Messages in this thread | | | Date | Tue, 6 Jan 1998 21:38:33 -0500 (EST) | From | linux kernel account <> | Subject | Re: lack of raw disk devices |
| |
On Tue, 6 Jan 1998 ralf@uni-koblenz.de wrote:
> On Sun, Jan 04, 1998 at 11:48:50AM -0500, linux kernel account wrote: > > > On Sun, 4 Jan 1998, Harald Milz wrote: > > > > > Marty Leisner (leisner@sdsp.mc.xerox.com) wrote: > > > > I understand linux doesn't have raw devices because > > > > we don't need it. > > > > > > We actually do. A couple of database product make use of raw devices if > > > they exist, e.g. Oracle. > > > > They are just looking for excuses.. They could use whats already there.. > > They just would have to deal with the fact that the OS is going to cache > > things in ways they dont like.. > > In a way that makes error recovery impossible for the database.
You can use O_SYNC to do synced writes.. That only leaves read problems.
> > They want to have complete input/output cach control.. Yes, that would be > > important if Linux was their prime target.. > > Customer calls Oracle (Sybase etc.): ``Your friggin database engine fucked > my entire 1TB database. I hope your competition does better ...'' > > Cite manual of xy-db ... ``Linux: We're sorry, if your box should crash, your > database might be fried'' ... > > Wouldn't that be good promotion?
Better then nothing, at least if they explain why.
> > > But they could just say 'the > > Linux version is beta/not as good and stable because of it' and when they > > decide that Linux was a platform they really like: They could make a > > friggen kernel module to provide a raw interface. > > That would be a horrible hack that'd break once a new kernelpatch is > released. And that's still the least problem.
Most people take what their distribution gives them.. And most commercial places would only dare support that single kernel.. I challenge you to by Caldara Linux and compile your own newer kernel and get support..
> This is not a vote for a blind implementation of raw devices. Raw devices > have their problems by design, mainly because of a lack of communication > between the buffercache and the userspace cache in the db engine. > > Ralf >
| |