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/cred.h> 45 #include <linux/errno.h> 46 #include <linux/mm.h> 47 #include <linux/memblock.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 #include <linux/sysctl.h> 59 60 #include <asm/page.h> 61 #include <asm/pgalloc.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 #include <xen/mem-reservation.h> 74 75 static int xen_hotplug_unpopulated; 76 77 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 78 79 static struct ctl_table balloon_table[] = { 80 { 81 .procname = "hotplug_unpopulated", 82 .data = &xen_hotplug_unpopulated, 83 .maxlen = sizeof(int), 84 .mode = 0644, 85 .proc_handler = proc_dointvec_minmax, 86 .extra1 = SYSCTL_ZERO, 87 .extra2 = SYSCTL_ONE, 88 }, 89 { } 90 }; 91 92 static struct ctl_table balloon_root[] = { 93 { 94 .procname = "balloon", 95 .mode = 0555, 96 .child = balloon_table, 97 }, 98 { } 99 }; 100 101 static struct ctl_table xen_root[] = { 102 { 103 .procname = "xen", 104 .mode = 0555, 105 .child = balloon_root, 106 }, 107 { } 108 }; 109 110 #endif 111 112 /* 113 * Use one extent per PAGE_SIZE to avoid to break down the page into 114 * multiple frame. 115 */ 116 #define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1) 117 118 /* 119 * balloon_process() state: 120 * 121 * BP_DONE: done or nothing to do, 122 * BP_WAIT: wait to be rescheduled, 123 * BP_EAGAIN: error, go to sleep, 124 * BP_ECANCELED: error, balloon operation canceled. 125 */ 126 127 enum bp_state { 128 BP_DONE, 129 BP_WAIT, 130 BP_EAGAIN, 131 BP_ECANCELED 132 }; 133 134 135 static DEFINE_MUTEX(balloon_mutex); 136 137 struct balloon_stats balloon_stats; 138 EXPORT_SYMBOL_GPL(balloon_stats); 139 140 /* We increase/decrease in batches which fit in a page */ 141 static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)]; 142 143 144 /* List of ballooned pages, threaded through the mem_map array. */ 145 static LIST_HEAD(ballooned_pages); 146 static DECLARE_WAIT_QUEUE_HEAD(balloon_wq); 147 148 /* Main work function, always executed in process context. */ 149 static void balloon_process(struct work_struct *work); 150 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process); 151 152 /* When ballooning out (allocating memory to return to Xen) we don't really 153 want the kernel to try too hard since that can trigger the oom killer. */ 154 #define GFP_BALLOON \ 155 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC) 156 157 /* balloon_append: add the given page to the balloon. */ 158 static void balloon_append(struct page *page) 159 { 160 __SetPageOffline(page); 161 162 /* Lowmem is re-populated first, so highmem pages go at list tail. */ 163 if (PageHighMem(page)) { 164 list_add_tail(&page->lru, &ballooned_pages); 165 balloon_stats.balloon_high++; 166 } else { 167 list_add(&page->lru, &ballooned_pages); 168 balloon_stats.balloon_low++; 169 } 170 wake_up(&balloon_wq); 171 } 172 173 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */ 174 static struct page *balloon_retrieve(bool require_lowmem) 175 { 176 struct page *page; 177 178 if (list_empty(&ballooned_pages)) 179 return NULL; 180 181 page = list_entry(ballooned_pages.next, struct page, lru); 182 if (require_lowmem && PageHighMem(page)) 183 return NULL; 184 list_del(&page->lru); 185 186 if (PageHighMem(page)) 187 balloon_stats.balloon_high--; 188 else 189 balloon_stats.balloon_low--; 190 191 __ClearPageOffline(page); 192 return page; 193 } 194 195 static struct page *balloon_next_page(struct page *page) 196 { 197 struct list_head *next = page->lru.next; 198 if (next == &ballooned_pages) 199 return NULL; 200 return list_entry(next, struct page, lru); 201 } 202 203 static enum bp_state update_schedule(enum bp_state state) 204 { 205 if (state == BP_WAIT) 206 return BP_WAIT; 207 208 if (state == BP_ECANCELED) 209 return BP_ECANCELED; 210 211 if (state == BP_DONE) { 212 balloon_stats.schedule_delay = 1; 213 balloon_stats.retry_count = 1; 214 return BP_DONE; 215 } 216 217 ++balloon_stats.retry_count; 218 219 if (balloon_stats.max_retry_count != RETRY_UNLIMITED && 220 balloon_stats.retry_count > balloon_stats.max_retry_count) { 221 balloon_stats.schedule_delay = 1; 222 balloon_stats.retry_count = 1; 223 return BP_ECANCELED; 224 } 225 226 balloon_stats.schedule_delay <<= 1; 227 228 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay) 229 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay; 230 231 return BP_EAGAIN; 232 } 233 234 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 235 static void release_memory_resource(struct resource *resource) 236 { 237 if (!resource) 238 return; 239 240 /* 241 * No need to reset region to identity mapped since we now 242 * know that no I/O can be in this region 243 */ 244 release_resource(resource); 245 kfree(resource); 246 } 247 248 static struct resource *additional_memory_resource(phys_addr_t size) 249 { 250 struct resource *res; 251 int ret; 252 253 res = kzalloc(sizeof(*res), GFP_KERNEL); 254 if (!res) 255 return NULL; 256 257 res->name = "System RAM"; 258 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY; 259 260 ret = allocate_resource(&iomem_resource, res, 261 size, 0, -1, 262 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL); 263 if (ret < 0) { 264 pr_err("Cannot allocate new System RAM resource\n"); 265 kfree(res); 266 return NULL; 267 } 268 269 #ifdef CONFIG_SPARSEMEM 270 { 271 unsigned long limit = 1UL << (MAX_PHYSMEM_BITS - PAGE_SHIFT); 272 unsigned long pfn = res->start >> PAGE_SHIFT; 273 274 if (pfn > limit) { 275 pr_err("New System RAM resource outside addressable RAM (%lu > %lu)\n", 276 pfn, limit); 277 release_memory_resource(res); 278 return NULL; 279 } 280 } 281 #endif 282 283 return res; 284 } 285 286 static enum bp_state reserve_additional_memory(void) 287 { 288 long credit; 289 struct resource *resource; 290 int nid, rc; 291 unsigned long balloon_hotplug; 292 293 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated 294 - balloon_stats.total_pages; 295 296 /* 297 * Already hotplugged enough pages? Wait for them to be 298 * onlined. 299 */ 300 if (credit <= 0) 301 return BP_WAIT; 302 303 balloon_hotplug = round_up(credit, PAGES_PER_SECTION); 304 305 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE); 306 if (!resource) 307 goto err; 308 309 nid = memory_add_physaddr_to_nid(resource->start); 310 311 #ifdef CONFIG_XEN_HAVE_PVMMU 312 /* 313 * We don't support PV MMU when Linux and Xen is using 314 * different page granularity. 315 */ 316 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE); 317 318 /* 319 * add_memory() will build page tables for the new memory so 320 * the p2m must contain invalid entries so the correct 321 * non-present PTEs will be written. 322 * 323 * If a failure occurs, the original (identity) p2m entries 324 * are not restored since this region is now known not to 325 * conflict with any devices. 326 */ 327 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 328 unsigned long pfn, i; 329 330 pfn = PFN_DOWN(resource->start); 331 for (i = 0; i < balloon_hotplug; i++) { 332 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) { 333 pr_warn("set_phys_to_machine() failed, no memory added\n"); 334 goto err; 335 } 336 } 337 } 338 #endif 339 340 /* 341 * add_memory_resource() will call online_pages() which in its turn 342 * will call xen_online_page() callback causing deadlock if we don't 343 * release balloon_mutex here. Unlocking here is safe because the 344 * callers drop the mutex before trying again. 345 */ 346 mutex_unlock(&balloon_mutex); 347 /* add_memory_resource() requires the device_hotplug lock */ 348 lock_device_hotplug(); 349 rc = add_memory_resource(nid, resource); 350 unlock_device_hotplug(); 351 mutex_lock(&balloon_mutex); 352 353 if (rc) { 354 pr_warn("Cannot add additional memory (%i)\n", rc); 355 goto err; 356 } 357 358 balloon_stats.total_pages += balloon_hotplug; 359 360 return BP_WAIT; 361 err: 362 release_memory_resource(resource); 363 return BP_ECANCELED; 364 } 365 366 static void xen_online_page(struct page *page, unsigned int order) 367 { 368 unsigned long i, size = (1 << order); 369 unsigned long start_pfn = page_to_pfn(page); 370 struct page *p; 371 372 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn); 373 mutex_lock(&balloon_mutex); 374 for (i = 0; i < size; i++) { 375 p = pfn_to_page(start_pfn + i); 376 balloon_append(p); 377 } 378 mutex_unlock(&balloon_mutex); 379 } 380 381 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v) 382 { 383 if (val == MEM_ONLINE) 384 schedule_delayed_work(&balloon_worker, 0); 385 386 return NOTIFY_OK; 387 } 388 389 static struct notifier_block xen_memory_nb = { 390 .notifier_call = xen_memory_notifier, 391 .priority = 0 392 }; 393 #else 394 static enum bp_state reserve_additional_memory(void) 395 { 396 balloon_stats.target_pages = balloon_stats.current_pages + 397 balloon_stats.target_unpopulated; 398 return BP_ECANCELED; 399 } 400 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */ 401 402 static long current_credit(void) 403 { 404 return balloon_stats.target_pages - balloon_stats.current_pages; 405 } 406 407 static bool balloon_is_inflated(void) 408 { 409 return balloon_stats.balloon_low || balloon_stats.balloon_high; 410 } 411 412 static enum bp_state increase_reservation(unsigned long nr_pages) 413 { 414 int rc; 415 unsigned long i; 416 struct page *page; 417 418 if (nr_pages > ARRAY_SIZE(frame_list)) 419 nr_pages = ARRAY_SIZE(frame_list); 420 421 page = list_first_entry_or_null(&ballooned_pages, struct page, lru); 422 for (i = 0; i < nr_pages; i++) { 423 if (!page) { 424 nr_pages = i; 425 break; 426 } 427 428 frame_list[i] = page_to_xen_pfn(page); 429 page = balloon_next_page(page); 430 } 431 432 rc = xenmem_reservation_increase(nr_pages, frame_list); 433 if (rc <= 0) 434 return BP_EAGAIN; 435 436 for (i = 0; i < rc; i++) { 437 page = balloon_retrieve(false); 438 BUG_ON(page == NULL); 439 440 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]); 441 442 /* Relinquish the page back to the allocator. */ 443 free_reserved_page(page); 444 } 445 446 balloon_stats.current_pages += rc; 447 448 return BP_DONE; 449 } 450 451 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp) 452 { 453 enum bp_state state = BP_DONE; 454 unsigned long i; 455 struct page *page, *tmp; 456 int ret; 457 LIST_HEAD(pages); 458 459 if (nr_pages > ARRAY_SIZE(frame_list)) 460 nr_pages = ARRAY_SIZE(frame_list); 461 462 for (i = 0; i < nr_pages; i++) { 463 page = alloc_page(gfp); 464 if (page == NULL) { 465 nr_pages = i; 466 state = BP_EAGAIN; 467 break; 468 } 469 adjust_managed_page_count(page, -1); 470 xenmem_reservation_scrub_page(page); 471 list_add(&page->lru, &pages); 472 } 473 474 /* 475 * Ensure that ballooned highmem pages don't have kmaps. 476 * 477 * Do this before changing the p2m as kmap_flush_unused() 478 * reads PTEs to obtain pages (and hence needs the original 479 * p2m entry). 480 */ 481 kmap_flush_unused(); 482 483 /* 484 * Setup the frame, update direct mapping, invalidate P2M, 485 * and add to balloon. 486 */ 487 i = 0; 488 list_for_each_entry_safe(page, tmp, &pages, lru) { 489 frame_list[i++] = xen_page_to_gfn(page); 490 491 xenmem_reservation_va_mapping_reset(1, &page); 492 493 list_del(&page->lru); 494 495 balloon_append(page); 496 } 497 498 flush_tlb_all(); 499 500 ret = xenmem_reservation_decrease(nr_pages, frame_list); 501 BUG_ON(ret != nr_pages); 502 503 balloon_stats.current_pages -= nr_pages; 504 505 return state; 506 } 507 508 /* 509 * As this is a work item it is guaranteed to run as a single instance only. 510 * We may of course race updates of the target counts (which are protected 511 * by the balloon lock), or with changes to the Xen hard limit, but we will 512 * recover from these in time. 513 */ 514 static void balloon_process(struct work_struct *work) 515 { 516 enum bp_state state = BP_DONE; 517 long credit; 518 519 520 do { 521 mutex_lock(&balloon_mutex); 522 523 credit = current_credit(); 524 525 if (credit > 0) { 526 if (balloon_is_inflated()) 527 state = increase_reservation(credit); 528 else 529 state = reserve_additional_memory(); 530 } 531 532 if (credit < 0) { 533 long n_pages; 534 535 n_pages = min(-credit, si_mem_available()); 536 state = decrease_reservation(n_pages, GFP_BALLOON); 537 if (state == BP_DONE && n_pages != -credit && 538 n_pages < totalreserve_pages) 539 state = BP_EAGAIN; 540 } 541 542 state = update_schedule(state); 543 544 mutex_unlock(&balloon_mutex); 545 546 cond_resched(); 547 548 } while (credit && state == BP_DONE); 549 550 /* Schedule more work if there is some still to be done. */ 551 if (state == BP_EAGAIN) 552 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ); 553 } 554 555 /* Resets the Xen limit, sets new target, and kicks off processing. */ 556 void balloon_set_new_target(unsigned long target) 557 { 558 /* No need for lock. Not read-modify-write updates. */ 559 balloon_stats.target_pages = target; 560 schedule_delayed_work(&balloon_worker, 0); 561 } 562 EXPORT_SYMBOL_GPL(balloon_set_new_target); 563 564 static int add_ballooned_pages(int nr_pages) 565 { 566 enum bp_state st; 567 568 if (xen_hotplug_unpopulated) { 569 st = reserve_additional_memory(); 570 if (st != BP_ECANCELED) { 571 mutex_unlock(&balloon_mutex); 572 wait_event(balloon_wq, 573 !list_empty(&ballooned_pages)); 574 mutex_lock(&balloon_mutex); 575 return 0; 576 } 577 } 578 579 if (si_mem_available() < nr_pages) 580 return -ENOMEM; 581 582 st = decrease_reservation(nr_pages, GFP_USER); 583 if (st != BP_DONE) 584 return -ENOMEM; 585 586 return 0; 587 } 588 589 /** 590 * alloc_xenballooned_pages - get pages that have been ballooned out 591 * @nr_pages: Number of pages to get 592 * @pages: pages returned 593 * @return 0 on success, error otherwise 594 */ 595 int alloc_xenballooned_pages(int nr_pages, struct page **pages) 596 { 597 int pgno = 0; 598 struct page *page; 599 int ret; 600 601 mutex_lock(&balloon_mutex); 602 603 balloon_stats.target_unpopulated += nr_pages; 604 605 while (pgno < nr_pages) { 606 page = balloon_retrieve(true); 607 if (page) { 608 pages[pgno++] = page; 609 #ifdef CONFIG_XEN_HAVE_PVMMU 610 /* 611 * We don't support PV MMU when Linux and Xen is using 612 * different page granularity. 613 */ 614 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE); 615 616 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 617 ret = xen_alloc_p2m_entry(page_to_pfn(page)); 618 if (ret < 0) 619 goto out_undo; 620 } 621 #endif 622 } else { 623 ret = add_ballooned_pages(nr_pages - pgno); 624 if (ret < 0) 625 goto out_undo; 626 } 627 } 628 mutex_unlock(&balloon_mutex); 629 return 0; 630 out_undo: 631 mutex_unlock(&balloon_mutex); 632 free_xenballooned_pages(pgno, pages); 633 return ret; 634 } 635 EXPORT_SYMBOL(alloc_xenballooned_pages); 636 637 /** 638 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages 639 * @nr_pages: Number of pages 640 * @pages: pages to return 641 */ 642 void free_xenballooned_pages(int nr_pages, struct page **pages) 643 { 644 int i; 645 646 mutex_lock(&balloon_mutex); 647 648 for (i = 0; i < nr_pages; i++) { 649 if (pages[i]) 650 balloon_append(pages[i]); 651 } 652 653 balloon_stats.target_unpopulated -= nr_pages; 654 655 /* The balloon may be too large now. Shrink it if needed. */ 656 if (current_credit()) 657 schedule_delayed_work(&balloon_worker, 0); 658 659 mutex_unlock(&balloon_mutex); 660 } 661 EXPORT_SYMBOL(free_xenballooned_pages); 662 663 #ifdef CONFIG_XEN_PV 664 static void __init balloon_add_region(unsigned long start_pfn, 665 unsigned long pages) 666 { 667 unsigned long pfn, extra_pfn_end; 668 669 /* 670 * If the amount of usable memory has been limited (e.g., with 671 * the 'mem' command line parameter), don't add pages beyond 672 * this limit. 673 */ 674 extra_pfn_end = min(max_pfn, start_pfn + pages); 675 676 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) { 677 /* totalram_pages and totalhigh_pages do not 678 include the boot-time balloon extension, so 679 don't subtract from it. */ 680 balloon_append(pfn_to_page(pfn)); 681 } 682 683 balloon_stats.total_pages += extra_pfn_end - start_pfn; 684 } 685 #endif 686 687 static int __init balloon_init(void) 688 { 689 if (!xen_domain()) 690 return -ENODEV; 691 692 pr_info("Initialising balloon driver\n"); 693 694 #ifdef CONFIG_XEN_PV 695 balloon_stats.current_pages = xen_pv_domain() 696 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn) 697 : get_num_physpages(); 698 #else 699 balloon_stats.current_pages = get_num_physpages(); 700 #endif 701 balloon_stats.target_pages = balloon_stats.current_pages; 702 balloon_stats.balloon_low = 0; 703 balloon_stats.balloon_high = 0; 704 balloon_stats.total_pages = balloon_stats.current_pages; 705 706 balloon_stats.schedule_delay = 1; 707 balloon_stats.max_schedule_delay = 32; 708 balloon_stats.retry_count = 1; 709 balloon_stats.max_retry_count = 4; 710 711 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 712 set_online_page_callback(&xen_online_page); 713 register_memory_notifier(&xen_memory_nb); 714 register_sysctl_table(xen_root); 715 #endif 716 717 #ifdef CONFIG_XEN_PV 718 { 719 int i; 720 721 /* 722 * Initialize the balloon with pages from the extra memory 723 * regions (see arch/x86/xen/setup.c). 724 */ 725 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) 726 if (xen_extra_mem[i].n_pfns) 727 balloon_add_region(xen_extra_mem[i].start_pfn, 728 xen_extra_mem[i].n_pfns); 729 } 730 #endif 731 732 /* Init the xen-balloon driver. */ 733 xen_balloon_init(); 734 735 return 0; 736 } 737 subsys_initcall(balloon_init); 738