Re: [PATCH v2 5/5] dax: fix clearing of holes in __dax_pmd_fault()

[Jan Kara <jack@suse.cz>]

On Thu 21-01-16 10:46:04, Ross Zwisler wrote:
> When the user reads from a DAX hole via a mmap we service page faults using
> zero-filled page cache pages.  These zero pages are also placed into the
> address_space radix tree.  When we get our first write for that space, we
> can allocate a PMD page worth of DAX storage to replace that hole.
> 
> When this happens we need to unmap the zero pages and remove them from the
> radix tree.  Prior to this patch we were unmapping *all* storage in our
> PMD's range, which is incorrect because it removed DAX entries as well on
> non-allocating page faults.
> 
> Instead, keep track of when get_block() actually gives us storage so that
> we can be sure to only remove zero pages that were covering holes.
> 
> Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
> Reported-by: Jan Kara <jack@suse.cz>

So the get_block() calls below are still racy so your allocation detection
doesn't quite work reliably anyway. For that we need a change in the fault
locking we were talking about and at that point we'll need to somewhat
update __dax_pmd_fault() anyway. So I'm not sure if this intermediate
change really makes sense... I'd rather first change the locking in the
filesystems and then fixup remaining issues with __dax_pmd_fault().

								Honza

> ---
>  fs/dax.c | 32 ++++++++++++++++++++++----------
>  1 file changed, 22 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/dax.c b/fs/dax.c
> index afacc30..263aed1 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -790,9 +790,9 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
>  	bool write = flags & FAULT_FLAG_WRITE;
>  	struct block_device *bdev;
>  	pgoff_t size, pgoff;
> -	loff_t lstart, lend;
>  	sector_t block;
>  	int error, result = 0;
> +	bool alloc = false;
>  
>  	/* dax pmd mappings require pfn_t_devmap() */
>  	if (!IS_ENABLED(CONFIG_FS_DAX_PMD))
> @@ -830,10 +830,17 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
>  	block = (sector_t)pgoff << (PAGE_SHIFT - blkbits);
>  
>  	bh.b_size = PMD_SIZE;
> -	if (get_block(inode, block, &bh, write) != 0)
> +
> +	if (get_block(inode, block, &bh, 0) != 0)
>  		return VM_FAULT_SIGBUS;
> +
> +	if (!buffer_mapped(&bh) && write) {
> +		if (get_block(inode, block, &bh, 1) != 0)
> +			return VM_FAULT_SIGBUS;
> +		alloc = true;
> +	}
> +
>  	bdev = bh.b_bdev;
> -	i_mmap_lock_read(mapping);
>  
>  	/*
>  	 * If the filesystem isn't willing to tell us the length of a hole,
> @@ -842,15 +849,20 @@ int __dax_pmd_fault(struct vm_area_struct *vma, unsigned long address,
>  	 */
>  	if (!buffer_size_valid(&bh) || bh.b_size < PMD_SIZE) {
>  		dax_pmd_dbg(&bh, address, "allocated block too small");
> -		goto fallback;
> +		return VM_FAULT_FALLBACK;
> +	}
> +
> +	/*
> +	 * If we allocated new storage, make sure no process has any
> +	 * zero pages covering this hole
> +	 */
> +	if (alloc) {
> +		loff_t lstart = pgoff << PAGE_SHIFT;
> +		loff_t lend = lstart + PMD_SIZE - 1; /* inclusive */
> +
> +		truncate_pagecache_range(inode, lstart, lend);
>  	}
>  
> -	/* make sure no process has any zero pages covering this hole */
> -	lstart = pgoff << PAGE_SHIFT;
> -	lend = lstart + PMD_SIZE - 1; /* inclusive */
> -	i_mmap_unlock_read(mapping);
> -	unmap_mapping_range(mapping, lstart, PMD_SIZE, 0);
> -	truncate_inode_pages_range(mapping, lstart, lend);
>  	i_mmap_lock_read(mapping);
>  
>  	/*
> -- 
> 2.5.0
> 
> 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR