fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <miklos@szeredi.hu>
   4
   5   This program can be distributed under the terms of the GNU GPL.
   6   See the file COPYING.
   7 */
   8
   9 #include "fuse_i.h"
  10
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/poll.h>
  14 #include <linux/sched/signal.h>
  15 #include <linux/uio.h>
  16 #include <linux/miscdevice.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/pipe_fs_i.h>
  21 #include <linux/swap.h>
  22 #include <linux/splice.h>
  23 #include <linux/sched.h>
  24
  25 MODULE_ALIAS_MISCDEV(FUSE_MINOR);
  26 MODULE_ALIAS("devname:fuse");
  27
  28 /* Ordinary requests have even IDs, while interrupts IDs are odd */
  29 #define FUSE_INT_REQ_BIT (1ULL << 0)
  30 #define FUSE_REQ_ID_STEP (1ULL << 1)
  31
  32 static struct kmem_cache *fuse_req_cachep;
  33
  34 static struct fuse_dev *fuse_get_dev(struct file *file)
  35 {
  36         /*
  37          * Lockless access is OK, because file->private data is set
  38          * once during mount and is valid until the file is released.
  39          */
  40         return READ_ONCE(file->private_data);
  41 }
  42
  43 static void fuse_request_init(struct fuse_req *req, struct page **pages,
  44                               struct fuse_page_desc *page_descs,
  45                               unsigned npages)
  46 {
  47         INIT_LIST_HEAD(&req->list);
  48         INIT_LIST_HEAD(&req->intr_entry);
  49         init_waitqueue_head(&req->waitq);
  50         refcount_set(&req->count, 1);
  51         req->pages = pages;
  52         req->page_descs = page_descs;
  53         req->max_pages = npages;
  54         __set_bit(FR_PENDING, &req->flags);
  55 }
  56
  57 static struct page **fuse_req_pages_alloc(unsigned int npages, gfp_t flags,
  58                                           struct fuse_page_desc **desc)
  59 {
  60         struct page **pages;
  61
  62         pages = kzalloc(npages * (sizeof(struct page *) +
  63                                   sizeof(struct fuse_page_desc)), flags);
  64         *desc = (void *) pages + npages * sizeof(struct page *);
  65
  66         return pages;
  67 }
  68
  69 static struct fuse_req *__fuse_request_alloc(unsigned npages, gfp_t flags)
  70 {
  71         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  72         if (req) {
  73                 struct page **pages = NULL;
  74                 struct fuse_page_desc *page_descs = NULL;
  75
  76                 WARN_ON(npages > FUSE_MAX_MAX_PAGES);
  77                 if (npages > FUSE_REQ_INLINE_PAGES) {
  78                         pages = fuse_req_pages_alloc(npages, flags,
  79                                                      &page_descs);
  80                         if (!pages) {
  81                                 kmem_cache_free(fuse_req_cachep, req);
  82                                 return NULL;
  83                         }
  84                 } else if (npages) {
  85                         pages = req->inline_pages;
  86                         page_descs = req->inline_page_descs;
  87                 }
  88
  89                 fuse_request_init(req, pages, page_descs, npages);
  90         }
  91         return req;
  92 }
  93
  94 struct fuse_req *fuse_request_alloc(unsigned npages)
  95 {
  96         return __fuse_request_alloc(npages, GFP_KERNEL);
  97 }
  98 EXPORT_SYMBOL_GPL(fuse_request_alloc);
  99
 100 struct fuse_req *fuse_request_alloc_nofs(unsigned npages)
 101 {
 102         return __fuse_request_alloc(npages, GFP_NOFS);
 103 }
 104
 105 static void fuse_req_pages_free(struct fuse_req *req)
 106 {
 107         if (req->pages != req->inline_pages)
 108                 kfree(req->pages);
 109 }
 110
 111 bool fuse_req_realloc_pages(struct fuse_conn *fc, struct fuse_req *req,
 112                             gfp_t flags)
 113 {
 114         struct page **pages;
 115         struct fuse_page_desc *page_descs;
 116         unsigned int npages = min_t(unsigned int,
 117                                     max_t(unsigned int, req->max_pages * 2,
 118                                           FUSE_DEFAULT_MAX_PAGES_PER_REQ),
 119                                     fc->max_pages);
 120         WARN_ON(npages <= req->max_pages);
 121
 122         pages = fuse_req_pages_alloc(npages, flags, &page_descs);
 123         if (!pages)
 124                 return false;
 125
 126         memcpy(pages, req->pages, sizeof(struct page *) * req->max_pages);
 127         memcpy(page_descs, req->page_descs,
 128                sizeof(struct fuse_page_desc) * req->max_pages);
 129         fuse_req_pages_free(req);
 130         req->pages = pages;
 131         req->page_descs = page_descs;
 132         req->max_pages = npages;
 133
 134         return true;
 135 }
 136
 137 void fuse_request_free(struct fuse_req *req)
 138 {
 139         fuse_req_pages_free(req);
 140         kmem_cache_free(fuse_req_cachep, req);
 141 }
 142
 143 void __fuse_get_request(struct fuse_req *req)
 144 {
 145         refcount_inc(&req->count);
 146 }
 147
 148 /* Must be called with > 1 refcount */
 149 static void __fuse_put_request(struct fuse_req *req)
 150 {
 151         refcount_dec(&req->count);
 152 }
 153
 154 void fuse_set_initialized(struct fuse_conn *fc)
 155 {
 156         /* Make sure stores before this are seen on another CPU */
 157         smp_wmb();
 158         fc->initialized = 1;
 159 }
 160
 161 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
 162 {
 163         return !fc->initialized || (for_background && fc->blocked);
 164 }
 165
 166 static void fuse_drop_waiting(struct fuse_conn *fc)
 167 {
 168         /*
 169          * lockess check of fc->connected is okay, because atomic_dec_and_test()
 170          * provides a memory barrier mached with the one in fuse_wait_aborted()
 171          * to ensure no wake-up is missed.
 172          */
 173         if (atomic_dec_and_test(&fc->num_waiting) &&
 174             !READ_ONCE(fc->connected)) {
 175                 /* wake up aborters */
 176                 wake_up_all(&fc->blocked_waitq);
 177         }
 178 }
 179
 180 static struct fuse_req *__fuse_get_req(struct fuse_conn *fc, unsigned npages,
 181                                        bool for_background)
 182 {
 183         struct fuse_req *req;
 184         int err;
 185         atomic_inc(&fc->num_waiting);
 186
 187         if (fuse_block_alloc(fc, for_background)) {
 188                 err = -EINTR;
 189                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 190                                 !fuse_block_alloc(fc, for_background)))
 191                         goto out;
 192         }
 193         /* Matches smp_wmb() in fuse_set_initialized() */
 194         smp_rmb();
 195
 196         err = -ENOTCONN;
 197         if (!fc->connected)
 198                 goto out;
 199
 200         err = -ECONNREFUSED;
 201         if (fc->conn_error)
 202                 goto out;
 203
 204         req = fuse_request_alloc(npages);
 205         err = -ENOMEM;
 206         if (!req) {
 207                 if (for_background)
 208                         wake_up(&fc->blocked_waitq);
 209                 goto out;
 210         }
 211
 212         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 213         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 214         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 215
 216         __set_bit(FR_WAITING, &req->flags);
 217         if (for_background)
 218                 __set_bit(FR_BACKGROUND, &req->flags);
 219
 220         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 221                      req->in.h.gid == ((gid_t)-1))) {
 222                 fuse_put_request(fc, req);
 223                 return ERR_PTR(-EOVERFLOW);
 224         }
 225         return req;
 226
 227  out:
 228         fuse_drop_waiting(fc);
 229         return ERR_PTR(err);
 230 }
 231
 232 struct fuse_req *fuse_get_req(struct fuse_conn *fc, unsigned npages)
 233 {
 234         return __fuse_get_req(fc, npages, false);
 235 }
 236 EXPORT_SYMBOL_GPL(fuse_get_req);
 237
 238 struct fuse_req *fuse_get_req_for_background(struct fuse_conn *fc,
 239                                              unsigned npages)
 240 {
 241         return __fuse_get_req(fc, npages, true);
 242 }
 243 EXPORT_SYMBOL_GPL(fuse_get_req_for_background);
 244
 245 /*
 246  * Return request in fuse_file->reserved_req.  However that may
 247  * currently be in use.  If that is the case, wait for it to become
 248  * available.
 249  */
 250 static struct fuse_req *get_reserved_req(struct fuse_conn *fc,
 251                                          struct file *file)
 252 {
 253         struct fuse_req *req = NULL;
 254         struct fuse_file *ff = file->private_data;
 255
 256         do {
 257                 wait_event(fc->reserved_req_waitq, ff->reserved_req);
 258                 spin_lock(&fc->lock);
 259                 if (ff->reserved_req) {
 260                         req = ff->reserved_req;
 261                         ff->reserved_req = NULL;
 262                         req->stolen_file = get_file(file);
 263                 }
 264                 spin_unlock(&fc->lock);
 265         } while (!req);
 266
 267         return req;
 268 }
 269
 270 /*
 271  * Put stolen request back into fuse_file->reserved_req
 272  */
 273 static void put_reserved_req(struct fuse_conn *fc, struct fuse_req *req)
 274 {
 275         struct file *file = req->stolen_file;
 276         struct fuse_file *ff = file->private_data;
 277
 278         WARN_ON(req->max_pages);
 279         spin_lock(&fc->lock);
 280         memset(req, 0, sizeof(*req));
 281         fuse_request_init(req, NULL, NULL, 0);
 282         BUG_ON(ff->reserved_req);
 283         ff->reserved_req = req;
 284         wake_up_all(&fc->reserved_req_waitq);
 285         spin_unlock(&fc->lock);
 286         fput(file);
 287 }
 288
 289 /*
 290  * Gets a requests for a file operation, always succeeds
 291  *
 292  * This is used for sending the FLUSH request, which must get to
 293  * userspace, due to POSIX locks which may need to be unlocked.
 294  *
 295  * If allocation fails due to OOM, use the reserved request in
 296  * fuse_file.
 297  *
 298  * This is very unlikely to deadlock accidentally, since the
 299  * filesystem should not have it's own file open.  If deadlock is
 300  * intentional, it can still be broken by "aborting" the filesystem.
 301  */
 302 struct fuse_req *fuse_get_req_nofail_nopages(struct fuse_conn *fc,
 303                                              struct file *file)
 304 {
 305         struct fuse_req *req;
 306
 307         atomic_inc(&fc->num_waiting);
 308         wait_event(fc->blocked_waitq, fc->initialized);
 309         /* Matches smp_wmb() in fuse_set_initialized() */
 310         smp_rmb();
 311         req = fuse_request_alloc(0);
 312         if (!req)
 313                 req = get_reserved_req(fc, file);
 314
 315         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 316         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 317         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 318
 319         __set_bit(FR_WAITING, &req->flags);
 320         __clear_bit(FR_BACKGROUND, &req->flags);
 321         return req;
 322 }
 323
 324 void fuse_put_request(struct fuse_conn *fc, struct fuse_req *req)
 325 {
 326         if (refcount_dec_and_test(&req->count)) {
 327                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 328                         /*
 329                          * We get here in the unlikely case that a background
 330                          * request was allocated but not sent
 331                          */
 332                         spin_lock(&fc->bg_lock);
 333                         if (!fc->blocked)
 334                                 wake_up(&fc->blocked_waitq);
 335                         spin_unlock(&fc->bg_lock);
 336                 }
 337
 338                 if (test_bit(FR_WAITING, &req->flags)) {
 339                         __clear_bit(FR_WAITING, &req->flags);
 340                         fuse_drop_waiting(fc);
 341                 }
 342
 343                 if (req->stolen_file)
 344                         put_reserved_req(fc, req);
 345                 else
 346                         fuse_request_free(req);
 347         }
 348 }
 349 EXPORT_SYMBOL_GPL(fuse_put_request);
 350
 351 static unsigned len_args(unsigned numargs, struct fuse_arg *args)
 352 {
 353         unsigned nbytes = 0;
 354         unsigned i;
 355
 356         for (i = 0; i < numargs; i++)
 357                 nbytes += args[i].size;
 358
 359         return nbytes;
 360 }
 361
 362 static u64 fuse_get_unique(struct fuse_iqueue *fiq)
 363 {
 364         fiq->reqctr += FUSE_REQ_ID_STEP;
 365         return fiq->reqctr;
 366 }
 367
 368 static unsigned int fuse_req_hash(u64 unique)
 369 {
 370         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 371 }
 372
 373 static void queue_request(struct fuse_iqueue *fiq, struct fuse_req *req)
 374 {
 375         req->in.h.len = sizeof(struct fuse_in_header) +
 376                 len_args(req->in.numargs, (struct fuse_arg *) req->in.args);
 377         list_add_tail(&req->list, &fiq->pending);
 378         wake_up_locked(&fiq->waitq);
 379         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 380 }
 381
 382 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 383                        u64 nodeid, u64 nlookup)
 384 {
 385         struct fuse_iqueue *fiq = &fc->iq;
 386
 387         forget->forget_one.nodeid = nodeid;
 388         forget->forget_one.nlookup = nlookup;
 389
 390         spin_lock(&fiq->waitq.lock);
 391         if (fiq->connected) {
 392                 fiq->forget_list_tail->next = forget;
 393                 fiq->forget_list_tail = forget;
 394                 wake_up_locked(&fiq->waitq);
 395                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 396         } else {
 397                 kfree(forget);
 398         }
 399         spin_unlock(&fiq->waitq.lock);
 400 }
 401
 402 static void flush_bg_queue(struct fuse_conn *fc)
 403 {
 404         struct fuse_iqueue *fiq = &fc->iq;
 405
 406         while (fc->active_background < fc->max_background &&
 407                !list_empty(&fc->bg_queue)) {
 408                 struct fuse_req *req;
 409
 410                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 411                 list_del(&req->list);
 412                 fc->active_background++;
 413                 spin_lock(&fiq->waitq.lock);
 414                 req->in.h.unique = fuse_get_unique(fiq);
 415                 queue_request(fiq, req);
 416                 spin_unlock(&fiq->waitq.lock);
 417         }
 418 }
 419
 420 /*
 421  * This function is called when a request is finished.  Either a reply
 422  * has arrived or it was aborted (and not yet sent) or some error
 423  * occurred during communication with userspace, or the device file
 424  * was closed.  The requester thread is woken up (if still waiting),
 425  * the 'end' callback is called if given, else the reference to the
 426  * request is released
 427  */
 428 static void request_end(struct fuse_conn *fc, struct fuse_req *req)
 429 {
 430         struct fuse_iqueue *fiq = &fc->iq;
 431
 432         if (test_and_set_bit(FR_FINISHED, &req->flags))
 433                 goto put_request;
 434         /*
 435          * test_and_set_bit() implies smp_mb() between bit
 436          * changing and below intr_entry check. Pairs with
 437          * smp_mb() from queue_interrupt().
 438          */
 439         if (!list_empty(&req->intr_entry)) {
 440                 spin_lock(&fiq->waitq.lock);
 441                 list_del_init(&req->intr_entry);
 442                 spin_unlock(&fiq->waitq.lock);
 443         }
 444         WARN_ON(test_bit(FR_PENDING, &req->flags));
 445         WARN_ON(test_bit(FR_SENT, &req->flags));
 446         if (test_bit(FR_BACKGROUND, &req->flags)) {
 447                 spin_lock(&fc->bg_lock);
 448                 clear_bit(FR_BACKGROUND, &req->flags);
 449                 if (fc->num_background == fc->max_background) {
 450                         fc->blocked = 0;
 451                         wake_up(&fc->blocked_waitq);
 452                 } else if (!fc->blocked) {
 453                         /*
 454                          * Wake up next waiter, if any.  It's okay to use
 455                          * waitqueue_active(), as we've already synced up
 456                          * fc->blocked with waiters with the wake_up() call
 457                          * above.
 458                          */
 459                         if (waitqueue_active(&fc->blocked_waitq))
 460                                 wake_up(&fc->blocked_waitq);
 461                 }
 462
 463                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 464                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 465                         clear_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 466                 }
 467                 fc->num_background--;
 468                 fc->active_background--;
 469                 flush_bg_queue(fc);
 470                 spin_unlock(&fc->bg_lock);
 471         }
 472         wake_up(&req->waitq);
 473         if (req->end)
 474                 req->end(fc, req);
 475 put_request:
 476         fuse_put_request(fc, req);
 477 }
 478
 479 static void queue_interrupt(struct fuse_iqueue *fiq, struct fuse_req *req)
 480 {
 481         spin_lock(&fiq->waitq.lock);
 482         if (list_empty(&req->intr_entry)) {
 483                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 484                 /*
 485                  * Pairs with smp_mb() implied by test_and_set_bit()
 486                  * from request_end().
 487                  */
 488                 smp_mb();
 489                 if (test_bit(FR_FINISHED, &req->flags)) {
 490                         list_del_init(&req->intr_entry);
 491                         spin_unlock(&fiq->waitq.lock);
 492                         return;
 493                 }
 494                 wake_up_locked(&fiq->waitq);
 495                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 496         }
 497         spin_unlock(&fiq->waitq.lock);
 498 }
 499
 500 static void request_wait_answer(struct fuse_conn *fc, struct fuse_req *req)
 501 {
 502         struct fuse_iqueue *fiq = &fc->iq;
 503         int err;
 504
 505         if (!fc->no_interrupt) {
 506                 /* Any signal may interrupt this */
 507                 err = wait_event_interruptible(req->waitq,
 508                                         test_bit(FR_FINISHED, &req->flags));
 509                 if (!err)
 510                         return;
 511
 512                 set_bit(FR_INTERRUPTED, &req->flags);
 513                 /* matches barrier in fuse_dev_do_read() */
 514                 smp_mb__after_atomic();
 515                 if (test_bit(FR_SENT, &req->flags))
 516                         queue_interrupt(fiq, req);
 517         }
 518
 519         if (!test_bit(FR_FORCE, &req->flags)) {
 520                 /* Only fatal signals may interrupt this */
 521                 err = wait_event_killable(req->waitq,
 522                                         test_bit(FR_FINISHED, &req->flags));
 523                 if (!err)
 524                         return;
 525
 526                 spin_lock(&fiq->waitq.lock);
 527                 /* Request is not yet in userspace, bail out */
 528                 if (test_bit(FR_PENDING, &req->flags)) {
 529                         list_del(&req->list);
 530                         spin_unlock(&fiq->waitq.lock);
 531                         __fuse_put_request(req);
 532                         req->out.h.error = -EINTR;
 533                         return;
 534                 }
 535                 spin_unlock(&fiq->waitq.lock);
 536         }
 537
 538         /*
 539          * Either request is already in userspace, or it was forced.
 540          * Wait it out.
 541          */
 542         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 543 }
 544
 545 static void __fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 546 {
 547         struct fuse_iqueue *fiq = &fc->iq;
 548
 549         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 550         spin_lock(&fiq->waitq.lock);
 551         if (!fiq->connected) {
 552                 spin_unlock(&fiq->waitq.lock);
 553                 req->out.h.error = -ENOTCONN;
 554         } else {
 555                 req->in.h.unique = fuse_get_unique(fiq);
 556                 queue_request(fiq, req);
 557                 /* acquire extra reference, since request is still needed
 558                    after request_end() */
 559                 __fuse_get_request(req);
 560                 spin_unlock(&fiq->waitq.lock);
 561
 562                 request_wait_answer(fc, req);
 563                 /* Pairs with smp_wmb() in request_end() */
 564                 smp_rmb();
 565         }
 566 }
 567
 568 void fuse_request_send(struct fuse_conn *fc, struct fuse_req *req)
 569 {
 570         __set_bit(FR_ISREPLY, &req->flags);
 571         if (!test_bit(FR_WAITING, &req->flags)) {
 572                 __set_bit(FR_WAITING, &req->flags);
 573                 atomic_inc(&fc->num_waiting);
 574         }
 575         __fuse_request_send(fc, req);
 576 }
 577 EXPORT_SYMBOL_GPL(fuse_request_send);
 578
 579 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 580 {
 581         if (fc->minor < 4 && args->in.h.opcode == FUSE_STATFS)
 582                 args->out.args[0].size = FUSE_COMPAT_STATFS_SIZE;
 583
 584         if (fc->minor < 9) {
 585                 switch (args->in.h.opcode) {
 586                 case FUSE_LOOKUP:
 587                 case FUSE_CREATE:
 588                 case FUSE_MKNOD:
 589                 case FUSE_MKDIR:
 590                 case FUSE_SYMLINK:
 591                 case FUSE_LINK:
 592                         args->out.args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 593                         break;
 594                 case FUSE_GETATTR:
 595                 case FUSE_SETATTR:
 596                         args->out.args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 597                         break;
 598                 }
 599         }
 600         if (fc->minor < 12) {
 601                 switch (args->in.h.opcode) {
 602                 case FUSE_CREATE:
 603                         args->in.args[0].size = sizeof(struct fuse_open_in);
 604                         break;
 605                 case FUSE_MKNOD:
 606                         args->in.args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 607                         break;
 608                 }
 609         }
 610 }
 611
 612 ssize_t fuse_simple_request(struct fuse_conn *fc, struct fuse_args *args)
 613 {
 614         struct fuse_req *req;
 615         ssize_t ret;
 616
 617         req = fuse_get_req(fc, 0);
 618         if (IS_ERR(req))
 619                 return PTR_ERR(req);
 620
 621         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 622         fuse_adjust_compat(fc, args);
 623
 624         req->in.h.opcode = args->in.h.opcode;
 625         req->in.h.nodeid = args->in.h.nodeid;
 626         req->in.numargs = args->in.numargs;
 627         memcpy(req->in.args, args->in.args,
 628                args->in.numargs * sizeof(struct fuse_in_arg));
 629         req->out.argvar = args->out.argvar;
 630         req->out.numargs = args->out.numargs;
 631         memcpy(req->out.args, args->out.args,
 632                args->out.numargs * sizeof(struct fuse_arg));
 633         fuse_request_send(fc, req);
 634         ret = req->out.h.error;
 635         if (!ret && args->out.argvar) {
 636                 BUG_ON(args->out.numargs != 1);
 637                 ret = req->out.args[0].size;
 638         }
 639         fuse_put_request(fc, req);
 640
 641         return ret;
 642 }
 643
 644 bool fuse_request_queue_background(struct fuse_conn *fc, struct fuse_req *req)
 645 {
 646         bool queued = false;
 647
 648         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 649         if (!test_bit(FR_WAITING, &req->flags)) {
 650                 __set_bit(FR_WAITING, &req->flags);
 651                 atomic_inc(&fc->num_waiting);
 652         }
 653         __set_bit(FR_ISREPLY, &req->flags);
 654         spin_lock(&fc->bg_lock);
 655         if (likely(fc->connected)) {
 656                 fc->num_background++;
 657                 if (fc->num_background == fc->max_background)
 658                         fc->blocked = 1;
 659                 if (fc->num_background == fc->congestion_threshold && fc->sb) {
 660                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_SYNC);
 661                         set_bdi_congested(fc->sb->s_bdi, BLK_RW_ASYNC);
 662                 }
 663                 list_add_tail(&req->list, &fc->bg_queue);
 664                 flush_bg_queue(fc);
 665                 queued = true;
 666         }
 667         spin_unlock(&fc->bg_lock);
 668
 669         return queued;
 670 }
 671
 672 void fuse_request_send_background(struct fuse_conn *fc, struct fuse_req *req)
 673 {
 674         WARN_ON(!req->end);
 675         if (!fuse_request_queue_background(fc, req)) {
 676                 req->out.h.error = -ENOTCONN;
 677                 req->end(fc, req);
 678                 fuse_put_request(fc, req);
 679         }
 680 }
 681 EXPORT_SYMBOL_GPL(fuse_request_send_background);
 682
 683 static int fuse_request_send_notify_reply(struct fuse_conn *fc,
 684                                           struct fuse_req *req, u64 unique)
 685 {
 686         int err = -ENODEV;
 687         struct fuse_iqueue *fiq = &fc->iq;
 688
 689         __clear_bit(FR_ISREPLY, &req->flags);
 690         req->in.h.unique = unique;
 691         spin_lock(&fiq->waitq.lock);
 692         if (fiq->connected) {
 693                 queue_request(fiq, req);
 694                 err = 0;
 695         }
 696         spin_unlock(&fiq->waitq.lock);
 697
 698         return err;
 699 }
 700
 701 void fuse_force_forget(struct file *file, u64 nodeid)
 702 {
 703         struct inode *inode = file_inode(file);
 704         struct fuse_conn *fc = get_fuse_conn(inode);
 705         struct fuse_req *req;
 706         struct fuse_forget_in inarg;
 707
 708         memset(&inarg, 0, sizeof(inarg));
 709         inarg.nlookup = 1;
 710         req = fuse_get_req_nofail_nopages(fc, file);
 711         req->in.h.opcode = FUSE_FORGET;
 712         req->in.h.nodeid = nodeid;
 713         req->in.numargs = 1;
 714         req->in.args[0].size = sizeof(inarg);
 715         req->in.args[0].value = &inarg;
 716         __clear_bit(FR_ISREPLY, &req->flags);
 717         __fuse_request_send(fc, req);
 718         /* ignore errors */
 719         fuse_put_request(fc, req);
 720 }
 721
 722 /*
 723  * Lock the request.  Up to the next unlock_request() there mustn't be
 724  * anything that could cause a page-fault.  If the request was already
 725  * aborted bail out.
 726  */
 727 static int lock_request(struct fuse_req *req)
 728 {
 729         int err = 0;
 730         if (req) {
 731                 spin_lock(&req->waitq.lock);
 732                 if (test_bit(FR_ABORTED, &req->flags))
 733                         err = -ENOENT;
 734                 else
 735                         set_bit(FR_LOCKED, &req->flags);
 736                 spin_unlock(&req->waitq.lock);
 737         }
 738         return err;
 739 }
 740
 741 /*
 742  * Unlock request.  If it was aborted while locked, caller is responsible
 743  * for unlocking and ending the request.
 744  */
 745 static int unlock_request(struct fuse_req *req)
 746 {
 747         int err = 0;
 748         if (req) {
 749                 spin_lock(&req->waitq.lock);
 750                 if (test_bit(FR_ABORTED, &req->flags))
 751                         err = -ENOENT;
 752                 else
 753                         clear_bit(FR_LOCKED, &req->flags);
 754                 spin_unlock(&req->waitq.lock);
 755         }
 756         return err;
 757 }
 758
 759 struct fuse_copy_state {
 760         int write;
 761         struct fuse_req *req;
 762         struct iov_iter *iter;
 763         struct pipe_buffer *pipebufs;
 764         struct pipe_buffer *currbuf;
 765         struct pipe_inode_info *pipe;
 766         unsigned long nr_segs;
 767         struct page *pg;
 768         unsigned len;
 769         unsigned offset;
 770         unsigned move_pages:1;
 771 };
 772
 773 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 774                            struct iov_iter *iter)
 775 {
 776         memset(cs, 0, sizeof(*cs));
 777         cs->write = write;
 778         cs->iter = iter;
 779 }
 780
 781 /* Unmap and put previous page of userspace buffer */
 782 static void fuse_copy_finish(struct fuse_copy_state *cs)
 783 {
 784         if (cs->currbuf) {
 785                 struct pipe_buffer *buf = cs->currbuf;
 786
 787                 if (cs->write)
 788                         buf->len = PAGE_SIZE - cs->len;
 789                 cs->currbuf = NULL;
 790         } else if (cs->pg) {
 791                 if (cs->write) {
 792                         flush_dcache_page(cs->pg);
 793                         set_page_dirty_lock(cs->pg);
 794                 }
 795                 put_page(cs->pg);
 796         }
 797         cs->pg = NULL;
 798 }
 799
 800 /*
 801  * Get another pagefull of userspace buffer, and map it to kernel
 802  * address space, and lock request
 803  */
 804 static int fuse_copy_fill(struct fuse_copy_state *cs)
 805 {
 806         struct page *page;
 807         int err;
 808
 809         err = unlock_request(cs->req);
 810         if (err)
 811                 return err;
 812
 813         fuse_copy_finish(cs);
 814         if (cs->pipebufs) {
 815                 struct pipe_buffer *buf = cs->pipebufs;
 816
 817                 if (!cs->write) {
 818                         err = pipe_buf_confirm(cs->pipe, buf);
 819                         if (err)
 820                                 return err;
 821
 822                         BUG_ON(!cs->nr_segs);
 823                         cs->currbuf = buf;
 824                         cs->pg = buf->page;
 825                         cs->offset = buf->offset;
 826                         cs->len = buf->len;
 827                         cs->pipebufs++;
 828                         cs->nr_segs--;
 829                 } else {
 830                         if (cs->nr_segs == cs->pipe->buffers)
 831                                 return -EIO;
 832
 833                         page = alloc_page(GFP_HIGHUSER);
 834                         if (!page)
 835                                 return -ENOMEM;
 836
 837                         buf->page = page;
 838                         buf->offset = 0;
 839                         buf->len = 0;
 840
 841                         cs->currbuf = buf;
 842                         cs->pg = page;
 843                         cs->offset = 0;
 844                         cs->len = PAGE_SIZE;
 845                         cs->pipebufs++;
 846                         cs->nr_segs++;
 847                 }
 848         } else {
 849                 size_t off;
 850                 err = iov_iter_get_pages(cs->iter, &page, PAGE_SIZE, 1, &off);
 851                 if (err < 0)
 852                         return err;
 853                 BUG_ON(!err);
 854                 cs->len = err;
 855                 cs->offset = off;
 856                 cs->pg = page;
 857                 iov_iter_advance(cs->iter, err);
 858         }
 859
 860         return lock_request(cs->req);
 861 }
 862
 863 /* Do as much copy to/from userspace buffer as we can */
 864 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 865 {
 866         unsigned ncpy = min(*size, cs->len);
 867         if (val) {
 868                 void *pgaddr = kmap_atomic(cs->pg);
 869                 void *buf = pgaddr + cs->offset;
 870
 871                 if (cs->write)
 872                         memcpy(buf, *val, ncpy);
 873                 else
 874                         memcpy(*val, buf, ncpy);
 875
 876                 kunmap_atomic(pgaddr);
 877                 *val += ncpy;
 878         }
 879         *size -= ncpy;
 880         cs->len -= ncpy;
 881         cs->offset += ncpy;
 882         return ncpy;
 883 }
 884
 885 static int fuse_check_page(struct page *page)
 886 {
 887         if (page_mapcount(page) ||
 888             page->mapping != NULL ||
 889             page_count(page) != 1 ||
 890             (page->flags & PAGE_FLAGS_CHECK_AT_PREP &
 891              ~(1 << PG_locked |
 892                1 << PG_referenced |
 893                1 << PG_uptodate |
 894                1 << PG_lru |
 895                1 << PG_active |
 896                1 << PG_reclaim))) {
 897                 printk(KERN_WARNING "fuse: trying to steal weird page\n");
 898                 printk(KERN_WARNING "  page=%p index=%li flags=%08lx, count=%i, mapcount=%i, mapping=%p\n", page, page->index, page->flags, page_count(page), page_mapcount(page), page->mapping);
 899                 return 1;
 900         }
 901         return 0;
 902 }
 903
 904 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 905 {
 906         int err;
 907         struct page *oldpage = *pagep;
 908         struct page *newpage;
 909         struct pipe_buffer *buf = cs->pipebufs;
 910
 911         err = unlock_request(cs->req);
 912         if (err)
 913                 return err;
 914
 915         fuse_copy_finish(cs);
 916
 917         err = pipe_buf_confirm(cs->pipe, buf);
 918         if (err)
 919                 return err;
 920
 921         BUG_ON(!cs->nr_segs);
 922         cs->currbuf = buf;
 923         cs->len = buf->len;
 924         cs->pipebufs++;
 925         cs->nr_segs--;
 926
 927         if (cs->len != PAGE_SIZE)
 928                 goto out_fallback;
 929
 930         if (pipe_buf_steal(cs->pipe, buf) != 0)
 931                 goto out_fallback;
 932
 933         newpage = buf->page;
 934
 935         if (!PageUptodate(newpage))
 936                 SetPageUptodate(newpage);
 937
 938         ClearPageMappedToDisk(newpage);
 939
 940         if (fuse_check_page(newpage) != 0)
 941                 goto out_fallback_unlock;
 942
 943         /*
 944          * This is a new and locked page, it shouldn't be mapped or
 945          * have any special flags on it
 946          */
 947         if (WARN_ON(page_mapped(oldpage)))
 948                 goto out_fallback_unlock;
 949         if (WARN_ON(page_has_private(oldpage)))
 950                 goto out_fallback_unlock;
 951         if (WARN_ON(PageDirty(oldpage) || PageWriteback(oldpage)))
 952                 goto out_fallback_unlock;
 953         if (WARN_ON(PageMlocked(oldpage)))
 954                 goto out_fallback_unlock;
 955
 956         err = replace_page_cache_page(oldpage, newpage, GFP_KERNEL);
 957         if (err) {
 958                 unlock_page(newpage);
 959                 return err;
 960         }
 961
 962         get_page(newpage);
 963
 964         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 965                 lru_cache_add_file(newpage);
 966
 967         err = 0;
 968         spin_lock(&cs->req->waitq.lock);
 969         if (test_bit(FR_ABORTED, &cs->req->flags))
 970                 err = -ENOENT;
 971         else
 972                 *pagep = newpage;
 973         spin_unlock(&cs->req->waitq.lock);
 974
 975         if (err) {
 976                 unlock_page(newpage);
 977                 put_page(newpage);
 978                 return err;
 979         }
 980
 981         unlock_page(oldpage);
 982         put_page(oldpage);
 983         cs->len = 0;
 984
 985         return 0;
 986
 987 out_fallback_unlock:
 988         unlock_page(newpage);
 989 out_fallback:
 990         cs->pg = buf->page;
 991         cs->offset = buf->offset;
 992
 993         err = lock_request(cs->req);
 994         if (err)
 995                 return err;
 996
 997         return 1;
 998 }
 999
1000 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
1001                          unsigned offset, unsigned count)
1002 {
1003         struct pipe_buffer *buf;
1004         int err;
1005
1006         if (cs->nr_segs == cs->pipe->buffers)
1007                 return -EIO;
1008
1009         err = unlock_request(cs->req);
1010         if (err)
1011                 return err;
1012
1013         fuse_copy_finish(cs);
1014
1015         buf = cs->pipebufs;
1016         get_page(page);
1017         buf->page = page;
1018         buf->offset = offset;
1019         buf->len = count;
1020
1021         cs->pipebufs++;
1022         cs->nr_segs++;
1023         cs->len = 0;
1024
1025         return 0;
1026 }
1027
1028 /*
1029  * Copy a page in the request to/from the userspace buffer.  Must be
1030  * done atomically
1031  */
1032 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
1033                           unsigned offset, unsigned count, int zeroing)
1034 {
1035         int err;
1036         struct page *page = *pagep;
1037
1038         if (page && zeroing && count < PAGE_SIZE)
1039                 clear_highpage(page);
1040
1041         while (count) {
1042                 if (cs->write && cs->pipebufs && page) {
1043                         return fuse_ref_page(cs, page, offset, count);
1044                 } else if (!cs->len) {
1045                         if (cs->move_pages && page &&
1046                             offset == 0 && count == PAGE_SIZE) {
1047                                 err = fuse_try_move_page(cs, pagep);
1048                                 if (err <= 0)
1049                                         return err;
1050                         } else {
1051                                 err = fuse_copy_fill(cs);
1052                                 if (err)
1053                                         return err;
1054                         }
1055                 }
1056                 if (page) {
1057                         void *mapaddr = kmap_atomic(page);
1058                         void *buf = mapaddr + offset;
1059                         offset += fuse_copy_do(cs, &buf, &count);
1060                         kunmap_atomic(mapaddr);
1061                 } else
1062                         offset += fuse_copy_do(cs, NULL, &count);
1063         }
1064         if (page && !cs->write)
1065                 flush_dcache_page(page);
1066         return 0;
1067 }
1068
1069 /* Copy pages in the request to/from userspace buffer */
1070 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
1071                            int zeroing)
1072 {
1073         unsigned i;
1074         struct fuse_req *req = cs->req;
1075
1076         for (i = 0; i < req->num_pages && (nbytes || zeroing); i++) {
1077                 int err;
1078                 unsigned offset = req->page_descs[i].offset;
1079                 unsigned count = min(nbytes, req->page_descs[i].length);
1080
1081                 err = fuse_copy_page(cs, &req->pages[i], offset, count,
1082                                      zeroing);
1083                 if (err)
1084                         return err;
1085
1086                 nbytes -= count;
1087         }
1088         return 0;
1089 }
1090
1091 /* Copy a single argument in the request to/from userspace buffer */
1092 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1093 {
1094         while (size) {
1095                 if (!cs->len) {
1096                         int err = fuse_copy_fill(cs);
1097                         if (err)
1098                                 return err;
1099                 }
1100                 fuse_copy_do(cs, &val, &size);
1101         }
1102         return 0;
1103 }
1104
1105 /* Copy request arguments to/from userspace buffer */
1106 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1107                           unsigned argpages, struct fuse_arg *args,
1108                           int zeroing)
1109 {
1110         int err = 0;
1111         unsigned i;
1112
1113         for (i = 0; !err && i < numargs; i++)  {
1114                 struct fuse_arg *arg = &args[i];
1115                 if (i == numargs - 1 && argpages)
1116                         err = fuse_copy_pages(cs, arg->size, zeroing);
1117                 else
1118                         err = fuse_copy_one(cs, arg->value, arg->size);
1119         }
1120         return err;
1121 }
1122
1123 static int forget_pending(struct fuse_iqueue *fiq)
1124 {
1125         return fiq->forget_list_head.next != NULL;
1126 }
1127
1128 static int request_pending(struct fuse_iqueue *fiq)
1129 {
1130         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1131                 forget_pending(fiq);
1132 }
1133
1134 /*
1135  * Transfer an interrupt request to userspace
1136  *
1137  * Unlike other requests this is assembled on demand, without a need
1138  * to allocate a separate fuse_req structure.
1139  *
1140  * Called with fiq->waitq.lock held, releases it
1141  */
1142 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1143                                struct fuse_copy_state *cs,
1144                                size_t nbytes, struct fuse_req *req)
1145 __releases(fiq->waitq.lock)
1146 {
1147         struct fuse_in_header ih;
1148         struct fuse_interrupt_in arg;
1149         unsigned reqsize = sizeof(ih) + sizeof(arg);
1150         int err;
1151
1152         list_del_init(&req->intr_entry);
1153         memset(&ih, 0, sizeof(ih));
1154         memset(&arg, 0, sizeof(arg));
1155         ih.len = reqsize;
1156         ih.opcode = FUSE_INTERRUPT;
1157         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1158         arg.unique = req->in.h.unique;
1159
1160         spin_unlock(&fiq->waitq.lock);
1161         if (nbytes < reqsize)
1162                 return -EINVAL;
1163
1164         err = fuse_copy_one(cs, &ih, sizeof(ih));
1165         if (!err)
1166                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1167         fuse_copy_finish(cs);
1168
1169         return err ? err : reqsize;
1170 }
1171
1172 static struct fuse_forget_link *dequeue_forget(struct fuse_iqueue *fiq,
1173                                                unsigned max,
1174                                                unsigned *countp)
1175 {
1176         struct fuse_forget_link *head = fiq->forget_list_head.next;
1177         struct fuse_forget_link **newhead = &head;
1178         unsigned count;
1179
1180         for (count = 0; *newhead != NULL && count < max; count++)
1181                 newhead = &(*newhead)->next;
1182
1183         fiq->forget_list_head.next = *newhead;
1184         *newhead = NULL;
1185         if (fiq->forget_list_head.next == NULL)
1186                 fiq->forget_list_tail = &fiq->forget_list_head;
1187
1188         if (countp != NULL)
1189                 *countp = count;
1190
1191         return head;
1192 }
1193
1194 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1195                                    struct fuse_copy_state *cs,
1196                                    size_t nbytes)
1197 __releases(fiq->waitq.lock)
1198 {
1199         int err;
1200         struct fuse_forget_link *forget = dequeue_forget(fiq, 1, NULL);
1201         struct fuse_forget_in arg = {
1202                 .nlookup = forget->forget_one.nlookup,
1203         };
1204         struct fuse_in_header ih = {
1205                 .opcode = FUSE_FORGET,
1206                 .nodeid = forget->forget_one.nodeid,
1207                 .unique = fuse_get_unique(fiq),
1208                 .len = sizeof(ih) + sizeof(arg),
1209         };
1210
1211         spin_unlock(&fiq->waitq.lock);
1212         kfree(forget);
1213         if (nbytes < ih.len)
1214                 return -EINVAL;
1215
1216         err = fuse_copy_one(cs, &ih, sizeof(ih));
1217         if (!err)
1218                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1219         fuse_copy_finish(cs);
1220
1221         if (err)
1222                 return err;
1223
1224         return ih.len;
1225 }
1226
1227 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1228                                    struct fuse_copy_state *cs, size_t nbytes)
1229 __releases(fiq->waitq.lock)
1230 {
1231         int err;
1232         unsigned max_forgets;
1233         unsigned count;
1234         struct fuse_forget_link *head;
1235         struct fuse_batch_forget_in arg = { .count = 0 };
1236         struct fuse_in_header ih = {
1237                 .opcode = FUSE_BATCH_FORGET,
1238                 .unique = fuse_get_unique(fiq),
1239                 .len = sizeof(ih) + sizeof(arg),
1240         };
1241
1242         if (nbytes < ih.len) {
1243                 spin_unlock(&fiq->waitq.lock);
1244                 return -EINVAL;
1245         }
1246
1247         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1248         head = dequeue_forget(fiq, max_forgets, &count);
1249         spin_unlock(&fiq->waitq.lock);
1250
1251         arg.count = count;
1252         ih.len += count * sizeof(struct fuse_forget_one);
1253         err = fuse_copy_one(cs, &ih, sizeof(ih));
1254         if (!err)
1255                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1256
1257         while (head) {
1258                 struct fuse_forget_link *forget = head;
1259
1260                 if (!err) {
1261                         err = fuse_copy_one(cs, &forget->forget_one,
1262                                             sizeof(forget->forget_one));
1263                 }
1264                 head = forget->next;
1265                 kfree(forget);
1266         }
1267
1268         fuse_copy_finish(cs);
1269
1270         if (err)
1271                 return err;
1272
1273         return ih.len;
1274 }
1275
1276 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1277                             struct fuse_copy_state *cs,
1278                             size_t nbytes)
1279 __releases(fiq->waitq.lock)
1280 {
1281         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1282                 return fuse_read_single_forget(fiq, cs, nbytes);
1283         else
1284                 return fuse_read_batch_forget(fiq, cs, nbytes);
1285 }
1286
1287 /*
1288  * Read a single request into the userspace filesystem's buffer.  This
1289  * function waits until a request is available, then removes it from
1290  * the pending list and copies request data to userspace buffer.  If
1291  * no reply is needed (FORGET) or request has been aborted or there
1292  * was an error during the copying then it's finished by calling
1293  * request_end().  Otherwise add it to the processing list, and set
1294  * the 'sent' flag.
1295  */
1296 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1297                                 struct fuse_copy_state *cs, size_t nbytes)
1298 {
1299         ssize_t err;
1300         struct fuse_conn *fc = fud->fc;
1301         struct fuse_iqueue *fiq = &fc->iq;
1302         struct fuse_pqueue *fpq = &fud->pq;
1303         struct fuse_req *req;
1304         struct fuse_in *in;
1305         unsigned reqsize;
1306         unsigned int hash;
1307
1308  restart:
1309         spin_lock(&fiq->waitq.lock);
1310         err = -EAGAIN;
1311         if ((file->f_flags & O_NONBLOCK) && fiq->connected &&
1312             !request_pending(fiq))
1313                 goto err_unlock;
1314
1315         err = wait_event_interruptible_exclusive_locked(fiq->waitq,
1316                                 !fiq->connected || request_pending(fiq));
1317         if (err)
1318                 goto err_unlock;
1319
1320         if (!fiq->connected) {
1321                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1322                 goto err_unlock;
1323         }
1324
1325         if (!list_empty(&fiq->interrupts)) {
1326                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1327                                  intr_entry);
1328                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1329         }
1330
1331         if (forget_pending(fiq)) {
1332                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1333                         return fuse_read_forget(fc, fiq, cs, nbytes);
1334
1335                 if (fiq->forget_batch <= -8)
1336                         fiq->forget_batch = 16;
1337         }
1338
1339         req = list_entry(fiq->pending.next, struct fuse_req, list);
1340         clear_bit(FR_PENDING, &req->flags);
1341         list_del_init(&req->list);
1342         spin_unlock(&fiq->waitq.lock);
1343
1344         in = &req->in;
1345         reqsize = in->h.len;
1346
1347         /* If request is too large, reply with an error and restart the read */
1348         if (nbytes < reqsize) {
1349                 req->out.h.error = -EIO;
1350                 /* SETXATTR is special, since it may contain too large data */
1351                 if (in->h.opcode == FUSE_SETXATTR)
1352                         req->out.h.error = -E2BIG;
1353                 request_end(fc, req);
1354                 goto restart;
1355         }
1356         spin_lock(&fpq->lock);
1357         list_add(&req->list, &fpq->io);
1358         spin_unlock(&fpq->lock);
1359         cs->req = req;
1360         err = fuse_copy_one(cs, &in->h, sizeof(in->h));
1361         if (!err)
1362                 err = fuse_copy_args(cs, in->numargs, in->argpages,
1363                                      (struct fuse_arg *) in->args, 0);
1364         fuse_copy_finish(cs);
1365         spin_lock(&fpq->lock);
1366         clear_bit(FR_LOCKED, &req->flags);
1367         if (!fpq->connected) {
1368                 err = (fc->aborted && fc->abort_err) ? -ECONNABORTED : -ENODEV;
1369                 goto out_end;
1370         }
1371         if (err) {
1372                 req->out.h.error = -EIO;
1373                 goto out_end;
1374         }
1375         if (!test_bit(FR_ISREPLY, &req->flags)) {
1376                 err = reqsize;
1377                 goto out_end;
1378         }
1379         hash = fuse_req_hash(req->in.h.unique);
1380         list_move_tail(&req->list, &fpq->processing[hash]);
1381         __fuse_get_request(req);
1382         set_bit(FR_SENT, &req->flags);
1383         spin_unlock(&fpq->lock);
1384         /* matches barrier in request_wait_answer() */
1385         smp_mb__after_atomic();
1386         if (test_bit(FR_INTERRUPTED, &req->flags))
1387                 queue_interrupt(fiq, req);
1388         fuse_put_request(fc, req);
1389
1390         return reqsize;
1391
1392 out_end:
1393         if (!test_bit(FR_PRIVATE, &req->flags))
1394                 list_del_init(&req->list);
1395         spin_unlock(&fpq->lock);
1396         request_end(fc, req);
1397         return err;
1398
1399  err_unlock:
1400         spin_unlock(&fiq->waitq.lock);
1401         return err;
1402 }
1403
1404 static int fuse_dev_open(struct inode *inode, struct file *file)
1405 {
1406         /*
1407          * The fuse device's file's private_data is used to hold
1408          * the fuse_conn(ection) when it is mounted, and is used to
1409          * keep track of whether the file has been mounted already.
1410          */
1411         file->private_data = NULL;
1412         return 0;
1413 }
1414
1415 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1416 {
1417         struct fuse_copy_state cs;
1418         struct file *file = iocb->ki_filp;
1419         struct fuse_dev *fud = fuse_get_dev(file);
1420
1421         if (!fud)
1422                 return -EPERM;
1423
1424         if (!iter_is_iovec(to))
1425                 return -EINVAL;
1426
1427         fuse_copy_init(&cs, 1, to);
1428
1429         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1430 }
1431
1432 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1433                                     struct pipe_inode_info *pipe,
1434                                     size_t len, unsigned int flags)
1435 {
1436         int total, ret;
1437         int page_nr = 0;
1438         struct pipe_buffer *bufs;
1439         struct fuse_copy_state cs;
1440         struct fuse_dev *fud = fuse_get_dev(in);
1441
1442         if (!fud)
1443                 return -EPERM;
1444
1445         bufs = kvmalloc_array(pipe->buffers, sizeof(struct pipe_buffer),
1446                               GFP_KERNEL);
1447         if (!bufs)
1448                 return -ENOMEM;
1449
1450         fuse_copy_init(&cs, 1, NULL);
1451         cs.pipebufs = bufs;
1452         cs.pipe = pipe;
1453         ret = fuse_dev_do_read(fud, in, &cs, len);
1454         if (ret < 0)
1455                 goto out;
1456
1457         if (pipe->nrbufs + cs.nr_segs > pipe->buffers) {
1458                 ret = -EIO;
1459                 goto out;
1460         }
1461
1462         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1463                 /*
1464                  * Need to be careful about this.  Having buf->ops in module
1465                  * code can Oops if the buffer persists after module unload.
1466                  */
1467                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1468                 bufs[page_nr].flags = 0;
1469                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1470                 if (unlikely(ret < 0))
1471                         break;
1472         }
1473         if (total)
1474                 ret = total;
1475 out:
1476         for (; page_nr < cs.nr_segs; page_nr++)
1477                 put_page(bufs[page_nr].page);
1478
1479         kvfree(bufs);
1480         return ret;
1481 }
1482
1483 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1484                             struct fuse_copy_state *cs)
1485 {
1486         struct fuse_notify_poll_wakeup_out outarg;
1487         int err = -EINVAL;
1488
1489         if (size != sizeof(outarg))
1490                 goto err;
1491
1492         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1493         if (err)
1494                 goto err;
1495
1496         fuse_copy_finish(cs);
1497         return fuse_notify_poll_wakeup(fc, &outarg);
1498
1499 err:
1500         fuse_copy_finish(cs);
1501         return err;
1502 }
1503
1504 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1505                                    struct fuse_copy_state *cs)
1506 {
1507         struct fuse_notify_inval_inode_out outarg;
1508         int err = -EINVAL;
1509
1510         if (size != sizeof(outarg))
1511                 goto err;
1512
1513         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1514         if (err)
1515                 goto err;
1516         fuse_copy_finish(cs);
1517
1518         down_read(&fc->killsb);
1519         err = -ENOENT;
1520         if (fc->sb) {
1521                 err = fuse_reverse_inval_inode(fc->sb, outarg.ino,
1522                                                outarg.off, outarg.len);
1523         }
1524         up_read(&fc->killsb);
1525         return err;
1526
1527 err:
1528         fuse_copy_finish(cs);
1529         return err;
1530 }
1531
1532 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1533                                    struct fuse_copy_state *cs)
1534 {
1535         struct fuse_notify_inval_entry_out outarg;
1536         int err = -ENOMEM;
1537         char *buf;
1538         struct qstr name;
1539
1540         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1541         if (!buf)
1542                 goto err;
1543
1544         err = -EINVAL;
1545         if (size < sizeof(outarg))
1546                 goto err;
1547
1548         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1549         if (err)
1550                 goto err;
1551
1552         err = -ENAMETOOLONG;
1553         if (outarg.namelen > FUSE_NAME_MAX)
1554                 goto err;
1555
1556         err = -EINVAL;
1557         if (size != sizeof(outarg) + outarg.namelen + 1)
1558                 goto err;
1559
1560         name.name = buf;
1561         name.len = outarg.namelen;
1562         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1563         if (err)
1564                 goto err;
1565         fuse_copy_finish(cs);
1566         buf[outarg.namelen] = 0;
1567
1568         down_read(&fc->killsb);
1569         err = -ENOENT;
1570         if (fc->sb)
1571                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent, 0, &name);
1572         up_read(&fc->killsb);
1573         kfree(buf);
1574         return err;
1575
1576 err:
1577         kfree(buf);
1578         fuse_copy_finish(cs);
1579         return err;
1580 }
1581
1582 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1583                               struct fuse_copy_state *cs)
1584 {
1585         struct fuse_notify_delete_out outarg;
1586         int err = -ENOMEM;
1587         char *buf;
1588         struct qstr name;
1589
1590         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1591         if (!buf)
1592                 goto err;
1593
1594         err = -EINVAL;
1595         if (size < sizeof(outarg))
1596                 goto err;
1597
1598         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1599         if (err)
1600                 goto err;
1601
1602         err = -ENAMETOOLONG;
1603         if (outarg.namelen > FUSE_NAME_MAX)
1604                 goto err;
1605
1606         err = -EINVAL;
1607         if (size != sizeof(outarg) + outarg.namelen + 1)
1608                 goto err;
1609
1610         name.name = buf;
1611         name.len = outarg.namelen;
1612         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1613         if (err)
1614                 goto err;
1615         fuse_copy_finish(cs);
1616         buf[outarg.namelen] = 0;
1617
1618         down_read(&fc->killsb);
1619         err = -ENOENT;
1620         if (fc->sb)
1621                 err = fuse_reverse_inval_entry(fc->sb, outarg.parent,
1622                                                outarg.child, &name);
1623         up_read(&fc->killsb);
1624         kfree(buf);
1625         return err;
1626
1627 err:
1628         kfree(buf);
1629         fuse_copy_finish(cs);
1630         return err;
1631 }
1632
1633 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1634                              struct fuse_copy_state *cs)
1635 {
1636         struct fuse_notify_store_out outarg;
1637         struct inode *inode;
1638         struct address_space *mapping;
1639         u64 nodeid;
1640         int err;
1641         pgoff_t index;
1642         unsigned int offset;
1643         unsigned int num;
1644         loff_t file_size;
1645         loff_t end;
1646
1647         err = -EINVAL;
1648         if (size < sizeof(outarg))
1649                 goto out_finish;
1650
1651         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1652         if (err)
1653                 goto out_finish;
1654
1655         err = -EINVAL;
1656         if (size - sizeof(outarg) != outarg.size)
1657                 goto out_finish;
1658
1659         nodeid = outarg.nodeid;
1660
1661         down_read(&fc->killsb);
1662
1663         err = -ENOENT;
1664         if (!fc->sb)
1665                 goto out_up_killsb;
1666
1667         inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1668         if (!inode)
1669                 goto out_up_killsb;
1670
1671         mapping = inode->i_mapping;
1672         index = outarg.offset >> PAGE_SHIFT;
1673         offset = outarg.offset & ~PAGE_MASK;
1674         file_size = i_size_read(inode);
1675         end = outarg.offset + outarg.size;
1676         if (end > file_size) {
1677                 file_size = end;
1678                 fuse_write_update_size(inode, file_size);
1679         }
1680
1681         num = outarg.size;
1682         while (num) {
1683                 struct page *page;
1684                 unsigned int this_num;
1685
1686                 err = -ENOMEM;
1687                 page = find_or_create_page(mapping, index,
1688                                            mapping_gfp_mask(mapping));
1689                 if (!page)
1690                         goto out_iput;
1691
1692                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1693                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1694                 if (!err && offset == 0 &&
1695                     (this_num == PAGE_SIZE || file_size == end))
1696                         SetPageUptodate(page);
1697                 unlock_page(page);
1698                 put_page(page);
1699
1700                 if (err)
1701                         goto out_iput;
1702
1703                 num -= this_num;
1704                 offset = 0;
1705                 index++;
1706         }
1707
1708         err = 0;
1709
1710 out_iput:
1711         iput(inode);
1712 out_up_killsb:
1713         up_read(&fc->killsb);
1714 out_finish:
1715         fuse_copy_finish(cs);
1716         return err;
1717 }
1718
1719 static void fuse_retrieve_end(struct fuse_conn *fc, struct fuse_req *req)
1720 {
1721         release_pages(req->pages, req->num_pages);
1722 }
1723
1724 static int fuse_retrieve(struct fuse_conn *fc, struct inode *inode,
1725                          struct fuse_notify_retrieve_out *outarg)
1726 {
1727         int err;
1728         struct address_space *mapping = inode->i_mapping;
1729         struct fuse_req *req;
1730         pgoff_t index;
1731         loff_t file_size;
1732         unsigned int num;
1733         unsigned int offset;
1734         size_t total_len = 0;
1735         unsigned int num_pages;
1736
1737         offset = outarg->offset & ~PAGE_MASK;
1738         file_size = i_size_read(inode);
1739
1740         num = outarg->size;
1741         if (outarg->offset > file_size)
1742                 num = 0;
1743         else if (outarg->offset + num > file_size)
1744                 num = file_size - outarg->offset;
1745
1746         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1747         num_pages = min(num_pages, fc->max_pages);
1748
1749         req = fuse_get_req(fc, num_pages);
1750         if (IS_ERR(req))
1751                 return PTR_ERR(req);
1752
1753         req->in.h.opcode = FUSE_NOTIFY_REPLY;
1754         req->in.h.nodeid = outarg->nodeid;
1755         req->in.numargs = 2;
1756         req->in.argpages = 1;
1757         req->end = fuse_retrieve_end;
1758
1759         index = outarg->offset >> PAGE_SHIFT;
1760
1761         while (num && req->num_pages < num_pages) {
1762                 struct page *page;
1763                 unsigned int this_num;
1764
1765                 page = find_get_page(mapping, index);
1766                 if (!page)
1767                         break;
1768
1769                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1770                 req->pages[req->num_pages] = page;
1771                 req->page_descs[req->num_pages].offset = offset;
1772                 req->page_descs[req->num_pages].length = this_num;
1773                 req->num_pages++;
1774
1775                 offset = 0;
1776                 num -= this_num;
1777                 total_len += this_num;
1778                 index++;
1779         }
1780         req->misc.retrieve_in.offset = outarg->offset;
1781         req->misc.retrieve_in.size = total_len;
1782         req->in.args[0].size = sizeof(req->misc.retrieve_in);
1783         req->in.args[0].value = &req->misc.retrieve_in;
1784         req->in.args[1].size = total_len;
1785
1786         err = fuse_request_send_notify_reply(fc, req, outarg->notify_unique);
1787         if (err) {
1788                 fuse_retrieve_end(fc, req);
1789                 fuse_put_request(fc, req);
1790         }
1791
1792         return err;
1793 }
1794
1795 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1796                                 struct fuse_copy_state *cs)
1797 {
1798         struct fuse_notify_retrieve_out outarg;
1799         struct inode *inode;
1800         int err;
1801
1802         err = -EINVAL;
1803         if (size != sizeof(outarg))
1804                 goto copy_finish;
1805
1806         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1807         if (err)
1808                 goto copy_finish;
1809
1810         fuse_copy_finish(cs);
1811
1812         down_read(&fc->killsb);
1813         err = -ENOENT;
1814         if (fc->sb) {
1815                 u64 nodeid = outarg.nodeid;
1816
1817                 inode = ilookup5(fc->sb, nodeid, fuse_inode_eq, &nodeid);
1818                 if (inode) {
1819                         err = fuse_retrieve(fc, inode, &outarg);
1820                         iput(inode);
1821                 }
1822         }
1823         up_read(&fc->killsb);
1824
1825         return err;
1826
1827 copy_finish:
1828         fuse_copy_finish(cs);
1829         return err;
1830 }
1831
1832 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1833                        unsigned int size, struct fuse_copy_state *cs)
1834 {
1835         /* Don't try to move pages (yet) */
1836         cs->move_pages = 0;
1837
1838         switch (code) {
1839         case FUSE_NOTIFY_POLL:
1840                 return fuse_notify_poll(fc, size, cs);
1841
1842         case FUSE_NOTIFY_INVAL_INODE:
1843                 return fuse_notify_inval_inode(fc, size, cs);
1844
1845         case FUSE_NOTIFY_INVAL_ENTRY:
1846                 return fuse_notify_inval_entry(fc, size, cs);
1847
1848         case FUSE_NOTIFY_STORE:
1849                 return fuse_notify_store(fc, size, cs);
1850
1851         case FUSE_NOTIFY_RETRIEVE:
1852                 return fuse_notify_retrieve(fc, size, cs);
1853
1854         case FUSE_NOTIFY_DELETE:
1855                 return fuse_notify_delete(fc, size, cs);
1856
1857         default:
1858                 fuse_copy_finish(cs);
1859                 return -EINVAL;
1860         }
1861 }
1862
1863 /* Look up request on processing list by unique ID */
1864 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1865 {
1866         unsigned int hash = fuse_req_hash(unique);
1867         struct fuse_req *req;
1868
1869         list_for_each_entry(req, &fpq->processing[hash], list) {
1870                 if (req->in.h.unique == unique)
1871                         return req;
1872         }
1873         return NULL;
1874 }
1875
1876 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_out *out,
1877                          unsigned nbytes)
1878 {
1879         unsigned reqsize = sizeof(struct fuse_out_header);
1880
1881         if (out->h.error)
1882                 return nbytes != reqsize ? -EINVAL : 0;
1883
1884         reqsize += len_args(out->numargs, out->args);
1885
1886         if (reqsize < nbytes || (reqsize > nbytes && !out->argvar))
1887                 return -EINVAL;
1888         else if (reqsize > nbytes) {
1889                 struct fuse_arg *lastarg = &out->args[out->numargs-1];
1890                 unsigned diffsize = reqsize - nbytes;
1891                 if (diffsize > lastarg->size)
1892                         return -EINVAL;
1893                 lastarg->size -= diffsize;
1894         }
1895         return fuse_copy_args(cs, out->numargs, out->argpages, out->args,
1896                               out->page_zeroing);
1897 }
1898
1899 /*
1900  * Write a single reply to a request.  First the header is copied from
1901  * the write buffer.  The request is then searched on the processing
1902  * list by the unique ID found in the header.  If found, then remove
1903  * it from the list and copy the rest of the buffer to the request.
1904  * The request is finished by calling request_end()
1905  */
1906 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1907                                  struct fuse_copy_state *cs, size_t nbytes)
1908 {
1909         int err;
1910         struct fuse_conn *fc = fud->fc;
1911         struct fuse_pqueue *fpq = &fud->pq;
1912         struct fuse_req *req;
1913         struct fuse_out_header oh;
1914
1915         if (nbytes < sizeof(struct fuse_out_header))
1916                 return -EINVAL;
1917
1918         err = fuse_copy_one(cs, &oh, sizeof(oh));
1919         if (err)
1920                 goto err_finish;
1921
1922         err = -EINVAL;
1923         if (oh.len != nbytes)
1924                 goto err_finish;
1925
1926         /*
1927          * Zero oh.unique indicates unsolicited notification message
1928          * and error contains notification code.
1929          */
1930         if (!oh.unique) {
1931                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1932                 return err ? err : nbytes;
1933         }
1934
1935         err = -EINVAL;
1936         if (oh.error <= -1000 || oh.error > 0)
1937                 goto err_finish;
1938
1939         spin_lock(&fpq->lock);
1940         err = -ENOENT;
1941         if (!fpq->connected)
1942                 goto err_unlock_pq;
1943
1944         req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1945         if (!req)
1946                 goto err_unlock_pq;
1947
1948         /* Is it an interrupt reply ID? */
1949         if (oh.unique & FUSE_INT_REQ_BIT) {
1950                 __fuse_get_request(req);
1951                 spin_unlock(&fpq->lock);
1952
1953                 err = -EINVAL;
1954                 if (nbytes != sizeof(struct fuse_out_header)) {
1955                         fuse_put_request(fc, req);
1956                         goto err_finish;
1957                 }
1958
1959                 if (oh.error == -ENOSYS)
1960                         fc->no_interrupt = 1;
1961                 else if (oh.error == -EAGAIN)
1962                         queue_interrupt(&fc->iq, req);
1963                 fuse_put_request(fc, req);
1964
1965                 fuse_copy_finish(cs);
1966                 return nbytes;
1967         }
1968
1969         clear_bit(FR_SENT, &req->flags);
1970         list_move(&req->list, &fpq->io);
1971         req->out.h = oh;
1972         set_bit(FR_LOCKED, &req->flags);
1973         spin_unlock(&fpq->lock);
1974         cs->req = req;
1975         if (!req->out.page_replace)
1976                 cs->move_pages = 0;
1977
1978         err = copy_out_args(cs, &req->out, nbytes);
1979         fuse_copy_finish(cs);
1980
1981         spin_lock(&fpq->lock);
1982         clear_bit(FR_LOCKED, &req->flags);
1983         if (!fpq->connected)
1984                 err = -ENOENT;
1985         else if (err)
1986                 req->out.h.error = -EIO;
1987         if (!test_bit(FR_PRIVATE, &req->flags))
1988                 list_del_init(&req->list);
1989         spin_unlock(&fpq->lock);
1990
1991         request_end(fc, req);
1992
1993         return err ? err : nbytes;
1994
1995  err_unlock_pq:
1996         spin_unlock(&fpq->lock);
1997  err_finish:
1998         fuse_copy_finish(cs);
1999         return err;
2000 }
2001
2002 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
2003 {
2004         struct fuse_copy_state cs;
2005         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
2006
2007         if (!fud)
2008                 return -EPERM;
2009
2010         if (!iter_is_iovec(from))
2011                 return -EINVAL;
2012
2013         fuse_copy_init(&cs, 0, from);
2014
2015         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2016 }
2017
2018 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2019                                      struct file *out, loff_t *ppos,
2020                                      size_t len, unsigned int flags)
2021 {
2022         unsigned nbuf;
2023         unsigned idx;
2024         struct pipe_buffer *bufs;
2025         struct fuse_copy_state cs;
2026         struct fuse_dev *fud;
2027         size_t rem;
2028         ssize_t ret;
2029
2030         fud = fuse_get_dev(out);
2031         if (!fud)
2032                 return -EPERM;
2033
2034         pipe_lock(pipe);
2035
2036         bufs = kvmalloc_array(pipe->nrbufs, sizeof(struct pipe_buffer),
2037                               GFP_KERNEL);
2038         if (!bufs) {
2039                 pipe_unlock(pipe);
2040                 return -ENOMEM;
2041         }
2042
2043         nbuf = 0;
2044         rem = 0;
2045         for (idx = 0; idx < pipe->nrbufs && rem < len; idx++)
2046                 rem += pipe->bufs[(pipe->curbuf + idx) & (pipe->buffers - 1)].len;
2047
2048         ret = -EINVAL;
2049         if (rem < len) {
2050                 pipe_unlock(pipe);
2051                 goto out;
2052         }
2053
2054         rem = len;
2055         while (rem) {
2056                 struct pipe_buffer *ibuf;
2057                 struct pipe_buffer *obuf;
2058
2059                 BUG_ON(nbuf >= pipe->buffers);
2060                 BUG_ON(!pipe->nrbufs);
2061                 ibuf = &pipe->bufs[pipe->curbuf];
2062                 obuf = &bufs[nbuf];
2063
2064                 if (rem >= ibuf->len) {
2065                         *obuf = *ibuf;
2066                         ibuf->ops = NULL;
2067                         pipe->curbuf = (pipe->curbuf + 1) & (pipe->buffers - 1);
2068                         pipe->nrbufs--;
2069                 } else {
2070                         pipe_buf_get(pipe, ibuf);
2071                         *obuf = *ibuf;
2072                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2073                         obuf->len = rem;
2074                         ibuf->offset += obuf->len;
2075                         ibuf->len -= obuf->len;
2076                 }
2077                 nbuf++;
2078                 rem -= obuf->len;
2079         }
2080         pipe_unlock(pipe);
2081
2082         fuse_copy_init(&cs, 0, NULL);
2083         cs.pipebufs = bufs;
2084         cs.nr_segs = nbuf;
2085         cs.pipe = pipe;
2086
2087         if (flags & SPLICE_F_MOVE)
2088                 cs.move_pages = 1;
2089
2090         ret = fuse_dev_do_write(fud, &cs, len);
2091
2092         pipe_lock(pipe);
2093         for (idx = 0; idx < nbuf; idx++)
2094                 pipe_buf_release(pipe, &bufs[idx]);
2095         pipe_unlock(pipe);
2096
2097 out:
2098         kvfree(bufs);
2099         return ret;
2100 }
2101
2102 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2103 {
2104         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2105         struct fuse_iqueue *fiq;
2106         struct fuse_dev *fud = fuse_get_dev(file);
2107
2108         if (!fud)
2109                 return EPOLLERR;
2110
2111         fiq = &fud->fc->iq;
2112         poll_wait(file, &fiq->waitq, wait);
2113
2114         spin_lock(&fiq->waitq.lock);
2115         if (!fiq->connected)
2116                 mask = EPOLLERR;
2117         else if (request_pending(fiq))
2118                 mask |= EPOLLIN | EPOLLRDNORM;
2119         spin_unlock(&fiq->waitq.lock);
2120
2121         return mask;
2122 }
2123
2124 /* Abort all requests on the given list (pending or processing) */
2125 static void end_requests(struct fuse_conn *fc, struct list_head *head)
2126 {
2127         while (!list_empty(head)) {
2128                 struct fuse_req *req;
2129                 req = list_entry(head->next, struct fuse_req, list);
2130                 req->out.h.error = -ECONNABORTED;
2131                 clear_bit(FR_SENT, &req->flags);
2132                 list_del_init(&req->list);
2133                 request_end(fc, req);
2134         }
2135 }
2136
2137 static void end_polls(struct fuse_conn *fc)
2138 {
2139         struct rb_node *p;
2140
2141         p = rb_first(&fc->polled_files);
2142
2143         while (p) {
2144                 struct fuse_file *ff;
2145                 ff = rb_entry(p, struct fuse_file, polled_node);
2146                 wake_up_interruptible_all(&ff->poll_wait);
2147
2148                 p = rb_next(p);
2149         }
2150 }
2151
2152 /*
2153  * Abort all requests.
2154  *
2155  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2156  * filesystem.
2157  *
2158  * The same effect is usually achievable through killing the filesystem daemon
2159  * and all users of the filesystem.  The exception is the combination of an
2160  * asynchronous request and the tricky deadlock (see
2161  * Documentation/filesystems/fuse.txt).
2162  *
2163  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2164  * requests, they should be finished off immediately.  Locked requests will be
2165  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2166  * requests.  It is possible that some request will finish before we can.  This
2167  * is OK, the request will in that case be removed from the list before we touch
2168  * it.
2169  */
2170 void fuse_abort_conn(struct fuse_conn *fc, bool is_abort)
2171 {
2172         struct fuse_iqueue *fiq = &fc->iq;
2173
2174         spin_lock(&fc->lock);
2175         if (fc->connected) {
2176                 struct fuse_dev *fud;
2177                 struct fuse_req *req, *next;
2178                 LIST_HEAD(to_end);
2179                 unsigned int i;
2180
2181                 /* Background queuing checks fc->connected under bg_lock */
2182                 spin_lock(&fc->bg_lock);
2183                 fc->connected = 0;
2184                 spin_unlock(&fc->bg_lock);
2185
2186                 fc->aborted = is_abort;
2187                 fuse_set_initialized(fc);
2188                 list_for_each_entry(fud, &fc->devices, entry) {
2189                         struct fuse_pqueue *fpq = &fud->pq;
2190
2191                         spin_lock(&fpq->lock);
2192                         fpq->connected = 0;
2193                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2194                                 req->out.h.error = -ECONNABORTED;
2195                                 spin_lock(&req->waitq.lock);
2196                                 set_bit(FR_ABORTED, &req->flags);
2197                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2198                                         set_bit(FR_PRIVATE, &req->flags);
2199                                         __fuse_get_request(req);
2200                                         list_move(&req->list, &to_end);
2201                                 }
2202                                 spin_unlock(&req->waitq.lock);
2203                         }
2204                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2205                                 list_splice_tail_init(&fpq->processing[i],
2206                                                       &to_end);
2207                         spin_unlock(&fpq->lock);
2208                 }
2209                 spin_lock(&fc->bg_lock);
2210                 fc->blocked = 0;
2211                 fc->max_background = UINT_MAX;
2212                 flush_bg_queue(fc);
2213                 spin_unlock(&fc->bg_lock);
2214
2215                 spin_lock(&fiq->waitq.lock);
2216                 fiq->connected = 0;
2217                 list_for_each_entry(req, &fiq->pending, list)
2218                         clear_bit(FR_PENDING, &req->flags);
2219                 list_splice_tail_init(&fiq->pending, &to_end);
2220                 while (forget_pending(fiq))
2221                         kfree(dequeue_forget(fiq, 1, NULL));
2222                 wake_up_all_locked(&fiq->waitq);
2223                 spin_unlock(&fiq->waitq.lock);
2224                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2225                 end_polls(fc);
2226                 wake_up_all(&fc->blocked_waitq);
2227                 spin_unlock(&fc->lock);
2228
2229                 end_requests(fc, &to_end);
2230         } else {
2231                 spin_unlock(&fc->lock);
2232         }
2233 }
2234 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2235
2236 void fuse_wait_aborted(struct fuse_conn *fc)
2237 {
2238         /* matches implicit memory barrier in fuse_drop_waiting() */
2239         smp_mb();
2240         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2241 }
2242
2243 int fuse_dev_release(struct inode *inode, struct file *file)
2244 {
2245         struct fuse_dev *fud = fuse_get_dev(file);
2246
2247         if (fud) {
2248                 struct fuse_conn *fc = fud->fc;
2249                 struct fuse_pqueue *fpq = &fud->pq;
2250                 LIST_HEAD(to_end);
2251                 unsigned int i;
2252
2253                 spin_lock(&fpq->lock);
2254                 WARN_ON(!list_empty(&fpq->io));
2255                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2256                         list_splice_init(&fpq->processing[i], &to_end);
2257                 spin_unlock(&fpq->lock);
2258
2259                 end_requests(fc, &to_end);
2260
2261                 /* Are we the last open device? */
2262                 if (atomic_dec_and_test(&fc->dev_count)) {
2263                         WARN_ON(fc->iq.fasync != NULL);
2264                         fuse_abort_conn(fc, false);
2265                 }
2266                 fuse_dev_free(fud);
2267         }
2268         return 0;
2269 }
2270 EXPORT_SYMBOL_GPL(fuse_dev_release);
2271
2272 static int fuse_dev_fasync(int fd, struct file *file, int on)
2273 {
2274         struct fuse_dev *fud = fuse_get_dev(file);
2275
2276         if (!fud)
2277                 return -EPERM;
2278
2279         /* No locking - fasync_helper does its own locking */
2280         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2281 }
2282
2283 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2284 {
2285         struct fuse_dev *fud;
2286
2287         if (new->private_data)
2288                 return -EINVAL;
2289
2290         fud = fuse_dev_alloc(fc);
2291         if (!fud)
2292                 return -ENOMEM;
2293
2294         new->private_data = fud;
2295         atomic_inc(&fc->dev_count);
2296
2297         return 0;
2298 }
2299
2300 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2301                            unsigned long arg)
2302 {
2303         int err = -ENOTTY;
2304
2305         if (cmd == FUSE_DEV_IOC_CLONE) {
2306                 int oldfd;
2307
2308                 err = -EFAULT;
2309                 if (!get_user(oldfd, (__u32 __user *) arg)) {
2310                         struct file *old = fget(oldfd);
2311
2312                         err = -EINVAL;
2313                         if (old) {
2314                                 struct fuse_dev *fud = NULL;
2315
2316                                 /*
2317                                  * Check against file->f_op because CUSE
2318                                  * uses the same ioctl handler.
2319                                  */
2320                                 if (old->f_op == file->f_op &&
2321                                     old->f_cred->user_ns == file->f_cred->user_ns)
2322                                         fud = fuse_get_dev(old);
2323
2324                                 if (fud) {
2325                                         mutex_lock(&fuse_mutex);
2326                                         err = fuse_device_clone(fud->fc, file);
2327                                         mutex_unlock(&fuse_mutex);
2328                                 }
2329                                 fput(old);
2330                         }
2331                 }
2332         }
2333         return err;
2334 }
2335
2336 const struct file_operations fuse_dev_operations = {
2337         .owner          = THIS_MODULE,
2338         .open           = fuse_dev_open,
2339         .llseek         = no_llseek,
2340         .read_iter      = fuse_dev_read,
2341         .splice_read    = fuse_dev_splice_read,
2342         .write_iter     = fuse_dev_write,
2343         .splice_write   = fuse_dev_splice_write,
2344         .poll           = fuse_dev_poll,
2345         .release        = fuse_dev_release,
2346         .fasync         = fuse_dev_fasync,
2347         .unlocked_ioctl = fuse_dev_ioctl,
2348         .compat_ioctl   = fuse_dev_ioctl,
2349 };
2350 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2351
2352 static struct miscdevice fuse_miscdevice = {
2353         .minor = FUSE_MINOR,
2354         .name  = "fuse",
2355         .fops = &fuse_dev_operations,
2356 };
2357
2358 int __init fuse_dev_init(void)
2359 {
2360         int err = -ENOMEM;
2361         fuse_req_cachep = kmem_cache_create("fuse_request",
2362                                             sizeof(struct fuse_req),
2363                                             0, 0, NULL);
2364         if (!fuse_req_cachep)
2365                 goto out;
2366
2367         err = misc_register(&fuse_miscdevice);
2368         if (err)
2369                 goto out_cache_clean;
2370
2371         return 0;
2372
2373  out_cache_clean:
2374         kmem_cache_destroy(fuse_req_cachep);
2375  out:
2376         return err;
2377 }
2378
2379 void fuse_dev_cleanup(void)
2380 {
2381         misc_deregister(&fuse_miscdevice);
2382         kmem_cache_destroy(fuse_req_cachep);
2383 }