From: Christoph Hellwig Subject: Re: fix block device size update serialization v2 Date: Thu, 27 Aug 2020 09:47:58 +0200 Message-ID: <20200827074758.GA8009@lst.de> References: <20200823091043.2600261-1-hch@lst.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Content-Disposition: inline In-Reply-To: <20200823091043.2600261-1-hch@lst.de> Sender: linux-block-owner@vger.kernel.org To: Jens Axboe Cc: Justin Sanders , Josef Bacik , Xianting Tian , linux-block@vger.kernel.org, dm-devel@redhat.com, Stefan Haberland , Jan Hoeppner , linux-kernel@vger.kernel.org, nbd@other.debian.org, linux-nvme@lists.infradead.org, linux-s390@vger.kernel.org List-Id: dm-devel.ids Jens, can you consider this for 5.9? It reliably fixes the reported hangs with nvme hotremoval that we've had for a few releases. On Sun, Aug 23, 2020 at 11:10:40AM +0200, Christoph Hellwig wrote: > Hi Jens, > > this series fixes how we update i_size for the block device inodes (and > thus the block device). Different helpers use two different locks > (bd_mutex and i_rwsem) to protect the update, and it appears device > mapper uses yet another internal lock. A lot of the drivers do the > update handcrafted in often crufty ways. And in addition to that mess > it turns out that the "main" lock, bd_mutex is pretty dead lock prone > vs other spots in the block layer that acquire it during revalidation > operations, as reported by Xianting. > > Fix all that by adding a dedicated spinlock just for the size updates. > > Changes since v1: > - don't call __invalidate_device under the new spinlock > - don't call into the file system code from the nvme removal code ---end quoted text---