On Thu, Nov 09, 2023 at 09:57:39PM -0800, Linus Torvalds wrote: > Anyway, what I'm actually getting at in a roundabout way is that maybe > we should make D_UNHASHED be another flag in d_flags, and *not* use > that d_hash.pprev field, and that would allow us to combine even more > of these tests in dput(), because now pretty much *all* of those > "retain_dentry()" checks would be about d_flags bits. > > Hmm? As it is, it has that odd combination of d_flags and that > d_unhashed() test, so it's testing two different fields. Hmm, indeed. The trouble is, we are getting tight on the ->d_flags bits. Only two unassigned bits left (0x08000000 and 0x80000000). DCACHE_COOKIE is defined (0x00002000), but unused. Should've been taken out when dcookie stuff went. DCACHE_DENTRY_KILLED might be mergable with DCACHE_MAY_FREE now; worth looking into. In effect, DCACHE_MAY_FREE is set iff we have both DCACHE_DENTRY_KILLED and DCACHE_SHRINK_LIST - and the only place that checks it is guaranteed to have had DCACHE_SHRINK_LIST. Actually, that's nice - in terms of dentry states we have refcount > 0 <=> Busy refcount == 0 <=> Retained refcount < 0 && !KILLED <=> Dying refcount < 0 && KILLED && !SHRINK_LIST <=> Freeing refcount < 0 && KILLED && SHRINK_LIST <=> Husk. <makes a note in the docs being written> DCACHE_FALLTRHU is odd - it's never checked (or set, for that matter); might be killable, might be intended for some overlayfs plans. DCACHE_GENOCIDE might become killable, what with selinuxfs patch I've got (apparently OK with selinux folks, will sort it out after -rc1). OK, it's not as awful as I thought - one more bit won't hurt. I'll go through the unlocked callers and see if any of those is sensitive to separating setting that flag from hash list removal. There might be dragons... > Anyway, I really don't think it matters much, but since you brought up > the whole suboptimal code generation.. FWIW, it's not all that suboptimal, at least with current gcc. The thing I'm really not sure about is whether that patch makes the whole thing easier to follow - probably need to let it sit around for a week or so, then look at it again; right now I don't trust my taste regarding that particular change, having spent too much time today mucking with it ;-/
