Linux Audio

Check our new training course

Loading...
v3.1
 
   1/* Copyright (C) 2009 Red Hat, Inc.
   2 * Copyright (C) 2006 Rusty Russell IBM Corporation
   3 *
   4 * Author: Michael S. Tsirkin <mst@redhat.com>
   5 *
   6 * Inspiration, some code, and most witty comments come from
   7 * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8 *
   9 * This work is licensed under the terms of the GNU GPL, version 2.
  10 *
  11 * Generic code for virtio server in host kernel.
  12 */
  13
  14#include <linux/eventfd.h>
  15#include <linux/vhost.h>
  16#include <linux/virtio_net.h>
  17#include <linux/mm.h>
  18#include <linux/mmu_context.h>
  19#include <linux/miscdevice.h>
  20#include <linux/mutex.h>
  21#include <linux/rcupdate.h>
  22#include <linux/poll.h>
  23#include <linux/file.h>
  24#include <linux/highmem.h>
  25#include <linux/slab.h>
 
  26#include <linux/kthread.h>
  27#include <linux/cgroup.h>
  28
  29#include <linux/net.h>
  30#include <linux/if_packet.h>
  31#include <linux/if_arp.h>
 
 
 
  32
  33#include "vhost.h"
  34
 
 
 
 
 
 
 
 
 
  35enum {
  36	VHOST_MEMORY_MAX_NREGIONS = 64,
  37	VHOST_MEMORY_F_LOG = 0x1,
  38};
  39
  40static unsigned vhost_zcopy_mask __read_mostly;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  41
  42#define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
  43#define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
 
  44
  45static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  46			    poll_table *pt)
  47{
  48	struct vhost_poll *poll;
  49
  50	poll = container_of(pt, struct vhost_poll, table);
  51	poll->wqh = wqh;
  52	add_wait_queue(wqh, &poll->wait);
  53}
  54
  55static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  56			     void *key)
  57{
  58	struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 
  59
  60	if (!((unsigned long)key & poll->mask))
  61		return 0;
  62
  63	vhost_poll_queue(poll);
 
 
 
 
  64	return 0;
  65}
  66
  67static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
  68{
  69	INIT_LIST_HEAD(&work->node);
  70	work->fn = fn;
  71	init_waitqueue_head(&work->done);
  72	work->flushing = 0;
  73	work->queue_seq = work->done_seq = 0;
  74}
 
  75
  76/* Init poll structure */
  77void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
  78		     unsigned long mask, struct vhost_dev *dev)
  79{
  80	init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
  81	init_poll_funcptr(&poll->table, vhost_poll_func);
  82	poll->mask = mask;
  83	poll->dev = dev;
 
  84
  85	vhost_work_init(&poll->work, fn);
  86}
 
  87
  88/* Start polling a file. We add ourselves to file's wait queue. The caller must
  89 * keep a reference to a file until after vhost_poll_stop is called. */
  90void vhost_poll_start(struct vhost_poll *poll, struct file *file)
  91{
  92	unsigned long mask;
 
 
 
  93
  94	mask = file->f_op->poll(file, &poll->table);
  95	if (mask)
  96		vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 
 
 
 
 
 
  97}
 
  98
  99/* Stop polling a file. After this function returns, it becomes safe to drop the
 100 * file reference. You must also flush afterwards. */
 101void vhost_poll_stop(struct vhost_poll *poll)
 102{
 103	remove_wait_queue(poll->wqh, &poll->wait);
 
 
 
 104}
 
 105
 106static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 107				unsigned seq)
 108{
 109	int left;
 110
 111	spin_lock_irq(&dev->work_lock);
 112	left = seq - work->done_seq;
 113	spin_unlock_irq(&dev->work_lock);
 114	return left <= 0;
 115}
 116
 117static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 118{
 119	unsigned seq;
 120	int flushing;
 121
 122	spin_lock_irq(&dev->work_lock);
 123	seq = work->queue_seq;
 124	work->flushing++;
 125	spin_unlock_irq(&dev->work_lock);
 126	wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 127	spin_lock_irq(&dev->work_lock);
 128	flushing = --work->flushing;
 129	spin_unlock_irq(&dev->work_lock);
 130	BUG_ON(flushing < 0);
 131}
 
 132
 133/* Flush any work that has been scheduled. When calling this, don't hold any
 134 * locks that are also used by the callback. */
 135void vhost_poll_flush(struct vhost_poll *poll)
 136{
 137	vhost_work_flush(poll->dev, &poll->work);
 138}
 
 139
 140static inline void vhost_work_queue(struct vhost_dev *dev,
 141				    struct vhost_work *work)
 142{
 143	unsigned long flags;
 
 144
 145	spin_lock_irqsave(&dev->work_lock, flags);
 146	if (list_empty(&work->node)) {
 147		list_add_tail(&work->node, &dev->work_list);
 148		work->queue_seq++;
 
 
 149		wake_up_process(dev->worker);
 150	}
 151	spin_unlock_irqrestore(&dev->work_lock, flags);
 152}
 
 
 
 
 
 
 
 
 153
 154void vhost_poll_queue(struct vhost_poll *poll)
 155{
 156	vhost_work_queue(poll->dev, &poll->work);
 157}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 158
 159static void vhost_vq_reset(struct vhost_dev *dev,
 160			   struct vhost_virtqueue *vq)
 161{
 162	vq->num = 1;
 163	vq->desc = NULL;
 164	vq->avail = NULL;
 165	vq->used = NULL;
 166	vq->last_avail_idx = 0;
 167	vq->avail_idx = 0;
 168	vq->last_used_idx = 0;
 169	vq->signalled_used = 0;
 170	vq->signalled_used_valid = false;
 171	vq->used_flags = 0;
 172	vq->log_used = false;
 173	vq->log_addr = -1ull;
 174	vq->vhost_hlen = 0;
 175	vq->sock_hlen = 0;
 176	vq->private_data = NULL;
 
 
 177	vq->log_base = NULL;
 178	vq->error_ctx = NULL;
 179	vq->error = NULL;
 180	vq->kick = NULL;
 181	vq->call_ctx = NULL;
 182	vq->call = NULL;
 183	vq->log_ctx = NULL;
 184	vq->upend_idx = 0;
 185	vq->done_idx = 0;
 186	vq->ubufs = NULL;
 
 
 
 
 187}
 188
 189static int vhost_worker(void *data)
 190{
 191	struct vhost_dev *dev = data;
 192	struct vhost_work *work = NULL;
 193	unsigned uninitialized_var(seq);
 194
 195	use_mm(dev->mm);
 196
 197	for (;;) {
 198		/* mb paired w/ kthread_stop */
 199		set_current_state(TASK_INTERRUPTIBLE);
 200
 201		spin_lock_irq(&dev->work_lock);
 202		if (work) {
 203			work->done_seq = seq;
 204			if (work->flushing)
 205				wake_up_all(&work->done);
 206		}
 207
 208		if (kthread_should_stop()) {
 209			spin_unlock_irq(&dev->work_lock);
 210			__set_current_state(TASK_RUNNING);
 211			break;
 212		}
 213		if (!list_empty(&dev->work_list)) {
 214			work = list_first_entry(&dev->work_list,
 215						struct vhost_work, node);
 216			list_del_init(&work->node);
 217			seq = work->queue_seq;
 218		} else
 219			work = NULL;
 220		spin_unlock_irq(&dev->work_lock);
 221
 222		if (work) {
 223			__set_current_state(TASK_RUNNING);
 224			work->fn(work);
 225		} else
 226			schedule();
 227
 
 
 
 
 
 
 
 
 
 
 
 
 228	}
 229	unuse_mm(dev->mm);
 230	return 0;
 231}
 232
 233static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 234{
 235	kfree(vq->indirect);
 236	vq->indirect = NULL;
 237	kfree(vq->log);
 238	vq->log = NULL;
 239	kfree(vq->heads);
 240	vq->heads = NULL;
 241	kfree(vq->ubuf_info);
 242	vq->ubuf_info = NULL;
 243}
 244
 245void vhost_enable_zcopy(int vq)
 246{
 247	vhost_zcopy_mask |= 0x1 << vq;
 248}
 249
 250/* Helper to allocate iovec buffers for all vqs. */
 251static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 252{
 
 253	int i;
 254	bool zcopy;
 255
 256	for (i = 0; i < dev->nvqs; ++i) {
 257		dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
 258					       UIO_MAXIOV, GFP_KERNEL);
 259		dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
 
 
 
 
 260					  GFP_KERNEL);
 261		dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
 262					    UIO_MAXIOV, GFP_KERNEL);
 263		zcopy = vhost_zcopy_mask & (0x1 << i);
 264		if (zcopy)
 265			dev->vqs[i].ubuf_info =
 266				kmalloc(sizeof *dev->vqs[i].ubuf_info *
 267					UIO_MAXIOV, GFP_KERNEL);
 268		if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
 269			!dev->vqs[i].heads ||
 270			(zcopy && !dev->vqs[i].ubuf_info))
 271			goto err_nomem;
 272	}
 273	return 0;
 274
 275err_nomem:
 276	for (; i >= 0; --i)
 277		vhost_vq_free_iovecs(&dev->vqs[i]);
 278	return -ENOMEM;
 279}
 280
 281static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 282{
 283	int i;
 284
 285	for (i = 0; i < dev->nvqs; ++i)
 286		vhost_vq_free_iovecs(&dev->vqs[i]);
 287}
 288
 289long vhost_dev_init(struct vhost_dev *dev,
 290		    struct vhost_virtqueue *vqs, int nvqs)
 291{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 292	int i;
 293
 294	dev->vqs = vqs;
 295	dev->nvqs = nvqs;
 296	mutex_init(&dev->mutex);
 297	dev->log_ctx = NULL;
 298	dev->log_file = NULL;
 299	dev->memory = NULL;
 300	dev->mm = NULL;
 301	spin_lock_init(&dev->work_lock);
 302	INIT_LIST_HEAD(&dev->work_list);
 303	dev->worker = NULL;
 
 
 
 
 
 
 
 
 
 
 
 304
 305	for (i = 0; i < dev->nvqs; ++i) {
 306		dev->vqs[i].log = NULL;
 307		dev->vqs[i].indirect = NULL;
 308		dev->vqs[i].heads = NULL;
 309		dev->vqs[i].ubuf_info = NULL;
 310		dev->vqs[i].dev = dev;
 311		mutex_init(&dev->vqs[i].mutex);
 312		vhost_vq_reset(dev, dev->vqs + i);
 313		if (dev->vqs[i].handle_kick)
 314			vhost_poll_init(&dev->vqs[i].poll,
 315					dev->vqs[i].handle_kick, POLLIN, dev);
 316	}
 317
 318	return 0;
 319}
 
 320
 321/* Caller should have device mutex */
 322long vhost_dev_check_owner(struct vhost_dev *dev)
 323{
 324	/* Are you the owner? If not, I don't think you mean to do that */
 325	return dev->mm == current->mm ? 0 : -EPERM;
 326}
 
 327
 328struct vhost_attach_cgroups_struct {
 329	struct vhost_work work;
 330	struct task_struct *owner;
 331	int ret;
 332};
 333
 334static void vhost_attach_cgroups_work(struct vhost_work *work)
 335{
 336	struct vhost_attach_cgroups_struct *s;
 337
 338	s = container_of(work, struct vhost_attach_cgroups_struct, work);
 339	s->ret = cgroup_attach_task_all(s->owner, current);
 340}
 341
 342static int vhost_attach_cgroups(struct vhost_dev *dev)
 343{
 344	struct vhost_attach_cgroups_struct attach;
 345
 346	attach.owner = current;
 347	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 348	vhost_work_queue(dev, &attach.work);
 349	vhost_work_flush(dev, &attach.work);
 350	return attach.ret;
 351}
 352
 353/* Caller should have device mutex */
 354static long vhost_dev_set_owner(struct vhost_dev *dev)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 355{
 356	struct task_struct *worker;
 357	int err;
 358
 359	/* Is there an owner already? */
 360	if (dev->mm) {
 361		err = -EBUSY;
 362		goto err_mm;
 363	}
 364
 365	/* No owner, become one */
 366	dev->mm = get_task_mm(current);
 367	worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 368	if (IS_ERR(worker)) {
 369		err = PTR_ERR(worker);
 370		goto err_worker;
 371	}
 372
 373	dev->worker = worker;
 374	wake_up_process(worker);	/* avoid contributing to loadavg */
 375
 376	err = vhost_attach_cgroups(dev);
 377	if (err)
 378		goto err_cgroup;
 
 
 
 
 
 
 
 
 
 
 
 
 
 379
 380	err = vhost_dev_alloc_iovecs(dev);
 381	if (err)
 382		goto err_cgroup;
 383
 384	return 0;
 385err_cgroup:
 386	kthread_stop(worker);
 387	dev->worker = NULL;
 
 
 388err_worker:
 389	if (dev->mm)
 390		mmput(dev->mm);
 391	dev->mm = NULL;
 392err_mm:
 393	return err;
 394}
 
 395
 396/* Caller should have device mutex */
 397long vhost_dev_reset_owner(struct vhost_dev *dev)
 398{
 399	struct vhost_memory *memory;
 
 
 400
 401	/* Restore memory to default empty mapping. */
 402	memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 403	if (!memory)
 404		return -ENOMEM;
 
 
 
 
 
 
 405
 406	vhost_dev_cleanup(dev);
 407
 408	memory->nregions = 0;
 409	RCU_INIT_POINTER(dev->memory, memory);
 410	return 0;
 
 
 
 411}
 
 412
 413/* In case of DMA done not in order in lower device driver for some reason.
 414 * upend_idx is used to track end of used idx, done_idx is used to track head
 415 * of used idx. Once lower device DMA done contiguously, we will signal KVM
 416 * guest used idx.
 417 */
 418int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
 419{
 420	int i;
 421	int j = 0;
 422
 423	for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 424		if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
 425			vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 426			vhost_add_used_and_signal(vq->dev, vq,
 427						  vq->heads[i].id, 0);
 428			++j;
 429		} else
 430			break;
 431	}
 432	if (j)
 433		vq->done_idx = i;
 434	return j;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 435}
 436
 437/* Caller should have device mutex */
 438void vhost_dev_cleanup(struct vhost_dev *dev)
 439{
 440	int i;
 441
 442	for (i = 0; i < dev->nvqs; ++i) {
 443		if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
 444			vhost_poll_stop(&dev->vqs[i].poll);
 445			vhost_poll_flush(&dev->vqs[i].poll);
 446		}
 447		/* Wait for all lower device DMAs done. */
 448		if (dev->vqs[i].ubufs)
 449			vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
 450
 451		/* Signal guest as appropriate. */
 452		vhost_zerocopy_signal_used(&dev->vqs[i]);
 453
 454		if (dev->vqs[i].error_ctx)
 455			eventfd_ctx_put(dev->vqs[i].error_ctx);
 456		if (dev->vqs[i].error)
 457			fput(dev->vqs[i].error);
 458		if (dev->vqs[i].kick)
 459			fput(dev->vqs[i].kick);
 460		if (dev->vqs[i].call_ctx)
 461			eventfd_ctx_put(dev->vqs[i].call_ctx);
 462		if (dev->vqs[i].call)
 463			fput(dev->vqs[i].call);
 464		vhost_vq_reset(dev, dev->vqs + i);
 465	}
 466	vhost_dev_free_iovecs(dev);
 467	if (dev->log_ctx)
 468		eventfd_ctx_put(dev->log_ctx);
 469	dev->log_ctx = NULL;
 470	if (dev->log_file)
 471		fput(dev->log_file);
 472	dev->log_file = NULL;
 473	/* No one will access memory at this point */
 474	kfree(rcu_dereference_protected(dev->memory,
 475					lockdep_is_held(&dev->mutex)));
 476	RCU_INIT_POINTER(dev->memory, NULL);
 477	WARN_ON(!list_empty(&dev->work_list));
 
 
 
 478	if (dev->worker) {
 479		kthread_stop(dev->worker);
 480		dev->worker = NULL;
 
 481	}
 482	if (dev->mm)
 483		mmput(dev->mm);
 484	dev->mm = NULL;
 485}
 
 486
 487static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 488{
 489	u64 a = addr / VHOST_PAGE_SIZE / 8;
 490
 491	/* Make sure 64 bit math will not overflow. */
 492	if (a > ULONG_MAX - (unsigned long)log_base ||
 493	    a + (unsigned long)log_base > ULONG_MAX)
 494		return 0;
 495
 496	return access_ok(VERIFY_WRITE, log_base + a,
 497			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 498}
 499
 
 
 
 
 
 
 
 
 
 
 
 
 500/* Caller should have vq mutex and device mutex. */
 501static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 502			       int log_all)
 503{
 504	int i;
 505
 506	if (!mem)
 507		return 0;
 508
 509	for (i = 0; i < mem->nregions; ++i) {
 510		struct vhost_memory_region *m = mem->regions + i;
 511		unsigned long a = m->userspace_addr;
 512		if (m->memory_size > ULONG_MAX)
 513			return 0;
 514		else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 515				    m->memory_size))
 516			return 0;
 
 517		else if (log_all && !log_access_ok(log_base,
 518						   m->guest_phys_addr,
 519						   m->memory_size))
 520			return 0;
 521	}
 522	return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 523}
 524
 525/* Can we switch to this memory table? */
 526/* Caller should have device mutex but not vq mutex */
 527static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 528			    int log_all)
 529{
 530	int i;
 531
 532	for (i = 0; i < d->nvqs; ++i) {
 533		int ok;
 534		mutex_lock(&d->vqs[i].mutex);
 
 
 
 535		/* If ring is inactive, will check when it's enabled. */
 536		if (d->vqs[i].private_data)
 537			ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
 538						 log_all);
 539		else
 540			ok = 1;
 541		mutex_unlock(&d->vqs[i].mutex);
 542		if (!ok)
 543			return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 544	}
 545	return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 546}
 547
 548static int vq_access_ok(struct vhost_dev *d, unsigned int num,
 549			struct vring_desc __user *desc,
 550			struct vring_avail __user *avail,
 551			struct vring_used __user *used)
 552{
 553	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 554	return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 555	       access_ok(VERIFY_READ, avail,
 556			 sizeof *avail + num * sizeof *avail->ring + s) &&
 557	       access_ok(VERIFY_WRITE, used,
 558			sizeof *used + num * sizeof *used->ring + s);
 559}
 560
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 561/* Can we log writes? */
 562/* Caller should have device mutex but not vq mutex */
 563int vhost_log_access_ok(struct vhost_dev *dev)
 
 
 
 
 
 
 
 
 
 564{
 565	struct vhost_memory *mp;
 
 
 
 566
 567	mp = rcu_dereference_protected(dev->memory,
 568				       lockdep_is_held(&dev->mutex));
 569	return memory_access_ok(dev, mp, 1);
 570}
 571
 572/* Verify access for write logging. */
 573/* Caller should have vq mutex and device mutex */
 574static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 575			    void __user *log_base)
 576{
 577	struct vhost_memory *mp;
 578	size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 579
 580	mp = rcu_dereference_protected(vq->dev->memory,
 581				       lockdep_is_held(&vq->mutex));
 582	return vq_memory_access_ok(log_base, mp,
 583			    vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 584		(!vq->log_used || log_access_ok(log_base, vq->log_addr,
 585					sizeof *vq->used +
 586					vq->num * sizeof *vq->used->ring + s));
 587}
 588
 589/* Can we start vq? */
 590/* Caller should have vq mutex and device mutex */
 591int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 592{
 593	return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 594		vq_log_access_ok(vq->dev, vq, vq->log_base);
 
 
 595}
 
 596
 597static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 598{
 599	struct vhost_memory mem, *newmem, *oldmem;
 
 
 600	unsigned long size = offsetof(struct vhost_memory, regions);
 
 601
 602	if (copy_from_user(&mem, m, size))
 603		return -EFAULT;
 604	if (mem.padding)
 605		return -EOPNOTSUPP;
 606	if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 607		return -E2BIG;
 608	newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 
 609	if (!newmem)
 610		return -ENOMEM;
 611
 612	memcpy(newmem, &mem, size);
 613	if (copy_from_user(newmem->regions, m->regions,
 614			   mem.nregions * sizeof *m->regions)) {
 615		kfree(newmem);
 616		return -EFAULT;
 617	}
 618
 619	if (!memory_access_ok(d, newmem,
 620			      vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 621		kfree(newmem);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 622		return -EFAULT;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 623	}
 624	oldmem = rcu_dereference_protected(d->memory,
 625					   lockdep_is_held(&d->mutex));
 626	rcu_assign_pointer(d->memory, newmem);
 627	synchronize_rcu();
 628	kfree(oldmem);
 
 
 629	return 0;
 630}
 631
 632static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
 
 
 
 633{
 634	struct file *eventfp, *filep = NULL,
 635		    *pollstart = NULL, *pollstop = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 636	struct eventfd_ctx *ctx = NULL;
 637	u32 __user *idxp = argp;
 638	struct vhost_virtqueue *vq;
 639	struct vhost_vring_state s;
 640	struct vhost_vring_file f;
 641	struct vhost_vring_addr a;
 642	u32 idx;
 643	long r;
 644
 645	r = get_user(idx, idxp);
 646	if (r < 0)
 647		return r;
 648	if (idx >= d->nvqs)
 649		return -ENOBUFS;
 650
 651	vq = d->vqs + idx;
 
 
 
 
 
 
 652
 653	mutex_lock(&vq->mutex);
 654
 655	switch (ioctl) {
 656	case VHOST_SET_VRING_NUM:
 657		/* Resizing ring with an active backend?
 658		 * You don't want to do that. */
 659		if (vq->private_data) {
 660			r = -EBUSY;
 661			break;
 662		}
 663		if (copy_from_user(&s, argp, sizeof s)) {
 664			r = -EFAULT;
 665			break;
 666		}
 667		if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 668			r = -EINVAL;
 669			break;
 670		}
 671		vq->num = s.num;
 672		break;
 673	case VHOST_SET_VRING_BASE:
 674		/* Moving base with an active backend?
 675		 * You don't want to do that. */
 676		if (vq->private_data) {
 677			r = -EBUSY;
 678			break;
 679		}
 680		if (copy_from_user(&s, argp, sizeof s)) {
 681			r = -EFAULT;
 682			break;
 683		}
 684		if (s.num > 0xffff) {
 685			r = -EINVAL;
 686			break;
 687		}
 688		vq->last_avail_idx = s.num;
 689		/* Forget the cached index value. */
 690		vq->avail_idx = vq->last_avail_idx;
 691		break;
 692	case VHOST_GET_VRING_BASE:
 693		s.index = idx;
 694		s.num = vq->last_avail_idx;
 695		if (copy_to_user(argp, &s, sizeof s))
 696			r = -EFAULT;
 697		break;
 698	case VHOST_SET_VRING_ADDR:
 699		if (copy_from_user(&a, argp, sizeof a)) {
 700			r = -EFAULT;
 701			break;
 702		}
 703		if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 704			r = -EOPNOTSUPP;
 705			break;
 706		}
 707		/* For 32bit, verify that the top 32bits of the user
 708		   data are set to zero. */
 709		if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 710		    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 711		    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 712			r = -EFAULT;
 713			break;
 714		}
 715		if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 716		    (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 717		    (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 718			r = -EINVAL;
 719			break;
 720		}
 721
 722		/* We only verify access here if backend is configured.
 723		 * If it is not, we don't as size might not have been setup.
 724		 * We will verify when backend is configured. */
 725		if (vq->private_data) {
 726			if (!vq_access_ok(d, vq->num,
 727				(void __user *)(unsigned long)a.desc_user_addr,
 728				(void __user *)(unsigned long)a.avail_user_addr,
 729				(void __user *)(unsigned long)a.used_user_addr)) {
 730				r = -EINVAL;
 731				break;
 732			}
 733
 734			/* Also validate log access for used ring if enabled. */
 735			if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 736			    !log_access_ok(vq->log_base, a.log_guest_addr,
 737					   sizeof *vq->used +
 738					   vq->num * sizeof *vq->used->ring)) {
 739				r = -EINVAL;
 740				break;
 741			}
 742		}
 743
 744		vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 745		vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 746		vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 747		vq->log_addr = a.log_guest_addr;
 748		vq->used = (void __user *)(unsigned long)a.used_user_addr;
 749		break;
 750	case VHOST_SET_VRING_KICK:
 751		if (copy_from_user(&f, argp, sizeof f)) {
 752			r = -EFAULT;
 753			break;
 754		}
 755		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 756		if (IS_ERR(eventfp)) {
 757			r = PTR_ERR(eventfp);
 758			break;
 759		}
 760		if (eventfp != vq->kick) {
 761			pollstop = filep = vq->kick;
 762			pollstart = vq->kick = eventfp;
 763		} else
 764			filep = eventfp;
 765		break;
 766	case VHOST_SET_VRING_CALL:
 767		if (copy_from_user(&f, argp, sizeof f)) {
 768			r = -EFAULT;
 769			break;
 770		}
 771		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 772		if (IS_ERR(eventfp)) {
 773			r = PTR_ERR(eventfp);
 774			break;
 775		}
 776		if (eventfp != vq->call) {
 777			filep = vq->call;
 778			ctx = vq->call_ctx;
 779			vq->call = eventfp;
 780			vq->call_ctx = eventfp ?
 781				eventfd_ctx_fileget(eventfp) : NULL;
 782		} else
 783			filep = eventfp;
 784		break;
 785	case VHOST_SET_VRING_ERR:
 786		if (copy_from_user(&f, argp, sizeof f)) {
 787			r = -EFAULT;
 788			break;
 789		}
 790		eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 791		if (IS_ERR(eventfp)) {
 792			r = PTR_ERR(eventfp);
 793			break;
 794		}
 795		if (eventfp != vq->error) {
 796			filep = vq->error;
 797			vq->error = eventfp;
 798			ctx = vq->error_ctx;
 799			vq->error_ctx = eventfp ?
 800				eventfd_ctx_fileget(eventfp) : NULL;
 801		} else
 802			filep = eventfp;
 
 
 
 
 
 
 
 
 
 
 
 
 803		break;
 804	default:
 805		r = -ENOIOCTLCMD;
 806	}
 807
 808	if (pollstop && vq->handle_kick)
 809		vhost_poll_stop(&vq->poll);
 810
 811	if (ctx)
 812		eventfd_ctx_put(ctx);
 813	if (filep)
 814		fput(filep);
 815
 816	if (pollstart && vq->handle_kick)
 817		vhost_poll_start(&vq->poll, vq->kick);
 818
 819	mutex_unlock(&vq->mutex);
 820
 821	if (pollstop && vq->handle_kick)
 822		vhost_poll_flush(&vq->poll);
 823	return r;
 824}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 825
 826/* Caller must have device mutex */
 827long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
 828{
 829	void __user *argp = (void __user *)arg;
 830	struct file *eventfp, *filep = NULL;
 831	struct eventfd_ctx *ctx = NULL;
 832	u64 p;
 833	long r;
 834	int i, fd;
 835
 836	/* If you are not the owner, you can become one */
 837	if (ioctl == VHOST_SET_OWNER) {
 838		r = vhost_dev_set_owner(d);
 839		goto done;
 840	}
 841
 842	/* You must be the owner to do anything else */
 843	r = vhost_dev_check_owner(d);
 844	if (r)
 845		goto done;
 846
 847	switch (ioctl) {
 848	case VHOST_SET_MEM_TABLE:
 849		r = vhost_set_memory(d, argp);
 850		break;
 851	case VHOST_SET_LOG_BASE:
 852		if (copy_from_user(&p, argp, sizeof p)) {
 853			r = -EFAULT;
 854			break;
 855		}
 856		if ((u64)(unsigned long)p != p) {
 857			r = -EFAULT;
 858			break;
 859		}
 860		for (i = 0; i < d->nvqs; ++i) {
 861			struct vhost_virtqueue *vq;
 862			void __user *base = (void __user *)(unsigned long)p;
 863			vq = d->vqs + i;
 864			mutex_lock(&vq->mutex);
 865			/* If ring is inactive, will check when it's enabled. */
 866			if (vq->private_data && !vq_log_access_ok(d, vq, base))
 867				r = -EFAULT;
 868			else
 869				vq->log_base = base;
 870			mutex_unlock(&vq->mutex);
 871		}
 872		break;
 873	case VHOST_SET_LOG_FD:
 874		r = get_user(fd, (int __user *)argp);
 875		if (r < 0)
 876			break;
 877		eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 878		if (IS_ERR(eventfp)) {
 879			r = PTR_ERR(eventfp);
 880			break;
 881		}
 882		if (eventfp != d->log_file) {
 883			filep = d->log_file;
 884			ctx = d->log_ctx;
 885			d->log_ctx = eventfp ?
 886				eventfd_ctx_fileget(eventfp) : NULL;
 887		} else
 888			filep = eventfp;
 889		for (i = 0; i < d->nvqs; ++i) {
 890			mutex_lock(&d->vqs[i].mutex);
 891			d->vqs[i].log_ctx = d->log_ctx;
 892			mutex_unlock(&d->vqs[i].mutex);
 893		}
 894		if (ctx)
 895			eventfd_ctx_put(ctx);
 896		if (filep)
 897			fput(filep);
 898		break;
 899	default:
 900		r = vhost_set_vring(d, ioctl, argp);
 901		break;
 902	}
 903done:
 904	return r;
 905}
 906
 907static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 908						     __u64 addr, __u32 len)
 909{
 910	struct vhost_memory_region *reg;
 911	int i;
 912
 913	/* linear search is not brilliant, but we really have on the order of 6
 914	 * regions in practice */
 915	for (i = 0; i < mem->nregions; ++i) {
 916		reg = mem->regions + i;
 917		if (reg->guest_phys_addr <= addr &&
 918		    reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 919			return reg;
 920	}
 921	return NULL;
 922}
 923
 924/* TODO: This is really inefficient.  We need something like get_user()
 925 * (instruction directly accesses the data, with an exception table entry
 926 * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 927 */
 928static int set_bit_to_user(int nr, void __user *addr)
 929{
 930	unsigned long log = (unsigned long)addr;
 931	struct page *page;
 932	void *base;
 933	int bit = nr + (log % PAGE_SIZE) * 8;
 934	int r;
 935
 936	r = get_user_pages_fast(log, 1, 1, &page);
 937	if (r < 0)
 938		return r;
 939	BUG_ON(r != 1);
 940	base = kmap_atomic(page, KM_USER0);
 941	set_bit(bit, base);
 942	kunmap_atomic(base, KM_USER0);
 943	set_page_dirty_lock(page);
 944	put_page(page);
 945	return 0;
 946}
 947
 948static int log_write(void __user *log_base,
 949		     u64 write_address, u64 write_length)
 950{
 951	u64 write_page = write_address / VHOST_PAGE_SIZE;
 952	int r;
 953
 954	if (!write_length)
 955		return 0;
 956	write_length += write_address % VHOST_PAGE_SIZE;
 957	for (;;) {
 958		u64 base = (u64)(unsigned long)log_base;
 959		u64 log = base + write_page / 8;
 960		int bit = write_page % 8;
 961		if ((u64)(unsigned long)log != log)
 962			return -EFAULT;
 963		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 964		if (r < 0)
 965			return r;
 966		if (write_length <= VHOST_PAGE_SIZE)
 967			break;
 968		write_length -= VHOST_PAGE_SIZE;
 969		write_page += 1;
 970	}
 971	return r;
 972}
 973
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 974int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 975		    unsigned int log_num, u64 len)
 976{
 977	int i, r;
 978
 979	/* Make sure data written is seen before log. */
 980	smp_wmb();
 
 
 
 
 
 
 
 
 
 
 
 981	for (i = 0; i < log_num; ++i) {
 982		u64 l = min(log[i].len, len);
 983		r = log_write(vq->log_base, log[i].addr, l);
 984		if (r < 0)
 985			return r;
 986		len -= l;
 987		if (!len) {
 988			if (vq->log_ctx)
 989				eventfd_signal(vq->log_ctx, 1);
 990			return 0;
 991		}
 992	}
 993	/* Length written exceeds what we have stored. This is a bug. */
 994	BUG();
 995	return 0;
 996}
 
 997
 998static int vhost_update_used_flags(struct vhost_virtqueue *vq)
 999{
1000	void __user *used;
1001	if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1002		return -EFAULT;
1003	if (unlikely(vq->log_used)) {
1004		/* Make sure the flag is seen before log. */
1005		smp_wmb();
1006		/* Log used flag write. */
1007		used = &vq->used->flags;
1008		log_write(vq->log_base, vq->log_addr +
1009			  (used - (void __user *)vq->used),
1010			  sizeof vq->used->flags);
1011		if (vq->log_ctx)
1012			eventfd_signal(vq->log_ctx, 1);
1013	}
1014	return 0;
1015}
1016
1017static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1018{
1019	if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1020		return -EFAULT;
1021	if (unlikely(vq->log_used)) {
1022		void __user *used;
1023		/* Make sure the event is seen before log. */
1024		smp_wmb();
1025		/* Log avail event write */
1026		used = vhost_avail_event(vq);
1027		log_write(vq->log_base, vq->log_addr +
1028			  (used - (void __user *)vq->used),
1029			  sizeof *vhost_avail_event(vq));
1030		if (vq->log_ctx)
1031			eventfd_signal(vq->log_ctx, 1);
1032	}
1033	return 0;
1034}
1035
1036int vhost_init_used(struct vhost_virtqueue *vq)
1037{
 
1038	int r;
 
 
1039	if (!vq->private_data)
1040		return 0;
1041
 
 
1042	r = vhost_update_used_flags(vq);
1043	if (r)
1044		return r;
1045	vq->signalled_used_valid = false;
1046	return get_user(vq->last_used_idx, &vq->used->idx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1047}
 
1048
1049static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1050			  struct iovec iov[], int iov_size)
1051{
1052	const struct vhost_memory_region *reg;
1053	struct vhost_memory *mem;
 
1054	struct iovec *_iov;
1055	u64 s = 0;
1056	int ret = 0;
1057
1058	rcu_read_lock();
1059
1060	mem = rcu_dereference(dev->memory);
1061	while ((u64)len > s) {
1062		u64 size;
1063		if (unlikely(ret >= iov_size)) {
1064			ret = -ENOBUFS;
1065			break;
1066		}
1067		reg = find_region(mem, addr, len);
1068		if (unlikely(!reg)) {
1069			ret = -EFAULT;
 
 
 
 
 
 
 
 
1070			break;
1071		}
 
1072		_iov = iov + ret;
1073		size = reg->memory_size - addr + reg->guest_phys_addr;
1074		_iov->iov_len = min((u64)len, size);
1075		_iov->iov_base = (void __user *)(unsigned long)
1076			(reg->userspace_addr + addr - reg->guest_phys_addr);
1077		s += size;
1078		addr += size;
1079		++ret;
1080	}
1081
1082	rcu_read_unlock();
 
1083	return ret;
1084}
1085
1086/* Each buffer in the virtqueues is actually a chain of descriptors.  This
1087 * function returns the next descriptor in the chain,
1088 * or -1U if we're at the end. */
1089static unsigned next_desc(struct vring_desc *desc)
1090{
1091	unsigned int next;
1092
1093	/* If this descriptor says it doesn't chain, we're done. */
1094	if (!(desc->flags & VRING_DESC_F_NEXT))
1095		return -1U;
1096
1097	/* Check they're not leading us off end of descriptors. */
1098	next = desc->next;
1099	/* Make sure compiler knows to grab that: we don't want it changing! */
1100	/* We will use the result as an index in an array, so most
1101	 * architectures only need a compiler barrier here. */
1102	read_barrier_depends();
1103
1104	return next;
1105}
1106
1107static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1108			struct iovec iov[], unsigned int iov_size,
1109			unsigned int *out_num, unsigned int *in_num,
1110			struct vhost_log *log, unsigned int *log_num,
1111			struct vring_desc *indirect)
1112{
1113	struct vring_desc desc;
1114	unsigned int i = 0, count, found = 0;
1115	int ret;
 
 
1116
1117	/* Sanity check */
1118	if (unlikely(indirect->len % sizeof desc)) {
1119		vq_err(vq, "Invalid length in indirect descriptor: "
1120		       "len 0x%llx not multiple of 0x%zx\n",
1121		       (unsigned long long)indirect->len,
1122		       sizeof desc);
1123		return -EINVAL;
1124	}
1125
1126	ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1127			     UIO_MAXIOV);
1128	if (unlikely(ret < 0)) {
1129		vq_err(vq, "Translation failure %d in indirect.\n", ret);
 
1130		return ret;
1131	}
1132
1133	/* We will use the result as an address to read from, so most
1134	 * architectures only need a compiler barrier here. */
1135	read_barrier_depends();
1136
1137	count = indirect->len / sizeof desc;
1138	/* Buffers are chained via a 16 bit next field, so
1139	 * we can have at most 2^16 of these. */
1140	if (unlikely(count > USHRT_MAX + 1)) {
1141		vq_err(vq, "Indirect buffer length too big: %d\n",
1142		       indirect->len);
1143		return -E2BIG;
1144	}
1145
1146	do {
1147		unsigned iov_count = *in_num + *out_num;
1148		if (unlikely(++found > count)) {
1149			vq_err(vq, "Loop detected: last one at %u "
1150			       "indirect size %u\n",
1151			       i, count);
1152			return -EINVAL;
1153		}
1154		if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1155					      vq->indirect, sizeof desc))) {
1156			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1157			       i, (size_t)indirect->addr + i * sizeof desc);
1158			return -EINVAL;
1159		}
1160		if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1161			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1162			       i, (size_t)indirect->addr + i * sizeof desc);
1163			return -EINVAL;
1164		}
1165
1166		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1167				     iov_size - iov_count);
 
 
 
 
 
 
1168		if (unlikely(ret < 0)) {
1169			vq_err(vq, "Translation failure %d indirect idx %d\n",
1170			       ret, i);
 
1171			return ret;
1172		}
1173		/* If this is an input descriptor, increment that count. */
1174		if (desc.flags & VRING_DESC_F_WRITE) {
1175			*in_num += ret;
1176			if (unlikely(log)) {
1177				log[*log_num].addr = desc.addr;
1178				log[*log_num].len = desc.len;
1179				++*log_num;
1180			}
1181		} else {
1182			/* If it's an output descriptor, they're all supposed
1183			 * to come before any input descriptors. */
1184			if (unlikely(*in_num)) {
1185				vq_err(vq, "Indirect descriptor "
1186				       "has out after in: idx %d\n", i);
1187				return -EINVAL;
1188			}
1189			*out_num += ret;
1190		}
1191	} while ((i = next_desc(&desc)) != -1);
1192	return 0;
1193}
1194
1195/* This looks in the virtqueue and for the first available buffer, and converts
1196 * it to an iovec for convenient access.  Since descriptors consist of some
1197 * number of output then some number of input descriptors, it's actually two
1198 * iovecs, but we pack them into one and note how many of each there were.
1199 *
1200 * This function returns the descriptor number found, or vq->num (which is
1201 * never a valid descriptor number) if none was found.  A negative code is
1202 * returned on error. */
1203int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1204		      struct iovec iov[], unsigned int iov_size,
1205		      unsigned int *out_num, unsigned int *in_num,
1206		      struct vhost_log *log, unsigned int *log_num)
1207{
1208	struct vring_desc desc;
1209	unsigned int i, head, found = 0;
1210	u16 last_avail_idx;
1211	int ret;
 
 
1212
1213	/* Check it isn't doing very strange things with descriptor numbers. */
1214	last_avail_idx = vq->last_avail_idx;
1215	if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1216		vq_err(vq, "Failed to access avail idx at %p\n",
1217		       &vq->avail->idx);
1218		return -EFAULT;
1219	}
1220
1221	if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1222		vq_err(vq, "Guest moved used index from %u to %u",
1223		       last_avail_idx, vq->avail_idx);
1224		return -EFAULT;
1225	}
 
 
1226
1227	/* If there's nothing new since last we looked, return invalid. */
1228	if (vq->avail_idx == last_avail_idx)
1229		return vq->num;
 
 
1230
1231	/* Only get avail ring entries after they have been exposed by guest. */
1232	smp_rmb();
 
 
 
 
 
 
 
 
 
1233
1234	/* Grab the next descriptor number they're advertising, and increment
1235	 * the index we've seen. */
1236	if (unlikely(__get_user(head,
1237				&vq->avail->ring[last_avail_idx % vq->num]))) {
1238		vq_err(vq, "Failed to read head: idx %d address %p\n",
1239		       last_avail_idx,
1240		       &vq->avail->ring[last_avail_idx % vq->num]);
1241		return -EFAULT;
1242	}
1243
 
 
1244	/* If their number is silly, that's an error. */
1245	if (unlikely(head >= vq->num)) {
1246		vq_err(vq, "Guest says index %u > %u is available",
1247		       head, vq->num);
1248		return -EINVAL;
1249	}
1250
1251	/* When we start there are none of either input nor output. */
1252	*out_num = *in_num = 0;
1253	if (unlikely(log))
1254		*log_num = 0;
1255
1256	i = head;
1257	do {
1258		unsigned iov_count = *in_num + *out_num;
1259		if (unlikely(i >= vq->num)) {
1260			vq_err(vq, "Desc index is %u > %u, head = %u",
1261			       i, vq->num, head);
1262			return -EINVAL;
1263		}
1264		if (unlikely(++found > vq->num)) {
1265			vq_err(vq, "Loop detected: last one at %u "
1266			       "vq size %u head %u\n",
1267			       i, vq->num, head);
1268			return -EINVAL;
1269		}
1270		ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1271		if (unlikely(ret)) {
1272			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1273			       i, vq->desc + i);
1274			return -EFAULT;
1275		}
1276		if (desc.flags & VRING_DESC_F_INDIRECT) {
1277			ret = get_indirect(dev, vq, iov, iov_size,
1278					   out_num, in_num,
1279					   log, log_num, &desc);
1280			if (unlikely(ret < 0)) {
1281				vq_err(vq, "Failure detected "
1282				       "in indirect descriptor at idx %d\n", i);
 
1283				return ret;
1284			}
1285			continue;
1286		}
1287
1288		ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1289				     iov_size - iov_count);
 
 
 
 
 
1290		if (unlikely(ret < 0)) {
1291			vq_err(vq, "Translation failure %d descriptor idx %d\n",
1292			       ret, i);
 
1293			return ret;
1294		}
1295		if (desc.flags & VRING_DESC_F_WRITE) {
1296			/* If this is an input descriptor,
1297			 * increment that count. */
1298			*in_num += ret;
1299			if (unlikely(log)) {
1300				log[*log_num].addr = desc.addr;
1301				log[*log_num].len = desc.len;
1302				++*log_num;
1303			}
1304		} else {
1305			/* If it's an output descriptor, they're all supposed
1306			 * to come before any input descriptors. */
1307			if (unlikely(*in_num)) {
1308				vq_err(vq, "Descriptor has out after in: "
1309				       "idx %d\n", i);
1310				return -EINVAL;
1311			}
1312			*out_num += ret;
1313		}
1314	} while ((i = next_desc(&desc)) != -1);
1315
1316	/* On success, increment avail index. */
1317	vq->last_avail_idx++;
1318
1319	/* Assume notifications from guest are disabled at this point,
1320	 * if they aren't we would need to update avail_event index. */
1321	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1322	return head;
1323}
 
1324
1325/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1326void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1327{
1328	vq->last_avail_idx -= n;
1329}
 
1330
1331/* After we've used one of their buffers, we tell them about it.  We'll then
1332 * want to notify the guest, using eventfd. */
1333int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1334{
1335	struct vring_used_elem __user *used;
 
 
 
1336
1337	/* The virtqueue contains a ring of used buffers.  Get a pointer to the
1338	 * next entry in that used ring. */
1339	used = &vq->used->ring[vq->last_used_idx % vq->num];
1340	if (__put_user(head, &used->id)) {
1341		vq_err(vq, "Failed to write used id");
1342		return -EFAULT;
1343	}
1344	if (__put_user(len, &used->len)) {
1345		vq_err(vq, "Failed to write used len");
1346		return -EFAULT;
1347	}
1348	/* Make sure buffer is written before we update index. */
1349	smp_wmb();
1350	if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1351		vq_err(vq, "Failed to increment used idx");
1352		return -EFAULT;
1353	}
1354	if (unlikely(vq->log_used)) {
1355		/* Make sure data is seen before log. */
1356		smp_wmb();
1357		/* Log used ring entry write. */
1358		log_write(vq->log_base,
1359			  vq->log_addr +
1360			   ((void __user *)used - (void __user *)vq->used),
1361			  sizeof *used);
1362		/* Log used index update. */
1363		log_write(vq->log_base,
1364			  vq->log_addr + offsetof(struct vring_used, idx),
1365			  sizeof vq->used->idx);
1366		if (vq->log_ctx)
1367			eventfd_signal(vq->log_ctx, 1);
1368	}
1369	vq->last_used_idx++;
1370	/* If the driver never bothers to signal in a very long while,
1371	 * used index might wrap around. If that happens, invalidate
1372	 * signalled_used index we stored. TODO: make sure driver
1373	 * signals at least once in 2^16 and remove this. */
1374	if (unlikely(vq->last_used_idx == vq->signalled_used))
1375		vq->signalled_used_valid = false;
1376	return 0;
1377}
 
1378
1379static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1380			    struct vring_used_elem *heads,
1381			    unsigned count)
1382{
1383	struct vring_used_elem __user *used;
1384	u16 old, new;
1385	int start;
1386
1387	start = vq->last_used_idx % vq->num;
1388	used = vq->used->ring + start;
1389	if (__copy_to_user(used, heads, count * sizeof *used)) {
1390		vq_err(vq, "Failed to write used");
1391		return -EFAULT;
1392	}
1393	if (unlikely(vq->log_used)) {
1394		/* Make sure data is seen before log. */
1395		smp_wmb();
1396		/* Log used ring entry write. */
1397		log_write(vq->log_base,
1398			  vq->log_addr +
1399			   ((void __user *)used - (void __user *)vq->used),
1400			  count * sizeof *used);
1401	}
1402	old = vq->last_used_idx;
1403	new = (vq->last_used_idx += count);
1404	/* If the driver never bothers to signal in a very long while,
1405	 * used index might wrap around. If that happens, invalidate
1406	 * signalled_used index we stored. TODO: make sure driver
1407	 * signals at least once in 2^16 and remove this. */
1408	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1409		vq->signalled_used_valid = false;
1410	return 0;
1411}
1412
1413/* After we've used one of their buffers, we tell them about it.  We'll then
1414 * want to notify the guest, using eventfd. */
1415int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1416		     unsigned count)
1417{
1418	int start, n, r;
1419
1420	start = vq->last_used_idx % vq->num;
1421	n = vq->num - start;
1422	if (n < count) {
1423		r = __vhost_add_used_n(vq, heads, n);
1424		if (r < 0)
1425			return r;
1426		heads += n;
1427		count -= n;
1428	}
1429	r = __vhost_add_used_n(vq, heads, count);
1430
1431	/* Make sure buffer is written before we update index. */
1432	smp_wmb();
1433	if (put_user(vq->last_used_idx, &vq->used->idx)) {
1434		vq_err(vq, "Failed to increment used idx");
1435		return -EFAULT;
1436	}
1437	if (unlikely(vq->log_used)) {
 
 
1438		/* Log used index update. */
1439		log_write(vq->log_base,
1440			  vq->log_addr + offsetof(struct vring_used, idx),
1441			  sizeof vq->used->idx);
1442		if (vq->log_ctx)
1443			eventfd_signal(vq->log_ctx, 1);
1444	}
1445	return r;
1446}
 
1447
1448static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1449{
1450	__u16 old, new, event;
 
1451	bool v;
1452	/* Flush out used index updates. This is paired
1453	 * with the barrier that the Guest executes when enabling
1454	 * interrupts. */
1455	smp_mb();
1456
1457	if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1458	    unlikely(vq->avail_idx == vq->last_avail_idx))
1459		return true;
1460
1461	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1462		__u16 flags;
1463		if (__get_user(flags, &vq->avail->flags)) {
1464			vq_err(vq, "Failed to get flags");
1465			return true;
1466		}
1467		return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1468	}
1469	old = vq->signalled_used;
1470	v = vq->signalled_used_valid;
1471	new = vq->signalled_used = vq->last_used_idx;
1472	vq->signalled_used_valid = true;
1473
1474	if (unlikely(!v))
1475		return true;
1476
1477	if (get_user(event, vhost_used_event(vq))) {
1478		vq_err(vq, "Failed to get used event idx");
1479		return true;
1480	}
1481	return vring_need_event(event, new, old);
1482}
1483
1484/* This actually signals the guest, using eventfd. */
1485void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1486{
1487	/* Signal the Guest tell them we used something up. */
1488	if (vq->call_ctx && vhost_notify(dev, vq))
1489		eventfd_signal(vq->call_ctx, 1);
1490}
 
1491
1492/* And here's the combo meal deal.  Supersize me! */
1493void vhost_add_used_and_signal(struct vhost_dev *dev,
1494			       struct vhost_virtqueue *vq,
1495			       unsigned int head, int len)
1496{
1497	vhost_add_used(vq, head, len);
1498	vhost_signal(dev, vq);
1499}
 
1500
1501/* multi-buffer version of vhost_add_used_and_signal */
1502void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1503				 struct vhost_virtqueue *vq,
1504				 struct vring_used_elem *heads, unsigned count)
1505{
1506	vhost_add_used_n(vq, heads, count);
1507	vhost_signal(dev, vq);
1508}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1509
1510/* OK, now we need to know about added descriptors. */
1511bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1512{
1513	u16 avail_idx;
1514	int r;
1515
1516	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1517		return false;
1518	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1519	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1520		r = vhost_update_used_flags(vq);
1521		if (r) {
1522			vq_err(vq, "Failed to enable notification at %p: %d\n",
1523			       &vq->used->flags, r);
1524			return false;
1525		}
1526	} else {
1527		r = vhost_update_avail_event(vq, vq->avail_idx);
1528		if (r) {
1529			vq_err(vq, "Failed to update avail event index at %p: %d\n",
1530			       vhost_avail_event(vq), r);
1531			return false;
1532		}
1533	}
1534	/* They could have slipped one in as we were doing that: make
1535	 * sure it's written, then check again. */
1536	smp_mb();
1537	r = __get_user(avail_idx, &vq->avail->idx);
1538	if (r) {
1539		vq_err(vq, "Failed to check avail idx at %p: %d\n",
1540		       &vq->avail->idx, r);
1541		return false;
1542	}
1543
1544	return avail_idx != vq->avail_idx;
1545}
 
1546
1547/* We don't need to be notified again. */
1548void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1549{
1550	int r;
1551
1552	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1553		return;
1554	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1555	if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1556		r = vhost_update_used_flags(vq);
1557		if (r)
1558			vq_err(vq, "Failed to enable notification at %p: %d\n",
1559			       &vq->used->flags, r);
1560	}
1561}
 
1562
1563static void vhost_zerocopy_done_signal(struct kref *kref)
 
1564{
1565	struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1566						    kref);
1567	wake_up(&ubufs->wait);
 
 
 
 
 
 
1568}
 
1569
1570struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1571					bool zcopy)
1572{
1573	struct vhost_ubuf_ref *ubufs;
1574	/* No zero copy backend? Nothing to count. */
1575	if (!zcopy)
1576		return NULL;
1577	ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1578	if (!ubufs)
1579		return ERR_PTR(-ENOMEM);
1580	kref_init(&ubufs->kref);
1581	init_waitqueue_head(&ubufs->wait);
1582	ubufs->vq = vq;
1583	return ubufs;
1584}
 
1585
1586void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
 
1587{
1588	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
 
 
 
 
 
 
 
 
 
 
1589}
 
1590
1591void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1592{
1593	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1594	wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1595	kfree(ubufs);
 
 
 
 
 
 
 
 
1596}
 
1597
1598void vhost_zerocopy_callback(void *arg)
1599{
1600	struct ubuf_info *ubuf = arg;
1601	struct vhost_ubuf_ref *ubufs = ubuf->arg;
1602	struct vhost_virtqueue *vq = ubufs->vq;
1603
1604	/* set len = 1 to mark this desc buffers done DMA */
1605	vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1606	kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1607}
v5.14.15
   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_work_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_work_dev_flush);
 247
 248/* Flush any work that has been scheduled. When calling this, don't hold any
 249 * locks that are also used by the callback. */
 250void vhost_poll_flush(struct vhost_poll *poll)
 251{
 252	vhost_work_dev_flush(poll->dev);
 253}
 254EXPORT_SYMBOL_GPL(vhost_poll_flush);
 255
 256void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 
 257{
 258	if (!dev->worker)
 259		return;
 260
 261	if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 262		/* We can only add the work to the list after we're
 263		 * sure it was not in the list.
 264		 * test_and_set_bit() implies a memory barrier.
 265		 */
 266		llist_add(&work->node, &dev->work_list);
 267		wake_up_process(dev->worker);
 268	}
 
 269}
 270EXPORT_SYMBOL_GPL(vhost_work_queue);
 271
 272/* A lockless hint for busy polling code to exit the loop */
 273bool vhost_has_work(struct vhost_dev *dev)
 274{
 275	return !llist_empty(&dev->work_list);
 276}
 277EXPORT_SYMBOL_GPL(vhost_has_work);
 278
 279void vhost_poll_queue(struct vhost_poll *poll)
 280{
 281	vhost_work_queue(poll->dev, &poll->work);
 282}
 283EXPORT_SYMBOL_GPL(vhost_poll_queue);
 284
 285static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 286{
 287	int j;
 288
 289	for (j = 0; j < VHOST_NUM_ADDRS; j++)
 290		vq->meta_iotlb[j] = NULL;
 291}
 292
 293static void vhost_vq_meta_reset(struct vhost_dev *d)
 294{
 295	int i;
 296
 297	for (i = 0; i < d->nvqs; ++i)
 298		__vhost_vq_meta_reset(d->vqs[i]);
 299}
 300
 301static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 302{
 303	call_ctx->ctx = NULL;
 304	memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 305}
 306
 307bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 308{
 309	return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 310}
 311EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 312
 313static void vhost_vq_reset(struct vhost_dev *dev,
 314			   struct vhost_virtqueue *vq)
 315{
 316	vq->num = 1;
 317	vq->desc = NULL;
 318	vq->avail = NULL;
 319	vq->used = NULL;
 320	vq->last_avail_idx = 0;
 321	vq->avail_idx = 0;
 322	vq->last_used_idx = 0;
 323	vq->signalled_used = 0;
 324	vq->signalled_used_valid = false;
 325	vq->used_flags = 0;
 326	vq->log_used = false;
 327	vq->log_addr = -1ull;
 
 
 328	vq->private_data = NULL;
 329	vq->acked_features = 0;
 330	vq->acked_backend_features = 0;
 331	vq->log_base = NULL;
 332	vq->error_ctx = NULL;
 
 333	vq->kick = NULL;
 
 
 334	vq->log_ctx = NULL;
 335	vhost_disable_cross_endian(vq);
 336	vhost_reset_is_le(vq);
 337	vq->busyloop_timeout = 0;
 338	vq->umem = NULL;
 339	vq->iotlb = NULL;
 340	vhost_vring_call_reset(&vq->call_ctx);
 341	__vhost_vq_meta_reset(vq);
 342}
 343
 344static int vhost_worker(void *data)
 345{
 346	struct vhost_dev *dev = data;
 347	struct vhost_work *work, *work_next;
 348	struct llist_node *node;
 349
 350	kthread_use_mm(dev->mm);
 351
 352	for (;;) {
 353		/* mb paired w/ kthread_stop */
 354		set_current_state(TASK_INTERRUPTIBLE);
 355
 
 
 
 
 
 
 
 356		if (kthread_should_stop()) {
 
 357			__set_current_state(TASK_RUNNING);
 358			break;
 359		}
 
 
 
 
 
 
 
 
 360
 361		node = llist_del_all(&dev->work_list);
 362		if (!node)
 
 
 363			schedule();
 364
 365		node = llist_reverse_order(node);
 366		/* make sure flag is seen after deletion */
 367		smp_wmb();
 368		llist_for_each_entry_safe(work, work_next, node, node) {
 369			clear_bit(VHOST_WORK_QUEUED, &work->flags);
 370			__set_current_state(TASK_RUNNING);
 371			kcov_remote_start_common(dev->kcov_handle);
 372			work->fn(work);
 373			kcov_remote_stop();
 374			if (need_resched())
 375				schedule();
 376		}
 377	}
 378	kthread_unuse_mm(dev->mm);
 379	return 0;
 380}
 381
 382static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 383{
 384	kfree(vq->indirect);
 385	vq->indirect = NULL;
 386	kfree(vq->log);
 387	vq->log = NULL;
 388	kfree(vq->heads);
 389	vq->heads = NULL;
 
 
 
 
 
 
 
 390}
 391
 392/* Helper to allocate iovec buffers for all vqs. */
 393static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 394{
 395	struct vhost_virtqueue *vq;
 396	int i;
 
 397
 398	for (i = 0; i < dev->nvqs; ++i) {
 399		vq = dev->vqs[i];
 400		vq->indirect = kmalloc_array(UIO_MAXIOV,
 401					     sizeof(*vq->indirect),
 402					     GFP_KERNEL);
 403		vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 404					GFP_KERNEL);
 405		vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 406					  GFP_KERNEL);
 407		if (!vq->indirect || !vq->log || !vq->heads)
 
 
 
 
 
 
 
 
 
 408			goto err_nomem;
 409	}
 410	return 0;
 411
 412err_nomem:
 413	for (; i >= 0; --i)
 414		vhost_vq_free_iovecs(dev->vqs[i]);
 415	return -ENOMEM;
 416}
 417
 418static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 419{
 420	int i;
 421
 422	for (i = 0; i < dev->nvqs; ++i)
 423		vhost_vq_free_iovecs(dev->vqs[i]);
 424}
 425
 426bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 427			  int pkts, int total_len)
 428{
 429	struct vhost_dev *dev = vq->dev;
 430
 431	if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 432	    pkts >= dev->weight) {
 433		vhost_poll_queue(&vq->poll);
 434		return true;
 435	}
 436
 437	return false;
 438}
 439EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 440
 441static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 442				   unsigned int num)
 443{
 444	size_t event __maybe_unused =
 445	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 446
 447	return sizeof(*vq->avail) +
 448	       sizeof(*vq->avail->ring) * num + event;
 449}
 450
 451static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 452				  unsigned int num)
 453{
 454	size_t event __maybe_unused =
 455	       vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 456
 457	return sizeof(*vq->used) +
 458	       sizeof(*vq->used->ring) * num + event;
 459}
 460
 461static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 462				  unsigned int num)
 463{
 464	return sizeof(*vq->desc) * num;
 465}
 466
 467void vhost_dev_init(struct vhost_dev *dev,
 468		    struct vhost_virtqueue **vqs, int nvqs,
 469		    int iov_limit, int weight, int byte_weight,
 470		    bool use_worker,
 471		    int (*msg_handler)(struct vhost_dev *dev,
 472				       struct vhost_iotlb_msg *msg))
 473{
 474	struct vhost_virtqueue *vq;
 475	int i;
 476
 477	dev->vqs = vqs;
 478	dev->nvqs = nvqs;
 479	mutex_init(&dev->mutex);
 480	dev->log_ctx = NULL;
 481	dev->umem = NULL;
 482	dev->iotlb = NULL;
 483	dev->mm = NULL;
 
 
 484	dev->worker = NULL;
 485	dev->iov_limit = iov_limit;
 486	dev->weight = weight;
 487	dev->byte_weight = byte_weight;
 488	dev->use_worker = use_worker;
 489	dev->msg_handler = msg_handler;
 490	init_llist_head(&dev->work_list);
 491	init_waitqueue_head(&dev->wait);
 492	INIT_LIST_HEAD(&dev->read_list);
 493	INIT_LIST_HEAD(&dev->pending_list);
 494	spin_lock_init(&dev->iotlb_lock);
 495
 496
 497	for (i = 0; i < dev->nvqs; ++i) {
 498		vq = dev->vqs[i];
 499		vq->log = NULL;
 500		vq->indirect = NULL;
 501		vq->heads = NULL;
 502		vq->dev = dev;
 503		mutex_init(&vq->mutex);
 504		vhost_vq_reset(dev, vq);
 505		if (vq->handle_kick)
 506			vhost_poll_init(&vq->poll, vq->handle_kick,
 507					EPOLLIN, dev);
 508	}
 
 
 509}
 510EXPORT_SYMBOL_GPL(vhost_dev_init);
 511
 512/* Caller should have device mutex */
 513long vhost_dev_check_owner(struct vhost_dev *dev)
 514{
 515	/* Are you the owner? If not, I don't think you mean to do that */
 516	return dev->mm == current->mm ? 0 : -EPERM;
 517}
 518EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 519
 520struct vhost_attach_cgroups_struct {
 521	struct vhost_work work;
 522	struct task_struct *owner;
 523	int ret;
 524};
 525
 526static void vhost_attach_cgroups_work(struct vhost_work *work)
 527{
 528	struct vhost_attach_cgroups_struct *s;
 529
 530	s = container_of(work, struct vhost_attach_cgroups_struct, work);
 531	s->ret = cgroup_attach_task_all(s->owner, current);
 532}
 533
 534static int vhost_attach_cgroups(struct vhost_dev *dev)
 535{
 536	struct vhost_attach_cgroups_struct attach;
 537
 538	attach.owner = current;
 539	vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 540	vhost_work_queue(dev, &attach.work);
 541	vhost_work_dev_flush(dev);
 542	return attach.ret;
 543}
 544
 545/* Caller should have device mutex */
 546bool vhost_dev_has_owner(struct vhost_dev *dev)
 547{
 548	return dev->mm;
 549}
 550EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 551
 552static void vhost_attach_mm(struct vhost_dev *dev)
 553{
 554	/* No owner, become one */
 555	if (dev->use_worker) {
 556		dev->mm = get_task_mm(current);
 557	} else {
 558		/* vDPA device does not use worker thead, so there's
 559		 * no need to hold the address space for mm. This help
 560		 * to avoid deadlock in the case of mmap() which may
 561		 * held the refcnt of the file and depends on release
 562		 * method to remove vma.
 563		 */
 564		dev->mm = current->mm;
 565		mmgrab(dev->mm);
 566	}
 567}
 568
 569static void vhost_detach_mm(struct vhost_dev *dev)
 570{
 571	if (!dev->mm)
 572		return;
 573
 574	if (dev->use_worker)
 575		mmput(dev->mm);
 576	else
 577		mmdrop(dev->mm);
 578
 579	dev->mm = NULL;
 580}
 581
 582/* Caller should have device mutex */
 583long vhost_dev_set_owner(struct vhost_dev *dev)
 584{
 585	struct task_struct *worker;
 586	int err;
 587
 588	/* Is there an owner already? */
 589	if (vhost_dev_has_owner(dev)) {
 590		err = -EBUSY;
 591		goto err_mm;
 592	}
 593
 594	vhost_attach_mm(dev);
 
 
 
 
 
 
 
 
 
 595
 596	dev->kcov_handle = kcov_common_handle();
 597	if (dev->use_worker) {
 598		worker = kthread_create(vhost_worker, dev,
 599					"vhost-%d", current->pid);
 600		if (IS_ERR(worker)) {
 601			err = PTR_ERR(worker);
 602			goto err_worker;
 603		}
 604
 605		dev->worker = worker;
 606		wake_up_process(worker); /* avoid contributing to loadavg */
 607
 608		err = vhost_attach_cgroups(dev);
 609		if (err)
 610			goto err_cgroup;
 611	}
 612
 613	err = vhost_dev_alloc_iovecs(dev);
 614	if (err)
 615		goto err_cgroup;
 616
 617	return 0;
 618err_cgroup:
 619	if (dev->worker) {
 620		kthread_stop(dev->worker);
 621		dev->worker = NULL;
 622	}
 623err_worker:
 624	vhost_detach_mm(dev);
 625	dev->kcov_handle = 0;
 
 626err_mm:
 627	return err;
 628}
 629EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 630
 631static struct vhost_iotlb *iotlb_alloc(void)
 
 632{
 633	return vhost_iotlb_alloc(max_iotlb_entries,
 634				 VHOST_IOTLB_FLAG_RETIRE);
 635}
 636
 637struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 638{
 639	return iotlb_alloc();
 640}
 641EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 642
 643/* Caller should have device mutex */
 644void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 645{
 646	int i;
 647
 648	vhost_dev_cleanup(dev);
 649
 650	dev->umem = umem;
 651	/* We don't need VQ locks below since vhost_dev_cleanup makes sure
 652	 * VQs aren't running.
 653	 */
 654	for (i = 0; i < dev->nvqs; ++i)
 655		dev->vqs[i]->umem = umem;
 656}
 657EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 658
 659void vhost_dev_stop(struct vhost_dev *dev)
 
 
 
 
 
 660{
 661	int i;
 
 662
 663	for (i = 0; i < dev->nvqs; ++i) {
 664		if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 665			vhost_poll_stop(&dev->vqs[i]->poll);
 666			vhost_poll_flush(&dev->vqs[i]->poll);
 667		}
 
 
 
 668	}
 669}
 670EXPORT_SYMBOL_GPL(vhost_dev_stop);
 671
 672static void vhost_clear_msg(struct vhost_dev *dev)
 673{
 674	struct vhost_msg_node *node, *n;
 675
 676	spin_lock(&dev->iotlb_lock);
 677
 678	list_for_each_entry_safe(node, n, &dev->read_list, node) {
 679		list_del(&node->node);
 680		kfree(node);
 681	}
 682
 683	list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 684		list_del(&node->node);
 685		kfree(node);
 686	}
 687
 688	spin_unlock(&dev->iotlb_lock);
 689}
 690
 
 691void vhost_dev_cleanup(struct vhost_dev *dev)
 692{
 693	int i;
 694
 695	for (i = 0; i < dev->nvqs; ++i) {
 696		if (dev->vqs[i]->error_ctx)
 697			eventfd_ctx_put(dev->vqs[i]->error_ctx);
 698		if (dev->vqs[i]->kick)
 699			fput(dev->vqs[i]->kick);
 700		if (dev->vqs[i]->call_ctx.ctx)
 701			eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 702		vhost_vq_reset(dev, dev->vqs[i]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 703	}
 704	vhost_dev_free_iovecs(dev);
 705	if (dev->log_ctx)
 706		eventfd_ctx_put(dev->log_ctx);
 707	dev->log_ctx = NULL;
 
 
 
 708	/* No one will access memory at this point */
 709	vhost_iotlb_free(dev->umem);
 710	dev->umem = NULL;
 711	vhost_iotlb_free(dev->iotlb);
 712	dev->iotlb = NULL;
 713	vhost_clear_msg(dev);
 714	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 715	WARN_ON(!llist_empty(&dev->work_list));
 716	if (dev->worker) {
 717		kthread_stop(dev->worker);
 718		dev->worker = NULL;
 719		dev->kcov_handle = 0;
 720	}
 721	vhost_detach_mm(dev);
 
 
 722}
 723EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 724
 725static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 726{
 727	u64 a = addr / VHOST_PAGE_SIZE / 8;
 728
 729	/* Make sure 64 bit math will not overflow. */
 730	if (a > ULONG_MAX - (unsigned long)log_base ||
 731	    a + (unsigned long)log_base > ULONG_MAX)
 732		return false;
 733
 734	return access_ok(log_base + a,
 735			 (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 736}
 737
 738/* Make sure 64 bit math will not overflow. */
 739static bool vhost_overflow(u64 uaddr, u64 size)
 740{
 741	if (uaddr > ULONG_MAX || size > ULONG_MAX)
 742		return true;
 743
 744	if (!size)
 745		return false;
 746
 747	return uaddr > ULONG_MAX - size + 1;
 748}
 749
 750/* Caller should have vq mutex and device mutex. */
 751static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 752				int log_all)
 753{
 754	struct vhost_iotlb_map *map;
 755
 756	if (!umem)
 757		return false;
 758
 759	list_for_each_entry(map, &umem->list, link) {
 760		unsigned long a = map->addr;
 761
 762		if (vhost_overflow(map->addr, map->size))
 763			return false;
 764
 765
 766		if (!access_ok((void __user *)a, map->size))
 767			return false;
 768		else if (log_all && !log_access_ok(log_base,
 769						   map->start,
 770						   map->size))
 771			return false;
 772	}
 773	return true;
 774}
 775
 776static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 777					       u64 addr, unsigned int size,
 778					       int type)
 779{
 780	const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 781
 782	if (!map)
 783		return NULL;
 784
 785	return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 786}
 787
 788/* Can we switch to this memory table? */
 789/* Caller should have device mutex but not vq mutex */
 790static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 791			     int log_all)
 792{
 793	int i;
 794
 795	for (i = 0; i < d->nvqs; ++i) {
 796		bool ok;
 797		bool log;
 798
 799		mutex_lock(&d->vqs[i]->mutex);
 800		log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 801		/* If ring is inactive, will check when it's enabled. */
 802		if (d->vqs[i]->private_data)
 803			ok = vq_memory_access_ok(d->vqs[i]->log_base,
 804						 umem, log);
 805		else
 806			ok = true;
 807		mutex_unlock(&d->vqs[i]->mutex);
 808		if (!ok)
 809			return false;
 810	}
 811	return true;
 812}
 813
 814static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 815			  struct iovec iov[], int iov_size, int access);
 816
 817static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 818			      const void *from, unsigned size)
 819{
 820	int ret;
 821
 822	if (!vq->iotlb)
 823		return __copy_to_user(to, from, size);
 824	else {
 825		/* This function should be called after iotlb
 826		 * prefetch, which means we're sure that all vq
 827		 * could be access through iotlb. So -EAGAIN should
 828		 * not happen in this case.
 829		 */
 830		struct iov_iter t;
 831		void __user *uaddr = vhost_vq_meta_fetch(vq,
 832				     (u64)(uintptr_t)to, size,
 833				     VHOST_ADDR_USED);
 834
 835		if (uaddr)
 836			return __copy_to_user(uaddr, from, size);
 837
 838		ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 839				     ARRAY_SIZE(vq->iotlb_iov),
 840				     VHOST_ACCESS_WO);
 841		if (ret < 0)
 842			goto out;
 843		iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 844		ret = copy_to_iter(from, size, &t);
 845		if (ret == size)
 846			ret = 0;
 847	}
 848out:
 849	return ret;
 850}
 851
 852static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 853				void __user *from, unsigned size)
 854{
 855	int ret;
 856
 857	if (!vq->iotlb)
 858		return __copy_from_user(to, from, size);
 859	else {
 860		/* This function should be called after iotlb
 861		 * prefetch, which means we're sure that vq
 862		 * could be access through iotlb. So -EAGAIN should
 863		 * not happen in this case.
 864		 */
 865		void __user *uaddr = vhost_vq_meta_fetch(vq,
 866				     (u64)(uintptr_t)from, size,
 867				     VHOST_ADDR_DESC);
 868		struct iov_iter f;
 869
 870		if (uaddr)
 871			return __copy_from_user(to, uaddr, size);
 872
 873		ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 874				     ARRAY_SIZE(vq->iotlb_iov),
 875				     VHOST_ACCESS_RO);
 876		if (ret < 0) {
 877			vq_err(vq, "IOTLB translation failure: uaddr "
 878			       "%p size 0x%llx\n", from,
 879			       (unsigned long long) size);
 880			goto out;
 881		}
 882		iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 883		ret = copy_from_iter(to, size, &f);
 884		if (ret == size)
 885			ret = 0;
 886	}
 887
 888out:
 889	return ret;
 890}
 891
 892static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 893					  void __user *addr, unsigned int size,
 894					  int type)
 895{
 896	int ret;
 897
 898	ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 899			     ARRAY_SIZE(vq->iotlb_iov),
 900			     VHOST_ACCESS_RO);
 901	if (ret < 0) {
 902		vq_err(vq, "IOTLB translation failure: uaddr "
 903			"%p size 0x%llx\n", addr,
 904			(unsigned long long) size);
 905		return NULL;
 906	}
 907
 908	if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 909		vq_err(vq, "Non atomic userspace memory access: uaddr "
 910			"%p size 0x%llx\n", addr,
 911			(unsigned long long) size);
 912		return NULL;
 913	}
 914
 915	return vq->iotlb_iov[0].iov_base;
 916}
 917
 918/* This function should be called after iotlb
 919 * prefetch, which means we're sure that vq
 920 * could be access through iotlb. So -EAGAIN should
 921 * not happen in this case.
 922 */
 923static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 924					    void __user *addr, unsigned int size,
 925					    int type)
 926{
 927	void __user *uaddr = vhost_vq_meta_fetch(vq,
 928			     (u64)(uintptr_t)addr, size, type);
 929	if (uaddr)
 930		return uaddr;
 931
 932	return __vhost_get_user_slow(vq, addr, size, type);
 933}
 934
 935#define vhost_put_user(vq, x, ptr)		\
 936({ \
 937	int ret; \
 938	if (!vq->iotlb) { \
 939		ret = __put_user(x, ptr); \
 940	} else { \
 941		__typeof__(ptr) to = \
 942			(__typeof__(ptr)) __vhost_get_user(vq, ptr,	\
 943					  sizeof(*ptr), VHOST_ADDR_USED); \
 944		if (to != NULL) \
 945			ret = __put_user(x, to); \
 946		else \
 947			ret = -EFAULT;	\
 948	} \
 949	ret; \
 950})
 951
 952static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 953{
 954	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 955			      vhost_avail_event(vq));
 956}
 957
 958static inline int vhost_put_used(struct vhost_virtqueue *vq,
 959				 struct vring_used_elem *head, int idx,
 960				 int count)
 961{
 962	return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 963				  count * sizeof(*head));
 964}
 965
 966static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 967
 968{
 969	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 970			      &vq->used->flags);
 971}
 972
 973static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 974
 975{
 976	return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 977			      &vq->used->idx);
 978}
 979
 980#define vhost_get_user(vq, x, ptr, type)		\
 981({ \
 982	int ret; \
 983	if (!vq->iotlb) { \
 984		ret = __get_user(x, ptr); \
 985	} else { \
 986		__typeof__(ptr) from = \
 987			(__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 988							   sizeof(*ptr), \
 989							   type); \
 990		if (from != NULL) \
 991			ret = __get_user(x, from); \
 992		else \
 993			ret = -EFAULT; \
 994	} \
 995	ret; \
 996})
 997
 998#define vhost_get_avail(vq, x, ptr) \
 999	vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1000
1001#define vhost_get_used(vq, x, ptr) \
1002	vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1003
1004static void vhost_dev_lock_vqs(struct vhost_dev *d)
1005{
1006	int i = 0;
1007	for (i = 0; i < d->nvqs; ++i)
1008		mutex_lock_nested(&d->vqs[i]->mutex, i);
1009}
1010
1011static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1012{
1013	int i = 0;
1014	for (i = 0; i < d->nvqs; ++i)
1015		mutex_unlock(&d->vqs[i]->mutex);
1016}
1017
1018static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1019				      __virtio16 *idx)
1020{
1021	return vhost_get_avail(vq, *idx, &vq->avail->idx);
1022}
1023
1024static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1025				       __virtio16 *head, int idx)
1026{
1027	return vhost_get_avail(vq, *head,
1028			       &vq->avail->ring[idx & (vq->num - 1)]);
1029}
1030
1031static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1032					__virtio16 *flags)
1033{
1034	return vhost_get_avail(vq, *flags, &vq->avail->flags);
1035}
1036
1037static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1038				       __virtio16 *event)
1039{
1040	return vhost_get_avail(vq, *event, vhost_used_event(vq));
1041}
1042
1043static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1044				     __virtio16 *idx)
1045{
1046	return vhost_get_used(vq, *idx, &vq->used->idx);
1047}
1048
1049static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1050				 struct vring_desc *desc, int idx)
1051{
1052	return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1053}
1054
1055static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1056				  struct vhost_iotlb_msg *msg)
1057{
1058	struct vhost_msg_node *node, *n;
1059
1060	spin_lock(&d->iotlb_lock);
1061
1062	list_for_each_entry_safe(node, n, &d->pending_list, node) {
1063		struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1064		if (msg->iova <= vq_msg->iova &&
1065		    msg->iova + msg->size - 1 >= vq_msg->iova &&
1066		    vq_msg->type == VHOST_IOTLB_MISS) {
1067			vhost_poll_queue(&node->vq->poll);
1068			list_del(&node->node);
1069			kfree(node);
1070		}
1071	}
1072
1073	spin_unlock(&d->iotlb_lock);
1074}
1075
1076static bool umem_access_ok(u64 uaddr, u64 size, int access)
1077{
1078	unsigned long a = uaddr;
1079
1080	/* Make sure 64 bit math will not overflow. */
1081	if (vhost_overflow(uaddr, size))
1082		return false;
1083
1084	if ((access & VHOST_ACCESS_RO) &&
1085	    !access_ok((void __user *)a, size))
1086		return false;
1087	if ((access & VHOST_ACCESS_WO) &&
1088	    !access_ok((void __user *)a, size))
1089		return false;
1090	return true;
1091}
1092
1093static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1094				   struct vhost_iotlb_msg *msg)
1095{
1096	int ret = 0;
1097
1098	mutex_lock(&dev->mutex);
1099	vhost_dev_lock_vqs(dev);
1100	switch (msg->type) {
1101	case VHOST_IOTLB_UPDATE:
1102		if (!dev->iotlb) {
1103			ret = -EFAULT;
1104			break;
1105		}
1106		if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1107			ret = -EFAULT;
1108			break;
1109		}
1110		vhost_vq_meta_reset(dev);
1111		if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1112					  msg->iova + msg->size - 1,
1113					  msg->uaddr, msg->perm)) {
1114			ret = -ENOMEM;
1115			break;
1116		}
1117		vhost_iotlb_notify_vq(dev, msg);
1118		break;
1119	case VHOST_IOTLB_INVALIDATE:
1120		if (!dev->iotlb) {
1121			ret = -EFAULT;
1122			break;
1123		}
1124		vhost_vq_meta_reset(dev);
1125		vhost_iotlb_del_range(dev->iotlb, msg->iova,
1126				      msg->iova + msg->size - 1);
1127		break;
1128	default:
1129		ret = -EINVAL;
1130		break;
1131	}
1132
1133	vhost_dev_unlock_vqs(dev);
1134	mutex_unlock(&dev->mutex);
1135
1136	return ret;
1137}
1138ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1139			     struct iov_iter *from)
1140{
1141	struct vhost_iotlb_msg msg;
1142	size_t offset;
1143	int type, ret;
1144
1145	ret = copy_from_iter(&type, sizeof(type), from);
1146	if (ret != sizeof(type)) {
1147		ret = -EINVAL;
1148		goto done;
1149	}
1150
1151	switch (type) {
1152	case VHOST_IOTLB_MSG:
1153		/* There maybe a hole after type for V1 message type,
1154		 * so skip it here.
1155		 */
1156		offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1157		break;
1158	case VHOST_IOTLB_MSG_V2:
1159		offset = sizeof(__u32);
1160		break;
1161	default:
1162		ret = -EINVAL;
1163		goto done;
1164	}
1165
1166	iov_iter_advance(from, offset);
1167	ret = copy_from_iter(&msg, sizeof(msg), from);
1168	if (ret != sizeof(msg)) {
1169		ret = -EINVAL;
1170		goto done;
1171	}
1172
1173	if (dev->msg_handler)
1174		ret = dev->msg_handler(dev, &msg);
1175	else
1176		ret = vhost_process_iotlb_msg(dev, &msg);
1177	if (ret) {
1178		ret = -EFAULT;
1179		goto done;
1180	}
1181
1182	ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1183	      sizeof(struct vhost_msg_v2);
1184done:
1185	return ret;
1186}
1187EXPORT_SYMBOL(vhost_chr_write_iter);
1188
1189__poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1190			    poll_table *wait)
1191{
1192	__poll_t mask = 0;
1193
1194	poll_wait(file, &dev->wait, wait);
1195
1196	if (!list_empty(&dev->read_list))
1197		mask |= EPOLLIN | EPOLLRDNORM;
1198
1199	return mask;
1200}
1201EXPORT_SYMBOL(vhost_chr_poll);
1202
1203ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1204			    int noblock)
1205{
1206	DEFINE_WAIT(wait);
1207	struct vhost_msg_node *node;
1208	ssize_t ret = 0;
1209	unsigned size = sizeof(struct vhost_msg);
1210
1211	if (iov_iter_count(to) < size)
1212		return 0;
1213
1214	while (1) {
1215		if (!noblock)
1216			prepare_to_wait(&dev->wait, &wait,
1217					TASK_INTERRUPTIBLE);
1218
1219		node = vhost_dequeue_msg(dev, &dev->read_list);
1220		if (node)
1221			break;
1222		if (noblock) {
1223			ret = -EAGAIN;
1224			break;
1225		}
1226		if (signal_pending(current)) {
1227			ret = -ERESTARTSYS;
1228			break;
1229		}
1230		if (!dev->iotlb) {
1231			ret = -EBADFD;
1232			break;
1233		}
1234
1235		schedule();
1236	}
1237
1238	if (!noblock)
1239		finish_wait(&dev->wait, &wait);
1240
1241	if (node) {
1242		struct vhost_iotlb_msg *msg;
1243		void *start = &node->msg;
1244
1245		switch (node->msg.type) {
1246		case VHOST_IOTLB_MSG:
1247			size = sizeof(node->msg);
1248			msg = &node->msg.iotlb;
1249			break;
1250		case VHOST_IOTLB_MSG_V2:
1251			size = sizeof(node->msg_v2);
1252			msg = &node->msg_v2.iotlb;
1253			break;
1254		default:
1255			BUG();
1256			break;
1257		}
1258
1259		ret = copy_to_iter(start, size, to);
1260		if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1261			kfree(node);
1262			return ret;
1263		}
1264		vhost_enqueue_msg(dev, &dev->pending_list, node);
1265	}
1266
1267	return ret;
1268}
1269EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1270
1271static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1272{
1273	struct vhost_dev *dev = vq->dev;
1274	struct vhost_msg_node *node;
1275	struct vhost_iotlb_msg *msg;
1276	bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1277
1278	node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1279	if (!node)
1280		return -ENOMEM;
1281
1282	if (v2) {
1283		node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1284		msg = &node->msg_v2.iotlb;
1285	} else {
1286		msg = &node->msg.iotlb;
1287	}
1288
1289	msg->type = VHOST_IOTLB_MISS;
1290	msg->iova = iova;
1291	msg->perm = access;
1292
1293	vhost_enqueue_msg(dev, &dev->read_list, node);
1294
1295	return 0;
1296}
1297
1298static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1299			 vring_desc_t __user *desc,
1300			 vring_avail_t __user *avail,
1301			 vring_used_t __user *used)
1302
1303{
1304	/* If an IOTLB device is present, the vring addresses are
1305	 * GIOVAs. Access validation occurs at prefetch time. */
1306	if (vq->iotlb)
1307		return true;
1308
1309	return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1310	       access_ok(avail, vhost_get_avail_size(vq, num)) &&
1311	       access_ok(used, vhost_get_used_size(vq, num));
1312}
1313
1314static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1315				 const struct vhost_iotlb_map *map,
1316				 int type)
1317{
1318	int access = (type == VHOST_ADDR_USED) ?
1319		     VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1320
1321	if (likely(map->perm & access))
1322		vq->meta_iotlb[type] = map;
 
 
1323}
1324
1325static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1326			    int access, u64 addr, u64 len, int type)
1327{
1328	const struct vhost_iotlb_map *map;
1329	struct vhost_iotlb *umem = vq->iotlb;
1330	u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1331
1332	if (vhost_vq_meta_fetch(vq, addr, len, type))
1333		return true;
1334
1335	while (len > s) {
1336		map = vhost_iotlb_itree_first(umem, addr, last);
1337		if (map == NULL || map->start > addr) {
1338			vhost_iotlb_miss(vq, addr, access);
1339			return false;
1340		} else if (!(map->perm & access)) {
1341			/* Report the possible access violation by
1342			 * request another translation from userspace.
1343			 */
1344			return false;
1345		}
1346
1347		size = map->size - addr + map->start;
1348
1349		if (orig_addr == addr && size >= len)
1350			vhost_vq_meta_update(vq, map, type);
1351
1352		s += size;
1353		addr += size;
1354	}
1355
1356	return true;
1357}
1358
1359int vq_meta_prefetch(struct vhost_virtqueue *vq)
1360{
1361	unsigned int num = vq->num;
1362
1363	if (!vq->iotlb)
1364		return 1;
1365
1366	return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1367			       vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1368	       iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1369			       vhost_get_avail_size(vq, num),
1370			       VHOST_ADDR_AVAIL) &&
1371	       iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1372			       vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1373}
1374EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1375
1376/* Can we log writes? */
1377/* Caller should have device mutex but not vq mutex */
1378bool vhost_log_access_ok(struct vhost_dev *dev)
1379{
1380	return memory_access_ok(dev, dev->umem, 1);
1381}
1382EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1383
1384static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1385				  void __user *log_base,
1386				  bool log_used,
1387				  u64 log_addr)
1388{
1389	/* If an IOTLB device is present, log_addr is a GIOVA that
1390	 * will never be logged by log_used(). */
1391	if (vq->iotlb)
1392		return true;
1393
1394	return !log_used || log_access_ok(log_base, log_addr,
1395					  vhost_get_used_size(vq, vq->num));
 
1396}
1397
1398/* Verify access for write logging. */
1399/* Caller should have vq mutex and device mutex */
1400static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1401			     void __user *log_base)
1402{
1403	return vq_memory_access_ok(log_base, vq->umem,
1404				   vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1405		vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
 
 
 
 
 
 
 
1406}
1407
1408/* Can we start vq? */
1409/* Caller should have vq mutex and device mutex */
1410bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1411{
1412	if (!vq_log_access_ok(vq, vq->log_base))
1413		return false;
1414
1415	return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1416}
1417EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1418
1419static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1420{
1421	struct vhost_memory mem, *newmem;
1422	struct vhost_memory_region *region;
1423	struct vhost_iotlb *newumem, *oldumem;
1424	unsigned long size = offsetof(struct vhost_memory, regions);
1425	int i;
1426
1427	if (copy_from_user(&mem, m, size))
1428		return -EFAULT;
1429	if (mem.padding)
1430		return -EOPNOTSUPP;
1431	if (mem.nregions > max_mem_regions)
1432		return -E2BIG;
1433	newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1434			GFP_KERNEL);
1435	if (!newmem)
1436		return -ENOMEM;
1437
1438	memcpy(newmem, &mem, size);
1439	if (copy_from_user(newmem->regions, m->regions,
1440			   flex_array_size(newmem, regions, mem.nregions))) {
1441		kvfree(newmem);
1442		return -EFAULT;
1443	}
1444
1445	newumem = iotlb_alloc();
1446	if (!newumem) {
1447		kvfree(newmem);
1448		return -ENOMEM;
1449	}
1450
1451	for (region = newmem->regions;
1452	     region < newmem->regions + mem.nregions;
1453	     region++) {
1454		if (vhost_iotlb_add_range(newumem,
1455					  region->guest_phys_addr,
1456					  region->guest_phys_addr +
1457					  region->memory_size - 1,
1458					  region->userspace_addr,
1459					  VHOST_MAP_RW))
1460			goto err;
1461	}
1462
1463	if (!memory_access_ok(d, newumem, 0))
1464		goto err;
1465
1466	oldumem = d->umem;
1467	d->umem = newumem;
1468
1469	/* All memory accesses are done under some VQ mutex. */
1470	for (i = 0; i < d->nvqs; ++i) {
1471		mutex_lock(&d->vqs[i]->mutex);
1472		d->vqs[i]->umem = newumem;
1473		mutex_unlock(&d->vqs[i]->mutex);
1474	}
1475
1476	kvfree(newmem);
1477	vhost_iotlb_free(oldumem);
1478	return 0;
1479
1480err:
1481	vhost_iotlb_free(newumem);
1482	kvfree(newmem);
1483	return -EFAULT;
1484}
1485
1486static long vhost_vring_set_num(struct vhost_dev *d,
1487				struct vhost_virtqueue *vq,
1488				void __user *argp)
1489{
1490	struct vhost_vring_state s;
1491
1492	/* Resizing ring with an active backend?
1493	 * You don't want to do that. */
1494	if (vq->private_data)
1495		return -EBUSY;
1496
1497	if (copy_from_user(&s, argp, sizeof s))
1498		return -EFAULT;
1499
1500	if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1501		return -EINVAL;
1502	vq->num = s.num;
1503
1504	return 0;
1505}
1506
1507static long vhost_vring_set_addr(struct vhost_dev *d,
1508				 struct vhost_virtqueue *vq,
1509				 void __user *argp)
1510{
1511	struct vhost_vring_addr a;
1512
1513	if (copy_from_user(&a, argp, sizeof a))
1514		return -EFAULT;
1515	if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1516		return -EOPNOTSUPP;
1517
1518	/* For 32bit, verify that the top 32bits of the user
1519	   data are set to zero. */
1520	if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1521	    (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1522	    (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1523		return -EFAULT;
1524
1525	/* Make sure it's safe to cast pointers to vring types. */
1526	BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1527	BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1528	if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1529	    (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1530	    (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1531		return -EINVAL;
1532
1533	/* We only verify access here if backend is configured.
1534	 * If it is not, we don't as size might not have been setup.
1535	 * We will verify when backend is configured. */
1536	if (vq->private_data) {
1537		if (!vq_access_ok(vq, vq->num,
1538			(void __user *)(unsigned long)a.desc_user_addr,
1539			(void __user *)(unsigned long)a.avail_user_addr,
1540			(void __user *)(unsigned long)a.used_user_addr))
1541			return -EINVAL;
1542
1543		/* Also validate log access for used ring if enabled. */
1544		if (!vq_log_used_access_ok(vq, vq->log_base,
1545				a.flags & (0x1 << VHOST_VRING_F_LOG),
1546				a.log_guest_addr))
1547			return -EINVAL;
1548	}
1549
1550	vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1551	vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1552	vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1553	vq->log_addr = a.log_guest_addr;
1554	vq->used = (void __user *)(unsigned long)a.used_user_addr;
1555
1556	return 0;
1557}
1558
1559static long vhost_vring_set_num_addr(struct vhost_dev *d,
1560				     struct vhost_virtqueue *vq,
1561				     unsigned int ioctl,
1562				     void __user *argp)
1563{
1564	long r;
1565
1566	mutex_lock(&vq->mutex);
1567
1568	switch (ioctl) {
1569	case VHOST_SET_VRING_NUM:
1570		r = vhost_vring_set_num(d, vq, argp);
1571		break;
1572	case VHOST_SET_VRING_ADDR:
1573		r = vhost_vring_set_addr(d, vq, argp);
1574		break;
1575	default:
1576		BUG();
1577	}
1578
1579	mutex_unlock(&vq->mutex);
1580
1581	return r;
1582}
1583long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1584{
1585	struct file *eventfp, *filep = NULL;
1586	bool pollstart = false, pollstop = false;
1587	struct eventfd_ctx *ctx = NULL;
1588	u32 __user *idxp = argp;
1589	struct vhost_virtqueue *vq;
1590	struct vhost_vring_state s;
1591	struct vhost_vring_file f;
 
1592	u32 idx;
1593	long r;
1594
1595	r = get_user(idx, idxp);
1596	if (r < 0)
1597		return r;
1598	if (idx >= d->nvqs)
1599		return -ENOBUFS;
1600
1601	idx = array_index_nospec(idx, d->nvqs);
1602	vq = d->vqs[idx];
1603
1604	if (ioctl == VHOST_SET_VRING_NUM ||
1605	    ioctl == VHOST_SET_VRING_ADDR) {
1606		return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1607	}
1608
1609	mutex_lock(&vq->mutex);
1610
1611	switch (ioctl) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1612	case VHOST_SET_VRING_BASE:
1613		/* Moving base with an active backend?
1614		 * You don't want to do that. */
1615		if (vq->private_data) {
1616			r = -EBUSY;
1617			break;
1618		}
1619		if (copy_from_user(&s, argp, sizeof s)) {
1620			r = -EFAULT;
1621			break;
1622		}
1623		if (s.num > 0xffff) {
1624			r = -EINVAL;
1625			break;
1626		}
1627		vq->last_avail_idx = s.num;
1628		/* Forget the cached index value. */
1629		vq->avail_idx = vq->last_avail_idx;
1630		break;
1631	case VHOST_GET_VRING_BASE:
1632		s.index = idx;
1633		s.num = vq->last_avail_idx;
1634		if (copy_to_user(argp, &s, sizeof s))
1635			r = -EFAULT;
1636		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1637	case VHOST_SET_VRING_KICK:
1638		if (copy_from_user(&f, argp, sizeof f)) {
1639			r = -EFAULT;
1640			break;
1641		}
1642		eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1643		if (IS_ERR(eventfp)) {
1644			r = PTR_ERR(eventfp);
1645			break;
1646		}
1647		if (eventfp != vq->kick) {
1648			pollstop = (filep = vq->kick) != NULL;
1649			pollstart = (vq->kick = eventfp) != NULL;
1650		} else
1651			filep = eventfp;
1652		break;
1653	case VHOST_SET_VRING_CALL:
1654		if (copy_from_user(&f, argp, sizeof f)) {
1655			r = -EFAULT;
1656			break;
1657		}
1658		ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1659		if (IS_ERR(ctx)) {
1660			r = PTR_ERR(ctx);
1661			break;
1662		}
1663
1664		swap(ctx, vq->call_ctx.ctx);
 
 
 
 
 
 
1665		break;
1666	case VHOST_SET_VRING_ERR:
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		swap(ctx, vq->error_ctx);
1677		break;
1678	case VHOST_SET_VRING_ENDIAN:
1679		r = vhost_set_vring_endian(vq, argp);
1680		break;
1681	case VHOST_GET_VRING_ENDIAN:
1682		r = vhost_get_vring_endian(vq, idx, argp);
1683		break;
1684	case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1685		if (copy_from_user(&s, argp, sizeof(s))) {
1686			r = -EFAULT;
1687			break;
1688		}
1689		vq->busyloop_timeout = s.num;
1690		break;
1691	case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1692		s.index = idx;
1693		s.num = vq->busyloop_timeout;
1694		if (copy_to_user(argp, &s, sizeof(s)))
1695			r = -EFAULT;
1696		break;
1697	default:
1698		r = -ENOIOCTLCMD;
1699	}
1700
1701	if (pollstop && vq->handle_kick)
1702		vhost_poll_stop(&vq->poll);
1703
1704	if (!IS_ERR_OR_NULL(ctx))
1705		eventfd_ctx_put(ctx);
1706	if (filep)
1707		fput(filep);
1708
1709	if (pollstart && vq->handle_kick)
1710		r = vhost_poll_start(&vq->poll, vq->kick);
1711
1712	mutex_unlock(&vq->mutex);
1713
1714	if (pollstop && vq->handle_kick)
1715		vhost_poll_flush(&vq->poll);
1716	return r;
1717}
1718EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1719
1720int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1721{
1722	struct vhost_iotlb *niotlb, *oiotlb;
1723	int i;
1724
1725	niotlb = iotlb_alloc();
1726	if (!niotlb)
1727		return -ENOMEM;
1728
1729	oiotlb = d->iotlb;
1730	d->iotlb = niotlb;
1731
1732	for (i = 0; i < d->nvqs; ++i) {
1733		struct vhost_virtqueue *vq = d->vqs[i];
1734
1735		mutex_lock(&vq->mutex);
1736		vq->iotlb = niotlb;
1737		__vhost_vq_meta_reset(vq);
1738		mutex_unlock(&vq->mutex);
1739	}
1740
1741	vhost_iotlb_free(oiotlb);
1742
1743	return 0;
1744}
1745EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1746
1747/* Caller must have device mutex */
1748long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1749{
1750	struct eventfd_ctx *ctx;
 
 
1751	u64 p;
1752	long r;
1753	int i, fd;
1754
1755	/* If you are not the owner, you can become one */
1756	if (ioctl == VHOST_SET_OWNER) {
1757		r = vhost_dev_set_owner(d);
1758		goto done;
1759	}
1760
1761	/* You must be the owner to do anything else */
1762	r = vhost_dev_check_owner(d);
1763	if (r)
1764		goto done;
1765
1766	switch (ioctl) {
1767	case VHOST_SET_MEM_TABLE:
1768		r = vhost_set_memory(d, argp);
1769		break;
1770	case VHOST_SET_LOG_BASE:
1771		if (copy_from_user(&p, argp, sizeof p)) {
1772			r = -EFAULT;
1773			break;
1774		}
1775		if ((u64)(unsigned long)p != p) {
1776			r = -EFAULT;
1777			break;
1778		}
1779		for (i = 0; i < d->nvqs; ++i) {
1780			struct vhost_virtqueue *vq;
1781			void __user *base = (void __user *)(unsigned long)p;
1782			vq = d->vqs[i];
1783			mutex_lock(&vq->mutex);
1784			/* If ring is inactive, will check when it's enabled. */
1785			if (vq->private_data && !vq_log_access_ok(vq, base))
1786				r = -EFAULT;
1787			else
1788				vq->log_base = base;
1789			mutex_unlock(&vq->mutex);
1790		}
1791		break;
1792	case VHOST_SET_LOG_FD:
1793		r = get_user(fd, (int __user *)argp);
1794		if (r < 0)
1795			break;
1796		ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1797		if (IS_ERR(ctx)) {
1798			r = PTR_ERR(ctx);
1799			break;
1800		}
1801		swap(ctx, d->log_ctx);
 
 
 
 
 
 
1802		for (i = 0; i < d->nvqs; ++i) {
1803			mutex_lock(&d->vqs[i]->mutex);
1804			d->vqs[i]->log_ctx = d->log_ctx;
1805			mutex_unlock(&d->vqs[i]->mutex);
1806		}
1807		if (ctx)
1808			eventfd_ctx_put(ctx);
 
 
1809		break;
1810	default:
1811		r = -ENOIOCTLCMD;
1812		break;
1813	}
1814done:
1815	return r;
1816}
1817EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1818
1819/* TODO: This is really inefficient.  We need something like get_user()
1820 * (instruction directly accesses the data, with an exception table entry
1821 * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1822 */
1823static int set_bit_to_user(int nr, void __user *addr)
1824{
1825	unsigned long log = (unsigned long)addr;
1826	struct page *page;
1827	void *base;
1828	int bit = nr + (log % PAGE_SIZE) * 8;
1829	int r;
1830
1831	r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1832	if (r < 0)
1833		return r;
1834	BUG_ON(r != 1);
1835	base = kmap_atomic(page);
1836	set_bit(bit, base);
1837	kunmap_atomic(base);
1838	unpin_user_pages_dirty_lock(&page, 1, true);
 
1839	return 0;
1840}
1841
1842static int log_write(void __user *log_base,
1843		     u64 write_address, u64 write_length)
1844{
1845	u64 write_page = write_address / VHOST_PAGE_SIZE;
1846	int r;
1847
1848	if (!write_length)
1849		return 0;
1850	write_length += write_address % VHOST_PAGE_SIZE;
1851	for (;;) {
1852		u64 base = (u64)(unsigned long)log_base;
1853		u64 log = base + write_page / 8;
1854		int bit = write_page % 8;
1855		if ((u64)(unsigned long)log != log)
1856			return -EFAULT;
1857		r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1858		if (r < 0)
1859			return r;
1860		if (write_length <= VHOST_PAGE_SIZE)
1861			break;
1862		write_length -= VHOST_PAGE_SIZE;
1863		write_page += 1;
1864	}
1865	return r;
1866}
1867
1868static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1869{
1870	struct vhost_iotlb *umem = vq->umem;
1871	struct vhost_iotlb_map *u;
1872	u64 start, end, l, min;
1873	int r;
1874	bool hit = false;
1875
1876	while (len) {
1877		min = len;
1878		/* More than one GPAs can be mapped into a single HVA. So
1879		 * iterate all possible umems here to be safe.
1880		 */
1881		list_for_each_entry(u, &umem->list, link) {
1882			if (u->addr > hva - 1 + len ||
1883			    u->addr - 1 + u->size < hva)
1884				continue;
1885			start = max(u->addr, hva);
1886			end = min(u->addr - 1 + u->size, hva - 1 + len);
1887			l = end - start + 1;
1888			r = log_write(vq->log_base,
1889				      u->start + start - u->addr,
1890				      l);
1891			if (r < 0)
1892				return r;
1893			hit = true;
1894			min = min(l, min);
1895		}
1896
1897		if (!hit)
1898			return -EFAULT;
1899
1900		len -= min;
1901		hva += min;
1902	}
1903
1904	return 0;
1905}
1906
1907static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1908{
1909	struct iovec *iov = vq->log_iov;
1910	int i, ret;
1911
1912	if (!vq->iotlb)
1913		return log_write(vq->log_base, vq->log_addr + used_offset, len);
1914
1915	ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1916			     len, iov, 64, VHOST_ACCESS_WO);
1917	if (ret < 0)
1918		return ret;
1919
1920	for (i = 0; i < ret; i++) {
1921		ret = log_write_hva(vq,	(uintptr_t)iov[i].iov_base,
1922				    iov[i].iov_len);
1923		if (ret)
1924			return ret;
1925	}
1926
1927	return 0;
1928}
1929
1930int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1931		    unsigned int log_num, u64 len, struct iovec *iov, int count)
1932{
1933	int i, r;
1934
1935	/* Make sure data written is seen before log. */
1936	smp_wmb();
1937
1938	if (vq->iotlb) {
1939		for (i = 0; i < count; i++) {
1940			r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1941					  iov[i].iov_len);
1942			if (r < 0)
1943				return r;
1944		}
1945		return 0;
1946	}
1947
1948	for (i = 0; i < log_num; ++i) {
1949		u64 l = min(log[i].len, len);
1950		r = log_write(vq->log_base, log[i].addr, l);
1951		if (r < 0)
1952			return r;
1953		len -= l;
1954		if (!len) {
1955			if (vq->log_ctx)
1956				eventfd_signal(vq->log_ctx, 1);
1957			return 0;
1958		}
1959	}
1960	/* Length written exceeds what we have stored. This is a bug. */
1961	BUG();
1962	return 0;
1963}
1964EXPORT_SYMBOL_GPL(vhost_log_write);
1965
1966static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1967{
1968	void __user *used;
1969	if (vhost_put_used_flags(vq))
1970		return -EFAULT;
1971	if (unlikely(vq->log_used)) {
1972		/* Make sure the flag is seen before log. */
1973		smp_wmb();
1974		/* Log used flag write. */
1975		used = &vq->used->flags;
1976		log_used(vq, (used - (void __user *)vq->used),
1977			 sizeof vq->used->flags);
 
1978		if (vq->log_ctx)
1979			eventfd_signal(vq->log_ctx, 1);
1980	}
1981	return 0;
1982}
1983
1984static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1985{
1986	if (vhost_put_avail_event(vq))
1987		return -EFAULT;
1988	if (unlikely(vq->log_used)) {
1989		void __user *used;
1990		/* Make sure the event is seen before log. */
1991		smp_wmb();
1992		/* Log avail event write */
1993		used = vhost_avail_event(vq);
1994		log_used(vq, (used - (void __user *)vq->used),
1995			 sizeof *vhost_avail_event(vq));
 
1996		if (vq->log_ctx)
1997			eventfd_signal(vq->log_ctx, 1);
1998	}
1999	return 0;
2000}
2001
2002int vhost_vq_init_access(struct vhost_virtqueue *vq)
2003{
2004	__virtio16 last_used_idx;
2005	int r;
2006	bool is_le = vq->is_le;
2007
2008	if (!vq->private_data)
2009		return 0;
2010
2011	vhost_init_is_le(vq);
2012
2013	r = vhost_update_used_flags(vq);
2014	if (r)
2015		goto err;
2016	vq->signalled_used_valid = false;
2017	if (!vq->iotlb &&
2018	    !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2019		r = -EFAULT;
2020		goto err;
2021	}
2022	r = vhost_get_used_idx(vq, &last_used_idx);
2023	if (r) {
2024		vq_err(vq, "Can't access used idx at %p\n",
2025		       &vq->used->idx);
2026		goto err;
2027	}
2028	vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2029	return 0;
2030
2031err:
2032	vq->is_le = is_le;
2033	return r;
2034}
2035EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2036
2037static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2038			  struct iovec iov[], int iov_size, int access)
2039{
2040	const struct vhost_iotlb_map *map;
2041	struct vhost_dev *dev = vq->dev;
2042	struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2043	struct iovec *_iov;
2044	u64 s = 0;
2045	int ret = 0;
2046
 
 
 
2047	while ((u64)len > s) {
2048		u64 size;
2049		if (unlikely(ret >= iov_size)) {
2050			ret = -ENOBUFS;
2051			break;
2052		}
2053
2054		map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2055		if (map == NULL || map->start > addr) {
2056			if (umem != dev->iotlb) {
2057				ret = -EFAULT;
2058				break;
2059			}
2060			ret = -EAGAIN;
2061			break;
2062		} else if (!(map->perm & access)) {
2063			ret = -EPERM;
2064			break;
2065		}
2066
2067		_iov = iov + ret;
2068		size = map->size - addr + map->start;
2069		_iov->iov_len = min((u64)len - s, size);
2070		_iov->iov_base = (void __user *)(unsigned long)
2071				 (map->addr + addr - map->start);
2072		s += size;
2073		addr += size;
2074		++ret;
2075	}
2076
2077	if (ret == -EAGAIN)
2078		vhost_iotlb_miss(vq, addr, access);
2079	return ret;
2080}
2081
2082/* Each buffer in the virtqueues is actually a chain of descriptors.  This
2083 * function returns the next descriptor in the chain,
2084 * or -1U if we're at the end. */
2085static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2086{
2087	unsigned int next;
2088
2089	/* If this descriptor says it doesn't chain, we're done. */
2090	if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2091		return -1U;
2092
2093	/* Check they're not leading us off end of descriptors. */
2094	next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
 
 
 
 
 
2095	return next;
2096}
2097
2098static int get_indirect(struct vhost_virtqueue *vq,
2099			struct iovec iov[], unsigned int iov_size,
2100			unsigned int *out_num, unsigned int *in_num,
2101			struct vhost_log *log, unsigned int *log_num,
2102			struct vring_desc *indirect)
2103{
2104	struct vring_desc desc;
2105	unsigned int i = 0, count, found = 0;
2106	u32 len = vhost32_to_cpu(vq, indirect->len);
2107	struct iov_iter from;
2108	int ret, access;
2109
2110	/* Sanity check */
2111	if (unlikely(len % sizeof desc)) {
2112		vq_err(vq, "Invalid length in indirect descriptor: "
2113		       "len 0x%llx not multiple of 0x%zx\n",
2114		       (unsigned long long)len,
2115		       sizeof desc);
2116		return -EINVAL;
2117	}
2118
2119	ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2120			     UIO_MAXIOV, VHOST_ACCESS_RO);
2121	if (unlikely(ret < 0)) {
2122		if (ret != -EAGAIN)
2123			vq_err(vq, "Translation failure %d in indirect.\n", ret);
2124		return ret;
2125	}
2126	iov_iter_init(&from, READ, vq->indirect, ret, len);
2127	count = len / sizeof desc;
 
 
 
 
2128	/* Buffers are chained via a 16 bit next field, so
2129	 * we can have at most 2^16 of these. */
2130	if (unlikely(count > USHRT_MAX + 1)) {
2131		vq_err(vq, "Indirect buffer length too big: %d\n",
2132		       indirect->len);
2133		return -E2BIG;
2134	}
2135
2136	do {
2137		unsigned iov_count = *in_num + *out_num;
2138		if (unlikely(++found > count)) {
2139			vq_err(vq, "Loop detected: last one at %u "
2140			       "indirect size %u\n",
2141			       i, count);
2142			return -EINVAL;
2143		}
2144		if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
 
2145			vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2146			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2147			return -EINVAL;
2148		}
2149		if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2150			vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2151			       i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2152			return -EINVAL;
2153		}
2154
2155		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2156			access = VHOST_ACCESS_WO;
2157		else
2158			access = VHOST_ACCESS_RO;
2159
2160		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2161				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2162				     iov_size - iov_count, access);
2163		if (unlikely(ret < 0)) {
2164			if (ret != -EAGAIN)
2165				vq_err(vq, "Translation failure %d indirect idx %d\n",
2166					ret, i);
2167			return ret;
2168		}
2169		/* If this is an input descriptor, increment that count. */
2170		if (access == VHOST_ACCESS_WO) {
2171			*in_num += ret;
2172			if (unlikely(log && ret)) {
2173				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2174				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2175				++*log_num;
2176			}
2177		} else {
2178			/* If it's an output descriptor, they're all supposed
2179			 * to come before any input descriptors. */
2180			if (unlikely(*in_num)) {
2181				vq_err(vq, "Indirect descriptor "
2182				       "has out after in: idx %d\n", i);
2183				return -EINVAL;
2184			}
2185			*out_num += ret;
2186		}
2187	} while ((i = next_desc(vq, &desc)) != -1);
2188	return 0;
2189}
2190
2191/* This looks in the virtqueue and for the first available buffer, and converts
2192 * it to an iovec for convenient access.  Since descriptors consist of some
2193 * number of output then some number of input descriptors, it's actually two
2194 * iovecs, but we pack them into one and note how many of each there were.
2195 *
2196 * This function returns the descriptor number found, or vq->num (which is
2197 * never a valid descriptor number) if none was found.  A negative code is
2198 * returned on error. */
2199int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2200		      struct iovec iov[], unsigned int iov_size,
2201		      unsigned int *out_num, unsigned int *in_num,
2202		      struct vhost_log *log, unsigned int *log_num)
2203{
2204	struct vring_desc desc;
2205	unsigned int i, head, found = 0;
2206	u16 last_avail_idx;
2207	__virtio16 avail_idx;
2208	__virtio16 ring_head;
2209	int ret, access;
2210
2211	/* Check it isn't doing very strange things with descriptor numbers. */
2212	last_avail_idx = vq->last_avail_idx;
 
 
 
 
 
2213
2214	if (vq->avail_idx == vq->last_avail_idx) {
2215		if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2216			vq_err(vq, "Failed to access avail idx at %p\n",
2217				&vq->avail->idx);
2218			return -EFAULT;
2219		}
2220		vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2221
2222		if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2223			vq_err(vq, "Guest moved used index from %u to %u",
2224				last_avail_idx, vq->avail_idx);
2225			return -EFAULT;
2226		}
2227
2228		/* If there's nothing new since last we looked, return
2229		 * invalid.
2230		 */
2231		if (vq->avail_idx == last_avail_idx)
2232			return vq->num;
2233
2234		/* Only get avail ring entries after they have been
2235		 * exposed by guest.
2236		 */
2237		smp_rmb();
2238	}
2239
2240	/* Grab the next descriptor number they're advertising, and increment
2241	 * the index we've seen. */
2242	if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
 
2243		vq_err(vq, "Failed to read head: idx %d address %p\n",
2244		       last_avail_idx,
2245		       &vq->avail->ring[last_avail_idx % vq->num]);
2246		return -EFAULT;
2247	}
2248
2249	head = vhost16_to_cpu(vq, ring_head);
2250
2251	/* If their number is silly, that's an error. */
2252	if (unlikely(head >= vq->num)) {
2253		vq_err(vq, "Guest says index %u > %u is available",
2254		       head, vq->num);
2255		return -EINVAL;
2256	}
2257
2258	/* When we start there are none of either input nor output. */
2259	*out_num = *in_num = 0;
2260	if (unlikely(log))
2261		*log_num = 0;
2262
2263	i = head;
2264	do {
2265		unsigned iov_count = *in_num + *out_num;
2266		if (unlikely(i >= vq->num)) {
2267			vq_err(vq, "Desc index is %u > %u, head = %u",
2268			       i, vq->num, head);
2269			return -EINVAL;
2270		}
2271		if (unlikely(++found > vq->num)) {
2272			vq_err(vq, "Loop detected: last one at %u "
2273			       "vq size %u head %u\n",
2274			       i, vq->num, head);
2275			return -EINVAL;
2276		}
2277		ret = vhost_get_desc(vq, &desc, i);
2278		if (unlikely(ret)) {
2279			vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2280			       i, vq->desc + i);
2281			return -EFAULT;
2282		}
2283		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2284			ret = get_indirect(vq, iov, iov_size,
2285					   out_num, in_num,
2286					   log, log_num, &desc);
2287			if (unlikely(ret < 0)) {
2288				if (ret != -EAGAIN)
2289					vq_err(vq, "Failure detected "
2290						"in indirect descriptor at idx %d\n", i);
2291				return ret;
2292			}
2293			continue;
2294		}
2295
2296		if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2297			access = VHOST_ACCESS_WO;
2298		else
2299			access = VHOST_ACCESS_RO;
2300		ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2301				     vhost32_to_cpu(vq, desc.len), iov + iov_count,
2302				     iov_size - iov_count, access);
2303		if (unlikely(ret < 0)) {
2304			if (ret != -EAGAIN)
2305				vq_err(vq, "Translation failure %d descriptor idx %d\n",
2306					ret, i);
2307			return ret;
2308		}
2309		if (access == VHOST_ACCESS_WO) {
2310			/* If this is an input descriptor,
2311			 * increment that count. */
2312			*in_num += ret;
2313			if (unlikely(log && ret)) {
2314				log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2315				log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2316				++*log_num;
2317			}
2318		} else {
2319			/* If it's an output descriptor, they're all supposed
2320			 * to come before any input descriptors. */
2321			if (unlikely(*in_num)) {
2322				vq_err(vq, "Descriptor has out after in: "
2323				       "idx %d\n", i);
2324				return -EINVAL;
2325			}
2326			*out_num += ret;
2327		}
2328	} while ((i = next_desc(vq, &desc)) != -1);
2329
2330	/* On success, increment avail index. */
2331	vq->last_avail_idx++;
2332
2333	/* Assume notifications from guest are disabled at this point,
2334	 * if they aren't we would need to update avail_event index. */
2335	BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2336	return head;
2337}
2338EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2339
2340/* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2341void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2342{
2343	vq->last_avail_idx -= n;
2344}
2345EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2346
2347/* After we've used one of their buffers, we tell them about it.  We'll then
2348 * want to notify the guest, using eventfd. */
2349int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2350{
2351	struct vring_used_elem heads = {
2352		cpu_to_vhost32(vq, head),
2353		cpu_to_vhost32(vq, len)
2354	};
2355
2356	return vhost_add_used_n(vq, &heads, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2357}
2358EXPORT_SYMBOL_GPL(vhost_add_used);
2359
2360static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2361			    struct vring_used_elem *heads,
2362			    unsigned count)
2363{
2364	vring_used_elem_t __user *used;
2365	u16 old, new;
2366	int start;
2367
2368	start = vq->last_used_idx & (vq->num - 1);
2369	used = vq->used->ring + start;
2370	if (vhost_put_used(vq, heads, start, count)) {
2371		vq_err(vq, "Failed to write used");
2372		return -EFAULT;
2373	}
2374	if (unlikely(vq->log_used)) {
2375		/* Make sure data is seen before log. */
2376		smp_wmb();
2377		/* Log used ring entry write. */
2378		log_used(vq, ((void __user *)used - (void __user *)vq->used),
2379			 count * sizeof *used);
 
 
2380	}
2381	old = vq->last_used_idx;
2382	new = (vq->last_used_idx += count);
2383	/* If the driver never bothers to signal in a very long while,
2384	 * used index might wrap around. If that happens, invalidate
2385	 * signalled_used index we stored. TODO: make sure driver
2386	 * signals at least once in 2^16 and remove this. */
2387	if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2388		vq->signalled_used_valid = false;
2389	return 0;
2390}
2391
2392/* After we've used one of their buffers, we tell them about it.  We'll then
2393 * want to notify the guest, using eventfd. */
2394int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2395		     unsigned count)
2396{
2397	int start, n, r;
2398
2399	start = vq->last_used_idx & (vq->num - 1);
2400	n = vq->num - start;
2401	if (n < count) {
2402		r = __vhost_add_used_n(vq, heads, n);
2403		if (r < 0)
2404			return r;
2405		heads += n;
2406		count -= n;
2407	}
2408	r = __vhost_add_used_n(vq, heads, count);
2409
2410	/* Make sure buffer is written before we update index. */
2411	smp_wmb();
2412	if (vhost_put_used_idx(vq)) {
2413		vq_err(vq, "Failed to increment used idx");
2414		return -EFAULT;
2415	}
2416	if (unlikely(vq->log_used)) {
2417		/* Make sure used idx is seen before log. */
2418		smp_wmb();
2419		/* Log used index update. */
2420		log_used(vq, offsetof(struct vring_used, idx),
2421			 sizeof vq->used->idx);
 
2422		if (vq->log_ctx)
2423			eventfd_signal(vq->log_ctx, 1);
2424	}
2425	return r;
2426}
2427EXPORT_SYMBOL_GPL(vhost_add_used_n);
2428
2429static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2430{
2431	__u16 old, new;
2432	__virtio16 event;
2433	bool v;
2434	/* Flush out used index updates. This is paired
2435	 * with the barrier that the Guest executes when enabling
2436	 * interrupts. */
2437	smp_mb();
2438
2439	if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2440	    unlikely(vq->avail_idx == vq->last_avail_idx))
2441		return true;
2442
2443	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2444		__virtio16 flags;
2445		if (vhost_get_avail_flags(vq, &flags)) {
2446			vq_err(vq, "Failed to get flags");
2447			return true;
2448		}
2449		return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2450	}
2451	old = vq->signalled_used;
2452	v = vq->signalled_used_valid;
2453	new = vq->signalled_used = vq->last_used_idx;
2454	vq->signalled_used_valid = true;
2455
2456	if (unlikely(!v))
2457		return true;
2458
2459	if (vhost_get_used_event(vq, &event)) {
2460		vq_err(vq, "Failed to get used event idx");
2461		return true;
2462	}
2463	return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2464}
2465
2466/* This actually signals the guest, using eventfd. */
2467void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2468{
2469	/* Signal the Guest tell them we used something up. */
2470	if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2471		eventfd_signal(vq->call_ctx.ctx, 1);
2472}
2473EXPORT_SYMBOL_GPL(vhost_signal);
2474
2475/* And here's the combo meal deal.  Supersize me! */
2476void vhost_add_used_and_signal(struct vhost_dev *dev,
2477			       struct vhost_virtqueue *vq,
2478			       unsigned int head, int len)
2479{
2480	vhost_add_used(vq, head, len);
2481	vhost_signal(dev, vq);
2482}
2483EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2484
2485/* multi-buffer version of vhost_add_used_and_signal */
2486void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2487				 struct vhost_virtqueue *vq,
2488				 struct vring_used_elem *heads, unsigned count)
2489{
2490	vhost_add_used_n(vq, heads, count);
2491	vhost_signal(dev, vq);
2492}
2493EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2494
2495/* return true if we're sure that avaiable ring is empty */
2496bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2497{
2498	__virtio16 avail_idx;
2499	int r;
2500
2501	if (vq->avail_idx != vq->last_avail_idx)
2502		return false;
2503
2504	r = vhost_get_avail_idx(vq, &avail_idx);
2505	if (unlikely(r))
2506		return false;
2507	vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2508
2509	return vq->avail_idx == vq->last_avail_idx;
2510}
2511EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2512
2513/* OK, now we need to know about added descriptors. */
2514bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2515{
2516	__virtio16 avail_idx;
2517	int r;
2518
2519	if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2520		return false;
2521	vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2522	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2523		r = vhost_update_used_flags(vq);
2524		if (r) {
2525			vq_err(vq, "Failed to enable notification at %p: %d\n",
2526			       &vq->used->flags, r);
2527			return false;
2528		}
2529	} else {
2530		r = vhost_update_avail_event(vq, vq->avail_idx);
2531		if (r) {
2532			vq_err(vq, "Failed to update avail event index at %p: %d\n",
2533			       vhost_avail_event(vq), r);
2534			return false;
2535		}
2536	}
2537	/* They could have slipped one in as we were doing that: make
2538	 * sure it's written, then check again. */
2539	smp_mb();
2540	r = vhost_get_avail_idx(vq, &avail_idx);
2541	if (r) {
2542		vq_err(vq, "Failed to check avail idx at %p: %d\n",
2543		       &vq->avail->idx, r);
2544		return false;
2545	}
2546
2547	return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2548}
2549EXPORT_SYMBOL_GPL(vhost_enable_notify);
2550
2551/* We don't need to be notified again. */
2552void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2553{
2554	int r;
2555
2556	if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2557		return;
2558	vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2559	if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2560		r = vhost_update_used_flags(vq);
2561		if (r)
2562			vq_err(vq, "Failed to disable notification at %p: %d\n",
2563			       &vq->used->flags, r);
2564	}
2565}
2566EXPORT_SYMBOL_GPL(vhost_disable_notify);
2567
2568/* Create a new message. */
2569struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2570{
2571	struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2572	if (!node)
2573		return NULL;
2574
2575	/* Make sure all padding within the structure is initialized. */
2576	memset(&node->msg, 0, sizeof node->msg);
2577	node->vq = vq;
2578	node->msg.type = type;
2579	return node;
2580}
2581EXPORT_SYMBOL_GPL(vhost_new_msg);
2582
2583void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2584		       struct vhost_msg_node *node)
2585{
2586	spin_lock(&dev->iotlb_lock);
2587	list_add_tail(&node->node, head);
2588	spin_unlock(&dev->iotlb_lock);
2589
2590	wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 
 
 
 
 
 
2591}
2592EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2593
2594struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2595					 struct list_head *head)
2596{
2597	struct vhost_msg_node *node = NULL;
2598
2599	spin_lock(&dev->iotlb_lock);
2600	if (!list_empty(head)) {
2601		node = list_first_entry(head, struct vhost_msg_node,
2602					node);
2603		list_del(&node->node);
2604	}
2605	spin_unlock(&dev->iotlb_lock);
2606
2607	return node;
2608}
2609EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2610
2611void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2612{
2613	struct vhost_virtqueue *vq;
2614	int i;
2615
2616	mutex_lock(&dev->mutex);
2617	for (i = 0; i < dev->nvqs; ++i) {
2618		vq = dev->vqs[i];
2619		mutex_lock(&vq->mutex);
2620		vq->acked_backend_features = features;
2621		mutex_unlock(&vq->mutex);
2622	}
2623	mutex_unlock(&dev->mutex);
2624}
2625EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2626
2627static int __init vhost_init(void)
2628{
2629	return 0;
2630}
 
2631
2632static void __exit vhost_exit(void)
2633{
 
2634}
2635
2636module_init(vhost_init);
2637module_exit(vhost_exit);
2638
2639MODULE_VERSION("0.0.1");
2640MODULE_LICENSE("GPL v2");
2641MODULE_AUTHOR("Michael S. Tsirkin");
2642MODULE_DESCRIPTION("Host kernel accelerator for virtio");