From: Christoph Hellwig Date: Wed, 15 Jan 2020 07:54:06 +0100 Subject: [Cluster-devel] RFC: hold i_rwsem until aio completes In-Reply-To: <20200114184707.GA10467@bombadil.infradead.org> References: <20200114161225.309792-1-hch@lst.de> <20200114184707.GA10467@bombadil.infradead.org> Message-ID: <20200115065406.GB21219@lst.de> List-Id: To: cluster-devel.redhat.com MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit On Tue, Jan 14, 2020 at 10:47:07AM -0800, Matthew Wilcox wrote: > It would be helpful if we could also use the same lockdep logic > for PageLocked. Again, it's a case where returning to userspace with > PageLock held is fine, because we're expecting an interrupt to come in > and drop the lock for us. Yes, this is a very typical pattern for I/O. Besides the page and buffer head bit locks it also applies to the semaphore in the xfs_buf structure and probably various other places that currently used hand crafted or legacy locking primitives to escape lockdep. > Perhaps the right answer is, from lockdep's point of view, to mark the > lock as being released at the point where we submit the I/O. Then > in the completion path release the lock without telling lockdep we > released it. That is similar to what the fsfreeze code does, but I don't think it is very optimal, as misses to track any dependencies after I/O submission, and at least some of the completions paths do take locks. But it might be a start.