2 * (C) 2001 Clemson University and The University of Chicago
4 * See COPYING in top-level directory.
8 * The ORANGEFS Linux kernel support allows ORANGEFS volumes to be mounted and
9 * accessed through the Linux VFS (i.e. using standard I/O system calls).
10 * This support is only needed on clients that wish to mount the file system.
15 * Declarations and macros for the ORANGEFS Linux kernel support.
18 #ifndef __ORANGEFSKERNEL_H
19 #define __ORANGEFSKERNEL_H
21 #include <linux/kernel.h>
22 #include <linux/moduleparam.h>
23 #include <linux/statfs.h>
24 #include <linux/backing-dev.h>
25 #include <linux/device.h>
26 #include <linux/mpage.h>
27 #include <linux/namei.h>
28 #include <linux/errno.h>
29 #include <linux/init.h>
30 #include <linux/module.h>
31 #include <linux/slab.h>
32 #include <linux/types.h>
34 #include <linux/vmalloc.h>
36 #include <linux/aio.h>
37 #include <linux/posix_acl.h>
38 #include <linux/posix_acl_xattr.h>
39 #include <linux/compat.h>
40 #include <linux/mount.h>
41 #include <linux/uaccess.h>
42 #include <linux/atomic.h>
43 #include <linux/uio.h>
44 #include <linux/sched.h>
46 #include <linux/wait.h>
47 #include <linux/dcache.h>
48 #include <linux/pagemap.h>
49 #include <linux/poll.h>
50 #include <linux/rwsem.h>
51 #include <linux/xattr.h>
52 #include <linux/exportfs.h>
54 #include <asm/unaligned.h>
56 #include "orangefs-dev-proto.h"
58 #ifdef ORANGEFS_KERNEL_DEBUG
59 #define ORANGEFS_DEFAULT_OP_TIMEOUT_SECS 10
61 #define ORANGEFS_DEFAULT_OP_TIMEOUT_SECS 20
64 #define ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS 30
66 #define ORANGEFS_DEFAULT_SLOT_TIMEOUT_SECS 900 /* 15 minutes */
68 #define ORANGEFS_REQDEVICE_NAME "pvfs2-req"
70 #define ORANGEFS_DEVREQ_MAGIC 0x20030529
71 #define ORANGEFS_LINK_MAX 0x000000FF
72 #define ORANGEFS_PURGE_RETRY_COUNT 0x00000005
73 #define ORANGEFS_SEEK_END 0x00000002
74 #define ORANGEFS_MAX_NUM_OPTIONS 0x00000004
75 #define ORANGEFS_MAX_MOUNT_OPT_LEN 0x00000080
76 #define ORANGEFS_MAX_FSKEY_LEN 64
78 #define MAX_DEV_REQ_UPSIZE (2 * sizeof(__s32) + \
79 sizeof(__u64) + sizeof(struct orangefs_upcall_s))
80 #define MAX_DEV_REQ_DOWNSIZE (2 * sizeof(__s32) + \
81 sizeof(__u64) + sizeof(struct orangefs_downcall_s))
84 * valid orangefs kernel operation states
86 * unknown - op was just initialized
87 * waiting - op is on request_list (upward bound)
88 * inprogr - op is in progress (waiting for downcall)
89 * serviced - op has matching downcall; ok
90 * purged - op has to start a timer since client-core
91 * exited uncleanly before servicing op
92 * given up - submitter has given up waiting for it
94 enum orangefs_vfs_op_states
{
95 OP_VFS_STATE_UNKNOWN
= 0,
96 OP_VFS_STATE_WAITING
= 1,
97 OP_VFS_STATE_INPROGR
= 2,
98 OP_VFS_STATE_SERVICED
= 4,
99 OP_VFS_STATE_PURGED
= 8,
100 OP_VFS_STATE_GIVEN_UP
= 16,
104 * Defines for controlling whether I/O upcalls are for async or sync operations
106 enum ORANGEFS_async_io_type
{
107 ORANGEFS_VFS_SYNC_IO
= 0,
108 ORANGEFS_VFS_ASYNC_IO
= 1,
112 * An array of client_debug_mask will be built to hold debug keyword/mask
113 * values fetched from userspace.
115 struct client_debug_mask
{
122 * orangefs kernel memory related flags
125 #if ((defined ORANGEFS_KERNEL_DEBUG) && (defined CONFIG_DEBUG_SLAB))
126 #define ORANGEFS_CACHE_CREATE_FLAGS SLAB_RED_ZONE
128 #define ORANGEFS_CACHE_CREATE_FLAGS 0
129 #endif /* ((defined ORANGEFS_KERNEL_DEBUG) && (defined CONFIG_DEBUG_SLAB)) */
131 #define ORANGEFS_GFP_FLAGS (GFP_KERNEL)
132 #define ORANGEFS_BUFMAP_GFP_FLAGS (GFP_KERNEL)
134 /* orangefs xattr and acl related defines */
135 #define ORANGEFS_XATTR_INDEX_POSIX_ACL_ACCESS 1
136 #define ORANGEFS_XATTR_INDEX_POSIX_ACL_DEFAULT 2
137 #define ORANGEFS_XATTR_INDEX_TRUSTED 3
138 #define ORANGEFS_XATTR_INDEX_DEFAULT 4
140 #define ORANGEFS_XATTR_NAME_ACL_ACCESS XATTR_NAME_POSIX_ACL_ACCESS
141 #define ORANGEFS_XATTR_NAME_ACL_DEFAULT XATTR_NAME_POSIX_ACL_DEFAULT
142 #define ORANGEFS_XATTR_NAME_TRUSTED_PREFIX "trusted."
143 #define ORANGEFS_XATTR_NAME_DEFAULT_PREFIX ""
145 /* these functions are defined in orangefs-utils.c */
146 int orangefs_prepare_cdm_array(char *debug_array_string
);
147 int orangefs_prepare_debugfs_help_string(int);
149 /* defined in orangefs-debugfs.c */
150 int orangefs_client_debug_init(void);
152 void debug_string_to_mask(char *, void *, int);
153 void do_c_mask(int, char *, struct client_debug_mask
**);
154 void do_k_mask(int, char *, __u64
**);
156 void debug_mask_to_string(void *, int);
157 void do_k_string(void *, int);
158 void do_c_string(void *, int);
159 int check_amalgam_keyword(void *, int);
160 int keyword_is_amalgam(char *);
162 /*these variables are defined in orangefs-mod.c */
163 extern char kernel_debug_string
[ORANGEFS_MAX_DEBUG_STRING_LEN
];
164 extern char client_debug_string
[ORANGEFS_MAX_DEBUG_STRING_LEN
];
165 extern char client_debug_array_string
[ORANGEFS_MAX_DEBUG_STRING_LEN
];
166 extern unsigned int kernel_mask_set_mod_init
;
168 extern int orangefs_init_acl(struct inode
*inode
, struct inode
*dir
);
169 extern const struct xattr_handler
*orangefs_xattr_handlers
[];
171 extern struct posix_acl
*orangefs_get_acl(struct inode
*inode
, int type
);
172 extern int orangefs_set_acl(struct inode
*inode
, struct posix_acl
*acl
, int type
);
175 * Redefine xtvec structure so that we could move helper functions out of
179 __kernel_off_t xtv_off
; /* must be off_t */
180 __kernel_size_t xtv_len
; /* must be size_t */
184 * orangefs data structures
186 struct orangefs_kernel_op_s
{
187 enum orangefs_vfs_op_states op_state
;
191 * Set uses_shared_memory to 1 if this operation uses shared memory.
192 * If true, then a retry on the op must also get a new shared memory
193 * buffer and re-populate it. Cancels don't care - it only matters
194 * for service_operation() retry logics and cancels don't go through
198 int uses_shared_memory
;
202 struct orangefs_upcall_s upcall
;
203 struct orangefs_downcall_s downcall
;
205 struct completion waitq
;
208 struct completion done
;
216 struct list_head list
;
219 #define set_op_state_waiting(op) ((op)->op_state = OP_VFS_STATE_WAITING)
220 #define set_op_state_inprogress(op) ((op)->op_state = OP_VFS_STATE_INPROGR)
221 #define set_op_state_given_up(op) ((op)->op_state = OP_VFS_STATE_GIVEN_UP)
222 static inline void set_op_state_serviced(struct orangefs_kernel_op_s
*op
)
224 op
->op_state
= OP_VFS_STATE_SERVICED
;
225 complete(&op
->waitq
);
228 #define op_state_waiting(op) ((op)->op_state & OP_VFS_STATE_WAITING)
229 #define op_state_in_progress(op) ((op)->op_state & OP_VFS_STATE_INPROGR)
230 #define op_state_serviced(op) ((op)->op_state & OP_VFS_STATE_SERVICED)
231 #define op_state_purged(op) ((op)->op_state & OP_VFS_STATE_PURGED)
232 #define op_state_given_up(op) ((op)->op_state & OP_VFS_STATE_GIVEN_UP)
233 #define op_is_cancel(op) ((op)->upcall.type == ORANGEFS_VFS_OP_CANCEL)
235 static inline void get_op(struct orangefs_kernel_op_s
*op
)
237 atomic_inc(&op
->ref_count
);
238 gossip_debug(GOSSIP_DEV_DEBUG
,
239 "(get) Alloced OP (%p:%llu)\n", op
, llu(op
->tag
));
242 void __op_release(struct orangefs_kernel_op_s
*op
);
244 static inline void op_release(struct orangefs_kernel_op_s
*op
)
246 if (atomic_dec_and_test(&op
->ref_count
)) {
247 gossip_debug(GOSSIP_DEV_DEBUG
,
248 "(put) Releasing OP (%p:%llu)\n", op
, llu((op
)->tag
));
253 extern void orangefs_bufmap_put(int);
254 static inline void put_cancel(struct orangefs_kernel_op_s
*op
)
256 orangefs_bufmap_put(op
->slot_to_free
);
260 static inline void set_op_state_purged(struct orangefs_kernel_op_s
*op
)
262 spin_lock(&op
->lock
);
263 if (unlikely(op_is_cancel(op
))) {
265 spin_unlock(&op
->lock
);
268 op
->op_state
|= OP_VFS_STATE_PURGED
;
269 complete(&op
->waitq
);
270 spin_unlock(&op
->lock
);
274 /* per inode private orangefs info */
275 struct orangefs_inode_s
{
276 struct orangefs_object_kref refn
;
277 char link_target
[ORANGEFS_NAME_MAX
];
280 * Reading/Writing Extended attributes need to acquire the appropriate
281 * reader/writer semaphore on the orangefs_inode_s structure.
283 struct rw_semaphore xattr_sem
;
285 struct inode vfs_inode
;
286 sector_t last_failed_block_index_read
;
289 * State of in-memory attributes not yet flushed to disk associated
292 unsigned long pinode_flags
;
295 #define P_ATIME_FLAG 0
296 #define P_MTIME_FLAG 1
297 #define P_CTIME_FLAG 2
298 #define P_MODE_FLAG 3
300 #define ClearAtimeFlag(pinode) clear_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
301 #define SetAtimeFlag(pinode) set_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
302 #define AtimeFlag(pinode) test_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
304 #define ClearMtimeFlag(pinode) clear_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
305 #define SetMtimeFlag(pinode) set_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
306 #define MtimeFlag(pinode) test_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
308 #define ClearCtimeFlag(pinode) clear_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
309 #define SetCtimeFlag(pinode) set_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
310 #define CtimeFlag(pinode) test_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
312 #define ClearModeFlag(pinode) clear_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
313 #define SetModeFlag(pinode) set_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
314 #define ModeFlag(pinode) test_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
316 /* per superblock private orangefs info */
317 struct orangefs_sb_info_s
{
318 struct orangefs_khandle root_khandle
;
322 #define ORANGEFS_OPT_INTR 0x01
323 #define ORANGEFS_OPT_LOCAL_LOCK 0x02
324 char devname
[ORANGEFS_MAX_SERVER_ADDR_LEN
];
325 struct super_block
*sb
;
327 struct list_head list
;
331 * structure that holds the state of any async I/O operation issued
332 * through the VFS. Needed especially to handle cancellation requests
333 * or even completion notification so that the VFS client-side daemon
334 * can free up its vfs_request slots.
336 struct orangefs_kiocb_s
{
337 /* the pointer to the task that initiated the AIO */
338 struct task_struct
*tsk
;
340 /* pointer to the kiocb that kicked this operation */
343 /* buffer index that was used for the I/O */
344 struct orangefs_bufmap
*bufmap
;
347 /* orangefs kernel operation type */
348 struct orangefs_kernel_op_s
*op
;
350 /* The user space buffers from/to which I/O is being staged */
353 /* number of elements in the iovector */
354 unsigned long nr_segs
;
356 /* set to indicate the type of the operation */
362 /* and the count in bytes */
363 size_t bytes_to_be_copied
;
365 ssize_t bytes_copied
;
369 struct orangefs_stats
{
370 unsigned long cache_hits
;
371 unsigned long cache_misses
;
373 unsigned long writes
;
376 extern struct orangefs_stats g_orangefs_stats
;
379 * NOTE: See Documentation/filesystems/porting for information
380 * on implementing FOO_I and properly accessing fs private data
382 static inline struct orangefs_inode_s
*ORANGEFS_I(struct inode
*inode
)
384 return container_of(inode
, struct orangefs_inode_s
, vfs_inode
);
387 static inline struct orangefs_sb_info_s
*ORANGEFS_SB(struct super_block
*sb
)
389 return (struct orangefs_sb_info_s
*) sb
->s_fs_info
;
392 /* ino_t descends from "unsigned long", 8 bytes, 64 bits. */
393 static inline ino_t
orangefs_khandle_to_ino(struct orangefs_khandle
*khandle
)
400 ihandle
.u
[0] = khandle
->u
[0] ^ khandle
->u
[4];
401 ihandle
.u
[1] = khandle
->u
[1] ^ khandle
->u
[5];
402 ihandle
.u
[2] = khandle
->u
[2] ^ khandle
->u
[6];
403 ihandle
.u
[3] = khandle
->u
[3] ^ khandle
->u
[7];
404 ihandle
.u
[4] = khandle
->u
[12] ^ khandle
->u
[8];
405 ihandle
.u
[5] = khandle
->u
[13] ^ khandle
->u
[9];
406 ihandle
.u
[6] = khandle
->u
[14] ^ khandle
->u
[10];
407 ihandle
.u
[7] = khandle
->u
[15] ^ khandle
->u
[11];
412 static inline struct orangefs_khandle
*get_khandle_from_ino(struct inode
*inode
)
414 return &(ORANGEFS_I(inode
)->refn
.khandle
);
417 static inline __s32
get_fsid_from_ino(struct inode
*inode
)
419 return ORANGEFS_I(inode
)->refn
.fs_id
;
422 static inline ino_t
get_ino_from_khandle(struct inode
*inode
)
424 struct orangefs_khandle
*khandle
;
427 khandle
= get_khandle_from_ino(inode
);
428 ino
= orangefs_khandle_to_ino(khandle
);
432 static inline ino_t
get_parent_ino_from_dentry(struct dentry
*dentry
)
434 return get_ino_from_khandle(dentry
->d_parent
->d_inode
);
437 static inline int is_root_handle(struct inode
*inode
)
439 gossip_debug(GOSSIP_DCACHE_DEBUG
,
440 "%s: root handle: %pU, this handle: %pU:\n",
442 &ORANGEFS_SB(inode
->i_sb
)->root_khandle
,
443 get_khandle_from_ino(inode
));
445 if (ORANGEFS_khandle_cmp(&(ORANGEFS_SB(inode
->i_sb
)->root_khandle
),
446 get_khandle_from_ino(inode
)))
452 static inline int match_handle(struct orangefs_khandle resp_handle
,
455 gossip_debug(GOSSIP_DCACHE_DEBUG
,
456 "%s: one handle: %pU, another handle:%pU:\n",
459 get_khandle_from_ino(inode
));
461 if (ORANGEFS_khandle_cmp(&resp_handle
, get_khandle_from_ino(inode
)))
468 * defined in orangefs-cache.c
470 int op_cache_initialize(void);
471 int op_cache_finalize(void);
472 struct orangefs_kernel_op_s
*op_alloc(__s32 type
);
473 void orangefs_new_tag(struct orangefs_kernel_op_s
*op
);
474 char *get_opname_string(struct orangefs_kernel_op_s
*new_op
);
476 int orangefs_inode_cache_initialize(void);
477 int orangefs_inode_cache_finalize(void);
480 * defined in orangefs-mod.c
482 void purge_inprogress_ops(void);
485 * defined in waitqueue.c
487 void purge_waiting_ops(void);
492 struct dentry
*orangefs_mount(struct file_system_type
*fst
,
497 void orangefs_kill_sb(struct super_block
*sb
);
498 int orangefs_remount(struct super_block
*sb
);
500 int fsid_key_table_initialize(void);
501 void fsid_key_table_finalize(void);
506 __u32
convert_to_orangefs_mask(unsigned long lite_mask
);
507 struct inode
*orangefs_new_inode(struct super_block
*sb
,
511 struct orangefs_object_kref
*ref
);
513 int orangefs_setattr(struct dentry
*dentry
, struct iattr
*iattr
);
515 int orangefs_getattr(struct vfsmount
*mnt
,
516 struct dentry
*dentry
,
517 struct kstat
*kstat
);
519 int orangefs_permission(struct inode
*inode
, int mask
);
524 int orangefs_setxattr(struct dentry
*dentry
,
530 ssize_t
orangefs_getxattr(struct dentry
*dentry
,
535 ssize_t
orangefs_listxattr(struct dentry
*dentry
, char *buffer
, size_t size
);
540 struct inode
*orangefs_iget(struct super_block
*sb
,
541 struct orangefs_object_kref
*ref
);
543 ssize_t
orangefs_inode_read(struct inode
*inode
,
544 struct iov_iter
*iter
,
546 loff_t readahead_size
);
549 * defined in devorangefs-req.c
551 int orangefs_dev_init(void);
552 void orangefs_dev_cleanup(void);
553 int is_daemon_in_service(void);
554 bool __is_daemon_in_service(void);
555 int fs_mount_pending(__s32 fsid
);
558 * defined in orangefs-utils.c
560 __s32
fsid_of_op(struct orangefs_kernel_op_s
*op
);
562 int orangefs_flush_inode(struct inode
*inode
);
564 ssize_t
orangefs_inode_getxattr(struct inode
*inode
,
570 int orangefs_inode_setxattr(struct inode
*inode
,
577 int orangefs_inode_getattr(struct inode
*inode
, __u32 mask
, int check
);
579 int orangefs_inode_setattr(struct inode
*inode
, struct iattr
*iattr
);
581 void orangefs_make_bad_inode(struct inode
*inode
);
583 int orangefs_unmount_sb(struct super_block
*sb
);
585 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s
*op
);
587 static inline __u64
orangefs_convert_time_field(const struct timespec
*ts
)
589 return (__u64
)ts
->tv_sec
;
592 int orangefs_normalize_to_errno(__s32 error_code
);
594 extern struct mutex devreq_mutex
;
595 extern struct mutex request_mutex
;
597 extern int op_timeout_secs
;
598 extern int slot_timeout_secs
;
599 extern struct list_head orangefs_superblocks
;
600 extern spinlock_t orangefs_superblocks_lock
;
601 extern struct list_head orangefs_request_list
;
602 extern spinlock_t orangefs_request_list_lock
;
603 extern wait_queue_head_t orangefs_request_list_waitq
;
604 extern struct list_head
*htable_ops_in_progress
;
605 extern spinlock_t htable_ops_in_progress_lock
;
606 extern int hash_table_size
;
608 extern const struct address_space_operations orangefs_address_operations
;
609 extern struct backing_dev_info orangefs_backing_dev_info
;
610 extern struct inode_operations orangefs_file_inode_operations
;
611 extern const struct file_operations orangefs_file_operations
;
612 extern struct inode_operations orangefs_symlink_inode_operations
;
613 extern struct inode_operations orangefs_dir_inode_operations
;
614 extern const struct file_operations orangefs_dir_operations
;
615 extern const struct dentry_operations orangefs_dentry_operations
;
616 extern const struct file_operations orangefs_devreq_file_operations
;
618 extern wait_queue_head_t orangefs_bufmap_init_waitq
;
621 * misc convenience macros
624 #define ORANGEFS_OP_INTERRUPTIBLE 1 /* service_operation() is interruptible */
625 #define ORANGEFS_OP_PRIORITY 2 /* service_operation() is high priority */
626 #define ORANGEFS_OP_CANCELLATION 4 /* this is a cancellation */
627 #define ORANGEFS_OP_NO_SEMAPHORE 8 /* don't acquire semaphore */
628 #define ORANGEFS_OP_ASYNC 16 /* Queue it, but don't wait */
630 int service_operation(struct orangefs_kernel_op_s
*op
,
634 #define get_interruptible_flag(inode) \
635 ((ORANGEFS_SB(inode->i_sb)->flags & ORANGEFS_OPT_INTR) ? \
636 ORANGEFS_OP_INTERRUPTIBLE : 0)
638 #define add_orangefs_sb(sb) \
640 gossip_debug(GOSSIP_SUPER_DEBUG, \
641 "Adding SB %p to orangefs superblocks\n", \
643 spin_lock(&orangefs_superblocks_lock); \
644 list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks); \
645 spin_unlock(&orangefs_superblocks_lock); \
648 #define remove_orangefs_sb(sb) \
650 struct list_head *tmp = NULL; \
651 struct list_head *tmp_safe = NULL; \
652 struct orangefs_sb_info_s *orangefs_sb = NULL; \
654 spin_lock(&orangefs_superblocks_lock); \
655 list_for_each_safe(tmp, tmp_safe, &orangefs_superblocks) { \
656 orangefs_sb = list_entry(tmp, \
657 struct orangefs_sb_info_s, \
659 if (orangefs_sb && (orangefs_sb->sb == sb)) { \
660 gossip_debug(GOSSIP_SUPER_DEBUG, \
661 "Removing SB %p from orangefs superblocks\n", \
663 list_del(&orangefs_sb->list); \
667 spin_unlock(&orangefs_superblocks_lock); \
670 #define orangefs_lock_inode(inode) spin_lock(&inode->i_lock)
671 #define orangefs_unlock_inode(inode) spin_unlock(&inode->i_lock)
673 #define fill_default_sys_attrs(sys_attr, type, mode) \
675 sys_attr.owner = from_kuid(current_user_ns(), current_fsuid()); \
676 sys_attr.group = from_kgid(current_user_ns(), current_fsgid()); \
678 sys_attr.perms = ORANGEFS_util_translate_mode(mode); \
679 sys_attr.objtype = type; \
680 sys_attr.mask = ORANGEFS_ATTR_SYS_ALL_SETABLE; \
683 #define orangefs_inode_lock(__i) mutex_lock(&(__i)->i_mutex)
685 #define orangefs_inode_unlock(__i) mutex_unlock(&(__i)->i_mutex)
687 static inline void orangefs_i_size_write(struct inode
*inode
, loff_t i_size
)
689 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
690 orangefs_inode_lock(inode
);
692 i_size_write(inode
, i_size
);
693 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
694 orangefs_inode_unlock(inode
);
698 static inline unsigned int diff(struct timeval
*end
, struct timeval
*begin
)
700 if (end
->tv_usec
< begin
->tv_usec
) {
701 end
->tv_usec
+= 1000000;
704 end
->tv_sec
-= begin
->tv_sec
;
705 end
->tv_usec
-= begin
->tv_usec
;
706 return (end
->tv_sec
* 1000000) + end
->tv_usec
;
709 #endif /* __ORANGEFSKERNEL_H */