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;
72 u32 head ____cacheline_aligned_in_smp;
73 u32 tail ____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;
179 struct io_uring_cqe cqes[16];
183 struct eventfd_ctx *cq_ev_fd;
184 unsigned int eventfd_async: 1;
190 struct io_alloc_cache {
191 struct hlist_head list;
192 unsigned int nr_cached;
196 /* const or read-mostly hot data */
198 struct percpu_ref refs;
200 struct io_rings *rings;
202 enum task_work_notify_mode notify_method;
203 unsigned int compat: 1;
204 unsigned int drain_next: 1;
205 unsigned int restricted: 1;
206 unsigned int off_timeout_used: 1;
207 unsigned int drain_active: 1;
208 unsigned int drain_disabled: 1;
209 unsigned int has_evfd: 1;
210 unsigned int syscall_iopoll: 1;
211 } ____cacheline_aligned_in_smp;
213 /* submission data */
215 struct mutex uring_lock;
218 * Ring buffer of indices into array of io_uring_sqe, which is
219 * mmapped by the application using the IORING_OFF_SQES offset.
221 * This indirection could e.g. be used to assign fixed
222 * io_uring_sqe entries to operations and only submit them to
223 * the queue when needed.
225 * The kernel modifies neither the indices array nor the entries
229 struct io_uring_sqe *sq_sqes;
230 unsigned cached_sq_head;
234 * Fixed resources fast path, should be accessed only under
235 * uring_lock, and updated through io_uring_register(2)
237 struct io_rsrc_node *rsrc_node;
238 int rsrc_cached_refs;
240 struct io_file_table file_table;
241 unsigned nr_user_files;
242 unsigned nr_user_bufs;
243 struct io_mapped_ubuf **user_bufs;
245 struct io_submit_state submit_state;
247 struct io_buffer_list *io_bl;
248 struct xarray io_bl_xa;
249 struct list_head io_buffers_cache;
251 struct io_hash_table cancel_table_locked;
252 struct list_head cq_overflow_list;
253 struct io_alloc_cache apoll_cache;
254 struct io_alloc_cache netmsg_cache;
255 } ____cacheline_aligned_in_smp;
257 /* IRQ completion list, under ->completion_lock */
258 struct io_wq_work_list locked_free_list;
259 unsigned int locked_free_nr;
261 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
262 struct io_sq_data *sq_data; /* if using sq thread polling */
264 struct wait_queue_head sqo_sq_wait;
265 struct list_head sqd_list;
267 unsigned long check_cq;
269 unsigned int file_alloc_start;
270 unsigned int file_alloc_end;
272 struct xarray personalities;
277 * We cache a range of free CQEs we can use, once exhausted it
278 * should go through a slower range setup, see __io_get_cqe()
280 struct io_uring_cqe *cqe_cached;
281 struct io_uring_cqe *cqe_sentinel;
283 unsigned cached_cq_tail;
285 struct io_ev_fd __rcu *io_ev_fd;
286 struct wait_queue_head cq_wait;
288 } ____cacheline_aligned_in_smp;
291 spinlock_t completion_lock;
294 * ->iopoll_list is protected by the ctx->uring_lock for
295 * io_uring instances that don't use IORING_SETUP_SQPOLL.
296 * For SQPOLL, only the single threaded io_sq_thread() will
297 * manipulate the list, hence no extra locking is needed there.
299 struct io_wq_work_list iopoll_list;
300 struct io_hash_table cancel_table;
301 bool poll_multi_queue;
303 struct llist_head work_llist;
305 struct list_head io_buffers_comp;
306 } ____cacheline_aligned_in_smp;
310 spinlock_t timeout_lock;
311 atomic_t cq_timeouts;
312 struct list_head timeout_list;
313 struct list_head ltimeout_list;
314 unsigned cq_last_tm_flush;
315 } ____cacheline_aligned_in_smp;
317 /* Keep this last, we don't need it for the fast path */
319 struct io_restriction restrictions;
320 struct task_struct *submitter_task;
322 /* slow path rsrc auxilary data, used by update/register */
323 struct io_rsrc_node *rsrc_backup_node;
324 struct io_mapped_ubuf *dummy_ubuf;
325 struct io_rsrc_data *file_data;
326 struct io_rsrc_data *buf_data;
328 struct delayed_work rsrc_put_work;
329 struct llist_head rsrc_put_llist;
330 struct list_head rsrc_ref_list;
331 spinlock_t rsrc_ref_lock;
333 struct list_head io_buffers_pages;
335 #if defined(CONFIG_UNIX)
336 struct socket *ring_sock;
338 /* hashed buffered write serialization */
339 struct io_wq_hash *hash_map;
341 /* Only used for accounting purposes */
342 struct user_struct *user;
343 struct mm_struct *mm_account;
345 /* ctx exit and cancelation */
346 struct llist_head fallback_llist;
347 struct delayed_work fallback_work;
348 struct work_struct exit_work;
349 struct list_head tctx_list;
350 struct completion ref_comp;
352 /* io-wq management, e.g. thread count */
354 bool iowq_limits_set;
356 struct list_head defer_list;
357 unsigned sq_thread_idle;
358 /* protected by ->completion_lock */
359 unsigned evfd_last_cq_tail;
363 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
364 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
365 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
366 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
367 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
368 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
369 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
371 /* first byte is taken by user flags, shift it to not overlap */
376 REQ_F_LINK_TIMEOUT_BIT,
377 REQ_F_NEED_CLEANUP_BIT,
379 REQ_F_BUFFER_SELECTED_BIT,
380 REQ_F_BUFFER_RING_BIT,
384 REQ_F_ARM_LTIMEOUT_BIT,
385 REQ_F_ASYNC_DATA_BIT,
386 REQ_F_SKIP_LINK_CQES_BIT,
387 REQ_F_SINGLE_POLL_BIT,
388 REQ_F_DOUBLE_POLL_BIT,
389 REQ_F_PARTIAL_IO_BIT,
390 REQ_F_CQE32_INIT_BIT,
391 REQ_F_APOLL_MULTISHOT_BIT,
392 REQ_F_CLEAR_POLLIN_BIT,
393 REQ_F_HASH_LOCKED_BIT,
394 /* keep async read/write and isreg together and in order */
395 REQ_F_SUPPORT_NOWAIT_BIT,
398 /* not a real bit, just to check we're not overflowing the space */
404 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
405 /* drain existing IO first */
406 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
408 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
409 /* doesn't sever on completion < 0 */
410 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
412 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
413 /* IOSQE_BUFFER_SELECT */
414 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
415 /* IOSQE_CQE_SKIP_SUCCESS */
416 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
418 /* fail rest of links */
419 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
420 /* on inflight list, should be cancelled and waited on exit reliably */
421 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
422 /* read/write uses file position */
423 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
424 /* must not punt to workers */
425 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
426 /* has or had linked timeout */
427 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
429 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
430 /* already went through poll handler */
431 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
432 /* buffer already selected */
433 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
434 /* buffer selected from ring, needs commit */
435 REQ_F_BUFFER_RING = BIT(REQ_F_BUFFER_RING_BIT),
436 /* caller should reissue async */
437 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
438 /* supports async reads/writes */
439 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
441 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
442 /* has creds assigned */
443 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
444 /* skip refcounting if not set */
445 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
446 /* there is a linked timeout that has to be armed */
447 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
448 /* ->async_data allocated */
449 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
450 /* don't post CQEs while failing linked requests */
451 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
452 /* single poll may be active */
453 REQ_F_SINGLE_POLL = BIT(REQ_F_SINGLE_POLL_BIT),
454 /* double poll may active */
455 REQ_F_DOUBLE_POLL = BIT(REQ_F_DOUBLE_POLL_BIT),
456 /* request has already done partial IO */
457 REQ_F_PARTIAL_IO = BIT(REQ_F_PARTIAL_IO_BIT),
458 /* fast poll multishot mode */
459 REQ_F_APOLL_MULTISHOT = BIT(REQ_F_APOLL_MULTISHOT_BIT),
460 /* ->extra1 and ->extra2 are initialised */
461 REQ_F_CQE32_INIT = BIT(REQ_F_CQE32_INIT_BIT),
462 /* recvmsg special flag, clear EPOLLIN */
463 REQ_F_CLEAR_POLLIN = BIT(REQ_F_CLEAR_POLLIN_BIT),
464 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
465 REQ_F_HASH_LOCKED = BIT(REQ_F_HASH_LOCKED_BIT),
468 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
470 struct io_task_work {
471 struct llist_node node;
472 io_req_tw_func_t func;
478 /* fd initially, then cflags for completion */
486 * Each request type overlays its private data structure on top of this one.
487 * They must not exceed this one in size.
491 /* each command gets 56 bytes of data */
495 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
497 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
499 #define io_kiocb_to_cmd(req, cmd_type) ( \
500 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
501 ((cmd_type *)&(req)->cmd) \
503 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
508 * NOTE! Each of the io_kiocb union members has the file pointer
509 * as the first entry in their struct definition. So you can
510 * access the file pointer through any of the sub-structs,
511 * or directly as just 'file' in this struct.
514 struct io_cmd_data cmd;
518 /* polled IO has completed */
521 * Can be either a fixed buffer index, or used with provided buffers.
522 * For the latter, before issue it points to the buffer group ID,
523 * and after selection it points to the buffer ID itself.
530 struct io_ring_ctx *ctx;
531 struct task_struct *task;
533 struct io_rsrc_node *rsrc_node;
536 /* store used ubuf, so we can prevent reloading */
537 struct io_mapped_ubuf *imu;
539 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
540 struct io_buffer *kbuf;
543 * stores buffer ID for ring provided buffers, valid IFF
544 * REQ_F_BUFFER_RING is set.
546 struct io_buffer_list *buf_list;
550 /* used by request caches, completion batching and iopoll */
551 struct io_wq_work_node comp_list;
552 /* cache ->apoll->events */
553 __poll_t apoll_events;
557 struct io_task_work io_task_work;
558 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
560 struct hlist_node hash_node;
566 /* internal polling, see IORING_FEAT_FAST_POLL */
567 struct async_poll *apoll;
568 /* opcode allocated if it needs to store data for async defer */
570 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
571 struct io_kiocb *link;
572 /* custom credentials, valid IFF REQ_F_CREDS is set */
573 const struct cred *creds;
574 struct io_wq_work work;
577 struct io_overflow_cqe {
578 struct list_head list;
579 struct io_uring_cqe cqe;