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 wirh 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;
208 unsigned short submit_nr;
209 unsigned int cqes_count;
210 struct blk_plug plug;
214 struct eventfd_ctx *cq_ev_fd;
215 unsigned int eventfd_async: 1;
221 struct io_alloc_cache {
222 struct io_wq_work_node list;
223 unsigned int nr_cached;
224 unsigned int max_cached;
229 /* const or read-mostly hot data */
232 unsigned int drain_next: 1;
233 unsigned int restricted: 1;
234 unsigned int off_timeout_used: 1;
235 unsigned int drain_active: 1;
236 unsigned int has_evfd: 1;
237 /* all CQEs should be posted only by the submitter task */
238 unsigned int task_complete: 1;
239 unsigned int lockless_cq: 1;
240 unsigned int syscall_iopoll: 1;
241 unsigned int poll_activated: 1;
242 unsigned int drain_disabled: 1;
243 unsigned int compat: 1;
244 unsigned int iowq_limits_set : 1;
246 struct task_struct *submitter_task;
247 struct io_rings *rings;
248 struct percpu_ref refs;
250 enum task_work_notify_mode notify_method;
251 unsigned sq_thread_idle;
252 } ____cacheline_aligned_in_smp;
254 /* submission data */
256 struct mutex uring_lock;
259 * Ring buffer of indices into array of io_uring_sqe, which is
260 * mmapped by the application using the IORING_OFF_SQES offset.
262 * This indirection could e.g. be used to assign fixed
263 * io_uring_sqe entries to operations and only submit them to
264 * the queue when needed.
266 * The kernel modifies neither the indices array nor the entries
270 struct io_uring_sqe *sq_sqes;
271 unsigned cached_sq_head;
275 * Fixed resources fast path, should be accessed only under
276 * uring_lock, and updated through io_uring_register(2)
278 struct io_rsrc_node *rsrc_node;
282 * ->iopoll_list is protected by the ctx->uring_lock for
283 * io_uring instances that don't use IORING_SETUP_SQPOLL.
284 * For SQPOLL, only the single threaded io_sq_thread() will
285 * manipulate the list, hence no extra locking is needed there.
287 bool poll_multi_queue;
288 struct io_wq_work_list iopoll_list;
290 struct io_file_table file_table;
291 struct io_mapped_ubuf **user_bufs;
292 unsigned nr_user_files;
293 unsigned nr_user_bufs;
295 struct io_submit_state submit_state;
297 struct io_buffer_list *io_bl;
298 struct xarray io_bl_xa;
300 struct io_hash_table cancel_table_locked;
301 struct io_alloc_cache apoll_cache;
302 struct io_alloc_cache netmsg_cache;
305 * Any cancelable uring_cmd is added to this list in
306 * ->uring_cmd() by io_uring_cmd_insert_cancelable()
308 struct hlist_head cancelable_uring_cmd;
309 } ____cacheline_aligned_in_smp;
313 * We cache a range of free CQEs we can use, once exhausted it
314 * should go through a slower range setup, see __io_get_cqe()
316 struct io_uring_cqe *cqe_cached;
317 struct io_uring_cqe *cqe_sentinel;
319 unsigned cached_cq_tail;
321 struct io_ev_fd __rcu *io_ev_fd;
323 } ____cacheline_aligned_in_smp;
326 * task_work and async notification delivery cacheline. Expected to
327 * regularly bounce b/w CPUs.
330 struct llist_head work_llist;
331 unsigned long check_cq;
333 atomic_t cq_timeouts;
334 struct wait_queue_head cq_wait;
335 } ____cacheline_aligned_in_smp;
339 spinlock_t timeout_lock;
340 struct list_head timeout_list;
341 struct list_head ltimeout_list;
342 unsigned cq_last_tm_flush;
343 } ____cacheline_aligned_in_smp;
345 struct io_uring_cqe completion_cqes[16];
347 spinlock_t completion_lock;
349 /* IRQ completion list, under ->completion_lock */
350 unsigned int locked_free_nr;
351 struct io_wq_work_list locked_free_list;
353 struct list_head io_buffers_comp;
354 struct list_head cq_overflow_list;
355 struct io_hash_table cancel_table;
357 struct hlist_head waitid_list;
360 struct hlist_head futex_list;
361 struct io_alloc_cache futex_cache;
364 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
365 struct io_sq_data *sq_data; /* if using sq thread polling */
367 struct wait_queue_head sqo_sq_wait;
368 struct list_head sqd_list;
370 unsigned int file_alloc_start;
371 unsigned int file_alloc_end;
373 struct list_head io_buffers_cache;
375 /* deferred free list, protected by ->uring_lock */
376 struct hlist_head io_buf_list;
378 /* Keep this last, we don't need it for the fast path */
379 struct wait_queue_head poll_wq;
380 struct io_restriction restrictions;
382 /* slow path rsrc auxilary data, used by update/register */
383 struct io_mapped_ubuf *dummy_ubuf;
384 struct io_rsrc_data *file_data;
385 struct io_rsrc_data *buf_data;
387 /* protected by ->uring_lock */
388 struct list_head rsrc_ref_list;
389 struct io_alloc_cache rsrc_node_cache;
390 struct wait_queue_head rsrc_quiesce_wq;
391 unsigned rsrc_quiesce;
394 struct xarray personalities;
396 /* hashed buffered write serialization */
397 struct io_wq_hash *hash_map;
399 /* Only used for accounting purposes */
400 struct user_struct *user;
401 struct mm_struct *mm_account;
403 /* ctx exit and cancelation */
404 struct llist_head fallback_llist;
405 struct delayed_work fallback_work;
406 struct work_struct exit_work;
407 struct list_head tctx_list;
408 struct completion ref_comp;
410 /* io-wq management, e.g. thread count */
413 struct callback_head poll_wq_task_work;
414 struct list_head defer_list;
416 #ifdef CONFIG_NET_RX_BUSY_POLL
417 struct list_head napi_list; /* track busy poll napi_id */
418 spinlock_t napi_lock; /* napi_list lock */
420 /* napi busy poll default timeout */
421 unsigned int napi_busy_poll_to;
422 bool napi_prefer_busy_poll;
425 DECLARE_HASHTABLE(napi_ht, 4);
428 /* protected by ->completion_lock */
429 unsigned evfd_last_cq_tail;
432 * If IORING_SETUP_NO_MMAP is used, then the below holds
433 * the gup'ed pages for the two rings, and the sqes.
435 unsigned short n_ring_pages;
436 unsigned short n_sqe_pages;
437 struct page **ring_pages;
438 struct page **sqe_pages;
442 /* ->uring_lock is taken, callbacks can use io_tw_lock to lock it */
447 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
448 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
449 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
450 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
451 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
452 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
453 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
455 /* first byte is taken by user flags, shift it to not overlap */
460 REQ_F_LINK_TIMEOUT_BIT,
461 REQ_F_NEED_CLEANUP_BIT,
463 REQ_F_BUFFER_SELECTED_BIT,
464 REQ_F_BUFFER_RING_BIT,
468 REQ_F_ARM_LTIMEOUT_BIT,
469 REQ_F_ASYNC_DATA_BIT,
470 REQ_F_SKIP_LINK_CQES_BIT,
471 REQ_F_SINGLE_POLL_BIT,
472 REQ_F_DOUBLE_POLL_BIT,
473 REQ_F_APOLL_MULTISHOT_BIT,
474 REQ_F_CLEAR_POLLIN_BIT,
475 REQ_F_HASH_LOCKED_BIT,
476 /* keep async read/write and isreg together and in order */
477 REQ_F_SUPPORT_NOWAIT_BIT,
479 REQ_F_POLL_NO_LAZY_BIT,
480 REQ_F_CANCEL_SEQ_BIT,
483 REQ_F_BL_NO_RECYCLE_BIT,
485 /* not a real bit, just to check we're not overflowing the space */
489 typedef u64 __bitwise io_req_flags_t;
490 #define IO_REQ_FLAG(bitno) ((__force io_req_flags_t) BIT_ULL((bitno)))
494 REQ_F_FIXED_FILE = IO_REQ_FLAG(REQ_F_FIXED_FILE_BIT),
495 /* drain existing IO first */
496 REQ_F_IO_DRAIN = IO_REQ_FLAG(REQ_F_IO_DRAIN_BIT),
498 REQ_F_LINK = IO_REQ_FLAG(REQ_F_LINK_BIT),
499 /* doesn't sever on completion < 0 */
500 REQ_F_HARDLINK = IO_REQ_FLAG(REQ_F_HARDLINK_BIT),
502 REQ_F_FORCE_ASYNC = IO_REQ_FLAG(REQ_F_FORCE_ASYNC_BIT),
503 /* IOSQE_BUFFER_SELECT */
504 REQ_F_BUFFER_SELECT = IO_REQ_FLAG(REQ_F_BUFFER_SELECT_BIT),
505 /* IOSQE_CQE_SKIP_SUCCESS */
506 REQ_F_CQE_SKIP = IO_REQ_FLAG(REQ_F_CQE_SKIP_BIT),
508 /* fail rest of links */
509 REQ_F_FAIL = IO_REQ_FLAG(REQ_F_FAIL_BIT),
510 /* on inflight list, should be cancelled and waited on exit reliably */
511 REQ_F_INFLIGHT = IO_REQ_FLAG(REQ_F_INFLIGHT_BIT),
512 /* read/write uses file position */
513 REQ_F_CUR_POS = IO_REQ_FLAG(REQ_F_CUR_POS_BIT),
514 /* must not punt to workers */
515 REQ_F_NOWAIT = IO_REQ_FLAG(REQ_F_NOWAIT_BIT),
516 /* has or had linked timeout */
517 REQ_F_LINK_TIMEOUT = IO_REQ_FLAG(REQ_F_LINK_TIMEOUT_BIT),
519 REQ_F_NEED_CLEANUP = IO_REQ_FLAG(REQ_F_NEED_CLEANUP_BIT),
520 /* already went through poll handler */
521 REQ_F_POLLED = IO_REQ_FLAG(REQ_F_POLLED_BIT),
522 /* buffer already selected */
523 REQ_F_BUFFER_SELECTED = IO_REQ_FLAG(REQ_F_BUFFER_SELECTED_BIT),
524 /* buffer selected from ring, needs commit */
525 REQ_F_BUFFER_RING = IO_REQ_FLAG(REQ_F_BUFFER_RING_BIT),
526 /* caller should reissue async */
527 REQ_F_REISSUE = IO_REQ_FLAG(REQ_F_REISSUE_BIT),
528 /* supports async reads/writes */
529 REQ_F_SUPPORT_NOWAIT = IO_REQ_FLAG(REQ_F_SUPPORT_NOWAIT_BIT),
531 REQ_F_ISREG = IO_REQ_FLAG(REQ_F_ISREG_BIT),
532 /* has creds assigned */
533 REQ_F_CREDS = IO_REQ_FLAG(REQ_F_CREDS_BIT),
534 /* skip refcounting if not set */
535 REQ_F_REFCOUNT = IO_REQ_FLAG(REQ_F_REFCOUNT_BIT),
536 /* there is a linked timeout that has to be armed */
537 REQ_F_ARM_LTIMEOUT = IO_REQ_FLAG(REQ_F_ARM_LTIMEOUT_BIT),
538 /* ->async_data allocated */
539 REQ_F_ASYNC_DATA = IO_REQ_FLAG(REQ_F_ASYNC_DATA_BIT),
540 /* don't post CQEs while failing linked requests */
541 REQ_F_SKIP_LINK_CQES = IO_REQ_FLAG(REQ_F_SKIP_LINK_CQES_BIT),
542 /* single poll may be active */
543 REQ_F_SINGLE_POLL = IO_REQ_FLAG(REQ_F_SINGLE_POLL_BIT),
544 /* double poll may active */
545 REQ_F_DOUBLE_POLL = IO_REQ_FLAG(REQ_F_DOUBLE_POLL_BIT),
546 /* fast poll multishot mode */
547 REQ_F_APOLL_MULTISHOT = IO_REQ_FLAG(REQ_F_APOLL_MULTISHOT_BIT),
548 /* recvmsg special flag, clear EPOLLIN */
549 REQ_F_CLEAR_POLLIN = IO_REQ_FLAG(REQ_F_CLEAR_POLLIN_BIT),
550 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
551 REQ_F_HASH_LOCKED = IO_REQ_FLAG(REQ_F_HASH_LOCKED_BIT),
552 /* don't use lazy poll wake for this request */
553 REQ_F_POLL_NO_LAZY = IO_REQ_FLAG(REQ_F_POLL_NO_LAZY_BIT),
554 /* cancel sequence is set and valid */
555 REQ_F_CANCEL_SEQ = IO_REQ_FLAG(REQ_F_CANCEL_SEQ_BIT),
556 /* file is pollable */
557 REQ_F_CAN_POLL = IO_REQ_FLAG(REQ_F_CAN_POLL_BIT),
558 /* buffer list was empty after selection of buffer */
559 REQ_F_BL_EMPTY = IO_REQ_FLAG(REQ_F_BL_EMPTY_BIT),
560 /* don't recycle provided buffers for this request */
561 REQ_F_BL_NO_RECYCLE = IO_REQ_FLAG(REQ_F_BL_NO_RECYCLE_BIT),
564 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
566 struct io_task_work {
567 struct llist_node node;
568 io_req_tw_func_t func;
574 /* fd initially, then cflags for completion */
582 * Each request type overlays its private data structure on top of this one.
583 * They must not exceed this one in size.
587 /* each command gets 56 bytes of data */
591 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
593 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
595 #define io_kiocb_to_cmd(req, cmd_type) ( \
596 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
597 ((cmd_type *)&(req)->cmd) \
599 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
604 * NOTE! Each of the io_kiocb union members has the file pointer
605 * as the first entry in their struct definition. So you can
606 * access the file pointer through any of the sub-structs,
607 * or directly as just 'file' in this struct.
610 struct io_cmd_data cmd;
614 /* polled IO has completed */
617 * Can be either a fixed buffer index, or used with provided buffers.
618 * For the latter, before issue it points to the buffer group ID,
619 * and after selection it points to the buffer ID itself.
626 io_req_flags_t flags;
630 struct io_ring_ctx *ctx;
631 struct task_struct *task;
634 /* store used ubuf, so we can prevent reloading */
635 struct io_mapped_ubuf *imu;
637 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
638 struct io_buffer *kbuf;
641 * stores buffer ID for ring provided buffers, valid IFF
642 * REQ_F_BUFFER_RING is set.
644 struct io_buffer_list *buf_list;
648 /* used by request caches, completion batching and iopoll */
649 struct io_wq_work_node comp_list;
650 /* cache ->apoll->events */
651 __poll_t apoll_events;
654 struct io_rsrc_node *rsrc_node;
658 struct io_task_work io_task_work;
659 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
660 struct hlist_node hash_node;
661 /* internal polling, see IORING_FEAT_FAST_POLL */
662 struct async_poll *apoll;
663 /* opcode allocated if it needs to store data for async defer */
665 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
666 struct io_kiocb *link;
667 /* custom credentials, valid IFF REQ_F_CREDS is set */
668 const struct cred *creds;
669 struct io_wq_work work;
677 struct io_overflow_cqe {
678 struct list_head list;
679 struct io_uring_cqe cqe;