1 #ifndef IO_URING_TYPES_H 2 #define IO_URING_TYPES_H 3 4 #include <linux/blkdev.h> 5 #include <linux/hashtable.h> 6 #include <linux/task_work.h> 7 #include <linux/bitmap.h> 8 #include <linux/llist.h> 9 #include <uapi/linux/io_uring.h> 10 11 struct iou_loop_params; 12 struct io_uring_bpf_ops; 13 14 enum { 15 /* 16 * A hint to not wake right away but delay until there are enough of 17 * tw's queued to match the number of CQEs the task is waiting for. 18 * 19 * Must not be used with requests generating more than one CQE. 20 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set. 21 */ 22 IOU_F_TWQ_LAZY_WAKE = 1, 23 }; 24 25 enum io_uring_cmd_flags { 26 IO_URING_F_COMPLETE_DEFER = 1, 27 IO_URING_F_UNLOCKED = 2, 28 /* the request is executed from poll, it should not be freed */ 29 IO_URING_F_MULTISHOT = 4, 30 /* executed by io-wq */ 31 IO_URING_F_IOWQ = 8, 32 /* executed inline from syscall */ 33 IO_URING_F_INLINE = 16, 34 /* int's last bit, sign checks are usually faster than a bit test */ 35 IO_URING_F_NONBLOCK = INT_MIN, 36 37 /* ctx state flags, for URING_CMD */ 38 IO_URING_F_SQE128 = (1 << 8), 39 IO_URING_F_CQE32 = (1 << 9), 40 IO_URING_F_IOPOLL = (1 << 10), 41 42 /* set when uring wants to cancel a previously issued command */ 43 IO_URING_F_CANCEL = (1 << 11), 44 IO_URING_F_COMPAT = (1 << 12), 45 }; 46 47 struct iou_loop_params; 48 49 struct io_wq_work_node { 50 struct io_wq_work_node *next; 51 }; 52 53 struct io_wq_work_list { 54 struct io_wq_work_node *first; 55 struct io_wq_work_node *last; 56 }; 57 58 /* 59 * Lockless multi-producer, single-consumer FIFO queue, see 60 * io_uring/mpscq.h for the implementation and rules. Defined here so 61 * that it can be embedded in io_ring_ctx. This is the producer side 62 * only - the consumer cursor is kept separately, on a cacheline that 63 * isn't dirtied by the producers. 64 */ 65 struct mpscq { 66 struct llist_node *tail; /* producers */ 67 struct llist_node stub; 68 }; 69 70 struct io_wq_work { 71 struct io_wq_work_node list; 72 atomic_t flags; 73 /* place it here instead of io_kiocb as it fills padding and saves 4B */ 74 int cancel_seq; 75 }; 76 77 struct io_rsrc_data { 78 unsigned int nr; 79 struct io_rsrc_node **nodes; 80 }; 81 82 struct io_file_table { 83 struct io_rsrc_data data; 84 unsigned long *bitmap; 85 unsigned int alloc_hint; 86 }; 87 88 struct io_hash_bucket { 89 struct hlist_head list; 90 } ____cacheline_aligned_in_smp; 91 92 struct io_hash_table { 93 struct io_hash_bucket *hbs; 94 unsigned hash_bits; 95 }; 96 97 struct io_mapped_region { 98 struct page **pages; 99 void *ptr; 100 unsigned nr_pages; 101 unsigned flags; 102 }; 103 104 /* 105 * Return value from io_buffer_list selection, to avoid stashing it in 106 * struct io_kiocb. For legacy/classic provided buffers, keeping a reference 107 * across execution contexts are fine. But for ring provided buffers, the 108 * list may go away as soon as ->uring_lock is dropped. As the io_kiocb 109 * persists, it's better to just keep the buffer local for those cases. 110 */ 111 struct io_br_sel { 112 struct io_buffer_list *buf_list; 113 /* 114 * Some selection parts return the user address, others return an error. 115 */ 116 union { 117 void __user *addr; 118 ssize_t val; 119 }; 120 }; 121 122 123 /* 124 * Arbitrary limit, can be raised if need be 125 */ 126 #define IO_RINGFD_REG_MAX 16 127 128 struct io_uring_task { 129 /* submission side */ 130 int cached_refs; 131 const struct io_ring_ctx *last; 132 struct task_struct *task; 133 struct io_wq *io_wq; 134 /* 135 * Consumer cursor for ->task_list. Only popped by the task itself, 136 * or by ->fallback_work once the task can no longer run task_work. 137 */ 138 struct llist_node *task_head; 139 struct file *registered_rings[IO_RINGFD_REG_MAX]; 140 141 struct xarray xa; 142 struct wait_queue_head wait; 143 atomic_t in_cancel; 144 atomic_t inflight_tracked; 145 struct percpu_counter inflight; 146 147 /* drains ->task_list once the task can no longer run task_work */ 148 struct work_struct fallback_work; 149 150 struct { /* task_work */ 151 struct mpscq task_list; 152 struct callback_head task_work; 153 } ____cacheline_aligned_in_smp; 154 }; 155 156 struct iou_vec { 157 union { 158 struct iovec *iovec; 159 struct bio_vec *bvec; 160 }; 161 unsigned nr; /* number of struct iovec it can hold */ 162 }; 163 164 struct io_uring { 165 u32 head; 166 u32 tail; 167 }; 168 169 /* 170 * This data is shared with the application through the mmap at offsets 171 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. 172 * 173 * The offsets to the member fields are published through struct 174 * io_sqring_offsets when calling io_uring_setup. 175 */ 176 struct io_rings { 177 /* 178 * Head and tail offsets into the ring; the offsets need to be 179 * masked to get valid indices. 180 * 181 * The kernel controls head of the sq ring and the tail of the cq ring, 182 * and the application controls tail of the sq ring and the head of the 183 * cq ring. 184 */ 185 struct io_uring sq, cq; 186 /* 187 * Bitmasks to apply to head and tail offsets (constant, equals 188 * ring_entries - 1) 189 */ 190 u32 sq_ring_mask, cq_ring_mask; 191 /* Ring sizes (constant, power of 2) */ 192 u32 sq_ring_entries, cq_ring_entries; 193 /* 194 * Number of invalid entries dropped by the kernel due to 195 * invalid index stored in array 196 * 197 * Written by the kernel, shouldn't be modified by the 198 * application (i.e. get number of "new events" by comparing to 199 * cached value). 200 * 201 * After a new SQ head value was read by the application this 202 * counter includes all submissions that were dropped reaching 203 * the new SQ head (and possibly more). 204 */ 205 u32 sq_dropped; 206 /* 207 * Runtime SQ flags 208 * 209 * Written by the kernel, shouldn't be modified by the 210 * application. 211 * 212 * The application needs a full memory barrier before checking 213 * for IORING_SQ_NEED_WAKEUP after updating the sq tail. 214 */ 215 atomic_t sq_flags; 216 /* 217 * Runtime CQ flags 218 * 219 * Written by the application, shouldn't be modified by the 220 * kernel. 221 */ 222 u32 cq_flags; 223 /* 224 * Number of completion events lost because the queue was full; 225 * this should be avoided by the application by making sure 226 * there are not more requests pending than there is space in 227 * the completion queue. 228 * 229 * Written by the kernel, shouldn't be modified by the 230 * application (i.e. get number of "new events" by comparing to 231 * cached value). 232 * 233 * As completion events come in out of order this counter is not 234 * ordered with any other data. 235 */ 236 u32 cq_overflow; 237 /* 238 * Ring buffer of completion events. 239 * 240 * The kernel writes completion events fresh every time they are 241 * produced, so the application is allowed to modify pending 242 * entries. 243 */ 244 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; 245 }; 246 247 struct io_bpf_filter; 248 struct io_bpf_filters { 249 refcount_t refs; /* ref for ->bpf_filters */ 250 spinlock_t lock; /* protects ->bpf_filters modifications */ 251 struct io_bpf_filter __rcu **filters; 252 struct rcu_head rcu_head; 253 }; 254 255 struct io_restriction { 256 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST); 257 DECLARE_BITMAP(sqe_op, IORING_OP_LAST); 258 struct io_bpf_filters *bpf_filters; 259 /* ->bpf_filters needs COW on modification */ 260 bool bpf_filters_cow; 261 u8 sqe_flags_allowed; 262 u8 sqe_flags_required; 263 /* IORING_OP_* restrictions exist */ 264 bool op_registered; 265 /* IORING_REGISTER_* restrictions exist */ 266 bool reg_registered; 267 }; 268 269 struct io_submit_link { 270 struct io_kiocb *head; 271 struct io_kiocb *last; 272 }; 273 274 struct io_submit_state { 275 /* inline/task_work completion list, under ->uring_lock */ 276 struct io_wq_work_node free_list; 277 /* batch completion logic */ 278 struct io_wq_work_list compl_reqs; 279 struct io_submit_link link; 280 281 bool plug_started; 282 bool need_plug; 283 bool cq_flush; 284 unsigned short submit_nr; 285 struct blk_plug plug; 286 }; 287 288 struct io_alloc_cache { 289 void **entries; 290 unsigned int nr_cached; 291 unsigned int max_cached; 292 unsigned int elem_size; 293 unsigned int init_clear; 294 }; 295 296 enum { 297 IO_RING_F_DRAIN_NEXT = BIT(0), 298 IO_RING_F_OP_RESTRICTED = BIT(1), 299 IO_RING_F_REG_RESTRICTED = BIT(2), 300 IO_RING_F_OFF_TIMEOUT_USED = BIT(3), 301 IO_RING_F_DRAIN_ACTIVE = BIT(4), 302 IO_RING_F_HAS_EVFD = BIT(5), 303 /* all CQEs should be posted only by the submitter task */ 304 IO_RING_F_TASK_COMPLETE = BIT(6), 305 IO_RING_F_LOCKLESS_CQ = BIT(7), 306 IO_RING_F_SYSCALL_IOPOLL = BIT(8), 307 IO_RING_F_POLL_ACTIVATED = BIT(9), 308 IO_RING_F_DRAIN_DISABLED = BIT(10), 309 IO_RING_F_COMPAT = BIT(11), 310 IO_RING_F_IOWQ_LIMITS_SET = BIT(12), 311 }; 312 313 struct iou_ctx {}; 314 315 struct io_ring_ctx { 316 /* const or read-mostly hot data */ 317 struct { 318 /* ring setup flags */ 319 unsigned int flags; 320 /* internal state flags IO_RING_F_* flags , mostly read-only */ 321 unsigned int int_flags; 322 323 struct task_struct *submitter_task; 324 struct io_rings *rings; 325 /* cache of ->restrictions.bpf_filters->filters */ 326 struct io_bpf_filter __rcu **bpf_filters; 327 struct percpu_ref refs; 328 329 clockid_t clockid; 330 enum tk_offsets clock_offset; 331 332 enum task_work_notify_mode notify_method; 333 unsigned sq_thread_idle; 334 } ____cacheline_aligned_in_smp; 335 336 /* submission data */ 337 struct { 338 struct mutex uring_lock; 339 340 /* 341 * Ring buffer of indices into array of io_uring_sqe, which is 342 * mmapped by the application using the IORING_OFF_SQES offset. 343 * 344 * This indirection could e.g. be used to assign fixed 345 * io_uring_sqe entries to operations and only submit them to 346 * the queue when needed. 347 * 348 * The kernel modifies neither the indices array nor the entries 349 * array. 350 */ 351 u32 *sq_array; 352 struct io_uring_sqe *sq_sqes; 353 unsigned cached_sq_head; 354 unsigned sq_entries; 355 356 /* 357 * Fixed resources fast path, should be accessed only under 358 * uring_lock, and updated through io_uring_register(2) 359 */ 360 atomic_t cancel_seq; 361 362 /* 363 * ->iopoll_list is protected by the ctx->uring_lock for 364 * io_uring instances that don't use IORING_SETUP_SQPOLL. 365 * For SQPOLL, only the single threaded io_sq_thread() will 366 * manipulate the list, hence no extra locking is needed there. 367 */ 368 bool poll_multi_queue; 369 struct list_head iopoll_list; 370 371 /* 372 * Consumer cursor for ->work_list, protected by ->uring_lock. 373 * Deliberately kept away from the producer side of the queue, 374 * as it's written for every popped entry, and the producer 375 * cacheline is contended enough as it is. 376 */ 377 struct llist_node *work_head; 378 379 struct io_file_table file_table; 380 struct io_rsrc_data buf_table; 381 struct io_alloc_cache node_cache; 382 struct io_alloc_cache imu_cache; 383 384 struct io_submit_state submit_state; 385 386 /* 387 * Modifications are protected by ->uring_lock and ->mmap_lock. 388 * The buffer list's io mapped region should be stable once 389 * published. 390 */ 391 struct xarray io_bl_xa; 392 393 struct io_hash_table cancel_table; 394 struct io_alloc_cache apoll_cache; 395 struct io_alloc_cache netmsg_cache; 396 struct io_alloc_cache rw_cache; 397 struct io_alloc_cache cmd_cache; 398 399 int (*loop_step)(struct iou_ctx *, 400 struct iou_loop_params *); 401 402 /* 403 * Any cancelable uring_cmd is added to this list in 404 * ->uring_cmd() by io_uring_cmd_insert_cancelable() 405 */ 406 struct hlist_head cancelable_uring_cmd; 407 /* 408 * For Hybrid IOPOLL, runtime in hybrid polling, without 409 * scheduling time 410 */ 411 u64 hybrid_poll_time; 412 } ____cacheline_aligned_in_smp; 413 414 struct { 415 /* 416 * We cache a range of free CQEs we can use, once exhausted it 417 * should go through a slower range setup, see __io_get_cqe() 418 */ 419 struct io_uring_cqe *cqe_cached; 420 struct io_uring_cqe *cqe_sentinel; 421 422 unsigned cached_cq_tail; 423 unsigned cq_entries; 424 struct io_ev_fd __rcu *io_ev_fd; 425 426 void *cq_wait_arg; 427 size_t cq_wait_size; 428 } ____cacheline_aligned_in_smp; 429 430 /* 431 * task_work and async notification delivery cacheline. Expected to 432 * regularly bounce b/w CPUs. 433 */ 434 struct { 435 struct io_rings __rcu *rings_rcu; 436 struct mpscq work_list; 437 unsigned long check_cq; 438 atomic_t cq_wait_nr; 439 atomic_t cq_timeouts; 440 struct wait_queue_head cq_wait; 441 } ____cacheline_aligned_in_smp; 442 443 /* timeouts */ 444 struct { 445 raw_spinlock_t timeout_lock; 446 struct list_head timeout_list; 447 struct list_head ltimeout_list; 448 unsigned cq_last_tm_flush; 449 } ____cacheline_aligned_in_smp; 450 451 spinlock_t completion_lock; 452 453 struct list_head cq_overflow_list; 454 455 struct hlist_head waitid_list; 456 457 #ifdef CONFIG_FUTEX 458 struct hlist_head futex_list; 459 struct io_alloc_cache futex_cache; 460 #endif 461 462 const struct cred *sq_creds; /* cred used for __io_sq_thread() */ 463 struct io_sq_data *sq_data; /* if using sq thread polling */ 464 465 struct wait_queue_head sqo_sq_wait; 466 struct list_head sqd_list; 467 468 unsigned int file_alloc_start; 469 unsigned int file_alloc_end; 470 471 /* Keep this last, we don't need it for the fast path */ 472 struct wait_queue_head poll_wq; 473 struct io_restriction restrictions; 474 475 /* Stores zcrx object pointers of type struct io_zcrx_ifq */ 476 struct xarray zcrx_ctxs; 477 478 /* Used for accounting references on pages in registered buffers */ 479 struct xarray hpage_acct; 480 481 u32 pers_next; 482 struct xarray personalities; 483 484 /* hashed buffered write serialization */ 485 struct io_wq_hash *hash_map; 486 487 /* Only used for accounting purposes */ 488 struct user_struct *user; 489 struct mm_struct *mm_account; 490 491 /* 492 * List of tctx nodes for this ctx, protected by tctx_lock. For 493 * cancelation purposes, nests under uring_lock. 494 */ 495 struct list_head tctx_list; 496 struct mutex tctx_lock; 497 498 /* ctx exit and cancelation */ 499 struct work_struct exit_work; 500 struct completion ref_comp; 501 502 /* io-wq management, e.g. thread count */ 503 u32 iowq_limits[2]; 504 505 struct callback_head poll_wq_task_work; 506 struct list_head defer_list; 507 unsigned nr_drained; 508 509 /* protected by ->completion_lock */ 510 unsigned nr_req_allocated; 511 512 #ifdef CONFIG_NET_RX_BUSY_POLL 513 struct list_head napi_list; /* track busy poll napi_id */ 514 spinlock_t napi_lock; /* napi_list lock */ 515 516 /* napi busy poll default timeout */ 517 ktime_t napi_busy_poll_dt; 518 bool napi_prefer_busy_poll; 519 u8 napi_track_mode; 520 521 DECLARE_HASHTABLE(napi_ht, 4); 522 #endif 523 524 struct io_uring_bpf_ops *bpf_ops; 525 526 /* 527 * Protection for resize vs mmap races - both the mmap and resize 528 * side will need to grab this lock, to prevent either side from 529 * being run concurrently with the other. 530 */ 531 struct mutex mmap_lock; 532 533 struct io_mapped_region sq_region; 534 struct io_mapped_region ring_region; 535 /* used for optimised request parameter and wait argument passing */ 536 struct io_mapped_region param_region; 537 }; 538 539 /* 540 * Token indicating function is called in task work context: 541 * ctx->uring_lock is held and any completions generated will be flushed. 542 * ONLY core io_uring.c should instantiate this struct. 543 */ 544 struct io_tw_state { 545 bool cancel; 546 }; 547 /* Alias to use in code that doesn't instantiate struct io_tw_state */ 548 typedef struct io_tw_state io_tw_token_t; 549 550 enum { 551 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, 552 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, 553 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, 554 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, 555 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, 556 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, 557 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT, 558 559 /* first byte is taken by user flags, shift it to not overlap */ 560 REQ_F_FAIL_BIT = 8, 561 REQ_F_INFLIGHT_BIT, 562 REQ_F_CUR_POS_BIT, 563 REQ_F_NOWAIT_BIT, 564 REQ_F_LINK_TIMEOUT_BIT, 565 REQ_F_NEED_CLEANUP_BIT, 566 REQ_F_POLLED_BIT, 567 REQ_F_HYBRID_IOPOLL_STATE_BIT, 568 REQ_F_BUFFER_SELECTED_BIT, 569 REQ_F_BUFFER_RING_BIT, 570 REQ_F_REISSUE_BIT, 571 REQ_F_CREDS_BIT, 572 REQ_F_REFCOUNT_BIT, 573 REQ_F_ARM_LTIMEOUT_BIT, 574 REQ_F_ASYNC_DATA_BIT, 575 REQ_F_SKIP_LINK_CQES_BIT, 576 REQ_F_SINGLE_POLL_BIT, 577 REQ_F_DOUBLE_POLL_BIT, 578 REQ_F_MULTISHOT_BIT, 579 REQ_F_APOLL_MULTISHOT_BIT, 580 REQ_F_CLEAR_POLLIN_BIT, 581 /* keep async read/write and isreg together and in order */ 582 REQ_F_SUPPORT_NOWAIT_BIT, 583 REQ_F_ISREG_BIT, 584 REQ_F_POLL_NO_LAZY_BIT, 585 REQ_F_CAN_POLL_BIT, 586 REQ_F_BL_EMPTY_BIT, 587 REQ_F_BL_NO_RECYCLE_BIT, 588 REQ_F_BUFFERS_COMMIT_BIT, 589 REQ_F_BUF_NODE_BIT, 590 REQ_F_BUF_MORE_BIT, 591 REQ_F_HAS_METADATA_BIT, 592 REQ_F_IMPORT_BUFFER_BIT, 593 REQ_F_SQE_COPIED_BIT, 594 REQ_F_IOPOLL_BIT, 595 596 /* not a real bit, just to check we're not overflowing the space */ 597 __REQ_F_LAST_BIT, 598 }; 599 600 typedef u64 __bitwise io_req_flags_t; 601 #define IO_REQ_FLAG(bitno) ((__force io_req_flags_t) BIT_ULL((bitno))) 602 603 enum { 604 /* ctx owns file */ 605 REQ_F_FIXED_FILE = IO_REQ_FLAG(REQ_F_FIXED_FILE_BIT), 606 /* drain existing IO first */ 607 REQ_F_IO_DRAIN = IO_REQ_FLAG(REQ_F_IO_DRAIN_BIT), 608 /* linked sqes */ 609 REQ_F_LINK = IO_REQ_FLAG(REQ_F_LINK_BIT), 610 /* doesn't sever on completion < 0 */ 611 REQ_F_HARDLINK = IO_REQ_FLAG(REQ_F_HARDLINK_BIT), 612 /* IOSQE_ASYNC */ 613 REQ_F_FORCE_ASYNC = IO_REQ_FLAG(REQ_F_FORCE_ASYNC_BIT), 614 /* IOSQE_BUFFER_SELECT */ 615 REQ_F_BUFFER_SELECT = IO_REQ_FLAG(REQ_F_BUFFER_SELECT_BIT), 616 /* IOSQE_CQE_SKIP_SUCCESS */ 617 REQ_F_CQE_SKIP = IO_REQ_FLAG(REQ_F_CQE_SKIP_BIT), 618 619 /* fail rest of links */ 620 REQ_F_FAIL = IO_REQ_FLAG(REQ_F_FAIL_BIT), 621 /* on inflight list, should be cancelled and waited on exit reliably */ 622 REQ_F_INFLIGHT = IO_REQ_FLAG(REQ_F_INFLIGHT_BIT), 623 /* read/write uses file position */ 624 REQ_F_CUR_POS = IO_REQ_FLAG(REQ_F_CUR_POS_BIT), 625 /* must not punt to workers */ 626 REQ_F_NOWAIT = IO_REQ_FLAG(REQ_F_NOWAIT_BIT), 627 /* has or had linked timeout */ 628 REQ_F_LINK_TIMEOUT = IO_REQ_FLAG(REQ_F_LINK_TIMEOUT_BIT), 629 /* needs cleanup */ 630 REQ_F_NEED_CLEANUP = IO_REQ_FLAG(REQ_F_NEED_CLEANUP_BIT), 631 /* already went through poll handler */ 632 REQ_F_POLLED = IO_REQ_FLAG(REQ_F_POLLED_BIT), 633 /* every req only blocks once in hybrid poll */ 634 REQ_F_IOPOLL_STATE = IO_REQ_FLAG(REQ_F_HYBRID_IOPOLL_STATE_BIT), 635 /* buffer already selected */ 636 REQ_F_BUFFER_SELECTED = IO_REQ_FLAG(REQ_F_BUFFER_SELECTED_BIT), 637 /* buffer selected from ring, needs commit */ 638 REQ_F_BUFFER_RING = IO_REQ_FLAG(REQ_F_BUFFER_RING_BIT), 639 /* caller should reissue async */ 640 REQ_F_REISSUE = IO_REQ_FLAG(REQ_F_REISSUE_BIT), 641 /* supports async reads/writes */ 642 REQ_F_SUPPORT_NOWAIT = IO_REQ_FLAG(REQ_F_SUPPORT_NOWAIT_BIT), 643 /* regular file */ 644 REQ_F_ISREG = IO_REQ_FLAG(REQ_F_ISREG_BIT), 645 /* has creds assigned */ 646 REQ_F_CREDS = IO_REQ_FLAG(REQ_F_CREDS_BIT), 647 /* skip refcounting if not set */ 648 REQ_F_REFCOUNT = IO_REQ_FLAG(REQ_F_REFCOUNT_BIT), 649 /* there is a linked timeout that has to be armed */ 650 REQ_F_ARM_LTIMEOUT = IO_REQ_FLAG(REQ_F_ARM_LTIMEOUT_BIT), 651 /* ->async_data allocated */ 652 REQ_F_ASYNC_DATA = IO_REQ_FLAG(REQ_F_ASYNC_DATA_BIT), 653 /* don't post CQEs while failing linked requests */ 654 REQ_F_SKIP_LINK_CQES = IO_REQ_FLAG(REQ_F_SKIP_LINK_CQES_BIT), 655 /* single poll may be active */ 656 REQ_F_SINGLE_POLL = IO_REQ_FLAG(REQ_F_SINGLE_POLL_BIT), 657 /* double poll may active */ 658 REQ_F_DOUBLE_POLL = IO_REQ_FLAG(REQ_F_DOUBLE_POLL_BIT), 659 /* request posts multiple completions, should be set at prep time */ 660 REQ_F_MULTISHOT = IO_REQ_FLAG(REQ_F_MULTISHOT_BIT), 661 /* fast poll multishot mode */ 662 REQ_F_APOLL_MULTISHOT = IO_REQ_FLAG(REQ_F_APOLL_MULTISHOT_BIT), 663 /* recvmsg special flag, clear EPOLLIN */ 664 REQ_F_CLEAR_POLLIN = IO_REQ_FLAG(REQ_F_CLEAR_POLLIN_BIT), 665 /* don't use lazy poll wake for this request */ 666 REQ_F_POLL_NO_LAZY = IO_REQ_FLAG(REQ_F_POLL_NO_LAZY_BIT), 667 /* file is pollable */ 668 REQ_F_CAN_POLL = IO_REQ_FLAG(REQ_F_CAN_POLL_BIT), 669 /* buffer list was empty after selection of buffer */ 670 REQ_F_BL_EMPTY = IO_REQ_FLAG(REQ_F_BL_EMPTY_BIT), 671 /* don't recycle provided buffers for this request */ 672 REQ_F_BL_NO_RECYCLE = IO_REQ_FLAG(REQ_F_BL_NO_RECYCLE_BIT), 673 /* buffer ring head needs incrementing on put */ 674 REQ_F_BUFFERS_COMMIT = IO_REQ_FLAG(REQ_F_BUFFERS_COMMIT_BIT), 675 /* buf node is valid */ 676 REQ_F_BUF_NODE = IO_REQ_FLAG(REQ_F_BUF_NODE_BIT), 677 /* incremental buffer consumption, more space available */ 678 REQ_F_BUF_MORE = IO_REQ_FLAG(REQ_F_BUF_MORE_BIT), 679 /* request has read/write metadata assigned */ 680 REQ_F_HAS_METADATA = IO_REQ_FLAG(REQ_F_HAS_METADATA_BIT), 681 /* 682 * For vectored fixed buffers, resolve iovec to registered buffers. 683 * For SEND_ZC, whether to import buffers (i.e. the first issue). 684 */ 685 REQ_F_IMPORT_BUFFER = IO_REQ_FLAG(REQ_F_IMPORT_BUFFER_BIT), 686 /* ->sqe_copy() has been called, if necessary */ 687 REQ_F_SQE_COPIED = IO_REQ_FLAG(REQ_F_SQE_COPIED_BIT), 688 /* request must be iopolled to completion (set in ->issue()) */ 689 REQ_F_IOPOLL = IO_REQ_FLAG(REQ_F_IOPOLL_BIT), 690 }; 691 692 struct io_tw_req { 693 struct io_kiocb *req; 694 }; 695 696 typedef void (*io_req_tw_func_t)(struct io_tw_req tw_req, io_tw_token_t tw); 697 698 struct io_task_work { 699 struct llist_node node; 700 io_req_tw_func_t func; 701 }; 702 703 struct io_cqe { 704 __u64 user_data; 705 __s32 res; 706 /* fd initially, then cflags for completion */ 707 union { 708 __u32 flags; 709 int fd; 710 }; 711 }; 712 713 /* 714 * Each request type overlays its private data structure on top of this one. 715 * They must not exceed this one in size. 716 */ 717 struct io_cmd_data { 718 struct file *file; 719 /* each command gets 56 bytes of data */ 720 __u8 data[56]; 721 }; 722 723 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz) 724 { 725 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data)); 726 } 727 #define io_kiocb_to_cmd(req, cmd_type) ( \ 728 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \ 729 ((cmd_type *)&(req)->cmd) \ 730 ) 731 732 static inline struct io_kiocb *cmd_to_io_kiocb(void *ptr) 733 { 734 return ptr; 735 } 736 737 struct io_kiocb { 738 union { 739 /* 740 * NOTE! Each of the io_kiocb union members has the file pointer 741 * as the first entry in their struct definition. So you can 742 * access the file pointer through any of the sub-structs, 743 * or directly as just 'file' in this struct. 744 */ 745 struct file *file; 746 struct io_cmd_data cmd; 747 }; 748 749 u8 opcode; 750 /* polled IO has completed */ 751 u8 iopoll_completed; 752 /* 753 * Can be either a fixed buffer index, or used with provided buffers. 754 * For the latter, it points to the selected buffer ID. 755 */ 756 u16 buf_index; 757 758 /* REQ_F_* flags */ 759 io_req_flags_t flags; 760 761 struct io_cqe cqe; 762 763 struct io_ring_ctx *ctx; 764 struct io_uring_task *tctx; 765 766 union { 767 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */ 768 struct io_buffer *kbuf; 769 770 struct io_rsrc_node *buf_node; 771 }; 772 773 union { 774 /* used by request caches, completion batching and iopoll */ 775 struct io_wq_work_node comp_list; 776 /* cache ->apoll->events */ 777 __poll_t apoll_events; 778 }; 779 780 struct io_rsrc_node *file_node; 781 782 atomic_t refs; 783 bool cancel_seq_set; 784 785 union { 786 struct io_task_work io_task_work; 787 /* For IOPOLL setup queues, with hybrid polling */ 788 u64 iopoll_start; 789 }; 790 791 union { 792 /* 793 * for polled requests, i.e. IORING_OP_POLL_ADD and async armed 794 * poll 795 */ 796 struct hlist_node hash_node; 797 /* IOPOLL completion handling */ 798 struct list_head iopoll_node; 799 /* for private io_kiocb freeing */ 800 struct rcu_head rcu_head; 801 }; 802 /* internal polling, see IORING_FEAT_FAST_POLL */ 803 struct async_poll *apoll; 804 /* opcode allocated if it needs to store data for async defer */ 805 void *async_data; 806 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */ 807 atomic_t poll_refs; 808 struct io_kiocb *link; 809 /* custom credentials, valid IFF REQ_F_CREDS is set */ 810 const struct cred *creds; 811 struct io_wq_work work; 812 813 struct io_big_cqe { 814 u64 extra1; 815 u64 extra2; 816 } big_cqe; 817 }; 818 819 struct io_overflow_cqe { 820 struct list_head list; 821 struct io_uring_cqe cqe; 822 }; 823 #endif 824