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;
40 struct io_hash_bucket {
42 struct hlist_head list;
43 } ____cacheline_aligned_in_smp;
45 struct io_hash_table {
46 struct io_hash_bucket *hbs;
51 * Arbitrary limit, can be raised if need be
53 #define IO_RINGFD_REG_MAX 16
55 struct io_uring_task {
58 const struct io_ring_ctx *last;
60 struct file *registered_rings[IO_RINGFD_REG_MAX];
63 struct wait_queue_head wait;
65 atomic_t inflight_tracked;
66 struct percpu_counter inflight;
68 struct { /* task_work */
69 struct llist_head task_list;
70 struct callback_head task_work;
71 } ____cacheline_aligned_in_smp;
75 u32 head ____cacheline_aligned_in_smp;
76 u32 tail ____cacheline_aligned_in_smp;
80 * This data is shared with the application through the mmap at offsets
81 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
83 * The offsets to the member fields are published through struct
84 * io_sqring_offsets when calling io_uring_setup.
88 * Head and tail offsets into the ring; the offsets need to be
89 * masked to get valid indices.
91 * The kernel controls head of the sq ring and the tail of the cq ring,
92 * and the application controls tail of the sq ring and the head of the
95 struct io_uring sq, cq;
97 * Bitmasks to apply to head and tail offsets (constant, equals
100 u32 sq_ring_mask, cq_ring_mask;
101 /* Ring sizes (constant, power of 2) */
102 u32 sq_ring_entries, cq_ring_entries;
104 * Number of invalid entries dropped by the kernel due to
105 * invalid index stored in array
107 * Written by the kernel, shouldn't be modified by the
108 * application (i.e. get number of "new events" by comparing to
111 * After a new SQ head value was read by the application this
112 * counter includes all submissions that were dropped reaching
113 * the new SQ head (and possibly more).
119 * Written by the kernel, shouldn't be modified by the
122 * The application needs a full memory barrier before checking
123 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
129 * Written by the application, shouldn't be modified by the
134 * Number of completion events lost because the queue was full;
135 * this should be avoided by the application by making sure
136 * there are not more requests pending than there is space in
137 * the completion queue.
139 * Written by the kernel, shouldn't be modified by the
140 * application (i.e. get number of "new events" by comparing to
143 * As completion events come in out of order this counter is not
144 * ordered with any other data.
148 * Ring buffer of completion events.
150 * The kernel writes completion events fresh every time they are
151 * produced, so the application is allowed to modify pending
154 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
157 struct io_restriction {
158 DECLARE_BITMAP(register_op, IORING_REGISTER_LAST);
159 DECLARE_BITMAP(sqe_op, IORING_OP_LAST);
160 u8 sqe_flags_allowed;
161 u8 sqe_flags_required;
165 struct io_submit_link {
166 struct io_kiocb *head;
167 struct io_kiocb *last;
170 struct io_submit_state {
171 /* inline/task_work completion list, under ->uring_lock */
172 struct io_wq_work_node free_list;
173 /* batch completion logic */
174 struct io_wq_work_list compl_reqs;
175 struct io_submit_link link;
179 unsigned short submit_nr;
180 struct blk_plug plug;
184 struct eventfd_ctx *cq_ev_fd;
185 unsigned int eventfd_async: 1;
191 struct io_alloc_cache {
192 struct hlist_head list;
193 unsigned int nr_cached;
197 /* const or read-mostly hot data */
199 struct percpu_ref refs;
201 struct io_rings *rings;
203 enum task_work_notify_mode notify_method;
204 unsigned int compat: 1;
205 unsigned int drain_next: 1;
206 unsigned int restricted: 1;
207 unsigned int off_timeout_used: 1;
208 unsigned int drain_active: 1;
209 unsigned int drain_disabled: 1;
210 unsigned int has_evfd: 1;
211 unsigned int syscall_iopoll: 1;
212 } ____cacheline_aligned_in_smp;
214 /* submission data */
216 struct mutex uring_lock;
219 * Ring buffer of indices into array of io_uring_sqe, which is
220 * mmapped by the application using the IORING_OFF_SQES offset.
222 * This indirection could e.g. be used to assign fixed
223 * io_uring_sqe entries to operations and only submit them to
224 * the queue when needed.
226 * The kernel modifies neither the indices array nor the entries
230 struct io_uring_sqe *sq_sqes;
231 unsigned cached_sq_head;
235 * Fixed resources fast path, should be accessed only under
236 * uring_lock, and updated through io_uring_register(2)
238 struct io_rsrc_node *rsrc_node;
239 int rsrc_cached_refs;
241 struct io_file_table file_table;
242 unsigned nr_user_files;
243 unsigned nr_user_bufs;
244 struct io_mapped_ubuf **user_bufs;
245 struct io_notif_slot *notif_slots;
246 unsigned nr_notif_slots;
248 struct io_submit_state submit_state;
250 struct io_buffer_list *io_bl;
251 struct xarray io_bl_xa;
252 struct list_head io_buffers_cache;
254 struct io_hash_table cancel_table_locked;
255 struct list_head cq_overflow_list;
256 struct io_alloc_cache apoll_cache;
257 struct io_alloc_cache netmsg_cache;
258 } ____cacheline_aligned_in_smp;
260 /* IRQ completion list, under ->completion_lock */
261 struct io_wq_work_list locked_free_list;
262 unsigned int locked_free_nr;
264 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
265 struct io_sq_data *sq_data; /* if using sq thread polling */
267 struct wait_queue_head sqo_sq_wait;
268 struct list_head sqd_list;
270 unsigned long check_cq;
272 unsigned int file_alloc_start;
273 unsigned int file_alloc_end;
275 struct xarray personalities;
280 * We cache a range of free CQEs we can use, once exhausted it
281 * should go through a slower range setup, see __io_get_cqe()
283 struct io_uring_cqe *cqe_cached;
284 struct io_uring_cqe *cqe_sentinel;
286 unsigned cached_cq_tail;
288 struct io_ev_fd __rcu *io_ev_fd;
289 struct wait_queue_head cq_wait;
291 } ____cacheline_aligned_in_smp;
294 spinlock_t completion_lock;
297 * ->iopoll_list is protected by the ctx->uring_lock for
298 * io_uring instances that don't use IORING_SETUP_SQPOLL.
299 * For SQPOLL, only the single threaded io_sq_thread() will
300 * manipulate the list, hence no extra locking is needed there.
302 struct io_wq_work_list iopoll_list;
303 struct io_hash_table cancel_table;
304 bool poll_multi_queue;
306 struct llist_head work_llist;
308 struct list_head io_buffers_comp;
309 } ____cacheline_aligned_in_smp;
313 spinlock_t timeout_lock;
314 atomic_t cq_timeouts;
315 struct list_head timeout_list;
316 struct list_head ltimeout_list;
317 unsigned cq_last_tm_flush;
318 } ____cacheline_aligned_in_smp;
320 /* Keep this last, we don't need it for the fast path */
322 struct io_restriction restrictions;
323 struct task_struct *submitter_task;
325 /* slow path rsrc auxilary data, used by update/register */
326 struct io_rsrc_node *rsrc_backup_node;
327 struct io_mapped_ubuf *dummy_ubuf;
328 struct io_rsrc_data *file_data;
329 struct io_rsrc_data *buf_data;
331 struct delayed_work rsrc_put_work;
332 struct llist_head rsrc_put_llist;
333 struct list_head rsrc_ref_list;
334 spinlock_t rsrc_ref_lock;
336 struct list_head io_buffers_pages;
338 #if defined(CONFIG_UNIX)
339 struct socket *ring_sock;
341 /* hashed buffered write serialization */
342 struct io_wq_hash *hash_map;
344 /* Only used for accounting purposes */
345 struct user_struct *user;
346 struct mm_struct *mm_account;
348 /* ctx exit and cancelation */
349 struct llist_head fallback_llist;
350 struct delayed_work fallback_work;
351 struct work_struct exit_work;
352 struct list_head tctx_list;
353 struct completion ref_comp;
355 /* io-wq management, e.g. thread count */
357 bool iowq_limits_set;
359 struct list_head defer_list;
360 unsigned sq_thread_idle;
361 /* protected by ->completion_lock */
362 unsigned evfd_last_cq_tail;
366 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
367 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
368 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
369 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
370 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
371 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
372 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
374 /* first byte is taken by user flags, shift it to not overlap */
379 REQ_F_LINK_TIMEOUT_BIT,
380 REQ_F_NEED_CLEANUP_BIT,
382 REQ_F_BUFFER_SELECTED_BIT,
383 REQ_F_BUFFER_RING_BIT,
387 REQ_F_ARM_LTIMEOUT_BIT,
388 REQ_F_ASYNC_DATA_BIT,
389 REQ_F_SKIP_LINK_CQES_BIT,
390 REQ_F_SINGLE_POLL_BIT,
391 REQ_F_DOUBLE_POLL_BIT,
392 REQ_F_PARTIAL_IO_BIT,
393 REQ_F_CQE32_INIT_BIT,
394 REQ_F_APOLL_MULTISHOT_BIT,
395 REQ_F_CLEAR_POLLIN_BIT,
396 REQ_F_HASH_LOCKED_BIT,
397 /* keep async read/write and isreg together and in order */
398 REQ_F_SUPPORT_NOWAIT_BIT,
401 /* not a real bit, just to check we're not overflowing the space */
407 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
408 /* drain existing IO first */
409 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
411 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
412 /* doesn't sever on completion < 0 */
413 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
415 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
416 /* IOSQE_BUFFER_SELECT */
417 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
418 /* IOSQE_CQE_SKIP_SUCCESS */
419 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
421 /* fail rest of links */
422 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
423 /* on inflight list, should be cancelled and waited on exit reliably */
424 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
425 /* read/write uses file position */
426 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
427 /* must not punt to workers */
428 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
429 /* has or had linked timeout */
430 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
432 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
433 /* already went through poll handler */
434 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
435 /* buffer already selected */
436 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
437 /* buffer selected from ring, needs commit */
438 REQ_F_BUFFER_RING = BIT(REQ_F_BUFFER_RING_BIT),
439 /* caller should reissue async */
440 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
441 /* supports async reads/writes */
442 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
444 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
445 /* has creds assigned */
446 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
447 /* skip refcounting if not set */
448 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
449 /* there is a linked timeout that has to be armed */
450 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
451 /* ->async_data allocated */
452 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
453 /* don't post CQEs while failing linked requests */
454 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
455 /* single poll may be active */
456 REQ_F_SINGLE_POLL = BIT(REQ_F_SINGLE_POLL_BIT),
457 /* double poll may active */
458 REQ_F_DOUBLE_POLL = BIT(REQ_F_DOUBLE_POLL_BIT),
459 /* request has already done partial IO */
460 REQ_F_PARTIAL_IO = BIT(REQ_F_PARTIAL_IO_BIT),
461 /* fast poll multishot mode */
462 REQ_F_APOLL_MULTISHOT = BIT(REQ_F_APOLL_MULTISHOT_BIT),
463 /* ->extra1 and ->extra2 are initialised */
464 REQ_F_CQE32_INIT = BIT(REQ_F_CQE32_INIT_BIT),
465 /* recvmsg special flag, clear EPOLLIN */
466 REQ_F_CLEAR_POLLIN = BIT(REQ_F_CLEAR_POLLIN_BIT),
467 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
468 REQ_F_HASH_LOCKED = BIT(REQ_F_HASH_LOCKED_BIT),
471 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
473 struct io_task_work {
474 struct llist_node node;
475 io_req_tw_func_t func;
481 /* fd initially, then cflags for completion */
489 * Each request type overlays its private data structure on top of this one.
490 * They must not exceed this one in size.
494 /* each command gets 56 bytes of data */
498 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
500 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
502 #define io_kiocb_to_cmd(req, cmd_type) ( \
503 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
504 ((cmd_type *)&(req)->cmd) \
506 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
511 * NOTE! Each of the io_kiocb union members has the file pointer
512 * as the first entry in their struct definition. So you can
513 * access the file pointer through any of the sub-structs,
514 * or directly as just 'file' in this struct.
517 struct io_cmd_data cmd;
521 /* polled IO has completed */
524 * Can be either a fixed buffer index, or used with provided buffers.
525 * For the latter, before issue it points to the buffer group ID,
526 * and after selection it points to the buffer ID itself.
533 struct io_ring_ctx *ctx;
534 struct task_struct *task;
536 struct io_rsrc_node *rsrc_node;
539 /* store used ubuf, so we can prevent reloading */
540 struct io_mapped_ubuf *imu;
542 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
543 struct io_buffer *kbuf;
546 * stores buffer ID for ring provided buffers, valid IFF
547 * REQ_F_BUFFER_RING is set.
549 struct io_buffer_list *buf_list;
553 /* used by request caches, completion batching and iopoll */
554 struct io_wq_work_node comp_list;
555 /* cache ->apoll->events */
556 __poll_t apoll_events;
560 struct io_task_work io_task_work;
561 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
563 struct hlist_node hash_node;
569 /* internal polling, see IORING_FEAT_FAST_POLL */
570 struct async_poll *apoll;
571 /* opcode allocated if it needs to store data for async defer */
573 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
574 struct io_kiocb *link;
575 /* custom credentials, valid IFF REQ_F_CREDS is set */
576 const struct cred *creds;
577 struct io_wq_work work;
580 struct io_overflow_cqe {
581 struct list_head list;
582 struct io_uring_cqe cqe;