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>
12 #include "filetable.h"
14 #ifndef CREATE_TRACE_POINTS
15 #include <trace/events/io_uring.h>
19 /* don't use deferred task_work */
20 IOU_F_TWQ_FORCE_NORMAL = 1,
23 * A hint to not wake right away but delay until there are enough of
24 * tw's queued to match the number of CQEs the task is waiting for.
26 * Must not be used wirh requests generating more than one CQE.
27 * It's also ignored unless IORING_SETUP_DEFER_TASKRUN is set.
29 IOU_F_TWQ_LAZY_WAKE = 2,
34 IOU_ISSUE_SKIP_COMPLETE = -EIOCBQUEUED,
37 * Intended only when both IO_URING_F_MULTISHOT is passed
38 * to indicate to the poll runner that multishot should be
39 * removed and the result is set on req->cqe.res.
41 IOU_STOP_MULTISHOT = -ECANCELED,
44 struct io_uring_cqe *__io_get_cqe(struct io_ring_ctx *ctx, bool overflow);
45 bool io_req_cqe_overflow(struct io_kiocb *req);
46 int io_run_task_work_sig(struct io_ring_ctx *ctx);
47 void io_req_defer_failed(struct io_kiocb *req, s32 res);
48 void io_req_complete_post(struct io_kiocb *req, unsigned issue_flags);
49 bool io_post_aux_cqe(struct io_ring_ctx *ctx, u64 user_data, s32 res, u32 cflags);
50 bool io_aux_cqe(const struct io_kiocb *req, bool defer, s32 res, u32 cflags,
52 void __io_commit_cqring_flush(struct io_ring_ctx *ctx);
54 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages);
56 struct file *io_file_get_normal(struct io_kiocb *req, int fd);
57 struct file *io_file_get_fixed(struct io_kiocb *req, int fd,
58 unsigned issue_flags);
60 void __io_req_task_work_add(struct io_kiocb *req, unsigned flags);
61 bool io_is_uring_fops(struct file *file);
62 bool io_alloc_async_data(struct io_kiocb *req);
63 void io_req_task_queue(struct io_kiocb *req);
64 void io_queue_iowq(struct io_kiocb *req, struct io_tw_state *ts_dont_use);
65 void io_req_task_complete(struct io_kiocb *req, struct io_tw_state *ts);
66 void io_req_task_queue_fail(struct io_kiocb *req, int ret);
67 void io_req_task_submit(struct io_kiocb *req, struct io_tw_state *ts);
68 void tctx_task_work(struct callback_head *cb);
69 __cold void io_uring_cancel_generic(bool cancel_all, struct io_sq_data *sqd);
70 int io_uring_alloc_task_context(struct task_struct *task,
71 struct io_ring_ctx *ctx);
73 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
76 int io_poll_issue(struct io_kiocb *req, struct io_tw_state *ts);
77 int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr);
78 int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin);
79 void io_free_batch_list(struct io_ring_ctx *ctx, struct io_wq_work_node *node);
80 int io_req_prep_async(struct io_kiocb *req);
82 struct io_wq_work *io_wq_free_work(struct io_wq_work *work);
83 void io_wq_submit_work(struct io_wq_work *work);
85 void io_free_req(struct io_kiocb *req);
86 void io_queue_next(struct io_kiocb *req);
87 void io_task_refs_refill(struct io_uring_task *tctx);
88 bool __io_alloc_req_refill(struct io_ring_ctx *ctx);
90 bool io_match_task_safe(struct io_kiocb *head, struct task_struct *task,
93 #define io_lockdep_assert_cq_locked(ctx) \
95 lockdep_assert(in_task()); \
97 if (ctx->flags & IORING_SETUP_IOPOLL) { \
98 lockdep_assert_held(&ctx->uring_lock); \
99 } else if (!ctx->task_complete) { \
100 lockdep_assert_held(&ctx->completion_lock); \
101 } else if (ctx->submitter_task->flags & PF_EXITING) { \
102 lockdep_assert(current_work()); \
104 lockdep_assert(current == ctx->submitter_task); \
108 static inline void io_req_task_work_add(struct io_kiocb *req)
110 __io_req_task_work_add(req, 0);
113 #define io_for_each_link(pos, head) \
114 for (pos = (head); pos; pos = pos->link)
116 void io_cq_unlock_post(struct io_ring_ctx *ctx);
118 static inline struct io_uring_cqe *io_get_cqe_overflow(struct io_ring_ctx *ctx,
121 io_lockdep_assert_cq_locked(ctx);
123 if (likely(ctx->cqe_cached < ctx->cqe_sentinel)) {
124 struct io_uring_cqe *cqe = ctx->cqe_cached;
126 ctx->cached_cq_tail++;
128 if (ctx->flags & IORING_SETUP_CQE32)
133 return __io_get_cqe(ctx, overflow);
136 static inline struct io_uring_cqe *io_get_cqe(struct io_ring_ctx *ctx)
138 return io_get_cqe_overflow(ctx, false);
141 static inline bool __io_fill_cqe_req(struct io_ring_ctx *ctx,
142 struct io_kiocb *req)
144 struct io_uring_cqe *cqe;
147 * If we can't get a cq entry, userspace overflowed the
148 * submission (by quite a lot). Increment the overflow count in
151 cqe = io_get_cqe(ctx);
155 trace_io_uring_complete(req->ctx, req, req->cqe.user_data,
156 req->cqe.res, req->cqe.flags,
157 (req->flags & REQ_F_CQE32_INIT) ? req->extra1 : 0,
158 (req->flags & REQ_F_CQE32_INIT) ? req->extra2 : 0);
160 memcpy(cqe, &req->cqe, sizeof(*cqe));
162 if (ctx->flags & IORING_SETUP_CQE32) {
163 u64 extra1 = 0, extra2 = 0;
165 if (req->flags & REQ_F_CQE32_INIT) {
166 extra1 = req->extra1;
167 extra2 = req->extra2;
170 WRITE_ONCE(cqe->big_cqe[0], extra1);
171 WRITE_ONCE(cqe->big_cqe[1], extra2);
176 static inline bool io_fill_cqe_req(struct io_ring_ctx *ctx,
177 struct io_kiocb *req)
179 if (likely(__io_fill_cqe_req(ctx, req)))
181 return io_req_cqe_overflow(req);
184 static inline void req_set_fail(struct io_kiocb *req)
186 req->flags |= REQ_F_FAIL;
187 if (req->flags & REQ_F_CQE_SKIP) {
188 req->flags &= ~REQ_F_CQE_SKIP;
189 req->flags |= REQ_F_SKIP_LINK_CQES;
193 static inline void io_req_set_res(struct io_kiocb *req, s32 res, u32 cflags)
196 req->cqe.flags = cflags;
199 static inline bool req_has_async_data(struct io_kiocb *req)
201 return req->flags & REQ_F_ASYNC_DATA;
204 static inline void io_put_file(struct file *file)
210 static inline void io_ring_submit_unlock(struct io_ring_ctx *ctx,
211 unsigned issue_flags)
213 lockdep_assert_held(&ctx->uring_lock);
214 if (issue_flags & IO_URING_F_UNLOCKED)
215 mutex_unlock(&ctx->uring_lock);
218 static inline void io_ring_submit_lock(struct io_ring_ctx *ctx,
219 unsigned issue_flags)
222 * "Normal" inline submissions always hold the uring_lock, since we
223 * grab it from the system call. Same is true for the SQPOLL offload.
224 * The only exception is when we've detached the request and issue it
225 * from an async worker thread, grab the lock for that case.
227 if (issue_flags & IO_URING_F_UNLOCKED)
228 mutex_lock(&ctx->uring_lock);
229 lockdep_assert_held(&ctx->uring_lock);
232 static inline void io_commit_cqring(struct io_ring_ctx *ctx)
234 /* order cqe stores with ring update */
235 smp_store_release(&ctx->rings->cq.tail, ctx->cached_cq_tail);
238 static inline void io_poll_wq_wake(struct io_ring_ctx *ctx)
240 if (wq_has_sleeper(&ctx->poll_wq))
241 __wake_up(&ctx->poll_wq, TASK_NORMAL, 0,
242 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
245 static inline void io_cqring_wake(struct io_ring_ctx *ctx)
248 * Trigger waitqueue handler on all waiters on our waitqueue. This
249 * won't necessarily wake up all the tasks, io_should_wake() will make
252 * Pass in EPOLLIN|EPOLL_URING_WAKE as the poll wakeup key. The latter
253 * set in the mask so that if we recurse back into our own poll
254 * waitqueue handlers, we know we have a dependency between eventfd or
255 * epoll and should terminate multishot poll at that point.
257 if (wq_has_sleeper(&ctx->cq_wait))
258 __wake_up(&ctx->cq_wait, TASK_NORMAL, 0,
259 poll_to_key(EPOLL_URING_WAKE | EPOLLIN));
262 static inline bool io_sqring_full(struct io_ring_ctx *ctx)
264 struct io_rings *r = ctx->rings;
266 return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == ctx->sq_entries;
269 static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx)
271 struct io_rings *rings = ctx->rings;
272 unsigned int entries;
274 /* make sure SQ entry isn't read before tail */
275 entries = smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head;
276 return min(entries, ctx->sq_entries);
279 static inline int io_run_task_work(void)
282 * Always check-and-clear the task_work notification signal. With how
283 * signaling works for task_work, we can find it set with nothing to
284 * run. We need to clear it for that case, like get_signal() does.
286 if (test_thread_flag(TIF_NOTIFY_SIGNAL))
287 clear_notify_signal();
289 * PF_IO_WORKER never returns to userspace, so check here if we have
290 * notify work that needs processing.
292 if (current->flags & PF_IO_WORKER &&
293 test_thread_flag(TIF_NOTIFY_RESUME)) {
294 __set_current_state(TASK_RUNNING);
295 resume_user_mode_work(NULL);
297 if (task_work_pending(current)) {
298 __set_current_state(TASK_RUNNING);
306 static inline bool io_task_work_pending(struct io_ring_ctx *ctx)
308 return task_work_pending(current) || !wq_list_empty(&ctx->work_llist);
311 static inline void io_tw_lock(struct io_ring_ctx *ctx, struct io_tw_state *ts)
314 mutex_lock(&ctx->uring_lock);
320 * Don't complete immediately but use deferred completion infrastructure.
321 * Protected by ->uring_lock and can only be used either with
322 * IO_URING_F_COMPLETE_DEFER or inside a tw handler holding the mutex.
324 static inline void io_req_complete_defer(struct io_kiocb *req)
325 __must_hold(&req->ctx->uring_lock)
327 struct io_submit_state *state = &req->ctx->submit_state;
329 lockdep_assert_held(&req->ctx->uring_lock);
331 wq_list_add_tail(&req->comp_list, &state->compl_reqs);
334 static inline void io_commit_cqring_flush(struct io_ring_ctx *ctx)
336 if (unlikely(ctx->off_timeout_used || ctx->drain_active ||
337 ctx->has_evfd || ctx->poll_activated))
338 __io_commit_cqring_flush(ctx);
341 static inline void io_get_task_refs(int nr)
343 struct io_uring_task *tctx = current->io_uring;
345 tctx->cached_refs -= nr;
346 if (unlikely(tctx->cached_refs < 0))
347 io_task_refs_refill(tctx);
350 static inline bool io_req_cache_empty(struct io_ring_ctx *ctx)
352 return !ctx->submit_state.free_list.next;
355 extern struct kmem_cache *req_cachep;
357 static inline struct io_kiocb *io_extract_req(struct io_ring_ctx *ctx)
359 struct io_kiocb *req;
361 req = container_of(ctx->submit_state.free_list.next, struct io_kiocb, comp_list);
362 kasan_unpoison_object_data(req_cachep, req);
363 wq_stack_extract(&ctx->submit_state.free_list);
367 static inline bool io_alloc_req(struct io_ring_ctx *ctx, struct io_kiocb **req)
369 if (unlikely(io_req_cache_empty(ctx))) {
370 if (!__io_alloc_req_refill(ctx))
373 *req = io_extract_req(ctx);
377 static inline bool io_allowed_defer_tw_run(struct io_ring_ctx *ctx)
379 return likely(ctx->submitter_task == current);
382 static inline bool io_allowed_run_tw(struct io_ring_ctx *ctx)
384 return likely(!(ctx->flags & IORING_SETUP_DEFER_TASKRUN) ||
385 ctx->submitter_task == current);
388 static inline void io_req_queue_tw_complete(struct io_kiocb *req, s32 res)
390 io_req_set_res(req, res, 0);
391 req->io_task_work.func = io_req_task_complete;
392 io_req_task_work_add(req);
396 * IORING_SETUP_SQE128 contexts allocate twice the normal SQE size for each
399 static inline size_t uring_sqe_size(struct io_ring_ctx *ctx)
401 if (ctx->flags & IORING_SETUP_SQE128)
402 return 2 * sizeof(struct io_uring_sqe);
403 return sizeof(struct io_uring_sqe);