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 static DEFINE_PER_CPU(struct page *, balloon_scratch_page); 96 97 98 /* List of ballooned pages, threaded through the mem_map array. */ 99 static LIST_HEAD(ballooned_pages); 100 101 /* Main work function, always executed in process context. */ 102 static void balloon_process(struct work_struct *work); 103 static DECLARE_DELAYED_WORK(balloon_worker, balloon_process); 104 105 /* When ballooning out (allocating memory to return to Xen) we don't really 106 want the kernel to try too hard since that can trigger the oom killer. */ 107 #define GFP_BALLOON \ 108 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC) 109 110 static void scrub_page(struct page *page) 111 { 112 #ifdef CONFIG_XEN_SCRUB_PAGES 113 clear_highpage(page); 114 #endif 115 } 116 117 /* balloon_append: add the given page to the balloon. */ 118 static void __balloon_append(struct page *page) 119 { 120 /* Lowmem is re-populated first, so highmem pages go at list tail. */ 121 if (PageHighMem(page)) { 122 list_add_tail(&page->lru, &ballooned_pages); 123 balloon_stats.balloon_high++; 124 } else { 125 list_add(&page->lru, &ballooned_pages); 126 balloon_stats.balloon_low++; 127 } 128 } 129 130 static void balloon_append(struct page *page) 131 { 132 __balloon_append(page); 133 adjust_managed_page_count(page, -1); 134 } 135 136 /* balloon_retrieve: rescue a page from the balloon, if it is not empty. */ 137 static struct page *balloon_retrieve(bool prefer_highmem) 138 { 139 struct page *page; 140 141 if (list_empty(&ballooned_pages)) 142 return NULL; 143 144 if (prefer_highmem) 145 page = list_entry(ballooned_pages.prev, struct page, lru); 146 else 147 page = list_entry(ballooned_pages.next, struct page, lru); 148 list_del(&page->lru); 149 150 if (PageHighMem(page)) 151 balloon_stats.balloon_high--; 152 else 153 balloon_stats.balloon_low--; 154 155 adjust_managed_page_count(page, 1); 156 157 return page; 158 } 159 160 static struct page *balloon_first_page(void) 161 { 162 if (list_empty(&ballooned_pages)) 163 return NULL; 164 return list_entry(ballooned_pages.next, struct page, lru); 165 } 166 167 static struct page *balloon_next_page(struct page *page) 168 { 169 struct list_head *next = page->lru.next; 170 if (next == &ballooned_pages) 171 return NULL; 172 return list_entry(next, struct page, lru); 173 } 174 175 static enum bp_state update_schedule(enum bp_state state) 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 long current_credit(void) 202 { 203 return balloon_stats.target_pages - balloon_stats.current_pages - 204 balloon_stats.hotplug_pages; 205 } 206 207 static bool balloon_is_inflated(void) 208 { 209 if (balloon_stats.balloon_low || balloon_stats.balloon_high || 210 balloon_stats.balloon_hotplug) 211 return true; 212 else 213 return false; 214 } 215 216 /* 217 * reserve_additional_memory() adds memory region of size >= credit above 218 * max_pfn. New region is section aligned and size is modified to be multiple 219 * of section size. Those features allow optimal use of address space and 220 * establish proper alignment when this function is called first time after 221 * boot (last section not fully populated at boot time contains unused memory 222 * pages with PG_reserved bit not set; online_pages_range() does not allow page 223 * onlining in whole range if first onlined page does not have PG_reserved 224 * bit set). Real size of added memory is established at page onlining stage. 225 */ 226 227 static enum bp_state reserve_additional_memory(long credit) 228 { 229 int nid, rc; 230 u64 hotplug_start_paddr; 231 unsigned long balloon_hotplug = credit; 232 233 hotplug_start_paddr = PFN_PHYS(SECTION_ALIGN_UP(max_pfn)); 234 balloon_hotplug = round_up(balloon_hotplug, PAGES_PER_SECTION); 235 nid = memory_add_physaddr_to_nid(hotplug_start_paddr); 236 237 rc = add_memory(nid, hotplug_start_paddr, balloon_hotplug << PAGE_SHIFT); 238 239 if (rc) { 240 pr_info("%s: add_memory() failed: %i\n", __func__, rc); 241 return BP_EAGAIN; 242 } 243 244 balloon_hotplug -= credit; 245 246 balloon_stats.hotplug_pages += credit; 247 balloon_stats.balloon_hotplug = balloon_hotplug; 248 249 return BP_DONE; 250 } 251 252 static void xen_online_page(struct page *page) 253 { 254 __online_page_set_limits(page); 255 256 mutex_lock(&balloon_mutex); 257 258 __balloon_append(page); 259 260 if (balloon_stats.hotplug_pages) 261 --balloon_stats.hotplug_pages; 262 else 263 --balloon_stats.balloon_hotplug; 264 265 mutex_unlock(&balloon_mutex); 266 } 267 268 static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v) 269 { 270 if (val == MEM_ONLINE) 271 schedule_delayed_work(&balloon_worker, 0); 272 273 return NOTIFY_OK; 274 } 275 276 static struct notifier_block xen_memory_nb = { 277 .notifier_call = xen_memory_notifier, 278 .priority = 0 279 }; 280 #else 281 static long current_credit(void) 282 { 283 unsigned long target = balloon_stats.target_pages; 284 285 target = min(target, 286 balloon_stats.current_pages + 287 balloon_stats.balloon_low + 288 balloon_stats.balloon_high); 289 290 return target - balloon_stats.current_pages; 291 } 292 293 static bool balloon_is_inflated(void) 294 { 295 if (balloon_stats.balloon_low || balloon_stats.balloon_high) 296 return true; 297 else 298 return false; 299 } 300 301 static enum bp_state reserve_additional_memory(long credit) 302 { 303 balloon_stats.target_pages = balloon_stats.current_pages; 304 return BP_DONE; 305 } 306 #endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */ 307 308 static enum bp_state increase_reservation(unsigned long nr_pages) 309 { 310 int rc; 311 unsigned long pfn, i; 312 struct page *page; 313 struct xen_memory_reservation reservation = { 314 .address_bits = 0, 315 .extent_order = 0, 316 .domid = DOMID_SELF 317 }; 318 319 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 320 if (!balloon_stats.balloon_low && !balloon_stats.balloon_high) { 321 nr_pages = min(nr_pages, balloon_stats.balloon_hotplug); 322 balloon_stats.hotplug_pages += nr_pages; 323 balloon_stats.balloon_hotplug -= nr_pages; 324 return BP_DONE; 325 } 326 #endif 327 328 if (nr_pages > ARRAY_SIZE(frame_list)) 329 nr_pages = ARRAY_SIZE(frame_list); 330 331 page = balloon_first_page(); 332 for (i = 0; i < nr_pages; i++) { 333 if (!page) { 334 nr_pages = i; 335 break; 336 } 337 frame_list[i] = page_to_pfn(page); 338 page = balloon_next_page(page); 339 } 340 341 set_xen_guest_handle(reservation.extent_start, frame_list); 342 reservation.nr_extents = nr_pages; 343 rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation); 344 if (rc <= 0) 345 return BP_EAGAIN; 346 347 for (i = 0; i < rc; i++) { 348 page = balloon_retrieve(false); 349 BUG_ON(page == NULL); 350 351 pfn = page_to_pfn(page); 352 353 set_phys_to_machine(pfn, frame_list[i]); 354 355 #ifdef CONFIG_XEN_HAVE_PVMMU 356 /* Link back into the page tables if not highmem. */ 357 if (xen_pv_domain() && !PageHighMem(page)) { 358 int ret; 359 ret = HYPERVISOR_update_va_mapping( 360 (unsigned long)__va(pfn << PAGE_SHIFT), 361 mfn_pte(frame_list[i], PAGE_KERNEL), 362 0); 363 BUG_ON(ret); 364 } 365 #endif 366 367 /* Relinquish the page back to the allocator. */ 368 __free_reserved_page(page); 369 } 370 371 balloon_stats.current_pages += rc; 372 373 return BP_DONE; 374 } 375 376 static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp) 377 { 378 enum bp_state state = BP_DONE; 379 unsigned long pfn, i; 380 struct page *page; 381 struct page *scratch_page; 382 int ret; 383 struct xen_memory_reservation reservation = { 384 .address_bits = 0, 385 .extent_order = 0, 386 .domid = DOMID_SELF 387 }; 388 389 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 390 if (balloon_stats.hotplug_pages) { 391 nr_pages = min(nr_pages, balloon_stats.hotplug_pages); 392 balloon_stats.hotplug_pages -= nr_pages; 393 balloon_stats.balloon_hotplug += nr_pages; 394 return BP_DONE; 395 } 396 #endif 397 398 if (nr_pages > ARRAY_SIZE(frame_list)) 399 nr_pages = ARRAY_SIZE(frame_list); 400 401 scratch_page = get_balloon_scratch_page(); 402 403 for (i = 0; i < nr_pages; i++) { 404 page = alloc_page(gfp); 405 if (page == NULL) { 406 nr_pages = i; 407 state = BP_EAGAIN; 408 break; 409 } 410 411 pfn = page_to_pfn(page); 412 frame_list[i] = pfn_to_mfn(pfn); 413 414 scrub_page(page); 415 416 #ifdef CONFIG_XEN_HAVE_PVMMU 417 if (xen_pv_domain() && !PageHighMem(page)) { 418 ret = HYPERVISOR_update_va_mapping( 419 (unsigned long)__va(pfn << PAGE_SHIFT), 420 pfn_pte(page_to_pfn(scratch_page), 421 PAGE_KERNEL_RO), 0); 422 BUG_ON(ret); 423 } 424 #endif 425 } 426 427 /* Ensure that ballooned highmem pages don't have kmaps. */ 428 kmap_flush_unused(); 429 flush_tlb_all(); 430 431 /* No more mappings: invalidate P2M and add to balloon. */ 432 for (i = 0; i < nr_pages; i++) { 433 pfn = mfn_to_pfn(frame_list[i]); 434 if (!xen_feature(XENFEAT_auto_translated_physmap)) { 435 unsigned long p; 436 p = page_to_pfn(scratch_page); 437 __set_phys_to_machine(pfn, pfn_to_mfn(p)); 438 } 439 balloon_append(pfn_to_page(pfn)); 440 } 441 442 put_balloon_scratch_page(); 443 444 set_xen_guest_handle(reservation.extent_start, frame_list); 445 reservation.nr_extents = nr_pages; 446 ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation); 447 BUG_ON(ret != nr_pages); 448 449 balloon_stats.current_pages -= nr_pages; 450 451 return state; 452 } 453 454 /* 455 * We avoid multiple worker processes conflicting via the balloon mutex. 456 * We may of course race updates of the target counts (which are protected 457 * by the balloon lock), or with changes to the Xen hard limit, but we will 458 * recover from these in time. 459 */ 460 static void balloon_process(struct work_struct *work) 461 { 462 enum bp_state state = BP_DONE; 463 long credit; 464 465 mutex_lock(&balloon_mutex); 466 467 do { 468 credit = current_credit(); 469 470 if (credit > 0) { 471 if (balloon_is_inflated()) 472 state = increase_reservation(credit); 473 else 474 state = reserve_additional_memory(credit); 475 } 476 477 if (credit < 0) 478 state = decrease_reservation(-credit, GFP_BALLOON); 479 480 state = update_schedule(state); 481 482 #ifndef CONFIG_PREEMPT 483 if (need_resched()) 484 schedule(); 485 #endif 486 } while (credit && state == BP_DONE); 487 488 /* Schedule more work if there is some still to be done. */ 489 if (state == BP_EAGAIN) 490 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ); 491 492 mutex_unlock(&balloon_mutex); 493 } 494 495 struct page *get_balloon_scratch_page(void) 496 { 497 struct page *ret = get_cpu_var(balloon_scratch_page); 498 BUG_ON(ret == NULL); 499 return ret; 500 } 501 502 void put_balloon_scratch_page(void) 503 { 504 put_cpu_var(balloon_scratch_page); 505 } 506 507 /* Resets the Xen limit, sets new target, and kicks off processing. */ 508 void balloon_set_new_target(unsigned long target) 509 { 510 /* No need for lock. Not read-modify-write updates. */ 511 balloon_stats.target_pages = target; 512 schedule_delayed_work(&balloon_worker, 0); 513 } 514 EXPORT_SYMBOL_GPL(balloon_set_new_target); 515 516 /** 517 * alloc_xenballooned_pages - get pages that have been ballooned out 518 * @nr_pages: Number of pages to get 519 * @pages: pages returned 520 * @highmem: allow highmem pages 521 * @return 0 on success, error otherwise 522 */ 523 int alloc_xenballooned_pages(int nr_pages, struct page **pages, bool highmem) 524 { 525 int pgno = 0; 526 struct page *page; 527 mutex_lock(&balloon_mutex); 528 while (pgno < nr_pages) { 529 page = balloon_retrieve(highmem); 530 if (page && (highmem || !PageHighMem(page))) { 531 pages[pgno++] = page; 532 } else { 533 enum bp_state st; 534 if (page) 535 balloon_append(page); 536 st = decrease_reservation(nr_pages - pgno, 537 highmem ? GFP_HIGHUSER : GFP_USER); 538 if (st != BP_DONE) 539 goto out_undo; 540 } 541 } 542 mutex_unlock(&balloon_mutex); 543 return 0; 544 out_undo: 545 while (pgno) 546 balloon_append(pages[--pgno]); 547 /* Free the memory back to the kernel soon */ 548 schedule_delayed_work(&balloon_worker, 0); 549 mutex_unlock(&balloon_mutex); 550 return -ENOMEM; 551 } 552 EXPORT_SYMBOL(alloc_xenballooned_pages); 553 554 /** 555 * free_xenballooned_pages - return pages retrieved with get_ballooned_pages 556 * @nr_pages: Number of pages 557 * @pages: pages to return 558 */ 559 void free_xenballooned_pages(int nr_pages, struct page **pages) 560 { 561 int i; 562 563 mutex_lock(&balloon_mutex); 564 565 for (i = 0; i < nr_pages; i++) { 566 if (pages[i]) 567 balloon_append(pages[i]); 568 } 569 570 /* The balloon may be too large now. Shrink it if needed. */ 571 if (current_credit()) 572 schedule_delayed_work(&balloon_worker, 0); 573 574 mutex_unlock(&balloon_mutex); 575 } 576 EXPORT_SYMBOL(free_xenballooned_pages); 577 578 static void __init balloon_add_region(unsigned long start_pfn, 579 unsigned long pages) 580 { 581 unsigned long pfn, extra_pfn_end; 582 struct page *page; 583 584 /* 585 * If the amount of usable memory has been limited (e.g., with 586 * the 'mem' command line parameter), don't add pages beyond 587 * this limit. 588 */ 589 extra_pfn_end = min(max_pfn, start_pfn + pages); 590 591 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++) { 592 page = pfn_to_page(pfn); 593 /* totalram_pages and totalhigh_pages do not 594 include the boot-time balloon extension, so 595 don't subtract from it. */ 596 __balloon_append(page); 597 } 598 } 599 600 static int __cpuinit balloon_cpu_notify(struct notifier_block *self, 601 unsigned long action, void *hcpu) 602 { 603 int cpu = (long)hcpu; 604 switch (action) { 605 case CPU_UP_PREPARE: 606 if (per_cpu(balloon_scratch_page, cpu) != NULL) 607 break; 608 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL); 609 if (per_cpu(balloon_scratch_page, cpu) == NULL) { 610 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu); 611 return NOTIFY_BAD; 612 } 613 break; 614 default: 615 break; 616 } 617 return NOTIFY_OK; 618 } 619 620 static struct notifier_block balloon_cpu_notifier __cpuinitdata = { 621 .notifier_call = balloon_cpu_notify, 622 }; 623 624 static int __init balloon_init(void) 625 { 626 int i, cpu; 627 628 if (!xen_domain()) 629 return -ENODEV; 630 631 for_each_online_cpu(cpu) 632 { 633 per_cpu(balloon_scratch_page, cpu) = alloc_page(GFP_KERNEL); 634 if (per_cpu(balloon_scratch_page, cpu) == NULL) { 635 pr_warn("Failed to allocate balloon_scratch_page for cpu %d\n", cpu); 636 return -ENOMEM; 637 } 638 } 639 register_cpu_notifier(&balloon_cpu_notifier); 640 641 pr_info("Initialising balloon driver\n"); 642 643 balloon_stats.current_pages = xen_pv_domain() 644 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn) 645 : max_pfn; 646 balloon_stats.target_pages = balloon_stats.current_pages; 647 balloon_stats.balloon_low = 0; 648 balloon_stats.balloon_high = 0; 649 650 balloon_stats.schedule_delay = 1; 651 balloon_stats.max_schedule_delay = 32; 652 balloon_stats.retry_count = 1; 653 balloon_stats.max_retry_count = RETRY_UNLIMITED; 654 655 #ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG 656 balloon_stats.hotplug_pages = 0; 657 balloon_stats.balloon_hotplug = 0; 658 659 set_online_page_callback(&xen_online_page); 660 register_memory_notifier(&xen_memory_nb); 661 #endif 662 663 /* 664 * Initialize the balloon with pages from the extra memory 665 * regions (see arch/x86/xen/setup.c). 666 */ 667 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) 668 if (xen_extra_mem[i].size) 669 balloon_add_region(PFN_UP(xen_extra_mem[i].start), 670 PFN_DOWN(xen_extra_mem[i].size)); 671 672 return 0; 673 } 674 675 subsys_initcall(balloon_init); 676 677 static int __init balloon_clear(void) 678 { 679 int cpu; 680 681 for_each_possible_cpu(cpu) 682 per_cpu(balloon_scratch_page, cpu) = NULL; 683 684 return 0; 685 } 686 early_initcall(balloon_clear); 687 688 MODULE_LICENSE("GPL"); 689