io_uring: enable LOOKUP_CACHED path resolution for filename lookups

Instead of being pessimistic and assume that path lookup will block, use
LOOKUP_CACHED to attempt just a cached lookup. This ensures that the
fast path is always done inline, and we only punt to async context if
IO is needed to satisfy the lookup.

For forced nonblock open attempts, mark the file O_NONBLOCK over the
actual ->open() call as well. We can safely clear this again before
doing fd_install(), so it'll never be user visible that we fiddled with
it.

This greatly improves the performance of file open where the dentry is
already cached:

ached		5.10-git	5.10-git+LOOKUP_CACHED	Speedup
---------------------------------------------------------------
33%		1,014,975	900,474			1.1x
89%		 545,466	292,937			1.9x
100%		 435,636	151,475			2.9x

The more cache hot we are, the faster the inline LOOKUP_CACHED
optimization helps. This is unsurprising and expected, as a thread
offload becomes a more dominant part of the total overhead. If we look
at io_uring tracing, doing an IORING_OP_OPENAT on a file that isn't in
the dentry cache will yield:

275.550481: io_uring_create: ring 00000000ddda6278, fd 3 sq size 8, cq size 16, flags 0
275.550491: io_uring_submit_sqe: ring 00000000ddda6278, op 18, data 0x0, non block 1, sq_thread 0
275.550498: io_uring_queue_async_work: ring 00000000ddda6278, request 00000000c0267d17, flags 69760, normal queue, work 000000003d683991
275.550502: io_uring_cqring_wait: ring 00000000ddda6278, min_events 1
275.550556: io_uring_complete: ring 00000000ddda6278, user_data 0x0, result 4

which shows a failed nonblock lookup, then punt to worker, and then we
complete with fd == 4. This takes 65 usec in total. Re-running the same
test case again:

281.253956: io_uring_create: ring 0000000008207252, fd 3 sq size 8, cq size 16, flags 0
281.253967: io_uring_submit_sqe: ring 0000000008207252, op 18, data 0x0, non block 1, sq_thread 0
281.253973: io_uring_complete: ring 0000000008207252, user_data 0x0, result 4

shows the same request completing inline, also returning fd == 4. This
takes 6 usec.
Signed-off-by: Jens Axboe <axboe@kernel.dk>

fs/io_uring.c

@@ -489,7 +489,6 @@ struct io_sr_msg {
 struct io_open {
 	struct file			*file;
 	int				dfd;
-	bool				ignore_nonblock;
 	struct filename			*filename;
 	struct open_how			how;
 	unsigned long			nofile;
@@ -4054,7 +4053,6 @@ static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe
 		return ret;
 	}
 	req->open.nofile = rlimit(RLIMIT_NOFILE);
-	req->open.ignore_nonblock = false;
 	req->flags |= REQ_F_NEED_CLEANUP;
 	return 0;
 }
@@ -4096,39 +4094,48 @@ static int io_openat2(struct io_kiocb *req, bool force_nonblock)
 {
 	struct open_flags op;
 	struct file *file;
+	bool nonblock_set;
+	bool resolve_nonblock;
 	int ret;
 
-	if (force_nonblock && !req->open.ignore_nonblock)
-		return -EAGAIN;
-
 	ret = build_open_flags(&req->open.how, &op);
 	if (ret)
 		goto err;
+	nonblock_set = op.open_flag & O_NONBLOCK;
+	resolve_nonblock = req->open.how.resolve & RESOLVE_CACHED;
+	if (force_nonblock) {
+		/*
+		 * Don't bother trying for O_TRUNC, O_CREAT, or O_TMPFILE open,
+		 * it'll always -EAGAIN
+		 */
+		if (req->open.how.flags & (O_TRUNC | O_CREAT | O_TMPFILE))
+			return -EAGAIN;
+		op.lookup_flags |= LOOKUP_CACHED;
+		op.open_flag |= O_NONBLOCK;
+	}
 
 	ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile);
 	if (ret < 0)
 		goto err;
 
 	file = do_filp_open(req->open.dfd, req->open.filename, &op);
+	/* only retry if RESOLVE_CACHED wasn't already set by application */
+	if ((!resolve_nonblock && force_nonblock) && file == ERR_PTR(-EAGAIN)) {
+		/*
+		 * We could hang on to this 'fd', but seems like marginal
+		 * gain for something that is now known to be a slower path.
+		 * So just put it, and we'll get a new one when we retry.
+		 */
+		put_unused_fd(ret);
+		return -EAGAIN;
+	}
+
 	if (IS_ERR(file)) {
 		put_unused_fd(ret);
 		ret = PTR_ERR(file);
-		/*
-		 * A work-around to ensure that /proc/self works that way
-		 * that it should - if we get -EOPNOTSUPP back, then assume
-		 * that proc_self_get_link() failed us because we're in async
-		 * context. We should be safe to retry this from the task
-		 * itself with force_nonblock == false set, as it should not
-		 * block on lookup. Would be nice to know this upfront and
-		 * avoid the async dance, but doesn't seem feasible.
-		 */
-		if (ret == -EOPNOTSUPP && io_wq_current_is_worker()) {
-			req->open.ignore_nonblock = true;
-			refcount_inc(&req->refs);
-			io_req_task_queue(req);
-			return 0;
-		}
 	} else {
+		if (force_nonblock && !nonblock_set)
+			file->f_flags &= ~O_NONBLOCK;
 		fsnotify_open(file);
 		fd_install(ret, file);
 	}