1 // SPDX-License-Identifier: GPL-2.0
3 * (C) 2001 Clemson University and The University of Chicago
4 * Copyright 2018 Omnibond Systems, L.L.C.
6 * See COPYING in top-level directory.
10 * Linux VFS file operations.
14 #include "orangefs-kernel.h"
15 #include "orangefs-bufmap.h"
17 #include <linux/pagemap.h>
19 static int flush_racache(struct inode *inode)
21 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
22 struct orangefs_kernel_op_s *new_op;
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);
30 new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
33 new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
35 ret = service_operation(new_op, "orangefs_flush_racache",
36 get_interruptible_flag(inode));
38 gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
46 * Post and wait for the I/O upcall to finish
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)
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;
62 new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
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;
71 populate_shared_memory:
72 /* get a shared buffer index */
73 buffer_index = orangefs_bufmap_get();
74 if (buffer_index < 0) {
76 gossip_debug(GOSSIP_FILE_DEBUG,
77 "%s: orangefs_bufmap_get failure (%zd)\n",
81 gossip_debug(GOSSIP_FILE_DEBUG,
82 "%s(%pU): GET op %p -> buffer_index %d\n",
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);
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.
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.
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.
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.
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?
118 open_for_write = file->f_mode & FMODE_WRITE;
119 open_for_read = file->f_mode & FMODE_READ;
122 open_for_read = 0; /* not relevant? */
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;
129 gossip_debug(GOSSIP_FILE_DEBUG,
130 "%s(%pU): offset: %llu total_size: %zd\n",
136 * Stage 1: copy the buffers into client-core's address space
138 if (type == ORANGEFS_IO_WRITE && total_size) {
139 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
142 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
143 __func__, (long)ret);
148 gossip_debug(GOSSIP_FILE_DEBUG,
149 "%s(%pU): Calling post_io_request with tag (%llu)\n",
154 /* Stage 2: Service the I/O operation */
155 ret = service_operation(new_op,
156 type == ORANGEFS_IO_WRITE ?
159 get_interruptible_flag(inode));
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.
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",
177 goto populate_shared_memory;
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.
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
193 switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
195 * If the op was waiting when the interrupt
196 * occurred, then the client-core did not
199 case OP_VFS_STATE_WAITING:
206 * If the op was in progress when the interrupt
207 * occurred, then the client-core was able to
210 case OP_VFS_STATE_INPROGR:
211 if (type == ORANGEFS_IO_READ)
217 gossip_err("%s: unexpected op state :%d:.\n",
223 gossip_debug(GOSSIP_FILE_DEBUG,
224 "%s: got EINTR, state:%d: %p\n",
229 gossip_err("%s: error in %s handle %pU, returning %zd\n",
231 type == ORANGEFS_IO_READ ?
232 "read from" : "write to",
235 if (orangefs_cancel_op_in_progress(new_op))
242 * Stage 3: Post copy buffers from client-core's address space
244 if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
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
252 * When reading, readahead_size will only be zero when
253 * we're doing O_DIRECT, otherwise we got here from
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.
262 min(new_op->downcall.resp.io.amt_complete,
265 copy_amount = new_op->downcall.resp.io.amt_complete;
267 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
270 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
271 __func__, (long)ret);
275 gossip_debug(GOSSIP_FILE_DEBUG,
276 "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
279 type == ORANGEFS_IO_READ ? "read" : "written",
280 (int)new_op->downcall.resp.io.amt_complete);
282 ret = new_op->downcall.resp.io.amt_complete;
285 if (buffer_index >= 0) {
286 if ((readahead_size) && (type == ORANGEFS_IO_READ)) {
288 *index_return = buffer_index;
289 gossip_debug(GOSSIP_FILE_DEBUG,
290 "%s: hold on to buffer_index :%d:\n",
291 __func__, buffer_index);
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);
304 int orangefs_revalidate_mapping(struct inode *inode)
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;
312 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
315 spin_lock(&inode->i_lock);
316 if (test_bit(1, bitlock)) {
317 spin_unlock(&inode->i_lock);
320 if (!time_before(jiffies, orangefs_inode->mapping_time))
322 spin_unlock(&inode->i_lock);
328 spin_unlock(&inode->i_lock);
330 unmap_mapping_range(mapping, 0, 0, 0);
331 ret = filemap_write_and_wait(mapping);
333 ret = invalidate_inode_pages2(mapping);
335 orangefs_inode->mapping_time = jiffies +
336 orangefs_cache_timeout_msecs*HZ/1000;
338 clear_bit(1, bitlock);
339 smp_mb__after_atomic();
340 wake_up_bit(bitlock, 1);
345 static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
346 struct iov_iter *iter)
349 orangefs_stats.reads++;
351 down_read(&file_inode(iocb->ki_filp)->i_rwsem);
352 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
356 ret = generic_file_read_iter(iocb, iter);
358 up_read(&file_inode(iocb->ki_filp)->i_rwsem);
362 static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
363 struct iov_iter *iter)
366 orangefs_stats.writes++;
368 if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
369 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
374 ret = generic_file_write_iter(iocb, iter);
378 static int orangefs_getflags(struct inode *inode, unsigned long *uval)
383 ret = orangefs_inode_getxattr(inode,
384 "user.pvfs2.meta_hint",
386 if (ret < 0 && ret != -ENODATA)
388 else if (ret == -ENODATA)
395 * Perform a miscellaneous operation on a file.
397 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
399 struct inode *inode = file_inode(file);
404 gossip_debug(GOSSIP_FILE_DEBUG,
405 "orangefs_ioctl: called with cmd %d\n",
409 * we understand some general ioctls on files, such as the immutable
412 if (cmd == FS_IOC_GETFLAGS) {
413 ret = orangefs_getflags(inode, &uval);
416 gossip_debug(GOSSIP_FILE_DEBUG,
417 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
418 (unsigned long long)uval);
419 return put_user(uval, (int __user *)arg);
420 } else if (cmd == FS_IOC_SETFLAGS) {
421 unsigned long old_uval;
424 if (get_user(uval, (int __user *)arg))
427 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
428 * is turned on for a file. The user is not allowed to turn
429 * on this bit, but the bit is present if the user first gets
430 * the flags and then updates the flags with some new
431 * settings. So, we ignore it in the following edit. bligon.
433 if ((uval & ~ORANGEFS_MIRROR_FL) &
434 (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
435 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
438 ret = orangefs_getflags(inode, &old_uval);
441 ret = vfs_ioc_setflags_prepare(inode, old_uval, uval);
445 gossip_debug(GOSSIP_FILE_DEBUG,
446 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
447 (unsigned long long)val);
448 ret = orangefs_inode_setxattr(inode,
449 "user.pvfs2.meta_hint",
450 &val, sizeof(val), 0);
456 static vm_fault_t orangefs_fault(struct vm_fault *vmf)
458 struct file *file = vmf->vma->vm_file;
460 ret = orangefs_inode_getattr(file->f_mapping->host,
461 ORANGEFS_GETATTR_SIZE);
465 gossip_err("%s: orangefs_inode_getattr failed, "
466 "ret:%d:.\n", __func__, ret);
467 return VM_FAULT_SIGBUS;
469 return filemap_fault(vmf);
472 static const struct vm_operations_struct orangefs_file_vm_ops = {
473 .fault = orangefs_fault,
474 .map_pages = filemap_map_pages,
475 .page_mkwrite = orangefs_page_mkwrite,
479 * Memory map a region of a file.
481 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
485 ret = orangefs_revalidate_mapping(file_inode(file));
489 gossip_debug(GOSSIP_FILE_DEBUG,
490 "orangefs_file_mmap: called on %s\n",
492 (char *)file->f_path.dentry->d_name.name :
495 /* set the sequential readahead hint */
496 vma->vm_flags |= VM_SEQ_READ;
497 vma->vm_flags &= ~VM_RAND_READ;
500 vma->vm_ops = &orangefs_file_vm_ops;
504 #define mapping_nrpages(idata) ((idata)->nrpages)
507 * Called to notify the module that there are no more references to
508 * this file (i.e. no processes have it open).
510 * \note Not called when each file is closed.
512 static int orangefs_file_release(struct inode *inode, struct file *file)
514 gossip_debug(GOSSIP_FILE_DEBUG,
515 "orangefs_file_release: called on %pD\n",
519 * remove all associated inode pages from the page cache and
520 * readahead cache (if any); this forces an expensive refresh of
521 * data for the next caller of mmap (or 'get_block' accesses)
523 if (file_inode(file) &&
524 file_inode(file)->i_mapping &&
525 mapping_nrpages(&file_inode(file)->i_data)) {
526 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
527 gossip_debug(GOSSIP_INODE_DEBUG,
528 "calling flush_racache on %pU\n",
529 get_khandle_from_ino(inode));
530 flush_racache(inode);
531 gossip_debug(GOSSIP_INODE_DEBUG,
532 "flush_racache finished\n");
540 * Push all data for a specific file onto permanent storage.
542 static int orangefs_fsync(struct file *file,
548 struct orangefs_inode_s *orangefs_inode =
549 ORANGEFS_I(file_inode(file));
550 struct orangefs_kernel_op_s *new_op = NULL;
552 ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
557 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
560 new_op->upcall.req.fsync.refn = orangefs_inode->refn;
562 ret = service_operation(new_op,
564 get_interruptible_flag(file_inode(file)));
566 gossip_debug(GOSSIP_FILE_DEBUG,
567 "orangefs_fsync got return value of %d\n",
575 * Change the file pointer position for an instance of an open file.
577 * \note If .llseek is overriden, we must acquire lock as described in
578 * Documentation/filesystems/locking.rst.
580 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
581 * require much changes to the FS
583 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
586 struct inode *inode = file_inode(file);
588 if (origin == SEEK_END) {
590 * revalidate the inode's file size.
591 * NOTE: We are only interested in file size here,
592 * so we set mask accordingly.
594 ret = orangefs_inode_getattr(file->f_mapping->host,
595 ORANGEFS_GETATTR_SIZE);
599 gossip_debug(GOSSIP_FILE_DEBUG,
600 "%s:%s:%d calling make bad inode\n",
608 gossip_debug(GOSSIP_FILE_DEBUG,
609 "orangefs_file_llseek: offset is %ld | origin is %d"
610 " | inode size is %lu\n",
613 (unsigned long)i_size_read(inode));
615 return generic_file_llseek(file, offset, origin);
619 * Support local locks (locks that only this kernel knows about)
620 * if Orangefs was mounted -o local_lock.
622 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
626 if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
627 if (cmd == F_GETLK) {
629 posix_test_lock(filp, fl);
631 rc = posix_lock_file(filp, fl, NULL);
638 static int orangefs_flush(struct file *file, fl_owner_t id)
641 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
642 * service_operation in orangefs_fsync.
644 * Do not send fsync to OrangeFS server on a close. Do send fsync
645 * on an explicit fsync call. This duplicates historical OrangeFS
650 r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
657 /** ORANGEFS implementation of VFS file operations */
658 const struct file_operations orangefs_file_operations = {
659 .llseek = orangefs_file_llseek,
660 .read_iter = orangefs_file_read_iter,
661 .write_iter = orangefs_file_write_iter,
662 .lock = orangefs_lock,
663 .unlocked_ioctl = orangefs_ioctl,
664 .mmap = orangefs_file_mmap,
665 .open = generic_file_open,
666 .flush = orangefs_flush,
667 .release = orangefs_file_release,
668 .fsync = orangefs_fsync,