fs/fuse/dev.c

   1 /*
   2   FUSE: Filesystem in Userspace
   3   Copyright (C) 2001-2008  Miklos Szeredi <[email protected]>
   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_mount *fm, struct fuse_req *req)
  44 {
  45         INIT_LIST_HEAD(&req->list);
  46         INIT_LIST_HEAD(&req->intr_entry);
  47         init_waitqueue_head(&req->waitq);
  48         refcount_set(&req->count, 1);
  49         __set_bit(FR_PENDING, &req->flags);
  50         req->fm = fm;
  51 }
  52
  53 static struct fuse_req *fuse_request_alloc(struct fuse_mount *fm, gfp_t flags)
  54 {
  55         struct fuse_req *req = kmem_cache_zalloc(fuse_req_cachep, flags);
  56         if (req)
  57                 fuse_request_init(fm, req);
  58
  59         return req;
  60 }
  61
  62 static void fuse_request_free(struct fuse_req *req)
  63 {
  64         kmem_cache_free(fuse_req_cachep, req);
  65 }
  66
  67 static void __fuse_get_request(struct fuse_req *req)
  68 {
  69         refcount_inc(&req->count);
  70 }
  71
  72 /* Must be called with > 1 refcount */
  73 static void __fuse_put_request(struct fuse_req *req)
  74 {
  75         refcount_dec(&req->count);
  76 }
  77
  78 void fuse_set_initialized(struct fuse_conn *fc)
  79 {
  80         /* Make sure stores before this are seen on another CPU */
  81         smp_wmb();
  82         fc->initialized = 1;
  83 }
  84
  85 static bool fuse_block_alloc(struct fuse_conn *fc, bool for_background)
  86 {
  87         return !fc->initialized || (for_background && fc->blocked);
  88 }
  89
  90 static void fuse_drop_waiting(struct fuse_conn *fc)
  91 {
  92         /*
  93          * lockess check of fc->connected is okay, because atomic_dec_and_test()
  94          * provides a memory barrier matched with the one in fuse_wait_aborted()
  95          * to ensure no wake-up is missed.
  96          */
  97         if (atomic_dec_and_test(&fc->num_waiting) &&
  98             !READ_ONCE(fc->connected)) {
  99                 /* wake up aborters */
 100                 wake_up_all(&fc->blocked_waitq);
 101         }
 102 }
 103
 104 static void fuse_put_request(struct fuse_req *req);
 105
 106 static struct fuse_req *fuse_get_req(struct fuse_mount *fm, bool for_background)
 107 {
 108         struct fuse_conn *fc = fm->fc;
 109         struct fuse_req *req;
 110         int err;
 111         atomic_inc(&fc->num_waiting);
 112
 113         if (fuse_block_alloc(fc, for_background)) {
 114                 err = -EINTR;
 115                 if (wait_event_killable_exclusive(fc->blocked_waitq,
 116                                 !fuse_block_alloc(fc, for_background)))
 117                         goto out;
 118         }
 119         /* Matches smp_wmb() in fuse_set_initialized() */
 120         smp_rmb();
 121
 122         err = -ENOTCONN;
 123         if (!fc->connected)
 124                 goto out;
 125
 126         err = -ECONNREFUSED;
 127         if (fc->conn_error)
 128                 goto out;
 129
 130         req = fuse_request_alloc(fm, GFP_KERNEL);
 131         err = -ENOMEM;
 132         if (!req) {
 133                 if (for_background)
 134                         wake_up(&fc->blocked_waitq);
 135                 goto out;
 136         }
 137
 138         req->in.h.uid = from_kuid(fc->user_ns, current_fsuid());
 139         req->in.h.gid = from_kgid(fc->user_ns, current_fsgid());
 140         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 141
 142         __set_bit(FR_WAITING, &req->flags);
 143         if (for_background)
 144                 __set_bit(FR_BACKGROUND, &req->flags);
 145
 146         if (unlikely(req->in.h.uid == ((uid_t)-1) ||
 147                      req->in.h.gid == ((gid_t)-1))) {
 148                 fuse_put_request(req);
 149                 return ERR_PTR(-EOVERFLOW);
 150         }
 151         return req;
 152
 153  out:
 154         fuse_drop_waiting(fc);
 155         return ERR_PTR(err);
 156 }
 157
 158 static void fuse_put_request(struct fuse_req *req)
 159 {
 160         struct fuse_conn *fc = req->fm->fc;
 161
 162         if (refcount_dec_and_test(&req->count)) {
 163                 if (test_bit(FR_BACKGROUND, &req->flags)) {
 164                         /*
 165                          * We get here in the unlikely case that a background
 166                          * request was allocated but not sent
 167                          */
 168                         spin_lock(&fc->bg_lock);
 169                         if (!fc->blocked)
 170                                 wake_up(&fc->blocked_waitq);
 171                         spin_unlock(&fc->bg_lock);
 172                 }
 173
 174                 if (test_bit(FR_WAITING, &req->flags)) {
 175                         __clear_bit(FR_WAITING, &req->flags);
 176                         fuse_drop_waiting(fc);
 177                 }
 178
 179                 fuse_request_free(req);
 180         }
 181 }
 182
 183 unsigned int fuse_len_args(unsigned int numargs, struct fuse_arg *args)
 184 {
 185         unsigned nbytes = 0;
 186         unsigned i;
 187
 188         for (i = 0; i < numargs; i++)
 189                 nbytes += args[i].size;
 190
 191         return nbytes;
 192 }
 193 EXPORT_SYMBOL_GPL(fuse_len_args);
 194
 195 u64 fuse_get_unique(struct fuse_iqueue *fiq)
 196 {
 197         fiq->reqctr += FUSE_REQ_ID_STEP;
 198         return fiq->reqctr;
 199 }
 200 EXPORT_SYMBOL_GPL(fuse_get_unique);
 201
 202 static unsigned int fuse_req_hash(u64 unique)
 203 {
 204         return hash_long(unique & ~FUSE_INT_REQ_BIT, FUSE_PQ_HASH_BITS);
 205 }
 206
 207 /*
 208  * A new request is available, wake fiq->waitq
 209  */
 210 static void fuse_dev_wake_and_unlock(struct fuse_iqueue *fiq)
 211 __releases(fiq->lock)
 212 {
 213         wake_up(&fiq->waitq);
 214         kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
 215         spin_unlock(&fiq->lock);
 216 }
 217
 218 const struct fuse_iqueue_ops fuse_dev_fiq_ops = {
 219         .wake_forget_and_unlock         = fuse_dev_wake_and_unlock,
 220         .wake_interrupt_and_unlock      = fuse_dev_wake_and_unlock,
 221         .wake_pending_and_unlock        = fuse_dev_wake_and_unlock,
 222 };
 223 EXPORT_SYMBOL_GPL(fuse_dev_fiq_ops);
 224
 225 static void queue_request_and_unlock(struct fuse_iqueue *fiq,
 226                                      struct fuse_req *req)
 227 __releases(fiq->lock)
 228 {
 229         req->in.h.len = sizeof(struct fuse_in_header) +
 230                 fuse_len_args(req->args->in_numargs,
 231                               (struct fuse_arg *) req->args->in_args);
 232         list_add_tail(&req->list, &fiq->pending);
 233         fiq->ops->wake_pending_and_unlock(fiq);
 234 }
 235
 236 void fuse_queue_forget(struct fuse_conn *fc, struct fuse_forget_link *forget,
 237                        u64 nodeid, u64 nlookup)
 238 {
 239         struct fuse_iqueue *fiq = &fc->iq;
 240
 241         forget->forget_one.nodeid = nodeid;
 242         forget->forget_one.nlookup = nlookup;
 243
 244         spin_lock(&fiq->lock);
 245         if (fiq->connected) {
 246                 fiq->forget_list_tail->next = forget;
 247                 fiq->forget_list_tail = forget;
 248                 fiq->ops->wake_forget_and_unlock(fiq);
 249         } else {
 250                 kfree(forget);
 251                 spin_unlock(&fiq->lock);
 252         }
 253 }
 254
 255 static void flush_bg_queue(struct fuse_conn *fc)
 256 {
 257         struct fuse_iqueue *fiq = &fc->iq;
 258
 259         while (fc->active_background < fc->max_background &&
 260                !list_empty(&fc->bg_queue)) {
 261                 struct fuse_req *req;
 262
 263                 req = list_first_entry(&fc->bg_queue, struct fuse_req, list);
 264                 list_del(&req->list);
 265                 fc->active_background++;
 266                 spin_lock(&fiq->lock);
 267                 req->in.h.unique = fuse_get_unique(fiq);
 268                 queue_request_and_unlock(fiq, req);
 269         }
 270 }
 271
 272 /*
 273  * This function is called when a request is finished.  Either a reply
 274  * has arrived or it was aborted (and not yet sent) or some error
 275  * occurred during communication with userspace, or the device file
 276  * was closed.  The requester thread is woken up (if still waiting),
 277  * the 'end' callback is called if given, else the reference to the
 278  * request is released
 279  */
 280 void fuse_request_end(struct fuse_req *req)
 281 {
 282         struct fuse_mount *fm = req->fm;
 283         struct fuse_conn *fc = fm->fc;
 284         struct fuse_iqueue *fiq = &fc->iq;
 285
 286         if (test_and_set_bit(FR_FINISHED, &req->flags))
 287                 goto put_request;
 288
 289         /*
 290          * test_and_set_bit() implies smp_mb() between bit
 291          * changing and below FR_INTERRUPTED check. Pairs with
 292          * smp_mb() from queue_interrupt().
 293          */
 294         if (test_bit(FR_INTERRUPTED, &req->flags)) {
 295                 spin_lock(&fiq->lock);
 296                 list_del_init(&req->intr_entry);
 297                 spin_unlock(&fiq->lock);
 298         }
 299         WARN_ON(test_bit(FR_PENDING, &req->flags));
 300         WARN_ON(test_bit(FR_SENT, &req->flags));
 301         if (test_bit(FR_BACKGROUND, &req->flags)) {
 302                 spin_lock(&fc->bg_lock);
 303                 clear_bit(FR_BACKGROUND, &req->flags);
 304                 if (fc->num_background == fc->max_background) {
 305                         fc->blocked = 0;
 306                         wake_up(&fc->blocked_waitq);
 307                 } else if (!fc->blocked) {
 308                         /*
 309                          * Wake up next waiter, if any.  It's okay to use
 310                          * waitqueue_active(), as we've already synced up
 311                          * fc->blocked with waiters with the wake_up() call
 312                          * above.
 313                          */
 314                         if (waitqueue_active(&fc->blocked_waitq))
 315                                 wake_up(&fc->blocked_waitq);
 316                 }
 317
 318                 fc->num_background--;
 319                 fc->active_background--;
 320                 flush_bg_queue(fc);
 321                 spin_unlock(&fc->bg_lock);
 322         } else {
 323                 /* Wake up waiter sleeping in request_wait_answer() */
 324                 wake_up(&req->waitq);
 325         }
 326
 327         if (test_bit(FR_ASYNC, &req->flags))
 328                 req->args->end(fm, req->args, req->out.h.error);
 329 put_request:
 330         fuse_put_request(req);
 331 }
 332 EXPORT_SYMBOL_GPL(fuse_request_end);
 333
 334 static int queue_interrupt(struct fuse_req *req)
 335 {
 336         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 337
 338         spin_lock(&fiq->lock);
 339         /* Check for we've sent request to interrupt this req */
 340         if (unlikely(!test_bit(FR_INTERRUPTED, &req->flags))) {
 341                 spin_unlock(&fiq->lock);
 342                 return -EINVAL;
 343         }
 344
 345         if (list_empty(&req->intr_entry)) {
 346                 list_add_tail(&req->intr_entry, &fiq->interrupts);
 347                 /*
 348                  * Pairs with smp_mb() implied by test_and_set_bit()
 349                  * from fuse_request_end().
 350                  */
 351                 smp_mb();
 352                 if (test_bit(FR_FINISHED, &req->flags)) {
 353                         list_del_init(&req->intr_entry);
 354                         spin_unlock(&fiq->lock);
 355                         return 0;
 356                 }
 357                 fiq->ops->wake_interrupt_and_unlock(fiq);
 358         } else {
 359                 spin_unlock(&fiq->lock);
 360         }
 361         return 0;
 362 }
 363
 364 static void request_wait_answer(struct fuse_req *req)
 365 {
 366         struct fuse_conn *fc = req->fm->fc;
 367         struct fuse_iqueue *fiq = &fc->iq;
 368         int err;
 369
 370         if (!fc->no_interrupt) {
 371                 /* Any signal may interrupt this */
 372                 err = wait_event_interruptible(req->waitq,
 373                                         test_bit(FR_FINISHED, &req->flags));
 374                 if (!err)
 375                         return;
 376
 377                 set_bit(FR_INTERRUPTED, &req->flags);
 378                 /* matches barrier in fuse_dev_do_read() */
 379                 smp_mb__after_atomic();
 380                 if (test_bit(FR_SENT, &req->flags))
 381                         queue_interrupt(req);
 382         }
 383
 384         if (!test_bit(FR_FORCE, &req->flags)) {
 385                 /* Only fatal signals may interrupt this */
 386                 err = wait_event_killable(req->waitq,
 387                                         test_bit(FR_FINISHED, &req->flags));
 388                 if (!err)
 389                         return;
 390
 391                 spin_lock(&fiq->lock);
 392                 /* Request is not yet in userspace, bail out */
 393                 if (test_bit(FR_PENDING, &req->flags)) {
 394                         list_del(&req->list);
 395                         spin_unlock(&fiq->lock);
 396                         __fuse_put_request(req);
 397                         req->out.h.error = -EINTR;
 398                         return;
 399                 }
 400                 spin_unlock(&fiq->lock);
 401         }
 402
 403         /*
 404          * Either request is already in userspace, or it was forced.
 405          * Wait it out.
 406          */
 407         wait_event(req->waitq, test_bit(FR_FINISHED, &req->flags));
 408 }
 409
 410 static void __fuse_request_send(struct fuse_req *req)
 411 {
 412         struct fuse_iqueue *fiq = &req->fm->fc->iq;
 413
 414         BUG_ON(test_bit(FR_BACKGROUND, &req->flags));
 415         spin_lock(&fiq->lock);
 416         if (!fiq->connected) {
 417                 spin_unlock(&fiq->lock);
 418                 req->out.h.error = -ENOTCONN;
 419         } else {
 420                 req->in.h.unique = fuse_get_unique(fiq);
 421                 /* acquire extra reference, since request is still needed
 422                    after fuse_request_end() */
 423                 __fuse_get_request(req);
 424                 queue_request_and_unlock(fiq, req);
 425
 426                 request_wait_answer(req);
 427                 /* Pairs with smp_wmb() in fuse_request_end() */
 428                 smp_rmb();
 429         }
 430 }
 431
 432 static void fuse_adjust_compat(struct fuse_conn *fc, struct fuse_args *args)
 433 {
 434         if (fc->minor < 4 && args->opcode == FUSE_STATFS)
 435                 args->out_args[0].size = FUSE_COMPAT_STATFS_SIZE;
 436
 437         if (fc->minor < 9) {
 438                 switch (args->opcode) {
 439                 case FUSE_LOOKUP:
 440                 case FUSE_CREATE:
 441                 case FUSE_MKNOD:
 442                 case FUSE_MKDIR:
 443                 case FUSE_SYMLINK:
 444                 case FUSE_LINK:
 445                         args->out_args[0].size = FUSE_COMPAT_ENTRY_OUT_SIZE;
 446                         break;
 447                 case FUSE_GETATTR:
 448                 case FUSE_SETATTR:
 449                         args->out_args[0].size = FUSE_COMPAT_ATTR_OUT_SIZE;
 450                         break;
 451                 }
 452         }
 453         if (fc->minor < 12) {
 454                 switch (args->opcode) {
 455                 case FUSE_CREATE:
 456                         args->in_args[0].size = sizeof(struct fuse_open_in);
 457                         break;
 458                 case FUSE_MKNOD:
 459                         args->in_args[0].size = FUSE_COMPAT_MKNOD_IN_SIZE;
 460                         break;
 461                 }
 462         }
 463 }
 464
 465 static void fuse_force_creds(struct fuse_req *req)
 466 {
 467         struct fuse_conn *fc = req->fm->fc;
 468
 469         req->in.h.uid = from_kuid_munged(fc->user_ns, current_fsuid());
 470         req->in.h.gid = from_kgid_munged(fc->user_ns, current_fsgid());
 471         req->in.h.pid = pid_nr_ns(task_pid(current), fc->pid_ns);
 472 }
 473
 474 static void fuse_args_to_req(struct fuse_req *req, struct fuse_args *args)
 475 {
 476         req->in.h.opcode = args->opcode;
 477         req->in.h.nodeid = args->nodeid;
 478         req->args = args;
 479         if (args->is_ext)
 480                 req->in.h.total_extlen = args->in_args[args->ext_idx].size / 8;
 481         if (args->end)
 482                 __set_bit(FR_ASYNC, &req->flags);
 483 }
 484
 485 ssize_t fuse_simple_request(struct fuse_mount *fm, struct fuse_args *args)
 486 {
 487         struct fuse_conn *fc = fm->fc;
 488         struct fuse_req *req;
 489         ssize_t ret;
 490
 491         if (args->force) {
 492                 atomic_inc(&fc->num_waiting);
 493                 req = fuse_request_alloc(fm, GFP_KERNEL | __GFP_NOFAIL);
 494
 495                 if (!args->nocreds)
 496                         fuse_force_creds(req);
 497
 498                 __set_bit(FR_WAITING, &req->flags);
 499                 __set_bit(FR_FORCE, &req->flags);
 500         } else {
 501                 WARN_ON(args->nocreds);
 502                 req = fuse_get_req(fm, false);
 503                 if (IS_ERR(req))
 504                         return PTR_ERR(req);
 505         }
 506
 507         /* Needs to be done after fuse_get_req() so that fc->minor is valid */
 508         fuse_adjust_compat(fc, args);
 509         fuse_args_to_req(req, args);
 510
 511         if (!args->noreply)
 512                 __set_bit(FR_ISREPLY, &req->flags);
 513         __fuse_request_send(req);
 514         ret = req->out.h.error;
 515         if (!ret && args->out_argvar) {
 516                 BUG_ON(args->out_numargs == 0);
 517                 ret = args->out_args[args->out_numargs - 1].size;
 518         }
 519         fuse_put_request(req);
 520
 521         return ret;
 522 }
 523
 524 static bool fuse_request_queue_background(struct fuse_req *req)
 525 {
 526         struct fuse_mount *fm = req->fm;
 527         struct fuse_conn *fc = fm->fc;
 528         bool queued = false;
 529
 530         WARN_ON(!test_bit(FR_BACKGROUND, &req->flags));
 531         if (!test_bit(FR_WAITING, &req->flags)) {
 532                 __set_bit(FR_WAITING, &req->flags);
 533                 atomic_inc(&fc->num_waiting);
 534         }
 535         __set_bit(FR_ISREPLY, &req->flags);
 536         spin_lock(&fc->bg_lock);
 537         if (likely(fc->connected)) {
 538                 fc->num_background++;
 539                 if (fc->num_background == fc->max_background)
 540                         fc->blocked = 1;
 541                 list_add_tail(&req->list, &fc->bg_queue);
 542                 flush_bg_queue(fc);
 543                 queued = true;
 544         }
 545         spin_unlock(&fc->bg_lock);
 546
 547         return queued;
 548 }
 549
 550 int fuse_simple_background(struct fuse_mount *fm, struct fuse_args *args,
 551                             gfp_t gfp_flags)
 552 {
 553         struct fuse_req *req;
 554
 555         if (args->force) {
 556                 WARN_ON(!args->nocreds);
 557                 req = fuse_request_alloc(fm, gfp_flags);
 558                 if (!req)
 559                         return -ENOMEM;
 560                 __set_bit(FR_BACKGROUND, &req->flags);
 561         } else {
 562                 WARN_ON(args->nocreds);
 563                 req = fuse_get_req(fm, true);
 564                 if (IS_ERR(req))
 565                         return PTR_ERR(req);
 566         }
 567
 568         fuse_args_to_req(req, args);
 569
 570         if (!fuse_request_queue_background(req)) {
 571                 fuse_put_request(req);
 572                 return -ENOTCONN;
 573         }
 574
 575         return 0;
 576 }
 577 EXPORT_SYMBOL_GPL(fuse_simple_background);
 578
 579 static int fuse_simple_notify_reply(struct fuse_mount *fm,
 580                                     struct fuse_args *args, u64 unique)
 581 {
 582         struct fuse_req *req;
 583         struct fuse_iqueue *fiq = &fm->fc->iq;
 584         int err = 0;
 585
 586         req = fuse_get_req(fm, false);
 587         if (IS_ERR(req))
 588                 return PTR_ERR(req);
 589
 590         __clear_bit(FR_ISREPLY, &req->flags);
 591         req->in.h.unique = unique;
 592
 593         fuse_args_to_req(req, args);
 594
 595         spin_lock(&fiq->lock);
 596         if (fiq->connected) {
 597                 queue_request_and_unlock(fiq, req);
 598         } else {
 599                 err = -ENODEV;
 600                 spin_unlock(&fiq->lock);
 601                 fuse_put_request(req);
 602         }
 603
 604         return err;
 605 }
 606
 607 /*
 608  * Lock the request.  Up to the next unlock_request() there mustn't be
 609  * anything that could cause a page-fault.  If the request was already
 610  * aborted bail out.
 611  */
 612 static int lock_request(struct fuse_req *req)
 613 {
 614         int err = 0;
 615         if (req) {
 616                 spin_lock(&req->waitq.lock);
 617                 if (test_bit(FR_ABORTED, &req->flags))
 618                         err = -ENOENT;
 619                 else
 620                         set_bit(FR_LOCKED, &req->flags);
 621                 spin_unlock(&req->waitq.lock);
 622         }
 623         return err;
 624 }
 625
 626 /*
 627  * Unlock request.  If it was aborted while locked, caller is responsible
 628  * for unlocking and ending the request.
 629  */
 630 static int unlock_request(struct fuse_req *req)
 631 {
 632         int err = 0;
 633         if (req) {
 634                 spin_lock(&req->waitq.lock);
 635                 if (test_bit(FR_ABORTED, &req->flags))
 636                         err = -ENOENT;
 637                 else
 638                         clear_bit(FR_LOCKED, &req->flags);
 639                 spin_unlock(&req->waitq.lock);
 640         }
 641         return err;
 642 }
 643
 644 struct fuse_copy_state {
 645         int write;
 646         struct fuse_req *req;
 647         struct iov_iter *iter;
 648         struct pipe_buffer *pipebufs;
 649         struct pipe_buffer *currbuf;
 650         struct pipe_inode_info *pipe;
 651         unsigned long nr_segs;
 652         struct page *pg;
 653         unsigned len;
 654         unsigned offset;
 655         unsigned move_pages:1;
 656 };
 657
 658 static void fuse_copy_init(struct fuse_copy_state *cs, int write,
 659                            struct iov_iter *iter)
 660 {
 661         memset(cs, 0, sizeof(*cs));
 662         cs->write = write;
 663         cs->iter = iter;
 664 }
 665
 666 /* Unmap and put previous page of userspace buffer */
 667 static void fuse_copy_finish(struct fuse_copy_state *cs)
 668 {
 669         if (cs->currbuf) {
 670                 struct pipe_buffer *buf = cs->currbuf;
 671
 672                 if (cs->write)
 673                         buf->len = PAGE_SIZE - cs->len;
 674                 cs->currbuf = NULL;
 675         } else if (cs->pg) {
 676                 if (cs->write) {
 677                         flush_dcache_page(cs->pg);
 678                         set_page_dirty_lock(cs->pg);
 679                 }
 680                 put_page(cs->pg);
 681         }
 682         cs->pg = NULL;
 683 }
 684
 685 /*
 686  * Get another pagefull of userspace buffer, and map it to kernel
 687  * address space, and lock request
 688  */
 689 static int fuse_copy_fill(struct fuse_copy_state *cs)
 690 {
 691         struct page *page;
 692         int err;
 693
 694         err = unlock_request(cs->req);
 695         if (err)
 696                 return err;
 697
 698         fuse_copy_finish(cs);
 699         if (cs->pipebufs) {
 700                 struct pipe_buffer *buf = cs->pipebufs;
 701
 702                 if (!cs->write) {
 703                         err = pipe_buf_confirm(cs->pipe, buf);
 704                         if (err)
 705                                 return err;
 706
 707                         BUG_ON(!cs->nr_segs);
 708                         cs->currbuf = buf;
 709                         cs->pg = buf->page;
 710                         cs->offset = buf->offset;
 711                         cs->len = buf->len;
 712                         cs->pipebufs++;
 713                         cs->nr_segs--;
 714                 } else {
 715                         if (cs->nr_segs >= cs->pipe->max_usage)
 716                                 return -EIO;
 717
 718                         page = alloc_page(GFP_HIGHUSER);
 719                         if (!page)
 720                                 return -ENOMEM;
 721
 722                         buf->page = page;
 723                         buf->offset = 0;
 724                         buf->len = 0;
 725
 726                         cs->currbuf = buf;
 727                         cs->pg = page;
 728                         cs->offset = 0;
 729                         cs->len = PAGE_SIZE;
 730                         cs->pipebufs++;
 731                         cs->nr_segs++;
 732                 }
 733         } else {
 734                 size_t off;
 735                 err = iov_iter_get_pages2(cs->iter, &page, PAGE_SIZE, 1, &off);
 736                 if (err < 0)
 737                         return err;
 738                 BUG_ON(!err);
 739                 cs->len = err;
 740                 cs->offset = off;
 741                 cs->pg = page;
 742         }
 743
 744         return lock_request(cs->req);
 745 }
 746
 747 /* Do as much copy to/from userspace buffer as we can */
 748 static int fuse_copy_do(struct fuse_copy_state *cs, void **val, unsigned *size)
 749 {
 750         unsigned ncpy = min(*size, cs->len);
 751         if (val) {
 752                 void *pgaddr = kmap_local_page(cs->pg);
 753                 void *buf = pgaddr + cs->offset;
 754
 755                 if (cs->write)
 756                         memcpy(buf, *val, ncpy);
 757                 else
 758                         memcpy(*val, buf, ncpy);
 759
 760                 kunmap_local(pgaddr);
 761                 *val += ncpy;
 762         }
 763         *size -= ncpy;
 764         cs->len -= ncpy;
 765         cs->offset += ncpy;
 766         return ncpy;
 767 }
 768
 769 static int fuse_check_folio(struct folio *folio)
 770 {
 771         if (folio_mapped(folio) ||
 772             folio->mapping != NULL ||
 773             (folio->flags & PAGE_FLAGS_CHECK_AT_PREP &
 774              ~(1 << PG_locked |
 775                1 << PG_referenced |
 776                1 << PG_uptodate |
 777                1 << PG_lru |
 778                1 << PG_active |
 779                1 << PG_workingset |
 780                1 << PG_reclaim |
 781                1 << PG_waiters |
 782                LRU_GEN_MASK | LRU_REFS_MASK))) {
 783                 dump_page(&folio->page, "fuse: trying to steal weird page");
 784                 return 1;
 785         }
 786         return 0;
 787 }
 788
 789 static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
 790 {
 791         int err;
 792         struct folio *oldfolio = page_folio(*pagep);
 793         struct folio *newfolio;
 794         struct pipe_buffer *buf = cs->pipebufs;
 795
 796         folio_get(oldfolio);
 797         err = unlock_request(cs->req);
 798         if (err)
 799                 goto out_put_old;
 800
 801         fuse_copy_finish(cs);
 802
 803         err = pipe_buf_confirm(cs->pipe, buf);
 804         if (err)
 805                 goto out_put_old;
 806
 807         BUG_ON(!cs->nr_segs);
 808         cs->currbuf = buf;
 809         cs->len = buf->len;
 810         cs->pipebufs++;
 811         cs->nr_segs--;
 812
 813         if (cs->len != PAGE_SIZE)
 814                 goto out_fallback;
 815
 816         if (!pipe_buf_try_steal(cs->pipe, buf))
 817                 goto out_fallback;
 818
 819         newfolio = page_folio(buf->page);
 820
 821         if (!folio_test_uptodate(newfolio))
 822                 folio_mark_uptodate(newfolio);
 823
 824         folio_clear_mappedtodisk(newfolio);
 825
 826         if (fuse_check_folio(newfolio) != 0)
 827                 goto out_fallback_unlock;
 828
 829         /*
 830          * This is a new and locked page, it shouldn't be mapped or
 831          * have any special flags on it
 832          */
 833         if (WARN_ON(folio_mapped(oldfolio)))
 834                 goto out_fallback_unlock;
 835         if (WARN_ON(folio_has_private(oldfolio)))
 836                 goto out_fallback_unlock;
 837         if (WARN_ON(folio_test_dirty(oldfolio) ||
 838                                 folio_test_writeback(oldfolio)))
 839                 goto out_fallback_unlock;
 840         if (WARN_ON(folio_test_mlocked(oldfolio)))
 841                 goto out_fallback_unlock;
 842
 843         replace_page_cache_folio(oldfolio, newfolio);
 844
 845         folio_get(newfolio);
 846
 847         if (!(buf->flags & PIPE_BUF_FLAG_LRU))
 848                 folio_add_lru(newfolio);
 849
 850         /*
 851          * Release while we have extra ref on stolen page.  Otherwise
 852          * anon_pipe_buf_release() might think the page can be reused.
 853          */
 854         pipe_buf_release(cs->pipe, buf);
 855
 856         err = 0;
 857         spin_lock(&cs->req->waitq.lock);
 858         if (test_bit(FR_ABORTED, &cs->req->flags))
 859                 err = -ENOENT;
 860         else
 861                 *pagep = &newfolio->page;
 862         spin_unlock(&cs->req->waitq.lock);
 863
 864         if (err) {
 865                 folio_unlock(newfolio);
 866                 folio_put(newfolio);
 867                 goto out_put_old;
 868         }
 869
 870         folio_unlock(oldfolio);
 871         /* Drop ref for ap->pages[] array */
 872         folio_put(oldfolio);
 873         cs->len = 0;
 874
 875         err = 0;
 876 out_put_old:
 877         /* Drop ref obtained in this function */
 878         folio_put(oldfolio);
 879         return err;
 880
 881 out_fallback_unlock:
 882         folio_unlock(newfolio);
 883 out_fallback:
 884         cs->pg = buf->page;
 885         cs->offset = buf->offset;
 886
 887         err = lock_request(cs->req);
 888         if (!err)
 889                 err = 1;
 890
 891         goto out_put_old;
 892 }
 893
 894 static int fuse_ref_page(struct fuse_copy_state *cs, struct page *page,
 895                          unsigned offset, unsigned count)
 896 {
 897         struct pipe_buffer *buf;
 898         int err;
 899
 900         if (cs->nr_segs >= cs->pipe->max_usage)
 901                 return -EIO;
 902
 903         get_page(page);
 904         err = unlock_request(cs->req);
 905         if (err) {
 906                 put_page(page);
 907                 return err;
 908         }
 909
 910         fuse_copy_finish(cs);
 911
 912         buf = cs->pipebufs;
 913         buf->page = page;
 914         buf->offset = offset;
 915         buf->len = count;
 916
 917         cs->pipebufs++;
 918         cs->nr_segs++;
 919         cs->len = 0;
 920
 921         return 0;
 922 }
 923
 924 /*
 925  * Copy a page in the request to/from the userspace buffer.  Must be
 926  * done atomically
 927  */
 928 static int fuse_copy_page(struct fuse_copy_state *cs, struct page **pagep,
 929                           unsigned offset, unsigned count, int zeroing)
 930 {
 931         int err;
 932         struct page *page = *pagep;
 933
 934         if (page && zeroing && count < PAGE_SIZE)
 935                 clear_highpage(page);
 936
 937         while (count) {
 938                 if (cs->write && cs->pipebufs && page) {
 939                         /*
 940                          * Can't control lifetime of pipe buffers, so always
 941                          * copy user pages.
 942                          */
 943                         if (cs->req->args->user_pages) {
 944                                 err = fuse_copy_fill(cs);
 945                                 if (err)
 946                                         return err;
 947                         } else {
 948                                 return fuse_ref_page(cs, page, offset, count);
 949                         }
 950                 } else if (!cs->len) {
 951                         if (cs->move_pages && page &&
 952                             offset == 0 && count == PAGE_SIZE) {
 953                                 err = fuse_try_move_page(cs, pagep);
 954                                 if (err <= 0)
 955                                         return err;
 956                         } else {
 957                                 err = fuse_copy_fill(cs);
 958                                 if (err)
 959                                         return err;
 960                         }
 961                 }
 962                 if (page) {
 963                         void *mapaddr = kmap_local_page(page);
 964                         void *buf = mapaddr + offset;
 965                         offset += fuse_copy_do(cs, &buf, &count);
 966                         kunmap_local(mapaddr);
 967                 } else
 968                         offset += fuse_copy_do(cs, NULL, &count);
 969         }
 970         if (page && !cs->write)
 971                 flush_dcache_page(page);
 972         return 0;
 973 }
 974
 975 /* Copy pages in the request to/from userspace buffer */
 976 static int fuse_copy_pages(struct fuse_copy_state *cs, unsigned nbytes,
 977                            int zeroing)
 978 {
 979         unsigned i;
 980         struct fuse_req *req = cs->req;
 981         struct fuse_args_pages *ap = container_of(req->args, typeof(*ap), args);
 982
 983
 984         for (i = 0; i < ap->num_pages && (nbytes || zeroing); i++) {
 985                 int err;
 986                 unsigned int offset = ap->descs[i].offset;
 987                 unsigned int count = min(nbytes, ap->descs[i].length);
 988
 989                 err = fuse_copy_page(cs, &ap->pages[i], offset, count, zeroing);
 990                 if (err)
 991                         return err;
 992
 993                 nbytes -= count;
 994         }
 995         return 0;
 996 }
 997
 998 /* Copy a single argument in the request to/from userspace buffer */
 999 static int fuse_copy_one(struct fuse_copy_state *cs, void *val, unsigned size)
1000 {
1001         while (size) {
1002                 if (!cs->len) {
1003                         int err = fuse_copy_fill(cs);
1004                         if (err)
1005                                 return err;
1006                 }
1007                 fuse_copy_do(cs, &val, &size);
1008         }
1009         return 0;
1010 }
1011
1012 /* Copy request arguments to/from userspace buffer */
1013 static int fuse_copy_args(struct fuse_copy_state *cs, unsigned numargs,
1014                           unsigned argpages, struct fuse_arg *args,
1015                           int zeroing)
1016 {
1017         int err = 0;
1018         unsigned i;
1019
1020         for (i = 0; !err && i < numargs; i++)  {
1021                 struct fuse_arg *arg = &args[i];
1022                 if (i == numargs - 1 && argpages)
1023                         err = fuse_copy_pages(cs, arg->size, zeroing);
1024                 else
1025                         err = fuse_copy_one(cs, arg->value, arg->size);
1026         }
1027         return err;
1028 }
1029
1030 static int forget_pending(struct fuse_iqueue *fiq)
1031 {
1032         return fiq->forget_list_head.next != NULL;
1033 }
1034
1035 static int request_pending(struct fuse_iqueue *fiq)
1036 {
1037         return !list_empty(&fiq->pending) || !list_empty(&fiq->interrupts) ||
1038                 forget_pending(fiq);
1039 }
1040
1041 /*
1042  * Transfer an interrupt request to userspace
1043  *
1044  * Unlike other requests this is assembled on demand, without a need
1045  * to allocate a separate fuse_req structure.
1046  *
1047  * Called with fiq->lock held, releases it
1048  */
1049 static int fuse_read_interrupt(struct fuse_iqueue *fiq,
1050                                struct fuse_copy_state *cs,
1051                                size_t nbytes, struct fuse_req *req)
1052 __releases(fiq->lock)
1053 {
1054         struct fuse_in_header ih;
1055         struct fuse_interrupt_in arg;
1056         unsigned reqsize = sizeof(ih) + sizeof(arg);
1057         int err;
1058
1059         list_del_init(&req->intr_entry);
1060         memset(&ih, 0, sizeof(ih));
1061         memset(&arg, 0, sizeof(arg));
1062         ih.len = reqsize;
1063         ih.opcode = FUSE_INTERRUPT;
1064         ih.unique = (req->in.h.unique | FUSE_INT_REQ_BIT);
1065         arg.unique = req->in.h.unique;
1066
1067         spin_unlock(&fiq->lock);
1068         if (nbytes < reqsize)
1069                 return -EINVAL;
1070
1071         err = fuse_copy_one(cs, &ih, sizeof(ih));
1072         if (!err)
1073                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1074         fuse_copy_finish(cs);
1075
1076         return err ? err : reqsize;
1077 }
1078
1079 struct fuse_forget_link *fuse_dequeue_forget(struct fuse_iqueue *fiq,
1080                                              unsigned int max,
1081                                              unsigned int *countp)
1082 {
1083         struct fuse_forget_link *head = fiq->forget_list_head.next;
1084         struct fuse_forget_link **newhead = &head;
1085         unsigned count;
1086
1087         for (count = 0; *newhead != NULL && count < max; count++)
1088                 newhead = &(*newhead)->next;
1089
1090         fiq->forget_list_head.next = *newhead;
1091         *newhead = NULL;
1092         if (fiq->forget_list_head.next == NULL)
1093                 fiq->forget_list_tail = &fiq->forget_list_head;
1094
1095         if (countp != NULL)
1096                 *countp = count;
1097
1098         return head;
1099 }
1100 EXPORT_SYMBOL(fuse_dequeue_forget);
1101
1102 static int fuse_read_single_forget(struct fuse_iqueue *fiq,
1103                                    struct fuse_copy_state *cs,
1104                                    size_t nbytes)
1105 __releases(fiq->lock)
1106 {
1107         int err;
1108         struct fuse_forget_link *forget = fuse_dequeue_forget(fiq, 1, NULL);
1109         struct fuse_forget_in arg = {
1110                 .nlookup = forget->forget_one.nlookup,
1111         };
1112         struct fuse_in_header ih = {
1113                 .opcode = FUSE_FORGET,
1114                 .nodeid = forget->forget_one.nodeid,
1115                 .unique = fuse_get_unique(fiq),
1116                 .len = sizeof(ih) + sizeof(arg),
1117         };
1118
1119         spin_unlock(&fiq->lock);
1120         kfree(forget);
1121         if (nbytes < ih.len)
1122                 return -EINVAL;
1123
1124         err = fuse_copy_one(cs, &ih, sizeof(ih));
1125         if (!err)
1126                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1127         fuse_copy_finish(cs);
1128
1129         if (err)
1130                 return err;
1131
1132         return ih.len;
1133 }
1134
1135 static int fuse_read_batch_forget(struct fuse_iqueue *fiq,
1136                                    struct fuse_copy_state *cs, size_t nbytes)
1137 __releases(fiq->lock)
1138 {
1139         int err;
1140         unsigned max_forgets;
1141         unsigned count;
1142         struct fuse_forget_link *head;
1143         struct fuse_batch_forget_in arg = { .count = 0 };
1144         struct fuse_in_header ih = {
1145                 .opcode = FUSE_BATCH_FORGET,
1146                 .unique = fuse_get_unique(fiq),
1147                 .len = sizeof(ih) + sizeof(arg),
1148         };
1149
1150         if (nbytes < ih.len) {
1151                 spin_unlock(&fiq->lock);
1152                 return -EINVAL;
1153         }
1154
1155         max_forgets = (nbytes - ih.len) / sizeof(struct fuse_forget_one);
1156         head = fuse_dequeue_forget(fiq, max_forgets, &count);
1157         spin_unlock(&fiq->lock);
1158
1159         arg.count = count;
1160         ih.len += count * sizeof(struct fuse_forget_one);
1161         err = fuse_copy_one(cs, &ih, sizeof(ih));
1162         if (!err)
1163                 err = fuse_copy_one(cs, &arg, sizeof(arg));
1164
1165         while (head) {
1166                 struct fuse_forget_link *forget = head;
1167
1168                 if (!err) {
1169                         err = fuse_copy_one(cs, &forget->forget_one,
1170                                             sizeof(forget->forget_one));
1171                 }
1172                 head = forget->next;
1173                 kfree(forget);
1174         }
1175
1176         fuse_copy_finish(cs);
1177
1178         if (err)
1179                 return err;
1180
1181         return ih.len;
1182 }
1183
1184 static int fuse_read_forget(struct fuse_conn *fc, struct fuse_iqueue *fiq,
1185                             struct fuse_copy_state *cs,
1186                             size_t nbytes)
1187 __releases(fiq->lock)
1188 {
1189         if (fc->minor < 16 || fiq->forget_list_head.next->next == NULL)
1190                 return fuse_read_single_forget(fiq, cs, nbytes);
1191         else
1192                 return fuse_read_batch_forget(fiq, cs, nbytes);
1193 }
1194
1195 /*
1196  * Read a single request into the userspace filesystem's buffer.  This
1197  * function waits until a request is available, then removes it from
1198  * the pending list and copies request data to userspace buffer.  If
1199  * no reply is needed (FORGET) or request has been aborted or there
1200  * was an error during the copying then it's finished by calling
1201  * fuse_request_end().  Otherwise add it to the processing list, and set
1202  * the 'sent' flag.
1203  */
1204 static ssize_t fuse_dev_do_read(struct fuse_dev *fud, struct file *file,
1205                                 struct fuse_copy_state *cs, size_t nbytes)
1206 {
1207         ssize_t err;
1208         struct fuse_conn *fc = fud->fc;
1209         struct fuse_iqueue *fiq = &fc->iq;
1210         struct fuse_pqueue *fpq = &fud->pq;
1211         struct fuse_req *req;
1212         struct fuse_args *args;
1213         unsigned reqsize;
1214         unsigned int hash;
1215
1216         /*
1217          * Require sane minimum read buffer - that has capacity for fixed part
1218          * of any request header + negotiated max_write room for data.
1219          *
1220          * Historically libfuse reserves 4K for fixed header room, but e.g.
1221          * GlusterFS reserves only 80 bytes
1222          *
1223          *      = `sizeof(fuse_in_header) + sizeof(fuse_write_in)`
1224          *
1225          * which is the absolute minimum any sane filesystem should be using
1226          * for header room.
1227          */
1228         if (nbytes < max_t(size_t, FUSE_MIN_READ_BUFFER,
1229                            sizeof(struct fuse_in_header) +
1230                            sizeof(struct fuse_write_in) +
1231                            fc->max_write))
1232                 return -EINVAL;
1233
1234  restart:
1235         for (;;) {
1236                 spin_lock(&fiq->lock);
1237                 if (!fiq->connected || request_pending(fiq))
1238                         break;
1239                 spin_unlock(&fiq->lock);
1240
1241                 if (file->f_flags & O_NONBLOCK)
1242                         return -EAGAIN;
1243                 err = wait_event_interruptible_exclusive(fiq->waitq,
1244                                 !fiq->connected || request_pending(fiq));
1245                 if (err)
1246                         return err;
1247         }
1248
1249         if (!fiq->connected) {
1250                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1251                 goto err_unlock;
1252         }
1253
1254         if (!list_empty(&fiq->interrupts)) {
1255                 req = list_entry(fiq->interrupts.next, struct fuse_req,
1256                                  intr_entry);
1257                 return fuse_read_interrupt(fiq, cs, nbytes, req);
1258         }
1259
1260         if (forget_pending(fiq)) {
1261                 if (list_empty(&fiq->pending) || fiq->forget_batch-- > 0)
1262                         return fuse_read_forget(fc, fiq, cs, nbytes);
1263
1264                 if (fiq->forget_batch <= -8)
1265                         fiq->forget_batch = 16;
1266         }
1267
1268         req = list_entry(fiq->pending.next, struct fuse_req, list);
1269         clear_bit(FR_PENDING, &req->flags);
1270         list_del_init(&req->list);
1271         spin_unlock(&fiq->lock);
1272
1273         args = req->args;
1274         reqsize = req->in.h.len;
1275
1276         /* If request is too large, reply with an error and restart the read */
1277         if (nbytes < reqsize) {
1278                 req->out.h.error = -EIO;
1279                 /* SETXATTR is special, since it may contain too large data */
1280                 if (args->opcode == FUSE_SETXATTR)
1281                         req->out.h.error = -E2BIG;
1282                 fuse_request_end(req);
1283                 goto restart;
1284         }
1285         spin_lock(&fpq->lock);
1286         /*
1287          *  Must not put request on fpq->io queue after having been shut down by
1288          *  fuse_abort_conn()
1289          */
1290         if (!fpq->connected) {
1291                 req->out.h.error = err = -ECONNABORTED;
1292                 goto out_end;
1293
1294         }
1295         list_add(&req->list, &fpq->io);
1296         spin_unlock(&fpq->lock);
1297         cs->req = req;
1298         err = fuse_copy_one(cs, &req->in.h, sizeof(req->in.h));
1299         if (!err)
1300                 err = fuse_copy_args(cs, args->in_numargs, args->in_pages,
1301                                      (struct fuse_arg *) args->in_args, 0);
1302         fuse_copy_finish(cs);
1303         spin_lock(&fpq->lock);
1304         clear_bit(FR_LOCKED, &req->flags);
1305         if (!fpq->connected) {
1306                 err = fc->aborted ? -ECONNABORTED : -ENODEV;
1307                 goto out_end;
1308         }
1309         if (err) {
1310                 req->out.h.error = -EIO;
1311                 goto out_end;
1312         }
1313         if (!test_bit(FR_ISREPLY, &req->flags)) {
1314                 err = reqsize;
1315                 goto out_end;
1316         }
1317         hash = fuse_req_hash(req->in.h.unique);
1318         list_move_tail(&req->list, &fpq->processing[hash]);
1319         __fuse_get_request(req);
1320         set_bit(FR_SENT, &req->flags);
1321         spin_unlock(&fpq->lock);
1322         /* matches barrier in request_wait_answer() */
1323         smp_mb__after_atomic();
1324         if (test_bit(FR_INTERRUPTED, &req->flags))
1325                 queue_interrupt(req);
1326         fuse_put_request(req);
1327
1328         return reqsize;
1329
1330 out_end:
1331         if (!test_bit(FR_PRIVATE, &req->flags))
1332                 list_del_init(&req->list);
1333         spin_unlock(&fpq->lock);
1334         fuse_request_end(req);
1335         return err;
1336
1337  err_unlock:
1338         spin_unlock(&fiq->lock);
1339         return err;
1340 }
1341
1342 static int fuse_dev_open(struct inode *inode, struct file *file)
1343 {
1344         /*
1345          * The fuse device's file's private_data is used to hold
1346          * the fuse_conn(ection) when it is mounted, and is used to
1347          * keep track of whether the file has been mounted already.
1348          */
1349         file->private_data = NULL;
1350         return 0;
1351 }
1352
1353 static ssize_t fuse_dev_read(struct kiocb *iocb, struct iov_iter *to)
1354 {
1355         struct fuse_copy_state cs;
1356         struct file *file = iocb->ki_filp;
1357         struct fuse_dev *fud = fuse_get_dev(file);
1358
1359         if (!fud)
1360                 return -EPERM;
1361
1362         if (!user_backed_iter(to))
1363                 return -EINVAL;
1364
1365         fuse_copy_init(&cs, 1, to);
1366
1367         return fuse_dev_do_read(fud, file, &cs, iov_iter_count(to));
1368 }
1369
1370 static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
1371                                     struct pipe_inode_info *pipe,
1372                                     size_t len, unsigned int flags)
1373 {
1374         int total, ret;
1375         int page_nr = 0;
1376         struct pipe_buffer *bufs;
1377         struct fuse_copy_state cs;
1378         struct fuse_dev *fud = fuse_get_dev(in);
1379
1380         if (!fud)
1381                 return -EPERM;
1382
1383         bufs = kvmalloc_array(pipe->max_usage, sizeof(struct pipe_buffer),
1384                               GFP_KERNEL);
1385         if (!bufs)
1386                 return -ENOMEM;
1387
1388         fuse_copy_init(&cs, 1, NULL);
1389         cs.pipebufs = bufs;
1390         cs.pipe = pipe;
1391         ret = fuse_dev_do_read(fud, in, &cs, len);
1392         if (ret < 0)
1393                 goto out;
1394
1395         if (pipe_occupancy(pipe->head, pipe->tail) + cs.nr_segs > pipe->max_usage) {
1396                 ret = -EIO;
1397                 goto out;
1398         }
1399
1400         for (ret = total = 0; page_nr < cs.nr_segs; total += ret) {
1401                 /*
1402                  * Need to be careful about this.  Having buf->ops in module
1403                  * code can Oops if the buffer persists after module unload.
1404                  */
1405                 bufs[page_nr].ops = &nosteal_pipe_buf_ops;
1406                 bufs[page_nr].flags = 0;
1407                 ret = add_to_pipe(pipe, &bufs[page_nr++]);
1408                 if (unlikely(ret < 0))
1409                         break;
1410         }
1411         if (total)
1412                 ret = total;
1413 out:
1414         for (; page_nr < cs.nr_segs; page_nr++)
1415                 put_page(bufs[page_nr].page);
1416
1417         kvfree(bufs);
1418         return ret;
1419 }
1420
1421 static int fuse_notify_poll(struct fuse_conn *fc, unsigned int size,
1422                             struct fuse_copy_state *cs)
1423 {
1424         struct fuse_notify_poll_wakeup_out outarg;
1425         int err = -EINVAL;
1426
1427         if (size != sizeof(outarg))
1428                 goto err;
1429
1430         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1431         if (err)
1432                 goto err;
1433
1434         fuse_copy_finish(cs);
1435         return fuse_notify_poll_wakeup(fc, &outarg);
1436
1437 err:
1438         fuse_copy_finish(cs);
1439         return err;
1440 }
1441
1442 static int fuse_notify_inval_inode(struct fuse_conn *fc, unsigned int size,
1443                                    struct fuse_copy_state *cs)
1444 {
1445         struct fuse_notify_inval_inode_out outarg;
1446         int err = -EINVAL;
1447
1448         if (size != sizeof(outarg))
1449                 goto err;
1450
1451         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1452         if (err)
1453                 goto err;
1454         fuse_copy_finish(cs);
1455
1456         down_read(&fc->killsb);
1457         err = fuse_reverse_inval_inode(fc, outarg.ino,
1458                                        outarg.off, outarg.len);
1459         up_read(&fc->killsb);
1460         return err;
1461
1462 err:
1463         fuse_copy_finish(cs);
1464         return err;
1465 }
1466
1467 static int fuse_notify_inval_entry(struct fuse_conn *fc, unsigned int size,
1468                                    struct fuse_copy_state *cs)
1469 {
1470         struct fuse_notify_inval_entry_out outarg;
1471         int err = -ENOMEM;
1472         char *buf;
1473         struct qstr name;
1474
1475         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1476         if (!buf)
1477                 goto err;
1478
1479         err = -EINVAL;
1480         if (size < sizeof(outarg))
1481                 goto err;
1482
1483         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1484         if (err)
1485                 goto err;
1486
1487         err = -ENAMETOOLONG;
1488         if (outarg.namelen > FUSE_NAME_MAX)
1489                 goto err;
1490
1491         err = -EINVAL;
1492         if (size != sizeof(outarg) + outarg.namelen + 1)
1493                 goto err;
1494
1495         name.name = buf;
1496         name.len = outarg.namelen;
1497         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1498         if (err)
1499                 goto err;
1500         fuse_copy_finish(cs);
1501         buf[outarg.namelen] = 0;
1502
1503         down_read(&fc->killsb);
1504         err = fuse_reverse_inval_entry(fc, outarg.parent, 0, &name, outarg.flags);
1505         up_read(&fc->killsb);
1506         kfree(buf);
1507         return err;
1508
1509 err:
1510         kfree(buf);
1511         fuse_copy_finish(cs);
1512         return err;
1513 }
1514
1515 static int fuse_notify_delete(struct fuse_conn *fc, unsigned int size,
1516                               struct fuse_copy_state *cs)
1517 {
1518         struct fuse_notify_delete_out outarg;
1519         int err = -ENOMEM;
1520         char *buf;
1521         struct qstr name;
1522
1523         buf = kzalloc(FUSE_NAME_MAX + 1, GFP_KERNEL);
1524         if (!buf)
1525                 goto err;
1526
1527         err = -EINVAL;
1528         if (size < sizeof(outarg))
1529                 goto err;
1530
1531         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1532         if (err)
1533                 goto err;
1534
1535         err = -ENAMETOOLONG;
1536         if (outarg.namelen > FUSE_NAME_MAX)
1537                 goto err;
1538
1539         err = -EINVAL;
1540         if (size != sizeof(outarg) + outarg.namelen + 1)
1541                 goto err;
1542
1543         name.name = buf;
1544         name.len = outarg.namelen;
1545         err = fuse_copy_one(cs, buf, outarg.namelen + 1);
1546         if (err)
1547                 goto err;
1548         fuse_copy_finish(cs);
1549         buf[outarg.namelen] = 0;
1550
1551         down_read(&fc->killsb);
1552         err = fuse_reverse_inval_entry(fc, outarg.parent, outarg.child, &name, 0);
1553         up_read(&fc->killsb);
1554         kfree(buf);
1555         return err;
1556
1557 err:
1558         kfree(buf);
1559         fuse_copy_finish(cs);
1560         return err;
1561 }
1562
1563 static int fuse_notify_store(struct fuse_conn *fc, unsigned int size,
1564                              struct fuse_copy_state *cs)
1565 {
1566         struct fuse_notify_store_out outarg;
1567         struct inode *inode;
1568         struct address_space *mapping;
1569         u64 nodeid;
1570         int err;
1571         pgoff_t index;
1572         unsigned int offset;
1573         unsigned int num;
1574         loff_t file_size;
1575         loff_t end;
1576
1577         err = -EINVAL;
1578         if (size < sizeof(outarg))
1579                 goto out_finish;
1580
1581         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1582         if (err)
1583                 goto out_finish;
1584
1585         err = -EINVAL;
1586         if (size - sizeof(outarg) != outarg.size)
1587                 goto out_finish;
1588
1589         nodeid = outarg.nodeid;
1590
1591         down_read(&fc->killsb);
1592
1593         err = -ENOENT;
1594         inode = fuse_ilookup(fc, nodeid,  NULL);
1595         if (!inode)
1596                 goto out_up_killsb;
1597
1598         mapping = inode->i_mapping;
1599         index = outarg.offset >> PAGE_SHIFT;
1600         offset = outarg.offset & ~PAGE_MASK;
1601         file_size = i_size_read(inode);
1602         end = outarg.offset + outarg.size;
1603         if (end > file_size) {
1604                 file_size = end;
1605                 fuse_write_update_attr(inode, file_size, outarg.size);
1606         }
1607
1608         num = outarg.size;
1609         while (num) {
1610                 struct page *page;
1611                 unsigned int this_num;
1612
1613                 err = -ENOMEM;
1614                 page = find_or_create_page(mapping, index,
1615                                            mapping_gfp_mask(mapping));
1616                 if (!page)
1617                         goto out_iput;
1618
1619                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1620                 err = fuse_copy_page(cs, &page, offset, this_num, 0);
1621                 if (!PageUptodate(page) && !err && offset == 0 &&
1622                     (this_num == PAGE_SIZE || file_size == end)) {
1623                         zero_user_segment(page, this_num, PAGE_SIZE);
1624                         SetPageUptodate(page);
1625                 }
1626                 unlock_page(page);
1627                 put_page(page);
1628
1629                 if (err)
1630                         goto out_iput;
1631
1632                 num -= this_num;
1633                 offset = 0;
1634                 index++;
1635         }
1636
1637         err = 0;
1638
1639 out_iput:
1640         iput(inode);
1641 out_up_killsb:
1642         up_read(&fc->killsb);
1643 out_finish:
1644         fuse_copy_finish(cs);
1645         return err;
1646 }
1647
1648 struct fuse_retrieve_args {
1649         struct fuse_args_pages ap;
1650         struct fuse_notify_retrieve_in inarg;
1651 };
1652
1653 static void fuse_retrieve_end(struct fuse_mount *fm, struct fuse_args *args,
1654                               int error)
1655 {
1656         struct fuse_retrieve_args *ra =
1657                 container_of(args, typeof(*ra), ap.args);
1658
1659         release_pages(ra->ap.pages, ra->ap.num_pages);
1660         kfree(ra);
1661 }
1662
1663 static int fuse_retrieve(struct fuse_mount *fm, struct inode *inode,
1664                          struct fuse_notify_retrieve_out *outarg)
1665 {
1666         int err;
1667         struct address_space *mapping = inode->i_mapping;
1668         pgoff_t index;
1669         loff_t file_size;
1670         unsigned int num;
1671         unsigned int offset;
1672         size_t total_len = 0;
1673         unsigned int num_pages;
1674         struct fuse_conn *fc = fm->fc;
1675         struct fuse_retrieve_args *ra;
1676         size_t args_size = sizeof(*ra);
1677         struct fuse_args_pages *ap;
1678         struct fuse_args *args;
1679
1680         offset = outarg->offset & ~PAGE_MASK;
1681         file_size = i_size_read(inode);
1682
1683         num = min(outarg->size, fc->max_write);
1684         if (outarg->offset > file_size)
1685                 num = 0;
1686         else if (outarg->offset + num > file_size)
1687                 num = file_size - outarg->offset;
1688
1689         num_pages = (num + offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
1690         num_pages = min(num_pages, fc->max_pages);
1691
1692         args_size += num_pages * (sizeof(ap->pages[0]) + sizeof(ap->descs[0]));
1693
1694         ra = kzalloc(args_size, GFP_KERNEL);
1695         if (!ra)
1696                 return -ENOMEM;
1697
1698         ap = &ra->ap;
1699         ap->pages = (void *) (ra + 1);
1700         ap->descs = (void *) (ap->pages + num_pages);
1701
1702         args = &ap->args;
1703         args->nodeid = outarg->nodeid;
1704         args->opcode = FUSE_NOTIFY_REPLY;
1705         args->in_numargs = 2;
1706         args->in_pages = true;
1707         args->end = fuse_retrieve_end;
1708
1709         index = outarg->offset >> PAGE_SHIFT;
1710
1711         while (num && ap->num_pages < num_pages) {
1712                 struct page *page;
1713                 unsigned int this_num;
1714
1715                 page = find_get_page(mapping, index);
1716                 if (!page)
1717                         break;
1718
1719                 this_num = min_t(unsigned, num, PAGE_SIZE - offset);
1720                 ap->pages[ap->num_pages] = page;
1721                 ap->descs[ap->num_pages].offset = offset;
1722                 ap->descs[ap->num_pages].length = this_num;
1723                 ap->num_pages++;
1724
1725                 offset = 0;
1726                 num -= this_num;
1727                 total_len += this_num;
1728                 index++;
1729         }
1730         ra->inarg.offset = outarg->offset;
1731         ra->inarg.size = total_len;
1732         args->in_args[0].size = sizeof(ra->inarg);
1733         args->in_args[0].value = &ra->inarg;
1734         args->in_args[1].size = total_len;
1735
1736         err = fuse_simple_notify_reply(fm, args, outarg->notify_unique);
1737         if (err)
1738                 fuse_retrieve_end(fm, args, err);
1739
1740         return err;
1741 }
1742
1743 static int fuse_notify_retrieve(struct fuse_conn *fc, unsigned int size,
1744                                 struct fuse_copy_state *cs)
1745 {
1746         struct fuse_notify_retrieve_out outarg;
1747         struct fuse_mount *fm;
1748         struct inode *inode;
1749         u64 nodeid;
1750         int err;
1751
1752         err = -EINVAL;
1753         if (size != sizeof(outarg))
1754                 goto copy_finish;
1755
1756         err = fuse_copy_one(cs, &outarg, sizeof(outarg));
1757         if (err)
1758                 goto copy_finish;
1759
1760         fuse_copy_finish(cs);
1761
1762         down_read(&fc->killsb);
1763         err = -ENOENT;
1764         nodeid = outarg.nodeid;
1765
1766         inode = fuse_ilookup(fc, nodeid, &fm);
1767         if (inode) {
1768                 err = fuse_retrieve(fm, inode, &outarg);
1769                 iput(inode);
1770         }
1771         up_read(&fc->killsb);
1772
1773         return err;
1774
1775 copy_finish:
1776         fuse_copy_finish(cs);
1777         return err;
1778 }
1779
1780 /*
1781  * Resending all processing queue requests.
1782  *
1783  * During a FUSE daemon panics and failover, it is possible for some inflight
1784  * requests to be lost and never returned. As a result, applications awaiting
1785  * replies would become stuck forever. To address this, we can use notification
1786  * to trigger resending of these pending requests to the FUSE daemon, ensuring
1787  * they are properly processed again.
1788  *
1789  * Please note that this strategy is applicable only to idempotent requests or
1790  * if the FUSE daemon takes careful measures to avoid processing duplicated
1791  * non-idempotent requests.
1792  */
1793 static void fuse_resend(struct fuse_conn *fc)
1794 {
1795         struct fuse_dev *fud;
1796         struct fuse_req *req, *next;
1797         struct fuse_iqueue *fiq = &fc->iq;
1798         LIST_HEAD(to_queue);
1799         unsigned int i;
1800
1801         spin_lock(&fc->lock);
1802         if (!fc->connected) {
1803                 spin_unlock(&fc->lock);
1804                 return;
1805         }
1806
1807         list_for_each_entry(fud, &fc->devices, entry) {
1808                 struct fuse_pqueue *fpq = &fud->pq;
1809
1810                 spin_lock(&fpq->lock);
1811                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
1812                         list_splice_tail_init(&fpq->processing[i], &to_queue);
1813                 spin_unlock(&fpq->lock);
1814         }
1815         spin_unlock(&fc->lock);
1816
1817         list_for_each_entry_safe(req, next, &to_queue, list) {
1818                 set_bit(FR_PENDING, &req->flags);
1819                 clear_bit(FR_SENT, &req->flags);
1820                 /* mark the request as resend request */
1821                 req->in.h.unique |= FUSE_UNIQUE_RESEND;
1822         }
1823
1824         spin_lock(&fiq->lock);
1825         /* iq and pq requests are both oldest to newest */
1826         list_splice(&to_queue, &fiq->pending);
1827         fiq->ops->wake_pending_and_unlock(fiq);
1828 }
1829
1830 static int fuse_notify_resend(struct fuse_conn *fc)
1831 {
1832         fuse_resend(fc);
1833         return 0;
1834 }
1835
1836 static int fuse_notify(struct fuse_conn *fc, enum fuse_notify_code code,
1837                        unsigned int size, struct fuse_copy_state *cs)
1838 {
1839         /* Don't try to move pages (yet) */
1840         cs->move_pages = 0;
1841
1842         switch (code) {
1843         case FUSE_NOTIFY_POLL:
1844                 return fuse_notify_poll(fc, size, cs);
1845
1846         case FUSE_NOTIFY_INVAL_INODE:
1847                 return fuse_notify_inval_inode(fc, size, cs);
1848
1849         case FUSE_NOTIFY_INVAL_ENTRY:
1850                 return fuse_notify_inval_entry(fc, size, cs);
1851
1852         case FUSE_NOTIFY_STORE:
1853                 return fuse_notify_store(fc, size, cs);
1854
1855         case FUSE_NOTIFY_RETRIEVE:
1856                 return fuse_notify_retrieve(fc, size, cs);
1857
1858         case FUSE_NOTIFY_DELETE:
1859                 return fuse_notify_delete(fc, size, cs);
1860
1861         case FUSE_NOTIFY_RESEND:
1862                 return fuse_notify_resend(fc);
1863
1864         default:
1865                 fuse_copy_finish(cs);
1866                 return -EINVAL;
1867         }
1868 }
1869
1870 /* Look up request on processing list by unique ID */
1871 static struct fuse_req *request_find(struct fuse_pqueue *fpq, u64 unique)
1872 {
1873         unsigned int hash = fuse_req_hash(unique);
1874         struct fuse_req *req;
1875
1876         list_for_each_entry(req, &fpq->processing[hash], list) {
1877                 if (req->in.h.unique == unique)
1878                         return req;
1879         }
1880         return NULL;
1881 }
1882
1883 static int copy_out_args(struct fuse_copy_state *cs, struct fuse_args *args,
1884                          unsigned nbytes)
1885 {
1886         unsigned reqsize = sizeof(struct fuse_out_header);
1887
1888         reqsize += fuse_len_args(args->out_numargs, args->out_args);
1889
1890         if (reqsize < nbytes || (reqsize > nbytes && !args->out_argvar))
1891                 return -EINVAL;
1892         else if (reqsize > nbytes) {
1893                 struct fuse_arg *lastarg = &args->out_args[args->out_numargs-1];
1894                 unsigned diffsize = reqsize - nbytes;
1895
1896                 if (diffsize > lastarg->size)
1897                         return -EINVAL;
1898                 lastarg->size -= diffsize;
1899         }
1900         return fuse_copy_args(cs, args->out_numargs, args->out_pages,
1901                               args->out_args, args->page_zeroing);
1902 }
1903
1904 /*
1905  * Write a single reply to a request.  First the header is copied from
1906  * the write buffer.  The request is then searched on the processing
1907  * list by the unique ID found in the header.  If found, then remove
1908  * it from the list and copy the rest of the buffer to the request.
1909  * The request is finished by calling fuse_request_end().
1910  */
1911 static ssize_t fuse_dev_do_write(struct fuse_dev *fud,
1912                                  struct fuse_copy_state *cs, size_t nbytes)
1913 {
1914         int err;
1915         struct fuse_conn *fc = fud->fc;
1916         struct fuse_pqueue *fpq = &fud->pq;
1917         struct fuse_req *req;
1918         struct fuse_out_header oh;
1919
1920         err = -EINVAL;
1921         if (nbytes < sizeof(struct fuse_out_header))
1922                 goto out;
1923
1924         err = fuse_copy_one(cs, &oh, sizeof(oh));
1925         if (err)
1926                 goto copy_finish;
1927
1928         err = -EINVAL;
1929         if (oh.len != nbytes)
1930                 goto copy_finish;
1931
1932         /*
1933          * Zero oh.unique indicates unsolicited notification message
1934          * and error contains notification code.
1935          */
1936         if (!oh.unique) {
1937                 err = fuse_notify(fc, oh.error, nbytes - sizeof(oh), cs);
1938                 goto out;
1939         }
1940
1941         err = -EINVAL;
1942         if (oh.error <= -512 || oh.error > 0)
1943                 goto copy_finish;
1944
1945         spin_lock(&fpq->lock);
1946         req = NULL;
1947         if (fpq->connected)
1948                 req = request_find(fpq, oh.unique & ~FUSE_INT_REQ_BIT);
1949
1950         err = -ENOENT;
1951         if (!req) {
1952                 spin_unlock(&fpq->lock);
1953                 goto copy_finish;
1954         }
1955
1956         /* Is it an interrupt reply ID? */
1957         if (oh.unique & FUSE_INT_REQ_BIT) {
1958                 __fuse_get_request(req);
1959                 spin_unlock(&fpq->lock);
1960
1961                 err = 0;
1962                 if (nbytes != sizeof(struct fuse_out_header))
1963                         err = -EINVAL;
1964                 else if (oh.error == -ENOSYS)
1965                         fc->no_interrupt = 1;
1966                 else if (oh.error == -EAGAIN)
1967                         err = queue_interrupt(req);
1968
1969                 fuse_put_request(req);
1970
1971                 goto copy_finish;
1972         }
1973
1974         clear_bit(FR_SENT, &req->flags);
1975         list_move(&req->list, &fpq->io);
1976         req->out.h = oh;
1977         set_bit(FR_LOCKED, &req->flags);
1978         spin_unlock(&fpq->lock);
1979         cs->req = req;
1980         if (!req->args->page_replace)
1981                 cs->move_pages = 0;
1982
1983         if (oh.error)
1984                 err = nbytes != sizeof(oh) ? -EINVAL : 0;
1985         else
1986                 err = copy_out_args(cs, req->args, nbytes);
1987         fuse_copy_finish(cs);
1988
1989         spin_lock(&fpq->lock);
1990         clear_bit(FR_LOCKED, &req->flags);
1991         if (!fpq->connected)
1992                 err = -ENOENT;
1993         else if (err)
1994                 req->out.h.error = -EIO;
1995         if (!test_bit(FR_PRIVATE, &req->flags))
1996                 list_del_init(&req->list);
1997         spin_unlock(&fpq->lock);
1998
1999         fuse_request_end(req);
2000 out:
2001         return err ? err : nbytes;
2002
2003 copy_finish:
2004         fuse_copy_finish(cs);
2005         goto out;
2006 }
2007
2008 static ssize_t fuse_dev_write(struct kiocb *iocb, struct iov_iter *from)
2009 {
2010         struct fuse_copy_state cs;
2011         struct fuse_dev *fud = fuse_get_dev(iocb->ki_filp);
2012
2013         if (!fud)
2014                 return -EPERM;
2015
2016         if (!user_backed_iter(from))
2017                 return -EINVAL;
2018
2019         fuse_copy_init(&cs, 0, from);
2020
2021         return fuse_dev_do_write(fud, &cs, iov_iter_count(from));
2022 }
2023
2024 static ssize_t fuse_dev_splice_write(struct pipe_inode_info *pipe,
2025                                      struct file *out, loff_t *ppos,
2026                                      size_t len, unsigned int flags)
2027 {
2028         unsigned int head, tail, mask, count;
2029         unsigned nbuf;
2030         unsigned idx;
2031         struct pipe_buffer *bufs;
2032         struct fuse_copy_state cs;
2033         struct fuse_dev *fud;
2034         size_t rem;
2035         ssize_t ret;
2036
2037         fud = fuse_get_dev(out);
2038         if (!fud)
2039                 return -EPERM;
2040
2041         pipe_lock(pipe);
2042
2043         head = pipe->head;
2044         tail = pipe->tail;
2045         mask = pipe->ring_size - 1;
2046         count = head - tail;
2047
2048         bufs = kvmalloc_array(count, sizeof(struct pipe_buffer), GFP_KERNEL);
2049         if (!bufs) {
2050                 pipe_unlock(pipe);
2051                 return -ENOMEM;
2052         }
2053
2054         nbuf = 0;
2055         rem = 0;
2056         for (idx = tail; idx != head && rem < len; idx++)
2057                 rem += pipe->bufs[idx & mask].len;
2058
2059         ret = -EINVAL;
2060         if (rem < len)
2061                 goto out_free;
2062
2063         rem = len;
2064         while (rem) {
2065                 struct pipe_buffer *ibuf;
2066                 struct pipe_buffer *obuf;
2067
2068                 if (WARN_ON(nbuf >= count || tail == head))
2069                         goto out_free;
2070
2071                 ibuf = &pipe->bufs[tail & mask];
2072                 obuf = &bufs[nbuf];
2073
2074                 if (rem >= ibuf->len) {
2075                         *obuf = *ibuf;
2076                         ibuf->ops = NULL;
2077                         tail++;
2078                         pipe->tail = tail;
2079                 } else {
2080                         if (!pipe_buf_get(pipe, ibuf))
2081                                 goto out_free;
2082
2083                         *obuf = *ibuf;
2084                         obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
2085                         obuf->len = rem;
2086                         ibuf->offset += obuf->len;
2087                         ibuf->len -= obuf->len;
2088                 }
2089                 nbuf++;
2090                 rem -= obuf->len;
2091         }
2092         pipe_unlock(pipe);
2093
2094         fuse_copy_init(&cs, 0, NULL);
2095         cs.pipebufs = bufs;
2096         cs.nr_segs = nbuf;
2097         cs.pipe = pipe;
2098
2099         if (flags & SPLICE_F_MOVE)
2100                 cs.move_pages = 1;
2101
2102         ret = fuse_dev_do_write(fud, &cs, len);
2103
2104         pipe_lock(pipe);
2105 out_free:
2106         for (idx = 0; idx < nbuf; idx++) {
2107                 struct pipe_buffer *buf = &bufs[idx];
2108
2109                 if (buf->ops)
2110                         pipe_buf_release(pipe, buf);
2111         }
2112         pipe_unlock(pipe);
2113
2114         kvfree(bufs);
2115         return ret;
2116 }
2117
2118 static __poll_t fuse_dev_poll(struct file *file, poll_table *wait)
2119 {
2120         __poll_t mask = EPOLLOUT | EPOLLWRNORM;
2121         struct fuse_iqueue *fiq;
2122         struct fuse_dev *fud = fuse_get_dev(file);
2123
2124         if (!fud)
2125                 return EPOLLERR;
2126
2127         fiq = &fud->fc->iq;
2128         poll_wait(file, &fiq->waitq, wait);
2129
2130         spin_lock(&fiq->lock);
2131         if (!fiq->connected)
2132                 mask = EPOLLERR;
2133         else if (request_pending(fiq))
2134                 mask |= EPOLLIN | EPOLLRDNORM;
2135         spin_unlock(&fiq->lock);
2136
2137         return mask;
2138 }
2139
2140 /* Abort all requests on the given list (pending or processing) */
2141 static void end_requests(struct list_head *head)
2142 {
2143         while (!list_empty(head)) {
2144                 struct fuse_req *req;
2145                 req = list_entry(head->next, struct fuse_req, list);
2146                 req->out.h.error = -ECONNABORTED;
2147                 clear_bit(FR_SENT, &req->flags);
2148                 list_del_init(&req->list);
2149                 fuse_request_end(req);
2150         }
2151 }
2152
2153 static void end_polls(struct fuse_conn *fc)
2154 {
2155         struct rb_node *p;
2156
2157         p = rb_first(&fc->polled_files);
2158
2159         while (p) {
2160                 struct fuse_file *ff;
2161                 ff = rb_entry(p, struct fuse_file, polled_node);
2162                 wake_up_interruptible_all(&ff->poll_wait);
2163
2164                 p = rb_next(p);
2165         }
2166 }
2167
2168 /*
2169  * Abort all requests.
2170  *
2171  * Emergency exit in case of a malicious or accidental deadlock, or just a hung
2172  * filesystem.
2173  *
2174  * The same effect is usually achievable through killing the filesystem daemon
2175  * and all users of the filesystem.  The exception is the combination of an
2176  * asynchronous request and the tricky deadlock (see
2177  * Documentation/filesystems/fuse.rst).
2178  *
2179  * Aborting requests under I/O goes as follows: 1: Separate out unlocked
2180  * requests, they should be finished off immediately.  Locked requests will be
2181  * finished after unlock; see unlock_request(). 2: Finish off the unlocked
2182  * requests.  It is possible that some request will finish before we can.  This
2183  * is OK, the request will in that case be removed from the list before we touch
2184  * it.
2185  */
2186 void fuse_abort_conn(struct fuse_conn *fc)
2187 {
2188         struct fuse_iqueue *fiq = &fc->iq;
2189
2190         spin_lock(&fc->lock);
2191         if (fc->connected) {
2192                 struct fuse_dev *fud;
2193                 struct fuse_req *req, *next;
2194                 LIST_HEAD(to_end);
2195                 unsigned int i;
2196
2197                 /* Background queuing checks fc->connected under bg_lock */
2198                 spin_lock(&fc->bg_lock);
2199                 fc->connected = 0;
2200                 spin_unlock(&fc->bg_lock);
2201
2202                 fuse_set_initialized(fc);
2203                 list_for_each_entry(fud, &fc->devices, entry) {
2204                         struct fuse_pqueue *fpq = &fud->pq;
2205
2206                         spin_lock(&fpq->lock);
2207                         fpq->connected = 0;
2208                         list_for_each_entry_safe(req, next, &fpq->io, list) {
2209                                 req->out.h.error = -ECONNABORTED;
2210                                 spin_lock(&req->waitq.lock);
2211                                 set_bit(FR_ABORTED, &req->flags);
2212                                 if (!test_bit(FR_LOCKED, &req->flags)) {
2213                                         set_bit(FR_PRIVATE, &req->flags);
2214                                         __fuse_get_request(req);
2215                                         list_move(&req->list, &to_end);
2216                                 }
2217                                 spin_unlock(&req->waitq.lock);
2218                         }
2219                         for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2220                                 list_splice_tail_init(&fpq->processing[i],
2221                                                       &to_end);
2222                         spin_unlock(&fpq->lock);
2223                 }
2224                 spin_lock(&fc->bg_lock);
2225                 fc->blocked = 0;
2226                 fc->max_background = UINT_MAX;
2227                 flush_bg_queue(fc);
2228                 spin_unlock(&fc->bg_lock);
2229
2230                 spin_lock(&fiq->lock);
2231                 fiq->connected = 0;
2232                 list_for_each_entry(req, &fiq->pending, list)
2233                         clear_bit(FR_PENDING, &req->flags);
2234                 list_splice_tail_init(&fiq->pending, &to_end);
2235                 while (forget_pending(fiq))
2236                         kfree(fuse_dequeue_forget(fiq, 1, NULL));
2237                 wake_up_all(&fiq->waitq);
2238                 spin_unlock(&fiq->lock);
2239                 kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
2240                 end_polls(fc);
2241                 wake_up_all(&fc->blocked_waitq);
2242                 spin_unlock(&fc->lock);
2243
2244                 end_requests(&to_end);
2245         } else {
2246                 spin_unlock(&fc->lock);
2247         }
2248 }
2249 EXPORT_SYMBOL_GPL(fuse_abort_conn);
2250
2251 void fuse_wait_aborted(struct fuse_conn *fc)
2252 {
2253         /* matches implicit memory barrier in fuse_drop_waiting() */
2254         smp_mb();
2255         wait_event(fc->blocked_waitq, atomic_read(&fc->num_waiting) == 0);
2256 }
2257
2258 int fuse_dev_release(struct inode *inode, struct file *file)
2259 {
2260         struct fuse_dev *fud = fuse_get_dev(file);
2261
2262         if (fud) {
2263                 struct fuse_conn *fc = fud->fc;
2264                 struct fuse_pqueue *fpq = &fud->pq;
2265                 LIST_HEAD(to_end);
2266                 unsigned int i;
2267
2268                 spin_lock(&fpq->lock);
2269                 WARN_ON(!list_empty(&fpq->io));
2270                 for (i = 0; i < FUSE_PQ_HASH_SIZE; i++)
2271                         list_splice_init(&fpq->processing[i], &to_end);
2272                 spin_unlock(&fpq->lock);
2273
2274                 end_requests(&to_end);
2275
2276                 /* Are we the last open device? */
2277                 if (atomic_dec_and_test(&fc->dev_count)) {
2278                         WARN_ON(fc->iq.fasync != NULL);
2279                         fuse_abort_conn(fc);
2280                 }
2281                 fuse_dev_free(fud);
2282         }
2283         return 0;
2284 }
2285 EXPORT_SYMBOL_GPL(fuse_dev_release);
2286
2287 static int fuse_dev_fasync(int fd, struct file *file, int on)
2288 {
2289         struct fuse_dev *fud = fuse_get_dev(file);
2290
2291         if (!fud)
2292                 return -EPERM;
2293
2294         /* No locking - fasync_helper does its own locking */
2295         return fasync_helper(fd, file, on, &fud->fc->iq.fasync);
2296 }
2297
2298 static int fuse_device_clone(struct fuse_conn *fc, struct file *new)
2299 {
2300         struct fuse_dev *fud;
2301
2302         if (new->private_data)
2303                 return -EINVAL;
2304
2305         fud = fuse_dev_alloc_install(fc);
2306         if (!fud)
2307                 return -ENOMEM;
2308
2309         new->private_data = fud;
2310         atomic_inc(&fc->dev_count);
2311
2312         return 0;
2313 }
2314
2315 static long fuse_dev_ioctl_clone(struct file *file, __u32 __user *argp)
2316 {
2317         int res;
2318         int oldfd;
2319         struct fuse_dev *fud = NULL;
2320         struct fd f;
2321
2322         if (get_user(oldfd, argp))
2323                 return -EFAULT;
2324
2325         f = fdget(oldfd);
2326         if (!f.file)
2327                 return -EINVAL;
2328
2329         /*
2330          * Check against file->f_op because CUSE
2331          * uses the same ioctl handler.
2332          */
2333         if (f.file->f_op == file->f_op)
2334                 fud = fuse_get_dev(f.file);
2335
2336         res = -EINVAL;
2337         if (fud) {
2338                 mutex_lock(&fuse_mutex);
2339                 res = fuse_device_clone(fud->fc, file);
2340                 mutex_unlock(&fuse_mutex);
2341         }
2342
2343         fdput(f);
2344         return res;
2345 }
2346
2347 static long fuse_dev_ioctl_backing_open(struct file *file,
2348                                         struct fuse_backing_map __user *argp)
2349 {
2350         struct fuse_dev *fud = fuse_get_dev(file);
2351         struct fuse_backing_map map;
2352
2353         if (!fud)
2354                 return -EPERM;
2355
2356         if (!IS_ENABLED(CONFIG_FUSE_PASSTHROUGH))
2357                 return -EOPNOTSUPP;
2358
2359         if (copy_from_user(&map, argp, sizeof(map)))
2360                 return -EFAULT;
2361
2362         return fuse_backing_open(fud->fc, &map);
2363 }
2364
2365 static long fuse_dev_ioctl_backing_close(struct file *file, __u32 __user *argp)
2366 {
2367         struct fuse_dev *fud = fuse_get_dev(file);
2368         int backing_id;
2369
2370         if (!fud)
2371                 return -EPERM;
2372
2373         if (!IS_ENABLED(CONFIG_FUSE_PASSTHROUGH))
2374                 return -EOPNOTSUPP;
2375
2376         if (get_user(backing_id, argp))
2377                 return -EFAULT;
2378
2379         return fuse_backing_close(fud->fc, backing_id);
2380 }
2381
2382 static long fuse_dev_ioctl(struct file *file, unsigned int cmd,
2383                            unsigned long arg)
2384 {
2385         void __user *argp = (void __user *)arg;
2386
2387         switch (cmd) {
2388         case FUSE_DEV_IOC_CLONE:
2389                 return fuse_dev_ioctl_clone(file, argp);
2390
2391         case FUSE_DEV_IOC_BACKING_OPEN:
2392                 return fuse_dev_ioctl_backing_open(file, argp);
2393
2394         case FUSE_DEV_IOC_BACKING_CLOSE:
2395                 return fuse_dev_ioctl_backing_close(file, argp);
2396
2397         default:
2398                 return -ENOTTY;
2399         }
2400 }
2401
2402 const struct file_operations fuse_dev_operations = {
2403         .owner          = THIS_MODULE,
2404         .open           = fuse_dev_open,
2405         .llseek         = no_llseek,
2406         .read_iter      = fuse_dev_read,
2407         .splice_read    = fuse_dev_splice_read,
2408         .write_iter     = fuse_dev_write,
2409         .splice_write   = fuse_dev_splice_write,
2410         .poll           = fuse_dev_poll,
2411         .release        = fuse_dev_release,
2412         .fasync         = fuse_dev_fasync,
2413         .unlocked_ioctl = fuse_dev_ioctl,
2414         .compat_ioctl   = compat_ptr_ioctl,
2415 };
2416 EXPORT_SYMBOL_GPL(fuse_dev_operations);
2417
2418 static struct miscdevice fuse_miscdevice = {
2419         .minor = FUSE_MINOR,
2420         .name  = "fuse",
2421         .fops = &fuse_dev_operations,
2422 };
2423
2424 int __init fuse_dev_init(void)
2425 {
2426         int err = -ENOMEM;
2427         fuse_req_cachep = kmem_cache_create("fuse_request",
2428                                             sizeof(struct fuse_req),
2429                                             0, 0, NULL);
2430         if (!fuse_req_cachep)
2431                 goto out;
2432
2433         err = misc_register(&fuse_miscdevice);
2434         if (err)
2435                 goto out_cache_clean;
2436
2437         return 0;
2438
2439  out_cache_clean:
2440         kmem_cache_destroy(fuse_req_cachep);
2441  out:
2442         return err;
2443 }
2444
2445 void fuse_dev_cleanup(void)
2446 {
2447         misc_deregister(&fuse_miscdevice);
2448         kmem_cache_destroy(fuse_req_cachep);
2449 }