On Wed, May 12, 2021 at 08:52:25AM -0400, Brian Foster wrote:

On Thu, May 06, 2021 at 05:20:50PM +1000, Dave Chinner wrote:

quoted

From: Dave Chinner <redacted>

xfs_dialloc_select_ag() does a lot of repetitive work. It first
calls xfs_ialloc_ag_select() to select the AG to start allocation
attempts in, which can do up to two entire loops across the perags
that inodes can be allocated in. This is simply checking if there is
spce available to allocate inodes in an AG, and it returns when it
finds the first candidate AG.

xfs_dialloc_select_ag() then does it's own iterative walk across
all the perags locking the AGIs and trying to allocate inodes from
the locked AG. It also doesn't limit the search to mp->m_maxagi,
so it will walk all AGs whether they can allocate inodes or not.

Hence if we are really low on inodes, we could do almost 3 entire
walks across the whole perag range before we find an allocation
group we can allocate inodes in or report ENOSPC.

Because xfs_ialloc_ag_select() returns on the first candidate AG it
finds, we can simply do these checks directly in
xfs_dialloc_select_ag() before we lock and try to allocate inodes.
This reduces the inode allocation pass down to 2 perag sweeps at
most - one for aligned inode cluster allocation and if we can't
allocate full, aligned inode clusters anywhere we'll do another pass
trying to do sparse inode cluster allocation.

This also removes a big chunk of duplicate code.

Signed-off-by: Dave Chinner <redacted>
---
 fs/xfs/libxfs/xfs_ialloc.c | 221 +++++++++++++------------------------
 1 file changed, 75 insertions(+), 146 deletions(-)

diff --git a/fs/xfs/libxfs/xfs_ialloc.c b/fs/xfs/libxfs/xfs_ialloc.c
index 872591e8f5cb..b22556556bba 100644
--- a/fs/xfs/libxfs/xfs_ialloc.c
+++ b/fs/xfs/libxfs/xfs_ialloc.c

...

quoted

@@ -1778,10 +1669,41 @@ xfs_dialloc_select_ag(
 				break;
 		}
 
+		if (!pag->pagi_freecount)
+			goto nextag;

It looks like this would never allow for allocation of new inode
chunks..?

quoted

+		if (!okalloc)
+			goto nextag;

I guess I never tested this patch in isolation, and I fixed the bug
in one of the following patches. This should be:

		if (!pag->pagi_freecount && !okalloc)
			goto nextag;

Fixed.

-Dave.
-- 
Dave Chinner
david@fromorbit.com