gntdev.c

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