io_uring: fix __tctx_task_work() ctx race

There is an unlikely but possible race using a freed context. That's
because req->task_work.func() can free a request, but we won't
necessarily find a completion in submit_state.comp and so all ctx refs
may be put by the time we do mutex_lock(&ctx->uring_ctx);

There are several reasons why it can miss going through
submit_state.comp: 1) req->task_work.func() didn't complete it itself,
but punted to iowq (e.g. reissue) and it got freed later, or a similar
situation with it overflowing and getting flushed by someone else, or
being submitted to IRQ completion, 2) As we don't hold the uring_lock,
someone else can do io_submit_flush_completions() and put our ref.
3) Bugs and code obscurities, e.g. failing to propagate issue_flags
properly.

One example is as follows

  CPU1                                  |  CPU2
=======================================================================
@req->task_work.func()                  |
  -> @req overflwed,                    |
     so submit_state.comp,nr==0         |
                                        | flush overflows, and free @req
                                        | ctx refs == 0, free it
ctx is dead, but we do                  |
	lock + flush + unlock           |

So take a ctx reference for each new ctx we see in __tctx_task_work(),
and do release it until we do all our flushing.

Fixes: 65453d1e

 ("io_uring: enable req cache for task_work items")
Reported-by:  <syzbot+a157ac7c03a56397f553@syzkaller.appspotmail.com>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
[axboe: fold in my one-liner and fix ref mismatch]
Signed-off-by: Jens Axboe <axboe@kernel.dk>