve(false);
 367                 BUG_ON(page == NULL);
 368
 369                 pfn = page_to_pfn(page);
 370
 371 #ifdef CONFIG_XEN_HAVE_PVMMU
 372                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 373                         set_phys_to_machine(pfn, frame_list[i]);
 374
 375                         /* Link back into the page tables if not highmem. */
 376                         if (!PageHighMem(page)) {
 377                                 int ret;
 378                                 ret = HYPERVISOR_update_va_mapping(
 379                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 380                                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 381                                                 0);
 382                                 BUG_ON(ret);
 383                         }
 384                 }
 385 #endif
 386
 387                 /* Relinquish the page back to the allocator. */
 388                 __free_reserved_page(page);
 389         }
 390
 391         balloon_stats.current_pages += rc;
 392
 393         return BP_DONE;
 394 }
 395
 396 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 397 {
 398         enum bp_state state = BP_DONE;
 399         unsigned long  pfn, i;
 400         struct page   *page;
 401         int ret;
 402         struct xen_memory_reservation reservation = {
 403                 .address_bits = 0,
 404                 .extent_order = 0,
 405                 .domid        = DOMID_SELF
 406         };
 407
 408 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 409         if (balloon_stats.hotplug_pages) {
 410                 nr_pages = min(nr_pages, balloon_stats.hotplug_pages);
 411                 balloon_stats.hotplug_pages -= nr_pages;
 412                 balloon_stats.balloon_hotplug += nr_pages;
 413                 return BP_DONE;
 414         }
 415 #endif
 416
 417         if (nr_pages > ARRAY_SIZE(frame_list))
 418                 nr_pages = ARRAY_SIZE(frame_list);
 419
 420         for (i = 0; i < nr_pages; i++) {
 421                 page = alloc_page(gfp);
 422                 if (page == NULL) {
 423                         nr_pages = i;
 424                         state = BP_EAGAIN;
 425                         break;
 426                 }
 427                 scrub_page(page);
 428
 429                 frame_list[i] = page_to_pfn(page);
 430         }
 431
 432         /*
 433          * Ensure that ballooned highmem pages don't have kmaps.
 434          *
 435          * Do this before changing the p2m as kmap_flush_unused()
 436          * reads PTEs to obtain pages (and hence needs the original
 437          * p2m entry).
 438          */
 439         kmap_flush_unused();
 440
 441         /* Update direct mapping, invalidate P2M, and add to balloon. */
 442         for (i = 0; i < nr_pages; i++) {
 443                 pfn = frame_list[i];
 444                 frame_list[i] = pfn_to_mfn(pfn);
 445                 page = pfn_to_page(pfn);
 446
 447 #ifdef CONFIG_XEN_HAVE_PVMMU
 448                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 449                         if (!PageHighMem(page)) {
 450                                 ret = HYPERVISOR_update_va_mapping(
 451                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 452                                                 __pte_ma(0), 0);
 453                                 BUG_ON(ret);
 454                         }
 455                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 456                 }
 457 #endif
 458
 459                 balloon_append(page);
 460         }
 461
 462         flush_tlb_all();
 463
 464         set_xen_guest_handle(reservation.extent_start, frame_list);
 465         reservation.nr_extents   = nr_pages;
 466         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 467         BUG_ON(ret != nr_pages);
 468
 469         balloon_stats.current_pages -= nr_pages;
 470
 471         return state;
 472 }
 473
 474 /*
 475  * As this is a work item it is guaranteed to run as a single instance only.
 476  * We may of course race updates of the target counts (which are protected
 477  * by the balloon lock), or with changes to the Xen hard limit, but we will
 478  * recover from these in time.
 479  */
 480 static void balloon_process(struct work_struct *work)
 481 {
 482         enum bp_state state = BP_DONE;
 483         long credit;
 484
 485
 486         do {
 487                 mutex_lock(&balloon_mutex);
 488
 489                 credit = current_credit();
 490
 491                 if (credit > 0) {
 492                         if (balloon_is_inflated())
 493                                 state = increase_reservation(credit);
 494                         else
 495                                 state = reserve_additional_memory(credit);
 496                 }
 497
 498                 if (credit < 0)
 499                         state = decrease_reservation(-credit, GFP_BALLOON);
 500
 501                 state = update_schedule(state);
 502
 503                 mutex_unlock(&balloon_mutex);
 504
 505                 cond_resched();
 506
 507         } while (credit && state == BP_DONE);
 508
 509         /* Schedule more work if there is some still to be done. */
 510         if (state == BP_EAGAIN)
 511                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 512 }
 513
 514 /* Resets the Xen limit, sets new target, and kicks off processing. */
 515 void balloon_set_new_target(unsigned long target)
 516 {
 517         /* No need for lock. Not read-modify-write updates. */
 518         balloon_stats.target_pages = target;
 519         schedule_delayed_work(&balloon_worker, 0);
 520 }
 521 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 522
 523 /**
 524  * alloc_xenballooned_pages - get pages that have been ballooned out
 525  * @nr_pages: Number of pages to get
 526  * @pages: pages returned
 527  * @highmem: allow highmem pages
 528  * @return 0 on success, error otherwise
 529  */
 530 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem)
 531 {
 532         int pgno = 0;
 533         struct page *page;
 534         mutex_lock(&balloon_mutex);
 535         while (pgno < nr_pages) {
 536                 page = balloon_retrieve(highmem);
 537                 if (page && (highmem || !PageHighMem(page))) {
 538                         pages[pgno++] = page;
 539                 } else {
 540                         enum bp_state st;
 541                         if (page)
 542                                 balloon_append(page);
 543                         st = decrease_reservation(nr_pages - pgno,
 544                                         highmem ? GFP_HIGHUSER : GFP_USER);
 545                         if (st != BP_DONE)
 546                                 goto out_undo;
 547                 }
 548         }
 549         mutex_unlock(&balloon_mutex);
 550         return 0;
 551  out_undo:
 552         while (pgno)
 553                 balloon_append(pages[--pgno]);
 554         /* Free the memory back to the kernel soon */
 555         schedule_delayed_work(&balloon_worker, 0);
 556         mutex_unlock(&balloon_mutex);
 557         return -ENOMEM;
 558 }
 559 EXPORT_SYMBOL(alloc_xenballooned_pages);
 560
 561 /**
 562  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 563  * @nr_pages: Number of pages
 564  * @pages: pages to return
 565  */
 566 void free_xenballooned_pages(int nr_pages, struct page **pages)
 567 {
 568         int i;
 569
 570         mutex_lock(&balloon_mutex);
 571
 572         for (i = 0; i < nr_pages; i++) {
 573                 if (pages[i])
 574                         balloon_append(pages[i]);
 575         }
 576
 577         /* The balloon may be too large now. Shrink it if needed. */
 578         if (current_credit())
 579                 schedule_delayed_work(&balloon_worker, 0);
 580
 581         mutex_unlock(&balloon_mutex);
 582 }
 583 EXPORT_SYMBOL(free_xenballooned_pages);
 584
 585 static void __init balloon_add_region(unsigned long start_pfn,
 586                                       unsigned long pages)
 587 {
 588         unsigned long pfn, extra_pfn_end;
 589         struct page *page;
 590
 591         /*
 592          * If the amount of usable memory has been limited (e.g., with
 593          * the 'mem' command line parameter), don't add pages beyond
 594          * this limit.
 595          */
 596         extra_pfn_end = min(max_pfn, start_pfn + pages);
 597
 598         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 599                 page = pfn_to_page(pfn);
 600                 /* totalram_pages and totalhigh_pages do not
 601                    include the boot-time balloon extension, so
 602                    don't subtract from it. */
 603                 __balloon_append(page);
 604         }
 605 }
 606
 607 static int __init balloon_init(void)
 608 {
 609         int i;
 610
 611         if (!xen_domain())
 612                 return -ENODEV;
 613
 614         pr_info("Initialising balloon driver\n");
 615
 616         balloon_stats.current_pages = xen_pv_domain()
 617                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 618                 : get_num_physpages();
 619         balloon_stats.target_pages  = balloon_stats.current_pages;
 620         balloon_stats.balloon_low   = 0;
 621         balloon_stats.balloon_high  = 0;
 622
 623         balloon_stats.schedule_delay = 1;
 624         balloon_stats.max_schedule_delay = 32;
 625         balloon_stats.retry_count = 1;
 626         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 627
 628 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 629         balloon_stats.hotplug_pages = 0;
 630         balloon_stats.balloon_hotplug = 0;
 631
 632         set_online_page_callback(&xen_online_page);
 633         register_memory_notifier(&xen_memory_nb);
 634 #endif
 635
 636         /*
 637          * Initialize the balloon with pages from the extra memory
 638          * regions (see arch/x86/xen/setup.c).
 639          */
 640         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 641                 if (xen_extra_mem[i].n_pfns)
 642                         balloon_add_region(xen_extra_mem[i].start_pfn,
 643                                            xen_extra_mem[i].n_pfns);
 644
 645         return 0;
 646 }
 647
 648 subsys_initcall(balloon_init);
 649
 650 MODULE_LICENSE("GPL");