1 /* 2 * (C) 2001 Clemson University and The University of Chicago 3 * 4 * See COPYING in top-level directory. 5 */ 6 7 /* 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. 11 * 12 */ 13 14 /* 15 * Declarations and macros for the ORANGEFS Linux kernel support. 16 */ 17 18 #ifndef __ORANGEFSKERNEL_H 19 #define __ORANGEFSKERNEL_H 20 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> 33 #include <linux/fs.h> 34 #include <linux/vmalloc.h> 35 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> 45 #include <linux/mm.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> 53 54 #include <asm/unaligned.h> 55 56 #include "orangefs-dev-proto.h" 57 58 #ifdef ORANGEFS_KERNEL_DEBUG 59 #define ORANGEFS_DEFAULT_OP_TIMEOUT_SECS 10 60 #else 61 #define ORANGEFS_DEFAULT_OP_TIMEOUT_SECS 20 62 #endif 63 64 #define ORANGEFS_BUFMAP_WAIT_TIMEOUT_SECS 30 65 66 #define ORANGEFS_DEFAULT_SLOT_TIMEOUT_SECS 900 /* 15 minutes */ 67 68 #define ORANGEFS_REQDEVICE_NAME "pvfs2-req" 69 70 #define ORANGEFS_DEVREQ_MAGIC 0x20030529 71 #define ORANGEFS_LINK_MAX 0x000000FF 72 #define ORANGEFS_PURGE_RETRY_COUNT 0x00000005 73 #define ORANGEFS_MAX_NUM_OPTIONS 0x00000004 74 #define ORANGEFS_MAX_MOUNT_OPT_LEN 0x00000080 75 #define ORANGEFS_MAX_FSKEY_LEN 64 76 77 #define MAX_DEV_REQ_UPSIZE (2 * sizeof(__s32) + \ 78 sizeof(__u64) + sizeof(struct orangefs_upcall_s)) 79 #define MAX_DEV_REQ_DOWNSIZE (2 * sizeof(__s32) + \ 80 sizeof(__u64) + sizeof(struct orangefs_downcall_s)) 81 82 /* 83 * valid orangefs kernel operation states 84 * 85 * unknown - op was just initialized 86 * waiting - op is on request_list (upward bound) 87 * inprogr - op is in progress (waiting for downcall) 88 * serviced - op has matching downcall; ok 89 * purged - op has to start a timer since client-core 90 * exited uncleanly before servicing op 91 * given up - submitter has given up waiting for it 92 */ 93 enum orangefs_vfs_op_states { 94 OP_VFS_STATE_UNKNOWN = 0, 95 OP_VFS_STATE_WAITING = 1, 96 OP_VFS_STATE_INPROGR = 2, 97 OP_VFS_STATE_SERVICED = 4, 98 OP_VFS_STATE_PURGED = 8, 99 OP_VFS_STATE_GIVEN_UP = 16, 100 }; 101 102 /* 103 * An array of client_debug_mask will be built to hold debug keyword/mask 104 * values fetched from userspace. 105 */ 106 struct client_debug_mask { 107 char *keyword; 108 __u64 mask1; 109 __u64 mask2; 110 }; 111 112 /* 113 * orangefs kernel memory related flags 114 */ 115 116 #if ((defined ORANGEFS_KERNEL_DEBUG) && (defined CONFIG_DEBUG_SLAB)) 117 #define ORANGEFS_CACHE_CREATE_FLAGS SLAB_RED_ZONE 118 #else 119 #define ORANGEFS_CACHE_CREATE_FLAGS 0 120 #endif /* ((defined ORANGEFS_KERNEL_DEBUG) && (defined CONFIG_DEBUG_SLAB)) */ 121 122 /* these functions are defined in orangefs-utils.c */ 123 int orangefs_prepare_cdm_array(char *debug_array_string); 124 int orangefs_prepare_debugfs_help_string(int); 125 126 /* defined in orangefs-debugfs.c */ 127 int orangefs_client_debug_init(void); 128 129 void debug_string_to_mask(char *, void *, int); 130 void do_c_mask(int, char *, struct client_debug_mask **); 131 void do_k_mask(int, char *, __u64 **); 132 133 void debug_mask_to_string(void *, int); 134 void do_k_string(void *, int); 135 void do_c_string(void *, int); 136 int check_amalgam_keyword(void *, int); 137 int keyword_is_amalgam(char *); 138 139 /*these variables are defined in orangefs-mod.c */ 140 extern char kernel_debug_string[ORANGEFS_MAX_DEBUG_STRING_LEN]; 141 extern char client_debug_string[ORANGEFS_MAX_DEBUG_STRING_LEN]; 142 extern char client_debug_array_string[ORANGEFS_MAX_DEBUG_STRING_LEN]; 143 extern unsigned int kernel_mask_set_mod_init; 144 145 extern int orangefs_init_acl(struct inode *inode, struct inode *dir); 146 extern const struct xattr_handler *orangefs_xattr_handlers[]; 147 148 extern struct posix_acl *orangefs_get_acl(struct inode *inode, int type); 149 extern int orangefs_set_acl(struct inode *inode, struct posix_acl *acl, int type); 150 151 /* 152 * Redefine xtvec structure so that we could move helper functions out of 153 * the define 154 */ 155 struct xtvec { 156 __kernel_off_t xtv_off; /* must be off_t */ 157 __kernel_size_t xtv_len; /* must be size_t */ 158 }; 159 160 /* 161 * orangefs data structures 162 */ 163 struct orangefs_kernel_op_s { 164 enum orangefs_vfs_op_states op_state; 165 __u64 tag; 166 167 /* 168 * Set uses_shared_memory to non zero if this operation uses 169 * shared memory. If true, then a retry on the op must also 170 * get a new shared memory buffer and re-populate it. 171 * Cancels don't care - it only matters for service_operation() 172 * retry logics and cancels don't go through it anymore. It 173 * safely stays non-zero when we use it as slot_to_free. 174 */ 175 union { 176 int uses_shared_memory; 177 int slot_to_free; 178 }; 179 180 struct orangefs_upcall_s upcall; 181 struct orangefs_downcall_s downcall; 182 183 struct completion waitq; 184 spinlock_t lock; 185 186 int attempts; 187 188 struct list_head list; 189 }; 190 191 #define set_op_state_waiting(op) ((op)->op_state = OP_VFS_STATE_WAITING) 192 #define set_op_state_inprogress(op) ((op)->op_state = OP_VFS_STATE_INPROGR) 193 #define set_op_state_given_up(op) ((op)->op_state = OP_VFS_STATE_GIVEN_UP) 194 static inline void set_op_state_serviced(struct orangefs_kernel_op_s *op) 195 { 196 op->op_state = OP_VFS_STATE_SERVICED; 197 complete(&op->waitq); 198 } 199 200 #define op_state_waiting(op) ((op)->op_state & OP_VFS_STATE_WAITING) 201 #define op_state_in_progress(op) ((op)->op_state & OP_VFS_STATE_INPROGR) 202 #define op_state_serviced(op) ((op)->op_state & OP_VFS_STATE_SERVICED) 203 #define op_state_purged(op) ((op)->op_state & OP_VFS_STATE_PURGED) 204 #define op_state_given_up(op) ((op)->op_state & OP_VFS_STATE_GIVEN_UP) 205 #define op_is_cancel(op) ((op)->upcall.type == ORANGEFS_VFS_OP_CANCEL) 206 207 void op_release(struct orangefs_kernel_op_s *op); 208 209 extern void orangefs_bufmap_put(int); 210 static inline void put_cancel(struct orangefs_kernel_op_s *op) 211 { 212 orangefs_bufmap_put(op->slot_to_free); 213 op_release(op); 214 } 215 216 static inline void set_op_state_purged(struct orangefs_kernel_op_s *op) 217 { 218 spin_lock(&op->lock); 219 if (unlikely(op_is_cancel(op))) { 220 list_del_init(&op->list); 221 spin_unlock(&op->lock); 222 put_cancel(op); 223 } else { 224 op->op_state |= OP_VFS_STATE_PURGED; 225 complete(&op->waitq); 226 spin_unlock(&op->lock); 227 } 228 } 229 230 /* per inode private orangefs info */ 231 struct orangefs_inode_s { 232 struct orangefs_object_kref refn; 233 char link_target[ORANGEFS_NAME_MAX]; 234 __s64 blksize; 235 /* 236 * Reading/Writing Extended attributes need to acquire the appropriate 237 * reader/writer semaphore on the orangefs_inode_s structure. 238 */ 239 struct rw_semaphore xattr_sem; 240 241 struct inode vfs_inode; 242 sector_t last_failed_block_index_read; 243 244 /* 245 * State of in-memory attributes not yet flushed to disk associated 246 * with this object 247 */ 248 unsigned long pinode_flags; 249 250 unsigned long getattr_time; 251 }; 252 253 #define P_ATIME_FLAG 0 254 #define P_MTIME_FLAG 1 255 #define P_CTIME_FLAG 2 256 #define P_MODE_FLAG 3 257 258 #define ClearAtimeFlag(pinode) clear_bit(P_ATIME_FLAG, &(pinode)->pinode_flags) 259 #define SetAtimeFlag(pinode) set_bit(P_ATIME_FLAG, &(pinode)->pinode_flags) 260 #define AtimeFlag(pinode) test_bit(P_ATIME_FLAG, &(pinode)->pinode_flags) 261 262 #define ClearMtimeFlag(pinode) clear_bit(P_MTIME_FLAG, &(pinode)->pinode_flags) 263 #define SetMtimeFlag(pinode) set_bit(P_MTIME_FLAG, &(pinode)->pinode_flags) 264 #define MtimeFlag(pinode) test_bit(P_MTIME_FLAG, &(pinode)->pinode_flags) 265 266 #define ClearCtimeFlag(pinode) clear_bit(P_CTIME_FLAG, &(pinode)->pinode_flags) 267 #define SetCtimeFlag(pinode) set_bit(P_CTIME_FLAG, &(pinode)->pinode_flags) 268 #define CtimeFlag(pinode) test_bit(P_CTIME_FLAG, &(pinode)->pinode_flags) 269 270 #define ClearModeFlag(pinode) clear_bit(P_MODE_FLAG, &(pinode)->pinode_flags) 271 #define SetModeFlag(pinode) set_bit(P_MODE_FLAG, &(pinode)->pinode_flags) 272 #define ModeFlag(pinode) test_bit(P_MODE_FLAG, &(pinode)->pinode_flags) 273 274 /* per superblock private orangefs info */ 275 struct orangefs_sb_info_s { 276 struct orangefs_khandle root_khandle; 277 __s32 fs_id; 278 int id; 279 int flags; 280 #define ORANGEFS_OPT_INTR 0x01 281 #define ORANGEFS_OPT_LOCAL_LOCK 0x02 282 char devname[ORANGEFS_MAX_SERVER_ADDR_LEN]; 283 struct super_block *sb; 284 int mount_pending; 285 struct list_head list; 286 }; 287 288 /* 289 * structure that holds the state of any async I/O operation issued 290 * through the VFS. Needed especially to handle cancellation requests 291 * or even completion notification so that the VFS client-side daemon 292 * can free up its vfs_request slots. 293 */ 294 struct orangefs_kiocb_s { 295 /* the pointer to the task that initiated the AIO */ 296 struct task_struct *tsk; 297 298 /* pointer to the kiocb that kicked this operation */ 299 struct kiocb *kiocb; 300 301 /* buffer index that was used for the I/O */ 302 struct orangefs_bufmap *bufmap; 303 int buffer_index; 304 305 /* orangefs kernel operation type */ 306 struct orangefs_kernel_op_s *op; 307 308 /* The user space buffers from/to which I/O is being staged */ 309 struct iovec *iov; 310 311 /* number of elements in the iovector */ 312 unsigned long nr_segs; 313 314 /* set to indicate the type of the operation */ 315 int rw; 316 317 /* file offset */ 318 loff_t offset; 319 320 /* and the count in bytes */ 321 size_t bytes_to_be_copied; 322 323 ssize_t bytes_copied; 324 int needs_cleanup; 325 }; 326 327 struct orangefs_stats { 328 unsigned long cache_hits; 329 unsigned long cache_misses; 330 unsigned long reads; 331 unsigned long writes; 332 }; 333 334 extern struct orangefs_stats g_orangefs_stats; 335 336 /* 337 * NOTE: See Documentation/filesystems/porting for information 338 * on implementing FOO_I and properly accessing fs private data 339 */ 340 static inline struct orangefs_inode_s *ORANGEFS_I(struct inode *inode) 341 { 342 return container_of(inode, struct orangefs_inode_s, vfs_inode); 343 } 344 345 static inline struct orangefs_sb_info_s *ORANGEFS_SB(struct super_block *sb) 346 { 347 return (struct orangefs_sb_info_s *) sb->s_fs_info; 348 } 349 350 /* ino_t descends from "unsigned long", 8 bytes, 64 bits. */ 351 static inline ino_t orangefs_khandle_to_ino(struct orangefs_khandle *khandle) 352 { 353 union { 354 unsigned char u[8]; 355 __u64 ino; 356 } ihandle; 357 358 ihandle.u[0] = khandle->u[0] ^ khandle->u[4]; 359 ihandle.u[1] = khandle->u[1] ^ khandle->u[5]; 360 ihandle.u[2] = khandle->u[2] ^ khandle->u[6]; 361 ihandle.u[3] = khandle->u[3] ^ khandle->u[7]; 362 ihandle.u[4] = khandle->u[12] ^ khandle->u[8]; 363 ihandle.u[5] = khandle->u[13] ^ khandle->u[9]; 364 ihandle.u[6] = khandle->u[14] ^ khandle->u[10]; 365 ihandle.u[7] = khandle->u[15] ^ khandle->u[11]; 366 367 return ihandle.ino; 368 } 369 370 static inline struct orangefs_khandle *get_khandle_from_ino(struct inode *inode) 371 { 372 return &(ORANGEFS_I(inode)->refn.khandle); 373 } 374 375 static inline __s32 get_fsid_from_ino(struct inode *inode) 376 { 377 return ORANGEFS_I(inode)->refn.fs_id; 378 } 379 380 static inline ino_t get_ino_from_khandle(struct inode *inode) 381 { 382 struct orangefs_khandle *khandle; 383 ino_t ino; 384 385 khandle = get_khandle_from_ino(inode); 386 ino = orangefs_khandle_to_ino(khandle); 387 return ino; 388 } 389 390 static inline ino_t get_parent_ino_from_dentry(struct dentry *dentry) 391 { 392 return get_ino_from_khandle(dentry->d_parent->d_inode); 393 } 394 395 static inline int is_root_handle(struct inode *inode) 396 { 397 gossip_debug(GOSSIP_DCACHE_DEBUG, 398 "%s: root handle: %pU, this handle: %pU:\n", 399 __func__, 400 &ORANGEFS_SB(inode->i_sb)->root_khandle, 401 get_khandle_from_ino(inode)); 402 403 if (ORANGEFS_khandle_cmp(&(ORANGEFS_SB(inode->i_sb)->root_khandle), 404 get_khandle_from_ino(inode))) 405 return 0; 406 else 407 return 1; 408 } 409 410 static inline int match_handle(struct orangefs_khandle resp_handle, 411 struct inode *inode) 412 { 413 gossip_debug(GOSSIP_DCACHE_DEBUG, 414 "%s: one handle: %pU, another handle:%pU:\n", 415 __func__, 416 &resp_handle, 417 get_khandle_from_ino(inode)); 418 419 if (ORANGEFS_khandle_cmp(&resp_handle, get_khandle_from_ino(inode))) 420 return 0; 421 else 422 return 1; 423 } 424 425 /* 426 * defined in orangefs-cache.c 427 */ 428 int op_cache_initialize(void); 429 int op_cache_finalize(void); 430 struct orangefs_kernel_op_s *op_alloc(__s32 type); 431 void orangefs_new_tag(struct orangefs_kernel_op_s *op); 432 char *get_opname_string(struct orangefs_kernel_op_s *new_op); 433 434 int orangefs_inode_cache_initialize(void); 435 int orangefs_inode_cache_finalize(void); 436 437 /* 438 * defined in orangefs-mod.c 439 */ 440 void purge_inprogress_ops(void); 441 442 /* 443 * defined in waitqueue.c 444 */ 445 void purge_waiting_ops(void); 446 447 /* 448 * defined in super.c 449 */ 450 struct dentry *orangefs_mount(struct file_system_type *fst, 451 int flags, 452 const char *devname, 453 void *data); 454 455 void orangefs_kill_sb(struct super_block *sb); 456 int orangefs_remount(struct orangefs_sb_info_s *); 457 458 int fsid_key_table_initialize(void); 459 void fsid_key_table_finalize(void); 460 461 /* 462 * defined in inode.c 463 */ 464 __u32 convert_to_orangefs_mask(unsigned long lite_mask); 465 struct inode *orangefs_new_inode(struct super_block *sb, 466 struct inode *dir, 467 int mode, 468 dev_t dev, 469 struct orangefs_object_kref *ref); 470 471 int orangefs_setattr(struct dentry *dentry, struct iattr *iattr); 472 473 int orangefs_getattr(struct vfsmount *mnt, 474 struct dentry *dentry, 475 struct kstat *kstat); 476 477 int orangefs_permission(struct inode *inode, int mask); 478 479 /* 480 * defined in xattr.c 481 */ 482 int orangefs_setxattr(struct dentry *dentry, 483 const char *name, 484 const void *value, 485 size_t size, 486 int flags); 487 488 ssize_t orangefs_getxattr(struct dentry *dentry, 489 const char *name, 490 void *buffer, 491 size_t size); 492 493 ssize_t orangefs_listxattr(struct dentry *dentry, char *buffer, size_t size); 494 495 /* 496 * defined in namei.c 497 */ 498 struct inode *orangefs_iget(struct super_block *sb, 499 struct orangefs_object_kref *ref); 500 501 ssize_t orangefs_inode_read(struct inode *inode, 502 struct iov_iter *iter, 503 loff_t *offset, 504 loff_t readahead_size); 505 506 /* 507 * defined in devorangefs-req.c 508 */ 509 int orangefs_dev_init(void); 510 void orangefs_dev_cleanup(void); 511 int is_daemon_in_service(void); 512 bool __is_daemon_in_service(void); 513 514 /* 515 * defined in orangefs-utils.c 516 */ 517 __s32 fsid_of_op(struct orangefs_kernel_op_s *op); 518 519 int orangefs_flush_inode(struct inode *inode); 520 521 ssize_t orangefs_inode_getxattr(struct inode *inode, 522 const char *name, 523 void *buffer, 524 size_t size); 525 526 int orangefs_inode_setxattr(struct inode *inode, 527 const char *name, 528 const void *value, 529 size_t size, 530 int flags); 531 532 int orangefs_inode_getattr(struct inode *inode, int new, int bypass); 533 534 int orangefs_inode_check_changed(struct inode *inode); 535 536 int orangefs_inode_setattr(struct inode *inode, struct iattr *iattr); 537 538 void orangefs_make_bad_inode(struct inode *inode); 539 540 int orangefs_unmount_sb(struct super_block *sb); 541 542 bool orangefs_cancel_op_in_progress(struct orangefs_kernel_op_s *op); 543 544 int orangefs_normalize_to_errno(__s32 error_code); 545 546 extern struct mutex devreq_mutex; 547 extern struct mutex request_mutex; 548 extern int debug; 549 extern int op_timeout_secs; 550 extern int slot_timeout_secs; 551 extern int dcache_timeout_msecs; 552 extern int getattr_timeout_msecs; 553 extern struct list_head orangefs_superblocks; 554 extern spinlock_t orangefs_superblocks_lock; 555 extern struct list_head orangefs_request_list; 556 extern spinlock_t orangefs_request_list_lock; 557 extern wait_queue_head_t orangefs_request_list_waitq; 558 extern struct list_head *htable_ops_in_progress; 559 extern spinlock_t htable_ops_in_progress_lock; 560 extern int hash_table_size; 561 562 extern const struct address_space_operations orangefs_address_operations; 563 extern struct backing_dev_info orangefs_backing_dev_info; 564 extern const struct inode_operations orangefs_file_inode_operations; 565 extern const struct file_operations orangefs_file_operations; 566 extern const struct inode_operations orangefs_symlink_inode_operations; 567 extern const struct inode_operations orangefs_dir_inode_operations; 568 extern const struct file_operations orangefs_dir_operations; 569 extern const struct dentry_operations orangefs_dentry_operations; 570 extern const struct file_operations orangefs_devreq_file_operations; 571 572 extern wait_queue_head_t orangefs_bufmap_init_waitq; 573 574 /* 575 * misc convenience macros 576 */ 577 578 #define ORANGEFS_OP_INTERRUPTIBLE 1 /* service_operation() is interruptible */ 579 #define ORANGEFS_OP_PRIORITY 2 /* service_operation() is high priority */ 580 #define ORANGEFS_OP_CANCELLATION 4 /* this is a cancellation */ 581 #define ORANGEFS_OP_NO_MUTEX 8 /* don't acquire request_mutex */ 582 #define ORANGEFS_OP_ASYNC 16 /* Queue it, but don't wait */ 583 584 int service_operation(struct orangefs_kernel_op_s *op, 585 const char *op_name, 586 int flags); 587 588 #define get_interruptible_flag(inode) \ 589 ((ORANGEFS_SB(inode->i_sb)->flags & ORANGEFS_OPT_INTR) ? \ 590 ORANGEFS_OP_INTERRUPTIBLE : 0) 591 592 #define fill_default_sys_attrs(sys_attr, type, mode) \ 593 do { \ 594 sys_attr.owner = from_kuid(&init_user_ns, current_fsuid()); \ 595 sys_attr.group = from_kgid(&init_user_ns, current_fsgid()); \ 596 sys_attr.perms = ORANGEFS_util_translate_mode(mode); \ 597 sys_attr.mtime = 0; \ 598 sys_attr.atime = 0; \ 599 sys_attr.ctime = 0; \ 600 sys_attr.mask = ORANGEFS_ATTR_SYS_ALL_SETABLE; \ 601 } while (0) 602 603 static inline void orangefs_i_size_write(struct inode *inode, loff_t i_size) 604 { 605 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP) 606 inode_lock(inode); 607 #endif 608 i_size_write(inode, i_size); 609 #if BITS_PER_LONG == 32 && defined(CONFIG_SMP) 610 inode_unlock(inode); 611 #endif 612 } 613 614 #endif /* __ORANGEFSKERNEL_H */ 615