site_archiver@lists.apple.com Delivered-To: darwin-dev@lists.apple.com Make, I want to point out following: this discussion is simlar to "is it possible to develop a high-performance OS in C++" I do not think it's relevant to darwin-dev any longer. Stacking is a very powerful obstraction. It enables a more flexible construction of a data flow. In some instances this comes in very handy. If a cost can be kept under 1-2% per layer for a random access, it is usually worth it. However to take advantage of a stackable design one has to explicitly exploit its benefits. I have some data to support my opinions. Please contact me directly if you would like to talk about this. Igor. -----Original Message----- From: Mike Smith <msmith@freebsd.org> To: Igor Shmukler <shmukler@mail.ru> Date: Thu, 18 Nov 2004 21:42:49 -0800 Subject: Re: Re[2]: Intercepting file system calls (read, write, open, close, etc)

On Nov 18, 2004, at 6:28 PM, Igor Shmukler wrote:

...

...

...

...

1. Try VFS Stacking ? Or over-riding V-node ops ?

Was support for VFS stacking removed? I was warned of such a year or two ago, and stopped my VFS efforts at that point (also trying to intercept file system calls).

Stacking filesystems in the BSD model doesn't work very well. It gets expensive very quickly, and locking can become a nightmare. It's also often not really what you want; the above is a good example of that.

The above being - ?

"also trying to intercept file system calls"

...

SunOS is a proof that BSD model and stackable VFS do work well together.

Opinions on that differ to a considerable degree.

...

The reason locking in an operating system such as a BSD would problematic is because VFS is very complex. It needed flattening back in 1998 right after VM became stable. Now to support MP kernel locking is more complex than ever. That does not mean that things ought to be this way. Implementing a stackable FS is relatively straightforward if you start from the right place - vnode interface. If done correctly stackable does not create any problems with locking that would not be present in a leaf design.

These are interesting assertions, but without any real application in the current context. I would caution anyone else reading this that your opinions are certainly not widely held.

In particular, fan-in, fan-out and partial unwinds are all necessary for stacking filesystems (if you don't want unwinds, you need reservations which are arguably worse). Locking these is painful at best, and nesting vnodes is never not expensive.

= Mike

http://Mail.Ru - лучшая почта с неограниченным объемом почтового ящика! _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev@lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.appl... This email sent to site_archiver@lists.apple.com