On 2021/03/01 21:52, Guido Trentalancia wrote: > On Mon, 01/03/2021 at 12.42 +0000, Damien Le Moal wrote: >> I checked the standards again. It turns out that SYNCHRONIZE CACHE >> command is >> optional in SBC... Aouch. Got so used to have that one on any drive >> that I >> thought it was mandatory. >> >> Well, it certainly is mandatory if the drive has a write cache, which >> seems to >> be the case for you. >> >> The problem with your patch though is that you disable write caching >> when you >> see an ILLEGAL REQUEST/INVALID OPCODE termination of synchronize >> cache. Which >> means that the drive was already used, written too and the cache has >> likely >> dirty data and I do not see any method to guarantee that that data >> makes it to >> persistent media before shutdown. Imagine if that was the synchronize >> cache sent >> before shutdown. > > As already said, I have tested the patch for over a year now and I have > never experienced the problem that you are foreseeing ! That may be true for your specific drive, but since we are in uncharted (non-standard) territory here, we cannot say that this will hold for all such weird drives out there. > The current alternative is data corruption each time that the drive is > mounted and the inability to use it. > > So, the patch is the way forward for using such drives plug and play > without cumbersome configuration such as disabling the write cache, > which advanced users can always make. > >> So the only reasonable solution for such drive is to use it with >> write cache >> disabled from the start. >> down. > > It works well even with write caching enabled. > > If you have an alternative patch to propose, go ahead otherwise I would > push for getting this merged and sorting out the issue for good. As mentioned, the alternative is a udev rule to disable write cache. Or if the drive supports that, permanently save WCE=0 setting in the drive config so that it always come up with write cache disabled. No kernel patch needed, and you will note that this is also exactly the same as what your patch does, without waiting for an error. > > Guido > -- Damien Le Moal Western Digital Research
