btrfs: possible unit mismatch on the async_delalloc_pages limit

btrfs: possible unit mismatch on the async_delalloc_pages limit

[Wu Fengguang]

Hi Chris,

I suspect the async_delalloc_pages limit checking in btrfs/inode.c

        atomic_read(&root->fs_info->async_delalloc_pages) < 5 * 1042 * 1024

never evaluate to FALSE because 5M pages are 20GB which is super
large. I tried to fix it with the following patch. I increased the
limit a bit, however still worry that the "corrected" limit is too
small to create some regressions.

What do you think?

Thanks,
Fengguang

---
 fs/btrfs/inode.c |   10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

--- linux.orig/fs/btrfs/inode.c	2009-11-07 19:28:07.000000000 +0800
+++ linux/fs/btrfs/inode.c	2009-11-07 19:28:08.000000000 +0800
@@ -50,6 +50,8 @@
 #include "compression.h"
 #include "locking.h"
 
+#define BTRFS_ASYNC_DELALLOC_PAGES_LIMIT	(64 << (20-PAGE_CACHE_SHIFT))
+
 struct btrfs_iget_args {
 	u64 ino;
 	struct btrfs_root *root;
@@ -856,7 +858,7 @@ static noinline void async_cow_submit(st
 	atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
 
 	if (atomic_read(&root->fs_info->async_delalloc_pages) <
-	    5 * 1042 * 1024 &&
+	    BTRFS_ASYNC_DELALLOC_PAGES_LIMIT / 2 &&
 	    waitqueue_active(&root->fs_info->async_submit_wait))
 		wake_up(&root->fs_info->async_submit_wait);
 
@@ -879,7 +881,6 @@ static int cow_file_range_async(struct i
 	struct btrfs_root *root = BTRFS_I(inode)->root;
 	unsigned long nr_pages;
 	u64 cur_end;
-	int limit = 10 * 1024 * 1042;
 
 	clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
 			 1, 0, NULL, GFP_NOFS);
@@ -910,10 +911,11 @@ static int cow_file_range_async(struct i
 		btrfs_queue_worker(&root->fs_info->delalloc_workers,
 				   &async_cow->work);
 
-		if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
+		if (atomic_read(&root->fs_info->async_delalloc_pages) >
+		    BTRFS_ASYNC_DELALLOC_PAGES_LIMIT) {
 			wait_event(root->fs_info->async_submit_wait,
 			   (atomic_read(&root->fs_info->async_delalloc_pages) <
-			    limit));
+			    BTRFS_ASYNC_DELALLOC_PAGES_LIMIT));
 		}
 
 		while (atomic_read(&root->fs_info->async_submit_draining) &&

Re: btrfs: possible unit mismatch on the async_delalloc_pages limit

[Chris Mason]

On Wed, Nov 18, 2009 at 04:51:32PM +0800, Wu Fengguang wrote:
> Hi Chris,
> 
> I suspect the async_delalloc_pages limit checking in btrfs/inode.c
> 
>         atomic_read(&root->fs_info->async_delalloc_pages) < 5 * 1042 * 1024
> 
> never evaluate to FALSE because 5M pages are 20GB which is super
> large. I tried to fix it with the following patch. I increased the
> limit a bit, however still worry that the "corrected" limit is too
> small to create some regressions.
> 
> What do you think?

Thanks, I'll give this a shot here.

-chris