2 * (C) 2001 Clemson University and The University of Chicago
4 * See COPYING in top-level directory.
8 * Linux VFS file operations.
12 #include "orangefs-kernel.h"
13 #include "orangefs-bufmap.h"
15 #include <linux/pagemap.h>
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
24 static int precopy_buffers(int buffer_index
,
25 struct iov_iter
*iter
,
30 * copy data from application/kernel by pulling it out
36 ret
= orangefs_bufmap_copy_from_iovec(iter
,
40 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
46 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
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
59 static int postcopy_buffers(int buffer_index
,
60 struct iov_iter
*iter
,
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.
70 ret
= orangefs_bufmap_copy_to_iovec(iter
,
74 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
82 * Post and wait for the I/O upcall to finish
84 static 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
)
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;
95 new_op
= op_alloc(ORANGEFS_VFS_OP_FILE_IO
);
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
;
104 populate_shared_memory
:
105 /* get a shared buffer index */
106 buffer_index
= orangefs_bufmap_get();
107 if (buffer_index
< 0) {
109 gossip_debug(GOSSIP_FILE_DEBUG
,
110 "%s: orangefs_bufmap_get failure (%zd)\n",
114 gossip_debug(GOSSIP_FILE_DEBUG
,
115 "%s(%pU): GET op %p -> buffer_index %d\n",
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
;
126 gossip_debug(GOSSIP_FILE_DEBUG
,
127 "%s(%pU): offset: %llu total_size: %zd\n",
133 * Stage 1: copy the buffers into client-core's address space
134 * precopy_buffers only pertains to writes.
136 if (type
== ORANGEFS_IO_WRITE
) {
137 ret
= precopy_buffers(buffer_index
,
144 gossip_debug(GOSSIP_FILE_DEBUG
,
145 "%s(%pU): Calling post_io_request with tag (%llu)\n",
150 /* Stage 2: Service the I/O operation */
151 ret
= service_operation(new_op
,
152 type
== ORANGEFS_IO_WRITE
?
155 get_interruptible_flag(inode
));
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.
166 if (ret
== -EAGAIN
&& op_state_purged(new_op
)) {
167 orangefs_bufmap_put(buffer_index
);
169 if (type
== ORANGEFS_IO_WRITE
)
171 gossip_debug(GOSSIP_FILE_DEBUG
,
172 "%s:going to repopulate_shared_memory.\n",
174 goto populate_shared_memory
;
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.
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
190 switch (new_op
->op_state
- OP_VFS_STATE_GIVEN_UP
) {
192 * If the op was waiting when the interrupt
193 * occurred, then the client-core did not
196 case OP_VFS_STATE_WAITING
:
203 * If the op was in progress when the interrupt
204 * occurred, then the client-core was able to
207 case OP_VFS_STATE_INPROGR
:
211 gossip_err("%s: unexpected op state :%d:.\n",
217 gossip_debug(GOSSIP_FILE_DEBUG
,
218 "%s: got EINTR, state:%d: %p\n",
223 gossip_err("%s: error in %s handle %pU, returning %zd\n",
225 type
== ORANGEFS_IO_READ
?
226 "read from" : "write to",
229 if (orangefs_cancel_op_in_progress(new_op
))
236 * Stage 3: Post copy buffers from client-core's address space
237 * postcopy_buffers only pertains to reads.
239 if (type
== ORANGEFS_IO_READ
) {
240 ret
= postcopy_buffers(buffer_index
,
242 new_op
->downcall
.resp
.io
.amt_complete
);
246 gossip_debug(GOSSIP_FILE_DEBUG
,
247 "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
250 type
== ORANGEFS_IO_READ
? "read" : "written",
251 (int)new_op
->downcall
.resp
.io
.amt_complete
);
253 ret
= new_op
->downcall
.resp
.io
.amt_complete
;
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
);
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.
274 static ssize_t
do_readv_writev(enum ORANGEFS_io_type type
, struct file
*file
,
275 loff_t
*offset
, struct iov_iter
*iter
)
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
;
284 gossip_debug(GOSSIP_FILE_DEBUG
,
285 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
290 if (type
== ORANGEFS_IO_WRITE
) {
291 gossip_debug(GOSSIP_FILE_DEBUG
,
292 "%s(%pU): proceeding with offset : %llu, "
305 while (iov_iter_count(iter
)) {
306 size_t each_count
= iov_iter_count(iter
);
309 /* how much to transfer in this loop iteration */
310 if (each_count
> orangefs_bufmap_size_query())
311 each_count
= orangefs_bufmap_size_query();
313 gossip_debug(GOSSIP_FILE_DEBUG
,
314 "%s(%pU): size of each_count(%d)\n",
318 gossip_debug(GOSSIP_FILE_DEBUG
,
319 "%s(%pU): BEFORE wait_for_io: offset is %d\n",
324 ret
= wait_for_direct_io(type
, inode
, offset
, iter
,
326 gossip_debug(GOSSIP_FILE_DEBUG
,
327 "%s(%pU): return from wait_for_io:%d\n",
339 gossip_debug(GOSSIP_FILE_DEBUG
,
340 "%s(%pU): AFTER wait_for_io: offset is %d\n",
346 * if we got a short I/O operations,
347 * fall out and return what we got so far
349 if (amt_complete
< each_count
)
357 if (type
== ORANGEFS_IO_READ
) {
360 SetMtimeFlag(orangefs_inode
);
361 inode
->i_mtime
= CURRENT_TIME
;
362 mark_inode_dirty_sync(inode
);
366 gossip_debug(GOSSIP_FILE_DEBUG
,
367 "%s(%pU): Value(%d) returned.\n",
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.
379 ssize_t
orangefs_inode_read(struct inode
*inode
,
380 struct iov_iter
*iter
,
382 loff_t readahead_size
)
384 struct orangefs_inode_s
*orangefs_inode
= ORANGEFS_I(inode
);
385 size_t count
= iov_iter_count(iter
);
387 ssize_t ret
= -EINVAL
;
389 g_orangefs_stats
.reads
++;
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
);
399 gossip_debug(GOSSIP_FILE_DEBUG
,
400 "%s(%pU) %zd@%llu\n",
402 &orangefs_inode
->refn
.khandle
,
406 ret
= wait_for_direct_io(ORANGEFS_IO_READ
, inode
, offset
, iter
,
407 count
, readahead_size
);
411 gossip_debug(GOSSIP_FILE_DEBUG
,
412 "%s(%pU): Value(%zd) returned.\n",
414 &orangefs_inode
->refn
.khandle
,
420 static ssize_t
orangefs_file_read_iter(struct kiocb
*iocb
, struct iov_iter
*iter
)
422 struct file
*file
= iocb
->ki_filp
;
423 loff_t pos
= *(&iocb
->ki_pos
);
426 BUG_ON(iocb
->private);
428 gossip_debug(GOSSIP_FILE_DEBUG
, "orangefs_file_read_iter\n");
430 g_orangefs_stats
.reads
++;
432 rc
= do_readv_writev(ORANGEFS_IO_READ
, file
, &pos
, iter
);
438 static ssize_t
orangefs_file_write_iter(struct kiocb
*iocb
, struct iov_iter
*iter
)
440 struct file
*file
= iocb
->ki_filp
;
444 BUG_ON(iocb
->private);
446 gossip_debug(GOSSIP_FILE_DEBUG
, "orangefs_file_write_iter\n");
448 mutex_lock(&file
->f_mapping
->host
->i_mutex
);
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);
456 gossip_err("%s: orangefs_inode_getattr failed, "
457 "rc:%zd:.\n", __func__
, rc
);
462 if (file
->f_pos
> i_size_read(file
->f_mapping
->host
))
463 orangefs_i_size_write(file
->f_mapping
->host
, file
->f_pos
);
465 rc
= generic_write_checks(iocb
, iter
);
468 gossip_err("%s: generic_write_checks failed, rc:%zd:.\n",
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
478 pos
= *(&iocb
->ki_pos
);
480 rc
= do_readv_writev(ORANGEFS_IO_WRITE
,
485 gossip_err("%s: do_readv_writev failed, rc:%zd:.\n",
491 g_orangefs_stats
.writes
++;
495 mutex_unlock(&file
->f_mapping
->host
->i_mutex
);
500 * Perform a miscellaneous operation on a file.
502 static long orangefs_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
508 gossip_debug(GOSSIP_FILE_DEBUG
,
509 "orangefs_ioctl: called with cmd %d\n",
513 * we understand some general ioctls on files, such as the immutable
516 if (cmd
== FS_IOC_GETFLAGS
) {
518 ret
= orangefs_inode_getxattr(file_inode(file
),
519 ORANGEFS_XATTR_NAME_DEFAULT_PREFIX
,
520 "user.pvfs2.meta_hint",
522 if (ret
< 0 && ret
!= -ENODATA
)
524 else if (ret
== -ENODATA
)
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
) {
533 if (get_user(uval
, (int __user
*)arg
))
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.
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");
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);
561 * Memory map a region of a file.
563 static int orangefs_file_mmap(struct file
*file
, struct vm_area_struct
*vma
)
565 gossip_debug(GOSSIP_FILE_DEBUG
,
566 "orangefs_file_mmap: called on %s\n",
568 (char *)file
->f_path
.dentry
->d_name
.name
:
571 /* set the sequential readahead hint */
572 vma
->vm_flags
|= VM_SEQ_READ
;
573 vma
->vm_flags
&= ~VM_RAND_READ
;
575 /* Use readonly mmap since we cannot support writable maps. */
576 return generic_file_readonly_mmap(file
, vma
);
579 #define mapping_nrpages(idata) ((idata)->nrpages)
582 * Called to notify the module that there are no more references to
583 * this file (i.e. no processes have it open).
585 * \note Not called when each file is closed.
587 static int orangefs_file_release(struct inode
*inode
, struct file
*file
)
589 gossip_debug(GOSSIP_FILE_DEBUG
,
590 "orangefs_file_release: called on %s\n",
591 file
->f_path
.dentry
->d_name
.name
);
593 orangefs_flush_inode(inode
);
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)
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
,
609 * Push all data for a specific file onto permanent storage.
611 static int orangefs_fsync(struct file
*file
,
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
;
622 filemap_write_and_wait_range(file
->f_mapping
, start
, end
);
624 new_op
= op_alloc(ORANGEFS_VFS_OP_FSYNC
);
627 new_op
->upcall
.req
.fsync
.refn
= orangefs_inode
->refn
;
629 ret
= service_operation(new_op
,
631 get_interruptible_flag(file
->f_path
.dentry
->d_inode
));
633 gossip_debug(GOSSIP_FILE_DEBUG
,
634 "orangefs_fsync got return value of %d\n",
639 orangefs_flush_inode(file
->f_path
.dentry
->d_inode
);
644 * Change the file pointer position for an instance of an open file.
646 * \note If .llseek is overriden, we must acquire lock as described in
647 * Documentation/filesystems/Locking.
649 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
650 * require much changes to the FS
652 static loff_t
orangefs_file_llseek(struct file
*file
, loff_t offset
, int origin
)
655 struct inode
*inode
= file_inode(file
);
657 if (origin
== SEEK_END
) {
659 * revalidate the inode's file size.
660 * NOTE: We are only interested in file size here,
661 * so we set mask accordingly.
663 ret
= orangefs_inode_getattr(file
->f_mapping
->host
, 0, 1);
667 gossip_debug(GOSSIP_FILE_DEBUG
,
668 "%s:%s:%d calling make bad inode\n",
676 gossip_debug(GOSSIP_FILE_DEBUG
,
677 "orangefs_file_llseek: offset is %ld | origin is %d"
678 " | inode size is %lu\n",
681 (unsigned long)i_size_read(inode
));
683 return generic_file_llseek(file
, offset
, origin
);
687 * Support local locks (locks that only this kernel knows about)
688 * if Orangefs was mounted -o local_lock.
690 static int orangefs_lock(struct file
*filp
, int cmd
, struct file_lock
*fl
)
694 if (ORANGEFS_SB(filp
->f_inode
->i_sb
)->flags
& ORANGEFS_OPT_LOCAL_LOCK
) {
695 if (cmd
== F_GETLK
) {
697 posix_test_lock(filp
, fl
);
699 rc
= posix_lock_file(filp
, fl
, NULL
);
706 /** ORANGEFS implementation of VFS file operations */
707 const 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
,