xref: /openbmc/linux/drivers/xen/gntdev.c (revision 2169e6da)
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/module.h>
26 #include <linux/kernel.h>
27 #include <linux/init.h>
28 #include <linux/miscdevice.h>
29 #include <linux/fs.h>
30 #include <linux/uaccess.h>
31 #include <linux/sched.h>
32 #include <linux/sched/mm.h>
33 #include <linux/spinlock.h>
34 #include <linux/slab.h>
35 #include <linux/highmem.h>
36 #include <linux/refcount.h>
37 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
38 #include <linux/of_device.h>
39 #endif
40 
41 #include <xen/xen.h>
42 #include <xen/grant_table.h>
43 #include <xen/balloon.h>
44 #include <xen/gntdev.h>
45 #include <xen/events.h>
46 #include <xen/page.h>
47 #include <asm/xen/hypervisor.h>
48 #include <asm/xen/hypercall.h>
49 
50 #include "gntdev-common.h"
51 #ifdef CONFIG_XEN_GNTDEV_DMABUF
52 #include "gntdev-dmabuf.h"
53 #endif
54 
55 MODULE_LICENSE("GPL");
56 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
57 	      "Gerd Hoffmann <kraxel@redhat.com>");
58 MODULE_DESCRIPTION("User-space granted page access driver");
59 
60 static int limit = 1024*1024;
61 module_param(limit, int, 0644);
62 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
63 		"the gntdev device");
64 
65 static atomic_t pages_mapped = ATOMIC_INIT(0);
66 
67 static int use_ptemod;
68 #define populate_freeable_maps use_ptemod
69 
70 static int unmap_grant_pages(struct gntdev_grant_map *map,
71 			     int offset, int pages);
72 
73 static struct miscdevice gntdev_miscdev;
74 
75 /* ------------------------------------------------------------------ */
76 
77 bool gntdev_account_mapped_pages(int count)
78 {
79 	return atomic_add_return(count, &pages_mapped) > limit;
80 }
81 
82 static void gntdev_print_maps(struct gntdev_priv *priv,
83 			      char *text, int text_index)
84 {
85 #ifdef DEBUG
86 	struct gntdev_grant_map *map;
87 
88 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
89 	list_for_each_entry(map, &priv->maps, next)
90 		pr_debug("  index %2d, count %2d %s\n",
91 		       map->index, map->count,
92 		       map->index == text_index && text ? text : "");
93 #endif
94 }
95 
96 static void gntdev_free_map(struct gntdev_grant_map *map)
97 {
98 	if (map == NULL)
99 		return;
100 
101 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
102 	if (map->dma_vaddr) {
103 		struct gnttab_dma_alloc_args args;
104 
105 		args.dev = map->dma_dev;
106 		args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
107 		args.nr_pages = map->count;
108 		args.pages = map->pages;
109 		args.frames = map->frames;
110 		args.vaddr = map->dma_vaddr;
111 		args.dev_bus_addr = map->dma_bus_addr;
112 
113 		gnttab_dma_free_pages(&args);
114 	} else
115 #endif
116 	if (map->pages)
117 		gnttab_free_pages(map->count, map->pages);
118 
119 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
120 	kfree(map->frames);
121 #endif
122 	kfree(map->pages);
123 	kfree(map->grants);
124 	kfree(map->map_ops);
125 	kfree(map->unmap_ops);
126 	kfree(map->kmap_ops);
127 	kfree(map->kunmap_ops);
128 	kfree(map);
129 }
130 
131 struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
132 					  int dma_flags)
133 {
134 	struct gntdev_grant_map *add;
135 	int i;
136 
137 	add = kzalloc(sizeof(*add), GFP_KERNEL);
138 	if (NULL == add)
139 		return NULL;
140 
141 	add->grants    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
142 	add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
143 	add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
144 	add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
145 	add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
146 	add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
147 	if (NULL == add->grants    ||
148 	    NULL == add->map_ops   ||
149 	    NULL == add->unmap_ops ||
150 	    NULL == add->kmap_ops  ||
151 	    NULL == add->kunmap_ops ||
152 	    NULL == add->pages)
153 		goto err;
154 
155 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
156 	add->dma_flags = dma_flags;
157 
158 	/*
159 	 * Check if this mapping is requested to be backed
160 	 * by a DMA buffer.
161 	 */
162 	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
163 		struct gnttab_dma_alloc_args args;
164 
165 		add->frames = kcalloc(count, sizeof(add->frames[0]),
166 				      GFP_KERNEL);
167 		if (!add->frames)
168 			goto err;
169 
170 		/* Remember the device, so we can free DMA memory. */
171 		add->dma_dev = priv->dma_dev;
172 
173 		args.dev = priv->dma_dev;
174 		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
175 		args.nr_pages = count;
176 		args.pages = add->pages;
177 		args.frames = add->frames;
178 
179 		if (gnttab_dma_alloc_pages(&args))
180 			goto err;
181 
182 		add->dma_vaddr = args.vaddr;
183 		add->dma_bus_addr = args.dev_bus_addr;
184 	} else
185 #endif
186 	if (gnttab_alloc_pages(count, add->pages))
187 		goto err;
188 
189 	for (i = 0; i < count; i++) {
190 		add->map_ops[i].handle = -1;
191 		add->unmap_ops[i].handle = -1;
192 		add->kmap_ops[i].handle = -1;
193 		add->kunmap_ops[i].handle = -1;
194 	}
195 
196 	add->index = 0;
197 	add->count = count;
198 	refcount_set(&add->users, 1);
199 
200 	return add;
201 
202 err:
203 	gntdev_free_map(add);
204 	return NULL;
205 }
206 
207 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
208 {
209 	struct gntdev_grant_map *map;
210 
211 	list_for_each_entry(map, &priv->maps, next) {
212 		if (add->index + add->count < map->index) {
213 			list_add_tail(&add->next, &map->next);
214 			goto done;
215 		}
216 		add->index = map->index + map->count;
217 	}
218 	list_add_tail(&add->next, &priv->maps);
219 
220 done:
221 	gntdev_print_maps(priv, "[new]", add->index);
222 }
223 
224 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
225 						      int index, int count)
226 {
227 	struct gntdev_grant_map *map;
228 
229 	list_for_each_entry(map, &priv->maps, next) {
230 		if (map->index != index)
231 			continue;
232 		if (count && map->count != count)
233 			continue;
234 		return map;
235 	}
236 	return NULL;
237 }
238 
239 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
240 {
241 	if (!map)
242 		return;
243 
244 	if (!refcount_dec_and_test(&map->users))
245 		return;
246 
247 	atomic_sub(map->count, &pages_mapped);
248 
249 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
250 		notify_remote_via_evtchn(map->notify.event);
251 		evtchn_put(map->notify.event);
252 	}
253 
254 	if (populate_freeable_maps && priv) {
255 		mutex_lock(&priv->lock);
256 		list_del(&map->next);
257 		mutex_unlock(&priv->lock);
258 	}
259 
260 	if (map->pages && !use_ptemod)
261 		unmap_grant_pages(map, 0, map->count);
262 	gntdev_free_map(map);
263 }
264 
265 /* ------------------------------------------------------------------ */
266 
267 static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
268 {
269 	struct gntdev_grant_map *map = data;
270 	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
271 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
272 	u64 pte_maddr;
273 
274 	BUG_ON(pgnr >= map->count);
275 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
276 
277 	/*
278 	 * Set the PTE as special to force get_user_pages_fast() fall
279 	 * back to the slow path.  If this is not supported as part of
280 	 * the grant map, it will be done afterwards.
281 	 */
282 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
283 		flags |= (1 << _GNTMAP_guest_avail0);
284 
285 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
286 			  map->grants[pgnr].ref,
287 			  map->grants[pgnr].domid);
288 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
289 			    -1 /* handle */);
290 	return 0;
291 }
292 
293 #ifdef CONFIG_X86
294 static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
295 {
296 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
297 	return 0;
298 }
299 #endif
300 
301 int gntdev_map_grant_pages(struct gntdev_grant_map *map)
302 {
303 	int i, err = 0;
304 
305 	if (!use_ptemod) {
306 		/* Note: it could already be mapped */
307 		if (map->map_ops[0].handle != -1)
308 			return 0;
309 		for (i = 0; i < map->count; i++) {
310 			unsigned long addr = (unsigned long)
311 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
312 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
313 				map->grants[i].ref,
314 				map->grants[i].domid);
315 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
316 				map->flags, -1 /* handle */);
317 		}
318 	} else {
319 		/*
320 		 * Setup the map_ops corresponding to the pte entries pointing
321 		 * to the kernel linear addresses of the struct pages.
322 		 * These ptes are completely different from the user ptes dealt
323 		 * with find_grant_ptes.
324 		 */
325 		for (i = 0; i < map->count; i++) {
326 			unsigned long address = (unsigned long)
327 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
328 			BUG_ON(PageHighMem(map->pages[i]));
329 
330 			gnttab_set_map_op(&map->kmap_ops[i], address,
331 				map->flags | GNTMAP_host_map,
332 				map->grants[i].ref,
333 				map->grants[i].domid);
334 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
335 				map->flags | GNTMAP_host_map, -1);
336 		}
337 	}
338 
339 	pr_debug("map %d+%d\n", map->index, map->count);
340 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
341 			map->pages, map->count);
342 	if (err)
343 		return err;
344 
345 	for (i = 0; i < map->count; i++) {
346 		if (map->map_ops[i].status) {
347 			err = -EINVAL;
348 			continue;
349 		}
350 
351 		map->unmap_ops[i].handle = map->map_ops[i].handle;
352 		if (use_ptemod)
353 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
354 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
355 		else if (map->dma_vaddr) {
356 			unsigned long bfn;
357 
358 			bfn = pfn_to_bfn(page_to_pfn(map->pages[i]));
359 			map->unmap_ops[i].dev_bus_addr = __pfn_to_phys(bfn);
360 		}
361 #endif
362 	}
363 	return err;
364 }
365 
366 static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
367 			       int pages)
368 {
369 	int i, err = 0;
370 	struct gntab_unmap_queue_data unmap_data;
371 
372 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
373 		int pgno = (map->notify.addr >> PAGE_SHIFT);
374 		if (pgno >= offset && pgno < offset + pages) {
375 			/* No need for kmap, pages are in lowmem */
376 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
377 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
378 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
379 		}
380 	}
381 
382 	unmap_data.unmap_ops = map->unmap_ops + offset;
383 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
384 	unmap_data.pages = map->pages + offset;
385 	unmap_data.count = pages;
386 
387 	err = gnttab_unmap_refs_sync(&unmap_data);
388 	if (err)
389 		return err;
390 
391 	for (i = 0; i < pages; i++) {
392 		if (map->unmap_ops[offset+i].status)
393 			err = -EINVAL;
394 		pr_debug("unmap handle=%d st=%d\n",
395 			map->unmap_ops[offset+i].handle,
396 			map->unmap_ops[offset+i].status);
397 		map->unmap_ops[offset+i].handle = -1;
398 	}
399 	return err;
400 }
401 
402 static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
403 			     int pages)
404 {
405 	int range, err = 0;
406 
407 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
408 
409 	/* It is possible the requested range will have a "hole" where we
410 	 * already unmapped some of the grants. Only unmap valid ranges.
411 	 */
412 	while (pages && !err) {
413 		while (pages && map->unmap_ops[offset].handle == -1) {
414 			offset++;
415 			pages--;
416 		}
417 		range = 0;
418 		while (range < pages) {
419 			if (map->unmap_ops[offset+range].handle == -1)
420 				break;
421 			range++;
422 		}
423 		err = __unmap_grant_pages(map, offset, range);
424 		offset += range;
425 		pages -= range;
426 	}
427 
428 	return err;
429 }
430 
431 /* ------------------------------------------------------------------ */
432 
433 static void gntdev_vma_open(struct vm_area_struct *vma)
434 {
435 	struct gntdev_grant_map *map = vma->vm_private_data;
436 
437 	pr_debug("gntdev_vma_open %p\n", vma);
438 	refcount_inc(&map->users);
439 }
440 
441 static void gntdev_vma_close(struct vm_area_struct *vma)
442 {
443 	struct gntdev_grant_map *map = vma->vm_private_data;
444 	struct file *file = vma->vm_file;
445 	struct gntdev_priv *priv = file->private_data;
446 
447 	pr_debug("gntdev_vma_close %p\n", vma);
448 	if (use_ptemod) {
449 		/* It is possible that an mmu notifier could be running
450 		 * concurrently, so take priv->lock to ensure that the vma won't
451 		 * vanishing during the unmap_grant_pages call, since we will
452 		 * spin here until that completes. Such a concurrent call will
453 		 * not do any unmapping, since that has been done prior to
454 		 * closing the vma, but it may still iterate the unmap_ops list.
455 		 */
456 		mutex_lock(&priv->lock);
457 		map->vma = NULL;
458 		mutex_unlock(&priv->lock);
459 	}
460 	vma->vm_private_data = NULL;
461 	gntdev_put_map(priv, map);
462 }
463 
464 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
465 						 unsigned long addr)
466 {
467 	struct gntdev_grant_map *map = vma->vm_private_data;
468 
469 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
470 }
471 
472 static const struct vm_operations_struct gntdev_vmops = {
473 	.open = gntdev_vma_open,
474 	.close = gntdev_vma_close,
475 	.find_special_page = gntdev_vma_find_special_page,
476 };
477 
478 /* ------------------------------------------------------------------ */
479 
480 static bool in_range(struct gntdev_grant_map *map,
481 			      unsigned long start, unsigned long end)
482 {
483 	if (!map->vma)
484 		return false;
485 	if (map->vma->vm_start >= end)
486 		return false;
487 	if (map->vma->vm_end <= start)
488 		return false;
489 
490 	return true;
491 }
492 
493 static int unmap_if_in_range(struct gntdev_grant_map *map,
494 			      unsigned long start, unsigned long end,
495 			      bool blockable)
496 {
497 	unsigned long mstart, mend;
498 	int err;
499 
500 	if (!in_range(map, start, end))
501 		return 0;
502 
503 	if (!blockable)
504 		return -EAGAIN;
505 
506 	mstart = max(start, map->vma->vm_start);
507 	mend   = min(end,   map->vma->vm_end);
508 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
509 			map->index, map->count,
510 			map->vma->vm_start, map->vma->vm_end,
511 			start, end, mstart, mend);
512 	err = unmap_grant_pages(map,
513 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
514 				(mend - mstart) >> PAGE_SHIFT);
515 	WARN_ON(err);
516 
517 	return 0;
518 }
519 
520 static int mn_invl_range_start(struct mmu_notifier *mn,
521 			       const struct mmu_notifier_range *range)
522 {
523 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
524 	struct gntdev_grant_map *map;
525 	int ret = 0;
526 
527 	if (mmu_notifier_range_blockable(range))
528 		mutex_lock(&priv->lock);
529 	else if (!mutex_trylock(&priv->lock))
530 		return -EAGAIN;
531 
532 	list_for_each_entry(map, &priv->maps, next) {
533 		ret = unmap_if_in_range(map, range->start, range->end,
534 					mmu_notifier_range_blockable(range));
535 		if (ret)
536 			goto out_unlock;
537 	}
538 	list_for_each_entry(map, &priv->freeable_maps, next) {
539 		ret = unmap_if_in_range(map, range->start, range->end,
540 					mmu_notifier_range_blockable(range));
541 		if (ret)
542 			goto out_unlock;
543 	}
544 
545 out_unlock:
546 	mutex_unlock(&priv->lock);
547 
548 	return ret;
549 }
550 
551 static void mn_release(struct mmu_notifier *mn,
552 		       struct mm_struct *mm)
553 {
554 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
555 	struct gntdev_grant_map *map;
556 	int err;
557 
558 	mutex_lock(&priv->lock);
559 	list_for_each_entry(map, &priv->maps, next) {
560 		if (!map->vma)
561 			continue;
562 		pr_debug("map %d+%d (%lx %lx)\n",
563 				map->index, map->count,
564 				map->vma->vm_start, map->vma->vm_end);
565 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
566 		WARN_ON(err);
567 	}
568 	list_for_each_entry(map, &priv->freeable_maps, next) {
569 		if (!map->vma)
570 			continue;
571 		pr_debug("map %d+%d (%lx %lx)\n",
572 				map->index, map->count,
573 				map->vma->vm_start, map->vma->vm_end);
574 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
575 		WARN_ON(err);
576 	}
577 	mutex_unlock(&priv->lock);
578 }
579 
580 static const struct mmu_notifier_ops gntdev_mmu_ops = {
581 	.release                = mn_release,
582 	.invalidate_range_start = mn_invl_range_start,
583 };
584 
585 /* ------------------------------------------------------------------ */
586 
587 static int gntdev_open(struct inode *inode, struct file *flip)
588 {
589 	struct gntdev_priv *priv;
590 	int ret = 0;
591 
592 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
593 	if (!priv)
594 		return -ENOMEM;
595 
596 	INIT_LIST_HEAD(&priv->maps);
597 	INIT_LIST_HEAD(&priv->freeable_maps);
598 	mutex_init(&priv->lock);
599 
600 #ifdef CONFIG_XEN_GNTDEV_DMABUF
601 	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
602 	if (IS_ERR(priv->dmabuf_priv)) {
603 		ret = PTR_ERR(priv->dmabuf_priv);
604 		kfree(priv);
605 		return ret;
606 	}
607 #endif
608 
609 	if (use_ptemod) {
610 		priv->mm = get_task_mm(current);
611 		if (!priv->mm) {
612 			kfree(priv);
613 			return -ENOMEM;
614 		}
615 		priv->mn.ops = &gntdev_mmu_ops;
616 		ret = mmu_notifier_register(&priv->mn, priv->mm);
617 		mmput(priv->mm);
618 	}
619 
620 	if (ret) {
621 		kfree(priv);
622 		return ret;
623 	}
624 
625 	flip->private_data = priv;
626 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
627 	priv->dma_dev = gntdev_miscdev.this_device;
628 
629 	/*
630 	 * The device is not spawn from a device tree, so arch_setup_dma_ops
631 	 * is not called, thus leaving the device with dummy DMA ops.
632 	 * Fix this by calling of_dma_configure() with a NULL node to set
633 	 * default DMA ops.
634 	 */
635 	of_dma_configure(priv->dma_dev, NULL, true);
636 #endif
637 	pr_debug("priv %p\n", priv);
638 
639 	return 0;
640 }
641 
642 static int gntdev_release(struct inode *inode, struct file *flip)
643 {
644 	struct gntdev_priv *priv = flip->private_data;
645 	struct gntdev_grant_map *map;
646 
647 	pr_debug("priv %p\n", priv);
648 
649 	mutex_lock(&priv->lock);
650 	while (!list_empty(&priv->maps)) {
651 		map = list_entry(priv->maps.next,
652 				 struct gntdev_grant_map, next);
653 		list_del(&map->next);
654 		gntdev_put_map(NULL /* already removed */, map);
655 	}
656 	WARN_ON(!list_empty(&priv->freeable_maps));
657 	mutex_unlock(&priv->lock);
658 
659 #ifdef CONFIG_XEN_GNTDEV_DMABUF
660 	gntdev_dmabuf_fini(priv->dmabuf_priv);
661 #endif
662 
663 	if (use_ptemod)
664 		mmu_notifier_unregister(&priv->mn, priv->mm);
665 
666 	kfree(priv);
667 	return 0;
668 }
669 
670 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
671 				       struct ioctl_gntdev_map_grant_ref __user *u)
672 {
673 	struct ioctl_gntdev_map_grant_ref op;
674 	struct gntdev_grant_map *map;
675 	int err;
676 
677 	if (copy_from_user(&op, u, sizeof(op)) != 0)
678 		return -EFAULT;
679 	pr_debug("priv %p, add %d\n", priv, op.count);
680 	if (unlikely(op.count <= 0))
681 		return -EINVAL;
682 
683 	err = -ENOMEM;
684 	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
685 	if (!map)
686 		return err;
687 
688 	if (unlikely(gntdev_account_mapped_pages(op.count))) {
689 		pr_debug("can't map: over limit\n");
690 		gntdev_put_map(NULL, map);
691 		return err;
692 	}
693 
694 	if (copy_from_user(map->grants, &u->refs,
695 			   sizeof(map->grants[0]) * op.count) != 0) {
696 		gntdev_put_map(NULL, map);
697 		return -EFAULT;
698 	}
699 
700 	mutex_lock(&priv->lock);
701 	gntdev_add_map(priv, map);
702 	op.index = map->index << PAGE_SHIFT;
703 	mutex_unlock(&priv->lock);
704 
705 	if (copy_to_user(u, &op, sizeof(op)) != 0)
706 		return -EFAULT;
707 
708 	return 0;
709 }
710 
711 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
712 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
713 {
714 	struct ioctl_gntdev_unmap_grant_ref op;
715 	struct gntdev_grant_map *map;
716 	int err = -ENOENT;
717 
718 	if (copy_from_user(&op, u, sizeof(op)) != 0)
719 		return -EFAULT;
720 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
721 
722 	mutex_lock(&priv->lock);
723 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
724 	if (map) {
725 		list_del(&map->next);
726 		if (populate_freeable_maps)
727 			list_add_tail(&map->next, &priv->freeable_maps);
728 		err = 0;
729 	}
730 	mutex_unlock(&priv->lock);
731 	if (map)
732 		gntdev_put_map(priv, map);
733 	return err;
734 }
735 
736 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
737 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
738 {
739 	struct ioctl_gntdev_get_offset_for_vaddr op;
740 	struct vm_area_struct *vma;
741 	struct gntdev_grant_map *map;
742 	int rv = -EINVAL;
743 
744 	if (copy_from_user(&op, u, sizeof(op)) != 0)
745 		return -EFAULT;
746 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
747 
748 	down_read(&current->mm->mmap_sem);
749 	vma = find_vma(current->mm, op.vaddr);
750 	if (!vma || vma->vm_ops != &gntdev_vmops)
751 		goto out_unlock;
752 
753 	map = vma->vm_private_data;
754 	if (!map)
755 		goto out_unlock;
756 
757 	op.offset = map->index << PAGE_SHIFT;
758 	op.count = map->count;
759 	rv = 0;
760 
761  out_unlock:
762 	up_read(&current->mm->mmap_sem);
763 
764 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
765 		return -EFAULT;
766 	return rv;
767 }
768 
769 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
770 {
771 	struct ioctl_gntdev_unmap_notify op;
772 	struct gntdev_grant_map *map;
773 	int rc;
774 	int out_flags;
775 	unsigned int out_event;
776 
777 	if (copy_from_user(&op, u, sizeof(op)))
778 		return -EFAULT;
779 
780 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
781 		return -EINVAL;
782 
783 	/* We need to grab a reference to the event channel we are going to use
784 	 * to send the notify before releasing the reference we may already have
785 	 * (if someone has called this ioctl twice). This is required so that
786 	 * it is possible to change the clear_byte part of the notification
787 	 * without disturbing the event channel part, which may now be the last
788 	 * reference to that event channel.
789 	 */
790 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
791 		if (evtchn_get(op.event_channel_port))
792 			return -EINVAL;
793 	}
794 
795 	out_flags = op.action;
796 	out_event = op.event_channel_port;
797 
798 	mutex_lock(&priv->lock);
799 
800 	list_for_each_entry(map, &priv->maps, next) {
801 		uint64_t begin = map->index << PAGE_SHIFT;
802 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
803 		if (op.index >= begin && op.index < end)
804 			goto found;
805 	}
806 	rc = -ENOENT;
807 	goto unlock_out;
808 
809  found:
810 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
811 			(map->flags & GNTMAP_readonly)) {
812 		rc = -EINVAL;
813 		goto unlock_out;
814 	}
815 
816 	out_flags = map->notify.flags;
817 	out_event = map->notify.event;
818 
819 	map->notify.flags = op.action;
820 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
821 	map->notify.event = op.event_channel_port;
822 
823 	rc = 0;
824 
825  unlock_out:
826 	mutex_unlock(&priv->lock);
827 
828 	/* Drop the reference to the event channel we did not save in the map */
829 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
830 		evtchn_put(out_event);
831 
832 	return rc;
833 }
834 
835 #define GNTDEV_COPY_BATCH 16
836 
837 struct gntdev_copy_batch {
838 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
839 	struct page *pages[GNTDEV_COPY_BATCH];
840 	s16 __user *status[GNTDEV_COPY_BATCH];
841 	unsigned int nr_ops;
842 	unsigned int nr_pages;
843 };
844 
845 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
846 			   bool writeable, unsigned long *gfn)
847 {
848 	unsigned long addr = (unsigned long)virt;
849 	struct page *page;
850 	unsigned long xen_pfn;
851 	int ret;
852 
853 	ret = get_user_pages_fast(addr, 1, writeable ? FOLL_WRITE : 0, &page);
854 	if (ret < 0)
855 		return ret;
856 
857 	batch->pages[batch->nr_pages++] = page;
858 
859 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
860 	*gfn = pfn_to_gfn(xen_pfn);
861 
862 	return 0;
863 }
864 
865 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
866 {
867 	unsigned int i;
868 
869 	for (i = 0; i < batch->nr_pages; i++)
870 		put_page(batch->pages[i]);
871 	batch->nr_pages = 0;
872 }
873 
874 static int gntdev_copy(struct gntdev_copy_batch *batch)
875 {
876 	unsigned int i;
877 
878 	gnttab_batch_copy(batch->ops, batch->nr_ops);
879 	gntdev_put_pages(batch);
880 
881 	/*
882 	 * For each completed op, update the status if the op failed
883 	 * and all previous ops for the segment were successful.
884 	 */
885 	for (i = 0; i < batch->nr_ops; i++) {
886 		s16 status = batch->ops[i].status;
887 		s16 old_status;
888 
889 		if (status == GNTST_okay)
890 			continue;
891 
892 		if (__get_user(old_status, batch->status[i]))
893 			return -EFAULT;
894 
895 		if (old_status != GNTST_okay)
896 			continue;
897 
898 		if (__put_user(status, batch->status[i]))
899 			return -EFAULT;
900 	}
901 
902 	batch->nr_ops = 0;
903 	return 0;
904 }
905 
906 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
907 				 struct gntdev_grant_copy_segment *seg,
908 				 s16 __user *status)
909 {
910 	uint16_t copied = 0;
911 
912 	/*
913 	 * Disallow local -> local copies since there is only space in
914 	 * batch->pages for one page per-op and this would be a very
915 	 * expensive memcpy().
916 	 */
917 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
918 		return -EINVAL;
919 
920 	/* Can't cross page if source/dest is a grant ref. */
921 	if (seg->flags & GNTCOPY_source_gref) {
922 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
923 			return -EINVAL;
924 	}
925 	if (seg->flags & GNTCOPY_dest_gref) {
926 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
927 			return -EINVAL;
928 	}
929 
930 	if (put_user(GNTST_okay, status))
931 		return -EFAULT;
932 
933 	while (copied < seg->len) {
934 		struct gnttab_copy *op;
935 		void __user *virt;
936 		size_t len, off;
937 		unsigned long gfn;
938 		int ret;
939 
940 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
941 			ret = gntdev_copy(batch);
942 			if (ret < 0)
943 				return ret;
944 		}
945 
946 		len = seg->len - copied;
947 
948 		op = &batch->ops[batch->nr_ops];
949 		op->flags = 0;
950 
951 		if (seg->flags & GNTCOPY_source_gref) {
952 			op->source.u.ref = seg->source.foreign.ref;
953 			op->source.domid = seg->source.foreign.domid;
954 			op->source.offset = seg->source.foreign.offset + copied;
955 			op->flags |= GNTCOPY_source_gref;
956 		} else {
957 			virt = seg->source.virt + copied;
958 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
959 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
960 
961 			ret = gntdev_get_page(batch, virt, false, &gfn);
962 			if (ret < 0)
963 				return ret;
964 
965 			op->source.u.gmfn = gfn;
966 			op->source.domid = DOMID_SELF;
967 			op->source.offset = off;
968 		}
969 
970 		if (seg->flags & GNTCOPY_dest_gref) {
971 			op->dest.u.ref = seg->dest.foreign.ref;
972 			op->dest.domid = seg->dest.foreign.domid;
973 			op->dest.offset = seg->dest.foreign.offset + copied;
974 			op->flags |= GNTCOPY_dest_gref;
975 		} else {
976 			virt = seg->dest.virt + copied;
977 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
978 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
979 
980 			ret = gntdev_get_page(batch, virt, true, &gfn);
981 			if (ret < 0)
982 				return ret;
983 
984 			op->dest.u.gmfn = gfn;
985 			op->dest.domid = DOMID_SELF;
986 			op->dest.offset = off;
987 		}
988 
989 		op->len = len;
990 		copied += len;
991 
992 		batch->status[batch->nr_ops] = status;
993 		batch->nr_ops++;
994 	}
995 
996 	return 0;
997 }
998 
999 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
1000 {
1001 	struct ioctl_gntdev_grant_copy copy;
1002 	struct gntdev_copy_batch batch;
1003 	unsigned int i;
1004 	int ret = 0;
1005 
1006 	if (copy_from_user(&copy, u, sizeof(copy)))
1007 		return -EFAULT;
1008 
1009 	batch.nr_ops = 0;
1010 	batch.nr_pages = 0;
1011 
1012 	for (i = 0; i < copy.count; i++) {
1013 		struct gntdev_grant_copy_segment seg;
1014 
1015 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
1016 			ret = -EFAULT;
1017 			goto out;
1018 		}
1019 
1020 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
1021 		if (ret < 0)
1022 			goto out;
1023 
1024 		cond_resched();
1025 	}
1026 	if (batch.nr_ops)
1027 		ret = gntdev_copy(&batch);
1028 	return ret;
1029 
1030   out:
1031 	gntdev_put_pages(&batch);
1032 	return ret;
1033 }
1034 
1035 static long gntdev_ioctl(struct file *flip,
1036 			 unsigned int cmd, unsigned long arg)
1037 {
1038 	struct gntdev_priv *priv = flip->private_data;
1039 	void __user *ptr = (void __user *)arg;
1040 
1041 	switch (cmd) {
1042 	case IOCTL_GNTDEV_MAP_GRANT_REF:
1043 		return gntdev_ioctl_map_grant_ref(priv, ptr);
1044 
1045 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
1046 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
1047 
1048 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
1049 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
1050 
1051 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
1052 		return gntdev_ioctl_notify(priv, ptr);
1053 
1054 	case IOCTL_GNTDEV_GRANT_COPY:
1055 		return gntdev_ioctl_grant_copy(priv, ptr);
1056 
1057 #ifdef CONFIG_XEN_GNTDEV_DMABUF
1058 	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
1059 		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
1060 
1061 	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
1062 		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
1063 
1064 	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
1065 		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
1066 
1067 	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
1068 		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
1069 #endif
1070 
1071 	default:
1072 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
1073 		return -ENOIOCTLCMD;
1074 	}
1075 
1076 	return 0;
1077 }
1078 
1079 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
1080 {
1081 	struct gntdev_priv *priv = flip->private_data;
1082 	int index = vma->vm_pgoff;
1083 	int count = vma_pages(vma);
1084 	struct gntdev_grant_map *map;
1085 	int err = -EINVAL;
1086 
1087 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
1088 		return -EINVAL;
1089 
1090 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
1091 			index, count, vma->vm_start, vma->vm_pgoff);
1092 
1093 	mutex_lock(&priv->lock);
1094 	map = gntdev_find_map_index(priv, index, count);
1095 	if (!map)
1096 		goto unlock_out;
1097 	if (use_ptemod && map->vma)
1098 		goto unlock_out;
1099 	if (use_ptemod && priv->mm != vma->vm_mm) {
1100 		pr_warn("Huh? Other mm?\n");
1101 		goto unlock_out;
1102 	}
1103 
1104 	refcount_inc(&map->users);
1105 
1106 	vma->vm_ops = &gntdev_vmops;
1107 
1108 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1109 
1110 	if (use_ptemod)
1111 		vma->vm_flags |= VM_DONTCOPY;
1112 
1113 	vma->vm_private_data = map;
1114 
1115 	if (use_ptemod)
1116 		map->vma = vma;
1117 
1118 	if (map->flags) {
1119 		if ((vma->vm_flags & VM_WRITE) &&
1120 				(map->flags & GNTMAP_readonly))
1121 			goto out_unlock_put;
1122 	} else {
1123 		map->flags = GNTMAP_host_map;
1124 		if (!(vma->vm_flags & VM_WRITE))
1125 			map->flags |= GNTMAP_readonly;
1126 	}
1127 
1128 	mutex_unlock(&priv->lock);
1129 
1130 	if (use_ptemod) {
1131 		map->pages_vm_start = vma->vm_start;
1132 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1133 					  vma->vm_end - vma->vm_start,
1134 					  find_grant_ptes, map);
1135 		if (err) {
1136 			pr_warn("find_grant_ptes() failure.\n");
1137 			goto out_put_map;
1138 		}
1139 	}
1140 
1141 	err = gntdev_map_grant_pages(map);
1142 	if (err)
1143 		goto out_put_map;
1144 
1145 	if (!use_ptemod) {
1146 		err = vm_map_pages(vma, map->pages, map->count);
1147 		if (err)
1148 			goto out_put_map;
1149 	} else {
1150 #ifdef CONFIG_X86
1151 		/*
1152 		 * If the PTEs were not made special by the grant map
1153 		 * hypercall, do so here.
1154 		 *
1155 		 * This is racy since the mapping is already visible
1156 		 * to userspace but userspace should be well-behaved
1157 		 * enough to not touch it until the mmap() call
1158 		 * returns.
1159 		 */
1160 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1161 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1162 					    vma->vm_end - vma->vm_start,
1163 					    set_grant_ptes_as_special, NULL);
1164 		}
1165 #endif
1166 	}
1167 
1168 	return 0;
1169 
1170 unlock_out:
1171 	mutex_unlock(&priv->lock);
1172 	return err;
1173 
1174 out_unlock_put:
1175 	mutex_unlock(&priv->lock);
1176 out_put_map:
1177 	if (use_ptemod) {
1178 		map->vma = NULL;
1179 		unmap_grant_pages(map, 0, map->count);
1180 	}
1181 	gntdev_put_map(priv, map);
1182 	return err;
1183 }
1184 
1185 static const struct file_operations gntdev_fops = {
1186 	.owner = THIS_MODULE,
1187 	.open = gntdev_open,
1188 	.release = gntdev_release,
1189 	.mmap = gntdev_mmap,
1190 	.unlocked_ioctl = gntdev_ioctl
1191 };
1192 
1193 static struct miscdevice gntdev_miscdev = {
1194 	.minor        = MISC_DYNAMIC_MINOR,
1195 	.name         = "xen/gntdev",
1196 	.fops         = &gntdev_fops,
1197 };
1198 
1199 /* ------------------------------------------------------------------ */
1200 
1201 static int __init gntdev_init(void)
1202 {
1203 	int err;
1204 
1205 	if (!xen_domain())
1206 		return -ENODEV;
1207 
1208 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1209 
1210 	err = misc_register(&gntdev_miscdev);
1211 	if (err != 0) {
1212 		pr_err("Could not register gntdev device\n");
1213 		return err;
1214 	}
1215 	return 0;
1216 }
1217 
1218 static void __exit gntdev_exit(void)
1219 {
1220 	misc_deregister(&gntdev_miscdev);
1221 }
1222 
1223 module_init(gntdev_init);
1224 module_exit(gntdev_exit);
1225 
1226 /* ------------------------------------------------------------------ */
1227