1 /******************************************************************************
2 * gntdev.c
3 *
4 * Device for accessing (in user-space) pages that have been granted by other
5 * domains.
6 *
7 * Copyright (c) 2006-2007, D G Murray.
8 * (c) 2009 Gerd Hoffmann <kraxel@redhat.com>
9 * (c) 2018 Oleksandr Andrushchenko, EPAM Systems Inc.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 */
20
21 #undef DEBUG
22
23 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
24
25 #include <linux/dma-mapping.h>
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/init.h>
29 #include <linux/miscdevice.h>
30 #include <linux/fs.h>
31 #include <linux/uaccess.h>
32 #include <linux/sched.h>
33 #include <linux/sched/mm.h>
34 #include <linux/spinlock.h>
35 #include <linux/slab.h>
36 #include <linux/highmem.h>
37 #include <linux/refcount.h>
38 #include <linux/workqueue.h>
39
40 #include <xen/xen.h>
41 #include <xen/grant_table.h>
42 #include <xen/balloon.h>
43 #include <xen/gntdev.h>
44 #include <xen/events.h>
45 #include <xen/page.h>
46 #include <asm/xen/hypervisor.h>
47 #include <asm/xen/hypercall.h>
48
49 #include "gntdev-common.h"
50 #ifdef CONFIG_XEN_GNTDEV_DMABUF
51 #include "gntdev-dmabuf.h"
52 #endif
53
54 MODULE_LICENSE("GPL");
55 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
56 "Gerd Hoffmann <kraxel@redhat.com>");
57 MODULE_DESCRIPTION("User-space granted page access driver");
58
59 #define GNTDEV_COPY_BATCH 16
60
61 struct gntdev_copy_batch {
62 struct gnttab_copy ops[GNTDEV_COPY_BATCH];
63 struct page *pages[GNTDEV_COPY_BATCH];
64 s16 __user *status[GNTDEV_COPY_BATCH];
65 unsigned int nr_ops;
66 unsigned int nr_pages;
67 bool writeable;
68 struct gntdev_copy_batch *next;
69 };
70
71 static unsigned int limit = 64*1024;
72 module_param(limit, uint, 0644);
73 MODULE_PARM_DESC(limit,
74 "Maximum number of grants that may be mapped by one mapping request");
75
76 /* True in PV mode, false otherwise */
77 static int use_ptemod;
78
79 static void unmap_grant_pages(struct gntdev_grant_map *map,
80 int offset, int pages);
81
82 static struct miscdevice gntdev_miscdev;
83
84 /* ------------------------------------------------------------------ */
85
gntdev_test_page_count(unsigned int count)86 bool gntdev_test_page_count(unsigned int count)
87 {
88 return !count || count > limit;
89 }
90
gntdev_print_maps(struct gntdev_priv * priv,char * text,int text_index)91 static void gntdev_print_maps(struct gntdev_priv *priv,
92 char *text, int text_index)
93 {
94 #ifdef DEBUG
95 struct gntdev_grant_map *map;
96
97 pr_debug("%s: maps list (priv %p)\n", __func__, priv);
98 list_for_each_entry(map, &priv->maps, next)
99 pr_debug(" index %2d, count %2d %s\n",
100 map->index, map->count,
101 map->index == text_index && text ? text : "");
102 #endif
103 }
104
gntdev_free_map(struct gntdev_grant_map * map)105 static void gntdev_free_map(struct gntdev_grant_map *map)
106 {
107 if (map == NULL)
108 return;
109
110 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
111 if (map->dma_vaddr) {
112 struct gnttab_dma_alloc_args args;
113
114 args.dev = map->dma_dev;
115 args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
116 args.nr_pages = map->count;
117 args.pages = map->pages;
118 args.frames = map->frames;
119 args.vaddr = map->dma_vaddr;
120 args.dev_bus_addr = map->dma_bus_addr;
121
122 gnttab_dma_free_pages(&args);
123 } else
124 #endif
125 if (map->pages)
126 gnttab_free_pages(map->count, map->pages);
127
128 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
129 kvfree(map->frames);
130 #endif
131 kvfree(map->pages);
132 kvfree(map->grants);
133 kvfree(map->map_ops);
134 kvfree(map->unmap_ops);
135 kvfree(map->kmap_ops);
136 kvfree(map->kunmap_ops);
137 kvfree(map->being_removed);
138 kfree(map);
139 }
140
gntdev_alloc_map(struct gntdev_priv * priv,int count,int dma_flags)141 struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
142 int dma_flags)
143 {
144 struct gntdev_grant_map *add;
145 int i;
146
147 add = kzalloc(sizeof(*add), GFP_KERNEL);
148 if (NULL == add)
149 return NULL;
150
151 add->grants = kvmalloc_array(count, sizeof(add->grants[0]),
152 GFP_KERNEL);
153 add->map_ops = kvmalloc_array(count, sizeof(add->map_ops[0]),
154 GFP_KERNEL);
155 add->unmap_ops = kvmalloc_array(count, sizeof(add->unmap_ops[0]),
156 GFP_KERNEL);
157 add->pages = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
158 add->being_removed =
159 kvcalloc(count, sizeof(add->being_removed[0]), GFP_KERNEL);
160 if (NULL == add->grants ||
161 NULL == add->map_ops ||
162 NULL == add->unmap_ops ||
163 NULL == add->pages ||
164 NULL == add->being_removed)
165 goto err;
166 if (use_ptemod) {
167 add->kmap_ops = kvmalloc_array(count, sizeof(add->kmap_ops[0]),
168 GFP_KERNEL);
169 add->kunmap_ops = kvmalloc_array(count, sizeof(add->kunmap_ops[0]),
170 GFP_KERNEL);
171 if (NULL == add->kmap_ops || NULL == add->kunmap_ops)
172 goto err;
173 }
174
175 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
176 add->dma_flags = dma_flags;
177
178 /*
179 * Check if this mapping is requested to be backed
180 * by a DMA buffer.
181 */
182 if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
183 struct gnttab_dma_alloc_args args;
184
185 add->frames = kvcalloc(count, sizeof(add->frames[0]),
186 GFP_KERNEL);
187 if (!add->frames)
188 goto err;
189
190 /* Remember the device, so we can free DMA memory. */
191 add->dma_dev = priv->dma_dev;
192
193 args.dev = priv->dma_dev;
194 args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
195 args.nr_pages = count;
196 args.pages = add->pages;
197 args.frames = add->frames;
198
199 if (gnttab_dma_alloc_pages(&args))
200 goto err;
201
202 add->dma_vaddr = args.vaddr;
203 add->dma_bus_addr = args.dev_bus_addr;
204 } else
205 #endif
206 if (gnttab_alloc_pages(count, add->pages))
207 goto err;
208
209 for (i = 0; i < count; i++) {
210 add->grants[i].domid = DOMID_INVALID;
211 add->grants[i].ref = INVALID_GRANT_REF;
212 add->map_ops[i].handle = INVALID_GRANT_HANDLE;
213 add->unmap_ops[i].handle = INVALID_GRANT_HANDLE;
214 if (use_ptemod) {
215 add->kmap_ops[i].handle = INVALID_GRANT_HANDLE;
216 add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE;
217 }
218 }
219
220 add->index = 0;
221 add->count = count;
222 refcount_set(&add->users, 1);
223
224 return add;
225
226 err:
227 gntdev_free_map(add);
228 return NULL;
229 }
230
gntdev_add_map(struct gntdev_priv * priv,struct gntdev_grant_map * add)231 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
232 {
233 struct gntdev_grant_map *map;
234
235 list_for_each_entry(map, &priv->maps, next) {
236 if (add->index + add->count < map->index) {
237 list_add_tail(&add->next, &map->next);
238 goto done;
239 }
240 add->index = map->index + map->count;
241 }
242 list_add_tail(&add->next, &priv->maps);
243
244 done:
245 gntdev_print_maps(priv, "[new]", add->index);
246 }
247
gntdev_find_map_index(struct gntdev_priv * priv,int index,int count)248 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
249 int index, int count)
250 {
251 struct gntdev_grant_map *map;
252
253 list_for_each_entry(map, &priv->maps, next) {
254 if (map->index != index)
255 continue;
256 if (count && map->count != count)
257 continue;
258 return map;
259 }
260 return NULL;
261 }
262
gntdev_put_map(struct gntdev_priv * priv,struct gntdev_grant_map * map)263 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
264 {
265 if (!map)
266 return;
267
268 if (!refcount_dec_and_test(&map->users))
269 return;
270
271 if (map->pages && !use_ptemod) {
272 /*
273 * Increment the reference count. This ensures that the
274 * subsequent call to unmap_grant_pages() will not wind up
275 * re-entering itself. It *can* wind up calling
276 * gntdev_put_map() recursively, but such calls will be with a
277 * reference count greater than 1, so they will return before
278 * this code is reached. The recursion depth is thus limited to
279 * 1. Do NOT use refcount_inc() here, as it will detect that
280 * the reference count is zero and WARN().
281 */
282 refcount_set(&map->users, 1);
283
284 /*
285 * Unmap the grants. This may or may not be asynchronous, so it
286 * is possible that the reference count is 1 on return, but it
287 * could also be greater than 1.
288 */
289 unmap_grant_pages(map, 0, map->count);
290
291 /* Check if the memory now needs to be freed */
292 if (!refcount_dec_and_test(&map->users))
293 return;
294
295 /*
296 * All pages have been returned to the hypervisor, so free the
297 * map.
298 */
299 }
300
301 if (use_ptemod && map->notifier_init)
302 mmu_interval_notifier_remove(&map->notifier);
303
304 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
305 notify_remote_via_evtchn(map->notify.event);
306 evtchn_put(map->notify.event);
307 }
308 gntdev_free_map(map);
309 }
310
311 /* ------------------------------------------------------------------ */
312
find_grant_ptes(pte_t * pte,unsigned long addr,void * data)313 static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
314 {
315 struct gntdev_grant_map *map = data;
316 unsigned int pgnr = (addr - map->pages_vm_start) >> PAGE_SHIFT;
317 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte |
318 (1 << _GNTMAP_guest_avail0);
319 u64 pte_maddr;
320
321 BUG_ON(pgnr >= map->count);
322 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
323
324 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
325 map->grants[pgnr].ref,
326 map->grants[pgnr].domid);
327 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
328 INVALID_GRANT_HANDLE);
329 return 0;
330 }
331
gntdev_map_grant_pages(struct gntdev_grant_map * map)332 int gntdev_map_grant_pages(struct gntdev_grant_map *map)
333 {
334 size_t alloced = 0;
335 int i, err = 0;
336
337 if (!use_ptemod) {
338 /* Note: it could already be mapped */
339 if (map->map_ops[0].handle != INVALID_GRANT_HANDLE)
340 return 0;
341 for (i = 0; i < map->count; i++) {
342 unsigned long addr = (unsigned long)
343 pfn_to_kaddr(page_to_pfn(map->pages[i]));
344 gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
345 map->grants[i].ref,
346 map->grants[i].domid);
347 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
348 map->flags, INVALID_GRANT_HANDLE);
349 }
350 } else {
351 /*
352 * Setup the map_ops corresponding to the pte entries pointing
353 * to the kernel linear addresses of the struct pages.
354 * These ptes are completely different from the user ptes dealt
355 * with find_grant_ptes.
356 * Note that GNTMAP_device_map isn't needed here: The
357 * dev_bus_addr output field gets consumed only from ->map_ops,
358 * and by not requesting it when mapping we also avoid needing
359 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra
360 * reference to the page in the hypervisor).
361 */
362 unsigned int flags = (map->flags & ~GNTMAP_device_map) |
363 GNTMAP_host_map;
364
365 for (i = 0; i < map->count; i++) {
366 unsigned long address = (unsigned long)
367 pfn_to_kaddr(page_to_pfn(map->pages[i]));
368 BUG_ON(PageHighMem(map->pages[i]));
369
370 gnttab_set_map_op(&map->kmap_ops[i], address, flags,
371 map->grants[i].ref,
372 map->grants[i].domid);
373 gnttab_set_unmap_op(&map->kunmap_ops[i], address,
374 flags, INVALID_GRANT_HANDLE);
375 }
376 }
377
378 pr_debug("map %d+%d\n", map->index, map->count);
379 err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages,
380 map->count);
381
382 for (i = 0; i < map->count; i++) {
383 if (map->map_ops[i].status == GNTST_okay) {
384 map->unmap_ops[i].handle = map->map_ops[i].handle;
385 alloced++;
386 } else if (!err)
387 err = -EINVAL;
388
389 if (map->flags & GNTMAP_device_map)
390 map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr;
391
392 if (use_ptemod) {
393 if (map->kmap_ops[i].status == GNTST_okay) {
394 alloced++;
395 map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
396 } else if (!err)
397 err = -EINVAL;
398 }
399 }
400 atomic_add(alloced, &map->live_grants);
401 return err;
402 }
403
__unmap_grant_pages_done(int result,struct gntab_unmap_queue_data * data)404 static void __unmap_grant_pages_done(int result,
405 struct gntab_unmap_queue_data *data)
406 {
407 unsigned int i;
408 struct gntdev_grant_map *map = data->data;
409 unsigned int offset = data->unmap_ops - map->unmap_ops;
410 int successful_unmaps = 0;
411 int live_grants;
412
413 for (i = 0; i < data->count; i++) {
414 if (map->unmap_ops[offset + i].status == GNTST_okay &&
415 map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
416 successful_unmaps++;
417
418 WARN_ON(map->unmap_ops[offset + i].status != GNTST_okay &&
419 map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
420 pr_debug("unmap handle=%d st=%d\n",
421 map->unmap_ops[offset+i].handle,
422 map->unmap_ops[offset+i].status);
423 map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
424 if (use_ptemod) {
425 if (map->kunmap_ops[offset + i].status == GNTST_okay &&
426 map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
427 successful_unmaps++;
428
429 WARN_ON(map->kunmap_ops[offset + i].status != GNTST_okay &&
430 map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
431 pr_debug("kunmap handle=%u st=%d\n",
432 map->kunmap_ops[offset+i].handle,
433 map->kunmap_ops[offset+i].status);
434 map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
435 }
436 }
437
438 /*
439 * Decrease the live-grant counter. This must happen after the loop to
440 * prevent premature reuse of the grants by gnttab_mmap().
441 */
442 live_grants = atomic_sub_return(successful_unmaps, &map->live_grants);
443 if (WARN_ON(live_grants < 0))
444 pr_err("%s: live_grants became negative (%d) after unmapping %d pages!\n",
445 __func__, live_grants, successful_unmaps);
446
447 /* Release reference taken by __unmap_grant_pages */
448 gntdev_put_map(NULL, map);
449 }
450
__unmap_grant_pages(struct gntdev_grant_map * map,int offset,int pages)451 static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
452 int pages)
453 {
454 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
455 int pgno = (map->notify.addr >> PAGE_SHIFT);
456
457 if (pgno >= offset && pgno < offset + pages) {
458 /* No need for kmap, pages are in lowmem */
459 uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
460
461 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
462 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
463 }
464 }
465
466 map->unmap_data.unmap_ops = map->unmap_ops + offset;
467 map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
468 map->unmap_data.pages = map->pages + offset;
469 map->unmap_data.count = pages;
470 map->unmap_data.done = __unmap_grant_pages_done;
471 map->unmap_data.data = map;
472 refcount_inc(&map->users); /* to keep map alive during async call below */
473
474 gnttab_unmap_refs_async(&map->unmap_data);
475 }
476
unmap_grant_pages(struct gntdev_grant_map * map,int offset,int pages)477 static void unmap_grant_pages(struct gntdev_grant_map *map, int offset,
478 int pages)
479 {
480 int range;
481
482 if (atomic_read(&map->live_grants) == 0)
483 return; /* Nothing to do */
484
485 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
486
487 /* It is possible the requested range will have a "hole" where we
488 * already unmapped some of the grants. Only unmap valid ranges.
489 */
490 while (pages) {
491 while (pages && map->being_removed[offset]) {
492 offset++;
493 pages--;
494 }
495 range = 0;
496 while (range < pages) {
497 if (map->being_removed[offset + range])
498 break;
499 map->being_removed[offset + range] = true;
500 range++;
501 }
502 if (range)
503 __unmap_grant_pages(map, offset, range);
504 offset += range;
505 pages -= range;
506 }
507 }
508
509 /* ------------------------------------------------------------------ */
510
gntdev_vma_open(struct vm_area_struct * vma)511 static void gntdev_vma_open(struct vm_area_struct *vma)
512 {
513 struct gntdev_grant_map *map = vma->vm_private_data;
514
515 pr_debug("gntdev_vma_open %p\n", vma);
516 refcount_inc(&map->users);
517 }
518
gntdev_vma_close(struct vm_area_struct * vma)519 static void gntdev_vma_close(struct vm_area_struct *vma)
520 {
521 struct gntdev_grant_map *map = vma->vm_private_data;
522 struct file *file = vma->vm_file;
523 struct gntdev_priv *priv = file->private_data;
524
525 pr_debug("gntdev_vma_close %p\n", vma);
526
527 vma->vm_private_data = NULL;
528 gntdev_put_map(priv, map);
529 }
530
gntdev_vma_find_special_page(struct vm_area_struct * vma,unsigned long addr)531 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
532 unsigned long addr)
533 {
534 struct gntdev_grant_map *map = vma->vm_private_data;
535
536 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
537 }
538
539 static const struct vm_operations_struct gntdev_vmops = {
540 .open = gntdev_vma_open,
541 .close = gntdev_vma_close,
542 .find_special_page = gntdev_vma_find_special_page,
543 };
544
545 /* ------------------------------------------------------------------ */
546
gntdev_invalidate(struct mmu_interval_notifier * mn,const struct mmu_notifier_range * range,unsigned long cur_seq)547 static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
548 const struct mmu_notifier_range *range,
549 unsigned long cur_seq)
550 {
551 struct gntdev_grant_map *map =
552 container_of(mn, struct gntdev_grant_map, notifier);
553 unsigned long mstart, mend;
554 unsigned long map_start, map_end;
555
556 if (!mmu_notifier_range_blockable(range))
557 return false;
558
559 map_start = map->pages_vm_start;
560 map_end = map->pages_vm_start + (map->count << PAGE_SHIFT);
561
562 /*
563 * If the VMA is split or otherwise changed the notifier is not
564 * updated, but we don't want to process VA's outside the modified
565 * VMA. FIXME: It would be much more understandable to just prevent
566 * modifying the VMA in the first place.
567 */
568 if (map_start >= range->end || map_end <= range->start)
569 return true;
570
571 mstart = max(range->start, map_start);
572 mend = min(range->end, map_end);
573 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
574 map->index, map->count, map_start, map_end,
575 range->start, range->end, mstart, mend);
576 unmap_grant_pages(map, (mstart - map_start) >> PAGE_SHIFT,
577 (mend - mstart) >> PAGE_SHIFT);
578
579 return true;
580 }
581
582 static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
583 .invalidate = gntdev_invalidate,
584 };
585
586 /* ------------------------------------------------------------------ */
587
gntdev_open(struct inode * inode,struct file * flip)588 static int gntdev_open(struct inode *inode, struct file *flip)
589 {
590 struct gntdev_priv *priv;
591
592 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
593 if (!priv)
594 return -ENOMEM;
595
596 INIT_LIST_HEAD(&priv->maps);
597 mutex_init(&priv->lock);
598
599 mutex_init(&priv->batch_lock);
600
601 #ifdef CONFIG_XEN_GNTDEV_DMABUF
602 priv->dmabuf_priv = gntdev_dmabuf_init(flip);
603 if (IS_ERR(priv->dmabuf_priv)) {
604 int ret = PTR_ERR(priv->dmabuf_priv);
605
606 kfree(priv);
607 return ret;
608 }
609 #endif
610
611 flip->private_data = priv;
612 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
613 priv->dma_dev = gntdev_miscdev.this_device;
614 dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
615 #endif
616 pr_debug("priv %p\n", priv);
617
618 return 0;
619 }
620
gntdev_release(struct inode * inode,struct file * flip)621 static int gntdev_release(struct inode *inode, struct file *flip)
622 {
623 struct gntdev_priv *priv = flip->private_data;
624 struct gntdev_grant_map *map;
625 struct gntdev_copy_batch *batch;
626
627 pr_debug("priv %p\n", priv);
628
629 mutex_lock(&priv->lock);
630 while (!list_empty(&priv->maps)) {
631 map = list_entry(priv->maps.next,
632 struct gntdev_grant_map, next);
633 list_del(&map->next);
634 gntdev_put_map(NULL /* already removed */, map);
635 }
636 mutex_unlock(&priv->lock);
637
638 mutex_lock(&priv->batch_lock);
639 while (priv->batch) {
640 batch = priv->batch;
641 priv->batch = batch->next;
642 kfree(batch);
643 }
644 mutex_unlock(&priv->batch_lock);
645
646 #ifdef CONFIG_XEN_GNTDEV_DMABUF
647 gntdev_dmabuf_fini(priv->dmabuf_priv);
648 #endif
649
650 kfree(priv);
651 return 0;
652 }
653
gntdev_ioctl_map_grant_ref(struct gntdev_priv * priv,struct ioctl_gntdev_map_grant_ref __user * u)654 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
655 struct ioctl_gntdev_map_grant_ref __user *u)
656 {
657 struct ioctl_gntdev_map_grant_ref op;
658 struct gntdev_grant_map *map;
659 int err;
660
661 if (copy_from_user(&op, u, sizeof(op)) != 0)
662 return -EFAULT;
663 pr_debug("priv %p, add %d\n", priv, op.count);
664 if (unlikely(gntdev_test_page_count(op.count)))
665 return -EINVAL;
666
667 err = -ENOMEM;
668 map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
669 if (!map)
670 return err;
671
672 if (copy_from_user(map->grants, &u->refs,
673 sizeof(map->grants[0]) * op.count) != 0) {
674 gntdev_put_map(NULL, map);
675 return -EFAULT;
676 }
677
678 mutex_lock(&priv->lock);
679 gntdev_add_map(priv, map);
680 op.index = map->index << PAGE_SHIFT;
681 mutex_unlock(&priv->lock);
682
683 if (copy_to_user(u, &op, sizeof(op)) != 0)
684 return -EFAULT;
685
686 return 0;
687 }
688
gntdev_ioctl_unmap_grant_ref(struct gntdev_priv * priv,struct ioctl_gntdev_unmap_grant_ref __user * u)689 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
690 struct ioctl_gntdev_unmap_grant_ref __user *u)
691 {
692 struct ioctl_gntdev_unmap_grant_ref op;
693 struct gntdev_grant_map *map;
694 int err = -ENOENT;
695
696 if (copy_from_user(&op, u, sizeof(op)) != 0)
697 return -EFAULT;
698 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
699
700 mutex_lock(&priv->lock);
701 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
702 if (map) {
703 list_del(&map->next);
704 err = 0;
705 }
706 mutex_unlock(&priv->lock);
707 if (map)
708 gntdev_put_map(priv, map);
709 return err;
710 }
711
gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv * priv,struct ioctl_gntdev_get_offset_for_vaddr __user * u)712 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
713 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
714 {
715 struct ioctl_gntdev_get_offset_for_vaddr op;
716 struct vm_area_struct *vma;
717 struct gntdev_grant_map *map;
718 int rv = -EINVAL;
719
720 if (copy_from_user(&op, u, sizeof(op)) != 0)
721 return -EFAULT;
722 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
723
724 mmap_read_lock(current->mm);
725 vma = find_vma(current->mm, op.vaddr);
726 if (!vma || vma->vm_ops != &gntdev_vmops)
727 goto out_unlock;
728
729 map = vma->vm_private_data;
730 if (!map)
731 goto out_unlock;
732
733 op.offset = map->index << PAGE_SHIFT;
734 op.count = map->count;
735 rv = 0;
736
737 out_unlock:
738 mmap_read_unlock(current->mm);
739
740 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
741 return -EFAULT;
742 return rv;
743 }
744
gntdev_ioctl_notify(struct gntdev_priv * priv,void __user * u)745 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
746 {
747 struct ioctl_gntdev_unmap_notify op;
748 struct gntdev_grant_map *map;
749 int rc;
750 int out_flags;
751 evtchn_port_t out_event;
752
753 if (copy_from_user(&op, u, sizeof(op)))
754 return -EFAULT;
755
756 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
757 return -EINVAL;
758
759 /* We need to grab a reference to the event channel we are going to use
760 * to send the notify before releasing the reference we may already have
761 * (if someone has called this ioctl twice). This is required so that
762 * it is possible to change the clear_byte part of the notification
763 * without disturbing the event channel part, which may now be the last
764 * reference to that event channel.
765 */
766 if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
767 if (evtchn_get(op.event_channel_port))
768 return -EINVAL;
769 }
770
771 out_flags = op.action;
772 out_event = op.event_channel_port;
773
774 mutex_lock(&priv->lock);
775
776 list_for_each_entry(map, &priv->maps, next) {
777 uint64_t begin = map->index << PAGE_SHIFT;
778 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
779 if (op.index >= begin && op.index < end)
780 goto found;
781 }
782 rc = -ENOENT;
783 goto unlock_out;
784
785 found:
786 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
787 (map->flags & GNTMAP_readonly)) {
788 rc = -EINVAL;
789 goto unlock_out;
790 }
791
792 out_flags = map->notify.flags;
793 out_event = map->notify.event;
794
795 map->notify.flags = op.action;
796 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
797 map->notify.event = op.event_channel_port;
798
799 rc = 0;
800
801 unlock_out:
802 mutex_unlock(&priv->lock);
803
804 /* Drop the reference to the event channel we did not save in the map */
805 if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
806 evtchn_put(out_event);
807
808 return rc;
809 }
810
gntdev_get_page(struct gntdev_copy_batch * batch,void __user * virt,unsigned long * gfn)811 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
812 unsigned long *gfn)
813 {
814 unsigned long addr = (unsigned long)virt;
815 struct page *page;
816 unsigned long xen_pfn;
817 int ret;
818
819 ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
820 if (ret < 0)
821 return ret;
822
823 batch->pages[batch->nr_pages++] = page;
824
825 xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
826 *gfn = pfn_to_gfn(xen_pfn);
827
828 return 0;
829 }
830
gntdev_put_pages(struct gntdev_copy_batch * batch)831 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
832 {
833 unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
834 batch->nr_pages = 0;
835 batch->writeable = false;
836 }
837
gntdev_copy(struct gntdev_copy_batch * batch)838 static int gntdev_copy(struct gntdev_copy_batch *batch)
839 {
840 unsigned int i;
841
842 gnttab_batch_copy(batch->ops, batch->nr_ops);
843 gntdev_put_pages(batch);
844
845 /*
846 * For each completed op, update the status if the op failed
847 * and all previous ops for the segment were successful.
848 */
849 for (i = 0; i < batch->nr_ops; i++) {
850 s16 status = batch->ops[i].status;
851 s16 old_status;
852
853 if (status == GNTST_okay)
854 continue;
855
856 if (__get_user(old_status, batch->status[i]))
857 return -EFAULT;
858
859 if (old_status != GNTST_okay)
860 continue;
861
862 if (__put_user(status, batch->status[i]))
863 return -EFAULT;
864 }
865
866 batch->nr_ops = 0;
867 return 0;
868 }
869
gntdev_grant_copy_seg(struct gntdev_copy_batch * batch,struct gntdev_grant_copy_segment * seg,s16 __user * status)870 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
871 struct gntdev_grant_copy_segment *seg,
872 s16 __user *status)
873 {
874 uint16_t copied = 0;
875
876 /*
877 * Disallow local -> local copies since there is only space in
878 * batch->pages for one page per-op and this would be a very
879 * expensive memcpy().
880 */
881 if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
882 return -EINVAL;
883
884 /* Can't cross page if source/dest is a grant ref. */
885 if (seg->flags & GNTCOPY_source_gref) {
886 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
887 return -EINVAL;
888 }
889 if (seg->flags & GNTCOPY_dest_gref) {
890 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
891 return -EINVAL;
892 }
893
894 if (put_user(GNTST_okay, status))
895 return -EFAULT;
896
897 while (copied < seg->len) {
898 struct gnttab_copy *op;
899 void __user *virt;
900 size_t len, off;
901 unsigned long gfn;
902 int ret;
903
904 if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
905 ret = gntdev_copy(batch);
906 if (ret < 0)
907 return ret;
908 }
909
910 len = seg->len - copied;
911
912 op = &batch->ops[batch->nr_ops];
913 op->flags = 0;
914
915 if (seg->flags & GNTCOPY_source_gref) {
916 op->source.u.ref = seg->source.foreign.ref;
917 op->source.domid = seg->source.foreign.domid;
918 op->source.offset = seg->source.foreign.offset + copied;
919 op->flags |= GNTCOPY_source_gref;
920 } else {
921 virt = seg->source.virt + copied;
922 off = (unsigned long)virt & ~XEN_PAGE_MASK;
923 len = min(len, (size_t)XEN_PAGE_SIZE - off);
924 batch->writeable = false;
925
926 ret = gntdev_get_page(batch, virt, &gfn);
927 if (ret < 0)
928 return ret;
929
930 op->source.u.gmfn = gfn;
931 op->source.domid = DOMID_SELF;
932 op->source.offset = off;
933 }
934
935 if (seg->flags & GNTCOPY_dest_gref) {
936 op->dest.u.ref = seg->dest.foreign.ref;
937 op->dest.domid = seg->dest.foreign.domid;
938 op->dest.offset = seg->dest.foreign.offset + copied;
939 op->flags |= GNTCOPY_dest_gref;
940 } else {
941 virt = seg->dest.virt + copied;
942 off = (unsigned long)virt & ~XEN_PAGE_MASK;
943 len = min(len, (size_t)XEN_PAGE_SIZE - off);
944 batch->writeable = true;
945
946 ret = gntdev_get_page(batch, virt, &gfn);
947 if (ret < 0)
948 return ret;
949
950 op->dest.u.gmfn = gfn;
951 op->dest.domid = DOMID_SELF;
952 op->dest.offset = off;
953 }
954
955 op->len = len;
956 copied += len;
957
958 batch->status[batch->nr_ops] = status;
959 batch->nr_ops++;
960 }
961
962 return 0;
963 }
964
gntdev_ioctl_grant_copy(struct gntdev_priv * priv,void __user * u)965 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
966 {
967 struct ioctl_gntdev_grant_copy copy;
968 struct gntdev_copy_batch *batch;
969 unsigned int i;
970 int ret = 0;
971
972 if (copy_from_user(©, u, sizeof(copy)))
973 return -EFAULT;
974
975 mutex_lock(&priv->batch_lock);
976 if (!priv->batch) {
977 batch = kmalloc(sizeof(*batch), GFP_KERNEL);
978 } else {
979 batch = priv->batch;
980 priv->batch = batch->next;
981 }
982 mutex_unlock(&priv->batch_lock);
983 if (!batch)
984 return -ENOMEM;
985
986 batch->nr_ops = 0;
987 batch->nr_pages = 0;
988
989 for (i = 0; i < copy.count; i++) {
990 struct gntdev_grant_copy_segment seg;
991
992 if (copy_from_user(&seg, ©.segments[i], sizeof(seg))) {
993 ret = -EFAULT;
994 gntdev_put_pages(batch);
995 goto out;
996 }
997
998 ret = gntdev_grant_copy_seg(batch, &seg, ©.segments[i].status);
999 if (ret < 0) {
1000 gntdev_put_pages(batch);
1001 goto out;
1002 }
1003
1004 cond_resched();
1005 }
1006 if (batch->nr_ops)
1007 ret = gntdev_copy(batch);
1008
1009 out:
1010 mutex_lock(&priv->batch_lock);
1011 batch->next = priv->batch;
1012 priv->batch = batch;
1013 mutex_unlock(&priv->batch_lock);
1014
1015 return ret;
1016 }
1017
gntdev_ioctl(struct file * flip,unsigned int cmd,unsigned long arg)1018 static long gntdev_ioctl(struct file *flip,
1019 unsigned int cmd, unsigned long arg)
1020 {
1021 struct gntdev_priv *priv = flip->private_data;
1022 void __user *ptr = (void __user *)arg;
1023
1024 switch (cmd) {
1025 case IOCTL_GNTDEV_MAP_GRANT_REF:
1026 return gntdev_ioctl_map_grant_ref(priv, ptr);
1027
1028 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
1029 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
1030
1031 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
1032 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
1033
1034 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
1035 return gntdev_ioctl_notify(priv, ptr);
1036
1037 case IOCTL_GNTDEV_GRANT_COPY:
1038 return gntdev_ioctl_grant_copy(priv, ptr);
1039
1040 #ifdef CONFIG_XEN_GNTDEV_DMABUF
1041 case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
1042 return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
1043
1044 case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
1045 return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
1046
1047 case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
1048 return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
1049
1050 case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
1051 return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
1052 #endif
1053
1054 default:
1055 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
1056 return -ENOIOCTLCMD;
1057 }
1058
1059 return 0;
1060 }
1061
gntdev_mmap(struct file * flip,struct vm_area_struct * vma)1062 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
1063 {
1064 struct gntdev_priv *priv = flip->private_data;
1065 int index = vma->vm_pgoff;
1066 int count = vma_pages(vma);
1067 struct gntdev_grant_map *map;
1068 int err = -EINVAL;
1069
1070 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
1071 return -EINVAL;
1072
1073 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
1074 index, count, vma->vm_start, vma->vm_pgoff);
1075
1076 mutex_lock(&priv->lock);
1077 map = gntdev_find_map_index(priv, index, count);
1078 if (!map)
1079 goto unlock_out;
1080 if (!atomic_add_unless(&map->in_use, 1, 1))
1081 goto unlock_out;
1082
1083 refcount_inc(&map->users);
1084
1085 vma->vm_ops = &gntdev_vmops;
1086
1087 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP);
1088
1089 if (use_ptemod)
1090 vm_flags_set(vma, VM_DONTCOPY);
1091
1092 vma->vm_private_data = map;
1093 if (map->flags) {
1094 if ((vma->vm_flags & VM_WRITE) &&
1095 (map->flags & GNTMAP_readonly))
1096 goto out_unlock_put;
1097 } else {
1098 map->flags = GNTMAP_host_map;
1099 if (!(vma->vm_flags & VM_WRITE))
1100 map->flags |= GNTMAP_readonly;
1101 }
1102
1103 map->pages_vm_start = vma->vm_start;
1104
1105 if (use_ptemod) {
1106 err = mmu_interval_notifier_insert_locked(
1107 &map->notifier, vma->vm_mm, vma->vm_start,
1108 vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1109 if (err)
1110 goto out_unlock_put;
1111
1112 map->notifier_init = true;
1113 }
1114 mutex_unlock(&priv->lock);
1115
1116 if (use_ptemod) {
1117 /*
1118 * gntdev takes the address of the PTE in find_grant_ptes() and
1119 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1120 * purpose of the notifier is to prevent the hypervisor pointer
1121 * to the PTE from going stale.
1122 *
1123 * Since this vma's mappings can't be touched without the
1124 * mmap_lock, and we are holding it now, there is no need for
1125 * the notifier_range locking pattern.
1126 */
1127 mmu_interval_read_begin(&map->notifier);
1128
1129 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1130 vma->vm_end - vma->vm_start,
1131 find_grant_ptes, map);
1132 if (err) {
1133 pr_warn("find_grant_ptes() failure.\n");
1134 goto out_put_map;
1135 }
1136 }
1137
1138 err = gntdev_map_grant_pages(map);
1139 if (err)
1140 goto out_put_map;
1141
1142 if (!use_ptemod) {
1143 err = vm_map_pages_zero(vma, map->pages, map->count);
1144 if (err)
1145 goto out_put_map;
1146 }
1147
1148 return 0;
1149
1150 unlock_out:
1151 mutex_unlock(&priv->lock);
1152 return err;
1153
1154 out_unlock_put:
1155 mutex_unlock(&priv->lock);
1156 out_put_map:
1157 if (use_ptemod)
1158 unmap_grant_pages(map, 0, map->count);
1159 gntdev_put_map(priv, map);
1160 return err;
1161 }
1162
1163 static const struct file_operations gntdev_fops = {
1164 .owner = THIS_MODULE,
1165 .open = gntdev_open,
1166 .release = gntdev_release,
1167 .mmap = gntdev_mmap,
1168 .unlocked_ioctl = gntdev_ioctl
1169 };
1170
1171 static struct miscdevice gntdev_miscdev = {
1172 .minor = MISC_DYNAMIC_MINOR,
1173 .name = "xen/gntdev",
1174 .fops = &gntdev_fops,
1175 };
1176
1177 /* ------------------------------------------------------------------ */
1178
gntdev_init(void)1179 static int __init gntdev_init(void)
1180 {
1181 int err;
1182
1183 if (!xen_domain())
1184 return -ENODEV;
1185
1186 use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1187
1188 err = misc_register(&gntdev_miscdev);
1189 if (err != 0) {
1190 pr_err("Could not register gntdev device\n");
1191 return err;
1192 }
1193 return 0;
1194 }
1195
gntdev_exit(void)1196 static void __exit gntdev_exit(void)
1197 {
1198 misc_deregister(&gntdev_miscdev);
1199 }
1200
1201 module_init(gntdev_init);
1202 module_exit(gntdev_exit);
1203
1204 /* ------------------------------------------------------------------ */
1205