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(void)
 132 {
 133         struct page *page;
 134
 135         if (list_empty(&ballooned_pages))
 136                 return NULL;
 137
 138         page = list_entry(ballooned_pages.next, struct page, lru);
 139         list_del(&page->lru);
 140
 141         if (PageHighMem(page)) {
 142                 balloon_stats.balloon_high--;
 143                 inc_totalhigh_pages();
 144         }
 145         else
 146                 balloon_stats.balloon_low--;
 147
 148         totalram_pages++;
 149
 150         return page;
 151 }
 152
 153 static struct page *balloon_first_page(void)
 154 {
 155         if (list_empty(&ballooned_pages))
 156                 return NULL;
 157         return list_entry(ballooned_pages.next, struct page, lru);
 158 }
 159
 160 static struct page *balloon_next_page(struct page *page)
 161 {
 162         struct list_head *next = page->lru.next;
 163         if (next == &ballooned_pages)
 164                 return NULL;
 165         return list_entry(next, struct page, lru);
 166 }
 167
 168 static enum bp_state update_schedule(enum bp_state state)
 169 {
 170         if (state == BP_DONE) {
 171                 balloon_stats.schedule_delay = 1;
 172                 balloon_stats.retry_count = 1;
 173                 return BP_DONE;
 174         }
 175
 176         ++balloon_stats.retry_count;
 177
 178         if (balloon_stats.max_retry_count != RETRY_UNLIMITED &&
 179                         balloon_stats.retry_count > balloon_stats.max_retry_count) {
 180                 balloon_stats.schedule_delay = 1;
 181                 balloon_stats.retry_count = 1;
 182                 return BP_ECANCELED;
 183         }
 184
 185         balloon_stats.schedule_delay <<= 1;
 186
 187         if (balloon_stats.schedule_delay > balloon_stats.max_schedule_delay)
 188                 balloon_stats.schedule_delay = balloon_stats.max_schedule_delay;
 189
 190         return BP_EAGAIN;
 191 }
 192
 193 static unsigned long current_target(void)
 194 {
 195         unsigned long target = balloon_stats.target_pages;
 196
 197         target = min(target,
 198                      balloon_stats.current_pages +
 199                      balloon_stats.balloon_low +
 200                      balloon_stats.balloon_high);
 201
 202         return target;
 203 }
 204
 205 static enum bp_state increase_reservation(unsigned long nr_pages)
 206 {
 207         int rc;
 208         unsigned long  pfn, i;
 209         struct page   *page;
 210         struct xen_memory_reservation reservation = {
 211                 .address_bits = 0,
 212                 .extent_order = 0,
 213                 .domid        = DOMID_SELF
 214         };
 215
 216         if (nr_pages > ARRAY_SIZE(frame_list))
 217                 nr_pages = ARRAY_SIZE(frame_list);
 218
 219         page = balloon_first_page();
 220         for (i = 0; i < nr_pages; i++) {
 221                 if (!page) {
 222                         nr_pages = i;
 223                         break;
 224                 }
 225                 frame_list[i] = page_to_pfn(page);
 226                 page = balloon_next_page(page);
 227         }
 228
 229         set_xen_guest_handle(reservation.extent_start, frame_list);
 230         reservation.nr_extents = nr_pages;
 231         rc = HYPERVISOR_memory_op(XENMEM_populate_physmap, &reservation);
 232         if (rc <= 0)
 233                 return BP_EAGAIN;
 234
 235         for (i = 0; i < rc; i++) {
 236                 page = balloon_retrieve();
 237                 BUG_ON(page == NULL);
 238
 239                 pfn = page_to_pfn(page);
 240                 BUG_ON(!xen_feature(XENFEAT_auto_translated_physmap) &&
 241                        phys_to_machine_mapping_valid(pfn));
 242
 243                 set_phys_to_machine(pfn, frame_list[i]);
 244
 245                 /* Link back into the page tables if not highmem. */
 246                 if (pfn < max_low_pfn) {
 247                         int ret;
 248                         ret = HYPERVISOR_update_va_mapping(
 249                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 250                                 mfn_pte(frame_list[i], PAGE_KERNEL),
 251                                 0);
 252                         BUG_ON(ret);
 253                 }
 254
 255                 /* Relinquish the page back to the allocator. */
 256                 ClearPageReserved(page);
 257                 init_page_count(page);
 258                 __free_page(page);
 259         }
 260
 261         balloon_stats.current_pages += rc;
 262
 263         return BP_DONE;
 264 }
 265
 266 static enum bp_state decrease_reservation(unsigned long nr_pages)
 267 {
 268         enum bp_state state = BP_DONE;
 269         unsigned long  pfn, i;
 270         struct page   *page;
 271         int ret;
 272         struct xen_memory_reservation reservation = {
 273                 .address_bits = 0,
 274                 .extent_order = 0,
 275                 .domid        = DOMID_SELF
 276         };
 277
 278         if (nr_pages > ARRAY_SIZE(frame_list))
 279                 nr_pages = ARRAY_SIZE(frame_list);
 280
 281         for (i = 0; i < nr_pages; i++) {
 282                 if ((page = alloc_page(GFP_BALLOON)) == NULL) {
 283                         nr_pages = i;
 284                         state = BP_EAGAIN;
 285                         break;
 286                 }
 287
 288                 pfn = page_to_pfn(page);
 289                 frame_list[i] = pfn_to_mfn(pfn);
 290
 291                 scrub_page(page);
 292
 293                 if (!PageHighMem(page)) {
 294                         ret = HYPERVISOR_update_va_mapping(
 295                                 (unsigned long)__va(pfn << PAGE_SHIFT),
 296                                 __pte_ma(0), 0);
 297                         BUG_ON(ret);
 298                 }
 299
 300         }
 301
 302         /* Ensure that ballooned highmem pages don't have kmaps. */
 303         kmap_flush_unused();
 304         flush_tlb_all();
 305
 306         /* No more mappings: invalidate P2M and add to balloon. */
 307         for (i = 0; i < nr_pages; i++) {
 308                 pfn = mfn_to_pfn(frame_list[i]);
 309                 set_phys_to_machine(pfn, INVALID_P2M_ENTRY);
 310                 balloon_append(pfn_to_page(pfn));
 311         }
 312
 313         set_xen_guest_handle(reservation.extent_start, frame_list);
 314         reservation.nr_extents   = nr_pages;
 315         ret = HYPERVISOR_memory_op(XENMEM_decrease_reservation, &reservation);
 316         BUG_ON(ret != nr_pages);
 317
 318         balloon_stats.current_pages -= nr_pages;
 319
 320         return state;
 321 }
 322
 323 /*
 324  * We avoid multiple worker processes conflicting via the balloon mutex.
 325  * We may of course race updates of the target counts (which are protected
 326  * by the balloon lock), or with changes to the Xen hard limit, but we will
 327  * recover from these in time.
 328  */
 329 static void balloon_process(struct work_struct *work)
 330 {
 331         enum bp_state state = BP_DONE;
 332         long credit;
 333
 334         mutex_lock(&balloon_mutex);
 335
 336         do {
 337                 credit = current_target() - balloon_stats.current_pages;
 338
 339                 if (credit > 0)
 340                         state = increase_reservation(credit);
 341
 342                 if (credit < 0)
 343                         state = decrease_reservation(-credit);
 344
 345                 state = update_schedule(state);
 346
 347 #ifndef CONFIG_PREEMPT
 348                 if (need_resched())
 349                         schedule();
 350 #endif
 351         } while (credit && state == BP_DONE);
 352
 353         /* Schedule more work if there is some still to be done. */
 354         if (state == BP_EAGAIN)
 355                 schedule_delayed_work(&balloon_worker, balloon_stats.schedule_delay * HZ);
 356
 357         mutex_unlock(&balloon_mutex);
 358 }
 359
 360 /* Resets the Xen limit, sets new target, and kicks off processing. */
 361 void balloon_set_new_target(unsigned long target)
 362 {
 363         /* No need for lock. Not read-modify-write updates. */
 364         balloon_stats.target_pages = target;
 365         schedule_delayed_work(&balloon_worker, 0);
 366 }
 367 EXPORT_SYMBOL_GPL(balloon_set_new_target);
 368
 369 static int __init balloon_init(void)
 370 {
 371         unsigned long pfn, extra_pfn_end;
 372         struct page *page;
 373
 374         if (!xen_pv_domain())
 375                 return -ENODEV;
 376
 377         pr_info("xen/balloon: Initialising balloon driver.\n");
 378
 379         balloon_stats.current_pages = min(xen_start_info->nr_pages, max_pfn);
 380         balloon_stats.target_pages  = balloon_stats.current_pages;
 381         balloon_stats.balloon_low   = 0;
 382         balloon_stats.balloon_high  = 0;
 383
 384         balloon_stats.schedule_delay = 1;
 385         balloon_stats.max_schedule_delay = 32;
 386         balloon_stats.retry_count = 1;
 387         balloon_stats.max_retry_count = RETRY_UNLIMITED;
 388
 389         /*
 390          * Initialise the balloon with excess memory space.  We need
 391          * to make sure we don't add memory which doesn't exist or
 392          * logically exist.  The E820 map can be trimmed to be smaller
 393          * than the amount of physical memory due to the mem= command
 394          * line parameter.  And if this is a 32-bit non-HIGHMEM kernel
 395          * on a system with memory which requires highmem to access,
 396          * don't try to use it.
 397          */
 398         extra_pfn_end = min(min(max_pfn, e820_end_of_ram_pfn()),
 399                             (unsigned long)PFN_DOWN(xen_extra_mem_start + xen_extra_mem_size));
 400         for (pfn = PFN_UP(xen_extra_mem_start);
 401              pfn < extra_pfn_end;
 402              pfn++) {
 403                 page = pfn_to_page(pfn);
 404                 /* totalram_pages doesn't include the boot-time
 405                    balloon extension, so don't subtract from it. */
 406                 __balloon_append(page);
 407         }
 408
 409         return 0;
 410 }
 411
 412 subsys_initcall(balloon_init);
 413
 414 MODULE_LICENSE("GPL");