Linux Audio

Check our new training course

Loading...
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * Virtio-mem device driver.
   4 *
   5 * Copyright Red Hat, Inc. 2020
   6 *
   7 * Author(s): David Hildenbrand <david@redhat.com>
   8 */
   9
  10#include <linux/virtio.h>
  11#include <linux/virtio_mem.h>
  12#include <linux/workqueue.h>
  13#include <linux/slab.h>
  14#include <linux/module.h>
  15#include <linux/mm.h>
  16#include <linux/memory_hotplug.h>
  17#include <linux/memory.h>
  18#include <linux/hrtimer.h>
  19#include <linux/crash_dump.h>
  20#include <linux/mutex.h>
  21#include <linux/bitmap.h>
  22#include <linux/lockdep.h>
  23#include <linux/log2.h>
  24
  25#include <acpi/acpi_numa.h>
  26
  27static bool unplug_online = true;
  28module_param(unplug_online, bool, 0644);
  29MODULE_PARM_DESC(unplug_online, "Try to unplug online memory");
  30
  31static bool force_bbm;
  32module_param(force_bbm, bool, 0444);
  33MODULE_PARM_DESC(force_bbm,
  34		"Force Big Block Mode. Default is 0 (auto-selection)");
  35
  36static unsigned long bbm_block_size;
  37module_param(bbm_block_size, ulong, 0444);
  38MODULE_PARM_DESC(bbm_block_size,
  39		 "Big Block size in bytes. Default is 0 (auto-detection).");
  40
 
 
 
 
 
  41/*
  42 * virtio-mem currently supports the following modes of operation:
  43 *
  44 * * Sub Block Mode (SBM): A Linux memory block spans 2..X subblocks (SB). The
  45 *   size of a Sub Block (SB) is determined based on the device block size, the
  46 *   pageblock size, and the maximum allocation granularity of the buddy.
  47 *   Subblocks within a Linux memory block might either be plugged or unplugged.
  48 *   Memory is added/removed to Linux MM in Linux memory block granularity.
  49 *
  50 * * Big Block Mode (BBM): A Big Block (BB) spans 1..X Linux memory blocks.
  51 *   Memory is added/removed to Linux MM in Big Block granularity.
  52 *
  53 * The mode is determined automatically based on the Linux memory block size
  54 * and the device block size.
  55 *
  56 * User space / core MM (auto onlining) is responsible for onlining added
  57 * Linux memory blocks - and for selecting a zone. Linux Memory Blocks are
  58 * always onlined separately, and all memory within a Linux memory block is
  59 * onlined to the same zone - virtio-mem relies on this behavior.
  60 */
  61
  62/*
  63 * State of a Linux memory block in SBM.
  64 */
  65enum virtio_mem_sbm_mb_state {
  66	/* Unplugged, not added to Linux. Can be reused later. */
  67	VIRTIO_MEM_SBM_MB_UNUSED = 0,
  68	/* (Partially) plugged, not added to Linux. Error on add_memory(). */
  69	VIRTIO_MEM_SBM_MB_PLUGGED,
  70	/* Fully plugged, fully added to Linux, offline. */
  71	VIRTIO_MEM_SBM_MB_OFFLINE,
  72	/* Partially plugged, fully added to Linux, offline. */
  73	VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
  74	/* Fully plugged, fully added to Linux, onlined to a kernel zone. */
  75	VIRTIO_MEM_SBM_MB_KERNEL,
  76	/* Partially plugged, fully added to Linux, online to a kernel zone */
  77	VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
  78	/* Fully plugged, fully added to Linux, onlined to ZONE_MOVABLE. */
  79	VIRTIO_MEM_SBM_MB_MOVABLE,
  80	/* Partially plugged, fully added to Linux, onlined to ZONE_MOVABLE. */
  81	VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
  82	VIRTIO_MEM_SBM_MB_COUNT
  83};
  84
  85/*
  86 * State of a Big Block (BB) in BBM, covering 1..X Linux memory blocks.
  87 */
  88enum virtio_mem_bbm_bb_state {
  89	/* Unplugged, not added to Linux. Can be reused later. */
  90	VIRTIO_MEM_BBM_BB_UNUSED = 0,
  91	/* Plugged, not added to Linux. Error on add_memory(). */
  92	VIRTIO_MEM_BBM_BB_PLUGGED,
  93	/* Plugged and added to Linux. */
  94	VIRTIO_MEM_BBM_BB_ADDED,
  95	/* All online parts are fake-offline, ready to remove. */
  96	VIRTIO_MEM_BBM_BB_FAKE_OFFLINE,
  97	VIRTIO_MEM_BBM_BB_COUNT
  98};
  99
 100struct virtio_mem {
 101	struct virtio_device *vdev;
 102
 103	/* We might first have to unplug all memory when starting up. */
 104	bool unplug_all_required;
 105
 106	/* Workqueue that processes the plug/unplug requests. */
 107	struct work_struct wq;
 108	atomic_t wq_active;
 109	atomic_t config_changed;
 110
 111	/* Virtqueue for guest->host requests. */
 112	struct virtqueue *vq;
 113
 114	/* Wait for a host response to a guest request. */
 115	wait_queue_head_t host_resp;
 116
 117	/* Space for one guest request and the host response. */
 118	struct virtio_mem_req req;
 119	struct virtio_mem_resp resp;
 120
 121	/* The current size of the device. */
 122	uint64_t plugged_size;
 123	/* The requested size of the device. */
 124	uint64_t requested_size;
 125
 126	/* The device block size (for communicating with the device). */
 127	uint64_t device_block_size;
 128	/* The determined node id for all memory of the device. */
 129	int nid;
 130	/* Physical start address of the memory region. */
 131	uint64_t addr;
 132	/* Maximum region size in bytes. */
 133	uint64_t region_size;
 134
 135	/* The parent resource for all memory added via this device. */
 136	struct resource *parent_resource;
 137	/*
 138	 * Copy of "System RAM (virtio_mem)" to be used for
 139	 * add_memory_driver_managed().
 140	 */
 141	const char *resource_name;
 142	/* Memory group identification. */
 143	int mgid;
 144
 145	/*
 146	 * We don't want to add too much memory if it's not getting onlined,
 147	 * to avoid running OOM. Besides this threshold, we allow to have at
 148	 * least two offline blocks at a time (whatever is bigger).
 149	 */
 150#define VIRTIO_MEM_DEFAULT_OFFLINE_THRESHOLD		(1024 * 1024 * 1024)
 151	atomic64_t offline_size;
 152	uint64_t offline_threshold;
 153
 154	/* If set, the driver is in SBM, otherwise in BBM. */
 155	bool in_sbm;
 156
 157	union {
 158		struct {
 159			/* Id of the first memory block of this device. */
 160			unsigned long first_mb_id;
 161			/* Id of the last usable memory block of this device. */
 162			unsigned long last_usable_mb_id;
 163			/* Id of the next memory bock to prepare when needed. */
 164			unsigned long next_mb_id;
 165
 166			/* The subblock size. */
 167			uint64_t sb_size;
 168			/* The number of subblocks per Linux memory block. */
 169			uint32_t sbs_per_mb;
 170
 171			/*
 172			 * Some of the Linux memory blocks tracked as "partially
 173			 * plugged" are completely unplugged and can be offlined
 174			 * and removed -- which previously failed.
 175			 */
 176			bool have_unplugged_mb;
 177
 178			/* Summary of all memory block states. */
 179			unsigned long mb_count[VIRTIO_MEM_SBM_MB_COUNT];
 180
 181			/*
 182			 * One byte state per memory block. Allocated via
 183			 * vmalloc(). Resized (alloc+copy+free) on demand.
 184			 *
 185			 * With 128 MiB memory blocks, we have states for 512
 186			 * GiB of memory in one 4 KiB page.
 187			 */
 188			uint8_t *mb_states;
 189
 190			/*
 191			 * Bitmap: one bit per subblock. Allocated similar to
 192			 * sbm.mb_states.
 193			 *
 194			 * A set bit means the corresponding subblock is
 195			 * plugged, otherwise it's unblocked.
 196			 *
 197			 * With 4 MiB subblocks, we manage 128 GiB of memory
 198			 * in one 4 KiB page.
 199			 */
 200			unsigned long *sb_states;
 201		} sbm;
 202
 203		struct {
 204			/* Id of the first big block of this device. */
 205			unsigned long first_bb_id;
 206			/* Id of the last usable big block of this device. */
 207			unsigned long last_usable_bb_id;
 208			/* Id of the next device bock to prepare when needed. */
 209			unsigned long next_bb_id;
 210
 211			/* Summary of all big block states. */
 212			unsigned long bb_count[VIRTIO_MEM_BBM_BB_COUNT];
 213
 214			/* One byte state per big block. See sbm.mb_states. */
 215			uint8_t *bb_states;
 216
 217			/* The block size used for plugging/adding/removing. */
 218			uint64_t bb_size;
 219		} bbm;
 220	};
 221
 222	/*
 223	 * Mutex that protects the sbm.mb_count, sbm.mb_states,
 224	 * sbm.sb_states, bbm.bb_count, and bbm.bb_states
 225	 *
 226	 * When this lock is held the pointers can't change, ONLINE and
 227	 * OFFLINE blocks can't change the state and no subblocks will get
 228	 * plugged/unplugged.
 229	 *
 230	 * In kdump mode, used to serialize requests, last_block_addr and
 231	 * last_block_plugged.
 232	 */
 233	struct mutex hotplug_mutex;
 234	bool hotplug_active;
 235
 236	/* An error occurred we cannot handle - stop processing requests. */
 237	bool broken;
 238
 239	/* Cached valued of is_kdump_kernel() when the device was probed. */
 240	bool in_kdump;
 241
 242	/* The driver is being removed. */
 243	spinlock_t removal_lock;
 244	bool removing;
 245
 246	/* Timer for retrying to plug/unplug memory. */
 247	struct hrtimer retry_timer;
 248	unsigned int retry_timer_ms;
 249#define VIRTIO_MEM_RETRY_TIMER_MIN_MS		50000
 250#define VIRTIO_MEM_RETRY_TIMER_MAX_MS		300000
 251
 252	/* Memory notifier (online/offline events). */
 253	struct notifier_block memory_notifier;
 254
 255#ifdef CONFIG_PROC_VMCORE
 256	/* vmcore callback for /proc/vmcore handling in kdump mode */
 257	struct vmcore_cb vmcore_cb;
 258	uint64_t last_block_addr;
 259	bool last_block_plugged;
 260#endif /* CONFIG_PROC_VMCORE */
 261
 262	/* Next device in the list of virtio-mem devices. */
 263	struct list_head next;
 264};
 265
 266/*
 267 * We have to share a single online_page callback among all virtio-mem
 268 * devices. We use RCU to iterate the list in the callback.
 269 */
 270static DEFINE_MUTEX(virtio_mem_mutex);
 271static LIST_HEAD(virtio_mem_devices);
 272
 273static void virtio_mem_online_page_cb(struct page *page, unsigned int order);
 274static void virtio_mem_fake_offline_going_offline(unsigned long pfn,
 275						  unsigned long nr_pages);
 276static void virtio_mem_fake_offline_cancel_offline(unsigned long pfn,
 277						   unsigned long nr_pages);
 278static void virtio_mem_retry(struct virtio_mem *vm);
 279static int virtio_mem_create_resource(struct virtio_mem *vm);
 280static void virtio_mem_delete_resource(struct virtio_mem *vm);
 281
 282/*
 283 * Register a virtio-mem device so it will be considered for the online_page
 284 * callback.
 285 */
 286static int register_virtio_mem_device(struct virtio_mem *vm)
 287{
 288	int rc = 0;
 289
 290	/* First device registers the callback. */
 291	mutex_lock(&virtio_mem_mutex);
 292	if (list_empty(&virtio_mem_devices))
 293		rc = set_online_page_callback(&virtio_mem_online_page_cb);
 294	if (!rc)
 295		list_add_rcu(&vm->next, &virtio_mem_devices);
 296	mutex_unlock(&virtio_mem_mutex);
 297
 298	return rc;
 299}
 300
 301/*
 302 * Unregister a virtio-mem device so it will no longer be considered for the
 303 * online_page callback.
 304 */
 305static void unregister_virtio_mem_device(struct virtio_mem *vm)
 306{
 307	/* Last device unregisters the callback. */
 308	mutex_lock(&virtio_mem_mutex);
 309	list_del_rcu(&vm->next);
 310	if (list_empty(&virtio_mem_devices))
 311		restore_online_page_callback(&virtio_mem_online_page_cb);
 312	mutex_unlock(&virtio_mem_mutex);
 313
 314	synchronize_rcu();
 315}
 316
 317/*
 318 * Calculate the memory block id of a given address.
 319 */
 320static unsigned long virtio_mem_phys_to_mb_id(unsigned long addr)
 321{
 322	return addr / memory_block_size_bytes();
 323}
 324
 325/*
 326 * Calculate the physical start address of a given memory block id.
 327 */
 328static unsigned long virtio_mem_mb_id_to_phys(unsigned long mb_id)
 329{
 330	return mb_id * memory_block_size_bytes();
 331}
 332
 333/*
 334 * Calculate the big block id of a given address.
 335 */
 336static unsigned long virtio_mem_phys_to_bb_id(struct virtio_mem *vm,
 337					      uint64_t addr)
 338{
 339	return addr / vm->bbm.bb_size;
 340}
 341
 342/*
 343 * Calculate the physical start address of a given big block id.
 344 */
 345static uint64_t virtio_mem_bb_id_to_phys(struct virtio_mem *vm,
 346					 unsigned long bb_id)
 347{
 348	return bb_id * vm->bbm.bb_size;
 349}
 350
 351/*
 352 * Calculate the subblock id of a given address.
 353 */
 354static unsigned long virtio_mem_phys_to_sb_id(struct virtio_mem *vm,
 355					      unsigned long addr)
 356{
 357	const unsigned long mb_id = virtio_mem_phys_to_mb_id(addr);
 358	const unsigned long mb_addr = virtio_mem_mb_id_to_phys(mb_id);
 359
 360	return (addr - mb_addr) / vm->sbm.sb_size;
 361}
 362
 363/*
 364 * Set the state of a big block, taking care of the state counter.
 365 */
 366static void virtio_mem_bbm_set_bb_state(struct virtio_mem *vm,
 367					unsigned long bb_id,
 368					enum virtio_mem_bbm_bb_state state)
 369{
 370	const unsigned long idx = bb_id - vm->bbm.first_bb_id;
 371	enum virtio_mem_bbm_bb_state old_state;
 372
 373	old_state = vm->bbm.bb_states[idx];
 374	vm->bbm.bb_states[idx] = state;
 375
 376	BUG_ON(vm->bbm.bb_count[old_state] == 0);
 377	vm->bbm.bb_count[old_state]--;
 378	vm->bbm.bb_count[state]++;
 379}
 380
 381/*
 382 * Get the state of a big block.
 383 */
 384static enum virtio_mem_bbm_bb_state virtio_mem_bbm_get_bb_state(struct virtio_mem *vm,
 385								unsigned long bb_id)
 386{
 387	return vm->bbm.bb_states[bb_id - vm->bbm.first_bb_id];
 388}
 389
 390/*
 391 * Prepare the big block state array for the next big block.
 392 */
 393static int virtio_mem_bbm_bb_states_prepare_next_bb(struct virtio_mem *vm)
 394{
 395	unsigned long old_bytes = vm->bbm.next_bb_id - vm->bbm.first_bb_id;
 396	unsigned long new_bytes = old_bytes + 1;
 397	int old_pages = PFN_UP(old_bytes);
 398	int new_pages = PFN_UP(new_bytes);
 399	uint8_t *new_array;
 400
 401	if (vm->bbm.bb_states && old_pages == new_pages)
 402		return 0;
 403
 404	new_array = vzalloc(new_pages * PAGE_SIZE);
 405	if (!new_array)
 406		return -ENOMEM;
 407
 408	mutex_lock(&vm->hotplug_mutex);
 409	if (vm->bbm.bb_states)
 410		memcpy(new_array, vm->bbm.bb_states, old_pages * PAGE_SIZE);
 411	vfree(vm->bbm.bb_states);
 412	vm->bbm.bb_states = new_array;
 413	mutex_unlock(&vm->hotplug_mutex);
 414
 415	return 0;
 416}
 417
 418#define virtio_mem_bbm_for_each_bb(_vm, _bb_id, _state) \
 419	for (_bb_id = vm->bbm.first_bb_id; \
 420	     _bb_id < vm->bbm.next_bb_id && _vm->bbm.bb_count[_state]; \
 421	     _bb_id++) \
 422		if (virtio_mem_bbm_get_bb_state(_vm, _bb_id) == _state)
 423
 424#define virtio_mem_bbm_for_each_bb_rev(_vm, _bb_id, _state) \
 425	for (_bb_id = vm->bbm.next_bb_id - 1; \
 426	     _bb_id >= vm->bbm.first_bb_id && _vm->bbm.bb_count[_state]; \
 427	     _bb_id--) \
 428		if (virtio_mem_bbm_get_bb_state(_vm, _bb_id) == _state)
 429
 430/*
 431 * Set the state of a memory block, taking care of the state counter.
 432 */
 433static void virtio_mem_sbm_set_mb_state(struct virtio_mem *vm,
 434					unsigned long mb_id, uint8_t state)
 435{
 436	const unsigned long idx = mb_id - vm->sbm.first_mb_id;
 437	uint8_t old_state;
 438
 439	old_state = vm->sbm.mb_states[idx];
 440	vm->sbm.mb_states[idx] = state;
 441
 442	BUG_ON(vm->sbm.mb_count[old_state] == 0);
 443	vm->sbm.mb_count[old_state]--;
 444	vm->sbm.mb_count[state]++;
 445}
 446
 447/*
 448 * Get the state of a memory block.
 449 */
 450static uint8_t virtio_mem_sbm_get_mb_state(struct virtio_mem *vm,
 451					   unsigned long mb_id)
 452{
 453	const unsigned long idx = mb_id - vm->sbm.first_mb_id;
 454
 455	return vm->sbm.mb_states[idx];
 456}
 457
 458/*
 459 * Prepare the state array for the next memory block.
 460 */
 461static int virtio_mem_sbm_mb_states_prepare_next_mb(struct virtio_mem *vm)
 462{
 463	int old_pages = PFN_UP(vm->sbm.next_mb_id - vm->sbm.first_mb_id);
 464	int new_pages = PFN_UP(vm->sbm.next_mb_id - vm->sbm.first_mb_id + 1);
 465	uint8_t *new_array;
 466
 467	if (vm->sbm.mb_states && old_pages == new_pages)
 468		return 0;
 469
 470	new_array = vzalloc(new_pages * PAGE_SIZE);
 471	if (!new_array)
 472		return -ENOMEM;
 473
 474	mutex_lock(&vm->hotplug_mutex);
 475	if (vm->sbm.mb_states)
 476		memcpy(new_array, vm->sbm.mb_states, old_pages * PAGE_SIZE);
 477	vfree(vm->sbm.mb_states);
 478	vm->sbm.mb_states = new_array;
 479	mutex_unlock(&vm->hotplug_mutex);
 480
 481	return 0;
 482}
 483
 484#define virtio_mem_sbm_for_each_mb(_vm, _mb_id, _state) \
 485	for (_mb_id = _vm->sbm.first_mb_id; \
 486	     _mb_id < _vm->sbm.next_mb_id && _vm->sbm.mb_count[_state]; \
 487	     _mb_id++) \
 488		if (virtio_mem_sbm_get_mb_state(_vm, _mb_id) == _state)
 489
 490#define virtio_mem_sbm_for_each_mb_rev(_vm, _mb_id, _state) \
 491	for (_mb_id = _vm->sbm.next_mb_id - 1; \
 492	     _mb_id >= _vm->sbm.first_mb_id && _vm->sbm.mb_count[_state]; \
 493	     _mb_id--) \
 494		if (virtio_mem_sbm_get_mb_state(_vm, _mb_id) == _state)
 495
 496/*
 497 * Calculate the bit number in the subblock bitmap for the given subblock
 498 * inside the given memory block.
 499 */
 500static int virtio_mem_sbm_sb_state_bit_nr(struct virtio_mem *vm,
 501					  unsigned long mb_id, int sb_id)
 502{
 503	return (mb_id - vm->sbm.first_mb_id) * vm->sbm.sbs_per_mb + sb_id;
 504}
 505
 506/*
 507 * Mark all selected subblocks plugged.
 508 *
 509 * Will not modify the state of the memory block.
 510 */
 511static void virtio_mem_sbm_set_sb_plugged(struct virtio_mem *vm,
 512					  unsigned long mb_id, int sb_id,
 513					  int count)
 514{
 515	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 516
 517	__bitmap_set(vm->sbm.sb_states, bit, count);
 518}
 519
 520/*
 521 * Mark all selected subblocks unplugged.
 522 *
 523 * Will not modify the state of the memory block.
 524 */
 525static void virtio_mem_sbm_set_sb_unplugged(struct virtio_mem *vm,
 526					    unsigned long mb_id, int sb_id,
 527					    int count)
 528{
 529	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 530
 531	__bitmap_clear(vm->sbm.sb_states, bit, count);
 532}
 533
 534/*
 535 * Test if all selected subblocks are plugged.
 536 */
 537static bool virtio_mem_sbm_test_sb_plugged(struct virtio_mem *vm,
 538					   unsigned long mb_id, int sb_id,
 539					   int count)
 540{
 541	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 542
 543	if (count == 1)
 544		return test_bit(bit, vm->sbm.sb_states);
 545
 546	/* TODO: Helper similar to bitmap_set() */
 547	return find_next_zero_bit(vm->sbm.sb_states, bit + count, bit) >=
 548	       bit + count;
 549}
 550
 551/*
 552 * Test if all selected subblocks are unplugged.
 553 */
 554static bool virtio_mem_sbm_test_sb_unplugged(struct virtio_mem *vm,
 555					     unsigned long mb_id, int sb_id,
 556					     int count)
 557{
 558	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 559
 560	/* TODO: Helper similar to bitmap_set() */
 561	return find_next_bit(vm->sbm.sb_states, bit + count, bit) >=
 562	       bit + count;
 563}
 564
 565/*
 566 * Find the first unplugged subblock. Returns vm->sbm.sbs_per_mb in case there is
 567 * none.
 568 */
 569static int virtio_mem_sbm_first_unplugged_sb(struct virtio_mem *vm,
 570					    unsigned long mb_id)
 571{
 572	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, 0);
 573
 574	return find_next_zero_bit(vm->sbm.sb_states,
 575				  bit + vm->sbm.sbs_per_mb, bit) - bit;
 576}
 577
 578/*
 579 * Prepare the subblock bitmap for the next memory block.
 580 */
 581static int virtio_mem_sbm_sb_states_prepare_next_mb(struct virtio_mem *vm)
 582{
 583	const unsigned long old_nb_mb = vm->sbm.next_mb_id - vm->sbm.first_mb_id;
 584	const unsigned long old_nb_bits = old_nb_mb * vm->sbm.sbs_per_mb;
 585	const unsigned long new_nb_bits = (old_nb_mb + 1) * vm->sbm.sbs_per_mb;
 586	int old_pages = PFN_UP(BITS_TO_LONGS(old_nb_bits) * sizeof(long));
 587	int new_pages = PFN_UP(BITS_TO_LONGS(new_nb_bits) * sizeof(long));
 588	unsigned long *new_bitmap, *old_bitmap;
 589
 590	if (vm->sbm.sb_states && old_pages == new_pages)
 591		return 0;
 592
 593	new_bitmap = vzalloc(new_pages * PAGE_SIZE);
 594	if (!new_bitmap)
 595		return -ENOMEM;
 596
 597	mutex_lock(&vm->hotplug_mutex);
 598	if (vm->sbm.sb_states)
 599		memcpy(new_bitmap, vm->sbm.sb_states, old_pages * PAGE_SIZE);
 600
 601	old_bitmap = vm->sbm.sb_states;
 602	vm->sbm.sb_states = new_bitmap;
 603	mutex_unlock(&vm->hotplug_mutex);
 604
 605	vfree(old_bitmap);
 606	return 0;
 607}
 608
 609/*
 610 * Test if we could add memory without creating too much offline memory -
 611 * to avoid running OOM if memory is getting onlined deferred.
 612 */
 613static bool virtio_mem_could_add_memory(struct virtio_mem *vm, uint64_t size)
 614{
 615	if (WARN_ON_ONCE(size > vm->offline_threshold))
 616		return false;
 617
 618	return atomic64_read(&vm->offline_size) + size <= vm->offline_threshold;
 619}
 620
 621/*
 622 * Try adding memory to Linux. Will usually only fail if out of memory.
 623 *
 624 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 625 * onlining code).
 626 *
 627 * Will not modify the state of memory blocks in virtio-mem.
 628 */
 629static int virtio_mem_add_memory(struct virtio_mem *vm, uint64_t addr,
 630				 uint64_t size)
 631{
 632	int rc;
 633
 634	/*
 635	 * When force-unloading the driver and we still have memory added to
 636	 * Linux, the resource name has to stay.
 637	 */
 638	if (!vm->resource_name) {
 639		vm->resource_name = kstrdup_const("System RAM (virtio_mem)",
 640						  GFP_KERNEL);
 641		if (!vm->resource_name)
 642			return -ENOMEM;
 643	}
 644
 645	dev_dbg(&vm->vdev->dev, "adding memory: 0x%llx - 0x%llx\n", addr,
 646		addr + size - 1);
 647	/* Memory might get onlined immediately. */
 648	atomic64_add(size, &vm->offline_size);
 649	rc = add_memory_driver_managed(vm->mgid, addr, size, vm->resource_name,
 650				       MHP_MERGE_RESOURCE | MHP_NID_IS_MGID);
 651	if (rc) {
 652		atomic64_sub(size, &vm->offline_size);
 653		dev_warn(&vm->vdev->dev, "adding memory failed: %d\n", rc);
 654		/*
 655		 * TODO: Linux MM does not properly clean up yet in all cases
 656		 * where adding of memory failed - especially on -ENOMEM.
 657		 */
 658	}
 659	return rc;
 660}
 661
 662/*
 663 * See virtio_mem_add_memory(): Try adding a single Linux memory block.
 664 */
 665static int virtio_mem_sbm_add_mb(struct virtio_mem *vm, unsigned long mb_id)
 666{
 667	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 668	const uint64_t size = memory_block_size_bytes();
 669
 670	return virtio_mem_add_memory(vm, addr, size);
 671}
 672
 673/*
 674 * See virtio_mem_add_memory(): Try adding a big block.
 675 */
 676static int virtio_mem_bbm_add_bb(struct virtio_mem *vm, unsigned long bb_id)
 677{
 678	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
 679	const uint64_t size = vm->bbm.bb_size;
 680
 681	return virtio_mem_add_memory(vm, addr, size);
 682}
 683
 684/*
 685 * Try removing memory from Linux. Will only fail if memory blocks aren't
 686 * offline.
 687 *
 688 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 689 * onlining code).
 690 *
 691 * Will not modify the state of memory blocks in virtio-mem.
 692 */
 693static int virtio_mem_remove_memory(struct virtio_mem *vm, uint64_t addr,
 694				    uint64_t size)
 695{
 696	int rc;
 697
 698	dev_dbg(&vm->vdev->dev, "removing memory: 0x%llx - 0x%llx\n", addr,
 699		addr + size - 1);
 700	rc = remove_memory(addr, size);
 701	if (!rc) {
 702		atomic64_sub(size, &vm->offline_size);
 703		/*
 704		 * We might have freed up memory we can now unplug, retry
 705		 * immediately instead of waiting.
 706		 */
 707		virtio_mem_retry(vm);
 708	} else {
 709		dev_dbg(&vm->vdev->dev, "removing memory failed: %d\n", rc);
 710	}
 711	return rc;
 712}
 713
 714/*
 715 * See virtio_mem_remove_memory(): Try removing a single Linux memory block.
 716 */
 717static int virtio_mem_sbm_remove_mb(struct virtio_mem *vm, unsigned long mb_id)
 718{
 719	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 720	const uint64_t size = memory_block_size_bytes();
 721
 722	return virtio_mem_remove_memory(vm, addr, size);
 723}
 724
 725/*
 726 * Try offlining and removing memory from Linux.
 727 *
 728 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 729 * onlining code).
 730 *
 731 * Will not modify the state of memory blocks in virtio-mem.
 732 */
 733static int virtio_mem_offline_and_remove_memory(struct virtio_mem *vm,
 734						uint64_t addr,
 735						uint64_t size)
 736{
 737	int rc;
 738
 739	dev_dbg(&vm->vdev->dev,
 740		"offlining and removing memory: 0x%llx - 0x%llx\n", addr,
 741		addr + size - 1);
 742
 743	rc = offline_and_remove_memory(addr, size);
 744	if (!rc) {
 745		atomic64_sub(size, &vm->offline_size);
 746		/*
 747		 * We might have freed up memory we can now unplug, retry
 748		 * immediately instead of waiting.
 749		 */
 750		virtio_mem_retry(vm);
 751		return 0;
 
 
 752	}
 753	dev_dbg(&vm->vdev->dev, "offlining and removing memory failed: %d\n", rc);
 754	/*
 755	 * We don't really expect this to fail, because we fake-offlined all
 756	 * memory already. But it could fail in corner cases.
 757	 */
 758	WARN_ON_ONCE(rc != -ENOMEM && rc != -EBUSY);
 759	return rc == -ENOMEM ? -ENOMEM : -EBUSY;
 760}
 761
 762/*
 763 * See virtio_mem_offline_and_remove_memory(): Try offlining and removing
 764 * a single Linux memory block.
 765 */
 766static int virtio_mem_sbm_offline_and_remove_mb(struct virtio_mem *vm,
 767						unsigned long mb_id)
 768{
 769	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 770	const uint64_t size = memory_block_size_bytes();
 771
 772	return virtio_mem_offline_and_remove_memory(vm, addr, size);
 773}
 774
 775/*
 776 * Try (offlining and) removing memory from Linux in case all subblocks are
 777 * unplugged. Can be called on online and offline memory blocks.
 778 *
 779 * May modify the state of memory blocks in virtio-mem.
 780 */
 781static int virtio_mem_sbm_try_remove_unplugged_mb(struct virtio_mem *vm,
 782						  unsigned long mb_id)
 783{
 784	int rc;
 785
 786	/*
 787	 * Once all subblocks of a memory block were unplugged, offline and
 788	 * remove it.
 789	 */
 790	if (!virtio_mem_sbm_test_sb_unplugged(vm, mb_id, 0, vm->sbm.sbs_per_mb))
 791		return 0;
 792
 793	/* offline_and_remove_memory() works for online and offline memory. */
 794	mutex_unlock(&vm->hotplug_mutex);
 795	rc = virtio_mem_sbm_offline_and_remove_mb(vm, mb_id);
 796	mutex_lock(&vm->hotplug_mutex);
 797	if (!rc)
 798		virtio_mem_sbm_set_mb_state(vm, mb_id,
 799					    VIRTIO_MEM_SBM_MB_UNUSED);
 800	return rc;
 801}
 802
 803/*
 804 * See virtio_mem_offline_and_remove_memory(): Try to offline and remove a
 805 * all Linux memory blocks covered by the big block.
 806 */
 807static int virtio_mem_bbm_offline_and_remove_bb(struct virtio_mem *vm,
 808						unsigned long bb_id)
 809{
 810	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
 811	const uint64_t size = vm->bbm.bb_size;
 812
 813	return virtio_mem_offline_and_remove_memory(vm, addr, size);
 814}
 815
 816/*
 817 * Trigger the workqueue so the device can perform its magic.
 818 */
 819static void virtio_mem_retry(struct virtio_mem *vm)
 820{
 821	unsigned long flags;
 822
 823	spin_lock_irqsave(&vm->removal_lock, flags);
 824	if (!vm->removing)
 825		queue_work(system_freezable_wq, &vm->wq);
 826	spin_unlock_irqrestore(&vm->removal_lock, flags);
 827}
 828
 829static int virtio_mem_translate_node_id(struct virtio_mem *vm, uint16_t node_id)
 830{
 831	int node = NUMA_NO_NODE;
 832
 833#if defined(CONFIG_ACPI_NUMA)
 834	if (virtio_has_feature(vm->vdev, VIRTIO_MEM_F_ACPI_PXM))
 835		node = pxm_to_node(node_id);
 836#endif
 837	return node;
 838}
 839
 840/*
 841 * Test if a virtio-mem device overlaps with the given range. Can be called
 842 * from (notifier) callbacks lockless.
 843 */
 844static bool virtio_mem_overlaps_range(struct virtio_mem *vm, uint64_t start,
 845				      uint64_t size)
 846{
 847	return start < vm->addr + vm->region_size && vm->addr < start + size;
 848}
 849
 850/*
 851 * Test if a virtio-mem device contains a given range. Can be called from
 852 * (notifier) callbacks lockless.
 853 */
 854static bool virtio_mem_contains_range(struct virtio_mem *vm, uint64_t start,
 855				      uint64_t size)
 856{
 857	return start >= vm->addr && start + size <= vm->addr + vm->region_size;
 858}
 859
 860static int virtio_mem_sbm_notify_going_online(struct virtio_mem *vm,
 861					      unsigned long mb_id)
 862{
 863	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 864	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
 865	case VIRTIO_MEM_SBM_MB_OFFLINE:
 866		return NOTIFY_OK;
 867	default:
 868		break;
 869	}
 870	dev_warn_ratelimited(&vm->vdev->dev,
 871			     "memory block onlining denied\n");
 872	return NOTIFY_BAD;
 873}
 874
 875static void virtio_mem_sbm_notify_offline(struct virtio_mem *vm,
 876					  unsigned long mb_id)
 877{
 878	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 879	case VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL:
 880	case VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL:
 881		virtio_mem_sbm_set_mb_state(vm, mb_id,
 882					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
 883		break;
 884	case VIRTIO_MEM_SBM_MB_KERNEL:
 885	case VIRTIO_MEM_SBM_MB_MOVABLE:
 886		virtio_mem_sbm_set_mb_state(vm, mb_id,
 887					    VIRTIO_MEM_SBM_MB_OFFLINE);
 888		break;
 889	default:
 890		BUG();
 891		break;
 892	}
 893}
 894
 895static void virtio_mem_sbm_notify_online(struct virtio_mem *vm,
 896					 unsigned long mb_id,
 897					 unsigned long start_pfn)
 898{
 899	const bool is_movable = is_zone_movable_page(pfn_to_page(start_pfn));
 900	int new_state;
 901
 902	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 903	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
 904		new_state = VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL;
 905		if (is_movable)
 906			new_state = VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL;
 907		break;
 908	case VIRTIO_MEM_SBM_MB_OFFLINE:
 909		new_state = VIRTIO_MEM_SBM_MB_KERNEL;
 910		if (is_movable)
 911			new_state = VIRTIO_MEM_SBM_MB_MOVABLE;
 912		break;
 913	default:
 914		BUG();
 915		break;
 916	}
 917	virtio_mem_sbm_set_mb_state(vm, mb_id, new_state);
 918}
 919
 920static void virtio_mem_sbm_notify_going_offline(struct virtio_mem *vm,
 921						unsigned long mb_id)
 922{
 923	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size);
 924	unsigned long pfn;
 925	int sb_id;
 926
 927	for (sb_id = 0; sb_id < vm->sbm.sbs_per_mb; sb_id++) {
 928		if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
 929			continue;
 930		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
 931			       sb_id * vm->sbm.sb_size);
 932		virtio_mem_fake_offline_going_offline(pfn, nr_pages);
 933	}
 934}
 935
 936static void virtio_mem_sbm_notify_cancel_offline(struct virtio_mem *vm,
 937						 unsigned long mb_id)
 938{
 939	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size);
 940	unsigned long pfn;
 941	int sb_id;
 942
 943	for (sb_id = 0; sb_id < vm->sbm.sbs_per_mb; sb_id++) {
 944		if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
 945			continue;
 946		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
 947			       sb_id * vm->sbm.sb_size);
 948		virtio_mem_fake_offline_cancel_offline(pfn, nr_pages);
 949	}
 950}
 951
 952static void virtio_mem_bbm_notify_going_offline(struct virtio_mem *vm,
 953						unsigned long bb_id,
 954						unsigned long pfn,
 955						unsigned long nr_pages)
 956{
 957	/*
 958	 * When marked as "fake-offline", all online memory of this device block
 959	 * is allocated by us. Otherwise, we don't have any memory allocated.
 960	 */
 961	if (virtio_mem_bbm_get_bb_state(vm, bb_id) !=
 962	    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE)
 963		return;
 964	virtio_mem_fake_offline_going_offline(pfn, nr_pages);
 965}
 966
 967static void virtio_mem_bbm_notify_cancel_offline(struct virtio_mem *vm,
 968						 unsigned long bb_id,
 969						 unsigned long pfn,
 970						 unsigned long nr_pages)
 971{
 972	if (virtio_mem_bbm_get_bb_state(vm, bb_id) !=
 973	    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE)
 974		return;
 975	virtio_mem_fake_offline_cancel_offline(pfn, nr_pages);
 976}
 977
 978/*
 979 * This callback will either be called synchronously from add_memory() or
 980 * asynchronously (e.g., triggered via user space). We have to be careful
 981 * with locking when calling add_memory().
 982 */
 983static int virtio_mem_memory_notifier_cb(struct notifier_block *nb,
 984					 unsigned long action, void *arg)
 985{
 986	struct virtio_mem *vm = container_of(nb, struct virtio_mem,
 987					     memory_notifier);
 988	struct memory_notify *mhp = arg;
 989	const unsigned long start = PFN_PHYS(mhp->start_pfn);
 990	const unsigned long size = PFN_PHYS(mhp->nr_pages);
 991	int rc = NOTIFY_OK;
 992	unsigned long id;
 993
 994	if (!virtio_mem_overlaps_range(vm, start, size))
 995		return NOTIFY_DONE;
 996
 997	if (vm->in_sbm) {
 998		id = virtio_mem_phys_to_mb_id(start);
 999		/*
1000		 * In SBM, we add memory in separate memory blocks - we expect
1001		 * it to be onlined/offlined in the same granularity. Bail out
1002		 * if this ever changes.
1003		 */
1004		if (WARN_ON_ONCE(size != memory_block_size_bytes() ||
1005				 !IS_ALIGNED(start, memory_block_size_bytes())))
1006			return NOTIFY_BAD;
1007	} else {
1008		id = virtio_mem_phys_to_bb_id(vm, start);
1009		/*
1010		 * In BBM, we only care about onlining/offlining happening
1011		 * within a single big block, we don't care about the
1012		 * actual granularity as we don't track individual Linux
1013		 * memory blocks.
1014		 */
1015		if (WARN_ON_ONCE(id != virtio_mem_phys_to_bb_id(vm, start + size - 1)))
1016			return NOTIFY_BAD;
1017	}
1018
1019	/*
1020	 * Avoid circular locking lockdep warnings. We lock the mutex
1021	 * e.g., in MEM_GOING_ONLINE and unlock it in MEM_ONLINE. The
1022	 * blocking_notifier_call_chain() has it's own lock, which gets unlocked
1023	 * between both notifier calls and will bail out. False positive.
1024	 */
1025	lockdep_off();
1026
1027	switch (action) {
1028	case MEM_GOING_OFFLINE:
1029		mutex_lock(&vm->hotplug_mutex);
1030		if (vm->removing) {
1031			rc = notifier_from_errno(-EBUSY);
1032			mutex_unlock(&vm->hotplug_mutex);
1033			break;
1034		}
1035		vm->hotplug_active = true;
1036		if (vm->in_sbm)
1037			virtio_mem_sbm_notify_going_offline(vm, id);
1038		else
1039			virtio_mem_bbm_notify_going_offline(vm, id,
1040							    mhp->start_pfn,
1041							    mhp->nr_pages);
1042		break;
1043	case MEM_GOING_ONLINE:
1044		mutex_lock(&vm->hotplug_mutex);
1045		if (vm->removing) {
1046			rc = notifier_from_errno(-EBUSY);
1047			mutex_unlock(&vm->hotplug_mutex);
1048			break;
1049		}
1050		vm->hotplug_active = true;
1051		if (vm->in_sbm)
1052			rc = virtio_mem_sbm_notify_going_online(vm, id);
1053		break;
1054	case MEM_OFFLINE:
1055		if (vm->in_sbm)
1056			virtio_mem_sbm_notify_offline(vm, id);
1057
1058		atomic64_add(size, &vm->offline_size);
1059		/*
1060		 * Trigger the workqueue. Now that we have some offline memory,
1061		 * maybe we can handle pending unplug requests.
1062		 */
1063		if (!unplug_online)
1064			virtio_mem_retry(vm);
1065
1066		vm->hotplug_active = false;
1067		mutex_unlock(&vm->hotplug_mutex);
1068		break;
1069	case MEM_ONLINE:
1070		if (vm->in_sbm)
1071			virtio_mem_sbm_notify_online(vm, id, mhp->start_pfn);
1072
1073		atomic64_sub(size, &vm->offline_size);
1074		/*
1075		 * Start adding more memory once we onlined half of our
1076		 * threshold. Don't trigger if it's possibly due to our actipn
1077		 * (e.g., us adding memory which gets onlined immediately from
1078		 * the core).
1079		 */
1080		if (!atomic_read(&vm->wq_active) &&
1081		    virtio_mem_could_add_memory(vm, vm->offline_threshold / 2))
1082			virtio_mem_retry(vm);
1083
1084		vm->hotplug_active = false;
1085		mutex_unlock(&vm->hotplug_mutex);
1086		break;
1087	case MEM_CANCEL_OFFLINE:
1088		if (!vm->hotplug_active)
1089			break;
1090		if (vm->in_sbm)
1091			virtio_mem_sbm_notify_cancel_offline(vm, id);
1092		else
1093			virtio_mem_bbm_notify_cancel_offline(vm, id,
1094							     mhp->start_pfn,
1095							     mhp->nr_pages);
1096		vm->hotplug_active = false;
1097		mutex_unlock(&vm->hotplug_mutex);
1098		break;
1099	case MEM_CANCEL_ONLINE:
1100		if (!vm->hotplug_active)
1101			break;
1102		vm->hotplug_active = false;
1103		mutex_unlock(&vm->hotplug_mutex);
1104		break;
1105	default:
1106		break;
1107	}
1108
1109	lockdep_on();
1110
1111	return rc;
1112}
1113
1114/*
1115 * Set a range of pages PG_offline. Remember pages that were never onlined
1116 * (via generic_online_page()) using PageDirty().
1117 */
1118static void virtio_mem_set_fake_offline(unsigned long pfn,
1119					unsigned long nr_pages, bool onlined)
1120{
1121	page_offline_begin();
1122	for (; nr_pages--; pfn++) {
1123		struct page *page = pfn_to_page(pfn);
1124
1125		__SetPageOffline(page);
1126		if (!onlined) {
1127			SetPageDirty(page);
1128			/* FIXME: remove after cleanups */
1129			ClearPageReserved(page);
1130		}
1131	}
1132	page_offline_end();
1133}
1134
1135/*
1136 * Clear PG_offline from a range of pages. If the pages were never onlined,
1137 * (via generic_online_page()), clear PageDirty().
1138 */
1139static void virtio_mem_clear_fake_offline(unsigned long pfn,
1140					  unsigned long nr_pages, bool onlined)
1141{
1142	for (; nr_pages--; pfn++) {
1143		struct page *page = pfn_to_page(pfn);
1144
1145		__ClearPageOffline(page);
1146		if (!onlined)
1147			ClearPageDirty(page);
1148	}
1149}
1150
1151/*
1152 * Release a range of fake-offline pages to the buddy, effectively
1153 * fake-onlining them.
1154 */
1155static void virtio_mem_fake_online(unsigned long pfn, unsigned long nr_pages)
1156{
1157	unsigned long order = MAX_PAGE_ORDER;
1158	unsigned long i;
1159
1160	/*
1161	 * We might get called for ranges that don't cover properly aligned
1162	 * MAX_PAGE_ORDER pages; however, we can only online properly aligned
1163	 * pages with an order of MAX_PAGE_ORDER at maximum.
1164	 */
1165	while (!IS_ALIGNED(pfn | nr_pages, 1 << order))
1166		order--;
1167
1168	for (i = 0; i < nr_pages; i += 1 << order) {
1169		struct page *page = pfn_to_page(pfn + i);
1170
1171		/*
1172		 * If the page is PageDirty(), it was kept fake-offline when
1173		 * onlining the memory block. Otherwise, it was allocated
1174		 * using alloc_contig_range(). All pages in a subblock are
1175		 * alike.
1176		 */
1177		if (PageDirty(page)) {
1178			virtio_mem_clear_fake_offline(pfn + i, 1 << order, false);
1179			generic_online_page(page, order);
1180		} else {
1181			virtio_mem_clear_fake_offline(pfn + i, 1 << order, true);
1182			free_contig_range(pfn + i, 1 << order);
1183			adjust_managed_page_count(page, 1 << order);
1184		}
1185	}
1186}
1187
1188/*
1189 * Try to allocate a range, marking pages fake-offline, effectively
1190 * fake-offlining them.
1191 */
1192static int virtio_mem_fake_offline(struct virtio_mem *vm, unsigned long pfn,
1193				   unsigned long nr_pages)
1194{
1195	const bool is_movable = is_zone_movable_page(pfn_to_page(pfn));
1196	int rc, retry_count;
1197
1198	/*
1199	 * TODO: We want an alloc_contig_range() mode that tries to allocate
1200	 * harder (e.g., dealing with temporarily pinned pages, PCP), especially
1201	 * with ZONE_MOVABLE. So for now, retry a couple of times with
1202	 * ZONE_MOVABLE before giving up - because that zone is supposed to give
1203	 * some guarantees.
1204	 */
1205	for (retry_count = 0; retry_count < 5; retry_count++) {
1206		/*
1207		 * If the config changed, stop immediately and go back to the
1208		 * main loop: avoid trying to keep unplugging if the device
1209		 * might have decided to not remove any more memory.
1210		 */
1211		if (atomic_read(&vm->config_changed))
1212			return -EAGAIN;
1213
1214		rc = alloc_contig_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE,
1215					GFP_KERNEL);
1216		if (rc == -ENOMEM)
1217			/* whoops, out of memory */
1218			return rc;
1219		else if (rc && !is_movable)
1220			break;
1221		else if (rc)
1222			continue;
1223
1224		virtio_mem_set_fake_offline(pfn, nr_pages, true);
1225		adjust_managed_page_count(pfn_to_page(pfn), -nr_pages);
1226		return 0;
1227	}
1228
1229	return -EBUSY;
1230}
1231
1232/*
1233 * Handle fake-offline pages when memory is going offline - such that the
1234 * pages can be skipped by mm-core when offlining.
1235 */
1236static void virtio_mem_fake_offline_going_offline(unsigned long pfn,
1237						  unsigned long nr_pages)
1238{
1239	struct page *page;
1240	unsigned long i;
1241
1242	/*
1243	 * Drop our reference to the pages so the memory can get offlined
1244	 * and add the unplugged pages to the managed page counters (so
1245	 * offlining code can correctly subtract them again).
1246	 */
1247	adjust_managed_page_count(pfn_to_page(pfn), nr_pages);
1248	/* Drop our reference to the pages so the memory can get offlined. */
1249	for (i = 0; i < nr_pages; i++) {
1250		page = pfn_to_page(pfn + i);
1251		if (WARN_ON(!page_ref_dec_and_test(page)))
1252			dump_page(page, "fake-offline page referenced");
1253	}
1254}
1255
1256/*
1257 * Handle fake-offline pages when memory offlining is canceled - to undo
1258 * what we did in virtio_mem_fake_offline_going_offline().
1259 */
1260static void virtio_mem_fake_offline_cancel_offline(unsigned long pfn,
1261						   unsigned long nr_pages)
1262{
1263	unsigned long i;
1264
1265	/*
1266	 * Get the reference we dropped when going offline and subtract the
1267	 * unplugged pages from the managed page counters.
1268	 */
1269	adjust_managed_page_count(pfn_to_page(pfn), -nr_pages);
1270	for (i = 0; i < nr_pages; i++)
1271		page_ref_inc(pfn_to_page(pfn + i));
1272}
1273
1274static void virtio_mem_online_page(struct virtio_mem *vm,
1275				   struct page *page, unsigned int order)
1276{
1277	const unsigned long start = page_to_phys(page);
1278	const unsigned long end = start + PFN_PHYS(1 << order);
1279	unsigned long addr, next, id, sb_id, count;
1280	bool do_online;
1281
1282	/*
1283	 * We can get called with any order up to MAX_PAGE_ORDER. If our subblock
1284	 * size is smaller than that and we have a mixture of plugged and
1285	 * unplugged subblocks within such a page, we have to process in
1286	 * smaller granularity. In that case we'll adjust the order exactly once
1287	 * within the loop.
1288	 */
1289	for (addr = start; addr < end; ) {
1290		next = addr + PFN_PHYS(1 << order);
1291
1292		if (vm->in_sbm) {
1293			id = virtio_mem_phys_to_mb_id(addr);
1294			sb_id = virtio_mem_phys_to_sb_id(vm, addr);
1295			count = virtio_mem_phys_to_sb_id(vm, next - 1) - sb_id + 1;
1296
1297			if (virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, count)) {
1298				/* Fully plugged. */
1299				do_online = true;
1300			} else if (count == 1 ||
1301				   virtio_mem_sbm_test_sb_unplugged(vm, id, sb_id, count)) {
1302				/* Fully unplugged. */
1303				do_online = false;
1304			} else {
1305				/*
1306				 * Mixture, process sub-blocks instead. This
1307				 * will be at least the size of a pageblock.
1308				 * We'll run into this case exactly once.
1309				 */
1310				order = ilog2(vm->sbm.sb_size) - PAGE_SHIFT;
1311				do_online = virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, 1);
1312				continue;
1313			}
1314		} else {
1315			/*
1316			 * If the whole block is marked fake offline, keep
1317			 * everything that way.
1318			 */
1319			id = virtio_mem_phys_to_bb_id(vm, addr);
1320			do_online = virtio_mem_bbm_get_bb_state(vm, id) !=
1321				    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE;
1322		}
1323
1324		if (do_online)
1325			generic_online_page(pfn_to_page(PFN_DOWN(addr)), order);
1326		else
1327			virtio_mem_set_fake_offline(PFN_DOWN(addr), 1 << order,
1328						    false);
1329		addr = next;
1330	}
1331}
1332
1333static void virtio_mem_online_page_cb(struct page *page, unsigned int order)
1334{
1335	const unsigned long addr = page_to_phys(page);
1336	struct virtio_mem *vm;
1337
1338	rcu_read_lock();
1339	list_for_each_entry_rcu(vm, &virtio_mem_devices, next) {
1340		/*
1341		 * Pages we're onlining will never cross memory blocks and,
1342		 * therefore, not virtio-mem devices.
1343		 */
1344		if (!virtio_mem_contains_range(vm, addr, PFN_PHYS(1 << order)))
1345			continue;
1346
1347		/*
1348		 * virtio_mem_set_fake_offline() might sleep. We can safely
1349		 * drop the RCU lock at this point because the device
1350		 * cannot go away. See virtio_mem_remove() how races
1351		 * between memory onlining and device removal are handled.
1352		 */
1353		rcu_read_unlock();
1354
1355		virtio_mem_online_page(vm, page, order);
1356		return;
1357	}
1358	rcu_read_unlock();
1359
1360	/* not virtio-mem memory, but e.g., a DIMM. online it */
1361	generic_online_page(page, order);
1362}
1363
1364static uint64_t virtio_mem_send_request(struct virtio_mem *vm,
1365					const struct virtio_mem_req *req)
1366{
1367	struct scatterlist *sgs[2], sg_req, sg_resp;
1368	unsigned int len;
1369	int rc;
1370
1371	/* don't use the request residing on the stack (vaddr) */
1372	vm->req = *req;
1373
1374	/* out: buffer for request */
1375	sg_init_one(&sg_req, &vm->req, sizeof(vm->req));
1376	sgs[0] = &sg_req;
1377
1378	/* in: buffer for response */
1379	sg_init_one(&sg_resp, &vm->resp, sizeof(vm->resp));
1380	sgs[1] = &sg_resp;
1381
1382	rc = virtqueue_add_sgs(vm->vq, sgs, 1, 1, vm, GFP_KERNEL);
1383	if (rc < 0)
1384		return rc;
1385
1386	virtqueue_kick(vm->vq);
1387
1388	/* wait for a response */
1389	wait_event(vm->host_resp, virtqueue_get_buf(vm->vq, &len));
1390
1391	return virtio16_to_cpu(vm->vdev, vm->resp.type);
1392}
1393
1394static int virtio_mem_send_plug_request(struct virtio_mem *vm, uint64_t addr,
1395					uint64_t size)
1396{
1397	const uint64_t nb_vm_blocks = size / vm->device_block_size;
1398	const struct virtio_mem_req req = {
1399		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_PLUG),
1400		.u.plug.addr = cpu_to_virtio64(vm->vdev, addr),
1401		.u.plug.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
1402	};
1403	int rc = -ENOMEM;
1404
1405	if (atomic_read(&vm->config_changed))
1406		return -EAGAIN;
1407
1408	dev_dbg(&vm->vdev->dev, "plugging memory: 0x%llx - 0x%llx\n", addr,
1409		addr + size - 1);
1410
1411	switch (virtio_mem_send_request(vm, &req)) {
1412	case VIRTIO_MEM_RESP_ACK:
1413		vm->plugged_size += size;
1414		return 0;
1415	case VIRTIO_MEM_RESP_NACK:
1416		rc = -EAGAIN;
1417		break;
1418	case VIRTIO_MEM_RESP_BUSY:
1419		rc = -ETXTBSY;
1420		break;
1421	case VIRTIO_MEM_RESP_ERROR:
1422		rc = -EINVAL;
1423		break;
1424	default:
1425		break;
1426	}
1427
1428	dev_dbg(&vm->vdev->dev, "plugging memory failed: %d\n", rc);
1429	return rc;
1430}
1431
1432static int virtio_mem_send_unplug_request(struct virtio_mem *vm, uint64_t addr,
1433					  uint64_t size)
1434{
1435	const uint64_t nb_vm_blocks = size / vm->device_block_size;
1436	const struct virtio_mem_req req = {
1437		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_UNPLUG),
1438		.u.unplug.addr = cpu_to_virtio64(vm->vdev, addr),
1439		.u.unplug.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
1440	};
1441	int rc = -ENOMEM;
1442
1443	if (atomic_read(&vm->config_changed))
1444		return -EAGAIN;
1445
1446	dev_dbg(&vm->vdev->dev, "unplugging memory: 0x%llx - 0x%llx\n", addr,
1447		addr + size - 1);
1448
1449	switch (virtio_mem_send_request(vm, &req)) {
1450	case VIRTIO_MEM_RESP_ACK:
1451		vm->plugged_size -= size;
1452		return 0;
1453	case VIRTIO_MEM_RESP_BUSY:
1454		rc = -ETXTBSY;
1455		break;
1456	case VIRTIO_MEM_RESP_ERROR:
1457		rc = -EINVAL;
1458		break;
1459	default:
1460		break;
1461	}
1462
1463	dev_dbg(&vm->vdev->dev, "unplugging memory failed: %d\n", rc);
1464	return rc;
1465}
1466
1467static int virtio_mem_send_unplug_all_request(struct virtio_mem *vm)
1468{
1469	const struct virtio_mem_req req = {
1470		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_UNPLUG_ALL),
1471	};
1472	int rc = -ENOMEM;
1473
1474	dev_dbg(&vm->vdev->dev, "unplugging all memory");
1475
1476	switch (virtio_mem_send_request(vm, &req)) {
1477	case VIRTIO_MEM_RESP_ACK:
1478		vm->unplug_all_required = false;
1479		vm->plugged_size = 0;
1480		/* usable region might have shrunk */
1481		atomic_set(&vm->config_changed, 1);
1482		return 0;
1483	case VIRTIO_MEM_RESP_BUSY:
1484		rc = -ETXTBSY;
1485		break;
1486	default:
1487		break;
1488	}
1489
1490	dev_dbg(&vm->vdev->dev, "unplugging all memory failed: %d\n", rc);
1491	return rc;
1492}
1493
1494/*
1495 * Plug selected subblocks. Updates the plugged state, but not the state
1496 * of the memory block.
1497 */
1498static int virtio_mem_sbm_plug_sb(struct virtio_mem *vm, unsigned long mb_id,
1499				  int sb_id, int count)
1500{
1501	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id) +
1502			      sb_id * vm->sbm.sb_size;
1503	const uint64_t size = count * vm->sbm.sb_size;
1504	int rc;
1505
1506	rc = virtio_mem_send_plug_request(vm, addr, size);
1507	if (!rc)
1508		virtio_mem_sbm_set_sb_plugged(vm, mb_id, sb_id, count);
1509	return rc;
1510}
1511
1512/*
1513 * Unplug selected subblocks. Updates the plugged state, but not the state
1514 * of the memory block.
1515 */
1516static int virtio_mem_sbm_unplug_sb(struct virtio_mem *vm, unsigned long mb_id,
1517				    int sb_id, int count)
1518{
1519	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id) +
1520			      sb_id * vm->sbm.sb_size;
1521	const uint64_t size = count * vm->sbm.sb_size;
1522	int rc;
1523
1524	rc = virtio_mem_send_unplug_request(vm, addr, size);
1525	if (!rc)
1526		virtio_mem_sbm_set_sb_unplugged(vm, mb_id, sb_id, count);
1527	return rc;
1528}
1529
1530/*
1531 * Request to unplug a big block.
1532 *
1533 * Will not modify the state of the big block.
1534 */
1535static int virtio_mem_bbm_unplug_bb(struct virtio_mem *vm, unsigned long bb_id)
1536{
1537	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
1538	const uint64_t size = vm->bbm.bb_size;
1539
1540	return virtio_mem_send_unplug_request(vm, addr, size);
1541}
1542
1543/*
1544 * Request to plug a big block.
1545 *
1546 * Will not modify the state of the big block.
1547 */
1548static int virtio_mem_bbm_plug_bb(struct virtio_mem *vm, unsigned long bb_id)
1549{
1550	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
1551	const uint64_t size = vm->bbm.bb_size;
1552
1553	return virtio_mem_send_plug_request(vm, addr, size);
1554}
1555
1556/*
1557 * Unplug the desired number of plugged subblocks of a offline or not-added
1558 * memory block. Will fail if any subblock cannot get unplugged (instead of
1559 * skipping it).
1560 *
1561 * Will not modify the state of the memory block.
1562 *
1563 * Note: can fail after some subblocks were unplugged.
1564 */
1565static int virtio_mem_sbm_unplug_any_sb_raw(struct virtio_mem *vm,
1566					    unsigned long mb_id, uint64_t *nb_sb)
1567{
1568	int sb_id, count;
1569	int rc;
1570
1571	sb_id = vm->sbm.sbs_per_mb - 1;
1572	while (*nb_sb) {
1573		/* Find the next candidate subblock */
1574		while (sb_id >= 0 &&
1575		       virtio_mem_sbm_test_sb_unplugged(vm, mb_id, sb_id, 1))
1576			sb_id--;
1577		if (sb_id < 0)
1578			break;
1579		/* Try to unplug multiple subblocks at a time */
1580		count = 1;
1581		while (count < *nb_sb && sb_id > 0 &&
1582		       virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id - 1, 1)) {
1583			count++;
1584			sb_id--;
1585		}
1586
1587		rc = virtio_mem_sbm_unplug_sb(vm, mb_id, sb_id, count);
1588		if (rc)
1589			return rc;
1590		*nb_sb -= count;
1591		sb_id--;
1592	}
1593
1594	return 0;
1595}
1596
1597/*
1598 * Unplug all plugged subblocks of an offline or not-added memory block.
1599 *
1600 * Will not modify the state of the memory block.
1601 *
1602 * Note: can fail after some subblocks were unplugged.
1603 */
1604static int virtio_mem_sbm_unplug_mb(struct virtio_mem *vm, unsigned long mb_id)
1605{
1606	uint64_t nb_sb = vm->sbm.sbs_per_mb;
1607
1608	return virtio_mem_sbm_unplug_any_sb_raw(vm, mb_id, &nb_sb);
1609}
1610
1611/*
1612 * Prepare tracking data for the next memory block.
1613 */
1614static int virtio_mem_sbm_prepare_next_mb(struct virtio_mem *vm,
1615					  unsigned long *mb_id)
1616{
1617	int rc;
1618
1619	if (vm->sbm.next_mb_id > vm->sbm.last_usable_mb_id)
1620		return -ENOSPC;
1621
1622	/* Resize the state array if required. */
1623	rc = virtio_mem_sbm_mb_states_prepare_next_mb(vm);
1624	if (rc)
1625		return rc;
1626
1627	/* Resize the subblock bitmap if required. */
1628	rc = virtio_mem_sbm_sb_states_prepare_next_mb(vm);
1629	if (rc)
1630		return rc;
1631
1632	vm->sbm.mb_count[VIRTIO_MEM_SBM_MB_UNUSED]++;
1633	*mb_id = vm->sbm.next_mb_id++;
1634	return 0;
1635}
1636
1637/*
1638 * Try to plug the desired number of subblocks and add the memory block
1639 * to Linux.
1640 *
1641 * Will modify the state of the memory block.
1642 */
1643static int virtio_mem_sbm_plug_and_add_mb(struct virtio_mem *vm,
1644					  unsigned long mb_id, uint64_t *nb_sb)
1645{
1646	const int count = min_t(int, *nb_sb, vm->sbm.sbs_per_mb);
1647	int rc;
1648
1649	if (WARN_ON_ONCE(!count))
1650		return -EINVAL;
1651
1652	/*
1653	 * Plug the requested number of subblocks before adding it to linux,
1654	 * so that onlining will directly online all plugged subblocks.
1655	 */
1656	rc = virtio_mem_sbm_plug_sb(vm, mb_id, 0, count);
1657	if (rc)
1658		return rc;
1659
1660	/*
1661	 * Mark the block properly offline before adding it to Linux,
1662	 * so the memory notifiers will find the block in the right state.
1663	 */
1664	if (count == vm->sbm.sbs_per_mb)
1665		virtio_mem_sbm_set_mb_state(vm, mb_id,
1666					    VIRTIO_MEM_SBM_MB_OFFLINE);
1667	else
1668		virtio_mem_sbm_set_mb_state(vm, mb_id,
1669					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
1670
1671	/* Add the memory block to linux - if that fails, try to unplug. */
1672	rc = virtio_mem_sbm_add_mb(vm, mb_id);
1673	if (rc) {
1674		int new_state = VIRTIO_MEM_SBM_MB_UNUSED;
1675
1676		if (virtio_mem_sbm_unplug_sb(vm, mb_id, 0, count))
1677			new_state = VIRTIO_MEM_SBM_MB_PLUGGED;
1678		virtio_mem_sbm_set_mb_state(vm, mb_id, new_state);
1679		return rc;
1680	}
1681
1682	*nb_sb -= count;
1683	return 0;
1684}
1685
1686/*
1687 * Try to plug the desired number of subblocks of a memory block that
1688 * is already added to Linux.
1689 *
1690 * Will modify the state of the memory block.
1691 *
1692 * Note: Can fail after some subblocks were successfully plugged.
1693 */
1694static int virtio_mem_sbm_plug_any_sb(struct virtio_mem *vm,
1695				      unsigned long mb_id, uint64_t *nb_sb)
1696{
1697	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
1698	unsigned long pfn, nr_pages;
1699	int sb_id, count;
1700	int rc;
1701
1702	if (WARN_ON_ONCE(!*nb_sb))
1703		return -EINVAL;
1704
1705	while (*nb_sb) {
1706		sb_id = virtio_mem_sbm_first_unplugged_sb(vm, mb_id);
1707		if (sb_id >= vm->sbm.sbs_per_mb)
1708			break;
1709		count = 1;
1710		while (count < *nb_sb &&
1711		       sb_id + count < vm->sbm.sbs_per_mb &&
1712		       !virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id + count, 1))
1713			count++;
1714
1715		rc = virtio_mem_sbm_plug_sb(vm, mb_id, sb_id, count);
1716		if (rc)
1717			return rc;
1718		*nb_sb -= count;
1719		if (old_state == VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL)
1720			continue;
1721
1722		/* fake-online the pages if the memory block is online */
1723		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
1724			       sb_id * vm->sbm.sb_size);
1725		nr_pages = PFN_DOWN(count * vm->sbm.sb_size);
1726		virtio_mem_fake_online(pfn, nr_pages);
1727	}
1728
1729	if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb))
1730		virtio_mem_sbm_set_mb_state(vm, mb_id, old_state - 1);
1731
1732	return 0;
1733}
1734
1735static int virtio_mem_sbm_plug_request(struct virtio_mem *vm, uint64_t diff)
1736{
1737	const int mb_states[] = {
1738		VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
1739		VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
1740		VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
1741	};
1742	uint64_t nb_sb = diff / vm->sbm.sb_size;
1743	unsigned long mb_id;
1744	int rc, i;
1745
1746	if (!nb_sb)
1747		return 0;
1748
1749	/* Don't race with onlining/offlining */
1750	mutex_lock(&vm->hotplug_mutex);
1751
1752	for (i = 0; i < ARRAY_SIZE(mb_states); i++) {
1753		virtio_mem_sbm_for_each_mb(vm, mb_id, mb_states[i]) {
1754			rc = virtio_mem_sbm_plug_any_sb(vm, mb_id, &nb_sb);
1755			if (rc || !nb_sb)
1756				goto out_unlock;
1757			cond_resched();
1758		}
1759	}
1760
1761	/*
1762	 * We won't be working on online/offline memory blocks from this point,
1763	 * so we can't race with memory onlining/offlining. Drop the mutex.
1764	 */
1765	mutex_unlock(&vm->hotplug_mutex);
1766
1767	/* Try to plug and add unused blocks */
1768	virtio_mem_sbm_for_each_mb(vm, mb_id, VIRTIO_MEM_SBM_MB_UNUSED) {
1769		if (!virtio_mem_could_add_memory(vm, memory_block_size_bytes()))
1770			return -ENOSPC;
1771
1772		rc = virtio_mem_sbm_plug_and_add_mb(vm, mb_id, &nb_sb);
1773		if (rc || !nb_sb)
1774			return rc;
1775		cond_resched();
1776	}
1777
1778	/* Try to prepare, plug and add new blocks */
1779	while (nb_sb) {
1780		if (!virtio_mem_could_add_memory(vm, memory_block_size_bytes()))
1781			return -ENOSPC;
1782
1783		rc = virtio_mem_sbm_prepare_next_mb(vm, &mb_id);
1784		if (rc)
1785			return rc;
1786		rc = virtio_mem_sbm_plug_and_add_mb(vm, mb_id, &nb_sb);
1787		if (rc)
1788			return rc;
1789		cond_resched();
1790	}
1791
1792	return 0;
1793out_unlock:
1794	mutex_unlock(&vm->hotplug_mutex);
1795	return rc;
1796}
1797
1798/*
1799 * Plug a big block and add it to Linux.
1800 *
1801 * Will modify the state of the big block.
1802 */
1803static int virtio_mem_bbm_plug_and_add_bb(struct virtio_mem *vm,
1804					  unsigned long bb_id)
1805{
1806	int rc;
1807
1808	if (WARN_ON_ONCE(virtio_mem_bbm_get_bb_state(vm, bb_id) !=
1809			 VIRTIO_MEM_BBM_BB_UNUSED))
1810		return -EINVAL;
1811
1812	rc = virtio_mem_bbm_plug_bb(vm, bb_id);
1813	if (rc)
1814		return rc;
1815	virtio_mem_bbm_set_bb_state(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED);
1816
1817	rc = virtio_mem_bbm_add_bb(vm, bb_id);
1818	if (rc) {
1819		if (!virtio_mem_bbm_unplug_bb(vm, bb_id))
1820			virtio_mem_bbm_set_bb_state(vm, bb_id,
1821						    VIRTIO_MEM_BBM_BB_UNUSED);
1822		else
1823			/* Retry from the main loop. */
1824			virtio_mem_bbm_set_bb_state(vm, bb_id,
1825						    VIRTIO_MEM_BBM_BB_PLUGGED);
1826		return rc;
1827	}
1828	return 0;
1829}
1830
1831/*
1832 * Prepare tracking data for the next big block.
1833 */
1834static int virtio_mem_bbm_prepare_next_bb(struct virtio_mem *vm,
1835					  unsigned long *bb_id)
1836{
1837	int rc;
1838
1839	if (vm->bbm.next_bb_id > vm->bbm.last_usable_bb_id)
1840		return -ENOSPC;
1841
1842	/* Resize the big block state array if required. */
1843	rc = virtio_mem_bbm_bb_states_prepare_next_bb(vm);
1844	if (rc)
1845		return rc;
1846
1847	vm->bbm.bb_count[VIRTIO_MEM_BBM_BB_UNUSED]++;
1848	*bb_id = vm->bbm.next_bb_id;
1849	vm->bbm.next_bb_id++;
1850	return 0;
1851}
1852
1853static int virtio_mem_bbm_plug_request(struct virtio_mem *vm, uint64_t diff)
1854{
1855	uint64_t nb_bb = diff / vm->bbm.bb_size;
1856	unsigned long bb_id;
1857	int rc;
1858
1859	if (!nb_bb)
1860		return 0;
1861
1862	/* Try to plug and add unused big blocks */
1863	virtio_mem_bbm_for_each_bb(vm, bb_id, VIRTIO_MEM_BBM_BB_UNUSED) {
1864		if (!virtio_mem_could_add_memory(vm, vm->bbm.bb_size))
1865			return -ENOSPC;
1866
1867		rc = virtio_mem_bbm_plug_and_add_bb(vm, bb_id);
1868		if (!rc)
1869			nb_bb--;
1870		if (rc || !nb_bb)
1871			return rc;
1872		cond_resched();
1873	}
1874
1875	/* Try to prepare, plug and add new big blocks */
1876	while (nb_bb) {
1877		if (!virtio_mem_could_add_memory(vm, vm->bbm.bb_size))
1878			return -ENOSPC;
1879
1880		rc = virtio_mem_bbm_prepare_next_bb(vm, &bb_id);
1881		if (rc)
1882			return rc;
1883		rc = virtio_mem_bbm_plug_and_add_bb(vm, bb_id);
1884		if (!rc)
1885			nb_bb--;
1886		if (rc)
1887			return rc;
1888		cond_resched();
1889	}
1890
1891	return 0;
1892}
1893
1894/*
1895 * Try to plug the requested amount of memory.
1896 */
1897static int virtio_mem_plug_request(struct virtio_mem *vm, uint64_t diff)
1898{
1899	if (vm->in_sbm)
1900		return virtio_mem_sbm_plug_request(vm, diff);
1901	return virtio_mem_bbm_plug_request(vm, diff);
1902}
1903
1904/*
1905 * Unplug the desired number of plugged subblocks of an offline memory block.
1906 * Will fail if any subblock cannot get unplugged (instead of skipping it).
1907 *
1908 * Will modify the state of the memory block. Might temporarily drop the
1909 * hotplug_mutex.
1910 *
1911 * Note: Can fail after some subblocks were successfully unplugged.
1912 */
1913static int virtio_mem_sbm_unplug_any_sb_offline(struct virtio_mem *vm,
1914						unsigned long mb_id,
1915						uint64_t *nb_sb)
1916{
1917	int rc;
1918
1919	rc = virtio_mem_sbm_unplug_any_sb_raw(vm, mb_id, nb_sb);
1920
1921	/* some subblocks might have been unplugged even on failure */
1922	if (!virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb))
1923		virtio_mem_sbm_set_mb_state(vm, mb_id,
1924					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
1925	if (rc)
1926		return rc;
1927
1928	if (virtio_mem_sbm_test_sb_unplugged(vm, mb_id, 0, vm->sbm.sbs_per_mb)) {
1929		/*
1930		 * Remove the block from Linux - this should never fail.
1931		 * Hinder the block from getting onlined by marking it
1932		 * unplugged. Temporarily drop the mutex, so
1933		 * any pending GOING_ONLINE requests can be serviced/rejected.
1934		 */
1935		virtio_mem_sbm_set_mb_state(vm, mb_id,
1936					    VIRTIO_MEM_SBM_MB_UNUSED);
1937
1938		mutex_unlock(&vm->hotplug_mutex);
1939		rc = virtio_mem_sbm_remove_mb(vm, mb_id);
1940		BUG_ON(rc);
1941		mutex_lock(&vm->hotplug_mutex);
1942	}
1943	return 0;
1944}
1945
1946/*
1947 * Unplug the given plugged subblocks of an online memory block.
1948 *
1949 * Will modify the state of the memory block.
1950 */
1951static int virtio_mem_sbm_unplug_sb_online(struct virtio_mem *vm,
1952					   unsigned long mb_id, int sb_id,
1953					   int count)
1954{
1955	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size) * count;
1956	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
1957	unsigned long start_pfn;
1958	int rc;
1959
1960	start_pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
1961			     sb_id * vm->sbm.sb_size);
1962
1963	rc = virtio_mem_fake_offline(vm, start_pfn, nr_pages);
1964	if (rc)
1965		return rc;
1966
1967	/* Try to unplug the allocated memory */
1968	rc = virtio_mem_sbm_unplug_sb(vm, mb_id, sb_id, count);
1969	if (rc) {
1970		/* Return the memory to the buddy. */
1971		virtio_mem_fake_online(start_pfn, nr_pages);
1972		return rc;
1973	}
1974
1975	switch (old_state) {
1976	case VIRTIO_MEM_SBM_MB_KERNEL:
1977		virtio_mem_sbm_set_mb_state(vm, mb_id,
1978					    VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL);
1979		break;
1980	case VIRTIO_MEM_SBM_MB_MOVABLE:
1981		virtio_mem_sbm_set_mb_state(vm, mb_id,
1982					    VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL);
1983		break;
1984	}
1985
1986	return 0;
1987}
1988
1989/*
1990 * Unplug the desired number of plugged subblocks of an online memory block.
1991 * Will skip subblock that are busy.
1992 *
1993 * Will modify the state of the memory block. Might temporarily drop the
1994 * hotplug_mutex.
1995 *
1996 * Note: Can fail after some subblocks were successfully unplugged. Can
1997 *       return 0 even if subblocks were busy and could not get unplugged.
1998 */
1999static int virtio_mem_sbm_unplug_any_sb_online(struct virtio_mem *vm,
2000					       unsigned long mb_id,
2001					       uint64_t *nb_sb)
2002{
2003	int rc, sb_id;
2004
2005	/* If possible, try to unplug the complete block in one shot. */
2006	if (*nb_sb >= vm->sbm.sbs_per_mb &&
2007	    virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb)) {
2008		rc = virtio_mem_sbm_unplug_sb_online(vm, mb_id, 0,
2009						     vm->sbm.sbs_per_mb);
2010		if (!rc) {
2011			*nb_sb -= vm->sbm.sbs_per_mb;
2012			goto unplugged;
2013		} else if (rc != -EBUSY)
2014			return rc;
2015	}
2016
2017	/* Fallback to single subblocks. */
2018	for (sb_id = vm->sbm.sbs_per_mb - 1; sb_id >= 0 && *nb_sb; sb_id--) {
2019		/* Find the next candidate subblock */
2020		while (sb_id >= 0 &&
2021		       !virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
2022			sb_id--;
2023		if (sb_id < 0)
2024			break;
2025
2026		rc = virtio_mem_sbm_unplug_sb_online(vm, mb_id, sb_id, 1);
2027		if (rc == -EBUSY)
2028			continue;
2029		else if (rc)
2030			return rc;
2031		*nb_sb -= 1;
2032	}
2033
2034unplugged:
2035	rc = virtio_mem_sbm_try_remove_unplugged_mb(vm, mb_id);
2036	if (rc)
2037		vm->sbm.have_unplugged_mb = 1;
2038	/* Ignore errors, this is not critical. We'll retry later. */
 
 
 
 
 
 
 
 
 
 
2039	return 0;
2040}
2041
2042/*
2043 * Unplug the desired number of plugged subblocks of a memory block that is
2044 * already added to Linux. Will skip subblock of online memory blocks that are
2045 * busy (by the OS). Will fail if any subblock that's not busy cannot get
2046 * unplugged.
2047 *
2048 * Will modify the state of the memory block. Might temporarily drop the
2049 * hotplug_mutex.
2050 *
2051 * Note: Can fail after some subblocks were successfully unplugged. Can
2052 *       return 0 even if subblocks were busy and could not get unplugged.
2053 */
2054static int virtio_mem_sbm_unplug_any_sb(struct virtio_mem *vm,
2055					unsigned long mb_id,
2056					uint64_t *nb_sb)
2057{
2058	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
2059
2060	switch (old_state) {
2061	case VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL:
2062	case VIRTIO_MEM_SBM_MB_KERNEL:
2063	case VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL:
2064	case VIRTIO_MEM_SBM_MB_MOVABLE:
2065		return virtio_mem_sbm_unplug_any_sb_online(vm, mb_id, nb_sb);
2066	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
2067	case VIRTIO_MEM_SBM_MB_OFFLINE:
2068		return virtio_mem_sbm_unplug_any_sb_offline(vm, mb_id, nb_sb);
2069	}
2070	return -EINVAL;
2071}
2072
2073static int virtio_mem_sbm_unplug_request(struct virtio_mem *vm, uint64_t diff)
2074{
2075	const int mb_states[] = {
2076		VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
2077		VIRTIO_MEM_SBM_MB_OFFLINE,
2078		VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
2079		VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
2080		VIRTIO_MEM_SBM_MB_MOVABLE,
2081		VIRTIO_MEM_SBM_MB_KERNEL,
2082	};
2083	uint64_t nb_sb = diff / vm->sbm.sb_size;
2084	unsigned long mb_id;
2085	int rc, i;
2086
2087	if (!nb_sb)
2088		return 0;
2089
2090	/*
2091	 * We'll drop the mutex a couple of times when it is safe to do so.
2092	 * This might result in some blocks switching the state (online/offline)
2093	 * and we could miss them in this run - we will retry again later.
2094	 */
2095	mutex_lock(&vm->hotplug_mutex);
2096
2097	/*
2098	 * We try unplug from partially plugged blocks first, to try removing
2099	 * whole memory blocks along with metadata. We prioritize ZONE_MOVABLE
2100	 * as it's more reliable to unplug memory and remove whole memory
2101	 * blocks, and we don't want to trigger a zone imbalances by
2102	 * accidentially removing too much kernel memory.
2103	 */
2104	for (i = 0; i < ARRAY_SIZE(mb_states); i++) {
2105		virtio_mem_sbm_for_each_mb_rev(vm, mb_id, mb_states[i]) {
2106			rc = virtio_mem_sbm_unplug_any_sb(vm, mb_id, &nb_sb);
2107			if (rc || !nb_sb)
2108				goto out_unlock;
2109			mutex_unlock(&vm->hotplug_mutex);
2110			cond_resched();
2111			mutex_lock(&vm->hotplug_mutex);
2112		}
2113		if (!unplug_online && i == 1) {
2114			mutex_unlock(&vm->hotplug_mutex);
2115			return 0;
2116		}
2117	}
2118
2119	mutex_unlock(&vm->hotplug_mutex);
2120	return nb_sb ? -EBUSY : 0;
2121out_unlock:
2122	mutex_unlock(&vm->hotplug_mutex);
2123	return rc;
2124}
2125
2126/*
2127 * Try to offline and remove a big block from Linux and unplug it. Will fail
2128 * with -EBUSY if some memory is busy and cannot get unplugged.
2129 *
2130 * Will modify the state of the memory block. Might temporarily drop the
2131 * hotplug_mutex.
2132 */
2133static int virtio_mem_bbm_offline_remove_and_unplug_bb(struct virtio_mem *vm,
2134						       unsigned long bb_id)
2135{
2136	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2137	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2138	unsigned long end_pfn = start_pfn + nr_pages;
2139	unsigned long pfn;
2140	struct page *page;
2141	int rc;
2142
2143	if (WARN_ON_ONCE(virtio_mem_bbm_get_bb_state(vm, bb_id) !=
2144			 VIRTIO_MEM_BBM_BB_ADDED))
2145		return -EINVAL;
2146
2147	/*
2148	 * Start by fake-offlining all memory. Once we marked the device
2149	 * block as fake-offline, all newly onlined memory will
2150	 * automatically be kept fake-offline. Protect from concurrent
2151	 * onlining/offlining until we have a consistent state.
2152	 */
2153	mutex_lock(&vm->hotplug_mutex);
2154	virtio_mem_bbm_set_bb_state(vm, bb_id, VIRTIO_MEM_BBM_BB_FAKE_OFFLINE);
 
 
2155
2156	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
2157		page = pfn_to_online_page(pfn);
2158		if (!page)
2159			continue;
2160
2161		rc = virtio_mem_fake_offline(vm, pfn, PAGES_PER_SECTION);
2162		if (rc) {
2163			end_pfn = pfn;
2164			goto rollback;
 
2165		}
 
2166	}
2167	mutex_unlock(&vm->hotplug_mutex);
2168
2169	rc = virtio_mem_bbm_offline_and_remove_bb(vm, bb_id);
2170	if (rc) {
2171		mutex_lock(&vm->hotplug_mutex);
2172		goto rollback;
 
 
 
2173	}
2174
2175	rc = virtio_mem_bbm_unplug_bb(vm, bb_id);
2176	if (rc)
2177		virtio_mem_bbm_set_bb_state(vm, bb_id,
2178					    VIRTIO_MEM_BBM_BB_PLUGGED);
2179	else
2180		virtio_mem_bbm_set_bb_state(vm, bb_id,
2181					    VIRTIO_MEM_BBM_BB_UNUSED);
2182	return rc;
2183
2184rollback:
2185	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
2186		page = pfn_to_online_page(pfn);
2187		if (!page)
2188			continue;
2189		virtio_mem_fake_online(pfn, PAGES_PER_SECTION);
2190	}
2191	virtio_mem_bbm_set_bb_state(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED);
2192	mutex_unlock(&vm->hotplug_mutex);
2193	return rc;
2194}
2195
2196/*
2197 * Test if a big block is completely offline.
2198 */
2199static bool virtio_mem_bbm_bb_is_offline(struct virtio_mem *vm,
2200					 unsigned long bb_id)
2201{
2202	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2203	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2204	unsigned long pfn;
2205
2206	for (pfn = start_pfn; pfn < start_pfn + nr_pages;
2207	     pfn += PAGES_PER_SECTION) {
2208		if (pfn_to_online_page(pfn))
2209			return false;
2210	}
2211
2212	return true;
2213}
2214
2215/*
2216 * Test if a big block is completely onlined to ZONE_MOVABLE (or offline).
2217 */
2218static bool virtio_mem_bbm_bb_is_movable(struct virtio_mem *vm,
2219					 unsigned long bb_id)
2220{
2221	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2222	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2223	struct page *page;
2224	unsigned long pfn;
2225
2226	for (pfn = start_pfn; pfn < start_pfn + nr_pages;
2227	     pfn += PAGES_PER_SECTION) {
2228		page = pfn_to_online_page(pfn);
2229		if (!page)
2230			continue;
2231		if (page_zonenum(page) != ZONE_MOVABLE)
2232			return false;
2233	}
2234
2235	return true;
2236}
2237
2238static int virtio_mem_bbm_unplug_request(struct virtio_mem *vm, uint64_t diff)
2239{
2240	uint64_t nb_bb = diff / vm->bbm.bb_size;
2241	uint64_t bb_id;
2242	int rc, i;
2243
2244	if (!nb_bb)
2245		return 0;
2246
2247	/*
2248	 * Try to unplug big blocks. Similar to SBM, start with offline
2249	 * big blocks.
2250	 */
2251	for (i = 0; i < 3; i++) {
2252		virtio_mem_bbm_for_each_bb_rev(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED) {
2253			cond_resched();
2254
2255			/*
2256			 * As we're holding no locks, these checks are racy,
2257			 * but we don't care.
2258			 */
2259			if (i == 0 && !virtio_mem_bbm_bb_is_offline(vm, bb_id))
2260				continue;
2261			if (i == 1 && !virtio_mem_bbm_bb_is_movable(vm, bb_id))
2262				continue;
2263			rc = virtio_mem_bbm_offline_remove_and_unplug_bb(vm, bb_id);
2264			if (rc == -EBUSY)
2265				continue;
2266			if (!rc)
2267				nb_bb--;
2268			if (rc || !nb_bb)
2269				return rc;
2270		}
2271		if (i == 0 && !unplug_online)
2272			return 0;
2273	}
2274
2275	return nb_bb ? -EBUSY : 0;
2276}
2277
2278/*
2279 * Try to unplug the requested amount of memory.
2280 */
2281static int virtio_mem_unplug_request(struct virtio_mem *vm, uint64_t diff)
2282{
2283	if (vm->in_sbm)
2284		return virtio_mem_sbm_unplug_request(vm, diff);
2285	return virtio_mem_bbm_unplug_request(vm, diff);
2286}
2287
2288/*
2289 * Try to unplug all blocks that couldn't be unplugged before, for example,
2290 * because the hypervisor was busy. Further, offline and remove any memory
2291 * blocks where we previously failed.
2292 */
2293static int virtio_mem_cleanup_pending_mb(struct virtio_mem *vm)
2294{
2295	unsigned long id;
2296	int rc = 0;
2297
2298	if (!vm->in_sbm) {
2299		virtio_mem_bbm_for_each_bb(vm, id,
2300					   VIRTIO_MEM_BBM_BB_PLUGGED) {
2301			rc = virtio_mem_bbm_unplug_bb(vm, id);
2302			if (rc)
2303				return rc;
2304			virtio_mem_bbm_set_bb_state(vm, id,
2305						    VIRTIO_MEM_BBM_BB_UNUSED);
2306		}
2307		return 0;
2308	}
2309
2310	virtio_mem_sbm_for_each_mb(vm, id, VIRTIO_MEM_SBM_MB_PLUGGED) {
2311		rc = virtio_mem_sbm_unplug_mb(vm, id);
2312		if (rc)
2313			return rc;
2314		virtio_mem_sbm_set_mb_state(vm, id,
2315					    VIRTIO_MEM_SBM_MB_UNUSED);
2316	}
2317
2318	if (!vm->sbm.have_unplugged_mb)
2319		return 0;
2320
2321	/*
2322	 * Let's retry (offlining and) removing completely unplugged Linux
2323	 * memory blocks.
2324	 */
2325	vm->sbm.have_unplugged_mb = false;
2326
2327	mutex_lock(&vm->hotplug_mutex);
2328	virtio_mem_sbm_for_each_mb(vm, id, VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL)
2329		rc |= virtio_mem_sbm_try_remove_unplugged_mb(vm, id);
2330	virtio_mem_sbm_for_each_mb(vm, id, VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL)
2331		rc |= virtio_mem_sbm_try_remove_unplugged_mb(vm, id);
2332	virtio_mem_sbm_for_each_mb(vm, id, VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL)
2333		rc |= virtio_mem_sbm_try_remove_unplugged_mb(vm, id);
2334	mutex_unlock(&vm->hotplug_mutex);
2335
2336	if (rc)
2337		vm->sbm.have_unplugged_mb = true;
2338	/* Ignore errors, this is not critical. We'll retry later. */
2339	return 0;
2340}
2341
2342/*
2343 * Update all parts of the config that could have changed.
2344 */
2345static void virtio_mem_refresh_config(struct virtio_mem *vm)
2346{
2347	const struct range pluggable_range = mhp_get_pluggable_range(true);
2348	uint64_t new_plugged_size, usable_region_size, end_addr;
2349
2350	/* the plugged_size is just a reflection of what _we_ did previously */
2351	virtio_cread_le(vm->vdev, struct virtio_mem_config, plugged_size,
2352			&new_plugged_size);
2353	if (WARN_ON_ONCE(new_plugged_size != vm->plugged_size))
2354		vm->plugged_size = new_plugged_size;
2355
2356	/* calculate the last usable memory block id */
2357	virtio_cread_le(vm->vdev, struct virtio_mem_config,
2358			usable_region_size, &usable_region_size);
2359	end_addr = min(vm->addr + usable_region_size - 1,
2360		       pluggable_range.end);
2361
2362	if (vm->in_sbm) {
2363		vm->sbm.last_usable_mb_id = virtio_mem_phys_to_mb_id(end_addr);
2364		if (!IS_ALIGNED(end_addr + 1, memory_block_size_bytes()))
2365			vm->sbm.last_usable_mb_id--;
2366	} else {
2367		vm->bbm.last_usable_bb_id = virtio_mem_phys_to_bb_id(vm,
2368								     end_addr);
2369		if (!IS_ALIGNED(end_addr + 1, vm->bbm.bb_size))
2370			vm->bbm.last_usable_bb_id--;
2371	}
2372	/*
2373	 * If we cannot plug any of our device memory (e.g., nothing in the
2374	 * usable region is addressable), the last usable memory block id will
2375	 * be smaller than the first usable memory block id. We'll stop
2376	 * attempting to add memory with -ENOSPC from our main loop.
2377	 */
2378
2379	/* see if there is a request to change the size */
2380	virtio_cread_le(vm->vdev, struct virtio_mem_config, requested_size,
2381			&vm->requested_size);
2382
2383	dev_info(&vm->vdev->dev, "plugged size: 0x%llx", vm->plugged_size);
2384	dev_info(&vm->vdev->dev, "requested size: 0x%llx", vm->requested_size);
2385}
2386
2387/*
2388 * Workqueue function for handling plug/unplug requests and config updates.
2389 */
2390static void virtio_mem_run_wq(struct work_struct *work)
2391{
2392	struct virtio_mem *vm = container_of(work, struct virtio_mem, wq);
2393	uint64_t diff;
2394	int rc;
2395
2396	if (unlikely(vm->in_kdump)) {
2397		dev_warn_once(&vm->vdev->dev,
2398			     "unexpected workqueue run in kdump kernel\n");
2399		return;
2400	}
2401
2402	hrtimer_cancel(&vm->retry_timer);
2403
2404	if (vm->broken)
2405		return;
2406
2407	atomic_set(&vm->wq_active, 1);
2408retry:
2409	rc = 0;
2410
2411	/* Make sure we start with a clean state if there are leftovers. */
2412	if (unlikely(vm->unplug_all_required))
2413		rc = virtio_mem_send_unplug_all_request(vm);
2414
2415	if (atomic_read(&vm->config_changed)) {
2416		atomic_set(&vm->config_changed, 0);
2417		virtio_mem_refresh_config(vm);
2418	}
2419
2420	/* Cleanup any leftovers from previous runs */
2421	if (!rc)
2422		rc = virtio_mem_cleanup_pending_mb(vm);
2423
2424	if (!rc && vm->requested_size != vm->plugged_size) {
2425		if (vm->requested_size > vm->plugged_size) {
2426			diff = vm->requested_size - vm->plugged_size;
2427			rc = virtio_mem_plug_request(vm, diff);
2428		} else {
2429			diff = vm->plugged_size - vm->requested_size;
2430			rc = virtio_mem_unplug_request(vm, diff);
2431		}
2432	}
2433
2434	/*
2435	 * Keep retrying to offline and remove completely unplugged Linux
2436	 * memory blocks.
2437	 */
2438	if (!rc && vm->in_sbm && vm->sbm.have_unplugged_mb)
2439		rc = -EBUSY;
2440
2441	switch (rc) {
2442	case 0:
2443		vm->retry_timer_ms = VIRTIO_MEM_RETRY_TIMER_MIN_MS;
2444		break;
2445	case -ENOSPC:
2446		/*
2447		 * We cannot add any more memory (alignment, physical limit)
2448		 * or we have too many offline memory blocks.
2449		 */
2450		break;
2451	case -ETXTBSY:
2452		/*
2453		 * The hypervisor cannot process our request right now
2454		 * (e.g., out of memory, migrating);
2455		 */
2456	case -EBUSY:
2457		/*
2458		 * We cannot free up any memory to unplug it (all plugged memory
2459		 * is busy).
2460		 */
2461	case -ENOMEM:
2462		/* Out of memory, try again later. */
2463		hrtimer_start(&vm->retry_timer, ms_to_ktime(vm->retry_timer_ms),
2464			      HRTIMER_MODE_REL);
2465		break;
2466	case -EAGAIN:
2467		/* Retry immediately (e.g., the config changed). */
2468		goto retry;
2469	default:
2470		/* Unknown error, mark as broken */
2471		dev_err(&vm->vdev->dev,
2472			"unknown error, marking device broken: %d\n", rc);
2473		vm->broken = true;
2474	}
2475
2476	atomic_set(&vm->wq_active, 0);
2477}
2478
2479static enum hrtimer_restart virtio_mem_timer_expired(struct hrtimer *timer)
2480{
2481	struct virtio_mem *vm = container_of(timer, struct virtio_mem,
2482					     retry_timer);
2483
2484	virtio_mem_retry(vm);
2485	vm->retry_timer_ms = min_t(unsigned int, vm->retry_timer_ms * 2,
2486				   VIRTIO_MEM_RETRY_TIMER_MAX_MS);
2487	return HRTIMER_NORESTART;
2488}
2489
2490static void virtio_mem_handle_response(struct virtqueue *vq)
2491{
2492	struct virtio_mem *vm = vq->vdev->priv;
2493
2494	wake_up(&vm->host_resp);
2495}
2496
2497static int virtio_mem_init_vq(struct virtio_mem *vm)
2498{
2499	struct virtqueue *vq;
2500
2501	vq = virtio_find_single_vq(vm->vdev, virtio_mem_handle_response,
2502				   "guest-request");
2503	if (IS_ERR(vq))
2504		return PTR_ERR(vq);
2505	vm->vq = vq;
2506
2507	return 0;
2508}
2509
2510static int virtio_mem_init_hotplug(struct virtio_mem *vm)
2511{
2512	const struct range pluggable_range = mhp_get_pluggable_range(true);
2513	uint64_t unit_pages, sb_size, addr;
2514	int rc;
2515
2516	/* bad device setup - warn only */
2517	if (!IS_ALIGNED(vm->addr, memory_block_size_bytes()))
2518		dev_warn(&vm->vdev->dev,
2519			 "The alignment of the physical start address can make some memory unusable.\n");
2520	if (!IS_ALIGNED(vm->addr + vm->region_size, memory_block_size_bytes()))
2521		dev_warn(&vm->vdev->dev,
2522			 "The alignment of the physical end address can make some memory unusable.\n");
2523	if (vm->addr < pluggable_range.start ||
2524	    vm->addr + vm->region_size - 1 > pluggable_range.end)
2525		dev_warn(&vm->vdev->dev,
2526			 "Some device memory is not addressable/pluggable. This can make some memory unusable.\n");
2527
2528	/* Prepare the offline threshold - make sure we can add two blocks. */
2529	vm->offline_threshold = max_t(uint64_t, 2 * memory_block_size_bytes(),
2530				      VIRTIO_MEM_DEFAULT_OFFLINE_THRESHOLD);
2531
2532	/*
2533	 * alloc_contig_range() works reliably with pageblock
2534	 * granularity on ZONE_NORMAL, use pageblock_nr_pages.
2535	 */
2536	sb_size = PAGE_SIZE * pageblock_nr_pages;
2537	sb_size = max_t(uint64_t, vm->device_block_size, sb_size);
2538
2539	if (sb_size < memory_block_size_bytes() && !force_bbm) {
2540		/* SBM: At least two subblocks per Linux memory block. */
2541		vm->in_sbm = true;
2542		vm->sbm.sb_size = sb_size;
2543		vm->sbm.sbs_per_mb = memory_block_size_bytes() /
2544				     vm->sbm.sb_size;
2545
2546		/* Round up to the next full memory block */
2547		addr = max_t(uint64_t, vm->addr, pluggable_range.start) +
2548		       memory_block_size_bytes() - 1;
2549		vm->sbm.first_mb_id = virtio_mem_phys_to_mb_id(addr);
2550		vm->sbm.next_mb_id = vm->sbm.first_mb_id;
2551	} else {
2552		/* BBM: At least one Linux memory block. */
2553		vm->bbm.bb_size = max_t(uint64_t, vm->device_block_size,
2554					memory_block_size_bytes());
2555
2556		if (bbm_block_size) {
2557			if (!is_power_of_2(bbm_block_size)) {
2558				dev_warn(&vm->vdev->dev,
2559					 "bbm_block_size is not a power of 2");
2560			} else if (bbm_block_size < vm->bbm.bb_size) {
2561				dev_warn(&vm->vdev->dev,
2562					 "bbm_block_size is too small");
2563			} else {
2564				vm->bbm.bb_size = bbm_block_size;
2565			}
2566		}
2567
2568		/* Round up to the next aligned big block */
2569		addr = max_t(uint64_t, vm->addr, pluggable_range.start) +
2570		       vm->bbm.bb_size - 1;
2571		vm->bbm.first_bb_id = virtio_mem_phys_to_bb_id(vm, addr);
2572		vm->bbm.next_bb_id = vm->bbm.first_bb_id;
2573
2574		/* Make sure we can add two big blocks. */
2575		vm->offline_threshold = max_t(uint64_t, 2 * vm->bbm.bb_size,
2576					      vm->offline_threshold);
2577	}
2578
2579	dev_info(&vm->vdev->dev, "memory block size: 0x%lx",
2580		 memory_block_size_bytes());
2581	if (vm->in_sbm)
2582		dev_info(&vm->vdev->dev, "subblock size: 0x%llx",
2583			 (unsigned long long)vm->sbm.sb_size);
2584	else
2585		dev_info(&vm->vdev->dev, "big block size: 0x%llx",
2586			 (unsigned long long)vm->bbm.bb_size);
2587
2588	/* create the parent resource for all memory */
2589	rc = virtio_mem_create_resource(vm);
2590	if (rc)
2591		return rc;
2592
2593	/* use a single dynamic memory group to cover the whole memory device */
2594	if (vm->in_sbm)
2595		unit_pages = PHYS_PFN(memory_block_size_bytes());
2596	else
2597		unit_pages = PHYS_PFN(vm->bbm.bb_size);
2598	rc = memory_group_register_dynamic(vm->nid, unit_pages);
2599	if (rc < 0)
2600		goto out_del_resource;
2601	vm->mgid = rc;
2602
2603	/*
2604	 * If we still have memory plugged, we have to unplug all memory first.
2605	 * Registering our parent resource makes sure that this memory isn't
2606	 * actually in use (e.g., trying to reload the driver).
2607	 */
2608	if (vm->plugged_size) {
2609		vm->unplug_all_required = true;
2610		dev_info(&vm->vdev->dev, "unplugging all memory is required\n");
2611	}
2612
2613	/* register callbacks */
2614	vm->memory_notifier.notifier_call = virtio_mem_memory_notifier_cb;
2615	rc = register_memory_notifier(&vm->memory_notifier);
2616	if (rc)
2617		goto out_unreg_group;
2618	rc = register_virtio_mem_device(vm);
2619	if (rc)
2620		goto out_unreg_mem;
2621
2622	return 0;
2623out_unreg_mem:
2624	unregister_memory_notifier(&vm->memory_notifier);
2625out_unreg_group:
2626	memory_group_unregister(vm->mgid);
2627out_del_resource:
2628	virtio_mem_delete_resource(vm);
2629	return rc;
2630}
2631
2632#ifdef CONFIG_PROC_VMCORE
2633static int virtio_mem_send_state_request(struct virtio_mem *vm, uint64_t addr,
2634					 uint64_t size)
2635{
2636	const uint64_t nb_vm_blocks = size / vm->device_block_size;
2637	const struct virtio_mem_req req = {
2638		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_STATE),
2639		.u.state.addr = cpu_to_virtio64(vm->vdev, addr),
2640		.u.state.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
2641	};
2642	int rc = -ENOMEM;
2643
2644	dev_dbg(&vm->vdev->dev, "requesting state: 0x%llx - 0x%llx\n", addr,
2645		addr + size - 1);
2646
2647	switch (virtio_mem_send_request(vm, &req)) {
2648	case VIRTIO_MEM_RESP_ACK:
2649		return virtio16_to_cpu(vm->vdev, vm->resp.u.state.state);
2650	case VIRTIO_MEM_RESP_ERROR:
2651		rc = -EINVAL;
2652		break;
2653	default:
2654		break;
2655	}
2656
2657	dev_dbg(&vm->vdev->dev, "requesting state failed: %d\n", rc);
2658	return rc;
2659}
2660
2661static bool virtio_mem_vmcore_pfn_is_ram(struct vmcore_cb *cb,
2662					 unsigned long pfn)
2663{
2664	struct virtio_mem *vm = container_of(cb, struct virtio_mem,
2665					     vmcore_cb);
2666	uint64_t addr = PFN_PHYS(pfn);
2667	bool is_ram;
2668	int rc;
2669
2670	if (!virtio_mem_contains_range(vm, addr, PAGE_SIZE))
2671		return true;
2672	if (!vm->plugged_size)
2673		return false;
2674
2675	/*
2676	 * We have to serialize device requests and access to the information
2677	 * about the block queried last.
2678	 */
2679	mutex_lock(&vm->hotplug_mutex);
2680
2681	addr = ALIGN_DOWN(addr, vm->device_block_size);
2682	if (addr != vm->last_block_addr) {
2683		rc = virtio_mem_send_state_request(vm, addr,
2684						   vm->device_block_size);
2685		/* On any kind of error, we're going to signal !ram. */
2686		if (rc == VIRTIO_MEM_STATE_PLUGGED)
2687			vm->last_block_plugged = true;
2688		else
2689			vm->last_block_plugged = false;
2690		vm->last_block_addr = addr;
2691	}
2692
2693	is_ram = vm->last_block_plugged;
2694	mutex_unlock(&vm->hotplug_mutex);
2695	return is_ram;
2696}
2697#endif /* CONFIG_PROC_VMCORE */
2698
2699static int virtio_mem_init_kdump(struct virtio_mem *vm)
2700{
2701#ifdef CONFIG_PROC_VMCORE
2702	dev_info(&vm->vdev->dev, "memory hot(un)plug disabled in kdump kernel\n");
2703	vm->vmcore_cb.pfn_is_ram = virtio_mem_vmcore_pfn_is_ram;
2704	register_vmcore_cb(&vm->vmcore_cb);
2705	return 0;
2706#else /* CONFIG_PROC_VMCORE */
2707	dev_warn(&vm->vdev->dev, "disabled in kdump kernel without vmcore\n");
2708	return -EBUSY;
2709#endif /* CONFIG_PROC_VMCORE */
2710}
2711
2712static int virtio_mem_init(struct virtio_mem *vm)
2713{
2714	uint16_t node_id;
2715
2716	if (!vm->vdev->config->get) {
2717		dev_err(&vm->vdev->dev, "config access disabled\n");
2718		return -EINVAL;
2719	}
2720
2721	/* Fetch all properties that can't change. */
2722	virtio_cread_le(vm->vdev, struct virtio_mem_config, plugged_size,
2723			&vm->plugged_size);
2724	virtio_cread_le(vm->vdev, struct virtio_mem_config, block_size,
2725			&vm->device_block_size);
2726	virtio_cread_le(vm->vdev, struct virtio_mem_config, node_id,
2727			&node_id);
2728	vm->nid = virtio_mem_translate_node_id(vm, node_id);
2729	virtio_cread_le(vm->vdev, struct virtio_mem_config, addr, &vm->addr);
2730	virtio_cread_le(vm->vdev, struct virtio_mem_config, region_size,
2731			&vm->region_size);
2732
2733	/* Determine the nid for the device based on the lowest address. */
2734	if (vm->nid == NUMA_NO_NODE)
2735		vm->nid = memory_add_physaddr_to_nid(vm->addr);
2736
2737	dev_info(&vm->vdev->dev, "start address: 0x%llx", vm->addr);
2738	dev_info(&vm->vdev->dev, "region size: 0x%llx", vm->region_size);
2739	dev_info(&vm->vdev->dev, "device block size: 0x%llx",
2740		 (unsigned long long)vm->device_block_size);
2741	if (vm->nid != NUMA_NO_NODE && IS_ENABLED(CONFIG_NUMA))
2742		dev_info(&vm->vdev->dev, "nid: %d", vm->nid);
2743
2744	/*
2745	 * We don't want to (un)plug or reuse any memory when in kdump. The
2746	 * memory is still accessible (but not exposed to Linux).
2747	 */
2748	if (vm->in_kdump)
2749		return virtio_mem_init_kdump(vm);
2750	return virtio_mem_init_hotplug(vm);
2751}
2752
2753static int virtio_mem_create_resource(struct virtio_mem *vm)
2754{
2755	/*
2756	 * When force-unloading the driver and removing the device, we
2757	 * could have a garbage pointer. Duplicate the string.
2758	 */
2759	const char *name = kstrdup(dev_name(&vm->vdev->dev), GFP_KERNEL);
2760
2761	if (!name)
2762		return -ENOMEM;
2763
2764	/* Disallow mapping device memory via /dev/mem completely. */
2765	vm->parent_resource = __request_mem_region(vm->addr, vm->region_size,
2766						   name, IORESOURCE_SYSTEM_RAM |
2767						   IORESOURCE_EXCLUSIVE);
2768	if (!vm->parent_resource) {
2769		kfree(name);
2770		dev_warn(&vm->vdev->dev, "could not reserve device region\n");
2771		dev_info(&vm->vdev->dev,
2772			 "reloading the driver is not supported\n");
2773		return -EBUSY;
2774	}
2775
2776	/* The memory is not actually busy - make add_memory() work. */
2777	vm->parent_resource->flags &= ~IORESOURCE_BUSY;
2778	return 0;
2779}
2780
2781static void virtio_mem_delete_resource(struct virtio_mem *vm)
2782{
2783	const char *name;
2784
2785	if (!vm->parent_resource)
2786		return;
2787
2788	name = vm->parent_resource->name;
2789	release_resource(vm->parent_resource);
2790	kfree(vm->parent_resource);
2791	kfree(name);
2792	vm->parent_resource = NULL;
2793}
2794
2795static int virtio_mem_range_has_system_ram(struct resource *res, void *arg)
2796{
2797	return 1;
2798}
2799
2800static bool virtio_mem_has_memory_added(struct virtio_mem *vm)
2801{
2802	const unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
2803
2804	return walk_iomem_res_desc(IORES_DESC_NONE, flags, vm->addr,
2805				   vm->addr + vm->region_size, NULL,
2806				   virtio_mem_range_has_system_ram) == 1;
2807}
2808
2809static int virtio_mem_probe(struct virtio_device *vdev)
2810{
2811	struct virtio_mem *vm;
2812	int rc;
2813
2814	BUILD_BUG_ON(sizeof(struct virtio_mem_req) != 24);
2815	BUILD_BUG_ON(sizeof(struct virtio_mem_resp) != 10);
2816
2817	vdev->priv = vm = kzalloc(sizeof(*vm), GFP_KERNEL);
2818	if (!vm)
2819		return -ENOMEM;
2820
2821	init_waitqueue_head(&vm->host_resp);
2822	vm->vdev = vdev;
2823	INIT_WORK(&vm->wq, virtio_mem_run_wq);
2824	mutex_init(&vm->hotplug_mutex);
2825	INIT_LIST_HEAD(&vm->next);
2826	spin_lock_init(&vm->removal_lock);
2827	hrtimer_init(&vm->retry_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2828	vm->retry_timer.function = virtio_mem_timer_expired;
2829	vm->retry_timer_ms = VIRTIO_MEM_RETRY_TIMER_MIN_MS;
2830	vm->in_kdump = is_kdump_kernel();
2831
2832	/* register the virtqueue */
2833	rc = virtio_mem_init_vq(vm);
2834	if (rc)
2835		goto out_free_vm;
2836
2837	/* initialize the device by querying the config */
2838	rc = virtio_mem_init(vm);
2839	if (rc)
2840		goto out_del_vq;
2841
2842	virtio_device_ready(vdev);
2843
2844	/* trigger a config update to start processing the requested_size */
2845	if (!vm->in_kdump) {
2846		atomic_set(&vm->config_changed, 1);
2847		queue_work(system_freezable_wq, &vm->wq);
2848	}
2849
2850	return 0;
2851out_del_vq:
2852	vdev->config->del_vqs(vdev);
2853out_free_vm:
2854	kfree(vm);
2855	vdev->priv = NULL;
2856
2857	return rc;
2858}
2859
2860static void virtio_mem_deinit_hotplug(struct virtio_mem *vm)
2861{
2862	unsigned long mb_id;
2863	int rc;
2864
2865	/*
2866	 * Make sure the workqueue won't be triggered anymore and no memory
2867	 * blocks can be onlined/offlined until we're finished here.
2868	 */
2869	mutex_lock(&vm->hotplug_mutex);
2870	spin_lock_irq(&vm->removal_lock);
2871	vm->removing = true;
2872	spin_unlock_irq(&vm->removal_lock);
2873	mutex_unlock(&vm->hotplug_mutex);
2874
2875	/* wait until the workqueue stopped */
2876	cancel_work_sync(&vm->wq);
2877	hrtimer_cancel(&vm->retry_timer);
2878
2879	if (vm->in_sbm) {
2880		/*
2881		 * After we unregistered our callbacks, user space can online
2882		 * partially plugged offline blocks. Make sure to remove them.
2883		 */
2884		virtio_mem_sbm_for_each_mb(vm, mb_id,
2885					   VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL) {
2886			rc = virtio_mem_sbm_remove_mb(vm, mb_id);
2887			BUG_ON(rc);
2888			virtio_mem_sbm_set_mb_state(vm, mb_id,
2889						    VIRTIO_MEM_SBM_MB_UNUSED);
2890		}
2891		/*
2892		 * After we unregistered our callbacks, user space can no longer
2893		 * offline partially plugged online memory blocks. No need to
2894		 * worry about them.
2895		 */
2896	}
2897
2898	/* unregister callbacks */
2899	unregister_virtio_mem_device(vm);
2900	unregister_memory_notifier(&vm->memory_notifier);
2901
2902	/*
2903	 * There is no way we could reliably remove all memory we have added to
2904	 * the system. And there is no way to stop the driver/device from going
2905	 * away. Warn at least.
2906	 */
2907	if (virtio_mem_has_memory_added(vm)) {
2908		dev_warn(&vm->vdev->dev,
2909			 "device still has system memory added\n");
2910	} else {
2911		virtio_mem_delete_resource(vm);
2912		kfree_const(vm->resource_name);
2913		memory_group_unregister(vm->mgid);
2914	}
2915
2916	/* remove all tracking data - no locking needed */
2917	if (vm->in_sbm) {
2918		vfree(vm->sbm.mb_states);
2919		vfree(vm->sbm.sb_states);
2920	} else {
2921		vfree(vm->bbm.bb_states);
2922	}
2923}
2924
2925static void virtio_mem_deinit_kdump(struct virtio_mem *vm)
2926{
2927#ifdef CONFIG_PROC_VMCORE
2928	unregister_vmcore_cb(&vm->vmcore_cb);
2929#endif /* CONFIG_PROC_VMCORE */
2930}
2931
2932static void virtio_mem_remove(struct virtio_device *vdev)
2933{
2934	struct virtio_mem *vm = vdev->priv;
2935
2936	if (vm->in_kdump)
2937		virtio_mem_deinit_kdump(vm);
2938	else
2939		virtio_mem_deinit_hotplug(vm);
2940
2941	/* reset the device and cleanup the queues */
2942	virtio_reset_device(vdev);
2943	vdev->config->del_vqs(vdev);
2944
2945	kfree(vm);
2946	vdev->priv = NULL;
2947}
2948
2949static void virtio_mem_config_changed(struct virtio_device *vdev)
2950{
2951	struct virtio_mem *vm = vdev->priv;
2952
2953	if (unlikely(vm->in_kdump))
2954		return;
2955
2956	atomic_set(&vm->config_changed, 1);
2957	virtio_mem_retry(vm);
2958}
2959
2960#ifdef CONFIG_PM_SLEEP
2961static int virtio_mem_freeze(struct virtio_device *vdev)
2962{
2963	/*
2964	 * When restarting the VM, all memory is usually unplugged. Don't
2965	 * allow to suspend/hibernate.
2966	 */
2967	dev_err(&vdev->dev, "save/restore not supported.\n");
2968	return -EPERM;
2969}
2970
2971static int virtio_mem_restore(struct virtio_device *vdev)
2972{
2973	return -EPERM;
2974}
2975#endif
2976
2977static unsigned int virtio_mem_features[] = {
2978#if defined(CONFIG_NUMA) && defined(CONFIG_ACPI_NUMA)
2979	VIRTIO_MEM_F_ACPI_PXM,
2980#endif
2981	VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE,
2982};
2983
2984static const struct virtio_device_id virtio_mem_id_table[] = {
2985	{ VIRTIO_ID_MEM, VIRTIO_DEV_ANY_ID },
2986	{ 0 },
2987};
2988
2989static struct virtio_driver virtio_mem_driver = {
2990	.feature_table = virtio_mem_features,
2991	.feature_table_size = ARRAY_SIZE(virtio_mem_features),
2992	.driver.name = KBUILD_MODNAME,
2993	.driver.owner = THIS_MODULE,
2994	.id_table = virtio_mem_id_table,
2995	.probe = virtio_mem_probe,
2996	.remove = virtio_mem_remove,
2997	.config_changed = virtio_mem_config_changed,
2998#ifdef CONFIG_PM_SLEEP
2999	.freeze	=	virtio_mem_freeze,
3000	.restore =	virtio_mem_restore,
3001#endif
3002};
3003
3004module_virtio_driver(virtio_mem_driver);
3005MODULE_DEVICE_TABLE(virtio, virtio_mem_id_table);
3006MODULE_AUTHOR("David Hildenbrand <david@redhat.com>");
3007MODULE_DESCRIPTION("Virtio-mem driver");
3008MODULE_LICENSE("GPL");
   1// SPDX-License-Identifier: GPL-2.0-or-later
   2/*
   3 * Virtio-mem device driver.
   4 *
   5 * Copyright Red Hat, Inc. 2020
   6 *
   7 * Author(s): David Hildenbrand <david@redhat.com>
   8 */
   9
  10#include <linux/virtio.h>
  11#include <linux/virtio_mem.h>
  12#include <linux/workqueue.h>
  13#include <linux/slab.h>
  14#include <linux/module.h>
  15#include <linux/mm.h>
  16#include <linux/memory_hotplug.h>
  17#include <linux/memory.h>
  18#include <linux/hrtimer.h>
  19#include <linux/crash_dump.h>
  20#include <linux/mutex.h>
  21#include <linux/bitmap.h>
  22#include <linux/lockdep.h>
  23#include <linux/log2.h>
  24
  25#include <acpi/acpi_numa.h>
  26
  27static bool unplug_online = true;
  28module_param(unplug_online, bool, 0644);
  29MODULE_PARM_DESC(unplug_online, "Try to unplug online memory");
  30
  31static bool force_bbm;
  32module_param(force_bbm, bool, 0444);
  33MODULE_PARM_DESC(force_bbm,
  34		"Force Big Block Mode. Default is 0 (auto-selection)");
  35
  36static unsigned long bbm_block_size;
  37module_param(bbm_block_size, ulong, 0444);
  38MODULE_PARM_DESC(bbm_block_size,
  39		 "Big Block size in bytes. Default is 0 (auto-detection).");
  40
  41static bool bbm_safe_unplug = true;
  42module_param(bbm_safe_unplug, bool, 0444);
  43MODULE_PARM_DESC(bbm_safe_unplug,
  44	     "Use a safe unplug mechanism in BBM, avoiding long/endless loops");
  45
  46/*
  47 * virtio-mem currently supports the following modes of operation:
  48 *
  49 * * Sub Block Mode (SBM): A Linux memory block spans 2..X subblocks (SB). The
  50 *   size of a Sub Block (SB) is determined based on the device block size, the
  51 *   pageblock size, and the maximum allocation granularity of the buddy.
  52 *   Subblocks within a Linux memory block might either be plugged or unplugged.
  53 *   Memory is added/removed to Linux MM in Linux memory block granularity.
  54 *
  55 * * Big Block Mode (BBM): A Big Block (BB) spans 1..X Linux memory blocks.
  56 *   Memory is added/removed to Linux MM in Big Block granularity.
  57 *
  58 * The mode is determined automatically based on the Linux memory block size
  59 * and the device block size.
  60 *
  61 * User space / core MM (auto onlining) is responsible for onlining added
  62 * Linux memory blocks - and for selecting a zone. Linux Memory Blocks are
  63 * always onlined separately, and all memory within a Linux memory block is
  64 * onlined to the same zone - virtio-mem relies on this behavior.
  65 */
  66
  67/*
  68 * State of a Linux memory block in SBM.
  69 */
  70enum virtio_mem_sbm_mb_state {
  71	/* Unplugged, not added to Linux. Can be reused later. */
  72	VIRTIO_MEM_SBM_MB_UNUSED = 0,
  73	/* (Partially) plugged, not added to Linux. Error on add_memory(). */
  74	VIRTIO_MEM_SBM_MB_PLUGGED,
  75	/* Fully plugged, fully added to Linux, offline. */
  76	VIRTIO_MEM_SBM_MB_OFFLINE,
  77	/* Partially plugged, fully added to Linux, offline. */
  78	VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
  79	/* Fully plugged, fully added to Linux, onlined to a kernel zone. */
  80	VIRTIO_MEM_SBM_MB_KERNEL,
  81	/* Partially plugged, fully added to Linux, online to a kernel zone */
  82	VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
  83	/* Fully plugged, fully added to Linux, onlined to ZONE_MOVABLE. */
  84	VIRTIO_MEM_SBM_MB_MOVABLE,
  85	/* Partially plugged, fully added to Linux, onlined to ZONE_MOVABLE. */
  86	VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
  87	VIRTIO_MEM_SBM_MB_COUNT
  88};
  89
  90/*
  91 * State of a Big Block (BB) in BBM, covering 1..X Linux memory blocks.
  92 */
  93enum virtio_mem_bbm_bb_state {
  94	/* Unplugged, not added to Linux. Can be reused later. */
  95	VIRTIO_MEM_BBM_BB_UNUSED = 0,
  96	/* Plugged, not added to Linux. Error on add_memory(). */
  97	VIRTIO_MEM_BBM_BB_PLUGGED,
  98	/* Plugged and added to Linux. */
  99	VIRTIO_MEM_BBM_BB_ADDED,
 100	/* All online parts are fake-offline, ready to remove. */
 101	VIRTIO_MEM_BBM_BB_FAKE_OFFLINE,
 102	VIRTIO_MEM_BBM_BB_COUNT
 103};
 104
 105struct virtio_mem {
 106	struct virtio_device *vdev;
 107
 108	/* We might first have to unplug all memory when starting up. */
 109	bool unplug_all_required;
 110
 111	/* Workqueue that processes the plug/unplug requests. */
 112	struct work_struct wq;
 113	atomic_t wq_active;
 114	atomic_t config_changed;
 115
 116	/* Virtqueue for guest->host requests. */
 117	struct virtqueue *vq;
 118
 119	/* Wait for a host response to a guest request. */
 120	wait_queue_head_t host_resp;
 121
 122	/* Space for one guest request and the host response. */
 123	struct virtio_mem_req req;
 124	struct virtio_mem_resp resp;
 125
 126	/* The current size of the device. */
 127	uint64_t plugged_size;
 128	/* The requested size of the device. */
 129	uint64_t requested_size;
 130
 131	/* The device block size (for communicating with the device). */
 132	uint64_t device_block_size;
 133	/* The determined node id for all memory of the device. */
 134	int nid;
 135	/* Physical start address of the memory region. */
 136	uint64_t addr;
 137	/* Maximum region size in bytes. */
 138	uint64_t region_size;
 139
 140	/* The parent resource for all memory added via this device. */
 141	struct resource *parent_resource;
 142	/*
 143	 * Copy of "System RAM (virtio_mem)" to be used for
 144	 * add_memory_driver_managed().
 145	 */
 146	const char *resource_name;
 147	/* Memory group identification. */
 148	int mgid;
 149
 150	/*
 151	 * We don't want to add too much memory if it's not getting onlined,
 152	 * to avoid running OOM. Besides this threshold, we allow to have at
 153	 * least two offline blocks at a time (whatever is bigger).
 154	 */
 155#define VIRTIO_MEM_DEFAULT_OFFLINE_THRESHOLD		(1024 * 1024 * 1024)
 156	atomic64_t offline_size;
 157	uint64_t offline_threshold;
 158
 159	/* If set, the driver is in SBM, otherwise in BBM. */
 160	bool in_sbm;
 161
 162	union {
 163		struct {
 164			/* Id of the first memory block of this device. */
 165			unsigned long first_mb_id;
 166			/* Id of the last usable memory block of this device. */
 167			unsigned long last_usable_mb_id;
 168			/* Id of the next memory bock to prepare when needed. */
 169			unsigned long next_mb_id;
 170
 171			/* The subblock size. */
 172			uint64_t sb_size;
 173			/* The number of subblocks per Linux memory block. */
 174			uint32_t sbs_per_mb;
 175
 
 
 
 
 
 
 
 176			/* Summary of all memory block states. */
 177			unsigned long mb_count[VIRTIO_MEM_SBM_MB_COUNT];
 178
 179			/*
 180			 * One byte state per memory block. Allocated via
 181			 * vmalloc(). Resized (alloc+copy+free) on demand.
 182			 *
 183			 * With 128 MiB memory blocks, we have states for 512
 184			 * GiB of memory in one 4 KiB page.
 185			 */
 186			uint8_t *mb_states;
 187
 188			/*
 189			 * Bitmap: one bit per subblock. Allocated similar to
 190			 * sbm.mb_states.
 191			 *
 192			 * A set bit means the corresponding subblock is
 193			 * plugged, otherwise it's unblocked.
 194			 *
 195			 * With 4 MiB subblocks, we manage 128 GiB of memory
 196			 * in one 4 KiB page.
 197			 */
 198			unsigned long *sb_states;
 199		} sbm;
 200
 201		struct {
 202			/* Id of the first big block of this device. */
 203			unsigned long first_bb_id;
 204			/* Id of the last usable big block of this device. */
 205			unsigned long last_usable_bb_id;
 206			/* Id of the next device bock to prepare when needed. */
 207			unsigned long next_bb_id;
 208
 209			/* Summary of all big block states. */
 210			unsigned long bb_count[VIRTIO_MEM_BBM_BB_COUNT];
 211
 212			/* One byte state per big block. See sbm.mb_states. */
 213			uint8_t *bb_states;
 214
 215			/* The block size used for plugging/adding/removing. */
 216			uint64_t bb_size;
 217		} bbm;
 218	};
 219
 220	/*
 221	 * Mutex that protects the sbm.mb_count, sbm.mb_states,
 222	 * sbm.sb_states, bbm.bb_count, and bbm.bb_states
 223	 *
 224	 * When this lock is held the pointers can't change, ONLINE and
 225	 * OFFLINE blocks can't change the state and no subblocks will get
 226	 * plugged/unplugged.
 227	 *
 228	 * In kdump mode, used to serialize requests, last_block_addr and
 229	 * last_block_plugged.
 230	 */
 231	struct mutex hotplug_mutex;
 232	bool hotplug_active;
 233
 234	/* An error occurred we cannot handle - stop processing requests. */
 235	bool broken;
 236
 237	/* Cached valued of is_kdump_kernel() when the device was probed. */
 238	bool in_kdump;
 239
 240	/* The driver is being removed. */
 241	spinlock_t removal_lock;
 242	bool removing;
 243
 244	/* Timer for retrying to plug/unplug memory. */
 245	struct hrtimer retry_timer;
 246	unsigned int retry_timer_ms;
 247#define VIRTIO_MEM_RETRY_TIMER_MIN_MS		50000
 248#define VIRTIO_MEM_RETRY_TIMER_MAX_MS		300000
 249
 250	/* Memory notifier (online/offline events). */
 251	struct notifier_block memory_notifier;
 252
 253#ifdef CONFIG_PROC_VMCORE
 254	/* vmcore callback for /proc/vmcore handling in kdump mode */
 255	struct vmcore_cb vmcore_cb;
 256	uint64_t last_block_addr;
 257	bool last_block_plugged;
 258#endif /* CONFIG_PROC_VMCORE */
 259
 260	/* Next device in the list of virtio-mem devices. */
 261	struct list_head next;
 262};
 263
 264/*
 265 * We have to share a single online_page callback among all virtio-mem
 266 * devices. We use RCU to iterate the list in the callback.
 267 */
 268static DEFINE_MUTEX(virtio_mem_mutex);
 269static LIST_HEAD(virtio_mem_devices);
 270
 271static void virtio_mem_online_page_cb(struct page *page, unsigned int order);
 272static void virtio_mem_fake_offline_going_offline(unsigned long pfn,
 273						  unsigned long nr_pages);
 274static void virtio_mem_fake_offline_cancel_offline(unsigned long pfn,
 275						   unsigned long nr_pages);
 276static void virtio_mem_retry(struct virtio_mem *vm);
 277static int virtio_mem_create_resource(struct virtio_mem *vm);
 278static void virtio_mem_delete_resource(struct virtio_mem *vm);
 279
 280/*
 281 * Register a virtio-mem device so it will be considered for the online_page
 282 * callback.
 283 */
 284static int register_virtio_mem_device(struct virtio_mem *vm)
 285{
 286	int rc = 0;
 287
 288	/* First device registers the callback. */
 289	mutex_lock(&virtio_mem_mutex);
 290	if (list_empty(&virtio_mem_devices))
 291		rc = set_online_page_callback(&virtio_mem_online_page_cb);
 292	if (!rc)
 293		list_add_rcu(&vm->next, &virtio_mem_devices);
 294	mutex_unlock(&virtio_mem_mutex);
 295
 296	return rc;
 297}
 298
 299/*
 300 * Unregister a virtio-mem device so it will no longer be considered for the
 301 * online_page callback.
 302 */
 303static void unregister_virtio_mem_device(struct virtio_mem *vm)
 304{
 305	/* Last device unregisters the callback. */
 306	mutex_lock(&virtio_mem_mutex);
 307	list_del_rcu(&vm->next);
 308	if (list_empty(&virtio_mem_devices))
 309		restore_online_page_callback(&virtio_mem_online_page_cb);
 310	mutex_unlock(&virtio_mem_mutex);
 311
 312	synchronize_rcu();
 313}
 314
 315/*
 316 * Calculate the memory block id of a given address.
 317 */
 318static unsigned long virtio_mem_phys_to_mb_id(unsigned long addr)
 319{
 320	return addr / memory_block_size_bytes();
 321}
 322
 323/*
 324 * Calculate the physical start address of a given memory block id.
 325 */
 326static unsigned long virtio_mem_mb_id_to_phys(unsigned long mb_id)
 327{
 328	return mb_id * memory_block_size_bytes();
 329}
 330
 331/*
 332 * Calculate the big block id of a given address.
 333 */
 334static unsigned long virtio_mem_phys_to_bb_id(struct virtio_mem *vm,
 335					      uint64_t addr)
 336{
 337	return addr / vm->bbm.bb_size;
 338}
 339
 340/*
 341 * Calculate the physical start address of a given big block id.
 342 */
 343static uint64_t virtio_mem_bb_id_to_phys(struct virtio_mem *vm,
 344					 unsigned long bb_id)
 345{
 346	return bb_id * vm->bbm.bb_size;
 347}
 348
 349/*
 350 * Calculate the subblock id of a given address.
 351 */
 352static unsigned long virtio_mem_phys_to_sb_id(struct virtio_mem *vm,
 353					      unsigned long addr)
 354{
 355	const unsigned long mb_id = virtio_mem_phys_to_mb_id(addr);
 356	const unsigned long mb_addr = virtio_mem_mb_id_to_phys(mb_id);
 357
 358	return (addr - mb_addr) / vm->sbm.sb_size;
 359}
 360
 361/*
 362 * Set the state of a big block, taking care of the state counter.
 363 */
 364static void virtio_mem_bbm_set_bb_state(struct virtio_mem *vm,
 365					unsigned long bb_id,
 366					enum virtio_mem_bbm_bb_state state)
 367{
 368	const unsigned long idx = bb_id - vm->bbm.first_bb_id;
 369	enum virtio_mem_bbm_bb_state old_state;
 370
 371	old_state = vm->bbm.bb_states[idx];
 372	vm->bbm.bb_states[idx] = state;
 373
 374	BUG_ON(vm->bbm.bb_count[old_state] == 0);
 375	vm->bbm.bb_count[old_state]--;
 376	vm->bbm.bb_count[state]++;
 377}
 378
 379/*
 380 * Get the state of a big block.
 381 */
 382static enum virtio_mem_bbm_bb_state virtio_mem_bbm_get_bb_state(struct virtio_mem *vm,
 383								unsigned long bb_id)
 384{
 385	return vm->bbm.bb_states[bb_id - vm->bbm.first_bb_id];
 386}
 387
 388/*
 389 * Prepare the big block state array for the next big block.
 390 */
 391static int virtio_mem_bbm_bb_states_prepare_next_bb(struct virtio_mem *vm)
 392{
 393	unsigned long old_bytes = vm->bbm.next_bb_id - vm->bbm.first_bb_id;
 394	unsigned long new_bytes = old_bytes + 1;
 395	int old_pages = PFN_UP(old_bytes);
 396	int new_pages = PFN_UP(new_bytes);
 397	uint8_t *new_array;
 398
 399	if (vm->bbm.bb_states && old_pages == new_pages)
 400		return 0;
 401
 402	new_array = vzalloc(new_pages * PAGE_SIZE);
 403	if (!new_array)
 404		return -ENOMEM;
 405
 406	mutex_lock(&vm->hotplug_mutex);
 407	if (vm->bbm.bb_states)
 408		memcpy(new_array, vm->bbm.bb_states, old_pages * PAGE_SIZE);
 409	vfree(vm->bbm.bb_states);
 410	vm->bbm.bb_states = new_array;
 411	mutex_unlock(&vm->hotplug_mutex);
 412
 413	return 0;
 414}
 415
 416#define virtio_mem_bbm_for_each_bb(_vm, _bb_id, _state) \
 417	for (_bb_id = vm->bbm.first_bb_id; \
 418	     _bb_id < vm->bbm.next_bb_id && _vm->bbm.bb_count[_state]; \
 419	     _bb_id++) \
 420		if (virtio_mem_bbm_get_bb_state(_vm, _bb_id) == _state)
 421
 422#define virtio_mem_bbm_for_each_bb_rev(_vm, _bb_id, _state) \
 423	for (_bb_id = vm->bbm.next_bb_id - 1; \
 424	     _bb_id >= vm->bbm.first_bb_id && _vm->bbm.bb_count[_state]; \
 425	     _bb_id--) \
 426		if (virtio_mem_bbm_get_bb_state(_vm, _bb_id) == _state)
 427
 428/*
 429 * Set the state of a memory block, taking care of the state counter.
 430 */
 431static void virtio_mem_sbm_set_mb_state(struct virtio_mem *vm,
 432					unsigned long mb_id, uint8_t state)
 433{
 434	const unsigned long idx = mb_id - vm->sbm.first_mb_id;
 435	uint8_t old_state;
 436
 437	old_state = vm->sbm.mb_states[idx];
 438	vm->sbm.mb_states[idx] = state;
 439
 440	BUG_ON(vm->sbm.mb_count[old_state] == 0);
 441	vm->sbm.mb_count[old_state]--;
 442	vm->sbm.mb_count[state]++;
 443}
 444
 445/*
 446 * Get the state of a memory block.
 447 */
 448static uint8_t virtio_mem_sbm_get_mb_state(struct virtio_mem *vm,
 449					   unsigned long mb_id)
 450{
 451	const unsigned long idx = mb_id - vm->sbm.first_mb_id;
 452
 453	return vm->sbm.mb_states[idx];
 454}
 455
 456/*
 457 * Prepare the state array for the next memory block.
 458 */
 459static int virtio_mem_sbm_mb_states_prepare_next_mb(struct virtio_mem *vm)
 460{
 461	int old_pages = PFN_UP(vm->sbm.next_mb_id - vm->sbm.first_mb_id);
 462	int new_pages = PFN_UP(vm->sbm.next_mb_id - vm->sbm.first_mb_id + 1);
 463	uint8_t *new_array;
 464
 465	if (vm->sbm.mb_states && old_pages == new_pages)
 466		return 0;
 467
 468	new_array = vzalloc(new_pages * PAGE_SIZE);
 469	if (!new_array)
 470		return -ENOMEM;
 471
 472	mutex_lock(&vm->hotplug_mutex);
 473	if (vm->sbm.mb_states)
 474		memcpy(new_array, vm->sbm.mb_states, old_pages * PAGE_SIZE);
 475	vfree(vm->sbm.mb_states);
 476	vm->sbm.mb_states = new_array;
 477	mutex_unlock(&vm->hotplug_mutex);
 478
 479	return 0;
 480}
 481
 482#define virtio_mem_sbm_for_each_mb(_vm, _mb_id, _state) \
 483	for (_mb_id = _vm->sbm.first_mb_id; \
 484	     _mb_id < _vm->sbm.next_mb_id && _vm->sbm.mb_count[_state]; \
 485	     _mb_id++) \
 486		if (virtio_mem_sbm_get_mb_state(_vm, _mb_id) == _state)
 487
 488#define virtio_mem_sbm_for_each_mb_rev(_vm, _mb_id, _state) \
 489	for (_mb_id = _vm->sbm.next_mb_id - 1; \
 490	     _mb_id >= _vm->sbm.first_mb_id && _vm->sbm.mb_count[_state]; \
 491	     _mb_id--) \
 492		if (virtio_mem_sbm_get_mb_state(_vm, _mb_id) == _state)
 493
 494/*
 495 * Calculate the bit number in the subblock bitmap for the given subblock
 496 * inside the given memory block.
 497 */
 498static int virtio_mem_sbm_sb_state_bit_nr(struct virtio_mem *vm,
 499					  unsigned long mb_id, int sb_id)
 500{
 501	return (mb_id - vm->sbm.first_mb_id) * vm->sbm.sbs_per_mb + sb_id;
 502}
 503
 504/*
 505 * Mark all selected subblocks plugged.
 506 *
 507 * Will not modify the state of the memory block.
 508 */
 509static void virtio_mem_sbm_set_sb_plugged(struct virtio_mem *vm,
 510					  unsigned long mb_id, int sb_id,
 511					  int count)
 512{
 513	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 514
 515	__bitmap_set(vm->sbm.sb_states, bit, count);
 516}
 517
 518/*
 519 * Mark all selected subblocks unplugged.
 520 *
 521 * Will not modify the state of the memory block.
 522 */
 523static void virtio_mem_sbm_set_sb_unplugged(struct virtio_mem *vm,
 524					    unsigned long mb_id, int sb_id,
 525					    int count)
 526{
 527	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 528
 529	__bitmap_clear(vm->sbm.sb_states, bit, count);
 530}
 531
 532/*
 533 * Test if all selected subblocks are plugged.
 534 */
 535static bool virtio_mem_sbm_test_sb_plugged(struct virtio_mem *vm,
 536					   unsigned long mb_id, int sb_id,
 537					   int count)
 538{
 539	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 540
 541	if (count == 1)
 542		return test_bit(bit, vm->sbm.sb_states);
 543
 544	/* TODO: Helper similar to bitmap_set() */
 545	return find_next_zero_bit(vm->sbm.sb_states, bit + count, bit) >=
 546	       bit + count;
 547}
 548
 549/*
 550 * Test if all selected subblocks are unplugged.
 551 */
 552static bool virtio_mem_sbm_test_sb_unplugged(struct virtio_mem *vm,
 553					     unsigned long mb_id, int sb_id,
 554					     int count)
 555{
 556	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, sb_id);
 557
 558	/* TODO: Helper similar to bitmap_set() */
 559	return find_next_bit(vm->sbm.sb_states, bit + count, bit) >=
 560	       bit + count;
 561}
 562
 563/*
 564 * Find the first unplugged subblock. Returns vm->sbm.sbs_per_mb in case there is
 565 * none.
 566 */
 567static int virtio_mem_sbm_first_unplugged_sb(struct virtio_mem *vm,
 568					    unsigned long mb_id)
 569{
 570	const int bit = virtio_mem_sbm_sb_state_bit_nr(vm, mb_id, 0);
 571
 572	return find_next_zero_bit(vm->sbm.sb_states,
 573				  bit + vm->sbm.sbs_per_mb, bit) - bit;
 574}
 575
 576/*
 577 * Prepare the subblock bitmap for the next memory block.
 578 */
 579static int virtio_mem_sbm_sb_states_prepare_next_mb(struct virtio_mem *vm)
 580{
 581	const unsigned long old_nb_mb = vm->sbm.next_mb_id - vm->sbm.first_mb_id;
 582	const unsigned long old_nb_bits = old_nb_mb * vm->sbm.sbs_per_mb;
 583	const unsigned long new_nb_bits = (old_nb_mb + 1) * vm->sbm.sbs_per_mb;
 584	int old_pages = PFN_UP(BITS_TO_LONGS(old_nb_bits) * sizeof(long));
 585	int new_pages = PFN_UP(BITS_TO_LONGS(new_nb_bits) * sizeof(long));
 586	unsigned long *new_bitmap, *old_bitmap;
 587
 588	if (vm->sbm.sb_states && old_pages == new_pages)
 589		return 0;
 590
 591	new_bitmap = vzalloc(new_pages * PAGE_SIZE);
 592	if (!new_bitmap)
 593		return -ENOMEM;
 594
 595	mutex_lock(&vm->hotplug_mutex);
 596	if (vm->sbm.sb_states)
 597		memcpy(new_bitmap, vm->sbm.sb_states, old_pages * PAGE_SIZE);
 598
 599	old_bitmap = vm->sbm.sb_states;
 600	vm->sbm.sb_states = new_bitmap;
 601	mutex_unlock(&vm->hotplug_mutex);
 602
 603	vfree(old_bitmap);
 604	return 0;
 605}
 606
 607/*
 608 * Test if we could add memory without creating too much offline memory -
 609 * to avoid running OOM if memory is getting onlined deferred.
 610 */
 611static bool virtio_mem_could_add_memory(struct virtio_mem *vm, uint64_t size)
 612{
 613	if (WARN_ON_ONCE(size > vm->offline_threshold))
 614		return false;
 615
 616	return atomic64_read(&vm->offline_size) + size <= vm->offline_threshold;
 617}
 618
 619/*
 620 * Try adding memory to Linux. Will usually only fail if out of memory.
 621 *
 622 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 623 * onlining code).
 624 *
 625 * Will not modify the state of memory blocks in virtio-mem.
 626 */
 627static int virtio_mem_add_memory(struct virtio_mem *vm, uint64_t addr,
 628				 uint64_t size)
 629{
 630	int rc;
 631
 632	/*
 633	 * When force-unloading the driver and we still have memory added to
 634	 * Linux, the resource name has to stay.
 635	 */
 636	if (!vm->resource_name) {
 637		vm->resource_name = kstrdup_const("System RAM (virtio_mem)",
 638						  GFP_KERNEL);
 639		if (!vm->resource_name)
 640			return -ENOMEM;
 641	}
 642
 643	dev_dbg(&vm->vdev->dev, "adding memory: 0x%llx - 0x%llx\n", addr,
 644		addr + size - 1);
 645	/* Memory might get onlined immediately. */
 646	atomic64_add(size, &vm->offline_size);
 647	rc = add_memory_driver_managed(vm->mgid, addr, size, vm->resource_name,
 648				       MHP_MERGE_RESOURCE | MHP_NID_IS_MGID);
 649	if (rc) {
 650		atomic64_sub(size, &vm->offline_size);
 651		dev_warn(&vm->vdev->dev, "adding memory failed: %d\n", rc);
 652		/*
 653		 * TODO: Linux MM does not properly clean up yet in all cases
 654		 * where adding of memory failed - especially on -ENOMEM.
 655		 */
 656	}
 657	return rc;
 658}
 659
 660/*
 661 * See virtio_mem_add_memory(): Try adding a single Linux memory block.
 662 */
 663static int virtio_mem_sbm_add_mb(struct virtio_mem *vm, unsigned long mb_id)
 664{
 665	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 666	const uint64_t size = memory_block_size_bytes();
 667
 668	return virtio_mem_add_memory(vm, addr, size);
 669}
 670
 671/*
 672 * See virtio_mem_add_memory(): Try adding a big block.
 673 */
 674static int virtio_mem_bbm_add_bb(struct virtio_mem *vm, unsigned long bb_id)
 675{
 676	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
 677	const uint64_t size = vm->bbm.bb_size;
 678
 679	return virtio_mem_add_memory(vm, addr, size);
 680}
 681
 682/*
 683 * Try removing memory from Linux. Will only fail if memory blocks aren't
 684 * offline.
 685 *
 686 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 687 * onlining code).
 688 *
 689 * Will not modify the state of memory blocks in virtio-mem.
 690 */
 691static int virtio_mem_remove_memory(struct virtio_mem *vm, uint64_t addr,
 692				    uint64_t size)
 693{
 694	int rc;
 695
 696	dev_dbg(&vm->vdev->dev, "removing memory: 0x%llx - 0x%llx\n", addr,
 697		addr + size - 1);
 698	rc = remove_memory(addr, size);
 699	if (!rc) {
 700		atomic64_sub(size, &vm->offline_size);
 701		/*
 702		 * We might have freed up memory we can now unplug, retry
 703		 * immediately instead of waiting.
 704		 */
 705		virtio_mem_retry(vm);
 706	} else {
 707		dev_dbg(&vm->vdev->dev, "removing memory failed: %d\n", rc);
 708	}
 709	return rc;
 710}
 711
 712/*
 713 * See virtio_mem_remove_memory(): Try removing a single Linux memory block.
 714 */
 715static int virtio_mem_sbm_remove_mb(struct virtio_mem *vm, unsigned long mb_id)
 716{
 717	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 718	const uint64_t size = memory_block_size_bytes();
 719
 720	return virtio_mem_remove_memory(vm, addr, size);
 721}
 722
 723/*
 724 * Try offlining and removing memory from Linux.
 725 *
 726 * Must not be called with the vm->hotplug_mutex held (possible deadlock with
 727 * onlining code).
 728 *
 729 * Will not modify the state of memory blocks in virtio-mem.
 730 */
 731static int virtio_mem_offline_and_remove_memory(struct virtio_mem *vm,
 732						uint64_t addr,
 733						uint64_t size)
 734{
 735	int rc;
 736
 737	dev_dbg(&vm->vdev->dev,
 738		"offlining and removing memory: 0x%llx - 0x%llx\n", addr,
 739		addr + size - 1);
 740
 741	rc = offline_and_remove_memory(addr, size);
 742	if (!rc) {
 743		atomic64_sub(size, &vm->offline_size);
 744		/*
 745		 * We might have freed up memory we can now unplug, retry
 746		 * immediately instead of waiting.
 747		 */
 748		virtio_mem_retry(vm);
 749	} else {
 750		dev_dbg(&vm->vdev->dev,
 751			"offlining and removing memory failed: %d\n", rc);
 752	}
 753	return rc;
 
 
 
 
 
 
 754}
 755
 756/*
 757 * See virtio_mem_offline_and_remove_memory(): Try offlining and removing
 758 * a single Linux memory block.
 759 */
 760static int virtio_mem_sbm_offline_and_remove_mb(struct virtio_mem *vm,
 761						unsigned long mb_id)
 762{
 763	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id);
 764	const uint64_t size = memory_block_size_bytes();
 765
 766	return virtio_mem_offline_and_remove_memory(vm, addr, size);
 767}
 768
 769/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 770 * See virtio_mem_offline_and_remove_memory(): Try to offline and remove a
 771 * all Linux memory blocks covered by the big block.
 772 */
 773static int virtio_mem_bbm_offline_and_remove_bb(struct virtio_mem *vm,
 774						unsigned long bb_id)
 775{
 776	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
 777	const uint64_t size = vm->bbm.bb_size;
 778
 779	return virtio_mem_offline_and_remove_memory(vm, addr, size);
 780}
 781
 782/*
 783 * Trigger the workqueue so the device can perform its magic.
 784 */
 785static void virtio_mem_retry(struct virtio_mem *vm)
 786{
 787	unsigned long flags;
 788
 789	spin_lock_irqsave(&vm->removal_lock, flags);
 790	if (!vm->removing)
 791		queue_work(system_freezable_wq, &vm->wq);
 792	spin_unlock_irqrestore(&vm->removal_lock, flags);
 793}
 794
 795static int virtio_mem_translate_node_id(struct virtio_mem *vm, uint16_t node_id)
 796{
 797	int node = NUMA_NO_NODE;
 798
 799#if defined(CONFIG_ACPI_NUMA)
 800	if (virtio_has_feature(vm->vdev, VIRTIO_MEM_F_ACPI_PXM))
 801		node = pxm_to_node(node_id);
 802#endif
 803	return node;
 804}
 805
 806/*
 807 * Test if a virtio-mem device overlaps with the given range. Can be called
 808 * from (notifier) callbacks lockless.
 809 */
 810static bool virtio_mem_overlaps_range(struct virtio_mem *vm, uint64_t start,
 811				      uint64_t size)
 812{
 813	return start < vm->addr + vm->region_size && vm->addr < start + size;
 814}
 815
 816/*
 817 * Test if a virtio-mem device contains a given range. Can be called from
 818 * (notifier) callbacks lockless.
 819 */
 820static bool virtio_mem_contains_range(struct virtio_mem *vm, uint64_t start,
 821				      uint64_t size)
 822{
 823	return start >= vm->addr && start + size <= vm->addr + vm->region_size;
 824}
 825
 826static int virtio_mem_sbm_notify_going_online(struct virtio_mem *vm,
 827					      unsigned long mb_id)
 828{
 829	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 830	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
 831	case VIRTIO_MEM_SBM_MB_OFFLINE:
 832		return NOTIFY_OK;
 833	default:
 834		break;
 835	}
 836	dev_warn_ratelimited(&vm->vdev->dev,
 837			     "memory block onlining denied\n");
 838	return NOTIFY_BAD;
 839}
 840
 841static void virtio_mem_sbm_notify_offline(struct virtio_mem *vm,
 842					  unsigned long mb_id)
 843{
 844	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 845	case VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL:
 846	case VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL:
 847		virtio_mem_sbm_set_mb_state(vm, mb_id,
 848					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
 849		break;
 850	case VIRTIO_MEM_SBM_MB_KERNEL:
 851	case VIRTIO_MEM_SBM_MB_MOVABLE:
 852		virtio_mem_sbm_set_mb_state(vm, mb_id,
 853					    VIRTIO_MEM_SBM_MB_OFFLINE);
 854		break;
 855	default:
 856		BUG();
 857		break;
 858	}
 859}
 860
 861static void virtio_mem_sbm_notify_online(struct virtio_mem *vm,
 862					 unsigned long mb_id,
 863					 unsigned long start_pfn)
 864{
 865	const bool is_movable = is_zone_movable_page(pfn_to_page(start_pfn));
 866	int new_state;
 867
 868	switch (virtio_mem_sbm_get_mb_state(vm, mb_id)) {
 869	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
 870		new_state = VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL;
 871		if (is_movable)
 872			new_state = VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL;
 873		break;
 874	case VIRTIO_MEM_SBM_MB_OFFLINE:
 875		new_state = VIRTIO_MEM_SBM_MB_KERNEL;
 876		if (is_movable)
 877			new_state = VIRTIO_MEM_SBM_MB_MOVABLE;
 878		break;
 879	default:
 880		BUG();
 881		break;
 882	}
 883	virtio_mem_sbm_set_mb_state(vm, mb_id, new_state);
 884}
 885
 886static void virtio_mem_sbm_notify_going_offline(struct virtio_mem *vm,
 887						unsigned long mb_id)
 888{
 889	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size);
 890	unsigned long pfn;
 891	int sb_id;
 892
 893	for (sb_id = 0; sb_id < vm->sbm.sbs_per_mb; sb_id++) {
 894		if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
 895			continue;
 896		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
 897			       sb_id * vm->sbm.sb_size);
 898		virtio_mem_fake_offline_going_offline(pfn, nr_pages);
 899	}
 900}
 901
 902static void virtio_mem_sbm_notify_cancel_offline(struct virtio_mem *vm,
 903						 unsigned long mb_id)
 904{
 905	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size);
 906	unsigned long pfn;
 907	int sb_id;
 908
 909	for (sb_id = 0; sb_id < vm->sbm.sbs_per_mb; sb_id++) {
 910		if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
 911			continue;
 912		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
 913			       sb_id * vm->sbm.sb_size);
 914		virtio_mem_fake_offline_cancel_offline(pfn, nr_pages);
 915	}
 916}
 917
 918static void virtio_mem_bbm_notify_going_offline(struct virtio_mem *vm,
 919						unsigned long bb_id,
 920						unsigned long pfn,
 921						unsigned long nr_pages)
 922{
 923	/*
 924	 * When marked as "fake-offline", all online memory of this device block
 925	 * is allocated by us. Otherwise, we don't have any memory allocated.
 926	 */
 927	if (virtio_mem_bbm_get_bb_state(vm, bb_id) !=
 928	    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE)
 929		return;
 930	virtio_mem_fake_offline_going_offline(pfn, nr_pages);
 931}
 932
 933static void virtio_mem_bbm_notify_cancel_offline(struct virtio_mem *vm,
 934						 unsigned long bb_id,
 935						 unsigned long pfn,
 936						 unsigned long nr_pages)
 937{
 938	if (virtio_mem_bbm_get_bb_state(vm, bb_id) !=
 939	    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE)
 940		return;
 941	virtio_mem_fake_offline_cancel_offline(pfn, nr_pages);
 942}
 943
 944/*
 945 * This callback will either be called synchronously from add_memory() or
 946 * asynchronously (e.g., triggered via user space). We have to be careful
 947 * with locking when calling add_memory().
 948 */
 949static int virtio_mem_memory_notifier_cb(struct notifier_block *nb,
 950					 unsigned long action, void *arg)
 951{
 952	struct virtio_mem *vm = container_of(nb, struct virtio_mem,
 953					     memory_notifier);
 954	struct memory_notify *mhp = arg;
 955	const unsigned long start = PFN_PHYS(mhp->start_pfn);
 956	const unsigned long size = PFN_PHYS(mhp->nr_pages);
 957	int rc = NOTIFY_OK;
 958	unsigned long id;
 959
 960	if (!virtio_mem_overlaps_range(vm, start, size))
 961		return NOTIFY_DONE;
 962
 963	if (vm->in_sbm) {
 964		id = virtio_mem_phys_to_mb_id(start);
 965		/*
 966		 * In SBM, we add memory in separate memory blocks - we expect
 967		 * it to be onlined/offlined in the same granularity. Bail out
 968		 * if this ever changes.
 969		 */
 970		if (WARN_ON_ONCE(size != memory_block_size_bytes() ||
 971				 !IS_ALIGNED(start, memory_block_size_bytes())))
 972			return NOTIFY_BAD;
 973	} else {
 974		id = virtio_mem_phys_to_bb_id(vm, start);
 975		/*
 976		 * In BBM, we only care about onlining/offlining happening
 977		 * within a single big block, we don't care about the
 978		 * actual granularity as we don't track individual Linux
 979		 * memory blocks.
 980		 */
 981		if (WARN_ON_ONCE(id != virtio_mem_phys_to_bb_id(vm, start + size - 1)))
 982			return NOTIFY_BAD;
 983	}
 984
 985	/*
 986	 * Avoid circular locking lockdep warnings. We lock the mutex
 987	 * e.g., in MEM_GOING_ONLINE and unlock it in MEM_ONLINE. The
 988	 * blocking_notifier_call_chain() has it's own lock, which gets unlocked
 989	 * between both notifier calls and will bail out. False positive.
 990	 */
 991	lockdep_off();
 992
 993	switch (action) {
 994	case MEM_GOING_OFFLINE:
 995		mutex_lock(&vm->hotplug_mutex);
 996		if (vm->removing) {
 997			rc = notifier_from_errno(-EBUSY);
 998			mutex_unlock(&vm->hotplug_mutex);
 999			break;
1000		}
1001		vm->hotplug_active = true;
1002		if (vm->in_sbm)
1003			virtio_mem_sbm_notify_going_offline(vm, id);
1004		else
1005			virtio_mem_bbm_notify_going_offline(vm, id,
1006							    mhp->start_pfn,
1007							    mhp->nr_pages);
1008		break;
1009	case MEM_GOING_ONLINE:
1010		mutex_lock(&vm->hotplug_mutex);
1011		if (vm->removing) {
1012			rc = notifier_from_errno(-EBUSY);
1013			mutex_unlock(&vm->hotplug_mutex);
1014			break;
1015		}
1016		vm->hotplug_active = true;
1017		if (vm->in_sbm)
1018			rc = virtio_mem_sbm_notify_going_online(vm, id);
1019		break;
1020	case MEM_OFFLINE:
1021		if (vm->in_sbm)
1022			virtio_mem_sbm_notify_offline(vm, id);
1023
1024		atomic64_add(size, &vm->offline_size);
1025		/*
1026		 * Trigger the workqueue. Now that we have some offline memory,
1027		 * maybe we can handle pending unplug requests.
1028		 */
1029		if (!unplug_online)
1030			virtio_mem_retry(vm);
1031
1032		vm->hotplug_active = false;
1033		mutex_unlock(&vm->hotplug_mutex);
1034		break;
1035	case MEM_ONLINE:
1036		if (vm->in_sbm)
1037			virtio_mem_sbm_notify_online(vm, id, mhp->start_pfn);
1038
1039		atomic64_sub(size, &vm->offline_size);
1040		/*
1041		 * Start adding more memory once we onlined half of our
1042		 * threshold. Don't trigger if it's possibly due to our actipn
1043		 * (e.g., us adding memory which gets onlined immediately from
1044		 * the core).
1045		 */
1046		if (!atomic_read(&vm->wq_active) &&
1047		    virtio_mem_could_add_memory(vm, vm->offline_threshold / 2))
1048			virtio_mem_retry(vm);
1049
1050		vm->hotplug_active = false;
1051		mutex_unlock(&vm->hotplug_mutex);
1052		break;
1053	case MEM_CANCEL_OFFLINE:
1054		if (!vm->hotplug_active)
1055			break;
1056		if (vm->in_sbm)
1057			virtio_mem_sbm_notify_cancel_offline(vm, id);
1058		else
1059			virtio_mem_bbm_notify_cancel_offline(vm, id,
1060							     mhp->start_pfn,
1061							     mhp->nr_pages);
1062		vm->hotplug_active = false;
1063		mutex_unlock(&vm->hotplug_mutex);
1064		break;
1065	case MEM_CANCEL_ONLINE:
1066		if (!vm->hotplug_active)
1067			break;
1068		vm->hotplug_active = false;
1069		mutex_unlock(&vm->hotplug_mutex);
1070		break;
1071	default:
1072		break;
1073	}
1074
1075	lockdep_on();
1076
1077	return rc;
1078}
1079
1080/*
1081 * Set a range of pages PG_offline. Remember pages that were never onlined
1082 * (via generic_online_page()) using PageDirty().
1083 */
1084static void virtio_mem_set_fake_offline(unsigned long pfn,
1085					unsigned long nr_pages, bool onlined)
1086{
1087	page_offline_begin();
1088	for (; nr_pages--; pfn++) {
1089		struct page *page = pfn_to_page(pfn);
1090
1091		__SetPageOffline(page);
1092		if (!onlined) {
1093			SetPageDirty(page);
1094			/* FIXME: remove after cleanups */
1095			ClearPageReserved(page);
1096		}
1097	}
1098	page_offline_end();
1099}
1100
1101/*
1102 * Clear PG_offline from a range of pages. If the pages were never onlined,
1103 * (via generic_online_page()), clear PageDirty().
1104 */
1105static void virtio_mem_clear_fake_offline(unsigned long pfn,
1106					  unsigned long nr_pages, bool onlined)
1107{
1108	for (; nr_pages--; pfn++) {
1109		struct page *page = pfn_to_page(pfn);
1110
1111		__ClearPageOffline(page);
1112		if (!onlined)
1113			ClearPageDirty(page);
1114	}
1115}
1116
1117/*
1118 * Release a range of fake-offline pages to the buddy, effectively
1119 * fake-onlining them.
1120 */
1121static void virtio_mem_fake_online(unsigned long pfn, unsigned long nr_pages)
1122{
1123	unsigned long order = MAX_ORDER - 1;
1124	unsigned long i;
1125
1126	/*
1127	 * We might get called for ranges that don't cover properly aligned
1128	 * MAX_ORDER - 1 pages; however, we can only online properly aligned
1129	 * pages with an order of MAX_ORDER - 1 at maximum.
1130	 */
1131	while (!IS_ALIGNED(pfn | nr_pages, 1 << order))
1132		order--;
1133
1134	for (i = 0; i < nr_pages; i += 1 << order) {
1135		struct page *page = pfn_to_page(pfn + i);
1136
1137		/*
1138		 * If the page is PageDirty(), it was kept fake-offline when
1139		 * onlining the memory block. Otherwise, it was allocated
1140		 * using alloc_contig_range(). All pages in a subblock are
1141		 * alike.
1142		 */
1143		if (PageDirty(page)) {
1144			virtio_mem_clear_fake_offline(pfn + i, 1 << order, false);
1145			generic_online_page(page, order);
1146		} else {
1147			virtio_mem_clear_fake_offline(pfn + i, 1 << order, true);
1148			free_contig_range(pfn + i, 1 << order);
1149			adjust_managed_page_count(page, 1 << order);
1150		}
1151	}
1152}
1153
1154/*
1155 * Try to allocate a range, marking pages fake-offline, effectively
1156 * fake-offlining them.
1157 */
1158static int virtio_mem_fake_offline(unsigned long pfn, unsigned long nr_pages)
 
1159{
1160	const bool is_movable = is_zone_movable_page(pfn_to_page(pfn));
1161	int rc, retry_count;
1162
1163	/*
1164	 * TODO: We want an alloc_contig_range() mode that tries to allocate
1165	 * harder (e.g., dealing with temporarily pinned pages, PCP), especially
1166	 * with ZONE_MOVABLE. So for now, retry a couple of times with
1167	 * ZONE_MOVABLE before giving up - because that zone is supposed to give
1168	 * some guarantees.
1169	 */
1170	for (retry_count = 0; retry_count < 5; retry_count++) {
 
 
 
 
 
 
 
 
1171		rc = alloc_contig_range(pfn, pfn + nr_pages, MIGRATE_MOVABLE,
1172					GFP_KERNEL);
1173		if (rc == -ENOMEM)
1174			/* whoops, out of memory */
1175			return rc;
1176		else if (rc && !is_movable)
1177			break;
1178		else if (rc)
1179			continue;
1180
1181		virtio_mem_set_fake_offline(pfn, nr_pages, true);
1182		adjust_managed_page_count(pfn_to_page(pfn), -nr_pages);
1183		return 0;
1184	}
1185
1186	return -EBUSY;
1187}
1188
1189/*
1190 * Handle fake-offline pages when memory is going offline - such that the
1191 * pages can be skipped by mm-core when offlining.
1192 */
1193static void virtio_mem_fake_offline_going_offline(unsigned long pfn,
1194						  unsigned long nr_pages)
1195{
1196	struct page *page;
1197	unsigned long i;
1198
1199	/*
1200	 * Drop our reference to the pages so the memory can get offlined
1201	 * and add the unplugged pages to the managed page counters (so
1202	 * offlining code can correctly subtract them again).
1203	 */
1204	adjust_managed_page_count(pfn_to_page(pfn), nr_pages);
1205	/* Drop our reference to the pages so the memory can get offlined. */
1206	for (i = 0; i < nr_pages; i++) {
1207		page = pfn_to_page(pfn + i);
1208		if (WARN_ON(!page_ref_dec_and_test(page)))
1209			dump_page(page, "fake-offline page referenced");
1210	}
1211}
1212
1213/*
1214 * Handle fake-offline pages when memory offlining is canceled - to undo
1215 * what we did in virtio_mem_fake_offline_going_offline().
1216 */
1217static void virtio_mem_fake_offline_cancel_offline(unsigned long pfn,
1218						   unsigned long nr_pages)
1219{
1220	unsigned long i;
1221
1222	/*
1223	 * Get the reference we dropped when going offline and subtract the
1224	 * unplugged pages from the managed page counters.
1225	 */
1226	adjust_managed_page_count(pfn_to_page(pfn), -nr_pages);
1227	for (i = 0; i < nr_pages; i++)
1228		page_ref_inc(pfn_to_page(pfn + i));
1229}
1230
1231static void virtio_mem_online_page(struct virtio_mem *vm,
1232				   struct page *page, unsigned int order)
1233{
1234	const unsigned long start = page_to_phys(page);
1235	const unsigned long end = start + PFN_PHYS(1 << order);
1236	unsigned long addr, next, id, sb_id, count;
1237	bool do_online;
1238
1239	/*
1240	 * We can get called with any order up to MAX_ORDER - 1. If our
1241	 * subblock size is smaller than that and we have a mixture of plugged
1242	 * and unplugged subblocks within such a page, we have to process in
1243	 * smaller granularity. In that case we'll adjust the order exactly once
1244	 * within the loop.
1245	 */
1246	for (addr = start; addr < end; ) {
1247		next = addr + PFN_PHYS(1 << order);
1248
1249		if (vm->in_sbm) {
1250			id = virtio_mem_phys_to_mb_id(addr);
1251			sb_id = virtio_mem_phys_to_sb_id(vm, addr);
1252			count = virtio_mem_phys_to_sb_id(vm, next - 1) - sb_id + 1;
1253
1254			if (virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, count)) {
1255				/* Fully plugged. */
1256				do_online = true;
1257			} else if (count == 1 ||
1258				   virtio_mem_sbm_test_sb_unplugged(vm, id, sb_id, count)) {
1259				/* Fully unplugged. */
1260				do_online = false;
1261			} else {
1262				/*
1263				 * Mixture, process sub-blocks instead. This
1264				 * will be at least the size of a pageblock.
1265				 * We'll run into this case exactly once.
1266				 */
1267				order = ilog2(vm->sbm.sb_size) - PAGE_SHIFT;
1268				do_online = virtio_mem_sbm_test_sb_plugged(vm, id, sb_id, 1);
1269				continue;
1270			}
1271		} else {
1272			/*
1273			 * If the whole block is marked fake offline, keep
1274			 * everything that way.
1275			 */
1276			id = virtio_mem_phys_to_bb_id(vm, addr);
1277			do_online = virtio_mem_bbm_get_bb_state(vm, id) !=
1278				    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE;
1279		}
1280
1281		if (do_online)
1282			generic_online_page(pfn_to_page(PFN_DOWN(addr)), order);
1283		else
1284			virtio_mem_set_fake_offline(PFN_DOWN(addr), 1 << order,
1285						    false);
1286		addr = next;
1287	}
1288}
1289
1290static void virtio_mem_online_page_cb(struct page *page, unsigned int order)
1291{
1292	const unsigned long addr = page_to_phys(page);
1293	struct virtio_mem *vm;
1294
1295	rcu_read_lock();
1296	list_for_each_entry_rcu(vm, &virtio_mem_devices, next) {
1297		/*
1298		 * Pages we're onlining will never cross memory blocks and,
1299		 * therefore, not virtio-mem devices.
1300		 */
1301		if (!virtio_mem_contains_range(vm, addr, PFN_PHYS(1 << order)))
1302			continue;
1303
1304		/*
1305		 * virtio_mem_set_fake_offline() might sleep. We can safely
1306		 * drop the RCU lock at this point because the device
1307		 * cannot go away. See virtio_mem_remove() how races
1308		 * between memory onlining and device removal are handled.
1309		 */
1310		rcu_read_unlock();
1311
1312		virtio_mem_online_page(vm, page, order);
1313		return;
1314	}
1315	rcu_read_unlock();
1316
1317	/* not virtio-mem memory, but e.g., a DIMM. online it */
1318	generic_online_page(page, order);
1319}
1320
1321static uint64_t virtio_mem_send_request(struct virtio_mem *vm,
1322					const struct virtio_mem_req *req)
1323{
1324	struct scatterlist *sgs[2], sg_req, sg_resp;
1325	unsigned int len;
1326	int rc;
1327
1328	/* don't use the request residing on the stack (vaddr) */
1329	vm->req = *req;
1330
1331	/* out: buffer for request */
1332	sg_init_one(&sg_req, &vm->req, sizeof(vm->req));
1333	sgs[0] = &sg_req;
1334
1335	/* in: buffer for response */
1336	sg_init_one(&sg_resp, &vm->resp, sizeof(vm->resp));
1337	sgs[1] = &sg_resp;
1338
1339	rc = virtqueue_add_sgs(vm->vq, sgs, 1, 1, vm, GFP_KERNEL);
1340	if (rc < 0)
1341		return rc;
1342
1343	virtqueue_kick(vm->vq);
1344
1345	/* wait for a response */
1346	wait_event(vm->host_resp, virtqueue_get_buf(vm->vq, &len));
1347
1348	return virtio16_to_cpu(vm->vdev, vm->resp.type);
1349}
1350
1351static int virtio_mem_send_plug_request(struct virtio_mem *vm, uint64_t addr,
1352					uint64_t size)
1353{
1354	const uint64_t nb_vm_blocks = size / vm->device_block_size;
1355	const struct virtio_mem_req req = {
1356		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_PLUG),
1357		.u.plug.addr = cpu_to_virtio64(vm->vdev, addr),
1358		.u.plug.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
1359	};
1360	int rc = -ENOMEM;
1361
1362	if (atomic_read(&vm->config_changed))
1363		return -EAGAIN;
1364
1365	dev_dbg(&vm->vdev->dev, "plugging memory: 0x%llx - 0x%llx\n", addr,
1366		addr + size - 1);
1367
1368	switch (virtio_mem_send_request(vm, &req)) {
1369	case VIRTIO_MEM_RESP_ACK:
1370		vm->plugged_size += size;
1371		return 0;
1372	case VIRTIO_MEM_RESP_NACK:
1373		rc = -EAGAIN;
1374		break;
1375	case VIRTIO_MEM_RESP_BUSY:
1376		rc = -ETXTBSY;
1377		break;
1378	case VIRTIO_MEM_RESP_ERROR:
1379		rc = -EINVAL;
1380		break;
1381	default:
1382		break;
1383	}
1384
1385	dev_dbg(&vm->vdev->dev, "plugging memory failed: %d\n", rc);
1386	return rc;
1387}
1388
1389static int virtio_mem_send_unplug_request(struct virtio_mem *vm, uint64_t addr,
1390					  uint64_t size)
1391{
1392	const uint64_t nb_vm_blocks = size / vm->device_block_size;
1393	const struct virtio_mem_req req = {
1394		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_UNPLUG),
1395		.u.unplug.addr = cpu_to_virtio64(vm->vdev, addr),
1396		.u.unplug.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
1397	};
1398	int rc = -ENOMEM;
1399
1400	if (atomic_read(&vm->config_changed))
1401		return -EAGAIN;
1402
1403	dev_dbg(&vm->vdev->dev, "unplugging memory: 0x%llx - 0x%llx\n", addr,
1404		addr + size - 1);
1405
1406	switch (virtio_mem_send_request(vm, &req)) {
1407	case VIRTIO_MEM_RESP_ACK:
1408		vm->plugged_size -= size;
1409		return 0;
1410	case VIRTIO_MEM_RESP_BUSY:
1411		rc = -ETXTBSY;
1412		break;
1413	case VIRTIO_MEM_RESP_ERROR:
1414		rc = -EINVAL;
1415		break;
1416	default:
1417		break;
1418	}
1419
1420	dev_dbg(&vm->vdev->dev, "unplugging memory failed: %d\n", rc);
1421	return rc;
1422}
1423
1424static int virtio_mem_send_unplug_all_request(struct virtio_mem *vm)
1425{
1426	const struct virtio_mem_req req = {
1427		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_UNPLUG_ALL),
1428	};
1429	int rc = -ENOMEM;
1430
1431	dev_dbg(&vm->vdev->dev, "unplugging all memory");
1432
1433	switch (virtio_mem_send_request(vm, &req)) {
1434	case VIRTIO_MEM_RESP_ACK:
1435		vm->unplug_all_required = false;
1436		vm->plugged_size = 0;
1437		/* usable region might have shrunk */
1438		atomic_set(&vm->config_changed, 1);
1439		return 0;
1440	case VIRTIO_MEM_RESP_BUSY:
1441		rc = -ETXTBSY;
1442		break;
1443	default:
1444		break;
1445	}
1446
1447	dev_dbg(&vm->vdev->dev, "unplugging all memory failed: %d\n", rc);
1448	return rc;
1449}
1450
1451/*
1452 * Plug selected subblocks. Updates the plugged state, but not the state
1453 * of the memory block.
1454 */
1455static int virtio_mem_sbm_plug_sb(struct virtio_mem *vm, unsigned long mb_id,
1456				  int sb_id, int count)
1457{
1458	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id) +
1459			      sb_id * vm->sbm.sb_size;
1460	const uint64_t size = count * vm->sbm.sb_size;
1461	int rc;
1462
1463	rc = virtio_mem_send_plug_request(vm, addr, size);
1464	if (!rc)
1465		virtio_mem_sbm_set_sb_plugged(vm, mb_id, sb_id, count);
1466	return rc;
1467}
1468
1469/*
1470 * Unplug selected subblocks. Updates the plugged state, but not the state
1471 * of the memory block.
1472 */
1473static int virtio_mem_sbm_unplug_sb(struct virtio_mem *vm, unsigned long mb_id,
1474				    int sb_id, int count)
1475{
1476	const uint64_t addr = virtio_mem_mb_id_to_phys(mb_id) +
1477			      sb_id * vm->sbm.sb_size;
1478	const uint64_t size = count * vm->sbm.sb_size;
1479	int rc;
1480
1481	rc = virtio_mem_send_unplug_request(vm, addr, size);
1482	if (!rc)
1483		virtio_mem_sbm_set_sb_unplugged(vm, mb_id, sb_id, count);
1484	return rc;
1485}
1486
1487/*
1488 * Request to unplug a big block.
1489 *
1490 * Will not modify the state of the big block.
1491 */
1492static int virtio_mem_bbm_unplug_bb(struct virtio_mem *vm, unsigned long bb_id)
1493{
1494	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
1495	const uint64_t size = vm->bbm.bb_size;
1496
1497	return virtio_mem_send_unplug_request(vm, addr, size);
1498}
1499
1500/*
1501 * Request to plug a big block.
1502 *
1503 * Will not modify the state of the big block.
1504 */
1505static int virtio_mem_bbm_plug_bb(struct virtio_mem *vm, unsigned long bb_id)
1506{
1507	const uint64_t addr = virtio_mem_bb_id_to_phys(vm, bb_id);
1508	const uint64_t size = vm->bbm.bb_size;
1509
1510	return virtio_mem_send_plug_request(vm, addr, size);
1511}
1512
1513/*
1514 * Unplug the desired number of plugged subblocks of a offline or not-added
1515 * memory block. Will fail if any subblock cannot get unplugged (instead of
1516 * skipping it).
1517 *
1518 * Will not modify the state of the memory block.
1519 *
1520 * Note: can fail after some subblocks were unplugged.
1521 */
1522static int virtio_mem_sbm_unplug_any_sb_raw(struct virtio_mem *vm,
1523					    unsigned long mb_id, uint64_t *nb_sb)
1524{
1525	int sb_id, count;
1526	int rc;
1527
1528	sb_id = vm->sbm.sbs_per_mb - 1;
1529	while (*nb_sb) {
1530		/* Find the next candidate subblock */
1531		while (sb_id >= 0 &&
1532		       virtio_mem_sbm_test_sb_unplugged(vm, mb_id, sb_id, 1))
1533			sb_id--;
1534		if (sb_id < 0)
1535			break;
1536		/* Try to unplug multiple subblocks at a time */
1537		count = 1;
1538		while (count < *nb_sb && sb_id > 0 &&
1539		       virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id - 1, 1)) {
1540			count++;
1541			sb_id--;
1542		}
1543
1544		rc = virtio_mem_sbm_unplug_sb(vm, mb_id, sb_id, count);
1545		if (rc)
1546			return rc;
1547		*nb_sb -= count;
1548		sb_id--;
1549	}
1550
1551	return 0;
1552}
1553
1554/*
1555 * Unplug all plugged subblocks of an offline or not-added memory block.
1556 *
1557 * Will not modify the state of the memory block.
1558 *
1559 * Note: can fail after some subblocks were unplugged.
1560 */
1561static int virtio_mem_sbm_unplug_mb(struct virtio_mem *vm, unsigned long mb_id)
1562{
1563	uint64_t nb_sb = vm->sbm.sbs_per_mb;
1564
1565	return virtio_mem_sbm_unplug_any_sb_raw(vm, mb_id, &nb_sb);
1566}
1567
1568/*
1569 * Prepare tracking data for the next memory block.
1570 */
1571static int virtio_mem_sbm_prepare_next_mb(struct virtio_mem *vm,
1572					  unsigned long *mb_id)
1573{
1574	int rc;
1575
1576	if (vm->sbm.next_mb_id > vm->sbm.last_usable_mb_id)
1577		return -ENOSPC;
1578
1579	/* Resize the state array if required. */
1580	rc = virtio_mem_sbm_mb_states_prepare_next_mb(vm);
1581	if (rc)
1582		return rc;
1583
1584	/* Resize the subblock bitmap if required. */
1585	rc = virtio_mem_sbm_sb_states_prepare_next_mb(vm);
1586	if (rc)
1587		return rc;
1588
1589	vm->sbm.mb_count[VIRTIO_MEM_SBM_MB_UNUSED]++;
1590	*mb_id = vm->sbm.next_mb_id++;
1591	return 0;
1592}
1593
1594/*
1595 * Try to plug the desired number of subblocks and add the memory block
1596 * to Linux.
1597 *
1598 * Will modify the state of the memory block.
1599 */
1600static int virtio_mem_sbm_plug_and_add_mb(struct virtio_mem *vm,
1601					  unsigned long mb_id, uint64_t *nb_sb)
1602{
1603	const int count = min_t(int, *nb_sb, vm->sbm.sbs_per_mb);
1604	int rc;
1605
1606	if (WARN_ON_ONCE(!count))
1607		return -EINVAL;
1608
1609	/*
1610	 * Plug the requested number of subblocks before adding it to linux,
1611	 * so that onlining will directly online all plugged subblocks.
1612	 */
1613	rc = virtio_mem_sbm_plug_sb(vm, mb_id, 0, count);
1614	if (rc)
1615		return rc;
1616
1617	/*
1618	 * Mark the block properly offline before adding it to Linux,
1619	 * so the memory notifiers will find the block in the right state.
1620	 */
1621	if (count == vm->sbm.sbs_per_mb)
1622		virtio_mem_sbm_set_mb_state(vm, mb_id,
1623					    VIRTIO_MEM_SBM_MB_OFFLINE);
1624	else
1625		virtio_mem_sbm_set_mb_state(vm, mb_id,
1626					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
1627
1628	/* Add the memory block to linux - if that fails, try to unplug. */
1629	rc = virtio_mem_sbm_add_mb(vm, mb_id);
1630	if (rc) {
1631		int new_state = VIRTIO_MEM_SBM_MB_UNUSED;
1632
1633		if (virtio_mem_sbm_unplug_sb(vm, mb_id, 0, count))
1634			new_state = VIRTIO_MEM_SBM_MB_PLUGGED;
1635		virtio_mem_sbm_set_mb_state(vm, mb_id, new_state);
1636		return rc;
1637	}
1638
1639	*nb_sb -= count;
1640	return 0;
1641}
1642
1643/*
1644 * Try to plug the desired number of subblocks of a memory block that
1645 * is already added to Linux.
1646 *
1647 * Will modify the state of the memory block.
1648 *
1649 * Note: Can fail after some subblocks were successfully plugged.
1650 */
1651static int virtio_mem_sbm_plug_any_sb(struct virtio_mem *vm,
1652				      unsigned long mb_id, uint64_t *nb_sb)
1653{
1654	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
1655	unsigned long pfn, nr_pages;
1656	int sb_id, count;
1657	int rc;
1658
1659	if (WARN_ON_ONCE(!*nb_sb))
1660		return -EINVAL;
1661
1662	while (*nb_sb) {
1663		sb_id = virtio_mem_sbm_first_unplugged_sb(vm, mb_id);
1664		if (sb_id >= vm->sbm.sbs_per_mb)
1665			break;
1666		count = 1;
1667		while (count < *nb_sb &&
1668		       sb_id + count < vm->sbm.sbs_per_mb &&
1669		       !virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id + count, 1))
1670			count++;
1671
1672		rc = virtio_mem_sbm_plug_sb(vm, mb_id, sb_id, count);
1673		if (rc)
1674			return rc;
1675		*nb_sb -= count;
1676		if (old_state == VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL)
1677			continue;
1678
1679		/* fake-online the pages if the memory block is online */
1680		pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
1681			       sb_id * vm->sbm.sb_size);
1682		nr_pages = PFN_DOWN(count * vm->sbm.sb_size);
1683		virtio_mem_fake_online(pfn, nr_pages);
1684	}
1685
1686	if (virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb))
1687		virtio_mem_sbm_set_mb_state(vm, mb_id, old_state - 1);
1688
1689	return 0;
1690}
1691
1692static int virtio_mem_sbm_plug_request(struct virtio_mem *vm, uint64_t diff)
1693{
1694	const int mb_states[] = {
1695		VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
1696		VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
1697		VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
1698	};
1699	uint64_t nb_sb = diff / vm->sbm.sb_size;
1700	unsigned long mb_id;
1701	int rc, i;
1702
1703	if (!nb_sb)
1704		return 0;
1705
1706	/* Don't race with onlining/offlining */
1707	mutex_lock(&vm->hotplug_mutex);
1708
1709	for (i = 0; i < ARRAY_SIZE(mb_states); i++) {
1710		virtio_mem_sbm_for_each_mb(vm, mb_id, mb_states[i]) {
1711			rc = virtio_mem_sbm_plug_any_sb(vm, mb_id, &nb_sb);
1712			if (rc || !nb_sb)
1713				goto out_unlock;
1714			cond_resched();
1715		}
1716	}
1717
1718	/*
1719	 * We won't be working on online/offline memory blocks from this point,
1720	 * so we can't race with memory onlining/offlining. Drop the mutex.
1721	 */
1722	mutex_unlock(&vm->hotplug_mutex);
1723
1724	/* Try to plug and add unused blocks */
1725	virtio_mem_sbm_for_each_mb(vm, mb_id, VIRTIO_MEM_SBM_MB_UNUSED) {
1726		if (!virtio_mem_could_add_memory(vm, memory_block_size_bytes()))
1727			return -ENOSPC;
1728
1729		rc = virtio_mem_sbm_plug_and_add_mb(vm, mb_id, &nb_sb);
1730		if (rc || !nb_sb)
1731			return rc;
1732		cond_resched();
1733	}
1734
1735	/* Try to prepare, plug and add new blocks */
1736	while (nb_sb) {
1737		if (!virtio_mem_could_add_memory(vm, memory_block_size_bytes()))
1738			return -ENOSPC;
1739
1740		rc = virtio_mem_sbm_prepare_next_mb(vm, &mb_id);
1741		if (rc)
1742			return rc;
1743		rc = virtio_mem_sbm_plug_and_add_mb(vm, mb_id, &nb_sb);
1744		if (rc)
1745			return rc;
1746		cond_resched();
1747	}
1748
1749	return 0;
1750out_unlock:
1751	mutex_unlock(&vm->hotplug_mutex);
1752	return rc;
1753}
1754
1755/*
1756 * Plug a big block and add it to Linux.
1757 *
1758 * Will modify the state of the big block.
1759 */
1760static int virtio_mem_bbm_plug_and_add_bb(struct virtio_mem *vm,
1761					  unsigned long bb_id)
1762{
1763	int rc;
1764
1765	if (WARN_ON_ONCE(virtio_mem_bbm_get_bb_state(vm, bb_id) !=
1766			 VIRTIO_MEM_BBM_BB_UNUSED))
1767		return -EINVAL;
1768
1769	rc = virtio_mem_bbm_plug_bb(vm, bb_id);
1770	if (rc)
1771		return rc;
1772	virtio_mem_bbm_set_bb_state(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED);
1773
1774	rc = virtio_mem_bbm_add_bb(vm, bb_id);
1775	if (rc) {
1776		if (!virtio_mem_bbm_unplug_bb(vm, bb_id))
1777			virtio_mem_bbm_set_bb_state(vm, bb_id,
1778						    VIRTIO_MEM_BBM_BB_UNUSED);
1779		else
1780			/* Retry from the main loop. */
1781			virtio_mem_bbm_set_bb_state(vm, bb_id,
1782						    VIRTIO_MEM_BBM_BB_PLUGGED);
1783		return rc;
1784	}
1785	return 0;
1786}
1787
1788/*
1789 * Prepare tracking data for the next big block.
1790 */
1791static int virtio_mem_bbm_prepare_next_bb(struct virtio_mem *vm,
1792					  unsigned long *bb_id)
1793{
1794	int rc;
1795
1796	if (vm->bbm.next_bb_id > vm->bbm.last_usable_bb_id)
1797		return -ENOSPC;
1798
1799	/* Resize the big block state array if required. */
1800	rc = virtio_mem_bbm_bb_states_prepare_next_bb(vm);
1801	if (rc)
1802		return rc;
1803
1804	vm->bbm.bb_count[VIRTIO_MEM_BBM_BB_UNUSED]++;
1805	*bb_id = vm->bbm.next_bb_id;
1806	vm->bbm.next_bb_id++;
1807	return 0;
1808}
1809
1810static int virtio_mem_bbm_plug_request(struct virtio_mem *vm, uint64_t diff)
1811{
1812	uint64_t nb_bb = diff / vm->bbm.bb_size;
1813	unsigned long bb_id;
1814	int rc;
1815
1816	if (!nb_bb)
1817		return 0;
1818
1819	/* Try to plug and add unused big blocks */
1820	virtio_mem_bbm_for_each_bb(vm, bb_id, VIRTIO_MEM_BBM_BB_UNUSED) {
1821		if (!virtio_mem_could_add_memory(vm, vm->bbm.bb_size))
1822			return -ENOSPC;
1823
1824		rc = virtio_mem_bbm_plug_and_add_bb(vm, bb_id);
1825		if (!rc)
1826			nb_bb--;
1827		if (rc || !nb_bb)
1828			return rc;
1829		cond_resched();
1830	}
1831
1832	/* Try to prepare, plug and add new big blocks */
1833	while (nb_bb) {
1834		if (!virtio_mem_could_add_memory(vm, vm->bbm.bb_size))
1835			return -ENOSPC;
1836
1837		rc = virtio_mem_bbm_prepare_next_bb(vm, &bb_id);
1838		if (rc)
1839			return rc;
1840		rc = virtio_mem_bbm_plug_and_add_bb(vm, bb_id);
1841		if (!rc)
1842			nb_bb--;
1843		if (rc)
1844			return rc;
1845		cond_resched();
1846	}
1847
1848	return 0;
1849}
1850
1851/*
1852 * Try to plug the requested amount of memory.
1853 */
1854static int virtio_mem_plug_request(struct virtio_mem *vm, uint64_t diff)
1855{
1856	if (vm->in_sbm)
1857		return virtio_mem_sbm_plug_request(vm, diff);
1858	return virtio_mem_bbm_plug_request(vm, diff);
1859}
1860
1861/*
1862 * Unplug the desired number of plugged subblocks of an offline memory block.
1863 * Will fail if any subblock cannot get unplugged (instead of skipping it).
1864 *
1865 * Will modify the state of the memory block. Might temporarily drop the
1866 * hotplug_mutex.
1867 *
1868 * Note: Can fail after some subblocks were successfully unplugged.
1869 */
1870static int virtio_mem_sbm_unplug_any_sb_offline(struct virtio_mem *vm,
1871						unsigned long mb_id,
1872						uint64_t *nb_sb)
1873{
1874	int rc;
1875
1876	rc = virtio_mem_sbm_unplug_any_sb_raw(vm, mb_id, nb_sb);
1877
1878	/* some subblocks might have been unplugged even on failure */
1879	if (!virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb))
1880		virtio_mem_sbm_set_mb_state(vm, mb_id,
1881					    VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL);
1882	if (rc)
1883		return rc;
1884
1885	if (virtio_mem_sbm_test_sb_unplugged(vm, mb_id, 0, vm->sbm.sbs_per_mb)) {
1886		/*
1887		 * Remove the block from Linux - this should never fail.
1888		 * Hinder the block from getting onlined by marking it
1889		 * unplugged. Temporarily drop the mutex, so
1890		 * any pending GOING_ONLINE requests can be serviced/rejected.
1891		 */
1892		virtio_mem_sbm_set_mb_state(vm, mb_id,
1893					    VIRTIO_MEM_SBM_MB_UNUSED);
1894
1895		mutex_unlock(&vm->hotplug_mutex);
1896		rc = virtio_mem_sbm_remove_mb(vm, mb_id);
1897		BUG_ON(rc);
1898		mutex_lock(&vm->hotplug_mutex);
1899	}
1900	return 0;
1901}
1902
1903/*
1904 * Unplug the given plugged subblocks of an online memory block.
1905 *
1906 * Will modify the state of the memory block.
1907 */
1908static int virtio_mem_sbm_unplug_sb_online(struct virtio_mem *vm,
1909					   unsigned long mb_id, int sb_id,
1910					   int count)
1911{
1912	const unsigned long nr_pages = PFN_DOWN(vm->sbm.sb_size) * count;
1913	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
1914	unsigned long start_pfn;
1915	int rc;
1916
1917	start_pfn = PFN_DOWN(virtio_mem_mb_id_to_phys(mb_id) +
1918			     sb_id * vm->sbm.sb_size);
1919
1920	rc = virtio_mem_fake_offline(start_pfn, nr_pages);
1921	if (rc)
1922		return rc;
1923
1924	/* Try to unplug the allocated memory */
1925	rc = virtio_mem_sbm_unplug_sb(vm, mb_id, sb_id, count);
1926	if (rc) {
1927		/* Return the memory to the buddy. */
1928		virtio_mem_fake_online(start_pfn, nr_pages);
1929		return rc;
1930	}
1931
1932	switch (old_state) {
1933	case VIRTIO_MEM_SBM_MB_KERNEL:
1934		virtio_mem_sbm_set_mb_state(vm, mb_id,
1935					    VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL);
1936		break;
1937	case VIRTIO_MEM_SBM_MB_MOVABLE:
1938		virtio_mem_sbm_set_mb_state(vm, mb_id,
1939					    VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL);
1940		break;
1941	}
1942
1943	return 0;
1944}
1945
1946/*
1947 * Unplug the desired number of plugged subblocks of an online memory block.
1948 * Will skip subblock that are busy.
1949 *
1950 * Will modify the state of the memory block. Might temporarily drop the
1951 * hotplug_mutex.
1952 *
1953 * Note: Can fail after some subblocks were successfully unplugged. Can
1954 *       return 0 even if subblocks were busy and could not get unplugged.
1955 */
1956static int virtio_mem_sbm_unplug_any_sb_online(struct virtio_mem *vm,
1957					       unsigned long mb_id,
1958					       uint64_t *nb_sb)
1959{
1960	int rc, sb_id;
1961
1962	/* If possible, try to unplug the complete block in one shot. */
1963	if (*nb_sb >= vm->sbm.sbs_per_mb &&
1964	    virtio_mem_sbm_test_sb_plugged(vm, mb_id, 0, vm->sbm.sbs_per_mb)) {
1965		rc = virtio_mem_sbm_unplug_sb_online(vm, mb_id, 0,
1966						     vm->sbm.sbs_per_mb);
1967		if (!rc) {
1968			*nb_sb -= vm->sbm.sbs_per_mb;
1969			goto unplugged;
1970		} else if (rc != -EBUSY)
1971			return rc;
1972	}
1973
1974	/* Fallback to single subblocks. */
1975	for (sb_id = vm->sbm.sbs_per_mb - 1; sb_id >= 0 && *nb_sb; sb_id--) {
1976		/* Find the next candidate subblock */
1977		while (sb_id >= 0 &&
1978		       !virtio_mem_sbm_test_sb_plugged(vm, mb_id, sb_id, 1))
1979			sb_id--;
1980		if (sb_id < 0)
1981			break;
1982
1983		rc = virtio_mem_sbm_unplug_sb_online(vm, mb_id, sb_id, 1);
1984		if (rc == -EBUSY)
1985			continue;
1986		else if (rc)
1987			return rc;
1988		*nb_sb -= 1;
1989	}
1990
1991unplugged:
1992	/*
1993	 * Once all subblocks of a memory block were unplugged, offline and
1994	 * remove it. This will usually not fail, as no memory is in use
1995	 * anymore - however some other notifiers might NACK the request.
1996	 */
1997	if (virtio_mem_sbm_test_sb_unplugged(vm, mb_id, 0, vm->sbm.sbs_per_mb)) {
1998		mutex_unlock(&vm->hotplug_mutex);
1999		rc = virtio_mem_sbm_offline_and_remove_mb(vm, mb_id);
2000		mutex_lock(&vm->hotplug_mutex);
2001		if (!rc)
2002			virtio_mem_sbm_set_mb_state(vm, mb_id,
2003						    VIRTIO_MEM_SBM_MB_UNUSED);
2004	}
2005
2006	return 0;
2007}
2008
2009/*
2010 * Unplug the desired number of plugged subblocks of a memory block that is
2011 * already added to Linux. Will skip subblock of online memory blocks that are
2012 * busy (by the OS). Will fail if any subblock that's not busy cannot get
2013 * unplugged.
2014 *
2015 * Will modify the state of the memory block. Might temporarily drop the
2016 * hotplug_mutex.
2017 *
2018 * Note: Can fail after some subblocks were successfully unplugged. Can
2019 *       return 0 even if subblocks were busy and could not get unplugged.
2020 */
2021static int virtio_mem_sbm_unplug_any_sb(struct virtio_mem *vm,
2022					unsigned long mb_id,
2023					uint64_t *nb_sb)
2024{
2025	const int old_state = virtio_mem_sbm_get_mb_state(vm, mb_id);
2026
2027	switch (old_state) {
2028	case VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL:
2029	case VIRTIO_MEM_SBM_MB_KERNEL:
2030	case VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL:
2031	case VIRTIO_MEM_SBM_MB_MOVABLE:
2032		return virtio_mem_sbm_unplug_any_sb_online(vm, mb_id, nb_sb);
2033	case VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL:
2034	case VIRTIO_MEM_SBM_MB_OFFLINE:
2035		return virtio_mem_sbm_unplug_any_sb_offline(vm, mb_id, nb_sb);
2036	}
2037	return -EINVAL;
2038}
2039
2040static int virtio_mem_sbm_unplug_request(struct virtio_mem *vm, uint64_t diff)
2041{
2042	const int mb_states[] = {
2043		VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL,
2044		VIRTIO_MEM_SBM_MB_OFFLINE,
2045		VIRTIO_MEM_SBM_MB_MOVABLE_PARTIAL,
2046		VIRTIO_MEM_SBM_MB_KERNEL_PARTIAL,
2047		VIRTIO_MEM_SBM_MB_MOVABLE,
2048		VIRTIO_MEM_SBM_MB_KERNEL,
2049	};
2050	uint64_t nb_sb = diff / vm->sbm.sb_size;
2051	unsigned long mb_id;
2052	int rc, i;
2053
2054	if (!nb_sb)
2055		return 0;
2056
2057	/*
2058	 * We'll drop the mutex a couple of times when it is safe to do so.
2059	 * This might result in some blocks switching the state (online/offline)
2060	 * and we could miss them in this run - we will retry again later.
2061	 */
2062	mutex_lock(&vm->hotplug_mutex);
2063
2064	/*
2065	 * We try unplug from partially plugged blocks first, to try removing
2066	 * whole memory blocks along with metadata. We prioritize ZONE_MOVABLE
2067	 * as it's more reliable to unplug memory and remove whole memory
2068	 * blocks, and we don't want to trigger a zone imbalances by
2069	 * accidentially removing too much kernel memory.
2070	 */
2071	for (i = 0; i < ARRAY_SIZE(mb_states); i++) {
2072		virtio_mem_sbm_for_each_mb_rev(vm, mb_id, mb_states[i]) {
2073			rc = virtio_mem_sbm_unplug_any_sb(vm, mb_id, &nb_sb);
2074			if (rc || !nb_sb)
2075				goto out_unlock;
2076			mutex_unlock(&vm->hotplug_mutex);
2077			cond_resched();
2078			mutex_lock(&vm->hotplug_mutex);
2079		}
2080		if (!unplug_online && i == 1) {
2081			mutex_unlock(&vm->hotplug_mutex);
2082			return 0;
2083		}
2084	}
2085
2086	mutex_unlock(&vm->hotplug_mutex);
2087	return nb_sb ? -EBUSY : 0;
2088out_unlock:
2089	mutex_unlock(&vm->hotplug_mutex);
2090	return rc;
2091}
2092
2093/*
2094 * Try to offline and remove a big block from Linux and unplug it. Will fail
2095 * with -EBUSY if some memory is busy and cannot get unplugged.
2096 *
2097 * Will modify the state of the memory block. Might temporarily drop the
2098 * hotplug_mutex.
2099 */
2100static int virtio_mem_bbm_offline_remove_and_unplug_bb(struct virtio_mem *vm,
2101						       unsigned long bb_id)
2102{
2103	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2104	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2105	unsigned long end_pfn = start_pfn + nr_pages;
2106	unsigned long pfn;
2107	struct page *page;
2108	int rc;
2109
2110	if (WARN_ON_ONCE(virtio_mem_bbm_get_bb_state(vm, bb_id) !=
2111			 VIRTIO_MEM_BBM_BB_ADDED))
2112		return -EINVAL;
2113
2114	if (bbm_safe_unplug) {
2115		/*
2116		 * Start by fake-offlining all memory. Once we marked the device
2117		 * block as fake-offline, all newly onlined memory will
2118		 * automatically be kept fake-offline. Protect from concurrent
2119		 * onlining/offlining until we have a consistent state.
2120		 */
2121		mutex_lock(&vm->hotplug_mutex);
2122		virtio_mem_bbm_set_bb_state(vm, bb_id,
2123					    VIRTIO_MEM_BBM_BB_FAKE_OFFLINE);
2124
2125		for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
2126			page = pfn_to_online_page(pfn);
2127			if (!page)
2128				continue;
2129
2130			rc = virtio_mem_fake_offline(pfn, PAGES_PER_SECTION);
2131			if (rc) {
2132				end_pfn = pfn;
2133				goto rollback_safe_unplug;
2134			}
2135		}
2136		mutex_unlock(&vm->hotplug_mutex);
2137	}
 
2138
2139	rc = virtio_mem_bbm_offline_and_remove_bb(vm, bb_id);
2140	if (rc) {
2141		if (bbm_safe_unplug) {
2142			mutex_lock(&vm->hotplug_mutex);
2143			goto rollback_safe_unplug;
2144		}
2145		return rc;
2146	}
2147
2148	rc = virtio_mem_bbm_unplug_bb(vm, bb_id);
2149	if (rc)
2150		virtio_mem_bbm_set_bb_state(vm, bb_id,
2151					    VIRTIO_MEM_BBM_BB_PLUGGED);
2152	else
2153		virtio_mem_bbm_set_bb_state(vm, bb_id,
2154					    VIRTIO_MEM_BBM_BB_UNUSED);
2155	return rc;
2156
2157rollback_safe_unplug:
2158	for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
2159		page = pfn_to_online_page(pfn);
2160		if (!page)
2161			continue;
2162		virtio_mem_fake_online(pfn, PAGES_PER_SECTION);
2163	}
2164	virtio_mem_bbm_set_bb_state(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED);
2165	mutex_unlock(&vm->hotplug_mutex);
2166	return rc;
2167}
2168
2169/*
2170 * Test if a big block is completely offline.
2171 */
2172static bool virtio_mem_bbm_bb_is_offline(struct virtio_mem *vm,
2173					 unsigned long bb_id)
2174{
2175	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2176	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2177	unsigned long pfn;
2178
2179	for (pfn = start_pfn; pfn < start_pfn + nr_pages;
2180	     pfn += PAGES_PER_SECTION) {
2181		if (pfn_to_online_page(pfn))
2182			return false;
2183	}
2184
2185	return true;
2186}
2187
2188/*
2189 * Test if a big block is completely onlined to ZONE_MOVABLE (or offline).
2190 */
2191static bool virtio_mem_bbm_bb_is_movable(struct virtio_mem *vm,
2192					 unsigned long bb_id)
2193{
2194	const unsigned long start_pfn = PFN_DOWN(virtio_mem_bb_id_to_phys(vm, bb_id));
2195	const unsigned long nr_pages = PFN_DOWN(vm->bbm.bb_size);
2196	struct page *page;
2197	unsigned long pfn;
2198
2199	for (pfn = start_pfn; pfn < start_pfn + nr_pages;
2200	     pfn += PAGES_PER_SECTION) {
2201		page = pfn_to_online_page(pfn);
2202		if (!page)
2203			continue;
2204		if (page_zonenum(page) != ZONE_MOVABLE)
2205			return false;
2206	}
2207
2208	return true;
2209}
2210
2211static int virtio_mem_bbm_unplug_request(struct virtio_mem *vm, uint64_t diff)
2212{
2213	uint64_t nb_bb = diff / vm->bbm.bb_size;
2214	uint64_t bb_id;
2215	int rc, i;
2216
2217	if (!nb_bb)
2218		return 0;
2219
2220	/*
2221	 * Try to unplug big blocks. Similar to SBM, start with offline
2222	 * big blocks.
2223	 */
2224	for (i = 0; i < 3; i++) {
2225		virtio_mem_bbm_for_each_bb_rev(vm, bb_id, VIRTIO_MEM_BBM_BB_ADDED) {
2226			cond_resched();
2227
2228			/*
2229			 * As we're holding no locks, these checks are racy,
2230			 * but we don't care.
2231			 */
2232			if (i == 0 && !virtio_mem_bbm_bb_is_offline(vm, bb_id))
2233				continue;
2234			if (i == 1 && !virtio_mem_bbm_bb_is_movable(vm, bb_id))
2235				continue;
2236			rc = virtio_mem_bbm_offline_remove_and_unplug_bb(vm, bb_id);
2237			if (rc == -EBUSY)
2238				continue;
2239			if (!rc)
2240				nb_bb--;
2241			if (rc || !nb_bb)
2242				return rc;
2243		}
2244		if (i == 0 && !unplug_online)
2245			return 0;
2246	}
2247
2248	return nb_bb ? -EBUSY : 0;
2249}
2250
2251/*
2252 * Try to unplug the requested amount of memory.
2253 */
2254static int virtio_mem_unplug_request(struct virtio_mem *vm, uint64_t diff)
2255{
2256	if (vm->in_sbm)
2257		return virtio_mem_sbm_unplug_request(vm, diff);
2258	return virtio_mem_bbm_unplug_request(vm, diff);
2259}
2260
2261/*
2262 * Try to unplug all blocks that couldn't be unplugged before, for example,
2263 * because the hypervisor was busy.
 
2264 */
2265static int virtio_mem_unplug_pending_mb(struct virtio_mem *vm)
2266{
2267	unsigned long id;
2268	int rc;
2269
2270	if (!vm->in_sbm) {
2271		virtio_mem_bbm_for_each_bb(vm, id,
2272					   VIRTIO_MEM_BBM_BB_PLUGGED) {
2273			rc = virtio_mem_bbm_unplug_bb(vm, id);
2274			if (rc)
2275				return rc;
2276			virtio_mem_bbm_set_bb_state(vm, id,
2277						    VIRTIO_MEM_BBM_BB_UNUSED);
2278		}
2279		return 0;
2280	}
2281
2282	virtio_mem_sbm_for_each_mb(vm, id, VIRTIO_MEM_SBM_MB_PLUGGED) {
2283		rc = virtio_mem_sbm_unplug_mb(vm, id);
2284		if (rc)
2285			return rc;
2286		virtio_mem_sbm_set_mb_state(vm, id,
2287					    VIRTIO_MEM_SBM_MB_UNUSED);
2288	}
2289
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2290	return 0;
2291}
2292
2293/*
2294 * Update all parts of the config that could have changed.
2295 */
2296static void virtio_mem_refresh_config(struct virtio_mem *vm)
2297{
2298	const struct range pluggable_range = mhp_get_pluggable_range(true);
2299	uint64_t new_plugged_size, usable_region_size, end_addr;
2300
2301	/* the plugged_size is just a reflection of what _we_ did previously */
2302	virtio_cread_le(vm->vdev, struct virtio_mem_config, plugged_size,
2303			&new_plugged_size);
2304	if (WARN_ON_ONCE(new_plugged_size != vm->plugged_size))
2305		vm->plugged_size = new_plugged_size;
2306
2307	/* calculate the last usable memory block id */
2308	virtio_cread_le(vm->vdev, struct virtio_mem_config,
2309			usable_region_size, &usable_region_size);
2310	end_addr = min(vm->addr + usable_region_size - 1,
2311		       pluggable_range.end);
2312
2313	if (vm->in_sbm) {
2314		vm->sbm.last_usable_mb_id = virtio_mem_phys_to_mb_id(end_addr);
2315		if (!IS_ALIGNED(end_addr + 1, memory_block_size_bytes()))
2316			vm->sbm.last_usable_mb_id--;
2317	} else {
2318		vm->bbm.last_usable_bb_id = virtio_mem_phys_to_bb_id(vm,
2319								     end_addr);
2320		if (!IS_ALIGNED(end_addr + 1, vm->bbm.bb_size))
2321			vm->bbm.last_usable_bb_id--;
2322	}
2323	/*
2324	 * If we cannot plug any of our device memory (e.g., nothing in the
2325	 * usable region is addressable), the last usable memory block id will
2326	 * be smaller than the first usable memory block id. We'll stop
2327	 * attempting to add memory with -ENOSPC from our main loop.
2328	 */
2329
2330	/* see if there is a request to change the size */
2331	virtio_cread_le(vm->vdev, struct virtio_mem_config, requested_size,
2332			&vm->requested_size);
2333
2334	dev_info(&vm->vdev->dev, "plugged size: 0x%llx", vm->plugged_size);
2335	dev_info(&vm->vdev->dev, "requested size: 0x%llx", vm->requested_size);
2336}
2337
2338/*
2339 * Workqueue function for handling plug/unplug requests and config updates.
2340 */
2341static void virtio_mem_run_wq(struct work_struct *work)
2342{
2343	struct virtio_mem *vm = container_of(work, struct virtio_mem, wq);
2344	uint64_t diff;
2345	int rc;
2346
2347	if (unlikely(vm->in_kdump)) {
2348		dev_warn_once(&vm->vdev->dev,
2349			     "unexpected workqueue run in kdump kernel\n");
2350		return;
2351	}
2352
2353	hrtimer_cancel(&vm->retry_timer);
2354
2355	if (vm->broken)
2356		return;
2357
2358	atomic_set(&vm->wq_active, 1);
2359retry:
2360	rc = 0;
2361
2362	/* Make sure we start with a clean state if there are leftovers. */
2363	if (unlikely(vm->unplug_all_required))
2364		rc = virtio_mem_send_unplug_all_request(vm);
2365
2366	if (atomic_read(&vm->config_changed)) {
2367		atomic_set(&vm->config_changed, 0);
2368		virtio_mem_refresh_config(vm);
2369	}
2370
2371	/* Unplug any leftovers from previous runs */
2372	if (!rc)
2373		rc = virtio_mem_unplug_pending_mb(vm);
2374
2375	if (!rc && vm->requested_size != vm->plugged_size) {
2376		if (vm->requested_size > vm->plugged_size) {
2377			diff = vm->requested_size - vm->plugged_size;
2378			rc = virtio_mem_plug_request(vm, diff);
2379		} else {
2380			diff = vm->plugged_size - vm->requested_size;
2381			rc = virtio_mem_unplug_request(vm, diff);
2382		}
2383	}
 
 
 
 
 
 
 
2384
2385	switch (rc) {
2386	case 0:
2387		vm->retry_timer_ms = VIRTIO_MEM_RETRY_TIMER_MIN_MS;
2388		break;
2389	case -ENOSPC:
2390		/*
2391		 * We cannot add any more memory (alignment, physical limit)
2392		 * or we have too many offline memory blocks.
2393		 */
2394		break;
2395	case -ETXTBSY:
2396		/*
2397		 * The hypervisor cannot process our request right now
2398		 * (e.g., out of memory, migrating);
2399		 */
2400	case -EBUSY:
2401		/*
2402		 * We cannot free up any memory to unplug it (all plugged memory
2403		 * is busy).
2404		 */
2405	case -ENOMEM:
2406		/* Out of memory, try again later. */
2407		hrtimer_start(&vm->retry_timer, ms_to_ktime(vm->retry_timer_ms),
2408			      HRTIMER_MODE_REL);
2409		break;
2410	case -EAGAIN:
2411		/* Retry immediately (e.g., the config changed). */
2412		goto retry;
2413	default:
2414		/* Unknown error, mark as broken */
2415		dev_err(&vm->vdev->dev,
2416			"unknown error, marking device broken: %d\n", rc);
2417		vm->broken = true;
2418	}
2419
2420	atomic_set(&vm->wq_active, 0);
2421}
2422
2423static enum hrtimer_restart virtio_mem_timer_expired(struct hrtimer *timer)
2424{
2425	struct virtio_mem *vm = container_of(timer, struct virtio_mem,
2426					     retry_timer);
2427
2428	virtio_mem_retry(vm);
2429	vm->retry_timer_ms = min_t(unsigned int, vm->retry_timer_ms * 2,
2430				   VIRTIO_MEM_RETRY_TIMER_MAX_MS);
2431	return HRTIMER_NORESTART;
2432}
2433
2434static void virtio_mem_handle_response(struct virtqueue *vq)
2435{
2436	struct virtio_mem *vm = vq->vdev->priv;
2437
2438	wake_up(&vm->host_resp);
2439}
2440
2441static int virtio_mem_init_vq(struct virtio_mem *vm)
2442{
2443	struct virtqueue *vq;
2444
2445	vq = virtio_find_single_vq(vm->vdev, virtio_mem_handle_response,
2446				   "guest-request");
2447	if (IS_ERR(vq))
2448		return PTR_ERR(vq);
2449	vm->vq = vq;
2450
2451	return 0;
2452}
2453
2454static int virtio_mem_init_hotplug(struct virtio_mem *vm)
2455{
2456	const struct range pluggable_range = mhp_get_pluggable_range(true);
2457	uint64_t unit_pages, sb_size, addr;
2458	int rc;
2459
2460	/* bad device setup - warn only */
2461	if (!IS_ALIGNED(vm->addr, memory_block_size_bytes()))
2462		dev_warn(&vm->vdev->dev,
2463			 "The alignment of the physical start address can make some memory unusable.\n");
2464	if (!IS_ALIGNED(vm->addr + vm->region_size, memory_block_size_bytes()))
2465		dev_warn(&vm->vdev->dev,
2466			 "The alignment of the physical end address can make some memory unusable.\n");
2467	if (vm->addr < pluggable_range.start ||
2468	    vm->addr + vm->region_size - 1 > pluggable_range.end)
2469		dev_warn(&vm->vdev->dev,
2470			 "Some device memory is not addressable/pluggable. This can make some memory unusable.\n");
2471
2472	/* Prepare the offline threshold - make sure we can add two blocks. */
2473	vm->offline_threshold = max_t(uint64_t, 2 * memory_block_size_bytes(),
2474				      VIRTIO_MEM_DEFAULT_OFFLINE_THRESHOLD);
2475
2476	/*
2477	 * alloc_contig_range() works reliably with pageblock
2478	 * granularity on ZONE_NORMAL, use pageblock_nr_pages.
2479	 */
2480	sb_size = PAGE_SIZE * pageblock_nr_pages;
2481	sb_size = max_t(uint64_t, vm->device_block_size, sb_size);
2482
2483	if (sb_size < memory_block_size_bytes() && !force_bbm) {
2484		/* SBM: At least two subblocks per Linux memory block. */
2485		vm->in_sbm = true;
2486		vm->sbm.sb_size = sb_size;
2487		vm->sbm.sbs_per_mb = memory_block_size_bytes() /
2488				     vm->sbm.sb_size;
2489
2490		/* Round up to the next full memory block */
2491		addr = max_t(uint64_t, vm->addr, pluggable_range.start) +
2492		       memory_block_size_bytes() - 1;
2493		vm->sbm.first_mb_id = virtio_mem_phys_to_mb_id(addr);
2494		vm->sbm.next_mb_id = vm->sbm.first_mb_id;
2495	} else {
2496		/* BBM: At least one Linux memory block. */
2497		vm->bbm.bb_size = max_t(uint64_t, vm->device_block_size,
2498					memory_block_size_bytes());
2499
2500		if (bbm_block_size) {
2501			if (!is_power_of_2(bbm_block_size)) {
2502				dev_warn(&vm->vdev->dev,
2503					 "bbm_block_size is not a power of 2");
2504			} else if (bbm_block_size < vm->bbm.bb_size) {
2505				dev_warn(&vm->vdev->dev,
2506					 "bbm_block_size is too small");
2507			} else {
2508				vm->bbm.bb_size = bbm_block_size;
2509			}
2510		}
2511
2512		/* Round up to the next aligned big block */
2513		addr = max_t(uint64_t, vm->addr, pluggable_range.start) +
2514		       vm->bbm.bb_size - 1;
2515		vm->bbm.first_bb_id = virtio_mem_phys_to_bb_id(vm, addr);
2516		vm->bbm.next_bb_id = vm->bbm.first_bb_id;
2517
2518		/* Make sure we can add two big blocks. */
2519		vm->offline_threshold = max_t(uint64_t, 2 * vm->bbm.bb_size,
2520					      vm->offline_threshold);
2521	}
2522
2523	dev_info(&vm->vdev->dev, "memory block size: 0x%lx",
2524		 memory_block_size_bytes());
2525	if (vm->in_sbm)
2526		dev_info(&vm->vdev->dev, "subblock size: 0x%llx",
2527			 (unsigned long long)vm->sbm.sb_size);
2528	else
2529		dev_info(&vm->vdev->dev, "big block size: 0x%llx",
2530			 (unsigned long long)vm->bbm.bb_size);
2531
2532	/* create the parent resource for all memory */
2533	rc = virtio_mem_create_resource(vm);
2534	if (rc)
2535		return rc;
2536
2537	/* use a single dynamic memory group to cover the whole memory device */
2538	if (vm->in_sbm)
2539		unit_pages = PHYS_PFN(memory_block_size_bytes());
2540	else
2541		unit_pages = PHYS_PFN(vm->bbm.bb_size);
2542	rc = memory_group_register_dynamic(vm->nid, unit_pages);
2543	if (rc < 0)
2544		goto out_del_resource;
2545	vm->mgid = rc;
2546
2547	/*
2548	 * If we still have memory plugged, we have to unplug all memory first.
2549	 * Registering our parent resource makes sure that this memory isn't
2550	 * actually in use (e.g., trying to reload the driver).
2551	 */
2552	if (vm->plugged_size) {
2553		vm->unplug_all_required = true;
2554		dev_info(&vm->vdev->dev, "unplugging all memory is required\n");
2555	}
2556
2557	/* register callbacks */
2558	vm->memory_notifier.notifier_call = virtio_mem_memory_notifier_cb;
2559	rc = register_memory_notifier(&vm->memory_notifier);
2560	if (rc)
2561		goto out_unreg_group;
2562	rc = register_virtio_mem_device(vm);
2563	if (rc)
2564		goto out_unreg_mem;
2565
2566	return 0;
2567out_unreg_mem:
2568	unregister_memory_notifier(&vm->memory_notifier);
2569out_unreg_group:
2570	memory_group_unregister(vm->mgid);
2571out_del_resource:
2572	virtio_mem_delete_resource(vm);
2573	return rc;
2574}
2575
2576#ifdef CONFIG_PROC_VMCORE
2577static int virtio_mem_send_state_request(struct virtio_mem *vm, uint64_t addr,
2578					 uint64_t size)
2579{
2580	const uint64_t nb_vm_blocks = size / vm->device_block_size;
2581	const struct virtio_mem_req req = {
2582		.type = cpu_to_virtio16(vm->vdev, VIRTIO_MEM_REQ_STATE),
2583		.u.state.addr = cpu_to_virtio64(vm->vdev, addr),
2584		.u.state.nb_blocks = cpu_to_virtio16(vm->vdev, nb_vm_blocks),
2585	};
2586	int rc = -ENOMEM;
2587
2588	dev_dbg(&vm->vdev->dev, "requesting state: 0x%llx - 0x%llx\n", addr,
2589		addr + size - 1);
2590
2591	switch (virtio_mem_send_request(vm, &req)) {
2592	case VIRTIO_MEM_RESP_ACK:
2593		return virtio16_to_cpu(vm->vdev, vm->resp.u.state.state);
2594	case VIRTIO_MEM_RESP_ERROR:
2595		rc = -EINVAL;
2596		break;
2597	default:
2598		break;
2599	}
2600
2601	dev_dbg(&vm->vdev->dev, "requesting state failed: %d\n", rc);
2602	return rc;
2603}
2604
2605static bool virtio_mem_vmcore_pfn_is_ram(struct vmcore_cb *cb,
2606					 unsigned long pfn)
2607{
2608	struct virtio_mem *vm = container_of(cb, struct virtio_mem,
2609					     vmcore_cb);
2610	uint64_t addr = PFN_PHYS(pfn);
2611	bool is_ram;
2612	int rc;
2613
2614	if (!virtio_mem_contains_range(vm, addr, PAGE_SIZE))
2615		return true;
2616	if (!vm->plugged_size)
2617		return false;
2618
2619	/*
2620	 * We have to serialize device requests and access to the information
2621	 * about the block queried last.
2622	 */
2623	mutex_lock(&vm->hotplug_mutex);
2624
2625	addr = ALIGN_DOWN(addr, vm->device_block_size);
2626	if (addr != vm->last_block_addr) {
2627		rc = virtio_mem_send_state_request(vm, addr,
2628						   vm->device_block_size);
2629		/* On any kind of error, we're going to signal !ram. */
2630		if (rc == VIRTIO_MEM_STATE_PLUGGED)
2631			vm->last_block_plugged = true;
2632		else
2633			vm->last_block_plugged = false;
2634		vm->last_block_addr = addr;
2635	}
2636
2637	is_ram = vm->last_block_plugged;
2638	mutex_unlock(&vm->hotplug_mutex);
2639	return is_ram;
2640}
2641#endif /* CONFIG_PROC_VMCORE */
2642
2643static int virtio_mem_init_kdump(struct virtio_mem *vm)
2644{
2645#ifdef CONFIG_PROC_VMCORE
2646	dev_info(&vm->vdev->dev, "memory hot(un)plug disabled in kdump kernel\n");
2647	vm->vmcore_cb.pfn_is_ram = virtio_mem_vmcore_pfn_is_ram;
2648	register_vmcore_cb(&vm->vmcore_cb);
2649	return 0;
2650#else /* CONFIG_PROC_VMCORE */
2651	dev_warn(&vm->vdev->dev, "disabled in kdump kernel without vmcore\n");
2652	return -EBUSY;
2653#endif /* CONFIG_PROC_VMCORE */
2654}
2655
2656static int virtio_mem_init(struct virtio_mem *vm)
2657{
2658	uint16_t node_id;
2659
2660	if (!vm->vdev->config->get) {
2661		dev_err(&vm->vdev->dev, "config access disabled\n");
2662		return -EINVAL;
2663	}
2664
2665	/* Fetch all properties that can't change. */
2666	virtio_cread_le(vm->vdev, struct virtio_mem_config, plugged_size,
2667			&vm->plugged_size);
2668	virtio_cread_le(vm->vdev, struct virtio_mem_config, block_size,
2669			&vm->device_block_size);
2670	virtio_cread_le(vm->vdev, struct virtio_mem_config, node_id,
2671			&node_id);
2672	vm->nid = virtio_mem_translate_node_id(vm, node_id);
2673	virtio_cread_le(vm->vdev, struct virtio_mem_config, addr, &vm->addr);
2674	virtio_cread_le(vm->vdev, struct virtio_mem_config, region_size,
2675			&vm->region_size);
2676
2677	/* Determine the nid for the device based on the lowest address. */
2678	if (vm->nid == NUMA_NO_NODE)
2679		vm->nid = memory_add_physaddr_to_nid(vm->addr);
2680
2681	dev_info(&vm->vdev->dev, "start address: 0x%llx", vm->addr);
2682	dev_info(&vm->vdev->dev, "region size: 0x%llx", vm->region_size);
2683	dev_info(&vm->vdev->dev, "device block size: 0x%llx",
2684		 (unsigned long long)vm->device_block_size);
2685	if (vm->nid != NUMA_NO_NODE && IS_ENABLED(CONFIG_NUMA))
2686		dev_info(&vm->vdev->dev, "nid: %d", vm->nid);
2687
2688	/*
2689	 * We don't want to (un)plug or reuse any memory when in kdump. The
2690	 * memory is still accessible (but not exposed to Linux).
2691	 */
2692	if (vm->in_kdump)
2693		return virtio_mem_init_kdump(vm);
2694	return virtio_mem_init_hotplug(vm);
2695}
2696
2697static int virtio_mem_create_resource(struct virtio_mem *vm)
2698{
2699	/*
2700	 * When force-unloading the driver and removing the device, we
2701	 * could have a garbage pointer. Duplicate the string.
2702	 */
2703	const char *name = kstrdup(dev_name(&vm->vdev->dev), GFP_KERNEL);
2704
2705	if (!name)
2706		return -ENOMEM;
2707
2708	/* Disallow mapping device memory via /dev/mem completely. */
2709	vm->parent_resource = __request_mem_region(vm->addr, vm->region_size,
2710						   name, IORESOURCE_SYSTEM_RAM |
2711						   IORESOURCE_EXCLUSIVE);
2712	if (!vm->parent_resource) {
2713		kfree(name);
2714		dev_warn(&vm->vdev->dev, "could not reserve device region\n");
2715		dev_info(&vm->vdev->dev,
2716			 "reloading the driver is not supported\n");
2717		return -EBUSY;
2718	}
2719
2720	/* The memory is not actually busy - make add_memory() work. */
2721	vm->parent_resource->flags &= ~IORESOURCE_BUSY;
2722	return 0;
2723}
2724
2725static void virtio_mem_delete_resource(struct virtio_mem *vm)
2726{
2727	const char *name;
2728
2729	if (!vm->parent_resource)
2730		return;
2731
2732	name = vm->parent_resource->name;
2733	release_resource(vm->parent_resource);
2734	kfree(vm->parent_resource);
2735	kfree(name);
2736	vm->parent_resource = NULL;
2737}
2738
2739static int virtio_mem_range_has_system_ram(struct resource *res, void *arg)
2740{
2741	return 1;
2742}
2743
2744static bool virtio_mem_has_memory_added(struct virtio_mem *vm)
2745{
2746	const unsigned long flags = IORESOURCE_SYSTEM_RAM | IORESOURCE_BUSY;
2747
2748	return walk_iomem_res_desc(IORES_DESC_NONE, flags, vm->addr,
2749				   vm->addr + vm->region_size, NULL,
2750				   virtio_mem_range_has_system_ram) == 1;
2751}
2752
2753static int virtio_mem_probe(struct virtio_device *vdev)
2754{
2755	struct virtio_mem *vm;
2756	int rc;
2757
2758	BUILD_BUG_ON(sizeof(struct virtio_mem_req) != 24);
2759	BUILD_BUG_ON(sizeof(struct virtio_mem_resp) != 10);
2760
2761	vdev->priv = vm = kzalloc(sizeof(*vm), GFP_KERNEL);
2762	if (!vm)
2763		return -ENOMEM;
2764
2765	init_waitqueue_head(&vm->host_resp);
2766	vm->vdev = vdev;
2767	INIT_WORK(&vm->wq, virtio_mem_run_wq);
2768	mutex_init(&vm->hotplug_mutex);
2769	INIT_LIST_HEAD(&vm->next);
2770	spin_lock_init(&vm->removal_lock);
2771	hrtimer_init(&vm->retry_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
2772	vm->retry_timer.function = virtio_mem_timer_expired;
2773	vm->retry_timer_ms = VIRTIO_MEM_RETRY_TIMER_MIN_MS;
2774	vm->in_kdump = is_kdump_kernel();
2775
2776	/* register the virtqueue */
2777	rc = virtio_mem_init_vq(vm);
2778	if (rc)
2779		goto out_free_vm;
2780
2781	/* initialize the device by querying the config */
2782	rc = virtio_mem_init(vm);
2783	if (rc)
2784		goto out_del_vq;
2785
2786	virtio_device_ready(vdev);
2787
2788	/* trigger a config update to start processing the requested_size */
2789	if (!vm->in_kdump) {
2790		atomic_set(&vm->config_changed, 1);
2791		queue_work(system_freezable_wq, &vm->wq);
2792	}
2793
2794	return 0;
2795out_del_vq:
2796	vdev->config->del_vqs(vdev);
2797out_free_vm:
2798	kfree(vm);
2799	vdev->priv = NULL;
2800
2801	return rc;
2802}
2803
2804static void virtio_mem_deinit_hotplug(struct virtio_mem *vm)
2805{
2806	unsigned long mb_id;
2807	int rc;
2808
2809	/*
2810	 * Make sure the workqueue won't be triggered anymore and no memory
2811	 * blocks can be onlined/offlined until we're finished here.
2812	 */
2813	mutex_lock(&vm->hotplug_mutex);
2814	spin_lock_irq(&vm->removal_lock);
2815	vm->removing = true;
2816	spin_unlock_irq(&vm->removal_lock);
2817	mutex_unlock(&vm->hotplug_mutex);
2818
2819	/* wait until the workqueue stopped */
2820	cancel_work_sync(&vm->wq);
2821	hrtimer_cancel(&vm->retry_timer);
2822
2823	if (vm->in_sbm) {
2824		/*
2825		 * After we unregistered our callbacks, user space can online
2826		 * partially plugged offline blocks. Make sure to remove them.
2827		 */
2828		virtio_mem_sbm_for_each_mb(vm, mb_id,
2829					   VIRTIO_MEM_SBM_MB_OFFLINE_PARTIAL) {
2830			rc = virtio_mem_sbm_remove_mb(vm, mb_id);
2831			BUG_ON(rc);
2832			virtio_mem_sbm_set_mb_state(vm, mb_id,
2833						    VIRTIO_MEM_SBM_MB_UNUSED);
2834		}
2835		/*
2836		 * After we unregistered our callbacks, user space can no longer
2837		 * offline partially plugged online memory blocks. No need to
2838		 * worry about them.
2839		 */
2840	}
2841
2842	/* unregister callbacks */
2843	unregister_virtio_mem_device(vm);
2844	unregister_memory_notifier(&vm->memory_notifier);
2845
2846	/*
2847	 * There is no way we could reliably remove all memory we have added to
2848	 * the system. And there is no way to stop the driver/device from going
2849	 * away. Warn at least.
2850	 */
2851	if (virtio_mem_has_memory_added(vm)) {
2852		dev_warn(&vm->vdev->dev,
2853			 "device still has system memory added\n");
2854	} else {
2855		virtio_mem_delete_resource(vm);
2856		kfree_const(vm->resource_name);
2857		memory_group_unregister(vm->mgid);
2858	}
2859
2860	/* remove all tracking data - no locking needed */
2861	if (vm->in_sbm) {
2862		vfree(vm->sbm.mb_states);
2863		vfree(vm->sbm.sb_states);
2864	} else {
2865		vfree(vm->bbm.bb_states);
2866	}
2867}
2868
2869static void virtio_mem_deinit_kdump(struct virtio_mem *vm)
2870{
2871#ifdef CONFIG_PROC_VMCORE
2872	unregister_vmcore_cb(&vm->vmcore_cb);
2873#endif /* CONFIG_PROC_VMCORE */
2874}
2875
2876static void virtio_mem_remove(struct virtio_device *vdev)
2877{
2878	struct virtio_mem *vm = vdev->priv;
2879
2880	if (vm->in_kdump)
2881		virtio_mem_deinit_kdump(vm);
2882	else
2883		virtio_mem_deinit_hotplug(vm);
2884
2885	/* reset the device and cleanup the queues */
2886	virtio_reset_device(vdev);
2887	vdev->config->del_vqs(vdev);
2888
2889	kfree(vm);
2890	vdev->priv = NULL;
2891}
2892
2893static void virtio_mem_config_changed(struct virtio_device *vdev)
2894{
2895	struct virtio_mem *vm = vdev->priv;
2896
2897	if (unlikely(vm->in_kdump))
2898		return;
2899
2900	atomic_set(&vm->config_changed, 1);
2901	virtio_mem_retry(vm);
2902}
2903
2904#ifdef CONFIG_PM_SLEEP
2905static int virtio_mem_freeze(struct virtio_device *vdev)
2906{
2907	/*
2908	 * When restarting the VM, all memory is usually unplugged. Don't
2909	 * allow to suspend/hibernate.
2910	 */
2911	dev_err(&vdev->dev, "save/restore not supported.\n");
2912	return -EPERM;
2913}
2914
2915static int virtio_mem_restore(struct virtio_device *vdev)
2916{
2917	return -EPERM;
2918}
2919#endif
2920
2921static unsigned int virtio_mem_features[] = {
2922#if defined(CONFIG_NUMA) && defined(CONFIG_ACPI_NUMA)
2923	VIRTIO_MEM_F_ACPI_PXM,
2924#endif
2925	VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE,
2926};
2927
2928static const struct virtio_device_id virtio_mem_id_table[] = {
2929	{ VIRTIO_ID_MEM, VIRTIO_DEV_ANY_ID },
2930	{ 0 },
2931};
2932
2933static struct virtio_driver virtio_mem_driver = {
2934	.feature_table = virtio_mem_features,
2935	.feature_table_size = ARRAY_SIZE(virtio_mem_features),
2936	.driver.name = KBUILD_MODNAME,
2937	.driver.owner = THIS_MODULE,
2938	.id_table = virtio_mem_id_table,
2939	.probe = virtio_mem_probe,
2940	.remove = virtio_mem_remove,
2941	.config_changed = virtio_mem_config_changed,
2942#ifdef CONFIG_PM_SLEEP
2943	.freeze	=	virtio_mem_freeze,
2944	.restore =	virtio_mem_restore,
2945#endif
2946};
2947
2948module_virtio_driver(virtio_mem_driver);
2949MODULE_DEVICE_TABLE(virtio, virtio_mem_id_table);
2950MODULE_AUTHOR("David Hildenbrand <david@redhat.com>");
2951MODULE_DESCRIPTION("Virtio-mem driver");
2952MODULE_LICENSE("GPL");