Linux Audio

Check our new training course

Loading...
v5.9
   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
  39#include <xen/xen.h>
  40#include <xen/grant_table.h>
  41#include <xen/balloon.h>
  42#include <xen/gntdev.h>
  43#include <xen/events.h>
  44#include <xen/page.h>
  45#include <asm/xen/hypervisor.h>
  46#include <asm/xen/hypercall.h>
  47
  48#include "gntdev-common.h"
  49#ifdef CONFIG_XEN_GNTDEV_DMABUF
  50#include "gntdev-dmabuf.h"
  51#endif
  52
  53MODULE_LICENSE("GPL");
  54MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
  55	      "Gerd Hoffmann <kraxel@redhat.com>");
  56MODULE_DESCRIPTION("User-space granted page access driver");
  57
  58static unsigned int limit = 64*1024;
  59module_param(limit, uint, 0644);
  60MODULE_PARM_DESC(limit,
  61	"Maximum number of grants that may be mapped by one mapping request");
  62
  63static int use_ptemod;
  64
  65static int unmap_grant_pages(struct gntdev_grant_map *map,
  66			     int offset, int pages);
  67
  68static struct miscdevice gntdev_miscdev;
  69
  70/* ------------------------------------------------------------------ */
  71
  72bool gntdev_test_page_count(unsigned int count)
  73{
  74	return !count || count > limit;
  75}
  76
  77static void gntdev_print_maps(struct gntdev_priv *priv,
  78			      char *text, int text_index)
  79{
  80#ifdef DEBUG
  81	struct gntdev_grant_map *map;
  82
  83	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
  84	list_for_each_entry(map, &priv->maps, next)
  85		pr_debug("  index %2d, count %2d %s\n",
  86		       map->index, map->count,
  87		       map->index == text_index && text ? text : "");
  88#endif
  89}
  90
  91static void gntdev_free_map(struct gntdev_grant_map *map)
  92{
  93	if (map == NULL)
  94		return;
  95
  96#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
  97	if (map->dma_vaddr) {
  98		struct gnttab_dma_alloc_args args;
  99
 100		args.dev = map->dma_dev;
 101		args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
 102		args.nr_pages = map->count;
 103		args.pages = map->pages;
 104		args.frames = map->frames;
 105		args.vaddr = map->dma_vaddr;
 106		args.dev_bus_addr = map->dma_bus_addr;
 107
 108		gnttab_dma_free_pages(&args);
 109	} else
 110#endif
 111	if (map->pages)
 112		gnttab_free_pages(map->count, map->pages);
 113
 114#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 115	kvfree(map->frames);
 116#endif
 117	kvfree(map->pages);
 118	kvfree(map->grants);
 119	kvfree(map->map_ops);
 120	kvfree(map->unmap_ops);
 121	kvfree(map->kmap_ops);
 122	kvfree(map->kunmap_ops);
 123	kfree(map);
 124}
 125
 126struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
 127					  int dma_flags)
 128{
 129	struct gntdev_grant_map *add;
 130	int i;
 131
 132	add = kzalloc(sizeof(*add), GFP_KERNEL);
 133	if (NULL == add)
 134		return NULL;
 135
 136	add->grants    = kvcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
 137	add->map_ops   = kvcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
 138	add->unmap_ops = kvcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
 139	add->kmap_ops  = kvcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
 140	add->kunmap_ops = kvcalloc(count,
 141				   sizeof(add->kunmap_ops[0]), GFP_KERNEL);
 142	add->pages     = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
 143	if (NULL == add->grants    ||
 144	    NULL == add->map_ops   ||
 145	    NULL == add->unmap_ops ||
 146	    NULL == add->kmap_ops  ||
 147	    NULL == add->kunmap_ops ||
 148	    NULL == add->pages)
 149		goto err;
 
 
 
 
 
 
 
 
 150
 151#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 152	add->dma_flags = dma_flags;
 153
 154	/*
 155	 * Check if this mapping is requested to be backed
 156	 * by a DMA buffer.
 157	 */
 158	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
 159		struct gnttab_dma_alloc_args args;
 160
 161		add->frames = kvcalloc(count, sizeof(add->frames[0]),
 162				       GFP_KERNEL);
 163		if (!add->frames)
 164			goto err;
 165
 166		/* Remember the device, so we can free DMA memory. */
 167		add->dma_dev = priv->dma_dev;
 168
 169		args.dev = priv->dma_dev;
 170		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
 171		args.nr_pages = count;
 172		args.pages = add->pages;
 173		args.frames = add->frames;
 174
 175		if (gnttab_dma_alloc_pages(&args))
 176			goto err;
 177
 178		add->dma_vaddr = args.vaddr;
 179		add->dma_bus_addr = args.dev_bus_addr;
 180	} else
 181#endif
 182	if (gnttab_alloc_pages(count, add->pages))
 183		goto err;
 184
 185	for (i = 0; i < count; i++) {
 186		add->map_ops[i].handle = -1;
 187		add->unmap_ops[i].handle = -1;
 188		add->kmap_ops[i].handle = -1;
 189		add->kunmap_ops[i].handle = -1;
 
 
 
 
 190	}
 191
 192	add->index = 0;
 193	add->count = count;
 194	refcount_set(&add->users, 1);
 195
 196	return add;
 197
 198err:
 199	gntdev_free_map(add);
 200	return NULL;
 201}
 202
 203void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
 204{
 205	struct gntdev_grant_map *map;
 206
 207	list_for_each_entry(map, &priv->maps, next) {
 208		if (add->index + add->count < map->index) {
 209			list_add_tail(&add->next, &map->next);
 210			goto done;
 211		}
 212		add->index = map->index + map->count;
 213	}
 214	list_add_tail(&add->next, &priv->maps);
 215
 216done:
 217	gntdev_print_maps(priv, "[new]", add->index);
 218}
 219
 220static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
 221						      int index, int count)
 222{
 223	struct gntdev_grant_map *map;
 224
 225	list_for_each_entry(map, &priv->maps, next) {
 226		if (map->index != index)
 227			continue;
 228		if (count && map->count != count)
 229			continue;
 230		return map;
 231	}
 232	return NULL;
 233}
 234
 235void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
 236{
 237	if (!map)
 238		return;
 239
 240	if (!refcount_dec_and_test(&map->users))
 241		return;
 242
 243	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
 244		notify_remote_via_evtchn(map->notify.event);
 245		evtchn_put(map->notify.event);
 246	}
 247
 248	if (map->pages && !use_ptemod)
 249		unmap_grant_pages(map, 0, map->count);
 250	gntdev_free_map(map);
 251}
 252
 253/* ------------------------------------------------------------------ */
 254
 255static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
 256{
 257	struct gntdev_grant_map *map = data;
 258	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
 259	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
 260	u64 pte_maddr;
 261
 262	BUG_ON(pgnr >= map->count);
 263	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
 264
 265	/*
 266	 * Set the PTE as special to force get_user_pages_fast() fall
 267	 * back to the slow path.  If this is not supported as part of
 268	 * the grant map, it will be done afterwards.
 269	 */
 270	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
 271		flags |= (1 << _GNTMAP_guest_avail0);
 272
 273	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
 274			  map->grants[pgnr].ref,
 275			  map->grants[pgnr].domid);
 276	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
 277			    -1 /* handle */);
 278	return 0;
 279}
 280
 281#ifdef CONFIG_X86
 282static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
 283{
 284	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
 285	return 0;
 286}
 287#endif
 288
 289int gntdev_map_grant_pages(struct gntdev_grant_map *map)
 290{
 291	int i, err = 0;
 292
 293	if (!use_ptemod) {
 294		/* Note: it could already be mapped */
 295		if (map->map_ops[0].handle != -1)
 296			return 0;
 297		for (i = 0; i < map->count; i++) {
 298			unsigned long addr = (unsigned long)
 299				pfn_to_kaddr(page_to_pfn(map->pages[i]));
 300			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
 301				map->grants[i].ref,
 302				map->grants[i].domid);
 303			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
 304				map->flags, -1 /* handle */);
 305		}
 306	} else {
 307		/*
 308		 * Setup the map_ops corresponding to the pte entries pointing
 309		 * to the kernel linear addresses of the struct pages.
 310		 * These ptes are completely different from the user ptes dealt
 311		 * with find_grant_ptes.
 
 
 
 
 
 312		 */
 
 
 
 313		for (i = 0; i < map->count; i++) {
 314			unsigned long address = (unsigned long)
 315				pfn_to_kaddr(page_to_pfn(map->pages[i]));
 316			BUG_ON(PageHighMem(map->pages[i]));
 317
 318			gnttab_set_map_op(&map->kmap_ops[i], address,
 319				map->flags | GNTMAP_host_map,
 320				map->grants[i].ref,
 321				map->grants[i].domid);
 322			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
 323				map->flags | GNTMAP_host_map, -1);
 324		}
 325	}
 326
 327	pr_debug("map %d+%d\n", map->index, map->count);
 328	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
 329			map->pages, map->count);
 330	if (err)
 331		return err;
 332
 333	for (i = 0; i < map->count; i++) {
 334		if (map->map_ops[i].status) {
 
 
 335			err = -EINVAL;
 336			continue;
 337		}
 338
 339		map->unmap_ops[i].handle = map->map_ops[i].handle;
 340		if (use_ptemod)
 341			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
 342#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 343		else if (map->dma_vaddr) {
 344			unsigned long bfn;
 345
 346			bfn = pfn_to_bfn(page_to_pfn(map->pages[i]));
 347			map->unmap_ops[i].dev_bus_addr = __pfn_to_phys(bfn);
 
 
 
 348		}
 349#endif
 350	}
 351	return err;
 352}
 353
 354static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
 355			       int pages)
 356{
 357	int i, err = 0;
 358	struct gntab_unmap_queue_data unmap_data;
 359
 360	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
 361		int pgno = (map->notify.addr >> PAGE_SHIFT);
 362		if (pgno >= offset && pgno < offset + pages) {
 363			/* No need for kmap, pages are in lowmem */
 364			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
 365			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
 366			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
 367		}
 368	}
 369
 370	unmap_data.unmap_ops = map->unmap_ops + offset;
 371	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
 372	unmap_data.pages = map->pages + offset;
 373	unmap_data.count = pages;
 374
 375	err = gnttab_unmap_refs_sync(&unmap_data);
 376	if (err)
 377		return err;
 378
 379	for (i = 0; i < pages; i++) {
 380		if (map->unmap_ops[offset+i].status)
 381			err = -EINVAL;
 382		pr_debug("unmap handle=%d st=%d\n",
 383			map->unmap_ops[offset+i].handle,
 384			map->unmap_ops[offset+i].status);
 385		map->unmap_ops[offset+i].handle = -1;
 
 
 
 
 
 
 
 
 386	}
 387	return err;
 388}
 389
 390static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
 391			     int pages)
 392{
 393	int range, err = 0;
 394
 395	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
 396
 397	/* It is possible the requested range will have a "hole" where we
 398	 * already unmapped some of the grants. Only unmap valid ranges.
 399	 */
 400	while (pages && !err) {
 401		while (pages && map->unmap_ops[offset].handle == -1) {
 
 402			offset++;
 403			pages--;
 404		}
 405		range = 0;
 406		while (range < pages) {
 407			if (map->unmap_ops[offset+range].handle == -1)
 
 408				break;
 409			range++;
 410		}
 411		err = __unmap_grant_pages(map, offset, range);
 412		offset += range;
 413		pages -= range;
 414	}
 415
 416	return err;
 417}
 418
 419/* ------------------------------------------------------------------ */
 420
 421static void gntdev_vma_open(struct vm_area_struct *vma)
 422{
 423	struct gntdev_grant_map *map = vma->vm_private_data;
 424
 425	pr_debug("gntdev_vma_open %p\n", vma);
 426	refcount_inc(&map->users);
 427}
 428
 429static void gntdev_vma_close(struct vm_area_struct *vma)
 430{
 431	struct gntdev_grant_map *map = vma->vm_private_data;
 432	struct file *file = vma->vm_file;
 433	struct gntdev_priv *priv = file->private_data;
 434
 435	pr_debug("gntdev_vma_close %p\n", vma);
 436	if (use_ptemod) {
 437		WARN_ON(map->vma != vma);
 438		mmu_interval_notifier_remove(&map->notifier);
 439		map->vma = NULL;
 440	}
 441	vma->vm_private_data = NULL;
 442	gntdev_put_map(priv, map);
 443}
 444
 445static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
 446						 unsigned long addr)
 447{
 448	struct gntdev_grant_map *map = vma->vm_private_data;
 449
 450	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
 451}
 452
 453static const struct vm_operations_struct gntdev_vmops = {
 454	.open = gntdev_vma_open,
 455	.close = gntdev_vma_close,
 456	.find_special_page = gntdev_vma_find_special_page,
 457};
 458
 459/* ------------------------------------------------------------------ */
 460
 461static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
 462			      const struct mmu_notifier_range *range,
 463			      unsigned long cur_seq)
 464{
 465	struct gntdev_grant_map *map =
 466		container_of(mn, struct gntdev_grant_map, notifier);
 467	unsigned long mstart, mend;
 468	int err;
 469
 470	if (!mmu_notifier_range_blockable(range))
 471		return false;
 472
 473	/*
 474	 * If the VMA is split or otherwise changed the notifier is not
 475	 * updated, but we don't want to process VA's outside the modified
 476	 * VMA. FIXME: It would be much more understandable to just prevent
 477	 * modifying the VMA in the first place.
 478	 */
 479	if (map->vma->vm_start >= range->end ||
 480	    map->vma->vm_end <= range->start)
 481		return true;
 482
 483	mstart = max(range->start, map->vma->vm_start);
 484	mend = min(range->end, map->vma->vm_end);
 485	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
 486			map->index, map->count,
 487			map->vma->vm_start, map->vma->vm_end,
 488			range->start, range->end, mstart, mend);
 489	err = unmap_grant_pages(map,
 490				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
 491				(mend - mstart) >> PAGE_SHIFT);
 492	WARN_ON(err);
 493
 494	return true;
 495}
 496
 497static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
 498	.invalidate = gntdev_invalidate,
 499};
 500
 501/* ------------------------------------------------------------------ */
 502
 503static int gntdev_open(struct inode *inode, struct file *flip)
 504{
 505	struct gntdev_priv *priv;
 506
 507	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 508	if (!priv)
 509		return -ENOMEM;
 510
 511	INIT_LIST_HEAD(&priv->maps);
 512	mutex_init(&priv->lock);
 513
 514#ifdef CONFIG_XEN_GNTDEV_DMABUF
 515	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
 516	if (IS_ERR(priv->dmabuf_priv)) {
 517		int ret = PTR_ERR(priv->dmabuf_priv);
 518
 519		kfree(priv);
 520		return ret;
 521	}
 522#endif
 523
 524	flip->private_data = priv;
 525#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 526	priv->dma_dev = gntdev_miscdev.this_device;
 527	dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
 528#endif
 529	pr_debug("priv %p\n", priv);
 530
 531	return 0;
 532}
 533
 534static int gntdev_release(struct inode *inode, struct file *flip)
 535{
 536	struct gntdev_priv *priv = flip->private_data;
 537	struct gntdev_grant_map *map;
 538
 539	pr_debug("priv %p\n", priv);
 540
 541	mutex_lock(&priv->lock);
 542	while (!list_empty(&priv->maps)) {
 543		map = list_entry(priv->maps.next,
 544				 struct gntdev_grant_map, next);
 545		list_del(&map->next);
 546		gntdev_put_map(NULL /* already removed */, map);
 547	}
 548	mutex_unlock(&priv->lock);
 549
 550#ifdef CONFIG_XEN_GNTDEV_DMABUF
 551	gntdev_dmabuf_fini(priv->dmabuf_priv);
 552#endif
 553
 554	kfree(priv);
 555	return 0;
 556}
 557
 558static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
 559				       struct ioctl_gntdev_map_grant_ref __user *u)
 560{
 561	struct ioctl_gntdev_map_grant_ref op;
 562	struct gntdev_grant_map *map;
 563	int err;
 564
 565	if (copy_from_user(&op, u, sizeof(op)) != 0)
 566		return -EFAULT;
 567	pr_debug("priv %p, add %d\n", priv, op.count);
 568	if (unlikely(gntdev_test_page_count(op.count)))
 569		return -EINVAL;
 570
 571	err = -ENOMEM;
 572	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
 573	if (!map)
 574		return err;
 575
 576	if (copy_from_user(map->grants, &u->refs,
 577			   sizeof(map->grants[0]) * op.count) != 0) {
 578		gntdev_put_map(NULL, map);
 579		return -EFAULT;
 580	}
 581
 582	mutex_lock(&priv->lock);
 583	gntdev_add_map(priv, map);
 584	op.index = map->index << PAGE_SHIFT;
 585	mutex_unlock(&priv->lock);
 586
 587	if (copy_to_user(u, &op, sizeof(op)) != 0)
 588		return -EFAULT;
 589
 590	return 0;
 591}
 592
 593static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
 594					 struct ioctl_gntdev_unmap_grant_ref __user *u)
 595{
 596	struct ioctl_gntdev_unmap_grant_ref op;
 597	struct gntdev_grant_map *map;
 598	int err = -ENOENT;
 599
 600	if (copy_from_user(&op, u, sizeof(op)) != 0)
 601		return -EFAULT;
 602	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
 603
 604	mutex_lock(&priv->lock);
 605	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
 606	if (map) {
 607		list_del(&map->next);
 608		err = 0;
 609	}
 610	mutex_unlock(&priv->lock);
 611	if (map)
 612		gntdev_put_map(priv, map);
 613	return err;
 614}
 615
 616static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
 617					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
 618{
 619	struct ioctl_gntdev_get_offset_for_vaddr op;
 620	struct vm_area_struct *vma;
 621	struct gntdev_grant_map *map;
 622	int rv = -EINVAL;
 623
 624	if (copy_from_user(&op, u, sizeof(op)) != 0)
 625		return -EFAULT;
 626	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
 627
 628	mmap_read_lock(current->mm);
 629	vma = find_vma(current->mm, op.vaddr);
 630	if (!vma || vma->vm_ops != &gntdev_vmops)
 631		goto out_unlock;
 632
 633	map = vma->vm_private_data;
 634	if (!map)
 635		goto out_unlock;
 636
 637	op.offset = map->index << PAGE_SHIFT;
 638	op.count = map->count;
 639	rv = 0;
 640
 641 out_unlock:
 642	mmap_read_unlock(current->mm);
 643
 644	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
 645		return -EFAULT;
 646	return rv;
 647}
 648
 649static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
 650{
 651	struct ioctl_gntdev_unmap_notify op;
 652	struct gntdev_grant_map *map;
 653	int rc;
 654	int out_flags;
 655	evtchn_port_t out_event;
 656
 657	if (copy_from_user(&op, u, sizeof(op)))
 658		return -EFAULT;
 659
 660	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
 661		return -EINVAL;
 662
 663	/* We need to grab a reference to the event channel we are going to use
 664	 * to send the notify before releasing the reference we may already have
 665	 * (if someone has called this ioctl twice). This is required so that
 666	 * it is possible to change the clear_byte part of the notification
 667	 * without disturbing the event channel part, which may now be the last
 668	 * reference to that event channel.
 669	 */
 670	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
 671		if (evtchn_get(op.event_channel_port))
 672			return -EINVAL;
 673	}
 674
 675	out_flags = op.action;
 676	out_event = op.event_channel_port;
 677
 678	mutex_lock(&priv->lock);
 679
 680	list_for_each_entry(map, &priv->maps, next) {
 681		uint64_t begin = map->index << PAGE_SHIFT;
 682		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
 683		if (op.index >= begin && op.index < end)
 684			goto found;
 685	}
 686	rc = -ENOENT;
 687	goto unlock_out;
 688
 689 found:
 690	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
 691			(map->flags & GNTMAP_readonly)) {
 692		rc = -EINVAL;
 693		goto unlock_out;
 694	}
 695
 696	out_flags = map->notify.flags;
 697	out_event = map->notify.event;
 698
 699	map->notify.flags = op.action;
 700	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
 701	map->notify.event = op.event_channel_port;
 702
 703	rc = 0;
 704
 705 unlock_out:
 706	mutex_unlock(&priv->lock);
 707
 708	/* Drop the reference to the event channel we did not save in the map */
 709	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
 710		evtchn_put(out_event);
 711
 712	return rc;
 713}
 714
 715#define GNTDEV_COPY_BATCH 16
 716
 717struct gntdev_copy_batch {
 718	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
 719	struct page *pages[GNTDEV_COPY_BATCH];
 720	s16 __user *status[GNTDEV_COPY_BATCH];
 721	unsigned int nr_ops;
 722	unsigned int nr_pages;
 
 723};
 724
 725static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
 726			   bool writeable, unsigned long *gfn)
 727{
 728	unsigned long addr = (unsigned long)virt;
 729	struct page *page;
 730	unsigned long xen_pfn;
 731	int ret;
 732
 733	ret = get_user_pages_fast(addr, 1, writeable ? FOLL_WRITE : 0, &page);
 734	if (ret < 0)
 735		return ret;
 736
 737	batch->pages[batch->nr_pages++] = page;
 738
 739	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
 740	*gfn = pfn_to_gfn(xen_pfn);
 741
 742	return 0;
 743}
 744
 745static void gntdev_put_pages(struct gntdev_copy_batch *batch)
 746{
 747	unsigned int i;
 748
 749	for (i = 0; i < batch->nr_pages; i++)
 750		put_page(batch->pages[i]);
 751	batch->nr_pages = 0;
 
 752}
 753
 754static int gntdev_copy(struct gntdev_copy_batch *batch)
 755{
 756	unsigned int i;
 757
 758	gnttab_batch_copy(batch->ops, batch->nr_ops);
 759	gntdev_put_pages(batch);
 760
 761	/*
 762	 * For each completed op, update the status if the op failed
 763	 * and all previous ops for the segment were successful.
 764	 */
 765	for (i = 0; i < batch->nr_ops; i++) {
 766		s16 status = batch->ops[i].status;
 767		s16 old_status;
 768
 769		if (status == GNTST_okay)
 770			continue;
 771
 772		if (__get_user(old_status, batch->status[i]))
 773			return -EFAULT;
 774
 775		if (old_status != GNTST_okay)
 776			continue;
 777
 778		if (__put_user(status, batch->status[i]))
 779			return -EFAULT;
 780	}
 781
 782	batch->nr_ops = 0;
 783	return 0;
 784}
 785
 786static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
 787				 struct gntdev_grant_copy_segment *seg,
 788				 s16 __user *status)
 789{
 790	uint16_t copied = 0;
 791
 792	/*
 793	 * Disallow local -> local copies since there is only space in
 794	 * batch->pages for one page per-op and this would be a very
 795	 * expensive memcpy().
 796	 */
 797	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
 798		return -EINVAL;
 799
 800	/* Can't cross page if source/dest is a grant ref. */
 801	if (seg->flags & GNTCOPY_source_gref) {
 802		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
 803			return -EINVAL;
 804	}
 805	if (seg->flags & GNTCOPY_dest_gref) {
 806		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
 807			return -EINVAL;
 808	}
 809
 810	if (put_user(GNTST_okay, status))
 811		return -EFAULT;
 812
 813	while (copied < seg->len) {
 814		struct gnttab_copy *op;
 815		void __user *virt;
 816		size_t len, off;
 817		unsigned long gfn;
 818		int ret;
 819
 820		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
 821			ret = gntdev_copy(batch);
 822			if (ret < 0)
 823				return ret;
 824		}
 825
 826		len = seg->len - copied;
 827
 828		op = &batch->ops[batch->nr_ops];
 829		op->flags = 0;
 830
 831		if (seg->flags & GNTCOPY_source_gref) {
 832			op->source.u.ref = seg->source.foreign.ref;
 833			op->source.domid = seg->source.foreign.domid;
 834			op->source.offset = seg->source.foreign.offset + copied;
 835			op->flags |= GNTCOPY_source_gref;
 836		} else {
 837			virt = seg->source.virt + copied;
 838			off = (unsigned long)virt & ~XEN_PAGE_MASK;
 839			len = min(len, (size_t)XEN_PAGE_SIZE - off);
 
 840
 841			ret = gntdev_get_page(batch, virt, false, &gfn);
 842			if (ret < 0)
 843				return ret;
 844
 845			op->source.u.gmfn = gfn;
 846			op->source.domid = DOMID_SELF;
 847			op->source.offset = off;
 848		}
 849
 850		if (seg->flags & GNTCOPY_dest_gref) {
 851			op->dest.u.ref = seg->dest.foreign.ref;
 852			op->dest.domid = seg->dest.foreign.domid;
 853			op->dest.offset = seg->dest.foreign.offset + copied;
 854			op->flags |= GNTCOPY_dest_gref;
 855		} else {
 856			virt = seg->dest.virt + copied;
 857			off = (unsigned long)virt & ~XEN_PAGE_MASK;
 858			len = min(len, (size_t)XEN_PAGE_SIZE - off);
 
 859
 860			ret = gntdev_get_page(batch, virt, true, &gfn);
 861			if (ret < 0)
 862				return ret;
 863
 864			op->dest.u.gmfn = gfn;
 865			op->dest.domid = DOMID_SELF;
 866			op->dest.offset = off;
 867		}
 868
 869		op->len = len;
 870		copied += len;
 871
 872		batch->status[batch->nr_ops] = status;
 873		batch->nr_ops++;
 874	}
 875
 876	return 0;
 877}
 878
 879static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
 880{
 881	struct ioctl_gntdev_grant_copy copy;
 882	struct gntdev_copy_batch batch;
 883	unsigned int i;
 884	int ret = 0;
 885
 886	if (copy_from_user(&copy, u, sizeof(copy)))
 887		return -EFAULT;
 888
 889	batch.nr_ops = 0;
 890	batch.nr_pages = 0;
 891
 892	for (i = 0; i < copy.count; i++) {
 893		struct gntdev_grant_copy_segment seg;
 894
 895		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
 896			ret = -EFAULT;
 897			goto out;
 898		}
 899
 900		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
 901		if (ret < 0)
 902			goto out;
 903
 904		cond_resched();
 905	}
 906	if (batch.nr_ops)
 907		ret = gntdev_copy(&batch);
 908	return ret;
 909
 910  out:
 911	gntdev_put_pages(&batch);
 912	return ret;
 913}
 914
 915static long gntdev_ioctl(struct file *flip,
 916			 unsigned int cmd, unsigned long arg)
 917{
 918	struct gntdev_priv *priv = flip->private_data;
 919	void __user *ptr = (void __user *)arg;
 920
 921	switch (cmd) {
 922	case IOCTL_GNTDEV_MAP_GRANT_REF:
 923		return gntdev_ioctl_map_grant_ref(priv, ptr);
 924
 925	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
 926		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
 927
 928	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
 929		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
 930
 931	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
 932		return gntdev_ioctl_notify(priv, ptr);
 933
 934	case IOCTL_GNTDEV_GRANT_COPY:
 935		return gntdev_ioctl_grant_copy(priv, ptr);
 936
 937#ifdef CONFIG_XEN_GNTDEV_DMABUF
 938	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
 939		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
 940
 941	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
 942		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
 943
 944	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
 945		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
 946
 947	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
 948		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
 949#endif
 950
 951	default:
 952		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
 953		return -ENOIOCTLCMD;
 954	}
 955
 956	return 0;
 957}
 958
 959static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
 960{
 961	struct gntdev_priv *priv = flip->private_data;
 962	int index = vma->vm_pgoff;
 963	int count = vma_pages(vma);
 964	struct gntdev_grant_map *map;
 965	int err = -EINVAL;
 966
 967	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
 968		return -EINVAL;
 969
 970	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
 971			index, count, vma->vm_start, vma->vm_pgoff);
 972
 973	mutex_lock(&priv->lock);
 974	map = gntdev_find_map_index(priv, index, count);
 975	if (!map)
 976		goto unlock_out;
 977	if (use_ptemod && map->vma)
 978		goto unlock_out;
 979	refcount_inc(&map->users);
 980
 981	vma->vm_ops = &gntdev_vmops;
 982
 983	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
 984
 985	if (use_ptemod)
 986		vma->vm_flags |= VM_DONTCOPY;
 987
 988	vma->vm_private_data = map;
 989	if (map->flags) {
 990		if ((vma->vm_flags & VM_WRITE) &&
 991				(map->flags & GNTMAP_readonly))
 992			goto out_unlock_put;
 993	} else {
 994		map->flags = GNTMAP_host_map;
 995		if (!(vma->vm_flags & VM_WRITE))
 996			map->flags |= GNTMAP_readonly;
 997	}
 998
 999	if (use_ptemod) {
1000		map->vma = vma;
1001		err = mmu_interval_notifier_insert_locked(
1002			&map->notifier, vma->vm_mm, vma->vm_start,
1003			vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1004		if (err)
 
1005			goto out_unlock_put;
 
1006	}
1007	mutex_unlock(&priv->lock);
1008
1009	if (use_ptemod) {
1010		/*
1011		 * gntdev takes the address of the PTE in find_grant_ptes() and
1012		 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1013		 * purpose of the notifier is to prevent the hypervisor pointer
1014		 * to the PTE from going stale.
1015		 *
1016		 * Since this vma's mappings can't be touched without the
1017		 * mmap_lock, and we are holding it now, there is no need for
1018		 * the notifier_range locking pattern.
1019		 */
1020		mmu_interval_read_begin(&map->notifier);
1021
1022		map->pages_vm_start = vma->vm_start;
1023		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1024					  vma->vm_end - vma->vm_start,
1025					  find_grant_ptes, map);
1026		if (err) {
1027			pr_warn("find_grant_ptes() failure.\n");
1028			goto out_put_map;
1029		}
1030	}
1031
1032	err = gntdev_map_grant_pages(map);
1033	if (err)
1034		goto out_put_map;
1035
1036	if (!use_ptemod) {
1037		err = vm_map_pages_zero(vma, map->pages, map->count);
1038		if (err)
1039			goto out_put_map;
1040	} else {
1041#ifdef CONFIG_X86
1042		/*
1043		 * If the PTEs were not made special by the grant map
1044		 * hypercall, do so here.
1045		 *
1046		 * This is racy since the mapping is already visible
1047		 * to userspace but userspace should be well-behaved
1048		 * enough to not touch it until the mmap() call
1049		 * returns.
1050		 */
1051		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1052			apply_to_page_range(vma->vm_mm, vma->vm_start,
1053					    vma->vm_end - vma->vm_start,
1054					    set_grant_ptes_as_special, NULL);
1055		}
1056#endif
1057	}
1058
1059	return 0;
1060
1061unlock_out:
1062	mutex_unlock(&priv->lock);
1063	return err;
1064
1065out_unlock_put:
1066	mutex_unlock(&priv->lock);
1067out_put_map:
1068	if (use_ptemod) {
1069		unmap_grant_pages(map, 0, map->count);
1070		if (map->vma) {
1071			mmu_interval_notifier_remove(&map->notifier);
1072			map->vma = NULL;
1073		}
1074	}
1075	gntdev_put_map(priv, map);
1076	return err;
1077}
1078
1079static const struct file_operations gntdev_fops = {
1080	.owner = THIS_MODULE,
1081	.open = gntdev_open,
1082	.release = gntdev_release,
1083	.mmap = gntdev_mmap,
1084	.unlocked_ioctl = gntdev_ioctl
1085};
1086
1087static struct miscdevice gntdev_miscdev = {
1088	.minor        = MISC_DYNAMIC_MINOR,
1089	.name         = "xen/gntdev",
1090	.fops         = &gntdev_fops,
1091};
1092
1093/* ------------------------------------------------------------------ */
1094
1095static int __init gntdev_init(void)
1096{
1097	int err;
1098
1099	if (!xen_domain())
1100		return -ENODEV;
1101
1102	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1103
1104	err = misc_register(&gntdev_miscdev);
1105	if (err != 0) {
1106		pr_err("Could not register gntdev device\n");
1107		return err;
1108	}
1109	return 0;
1110}
1111
1112static void __exit gntdev_exit(void)
1113{
1114	misc_deregister(&gntdev_miscdev);
1115}
1116
1117module_init(gntdev_init);
1118module_exit(gntdev_exit);
1119
1120/* ------------------------------------------------------------------ */
v5.14.15
   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
  39#include <xen/xen.h>
  40#include <xen/grant_table.h>
  41#include <xen/balloon.h>
  42#include <xen/gntdev.h>
  43#include <xen/events.h>
  44#include <xen/page.h>
  45#include <asm/xen/hypervisor.h>
  46#include <asm/xen/hypercall.h>
  47
  48#include "gntdev-common.h"
  49#ifdef CONFIG_XEN_GNTDEV_DMABUF
  50#include "gntdev-dmabuf.h"
  51#endif
  52
  53MODULE_LICENSE("GPL");
  54MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
  55	      "Gerd Hoffmann <kraxel@redhat.com>");
  56MODULE_DESCRIPTION("User-space granted page access driver");
  57
  58static unsigned int limit = 64*1024;
  59module_param(limit, uint, 0644);
  60MODULE_PARM_DESC(limit,
  61	"Maximum number of grants that may be mapped by one mapping request");
  62
  63static int use_ptemod;
  64
  65static int unmap_grant_pages(struct gntdev_grant_map *map,
  66			     int offset, int pages);
  67
  68static struct miscdevice gntdev_miscdev;
  69
  70/* ------------------------------------------------------------------ */
  71
  72bool gntdev_test_page_count(unsigned int count)
  73{
  74	return !count || count > limit;
  75}
  76
  77static void gntdev_print_maps(struct gntdev_priv *priv,
  78			      char *text, int text_index)
  79{
  80#ifdef DEBUG
  81	struct gntdev_grant_map *map;
  82
  83	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
  84	list_for_each_entry(map, &priv->maps, next)
  85		pr_debug("  index %2d, count %2d %s\n",
  86		       map->index, map->count,
  87		       map->index == text_index && text ? text : "");
  88#endif
  89}
  90
  91static void gntdev_free_map(struct gntdev_grant_map *map)
  92{
  93	if (map == NULL)
  94		return;
  95
  96#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
  97	if (map->dma_vaddr) {
  98		struct gnttab_dma_alloc_args args;
  99
 100		args.dev = map->dma_dev;
 101		args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
 102		args.nr_pages = map->count;
 103		args.pages = map->pages;
 104		args.frames = map->frames;
 105		args.vaddr = map->dma_vaddr;
 106		args.dev_bus_addr = map->dma_bus_addr;
 107
 108		gnttab_dma_free_pages(&args);
 109	} else
 110#endif
 111	if (map->pages)
 112		gnttab_free_pages(map->count, map->pages);
 113
 114#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 115	kvfree(map->frames);
 116#endif
 117	kvfree(map->pages);
 118	kvfree(map->grants);
 119	kvfree(map->map_ops);
 120	kvfree(map->unmap_ops);
 121	kvfree(map->kmap_ops);
 122	kvfree(map->kunmap_ops);
 123	kfree(map);
 124}
 125
 126struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
 127					  int dma_flags)
 128{
 129	struct gntdev_grant_map *add;
 130	int i;
 131
 132	add = kzalloc(sizeof(*add), GFP_KERNEL);
 133	if (NULL == add)
 134		return NULL;
 135
 136	add->grants    = kvmalloc_array(count, sizeof(add->grants[0]),
 137					GFP_KERNEL);
 138	add->map_ops   = kvmalloc_array(count, sizeof(add->map_ops[0]),
 139					GFP_KERNEL);
 140	add->unmap_ops = kvmalloc_array(count, sizeof(add->unmap_ops[0]),
 141					GFP_KERNEL);
 142	add->pages     = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
 143	if (NULL == add->grants    ||
 144	    NULL == add->map_ops   ||
 145	    NULL == add->unmap_ops ||
 
 
 146	    NULL == add->pages)
 147		goto err;
 148	if (use_ptemod) {
 149		add->kmap_ops   = kvmalloc_array(count, sizeof(add->kmap_ops[0]),
 150						 GFP_KERNEL);
 151		add->kunmap_ops = kvmalloc_array(count, sizeof(add->kunmap_ops[0]),
 152						 GFP_KERNEL);
 153		if (NULL == add->kmap_ops || NULL == add->kunmap_ops)
 154			goto err;
 155	}
 156
 157#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 158	add->dma_flags = dma_flags;
 159
 160	/*
 161	 * Check if this mapping is requested to be backed
 162	 * by a DMA buffer.
 163	 */
 164	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
 165		struct gnttab_dma_alloc_args args;
 166
 167		add->frames = kvcalloc(count, sizeof(add->frames[0]),
 168				       GFP_KERNEL);
 169		if (!add->frames)
 170			goto err;
 171
 172		/* Remember the device, so we can free DMA memory. */
 173		add->dma_dev = priv->dma_dev;
 174
 175		args.dev = priv->dma_dev;
 176		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
 177		args.nr_pages = count;
 178		args.pages = add->pages;
 179		args.frames = add->frames;
 180
 181		if (gnttab_dma_alloc_pages(&args))
 182			goto err;
 183
 184		add->dma_vaddr = args.vaddr;
 185		add->dma_bus_addr = args.dev_bus_addr;
 186	} else
 187#endif
 188	if (gnttab_alloc_pages(count, add->pages))
 189		goto err;
 190
 191	for (i = 0; i < count; i++) {
 192		add->grants[i].domid = DOMID_INVALID;
 193		add->grants[i].ref = INVALID_GRANT_REF;
 194		add->map_ops[i].handle = INVALID_GRANT_HANDLE;
 195		add->unmap_ops[i].handle = INVALID_GRANT_HANDLE;
 196		if (use_ptemod) {
 197			add->kmap_ops[i].handle = INVALID_GRANT_HANDLE;
 198			add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE;
 199		}
 200	}
 201
 202	add->index = 0;
 203	add->count = count;
 204	refcount_set(&add->users, 1);
 205
 206	return add;
 207
 208err:
 209	gntdev_free_map(add);
 210	return NULL;
 211}
 212
 213void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
 214{
 215	struct gntdev_grant_map *map;
 216
 217	list_for_each_entry(map, &priv->maps, next) {
 218		if (add->index + add->count < map->index) {
 219			list_add_tail(&add->next, &map->next);
 220			goto done;
 221		}
 222		add->index = map->index + map->count;
 223	}
 224	list_add_tail(&add->next, &priv->maps);
 225
 226done:
 227	gntdev_print_maps(priv, "[new]", add->index);
 228}
 229
 230static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
 231						      int index, int count)
 232{
 233	struct gntdev_grant_map *map;
 234
 235	list_for_each_entry(map, &priv->maps, next) {
 236		if (map->index != index)
 237			continue;
 238		if (count && map->count != count)
 239			continue;
 240		return map;
 241	}
 242	return NULL;
 243}
 244
 245void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
 246{
 247	if (!map)
 248		return;
 249
 250	if (!refcount_dec_and_test(&map->users))
 251		return;
 252
 253	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
 254		notify_remote_via_evtchn(map->notify.event);
 255		evtchn_put(map->notify.event);
 256	}
 257
 258	if (map->pages && !use_ptemod)
 259		unmap_grant_pages(map, 0, map->count);
 260	gntdev_free_map(map);
 261}
 262
 263/* ------------------------------------------------------------------ */
 264
 265static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
 266{
 267	struct gntdev_grant_map *map = data;
 268	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
 269	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
 270	u64 pte_maddr;
 271
 272	BUG_ON(pgnr >= map->count);
 273	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
 274
 275	/*
 276	 * Set the PTE as special to force get_user_pages_fast() fall
 277	 * back to the slow path.  If this is not supported as part of
 278	 * the grant map, it will be done afterwards.
 279	 */
 280	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
 281		flags |= (1 << _GNTMAP_guest_avail0);
 282
 283	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
 284			  map->grants[pgnr].ref,
 285			  map->grants[pgnr].domid);
 286	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
 287			    INVALID_GRANT_HANDLE);
 288	return 0;
 289}
 290
 291#ifdef CONFIG_X86
 292static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
 293{
 294	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
 295	return 0;
 296}
 297#endif
 298
 299int gntdev_map_grant_pages(struct gntdev_grant_map *map)
 300{
 301	int i, err = 0;
 302
 303	if (!use_ptemod) {
 304		/* Note: it could already be mapped */
 305		if (map->map_ops[0].handle != INVALID_GRANT_HANDLE)
 306			return 0;
 307		for (i = 0; i < map->count; i++) {
 308			unsigned long addr = (unsigned long)
 309				pfn_to_kaddr(page_to_pfn(map->pages[i]));
 310			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
 311				map->grants[i].ref,
 312				map->grants[i].domid);
 313			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
 314				map->flags, INVALID_GRANT_HANDLE);
 315		}
 316	} else {
 317		/*
 318		 * Setup the map_ops corresponding to the pte entries pointing
 319		 * to the kernel linear addresses of the struct pages.
 320		 * These ptes are completely different from the user ptes dealt
 321		 * with find_grant_ptes.
 322		 * Note that GNTMAP_device_map isn't needed here: The
 323		 * dev_bus_addr output field gets consumed only from ->map_ops,
 324		 * and by not requesting it when mapping we also avoid needing
 325		 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra
 326		 * reference to the page in the hypervisor).
 327		 */
 328		unsigned int flags = (map->flags & ~GNTMAP_device_map) |
 329				     GNTMAP_host_map;
 330
 331		for (i = 0; i < map->count; i++) {
 332			unsigned long address = (unsigned long)
 333				pfn_to_kaddr(page_to_pfn(map->pages[i]));
 334			BUG_ON(PageHighMem(map->pages[i]));
 335
 336			gnttab_set_map_op(&map->kmap_ops[i], address, flags,
 
 337				map->grants[i].ref,
 338				map->grants[i].domid);
 339			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
 340				flags, INVALID_GRANT_HANDLE);
 341		}
 342	}
 343
 344	pr_debug("map %d+%d\n", map->index, map->count);
 345	err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages,
 346			map->count);
 
 
 347
 348	for (i = 0; i < map->count; i++) {
 349		if (map->map_ops[i].status == GNTST_okay)
 350			map->unmap_ops[i].handle = map->map_ops[i].handle;
 351		else if (!err)
 352			err = -EINVAL;
 
 
 353
 354		if (map->flags & GNTMAP_device_map)
 355			map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr;
 
 
 
 
 356
 357		if (use_ptemod) {
 358			if (map->kmap_ops[i].status == GNTST_okay)
 359				map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
 360			else if (!err)
 361				err = -EINVAL;
 362		}
 
 363	}
 364	return err;
 365}
 366
 367static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
 368			       int pages)
 369{
 370	int i, err = 0;
 371	struct gntab_unmap_queue_data unmap_data;
 372
 373	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
 374		int pgno = (map->notify.addr >> PAGE_SHIFT);
 375		if (pgno >= offset && pgno < offset + pages) {
 376			/* No need for kmap, pages are in lowmem */
 377			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
 378			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
 379			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
 380		}
 381	}
 382
 383	unmap_data.unmap_ops = map->unmap_ops + offset;
 384	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
 385	unmap_data.pages = map->pages + offset;
 386	unmap_data.count = pages;
 387
 388	err = gnttab_unmap_refs_sync(&unmap_data);
 389	if (err)
 390		return err;
 391
 392	for (i = 0; i < pages; i++) {
 393		if (map->unmap_ops[offset+i].status)
 394			err = -EINVAL;
 395		pr_debug("unmap handle=%d st=%d\n",
 396			map->unmap_ops[offset+i].handle,
 397			map->unmap_ops[offset+i].status);
 398		map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
 399		if (use_ptemod) {
 400			if (map->kunmap_ops[offset+i].status)
 401				err = -EINVAL;
 402			pr_debug("kunmap handle=%u st=%d\n",
 403				 map->kunmap_ops[offset+i].handle,
 404				 map->kunmap_ops[offset+i].status);
 405			map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
 406		}
 407	}
 408	return err;
 409}
 410
 411static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
 412			     int pages)
 413{
 414	int range, err = 0;
 415
 416	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
 417
 418	/* It is possible the requested range will have a "hole" where we
 419	 * already unmapped some of the grants. Only unmap valid ranges.
 420	 */
 421	while (pages && !err) {
 422		while (pages &&
 423		       map->unmap_ops[offset].handle == INVALID_GRANT_HANDLE) {
 424			offset++;
 425			pages--;
 426		}
 427		range = 0;
 428		while (range < pages) {
 429			if (map->unmap_ops[offset + range].handle ==
 430			    INVALID_GRANT_HANDLE)
 431				break;
 432			range++;
 433		}
 434		err = __unmap_grant_pages(map, offset, range);
 435		offset += range;
 436		pages -= range;
 437	}
 438
 439	return err;
 440}
 441
 442/* ------------------------------------------------------------------ */
 443
 444static void gntdev_vma_open(struct vm_area_struct *vma)
 445{
 446	struct gntdev_grant_map *map = vma->vm_private_data;
 447
 448	pr_debug("gntdev_vma_open %p\n", vma);
 449	refcount_inc(&map->users);
 450}
 451
 452static void gntdev_vma_close(struct vm_area_struct *vma)
 453{
 454	struct gntdev_grant_map *map = vma->vm_private_data;
 455	struct file *file = vma->vm_file;
 456	struct gntdev_priv *priv = file->private_data;
 457
 458	pr_debug("gntdev_vma_close %p\n", vma);
 459	if (use_ptemod) {
 460		WARN_ON(map->vma != vma);
 461		mmu_interval_notifier_remove(&map->notifier);
 462		map->vma = NULL;
 463	}
 464	vma->vm_private_data = NULL;
 465	gntdev_put_map(priv, map);
 466}
 467
 468static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
 469						 unsigned long addr)
 470{
 471	struct gntdev_grant_map *map = vma->vm_private_data;
 472
 473	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
 474}
 475
 476static const struct vm_operations_struct gntdev_vmops = {
 477	.open = gntdev_vma_open,
 478	.close = gntdev_vma_close,
 479	.find_special_page = gntdev_vma_find_special_page,
 480};
 481
 482/* ------------------------------------------------------------------ */
 483
 484static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
 485			      const struct mmu_notifier_range *range,
 486			      unsigned long cur_seq)
 487{
 488	struct gntdev_grant_map *map =
 489		container_of(mn, struct gntdev_grant_map, notifier);
 490	unsigned long mstart, mend;
 491	int err;
 492
 493	if (!mmu_notifier_range_blockable(range))
 494		return false;
 495
 496	/*
 497	 * If the VMA is split or otherwise changed the notifier is not
 498	 * updated, but we don't want to process VA's outside the modified
 499	 * VMA. FIXME: It would be much more understandable to just prevent
 500	 * modifying the VMA in the first place.
 501	 */
 502	if (map->vma->vm_start >= range->end ||
 503	    map->vma->vm_end <= range->start)
 504		return true;
 505
 506	mstart = max(range->start, map->vma->vm_start);
 507	mend = min(range->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			range->start, range->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 true;
 518}
 519
 520static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
 521	.invalidate = gntdev_invalidate,
 522};
 523
 524/* ------------------------------------------------------------------ */
 525
 526static int gntdev_open(struct inode *inode, struct file *flip)
 527{
 528	struct gntdev_priv *priv;
 529
 530	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 531	if (!priv)
 532		return -ENOMEM;
 533
 534	INIT_LIST_HEAD(&priv->maps);
 535	mutex_init(&priv->lock);
 536
 537#ifdef CONFIG_XEN_GNTDEV_DMABUF
 538	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
 539	if (IS_ERR(priv->dmabuf_priv)) {
 540		int ret = PTR_ERR(priv->dmabuf_priv);
 541
 542		kfree(priv);
 543		return ret;
 544	}
 545#endif
 546
 547	flip->private_data = priv;
 548#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
 549	priv->dma_dev = gntdev_miscdev.this_device;
 550	dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
 551#endif
 552	pr_debug("priv %p\n", priv);
 553
 554	return 0;
 555}
 556
 557static int gntdev_release(struct inode *inode, struct file *flip)
 558{
 559	struct gntdev_priv *priv = flip->private_data;
 560	struct gntdev_grant_map *map;
 561
 562	pr_debug("priv %p\n", priv);
 563
 564	mutex_lock(&priv->lock);
 565	while (!list_empty(&priv->maps)) {
 566		map = list_entry(priv->maps.next,
 567				 struct gntdev_grant_map, next);
 568		list_del(&map->next);
 569		gntdev_put_map(NULL /* already removed */, map);
 570	}
 571	mutex_unlock(&priv->lock);
 572
 573#ifdef CONFIG_XEN_GNTDEV_DMABUF
 574	gntdev_dmabuf_fini(priv->dmabuf_priv);
 575#endif
 576
 577	kfree(priv);
 578	return 0;
 579}
 580
 581static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
 582				       struct ioctl_gntdev_map_grant_ref __user *u)
 583{
 584	struct ioctl_gntdev_map_grant_ref op;
 585	struct gntdev_grant_map *map;
 586	int err;
 587
 588	if (copy_from_user(&op, u, sizeof(op)) != 0)
 589		return -EFAULT;
 590	pr_debug("priv %p, add %d\n", priv, op.count);
 591	if (unlikely(gntdev_test_page_count(op.count)))
 592		return -EINVAL;
 593
 594	err = -ENOMEM;
 595	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
 596	if (!map)
 597		return err;
 598
 599	if (copy_from_user(map->grants, &u->refs,
 600			   sizeof(map->grants[0]) * op.count) != 0) {
 601		gntdev_put_map(NULL, map);
 602		return -EFAULT;
 603	}
 604
 605	mutex_lock(&priv->lock);
 606	gntdev_add_map(priv, map);
 607	op.index = map->index << PAGE_SHIFT;
 608	mutex_unlock(&priv->lock);
 609
 610	if (copy_to_user(u, &op, sizeof(op)) != 0)
 611		return -EFAULT;
 612
 613	return 0;
 614}
 615
 616static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
 617					 struct ioctl_gntdev_unmap_grant_ref __user *u)
 618{
 619	struct ioctl_gntdev_unmap_grant_ref op;
 620	struct gntdev_grant_map *map;
 621	int err = -ENOENT;
 622
 623	if (copy_from_user(&op, u, sizeof(op)) != 0)
 624		return -EFAULT;
 625	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
 626
 627	mutex_lock(&priv->lock);
 628	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
 629	if (map) {
 630		list_del(&map->next);
 631		err = 0;
 632	}
 633	mutex_unlock(&priv->lock);
 634	if (map)
 635		gntdev_put_map(priv, map);
 636	return err;
 637}
 638
 639static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
 640					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
 641{
 642	struct ioctl_gntdev_get_offset_for_vaddr op;
 643	struct vm_area_struct *vma;
 644	struct gntdev_grant_map *map;
 645	int rv = -EINVAL;
 646
 647	if (copy_from_user(&op, u, sizeof(op)) != 0)
 648		return -EFAULT;
 649	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
 650
 651	mmap_read_lock(current->mm);
 652	vma = find_vma(current->mm, op.vaddr);
 653	if (!vma || vma->vm_ops != &gntdev_vmops)
 654		goto out_unlock;
 655
 656	map = vma->vm_private_data;
 657	if (!map)
 658		goto out_unlock;
 659
 660	op.offset = map->index << PAGE_SHIFT;
 661	op.count = map->count;
 662	rv = 0;
 663
 664 out_unlock:
 665	mmap_read_unlock(current->mm);
 666
 667	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
 668		return -EFAULT;
 669	return rv;
 670}
 671
 672static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
 673{
 674	struct ioctl_gntdev_unmap_notify op;
 675	struct gntdev_grant_map *map;
 676	int rc;
 677	int out_flags;
 678	evtchn_port_t out_event;
 679
 680	if (copy_from_user(&op, u, sizeof(op)))
 681		return -EFAULT;
 682
 683	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
 684		return -EINVAL;
 685
 686	/* We need to grab a reference to the event channel we are going to use
 687	 * to send the notify before releasing the reference we may already have
 688	 * (if someone has called this ioctl twice). This is required so that
 689	 * it is possible to change the clear_byte part of the notification
 690	 * without disturbing the event channel part, which may now be the last
 691	 * reference to that event channel.
 692	 */
 693	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
 694		if (evtchn_get(op.event_channel_port))
 695			return -EINVAL;
 696	}
 697
 698	out_flags = op.action;
 699	out_event = op.event_channel_port;
 700
 701	mutex_lock(&priv->lock);
 702
 703	list_for_each_entry(map, &priv->maps, next) {
 704		uint64_t begin = map->index << PAGE_SHIFT;
 705		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
 706		if (op.index >= begin && op.index < end)
 707			goto found;
 708	}
 709	rc = -ENOENT;
 710	goto unlock_out;
 711
 712 found:
 713	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
 714			(map->flags & GNTMAP_readonly)) {
 715		rc = -EINVAL;
 716		goto unlock_out;
 717	}
 718
 719	out_flags = map->notify.flags;
 720	out_event = map->notify.event;
 721
 722	map->notify.flags = op.action;
 723	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
 724	map->notify.event = op.event_channel_port;
 725
 726	rc = 0;
 727
 728 unlock_out:
 729	mutex_unlock(&priv->lock);
 730
 731	/* Drop the reference to the event channel we did not save in the map */
 732	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
 733		evtchn_put(out_event);
 734
 735	return rc;
 736}
 737
 738#define GNTDEV_COPY_BATCH 16
 739
 740struct gntdev_copy_batch {
 741	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
 742	struct page *pages[GNTDEV_COPY_BATCH];
 743	s16 __user *status[GNTDEV_COPY_BATCH];
 744	unsigned int nr_ops;
 745	unsigned int nr_pages;
 746	bool writeable;
 747};
 748
 749static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
 750				unsigned long *gfn)
 751{
 752	unsigned long addr = (unsigned long)virt;
 753	struct page *page;
 754	unsigned long xen_pfn;
 755	int ret;
 756
 757	ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
 758	if (ret < 0)
 759		return ret;
 760
 761	batch->pages[batch->nr_pages++] = page;
 762
 763	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
 764	*gfn = pfn_to_gfn(xen_pfn);
 765
 766	return 0;
 767}
 768
 769static void gntdev_put_pages(struct gntdev_copy_batch *batch)
 770{
 771	unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
 
 
 
 772	batch->nr_pages = 0;
 773	batch->writeable = false;
 774}
 775
 776static int gntdev_copy(struct gntdev_copy_batch *batch)
 777{
 778	unsigned int i;
 779
 780	gnttab_batch_copy(batch->ops, batch->nr_ops);
 781	gntdev_put_pages(batch);
 782
 783	/*
 784	 * For each completed op, update the status if the op failed
 785	 * and all previous ops for the segment were successful.
 786	 */
 787	for (i = 0; i < batch->nr_ops; i++) {
 788		s16 status = batch->ops[i].status;
 789		s16 old_status;
 790
 791		if (status == GNTST_okay)
 792			continue;
 793
 794		if (__get_user(old_status, batch->status[i]))
 795			return -EFAULT;
 796
 797		if (old_status != GNTST_okay)
 798			continue;
 799
 800		if (__put_user(status, batch->status[i]))
 801			return -EFAULT;
 802	}
 803
 804	batch->nr_ops = 0;
 805	return 0;
 806}
 807
 808static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
 809				 struct gntdev_grant_copy_segment *seg,
 810				 s16 __user *status)
 811{
 812	uint16_t copied = 0;
 813
 814	/*
 815	 * Disallow local -> local copies since there is only space in
 816	 * batch->pages for one page per-op and this would be a very
 817	 * expensive memcpy().
 818	 */
 819	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
 820		return -EINVAL;
 821
 822	/* Can't cross page if source/dest is a grant ref. */
 823	if (seg->flags & GNTCOPY_source_gref) {
 824		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
 825			return -EINVAL;
 826	}
 827	if (seg->flags & GNTCOPY_dest_gref) {
 828		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
 829			return -EINVAL;
 830	}
 831
 832	if (put_user(GNTST_okay, status))
 833		return -EFAULT;
 834
 835	while (copied < seg->len) {
 836		struct gnttab_copy *op;
 837		void __user *virt;
 838		size_t len, off;
 839		unsigned long gfn;
 840		int ret;
 841
 842		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
 843			ret = gntdev_copy(batch);
 844			if (ret < 0)
 845				return ret;
 846		}
 847
 848		len = seg->len - copied;
 849
 850		op = &batch->ops[batch->nr_ops];
 851		op->flags = 0;
 852
 853		if (seg->flags & GNTCOPY_source_gref) {
 854			op->source.u.ref = seg->source.foreign.ref;
 855			op->source.domid = seg->source.foreign.domid;
 856			op->source.offset = seg->source.foreign.offset + copied;
 857			op->flags |= GNTCOPY_source_gref;
 858		} else {
 859			virt = seg->source.virt + copied;
 860			off = (unsigned long)virt & ~XEN_PAGE_MASK;
 861			len = min(len, (size_t)XEN_PAGE_SIZE - off);
 862			batch->writeable = false;
 863
 864			ret = gntdev_get_page(batch, virt, &gfn);
 865			if (ret < 0)
 866				return ret;
 867
 868			op->source.u.gmfn = gfn;
 869			op->source.domid = DOMID_SELF;
 870			op->source.offset = off;
 871		}
 872
 873		if (seg->flags & GNTCOPY_dest_gref) {
 874			op->dest.u.ref = seg->dest.foreign.ref;
 875			op->dest.domid = seg->dest.foreign.domid;
 876			op->dest.offset = seg->dest.foreign.offset + copied;
 877			op->flags |= GNTCOPY_dest_gref;
 878		} else {
 879			virt = seg->dest.virt + copied;
 880			off = (unsigned long)virt & ~XEN_PAGE_MASK;
 881			len = min(len, (size_t)XEN_PAGE_SIZE - off);
 882			batch->writeable = true;
 883
 884			ret = gntdev_get_page(batch, virt, &gfn);
 885			if (ret < 0)
 886				return ret;
 887
 888			op->dest.u.gmfn = gfn;
 889			op->dest.domid = DOMID_SELF;
 890			op->dest.offset = off;
 891		}
 892
 893		op->len = len;
 894		copied += len;
 895
 896		batch->status[batch->nr_ops] = status;
 897		batch->nr_ops++;
 898	}
 899
 900	return 0;
 901}
 902
 903static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
 904{
 905	struct ioctl_gntdev_grant_copy copy;
 906	struct gntdev_copy_batch batch;
 907	unsigned int i;
 908	int ret = 0;
 909
 910	if (copy_from_user(&copy, u, sizeof(copy)))
 911		return -EFAULT;
 912
 913	batch.nr_ops = 0;
 914	batch.nr_pages = 0;
 915
 916	for (i = 0; i < copy.count; i++) {
 917		struct gntdev_grant_copy_segment seg;
 918
 919		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
 920			ret = -EFAULT;
 921			goto out;
 922		}
 923
 924		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
 925		if (ret < 0)
 926			goto out;
 927
 928		cond_resched();
 929	}
 930	if (batch.nr_ops)
 931		ret = gntdev_copy(&batch);
 932	return ret;
 933
 934  out:
 935	gntdev_put_pages(&batch);
 936	return ret;
 937}
 938
 939static long gntdev_ioctl(struct file *flip,
 940			 unsigned int cmd, unsigned long arg)
 941{
 942	struct gntdev_priv *priv = flip->private_data;
 943	void __user *ptr = (void __user *)arg;
 944
 945	switch (cmd) {
 946	case IOCTL_GNTDEV_MAP_GRANT_REF:
 947		return gntdev_ioctl_map_grant_ref(priv, ptr);
 948
 949	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
 950		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
 951
 952	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
 953		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
 954
 955	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
 956		return gntdev_ioctl_notify(priv, ptr);
 957
 958	case IOCTL_GNTDEV_GRANT_COPY:
 959		return gntdev_ioctl_grant_copy(priv, ptr);
 960
 961#ifdef CONFIG_XEN_GNTDEV_DMABUF
 962	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
 963		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
 964
 965	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
 966		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
 967
 968	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
 969		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
 970
 971	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
 972		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
 973#endif
 974
 975	default:
 976		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
 977		return -ENOIOCTLCMD;
 978	}
 979
 980	return 0;
 981}
 982
 983static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
 984{
 985	struct gntdev_priv *priv = flip->private_data;
 986	int index = vma->vm_pgoff;
 987	int count = vma_pages(vma);
 988	struct gntdev_grant_map *map;
 989	int err = -EINVAL;
 990
 991	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
 992		return -EINVAL;
 993
 994	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
 995			index, count, vma->vm_start, vma->vm_pgoff);
 996
 997	mutex_lock(&priv->lock);
 998	map = gntdev_find_map_index(priv, index, count);
 999	if (!map)
1000		goto unlock_out;
1001	if (use_ptemod && map->vma)
1002		goto unlock_out;
1003	refcount_inc(&map->users);
1004
1005	vma->vm_ops = &gntdev_vmops;
1006
1007	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1008
1009	if (use_ptemod)
1010		vma->vm_flags |= VM_DONTCOPY;
1011
1012	vma->vm_private_data = map;
1013	if (map->flags) {
1014		if ((vma->vm_flags & VM_WRITE) &&
1015				(map->flags & GNTMAP_readonly))
1016			goto out_unlock_put;
1017	} else {
1018		map->flags = GNTMAP_host_map;
1019		if (!(vma->vm_flags & VM_WRITE))
1020			map->flags |= GNTMAP_readonly;
1021	}
1022
1023	if (use_ptemod) {
1024		map->vma = vma;
1025		err = mmu_interval_notifier_insert_locked(
1026			&map->notifier, vma->vm_mm, vma->vm_start,
1027			vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1028		if (err) {
1029			map->vma = NULL;
1030			goto out_unlock_put;
1031		}
1032	}
1033	mutex_unlock(&priv->lock);
1034
1035	if (use_ptemod) {
1036		/*
1037		 * gntdev takes the address of the PTE in find_grant_ptes() and
1038		 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1039		 * purpose of the notifier is to prevent the hypervisor pointer
1040		 * to the PTE from going stale.
1041		 *
1042		 * Since this vma's mappings can't be touched without the
1043		 * mmap_lock, and we are holding it now, there is no need for
1044		 * the notifier_range locking pattern.
1045		 */
1046		mmu_interval_read_begin(&map->notifier);
1047
1048		map->pages_vm_start = vma->vm_start;
1049		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1050					  vma->vm_end - vma->vm_start,
1051					  find_grant_ptes, map);
1052		if (err) {
1053			pr_warn("find_grant_ptes() failure.\n");
1054			goto out_put_map;
1055		}
1056	}
1057
1058	err = gntdev_map_grant_pages(map);
1059	if (err)
1060		goto out_put_map;
1061
1062	if (!use_ptemod) {
1063		err = vm_map_pages_zero(vma, map->pages, map->count);
1064		if (err)
1065			goto out_put_map;
1066	} else {
1067#ifdef CONFIG_X86
1068		/*
1069		 * If the PTEs were not made special by the grant map
1070		 * hypercall, do so here.
1071		 *
1072		 * This is racy since the mapping is already visible
1073		 * to userspace but userspace should be well-behaved
1074		 * enough to not touch it until the mmap() call
1075		 * returns.
1076		 */
1077		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1078			apply_to_page_range(vma->vm_mm, vma->vm_start,
1079					    vma->vm_end - vma->vm_start,
1080					    set_grant_ptes_as_special, NULL);
1081		}
1082#endif
1083	}
1084
1085	return 0;
1086
1087unlock_out:
1088	mutex_unlock(&priv->lock);
1089	return err;
1090
1091out_unlock_put:
1092	mutex_unlock(&priv->lock);
1093out_put_map:
1094	if (use_ptemod) {
1095		unmap_grant_pages(map, 0, map->count);
1096		if (map->vma) {
1097			mmu_interval_notifier_remove(&map->notifier);
1098			map->vma = NULL;
1099		}
1100	}
1101	gntdev_put_map(priv, map);
1102	return err;
1103}
1104
1105static const struct file_operations gntdev_fops = {
1106	.owner = THIS_MODULE,
1107	.open = gntdev_open,
1108	.release = gntdev_release,
1109	.mmap = gntdev_mmap,
1110	.unlocked_ioctl = gntdev_ioctl
1111};
1112
1113static struct miscdevice gntdev_miscdev = {
1114	.minor        = MISC_DYNAMIC_MINOR,
1115	.name         = "xen/gntdev",
1116	.fops         = &gntdev_fops,
1117};
1118
1119/* ------------------------------------------------------------------ */
1120
1121static int __init gntdev_init(void)
1122{
1123	int err;
1124
1125	if (!xen_domain())
1126		return -ENODEV;
1127
1128	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1129
1130	err = misc_register(&gntdev_miscdev);
1131	if (err != 0) {
1132		pr_err("Could not register gntdev device\n");
1133		return err;
1134	}
1135	return 0;
1136}
1137
1138static void __exit gntdev_exit(void)
1139{
1140	misc_deregister(&gntdev_miscdev);
1141}
1142
1143module_init(gntdev_init);
1144module_exit(gntdev_exit);
1145
1146/* ------------------------------------------------------------------ */