svn commit: r289565 - head/sys/ofed/include/linux

[Konstantin Belousov]

On Mon, Oct 19, 2015 at 10:56:32AM +0000, Hans Petter Selasky wrote:
> Author: hselasky
> Date: Mon Oct 19 10:56:32 2015
> New Revision: 289565
> URL: https://svnweb.freebsd.org/changeset/base/289565
> 
> Log:
>   Merge LinuxKPI changes from DragonflyBSD:
>   - Implement pagefault_disable() and pagefault_enable().
>   
>   Sponsored by:	Mellanox Technologies
> 
> Modified:
>   head/sys/ofed/include/linux/uaccess.h
> 
> Modified: head/sys/ofed/include/linux/uaccess.h
> ==============================================================================
> --- head/sys/ofed/include/linux/uaccess.h	Mon Oct 19 10:54:24 2015	(r289564)
> +++ head/sys/ofed/include/linux/uaccess.h	Mon Oct 19 10:56:32 2015	(r289565)
> @@ -2,7 +2,8 @@
>   * Copyright (c) 2010 Isilon Systems, Inc.
>   * Copyright (c) 2010 iX Systems, Inc.
>   * Copyright (c) 2010 Panasas, Inc.
> - * Copyright (c) 2013, 2014 Mellanox Technologies, Ltd.
> + * Copyright (c) 2013-2015 Mellanox Technologies, Ltd.
> + * Copyright (c) 2015 Fran??ois Tigeot
>   * All rights reserved.
>   *
>   * Redistribution and use in source and binary forms, with or without
> @@ -32,4 +33,14 @@
>  #define	get_user(_x, _p)	-copyin((_p), &(_x), sizeof(*(_p)))
>  #define	put_user(_x, _p)	-copyout(&(_x), (_p), sizeof(*(_p)))
>  
> +static inline void pagefault_disable(void)
> +{
> +	curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR);
> +}
> +
> +static inline void pagefault_enable(void)
> +{
> +	curthread_pflags_restore(~(TDP_NOFAULTING | TDP_RESETSPUR));
> +}
> +
>  #endif	/* _LINUX_UACCESS_H_ */

This is wrong on many counts.  First, we already provide the wrappers to
implement the nofaulting behaviour, see vm_fault_disable_pagefaults()
and vm_fault_enable_pagefaults().

Second, your implementation ignores possible recursion on the
state, unconditionally re-enabling faulting, and more seriously, it
unconditionally obliterates the spurious faults tracker.

The return value from vm_fault_disable_pagefaults() is there for the reason,
and if Linux KPI does not have a place to store the state, some replacement
must be implemented instead of silently corrupting the flags on enable.