md: stop using do_sync_mapping_range
PATCH 1/1 md: stop using do_sync_mapping_range
It's a very awkward way to write out all data and wait for it, so just call filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.
drivers/md/bitmap.c +1 -4
@@ -1621,10 +1621,7 @@ int bitmap_create(mddev_t *mddev) bitmap->offset = mddev->bitmap_offset;if (file) {get_file (file); -do_sync_mapping_range (file->f_mapping,0 ,LLONG_MAX , -SYNC_FILE_RANGE_WAIT_BEFORE | -SYNC_FILE_RANGE_WRITE | -SYNC_FILE_RANGE_WAIT_AFTER ); +filemap_write_and_wait (file->f_mapping); }/* read superblock from bitmap file (this sets bitmap->chunksize) */ err =bitmap_read_sb (bitmap);
Quoted 9 lines · Christoph Hellwig
drivers/md/bitmap.c +1
-4-do_sync_mapping_range (file->f_mapping,0 ,LLONG_MAX , -SYNC_FILE_RANGE_WAIT_BEFORE | -SYNC_FILE_RANGE_WRITE | -SYNC_FILE_RANGE_WAIT_AFTER ); +filemap_write_and_wait (file->f_mapping);It's a very awkward way to write out all data and wait for it, so just call
filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.
The SYNC_FILE_RANGE_WAIT_BEFORE is to make sure that writes which were started before the most recent dirtying are completed before initiating a second write on the same pages to make sure the most recently dirty data is written.
Does filemap_write_and_wait() do that?
Fwiw, It's partially redundant: it isn't necessary to have a second write when the first write request was queued in the elevator but hadn't yet reached the device.
-- Jamie
Quoted 3 lines · On Wednesday September 23, [email protected] wrote:
It's a very awkward way to write out all data and wait for it, so just call
filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.
When md/bitmap accesses a file, it uses bmap to find addresses and then submit_bh to do IO, so it completely by-passes the page cache. So this code is present to ensure that the page cache has no dirty pages for the file before we start using the file.
I don't recall exactly why I used do_sync_mapping_range. I suspect that I looked at what "sys_fsync" used (do_fsync?) and found that wasn't exported, so I looked at what sys_sync_file_range used, found that was exported, and so used that.
Looking at the current state of the VFS, I think I would rather use vfs_fsync.
So would you be happy with something like the following?
Thanks,
NeilBrown
drivers/md/bitmap.c +5 -4
@@ -1624,10 +1624,11 @@ int bitmap_create(mddev_t *mddev) bitmap->offset = mddev->bitmap_offset;if (file) {get_file (file); -do_sync_mapping_range (file->f_mapping,0 ,LLONG_MAX , -SYNC_FILE_RANGE_WAIT_BEFORE | -SYNC_FILE_RANGE_WRITE | -SYNC_FILE_RANGE_WAIT_AFTER ); +/* As future accesses to this file will use bmap, +* and bypass the page cache, we must sync the file +* first. +*/ +vfs_fsync (file, file->f_dentry,1 ); }/* read superblock from bitmap file (this sets bitmap->chunksize) */ err =bitmap_read_sb (bitmap);
Quoted 16 lines · Jamie Lokier
Quoted 9 lines · Christoph Hellwig
drivers/md/bitmap.c +1
-4-do_sync_mapping_range (file->f_mapping,0 ,LLONG_MAX , -SYNC_FILE_RANGE_WAIT_BEFORE | -SYNC_FILE_RANGE_WRITE | -SYNC_FILE_RANGE_WAIT_AFTER ); +filemap_write_and_wait (file->f_mapping);It's a very awkward way to write out all data and wait for it, so just call
filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.The
SYNC_FILE_RANGE_WAIT_BEFOREis to make sure that writes which were started before the most recent dirtying are completed before initiating a second write on the same pages to make sure the most recently dirty data is written.
Quoted 1 line
Does
filemap_write_and_wait()do that?
Yes, for WB_SYNC_ALL syncs write_cache_pages or the fs-specific equivalent waits for pages marked writeback before starting starting the writeout.
Quoted 10 lines · Neil Brown
Quoted 3 lines · On Wednesday September 23, [email protected] wrote:
It's a very awkward way to write out all data and wait for it, so just call
filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.When md/bitmap accesses a file, it uses bmap to find addresses and then
submit_bhto do IO, so it completely by-passes the page cache. So this code is present to ensure that the page cache has no dirty pages for the file before we start using the file.
Quoted 7 lines
I don't recall exactly why I used
do_sync_mapping_range. I suspect that I looked at what "sys_fsync" used (do_fsync?) and found that wasn't exported, so I looked at whatsys_sync_file_rangeused, found that was exported, and so used that.Looking at the current state of the VFS, I think I would rather use
vfs_fsync.
Using vfs_fsync is better, but using bmap is extremly dangerous, I'm kinda suprised how this managed to get sneaked in without notice. bmap really is just an advisory interface as the mappings by change underneath all the time. Some modern filesystems like btrfs really can't implement it at all.
Quoted 1 line
So would you be happy with something like the following?
It's at least a small improvement..
Quoted 16 lines · Christoph Hellwig
Quoted 15 lines · Neil Brown
Quoted 3 lines · On Wednesday September 23, [email protected] wrote:
It's a very awkward way to write out all data and wait for it, so just call
filemap_write_and_wait. I still can't figure what the point of all this is, so a comment would surely be helpful.When md/bitmap accesses a file, it uses bmap to find addresses and then
submit_bhto do IO, so it completely by-passes the page cache. So this code is present to ensure that the page cache has no dirty pages for the file before we start using the file.I don't recall exactly why I used
do_sync_mapping_range. I suspect that I looked at what "sys_fsync" used (do_fsync?) and found that wasn't exported, so I looked at whatsys_sync_file_rangeused, found that was exported, and so used that.
Quoted 9 lines
Quoted 2 lines
Looking at the current state of the VFS, I think I would rather use
vfs_fsync.Using
vfs_fsyncis better, but using bmap is extremly dangerous, I'm kinda suprised how this managed to get sneaked in without notice. bmap really is just an advisory interface as the mappings by change underneath all the time. Some modern filesystems like btrfs really can't implement it at all.
I used bmap because I wanted writeout characteristics similar to swapout, and mm/swapfile.c uses bmap. I copied the ideas from that code. If bmap is unsafe for bitmap files, then presumably it is unsafe for swap files too.
This is exactly the argument that I used when Andrew Morton queried the code on the way in.
The swap-over-NFS patches introduce a new interface to the filesystem that supports swapout (i.e. writeout that is guaranteed not to block on mem allocation etc). If they ever get upstream, I will change bitmap files to use that interface when it is present.
Quoted 4 lines
Quoted 2 lines
So would you be happy with something like the following?
It's at least a small improvement..
Thanks.
NeilBrown