orangefs: hopefully saner op refcounting and locking
[deliverable/linux.git] / fs / orangefs / file.c
1 /*
2 * (C) 2001 Clemson University and The University of Chicago
3 *
4 * See COPYING in top-level directory.
5 */
6
7 /*
8 * Linux VFS file operations.
9 */
10
11 #include "protocol.h"
12 #include "orangefs-kernel.h"
13 #include "orangefs-bufmap.h"
14 #include <linux/fs.h>
15 #include <linux/pagemap.h>
16
17 #define wake_up_daemon_for_return(op) \
18 do { \
19 spin_lock(&op->lock); \
20 op->io_completed = 1; \
21 spin_unlock(&op->lock); \
22 wake_up_interruptible(&op->io_completion_waitq);\
23 } while (0)
24
25 /*
26 * Copy to client-core's address space from the buffers specified
27 * by the iovec upto total_size bytes.
28 * NOTE: the iovector can either contain addresses which
29 * can futher be kernel-space or user-space addresses.
30 * or it can pointers to struct page's
31 */
32 static int precopy_buffers(struct orangefs_bufmap *bufmap,
33 int buffer_index,
34 struct iov_iter *iter,
35 size_t total_size)
36 {
37 int ret = 0;
38 /*
39 * copy data from application/kernel by pulling it out
40 * of the iovec.
41 */
42
43
44 if (total_size) {
45 ret = orangefs_bufmap_copy_from_iovec(bufmap,
46 iter,
47 buffer_index,
48 total_size);
49 if (ret < 0)
50 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
51 __func__,
52 (long)ret);
53 }
54
55 if (ret < 0)
56 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
57 __func__,
58 (long)ret);
59 return ret;
60 }
61
62 /*
63 * Copy from client-core's address space to the buffers specified
64 * by the iovec upto total_size bytes.
65 * NOTE: the iovector can either contain addresses which
66 * can futher be kernel-space or user-space addresses.
67 * or it can pointers to struct page's
68 */
69 static int postcopy_buffers(struct orangefs_bufmap *bufmap,
70 int buffer_index,
71 struct iov_iter *iter,
72 size_t total_size)
73 {
74 int ret = 0;
75 /*
76 * copy data to application/kernel by pushing it out to
77 * the iovec. NOTE; target buffers can be addresses or
78 * struct page pointers.
79 */
80 if (total_size) {
81 ret = orangefs_bufmap_copy_to_iovec(bufmap,
82 iter,
83 buffer_index,
84 total_size);
85 if (ret < 0)
86 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
87 __func__,
88 (long)ret);
89 }
90 return ret;
91 }
92
93 /*
94 * Post and wait for the I/O upcall to finish
95 */
96 static ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
97 loff_t *offset, struct iov_iter *iter,
98 size_t total_size, loff_t readahead_size)
99 {
100 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
101 struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
102 struct orangefs_bufmap *bufmap = NULL;
103 struct orangefs_kernel_op_s *new_op = NULL;
104 int buffer_index = -1;
105 ssize_t ret;
106
107 new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
108 if (!new_op)
109 return -ENOMEM;
110
111 /* synchronous I/O */
112 new_op->upcall.req.io.async_vfs_io = ORANGEFS_VFS_SYNC_IO;
113 new_op->upcall.req.io.readahead_size = readahead_size;
114 new_op->upcall.req.io.io_type = type;
115 new_op->upcall.req.io.refn = orangefs_inode->refn;
116
117 populate_shared_memory:
118 /* get a shared buffer index */
119 ret = orangefs_bufmap_get(&bufmap, &buffer_index);
120 if (ret < 0) {
121 gossip_debug(GOSSIP_FILE_DEBUG,
122 "%s: orangefs_bufmap_get failure (%ld)\n",
123 __func__, (long)ret);
124 goto out;
125 }
126 gossip_debug(GOSSIP_FILE_DEBUG,
127 "%s(%pU): GET op %p -> buffer_index %d\n",
128 __func__,
129 handle,
130 new_op,
131 buffer_index);
132
133 new_op->uses_shared_memory = 1;
134 new_op->upcall.req.io.buf_index = buffer_index;
135 new_op->upcall.req.io.count = total_size;
136 new_op->upcall.req.io.offset = *offset;
137
138 gossip_debug(GOSSIP_FILE_DEBUG,
139 "%s(%pU): offset: %llu total_size: %zd\n",
140 __func__,
141 handle,
142 llu(*offset),
143 total_size);
144 /*
145 * Stage 1: copy the buffers into client-core's address space
146 * precopy_buffers only pertains to writes.
147 */
148 if (type == ORANGEFS_IO_WRITE) {
149 ret = precopy_buffers(bufmap,
150 buffer_index,
151 iter,
152 total_size);
153 if (ret < 0)
154 goto out;
155 }
156
157 gossip_debug(GOSSIP_FILE_DEBUG,
158 "%s(%pU): Calling post_io_request with tag (%llu)\n",
159 __func__,
160 handle,
161 llu(new_op->tag));
162
163 /* Stage 2: Service the I/O operation */
164 ret = service_operation(new_op,
165 type == ORANGEFS_IO_WRITE ?
166 "file_write" :
167 "file_read",
168 get_interruptible_flag(inode));
169
170 /*
171 * If service_operation() returns -EAGAIN #and# the operation was
172 * purged from orangefs_request_list or htable_ops_in_progress, then
173 * we know that the client was restarted, causing the shared memory
174 * area to be wiped clean. To restart a write operation in this
175 * case, we must re-copy the data from the user's iovec to a NEW
176 * shared memory location. To restart a read operation, we must get
177 * a new shared memory location.
178 */
179 if (ret == -EAGAIN && op_state_purged(new_op)) {
180 orangefs_bufmap_put(bufmap, buffer_index);
181 gossip_debug(GOSSIP_FILE_DEBUG,
182 "%s:going to repopulate_shared_memory.\n",
183 __func__);
184 goto populate_shared_memory;
185 }
186
187 if (ret < 0) {
188 handle_io_error();
189 /*
190 * don't write an error to syslog on signaled operation
191 * termination unless we've got debugging turned on, as
192 * this can happen regularly (i.e. ctrl-c)
193 */
194 if (ret == -EINTR)
195 gossip_debug(GOSSIP_FILE_DEBUG,
196 "%s: returning error %ld\n", __func__,
197 (long)ret);
198 else
199 gossip_err("%s: error in %s handle %pU, returning %zd\n",
200 __func__,
201 type == ORANGEFS_IO_READ ?
202 "read from" : "write to",
203 handle, ret);
204 goto out;
205 }
206
207 /*
208 * Stage 3: Post copy buffers from client-core's address space
209 * postcopy_buffers only pertains to reads.
210 */
211 if (type == ORANGEFS_IO_READ) {
212 ret = postcopy_buffers(bufmap,
213 buffer_index,
214 iter,
215 new_op->downcall.resp.io.amt_complete);
216 if (ret < 0) {
217 /*
218 * put error codes in downcall so that handle_io_error()
219 * preserves it properly
220 */
221 new_op->downcall.status = ret;
222 handle_io_error();
223 goto out;
224 }
225 }
226 gossip_debug(GOSSIP_FILE_DEBUG,
227 "%s(%pU): Amount written as returned by the sys-io call:%d\n",
228 __func__,
229 handle,
230 (int)new_op->downcall.resp.io.amt_complete);
231
232 ret = new_op->downcall.resp.io.amt_complete;
233
234 /*
235 * tell the device file owner waiting on I/O that this read has
236 * completed and it can return now.
237 */
238 wake_up_daemon_for_return(new_op);
239
240 out:
241 if (buffer_index >= 0) {
242 orangefs_bufmap_put(bufmap, buffer_index);
243 gossip_debug(GOSSIP_FILE_DEBUG,
244 "%s(%pU): PUT buffer_index %d\n",
245 __func__, handle, buffer_index);
246 buffer_index = -1;
247 }
248 op_release(new_op);
249 return ret;
250 }
251
252 /*
253 * Common entry point for read/write/readv/writev
254 * This function will dispatch it to either the direct I/O
255 * or buffered I/O path depending on the mount options and/or
256 * augmented/extended metadata attached to the file.
257 * Note: File extended attributes override any mount options.
258 */
259 static ssize_t do_readv_writev(enum ORANGEFS_io_type type, struct file *file,
260 loff_t *offset, struct iov_iter *iter)
261 {
262 struct inode *inode = file->f_mapping->host;
263 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
264 struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
265 size_t count = iov_iter_count(iter);
266 ssize_t total_count = 0;
267 ssize_t ret = -EINVAL;
268
269 gossip_debug(GOSSIP_FILE_DEBUG,
270 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
271 __func__,
272 handle,
273 (int)count);
274
275 if (type == ORANGEFS_IO_WRITE) {
276 gossip_debug(GOSSIP_FILE_DEBUG,
277 "%s(%pU): proceeding with offset : %llu, "
278 "size %d\n",
279 __func__,
280 handle,
281 llu(*offset),
282 (int)count);
283 }
284
285 if (count == 0) {
286 ret = 0;
287 goto out;
288 }
289
290 while (iov_iter_count(iter)) {
291 size_t each_count = iov_iter_count(iter);
292 size_t amt_complete;
293
294 /* how much to transfer in this loop iteration */
295 if (each_count > orangefs_bufmap_size_query())
296 each_count = orangefs_bufmap_size_query();
297
298 gossip_debug(GOSSIP_FILE_DEBUG,
299 "%s(%pU): size of each_count(%d)\n",
300 __func__,
301 handle,
302 (int)each_count);
303 gossip_debug(GOSSIP_FILE_DEBUG,
304 "%s(%pU): BEFORE wait_for_io: offset is %d\n",
305 __func__,
306 handle,
307 (int)*offset);
308
309 ret = wait_for_direct_io(type, inode, offset, iter,
310 each_count, 0);
311 gossip_debug(GOSSIP_FILE_DEBUG,
312 "%s(%pU): return from wait_for_io:%d\n",
313 __func__,
314 handle,
315 (int)ret);
316
317 if (ret < 0)
318 goto out;
319
320 *offset += ret;
321 total_count += ret;
322 amt_complete = ret;
323
324 gossip_debug(GOSSIP_FILE_DEBUG,
325 "%s(%pU): AFTER wait_for_io: offset is %d\n",
326 __func__,
327 handle,
328 (int)*offset);
329
330 /*
331 * if we got a short I/O operations,
332 * fall out and return what we got so far
333 */
334 if (amt_complete < each_count)
335 break;
336 } /*end while */
337
338 if (total_count > 0)
339 ret = total_count;
340 out:
341 if (ret > 0) {
342 if (type == ORANGEFS_IO_READ) {
343 file_accessed(file);
344 } else {
345 SetMtimeFlag(orangefs_inode);
346 inode->i_mtime = CURRENT_TIME;
347 mark_inode_dirty_sync(inode);
348 }
349 }
350
351 gossip_debug(GOSSIP_FILE_DEBUG,
352 "%s(%pU): Value(%d) returned.\n",
353 __func__,
354 handle,
355 (int)ret);
356
357 return ret;
358 }
359
360 /*
361 * Read data from a specified offset in a file (referenced by inode).
362 * Data may be placed either in a user or kernel buffer.
363 */
364 ssize_t orangefs_inode_read(struct inode *inode,
365 struct iov_iter *iter,
366 loff_t *offset,
367 loff_t readahead_size)
368 {
369 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
370 size_t count = iov_iter_count(iter);
371 size_t bufmap_size;
372 ssize_t ret = -EINVAL;
373
374 g_orangefs_stats.reads++;
375
376 bufmap_size = orangefs_bufmap_size_query();
377 if (count > bufmap_size) {
378 gossip_debug(GOSSIP_FILE_DEBUG,
379 "%s: count is too large (%zd/%zd)!\n",
380 __func__, count, bufmap_size);
381 return -EINVAL;
382 }
383
384 gossip_debug(GOSSIP_FILE_DEBUG,
385 "%s(%pU) %zd@%llu\n",
386 __func__,
387 &orangefs_inode->refn.khandle,
388 count,
389 llu(*offset));
390
391 ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, offset, iter,
392 count, readahead_size);
393 if (ret > 0)
394 *offset += ret;
395
396 gossip_debug(GOSSIP_FILE_DEBUG,
397 "%s(%pU): Value(%zd) returned.\n",
398 __func__,
399 &orangefs_inode->refn.khandle,
400 ret);
401
402 return ret;
403 }
404
405 static ssize_t orangefs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
406 {
407 struct file *file = iocb->ki_filp;
408 loff_t pos = *(&iocb->ki_pos);
409 ssize_t rc = 0;
410
411 BUG_ON(iocb->private);
412
413 gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_read_iter\n");
414
415 g_orangefs_stats.reads++;
416
417 rc = do_readv_writev(ORANGEFS_IO_READ, file, &pos, iter);
418 iocb->ki_pos = pos;
419
420 return rc;
421 }
422
423 static ssize_t orangefs_file_write_iter(struct kiocb *iocb, struct iov_iter *iter)
424 {
425 struct file *file = iocb->ki_filp;
426 loff_t pos;
427 ssize_t rc;
428
429 BUG_ON(iocb->private);
430
431 gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_write_iter\n");
432
433 mutex_lock(&file->f_mapping->host->i_mutex);
434
435 /* Make sure generic_write_checks sees an up to date inode size. */
436 if (file->f_flags & O_APPEND) {
437 rc = orangefs_inode_getattr(file->f_mapping->host,
438 ORANGEFS_ATTR_SYS_SIZE);
439 if (rc) {
440 gossip_err("%s: orangefs_inode_getattr failed, rc:%zd:.\n",
441 __func__, rc);
442 goto out;
443 }
444 }
445
446 if (file->f_pos > i_size_read(file->f_mapping->host))
447 orangefs_i_size_write(file->f_mapping->host, file->f_pos);
448
449 rc = generic_write_checks(iocb, iter);
450
451 if (rc <= 0) {
452 gossip_err("%s: generic_write_checks failed, rc:%zd:.\n",
453 __func__, rc);
454 goto out;
455 }
456
457 /*
458 * if we are appending, generic_write_checks would have updated
459 * pos to the end of the file, so we will wait till now to set
460 * pos...
461 */
462 pos = *(&iocb->ki_pos);
463
464 rc = do_readv_writev(ORANGEFS_IO_WRITE,
465 file,
466 &pos,
467 iter);
468 if (rc < 0) {
469 gossip_err("%s: do_readv_writev failed, rc:%zd:.\n",
470 __func__, rc);
471 goto out;
472 }
473
474 iocb->ki_pos = pos;
475 g_orangefs_stats.writes++;
476
477 out:
478
479 mutex_unlock(&file->f_mapping->host->i_mutex);
480 return rc;
481 }
482
483 /*
484 * Perform a miscellaneous operation on a file.
485 */
486 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
487 {
488 int ret = -ENOTTY;
489 __u64 val = 0;
490 unsigned long uval;
491
492 gossip_debug(GOSSIP_FILE_DEBUG,
493 "orangefs_ioctl: called with cmd %d\n",
494 cmd);
495
496 /*
497 * we understand some general ioctls on files, such as the immutable
498 * and append flags
499 */
500 if (cmd == FS_IOC_GETFLAGS) {
501 val = 0;
502 ret = orangefs_inode_getxattr(file_inode(file),
503 ORANGEFS_XATTR_NAME_DEFAULT_PREFIX,
504 "user.pvfs2.meta_hint",
505 &val, sizeof(val));
506 if (ret < 0 && ret != -ENODATA)
507 return ret;
508 else if (ret == -ENODATA)
509 val = 0;
510 uval = val;
511 gossip_debug(GOSSIP_FILE_DEBUG,
512 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
513 (unsigned long long)uval);
514 return put_user(uval, (int __user *)arg);
515 } else if (cmd == FS_IOC_SETFLAGS) {
516 ret = 0;
517 if (get_user(uval, (int __user *)arg))
518 return -EFAULT;
519 /*
520 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
521 * is turned on for a file. The user is not allowed to turn
522 * on this bit, but the bit is present if the user first gets
523 * the flags and then updates the flags with some new
524 * settings. So, we ignore it in the following edit. bligon.
525 */
526 if ((uval & ~ORANGEFS_MIRROR_FL) &
527 (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
528 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
529 return -EINVAL;
530 }
531 val = uval;
532 gossip_debug(GOSSIP_FILE_DEBUG,
533 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
534 (unsigned long long)val);
535 ret = orangefs_inode_setxattr(file_inode(file),
536 ORANGEFS_XATTR_NAME_DEFAULT_PREFIX,
537 "user.pvfs2.meta_hint",
538 &val, sizeof(val), 0);
539 }
540
541 return ret;
542 }
543
544 /*
545 * Memory map a region of a file.
546 */
547 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
548 {
549 gossip_debug(GOSSIP_FILE_DEBUG,
550 "orangefs_file_mmap: called on %s\n",
551 (file ?
552 (char *)file->f_path.dentry->d_name.name :
553 (char *)"Unknown"));
554
555 /* set the sequential readahead hint */
556 vma->vm_flags |= VM_SEQ_READ;
557 vma->vm_flags &= ~VM_RAND_READ;
558
559 /* Use readonly mmap since we cannot support writable maps. */
560 return generic_file_readonly_mmap(file, vma);
561 }
562
563 #define mapping_nrpages(idata) ((idata)->nrpages)
564
565 /*
566 * Called to notify the module that there are no more references to
567 * this file (i.e. no processes have it open).
568 *
569 * \note Not called when each file is closed.
570 */
571 static int orangefs_file_release(struct inode *inode, struct file *file)
572 {
573 gossip_debug(GOSSIP_FILE_DEBUG,
574 "orangefs_file_release: called on %s\n",
575 file->f_path.dentry->d_name.name);
576
577 orangefs_flush_inode(inode);
578
579 /*
580 * remove all associated inode pages from the page cache and mmap
581 * readahead cache (if any); this forces an expensive refresh of
582 * data for the next caller of mmap (or 'get_block' accesses)
583 */
584 if (file->f_path.dentry->d_inode &&
585 file->f_path.dentry->d_inode->i_mapping &&
586 mapping_nrpages(&file->f_path.dentry->d_inode->i_data))
587 truncate_inode_pages(file->f_path.dentry->d_inode->i_mapping,
588 0);
589 return 0;
590 }
591
592 /*
593 * Push all data for a specific file onto permanent storage.
594 */
595 static int orangefs_fsync(struct file *file,
596 loff_t start,
597 loff_t end,
598 int datasync)
599 {
600 int ret = -EINVAL;
601 struct orangefs_inode_s *orangefs_inode =
602 ORANGEFS_I(file->f_path.dentry->d_inode);
603 struct orangefs_kernel_op_s *new_op = NULL;
604
605 /* required call */
606 filemap_write_and_wait_range(file->f_mapping, start, end);
607
608 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
609 if (!new_op)
610 return -ENOMEM;
611 new_op->upcall.req.fsync.refn = orangefs_inode->refn;
612
613 ret = service_operation(new_op,
614 "orangefs_fsync",
615 get_interruptible_flag(file->f_path.dentry->d_inode));
616
617 gossip_debug(GOSSIP_FILE_DEBUG,
618 "orangefs_fsync got return value of %d\n",
619 ret);
620
621 op_release(new_op);
622
623 orangefs_flush_inode(file->f_path.dentry->d_inode);
624 return ret;
625 }
626
627 /*
628 * Change the file pointer position for an instance of an open file.
629 *
630 * \note If .llseek is overriden, we must acquire lock as described in
631 * Documentation/filesystems/Locking.
632 *
633 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
634 * require much changes to the FS
635 */
636 static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
637 {
638 int ret = -EINVAL;
639 struct inode *inode = file->f_path.dentry->d_inode;
640
641 if (!inode) {
642 gossip_err("orangefs_file_llseek: invalid inode (NULL)\n");
643 return ret;
644 }
645
646 if (origin == ORANGEFS_SEEK_END) {
647 /*
648 * revalidate the inode's file size.
649 * NOTE: We are only interested in file size here,
650 * so we set mask accordingly.
651 */
652 ret = orangefs_inode_getattr(inode, ORANGEFS_ATTR_SYS_SIZE);
653 if (ret) {
654 gossip_debug(GOSSIP_FILE_DEBUG,
655 "%s:%s:%d calling make bad inode\n",
656 __FILE__,
657 __func__,
658 __LINE__);
659 orangefs_make_bad_inode(inode);
660 return ret;
661 }
662 }
663
664 gossip_debug(GOSSIP_FILE_DEBUG,
665 "orangefs_file_llseek: offset is %ld | origin is %d"
666 " | inode size is %lu\n",
667 (long)offset,
668 origin,
669 (unsigned long)file->f_path.dentry->d_inode->i_size);
670
671 return generic_file_llseek(file, offset, origin);
672 }
673
674 /*
675 * Support local locks (locks that only this kernel knows about)
676 * if Orangefs was mounted -o local_lock.
677 */
678 static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
679 {
680 int rc = -EINVAL;
681
682 if (ORANGEFS_SB(filp->f_inode->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
683 if (cmd == F_GETLK) {
684 rc = 0;
685 posix_test_lock(filp, fl);
686 } else {
687 rc = posix_lock_file(filp, fl, NULL);
688 }
689 }
690
691 return rc;
692 }
693
694 /** ORANGEFS implementation of VFS file operations */
695 const struct file_operations orangefs_file_operations = {
696 .llseek = orangefs_file_llseek,
697 .read_iter = orangefs_file_read_iter,
698 .write_iter = orangefs_file_write_iter,
699 .lock = orangefs_lock,
700 .unlocked_ioctl = orangefs_ioctl,
701 .mmap = orangefs_file_mmap,
702 .open = generic_file_open,
703 .release = orangefs_file_release,
704 .fsync = orangefs_fsync,
705 };
This page took 0.085581 seconds and 5 git commands to generate.