drivers/xen/balloon.c

   1 /******************************************************************************
   2  * balloon.c
   3  *
   4  * Xen balloon driver - enables returning/claiming memory to/from Xen.
   5  *
   6  * Copyright (c) 2003, B Dragovic
   7  * Copyright (c) 2003-2004, M Williamson, K Fraser
   8  * Copyright (c) 2005 Dan M. Smith, IBM Corporation
   9  *
  10  * This program is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation; or, when distributed
  13  * separately from the Linux kernel or incorporated into other
  14  * software packages, subject to the following license:
  15  *
  16  * Permission is hereby granted, free of charge, to any person obtaining a copy
  17  * of this source file (the "Software"), to deal in the Software without
  18  * restriction, including without limitation the rights to use, copy, modify,
  19  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  20  * and to permit persons to whom the Software is furnished to do so, subject to
  21  * the following conditions:
  22  *
  23  * The above copyright notice and this permission notice shall be included in
  24  * all copies or substantial portions of the Software.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  27  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  28  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  29  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  30  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  31  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  32  * IN THE SOFTWARE.
  33  */
  34
  35 #include <linux/kernel.h>
  36 #include <linux/module.h>
  37 #include <linux/sched.h>
  38 #include <linux/errno.h>
  39 #include <linux/mm.h>
  40 #include <linux/bootmem.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/highmem.h>
  43 #include <linux/mutex.h>
  44 #include <linux/list.h>
  45 #include <linux/sysdev.h>
  46
  47 #include <asm/page.h>
  48 #include <asm/pgalloc.h>
  49 #include <asm/pgtable.h>
  50 #include <asm/uaccess.h>
  51 #include <asm/tlb.h>
  52
  53 #include <asm/xen/hypervisor.h>
  54 #include <asm/xen/hypercall.h>
  55
  56 #include <xen/xen.h>
  57 #include <xen/interface/xen.h>
  58 #include <xen/interface/memory.h>
  59 #include <xen/xenbus.h>
  60 #include <xen/features.h>
  61 #include <xen/page.h>
  62
  63 #define PAGES2KB(_p) ((_p)<<(PAGE_SHIFT-10))
  64
  65 #define BALLOON_CLASS_NAME "xen_memory"
  66
  67 struct balloon_stats {
  68         /* We aim for 'current allocation' == 'target allocation'. */
  69         unsigned long current_pages;
  70         unsigned long target_pages;
  71         /* We may hit the hard limit in Xen. If we do then we remember it. */
  72         unsigned long hard_limit;
  73         /*
  74          * Drivers may alter the memory reservation independently, but they
  75          * must inform the balloon driver so we avoid hitting the hard limit.
  76          */
  77         unsigned long driver_pages;
  78         /* Number of pages in high- and low-memory balloons. */
  79         unsigned long balloon_low;
  80         unsigned long balloon_high;
  81 };
  82
  83 static DEFINE_MUTEX(balloon_mutex);
  84
  85 static struct sys_device balloon_sysdev;
  86
  87 static int register_balloon(struct sys_device *sysdev);
  88
  89 /*
  90  * Protects atomic reservation decrease/increase against concurrent increases.
  91  * Also protects non-atomic updates of current_pages and driver_pages, and
  92  * balloon lists.
  93  */
  94 static DEFINE_SPINLOCK(balloon_lock);
  95
  96 static struct balloon_stats balloon_stats;
  97
  98 /* We increase/decrease in batches which fit in a page */
  99 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
 100
 101 #ifdef CONFIG_HIGHMEM
 102 #define inc_totalhigh_pages() (totalhigh_pages++)
 103 #define dec_totalhigh_pages() (totalhigh_pages--)
 104 #else
 105 #define inc_totalhigh_pages() do {} while(0)
 106 #define dec_totalhigh_pages() do {} while(0)
 107 #endif
 108
 109 /* List of ballooned pages, threaded through the mem_map array. */
 110 static LIST_HEAD(ballooned_pages);
 111
 112 /* Main work function, always executed in process context. */
 113 static void balloon_process(struct work_struct *work);
 114 static DECLARE_WORK(balloon_worker, balloon_process);
 115 static struct timer_list balloon_timer;
 116
 117 /* When ballooning out (allocating memory to return to Xen) we don't really
 118    want the kernel to try too hard since that can trigger the oom killer. */
 119 #define GFP_BALLOON \
 120         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 121
 122 static void scrub_page(struct page *page)
 123 {
 124 #ifdef CONFIG_XEN_SCRUB_PAGES
 125         clear_highpage(page);
 126 #endif
 127 }
 128
 129 /* balloon_append: add the given page to the balloon. */
 130 static void balloon_append(struct page *page)
 131 {
 132         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 133         if (PageHighMem(page)) {
 134                 list_add_tail(&page->lru, &ballooned_pages);
 135                 balloon_stats.balloon_high++;
 136                 dec_totalhigh_pages();
 137         } else {
 138                 list_add(&page->lru, &ballooned_pages);
 139                 balloon_stats.balloon_low++;
 140         }
 141 }
 142
 143 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 144 static struct page *balloon_retrieve(void)
 145 {
 146         struct page *page;
 147
 148         if (list_empty(&ballooned_pages))
 149                 return NULL;
 150
 151         page = list_entry(ballooned_pages.next, struct page, lru);
 152         list_del(&page->lru);
 153
 154         if (PageHighMem(page)) {
 155                 balloon_stats.balloon_high--;
 156                 inc_totalhigh_pages();
 157         }
 158         else
 159                 balloon_stats.balloon_low--;
 160
 161         return page;
 162 }
 163
 164 static struct page *balloon_first_page(void)
 165 {
 166         if (list_empty(&ballooned_pages))
 167                 return NULL;
 168         return list_entry(ballooned_pages.next, struct page, lru);
 169 }
 170
 171 static struct page *balloon_next_page(struct page *page)
 172 {
 173         struct list_head *next = page->lru.next;
 174         if (next == &ballooned_pages)
 175                 return NULL;
 176         return list_entry(next, struct page, lru);
 177 }
 178
 179 static void balloon_alarm(unsigned long unused)
 180 {
 181         schedule_work(&balloon_worker);
 182 }
 183
 184 static unsigned long current_target(void)
 185 {
 186         unsigned long target = min(balloon_stats.target_pages, balloon_stats.hard_limit);
 187
 188         target = min(target,
 189                      balloon_stats.current_pages +
 190                      balloon_stats.balloon_low +
 191                      balloon_stats.balloon_high);
 192
 193         return target;
 194 }
 195
 196 static int increase_reservation(unsigned long nr_pages)
 197 {
 198         unsigned long  pfn, i, flags;
 199         struct page   *page;
 200         long           rc;
 201         struct xen_memory_reservation reservation = {
 202                 .address_bits = 0,
 203                 .extent_order = 0,
 204                 .domid        = DOMID_SELF
 205         };
 206
 207         if (nr_pages > ARRAY_SIZE(frame_list))
 208                 nr_pages = ARRAY_SIZE(frame_list);
 209
 210         spin_lock_irqsave(&balloon_lock, flags);
 211
 212         page = balloon_first_page();
 213         for (i = 0; i < nr_pages; i++) {
 214                 BUG_ON(page == NULL);
 215                 frame_list[i] = page_to_pfn(page);
 216                 page = balloon_next_page(page);
 217         }
 218
 219         set_xen_guest_handle(reservation.extent_start, frame_list);
 220         reservation.nr_extents = nr_pages;
 221         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 222         if (rc < nr_pages) {
 223                 if (rc > 0) {
 224                         int ret;
 225
 226                         /* We hit the Xen hard limit: reprobe. */
 227                         reservation.nr_extents = rc;
 228                         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation,
 229                                                    &reservation);
 230                         BUG_ON(ret != rc);
 231                 }
 232                 if (rc >= 0)
 233                         balloon_stats.hard_limit = (balloon_stats.current_pages + rc -
 234                                                     balloon_stats.driver_pages);
 235                 goto out;
 236         }
 237
 238         for (i = 0; i < nr_pages; i++) {
 239                 page = balloon_retrieve();
 240                 BUG_ON(page == NULL);
 241
 242                 pfn = page_to_pfn(page);
 243                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
 244                        phys_to_machine_mapping_valid(pfn));
 245
 246                 set_phys_to_machine(pfn, frame_list[i]);
 247
 248                 /* Link back into the page tables if not highmem. */
 249                 if (pfn < max_low_pfn) {
 250                         int ret;
 251                         ret = HYPERVISOR_update_va_mapping(
 252                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 253                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 254                                 0);
 255                         BUG_ON(ret);
 256                 }
 257
 258                 /* Relinquish the page back to the allocator. */
 259                 ClearPageReserved(page);
 260                 init_page_count(page);
 261                 __free_page(page);
 262         }
 263
 264         balloon_stats.current_pages += nr_pages;
 265         totalram_pages = balloon_stats.current_pages;
 266
 267  out:
 268         spin_unlock_irqrestore(&balloon_lock, flags);
 269
 270         return 0;
 271 }
 272
 273 static int decrease_reservation(unsigned long nr_pages)
 274 {
 275         unsigned long  pfn, i, flags;
 276         struct page   *page;
 277         int            need_sleep = 0;
 278         int ret;
 279         struct xen_memory_reservation reservation = {
 280                 .address_bits = 0,
 281                 .extent_order = 0,
 282                 .domid        = DOMID_SELF
 283         };
 284
 285         if (nr_pages > ARRAY_SIZE(frame_list))
 286                 nr_pages = ARRAY_SIZE(frame_list);
 287
 288         for (i = 0; i < nr_pages; i++) {
 289                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
 290                         nr_pages = i;
 291                         need_sleep = 1;
 292                         break;
 293                 }
 294
 295                 pfn = page_to_pfn(page);
 296                 frame_list[i] = pfn_to_mfn(pfn);
 297
 298                 scrub_page(page);
 299
 300                 if (!PageHighMem(page)) {
 301                         ret = HYPERVISOR_update_va_mapping(
 302                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 303                                 __pte_ma(0), 0);
 304                         BUG_ON(ret);
 305                 }
 306
 307         }
 308
 309         /* Ensure that ballooned highmem pages don't have kmaps. */
 310         kmap_flush_unused();
 311         flush_tlb_all();
 312
 313         spin_lock_irqsave(&balloon_lock, flags);
 314
 315         /* No more mappings: invalidate P2M and add to balloon. */
 316         for (i = 0; i < nr_pages; i++) {
 317                 pfn = mfn_to_pfn(frame_list[i]);
 318                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 319                 balloon_append(pfn_to_page(pfn));
 320         }
 321
 322         set_xen_guest_handle(reservation.extent_start, frame_list);
 323         reservation.nr_extents   = nr_pages;
 324         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 325         BUG_ON(ret != nr_pages);
 326
 327         balloon_stats.current_pages -= nr_pages;
 328         totalram_pages = balloon_stats.current_pages;
 329
 330         spin_unlock_irqrestore(&balloon_lock, flags);
 331
 332         return need_sleep;
 333 }
 334
 335 /*
 336  * We avoid multiple worker processes conflicting via the balloon mutex.
 337  * We may of course race updates of the target counts (which are protected
 338  * by the balloon lock), or with changes to the Xen hard limit, but we will
 339  * recover from these in time.
 340  */
 341 static void balloon_process(struct work_struct *work)
 342 {
 343         int need_sleep = 0;
 344         long credit;
 345
 346         mutex_lock(&balloon_mutex);
 347
 348         do {
 349                 credit = current_target() - balloon_stats.current_pages;
 350                 if (credit > 0)
 351                         need_sleep = (increase_reservation(credit) != 0);
 352                 if (credit < 0)
 353                         need_sleep = (decrease_reservation(-credit) != 0);
 354
 355 #ifndef CONFIG_PREEMPT
 356                 if (need_resched())
 357                         schedule();
 358 #endif
 359         } while ((credit != 0) && !need_sleep);
 360
 361         /* Schedule more work if there is some still to be done. */
 362         if (current_target() != balloon_stats.current_pages)
 363                 mod_timer(&balloon_timer, jiffies + HZ);
 364
 365         mutex_unlock(&balloon_mutex);
 366 }
 367
 368 /* Resets the Xen limit, sets new target, and kicks off processing. */
 369 static void balloon_set_new_target(unsigned long target)
 370 {
 371         /* No need for lock. Not read-modify-write updates. */
 372         balloon_stats.hard_limit   = ~0UL;
 373         balloon_stats.target_pages = target;
 374         schedule_work(&balloon_worker);
 375 }
 376
 377 static struct xenbus_watch target_watch =
 378 {
 379         .node = "memory/target"
 380 };
 381
 382 /* React to a change in the target key */
 383 static void watch_target(struct xenbus_watch *watch,
 384                          const char **vec, unsigned int len)
 385 {
 386         unsigned long long new_target;
 387         int err;
 388
 389         err = xenbus_scanf(XBT_NIL, "memory", "target", "%llu", &new_target);
 390         if (err != 1) {
 391                 /* This is ok (for domain0 at least) - so just return */
 392                 return;
 393         }
 394
 395         /* The given memory/target value is in KiB, so it needs converting to
 396          * pages. PAGE_SHIFT converts bytes to pages, hence PAGE_SHIFT - 10.
 397          */
 398         balloon_set_new_target(new_target >> (PAGE_SHIFT - 10));
 399 }
 400
 401 static int balloon_init_watcher(struct notifier_block *notifier,
 402                                 unsigned long event,
 403                                 void *data)
 404 {
 405         int err;
 406
 407         err = register_xenbus_watch(&target_watch);
 408         if (err)
 409                 printk(KERN_ERR "Failed to set balloon watcher\n");
 410
 411         return NOTIFY_DONE;
 412 }
 413
 414 static struct notifier_block xenstore_notifier;
 415
 416 static int __init balloon_init(void)
 417 {
 418         unsigned long pfn;
 419         struct page *page;
 420
 421         if (!xen_pv_domain())
 422                 return -ENODEV;
 423
 424         pr_info("xen_balloon: Initialising balloon driver.\n");
 425
 426         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
 427         totalram_pages   = balloon_stats.current_pages;
 428         balloon_stats.target_pages  = balloon_stats.current_pages;
 429         balloon_stats.balloon_low   = 0;
 430         balloon_stats.balloon_high  = 0;
 431         balloon_stats.driver_pages  = 0UL;
 432         balloon_stats.hard_limit    = ~0UL;
 433
 434         init_timer(&balloon_timer);
 435         balloon_timer.data = 0;
 436         balloon_timer.function = balloon_alarm;
 437
 438         register_balloon(&balloon_sysdev);
 439
 440         /* Initialise the balloon with excess memory space. */
 441         for (pfn = xen_start_info->nr_pages; pfn < max_pfn; pfn++) {
 442                 page = pfn_to_page(pfn);
 443                 if (!PageReserved(page))
 444                         balloon_append(page);
 445         }
 446
 447         target_watch.callback = watch_target;
 448         xenstore_notifier.notifier_call = balloon_init_watcher;
 449
 450         register_xenstore_notifier(&xenstore_notifier);
 451
 452         return 0;
 453 }
 454
 455 subsys_initcall(balloon_init);
 456
 457 static void balloon_exit(void)
 458 {
 459     /* XXX - release balloon here */
 460     return;
 461 }
 462
 463 module_exit(balloon_exit);
 464
 465 #define BALLOON_SHOW(name, format, args...)                             \
 466         static ssize_t show_##name(struct sys_device *dev,              \
 467                                    struct sysdev_attribute *attr,       \
 468                                    char *buf)                           \
 469         {                                                               \
 470                 return sprintf(buf, format, ##args);                    \
 471         }                                                               \
 472         static SYSDEV_ATTR(name, S_IRUGO, show_##name, NULL)
 473
 474 BALLOON_SHOW(current_kb, "%lu\n", PAGES2KB(balloon_stats.current_pages));
 475 BALLOON_SHOW(low_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_low));
 476 BALLOON_SHOW(high_kb, "%lu\n", PAGES2KB(balloon_stats.balloon_high));
 477 BALLOON_SHOW(hard_limit_kb,
 478              (balloon_stats.hard_limit!=~0UL) ? "%lu\n" : "???\n",
 479              (balloon_stats.hard_limit!=~0UL) ? PAGES2KB(balloon_stats.hard_limit) : 0);
 480 BALLOON_SHOW(driver_kb, "%lu\n", PAGES2KB(balloon_stats.driver_pages));
 481
 482 static ssize_t show_target_kb(struct sys_device *dev, struct sysdev_attribute *attr,
 483                               char *buf)
 484 {
 485         return sprintf(buf, "%lu\n", PAGES2KB(balloon_stats.target_pages));
 486 }
 487
 488 static ssize_t store_target_kb(struct sys_device *dev,
 489                                struct sysdev_attribute *attr,
 490                                const char *buf,
 491                                size_t count)
 492 {
 493         char *endchar;
 494         unsigned long long target_bytes;
 495
 496         if (!capable(CAP_SYS_ADMIN))
 497                 return -EPERM;
 498
 499         target_bytes = simple_strtoull(buf, &endchar, 0) * 1024;
 500
 501         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
 502
 503         return count;
 504 }
 505
 506 static SYSDEV_ATTR(target_kb, S_IRUGO | S_IWUSR,
 507                    show_target_kb, store_target_kb);
 508
 509
 510 static ssize_t show_target(struct sys_device *dev, struct sysdev_attribute *attr,
 511                               char *buf)
 512 {
 513         return sprintf(buf, "%llu\n",
 514                        (unsigned long long)balloon_stats.target_pages
 515                        << PAGE_SHIFT);
 516 }
 517
 518 static ssize_t store_target(struct sys_device *dev,
 519                             struct sysdev_attribute *attr,
 520                             const char *buf,
 521                             size_t count)
 522 {
 523         char *endchar;
 524         unsigned long long target_bytes;
 525
 526         if (!capable(CAP_SYS_ADMIN))
 527                 return -EPERM;
 528
 529         target_bytes = memparse(buf, &endchar);
 530
 531         balloon_set_new_target(target_bytes >> PAGE_SHIFT);
 532
 533         return count;
 534 }
 535
 536 static SYSDEV_ATTR(target, S_IRUGO | S_IWUSR,
 537                    show_target, store_target);
 538
 539
 540 static struct sysdev_attribute *balloon_attrs[] = {
 541         &attr_target_kb,
 542         &attr_target,
 543 };
 544
 545 static struct attribute *balloon_info_attrs[] = {
 546         &attr_current_kb.attr,
 547         &attr_low_kb.attr,
 548         &attr_high_kb.attr,
 549         &attr_hard_limit_kb.attr,
 550         &attr_driver_kb.attr,
 551         NULL
 552 };
 553
 554 static struct attribute_group balloon_info_group = {
 555         .name = "info",
 556         .attrs = balloon_info_attrs,
 557 };
 558
 559 static struct sysdev_class balloon_sysdev_class = {
 560         .name = BALLOON_CLASS_NAME,
 561 };
 562
 563 static int register_balloon(struct sys_device *sysdev)
 564 {
 565         int i, error;
 566
 567         error = sysdev_class_register(&balloon_sysdev_class);
 568         if (error)
 569                 return error;
 570
 571         sysdev->id = 0;
 572         sysdev->cls = &balloon_sysdev_class;
 573
 574         error = sysdev_register(sysdev);
 575         if (error) {
 576                 sysdev_class_unregister(&balloon_sysdev_class);
 577                 return error;
 578         }
 579
 580         for (i = 0; i < ARRAY_SIZE(balloon_attrs); i++) {
 581                 error = sysdev_create_file(sysdev, balloon_attrs[i]);
 582                 if (error)
 583                         goto fail;
 584         }
 585
 586         error = sysfs_create_group(&sysdev->kobj, &balloon_info_group);
 587         if (error)
 588                 goto fail;
 589
 590         return 0;
 591
 592  fail:
 593         while (--i >= 0)
 594                 sysdev_remove_file(sysdev, balloon_attrs[i]);
 595         sysdev_unregister(sysdev);
 596         sysdev_class_unregister(&balloon_sysdev_class);
 597         return error;
 598 }
 599
 600 MODULE_LICENSE("GPL");