drivers/misc/vmw_balloon.c

   1 /*
   2  * VMware Balloon driver.
   3  *
   4  * Copyright (C) 2000-2013, VMware, Inc. All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License as published by the
   8  * Free Software Foundation; version 2 of the License and no later version.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  13  * NON INFRINGEMENT.  See the GNU General Public License for more
  14  * details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Maintained by:       Xavier Deguillard <xdeguillard@vmware.com>
  21  *                      Philip Moltmann <moltmann@vmware.com>
  22  */
  23
  24 /*
  25  * This is VMware physical memory management driver for Linux. The driver
  26  * acts like a "balloon" that can be inflated to reclaim physical pages by
  27  * reserving them in the guest and invalidating them in the monitor,
  28  * freeing up the underlying machine pages so they can be allocated to
  29  * other guests.  The balloon can also be deflated to allow the guest to
  30  * use more physical memory. Higher level policies can control the sizes
  31  * of balloons in VMs in order to manage physical memory resources.
  32  */
  33
  34 //#define DEBUG
  35 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  36
  37 #include <linux/types.h>
  38 #include <linux/kernel.h>
  39 #include <linux/mm.h>
  40 #include <linux/vmalloc.h>
  41 #include <linux/sched.h>
  42 #include <linux/module.h>
  43 #include <linux/workqueue.h>
  44 #include <linux/debugfs.h>
  45 #include <linux/seq_file.h>
  46 #include <asm/hypervisor.h>
  47
  48 MODULE_AUTHOR("VMware, Inc.");
  49 MODULE_DESCRIPTION("VMware Memory Control (Balloon) Driver");
  50 MODULE_VERSION("1.3.4.0-k");
  51 MODULE_ALIAS("dmi:*:svnVMware*:*");
  52 MODULE_ALIAS("vmware_vmmemctl");
  53 MODULE_LICENSE("GPL");
  54
  55 /*
  56  * Various constants controlling rate of inflaint/deflating balloon,
  57  * measured in pages.
  58  */
  59
  60 /*
  61  * Rates of memory allocaton when guest experiences memory pressure
  62  * (driver performs sleeping allocations).
  63  */
  64 #define VMW_BALLOON_RATE_ALLOC_MIN      512U
  65 #define VMW_BALLOON_RATE_ALLOC_MAX      2048U
  66 #define VMW_BALLOON_RATE_ALLOC_INC      16U
  67
  68 /*
  69  * When guest is under memory pressure, use a reduced page allocation
  70  * rate for next several cycles.
  71  */
  72 #define VMW_BALLOON_SLOW_CYCLES         4
  73
  74 /*
  75  * Use __GFP_HIGHMEM to allow pages from HIGHMEM zone. We don't
  76  * allow wait (__GFP_WAIT) for NOSLEEP page allocations. Use
  77  * __GFP_NOWARN, to suppress page allocation failure warnings.
  78  */
  79 #define VMW_PAGE_ALLOC_NOSLEEP          (__GFP_HIGHMEM|__GFP_NOWARN)
  80
  81 /*
  82  * Use GFP_HIGHUSER when executing in a separate kernel thread
  83  * context and allocation can sleep.  This is less stressful to
  84  * the guest memory system, since it allows the thread to block
  85  * while memory is reclaimed, and won't take pages from emergency
  86  * low-memory pools.
  87  */
  88 #define VMW_PAGE_ALLOC_CANSLEEP         (GFP_HIGHUSER)
  89
  90 /* Maximum number of refused pages we accumulate during inflation cycle */
  91 #define VMW_BALLOON_MAX_REFUSED         16
  92
  93 /*
  94  * Hypervisor communication port definitions.
  95  */
  96 #define VMW_BALLOON_HV_PORT             0x5670
  97 #define VMW_BALLOON_HV_MAGIC            0x456c6d6f
  98 #define VMW_BALLOON_GUEST_ID            1       /* Linux */
  99
 100 enum vmwballoon_capabilities {
 101         /*
 102          * Bit 0 is reserved and not associated to any capability.
 103          */
 104         VMW_BALLOON_BASIC_CMDS          = (1 << 1),
 105         VMW_BALLOON_BATCHED_CMDS        = (1 << 2)
 106 };
 107
 108 #define VMW_BALLOON_CAPABILITIES        (VMW_BALLOON_BASIC_CMDS \
 109                                         | VMW_BALLOON_BATCHED_CMDS)
 110
 111 /*
 112  * Backdoor commands availability:
 113  *
 114  * START, GET_TARGET and GUEST_ID are always available,
 115  *
 116  * VMW_BALLOON_BASIC_CMDS:
 117  *      LOCK and UNLOCK commands,
 118  * VMW_BALLOON_BATCHED_CMDS:
 119  *      BATCHED_LOCK and BATCHED_UNLOCK commands.
 120  */
 121 #define VMW_BALLOON_CMD_START           0
 122 #define VMW_BALLOON_CMD_GET_TARGET      1
 123 #define VMW_BALLOON_CMD_LOCK            2
 124 #define VMW_BALLOON_CMD_UNLOCK          3
 125 #define VMW_BALLOON_CMD_GUEST_ID        4
 126 #define VMW_BALLOON_CMD_BATCHED_LOCK    6
 127 #define VMW_BALLOON_CMD_BATCHED_UNLOCK  7
 128
 129 /* error codes */
 130 #define VMW_BALLOON_SUCCESS                     0
 131 #define VMW_BALLOON_FAILURE                     -1
 132 #define VMW_BALLOON_ERROR_CMD_INVALID           1
 133 #define VMW_BALLOON_ERROR_PPN_INVALID           2
 134 #define VMW_BALLOON_ERROR_PPN_LOCKED            3
 135 #define VMW_BALLOON_ERROR_PPN_UNLOCKED          4
 136 #define VMW_BALLOON_ERROR_PPN_PINNED            5
 137 #define VMW_BALLOON_ERROR_PPN_NOTNEEDED         6
 138 #define VMW_BALLOON_ERROR_RESET                 7
 139 #define VMW_BALLOON_ERROR_BUSY                  8
 140
 141 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES   (0x03000000)
 142
 143 /* Batch page description */
 144
 145 /*
 146  * Layout of a page in the batch page:
 147  *
 148  * +-------------+----------+--------+
 149  * |             |          |        |
 150  * | Page number | Reserved | Status |
 151  * |             |          |        |
 152  * +-------------+----------+--------+
 153  * 64  PAGE_SHIFT          6         0
 154  *
 155  * For now only 4K pages are supported, but we can easily support large pages
 156  * by using bits in the reserved field.
 157  *
 158  * The reserved field should be set to 0.
 159  */
 160 #define VMW_BALLOON_BATCH_MAX_PAGES     (PAGE_SIZE / sizeof(u64))
 161 #define VMW_BALLOON_BATCH_STATUS_MASK   ((1UL << 5) - 1)
 162 #define VMW_BALLOON_BATCH_PAGE_MASK     (~((1UL << PAGE_SHIFT) - 1))
 163
 164 struct vmballoon_batch_page {
 165         u64 pages[VMW_BALLOON_BATCH_MAX_PAGES];
 166 };
 167
 168 static u64 vmballoon_batch_get_pa(struct vmballoon_batch_page *batch, int idx)
 169 {
 170         return batch->pages[idx] & VMW_BALLOON_BATCH_PAGE_MASK;
 171 }
 172
 173 static int vmballoon_batch_get_status(struct vmballoon_batch_page *batch,
 174                                 int idx)
 175 {
 176         return (int)(batch->pages[idx] & VMW_BALLOON_BATCH_STATUS_MASK);
 177 }
 178
 179 static void vmballoon_batch_set_pa(struct vmballoon_batch_page *batch, int idx,
 180                                 u64 pa)
 181 {
 182         batch->pages[idx] = pa;
 183 }
 184
 185
 186 #define VMWARE_BALLOON_CMD(cmd, arg1, arg2, result)             \
 187 ({                                                              \
 188         unsigned long __status, __dummy1, __dummy2, __dummy3;   \
 189         __asm__ __volatile__ ("inl %%dx" :                      \
 190                 "=a"(__status),                                 \
 191                 "=c"(__dummy1),                                 \
 192                 "=d"(__dummy2),                                 \
 193                 "=b"(result),                                   \
 194                 "=S" (__dummy3) :                               \
 195                 "0"(VMW_BALLOON_HV_MAGIC),                      \
 196                 "1"(VMW_BALLOON_CMD_##cmd),                     \
 197                 "2"(VMW_BALLOON_HV_PORT),                       \
 198                 "3"(arg1),                                      \
 199                 "4" (arg2) :                                    \
 200                 "memory");                                      \
 201         if (VMW_BALLOON_CMD_##cmd == VMW_BALLOON_CMD_START)     \
 202                 result = __dummy1;                              \
 203         result &= -1UL;                                         \
 204         __status & -1UL;                                        \
 205 })
 206
 207 #ifdef CONFIG_DEBUG_FS
 208 struct vmballoon_stats {
 209         unsigned int timer;
 210
 211         /* allocation statistics */
 212         unsigned int alloc;
 213         unsigned int alloc_fail;
 214         unsigned int sleep_alloc;
 215         unsigned int sleep_alloc_fail;
 216         unsigned int refused_alloc;
 217         unsigned int refused_free;
 218         unsigned int free;
 219
 220         /* monitor operations */
 221         unsigned int lock;
 222         unsigned int lock_fail;
 223         unsigned int unlock;
 224         unsigned int unlock_fail;
 225         unsigned int target;
 226         unsigned int target_fail;
 227         unsigned int start;
 228         unsigned int start_fail;
 229         unsigned int guest_type;
 230         unsigned int guest_type_fail;
 231 };
 232
 233 #define STATS_INC(stat) (stat)++
 234 #else
 235 #define STATS_INC(stat)
 236 #endif
 237
 238 struct vmballoon;
 239
 240 struct vmballoon_ops {
 241         void (*add_page)(struct vmballoon *b, int idx, struct page *p);
 242         int (*lock)(struct vmballoon *b, unsigned int num_pages,
 243                                                 unsigned int *target);
 244         int (*unlock)(struct vmballoon *b, unsigned int num_pages,
 245                                                 unsigned int *target);
 246 };
 247
 248 struct vmballoon {
 249
 250         /* list of reserved physical pages */
 251         struct list_head pages;
 252
 253         /* transient list of non-balloonable pages */
 254         struct list_head refused_pages;
 255         unsigned int n_refused_pages;
 256
 257         /* balloon size in pages */
 258         unsigned int size;
 259         unsigned int target;
 260
 261         /* reset flag */
 262         bool reset_required;
 263
 264         /* adjustment rates (pages per second) */
 265         unsigned int rate_alloc;
 266
 267         /* slowdown page allocations for next few cycles */
 268         unsigned int slow_allocation_cycles;
 269
 270         unsigned long capabilities;
 271
 272         struct vmballoon_batch_page *batch_page;
 273         unsigned int batch_max_pages;
 274         struct page *page;
 275
 276         const struct vmballoon_ops *ops;
 277
 278 #ifdef CONFIG_DEBUG_FS
 279         /* statistics */
 280         struct vmballoon_stats stats;
 281
 282         /* debugfs file exporting statistics */
 283         struct dentry *dbg_entry;
 284 #endif
 285
 286         struct sysinfo sysinfo;
 287
 288         struct delayed_work dwork;
 289 };
 290
 291 static struct vmballoon balloon;
 292
 293 /*
 294  * Send "start" command to the host, communicating supported version
 295  * of the protocol.
 296  */
 297 static bool vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
 298 {
 299         unsigned long status, capabilities, dummy = 0;
 300
 301         STATS_INC(b->stats.start);
 302
 303         status = VMWARE_BALLOON_CMD(START, req_caps, dummy, capabilities);
 304
 305         switch (status) {
 306         case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
 307                 b->capabilities = capabilities;
 308                 return true;
 309         case VMW_BALLOON_SUCCESS:
 310                 b->capabilities = VMW_BALLOON_BASIC_CMDS;
 311                 return true;
 312         }
 313
 314         pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 315         STATS_INC(b->stats.start_fail);
 316         return false;
 317 }
 318
 319 static bool vmballoon_check_status(struct vmballoon *b, unsigned long status)
 320 {
 321         switch (status) {
 322         case VMW_BALLOON_SUCCESS:
 323                 return true;
 324
 325         case VMW_BALLOON_ERROR_RESET:
 326                 b->reset_required = true;
 327                 /* fall through */
 328
 329         default:
 330                 return false;
 331         }
 332 }
 333
 334 /*
 335  * Communicate guest type to the host so that it can adjust ballooning
 336  * algorithm to the one most appropriate for the guest. This command
 337  * is normally issued after sending "start" command and is part of
 338  * standard reset sequence.
 339  */
 340 static bool vmballoon_send_guest_id(struct vmballoon *b)
 341 {
 342         unsigned long status, dummy = 0;
 343
 344         status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy,
 345                                 dummy);
 346
 347         STATS_INC(b->stats.guest_type);
 348
 349         if (vmballoon_check_status(b, status))
 350                 return true;
 351
 352         pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 353         STATS_INC(b->stats.guest_type_fail);
 354         return false;
 355 }
 356
 357 /*
 358  * Retrieve desired balloon size from the host.
 359  */
 360 static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target)
 361 {
 362         unsigned long status;
 363         unsigned long target;
 364         unsigned long limit;
 365         unsigned long dummy = 0;
 366         u32 limit32;
 367
 368         /*
 369          * si_meminfo() is cheap. Moreover, we want to provide dynamic
 370          * max balloon size later. So let us call si_meminfo() every
 371          * iteration.
 372          */
 373         si_meminfo(&b->sysinfo);
 374         limit = b->sysinfo.totalram;
 375
 376         /* Ensure limit fits in 32-bits */
 377         limit32 = (u32)limit;
 378         if (limit != limit32)
 379                 return false;
 380
 381         /* update stats */
 382         STATS_INC(b->stats.target);
 383
 384         status = VMWARE_BALLOON_CMD(GET_TARGET, limit, dummy, target);
 385         if (vmballoon_check_status(b, status)) {
 386                 *new_target = target;
 387                 return true;
 388         }
 389
 390         pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 391         STATS_INC(b->stats.target_fail);
 392         return false;
 393 }
 394
 395 /*
 396  * Notify the host about allocated page so that host can use it without
 397  * fear that guest will need it. Host may reject some pages, we need to
 398  * check the return value and maybe submit a different page.
 399  */
 400 static int vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn,
 401                                 unsigned int *hv_status, unsigned int *target)
 402 {
 403         unsigned long status, dummy = 0;
 404         u32 pfn32;
 405
 406         pfn32 = (u32)pfn;
 407         if (pfn32 != pfn)
 408                 return -1;
 409
 410         STATS_INC(b->stats.lock);
 411
 412         *hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy, *target);
 413         if (vmballoon_check_status(b, status))
 414                 return 0;
 415
 416         pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
 417         STATS_INC(b->stats.lock_fail);
 418         return 1;
 419 }
 420
 421 static int vmballoon_send_batched_lock(struct vmballoon *b,
 422                                 unsigned int num_pages, unsigned int *target)
 423 {
 424         unsigned long status;
 425         unsigned long pfn = page_to_pfn(b->page);
 426
 427         STATS_INC(b->stats.lock);
 428
 429         status = VMWARE_BALLOON_CMD(BATCHED_LOCK, pfn, num_pages, *target);
 430         if (vmballoon_check_status(b, status))
 431                 return 0;
 432
 433         pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
 434         STATS_INC(b->stats.lock_fail);
 435         return 1;
 436 }
 437
 438 /*
 439  * Notify the host that guest intends to release given page back into
 440  * the pool of available (to the guest) pages.
 441  */
 442 static bool vmballoon_send_unlock_page(struct vmballoon *b, unsigned long pfn,
 443                                                         unsigned int *target)
 444 {
 445         unsigned long status, dummy = 0;
 446         u32 pfn32;
 447
 448         pfn32 = (u32)pfn;
 449         if (pfn32 != pfn)
 450                 return false;
 451
 452         STATS_INC(b->stats.unlock);
 453
 454         status = VMWARE_BALLOON_CMD(UNLOCK, pfn, dummy, *target);
 455         if (vmballoon_check_status(b, status))
 456                 return true;
 457
 458         pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
 459         STATS_INC(b->stats.unlock_fail);
 460         return false;
 461 }
 462
 463 static bool vmballoon_send_batched_unlock(struct vmballoon *b,
 464                                 unsigned int num_pages, unsigned int *target)
 465 {
 466         unsigned long status;
 467         unsigned long pfn = page_to_pfn(b->page);
 468
 469         STATS_INC(b->stats.unlock);
 470
 471         status = VMWARE_BALLOON_CMD(BATCHED_UNLOCK, pfn, num_pages, *target);
 472         if (vmballoon_check_status(b, status))
 473                 return true;
 474
 475         pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
 476         STATS_INC(b->stats.unlock_fail);
 477         return false;
 478 }
 479
 480 /*
 481  * Quickly release all pages allocated for the balloon. This function is
 482  * called when host decides to "reset" balloon for one reason or another.
 483  * Unlike normal "deflate" we do not (shall not) notify host of the pages
 484  * being released.
 485  */
 486 static void vmballoon_pop(struct vmballoon *b)
 487 {
 488         struct page *page, *next;
 489
 490         list_for_each_entry_safe(page, next, &b->pages, lru) {
 491                 list_del(&page->lru);
 492                 __free_page(page);
 493                 STATS_INC(b->stats.free);
 494                 b->size--;
 495                 cond_resched();
 496         }
 497
 498         if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
 499                 if (b->batch_page)
 500                         vunmap(b->batch_page);
 501
 502                 if (b->page)
 503                         __free_page(b->page);
 504         }
 505 }
 506
 507 /*
 508  * Notify the host of a ballooned page. If host rejects the page put it on the
 509  * refuse list, those refused page are then released at the end of the
 510  * inflation cycle.
 511  */
 512 static int vmballoon_lock_page(struct vmballoon *b, unsigned int num_pages,
 513                                                         unsigned int *target)
 514 {
 515         int locked, hv_status;
 516         struct page *page = b->page;
 517
 518         locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status,
 519                                                                 target);
 520         if (locked > 0) {
 521                 STATS_INC(b->stats.refused_alloc);
 522
 523                 if (hv_status == VMW_BALLOON_ERROR_RESET ||
 524                                 hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED) {
 525                         __free_page(page);
 526                         return -EIO;
 527                 }
 528
 529                 /*
 530                  * Place page on the list of non-balloonable pages
 531                  * and retry allocation, unless we already accumulated
 532                  * too many of them, in which case take a breather.
 533                  */
 534                 if (b->n_refused_pages < VMW_BALLOON_MAX_REFUSED) {
 535                         b->n_refused_pages++;
 536                         list_add(&page->lru, &b->refused_pages);
 537                 } else {
 538                         __free_page(page);
 539                 }
 540                 return -EIO;
 541         }
 542
 543         /* track allocated page */
 544         list_add(&page->lru, &b->pages);
 545
 546         /* update balloon size */
 547         b->size++;
 548
 549         return 0;
 550 }
 551
 552 static int vmballoon_lock_batched_page(struct vmballoon *b,
 553                                 unsigned int num_pages, unsigned int *target)
 554 {
 555         int locked, i;
 556
 557         locked = vmballoon_send_batched_lock(b, num_pages, target);
 558         if (locked > 0) {
 559                 for (i = 0; i < num_pages; i++) {
 560                         u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 561                         struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 562
 563                         __free_page(p);
 564                 }
 565
 566                 return -EIO;
 567         }
 568
 569         for (i = 0; i < num_pages; i++) {
 570                 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 571                 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 572
 573                 locked = vmballoon_batch_get_status(b->batch_page, i);
 574
 575                 switch (locked) {
 576                 case VMW_BALLOON_SUCCESS:
 577                         list_add(&p->lru, &b->pages);
 578                         b->size++;
 579                         break;
 580                 case VMW_BALLOON_ERROR_PPN_PINNED:
 581                 case VMW_BALLOON_ERROR_PPN_INVALID:
 582                         if (b->n_refused_pages < VMW_BALLOON_MAX_REFUSED) {
 583                                 list_add(&p->lru, &b->refused_pages);
 584                                 b->n_refused_pages++;
 585                                 break;
 586                         }
 587                         /* Fallthrough */
 588                 case VMW_BALLOON_ERROR_RESET:
 589                 case VMW_BALLOON_ERROR_PPN_NOTNEEDED:
 590                         __free_page(p);
 591                         break;
 592                 default:
 593                         /* This should never happen */
 594                         WARN_ON_ONCE(true);
 595                 }
 596         }
 597
 598         return 0;
 599 }
 600
 601 /*
 602  * Release the page allocated for the balloon. Note that we first notify
 603  * the host so it can make sure the page will be available for the guest
 604  * to use, if needed.
 605  */
 606 static int vmballoon_unlock_page(struct vmballoon *b, unsigned int num_pages,
 607                                                         unsigned int *target)
 608 {
 609         struct page *page = b->page;
 610
 611         if (!vmballoon_send_unlock_page(b, page_to_pfn(page), target)) {
 612                 list_add(&page->lru, &b->pages);
 613                 return -EIO;
 614         }
 615
 616         /* deallocate page */
 617         __free_page(page);
 618         STATS_INC(b->stats.free);
 619
 620         /* update balloon size */
 621         b->size--;
 622
 623         return 0;
 624 }
 625
 626 static int vmballoon_unlock_batched_page(struct vmballoon *b,
 627                                 unsigned int num_pages, unsigned int *target)
 628 {
 629         int locked, i, ret = 0;
 630         bool hv_success;
 631
 632         hv_success = vmballoon_send_batched_unlock(b, num_pages, target);
 633         if (!hv_success)
 634                 ret = -EIO;
 635
 636         for (i = 0; i < num_pages; i++) {
 637                 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 638                 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 639
 640                 locked = vmballoon_batch_get_status(b->batch_page, i);
 641                 if (!hv_success || locked != VMW_BALLOON_SUCCESS) {
 642                         /*
 643                          * That page wasn't successfully unlocked by the
 644                          * hypervisor, re-add it to the list of pages owned by
 645                          * the balloon driver.
 646                          */
 647                         list_add(&p->lru, &b->pages);
 648                 } else {
 649                         /* deallocate page */
 650                         __free_page(p);
 651                         STATS_INC(b->stats.free);
 652
 653                         /* update balloon size */
 654                         b->size--;
 655                 }
 656         }
 657
 658         return ret;
 659 }
 660
 661 /*
 662  * Release pages that were allocated while attempting to inflate the
 663  * balloon but were refused by the host for one reason or another.
 664  */
 665 static void vmballoon_release_refused_pages(struct vmballoon *b)
 666 {
 667         struct page *page, *next;
 668
 669         list_for_each_entry_safe(page, next, &b->refused_pages, lru) {
 670                 list_del(&page->lru);
 671                 __free_page(page);
 672                 STATS_INC(b->stats.refused_free);
 673         }
 674
 675         b->n_refused_pages = 0;
 676 }
 677
 678 static void vmballoon_add_page(struct vmballoon *b, int idx, struct page *p)
 679 {
 680         b->page = p;
 681 }
 682
 683 static void vmballoon_add_batched_page(struct vmballoon *b, int idx,
 684                                 struct page *p)
 685 {
 686         vmballoon_batch_set_pa(b->batch_page, idx,
 687                         (u64)page_to_pfn(p) << PAGE_SHIFT);
 688 }
 689
 690 /*
 691  * Inflate the balloon towards its target size. Note that we try to limit
 692  * the rate of allocation to make sure we are not choking the rest of the
 693  * system.
 694  */
 695 static void vmballoon_inflate(struct vmballoon *b)
 696 {
 697         unsigned rate;
 698         unsigned int allocations = 0;
 699         unsigned int num_pages = 0;
 700         int error = 0;
 701         gfp_t flags = VMW_PAGE_ALLOC_NOSLEEP;
 702
 703         pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
 704
 705         /*
 706          * First try NOSLEEP page allocations to inflate balloon.
 707          *
 708          * If we do not throttle nosleep allocations, we can drain all
 709          * free pages in the guest quickly (if the balloon target is high).
 710          * As a side-effect, draining free pages helps to inform (force)
 711          * the guest to start swapping if balloon target is not met yet,
 712          * which is a desired behavior. However, balloon driver can consume
 713          * all available CPU cycles if too many pages are allocated in a
 714          * second. Therefore, we throttle nosleep allocations even when
 715          * the guest is not under memory pressure. OTOH, if we have already
 716          * predicted that the guest is under memory pressure, then we
 717          * slowdown page allocations considerably.
 718          */
 719
 720         /*
 721          * Start with no sleep allocation rate which may be higher
 722          * than sleeping allocation rate.
 723          */
 724         rate = b->slow_allocation_cycles ? b->rate_alloc : UINT_MAX;
 725
 726         pr_debug("%s - goal: %d, no-sleep rate: %u, sleep rate: %d\n",
 727                  __func__, b->target - b->size, rate, b->rate_alloc);
 728
 729         while (!b->reset_required &&
 730                 b->size < b->target && num_pages < b->target - b->size) {
 731                 struct page *page;
 732
 733                 if (flags == VMW_PAGE_ALLOC_NOSLEEP)
 734                         STATS_INC(b->stats.alloc);
 735                 else
 736                         STATS_INC(b->stats.sleep_alloc);
 737
 738                 page = alloc_page(flags);
 739                 if (!page) {
 740                         if (flags == VMW_PAGE_ALLOC_CANSLEEP) {
 741                                 /*
 742                                  * CANSLEEP page allocation failed, so guest
 743                                  * is under severe memory pressure. Quickly
 744                                  * decrease allocation rate.
 745                                  */
 746                                 b->rate_alloc = max(b->rate_alloc / 2,
 747                                                     VMW_BALLOON_RATE_ALLOC_MIN);
 748                                 STATS_INC(b->stats.sleep_alloc_fail);
 749                                 break;
 750                         }
 751                         STATS_INC(b->stats.alloc_fail);
 752
 753                         /*
 754                          * NOSLEEP page allocation failed, so the guest is
 755                          * under memory pressure. Let us slow down page
 756                          * allocations for next few cycles so that the guest
 757                          * gets out of memory pressure. Also, if we already
 758                          * allocated b->rate_alloc pages, let's pause,
 759                          * otherwise switch to sleeping allocations.
 760                          */
 761                         b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES;
 762
 763                         if (allocations >= b->rate_alloc)
 764                                 break;
 765
 766                         flags = VMW_PAGE_ALLOC_CANSLEEP;
 767                         /* Lower rate for sleeping allocations. */
 768                         rate = b->rate_alloc;
 769                         continue;
 770                 }
 771
 772                 b->ops->add_page(b, num_pages++, page);
 773                 if (num_pages == b->batch_max_pages) {
 774                         error = b->ops->lock(b, num_pages, &b->target);
 775                         num_pages = 0;
 776                         if (error)
 777                                 break;
 778                 }
 779
 780                 cond_resched();
 781
 782                 if (allocations >= rate) {
 783                         /* We allocated enough pages, let's take a break. */
 784                         break;
 785                 }
 786         }
 787
 788         if (num_pages > 0)
 789                 b->ops->lock(b, num_pages, &b->target);
 790
 791         /*
 792          * We reached our goal without failures so try increasing
 793          * allocation rate.
 794          */
 795         if (error == 0 && allocations >= b->rate_alloc) {
 796                 unsigned int mult = allocations / b->rate_alloc;
 797
 798                 b->rate_alloc =
 799                         min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC,
 800                             VMW_BALLOON_RATE_ALLOC_MAX);
 801         }
 802
 803         vmballoon_release_refused_pages(b);
 804 }
 805
 806 /*
 807  * Decrease the size of the balloon allowing guest to use more memory.
 808  */
 809 static void vmballoon_deflate(struct vmballoon *b)
 810 {
 811         struct page *page, *next;
 812         unsigned int i = 0;
 813         unsigned int num_pages = 0;
 814         int error;
 815
 816         pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
 817
 818         /* free pages to reach target */
 819         list_for_each_entry_safe(page, next, &b->pages, lru) {
 820                 list_del(&page->lru);
 821                 b->ops->add_page(b, num_pages++, page);
 822
 823
 824                 if (num_pages == b->batch_max_pages) {
 825                         error = b->ops->unlock(b, num_pages, &b->target);
 826                         num_pages = 0;
 827                         if (error)
 828                                 return;
 829                 }
 830
 831                 if (b->reset_required || ++i >= b->size - b->target)
 832                         break;
 833
 834                 cond_resched();
 835         }
 836
 837         if (num_pages > 0)
 838                 b->ops->unlock(b, num_pages, &b->target);
 839 }
 840
 841 static const struct vmballoon_ops vmballoon_basic_ops = {
 842         .add_page = vmballoon_add_page,
 843         .lock = vmballoon_lock_page,
 844         .unlock = vmballoon_unlock_page
 845 };
 846
 847 static const struct vmballoon_ops vmballoon_batched_ops = {
 848         .add_page = vmballoon_add_batched_page,
 849         .lock = vmballoon_lock_batched_page,
 850         .unlock = vmballoon_unlock_batched_page
 851 };
 852
 853 static bool vmballoon_init_batching(struct vmballoon *b)
 854 {
 855         b->page = alloc_page(VMW_PAGE_ALLOC_NOSLEEP);
 856         if (!b->page)
 857                 return false;
 858
 859         b->batch_page = vmap(&b->page, 1, VM_MAP, PAGE_KERNEL);
 860         if (!b->batch_page) {
 861                 __free_page(b->page);
 862                 return false;
 863         }
 864
 865         return true;
 866 }
 867
 868 /*
 869  * Perform standard reset sequence by popping the balloon (in case it
 870  * is not  empty) and then restarting protocol. This operation normally
 871  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
 872  */
 873 static void vmballoon_reset(struct vmballoon *b)
 874 {
 875         /* free all pages, skipping monitor unlock */
 876         vmballoon_pop(b);
 877
 878         if (!vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
 879                 return;
 880
 881         if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
 882                 b->ops = &vmballoon_batched_ops;
 883                 b->batch_max_pages = VMW_BALLOON_BATCH_MAX_PAGES;
 884                 if (!vmballoon_init_batching(b)) {
 885                         /*
 886                          * We failed to initialize batching, inform the monitor
 887                          * about it by sending a null capability.
 888                          *
 889                          * The guest will retry in one second.
 890                          */
 891                         vmballoon_send_start(b, 0);
 892                         return;
 893                 }
 894         } else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
 895                 b->ops = &vmballoon_basic_ops;
 896                 b->batch_max_pages = 1;
 897         }
 898
 899         b->reset_required = false;
 900         if (!vmballoon_send_guest_id(b))
 901                 pr_err("failed to send guest ID to the host\n");
 902 }
 903
 904 /*
 905  * Balloon work function: reset protocol, if needed, get the new size and
 906  * adjust balloon as needed. Repeat in 1 sec.
 907  */
 908 static void vmballoon_work(struct work_struct *work)
 909 {
 910         struct delayed_work *dwork = to_delayed_work(work);
 911         struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
 912         unsigned int target;
 913
 914         STATS_INC(b->stats.timer);
 915
 916         if (b->reset_required)
 917                 vmballoon_reset(b);
 918
 919         if (b->slow_allocation_cycles > 0)
 920                 b->slow_allocation_cycles--;
 921
 922         if (vmballoon_send_get_target(b, &target)) {
 923                 /* update target, adjust size */
 924                 b->target = target;
 925
 926                 if (b->size < target)
 927                         vmballoon_inflate(b);
 928                 else if (b->size > target)
 929                         vmballoon_deflate(b);
 930         }
 931
 932         /*
 933          * We are using a freezable workqueue so that balloon operations are
 934          * stopped while the system transitions to/from sleep/hibernation.
 935          */
 936         queue_delayed_work(system_freezable_wq,
 937                            dwork, round_jiffies_relative(HZ));
 938 }
 939
 940 /*
 941  * DEBUGFS Interface
 942  */
 943 #ifdef CONFIG_DEBUG_FS
 944
 945 static int vmballoon_debug_show(struct seq_file *f, void *offset)
 946 {
 947         struct vmballoon *b = f->private;
 948         struct vmballoon_stats *stats = &b->stats;
 949
 950         /* format capabilities info */
 951         seq_printf(f,
 952                    "balloon capabilities:   %#4x\n"
 953                    "used capabilities:      %#4lx\n",
 954                    VMW_BALLOON_CAPABILITIES, b->capabilities);
 955
 956         /* format size info */
 957         seq_printf(f,
 958                    "target:             %8d pages\n"
 959                    "current:            %8d pages\n",
 960                    b->target, b->size);
 961
 962         /* format rate info */
 963         seq_printf(f,
 964                    "rateSleepAlloc:     %8d pages/sec\n",
 965                    b->rate_alloc);
 966
 967         seq_printf(f,
 968                    "\n"
 969                    "timer:              %8u\n"
 970                    "start:              %8u (%4u failed)\n"
 971                    "guestType:          %8u (%4u failed)\n"
 972                    "lock:               %8u (%4u failed)\n"
 973                    "unlock:             %8u (%4u failed)\n"
 974                    "target:             %8u (%4u failed)\n"
 975                    "primNoSleepAlloc:   %8u (%4u failed)\n"
 976                    "primCanSleepAlloc:  %8u (%4u failed)\n"
 977                    "primFree:           %8u\n"
 978                    "errAlloc:           %8u\n"
 979                    "errFree:            %8u\n",
 980                    stats->timer,
 981                    stats->start, stats->start_fail,
 982                    stats->guest_type, stats->guest_type_fail,
 983                    stats->lock,  stats->lock_fail,
 984                    stats->unlock, stats->unlock_fail,
 985                    stats->target, stats->target_fail,
 986                    stats->alloc, stats->alloc_fail,
 987                    stats->sleep_alloc, stats->sleep_alloc_fail,
 988                    stats->free,
 989                    stats->refused_alloc, stats->refused_free);
 990
 991         return 0;
 992 }
 993
 994 static int vmballoon_debug_open(struct inode *inode, struct file *file)
 995 {
 996         return single_open(file, vmballoon_debug_show, inode->i_private);
 997 }
 998
 999 static const struct file_operations vmballoon_debug_fops = {
1000         .owner          = THIS_MODULE,
1001         .open           = vmballoon_debug_open,
1002         .read           = seq_read,
1003         .llseek         = seq_lseek,
1004         .release        = single_release,
1005 };
1006
1007 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1008 {
1009         int error;
1010
1011         b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1012                                            &vmballoon_debug_fops);
1013         if (IS_ERR(b->dbg_entry)) {
1014                 error = PTR_ERR(b->dbg_entry);
1015                 pr_err("failed to create debugfs entry, error: %d\n", error);
1016                 return error;
1017         }
1018
1019         return 0;
1020 }
1021
1022 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1023 {
1024         debugfs_remove(b->dbg_entry);
1025 }
1026
1027 #else
1028
1029 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1030 {
1031         return 0;
1032 }
1033
1034 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1035 {
1036 }
1037
1038 #endif  /* CONFIG_DEBUG_FS */
1039
1040 static int __init vmballoon_init(void)
1041 {
1042         int error;
1043
1044         /*
1045          * Check if we are running on VMware's hypervisor and bail out
1046          * if we are not.
1047          */
1048         if (x86_hyper != &x86_hyper_vmware)
1049                 return -ENODEV;
1050
1051         INIT_LIST_HEAD(&balloon.pages);
1052         INIT_LIST_HEAD(&balloon.refused_pages);
1053
1054         /* initialize rates */
1055         balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
1056
1057         INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1058
1059         /*
1060          * Start balloon.
1061          */
1062         if (!vmballoon_send_start(&balloon, VMW_BALLOON_CAPABILITIES)) {
1063                 pr_err("failed to send start command to the host\n");
1064                 return -EIO;
1065         }
1066
1067         if ((balloon.capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1068                 balloon.ops = &vmballoon_batched_ops;
1069                 balloon.batch_max_pages = VMW_BALLOON_BATCH_MAX_PAGES;
1070                 if (!vmballoon_init_batching(&balloon)) {
1071                         pr_err("failed to init batching\n");
1072                         return -EIO;
1073                 }
1074         } else if ((balloon.capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1075                 balloon.ops = &vmballoon_basic_ops;
1076                 balloon.batch_max_pages = 1;
1077         }
1078
1079         if (!vmballoon_send_guest_id(&balloon)) {
1080                 pr_err("failed to send guest ID to the host\n");
1081                 return -EIO;
1082         }
1083
1084         error = vmballoon_debugfs_init(&balloon);
1085         if (error)
1086                 return error;
1087
1088         queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1089
1090         return 0;
1091 }
1092 module_init(vmballoon_init);
1093
1094 static void __exit vmballoon_exit(void)
1095 {
1096         cancel_delayed_work_sync(&balloon.dwork);
1097
1098         vmballoon_debugfs_exit(&balloon);
1099
1100         /*
1101          * Deallocate all reserved memory, and reset connection with monitor.
1102          * Reset connection before deallocating memory to avoid potential for
1103          * additional spurious resets from guest touching deallocated pages.
1104          */
1105         vmballoon_send_start(&balloon, VMW_BALLOON_CAPABILITIES);
1106         vmballoon_pop(&balloon);
1107 }
1108 module_exit(vmballoon_exit);