1/******************************************************************************
2 * gntalloc.c
3 *
4 * Device for creating grant references (in user-space) that may be shared
5 * with other domains.
6 *
7 * This program is distributed in the hope that it will be useful,
8 * but WITHOUT ANY WARRANTY; without even the implied warranty of
9 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
10 * GNU General Public License for more details.
11 *
12 * You should have received a copy of the GNU General Public License
13 * along with this program; if not, write to the Free Software
14 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
15 */
16
17/*
18 * This driver exists to allow userspace programs in Linux to allocate kernel
19 * memory that will later be shared with another domain.  Without this device,
20 * Linux userspace programs cannot create grant references.
21 *
22 * How this stuff works:
23 *   X -> granting a page to Y
24 *   Y -> mapping the grant from X
25 *
26 *   1. X uses the gntalloc device to allocate a page of kernel memory, P.
27 *   2. X creates an entry in the grant table that says domid(Y) can access P.
28 *      This is done without a hypercall unless the grant table needs expansion.
29 *   3. X gives the grant reference identifier, GREF, to Y.
30 *   4. Y maps the page, either directly into kernel memory for use in a backend
31 *      driver, or via a the gntdev device to map into the address space of an
32 *      application running in Y. This is the first point at which Xen does any
33 *      tracking of the page.
34 *   5. A program in X mmap()s a segment of the gntalloc device that corresponds
35 *      to the shared page, and can now communicate with Y over the shared page.
36 *
37 *
38 * NOTE TO USERSPACE LIBRARIES:
39 *   The grant allocation and mmap()ing are, naturally, two separate operations.
40 *   You set up the sharing by calling the create ioctl() and then the mmap().
41 *   Teardown requires munmap() and either close() or ioctl().
42 *
43 * WARNING: Since Xen does not allow a guest to forcibly end the use of a grant
44 * reference, this device can be used to consume kernel memory by leaving grant
45 * references mapped by another domain when an application exits. Therefore,
46 * there is a global limit on the number of pages that can be allocated. When
47 * all references to the page are unmapped, it will be freed during the next
48 * grant operation.
49 */
50
51#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
52
53#include <linux/atomic.h>
54#include <linux/module.h>
55#include <linux/miscdevice.h>
56#include <linux/kernel.h>
57#include <linux/init.h>
58#include <linux/slab.h>
59#include <linux/fs.h>
60#include <linux/device.h>
61#include <linux/mm.h>
62#include <linux/uaccess.h>
63#include <linux/types.h>
64#include <linux/list.h>
65#include <linux/highmem.h>
66
67#include <xen/xen.h>
68#include <xen/page.h>
69#include <xen/grant_table.h>
70#include <xen/gntalloc.h>
71#include <xen/events.h>
72
73static int limit = 1024;
74module_param(limit, int, 0644);
75MODULE_PARM_DESC(limit, "Maximum number of grants that may be allocated by "
76		"the gntalloc device");
77
78static LIST_HEAD(gref_list);
79static DEFINE_MUTEX(gref_mutex);
80static int gref_size;
81
82struct notify_info {
83	uint16_t pgoff:12;    /* Bits 0-11: Offset of the byte to clear */
84	uint16_t flags:2;     /* Bits 12-13: Unmap notification flags */
85	int event;            /* Port (event channel) to notify */
86};
87
88/* Metadata on a grant reference. */
89struct gntalloc_gref {
90	struct list_head next_gref;  /* list entry gref_list */
91	struct list_head next_file;  /* list entry file->list, if open */
92	struct page *page;	     /* The shared page */
93	uint64_t file_index;         /* File offset for mmap() */
94	unsigned int users;          /* Use count - when zero, waiting on Xen */
95	grant_ref_t gref_id;         /* The grant reference number */
96	struct notify_info notify;   /* Unmap notification */
97};
98
99struct gntalloc_file_private_data {
100	struct list_head list;
101	uint64_t index;
102};
103
104struct gntalloc_vma_private_data {
105	struct gntalloc_gref *gref;
106	int users;
107	int count;
108};
109
110static void __del_gref(struct gntalloc_gref *gref);
111
112static void do_cleanup(void)
113{
114	struct gntalloc_gref *gref, *n;
115	list_for_each_entry_safe(gref, n, &gref_list, next_gref) {
116		if (!gref->users)
117			__del_gref(gref);
118	}
119}
120
121static int add_grefs(struct ioctl_gntalloc_alloc_gref *op,
122	uint32_t *gref_ids, struct gntalloc_file_private_data *priv)
123{
124	int i, rc, readonly;
125	LIST_HEAD(queue_gref);
126	LIST_HEAD(queue_file);
127	struct gntalloc_gref *gref, *next;
128
129	readonly = !(op->flags & GNTALLOC_FLAG_WRITABLE);
130	rc = -ENOMEM;
131	for (i = 0; i < op->count; i++) {
132		gref = kzalloc(sizeof(*gref), GFP_KERNEL);
133		if (!gref)
134			goto undo;
135		list_add_tail(&gref->next_gref, &queue_gref);
136		list_add_tail(&gref->next_file, &queue_file);
137		gref->users = 1;
138		gref->file_index = op->index + i * PAGE_SIZE;
139		gref->page = alloc_page(GFP_KERNEL|__GFP_ZERO);
140		if (!gref->page)
141			goto undo;
142
143		/* Grant foreign access to the page. */
144		rc = gnttab_grant_foreign_access(op->domid,
145						 xen_page_to_gfn(gref->page),
146						 readonly);
147		if (rc < 0)
148			goto undo;
149		gref_ids[i] = gref->gref_id = rc;
150	}
151
152	/* Add to gref lists. */
153	mutex_lock(&gref_mutex);
154	list_splice_tail(&queue_gref, &gref_list);
155	list_splice_tail(&queue_file, &priv->list);
156	mutex_unlock(&gref_mutex);
157
158	return 0;
159
160undo:
161	mutex_lock(&gref_mutex);
162	gref_size -= (op->count - i);
163
164	list_for_each_entry_safe(gref, next, &queue_file, next_file) {
165		list_del(&gref->next_file);
166		__del_gref(gref);
167	}
168
169	/* It's possible for the target domain to map the just-allocated grant
170	 * references by blindly guessing their IDs; if this is done, then
171	 * __del_gref will leave them in the queue_gref list. They need to be
172	 * added to the global list so that we can free them when they are no
173	 * longer referenced.
174	 */
175	if (unlikely(!list_empty(&queue_gref)))
176		list_splice_tail(&queue_gref, &gref_list);
177	mutex_unlock(&gref_mutex);
178	return rc;
179}
180
181static void __del_gref(struct gntalloc_gref *gref)
182{
183	if (gref->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
184		uint8_t *tmp = kmap(gref->page);
185		tmp[gref->notify.pgoff] = 0;
186		kunmap(gref->page);
187	}
188	if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
189		notify_remote_via_evtchn(gref->notify.event);
190		evtchn_put(gref->notify.event);
191	}
192
193	gref->notify.flags = 0;
194
195	if (gref->gref_id) {
196		if (gnttab_query_foreign_access(gref->gref_id))
197			return;
198
199		if (!gnttab_end_foreign_access_ref(gref->gref_id, 0))
200			return;
201
202		gnttab_free_grant_reference(gref->gref_id);
203	}
204
205	gref_size--;
206	list_del(&gref->next_gref);
207
208	if (gref->page)
209		__free_page(gref->page);
210
211	kfree(gref);
212}
213
214/* finds contiguous grant references in a file, returns the first */
215static struct gntalloc_gref *find_grefs(struct gntalloc_file_private_data *priv,
216		uint64_t index, uint32_t count)
217{
218	struct gntalloc_gref *rv = NULL, *gref;
219	list_for_each_entry(gref, &priv->list, next_file) {
220		if (gref->file_index == index && !rv)
221			rv = gref;
222		if (rv) {
223			if (gref->file_index != index)
224				return NULL;
225			index += PAGE_SIZE;
226			count--;
227			if (count == 0)
228				return rv;
229		}
230	}
231	return NULL;
232}
233
234/*
235 * -------------------------------------
236 *  File operations.
237 * -------------------------------------
238 */
239static int gntalloc_open(struct inode *inode, struct file *filp)
240{
241	struct gntalloc_file_private_data *priv;
242
243	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
244	if (!priv)
245		goto out_nomem;
246	INIT_LIST_HEAD(&priv->list);
247
248	filp->private_data = priv;
249
250	pr_debug("%s: priv %p\n", __func__, priv);
251
252	return 0;
253
254out_nomem:
255	return -ENOMEM;
256}
257
258static int gntalloc_release(struct inode *inode, struct file *filp)
259{
260	struct gntalloc_file_private_data *priv = filp->private_data;
261	struct gntalloc_gref *gref;
262
263	pr_debug("%s: priv %p\n", __func__, priv);
264
265	mutex_lock(&gref_mutex);
266	while (!list_empty(&priv->list)) {
267		gref = list_entry(priv->list.next,
268			struct gntalloc_gref, next_file);
269		list_del(&gref->next_file);
270		gref->users--;
271		if (gref->users == 0)
272			__del_gref(gref);
273	}
274	kfree(priv);
275	mutex_unlock(&gref_mutex);
276
277	return 0;
278}
279
280static long gntalloc_ioctl_alloc(struct gntalloc_file_private_data *priv,
281		struct ioctl_gntalloc_alloc_gref __user *arg)
282{
283	int rc = 0;
284	struct ioctl_gntalloc_alloc_gref op;
285	uint32_t *gref_ids;
286
287	pr_debug("%s: priv %p\n", __func__, priv);
288
289	if (copy_from_user(&op, arg, sizeof(op))) {
290		rc = -EFAULT;
291		goto out;
292	}
293
294	gref_ids = kcalloc(op.count, sizeof(gref_ids[0]), GFP_TEMPORARY);
295	if (!gref_ids) {
296		rc = -ENOMEM;
297		goto out;
298	}
299
300	mutex_lock(&gref_mutex);
301	/* Clean up pages that were at zero (local) users but were still mapped
302	 * by remote domains. Since those pages count towards the limit that we
303	 * are about to enforce, removing them here is a good idea.
304	 */
305	do_cleanup();
306	if (gref_size + op.count > limit) {
307		mutex_unlock(&gref_mutex);
308		rc = -ENOSPC;
309		goto out_free;
310	}
311	gref_size += op.count;
312	op.index = priv->index;
313	priv->index += op.count * PAGE_SIZE;
314	mutex_unlock(&gref_mutex);
315
316	rc = add_grefs(&op, gref_ids, priv);
317	if (rc < 0)
318		goto out_free;
319
320	/* Once we finish add_grefs, it is unsafe to touch the new reference,
321	 * since it is possible for a concurrent ioctl to remove it (by guessing
322	 * its index). If the userspace application doesn't provide valid memory
323	 * to write the IDs to, then it will need to close the file in order to
324	 * release - which it will do by segfaulting when it tries to access the
325	 * IDs to close them.
326	 */
327	if (copy_to_user(arg, &op, sizeof(op))) {
328		rc = -EFAULT;
329		goto out_free;
330	}
331	if (copy_to_user(arg->gref_ids, gref_ids,
332			sizeof(gref_ids[0]) * op.count)) {
333		rc = -EFAULT;
334		goto out_free;
335	}
336
337out_free:
338	kfree(gref_ids);
339out:
340	return rc;
341}
342
343static long gntalloc_ioctl_dealloc(struct gntalloc_file_private_data *priv,
344		void __user *arg)
345{
346	int i, rc = 0;
347	struct ioctl_gntalloc_dealloc_gref op;
348	struct gntalloc_gref *gref, *n;
349
350	pr_debug("%s: priv %p\n", __func__, priv);
351
352	if (copy_from_user(&op, arg, sizeof(op))) {
353		rc = -EFAULT;
354		goto dealloc_grant_out;
355	}
356
357	mutex_lock(&gref_mutex);
358	gref = find_grefs(priv, op.index, op.count);
359	if (gref) {
360		/* Remove from the file list only, and decrease reference count.
361		 * The later call to do_cleanup() will remove from gref_list and
362		 * free the memory if the pages aren't mapped anywhere.
363		 */
364		for (i = 0; i < op.count; i++) {
365			n = list_entry(gref->next_file.next,
366				struct gntalloc_gref, next_file);
367			list_del(&gref->next_file);
368			gref->users--;
369			gref = n;
370		}
371	} else {
372		rc = -EINVAL;
373	}
374
375	do_cleanup();
376
377	mutex_unlock(&gref_mutex);
378dealloc_grant_out:
379	return rc;
380}
381
382static long gntalloc_ioctl_unmap_notify(struct gntalloc_file_private_data *priv,
383		void __user *arg)
384{
385	struct ioctl_gntalloc_unmap_notify op;
386	struct gntalloc_gref *gref;
387	uint64_t index;
388	int pgoff;
389	int rc;
390
391	if (copy_from_user(&op, arg, sizeof(op)))
392		return -EFAULT;
393
394	index = op.index & ~(PAGE_SIZE - 1);
395	pgoff = op.index & (PAGE_SIZE - 1);
396
397	mutex_lock(&gref_mutex);
398
399	gref = find_grefs(priv, index, 1);
400	if (!gref) {
401		rc = -ENOENT;
402		goto unlock_out;
403	}
404
405	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) {
406		rc = -EINVAL;
407		goto unlock_out;
408	}
409
410	/* We need to grab a reference to the event channel we are going to use
411	 * to send the notify before releasing the reference we may already have
412	 * (if someone has called this ioctl twice). This is required so that
413	 * it is possible to change the clear_byte part of the notification
414	 * without disturbing the event channel part, which may now be the last
415	 * reference to that event channel.
416	 */
417	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
418		if (evtchn_get(op.event_channel_port)) {
419			rc = -EINVAL;
420			goto unlock_out;
421		}
422	}
423
424	if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT)
425		evtchn_put(gref->notify.event);
426
427	gref->notify.flags = op.action;
428	gref->notify.pgoff = pgoff;
429	gref->notify.event = op.event_channel_port;
430	rc = 0;
431
432 unlock_out:
433	mutex_unlock(&gref_mutex);
434	return rc;
435}
436
437static long gntalloc_ioctl(struct file *filp, unsigned int cmd,
438		unsigned long arg)
439{
440	struct gntalloc_file_private_data *priv = filp->private_data;
441
442	switch (cmd) {
443	case IOCTL_GNTALLOC_ALLOC_GREF:
444		return gntalloc_ioctl_alloc(priv, (void __user *)arg);
445
446	case IOCTL_GNTALLOC_DEALLOC_GREF:
447		return gntalloc_ioctl_dealloc(priv, (void __user *)arg);
448
449	case IOCTL_GNTALLOC_SET_UNMAP_NOTIFY:
450		return gntalloc_ioctl_unmap_notify(priv, (void __user *)arg);
451
452	default:
453		return -ENOIOCTLCMD;
454	}
455
456	return 0;
457}
458
459static void gntalloc_vma_open(struct vm_area_struct *vma)
460{
461	struct gntalloc_vma_private_data *priv = vma->vm_private_data;
462
463	if (!priv)
464		return;
465
466	mutex_lock(&gref_mutex);
467	priv->users++;
468	mutex_unlock(&gref_mutex);
469}
470
471static void gntalloc_vma_close(struct vm_area_struct *vma)
472{
473	struct gntalloc_vma_private_data *priv = vma->vm_private_data;
474	struct gntalloc_gref *gref, *next;
475	int i;
476
477	if (!priv)
478		return;
479
480	mutex_lock(&gref_mutex);
481	priv->users--;
482	if (priv->users == 0) {
483		gref = priv->gref;
484		for (i = 0; i < priv->count; i++) {
485			gref->users--;
486			next = list_entry(gref->next_gref.next,
487					  struct gntalloc_gref, next_gref);
488			if (gref->users == 0)
489				__del_gref(gref);
490			gref = next;
491		}
492		kfree(priv);
493	}
494	mutex_unlock(&gref_mutex);
495}
496
497static const struct vm_operations_struct gntalloc_vmops = {
498	.open = gntalloc_vma_open,
499	.close = gntalloc_vma_close,
500};
501
502static int gntalloc_mmap(struct file *filp, struct vm_area_struct *vma)
503{
504	struct gntalloc_file_private_data *priv = filp->private_data;
505	struct gntalloc_vma_private_data *vm_priv;
506	struct gntalloc_gref *gref;
507	int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
508	int rv, i;
509
510	if (!(vma->vm_flags & VM_SHARED)) {
511		pr_err("%s: Mapping must be shared\n", __func__);
512		return -EINVAL;
513	}
514
515	vm_priv = kmalloc(sizeof(*vm_priv), GFP_KERNEL);
516	if (!vm_priv)
517		return -ENOMEM;
518
519	mutex_lock(&gref_mutex);
520
521	pr_debug("%s: priv %p,%p, page %lu+%d\n", __func__,
522		       priv, vm_priv, vma->vm_pgoff, count);
523
524	gref = find_grefs(priv, vma->vm_pgoff << PAGE_SHIFT, count);
525	if (gref == NULL) {
526		rv = -ENOENT;
527		pr_debug("%s: Could not find grant reference",
528				__func__);
529		kfree(vm_priv);
530		goto out_unlock;
531	}
532
533	vm_priv->gref = gref;
534	vm_priv->users = 1;
535	vm_priv->count = count;
536
537	vma->vm_private_data = vm_priv;
538
539	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
540
541	vma->vm_ops = &gntalloc_vmops;
542
543	for (i = 0; i < count; i++) {
544		gref->users++;
545		rv = vm_insert_page(vma, vma->vm_start + i * PAGE_SIZE,
546				gref->page);
547		if (rv)
548			goto out_unlock;
549
550		gref = list_entry(gref->next_file.next,
551				struct gntalloc_gref, next_file);
552	}
553	rv = 0;
554
555out_unlock:
556	mutex_unlock(&gref_mutex);
557	return rv;
558}
559
560static const struct file_operations gntalloc_fops = {
561	.owner = THIS_MODULE,
562	.open = gntalloc_open,
563	.release = gntalloc_release,
564	.unlocked_ioctl = gntalloc_ioctl,
565	.mmap = gntalloc_mmap
566};
567
568/*
569 * -------------------------------------
570 * Module creation/destruction.
571 * -------------------------------------
572 */
573static struct miscdevice gntalloc_miscdev = {
574	.minor	= MISC_DYNAMIC_MINOR,
575	.name	= "xen/gntalloc",
576	.fops	= &gntalloc_fops,
577};
578
579static int __init gntalloc_init(void)
580{
581	int err;
582
583	if (!xen_domain())
584		return -ENODEV;
585
586	err = misc_register(&gntalloc_miscdev);
587	if (err != 0) {
588		pr_err("Could not register misc gntalloc device\n");
589		return err;
590	}
591
592	pr_debug("Created grant allocation device at %d,%d\n",
593			MISC_MAJOR, gntalloc_miscdev.minor);
594
595	return 0;
596}
597
598static void __exit gntalloc_exit(void)
599{
600	misc_deregister(&gntalloc_miscdev);
601}
602
603module_init(gntalloc_init);
604module_exit(gntalloc_exit);
605
606MODULE_LICENSE("GPL");
607MODULE_AUTHOR("Carter Weatherly <carter.weatherly@jhuapl.edu>, "
608		"Daniel De Graaf <dgdegra@tycho.nsa.gov>");
609MODULE_DESCRIPTION("User-space grant reference allocator driver");
610