[v2,2/6] dax: Use copy_from_iter_nocache
diff mbox

Message ID 1435934443-17090-3-git-send-email-matthew.r.wilcox@intel.com
State New
Headers show

Commit Message

Wilcox, Matthew R July 3, 2015, 2:40 p.m. UTC
From: Matthew Wilcox <willy@linux.intel.com>

When userspace does a write, there's no need for the written data to
pollute the CPU cache.  This matches the original XIP code.

Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
---
 fs/dax.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Boaz Harrosh July 5, 2015, 1:11 p.m. UTC | #1
On 07/03/2015 05:40 PM, Matthew Wilcox wrote:
> From: Matthew Wilcox <willy@linux.intel.com>
> 
> When userspace does a write, there's no need for the written data to
> pollute the CPU cache.  This matches the original XIP code.
> 
> Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
> ---
>  fs/dax.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/dax.c b/fs/dax.c
> index 99b5fbc..eaa9e06 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -155,7 +155,7 @@ static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
>  		}
>  
>  		if (iov_iter_rw(iter) == WRITE)
> -			len = copy_from_iter(addr, max - pos, iter);
> +			len = copy_from_iter_nocache(addr, max - pos, iter);
>  		else if (!hole)
>  			len = copy_to_iter(addr, max - pos, iter);
>  		else
> 

With the current ioremap_nocache at pmem none of this matters for pmem.

For brd yes, so We've been conducting some measurements and regular ext4
(no DAX) benchmark gives 6-16% increase in performance with this above.
And DAX is almost x2 then no DAX.
Is why the network guys been using this for a long time. So I'd say this
is a good default for any page-cache writes. (Think about it it makes sense,
we will 95% of the time flush these to real memory before DMA)

For pmem with any sane cached mapping (We use page-stuct-pmem actually)
Then DAX, for it to actually work (persist) with pmem, needs this:

static size_t copy_from_iter_nt(void *addr, size_t bytes, struct iov_iter *ii)
{
	size_t ret = copy_from_iter_nocache(addr, bytes, ii);

	if (unlikely((ii->type & ITER_BVEC) || (ii->type & ITER_KVEC))) {
		/* FIXME: copy_from_iter_nocache did regular copy for Kernel
		 * buffers (BVEC or KVEC). Before we fix it do cl_flush
		 * for now.
		 */
		cl_flush(addr, bytes, false);
	} else {
		/* copy_from_iter_nocache only persists in 8-byte aligned words.
		 * Lets persist remaining unaligned edges.
		 */
		if (unlikely((ulong)addr & 0x7))
			cl_flush(addr, 1, false);
		if (unlikely((ulong)(addr + bytes) & 0x7))
			cl_flush((addr + bytes), 1, false);
	}

	return ret;
}

This is based on an not-in-kernel cl_flush().

The first part FIXME above could be fixed with Dan's memcpy_persistent() patches

Cheers
Boaz

--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/fs/dax.c b/fs/dax.c
index 99b5fbc..eaa9e06 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -155,7 +155,7 @@  static ssize_t dax_io(struct inode *inode, struct iov_iter *iter,
 		}
 
 		if (iov_iter_rw(iter) == WRITE)
-			len = copy_from_iter(addr, max - pos, iter);
+			len = copy_from_iter_nocache(addr, max - pos, iter);
 		else if (!hole)
 			len = copy_to_iter(addr, max - pos, iter);
 		else