From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from smtp.kernel.org (aws-us-west-2-korg-mail-1.web.codeaurora.org [10.30.226.201]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.subspace.kernel.org (Postfix) with ESMTPS id 96517241667 for ; Thu, 17 Jul 2025 23:26:36 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=10.30.226.201 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1752794799; cv=none; b=AyhR5rJaeTAytNWVWS3XxkKIytlNZt3qJc8hm3J/nQLfPc7JwJ1ohTzrxAFGcQT65sq90gjqkerPT/pl6zlQk5h9ukBJBsibCQ8kY8Kp+bgkzjDPFD8aDIH8j6A9ef9Jmq2/szH1aPyleEFtTB1PB6WmbnhMpFXcgb1dorRdUEw= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1752794799; c=relaxed/simple; bh=AEYPHnhz1GLfPcMcRaLMGL99irRtb/YWuN1tn/K3BsA=; h=Date:Subject:From:To:Cc:Message-ID:In-Reply-To:References: MIME-Version:Content-Type; b=D0/Zn44P+JE3584LSJ06KQtH8K/R1/P+2mj9fgMHIguVO1O032gu9rY0PhkM8eKvTLRj3QXfulhDftfDqZImKgo9r3EV976ijjz/iEiGcYC9P3MUHsYjlORzqxS7O2D/7mfGkj97myaw/WUhy3frS6QQ5Kmu6AjlxmwSuUL2s3c= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b=YNBTF2VD; arc=none smtp.client-ip=10.30.226.201 Authentication-Results: smtp.subspace.kernel.org; dkim=pass (2048-bit key) header.d=kernel.org header.i=@kernel.org header.b="YNBTF2VD" Received: by smtp.kernel.org (Postfix) with ESMTPSA id 13EC5C4CEE3; Thu, 17 Jul 2025 23:26:36 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=kernel.org; s=k20201202; t=1752794796; bh=AEYPHnhz1GLfPcMcRaLMGL99irRtb/YWuN1tn/K3BsA=; h=Date:Subject:From:To:Cc:In-Reply-To:References:From; b=YNBTF2VD919Hv9vIT0TrE1jPOl6ubiJ3xzKCQhZrXpZwI3iRwyabeJAM3eX3PEdE3 tjdJ3mLdRciXtnz0TqPInz3gafZu7NnT+OFK645n3Iv1jcfn7m6Xr5S7G/cRpE+Y8I rVIQ4lV/KmJ832iMiMq7iep9bFo1mIFpfR22jh9nZtedgIimopYOnbicALwHQS6G2f OVFiKJifFWBYPKGKWJ8MByJQBOfHIZ0avdqieoFUcOJkJ9ZZNOnRJnL8AW5WyjQiwJ lUaROjMEfGnGeonXnTLfouLQEwOBwYrOhXBxjg6JIXrqy0N/YBO4f1NTtdgMBvevWd a+k0hfzGBoHkA== Date: Thu, 17 Jul 2025 16:26:35 -0700 Subject: [PATCH 1/7] fuse: fix livelock in synchronous file put from fuseblk workers From: "Darrick J. Wong" To: djwong@kernel.org Cc: linux-fsdevel@vger.kernel.org, neal@gompa.dev, John@groves.net, miklos@szeredi.hu, bernd@bsbernd.com, joannelkoong@gmail.com Message-ID: <175279449480.710975.8767104113054819157.stgit@frogsfrogsfrogs> In-Reply-To: <175279449418.710975.17923641852675480305.stgit@frogsfrogsfrogs> References: <175279449418.710975.17923641852675480305.stgit@frogsfrogsfrogs> Precedence: bulk X-Mailing-List: linux-fsdevel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 7bit From: Darrick J. Wong I observed a hang when running generic/323 against a fuseblk server. This test opens a file, initiates a lot of AIO writes to that file descriptor, and closes the file descriptor before the writes complete. Unsurprisingly, the AIO exerciser threads are mostly stuck waiting for responses from the fuseblk server: # cat /proc/372265/task/372313/stack [<0>] request_wait_answer+0x1fe/0x2a0 [fuse] [<0>] __fuse_simple_request+0xd3/0x2b0 [fuse] [<0>] fuse_do_getattr+0xfc/0x1f0 [fuse] [<0>] fuse_file_read_iter+0xbe/0x1c0 [fuse] [<0>] aio_read+0x130/0x1e0 [<0>] io_submit_one+0x542/0x860 [<0>] __x64_sys_io_submit+0x98/0x1a0 [<0>] do_syscall_64+0x37/0xf0 [<0>] entry_SYSCALL_64_after_hwframe+0x4b/0x53 But the /weird/ part is that the fuseblk server threads are waiting for responses from itself: # cat /proc/372210/task/372232/stack [<0>] request_wait_answer+0x1fe/0x2a0 [fuse] [<0>] __fuse_simple_request+0xd3/0x2b0 [fuse] [<0>] fuse_file_put+0x9a/0xd0 [fuse] [<0>] fuse_release+0x36/0x50 [fuse] [<0>] __fput+0xec/0x2b0 [<0>] task_work_run+0x55/0x90 [<0>] syscall_exit_to_user_mode+0xe9/0x100 [<0>] do_syscall_64+0x43/0xf0 [<0>] entry_SYSCALL_64_after_hwframe+0x4b/0x53 The fuseblk server is fuse2fs so there's nothing all that exciting in the server itself. So why is the fuse server calling fuse_file_put? The commit message for the fstest sheds some light on that: "By closing the file descriptor before calling io_destroy, you pretty much guarantee that the last put on the ioctx will be done in interrupt context (during I/O completion). Aha. AIO fgets a new struct file from the fd when it queues the ioctx. The completion of the FUSE_WRITE command from userspace causes the fuse server to call the AIO completion function. The completion puts the struct file, queuing a delayed fput to the fuse server task. When the fuse server task returns to userspace, it has to run the delayed fput, which in the case of a fuseblk server, it does synchronously. Sending the FUSE_RELEASE command sychronously from fuse server threads is a bad idea because a client program can initiate enough simultaneous AIOs such that all the fuse server threads end up in delayed_fput, and now there aren't any threads left to handle the queued fuse commands. Fix this by only using synchronous fputs for fuseblk servers if the process doesn't have PF_LOCAL_THROTTLE. Hopefully the fuseblk server had the good sense to call PR_SET_IO_FLUSHER to mark itself as a filesystem server. Signed-off-by: "Darrick J. Wong" --- fs/fuse/file.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) diff --git a/fs/fuse/file.c b/fs/fuse/file.c index 47006d0753f1cd..ee79cb7bc05805 100644 --- a/fs/fuse/file.c +++ b/fs/fuse/file.c @@ -355,8 +355,16 @@ void fuse_file_release(struct inode *inode, struct fuse_file *ff, * Make the release synchronous if this is a fuseblk mount, * synchronous RELEASE is allowed (and desirable) in this case * because the server can be trusted not to screw up. + * + * If we're a LOCAL_THROTTLE thread, use the asynchronous put + * because the current thread might be a fuse server. This can + * happen if a process starts some aio and closes the fd before + * the aio completes. Since aio takes its own ref to the file, + * the IO completion has to drop the ref, which is how the fuse + * server can end up closing its own clients' files. */ - fuse_file_put(ff, ff->fm->fc->destroy); + fuse_file_put(ff, ff->fm->fc->destroy && + (current->flags & PF_LOCAL_THROTTLE) == 0); } void fuse_release_common(struct file *file, bool isdir)