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