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 #include <linux/slab.h>
  58
  59 #include <asm/page.h>
  60 #include <asm/pgalloc.h>
  61 #include <asm/pgtable.h>
  62 #include <asm/tlb.h>
  63
  64 #include <asm/xen/hypervisor.h>
  65 #include <asm/xen/hypercall.h>
  66
  67 #include <xen/xen.h>
  68 #include <xen/interface/xen.h>
  69 #include <xen/interface/memory.h>
  70 #include <xen/balloon.h>
  71 #include <xen/features.h>
  72 #include <xen/page.h>
  73
  74 /*
  75  * balloon_process() state:
  76  *
  77  * BP_DONE: done or nothing to do,
  78  * BP_WAIT: wait to be rescheduled,
  79  * BP_EAGAIN: error, go to sleep,
  80  * BP_ECANCELED: error, balloon operation canceled.
  81  */
  82
  83 enum bp_state {
  84         BP_DONE,
  85         BP_WAIT,
  86         BP_EAGAIN,
  87         BP_ECANCELED
  88 };
  89
  90
  91 static DEFINE_MUTEX(balloon_mutex);
  92
  93 struct balloon_stats balloon_stats;
  94 EXPORT_SYMBOL_GPL(balloon_stats);
  95
  96 /* We increase/decrease in batches which fit in a page */
  97 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(unsigned long)];
  98
  99
 100 /* List of ballooned pages, threaded through the mem_map array. */
 101 static LIST_HEAD(ballooned_pages);
 102
 103 /* Main work function, always executed in process context. */
 104 static void balloon_process(struct work_struct *work);
 105 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
 106
 107 /* When ballooning out (allocating memory to return to Xen) we don't really
 108    want the kernel to try too hard since that can trigger the oom killer. */
 109 #define GFP_BALLOON \
 110         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 111
 112 static void scrub_page(struct page *page)
 113 {
 114 #ifdef CONFIG_XEN_SCRUB_PAGES
 115         clear_highpage(page);
 116 #endif
 117 }
 118
 119 /* balloon_append: add the given page to the balloon. */
 120 static void __balloon_append(struct page *page)
 121 {
 122         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 123         if (PageHighMem(page)) {
 124                 list_add_tail(&page->lru, &ballooned_pages);
 125                 balloon_stats.balloon_high++;
 126         } else {
 127                 list_add(&page->lru, &ballooned_pages);
 128                 balloon_stats.balloon_low++;
 129         }
 130 }
 131
 132 static void balloon_append(struct page *page)
 133 {
 134         __balloon_append(page);
 135         adjust_managed_page_count(page, -1);
 136 }
 137
 138 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 139 static struct page *balloon_retrieve(bool require_lowmem)
 140 {
 141         struct page *page;
 142
 143         if (list_empty(&ballooned_pages))
 144                 return NULL;
 145
 146         page = list_entry(ballooned_pages.next, struct page, lru);
 147         if (require_lowmem && PageHighMem(page))
 148                 return NULL;
 149         list_del(&page->lru);
 150
 151         if (PageHighMem(page))
 152                 balloon_stats.balloon_high--;
 153         else
 154                 balloon_stats.balloon_low--;
 155
 156         adjust_managed_page_count(page, 1);
 157
 158         return page;
 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_WAIT)
 172                 return BP_WAIT;
 173
 174         if (state == BP_ECANCELED)
 175                 return BP_ECANCELED;
 176
 177         if (state == BP_DONE) {
 178                 balloon_stats.schedule_delay = 1;
 179                 balloon_stats.retry_count = 1;
 180                 return BP_DONE;
 181         }
 182
 183         ++balloon_stats.retry_count;
 184
 185         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 186                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 187                 balloon_stats.schedule_delay = 1;
 188                 balloon_stats.retry_count = 1;
 189                 return BP_ECANCELED;
 190         }
 191
 192         balloon_stats.schedule_delay <<= 1;
 193
 194         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 195                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 196
 197         return BP_EAGAIN;
 198 }
 199
 200 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 201 static struct resource *additional_memory_resource(phys_addr_t size)
 202 {
 203         struct resource *res;
 204         int ret;
 205
 206         res = kzalloc(sizeof(*res), GFP_KERNEL);
 207         if (!res)
 208                 return NULL;
 209
 210         res->name = "System RAM";
 211         res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 212
 213         ret = allocate_resource(&iomem_resource, res,
 214                                 size, 0, -1,
 215                                 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);
 216         if (ret < 0) {
 217                 pr_err("Cannot allocate new System RAM resource\n");
 218                 kfree(res);
 219                 return NULL;
 220         }
 221
 222         return res;
 223 }
 224
 225 static void release_memory_resource(struct resource *resource)
 226 {
 227         if (!resource)
 228                 return;
 229
 230         /*
 231          * No need to reset region to identity mapped since we now
 232          * know that no I/O can be in this region
 233          */
 234         release_resource(resource);
 235         kfree(resource);
 236 }
 237
 238 static enum bp_state reserve_additional_memory(void)
 239 {
 240         long credit;
 241         struct resource *resource;
 242         int nid, rc;
 243         unsigned long balloon_hotplug;
 244
 245         credit = balloon_stats.target_pages - balloon_stats.total_pages;
 246
 247         /*
 248          * Already hotplugged enough pages?  Wait for them to be
 249          * onlined.
 250          */
 251         if (credit <= 0)
 252                 return BP_WAIT;
 253
 254         balloon_hotplug = round_up(credit, PAGES_PER_SECTION);
 255
 256         resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);
 257         if (!resource)
 258                 goto err;
 259
 260         nid = memory_add_physaddr_to_nid(resource->start);
 261
 262 #ifdef CONFIG_XEN_HAVE_PVMMU
 263         /*
 264          * add_memory() will build page tables for the new memory so
 265          * the p2m must contain invalid entries so the correct
 266          * non-present PTEs will be written.
 267          *
 268          * If a failure occurs, the original (identity) p2m entries
 269          * are not restored since this region is now known not to
 270          * conflict with any devices.
 271          */
 272         if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 273                 unsigned long pfn, i;
 274
 275                 pfn = PFN_DOWN(resource->start);
 276                 for (i = 0; i < balloon_hotplug; i++) {
 277                         if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {
 278                                 pr_warn("set_phys_to_machine() failed, no memory added\n");
 279                                 goto err;
 280                         }
 281                 }
 282         }
 283 #endif
 284
 285         rc = add_memory_resource(nid, resource);
 286         if (rc) {
 287                 pr_warn("Cannot add additional memory (%i)\n", rc);
 288                 goto err;
 289         }
 290
 291         balloon_stats.total_pages += balloon_hotplug;
 292
 293         return BP_WAIT;
 294   err:
 295         release_memory_resource(resource);
 296         return BP_ECANCELED;
 297 }
 298
 299 static void xen_online_page(struct page *page)
 300 {
 301         __online_page_set_limits(page);
 302
 303         mutex_lock(&balloon_mutex);
 304
 305         __balloon_append(page);
 306
 307         mutex_unlock(&balloon_mutex);
 308 }
 309
 310 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)
 311 {
 312         if (val == MEM_ONLINE)
 313                 schedule_delayed_work(&balloon_worker, 0);
 314
 315         return NOTIFY_OK;
 316 }
 317
 318 static struct notifier_block xen_memory_nb = {
 319         .notifier_call = xen_memory_notifier,
 320         .priority = 0
 321 };
 322 #else
 323 static enum bp_state reserve_additional_memory(void)
 324 {
 325         balloon_stats.target_pages = balloon_stats.current_pages;
 326         return BP_DONE;
 327 }
 328 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */
 329
 330 static long current_credit(void)
 331 {
 332         return balloon_stats.target_pages - balloon_stats.current_pages;
 333 }
 334
 335 static bool balloon_is_inflated(void)
 336 {
 337         return balloon_stats.balloon_low || balloon_stats.balloon_high;
 338 }
 339
 340 static enum bp_state increase_reservation(unsigned long nr_pages)
 341 {
 342         int rc;
 343         unsigned long  pfn, i;
 344         struct page   *page;
 345         struct xen_memory_reservation reservation = {
 346                 .address_bits = 0,
 347                 .extent_order = 0,
 348                 .domid        = DOMID_SELF
 349         };
 350
 351         if (nr_pages > ARRAY_SIZE(frame_list))
 352                 nr_pages = ARRAY_SIZE(frame_list);
 353
 354         page = list_first_entry_or_null(&ballooned_pages, struct page, lru);
 355         for (i = 0; i < nr_pages; i++) {
 356                 if (!page) {
 357                         nr_pages = i;
 358                         break;
 359                 }
 360                 frame_list[i] = page_to_pfn(page);
 361                 page = balloon_next_page(page);
 362         }
 363
 364         set_xen_guest_handle(reservation.extent_start, frame_list);
 365         reservation.nr_extents = nr_pages;
 366         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 367         if (rc <= 0)
 368                 return BP_EAGAIN;
 369
 370         for (i = 0; i < rc; i++) {
 371                 page = balloon_retrieve(false);
 372                 BUG_ON(page == NULL);
 373
 374                 pfn = page_to_pfn(page);
 375
 376 #ifdef CONFIG_XEN_HAVE_PVMMU
 377                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 378                         set_phys_to_machine(pfn, frame_list[i]);
 379
 380                         /* Link back into the page tables if not highmem. */
 381                         if (!PageHighMem(page)) {
 382                                 int ret;
 383                                 ret = HYPERVISOR_update_va_mapping(
 384                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 385                                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 386                                                 0);
 387                                 BUG_ON(ret);
 388                         }
 389                 }
 390 #endif
 391
 392                 /* Relinquish the page back to the allocator. */
 393                 __free_reserved_page(page);
 394         }
 395
 396         balloon_stats.current_pages += rc;
 397
 398         return BP_DONE;
 399 }
 400
 401 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 402 {
 403         enum bp_state state = BP_DONE;
 404         unsigned long  pfn, i;
 405         struct page   *page;
 406         int ret;
 407         struct xen_memory_reservation reservation = {
 408                 .address_bits = 0,
 409                 .extent_order = 0,
 410                 .domid        = DOMID_SELF
 411         };
 412
 413         if (nr_pages > ARRAY_SIZE(frame_list))
 414                 nr_pages = ARRAY_SIZE(frame_list);
 415
 416         for (i = 0; i < nr_pages; i++) {
 417                 page = alloc_page(gfp);
 418                 if (page == NULL) {
 419                         nr_pages = i;
 420                         state = BP_EAGAIN;
 421                         break;
 422                 }
 423                 scrub_page(page);
 424
 425                 frame_list[i] = page_to_pfn(page);
 426         }
 427
 428         /*
 429          * Ensure that ballooned highmem pages don't have kmaps.
 430          *
 431          * Do this before changing the p2m as kmap_flush_unused()
 432          * reads PTEs to obtain pages (and hence needs the original
 433          * p2m entry).
 434          */
 435         kmap_flush_unused();
 436
 437         /* Update direct mapping, invalidate P2M, and add to balloon. */
 438         for (i = 0; i < nr_pages; i++) {
 439                 pfn = frame_list[i];
 440                 frame_list[i] = pfn_to_gfn(pfn);
 441                 page = pfn_to_page(pfn);
 442
 443 #ifdef CONFIG_XEN_HAVE_PVMMU
 444                 if (!xen_feature(XENFEAT_auto_translated_physmap)) {
 445                         if (!PageHighMem(page)) {
 446                                 ret = HYPERVISOR_update_va_mapping(
 447                                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 448                                                 __pte_ma(0), 0);
 449                                 BUG_ON(ret);
 450                         }
 451                         __set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 452                 }
 453 #endif
 454
 455                 balloon_append(page);
 456         }
 457
 458         flush_tlb_all();
 459
 460         set_xen_guest_handle(reservation.extent_start, frame_list);
 461         reservation.nr_extents   = nr_pages;
 462         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 463         BUG_ON(ret != nr_pages);
 464
 465         balloon_stats.current_pages -= nr_pages;
 466
 467         return state;
 468 }
 469
 470 /*
 471  * As this is a work item it is guaranteed to run as a single instance only.
 472  * We may of course race updates of the target counts (which are protected
 473  * by the balloon lock), or with changes to the Xen hard limit, but we will
 474  * recover from these in time.
 475  */
 476 static void balloon_process(struct work_struct *work)
 477 {
 478         enum bp_state state = BP_DONE;
 479         long credit;
 480
 481
 482         do {
 483                 mutex_lock(&balloon_mutex);
 484
 485                 credit = current_credit();
 486
 487                 if (credit > 0) {
 488                         if (balloon_is_inflated())
 489                                 state = increase_reservation(credit);
 490                         else
 491                                 state = reserve_additional_memory();
 492                 }
 493
 494                 if (credit < 0)
 495                         state = decrease_reservation(-credit, GFP_BALLOON);
 496
 497                 state = update_schedule(state);
 498
 499                 mutex_unlock(&balloon_mutex);
 500
 501                 cond_resched();
 502
 503         } while (credit && state == BP_DONE);
 504
 505         /* Schedule more work if there is some still to be done. */
 506         if (state == BP_EAGAIN)
 507                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 508 }
 509
 510 /* Resets the Xen limit, sets new target, and kicks off processing. */
 511 void balloon_set_new_target(unsigned long target)
 512 {
 513         /* No need for lock. Not read-modify-write updates. */
 514         balloon_stats.target_pages = target;
 515         schedule_delayed_work(&balloon_worker, 0);
 516 }
 517 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 518
 519 /**
 520  * alloc_xenballooned_pages - get pages that have been ballooned out
 521  * @nr_pages: Number of pages to get
 522  * @pages: pages returned
 523  * @return 0 on success, error otherwise
 524  */
 525 int alloc_xenballooned_pages(int nr_pages, struct page **pages)
 526 {
 527         int pgno = 0;
 528         struct page *page;
 529         mutex_lock(&balloon_mutex);
 530         while (pgno < nr_pages) {
 531                 page = balloon_retrieve(true);
 532                 if (page) {
 533                         pages[pgno++] = page;
 534                 } else {
 535                         enum bp_state st;
 536                         st = decrease_reservation(nr_pages - pgno, GFP_USER);
 537                         if (st != BP_DONE)
 538                                 goto out_undo;
 539                 }
 540         }
 541         mutex_unlock(&balloon_mutex);
 542         return 0;
 543  out_undo:
 544         while (pgno)
 545                 balloon_append(pages[--pgno]);
 546         /* Free the memory back to the kernel soon */
 547         schedule_delayed_work(&balloon_worker, 0);
 548         mutex_unlock(&balloon_mutex);
 549         return -ENOMEM;
 550 }
 551 EXPORT_SYMBOL(alloc_xenballooned_pages);
 552
 553 /**
 554  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 555  * @nr_pages: Number of pages
 556  * @pages: pages to return
 557  */
 558 void free_xenballooned_pages(int nr_pages, struct page **pages)
 559 {
 560         int i;
 561
 562         mutex_lock(&balloon_mutex);
 563
 564         for (i = 0; i < nr_pages; i++) {
 565                 if (pages[i])
 566                         balloon_append(pages[i]);
 567         }
 568
 569         /* The balloon may be too large now. Shrink it if needed. */
 570         if (current_credit())
 571                 schedule_delayed_work(&balloon_worker, 0);
 572
 573         mutex_unlock(&balloon_mutex);
 574 }
 575 EXPORT_SYMBOL(free_xenballooned_pages);
 576
 577 static void __init balloon_add_region(unsigned long start_pfn,
 578                                       unsigned long pages)
 579 {
 580         unsigned long pfn, extra_pfn_end;
 581         struct page *page;
 582
 583         /*
 584          * If the amount of usable memory has been limited (e.g., with
 585          * the 'mem' command line parameter), don't add pages beyond
 586          * this limit.
 587          */
 588         extra_pfn_end = min(max_pfn, start_pfn + pages);
 589
 590         for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) {
 591                 page = pfn_to_page(pfn);
 592                 /* totalram_pages and totalhigh_pages do not
 593                    include the boot-time balloon extension, so
 594                    don't subtract from it. */
 595                 __balloon_append(page);
 596         }
 597
 598         balloon_stats.total_pages += extra_pfn_end - start_pfn;
 599 }
 600
 601 static int __init balloon_init(void)
 602 {
 603         int i;
 604
 605         if (!xen_domain())
 606                 return -ENODEV;
 607
 608         pr_info("Initialising balloon driver\n");
 609
 610         balloon_stats.current_pages = xen_pv_domain()
 611                 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)
 612                 : get_num_physpages();
 613         balloon_stats.target_pages  = balloon_stats.current_pages;
 614         balloon_stats.balloon_low   = 0;
 615         balloon_stats.balloon_high  = 0;
 616         balloon_stats.total_pages   = balloon_stats.current_pages;
 617
 618         balloon_stats.schedule_delay = 1;
 619         balloon_stats.max_schedule_delay = 32;
 620         balloon_stats.retry_count = 1;
 621         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 622
 623 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG
 624         set_online_page_callback(&xen_online_page);
 625         register_memory_notifier(&xen_memory_nb);
 626 #endif
 627
 628         /*
 629          * Initialize the balloon with pages from the extra memory
 630          * regions (see arch/x86/xen/setup.c).
 631          */
 632         for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++)
 633                 if (xen_extra_mem[i].n_pfns)
 634                         balloon_add_region(xen_extra_mem[i].start_pfn,
 635                                            xen_extra_mem[i].n_pfns);
 636
 637         return 0;
 638 }
 639
 640 subsys_initcall(balloon_init);
 641
 642 MODULE_LICENSE("GPL");