Re: [RFC][PATCH] problem of cont_prepare_write()

[Andrew Morton <akpm@osdl.org>]

OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> wrote:
>
>  If I do the following operation on fatfs, and my box under heavy load,
> 
>  	open("testfile", O_CREAT | O_TRUNC | O_RDWR, 0664);
>  	lseek(fd, 500*1024*1024 - 1, SEEK_SET);
>  	write(fd, "\0", 1);
> 
>  In cont_prepare_write(), kernel fills the hole by zero cleared page.
> 
>  fs/buffer.c:cont_prepare_write:2210,
>  	while(page->index > (pgpos = *bytes>>PAGE_CACHE_SHIFT)) {
> 
>  		[...]	
> 
>  		status = __block_prepare_write(inode, new_page, zerofrom,
>  						PAGE_CACHE_SIZE, get_block);
>  		if (status)
>  			goto out_unmap;
>  		kaddr = kmap_atomic(new_page, KM_USER0);
>  		memset(kaddr+zerofrom, 0, PAGE_CACHE_SIZE-zerofrom);
>  		flush_dcache_page(new_page);
>  		kunmap_atomic(kaddr, KM_USER0);
>  		__block_commit_write(inode, new_page,
>  				zerofrom, PAGE_CACHE_SIZE);
>  		unlock_page(new_page);
>  		page_cache_release(new_page);
>  	}
> 
>  But until ->commit_write(), kernel doesn't update the ->i_size. Then,
>  if kernel writes out that hole page before updates of ->i_size, dirty
>  flag of buffer_head is cleared in __block_write_full_page(). So hole
>  page was not writed to disk.

But the page remains locked across both the ->prepare_write() and
->commit_write() operations.  So writeback cannot get in there to call
->writepage().

The page lock should correctly synchronise the prepare_write/commit_write
and writeback functions.