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  *
   8  * This program is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU General Public License version 2
  10  * as published by the Free Software Foundation; or, when distributed
  11  * separately from the Linux kernel or incorporated into other
  12  * software packages, subject to the following license:
  13  *
  14  * Permission is hereby granted, free of charge, to any person obtaining a copy
  15  * of this source file (the "Software"), to deal in the Software without
  16  * restriction, including without limitation the rights to use, copy, modify,
  17  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  18  * and to permit persons to whom the Software is furnished to do so, subject to
  19  * the following conditions:
  20  *
  21  * The above copyright notice and this permission notice shall be included in
  22  * all copies or substantial portions of the Software.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  25  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  26  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  27  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  28  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  29  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  30  * IN THE SOFTWARE.
  31  */
  32
  33 #include <linux/kernel.h>
  34 #include <linux/sched.h>
  35 #include <linux/errno.h>
  36 #include <linux/mm.h>
  37 #include <linux/bootmem.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/highmem.h>
  40 #include <linux/mutex.h>
  41 #include <linux/list.h>
  42 #include <linux/gfp.h>
  43
  44 #include <asm/page.h>
  45 #include <asm/pgalloc.h>
  46 #include <asm/pgtable.h>
  47 #include <asm/tlb.h>
  48 #include <asm/e820.h>
  49
  50 #include <asm/xen/hypervisor.h>
  51 #include <asm/xen/hypercall.h>
  52
  53 #include <xen/xen.h>
  54 #include <xen/interface/xen.h>
  55 #include <xen/interface/memory.h>
  56 #include <xen/balloon.h>
  57 #include <xen/features.h>
  58 #include <xen/page.h>
  59
  60 /*
  61  * balloon_process() state:
  62  *
  63  * BP_DONE: done or nothing to do,
  64  * BP_EAGAIN: error, go to sleep,
  65  * BP_ECANCELED: error, balloon operation canceled.
  66  */
  67
  68 enum bp_state {
  69         BP_DONE,
  70         BP_EAGAIN,
  71         BP_ECANCELED
  72 };
  73
  74
  75 static DEFINE_MUTEX(balloon_mutex);
  76
  77 struct balloon_stats balloon_stats;
  78 EXPORT_SYMBOL_GPL(balloon_stats);
  79
  80 /* We increase/decrease in batches which fit in a page */
  81 static unsigned long frame_list[PAGE_SIZE / sizeof(unsigned long)];
  82
  83 #ifdef CONFIG_HIGHMEM
  84 #define inc_totalhigh_pages() (totalhigh_pages++)
  85 #define dec_totalhigh_pages() (totalhigh_pages--)
  86 #else
  87 #define inc_totalhigh_pages() do {} while(0)
  88 #define dec_totalhigh_pages() do {} while(0)
  89 #endif
  90
  91 /* List of ballooned pages, threaded through the mem_map array. */
  92 static LIST_HEAD(ballooned_pages);
  93
  94 /* Main work function, always executed in process context. */
  95 static void balloon_process(struct work_struct *work);
  96 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process);
  97
  98 /* When ballooning out (allocating memory to return to Xen) we don't really
  99    want the kernel to try too hard since that can trigger the oom killer. */
 100 #define GFP_BALLOON \
 101         (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)
 102
 103 static void scrub_page(struct page *page)
 104 {
 105 #ifdef CONFIG_XEN_SCRUB_PAGES
 106         clear_highpage(page);
 107 #endif
 108 }
 109
 110 /* balloon_append: add the given page to the balloon. */
 111 static void __balloon_append(struct page *page)
 112 {
 113         /* Lowmem is re-populated first, so highmem pages go at list tail. */
 114         if (PageHighMem(page)) {
 115                 list_add_tail(&page->lru, &ballooned_pages);
 116                 balloon_stats.balloon_high++;
 117                 dec_totalhigh_pages();
 118         } else {
 119                 list_add(&page->lru, &ballooned_pages);
 120                 balloon_stats.balloon_low++;
 121         }
 122 }
 123
 124 static void balloon_append(struct page *page)
 125 {
 126         __balloon_append(page);
 127         totalram_pages--;
 128 }
 129
 130 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */
 131 static struct page *balloon_retrieve(bool prefer_highmem)
 132 {
 133         struct page *page;
 134
 135         if (list_empty(&ballooned_pages))
 136                 return NULL;
 137
 138         if (prefer_highmem)
 139                 page = list_entry(ballooned_pages.prev, struct page, lru);
 140         else
 141                 page = list_entry(ballooned_pages.next, struct page, lru);
 142         list_del(&page->lru);
 143
 144         if (PageHighMem(page)) {
 145                 balloon_stats.balloon_high--;
 146                 inc_totalhigh_pages();
 147         }
 148         else
 149                 balloon_stats.balloon_low--;
 150
 151         totalram_pages++;
 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 static unsigned long current_target(void)
 197 {
 198         unsigned long target = balloon_stats.target_pages;
 199
 200         target = min(target,
 201                      balloon_stats.current_pages +
 202                      balloon_stats.balloon_low +
 203                      balloon_stats.balloon_high);
 204
 205         return target;
 206 }
 207
 208 static enum bp_state increase_reservation(unsigned long nr_pages)
 209 {
 210         int rc;
 211         unsigned long  pfn, i;
 212         struct page   *page;
 213         struct xen_memory_reservation reservation = {
 214                 .address_bits = 0,
 215                 .extent_order = 0,
 216                 .domid        = DOMID_SELF
 217         };
 218
 219         if (nr_pages > ARRAY_SIZE(frame_list))
 220                 nr_pages = ARRAY_SIZE(frame_list);
 221
 222         page = balloon_first_page();
 223         for (i = 0; i < nr_pages; i++) {
 224                 if (!page) {
 225                         nr_pages = i;
 226                         break;
 227                 }
 228                 frame_list[i] = page_to_pfn(page);
 229                 page = balloon_next_page(page);
 230         }
 231
 232         set_xen_guest_handle(reservation.extent_start, frame_list);
 233         reservation.nr_extents = nr_pages;
 234         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 235         if (rc <= 0)
 236                 return BP_EAGAIN;
 237
 238         for (i = 0; i < rc; i++) {
 239                 page = balloon_retrieve(false);
 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 += rc;
 265
 266         return BP_DONE;
 267 }
 268
 269 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)
 270 {
 271         enum bp_state state = BP_DONE;
 272         unsigned long  pfn, i;
 273         struct page   *page;
 274         int ret;
 275         struct xen_memory_reservation reservation = {
 276                 .address_bits = 0,
 277                 .extent_order = 0,
 278                 .domid        = DOMID_SELF
 279         };
 280
 281         if (nr_pages > ARRAY_SIZE(frame_list))
 282                 nr_pages = ARRAY_SIZE(frame_list);
 283
 284         for (i = 0; i < nr_pages; i++) {
 285                 if ((page = alloc_page(gfp)) == NULL) {
 286                         nr_pages = i;
 287                         state = BP_EAGAIN;
 288                         break;
 289                 }
 290
 291                 pfn = page_to_pfn(page);
 292                 frame_list[i] = pfn_to_mfn(pfn);
 293
 294                 scrub_page(page);
 295
 296                 if (!PageHighMem(page)) {
 297                         ret = HYPERVISOR_update_va_mapping(
 298                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 299                                 __pte_ma(0), 0);
 300                         BUG_ON(ret);
 301                 }
 302
 303         }
 304
 305         /* Ensure that ballooned highmem pages don't have kmaps. */
 306         kmap_flush_unused();
 307         flush_tlb_all();
 308
 309         /* No more mappings: invalidate P2M and add to balloon. */
 310         for (i = 0; i < nr_pages; i++) {
 311                 pfn = mfn_to_pfn(frame_list[i]);
 312                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 313                 balloon_append(pfn_to_page(pfn));
 314         }
 315
 316         set_xen_guest_handle(reservation.extent_start, frame_list);
 317         reservation.nr_extents   = nr_pages;
 318         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 319         BUG_ON(ret != nr_pages);
 320
 321         balloon_stats.current_pages -= nr_pages;
 322
 323         return state;
 324 }
 325
 326 /*
 327  * We avoid multiple worker processes conflicting via the balloon mutex.
 328  * We may of course race updates of the target counts (which are protected
 329  * by the balloon lock), or with changes to the Xen hard limit, but we will
 330  * recover from these in time.
 331  */
 332 static void balloon_process(struct work_struct *work)
 333 {
 334         enum bp_state state = BP_DONE;
 335         long credit;
 336
 337         mutex_lock(&balloon_mutex);
 338
 339         do {
 340                 credit = current_target() - balloon_stats.current_pages;
 341
 342                 if (credit > 0)
 343                         state = increase_reservation(credit);
 344
 345                 if (credit < 0)
 346                         state = decrease_reservation(-credit, GFP_BALLOON);
 347
 348                 state = update_schedule(state);
 349
 350 #ifndef CONFIG_PREEMPT
 351                 if (need_resched())
 352                         schedule();
 353 #endif
 354         } while (credit && state == BP_DONE);
 355
 356         /* Schedule more work if there is some still to be done. */
 357         if (state == BP_EAGAIN)
 358                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 359
 360         mutex_unlock(&balloon_mutex);
 361 }
 362
 363 /* Resets the Xen limit, sets new target, and kicks off processing. */
 364 void balloon_set_new_target(unsigned long target)
 365 {
 366         /* No need for lock. Not read-modify-write updates. */
 367         balloon_stats.target_pages = target;
 368         schedule_delayed_work(&balloon_worker, 0);
 369 }
 370 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 371
 372 /**
 373  * alloc_xenballooned_pages - get pages that have been ballooned out
 374  * @nr_pages: Number of pages to get
 375  * @pages: pages returned
 376  * @return 0 on success, error otherwise
 377  */
 378 int alloc_xenballooned_pages(int nr_pages, struct page** pages)
 379 {
 380         int pgno = 0;
 381         struct page* page;
 382         mutex_lock(&balloon_mutex);
 383         while (pgno < nr_pages) {
 384                 page = balloon_retrieve(true);
 385                 if (page) {
 386                         pages[pgno++] = page;
 387                 } else {
 388                         enum bp_state st;
 389                         st = decrease_reservation(nr_pages - pgno, GFP_HIGHUSER);
 390                         if (st != BP_DONE)
 391                                 goto out_undo;
 392                 }
 393         }
 394         mutex_unlock(&balloon_mutex);
 395         return 0;
 396  out_undo:
 397         while (pgno)
 398                 balloon_append(pages[--pgno]);
 399         /* Free the memory back to the kernel soon */
 400         schedule_delayed_work(&balloon_worker, 0);
 401         mutex_unlock(&balloon_mutex);
 402         return -ENOMEM;
 403 }
 404 EXPORT_SYMBOL(alloc_xenballooned_pages);
 405
 406 /**
 407  * free_xenballooned_pages - return pages retrieved with get_ballooned_pages
 408  * @nr_pages: Number of pages
 409  * @pages: pages to return
 410  */
 411 void free_xenballooned_pages(int nr_pages, struct page** pages)
 412 {
 413         int i;
 414
 415         mutex_lock(&balloon_mutex);
 416
 417         for (i = 0; i < nr_pages; i++) {
 418                 if (pages[i])
 419                         balloon_append(pages[i]);
 420         }
 421
 422         /* The balloon may be too large now. Shrink it if needed. */
 423         if (current_target() != balloon_stats.current_pages)
 424                 schedule_delayed_work(&balloon_worker, 0);
 425
 426         mutex_unlock(&balloon_mutex);
 427 }
 428 EXPORT_SYMBOL(free_xenballooned_pages);
 429
 430 static int __init balloon_init(void)
 431 {
 432         unsigned long pfn, extra_pfn_end;
 433         struct page *page;
 434
 435         if (!xen_pv_domain())
 436                 return -ENODEV;
 437
 438         pr_info("xen/balloon: Initialising balloon driver.\n");
 439
 440         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
 441         balloon_stats.target_pages  = balloon_stats.current_pages;
 442         balloon_stats.balloon_low   = 0;
 443         balloon_stats.balloon_high  = 0;
 444
 445         balloon_stats.schedule_delay = 1;
 446         balloon_stats.max_schedule_delay = 32;
 447         balloon_stats.retry_count = 1;
 448         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 449
 450         /*
 451          * Initialise the balloon with excess memory space.  We need
 452          * to make sure we don't add memory which doesn't exist or
 453          * logically exist.  The E820 map can be trimmed to be smaller
 454          * than the amount of physical memory due to the mem= command
 455          * line parameter.  And if this is a 32-bit non-HIGHMEM kernel
 456          * on a system with memory which requires highmem to access,
 457          * don't try to use it.
 458          */
 459         extra_pfn_end = min(min(max_pfn, e820_end_of_ram_pfn()),
 460                             (unsigned long)PFN_DOWN(xen_extra_mem_start + xen_extra_mem_size));
 461         for (pfn = PFN_UP(xen_extra_mem_start);
 462              pfn < extra_pfn_end;
 463              pfn++) {
 464                 page = pfn_to_page(pfn);
 465                 /* totalram_pages doesn't include the boot-time
 466                    balloon extension, so don't subtract from it. */
 467                 __balloon_append(page);
 468         }
 469
 470         return 0;
 471 }
 472
 473 subsys_initcall(balloon_init);
 474
 475 MODULE_LICENSE("GPL");