Re: sys_write() racy for multi-threaded append?

[Eric Dumazet <dada1@cosmosbay.com>]

Michael K. Edwards a écrit :
> On 3/8/07, Eric Dumazet <dada1@cosmosbay.com> wrote:
>> Absolutely not. We dont want to slow down kernel 'just in case a fool 
>> might
>> want to do crazy things'
> 
> Actually, I think it would make the kernel (negligibly) faster to bump
> f_pos before the vfs_write() call.  Unless fget_light sets fput_needed
> or the write doesn't complete cleanly, you won't have to touch the
> file table entry again after vfs_write() returns.  You can adjust
> vfs_write to grab f_dentry out of the file before going into
> do_sync_write.  do_sync_write is done with the struct file before it
> goes into the aio_write() loop.  Result: you probably save at least an
> L1 cache miss, unless the aio_write loop is so frugal with L1 cache
> that it doesn't manage to evict the struct file.
> 
> Patch to follow.

Dont even try, you *cannot* do that, without breaking the standards, or 
without a performance drop.

The only safe way would be to lock the file during the whole read()/write() 
syscall, and we dont want this (this would be more expensive than current)
Dont forget 'file' may be some sockets/tty/whatever, not a regular file.

Standards are saying :

If an error occurs, file pointer remains unchanged.

You cannot know for sure how many bytes will be written, since write() can 
returns a count that is different than buflen.

So you cannot update fpos before calling vfs_write()

About your L1 'miss', dont forget that multi-threaded apps are going to 
atomic_dec_and_test(&file->f_count) anyway when fput() is done at the end of 
syscall. And you were concerned about multi-threaded apps, didnt you ?