1 #ifndef IOU_CORE_H 2 #define IOU_CORE_H 3 4 #include <linux/errno.h> 5 #include <linux/lockdep.h> 6 #include <linux/resume_user_mode.h> 7 #include <linux/kasan.h> 8 #include <linux/io_uring_types.h> 9 #include <uapi/linux/eventpoll.h> 10 #include "io-wq.h" 11 #include "slist.h" 12 #include "filetable.h" 13 14 #ifndef CREATE_TRACE_POINTS 15 #include <trace/events/io_uring.h> 16 #endif 17 18 enum { 19 IOU_OK = 0, 20 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED, 21 22 /* 23 * Intended only when both IO_URING_F_MULTISHOT is passed 24 * to indicate to the poll runner that multishot should be 25 * removed and the result is set on req->cqe.res. 26 */ 27 IOU_STOP_MULTISHOT = -ECANCELED, 28 }; 29 30 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow); 31 bool io_req_cqe_overflow(struct io_kiocb *req); 32 int io_run_task_work_sig(struct io_ring_ctx *ctx); 33 void io_req_defer_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 io_aux_cqe(struct io_ring_ctx *ctx, bool defer, u64 user_data, s32 res, u32 cflags, 37 bool allow_overflow); 38 void __io_commit_cqring_flush(struct io_ring_ctx *ctx); 39 40 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages); 41 42 struct file *io_file_get_normal(struct io_kiocb *req, int fd); 43 struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 44 unsigned issue_flags); 45 46 static inline bool io_req_ffs_set(struct io_kiocb *req) 47 { 48 return req->flags & REQ_F_FIXED_FILE; 49 } 50 51 void __io_req_task_work_add(struct io_kiocb *req, bool allow_local); 52 bool io_is_uring_fops(struct file *file); 53 bool io_alloc_async_data(struct io_kiocb *req); 54 void io_req_task_queue(struct io_kiocb *req); 55 void io_queue_iowq(struct io_kiocb *req, bool *dont_use); 56 void io_req_task_complete(struct io_kiocb *req, bool *locked); 57 void io_req_task_queue_fail(struct io_kiocb *req, int ret); 58 void io_req_task_submit(struct io_kiocb *req, bool *locked); 59 void tctx_task_work(struct callback_head *cb); 60 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); 61 int io_uring_alloc_task_context(struct task_struct *task, 62 struct io_ring_ctx *ctx); 63 64 int io_poll_issue(struct io_kiocb *req, bool *locked); 65 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr); 66 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin); 67 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node); 68 int io_req_prep_async(struct io_kiocb *req); 69 70 struct io_wq_work *io_wq_free_work(struct io_wq_work *work); 71 void io_wq_submit_work(struct io_wq_work *work); 72 73 void io_free_req(struct io_kiocb *req); 74 void io_queue_next(struct io_kiocb *req); 75 void io_task_refs_refill(struct io_uring_task *tctx); 76 bool __io_alloc_req_refill(struct io_ring_ctx *ctx); 77 78 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 79 bool cancel_all); 80 81 #define io_lockdep_assert_cq_locked(ctx) \ 82 do { \ 83 if (ctx->flags & IORING_SETUP_IOPOLL) { \ 84 lockdep_assert_held(&ctx->uring_lock); \ 85 } else if (!ctx->task_complete) { \ 86 lockdep_assert_held(&ctx->completion_lock); \ 87 } else if (ctx->submitter_task->flags & PF_EXITING) { \ 88 lockdep_assert(current_work()); \ 89 } else { \ 90 lockdep_assert(current == ctx->submitter_task); \ 91 } \ 92 } while (0) 93 94 static inline void io_req_task_work_add(struct io_kiocb *req) 95 { 96 __io_req_task_work_add(req, true); 97 } 98 99 #define io_for_each_link(pos, head) \ 100 for (pos = (head); pos; pos = pos->link) 101 102 void io_cq_unlock_post(struct io_ring_ctx *ctx); 103 104 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx, 105 bool overflow) 106 { 107 io_lockdep_assert_cq_locked(ctx); 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 false; 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 bool io_fill_cqe_req(struct io_ring_ctx *ctx, 163 struct io_kiocb *req) 164 { 165 if (likely(__io_fill_cqe_req(ctx, req))) 166 return true; 167 return io_req_cqe_overflow(req); 168 } 169 170 static inline void req_set_fail(struct io_kiocb *req) 171 { 172 req->flags |= REQ_F_FAIL; 173 if (req->flags & REQ_F_CQE_SKIP) { 174 req->flags &= ~REQ_F_CQE_SKIP; 175 req->flags |= REQ_F_SKIP_LINK_CQES; 176 } 177 } 178 179 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags) 180 { 181 req->cqe.res = res; 182 req->cqe.flags = cflags; 183 } 184 185 static inline bool req_has_async_data(struct io_kiocb *req) 186 { 187 return req->flags & REQ_F_ASYNC_DATA; 188 } 189 190 static inline void io_put_file(struct file *file) 191 { 192 if (file) 193 fput(file); 194 } 195 196 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx, 197 unsigned issue_flags) 198 { 199 lockdep_assert_held(&ctx->uring_lock); 200 if (issue_flags & IO_URING_F_UNLOCKED) 201 mutex_unlock(&ctx->uring_lock); 202 } 203 204 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx, 205 unsigned issue_flags) 206 { 207 /* 208 * "Normal" inline submissions always hold the uring_lock, since we 209 * grab it from the system call. Same is true for the SQPOLL offload. 210 * The only exception is when we've detached the request and issue it 211 * from an async worker thread, grab the lock for that case. 212 */ 213 if (issue_flags & IO_URING_F_UNLOCKED) 214 mutex_lock(&ctx->uring_lock); 215 lockdep_assert_held(&ctx->uring_lock); 216 } 217 218 static inline void io_commit_cqring(struct io_ring_ctx *ctx) 219 { 220 /* order cqe stores with ring update */ 221 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); 222 } 223 224 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx) 225 { 226 if (wq_has_sleeper(&ctx->poll_wq)) 227 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0, 228 poll_to_key(EPOLL_URING_WAKE | EPOLLIN)); 229 } 230 231 /* requires smb_mb() prior, see wq_has_sleeper() */ 232 static inline void __io_cqring_wake(struct io_ring_ctx *ctx) 233 { 234 /* 235 * Trigger waitqueue handler on all waiters on our waitqueue. This 236 * won't necessarily wake up all the tasks, io_should_wake() will make 237 * that decision. 238 * 239 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter 240 * set in the mask so that if we recurse back into our own poll 241 * waitqueue handlers, we know we have a dependency between eventfd or 242 * epoll and should terminate multishot poll at that point. 243 */ 244 if (waitqueue_active(&ctx->cq_wait)) 245 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0, 246 poll_to_key(EPOLL_URING_WAKE | EPOLLIN)); 247 } 248 249 static inline void io_cqring_wake(struct io_ring_ctx *ctx) 250 { 251 smp_mb(); 252 __io_cqring_wake(ctx); 253 } 254 255 static inline bool io_sqring_full(struct io_ring_ctx *ctx) 256 { 257 struct io_rings *r = ctx->rings; 258 259 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; 260 } 261 262 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) 263 { 264 struct io_rings *rings = ctx->rings; 265 266 /* make sure SQ entry isn't read before tail */ 267 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; 268 } 269 270 static inline int io_run_task_work(void) 271 { 272 /* 273 * Always check-and-clear the task_work notification signal. With how 274 * signaling works for task_work, we can find it set with nothing to 275 * run. We need to clear it for that case, like get_signal() does. 276 */ 277 if (test_thread_flag(TIF_NOTIFY_SIGNAL)) 278 clear_notify_signal(); 279 /* 280 * PF_IO_WORKER never returns to userspace, so check here if we have 281 * notify work that needs processing. 282 */ 283 if (current->flags & PF_IO_WORKER && 284 test_thread_flag(TIF_NOTIFY_RESUME)) { 285 __set_current_state(TASK_RUNNING); 286 resume_user_mode_work(NULL); 287 } 288 if (task_work_pending(current)) { 289 __set_current_state(TASK_RUNNING); 290 task_work_run(); 291 return 1; 292 } 293 294 return 0; 295 } 296 297 static inline bool io_task_work_pending(struct io_ring_ctx *ctx) 298 { 299 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist); 300 } 301 302 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked) 303 { 304 if (!*locked) { 305 mutex_lock(&ctx->uring_lock); 306 *locked = true; 307 } 308 } 309 310 /* 311 * Don't complete immediately but use deferred completion infrastructure. 312 * Protected by ->uring_lock and can only be used either with 313 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex. 314 */ 315 static inline void io_req_complete_defer(struct io_kiocb *req) 316 __must_hold(&req->ctx->uring_lock) 317 { 318 struct io_submit_state *state = &req->ctx->submit_state; 319 320 lockdep_assert_held(&req->ctx->uring_lock); 321 322 wq_list_add_tail(&req->comp_list, &state->compl_reqs); 323 } 324 325 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx) 326 { 327 if (unlikely(ctx->off_timeout_used || ctx->drain_active || 328 ctx->has_evfd || ctx->poll_activated)) 329 __io_commit_cqring_flush(ctx); 330 } 331 332 static inline void io_get_task_refs(int nr) 333 { 334 struct io_uring_task *tctx = current->io_uring; 335 336 tctx->cached_refs -= nr; 337 if (unlikely(tctx->cached_refs < 0)) 338 io_task_refs_refill(tctx); 339 } 340 341 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx) 342 { 343 return !ctx->submit_state.free_list.next; 344 } 345 346 extern struct kmem_cache *req_cachep; 347 348 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx) 349 { 350 struct io_kiocb *req; 351 352 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list); 353 kasan_unpoison_object_data(req_cachep, req); 354 wq_stack_extract(&ctx->submit_state.free_list); 355 return req; 356 } 357 358 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req) 359 { 360 if (unlikely(io_req_cache_empty(ctx))) { 361 if (!__io_alloc_req_refill(ctx)) 362 return false; 363 } 364 *req = io_extract_req(ctx); 365 return true; 366 } 367 368 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx) 369 { 370 return likely(ctx->submitter_task == current); 371 } 372 373 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx) 374 { 375 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) || 376 ctx->submitter_task == current); 377 } 378 379 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res) 380 { 381 io_req_set_res(req, res, 0); 382 req->io_task_work.func = io_req_task_complete; 383 io_req_task_work_add(req); 384 } 385 386 #endif 387