On Sat, 2024-08-03 at 06:44 +0200, Mateusz Guzik wrote: > On Fri, Aug 02, 2024 at 05:45:04PM -0400, Jeff Layton wrote: > > In a later patch, we want to change the open(..., O_CREAT) codepath to > > avoid taking the inode->i_rwsem for write when the dentry already exists. > > When we tested that initially, the performance devolved significantly > > due to contention for the parent's d_lockref spinlock. > > > > There are two problems with lockrefs today: First, once any concurrent > > task takes the spinlock, they all end up taking the spinlock, which is > > much more costly than a single cmpxchg operation. The second problem is > > that once any task fails to cmpxchg 100 times, it falls back to the > > spinlock. The upshot there is that even moderate contention can cause a > > fallback to serialized spinlocking, which worsens performance. > > > > This patch changes CMPXCHG_LOOP in 2 ways: > > > > First, change the loop to spin instead of falling back to a locked > > codepath when the spinlock is held. Once the lock is released, allow the > > task to continue trying its cmpxchg loop as before instead of taking the > > lock. Second, don't allow the cmpxchg loop to give up after 100 retries. > > Just continue infinitely. > > > > This greatly reduces contention on the lockref when there are large > > numbers of concurrent increments and decrements occurring. > > > > This was already tried by me and it unfortunately can reduce performance. > > Key problem is that in some corner cases the lock can be continuously > held and be queued on, making the fast path always fail and making all > the spins actively waste time (and notably pull on the cacheline). > The cacheline contention does seem like a real problem with this approach. > See this for more details: > https://lore.kernel.org/oe-lkp/lv7ykdnn2nrci3orajf7ev64afxqdw2d65bcpu2mfaqbkvv4ke@hzxat7utjnvx/ > > However, I *suspect* in the case you are optimizing here (open + O_CREAT > of an existing file) lockref on the parent can be avoided altogether > with some hackery and that's what should be done here. > Unfortunately I don't think we can in this codepath: -------------------8<---------------------- if (!(open_flag & O_CREAT)) { ... } else { /* create side of things */ if (nd->flags & LOOKUP_RCU) { if (!try_to_unlazy(nd)) return ERR_PTR(-ECHILD); } audit_inode(nd->name, dir, AUDIT_INODE_PARENT); /* trailing slashes? */ if (unlikely(nd->last.name[nd->last.len])) return ERR_PTR(-EISDIR); } -------------------8<---------------------- The problem here is the audit_inode call, which can do a GFP_KERNEL allocation. We can't stay in RCU mode for that, and we need a reference to "dir" (at least with the current way audit works). > When it comes to lockref in vfs in general, most uses can be elided with > some hackery (see the above thread) which is in early WIP (the LSMs are > a massive headache). > > For open calls which *do* need to take a real ref the hackery does not > help of course. > > This is where I think decoupling ref from the lock is the best way > forward. For that to work the dentry must hang around after the last > unref (already done thanks to RCU and dput even explicitly handles that > already!) and there needs to be a way to block new refs atomically -- > can be done with cmpxchg from a 0-ref state to a flag blocking new refs > coming in. I have that as a WIP as well. > These both sound very interesting. FWIW, Josef also started looking at decoupling the refcount and lock, but I don't think he's gotten very far yet. I'm happy to help test some of this too if you get to that point. The 4th patch in this RFC series really amps up the contention for the lockref once the i_rwsem isn't being touched. > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx> > > --- > > lib/lockref.c | 85 ++++++++++++++++++++++------------------------------------- > > 1 file changed, 32 insertions(+), 53 deletions(-) > > > > diff --git a/lib/lockref.c b/lib/lockref.c > > index 2afe4c5d8919..b76941043fe9 100644 > > --- a/lib/lockref.c > > +++ b/lib/lockref.c > > @@ -8,22 +8,25 @@ > > * Note that the "cmpxchg()" reloads the "old" value for the > > * failure case. > > */ > > -#define CMPXCHG_LOOP(CODE, SUCCESS) do { \ > > - int retry = 100; \ > > - struct lockref old; \ > > - BUILD_BUG_ON(sizeof(old) != 8); \ > > - old.lock_count = READ_ONCE(lockref->lock_count); \ > > - while (likely(arch_spin_value_unlocked(old.lock.rlock.raw_lock))) { \ > > - struct lockref new = old; \ > > - CODE \ > > - if (likely(try_cmpxchg64_relaxed(&lockref->lock_count, \ > > - &old.lock_count, \ > > - new.lock_count))) { \ > > - SUCCESS; \ > > - } \ > > - if (!--retry) \ > > - break; \ > > - } \ > > +#define CMPXCHG_LOOP(CODE, SUCCESS) do { \ > > + struct lockref old; \ > > + BUILD_BUG_ON(sizeof(old) != 8); \ > > + old.lock_count = READ_ONCE(lockref->lock_count); \ > > + for (;;) { \ > > + struct lockref new = old; \ > > + \ > > + if (likely(arch_spin_value_unlocked(old.lock.rlock.raw_lock))) { \ > > + CODE \ > > + if (likely(try_cmpxchg64_relaxed(&lockref->lock_count, \ > > + &old.lock_count, \ > > + new.lock_count))) { \ > > + SUCCESS; \ > > + } \ > > + } else { \ > > + cpu_relax(); \ > > + old.lock_count = READ_ONCE(lockref->lock_count); \ > > + } \ > > + } \ > > } while (0) > > > > #else > > @@ -46,10 +49,8 @@ void lockref_get(struct lockref *lockref) > > , > > return; > > ); > > - > > - spin_lock(&lockref->lock); > > - lockref->count++; > > - spin_unlock(&lockref->lock); > > + /* should never get here */ > > + WARN_ON_ONCE(1); > > } > > EXPORT_SYMBOL(lockref_get); > > > > @@ -60,8 +61,6 @@ EXPORT_SYMBOL(lockref_get); > > */ > > int lockref_get_not_zero(struct lockref *lockref) > > { > > - int retval; > > - > > CMPXCHG_LOOP( > > new.count++; > > if (old.count <= 0) > > @@ -69,15 +68,9 @@ int lockref_get_not_zero(struct lockref *lockref) > > , > > return 1; > > ); > > - > > - spin_lock(&lockref->lock); > > - retval = 0; > > - if (lockref->count > 0) { > > - lockref->count++; > > - retval = 1; > > - } > > - spin_unlock(&lockref->lock); > > - return retval; > > + /* should never get here */ > > + WARN_ON_ONCE(1); > > + return -1; > > } > > EXPORT_SYMBOL(lockref_get_not_zero); > > > > @@ -88,8 +81,6 @@ EXPORT_SYMBOL(lockref_get_not_zero); > > */ > > int lockref_put_not_zero(struct lockref *lockref) > > { > > - int retval; > > - > > CMPXCHG_LOOP( > > new.count--; > > if (old.count <= 1) > > @@ -97,15 +88,9 @@ int lockref_put_not_zero(struct lockref *lockref) > > , > > return 1; > > ); > > - > > - spin_lock(&lockref->lock); > > - retval = 0; > > - if (lockref->count > 1) { > > - lockref->count--; > > - retval = 1; > > - } > > - spin_unlock(&lockref->lock); > > - return retval; > > + /* should never get here */ > > + WARN_ON_ONCE(1); > > + return -1; > > } > > EXPORT_SYMBOL(lockref_put_not_zero); > > > > @@ -125,6 +110,8 @@ int lockref_put_return(struct lockref *lockref) > > , > > return new.count; > > ); > > + /* should never get here */ > > + WARN_ON_ONCE(1); > > return -1; > > } > > EXPORT_SYMBOL(lockref_put_return); > > @@ -171,8 +158,6 @@ EXPORT_SYMBOL(lockref_mark_dead); > > */ > > int lockref_get_not_dead(struct lockref *lockref) > > { > > - int retval; > > - > > CMPXCHG_LOOP( > > new.count++; > > if (old.count < 0) > > @@ -180,14 +165,8 @@ int lockref_get_not_dead(struct lockref *lockref) > > , > > return 1; > > ); > > - > > - spin_lock(&lockref->lock); > > - retval = 0; > > - if (lockref->count >= 0) { > > - lockref->count++; > > - retval = 1; > > - } > > - spin_unlock(&lockref->lock); > > - return retval; > > + /* should never get here */ > > + WARN_ON_ONCE(1); > > + return -1; > > } > > EXPORT_SYMBOL(lockref_get_not_dead); > > > > -- > > 2.45.2 > > -- Jeff Layton <jlayton@xxxxxxxxxx>
