fs/orangefs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * (C) 2001 Clemson University and The University of Chicago
   4  * Copyright 2018 Omnibond Systems, L.L.C.
   5  *
   6  * See COPYING in top-level directory.
   7  */
   8
   9 /*
  10  *  Linux VFS file operations.
  11  */
  12
  13 #include "protocol.h"
  14 #include "orangefs-kernel.h"
  15 #include "orangefs-bufmap.h"
  16 #include <linux/fs.h>
  17 #include <linux/pagemap.h>
  18
  19 static int flush_racache(struct inode *inode)
  20 {
  21         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
  22         struct orangefs_kernel_op_s *new_op;
  23         int ret;
  24
  25         gossip_debug(GOSSIP_UTILS_DEBUG,
  26             "%s: %pU: Handle is %pU | fs_id %d\n", __func__,
  27             get_khandle_from_ino(inode), &orangefs_inode->refn.khandle,
  28             orangefs_inode->refn.fs_id);
  29
  30         new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
  31         if (!new_op)
  32                 return -ENOMEM;
  33         new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
  34
  35         ret = service_operation(new_op, "orangefs_flush_racache",
  36             get_interruptible_flag(inode));
  37
  38         gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
  39             __func__, ret);
  40
  41         op_release(new_op);
  42         return ret;
  43 }
  44
  45 /*
  46  * Post and wait for the I/O upcall to finish
  47  */
  48 ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
  49         loff_t *offset, struct iov_iter *iter, size_t total_size,
  50         loff_t readahead_size, struct orangefs_write_range *wr,
  51         int *index_return, struct file *file)
  52 {
  53         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
  54         struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
  55         struct orangefs_kernel_op_s *new_op = NULL;
  56         int buffer_index;
  57         ssize_t ret;
  58         size_t copy_amount;
  59         int open_for_read;
  60         int open_for_write;
  61
  62         new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
  63         if (!new_op)
  64                 return -ENOMEM;
  65
  66         /* synchronous I/O */
  67         new_op->upcall.req.io.readahead_size = readahead_size;
  68         new_op->upcall.req.io.io_type = type;
  69         new_op->upcall.req.io.refn = orangefs_inode->refn;
  70
  71 populate_shared_memory:
  72         /* get a shared buffer index */
  73         buffer_index = orangefs_bufmap_get();
  74         if (buffer_index < 0) {
  75                 ret = buffer_index;
  76                 gossip_debug(GOSSIP_FILE_DEBUG,
  77                              "%s: orangefs_bufmap_get failure (%zd)\n",
  78                              __func__, ret);
  79                 goto out;
  80         }
  81         gossip_debug(GOSSIP_FILE_DEBUG,
  82                      "%s(%pU): GET op %p -> buffer_index %d\n",
  83                      __func__,
  84                      handle,
  85                      new_op,
  86                      buffer_index);
  87
  88         new_op->uses_shared_memory = 1;
  89         new_op->upcall.req.io.buf_index = buffer_index;
  90         new_op->upcall.req.io.count = total_size;
  91         new_op->upcall.req.io.offset = *offset;
  92         if (type == ORANGEFS_IO_WRITE && wr) {
  93                 new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
  94                 new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
  95         }
  96         /*
  97          * Orangefs has no open, and orangefs checks file permissions
  98          * on each file access. Posix requires that file permissions
  99          * be checked on open and nowhere else. Orangefs-through-the-kernel
 100          * needs to seem posix compliant.
 101          *
 102          * The VFS opens files, even if the filesystem provides no
 103          * method. We can see if a file was successfully opened for
 104          * read and or for write by looking at file->f_mode.
 105          *
 106          * When writes are flowing from the page cache, file is no
 107          * longer available. We can trust the VFS to have checked
 108          * file->f_mode before writing to the page cache.
 109          *
 110          * The mode of a file might change between when it is opened
 111          * and IO commences, or it might be created with an arbitrary mode.
 112          *
 113          * We'll make sure we don't hit EACCES during the IO stage by
 114          * using UID 0. Some of the time we have access without changing
 115          * to UID 0 - how to check?
 116          */
 117         if (file) {
 118                 open_for_write = file->f_mode & FMODE_WRITE;
 119                 open_for_read = file->f_mode & FMODE_READ;
 120         } else {
 121                 open_for_write = 1;
 122                 open_for_read = 0; /* not relevant? */
 123         }
 124         if ((type == ORANGEFS_IO_WRITE) && open_for_write)
 125                 new_op->upcall.uid = 0;
 126         if ((type == ORANGEFS_IO_READ) && open_for_read)
 127                 new_op->upcall.uid = 0;
 128
 129         gossip_debug(GOSSIP_FILE_DEBUG,
 130                      "%s(%pU): offset: %llu total_size: %zd\n",
 131                      __func__,
 132                      handle,
 133                      llu(*offset),
 134                      total_size);
 135         /*
 136          * Stage 1: copy the buffers into client-core's address space
 137          */
 138         if (type == ORANGEFS_IO_WRITE && total_size) {
 139                 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
 140                     total_size);
 141                 if (ret < 0) {
 142                         gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
 143                             __func__, (long)ret);
 144                         goto out;
 145                 }
 146         }
 147
 148         gossip_debug(GOSSIP_FILE_DEBUG,
 149                      "%s(%pU): Calling post_io_request with tag (%llu)\n",
 150                      __func__,
 151                      handle,
 152                      llu(new_op->tag));
 153
 154         /* Stage 2: Service the I/O operation */
 155         ret = service_operation(new_op,
 156                                 type == ORANGEFS_IO_WRITE ?
 157                                         "file_write" :
 158                                         "file_read",
 159                                 get_interruptible_flag(inode));
 160
 161         /*
 162          * If service_operation() returns -EAGAIN #and# the operation was
 163          * purged from orangefs_request_list or htable_ops_in_progress, then
 164          * we know that the client was restarted, causing the shared memory
 165          * area to be wiped clean.  To restart a  write operation in this
 166          * case, we must re-copy the data from the user's iovec to a NEW
 167          * shared memory location. To restart a read operation, we must get
 168          * a new shared memory location.
 169          */
 170         if (ret == -EAGAIN && op_state_purged(new_op)) {
 171                 orangefs_bufmap_put(buffer_index);
 172                 if (type == ORANGEFS_IO_WRITE)
 173                         iov_iter_revert(iter, total_size);
 174                 gossip_debug(GOSSIP_FILE_DEBUG,
 175                              "%s:going to repopulate_shared_memory.\n",
 176                              __func__);
 177                 goto populate_shared_memory;
 178         }
 179
 180         if (ret < 0) {
 181                 if (ret == -EINTR) {
 182                         /*
 183                          * We can't return EINTR if any data was written,
 184                          * it's not POSIX. It is minimally acceptable
 185                          * to give a partial write, the way NFS does.
 186                          *
 187                          * It would be optimal to return all or nothing,
 188                          * but if a userspace write is bigger than
 189                          * an IO buffer, and the interrupt occurs
 190                          * between buffer writes, that would not be
 191                          * possible.
 192                          */
 193                         switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
 194                         /*
 195                          * If the op was waiting when the interrupt
 196                          * occurred, then the client-core did not
 197                          * trigger the write.
 198                          */
 199                         case OP_VFS_STATE_WAITING:
 200                                 if (*offset == 0)
 201                                         ret = -EINTR;
 202                                 else
 203                                         ret = 0;
 204                                 break;
 205                         /*
 206                          * If the op was in progress when the interrupt
 207                          * occurred, then the client-core was able to
 208                          * trigger the write.
 209                          */
 210                         case OP_VFS_STATE_INPROGR:
 211                                 if (type == ORANGEFS_IO_READ)
 212                                         ret = -EINTR;
 213                                 else
 214                                         ret = total_size;
 215                                 break;
 216                         default:
 217                                 gossip_err("%s: unexpected op state :%d:.\n",
 218                                            __func__,
 219                                            new_op->op_state);
 220                                 ret = 0;
 221                                 break;
 222                         }
 223                         gossip_debug(GOSSIP_FILE_DEBUG,
 224                                      "%s: got EINTR, state:%d: %p\n",
 225                                      __func__,
 226                                      new_op->op_state,
 227                                      new_op);
 228                 } else {
 229                         gossip_err("%s: error in %s handle %pU, returning %zd\n",
 230                                 __func__,
 231                                 type == ORANGEFS_IO_READ ?
 232                                         "read from" : "write to",
 233                                 handle, ret);
 234                 }
 235                 if (orangefs_cancel_op_in_progress(new_op))
 236                         return ret;
 237
 238                 goto out;
 239         }
 240
 241         /*
 242          * Stage 3: Post copy buffers from client-core's address space
 243          */
 244         if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
 245                 /*
 246                  * NOTE: the iovector can either contain addresses which
 247                  *       can futher be kernel-space or user-space addresses.
 248                  *       or it can pointers to struct page's
 249                  */
 250
 251                 /*
 252                  * When reading, readahead_size will only be zero when
 253                  * we're doing O_DIRECT, otherwise we got here from
 254                  * orangefs_readpage.
 255                  *
 256                  * If we got here from orangefs_readpage we want to
 257                  * copy either a page or the whole file into the io
 258                  * vector, whichever is smaller.
 259                  */
 260                 if (readahead_size)
 261                         copy_amount =
 262                                 min(new_op->downcall.resp.io.amt_complete,
 263                                         (__s64)PAGE_SIZE);
 264                 else
 265                         copy_amount = new_op->downcall.resp.io.amt_complete;
 266
 267                 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
 268                         copy_amount);
 269                 if (ret < 0) {
 270                         gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
 271                             __func__, (long)ret);
 272                         goto out;
 273                 }
 274         }
 275         gossip_debug(GOSSIP_FILE_DEBUG,
 276             "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
 277             __func__,
 278             handle,
 279             type == ORANGEFS_IO_READ ?  "read" : "written",
 280             (int)new_op->downcall.resp.io.amt_complete);
 281
 282         ret = new_op->downcall.resp.io.amt_complete;
 283
 284 out:
 285         if (buffer_index >= 0) {
 286                 if ((readahead_size) && (type == ORANGEFS_IO_READ)) {
 287                         /* readpage */
 288                         *index_return = buffer_index;
 289                         gossip_debug(GOSSIP_FILE_DEBUG,
 290                                 "%s: hold on to buffer_index :%d:\n",
 291                                 __func__, buffer_index);
 292                 } else {
 293                         /* O_DIRECT */
 294                         orangefs_bufmap_put(buffer_index);
 295                         gossip_debug(GOSSIP_FILE_DEBUG,
 296                                 "%s(%pU): PUT buffer_index %d\n",
 297                                 __func__, handle, buffer_index);
 298                 }
 299         }
 300         op_release(new_op);
 301         return ret;
 302 }
 303
 304 int orangefs_revalidate_mapping(struct inode *inode)
 305 {
 306         struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 307         struct address_space *mapping = inode->i_mapping;
 308         unsigned long *bitlock = &orangefs_inode->bitlock;
 309         int ret;
 310
 311         while (1) {
 312                 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
 313                 if (ret)
 314                         return ret;
 315                 spin_lock(&inode->i_lock);
 316                 if (test_bit(1, bitlock)) {
 317                         spin_unlock(&inode->i_lock);
 318                         continue;
 319                 }
 320                 if (!time_before(jiffies, orangefs_inode->mapping_time))
 321                         break;
 322                 spin_unlock(&inode->i_lock);
 323                 return 0;
 324         }
 325
 326         set_bit(1, bitlock);
 327         smp_wmb();
 328         spin_unlock(&inode->i_lock);
 329
 330         unmap_mapping_range(mapping, 0, 0, 0);
 331         ret = filemap_write_and_wait(mapping);
 332         if (!ret)
 333                 ret = invalidate_inode_pages2(mapping);
 334
 335         orangefs_inode->mapping_time = jiffies +
 336             orangefs_cache_timeout_msecs*HZ/1000;
 337
 338         clear_bit(1, bitlock);
 339         smp_mb__after_atomic();
 340         wake_up_bit(bitlock, 1);
 341
 342         return ret;
 343 }
 344
 345 static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
 346     struct iov_iter *iter)
 347 {
 348         int ret;
 349         orangefs_stats.reads++;
 350
 351         down_read(&file_inode(iocb->ki_filp)->i_rwsem);
 352         ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
 353         if (ret)
 354                 goto out;
 355
 356         ret = generic_file_read_iter(iocb, iter);
 357 out:
 358         up_read(&file_inode(iocb->ki_filp)->i_rwsem);
 359         return ret;
 360 }
 361
 362 static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
 363     struct iov_iter *iter)
 364 {
 365         int ret;
 366         orangefs_stats.writes++;
 367
 368         if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
 369                 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
 370                 if (ret)
 371                         return ret;
 372         }
 373
 374         ret = generic_file_write_iter(iocb, iter);
 375         return ret;
 376 }
 377
 378 static vm_fault_t orangefs_fault(struct vm_fault *vmf)
 379 {
 380         struct file *file = vmf->vma->vm_file;
 381         int ret;
 382         ret = orangefs_inode_getattr(file->f_mapping->host,
 383             ORANGEFS_GETATTR_SIZE);
 384         if (ret == -ESTALE)
 385                 ret = -EIO;
 386         if (ret) {
 387                 gossip_err("%s: orangefs_inode_getattr failed, "
 388                     "ret:%d:.\n", __func__, ret);
 389                 return VM_FAULT_SIGBUS;
 390         }
 391         return filemap_fault(vmf);
 392 }
 393
 394 static const struct vm_operations_struct orangefs_file_vm_ops = {
 395         .fault = orangefs_fault,
 396         .map_pages = filemap_map_pages,
 397         .page_mkwrite = orangefs_page_mkwrite,
 398 };
 399
 400 /*
 401  * Memory map a region of a file.
 402  */
 403 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
 404 {
 405         int ret;
 406
 407         ret = orangefs_revalidate_mapping(file_inode(file));
 408         if (ret)
 409                 return ret;
 410
 411         gossip_debug(GOSSIP_FILE_DEBUG,
 412                      "orangefs_file_mmap: called on %pD\n", file);
 413
 414         /* set the sequential readahead hint */
 415         vma->vm_flags |= VM_SEQ_READ;
 416         vma->vm_flags &= ~VM_RAND_READ;
 417
 418         file_accessed(file);
 419         vma->vm_ops = &orangefs_file_vm_ops;
 420         return 0;
 421 }
 422
 423 #define mapping_nrpages(idata) ((idata)->nrpages)
 424
 425 /*
 426  * Called to notify the module that there are no more references to
 427  * this file (i.e. no processes have it open).
 428  *
 429  * \note Not called when each file is closed.
 430  */
 431 static int orangefs_file_release(struct inode *inode, struct file *file)
 432 {
 433         gossip_debug(GOSSIP_FILE_DEBUG,
 434                      "orangefs_file_release: called on %pD\n",
 435                      file);
 436
 437         /*
 438          * remove all associated inode pages from the page cache and
 439          * readahead cache (if any); this forces an expensive refresh of
 440          * data for the next caller of mmap (or 'get_block' accesses)
 441          */
 442         if (file_inode(file) &&
 443             file_inode(file)->i_mapping &&
 444             mapping_nrpages(&file_inode(file)->i_data)) {
 445                 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
 446                         gossip_debug(GOSSIP_INODE_DEBUG,
 447                             "calling flush_racache on %pU\n",
 448                             get_khandle_from_ino(inode));
 449                         flush_racache(inode);
 450                         gossip_debug(GOSSIP_INODE_DEBUG,
 451                             "flush_racache finished\n");
 452                 }
 453
 454         }
 455         return 0;
 456 }
 457
 458 /*
 459  * Push all data for a specific file onto permanent storage.
 460  */
 461 static int orangefs_fsync(struct file *file,
 462                        loff_t start,
 463                        loff_t end,
 464                        int datasync)
 465 {
 466         int ret;
 467         struct orangefs_inode_s *orangefs_inode =
 468                 ORANGEFS_I(file_inode(file));
 469         struct orangefs_kernel_op_s *new_op = NULL;
 470
 471         ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
 472             start, end);
 473         if (ret < 0)
 474                 return ret;
 475
 476         new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
 477         if (!new_op)
 478                 return -ENOMEM;
 479         new_op->upcall.req.fsync.refn = orangefs_inode->refn;
 480
 481         ret = service_operation(new_op,
 482                         "orangefs_fsync",
 483                         get_interruptible_flag(file_inode(file)));
 484
 485         gossip_debug(GOSSIP_FILE_DEBUG,
 486                      "orangefs_fsync got return value of %d\n",
 487                      ret);
 488
 489         op_release(new_op);
 490         return ret;
 491 }
 492
 493 /*
 494  * Change the file pointer position for an instance of an open file.
 495  *
 496  * \note If .llseek is overriden, we must acquire lock as described in
 497  *       Documentation/filesystems/locking.rst.
 498  *
 499  * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
 500  * require much changes to the FS
 501  */
 502 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
 503 {
 504         int ret = -EINVAL;
 505         struct inode *inode = file_inode(file);
 506
 507         if (origin == SEEK_END) {
 508                 /*
 509                  * revalidate the inode's file size.
 510                  * NOTE: We are only interested in file size here,
 511                  * so we set mask accordingly.
 512                  */
 513                 ret = orangefs_inode_getattr(file->f_mapping->host,
 514                     ORANGEFS_GETATTR_SIZE);
 515                 if (ret == -ESTALE)
 516                         ret = -EIO;
 517                 if (ret) {
 518                         gossip_debug(GOSSIP_FILE_DEBUG,
 519                                      "%s:%s:%d calling make bad inode\n",
 520                                      __FILE__,
 521                                      __func__,
 522                                      __LINE__);
 523                         return ret;
 524                 }
 525         }
 526
 527         gossip_debug(GOSSIP_FILE_DEBUG,
 528                      "orangefs_file_llseek: offset is %ld | origin is %d"
 529                      " | inode size is %lu\n",
 530                      (long)offset,
 531                      origin,
 532                      (unsigned long)i_size_read(inode));
 533
 534         return generic_file_llseek(file, offset, origin);
 535 }
 536
 537 /*
 538  * Support local locks (locks that only this kernel knows about)
 539  * if Orangefs was mounted -o local_lock.
 540  */
 541 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
 542 {
 543         int rc = -EINVAL;
 544
 545         if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
 546                 if (cmd == F_GETLK) {
 547                         rc = 0;
 548                         posix_test_lock(filp, fl);
 549                 } else {
 550                         rc = posix_lock_file(filp, fl, NULL);
 551                 }
 552         }
 553
 554         return rc;
 555 }
 556
 557 static int orangefs_flush(struct file *file, fl_owner_t id)
 558 {
 559         /*
 560          * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
 561          * service_operation in orangefs_fsync.
 562          *
 563          * Do not send fsync to OrangeFS server on a close.  Do send fsync
 564          * on an explicit fsync call.  This duplicates historical OrangeFS
 565          * behavior.
 566          */
 567         int r;
 568
 569         r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
 570         if (r > 0)
 571                 return 0;
 572         else
 573                 return r;
 574 }
 575
 576 /** ORANGEFS implementation of VFS file operations */
 577 const struct file_operations orangefs_file_operations = {
 578         .llseek         = orangefs_file_llseek,
 579         .read_iter      = orangefs_file_read_iter,
 580         .write_iter     = orangefs_file_write_iter,
 581         .lock           = orangefs_lock,
 582         .mmap           = orangefs_file_mmap,
 583         .open           = generic_file_open,
 584         .splice_read    = generic_file_splice_read,
 585         .splice_write   = iter_file_splice_write,
 586         .flush          = orangefs_flush,
 587         .release        = orangefs_file_release,
 588         .fsync          = orangefs_fsync,
 589 };