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("1.4.0.0-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         VMW_BALLOON_BATCHED_2M_CMDS     = (1 << 3),
 107 };
 108
 109 #define VMW_BALLOON_CAPABILITIES        (VMW_BALLOON_BASIC_CMDS \
 110                                         | VMW_BALLOON_BATCHED_CMDS \
 111                                         | VMW_BALLOON_BATCHED_2M_CMDS)
 112
 113 #define VMW_BALLOON_2M_SHIFT            (9)
 114 #define VMW_BALLOON_NUM_PAGE_SIZES      (2)
 115
 116 /*
 117  * Backdoor commands availability:
 118  *
 119  * START, GET_TARGET and GUEST_ID are always available,
 120  *
 121  * VMW_BALLOON_BASIC_CMDS:
 122  *      LOCK and UNLOCK commands,
 123  * VMW_BALLOON_BATCHED_CMDS:
 124  *      BATCHED_LOCK and BATCHED_UNLOCK commands.
 125  * VMW BALLOON_BATCHED_2M_CMDS:
 126  *      BATCHED_2M_LOCK and BATCHED_2M_UNLOCK commands.
 127  */
 128 #define VMW_BALLOON_CMD_START                   0
 129 #define VMW_BALLOON_CMD_GET_TARGET              1
 130 #define VMW_BALLOON_CMD_LOCK                    2
 131 #define VMW_BALLOON_CMD_UNLOCK                  3
 132 #define VMW_BALLOON_CMD_GUEST_ID                4
 133 #define VMW_BALLOON_CMD_BATCHED_LOCK            6
 134 #define VMW_BALLOON_CMD_BATCHED_UNLOCK          7
 135 #define VMW_BALLOON_CMD_BATCHED_2M_LOCK         8
 136 #define VMW_BALLOON_CMD_BATCHED_2M_UNLOCK       9
 137
 138
 139 /* error codes */
 140 #define VMW_BALLOON_SUCCESS                     0
 141 #define VMW_BALLOON_FAILURE                     -1
 142 #define VMW_BALLOON_ERROR_CMD_INVALID           1
 143 #define VMW_BALLOON_ERROR_PPN_INVALID           2
 144 #define VMW_BALLOON_ERROR_PPN_LOCKED            3
 145 #define VMW_BALLOON_ERROR_PPN_UNLOCKED          4
 146 #define VMW_BALLOON_ERROR_PPN_PINNED            5
 147 #define VMW_BALLOON_ERROR_PPN_NOTNEEDED         6
 148 #define VMW_BALLOON_ERROR_RESET                 7
 149 #define VMW_BALLOON_ERROR_BUSY                  8
 150
 151 #define VMW_BALLOON_SUCCESS_WITH_CAPABILITIES   (0x03000000)
 152
 153 /* Batch page description */
 154
 155 /*
 156  * Layout of a page in the batch page:
 157  *
 158  * +-------------+----------+--------+
 159  * |             |          |        |
 160  * | Page number | Reserved | Status |
 161  * |             |          |        |
 162  * +-------------+----------+--------+
 163  * 64  PAGE_SHIFT          6         0
 164  *
 165  * The reserved field should be set to 0.
 166  */
 167 #define VMW_BALLOON_BATCH_MAX_PAGES     (PAGE_SIZE / sizeof(u64))
 168 #define VMW_BALLOON_BATCH_STATUS_MASK   ((1UL << 5) - 1)
 169 #define VMW_BALLOON_BATCH_PAGE_MASK     (~((1UL << PAGE_SHIFT) - 1))
 170
 171 struct vmballoon_batch_page {
 172         u64 pages[VMW_BALLOON_BATCH_MAX_PAGES];
 173 };
 174
 175 static u64 vmballoon_batch_get_pa(struct vmballoon_batch_page *batch, int idx)
 176 {
 177         return batch->pages[idx] & VMW_BALLOON_BATCH_PAGE_MASK;
 178 }
 179
 180 static int vmballoon_batch_get_status(struct vmballoon_batch_page *batch,
 181                                 int idx)
 182 {
 183         return (int)(batch->pages[idx] & VMW_BALLOON_BATCH_STATUS_MASK);
 184 }
 185
 186 static void vmballoon_batch_set_pa(struct vmballoon_batch_page *batch, int idx,
 187                                 u64 pa)
 188 {
 189         batch->pages[idx] = pa;
 190 }
 191
 192
 193 #define VMWARE_BALLOON_CMD(cmd, arg1, arg2, result)             \
 194 ({                                                              \
 195         unsigned long __status, __dummy1, __dummy2, __dummy3;   \
 196         __asm__ __volatile__ ("inl %%dx" :                      \
 197                 "=a"(__status),                                 \
 198                 "=c"(__dummy1),                                 \
 199                 "=d"(__dummy2),                                 \
 200                 "=b"(result),                                   \
 201                 "=S" (__dummy3) :                               \
 202                 "0"(VMW_BALLOON_HV_MAGIC),                      \
 203                 "1"(VMW_BALLOON_CMD_##cmd),                     \
 204                 "2"(VMW_BALLOON_HV_PORT),                       \
 205                 "3"(arg1),                                      \
 206                 "4" (arg2) :                                    \
 207                 "memory");                                      \
 208         if (VMW_BALLOON_CMD_##cmd == VMW_BALLOON_CMD_START)     \
 209                 result = __dummy1;                              \
 210         result &= -1UL;                                         \
 211         __status & -1UL;                                        \
 212 })
 213
 214 #ifdef CONFIG_DEBUG_FS
 215 struct vmballoon_stats {
 216         unsigned int timer;
 217
 218         /* allocation statistics */
 219         unsigned int alloc[VMW_BALLOON_NUM_PAGE_SIZES];
 220         unsigned int alloc_fail[VMW_BALLOON_NUM_PAGE_SIZES];
 221         unsigned int sleep_alloc;
 222         unsigned int sleep_alloc_fail;
 223         unsigned int refused_alloc[VMW_BALLOON_NUM_PAGE_SIZES];
 224         unsigned int refused_free[VMW_BALLOON_NUM_PAGE_SIZES];
 225         unsigned int free[VMW_BALLOON_NUM_PAGE_SIZES];
 226
 227         /* monitor operations */
 228         unsigned int lock[VMW_BALLOON_NUM_PAGE_SIZES];
 229         unsigned int lock_fail[VMW_BALLOON_NUM_PAGE_SIZES];
 230         unsigned int unlock[VMW_BALLOON_NUM_PAGE_SIZES];
 231         unsigned int unlock_fail[VMW_BALLOON_NUM_PAGE_SIZES];
 232         unsigned int target;
 233         unsigned int target_fail;
 234         unsigned int start;
 235         unsigned int start_fail;
 236         unsigned int guest_type;
 237         unsigned int guest_type_fail;
 238 };
 239
 240 #define STATS_INC(stat) (stat)++
 241 #else
 242 #define STATS_INC(stat)
 243 #endif
 244
 245 struct vmballoon;
 246
 247 struct vmballoon_ops {
 248         void (*add_page)(struct vmballoon *b, int idx, struct page *p);
 249         int (*lock)(struct vmballoon *b, unsigned int num_pages,
 250                         bool is_2m_pages, unsigned int *target);
 251         int (*unlock)(struct vmballoon *b, unsigned int num_pages,
 252                         bool is_2m_pages, unsigned int *target);
 253 };
 254
 255 struct vmballoon_page_size {
 256         /* list of reserved physical pages */
 257         struct list_head pages;
 258
 259         /* transient list of non-balloonable pages */
 260         struct list_head refused_pages;
 261         unsigned int n_refused_pages;
 262 };
 263
 264 struct vmballoon {
 265         struct vmballoon_page_size page_sizes[VMW_BALLOON_NUM_PAGE_SIZES];
 266
 267         /* supported page sizes. 1 == 4k pages only, 2 == 4k and 2m pages */
 268         unsigned supported_page_sizes;
 269
 270         /* balloon size in pages */
 271         unsigned int size;
 272         unsigned int target;
 273
 274         /* reset flag */
 275         bool reset_required;
 276
 277         /* adjustment rates (pages per second) */
 278         unsigned int rate_alloc;
 279
 280         /* slowdown page allocations for next few cycles */
 281         unsigned int slow_allocation_cycles;
 282
 283         unsigned long capabilities;
 284
 285         struct vmballoon_batch_page *batch_page;
 286         unsigned int batch_max_pages;
 287         struct page *page;
 288
 289         const struct vmballoon_ops *ops;
 290
 291 #ifdef CONFIG_DEBUG_FS
 292         /* statistics */
 293         struct vmballoon_stats stats;
 294
 295         /* debugfs file exporting statistics */
 296         struct dentry *dbg_entry;
 297 #endif
 298
 299         struct sysinfo sysinfo;
 300
 301         struct delayed_work dwork;
 302 };
 303
 304 static struct vmballoon balloon;
 305
 306 /*
 307  * Send "start" command to the host, communicating supported version
 308  * of the protocol.
 309  */
 310 static bool vmballoon_send_start(struct vmballoon *b, unsigned long req_caps)
 311 {
 312         unsigned long status, capabilities, dummy = 0;
 313         bool success;
 314
 315         STATS_INC(b->stats.start);
 316
 317         status = VMWARE_BALLOON_CMD(START, req_caps, dummy, capabilities);
 318
 319         switch (status) {
 320         case VMW_BALLOON_SUCCESS_WITH_CAPABILITIES:
 321                 b->capabilities = capabilities;
 322                 success = true;
 323                 break;
 324         case VMW_BALLOON_SUCCESS:
 325                 b->capabilities = VMW_BALLOON_BASIC_CMDS;
 326                 success = true;
 327                 break;
 328         default:
 329                 success = false;
 330         }
 331
 332         if (b->capabilities & VMW_BALLOON_BATCHED_2M_CMDS)
 333                 b->supported_page_sizes = 2;
 334         else
 335                 b->supported_page_sizes = 1;
 336
 337         if (!success) {
 338                 pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 339                 STATS_INC(b->stats.start_fail);
 340         }
 341         return success;
 342 }
 343
 344 static bool vmballoon_check_status(struct vmballoon *b, unsigned long status)
 345 {
 346         switch (status) {
 347         case VMW_BALLOON_SUCCESS:
 348                 return true;
 349
 350         case VMW_BALLOON_ERROR_RESET:
 351                 b->reset_required = true;
 352                 /* fall through */
 353
 354         default:
 355                 return false;
 356         }
 357 }
 358
 359 /*
 360  * Communicate guest type to the host so that it can adjust ballooning
 361  * algorithm to the one most appropriate for the guest. This command
 362  * is normally issued after sending "start" command and is part of
 363  * standard reset sequence.
 364  */
 365 static bool vmballoon_send_guest_id(struct vmballoon *b)
 366 {
 367         unsigned long status, dummy = 0;
 368
 369         status = VMWARE_BALLOON_CMD(GUEST_ID, VMW_BALLOON_GUEST_ID, dummy,
 370                                 dummy);
 371
 372         STATS_INC(b->stats.guest_type);
 373
 374         if (vmballoon_check_status(b, status))
 375                 return true;
 376
 377         pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 378         STATS_INC(b->stats.guest_type_fail);
 379         return false;
 380 }
 381
 382 static u16 vmballoon_page_size(bool is_2m_page)
 383 {
 384         if (is_2m_page)
 385                 return 1 << VMW_BALLOON_2M_SHIFT;
 386
 387         return 1;
 388 }
 389
 390 /*
 391  * Retrieve desired balloon size from the host.
 392  */
 393 static bool vmballoon_send_get_target(struct vmballoon *b, u32 *new_target)
 394 {
 395         unsigned long status;
 396         unsigned long target;
 397         unsigned long limit;
 398         unsigned long dummy = 0;
 399         u32 limit32;
 400
 401         /*
 402          * si_meminfo() is cheap. Moreover, we want to provide dynamic
 403          * max balloon size later. So let us call si_meminfo() every
 404          * iteration.
 405          */
 406         si_meminfo(&b->sysinfo);
 407         limit = b->sysinfo.totalram;
 408
 409         /* Ensure limit fits in 32-bits */
 410         limit32 = (u32)limit;
 411         if (limit != limit32)
 412                 return false;
 413
 414         /* update stats */
 415         STATS_INC(b->stats.target);
 416
 417         status = VMWARE_BALLOON_CMD(GET_TARGET, limit, dummy, target);
 418         if (vmballoon_check_status(b, status)) {
 419                 *new_target = target;
 420                 return true;
 421         }
 422
 423         pr_debug("%s - failed, hv returns %ld\n", __func__, status);
 424         STATS_INC(b->stats.target_fail);
 425         return false;
 426 }
 427
 428 /*
 429  * Notify the host about allocated page so that host can use it without
 430  * fear that guest will need it. Host may reject some pages, we need to
 431  * check the return value and maybe submit a different page.
 432  */
 433 static int vmballoon_send_lock_page(struct vmballoon *b, unsigned long pfn,
 434                                 unsigned int *hv_status, unsigned int *target)
 435 {
 436         unsigned long status, dummy = 0;
 437         u32 pfn32;
 438
 439         pfn32 = (u32)pfn;
 440         if (pfn32 != pfn)
 441                 return -1;
 442
 443         STATS_INC(b->stats.lock[false]);
 444
 445         *hv_status = status = VMWARE_BALLOON_CMD(LOCK, pfn, dummy, *target);
 446         if (vmballoon_check_status(b, status))
 447                 return 0;
 448
 449         pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
 450         STATS_INC(b->stats.lock_fail[false]);
 451         return 1;
 452 }
 453
 454 static int vmballoon_send_batched_lock(struct vmballoon *b,
 455                 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
 456 {
 457         unsigned long status;
 458         unsigned long pfn = page_to_pfn(b->page);
 459
 460         STATS_INC(b->stats.lock[is_2m_pages]);
 461
 462         if (is_2m_pages)
 463                 status = VMWARE_BALLOON_CMD(BATCHED_2M_LOCK, pfn, num_pages,
 464                                 *target);
 465         else
 466                 status = VMWARE_BALLOON_CMD(BATCHED_LOCK, pfn, num_pages,
 467                                 *target);
 468
 469         if (vmballoon_check_status(b, status))
 470                 return 0;
 471
 472         pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
 473         STATS_INC(b->stats.lock_fail[is_2m_pages]);
 474         return 1;
 475 }
 476
 477 /*
 478  * Notify the host that guest intends to release given page back into
 479  * the pool of available (to the guest) pages.
 480  */
 481 static bool vmballoon_send_unlock_page(struct vmballoon *b, unsigned long pfn,
 482                                                         unsigned int *target)
 483 {
 484         unsigned long status, dummy = 0;
 485         u32 pfn32;
 486
 487         pfn32 = (u32)pfn;
 488         if (pfn32 != pfn)
 489                 return false;
 490
 491         STATS_INC(b->stats.unlock[false]);
 492
 493         status = VMWARE_BALLOON_CMD(UNLOCK, pfn, dummy, *target);
 494         if (vmballoon_check_status(b, status))
 495                 return true;
 496
 497         pr_debug("%s - ppn %lx, hv returns %ld\n", __func__, pfn, status);
 498         STATS_INC(b->stats.unlock_fail[false]);
 499         return false;
 500 }
 501
 502 static bool vmballoon_send_batched_unlock(struct vmballoon *b,
 503                 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
 504 {
 505         unsigned long status;
 506         unsigned long pfn = page_to_pfn(b->page);
 507
 508         STATS_INC(b->stats.unlock[is_2m_pages]);
 509
 510         if (is_2m_pages)
 511                 status = VMWARE_BALLOON_CMD(BATCHED_2M_UNLOCK, pfn, num_pages,
 512                                 *target);
 513         else
 514                 status = VMWARE_BALLOON_CMD(BATCHED_UNLOCK, pfn, num_pages,
 515                                 *target);
 516
 517         if (vmballoon_check_status(b, status))
 518                 return true;
 519
 520         pr_debug("%s - batch ppn %lx, hv returns %ld\n", __func__, pfn, status);
 521         STATS_INC(b->stats.unlock_fail[is_2m_pages]);
 522         return false;
 523 }
 524
 525 static struct page *vmballoon_alloc_page(gfp_t flags, bool is_2m_page)
 526 {
 527         if (is_2m_page)
 528                 return alloc_pages(flags, VMW_BALLOON_2M_SHIFT);
 529
 530         return alloc_page(flags);
 531 }
 532
 533 static void vmballoon_free_page(struct page *page, bool is_2m_page)
 534 {
 535         if (is_2m_page)
 536                 __free_pages(page, VMW_BALLOON_2M_SHIFT);
 537         else
 538                 __free_page(page);
 539 }
 540
 541 /*
 542  * Quickly release all pages allocated for the balloon. This function is
 543  * called when host decides to "reset" balloon for one reason or another.
 544  * Unlike normal "deflate" we do not (shall not) notify host of the pages
 545  * being released.
 546  */
 547 static void vmballoon_pop(struct vmballoon *b)
 548 {
 549         struct page *page, *next;
 550         unsigned is_2m_pages;
 551
 552         for (is_2m_pages = 0; is_2m_pages < VMW_BALLOON_NUM_PAGE_SIZES;
 553                         is_2m_pages++) {
 554                 struct vmballoon_page_size *page_size =
 555                                 &b->page_sizes[is_2m_pages];
 556                 u16 size_per_page = vmballoon_page_size(is_2m_pages);
 557
 558                 list_for_each_entry_safe(page, next, &page_size->pages, lru) {
 559                         list_del(&page->lru);
 560                         vmballoon_free_page(page, is_2m_pages);
 561                         STATS_INC(b->stats.free[is_2m_pages]);
 562                         b->size -= size_per_page;
 563                         cond_resched();
 564                 }
 565         }
 566
 567         if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
 568                 if (b->batch_page)
 569                         vunmap(b->batch_page);
 570
 571                 if (b->page)
 572                         __free_page(b->page);
 573         }
 574 }
 575
 576 /*
 577  * Notify the host of a ballooned page. If host rejects the page put it on the
 578  * refuse list, those refused page are then released at the end of the
 579  * inflation cycle.
 580  */
 581 static int vmballoon_lock_page(struct vmballoon *b, unsigned int num_pages,
 582                                 bool is_2m_pages, unsigned int *target)
 583 {
 584         int locked, hv_status;
 585         struct page *page = b->page;
 586         struct vmballoon_page_size *page_size = &b->page_sizes[false];
 587
 588         /* is_2m_pages can never happen as 2m pages support implies batching */
 589
 590         locked = vmballoon_send_lock_page(b, page_to_pfn(page), &hv_status,
 591                                                                 target);
 592         if (locked > 0) {
 593                 STATS_INC(b->stats.refused_alloc[false]);
 594
 595                 if (hv_status == VMW_BALLOON_ERROR_RESET ||
 596                                 hv_status == VMW_BALLOON_ERROR_PPN_NOTNEEDED) {
 597                         vmballoon_free_page(page, false);
 598                         return -EIO;
 599                 }
 600
 601                 /*
 602                  * Place page on the list of non-balloonable pages
 603                  * and retry allocation, unless we already accumulated
 604                  * too many of them, in which case take a breather.
 605                  */
 606                 if (page_size->n_refused_pages < VMW_BALLOON_MAX_REFUSED) {
 607                         page_size->n_refused_pages++;
 608                         list_add(&page->lru, &page_size->refused_pages);
 609                 } else {
 610                         vmballoon_free_page(page, false);
 611                 }
 612                 return -EIO;
 613         }
 614
 615         /* track allocated page */
 616         list_add(&page->lru, &page_size->pages);
 617
 618         /* update balloon size */
 619         b->size++;
 620
 621         return 0;
 622 }
 623
 624 static int vmballoon_lock_batched_page(struct vmballoon *b,
 625                 unsigned int num_pages, bool is_2m_pages, unsigned int *target)
 626 {
 627         int locked, i;
 628         u16 size_per_page = vmballoon_page_size(is_2m_pages);
 629
 630         locked = vmballoon_send_batched_lock(b, num_pages, is_2m_pages,
 631                         target);
 632         if (locked > 0) {
 633                 for (i = 0; i < num_pages; i++) {
 634                         u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 635                         struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 636
 637                         vmballoon_free_page(p, is_2m_pages);
 638                 }
 639
 640                 return -EIO;
 641         }
 642
 643         for (i = 0; i < num_pages; i++) {
 644                 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 645                 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 646                 struct vmballoon_page_size *page_size =
 647                                 &b->page_sizes[is_2m_pages];
 648
 649                 locked = vmballoon_batch_get_status(b->batch_page, i);
 650
 651                 switch (locked) {
 652                 case VMW_BALLOON_SUCCESS:
 653                         list_add(&p->lru, &page_size->pages);
 654                         b->size += size_per_page;
 655                         break;
 656                 case VMW_BALLOON_ERROR_PPN_PINNED:
 657                 case VMW_BALLOON_ERROR_PPN_INVALID:
 658                         if (page_size->n_refused_pages
 659                                         < VMW_BALLOON_MAX_REFUSED) {
 660                                 list_add(&p->lru, &page_size->refused_pages);
 661                                 page_size->n_refused_pages++;
 662                                 break;
 663                         }
 664                         /* Fallthrough */
 665                 case VMW_BALLOON_ERROR_RESET:
 666                 case VMW_BALLOON_ERROR_PPN_NOTNEEDED:
 667                         vmballoon_free_page(p, is_2m_pages);
 668                         break;
 669                 default:
 670                         /* This should never happen */
 671                         WARN_ON_ONCE(true);
 672                 }
 673         }
 674
 675         return 0;
 676 }
 677
 678 /*
 679  * Release the page allocated for the balloon. Note that we first notify
 680  * the host so it can make sure the page will be available for the guest
 681  * to use, if needed.
 682  */
 683 static int vmballoon_unlock_page(struct vmballoon *b, unsigned int num_pages,
 684                 bool is_2m_pages, unsigned int *target)
 685 {
 686         struct page *page = b->page;
 687         struct vmballoon_page_size *page_size = &b->page_sizes[false];
 688
 689         /* is_2m_pages can never happen as 2m pages support implies batching */
 690
 691         if (!vmballoon_send_unlock_page(b, page_to_pfn(page), target)) {
 692                 list_add(&page->lru, &page_size->pages);
 693                 return -EIO;
 694         }
 695
 696         /* deallocate page */
 697         vmballoon_free_page(page, false);
 698         STATS_INC(b->stats.free[false]);
 699
 700         /* update balloon size */
 701         b->size--;
 702
 703         return 0;
 704 }
 705
 706 static int vmballoon_unlock_batched_page(struct vmballoon *b,
 707                                 unsigned int num_pages, bool is_2m_pages,
 708                                 unsigned int *target)
 709 {
 710         int locked, i, ret = 0;
 711         bool hv_success;
 712         u16 size_per_page = vmballoon_page_size(is_2m_pages);
 713
 714         hv_success = vmballoon_send_batched_unlock(b, num_pages, is_2m_pages,
 715                         target);
 716         if (!hv_success)
 717                 ret = -EIO;
 718
 719         for (i = 0; i < num_pages; i++) {
 720                 u64 pa = vmballoon_batch_get_pa(b->batch_page, i);
 721                 struct page *p = pfn_to_page(pa >> PAGE_SHIFT);
 722                 struct vmballoon_page_size *page_size =
 723                                 &b->page_sizes[is_2m_pages];
 724
 725                 locked = vmballoon_batch_get_status(b->batch_page, i);
 726                 if (!hv_success || locked != VMW_BALLOON_SUCCESS) {
 727                         /*
 728                          * That page wasn't successfully unlocked by the
 729                          * hypervisor, re-add it to the list of pages owned by
 730                          * the balloon driver.
 731                          */
 732                         list_add(&p->lru, &page_size->pages);
 733                 } else {
 734                         /* deallocate page */
 735                         vmballoon_free_page(p, is_2m_pages);
 736                         STATS_INC(b->stats.free[is_2m_pages]);
 737
 738                         /* update balloon size */
 739                         b->size -= size_per_page;
 740                 }
 741         }
 742
 743         return ret;
 744 }
 745
 746 /*
 747  * Release pages that were allocated while attempting to inflate the
 748  * balloon but were refused by the host for one reason or another.
 749  */
 750 static void vmballoon_release_refused_pages(struct vmballoon *b,
 751                 bool is_2m_pages)
 752 {
 753         struct page *page, *next;
 754         struct vmballoon_page_size *page_size =
 755                         &b->page_sizes[is_2m_pages];
 756
 757         list_for_each_entry_safe(page, next, &page_size->refused_pages, lru) {
 758                 list_del(&page->lru);
 759                 vmballoon_free_page(page, is_2m_pages);
 760                 STATS_INC(b->stats.refused_free[is_2m_pages]);
 761         }
 762
 763         page_size->n_refused_pages = 0;
 764 }
 765
 766 static void vmballoon_add_page(struct vmballoon *b, int idx, struct page *p)
 767 {
 768         b->page = p;
 769 }
 770
 771 static void vmballoon_add_batched_page(struct vmballoon *b, int idx,
 772                                 struct page *p)
 773 {
 774         vmballoon_batch_set_pa(b->batch_page, idx,
 775                         (u64)page_to_pfn(p) << PAGE_SHIFT);
 776 }
 777
 778 /*
 779  * Inflate the balloon towards its target size. Note that we try to limit
 780  * the rate of allocation to make sure we are not choking the rest of the
 781  * system.
 782  */
 783 static void vmballoon_inflate(struct vmballoon *b)
 784 {
 785         unsigned rate;
 786         unsigned int allocations = 0;
 787         unsigned int num_pages = 0;
 788         int error = 0;
 789         gfp_t flags = VMW_PAGE_ALLOC_NOSLEEP;
 790         bool is_2m_pages;
 791
 792         pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
 793
 794         /*
 795          * First try NOSLEEP page allocations to inflate balloon.
 796          *
 797          * If we do not throttle nosleep allocations, we can drain all
 798          * free pages in the guest quickly (if the balloon target is high).
 799          * As a side-effect, draining free pages helps to inform (force)
 800          * the guest to start swapping if balloon target is not met yet,
 801          * which is a desired behavior. However, balloon driver can consume
 802          * all available CPU cycles if too many pages are allocated in a
 803          * second. Therefore, we throttle nosleep allocations even when
 804          * the guest is not under memory pressure. OTOH, if we have already
 805          * predicted that the guest is under memory pressure, then we
 806          * slowdown page allocations considerably.
 807          */
 808
 809         /*
 810          * Start with no sleep allocation rate which may be higher
 811          * than sleeping allocation rate.
 812          */
 813         if (b->slow_allocation_cycles) {
 814                 rate = b->rate_alloc;
 815                 is_2m_pages = false;
 816         } else {
 817                 rate = UINT_MAX;
 818                 is_2m_pages =
 819                         b->supported_page_sizes == VMW_BALLOON_NUM_PAGE_SIZES;
 820         }
 821
 822         pr_debug("%s - goal: %d, no-sleep rate: %u, sleep rate: %d\n",
 823                  __func__, b->target - b->size, rate, b->rate_alloc);
 824
 825         while (!b->reset_required &&
 826                 b->size + num_pages * vmballoon_page_size(is_2m_pages)
 827                 < b->target) {
 828                 struct page *page;
 829
 830                 if (flags == VMW_PAGE_ALLOC_NOSLEEP)
 831                         STATS_INC(b->stats.alloc[is_2m_pages]);
 832                 else
 833                         STATS_INC(b->stats.sleep_alloc);
 834
 835                 page = vmballoon_alloc_page(flags, is_2m_pages);
 836                 if (!page) {
 837                         STATS_INC(b->stats.alloc_fail[is_2m_pages]);
 838
 839                         if (is_2m_pages) {
 840                                 b->ops->lock(b, num_pages, true, &b->target);
 841
 842                                 /*
 843                                  * ignore errors from locking as we now switch
 844                                  * to 4k pages and we might get different
 845                                  * errors.
 846                                  */
 847
 848                                 num_pages = 0;
 849                                 is_2m_pages = false;
 850                                 continue;
 851                         }
 852
 853                         if (flags == VMW_PAGE_ALLOC_CANSLEEP) {
 854                                 /*
 855                                  * CANSLEEP page allocation failed, so guest
 856                                  * is under severe memory pressure. Quickly
 857                                  * decrease allocation rate.
 858                                  */
 859                                 b->rate_alloc = max(b->rate_alloc / 2,
 860                                                     VMW_BALLOON_RATE_ALLOC_MIN);
 861                                 STATS_INC(b->stats.sleep_alloc_fail);
 862                                 break;
 863                         }
 864
 865                         /*
 866                          * NOSLEEP page allocation failed, so the guest is
 867                          * under memory pressure. Let us slow down page
 868                          * allocations for next few cycles so that the guest
 869                          * gets out of memory pressure. Also, if we already
 870                          * allocated b->rate_alloc pages, let's pause,
 871                          * otherwise switch to sleeping allocations.
 872                          */
 873                         b->slow_allocation_cycles = VMW_BALLOON_SLOW_CYCLES;
 874
 875                         if (allocations >= b->rate_alloc)
 876                                 break;
 877
 878                         flags = VMW_PAGE_ALLOC_CANSLEEP;
 879                         /* Lower rate for sleeping allocations. */
 880                         rate = b->rate_alloc;
 881                         continue;
 882                 }
 883
 884                 b->ops->add_page(b, num_pages++, page);
 885                 if (num_pages == b->batch_max_pages) {
 886                         error = b->ops->lock(b, num_pages, is_2m_pages,
 887                                         &b->target);
 888                         num_pages = 0;
 889                         if (error)
 890                                 break;
 891                 }
 892
 893                 cond_resched();
 894
 895                 if (allocations >= rate) {
 896                         /* We allocated enough pages, let's take a break. */
 897                         break;
 898                 }
 899         }
 900
 901         if (num_pages > 0)
 902                 b->ops->lock(b, num_pages, is_2m_pages, &b->target);
 903
 904         /*
 905          * We reached our goal without failures so try increasing
 906          * allocation rate.
 907          */
 908         if (error == 0 && allocations >= b->rate_alloc) {
 909                 unsigned int mult = allocations / b->rate_alloc;
 910
 911                 b->rate_alloc =
 912                         min(b->rate_alloc + mult * VMW_BALLOON_RATE_ALLOC_INC,
 913                             VMW_BALLOON_RATE_ALLOC_MAX);
 914         }
 915
 916         vmballoon_release_refused_pages(b, true);
 917         vmballoon_release_refused_pages(b, false);
 918 }
 919
 920 /*
 921  * Decrease the size of the balloon allowing guest to use more memory.
 922  */
 923 static void vmballoon_deflate(struct vmballoon *b)
 924 {
 925         unsigned is_2m_pages;
 926
 927         pr_debug("%s - size: %d, target %d\n", __func__, b->size, b->target);
 928
 929         /* free pages to reach target */
 930         for (is_2m_pages = 0; is_2m_pages < b->supported_page_sizes;
 931                         is_2m_pages++) {
 932                 struct page *page, *next;
 933                 unsigned int num_pages = 0;
 934                 struct vmballoon_page_size *page_size =
 935                                 &b->page_sizes[is_2m_pages];
 936
 937                 list_for_each_entry_safe(page, next, &page_size->pages, lru) {
 938                         if (b->reset_required ||
 939                                 (b->target > 0 &&
 940                                         b->size - num_pages
 941                                         * vmballoon_page_size(is_2m_pages)
 942                                 < b->target + vmballoon_page_size(true)))
 943                                 break;
 944
 945                         list_del(&page->lru);
 946                         b->ops->add_page(b, num_pages++, page);
 947
 948                         if (num_pages == b->batch_max_pages) {
 949                                 int error;
 950
 951                                 error = b->ops->unlock(b, num_pages,
 952                                                 is_2m_pages, &b->target);
 953                                 num_pages = 0;
 954                                 if (error)
 955                                         return;
 956                         }
 957
 958                         cond_resched();
 959                 }
 960
 961                 if (num_pages > 0)
 962                         b->ops->unlock(b, num_pages, is_2m_pages, &b->target);
 963         }
 964 }
 965
 966 static const struct vmballoon_ops vmballoon_basic_ops = {
 967         .add_page = vmballoon_add_page,
 968         .lock = vmballoon_lock_page,
 969         .unlock = vmballoon_unlock_page
 970 };
 971
 972 static const struct vmballoon_ops vmballoon_batched_ops = {
 973         .add_page = vmballoon_add_batched_page,
 974         .lock = vmballoon_lock_batched_page,
 975         .unlock = vmballoon_unlock_batched_page
 976 };
 977
 978 static bool vmballoon_init_batching(struct vmballoon *b)
 979 {
 980         b->page = alloc_page(VMW_PAGE_ALLOC_NOSLEEP);
 981         if (!b->page)
 982                 return false;
 983
 984         b->batch_page = vmap(&b->page, 1, VM_MAP, PAGE_KERNEL);
 985         if (!b->batch_page) {
 986                 __free_page(b->page);
 987                 return false;
 988         }
 989
 990         return true;
 991 }
 992
 993 /*
 994  * Perform standard reset sequence by popping the balloon (in case it
 995  * is not  empty) and then restarting protocol. This operation normally
 996  * happens when host responds with VMW_BALLOON_ERROR_RESET to a command.
 997  */
 998 static void vmballoon_reset(struct vmballoon *b)
 999 {
1000         /* free all pages, skipping monitor unlock */
1001         vmballoon_pop(b);
1002
1003         if (!vmballoon_send_start(b, VMW_BALLOON_CAPABILITIES))
1004                 return;
1005
1006         if ((b->capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1007                 b->ops = &vmballoon_batched_ops;
1008                 b->batch_max_pages = VMW_BALLOON_BATCH_MAX_PAGES;
1009                 if (!vmballoon_init_batching(b)) {
1010                         /*
1011                          * We failed to initialize batching, inform the monitor
1012                          * about it by sending a null capability.
1013                          *
1014                          * The guest will retry in one second.
1015                          */
1016                         vmballoon_send_start(b, 0);
1017                         return;
1018                 }
1019         } else if ((b->capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1020                 b->ops = &vmballoon_basic_ops;
1021                 b->batch_max_pages = 1;
1022         }
1023
1024         b->reset_required = false;
1025         if (!vmballoon_send_guest_id(b))
1026                 pr_err("failed to send guest ID to the host\n");
1027 }
1028
1029 /*
1030  * Balloon work function: reset protocol, if needed, get the new size and
1031  * adjust balloon as needed. Repeat in 1 sec.
1032  */
1033 static void vmballoon_work(struct work_struct *work)
1034 {
1035         struct delayed_work *dwork = to_delayed_work(work);
1036         struct vmballoon *b = container_of(dwork, struct vmballoon, dwork);
1037         unsigned int target;
1038
1039         STATS_INC(b->stats.timer);
1040
1041         if (b->reset_required)
1042                 vmballoon_reset(b);
1043
1044         if (b->slow_allocation_cycles > 0)
1045                 b->slow_allocation_cycles--;
1046
1047         if (vmballoon_send_get_target(b, &target)) {
1048                 /* update target, adjust size */
1049                 b->target = target;
1050
1051                 if (b->size < target)
1052                         vmballoon_inflate(b);
1053                 else if (target == 0 ||
1054                                 b->size > target + vmballoon_page_size(true))
1055                         vmballoon_deflate(b);
1056         }
1057
1058         /*
1059          * We are using a freezable workqueue so that balloon operations are
1060          * stopped while the system transitions to/from sleep/hibernation.
1061          */
1062         queue_delayed_work(system_freezable_wq,
1063                            dwork, round_jiffies_relative(HZ));
1064 }
1065
1066 /*
1067  * DEBUGFS Interface
1068  */
1069 #ifdef CONFIG_DEBUG_FS
1070
1071 static int vmballoon_debug_show(struct seq_file *f, void *offset)
1072 {
1073         struct vmballoon *b = f->private;
1074         struct vmballoon_stats *stats = &b->stats;
1075
1076         /* format capabilities info */
1077         seq_printf(f,
1078                    "balloon capabilities:   %#4x\n"
1079                    "used capabilities:      %#4lx\n",
1080                    VMW_BALLOON_CAPABILITIES, b->capabilities);
1081
1082         /* format size info */
1083         seq_printf(f,
1084                    "target:             %8d pages\n"
1085                    "current:            %8d pages\n",
1086                    b->target, b->size);
1087
1088         /* format rate info */
1089         seq_printf(f,
1090                    "rateSleepAlloc:     %8d pages/sec\n",
1091                    b->rate_alloc);
1092
1093         seq_printf(f,
1094                    "\n"
1095                    "timer:              %8u\n"
1096                    "start:              %8u (%4u failed)\n"
1097                    "guestType:          %8u (%4u failed)\n"
1098                    "2m-lock:            %8u (%4u failed)\n"
1099                    "lock:               %8u (%4u failed)\n"
1100                    "2m-unlock:          %8u (%4u failed)\n"
1101                    "unlock:             %8u (%4u failed)\n"
1102                    "target:             %8u (%4u failed)\n"
1103                    "prim2mAlloc:        %8u (%4u failed)\n"
1104                    "primNoSleepAlloc:   %8u (%4u failed)\n"
1105                    "primCanSleepAlloc:  %8u (%4u failed)\n"
1106                    "prim2mFree:         %8u\n"
1107                    "primFree:           %8u\n"
1108                    "err2mAlloc:         %8u\n"
1109                    "errAlloc:           %8u\n"
1110                    "err2mFree:          %8u\n"
1111                    "errFree:            %8u\n",
1112                    stats->timer,
1113                    stats->start, stats->start_fail,
1114                    stats->guest_type, stats->guest_type_fail,
1115                    stats->lock[true],  stats->lock_fail[true],
1116                    stats->lock[false],  stats->lock_fail[false],
1117                    stats->unlock[true], stats->unlock_fail[true],
1118                    stats->unlock[false], stats->unlock_fail[false],
1119                    stats->target, stats->target_fail,
1120                    stats->alloc[true], stats->alloc_fail[true],
1121                    stats->alloc[false], stats->alloc_fail[false],
1122                    stats->sleep_alloc, stats->sleep_alloc_fail,
1123                    stats->free[true],
1124                    stats->free[false],
1125                    stats->refused_alloc[true], stats->refused_alloc[false],
1126                    stats->refused_free[true], stats->refused_free[false]);
1127
1128         return 0;
1129 }
1130
1131 static int vmballoon_debug_open(struct inode *inode, struct file *file)
1132 {
1133         return single_open(file, vmballoon_debug_show, inode->i_private);
1134 }
1135
1136 static const struct file_operations vmballoon_debug_fops = {
1137         .owner          = THIS_MODULE,
1138         .open           = vmballoon_debug_open,
1139         .read           = seq_read,
1140         .llseek         = seq_lseek,
1141         .release        = single_release,
1142 };
1143
1144 static int __init vmballoon_debugfs_init(struct vmballoon *b)
1145 {
1146         int error;
1147
1148         b->dbg_entry = debugfs_create_file("vmmemctl", S_IRUGO, NULL, b,
1149                                            &vmballoon_debug_fops);
1150         if (IS_ERR(b->dbg_entry)) {
1151                 error = PTR_ERR(b->dbg_entry);
1152                 pr_err("failed to create debugfs entry, error: %d\n", error);
1153                 return error;
1154         }
1155
1156         return 0;
1157 }
1158
1159 static void __exit vmballoon_debugfs_exit(struct vmballoon *b)
1160 {
1161         debugfs_remove(b->dbg_entry);
1162 }
1163
1164 #else
1165
1166 static inline int vmballoon_debugfs_init(struct vmballoon *b)
1167 {
1168         return 0;
1169 }
1170
1171 static inline void vmballoon_debugfs_exit(struct vmballoon *b)
1172 {
1173 }
1174
1175 #endif  /* CONFIG_DEBUG_FS */
1176
1177 static int __init vmballoon_init(void)
1178 {
1179         int error;
1180         unsigned is_2m_pages;
1181         /*
1182          * Check if we are running on VMware's hypervisor and bail out
1183          * if we are not.
1184          */
1185         if (x86_hyper != &x86_hyper_vmware)
1186                 return -ENODEV;
1187
1188         for (is_2m_pages = 0; is_2m_pages < VMW_BALLOON_NUM_PAGE_SIZES;
1189                         is_2m_pages++) {
1190                 INIT_LIST_HEAD(&balloon.page_sizes[is_2m_pages].pages);
1191                 INIT_LIST_HEAD(&balloon.page_sizes[is_2m_pages].refused_pages);
1192         }
1193
1194         /* initialize rates */
1195         balloon.rate_alloc = VMW_BALLOON_RATE_ALLOC_MAX;
1196
1197         INIT_DELAYED_WORK(&balloon.dwork, vmballoon_work);
1198
1199         /*
1200          * Start balloon.
1201          */
1202         if (!vmballoon_send_start(&balloon, VMW_BALLOON_CAPABILITIES)) {
1203                 pr_err("failed to send start command to the host\n");
1204                 return -EIO;
1205         }
1206
1207         if ((balloon.capabilities & VMW_BALLOON_BATCHED_CMDS) != 0) {
1208                 balloon.ops = &vmballoon_batched_ops;
1209                 balloon.batch_max_pages = VMW_BALLOON_BATCH_MAX_PAGES;
1210                 if (!vmballoon_init_batching(&balloon)) {
1211                         pr_err("failed to init batching\n");
1212                         return -EIO;
1213                 }
1214         } else if ((balloon.capabilities & VMW_BALLOON_BASIC_CMDS) != 0) {
1215                 balloon.ops = &vmballoon_basic_ops;
1216                 balloon.batch_max_pages = 1;
1217         }
1218
1219         if (!vmballoon_send_guest_id(&balloon)) {
1220                 pr_err("failed to send guest ID to the host\n");
1221                 return -EIO;
1222         }
1223
1224         error = vmballoon_debugfs_init(&balloon);
1225         if (error)
1226                 return error;
1227
1228         queue_delayed_work(system_freezable_wq, &balloon.dwork, 0);
1229
1230         return 0;
1231 }
1232 module_init(vmballoon_init);
1233
1234 static void __exit vmballoon_exit(void)
1235 {
1236         cancel_delayed_work_sync(&balloon.dwork);
1237
1238         vmballoon_debugfs_exit(&balloon);
1239
1240         /*
1241          * Deallocate all reserved memory, and reset connection with monitor.
1242          * Reset connection before deallocating memory to avoid potential for
1243          * additional spurious resets from guest touching deallocated pages.
1244          */
1245         vmballoon_send_start(&balloon, VMW_BALLOON_CAPABILITIES);
1246         vmballoon_pop(&balloon);
1247 }
1248 module_exit(vmballoon_exit);