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 struct orangefs_read_options *ro; 350 351 orangefs_stats.reads++; 352 353 /* 354 * Remember how they set "count" in read(2) or pread(2) or whatever - 355 * users can use count as a knob to control orangefs io size and later 356 * we can try to help them fill as many pages as possible in readpage. 357 */ 358 if (!iocb->ki_filp->private_data) { 359 iocb->ki_filp->private_data = kmalloc(sizeof *ro, GFP_KERNEL); 360 if (!iocb->ki_filp->private_data) 361 return(ENOMEM); 362 ro = iocb->ki_filp->private_data; 363 ro->blksiz = iter->count; 364 } 365 366 down_read(&file_inode(iocb->ki_filp)->i_rwsem); 367 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp)); 368 if (ret) 369 goto out; 370 371 ret = generic_file_read_iter(iocb, iter); 372 out: 373 up_read(&file_inode(iocb->ki_filp)->i_rwsem); 374 return ret; 375 } 376 377 static ssize_t orangefs_file_write_iter(struct kiocb *iocb, 378 struct iov_iter *iter) 379 { 380 int ret; 381 orangefs_stats.writes++; 382 383 if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) { 384 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp)); 385 if (ret) 386 return ret; 387 } 388 389 ret = generic_file_write_iter(iocb, iter); 390 return ret; 391 } 392 393 static int orangefs_getflags(struct inode *inode, unsigned long *uval) 394 { 395 __u64 val = 0; 396 int ret; 397 398 ret = orangefs_inode_getxattr(inode, 399 "user.pvfs2.meta_hint", 400 &val, sizeof(val)); 401 if (ret < 0 && ret != -ENODATA) 402 return ret; 403 else if (ret == -ENODATA) 404 val = 0; 405 *uval = val; 406 return 0; 407 } 408 409 /* 410 * Perform a miscellaneous operation on a file. 411 */ 412 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg) 413 { 414 struct inode *inode = file_inode(file); 415 int ret = -ENOTTY; 416 __u64 val = 0; 417 unsigned long uval; 418 419 gossip_debug(GOSSIP_FILE_DEBUG, 420 "orangefs_ioctl: called with cmd %d\n", 421 cmd); 422 423 /* 424 * we understand some general ioctls on files, such as the immutable 425 * and append flags 426 */ 427 if (cmd == FS_IOC_GETFLAGS) { 428 ret = orangefs_getflags(inode, &uval); 429 if (ret) 430 return ret; 431 gossip_debug(GOSSIP_FILE_DEBUG, 432 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n", 433 (unsigned long long)uval); 434 return put_user(uval, (int __user *)arg); 435 } else if (cmd == FS_IOC_SETFLAGS) { 436 unsigned long old_uval; 437 438 ret = 0; 439 if (get_user(uval, (int __user *)arg)) 440 return -EFAULT; 441 /* 442 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode 443 * is turned on for a file. The user is not allowed to turn 444 * on this bit, but the bit is present if the user first gets 445 * the flags and then updates the flags with some new 446 * settings. So, we ignore it in the following edit. bligon. 447 */ 448 if ((uval & ~ORANGEFS_MIRROR_FL) & 449 (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) { 450 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n"); 451 return -EINVAL; 452 } 453 ret = orangefs_getflags(inode, &old_uval); 454 if (ret) 455 return ret; 456 ret = vfs_ioc_setflags_prepare(inode, old_uval, uval); 457 if (ret) 458 return ret; 459 val = uval; 460 gossip_debug(GOSSIP_FILE_DEBUG, 461 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n", 462 (unsigned long long)val); 463 ret = orangefs_inode_setxattr(inode, 464 "user.pvfs2.meta_hint", 465 &val, sizeof(val), 0); 466 } 467 468 return ret; 469 } 470 471 static vm_fault_t orangefs_fault(struct vm_fault *vmf) 472 { 473 struct file *file = vmf->vma->vm_file; 474 int ret; 475 ret = orangefs_inode_getattr(file->f_mapping->host, 476 ORANGEFS_GETATTR_SIZE); 477 if (ret == -ESTALE) 478 ret = -EIO; 479 if (ret) { 480 gossip_err("%s: orangefs_inode_getattr failed, " 481 "ret:%d:.\n", __func__, ret); 482 return VM_FAULT_SIGBUS; 483 } 484 return filemap_fault(vmf); 485 } 486 487 static const struct vm_operations_struct orangefs_file_vm_ops = { 488 .fault = orangefs_fault, 489 .map_pages = filemap_map_pages, 490 .page_mkwrite = orangefs_page_mkwrite, 491 }; 492 493 /* 494 * Memory map a region of a file. 495 */ 496 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma) 497 { 498 int ret; 499 500 ret = orangefs_revalidate_mapping(file_inode(file)); 501 if (ret) 502 return ret; 503 504 gossip_debug(GOSSIP_FILE_DEBUG, 505 "orangefs_file_mmap: called on %s\n", 506 (file ? 507 (char *)file->f_path.dentry->d_name.name : 508 (char *)"Unknown")); 509 510 /* set the sequential readahead hint */ 511 vma->vm_flags |= VM_SEQ_READ; 512 vma->vm_flags &= ~VM_RAND_READ; 513 514 file_accessed(file); 515 vma->vm_ops = &orangefs_file_vm_ops; 516 return 0; 517 } 518 519 #define mapping_nrpages(idata) ((idata)->nrpages) 520 521 /* 522 * Called to notify the module that there are no more references to 523 * this file (i.e. no processes have it open). 524 * 525 * \note Not called when each file is closed. 526 */ 527 static int orangefs_file_release(struct inode *inode, struct file *file) 528 { 529 gossip_debug(GOSSIP_FILE_DEBUG, 530 "orangefs_file_release: called on %pD\n", 531 file); 532 533 /* 534 * remove all associated inode pages from the page cache and 535 * readahead cache (if any); this forces an expensive refresh of 536 * data for the next caller of mmap (or 'get_block' accesses) 537 */ 538 if (file_inode(file) && 539 file_inode(file)->i_mapping && 540 mapping_nrpages(&file_inode(file)->i_data)) { 541 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) { 542 gossip_debug(GOSSIP_INODE_DEBUG, 543 "calling flush_racache on %pU\n", 544 get_khandle_from_ino(inode)); 545 flush_racache(inode); 546 gossip_debug(GOSSIP_INODE_DEBUG, 547 "flush_racache finished\n"); 548 } 549 550 } 551 return 0; 552 } 553 554 /* 555 * Push all data for a specific file onto permanent storage. 556 */ 557 static int orangefs_fsync(struct file *file, 558 loff_t start, 559 loff_t end, 560 int datasync) 561 { 562 int ret; 563 struct orangefs_inode_s *orangefs_inode = 564 ORANGEFS_I(file_inode(file)); 565 struct orangefs_kernel_op_s *new_op = NULL; 566 567 ret = filemap_write_and_wait_range(file_inode(file)->i_mapping, 568 start, end); 569 if (ret < 0) 570 return ret; 571 572 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC); 573 if (!new_op) 574 return -ENOMEM; 575 new_op->upcall.req.fsync.refn = orangefs_inode->refn; 576 577 ret = service_operation(new_op, 578 "orangefs_fsync", 579 get_interruptible_flag(file_inode(file))); 580 581 gossip_debug(GOSSIP_FILE_DEBUG, 582 "orangefs_fsync got return value of %d\n", 583 ret); 584 585 op_release(new_op); 586 return ret; 587 } 588 589 /* 590 * Change the file pointer position for an instance of an open file. 591 * 592 * \note If .llseek is overriden, we must acquire lock as described in 593 * Documentation/filesystems/locking.rst. 594 * 595 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would 596 * require much changes to the FS 597 */ 598 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin) 599 { 600 int ret = -EINVAL; 601 struct inode *inode = file_inode(file); 602 603 if (origin == SEEK_END) { 604 /* 605 * revalidate the inode's file size. 606 * NOTE: We are only interested in file size here, 607 * so we set mask accordingly. 608 */ 609 ret = orangefs_inode_getattr(file->f_mapping->host, 610 ORANGEFS_GETATTR_SIZE); 611 if (ret == -ESTALE) 612 ret = -EIO; 613 if (ret) { 614 gossip_debug(GOSSIP_FILE_DEBUG, 615 "%s:%s:%d calling make bad inode\n", 616 __FILE__, 617 __func__, 618 __LINE__); 619 return ret; 620 } 621 } 622 623 gossip_debug(GOSSIP_FILE_DEBUG, 624 "orangefs_file_llseek: offset is %ld | origin is %d" 625 " | inode size is %lu\n", 626 (long)offset, 627 origin, 628 (unsigned long)i_size_read(inode)); 629 630 return generic_file_llseek(file, offset, origin); 631 } 632 633 /* 634 * Support local locks (locks that only this kernel knows about) 635 * if Orangefs was mounted -o local_lock. 636 */ 637 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl) 638 { 639 int rc = -EINVAL; 640 641 if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) { 642 if (cmd == F_GETLK) { 643 rc = 0; 644 posix_test_lock(filp, fl); 645 } else { 646 rc = posix_lock_file(filp, fl, NULL); 647 } 648 } 649 650 return rc; 651 } 652 653 static int orangefs_file_open(struct inode * inode, struct file *file) 654 { 655 file->private_data = NULL; 656 return generic_file_open(inode, file); 657 } 658 659 static int orangefs_flush(struct file *file, fl_owner_t id) 660 { 661 /* 662 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the 663 * service_operation in orangefs_fsync. 664 * 665 * Do not send fsync to OrangeFS server on a close. Do send fsync 666 * on an explicit fsync call. This duplicates historical OrangeFS 667 * behavior. 668 */ 669 struct inode *inode = file->f_mapping->host; 670 int r; 671 672 kfree(file->private_data); 673 file->private_data = NULL; 674 675 if (inode->i_state & I_DIRTY_TIME) { 676 spin_lock(&inode->i_lock); 677 inode->i_state &= ~I_DIRTY_TIME; 678 spin_unlock(&inode->i_lock); 679 mark_inode_dirty_sync(inode); 680 } 681 682 r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX); 683 if (r > 0) 684 return 0; 685 else 686 return r; 687 } 688 689 /** ORANGEFS implementation of VFS file operations */ 690 const struct file_operations orangefs_file_operations = { 691 .llseek = orangefs_file_llseek, 692 .read_iter = orangefs_file_read_iter, 693 .write_iter = orangefs_file_write_iter, 694 .lock = orangefs_lock, 695 .unlocked_ioctl = orangefs_ioctl, 696 .mmap = orangefs_file_mmap, 697 .open = orangefs_file_open, 698 .flush = orangefs_flush, 699 .release = orangefs_file_release, 700 .fsync = orangefs_fsync, 701 }; 702