Linux Audio

Check our new training course

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