On Wed, Mar 7, 2018 at 10:43 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote: [...] >> >> >> >> Q: why do you set samedir: = false here? >> >> A: because other hardlink aliasses cannot follow a relative redirect, right? >> > >> > Right. If we create a hardlink later then it will need absolute redirect >> > if both dentries are not in same dir. >> > >> >> >> >> This is a subtle detail that should be documented, >> > >> > Ok, will do. >> > >> >> but also >> >> maybe do use samedir if nlink == 1? >> > >> > Hmm.., so initially we could put a relative redirct (if nlink=1) and later >> > if we create a link, we could replace relative redirect with an absolute >> > redirect? I see we already have logic to do that for the case of rename. >> > >> > Now only thing I need to figure out in ovl_link() whethre two dentries >> > are in same dir or not. I am assuming I can just check parent dentry >> > pointers and see if these two have same parent or not. >> >> Yes or we can just convert to absolute path anyway for nlink > 1. >> > >> > In fact, we probably don't even have to check for nlink=1. Only when >> > we create a upper hard link, then we need to make sure we replace relative >> > hardlink with absolute one. I will play with it and see how it goes. >> > >> >> Yes. alomst true. but we do need to check for lower nlink > 1, >> because in that case (when index=on) upper hardlinks are created on >> copy up not only on ovl_link(), so easiest is to just start with >> absolute redirect >> on rename of lower hardlink. > > Hmm..., even in that case it should work. For example, say foo.txt and > bar.txt are hardlinked in lower. hence nlink=2. And now I rename foo.txt > to foo-upper.txt, then it will be copied up (with index hardlinked) and > then a redirect will be set (foo.txt). And now, bar.txt can be looked up > without redirct and foo-upper.txt can be looked up with redirect. So it > should work even in nlink>1 in lower. Can you give a specific example > where it will be broken. > Suppose bar/bar.txt and foo/foo.txt. On metacopy up of bar.txt you need to convert to absolute redirect. Because (with index=on) bar.txt will be "linked up" to foo-upper.txt, which has a redirect xattr of "foo.txt". On cold cache lookup of bar/bar.txt, metacopy lookup will follow "foo.txt" redirect to "bar/foo.txt", which does not exist. > Having said that I don't mind to always set absolute redirect whenever > nlink > 1 (both in lower and upper) and that simplifies the logic a bit. > Yeh, we optimize the common case of nlink == 1 and same dir, because absolute redirect is heavy on lookup, and we leave the less common case of upper or lower hardlink simple and not optimized. Thanks, Amir. -- To unsubscribe from this list: send the line "unsubscribe linux-unionfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
