Re: [PATCH] mm/fs: don't keep pages when receiving a pending SIGKILL in __get_user_pages()

[David Rientjes <rientjes@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1900 bytes --]

On Wed, 15 Jan 2014, Xishi Qiu wrote:

> In the process IO direction, dio_refill_pages will call get_user_pages_fast 
> to map the page from user space. If ret is less than 0 and IO is write, the 
> function will create a zero page to fill data. This may work for some file 
> system, but in some device operate we prefer whole write or fail, not half 
> data half zero, e.g. fs metadata, like inode, identy.
> This happens often when kill a process which is doing direct IO. Consider 
> the following cases, the process A is doing IO process, may enter __get_user_pages 
> function, if other processes send process A SIG_KILL, A will enter the 
> following branches 
> 		/*
> 		 * If we have a pending SIGKILL, don't keep faulting
> 		 * pages and potentially allocating memory.
> 		 */
> 		if (unlikely(fatal_signal_pending(current)))
> 			return i ? i : -ERESTARTSYS;
> Return current pages. direct IO will write the pages, the subsequent pages 
> which can?t get will use zero page instead. 
> This patch will modify this judgment, if receive SIG_KILL, release pages and 
> return an error. Direct IO will find no blocks_available and return error 
> direct, rather than half IO data and half zero page.
> 
> Signed-off-by: Xishi Qiu <qiuxishi@huawei.com>
> Signed-off-by: Bin Yang <robin.yb@huawei.com>

It's scary to change the behavior of gup when some callers may want the 
exact opposite of what you're intending here, which is sane fallback by 
mapping the zero page.  In fact, gup never does put_page() itself and 
__get_user_pages() always returns the number of pages pinned and may not 
equal what is passed.

So, this definitely isn't the right solution for a special-case direct IO.  
Instead, it would be better to code this directly in the caller and 
compare the return value with nr_pages in dio_refill_pages() and then do 
the put_page() itself before falling back to ZERO_PAGE().