drivers/xen/gntalloc.c

   1 /******************************************************************************
   2  * gntalloc.c
   3  *
   4  * Device for creating grant references (in user-space) that may be shared
   5  * with other domains.
   6  *
   7  * This program is distributed in the hope that it will be useful,
   8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  10  * GNU General Public License for more details.
  11  *
  12  * You should have received a copy of the GNU General Public License
  13  * along with this program; if not, write to the Free Software
  14  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  15  */
  16
  17 /*
  18  * This driver exists to allow userspace programs in Linux to allocate kernel
  19  * memory that will later be shared with another domain.  Without this device,
  20  * Linux userspace programs cannot create grant references.
  21  *
  22  * How this stuff works:
  23  *   X -> granting a page to Y
  24  *   Y -> mapping the grant from X
  25  *
  26  *   1. X uses the gntalloc device to allocate a page of kernel memory, P.
  27  *   2. X creates an entry in the grant table that says domid(Y) can access P.
  28  *      This is done without a hypercall unless the grant table needs expansion.
  29  *   3. X gives the grant reference identifier, GREF, to Y.
  30  *   4. Y maps the page, either directly into kernel memory for use in a backend
  31  *      driver, or via a the gntdev device to map into the address space of an
  32  *      application running in Y. This is the first point at which Xen does any
  33  *      tracking of the page.
  34  *   5. A program in X mmap()s a segment of the gntalloc device that corresponds
  35  *      to the shared page, and can now communicate with Y over the shared page.
  36  *
  37  *
  38  * NOTE TO USERSPACE LIBRARIES:
  39  *   The grant allocation and mmap()ing are, naturally, two separate operations.
  40  *   You set up the sharing by calling the create ioctl() and then the mmap().
  41  *   Teardown requires munmap() and either close() or ioctl().
  42  *
  43  * WARNING: Since Xen does not allow a guest to forcibly end the use of a grant
  44  * reference, this device can be used to consume kernel memory by leaving grant
  45  * references mapped by another domain when an application exits. Therefore,
  46  * there is a global limit on the number of pages that can be allocated. When
  47  * all references to the page are unmapped, it will be freed during the next
  48  * grant operation.
  49  */
  50
  51 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  52
  53 #include <linux/atomic.h>
  54 #include <linux/module.h>
  55 #include <linux/miscdevice.h>
  56 #include <linux/kernel.h>
  57 #include <linux/init.h>
  58 #include <linux/slab.h>
  59 #include <linux/fs.h>
  60 #include <linux/device.h>
  61 #include <linux/mm.h>
  62 #include <linux/uaccess.h>
  63 #include <linux/types.h>
  64 #include <linux/list.h>
  65 #include <linux/highmem.h>
  66
  67 #include <xen/xen.h>
  68 #include <xen/page.h>
  69 #include <xen/grant_table.h>
  70 #include <xen/gntalloc.h>
  71 #include <xen/events.h>
  72
  73 static int limit = 1024;
  74 module_param(limit, int, 0644);
  75 MODULE_PARM_DESC(limit, "Maximum number of grants that may be allocated by "
  76                 "the gntalloc device");
  77
  78 static LIST_HEAD(gref_list);
  79 static DEFINE_MUTEX(gref_mutex);
  80 static int gref_size;
  81
  82 struct notify_info {
  83         uint16_t pgoff:12;    /* Bits 0-11: Offset of the byte to clear */
  84         uint16_t flags:2;     /* Bits 12-13: Unmap notification flags */
  85         int event;            /* Port (event channel) to notify */
  86 };
  87
  88 /* Metadata on a grant reference. */
  89 struct gntalloc_gref {
  90         struct list_head next_gref;  /* list entry gref_list */
  91         struct list_head next_file;  /* list entry file->list, if open */
  92         struct page *page;           /* The shared page */
  93         uint64_t file_index;         /* File offset for mmap() */
  94         unsigned int users;          /* Use count - when zero, waiting on Xen */
  95         grant_ref_t gref_id;         /* The grant reference number */
  96         struct notify_info notify;   /* Unmap notification */
  97 };
  98
  99 struct gntalloc_file_private_data {
 100         struct list_head list;
 101         uint64_t index;
 102 };
 103
 104 struct gntalloc_vma_private_data {
 105         struct gntalloc_gref *gref;
 106         int users;
 107         int count;
 108 };
 109
 110 static void __del_gref(struct gntalloc_gref *gref);
 111
 112 static void do_cleanup(void)
 113 {
 114         struct gntalloc_gref *gref, *n;
 115         list_for_each_entry_safe(gref, n, &gref_list, next_gref) {
 116                 if (!gref->users)
 117                         __del_gref(gref);
 118         }
 119 }
 120
 121 static int add_grefs(struct ioctl_gntalloc_alloc_gref *op,
 122         uint32_t *gref_ids, struct gntalloc_file_private_data *priv)
 123 {
 124         int i, rc, readonly;
 125         LIST_HEAD(queue_gref);
 126         LIST_HEAD(queue_file);
 127         struct gntalloc_gref *gref;
 128
 129         readonly = !(op->flags & GNTALLOC_FLAG_WRITABLE);
 130         rc = -ENOMEM;
 131         for (i = 0; i < op->count; i++) {
 132                 gref = kzalloc(sizeof(*gref), GFP_KERNEL);
 133                 if (!gref)
 134                         goto undo;
 135                 list_add_tail(&gref->next_gref, &queue_gref);
 136                 list_add_tail(&gref->next_file, &queue_file);
 137                 gref->users = 1;
 138                 gref->file_index = op->index + i * PAGE_SIZE;
 139                 gref->page = alloc_page(GFP_KERNEL|__GFP_ZERO);
 140                 if (!gref->page)
 141                         goto undo;
 142
 143                 /* Grant foreign access to the page. */
 144                 gref->gref_id = gnttab_grant_foreign_access(op->domid,
 145                         pfn_to_mfn(page_to_pfn(gref->page)), readonly);
 146                 if ((int)gref->gref_id < 0) {
 147                         rc = gref->gref_id;
 148                         goto undo;
 149                 }
 150                 gref_ids[i] = gref->gref_id;
 151         }
 152
 153         /* Add to gref lists. */
 154         mutex_lock(&gref_mutex);
 155         list_splice_tail(&queue_gref, &gref_list);
 156         list_splice_tail(&queue_file, &priv->list);
 157         mutex_unlock(&gref_mutex);
 158
 159         return 0;
 160
 161 undo:
 162         mutex_lock(&gref_mutex);
 163         gref_size -= (op->count - i);
 164
 165         list_for_each_entry(gref, &queue_file, next_file) {
 166                 /* __del_gref does not remove from queue_file */
 167                 __del_gref(gref);
 168         }
 169
 170         /* It's possible for the target domain to map the just-allocated grant
 171          * references by blindly guessing their IDs; if this is done, then
 172          * __del_gref will leave them in the queue_gref list. They need to be
 173          * added to the global list so that we can free them when they are no
 174          * longer referenced.
 175          */
 176         if (unlikely(!list_empty(&queue_gref)))
 177                 list_splice_tail(&queue_gref, &gref_list);
 178         mutex_unlock(&gref_mutex);
 179         return rc;
 180 }
 181
 182 static void __del_gref(struct gntalloc_gref *gref)
 183 {
 184         if (gref->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
 185                 uint8_t *tmp = kmap(gref->page);
 186                 tmp[gref->notify.pgoff] = 0;
 187                 kunmap(gref->page);
 188         }
 189         if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
 190                 notify_remote_via_evtchn(gref->notify.event);
 191                 evtchn_put(gref->notify.event);
 192         }
 193
 194         gref->notify.flags = 0;
 195
 196         if (gref->gref_id > 0) {
 197                 if (gnttab_query_foreign_access(gref->gref_id))
 198                         return;
 199
 200                 if (!gnttab_end_foreign_access_ref(gref->gref_id, 0))
 201                         return;
 202
 203                 gnttab_free_grant_reference(gref->gref_id);
 204         }
 205
 206         gref_size--;
 207         list_del(&gref->next_gref);
 208
 209         if (gref->page)
 210                 __free_page(gref->page);
 211
 212         kfree(gref);
 213 }
 214
 215 /* finds contiguous grant references in a file, returns the first */
 216 static struct gntalloc_gref *find_grefs(struct gntalloc_file_private_data *priv,
 217                 uint64_t index, uint32_t count)
 218 {
 219         struct gntalloc_gref *rv = NULL, *gref;
 220         list_for_each_entry(gref, &priv->list, next_file) {
 221                 if (gref->file_index == index && !rv)
 222                         rv = gref;
 223                 if (rv) {
 224                         if (gref->file_index != index)
 225                                 return NULL;
 226                         index += PAGE_SIZE;
 227                         count--;
 228                         if (count == 0)
 229                                 return rv;
 230                 }
 231         }
 232         return NULL;
 233 }
 234
 235 /*
 236  * -------------------------------------
 237  *  File operations.
 238  * -------------------------------------
 239  */
 240 static int gntalloc_open(struct inode *inode, struct file *filp)
 241 {
 242         struct gntalloc_file_private_data *priv;
 243
 244         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 245         if (!priv)
 246                 goto out_nomem;
 247         INIT_LIST_HEAD(&priv->list);
 248
 249         filp->private_data = priv;
 250
 251         pr_debug("%s: priv %p\n", __func__, priv);
 252
 253         return 0;
 254
 255 out_nomem:
 256         return -ENOMEM;
 257 }
 258
 259 static int gntalloc_release(struct inode *inode, struct file *filp)
 260 {
 261         struct gntalloc_file_private_data *priv = filp->private_data;
 262         struct gntalloc_gref *gref;
 263
 264         pr_debug("%s: priv %p\n", __func__, priv);
 265
 266         mutex_lock(&gref_mutex);
 267         while (!list_empty(&priv->list)) {
 268                 gref = list_entry(priv->list.next,
 269                         struct gntalloc_gref, next_file);
 270                 list_del(&gref->next_file);
 271                 gref->users--;
 272                 if (gref->users == 0)
 273                         __del_gref(gref);
 274         }
 275         kfree(priv);
 276         mutex_unlock(&gref_mutex);
 277
 278         return 0;
 279 }
 280
 281 static long gntalloc_ioctl_alloc(struct gntalloc_file_private_data *priv,
 282                 struct ioctl_gntalloc_alloc_gref __user *arg)
 283 {
 284         int rc = 0;
 285         struct ioctl_gntalloc_alloc_gref op;
 286         uint32_t *gref_ids;
 287
 288         pr_debug("%s: priv %p\n", __func__, priv);
 289
 290         if (copy_from_user(&op, arg, sizeof(op))) {
 291                 rc = -EFAULT;
 292                 goto out;
 293         }
 294
 295         gref_ids = kcalloc(op.count, sizeof(gref_ids[0]), GFP_TEMPORARY);
 296         if (!gref_ids) {
 297                 rc = -ENOMEM;
 298                 goto out;
 299         }
 300
 301         mutex_lock(&gref_mutex);
 302         /* Clean up pages that were at zero (local) users but were still mapped
 303          * by remote domains. Since those pages count towards the limit that we
 304          * are about to enforce, removing them here is a good idea.
 305          */
 306         do_cleanup();
 307         if (gref_size + op.count > limit) {
 308                 mutex_unlock(&gref_mutex);
 309                 rc = -ENOSPC;
 310                 goto out_free;
 311         }
 312         gref_size += op.count;
 313         op.index = priv->index;
 314         priv->index += op.count * PAGE_SIZE;
 315         mutex_unlock(&gref_mutex);
 316
 317         rc = add_grefs(&op, gref_ids, priv);
 318         if (rc < 0)
 319                 goto out_free;
 320
 321         /* Once we finish add_grefs, it is unsafe to touch the new reference,
 322          * since it is possible for a concurrent ioctl to remove it (by guessing
 323          * its index). If the userspace application doesn't provide valid memory
 324          * to write the IDs to, then it will need to close the file in order to
 325          * release - which it will do by segfaulting when it tries to access the
 326          * IDs to close them.
 327          */
 328         if (copy_to_user(arg, &op, sizeof(op))) {
 329                 rc = -EFAULT;
 330                 goto out_free;
 331         }
 332         if (copy_to_user(arg->gref_ids, gref_ids,
 333                         sizeof(gref_ids[0]) * op.count)) {
 334                 rc = -EFAULT;
 335                 goto out_free;
 336         }
 337
 338 out_free:
 339         kfree(gref_ids);
 340 out:
 341         return rc;
 342 }
 343
 344 static long gntalloc_ioctl_dealloc(struct gntalloc_file_private_data *priv,
 345                 void __user *arg)
 346 {
 347         int i, rc = 0;
 348         struct ioctl_gntalloc_dealloc_gref op;
 349         struct gntalloc_gref *gref, *n;
 350
 351         pr_debug("%s: priv %p\n", __func__, priv);
 352
 353         if (copy_from_user(&op, arg, sizeof(op))) {
 354                 rc = -EFAULT;
 355                 goto dealloc_grant_out;
 356         }
 357
 358         mutex_lock(&gref_mutex);
 359         gref = find_grefs(priv, op.index, op.count);
 360         if (gref) {
 361                 /* Remove from the file list only, and decrease reference count.
 362                  * The later call to do_cleanup() will remove from gref_list and
 363                  * free the memory if the pages aren't mapped anywhere.
 364                  */
 365                 for (i = 0; i < op.count; i++) {
 366                         n = list_entry(gref->next_file.next,
 367                                 struct gntalloc_gref, next_file);
 368                         list_del(&gref->next_file);
 369                         gref->users--;
 370                         gref = n;
 371                 }
 372         } else {
 373                 rc = -EINVAL;
 374         }
 375
 376         do_cleanup();
 377
 378         mutex_unlock(&gref_mutex);
 379 dealloc_grant_out:
 380         return rc;
 381 }
 382
 383 static long gntalloc_ioctl_unmap_notify(struct gntalloc_file_private_data *priv,
 384                 void __user *arg)
 385 {
 386         struct ioctl_gntalloc_unmap_notify op;
 387         struct gntalloc_gref *gref;
 388         uint64_t index;
 389         int pgoff;
 390         int rc;
 391
 392         if (copy_from_user(&op, arg, sizeof(op)))
 393                 return -EFAULT;
 394
 395         index = op.index & ~(PAGE_SIZE - 1);
 396         pgoff = op.index & (PAGE_SIZE - 1);
 397
 398         mutex_lock(&gref_mutex);
 399
 400         gref = find_grefs(priv, index, 1);
 401         if (!gref) {
 402                 rc = -ENOENT;
 403                 goto unlock_out;
 404         }
 405
 406         if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) {
 407                 rc = -EINVAL;
 408                 goto unlock_out;
 409         }
 410
 411         /* We need to grab a reference to the event channel we are going to use
 412          * to send the notify before releasing the reference we may already have
 413          * (if someone has called this ioctl twice). This is required so that
 414          * it is possible to change the clear_byte part of the notification
 415          * without disturbing the event channel part, which may now be the last
 416          * reference to that event channel.
 417          */
 418         if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
 419                 if (evtchn_get(op.event_channel_port)) {
 420                         rc = -EINVAL;
 421                         goto unlock_out;
 422                 }
 423         }
 424
 425         if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT)
 426                 evtchn_put(gref->notify.event);
 427
 428         gref->notify.flags = op.action;
 429         gref->notify.pgoff = pgoff;
 430         gref->notify.event = op.event_channel_port;
 431         rc = 0;
 432
 433  unlock_out:
 434         mutex_unlock(&gref_mutex);
 435         return rc;
 436 }
 437
 438 static long gntalloc_ioctl(struct file *filp, unsigned int cmd,
 439                 unsigned long arg)
 440 {
 441         struct gntalloc_file_private_data *priv = filp->private_data;
 442
 443         switch (cmd) {
 444         case IOCTL_GNTALLOC_ALLOC_GREF:
 445                 return gntalloc_ioctl_alloc(priv, (void __user *)arg);
 446
 447         case IOCTL_GNTALLOC_DEALLOC_GREF:
 448                 return gntalloc_ioctl_dealloc(priv, (void __user *)arg);
 449
 450         case IOCTL_GNTALLOC_SET_UNMAP_NOTIFY:
 451                 return gntalloc_ioctl_unmap_notify(priv, (void __user *)arg);
 452
 453         default:
 454                 return -ENOIOCTLCMD;
 455         }
 456
 457         return 0;
 458 }
 459
 460 static void gntalloc_vma_open(struct vm_area_struct *vma)
 461 {
 462         struct gntalloc_vma_private_data *priv = vma->vm_private_data;
 463
 464         if (!priv)
 465                 return;
 466
 467         mutex_lock(&gref_mutex);
 468         priv->users++;
 469         mutex_unlock(&gref_mutex);
 470 }
 471
 472 static void gntalloc_vma_close(struct vm_area_struct *vma)
 473 {
 474         struct gntalloc_vma_private_data *priv = vma->vm_private_data;
 475         struct gntalloc_gref *gref, *next;
 476         int i;
 477
 478         if (!priv)
 479                 return;
 480
 481         mutex_lock(&gref_mutex);
 482         priv->users--;
 483         if (priv->users == 0) {
 484                 gref = priv->gref;
 485                 for (i = 0; i < priv->count; i++) {
 486                         gref->users--;
 487                         next = list_entry(gref->next_gref.next,
 488                                           struct gntalloc_gref, next_gref);
 489                         if (gref->users == 0)
 490                                 __del_gref(gref);
 491                         gref = next;
 492                 }
 493                 kfree(priv);
 494         }
 495         mutex_unlock(&gref_mutex);
 496 }
 497
 498 static struct vm_operations_struct gntalloc_vmops = {
 499         .open = gntalloc_vma_open,
 500         .close = gntalloc_vma_close,
 501 };
 502
 503 static int gntalloc_mmap(struct file *filp, struct vm_area_struct *vma)
 504 {
 505         struct gntalloc_file_private_data *priv = filp->private_data;
 506         struct gntalloc_vma_private_data *vm_priv;
 507         struct gntalloc_gref *gref;
 508         int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 509         int rv, i;
 510
 511         if (!(vma->vm_flags & VM_SHARED)) {
 512                 pr_err("%s: Mapping must be shared\n", __func__);
 513                 return -EINVAL;
 514         }
 515
 516         vm_priv = kmalloc(sizeof(*vm_priv), GFP_KERNEL);
 517         if (!vm_priv)
 518                 return -ENOMEM;
 519
 520         mutex_lock(&gref_mutex);
 521
 522         pr_debug("%s: priv %p,%p, page %lu+%d\n", __func__,
 523                        priv, vm_priv, vma->vm_pgoff, count);
 524
 525         gref = find_grefs(priv, vma->vm_pgoff << PAGE_SHIFT, count);
 526         if (gref == NULL) {
 527                 rv = -ENOENT;
 528                 pr_debug("%s: Could not find grant reference",
 529                                 __func__);
 530                 kfree(vm_priv);
 531                 goto out_unlock;
 532         }
 533
 534         vm_priv->gref = gref;
 535         vm_priv->users = 1;
 536         vm_priv->count = count;
 537
 538         vma->vm_private_data = vm_priv;
 539
 540         vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
 541
 542         vma->vm_ops = &gntalloc_vmops;
 543
 544         for (i = 0; i < count; i++) {
 545                 gref->users++;
 546                 rv = vm_insert_page(vma, vma->vm_start + i * PAGE_SIZE,
 547                                 gref->page);
 548                 if (rv)
 549                         goto out_unlock;
 550
 551                 gref = list_entry(gref->next_file.next,
 552                                 struct gntalloc_gref, next_file);
 553         }
 554         rv = 0;
 555
 556 out_unlock:
 557         mutex_unlock(&gref_mutex);
 558         return rv;
 559 }
 560
 561 static const struct file_operations gntalloc_fops = {
 562         .owner = THIS_MODULE,
 563         .open = gntalloc_open,
 564         .release = gntalloc_release,
 565         .unlocked_ioctl = gntalloc_ioctl,
 566         .mmap = gntalloc_mmap
 567 };
 568
 569 /*
 570  * -------------------------------------
 571  * Module creation/destruction.
 572  * -------------------------------------
 573  */
 574 static struct miscdevice gntalloc_miscdev = {
 575         .minor  = MISC_DYNAMIC_MINOR,
 576         .name   = "xen/gntalloc",
 577         .fops   = &gntalloc_fops,
 578 };
 579
 580 static int __init gntalloc_init(void)
 581 {
 582         int err;
 583
 584         if (!xen_domain())
 585                 return -ENODEV;
 586
 587         err = misc_register(&gntalloc_miscdev);
 588         if (err != 0) {
 589                 pr_err("Could not register misc gntalloc device\n");
 590                 return err;
 591         }
 592
 593         pr_debug("Created grant allocation device at %d,%d\n",
 594                         MISC_MAJOR, gntalloc_miscdev.minor);
 595
 596         return 0;
 597 }
 598
 599 static void __exit gntalloc_exit(void)
 600 {
 601         misc_deregister(&gntalloc_miscdev);
 602 }
 603
 604 module_init(gntalloc_init);
 605 module_exit(gntalloc_exit);
 606
 607 MODULE_LICENSE("GPL");
 608 MODULE_AUTHOR("Carter Weatherly <carter.weatherly@jhuapl.edu>, "
 609                 "Daniel De Graaf <dgdegra@tycho.nsa.gov>");
 610 MODULE_DESCRIPTION("User-space grant reference allocator driver");