| Message ID | 20230322221714.2702819-2-neal@gompa.dev (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Enforce 4k sectorize by default for mkfs | expand |
Comparing btrfs sectorsize=64K with sectorsize=4K on an aarch64 virtual host with PAGESIZE=64k shows that switching to sectorsize=4K by default for buffered IO has a low impact, while the direct IO performance is improved by roughly 21% to 152% for various fio block sizes as shown below. aarch64 PAGESIZE=64K ==================== Buffered IO: ============ FIO bs sectorsize=64k sectorsize=4k diff K MB/s MB/s PC 4 752 755 8 783 832 +6 64 1066 1173 +10 128 1120 1098 +2 256 1112 1079 -3 Dierct IO: ============ FIO bs sectorsize=64k sectorsize=4k diff K MB/s MB/s PC 4 54 106 +96 8 107 270 +152 64 737 894 +21 128 862 1130 +31 256 846 1103 +30 FIO config file: [global] directory=/mnt/scratch allrandrepeat=1 readwrite=readwrite ioengine=io_uring iodepth=256 end_fsync=1 fallocate=none group_reporting gtod_reduce=1 numjobs=8 size=10G stonewall [fio-direct-4k] direct=1 <-- changed as appropriate bs=4k <---changed as appropriate On 23/03/2023 06:17, Neal Gompa wrote: > We have had working subpage support in Btrfs for many cycles now. > Generally, we do not want people creating filesystems by default > with non-4k sectorsizes since it creates portability problems. > > Signed-off-by: Neal Gompa <neal@gompa.dev> > > Reviewed-by: Anand Jain <anand.jain@oracle.com> > Reviewed-by: Qu Wenruo <wqu@suse.com> > Reviewed-by: Josef Bacik <josef@toxicpanda.com> > > --- > Documentation/Subpage.rst | 15 ++++++++------- > Documentation/mkfs.btrfs.rst | 13 +++++++++---- > mkfs/main.c | 2 +- > 3 files changed, 18 insertions(+), 12 deletions(-) > > diff --git a/Documentation/Subpage.rst b/Documentation/Subpage.rst > index 21a495d5..39ef7d6d 100644 > --- a/Documentation/Subpage.rst > +++ b/Documentation/Subpage.rst > @@ -9,17 +9,18 @@ to the exactly same size of the block and page. On x86_64 this is typically > pages, like 64KiB on 64bit ARM or PowerPC. This means filesystems created > with 64KiB sector size cannot be mounted on a system with 4KiB page size. > > -While with subpage support, systems with 64KiB page size can create (still needs > -"-s 4k" option for mkfs.btrfs) and mount filesystems with 4KiB sectorsize, > -allowing us to push 4KiB sectorsize as default sectorsize for all platforms in the > -near future. > +Since v6.3, filesystems are created with a 4KiB sectorsize by default, > +though it remains possible to create filesystems with other page sizes > +(such as 64KiB with the "-s 64k" option for mkfs.btrfs). This ensures that > +new filesystems are compatible across other architecture variants using > +larger page sizes. > > Requirements, limitations > ------------------------- > > -The initial subpage support has been added in v5.15, although it's still > -considered as experimental at the time of writing (v5.18), most features are > -already working without problems. > +The initial subpage support has been added in v5.15. Most features are > +already working without problems. Subpage support is used by default > +for systems with a non-4KiB page size since v6.3. > > End users can mount filesystems with 4KiB sectorsize and do their usual > workload, while should not notice any obvious change, as long as the initial > diff --git a/Documentation/mkfs.btrfs.rst b/Documentation/mkfs.btrfs.rst > index ba7227b3..16abf0ca 100644 > --- a/Documentation/mkfs.btrfs.rst > +++ b/Documentation/mkfs.btrfs.rst > @@ -116,10 +116,15 @@ OPTIONS > -s|--sectorsize <size> > Specify the sectorsize, the minimum data block allocation unit. > > - The default value is the page size and is autodetected. If the sectorsize > - differs from the page size, the created filesystem may not be mountable by the > - running kernel. Therefore it is not recommended to use this option unless you > - are going to mount it on a system with the appropriate page size. > + By default, the value is 4KiB, but it can be manually set to match the > + system page size. However, if the sector size is different from the page > + size, the resulting filesystem may not be mountable by the current > + kernel, apart from the default 4KiB. Hence, using this option is not > + advised unless you intend to mount it on a system with the suitable > + page size. > + > + .. note:: > + Versions prior to 6.3 set the sectorsize matching to the page size. > > -L|--label <string> > Specify a label for the filesystem. The *string* should be less than 256 > diff --git a/mkfs/main.c b/mkfs/main.c > index f5e34cbd..5e1834d7 100644 > --- a/mkfs/main.c > +++ b/mkfs/main.c > @@ -1207,7 +1207,7 @@ int BOX_MAIN(mkfs)(int argc, char **argv) > } > > if (!sectorsize) > - sectorsize = (u32)sysconf(_SC_PAGESIZE); > + sectorsize = (u32)SZ_4K; > if (btrfs_check_sectorsize(sectorsize)) > goto error; >
[ fixed table format ] Comparing btrfs sectorsize=64K with sectorsize=4K on an aarch64 virtual host with PAGESIZE=64k shows that switching to sectorsize (ss) =4K by default for buffered IO has a low impact, while the direct IO performance is improved by roughly 21% to 152% for various fio block sizes as shown below. aarch64 PAGESIZE=64K ==================== Buffered IO: =========== FIO bs ss=64k ss=4k diff K MB/s MB/s % 4 752 755 8 783 832 +6 64 1066 1173 +10 128 1120 1098 +2 256 1112 1079 -3 Dierct IO: ========= FIO bs ss=64k ss=4k diff K MB/s MB/s % 4 54 106 +96 8 107 270 +152 64 737 894 +21 128 862 1130 +31 256 846 1103 +30 FIO config file: [global] directory=/mnt/scratch allrandrepeat=1 readwrite=readwrite ioengine=io_uring iodepth=256 end_fsync=1 fallocate=none group_reporting gtod_reduce=1 numjobs=8 size=10G stonewall [fio-direct-4k] direct=1 <-- changed as appropriate bs=4k <---changed as appropriate On 23/03/2023 06:17, Neal Gompa wrote: > We have had working subpage support in Btrfs for many cycles now. > Generally, we do not want people creating filesystems by default > with non-4k sectorsizes since it creates portability problems. > > Signed-off-by: Neal Gompa <neal@gompa.dev> > > Reviewed-by: Anand Jain <anand.jain@oracle.com> > Reviewed-by: Qu Wenruo <wqu@suse.com> > Reviewed-by: Josef Bacik <josef@toxicpanda.com> > > --- > Documentation/Subpage.rst | 15 ++++++++------- > Documentation/mkfs.btrfs.rst | 13 +++++++++---- > mkfs/main.c | 2 +- > 3 files changed, 18 insertions(+), 12 deletions(-) > > diff --git a/Documentation/Subpage.rst b/Documentation/Subpage.rst > index 21a495d5..39ef7d6d 100644 > --- a/Documentation/Subpage.rst > +++ b/Documentation/Subpage.rst > @@ -9,17 +9,18 @@ to the exactly same size of the block and page. On x86_64 this is typically > pages, like 64KiB on 64bit ARM or PowerPC. This means filesystems created > with 64KiB sector size cannot be mounted on a system with 4KiB page size. > > -While with subpage support, systems with 64KiB page size can create (still needs > -"-s 4k" option for mkfs.btrfs) and mount filesystems with 4KiB sectorsize, > -allowing us to push 4KiB sectorsize as default sectorsize for all platforms in the > -near future. > +Since v6.3, filesystems are created with a 4KiB sectorsize by default, > +though it remains possible to create filesystems with other page sizes > +(such as 64KiB with the "-s 64k" option for mkfs.btrfs). This ensures that > +new filesystems are compatible across other architecture variants using > +larger page sizes. > > Requirements, limitations > ------------------------- > > -The initial subpage support has been added in v5.15, although it's still > -considered as experimental at the time of writing (v5.18), most features are > -already working without problems. > +The initial subpage support has been added in v5.15. Most features are > +already working without problems. Subpage support is used by default > +for systems with a non-4KiB page size since v6.3. > > End users can mount filesystems with 4KiB sectorsize and do their usual > workload, while should not notice any obvious change, as long as the initial > diff --git a/Documentation/mkfs.btrfs.rst b/Documentation/mkfs.btrfs.rst > index ba7227b3..16abf0ca 100644 > --- a/Documentation/mkfs.btrfs.rst > +++ b/Documentation/mkfs.btrfs.rst > @@ -116,10 +116,15 @@ OPTIONS > -s|--sectorsize <size> > Specify the sectorsize, the minimum data block allocation unit. > > - The default value is the page size and is autodetected. If the sectorsize > - differs from the page size, the created filesystem may not be mountable by the > - running kernel. Therefore it is not recommended to use this option unless you > - are going to mount it on a system with the appropriate page size. > + By default, the value is 4KiB, but it can be manually set to match the > + system page size. However, if the sector size is different from the page > + size, the resulting filesystem may not be mountable by the current > + kernel, apart from the default 4KiB. Hence, using this option is not > + advised unless you intend to mount it on a system with the suitable > + page size. > + > + .. note:: > + Versions prior to 6.3 set the sectorsize matching to the page size. > > -L|--label <string> > Specify a label for the filesystem. The *string* should be less than 256 > diff --git a/mkfs/main.c b/mkfs/main.c > index f5e34cbd..5e1834d7 100644 > --- a/mkfs/main.c > +++ b/mkfs/main.c > @@ -1207,7 +1207,7 @@ int BOX_MAIN(mkfs)(int argc, char **argv) > } > > if (!sectorsize) > - sectorsize = (u32)sysconf(_SC_PAGESIZE); > + sectorsize = (u32)SZ_4K; > if (btrfs_check_sectorsize(sectorsize)) > goto error; >
On Sat, Apr 1, 2023 at 1:31 AM Anand Jain <anand.jain@oracle.com> wrote: > > > [ fixed table format ] > > Comparing btrfs sectorsize=64K with sectorsize=4K on an aarch64 > virtual host with PAGESIZE=64k shows that switching to sectorsize > (ss) =4K by default for buffered IO has a low impact, while the > direct IO performance is improved by roughly 21% to 152% for > various fio block sizes as shown below. > > > aarch64 PAGESIZE=64K > ==================== > > Buffered IO: > =========== > > FIO bs ss=64k ss=4k diff > K MB/s MB/s % > 4 752 755 > 8 783 832 +6 > 64 1066 1173 +10 > 128 1120 1098 +2 > 256 1112 1079 -3 > > > Dierct IO: > ========= > > FIO bs ss=64k ss=4k diff > K MB/s MB/s % > 4 54 106 +96 > 8 107 270 +152 > 64 737 894 +21 > 128 862 1130 +31 > 256 846 1103 +30 > > > FIO config file: > > [global] > directory=/mnt/scratch > allrandrepeat=1 > readwrite=readwrite > ioengine=io_uring > iodepth=256 > end_fsync=1 > fallocate=none > group_reporting > gtod_reduce=1 > numjobs=8 > size=10G > stonewall > > [fio-direct-4k] > direct=1 <-- changed as appropriate > bs=4k <---changed as appropriate > Well, it's nice to see that there are performance advantages to using 4k block sizes everywhere. :) Can we get this committed to btrfs-progs then? -- 真実はいつも一つ！/ Always, there's only one truth!
diff --git a/Documentation/Subpage.rst b/Documentation/Subpage.rst index 21a495d5..39ef7d6d 100644 --- a/Documentation/Subpage.rst +++ b/Documentation/Subpage.rst @@ -9,17 +9,18 @@ to the exactly same size of the block and page. On x86_64 this is typically pages, like 64KiB on 64bit ARM or PowerPC. This means filesystems created with 64KiB sector size cannot be mounted on a system with 4KiB page size. -While with subpage support, systems with 64KiB page size can create (still needs -"-s 4k" option for mkfs.btrfs) and mount filesystems with 4KiB sectorsize, -allowing us to push 4KiB sectorsize as default sectorsize for all platforms in the -near future. +Since v6.3, filesystems are created with a 4KiB sectorsize by default, +though it remains possible to create filesystems with other page sizes +(such as 64KiB with the "-s 64k" option for mkfs.btrfs). This ensures that +new filesystems are compatible across other architecture variants using +larger page sizes. Requirements, limitations ------------------------- -The initial subpage support has been added in v5.15, although it's still -considered as experimental at the time of writing (v5.18), most features are -already working without problems. +The initial subpage support has been added in v5.15. Most features are +already working without problems. Subpage support is used by default +for systems with a non-4KiB page size since v6.3. End users can mount filesystems with 4KiB sectorsize and do their usual workload, while should not notice any obvious change, as long as the initial diff --git a/Documentation/mkfs.btrfs.rst b/Documentation/mkfs.btrfs.rst index ba7227b3..16abf0ca 100644 --- a/Documentation/mkfs.btrfs.rst +++ b/Documentation/mkfs.btrfs.rst @@ -116,10 +116,15 @@ OPTIONS -s|--sectorsize <size> Specify the sectorsize, the minimum data block allocation unit. - The default value is the page size and is autodetected. If the sectorsize - differs from the page size, the created filesystem may not be mountable by the - running kernel. Therefore it is not recommended to use this option unless you - are going to mount it on a system with the appropriate page size. + By default, the value is 4KiB, but it can be manually set to match the + system page size. However, if the sector size is different from the page + size, the resulting filesystem may not be mountable by the current + kernel, apart from the default 4KiB. Hence, using this option is not + advised unless you intend to mount it on a system with the suitable + page size. + + .. note:: + Versions prior to 6.3 set the sectorsize matching to the page size. -L|--label <string> Specify a label for the filesystem. The *string* should be less than 256 diff --git a/mkfs/main.c b/mkfs/main.c index f5e34cbd..5e1834d7 100644 --- a/mkfs/main.c +++ b/mkfs/main.c @@ -1207,7 +1207,7 @@ int BOX_MAIN(mkfs)(int argc, char **argv) } if (!sectorsize) - sectorsize = (u32)sysconf(_SC_PAGESIZE); + sectorsize = (u32)SZ_4K; if (btrfs_check_sectorsize(sectorsize)) goto error;