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
;
214 struct list_head list
;
217 #define set_op_state_waiting(op) ((op)->op_state = OP_VFS_STATE_WAITING)
218 #define set_op_state_inprogress(op) ((op)->op_state = OP_VFS_STATE_INPROGR)
219 #define set_op_state_given_up(op) ((op)->op_state = OP_VFS_STATE_GIVEN_UP)
220 static inline void set_op_state_serviced(struct orangefs_kernel_op_s
*op
)
222 op
->op_state
= OP_VFS_STATE_SERVICED
;
223 complete(&op
->waitq
);
226 #define op_state_waiting(op) ((op)->op_state & OP_VFS_STATE_WAITING)
227 #define op_state_in_progress(op) ((op)->op_state & OP_VFS_STATE_INPROGR)
228 #define op_state_serviced(op) ((op)->op_state & OP_VFS_STATE_SERVICED)
229 #define op_state_purged(op) ((op)->op_state & OP_VFS_STATE_PURGED)
230 #define op_state_given_up(op) ((op)->op_state & OP_VFS_STATE_GIVEN_UP)
231 #define op_is_cancel(op) ((op)->upcall.type == ORANGEFS_VFS_OP_CANCEL)
233 static inline void get_op(struct orangefs_kernel_op_s
*op
)
235 atomic_inc(&op
->ref_count
);
236 gossip_debug(GOSSIP_DEV_DEBUG
,
237 "(get) Alloced OP (%p:%llu)\n", op
, llu(op
->tag
));
240 void __op_release(struct orangefs_kernel_op_s
*op
);
242 static inline void op_release(struct orangefs_kernel_op_s
*op
)
244 if (atomic_dec_and_test(&op
->ref_count
)) {
245 gossip_debug(GOSSIP_DEV_DEBUG
,
246 "(put) Releasing OP (%p:%llu)\n", op
, llu((op
)->tag
));
251 extern void orangefs_bufmap_put(int);
252 static inline void put_cancel(struct orangefs_kernel_op_s
*op
)
254 orangefs_bufmap_put(op
->slot_to_free
);
258 static inline void set_op_state_purged(struct orangefs_kernel_op_s
*op
)
260 spin_lock(&op
->lock
);
261 if (unlikely(op_is_cancel(op
))) {
262 list_del_init(&op
->list
);
263 spin_unlock(&op
->lock
);
266 op
->op_state
|= OP_VFS_STATE_PURGED
;
267 complete(&op
->waitq
);
268 spin_unlock(&op
->lock
);
272 /* per inode private orangefs info */
273 struct orangefs_inode_s
{
274 struct orangefs_object_kref refn
;
275 char link_target
[ORANGEFS_NAME_MAX
];
278 * Reading/Writing Extended attributes need to acquire the appropriate
279 * reader/writer semaphore on the orangefs_inode_s structure.
281 struct rw_semaphore xattr_sem
;
283 struct inode vfs_inode
;
284 sector_t last_failed_block_index_read
;
287 * State of in-memory attributes not yet flushed to disk associated
290 unsigned long pinode_flags
;
293 #define P_ATIME_FLAG 0
294 #define P_MTIME_FLAG 1
295 #define P_CTIME_FLAG 2
296 #define P_MODE_FLAG 3
298 #define ClearAtimeFlag(pinode) clear_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
299 #define SetAtimeFlag(pinode) set_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
300 #define AtimeFlag(pinode) test_bit(P_ATIME_FLAG, &(pinode)->pinode_flags)
302 #define ClearMtimeFlag(pinode) clear_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
303 #define SetMtimeFlag(pinode) set_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
304 #define MtimeFlag(pinode) test_bit(P_MTIME_FLAG, &(pinode)->pinode_flags)
306 #define ClearCtimeFlag(pinode) clear_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
307 #define SetCtimeFlag(pinode) set_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
308 #define CtimeFlag(pinode) test_bit(P_CTIME_FLAG, &(pinode)->pinode_flags)
310 #define ClearModeFlag(pinode) clear_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
311 #define SetModeFlag(pinode) set_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
312 #define ModeFlag(pinode) test_bit(P_MODE_FLAG, &(pinode)->pinode_flags)
314 /* per superblock private orangefs info */
315 struct orangefs_sb_info_s
{
316 struct orangefs_khandle root_khandle
;
320 #define ORANGEFS_OPT_INTR 0x01
321 #define ORANGEFS_OPT_LOCAL_LOCK 0x02
322 char devname
[ORANGEFS_MAX_SERVER_ADDR_LEN
];
323 struct super_block
*sb
;
325 struct list_head list
;
329 * structure that holds the state of any async I/O operation issued
330 * through the VFS. Needed especially to handle cancellation requests
331 * or even completion notification so that the VFS client-side daemon
332 * can free up its vfs_request slots.
334 struct orangefs_kiocb_s
{
335 /* the pointer to the task that initiated the AIO */
336 struct task_struct
*tsk
;
338 /* pointer to the kiocb that kicked this operation */
341 /* buffer index that was used for the I/O */
342 struct orangefs_bufmap
*bufmap
;
345 /* orangefs kernel operation type */
346 struct orangefs_kernel_op_s
*op
;
348 /* The user space buffers from/to which I/O is being staged */
351 /* number of elements in the iovector */
352 unsigned long nr_segs
;
354 /* set to indicate the type of the operation */
360 /* and the count in bytes */
361 size_t bytes_to_be_copied
;
363 ssize_t bytes_copied
;
367 struct orangefs_stats
{
368 unsigned long cache_hits
;
369 unsigned long cache_misses
;
371 unsigned long writes
;
374 extern struct orangefs_stats g_orangefs_stats
;
377 * NOTE: See Documentation/filesystems/porting for information
378 * on implementing FOO_I and properly accessing fs private data
380 static inline struct orangefs_inode_s
*ORANGEFS_I(struct inode
*inode
)
382 return container_of(inode
, struct orangefs_inode_s
, vfs_inode
);
385 static inline struct orangefs_sb_info_s
*ORANGEFS_SB(struct super_block
*sb
)
387 return (struct orangefs_sb_info_s
*) sb
->s_fs_info
;
390 /* ino_t descends from "unsigned long", 8 bytes, 64 bits. */
391 static inline ino_t
orangefs_khandle_to_ino(struct orangefs_khandle
*khandle
)
398 ihandle
.u
[0] = khandle
->u
[0] ^ khandle
->u
[4];
399 ihandle
.u
[1] = khandle
->u
[1] ^ khandle
->u
[5];
400 ihandle
.u
[2] = khandle
->u
[2] ^ khandle
->u
[6];
401 ihandle
.u
[3] = khandle
->u
[3] ^ khandle
->u
[7];
402 ihandle
.u
[4] = khandle
->u
[12] ^ khandle
->u
[8];
403 ihandle
.u
[5] = khandle
->u
[13] ^ khandle
->u
[9];
404 ihandle
.u
[6] = khandle
->u
[14] ^ khandle
->u
[10];
405 ihandle
.u
[7] = khandle
->u
[15] ^ khandle
->u
[11];
410 static inline struct orangefs_khandle
*get_khandle_from_ino(struct inode
*inode
)
412 return &(ORANGEFS_I(inode
)->refn
.khandle
);
415 static inline __s32
get_fsid_from_ino(struct inode
*inode
)
417 return ORANGEFS_I(inode
)->refn
.fs_id
;
420 static inline ino_t
get_ino_from_khandle(struct inode
*inode
)
422 struct orangefs_khandle
*khandle
;
425 khandle
= get_khandle_from_ino(inode
);
426 ino
= orangefs_khandle_to_ino(khandle
);
430 static inline ino_t
get_parent_ino_from_dentry(struct dentry
*dentry
)
432 return get_ino_from_khandle(dentry
->d_parent
->d_inode
);
435 static inline int is_root_handle(struct inode
*inode
)
437 gossip_debug(GOSSIP_DCACHE_DEBUG
,
438 "%s: root handle: %pU, this handle: %pU:\n",
440 &ORANGEFS_SB(inode
->i_sb
)->root_khandle
,
441 get_khandle_from_ino(inode
));
443 if (ORANGEFS_khandle_cmp(&(ORANGEFS_SB(inode
->i_sb
)->root_khandle
),
444 get_khandle_from_ino(inode
)))
450 static inline int match_handle(struct orangefs_khandle resp_handle
,
453 gossip_debug(GOSSIP_DCACHE_DEBUG
,
454 "%s: one handle: %pU, another handle:%pU:\n",
457 get_khandle_from_ino(inode
));
459 if (ORANGEFS_khandle_cmp(&resp_handle
, get_khandle_from_ino(inode
)))
466 * defined in orangefs-cache.c
468 int op_cache_initialize(void);
469 int op_cache_finalize(void);
470 struct orangefs_kernel_op_s
*op_alloc(__s32 type
);
471 void orangefs_new_tag(struct orangefs_kernel_op_s
*op
);
472 char *get_opname_string(struct orangefs_kernel_op_s
*new_op
);
474 int orangefs_inode_cache_initialize(void);
475 int orangefs_inode_cache_finalize(void);
478 * defined in orangefs-mod.c
480 void purge_inprogress_ops(void);
483 * defined in waitqueue.c
485 void purge_waiting_ops(void);
490 struct dentry
*orangefs_mount(struct file_system_type
*fst
,
495 void orangefs_kill_sb(struct super_block
*sb
);
496 int orangefs_remount(struct super_block
*sb
);
498 int fsid_key_table_initialize(void);
499 void fsid_key_table_finalize(void);
504 __u32
convert_to_orangefs_mask(unsigned long lite_mask
);
505 struct inode
*orangefs_new_inode(struct super_block
*sb
,
509 struct orangefs_object_kref
*ref
);
511 int orangefs_setattr(struct dentry
*dentry
, struct iattr
*iattr
);
513 int orangefs_getattr(struct vfsmount
*mnt
,
514 struct dentry
*dentry
,
515 struct kstat
*kstat
);
517 int orangefs_permission(struct inode
*inode
, int mask
);
522 int orangefs_setxattr(struct dentry
*dentry
,
528 ssize_t
orangefs_getxattr(struct dentry
*dentry
,
533 ssize_t
orangefs_listxattr(struct dentry
*dentry
, char *buffer
, size_t size
);
538 struct inode
*orangefs_iget(struct super_block
*sb
,
539 struct orangefs_object_kref
*ref
);
541 ssize_t
orangefs_inode_read(struct inode
*inode
,
542 struct iov_iter
*iter
,
544 loff_t readahead_size
);
547 * defined in devorangefs-req.c
549 int orangefs_dev_init(void);
550 void orangefs_dev_cleanup(void);
551 int is_daemon_in_service(void);
552 bool __is_daemon_in_service(void);
553 int fs_mount_pending(__s32 fsid
);
556 * defined in orangefs-utils.c
558 __s32
fsid_of_op(struct orangefs_kernel_op_s
*op
);
560 int orangefs_flush_inode(struct inode
*inode
);
562 ssize_t
orangefs_inode_getxattr(struct inode
*inode
,
568 int orangefs_inode_setxattr(struct inode
*inode
,
575 int orangefs_inode_getattr(struct inode
*inode
, __u32 mask
, int check
);
577 int orangefs_inode_setattr(struct inode
*inode
, struct iattr
*iattr
);
579 void orangefs_make_bad_inode(struct inode
*inode
);
581 int orangefs_unmount_sb(struct super_block
*sb
);
583 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s
*op
);
585 static inline __u64
orangefs_convert_time_field(const struct timespec
*ts
)
587 return (__u64
)ts
->tv_sec
;
590 int orangefs_normalize_to_errno(__s32 error_code
);
592 extern struct mutex devreq_mutex
;
593 extern struct mutex request_mutex
;
595 extern int op_timeout_secs
;
596 extern int slot_timeout_secs
;
597 extern struct list_head orangefs_superblocks
;
598 extern spinlock_t orangefs_superblocks_lock
;
599 extern struct list_head orangefs_request_list
;
600 extern spinlock_t orangefs_request_list_lock
;
601 extern wait_queue_head_t orangefs_request_list_waitq
;
602 extern struct list_head
*htable_ops_in_progress
;
603 extern spinlock_t htable_ops_in_progress_lock
;
604 extern int hash_table_size
;
606 extern const struct address_space_operations orangefs_address_operations
;
607 extern struct backing_dev_info orangefs_backing_dev_info
;
608 extern struct inode_operations orangefs_file_inode_operations
;
609 extern const struct file_operations orangefs_file_operations
;
610 extern struct inode_operations orangefs_symlink_inode_operations
;
611 extern struct inode_operations orangefs_dir_inode_operations
;
612 extern const struct file_operations orangefs_dir_operations
;
613 extern const struct dentry_operations orangefs_dentry_operations
;
614 extern const struct file_operations orangefs_devreq_file_operations
;
616 extern wait_queue_head_t orangefs_bufmap_init_waitq
;
619 * misc convenience macros
622 #define ORANGEFS_OP_INTERRUPTIBLE 1 /* service_operation() is interruptible */
623 #define ORANGEFS_OP_PRIORITY 2 /* service_operation() is high priority */
624 #define ORANGEFS_OP_CANCELLATION 4 /* this is a cancellation */
625 #define ORANGEFS_OP_NO_SEMAPHORE 8 /* don't acquire semaphore */
626 #define ORANGEFS_OP_ASYNC 16 /* Queue it, but don't wait */
628 int service_operation(struct orangefs_kernel_op_s
*op
,
632 #define get_interruptible_flag(inode) \
633 ((ORANGEFS_SB(inode->i_sb)->flags & ORANGEFS_OPT_INTR) ? \
634 ORANGEFS_OP_INTERRUPTIBLE : 0)
636 #define add_orangefs_sb(sb) \
638 gossip_debug(GOSSIP_SUPER_DEBUG, \
639 "Adding SB %p to orangefs superblocks\n", \
641 spin_lock(&orangefs_superblocks_lock); \
642 list_add_tail(&ORANGEFS_SB(sb)->list, &orangefs_superblocks); \
643 spin_unlock(&orangefs_superblocks_lock); \
646 #define remove_orangefs_sb(sb) \
648 struct list_head *tmp = NULL; \
649 struct list_head *tmp_safe = NULL; \
650 struct orangefs_sb_info_s *orangefs_sb = NULL; \
652 spin_lock(&orangefs_superblocks_lock); \
653 list_for_each_safe(tmp, tmp_safe, &orangefs_superblocks) { \
654 orangefs_sb = list_entry(tmp, \
655 struct orangefs_sb_info_s, \
657 if (orangefs_sb && (orangefs_sb->sb == sb)) { \
658 gossip_debug(GOSSIP_SUPER_DEBUG, \
659 "Removing SB %p from orangefs superblocks\n", \
661 list_del(&orangefs_sb->list); \
665 spin_unlock(&orangefs_superblocks_lock); \
668 #define orangefs_lock_inode(inode) spin_lock(&inode->i_lock)
669 #define orangefs_unlock_inode(inode) spin_unlock(&inode->i_lock)
671 #define fill_default_sys_attrs(sys_attr, type, mode) \
673 sys_attr.owner = from_kuid(current_user_ns(), current_fsuid()); \
674 sys_attr.group = from_kgid(current_user_ns(), current_fsgid()); \
676 sys_attr.perms = ORANGEFS_util_translate_mode(mode); \
677 sys_attr.objtype = type; \
678 sys_attr.mask = ORANGEFS_ATTR_SYS_ALL_SETABLE; \
681 #define orangefs_inode_lock(__i) mutex_lock(&(__i)->i_mutex)
683 #define orangefs_inode_unlock(__i) mutex_unlock(&(__i)->i_mutex)
685 static inline void orangefs_i_size_write(struct inode
*inode
, loff_t i_size
)
687 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
688 orangefs_inode_lock(inode
);
690 i_size_write(inode
, i_size
);
691 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP)
692 orangefs_inode_unlock(inode
);
696 static inline unsigned int diff(struct timeval
*end
, struct timeval
*begin
)
698 if (end
->tv_usec
< begin
->tv_usec
) {
699 end
->tv_usec
+= 1000000;
702 end
->tv_sec
-= begin
->tv_sec
;
703 end
->tv_usec
-= begin
->tv_usec
;
704 return (end
->tv_sec
* 1000000) + end
->tv_usec
;
707 #endif /* __ORANGEFSKERNEL_H */