improve buffered write performance with fgf order hint
cover improve buffered write performance with fgf order hint
From: Pankaj Raghav <[email protected]>
Filesystems that support large folios can set the fgf order before buffered write(see XFS iomap_write_begin()) that can provide as a hint to page cache to allocate large folios, if possible.
The first patch is a minor cleanup.
The second patch sets fgf order before starting the buffered write.
I tested the performance on Samsung SSD 990 pro on a system with 64GB RAM as follows:
$ bcachefs format -f /dev/nvme0n1;
$ mount -t bcachefs /dev/nvme0n1 /mnt
$ fio --name=bcachefs --filename=/mnt/test --size=100G \
--ioengine=io_uring --iodepth=16 --rw=write --bs=128kI measured the BW(MB/s) across three runs on 6.10-rc3:
Before patches: 2730
After patches: 3430 (1.25x boost)
With -o no_data_io mount option:
Before patches: 2920
After patches: 4630 (1.5x boost)
I was not able to test the patches with ktest due to some issue with debian(some broken package issue). Maybe Kent can run it in his CI while I try to fix ktest locally?
Pankaj Raghav (2):
bcachefs: use FGP_WRITEBEGIN instead of combining individual flags
bcachefs: set fgf order hint before starting a buffered write
fs/bcachefs/fs-io-buffered.c | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-)
base-commit: 03d44168cbd7fc57d5de56a3730427db758fc7f6
Signature · 1 line
2.44.1
PATCH 1/2 bcachefs: use FGP_WRITEBEGIN instead of combining individual flags
From: Pankaj Raghav <[email protected]>
Use FGP_WRITEBEGIN to avoid repeating the individual FGP flags before starting a buffered write.
--- fs/bcachefs/fs-io-buffered.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
fs/bcachefs/fs-io-buffered.c +4 -6
@@ -677,9 +677,8 @@ int bch2_write_begin(struct file *file, struct address_space *mapping,bch2_pagecache_add_get (inode); - folio =__filemap_get_folio (mapping, pos >>PAGE_SHIFT , -FGP_LOCK |FGP_WRITE |FGP_CREAT |FGP_STABLE , -mapping_gfp_mask (mapping)); + folio =__filemap_get_folio (mapping, pos >>PAGE_SHIFT ,FGP_WRITEBEGIN , +mapping_gfp_mask (mapping));if (IS_ERR_OR_NULL (folio)) goto err_unlock;
@@ -820,9 +819,8 @@ static int __bch2_buffered_write(struct bch_inode_info *inode,darray_init (&fs); ret =bch2_filemap_get_contig_folios_d (mapping, pos, end, -FGP_LOCK |FGP_WRITE |FGP_STABLE |FGP_CREAT , -mapping_gfp_mask (mapping), - &fs); +FGP_WRITEBEGIN , +mapping_gfp_mask (mapping), &fs);if (ret) goto out;
Signature · 1 line
2.44.1
Re: [PATCH 1/2] bcachefs: use FGP_WRITEBEGIN instead of combining individual flags
fs/bcachefs/fs-io-buffered.c static int __bch2_buffered_write(struct bch_inode_info *inode, · line 823
Quoted 6 lines · Pankaj Raghav (Samsung)
fs/bcachefs/fs-io-buffered.c +2
-3ret =bch2_filemap_get_contig_folios_d (mapping, pos, end, -FGP_LOCK |FGP_WRITE |FGP_STABLE |FGP_CREAT , -mapping_gfp_mask (mapping), - &fs); +FGP_WRITEBEGIN , +mapping_gfp_mask (mapping), &fs);
Don't change the indentation here. In the next patch it makes the lines too long.
In general, don't change the indentation. It's been one of the biggest pains of the folio conversion. "Oh, you changed the name of this function and now the arguments don't line up". Don't line up the arguments with the opening paren.
Re: [PATCH 1/2] bcachefs: use FGP_WRITEBEGIN instead of combining individual flags
Quoted 10 lines · Matthew Wilcox
fs/bcachefs/fs-io-buffered.c static int __bch2_buffered_write(struct bch_inode_info *inode, · line 823
Quoted 6 lines · Pankaj Raghav (Samsung)
fs/bcachefs/fs-io-buffered.c +2
-3ret =bch2_filemap_get_contig_folios_d (mapping, pos, end, -FGP_LOCK |FGP_WRITE |FGP_STABLE |FGP_CREAT , -mapping_gfp_mask (mapping), - &fs); +FGP_WRITEBEGIN , +mapping_gfp_mask (mapping), &fs);Don't change the indentation here. In the next patch it makes the lines too long.
I used clangd to do the formatting. I will probably stick without changing the format in this patch
Thanks.
Quoted 5 lines
In general, don't change the indentation. It's been one of the biggest pains of the folio conversion. "Oh, you changed the name of this function and now the arguments don't line up". Don't line up the arguments with the opening paren.
PATCH 2/2 bcachefs: set fgf order hint before starting a buffered write
In reply to Pankaj Raghav (Samsung)From: Pankaj Raghav <[email protected]>
Set the preferred folio order in the fgp_flags by calling fgf_set_order(). Page cache will try to allocate large folio of the preferred order whenever possible instead of allocating multiple 0 order folios.
This improves the buffered write performance up to 1.25x with default mount options and up to 1.57x when mounted with no_data_io option with the following fio workload:
fio --name=bcachefs --filename=/mnt/test --size=100G \
--ioengine=io_uring --iodepth=16 --rw=write --bs=128k--- fs/bcachefs/fs-io-buffered.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
fs/bcachefs/fs-io-buffered.c +3 -2
@@ -677,7 +677,8 @@ int bch2_write_begin(struct file *file, struct address_space *mapping,bch2_pagecache_add_get (inode); - folio =__filemap_get_folio (mapping, pos >>PAGE_SHIFT ,FGP_WRITEBEGIN , + folio =__filemap_get_folio (mapping, pos >>PAGE_SHIFT , +FGP_WRITEBEGIN |fgf_set_order (len),mapping_gfp_mask (mapping));if (IS_ERR_OR_NULL (folio)) goto err_unlock;
@@ -819,7 +820,7 @@ static int __bch2_buffered_write(struct bch_inode_info *inode,darray_init (&fs); ret =bch2_filemap_get_contig_folios_d (mapping, pos, end, -FGP_WRITEBEGIN , +FGP_WRITEBEGIN |fgf_set_order (len),mapping_gfp_mask (mapping), &fs);if (ret) goto out;
Signature · 1 line
2.44.1
Quoted 21 lines · Pankaj Raghav (Samsung)
From: Pankaj Raghav <[email protected]>
Filesystems that support large folios can set the fgf order before buffered
write(see XFS iomap_write_begin())that can provide as a hint to page cache to allocate large folios, if possible.The first patch is a minor cleanup.
The second patch sets fgf order before starting the buffered write.I tested the performance on Samsung SSD 990 pro on a system with 64GB RAM as follows:
$ bcachefs format -f /dev/nvme0n1; $ mount -t bcachefs /dev/nvme0n1 /mnt $ fio --name=bcachefs --filename=/mnt/test --size=100G \ --ioengine=io_uring --iodepth=16 --rw=write --bs=128kI measured the BW(MB/s) across three runs on 6.10-rc3:
Before patches: 2730
After patches: 3430 (1.25x boost)
Quoted 7 lines
With -o
no_data_iomount option:
Before patches: 2920
After patches: 4630 (1.5x boost)I was not able to test the patches with ktest due to some issue with debian(some broken package issue). Maybe Kent can run it in his CI while I try to fix ktest locally?
It's in my testing branch, results will be showing up here:
https://evilpiepirate.org/~testdashboard/ci?branch=bcachefs-testing