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