786 lines · c
1/******************************************************************************2 * Xen balloon driver - enables returning/claiming memory to/from Xen.3 *4 * Copyright (c) 2003, B Dragovic5 * Copyright (c) 2003-2004, M Williamson, K Fraser6 * Copyright (c) 2005 Dan M. Smith, IBM Corporation7 * Copyright (c) 2010 Daniel Kiper8 *9 * Memory hotplug support was written by Daniel Kiper. Work on10 * it was sponsored by Google under Google Summer of Code 201011 * program. Jeremy Fitzhardinge from Citrix was the mentor for12 * this project.13 *14 * This program is free software; you can redistribute it and/or15 * modify it under the terms of the GNU General Public License version 216 * as published by the Free Software Foundation; or, when distributed17 * separately from the Linux kernel or incorporated into other18 * software packages, subject to the following license:19 *20 * Permission is hereby granted, free of charge, to any person obtaining a copy21 * of this source file (the "Software"), to deal in the Software without22 * 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 to25 * the following conditions:26 *27 * The above copyright notice and this permission notice shall be included in28 * all copies or substantial portions of the Software.29 *30 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR31 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,32 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE33 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER34 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING35 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS36 * IN THE SOFTWARE.37 */38 39#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt40 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/freezer.h>47#include <linux/kthread.h>48#include <linux/mm.h>49#include <linux/memblock.h>50#include <linux/pagemap.h>51#include <linux/highmem.h>52#include <linux/mutex.h>53#include <linux/list.h>54#include <linux/gfp.h>55#include <linux/notifier.h>56#include <linux/memory.h>57#include <linux/memory_hotplug.h>58#include <linux/percpu-defs.h>59#include <linux/slab.h>60#include <linux/sysctl.h>61#include <linux/moduleparam.h>62#include <linux/jiffies.h>63 64#include <asm/page.h>65#include <asm/tlb.h>66 67#include <asm/xen/hypervisor.h>68#include <asm/xen/hypercall.h>69 70#include <xen/xen.h>71#include <xen/interface/xen.h>72#include <xen/interface/memory.h>73#include <xen/balloon.h>74#include <xen/features.h>75#include <xen/page.h>76#include <xen/mem-reservation.h>77 78#undef MODULE_PARAM_PREFIX79#define MODULE_PARAM_PREFIX "xen."80 81static uint __read_mostly balloon_boot_timeout = 180;82module_param(balloon_boot_timeout, uint, 0444);83 84#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG85static int xen_hotplug_unpopulated;86 87static struct ctl_table balloon_table[] = {88 {89 .procname = "hotplug_unpopulated",90 .data = &xen_hotplug_unpopulated,91 .maxlen = sizeof(int),92 .mode = 0644,93 .proc_handler = proc_dointvec_minmax,94 .extra1 = SYSCTL_ZERO,95 .extra2 = SYSCTL_ONE,96 },97};98 99#else100#define xen_hotplug_unpopulated 0101#endif102 103/*104 * Use one extent per PAGE_SIZE to avoid to break down the page into105 * multiple frame.106 */107#define EXTENT_ORDER (fls(XEN_PFN_PER_PAGE) - 1)108 109/*110 * balloon_thread() state:111 *112 * BP_DONE: done or nothing to do,113 * BP_WAIT: wait to be rescheduled,114 * BP_EAGAIN: error, go to sleep,115 * BP_ECANCELED: error, balloon operation canceled.116 */117 118static enum bp_state {119 BP_DONE,120 BP_WAIT,121 BP_EAGAIN,122 BP_ECANCELED123} balloon_state = BP_DONE;124 125/* Main waiting point for xen-balloon thread. */126static DECLARE_WAIT_QUEUE_HEAD(balloon_thread_wq);127 128static DEFINE_MUTEX(balloon_mutex);129 130struct balloon_stats balloon_stats;131EXPORT_SYMBOL_GPL(balloon_stats);132 133/* We increase/decrease in batches which fit in a page */134static xen_pfn_t frame_list[PAGE_SIZE / sizeof(xen_pfn_t)];135 136 137/* List of ballooned pages, threaded through the mem_map array. */138static LIST_HEAD(ballooned_pages);139static DECLARE_WAIT_QUEUE_HEAD(balloon_wq);140 141/* When ballooning out (allocating memory to return to Xen) we don't really142 want the kernel to try too hard since that can trigger the oom killer. */143#define GFP_BALLOON \144 (GFP_HIGHUSER | __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC)145 146/* balloon_append: add the given page to the balloon. */147static void balloon_append(struct page *page)148{149 if (!PageOffline(page))150 __SetPageOffline(page);151 152 /* Lowmem is re-populated first, so highmem pages go at list tail. */153 if (PageHighMem(page)) {154 list_add_tail(&page->lru, &ballooned_pages);155 balloon_stats.balloon_high++;156 } else {157 list_add(&page->lru, &ballooned_pages);158 balloon_stats.balloon_low++;159 }160 wake_up(&balloon_wq);161}162 163/* balloon_retrieve: rescue a page from the balloon, if it is not empty. */164static struct page *balloon_retrieve(bool require_lowmem)165{166 struct page *page;167 168 if (list_empty(&ballooned_pages))169 return NULL;170 171 page = list_entry(ballooned_pages.next, struct page, lru);172 if (require_lowmem && PageHighMem(page))173 return NULL;174 list_del(&page->lru);175 176 if (PageHighMem(page))177 balloon_stats.balloon_high--;178 else179 balloon_stats.balloon_low--;180 181 __ClearPageOffline(page);182 return page;183}184 185static struct page *balloon_next_page(struct page *page)186{187 struct list_head *next = page->lru.next;188 if (next == &ballooned_pages)189 return NULL;190 return list_entry(next, struct page, lru);191}192 193static void update_schedule(void)194{195 if (balloon_state == BP_WAIT || balloon_state == BP_ECANCELED)196 return;197 198 if (balloon_state == BP_DONE) {199 balloon_stats.schedule_delay = 1;200 balloon_stats.retry_count = 1;201 return;202 }203 204 ++balloon_stats.retry_count;205 206 if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&207 balloon_stats.retry_count > balloon_stats.max_retry_count) {208 balloon_stats.schedule_delay = 1;209 balloon_stats.retry_count = 1;210 balloon_state = BP_ECANCELED;211 return;212 }213 214 balloon_stats.schedule_delay <<= 1;215 216 if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)217 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;218 219 balloon_state = BP_EAGAIN;220}221 222#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG223static void release_memory_resource(struct resource *resource)224{225 if (!resource)226 return;227 228 /*229 * No need to reset region to identity mapped since we now230 * know that no I/O can be in this region231 */232 release_resource(resource);233 kfree(resource);234}235 236static struct resource *additional_memory_resource(phys_addr_t size)237{238 struct resource *res;239 int ret;240 241 res = kzalloc(sizeof(*res), GFP_KERNEL);242 if (!res)243 return NULL;244 245 res->name = "System RAM";246 res->flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;247 248 ret = allocate_resource(&iomem_resource, res,249 size, 0, -1,250 PAGES_PER_SECTION * PAGE_SIZE, NULL, NULL);251 if (ret < 0) {252 pr_err("Cannot allocate new System RAM resource\n");253 kfree(res);254 return NULL;255 }256 257 return res;258}259 260static enum bp_state reserve_additional_memory(void)261{262 long credit;263 struct resource *resource;264 int nid, rc;265 unsigned long balloon_hotplug;266 267 credit = balloon_stats.target_pages + balloon_stats.target_unpopulated268 - balloon_stats.total_pages;269 270 /*271 * Already hotplugged enough pages? Wait for them to be272 * onlined.273 */274 if (credit <= 0)275 return BP_WAIT;276 277 balloon_hotplug = round_up(credit, PAGES_PER_SECTION);278 279 resource = additional_memory_resource(balloon_hotplug * PAGE_SIZE);280 if (!resource)281 goto err;282 283 nid = memory_add_physaddr_to_nid(resource->start);284 285#ifdef CONFIG_XEN_HAVE_PVMMU286 /*287 * We don't support PV MMU when Linux and Xen is using288 * different page granularity.289 */290 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);291 292 /*293 * add_memory() will build page tables for the new memory so294 * the p2m must contain invalid entries so the correct295 * non-present PTEs will be written.296 *297 * If a failure occurs, the original (identity) p2m entries298 * are not restored since this region is now known not to299 * conflict with any devices.300 */ 301 if (!xen_feature(XENFEAT_auto_translated_physmap)) {302 unsigned long pfn, i;303 304 pfn = PFN_DOWN(resource->start);305 for (i = 0; i < balloon_hotplug; i++) {306 if (!set_phys_to_machine(pfn + i, INVALID_P2M_ENTRY)) {307 pr_warn("set_phys_to_machine() failed, no memory added\n");308 goto err;309 }310 }311 }312#endif313 314 /*315 * add_memory_resource() will call online_pages() which in its turn316 * will call xen_online_page() callback causing deadlock if we don't317 * release balloon_mutex here. Unlocking here is safe because the318 * callers drop the mutex before trying again.319 */320 mutex_unlock(&balloon_mutex);321 /* add_memory_resource() requires the device_hotplug lock */322 lock_device_hotplug();323 rc = add_memory_resource(nid, resource, MHP_MERGE_RESOURCE);324 unlock_device_hotplug();325 mutex_lock(&balloon_mutex);326 327 if (rc) {328 pr_warn("Cannot add additional memory (%i)\n", rc);329 goto err;330 }331 332 balloon_stats.total_pages += balloon_hotplug;333 334 return BP_WAIT;335 err:336 release_memory_resource(resource);337 return BP_ECANCELED;338}339 340static void xen_online_page(struct page *page, unsigned int order)341{342 unsigned long i, size = (1 << order);343 unsigned long start_pfn = page_to_pfn(page);344 struct page *p;345 346 pr_debug("Online %lu pages starting at pfn 0x%lx\n", size, start_pfn);347 mutex_lock(&balloon_mutex);348 for (i = 0; i < size; i++) {349 p = pfn_to_page(start_pfn + i);350 balloon_append(p);351 }352 mutex_unlock(&balloon_mutex);353}354 355static int xen_memory_notifier(struct notifier_block *nb, unsigned long val, void *v)356{357 if (val == MEM_ONLINE)358 wake_up(&balloon_thread_wq);359 360 return NOTIFY_OK;361}362 363static struct notifier_block xen_memory_nb = {364 .notifier_call = xen_memory_notifier,365 .priority = 0366};367#else368static enum bp_state reserve_additional_memory(void)369{370 balloon_stats.target_pages = balloon_stats.current_pages +371 balloon_stats.target_unpopulated;372 return BP_ECANCELED;373}374#endif /* CONFIG_XEN_BALLOON_MEMORY_HOTPLUG */375 376static long current_credit(void)377{378 return balloon_stats.target_pages - balloon_stats.current_pages;379}380 381static bool balloon_is_inflated(void)382{383 return balloon_stats.balloon_low || balloon_stats.balloon_high;384}385 386static enum bp_state increase_reservation(unsigned long nr_pages)387{388 int rc;389 unsigned long i;390 struct page *page;391 392 if (nr_pages > ARRAY_SIZE(frame_list))393 nr_pages = ARRAY_SIZE(frame_list);394 395 page = list_first_entry_or_null(&ballooned_pages, struct page, lru);396 for (i = 0; i < nr_pages; i++) {397 if (!page) {398 nr_pages = i;399 break;400 }401 402 frame_list[i] = page_to_xen_pfn(page);403 page = balloon_next_page(page);404 }405 406 rc = xenmem_reservation_increase(nr_pages, frame_list);407 if (rc <= 0)408 return BP_EAGAIN;409 410 for (i = 0; i < rc; i++) {411 page = balloon_retrieve(false);412 BUG_ON(page == NULL);413 414 xenmem_reservation_va_mapping_update(1, &page, &frame_list[i]);415 416 /*417 * Relinquish the page back to the allocator. Note that418 * some pages, including ones added via xen_online_page(), might419 * not be marked reserved; free_reserved_page() will handle that.420 */421 free_reserved_page(page);422 }423 424 balloon_stats.current_pages += rc;425 426 return BP_DONE;427}428 429static enum bp_state decrease_reservation(unsigned long nr_pages, gfp_t gfp)430{431 enum bp_state state = BP_DONE;432 unsigned long i;433 struct page *page, *tmp;434 int ret;435 LIST_HEAD(pages);436 437 if (nr_pages > ARRAY_SIZE(frame_list))438 nr_pages = ARRAY_SIZE(frame_list);439 440 for (i = 0; i < nr_pages; i++) {441 page = alloc_page(gfp);442 if (page == NULL) {443 nr_pages = i;444 state = BP_EAGAIN;445 break;446 }447 adjust_managed_page_count(page, -1);448 xenmem_reservation_scrub_page(page);449 list_add(&page->lru, &pages);450 }451 452 /*453 * Ensure that ballooned highmem pages don't have kmaps.454 *455 * Do this before changing the p2m as kmap_flush_unused()456 * reads PTEs to obtain pages (and hence needs the original457 * p2m entry).458 */459 kmap_flush_unused();460 461 /*462 * Setup the frame, update direct mapping, invalidate P2M,463 * and add to balloon.464 */465 i = 0;466 list_for_each_entry_safe(page, tmp, &pages, lru) {467 frame_list[i++] = xen_page_to_gfn(page);468 469 xenmem_reservation_va_mapping_reset(1, &page);470 471 list_del(&page->lru);472 473 balloon_append(page);474 }475 476 flush_tlb_all();477 478 ret = xenmem_reservation_decrease(nr_pages, frame_list);479 BUG_ON(ret != nr_pages);480 481 balloon_stats.current_pages -= nr_pages;482 483 return state;484}485 486/*487 * Stop waiting if either state is BP_DONE and ballooning action is488 * needed, or if the credit has changed while state is not BP_DONE.489 */490static bool balloon_thread_cond(long credit)491{492 if (balloon_state == BP_DONE)493 credit = 0;494 495 return current_credit() != credit || kthread_should_stop();496}497 498/*499 * As this is a kthread it is guaranteed to run as a single instance only.500 * We may of course race updates of the target counts (which are protected501 * by the balloon lock), or with changes to the Xen hard limit, but we will502 * recover from these in time.503 */504static int balloon_thread(void *unused)505{506 long credit;507 unsigned long timeout;508 509 set_freezable();510 for (;;) {511 switch (balloon_state) {512 case BP_DONE:513 case BP_ECANCELED:514 timeout = 3600 * HZ;515 break;516 case BP_EAGAIN:517 timeout = balloon_stats.schedule_delay * HZ;518 break;519 case BP_WAIT:520 timeout = HZ;521 break;522 }523 524 credit = current_credit();525 526 wait_event_freezable_timeout(balloon_thread_wq,527 balloon_thread_cond(credit), timeout);528 529 if (kthread_should_stop())530 return 0;531 532 mutex_lock(&balloon_mutex);533 534 credit = current_credit();535 536 if (credit > 0) {537 if (balloon_is_inflated())538 balloon_state = increase_reservation(credit);539 else540 balloon_state = reserve_additional_memory();541 }542 543 if (credit < 0) {544 long n_pages;545 546 n_pages = min(-credit, si_mem_available());547 balloon_state = decrease_reservation(n_pages,548 GFP_BALLOON);549 if (balloon_state == BP_DONE && n_pages != -credit &&550 n_pages < totalreserve_pages)551 balloon_state = BP_EAGAIN;552 }553 554 update_schedule();555 556 mutex_unlock(&balloon_mutex);557 558 cond_resched();559 }560}561 562/* Resets the Xen limit, sets new target, and kicks off processing. */563void balloon_set_new_target(unsigned long target)564{565 /* No need for lock. Not read-modify-write updates. */566 balloon_stats.target_pages = target;567 wake_up(&balloon_thread_wq);568}569EXPORT_SYMBOL_GPL(balloon_set_new_target);570 571static int add_ballooned_pages(unsigned int nr_pages)572{573 enum bp_state st;574 575 if (xen_hotplug_unpopulated) {576 st = reserve_additional_memory();577 if (st != BP_ECANCELED) {578 int rc;579 580 mutex_unlock(&balloon_mutex);581 rc = wait_event_interruptible(balloon_wq,582 !list_empty(&ballooned_pages));583 mutex_lock(&balloon_mutex);584 return rc ? -ENOMEM : 0;585 }586 }587 588 if (si_mem_available() < nr_pages)589 return -ENOMEM;590 591 st = decrease_reservation(nr_pages, GFP_USER);592 if (st != BP_DONE)593 return -ENOMEM;594 595 return 0;596}597 598/**599 * xen_alloc_ballooned_pages - get pages that have been ballooned out600 * @nr_pages: Number of pages to get601 * @pages: pages returned602 * @return 0 on success, error otherwise603 */604int xen_alloc_ballooned_pages(unsigned int nr_pages, struct page **pages)605{606 unsigned int pgno = 0;607 struct page *page;608 int ret;609 610 mutex_lock(&balloon_mutex);611 612 balloon_stats.target_unpopulated += nr_pages;613 614 while (pgno < nr_pages) {615 page = balloon_retrieve(true);616 if (page) {617 pages[pgno++] = page;618#ifdef CONFIG_XEN_HAVE_PVMMU619 /*620 * We don't support PV MMU when Linux and Xen is using621 * different page granularity.622 */623 BUILD_BUG_ON(XEN_PAGE_SIZE != PAGE_SIZE);624 625 if (!xen_feature(XENFEAT_auto_translated_physmap)) {626 ret = xen_alloc_p2m_entry(page_to_pfn(page));627 if (ret < 0)628 goto out_undo;629 }630#endif631 } else {632 ret = add_ballooned_pages(nr_pages - pgno);633 if (ret < 0)634 goto out_undo;635 }636 }637 mutex_unlock(&balloon_mutex);638 return 0;639 out_undo:640 mutex_unlock(&balloon_mutex);641 xen_free_ballooned_pages(pgno, pages);642 /*643 * NB: xen_free_ballooned_pages will only subtract pgno pages, but since644 * target_unpopulated is incremented with nr_pages at the start we need645 * to remove the remaining ones also, or accounting will be screwed.646 */647 balloon_stats.target_unpopulated -= nr_pages - pgno;648 return ret;649}650EXPORT_SYMBOL(xen_alloc_ballooned_pages);651 652/**653 * xen_free_ballooned_pages - return pages retrieved with get_ballooned_pages654 * @nr_pages: Number of pages655 * @pages: pages to return656 */657void xen_free_ballooned_pages(unsigned int nr_pages, struct page **pages)658{659 unsigned int i;660 661 mutex_lock(&balloon_mutex);662 663 for (i = 0; i < nr_pages; i++) {664 if (pages[i])665 balloon_append(pages[i]);666 }667 668 balloon_stats.target_unpopulated -= nr_pages;669 670 /* The balloon may be too large now. Shrink it if needed. */671 if (current_credit())672 wake_up(&balloon_thread_wq);673 674 mutex_unlock(&balloon_mutex);675}676EXPORT_SYMBOL(xen_free_ballooned_pages);677 678static void __init balloon_add_regions(void)679{680 unsigned long start_pfn, pages;681 unsigned long pfn, extra_pfn_end;682 unsigned int i;683 684 for (i = 0; i < XEN_EXTRA_MEM_MAX_REGIONS; i++) {685 pages = xen_extra_mem[i].n_pfns;686 if (!pages)687 continue;688 689 start_pfn = xen_extra_mem[i].start_pfn;690 691 /*692 * If the amount of usable memory has been limited (e.g., with693 * the 'mem' command line parameter), don't add pages beyond694 * this limit.695 */696 extra_pfn_end = min(max_pfn, start_pfn + pages);697 698 for (pfn = start_pfn; pfn < extra_pfn_end; pfn++)699 balloon_append(pfn_to_page(pfn));700 701 balloon_stats.total_pages += extra_pfn_end - start_pfn;702 }703}704 705static int __init balloon_init(void)706{707 struct task_struct *task;708 709 if (!xen_domain())710 return -ENODEV;711 712 pr_info("Initialising balloon driver\n");713 714#ifdef CONFIG_XEN_PV715 balloon_stats.current_pages = xen_pv_domain()716 ? min(xen_start_info->nr_pages - xen_released_pages, max_pfn)717 : get_num_physpages();718#else719 balloon_stats.current_pages = get_num_physpages();720#endif721 balloon_stats.target_pages = balloon_stats.current_pages;722 balloon_stats.balloon_low = 0;723 balloon_stats.balloon_high = 0;724 balloon_stats.total_pages = balloon_stats.current_pages;725 726 balloon_stats.schedule_delay = 1;727 balloon_stats.max_schedule_delay = 32;728 balloon_stats.retry_count = 1;729 balloon_stats.max_retry_count = 4;730 731#ifdef CONFIG_XEN_BALLOON_MEMORY_HOTPLUG732 set_online_page_callback(&xen_online_page);733 register_memory_notifier(&xen_memory_nb);734 register_sysctl_init("xen/balloon", balloon_table);735#endif736 737 balloon_add_regions();738 739 task = kthread_run(balloon_thread, NULL, "xen-balloon");740 if (IS_ERR(task)) {741 pr_err("xen-balloon thread could not be started, ballooning will not work!\n");742 return PTR_ERR(task);743 }744 745 /* Init the xen-balloon driver. */746 xen_balloon_init();747 748 return 0;749}750subsys_initcall(balloon_init);751 752static int __init balloon_wait_finish(void)753{754 long credit, last_credit = 0;755 unsigned long last_changed = 0;756 757 if (!xen_domain())758 return -ENODEV;759 760 /* PV guests don't need to wait. */761 if (xen_pv_domain() || !current_credit())762 return 0;763 764 pr_notice("Waiting for initial ballooning down having finished.\n");765 766 while ((credit = current_credit()) < 0) {767 if (credit != last_credit) {768 last_changed = jiffies;769 last_credit = credit;770 }771 if (balloon_state == BP_ECANCELED) {772 pr_warn_once("Initial ballooning failed, %ld pages need to be freed.\n",773 -credit);774 if (time_is_before_eq_jiffies(last_changed + HZ * balloon_boot_timeout))775 panic("Initial ballooning failed!\n");776 }777 778 schedule_timeout_interruptible(HZ / 10);779 }780 781 pr_notice("Initial ballooning down finished.\n");782 783 return 0;784}785late_initcall_sync(balloon_wait_finish);786