Linux Audio

Check our new training course

Loading...
v6.13.7
  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/filelock.h>
 18#include <linux/pagemap.h>
 19
 20static int flush_racache(struct inode *inode)
 21{
 22	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 23	struct orangefs_kernel_op_s *new_op;
 24	int ret;
 25
 26	gossip_debug(GOSSIP_UTILS_DEBUG,
 27	    "%s: %pU: Handle is %pU | fs_id %d\n", __func__,
 28	    get_khandle_from_ino(inode), &orangefs_inode->refn.khandle,
 29	    orangefs_inode->refn.fs_id);
 30
 31	new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
 32	if (!new_op)
 33		return -ENOMEM;
 34	new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
 35
 36	ret = service_operation(new_op, "orangefs_flush_racache",
 37	    get_interruptible_flag(inode));
 38
 39	gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
 40	    __func__, ret);
 41
 42	op_release(new_op);
 43	return ret;
 44}
 45
 46/*
 47 * Post and wait for the I/O upcall to finish
 48 */
 49ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
 50	loff_t *offset, struct iov_iter *iter, size_t total_size,
 51	loff_t readahead_size, struct orangefs_write_range *wr,
 52	int *index_return, struct file *file)
 53{
 54	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 55	struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
 56	struct orangefs_kernel_op_s *new_op = NULL;
 57	int buffer_index;
 58	ssize_t ret;
 59	size_t copy_amount;
 60	int open_for_read;
 61	int open_for_write;
 62
 63	new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
 64	if (!new_op)
 65		return -ENOMEM;
 66
 67	/* synchronous I/O */
 68	new_op->upcall.req.io.readahead_size = readahead_size;
 69	new_op->upcall.req.io.io_type = type;
 70	new_op->upcall.req.io.refn = orangefs_inode->refn;
 71
 72populate_shared_memory:
 73	/* get a shared buffer index */
 74	buffer_index = orangefs_bufmap_get();
 75	if (buffer_index < 0) {
 76		ret = buffer_index;
 77		gossip_debug(GOSSIP_FILE_DEBUG,
 78			     "%s: orangefs_bufmap_get failure (%zd)\n",
 79			     __func__, ret);
 80		goto out;
 81	}
 82	gossip_debug(GOSSIP_FILE_DEBUG,
 83		     "%s(%pU): GET op %p -> buffer_index %d\n",
 84		     __func__,
 85		     handle,
 86		     new_op,
 87		     buffer_index);
 88
 89	new_op->uses_shared_memory = 1;
 90	new_op->upcall.req.io.buf_index = buffer_index;
 91	new_op->upcall.req.io.count = total_size;
 92	new_op->upcall.req.io.offset = *offset;
 93	if (type == ORANGEFS_IO_WRITE && wr) {
 94		new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
 95		new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
 96	}
 97	/*
 98	 * Orangefs has no open, and orangefs checks file permissions
 99	 * on each file access. Posix requires that file permissions
100	 * be checked on open and nowhere else. Orangefs-through-the-kernel
101	 * needs to seem posix compliant.
102	 *
103	 * The VFS opens files, even if the filesystem provides no
104	 * method. We can see if a file was successfully opened for
105	 * read and or for write by looking at file->f_mode.
106	 *
107	 * When writes are flowing from the page cache, file is no
108	 * longer available. We can trust the VFS to have checked
109	 * file->f_mode before writing to the page cache.
110	 *
111	 * The mode of a file might change between when it is opened
112	 * and IO commences, or it might be created with an arbitrary mode.
113	 *
114	 * We'll make sure we don't hit EACCES during the IO stage by
115	 * using UID 0. Some of the time we have access without changing
116	 * to UID 0 - how to check?
117	 */
118	if (file) {
119		open_for_write = file->f_mode & FMODE_WRITE;
120		open_for_read = file->f_mode & FMODE_READ;
121	} else {
122		open_for_write = 1;
123		open_for_read = 0; /* not relevant? */
124	}
125	if ((type == ORANGEFS_IO_WRITE) && open_for_write)
126		new_op->upcall.uid = 0;
127	if ((type == ORANGEFS_IO_READ) && open_for_read)
128		new_op->upcall.uid = 0;
129
130	gossip_debug(GOSSIP_FILE_DEBUG,
131		     "%s(%pU): offset: %llu total_size: %zd\n",
132		     __func__,
133		     handle,
134		     llu(*offset),
135		     total_size);
136	/*
137	 * Stage 1: copy the buffers into client-core's address space
138	 */
139	if (type == ORANGEFS_IO_WRITE && total_size) {
140		ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
141		    total_size);
142		if (ret < 0) {
143			gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
144			    __func__, (long)ret);
145			goto out;
146		}
147	}
148
149	gossip_debug(GOSSIP_FILE_DEBUG,
150		     "%s(%pU): Calling post_io_request with tag (%llu)\n",
151		     __func__,
152		     handle,
153		     llu(new_op->tag));
154
155	/* Stage 2: Service the I/O operation */
156	ret = service_operation(new_op,
157				type == ORANGEFS_IO_WRITE ?
158					"file_write" :
159					"file_read",
160				get_interruptible_flag(inode));
161
162	/*
163	 * If service_operation() returns -EAGAIN #and# the operation was
164	 * purged from orangefs_request_list or htable_ops_in_progress, then
165	 * we know that the client was restarted, causing the shared memory
166	 * area to be wiped clean.  To restart a  write operation in this
167	 * case, we must re-copy the data from the user's iovec to a NEW
168	 * shared memory location. To restart a read operation, we must get
169	 * a new shared memory location.
170	 */
171	if (ret == -EAGAIN && op_state_purged(new_op)) {
172		orangefs_bufmap_put(buffer_index);
 
173		if (type == ORANGEFS_IO_WRITE)
174			iov_iter_revert(iter, total_size);
175		gossip_debug(GOSSIP_FILE_DEBUG,
176			     "%s:going to repopulate_shared_memory.\n",
177			     __func__);
178		goto populate_shared_memory;
179	}
180
181	if (ret < 0) {
182		if (ret == -EINTR) {
183			/*
184			 * We can't return EINTR if any data was written,
185			 * it's not POSIX. It is minimally acceptable
186			 * to give a partial write, the way NFS does.
187			 *
188			 * It would be optimal to return all or nothing,
189			 * but if a userspace write is bigger than
190			 * an IO buffer, and the interrupt occurs
191			 * between buffer writes, that would not be
192			 * possible.
193			 */
194			switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
195			/*
196			 * If the op was waiting when the interrupt
197			 * occurred, then the client-core did not
198			 * trigger the write.
199			 */
200			case OP_VFS_STATE_WAITING:
201				if (*offset == 0)
202					ret = -EINTR;
203				else
204					ret = 0;
205				break;
206			/*
207			 * If the op was in progress when the interrupt
208			 * occurred, then the client-core was able to
209			 * trigger the write.
210			 */
211			case OP_VFS_STATE_INPROGR:
212				if (type == ORANGEFS_IO_READ)
213					ret = -EINTR;
214				else
215					ret = total_size;
216				break;
217			default:
218				gossip_err("%s: unexpected op state :%d:.\n",
219					   __func__,
220					   new_op->op_state);
221				ret = 0;
222				break;
223			}
224			gossip_debug(GOSSIP_FILE_DEBUG,
225				     "%s: got EINTR, state:%d: %p\n",
226				     __func__,
227				     new_op->op_state,
228				     new_op);
229		} else {
230			gossip_err("%s: error in %s handle %pU, returning %zd\n",
231				__func__,
232				type == ORANGEFS_IO_READ ?
233					"read from" : "write to",
234				handle, ret);
235		}
236		if (orangefs_cancel_op_in_progress(new_op))
237			return ret;
238
239		goto out;
240	}
241
242	/*
243	 * Stage 3: Post copy buffers from client-core's address space
244	 */
245	if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
246		/*
247		 * NOTE: the iovector can either contain addresses which
248		 *       can futher be kernel-space or user-space addresses.
249		 *       or it can pointers to struct page's
250		 */
251
252		copy_amount = new_op->downcall.resp.io.amt_complete;
253
254		ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
255			copy_amount);
256		if (ret < 0) {
257			gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
258			    __func__, (long)ret);
259			goto out;
260		}
261	}
262	gossip_debug(GOSSIP_FILE_DEBUG,
263	    "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
264	    __func__,
265	    handle,
266	    type == ORANGEFS_IO_READ ?  "read" : "written",
267	    (int)new_op->downcall.resp.io.amt_complete);
268
269	ret = new_op->downcall.resp.io.amt_complete;
270
271out:
272	if (buffer_index >= 0) {
273		orangefs_bufmap_put(buffer_index);
274		gossip_debug(GOSSIP_FILE_DEBUG,
275			"%s(%pU): PUT buffer_index %d\n",
276			__func__, handle, buffer_index);
 
277	}
278	op_release(new_op);
279	return ret;
280}
281
282int orangefs_revalidate_mapping(struct inode *inode)
 
 
 
 
 
 
 
 
283{
 
284	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
285	struct address_space *mapping = inode->i_mapping;
286	unsigned long *bitlock = &orangefs_inode->bitlock;
287	int ret;
 
288
289	while (1) {
290		ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
291		if (ret)
292			return ret;
293		spin_lock(&inode->i_lock);
294		if (test_bit(1, bitlock)) {
295			spin_unlock(&inode->i_lock);
296			continue;
297		}
298		if (!time_before(jiffies, orangefs_inode->mapping_time))
299			break;
300		spin_unlock(&inode->i_lock);
301		return 0;
 
 
 
 
 
 
302	}
303
304	set_bit(1, bitlock);
305	smp_wmb();
306	spin_unlock(&inode->i_lock);
 
 
 
 
307
308	unmap_mapping_range(mapping, 0, 0, 0);
309	ret = filemap_write_and_wait(mapping);
310	if (!ret)
311		ret = invalidate_inode_pages2(mapping);
 
 
 
 
 
 
312
313	orangefs_inode->mapping_time = jiffies +
314	    orangefs_cache_timeout_msecs*HZ/1000;
 
 
 
 
 
315
316	clear_bit(1, bitlock);
317	smp_mb__after_atomic();
318	wake_up_bit(bitlock, 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
319
320	return ret;
321}
322
323static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
324    struct iov_iter *iter)
 
 
 
 
 
 
325{
326	int ret;
 
 
 
 
327	orangefs_stats.reads++;
328
329	down_read(&file_inode(iocb->ki_filp)->i_rwsem);
330	ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
331	if (ret)
332		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
333
334	ret = generic_file_read_iter(iocb, iter);
335out:
336	up_read(&file_inode(iocb->ki_filp)->i_rwsem);
337	return ret;
338}
339
340static ssize_t orangefs_file_splice_read(struct file *in, loff_t *ppos,
341					 struct pipe_inode_info *pipe,
342					 size_t len, unsigned int flags)
343{
344	struct inode *inode = file_inode(in);
345	ssize_t ret;
 
 
 
 
 
346
347	orangefs_stats.reads++;
348
349	down_read(&inode->i_rwsem);
350	ret = orangefs_revalidate_mapping(inode);
351	if (ret)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
352		goto out;
 
 
 
 
353
354	ret = filemap_splice_read(in, ppos, pipe, len, flags);
355out:
356	up_read(&inode->i_rwsem);
357	return ret;
 
358}
359
360static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
361    struct iov_iter *iter)
 
 
362{
363	int ret;
364	orangefs_stats.writes++;
 
 
 
 
 
365
366	if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
367		ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
368		if (ret)
 
 
 
 
 
 
 
369			return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
370	}
371
372	ret = generic_file_write_iter(iocb, iter);
373	return ret;
374}
375
376static vm_fault_t orangefs_fault(struct vm_fault *vmf)
377{
378	struct file *file = vmf->vma->vm_file;
379	int ret;
380	ret = orangefs_inode_getattr(file->f_mapping->host,
381	    ORANGEFS_GETATTR_SIZE);
382	if (ret == -ESTALE)
383		ret = -EIO;
384	if (ret) {
385		gossip_err("%s: orangefs_inode_getattr failed, "
386		    "ret:%d:.\n", __func__, ret);
387		return VM_FAULT_SIGBUS;
388	}
389	return filemap_fault(vmf);
390}
391
392static const struct vm_operations_struct orangefs_file_vm_ops = {
393	.fault = orangefs_fault,
394	.map_pages = filemap_map_pages,
395	.page_mkwrite = orangefs_page_mkwrite,
396};
397
398/*
399 * Memory map a region of a file.
400 */
401static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
402{
403	int ret;
404
405	ret = orangefs_revalidate_mapping(file_inode(file));
406	if (ret)
407		return ret;
408
409	gossip_debug(GOSSIP_FILE_DEBUG,
410		     "orangefs_file_mmap: called on %pD\n", file);
 
 
 
 
 
 
411
412	/* set the sequential readahead hint */
413	vm_flags_mod(vma, VM_SEQ_READ, VM_RAND_READ);
 
414
415	file_accessed(file);
416	vma->vm_ops = &orangefs_file_vm_ops;
417	return 0;
418}
419
420#define mapping_nrpages(idata) ((idata)->nrpages)
421
422/*
423 * Called to notify the module that there are no more references to
424 * this file (i.e. no processes have it open).
425 *
426 * \note Not called when each file is closed.
427 */
428static int orangefs_file_release(struct inode *inode, struct file *file)
429{
430	gossip_debug(GOSSIP_FILE_DEBUG,
431		     "orangefs_file_release: called on %pD\n",
432		     file);
433
434	/*
435	 * remove all associated inode pages from the page cache and
436	 * readahead cache (if any); this forces an expensive refresh of
437	 * data for the next caller of mmap (or 'get_block' accesses)
438	 */
439	if (mapping_nrpages(file->f_mapping)) {
 
 
440		if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
441			gossip_debug(GOSSIP_INODE_DEBUG,
442			    "calling flush_racache on %pU\n",
443			    get_khandle_from_ino(inode));
444			flush_racache(inode);
445			gossip_debug(GOSSIP_INODE_DEBUG,
446			    "flush_racache finished\n");
447		}
448
 
449	}
450	return 0;
451}
452
453/*
454 * Push all data for a specific file onto permanent storage.
455 */
456static int orangefs_fsync(struct file *file,
457		       loff_t start,
458		       loff_t end,
459		       int datasync)
460{
461	int ret;
462	struct orangefs_inode_s *orangefs_inode =
463		ORANGEFS_I(file_inode(file));
464	struct orangefs_kernel_op_s *new_op = NULL;
465
466	ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
467	    start, end);
468	if (ret < 0)
469		return ret;
470
471	new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
472	if (!new_op)
473		return -ENOMEM;
474	new_op->upcall.req.fsync.refn = orangefs_inode->refn;
475
476	ret = service_operation(new_op,
477			"orangefs_fsync",
478			get_interruptible_flag(file_inode(file)));
479
480	gossip_debug(GOSSIP_FILE_DEBUG,
481		     "orangefs_fsync got return value of %d\n",
482		     ret);
483
484	op_release(new_op);
485	return ret;
486}
487
488/*
489 * Change the file pointer position for an instance of an open file.
490 *
491 * \note If .llseek is overriden, we must acquire lock as described in
492 *       Documentation/filesystems/locking.rst.
493 *
494 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
495 * require much changes to the FS
496 */
497static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
498{
499	int ret = -EINVAL;
500	struct inode *inode = file_inode(file);
501
502	if (origin == SEEK_END) {
503		/*
504		 * revalidate the inode's file size.
505		 * NOTE: We are only interested in file size here,
506		 * so we set mask accordingly.
507		 */
508		ret = orangefs_inode_getattr(file->f_mapping->host,
509		    ORANGEFS_GETATTR_SIZE);
510		if (ret == -ESTALE)
511			ret = -EIO;
512		if (ret) {
513			gossip_debug(GOSSIP_FILE_DEBUG,
514				     "%s:%s:%d calling make bad inode\n",
515				     __FILE__,
516				     __func__,
517				     __LINE__);
518			return ret;
519		}
520	}
521
522	gossip_debug(GOSSIP_FILE_DEBUG,
523		     "orangefs_file_llseek: offset is %ld | origin is %d"
524		     " | inode size is %lu\n",
525		     (long)offset,
526		     origin,
527		     (unsigned long)i_size_read(inode));
528
529	return generic_file_llseek(file, offset, origin);
530}
531
532/*
533 * Support local locks (locks that only this kernel knows about)
534 * if Orangefs was mounted -o local_lock.
535 */
536static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
537{
538	int rc = -EINVAL;
539
540	if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
541		if (cmd == F_GETLK) {
542			rc = 0;
543			posix_test_lock(filp, fl);
544		} else {
545			rc = posix_lock_file(filp, fl, NULL);
546		}
547	}
548
549	return rc;
550}
551
552static int orangefs_flush(struct file *file, fl_owner_t id)
553{
554	/*
555	 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
556	 * service_operation in orangefs_fsync.
557	 *
558	 * Do not send fsync to OrangeFS server on a close.  Do send fsync
559	 * on an explicit fsync call.  This duplicates historical OrangeFS
560	 * behavior.
561	 */
562	int r;
563
564	r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
565	if (r > 0)
566		return 0;
567	else
568		return r;
569}
570
571/** ORANGEFS implementation of VFS file operations */
572const struct file_operations orangefs_file_operations = {
573	.llseek		= orangefs_file_llseek,
574	.read_iter	= orangefs_file_read_iter,
575	.write_iter	= orangefs_file_write_iter,
576	.lock		= orangefs_lock,
 
577	.mmap		= orangefs_file_mmap,
578	.open		= generic_file_open,
579	.splice_read    = orangefs_file_splice_read,
580	.splice_write   = iter_file_splice_write,
581	.flush		= orangefs_flush,
582	.release	= orangefs_file_release,
583	.fsync		= orangefs_fsync,
584};
v4.17
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * (C) 2001 Clemson University and The University of Chicago
 
  4 *
  5 * See COPYING in top-level directory.
  6 */
  7
  8/*
  9 *  Linux VFS file operations.
 10 */
 11
 12#include "protocol.h"
 13#include "orangefs-kernel.h"
 14#include "orangefs-bufmap.h"
 15#include <linux/fs.h>
 
 16#include <linux/pagemap.h>
 17
 18static int flush_racache(struct inode *inode)
 19{
 20	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 21	struct orangefs_kernel_op_s *new_op;
 22	int ret;
 23
 24	gossip_debug(GOSSIP_UTILS_DEBUG,
 25	    "%s: %pU: Handle is %pU | fs_id %d\n", __func__,
 26	    get_khandle_from_ino(inode), &orangefs_inode->refn.khandle,
 27	    orangefs_inode->refn.fs_id);
 28
 29	new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
 30	if (!new_op)
 31		return -ENOMEM;
 32	new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
 33
 34	ret = service_operation(new_op, "orangefs_flush_racache",
 35	    get_interruptible_flag(inode));
 36
 37	gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
 38	    __func__, ret);
 39
 40	op_release(new_op);
 41	return ret;
 42}
 43
 44/*
 45 * Post and wait for the I/O upcall to finish
 46 */
 47static 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)
 
 50{
 51	struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
 52	struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
 53	struct orangefs_kernel_op_s *new_op = NULL;
 54	int buffer_index = -1;
 55	ssize_t ret;
 
 
 
 56
 57	new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
 58	if (!new_op)
 59		return -ENOMEM;
 60
 61	/* synchronous I/O */
 62	new_op->upcall.req.io.readahead_size = readahead_size;
 63	new_op->upcall.req.io.io_type = type;
 64	new_op->upcall.req.io.refn = orangefs_inode->refn;
 65
 66populate_shared_memory:
 67	/* get a shared buffer index */
 68	buffer_index = orangefs_bufmap_get();
 69	if (buffer_index < 0) {
 70		ret = buffer_index;
 71		gossip_debug(GOSSIP_FILE_DEBUG,
 72			     "%s: orangefs_bufmap_get failure (%zd)\n",
 73			     __func__, ret);
 74		goto out;
 75	}
 76	gossip_debug(GOSSIP_FILE_DEBUG,
 77		     "%s(%pU): GET op %p -> buffer_index %d\n",
 78		     __func__,
 79		     handle,
 80		     new_op,
 81		     buffer_index);
 82
 83	new_op->uses_shared_memory = 1;
 84	new_op->upcall.req.io.buf_index = buffer_index;
 85	new_op->upcall.req.io.count = total_size;
 86	new_op->upcall.req.io.offset = *offset;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 87
 88	gossip_debug(GOSSIP_FILE_DEBUG,
 89		     "%s(%pU): offset: %llu total_size: %zd\n",
 90		     __func__,
 91		     handle,
 92		     llu(*offset),
 93		     total_size);
 94	/*
 95	 * Stage 1: copy the buffers into client-core's address space
 96	 */
 97	if (type == ORANGEFS_IO_WRITE && total_size) {
 98		ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
 99		    total_size);
100		if (ret < 0) {
101			gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
102			    __func__, (long)ret);
103			goto out;
104		}
105	}
106
107	gossip_debug(GOSSIP_FILE_DEBUG,
108		     "%s(%pU): Calling post_io_request with tag (%llu)\n",
109		     __func__,
110		     handle,
111		     llu(new_op->tag));
112
113	/* Stage 2: Service the I/O operation */
114	ret = service_operation(new_op,
115				type == ORANGEFS_IO_WRITE ?
116					"file_write" :
117					"file_read",
118				get_interruptible_flag(inode));
119
120	/*
121	 * If service_operation() returns -EAGAIN #and# the operation was
122	 * purged from orangefs_request_list or htable_ops_in_progress, then
123	 * we know that the client was restarted, causing the shared memory
124	 * area to be wiped clean.  To restart a  write operation in this
125	 * case, we must re-copy the data from the user's iovec to a NEW
126	 * shared memory location. To restart a read operation, we must get
127	 * a new shared memory location.
128	 */
129	if (ret == -EAGAIN && op_state_purged(new_op)) {
130		orangefs_bufmap_put(buffer_index);
131		buffer_index = -1;
132		if (type == ORANGEFS_IO_WRITE)
133			iov_iter_revert(iter, total_size);
134		gossip_debug(GOSSIP_FILE_DEBUG,
135			     "%s:going to repopulate_shared_memory.\n",
136			     __func__);
137		goto populate_shared_memory;
138	}
139
140	if (ret < 0) {
141		if (ret == -EINTR) {
142			/*
143			 * We can't return EINTR if any data was written,
144			 * it's not POSIX. It is minimally acceptable
145			 * to give a partial write, the way NFS does.
146			 *
147			 * It would be optimal to return all or nothing,
148			 * but if a userspace write is bigger than
149			 * an IO buffer, and the interrupt occurs
150			 * between buffer writes, that would not be
151			 * possible.
152			 */
153			switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
154			/*
155			 * If the op was waiting when the interrupt
156			 * occurred, then the client-core did not
157			 * trigger the write.
158			 */
159			case OP_VFS_STATE_WAITING:
160				if (*offset == 0)
161					ret = -EINTR;
162				else
163					ret = 0;
164				break;
165			/* 
166			 * If the op was in progress when the interrupt
167			 * occurred, then the client-core was able to
168			 * trigger the write.
169			 */
170			case OP_VFS_STATE_INPROGR:
171				ret = total_size;
 
 
 
172				break;
173			default:
174				gossip_err("%s: unexpected op state :%d:.\n",
175					   __func__,
176					   new_op->op_state);
177				ret = 0;
178				break;
179			}
180			gossip_debug(GOSSIP_FILE_DEBUG,
181				     "%s: got EINTR, state:%d: %p\n",
182				     __func__,
183				     new_op->op_state,
184				     new_op);
185		} else {
186			gossip_err("%s: error in %s handle %pU, returning %zd\n",
187				__func__,
188				type == ORANGEFS_IO_READ ?
189					"read from" : "write to",
190				handle, ret);
191		}
192		if (orangefs_cancel_op_in_progress(new_op))
193			return ret;
194
195		goto out;
196	}
197
198	/*
199	 * Stage 3: Post copy buffers from client-core's address space
200	 */
201	if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
202		/*
203		 * NOTE: the iovector can either contain addresses which
204		 *       can futher be kernel-space or user-space addresses.
205		 *       or it can pointers to struct page's
206		 */
 
 
 
207		ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
208		    new_op->downcall.resp.io.amt_complete);
209		if (ret < 0) {
210			gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
211			    __func__, (long)ret);
212			goto out;
213		}
214	}
215	gossip_debug(GOSSIP_FILE_DEBUG,
216	    "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
217	    __func__,
218	    handle,
219	    type == ORANGEFS_IO_READ ?  "read" : "written",
220	    (int)new_op->downcall.resp.io.amt_complete);
221
222	ret = new_op->downcall.resp.io.amt_complete;
223
224out:
225	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;
231	}
232	op_release(new_op);
233	return ret;
234}
235
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 */
243static ssize_t do_readv_writev(enum ORANGEFS_io_type type, struct file *file,
244		loff_t *offset, struct iov_iter *iter)
245{
246	struct inode *inode = file->f_mapping->host;
247	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;
252
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)
319			break;
320	} /*end while */
321
322out:
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
347	return ret;
348}
349
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 */
354ssize_t orangefs_inode_read(struct inode *inode,
355			    struct iov_iter *iter,
356			    loff_t *offset,
357			    loff_t readahead_size)
358{
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
395static 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
413static 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;
464	orangefs_stats.writes++;
465
 
466out:
467
468	inode_unlock(file->f_mapping->host);
469	return rc;
470}
471
472/*
473 * Perform a miscellaneous operation on a file.
474 */
475static long orangefs_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
476{
477	int ret = -ENOTTY;
478	__u64 val = 0;
479	unsigned long uval;
480
481	gossip_debug(GOSSIP_FILE_DEBUG,
482		     "orangefs_ioctl: called with cmd %d\n",
483		     cmd);
484
485	/*
486	 * we understand some general ioctls on files, such as the immutable
487	 * and append flags
488	 */
489	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)
495			return ret;
496		else if (ret == -ENODATA)
497			val = 0;
498		uval = val;
499		gossip_debug(GOSSIP_FILE_DEBUG,
500			     "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
501			     (unsigned long long)uval);
502		return put_user(uval, (int __user *)arg);
503	} else if (cmd == FS_IOC_SETFLAGS) {
504		ret = 0;
505		if (get_user(uval, (int __user *)arg))
506			return -EFAULT;
507		/*
508		 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
509		 * is turned on for a file. The user is not allowed to turn
510		 * on this bit, but the bit is present if the user first gets
511		 * the flags and then updates the flags with some new
512		 * settings. So, we ignore it in the following edit. bligon.
513		 */
514		if ((uval & ~ORANGEFS_MIRROR_FL) &
515		    (~(FS_IMMUTABLE_FL | FS_APPEND_FL | FS_NOATIME_FL))) {
516			gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
517			return -EINVAL;
518		}
519		val = uval;
520		gossip_debug(GOSSIP_FILE_DEBUG,
521			     "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
522			     (unsigned long long)val);
523		ret = orangefs_inode_setxattr(file_inode(file),
524					      "user.pvfs2.meta_hint",
525					      &val, sizeof(val), 0);
526	}
527
 
528	return ret;
529}
530
531static int orangefs_fault(struct vm_fault *vmf)
532{
533	struct file *file = vmf->vma->vm_file;
534	int rc;
535	rc = orangefs_inode_getattr(file->f_mapping->host, 0, 1,
536	    STATX_SIZE);
537	if (rc == -ESTALE)
538		rc = -EIO;
539	if (rc) {
540		gossip_err("%s: orangefs_inode_getattr failed, "
541		    "rc:%d:.\n", __func__, rc);
542		return rc;
543	}
544	return filemap_fault(vmf);
545}
546
547const struct vm_operations_struct orangefs_file_vm_ops = {
548	.fault = orangefs_fault,
549	.map_pages = filemap_map_pages,
550	.page_mkwrite = filemap_page_mkwrite,
551};
552
553/*
554 * Memory map a region of a file.
555 */
556static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
557{
 
 
 
 
 
 
558	gossip_debug(GOSSIP_FILE_DEBUG,
559		     "orangefs_file_mmap: called on %s\n",
560		     (file ?
561			(char *)file->f_path.dentry->d_name.name :
562			(char *)"Unknown"));
563
564	if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
565		return -EINVAL;
566
567	/* set the sequential readahead hint */
568	vma->vm_flags |= VM_SEQ_READ;
569	vma->vm_flags &= ~VM_RAND_READ;
570
571	file_accessed(file);
572	vma->vm_ops = &orangefs_file_vm_ops;
573	return 0;
574}
575
576#define mapping_nrpages(idata) ((idata)->nrpages)
577
578/*
579 * Called to notify the module that there are no more references to
580 * this file (i.e. no processes have it open).
581 *
582 * \note Not called when each file is closed.
583 */
584static int orangefs_file_release(struct inode *inode, struct file *file)
585{
586	gossip_debug(GOSSIP_FILE_DEBUG,
587		     "orangefs_file_release: called on %pD\n",
588		     file);
589
590	/*
591	 * remove all associated inode pages from the page cache and
592	 * readahead cache (if any); this forces an expensive refresh of
593	 * data for the next caller of mmap (or 'get_block' accesses)
594	 */
595	if (file_inode(file) &&
596	    file_inode(file)->i_mapping &&
597	    mapping_nrpages(&file_inode(file)->i_data)) {
598		if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
599			gossip_debug(GOSSIP_INODE_DEBUG,
600			    "calling flush_racache on %pU\n",
601			    get_khandle_from_ino(inode));
602			flush_racache(inode);
603			gossip_debug(GOSSIP_INODE_DEBUG,
604			    "flush_racache finished\n");
605		}
606		truncate_inode_pages(file_inode(file)->i_mapping,
607				     0);
608	}
609	return 0;
610}
611
612/*
613 * Push all data for a specific file onto permanent storage.
614 */
615static int orangefs_fsync(struct file *file,
616		       loff_t start,
617		       loff_t end,
618		       int datasync)
619{
620	int ret;
621	struct orangefs_inode_s *orangefs_inode =
622		ORANGEFS_I(file_inode(file));
623	struct orangefs_kernel_op_s *new_op = NULL;
624
 
 
 
 
 
625	new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
626	if (!new_op)
627		return -ENOMEM;
628	new_op->upcall.req.fsync.refn = orangefs_inode->refn;
629
630	ret = service_operation(new_op,
631			"orangefs_fsync",
632			get_interruptible_flag(file_inode(file)));
633
634	gossip_debug(GOSSIP_FILE_DEBUG,
635		     "orangefs_fsync got return value of %d\n",
636		     ret);
637
638	op_release(new_op);
639	return ret;
640}
641
642/*
643 * Change the file pointer position for an instance of an open file.
644 *
645 * \note If .llseek is overriden, we must acquire lock as described in
646 *       Documentation/filesystems/Locking.
647 *
648 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
649 * require much changes to the FS
650 */
651static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
652{
653	int ret = -EINVAL;
654	struct inode *inode = file_inode(file);
655
656	if (origin == SEEK_END) {
657		/*
658		 * revalidate the inode's file size.
659		 * NOTE: We are only interested in file size here,
660		 * so we set mask accordingly.
661		 */
662		ret = orangefs_inode_getattr(file->f_mapping->host, 0, 1,
663		    STATX_SIZE);
664		if (ret == -ESTALE)
665			ret = -EIO;
666		if (ret) {
667			gossip_debug(GOSSIP_FILE_DEBUG,
668				     "%s:%s:%d calling make bad inode\n",
669				     __FILE__,
670				     __func__,
671				     __LINE__);
672			return ret;
673		}
674	}
675
676	gossip_debug(GOSSIP_FILE_DEBUG,
677		     "orangefs_file_llseek: offset is %ld | origin is %d"
678		     " | inode size is %lu\n",
679		     (long)offset,
680		     origin,
681		     (unsigned long)i_size_read(inode));
682
683	return generic_file_llseek(file, offset, origin);
684}
685
686/*
687 * Support local locks (locks that only this kernel knows about)
688 * if Orangefs was mounted -o local_lock.
689 */
690static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
691{
692	int rc = -EINVAL;
693
694	if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
695		if (cmd == F_GETLK) {
696			rc = 0;
697			posix_test_lock(filp, fl);
698		} else {
699			rc = posix_lock_file(filp, fl, NULL);
700		}
701	}
702
703	return rc;
704}
705
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
706/** ORANGEFS implementation of VFS file operations */
707const struct file_operations orangefs_file_operations = {
708	.llseek		= orangefs_file_llseek,
709	.read_iter	= orangefs_file_read_iter,
710	.write_iter	= orangefs_file_write_iter,
711	.lock		= orangefs_lock,
712	.unlocked_ioctl	= orangefs_ioctl,
713	.mmap		= orangefs_file_mmap,
714	.open		= generic_file_open,
 
 
 
715	.release	= orangefs_file_release,
716	.fsync		= orangefs_fsync,
717};