"J. Bruce Fields" <bfields@xxxxxxxxxxxx> writes: > On Tue, Feb 25, 2014 at 02:03:36PM -0800, Eric W. Biederman wrote: >> "J. Bruce Fields" <bfields@xxxxxxxxxxxx> writes: >> >> > On Mon, Feb 24, 2014 at 04:01:29PM -0800, Eric W. Biederman wrote: >> >> Miklos Szeredi <miklos@xxxxxxxxxx> writes: >> >> >> >> > >> >> > You can optimize this by including the negative check within the above d_locked >> >> > region and calling __d_drop() instead. >> >> >> >> For this patch just moving the code and not changing it is the corret >> >> thing to do because it helps with review and understanding the code. >> >> >> >> There are two ways I could see going with optimizing the preamble. >> >> Simply dropping the d_lock from around the d_unhashed test as a pointer >> >> dereference should be atomic, and the test is racy against >> >> d_materialise_unique. >> > >> > Could you explain? What's the race, and what are the consequences? > > Actually I was just confused as to whether the above was "is racy" was > claiming the existance of some bug. > > I believe I should have read the above as more like "the test is already > racy against d_materialise_unique, but it's a harmless race, and > dropping the d_lock wouldn't make it any worse". > >> >> (We don't always hold the parent directories inode mutex when d_invalidate is called). >> >> d_unhashed is not a permanent condition because of d_materialise_unique, >> and d_splice_alias. >> >> d_invalidate can be called on an unhashed dentry in one of two ways >> (either d_revalidate dropped the dentry or another routine that drops >> the dentry beat the current invocation of d_invalidate to the job). >> >> >> There are 3 places d_revalidate is called. >> >> Once on the rcu path with with the appropriate flag set. >> >> Once without out the parent i_mutex held, just off of the rcu path, >> on that path d_invalidate is when d_revalidate fails. >> >> Once during lookup with the parent directory i_mutex held. >> >> >> Because the parent direcories i_mutex is not always held accross >> d_revalidate and the following d_invalidate it happens that d_invalidate >> is not always an atomic operation. >> >> >> At worst the race results in a dentry that is dropped when it could be >> hashed, > > Because somebody not holding the i_mutex calls d_invalidate based on old > information and unhashes something that > d_materialise_unique/d_splice_alias just hashed? More likely today somebody not holding i_mutex and not in rcu context calls d_revalidate. d_revalidate drops the dentry and before we d_invalidate d_materialise_unique/d_splice_alias rehashes it. After my changes it looks like it takes 3 processes two instances of d_invalidate and a instance of d_materialise_unique/d_spliace_alias to trigger this case. In either case the window is very small and the outcome is effectively harmless. So I don't see this as a problem. >> that we will resurrect next time someone attempts to look it >> up and d_materialise_unique/d_splice_alias is called. > > OK. > >> None of that really matters for optimizing d_invalidate, but it is part >> of the background in which d_invalidate lives. All that is significant >> in d_invalidate is knowing that d_materialise_unique, and possibly >> d_splice_alias may run concurrently with d_invalidate. It is unlikely >> and essentially harmless. >> >> >> After my patchset (because I removed all of the d_drop's from >> .d_revalidate) the only race that should remain is between two parallel >> calls of d_invalidate. Which probably means we can remove the test for >> d_unhashed altogether. >> >> Right now I just want to make this first big step and make certain the >> code is solid. After that optimization is easy. > > Thanks for the explanation! Welcome. Eric -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
