On Fri, Jan 10, 2020 at 02:20:55PM +0800, Ian Kent wrote: > Yeah, autofs ->d_automount() doesn't return -EISDIR, by the time > we get there it's not relevant any more, so that check looks > redundant. I'm not aware of any other fs automount implementation > that needs that EISDIR pass-thru function. > > I didn't notice it at the time of the merge, sorry about that. > > While we're at it that: > if (!path->dentry->d_op || !path->dentry->d_op->d_automount) > return -EREMOTE; > > at the top of follow_automount() isn't going to be be relevant > for autofs because ->d_automount() really must always be defined > for it. > > But, at the time of the merge, I didn't object to it because > there were (are) other file systems that use the VFS automount > function which may accidentally not define the method. OK... > > Unfortunately, there are other interesting questions related to > > autofs-specific bits (->d_manage()) and the timezone-related fun > > is, of course, still there. I hope to sort that out today or > > tomorrow, at least enough to do a reasonable set of backportable > > fixes to put in front of follow_managed()/step_into() queue. > > Oh, well... > > Yeah, I know it slows you down but I kink-off like having a chance Nice typo, that ;-) > to look at what's going and think about your questions before trying > to answer them, rather than replying prematurely, as I usually do ... > > It's been a bit of a busy day so far but I'm getting to look into > the questions you've asked. Here's a bit more of those (I might've missed some of your replies on IRC; my apologies if that's the case): 1) AFAICS, -EISDIR from ->d_manage() actually means "don't even try ->d_automount() here". If its effect can be delayed until the decision to call ->d_automount(), the things seem to get simpler. Is it ever returned in situation when the sucker _is_ overmounted? 2) can autofs_d_automount() ever be called for a daemon? Looks like it shouldn't be... 3) is _anything_ besides root directory ever created in direct autofs superblocks by anyone? If not, why does autofs_lookup() even bother to do anything there? IOW, why not have it return ERR_PTR(-ENOENT) immediately for direct ones? Or am I missing something and it is, in fact, possible to have the daemon create something in those? 4) Symlinks look like they should qualify for parent being non-empty; at least autofs_d_manage() seems to think so (simple_empty() use). So shouldn't we remove the trap from its parent on symlink/restore on unlink if parent gets empty? For version 4 or earlier, that is. Or is it simply that daemon only creates symlinks in root directory? Anyway, intermediate state of the series is in #work.namei right now, and some _very_ interesting possibilities open up. It definitely needs more massage around __follow_mount_rcu() (as it is, the fastpath in there is still too twisted). Said that * call graph is less convoluted * follow_managed() calls are folded into step_into(). Interface: int step_into(nd, flags, dentry, inode, seq), with inode/seq used only if we are in RCU mode. * ".." still doesn't use that; it probably ought to. * lookup_fast() doesn't take path - nd, &inode, &seq and returns dentry * lookup_open() and fs/namei.c:atomic_open() get similar treatment - don't take path, return dentry. * calls of follow_managed()/step_into() combination returning 1 are always followed by get_link(), and very shortly, at that. So much that we can realistically merge pick_link() (in the end of step_into()) with get_link(). That merge is NOT done in this branch yet. The last one promises to get rid of a rather unpleasant group of calling conventions. Right now we have several functions (step_into()/ walk_component()/lookup_last()/do_last()) with the following calling conventions: -E... => error 0 => non-symlink or symlink not followed; nd->path points to it 1 => picked a symlink to follow; its mount/dentry/seq has been pushed on nd->stack[]; its inode is stashed into nd->link_inode for subsequent get_link() to pick. nd->path is left unchanged. That way all of those become ERR_PTR(-E...) => error NULL => non-symlink, symlink not followed or a pure jump (bare "/" or procfs ones); nd->path points to where we end up string => symlink being followed; the sucker's pushed to stack, initial jump (if any) has been handled and the string returned is what we need to traverse. IMO it's less arbitrary that way. More importantly, the separation between step_into() committing to symlink traversal and (inevitably following) get_link() is gone - it's one operation after that change. No nd->link_inode either - it's only needed to carry the information from pick_link() to the next get_link(). Loops turn into while (!(err = link_path_walk(nd, s)) && (s = lookup_last(nd)) != NULL) ; and while (!(err = link_path_walk(nd, s)) && (s = do_last(nd, file, op)) != NULL) ; trailing_symlink() goes away (folded into pick_link()/get_link() combo, conditional upon nd->depth at the entry). And in link_path_walk() we'll have if (unlikely(!*name)) { /* pathname body, done */ if (!nd->depth) return 0; name = nd->stack[nd->depth - 1].name; /* trailing symlink, done */ if (!name) return 0; /* last component of nested symlink */ s = walk_component(nd, WALK_FOLLOW); } else { /* not the last component */ s = walk_component(nd, WALK_FOLLOW | WALK_MORE); } if (s) { if (IS_ERR(s)) return PTR_ERR(s); /* a symlink to follow */ nd->stack[nd->depth - 1].name = name; name = s; continue; } Anyway, before I try that one I'm going to fold path_openat2() into that series - that step is definitely going to require some massage there; it's too close to get_link() changes done in Aleksa's series. If we do that, we get a single primitive for "here's the result of lookup; traverse mounts and either move into the result or, if it's a symlink that needs to be traversed, start the symlink traversal - jump into the base position for it (if needed) and return the pathname that needs to be handled". As it is, mainline has that logics spread over about a dozen locations... Diffstat at the moment: fs/autofs/dev-ioctl.c | 6 +- fs/internal.h | 1 - fs/namei.c | 460 ++++++++++++++------------------------------------ fs/namespace.c | 97 +++++++---- fs/nfs/nfstrace.h | 2 - fs/open.c | 4 +- include/linux/namei.h | 3 +- 7 files changed, 197 insertions(+), 376 deletions(-) In the current form the sucker appears to work (so far - about 30% into the usual xfstests run) without visible slowdowns...