Re: [PATCH] x86 spinlock: Fix memory corruption on completing completions

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 02/06/2015 06:49 AM, Raghavendra K T wrote:
> Paravirt spinlock clears slowpath flag after doing unlock.
> As explained by Linus currently it does:
>                 prev = *lock;
>                 add_smp(&lock->tickets.head, TICKET_LOCK_INC);
>
>                 /* add_smp() is a full mb() */
>
>                 if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
>                         __ticket_unlock_slowpath(lock, prev);
>
>
> which is *exactly* the kind of things you cannot do with spinlocks,
> because after you've done the "add_smp()" and released the spinlock
> for the fast-path, you can't access the spinlock any more.  Exactly
> because a fast-path lock might come in, and release the whole data
> structure.

Yeah, that's an embarrasingly obvious bug in retrospect.

> Linus suggested that we should not do any writes to lock after unlock(),
> and we can move slowpath clearing to fastpath lock.

Yep, that seems like a sound approach.

> However it brings additional case to be handled, viz., slowpath still
> could be set when somebody does arch_trylock. Handle that too by ignoring
> slowpath flag during lock availability check.
>
> Reported-by: Sasha Levin <sasha.levin@oracle.com>
> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
> Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
> ---
>  arch/x86/include/asm/spinlock.h | 70 ++++++++++++++++++++---------------------
>  1 file changed, 34 insertions(+), 36 deletions(-)
>
> diff --git a/arch/x86/include/asm/spinlock.h b/arch/x86/include/asm/spinlock.h
> index 625660f..0829f86 100644
> --- a/arch/x86/include/asm/spinlock.h
> +++ b/arch/x86/include/asm/spinlock.h
> @@ -49,6 +49,23 @@ static inline void __ticket_enter_slowpath(arch_spinlock_t *lock)
>  	set_bit(0, (volatile unsigned long *)&lock->tickets.tail);
>  }
>  
> +static inline void __ticket_check_and_clear_slowpath(arch_spinlock_t *lock)
> +{
> +	arch_spinlock_t old, new;
> +	__ticket_t diff;
> +
> +	old.tickets = READ_ONCE(lock->tickets);

Couldn't the caller pass in the lock state that it read rather than
re-reading it?

> +	diff = (old.tickets.tail & ~TICKET_SLOWPATH_FLAG) - old.tickets.head;
> +
> +	/* try to clear slowpath flag when there are no contenders */
> +	if ((old.tickets.tail & TICKET_SLOWPATH_FLAG) &&
> +		(diff == TICKET_LOCK_INC)) {
> +		new = old;
> +		new.tickets.tail &= ~TICKET_SLOWPATH_FLAG;
> +		cmpxchg(&lock->head_tail, old.head_tail, new.head_tail);
> +	}
> +}
> +
>  #else  /* !CONFIG_PARAVIRT_SPINLOCKS */
>  static __always_inline void __ticket_lock_spinning(arch_spinlock_t *lock,
>  							__ticket_t ticket)
> @@ -59,6 +76,10 @@ static inline void __ticket_unlock_kick(arch_spinlock_t *lock,
>  {
>  }
>  
> +static inline void __ticket_check_and_clear_slowpath(arch_spinlock_t *lock)
> +{
> +}
> +
>  #endif /* CONFIG_PARAVIRT_SPINLOCKS */
>  
>  static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock)
> @@ -84,7 +105,7 @@ static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
>  	register struct __raw_tickets inc = { .tail = TICKET_LOCK_INC };
>  
>  	inc = xadd(&lock->tickets, inc);
> -	if (likely(inc.head == inc.tail))
> +	if (likely(inc.head == (inc.tail & ~TICKET_SLOWPATH_FLAG)))

The intent of this conditional was to be the quickest possible path when
taking a fastpath lock, with the code below being used for all slowpath
locks (free or taken). So I don't think masking out SLOWPATH_FLAG is
necessary here.

>  		goto out;
>  
>  	inc.tail &= ~TICKET_SLOWPATH_FLAG;
> @@ -98,7 +119,10 @@ static __always_inline void arch_spin_lock(arch_spinlock_t *lock)
>  		} while (--count);
>  		__ticket_lock_spinning(lock, inc.tail);
>  	}
> -out:	barrier();	/* make sure nothing creeps before the lock is taken */
> +out:
> +	__ticket_check_and_clear_slowpath(lock);
> +
> +	barrier();	/* make sure nothing creeps before the lock is taken */

Which means that if "goto out" path is only ever used for fastpath
locks, you can limit calling __ticket_check_and_clear_slowpath() to the
slowpath case.

>  }
>  
>  static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
> @@ -115,47 +139,21 @@ static __always_inline int arch_spin_trylock(arch_spinlock_t *lock)
>  	return cmpxchg(&lock->head_tail, old.head_tail, new.head_tail) == old.head_tail;
>  }
>  
> -static inline void __ticket_unlock_slowpath(arch_spinlock_t *lock,
> -					    arch_spinlock_t old)
> -{
> -	arch_spinlock_t new;
> -
> -	BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS);
> -
> -	/* Perform the unlock on the "before" copy */
> -	old.tickets.head += TICKET_LOCK_INC;

NB (see below)

> -
> -	/* Clear the slowpath flag */
> -	new.head_tail = old.head_tail & ~(TICKET_SLOWPATH_FLAG << TICKET_SHIFT);
> -
> -	/*
> -	 * If the lock is uncontended, clear the flag - use cmpxchg in
> -	 * case it changes behind our back though.
> -	 */
> -	if (new.tickets.head != new.tickets.tail ||
> -	    cmpxchg(&lock->head_tail, old.head_tail,
> -					new.head_tail) != old.head_tail) {
> -		/*
> -		 * Lock still has someone queued for it, so wake up an
> -		 * appropriate waiter.
> -		 */
> -		__ticket_unlock_kick(lock, old.tickets.head);
> -	}
> -}
> -
>  static __always_inline void arch_spin_unlock(arch_spinlock_t *lock)
>  {
>  	if (TICKET_SLOWPATH_FLAG &&
> -	    static_key_false(&paravirt_ticketlocks_enabled)) {
> -		arch_spinlock_t prev;
> +		static_key_false(&paravirt_ticketlocks_enabled)) {
> +		__ticket_t prev_head;
>  
> -		prev = *lock;
> +		prev_head = lock->tickets.head;
>  		add_smp(&lock->tickets.head, TICKET_LOCK_INC);
>  
>  		/* add_smp() is a full mb() */
>  
> -		if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG))
> -			__ticket_unlock_slowpath(lock, prev);
> +		if (unlikely(lock->tickets.tail & TICKET_SLOWPATH_FLAG)) {

So we're OK with still having a ("speculative"?) read-after-unlock here?
I guess the only way to avoid it is to make the add_smp an xadd, but
that's pretty expensive even compared to a locked add (at least last
time I checked, which was at least a couple of microarchitectures ago).
An unlocked add followed by lfence should also do the trick, but that
was also much worse in practice.

> +			BUILD_BUG_ON(((__ticket_t)NR_CPUS) != NR_CPUS);
> +			__ticket_unlock_kick(lock, prev_head);

Should be "prev_head + TICKET_LOCK_INC" to match the previous code,
otherwise it won't find the CPU waiting for the new head.

    J

> +		}
>  	} else
>  		__add(&lock->tickets.head, TICKET_LOCK_INC, UNLOCK_LOCK_PREFIX);
>  }
> @@ -164,7 +162,7 @@ static inline int arch_spin_is_locked(arch_spinlock_t *lock)
>  {
>  	struct __raw_tickets tmp = READ_ONCE(lock->tickets);
>  
> -	return tmp.tail != tmp.head;
> +	return (tmp.tail & ~TICKET_SLOWPATH_FLAG) != tmp.head;
>  }
>  
>  static inline int arch_spin_is_contended(arch_spinlock_t *lock)