gntalloc.c 15.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
/******************************************************************************
 * gntalloc.c
 *
 * Device for creating grant references (in user-space) that may be shared
 * with other domains.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */

/*
 * This driver exists to allow userspace programs in Linux to allocate kernel
 * memory that will later be shared with another domain.  Without this device,
 * Linux userspace programs cannot create grant references.
 *
 * How this stuff works:
 *   X -> granting a page to Y
 *   Y -> mapping the grant from X
 *
 *   1. X uses the gntalloc device to allocate a page of kernel memory, P.
 *   2. X creates an entry in the grant table that says domid(Y) can access P.
 *      This is done without a hypercall unless the grant table needs expansion.
 *   3. X gives the grant reference identifier, GREF, to Y.
 *   4. Y maps the page, either directly into kernel memory for use in a backend
 *      driver, or via a the gntdev device to map into the address space of an
 *      application running in Y. This is the first point at which Xen does any
 *      tracking of the page.
 *   5. A program in X mmap()s a segment of the gntalloc device that corresponds
 *      to the shared page, and can now communicate with Y over the shared page.
 *
 *
 * NOTE TO USERSPACE LIBRARIES:
 *   The grant allocation and mmap()ing are, naturally, two separate operations.
 *   You set up the sharing by calling the create ioctl() and then the mmap().
 *   Teardown requires munmap() and either close() or ioctl().
 *
 * WARNING: Since Xen does not allow a guest to forcibly end the use of a grant
 * reference, this device can be used to consume kernel memory by leaving grant
 * references mapped by another domain when an application exits. Therefore,
 * there is a global limit on the number of pages that can be allocated. When
 * all references to the page are unmapped, it will be freed during the next
 * grant operation.
 */

#include <linux/atomic.h>
#include <linux/module.h>
#include <linux/miscdevice.h>
#include <linux/kernel.h>
#include <linux/init.h>
#include <linux/slab.h>
#include <linux/fs.h>
#include <linux/device.h>
#include <linux/mm.h>
#include <linux/uaccess.h>
#include <linux/types.h>
#include <linux/list.h>
63
#include <linux/highmem.h>
64 65 66 67 68

#include <xen/xen.h>
#include <xen/page.h>
#include <xen/grant_table.h>
#include <xen/gntalloc.h>
69
#include <xen/events.h>
70 71 72 73 74 75 76

static int limit = 1024;
module_param(limit, int, 0644);
MODULE_PARM_DESC(limit, "Maximum number of grants that may be allocated by "
		"the gntalloc device");

static LIST_HEAD(gref_list);
77
static DEFINE_MUTEX(gref_mutex);
78 79
static int gref_size;

80 81 82 83 84 85
struct notify_info {
	uint16_t pgoff:12;    /* Bits 0-11: Offset of the byte to clear */
	uint16_t flags:2;     /* Bits 12-13: Unmap notification flags */
	int event;            /* Port (event channel) to notify */
};

86 87 88 89 90 91 92 93
/* Metadata on a grant reference. */
struct gntalloc_gref {
	struct list_head next_gref;  /* list entry gref_list */
	struct list_head next_file;  /* list entry file->list, if open */
	struct page *page;	     /* The shared page */
	uint64_t file_index;         /* File offset for mmap() */
	unsigned int users;          /* Use count - when zero, waiting on Xen */
	grant_ref_t gref_id;         /* The grant reference number */
94
	struct notify_info notify;   /* Unmap notification */
95 96 97 98 99 100 101
};

struct gntalloc_file_private_data {
	struct list_head list;
	uint64_t index;
};

102 103 104 105 106 107
struct gntalloc_vma_private_data {
	struct gntalloc_gref *gref;
	int users;
	int count;
};

108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
static void __del_gref(struct gntalloc_gref *gref);

static void do_cleanup(void)
{
	struct gntalloc_gref *gref, *n;
	list_for_each_entry_safe(gref, n, &gref_list, next_gref) {
		if (!gref->users)
			__del_gref(gref);
	}
}

static int add_grefs(struct ioctl_gntalloc_alloc_gref *op,
	uint32_t *gref_ids, struct gntalloc_file_private_data *priv)
{
	int i, rc, readonly;
	LIST_HEAD(queue_gref);
	LIST_HEAD(queue_file);
	struct gntalloc_gref *gref;

	readonly = !(op->flags & GNTALLOC_FLAG_WRITABLE);
	rc = -ENOMEM;
	for (i = 0; i < op->count; i++) {
		gref = kzalloc(sizeof(*gref), GFP_KERNEL);
		if (!gref)
			goto undo;
		list_add_tail(&gref->next_gref, &queue_gref);
		list_add_tail(&gref->next_file, &queue_file);
		gref->users = 1;
		gref->file_index = op->index + i * PAGE_SIZE;
		gref->page = alloc_page(GFP_KERNEL|__GFP_ZERO);
		if (!gref->page)
			goto undo;

		/* Grant foreign access to the page. */
		gref->gref_id = gnttab_grant_foreign_access(op->domid,
			pfn_to_mfn(page_to_pfn(gref->page)), readonly);
144
		if ((int)gref->gref_id < 0) {
145 146 147 148 149 150 151
			rc = gref->gref_id;
			goto undo;
		}
		gref_ids[i] = gref->gref_id;
	}

	/* Add to gref lists. */
152
	mutex_lock(&gref_mutex);
153 154
	list_splice_tail(&queue_gref, &gref_list);
	list_splice_tail(&queue_file, &priv->list);
155
	mutex_unlock(&gref_mutex);
156 157 158 159

	return 0;

undo:
160
	mutex_lock(&gref_mutex);
161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
	gref_size -= (op->count - i);

	list_for_each_entry(gref, &queue_file, next_file) {
		/* __del_gref does not remove from queue_file */
		__del_gref(gref);
	}

	/* It's possible for the target domain to map the just-allocated grant
	 * references by blindly guessing their IDs; if this is done, then
	 * __del_gref will leave them in the queue_gref list. They need to be
	 * added to the global list so that we can free them when they are no
	 * longer referenced.
	 */
	if (unlikely(!list_empty(&queue_gref)))
		list_splice_tail(&queue_gref, &gref_list);
176
	mutex_unlock(&gref_mutex);
177 178 179 180 181
	return rc;
}

static void __del_gref(struct gntalloc_gref *gref)
{
182 183 184 185 186
	if (gref->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
		uint8_t *tmp = kmap(gref->page);
		tmp[gref->notify.pgoff] = 0;
		kunmap(gref->page);
	}
187
	if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
188
		notify_remote_via_evtchn(gref->notify.event);
189 190
		evtchn_put(gref->notify.event);
	}
191 192 193

	gref->notify.flags = 0;

194 195 196 197 198 199
	if (gref->gref_id > 0) {
		if (gnttab_query_foreign_access(gref->gref_id))
			return;

		if (!gnttab_end_foreign_access_ref(gref->gref_id, 0))
			return;
200 201

		gnttab_free_grant_reference(gref->gref_id);
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
	}

	gref_size--;
	list_del(&gref->next_gref);

	if (gref->page)
		__free_page(gref->page);

	kfree(gref);
}

/* finds contiguous grant references in a file, returns the first */
static struct gntalloc_gref *find_grefs(struct gntalloc_file_private_data *priv,
		uint64_t index, uint32_t count)
{
	struct gntalloc_gref *rv = NULL, *gref;
	list_for_each_entry(gref, &priv->list, next_file) {
		if (gref->file_index == index && !rv)
			rv = gref;
		if (rv) {
			if (gref->file_index != index)
				return NULL;
			index += PAGE_SIZE;
			count--;
			if (count == 0)
				return rv;
		}
	}
	return NULL;
}

/*
 * -------------------------------------
 *  File operations.
 * -------------------------------------
 */
static int gntalloc_open(struct inode *inode, struct file *filp)
{
	struct gntalloc_file_private_data *priv;

	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
	if (!priv)
		goto out_nomem;
	INIT_LIST_HEAD(&priv->list);

	filp->private_data = priv;

	pr_debug("%s: priv %p\n", __func__, priv);

	return 0;

out_nomem:
	return -ENOMEM;
}

static int gntalloc_release(struct inode *inode, struct file *filp)
{
	struct gntalloc_file_private_data *priv = filp->private_data;
	struct gntalloc_gref *gref;

	pr_debug("%s: priv %p\n", __func__, priv);

264
	mutex_lock(&gref_mutex);
265 266 267 268 269 270 271 272 273
	while (!list_empty(&priv->list)) {
		gref = list_entry(priv->list.next,
			struct gntalloc_gref, next_file);
		list_del(&gref->next_file);
		gref->users--;
		if (gref->users == 0)
			__del_gref(gref);
	}
	kfree(priv);
274
	mutex_unlock(&gref_mutex);
275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292

	return 0;
}

static long gntalloc_ioctl_alloc(struct gntalloc_file_private_data *priv,
		struct ioctl_gntalloc_alloc_gref __user *arg)
{
	int rc = 0;
	struct ioctl_gntalloc_alloc_gref op;
	uint32_t *gref_ids;

	pr_debug("%s: priv %p\n", __func__, priv);

	if (copy_from_user(&op, arg, sizeof(op))) {
		rc = -EFAULT;
		goto out;
	}

293
	gref_ids = kcalloc(op.count, sizeof(gref_ids[0]), GFP_TEMPORARY);
294 295 296 297 298
	if (!gref_ids) {
		rc = -ENOMEM;
		goto out;
	}

299
	mutex_lock(&gref_mutex);
300 301 302 303 304 305
	/* Clean up pages that were at zero (local) users but were still mapped
	 * by remote domains. Since those pages count towards the limit that we
	 * are about to enforce, removing them here is a good idea.
	 */
	do_cleanup();
	if (gref_size + op.count > limit) {
306
		mutex_unlock(&gref_mutex);
307 308 309 310 311 312
		rc = -ENOSPC;
		goto out_free;
	}
	gref_size += op.count;
	op.index = priv->index;
	priv->index += op.count * PAGE_SIZE;
313
	mutex_unlock(&gref_mutex);
314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355

	rc = add_grefs(&op, gref_ids, priv);
	if (rc < 0)
		goto out_free;

	/* Once we finish add_grefs, it is unsafe to touch the new reference,
	 * since it is possible for a concurrent ioctl to remove it (by guessing
	 * its index). If the userspace application doesn't provide valid memory
	 * to write the IDs to, then it will need to close the file in order to
	 * release - which it will do by segfaulting when it tries to access the
	 * IDs to close them.
	 */
	if (copy_to_user(arg, &op, sizeof(op))) {
		rc = -EFAULT;
		goto out_free;
	}
	if (copy_to_user(arg->gref_ids, gref_ids,
			sizeof(gref_ids[0]) * op.count)) {
		rc = -EFAULT;
		goto out_free;
	}

out_free:
	kfree(gref_ids);
out:
	return rc;
}

static long gntalloc_ioctl_dealloc(struct gntalloc_file_private_data *priv,
		void __user *arg)
{
	int i, rc = 0;
	struct ioctl_gntalloc_dealloc_gref op;
	struct gntalloc_gref *gref, *n;

	pr_debug("%s: priv %p\n", __func__, priv);

	if (copy_from_user(&op, arg, sizeof(op))) {
		rc = -EFAULT;
		goto dealloc_grant_out;
	}

356
	mutex_lock(&gref_mutex);
357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
	gref = find_grefs(priv, op.index, op.count);
	if (gref) {
		/* Remove from the file list only, and decrease reference count.
		 * The later call to do_cleanup() will remove from gref_list and
		 * free the memory if the pages aren't mapped anywhere.
		 */
		for (i = 0; i < op.count; i++) {
			n = list_entry(gref->next_file.next,
				struct gntalloc_gref, next_file);
			list_del(&gref->next_file);
			gref->users--;
			gref = n;
		}
	} else {
		rc = -EINVAL;
	}

	do_cleanup();

376
	mutex_unlock(&gref_mutex);
377 378 379 380
dealloc_grant_out:
	return rc;
}

381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
static long gntalloc_ioctl_unmap_notify(struct gntalloc_file_private_data *priv,
		void __user *arg)
{
	struct ioctl_gntalloc_unmap_notify op;
	struct gntalloc_gref *gref;
	uint64_t index;
	int pgoff;
	int rc;

	if (copy_from_user(&op, arg, sizeof(op)))
		return -EFAULT;

	index = op.index & ~(PAGE_SIZE - 1);
	pgoff = op.index & (PAGE_SIZE - 1);

396
	mutex_lock(&gref_mutex);
397 398 399 400 401 402 403 404 405 406 407 408

	gref = find_grefs(priv, index, 1);
	if (!gref) {
		rc = -ENOENT;
		goto unlock_out;
	}

	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) {
		rc = -EINVAL;
		goto unlock_out;
	}

409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425
	/* We need to grab a reference to the event channel we are going to use
	 * to send the notify before releasing the reference we may already have
	 * (if someone has called this ioctl twice). This is required so that
	 * it is possible to change the clear_byte part of the notification
	 * without disturbing the event channel part, which may now be the last
	 * reference to that event channel.
	 */
	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
		if (evtchn_get(op.event_channel_port)) {
			rc = -EINVAL;
			goto unlock_out;
		}
	}

	if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT)
		evtchn_put(gref->notify.event);

426 427 428 429
	gref->notify.flags = op.action;
	gref->notify.pgoff = pgoff;
	gref->notify.event = op.event_channel_port;
	rc = 0;
430

431
 unlock_out:
432
	mutex_unlock(&gref_mutex);
433 434 435
	return rc;
}

436 437 438 439 440 441 442 443 444 445 446 447
static long gntalloc_ioctl(struct file *filp, unsigned int cmd,
		unsigned long arg)
{
	struct gntalloc_file_private_data *priv = filp->private_data;

	switch (cmd) {
	case IOCTL_GNTALLOC_ALLOC_GREF:
		return gntalloc_ioctl_alloc(priv, (void __user *)arg);

	case IOCTL_GNTALLOC_DEALLOC_GREF:
		return gntalloc_ioctl_dealloc(priv, (void __user *)arg);

448 449 450
	case IOCTL_GNTALLOC_SET_UNMAP_NOTIFY:
		return gntalloc_ioctl_unmap_notify(priv, (void __user *)arg);

451 452 453 454 455 456 457
	default:
		return -ENOIOCTLCMD;
	}

	return 0;
}

458 459
static void gntalloc_vma_open(struct vm_area_struct *vma)
{
460 461 462
	struct gntalloc_vma_private_data *priv = vma->vm_private_data;

	if (!priv)
463 464
		return;

465
	mutex_lock(&gref_mutex);
466
	priv->users++;
467
	mutex_unlock(&gref_mutex);
468 469
}

470 471
static void gntalloc_vma_close(struct vm_area_struct *vma)
{
472 473 474 475 476
	struct gntalloc_vma_private_data *priv = vma->vm_private_data;
	struct gntalloc_gref *gref, *next;
	int i;

	if (!priv)
477 478
		return;

479
	mutex_lock(&gref_mutex);
480 481 482 483 484 485 486 487 488 489 490 491 492
	priv->users--;
	if (priv->users == 0) {
		gref = priv->gref;
		for (i = 0; i < priv->count; i++) {
			gref->users--;
			next = list_entry(gref->next_gref.next,
					  struct gntalloc_gref, next_gref);
			if (gref->users == 0)
				__del_gref(gref);
			gref = next;
		}
		kfree(priv);
	}
493
	mutex_unlock(&gref_mutex);
494 495 496
}

static struct vm_operations_struct gntalloc_vmops = {
497
	.open = gntalloc_vma_open,
498 499 500 501 502 503
	.close = gntalloc_vma_close,
};

static int gntalloc_mmap(struct file *filp, struct vm_area_struct *vma)
{
	struct gntalloc_file_private_data *priv = filp->private_data;
504
	struct gntalloc_vma_private_data *vm_priv;
505 506 507 508 509 510 511 512 513
	struct gntalloc_gref *gref;
	int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
	int rv, i;

	if (!(vma->vm_flags & VM_SHARED)) {
		printk(KERN_ERR "%s: Mapping must be shared.\n", __func__);
		return -EINVAL;
	}

514 515 516 517
	vm_priv = kmalloc(sizeof(*vm_priv), GFP_KERNEL);
	if (!vm_priv)
		return -ENOMEM;

518
	mutex_lock(&gref_mutex);
519 520 521 522

	pr_debug("%s: priv %p,%p, page %lu+%d\n", __func__,
		       priv, vm_priv, vma->vm_pgoff, count);

523 524 525 526 527
	gref = find_grefs(priv, vma->vm_pgoff << PAGE_SHIFT, count);
	if (gref == NULL) {
		rv = -ENOENT;
		pr_debug("%s: Could not find grant reference",
				__func__);
528
		kfree(vm_priv);
529 530 531
		goto out_unlock;
	}

532 533 534 535 536
	vm_priv->gref = gref;
	vm_priv->users = 1;
	vm_priv->count = count;

	vma->vm_private_data = vm_priv;
537

538
	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554

	vma->vm_ops = &gntalloc_vmops;

	for (i = 0; i < count; i++) {
		gref->users++;
		rv = vm_insert_page(vma, vma->vm_start + i * PAGE_SIZE,
				gref->page);
		if (rv)
			goto out_unlock;

		gref = list_entry(gref->next_file.next,
				struct gntalloc_gref, next_file);
	}
	rv = 0;

out_unlock:
555
	mutex_unlock(&gref_mutex);
556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608
	return rv;
}

static const struct file_operations gntalloc_fops = {
	.owner = THIS_MODULE,
	.open = gntalloc_open,
	.release = gntalloc_release,
	.unlocked_ioctl = gntalloc_ioctl,
	.mmap = gntalloc_mmap
};

/*
 * -------------------------------------
 * Module creation/destruction.
 * -------------------------------------
 */
static struct miscdevice gntalloc_miscdev = {
	.minor	= MISC_DYNAMIC_MINOR,
	.name	= "xen/gntalloc",
	.fops	= &gntalloc_fops,
};

static int __init gntalloc_init(void)
{
	int err;

	if (!xen_domain())
		return -ENODEV;

	err = misc_register(&gntalloc_miscdev);
	if (err != 0) {
		printk(KERN_ERR "Could not register misc gntalloc device\n");
		return err;
	}

	pr_debug("Created grant allocation device at %d,%d\n",
			MISC_MAJOR, gntalloc_miscdev.minor);

	return 0;
}

static void __exit gntalloc_exit(void)
{
	misc_deregister(&gntalloc_miscdev);
}

module_init(gntalloc_init);
module_exit(gntalloc_exit);

MODULE_LICENSE("GPL");
MODULE_AUTHOR("Carter Weatherly <carter.weatherly@jhuapl.edu>, "
		"Daniel De Graaf <dgdegra@tycho.nsa.gov>");
MODULE_DESCRIPTION("User-space grant reference allocator driver");