On 4/12/17 10:15 AM, Christoph Hellwig wrote: > On Tue, Apr 11, 2017 at 11:48:19AM -0500, Eric Sandeen wrote: >> As an experiment, though, if you want to play, it might be interesting to >> mkfs.xfs it with a 4096 byte sector size, and see if that makes it happier. >> By default, xfs is doing metadata IO in 512 chunks, something other filesystems >> won't do by default. >> >> I guess you don't know how to provoke the corruption, though, to be able >> to run that test reliably... > > Btw, it might be a good idea to move to 4k sector size as the default, > on pretty much any modern hardware sector sizes are 4k or larger > internally, and 512 byte writes will always involve read-modify-write > cycles. And unlike SATA or SCSI disks NVMe doesn't have a physical > block size attribute, so we can't even look at that. Is it safe to do that on a device that /actually/ has only 512 sectors? I /think/ Brian's tear detection helps, but </handwave> is it legit to do metadata IO larger than the fundamental IO size of the storage? -Eric -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
