On 14/11/2024 13:52, Christian Brauner wrote: > I think you need at least something like the following completely > untested draft on top: > > - the pidfs_finish_open_by_handle_at() is somewhat of a clutch to handle > thread vs thread-group pidfds but it works. > > - In contrast to pidfd_open() that uses dentry_open() to create a pidfd > open_by_handle_at() uses file_open_root(). That's overall fine but > makes pidfds subject to security hooks which they aren't via > pidfd_open(). It also necessitats pidfs_finish_open_by_handle_at(). > There's probably other solutions I'm not currently seeing. These two concerns combined with the special flag make me wonder if pidfs is so much of a special snowflake we should just special case it up front and skip all of the shared handle decode logic? > - The exportfs_decode_fh_raw() call that's used to decode the pidfd is > passed vfs_dentry_acceptable() as acceptability callback. For pidfds > we don't need any of that functionality and we don't need any of the > disconnected dentry handling logic. So the easiest way to fix that is > to rely on EXPORT_OP_UNRESTRICTED_OPEN to skip everything. That in > turns means the only acceptability we have is the nop->fh_to_dentry() > callback for pidfs. With the current logic we go exportfs_decode_fh_raw(...) -> find_acceptable_alias(result) -> vfs_dentry_acceptable(context, result). vfs_dentry_acceptable immediately returns 1 if ctx->flags is 0, which will always be the case if EXPORT_OP_UNRESTRICTED_OPEN was set, so we immediately fall back out of the call tree and return result. So I'm not 100% sure we actually need this special case but I'm not opposed. > - This all really needs rigorous selftests before we can even think of > merging any of this.
