ra->prev_pos tracks the last visited byte in the previous read request.
It is used to check whether it is sequential read in
ondemand_readahead and thus affects the readahead window.
From commit 06c0444290ce ("mm/filemap.c: generic_file_buffered_read()
now uses find_get_pages_contig"),update logic of prev_pos is changed.
It updates prev_pos after each returns from filemap_get_pages.
But the read request from user may be not fully completed
at this point.
The updated prev_pos impacts the subsequent readahead window.
The real problem is performance drop of fsck_msdos between linux-5.4
and linux-5.15(also linux-6.4).
Comparing to linux-5.4,It spends about 110% time and read 140% pages.
The read pattern of fsck_msdos is not fully sequential.
Simplified read pattern of fsck_msdos likes below:
1.read at page offset 0xa,size 0x1000
2.read at other page offset like 0x20,size 0x1000
3.read at page offset 0xa,size 0x4000
4.read at page offset 0xe,size 0x1000
Here is the read status on linux-6.4:
1.after read at page offset 0xa,size 0x1000
->page ofs 0xa go into pagecache
2.after read at page offset 0x20,size 0x1000
->page ofs 0x20 go into pagecache
3.read at page offset 0xa,size 0x4000
->filemap_get_pages read ofs 0xa from pagecache and returns
->prev_pos is updated to 0xb and goto next loop
->filemap_get_pages tends to read ofs 0xb,size 0x3000
->initial_readahead case in ondemand_readahead since prev_pos is
the same as request ofs.
->read 8 pages while async size is 5 pages
(PageReadahead flag at page 0xe)
4.read at page offset 0xe,size 0x1000
->hit page 0xe with PageReadahead flag set,double the ra_size.
read 16 pages while async size is 16 pages
Now it reads 24 pages while actually uses 5 pages
on linux-5.4:
1.the same as 6.4
2.the same as 6.4
3.read at page offset 0xa,size 0x4000
->read ofs 0xa from pagecache
->read ofs 0xb,size 0x3000 using page_cache_sync_readahead
read 3 pages
->prev_pos is updated to 0xd before generic_file_buffered_read
returns
4.read at page offset 0xe,size 0x1000
->initial_readahead case in ondemand_readahead since
request ofs-prev_pos==1
->read 4 pages while async size is 3 pages
Now it reads 7 pages while actually uses 5 pages.
In above demo,the initial_readahead case is triggered by offset
of user request on linux-5.4.
While it may be triggered by update logic of prev_pos on linux-6.4.
To fix the performance drop,update prev_pos after finishing one read
request.
Signed-off-by: Haibo Li <[email protected]>
---
mm/filemap.c | 9 +++++----
1 file changed, 5 insertions(+), 4 deletions(-)
diff --git a/mm/filemap.c b/mm/filemap.c
index 83dda76d1fc3..16b2054eee71 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2670,6 +2670,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
int i, error = 0;
bool writably_mapped;
loff_t isize, end_offset;
+ loff_t last_pos = ra->prev_pos;
if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
return 0;
@@ -2721,8 +2722,8 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
* When a read accesses the same folio several times, only
* mark it as accessed the first time.
*/
- if (!pos_same_folio(iocb->ki_pos, ra->prev_pos - 1,
- fbatch.folios[0]))
+ if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
+ fbatch.folios[0]))
folio_mark_accessed(fbatch.folios[0]);
for (i = 0; i < folio_batch_count(&fbatch); i++) {
@@ -2749,7 +2750,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
already_read += copied;
iocb->ki_pos += copied;
- ra->prev_pos = iocb->ki_pos;
+ last_pos = iocb->ki_pos;
if (copied < bytes) {
error = -EFAULT;
@@ -2763,7 +2764,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
} while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
file_accessed(filp);
-
+ ra->prev_pos = last_pos;
return already_read ? already_read : error;
}
EXPORT_SYMBOL_GPL(filemap_read);
--
2.25.1
On Wed 28-06-23 19:02:20, Haibo Li wrote:
> ra->prev_pos tracks the last visited byte in the previous read request.
> It is used to check whether it is sequential read in
> ondemand_readahead and thus affects the readahead window.
>
> From commit 06c0444290ce ("mm/filemap.c: generic_file_buffered_read()
> now uses find_get_pages_contig"),update logic of prev_pos is changed.
> It updates prev_pos after each returns from filemap_get_pages.
> But the read request from user may be not fully completed
> at this point.
> The updated prev_pos impacts the subsequent readahead window.
>
> The real problem is performance drop of fsck_msdos between linux-5.4
> and linux-5.15(also linux-6.4).
> Comparing to linux-5.4,It spends about 110% time and read 140% pages.
> The read pattern of fsck_msdos is not fully sequential.
>
> Simplified read pattern of fsck_msdos likes below:
> 1.read at page offset 0xa,size 0x1000
> 2.read at other page offset like 0x20,size 0x1000
> 3.read at page offset 0xa,size 0x4000
> 4.read at page offset 0xe,size 0x1000
>
> Here is the read status on linux-6.4:
> 1.after read at page offset 0xa,size 0x1000
> ->page ofs 0xa go into pagecache
> 2.after read at page offset 0x20,size 0x1000
> ->page ofs 0x20 go into pagecache
> 3.read at page offset 0xa,size 0x4000
> ->filemap_get_pages read ofs 0xa from pagecache and returns
> ->prev_pos is updated to 0xb and goto next loop
> ->filemap_get_pages tends to read ofs 0xb,size 0x3000
> ->initial_readahead case in ondemand_readahead since prev_pos is
> the same as request ofs.
> ->read 8 pages while async size is 5 pages
> (PageReadahead flag at page 0xe)
> 4.read at page offset 0xe,size 0x1000
> ->hit page 0xe with PageReadahead flag set,double the ra_size.
> read 16 pages while async size is 16 pages
> Now it reads 24 pages while actually uses 5 pages
>
> on linux-5.4:
> 1.the same as 6.4
> 2.the same as 6.4
> 3.read at page offset 0xa,size 0x4000
> ->read ofs 0xa from pagecache
> ->read ofs 0xb,size 0x3000 using page_cache_sync_readahead
> read 3 pages
> ->prev_pos is updated to 0xd before generic_file_buffered_read
> returns
> 4.read at page offset 0xe,size 0x1000
> ->initial_readahead case in ondemand_readahead since
> request ofs-prev_pos==1
> ->read 4 pages while async size is 3 pages
>
> Now it reads 7 pages while actually uses 5 pages.
>
> In above demo,the initial_readahead case is triggered by offset
> of user request on linux-5.4.
> While it may be triggered by update logic of prev_pos on linux-6.4.
>
> To fix the performance drop,update prev_pos after finishing one read
> request.
>
> Signed-off-by: Haibo Li <[email protected]>
Sorry for the delayed reply. This seems to have fallen through the cracks.
So if I understand your analysis right, you are complaining that random
read larger than 1 page gets misdetected as sequential read and so "larger
than necessary" readahead happens. I tend to agree with your opinion and your
solution looks good to me. Feel free to add:
Reviewed-by: Jan Kara <[email protected]>
Willy, any opinion? Andrew, can you pickup the patch if Willy doesn't
object?
Honza
> ---
> mm/filemap.c | 9 +++++----
> 1 file changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/mm/filemap.c b/mm/filemap.c
> index 83dda76d1fc3..16b2054eee71 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -2670,6 +2670,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
> int i, error = 0;
> bool writably_mapped;
> loff_t isize, end_offset;
> + loff_t last_pos = ra->prev_pos;
>
> if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
> return 0;
> @@ -2721,8 +2722,8 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
> * When a read accesses the same folio several times, only
> * mark it as accessed the first time.
> */
> - if (!pos_same_folio(iocb->ki_pos, ra->prev_pos - 1,
> - fbatch.folios[0]))
> + if (!pos_same_folio(iocb->ki_pos, last_pos - 1,
> + fbatch.folios[0]))
> folio_mark_accessed(fbatch.folios[0]);
>
> for (i = 0; i < folio_batch_count(&fbatch); i++) {
> @@ -2749,7 +2750,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
>
> already_read += copied;
> iocb->ki_pos += copied;
> - ra->prev_pos = iocb->ki_pos;
> + last_pos = iocb->ki_pos;
>
> if (copied < bytes) {
> error = -EFAULT;
> @@ -2763,7 +2764,7 @@ ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
> } while (iov_iter_count(iter) && iocb->ki_pos < isize && !error);
>
> file_accessed(filp);
> -
> + ra->prev_pos = last_pos;
> return already_read ? already_read : error;
> }
> EXPORT_SYMBOL_GPL(filemap_read);
> --
> 2.25.1
>
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Wed, 9 Aug 2023 18:44:46 +0200 Jan Kara <[email protected]> wrote:
> Willy, any opinion? Andrew, can you pickup the patch if Willy doesn't
> object?
I added it to mm.git on July 2.
On Wed 09-08-23 11:45:07, Andrew Morton wrote:
> On Wed, 9 Aug 2023 18:44:46 +0200 Jan Kara <[email protected]> wrote:
>
> > Willy, any opinion? Andrew, can you pickup the patch if Willy doesn't
> > object?
>
> I added it to mm.git on July 2.
Ah, sorry for the noise then. I should have checked that. Thanks!
Honza
--
Jan Kara <[email protected]>
SUSE Labs, CR