1 #ifndef IO_URING_TYPES_H
2 #define IO_URING_TYPES_H
4 #include <linux/blkdev.h>
5 #include <linux/task_work.h>
6 #include <linux/bitmap.h>
7 #include <linux/llist.h>
8 #include <uapi/linux/io_uring.h>
10 struct io_wq_work_node {
11 struct io_wq_work_node *next;
14 struct io_wq_work_list {
15 struct io_wq_work_node *first;
16 struct io_wq_work_node *last;
20 struct io_wq_work_node list;
22 /* place it here instead of io_kiocb as it fills padding and saves 4B */
26 struct io_fixed_file {
27 /* file * with additional FFS_* flags */
28 unsigned long file_ptr;
31 struct io_file_table {
32 struct io_fixed_file *files;
33 unsigned long *bitmap;
34 unsigned int alloc_hint;
37 struct io_hash_bucket {
39 struct hlist_head list;
40 } ____cacheline_aligned_in_smp;
42 struct io_hash_table {
43 struct io_hash_bucket *hbs;
48 * Arbitrary limit, can be raised if need be
50 #define IO_RINGFD_REG_MAX 16
52 struct io_uring_task {
55 const struct io_ring_ctx *last;
57 struct file *registered_rings[IO_RINGFD_REG_MAX];
60 struct wait_queue_head wait;
62 atomic_t inflight_tracked;
63 struct percpu_counter inflight;
65 struct { /* task_work */
66 struct llist_head task_list;
67 struct callback_head task_work;
68 } ____cacheline_aligned_in_smp;
77 * This data is shared with the application through the mmap at offsets
78 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
80 * The offsets to the member fields are published through struct
81 * io_sqring_offsets when calling io_uring_setup.
85 * Head and tail offsets into the ring; the offsets need to be
86 * masked to get valid indices.
88 * The kernel controls head of the sq ring and the tail of the cq ring,
89 * and the application controls tail of the sq ring and the head of the
92 struct io_uring sq, cq;
94 * Bitmasks to apply to head and tail offsets (constant, equals
97 u32 sq_ring_mask, cq_ring_mask;
98 /* Ring sizes (constant, power of 2) */
99 u32 sq_ring_entries, cq_ring_entries;
101 * Number of invalid entries dropped by the kernel due to
102 * invalid index stored in array
104 * Written by the kernel, shouldn't be modified by the
105 * application (i.e. get number of "new events" by comparing to
108 * After a new SQ head value was read by the application this
109 * counter includes all submissions that were dropped reaching
110 * the new SQ head (and possibly more).
116 * Written by the kernel, shouldn't be modified by the
119 * The application needs a full memory barrier before checking
120 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
126 * Written by the application, shouldn't be modified by the
131 * Number of completion events lost because the queue was full;
132 * this should be avoided by the application by making sure
133 * there are not more requests pending than there is space in
134 * the completion queue.
136 * Written by the kernel, shouldn't be modified by the
137 * application (i.e. get number of "new events" by comparing to
140 * As completion events come in out of order this counter is not
141 * ordered with any other data.
145 * Ring buffer of completion events.
147 * The kernel writes completion events fresh every time they are
148 * produced, so the application is allowed to modify pending
151 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
154 struct io_restriction {
155 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
156 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
157 u8 sqe_flags_allowed;
158 u8 sqe_flags_required;
162 struct io_submit_link {
163 struct io_kiocb *head;
164 struct io_kiocb *last;
167 struct io_submit_state {
168 /* inline/task_work completion list, under ->uring_lock */
169 struct io_wq_work_node free_list;
170 /* batch completion logic */
171 struct io_wq_work_list compl_reqs;
172 struct io_submit_link link;
176 unsigned short submit_nr;
177 unsigned int cqes_count;
178 struct blk_plug plug;
182 struct eventfd_ctx *cq_ev_fd;
183 unsigned int eventfd_async: 1;
189 struct io_alloc_cache {
190 struct io_wq_work_node list;
191 unsigned int nr_cached;
192 unsigned int max_cached;
197 /* const or read-mostly hot data */
200 unsigned int drain_next: 1;
201 unsigned int restricted: 1;
202 unsigned int off_timeout_used: 1;
203 unsigned int drain_active: 1;
204 unsigned int has_evfd: 1;
205 /* all CQEs should be posted only by the submitter task */
206 unsigned int task_complete: 1;
207 unsigned int lockless_cq: 1;
208 unsigned int syscall_iopoll: 1;
209 unsigned int poll_activated: 1;
210 unsigned int drain_disabled: 1;
211 unsigned int compat: 1;
213 struct task_struct *submitter_task;
214 struct io_rings *rings;
215 struct percpu_ref refs;
217 enum task_work_notify_mode notify_method;
218 } ____cacheline_aligned_in_smp;
220 /* submission data */
222 struct mutex uring_lock;
225 * Ring buffer of indices into array of io_uring_sqe, which is
226 * mmapped by the application using the IORING_OFF_SQES offset.
228 * This indirection could e.g. be used to assign fixed
229 * io_uring_sqe entries to operations and only submit them to
230 * the queue when needed.
232 * The kernel modifies neither the indices array nor the entries
236 struct io_uring_sqe *sq_sqes;
237 unsigned cached_sq_head;
241 * Fixed resources fast path, should be accessed only under
242 * uring_lock, and updated through io_uring_register(2)
244 struct io_rsrc_node *rsrc_node;
246 struct io_file_table file_table;
247 unsigned nr_user_files;
248 unsigned nr_user_bufs;
249 struct io_mapped_ubuf **user_bufs;
251 struct io_submit_state submit_state;
253 struct io_buffer_list *io_bl;
254 struct xarray io_bl_xa;
256 struct io_hash_table cancel_table_locked;
257 struct io_alloc_cache apoll_cache;
258 struct io_alloc_cache netmsg_cache;
261 * ->iopoll_list is protected by the ctx->uring_lock for
262 * io_uring instances that don't use IORING_SETUP_SQPOLL.
263 * For SQPOLL, only the single threaded io_sq_thread() will
264 * manipulate the list, hence no extra locking is needed there.
266 struct io_wq_work_list iopoll_list;
267 bool poll_multi_queue;
270 * Any cancelable uring_cmd is added to this list in
271 * ->uring_cmd() by io_uring_cmd_insert_cancelable()
273 struct hlist_head cancelable_uring_cmd;
274 } ____cacheline_aligned_in_smp;
278 * We cache a range of free CQEs we can use, once exhausted it
279 * should go through a slower range setup, see __io_get_cqe()
281 struct io_uring_cqe *cqe_cached;
282 struct io_uring_cqe *cqe_sentinel;
284 unsigned cached_cq_tail;
286 struct io_ev_fd __rcu *io_ev_fd;
288 } ____cacheline_aligned_in_smp;
291 * task_work and async notification delivery cacheline. Expected to
292 * regularly bounce b/w CPUs.
295 struct llist_head work_llist;
296 unsigned long check_cq;
298 atomic_t cq_timeouts;
299 struct wait_queue_head cq_wait;
300 } ____cacheline_aligned_in_smp;
304 spinlock_t timeout_lock;
305 struct list_head timeout_list;
306 struct list_head ltimeout_list;
307 unsigned cq_last_tm_flush;
308 } ____cacheline_aligned_in_smp;
310 struct io_uring_cqe completion_cqes[16];
312 spinlock_t completion_lock;
314 /* IRQ completion list, under ->completion_lock */
315 struct io_wq_work_list locked_free_list;
316 unsigned int locked_free_nr;
318 struct list_head io_buffers_comp;
319 struct list_head cq_overflow_list;
320 struct io_hash_table cancel_table;
322 struct hlist_head waitid_list;
325 struct hlist_head futex_list;
326 struct io_alloc_cache futex_cache;
329 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
330 struct io_sq_data *sq_data; /* if using sq thread polling */
332 struct wait_queue_head sqo_sq_wait;
333 struct list_head sqd_list;
335 unsigned int file_alloc_start;
336 unsigned int file_alloc_end;
338 struct xarray personalities;
341 struct list_head io_buffers_cache;
343 /* deferred free list, protected by ->uring_lock */
344 struct hlist_head io_buf_list;
346 /* Keep this last, we don't need it for the fast path */
347 struct wait_queue_head poll_wq;
348 struct io_restriction restrictions;
350 /* slow path rsrc auxilary data, used by update/register */
351 struct io_mapped_ubuf *dummy_ubuf;
352 struct io_rsrc_data *file_data;
353 struct io_rsrc_data *buf_data;
355 /* protected by ->uring_lock */
356 struct list_head rsrc_ref_list;
357 struct io_alloc_cache rsrc_node_cache;
358 struct wait_queue_head rsrc_quiesce_wq;
359 unsigned rsrc_quiesce;
361 #if defined(CONFIG_UNIX)
362 struct socket *ring_sock;
364 /* hashed buffered write serialization */
365 struct io_wq_hash *hash_map;
367 /* Only used for accounting purposes */
368 struct user_struct *user;
369 struct mm_struct *mm_account;
371 /* ctx exit and cancelation */
372 struct llist_head fallback_llist;
373 struct delayed_work fallback_work;
374 struct work_struct exit_work;
375 struct list_head tctx_list;
376 struct completion ref_comp;
378 /* io-wq management, e.g. thread count */
380 bool iowq_limits_set;
382 struct callback_head poll_wq_task_work;
383 struct list_head defer_list;
384 unsigned sq_thread_idle;
385 /* protected by ->completion_lock */
386 unsigned evfd_last_cq_tail;
389 * If IORING_SETUP_NO_MMAP is used, then the below holds
390 * the gup'ed pages for the two rings, and the sqes.
392 unsigned short n_ring_pages;
393 unsigned short n_sqe_pages;
394 struct page **ring_pages;
395 struct page **sqe_pages;
399 /* ->uring_lock is taken, callbacks can use io_tw_lock to lock it */
404 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
405 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
406 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
407 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
408 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
409 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
410 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
412 /* first byte is taken by user flags, shift it to not overlap */
417 REQ_F_LINK_TIMEOUT_BIT,
418 REQ_F_NEED_CLEANUP_BIT,
420 REQ_F_BUFFER_SELECTED_BIT,
421 REQ_F_BUFFER_RING_BIT,
425 REQ_F_ARM_LTIMEOUT_BIT,
426 REQ_F_ASYNC_DATA_BIT,
427 REQ_F_SKIP_LINK_CQES_BIT,
428 REQ_F_SINGLE_POLL_BIT,
429 REQ_F_DOUBLE_POLL_BIT,
430 REQ_F_PARTIAL_IO_BIT,
431 REQ_F_APOLL_MULTISHOT_BIT,
432 REQ_F_CLEAR_POLLIN_BIT,
433 REQ_F_HASH_LOCKED_BIT,
434 /* keep async read/write and isreg together and in order */
435 REQ_F_SUPPORT_NOWAIT_BIT,
438 /* not a real bit, just to check we're not overflowing the space */
444 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
445 /* drain existing IO first */
446 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
448 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
449 /* doesn't sever on completion < 0 */
450 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
452 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
453 /* IOSQE_BUFFER_SELECT */
454 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
455 /* IOSQE_CQE_SKIP_SUCCESS */
456 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
458 /* fail rest of links */
459 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
460 /* on inflight list, should be cancelled and waited on exit reliably */
461 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
462 /* read/write uses file position */
463 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
464 /* must not punt to workers */
465 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
466 /* has or had linked timeout */
467 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
469 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
470 /* already went through poll handler */
471 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
472 /* buffer already selected */
473 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
474 /* buffer selected from ring, needs commit */
475 REQ_F_BUFFER_RING = BIT(REQ_F_BUFFER_RING_BIT),
476 /* caller should reissue async */
477 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
478 /* supports async reads/writes */
479 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
481 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
482 /* has creds assigned */
483 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
484 /* skip refcounting if not set */
485 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
486 /* there is a linked timeout that has to be armed */
487 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
488 /* ->async_data allocated */
489 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
490 /* don't post CQEs while failing linked requests */
491 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
492 /* single poll may be active */
493 REQ_F_SINGLE_POLL = BIT(REQ_F_SINGLE_POLL_BIT),
494 /* double poll may active */
495 REQ_F_DOUBLE_POLL = BIT(REQ_F_DOUBLE_POLL_BIT),
496 /* request has already done partial IO */
497 REQ_F_PARTIAL_IO = BIT(REQ_F_PARTIAL_IO_BIT),
498 /* fast poll multishot mode */
499 REQ_F_APOLL_MULTISHOT = BIT(REQ_F_APOLL_MULTISHOT_BIT),
500 /* recvmsg special flag, clear EPOLLIN */
501 REQ_F_CLEAR_POLLIN = BIT(REQ_F_CLEAR_POLLIN_BIT),
502 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
503 REQ_F_HASH_LOCKED = BIT(REQ_F_HASH_LOCKED_BIT),
506 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, struct io_tw_state *ts);
508 struct io_task_work {
509 struct llist_node node;
510 io_req_tw_func_t func;
516 /* fd initially, then cflags for completion */
524 * Each request type overlays its private data structure on top of this one.
525 * They must not exceed this one in size.
529 /* each command gets 56 bytes of data */
533 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
535 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
537 #define io_kiocb_to_cmd(req, cmd_type) ( \
538 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
539 ((cmd_type *)&(req)->cmd) \
541 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
546 * NOTE! Each of the io_kiocb union members has the file pointer
547 * as the first entry in their struct definition. So you can
548 * access the file pointer through any of the sub-structs,
549 * or directly as just 'file' in this struct.
552 struct io_cmd_data cmd;
556 /* polled IO has completed */
559 * Can be either a fixed buffer index, or used with provided buffers.
560 * For the latter, before issue it points to the buffer group ID,
561 * and after selection it points to the buffer ID itself.
568 struct io_ring_ctx *ctx;
569 struct task_struct *task;
571 struct io_rsrc_node *rsrc_node;
574 /* store used ubuf, so we can prevent reloading */
575 struct io_mapped_ubuf *imu;
577 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
578 struct io_buffer *kbuf;
581 * stores buffer ID for ring provided buffers, valid IFF
582 * REQ_F_BUFFER_RING is set.
584 struct io_buffer_list *buf_list;
588 /* used by request caches, completion batching and iopoll */
589 struct io_wq_work_node comp_list;
590 /* cache ->apoll->events */
591 __poll_t apoll_events;
595 struct io_task_work io_task_work;
597 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
598 struct hlist_node hash_node;
599 /* internal polling, see IORING_FEAT_FAST_POLL */
600 struct async_poll *apoll;
601 /* opcode allocated if it needs to store data for async defer */
603 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
604 struct io_kiocb *link;
605 /* custom credentials, valid IFF REQ_F_CREDS is set */
606 const struct cred *creds;
607 struct io_wq_work work;
615 struct io_overflow_cqe {
616 struct list_head list;
617 struct io_uring_cqe cqe;