linux/fs/fuse/file.c
<<
>>
Prefs
   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/pagemap.h>
  12#include <linux/slab.h>
  13#include <linux/kernel.h>
  14#include <linux/sched.h>
  15#include <linux/module.h>
  16#include <linux/compat.h>
  17#include <linux/swap.h>
  18#include <linux/aio.h>
  19#include <linux/falloc.h>
  20
  21static const struct file_operations fuse_direct_io_file_operations;
  22
  23static int fuse_send_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
  24                          int opcode, struct fuse_open_out *outargp)
  25{
  26        struct fuse_open_in inarg;
  27        struct fuse_req *req;
  28        int err;
  29
  30        req = fuse_get_req_nopages(fc);
  31        if (IS_ERR(req))
  32                return PTR_ERR(req);
  33
  34        memset(&inarg, 0, sizeof(inarg));
  35        inarg.flags = file->f_flags & ~(O_CREAT | O_EXCL | O_NOCTTY);
  36        if (!fc->atomic_o_trunc)
  37                inarg.flags &= ~O_TRUNC;
  38        req->in.h.opcode = opcode;
  39        req->in.h.nodeid = nodeid;
  40        req->in.numargs = 1;
  41        req->in.args[0].size = sizeof(inarg);
  42        req->in.args[0].value = &inarg;
  43        req->out.numargs = 1;
  44        req->out.args[0].size = sizeof(*outargp);
  45        req->out.args[0].value = outargp;
  46        fuse_request_send(fc, req);
  47        err = req->out.h.error;
  48        fuse_put_request(fc, req);
  49
  50        return err;
  51}
  52
  53struct fuse_file *fuse_file_alloc(struct fuse_conn *fc)
  54{
  55        struct fuse_file *ff;
  56
  57        ff = kmalloc(sizeof(struct fuse_file), GFP_KERNEL);
  58        if (unlikely(!ff))
  59                return NULL;
  60
  61        ff->fc = fc;
  62        ff->reserved_req = fuse_request_alloc(0);
  63        if (unlikely(!ff->reserved_req)) {
  64                kfree(ff);
  65                return NULL;
  66        }
  67
  68        INIT_LIST_HEAD(&ff->write_entry);
  69        atomic_set(&ff->count, 0);
  70        RB_CLEAR_NODE(&ff->polled_node);
  71        init_waitqueue_head(&ff->poll_wait);
  72
  73        spin_lock(&fc->lock);
  74        ff->kh = ++fc->khctr;
  75        spin_unlock(&fc->lock);
  76
  77        return ff;
  78}
  79
  80void fuse_file_free(struct fuse_file *ff)
  81{
  82        fuse_request_free(ff->reserved_req);
  83        kfree(ff);
  84}
  85
  86struct fuse_file *fuse_file_get(struct fuse_file *ff)
  87{
  88        atomic_inc(&ff->count);
  89        return ff;
  90}
  91
  92static void fuse_release_async(struct work_struct *work)
  93{
  94        struct fuse_req *req;
  95        struct fuse_conn *fc;
  96        struct path path;
  97
  98        req = container_of(work, struct fuse_req, misc.release.work);
  99        path = req->misc.release.path;
 100        fc = get_fuse_conn(path.dentry->d_inode);
 101
 102        fuse_put_request(fc, req);
 103        path_put(&path);
 104}
 105
 106static void fuse_release_end(struct fuse_conn *fc, struct fuse_req *req)
 107{
 108        if (fc->destroy_req) {
 109                /*
 110                 * If this is a fuseblk mount, then it's possible that
 111                 * releasing the path will result in releasing the
 112                 * super block and sending the DESTROY request.  If
 113                 * the server is single threaded, this would hang.
 114                 * For this reason do the path_put() in a separate
 115                 * thread.
 116                 */
 117                atomic_inc(&req->count);
 118                INIT_WORK(&req->misc.release.work, fuse_release_async);
 119                schedule_work(&req->misc.release.work);
 120        } else {
 121                path_put(&req->misc.release.path);
 122        }
 123}
 124
 125static void fuse_file_put(struct fuse_file *ff, bool sync)
 126{
 127        if (atomic_dec_and_test(&ff->count)) {
 128                struct fuse_req *req = ff->reserved_req;
 129
 130                if (ff->fc->no_open) {
 131                        /*
 132                         * Drop the release request when client does not
 133                         * implement 'open'
 134                         */
 135                        req->background = 0;
 136                        path_put(&req->misc.release.path);
 137                        fuse_put_request(ff->fc, req);
 138                } else if (sync) {
 139                        req->background = 0;
 140                        fuse_request_send(ff->fc, req);
 141                        path_put(&req->misc.release.path);
 142                        fuse_put_request(ff->fc, req);
 143                } else {
 144                        req->end = fuse_release_end;
 145                        req->background = 1;
 146                        fuse_request_send_background(ff->fc, req);
 147                }
 148                kfree(ff);
 149        }
 150}
 151
 152int fuse_do_open(struct fuse_conn *fc, u64 nodeid, struct file *file,
 153                 bool isdir)
 154{
 155        struct fuse_file *ff;
 156        int opcode = isdir ? FUSE_OPENDIR : FUSE_OPEN;
 157
 158        ff = fuse_file_alloc(fc);
 159        if (!ff)
 160                return -ENOMEM;
 161
 162        ff->fh = 0;
 163        ff->open_flags = FOPEN_KEEP_CACHE; /* Default for no-open */
 164        if (!fc->no_open || isdir) {
 165                struct fuse_open_out outarg;
 166                int err;
 167
 168                err = fuse_send_open(fc, nodeid, file, opcode, &outarg);
 169                if (!err) {
 170                        ff->fh = outarg.fh;
 171                        ff->open_flags = outarg.open_flags;
 172
 173                } else if (err != -ENOSYS || isdir) {
 174                        fuse_file_free(ff);
 175                        return err;
 176                } else {
 177                        fc->no_open = 1;
 178                }
 179        }
 180
 181        if (isdir)
 182                ff->open_flags &= ~FOPEN_DIRECT_IO;
 183
 184        ff->nodeid = nodeid;
 185        file->private_data = fuse_file_get(ff);
 186
 187        return 0;
 188}
 189EXPORT_SYMBOL_GPL(fuse_do_open);
 190
 191static void fuse_link_write_file(struct file *file)
 192{
 193        struct inode *inode = file_inode(file);
 194        struct fuse_conn *fc = get_fuse_conn(inode);
 195        struct fuse_inode *fi = get_fuse_inode(inode);
 196        struct fuse_file *ff = file->private_data;
 197        /*
 198         * file may be written through mmap, so chain it onto the
 199         * inodes's write_file list
 200         */
 201        spin_lock(&fc->lock);
 202        if (list_empty(&ff->write_entry))
 203                list_add(&ff->write_entry, &fi->write_files);
 204        spin_unlock(&fc->lock);
 205}
 206
 207void fuse_finish_open(struct inode *inode, struct file *file)
 208{
 209        struct fuse_file *ff = file->private_data;
 210        struct fuse_conn *fc = get_fuse_conn(inode);
 211
 212        if (ff->open_flags & FOPEN_DIRECT_IO)
 213                file->f_op = &fuse_direct_io_file_operations;
 214        if (!(ff->open_flags & FOPEN_KEEP_CACHE))
 215                invalidate_inode_pages2(inode->i_mapping);
 216        if (ff->open_flags & FOPEN_NONSEEKABLE)
 217                nonseekable_open(inode, file);
 218        if (fc->atomic_o_trunc && (file->f_flags & O_TRUNC)) {
 219                struct fuse_inode *fi = get_fuse_inode(inode);
 220
 221                spin_lock(&fc->lock);
 222                fi->attr_version = ++fc->attr_version;
 223                i_size_write(inode, 0);
 224                spin_unlock(&fc->lock);
 225                fuse_invalidate_attr(inode);
 226                if (fc->writeback_cache)
 227                        file_update_time(file);
 228        }
 229        if ((file->f_mode & FMODE_WRITE) && fc->writeback_cache)
 230                fuse_link_write_file(file);
 231}
 232
 233int fuse_open_common(struct inode *inode, struct file *file, bool isdir)
 234{
 235        struct fuse_conn *fc = get_fuse_conn(inode);
 236        int err;
 237        bool lock_inode = (file->f_flags & O_TRUNC) &&
 238                          fc->atomic_o_trunc &&
 239                          fc->writeback_cache;
 240
 241        err = generic_file_open(inode, file);
 242        if (err)
 243                return err;
 244
 245        if (lock_inode)
 246                mutex_lock(&inode->i_mutex);
 247
 248        err = fuse_do_open(fc, get_node_id(inode), file, isdir);
 249
 250        if (!err)
 251                fuse_finish_open(inode, file);
 252
 253        if (lock_inode)
 254                mutex_unlock(&inode->i_mutex);
 255
 256        return err;
 257}
 258
 259static void fuse_prepare_release(struct fuse_file *ff, int flags, int opcode)
 260{
 261        struct fuse_conn *fc = ff->fc;
 262        struct fuse_req *req = ff->reserved_req;
 263        struct fuse_release_in *inarg = &req->misc.release.in;
 264
 265        spin_lock(&fc->lock);
 266        list_del(&ff->write_entry);
 267        if (!RB_EMPTY_NODE(&ff->polled_node))
 268                rb_erase(&ff->polled_node, &fc->polled_files);
 269        spin_unlock(&fc->lock);
 270
 271        wake_up_interruptible_all(&ff->poll_wait);
 272
 273        inarg->fh = ff->fh;
 274        inarg->flags = flags;
 275        req->in.h.opcode = opcode;
 276        req->in.h.nodeid = ff->nodeid;
 277        req->in.numargs = 1;
 278        req->in.args[0].size = sizeof(struct fuse_release_in);
 279        req->in.args[0].value = inarg;
 280}
 281
 282void fuse_release_common(struct file *file, int opcode)
 283{
 284        struct fuse_file *ff;
 285        struct fuse_req *req;
 286
 287        ff = file->private_data;
 288        if (unlikely(!ff))
 289                return;
 290
 291        req = ff->reserved_req;
 292        fuse_prepare_release(ff, file->f_flags, opcode);
 293
 294        if (ff->flock) {
 295                struct fuse_release_in *inarg = &req->misc.release.in;
 296                inarg->release_flags |= FUSE_RELEASE_FLOCK_UNLOCK;
 297                inarg->lock_owner = fuse_lock_owner_id(ff->fc,
 298                                                       (fl_owner_t) file);
 299        }
 300        /* Hold vfsmount and dentry until release is finished */
 301        path_get(&file->f_path);
 302        req->misc.release.path = file->f_path;
 303
 304        /*
 305         * Normally this will send the RELEASE request, however if
 306         * some asynchronous READ or WRITE requests are outstanding,
 307         * the sending will be delayed.
 308         *
 309         * Make the release synchronous if this is a fuseblk mount,
 310         * synchronous RELEASE is allowed (and desirable) in this case
 311         * because the server can be trusted not to screw up.
 312         */
 313        fuse_file_put(ff, ff->fc->destroy_req != NULL);
 314}
 315
 316static int fuse_open(struct inode *inode, struct file *file)
 317{
 318        return fuse_open_common(inode, file, false);
 319}
 320
 321static int fuse_release(struct inode *inode, struct file *file)
 322{
 323        struct fuse_conn *fc = get_fuse_conn(inode);
 324
 325        /* see fuse_vma_close() for !writeback_cache case */
 326        if (fc->writeback_cache)
 327                write_inode_now(inode, 1);
 328
 329        fuse_release_common(file, FUSE_RELEASE);
 330
 331        /* return value is ignored by VFS */
 332        return 0;
 333}
 334
 335void fuse_sync_release(struct fuse_file *ff, int flags)
 336{
 337        WARN_ON(atomic_read(&ff->count) > 1);
 338        fuse_prepare_release(ff, flags, FUSE_RELEASE);
 339        ff->reserved_req->force = 1;
 340        ff->reserved_req->background = 0;
 341        fuse_request_send(ff->fc, ff->reserved_req);
 342        fuse_put_request(ff->fc, ff->reserved_req);
 343        kfree(ff);
 344}
 345EXPORT_SYMBOL_GPL(fuse_sync_release);
 346
 347/*
 348 * Scramble the ID space with XTEA, so that the value of the files_struct
 349 * pointer is not exposed to userspace.
 350 */
 351u64 fuse_lock_owner_id(struct fuse_conn *fc, fl_owner_t id)
 352{
 353        u32 *k = fc->scramble_key;
 354        u64 v = (unsigned long) id;
 355        u32 v0 = v;
 356        u32 v1 = v >> 32;
 357        u32 sum = 0;
 358        int i;
 359
 360        for (i = 0; i < 32; i++) {
 361                v0 += ((v1 << 4 ^ v1 >> 5) + v1) ^ (sum + k[sum & 3]);
 362                sum += 0x9E3779B9;
 363                v1 += ((v0 << 4 ^ v0 >> 5) + v0) ^ (sum + k[sum>>11 & 3]);
 364        }
 365
 366        return (u64) v0 + ((u64) v1 << 32);
 367}
 368
 369/*
 370 * Check if any page in a range is under writeback
 371 *
 372 * This is currently done by walking the list of writepage requests
 373 * for the inode, which can be pretty inefficient.
 374 */
 375static bool fuse_range_is_writeback(struct inode *inode, pgoff_t idx_from,
 376                                   pgoff_t idx_to)
 377{
 378        struct fuse_conn *fc = get_fuse_conn(inode);
 379        struct fuse_inode *fi = get_fuse_inode(inode);
 380        struct fuse_req *req;
 381        bool found = false;
 382
 383        spin_lock(&fc->lock);
 384        list_for_each_entry(req, &fi->writepages, writepages_entry) {
 385                pgoff_t curr_index;
 386
 387                BUG_ON(req->inode != inode);
 388                curr_index = req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
 389                if (idx_from < curr_index + req->num_pages &&
 390                    curr_index <= idx_to) {
 391                        found = true;
 392                        break;
 393                }
 394        }
 395        spin_unlock(&fc->lock);
 396
 397        return found;
 398}
 399
 400static inline bool fuse_page_is_writeback(struct inode *inode, pgoff_t index)
 401{
 402        return fuse_range_is_writeback(inode, index, index);
 403}
 404
 405/*
 406 * Wait for page writeback to be completed.
 407 *
 408 * Since fuse doesn't rely on the VM writeback tracking, this has to
 409 * use some other means.
 410 */
 411static int fuse_wait_on_page_writeback(struct inode *inode, pgoff_t index)
 412{
 413        struct fuse_inode *fi = get_fuse_inode(inode);
 414
 415        wait_event(fi->page_waitq, !fuse_page_is_writeback(inode, index));
 416        return 0;
 417}
 418
 419/*
 420 * Wait for all pending writepages on the inode to finish.
 421 *
 422 * This is currently done by blocking further writes with FUSE_NOWRITE
 423 * and waiting for all sent writes to complete.
 424 *
 425 * This must be called under i_mutex, otherwise the FUSE_NOWRITE usage
 426 * could conflict with truncation.
 427 */
 428static void fuse_sync_writes(struct inode *inode)
 429{
 430        fuse_set_nowrite(inode);
 431        fuse_release_nowrite(inode);
 432}
 433
 434static int fuse_flush(struct file *file, fl_owner_t id)
 435{
 436        struct inode *inode = file_inode(file);
 437        struct fuse_conn *fc = get_fuse_conn(inode);
 438        struct fuse_file *ff = file->private_data;
 439        struct fuse_req *req;
 440        struct fuse_flush_in inarg;
 441        int err;
 442
 443        if (is_bad_inode(inode))
 444                return -EIO;
 445
 446        if (fc->no_flush)
 447                return 0;
 448
 449        err = write_inode_now(inode, 1);
 450        if (err)
 451                return err;
 452
 453        mutex_lock(&inode->i_mutex);
 454        fuse_sync_writes(inode);
 455        mutex_unlock(&inode->i_mutex);
 456
 457        req = fuse_get_req_nofail_nopages(fc, file);
 458        memset(&inarg, 0, sizeof(inarg));
 459        inarg.fh = ff->fh;
 460        inarg.lock_owner = fuse_lock_owner_id(fc, id);
 461        req->in.h.opcode = FUSE_FLUSH;
 462        req->in.h.nodeid = get_node_id(inode);
 463        req->in.numargs = 1;
 464        req->in.args[0].size = sizeof(inarg);
 465        req->in.args[0].value = &inarg;
 466        req->force = 1;
 467        fuse_request_send(fc, req);
 468        err = req->out.h.error;
 469        fuse_put_request(fc, req);
 470        if (err == -ENOSYS) {
 471                fc->no_flush = 1;
 472                err = 0;
 473        }
 474        return err;
 475}
 476
 477int fuse_fsync_common(struct file *file, loff_t start, loff_t end,
 478                      int datasync, int isdir)
 479{
 480        struct inode *inode = file->f_mapping->host;
 481        struct fuse_conn *fc = get_fuse_conn(inode);
 482        struct fuse_file *ff = file->private_data;
 483        struct fuse_req *req;
 484        struct fuse_fsync_in inarg;
 485        int err;
 486
 487        if (is_bad_inode(inode))
 488                return -EIO;
 489
 490        mutex_lock(&inode->i_mutex);
 491
 492        /*
 493         * Start writeback against all dirty pages of the inode, then
 494         * wait for all outstanding writes, before sending the FSYNC
 495         * request.
 496         */
 497        err = filemap_write_and_wait_range(inode->i_mapping, start, end);
 498        if (err)
 499                goto out;
 500
 501        fuse_sync_writes(inode);
 502        err = sync_inode_metadata(inode, 1);
 503        if (err)
 504                goto out;
 505
 506        if ((!isdir && fc->no_fsync) || (isdir && fc->no_fsyncdir))
 507                goto out;
 508
 509        req = fuse_get_req_nopages(fc);
 510        if (IS_ERR(req)) {
 511                err = PTR_ERR(req);
 512                goto out;
 513        }
 514
 515        memset(&inarg, 0, sizeof(inarg));
 516        inarg.fh = ff->fh;
 517        inarg.fsync_flags = datasync ? 1 : 0;
 518        req->in.h.opcode = isdir ? FUSE_FSYNCDIR : FUSE_FSYNC;
 519        req->in.h.nodeid = get_node_id(inode);
 520        req->in.numargs = 1;
 521        req->in.args[0].size = sizeof(inarg);
 522        req->in.args[0].value = &inarg;
 523        fuse_request_send(fc, req);
 524        err = req->out.h.error;
 525        fuse_put_request(fc, req);
 526        if (err == -ENOSYS) {
 527                if (isdir)
 528                        fc->no_fsyncdir = 1;
 529                else
 530                        fc->no_fsync = 1;
 531                err = 0;
 532        }
 533out:
 534        mutex_unlock(&inode->i_mutex);
 535        return err;
 536}
 537
 538static int fuse_fsync(struct file *file, loff_t start, loff_t end,
 539                      int datasync)
 540{
 541        return fuse_fsync_common(file, start, end, datasync, 0);
 542}
 543
 544void fuse_read_fill(struct fuse_req *req, struct file *file, loff_t pos,
 545                    size_t count, int opcode)
 546{
 547        struct fuse_read_in *inarg = &req->misc.read.in;
 548        struct fuse_file *ff = file->private_data;
 549
 550        inarg->fh = ff->fh;
 551        inarg->offset = pos;
 552        inarg->size = count;
 553        inarg->flags = file->f_flags;
 554        req->in.h.opcode = opcode;
 555        req->in.h.nodeid = ff->nodeid;
 556        req->in.numargs = 1;
 557        req->in.args[0].size = sizeof(struct fuse_read_in);
 558        req->in.args[0].value = inarg;
 559        req->out.argvar = 1;
 560        req->out.numargs = 1;
 561        req->out.args[0].size = count;
 562}
 563
 564static void fuse_release_user_pages(struct fuse_req *req, int write)
 565{
 566        unsigned i;
 567
 568        for (i = 0; i < req->num_pages; i++) {
 569                struct page *page = req->pages[i];
 570                if (write)
 571                        set_page_dirty_lock(page);
 572                put_page(page);
 573        }
 574}
 575
 576/**
 577 * In case of short read, the caller sets 'pos' to the position of
 578 * actual end of fuse request in IO request. Otherwise, if bytes_requested
 579 * == bytes_transferred or rw == WRITE, the caller sets 'pos' to -1.
 580 *
 581 * An example:
 582 * User requested DIO read of 64K. It was splitted into two 32K fuse requests,
 583 * both submitted asynchronously. The first of them was ACKed by userspace as
 584 * fully completed (req->out.args[0].size == 32K) resulting in pos == -1. The
 585 * second request was ACKed as short, e.g. only 1K was read, resulting in
 586 * pos == 33K.
 587 *
 588 * Thus, when all fuse requests are completed, the minimal non-negative 'pos'
 589 * will be equal to the length of the longest contiguous fragment of
 590 * transferred data starting from the beginning of IO request.
 591 */
 592static void fuse_aio_complete(struct fuse_io_priv *io, int err, ssize_t pos)
 593{
 594        int left;
 595
 596        spin_lock(&io->lock);
 597        if (err)
 598                io->err = io->err ? : err;
 599        else if (pos >= 0 && (io->bytes < 0 || pos < io->bytes))
 600                io->bytes = pos;
 601
 602        left = --io->reqs;
 603        spin_unlock(&io->lock);
 604
 605        if (!left) {
 606                long res;
 607
 608                if (io->err)
 609                        res = io->err;
 610                else if (io->bytes >= 0 && io->write)
 611                        res = -EIO;
 612                else {
 613                        res = io->bytes < 0 ? io->size : io->bytes;
 614
 615                        if (!is_sync_kiocb(io->iocb)) {
 616                                struct inode *inode = file_inode(io->iocb->ki_filp);
 617                                struct fuse_conn *fc = get_fuse_conn(inode);
 618                                struct fuse_inode *fi = get_fuse_inode(inode);
 619
 620                                spin_lock(&fc->lock);
 621                                fi->attr_version = ++fc->attr_version;
 622                                spin_unlock(&fc->lock);
 623                        }
 624                }
 625
 626                aio_complete(io->iocb, res, 0);
 627                kfree(io);
 628        }
 629}
 630
 631static void fuse_aio_complete_req(struct fuse_conn *fc, struct fuse_req *req)
 632{
 633        struct fuse_io_priv *io = req->io;
 634        ssize_t pos = -1;
 635
 636        fuse_release_user_pages(req, !io->write);
 637
 638        if (io->write) {
 639                if (req->misc.write.in.size != req->misc.write.out.size)
 640                        pos = req->misc.write.in.offset - io->offset +
 641                                req->misc.write.out.size;
 642        } else {
 643                if (req->misc.read.in.size != req->out.args[0].size)
 644                        pos = req->misc.read.in.offset - io->offset +
 645                                req->out.args[0].size;
 646        }
 647
 648        fuse_aio_complete(io, req->out.h.error, pos);
 649}
 650
 651static size_t fuse_async_req_send(struct fuse_conn *fc, struct fuse_req *req,
 652                size_t num_bytes, struct fuse_io_priv *io)
 653{
 654        spin_lock(&io->lock);
 655        io->size += num_bytes;
 656        io->reqs++;
 657        spin_unlock(&io->lock);
 658
 659        req->io = io;
 660        req->end = fuse_aio_complete_req;
 661
 662        __fuse_get_request(req);
 663        fuse_request_send_background(fc, req);
 664
 665        return num_bytes;
 666}
 667
 668static size_t fuse_send_read(struct fuse_req *req, struct fuse_io_priv *io,
 669                             loff_t pos, size_t count, fl_owner_t owner)
 670{
 671        struct file *file = io->file;
 672        struct fuse_file *ff = file->private_data;
 673        struct fuse_conn *fc = ff->fc;
 674
 675        fuse_read_fill(req, file, pos, count, FUSE_READ);
 676        if (owner != NULL) {
 677                struct fuse_read_in *inarg = &req->misc.read.in;
 678
 679                inarg->read_flags |= FUSE_READ_LOCKOWNER;
 680                inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 681        }
 682
 683        if (io->async)
 684                return fuse_async_req_send(fc, req, count, io);
 685
 686        fuse_request_send(fc, req);
 687        return req->out.args[0].size;
 688}
 689
 690static void fuse_read_update_size(struct inode *inode, loff_t size,
 691                                  u64 attr_ver)
 692{
 693        struct fuse_conn *fc = get_fuse_conn(inode);
 694        struct fuse_inode *fi = get_fuse_inode(inode);
 695
 696        spin_lock(&fc->lock);
 697        if (attr_ver == fi->attr_version && size < inode->i_size &&
 698            !test_bit(FUSE_I_SIZE_UNSTABLE, &fi->state)) {
 699                fi->attr_version = ++fc->attr_version;
 700                i_size_write(inode, size);
 701        }
 702        spin_unlock(&fc->lock);
 703}
 704
 705static void fuse_short_read(struct fuse_req *req, struct inode *inode,
 706                            u64 attr_ver)
 707{
 708        size_t num_read = req->out.args[0].size;
 709        struct fuse_conn *fc = get_fuse_conn(inode);
 710
 711        if (fc->writeback_cache) {
 712                /*
 713                 * A hole in a file. Some data after the hole are in page cache,
 714                 * but have not reached the client fs yet. So, the hole is not
 715                 * present there.
 716                 */
 717                int i;
 718                int start_idx = num_read >> PAGE_CACHE_SHIFT;
 719                size_t off = num_read & (PAGE_CACHE_SIZE - 1);
 720
 721                for (i = start_idx; i < req->num_pages; i++) {
 722                        zero_user_segment(req->pages[i], off, PAGE_CACHE_SIZE);
 723                        off = 0;
 724                }
 725        } else {
 726                loff_t pos = page_offset(req->pages[0]) + num_read;
 727                fuse_read_update_size(inode, pos, attr_ver);
 728        }
 729}
 730
 731static int fuse_do_readpage(struct file *file, struct page *page)
 732{
 733        struct fuse_io_priv io = { .async = 0, .file = file };
 734        struct inode *inode = page->mapping->host;
 735        struct fuse_conn *fc = get_fuse_conn(inode);
 736        struct fuse_req *req;
 737        size_t num_read;
 738        loff_t pos = page_offset(page);
 739        size_t count = PAGE_CACHE_SIZE;
 740        u64 attr_ver;
 741        int err;
 742
 743        /*
 744         * Page writeback can extend beyond the lifetime of the
 745         * page-cache page, so make sure we read a properly synced
 746         * page.
 747         */
 748        fuse_wait_on_page_writeback(inode, page->index);
 749
 750        req = fuse_get_req(fc, 1);
 751        if (IS_ERR(req))
 752                return PTR_ERR(req);
 753
 754        attr_ver = fuse_get_attr_version(fc);
 755
 756        req->out.page_zeroing = 1;
 757        req->out.argpages = 1;
 758        req->num_pages = 1;
 759        req->pages[0] = page;
 760        req->page_descs[0].length = count;
 761        num_read = fuse_send_read(req, &io, pos, count, NULL);
 762        err = req->out.h.error;
 763
 764        if (!err) {
 765                /*
 766                 * Short read means EOF.  If file size is larger, truncate it
 767                 */
 768                if (num_read < count)
 769                        fuse_short_read(req, inode, attr_ver);
 770
 771                SetPageUptodate(page);
 772        }
 773
 774        fuse_put_request(fc, req);
 775
 776        return err;
 777}
 778
 779static int fuse_readpage(struct file *file, struct page *page)
 780{
 781        struct inode *inode = page->mapping->host;
 782        int err;
 783
 784        err = -EIO;
 785        if (is_bad_inode(inode))
 786                goto out;
 787
 788        err = fuse_do_readpage(file, page);
 789        fuse_invalidate_atime(inode);
 790 out:
 791        unlock_page(page);
 792        return err;
 793}
 794
 795static void fuse_readpages_end(struct fuse_conn *fc, struct fuse_req *req)
 796{
 797        int i;
 798        size_t count = req->misc.read.in.size;
 799        size_t num_read = req->out.args[0].size;
 800        struct address_space *mapping = NULL;
 801
 802        for (i = 0; mapping == NULL && i < req->num_pages; i++)
 803                mapping = req->pages[i]->mapping;
 804
 805        if (mapping) {
 806                struct inode *inode = mapping->host;
 807
 808                /*
 809                 * Short read means EOF. If file size is larger, truncate it
 810                 */
 811                if (!req->out.h.error && num_read < count)
 812                        fuse_short_read(req, inode, req->misc.read.attr_ver);
 813
 814                fuse_invalidate_atime(inode);
 815        }
 816
 817        for (i = 0; i < req->num_pages; i++) {
 818                struct page *page = req->pages[i];
 819                if (!req->out.h.error)
 820                        SetPageUptodate(page);
 821                else
 822                        SetPageError(page);
 823                unlock_page(page);
 824                page_cache_release(page);
 825        }
 826        if (req->ff)
 827                fuse_file_put(req->ff, false);
 828}
 829
 830static void fuse_send_readpages(struct fuse_req *req, struct file *file)
 831{
 832        struct fuse_file *ff = file->private_data;
 833        struct fuse_conn *fc = ff->fc;
 834        loff_t pos = page_offset(req->pages[0]);
 835        size_t count = req->num_pages << PAGE_CACHE_SHIFT;
 836
 837        req->out.argpages = 1;
 838        req->out.page_zeroing = 1;
 839        req->out.page_replace = 1;
 840        fuse_read_fill(req, file, pos, count, FUSE_READ);
 841        req->misc.read.attr_ver = fuse_get_attr_version(fc);
 842        if (fc->async_read) {
 843                req->ff = fuse_file_get(ff);
 844                req->end = fuse_readpages_end;
 845                fuse_request_send_background(fc, req);
 846        } else {
 847                fuse_request_send(fc, req);
 848                fuse_readpages_end(fc, req);
 849                fuse_put_request(fc, req);
 850        }
 851}
 852
 853struct fuse_fill_data {
 854        struct fuse_req *req;
 855        struct file *file;
 856        struct inode *inode;
 857        unsigned nr_pages;
 858};
 859
 860static int fuse_readpages_fill(void *_data, struct page *page)
 861{
 862        struct fuse_fill_data *data = _data;
 863        struct fuse_req *req = data->req;
 864        struct inode *inode = data->inode;
 865        struct fuse_conn *fc = get_fuse_conn(inode);
 866
 867        fuse_wait_on_page_writeback(inode, page->index);
 868
 869        if (req->num_pages &&
 870            (req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
 871             (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_read ||
 872             req->pages[req->num_pages - 1]->index + 1 != page->index)) {
 873                int nr_alloc = min_t(unsigned, data->nr_pages,
 874                                     FUSE_MAX_PAGES_PER_REQ);
 875                fuse_send_readpages(req, data->file);
 876                if (fc->async_read)
 877                        req = fuse_get_req_for_background(fc, nr_alloc);
 878                else
 879                        req = fuse_get_req(fc, nr_alloc);
 880
 881                data->req = req;
 882                if (IS_ERR(req)) {
 883                        unlock_page(page);
 884                        return PTR_ERR(req);
 885                }
 886        }
 887
 888        if (WARN_ON(req->num_pages >= req->max_pages)) {
 889                fuse_put_request(fc, req);
 890                return -EIO;
 891        }
 892
 893        page_cache_get(page);
 894        req->pages[req->num_pages] = page;
 895        req->page_descs[req->num_pages].length = PAGE_SIZE;
 896        req->num_pages++;
 897        data->nr_pages--;
 898        return 0;
 899}
 900
 901static int fuse_readpages(struct file *file, struct address_space *mapping,
 902                          struct list_head *pages, unsigned nr_pages)
 903{
 904        struct inode *inode = mapping->host;
 905        struct fuse_conn *fc = get_fuse_conn(inode);
 906        struct fuse_fill_data data;
 907        int err;
 908        int nr_alloc = min_t(unsigned, nr_pages, FUSE_MAX_PAGES_PER_REQ);
 909
 910        err = -EIO;
 911        if (is_bad_inode(inode))
 912                goto out;
 913
 914        data.file = file;
 915        data.inode = inode;
 916        if (fc->async_read)
 917                data.req = fuse_get_req_for_background(fc, nr_alloc);
 918        else
 919                data.req = fuse_get_req(fc, nr_alloc);
 920        data.nr_pages = nr_pages;
 921        err = PTR_ERR(data.req);
 922        if (IS_ERR(data.req))
 923                goto out;
 924
 925        err = read_cache_pages(mapping, pages, fuse_readpages_fill, &data);
 926        if (!err) {
 927                if (data.req->num_pages)
 928                        fuse_send_readpages(data.req, file);
 929                else
 930                        fuse_put_request(fc, data.req);
 931        }
 932out:
 933        return err;
 934}
 935
 936static ssize_t fuse_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 937{
 938        struct inode *inode = iocb->ki_filp->f_mapping->host;
 939        struct fuse_conn *fc = get_fuse_conn(inode);
 940
 941        /*
 942         * In auto invalidate mode, always update attributes on read.
 943         * Otherwise, only update if we attempt to read past EOF (to ensure
 944         * i_size is up to date).
 945         */
 946        if (fc->auto_inval_data ||
 947            (iocb->ki_pos + iov_iter_count(to) > i_size_read(inode))) {
 948                int err;
 949                err = fuse_update_attributes(inode, NULL, iocb->ki_filp, NULL);
 950                if (err)
 951                        return err;
 952        }
 953
 954        return generic_file_read_iter(iocb, to);
 955}
 956
 957static void fuse_write_fill(struct fuse_req *req, struct fuse_file *ff,
 958                            loff_t pos, size_t count)
 959{
 960        struct fuse_write_in *inarg = &req->misc.write.in;
 961        struct fuse_write_out *outarg = &req->misc.write.out;
 962
 963        inarg->fh = ff->fh;
 964        inarg->offset = pos;
 965        inarg->size = count;
 966        req->in.h.opcode = FUSE_WRITE;
 967        req->in.h.nodeid = ff->nodeid;
 968        req->in.numargs = 2;
 969        if (ff->fc->minor < 9)
 970                req->in.args[0].size = FUSE_COMPAT_WRITE_IN_SIZE;
 971        else
 972                req->in.args[0].size = sizeof(struct fuse_write_in);
 973        req->in.args[0].value = inarg;
 974        req->in.args[1].size = count;
 975        req->out.numargs = 1;
 976        req->out.args[0].size = sizeof(struct fuse_write_out);
 977        req->out.args[0].value = outarg;
 978}
 979
 980static size_t fuse_send_write(struct fuse_req *req, struct fuse_io_priv *io,
 981                              loff_t pos, size_t count, fl_owner_t owner)
 982{
 983        struct file *file = io->file;
 984        struct fuse_file *ff = file->private_data;
 985        struct fuse_conn *fc = ff->fc;
 986        struct fuse_write_in *inarg = &req->misc.write.in;
 987
 988        fuse_write_fill(req, ff, pos, count);
 989        inarg->flags = file->f_flags;
 990        if (owner != NULL) {
 991                inarg->write_flags |= FUSE_WRITE_LOCKOWNER;
 992                inarg->lock_owner = fuse_lock_owner_id(fc, owner);
 993        }
 994
 995        if (io->async)
 996                return fuse_async_req_send(fc, req, count, io);
 997
 998        fuse_request_send(fc, req);
 999        return req->misc.write.out.size;
1000}
1001
1002bool fuse_write_update_size(struct inode *inode, loff_t pos)
1003{
1004        struct fuse_conn *fc = get_fuse_conn(inode);
1005        struct fuse_inode *fi = get_fuse_inode(inode);
1006        bool ret = false;
1007
1008        spin_lock(&fc->lock);
1009        fi->attr_version = ++fc->attr_version;
1010        if (pos > inode->i_size) {
1011                i_size_write(inode, pos);
1012                ret = true;
1013        }
1014        spin_unlock(&fc->lock);
1015
1016        return ret;
1017}
1018
1019static size_t fuse_send_write_pages(struct fuse_req *req, struct file *file,
1020                                    struct inode *inode, loff_t pos,
1021                                    size_t count)
1022{
1023        size_t res;
1024        unsigned offset;
1025        unsigned i;
1026        struct fuse_io_priv io = { .async = 0, .file = file };
1027
1028        for (i = 0; i < req->num_pages; i++)
1029                fuse_wait_on_page_writeback(inode, req->pages[i]->index);
1030
1031        res = fuse_send_write(req, &io, pos, count, NULL);
1032
1033        offset = req->page_descs[0].offset;
1034        count = res;
1035        for (i = 0; i < req->num_pages; i++) {
1036                struct page *page = req->pages[i];
1037
1038                if (!req->out.h.error && !offset && count >= PAGE_CACHE_SIZE)
1039                        SetPageUptodate(page);
1040
1041                if (count > PAGE_CACHE_SIZE - offset)
1042                        count -= PAGE_CACHE_SIZE - offset;
1043                else
1044                        count = 0;
1045                offset = 0;
1046
1047                unlock_page(page);
1048                page_cache_release(page);
1049        }
1050
1051        return res;
1052}
1053
1054static ssize_t fuse_fill_write_pages(struct fuse_req *req,
1055                               struct address_space *mapping,
1056                               struct iov_iter *ii, loff_t pos)
1057{
1058        struct fuse_conn *fc = get_fuse_conn(mapping->host);
1059        unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1060        size_t count = 0;
1061        int err;
1062
1063        req->in.argpages = 1;
1064        req->page_descs[0].offset = offset;
1065
1066        do {
1067                size_t tmp;
1068                struct page *page;
1069                pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1070                size_t bytes = min_t(size_t, PAGE_CACHE_SIZE - offset,
1071                                     iov_iter_count(ii));
1072
1073                bytes = min_t(size_t, bytes, fc->max_write - count);
1074
1075 again:
1076                err = -EFAULT;
1077                if (iov_iter_fault_in_readable(ii, bytes))
1078                        break;
1079
1080                err = -ENOMEM;
1081                page = grab_cache_page_write_begin(mapping, index, 0);
1082                if (!page)
1083                        break;
1084
1085                if (mapping_writably_mapped(mapping))
1086                        flush_dcache_page(page);
1087
1088                tmp = iov_iter_copy_from_user_atomic(page, ii, offset, bytes);
1089                flush_dcache_page(page);
1090
1091                if (!tmp) {
1092                        unlock_page(page);
1093                        page_cache_release(page);
1094                        bytes = min(bytes, iov_iter_single_seg_count(ii));
1095                        goto again;
1096                }
1097
1098                err = 0;
1099                req->pages[req->num_pages] = page;
1100                req->page_descs[req->num_pages].length = tmp;
1101                req->num_pages++;
1102
1103                iov_iter_advance(ii, tmp);
1104                count += tmp;
1105                pos += tmp;
1106                offset += tmp;
1107                if (offset == PAGE_CACHE_SIZE)
1108                        offset = 0;
1109
1110                if (!fc->big_writes)
1111                        break;
1112        } while (iov_iter_count(ii) && count < fc->max_write &&
1113                 req->num_pages < req->max_pages && offset == 0);
1114
1115        return count > 0 ? count : err;
1116}
1117
1118static inline unsigned fuse_wr_pages(loff_t pos, size_t len)
1119{
1120        return min_t(unsigned,
1121                     ((pos + len - 1) >> PAGE_CACHE_SHIFT) -
1122                     (pos >> PAGE_CACHE_SHIFT) + 1,
1123                     FUSE_MAX_PAGES_PER_REQ);
1124}
1125
1126static ssize_t fuse_perform_write(struct file *file,
1127                                  struct address_space *mapping,
1128                                  struct iov_iter *ii, loff_t pos)
1129{
1130        struct inode *inode = mapping->host;
1131        struct fuse_conn *fc = get_fuse_conn(inode);
1132        struct fuse_inode *fi = get_fuse_inode(inode);
1133        int err = 0;
1134        ssize_t res = 0;
1135
1136        if (is_bad_inode(inode))
1137                return -EIO;
1138
1139        if (inode->i_size < pos + iov_iter_count(ii))
1140                set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1141
1142        do {
1143                struct fuse_req *req;
1144                ssize_t count;
1145                unsigned nr_pages = fuse_wr_pages(pos, iov_iter_count(ii));
1146
1147                req = fuse_get_req(fc, nr_pages);
1148                if (IS_ERR(req)) {
1149                        err = PTR_ERR(req);
1150                        break;
1151                }
1152
1153                count = fuse_fill_write_pages(req, mapping, ii, pos);
1154                if (count <= 0) {
1155                        err = count;
1156                } else {
1157                        size_t num_written;
1158
1159                        num_written = fuse_send_write_pages(req, file, inode,
1160                                                            pos, count);
1161                        err = req->out.h.error;
1162                        if (!err) {
1163                                res += num_written;
1164                                pos += num_written;
1165
1166                                /* break out of the loop on short write */
1167                                if (num_written != count)
1168                                        err = -EIO;
1169                        }
1170                }
1171                fuse_put_request(fc, req);
1172        } while (!err && iov_iter_count(ii));
1173
1174        if (res > 0)
1175                fuse_write_update_size(inode, pos);
1176
1177        clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
1178        fuse_invalidate_attr(inode);
1179
1180        return res > 0 ? res : err;
1181}
1182
1183static ssize_t fuse_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1184{
1185        struct file *file = iocb->ki_filp;
1186        struct address_space *mapping = file->f_mapping;
1187        size_t count = iov_iter_count(from);
1188        ssize_t written = 0;
1189        ssize_t written_buffered = 0;
1190        struct inode *inode = mapping->host;
1191        ssize_t err;
1192        loff_t endbyte = 0;
1193        loff_t pos = iocb->ki_pos;
1194
1195        if (get_fuse_conn(inode)->writeback_cache) {
1196                /* Update size (EOF optimization) and mode (SUID clearing) */
1197                err = fuse_update_attributes(mapping->host, NULL, file, NULL);
1198                if (err)
1199                        return err;
1200
1201                return generic_file_write_iter(iocb, from);
1202        }
1203
1204        mutex_lock(&inode->i_mutex);
1205
1206        /* We can write back this queue in page reclaim */
1207        current->backing_dev_info = mapping->backing_dev_info;
1208
1209        err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1210        if (err)
1211                goto out;
1212
1213        if (count == 0)
1214                goto out;
1215
1216        iov_iter_truncate(from, count);
1217        err = file_remove_suid(file);
1218        if (err)
1219                goto out;
1220
1221        err = file_update_time(file);
1222        if (err)
1223                goto out;
1224
1225        if (file->f_flags & O_DIRECT) {
1226                written = generic_file_direct_write(iocb, from, pos);
1227                if (written < 0 || !iov_iter_count(from))
1228                        goto out;
1229
1230                pos += written;
1231
1232                written_buffered = fuse_perform_write(file, mapping, from, pos);
1233                if (written_buffered < 0) {
1234                        err = written_buffered;
1235                        goto out;
1236                }
1237                endbyte = pos + written_buffered - 1;
1238
1239                err = filemap_write_and_wait_range(file->f_mapping, pos,
1240                                                   endbyte);
1241                if (err)
1242                        goto out;
1243
1244                invalidate_mapping_pages(file->f_mapping,
1245                                         pos >> PAGE_CACHE_SHIFT,
1246                                         endbyte >> PAGE_CACHE_SHIFT);
1247
1248                written += written_buffered;
1249                iocb->ki_pos = pos + written_buffered;
1250        } else {
1251                written = fuse_perform_write(file, mapping, from, pos);
1252                if (written >= 0)
1253                        iocb->ki_pos = pos + written;
1254        }
1255out:
1256        current->backing_dev_info = NULL;
1257        mutex_unlock(&inode->i_mutex);
1258
1259        return written ? written : err;
1260}
1261
1262static inline void fuse_page_descs_length_init(struct fuse_req *req,
1263                unsigned index, unsigned nr_pages)
1264{
1265        int i;
1266
1267        for (i = index; i < index + nr_pages; i++)
1268                req->page_descs[i].length = PAGE_SIZE -
1269                        req->page_descs[i].offset;
1270}
1271
1272static inline unsigned long fuse_get_user_addr(const struct iov_iter *ii)
1273{
1274        return (unsigned long)ii->iov->iov_base + ii->iov_offset;
1275}
1276
1277static inline size_t fuse_get_frag_size(const struct iov_iter *ii,
1278                                        size_t max_size)
1279{
1280        return min(iov_iter_single_seg_count(ii), max_size);
1281}
1282
1283static int fuse_get_user_pages(struct fuse_req *req, struct iov_iter *ii,
1284                               size_t *nbytesp, int write)
1285{
1286        size_t nbytes = 0;  /* # bytes already packed in req */
1287
1288        /* Special case for kernel I/O: can copy directly into the buffer */
1289        if (ii->type & ITER_KVEC) {
1290                unsigned long user_addr = fuse_get_user_addr(ii);
1291                size_t frag_size = fuse_get_frag_size(ii, *nbytesp);
1292
1293                if (write)
1294                        req->in.args[1].value = (void *) user_addr;
1295                else
1296                        req->out.args[0].value = (void *) user_addr;
1297
1298                iov_iter_advance(ii, frag_size);
1299                *nbytesp = frag_size;
1300                return 0;
1301        }
1302
1303        while (nbytes < *nbytesp && req->num_pages < req->max_pages) {
1304                unsigned npages;
1305                size_t start;
1306                unsigned n = req->max_pages - req->num_pages;
1307                ssize_t ret = iov_iter_get_pages(ii,
1308                                        &req->pages[req->num_pages],
1309                                        n * PAGE_SIZE, &start);
1310                if (ret < 0)
1311                        return ret;
1312
1313                iov_iter_advance(ii, ret);
1314                nbytes += ret;
1315
1316                ret += start;
1317                npages = (ret + PAGE_SIZE - 1) / PAGE_SIZE;
1318
1319                req->page_descs[req->num_pages].offset = start;
1320                fuse_page_descs_length_init(req, req->num_pages, npages);
1321
1322                req->num_pages += npages;
1323                req->page_descs[req->num_pages - 1].length -=
1324                        (PAGE_SIZE - ret) & (PAGE_SIZE - 1);
1325        }
1326
1327        if (write)
1328                req->in.argpages = 1;
1329        else
1330                req->out.argpages = 1;
1331
1332        *nbytesp = nbytes;
1333
1334        return 0;
1335}
1336
1337static inline int fuse_iter_npages(const struct iov_iter *ii_p)
1338{
1339        return iov_iter_npages(ii_p, FUSE_MAX_PAGES_PER_REQ);
1340}
1341
1342ssize_t fuse_direct_io(struct fuse_io_priv *io, struct iov_iter *iter,
1343                       loff_t *ppos, int flags)
1344{
1345        int write = flags & FUSE_DIO_WRITE;
1346        int cuse = flags & FUSE_DIO_CUSE;
1347        struct file *file = io->file;
1348        struct inode *inode = file->f_mapping->host;
1349        struct fuse_file *ff = file->private_data;
1350        struct fuse_conn *fc = ff->fc;
1351        size_t nmax = write ? fc->max_write : fc->max_read;
1352        loff_t pos = *ppos;
1353        size_t count = iov_iter_count(iter);
1354        pgoff_t idx_from = pos >> PAGE_CACHE_SHIFT;
1355        pgoff_t idx_to = (pos + count - 1) >> PAGE_CACHE_SHIFT;
1356        ssize_t res = 0;
1357        struct fuse_req *req;
1358
1359        if (io->async)
1360                req = fuse_get_req_for_background(fc, fuse_iter_npages(iter));
1361        else
1362                req = fuse_get_req(fc, fuse_iter_npages(iter));
1363        if (IS_ERR(req))
1364                return PTR_ERR(req);
1365
1366        if (!cuse && fuse_range_is_writeback(inode, idx_from, idx_to)) {
1367                if (!write)
1368                        mutex_lock(&inode->i_mutex);
1369                fuse_sync_writes(inode);
1370                if (!write)
1371                        mutex_unlock(&inode->i_mutex);
1372        }
1373
1374        while (count) {
1375                size_t nres;
1376                fl_owner_t owner = current->files;
1377                size_t nbytes = min(count, nmax);
1378                int err = fuse_get_user_pages(req, iter, &nbytes, write);
1379                if (err) {
1380                        res = err;
1381                        break;
1382                }
1383
1384                if (write)
1385                        nres = fuse_send_write(req, io, pos, nbytes, owner);
1386                else
1387                        nres = fuse_send_read(req, io, pos, nbytes, owner);
1388
1389                if (!io->async)
1390                        fuse_release_user_pages(req, !write);
1391                if (req->out.h.error) {
1392                        if (!res)
1393                                res = req->out.h.error;
1394                        break;
1395                } else if (nres > nbytes) {
1396                        res = -EIO;
1397                        break;
1398                }
1399                count -= nres;
1400                res += nres;
1401                pos += nres;
1402                if (nres != nbytes)
1403                        break;
1404                if (count) {
1405                        fuse_put_request(fc, req);
1406                        if (io->async)
1407                                req = fuse_get_req_for_background(fc,
1408                                        fuse_iter_npages(iter));
1409                        else
1410                                req = fuse_get_req(fc, fuse_iter_npages(iter));
1411                        if (IS_ERR(req))
1412                                break;
1413                }
1414        }
1415        if (!IS_ERR(req))
1416                fuse_put_request(fc, req);
1417        if (res > 0)
1418                *ppos = pos;
1419
1420        return res;
1421}
1422EXPORT_SYMBOL_GPL(fuse_direct_io);
1423
1424static ssize_t __fuse_direct_read(struct fuse_io_priv *io,
1425                                  struct iov_iter *iter,
1426                                  loff_t *ppos)
1427{
1428        ssize_t res;
1429        struct file *file = io->file;
1430        struct inode *inode = file_inode(file);
1431
1432        if (is_bad_inode(inode))
1433                return -EIO;
1434
1435        res = fuse_direct_io(io, iter, ppos, 0);
1436
1437        fuse_invalidate_attr(inode);
1438
1439        return res;
1440}
1441
1442static ssize_t fuse_direct_read(struct file *file, char __user *buf,
1443                                     size_t count, loff_t *ppos)
1444{
1445        struct fuse_io_priv io = { .async = 0, .file = file };
1446        struct iovec iov = { .iov_base = buf, .iov_len = count };
1447        struct iov_iter ii;
1448        iov_iter_init(&ii, READ, &iov, 1, count);
1449        return __fuse_direct_read(&io, &ii, ppos);
1450}
1451
1452static ssize_t __fuse_direct_write(struct fuse_io_priv *io,
1453                                   struct iov_iter *iter,
1454                                   loff_t *ppos)
1455{
1456        struct file *file = io->file;
1457        struct inode *inode = file_inode(file);
1458        size_t count = iov_iter_count(iter);
1459        ssize_t res;
1460
1461
1462        res = generic_write_checks(file, ppos, &count, 0);
1463        if (!res) {
1464                iov_iter_truncate(iter, count);
1465                res = fuse_direct_io(io, iter, ppos, FUSE_DIO_WRITE);
1466        }
1467
1468        fuse_invalidate_attr(inode);
1469
1470        return res;
1471}
1472
1473static ssize_t fuse_direct_write(struct file *file, const char __user *buf,
1474                                 size_t count, loff_t *ppos)
1475{
1476        struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = count };
1477        struct inode *inode = file_inode(file);
1478        ssize_t res;
1479        struct fuse_io_priv io = { .async = 0, .file = file };
1480        struct iov_iter ii;
1481        iov_iter_init(&ii, WRITE, &iov, 1, count);
1482
1483        if (is_bad_inode(inode))
1484                return -EIO;
1485
1486        /* Don't allow parallel writes to the same file */
1487        mutex_lock(&inode->i_mutex);
1488        res = __fuse_direct_write(&io, &ii, ppos);
1489        if (res > 0)
1490                fuse_write_update_size(inode, *ppos);
1491        mutex_unlock(&inode->i_mutex);
1492
1493        return res;
1494}
1495
1496static void fuse_writepage_free(struct fuse_conn *fc, struct fuse_req *req)
1497{
1498        int i;
1499
1500        for (i = 0; i < req->num_pages; i++)
1501                __free_page(req->pages[i]);
1502
1503        if (req->ff)
1504                fuse_file_put(req->ff, false);
1505}
1506
1507static void fuse_writepage_finish(struct fuse_conn *fc, struct fuse_req *req)
1508{
1509        struct inode *inode = req->inode;
1510        struct fuse_inode *fi = get_fuse_inode(inode);
1511        struct backing_dev_info *bdi = inode->i_mapping->backing_dev_info;
1512        int i;
1513
1514        list_del(&req->writepages_entry);
1515        for (i = 0; i < req->num_pages; i++) {
1516                dec_bdi_stat(bdi, BDI_WRITEBACK);
1517                dec_zone_page_state(req->pages[i], NR_WRITEBACK_TEMP);
1518                bdi_writeout_inc(bdi);
1519        }
1520        wake_up(&fi->page_waitq);
1521}
1522
1523/* Called under fc->lock, may release and reacquire it */
1524static void fuse_send_writepage(struct fuse_conn *fc, struct fuse_req *req,
1525                                loff_t size)
1526__releases(fc->lock)
1527__acquires(fc->lock)
1528{
1529        struct fuse_inode *fi = get_fuse_inode(req->inode);
1530        struct fuse_write_in *inarg = &req->misc.write.in;
1531        __u64 data_size = req->num_pages * PAGE_CACHE_SIZE;
1532
1533        if (!fc->connected)
1534                goto out_free;
1535
1536        if (inarg->offset + data_size <= size) {
1537                inarg->size = data_size;
1538        } else if (inarg->offset < size) {
1539                inarg->size = size - inarg->offset;
1540        } else {
1541                /* Got truncated off completely */
1542                goto out_free;
1543        }
1544
1545        req->in.args[1].size = inarg->size;
1546        fi->writectr++;
1547        fuse_request_send_background_locked(fc, req);
1548        return;
1549
1550 out_free:
1551        fuse_writepage_finish(fc, req);
1552        spin_unlock(&fc->lock);
1553        fuse_writepage_free(fc, req);
1554        fuse_put_request(fc, req);
1555        spin_lock(&fc->lock);
1556}
1557
1558/*
1559 * If fi->writectr is positive (no truncate or fsync going on) send
1560 * all queued writepage requests.
1561 *
1562 * Called with fc->lock
1563 */
1564void fuse_flush_writepages(struct inode *inode)
1565__releases(fc->lock)
1566__acquires(fc->lock)
1567{
1568        struct fuse_conn *fc = get_fuse_conn(inode);
1569        struct fuse_inode *fi = get_fuse_inode(inode);
1570        size_t crop = i_size_read(inode);
1571        struct fuse_req *req;
1572
1573        while (fi->writectr >= 0 && !list_empty(&fi->queued_writes)) {
1574                req = list_entry(fi->queued_writes.next, struct fuse_req, list);
1575                list_del_init(&req->list);
1576                fuse_send_writepage(fc, req, crop);
1577        }
1578}
1579
1580static void fuse_writepage_end(struct fuse_conn *fc, struct fuse_req *req)
1581{
1582        struct inode *inode = req->inode;
1583        struct fuse_inode *fi = get_fuse_inode(inode);
1584
1585        mapping_set_error(inode->i_mapping, req->out.h.error);
1586        spin_lock(&fc->lock);
1587        while (req->misc.write.next) {
1588                struct fuse_conn *fc = get_fuse_conn(inode);
1589                struct fuse_write_in *inarg = &req->misc.write.in;
1590                struct fuse_req *next = req->misc.write.next;
1591                req->misc.write.next = next->misc.write.next;
1592                next->misc.write.next = NULL;
1593                next->ff = fuse_file_get(req->ff);
1594                list_add(&next->writepages_entry, &fi->writepages);
1595
1596                /*
1597                 * Skip fuse_flush_writepages() to make it easy to crop requests
1598                 * based on primary request size.
1599                 *
1600                 * 1st case (trivial): there are no concurrent activities using
1601                 * fuse_set/release_nowrite.  Then we're on safe side because
1602                 * fuse_flush_writepages() would call fuse_send_writepage()
1603                 * anyway.
1604                 *
1605                 * 2nd case: someone called fuse_set_nowrite and it is waiting
1606                 * now for completion of all in-flight requests.  This happens
1607                 * rarely and no more than once per page, so this should be
1608                 * okay.
1609                 *
1610                 * 3rd case: someone (e.g. fuse_do_setattr()) is in the middle
1611                 * of fuse_set_nowrite..fuse_release_nowrite section.  The fact
1612                 * that fuse_set_nowrite returned implies that all in-flight
1613                 * requests were completed along with all of their secondary
1614                 * requests.  Further primary requests are blocked by negative
1615                 * writectr.  Hence there cannot be any in-flight requests and
1616                 * no invocations of fuse_writepage_end() while we're in
1617                 * fuse_set_nowrite..fuse_release_nowrite section.
1618                 */
1619                fuse_send_writepage(fc, next, inarg->offset + inarg->size);
1620        }
1621        fi->writectr--;
1622        fuse_writepage_finish(fc, req);
1623        spin_unlock(&fc->lock);
1624        fuse_writepage_free(fc, req);
1625}
1626
1627static struct fuse_file *__fuse_write_file_get(struct fuse_conn *fc,
1628                                               struct fuse_inode *fi)
1629{
1630        struct fuse_file *ff = NULL;
1631
1632        spin_lock(&fc->lock);
1633        if (!list_empty(&fi->write_files)) {
1634                ff = list_entry(fi->write_files.next, struct fuse_file,
1635                                write_entry);
1636                fuse_file_get(ff);
1637        }
1638        spin_unlock(&fc->lock);
1639
1640        return ff;
1641}
1642
1643static struct fuse_file *fuse_write_file_get(struct fuse_conn *fc,
1644                                             struct fuse_inode *fi)
1645{
1646        struct fuse_file *ff = __fuse_write_file_get(fc, fi);
1647        WARN_ON(!ff);
1648        return ff;
1649}
1650
1651int fuse_write_inode(struct inode *inode, struct writeback_control *wbc)
1652{
1653        struct fuse_conn *fc = get_fuse_conn(inode);
1654        struct fuse_inode *fi = get_fuse_inode(inode);
1655        struct fuse_file *ff;
1656        int err;
1657
1658        ff = __fuse_write_file_get(fc, fi);
1659        err = fuse_flush_times(inode, ff);
1660        if (ff)
1661                fuse_file_put(ff, 0);
1662
1663        return err;
1664}
1665
1666static int fuse_writepage_locked(struct page *page)
1667{
1668        struct address_space *mapping = page->mapping;
1669        struct inode *inode = mapping->host;
1670        struct fuse_conn *fc = get_fuse_conn(inode);
1671        struct fuse_inode *fi = get_fuse_inode(inode);
1672        struct fuse_req *req;
1673        struct page *tmp_page;
1674        int error = -ENOMEM;
1675
1676        set_page_writeback(page);
1677
1678        req = fuse_request_alloc_nofs(1);
1679        if (!req)
1680                goto err;
1681
1682        req->background = 1; /* writeback always goes to bg_queue */
1683        tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1684        if (!tmp_page)
1685                goto err_free;
1686
1687        error = -EIO;
1688        req->ff = fuse_write_file_get(fc, fi);
1689        if (!req->ff)
1690                goto err_nofile;
1691
1692        fuse_write_fill(req, req->ff, page_offset(page), 0);
1693
1694        copy_highpage(tmp_page, page);
1695        req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1696        req->misc.write.next = NULL;
1697        req->in.argpages = 1;
1698        req->num_pages = 1;
1699        req->pages[0] = tmp_page;
1700        req->page_descs[0].offset = 0;
1701        req->page_descs[0].length = PAGE_SIZE;
1702        req->end = fuse_writepage_end;
1703        req->inode = inode;
1704
1705        inc_bdi_stat(mapping->backing_dev_info, BDI_WRITEBACK);
1706        inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1707
1708        spin_lock(&fc->lock);
1709        list_add(&req->writepages_entry, &fi->writepages);
1710        list_add_tail(&req->list, &fi->queued_writes);
1711        fuse_flush_writepages(inode);
1712        spin_unlock(&fc->lock);
1713
1714        end_page_writeback(page);
1715
1716        return 0;
1717
1718err_nofile:
1719        __free_page(tmp_page);
1720err_free:
1721        fuse_request_free(req);
1722err:
1723        end_page_writeback(page);
1724        return error;
1725}
1726
1727static int fuse_writepage(struct page *page, struct writeback_control *wbc)
1728{
1729        int err;
1730
1731        if (fuse_page_is_writeback(page->mapping->host, page->index)) {
1732                /*
1733                 * ->writepages() should be called for sync() and friends.  We
1734                 * should only get here on direct reclaim and then we are
1735                 * allowed to skip a page which is already in flight
1736                 */
1737                WARN_ON(wbc->sync_mode == WB_SYNC_ALL);
1738
1739                redirty_page_for_writepage(wbc, page);
1740                return 0;
1741        }
1742
1743        err = fuse_writepage_locked(page);
1744        unlock_page(page);
1745
1746        return err;
1747}
1748
1749struct fuse_fill_wb_data {
1750        struct fuse_req *req;
1751        struct fuse_file *ff;
1752        struct inode *inode;
1753        struct page **orig_pages;
1754};
1755
1756static void fuse_writepages_send(struct fuse_fill_wb_data *data)
1757{
1758        struct fuse_req *req = data->req;
1759        struct inode *inode = data->inode;
1760        struct fuse_conn *fc = get_fuse_conn(inode);
1761        struct fuse_inode *fi = get_fuse_inode(inode);
1762        int num_pages = req->num_pages;
1763        int i;
1764
1765        req->ff = fuse_file_get(data->ff);
1766        spin_lock(&fc->lock);
1767        list_add_tail(&req->list, &fi->queued_writes);
1768        fuse_flush_writepages(inode);
1769        spin_unlock(&fc->lock);
1770
1771        for (i = 0; i < num_pages; i++)
1772                end_page_writeback(data->orig_pages[i]);
1773}
1774
1775static bool fuse_writepage_in_flight(struct fuse_req *new_req,
1776                                     struct page *page)
1777{
1778        struct fuse_conn *fc = get_fuse_conn(new_req->inode);
1779        struct fuse_inode *fi = get_fuse_inode(new_req->inode);
1780        struct fuse_req *tmp;
1781        struct fuse_req *old_req;
1782        bool found = false;
1783        pgoff_t curr_index;
1784
1785        BUG_ON(new_req->num_pages != 0);
1786
1787        spin_lock(&fc->lock);
1788        list_del(&new_req->writepages_entry);
1789        list_for_each_entry(old_req, &fi->writepages, writepages_entry) {
1790                BUG_ON(old_req->inode != new_req->inode);
1791                curr_index = old_req->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1792                if (curr_index <= page->index &&
1793                    page->index < curr_index + old_req->num_pages) {
1794                        found = true;
1795                        break;
1796                }
1797        }
1798        if (!found) {
1799                list_add(&new_req->writepages_entry, &fi->writepages);
1800                goto out_unlock;
1801        }
1802
1803        new_req->num_pages = 1;
1804        for (tmp = old_req; tmp != NULL; tmp = tmp->misc.write.next) {
1805                BUG_ON(tmp->inode != new_req->inode);
1806                curr_index = tmp->misc.write.in.offset >> PAGE_CACHE_SHIFT;
1807                if (tmp->num_pages == 1 &&
1808                    curr_index == page->index) {
1809                        old_req = tmp;
1810                }
1811        }
1812
1813        if (old_req->num_pages == 1 && (old_req->state == FUSE_REQ_INIT ||
1814                                        old_req->state == FUSE_REQ_PENDING)) {
1815                struct backing_dev_info *bdi = page->mapping->backing_dev_info;
1816
1817                copy_highpage(old_req->pages[0], page);
1818                spin_unlock(&fc->lock);
1819
1820                dec_bdi_stat(bdi, BDI_WRITEBACK);
1821                dec_zone_page_state(page, NR_WRITEBACK_TEMP);
1822                bdi_writeout_inc(bdi);
1823                fuse_writepage_free(fc, new_req);
1824                fuse_request_free(new_req);
1825                goto out;
1826        } else {
1827                new_req->misc.write.next = old_req->misc.write.next;
1828                old_req->misc.write.next = new_req;
1829        }
1830out_unlock:
1831        spin_unlock(&fc->lock);
1832out:
1833        return found;
1834}
1835
1836static int fuse_writepages_fill(struct page *page,
1837                struct writeback_control *wbc, void *_data)
1838{
1839        struct fuse_fill_wb_data *data = _data;
1840        struct fuse_req *req = data->req;
1841        struct inode *inode = data->inode;
1842        struct fuse_conn *fc = get_fuse_conn(inode);
1843        struct page *tmp_page;
1844        bool is_writeback;
1845        int err;
1846
1847        if (!data->ff) {
1848                err = -EIO;
1849                data->ff = fuse_write_file_get(fc, get_fuse_inode(inode));
1850                if (!data->ff)
1851                        goto out_unlock;
1852        }
1853
1854        /*
1855         * Being under writeback is unlikely but possible.  For example direct
1856         * read to an mmaped fuse file will set the page dirty twice; once when
1857         * the pages are faulted with get_user_pages(), and then after the read
1858         * completed.
1859         */
1860        is_writeback = fuse_page_is_writeback(inode, page->index);
1861
1862        if (req && req->num_pages &&
1863            (is_writeback || req->num_pages == FUSE_MAX_PAGES_PER_REQ ||
1864             (req->num_pages + 1) * PAGE_CACHE_SIZE > fc->max_write ||
1865             data->orig_pages[req->num_pages - 1]->index + 1 != page->index)) {
1866                fuse_writepages_send(data);
1867                data->req = NULL;
1868        }
1869        err = -ENOMEM;
1870        tmp_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM);
1871        if (!tmp_page)
1872                goto out_unlock;
1873
1874        /*
1875         * The page must not be redirtied until the writeout is completed
1876         * (i.e. userspace has sent a reply to the write request).  Otherwise
1877         * there could be more than one temporary page instance for each real
1878         * page.
1879         *
1880         * This is ensured by holding the page lock in page_mkwrite() while
1881         * checking fuse_page_is_writeback().  We already hold the page lock
1882         * since clear_page_dirty_for_io() and keep it held until we add the
1883         * request to the fi->writepages list and increment req->num_pages.
1884         * After this fuse_page_is_writeback() will indicate that the page is
1885         * under writeback, so we can release the page lock.
1886         */
1887        if (data->req == NULL) {
1888                struct fuse_inode *fi = get_fuse_inode(inode);
1889
1890                err = -ENOMEM;
1891                req = fuse_request_alloc_nofs(FUSE_MAX_PAGES_PER_REQ);
1892                if (!req) {
1893                        __free_page(tmp_page);
1894                        goto out_unlock;
1895                }
1896
1897                fuse_write_fill(req, data->ff, page_offset(page), 0);
1898                req->misc.write.in.write_flags |= FUSE_WRITE_CACHE;
1899                req->misc.write.next = NULL;
1900                req->in.argpages = 1;
1901                req->background = 1;
1902                req->num_pages = 0;
1903                req->end = fuse_writepage_end;
1904                req->inode = inode;
1905
1906                spin_lock(&fc->lock);
1907                list_add(&req->writepages_entry, &fi->writepages);
1908                spin_unlock(&fc->lock);
1909
1910                data->req = req;
1911        }
1912        set_page_writeback(page);
1913
1914        copy_highpage(tmp_page, page);
1915        req->pages[req->num_pages] = tmp_page;
1916        req->page_descs[req->num_pages].offset = 0;
1917        req->page_descs[req->num_pages].length = PAGE_SIZE;
1918
1919        inc_bdi_stat(page->mapping->backing_dev_info, BDI_WRITEBACK);
1920        inc_zone_page_state(tmp_page, NR_WRITEBACK_TEMP);
1921
1922        err = 0;
1923        if (is_writeback && fuse_writepage_in_flight(req, page)) {
1924                end_page_writeback(page);
1925                data->req = NULL;
1926                goto out_unlock;
1927        }
1928        data->orig_pages[req->num_pages] = page;
1929
1930        /*
1931         * Protected by fc->lock against concurrent access by
1932         * fuse_page_is_writeback().
1933         */
1934        spin_lock(&fc->lock);
1935        req->num_pages++;
1936        spin_unlock(&fc->lock);
1937
1938out_unlock:
1939        unlock_page(page);
1940
1941        return err;
1942}
1943
1944static int fuse_writepages(struct address_space *mapping,
1945                           struct writeback_control *wbc)
1946{
1947        struct inode *inode = mapping->host;
1948        struct fuse_fill_wb_data data;
1949        int err;
1950
1951        err = -EIO;
1952        if (is_bad_inode(inode))
1953                goto out;
1954
1955        data.inode = inode;
1956        data.req = NULL;
1957        data.ff = NULL;
1958
1959        err = -ENOMEM;
1960        data.orig_pages = kcalloc(FUSE_MAX_PAGES_PER_REQ,
1961                                  sizeof(struct page *),
1962                                  GFP_NOFS);
1963        if (!data.orig_pages)
1964                goto out;
1965
1966        err = write_cache_pages(mapping, wbc, fuse_writepages_fill, &data);
1967        if (data.req) {
1968                /* Ignore errors if we can write at least one page */
1969                BUG_ON(!data.req->num_pages);
1970                fuse_writepages_send(&data);
1971                err = 0;
1972        }
1973        if (data.ff)
1974                fuse_file_put(data.ff, false);
1975
1976        kfree(data.orig_pages);
1977out:
1978        return err;
1979}
1980
1981/*
1982 * It's worthy to make sure that space is reserved on disk for the write,
1983 * but how to implement it without killing performance need more thinking.
1984 */
1985static int fuse_write_begin(struct file *file, struct address_space *mapping,
1986                loff_t pos, unsigned len, unsigned flags,
1987                struct page **pagep, void **fsdata)
1988{
1989        pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1990        struct fuse_conn *fc = get_fuse_conn(file->f_dentry->d_inode);
1991        struct page *page;
1992        loff_t fsize;
1993        int err = -ENOMEM;
1994
1995        WARN_ON(!fc->writeback_cache);
1996
1997        page = grab_cache_page_write_begin(mapping, index, flags);
1998        if (!page)
1999                goto error;
2000
2001        fuse_wait_on_page_writeback(mapping->host, page->index);
2002
2003        if (PageUptodate(page) || len == PAGE_CACHE_SIZE)
2004                goto success;
2005        /*
2006         * Check if the start this page comes after the end of file, in which
2007         * case the readpage can be optimized away.
2008         */
2009        fsize = i_size_read(mapping->host);
2010        if (fsize <= (pos & PAGE_CACHE_MASK)) {
2011                size_t off = pos & ~PAGE_CACHE_MASK;
2012                if (off)
2013                        zero_user_segment(page, 0, off);
2014                goto success;
2015        }
2016        err = fuse_do_readpage(file, page);
2017        if (err)
2018                goto cleanup;
2019success:
2020        *pagep = page;
2021        return 0;
2022
2023cleanup:
2024        unlock_page(page);
2025        page_cache_release(page);
2026error:
2027        return err;
2028}
2029
2030static int fuse_write_end(struct file *file, struct address_space *mapping,
2031                loff_t pos, unsigned len, unsigned copied,
2032                struct page *page, void *fsdata)
2033{
2034        struct inode *inode = page->mapping->host;
2035
2036        if (!PageUptodate(page)) {
2037                /* Zero any unwritten bytes at the end of the page */
2038                size_t endoff = (pos + copied) & ~PAGE_CACHE_MASK;
2039                if (endoff)
2040                        zero_user_segment(page, endoff, PAGE_CACHE_SIZE);
2041                SetPageUptodate(page);
2042        }
2043
2044        fuse_write_update_size(inode, pos + copied);
2045        set_page_dirty(page);
2046        unlock_page(page);
2047        page_cache_release(page);
2048
2049        return copied;
2050}
2051
2052static int fuse_launder_page(struct page *page)
2053{
2054        int err = 0;
2055        if (clear_page_dirty_for_io(page)) {
2056                struct inode *inode = page->mapping->host;
2057                err = fuse_writepage_locked(page);
2058                if (!err)
2059                        fuse_wait_on_page_writeback(inode, page->index);
2060        }
2061        return err;
2062}
2063
2064/*
2065 * Write back dirty pages now, because there may not be any suitable
2066 * open files later
2067 */
2068static void fuse_vma_close(struct vm_area_struct *vma)
2069{
2070        filemap_write_and_wait(vma->vm_file->f_mapping);
2071}
2072
2073/*
2074 * Wait for writeback against this page to complete before allowing it
2075 * to be marked dirty again, and hence written back again, possibly
2076 * before the previous writepage completed.
2077 *
2078 * Block here, instead of in ->writepage(), so that the userspace fs
2079 * can only block processes actually operating on the filesystem.
2080 *
2081 * Otherwise unprivileged userspace fs would be able to block
2082 * unrelated:
2083 *
2084 * - page migration
2085 * - sync(2)
2086 * - try_to_free_pages() with order > PAGE_ALLOC_COSTLY_ORDER
2087 */
2088static int fuse_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2089{
2090        struct page *page = vmf->page;
2091        struct inode *inode = file_inode(vma->vm_file);
2092
2093        file_update_time(vma->vm_file);
2094        lock_page(page);
2095        if (page->mapping != inode->i_mapping) {
2096                unlock_page(page);
2097                return VM_FAULT_NOPAGE;
2098        }
2099
2100        fuse_wait_on_page_writeback(inode, page->index);
2101        return VM_FAULT_LOCKED;
2102}
2103
2104static const struct vm_operations_struct fuse_file_vm_ops = {
2105        .close          = fuse_vma_close,
2106        .fault          = filemap_fault,
2107        .map_pages      = filemap_map_pages,
2108        .page_mkwrite   = fuse_page_mkwrite,
2109        .remap_pages    = generic_file_remap_pages,
2110};
2111
2112static int fuse_file_mmap(struct file *file, struct vm_area_struct *vma)
2113{
2114        if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2115                fuse_link_write_file(file);
2116
2117        file_accessed(file);
2118        vma->vm_ops = &fuse_file_vm_ops;
2119        return 0;
2120}
2121
2122static int fuse_direct_mmap(struct file *file, struct vm_area_struct *vma)
2123{
2124        /* Can't provide the coherency needed for MAP_SHARED */
2125        if (vma->vm_flags & VM_MAYSHARE)
2126                return -ENODEV;
2127
2128        invalidate_inode_pages2(file->f_mapping);
2129
2130        return generic_file_mmap(file, vma);
2131}
2132
2133static int convert_fuse_file_lock(const struct fuse_file_lock *ffl,
2134                                  struct file_lock *fl)
2135{
2136        switch (ffl->type) {
2137        case F_UNLCK:
2138                break;
2139
2140        case F_RDLCK:
2141        case F_WRLCK:
2142                if (ffl->start > OFFSET_MAX || ffl->end > OFFSET_MAX ||
2143                    ffl->end < ffl->start)
2144                        return -EIO;
2145
2146                fl->fl_start = ffl->start;
2147                fl->fl_end = ffl->end;
2148                fl->fl_pid = ffl->pid;
2149                break;
2150
2151        default:
2152                return -EIO;
2153        }
2154        fl->fl_type = ffl->type;
2155        return 0;
2156}
2157
2158static void fuse_lk_fill(struct fuse_req *req, struct file *file,
2159                         const struct file_lock *fl, int opcode, pid_t pid,
2160                         int flock)
2161{
2162        struct inode *inode = file_inode(file);
2163        struct fuse_conn *fc = get_fuse_conn(inode);
2164        struct fuse_file *ff = file->private_data;
2165        struct fuse_lk_in *arg = &req->misc.lk_in;
2166
2167        arg->fh = ff->fh;
2168        arg->owner = fuse_lock_owner_id(fc, fl->fl_owner);
2169        arg->lk.start = fl->fl_start;
2170        arg->lk.end = fl->fl_end;
2171        arg->lk.type = fl->fl_type;
2172        arg->lk.pid = pid;
2173        if (flock)
2174                arg->lk_flags |= FUSE_LK_FLOCK;
2175        req->in.h.opcode = opcode;
2176        req->in.h.nodeid = get_node_id(inode);
2177        req->in.numargs = 1;
2178        req->in.args[0].size = sizeof(*arg);
2179        req->in.args[0].value = arg;
2180}
2181
2182static int fuse_getlk(struct file *file, struct file_lock *fl)
2183{
2184        struct inode *inode = file_inode(file);
2185        struct fuse_conn *fc = get_fuse_conn(inode);
2186        struct fuse_req *req;
2187        struct fuse_lk_out outarg;
2188        int err;
2189
2190        req = fuse_get_req_nopages(fc);
2191        if (IS_ERR(req))
2192                return PTR_ERR(req);
2193
2194        fuse_lk_fill(req, file, fl, FUSE_GETLK, 0, 0);
2195        req->out.numargs = 1;
2196        req->out.args[0].size = sizeof(outarg);
2197        req->out.args[0].value = &outarg;
2198        fuse_request_send(fc, req);
2199        err = req->out.h.error;
2200        fuse_put_request(fc, req);
2201        if (!err)
2202                err = convert_fuse_file_lock(&outarg.lk, fl);
2203
2204        return err;
2205}
2206
2207static int fuse_setlk(struct file *file, struct file_lock *fl, int flock)
2208{
2209        struct inode *inode = file_inode(file);
2210        struct fuse_conn *fc = get_fuse_conn(inode);
2211        struct fuse_req *req;
2212        int opcode = (fl->fl_flags & FL_SLEEP) ? FUSE_SETLKW : FUSE_SETLK;
2213        pid_t pid = fl->fl_type != F_UNLCK ? current->tgid : 0;
2214        int err;
2215
2216        if (fl->fl_lmops && fl->fl_lmops->lm_grant) {
2217                /* NLM needs asynchronous locks, which we don't support yet */
2218                return -ENOLCK;
2219        }
2220
2221        /* Unlock on close is handled by the flush method */
2222        if (fl->fl_flags & FL_CLOSE)
2223                return 0;
2224
2225        req = fuse_get_req_nopages(fc);
2226        if (IS_ERR(req))
2227                return PTR_ERR(req);
2228
2229        fuse_lk_fill(req, file, fl, opcode, pid, flock);
2230        fuse_request_send(fc, req);
2231        err = req->out.h.error;
2232        /* locking is restartable */
2233        if (err == -EINTR)
2234                err = -ERESTARTSYS;
2235        fuse_put_request(fc, req);
2236        return err;
2237}
2238
2239static int fuse_file_lock(struct file *file, int cmd, struct file_lock *fl)
2240{
2241        struct inode *inode = file_inode(file);
2242        struct fuse_conn *fc = get_fuse_conn(inode);
2243        int err;
2244
2245        if (cmd == F_CANCELLK) {
2246                err = 0;
2247        } else if (cmd == F_GETLK) {
2248                if (fc->no_lock) {
2249                        posix_test_lock(file, fl);
2250                        err = 0;
2251                } else
2252                        err = fuse_getlk(file, fl);
2253        } else {
2254                if (fc->no_lock)
2255                        err = posix_lock_file(file, fl, NULL);
2256                else
2257                        err = fuse_setlk(file, fl, 0);
2258        }
2259        return err;
2260}
2261
2262static int fuse_file_flock(struct file *file, int cmd, struct file_lock *fl)
2263{
2264        struct inode *inode = file_inode(file);
2265        struct fuse_conn *fc = get_fuse_conn(inode);
2266        int err;
2267
2268        if (fc->no_flock) {
2269                err = flock_lock_file_wait(file, fl);
2270        } else {
2271                struct fuse_file *ff = file->private_data;
2272
2273                /* emulate flock with POSIX locks */
2274                ff->flock = true;
2275                err = fuse_setlk(file, fl, 1);
2276        }
2277
2278        return err;
2279}
2280
2281static sector_t fuse_bmap(struct address_space *mapping, sector_t block)
2282{
2283        struct inode *inode = mapping->host;
2284        struct fuse_conn *fc = get_fuse_conn(inode);
2285        struct fuse_req *req;
2286        struct fuse_bmap_in inarg;
2287        struct fuse_bmap_out outarg;
2288        int err;
2289
2290        if (!inode->i_sb->s_bdev || fc->no_bmap)
2291                return 0;
2292
2293        req = fuse_get_req_nopages(fc);
2294        if (IS_ERR(req))
2295                return 0;
2296
2297        memset(&inarg, 0, sizeof(inarg));
2298        inarg.block = block;
2299        inarg.blocksize = inode->i_sb->s_blocksize;
2300        req->in.h.opcode = FUSE_BMAP;
2301        req->in.h.nodeid = get_node_id(inode);
2302        req->in.numargs = 1;
2303        req->in.args[0].size = sizeof(inarg);
2304        req->in.args[0].value = &inarg;
2305        req->out.numargs = 1;
2306        req->out.args[0].size = sizeof(outarg);
2307        req->out.args[0].value = &outarg;
2308        fuse_request_send(fc, req);
2309        err = req->out.h.error;
2310        fuse_put_request(fc, req);
2311        if (err == -ENOSYS)
2312                fc->no_bmap = 1;
2313
2314        return err ? 0 : outarg.block;
2315}
2316
2317static loff_t fuse_file_llseek(struct file *file, loff_t offset, int whence)
2318{
2319        loff_t retval;
2320        struct inode *inode = file_inode(file);
2321
2322        /* No i_mutex protection necessary for SEEK_CUR and SEEK_SET */
2323        if (whence == SEEK_CUR || whence == SEEK_SET)
2324                return generic_file_llseek(file, offset, whence);
2325
2326        mutex_lock(&inode->i_mutex);
2327        retval = fuse_update_attributes(inode, NULL, file, NULL);
2328        if (!retval)
2329                retval = generic_file_llseek(file, offset, whence);
2330        mutex_unlock(&inode->i_mutex);
2331
2332        return retval;
2333}
2334
2335static int fuse_ioctl_copy_user(struct page **pages, struct iovec *iov,
2336                        unsigned int nr_segs, size_t bytes, bool to_user)
2337{
2338        struct iov_iter ii;
2339        int page_idx = 0;
2340
2341        if (!bytes)
2342                return 0;
2343
2344        iov_iter_init(&ii, to_user ? READ : WRITE, iov, nr_segs, bytes);
2345
2346        while (iov_iter_count(&ii)) {
2347                struct page *page = pages[page_idx++];
2348                size_t todo = min_t(size_t, PAGE_SIZE, iov_iter_count(&ii));
2349                void *kaddr;
2350
2351                kaddr = kmap(page);
2352
2353                while (todo) {
2354                        char __user *uaddr = ii.iov->iov_base + ii.iov_offset;
2355                        size_t iov_len = ii.iov->iov_len - ii.iov_offset;
2356                        size_t copy = min(todo, iov_len);
2357                        size_t left;
2358
2359                        if (!to_user)
2360                                left = copy_from_user(kaddr, uaddr, copy);
2361                        else
2362                                left = copy_to_user(uaddr, kaddr, copy);
2363
2364                        if (unlikely(left))
2365                                return -EFAULT;
2366
2367                        iov_iter_advance(&ii, copy);
2368                        todo -= copy;
2369                        kaddr += copy;
2370                }
2371
2372                kunmap(page);
2373        }
2374
2375        return 0;
2376}
2377
2378/*
2379 * CUSE servers compiled on 32bit broke on 64bit kernels because the
2380 * ABI was defined to be 'struct iovec' which is different on 32bit
2381 * and 64bit.  Fortunately we can determine which structure the server
2382 * used from the size of the reply.
2383 */
2384static int fuse_copy_ioctl_iovec_old(struct iovec *dst, void *src,
2385                                     size_t transferred, unsigned count,
2386                                     bool is_compat)
2387{
2388#ifdef CONFIG_COMPAT
2389        if (count * sizeof(struct compat_iovec) == transferred) {
2390                struct compat_iovec *ciov = src;
2391                unsigned i;
2392
2393                /*
2394                 * With this interface a 32bit server cannot support
2395                 * non-compat (i.e. ones coming from 64bit apps) ioctl
2396                 * requests
2397                 */
2398                if (!is_compat)
2399                        return -EINVAL;
2400
2401                for (i = 0; i < count; i++) {
2402                        dst[i].iov_base = compat_ptr(ciov[i].iov_base);
2403                        dst[i].iov_len = ciov[i].iov_len;
2404                }
2405                return 0;
2406        }
2407#endif
2408
2409        if (count * sizeof(struct iovec) != transferred)
2410                return -EIO;
2411
2412        memcpy(dst, src, transferred);
2413        return 0;
2414}
2415
2416/* Make sure iov_length() won't overflow */
2417static int fuse_verify_ioctl_iov(struct iovec *iov, size_t count)
2418{
2419        size_t n;
2420        u32 max = FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT;
2421
2422        for (n = 0; n < count; n++, iov++) {
2423                if (iov->iov_len > (size_t) max)
2424                        return -ENOMEM;
2425                max -= iov->iov_len;
2426        }
2427        return 0;
2428}
2429
2430static int fuse_copy_ioctl_iovec(struct fuse_conn *fc, struct iovec *dst,
2431                                 void *src, size_t transferred, unsigned count,
2432                                 bool is_compat)
2433{
2434        unsigned i;
2435        struct fuse_ioctl_iovec *fiov = src;
2436
2437        if (fc->minor < 16) {
2438                return fuse_copy_ioctl_iovec_old(dst, src, transferred,
2439                                                 count, is_compat);
2440        }
2441
2442        if (count * sizeof(struct fuse_ioctl_iovec) != transferred)
2443                return -EIO;
2444
2445        for (i = 0; i < count; i++) {
2446                /* Did the server supply an inappropriate value? */
2447                if (fiov[i].base != (unsigned long) fiov[i].base ||
2448                    fiov[i].len != (unsigned long) fiov[i].len)
2449                        return -EIO;
2450
2451                dst[i].iov_base = (void __user *) (unsigned long) fiov[i].base;
2452                dst[i].iov_len = (size_t) fiov[i].len;
2453
2454#ifdef CONFIG_COMPAT
2455                if (is_compat &&
2456                    (ptr_to_compat(dst[i].iov_base) != fiov[i].base ||
2457                     (compat_size_t) dst[i].iov_len != fiov[i].len))
2458                        return -EIO;
2459#endif
2460        }
2461
2462        return 0;
2463}
2464
2465
2466/*
2467 * For ioctls, there is no generic way to determine how much memory
2468 * needs to be read and/or written.  Furthermore, ioctls are allowed
2469 * to dereference the passed pointer, so the parameter requires deep
2470 * copying but FUSE has no idea whatsoever about what to copy in or
2471 * out.
2472 *
2473 * This is solved by allowing FUSE server to retry ioctl with
2474 * necessary in/out iovecs.  Let's assume the ioctl implementation
2475 * needs to read in the following structure.
2476 *
2477 * struct a {
2478 *      char    *buf;
2479 *      size_t  buflen;
2480 * }
2481 *
2482 * On the first callout to FUSE server, inarg->in_size and
2483 * inarg->out_size will be NULL; then, the server completes the ioctl
2484 * with FUSE_IOCTL_RETRY set in out->flags, out->in_iovs set to 1 and
2485 * the actual iov array to
2486 *
2487 * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a) } }
2488 *
2489 * which tells FUSE to copy in the requested area and retry the ioctl.
2490 * On the second round, the server has access to the structure and
2491 * from that it can tell what to look for next, so on the invocation,
2492 * it sets FUSE_IOCTL_RETRY, out->in_iovs to 2 and iov array to
2493 *
2494 * { { .iov_base = inarg.arg,   .iov_len = sizeof(struct a)     },
2495 *   { .iov_base = a.buf,       .iov_len = a.buflen             } }
2496 *
2497 * FUSE will copy both struct a and the pointed buffer from the
2498 * process doing the ioctl and retry ioctl with both struct a and the
2499 * buffer.
2500 *
2501 * This time, FUSE server has everything it needs and completes ioctl
2502 * without FUSE_IOCTL_RETRY which finishes the ioctl call.
2503 *
2504 * Copying data out works the same way.
2505 *
2506 * Note that if FUSE_IOCTL_UNRESTRICTED is clear, the kernel
2507 * automatically initializes in and out iovs by decoding @cmd with
2508 * _IOC_* macros and the server is not allowed to request RETRY.  This
2509 * limits ioctl data transfers to well-formed ioctls and is the forced
2510 * behavior for all FUSE servers.
2511 */
2512long fuse_do_ioctl(struct file *file, unsigned int cmd, unsigned long arg,
2513                   unsigned int flags)
2514{
2515        struct fuse_file *ff = file->private_data;
2516        struct fuse_conn *fc = ff->fc;
2517        struct fuse_ioctl_in inarg = {
2518                .fh = ff->fh,
2519                .cmd = cmd,
2520                .arg = arg,
2521                .flags = flags
2522        };
2523        struct fuse_ioctl_out outarg;
2524        struct fuse_req *req = NULL;
2525        struct page **pages = NULL;
2526        struct iovec *iov_page = NULL;
2527        struct iovec *in_iov = NULL, *out_iov = NULL;
2528        unsigned int in_iovs = 0, out_iovs = 0, num_pages = 0, max_pages;
2529        size_t in_size, out_size, transferred;
2530        int err;
2531
2532#if BITS_PER_LONG == 32
2533        inarg.flags |= FUSE_IOCTL_32BIT;
2534#else
2535        if (flags & FUSE_IOCTL_COMPAT)
2536                inarg.flags |= FUSE_IOCTL_32BIT;
2537#endif
2538
2539        /* assume all the iovs returned by client always fits in a page */
2540        BUILD_BUG_ON(sizeof(struct fuse_ioctl_iovec) * FUSE_IOCTL_MAX_IOV > PAGE_SIZE);
2541
2542        err = -ENOMEM;
2543        pages = kcalloc(FUSE_MAX_PAGES_PER_REQ, sizeof(pages[0]), GFP_KERNEL);
2544        iov_page = (struct iovec *) __get_free_page(GFP_KERNEL);
2545        if (!pages || !iov_page)
2546                goto out;
2547
2548        /*
2549         * If restricted, initialize IO parameters as encoded in @cmd.
2550         * RETRY from server is not allowed.
2551         */
2552        if (!(flags & FUSE_IOCTL_UNRESTRICTED)) {
2553                struct iovec *iov = iov_page;
2554
2555                iov->iov_base = (void __user *)arg;
2556                iov->iov_len = _IOC_SIZE(cmd);
2557
2558                if (_IOC_DIR(cmd) & _IOC_WRITE) {
2559                        in_iov = iov;
2560                        in_iovs = 1;
2561                }
2562
2563                if (_IOC_DIR(cmd) & _IOC_READ) {
2564                        out_iov = iov;
2565                        out_iovs = 1;
2566                }
2567        }
2568
2569 retry:
2570        inarg.in_size = in_size = iov_length(in_iov, in_iovs);
2571        inarg.out_size = out_size = iov_length(out_iov, out_iovs);
2572
2573        /*
2574         * Out data can be used either for actual out data or iovs,
2575         * make sure there always is at least one page.
2576         */
2577        out_size = max_t(size_t, out_size, PAGE_SIZE);
2578        max_pages = DIV_ROUND_UP(max(in_size, out_size), PAGE_SIZE);
2579
2580        /* make sure there are enough buffer pages and init request with them */
2581        err = -ENOMEM;
2582        if (max_pages > FUSE_MAX_PAGES_PER_REQ)
2583                goto out;
2584        while (num_pages < max_pages) {
2585                pages[num_pages] = alloc_page(GFP_KERNEL | __GFP_HIGHMEM);
2586                if (!pages[num_pages])
2587                        goto out;
2588                num_pages++;
2589        }
2590
2591        req = fuse_get_req(fc, num_pages);
2592        if (IS_ERR(req)) {
2593                err = PTR_ERR(req);
2594                req = NULL;
2595                goto out;
2596        }
2597        memcpy(req->pages, pages, sizeof(req->pages[0]) * num_pages);
2598        req->num_pages = num_pages;
2599        fuse_page_descs_length_init(req, 0, req->num_pages);
2600
2601        /* okay, let's send it to the client */
2602        req->in.h.opcode = FUSE_IOCTL;
2603        req->in.h.nodeid = ff->nodeid;
2604        req->in.numargs = 1;
2605        req->in.args[0].size = sizeof(inarg);
2606        req->in.args[0].value = &inarg;
2607        if (in_size) {
2608                req->in.numargs++;
2609                req->in.args[1].size = in_size;
2610                req->in.argpages = 1;
2611
2612                err = fuse_ioctl_copy_user(pages, in_iov, in_iovs, in_size,
2613                                           false);
2614                if (err)
2615                        goto out;
2616        }
2617
2618        req->out.numargs = 2;
2619        req->out.args[0].size = sizeof(outarg);
2620        req->out.args[0].value = &outarg;
2621        req->out.args[1].size = out_size;
2622        req->out.argpages = 1;
2623        req->out.argvar = 1;
2624
2625        fuse_request_send(fc, req);
2626        err = req->out.h.error;
2627        transferred = req->out.args[1].size;
2628        fuse_put_request(fc, req);
2629        req = NULL;
2630        if (err)
2631                goto out;
2632
2633        /* did it ask for retry? */
2634        if (outarg.flags & FUSE_IOCTL_RETRY) {
2635                void *vaddr;
2636
2637                /* no retry if in restricted mode */
2638                err = -EIO;
2639                if (!(flags & FUSE_IOCTL_UNRESTRICTED))
2640                        goto out;
2641
2642                in_iovs = outarg.in_iovs;
2643                out_iovs = outarg.out_iovs;
2644
2645                /*
2646                 * Make sure things are in boundary, separate checks
2647                 * are to protect against overflow.
2648                 */
2649                err = -ENOMEM;
2650                if (in_iovs > FUSE_IOCTL_MAX_IOV ||
2651                    out_iovs > FUSE_IOCTL_MAX_IOV ||
2652                    in_iovs + out_iovs > FUSE_IOCTL_MAX_IOV)
2653                        goto out;
2654
2655                vaddr = kmap_atomic(pages[0]);
2656                err = fuse_copy_ioctl_iovec(fc, iov_page, vaddr,
2657                                            transferred, in_iovs + out_iovs,
2658                                            (flags & FUSE_IOCTL_COMPAT) != 0);
2659                kunmap_atomic(vaddr);
2660                if (err)
2661                        goto out;
2662
2663                in_iov = iov_page;
2664                out_iov = in_iov + in_iovs;
2665
2666                err = fuse_verify_ioctl_iov(in_iov, in_iovs);
2667                if (err)
2668                        goto out;
2669
2670                err = fuse_verify_ioctl_iov(out_iov, out_iovs);
2671                if (err)
2672                        goto out;
2673
2674                goto retry;
2675        }
2676
2677        err = -EIO;
2678        if (transferred > inarg.out_size)
2679                goto out;
2680
2681        err = fuse_ioctl_copy_user(pages, out_iov, out_iovs, transferred, true);
2682 out:
2683        if (req)
2684                fuse_put_request(fc, req);
2685        free_page((unsigned long) iov_page);
2686        while (num_pages)
2687                __free_page(pages[--num_pages]);
2688        kfree(pages);
2689
2690        return err ? err : outarg.result;
2691}
2692EXPORT_SYMBOL_GPL(fuse_do_ioctl);
2693
2694long fuse_ioctl_common(struct file *file, unsigned int cmd,
2695                       unsigned long arg, unsigned int flags)
2696{
2697        struct inode *inode = file_inode(file);
2698        struct fuse_conn *fc = get_fuse_conn(inode);
2699
2700        if (!fuse_allow_current_process(fc))
2701                return -EACCES;
2702
2703        if (is_bad_inode(inode))
2704                return -EIO;
2705
2706        return fuse_do_ioctl(file, cmd, arg, flags);
2707}
2708
2709static long fuse_file_ioctl(struct file *file, unsigned int cmd,
2710                            unsigned long arg)
2711{
2712        return fuse_ioctl_common(file, cmd, arg, 0);
2713}
2714
2715static long fuse_file_compat_ioctl(struct file *file, unsigned int cmd,
2716                                   unsigned long arg)
2717{
2718        return fuse_ioctl_common(file, cmd, arg, FUSE_IOCTL_COMPAT);
2719}
2720
2721/*
2722 * All files which have been polled are linked to RB tree
2723 * fuse_conn->polled_files which is indexed by kh.  Walk the tree and
2724 * find the matching one.
2725 */
2726static struct rb_node **fuse_find_polled_node(struct fuse_conn *fc, u64 kh,
2727                                              struct rb_node **parent_out)
2728{
2729        struct rb_node **link = &fc->polled_files.rb_node;
2730        struct rb_node *last = NULL;
2731
2732        while (*link) {
2733                struct fuse_file *ff;
2734
2735                last = *link;
2736                ff = rb_entry(last, struct fuse_file, polled_node);
2737
2738                if (kh < ff->kh)
2739                        link = &last->rb_left;
2740                else if (kh > ff->kh)
2741                        link = &last->rb_right;
2742                else
2743                        return link;
2744        }
2745
2746        if (parent_out)
2747                *parent_out = last;
2748        return link;
2749}
2750
2751/*
2752 * The file is about to be polled.  Make sure it's on the polled_files
2753 * RB tree.  Note that files once added to the polled_files tree are
2754 * not removed before the file is released.  This is because a file
2755 * polled once is likely to be polled again.
2756 */
2757static void fuse_register_polled_file(struct fuse_conn *fc,
2758                                      struct fuse_file *ff)
2759{
2760        spin_lock(&fc->lock);
2761        if (RB_EMPTY_NODE(&ff->polled_node)) {
2762                struct rb_node **link, *uninitialized_var(parent);
2763
2764                link = fuse_find_polled_node(fc, ff->kh, &parent);
2765                BUG_ON(*link);
2766                rb_link_node(&ff->polled_node, parent, link);
2767                rb_insert_color(&ff->polled_node, &fc->polled_files);
2768        }
2769        spin_unlock(&fc->lock);
2770}
2771
2772unsigned fuse_file_poll(struct file *file, poll_table *wait)
2773{
2774        struct fuse_file *ff = file->private_data;
2775        struct fuse_conn *fc = ff->fc;
2776        struct fuse_poll_in inarg = { .fh = ff->fh, .kh = ff->kh };
2777        struct fuse_poll_out outarg;
2778        struct fuse_req *req;
2779        int err;
2780
2781        if (fc->no_poll)
2782                return DEFAULT_POLLMASK;
2783
2784        poll_wait(file, &ff->poll_wait, wait);
2785        inarg.events = (__u32)poll_requested_events(wait);
2786
2787        /*
2788         * Ask for notification iff there's someone waiting for it.
2789         * The client may ignore the flag and always notify.
2790         */
2791        if (waitqueue_active(&ff->poll_wait)) {
2792                inarg.flags |= FUSE_POLL_SCHEDULE_NOTIFY;
2793                fuse_register_polled_file(fc, ff);
2794        }
2795
2796        req = fuse_get_req_nopages(fc);
2797        if (IS_ERR(req))
2798                return POLLERR;
2799
2800        req->in.h.opcode = FUSE_POLL;
2801        req->in.h.nodeid = ff->nodeid;
2802        req->in.numargs = 1;
2803        req->in.args[0].size = sizeof(inarg);
2804        req->in.args[0].value = &inarg;
2805        req->out.numargs = 1;
2806        req->out.args[0].size = sizeof(outarg);
2807        req->out.args[0].value = &outarg;
2808        fuse_request_send(fc, req);
2809        err = req->out.h.error;
2810        fuse_put_request(fc, req);
2811
2812        if (!err)
2813                return outarg.revents;
2814        if (err == -ENOSYS) {
2815                fc->no_poll = 1;
2816                return DEFAULT_POLLMASK;
2817        }
2818        return POLLERR;
2819}
2820EXPORT_SYMBOL_GPL(fuse_file_poll);
2821
2822/*
2823 * This is called from fuse_handle_notify() on FUSE_NOTIFY_POLL and
2824 * wakes up the poll waiters.
2825 */
2826int fuse_notify_poll_wakeup(struct fuse_conn *fc,
2827                            struct fuse_notify_poll_wakeup_out *outarg)
2828{
2829        u64 kh = outarg->kh;
2830        struct rb_node **link;
2831
2832        spin_lock(&fc->lock);
2833
2834        link = fuse_find_polled_node(fc, kh, NULL);
2835        if (*link) {
2836                struct fuse_file *ff;
2837
2838                ff = rb_entry(*link, struct fuse_file, polled_node);
2839                wake_up_interruptible_sync(&ff->poll_wait);
2840        }
2841
2842        spin_unlock(&fc->lock);
2843        return 0;
2844}
2845
2846static void fuse_do_truncate(struct file *file)
2847{
2848        struct inode *inode = file->f_mapping->host;
2849        struct iattr attr;
2850
2851        attr.ia_valid = ATTR_SIZE;
2852        attr.ia_size = i_size_read(inode);
2853
2854        attr.ia_file = file;
2855        attr.ia_valid |= ATTR_FILE;
2856
2857        fuse_do_setattr(inode, &attr, file);
2858}
2859
2860static inline loff_t fuse_round_up(loff_t off)
2861{
2862        return round_up(off, FUSE_MAX_PAGES_PER_REQ << PAGE_SHIFT);
2863}
2864
2865static ssize_t
2866fuse_direct_IO(int rw, struct kiocb *iocb, struct iov_iter *iter,
2867                        loff_t offset)
2868{
2869        ssize_t ret = 0;
2870        struct file *file = iocb->ki_filp;
2871        struct fuse_file *ff = file->private_data;
2872        bool async_dio = ff->fc->async_dio;
2873        loff_t pos = 0;
2874        struct inode *inode;
2875        loff_t i_size;
2876        size_t count = iov_iter_count(iter);
2877        struct fuse_io_priv *io;
2878
2879        pos = offset;
2880        inode = file->f_mapping->host;
2881        i_size = i_size_read(inode);
2882
2883        if ((rw == READ) && (offset > i_size))
2884                return 0;
2885
2886        /* optimization for short read */
2887        if (async_dio && rw != WRITE && offset + count > i_size) {
2888                if (offset >= i_size)
2889                        return 0;
2890                count = min_t(loff_t, count, fuse_round_up(i_size - offset));
2891                iov_iter_truncate(iter, count);
2892        }
2893
2894        io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL);
2895        if (!io)
2896                return -ENOMEM;
2897        spin_lock_init(&io->lock);
2898        io->reqs = 1;
2899        io->bytes = -1;
2900        io->size = 0;
2901        io->offset = offset;
2902        io->write = (rw == WRITE);
2903        io->err = 0;
2904        io->file = file;
2905        /*
2906         * By default, we want to optimize all I/Os with async request
2907         * submission to the client filesystem if supported.
2908         */
2909        io->async = async_dio;
2910        io->iocb = iocb;
2911
2912        /*
2913         * We cannot asynchronously extend the size of a file. We have no method
2914         * to wait on real async I/O requests, so we must submit this request
2915         * synchronously.
2916         */
2917        if (!is_sync_kiocb(iocb) && (offset + count > i_size) && rw == WRITE)
2918                io->async = false;
2919
2920        if (rw == WRITE)
2921                ret = __fuse_direct_write(io, iter, &pos);
2922        else
2923                ret = __fuse_direct_read(io, iter, &pos);
2924
2925        if (io->async) {
2926                fuse_aio_complete(io, ret < 0 ? ret : 0, -1);
2927
2928                /* we have a non-extending, async request, so return */
2929                if (!is_sync_kiocb(iocb))
2930                        return -EIOCBQUEUED;
2931
2932                ret = wait_on_sync_kiocb(iocb);
2933        } else {
2934                kfree(io);
2935        }
2936
2937        if (rw == WRITE) {
2938                if (ret > 0)
2939                        fuse_write_update_size(inode, pos);
2940                else if (ret < 0 && offset + count > i_size)
2941                        fuse_do_truncate(file);
2942        }
2943
2944        return ret;
2945}
2946
2947static long fuse_file_fallocate(struct file *file, int mode, loff_t offset,
2948                                loff_t length)
2949{
2950        struct fuse_file *ff = file->private_data;
2951        struct inode *inode = file->f_inode;
2952        struct fuse_inode *fi = get_fuse_inode(inode);
2953        struct fuse_conn *fc = ff->fc;
2954        struct fuse_req *req;
2955        struct fuse_fallocate_in inarg = {
2956                .fh = ff->fh,
2957                .offset = offset,
2958                .length = length,
2959                .mode = mode
2960        };
2961        int err;
2962        bool lock_inode = !(mode & FALLOC_FL_KEEP_SIZE) ||
2963                           (mode & FALLOC_FL_PUNCH_HOLE);
2964
2965        if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2966                return -EOPNOTSUPP;
2967
2968        if (fc->no_fallocate)
2969                return -EOPNOTSUPP;
2970
2971        if (lock_inode) {
2972                mutex_lock(&inode->i_mutex);
2973                if (mode & FALLOC_FL_PUNCH_HOLE) {
2974                        loff_t endbyte = offset + length - 1;
2975                        err = filemap_write_and_wait_range(inode->i_mapping,
2976                                                           offset, endbyte);
2977                        if (err)
2978                                goto out;
2979
2980                        fuse_sync_writes(inode);
2981                }
2982        }
2983
2984        if (!(mode & FALLOC_FL_KEEP_SIZE))
2985                set_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
2986
2987        req = fuse_get_req_nopages(fc);
2988        if (IS_ERR(req)) {
2989                err = PTR_ERR(req);
2990                goto out;
2991        }
2992
2993        req->in.h.opcode = FUSE_FALLOCATE;
2994        req->in.h.nodeid = ff->nodeid;
2995        req->in.numargs = 1;
2996        req->in.args[0].size = sizeof(inarg);
2997        req->in.args[0].value = &inarg;
2998        fuse_request_send(fc, req);
2999        err = req->out.h.error;
3000        if (err == -ENOSYS) {
3001                fc->no_fallocate = 1;
3002                err = -EOPNOTSUPP;
3003        }
3004        fuse_put_request(fc, req);
3005
3006        if (err)
3007                goto out;
3008
3009        /* we could have extended the file */
3010        if (!(mode & FALLOC_FL_KEEP_SIZE)) {
3011                bool changed = fuse_write_update_size(inode, offset + length);
3012
3013                if (changed && fc->writeback_cache)
3014                        file_update_time(file);
3015        }
3016
3017        if (mode & FALLOC_FL_PUNCH_HOLE)
3018                truncate_pagecache_range(inode, offset, offset + length - 1);
3019
3020        fuse_invalidate_attr(inode);
3021
3022out:
3023        if (!(mode & FALLOC_FL_KEEP_SIZE))
3024                clear_bit(FUSE_I_SIZE_UNSTABLE, &fi->state);
3025
3026        if (lock_inode)
3027                mutex_unlock(&inode->i_mutex);
3028
3029        return err;
3030}
3031
3032static const struct file_operations fuse_file_operations = {
3033        .llseek         = fuse_file_llseek,
3034        .read           = new_sync_read,
3035        .read_iter      = fuse_file_read_iter,
3036        .write          = new_sync_write,
3037        .write_iter     = fuse_file_write_iter,
3038        .mmap           = fuse_file_mmap,
3039        .open           = fuse_open,
3040        .flush          = fuse_flush,
3041        .release        = fuse_release,
3042        .fsync          = fuse_fsync,
3043        .lock           = fuse_file_lock,
3044        .flock          = fuse_file_flock,
3045        .splice_read    = generic_file_splice_read,
3046        .unlocked_ioctl = fuse_file_ioctl,
3047        .compat_ioctl   = fuse_file_compat_ioctl,
3048        .poll           = fuse_file_poll,
3049        .fallocate      = fuse_file_fallocate,
3050};
3051
3052static const struct file_operations fuse_direct_io_file_operations = {
3053        .llseek         = fuse_file_llseek,
3054        .read           = fuse_direct_read,
3055        .write          = fuse_direct_write,
3056        .mmap           = fuse_direct_mmap,
3057        .open           = fuse_open,
3058        .flush          = fuse_flush,
3059        .release        = fuse_release,
3060        .fsync          = fuse_fsync,
3061        .lock           = fuse_file_lock,
3062        .flock          = fuse_file_flock,
3063        .unlocked_ioctl = fuse_file_ioctl,
3064        .compat_ioctl   = fuse_file_compat_ioctl,
3065        .poll           = fuse_file_poll,
3066        .fallocate      = fuse_file_fallocate,
3067        /* no splice_read */
3068};
3069
3070static const struct address_space_operations fuse_file_aops  = {
3071        .readpage       = fuse_readpage,
3072        .writepage      = fuse_writepage,
3073        .writepages     = fuse_writepages,
3074        .launder_page   = fuse_launder_page,
3075        .readpages      = fuse_readpages,
3076        .set_page_dirty = __set_page_dirty_nobuffers,
3077        .bmap           = fuse_bmap,
3078        .direct_IO      = fuse_direct_IO,
3079        .write_begin    = fuse_write_begin,
3080        .write_end      = fuse_write_end,
3081};
3082
3083void fuse_init_file_inode(struct inode *inode)
3084{
3085        inode->i_fop = &fuse_file_operations;
3086        inode->i_data.a_ops = &fuse_file_aops;
3087}
3088
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.