if (mm) {
kthread_unuse_mm(mm);
mmput(mm);
+ current->mm = NULL;
}
}
static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx)
{
- if (!current->mm) {
- if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) ||
- !ctx->sqo_task->mm ||
- !mmget_not_zero(ctx->sqo_task->mm)))
- return -EFAULT;
- kthread_use_mm(ctx->sqo_task->mm);
+ struct mm_struct *mm;
+
+ if (current->mm)
+ return 0;
+
+ /* Should never happen */
+ if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL)))
+ return -EFAULT;
+
+ task_lock(ctx->sqo_task);
+ mm = ctx->sqo_task->mm;
+ if (unlikely(!mm || !mmget_not_zero(mm)))
+ mm = NULL;
+ task_unlock(ctx->sqo_task);
+
+ if (mm) {
+ kthread_use_mm(mm);
+ return 0;
}
- return 0;
+ return -EFAULT;
}
static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx,
/* add one for this request */
refcount_inc(&id->count);
- /* drop old identity, assign new one. one ref for req, one for tctx */
- if (req->work.identity != tctx->identity &&
- refcount_sub_and_test(2, &req->work.identity->count))
+ /* drop tctx and req identity references, if needed */
+ if (tctx->identity != &tctx->__identity &&
+ refcount_dec_and_test(&tctx->identity->count))
+ kfree(tctx->identity);
+ if (req->work.identity != &tctx->__identity &&
+ refcount_dec_and_test(&req->work.identity->count))
kfree(req->work.identity);
req->work.identity = id;
io_req_init_async(req);
id = req->work.identity;
+ if (req->flags & REQ_F_FORCE_ASYNC)
+ req->work.flags |= IO_WQ_WORK_CONCURRENT;
+
if (req->flags & REQ_F_ISREG) {
if (def->hash_reg_file || (ctx->flags & IORING_SETUP_IOPOLL))
io_wq_hash_work(&req->work, file_inode(req->file));
WRITE_ONCE(cqe->user_data, req->user_data);
WRITE_ONCE(cqe->res, res);
WRITE_ONCE(cqe->flags, cflags);
- } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) {
+ } else if (ctx->cq_overflow_flushed ||
+ atomic_read(&req->task->io_uring->in_idle)) {
/*
* If we're in ring overflow flush mode, or in task cancel mode,
* then we cannot store the request for later flushing, we need
io_dismantle_req(req);
percpu_counter_dec(&tctx->inflight);
- if (tctx->in_idle)
+ if (atomic_read(&tctx->in_idle))
wake_up(&tctx->wait);
put_task_struct(req->task);
percpu_ref_put(&ctx->refs);
}
-static bool io_link_cancel_timeout(struct io_kiocb *req)
+static void io_kill_linked_timeout(struct io_kiocb *req)
{
- struct io_timeout_data *io = req->async_data;
struct io_ring_ctx *ctx = req->ctx;
- int ret;
-
- ret = hrtimer_try_to_cancel(&io->timer);
- if (ret != -1) {
- io_cqring_fill_event(req, -ECANCELED);
- io_commit_cqring(ctx);
- req->flags &= ~REQ_F_LINK_HEAD;
- io_put_req_deferred(req, 1);
- return true;
- }
-
- return false;
-}
-
-static bool __io_kill_linked_timeout(struct io_kiocb *req)
-{
struct io_kiocb *link;
- bool wake_ev;
+ bool cancelled = false;
+ unsigned long flags;
- if (list_empty(&req->link_list))
- return false;
- link = list_first_entry(&req->link_list, struct io_kiocb, link_list);
- if (link->opcode != IORING_OP_LINK_TIMEOUT)
- return false;
+ spin_lock_irqsave(&ctx->completion_lock, flags);
+ link = list_first_entry_or_null(&req->link_list, struct io_kiocb,
+ link_list);
/*
* Can happen if a linked timeout fired and link had been like
* req -> link t-out -> link t-out [-> ...]
*/
- if (!(link->flags & REQ_F_LTIMEOUT_ACTIVE))
- return false;
+ if (link && (link->flags & REQ_F_LTIMEOUT_ACTIVE)) {
+ struct io_timeout_data *io = link->async_data;
+ int ret;
- list_del_init(&link->link_list);
- wake_ev = io_link_cancel_timeout(link);
+ list_del_init(&link->link_list);
+ ret = hrtimer_try_to_cancel(&io->timer);
+ if (ret != -1) {
+ io_cqring_fill_event(link, -ECANCELED);
+ io_commit_cqring(ctx);
+ cancelled = true;
+ }
+ }
req->flags &= ~REQ_F_LINK_TIMEOUT;
- return wake_ev;
-}
-
-static void io_kill_linked_timeout(struct io_kiocb *req)
-{
- struct io_ring_ctx *ctx = req->ctx;
- unsigned long flags;
- bool wake_ev;
-
- spin_lock_irqsave(&ctx->completion_lock, flags);
- wake_ev = __io_kill_linked_timeout(req);
spin_unlock_irqrestore(&ctx->completion_lock, flags);
- if (wake_ev)
+ if (cancelled) {
io_cqring_ev_posted(ctx);
+ io_put_req(link);
+ }
}
static struct io_kiocb *io_req_link_next(struct io_kiocb *req)
/* make sure double remove sees this as being gone */
wait->private = NULL;
spin_unlock(&poll->head->lock);
- if (!done)
- __io_async_wake(req, poll, mask, io_poll_task_func);
+ if (!done) {
+ /* use wait func handler, so it matches the rq type */
+ poll->wait.func(&poll->wait, mode, sync, key);
+ }
}
refcount_dec(&req->refs);
return 1;
static void __io_queue_sqe(struct io_kiocb *req, struct io_comp_state *cs)
{
struct io_kiocb *linked_timeout;
- struct io_kiocb *nxt;
const struct cred *old_creds = NULL;
int ret;
*/
if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) {
if (!io_arm_poll_handler(req)) {
-punt:
/*
* Queued up for async execution, worker will release
* submit reference when the iocb is actually submitted.
if (linked_timeout)
io_queue_linked_timeout(linked_timeout);
- goto exit;
- }
+ } else if (likely(!ret)) {
+ /* drop submission reference */
+ req = io_put_req_find_next(req);
+ if (linked_timeout)
+ io_queue_linked_timeout(linked_timeout);
- if (unlikely(ret)) {
+ if (req) {
+ if (!(req->flags & REQ_F_FORCE_ASYNC))
+ goto again;
+ io_queue_async_work(req);
+ }
+ } else {
/* un-prep timeout, so it'll be killed as any other linked */
req->flags &= ~REQ_F_LINK_TIMEOUT;
req_set_fail_links(req);
io_put_req(req);
io_req_complete(req, ret);
- goto exit;
}
- /* drop submission reference */
- nxt = io_put_req_find_next(req);
- if (linked_timeout)
- io_queue_linked_timeout(linked_timeout);
-
- if (nxt) {
- req = nxt;
-
- if (req->flags & REQ_F_FORCE_ASYNC) {
- linked_timeout = NULL;
- goto punt;
- }
- goto again;
- }
-exit:
if (old_creds)
revert_creds(old_creds);
}
if (unlikely(ret))
goto fail_req;
}
-
- /*
- * Never try inline submit of IOSQE_ASYNC is set, go straight
- * to async execution.
- */
- io_req_init_async(req);
- req->work.flags |= IO_WQ_WORK_CONCURRENT;
io_queue_async_work(req);
} else {
if (sqe) {
xa_init(&tctx->xa);
init_waitqueue_head(&tctx->wait);
tctx->last = NULL;
- tctx->in_idle = 0;
+ atomic_set(&tctx->in_idle, 0);
+ tctx->sqpoll = false;
io_init_identity(&tctx->__identity);
tctx->identity = &tctx->__identity;
task->io_uring = tctx;
{
struct task_struct *task = current;
- if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data)
+ if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
task = ctx->sq_data->thread;
+ atomic_inc(&task->io_uring->in_idle);
+ io_sq_thread_park(ctx->sq_data);
+ }
io_cqring_overflow_flush(ctx, true, task, files);
io_run_task_work();
cond_resched();
}
+
+ if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) {
+ atomic_dec(&task->io_uring->in_idle);
+ /*
+ * If the files that are going away are the ones in the thread
+ * identity, clear them out.
+ */
+ if (task->io_uring->identity->files == files)
+ task->io_uring->identity->files = NULL;
+ io_sq_thread_unpark(ctx->sq_data);
+ }
}
/*
* Note that this task has used io_uring. We use it for cancelation purposes.
*/
-static int io_uring_add_task_file(struct file *file)
+static int io_uring_add_task_file(struct io_ring_ctx *ctx, struct file *file)
{
struct io_uring_task *tctx = current->io_uring;
tctx->last = file;
}
+ /*
+ * This is race safe in that the task itself is doing this, hence it
+ * cannot be going through the exit/cancel paths at the same time.
+ * This cannot be modified while exit/cancel is running.
+ */
+ if (!tctx->sqpoll && (ctx->flags & IORING_SETUP_SQPOLL))
+ tctx->sqpoll = true;
+
return 0;
}
unsigned long index;
/* make sure overflow events are dropped */
- tctx->in_idle = true;
+ atomic_inc(&tctx->in_idle);
xa_for_each(&tctx->xa, index, file) {
struct io_ring_ctx *ctx = file->private_data;
if (files)
io_uring_del_task_file(file);
}
+
+ atomic_dec(&tctx->in_idle);
+}
+
+static s64 tctx_inflight(struct io_uring_task *tctx)
+{
+ unsigned long index;
+ struct file *file;
+ s64 inflight;
+
+ inflight = percpu_counter_sum(&tctx->inflight);
+ if (!tctx->sqpoll)
+ return inflight;
+
+ /*
+ * If we have SQPOLL rings, then we need to iterate and find them, and
+ * add the pending count for those.
+ */
+ xa_for_each(&tctx->xa, index, file) {
+ struct io_ring_ctx *ctx = file->private_data;
+
+ if (ctx->flags & IORING_SETUP_SQPOLL) {
+ struct io_uring_task *__tctx = ctx->sqo_task->io_uring;
+
+ inflight += percpu_counter_sum(&__tctx->inflight);
+ }
+ }
+
+ return inflight;
}
/*
s64 inflight;
/* make sure overflow events are dropped */
- tctx->in_idle = true;
+ atomic_inc(&tctx->in_idle);
do {
/* read completions before cancelations */
- inflight = percpu_counter_sum(&tctx->inflight);
+ inflight = tctx_inflight(tctx);
if (!inflight)
break;
__io_uring_files_cancel(NULL);
* If we've seen completions, retry. This avoids a race where
* a completion comes in before we did prepare_to_wait().
*/
- if (inflight != percpu_counter_sum(&tctx->inflight))
+ if (inflight != tctx_inflight(tctx))
continue;
schedule();
} while (1);
finish_wait(&tctx->wait, &wait);
- tctx->in_idle = false;
+ atomic_dec(&tctx->in_idle);
}
static int io_uring_flush(struct file *file, void *data)
io_sqpoll_wait_sq(ctx);
submitted = to_submit;
} else if (to_submit) {
- ret = io_uring_add_task_file(f.file);
+ ret = io_uring_add_task_file(ctx, f.file);
if (unlikely(ret))
goto out;
mutex_lock(&ctx->uring_lock);
#if defined(CONFIG_UNIX)
ctx->ring_sock->file = file;
#endif
- if (unlikely(io_uring_add_task_file(file))) {
+ if (unlikely(io_uring_add_task_file(ctx, file))) {
file = ERR_PTR(-ENOMEM);
goto err_fd;
}