commit 9dc55f1389f9 ("iomap: add support for sub-pagesize buffered I/O
without buffer heads") replace the per-block structure buffer_head with
the per-page structure iomap_page. However, iomap_page can't track the
dirty state of sub pages, which will cause performance issue since sub
pages will be writeback even if they are not dirty.
For example, if block size is 4k and page size is 64k:
dd if=/dev/zero of=testfile bs=4k count=16 oflag=sync
With buffer_head implementation, the above dd cmd will writeback 4k in
each round. However, with iomap_page implementation, the range of
writeback in each round is from the start of the page to the end offset
we just wrote.
Thus add support to track dirty state for sub pages in iomap_page.
Signed-off-by: Yu Kuai <[email protected]>
---
fs/iomap/buffered-io.c | 51 +++++++++++++++++++++++++++++++++++++++++-
1 file changed, 50 insertions(+), 1 deletion(-)
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index bcfc288dba3f..ac2676146b98 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -29,7 +29,9 @@ struct iomap_page {
atomic_t read_count;
atomic_t write_count;
spinlock_t uptodate_lock;
+ spinlock_t dirty_lock;
DECLARE_BITMAP(uptodate, PAGE_SIZE / 512);
+ DECLARE_BITMAP(dirty, PAGE_SIZE / 512);
};
static inline struct iomap_page *to_iomap_page(struct page *page)
@@ -53,7 +55,9 @@ iomap_page_create(struct inode *inode, struct page *page)
atomic_set(&iop->read_count, 0);
atomic_set(&iop->write_count, 0);
spin_lock_init(&iop->uptodate_lock);
+ spin_lock_init(&iop->dirty_lock);
bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
+ bitmap_zero(iop->dirty, PAGE_SIZE / SECTOR_SIZE);
/*
* migrate_page_move_mapping() assumes that pages with private data have
@@ -135,6 +139,44 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
*lenp = plen;
}
+static void
+iomap_iop_set_or_clear_range_dirty(
+ struct page *page,
+ unsigned int off,
+ unsigned int len,
+ bool is_set)
+{
+ struct iomap_page *iop = to_iomap_page(page);
+ struct inode *inode = page->mapping->host;
+ unsigned int first = off >> inode->i_blkbits;
+ unsigned int last = (off + len - 1) >> inode->i_blkbits;
+ unsigned long flags;
+ unsigned int i;
+
+ spin_lock_irqsave(&iop->dirty_lock, flags);
+ for (i = first; i <= last; i++)
+ if (is_set)
+ set_bit(i, iop->dirty);
+ else
+ clear_bit(i, iop->dirty);
+ spin_unlock_irqrestore(&iop->dirty_lock, flags);
+}
+
+static void
+iomap_set_or_clear_range_dirty(
+ struct page *page,
+ unsigned int off,
+ unsigned int len,
+ bool is_set)
+{
+ if (PageError(page))
+ return;
+
+ if (page_has_private(page))
+ iomap_iop_set_or_clear_range_dirty(
+ page, off, len, is_set);
+}
+
static void
iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
{
@@ -705,6 +747,8 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
if (unlikely(copied < len && !PageUptodate(page)))
return 0;
iomap_set_range_uptodate(page, offset_in_page(pos), len);
+ iomap_set_or_clear_range_dirty(
+ page, offset_in_page(pos), len, true);
iomap_set_page_dirty(page);
return copied;
}
@@ -1030,6 +1074,8 @@ iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
WARN_ON_ONCE(!PageUptodate(page));
iomap_page_create(inode, page);
set_page_dirty(page);
+ iomap_set_or_clear_range_dirty(
+ page, offset_in_page(pos), length, true);
}
return length;
@@ -1386,7 +1432,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
for (i = 0, file_offset = page_offset(page);
i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
i++, file_offset += len) {
- if (iop && !test_bit(i, iop->uptodate))
+ if (iop && (!test_bit(i, iop->uptodate) ||
+ !test_bit(i, iop->dirty)))
continue;
error = wpc->ops->map_blocks(wpc, inode, file_offset);
@@ -1435,6 +1482,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
*/
set_page_writeback_keepwrite(page);
} else {
+ iomap_set_or_clear_range_dirty(
+ page, 0, end_offset - page_offset(page) + 1, false);
clear_page_dirty_for_io(page);
set_page_writeback(page);
}
--
2.25.4
Hi Kuai,
On Thu, Jul 30, 2020 at 09:19:01AM +0800, Yu Kuai wrote:
> commit 9dc55f1389f9 ("iomap: add support for sub-pagesize buffered I/O
> without buffer heads") replace the per-block structure buffer_head with
> the per-page structure iomap_page. However, iomap_page can't track the
> dirty state of sub pages, which will cause performance issue since sub
> pages will be writeback even if they are not dirty.
>
> For example, if block size is 4k and page size is 64k:
>
> dd if=/dev/zero of=testfile bs=4k count=16 oflag=sync
>
> With buffer_head implementation, the above dd cmd will writeback 4k in
> each round. However, with iomap_page implementation, the range of
> writeback in each round is from the start of the page to the end offset
> we just wrote.
>
> Thus add support to track dirty state for sub pages in iomap_page.
AFAIK, the current focus is also on the numbers in the original
discussion thread, so it'd be better to show some numbers with
large PAGE_SIZE on this with some workloads as well.
https://lore.kernel.org/r/[email protected]
e.g. My guess is if the dirty blocks in the page are highly fragmented, maybe
it'd be better to writeback the whole page in an IO rather than individual blocks.
At a very quick glance, the approach looks good to me.
Thanks,
Gao Xiang
>
> Signed-off-by: Yu Kuai <[email protected]>
> ---
> fs/iomap/buffered-io.c | 51 +++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 50 insertions(+), 1 deletion(-)
>
> diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
> index bcfc288dba3f..ac2676146b98 100644
> --- a/fs/iomap/buffered-io.c
> +++ b/fs/iomap/buffered-io.c
> @@ -29,7 +29,9 @@ struct iomap_page {
> atomic_t read_count;
> atomic_t write_count;
> spinlock_t uptodate_lock;
> + spinlock_t dirty_lock;
> DECLARE_BITMAP(uptodate, PAGE_SIZE / 512);
> + DECLARE_BITMAP(dirty, PAGE_SIZE / 512);
> };
>
> static inline struct iomap_page *to_iomap_page(struct page *page)
> @@ -53,7 +55,9 @@ iomap_page_create(struct inode *inode, struct page *page)
> atomic_set(&iop->read_count, 0);
> atomic_set(&iop->write_count, 0);
> spin_lock_init(&iop->uptodate_lock);
> + spin_lock_init(&iop->dirty_lock);
> bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
> + bitmap_zero(iop->dirty, PAGE_SIZE / SECTOR_SIZE);
>
> /*
> * migrate_page_move_mapping() assumes that pages with private data have
> @@ -135,6 +139,44 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
> *lenp = plen;
> }
>
> +static void
> +iomap_iop_set_or_clear_range_dirty(
> + struct page *page,
> + unsigned int off,
> + unsigned int len,
> + bool is_set)
> +{
> + struct iomap_page *iop = to_iomap_page(page);
> + struct inode *inode = page->mapping->host;
> + unsigned int first = off >> inode->i_blkbits;
> + unsigned int last = (off + len - 1) >> inode->i_blkbits;
> + unsigned long flags;
> + unsigned int i;
> +
> + spin_lock_irqsave(&iop->dirty_lock, flags);
> + for (i = first; i <= last; i++)
> + if (is_set)
> + set_bit(i, iop->dirty);
> + else
> + clear_bit(i, iop->dirty);
> + spin_unlock_irqrestore(&iop->dirty_lock, flags);
> +}
> +
> +static void
> +iomap_set_or_clear_range_dirty(
> + struct page *page,
> + unsigned int off,
> + unsigned int len,
> + bool is_set)
> +{
> + if (PageError(page))
> + return;
> +
> + if (page_has_private(page))
> + iomap_iop_set_or_clear_range_dirty(
> + page, off, len, is_set);
3> +}
> +
> static void
> iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
> {
> @@ -705,6 +747,8 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
> if (unlikely(copied < len && !PageUptodate(page)))
> return 0;
> iomap_set_range_uptodate(page, offset_in_page(pos), len);
> + iomap_set_or_clear_range_dirty(
> + page, offset_in_page(pos), len, true);
> iomap_set_page_dirty(page);
> return copied;
> }
> @@ -1030,6 +1074,8 @@ iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
> WARN_ON_ONCE(!PageUptodate(page));
> iomap_page_create(inode, page);
> set_page_dirty(page);
> + iomap_set_or_clear_range_dirty(
> + page, offset_in_page(pos), length, true);
> }
>
> return length;
> @@ -1386,7 +1432,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
> for (i = 0, file_offset = page_offset(page);
> i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
> i++, file_offset += len) {
> - if (iop && !test_bit(i, iop->uptodate))
> + if (iop && (!test_bit(i, iop->uptodate) ||
> + !test_bit(i, iop->dirty)))
> continue;
>
> error = wpc->ops->map_blocks(wpc, inode, file_offset);
> @@ -1435,6 +1482,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
> */
> set_page_writeback_keepwrite(page);
> } else {
> + iomap_set_or_clear_range_dirty(
> + page, 0, end_offset - page_offset(page) + 1, false);
> clear_page_dirty_for_io(page);
> set_page_writeback(page);
> }
> --
> 2.25.4
>
On 2020/7/30 10:27, Gao Xiang wrote:
> Hi Kuai,
>
> On Thu, Jul 30, 2020 at 09:19:01AM +0800, Yu Kuai wrote:
>> commit 9dc55f1389f9 ("iomap: add support for sub-pagesize buffered I/O
>> without buffer heads") replace the per-block structure buffer_head with
>> the per-page structure iomap_page. However, iomap_page can't track the
>> dirty state of sub pages, which will cause performance issue since sub
>> pages will be writeback even if they are not dirty.
>>
>> For example, if block size is 4k and page size is 64k:
>>
>> dd if=/dev/zero of=testfile bs=4k count=16 oflag=sync
>>
>> With buffer_head implementation, the above dd cmd will writeback 4k in
>> each round. However, with iomap_page implementation, the range of
>> writeback in each round is from the start of the page to the end offset
>> we just wrote.
>>
>> Thus add support to track dirty state for sub pages in iomap_page.
>
> AFAIK, the current focus is also on the numbers in the original
> discussion thread, so it'd be better to show some numbers with
> large PAGE_SIZE on this with some workloads as well.
>
> https://lore.kernel.org/r/[email protected]
>
Hi, Xiang!
The problem was found by iozone to test 4k sequintail write in my
case, thanks for pointing out the discussion thread. I'll test it if
this patch have any effect on that situation.
Thanks,
Yu Kuai
> e.g. My guess is if the dirty blocks in the page are highly fragmented, maybe
> it'd be better to writeback the whole page in an IO rather than individual blocks.
>
> At a very quick glance, the approach looks good to me.
>
> Thanks,
> Gao Xiang
>
>>
>> Signed-off-by: Yu Kuai <[email protected]>
>> ---
>> fs/iomap/buffered-io.c | 51 +++++++++++++++++++++++++++++++++++++++++-
>> 1 file changed, 50 insertions(+), 1 deletion(-)
>>
>> diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
>> index bcfc288dba3f..ac2676146b98 100644
>> --- a/fs/iomap/buffered-io.c
>> +++ b/fs/iomap/buffered-io.c
>> @@ -29,7 +29,9 @@ struct iomap_page {
>> atomic_t read_count;
>> atomic_t write_count;
>> spinlock_t uptodate_lock;
>> + spinlock_t dirty_lock;
>> DECLARE_BITMAP(uptodate, PAGE_SIZE / 512);
>> + DECLARE_BITMAP(dirty, PAGE_SIZE / 512);
>> };
>>
>> static inline struct iomap_page *to_iomap_page(struct page *page)
>> @@ -53,7 +55,9 @@ iomap_page_create(struct inode *inode, struct page *page)
>> atomic_set(&iop->read_count, 0);
>> atomic_set(&iop->write_count, 0);
>> spin_lock_init(&iop->uptodate_lock);
>> + spin_lock_init(&iop->dirty_lock);
>> bitmap_zero(iop->uptodate, PAGE_SIZE / SECTOR_SIZE);
>> + bitmap_zero(iop->dirty, PAGE_SIZE / SECTOR_SIZE);
>>
>> /*
>> * migrate_page_move_mapping() assumes that pages with private data have
>> @@ -135,6 +139,44 @@ iomap_adjust_read_range(struct inode *inode, struct iomap_page *iop,
>> *lenp = plen;
>> }
>>
>> +static void
>> +iomap_iop_set_or_clear_range_dirty(
>> + struct page *page,
>> + unsigned int off,
>> + unsigned int len,
>> + bool is_set)
>> +{
>> + struct iomap_page *iop = to_iomap_page(page);
>> + struct inode *inode = page->mapping->host;
>> + unsigned int first = off >> inode->i_blkbits;
>> + unsigned int last = (off + len - 1) >> inode->i_blkbits;
>> + unsigned long flags;
>> + unsigned int i;
>> +
>> + spin_lock_irqsave(&iop->dirty_lock, flags);
>> + for (i = first; i <= last; i++)
>> + if (is_set)
>> + set_bit(i, iop->dirty);
>> + else
>> + clear_bit(i, iop->dirty);
>> + spin_unlock_irqrestore(&iop->dirty_lock, flags);
>> +}
>> +
>> +static void
>> +iomap_set_or_clear_range_dirty(
>> + struct page *page,
>> + unsigned int off,
>> + unsigned int len,
>> + bool is_set)
>> +{
>> + if (PageError(page))
>> + return;
>> +
>> + if (page_has_private(page))
>> + iomap_iop_set_or_clear_range_dirty(
>> + page, off, len, is_set);
> 3> +}
>> +
>> static void
>> iomap_iop_set_range_uptodate(struct page *page, unsigned off, unsigned len)
>> {
>> @@ -705,6 +747,8 @@ __iomap_write_end(struct inode *inode, loff_t pos, unsigned len,
>> if (unlikely(copied < len && !PageUptodate(page)))
>> return 0;
>> iomap_set_range_uptodate(page, offset_in_page(pos), len);
>> + iomap_set_or_clear_range_dirty(
>> + page, offset_in_page(pos), len, true);
>> iomap_set_page_dirty(page);
>> return copied;
>> }
>> @@ -1030,6 +1074,8 @@ iomap_page_mkwrite_actor(struct inode *inode, loff_t pos, loff_t length,
>> WARN_ON_ONCE(!PageUptodate(page));
>> iomap_page_create(inode, page);
>> set_page_dirty(page);
>> + iomap_set_or_clear_range_dirty(
>> + page, offset_in_page(pos), length, true);
>> }
>>
>> return length;
>> @@ -1386,7 +1432,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
>> for (i = 0, file_offset = page_offset(page);
>> i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
>> i++, file_offset += len) {
>> - if (iop && !test_bit(i, iop->uptodate))
>> + if (iop && (!test_bit(i, iop->uptodate) ||
>> + !test_bit(i, iop->dirty)))
>> continue;
>>
>> error = wpc->ops->map_blocks(wpc, inode, file_offset);
>> @@ -1435,6 +1482,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
>> */
>> set_page_writeback_keepwrite(page);
>> } else {
>> + iomap_set_or_clear_range_dirty(
>> + page, 0, end_offset - page_offset(page) + 1, false);
>> clear_page_dirty_for_io(page);
>> set_page_writeback(page);
>> }
>> --
>> 2.25.4
>>
>
>
> .
>
On Thu, Jul 30, 2020 at 09:19:01AM +0800, Yu Kuai wrote:
> +++ b/fs/iomap/buffered-io.c
> @@ -29,7 +29,9 @@ struct iomap_page {
> atomic_t read_count;
> atomic_t write_count;
> spinlock_t uptodate_lock;
> + spinlock_t dirty_lock;
No need for a separate spinlock. Just rename uptodate_lock. Maybe
'bitmap_lock'.
> DECLARE_BITMAP(uptodate, PAGE_SIZE / 512);
> + DECLARE_BITMAP(dirty, PAGE_SIZE / 512);
This is inefficient and poses difficulties for the THP patchset.
Maybe let the discussion on removing the ->uptodate array finish
before posting another patch for review?
> +static void
> +iomap_iop_set_or_clear_range_dirty(
> + struct page *page,
> + unsigned int off,
> + unsigned int len,
> + bool is_set)
Please follow normal kernel programming style. This isn't XFS.
Also 'set or clear' with a bool to indicate which to do is horrible
style. Separate functions!
> @@ -1386,7 +1432,8 @@ iomap_writepage_map(struct iomap_writepage_ctx *wpc,
> for (i = 0, file_offset = page_offset(page);
> i < (PAGE_SIZE >> inode->i_blkbits) && file_offset < end_offset;
> i++, file_offset += len) {
> - if (iop && !test_bit(i, iop->uptodate))
> + if (iop && (!test_bit(i, iop->uptodate) ||
> + !test_bit(i, iop->dirty)))
> continue;
Surely we don't need to test ->uptodate here at all. Why would we write
back a block which isn't dirty?
On 2020/7/30 11:19, Matthew Wilcox wrote:
> Maybe let the discussion on removing the ->uptodate array finish
> before posting another patch for review?
Hi, Matthew!
Of course, I missed the discussion thread before sending this path.
And thanks for your suggestions.
Best regards,
Yu Kuai
On Thu, Jul 30, 2020 at 04:19:34AM +0100, Matthew Wilcox wrote:
> On Thu, Jul 30, 2020 at 09:19:01AM +0800, Yu Kuai wrote:
> > +++ b/fs/iomap/buffered-io.c
> > @@ -29,7 +29,9 @@ struct iomap_page {
> > atomic_t read_count;
> > atomic_t write_count;
> > spinlock_t uptodate_lock;
> > + spinlock_t dirty_lock;
>
> No need for a separate spinlock. Just rename uptodate_lock. Maybe
> 'bitmap_lock'.
Agreed.
>
> > DECLARE_BITMAP(uptodate, PAGE_SIZE / 512);
> > + DECLARE_BITMAP(dirty, PAGE_SIZE / 512);
>
> This is inefficient and poses difficulties for the THP patchset.
> Maybe let the discussion on removing the ->uptodate array finish
> before posting another patch for review?
I really don't think we can kill the uptodate bit. But what we can
do is have on bitmap array (flex size as in your prep patches) and just
alternating bits for uptodate and dirty.