On Sat, Mar 26, 2011 at 2:17 AM, Andrei Warkentin <andreiw@xxxxxxxxxxxx> wrote: > Hi Arnd, > > On Fri, Mar 25, 2011 at 10:14 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote: >> On Thursday 24 March 2011, Andrei Warkentin wrote: >>> >>> Beyond REQ_FUA/REQ_META, this was meant to be used by a following patch that aimed >>> to reduce write amplification issues in cards employing a small (usually flash page-sized) >>> buffer and a large (usually erase-block sized) buffer, at the expense of performance. >> >> Looks good to me, but I don't really understand some of the block layer >> specifics here. One question: >> How does this work when you have a flush that does not directly follow >> a REQ_FUA or REQ_META request? I would assume that we still need to >> flush in some way, which you don't seem to do here. > > In much the same way as before - REQ_FLUSH does nothing as it did > nothing before. The only reason I'm handing REQ_FLUSH, is because you > can't register for REQ_FUA without registering for REQ_FLUSH, which > does make sense - after all, if REQ_FUA is used to ensure data is > committed immediately, then by default it isn't, and you need a flush > operation if you want to commit the data at a particular checkpoint > You are correct in stating that it is a likely issue, but it is an > issue even without this patch. Unfortunately, there is nothing (that > I saw) in the latest JEDEC MMC specification that says how a card's > internal buffers might be forced to be flushed to non-volatile > storage. Clearly, any recent card which tries to play optimization > games in order to reduce write amplification and flash wearing is > going to have an internal buffer, so the inability to handle REQ_FLUSH > is already a potential issue for these cards. > > Looking at the 4.41 JEDEC spec, CMD0, CMD15 and hardware reset are all > non-ways to ensure flush. In fact they will terminate the programming > operation resulting in an unknown state. > > I'm working with a flash vendor to see if I missed something or if > there is some way to flush internal buffers, and if that way is > generic enough to apply to all cards. > > A > Sorry, I totally forgot to add. REQ_FLUSH means - flush all write back caches internal to the disk hardware. A -- To unsubscribe from this list: send the line "unsubscribe linux-mmc" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
