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(struct orangefs_bufmap
*bufmap
,
26 struct iov_iter
*iter
,
31 * copy data from application/kernel by pulling it out
37 ret
= orangefs_bufmap_copy_from_iovec(bufmap
,
42 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
48 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
55 * Copy from client-core's address space to the buffers specified
56 * by the iovec upto total_size bytes.
57 * NOTE: the iovector can either contain addresses which
58 * can futher be kernel-space or user-space addresses.
59 * or it can pointers to struct page's
61 static int postcopy_buffers(struct orangefs_bufmap
*bufmap
,
63 struct iov_iter
*iter
,
68 * copy data to application/kernel by pushing it out to
69 * the iovec. NOTE; target buffers can be addresses or
70 * struct page pointers.
73 ret
= orangefs_bufmap_copy_to_iovec(bufmap
,
78 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
86 * handles two possible error cases, depending on context.
88 * by design, our vfs i/o errors need to be handled in one of two ways,
89 * depending on where the error occured.
91 * if the error happens in the waitqueue code because we either timed
92 * out or a signal was raised while waiting, we need to cancel the
93 * userspace i/o operation and free the op manually. this is done to
94 * avoid having the device start writing application data to our shared
95 * bufmap pages without us expecting it.
97 * FIXME: POSSIBLE OPTIMIZATION:
98 * However, if we timed out or if we got a signal AND our upcall was never
99 * picked off the queue (i.e. we were in OP_VFS_STATE_WAITING), then we don't
100 * need to send a cancellation upcall. The way we can handle this is
101 * set error_exit to 2 in such cases and 1 whenever cancellation has to be
102 * sent and have handle_error
103 * take care of this situation as well..
105 * if a orangefs sysint level error occured and i/o has been completed,
106 * there is no need to cancel the operation, as the user has finished
107 * using the bufmap page and so there is no danger in this case. in
108 * this case, we wake up the device normally so that it may free the
111 * note the only reason this is a macro is because both read and write
112 * cases need the exact same handling code.
114 #define handle_io_error() \
116 if (!op_state_serviced(new_op)) { \
117 orangefs_cancel_op_in_progress(new_op->tag); \
119 complete(&new_op->done); \
121 orangefs_bufmap_put(bufmap, buffer_index); \
126 * Post and wait for the I/O upcall to finish
128 static ssize_t
wait_for_direct_io(enum ORANGEFS_io_type type
, struct inode
*inode
,
129 loff_t
*offset
, struct iov_iter
*iter
,
130 size_t total_size
, loff_t readahead_size
)
132 struct orangefs_inode_s
*orangefs_inode
= ORANGEFS_I(inode
);
133 struct orangefs_khandle
*handle
= &orangefs_inode
->refn
.khandle
;
134 struct orangefs_bufmap
*bufmap
= NULL
;
135 struct orangefs_kernel_op_s
*new_op
= NULL
;
136 int buffer_index
= -1;
139 new_op
= op_alloc(ORANGEFS_VFS_OP_FILE_IO
);
143 /* synchronous I/O */
144 new_op
->upcall
.req
.io
.async_vfs_io
= ORANGEFS_VFS_SYNC_IO
;
145 new_op
->upcall
.req
.io
.readahead_size
= readahead_size
;
146 new_op
->upcall
.req
.io
.io_type
= type
;
147 new_op
->upcall
.req
.io
.refn
= orangefs_inode
->refn
;
149 populate_shared_memory
:
150 /* get a shared buffer index */
151 ret
= orangefs_bufmap_get(&bufmap
, &buffer_index
);
153 gossip_debug(GOSSIP_FILE_DEBUG
,
154 "%s: orangefs_bufmap_get failure (%ld)\n",
155 __func__
, (long)ret
);
158 gossip_debug(GOSSIP_FILE_DEBUG
,
159 "%s(%pU): GET op %p -> buffer_index %d\n",
165 new_op
->uses_shared_memory
= 1;
166 new_op
->upcall
.req
.io
.buf_index
= buffer_index
;
167 new_op
->upcall
.req
.io
.count
= total_size
;
168 new_op
->upcall
.req
.io
.offset
= *offset
;
170 gossip_debug(GOSSIP_FILE_DEBUG
,
171 "%s(%pU): offset: %llu total_size: %zd\n",
177 * Stage 1: copy the buffers into client-core's address space
178 * precopy_buffers only pertains to writes.
180 if (type
== ORANGEFS_IO_WRITE
) {
181 ret
= precopy_buffers(bufmap
,
189 gossip_debug(GOSSIP_FILE_DEBUG
,
190 "%s(%pU): Calling post_io_request with tag (%llu)\n",
195 /* Stage 2: Service the I/O operation */
196 ret
= service_operation(new_op
,
197 type
== ORANGEFS_IO_WRITE
?
200 get_interruptible_flag(inode
));
203 * If service_operation() returns -EAGAIN #and# the operation was
204 * purged from orangefs_request_list or htable_ops_in_progress, then
205 * we know that the client was restarted, causing the shared memory
206 * area to be wiped clean. To restart a write operation in this
207 * case, we must re-copy the data from the user's iovec to a NEW
208 * shared memory location. To restart a read operation, we must get
209 * a new shared memory location.
211 if (ret
== -EAGAIN
&& op_state_purged(new_op
)) {
212 orangefs_bufmap_put(bufmap
, buffer_index
);
213 gossip_debug(GOSSIP_FILE_DEBUG
,
214 "%s:going to repopulate_shared_memory.\n",
216 goto populate_shared_memory
;
222 * don't write an error to syslog on signaled operation
223 * termination unless we've got debugging turned on, as
224 * this can happen regularly (i.e. ctrl-c)
227 gossip_debug(GOSSIP_FILE_DEBUG
,
228 "%s: returning error %ld\n", __func__
,
231 gossip_err("%s: error in %s handle %pU, returning %zd\n",
233 type
== ORANGEFS_IO_READ
?
234 "read from" : "write to",
240 * Stage 3: Post copy buffers from client-core's address space
241 * postcopy_buffers only pertains to reads.
243 if (type
== ORANGEFS_IO_READ
) {
244 ret
= postcopy_buffers(bufmap
,
247 new_op
->downcall
.resp
.io
.amt_complete
);
250 * put error codes in downcall so that handle_io_error()
251 * preserves it properly
253 new_op
->downcall
.status
= ret
;
258 gossip_debug(GOSSIP_FILE_DEBUG
,
259 "%s(%pU): Amount written as returned by the sys-io call:%d\n",
262 (int)new_op
->downcall
.resp
.io
.amt_complete
);
264 ret
= new_op
->downcall
.resp
.io
.amt_complete
;
267 * tell the device file owner waiting on I/O that this read has
268 * completed and it can return now.
270 complete(&new_op
->done
);
273 if (buffer_index
>= 0) {
274 orangefs_bufmap_put(bufmap
, buffer_index
);
275 gossip_debug(GOSSIP_FILE_DEBUG
,
276 "%s(%pU): PUT buffer_index %d\n",
277 __func__
, handle
, buffer_index
);
285 * Common entry point for read/write/readv/writev
286 * This function will dispatch it to either the direct I/O
287 * or buffered I/O path depending on the mount options and/or
288 * augmented/extended metadata attached to the file.
289 * Note: File extended attributes override any mount options.
291 static ssize_t
do_readv_writev(enum ORANGEFS_io_type type
, struct file
*file
,
292 loff_t
*offset
, struct iov_iter
*iter
)
294 struct inode
*inode
= file
->f_mapping
->host
;
295 struct orangefs_inode_s
*orangefs_inode
= ORANGEFS_I(inode
);
296 struct orangefs_khandle
*handle
= &orangefs_inode
->refn
.khandle
;
297 size_t count
= iov_iter_count(iter
);
298 ssize_t total_count
= 0;
299 ssize_t ret
= -EINVAL
;
301 gossip_debug(GOSSIP_FILE_DEBUG
,
302 "%s-BEGIN(%pU): count(%d) after estimate_max_iovecs.\n",
307 if (type
== ORANGEFS_IO_WRITE
) {
308 gossip_debug(GOSSIP_FILE_DEBUG
,
309 "%s(%pU): proceeding with offset : %llu, "
322 while (iov_iter_count(iter
)) {
323 size_t each_count
= iov_iter_count(iter
);
326 /* how much to transfer in this loop iteration */
327 if (each_count
> orangefs_bufmap_size_query())
328 each_count
= orangefs_bufmap_size_query();
330 gossip_debug(GOSSIP_FILE_DEBUG
,
331 "%s(%pU): size of each_count(%d)\n",
335 gossip_debug(GOSSIP_FILE_DEBUG
,
336 "%s(%pU): BEFORE wait_for_io: offset is %d\n",
341 ret
= wait_for_direct_io(type
, inode
, offset
, iter
,
343 gossip_debug(GOSSIP_FILE_DEBUG
,
344 "%s(%pU): return from wait_for_io:%d\n",
356 gossip_debug(GOSSIP_FILE_DEBUG
,
357 "%s(%pU): AFTER wait_for_io: offset is %d\n",
363 * if we got a short I/O operations,
364 * fall out and return what we got so far
366 if (amt_complete
< each_count
)
374 if (type
== ORANGEFS_IO_READ
) {
377 SetMtimeFlag(orangefs_inode
);
378 inode
->i_mtime
= CURRENT_TIME
;
379 mark_inode_dirty_sync(inode
);
383 gossip_debug(GOSSIP_FILE_DEBUG
,
384 "%s(%pU): Value(%d) returned.\n",
393 * Read data from a specified offset in a file (referenced by inode).
394 * Data may be placed either in a user or kernel buffer.
396 ssize_t
orangefs_inode_read(struct inode
*inode
,
397 struct iov_iter
*iter
,
399 loff_t readahead_size
)
401 struct orangefs_inode_s
*orangefs_inode
= ORANGEFS_I(inode
);
402 size_t count
= iov_iter_count(iter
);
404 ssize_t ret
= -EINVAL
;
406 g_orangefs_stats
.reads
++;
408 bufmap_size
= orangefs_bufmap_size_query();
409 if (count
> bufmap_size
) {
410 gossip_debug(GOSSIP_FILE_DEBUG
,
411 "%s: count is too large (%zd/%zd)!\n",
412 __func__
, count
, bufmap_size
);
416 gossip_debug(GOSSIP_FILE_DEBUG
,
417 "%s(%pU) %zd@%llu\n",
419 &orangefs_inode
->refn
.khandle
,
423 ret
= wait_for_direct_io(ORANGEFS_IO_READ
, inode
, offset
, iter
,
424 count
, readahead_size
);
428 gossip_debug(GOSSIP_FILE_DEBUG
,
429 "%s(%pU): Value(%zd) returned.\n",
431 &orangefs_inode
->refn
.khandle
,
437 static ssize_t
orangefs_file_read_iter(struct kiocb
*iocb
, struct iov_iter
*iter
)
439 struct file
*file
= iocb
->ki_filp
;
440 loff_t pos
= *(&iocb
->ki_pos
);
443 BUG_ON(iocb
->private);
445 gossip_debug(GOSSIP_FILE_DEBUG
, "orangefs_file_read_iter\n");
447 g_orangefs_stats
.reads
++;
449 rc
= do_readv_writev(ORANGEFS_IO_READ
, file
, &pos
, iter
);
455 static ssize_t
orangefs_file_write_iter(struct kiocb
*iocb
, struct iov_iter
*iter
)
457 struct file
*file
= iocb
->ki_filp
;
461 BUG_ON(iocb
->private);
463 gossip_debug(GOSSIP_FILE_DEBUG
, "orangefs_file_write_iter\n");
465 mutex_lock(&file
->f_mapping
->host
->i_mutex
);
467 /* Make sure generic_write_checks sees an up to date inode size. */
468 if (file
->f_flags
& O_APPEND
) {
469 rc
= orangefs_inode_getattr(file
->f_mapping
->host
,
470 ORANGEFS_ATTR_SYS_SIZE
, 0);
472 gossip_err("%s: orangefs_inode_getattr failed, rc:%zd:.\n",
478 if (file
->f_pos
> i_size_read(file
->f_mapping
->host
))
479 orangefs_i_size_write(file
->f_mapping
->host
, file
->f_pos
);
481 rc
= generic_write_checks(iocb
, iter
);
484 gossip_err("%s: generic_write_checks failed, rc:%zd:.\n",
490 * if we are appending, generic_write_checks would have updated
491 * pos to the end of the file, so we will wait till now to set
494 pos
= *(&iocb
->ki_pos
);
496 rc
= do_readv_writev(ORANGEFS_IO_WRITE
,
501 gossip_err("%s: do_readv_writev failed, rc:%zd:.\n",
507 g_orangefs_stats
.writes
++;
511 mutex_unlock(&file
->f_mapping
->host
->i_mutex
);
516 * Perform a miscellaneous operation on a file.
518 static long orangefs_ioctl(struct file
*file
, unsigned int cmd
, unsigned long arg
)
524 gossip_debug(GOSSIP_FILE_DEBUG
,
525 "orangefs_ioctl: called with cmd %d\n",
529 * we understand some general ioctls on files, such as the immutable
532 if (cmd
== FS_IOC_GETFLAGS
) {
534 ret
= orangefs_inode_getxattr(file_inode(file
),
535 ORANGEFS_XATTR_NAME_DEFAULT_PREFIX
,
536 "user.pvfs2.meta_hint",
538 if (ret
< 0 && ret
!= -ENODATA
)
540 else if (ret
== -ENODATA
)
543 gossip_debug(GOSSIP_FILE_DEBUG
,
544 "orangefs_ioctl: FS_IOC_GETFLAGS: %llu\n",
545 (unsigned long long)uval
);
546 return put_user(uval
, (int __user
*)arg
);
547 } else if (cmd
== FS_IOC_SETFLAGS
) {
549 if (get_user(uval
, (int __user
*)arg
))
552 * ORANGEFS_MIRROR_FL is set internally when the mirroring mode
553 * is turned on for a file. The user is not allowed to turn
554 * on this bit, but the bit is present if the user first gets
555 * the flags and then updates the flags with some new
556 * settings. So, we ignore it in the following edit. bligon.
558 if ((uval
& ~ORANGEFS_MIRROR_FL
) &
559 (~(FS_IMMUTABLE_FL
| FS_APPEND_FL
| FS_NOATIME_FL
))) {
560 gossip_err("orangefs_ioctl: the FS_IOC_SETFLAGS only supports setting one of FS_IMMUTABLE_FL|FS_APPEND_FL|FS_NOATIME_FL\n");
564 gossip_debug(GOSSIP_FILE_DEBUG
,
565 "orangefs_ioctl: FS_IOC_SETFLAGS: %llu\n",
566 (unsigned long long)val
);
567 ret
= orangefs_inode_setxattr(file_inode(file
),
568 ORANGEFS_XATTR_NAME_DEFAULT_PREFIX
,
569 "user.pvfs2.meta_hint",
570 &val
, sizeof(val
), 0);
577 * Memory map a region of a file.
579 static int orangefs_file_mmap(struct file
*file
, struct vm_area_struct
*vma
)
581 gossip_debug(GOSSIP_FILE_DEBUG
,
582 "orangefs_file_mmap: called on %s\n",
584 (char *)file
->f_path
.dentry
->d_name
.name
:
587 /* set the sequential readahead hint */
588 vma
->vm_flags
|= VM_SEQ_READ
;
589 vma
->vm_flags
&= ~VM_RAND_READ
;
591 /* Use readonly mmap since we cannot support writable maps. */
592 return generic_file_readonly_mmap(file
, vma
);
595 #define mapping_nrpages(idata) ((idata)->nrpages)
598 * Called to notify the module that there are no more references to
599 * this file (i.e. no processes have it open).
601 * \note Not called when each file is closed.
603 static int orangefs_file_release(struct inode
*inode
, struct file
*file
)
605 gossip_debug(GOSSIP_FILE_DEBUG
,
606 "orangefs_file_release: called on %s\n",
607 file
->f_path
.dentry
->d_name
.name
);
609 orangefs_flush_inode(inode
);
612 * remove all associated inode pages from the page cache and mmap
613 * readahead cache (if any); this forces an expensive refresh of
614 * data for the next caller of mmap (or 'get_block' accesses)
616 if (file
->f_path
.dentry
->d_inode
&&
617 file
->f_path
.dentry
->d_inode
->i_mapping
&&
618 mapping_nrpages(&file
->f_path
.dentry
->d_inode
->i_data
))
619 truncate_inode_pages(file
->f_path
.dentry
->d_inode
->i_mapping
,
625 * Push all data for a specific file onto permanent storage.
627 static int orangefs_fsync(struct file
*file
,
633 struct orangefs_inode_s
*orangefs_inode
=
634 ORANGEFS_I(file
->f_path
.dentry
->d_inode
);
635 struct orangefs_kernel_op_s
*new_op
= NULL
;
638 filemap_write_and_wait_range(file
->f_mapping
, start
, end
);
640 new_op
= op_alloc(ORANGEFS_VFS_OP_FSYNC
);
643 new_op
->upcall
.req
.fsync
.refn
= orangefs_inode
->refn
;
645 ret
= service_operation(new_op
,
647 get_interruptible_flag(file
->f_path
.dentry
->d_inode
));
649 gossip_debug(GOSSIP_FILE_DEBUG
,
650 "orangefs_fsync got return value of %d\n",
655 orangefs_flush_inode(file
->f_path
.dentry
->d_inode
);
660 * Change the file pointer position for an instance of an open file.
662 * \note If .llseek is overriden, we must acquire lock as described in
663 * Documentation/filesystems/Locking.
665 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
666 * require much changes to the FS
668 static loff_t
orangefs_file_llseek(struct file
*file
, loff_t offset
, int origin
)
671 struct inode
*inode
= file
->f_path
.dentry
->d_inode
;
674 gossip_err("orangefs_file_llseek: invalid inode (NULL)\n");
678 if (origin
== ORANGEFS_SEEK_END
) {
680 * revalidate the inode's file size.
681 * NOTE: We are only interested in file size here,
682 * so we set mask accordingly.
684 ret
= orangefs_inode_getattr(inode
, ORANGEFS_ATTR_SYS_SIZE
, 0);
686 gossip_debug(GOSSIP_FILE_DEBUG
,
687 "%s:%s:%d calling make bad inode\n",
691 orangefs_make_bad_inode(inode
);
696 gossip_debug(GOSSIP_FILE_DEBUG
,
697 "orangefs_file_llseek: offset is %ld | origin is %d"
698 " | inode size is %lu\n",
701 (unsigned long)file
->f_path
.dentry
->d_inode
->i_size
);
703 return generic_file_llseek(file
, offset
, origin
);
707 * Support local locks (locks that only this kernel knows about)
708 * if Orangefs was mounted -o local_lock.
710 static int orangefs_lock(struct file
*filp
, int cmd
, struct file_lock
*fl
)
714 if (ORANGEFS_SB(filp
->f_inode
->i_sb
)->flags
& ORANGEFS_OPT_LOCAL_LOCK
) {
715 if (cmd
== F_GETLK
) {
717 posix_test_lock(filp
, fl
);
719 rc
= posix_lock_file(filp
, fl
, NULL
);
726 /** ORANGEFS implementation of VFS file operations */
727 const struct file_operations orangefs_file_operations
= {
728 .llseek
= orangefs_file_llseek
,
729 .read_iter
= orangefs_file_read_iter
,
730 .write_iter
= orangefs_file_write_iter
,
731 .lock
= orangefs_lock
,
732 .unlocked_ioctl
= orangefs_ioctl
,
733 .mmap
= orangefs_file_mmap
,
734 .open
= generic_file_open
,
735 .release
= orangefs_file_release
,
736 .fsync
= orangefs_fsync
,