On Mon, 20 Mar 2023 at 07:25, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > > On Thu, Mar 16, 2023 at 09:12:35PM +0200, Adrian Hunter wrote: > > Historically file systems have assumed that sectors are updated > > atomically i.e. there is never a sector with a mixture of > > old and new data. > > Yes. Not just file systems, but also all kinds of applications. > > > The eMMC spec does not guarantee that, > > except for the eMMC "reliable write" operation. > > Neither to ATA or SCSI, but applications and file systems always very > much expected it, so withou it storage devices would be considered > fault. Only NVMe actually finally made it part of the standard. Even if the standard doesn't say, it's perfectly possible that the storage device implements it. Hence, $subject patch isn't changing anything in regards to REQ_FUA, unless the eMMC device/vendor has agreed to this (via the MMC card quirks). > > > So the paragraph > > above is informing the potential benefit of reliable write instead > > of cache flush. > > But these are completely separate issue. Torn writes are completely > unrelated to cache flushes. You can indeed work around torn writes > by checksums, but not the lack of cache flushes or vice versa. It's not a separate issue for eMMC. Please read the complete commit message for further clarifications in this regard. > > > > Note, it is not that eMMC cannot avoid torn sectors, it is that > > the specification does not mandate that they do. > > If devices tear writes it will break not only various file systems, > but more importantly applications, at least on file systems without > data checksum (aka all except for btrfs, and even that has a nodatacsum > option). Yes, you are correct. Again, the card quirk (as suggested in $subject patch) helps us to manage eMMC cards in different ways. We should not avoid REQ_FUA (reliable writes) for an eMMC that actually needs it. > > > However, the issue has been raised that reliable write is not > > needed to provide sufficient assurance of data integrity, and that > > in fact, cache flush can be used instead and perform better. > > It does not. Can you please elaborate on this? The tests we have done so far indicate that performance differs based upon what eMMC we are using, for example. Kind regards Uffe
