On Fri, 8 Oct 1999, Manfred Spraul wrote:
> Alexander Viro wrote:
> >All stuff with truncate()/write() is racey right now,
> > AFAICS. The next thing I'm going to do is a (dumb) implementation of
> > blocking rwlocks. write() being reader and truncate() - writer.
>
> IMHO that's the wrong approach:
>
> 1) overlapping concurrent write operations violate Nfsv2.
Yes, that is serious.
> 2) the pipe code needs O_NONBLOCK characteristics, ie no synchonization
> is allowed. The 2.2 code uses an ugly hack: sys_write() acquires i_sem,
> and inside the pipe code the semaphore is dropped.
Huh? It's not a sys_write(), it's generic_file_write(). Pipes do not come
even close to that place. And I'm talking about 2.3 - 2.2 is _way_ too
different.
> 3) do not forget f_pos when you implement a synchonization: even
> concurrent reads are not allowed if both use the same file pointer.
Ditto. It's a separate issue.
> What about these ideas:
> a) rename i_sem (*), and add public functions such as
> inode_lock_exclusive(), inode_lock_truncate(), inode_lock_write() etc.
>
> Currently, everyone directly accesses the VFS synchonization (nfsd,
> drivers/char/tty_io.c; arch/*/kernel/sys*.c).
>
> b) the requirements of the VFS users differ completely, add flags:
> 0: synchonize nothing, eg pipes.
> I_FPOS_SYNC: synchonize f_pos access, nothing else
> I_FILE_SYNC: multiple read, non-overlapping write, truncate.
>
> Since it will be hidden by inode_lock_??(), it will be easy to optimize
> the implementation as needed.
>
> Eg I'm pretty sure that Stephen will not like it if raw-io is slowed
> down by a multiple reader-single writer synchonization.
Erm? Writer access is for truncate() and I don't see where the raw-io
comes into the picture.
