linux-kernel - Re: [PATCH] mm:Avoid soft lockup due to possible attempt of double locking object's lock in __delete

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <20160831075421.GA15732@e104818-lin.cambridge.arm.com>
Date:   Wed, 31 Aug 2016 08:54:21 +0100
From:   Catalin Marinas <catalin.marinas@....com>
To:     Nicholas Krause <xerofoify@...il.com>
Cc:     linux-mm@...ck.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] mm:Avoid soft lockup due to possible attempt of double
 locking object's lock in __delete_object

On Tue, Aug 30, 2016 at 02:35:12PM -0400, Nicholas Krause wrote:
> This fixes a issue in the current locking logic of the function,
> __delete_object where we are trying to attempt to lock the passed
> object structure's spinlock again after being previously held
> elsewhere by the kmemleak code. Fix this by instead of assuming
> we are the only one contending for the object's lock their are
> possible other users and create two branches, one where we get
> the lock when calling spin_trylock_irqsave on the object's lock
> and the other when the lock is held else where by kmemleak.

Have you actually got a deadlock that requires this fix?

> --- a/mm/kmemleak.c
> +++ b/mm/kmemleak.c
> @@ -631,12 +631,19 @@ static void __delete_object(struct kmemleak_object *object)
>  
>  	/*
>  	 * Locking here also ensures that the corresponding memory block
> -	 * cannot be freed when it is being scanned.
> +	 * cannot be freed when it is being scanned. Further more the
> +	 * object's lock may have been previously holded by another holder
> +	 * in the kmemleak code, therefore attempt to lock the object's lock
> +	 * before holding it and unlocking it.
>  	 */
> -	spin_lock_irqsave(&object->lock, flags);
> -	object->flags &= ~OBJECT_ALLOCATED;
> -	spin_unlock_irqrestore(&object->lock, flags);
> -	put_object(object);
> +	if (spin_trylock_irqsave(&object->lock, flags)) {
> +		object->flags &= ~OBJECT_ALLOCATED;
> +		spin_unlock_irqrestore(&object->lock, flags);
> +		put_object(object);
> +	} else {
> +		object->flags &= ~OBJECT_ALLOCATED;
> +		put_object(object);
> +	}

NAK. This lock here is needed, as described in the comment, to prevent
an object being freed while it is being scanned. The scan_object()
function acquires the same lock and checks for OBJECT_ALLOCATED before
accessing the memory (which could be vmalloc'ed for example, so freeing
would cause a page fault).

-- 
Catalin