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/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         /*
  20          * A hint to not wake right away but delay until there are enough of
  21          * tw's queued to match the number of CQEs the task is waiting for.
  22          *
  23          * Must not be used wirh requests generating more than one CQE.
  24          * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
  25          */
  26         IOU_F_TWQ_LAZY_WAKE                     = 1,
  27 };
  28
  29 enum {
  30         IOU_OK                  = 0,
  31         IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
  32
  33         /*
  34          * Intended only when both IO_URING_F_MULTISHOT is passed
  35          * to indicate to the poll runner that multishot should be
  36          * removed and the result is set on req->cqe.res.
  37          */
  38         IOU_STOP_MULTISHOT      = -ECANCELED,
  39 };
  40
  41 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow);
  42 bool io_req_cqe_overflow(struct io_kiocb *req);
  43 int io_run_task_work_sig(struct io_ring_ctx *ctx);
  44 void io_req_defer_failed(struct io_kiocb *req, s32 res);
  45 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
  46 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
  47 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
  48                 bool allow_overflow);
  49 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
  50
  51 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
  52
  53 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
  54 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
  55                                unsigned issue_flags);
  56
  57 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
  58 bool io_is_uring_fops(struct file *file);
  59 bool io_alloc_async_data(struct io_kiocb *req);
  60 void io_req_task_queue(struct io_kiocb *req);
  61 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use);
  62 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
  63 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
  64 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
  65 void tctx_task_work(struct callback_head *cb);
  66 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
  67 int io_uring_alloc_task_context(struct task_struct *task,
  68                                 struct io_ring_ctx *ctx);
  69
  70 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
  71                                      int start, int end);
  72
  73 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
  74 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
  75 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
  76 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
  77 int io_req_prep_async(struct io_kiocb *req);
  78
  79 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
  80 void io_wq_submit_work(struct io_wq_work *work);
  81
  82 void io_free_req(struct io_kiocb *req);
  83 void io_queue_next(struct io_kiocb *req);
  84 void io_task_refs_refill(struct io_uring_task *tctx);
  85 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
  86
  87 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
  88                         bool cancel_all);
  89
  90 #define io_lockdep_assert_cq_locked(ctx)                                \
  91         do {                                                            \
  92                 lockdep_assert(in_task());                              \
  93                                                                         \
  94                 if (ctx->flags & IORING_SETUP_IOPOLL) {                 \
  95                         lockdep_assert_held(&ctx->uring_lock);          \
  96                 } else if (!ctx->task_complete) {                       \
  97                         lockdep_assert_held(&ctx->completion_lock);     \
  98                 } else if (ctx->submitter_task->flags & PF_EXITING) {   \
  99                         lockdep_assert(current_work());                 \
 100                 } else {                                                \
 101                         lockdep_assert(current == ctx->submitter_task); \
 102                 }                                                       \
 103         } while (0)
 104
 105 static inline void io_req_task_work_add(struct io_kiocb *req)
 106 {
 107         __io_req_task_work_add(req, 0);
 108 }
 109
 110 #define io_for_each_link(pos, head) \
 111         for (pos = (head); pos; pos = pos->link)
 112
 113 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
 114                                                        bool overflow)
 115 {
 116         io_lockdep_assert_cq_locked(ctx);
 117
 118         if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
 119                 struct io_uring_cqe *cqe = ctx->cqe_cached;
 120
 121                 ctx->cached_cq_tail++;
 122                 ctx->cqe_cached++;
 123                 if (ctx->flags & IORING_SETUP_CQE32)
 124                         ctx->cqe_cached++;
 125                 return cqe;
 126         }
 127
 128         return __io_get_cqe(ctx, overflow);
 129 }
 130
 131 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
 132 {
 133         return io_get_cqe_overflow(ctx, false);
 134 }
 135
 136 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
 137                                      struct io_kiocb *req)
 138 {
 139         struct io_uring_cqe *cqe;
 140
 141         /*
 142          * If we can't get a cq entry, userspace overflowed the
 143          * submission (by quite a lot). Increment the overflow count in
 144          * the ring.
 145          */
 146         cqe = io_get_cqe(ctx);
 147         if (unlikely(!cqe))
 148                 return false;
 149
 150         trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
 151                                 req->cqe.res, req->cqe.flags,
 152                                 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
 153                                 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
 154
 155         memcpy(cqe, &req->cqe, sizeof(*cqe));
 156
 157         if (ctx->flags & IORING_SETUP_CQE32) {
 158                 u64 extra1 = 0, extra2 = 0;
 159
 160                 if (req->flags & REQ_F_CQE32_INIT) {
 161                         extra1 = req->extra1;
 162                         extra2 = req->extra2;
 163                 }
 164
 165                 WRITE_ONCE(cqe->big_cqe[0], extra1);
 166                 WRITE_ONCE(cqe->big_cqe[1], extra2);
 167         }
 168         return true;
 169 }
 170
 171 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
 172                                    struct io_kiocb *req)
 173 {
 174         if (likely(__io_fill_cqe_req(ctx, req)))
 175                 return true;
 176         return io_req_cqe_overflow(req);
 177 }
 178
 179 static inline void req_set_fail(struct io_kiocb *req)
 180 {
 181         req->flags |= REQ_F_FAIL;
 182         if (req->flags & REQ_F_CQE_SKIP) {
 183                 req->flags &= ~REQ_F_CQE_SKIP;
 184                 req->flags |= REQ_F_SKIP_LINK_CQES;
 185         }
 186 }
 187
 188 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
 189 {
 190         req->cqe.res = res;
 191         req->cqe.flags = cflags;
 192 }
 193
 194 static inline bool req_has_async_data(struct io_kiocb *req)
 195 {
 196         return req->flags & REQ_F_ASYNC_DATA;
 197 }
 198
 199 static inline void io_put_file(struct file *file)
 200 {
 201         if (file)
 202                 fput(file);
 203 }
 204
 205 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
 206                                          unsigned issue_flags)
 207 {
 208         lockdep_assert_held(&ctx->uring_lock);
 209         if (issue_flags & IO_URING_F_UNLOCKED)
 210                 mutex_unlock(&ctx->uring_lock);
 211 }
 212
 213 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
 214                                        unsigned issue_flags)
 215 {
 216         /*
 217          * "Normal" inline submissions always hold the uring_lock, since we
 218          * grab it from the system call. Same is true for the SQPOLL offload.
 219          * The only exception is when we've detached the request and issue it
 220          * from an async worker thread, grab the lock for that case.
 221          */
 222         if (issue_flags & IO_URING_F_UNLOCKED)
 223                 mutex_lock(&ctx->uring_lock);
 224         lockdep_assert_held(&ctx->uring_lock);
 225 }
 226
 227 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
 228 {
 229         /* order cqe stores with ring update */
 230         smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
 231 }
 232
 233 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
 234 {
 235         if (wq_has_sleeper(&ctx->poll_wq))
 236                 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
 237                                 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
 238 }
 239
 240 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
 241 {
 242         /*
 243          * Trigger waitqueue handler on all waiters on our waitqueue. This
 244          * won't necessarily wake up all the tasks, io_should_wake() will make
 245          * that decision.
 246          *
 247          * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
 248          * set in the mask so that if we recurse back into our own poll
 249          * waitqueue handlers, we know we have a dependency between eventfd or
 250          * epoll and should terminate multishot poll at that point.
 251          */
 252         if (wq_has_sleeper(&ctx->cq_wait))
 253                 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
 254                                 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
 255 }
 256
 257 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
 258 {
 259         struct io_rings *r = ctx->rings;
 260
 261         return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
 262 }
 263
 264 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
 265 {
 266         struct io_rings *rings = ctx->rings;
 267         unsigned int entries;
 268
 269         /* make sure SQ entry isn't read before tail */
 270         entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
 271         return min(entries, ctx->sq_entries);
 272 }
 273
 274 static inline int io_run_task_work(void)
 275 {
 276         /*
 277          * Always check-and-clear the task_work notification signal. With how
 278          * signaling works for task_work, we can find it set with nothing to
 279          * run. We need to clear it for that case, like get_signal() does.
 280          */
 281         if (test_thread_flag(TIF_NOTIFY_SIGNAL))
 282                 clear_notify_signal();
 283         /*
 284          * PF_IO_WORKER never returns to userspace, so check here if we have
 285          * notify work that needs processing.
 286          */
 287         if (current->flags & PF_IO_WORKER &&
 288             test_thread_flag(TIF_NOTIFY_RESUME)) {
 289                 __set_current_state(TASK_RUNNING);
 290                 resume_user_mode_work(NULL);
 291         }
 292         if (task_work_pending(current)) {
 293                 __set_current_state(TASK_RUNNING);
 294                 task_work_run();
 295                 return 1;
 296         }
 297
 298         return 0;
 299 }
 300
 301 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
 302 {
 303         return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
 304 }
 305
 306 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
 307 {
 308         if (!ts->locked) {
 309                 mutex_lock(&ctx->uring_lock);
 310                 ts->locked = true;
 311         }
 312 }
 313
 314 /*
 315  * Don't complete immediately but use deferred completion infrastructure.
 316  * Protected by ->uring_lock and can only be used either with
 317  * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
 318  */
 319 static inline void io_req_complete_defer(struct io_kiocb *req)
 320         __must_hold(&req->ctx->uring_lock)
 321 {
 322         struct io_submit_state *state = &req->ctx->submit_state;
 323
 324         lockdep_assert_held(&req->ctx->uring_lock);
 325
 326         wq_list_add_tail(&req->comp_list, &state->compl_reqs);
 327 }
 328
 329 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
 330 {
 331         if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
 332                      ctx->has_evfd || ctx->poll_activated))
 333                 __io_commit_cqring_flush(ctx);
 334 }
 335
 336 static inline void io_get_task_refs(int nr)
 337 {
 338         struct io_uring_task *tctx = current->io_uring;
 339
 340         tctx->cached_refs -= nr;
 341         if (unlikely(tctx->cached_refs < 0))
 342                 io_task_refs_refill(tctx);
 343 }
 344
 345 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
 346 {
 347         return !ctx->submit_state.free_list.next;
 348 }
 349
 350 extern struct kmem_cache *req_cachep;
 351
 352 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
 353 {
 354         struct io_kiocb *req;
 355
 356         req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
 357         kasan_unpoison_object_data(req_cachep, req);
 358         wq_stack_extract(&ctx->submit_state.free_list);
 359         return req;
 360 }
 361
 362 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
 363 {
 364         if (unlikely(io_req_cache_empty(ctx))) {
 365                 if (!__io_alloc_req_refill(ctx))
 366                         return false;
 367         }
 368         *req = io_extract_req(ctx);
 369         return true;
 370 }
 371
 372 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
 373 {
 374         return likely(ctx->submitter_task == current);
 375 }
 376
 377 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
 378 {
 379         return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
 380                       ctx->submitter_task == current);
 381 }
 382
 383 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
 384 {
 385         io_req_set_res(req, res, 0);
 386         req->io_task_work.func = io_req_task_complete;
 387         io_req_task_work_add(req);
 388 }
 389
 390 /*
 391  * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
 392  * slot.
 393  */
 394 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
 395 {
 396         if (ctx->flags & IORING_SETUP_SQE128)
 397                 return 2 * sizeof(struct io_uring_sqe);
 398         return sizeof(struct io_uring_sqe);
 399 }
 400 #endif