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
flush_racache(struct inode * inode)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 */
wait_for_direct_io(enum ORANGEFS_io_type type,struct inode * inode,loff_t * offset,struct iov_iter * iter,size_t total_size,loff_t readahead_size,struct orangefs_write_range * wr,int * index_return)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)
51 {
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;
56 ssize_t ret;
57 size_t copy_amount;
58
59 new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
60 if (!new_op)
61 return -ENOMEM;
62
63 /* synchronous I/O */
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;
67
68 populate_shared_memory:
69 /* get a shared buffer index */
70 buffer_index = orangefs_bufmap_get();
71 if (buffer_index < 0) {
72 ret = buffer_index;
73 gossip_debug(GOSSIP_FILE_DEBUG,
74 "%s: orangefs_bufmap_get failure (%zd)\n",
75 __func__, ret);
76 goto out;
77 }
78 gossip_debug(GOSSIP_FILE_DEBUG,
79 "%s(%pU): GET op %p -> buffer_index %d\n",
80 __func__,
81 handle,
82 new_op,
83 buffer_index);
84
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);
92 }
93
94 gossip_debug(GOSSIP_FILE_DEBUG,
95 "%s(%pU): offset: %llu total_size: %zd\n",
96 __func__,
97 handle,
98 llu(*offset),
99 total_size);
100 /*
101 * Stage 1: copy the buffers into client-core's address space
102 */
103 if (type == ORANGEFS_IO_WRITE && total_size) {
104 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
105 total_size);
106 if (ret < 0) {
107 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
108 __func__, (long)ret);
109 goto out;
110 }
111 }
112
113 gossip_debug(GOSSIP_FILE_DEBUG,
114 "%s(%pU): Calling post_io_request with tag (%llu)\n",
115 __func__,
116 handle,
117 llu(new_op->tag));
118
119 /* Stage 2: Service the I/O operation */
120 ret = service_operation(new_op,
121 type == ORANGEFS_IO_WRITE ?
122 "file_write" :
123 "file_read",
124 get_interruptible_flag(inode));
125
126 /*
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.
134 */
135 if (ret == -EAGAIN && op_state_purged(new_op)) {
136 orangefs_bufmap_put(buffer_index);
137 if (type == ORANGEFS_IO_WRITE)
138 iov_iter_revert(iter, total_size);
139 gossip_debug(GOSSIP_FILE_DEBUG,
140 "%s:going to repopulate_shared_memory.\n",
141 __func__);
142 goto populate_shared_memory;
143 }
144
145 if (ret < 0) {
146 if (ret == -EINTR) {
147 /*
148 * We can't return EINTR if any data was written,
149 * it's not POSIX. It is minimally acceptable
150 * to give a partial write, the way NFS does.
151 *
152 * It would be optimal to return all or nothing,
153 * but if a userspace write is bigger than
154 * an IO buffer, and the interrupt occurs
155 * between buffer writes, that would not be
156 * possible.
157 */
158 switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
159 /*
160 * If the op was waiting when the interrupt
161 * occurred, then the client-core did not
162 * trigger the write.
163 */
164 case OP_VFS_STATE_WAITING:
165 if (*offset == 0)
166 ret = -EINTR;
167 else
168 ret = 0;
169 break;
170 /*
171 * If the op was in progress when the interrupt
172 * occurred, then the client-core was able to
173 * trigger the write.
174 */
175 case OP_VFS_STATE_INPROGR:
176 if (type == ORANGEFS_IO_READ)
177 ret = -EINTR;
178 else
179 ret = total_size;
180 break;
181 default:
182 gossip_err("%s: unexpected op state :%d:.\n",
183 __func__,
184 new_op->op_state);
185 ret = 0;
186 break;
187 }
188 gossip_debug(GOSSIP_FILE_DEBUG,
189 "%s: got EINTR, state:%d: %p\n",
190 __func__,
191 new_op->op_state,
192 new_op);
193 } else {
194 gossip_err("%s: error in %s handle %pU, returning %zd\n",
195 __func__,
196 type == ORANGEFS_IO_READ ?
197 "read from" : "write to",
198 handle, ret);
199 }
200 if (orangefs_cancel_op_in_progress(new_op))
201 return ret;
202
203 goto out;
204 }
205
206 /*
207 * Stage 3: Post copy buffers from client-core's address space
208 */
209 if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
210 /*
211 * NOTE: the iovector can either contain addresses which
212 * can futher be kernel-space or user-space addresses.
213 * or it can pointers to struct page's
214 */
215
216 /*
217 * When reading, readahead_size will only be zero when
218 * we're doing O_DIRECT, otherwise we got here from
219 * orangefs_readpage.
220 *
221 * If we got here from orangefs_readpage we want to
222 * copy either a page or the whole file into the io
223 * vector, whichever is smaller.
224 */
225 if (readahead_size)
226 copy_amount =
227 min(new_op->downcall.resp.io.amt_complete,
228 (__s64)PAGE_SIZE);
229 else
230 copy_amount = new_op->downcall.resp.io.amt_complete;
231
232 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
233 copy_amount);
234 if (ret < 0) {
235 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
236 __func__, (long)ret);
237 goto out;
238 }
239 }
240 gossip_debug(GOSSIP_FILE_DEBUG,
241 "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
242 __func__,
243 handle,
244 type == ORANGEFS_IO_READ ? "read" : "written",
245 (int)new_op->downcall.resp.io.amt_complete);
246
247 ret = new_op->downcall.resp.io.amt_complete;
248
249 out:
250 if (buffer_index >= 0) {
251 if ((readahead_size) && (type == ORANGEFS_IO_READ)) {
252 /* readpage */
253 *index_return = buffer_index;
254 gossip_debug(GOSSIP_FILE_DEBUG,
255 "%s: hold on to buffer_index :%d:\n",
256 __func__, buffer_index);
257 } else {
258 /* O_DIRECT */
259 orangefs_bufmap_put(buffer_index);
260 gossip_debug(GOSSIP_FILE_DEBUG,
261 "%s(%pU): PUT buffer_index %d\n",
262 __func__, handle, buffer_index);
263 }
264 }
265 op_release(new_op);
266 return ret;
267 }
268
orangefs_revalidate_mapping(struct inode * inode)269 int orangefs_revalidate_mapping(struct inode *inode)
270 {
271 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
272 struct address_space *mapping = inode->i_mapping;
273 unsigned long *bitlock = &orangefs_inode->bitlock;
274 int ret;
275
276 while (1) {
277 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
278 if (ret)
279 return ret;
280 spin_lock(&inode->i_lock);
281 if (test_bit(1, bitlock)) {
282 spin_unlock(&inode->i_lock);
283 continue;
284 }
285 if (!time_before(jiffies, orangefs_inode->mapping_time))
286 break;
287 spin_unlock(&inode->i_lock);
288 return 0;
289 }
290
291 set_bit(1, bitlock);
292 smp_wmb();
293 spin_unlock(&inode->i_lock);
294
295 unmap_mapping_range(mapping, 0, 0, 0);
296 ret = filemap_write_and_wait(mapping);
297 if (!ret)
298 ret = invalidate_inode_pages2(mapping);
299
300 orangefs_inode->mapping_time = jiffies +
301 orangefs_cache_timeout_msecs*HZ/1000;
302
303 clear_bit(1, bitlock);
304 smp_mb__after_atomic();
305 wake_up_bit(bitlock, 1);
306
307 return ret;
308 }
309
orangefs_file_read_iter(struct kiocb * iocb,struct iov_iter * iter)310 static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
311 struct iov_iter *iter)
312 {
313 int ret;
314 struct orangefs_read_options *ro;
315
316 orangefs_stats.reads++;
317
318 /*
319 * Remember how they set "count" in read(2) or pread(2) or whatever -
320 * users can use count as a knob to control orangefs io size and later
321 * we can try to help them fill as many pages as possible in readpage.
322 */
323 if (!iocb->ki_filp->private_data) {
324 iocb->ki_filp->private_data = kmalloc(sizeof *ro, GFP_KERNEL);
325 if (!iocb->ki_filp->private_data)
326 return(ENOMEM);
327 ro = iocb->ki_filp->private_data;
328 ro->blksiz = iter->count;
329 }
330
331 down_read(&file_inode(iocb->ki_filp)->i_rwsem);
332 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
333 if (ret)
334 goto out;
335
336 ret = generic_file_read_iter(iocb, iter);
337 out:
338 up_read(&file_inode(iocb->ki_filp)->i_rwsem);
339 return ret;
340 }
341
orangefs_file_write_iter(struct kiocb * iocb,struct iov_iter * iter)342 static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
343 struct iov_iter *iter)
344 {
345 int ret;
346 orangefs_stats.writes++;
347
348 if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
349 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
350 if (ret)
351 return ret;
352 }
353
354 ret = generic_file_write_iter(iocb, iter);
355 return ret;
356 }
357
orangefs_getflags(struct inode * inode,unsigned long * uval)358 static int orangefs_getflags(struct inode *inode, unsigned long *uval)
359 {
360 __u64 val = 0;
361 int ret;
362
363 ret = orangefs_inode_getxattr(inode,
364 "user.pvfs2.meta_hint",
365 &val, sizeof(val));
366 if (ret < 0 && ret != -ENODATA)
367 return ret;
368 else if (ret == -ENODATA)
369 val = 0;
370 *uval = val;
371 return 0;
372 }
373
374 /*
375 * Perform a miscellaneous operation on a file.
376 */
orangefs_ioctl(struct file * file,unsigned int cmd,unsigned long arg)377 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
378 {
379 struct inode *inode = file_inode(file);
380 int ret = -ENOTTY;
381 __u64 val = 0;
382 unsigned long uval;
383
384 gossip_debug(GOSSIP_FILE_DEBUG,
385 "orangefs_ioctl: called with cmd %d\n",
386 cmd);
387
388 /*
389 * we understand some general ioctls on files, such as the immutable
390 * and append flags
391 */
392 if (cmd == FS_IOC_GETFLAGS) {
393 ret = orangefs_getflags(inode, &uval);
394 if (ret)
395 return ret;
396 gossip_debug(GOSSIP_FILE_DEBUG,
397 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
398 (unsigned long long)uval);
399 return put_user(uval, (int __user *)arg);
400 } else if (cmd == FS_IOC_SETFLAGS) {
401 unsigned long old_uval;
402
403 ret = 0;
404 if (get_user(uval, (int __user *)arg))
405 return -EFAULT;
406 /*
407 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
408 * is turned on for a file. The user is not allowed to turn
409 * on this bit, but the bit is present if the user first gets
410 * the flags and then updates the flags with some new
411 * settings. So, we ignore it in the following edit. bligon.
412 */
413 if ((uval & ~ORANGEFS_MIRROR_FL) &
414 (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
415 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
416 return -EINVAL;
417 }
418 ret = orangefs_getflags(inode, &old_uval);
419 if (ret)
420 return ret;
421 ret = vfs_ioc_setflags_prepare(inode, old_uval, uval);
422 if (ret)
423 return ret;
424 val = uval;
425 gossip_debug(GOSSIP_FILE_DEBUG,
426 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
427 (unsigned long long)val);
428 ret = orangefs_inode_setxattr(inode,
429 "user.pvfs2.meta_hint",
430 &val, sizeof(val), 0);
431 }
432
433 return ret;
434 }
435
orangefs_fault(struct vm_fault * vmf)436 static vm_fault_t orangefs_fault(struct vm_fault *vmf)
437 {
438 struct file *file = vmf->vma->vm_file;
439 int ret;
440 ret = orangefs_inode_getattr(file->f_mapping->host,
441 ORANGEFS_GETATTR_SIZE);
442 if (ret == -ESTALE)
443 ret = -EIO;
444 if (ret) {
445 gossip_err("%s: orangefs_inode_getattr failed, "
446 "ret:%d:.\n", __func__, ret);
447 return VM_FAULT_SIGBUS;
448 }
449 return filemap_fault(vmf);
450 }
451
452 static const struct vm_operations_struct orangefs_file_vm_ops = {
453 .fault = orangefs_fault,
454 .map_pages = filemap_map_pages,
455 .page_mkwrite = orangefs_page_mkwrite,
456 };
457
458 /*
459 * Memory map a region of a file.
460 */
orangefs_file_mmap(struct file * file,struct vm_area_struct * vma)461 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
462 {
463 int ret;
464
465 ret = orangefs_revalidate_mapping(file_inode(file));
466 if (ret)
467 return ret;
468
469 gossip_debug(GOSSIP_FILE_DEBUG,
470 "orangefs_file_mmap: called on %s\n",
471 (file ?
472 (char *)file->f_path.dentry->d_name.name :
473 (char *)"Unknown"));
474
475 /* set the sequential readahead hint */
476 vma->vm_flags |= VM_SEQ_READ;
477 vma->vm_flags &= ~VM_RAND_READ;
478
479 file_accessed(file);
480 vma->vm_ops = &orangefs_file_vm_ops;
481 return 0;
482 }
483
484 #define mapping_nrpages(idata) ((idata)->nrpages)
485
486 /*
487 * Called to notify the module that there are no more references to
488 * this file (i.e. no processes have it open).
489 *
490 * \note Not called when each file is closed.
491 */
orangefs_file_release(struct inode * inode,struct file * file)492 static int orangefs_file_release(struct inode *inode, struct file *file)
493 {
494 gossip_debug(GOSSIP_FILE_DEBUG,
495 "orangefs_file_release: called on %pD\n",
496 file);
497
498 /*
499 * remove all associated inode pages from the page cache and
500 * readahead cache (if any); this forces an expensive refresh of
501 * data for the next caller of mmap (or 'get_block' accesses)
502 */
503 if (file_inode(file) &&
504 file_inode(file)->i_mapping &&
505 mapping_nrpages(&file_inode(file)->i_data)) {
506 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
507 gossip_debug(GOSSIP_INODE_DEBUG,
508 "calling flush_racache on %pU\n",
509 get_khandle_from_ino(inode));
510 flush_racache(inode);
511 gossip_debug(GOSSIP_INODE_DEBUG,
512 "flush_racache finished\n");
513 }
514
515 }
516 return 0;
517 }
518
519 /*
520 * Push all data for a specific file onto permanent storage.
521 */
orangefs_fsync(struct file * file,loff_t start,loff_t end,int datasync)522 static int orangefs_fsync(struct file *file,
523 loff_t start,
524 loff_t end,
525 int datasync)
526 {
527 int ret;
528 struct orangefs_inode_s *orangefs_inode =
529 ORANGEFS_I(file_inode(file));
530 struct orangefs_kernel_op_s *new_op = NULL;
531
532 ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
533 start, end);
534 if (ret < 0)
535 return ret;
536
537 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
538 if (!new_op)
539 return -ENOMEM;
540 new_op->upcall.req.fsync.refn = orangefs_inode->refn;
541
542 ret = service_operation(new_op,
543 "orangefs_fsync",
544 get_interruptible_flag(file_inode(file)));
545
546 gossip_debug(GOSSIP_FILE_DEBUG,
547 "orangefs_fsync got return value of %d\n",
548 ret);
549
550 op_release(new_op);
551 return ret;
552 }
553
554 /*
555 * Change the file pointer position for an instance of an open file.
556 *
557 * \note If .llseek is overriden, we must acquire lock as described in
558 * Documentation/filesystems/locking.rst.
559 *
560 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
561 * require much changes to the FS
562 */
orangefs_file_llseek(struct file * file,loff_t offset,int origin)563 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
564 {
565 int ret = -EINVAL;
566 struct inode *inode = file_inode(file);
567
568 if (origin == SEEK_END) {
569 /*
570 * revalidate the inode's file size.
571 * NOTE: We are only interested in file size here,
572 * so we set mask accordingly.
573 */
574 ret = orangefs_inode_getattr(file->f_mapping->host,
575 ORANGEFS_GETATTR_SIZE);
576 if (ret == -ESTALE)
577 ret = -EIO;
578 if (ret) {
579 gossip_debug(GOSSIP_FILE_DEBUG,
580 "%s:%s:%d calling make bad inode\n",
581 __FILE__,
582 __func__,
583 __LINE__);
584 return ret;
585 }
586 }
587
588 gossip_debug(GOSSIP_FILE_DEBUG,
589 "orangefs_file_llseek: offset is %ld | origin is %d"
590 " | inode size is %lu\n",
591 (long)offset,
592 origin,
593 (unsigned long)i_size_read(inode));
594
595 return generic_file_llseek(file, offset, origin);
596 }
597
598 /*
599 * Support local locks (locks that only this kernel knows about)
600 * if Orangefs was mounted -o local_lock.
601 */
orangefs_lock(struct file * filp,int cmd,struct file_lock * fl)602 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
603 {
604 int rc = -EINVAL;
605
606 if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
607 if (cmd == F_GETLK) {
608 rc = 0;
609 posix_test_lock(filp, fl);
610 } else {
611 rc = posix_lock_file(filp, fl, NULL);
612 }
613 }
614
615 return rc;
616 }
617
orangefs_file_open(struct inode * inode,struct file * file)618 static int orangefs_file_open(struct inode * inode, struct file *file)
619 {
620 file->private_data = NULL;
621 return generic_file_open(inode, file);
622 }
623
orangefs_flush(struct file * file,fl_owner_t id)624 static int orangefs_flush(struct file *file, fl_owner_t id)
625 {
626 /*
627 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
628 * service_operation in orangefs_fsync.
629 *
630 * Do not send fsync to OrangeFS server on a close. Do send fsync
631 * on an explicit fsync call. This duplicates historical OrangeFS
632 * behavior.
633 */
634 struct inode *inode = file->f_mapping->host;
635 int r;
636
637 kfree(file->private_data);
638 file->private_data = NULL;
639
640 if (inode->i_state & I_DIRTY_TIME) {
641 spin_lock(&inode->i_lock);
642 inode->i_state &= ~I_DIRTY_TIME;
643 spin_unlock(&inode->i_lock);
644 mark_inode_dirty_sync(inode);
645 }
646
647 r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
648 if (r > 0)
649 return 0;
650 else
651 return r;
652 }
653
654 /** ORANGEFS implementation of VFS file operations */
655 const struct file_operations orangefs_file_operations = {
656 .llseek = orangefs_file_llseek,
657 .read_iter = orangefs_file_read_iter,
658 .write_iter = orangefs_file_write_iter,
659 .lock = orangefs_lock,
660 .unlocked_ioctl = orangefs_ioctl,
661 .mmap = orangefs_file_mmap,
662 .open = orangefs_file_open,
663 .flush = orangefs_flush,
664 .release = orangefs_file_release,
665 .fsync = orangefs_fsync,
666 };
667