1 #ifndef IO_URING_TYPES_H
2 #define IO_URING_TYPES_H
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>
13 * A hint to not wake right away but delay until there are enough of
14 * tw's queued to match the number of CQEs the task is waiting for.
16 * Must not be used with requests generating more than one CQE.
17 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
19 IOU_F_TWQ_LAZY_WAKE = 1,
22 enum io_uring_cmd_flags {
23 IO_URING_F_COMPLETE_DEFER = 1,
24 IO_URING_F_UNLOCKED = 2,
25 /* the request is executed from poll, it should not be freed */
26 IO_URING_F_MULTISHOT = 4,
27 /* executed by io-wq */
29 /* int's last bit, sign checks are usually faster than a bit test */
30 IO_URING_F_NONBLOCK = INT_MIN,
32 /* ctx state flags, for URING_CMD */
33 IO_URING_F_SQE128 = (1 << 8),
34 IO_URING_F_CQE32 = (1 << 9),
35 IO_URING_F_IOPOLL = (1 << 10),
37 /* set when uring wants to cancel a previously issued command */
38 IO_URING_F_CANCEL = (1 << 11),
39 IO_URING_F_COMPAT = (1 << 12),
42 struct io_wq_work_node {
43 struct io_wq_work_node *next;
46 struct io_wq_work_list {
47 struct io_wq_work_node *first;
48 struct io_wq_work_node *last;
52 struct io_wq_work_node list;
54 /* place it here instead of io_kiocb as it fills padding and saves 4B */
58 struct io_fixed_file {
59 /* file * with additional FFS_* flags */
60 unsigned long file_ptr;
63 struct io_file_table {
64 struct io_fixed_file *files;
65 unsigned long *bitmap;
66 unsigned int alloc_hint;
69 struct io_hash_bucket {
71 struct hlist_head list;
72 } ____cacheline_aligned_in_smp;
74 struct io_hash_table {
75 struct io_hash_bucket *hbs;
80 * Arbitrary limit, can be raised if need be
82 #define IO_RINGFD_REG_MAX 16
84 struct io_uring_task {
87 const struct io_ring_ctx *last;
89 struct file *registered_rings[IO_RINGFD_REG_MAX];
92 struct wait_queue_head wait;
94 atomic_t inflight_tracked;
95 struct percpu_counter inflight;
97 struct { /* task_work */
98 struct llist_head task_list;
99 struct callback_head task_work;
100 } ____cacheline_aligned_in_smp;
109 * This data is shared with the application through the mmap at offsets
110 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
112 * The offsets to the member fields are published through struct
113 * io_sqring_offsets when calling io_uring_setup.
117 * Head and tail offsets into the ring; the offsets need to be
118 * masked to get valid indices.
120 * The kernel controls head of the sq ring and the tail of the cq ring,
121 * and the application controls tail of the sq ring and the head of the
124 struct io_uring sq, cq;
126 * Bitmasks to apply to head and tail offsets (constant, equals
129 u32 sq_ring_mask, cq_ring_mask;
130 /* Ring sizes (constant, power of 2) */
131 u32 sq_ring_entries, cq_ring_entries;
133 * Number of invalid entries dropped by the kernel due to
134 * invalid index stored in array
136 * Written by the kernel, shouldn't be modified by the
137 * application (i.e. get number of "new events" by comparing to
140 * After a new SQ head value was read by the application this
141 * counter includes all submissions that were dropped reaching
142 * the new SQ head (and possibly more).
148 * Written by the kernel, shouldn't be modified by the
151 * The application needs a full memory barrier before checking
152 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
158 * Written by the application, shouldn't be modified by the
163 * Number of completion events lost because the queue was full;
164 * this should be avoided by the application by making sure
165 * there are not more requests pending than there is space in
166 * the completion queue.
168 * Written by the kernel, shouldn't be modified by the
169 * application (i.e. get number of "new events" by comparing to
172 * As completion events come in out of order this counter is not
173 * ordered with any other data.
177 * Ring buffer of completion events.
179 * The kernel writes completion events fresh every time they are
180 * produced, so the application is allowed to modify pending
183 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
186 struct io_restriction {
187 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
188 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
189 u8 sqe_flags_allowed;
190 u8 sqe_flags_required;
194 struct io_submit_link {
195 struct io_kiocb *head;
196 struct io_kiocb *last;
199 struct io_submit_state {
200 /* inline/task_work completion list, under ->uring_lock */
201 struct io_wq_work_node free_list;
202 /* batch completion logic */
203 struct io_wq_work_list compl_reqs;
204 struct io_submit_link link;
209 unsigned short submit_nr;
210 struct blk_plug plug;
213 struct io_alloc_cache {
215 unsigned int nr_cached;
216 unsigned int max_cached;
221 /* const or read-mostly hot data */
224 unsigned int drain_next: 1;
225 unsigned int restricted: 1;
226 unsigned int off_timeout_used: 1;
227 unsigned int drain_active: 1;
228 unsigned int has_evfd: 1;
229 /* all CQEs should be posted only by the submitter task */
230 unsigned int task_complete: 1;
231 unsigned int lockless_cq: 1;
232 unsigned int syscall_iopoll: 1;
233 unsigned int poll_activated: 1;
234 unsigned int drain_disabled: 1;
235 unsigned int compat: 1;
236 unsigned int iowq_limits_set : 1;
238 struct task_struct *submitter_task;
239 struct io_rings *rings;
240 struct percpu_ref refs;
242 enum task_work_notify_mode notify_method;
243 unsigned sq_thread_idle;
244 } ____cacheline_aligned_in_smp;
246 /* submission data */
248 struct mutex uring_lock;
251 * Ring buffer of indices into array of io_uring_sqe, which is
252 * mmapped by the application using the IORING_OFF_SQES offset.
254 * This indirection could e.g. be used to assign fixed
255 * io_uring_sqe entries to operations and only submit them to
256 * the queue when needed.
258 * The kernel modifies neither the indices array nor the entries
262 struct io_uring_sqe *sq_sqes;
263 unsigned cached_sq_head;
267 * Fixed resources fast path, should be accessed only under
268 * uring_lock, and updated through io_uring_register(2)
270 struct io_rsrc_node *rsrc_node;
274 * ->iopoll_list is protected by the ctx->uring_lock for
275 * io_uring instances that don't use IORING_SETUP_SQPOLL.
276 * For SQPOLL, only the single threaded io_sq_thread() will
277 * manipulate the list, hence no extra locking is needed there.
279 bool poll_multi_queue;
280 struct io_wq_work_list iopoll_list;
282 struct io_file_table file_table;
283 struct io_mapped_ubuf **user_bufs;
284 unsigned nr_user_files;
285 unsigned nr_user_bufs;
287 struct io_submit_state submit_state;
289 struct xarray io_bl_xa;
291 struct io_hash_table cancel_table_locked;
292 struct io_alloc_cache apoll_cache;
293 struct io_alloc_cache netmsg_cache;
294 struct io_alloc_cache rw_cache;
295 struct io_alloc_cache uring_cache;
298 * Any cancelable uring_cmd is added to this list in
299 * ->uring_cmd() by io_uring_cmd_insert_cancelable()
301 struct hlist_head cancelable_uring_cmd;
302 } ____cacheline_aligned_in_smp;
306 * We cache a range of free CQEs we can use, once exhausted it
307 * should go through a slower range setup, see __io_get_cqe()
309 struct io_uring_cqe *cqe_cached;
310 struct io_uring_cqe *cqe_sentinel;
312 unsigned cached_cq_tail;
314 struct io_ev_fd __rcu *io_ev_fd;
316 } ____cacheline_aligned_in_smp;
319 * task_work and async notification delivery cacheline. Expected to
320 * regularly bounce b/w CPUs.
323 struct llist_head work_llist;
324 unsigned long check_cq;
326 atomic_t cq_timeouts;
327 struct wait_queue_head cq_wait;
328 } ____cacheline_aligned_in_smp;
332 spinlock_t timeout_lock;
333 struct list_head timeout_list;
334 struct list_head ltimeout_list;
335 unsigned cq_last_tm_flush;
336 } ____cacheline_aligned_in_smp;
338 spinlock_t completion_lock;
340 struct list_head io_buffers_comp;
341 struct list_head cq_overflow_list;
342 struct io_hash_table cancel_table;
344 struct hlist_head waitid_list;
347 struct hlist_head futex_list;
348 struct io_alloc_cache futex_cache;
351 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
352 struct io_sq_data *sq_data; /* if using sq thread polling */
354 struct wait_queue_head sqo_sq_wait;
355 struct list_head sqd_list;
357 unsigned int file_alloc_start;
358 unsigned int file_alloc_end;
360 struct list_head io_buffers_cache;
362 /* Keep this last, we don't need it for the fast path */
363 struct wait_queue_head poll_wq;
364 struct io_restriction restrictions;
366 /* slow path rsrc auxilary data, used by update/register */
367 struct io_rsrc_data *file_data;
368 struct io_rsrc_data *buf_data;
370 /* protected by ->uring_lock */
371 struct list_head rsrc_ref_list;
372 struct io_alloc_cache rsrc_node_cache;
373 struct wait_queue_head rsrc_quiesce_wq;
374 unsigned rsrc_quiesce;
377 struct xarray personalities;
379 /* hashed buffered write serialization */
380 struct io_wq_hash *hash_map;
382 /* Only used for accounting purposes */
383 struct user_struct *user;
384 struct mm_struct *mm_account;
386 /* ctx exit and cancelation */
387 struct llist_head fallback_llist;
388 struct delayed_work fallback_work;
389 struct work_struct exit_work;
390 struct list_head tctx_list;
391 struct completion ref_comp;
393 /* io-wq management, e.g. thread count */
396 struct callback_head poll_wq_task_work;
397 struct list_head defer_list;
399 struct io_alloc_cache msg_cache;
402 #ifdef CONFIG_NET_RX_BUSY_POLL
403 struct list_head napi_list; /* track busy poll napi_id */
404 spinlock_t napi_lock; /* napi_list lock */
406 /* napi busy poll default timeout */
407 unsigned int napi_busy_poll_to;
408 bool napi_prefer_busy_poll;
411 DECLARE_HASHTABLE(napi_ht, 4);
414 /* protected by ->completion_lock */
415 unsigned evfd_last_cq_tail;
418 * If IORING_SETUP_NO_MMAP is used, then the below holds
419 * the gup'ed pages for the two rings, and the sqes.
421 unsigned short n_ring_pages;
422 unsigned short n_sqe_pages;
423 struct page **ring_pages;
424 struct page **sqe_pages;
431 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
432 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
433 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
434 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
435 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
436 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
437 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
439 /* first byte is taken by user flags, shift it to not overlap */
444 REQ_F_LINK_TIMEOUT_BIT,
445 REQ_F_NEED_CLEANUP_BIT,
447 REQ_F_BUFFER_SELECTED_BIT,
448 REQ_F_BUFFER_RING_BIT,
452 REQ_F_ARM_LTIMEOUT_BIT,
453 REQ_F_ASYNC_DATA_BIT,
454 REQ_F_SKIP_LINK_CQES_BIT,
455 REQ_F_SINGLE_POLL_BIT,
456 REQ_F_DOUBLE_POLL_BIT,
457 REQ_F_APOLL_MULTISHOT_BIT,
458 REQ_F_CLEAR_POLLIN_BIT,
459 REQ_F_HASH_LOCKED_BIT,
460 /* keep async read/write and isreg together and in order */
461 REQ_F_SUPPORT_NOWAIT_BIT,
463 REQ_F_POLL_NO_LAZY_BIT,
464 REQ_F_CANCEL_SEQ_BIT,
467 REQ_F_BL_NO_RECYCLE_BIT,
468 REQ_F_BUFFERS_COMMIT_BIT,
470 /* not a real bit, just to check we're not overflowing the space */
474 typedef u64 __bitwise io_req_flags_t;
475 #define IO_REQ_FLAG(bitno) ((__force io_req_flags_t) BIT_ULL((bitno)))
479 REQ_F_FIXED_FILE = IO_REQ_FLAG(REQ_F_FIXED_FILE_BIT),
480 /* drain existing IO first */
481 REQ_F_IO_DRAIN = IO_REQ_FLAG(REQ_F_IO_DRAIN_BIT),
483 REQ_F_LINK = IO_REQ_FLAG(REQ_F_LINK_BIT),
484 /* doesn't sever on completion < 0 */
485 REQ_F_HARDLINK = IO_REQ_FLAG(REQ_F_HARDLINK_BIT),
487 REQ_F_FORCE_ASYNC = IO_REQ_FLAG(REQ_F_FORCE_ASYNC_BIT),
488 /* IOSQE_BUFFER_SELECT */
489 REQ_F_BUFFER_SELECT = IO_REQ_FLAG(REQ_F_BUFFER_SELECT_BIT),
490 /* IOSQE_CQE_SKIP_SUCCESS */
491 REQ_F_CQE_SKIP = IO_REQ_FLAG(REQ_F_CQE_SKIP_BIT),
493 /* fail rest of links */
494 REQ_F_FAIL = IO_REQ_FLAG(REQ_F_FAIL_BIT),
495 /* on inflight list, should be cancelled and waited on exit reliably */
496 REQ_F_INFLIGHT = IO_REQ_FLAG(REQ_F_INFLIGHT_BIT),
497 /* read/write uses file position */
498 REQ_F_CUR_POS = IO_REQ_FLAG(REQ_F_CUR_POS_BIT),
499 /* must not punt to workers */
500 REQ_F_NOWAIT = IO_REQ_FLAG(REQ_F_NOWAIT_BIT),
501 /* has or had linked timeout */
502 REQ_F_LINK_TIMEOUT = IO_REQ_FLAG(REQ_F_LINK_TIMEOUT_BIT),
504 REQ_F_NEED_CLEANUP = IO_REQ_FLAG(REQ_F_NEED_CLEANUP_BIT),
505 /* already went through poll handler */
506 REQ_F_POLLED = IO_REQ_FLAG(REQ_F_POLLED_BIT),
507 /* buffer already selected */
508 REQ_F_BUFFER_SELECTED = IO_REQ_FLAG(REQ_F_BUFFER_SELECTED_BIT),
509 /* buffer selected from ring, needs commit */
510 REQ_F_BUFFER_RING = IO_REQ_FLAG(REQ_F_BUFFER_RING_BIT),
511 /* caller should reissue async */
512 REQ_F_REISSUE = IO_REQ_FLAG(REQ_F_REISSUE_BIT),
513 /* supports async reads/writes */
514 REQ_F_SUPPORT_NOWAIT = IO_REQ_FLAG(REQ_F_SUPPORT_NOWAIT_BIT),
516 REQ_F_ISREG = IO_REQ_FLAG(REQ_F_ISREG_BIT),
517 /* has creds assigned */
518 REQ_F_CREDS = IO_REQ_FLAG(REQ_F_CREDS_BIT),
519 /* skip refcounting if not set */
520 REQ_F_REFCOUNT = IO_REQ_FLAG(REQ_F_REFCOUNT_BIT),
521 /* there is a linked timeout that has to be armed */
522 REQ_F_ARM_LTIMEOUT = IO_REQ_FLAG(REQ_F_ARM_LTIMEOUT_BIT),
523 /* ->async_data allocated */
524 REQ_F_ASYNC_DATA = IO_REQ_FLAG(REQ_F_ASYNC_DATA_BIT),
525 /* don't post CQEs while failing linked requests */
526 REQ_F_SKIP_LINK_CQES = IO_REQ_FLAG(REQ_F_SKIP_LINK_CQES_BIT),
527 /* single poll may be active */
528 REQ_F_SINGLE_POLL = IO_REQ_FLAG(REQ_F_SINGLE_POLL_BIT),
529 /* double poll may active */
530 REQ_F_DOUBLE_POLL = IO_REQ_FLAG(REQ_F_DOUBLE_POLL_BIT),
531 /* fast poll multishot mode */
532 REQ_F_APOLL_MULTISHOT = IO_REQ_FLAG(REQ_F_APOLL_MULTISHOT_BIT),
533 /* recvmsg special flag, clear EPOLLIN */
534 REQ_F_CLEAR_POLLIN = IO_REQ_FLAG(REQ_F_CLEAR_POLLIN_BIT),
535 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
536 REQ_F_HASH_LOCKED = IO_REQ_FLAG(REQ_F_HASH_LOCKED_BIT),
537 /* don't use lazy poll wake for this request */
538 REQ_F_POLL_NO_LAZY = IO_REQ_FLAG(REQ_F_POLL_NO_LAZY_BIT),
539 /* cancel sequence is set and valid */
540 REQ_F_CANCEL_SEQ = IO_REQ_FLAG(REQ_F_CANCEL_SEQ_BIT),
541 /* file is pollable */
542 REQ_F_CAN_POLL = IO_REQ_FLAG(REQ_F_CAN_POLL_BIT),
543 /* buffer list was empty after selection of buffer */
544 REQ_F_BL_EMPTY = IO_REQ_FLAG(REQ_F_BL_EMPTY_BIT),
545 /* don't recycle provided buffers for this request */
546 REQ_F_BL_NO_RECYCLE = IO_REQ_FLAG(REQ_F_BL_NO_RECYCLE_BIT),
547 /* buffer ring head needs incrementing on put */
548 REQ_F_BUFFERS_COMMIT = IO_REQ_FLAG(REQ_F_BUFFERS_COMMIT_BIT),
551 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
553 struct io_task_work {
554 struct llist_node node;
555 io_req_tw_func_t func;
561 /* fd initially, then cflags for completion */
569 * Each request type overlays its private data structure on top of this one.
570 * They must not exceed this one in size.
574 /* each command gets 56 bytes of data */
578 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
580 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
582 #define io_kiocb_to_cmd(req, cmd_type) ( \
583 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
584 ((cmd_type *)&(req)->cmd) \
586 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
591 * NOTE! Each of the io_kiocb union members has the file pointer
592 * as the first entry in their struct definition. So you can
593 * access the file pointer through any of the sub-structs,
594 * or directly as just 'file' in this struct.
597 struct io_cmd_data cmd;
601 /* polled IO has completed */
604 * Can be either a fixed buffer index, or used with provided buffers.
605 * For the latter, before issue it points to the buffer group ID,
606 * and after selection it points to the buffer ID itself.
613 io_req_flags_t flags;
617 struct io_ring_ctx *ctx;
618 struct task_struct *task;
621 /* store used ubuf, so we can prevent reloading */
622 struct io_mapped_ubuf *imu;
624 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
625 struct io_buffer *kbuf;
628 * stores buffer ID for ring provided buffers, valid IFF
629 * REQ_F_BUFFER_RING is set.
631 struct io_buffer_list *buf_list;
635 /* used by request caches, completion batching and iopoll */
636 struct io_wq_work_node comp_list;
637 /* cache ->apoll->events */
638 __poll_t apoll_events;
641 struct io_rsrc_node *rsrc_node;
645 struct io_task_work io_task_work;
646 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
647 struct hlist_node hash_node;
648 /* internal polling, see IORING_FEAT_FAST_POLL */
649 struct async_poll *apoll;
650 /* opcode allocated if it needs to store data for async defer */
652 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
654 struct io_kiocb *link;
655 /* custom credentials, valid IFF REQ_F_CREDS is set */
656 const struct cred *creds;
657 struct io_wq_work work;
665 struct io_overflow_cqe {
666 struct list_head list;
667 struct io_uring_cqe cqe;