Message ID | 20240124142645.9334-1-john.g.garry@oracle.com (mailing list archive) |
---|---|
Headers | show |
Series | block atomic writes for XFS | expand |
On Wed, Jan 24, 2024 at 02:26:39PM +0000, John Garry wrote: > This series expands atomic write support to filesystems, specifically > XFS. Since XFS rtvol supports extent alignment already, support will > initially be added there. When XFS forcealign feature is merged, then we > can similarly support atomic writes for a non-rtvol filesystem. Hi John, Along with rtvol check, we can also have a simple check to see if the FS blocksize itself is big enough to satisfy the atomic requirements. For eg on machines with 64K page, we can have say 16k or 64k block sizes which should be able to provide required allocation behavior for atomic writes. In such cases we don't need rtvol. Regards, ojaswin
On 09/02/2024 07:14, Ojaswin Mujoo wrote: > On Wed, Jan 24, 2024 at 02:26:39PM +0000, John Garry wrote: >> This series expands atomic write support to filesystems, specifically >> XFS. Since XFS rtvol supports extent alignment already, support will >> initially be added there. When XFS forcealign feature is merged, then we >> can similarly support atomic writes for a non-rtvol filesystem. > > Hi John, > > Along with rtvol check, we can also have a simple check to see if the > FS blocksize itself is big enough to satisfy the atomic requirements. > For eg on machines with 64K page, we can have say 16k or 64k block sizes > which should be able to provide required allocation behavior for atomic > writes. In such cases we don't need rtvol. > I suppose we could do, but I would rather just concentrate on rtvol support initially, and there we do report atomic write unit min = FS block size (even if rt extsize is unset). In addition, I plan to initially just support atomic write unit min = FS block size (for both rtvol and !rtvol). Thanks, John
On Fri, Feb 09, 2024 at 09:22:20AM +0000, John Garry wrote: > On 09/02/2024 07:14, Ojaswin Mujoo wrote: > > On Wed, Jan 24, 2024 at 02:26:39PM +0000, John Garry wrote: > > > This series expands atomic write support to filesystems, specifically > > > XFS. Since XFS rtvol supports extent alignment already, support will > > > initially be added there. When XFS forcealign feature is merged, then we > > > can similarly support atomic writes for a non-rtvol filesystem. > > > > Hi John, > > > > Along with rtvol check, we can also have a simple check to see if the > > FS blocksize itself is big enough to satisfy the atomic requirements. > > For eg on machines with 64K page, we can have say 16k or 64k block sizes > > which should be able to provide required allocation behavior for atomic > > writes. In such cases we don't need rtvol. > > > I suppose we could do, but I would rather just concentrate on rtvol support > initially, and there we do report atomic write unit min = FS block size > (even if rt extsize is unset). Okay understood. Thanks, ojaswin > > In addition, I plan to initially just support atomic write unit min = FS > block size (for both rtvol and !rtvol). > > Thanks, > John
From reading the series and the discussions with Darrick and Dave I'm coming more and more back to my initial position that tying this user visible feature to hardware limits is wrong and will just keep on creating ever more painpoints in the future. Based on that I suspect that doing proper software only atomic writes using the swapext log item and selective always COW mode and making that work should be the first step. We can then avoid that overhead for properly aligned writs if the hardware supports it. For your Oracle DB loads you'll set the alignment hints and maybe even check with fiemap that everything is fine and will get the offload, but we also provide a nice and useful API for less performance critical applications that don't have to care about all these details.
John Garry <john.g.garry@oracle.com> writes: > This series expands atomic write support to filesystems, specifically > XFS. Since XFS rtvol supports extent alignment already, support will > initially be added there. When XFS forcealign feature is merged, then we > can similarly support atomic writes for a non-rtvol filesystem. > > Flag FS_XFLAG_ATOMICWRITES is added as an enabling flag for atomic writes. > > For XFS rtvol, support can be enabled through xfs_io command: > $xfs_io -c "chattr +W" filename > $xfs_io -c "lsattr -v" filename > [realtime, atomic-writes] filename Hi John, I first took your block atomic write patch series [1] and then applied this series on top. I also compiled xfsprogs with chattr atomic write support from [2]. [1]: https://lore.kernel.org/linux-nvme/20240124113841.31824-1-john.g.garry@oracle.com/T/#m4ad28b480a8e12eb51467e17208d98ca50041ff2 [2]: https://github.com/johnpgarry/xfsprogs-dev/commits/atomicwrites/ But while setting +W attr, I see an Invalid argument error. Is there anything I need to do first? root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "chattr +W" /mnt1/test/f1 xfs_io: cannot set flags on /mnt1/test/f1: Invalid argument root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "lsattr -v" /mnt1/test/f1 [realtime] /mnt1/test/f1 > > The FS needs to be formatted with a specific extent alignment size, like: > mkf.xfs -r rtdev=/dev/sdb,extsize=16K -d rtinherit=1 /dev/sda > > This enables 16K atomic write support. There are no checks whether the > underlying HW actually supports that for enabling atomic writes with > xfs_io, though, so statx needs to be issued for a file to know atomic > write limits. > Here you say that xfs_io does not check whether underlying HW actually supports atomic writes or not. So I am assuming xfs_io -c "chattr +W" should have just worked? Sorry, I am still in the process of going over the patches, but I thought let me anyways ask this first. -ritesh
On 13/02/2024 07:45, Ritesh Harjani (IBM) wrote: > John Garry <john.g.garry@oracle.com> writes: > >> This series expands atomic write support to filesystems, specifically >> XFS. Since XFS rtvol supports extent alignment already, support will >> initially be added there. When XFS forcealign feature is merged, then we >> can similarly support atomic writes for a non-rtvol filesystem. >> >> Flag FS_XFLAG_ATOMICWRITES is added as an enabling flag for atomic writes. >> >> For XFS rtvol, support can be enabled through xfs_io command: >> $xfs_io -c "chattr +W" filename >> $xfs_io -c "lsattr -v" filename >> [realtime, atomic-writes] filename > > Hi John, > > I first took your block atomic write patch series [1] and then applied this > series on top. I also compiled xfsprogs with chattr atomic write support from [2]. > > [1]: https://lore.kernel.org/linux-nvme/20240124113841.31824-1-john.g.garry@oracle.com/T/#m4ad28b480a8e12eb51467e17208d98ca50041ff2 > [2]: https://github.com/johnpgarry/xfsprogs-dev/commits/atomicwrites/ > > > But while setting +W attr, I see an Invalid argument error. Is there > anything I need to do first? > > root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "chattr +W" /mnt1/test/f1 > xfs_io: cannot set flags on /mnt1/test/f1: Invalid argument > > root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "lsattr -v" /mnt1/test/f1 > [realtime] /mnt1/test/f1 Can you provide your full steps? I'm doing something like: # /mkfs.xfs -r rtdev=/dev/sdb,extsize=16k -d rtinherit=1 /dev/sda meta-data=/dev/sda isize=512 agcount=4, agsize=22400 blks = sectsz=512 attr=2, projid32bit=1 = crc=1 finobt=1, sparse=1, rmapbt=0 = reflink=0 bigtime=1 inobtcount=1 nrext64=0 data = bsize=4096 blocks=89600, imaxpct=25 = sunit=0 swidth=0 blks naming =version 2 bsize=4096 ascii-ci=0, ftype=1 log =internal log bsize=4096 blocks=16384, version=2 = sectsz=512 sunit=0 blks, lazy-count=1 realtime =/dev/sdb extsz=16384 blocks=89600, rtextents=22400 # mount /dev/sda mnt -o rtdev=/dev/sdb [ 5.553482] XFS (sda): EXPERIMENTAL atomic writes feature in use. Use at your own risk! [ 5.556752] XFS (sda): Mounting V5 Filesystem 6e0820e6-4d44-4c3e-89f2-21b4d4480f88 [ 5.602315] XFS (sda): Ending clean mount # # touch mnt/file # /xfs_io -c "lsattr -v" mnt/file [realtime] mnt/file # # # /xfs_io -c "chattr +W" mnt/file # /xfs_io -c "lsattr -v" mnt/file [realtime, atomic-writes] mnt/file And then we can check limits: # /test-statx -a /root/mnt/file dump_statx results=9fff Size: 0 Blocks: 0 IO Block: 16384 regular file Device: 08:00 Inode: 131 Links: 1 Access: (0644/-rw-r--r--) Uid: 0 Gid: 0 Access: 2024-02-13 08:31:51.962900974+0000 Modify: 2024-02-13 08:31:51.962900974+0000 Change: 2024-02-13 08:31:51.969900974+0000 Birth: 2024-02-13 08:31:51.962900974+0000 stx_attributes_mask=0x603070 STATX_ATTR_WRITE_ATOMIC set unit min: 4096 unit max: 16384 segments max: 1 Attributes: 0000000000400000 (........ ........ ........ ........ ........ .?-..... ..--.... .---....) # # Does xfs_io have a statx function? If so, I can add support for atomic writes for statx there. In the meantime, that test-statx code is also on my branch, and can be run on the block device file (to sanity check that the rtvol device supports atomic writes). Thanks, John
John Garry <john.g.garry@oracle.com> writes: > On 13/02/2024 07:45, Ritesh Harjani (IBM) wrote: >> John Garry <john.g.garry@oracle.com> writes: >> >>> This series expands atomic write support to filesystems, specifically >>> XFS. Since XFS rtvol supports extent alignment already, support will >>> initially be added there. When XFS forcealign feature is merged, then we >>> can similarly support atomic writes for a non-rtvol filesystem. >>> >>> Flag FS_XFLAG_ATOMICWRITES is added as an enabling flag for atomic writes. >>> >>> For XFS rtvol, support can be enabled through xfs_io command: >>> $xfs_io -c "chattr +W" filename >>> $xfs_io -c "lsattr -v" filename >>> [realtime, atomic-writes] filename >> >> Hi John, >> >> I first took your block atomic write patch series [1] and then applied this >> series on top. I also compiled xfsprogs with chattr atomic write support from [2]. >> >> [1]: https://lore.kernel.org/linux-nvme/20240124113841.31824-1-john.g.garry@oracle.com/T/#m4ad28b480a8e12eb51467e17208d98ca50041ff2 >> [2]: https://github.com/johnpgarry/xfsprogs-dev/commits/atomicwrites/ >> >> >> But while setting +W attr, I see an Invalid argument error. Is there >> anything I need to do first? >> >> root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "chattr +W" /mnt1/test/f1 >> xfs_io: cannot set flags on /mnt1/test/f1: Invalid argument >> >> root@ubuntu:~# /root/xt/xfsprogs-dev/io/xfs_io -c "lsattr -v" /mnt1/test/f1 >> [realtime] /mnt1/test/f1 > > Can you provide your full steps? > > I'm doing something like: > > # /mkfs.xfs -r rtdev=/dev/sdb,extsize=16k -d rtinherit=1 /dev/sda > meta-data=/dev/sda isize=512 agcount=4, agsize=22400 blks > = sectsz=512 attr=2, projid32bit=1 > = crc=1 finobt=1, sparse=1, rmapbt=0 > = reflink=0 bigtime=1 inobtcount=1 > nrext64=0 > data = bsize=4096 blocks=89600, imaxpct=25 > = sunit=0 swidth=0 blks > naming =version 2 bsize=4096 ascii-ci=0, ftype=1 > log =internal log bsize=4096 blocks=16384, version=2 > = sectsz=512 sunit=0 blks, lazy-count=1 > realtime =/dev/sdb extsz=16384 blocks=89600, rtextents=22400 > # mount /dev/sda mnt -o rtdev=/dev/sdb > [ 5.553482] XFS (sda): EXPERIMENTAL atomic writes feature in use. Use > at your own risk! My bad, I missed to see your xfsprogs change involve setting this feature flag as well during mkfs time itself. I wasn't using the right mkfs utility. > [ 5.556752] XFS (sda): Mounting V5 Filesystem > 6e0820e6-4d44-4c3e-89f2-21b4d4480f88 > [ 5.602315] XFS (sda): Ending clean mount > # > # touch mnt/file > # /xfs_io -c "lsattr -v" mnt/file > [realtime] mnt/file > # > # > # /xfs_io -c "chattr +W" mnt/file > # /xfs_io -c "lsattr -v" mnt/file > [realtime, atomic-writes] mnt/file > Yup, this seems to work fine. Thanks! > And then we can check limits: > > # /test-statx -a /root/mnt/file > dump_statx results=9fff > Size: 0 Blocks: 0 IO Block: 16384 regular file > Device: 08:00 Inode: 131 Links: 1 > Access: (0644/-rw-r--r--) Uid: 0 Gid: 0 > Access: 2024-02-13 08:31:51.962900974+0000 > Modify: 2024-02-13 08:31:51.962900974+0000 > Change: 2024-02-13 08:31:51.969900974+0000 > Birth: 2024-02-13 08:31:51.962900974+0000 > stx_attributes_mask=0x603070 > STATX_ATTR_WRITE_ATOMIC set > unit min: 4096 > unit max: 16384 > segments max: 1 > Attributes: 0000000000400000 (........ ........ ........ ........ > ........ .?-..... ..--.... .---....) > # > # > > Does xfs_io have a statx function? If so, I can add support for atomic > writes for statx there. In the meantime, that test-statx code is also on > my branch, and can be run on the block device file (to sanity check that > the rtvol device supports atomic writes). > > Thanks, > John
On Tue, Feb 13, 2024 at 08:22:37AM +0100, Christoph Hellwig wrote: > From reading the series and the discussions with Darrick and Dave > I'm coming more and more back to my initial position that tying this > user visible feature to hardware limits is wrong and will just keep > on creating ever more painpoints in the future. > > Based on that I suspect that doing proper software only atomic writes > using the swapext log item and selective always COW mode Er, what are you thinking w.r.t. swapext and sometimescow? swapext doesn't currently handle COW forks at all, and it can only exchange between two of the same type of fork (e.g. both data forks or both attr forks, no mixing). Or will that be your next suggestion whenever I get back to fiddling with the online fsck patches? ;) > and making that > work should be the first step. We can then avoid that overhead for > properly aligned writs if the hardware supports it. For your Oracle > DB loads you'll set the alignment hints and maybe even check with > fiemap that everything is fine and will get the offload, but we also > provide a nice and useful API for less performance critical applications > that don't have to care about all these details. I suspect they might want to fail-fast (back to standard WAL mode or whatever) if the hardware support isn't available. --D
On Tue, Feb 13, 2024 at 08:41:10AM +0000, John Garry wrote: > On 13/02/2024 07:45, Ritesh Harjani (IBM) wrote: > > John Garry <john.g.garry@oracle.com> writes: > > Does xfs_io have a statx function? Yes, it's right there in the man page: statx [ -v|-r ][ -m basic | -m all | -m <mask> ][ -FD ] Selected statistics from stat(2) and the XFS_IOC_GETXATTR system call on the current file. -v Show timestamps. -r Dump raw statx structure values. -m basic Set the field mask for the statx call to STATX_BASIC_STATS. -m all Set the the field mask for the statx call to STATX_ALL (default). -m <mask> Specify a numeric field mask for the statx call. -F Force the attributes to be synced with the server. -D Don't sync attributes with the server. -Dave.
On Tue, Feb 13, 2024 at 08:22:37AM +0100, Christoph Hellwig wrote: > From reading the series and the discussions with Darrick and Dave > I'm coming more and more back to my initial position that tying this > user visible feature to hardware limits is wrong and will just keep > on creating ever more painpoints in the future. Yes, that's pretty much what I've been trying to say from the start. The functionality atomic writes need from the filesystem is for extent alignment constraints to be applied to all extent manipulations, not just allocation. This is the same functionality that DAX based XFS filesystems need to guarantee PMD aligned extents. IOWs, the required filesystem extent alignment functionality is not specific to atomic writes and it is not specific to a particular type of storage hardware. If we implement the generic extent alignment constraints properly, everything else from there is just a matter of configuring the filesystem geometry to match the underlying hardware capability. mkfs can do that for us, like it already does for RAID storage... -Dave.
On Wed, Feb 14, 2024 at 10:50:22AM +1100, Dave Chinner wrote: > The functionality atomic writes need from the filesystem is for > extent alignment constraints to be applied to all extent > manipulations, not just allocation. This is the same functionality > that DAX based XFS filesystems need to guarantee PMD aligned > extents. > > IOWs, the required filesystem extent alignment functionality is not > specific to atomic writes and it is not specific to a particular > type of storage hardware. > > If we implement the generic extent alignment constraints properly, > everything else from there is just a matter of configuring the > filesystem geometry to match the underlying hardware capability. > mkfs can do that for us, like it already does for RAID storage... Agreed. But the one thing making atomic writes odd right now is that it absolutely is required for operation right now, while for other features is is somewhere between nice and important to have and not a deal breaker. So eithe we need to figure out a somewhat generic and not totally XFS implementation specific user space interface to do the force alignment (which this series tries to do). Or we make atomic writes like the other features and ensure they still work without the proper alignment if they suck. Doing that was my initial gut feeling, and looking at other approaches just makes me tend even stronger towards that. > > -Dave. > -- > Dave Chinner > david@fromorbit.com ---end quoted text---
On Tue, Feb 13, 2024 at 09:55:49AM -0800, Darrick J. Wong wrote: > On Tue, Feb 13, 2024 at 08:22:37AM +0100, Christoph Hellwig wrote: > > From reading the series and the discussions with Darrick and Dave > > I'm coming more and more back to my initial position that tying this > > user visible feature to hardware limits is wrong and will just keep > > on creating ever more painpoints in the future. > > > > Based on that I suspect that doing proper software only atomic writes > > using the swapext log item and selective always COW mode > > Er, what are you thinking w.r.t. swapext and sometimescow? What do you mean with sometimescow? Just normal reflinked inodes? > swapext > doesn't currently handle COW forks at all, and it can only exchange > between two of the same type of fork (e.g. both data forks or both attr > forks, no mixing). > > Or will that be your next suggestion whenever I get back to fiddling > with the online fsck patches? ;) Let's take a step back. If we want atomic write semantics without hardware offload, what we need is to allocate new blocks and atomically swap them into the data fork. Basicall an atomic version of xfs_reflink_end_cow. But yes, the details of the current swapext item might not be an exact fit, maybe it's just shared infrastructure and concepts. I'm not planning to make you do it, because such a log item would generally be pretty useful for always COW mode. > > and making that > > work should be the first step. We can then avoid that overhead for > > properly aligned writs if the hardware supports it. For your Oracle > > DB loads you'll set the alignment hints and maybe even check with > > fiemap that everything is fine and will get the offload, but we also > > provide a nice and useful API for less performance critical applications > > that don't have to care about all these details. > > I suspect they might want to fail-fast (back to standard WAL mode or > whatever) if the hardware support isn't available. Maybe for your particular DB use case. But there's plenty of applications that just want atomic writes without building their own infrastruture, including some that want pretty large chunks. Also if a file system supports logging data (which I have an XFS early prototype for that I plan to finish), we can even do the small double writes more efficiently than the application, all through the same interface.
On 13/02/2024 22:49, Dave Chinner wrote: >> Does xfs_io have a statx function? > Yes, it's right there in the man page: > > statx [ -v|-r ][ -m basic | -m all | -m <mask> ][ -FD ] > Selected statistics from stat(2) and the XFS_IOC_GETXATTR system call on the current file. > -v Show timestamps. > -r Dump raw statx structure values. > -m basic > Set the field mask for the statx call to STATX_BASIC_STATS. > -m all > Set the the field mask for the statx call to STATX_ALL (default). > -m <mask> > Specify a numeric field mask for the statx call. > -F Force the attributes to be synced with the server. > -D Don't sync attributes with the server. ok, I can check that out and look to add any support required for atomic writes extension. Thanks, John
On Wed, Feb 14, 2024 at 08:45:59AM +0100, Christoph Hellwig wrote: > On Tue, Feb 13, 2024 at 09:55:49AM -0800, Darrick J. Wong wrote: > > On Tue, Feb 13, 2024 at 08:22:37AM +0100, Christoph Hellwig wrote: > > > From reading the series and the discussions with Darrick and Dave > > > I'm coming more and more back to my initial position that tying this > > > user visible feature to hardware limits is wrong and will just keep > > > on creating ever more painpoints in the future. > > > > > > Based on that I suspect that doing proper software only atomic writes > > > using the swapext log item and selective always COW mode > > > > Er, what are you thinking w.r.t. swapext and sometimescow? > > What do you mean with sometimescow? Just normal reflinked inodes? > > > swapext > > doesn't currently handle COW forks at all, and it can only exchange > > between two of the same type of fork (e.g. both data forks or both attr > > forks, no mixing). > > > > Or will that be your next suggestion whenever I get back to fiddling > > with the online fsck patches? ;) > > Let's take a step back. If we want atomic write semantics without > hardware offload, what we need is to allocate new blocks and atomically > swap them into the data fork. Basicall an atomic version of > xfs_reflink_end_cow. But yes, the details of the current swapext > item might not be an exact fit, maybe it's just shared infrastructure > and concepts. Hmm. For rt reflink (whenever I get back to that, ha) I've been starting to think that yes, we actually /do/ want to have a log item that tracks the progress of remap and cow operations. That would solve the problem of someone wanting to reflink a semi-written rtx. That said, it might complicate the reflink code quite a bit since right now it writes zeroes to the unwritten parts of an rt file's rtx so that there's only one mapping record for the whole rtx, and then it remaps them. That's most of why I haven't bothered to implement that solution. > I'm not planning to make you do it, because such a log item would > generally be pretty useful for always COW mode. One other thing -- while I was refactoring the swapext code into exch{range,maps}, it occurred to me that doing an exchange between the cow and data forks isn't possible because log recovery won't be able to do anything. There's no ondisk metadata to map a cow staging extent back to the file it came from, which means we can't generally resume an exchange operation. However for a small write I guess you could simply queue all the log intent items for all the changes needed and commit that. > > > and making that > > > work should be the first step. We can then avoid that overhead for > > > properly aligned writs if the hardware supports it. For your Oracle > > > DB loads you'll set the alignment hints and maybe even check with > > > fiemap that everything is fine and will get the offload, but we also > > > provide a nice and useful API for less performance critical applications > > > that don't have to care about all these details. > > > > I suspect they might want to fail-fast (back to standard WAL mode or > > whatever) if the hardware support isn't available. > > Maybe for your particular DB use case. But there's plenty of > applications that just want atomic writes without building their > own infrastruture, including some that want pretty large chunks. > > Also if a file system supports logging data (which I have an > XFS early prototype for that I plan to finish), we can even do > the small double writes more efficiently than the application, > all through the same interface. Heh. Ted's been trying to kill data=journal. Now we've found a use for it after all. :) --D
On Wed, Feb 21, 2024 at 08:56:15AM -0800, Darrick J. Wong wrote: > Hmm. For rt reflink (whenever I get back to that, ha) I've been > starting to think that yes, we actually /do/ want to have a log item > that tracks the progress of remap and cow operations. That would solve > the problem of someone wanting to reflink a semi-written rtx. > > That said, it might complicate the reflink code quite a bit since right > now it writes zeroes to the unwritten parts of an rt file's rtx so that > there's only one mapping record for the whole rtx, and then it remaps > them. That's most of why I haven't bothered to implement that solution. I'm still not sure that supporting reflinks for rtextsize > 1 is a good idea.. > > I'm not planning to make you do it, because such a log item would > > generally be pretty useful for always COW mode. > > One other thing -- while I was refactoring the swapext code into > exch{range,maps}, it occurred to me that doing an exchange between the > cow and data forks isn't possible because log recovery won't be able to > do anything. There's no ondisk metadata to map a cow staging extent > back to the file it came from, which means we can't generally resume an > exchange operation. Yeah. > > Also if a file system supports logging data (which I have an > > XFS early prototype for that I plan to finish), we can even do > > the small double writes more efficiently than the application, > > all through the same interface. > > Heh. Ted's been trying to kill data=journal. Now we've found a use for > it after all. :) Well.. unconditional logging of data just seems like a really bad idea. Using it as an optimization for very small and/or synchronous writes is a pretty common technique.