forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-01-05 071106ecf68c401173c58808b1cf5f68cc50d390 
 kernel/fs/orangefs/file.c
....@@ -1,6 +1,7 @@
11 // SPDX-License-Identifier: GPL-2.0
22 /*
33  * (C) 2001 Clemson University and The University of Chicago
4
+ * Copyright 2018 Omnibond Systems, L.L.C.
45  *
56  * See COPYING in top-level directory.
67  */
....@@ -44,15 +45,19 @@
4445 /*
4546  * Post and wait for the I/O upcall to finish
4647  */
47
-static ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
48
-		loff_t *offset, struct iov_iter *iter,
49
-		size_t total_size, loff_t readahead_size)
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)
5052 {
5153 	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
5254 	struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
5355 	struct orangefs_kernel_op_s *new_op = NULL;
54
-	int buffer_index = -1;
56
+	int buffer_index;
5557 	ssize_t ret;
58
+	size_t copy_amount;
59
+	int open_for_read;
60
+	int open_for_write;
5661 
5762 	new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
5863 	if (!new_op)
....@@ -84,6 +89,42 @@
8489 	new_op->upcall.req.io.buf_index = buffer_index;
8590 	new_op->upcall.req.io.count = total_size;
8691 	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;
87128 
88129 	gossip_debug(GOSSIP_FILE_DEBUG,
89130 		     "%s(%pU): offset: %llu total_size: %zd\n",
....@@ -128,7 +169,6 @@
128169 	 */
129170 	if (ret == -EAGAIN && op_state_purged(new_op)) {
130171 		orangefs_bufmap_put(buffer_index);
131
-		buffer_index = -1;
132172 		if (type == ORANGEFS_IO_WRITE)
133173 			iov_iter_revert(iter, total_size);
134174 		gossip_debug(GOSSIP_FILE_DEBUG,
....@@ -168,7 +208,10 @@
168208 			 * trigger the write.
169209 			 */
170210 			case OP_VFS_STATE_INPROGR:
171
-				ret = total_size;
211
+				if (type == ORANGEFS_IO_READ)
212
+					ret = -EINTR;
213
+				else
214
+					ret = total_size;
172215 				break;
173216 			default:
174217 				gossip_err("%s: unexpected op state :%d:.\n",
....@@ -204,8 +247,25 @@
204247 		 *       can futher be kernel-space or user-space addresses.
205248 		 *       or it can pointers to struct page's
206249 		 */
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
+
207267 		ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
208
-		    new_op->downcall.resp.io.amt_complete);
268
+			copy_amount);
209269 		if (ret < 0) {
210270 			gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
211271 			    __func__, (long)ret);
....@@ -223,250 +283,112 @@
223283 
224284 out:
225285 	if (buffer_index >= 0) {
226
-		orangefs_bufmap_put(buffer_index);
227
-		gossip_debug(GOSSIP_FILE_DEBUG,
228
-			     "%s(%pU): PUT buffer_index %d\n",
229
-			     __func__, handle, buffer_index);
230
-		buffer_index = -1;
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
+		}
231299 	}
232300 	op_release(new_op);
233301 	return ret;
234302 }
235303 
236
-/*
237
- * Common entry point for read/write/readv/writev
238
- * This function will dispatch it to either the direct I/O
239
- * or buffered I/O path depending on the mount options and/or
240
- * augmented/extended metadata attached to the file.
241
- * Note: File extended attributes override any mount options.
242
- */
243
-static ssize_t do_readv_writev(enum ORANGEFS_io_type type, struct file *file,
244
-		loff_t *offset, struct iov_iter *iter)
304
+int orangefs_revalidate_mapping(struct inode *inode)
245305 {
246
-	struct inode *inode = file->f_mapping->host;
247306 	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
248
-	struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
249
-	size_t count = iov_iter_count(iter);
250
-	ssize_t total_count = 0;
251
-	ssize_t ret = -EINVAL;
307
+	struct address_space *mapping = inode->i_mapping;
308
+	unsigned long *bitlock = &orangefs_inode->bitlock;
309
+	int ret;
252310 
253
-	gossip_debug(GOSSIP_FILE_DEBUG,
254
-		"%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
255
-		__func__,
256
-		handle,
257
-		(int)count);
258
-
259
-	if (type == ORANGEFS_IO_WRITE) {
260
-		gossip_debug(GOSSIP_FILE_DEBUG,
261
-			     "%s(%pU): proceeding with offset : %llu, "
262
-			     "size %d\n",
263
-			     __func__,
264
-			     handle,
265
-			     llu(*offset),
266
-			     (int)count);
267
-	}
268
-
269
-	if (count == 0) {
270
-		ret = 0;
271
-		goto out;
272
-	}
273
-
274
-	while (iov_iter_count(iter)) {
275
-		size_t each_count = iov_iter_count(iter);
276
-		size_t amt_complete;
277
-
278
-		/* how much to transfer in this loop iteration */
279
-		if (each_count > orangefs_bufmap_size_query())
280
-			each_count = orangefs_bufmap_size_query();
281
-
282
-		gossip_debug(GOSSIP_FILE_DEBUG,
283
-			     "%s(%pU): size of each_count(%d)\n",
284
-			     __func__,
285
-			     handle,
286
-			     (int)each_count);
287
-		gossip_debug(GOSSIP_FILE_DEBUG,
288
-			     "%s(%pU): BEFORE wait_for_io: offset is %d\n",
289
-			     __func__,
290
-			     handle,
291
-			     (int)*offset);
292
-
293
-		ret = wait_for_direct_io(type, inode, offset, iter,
294
-				each_count, 0);
295
-		gossip_debug(GOSSIP_FILE_DEBUG,
296
-			     "%s(%pU): return from wait_for_io:%d\n",
297
-			     __func__,
298
-			     handle,
299
-			     (int)ret);
300
-
301
-		if (ret < 0)
302
-			goto out;
303
-
304
-		*offset += ret;
305
-		total_count += ret;
306
-		amt_complete = ret;
307
-
308
-		gossip_debug(GOSSIP_FILE_DEBUG,
309
-			     "%s(%pU): AFTER wait_for_io: offset is %d\n",
310
-			     __func__,
311
-			     handle,
312
-			     (int)*offset);
313
-
314
-		/*
315
-		 * if we got a short I/O operations,
316
-		 * fall out and return what we got so far
317
-		 */
318
-		if (amt_complete < each_count)
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))
319321 			break;
320
-	} /*end while */
322
+		spin_unlock(&inode->i_lock);
323
+		return 0;
324
+	}
321325 
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
+	orangefs_stats.reads++;
350
+
351
+	down_read(&file_inode(iocb->ki_filp)->i_rwsem);
352
+	ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
353
+	if (ret)
354
+		goto out;
355
+
356
+	ret = generic_file_read_iter(iocb, iter);
322357 out:
323
-	if (total_count > 0)
324
-		ret = total_count;
325
-	if (ret > 0) {
326
-		if (type == ORANGEFS_IO_READ) {
327
-			file_accessed(file);
328
-		} else {
329
-			file_update_time(file);
330
-			/*
331
-			 * Must invalidate to ensure write loop doesn't
332
-			 * prevent kernel from reading updated
333
-			 * attribute.  Size probably changed because of
334
-			 * the write, and other clients could update
335
-			 * any other attribute.
336
-			 */
337
-			orangefs_inode->getattr_time = jiffies - 1;
338
-		}
339
-	}
340
-
341
-	gossip_debug(GOSSIP_FILE_DEBUG,
342
-		     "%s(%pU): Value(%d) returned.\n",
343
-		     __func__,
344
-		     handle,
345
-		     (int)ret);
346
-
358
+	up_read(&file_inode(iocb->ki_filp)->i_rwsem);
347359 	return ret;
348360 }
349361 
350
-/*
351
- * Read data from a specified offset in a file (referenced by inode).
352
- * Data may be placed either in a user or kernel buffer.
353
- */
354
-ssize_t orangefs_inode_read(struct inode *inode,
355
-			    struct iov_iter *iter,
356
-			    loff_t *offset,
357
-			    loff_t readahead_size)
362
+static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
363
+    struct iov_iter *iter)
358364 {
359
-	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
360
-	size_t count = iov_iter_count(iter);
361
-	size_t bufmap_size;
362
-	ssize_t ret = -EINVAL;
363
-
364
-	orangefs_stats.reads++;
365
-
366
-	bufmap_size = orangefs_bufmap_size_query();
367
-	if (count > bufmap_size) {
368
-		gossip_debug(GOSSIP_FILE_DEBUG,
369
-			     "%s: count is too large (%zd/%zd)!\n",
370
-			     __func__, count, bufmap_size);
371
-		return -EINVAL;
372
-	}
373
-
374
-	gossip_debug(GOSSIP_FILE_DEBUG,
375
-		     "%s(%pU) %zd@%llu\n",
376
-		     __func__,
377
-		     &orangefs_inode->refn.khandle,
378
-		     count,
379
-		     llu(*offset));
380
-
381
-	ret = wait_for_direct_io(ORANGEFS_IO_READ, inode, offset, iter,
382
-			count, readahead_size);
383
-	if (ret > 0)
384
-		*offset += ret;
385
-
386
-	gossip_debug(GOSSIP_FILE_DEBUG,
387
-		     "%s(%pU): Value(%zd) returned.\n",
388
-		     __func__,
389
-		     &orangefs_inode->refn.khandle,
390
-		     ret);
391
-
392
-	return ret;
393
-}
394
-
395
-static ssize_t orangefs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
396
-{
397
-	struct file *file = iocb->ki_filp;
398
-	loff_t pos = iocb->ki_pos;
399
-	ssize_t rc = 0;
400
-
401
-	BUG_ON(iocb->private);
402
-
403
-	gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_read_iter\n");
404
-
405
-	orangefs_stats.reads++;
406
-
407
-	rc = do_readv_writev(ORANGEFS_IO_READ, file, &pos, iter);
408
-	iocb->ki_pos = pos;
409
-
410
-	return rc;
411
-}
412
-
413
-static ssize_t orangefs_file_write_iter(struct kiocb *iocb, struct iov_iter *iter)
414
-{
415
-	struct file *file = iocb->ki_filp;
416
-	loff_t pos;
417
-	ssize_t rc;
418
-
419
-	BUG_ON(iocb->private);
420
-
421
-	gossip_debug(GOSSIP_FILE_DEBUG, "orangefs_file_write_iter\n");
422
-
423
-	inode_lock(file->f_mapping->host);
424
-
425
-	/* Make sure generic_write_checks sees an up to date inode size. */
426
-	if (file->f_flags & O_APPEND) {
427
-		rc = orangefs_inode_getattr(file->f_mapping->host, 0, 1,
428
-		    STATX_SIZE);
429
-		if (rc == -ESTALE)
430
-			rc = -EIO;
431
-		if (rc) {
432
-			gossip_err("%s: orangefs_inode_getattr failed, "
433
-			    "rc:%zd:.\n", __func__, rc);
434
-			goto out;
435
-		}
436
-	}
437
-
438
-	rc = generic_write_checks(iocb, iter);
439
-
440
-	if (rc <= 0) {
441
-		gossip_err("%s: generic_write_checks failed, rc:%zd:.\n",
442
-			   __func__, rc);
443
-		goto out;
444
-	}
445
-
446
-	/*
447
-	 * if we are appending, generic_write_checks would have updated
448
-	 * pos to the end of the file, so we will wait till now to set
449
-	 * pos...
450
-	 */
451
-	pos = iocb->ki_pos;
452
-
453
-	rc = do_readv_writev(ORANGEFS_IO_WRITE,
454
-			     file,
455
-			     &pos,
456
-			     iter);
457
-	if (rc < 0) {
458
-		gossip_err("%s: do_readv_writev failed, rc:%zd:.\n",
459
-			   __func__, rc);
460
-		goto out;
461
-	}
462
-
463
-	iocb->ki_pos = pos;
365
+	int ret;
464366 	orangefs_stats.writes++;
465367 
466
-out:
368
+	if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
369
+		ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
370
+		if (ret)
371
+			return ret;
372
+	}
467373 
468
-	inode_unlock(file->f_mapping->host);
469
-	return rc;
374
+	ret = generic_file_write_iter(iocb, iter);
375
+	return ret;
376
+}
377
+
378
+static int orangefs_getflags(struct inode *inode, unsigned long *uval)
379
+{
380
+	__u64 val = 0;
381
+	int ret;
382
+
383
+	ret = orangefs_inode_getxattr(inode,
384
+				      "user.pvfs2.meta_hint",
385
+				      &val, sizeof(val));
386
+	if (ret < 0 && ret != -ENODATA)
387
+		return ret;
388
+	else if (ret == -ENODATA)
389
+		val = 0;
390
+	*uval = val;
391
+	return 0;
470392 }
471393 
472394 /*
....@@ -474,6 +396,7 @@
474396  */
475397 static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
476398 {
399
+	struct inode *inode = file_inode(file);
477400 	int ret = -ENOTTY;
478401 	__u64 val = 0;
479402 	unsigned long uval;
....@@ -487,20 +410,16 @@
487410 	 * and append flags
488411 	 */
489412 	if (cmd == FS_IOC_GETFLAGS) {
490
-		val = 0;
491
-		ret = orangefs_inode_getxattr(file_inode(file),
492
-					      "user.pvfs2.meta_hint",
493
-					      &val, sizeof(val));
494
-		if (ret < 0 && ret != -ENODATA)
413
+		ret = orangefs_getflags(inode, &uval);
414
+		if (ret)
495415 			return ret;
496
-		else if (ret == -ENODATA)
497
-			val = 0;
498
-		uval = val;
499416 		gossip_debug(GOSSIP_FILE_DEBUG,
500417 			     "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
501418 			     (unsigned long long)uval);
502419 		return put_user(uval, (int __user *)arg);
503420 	} else if (cmd == FS_IOC_SETFLAGS) {
421
+		unsigned long old_uval;
422
+
504423 		ret = 0;
505424 		if (get_user(uval, (int __user *)arg))
506425 			return -EFAULT;
....@@ -516,11 +435,17 @@
516435 			gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
517436 			return -EINVAL;
518437 		}
438
+		ret = orangefs_getflags(inode, &old_uval);
439
+		if (ret)
440
+			return ret;
441
+		ret = vfs_ioc_setflags_prepare(inode, old_uval, uval);
442
+		if (ret)
443
+			return ret;
519444 		val = uval;
520445 		gossip_debug(GOSSIP_FILE_DEBUG,
521446 			     "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
522447 			     (unsigned long long)val);
523
-		ret = orangefs_inode_setxattr(file_inode(file),
448
+		ret = orangefs_inode_setxattr(inode,
524449 					      "user.pvfs2.meta_hint",
525450 					      &val, sizeof(val), 0);
526451 	}
....@@ -532,14 +457,13 @@
532457 {
533458 	struct file *file = vmf->vma->vm_file;
534459 	int ret;
535
-
536
-	ret = orangefs_inode_getattr(file->f_mapping->host, 0, 1,
537
-	    STATX_SIZE);
460
+	ret = orangefs_inode_getattr(file->f_mapping->host,
461
+	    ORANGEFS_GETATTR_SIZE);
538462 	if (ret == -ESTALE)
539463 		ret = -EIO;
540464 	if (ret) {
541
-		gossip_err("%s: orangefs_inode_getattr failed, ret:%d:.\n",
542
-				__func__, ret);
465
+		gossip_err("%s: orangefs_inode_getattr failed, "
466
+		    "ret:%d:.\n", __func__, ret);
543467 		return VM_FAULT_SIGBUS;
544468 	}
545469 	return filemap_fault(vmf);
....@@ -548,7 +472,7 @@
548472 static const struct vm_operations_struct orangefs_file_vm_ops = {
549473 	.fault = orangefs_fault,
550474 	.map_pages = filemap_map_pages,
551
-	.page_mkwrite = filemap_page_mkwrite,
475
+	.page_mkwrite = orangefs_page_mkwrite,
552476 };
553477 
554478 /*
....@@ -556,14 +480,17 @@
556480  */
557481 static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
558482 {
483
+	int ret;
484
+
485
+	ret = orangefs_revalidate_mapping(file_inode(file));
486
+	if (ret)
487
+		return ret;
488
+
559489 	gossip_debug(GOSSIP_FILE_DEBUG,
560490 		     "orangefs_file_mmap: called on %s\n",
561491 		     (file ?
562492 			(char *)file->f_path.dentry->d_name.name :
563493 			(char *)"Unknown"));
564
-
565
-	if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
566
-		return -EINVAL;
567494 
568495 	/* set the sequential readahead hint */
569496 	vma->vm_flags |= VM_SEQ_READ;
....@@ -604,8 +531,7 @@
604531 			gossip_debug(GOSSIP_INODE_DEBUG,
605532 			    "flush_racache finished\n");
606533 		}
607
-		truncate_inode_pages(file_inode(file)->i_mapping,
608
-				     0);
534
+
609535 	}
610536 	return 0;
611537 }
....@@ -622,6 +548,11 @@
622548 	struct orangefs_inode_s *orangefs_inode =
623549 		ORANGEFS_I(file_inode(file));
624550 	struct orangefs_kernel_op_s *new_op = NULL;
551
+
552
+	ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
553
+	    start, end);
554
+	if (ret < 0)
555
+		return ret;
625556 
626557 	new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
627558 	if (!new_op)
....@@ -644,7 +575,7 @@
644575  * Change the file pointer position for an instance of an open file.
645576  *
646577  * \note If .llseek is overriden, we must acquire lock as described in
647
- *       Documentation/filesystems/Locking.
578
+ *       Documentation/filesystems/locking.rst.
648579  *
649580  * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
650581  * require much changes to the FS
....@@ -660,8 +591,8 @@
660591 		 * NOTE: We are only interested in file size here,
661592 		 * so we set mask accordingly.
662593 		 */
663
-		ret = orangefs_inode_getattr(file->f_mapping->host, 0, 1,
664
-		    STATX_SIZE);
594
+		ret = orangefs_inode_getattr(file->f_mapping->host,
595
+		    ORANGEFS_GETATTR_SIZE);
665596 		if (ret == -ESTALE)
666597 			ret = -EIO;
667598 		if (ret) {
....@@ -704,6 +635,25 @@
704635 	return rc;
705636 }
706637 
638
+static int orangefs_flush(struct file *file, fl_owner_t id)
639
+{
640
+	/*
641
+	 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
642
+	 * service_operation in orangefs_fsync.
643
+	 *
644
+	 * Do not send fsync to OrangeFS server on a close.  Do send fsync
645
+	 * on an explicit fsync call.  This duplicates historical OrangeFS
646
+	 * behavior.
647
+	 */
648
+	int r;
649
+
650
+	r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
651
+	if (r > 0)
652
+		return 0;
653
+	else
654
+		return r;
655
+}
656
+
707657 /** ORANGEFS implementation of VFS file operations */
708658 const struct file_operations orangefs_file_operations = {
709659 	.llseek		= orangefs_file_llseek,
....@@ -713,6 +663,7 @@
713663 	.unlocked_ioctl	= orangefs_ioctl,
714664 	.mmap		= orangefs_file_mmap,
715665 	.open		= generic_file_open,
666
+	.flush		= orangefs_flush,
716667 	.release	= orangefs_file_release,
717668 	.fsync		= orangefs_fsync,
718669 };