On Sat, Nov 11, 2023 at 08:31:11PM +0200, Amir Goldstein wrote: > > in ovl_lookup(), and in case we have d_splice_alias() return a non-NULL > > dentry we can simply copy it there. Sure, somebody might race with > > us, pick dentry from hash and call ->d_revalidate() before we notice that > > DCACHE_OP_REVALIDATE could be cleaned. So what? That call of ->d_revalidate() > > will find nothing to do and return 1. Which is the effect of having > > DCACHE_OP_REVALIDATE cleared, except for pointless method call. Anyone > > who finds that dentry after the flag is cleared will skip the call. > > IOW, that race is harmless. > > Just a minute. > Do you know that ovl_obtain_alias() is *only* used to obtain a disconnected > non-dir overlayfs dentry? D'oh... > I think that makes all the analysis regarding race with d_splice_alias() > moot. Right? Right you are. > Do DCACHE_OP_*REVALIDATE even matter for a disconnected > non-dir dentry? As long as nothing picks it via d_find_any_alias() and moves it somewhere manually. The former might happen, the latter, AFAICS, doesn't - nothing like d_move() anywhere in sight... > You are missing that the OVL_E_UPPER_ALIAS flag is a property of > the overlay dentry, not a property of the inode. > > N lower hardlinks, the first copy up created an upper inode > all the rest of the N upper aliases to that upper inode are > created lazily. > > However, for obvious reasons, OVL_E_UPPER_ALIAS is not > well defined for a disconnected overlay dentry. > There should not be any code (I hope) that cares about > OVL_E_UPPER_ALIAS for a disconnected overlay dentry, > so I *think* ovl_dentry_set_upper_alias() in this code is moot. > > I need to look closer to verify, but please confirm my assumption > regarding the irrelevance of DCACHE_OP_*REVALIDATE for a > disconnected non-dir dentry. Correct; we only care if it gets reconnected to the main tree. The fact that it's only for non-directories simplifies life a lot there. Sorry, got confused by the work you do with ->d_flags and hadn't stopped to ask whether it's needed in the first place in there. OK, so... are there any reasons why simply calling d_obtain_alias() wouldn't do the right thing these days?
