On Tue, 2004-08-24 at 09:31, Kenneth Porter wrote: > What is barrier support? ok it's like this: nowadays storage (think "disks" but also "raid arrays") can handle more than 1 outstanding IO at a time. For example for scsi it's not uncommon to have 250 IO's outstanding to one disk at a time. This allows the disk to optimize the IO, for example by doing commands that are sequential on the platters as one big operation. The result of these optimisations is that IO's can and will happen in a different order than the operating system has submitted them. This would be a problem for journalling filesystems, because for example the journal update that gets submitted before the actual metadata update can in practice happen AFTER the actual metadata update. If there is a power failure this would lead to a broken journal/metadata combination (normally you recover from a metadata problem by replaying the journal.... except the journal is out of date due to the reorder). Currently, journalling filesystems deal with this by waiting for the journal IO to entirely complete before submitting the metadata IO, eg if it's already on the platter there is no way to reorder ;) Barriers (or in scsi terms "ordered tags") basically tell the disk "don't reorder past THIS point". In principle this alleviates the need for the journalling filesystem to wait on the first IO before submitting the second IO (the metadata one). In practice it's sort of an unknown of how well actual disks honor this... and if they do, it's mostly unknown how expensive it is for the disks to do such tag, so it's unsure if there is an actual performance gain. At least now there's the infrastructure so that people can play with it; enabling it by default from the start sounds like a bad idea. Of course we'll be looking at this, but it's a longer term project.
Description: This is a digitally signed message part