In the append-write scenario, after ensuring that the dirty inode can be
seen by the writeback process, there is no need to execute
"mark_inode_dirty" for every write. Instead, we can rely on
"ext4_mark_inode_dirty" executed when updating i_disksize in
"mpage_map_and_submit_extent" to ensure data consistency, which can
significantly improve performance in high-frequency append-write
scenarios.
In test scenarios of Kafka version 2.6.2, using packet size of 2K
resulted in a 10% performance improvement.
Signed-off-by: Liu Song <[email protected]>
---
fs/ext4/inode.c | 58 +++++++++++++++++++++++++++++++++++++++++++++++--
1 file changed, 56 insertions(+), 2 deletions(-)
diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
index 9d9f414f99fe..d1aa775c9936 100644
--- a/fs/ext4/inode.c
+++ b/fs/ext4/inode.c
@@ -3128,6 +3128,57 @@ static int ext4_da_should_update_i_disksize(struct page *page,
return 1;
}
+/*
+ * Copy from generic_write_end, add conditions to execute mark_inode_dirty
+ * to avoid additional overhead caused by frequent dirty inode operations
+ */
+static int ext4_da_generic_write_end(struct file *file, struct address_space *mapping,
+ loff_t pos, unsigned len, unsigned copied, bool need_dirty,
+ struct page *page, void *fsdata)
+{
+ struct inode *inode = mapping->host;
+ loff_t old_size = inode->i_size;
+ bool i_size_changed = false;
+ int was_dirty;
+
+ copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
+
+ /*
+ * No need to use i_size_read() here, the i_size cannot change under us
+ * because we hold i_rwsem.
+ *
+ * But it's important to update i_size while still holding page lock:
+ * page writeout could otherwise come in and zero beyond i_size.
+ */
+ if (pos + copied > inode->i_size) {
+ i_size_write(inode, pos + copied);
+ i_size_changed = true;
+ }
+
+ unlock_page(page);
+ put_page(page);
+
+ if (old_size < pos)
+ pagecache_isize_extended(inode, old_size, pos);
+
+ /*
+ * In the append-write scenario, if the inode is marked as dirty,
+ * it is ensured that the inode will be seen by the writeback process.
+ * In the ext4_writepages process, when updating i_disksize,
+ * corresponding metadata updates are also performed.
+ * Therefore, it is unnecessary to repeatedly execute mark_inode_dirty
+ * to improve performance.
+ */
+ if (i_size_changed) {
+ spin_lock(&inode->i_lock);
+ was_dirty = inode->i_state & I_DIRTY;
+ spin_unlock(&inode->i_lock);
+ if (!was_dirty || need_dirty)
+ mark_inode_dirty(inode);
+ }
+ return copied;
+}
+
static int ext4_da_write_end(struct file *file,
struct address_space *mapping,
loff_t pos, unsigned len, unsigned copied,
@@ -3137,6 +3188,7 @@ static int ext4_da_write_end(struct file *file,
loff_t new_i_size;
unsigned long start, end;
int write_mode = (int)(unsigned long)fsdata;
+ bool need_dirty = false;
if (write_mode == FALL_BACK_TO_NONDELALLOC)
return ext4_write_end(file, mapping, pos,
@@ -3169,10 +3221,12 @@ static int ext4_da_write_end(struct file *file,
*/
new_i_size = pos + copied;
if (copied && new_i_size > inode->i_size &&
- ext4_da_should_update_i_disksize(page, end))
+ ext4_da_should_update_i_disksize(page, end)) {
ext4_update_i_disksize(inode, new_i_size);
+ need_dirty = true;
+ }
- return generic_write_end(file, mapping, pos, len, copied, page, fsdata);
+ return ext4_da_generic_write_end(file, mapping, pos, len, copied, need_dirty, page, fsdata);
}
/*
--
2.19.1.6.gb485710b
Hi,
Any comments on this? Considering the most extreme scenario
where a file is appended with only one byte each time,
"jbd2_journal_dirty_metadata" will be executed because i_size
has changed. Therefore, it is better to update raw_inode's size
after the writeback is completed, which can also ensure the
consistency of metadata and data, and obtain better performance
benefits in high-frequency append write scenarios.
Thanks
On 2023/8/9 20:25, Jan Kara wrote:
> On Tue 11-07-23 11:42:56, Liu Song wrote:
>> In the append-write scenario, after ensuring that the dirty inode can be
>> seen by the writeback process, there is no need to execute
>> "mark_inode_dirty" for every write. Instead, we can rely on
>> "ext4_mark_inode_dirty" executed when updating i_disksize in
>> "mpage_map_and_submit_extent" to ensure data consistency, which can
>> significantly improve performance in high-frequency append-write
>> scenarios.
>>
>> In test scenarios of Kafka version 2.6.2, using packet size of 2K
>> resulted in a 10% performance improvement.
>>
>> Signed-off-by: Liu Song <[email protected]>
> Overall the benefit looks interesting. Nice work.
>
>> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
>> index 9d9f414f99fe..d1aa775c9936 100644
>> --- a/fs/ext4/inode.c
>> +++ b/fs/ext4/inode.c
>> @@ -3128,6 +3128,57 @@ static int ext4_da_should_update_i_disksize(struct page *page,
>> return 1;
>> }
>>
>> +/*
>> + * Copy from generic_write_end, add conditions to execute mark_inode_dirty
>> + * to avoid additional overhead caused by frequent dirty inode operations
>> + */
>> +static int ext4_da_generic_write_end(struct file *file, struct address_space *mapping,
>> + loff_t pos, unsigned len, unsigned copied, bool need_dirty,
>> + struct page *page, void *fsdata)
>> +{
>> + struct inode *inode = mapping->host;
>> + loff_t old_size = inode->i_size;
>> + bool i_size_changed = false;
>> + int was_dirty;
>> +
>> + copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
>> +
>> + /*
>> + * No need to use i_size_read() here, the i_size cannot change under us
>> + * because we hold i_rwsem.
>> + *
>> + * But it's important to update i_size while still holding page lock:
>> + * page writeout could otherwise come in and zero beyond i_size.
>> + */
>> + if (pos + copied > inode->i_size) {
>> + i_size_write(inode, pos + copied);
>> + i_size_changed = true;
>> + }
>> +
>> + unlock_page(page);
>> + put_page(page);
>> +
>> + if (old_size < pos)
>> + pagecache_isize_extended(inode, old_size, pos);
>> +
> I dislike the duplication of generic_write_end() but exporting a variant of
> generic_write_end() not doing the dirtying (and additionally returning the
> i_size_changed value) doesn't look appealing either. So I guess I'll just
> live with this...
>
>> + /*
>> + * In the append-write scenario, if the inode is marked as dirty,
>> + * it is ensured that the inode will be seen by the writeback process.
>> + * In the ext4_writepages process, when updating i_disksize,
>> + * corresponding metadata updates are also performed.
>> + * Therefore, it is unnecessary to repeatedly execute mark_inode_dirty
>> + * to improve performance.
>> + */
> Note that block_write_end() will mark the inode as dirty with I_DIRTY_PAGES
> flag (which all that's needed to trigger page writeback). Since i_size is
> never really written to disk by ext4, there's simply never a need to mark
> inode dirty because of that. Also if inode timestamps were updated inode
> was already properly marked dirty. Hence we can just mostly drop the code
> below.
>
>> + if (i_size_changed) {
>> + spin_lock(&inode->i_lock);
>> + was_dirty = inode->i_state & I_DIRTY;
>> + spin_unlock(&inode->i_lock);
>> + if (!was_dirty || need_dirty)
>> + mark_inode_dirty(inode);
>> + }
>> + return copied;
>> +}
>> +
>> static int ext4_da_write_end(struct file *file,
>> struct address_space *mapping,
>> loff_t pos, unsigned len, unsigned copied,
>> @@ -3137,6 +3188,7 @@ static int ext4_da_write_end(struct file *file,
>> loff_t new_i_size;
>> unsigned long start, end;
>> int write_mode = (int)(unsigned long)fsdata;
>> + bool need_dirty = false;
>>
>> if (write_mode == FALL_BACK_TO_NONDELALLOC)
>> return ext4_write_end(file, mapping, pos,
>> @@ -3169,10 +3221,12 @@ static int ext4_da_write_end(struct file *file,
>> */
>> new_i_size = pos + copied;
>> if (copied && new_i_size > inode->i_size &&
>> - ext4_da_should_update_i_disksize(page, end))
>> + ext4_da_should_update_i_disksize(page, end)) {
>> ext4_update_i_disksize(inode, new_i_size);
>> + need_dirty = true;
>> + }
Thank you very much for your response, and the implementation of
"ext4_da_do_write_end" is great.
I will refer to your suggestions and send a V2 version of the patch.
Thanks
> So when we create our own new helper function anyway I'd just move all this
> logic for delalloc write end there. Something like:
>
> static int ext4_da_do_write_end(struct address_space *mapping,
> loff_t pos, unsigned len, unsigned copied,
> struct page *page)
> {
> struct inode *inode = mapping->host;
> loff_t old_size = inode->i_size;
> bool disksize_changed = false;
> loff_t new_i_size;
>
> copied = block_write_end(NULL, mapping, pos, len, copied, page, NULL);
>
> new_i_size = pos + copied;
> /*
> * It's important to update i_size while still holding page lock:
> * page writeout could otherwise come in and zero beyond i_size.
> *
> * Since we are holding inode lock, we are sure i_disksize <=
> * i_size. We also know that if i_disksize < i_size, there are
> * delalloc writes pending in the range upto i_size. If the end of
> * the current write is <= i_size, there's no need to touch
> * i_disksize since writeback will push i_disksize upto i_size
> * eventually. If the end of the current write is > i_size and
> * inside an allocated block (ext4_da_should_update_i_disksize()
> * check), we need to update i_disksize here as certain
> * ext4_writepages() paths not allocating blocks update i_disksize.
> */
> if (new_i_size > inode->i_size) {
> unsigned long end;
>
> i_size_write(inode, new_i_size);
> end = (new_i_size - 1) & (PAGE_SIZE - 1);
> if (copied && ext4_da_should_update_i_disksize(folio, end)) {
> ext4_update_i_disksize(inode, new_i_size);
> disksize_changed = true;
> }
> }
>
> unlock_page(page);
> put_page(page);
>
> if (old_size < pos)
> pagecache_isize_extended(inode, old_size, pos);
>
> if (disksize_changed) {
> handle_t *handle;
>
> handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
> if (IS_ERR(handle))
> return PTR_ERR(handle);
> ext4_mark_inode_dirty(handle, inode);
> ext4_journal_stop(handle);
> }
>
> return copied;
> }
>
>
> Honza