io_uring/io_uring.h

   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_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_put_task(struct task_struct *task, int nr);
  76 void io_task_refs_refill(struct io_uring_task *tctx);
  77 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
  78
  79 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
  80                         bool cancel_all);
  81
  82 static inline void io_req_task_work_add(struct io_kiocb *req)
  83 {
  84         __io_req_task_work_add(req, true);
  85 }
  86
  87 #define io_for_each_link(pos, head) \
  88         for (pos = (head); pos; pos = pos->link)
  89
  90 void io_cq_unlock_post(struct io_ring_ctx *ctx);
  91
  92 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
  93                                                        bool overflow)
  94 {
  95         if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
  96                 struct io_uring_cqe *cqe = ctx->cqe_cached;
  97
  98                 ctx->cached_cq_tail++;
  99                 ctx->cqe_cached++;
 100                 if (ctx->flags & IORING_SETUP_CQE32)
 101                         ctx->cqe_cached++;
 102                 return cqe;
 103         }
 104
 105         return __io_get_cqe(ctx, overflow);
 106 }
 107
 108 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
 109 {
 110         return io_get_cqe_overflow(ctx, false);
 111 }
 112
 113 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
 114                                      struct io_kiocb *req)
 115 {
 116         struct io_uring_cqe *cqe;
 117
 118         /*
 119          * If we can't get a cq entry, userspace overflowed the
 120          * submission (by quite a lot). Increment the overflow count in
 121          * the ring.
 122          */
 123         cqe = io_get_cqe(ctx);
 124         if (unlikely(!cqe))
 125                 return false;
 126
 127         trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
 128                                 req->cqe.res, req->cqe.flags,
 129                                 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
 130                                 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
 131
 132         memcpy(cqe, &req->cqe, sizeof(*cqe));
 133
 134         if (ctx->flags & IORING_SETUP_CQE32) {
 135                 u64 extra1 = 0, extra2 = 0;
 136
 137                 if (req->flags & REQ_F_CQE32_INIT) {
 138                         extra1 = req->extra1;
 139                         extra2 = req->extra2;
 140                 }
 141
 142                 WRITE_ONCE(cqe->big_cqe[0], extra1);
 143                 WRITE_ONCE(cqe->big_cqe[1], extra2);
 144         }
 145         return true;
 146 }
 147
 148 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
 149                                    struct io_kiocb *req)
 150 {
 151         if (likely(__io_fill_cqe_req(ctx, req)))
 152                 return true;
 153         return io_req_cqe_overflow(req);
 154 }
 155
 156 static inline void req_set_fail(struct io_kiocb *req)
 157 {
 158         req->flags |= REQ_F_FAIL;
 159         if (req->flags & REQ_F_CQE_SKIP) {
 160                 req->flags &= ~REQ_F_CQE_SKIP;
 161                 req->flags |= REQ_F_SKIP_LINK_CQES;
 162         }
 163 }
 164
 165 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
 166 {
 167         req->cqe.res = res;
 168         req->cqe.flags = cflags;
 169 }
 170
 171 static inline bool req_has_async_data(struct io_kiocb *req)
 172 {
 173         return req->flags & REQ_F_ASYNC_DATA;
 174 }
 175
 176 static inline void io_put_file(struct file *file)
 177 {
 178         if (file)
 179                 fput(file);
 180 }
 181
 182 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
 183                                          unsigned issue_flags)
 184 {
 185         lockdep_assert_held(&ctx->uring_lock);
 186         if (issue_flags & IO_URING_F_UNLOCKED)
 187                 mutex_unlock(&ctx->uring_lock);
 188 }
 189
 190 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
 191                                        unsigned issue_flags)
 192 {
 193         /*
 194          * "Normal" inline submissions always hold the uring_lock, since we
 195          * grab it from the system call. Same is true for the SQPOLL offload.
 196          * The only exception is when we've detached the request and issue it
 197          * from an async worker thread, grab the lock for that case.
 198          */
 199         if (issue_flags & IO_URING_F_UNLOCKED)
 200                 mutex_lock(&ctx->uring_lock);
 201         lockdep_assert_held(&ctx->uring_lock);
 202 }
 203
 204 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
 205 {
 206         /* order cqe stores with ring update */
 207         smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
 208 }
 209
 210 /* requires smb_mb() prior, see wq_has_sleeper() */
 211 static inline void __io_cqring_wake(struct io_ring_ctx *ctx)
 212 {
 213         /*
 214          * Trigger waitqueue handler on all waiters on our waitqueue. This
 215          * won't necessarily wake up all the tasks, io_should_wake() will make
 216          * that decision.
 217          *
 218          * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
 219          * set in the mask so that if we recurse back into our own poll
 220          * waitqueue handlers, we know we have a dependency between eventfd or
 221          * epoll and should terminate multishot poll at that point.
 222          */
 223         if (waitqueue_active(&ctx->cq_wait))
 224                 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
 225                                 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
 226 }
 227
 228 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
 229 {
 230         smp_mb();
 231         __io_cqring_wake(ctx);
 232 }
 233
 234 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
 235 {
 236         struct io_rings *r = ctx->rings;
 237
 238         return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
 239 }
 240
 241 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
 242 {
 243         struct io_rings *rings = ctx->rings;
 244
 245         /* make sure SQ entry isn't read before tail */
 246         return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
 247 }
 248
 249 static inline int io_run_task_work(void)
 250 {
 251         /*
 252          * Always check-and-clear the task_work notification signal. With how
 253          * signaling works for task_work, we can find it set with nothing to
 254          * run. We need to clear it for that case, like get_signal() does.
 255          */
 256         if (test_thread_flag(TIF_NOTIFY_SIGNAL))
 257                 clear_notify_signal();
 258         if (task_work_pending(current)) {
 259                 __set_current_state(TASK_RUNNING);
 260                 task_work_run();
 261                 return 1;
 262         }
 263
 264         return 0;
 265 }
 266
 267 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
 268 {
 269         return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
 270 }
 271
 272 static inline int io_run_task_work_ctx(struct io_ring_ctx *ctx)
 273 {
 274         int ret = 0;
 275         int ret2;
 276
 277         if (ctx->flags & IORING_SETUP_DEFER_TASKRUN)
 278                 ret = io_run_local_work(ctx);
 279
 280         /* want to run this after in case more is added */
 281         ret2 = io_run_task_work();
 282
 283         /* Try propagate error in favour of if tasks were run,
 284          * but still make sure to run them if requested
 285          */
 286         if (ret >= 0)
 287                 ret += ret2;
 288
 289         return ret;
 290 }
 291
 292 static inline int io_run_local_work_locked(struct io_ring_ctx *ctx)
 293 {
 294         bool locked;
 295         int ret;
 296
 297         if (llist_empty(&ctx->work_llist))
 298                 return 0;
 299
 300         locked = true;
 301         ret = __io_run_local_work(ctx, &locked);
 302         /* shouldn't happen! */
 303         if (WARN_ON_ONCE(!locked))
 304                 mutex_lock(&ctx->uring_lock);
 305         return ret;
 306 }
 307
 308 static inline void io_tw_lock(struct io_ring_ctx *ctx, bool *locked)
 309 {
 310         if (!*locked) {
 311                 mutex_lock(&ctx->uring_lock);
 312                 *locked = true;
 313         }
 314 }
 315
 316 /*
 317  * Don't complete immediately but use deferred completion infrastructure.
 318  * Protected by ->uring_lock and can only be used either with
 319  * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
 320  */
 321 static inline void io_req_complete_defer(struct io_kiocb *req)
 322         __must_hold(&req->ctx->uring_lock)
 323 {
 324         struct io_submit_state *state = &req->ctx->submit_state;
 325
 326         lockdep_assert_held(&req->ctx->uring_lock);
 327
 328         wq_list_add_tail(&req->comp_list, &state->compl_reqs);
 329 }
 330
 331 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
 332 {
 333         if (unlikely(ctx->off_timeout_used || ctx->drain_active || ctx->has_evfd))
 334                 __io_commit_cqring_flush(ctx);
 335 }
 336
 337 /* must to be called somewhat shortly after putting a request */
 338 static inline void io_put_task(struct task_struct *task, int nr)
 339 {
 340         if (likely(task == current))
 341                 task->io_uring->cached_refs += nr;
 342         else
 343                 __io_put_task(task, nr);
 344 }
 345
 346 static inline void io_get_task_refs(int nr)
 347 {
 348         struct io_uring_task *tctx = current->io_uring;
 349
 350         tctx->cached_refs -= nr;
 351         if (unlikely(tctx->cached_refs < 0))
 352                 io_task_refs_refill(tctx);
 353 }
 354
 355 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
 356 {
 357         return !ctx->submit_state.free_list.next;
 358 }
 359
 360 static inline bool io_alloc_req_refill(struct io_ring_ctx *ctx)
 361 {
 362         if (unlikely(io_req_cache_empty(ctx)))
 363                 return __io_alloc_req_refill(ctx);
 364         return true;
 365 }
 366
 367 static inline struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx)
 368 {
 369         struct io_wq_work_node *node;
 370
 371         node = wq_stack_extract(&ctx->submit_state.free_list);
 372         return container_of(node, struct io_kiocb, comp_list);
 373 }
 374
 375 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
 376 {
 377         return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
 378                       ctx->submitter_task == current);
 379 }
 380
 381 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
 382 {
 383         io_req_set_res(req, res, 0);
 384         req->io_task_work.func = io_req_task_complete;
 385         io_req_task_work_add(req);
 386 }
 387
 388 #endif