1 #ifndef IO_URING_TYPES_H
2 #define IO_URING_TYPES_H
3
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>
9
10 struct io_wq_work_node {
11 struct io_wq_work_node *next;
12 };
13
14 struct io_wq_work_list {
15 struct io_wq_work_node *first;
16 struct io_wq_work_node *last;
17 };
18
19 struct io_wq_work {
20 struct io_wq_work_node list;
21 unsigned flags;
22 /* place it here instead of io_kiocb as it fills padding and saves 4B */
23 int cancel_seq;
24 };
25
26 struct io_fixed_file {
27 /* file * with additional FFS_* flags */
28 unsigned long file_ptr;
29 };
30
31 struct io_file_table {
32 struct io_fixed_file *files;
33 unsigned long *bitmap;
34 unsigned int alloc_hint;
35 };
36
37 struct io_hash_bucket {
38 spinlock_t lock;
39 struct hlist_head list;
40 } ____cacheline_aligned_in_smp;
41
42 struct io_hash_table {
43 struct io_hash_bucket *hbs;
44 unsigned hash_bits;
45 };
46
47 /*
48 * Arbitrary limit, can be raised if need be
49 */
50 #define IO_RINGFD_REG_MAX 16
51
52 struct io_uring_task {
53 /* submission side */
54 int cached_refs;
55 const struct io_ring_ctx *last;
56 struct io_wq *io_wq;
57 struct file *registered_rings[IO_RINGFD_REG_MAX];
58
59 struct xarray xa;
60 struct wait_queue_head wait;
61 atomic_t in_cancel;
62 atomic_t inflight_tracked;
63 struct percpu_counter inflight;
64
65 struct { /* task_work */
66 struct llist_head task_list;
67 struct callback_head task_work;
68 } ____cacheline_aligned_in_smp;
69 };
70
71 struct io_uring {
72 u32 head ____cacheline_aligned_in_smp;
73 u32 tail ____cacheline_aligned_in_smp;
74 };
75
76 /*
77 * This data is shared with the application through the mmap at offsets
78 * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING.
79 *
80 * The offsets to the member fields are published through struct
81 * io_sqring_offsets when calling io_uring_setup.
82 */
83 struct io_rings {
84 /*
85 * Head and tail offsets into the ring; the offsets need to be
86 * masked to get valid indices.
87 *
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
90 * cq ring.
91 */
92 struct io_uring sq, cq;
93 /*
94 * Bitmasks to apply to head and tail offsets (constant, equals
95 * ring_entries - 1)
96 */
97 u32 sq_ring_mask, cq_ring_mask;
98 /* Ring sizes (constant, power of 2) */
99 u32 sq_ring_entries, cq_ring_entries;
100 /*
101 * Number of invalid entries dropped by the kernel due to
102 * invalid index stored in array
103 *
104 * Written by the kernel, shouldn't be modified by the
105 * application (i.e. get number of "new events" by comparing to
106 * cached value).
107 *
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).
111 */
112 u32 sq_dropped;
113 /*
114 * Runtime SQ flags
115 *
116 * Written by the kernel, shouldn't be modified by the
117 * application.
118 *
119 * The application needs a full memory barrier before checking
120 * for IORING_SQ_NEED_WAKEUP after updating the sq tail.
121 */
122 atomic_t sq_flags;
123 /*
124 * Runtime CQ flags
125 *
126 * Written by the application, shouldn't be modified by the
127 * kernel.
128 */
129 u32 cq_flags;
130 /*
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.
135 *
136 * Written by the kernel, shouldn't be modified by the
137 * application (i.e. get number of "new events" by comparing to
138 * cached value).
139 *
140 * As completion events come in out of order this counter is not
141 * ordered with any other data.
142 */
143 u32 cq_overflow;
144 /*
145 * Ring buffer of completion events.
146 *
147 * The kernel writes completion events fresh every time they are
148 * produced, so the application is allowed to modify pending
149 * entries.
150 */
151 struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp;
152 };
153
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;
159 bool registered;
160 };
161
162 struct io_submit_link {
163 struct io_kiocb *head;
164 struct io_kiocb *last;
165 };
166
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;
173
174 bool plug_started;
175 bool need_plug;
176 unsigned short submit_nr;
177 unsigned int cqes_count;
178 struct blk_plug plug;
179 struct io_uring_cqe cqes[16];
180 };
181
182 struct io_ev_fd {
183 struct eventfd_ctx *cq_ev_fd;
184 unsigned int eventfd_async: 1;
185 struct rcu_head rcu;
186 atomic_t refs;
187 atomic_t ops;
188 };
189
190 struct io_alloc_cache {
191 struct hlist_head list;
192 unsigned int nr_cached;
193 };
194
195 struct io_ring_ctx {
196 /* const or read-mostly hot data */
197 struct {
198 unsigned int flags;
199 unsigned int drain_next: 1;
200 unsigned int restricted: 1;
201 unsigned int off_timeout_used: 1;
202 unsigned int drain_active: 1;
203 unsigned int has_evfd: 1;
204 /* all CQEs should be posted only by the submitter task */
205 unsigned int task_complete: 1;
206 unsigned int syscall_iopoll: 1;
207 unsigned int poll_activated: 1;
208 unsigned int drain_disabled: 1;
209 unsigned int compat: 1;
210
211 enum task_work_notify_mode notify_method;
212 struct io_rings *rings;
213 struct task_struct *submitter_task;
214 struct percpu_ref refs;
215 } ____cacheline_aligned_in_smp;
216
217 /* submission data */
218 struct {
219 struct mutex uring_lock;
220
221 /*
222 * Ring buffer of indices into array of io_uring_sqe, which is
223 * mmapped by the application using the IORING_OFF_SQES offset.
224 *
225 * This indirection could e.g. be used to assign fixed
226 * io_uring_sqe entries to operations and only submit them to
227 * the queue when needed.
228 *
229 * The kernel modifies neither the indices array nor the entries
230 * array.
231 */
232 u32 *sq_array;
233 struct io_uring_sqe *sq_sqes;
234 unsigned cached_sq_head;
235 unsigned sq_entries;
236
237 /*
238 * Fixed resources fast path, should be accessed only under
239 * uring_lock, and updated through io_uring_register(2)
240 */
241 struct io_rsrc_node *rsrc_node;
242 int rsrc_cached_refs;
243 atomic_t cancel_seq;
244 struct io_file_table file_table;
245 unsigned nr_user_files;
246 unsigned nr_user_bufs;
247 struct io_mapped_ubuf **user_bufs;
248
249 struct io_submit_state submit_state;
250
251 struct io_buffer_list *io_bl;
252 struct xarray io_bl_xa;
253 struct list_head io_buffers_cache;
254
255 struct io_hash_table cancel_table_locked;
256 struct list_head cq_overflow_list;
257 struct io_alloc_cache apoll_cache;
258 struct io_alloc_cache netmsg_cache;
259 } ____cacheline_aligned_in_smp;
260
261 /* IRQ completion list, under ->completion_lock */
262 struct io_wq_work_list locked_free_list;
263 unsigned int locked_free_nr;
264
265 const struct cred *sq_creds; /* cred used for __io_sq_thread() */
266 struct io_sq_data *sq_data; /* if using sq thread polling */
267
268 struct wait_queue_head sqo_sq_wait;
269 struct list_head sqd_list;
270
271 unsigned long check_cq;
272
273 unsigned int file_alloc_start;
274 unsigned int file_alloc_end;
275
276 struct xarray personalities;
277 u32 pers_next;
278
279 struct {
280 /*
281 * We cache a range of free CQEs we can use, once exhausted it
282 * should go through a slower range setup, see __io_get_cqe()
283 */
284 struct io_uring_cqe *cqe_cached;
285 struct io_uring_cqe *cqe_sentinel;
286
287 unsigned cached_cq_tail;
288 unsigned cq_entries;
289 struct io_ev_fd __rcu *io_ev_fd;
290 struct wait_queue_head cq_wait;
291 unsigned cq_extra;
292 } ____cacheline_aligned_in_smp;
293
294 struct {
295 spinlock_t completion_lock;
296
297 bool poll_multi_queue;
298 bool cq_waiting;
299
300 /*
301 * ->iopoll_list is protected by the ctx->uring_lock for
302 * io_uring instances that don't use IORING_SETUP_SQPOLL.
303 * For SQPOLL, only the single threaded io_sq_thread() will
304 * manipulate the list, hence no extra locking is needed there.
305 */
306 struct io_wq_work_list iopoll_list;
307 struct io_hash_table cancel_table;
308
309 struct llist_head work_llist;
310
311 struct list_head io_buffers_comp;
312 } ____cacheline_aligned_in_smp;
313
314 /* timeouts */
315 struct {
316 spinlock_t timeout_lock;
317 atomic_t cq_timeouts;
318 struct list_head timeout_list;
319 struct list_head ltimeout_list;
320 unsigned cq_last_tm_flush;
321 } ____cacheline_aligned_in_smp;
322
323 /* Keep this last, we don't need it for the fast path */
324 struct wait_queue_head poll_wq;
325 struct io_restriction restrictions;
326
327 /* slow path rsrc auxilary data, used by update/register */
328 struct io_rsrc_node *rsrc_backup_node;
329 struct io_mapped_ubuf *dummy_ubuf;
330 struct io_rsrc_data *file_data;
331 struct io_rsrc_data *buf_data;
332
333 struct delayed_work rsrc_put_work;
334 struct callback_head rsrc_put_tw;
335 struct llist_head rsrc_put_llist;
336 struct list_head rsrc_ref_list;
337 spinlock_t rsrc_ref_lock;
338
339 struct list_head io_buffers_pages;
340
341 #if defined(CONFIG_UNIX)
342 struct socket *ring_sock;
343 #endif
344 /* hashed buffered write serialization */
345 struct io_wq_hash *hash_map;
346
347 /* Only used for accounting purposes */
348 struct user_struct *user;
349 struct mm_struct *mm_account;
350
351 /* ctx exit and cancelation */
352 struct llist_head fallback_llist;
353 struct delayed_work fallback_work;
354 struct work_struct exit_work;
355 struct list_head tctx_list;
356 struct completion ref_comp;
357
358 /* io-wq management, e.g. thread count */
359 u32 iowq_limits[2];
360 bool iowq_limits_set;
361
362 struct callback_head poll_wq_task_work;
363 struct list_head defer_list;
364 unsigned sq_thread_idle;
365 /* protected by ->completion_lock */
366 unsigned evfd_last_cq_tail;
367 };
368
369 enum {
370 REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT,
371 REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT,
372 REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT,
373 REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT,
374 REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT,
375 REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT,
376 REQ_F_CQE_SKIP_BIT = IOSQE_CQE_SKIP_SUCCESS_BIT,
377
378 /* first byte is taken by user flags, shift it to not overlap */
379 REQ_F_FAIL_BIT = 8,
380 REQ_F_INFLIGHT_BIT,
381 REQ_F_CUR_POS_BIT,
382 REQ_F_NOWAIT_BIT,
383 REQ_F_LINK_TIMEOUT_BIT,
384 REQ_F_NEED_CLEANUP_BIT,
385 REQ_F_POLLED_BIT,
386 REQ_F_BUFFER_SELECTED_BIT,
387 REQ_F_BUFFER_RING_BIT,
388 REQ_F_REISSUE_BIT,
389 REQ_F_CREDS_BIT,
390 REQ_F_REFCOUNT_BIT,
391 REQ_F_ARM_LTIMEOUT_BIT,
392 REQ_F_ASYNC_DATA_BIT,
393 REQ_F_SKIP_LINK_CQES_BIT,
394 REQ_F_SINGLE_POLL_BIT,
395 REQ_F_DOUBLE_POLL_BIT,
396 REQ_F_PARTIAL_IO_BIT,
397 REQ_F_CQE32_INIT_BIT,
398 REQ_F_APOLL_MULTISHOT_BIT,
399 REQ_F_CLEAR_POLLIN_BIT,
400 REQ_F_HASH_LOCKED_BIT,
401 /* keep async read/write and isreg together and in order */
402 REQ_F_SUPPORT_NOWAIT_BIT,
403 REQ_F_ISREG_BIT,
404
405 /* not a real bit, just to check we're not overflowing the space */
406 __REQ_F_LAST_BIT,
407 };
408
409 enum {
410 /* ctx owns file */
411 REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT),
412 /* drain existing IO first */
413 REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT),
414 /* linked sqes */
415 REQ_F_LINK = BIT(REQ_F_LINK_BIT),
416 /* doesn't sever on completion < 0 */
417 REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT),
418 /* IOSQE_ASYNC */
419 REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT),
420 /* IOSQE_BUFFER_SELECT */
421 REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT),
422 /* IOSQE_CQE_SKIP_SUCCESS */
423 REQ_F_CQE_SKIP = BIT(REQ_F_CQE_SKIP_BIT),
424
425 /* fail rest of links */
426 REQ_F_FAIL = BIT(REQ_F_FAIL_BIT),
427 /* on inflight list, should be cancelled and waited on exit reliably */
428 REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT),
429 /* read/write uses file position */
430 REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT),
431 /* must not punt to workers */
432 REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT),
433 /* has or had linked timeout */
434 REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT),
435 /* needs cleanup */
436 REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT),
437 /* already went through poll handler */
438 REQ_F_POLLED = BIT(REQ_F_POLLED_BIT),
439 /* buffer already selected */
440 REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT),
441 /* buffer selected from ring, needs commit */
442 REQ_F_BUFFER_RING = BIT(REQ_F_BUFFER_RING_BIT),
443 /* caller should reissue async */
444 REQ_F_REISSUE = BIT(REQ_F_REISSUE_BIT),
445 /* supports async reads/writes */
446 REQ_F_SUPPORT_NOWAIT = BIT(REQ_F_SUPPORT_NOWAIT_BIT),
447 /* regular file */
448 REQ_F_ISREG = BIT(REQ_F_ISREG_BIT),
449 /* has creds assigned */
450 REQ_F_CREDS = BIT(REQ_F_CREDS_BIT),
451 /* skip refcounting if not set */
452 REQ_F_REFCOUNT = BIT(REQ_F_REFCOUNT_BIT),
453 /* there is a linked timeout that has to be armed */
454 REQ_F_ARM_LTIMEOUT = BIT(REQ_F_ARM_LTIMEOUT_BIT),
455 /* ->async_data allocated */
456 REQ_F_ASYNC_DATA = BIT(REQ_F_ASYNC_DATA_BIT),
457 /* don't post CQEs while failing linked requests */
458 REQ_F_SKIP_LINK_CQES = BIT(REQ_F_SKIP_LINK_CQES_BIT),
459 /* single poll may be active */
460 REQ_F_SINGLE_POLL = BIT(REQ_F_SINGLE_POLL_BIT),
461 /* double poll may active */
462 REQ_F_DOUBLE_POLL = BIT(REQ_F_DOUBLE_POLL_BIT),
463 /* request has already done partial IO */
464 REQ_F_PARTIAL_IO = BIT(REQ_F_PARTIAL_IO_BIT),
465 /* fast poll multishot mode */
466 REQ_F_APOLL_MULTISHOT = BIT(REQ_F_APOLL_MULTISHOT_BIT),
467 /* ->extra1 and ->extra2 are initialised */
468 REQ_F_CQE32_INIT = BIT(REQ_F_CQE32_INIT_BIT),
469 /* recvmsg special flag, clear EPOLLIN */
470 REQ_F_CLEAR_POLLIN = BIT(REQ_F_CLEAR_POLLIN_BIT),
471 /* hashed into ->cancel_hash_locked, protected by ->uring_lock */
472 REQ_F_HASH_LOCKED = BIT(REQ_F_HASH_LOCKED_BIT),
473 };
474
475 typedef void (*io_req_tw_func_t)(struct io_kiocb *req, bool *locked);
476
477 struct io_task_work {
478 struct llist_node node;
479 io_req_tw_func_t func;
480 };
481
482 struct io_cqe {
483 __u64 user_data;
484 __s32 res;
485 /* fd initially, then cflags for completion */
486 union {
487 __u32 flags;
488 int fd;
489 };
490 };
491
492 /*
493 * Each request type overlays its private data structure on top of this one.
494 * They must not exceed this one in size.
495 */
496 struct io_cmd_data {
497 struct file *file;
498 /* each command gets 56 bytes of data */
499 __u8 data[56];
500 };
501
io_kiocb_cmd_sz_check(size_t cmd_sz)502 static inline void io_kiocb_cmd_sz_check(size_t cmd_sz)
503 {
504 BUILD_BUG_ON(cmd_sz > sizeof(struct io_cmd_data));
505 }
506 #define io_kiocb_to_cmd(req, cmd_type) ( \
507 io_kiocb_cmd_sz_check(sizeof(cmd_type)) , \
508 ((cmd_type *)&(req)->cmd) \
509 )
510 #define cmd_to_io_kiocb(ptr) ((struct io_kiocb *) ptr)
511
512 struct io_kiocb {
513 union {
514 /*
515 * NOTE! Each of the io_kiocb union members has the file pointer
516 * as the first entry in their struct definition. So you can
517 * access the file pointer through any of the sub-structs,
518 * or directly as just 'file' in this struct.
519 */
520 struct file *file;
521 struct io_cmd_data cmd;
522 };
523
524 u8 opcode;
525 /* polled IO has completed */
526 u8 iopoll_completed;
527 /*
528 * Can be either a fixed buffer index, or used with provided buffers.
529 * For the latter, before issue it points to the buffer group ID,
530 * and after selection it points to the buffer ID itself.
531 */
532 u16 buf_index;
533 unsigned int flags;
534
535 struct io_cqe cqe;
536
537 struct io_ring_ctx *ctx;
538 struct task_struct *task;
539
540 struct io_rsrc_node *rsrc_node;
541
542 union {
543 /* store used ubuf, so we can prevent reloading */
544 struct io_mapped_ubuf *imu;
545
546 /* stores selected buf, valid IFF REQ_F_BUFFER_SELECTED is set */
547 struct io_buffer *kbuf;
548
549 /*
550 * stores buffer ID for ring provided buffers, valid IFF
551 * REQ_F_BUFFER_RING is set.
552 */
553 struct io_buffer_list *buf_list;
554 };
555
556 union {
557 /* used by request caches, completion batching and iopoll */
558 struct io_wq_work_node comp_list;
559 /* cache ->apoll->events */
560 __poll_t apoll_events;
561 };
562 atomic_t refs;
563 atomic_t poll_refs;
564 struct io_task_work io_task_work;
565 /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */
566 union {
567 struct hlist_node hash_node;
568 struct {
569 u64 extra1;
570 u64 extra2;
571 };
572 };
573 /* internal polling, see IORING_FEAT_FAST_POLL */
574 struct async_poll *apoll;
575 /* opcode allocated if it needs to store data for async defer */
576 void *async_data;
577 /* linked requests, IFF REQ_F_HARDLINK or REQ_F_LINK are set */
578 struct io_kiocb *link;
579 /* custom credentials, valid IFF REQ_F_CREDS is set */
580 const struct cred *creds;
581 struct io_wq_work work;
582 };
583
584 struct io_overflow_cqe {
585 struct list_head list;
586 struct io_uring_cqe cqe;
587 };
588
589 #endif
590