blob: a7ffdfe19fc9b06141792bde2d3ee10f61f10dbb [file] [log] [blame]
Daniel De Graafdd314052011-02-07 17:23:05 -05001/******************************************************************************
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#include <linux/atomic.h>
52#include <linux/module.h>
53#include <linux/miscdevice.h>
54#include <linux/kernel.h>
55#include <linux/init.h>
56#include <linux/slab.h>
57#include <linux/fs.h>
58#include <linux/device.h>
59#include <linux/mm.h>
60#include <linux/uaccess.h>
61#include <linux/types.h>
62#include <linux/list.h>
Daniel De Graafbdc612d2011-02-03 12:19:04 -050063#include <linux/highmem.h>
Daniel De Graafdd314052011-02-07 17:23:05 -050064
65#include <xen/xen.h>
66#include <xen/page.h>
67#include <xen/grant_table.h>
68#include <xen/gntalloc.h>
Daniel De Graafbdc612d2011-02-03 12:19:04 -050069#include <xen/events.h>
Daniel De Graafdd314052011-02-07 17:23:05 -050070
71static int limit = 1024;
72module_param(limit, int, 0644);
73MODULE_PARM_DESC(limit, "Maximum number of grants that may be allocated by "
74 "the gntalloc device");
75
76static LIST_HEAD(gref_list);
77static DEFINE_SPINLOCK(gref_lock);
78static int gref_size;
79
Daniel De Graafbdc612d2011-02-03 12:19:04 -050080struct notify_info {
81 uint16_t pgoff:12; /* Bits 0-11: Offset of the byte to clear */
82 uint16_t flags:2; /* Bits 12-13: Unmap notification flags */
83 int event; /* Port (event channel) to notify */
84};
85
Daniel De Graafdd314052011-02-07 17:23:05 -050086/* Metadata on a grant reference. */
87struct gntalloc_gref {
88 struct list_head next_gref; /* list entry gref_list */
89 struct list_head next_file; /* list entry file->list, if open */
90 struct page *page; /* The shared page */
91 uint64_t file_index; /* File offset for mmap() */
92 unsigned int users; /* Use count - when zero, waiting on Xen */
93 grant_ref_t gref_id; /* The grant reference number */
Daniel De Graafbdc612d2011-02-03 12:19:04 -050094 struct notify_info notify; /* Unmap notification */
Daniel De Graafdd314052011-02-07 17:23:05 -050095};
96
97struct gntalloc_file_private_data {
98 struct list_head list;
99 uint64_t index;
100};
101
102static void __del_gref(struct gntalloc_gref *gref);
103
104static void do_cleanup(void)
105{
106 struct gntalloc_gref *gref, *n;
107 list_for_each_entry_safe(gref, n, &gref_list, next_gref) {
108 if (!gref->users)
109 __del_gref(gref);
110 }
111}
112
113static int add_grefs(struct ioctl_gntalloc_alloc_gref *op,
114 uint32_t *gref_ids, struct gntalloc_file_private_data *priv)
115{
116 int i, rc, readonly;
117 LIST_HEAD(queue_gref);
118 LIST_HEAD(queue_file);
119 struct gntalloc_gref *gref;
120
121 readonly = !(op->flags & GNTALLOC_FLAG_WRITABLE);
122 rc = -ENOMEM;
123 for (i = 0; i < op->count; i++) {
124 gref = kzalloc(sizeof(*gref), GFP_KERNEL);
125 if (!gref)
126 goto undo;
127 list_add_tail(&gref->next_gref, &queue_gref);
128 list_add_tail(&gref->next_file, &queue_file);
129 gref->users = 1;
130 gref->file_index = op->index + i * PAGE_SIZE;
131 gref->page = alloc_page(GFP_KERNEL|__GFP_ZERO);
132 if (!gref->page)
133 goto undo;
134
135 /* Grant foreign access to the page. */
136 gref->gref_id = gnttab_grant_foreign_access(op->domid,
137 pfn_to_mfn(page_to_pfn(gref->page)), readonly);
138 if (gref->gref_id < 0) {
139 rc = gref->gref_id;
140 goto undo;
141 }
142 gref_ids[i] = gref->gref_id;
143 }
144
145 /* Add to gref lists. */
146 spin_lock(&gref_lock);
147 list_splice_tail(&queue_gref, &gref_list);
148 list_splice_tail(&queue_file, &priv->list);
149 spin_unlock(&gref_lock);
150
151 return 0;
152
153undo:
154 spin_lock(&gref_lock);
155 gref_size -= (op->count - i);
156
157 list_for_each_entry(gref, &queue_file, next_file) {
158 /* __del_gref does not remove from queue_file */
159 __del_gref(gref);
160 }
161
162 /* It's possible for the target domain to map the just-allocated grant
163 * references by blindly guessing their IDs; if this is done, then
164 * __del_gref will leave them in the queue_gref list. They need to be
165 * added to the global list so that we can free them when they are no
166 * longer referenced.
167 */
168 if (unlikely(!list_empty(&queue_gref)))
169 list_splice_tail(&queue_gref, &gref_list);
170 spin_unlock(&gref_lock);
171 return rc;
172}
173
174static void __del_gref(struct gntalloc_gref *gref)
175{
Daniel De Graafbdc612d2011-02-03 12:19:04 -0500176 if (gref->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
177 uint8_t *tmp = kmap(gref->page);
178 tmp[gref->notify.pgoff] = 0;
179 kunmap(gref->page);
180 }
181 if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT)
182 notify_remote_via_evtchn(gref->notify.event);
183
184 gref->notify.flags = 0;
185
Daniel De Graafdd314052011-02-07 17:23:05 -0500186 if (gref->gref_id > 0) {
187 if (gnttab_query_foreign_access(gref->gref_id))
188 return;
189
190 if (!gnttab_end_foreign_access_ref(gref->gref_id, 0))
191 return;
192 }
193
194 gref_size--;
195 list_del(&gref->next_gref);
196
197 if (gref->page)
198 __free_page(gref->page);
199
200 kfree(gref);
201}
202
203/* finds contiguous grant references in a file, returns the first */
204static struct gntalloc_gref *find_grefs(struct gntalloc_file_private_data *priv,
205 uint64_t index, uint32_t count)
206{
207 struct gntalloc_gref *rv = NULL, *gref;
208 list_for_each_entry(gref, &priv->list, next_file) {
209 if (gref->file_index == index && !rv)
210 rv = gref;
211 if (rv) {
212 if (gref->file_index != index)
213 return NULL;
214 index += PAGE_SIZE;
215 count--;
216 if (count == 0)
217 return rv;
218 }
219 }
220 return NULL;
221}
222
223/*
224 * -------------------------------------
225 * File operations.
226 * -------------------------------------
227 */
228static int gntalloc_open(struct inode *inode, struct file *filp)
229{
230 struct gntalloc_file_private_data *priv;
231
232 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
233 if (!priv)
234 goto out_nomem;
235 INIT_LIST_HEAD(&priv->list);
236
237 filp->private_data = priv;
238
239 pr_debug("%s: priv %p\n", __func__, priv);
240
241 return 0;
242
243out_nomem:
244 return -ENOMEM;
245}
246
247static int gntalloc_release(struct inode *inode, struct file *filp)
248{
249 struct gntalloc_file_private_data *priv = filp->private_data;
250 struct gntalloc_gref *gref;
251
252 pr_debug("%s: priv %p\n", __func__, priv);
253
254 spin_lock(&gref_lock);
255 while (!list_empty(&priv->list)) {
256 gref = list_entry(priv->list.next,
257 struct gntalloc_gref, next_file);
258 list_del(&gref->next_file);
259 gref->users--;
260 if (gref->users == 0)
261 __del_gref(gref);
262 }
263 kfree(priv);
264 spin_unlock(&gref_lock);
265
266 return 0;
267}
268
269static long gntalloc_ioctl_alloc(struct gntalloc_file_private_data *priv,
270 struct ioctl_gntalloc_alloc_gref __user *arg)
271{
272 int rc = 0;
273 struct ioctl_gntalloc_alloc_gref op;
274 uint32_t *gref_ids;
275
276 pr_debug("%s: priv %p\n", __func__, priv);
277
278 if (copy_from_user(&op, arg, sizeof(op))) {
279 rc = -EFAULT;
280 goto out;
281 }
282
283 gref_ids = kzalloc(sizeof(gref_ids[0]) * op.count, GFP_TEMPORARY);
284 if (!gref_ids) {
285 rc = -ENOMEM;
286 goto out;
287 }
288
289 spin_lock(&gref_lock);
290 /* Clean up pages that were at zero (local) users but were still mapped
291 * by remote domains. Since those pages count towards the limit that we
292 * are about to enforce, removing them here is a good idea.
293 */
294 do_cleanup();
295 if (gref_size + op.count > limit) {
296 spin_unlock(&gref_lock);
297 rc = -ENOSPC;
298 goto out_free;
299 }
300 gref_size += op.count;
301 op.index = priv->index;
302 priv->index += op.count * PAGE_SIZE;
303 spin_unlock(&gref_lock);
304
305 rc = add_grefs(&op, gref_ids, priv);
306 if (rc < 0)
307 goto out_free;
308
309 /* Once we finish add_grefs, it is unsafe to touch the new reference,
310 * since it is possible for a concurrent ioctl to remove it (by guessing
311 * its index). If the userspace application doesn't provide valid memory
312 * to write the IDs to, then it will need to close the file in order to
313 * release - which it will do by segfaulting when it tries to access the
314 * IDs to close them.
315 */
316 if (copy_to_user(arg, &op, sizeof(op))) {
317 rc = -EFAULT;
318 goto out_free;
319 }
320 if (copy_to_user(arg->gref_ids, gref_ids,
321 sizeof(gref_ids[0]) * op.count)) {
322 rc = -EFAULT;
323 goto out_free;
324 }
325
326out_free:
327 kfree(gref_ids);
328out:
329 return rc;
330}
331
332static long gntalloc_ioctl_dealloc(struct gntalloc_file_private_data *priv,
333 void __user *arg)
334{
335 int i, rc = 0;
336 struct ioctl_gntalloc_dealloc_gref op;
337 struct gntalloc_gref *gref, *n;
338
339 pr_debug("%s: priv %p\n", __func__, priv);
340
341 if (copy_from_user(&op, arg, sizeof(op))) {
342 rc = -EFAULT;
343 goto dealloc_grant_out;
344 }
345
346 spin_lock(&gref_lock);
347 gref = find_grefs(priv, op.index, op.count);
348 if (gref) {
349 /* Remove from the file list only, and decrease reference count.
350 * The later call to do_cleanup() will remove from gref_list and
351 * free the memory if the pages aren't mapped anywhere.
352 */
353 for (i = 0; i < op.count; i++) {
354 n = list_entry(gref->next_file.next,
355 struct gntalloc_gref, next_file);
356 list_del(&gref->next_file);
357 gref->users--;
358 gref = n;
359 }
360 } else {
361 rc = -EINVAL;
362 }
363
364 do_cleanup();
365
366 spin_unlock(&gref_lock);
367dealloc_grant_out:
368 return rc;
369}
370
Daniel De Graafbdc612d2011-02-03 12:19:04 -0500371static long gntalloc_ioctl_unmap_notify(struct gntalloc_file_private_data *priv,
372 void __user *arg)
373{
374 struct ioctl_gntalloc_unmap_notify op;
375 struct gntalloc_gref *gref;
376 uint64_t index;
377 int pgoff;
378 int rc;
379
380 if (copy_from_user(&op, arg, sizeof(op)))
381 return -EFAULT;
382
383 index = op.index & ~(PAGE_SIZE - 1);
384 pgoff = op.index & (PAGE_SIZE - 1);
385
386 spin_lock(&gref_lock);
387
388 gref = find_grefs(priv, index, 1);
389 if (!gref) {
390 rc = -ENOENT;
391 goto unlock_out;
392 }
393
394 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) {
395 rc = -EINVAL;
396 goto unlock_out;
397 }
398
399 gref->notify.flags = op.action;
400 gref->notify.pgoff = pgoff;
401 gref->notify.event = op.event_channel_port;
402 rc = 0;
403 unlock_out:
404 spin_unlock(&gref_lock);
405 return rc;
406}
407
Daniel De Graafdd314052011-02-07 17:23:05 -0500408static long gntalloc_ioctl(struct file *filp, unsigned int cmd,
409 unsigned long arg)
410{
411 struct gntalloc_file_private_data *priv = filp->private_data;
412
413 switch (cmd) {
414 case IOCTL_GNTALLOC_ALLOC_GREF:
415 return gntalloc_ioctl_alloc(priv, (void __user *)arg);
416
417 case IOCTL_GNTALLOC_DEALLOC_GREF:
418 return gntalloc_ioctl_dealloc(priv, (void __user *)arg);
419
Daniel De Graafbdc612d2011-02-03 12:19:04 -0500420 case IOCTL_GNTALLOC_SET_UNMAP_NOTIFY:
421 return gntalloc_ioctl_unmap_notify(priv, (void __user *)arg);
422
Daniel De Graafdd314052011-02-07 17:23:05 -0500423 default:
424 return -ENOIOCTLCMD;
425 }
426
427 return 0;
428}
429
430static void gntalloc_vma_close(struct vm_area_struct *vma)
431{
432 struct gntalloc_gref *gref = vma->vm_private_data;
433 if (!gref)
434 return;
435
436 spin_lock(&gref_lock);
437 gref->users--;
438 if (gref->users == 0)
439 __del_gref(gref);
440 spin_unlock(&gref_lock);
441}
442
443static struct vm_operations_struct gntalloc_vmops = {
444 .close = gntalloc_vma_close,
445};
446
447static int gntalloc_mmap(struct file *filp, struct vm_area_struct *vma)
448{
449 struct gntalloc_file_private_data *priv = filp->private_data;
450 struct gntalloc_gref *gref;
451 int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
452 int rv, i;
453
454 pr_debug("%s: priv %p, page %lu+%d\n", __func__,
455 priv, vma->vm_pgoff, count);
456
457 if (!(vma->vm_flags & VM_SHARED)) {
458 printk(KERN_ERR "%s: Mapping must be shared.\n", __func__);
459 return -EINVAL;
460 }
461
462 spin_lock(&gref_lock);
463 gref = find_grefs(priv, vma->vm_pgoff << PAGE_SHIFT, count);
464 if (gref == NULL) {
465 rv = -ENOENT;
466 pr_debug("%s: Could not find grant reference",
467 __func__);
468 goto out_unlock;
469 }
470
471 vma->vm_private_data = gref;
472
473 vma->vm_flags |= VM_RESERVED;
474 vma->vm_flags |= VM_DONTCOPY;
475 vma->vm_flags |= VM_PFNMAP | VM_PFN_AT_MMAP;
476
477 vma->vm_ops = &gntalloc_vmops;
478
479 for (i = 0; i < count; i++) {
480 gref->users++;
481 rv = vm_insert_page(vma, vma->vm_start + i * PAGE_SIZE,
482 gref->page);
483 if (rv)
484 goto out_unlock;
485
486 gref = list_entry(gref->next_file.next,
487 struct gntalloc_gref, next_file);
488 }
489 rv = 0;
490
491out_unlock:
492 spin_unlock(&gref_lock);
493 return rv;
494}
495
496static const struct file_operations gntalloc_fops = {
497 .owner = THIS_MODULE,
498 .open = gntalloc_open,
499 .release = gntalloc_release,
500 .unlocked_ioctl = gntalloc_ioctl,
501 .mmap = gntalloc_mmap
502};
503
504/*
505 * -------------------------------------
506 * Module creation/destruction.
507 * -------------------------------------
508 */
509static struct miscdevice gntalloc_miscdev = {
510 .minor = MISC_DYNAMIC_MINOR,
511 .name = "xen/gntalloc",
512 .fops = &gntalloc_fops,
513};
514
515static int __init gntalloc_init(void)
516{
517 int err;
518
519 if (!xen_domain())
520 return -ENODEV;
521
522 err = misc_register(&gntalloc_miscdev);
523 if (err != 0) {
524 printk(KERN_ERR "Could not register misc gntalloc device\n");
525 return err;
526 }
527
528 pr_debug("Created grant allocation device at %d,%d\n",
529 MISC_MAJOR, gntalloc_miscdev.minor);
530
531 return 0;
532}
533
534static void __exit gntalloc_exit(void)
535{
536 misc_deregister(&gntalloc_miscdev);
537}
538
539module_init(gntalloc_init);
540module_exit(gntalloc_exit);
541
542MODULE_LICENSE("GPL");
543MODULE_AUTHOR("Carter Weatherly <carter.weatherly@jhuapl.edu>, "
544 "Daniel De Graaf <dgdegra@tycho.nsa.gov>");
545MODULE_DESCRIPTION("User-space grant reference allocator driver");