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 73static int limit = 1024; 74module_param(limit, int, 0644); 75MODULE_PARM_DESC(limit, "Maximum number of grants that may be allocated by " 76 "the gntalloc device"); 77 78static LIST_HEAD(gref_list); 79static DEFINE_MUTEX(gref_mutex); 80static int gref_size; 81 82struct 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. */ 89struct 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 99struct gntalloc_file_private_data { 100 struct list_head list; 101 uint64_t index; 102}; 103 104struct gntalloc_vma_private_data { 105 struct gntalloc_gref *gref; 106 int users; 107 int count; 108}; 109 110static void __del_gref(struct gntalloc_gref *gref); 111 112static 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 121static 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, *next; 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 rc = gnttab_grant_foreign_access(op->domid, 145 pfn_to_mfn(page_to_pfn(gref->page)), readonly); 146 if (rc < 0) 147 goto undo; 148 gref_ids[i] = gref->gref_id = rc; 149 } 150 151 /* Add to gref lists. */ 152 mutex_lock(&gref_mutex); 153 list_splice_tail(&queue_gref, &gref_list); 154 list_splice_tail(&queue_file, &priv->list); 155 mutex_unlock(&gref_mutex); 156 157 return 0; 158 159undo: 160 mutex_lock(&gref_mutex); 161 gref_size -= (op->count - i); 162 163 list_for_each_entry_safe(gref, next, &queue_file, next_file) { 164 list_del(&gref->next_file); 165 __del_gref(gref); 166 } 167 168 /* It's possible for the target domain to map the just-allocated grant 169 * references by blindly guessing their IDs; if this is done, then 170 * __del_gref will leave them in the queue_gref list. They need to be 171 * added to the global list so that we can free them when they are no 172 * longer referenced. 173 */ 174 if (unlikely(!list_empty(&queue_gref))) 175 list_splice_tail(&queue_gref, &gref_list); 176 mutex_unlock(&gref_mutex); 177 return rc; 178} 179 180static void __del_gref(struct gntalloc_gref *gref) 181{ 182 if (gref->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) { 183 uint8_t *tmp = kmap(gref->page); 184 tmp[gref->notify.pgoff] = 0; 185 kunmap(gref->page); 186 } 187 if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT) { 188 notify_remote_via_evtchn(gref->notify.event); 189 evtchn_put(gref->notify.event); 190 } 191 192 gref->notify.flags = 0; 193 194 if (gref->gref_id) { 195 if (gnttab_query_foreign_access(gref->gref_id)) 196 return; 197 198 if (!gnttab_end_foreign_access_ref(gref->gref_id, 0)) 199 return; 200 201 gnttab_free_grant_reference(gref->gref_id); 202 } 203 204 gref_size--; 205 list_del(&gref->next_gref); 206 207 if (gref->page) 208 __free_page(gref->page); 209 210 kfree(gref); 211} 212 213/* finds contiguous grant references in a file, returns the first */ 214static struct gntalloc_gref *find_grefs(struct gntalloc_file_private_data *priv, 215 uint64_t index, uint32_t count) 216{ 217 struct gntalloc_gref *rv = NULL, *gref; 218 list_for_each_entry(gref, &priv->list, next_file) { 219 if (gref->file_index == index && !rv) 220 rv = gref; 221 if (rv) { 222 if (gref->file_index != index) 223 return NULL; 224 index += PAGE_SIZE; 225 count--; 226 if (count == 0) 227 return rv; 228 } 229 } 230 return NULL; 231} 232 233/* 234 * ------------------------------------- 235 * File operations. 236 * ------------------------------------- 237 */ 238static int gntalloc_open(struct inode *inode, struct file *filp) 239{ 240 struct gntalloc_file_private_data *priv; 241 242 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 243 if (!priv) 244 goto out_nomem; 245 INIT_LIST_HEAD(&priv->list); 246 247 filp->private_data = priv; 248 249 pr_debug("%s: priv %p\n", __func__, priv); 250 251 return 0; 252 253out_nomem: 254 return -ENOMEM; 255} 256 257static int gntalloc_release(struct inode *inode, struct file *filp) 258{ 259 struct gntalloc_file_private_data *priv = filp->private_data; 260 struct gntalloc_gref *gref; 261 262 pr_debug("%s: priv %p\n", __func__, priv); 263 264 mutex_lock(&gref_mutex); 265 while (!list_empty(&priv->list)) { 266 gref = list_entry(priv->list.next, 267 struct gntalloc_gref, next_file); 268 list_del(&gref->next_file); 269 gref->users--; 270 if (gref->users == 0) 271 __del_gref(gref); 272 } 273 kfree(priv); 274 mutex_unlock(&gref_mutex); 275 276 return 0; 277} 278 279static long gntalloc_ioctl_alloc(struct gntalloc_file_private_data *priv, 280 struct ioctl_gntalloc_alloc_gref __user *arg) 281{ 282 int rc = 0; 283 struct ioctl_gntalloc_alloc_gref op; 284 uint32_t *gref_ids; 285 286 pr_debug("%s: priv %p\n", __func__, priv); 287 288 if (copy_from_user(&op, arg, sizeof(op))) { 289 rc = -EFAULT; 290 goto out; 291 } 292 293 gref_ids = kcalloc(op.count, sizeof(gref_ids[0]), GFP_TEMPORARY); 294 if (!gref_ids) { 295 rc = -ENOMEM; 296 goto out; 297 } 298 299 mutex_lock(&gref_mutex); 300 /* Clean up pages that were at zero (local) users but were still mapped 301 * by remote domains. Since those pages count towards the limit that we 302 * are about to enforce, removing them here is a good idea. 303 */ 304 do_cleanup(); 305 if (gref_size + op.count > limit) { 306 mutex_unlock(&gref_mutex); 307 rc = -ENOSPC; 308 goto out_free; 309 } 310 gref_size += op.count; 311 op.index = priv->index; 312 priv->index += op.count * PAGE_SIZE; 313 mutex_unlock(&gref_mutex); 314 315 rc = add_grefs(&op, gref_ids, priv); 316 if (rc < 0) 317 goto out_free; 318 319 /* Once we finish add_grefs, it is unsafe to touch the new reference, 320 * since it is possible for a concurrent ioctl to remove it (by guessing 321 * its index). If the userspace application doesn't provide valid memory 322 * to write the IDs to, then it will need to close the file in order to 323 * release - which it will do by segfaulting when it tries to access the 324 * IDs to close them. 325 */ 326 if (copy_to_user(arg, &op, sizeof(op))) { 327 rc = -EFAULT; 328 goto out_free; 329 } 330 if (copy_to_user(arg->gref_ids, gref_ids, 331 sizeof(gref_ids[0]) * op.count)) { 332 rc = -EFAULT; 333 goto out_free; 334 } 335 336out_free: 337 kfree(gref_ids); 338out: 339 return rc; 340} 341 342static long gntalloc_ioctl_dealloc(struct gntalloc_file_private_data *priv, 343 void __user *arg) 344{ 345 int i, rc = 0; 346 struct ioctl_gntalloc_dealloc_gref op; 347 struct gntalloc_gref *gref, *n; 348 349 pr_debug("%s: priv %p\n", __func__, priv); 350 351 if (copy_from_user(&op, arg, sizeof(op))) { 352 rc = -EFAULT; 353 goto dealloc_grant_out; 354 } 355 356 mutex_lock(&gref_mutex); 357 gref = find_grefs(priv, op.index, op.count); 358 if (gref) { 359 /* Remove from the file list only, and decrease reference count. 360 * The later call to do_cleanup() will remove from gref_list and 361 * free the memory if the pages aren't mapped anywhere. 362 */ 363 for (i = 0; i < op.count; i++) { 364 n = list_entry(gref->next_file.next, 365 struct gntalloc_gref, next_file); 366 list_del(&gref->next_file); 367 gref->users--; 368 gref = n; 369 } 370 } else { 371 rc = -EINVAL; 372 } 373 374 do_cleanup(); 375 376 mutex_unlock(&gref_mutex); 377dealloc_grant_out: 378 return rc; 379} 380 381static long gntalloc_ioctl_unmap_notify(struct gntalloc_file_private_data *priv, 382 void __user *arg) 383{ 384 struct ioctl_gntalloc_unmap_notify op; 385 struct gntalloc_gref *gref; 386 uint64_t index; 387 int pgoff; 388 int rc; 389 390 if (copy_from_user(&op, arg, sizeof(op))) 391 return -EFAULT; 392 393 index = op.index & ~(PAGE_SIZE - 1); 394 pgoff = op.index & (PAGE_SIZE - 1); 395 396 mutex_lock(&gref_mutex); 397 398 gref = find_grefs(priv, index, 1); 399 if (!gref) { 400 rc = -ENOENT; 401 goto unlock_out; 402 } 403 404 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT)) { 405 rc = -EINVAL; 406 goto unlock_out; 407 } 408 409 /* We need to grab a reference to the event channel we are going to use 410 * to send the notify before releasing the reference we may already have 411 * (if someone has called this ioctl twice). This is required so that 412 * it is possible to change the clear_byte part of the notification 413 * without disturbing the event channel part, which may now be the last 414 * reference to that event channel. 415 */ 416 if (op.action & UNMAP_NOTIFY_SEND_EVENT) { 417 if (evtchn_get(op.event_channel_port)) { 418 rc = -EINVAL; 419 goto unlock_out; 420 } 421 } 422 423 if (gref->notify.flags & UNMAP_NOTIFY_SEND_EVENT) 424 evtchn_put(gref->notify.event); 425 426 gref->notify.flags = op.action; 427 gref->notify.pgoff = pgoff; 428 gref->notify.event = op.event_channel_port; 429 rc = 0; 430 431 unlock_out: 432 mutex_unlock(&gref_mutex); 433 return rc; 434} 435 436static long gntalloc_ioctl(struct file *filp, unsigned int cmd, 437 unsigned long arg) 438{ 439 struct gntalloc_file_private_data *priv = filp->private_data; 440 441 switch (cmd) { 442 case IOCTL_GNTALLOC_ALLOC_GREF: 443 return gntalloc_ioctl_alloc(priv, (void __user *)arg); 444 445 case IOCTL_GNTALLOC_DEALLOC_GREF: 446 return gntalloc_ioctl_dealloc(priv, (void __user *)arg); 447 448 case IOCTL_GNTALLOC_SET_UNMAP_NOTIFY: 449 return gntalloc_ioctl_unmap_notify(priv, (void __user *)arg); 450 451 default: 452 return -ENOIOCTLCMD; 453 } 454 455 return 0; 456} 457 458static void gntalloc_vma_open(struct vm_area_struct *vma) 459{ 460 struct gntalloc_vma_private_data *priv = vma->vm_private_data; 461 462 if (!priv) 463 return; 464 465 mutex_lock(&gref_mutex); 466 priv->users++; 467 mutex_unlock(&gref_mutex); 468} 469 470static void gntalloc_vma_close(struct vm_area_struct *vma) 471{ 472 struct gntalloc_vma_private_data *priv = vma->vm_private_data; 473 struct gntalloc_gref *gref, *next; 474 int i; 475 476 if (!priv) 477 return; 478 479 mutex_lock(&gref_mutex); 480 priv->users--; 481 if (priv->users == 0) { 482 gref = priv->gref; 483 for (i = 0; i < priv->count; i++) { 484 gref->users--; 485 next = list_entry(gref->next_gref.next, 486 struct gntalloc_gref, next_gref); 487 if (gref->users == 0) 488 __del_gref(gref); 489 gref = next; 490 } 491 kfree(priv); 492 } 493 mutex_unlock(&gref_mutex); 494} 495 496static struct vm_operations_struct gntalloc_vmops = { 497 .open = gntalloc_vma_open, 498 .close = gntalloc_vma_close, 499}; 500 501static int gntalloc_mmap(struct file *filp, struct vm_area_struct *vma) 502{ 503 struct gntalloc_file_private_data *priv = filp->private_data; 504 struct gntalloc_vma_private_data *vm_priv; 505 struct gntalloc_gref *gref; 506 int count = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 507 int rv, i; 508 509 if (!(vma->vm_flags & VM_SHARED)) { 510 pr_err("%s: Mapping must be shared\n", __func__); 511 return -EINVAL; 512 } 513 514 vm_priv = kmalloc(sizeof(*vm_priv), GFP_KERNEL); 515 if (!vm_priv) 516 return -ENOMEM; 517 518 mutex_lock(&gref_mutex); 519 520 pr_debug("%s: priv %p,%p, page %lu+%d\n", __func__, 521 priv, vm_priv, vma->vm_pgoff, count); 522 523 gref = find_grefs(priv, vma->vm_pgoff << PAGE_SHIFT, count); 524 if (gref == NULL) { 525 rv = -ENOENT; 526 pr_debug("%s: Could not find grant reference", 527 __func__); 528 kfree(vm_priv); 529 goto out_unlock; 530 } 531 532 vm_priv->gref = gref; 533 vm_priv->users = 1; 534 vm_priv->count = count; 535 536 vma->vm_private_data = vm_priv; 537 538 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP; 539 540 vma->vm_ops = &gntalloc_vmops; 541 542 for (i = 0; i < count; i++) { 543 gref->users++; 544 rv = vm_insert_page(vma, vma->vm_start + i * PAGE_SIZE, 545 gref->page); 546 if (rv) 547 goto out_unlock; 548 549 gref = list_entry(gref->next_file.next, 550 struct gntalloc_gref, next_file); 551 } 552 rv = 0; 553 554out_unlock: 555 mutex_unlock(&gref_mutex); 556 return rv; 557} 558 559static const struct file_operations gntalloc_fops = { 560 .owner = THIS_MODULE, 561 .open = gntalloc_open, 562 .release = gntalloc_release, 563 .unlocked_ioctl = gntalloc_ioctl, 564 .mmap = gntalloc_mmap 565}; 566 567/* 568 * ------------------------------------- 569 * Module creation/destruction. 570 * ------------------------------------- 571 */ 572static struct miscdevice gntalloc_miscdev = { 573 .minor = MISC_DYNAMIC_MINOR, 574 .name = "xen/gntalloc", 575 .fops = &gntalloc_fops, 576}; 577 578static int __init gntalloc_init(void) 579{ 580 int err; 581 582 if (!xen_domain()) 583 return -ENODEV; 584 585 err = misc_register(&gntalloc_miscdev); 586 if (err != 0) { 587 pr_err("Could not register misc gntalloc device\n"); 588 return err; 589 } 590 591 pr_debug("Created grant allocation device at %d,%d\n", 592 MISC_MAJOR, gntalloc_miscdev.minor); 593 594 return 0; 595} 596 597static void __exit gntalloc_exit(void) 598{ 599 misc_deregister(&gntalloc_miscdev); 600} 601 602module_init(gntalloc_init); 603module_exit(gntalloc_exit); 604 605MODULE_LICENSE("GPL"); 606MODULE_AUTHOR("Carter Weatherly <carter.weatherly@jhuapl.edu>, " 607 "Daniel De Graaf <dgdegra@tycho.nsa.gov>"); 608MODULE_DESCRIPTION("User-space grant reference allocator driver"); 609