Skip to contentSkip to conversation

ext4: deaccount delayed allocations at freeing inode in ext4_evict_inode()

PATCH In review Konstantin Khlebnikov 4 participants 7 messages Oct 29, 2019 – Nov 15, 2019 ext4 file system

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ext4: deaccount delayed allocations at freeing inode in ext4_evict_inode()

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

---
 fs/ext4/inode.c |    9 +++++++++
 1 file changed, 9 insertions(+)
fs/ext4/inode.c +9 -0
index 516faa280ced..580898145e8f 100644
@@ -293,6 +293,15 @@ void ext4_evict_inode(struct inode *inode)
 				   inode->i_ino, err);
 			goto stop_handle;
 		}
+	} else if (EXT4_I(inode)->i_reserved_data_blocks) {
+		/* Deaccount reserve if inode has only delayed allocations. */
+		err = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+		if (err) {
+			ext4_warning(inode->i_sb,
+				     "couldn't remove extents %lu (err %d)",
+				     inode->i_ino, err);
+			goto stop_handle;
+		}
 	}
 
 	/* Remove xattr references. */
#1 lore raw
Quoted 7 lines · Konstantin Khlebnikov

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

FYI, perf cannot correctly parse related perf events without this:

https://lore.kernel.org/lkml/157228145325.7530.4974461761228678289.stgit@buzz/

#2 lore raw
In reply to Konstantin Khlebnikov

[email protected] into Cc.

Quoted 34 lines · Konstantin Khlebnikov

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

---
  fs/ext4/inode.c |    9 +++++++++
  1 file changed, 9 insertions(+)
fs/ext4/inode.c +9 -0
index 516faa280ced..580898145e8f 100644
@@ -293,6 +293,15 @@ void ext4_evict_inode(struct inode *inode)
  				   inode->i_ino, err);
  			goto stop_handle;
  		}
+	} else if (EXT4_I(inode)->i_reserved_data_blocks) {
+		/* Deaccount reserve if inode has only delayed allocations. */
+		err = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+		if (err) {
+			ext4_warning(inode->i_sb,
+				     "couldn't remove extents %lu (err %d)",
+				     inode->i_ino, err);
+			goto stop_handle;
+		}
  	}
 
  	/* Remove xattr references. */
#3 lore raw
In reply to Konstantin Khlebnikov
Quoted 7 lines · Konstantin Khlebnikov

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

Sorry, I may have missed it, but could you please help understand what recent changes in delayed allocation make this break or worse?

A silly query, since I couldn't figure it out. Maybe the code has been
there ever since like this:-
So why can't we just move drop_dquot later after the ext4_es_remove_extent() (in function ext4_clear_inode)? Any known problems around that?

-ritesh

Quoted 27 lines
---
  fs/ext4/inode.c |    9 +++++++++
  1 file changed, 9 insertions(+)
fs/ext4/inode.c +9 -0
index 516faa280ced..580898145e8f 100644
@@ -293,6 +293,15 @@ void ext4_evict_inode(struct inode *inode)
  				   inode->i_ino, err);
  			goto stop_handle;
  		}
+	} else if (EXT4_I(inode)->i_reserved_data_blocks) {
+		/* Deaccount reserve if inode has only delayed allocations. */
+		err = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+		if (err) {
+			ext4_warning(inode->i_sb,
+				     "couldn't remove extents %lu (err %d)",
+				     inode->i_ino, err);
+			goto stop_handle;
+		}
  	}
 
  	/* Remove xattr references. */
#4 lore raw
Quoted 8 lines · Ritesh Harjani
Quoted 5 lines · Konstantin Khlebnikov

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

Quoted 5 lines
Quoted 2 lines

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

Sorry, I may have missed it, but could you please help understand what recent changes in delayed allocation make this break or worse?

I don't see problem for 4.19. Haven't bisected yet. Most likely this is around 'reserved cluster accounting'.

I suspect before these changes something always triggered da before unlink and space usage committed and then truncated at eviction.

Quoted 7 lines

A silly query, since I couldn't figure it out. Maybe the code has been there ever since like this:-

So why can't we just move drop_dquot later after the ext4_es_remove_extent() (in function ext4_clear_inode)? Any known problems around that?

Clear_inode is called also when inode evicts from cache while it has nlinks and stays at disk. I'm not sure how this must interact with reserves.

Quoted 32 lines

-ritesh

Quoted 27 lines

---
  fs/ext4/inode.c |    9 +++++++++

  1 file changed, 9 insertions(+)
fs/ext4/inode.c +9 -0
index 516faa280ced..580898145e8f 100644
@@ -293,6 +293,15 @@ void ext4_evict_inode(struct inode *inode)
                      inode->i_ino, err);
               goto stop_handle;
           }
+    } else if (EXT4_I(inode)->i_reserved_data_blocks) {
+        /* Deaccount reserve if inode has only delayed allocations. */
+        err = ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+        if (err) {
+            ext4_warning(inode->i_sb,
+                     "couldn't remove extents %lu (err %d)",
+                     inode->i_ino, err);
+            goto stop_handle;
+        }
       }
 
       /* Remove xattr references. */
reviewer
#5 lore raw
Quoted 12 lines · Konstantin Khlebnikov
Quoted 11 lines · Ritesh Harjani
Quoted 7 lines · Konstantin Khlebnikov

If inode->i_blocks is zero then ext4_evict_inode() skips ext4_truncate(). Delayed allocation extents are freed later in ext4_clear_inode() but this happens when quota reference is already dropped. This leads to leak of reserved space in quota block, which disappears after umount-mount.

This seems broken for a long time but worked somehow until recent changes in delayed allocation.

Quoted 8 lines
Quoted 2 lines

Sorry, I may have missed it, but could you please help understand what recent changes in delayed allocation make this break or worse?

I don't see problem for 4.19. Haven't bisected yet. Most likely this is around 'reserved cluster accounting'.

I suspect before these changes something always triggered da before unlink and space usage committed and then truncated at eviction.

Yes, I think it's commit 8fcc3a580651 "ext4: rework reserved cluster accounting when invalidating pages". Because that commit moved releasing of reserved space from page invalidation time to extent status tree eviction time. Does attached patch fix the problem for you?

Quoted 8 lines
Quoted 5 lines

A silly query, since I couldn't figure it out. Maybe the code has been there ever since like this:-

So why can't we just move drop_dquot later after the ext4_es_remove_extent() (in function ext4_clear_inode)? Any known problems around that?

Clear_inode is called also when inode evicts from cache while it has nlinks and stays at disk. I'm not sure how this must interact with reserves.

In that case all data should be written out for such inode and thus there should be no reserves...

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR

From ee27836b579d3bf750d45cd7081d3433ea6fedd5 Mon Sep 17 00:00:00 2001
From: Jan Kara <[email protected]>
Date: Fri, 8 Nov 2019 12:45:11 +0100
Subject: [PATCH] ext4: Fix leak of quota reservations

Commit 8fcc3a580651 ("ext4: rework reserved cluster accounting when invalidating pages") moved freeing of delayed allocation reservations from dirty page invalidation time to time when we evict corresponding status extent from extent status tree. For inodes which don't have any blocks allocated this may actually happen only in ext4_clear_blocks() which is after we've dropped references to quota structures from the inode. Thus reservation of quota leaked. Fix the problem by clearing quota information from the inode only after evicting extent status tree in ext4_clear_inode().

Fixes 8fcc3a580651 ("ext4: rework reserved cluster accounting when invalidating pages")
---
 fs/ext4/ialloc.c | 5 -----
 fs/ext4/super.c  | 2 +-
 2 files changed, 1 insertion(+), 6 deletions(-)
fs/ext4/ialloc.c +0 -5
index 764ff4c56233..564e2ceb8417 100644
@@ -265,13 +265,8 @@ void ext4_free_inode(handle_t *handle, struct inode *inode)
 	ext4_debug("freeing inode %lu\n", ino);
 	trace_ext4_free_inode(inode);
 
-	/*
-	 * Note: we must free any quota before locking the superblock,
-	 * as writing the quota to disk may need the lock as well.
-	 */
 	dquot_initialize(inode);
 	dquot_free_inode(inode);
-	dquot_drop(inode);
 
 	is_directory = S_ISDIR(inode->i_mode);
 
fs/ext4/super.c +1 -1
index dd654e53ba3d..9589f09a40f6 100644
@@ -1172,9 +1172,9 @@ void ext4_clear_inode(struct inode *inode)
 {
 	invalidate_inode_buffers(inode);
 	clear_inode(inode);
-	dquot_drop(inode);
 	ext4_discard_preallocations(inode);
 	ext4_es_remove_extent(inode, 0, EXT_MAX_BLOCKS);
+	dquot_drop(inode);
 	if (EXT4_I(inode)->jinode) {
 		jbd2_journal_release_jbd_inode(EXT4_JOURNAL(inode),
 					       EXT4_I(inode)->jinode);
Signature · 1 line
2.16.4
maintainer
#6 lore raw
Quoted 16 lines

From ee27836b579d3bf750d45cd7081d3433ea6fedd5 Mon Sep 17 00:00:00 2001
From: Jan Kara <[email protected]>
Date: Fri, 8 Nov 2019 12:45:11 +0100
Subject: [PATCH] ext4: Fix leak of quota reservations

Commit 8fcc3a580651 ("ext4: rework reserved cluster accounting when invalidating pages") moved freeing of delayed allocation reservations from dirty page invalidation time to time when we evict corresponding status extent from extent status tree. For inodes which don't have any blocks allocated this may actually happen only in ext4_clear_blocks() which is after we've dropped references to quota structures from the inode. Thus reservation of quota leaked. Fix the problem by clearing quota information from the inode only after evicting extent status tree in ext4_clear_inode().

Quoted 3 lines
Fixes 8fcc3a580651 ("ext4: rework reserved cluster accounting when invalidating pages")

OK, I've applied this patch.

- Ted