| Message ID | 20250430152554.23646-1-jarkko@kernel.org |
|---|---|
| State | New |
| Headers | show |
| Series | KEYS: Reduce smp_mb() calls in key_put() | expand |
On Wed, Apr 30, 2025 at 06:25:53PM +0300, Jarkko Sakkinen wrote: > Rely only on the memory ordering of spin_unlock() when setting > KEY_FLAG_FINAL_PUT under key->user->lock in key_put(). > > Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org> > --- > security/keys/key.c | 6 ++++-- > 1 file changed, 4 insertions(+), 2 deletions(-) > > diff --git a/security/keys/key.c b/security/keys/key.c > index 7198cd2ac3a3..aecbd624612d 100644 > --- a/security/keys/key.c > +++ b/security/keys/key.c > @@ -656,10 +656,12 @@ void key_put(struct key *key) > spin_lock_irqsave(&key->user->lock, flags); > key->user->qnkeys--; > key->user->qnbytes -= key->quotalen; > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > spin_unlock_irqrestore(&key->user->lock, flags); > + } else { > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > + smp_mb(); /* key->user before FINAL_PUT set. */ > } > - smp_mb(); /* key->user before FINAL_PUT set. */ > - set_bit(KEY_FLAG_FINAL_PUT, &key->flags); Oops, my bad (order swap), sorry. Should have been: spin_unlock_irqrestore(&key->user->lock, flags); } else { smp_mb(); /* key->user before FINAL_PUT set. */ } set_bit(KEY_FLAG_FINAL_PUT, &key->flags); Should spin_lock()/unlock() be good enough or what good does smp_mb() do in that branch? Just checking if I'm missing something before sending fixed version. BR, Jarkko
On Sat, May 03, 2025 at 05:39:16PM +0300, Jarkko Sakkinen wrote: > On Wed, Apr 30, 2025 at 06:25:53PM +0300, Jarkko Sakkinen wrote: > > Rely only on the memory ordering of spin_unlock() when setting > > KEY_FLAG_FINAL_PUT under key->user->lock in key_put(). > > > > Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org> > > --- > > security/keys/key.c | 6 ++++-- > > 1 file changed, 4 insertions(+), 2 deletions(-) > > > > diff --git a/security/keys/key.c b/security/keys/key.c > > index 7198cd2ac3a3..aecbd624612d 100644 > > --- a/security/keys/key.c > > +++ b/security/keys/key.c > > @@ -656,10 +656,12 @@ void key_put(struct key *key) > > spin_lock_irqsave(&key->user->lock, flags); > > key->user->qnkeys--; > > key->user->qnbytes -= key->quotalen; > > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > spin_unlock_irqrestore(&key->user->lock, flags); > > + } else { > > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > + smp_mb(); /* key->user before FINAL_PUT set. */ > > } > > - smp_mb(); /* key->user before FINAL_PUT set. */ > > - set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > Oops, my bad (order swap), sorry. Should have been: > > spin_unlock_irqrestore(&key->user->lock, flags); > } else { > smp_mb(); /* key->user before FINAL_PUT set. */ You can use smp_mb__before_atomic here as it is equivalent to smp_mb in this situation. > } > set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > Should spin_lock()/unlock() be good enough or what good does smp_mb() do > in that branch? Just checking if I'm missing something before sending > fixed version. I don't think spin_unlock alone is enough to replace an smp_mb. A spin_lock + spin_unlock would be enough though. However, looking at the bigger picture this smp_mb looks bogus. What exactly is it protecting against? The race condition that this is supposed to fix should have been dealt with by the set_bit/test_bit of FINAL_PUT alone. I don't see any point in having this smb_mb at all. Cheers,
Jarkko Sakkinen <jarkko@kernel.org> wrote: > Oops, my bad (order swap), sorry. Should have been: > > spin_unlock_irqrestore(&key->user->lock, flags); > } else { > smp_mb(); /* key->user before FINAL_PUT set. */ > } > set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > Should spin_lock()/unlock() be good enough or what good does smp_mb() do > in that branch? Just checking if I'm missing something before sending > fixed version. spin_unlock() is semi-permeable, so stuff after it can leak into the inside of it up as far as the spin_lock(). With your change, the garbage collector can no longer guarantee that key_put() will have done with accessing key->user when it sees KEY_FLAG_FINAL_PUT is set. So, NAK on this patch, I think. If you want a second opinion, I'd suggest waving it in front of Paul McKenney. Possibly we only need smp_mb() in the IN_QUOTA branch in key_put(). David
On Sat, May 03, 2025 at 11:19:21PM +0100, David Howells wrote: > > Possibly we only need smp_mb() in the IN_QUOTA branch in key_put(). Just change the smp_mb to smp_mb__before_atomic, at least on x86 it just disappears because set_bit is already a serialising operation. Or even better, reverse the FINAL_PUT bit and call it ALIVE, so that you can use test_bit_acquire and clear_bit_unlock. Cheers,
Herbert Xu <herbert@gondor.apana.org.au> wrote:
> + key->flags |= KEY_FLAG_DONT_GC_YET;
You need __set_bit() or 1<<N.
Also, don't really like the name, but that's just bikeshedding. I think I'd
lean more to your initial suggestion of KEY_FLAG_ALIVE.
David
On Sat, May 03, 2025 at 11:19:21PM +0100, David Howells wrote: > Jarkko Sakkinen <jarkko@kernel.org> wrote: > > > Oops, my bad (order swap), sorry. Should have been: > > > > spin_unlock_irqrestore(&key->user->lock, flags); > > } else { > > smp_mb(); /* key->user before FINAL_PUT set. */ > > } > > set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > > > Should spin_lock()/unlock() be good enough or what good does smp_mb() do > > in that branch? Just checking if I'm missing something before sending > > fixed version. > > spin_unlock() is semi-permeable, so stuff after it can leak into the inside of > it up as far as the spin_lock(). With your change, the garbage collector can > no longer guarantee that key_put() will have done with accessing key->user > when it sees KEY_FLAG_FINAL_PUT is set. > > So, NAK on this patch, I think. If you want a second opinion, I'd suggest > waving it in front of Paul McKenney. Fair enough. If I revisit this in a way or another, I'll cc to him for comments but for this I'll buy what you said. > > Possibly we only need smp_mb() in the IN_QUOTA branch in key_put(). > > David Thank you for the comments. BR, Jarkko
On Sat, May 03, 2025 at 11:02:57PM +0800, Herbert Xu wrote: > On Sat, May 03, 2025 at 05:39:16PM +0300, Jarkko Sakkinen wrote: > > On Wed, Apr 30, 2025 at 06:25:53PM +0300, Jarkko Sakkinen wrote: > > > Rely only on the memory ordering of spin_unlock() when setting > > > KEY_FLAG_FINAL_PUT under key->user->lock in key_put(). > > > > > > Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org> > > > --- > > > security/keys/key.c | 6 ++++-- > > > 1 file changed, 4 insertions(+), 2 deletions(-) > > > > > > diff --git a/security/keys/key.c b/security/keys/key.c > > > index 7198cd2ac3a3..aecbd624612d 100644 > > > --- a/security/keys/key.c > > > +++ b/security/keys/key.c > > > @@ -656,10 +656,12 @@ void key_put(struct key *key) > > > spin_lock_irqsave(&key->user->lock, flags); > > > key->user->qnkeys--; > > > key->user->qnbytes -= key->quotalen; > > > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > > spin_unlock_irqrestore(&key->user->lock, flags); > > > + } else { > > > + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > > + smp_mb(); /* key->user before FINAL_PUT set. */ > > > } > > > - smp_mb(); /* key->user before FINAL_PUT set. */ > > > - set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > > > Oops, my bad (order swap), sorry. Should have been: > > > > spin_unlock_irqrestore(&key->user->lock, flags); > > } else { > > smp_mb(); /* key->user before FINAL_PUT set. */ > > You can use smp_mb__before_atomic here as it is equivalent to > smp_mb in this situation. > > > } > > set_bit(KEY_FLAG_FINAL_PUT, &key->flags); > > > > Should spin_lock()/unlock() be good enough or what good does smp_mb() do > > in that branch? Just checking if I'm missing something before sending > > fixed version. > > I don't think spin_unlock alone is enough to replace an smp_mb. > A spin_lock + spin_unlock would be enough though. > > However, looking at the bigger picture this smp_mb looks bogus. > What exactly is it protecting against? > > The race condition that this is supposed to fix should have been > dealt with by the set_bit/test_bit of FINAL_PUT alone. I don't > see any point in having this smb_mb at all. smp_mb() there makes sure that key->user change don't spill between key_put() and gc. GC pairs smp_mb() in key_put() after FINAL_PUT to make sure that also in its side key->user changes have been walled before moving the key as part of unrefenced keys. See also [1]. It cleared this up for me. Here user->lock easily misleads to overlook the actual synchronization scheme. > > Cheers, > -- > Email: Herbert Xu <herbert@gondor.apana.org.au> > Home Page: http://gondor.apana.org.au/~herbert/ > PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt [1] https://lore.kernel.org/keyrings/1121543.1746310761@warthog.procyon.org.uk/ BR, Jarkko
diff --git a/security/keys/key.c b/security/keys/key.c index 7198cd2ac3a3..aecbd624612d 100644 --- a/security/keys/key.c +++ b/security/keys/key.c @@ -656,10 +656,12 @@ void key_put(struct key *key) spin_lock_irqsave(&key->user->lock, flags); key->user->qnkeys--; key->user->qnbytes -= key->quotalen; + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); spin_unlock_irqrestore(&key->user->lock, flags); + } else { + set_bit(KEY_FLAG_FINAL_PUT, &key->flags); + smp_mb(); /* key->user before FINAL_PUT set. */ } - smp_mb(); /* key->user before FINAL_PUT set. */ - set_bit(KEY_FLAG_FINAL_PUT, &key->flags); schedule_work(&key_gc_work); } }
Rely only on the memory ordering of spin_unlock() when setting KEY_FLAG_FINAL_PUT under key->user->lock in key_put(). Signed-off-by: Jarkko Sakkinen <jarkko@kernel.org> --- security/keys/key.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)