Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (C) 2009 Red Hat, Inc.
   3 * Copyright (C) 2006 Rusty Russell IBM Corporation
   4 *
   5 * Author: Michael S. Tsirkin <mst@redhat.com>
   6 *
   7 * Inspiration, some code, and most witty comments come from
   8 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9 *
  10 * Generic code for virtio server in host kernel.
  11 */
  12
  13#include <linux/eventfd.h>
  14#include <linux/vhost.h>
  15#include <linux/uio.h>
  16#include <linux/mm.h>
  17#include <linux/miscdevice.h>
  18#include <linux/mutex.h>
  19#include <linux/poll.h>
  20#include <linux/file.h>
  21#include <linux/highmem.h>
  22#include <linux/slab.h>
  23#include <linux/vmalloc.h>
  24#include <linux/kthread.h>
 
  25#include <linux/module.h>
  26#include <linux/sort.h>
  27#include <linux/sched/mm.h>
  28#include <linux/sched/signal.h>
  29#include <linux/sched/vhost_task.h>
  30#include <linux/interval_tree_generic.h>
  31#include <linux/nospec.h>
  32#include <linux/kcov.h>
  33
  34#include "vhost.h"
  35
  36static ushort max_mem_regions = 64;
  37module_param(max_mem_regions, ushort, 0444);
  38MODULE_PARM_DESC(max_mem_regions,
  39	"Maximum number of memory regions in memory map. (default: 64)");
  40static int max_iotlb_entries = 2048;
  41module_param(max_iotlb_entries, int, 0444);
  42MODULE_PARM_DESC(max_iotlb_entries,
  43	"Maximum number of iotlb entries. (default: 2048)");
  44
  45enum {
  46	VHOST_MEMORY_F_LOG = 0x1,
  47};
  48
  49#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  53static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  54{
  55	vq->user_be = !virtio_legacy_is_little_endian();
  56}
  57
  58static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  59{
  60	vq->user_be = true;
  61}
  62
  63static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  64{
  65	vq->user_be = false;
  66}
  67
  68static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  69{
  70	struct vhost_vring_state s;
  71
  72	if (vq->private_data)
  73		return -EBUSY;
  74
  75	if (copy_from_user(&s, argp, sizeof(s)))
  76		return -EFAULT;
  77
  78	if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  79	    s.num != VHOST_VRING_BIG_ENDIAN)
  80		return -EINVAL;
  81
  82	if (s.num == VHOST_VRING_BIG_ENDIAN)
  83		vhost_enable_cross_endian_big(vq);
  84	else
  85		vhost_enable_cross_endian_little(vq);
  86
  87	return 0;
  88}
  89
  90static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  91				   int __user *argp)
  92{
  93	struct vhost_vring_state s = {
  94		.index = idx,
  95		.num = vq->user_be
  96	};
  97
  98	if (copy_to_user(argp, &s, sizeof(s)))
  99		return -EFAULT;
 100
 101	return 0;
 102}
 103
 104static void vhost_init_is_le(struct vhost_virtqueue *vq)
 105{
 106	/* Note for legacy virtio: user_be is initialized at reset time
 107	 * according to the host endianness. If userspace does not set an
 108	 * explicit endianness, the default behavior is native endian, as
 109	 * expected by legacy virtio.
 110	 */
 111	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 112}
 113#else
 114static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 115{
 116}
 117
 118static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 119{
 120	return -ENOIOCTLCMD;
 121}
 122
 123static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 124				   int __user *argp)
 125{
 126	return -ENOIOCTLCMD;
 127}
 128
 129static void vhost_init_is_le(struct vhost_virtqueue *vq)
 130{
 131	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 132		|| virtio_legacy_is_little_endian();
 133}
 134#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 135
 136static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 137{
 138	vhost_init_is_le(vq);
 139}
 140
 141struct vhost_flush_struct {
 142	struct vhost_work work;
 143	struct completion wait_event;
 144};
 145
 146static void vhost_flush_work(struct vhost_work *work)
 147{
 148	struct vhost_flush_struct *s;
 149
 150	s = container_of(work, struct vhost_flush_struct, work);
 151	complete(&s->wait_event);
 152}
 153
 154static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 155			    poll_table *pt)
 156{
 157	struct vhost_poll *poll;
 158
 159	poll = container_of(pt, struct vhost_poll, table);
 160	poll->wqh = wqh;
 161	add_wait_queue(wqh, &poll->wait);
 162}
 163
 164static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 165			     void *key)
 166{
 167	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 168	struct vhost_work *work = &poll->work;
 169
 170	if (!(key_to_poll(key) & poll->mask))
 171		return 0;
 172
 173	if (!poll->dev->use_worker)
 174		work->fn(work);
 175	else
 176		vhost_poll_queue(poll);
 177
 178	return 0;
 179}
 180
 181void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182{
 183	clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184	work->fn = fn;
 185}
 186EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188/* Init poll structure */
 189void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190		     __poll_t mask, struct vhost_dev *dev,
 191		     struct vhost_virtqueue *vq)
 192{
 193	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 194	init_poll_funcptr(&poll->table, vhost_poll_func);
 195	poll->mask = mask;
 196	poll->dev = dev;
 197	poll->wqh = NULL;
 198	poll->vq = vq;
 199
 200	vhost_work_init(&poll->work, fn);
 201}
 202EXPORT_SYMBOL_GPL(vhost_poll_init);
 203
 204/* Start polling a file. We add ourselves to file's wait queue. The caller must
 205 * keep a reference to a file until after vhost_poll_stop is called. */
 206int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 207{
 208	__poll_t mask;
 209
 210	if (poll->wqh)
 211		return 0;
 212
 213	mask = vfs_poll(file, &poll->table);
 214	if (mask)
 215		vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 216	if (mask & EPOLLERR) {
 217		vhost_poll_stop(poll);
 218		return -EINVAL;
 219	}
 220
 221	return 0;
 222}
 223EXPORT_SYMBOL_GPL(vhost_poll_start);
 224
 225/* Stop polling a file. After this function returns, it becomes safe to drop the
 226 * file reference. You must also flush afterwards. */
 227void vhost_poll_stop(struct vhost_poll *poll)
 228{
 229	if (poll->wqh) {
 230		remove_wait_queue(poll->wqh, &poll->wait);
 231		poll->wqh = NULL;
 232	}
 233}
 234EXPORT_SYMBOL_GPL(vhost_poll_stop);
 235
 236static void vhost_worker_queue(struct vhost_worker *worker,
 237			       struct vhost_work *work)
 238{
 239	if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 240		/* We can only add the work to the list after we're
 241		 * sure it was not in the list.
 242		 * test_and_set_bit() implies a memory barrier.
 243		 */
 244		llist_add(&work->node, &worker->work_list);
 245		vhost_task_wake(worker->vtsk);
 246	}
 247}
 248
 249bool vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work)
 250{
 251	struct vhost_worker *worker;
 252	bool queued = false;
 253
 254	rcu_read_lock();
 255	worker = rcu_dereference(vq->worker);
 256	if (worker) {
 257		queued = true;
 258		vhost_worker_queue(worker, work);
 259	}
 260	rcu_read_unlock();
 261
 262	return queued;
 263}
 264EXPORT_SYMBOL_GPL(vhost_vq_work_queue);
 265
 266/**
 267 * __vhost_worker_flush - flush a worker
 268 * @worker: worker to flush
 269 *
 270 * The worker's flush_mutex must be held.
 271 */
 272static void __vhost_worker_flush(struct vhost_worker *worker)
 273{
 274	struct vhost_flush_struct flush;
 275
 276	if (!worker->attachment_cnt || worker->killed)
 277		return;
 278
 279	init_completion(&flush.wait_event);
 280	vhost_work_init(&flush.work, vhost_flush_work);
 281
 282	vhost_worker_queue(worker, &flush.work);
 283	/*
 284	 * Drop mutex in case our worker is killed and it needs to take the
 285	 * mutex to force cleanup.
 286	 */
 287	mutex_unlock(&worker->mutex);
 288	wait_for_completion(&flush.wait_event);
 289	mutex_lock(&worker->mutex);
 290}
 291
 292static void vhost_worker_flush(struct vhost_worker *worker)
 293{
 294	mutex_lock(&worker->mutex);
 295	__vhost_worker_flush(worker);
 296	mutex_unlock(&worker->mutex);
 297}
 298
 299void vhost_dev_flush(struct vhost_dev *dev)
 300{
 301	struct vhost_worker *worker;
 302	unsigned long i;
 303
 304	xa_for_each(&dev->worker_xa, i, worker)
 305		vhost_worker_flush(worker);
 306}
 307EXPORT_SYMBOL_GPL(vhost_dev_flush);
 308
 309/* A lockless hint for busy polling code to exit the loop */
 310bool vhost_vq_has_work(struct vhost_virtqueue *vq)
 311{
 312	struct vhost_worker *worker;
 313	bool has_work = false;
 314
 315	rcu_read_lock();
 316	worker = rcu_dereference(vq->worker);
 317	if (worker && !llist_empty(&worker->work_list))
 318		has_work = true;
 319	rcu_read_unlock();
 320
 321	return has_work;
 322}
 323EXPORT_SYMBOL_GPL(vhost_vq_has_work);
 324
 325void vhost_poll_queue(struct vhost_poll *poll)
 326{
 327	vhost_vq_work_queue(poll->vq, &poll->work);
 328}
 329EXPORT_SYMBOL_GPL(vhost_poll_queue);
 330
 331static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 332{
 333	int j;
 334
 335	for (j = 0; j < VHOST_NUM_ADDRS; j++)
 336		vq->meta_iotlb[j] = NULL;
 337}
 338
 339static void vhost_vq_meta_reset(struct vhost_dev *d)
 340{
 341	int i;
 342
 343	for (i = 0; i < d->nvqs; ++i)
 344		__vhost_vq_meta_reset(d->vqs[i]);
 345}
 346
 347static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 348{
 349	call_ctx->ctx = NULL;
 350	memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 351}
 352
 353bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 354{
 355	return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 356}
 357EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 358
 359static void vhost_vq_reset(struct vhost_dev *dev,
 360			   struct vhost_virtqueue *vq)
 361{
 362	vq->num = 1;
 363	vq->desc = NULL;
 364	vq->avail = NULL;
 365	vq->used = NULL;
 366	vq->last_avail_idx = 0;
 367	vq->avail_idx = 0;
 368	vq->last_used_idx = 0;
 369	vq->signalled_used = 0;
 370	vq->signalled_used_valid = false;
 371	vq->used_flags = 0;
 372	vq->log_used = false;
 373	vq->log_addr = -1ull;
 374	vq->private_data = NULL;
 375	vq->acked_features = 0;
 376	vq->acked_backend_features = 0;
 377	vq->log_base = NULL;
 378	vq->error_ctx = NULL;
 379	vq->kick = NULL;
 380	vq->log_ctx = NULL;
 381	vhost_disable_cross_endian(vq);
 382	vhost_reset_is_le(vq);
 383	vq->busyloop_timeout = 0;
 384	vq->umem = NULL;
 385	vq->iotlb = NULL;
 386	rcu_assign_pointer(vq->worker, NULL);
 387	vhost_vring_call_reset(&vq->call_ctx);
 388	__vhost_vq_meta_reset(vq);
 389}
 390
 391static bool vhost_run_work_list(void *data)
 392{
 393	struct vhost_worker *worker = data;
 394	struct vhost_work *work, *work_next;
 395	struct llist_node *node;
 396
 397	node = llist_del_all(&worker->work_list);
 398	if (node) {
 399		__set_current_state(TASK_RUNNING);
 
 
 
 
 
 
 
 
 
 
 
 400
 401		node = llist_reverse_order(node);
 402		/* make sure flag is seen after deletion */
 403		smp_wmb();
 404		llist_for_each_entry_safe(work, work_next, node, node) {
 405			clear_bit(VHOST_WORK_QUEUED, &work->flags);
 406			kcov_remote_start_common(worker->kcov_handle);
 
 407			work->fn(work);
 408			kcov_remote_stop();
 409			cond_resched();
 410		}
 411	}
 412
 413	return !!node;
 414}
 415
 416static void vhost_worker_killed(void *data)
 417{
 418	struct vhost_worker *worker = data;
 419	struct vhost_dev *dev = worker->dev;
 420	struct vhost_virtqueue *vq;
 421	int i, attach_cnt = 0;
 422
 423	mutex_lock(&worker->mutex);
 424	worker->killed = true;
 425
 426	for (i = 0; i < dev->nvqs; i++) {
 427		vq = dev->vqs[i];
 428
 429		mutex_lock(&vq->mutex);
 430		if (worker ==
 431		    rcu_dereference_check(vq->worker,
 432					  lockdep_is_held(&vq->mutex))) {
 433			rcu_assign_pointer(vq->worker, NULL);
 434			attach_cnt++;
 435		}
 436		mutex_unlock(&vq->mutex);
 437	}
 438
 439	worker->attachment_cnt -= attach_cnt;
 440	if (attach_cnt)
 441		synchronize_rcu();
 442	/*
 443	 * Finish vhost_worker_flush calls and any other works that snuck in
 444	 * before the synchronize_rcu.
 445	 */
 446	vhost_run_work_list(worker);
 447	mutex_unlock(&worker->mutex);
 448}
 449
 450static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 451{
 452	kfree(vq->indirect);
 453	vq->indirect = NULL;
 454	kfree(vq->log);
 455	vq->log = NULL;
 456	kfree(vq->heads);
 457	vq->heads = NULL;
 458}
 459
 460/* Helper to allocate iovec buffers for all vqs. */
 461static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 462{
 463	struct vhost_virtqueue *vq;
 464	int i;
 465
 466	for (i = 0; i < dev->nvqs; ++i) {
 467		vq = dev->vqs[i];
 468		vq->indirect = kmalloc_array(UIO_MAXIOV,
 469					     sizeof(*vq->indirect),
 470					     GFP_KERNEL);
 471		vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 472					GFP_KERNEL);
 473		vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 474					  GFP_KERNEL);
 475		if (!vq->indirect || !vq->log || !vq->heads)
 476			goto err_nomem;
 477	}
 478	return 0;
 479
 480err_nomem:
 481	for (; i >= 0; --i)
 482		vhost_vq_free_iovecs(dev->vqs[i]);
 483	return -ENOMEM;
 484}
 485
 486static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 487{
 488	int i;
 489
 490	for (i = 0; i < dev->nvqs; ++i)
 491		vhost_vq_free_iovecs(dev->vqs[i]);
 492}
 493
 494bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 495			  int pkts, int total_len)
 496{
 497	struct vhost_dev *dev = vq->dev;
 498
 499	if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 500	    pkts >= dev->weight) {
 501		vhost_poll_queue(&vq->poll);
 502		return true;
 503	}
 504
 505	return false;
 506}
 507EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 508
 509static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 510				   unsigned int num)
 511{
 512	size_t event __maybe_unused =
 513	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 514
 515	return size_add(struct_size(vq->avail, ring, num), event);
 
 516}
 517
 518static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 519				  unsigned int num)
 520{
 521	size_t event __maybe_unused =
 522	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 523
 524	return size_add(struct_size(vq->used, ring, num), event);
 
 525}
 526
 527static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 528				  unsigned int num)
 529{
 530	return sizeof(*vq->desc) * num;
 531}
 532
 533void vhost_dev_init(struct vhost_dev *dev,
 534		    struct vhost_virtqueue **vqs, int nvqs,
 535		    int iov_limit, int weight, int byte_weight,
 536		    bool use_worker,
 537		    int (*msg_handler)(struct vhost_dev *dev, u32 asid,
 538				       struct vhost_iotlb_msg *msg))
 539{
 540	struct vhost_virtqueue *vq;
 541	int i;
 542
 543	dev->vqs = vqs;
 544	dev->nvqs = nvqs;
 545	mutex_init(&dev->mutex);
 546	dev->log_ctx = NULL;
 547	dev->umem = NULL;
 548	dev->iotlb = NULL;
 549	dev->mm = NULL;
 
 550	dev->iov_limit = iov_limit;
 551	dev->weight = weight;
 552	dev->byte_weight = byte_weight;
 553	dev->use_worker = use_worker;
 554	dev->msg_handler = msg_handler;
 
 555	init_waitqueue_head(&dev->wait);
 556	INIT_LIST_HEAD(&dev->read_list);
 557	INIT_LIST_HEAD(&dev->pending_list);
 558	spin_lock_init(&dev->iotlb_lock);
 559	xa_init_flags(&dev->worker_xa, XA_FLAGS_ALLOC);
 560
 561	for (i = 0; i < dev->nvqs; ++i) {
 562		vq = dev->vqs[i];
 563		vq->log = NULL;
 564		vq->indirect = NULL;
 565		vq->heads = NULL;
 566		vq->dev = dev;
 567		mutex_init(&vq->mutex);
 568		vhost_vq_reset(dev, vq);
 569		if (vq->handle_kick)
 570			vhost_poll_init(&vq->poll, vq->handle_kick,
 571					EPOLLIN, dev, vq);
 572	}
 573}
 574EXPORT_SYMBOL_GPL(vhost_dev_init);
 575
 576/* Caller should have device mutex */
 577long vhost_dev_check_owner(struct vhost_dev *dev)
 578{
 579	/* Are you the owner? If not, I don't think you mean to do that */
 580	return dev->mm == current->mm ? 0 : -EPERM;
 581}
 582EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 583
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 584/* Caller should have device mutex */
 585bool vhost_dev_has_owner(struct vhost_dev *dev)
 586{
 587	return dev->mm;
 588}
 589EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 590
 591static void vhost_attach_mm(struct vhost_dev *dev)
 592{
 593	/* No owner, become one */
 594	if (dev->use_worker) {
 595		dev->mm = get_task_mm(current);
 596	} else {
 597		/* vDPA device does not use worker thead, so there's
 598		 * no need to hold the address space for mm. This help
 599		 * to avoid deadlock in the case of mmap() which may
 600		 * held the refcnt of the file and depends on release
 601		 * method to remove vma.
 602		 */
 603		dev->mm = current->mm;
 604		mmgrab(dev->mm);
 605	}
 606}
 607
 608static void vhost_detach_mm(struct vhost_dev *dev)
 609{
 610	if (!dev->mm)
 611		return;
 612
 613	if (dev->use_worker)
 614		mmput(dev->mm);
 615	else
 616		mmdrop(dev->mm);
 617
 618	dev->mm = NULL;
 619}
 620
 621static void vhost_worker_destroy(struct vhost_dev *dev,
 622				 struct vhost_worker *worker)
 623{
 624	if (!worker)
 625		return;
 626
 627	WARN_ON(!llist_empty(&worker->work_list));
 628	xa_erase(&dev->worker_xa, worker->id);
 629	vhost_task_stop(worker->vtsk);
 630	kfree(worker);
 631}
 632
 633static void vhost_workers_free(struct vhost_dev *dev)
 634{
 635	struct vhost_worker *worker;
 636	unsigned long i;
 637
 638	if (!dev->use_worker)
 639		return;
 640
 641	for (i = 0; i < dev->nvqs; i++)
 642		rcu_assign_pointer(dev->vqs[i]->worker, NULL);
 643	/*
 644	 * Free the default worker we created and cleanup workers userspace
 645	 * created but couldn't clean up (it forgot or crashed).
 646	 */
 647	xa_for_each(&dev->worker_xa, i, worker)
 648		vhost_worker_destroy(dev, worker);
 649	xa_destroy(&dev->worker_xa);
 650}
 651
 652static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
 653{
 654	struct vhost_worker *worker;
 655	struct vhost_task *vtsk;
 656	char name[TASK_COMM_LEN];
 657	int ret;
 658	u32 id;
 659
 660	worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
 661	if (!worker)
 662		return NULL;
 663
 664	worker->dev = dev;
 665	snprintf(name, sizeof(name), "vhost-%d", current->pid);
 666
 667	vtsk = vhost_task_create(vhost_run_work_list, vhost_worker_killed,
 668				 worker, name);
 669	if (!vtsk)
 670		goto free_worker;
 671
 672	mutex_init(&worker->mutex);
 673	init_llist_head(&worker->work_list);
 674	worker->kcov_handle = kcov_common_handle();
 675	worker->vtsk = vtsk;
 676
 677	vhost_task_start(vtsk);
 678
 679	ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
 680	if (ret < 0)
 681		goto stop_worker;
 682	worker->id = id;
 683
 684	return worker;
 685
 686stop_worker:
 687	vhost_task_stop(vtsk);
 688free_worker:
 689	kfree(worker);
 690	return NULL;
 691}
 692
 693/* Caller must have device mutex */
 694static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 695				     struct vhost_worker *worker)
 696{
 697	struct vhost_worker *old_worker;
 698
 699	mutex_lock(&worker->mutex);
 700	if (worker->killed) {
 701		mutex_unlock(&worker->mutex);
 702		return;
 703	}
 704
 705	mutex_lock(&vq->mutex);
 706
 707	old_worker = rcu_dereference_check(vq->worker,
 708					   lockdep_is_held(&vq->mutex));
 709	rcu_assign_pointer(vq->worker, worker);
 710	worker->attachment_cnt++;
 711
 712	if (!old_worker) {
 713		mutex_unlock(&vq->mutex);
 714		mutex_unlock(&worker->mutex);
 715		return;
 716	}
 717	mutex_unlock(&vq->mutex);
 718	mutex_unlock(&worker->mutex);
 719
 720	/*
 721	 * Take the worker mutex to make sure we see the work queued from
 722	 * device wide flushes which doesn't use RCU for execution.
 723	 */
 724	mutex_lock(&old_worker->mutex);
 725	if (old_worker->killed) {
 726		mutex_unlock(&old_worker->mutex);
 727		return;
 728	}
 729
 730	/*
 731	 * We don't want to call synchronize_rcu for every vq during setup
 732	 * because it will slow down VM startup. If we haven't done
 733	 * VHOST_SET_VRING_KICK and not done the driver specific
 734	 * SET_ENDPOINT/RUNNUNG then we can skip the sync since there will
 735	 * not be any works queued for scsi and net.
 736	 */
 737	mutex_lock(&vq->mutex);
 738	if (!vhost_vq_get_backend(vq) && !vq->kick) {
 739		mutex_unlock(&vq->mutex);
 740
 741		old_worker->attachment_cnt--;
 742		mutex_unlock(&old_worker->mutex);
 743		/*
 744		 * vsock can queue anytime after VHOST_VSOCK_SET_GUEST_CID.
 745		 * Warn if it adds support for multiple workers but forgets to
 746		 * handle the early queueing case.
 747		 */
 748		WARN_ON(!old_worker->attachment_cnt &&
 749			!llist_empty(&old_worker->work_list));
 750		return;
 751	}
 752	mutex_unlock(&vq->mutex);
 753
 754	/* Make sure new vq queue/flush/poll calls see the new worker */
 755	synchronize_rcu();
 756	/* Make sure whatever was queued gets run */
 757	__vhost_worker_flush(old_worker);
 758	old_worker->attachment_cnt--;
 759	mutex_unlock(&old_worker->mutex);
 760}
 761
 762 /* Caller must have device mutex */
 763static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 764				  struct vhost_vring_worker *info)
 765{
 766	unsigned long index = info->worker_id;
 767	struct vhost_dev *dev = vq->dev;
 768	struct vhost_worker *worker;
 769
 770	if (!dev->use_worker)
 771		return -EINVAL;
 772
 773	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
 774	if (!worker || worker->id != info->worker_id)
 775		return -ENODEV;
 776
 777	__vhost_vq_attach_worker(vq, worker);
 778	return 0;
 779}
 780
 781/* Caller must have device mutex */
 782static int vhost_new_worker(struct vhost_dev *dev,
 783			    struct vhost_worker_state *info)
 784{
 785	struct vhost_worker *worker;
 786
 787	worker = vhost_worker_create(dev);
 788	if (!worker)
 789		return -ENOMEM;
 790
 791	info->worker_id = worker->id;
 792	return 0;
 793}
 794
 795/* Caller must have device mutex */
 796static int vhost_free_worker(struct vhost_dev *dev,
 797			     struct vhost_worker_state *info)
 798{
 799	unsigned long index = info->worker_id;
 800	struct vhost_worker *worker;
 801
 802	worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
 803	if (!worker || worker->id != info->worker_id)
 804		return -ENODEV;
 805
 806	mutex_lock(&worker->mutex);
 807	if (worker->attachment_cnt || worker->killed) {
 808		mutex_unlock(&worker->mutex);
 809		return -EBUSY;
 810	}
 811	/*
 812	 * A flush might have raced and snuck in before attachment_cnt was set
 813	 * to zero. Make sure flushes are flushed from the queue before
 814	 * freeing.
 815	 */
 816	__vhost_worker_flush(worker);
 817	mutex_unlock(&worker->mutex);
 818
 819	vhost_worker_destroy(dev, worker);
 820	return 0;
 821}
 822
 823static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
 824				  struct vhost_virtqueue **vq, u32 *id)
 825{
 826	u32 __user *idxp = argp;
 827	u32 idx;
 828	long r;
 829
 830	r = get_user(idx, idxp);
 831	if (r < 0)
 832		return r;
 833
 834	if (idx >= dev->nvqs)
 835		return -ENOBUFS;
 836
 837	idx = array_index_nospec(idx, dev->nvqs);
 838
 839	*vq = dev->vqs[idx];
 840	*id = idx;
 841	return 0;
 842}
 843
 844/* Caller must have device mutex */
 845long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
 846			void __user *argp)
 847{
 848	struct vhost_vring_worker ring_worker;
 849	struct vhost_worker_state state;
 850	struct vhost_worker *worker;
 851	struct vhost_virtqueue *vq;
 852	long ret;
 853	u32 idx;
 854
 855	if (!dev->use_worker)
 856		return -EINVAL;
 857
 858	if (!vhost_dev_has_owner(dev))
 859		return -EINVAL;
 860
 861	ret = vhost_dev_check_owner(dev);
 862	if (ret)
 863		return ret;
 864
 865	switch (ioctl) {
 866	/* dev worker ioctls */
 867	case VHOST_NEW_WORKER:
 868		ret = vhost_new_worker(dev, &state);
 869		if (!ret && copy_to_user(argp, &state, sizeof(state)))
 870			ret = -EFAULT;
 871		return ret;
 872	case VHOST_FREE_WORKER:
 873		if (copy_from_user(&state, argp, sizeof(state)))
 874			return -EFAULT;
 875		return vhost_free_worker(dev, &state);
 876	/* vring worker ioctls */
 877	case VHOST_ATTACH_VRING_WORKER:
 878	case VHOST_GET_VRING_WORKER:
 879		break;
 880	default:
 881		return -ENOIOCTLCMD;
 882	}
 883
 884	ret = vhost_get_vq_from_user(dev, argp, &vq, &idx);
 885	if (ret)
 886		return ret;
 887
 888	switch (ioctl) {
 889	case VHOST_ATTACH_VRING_WORKER:
 890		if (copy_from_user(&ring_worker, argp, sizeof(ring_worker))) {
 891			ret = -EFAULT;
 892			break;
 893		}
 894
 895		ret = vhost_vq_attach_worker(vq, &ring_worker);
 896		break;
 897	case VHOST_GET_VRING_WORKER:
 898		worker = rcu_dereference_check(vq->worker,
 899					       lockdep_is_held(&dev->mutex));
 900		if (!worker) {
 901			ret = -EINVAL;
 902			break;
 903		}
 904
 905		ring_worker.index = idx;
 906		ring_worker.worker_id = worker->id;
 907
 908		if (copy_to_user(argp, &ring_worker, sizeof(ring_worker)))
 909			ret = -EFAULT;
 910		break;
 911	default:
 912		ret = -ENOIOCTLCMD;
 913		break;
 914	}
 915
 916	return ret;
 917}
 918EXPORT_SYMBOL_GPL(vhost_worker_ioctl);
 919
 920/* Caller should have device mutex */
 921long vhost_dev_set_owner(struct vhost_dev *dev)
 922{
 923	struct vhost_worker *worker;
 924	int err, i;
 925
 926	/* Is there an owner already? */
 927	if (vhost_dev_has_owner(dev)) {
 928		err = -EBUSY;
 929		goto err_mm;
 930	}
 931
 932	vhost_attach_mm(dev);
 933
 934	err = vhost_dev_alloc_iovecs(dev);
 935	if (err)
 936		goto err_iovecs;
 937
 938	if (dev->use_worker) {
 939		/*
 940		 * This should be done last, because vsock can queue work
 941		 * before VHOST_SET_OWNER so it simplifies the failure path
 942		 * below since we don't have to worry about vsock queueing
 943		 * while we free the worker.
 944		 */
 945		worker = vhost_worker_create(dev);
 946		if (!worker) {
 947			err = -ENOMEM;
 948			goto err_worker;
 949		}
 950
 951		for (i = 0; i < dev->nvqs; i++)
 952			__vhost_vq_attach_worker(dev->vqs[i], worker);
 
 
 
 
 953	}
 954
 955	return 0;
 
 
 956
 
 
 
 
 
 
 957err_worker:
 958	vhost_dev_free_iovecs(dev);
 959err_iovecs:
 960	vhost_detach_mm(dev);
 
 961err_mm:
 962	return err;
 963}
 964EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 965
 966static struct vhost_iotlb *iotlb_alloc(void)
 967{
 968	return vhost_iotlb_alloc(max_iotlb_entries,
 969				 VHOST_IOTLB_FLAG_RETIRE);
 970}
 971
 972struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 973{
 974	return iotlb_alloc();
 975}
 976EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 977
 978/* Caller should have device mutex */
 979void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 980{
 981	int i;
 982
 983	vhost_dev_cleanup(dev);
 984
 985	dev->umem = umem;
 986	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
 987	 * VQs aren't running.
 988	 */
 989	for (i = 0; i < dev->nvqs; ++i)
 990		dev->vqs[i]->umem = umem;
 991}
 992EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 993
 994void vhost_dev_stop(struct vhost_dev *dev)
 995{
 996	int i;
 997
 998	for (i = 0; i < dev->nvqs; ++i) {
 999		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
1000			vhost_poll_stop(&dev->vqs[i]->poll);
1001	}
1002
1003	vhost_dev_flush(dev);
1004}
1005EXPORT_SYMBOL_GPL(vhost_dev_stop);
1006
1007void vhost_clear_msg(struct vhost_dev *dev)
1008{
1009	struct vhost_msg_node *node, *n;
1010
1011	spin_lock(&dev->iotlb_lock);
1012
1013	list_for_each_entry_safe(node, n, &dev->read_list, node) {
1014		list_del(&node->node);
1015		kfree(node);
1016	}
1017
1018	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
1019		list_del(&node->node);
1020		kfree(node);
1021	}
1022
1023	spin_unlock(&dev->iotlb_lock);
1024}
1025EXPORT_SYMBOL_GPL(vhost_clear_msg);
1026
1027void vhost_dev_cleanup(struct vhost_dev *dev)
1028{
1029	int i;
1030
1031	for (i = 0; i < dev->nvqs; ++i) {
1032		if (dev->vqs[i]->error_ctx)
1033			eventfd_ctx_put(dev->vqs[i]->error_ctx);
1034		if (dev->vqs[i]->kick)
1035			fput(dev->vqs[i]->kick);
1036		if (dev->vqs[i]->call_ctx.ctx)
1037			eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
1038		vhost_vq_reset(dev, dev->vqs[i]);
1039	}
1040	vhost_dev_free_iovecs(dev);
1041	if (dev->log_ctx)
1042		eventfd_ctx_put(dev->log_ctx);
1043	dev->log_ctx = NULL;
1044	/* No one will access memory at this point */
1045	vhost_iotlb_free(dev->umem);
1046	dev->umem = NULL;
1047	vhost_iotlb_free(dev->iotlb);
1048	dev->iotlb = NULL;
1049	vhost_clear_msg(dev);
1050	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
1051	vhost_workers_free(dev);
 
 
 
 
 
1052	vhost_detach_mm(dev);
1053}
1054EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
1055
1056static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
1057{
1058	u64 a = addr / VHOST_PAGE_SIZE / 8;
1059
1060	/* Make sure 64 bit math will not overflow. */
1061	if (a > ULONG_MAX - (unsigned long)log_base ||
1062	    a + (unsigned long)log_base > ULONG_MAX)
1063		return false;
1064
1065	return access_ok(log_base + a,
1066			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
1067}
1068
1069/* Make sure 64 bit math will not overflow. */
1070static bool vhost_overflow(u64 uaddr, u64 size)
1071{
1072	if (uaddr > ULONG_MAX || size > ULONG_MAX)
1073		return true;
1074
1075	if (!size)
1076		return false;
1077
1078	return uaddr > ULONG_MAX - size + 1;
1079}
1080
1081/* Caller should have vq mutex and device mutex. */
1082static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
1083				int log_all)
1084{
1085	struct vhost_iotlb_map *map;
1086
1087	if (!umem)
1088		return false;
1089
1090	list_for_each_entry(map, &umem->list, link) {
1091		unsigned long a = map->addr;
1092
1093		if (vhost_overflow(map->addr, map->size))
1094			return false;
1095
1096
1097		if (!access_ok((void __user *)a, map->size))
1098			return false;
1099		else if (log_all && !log_access_ok(log_base,
1100						   map->start,
1101						   map->size))
1102			return false;
1103	}
1104	return true;
1105}
1106
1107static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
1108					       u64 addr, unsigned int size,
1109					       int type)
1110{
1111	const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
1112
1113	if (!map)
1114		return NULL;
1115
1116	return (void __user *)(uintptr_t)(map->addr + addr - map->start);
1117}
1118
1119/* Can we switch to this memory table? */
1120/* Caller should have device mutex but not vq mutex */
1121static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
1122			     int log_all)
1123{
1124	int i;
1125
1126	for (i = 0; i < d->nvqs; ++i) {
1127		bool ok;
1128		bool log;
1129
1130		mutex_lock(&d->vqs[i]->mutex);
1131		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
1132		/* If ring is inactive, will check when it's enabled. */
1133		if (d->vqs[i]->private_data)
1134			ok = vq_memory_access_ok(d->vqs[i]->log_base,
1135						 umem, log);
1136		else
1137			ok = true;
1138		mutex_unlock(&d->vqs[i]->mutex);
1139		if (!ok)
1140			return false;
1141	}
1142	return true;
1143}
1144
1145static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1146			  struct iovec iov[], int iov_size, int access);
1147
1148static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
1149			      const void *from, unsigned size)
1150{
1151	int ret;
1152
1153	if (!vq->iotlb)
1154		return __copy_to_user(to, from, size);
1155	else {
1156		/* This function should be called after iotlb
1157		 * prefetch, which means we're sure that all vq
1158		 * could be access through iotlb. So -EAGAIN should
1159		 * not happen in this case.
1160		 */
1161		struct iov_iter t;
1162		void __user *uaddr = vhost_vq_meta_fetch(vq,
1163				     (u64)(uintptr_t)to, size,
1164				     VHOST_ADDR_USED);
1165
1166		if (uaddr)
1167			return __copy_to_user(uaddr, from, size);
1168
1169		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
1170				     ARRAY_SIZE(vq->iotlb_iov),
1171				     VHOST_ACCESS_WO);
1172		if (ret < 0)
1173			goto out;
1174		iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
1175		ret = copy_to_iter(from, size, &t);
1176		if (ret == size)
1177			ret = 0;
1178	}
1179out:
1180	return ret;
1181}
1182
1183static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
1184				void __user *from, unsigned size)
1185{
1186	int ret;
1187
1188	if (!vq->iotlb)
1189		return __copy_from_user(to, from, size);
1190	else {
1191		/* This function should be called after iotlb
1192		 * prefetch, which means we're sure that vq
1193		 * could be access through iotlb. So -EAGAIN should
1194		 * not happen in this case.
1195		 */
1196		void __user *uaddr = vhost_vq_meta_fetch(vq,
1197				     (u64)(uintptr_t)from, size,
1198				     VHOST_ADDR_DESC);
1199		struct iov_iter f;
1200
1201		if (uaddr)
1202			return __copy_from_user(to, uaddr, size);
1203
1204		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
1205				     ARRAY_SIZE(vq->iotlb_iov),
1206				     VHOST_ACCESS_RO);
1207		if (ret < 0) {
1208			vq_err(vq, "IOTLB translation failure: uaddr "
1209			       "%p size 0x%llx\n", from,
1210			       (unsigned long long) size);
1211			goto out;
1212		}
1213		iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
1214		ret = copy_from_iter(to, size, &f);
1215		if (ret == size)
1216			ret = 0;
1217	}
1218
1219out:
1220	return ret;
1221}
1222
1223static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
1224					  void __user *addr, unsigned int size,
1225					  int type)
1226{
1227	int ret;
1228
1229	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
1230			     ARRAY_SIZE(vq->iotlb_iov),
1231			     VHOST_ACCESS_RO);
1232	if (ret < 0) {
1233		vq_err(vq, "IOTLB translation failure: uaddr "
1234			"%p size 0x%llx\n", addr,
1235			(unsigned long long) size);
1236		return NULL;
1237	}
1238
1239	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
1240		vq_err(vq, "Non atomic userspace memory access: uaddr "
1241			"%p size 0x%llx\n", addr,
1242			(unsigned long long) size);
1243		return NULL;
1244	}
1245
1246	return vq->iotlb_iov[0].iov_base;
1247}
1248
1249/* This function should be called after iotlb
1250 * prefetch, which means we're sure that vq
1251 * could be access through iotlb. So -EAGAIN should
1252 * not happen in this case.
1253 */
1254static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
1255					    void __user *addr, unsigned int size,
1256					    int type)
1257{
1258	void __user *uaddr = vhost_vq_meta_fetch(vq,
1259			     (u64)(uintptr_t)addr, size, type);
1260	if (uaddr)
1261		return uaddr;
1262
1263	return __vhost_get_user_slow(vq, addr, size, type);
1264}
1265
1266#define vhost_put_user(vq, x, ptr)		\
1267({ \
1268	int ret; \
1269	if (!vq->iotlb) { \
1270		ret = __put_user(x, ptr); \
1271	} else { \
1272		__typeof__(ptr) to = \
1273			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
1274					  sizeof(*ptr), VHOST_ADDR_USED); \
1275		if (to != NULL) \
1276			ret = __put_user(x, to); \
1277		else \
1278			ret = -EFAULT;	\
1279	} \
1280	ret; \
1281})
1282
1283static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
1284{
1285	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1286			      vhost_avail_event(vq));
1287}
1288
1289static inline int vhost_put_used(struct vhost_virtqueue *vq,
1290				 struct vring_used_elem *head, int idx,
1291				 int count)
1292{
1293	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
1294				  count * sizeof(*head));
1295}
1296
1297static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
1298
1299{
1300	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1301			      &vq->used->flags);
1302}
1303
1304static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
1305
1306{
1307	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
1308			      &vq->used->idx);
1309}
1310
1311#define vhost_get_user(vq, x, ptr, type)		\
1312({ \
1313	int ret; \
1314	if (!vq->iotlb) { \
1315		ret = __get_user(x, ptr); \
1316	} else { \
1317		__typeof__(ptr) from = \
1318			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
1319							   sizeof(*ptr), \
1320							   type); \
1321		if (from != NULL) \
1322			ret = __get_user(x, from); \
1323		else \
1324			ret = -EFAULT; \
1325	} \
1326	ret; \
1327})
1328
1329#define vhost_get_avail(vq, x, ptr) \
1330	vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1331
1332#define vhost_get_used(vq, x, ptr) \
1333	vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1334
1335static void vhost_dev_lock_vqs(struct vhost_dev *d)
1336{
1337	int i = 0;
1338	for (i = 0; i < d->nvqs; ++i)
1339		mutex_lock_nested(&d->vqs[i]->mutex, i);
1340}
1341
1342static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1343{
1344	int i = 0;
1345	for (i = 0; i < d->nvqs; ++i)
1346		mutex_unlock(&d->vqs[i]->mutex);
1347}
1348
1349static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq)
 
1350{
1351	__virtio16 idx;
1352	int r;
1353
1354	r = vhost_get_avail(vq, idx, &vq->avail->idx);
1355	if (unlikely(r < 0)) {
1356		vq_err(vq, "Failed to access available index at %p (%d)\n",
1357		       &vq->avail->idx, r);
1358		return r;
1359	}
1360
1361	/* Check it isn't doing very strange thing with available indexes */
1362	vq->avail_idx = vhost16_to_cpu(vq, idx);
1363	if (unlikely((u16)(vq->avail_idx - vq->last_avail_idx) > vq->num)) {
1364		vq_err(vq, "Invalid available index change from %u to %u",
1365		       vq->last_avail_idx, vq->avail_idx);
1366		return -EINVAL;
1367	}
1368
1369	/* We're done if there is nothing new */
1370	if (vq->avail_idx == vq->last_avail_idx)
1371		return 0;
1372
1373	/*
1374	 * We updated vq->avail_idx so we need a memory barrier between
1375	 * the index read above and the caller reading avail ring entries.
1376	 */
1377	smp_rmb();
1378	return 1;
1379}
1380
1381static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1382				       __virtio16 *head, int idx)
1383{
1384	return vhost_get_avail(vq, *head,
1385			       &vq->avail->ring[idx & (vq->num - 1)]);
1386}
1387
1388static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1389					__virtio16 *flags)
1390{
1391	return vhost_get_avail(vq, *flags, &vq->avail->flags);
1392}
1393
1394static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1395				       __virtio16 *event)
1396{
1397	return vhost_get_avail(vq, *event, vhost_used_event(vq));
1398}
1399
1400static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1401				     __virtio16 *idx)
1402{
1403	return vhost_get_used(vq, *idx, &vq->used->idx);
1404}
1405
1406static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1407				 struct vring_desc *desc, int idx)
1408{
1409	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1410}
1411
1412static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1413				  struct vhost_iotlb_msg *msg)
1414{
1415	struct vhost_msg_node *node, *n;
1416
1417	spin_lock(&d->iotlb_lock);
1418
1419	list_for_each_entry_safe(node, n, &d->pending_list, node) {
1420		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1421		if (msg->iova <= vq_msg->iova &&
1422		    msg->iova + msg->size - 1 >= vq_msg->iova &&
1423		    vq_msg->type == VHOST_IOTLB_MISS) {
1424			vhost_poll_queue(&node->vq->poll);
1425			list_del(&node->node);
1426			kfree(node);
1427		}
1428	}
1429
1430	spin_unlock(&d->iotlb_lock);
1431}
1432
1433static bool umem_access_ok(u64 uaddr, u64 size, int access)
1434{
1435	unsigned long a = uaddr;
1436
1437	/* Make sure 64 bit math will not overflow. */
1438	if (vhost_overflow(uaddr, size))
1439		return false;
1440
1441	if ((access & VHOST_ACCESS_RO) &&
1442	    !access_ok((void __user *)a, size))
1443		return false;
1444	if ((access & VHOST_ACCESS_WO) &&
1445	    !access_ok((void __user *)a, size))
1446		return false;
1447	return true;
1448}
1449
1450static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1451				   struct vhost_iotlb_msg *msg)
1452{
1453	int ret = 0;
1454
1455	if (asid != 0)
1456		return -EINVAL;
1457
1458	mutex_lock(&dev->mutex);
1459	vhost_dev_lock_vqs(dev);
1460	switch (msg->type) {
1461	case VHOST_IOTLB_UPDATE:
1462		if (!dev->iotlb) {
1463			ret = -EFAULT;
1464			break;
1465		}
1466		if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1467			ret = -EFAULT;
1468			break;
1469		}
1470		vhost_vq_meta_reset(dev);
1471		if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1472					  msg->iova + msg->size - 1,
1473					  msg->uaddr, msg->perm)) {
1474			ret = -ENOMEM;
1475			break;
1476		}
1477		vhost_iotlb_notify_vq(dev, msg);
1478		break;
1479	case VHOST_IOTLB_INVALIDATE:
1480		if (!dev->iotlb) {
1481			ret = -EFAULT;
1482			break;
1483		}
1484		vhost_vq_meta_reset(dev);
1485		vhost_iotlb_del_range(dev->iotlb, msg->iova,
1486				      msg->iova + msg->size - 1);
1487		break;
1488	default:
1489		ret = -EINVAL;
1490		break;
1491	}
1492
1493	vhost_dev_unlock_vqs(dev);
1494	mutex_unlock(&dev->mutex);
1495
1496	return ret;
1497}
1498ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1499			     struct iov_iter *from)
1500{
1501	struct vhost_iotlb_msg msg;
1502	size_t offset;
1503	int type, ret;
1504	u32 asid = 0;
1505
1506	ret = copy_from_iter(&type, sizeof(type), from);
1507	if (ret != sizeof(type)) {
1508		ret = -EINVAL;
1509		goto done;
1510	}
1511
1512	switch (type) {
1513	case VHOST_IOTLB_MSG:
1514		/* There maybe a hole after type for V1 message type,
1515		 * so skip it here.
1516		 */
1517		offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1518		break;
1519	case VHOST_IOTLB_MSG_V2:
1520		if (vhost_backend_has_feature(dev->vqs[0],
1521					      VHOST_BACKEND_F_IOTLB_ASID)) {
1522			ret = copy_from_iter(&asid, sizeof(asid), from);
1523			if (ret != sizeof(asid)) {
1524				ret = -EINVAL;
1525				goto done;
1526			}
1527			offset = 0;
1528		} else
1529			offset = sizeof(__u32);
1530		break;
1531	default:
1532		ret = -EINVAL;
1533		goto done;
1534	}
1535
1536	iov_iter_advance(from, offset);
1537	ret = copy_from_iter(&msg, sizeof(msg), from);
1538	if (ret != sizeof(msg)) {
1539		ret = -EINVAL;
1540		goto done;
1541	}
1542
1543	if (msg.type == VHOST_IOTLB_UPDATE && msg.size == 0) {
 
 
1544		ret = -EINVAL;
1545		goto done;
1546	}
1547
1548	if (dev->msg_handler)
1549		ret = dev->msg_handler(dev, asid, &msg);
1550	else
1551		ret = vhost_process_iotlb_msg(dev, asid, &msg);
1552	if (ret) {
1553		ret = -EFAULT;
1554		goto done;
1555	}
1556
1557	ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1558	      sizeof(struct vhost_msg_v2);
1559done:
1560	return ret;
1561}
1562EXPORT_SYMBOL(vhost_chr_write_iter);
1563
1564__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1565			    poll_table *wait)
1566{
1567	__poll_t mask = 0;
1568
1569	poll_wait(file, &dev->wait, wait);
1570
1571	if (!list_empty(&dev->read_list))
1572		mask |= EPOLLIN | EPOLLRDNORM;
1573
1574	return mask;
1575}
1576EXPORT_SYMBOL(vhost_chr_poll);
1577
1578ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1579			    int noblock)
1580{
1581	DEFINE_WAIT(wait);
1582	struct vhost_msg_node *node;
1583	ssize_t ret = 0;
1584	unsigned size = sizeof(struct vhost_msg);
1585
1586	if (iov_iter_count(to) < size)
1587		return 0;
1588
1589	while (1) {
1590		if (!noblock)
1591			prepare_to_wait(&dev->wait, &wait,
1592					TASK_INTERRUPTIBLE);
1593
1594		node = vhost_dequeue_msg(dev, &dev->read_list);
1595		if (node)
1596			break;
1597		if (noblock) {
1598			ret = -EAGAIN;
1599			break;
1600		}
1601		if (signal_pending(current)) {
1602			ret = -ERESTARTSYS;
1603			break;
1604		}
1605		if (!dev->iotlb) {
1606			ret = -EBADFD;
1607			break;
1608		}
1609
1610		schedule();
1611	}
1612
1613	if (!noblock)
1614		finish_wait(&dev->wait, &wait);
1615
1616	if (node) {
1617		struct vhost_iotlb_msg *msg;
1618		void *start = &node->msg;
1619
1620		switch (node->msg.type) {
1621		case VHOST_IOTLB_MSG:
1622			size = sizeof(node->msg);
1623			msg = &node->msg.iotlb;
1624			break;
1625		case VHOST_IOTLB_MSG_V2:
1626			size = sizeof(node->msg_v2);
1627			msg = &node->msg_v2.iotlb;
1628			break;
1629		default:
1630			BUG();
1631			break;
1632		}
1633
1634		ret = copy_to_iter(start, size, to);
1635		if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1636			kfree(node);
1637			return ret;
1638		}
1639		vhost_enqueue_msg(dev, &dev->pending_list, node);
1640	}
1641
1642	return ret;
1643}
1644EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1645
1646static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1647{
1648	struct vhost_dev *dev = vq->dev;
1649	struct vhost_msg_node *node;
1650	struct vhost_iotlb_msg *msg;
1651	bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1652
1653	node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1654	if (!node)
1655		return -ENOMEM;
1656
1657	if (v2) {
1658		node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1659		msg = &node->msg_v2.iotlb;
1660	} else {
1661		msg = &node->msg.iotlb;
1662	}
1663
1664	msg->type = VHOST_IOTLB_MISS;
1665	msg->iova = iova;
1666	msg->perm = access;
1667
1668	vhost_enqueue_msg(dev, &dev->read_list, node);
1669
1670	return 0;
1671}
1672
1673static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1674			 vring_desc_t __user *desc,
1675			 vring_avail_t __user *avail,
1676			 vring_used_t __user *used)
1677
1678{
1679	/* If an IOTLB device is present, the vring addresses are
1680	 * GIOVAs. Access validation occurs at prefetch time. */
1681	if (vq->iotlb)
1682		return true;
1683
1684	return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1685	       access_ok(avail, vhost_get_avail_size(vq, num)) &&
1686	       access_ok(used, vhost_get_used_size(vq, num));
1687}
1688
1689static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1690				 const struct vhost_iotlb_map *map,
1691				 int type)
1692{
1693	int access = (type == VHOST_ADDR_USED) ?
1694		     VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1695
1696	if (likely(map->perm & access))
1697		vq->meta_iotlb[type] = map;
1698}
1699
1700static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1701			    int access, u64 addr, u64 len, int type)
1702{
1703	const struct vhost_iotlb_map *map;
1704	struct vhost_iotlb *umem = vq->iotlb;
1705	u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1706
1707	if (vhost_vq_meta_fetch(vq, addr, len, type))
1708		return true;
1709
1710	while (len > s) {
1711		map = vhost_iotlb_itree_first(umem, addr, last);
1712		if (map == NULL || map->start > addr) {
1713			vhost_iotlb_miss(vq, addr, access);
1714			return false;
1715		} else if (!(map->perm & access)) {
1716			/* Report the possible access violation by
1717			 * request another translation from userspace.
1718			 */
1719			return false;
1720		}
1721
1722		size = map->size - addr + map->start;
1723
1724		if (orig_addr == addr && size >= len)
1725			vhost_vq_meta_update(vq, map, type);
1726
1727		s += size;
1728		addr += size;
1729	}
1730
1731	return true;
1732}
1733
1734int vq_meta_prefetch(struct vhost_virtqueue *vq)
1735{
1736	unsigned int num = vq->num;
1737
1738	if (!vq->iotlb)
1739		return 1;
1740
1741	return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1742			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1743	       iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1744			       vhost_get_avail_size(vq, num),
1745			       VHOST_ADDR_AVAIL) &&
1746	       iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1747			       vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1748}
1749EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1750
1751/* Can we log writes? */
1752/* Caller should have device mutex but not vq mutex */
1753bool vhost_log_access_ok(struct vhost_dev *dev)
1754{
1755	return memory_access_ok(dev, dev->umem, 1);
1756}
1757EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1758
1759static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1760				  void __user *log_base,
1761				  bool log_used,
1762				  u64 log_addr)
1763{
1764	/* If an IOTLB device is present, log_addr is a GIOVA that
1765	 * will never be logged by log_used(). */
1766	if (vq->iotlb)
1767		return true;
1768
1769	return !log_used || log_access_ok(log_base, log_addr,
1770					  vhost_get_used_size(vq, vq->num));
1771}
1772
1773/* Verify access for write logging. */
1774/* Caller should have vq mutex and device mutex */
1775static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1776			     void __user *log_base)
1777{
1778	return vq_memory_access_ok(log_base, vq->umem,
1779				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1780		vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1781}
1782
1783/* Can we start vq? */
1784/* Caller should have vq mutex and device mutex */
1785bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1786{
1787	if (!vq_log_access_ok(vq, vq->log_base))
1788		return false;
1789
1790	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1791}
1792EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1793
1794static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1795{
1796	struct vhost_memory mem, *newmem;
1797	struct vhost_memory_region *region;
1798	struct vhost_iotlb *newumem, *oldumem;
1799	unsigned long size = offsetof(struct vhost_memory, regions);
1800	int i;
1801
1802	if (copy_from_user(&mem, m, size))
1803		return -EFAULT;
1804	if (mem.padding)
1805		return -EOPNOTSUPP;
1806	if (mem.nregions > max_mem_regions)
1807		return -E2BIG;
1808	newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1809			GFP_KERNEL);
1810	if (!newmem)
1811		return -ENOMEM;
1812
1813	memcpy(newmem, &mem, size);
1814	if (copy_from_user(newmem->regions, m->regions,
1815			   flex_array_size(newmem, regions, mem.nregions))) {
1816		kvfree(newmem);
1817		return -EFAULT;
1818	}
1819
1820	newumem = iotlb_alloc();
1821	if (!newumem) {
1822		kvfree(newmem);
1823		return -ENOMEM;
1824	}
1825
1826	for (region = newmem->regions;
1827	     region < newmem->regions + mem.nregions;
1828	     region++) {
1829		if (vhost_iotlb_add_range(newumem,
1830					  region->guest_phys_addr,
1831					  region->guest_phys_addr +
1832					  region->memory_size - 1,
1833					  region->userspace_addr,
1834					  VHOST_MAP_RW))
1835			goto err;
1836	}
1837
1838	if (!memory_access_ok(d, newumem, 0))
1839		goto err;
1840
1841	oldumem = d->umem;
1842	d->umem = newumem;
1843
1844	/* All memory accesses are done under some VQ mutex. */
1845	for (i = 0; i < d->nvqs; ++i) {
1846		mutex_lock(&d->vqs[i]->mutex);
1847		d->vqs[i]->umem = newumem;
1848		mutex_unlock(&d->vqs[i]->mutex);
1849	}
1850
1851	kvfree(newmem);
1852	vhost_iotlb_free(oldumem);
1853	return 0;
1854
1855err:
1856	vhost_iotlb_free(newumem);
1857	kvfree(newmem);
1858	return -EFAULT;
1859}
1860
1861static long vhost_vring_set_num(struct vhost_dev *d,
1862				struct vhost_virtqueue *vq,
1863				void __user *argp)
1864{
1865	struct vhost_vring_state s;
1866
1867	/* Resizing ring with an active backend?
1868	 * You don't want to do that. */
1869	if (vq->private_data)
1870		return -EBUSY;
1871
1872	if (copy_from_user(&s, argp, sizeof s))
1873		return -EFAULT;
1874
1875	if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1876		return -EINVAL;
1877	vq->num = s.num;
1878
1879	return 0;
1880}
1881
1882static long vhost_vring_set_addr(struct vhost_dev *d,
1883				 struct vhost_virtqueue *vq,
1884				 void __user *argp)
1885{
1886	struct vhost_vring_addr a;
1887
1888	if (copy_from_user(&a, argp, sizeof a))
1889		return -EFAULT;
1890	if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1891		return -EOPNOTSUPP;
1892
1893	/* For 32bit, verify that the top 32bits of the user
1894	   data are set to zero. */
1895	if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1896	    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1897	    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1898		return -EFAULT;
1899
1900	/* Make sure it's safe to cast pointers to vring types. */
1901	BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1902	BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1903	if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1904	    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1905	    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1906		return -EINVAL;
1907
1908	/* We only verify access here if backend is configured.
1909	 * If it is not, we don't as size might not have been setup.
1910	 * We will verify when backend is configured. */
1911	if (vq->private_data) {
1912		if (!vq_access_ok(vq, vq->num,
1913			(void __user *)(unsigned long)a.desc_user_addr,
1914			(void __user *)(unsigned long)a.avail_user_addr,
1915			(void __user *)(unsigned long)a.used_user_addr))
1916			return -EINVAL;
1917
1918		/* Also validate log access for used ring if enabled. */
1919		if (!vq_log_used_access_ok(vq, vq->log_base,
1920				a.flags & (0x1 << VHOST_VRING_F_LOG),
1921				a.log_guest_addr))
1922			return -EINVAL;
1923	}
1924
1925	vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1926	vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1927	vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1928	vq->log_addr = a.log_guest_addr;
1929	vq->used = (void __user *)(unsigned long)a.used_user_addr;
1930
1931	return 0;
1932}
1933
1934static long vhost_vring_set_num_addr(struct vhost_dev *d,
1935				     struct vhost_virtqueue *vq,
1936				     unsigned int ioctl,
1937				     void __user *argp)
1938{
1939	long r;
1940
1941	mutex_lock(&vq->mutex);
1942
1943	switch (ioctl) {
1944	case VHOST_SET_VRING_NUM:
1945		r = vhost_vring_set_num(d, vq, argp);
1946		break;
1947	case VHOST_SET_VRING_ADDR:
1948		r = vhost_vring_set_addr(d, vq, argp);
1949		break;
1950	default:
1951		BUG();
1952	}
1953
1954	mutex_unlock(&vq->mutex);
1955
1956	return r;
1957}
1958long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1959{
1960	struct file *eventfp, *filep = NULL;
1961	bool pollstart = false, pollstop = false;
1962	struct eventfd_ctx *ctx = NULL;
 
1963	struct vhost_virtqueue *vq;
1964	struct vhost_vring_state s;
1965	struct vhost_vring_file f;
1966	u32 idx;
1967	long r;
1968
1969	r = vhost_get_vq_from_user(d, argp, &vq, &idx);
1970	if (r < 0)
1971		return r;
 
 
 
 
 
1972
1973	if (ioctl == VHOST_SET_VRING_NUM ||
1974	    ioctl == VHOST_SET_VRING_ADDR) {
1975		return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1976	}
1977
1978	mutex_lock(&vq->mutex);
1979
1980	switch (ioctl) {
1981	case VHOST_SET_VRING_BASE:
1982		/* Moving base with an active backend?
1983		 * You don't want to do that. */
1984		if (vq->private_data) {
1985			r = -EBUSY;
1986			break;
1987		}
1988		if (copy_from_user(&s, argp, sizeof s)) {
1989			r = -EFAULT;
1990			break;
1991		}
1992		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1993			vq->last_avail_idx = s.num & 0xffff;
1994			vq->last_used_idx = (s.num >> 16) & 0xffff;
1995		} else {
1996			if (s.num > 0xffff) {
1997				r = -EINVAL;
1998				break;
1999			}
2000			vq->last_avail_idx = s.num;
2001		}
 
2002		/* Forget the cached index value. */
2003		vq->avail_idx = vq->last_avail_idx;
2004		break;
2005	case VHOST_GET_VRING_BASE:
2006		s.index = idx;
2007		if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
2008			s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
2009		else
2010			s.num = vq->last_avail_idx;
2011		if (copy_to_user(argp, &s, sizeof s))
2012			r = -EFAULT;
2013		break;
2014	case VHOST_SET_VRING_KICK:
2015		if (copy_from_user(&f, argp, sizeof f)) {
2016			r = -EFAULT;
2017			break;
2018		}
2019		eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
2020		if (IS_ERR(eventfp)) {
2021			r = PTR_ERR(eventfp);
2022			break;
2023		}
2024		if (eventfp != vq->kick) {
2025			pollstop = (filep = vq->kick) != NULL;
2026			pollstart = (vq->kick = eventfp) != NULL;
2027		} else
2028			filep = eventfp;
2029		break;
2030	case VHOST_SET_VRING_CALL:
2031		if (copy_from_user(&f, argp, sizeof f)) {
2032			r = -EFAULT;
2033			break;
2034		}
2035		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
2036		if (IS_ERR(ctx)) {
2037			r = PTR_ERR(ctx);
2038			break;
2039		}
2040
2041		swap(ctx, vq->call_ctx.ctx);
2042		break;
2043	case VHOST_SET_VRING_ERR:
2044		if (copy_from_user(&f, argp, sizeof f)) {
2045			r = -EFAULT;
2046			break;
2047		}
2048		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
2049		if (IS_ERR(ctx)) {
2050			r = PTR_ERR(ctx);
2051			break;
2052		}
2053		swap(ctx, vq->error_ctx);
2054		break;
2055	case VHOST_SET_VRING_ENDIAN:
2056		r = vhost_set_vring_endian(vq, argp);
2057		break;
2058	case VHOST_GET_VRING_ENDIAN:
2059		r = vhost_get_vring_endian(vq, idx, argp);
2060		break;
2061	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
2062		if (copy_from_user(&s, argp, sizeof(s))) {
2063			r = -EFAULT;
2064			break;
2065		}
2066		vq->busyloop_timeout = s.num;
2067		break;
2068	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
2069		s.index = idx;
2070		s.num = vq->busyloop_timeout;
2071		if (copy_to_user(argp, &s, sizeof(s)))
2072			r = -EFAULT;
2073		break;
2074	default:
2075		r = -ENOIOCTLCMD;
2076	}
2077
2078	if (pollstop && vq->handle_kick)
2079		vhost_poll_stop(&vq->poll);
2080
2081	if (!IS_ERR_OR_NULL(ctx))
2082		eventfd_ctx_put(ctx);
2083	if (filep)
2084		fput(filep);
2085
2086	if (pollstart && vq->handle_kick)
2087		r = vhost_poll_start(&vq->poll, vq->kick);
2088
2089	mutex_unlock(&vq->mutex);
2090
2091	if (pollstop && vq->handle_kick)
2092		vhost_dev_flush(vq->poll.dev);
2093	return r;
2094}
2095EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
2096
2097int vhost_init_device_iotlb(struct vhost_dev *d)
2098{
2099	struct vhost_iotlb *niotlb, *oiotlb;
2100	int i;
2101
2102	niotlb = iotlb_alloc();
2103	if (!niotlb)
2104		return -ENOMEM;
2105
2106	oiotlb = d->iotlb;
2107	d->iotlb = niotlb;
2108
2109	for (i = 0; i < d->nvqs; ++i) {
2110		struct vhost_virtqueue *vq = d->vqs[i];
2111
2112		mutex_lock(&vq->mutex);
2113		vq->iotlb = niotlb;
2114		__vhost_vq_meta_reset(vq);
2115		mutex_unlock(&vq->mutex);
2116	}
2117
2118	vhost_iotlb_free(oiotlb);
2119
2120	return 0;
2121}
2122EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
2123
2124/* Caller must have device mutex */
2125long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
2126{
2127	struct eventfd_ctx *ctx;
2128	u64 p;
2129	long r;
2130	int i, fd;
2131
2132	/* If you are not the owner, you can become one */
2133	if (ioctl == VHOST_SET_OWNER) {
2134		r = vhost_dev_set_owner(d);
2135		goto done;
2136	}
2137
2138	/* You must be the owner to do anything else */
2139	r = vhost_dev_check_owner(d);
2140	if (r)
2141		goto done;
2142
2143	switch (ioctl) {
2144	case VHOST_SET_MEM_TABLE:
2145		r = vhost_set_memory(d, argp);
2146		break;
2147	case VHOST_SET_LOG_BASE:
2148		if (copy_from_user(&p, argp, sizeof p)) {
2149			r = -EFAULT;
2150			break;
2151		}
2152		if ((u64)(unsigned long)p != p) {
2153			r = -EFAULT;
2154			break;
2155		}
2156		for (i = 0; i < d->nvqs; ++i) {
2157			struct vhost_virtqueue *vq;
2158			void __user *base = (void __user *)(unsigned long)p;
2159			vq = d->vqs[i];
2160			mutex_lock(&vq->mutex);
2161			/* If ring is inactive, will check when it's enabled. */
2162			if (vq->private_data && !vq_log_access_ok(vq, base))
2163				r = -EFAULT;
2164			else
2165				vq->log_base = base;
2166			mutex_unlock(&vq->mutex);
2167		}
2168		break;
2169	case VHOST_SET_LOG_FD:
2170		r = get_user(fd, (int __user *)argp);
2171		if (r < 0)
2172			break;
2173		ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
2174		if (IS_ERR(ctx)) {
2175			r = PTR_ERR(ctx);
2176			break;
2177		}
2178		swap(ctx, d->log_ctx);
2179		for (i = 0; i < d->nvqs; ++i) {
2180			mutex_lock(&d->vqs[i]->mutex);
2181			d->vqs[i]->log_ctx = d->log_ctx;
2182			mutex_unlock(&d->vqs[i]->mutex);
2183		}
2184		if (ctx)
2185			eventfd_ctx_put(ctx);
2186		break;
2187	default:
2188		r = -ENOIOCTLCMD;
2189		break;
2190	}
2191done:
2192	return r;
2193}
2194EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
2195
2196/* TODO: This is really inefficient.  We need something like get_user()
2197 * (instruction directly accesses the data, with an exception table entry
2198 * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
2199 */
2200static int set_bit_to_user(int nr, void __user *addr)
2201{
2202	unsigned long log = (unsigned long)addr;
2203	struct page *page;
2204	void *base;
2205	int bit = nr + (log % PAGE_SIZE) * 8;
2206	int r;
2207
2208	r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
2209	if (r < 0)
2210		return r;
2211	BUG_ON(r != 1);
2212	base = kmap_atomic(page);
2213	set_bit(bit, base);
2214	kunmap_atomic(base);
2215	unpin_user_pages_dirty_lock(&page, 1, true);
2216	return 0;
2217}
2218
2219static int log_write(void __user *log_base,
2220		     u64 write_address, u64 write_length)
2221{
2222	u64 write_page = write_address / VHOST_PAGE_SIZE;
2223	int r;
2224
2225	if (!write_length)
2226		return 0;
2227	write_length += write_address % VHOST_PAGE_SIZE;
2228	for (;;) {
2229		u64 base = (u64)(unsigned long)log_base;
2230		u64 log = base + write_page / 8;
2231		int bit = write_page % 8;
2232		if ((u64)(unsigned long)log != log)
2233			return -EFAULT;
2234		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
2235		if (r < 0)
2236			return r;
2237		if (write_length <= VHOST_PAGE_SIZE)
2238			break;
2239		write_length -= VHOST_PAGE_SIZE;
2240		write_page += 1;
2241	}
2242	return r;
2243}
2244
2245static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
2246{
2247	struct vhost_iotlb *umem = vq->umem;
2248	struct vhost_iotlb_map *u;
2249	u64 start, end, l, min;
2250	int r;
2251	bool hit = false;
2252
2253	while (len) {
2254		min = len;
2255		/* More than one GPAs can be mapped into a single HVA. So
2256		 * iterate all possible umems here to be safe.
2257		 */
2258		list_for_each_entry(u, &umem->list, link) {
2259			if (u->addr > hva - 1 + len ||
2260			    u->addr - 1 + u->size < hva)
2261				continue;
2262			start = max(u->addr, hva);
2263			end = min(u->addr - 1 + u->size, hva - 1 + len);
2264			l = end - start + 1;
2265			r = log_write(vq->log_base,
2266				      u->start + start - u->addr,
2267				      l);
2268			if (r < 0)
2269				return r;
2270			hit = true;
2271			min = min(l, min);
2272		}
2273
2274		if (!hit)
2275			return -EFAULT;
2276
2277		len -= min;
2278		hva += min;
2279	}
2280
2281	return 0;
2282}
2283
2284static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
2285{
2286	struct iovec *iov = vq->log_iov;
2287	int i, ret;
2288
2289	if (!vq->iotlb)
2290		return log_write(vq->log_base, vq->log_addr + used_offset, len);
2291
2292	ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
2293			     len, iov, 64, VHOST_ACCESS_WO);
2294	if (ret < 0)
2295		return ret;
2296
2297	for (i = 0; i < ret; i++) {
2298		ret = log_write_hva(vq,	(uintptr_t)iov[i].iov_base,
2299				    iov[i].iov_len);
2300		if (ret)
2301			return ret;
2302	}
2303
2304	return 0;
2305}
2306
2307int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
2308		    unsigned int log_num, u64 len, struct iovec *iov, int count)
2309{
2310	int i, r;
2311
2312	/* Make sure data written is seen before log. */
2313	smp_wmb();
2314
2315	if (vq->iotlb) {
2316		for (i = 0; i < count; i++) {
2317			r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2318					  iov[i].iov_len);
2319			if (r < 0)
2320				return r;
2321		}
2322		return 0;
2323	}
2324
2325	for (i = 0; i < log_num; ++i) {
2326		u64 l = min(log[i].len, len);
2327		r = log_write(vq->log_base, log[i].addr, l);
2328		if (r < 0)
2329			return r;
2330		len -= l;
2331		if (!len) {
2332			if (vq->log_ctx)
2333				eventfd_signal(vq->log_ctx);
2334			return 0;
2335		}
2336	}
2337	/* Length written exceeds what we have stored. This is a bug. */
2338	BUG();
2339	return 0;
2340}
2341EXPORT_SYMBOL_GPL(vhost_log_write);
2342
2343static int vhost_update_used_flags(struct vhost_virtqueue *vq)
2344{
2345	void __user *used;
2346	if (vhost_put_used_flags(vq))
2347		return -EFAULT;
2348	if (unlikely(vq->log_used)) {
2349		/* Make sure the flag is seen before log. */
2350		smp_wmb();
2351		/* Log used flag write. */
2352		used = &vq->used->flags;
2353		log_used(vq, (used - (void __user *)vq->used),
2354			 sizeof vq->used->flags);
2355		if (vq->log_ctx)
2356			eventfd_signal(vq->log_ctx);
2357	}
2358	return 0;
2359}
2360
2361static int vhost_update_avail_event(struct vhost_virtqueue *vq)
2362{
2363	if (vhost_put_avail_event(vq))
2364		return -EFAULT;
2365	if (unlikely(vq->log_used)) {
2366		void __user *used;
2367		/* Make sure the event is seen before log. */
2368		smp_wmb();
2369		/* Log avail event write */
2370		used = vhost_avail_event(vq);
2371		log_used(vq, (used - (void __user *)vq->used),
2372			 sizeof *vhost_avail_event(vq));
2373		if (vq->log_ctx)
2374			eventfd_signal(vq->log_ctx);
2375	}
2376	return 0;
2377}
2378
2379int vhost_vq_init_access(struct vhost_virtqueue *vq)
2380{
2381	__virtio16 last_used_idx;
2382	int r;
2383	bool is_le = vq->is_le;
2384
2385	if (!vq->private_data)
2386		return 0;
2387
2388	vhost_init_is_le(vq);
2389
2390	r = vhost_update_used_flags(vq);
2391	if (r)
2392		goto err;
2393	vq->signalled_used_valid = false;
2394	if (!vq->iotlb &&
2395	    !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2396		r = -EFAULT;
2397		goto err;
2398	}
2399	r = vhost_get_used_idx(vq, &last_used_idx);
2400	if (r) {
2401		vq_err(vq, "Can't access used idx at %p\n",
2402		       &vq->used->idx);
2403		goto err;
2404	}
2405	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2406	return 0;
2407
2408err:
2409	vq->is_le = is_le;
2410	return r;
2411}
2412EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2413
2414static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2415			  struct iovec iov[], int iov_size, int access)
2416{
2417	const struct vhost_iotlb_map *map;
2418	struct vhost_dev *dev = vq->dev;
2419	struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2420	struct iovec *_iov;
2421	u64 s = 0, last = addr + len - 1;
2422	int ret = 0;
2423
2424	while ((u64)len > s) {
2425		u64 size;
2426		if (unlikely(ret >= iov_size)) {
2427			ret = -ENOBUFS;
2428			break;
2429		}
2430
2431		map = vhost_iotlb_itree_first(umem, addr, last);
2432		if (map == NULL || map->start > addr) {
2433			if (umem != dev->iotlb) {
2434				ret = -EFAULT;
2435				break;
2436			}
2437			ret = -EAGAIN;
2438			break;
2439		} else if (!(map->perm & access)) {
2440			ret = -EPERM;
2441			break;
2442		}
2443
2444		_iov = iov + ret;
2445		size = map->size - addr + map->start;
2446		_iov->iov_len = min((u64)len - s, size);
2447		_iov->iov_base = (void __user *)(unsigned long)
2448				 (map->addr + addr - map->start);
2449		s += size;
2450		addr += size;
2451		++ret;
2452	}
2453
2454	if (ret == -EAGAIN)
2455		vhost_iotlb_miss(vq, addr, access);
2456	return ret;
2457}
2458
2459/* Each buffer in the virtqueues is actually a chain of descriptors.  This
2460 * function returns the next descriptor in the chain,
2461 * or -1U if we're at the end. */
2462static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2463{
2464	unsigned int next;
2465
2466	/* If this descriptor says it doesn't chain, we're done. */
2467	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2468		return -1U;
2469
2470	/* Check they're not leading us off end of descriptors. */
2471	next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2472	return next;
2473}
2474
2475static int get_indirect(struct vhost_virtqueue *vq,
2476			struct iovec iov[], unsigned int iov_size,
2477			unsigned int *out_num, unsigned int *in_num,
2478			struct vhost_log *log, unsigned int *log_num,
2479			struct vring_desc *indirect)
2480{
2481	struct vring_desc desc;
2482	unsigned int i = 0, count, found = 0;
2483	u32 len = vhost32_to_cpu(vq, indirect->len);
2484	struct iov_iter from;
2485	int ret, access;
2486
2487	/* Sanity check */
2488	if (unlikely(len % sizeof desc)) {
2489		vq_err(vq, "Invalid length in indirect descriptor: "
2490		       "len 0x%llx not multiple of 0x%zx\n",
2491		       (unsigned long long)len,
2492		       sizeof desc);
2493		return -EINVAL;
2494	}
2495
2496	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2497			     UIO_MAXIOV, VHOST_ACCESS_RO);
2498	if (unlikely(ret < 0)) {
2499		if (ret != -EAGAIN)
2500			vq_err(vq, "Translation failure %d in indirect.\n", ret);
2501		return ret;
2502	}
2503	iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2504	count = len / sizeof desc;
2505	/* Buffers are chained via a 16 bit next field, so
2506	 * we can have at most 2^16 of these. */
2507	if (unlikely(count > USHRT_MAX + 1)) {
2508		vq_err(vq, "Indirect buffer length too big: %d\n",
2509		       indirect->len);
2510		return -E2BIG;
2511	}
2512
2513	do {
2514		unsigned iov_count = *in_num + *out_num;
2515		if (unlikely(++found > count)) {
2516			vq_err(vq, "Loop detected: last one at %u "
2517			       "indirect size %u\n",
2518			       i, count);
2519			return -EINVAL;
2520		}
2521		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2522			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2523			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2524			return -EINVAL;
2525		}
2526		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2527			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2528			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2529			return -EINVAL;
2530		}
2531
2532		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2533			access = VHOST_ACCESS_WO;
2534		else
2535			access = VHOST_ACCESS_RO;
2536
2537		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2538				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2539				     iov_size - iov_count, access);
2540		if (unlikely(ret < 0)) {
2541			if (ret != -EAGAIN)
2542				vq_err(vq, "Translation failure %d indirect idx %d\n",
2543					ret, i);
2544			return ret;
2545		}
2546		/* If this is an input descriptor, increment that count. */
2547		if (access == VHOST_ACCESS_WO) {
2548			*in_num += ret;
2549			if (unlikely(log && ret)) {
2550				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2551				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2552				++*log_num;
2553			}
2554		} else {
2555			/* If it's an output descriptor, they're all supposed
2556			 * to come before any input descriptors. */
2557			if (unlikely(*in_num)) {
2558				vq_err(vq, "Indirect descriptor "
2559				       "has out after in: idx %d\n", i);
2560				return -EINVAL;
2561			}
2562			*out_num += ret;
2563		}
2564	} while ((i = next_desc(vq, &desc)) != -1);
2565	return 0;
2566}
2567
2568/* This looks in the virtqueue and for the first available buffer, and converts
2569 * it to an iovec for convenient access.  Since descriptors consist of some
2570 * number of output then some number of input descriptors, it's actually two
2571 * iovecs, but we pack them into one and note how many of each there were.
2572 *
2573 * This function returns the descriptor number found, or vq->num (which is
2574 * never a valid descriptor number) if none was found.  A negative code is
2575 * returned on error. */
2576int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2577		      struct iovec iov[], unsigned int iov_size,
2578		      unsigned int *out_num, unsigned int *in_num,
2579		      struct vhost_log *log, unsigned int *log_num)
2580{
2581	struct vring_desc desc;
2582	unsigned int i, head, found = 0;
2583	u16 last_avail_idx = vq->last_avail_idx;
 
2584	__virtio16 ring_head;
2585	int ret, access;
2586
 
 
 
2587	if (vq->avail_idx == vq->last_avail_idx) {
2588		ret = vhost_get_avail_idx(vq);
2589		if (unlikely(ret < 0))
2590			return ret;
 
 
 
 
 
 
 
 
 
2591
2592		if (!ret)
 
 
 
2593			return vq->num;
 
 
 
 
 
2594	}
2595
2596	/* Grab the next descriptor number they're advertising, and increment
2597	 * the index we've seen. */
2598	if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2599		vq_err(vq, "Failed to read head: idx %d address %p\n",
2600		       last_avail_idx,
2601		       &vq->avail->ring[last_avail_idx % vq->num]);
2602		return -EFAULT;
2603	}
2604
2605	head = vhost16_to_cpu(vq, ring_head);
2606
2607	/* If their number is silly, that's an error. */
2608	if (unlikely(head >= vq->num)) {
2609		vq_err(vq, "Guest says index %u > %u is available",
2610		       head, vq->num);
2611		return -EINVAL;
2612	}
2613
2614	/* When we start there are none of either input nor output. */
2615	*out_num = *in_num = 0;
2616	if (unlikely(log))
2617		*log_num = 0;
2618
2619	i = head;
2620	do {
2621		unsigned iov_count = *in_num + *out_num;
2622		if (unlikely(i >= vq->num)) {
2623			vq_err(vq, "Desc index is %u > %u, head = %u",
2624			       i, vq->num, head);
2625			return -EINVAL;
2626		}
2627		if (unlikely(++found > vq->num)) {
2628			vq_err(vq, "Loop detected: last one at %u "
2629			       "vq size %u head %u\n",
2630			       i, vq->num, head);
2631			return -EINVAL;
2632		}
2633		ret = vhost_get_desc(vq, &desc, i);
2634		if (unlikely(ret)) {
2635			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2636			       i, vq->desc + i);
2637			return -EFAULT;
2638		}
2639		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2640			ret = get_indirect(vq, iov, iov_size,
2641					   out_num, in_num,
2642					   log, log_num, &desc);
2643			if (unlikely(ret < 0)) {
2644				if (ret != -EAGAIN)
2645					vq_err(vq, "Failure detected "
2646						"in indirect descriptor at idx %d\n", i);
2647				return ret;
2648			}
2649			continue;
2650		}
2651
2652		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2653			access = VHOST_ACCESS_WO;
2654		else
2655			access = VHOST_ACCESS_RO;
2656		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2657				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2658				     iov_size - iov_count, access);
2659		if (unlikely(ret < 0)) {
2660			if (ret != -EAGAIN)
2661				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2662					ret, i);
2663			return ret;
2664		}
2665		if (access == VHOST_ACCESS_WO) {
2666			/* If this is an input descriptor,
2667			 * increment that count. */
2668			*in_num += ret;
2669			if (unlikely(log && ret)) {
2670				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2671				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2672				++*log_num;
2673			}
2674		} else {
2675			/* If it's an output descriptor, they're all supposed
2676			 * to come before any input descriptors. */
2677			if (unlikely(*in_num)) {
2678				vq_err(vq, "Descriptor has out after in: "
2679				       "idx %d\n", i);
2680				return -EINVAL;
2681			}
2682			*out_num += ret;
2683		}
2684	} while ((i = next_desc(vq, &desc)) != -1);
2685
2686	/* On success, increment avail index. */
2687	vq->last_avail_idx++;
2688
2689	/* Assume notifications from guest are disabled at this point,
2690	 * if they aren't we would need to update avail_event index. */
2691	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2692	return head;
2693}
2694EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2695
2696/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2697void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2698{
2699	vq->last_avail_idx -= n;
2700}
2701EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2702
2703/* After we've used one of their buffers, we tell them about it.  We'll then
2704 * want to notify the guest, using eventfd. */
2705int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2706{
2707	struct vring_used_elem heads = {
2708		cpu_to_vhost32(vq, head),
2709		cpu_to_vhost32(vq, len)
2710	};
2711
2712	return vhost_add_used_n(vq, &heads, 1);
2713}
2714EXPORT_SYMBOL_GPL(vhost_add_used);
2715
2716static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2717			    struct vring_used_elem *heads,
2718			    unsigned count)
2719{
2720	vring_used_elem_t __user *used;
2721	u16 old, new;
2722	int start;
2723
2724	start = vq->last_used_idx & (vq->num - 1);
2725	used = vq->used->ring + start;
2726	if (vhost_put_used(vq, heads, start, count)) {
2727		vq_err(vq, "Failed to write used");
2728		return -EFAULT;
2729	}
2730	if (unlikely(vq->log_used)) {
2731		/* Make sure data is seen before log. */
2732		smp_wmb();
2733		/* Log used ring entry write. */
2734		log_used(vq, ((void __user *)used - (void __user *)vq->used),
2735			 count * sizeof *used);
2736	}
2737	old = vq->last_used_idx;
2738	new = (vq->last_used_idx += count);
2739	/* If the driver never bothers to signal in a very long while,
2740	 * used index might wrap around. If that happens, invalidate
2741	 * signalled_used index we stored. TODO: make sure driver
2742	 * signals at least once in 2^16 and remove this. */
2743	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2744		vq->signalled_used_valid = false;
2745	return 0;
2746}
2747
2748/* After we've used one of their buffers, we tell them about it.  We'll then
2749 * want to notify the guest, using eventfd. */
2750int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2751		     unsigned count)
2752{
2753	int start, n, r;
2754
2755	start = vq->last_used_idx & (vq->num - 1);
2756	n = vq->num - start;
2757	if (n < count) {
2758		r = __vhost_add_used_n(vq, heads, n);
2759		if (r < 0)
2760			return r;
2761		heads += n;
2762		count -= n;
2763	}
2764	r = __vhost_add_used_n(vq, heads, count);
2765
2766	/* Make sure buffer is written before we update index. */
2767	smp_wmb();
2768	if (vhost_put_used_idx(vq)) {
2769		vq_err(vq, "Failed to increment used idx");
2770		return -EFAULT;
2771	}
2772	if (unlikely(vq->log_used)) {
2773		/* Make sure used idx is seen before log. */
2774		smp_wmb();
2775		/* Log used index update. */
2776		log_used(vq, offsetof(struct vring_used, idx),
2777			 sizeof vq->used->idx);
2778		if (vq->log_ctx)
2779			eventfd_signal(vq->log_ctx);
2780	}
2781	return r;
2782}
2783EXPORT_SYMBOL_GPL(vhost_add_used_n);
2784
2785static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2786{
2787	__u16 old, new;
2788	__virtio16 event;
2789	bool v;
2790	/* Flush out used index updates. This is paired
2791	 * with the barrier that the Guest executes when enabling
2792	 * interrupts. */
2793	smp_mb();
2794
2795	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2796	    unlikely(vq->avail_idx == vq->last_avail_idx))
2797		return true;
2798
2799	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2800		__virtio16 flags;
2801		if (vhost_get_avail_flags(vq, &flags)) {
2802			vq_err(vq, "Failed to get flags");
2803			return true;
2804		}
2805		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2806	}
2807	old = vq->signalled_used;
2808	v = vq->signalled_used_valid;
2809	new = vq->signalled_used = vq->last_used_idx;
2810	vq->signalled_used_valid = true;
2811
2812	if (unlikely(!v))
2813		return true;
2814
2815	if (vhost_get_used_event(vq, &event)) {
2816		vq_err(vq, "Failed to get used event idx");
2817		return true;
2818	}
2819	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2820}
2821
2822/* This actually signals the guest, using eventfd. */
2823void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2824{
2825	/* Signal the Guest tell them we used something up. */
2826	if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2827		eventfd_signal(vq->call_ctx.ctx);
2828}
2829EXPORT_SYMBOL_GPL(vhost_signal);
2830
2831/* And here's the combo meal deal.  Supersize me! */
2832void vhost_add_used_and_signal(struct vhost_dev *dev,
2833			       struct vhost_virtqueue *vq,
2834			       unsigned int head, int len)
2835{
2836	vhost_add_used(vq, head, len);
2837	vhost_signal(dev, vq);
2838}
2839EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2840
2841/* multi-buffer version of vhost_add_used_and_signal */
2842void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2843				 struct vhost_virtqueue *vq,
2844				 struct vring_used_elem *heads, unsigned count)
2845{
2846	vhost_add_used_n(vq, heads, count);
2847	vhost_signal(dev, vq);
2848}
2849EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2850
2851/* return true if we're sure that avaiable ring is empty */
2852bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2853{
 
2854	int r;
2855
2856	if (vq->avail_idx != vq->last_avail_idx)
2857		return false;
2858
2859	r = vhost_get_avail_idx(vq);
 
 
 
2860
2861	/* Note: we treat error as non-empty here */
2862	return r == 0;
2863}
2864EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2865
2866/* OK, now we need to know about added descriptors. */
2867bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2868{
 
2869	int r;
2870
2871	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2872		return false;
2873	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2874	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2875		r = vhost_update_used_flags(vq);
2876		if (r) {
2877			vq_err(vq, "Failed to enable notification at %p: %d\n",
2878			       &vq->used->flags, r);
2879			return false;
2880		}
2881	} else {
2882		r = vhost_update_avail_event(vq);
2883		if (r) {
2884			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2885			       vhost_avail_event(vq), r);
2886			return false;
2887		}
2888	}
2889	/* They could have slipped one in as we were doing that: make
2890	 * sure it's written, then check again. */
2891	smp_mb();
2892
2893	r = vhost_get_avail_idx(vq);
2894	/* Note: we treat error as empty here */
2895	if (unlikely(r < 0))
2896		return false;
 
 
2897
2898	return r;
2899}
2900EXPORT_SYMBOL_GPL(vhost_enable_notify);
2901
2902/* We don't need to be notified again. */
2903void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2904{
2905	int r;
2906
2907	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2908		return;
2909	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2910	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2911		r = vhost_update_used_flags(vq);
2912		if (r)
2913			vq_err(vq, "Failed to disable notification at %p: %d\n",
2914			       &vq->used->flags, r);
2915	}
2916}
2917EXPORT_SYMBOL_GPL(vhost_disable_notify);
2918
2919/* Create a new message. */
2920struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2921{
2922	/* Make sure all padding within the structure is initialized. */
2923	struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2924	if (!node)
2925		return NULL;
2926
 
 
2927	node->vq = vq;
2928	node->msg.type = type;
2929	return node;
2930}
2931EXPORT_SYMBOL_GPL(vhost_new_msg);
2932
2933void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2934		       struct vhost_msg_node *node)
2935{
2936	spin_lock(&dev->iotlb_lock);
2937	list_add_tail(&node->node, head);
2938	spin_unlock(&dev->iotlb_lock);
2939
2940	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2941}
2942EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2943
2944struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2945					 struct list_head *head)
2946{
2947	struct vhost_msg_node *node = NULL;
2948
2949	spin_lock(&dev->iotlb_lock);
2950	if (!list_empty(head)) {
2951		node = list_first_entry(head, struct vhost_msg_node,
2952					node);
2953		list_del(&node->node);
2954	}
2955	spin_unlock(&dev->iotlb_lock);
2956
2957	return node;
2958}
2959EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2960
2961void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2962{
2963	struct vhost_virtqueue *vq;
2964	int i;
2965
2966	mutex_lock(&dev->mutex);
2967	for (i = 0; i < dev->nvqs; ++i) {
2968		vq = dev->vqs[i];
2969		mutex_lock(&vq->mutex);
2970		vq->acked_backend_features = features;
2971		mutex_unlock(&vq->mutex);
2972	}
2973	mutex_unlock(&dev->mutex);
2974}
2975EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2976
2977static int __init vhost_init(void)
2978{
2979	return 0;
2980}
2981
2982static void __exit vhost_exit(void)
2983{
2984}
2985
2986module_init(vhost_init);
2987module_exit(vhost_exit);
2988
2989MODULE_VERSION("0.0.1");
2990MODULE_LICENSE("GPL v2");
2991MODULE_AUTHOR("Michael S. Tsirkin");
2992MODULE_DESCRIPTION("Host kernel accelerator for virtio");
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (C) 2009 Red Hat, Inc.
   3 * Copyright (C) 2006 Rusty Russell IBM Corporation
   4 *
   5 * Author: Michael S. Tsirkin <mst@redhat.com>
   6 *
   7 * Inspiration, some code, and most witty comments come from
   8 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9 *
  10 * Generic code for virtio server in host kernel.
  11 */
  12
  13#include <linux/eventfd.h>
  14#include <linux/vhost.h>
  15#include <linux/uio.h>
  16#include <linux/mm.h>
  17#include <linux/miscdevice.h>
  18#include <linux/mutex.h>
  19#include <linux/poll.h>
  20#include <linux/file.h>
  21#include <linux/highmem.h>
  22#include <linux/slab.h>
  23#include <linux/vmalloc.h>
  24#include <linux/kthread.h>
  25#include <linux/cgroup.h>
  26#include <linux/module.h>
  27#include <linux/sort.h>
  28#include <linux/sched/mm.h>
  29#include <linux/sched/signal.h>
 
  30#include <linux/interval_tree_generic.h>
  31#include <linux/nospec.h>
  32#include <linux/kcov.h>
  33
  34#include "vhost.h"
  35
  36static ushort max_mem_regions = 64;
  37module_param(max_mem_regions, ushort, 0444);
  38MODULE_PARM_DESC(max_mem_regions,
  39	"Maximum number of memory regions in memory map. (default: 64)");
  40static int max_iotlb_entries = 2048;
  41module_param(max_iotlb_entries, int, 0444);
  42MODULE_PARM_DESC(max_iotlb_entries,
  43	"Maximum number of iotlb entries. (default: 2048)");
  44
  45enum {
  46	VHOST_MEMORY_F_LOG = 0x1,
  47};
  48
  49#define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50#define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52#ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  53static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  54{
  55	vq->user_be = !virtio_legacy_is_little_endian();
  56}
  57
  58static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  59{
  60	vq->user_be = true;
  61}
  62
  63static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  64{
  65	vq->user_be = false;
  66}
  67
  68static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  69{
  70	struct vhost_vring_state s;
  71
  72	if (vq->private_data)
  73		return -EBUSY;
  74
  75	if (copy_from_user(&s, argp, sizeof(s)))
  76		return -EFAULT;
  77
  78	if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  79	    s.num != VHOST_VRING_BIG_ENDIAN)
  80		return -EINVAL;
  81
  82	if (s.num == VHOST_VRING_BIG_ENDIAN)
  83		vhost_enable_cross_endian_big(vq);
  84	else
  85		vhost_enable_cross_endian_little(vq);
  86
  87	return 0;
  88}
  89
  90static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  91				   int __user *argp)
  92{
  93	struct vhost_vring_state s = {
  94		.index = idx,
  95		.num = vq->user_be
  96	};
  97
  98	if (copy_to_user(argp, &s, sizeof(s)))
  99		return -EFAULT;
 100
 101	return 0;
 102}
 103
 104static void vhost_init_is_le(struct vhost_virtqueue *vq)
 105{
 106	/* Note for legacy virtio: user_be is initialized at reset time
 107	 * according to the host endianness. If userspace does not set an
 108	 * explicit endianness, the default behavior is native endian, as
 109	 * expected by legacy virtio.
 110	 */
 111	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 112}
 113#else
 114static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 115{
 116}
 117
 118static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 119{
 120	return -ENOIOCTLCMD;
 121}
 122
 123static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 124				   int __user *argp)
 125{
 126	return -ENOIOCTLCMD;
 127}
 128
 129static void vhost_init_is_le(struct vhost_virtqueue *vq)
 130{
 131	vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 132		|| virtio_legacy_is_little_endian();
 133}
 134#endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 135
 136static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 137{
 138	vhost_init_is_le(vq);
 139}
 140
 141struct vhost_flush_struct {
 142	struct vhost_work work;
 143	struct completion wait_event;
 144};
 145
 146static void vhost_flush_work(struct vhost_work *work)
 147{
 148	struct vhost_flush_struct *s;
 149
 150	s = container_of(work, struct vhost_flush_struct, work);
 151	complete(&s->wait_event);
 152}
 153
 154static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 155			    poll_table *pt)
 156{
 157	struct vhost_poll *poll;
 158
 159	poll = container_of(pt, struct vhost_poll, table);
 160	poll->wqh = wqh;
 161	add_wait_queue(wqh, &poll->wait);
 162}
 163
 164static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 165			     void *key)
 166{
 167	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 168	struct vhost_work *work = &poll->work;
 169
 170	if (!(key_to_poll(key) & poll->mask))
 171		return 0;
 172
 173	if (!poll->dev->use_worker)
 174		work->fn(work);
 175	else
 176		vhost_poll_queue(poll);
 177
 178	return 0;
 179}
 180
 181void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182{
 183	clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184	work->fn = fn;
 185}
 186EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188/* Init poll structure */
 189void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190		     __poll_t mask, struct vhost_dev *dev)
 
 191{
 192	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193	init_poll_funcptr(&poll->table, vhost_poll_func);
 194	poll->mask = mask;
 195	poll->dev = dev;
 196	poll->wqh = NULL;
 
 197
 198	vhost_work_init(&poll->work, fn);
 199}
 200EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202/* Start polling a file. We add ourselves to file's wait queue. The caller must
 203 * keep a reference to a file until after vhost_poll_stop is called. */
 204int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205{
 206	__poll_t mask;
 207
 208	if (poll->wqh)
 209		return 0;
 210
 211	mask = vfs_poll(file, &poll->table);
 212	if (mask)
 213		vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214	if (mask & EPOLLERR) {
 215		vhost_poll_stop(poll);
 216		return -EINVAL;
 217	}
 218
 219	return 0;
 220}
 221EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223/* Stop polling a file. After this function returns, it becomes safe to drop the
 224 * file reference. You must also flush afterwards. */
 225void vhost_poll_stop(struct vhost_poll *poll)
 226{
 227	if (poll->wqh) {
 228		remove_wait_queue(poll->wqh, &poll->wait);
 229		poll->wqh = NULL;
 230	}
 231}
 232EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234void vhost_dev_flush(struct vhost_dev *dev)
 
 235{
 236	struct vhost_flush_struct flush;
 
 
 
 
 
 
 
 
 237
 238	if (dev->worker) {
 239		init_completion(&flush.wait_event);
 240		vhost_work_init(&flush.work, vhost_flush_work);
 
 241
 242		vhost_work_queue(dev, &flush.work);
 243		wait_for_completion(&flush.wait_event);
 
 
 
 244	}
 
 
 
 245}
 246EXPORT_SYMBOL_GPL(vhost_dev_flush);
 247
 248void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 
 
 
 
 
 
 249{
 250	if (!dev->worker)
 
 
 251		return;
 252
 253	if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 254		/* We can only add the work to the list after we're
 255		 * sure it was not in the list.
 256		 * test_and_set_bit() implies a memory barrier.
 257		 */
 258		llist_add(&work->node, &dev->work_list);
 259		wake_up_process(dev->worker);
 260	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 261}
 262EXPORT_SYMBOL_GPL(vhost_work_queue);
 263
 264/* A lockless hint for busy polling code to exit the loop */
 265bool vhost_has_work(struct vhost_dev *dev)
 266{
 267	return !llist_empty(&dev->work_list);
 
 
 
 
 
 
 
 
 
 268}
 269EXPORT_SYMBOL_GPL(vhost_has_work);
 270
 271void vhost_poll_queue(struct vhost_poll *poll)
 272{
 273	vhost_work_queue(poll->dev, &poll->work);
 274}
 275EXPORT_SYMBOL_GPL(vhost_poll_queue);
 276
 277static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 278{
 279	int j;
 280
 281	for (j = 0; j < VHOST_NUM_ADDRS; j++)
 282		vq->meta_iotlb[j] = NULL;
 283}
 284
 285static void vhost_vq_meta_reset(struct vhost_dev *d)
 286{
 287	int i;
 288
 289	for (i = 0; i < d->nvqs; ++i)
 290		__vhost_vq_meta_reset(d->vqs[i]);
 291}
 292
 293static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 294{
 295	call_ctx->ctx = NULL;
 296	memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 297}
 298
 299bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 300{
 301	return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 302}
 303EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 304
 305static void vhost_vq_reset(struct vhost_dev *dev,
 306			   struct vhost_virtqueue *vq)
 307{
 308	vq->num = 1;
 309	vq->desc = NULL;
 310	vq->avail = NULL;
 311	vq->used = NULL;
 312	vq->last_avail_idx = 0;
 313	vq->avail_idx = 0;
 314	vq->last_used_idx = 0;
 315	vq->signalled_used = 0;
 316	vq->signalled_used_valid = false;
 317	vq->used_flags = 0;
 318	vq->log_used = false;
 319	vq->log_addr = -1ull;
 320	vq->private_data = NULL;
 321	vq->acked_features = 0;
 322	vq->acked_backend_features = 0;
 323	vq->log_base = NULL;
 324	vq->error_ctx = NULL;
 325	vq->kick = NULL;
 326	vq->log_ctx = NULL;
 327	vhost_disable_cross_endian(vq);
 328	vhost_reset_is_le(vq);
 329	vq->busyloop_timeout = 0;
 330	vq->umem = NULL;
 331	vq->iotlb = NULL;
 
 332	vhost_vring_call_reset(&vq->call_ctx);
 333	__vhost_vq_meta_reset(vq);
 334}
 335
 336static int vhost_worker(void *data)
 337{
 338	struct vhost_dev *dev = data;
 339	struct vhost_work *work, *work_next;
 340	struct llist_node *node;
 341
 342	kthread_use_mm(dev->mm);
 343
 344	for (;;) {
 345		/* mb paired w/ kthread_stop */
 346		set_current_state(TASK_INTERRUPTIBLE);
 347
 348		if (kthread_should_stop()) {
 349			__set_current_state(TASK_RUNNING);
 350			break;
 351		}
 352
 353		node = llist_del_all(&dev->work_list);
 354		if (!node)
 355			schedule();
 356
 357		node = llist_reverse_order(node);
 358		/* make sure flag is seen after deletion */
 359		smp_wmb();
 360		llist_for_each_entry_safe(work, work_next, node, node) {
 361			clear_bit(VHOST_WORK_QUEUED, &work->flags);
 362			__set_current_state(TASK_RUNNING);
 363			kcov_remote_start_common(dev->kcov_handle);
 364			work->fn(work);
 365			kcov_remote_stop();
 366			if (need_resched())
 367				schedule();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 368		}
 
 369	}
 370	kthread_unuse_mm(dev->mm);
 371	return 0;
 
 
 
 
 
 
 
 
 372}
 373
 374static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 375{
 376	kfree(vq->indirect);
 377	vq->indirect = NULL;
 378	kfree(vq->log);
 379	vq->log = NULL;
 380	kfree(vq->heads);
 381	vq->heads = NULL;
 382}
 383
 384/* Helper to allocate iovec buffers for all vqs. */
 385static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 386{
 387	struct vhost_virtqueue *vq;
 388	int i;
 389
 390	for (i = 0; i < dev->nvqs; ++i) {
 391		vq = dev->vqs[i];
 392		vq->indirect = kmalloc_array(UIO_MAXIOV,
 393					     sizeof(*vq->indirect),
 394					     GFP_KERNEL);
 395		vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 396					GFP_KERNEL);
 397		vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 398					  GFP_KERNEL);
 399		if (!vq->indirect || !vq->log || !vq->heads)
 400			goto err_nomem;
 401	}
 402	return 0;
 403
 404err_nomem:
 405	for (; i >= 0; --i)
 406		vhost_vq_free_iovecs(dev->vqs[i]);
 407	return -ENOMEM;
 408}
 409
 410static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 411{
 412	int i;
 413
 414	for (i = 0; i < dev->nvqs; ++i)
 415		vhost_vq_free_iovecs(dev->vqs[i]);
 416}
 417
 418bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 419			  int pkts, int total_len)
 420{
 421	struct vhost_dev *dev = vq->dev;
 422
 423	if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 424	    pkts >= dev->weight) {
 425		vhost_poll_queue(&vq->poll);
 426		return true;
 427	}
 428
 429	return false;
 430}
 431EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 432
 433static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 434				   unsigned int num)
 435{
 436	size_t event __maybe_unused =
 437	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 438
 439	return sizeof(*vq->avail) +
 440	       sizeof(*vq->avail->ring) * num + event;
 441}
 442
 443static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 444				  unsigned int num)
 445{
 446	size_t event __maybe_unused =
 447	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 448
 449	return sizeof(*vq->used) +
 450	       sizeof(*vq->used->ring) * num + event;
 451}
 452
 453static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 454				  unsigned int num)
 455{
 456	return sizeof(*vq->desc) * num;
 457}
 458
 459void vhost_dev_init(struct vhost_dev *dev,
 460		    struct vhost_virtqueue **vqs, int nvqs,
 461		    int iov_limit, int weight, int byte_weight,
 462		    bool use_worker,
 463		    int (*msg_handler)(struct vhost_dev *dev, u32 asid,
 464				       struct vhost_iotlb_msg *msg))
 465{
 466	struct vhost_virtqueue *vq;
 467	int i;
 468
 469	dev->vqs = vqs;
 470	dev->nvqs = nvqs;
 471	mutex_init(&dev->mutex);
 472	dev->log_ctx = NULL;
 473	dev->umem = NULL;
 474	dev->iotlb = NULL;
 475	dev->mm = NULL;
 476	dev->worker = NULL;
 477	dev->iov_limit = iov_limit;
 478	dev->weight = weight;
 479	dev->byte_weight = byte_weight;
 480	dev->use_worker = use_worker;
 481	dev->msg_handler = msg_handler;
 482	init_llist_head(&dev->work_list);
 483	init_waitqueue_head(&dev->wait);
 484	INIT_LIST_HEAD(&dev->read_list);
 485	INIT_LIST_HEAD(&dev->pending_list);
 486	spin_lock_init(&dev->iotlb_lock);
 487
 488
 489	for (i = 0; i < dev->nvqs; ++i) {
 490		vq = dev->vqs[i];
 491		vq->log = NULL;
 492		vq->indirect = NULL;
 493		vq->heads = NULL;
 494		vq->dev = dev;
 495		mutex_init(&vq->mutex);
 496		vhost_vq_reset(dev, vq);
 497		if (vq->handle_kick)
 498			vhost_poll_init(&vq->poll, vq->handle_kick,
 499					EPOLLIN, dev);
 500	}
 501}
 502EXPORT_SYMBOL_GPL(vhost_dev_init);
 503
 504/* Caller should have device mutex */
 505long vhost_dev_check_owner(struct vhost_dev *dev)
 506{
 507	/* Are you the owner? If not, I don't think you mean to do that */
 508	return dev->mm == current->mm ? 0 : -EPERM;
 509}
 510EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 511
 512struct vhost_attach_cgroups_struct {
 513	struct vhost_work work;
 514	struct task_struct *owner;
 515	int ret;
 516};
 517
 518static void vhost_attach_cgroups_work(struct vhost_work *work)
 519{
 520	struct vhost_attach_cgroups_struct *s;
 521
 522	s = container_of(work, struct vhost_attach_cgroups_struct, work);
 523	s->ret = cgroup_attach_task_all(s->owner, current);
 524}
 525
 526static int vhost_attach_cgroups(struct vhost_dev *dev)
 527{
 528	struct vhost_attach_cgroups_struct attach;
 529
 530	attach.owner = current;
 531	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 532	vhost_work_queue(dev, &attach.work);
 533	vhost_dev_flush(dev);
 534	return attach.ret;
 535}
 536
 537/* Caller should have device mutex */
 538bool vhost_dev_has_owner(struct vhost_dev *dev)
 539{
 540	return dev->mm;
 541}
 542EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 543
 544static void vhost_attach_mm(struct vhost_dev *dev)
 545{
 546	/* No owner, become one */
 547	if (dev->use_worker) {
 548		dev->mm = get_task_mm(current);
 549	} else {
 550		/* vDPA device does not use worker thead, so there's
 551		 * no need to hold the address space for mm. This help
 552		 * to avoid deadlock in the case of mmap() which may
 553		 * held the refcnt of the file and depends on release
 554		 * method to remove vma.
 555		 */
 556		dev->mm = current->mm;
 557		mmgrab(dev->mm);
 558	}
 559}
 560
 561static void vhost_detach_mm(struct vhost_dev *dev)
 562{
 563	if (!dev->mm)
 564		return;
 565
 566	if (dev->use_worker)
 567		mmput(dev->mm);
 568	else
 569		mmdrop(dev->mm);
 570
 571	dev->mm = NULL;
 572}
 573
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 574/* Caller should have device mutex */
 575long vhost_dev_set_owner(struct vhost_dev *dev)
 576{
 577	struct task_struct *worker;
 578	int err;
 579
 580	/* Is there an owner already? */
 581	if (vhost_dev_has_owner(dev)) {
 582		err = -EBUSY;
 583		goto err_mm;
 584	}
 585
 586	vhost_attach_mm(dev);
 587
 588	dev->kcov_handle = kcov_common_handle();
 
 
 
 589	if (dev->use_worker) {
 590		worker = kthread_create(vhost_worker, dev,
 591					"vhost-%d", current->pid);
 592		if (IS_ERR(worker)) {
 593			err = PTR_ERR(worker);
 
 
 
 
 
 594			goto err_worker;
 595		}
 596
 597		dev->worker = worker;
 598		wake_up_process(worker); /* avoid contributing to loadavg */
 599
 600		err = vhost_attach_cgroups(dev);
 601		if (err)
 602			goto err_cgroup;
 603	}
 604
 605	err = vhost_dev_alloc_iovecs(dev);
 606	if (err)
 607		goto err_cgroup;
 608
 609	return 0;
 610err_cgroup:
 611	if (dev->worker) {
 612		kthread_stop(dev->worker);
 613		dev->worker = NULL;
 614	}
 615err_worker:
 
 
 616	vhost_detach_mm(dev);
 617	dev->kcov_handle = 0;
 618err_mm:
 619	return err;
 620}
 621EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 622
 623static struct vhost_iotlb *iotlb_alloc(void)
 624{
 625	return vhost_iotlb_alloc(max_iotlb_entries,
 626				 VHOST_IOTLB_FLAG_RETIRE);
 627}
 628
 629struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 630{
 631	return iotlb_alloc();
 632}
 633EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 634
 635/* Caller should have device mutex */
 636void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 637{
 638	int i;
 639
 640	vhost_dev_cleanup(dev);
 641
 642	dev->umem = umem;
 643	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
 644	 * VQs aren't running.
 645	 */
 646	for (i = 0; i < dev->nvqs; ++i)
 647		dev->vqs[i]->umem = umem;
 648}
 649EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 650
 651void vhost_dev_stop(struct vhost_dev *dev)
 652{
 653	int i;
 654
 655	for (i = 0; i < dev->nvqs; ++i) {
 656		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
 657			vhost_poll_stop(&dev->vqs[i]->poll);
 658	}
 659
 660	vhost_dev_flush(dev);
 661}
 662EXPORT_SYMBOL_GPL(vhost_dev_stop);
 663
 664void vhost_clear_msg(struct vhost_dev *dev)
 665{
 666	struct vhost_msg_node *node, *n;
 667
 668	spin_lock(&dev->iotlb_lock);
 669
 670	list_for_each_entry_safe(node, n, &dev->read_list, node) {
 671		list_del(&node->node);
 672		kfree(node);
 673	}
 674
 675	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 676		list_del(&node->node);
 677		kfree(node);
 678	}
 679
 680	spin_unlock(&dev->iotlb_lock);
 681}
 682EXPORT_SYMBOL_GPL(vhost_clear_msg);
 683
 684void vhost_dev_cleanup(struct vhost_dev *dev)
 685{
 686	int i;
 687
 688	for (i = 0; i < dev->nvqs; ++i) {
 689		if (dev->vqs[i]->error_ctx)
 690			eventfd_ctx_put(dev->vqs[i]->error_ctx);
 691		if (dev->vqs[i]->kick)
 692			fput(dev->vqs[i]->kick);
 693		if (dev->vqs[i]->call_ctx.ctx)
 694			eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 695		vhost_vq_reset(dev, dev->vqs[i]);
 696	}
 697	vhost_dev_free_iovecs(dev);
 698	if (dev->log_ctx)
 699		eventfd_ctx_put(dev->log_ctx);
 700	dev->log_ctx = NULL;
 701	/* No one will access memory at this point */
 702	vhost_iotlb_free(dev->umem);
 703	dev->umem = NULL;
 704	vhost_iotlb_free(dev->iotlb);
 705	dev->iotlb = NULL;
 706	vhost_clear_msg(dev);
 707	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 708	WARN_ON(!llist_empty(&dev->work_list));
 709	if (dev->worker) {
 710		kthread_stop(dev->worker);
 711		dev->worker = NULL;
 712		dev->kcov_handle = 0;
 713	}
 714	vhost_detach_mm(dev);
 715}
 716EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 717
 718static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 719{
 720	u64 a = addr / VHOST_PAGE_SIZE / 8;
 721
 722	/* Make sure 64 bit math will not overflow. */
 723	if (a > ULONG_MAX - (unsigned long)log_base ||
 724	    a + (unsigned long)log_base > ULONG_MAX)
 725		return false;
 726
 727	return access_ok(log_base + a,
 728			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 729}
 730
 731/* Make sure 64 bit math will not overflow. */
 732static bool vhost_overflow(u64 uaddr, u64 size)
 733{
 734	if (uaddr > ULONG_MAX || size > ULONG_MAX)
 735		return true;
 736
 737	if (!size)
 738		return false;
 739
 740	return uaddr > ULONG_MAX - size + 1;
 741}
 742
 743/* Caller should have vq mutex and device mutex. */
 744static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 745				int log_all)
 746{
 747	struct vhost_iotlb_map *map;
 748
 749	if (!umem)
 750		return false;
 751
 752	list_for_each_entry(map, &umem->list, link) {
 753		unsigned long a = map->addr;
 754
 755		if (vhost_overflow(map->addr, map->size))
 756			return false;
 757
 758
 759		if (!access_ok((void __user *)a, map->size))
 760			return false;
 761		else if (log_all && !log_access_ok(log_base,
 762						   map->start,
 763						   map->size))
 764			return false;
 765	}
 766	return true;
 767}
 768
 769static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 770					       u64 addr, unsigned int size,
 771					       int type)
 772{
 773	const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 774
 775	if (!map)
 776		return NULL;
 777
 778	return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 779}
 780
 781/* Can we switch to this memory table? */
 782/* Caller should have device mutex but not vq mutex */
 783static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 784			     int log_all)
 785{
 786	int i;
 787
 788	for (i = 0; i < d->nvqs; ++i) {
 789		bool ok;
 790		bool log;
 791
 792		mutex_lock(&d->vqs[i]->mutex);
 793		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 794		/* If ring is inactive, will check when it's enabled. */
 795		if (d->vqs[i]->private_data)
 796			ok = vq_memory_access_ok(d->vqs[i]->log_base,
 797						 umem, log);
 798		else
 799			ok = true;
 800		mutex_unlock(&d->vqs[i]->mutex);
 801		if (!ok)
 802			return false;
 803	}
 804	return true;
 805}
 806
 807static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 808			  struct iovec iov[], int iov_size, int access);
 809
 810static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 811			      const void *from, unsigned size)
 812{
 813	int ret;
 814
 815	if (!vq->iotlb)
 816		return __copy_to_user(to, from, size);
 817	else {
 818		/* This function should be called after iotlb
 819		 * prefetch, which means we're sure that all vq
 820		 * could be access through iotlb. So -EAGAIN should
 821		 * not happen in this case.
 822		 */
 823		struct iov_iter t;
 824		void __user *uaddr = vhost_vq_meta_fetch(vq,
 825				     (u64)(uintptr_t)to, size,
 826				     VHOST_ADDR_USED);
 827
 828		if (uaddr)
 829			return __copy_to_user(uaddr, from, size);
 830
 831		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 832				     ARRAY_SIZE(vq->iotlb_iov),
 833				     VHOST_ACCESS_WO);
 834		if (ret < 0)
 835			goto out;
 836		iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
 837		ret = copy_to_iter(from, size, &t);
 838		if (ret == size)
 839			ret = 0;
 840	}
 841out:
 842	return ret;
 843}
 844
 845static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 846				void __user *from, unsigned size)
 847{
 848	int ret;
 849
 850	if (!vq->iotlb)
 851		return __copy_from_user(to, from, size);
 852	else {
 853		/* This function should be called after iotlb
 854		 * prefetch, which means we're sure that vq
 855		 * could be access through iotlb. So -EAGAIN should
 856		 * not happen in this case.
 857		 */
 858		void __user *uaddr = vhost_vq_meta_fetch(vq,
 859				     (u64)(uintptr_t)from, size,
 860				     VHOST_ADDR_DESC);
 861		struct iov_iter f;
 862
 863		if (uaddr)
 864			return __copy_from_user(to, uaddr, size);
 865
 866		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 867				     ARRAY_SIZE(vq->iotlb_iov),
 868				     VHOST_ACCESS_RO);
 869		if (ret < 0) {
 870			vq_err(vq, "IOTLB translation failure: uaddr "
 871			       "%p size 0x%llx\n", from,
 872			       (unsigned long long) size);
 873			goto out;
 874		}
 875		iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
 876		ret = copy_from_iter(to, size, &f);
 877		if (ret == size)
 878			ret = 0;
 879	}
 880
 881out:
 882	return ret;
 883}
 884
 885static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 886					  void __user *addr, unsigned int size,
 887					  int type)
 888{
 889	int ret;
 890
 891	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 892			     ARRAY_SIZE(vq->iotlb_iov),
 893			     VHOST_ACCESS_RO);
 894	if (ret < 0) {
 895		vq_err(vq, "IOTLB translation failure: uaddr "
 896			"%p size 0x%llx\n", addr,
 897			(unsigned long long) size);
 898		return NULL;
 899	}
 900
 901	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 902		vq_err(vq, "Non atomic userspace memory access: uaddr "
 903			"%p size 0x%llx\n", addr,
 904			(unsigned long long) size);
 905		return NULL;
 906	}
 907
 908	return vq->iotlb_iov[0].iov_base;
 909}
 910
 911/* This function should be called after iotlb
 912 * prefetch, which means we're sure that vq
 913 * could be access through iotlb. So -EAGAIN should
 914 * not happen in this case.
 915 */
 916static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 917					    void __user *addr, unsigned int size,
 918					    int type)
 919{
 920	void __user *uaddr = vhost_vq_meta_fetch(vq,
 921			     (u64)(uintptr_t)addr, size, type);
 922	if (uaddr)
 923		return uaddr;
 924
 925	return __vhost_get_user_slow(vq, addr, size, type);
 926}
 927
 928#define vhost_put_user(vq, x, ptr)		\
 929({ \
 930	int ret; \
 931	if (!vq->iotlb) { \
 932		ret = __put_user(x, ptr); \
 933	} else { \
 934		__typeof__(ptr) to = \
 935			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
 936					  sizeof(*ptr), VHOST_ADDR_USED); \
 937		if (to != NULL) \
 938			ret = __put_user(x, to); \
 939		else \
 940			ret = -EFAULT;	\
 941	} \
 942	ret; \
 943})
 944
 945static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 946{
 947	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 948			      vhost_avail_event(vq));
 949}
 950
 951static inline int vhost_put_used(struct vhost_virtqueue *vq,
 952				 struct vring_used_elem *head, int idx,
 953				 int count)
 954{
 955	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 956				  count * sizeof(*head));
 957}
 958
 959static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 960
 961{
 962	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 963			      &vq->used->flags);
 964}
 965
 966static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 967
 968{
 969	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 970			      &vq->used->idx);
 971}
 972
 973#define vhost_get_user(vq, x, ptr, type)		\
 974({ \
 975	int ret; \
 976	if (!vq->iotlb) { \
 977		ret = __get_user(x, ptr); \
 978	} else { \
 979		__typeof__(ptr) from = \
 980			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 981							   sizeof(*ptr), \
 982							   type); \
 983		if (from != NULL) \
 984			ret = __get_user(x, from); \
 985		else \
 986			ret = -EFAULT; \
 987	} \
 988	ret; \
 989})
 990
 991#define vhost_get_avail(vq, x, ptr) \
 992	vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 993
 994#define vhost_get_used(vq, x, ptr) \
 995	vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 996
 997static void vhost_dev_lock_vqs(struct vhost_dev *d)
 998{
 999	int i = 0;
1000	for (i = 0; i < d->nvqs; ++i)
1001		mutex_lock_nested(&d->vqs[i]->mutex, i);
1002}
1003
1004static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1005{
1006	int i = 0;
1007	for (i = 0; i < d->nvqs; ++i)
1008		mutex_unlock(&d->vqs[i]->mutex);
1009}
1010
1011static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1012				      __virtio16 *idx)
1013{
1014	return vhost_get_avail(vq, *idx, &vq->avail->idx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1015}
1016
1017static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1018				       __virtio16 *head, int idx)
1019{
1020	return vhost_get_avail(vq, *head,
1021			       &vq->avail->ring[idx & (vq->num - 1)]);
1022}
1023
1024static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1025					__virtio16 *flags)
1026{
1027	return vhost_get_avail(vq, *flags, &vq->avail->flags);
1028}
1029
1030static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1031				       __virtio16 *event)
1032{
1033	return vhost_get_avail(vq, *event, vhost_used_event(vq));
1034}
1035
1036static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1037				     __virtio16 *idx)
1038{
1039	return vhost_get_used(vq, *idx, &vq->used->idx);
1040}
1041
1042static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1043				 struct vring_desc *desc, int idx)
1044{
1045	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1046}
1047
1048static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1049				  struct vhost_iotlb_msg *msg)
1050{
1051	struct vhost_msg_node *node, *n;
1052
1053	spin_lock(&d->iotlb_lock);
1054
1055	list_for_each_entry_safe(node, n, &d->pending_list, node) {
1056		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1057		if (msg->iova <= vq_msg->iova &&
1058		    msg->iova + msg->size - 1 >= vq_msg->iova &&
1059		    vq_msg->type == VHOST_IOTLB_MISS) {
1060			vhost_poll_queue(&node->vq->poll);
1061			list_del(&node->node);
1062			kfree(node);
1063		}
1064	}
1065
1066	spin_unlock(&d->iotlb_lock);
1067}
1068
1069static bool umem_access_ok(u64 uaddr, u64 size, int access)
1070{
1071	unsigned long a = uaddr;
1072
1073	/* Make sure 64 bit math will not overflow. */
1074	if (vhost_overflow(uaddr, size))
1075		return false;
1076
1077	if ((access & VHOST_ACCESS_RO) &&
1078	    !access_ok((void __user *)a, size))
1079		return false;
1080	if ((access & VHOST_ACCESS_WO) &&
1081	    !access_ok((void __user *)a, size))
1082		return false;
1083	return true;
1084}
1085
1086static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1087				   struct vhost_iotlb_msg *msg)
1088{
1089	int ret = 0;
1090
1091	if (asid != 0)
1092		return -EINVAL;
1093
1094	mutex_lock(&dev->mutex);
1095	vhost_dev_lock_vqs(dev);
1096	switch (msg->type) {
1097	case VHOST_IOTLB_UPDATE:
1098		if (!dev->iotlb) {
1099			ret = -EFAULT;
1100			break;
1101		}
1102		if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1103			ret = -EFAULT;
1104			break;
1105		}
1106		vhost_vq_meta_reset(dev);
1107		if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1108					  msg->iova + msg->size - 1,
1109					  msg->uaddr, msg->perm)) {
1110			ret = -ENOMEM;
1111			break;
1112		}
1113		vhost_iotlb_notify_vq(dev, msg);
1114		break;
1115	case VHOST_IOTLB_INVALIDATE:
1116		if (!dev->iotlb) {
1117			ret = -EFAULT;
1118			break;
1119		}
1120		vhost_vq_meta_reset(dev);
1121		vhost_iotlb_del_range(dev->iotlb, msg->iova,
1122				      msg->iova + msg->size - 1);
1123		break;
1124	default:
1125		ret = -EINVAL;
1126		break;
1127	}
1128
1129	vhost_dev_unlock_vqs(dev);
1130	mutex_unlock(&dev->mutex);
1131
1132	return ret;
1133}
1134ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1135			     struct iov_iter *from)
1136{
1137	struct vhost_iotlb_msg msg;
1138	size_t offset;
1139	int type, ret;
1140	u32 asid = 0;
1141
1142	ret = copy_from_iter(&type, sizeof(type), from);
1143	if (ret != sizeof(type)) {
1144		ret = -EINVAL;
1145		goto done;
1146	}
1147
1148	switch (type) {
1149	case VHOST_IOTLB_MSG:
1150		/* There maybe a hole after type for V1 message type,
1151		 * so skip it here.
1152		 */
1153		offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1154		break;
1155	case VHOST_IOTLB_MSG_V2:
1156		if (vhost_backend_has_feature(dev->vqs[0],
1157					      VHOST_BACKEND_F_IOTLB_ASID)) {
1158			ret = copy_from_iter(&asid, sizeof(asid), from);
1159			if (ret != sizeof(asid)) {
1160				ret = -EINVAL;
1161				goto done;
1162			}
1163			offset = 0;
1164		} else
1165			offset = sizeof(__u32);
1166		break;
1167	default:
1168		ret = -EINVAL;
1169		goto done;
1170	}
1171
1172	iov_iter_advance(from, offset);
1173	ret = copy_from_iter(&msg, sizeof(msg), from);
1174	if (ret != sizeof(msg)) {
1175		ret = -EINVAL;
1176		goto done;
1177	}
1178
1179	if ((msg.type == VHOST_IOTLB_UPDATE ||
1180	     msg.type == VHOST_IOTLB_INVALIDATE) &&
1181	     msg.size == 0) {
1182		ret = -EINVAL;
1183		goto done;
1184	}
1185
1186	if (dev->msg_handler)
1187		ret = dev->msg_handler(dev, asid, &msg);
1188	else
1189		ret = vhost_process_iotlb_msg(dev, asid, &msg);
1190	if (ret) {
1191		ret = -EFAULT;
1192		goto done;
1193	}
1194
1195	ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1196	      sizeof(struct vhost_msg_v2);
1197done:
1198	return ret;
1199}
1200EXPORT_SYMBOL(vhost_chr_write_iter);
1201
1202__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1203			    poll_table *wait)
1204{
1205	__poll_t mask = 0;
1206
1207	poll_wait(file, &dev->wait, wait);
1208
1209	if (!list_empty(&dev->read_list))
1210		mask |= EPOLLIN | EPOLLRDNORM;
1211
1212	return mask;
1213}
1214EXPORT_SYMBOL(vhost_chr_poll);
1215
1216ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1217			    int noblock)
1218{
1219	DEFINE_WAIT(wait);
1220	struct vhost_msg_node *node;
1221	ssize_t ret = 0;
1222	unsigned size = sizeof(struct vhost_msg);
1223
1224	if (iov_iter_count(to) < size)
1225		return 0;
1226
1227	while (1) {
1228		if (!noblock)
1229			prepare_to_wait(&dev->wait, &wait,
1230					TASK_INTERRUPTIBLE);
1231
1232		node = vhost_dequeue_msg(dev, &dev->read_list);
1233		if (node)
1234			break;
1235		if (noblock) {
1236			ret = -EAGAIN;
1237			break;
1238		}
1239		if (signal_pending(current)) {
1240			ret = -ERESTARTSYS;
1241			break;
1242		}
1243		if (!dev->iotlb) {
1244			ret = -EBADFD;
1245			break;
1246		}
1247
1248		schedule();
1249	}
1250
1251	if (!noblock)
1252		finish_wait(&dev->wait, &wait);
1253
1254	if (node) {
1255		struct vhost_iotlb_msg *msg;
1256		void *start = &node->msg;
1257
1258		switch (node->msg.type) {
1259		case VHOST_IOTLB_MSG:
1260			size = sizeof(node->msg);
1261			msg = &node->msg.iotlb;
1262			break;
1263		case VHOST_IOTLB_MSG_V2:
1264			size = sizeof(node->msg_v2);
1265			msg = &node->msg_v2.iotlb;
1266			break;
1267		default:
1268			BUG();
1269			break;
1270		}
1271
1272		ret = copy_to_iter(start, size, to);
1273		if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1274			kfree(node);
1275			return ret;
1276		}
1277		vhost_enqueue_msg(dev, &dev->pending_list, node);
1278	}
1279
1280	return ret;
1281}
1282EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1283
1284static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1285{
1286	struct vhost_dev *dev = vq->dev;
1287	struct vhost_msg_node *node;
1288	struct vhost_iotlb_msg *msg;
1289	bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1290
1291	node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1292	if (!node)
1293		return -ENOMEM;
1294
1295	if (v2) {
1296		node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1297		msg = &node->msg_v2.iotlb;
1298	} else {
1299		msg = &node->msg.iotlb;
1300	}
1301
1302	msg->type = VHOST_IOTLB_MISS;
1303	msg->iova = iova;
1304	msg->perm = access;
1305
1306	vhost_enqueue_msg(dev, &dev->read_list, node);
1307
1308	return 0;
1309}
1310
1311static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1312			 vring_desc_t __user *desc,
1313			 vring_avail_t __user *avail,
1314			 vring_used_t __user *used)
1315
1316{
1317	/* If an IOTLB device is present, the vring addresses are
1318	 * GIOVAs. Access validation occurs at prefetch time. */
1319	if (vq->iotlb)
1320		return true;
1321
1322	return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1323	       access_ok(avail, vhost_get_avail_size(vq, num)) &&
1324	       access_ok(used, vhost_get_used_size(vq, num));
1325}
1326
1327static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1328				 const struct vhost_iotlb_map *map,
1329				 int type)
1330{
1331	int access = (type == VHOST_ADDR_USED) ?
1332		     VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1333
1334	if (likely(map->perm & access))
1335		vq->meta_iotlb[type] = map;
1336}
1337
1338static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1339			    int access, u64 addr, u64 len, int type)
1340{
1341	const struct vhost_iotlb_map *map;
1342	struct vhost_iotlb *umem = vq->iotlb;
1343	u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1344
1345	if (vhost_vq_meta_fetch(vq, addr, len, type))
1346		return true;
1347
1348	while (len > s) {
1349		map = vhost_iotlb_itree_first(umem, addr, last);
1350		if (map == NULL || map->start > addr) {
1351			vhost_iotlb_miss(vq, addr, access);
1352			return false;
1353		} else if (!(map->perm & access)) {
1354			/* Report the possible access violation by
1355			 * request another translation from userspace.
1356			 */
1357			return false;
1358		}
1359
1360		size = map->size - addr + map->start;
1361
1362		if (orig_addr == addr && size >= len)
1363			vhost_vq_meta_update(vq, map, type);
1364
1365		s += size;
1366		addr += size;
1367	}
1368
1369	return true;
1370}
1371
1372int vq_meta_prefetch(struct vhost_virtqueue *vq)
1373{
1374	unsigned int num = vq->num;
1375
1376	if (!vq->iotlb)
1377		return 1;
1378
1379	return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1380			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1381	       iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1382			       vhost_get_avail_size(vq, num),
1383			       VHOST_ADDR_AVAIL) &&
1384	       iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1385			       vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1386}
1387EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1388
1389/* Can we log writes? */
1390/* Caller should have device mutex but not vq mutex */
1391bool vhost_log_access_ok(struct vhost_dev *dev)
1392{
1393	return memory_access_ok(dev, dev->umem, 1);
1394}
1395EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1396
1397static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1398				  void __user *log_base,
1399				  bool log_used,
1400				  u64 log_addr)
1401{
1402	/* If an IOTLB device is present, log_addr is a GIOVA that
1403	 * will never be logged by log_used(). */
1404	if (vq->iotlb)
1405		return true;
1406
1407	return !log_used || log_access_ok(log_base, log_addr,
1408					  vhost_get_used_size(vq, vq->num));
1409}
1410
1411/* Verify access for write logging. */
1412/* Caller should have vq mutex and device mutex */
1413static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1414			     void __user *log_base)
1415{
1416	return vq_memory_access_ok(log_base, vq->umem,
1417				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1418		vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1419}
1420
1421/* Can we start vq? */
1422/* Caller should have vq mutex and device mutex */
1423bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1424{
1425	if (!vq_log_access_ok(vq, vq->log_base))
1426		return false;
1427
1428	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1429}
1430EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1431
1432static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1433{
1434	struct vhost_memory mem, *newmem;
1435	struct vhost_memory_region *region;
1436	struct vhost_iotlb *newumem, *oldumem;
1437	unsigned long size = offsetof(struct vhost_memory, regions);
1438	int i;
1439
1440	if (copy_from_user(&mem, m, size))
1441		return -EFAULT;
1442	if (mem.padding)
1443		return -EOPNOTSUPP;
1444	if (mem.nregions > max_mem_regions)
1445		return -E2BIG;
1446	newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1447			GFP_KERNEL);
1448	if (!newmem)
1449		return -ENOMEM;
1450
1451	memcpy(newmem, &mem, size);
1452	if (copy_from_user(newmem->regions, m->regions,
1453			   flex_array_size(newmem, regions, mem.nregions))) {
1454		kvfree(newmem);
1455		return -EFAULT;
1456	}
1457
1458	newumem = iotlb_alloc();
1459	if (!newumem) {
1460		kvfree(newmem);
1461		return -ENOMEM;
1462	}
1463
1464	for (region = newmem->regions;
1465	     region < newmem->regions + mem.nregions;
1466	     region++) {
1467		if (vhost_iotlb_add_range(newumem,
1468					  region->guest_phys_addr,
1469					  region->guest_phys_addr +
1470					  region->memory_size - 1,
1471					  region->userspace_addr,
1472					  VHOST_MAP_RW))
1473			goto err;
1474	}
1475
1476	if (!memory_access_ok(d, newumem, 0))
1477		goto err;
1478
1479	oldumem = d->umem;
1480	d->umem = newumem;
1481
1482	/* All memory accesses are done under some VQ mutex. */
1483	for (i = 0; i < d->nvqs; ++i) {
1484		mutex_lock(&d->vqs[i]->mutex);
1485		d->vqs[i]->umem = newumem;
1486		mutex_unlock(&d->vqs[i]->mutex);
1487	}
1488
1489	kvfree(newmem);
1490	vhost_iotlb_free(oldumem);
1491	return 0;
1492
1493err:
1494	vhost_iotlb_free(newumem);
1495	kvfree(newmem);
1496	return -EFAULT;
1497}
1498
1499static long vhost_vring_set_num(struct vhost_dev *d,
1500				struct vhost_virtqueue *vq,
1501				void __user *argp)
1502{
1503	struct vhost_vring_state s;
1504
1505	/* Resizing ring with an active backend?
1506	 * You don't want to do that. */
1507	if (vq->private_data)
1508		return -EBUSY;
1509
1510	if (copy_from_user(&s, argp, sizeof s))
1511		return -EFAULT;
1512
1513	if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1514		return -EINVAL;
1515	vq->num = s.num;
1516
1517	return 0;
1518}
1519
1520static long vhost_vring_set_addr(struct vhost_dev *d,
1521				 struct vhost_virtqueue *vq,
1522				 void __user *argp)
1523{
1524	struct vhost_vring_addr a;
1525
1526	if (copy_from_user(&a, argp, sizeof a))
1527		return -EFAULT;
1528	if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1529		return -EOPNOTSUPP;
1530
1531	/* For 32bit, verify that the top 32bits of the user
1532	   data are set to zero. */
1533	if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1534	    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1535	    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1536		return -EFAULT;
1537
1538	/* Make sure it's safe to cast pointers to vring types. */
1539	BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1540	BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1541	if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1542	    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1543	    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1544		return -EINVAL;
1545
1546	/* We only verify access here if backend is configured.
1547	 * If it is not, we don't as size might not have been setup.
1548	 * We will verify when backend is configured. */
1549	if (vq->private_data) {
1550		if (!vq_access_ok(vq, vq->num,
1551			(void __user *)(unsigned long)a.desc_user_addr,
1552			(void __user *)(unsigned long)a.avail_user_addr,
1553			(void __user *)(unsigned long)a.used_user_addr))
1554			return -EINVAL;
1555
1556		/* Also validate log access for used ring if enabled. */
1557		if (!vq_log_used_access_ok(vq, vq->log_base,
1558				a.flags & (0x1 << VHOST_VRING_F_LOG),
1559				a.log_guest_addr))
1560			return -EINVAL;
1561	}
1562
1563	vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1564	vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1565	vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1566	vq->log_addr = a.log_guest_addr;
1567	vq->used = (void __user *)(unsigned long)a.used_user_addr;
1568
1569	return 0;
1570}
1571
1572static long vhost_vring_set_num_addr(struct vhost_dev *d,
1573				     struct vhost_virtqueue *vq,
1574				     unsigned int ioctl,
1575				     void __user *argp)
1576{
1577	long r;
1578
1579	mutex_lock(&vq->mutex);
1580
1581	switch (ioctl) {
1582	case VHOST_SET_VRING_NUM:
1583		r = vhost_vring_set_num(d, vq, argp);
1584		break;
1585	case VHOST_SET_VRING_ADDR:
1586		r = vhost_vring_set_addr(d, vq, argp);
1587		break;
1588	default:
1589		BUG();
1590	}
1591
1592	mutex_unlock(&vq->mutex);
1593
1594	return r;
1595}
1596long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1597{
1598	struct file *eventfp, *filep = NULL;
1599	bool pollstart = false, pollstop = false;
1600	struct eventfd_ctx *ctx = NULL;
1601	u32 __user *idxp = argp;
1602	struct vhost_virtqueue *vq;
1603	struct vhost_vring_state s;
1604	struct vhost_vring_file f;
1605	u32 idx;
1606	long r;
1607
1608	r = get_user(idx, idxp);
1609	if (r < 0)
1610		return r;
1611	if (idx >= d->nvqs)
1612		return -ENOBUFS;
1613
1614	idx = array_index_nospec(idx, d->nvqs);
1615	vq = d->vqs[idx];
1616
1617	if (ioctl == VHOST_SET_VRING_NUM ||
1618	    ioctl == VHOST_SET_VRING_ADDR) {
1619		return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1620	}
1621
1622	mutex_lock(&vq->mutex);
1623
1624	switch (ioctl) {
1625	case VHOST_SET_VRING_BASE:
1626		/* Moving base with an active backend?
1627		 * You don't want to do that. */
1628		if (vq->private_data) {
1629			r = -EBUSY;
1630			break;
1631		}
1632		if (copy_from_user(&s, argp, sizeof s)) {
1633			r = -EFAULT;
1634			break;
1635		}
1636		if (s.num > 0xffff) {
1637			r = -EINVAL;
1638			break;
 
 
 
 
 
 
1639		}
1640		vq->last_avail_idx = s.num;
1641		/* Forget the cached index value. */
1642		vq->avail_idx = vq->last_avail_idx;
1643		break;
1644	case VHOST_GET_VRING_BASE:
1645		s.index = idx;
1646		s.num = vq->last_avail_idx;
 
 
 
1647		if (copy_to_user(argp, &s, sizeof s))
1648			r = -EFAULT;
1649		break;
1650	case VHOST_SET_VRING_KICK:
1651		if (copy_from_user(&f, argp, sizeof f)) {
1652			r = -EFAULT;
1653			break;
1654		}
1655		eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1656		if (IS_ERR(eventfp)) {
1657			r = PTR_ERR(eventfp);
1658			break;
1659		}
1660		if (eventfp != vq->kick) {
1661			pollstop = (filep = vq->kick) != NULL;
1662			pollstart = (vq->kick = eventfp) != NULL;
1663		} else
1664			filep = eventfp;
1665		break;
1666	case VHOST_SET_VRING_CALL:
1667		if (copy_from_user(&f, argp, sizeof f)) {
1668			r = -EFAULT;
1669			break;
1670		}
1671		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1672		if (IS_ERR(ctx)) {
1673			r = PTR_ERR(ctx);
1674			break;
1675		}
1676
1677		swap(ctx, vq->call_ctx.ctx);
1678		break;
1679	case VHOST_SET_VRING_ERR:
1680		if (copy_from_user(&f, argp, sizeof f)) {
1681			r = -EFAULT;
1682			break;
1683		}
1684		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1685		if (IS_ERR(ctx)) {
1686			r = PTR_ERR(ctx);
1687			break;
1688		}
1689		swap(ctx, vq->error_ctx);
1690		break;
1691	case VHOST_SET_VRING_ENDIAN:
1692		r = vhost_set_vring_endian(vq, argp);
1693		break;
1694	case VHOST_GET_VRING_ENDIAN:
1695		r = vhost_get_vring_endian(vq, idx, argp);
1696		break;
1697	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1698		if (copy_from_user(&s, argp, sizeof(s))) {
1699			r = -EFAULT;
1700			break;
1701		}
1702		vq->busyloop_timeout = s.num;
1703		break;
1704	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1705		s.index = idx;
1706		s.num = vq->busyloop_timeout;
1707		if (copy_to_user(argp, &s, sizeof(s)))
1708			r = -EFAULT;
1709		break;
1710	default:
1711		r = -ENOIOCTLCMD;
1712	}
1713
1714	if (pollstop && vq->handle_kick)
1715		vhost_poll_stop(&vq->poll);
1716
1717	if (!IS_ERR_OR_NULL(ctx))
1718		eventfd_ctx_put(ctx);
1719	if (filep)
1720		fput(filep);
1721
1722	if (pollstart && vq->handle_kick)
1723		r = vhost_poll_start(&vq->poll, vq->kick);
1724
1725	mutex_unlock(&vq->mutex);
1726
1727	if (pollstop && vq->handle_kick)
1728		vhost_dev_flush(vq->poll.dev);
1729	return r;
1730}
1731EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1732
1733int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1734{
1735	struct vhost_iotlb *niotlb, *oiotlb;
1736	int i;
1737
1738	niotlb = iotlb_alloc();
1739	if (!niotlb)
1740		return -ENOMEM;
1741
1742	oiotlb = d->iotlb;
1743	d->iotlb = niotlb;
1744
1745	for (i = 0; i < d->nvqs; ++i) {
1746		struct vhost_virtqueue *vq = d->vqs[i];
1747
1748		mutex_lock(&vq->mutex);
1749		vq->iotlb = niotlb;
1750		__vhost_vq_meta_reset(vq);
1751		mutex_unlock(&vq->mutex);
1752	}
1753
1754	vhost_iotlb_free(oiotlb);
1755
1756	return 0;
1757}
1758EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1759
1760/* Caller must have device mutex */
1761long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1762{
1763	struct eventfd_ctx *ctx;
1764	u64 p;
1765	long r;
1766	int i, fd;
1767
1768	/* If you are not the owner, you can become one */
1769	if (ioctl == VHOST_SET_OWNER) {
1770		r = vhost_dev_set_owner(d);
1771		goto done;
1772	}
1773
1774	/* You must be the owner to do anything else */
1775	r = vhost_dev_check_owner(d);
1776	if (r)
1777		goto done;
1778
1779	switch (ioctl) {
1780	case VHOST_SET_MEM_TABLE:
1781		r = vhost_set_memory(d, argp);
1782		break;
1783	case VHOST_SET_LOG_BASE:
1784		if (copy_from_user(&p, argp, sizeof p)) {
1785			r = -EFAULT;
1786			break;
1787		}
1788		if ((u64)(unsigned long)p != p) {
1789			r = -EFAULT;
1790			break;
1791		}
1792		for (i = 0; i < d->nvqs; ++i) {
1793			struct vhost_virtqueue *vq;
1794			void __user *base = (void __user *)(unsigned long)p;
1795			vq = d->vqs[i];
1796			mutex_lock(&vq->mutex);
1797			/* If ring is inactive, will check when it's enabled. */
1798			if (vq->private_data && !vq_log_access_ok(vq, base))
1799				r = -EFAULT;
1800			else
1801				vq->log_base = base;
1802			mutex_unlock(&vq->mutex);
1803		}
1804		break;
1805	case VHOST_SET_LOG_FD:
1806		r = get_user(fd, (int __user *)argp);
1807		if (r < 0)
1808			break;
1809		ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1810		if (IS_ERR(ctx)) {
1811			r = PTR_ERR(ctx);
1812			break;
1813		}
1814		swap(ctx, d->log_ctx);
1815		for (i = 0; i < d->nvqs; ++i) {
1816			mutex_lock(&d->vqs[i]->mutex);
1817			d->vqs[i]->log_ctx = d->log_ctx;
1818			mutex_unlock(&d->vqs[i]->mutex);
1819		}
1820		if (ctx)
1821			eventfd_ctx_put(ctx);
1822		break;
1823	default:
1824		r = -ENOIOCTLCMD;
1825		break;
1826	}
1827done:
1828	return r;
1829}
1830EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1831
1832/* TODO: This is really inefficient.  We need something like get_user()
1833 * (instruction directly accesses the data, with an exception table entry
1834 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1835 */
1836static int set_bit_to_user(int nr, void __user *addr)
1837{
1838	unsigned long log = (unsigned long)addr;
1839	struct page *page;
1840	void *base;
1841	int bit = nr + (log % PAGE_SIZE) * 8;
1842	int r;
1843
1844	r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1845	if (r < 0)
1846		return r;
1847	BUG_ON(r != 1);
1848	base = kmap_atomic(page);
1849	set_bit(bit, base);
1850	kunmap_atomic(base);
1851	unpin_user_pages_dirty_lock(&page, 1, true);
1852	return 0;
1853}
1854
1855static int log_write(void __user *log_base,
1856		     u64 write_address, u64 write_length)
1857{
1858	u64 write_page = write_address / VHOST_PAGE_SIZE;
1859	int r;
1860
1861	if (!write_length)
1862		return 0;
1863	write_length += write_address % VHOST_PAGE_SIZE;
1864	for (;;) {
1865		u64 base = (u64)(unsigned long)log_base;
1866		u64 log = base + write_page / 8;
1867		int bit = write_page % 8;
1868		if ((u64)(unsigned long)log != log)
1869			return -EFAULT;
1870		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1871		if (r < 0)
1872			return r;
1873		if (write_length <= VHOST_PAGE_SIZE)
1874			break;
1875		write_length -= VHOST_PAGE_SIZE;
1876		write_page += 1;
1877	}
1878	return r;
1879}
1880
1881static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1882{
1883	struct vhost_iotlb *umem = vq->umem;
1884	struct vhost_iotlb_map *u;
1885	u64 start, end, l, min;
1886	int r;
1887	bool hit = false;
1888
1889	while (len) {
1890		min = len;
1891		/* More than one GPAs can be mapped into a single HVA. So
1892		 * iterate all possible umems here to be safe.
1893		 */
1894		list_for_each_entry(u, &umem->list, link) {
1895			if (u->addr > hva - 1 + len ||
1896			    u->addr - 1 + u->size < hva)
1897				continue;
1898			start = max(u->addr, hva);
1899			end = min(u->addr - 1 + u->size, hva - 1 + len);
1900			l = end - start + 1;
1901			r = log_write(vq->log_base,
1902				      u->start + start - u->addr,
1903				      l);
1904			if (r < 0)
1905				return r;
1906			hit = true;
1907			min = min(l, min);
1908		}
1909
1910		if (!hit)
1911			return -EFAULT;
1912
1913		len -= min;
1914		hva += min;
1915	}
1916
1917	return 0;
1918}
1919
1920static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1921{
1922	struct iovec *iov = vq->log_iov;
1923	int i, ret;
1924
1925	if (!vq->iotlb)
1926		return log_write(vq->log_base, vq->log_addr + used_offset, len);
1927
1928	ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1929			     len, iov, 64, VHOST_ACCESS_WO);
1930	if (ret < 0)
1931		return ret;
1932
1933	for (i = 0; i < ret; i++) {
1934		ret = log_write_hva(vq,	(uintptr_t)iov[i].iov_base,
1935				    iov[i].iov_len);
1936		if (ret)
1937			return ret;
1938	}
1939
1940	return 0;
1941}
1942
1943int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1944		    unsigned int log_num, u64 len, struct iovec *iov, int count)
1945{
1946	int i, r;
1947
1948	/* Make sure data written is seen before log. */
1949	smp_wmb();
1950
1951	if (vq->iotlb) {
1952		for (i = 0; i < count; i++) {
1953			r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1954					  iov[i].iov_len);
1955			if (r < 0)
1956				return r;
1957		}
1958		return 0;
1959	}
1960
1961	for (i = 0; i < log_num; ++i) {
1962		u64 l = min(log[i].len, len);
1963		r = log_write(vq->log_base, log[i].addr, l);
1964		if (r < 0)
1965			return r;
1966		len -= l;
1967		if (!len) {
1968			if (vq->log_ctx)
1969				eventfd_signal(vq->log_ctx, 1);
1970			return 0;
1971		}
1972	}
1973	/* Length written exceeds what we have stored. This is a bug. */
1974	BUG();
1975	return 0;
1976}
1977EXPORT_SYMBOL_GPL(vhost_log_write);
1978
1979static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1980{
1981	void __user *used;
1982	if (vhost_put_used_flags(vq))
1983		return -EFAULT;
1984	if (unlikely(vq->log_used)) {
1985		/* Make sure the flag is seen before log. */
1986		smp_wmb();
1987		/* Log used flag write. */
1988		used = &vq->used->flags;
1989		log_used(vq, (used - (void __user *)vq->used),
1990			 sizeof vq->used->flags);
1991		if (vq->log_ctx)
1992			eventfd_signal(vq->log_ctx, 1);
1993	}
1994	return 0;
1995}
1996
1997static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1998{
1999	if (vhost_put_avail_event(vq))
2000		return -EFAULT;
2001	if (unlikely(vq->log_used)) {
2002		void __user *used;
2003		/* Make sure the event is seen before log. */
2004		smp_wmb();
2005		/* Log avail event write */
2006		used = vhost_avail_event(vq);
2007		log_used(vq, (used - (void __user *)vq->used),
2008			 sizeof *vhost_avail_event(vq));
2009		if (vq->log_ctx)
2010			eventfd_signal(vq->log_ctx, 1);
2011	}
2012	return 0;
2013}
2014
2015int vhost_vq_init_access(struct vhost_virtqueue *vq)
2016{
2017	__virtio16 last_used_idx;
2018	int r;
2019	bool is_le = vq->is_le;
2020
2021	if (!vq->private_data)
2022		return 0;
2023
2024	vhost_init_is_le(vq);
2025
2026	r = vhost_update_used_flags(vq);
2027	if (r)
2028		goto err;
2029	vq->signalled_used_valid = false;
2030	if (!vq->iotlb &&
2031	    !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2032		r = -EFAULT;
2033		goto err;
2034	}
2035	r = vhost_get_used_idx(vq, &last_used_idx);
2036	if (r) {
2037		vq_err(vq, "Can't access used idx at %p\n",
2038		       &vq->used->idx);
2039		goto err;
2040	}
2041	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2042	return 0;
2043
2044err:
2045	vq->is_le = is_le;
2046	return r;
2047}
2048EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2049
2050static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2051			  struct iovec iov[], int iov_size, int access)
2052{
2053	const struct vhost_iotlb_map *map;
2054	struct vhost_dev *dev = vq->dev;
2055	struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2056	struct iovec *_iov;
2057	u64 s = 0, last = addr + len - 1;
2058	int ret = 0;
2059
2060	while ((u64)len > s) {
2061		u64 size;
2062		if (unlikely(ret >= iov_size)) {
2063			ret = -ENOBUFS;
2064			break;
2065		}
2066
2067		map = vhost_iotlb_itree_first(umem, addr, last);
2068		if (map == NULL || map->start > addr) {
2069			if (umem != dev->iotlb) {
2070				ret = -EFAULT;
2071				break;
2072			}
2073			ret = -EAGAIN;
2074			break;
2075		} else if (!(map->perm & access)) {
2076			ret = -EPERM;
2077			break;
2078		}
2079
2080		_iov = iov + ret;
2081		size = map->size - addr + map->start;
2082		_iov->iov_len = min((u64)len - s, size);
2083		_iov->iov_base = (void __user *)(unsigned long)
2084				 (map->addr + addr - map->start);
2085		s += size;
2086		addr += size;
2087		++ret;
2088	}
2089
2090	if (ret == -EAGAIN)
2091		vhost_iotlb_miss(vq, addr, access);
2092	return ret;
2093}
2094
2095/* Each buffer in the virtqueues is actually a chain of descriptors.  This
2096 * function returns the next descriptor in the chain,
2097 * or -1U if we're at the end. */
2098static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2099{
2100	unsigned int next;
2101
2102	/* If this descriptor says it doesn't chain, we're done. */
2103	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2104		return -1U;
2105
2106	/* Check they're not leading us off end of descriptors. */
2107	next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2108	return next;
2109}
2110
2111static int get_indirect(struct vhost_virtqueue *vq,
2112			struct iovec iov[], unsigned int iov_size,
2113			unsigned int *out_num, unsigned int *in_num,
2114			struct vhost_log *log, unsigned int *log_num,
2115			struct vring_desc *indirect)
2116{
2117	struct vring_desc desc;
2118	unsigned int i = 0, count, found = 0;
2119	u32 len = vhost32_to_cpu(vq, indirect->len);
2120	struct iov_iter from;
2121	int ret, access;
2122
2123	/* Sanity check */
2124	if (unlikely(len % sizeof desc)) {
2125		vq_err(vq, "Invalid length in indirect descriptor: "
2126		       "len 0x%llx not multiple of 0x%zx\n",
2127		       (unsigned long long)len,
2128		       sizeof desc);
2129		return -EINVAL;
2130	}
2131
2132	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2133			     UIO_MAXIOV, VHOST_ACCESS_RO);
2134	if (unlikely(ret < 0)) {
2135		if (ret != -EAGAIN)
2136			vq_err(vq, "Translation failure %d in indirect.\n", ret);
2137		return ret;
2138	}
2139	iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2140	count = len / sizeof desc;
2141	/* Buffers are chained via a 16 bit next field, so
2142	 * we can have at most 2^16 of these. */
2143	if (unlikely(count > USHRT_MAX + 1)) {
2144		vq_err(vq, "Indirect buffer length too big: %d\n",
2145		       indirect->len);
2146		return -E2BIG;
2147	}
2148
2149	do {
2150		unsigned iov_count = *in_num + *out_num;
2151		if (unlikely(++found > count)) {
2152			vq_err(vq, "Loop detected: last one at %u "
2153			       "indirect size %u\n",
2154			       i, count);
2155			return -EINVAL;
2156		}
2157		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2158			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2159			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2160			return -EINVAL;
2161		}
2162		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2163			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2164			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2165			return -EINVAL;
2166		}
2167
2168		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2169			access = VHOST_ACCESS_WO;
2170		else
2171			access = VHOST_ACCESS_RO;
2172
2173		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2174				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2175				     iov_size - iov_count, access);
2176		if (unlikely(ret < 0)) {
2177			if (ret != -EAGAIN)
2178				vq_err(vq, "Translation failure %d indirect idx %d\n",
2179					ret, i);
2180			return ret;
2181		}
2182		/* If this is an input descriptor, increment that count. */
2183		if (access == VHOST_ACCESS_WO) {
2184			*in_num += ret;
2185			if (unlikely(log && ret)) {
2186				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2187				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2188				++*log_num;
2189			}
2190		} else {
2191			/* If it's an output descriptor, they're all supposed
2192			 * to come before any input descriptors. */
2193			if (unlikely(*in_num)) {
2194				vq_err(vq, "Indirect descriptor "
2195				       "has out after in: idx %d\n", i);
2196				return -EINVAL;
2197			}
2198			*out_num += ret;
2199		}
2200	} while ((i = next_desc(vq, &desc)) != -1);
2201	return 0;
2202}
2203
2204/* This looks in the virtqueue and for the first available buffer, and converts
2205 * it to an iovec for convenient access.  Since descriptors consist of some
2206 * number of output then some number of input descriptors, it's actually two
2207 * iovecs, but we pack them into one and note how many of each there were.
2208 *
2209 * This function returns the descriptor number found, or vq->num (which is
2210 * never a valid descriptor number) if none was found.  A negative code is
2211 * returned on error. */
2212int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2213		      struct iovec iov[], unsigned int iov_size,
2214		      unsigned int *out_num, unsigned int *in_num,
2215		      struct vhost_log *log, unsigned int *log_num)
2216{
2217	struct vring_desc desc;
2218	unsigned int i, head, found = 0;
2219	u16 last_avail_idx;
2220	__virtio16 avail_idx;
2221	__virtio16 ring_head;
2222	int ret, access;
2223
2224	/* Check it isn't doing very strange things with descriptor numbers. */
2225	last_avail_idx = vq->last_avail_idx;
2226
2227	if (vq->avail_idx == vq->last_avail_idx) {
2228		if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2229			vq_err(vq, "Failed to access avail idx at %p\n",
2230				&vq->avail->idx);
2231			return -EFAULT;
2232		}
2233		vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2234
2235		if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2236			vq_err(vq, "Guest moved used index from %u to %u",
2237				last_avail_idx, vq->avail_idx);
2238			return -EFAULT;
2239		}
2240
2241		/* If there's nothing new since last we looked, return
2242		 * invalid.
2243		 */
2244		if (vq->avail_idx == last_avail_idx)
2245			return vq->num;
2246
2247		/* Only get avail ring entries after they have been
2248		 * exposed by guest.
2249		 */
2250		smp_rmb();
2251	}
2252
2253	/* Grab the next descriptor number they're advertising, and increment
2254	 * the index we've seen. */
2255	if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2256		vq_err(vq, "Failed to read head: idx %d address %p\n",
2257		       last_avail_idx,
2258		       &vq->avail->ring[last_avail_idx % vq->num]);
2259		return -EFAULT;
2260	}
2261
2262	head = vhost16_to_cpu(vq, ring_head);
2263
2264	/* If their number is silly, that's an error. */
2265	if (unlikely(head >= vq->num)) {
2266		vq_err(vq, "Guest says index %u > %u is available",
2267		       head, vq->num);
2268		return -EINVAL;
2269	}
2270
2271	/* When we start there are none of either input nor output. */
2272	*out_num = *in_num = 0;
2273	if (unlikely(log))
2274		*log_num = 0;
2275
2276	i = head;
2277	do {
2278		unsigned iov_count = *in_num + *out_num;
2279		if (unlikely(i >= vq->num)) {
2280			vq_err(vq, "Desc index is %u > %u, head = %u",
2281			       i, vq->num, head);
2282			return -EINVAL;
2283		}
2284		if (unlikely(++found > vq->num)) {
2285			vq_err(vq, "Loop detected: last one at %u "
2286			       "vq size %u head %u\n",
2287			       i, vq->num, head);
2288			return -EINVAL;
2289		}
2290		ret = vhost_get_desc(vq, &desc, i);
2291		if (unlikely(ret)) {
2292			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2293			       i, vq->desc + i);
2294			return -EFAULT;
2295		}
2296		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2297			ret = get_indirect(vq, iov, iov_size,
2298					   out_num, in_num,
2299					   log, log_num, &desc);
2300			if (unlikely(ret < 0)) {
2301				if (ret != -EAGAIN)
2302					vq_err(vq, "Failure detected "
2303						"in indirect descriptor at idx %d\n", i);
2304				return ret;
2305			}
2306			continue;
2307		}
2308
2309		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2310			access = VHOST_ACCESS_WO;
2311		else
2312			access = VHOST_ACCESS_RO;
2313		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2314				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2315				     iov_size - iov_count, access);
2316		if (unlikely(ret < 0)) {
2317			if (ret != -EAGAIN)
2318				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2319					ret, i);
2320			return ret;
2321		}
2322		if (access == VHOST_ACCESS_WO) {
2323			/* If this is an input descriptor,
2324			 * increment that count. */
2325			*in_num += ret;
2326			if (unlikely(log && ret)) {
2327				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2328				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2329				++*log_num;
2330			}
2331		} else {
2332			/* If it's an output descriptor, they're all supposed
2333			 * to come before any input descriptors. */
2334			if (unlikely(*in_num)) {
2335				vq_err(vq, "Descriptor has out after in: "
2336				       "idx %d\n", i);
2337				return -EINVAL;
2338			}
2339			*out_num += ret;
2340		}
2341	} while ((i = next_desc(vq, &desc)) != -1);
2342
2343	/* On success, increment avail index. */
2344	vq->last_avail_idx++;
2345
2346	/* Assume notifications from guest are disabled at this point,
2347	 * if they aren't we would need to update avail_event index. */
2348	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2349	return head;
2350}
2351EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2352
2353/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2354void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2355{
2356	vq->last_avail_idx -= n;
2357}
2358EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2359
2360/* After we've used one of their buffers, we tell them about it.  We'll then
2361 * want to notify the guest, using eventfd. */
2362int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2363{
2364	struct vring_used_elem heads = {
2365		cpu_to_vhost32(vq, head),
2366		cpu_to_vhost32(vq, len)
2367	};
2368
2369	return vhost_add_used_n(vq, &heads, 1);
2370}
2371EXPORT_SYMBOL_GPL(vhost_add_used);
2372
2373static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2374			    struct vring_used_elem *heads,
2375			    unsigned count)
2376{
2377	vring_used_elem_t __user *used;
2378	u16 old, new;
2379	int start;
2380
2381	start = vq->last_used_idx & (vq->num - 1);
2382	used = vq->used->ring + start;
2383	if (vhost_put_used(vq, heads, start, count)) {
2384		vq_err(vq, "Failed to write used");
2385		return -EFAULT;
2386	}
2387	if (unlikely(vq->log_used)) {
2388		/* Make sure data is seen before log. */
2389		smp_wmb();
2390		/* Log used ring entry write. */
2391		log_used(vq, ((void __user *)used - (void __user *)vq->used),
2392			 count * sizeof *used);
2393	}
2394	old = vq->last_used_idx;
2395	new = (vq->last_used_idx += count);
2396	/* If the driver never bothers to signal in a very long while,
2397	 * used index might wrap around. If that happens, invalidate
2398	 * signalled_used index we stored. TODO: make sure driver
2399	 * signals at least once in 2^16 and remove this. */
2400	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2401		vq->signalled_used_valid = false;
2402	return 0;
2403}
2404
2405/* After we've used one of their buffers, we tell them about it.  We'll then
2406 * want to notify the guest, using eventfd. */
2407int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2408		     unsigned count)
2409{
2410	int start, n, r;
2411
2412	start = vq->last_used_idx & (vq->num - 1);
2413	n = vq->num - start;
2414	if (n < count) {
2415		r = __vhost_add_used_n(vq, heads, n);
2416		if (r < 0)
2417			return r;
2418		heads += n;
2419		count -= n;
2420	}
2421	r = __vhost_add_used_n(vq, heads, count);
2422
2423	/* Make sure buffer is written before we update index. */
2424	smp_wmb();
2425	if (vhost_put_used_idx(vq)) {
2426		vq_err(vq, "Failed to increment used idx");
2427		return -EFAULT;
2428	}
2429	if (unlikely(vq->log_used)) {
2430		/* Make sure used idx is seen before log. */
2431		smp_wmb();
2432		/* Log used index update. */
2433		log_used(vq, offsetof(struct vring_used, idx),
2434			 sizeof vq->used->idx);
2435		if (vq->log_ctx)
2436			eventfd_signal(vq->log_ctx, 1);
2437	}
2438	return r;
2439}
2440EXPORT_SYMBOL_GPL(vhost_add_used_n);
2441
2442static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2443{
2444	__u16 old, new;
2445	__virtio16 event;
2446	bool v;
2447	/* Flush out used index updates. This is paired
2448	 * with the barrier that the Guest executes when enabling
2449	 * interrupts. */
2450	smp_mb();
2451
2452	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2453	    unlikely(vq->avail_idx == vq->last_avail_idx))
2454		return true;
2455
2456	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2457		__virtio16 flags;
2458		if (vhost_get_avail_flags(vq, &flags)) {
2459			vq_err(vq, "Failed to get flags");
2460			return true;
2461		}
2462		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2463	}
2464	old = vq->signalled_used;
2465	v = vq->signalled_used_valid;
2466	new = vq->signalled_used = vq->last_used_idx;
2467	vq->signalled_used_valid = true;
2468
2469	if (unlikely(!v))
2470		return true;
2471
2472	if (vhost_get_used_event(vq, &event)) {
2473		vq_err(vq, "Failed to get used event idx");
2474		return true;
2475	}
2476	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2477}
2478
2479/* This actually signals the guest, using eventfd. */
2480void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2481{
2482	/* Signal the Guest tell them we used something up. */
2483	if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2484		eventfd_signal(vq->call_ctx.ctx, 1);
2485}
2486EXPORT_SYMBOL_GPL(vhost_signal);
2487
2488/* And here's the combo meal deal.  Supersize me! */
2489void vhost_add_used_and_signal(struct vhost_dev *dev,
2490			       struct vhost_virtqueue *vq,
2491			       unsigned int head, int len)
2492{
2493	vhost_add_used(vq, head, len);
2494	vhost_signal(dev, vq);
2495}
2496EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2497
2498/* multi-buffer version of vhost_add_used_and_signal */
2499void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2500				 struct vhost_virtqueue *vq,
2501				 struct vring_used_elem *heads, unsigned count)
2502{
2503	vhost_add_used_n(vq, heads, count);
2504	vhost_signal(dev, vq);
2505}
2506EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2507
2508/* return true if we're sure that avaiable ring is empty */
2509bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2510{
2511	__virtio16 avail_idx;
2512	int r;
2513
2514	if (vq->avail_idx != vq->last_avail_idx)
2515		return false;
2516
2517	r = vhost_get_avail_idx(vq, &avail_idx);
2518	if (unlikely(r))
2519		return false;
2520	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2521
2522	return vq->avail_idx == vq->last_avail_idx;
 
2523}
2524EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2525
2526/* OK, now we need to know about added descriptors. */
2527bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2528{
2529	__virtio16 avail_idx;
2530	int r;
2531
2532	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2533		return false;
2534	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2535	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2536		r = vhost_update_used_flags(vq);
2537		if (r) {
2538			vq_err(vq, "Failed to enable notification at %p: %d\n",
2539			       &vq->used->flags, r);
2540			return false;
2541		}
2542	} else {
2543		r = vhost_update_avail_event(vq);
2544		if (r) {
2545			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2546			       vhost_avail_event(vq), r);
2547			return false;
2548		}
2549	}
2550	/* They could have slipped one in as we were doing that: make
2551	 * sure it's written, then check again. */
2552	smp_mb();
2553	r = vhost_get_avail_idx(vq, &avail_idx);
2554	if (r) {
2555		vq_err(vq, "Failed to check avail idx at %p: %d\n",
2556		       &vq->avail->idx, r);
2557		return false;
2558	}
2559	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2560
2561	return vq->avail_idx != vq->last_avail_idx;
2562}
2563EXPORT_SYMBOL_GPL(vhost_enable_notify);
2564
2565/* We don't need to be notified again. */
2566void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2567{
2568	int r;
2569
2570	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2571		return;
2572	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2573	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2574		r = vhost_update_used_flags(vq);
2575		if (r)
2576			vq_err(vq, "Failed to disable notification at %p: %d\n",
2577			       &vq->used->flags, r);
2578	}
2579}
2580EXPORT_SYMBOL_GPL(vhost_disable_notify);
2581
2582/* Create a new message. */
2583struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2584{
2585	struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
 
2586	if (!node)
2587		return NULL;
2588
2589	/* Make sure all padding within the structure is initialized. */
2590	memset(&node->msg, 0, sizeof node->msg);
2591	node->vq = vq;
2592	node->msg.type = type;
2593	return node;
2594}
2595EXPORT_SYMBOL_GPL(vhost_new_msg);
2596
2597void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2598		       struct vhost_msg_node *node)
2599{
2600	spin_lock(&dev->iotlb_lock);
2601	list_add_tail(&node->node, head);
2602	spin_unlock(&dev->iotlb_lock);
2603
2604	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2605}
2606EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2607
2608struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2609					 struct list_head *head)
2610{
2611	struct vhost_msg_node *node = NULL;
2612
2613	spin_lock(&dev->iotlb_lock);
2614	if (!list_empty(head)) {
2615		node = list_first_entry(head, struct vhost_msg_node,
2616					node);
2617		list_del(&node->node);
2618	}
2619	spin_unlock(&dev->iotlb_lock);
2620
2621	return node;
2622}
2623EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2624
2625void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2626{
2627	struct vhost_virtqueue *vq;
2628	int i;
2629
2630	mutex_lock(&dev->mutex);
2631	for (i = 0; i < dev->nvqs; ++i) {
2632		vq = dev->vqs[i];
2633		mutex_lock(&vq->mutex);
2634		vq->acked_backend_features = features;
2635		mutex_unlock(&vq->mutex);
2636	}
2637	mutex_unlock(&dev->mutex);
2638}
2639EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2640
2641static int __init vhost_init(void)
2642{
2643	return 0;
2644}
2645
2646static void __exit vhost_exit(void)
2647{
2648}
2649
2650module_init(vhost_init);
2651module_exit(vhost_exit);
2652
2653MODULE_VERSION("0.0.1");
2654MODULE_LICENSE("GPL v2");
2655MODULE_AUTHOR("Michael S. Tsirkin");
2656MODULE_DESCRIPTION("Host kernel accelerator for virtio");