1 #ifndef IOU_CORE_H 2 #define IOU_CORE_H 3 4 #include <linux/errno.h> 5 #include <linux/lockdep.h> 6 #include <linux/io_uring_types.h> 7 #include <uapi/linux/eventpoll.h> 8 #include "io-wq.h" 9 #include "slist.h" 10 #include "filetable.h" 11 12 #ifndef CREATE_TRACE_POINTS 13 #include <trace/events/io_uring.h> 14 #endif 15 16 enum { 17 IOU_OK = 0, 18 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED, 19 20 /* 21 * Intended only when both IO_URING_F_MULTISHOT is passed 22 * to indicate to the poll runner that multishot should be 23 * removed and the result is set on req->cqe.res. 24 */ 25 IOU_STOP_MULTISHOT = -ECANCELED, 26 }; 27 28 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow); 29 bool io_req_cqe_overflow(struct io_kiocb *req); 30 int io_run_task_work_sig(struct io_ring_ctx *ctx); 31 int __io_run_local_work(struct io_ring_ctx *ctx, bool *locked); 32 int io_run_local_work(struct io_ring_ctx *ctx); 33 void io_req_complete_failed(struct io_kiocb *req, s32 res); 34 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags); 35 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, 36 bool allow_overflow); 37 bool io_fill_cqe_aux(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags, 38 bool allow_overflow); 39 void __io_commit_cqring_flush(struct io_ring_ctx *ctx); 40 41 static inline void io_req_complete_post_tw(struct io_kiocb *req, bool *locked) 42 { 43 unsigned flags = *locked ? 0 : IO_URING_F_UNLOCKED; 44 45 io_req_complete_post(req, flags); 46 } 47 48 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages); 49 50 struct file *io_file_get_normal(struct io_kiocb *req, int fd); 51 struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 52 unsigned issue_flags); 53 54 static inline bool io_req_ffs_set(struct io_kiocb *req) 55 { 56 return req->flags & REQ_F_FIXED_FILE; 57 } 58 59 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local); 60 bool io_is_uring_fops(struct file *file); 61 bool io_alloc_async_data(struct io_kiocb *req); 62 void io_req_task_queue(struct io_kiocb *req); 63 void io_queue_iowq(struct io_kiocb *req, bool *dont_use); 64 void io_req_task_complete(struct io_kiocb *req, bool *locked); 65 void io_req_task_queue_fail(struct io_kiocb *req, int ret); 66 void io_req_task_submit(struct io_kiocb *req, bool *locked); 67 void tctx_task_work(struct callback_head *cb); 68 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); 69 int io_uring_alloc_task_context(struct task_struct *task, 70 struct io_ring_ctx *ctx); 71 72 int io_poll_issue(struct io_kiocb *req, bool *locked); 73 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr); 74 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin); 75 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node); 76 int io_req_prep_async(struct io_kiocb *req); 77 78 struct io_wq_work *io_wq_free_work(struct io_wq_work *work); 79 void io_wq_submit_work(struct io_wq_work *work); 80 81 void io_free_req(struct io_kiocb *req); 82 void io_queue_next(struct io_kiocb *req); 83 void __io_put_task(struct task_struct *task, int nr); 84 void io_task_refs_refill(struct io_uring_task *tctx); 85 bool __io_alloc_req_refill(struct io_ring_ctx *ctx); 86 87 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 88 bool cancel_all); 89 90 static inline void io_req_task_work_add(struct io_kiocb *req) 91 { 92 __io_req_task_work_add(req, true); 93 } 94 95 #define io_for_each_link(pos, head) \ 96 for (pos = (head); pos; pos = pos->link) 97 98 static inline void io_cq_lock(struct io_ring_ctx *ctx) 99 __acquires(ctx->completion_lock) 100 { 101 spin_lock(&ctx->completion_lock); 102 } 103 104 void io_cq_unlock_post(struct io_ring_ctx *ctx); 105 106 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx, 107 bool overflow) 108 { 109 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) { 110 struct io_uring_cqe *cqe = ctx->cqe_cached; 111 112 ctx->cached_cq_tail++; 113 ctx->cqe_cached++; 114 if (ctx->flags & IORING_SETUP_CQE32) 115 ctx->cqe_cached++; 116 return cqe; 117 } 118 119 return __io_get_cqe(ctx, overflow); 120 } 121 122 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) 123 { 124 return io_get_cqe_overflow(ctx, false); 125 } 126 127 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx, 128 struct io_kiocb *req) 129 { 130 struct io_uring_cqe *cqe; 131 132 /* 133 * If we can't get a cq entry, userspace overflowed the 134 * submission (by quite a lot). Increment the overflow count in 135 * the ring. 136 */ 137 cqe = io_get_cqe(ctx); 138 if (unlikely(!cqe)) 139 return io_req_cqe_overflow(req); 140 141 trace_io_uring_complete(req->ctx, req, req->cqe.user_data, 142 req->cqe.res, req->cqe.flags, 143 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0, 144 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0); 145 146 memcpy(cqe, &req->cqe, sizeof(*cqe)); 147 148 if (ctx->flags & IORING_SETUP_CQE32) { 149 u64 extra1 = 0, extra2 = 0; 150 151 if (req->flags & REQ_F_CQE32_INIT) { 152 extra1 = req->extra1; 153 extra2 = req->extra2; 154 } 155 156 WRITE_ONCE(cqe->big_cqe[0], extra1); 157 WRITE_ONCE(cqe->big_cqe[1], extra2); 158 } 159 return true; 160 } 161 162 static inline void req_set_fail(struct io_kiocb *req) 163 { 164 req->flags |= REQ_F_FAIL; 165 if (req->flags & REQ_F_CQE_SKIP) { 166 req->flags &= ~REQ_F_CQE_SKIP; 167 req->flags |= REQ_F_SKIP_LINK_CQES; 168 } 169 } 170 171 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags) 172 { 173 req->cqe.res = res; 174 req->cqe.flags = cflags; 175 } 176 177 static inline bool req_has_async_data(struct io_kiocb *req) 178 { 179 return req->flags & REQ_F_ASYNC_DATA; 180 } 181 182 static inline void io_put_file(struct file *file) 183 { 184 if (file) 185 fput(file); 186 } 187 188 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx, 189 unsigned issue_flags) 190 { 191 lockdep_assert_held(&ctx->uring_lock); 192 if (issue_flags & IO_URING_F_UNLOCKED) 193 mutex_unlock(&ctx->uring_lock); 194 } 195 196 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx, 197 unsigned issue_flags) 198 { 199 /* 200 * "Normal" inline submissions always hold the uring_lock, since we 201 * grab it from the system call. Same is true for the SQPOLL offload. 202 * The only exception is when we've detached the request and issue it 203 * from an async worker thread, grab the lock for that case. 204 */ 205 if (issue_flags & IO_URING_F_UNLOCKED) 206 mutex_lock(&ctx->uring_lock); 207 lockdep_assert_held(&ctx->uring_lock); 208 } 209 210 static inline void io_commit_cqring(struct io_ring_ctx *ctx) 211 { 212 /* order cqe stores with ring update */ 213 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); 214 } 215 216 /* requires smb_mb() prior, see wq_has_sleeper() */ 217 static inline void __io_cqring_wake(struct io_ring_ctx *ctx) 218 { 219 /* 220 * Trigger waitqueue handler on all waiters on our waitqueue. This 221 * won't necessarily wake up all the tasks, io_should_wake() will make 222 * that decision. 223 * 224 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter 225 * set in the mask so that if we recurse back into our own poll 226 * waitqueue handlers, we know we have a dependency between eventfd or 227 * epoll and should terminate multishot poll at that point. 228 */ 229 if (waitqueue_active(&ctx->cq_wait)) 230 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0, 231 poll_to_key(EPOLL_URING_WAKE | EPOLLIN)); 232 } 233 234 static inline void io_cqring_wake(struct io_ring_ctx *ctx) 235 { 236 smp_mb(); 237 __io_cqring_wake(ctx); 238 } 239 240 static inline bool io_sqring_full(struct io_ring_ctx *ctx) 241 { 242 struct io_rings *r = ctx->rings; 243 244 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; 245 } 246 247 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) 248 { 249 struct io_rings *rings = ctx->rings; 250 251 /* make sure SQ entry isn't read before tail */ 252 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; 253 } 254 255 static inline int io_run_task_work(void) 256 { 257 if (task_work_pending(current)) { 258 if (test_thread_flag(TIF_NOTIFY_SIGNAL)) 259 clear_notify_signal(); 260 __set_current_state(TASK_RUNNING); 261 task_work_run(); 262 return 1; 263 } 264 265 return 0; 266 } 267 268 static inline bool io_task_work_pending(struct io_ring_ctx *ctx) 269 { 270 return test_thread_flag(TIF_NOTIFY_SIGNAL) || 271 !wq_list_empty(&ctx->work_llist); 272 } 273 274 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx) 275 { 276 int ret = 0; 277 int ret2; 278 279 if (ctx->flags & IORING_SETUP_DEFER_TASKRUN) 280 ret = io_run_local_work(ctx); 281 282 /* want to run this after in case more is added */ 283 ret2 = io_run_task_work(); 284 285 /* Try propagate error in favour of if tasks were run, 286 * but still make sure to run them if requested 287 */ 288 if (ret >= 0) 289 ret += ret2; 290 291 return ret; 292 } 293 294 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx) 295 { 296 bool locked; 297 int ret; 298 299 if (llist_empty(&ctx->work_llist)) 300 return 0; 301 302 locked = true; 303 ret = __io_run_local_work(ctx, &locked); 304 /* shouldn't happen! */ 305 if (WARN_ON_ONCE(!locked)) 306 mutex_lock(&ctx->uring_lock); 307 return ret; 308 } 309 310 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked) 311 { 312 if (!*locked) { 313 mutex_lock(&ctx->uring_lock); 314 *locked = true; 315 } 316 } 317 318 /* 319 * Don't complete immediately but use deferred completion infrastructure. 320 * Protected by ->uring_lock and can only be used either with 321 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex. 322 */ 323 static inline void io_req_complete_defer(struct io_kiocb *req) 324 __must_hold(&req->ctx->uring_lock) 325 { 326 struct io_submit_state *state = &req->ctx->submit_state; 327 328 lockdep_assert_held(&req->ctx->uring_lock); 329 330 wq_list_add_tail(&req->comp_list, &state->compl_reqs); 331 } 332 333 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx) 334 { 335 if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd)) 336 __io_commit_cqring_flush(ctx); 337 } 338 339 /* must to be called somewhat shortly after putting a request */ 340 static inline void io_put_task(struct task_struct *task, int nr) 341 { 342 if (likely(task == current)) 343 task->io_uring->cached_refs += nr; 344 else 345 __io_put_task(task, nr); 346 } 347 348 static inline void io_get_task_refs(int nr) 349 { 350 struct io_uring_task *tctx = current->io_uring; 351 352 tctx->cached_refs -= nr; 353 if (unlikely(tctx->cached_refs < 0)) 354 io_task_refs_refill(tctx); 355 } 356 357 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx) 358 { 359 return !ctx->submit_state.free_list.next; 360 } 361 362 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx) 363 { 364 if (unlikely(io_req_cache_empty(ctx))) 365 return __io_alloc_req_refill(ctx); 366 return true; 367 } 368 369 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx) 370 { 371 struct io_wq_work_node *node; 372 373 node = wq_stack_extract(&ctx->submit_state.free_list); 374 return container_of(node, struct io_kiocb, comp_list); 375 } 376 377 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx) 378 { 379 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) || 380 ctx->submitter_task == current); 381 } 382 383 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res) 384 { 385 io_req_set_res(req, res, 0); 386 req->io_task_work.func = io_req_task_complete; 387 io_req_task_work_add(req); 388 } 389 390 #endif 391