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, int *index_return)
52 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
53 struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
54 struct orangefs_kernel_op_s *new_op = NULL;
55 int buffer_index = -1;
59 new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
64 new_op->upcall.req.io.readahead_size = readahead_size;
65 new_op->upcall.req.io.io_type = type;
66 new_op->upcall.req.io.refn = orangefs_inode->refn;
68 populate_shared_memory:
69 /* get a shared buffer index */
70 buffer_index = orangefs_bufmap_get();
71 if (buffer_index < 0) {
73 gossip_debug(GOSSIP_FILE_DEBUG,
74 "%s: orangefs_bufmap_get failure (%zd)\n",
78 gossip_debug(GOSSIP_FILE_DEBUG,
79 "%s(%pU): GET op %p -> buffer_index %d\n",
85 new_op->uses_shared_memory = 1;
86 new_op->upcall.req.io.buf_index = buffer_index;
87 new_op->upcall.req.io.count = total_size;
88 new_op->upcall.req.io.offset = *offset;
89 if (type == ORANGEFS_IO_WRITE && wr) {
90 new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
91 new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
94 gossip_debug(GOSSIP_FILE_DEBUG,
95 "%s(%pU): offset: %llu total_size: %zd\n",
101 * Stage 1: copy the buffers into client-core's address space
103 if (type == ORANGEFS_IO_WRITE && total_size) {
104 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
107 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
108 __func__, (long)ret);
113 gossip_debug(GOSSIP_FILE_DEBUG,
114 "%s(%pU): Calling post_io_request with tag (%llu)\n",
119 /* Stage 2: Service the I/O operation */
120 ret = service_operation(new_op,
121 type == ORANGEFS_IO_WRITE ?
124 get_interruptible_flag(inode));
127 * If service_operation() returns -EAGAIN #and# the operation was
128 * purged from orangefs_request_list or htable_ops_in_progress, then
129 * we know that the client was restarted, causing the shared memory
130 * area to be wiped clean. To restart a write operation in this
131 * case, we must re-copy the data from the user's iovec to a NEW
132 * shared memory location. To restart a read operation, we must get
133 * a new shared memory location.
135 if (ret == -EAGAIN && op_state_purged(new_op)) {
136 orangefs_bufmap_put(buffer_index);
138 if (type == ORANGEFS_IO_WRITE)
139 iov_iter_revert(iter, total_size);
140 gossip_debug(GOSSIP_FILE_DEBUG,
141 "%s:going to repopulate_shared_memory.\n",
143 goto populate_shared_memory;
149 * We can't return EINTR if any data was written,
150 * it's not POSIX. It is minimally acceptable
151 * to give a partial write, the way NFS does.
153 * It would be optimal to return all or nothing,
154 * but if a userspace write is bigger than
155 * an IO buffer, and the interrupt occurs
156 * between buffer writes, that would not be
159 switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
161 * If the op was waiting when the interrupt
162 * occurred, then the client-core did not
165 case OP_VFS_STATE_WAITING:
172 * If the op was in progress when the interrupt
173 * occurred, then the client-core was able to
176 case OP_VFS_STATE_INPROGR:
177 if (type == ORANGEFS_IO_READ)
183 gossip_err("%s: unexpected op state :%d:.\n",
189 gossip_debug(GOSSIP_FILE_DEBUG,
190 "%s: got EINTR, state:%d: %p\n",
195 gossip_err("%s: error in %s handle %pU, returning %zd\n",
197 type == ORANGEFS_IO_READ ?
198 "read from" : "write to",
201 if (orangefs_cancel_op_in_progress(new_op))
208 * Stage 3: Post copy buffers from client-core's address space
210 if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
212 * NOTE: the iovector can either contain addresses which
213 * can futher be kernel-space or user-space addresses.
214 * or it can pointers to struct page's
218 * When reading, readahead_size will only be zero when
219 * we're doing O_DIRECT, otherwise we got here from
222 * If we got here from orangefs_readpage we want to
223 * copy either a page or the whole file into the io
224 * vector, whichever is smaller.
228 min(new_op->downcall.resp.io.amt_complete,
231 copy_amount = new_op->downcall.resp.io.amt_complete;
233 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
236 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
237 __func__, (long)ret);
241 gossip_debug(GOSSIP_FILE_DEBUG,
242 "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
245 type == ORANGEFS_IO_READ ? "read" : "written",
246 (int)new_op->downcall.resp.io.amt_complete);
248 ret = new_op->downcall.resp.io.amt_complete;
251 if (buffer_index >= 0) {
252 if ((readahead_size) && (type == ORANGEFS_IO_READ)) {
254 *index_return = buffer_index;
255 gossip_debug(GOSSIP_FILE_DEBUG,
256 "%s: hold on to buffer_index :%d:\n",
257 __func__, buffer_index);
260 orangefs_bufmap_put(buffer_index);
261 gossip_debug(GOSSIP_FILE_DEBUG,
262 "%s(%pU): PUT buffer_index %d\n",
263 __func__, handle, buffer_index);
271 int orangefs_revalidate_mapping(struct inode *inode)
273 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
274 struct address_space *mapping = inode->i_mapping;
275 unsigned long *bitlock = &orangefs_inode->bitlock;
279 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
282 spin_lock(&inode->i_lock);
283 if (test_bit(1, bitlock)) {
284 spin_unlock(&inode->i_lock);
287 if (!time_before(jiffies, orangefs_inode->mapping_time))
289 spin_unlock(&inode->i_lock);
295 spin_unlock(&inode->i_lock);
297 unmap_mapping_range(mapping, 0, 0, 0);
298 ret = filemap_write_and_wait(mapping);
300 ret = invalidate_inode_pages2(mapping);
302 orangefs_inode->mapping_time = jiffies +
303 orangefs_cache_timeout_msecs*HZ/1000;
305 clear_bit(1, bitlock);
306 smp_mb__after_atomic();
307 wake_up_bit(bitlock, 1);
312 static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
313 struct iov_iter *iter)
316 struct orangefs_read_options *ro;
318 orangefs_stats.reads++;
321 * Remember how they set "count" in read(2) or pread(2) or whatever -
322 * users can use count as a knob to control orangefs io size and later
323 * we can try to help them fill as many pages as possible in readpage.
325 if (!iocb->ki_filp->private_data) {
326 iocb->ki_filp->private_data = kmalloc(sizeof *ro, GFP_KERNEL);
327 if (!iocb->ki_filp->private_data)
329 ro = iocb->ki_filp->private_data;
330 ro->blksiz = iter->count;
333 down_read(&file_inode(iocb->ki_filp)->i_rwsem);
334 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
338 ret = generic_file_read_iter(iocb, iter);
340 up_read(&file_inode(iocb->ki_filp)->i_rwsem);
344 static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
345 struct iov_iter *iter)
348 orangefs_stats.writes++;
350 if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
351 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
356 ret = generic_file_write_iter(iocb, iter);
360 static int orangefs_getflags(struct inode *inode, unsigned long *uval)
365 ret = orangefs_inode_getxattr(inode,
366 "user.pvfs2.meta_hint",
368 if (ret < 0 && ret != -ENODATA)
370 else if (ret == -ENODATA)
377 * Perform a miscellaneous operation on a file.
379 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
381 struct inode *inode = file_inode(file);
386 gossip_debug(GOSSIP_FILE_DEBUG,
387 "orangefs_ioctl: called with cmd %d\n",
391 * we understand some general ioctls on files, such as the immutable
394 if (cmd == FS_IOC_GETFLAGS) {
395 ret = orangefs_getflags(inode, &uval);
398 gossip_debug(GOSSIP_FILE_DEBUG,
399 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
400 (unsigned long long)uval);
401 return put_user(uval, (int __user *)arg);
402 } else if (cmd == FS_IOC_SETFLAGS) {
403 unsigned long old_uval;
406 if (get_user(uval, (int __user *)arg))
409 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
410 * is turned on for a file. The user is not allowed to turn
411 * on this bit, but the bit is present if the user first gets
412 * the flags and then updates the flags with some new
413 * settings. So, we ignore it in the following edit. bligon.
415 if ((uval & ~ORANGEFS_MIRROR_FL) &
416 (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
417 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
420 ret = orangefs_getflags(inode, &old_uval);
423 ret = vfs_ioc_setflags_prepare(inode, old_uval, uval);
427 gossip_debug(GOSSIP_FILE_DEBUG,
428 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
429 (unsigned long long)val);
430 ret = orangefs_inode_setxattr(inode,
431 "user.pvfs2.meta_hint",
432 &val, sizeof(val), 0);
438 static vm_fault_t orangefs_fault(struct vm_fault *vmf)
440 struct file *file = vmf->vma->vm_file;
442 ret = orangefs_inode_getattr(file->f_mapping->host,
443 ORANGEFS_GETATTR_SIZE);
447 gossip_err("%s: orangefs_inode_getattr failed, "
448 "ret:%d:.\n", __func__, ret);
449 return VM_FAULT_SIGBUS;
451 return filemap_fault(vmf);
454 static const struct vm_operations_struct orangefs_file_vm_ops = {
455 .fault = orangefs_fault,
456 .map_pages = filemap_map_pages,
457 .page_mkwrite = orangefs_page_mkwrite,
461 * Memory map a region of a file.
463 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
467 ret = orangefs_revalidate_mapping(file_inode(file));
471 gossip_debug(GOSSIP_FILE_DEBUG,
472 "orangefs_file_mmap: called on %s\n",
474 (char *)file->f_path.dentry->d_name.name :
477 /* set the sequential readahead hint */
478 vma->vm_flags |= VM_SEQ_READ;
479 vma->vm_flags &= ~VM_RAND_READ;
482 vma->vm_ops = &orangefs_file_vm_ops;
486 #define mapping_nrpages(idata) ((idata)->nrpages)
489 * Called to notify the module that there are no more references to
490 * this file (i.e. no processes have it open).
492 * \note Not called when each file is closed.
494 static int orangefs_file_release(struct inode *inode, struct file *file)
496 gossip_debug(GOSSIP_FILE_DEBUG,
497 "orangefs_file_release: called on %pD\n",
501 * remove all associated inode pages from the page cache and
502 * readahead cache (if any); this forces an expensive refresh of
503 * data for the next caller of mmap (or 'get_block' accesses)
505 if (file_inode(file) &&
506 file_inode(file)->i_mapping &&
507 mapping_nrpages(&file_inode(file)->i_data)) {
508 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
509 gossip_debug(GOSSIP_INODE_DEBUG,
510 "calling flush_racache on %pU\n",
511 get_khandle_from_ino(inode));
512 flush_racache(inode);
513 gossip_debug(GOSSIP_INODE_DEBUG,
514 "flush_racache finished\n");
522 * Push all data for a specific file onto permanent storage.
524 static int orangefs_fsync(struct file *file,
530 struct orangefs_inode_s *orangefs_inode =
531 ORANGEFS_I(file_inode(file));
532 struct orangefs_kernel_op_s *new_op = NULL;
534 ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
539 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
542 new_op->upcall.req.fsync.refn = orangefs_inode->refn;
544 ret = service_operation(new_op,
546 get_interruptible_flag(file_inode(file)));
548 gossip_debug(GOSSIP_FILE_DEBUG,
549 "orangefs_fsync got return value of %d\n",
557 * Change the file pointer position for an instance of an open file.
559 * \note If .llseek is overriden, we must acquire lock as described in
560 * Documentation/filesystems/Locking.
562 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
563 * require much changes to the FS
565 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
568 struct inode *inode = file_inode(file);
570 if (origin == SEEK_END) {
572 * revalidate the inode's file size.
573 * NOTE: We are only interested in file size here,
574 * so we set mask accordingly.
576 ret = orangefs_inode_getattr(file->f_mapping->host,
577 ORANGEFS_GETATTR_SIZE);
581 gossip_debug(GOSSIP_FILE_DEBUG,
582 "%s:%s:%d calling make bad inode\n",
590 gossip_debug(GOSSIP_FILE_DEBUG,
591 "orangefs_file_llseek: offset is %ld | origin is %d"
592 " | inode size is %lu\n",
595 (unsigned long)i_size_read(inode));
597 return generic_file_llseek(file, offset, origin);
601 * Support local locks (locks that only this kernel knows about)
602 * if Orangefs was mounted -o local_lock.
604 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
608 if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
609 if (cmd == F_GETLK) {
611 posix_test_lock(filp, fl);
613 rc = posix_lock_file(filp, fl, NULL);
620 static int orangefs_file_open(struct inode * inode, struct file *file)
622 file->private_data = NULL;
623 return generic_file_open(inode, file);
626 static int orangefs_flush(struct file *file, fl_owner_t id)
629 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
630 * service_operation in orangefs_fsync.
632 * Do not send fsync to OrangeFS server on a close. Do send fsync
633 * on an explicit fsync call. This duplicates historical OrangeFS
636 struct inode *inode = file->f_mapping->host;
639 kfree(file->private_data);
640 file->private_data = NULL;
642 if (inode->i_state & I_DIRTY_TIME) {
643 spin_lock(&inode->i_lock);
644 inode->i_state &= ~I_DIRTY_TIME;
645 spin_unlock(&inode->i_lock);
646 mark_inode_dirty_sync(inode);
649 r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
656 /** ORANGEFS implementation of VFS file operations */
657 const struct file_operations orangefs_file_operations = {
658 .llseek = orangefs_file_llseek,
659 .read_iter = orangefs_file_read_iter,
660 .write_iter = orangefs_file_write_iter,
661 .lock = orangefs_lock,
662 .unlocked_ioctl = orangefs_ioctl,
663 .mmap = orangefs_file_mmap,
664 .open = orangefs_file_open,
665 .flush = orangefs_flush,
666 .release = orangefs_file_release,
667 .fsync = orangefs_fsync,