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 "filetable.h" 9 10 #ifndef CREATE_TRACE_POINTS 11 #include <trace/events/io_uring.h> 12 #endif 13 14 enum { 15 IOU_OK = 0, 16 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED, 17 }; 18 19 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx); 20 bool io_req_cqe_overflow(struct io_kiocb *req); 21 22 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx) 23 { 24 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) { 25 struct io_uring_cqe *cqe = ctx->cqe_cached; 26 27 ctx->cached_cq_tail++; 28 ctx->cqe_cached++; 29 if (ctx->flags & IORING_SETUP_CQE32) 30 ctx->cqe_cached++; 31 return cqe; 32 } 33 34 return __io_get_cqe(ctx); 35 } 36 37 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx, 38 struct io_kiocb *req) 39 { 40 struct io_uring_cqe *cqe; 41 42 trace_io_uring_complete(req->ctx, req, req->cqe.user_data, 43 req->cqe.res, req->cqe.flags, 44 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0, 45 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0); 46 /* 47 * If we can't get a cq entry, userspace overflowed the 48 * submission (by quite a lot). Increment the overflow count in 49 * the ring. 50 */ 51 cqe = io_get_cqe(ctx); 52 if (unlikely(!cqe)) 53 return io_req_cqe_overflow(req); 54 memcpy(cqe, &req->cqe, sizeof(*cqe)); 55 56 if (ctx->flags & IORING_SETUP_CQE32) { 57 u64 extra1 = 0, extra2 = 0; 58 59 if (req->flags & REQ_F_CQE32_INIT) { 60 extra1 = req->extra1; 61 extra2 = req->extra2; 62 } 63 64 WRITE_ONCE(cqe->big_cqe[0], extra1); 65 WRITE_ONCE(cqe->big_cqe[1], extra2); 66 } 67 return true; 68 } 69 70 static inline void req_set_fail(struct io_kiocb *req) 71 { 72 req->flags |= REQ_F_FAIL; 73 if (req->flags & REQ_F_CQE_SKIP) { 74 req->flags &= ~REQ_F_CQE_SKIP; 75 req->flags |= REQ_F_SKIP_LINK_CQES; 76 } 77 } 78 79 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags) 80 { 81 req->cqe.res = res; 82 req->cqe.flags = cflags; 83 } 84 85 static inline bool req_has_async_data(struct io_kiocb *req) 86 { 87 return req->flags & REQ_F_ASYNC_DATA; 88 } 89 90 static inline void io_put_file(struct file *file) 91 { 92 if (file) 93 fput(file); 94 } 95 96 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx, 97 unsigned issue_flags) 98 { 99 lockdep_assert_held(&ctx->uring_lock); 100 if (issue_flags & IO_URING_F_UNLOCKED) 101 mutex_unlock(&ctx->uring_lock); 102 } 103 104 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx, 105 unsigned issue_flags) 106 { 107 /* 108 * "Normal" inline submissions always hold the uring_lock, since we 109 * grab it from the system call. Same is true for the SQPOLL offload. 110 * The only exception is when we've detached the request and issue it 111 * from an async worker thread, grab the lock for that case. 112 */ 113 if (issue_flags & IO_URING_F_UNLOCKED) 114 mutex_lock(&ctx->uring_lock); 115 lockdep_assert_held(&ctx->uring_lock); 116 } 117 118 static inline void io_commit_cqring(struct io_ring_ctx *ctx) 119 { 120 /* order cqe stores with ring update */ 121 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail); 122 } 123 124 static inline void io_cqring_wake(struct io_ring_ctx *ctx) 125 { 126 /* 127 * wake_up_all() may seem excessive, but io_wake_function() and 128 * io_should_wake() handle the termination of the loop and only 129 * wake as many waiters as we need to. 130 */ 131 if (wq_has_sleeper(&ctx->cq_wait)) 132 wake_up_all(&ctx->cq_wait); 133 } 134 135 static inline bool io_sqring_full(struct io_ring_ctx *ctx) 136 { 137 struct io_rings *r = ctx->rings; 138 139 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries; 140 } 141 142 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) 143 { 144 struct io_rings *rings = ctx->rings; 145 146 /* make sure SQ entry isn't read before tail */ 147 return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; 148 } 149 150 static inline bool io_run_task_work(void) 151 { 152 if (test_thread_flag(TIF_NOTIFY_SIGNAL) || task_work_pending(current)) { 153 __set_current_state(TASK_RUNNING); 154 clear_notify_signal(); 155 if (task_work_pending(current)) 156 task_work_run(); 157 return true; 158 } 159 160 return false; 161 } 162 163 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked) 164 { 165 if (!*locked) { 166 mutex_lock(&ctx->uring_lock); 167 *locked = true; 168 } 169 } 170 171 static inline void io_req_add_compl_list(struct io_kiocb *req) 172 { 173 struct io_submit_state *state = &req->ctx->submit_state; 174 175 if (!(req->flags & REQ_F_CQE_SKIP)) 176 state->flush_cqes = true; 177 wq_list_add_tail(&req->comp_list, &state->compl_reqs); 178 } 179 180 int io_run_task_work_sig(void); 181 void io_req_complete_failed(struct io_kiocb *req, s32 res); 182 void __io_req_complete(struct io_kiocb *req, unsigned issue_flags); 183 void io_req_complete_post(struct io_kiocb *req); 184 void __io_req_complete_post(struct io_kiocb *req); 185 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags); 186 void io_cqring_ev_posted(struct io_ring_ctx *ctx); 187 void __io_commit_cqring_flush(struct io_ring_ctx *ctx); 188 189 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages); 190 191 struct file *io_file_get_normal(struct io_kiocb *req, int fd); 192 struct file *io_file_get_fixed(struct io_kiocb *req, int fd, 193 unsigned issue_flags); 194 195 bool io_is_uring_fops(struct file *file); 196 bool io_alloc_async_data(struct io_kiocb *req); 197 void io_req_task_work_add(struct io_kiocb *req); 198 void io_req_task_prio_work_add(struct io_kiocb *req); 199 void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags); 200 void io_req_task_queue(struct io_kiocb *req); 201 void io_queue_iowq(struct io_kiocb *req, bool *dont_use); 202 void io_req_task_complete(struct io_kiocb *req, bool *locked); 203 void io_req_task_queue_fail(struct io_kiocb *req, int ret); 204 void io_req_task_submit(struct io_kiocb *req, bool *locked); 205 void tctx_task_work(struct callback_head *cb); 206 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd); 207 int io_uring_alloc_task_context(struct task_struct *task, 208 struct io_ring_ctx *ctx); 209 210 int io_poll_issue(struct io_kiocb *req, bool *locked); 211 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr); 212 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin); 213 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node); 214 int io_req_prep_async(struct io_kiocb *req); 215 216 struct io_wq_work *io_wq_free_work(struct io_wq_work *work); 217 void io_wq_submit_work(struct io_wq_work *work); 218 219 void io_free_req(struct io_kiocb *req); 220 void io_queue_next(struct io_kiocb *req); 221 222 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task, 223 bool cancel_all); 224 225 #define io_for_each_link(pos, head) \ 226 for (pos = (head); pos; pos = pos->link) 227 228 #endif 229