On Wed, 28 Feb 2024 at 22:30, Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
>
> Non append, non extending buffered writes can now avoid taking the inode
> lock.
I think this is buggy.
I think you still need to take the inode lock *shared* for the writes,
because otherwise you can have somebody else that truncates the file
and now you will do a write past the end of the size of the file. That
will cause a lot of issues.
So it's not a "inode_lock or not" situation. I think it's a
"inode_lock vs inode_locks_shared" situation.
Note that the reading side isn't all that critical - if a read races
with a truncate, at worst it will read some zeroes because we used the
old length and the page cache got cleared in the meantime.
But the writing side ends up having actual consistency issues on disk.
You don't want to have a truncate that removes the pages past the end
of the new size and clears the end of the new last page, and race with
another write that used the old size and *thought* it was writing to
the middle of the file, but is now actually accessing a folio that is
past the end of the whole file and writing to it.
There may be some reason that I'm missing that would make this a
non-issue, but I really think you want to get the inode lock at least
shared for the duration of the write.
Also note that for similar reasons, you can't just look at "will I
extend the file" and take the lock non-shared. No, in order to
actually trust the size, you need to *hold* the lock, so the logic
needs to be something like
- take the lock exclusively if O_APPEND or if it *looks* like you
might extend the file size.
- otherwise, take the shared lock, and THEN RE-CHECK. The file size
might have changed, so now you need to double-check that you're really
not going to extend the size of the file, and if you are, you need to
go back and take the inode lock exclusively after all.
Again - I haven't thought a ton about this, so maybe there's some
trick to it, but the above is what my naive thinking says the rule has
to be. Writes are different from reads.
Linus
