| Message ID | 20200329174714.32416-1-chaitanya.kulkarni@wdc.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | block: Add support for REQ_OP_ASSIGN_RANGE | expand |
Chaitanya, > This patchset introduces REQ_OP_ASSIGN_RANGE, which is going > to be used for forwarding user's fallocate(0) requests into > block device internals. s/assign_range/allocate/g
On 3/31/20 7:31 PM, Martin K. Petersen wrote: > Chaitanya, > >> This patchset introduces REQ_OP_ASSIGN_RANGE, which is going >> to be used for forwarding user's fallocate(0) requests into >> block device internals. > s/assign_range/allocate/g > Okay, will send out the V2.
On 29/03/2020 20.47, Chaitanya Kulkarni wrote: > Hi, > > This patch-series is based on the original RFC patch series:- > https://www.spinics.net/lists/linux-block/msg47933.html. > > I've designed a rough testcase based on the information present > in the mailing list archive for original RFC, it may need > some corrections from the author. > > If anyone is interested, test results are at the end of this patch. > > Following is the original cover-letter :- > > Information about continuous extent placement may be useful > for some block devices. Say, distributed network filesystems, > which provide block device interface, may use this information > for better blocks placement over the nodes in their cluster, > and for better performance. Block devices, which map a file > on another filesystem (loop), may request the same length extent > on underlining filesystem for less fragmentation and for batching > allocation requests. Also, hypervisors like QEMU may use this > information for optimization of cluster allocations. > > This patchset introduces REQ_OP_ASSIGN_RANGE, which is going > to be used for forwarding user's fallocate(0) requests into > block device internals. It rather similar to existing > REQ_OP_DISCARD, REQ_OP_WRITE_ZEROES, etc. The corresponding > exported primitive is called blkdev_issue_assign_range(). What exact semantics of that? It may/must preserve present data or may/must discard them, or may fill range with random garbage? Obviously I prefer weakest one - may discard data, may return garbage, may do nothing. I.e. lower layer could reuse blocks without zeroing, for encrypted storage this is even safe. So, this works as third type of dicasrd in addtion to REQ_OP_DISCARD and REQ_OP_SECURE_ERASE. > See [1/3] for the details. > > Patch [2/3] teaches loop driver to handle REQ_OP_ASSIGN_RANGE > requests by calling fallocate(0). > > Patch [3/3] makes ext4 to notify a block device about fallocate(0). > > Here is a simple test I did: > https://gist.github.com/tkhai/5b788651cdb74c1dbff3500745878856 > > I attached a file on ext4 to loop. Then, created ext4 partition > on loop device and started the test in the partition. Direct-io > is enabled on loop. > > The test fallocates 4G file and writes from some offset with > given step, then it chooses another offset and repeats. After > the test all the blocks in the file become written. > > The results shows that batching extents-assigning requests improves > the performance: > > Before patchset: real ~ 1min 27sec > After patchset: real ~ 1min 16sec (18% better) > > Ordinary fallocate() before writes improves the performance > by batching the requests. These results just show, the same > is in case of forwarding extents information to underlining > filesystem. > > Regards, > Chaitanya > > Changes from RFC:- > > 1. Add missing plumbing for REQ_OP_ASSIGN_RANGE similar to write-zeores. > 2. Add a prep patch to create a helper to submit payloadless bios. > 3. Design a testcases around the description present in the > cover-letter. > > Chaitanya Kulkarni (1): > block: create payloadless issue bio helper > > Kirill Tkhai (3): > block: Add support for REQ_OP_ASSIGN_RANGE > loop: Forward REQ_OP_ASSIGN_RANGE into fallocate(0) > ext4: Notify block device about alloc-assigned blk > > block/blk-core.c | 5 ++ > block/blk-lib.c | 115 +++++++++++++++++++++++++++++++------- > block/blk-merge.c | 21 +++++++ > block/blk-settings.c | 19 +++++++ > block/blk-zoned.c | 1 + > block/bounce.c | 1 + > drivers/block/loop.c | 5 ++ > fs/ext4/ext4.h | 2 + > fs/ext4/extents.c | 12 +++- > include/linux/bio.h | 9 ++- > include/linux/blk_types.h | 2 + > include/linux/blkdev.h | 34 +++++++++++ > 12 files changed, 201 insertions(+), 25 deletions(-) > > 1. Setup :- > ----------- > # git log --oneline -5 > c64a4c781915 (HEAD -> req-op-assign-range) ext4: Notify block device about alloc-assigned blk > 000cbc6720a4 loop: Forward REQ_OP_ASSIGN_RANGE into fallocate(0) > 89ceed8cac80 block: Add support for REQ_OP_ASSIGN_RANGE > a798743e87e7 block: create payloadless issue bio helper > b53df2e7442c (tag: block-5.6-2020-03-13) block: Fix partition support for host aware zoned block devices > > # cat /proc/kallsyms | grep -i blkdev_issue_assign_range > ffffffffa3264a80 T blkdev_issue_assign_range > ffffffffa4027184 r __ksymtab_blkdev_issue_assign_range > ffffffffa40524be r __kstrtabns_blkdev_issue_assign_range > ffffffffa405a8eb r __kstrtab_blkdev_issue_assign_range > > 2. Test program, will be moved to blktest once code is upstream :- > ----------------- > #define _GNU_SOURCE > #include <sys/types.h> > #include <unistd.h> > #include <stdlib.h> > #include <stdio.h> > #include <fcntl.h> > #include <errno.h> > > #define BLOCK_SIZE 4096 > #define STEP (BLOCK_SIZE * 16) > #define SIZE (1024 * 1024 * 1024ULL) > > int main(int argc, char *argv[]) > { > int fd, step, ret = 0; > unsigned long i; > void *buf; > > if (posix_memalign(&buf, BLOCK_SIZE, SIZE)) { > perror("alloc"); > exit(1); > } > > fd = open("/mnt/loop0/file.img", O_RDWR | O_CREAT | O_DIRECT); > if (fd < 0) { > perror("open"); > exit(1); > } > > if (ftruncate(fd, SIZE)) { > perror("ftruncate"); > exit(1); > } > > ret = fallocate(fd, 0, 0, SIZE); > if (ret) { > perror("fallocate"); > exit(1); > } > > for (step = STEP - BLOCK_SIZE; step >= 0; step -= BLOCK_SIZE) { > printf("step=%u\n", step); > for (i = step; i < SIZE; i += STEP) { > errno = 0; > if (pwrite(fd, buf, BLOCK_SIZE, i) != BLOCK_SIZE) { > perror("pwrite"); > exit(1); > } > } > > if (fsync(fd)) { > perror("fsync"); > exit(1); > } > } > return 0; > } > > 3. Test script, will be moved to blktests once code is upstream :- > ------------------------------------------------------------------ > # cat req_op_assign_test.sh > #!/bin/bash -x > > NULLB_FILE="/mnt/backend/data" > NULLB_MNT="/mnt/backend" > LOOP_MNT="/mnt/loop0" > > delete_loop() > { > umount ${LOOP_MNT} > losetup -D > sleep 3 > } > > delete_nullb() > { > umount ${NULLB_MNT} > echo 1 > config/nullb/nullb0/power > rmdir config/nullb/nullb0 > sleep 3 > } > > unload_modules() > { > rmmod drivers/block/loop.ko > rmmod fs/ext4/ext4.ko > rmmod drivers/block/null_blk.ko > lsmod | grep -e ext4 -e loop -e null_blk > } > > unload() > { > delete_loop > delete_nullb > unload_modules > } > > load_ext4() > { > make -j $(nproc) M=fs/ext4 modules > local src=fs/ext4/ > local dest=/lib/modules/`uname -r`/kernel/fs/ext4 > \cp ${src}/ext4.ko ${dest}/ > > modprobe mbcache > modprobe jbd2 > sleep 1 > insmod fs/ext4/ext4.ko > sleep 1 > } > > load_nullb() > { > local src=drivers/block/ > local dest=/lib/modules/`uname -r`/kernel/drivers/block > \cp ${src}/null_blk.ko ${dest}/ > > modprobe null_blk nr_devices=0 > sleep 1 > > mkdir config/nullb/nullb0 > tree config/nullb/nullb0 > > echo 1 > config/nullb/nullb0/memory_backed > echo 512 > config/nullb/nullb0/blocksize > > # 20 GB > echo 20480 > config/nullb/nullb0/size > echo 1 > config/nullb/nullb0/power > sleep 2 > IDX=`cat config/nullb/nullb0/index` > lsblk | grep null${IDX} > sleep 1 > > mkfs.ext4 /dev/nullb0 > mount /dev/nullb0 ${NULLB_MNT} > sleep 1 > mount | grep nullb > > # 10 GB > dd if=/dev/zero of=${NULLB_FILE} count=2621440 bs=4096 > } > > load_loop() > { > local src=drivers/block/ > local dest=/lib/modules/`uname -r`/kernel/drivers/block > \cp ${src}/loop.ko ${dest}/ > > insmod drivers/block/loop.ko max_loop=1 > sleep 3 > /root/util-linux/losetup --direct-io=off /dev/loop0 ${NULLB_FILE} > sleep 3 > /root/util-linux/losetup > ls -l /dev/loop* > dmesg -c > mkfs.ext4 /dev/loop0 > mount /dev/loop0 ${LOOP_MNT} > mount | grep loop0 > } > > load() > { > make -j $(nproc) M=drivers/block modules > > load_ext4 > load_nullb > load_loop > sleep 1 > sync > sync > sync > } > > unload > load > time ./test > > 4. Test Results :- > ------------------ > > # ./req_op_assign_test.sh > + NULLB_FILE=/mnt/backend/data > + NULLB_MNT=/mnt/backend > + LOOP_MNT=/mnt/loop0 > + unload > + delete_loop > + umount /mnt/loop0 > + losetup -D > + sleep 3 > + delete_nullb > + umount /mnt/backend > + echo 1 > + rmdir config/nullb/nullb0 > + sleep 3 > + unload_modules > + rmmod drivers/block/loop.ko > + rmmod fs/ext4/ext4.ko > + rmmod drivers/block/null_blk.ko > + lsmod > + grep -e ext4 -e loop -e null_blk > + load > ++ nproc > + make -j 32 M=drivers/block modules > CC [M] drivers/block/loop.o > MODPOST 11 modules > CC [M] drivers/block/loop.mod.o > LD [M] drivers/block/loop.ko > + load_ext4 > ++ nproc > + make -j 32 M=fs/ext4 modules > CC [M] fs/ext4/balloc.o > CC [M] fs/ext4/bitmap.o > CC [M] fs/ext4/block_validity.o > CC [M] fs/ext4/dir.o > CC [M] fs/ext4/ext4_jbd2.o > CC [M] fs/ext4/extents.o > CC [M] fs/ext4/extents_status.o > CC [M] fs/ext4/file.o > CC [M] fs/ext4/fsmap.o > CC [M] fs/ext4/fsync.o > CC [M] fs/ext4/hash.o > CC [M] fs/ext4/ialloc.o > CC [M] fs/ext4/indirect.o > CC [M] fs/ext4/inline.o > CC [M] fs/ext4/inode.o > CC [M] fs/ext4/ioctl.o > CC [M] fs/ext4/mballoc.o > CC [M] fs/ext4/migrate.o > CC [M] fs/ext4/mmp.o > CC [M] fs/ext4/move_extent.o > CC [M] fs/ext4/namei.o > CC [M] fs/ext4/page-io.o > CC [M] fs/ext4/readpage.o > CC [M] fs/ext4/resize.o > CC [M] fs/ext4/super.o > CC [M] fs/ext4/symlink.o > CC [M] fs/ext4/sysfs.o > CC [M] fs/ext4/xattr.o > CC [M] fs/ext4/xattr_trusted.o > CC [M] fs/ext4/xattr_user.o > CC [M] fs/ext4/acl.o > CC [M] fs/ext4/xattr_security.o > LD [M] fs/ext4/ext4.o > MODPOST 1 modules > LD [M] fs/ext4/ext4.ko > + local src=fs/ext4/ > ++ uname -r > + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/fs/ext4 > + cp fs/ext4//ext4.ko /lib/modules/5.6.0-rc3lbk+/kernel/fs/ext4/ > + modprobe mbcache > + modprobe jbd2 > + sleep 1 > + insmod fs/ext4/ext4.ko > + sleep 1 > + load_nullb > + local src=drivers/block/ > ++ uname -r > + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/drivers/block > + cp drivers/block//null_blk.ko /lib/modules/5.6.0-rc3lbk+/kernel/drivers/block/ > + modprobe null_blk nr_devices=0 > + sleep 1 > + mkdir config/nullb/nullb0 > + tree config/nullb/nullb0 > config/nullb/nullb0 > ├── badblocks > ├── blocking > ├── blocksize > ├── cache_size > ├── completion_nsec > ├── discard > ├── home_node > ├── hw_queue_depth > ├── index > ├── irqmode > ├── mbps > ├── memory_backed > ├── power > ├── queue_mode > ├── size > ├── submit_queues > ├── use_per_node_hctx > ├── zoned > ├── zone_nr_conv > └── zone_size > > 0 directories, 20 files > + echo 1 > + echo 512 > + echo 20480 > + echo 1 > + sleep 2 > ++ cat config/nullb/nullb0/index > + IDX=0 > + lsblk > + grep null0 > + sleep 1 > + mkfs.ext4 /dev/nullb0 > mke2fs 1.42.9 (28-Dec-2013) > Filesystem label= > OS type: Linux > Block size=4096 (log=2) > Fragment size=4096 (log=2) > Stride=0 blocks, Stripe width=0 blocks > 1310720 inodes, 5242880 blocks > 262144 blocks (5.00%) reserved for the super user > First data block=0 > Maximum filesystem blocks=2153775104 > 160 block groups > 32768 blocks per group, 32768 fragments per group > 8192 inodes per group > Superblock backups stored on blocks: > 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, > 4096000 > > Allocating group tables: done > Writing inode tables: done > Creating journal (32768 blocks): done > Writing superblocks and filesystem accounting information: done > > + mount /dev/nullb0 /mnt/backend > + sleep 1 > + mount > + grep nullb > /dev/nullb0 on /mnt/backend type ext4 (rw,relatime,seclabel) > + dd if=/dev/zero of=/mnt/backend/data count=2621440 bs=4096 > 2621440+0 records in > 2621440+0 records out > 10737418240 bytes (11 GB) copied, 27.4579 s, 391 MB/s > + load_loop > + local src=drivers/block/ > ++ uname -r > + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/drivers/block > + cp drivers/block//loop.ko /lib/modules/5.6.0-rc3lbk+/kernel/drivers/block/ > + insmod drivers/block/loop.ko max_loop=1 > + sleep 3 > + /root/util-linux/losetup --direct-io=off /dev/loop0 /mnt/backend/data > + sleep 3 > + /root/util-linux/losetup > NAME SIZELIMIT OFFSET AUTOCLEAR RO BACK-FILE DIO LOG-SEC > /dev/loop0 0 0 0 0 /mnt/backend/data 0 512 > + ls -l /dev/loop0 /dev/loop-control > brw-rw----. 1 root disk 7, 0 Mar 29 10:28 /dev/loop0 > crw-rw----. 1 root disk 10, 237 Mar 29 10:28 /dev/loop-control > + dmesg -c > [42963.967060] null_blk: module loaded > [42968.419481] EXT4-fs (nullb0): mounted filesystem with ordered data mode. Opts: (null) > [42996.928141] loop: module loaded > + mkfs.ext4 /dev/loop0 > mke2fs 1.42.9 (28-Dec-2013) > Discarding device blocks: done > Filesystem label= > OS type: Linux > Block size=4096 (log=2) > Fragment size=4096 (log=2) > Stride=0 blocks, Stripe width=0 blocks > 655360 inodes, 2621440 blocks > 131072 blocks (5.00%) reserved for the super user > First data block=0 > Maximum filesystem blocks=2151677952 > 80 block groups > 32768 blocks per group, 32768 fragments per group > 8192 inodes per group > Superblock backups stored on blocks: > 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632 > > Allocating group tables: done > Writing inode tables: done > Creating journal (32768 blocks): done > Writing superblocks and filesystem accounting information: done > > + mount /dev/loop0 /mnt/loop0 > + mount > + grep loop0 > /dev/loop0 on /mnt/loop0 type ext4 (rw,relatime,seclabel) > + sleep 1 > + sync > + sync > + sync > + ./test > step=61440 > step=57344 > step=53248 > step=49152 > step=45056 > step=40960 > step=36864 > step=32768 > step=28672 > step=24576 > step=20480 > step=16384 > step=12288 > step=8192 > step=4096 > step=0 > > real 9m34.472s > user 0m0.062s > sys 0m5.783s >
Konstantin, >> The corresponding exported primitive is called >> blkdev_issue_assign_range(). > > What exact semantics of that? REQ_OP_ALLOCATE will be used to compel a device to allocate a block range. What a given block contains after successful allocation is undefined (depends on the device implementation). For block allocation with deterministic zeroing, one must keep using REQ_OP_WRITE_ZEROES with the NOUNMAP flag set.
On 02/04/2020 05.29, Martin K. Petersen wrote: > > Konstantin, > >>> The corresponding exported primitive is called >>> blkdev_issue_assign_range(). >> >> What exact semantics of that? > > REQ_OP_ALLOCATE will be used to compel a device to allocate a block > range. What a given block contains after successful allocation is > undefined (depends on the device implementation). Ok. Then REQ_OP_ALLOCATE should be accounted as discard rather than write. That's decided by helper op_is_discard() which is used only by statistics. It seems REQ_OP_SECURE_ERASE also should be accounted in this way. > > For block allocation with deterministic zeroing, one must keep using > REQ_OP_WRITE_ZEROES with the NOUNMAP flag set. >
On Sun, Mar 29, 2020 at 10:47:10AM -0700, Chaitanya Kulkarni wrote: > Hi, > > This patch-series is based on the original RFC patch series:- > https://www.spinics.net/lists/linux-block/msg47933.html. > > I've designed a rough testcase based on the information present > in the mailing list archive for original RFC, it may need > some corrections from the author. > > If anyone is interested, test results are at the end of this patch. > > Following is the original cover-letter :- > > Information about continuous extent placement may be useful > for some block devices. Say, distributed network filesystems, > which provide block device interface, may use this information > for better blocks placement over the nodes in their cluster, > and for better performance. Block devices, which map a file > on another filesystem (loop), may request the same length extent > on underlining filesystem for less fragmentation and for batching > allocation requests. Also, hypervisors like QEMU may use this > information for optimization of cluster allocations. > > This patchset introduces REQ_OP_ASSIGN_RANGE, which is going > to be used for forwarding user's fallocate(0) requests into > block device internals. It rather similar to existing > REQ_OP_DISCARD, REQ_OP_WRITE_ZEROES, etc. The corresponding > exported primitive is called blkdev_issue_assign_range(). > See [1/3] for the details. > > Patch [2/3] teaches loop driver to handle REQ_OP_ASSIGN_RANGE > requests by calling fallocate(0). > > Patch [3/3] makes ext4 to notify a block device about fallocate(0). Ok, so ext4 has a very limited max allocation size for an extent, so I expect this won't cause huge latency problems. However, what happens when we use XFS, have a 64kB block size, and fallocate() is allocating disk space in continguous 100GB extents and passing those down to the block device? How does this get split by dm devices? Are raid stripes going to dice this into separate stripe unit sized bios, so instead of single large requests we end up with hundreds or thousands or tiny allocation requests being issued? I know that for the loop device, it is going to serialise all IO to the backing file while fallocate is run on it. Hence if you have concurrent IO running, any REQ_OP_ASSIGN_RANGE is going to cause an significant, measurable latency hit to all those IOs in flight. How are we expecting hardware to behave here? Is this a queued command in the scsi/nvme/sata protocols? Or is this, for the moment, just a special snowflake that we can't actually use in production because the hardware just can't handle what we throw at it? IOWs, what sort of latency issues is this operation going to cause on real hardware? Is this going to be like discard? i.e. where we end up not using it at all because so few devices actually handle the massive stream of operations the filesystem will end up sending the device(s) in the course of normal operations? Cheers, Dave.
Hi Dave! > Ok, so ext4 has a very limited max allocation size for an extent, so > I expect this won't cause huge latency problems. However, what > happens when we use XFS, have a 64kB block size, and fallocate() is > allocating disk space in continguous 100GB extents and passing those > down to the block device? Depends on the device. > How does this get split by dm devices? Are raid stripes going to dice > this into separate stripe unit sized bios, so instead of single large > requests we end up with hundreds or thousands or tiny allocation > requests being issued? There is nothing special about this operation. It needs to be handled the same way as all other splits. I.e. ideally coalesced at the bottom of the stack so we can issue larger, contiguous commands to the hardware. > How are we expecting hardware to behave here? Is this a queued > command in the scsi/nvme/sata protocols? Or is this, for the moment, > just a special snowflake that we can't actually use in production > because the hardware just can't handle what we throw at it? For now it's SCSI and queued. Only found in high-end thinly provisioned storage arrays and not in your average SSD. The performance expectation for REQ_OP_ALLOCATE is that it is faster than a write to the same block range since the device potentially needs to do less work. I.e. the device simply needs to decrement the free space and mark the LBAs reserved in a map. It doesn't need to write all the blocks to zero them. If you want zeroed blocks, use REQ_OP_WRITE_ZEROES. > IOWs, what sort of latency issues is this operation going to cause > on real hardware? Is this going to be like discard? i.e. where we > end up not using it at all because so few devices actually handle > the massive stream of operations the filesystem will end up sending > the device(s) in the course of normal operations? The intended use case, from a SCSI perspective, is that on a thinly provisioned device you can use this operation to preallocate blocks so that future writes to the LBAs in question will not fail due to the device being out of space. I.e. you would use this to pin down block ranges where you can not tolerate write failures. The advantage over writing the blocks individually is that dedup won't apply and that the device doesn't actually have to go write all the individual blocks.
On Thu, Apr 02, 2020 at 09:34:43PM -0400, Martin K. Petersen wrote: > > Hi Dave! > > > Ok, so ext4 has a very limited max allocation size for an extent, so > > I expect this won't cause huge latency problems. However, what > > happens when we use XFS, have a 64kB block size, and fallocate() is > > allocating disk space in continguous 100GB extents and passing those > > down to the block device? > > Depends on the device. Great. :( > > How does this get split by dm devices? Are raid stripes going to dice > > this into separate stripe unit sized bios, so instead of single large > > requests we end up with hundreds or thousands or tiny allocation > > requests being issued? > > There is nothing special about this operation. It needs to be handled > the same way as all other splits. I.e. ideally coalesced at the bottom > of the stack so we can issue larger, contiguous commands to the > hardware. > > > How are we expecting hardware to behave here? Is this a queued > > command in the scsi/nvme/sata protocols? Or is this, for the moment, > > just a special snowflake that we can't actually use in production > > because the hardware just can't handle what we throw at it? > > For now it's SCSI and queued. Only found in high-end thinly provisioned > storage arrays and not in your average SSD. So it's a special snowflake :) > The performance expectation for REQ_OP_ALLOCATE is that it is faster > than a write to the same block range since the device potentially needs > to do less work. I.e. the device simply needs to decrement the free > space and mark the LBAs reserved in a map. It doesn't need to write all > the blocks to zero them. If you want zeroed blocks, use > REQ_OP_WRITE_ZEROES. I suspect that the implications of wiring filesystems directly up to this hasn't been thought through entirely.... > > IOWs, what sort of latency issues is this operation going to cause > > on real hardware? Is this going to be like discard? i.e. where we > > end up not using it at all because so few devices actually handle > > the massive stream of operations the filesystem will end up sending > > the device(s) in the course of normal operations? > > The intended use case, from a SCSI perspective, is that on a thinly > provisioned device you can use this operation to preallocate blocks so > that future writes to the LBAs in question will not fail due to the > device being out of space. I.e. you would use this to pin down block > ranges where you can not tolerate write failures. The advantage over > writing the blocks individually is that dedup won't apply and that the > device doesn't actually have to go write all the individual blocks. .... because when backed by thinp storage, plumbing user level fallocate() straight through from the filesystem introduces a trivial, user level storage DOS vector.... i.e. a user can just fallocate a bunch of files and, because the filesystem can do that instantly, can also run the back end array out of space almost instantly. Storage admins are going to love this! Cheers, Dave.
Dave, > .... because when backed by thinp storage, plumbing user level > fallocate() straight through from the filesystem introduces a > trivial, user level storage DOS vector.... > > i.e. a user can just fallocate a bunch of files and, because the > filesystem can do that instantly, can also run the back end array > out of space almost instantly. Storage admins are going to love > this! In the standards space, the allocation concept was mainly aimed at protecting filesystem internals against out-of-space conditions on devices that dedup identical blocks and where simply zeroing the blocks therefore is ineffective. So far we have mainly been talking about fallocate on block devices. How XFS decides to enforce space allocation policy and potentially leverage this plumbing is entirely up to you.
On Thu, Apr 02, 2020 at 08:45:30PM -0700, Martin K. Petersen wrote: > > Dave, > > > .... because when backed by thinp storage, plumbing user level > > fallocate() straight through from the filesystem introduces a > > trivial, user level storage DOS vector.... > > > > i.e. a user can just fallocate a bunch of files and, because the > > filesystem can do that instantly, can also run the back end array > > out of space almost instantly. Storage admins are going to love > > this! > > In the standards space, the allocation concept was mainly aimed at > protecting filesystem internals against out-of-space conditions on > devices that dedup identical blocks and where simply zeroing the blocks > therefore is ineffective. Um, so we're supposed to use space allocation before overwriting existing metadata in the filesystem? So that the underlying storage can reserve space for it before we write it? Which would mean we have to issue a space allocation before we dirty the metadata, which means before we dirty any metadata in a transaction. Which means we'll basically have to redesign the filesystems from the ground up, yes? > So far we have mainly been talking about fallocate on block devices. You might be talking about filesystem metadata and block devices, but this patchset ends up connecting ext4's user data fallocate() to the block device, thereby allowing users to reserve space directly in the underlying block device and directly exposing this issue to userspace. I can only go on what is presented to me in patches - this patchset nothing to do with filesystem metadata nor preventing ENOSPC issues with internal filesystem updates. XFS is no different to ext4 or btrfs here - the filesystem doesn't matter because all of them can fallocate() terabytes of space in a second or two these days.... > How XFS decides to enforce space allocation policy and potentially > leverage this plumbing is entirely up to you. Do I understand this correctly? i.e. that it is the filesystem's responsibility to prevent users from preallocating more space than exists in an underlying storage pool that has been intentionally hidden from the filesystem so it can be underprovisioned? IOWs, I'm struggling to understand exactly how the "standards space" think filesystems are supposed to be using this feature whilst also preventing unprivileged exhaustion of a underprovisioned storage pool they know nothing about. Cheers, Dave.
Hi Dave! >> In the standards space, the allocation concept was mainly aimed at >> protecting filesystem internals against out-of-space conditions on >> devices that dedup identical blocks and where simply zeroing the blocks >> therefore is ineffective. > Um, so we're supposed to use space allocation before overwriting > existing metadata in the filesystem? Not before overwriting, no. Once you have allocated an LBA it remains allocated until you discard it. > So that the underlying storage can reserve space for it before we > write it? Which would mean we have to issue a space allocation before > we dirty the metadata, which means before we dirty any metadata in a > transaction. Which means we'll basically have to redesign the > filesystems from the ground up, yes? My understanding is that this facility was aimed at filesystems that do not dynamically allocate metadata. The intent was that mkfs would preallocate the metadata LBA ranges, not the filesystem. For filesystems that allocate metadata dynamically, then yes, an additional step is required if you want to pin the LBAs. > You might be talking about filesystem metadata and block devices, > but this patchset ends up connecting ext4's user data fallocate() to > the block device, thereby allowing users to reserve space directly > in the underlying block device and directly exposing this issue to > userspace. I missed that Chaitanya's repost of this series included the ext4 patch. Sorry! >> How XFS decides to enforce space allocation policy and potentially >> leverage this plumbing is entirely up to you. > > Do I understand this correctly? i.e. that it is the filesystem's > responsibility to prevent users from preallocating more space than > exists in an underlying storage pool that has been intentionally > hidden from the filesystem so it can be underprovisioned? No. But as an administrative policy it is useful to prevent runaway applications from writing a petabyte of random garbage to media. My point was that it is up to you and the other filesystem developers to decide how you want to leverage the low-level allocation capability and how you want to provide it to processes. And whether CAP_SYS_ADMIN, ulimit, or something else is the appropriate policy interface for this. In terms of thin provisioning and space management there are various thresholds that may be reported by the device. In past discussions there haven't been much interest in getting these exposed. It is also unclear to me whether it is actually beneficial to send low space warnings to hundreds or thousands of hosts attached to an array. In many cases the individual server admins are not even the right audience. The most common notification mechanism is a message to the storage array admin saying "click here to buy more disk". If you feel there is merit in having the kernel emit the threshold warnings you could use as a feedback mechanism, I can absolutely look into that.
On Wed, Apr 08, 2020 at 12:10:12AM -0400, Martin K. Petersen wrote: > > Hi Dave! > > >> In the standards space, the allocation concept was mainly aimed at > >> protecting filesystem internals against out-of-space conditions on > >> devices that dedup identical blocks and where simply zeroing the blocks > >> therefore is ineffective. > > > Um, so we're supposed to use space allocation before overwriting > > existing metadata in the filesystem? > > Not before overwriting, no. Once you have allocated an LBA it remains > allocated until you discard it. That is not a consistent argument. If the data has been deduped and we overwrite, the storage array has to allocate new physical space for an overwrite to an existing LBA. i.e. deduped data has multiple LBAs pointing to the same physical storage. Any overwrite of an LBA that maps to mulitply referenced physical storage requires the storage array to allocate new physical space for that overwrite. i.e. allocation is not determined by whether the LBA has been written to, "pinned" or not - it's whether the act of writing to that LBA requires the storage to allocate new space to allow the write to proceed. That's my point here - one particular shared data overwrite case is being special cased by preallocation (avoiding dedupe of zero filled data) to prevent ENOSPC, ignoring all the other cases where we overwrite shared non-zero data and will also require new physical space for the new data. In all those cases, the storage has to take the same action - allocation on overwrite - and so all of them are susceptible to ENOSPC. > > So that the underlying storage can reserve space for it before we > > write it? Which would mean we have to issue a space allocation before > > we dirty the metadata, which means before we dirty any metadata in a > > transaction. Which means we'll basically have to redesign the > > filesystems from the ground up, yes? > > My understanding is that this facility was aimed at filesystems that do > not dynamically allocate metadata. The intent was that mkfs would > preallocate the metadata LBA ranges, not the filesystem. For filesystems > that allocate metadata dynamically, then yes, an additional step is > required if you want to pin the LBAs. Ok, so you are confirming what I thought: it's almost completely useless to us. i.e. this requires issuing IO to "reserve" space whilst preserving data before every metadata object goes from clean to dirty in memory. But the problem with that is we don't know how much metadata we are going to dirty in any specific operation. Worse is that we don't know exactly *what* metadata we will modify until we walk structures and do lookups, which often happen after we've dirtied other structures. An ENOSPC from a space reservation at that point is fatal to the filesystem anyway, so there's no point in even trying to do this. Like I said, functionality like this cannot be retrofitted to existing filesysetms. IOWs, this is pretty much useless functionality for the filesystem layer, and if the only use is for some mythical filesystem with completely static metadata then the standards space really jumped the shark on this one.... > > You might be talking about filesystem metadata and block devices, > > but this patchset ends up connecting ext4's user data fallocate() to > > the block device, thereby allowing users to reserve space directly > > in the underlying block device and directly exposing this issue to > > userspace. > > I missed that Chaitanya's repost of this series included the ext4 patch. > Sorry! > > >> How XFS decides to enforce space allocation policy and potentially > >> leverage this plumbing is entirely up to you. > > > > Do I understand this correctly? i.e. that it is the filesystem's > > responsibility to prevent users from preallocating more space than > > exists in an underlying storage pool that has been intentionally > > hidden from the filesystem so it can be underprovisioned? > > No. But as an administrative policy it is useful to prevent runaway > applications from writing a petabyte of random garbage to media. My > point was that it is up to you and the other filesystem developers to > decide how you want to leverage the low-level allocation capability and > how you want to provide it to processes. And whether CAP_SYS_ADMIN, > ulimit, or something else is the appropriate policy interface for this. My cynical translation: the storage standards space haven't given any thought to how it can be used and/or administered in the real world. Pass the buck - let the filesystem people work that out. What I'm hearing is that this wasn't designed for typical filesystem use, it wasn't designed for typical user application use, and how to prevent abuse wasn't thought about at all. That sounds like a big fat NACK to me.... > In terms of thin provisioning and space management there are various > thresholds that may be reported by the device. In past discussions there > haven't been much interest in getting these exposed. It is also unclear > to me whether it is actually beneficial to send low space warnings to > hundreds or thousands of hosts attached to an array. In many cases the > individual server admins are not even the right audience. The most > common notification mechanism is a message to the storage array admin > saying "click here to buy more disk". Notifications are not relevant to preallocation functionality at all. -Dave.
Dave, >> Not before overwriting, no. Once you have allocated an LBA it remains >> allocated until you discard it. > Ok, so you are confirming what I thought: it's almost completely > useless to us. > > i.e. this requires issuing IO to "reserve" space whilst preserving > data before every metadata object goes from clean to dirty in memory. You can only reserve the space prior to writing a block for the first time. Once an LBA has been written ("Mapped" in the SCSI state machine), it remains allocated until it is explicitly deallocated (via a discard/Unmap operation). This part of the SCSI spec was written eons ago under the assumption that when a physical resource backing a given LBA had been established, you could write the block over and over without having to allocate new space. This used to be true, but obviously the introduction of de-duplication blew a major hole in that. I have been perusing the spec over and over trying to understand how block provisioning state transitions are defined when dedup is in the picture. However, much is left unexplained. As a result, I reached out to various folks. Including the people who worked on this feature in the standards way back. And the response that I get from them is that allocation operation got irreparably broken when support for de-duplication was added to the spec. Nobody attempted to fix the state transitions since most vendors only cared about deallocation. Consequently specifying the exact behavior of the allocation operation in the context of dedup fell by the wayside. The recommendation I got was that we should not rely on this feature despite it being advertised as supported by the storage. I looked at whether it was feasible to support it on non-dedup devices only, but it does not look like it's worthwhile to pursue. And as a result there is no need for block layer allocation operation to have parity with SCSI. Although we may want to keep NVMe in mind when defining the semantics.
Hi, This patch-series is based on the original RFC patch series:- https://www.spinics.net/lists/linux-block/msg47933.html. I've designed a rough testcase based on the information present in the mailing list archive for original RFC, it may need some corrections from the author. If anyone is interested, test results are at the end of this patch. Following is the original cover-letter :- Information about continuous extent placement may be useful for some block devices. Say, distributed network filesystems, which provide block device interface, may use this information for better blocks placement over the nodes in their cluster, and for better performance. Block devices, which map a file on another filesystem (loop), may request the same length extent on underlining filesystem for less fragmentation and for batching allocation requests. Also, hypervisors like QEMU may use this information for optimization of cluster allocations. This patchset introduces REQ_OP_ASSIGN_RANGE, which is going to be used for forwarding user's fallocate(0) requests into block device internals. It rather similar to existing REQ_OP_DISCARD, REQ_OP_WRITE_ZEROES, etc. The corresponding exported primitive is called blkdev_issue_assign_range(). See [1/3] for the details. Patch [2/3] teaches loop driver to handle REQ_OP_ASSIGN_RANGE requests by calling fallocate(0). Patch [3/3] makes ext4 to notify a block device about fallocate(0). Here is a simple test I did: https://gist.github.com/tkhai/5b788651cdb74c1dbff3500745878856 I attached a file on ext4 to loop. Then, created ext4 partition on loop device and started the test in the partition. Direct-io is enabled on loop. The test fallocates 4G file and writes from some offset with given step, then it chooses another offset and repeats. After the test all the blocks in the file become written. The results shows that batching extents-assigning requests improves the performance: Before patchset: real ~ 1min 27sec After patchset: real ~ 1min 16sec (18% better) Ordinary fallocate() before writes improves the performance by batching the requests. These results just show, the same is in case of forwarding extents information to underlining filesystem. Regards, Chaitanya Changes from RFC:- 1. Add missing plumbing for REQ_OP_ASSIGN_RANGE similar to write-zeores. 2. Add a prep patch to create a helper to submit payloadless bios. 3. Design a testcases around the description present in the cover-letter. Chaitanya Kulkarni (1): block: create payloadless issue bio helper Kirill Tkhai (3): block: Add support for REQ_OP_ASSIGN_RANGE loop: Forward REQ_OP_ASSIGN_RANGE into fallocate(0) ext4: Notify block device about alloc-assigned blk block/blk-core.c | 5 ++ block/blk-lib.c | 115 +++++++++++++++++++++++++++++++------- block/blk-merge.c | 21 +++++++ block/blk-settings.c | 19 +++++++ block/blk-zoned.c | 1 + block/bounce.c | 1 + drivers/block/loop.c | 5 ++ fs/ext4/ext4.h | 2 + fs/ext4/extents.c | 12 +++- include/linux/bio.h | 9 ++- include/linux/blk_types.h | 2 + include/linux/blkdev.h | 34 +++++++++++ 12 files changed, 201 insertions(+), 25 deletions(-) 1. Setup :- ----------- # git log --oneline -5 c64a4c781915 (HEAD -> req-op-assign-range) ext4: Notify block device about alloc-assigned blk 000cbc6720a4 loop: Forward REQ_OP_ASSIGN_RANGE into fallocate(0) 89ceed8cac80 block: Add support for REQ_OP_ASSIGN_RANGE a798743e87e7 block: create payloadless issue bio helper b53df2e7442c (tag: block-5.6-2020-03-13) block: Fix partition support for host aware zoned block devices # cat /proc/kallsyms | grep -i blkdev_issue_assign_range ffffffffa3264a80 T blkdev_issue_assign_range ffffffffa4027184 r __ksymtab_blkdev_issue_assign_range ffffffffa40524be r __kstrtabns_blkdev_issue_assign_range ffffffffa405a8eb r __kstrtab_blkdev_issue_assign_range 2. Test program, will be moved to blktest once code is upstream :- ----------------- #define _GNU_SOURCE #include <sys/types.h> #include <unistd.h> #include <stdlib.h> #include <stdio.h> #include <fcntl.h> #include <errno.h> #define BLOCK_SIZE 4096 #define STEP (BLOCK_SIZE * 16) #define SIZE (1024 * 1024 * 1024ULL) int main(int argc, char *argv[]) { int fd, step, ret = 0; unsigned long i; void *buf; if (posix_memalign(&buf, BLOCK_SIZE, SIZE)) { perror("alloc"); exit(1); } fd = open("/mnt/loop0/file.img", O_RDWR | O_CREAT | O_DIRECT); if (fd < 0) { perror("open"); exit(1); } if (ftruncate(fd, SIZE)) { perror("ftruncate"); exit(1); } ret = fallocate(fd, 0, 0, SIZE); if (ret) { perror("fallocate"); exit(1); } for (step = STEP - BLOCK_SIZE; step >= 0; step -= BLOCK_SIZE) { printf("step=%u\n", step); for (i = step; i < SIZE; i += STEP) { errno = 0; if (pwrite(fd, buf, BLOCK_SIZE, i) != BLOCK_SIZE) { perror("pwrite"); exit(1); } } if (fsync(fd)) { perror("fsync"); exit(1); } } return 0; } 3. Test script, will be moved to blktests once code is upstream :- ------------------------------------------------------------------ # cat req_op_assign_test.sh #!/bin/bash -x NULLB_FILE="/mnt/backend/data" NULLB_MNT="/mnt/backend" LOOP_MNT="/mnt/loop0" delete_loop() { umount ${LOOP_MNT} losetup -D sleep 3 } delete_nullb() { umount ${NULLB_MNT} echo 1 > config/nullb/nullb0/power rmdir config/nullb/nullb0 sleep 3 } unload_modules() { rmmod drivers/block/loop.ko rmmod fs/ext4/ext4.ko rmmod drivers/block/null_blk.ko lsmod | grep -e ext4 -e loop -e null_blk } unload() { delete_loop delete_nullb unload_modules } load_ext4() { make -j $(nproc) M=fs/ext4 modules local src=fs/ext4/ local dest=/lib/modules/`uname -r`/kernel/fs/ext4 \cp ${src}/ext4.ko ${dest}/ modprobe mbcache modprobe jbd2 sleep 1 insmod fs/ext4/ext4.ko sleep 1 } load_nullb() { local src=drivers/block/ local dest=/lib/modules/`uname -r`/kernel/drivers/block \cp ${src}/null_blk.ko ${dest}/ modprobe null_blk nr_devices=0 sleep 1 mkdir config/nullb/nullb0 tree config/nullb/nullb0 echo 1 > config/nullb/nullb0/memory_backed echo 512 > config/nullb/nullb0/blocksize # 20 GB echo 20480 > config/nullb/nullb0/size echo 1 > config/nullb/nullb0/power sleep 2 IDX=`cat config/nullb/nullb0/index` lsblk | grep null${IDX} sleep 1 mkfs.ext4 /dev/nullb0 mount /dev/nullb0 ${NULLB_MNT} sleep 1 mount | grep nullb # 10 GB dd if=/dev/zero of=${NULLB_FILE} count=2621440 bs=4096 } load_loop() { local src=drivers/block/ local dest=/lib/modules/`uname -r`/kernel/drivers/block \cp ${src}/loop.ko ${dest}/ insmod drivers/block/loop.ko max_loop=1 sleep 3 /root/util-linux/losetup --direct-io=off /dev/loop0 ${NULLB_FILE} sleep 3 /root/util-linux/losetup ls -l /dev/loop* dmesg -c mkfs.ext4 /dev/loop0 mount /dev/loop0 ${LOOP_MNT} mount | grep loop0 } load() { make -j $(nproc) M=drivers/block modules load_ext4 load_nullb load_loop sleep 1 sync sync sync } unload load time ./test 4. Test Results :- ------------------ # ./req_op_assign_test.sh + NULLB_FILE=/mnt/backend/data + NULLB_MNT=/mnt/backend + LOOP_MNT=/mnt/loop0 + unload + delete_loop + umount /mnt/loop0 + losetup -D + sleep 3 + delete_nullb + umount /mnt/backend + echo 1 + rmdir config/nullb/nullb0 + sleep 3 + unload_modules + rmmod drivers/block/loop.ko + rmmod fs/ext4/ext4.ko + rmmod drivers/block/null_blk.ko + lsmod + grep -e ext4 -e loop -e null_blk + load ++ nproc + make -j 32 M=drivers/block modules CC [M] drivers/block/loop.o MODPOST 11 modules CC [M] drivers/block/loop.mod.o LD [M] drivers/block/loop.ko + load_ext4 ++ nproc + make -j 32 M=fs/ext4 modules CC [M] fs/ext4/balloc.o CC [M] fs/ext4/bitmap.o CC [M] fs/ext4/block_validity.o CC [M] fs/ext4/dir.o CC [M] fs/ext4/ext4_jbd2.o CC [M] fs/ext4/extents.o CC [M] fs/ext4/extents_status.o CC [M] fs/ext4/file.o CC [M] fs/ext4/fsmap.o CC [M] fs/ext4/fsync.o CC [M] fs/ext4/hash.o CC [M] fs/ext4/ialloc.o CC [M] fs/ext4/indirect.o CC [M] fs/ext4/inline.o CC [M] fs/ext4/inode.o CC [M] fs/ext4/ioctl.o CC [M] fs/ext4/mballoc.o CC [M] fs/ext4/migrate.o CC [M] fs/ext4/mmp.o CC [M] fs/ext4/move_extent.o CC [M] fs/ext4/namei.o CC [M] fs/ext4/page-io.o CC [M] fs/ext4/readpage.o CC [M] fs/ext4/resize.o CC [M] fs/ext4/super.o CC [M] fs/ext4/symlink.o CC [M] fs/ext4/sysfs.o CC [M] fs/ext4/xattr.o CC [M] fs/ext4/xattr_trusted.o CC [M] fs/ext4/xattr_user.o CC [M] fs/ext4/acl.o CC [M] fs/ext4/xattr_security.o LD [M] fs/ext4/ext4.o MODPOST 1 modules LD [M] fs/ext4/ext4.ko + local src=fs/ext4/ ++ uname -r + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/fs/ext4 + cp fs/ext4//ext4.ko /lib/modules/5.6.0-rc3lbk+/kernel/fs/ext4/ + modprobe mbcache + modprobe jbd2 + sleep 1 + insmod fs/ext4/ext4.ko + sleep 1 + load_nullb + local src=drivers/block/ ++ uname -r + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/drivers/block + cp drivers/block//null_blk.ko /lib/modules/5.6.0-rc3lbk+/kernel/drivers/block/ + modprobe null_blk nr_devices=0 + sleep 1 + mkdir config/nullb/nullb0 + tree config/nullb/nullb0 config/nullb/nullb0 ├── badblocks ├── blocking ├── blocksize ├── cache_size ├── completion_nsec ├── discard ├── home_node ├── hw_queue_depth ├── index ├── irqmode ├── mbps ├── memory_backed ├── power ├── queue_mode ├── size ├── submit_queues ├── use_per_node_hctx ├── zoned ├── zone_nr_conv └── zone_size 0 directories, 20 files + echo 1 + echo 512 + echo 20480 + echo 1 + sleep 2 ++ cat config/nullb/nullb0/index + IDX=0 + lsblk + grep null0 + sleep 1 + mkfs.ext4 /dev/nullb0 mke2fs 1.42.9 (28-Dec-2013) Filesystem label= OS type: Linux Block size=4096 (log=2) Fragment size=4096 (log=2) Stride=0 blocks, Stripe width=0 blocks 1310720 inodes, 5242880 blocks 262144 blocks (5.00%) reserved for the super user First data block=0 Maximum filesystem blocks=2153775104 160 block groups 32768 blocks per group, 32768 fragments per group 8192 inodes per group Superblock backups stored on blocks: 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, 4096000 Allocating group tables: done Writing inode tables: done Creating journal (32768 blocks): done Writing superblocks and filesystem accounting information: done + mount /dev/nullb0 /mnt/backend + sleep 1 + mount + grep nullb /dev/nullb0 on /mnt/backend type ext4 (rw,relatime,seclabel) + dd if=/dev/zero of=/mnt/backend/data count=2621440 bs=4096 2621440+0 records in 2621440+0 records out 10737418240 bytes (11 GB) copied, 27.4579 s, 391 MB/s + load_loop + local src=drivers/block/ ++ uname -r + local dest=/lib/modules/5.6.0-rc3lbk+/kernel/drivers/block + cp drivers/block//loop.ko /lib/modules/5.6.0-rc3lbk+/kernel/drivers/block/ + insmod drivers/block/loop.ko max_loop=1 + sleep 3 + /root/util-linux/losetup --direct-io=off /dev/loop0 /mnt/backend/data + sleep 3 + /root/util-linux/losetup NAME SIZELIMIT OFFSET AUTOCLEAR RO BACK-FILE DIO LOG-SEC /dev/loop0 0 0 0 0 /mnt/backend/data 0 512 + ls -l /dev/loop0 /dev/loop-control brw-rw----. 1 root disk 7, 0 Mar 29 10:28 /dev/loop0 crw-rw----. 1 root disk 10, 237 Mar 29 10:28 /dev/loop-control + dmesg -c [42963.967060] null_blk: module loaded [42968.419481] EXT4-fs (nullb0): mounted filesystem with ordered data mode. Opts: (null) [42996.928141] loop: module loaded + mkfs.ext4 /dev/loop0 mke2fs 1.42.9 (28-Dec-2013) Discarding device blocks: done Filesystem label= OS type: Linux Block size=4096 (log=2) Fragment size=4096 (log=2) Stride=0 blocks, Stripe width=0 blocks 655360 inodes, 2621440 blocks 131072 blocks (5.00%) reserved for the super user First data block=0 Maximum filesystem blocks=2151677952 80 block groups 32768 blocks per group, 32768 fragments per group 8192 inodes per group Superblock backups stored on blocks: 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632 Allocating group tables: done Writing inode tables: done Creating journal (32768 blocks): done Writing superblocks and filesystem accounting information: done + mount /dev/loop0 /mnt/loop0 + mount + grep loop0 /dev/loop0 on /mnt/loop0 type ext4 (rw,relatime,seclabel) + sleep 1 + sync + sync + sync + ./test step=61440 step=57344 step=53248 step=49152 step=45056 step=40960 step=36864 step=32768 step=28672 step=24576 step=20480 step=16384 step=12288 step=8192 step=4096 step=0 real 9m34.472s user 0m0.062s sys 0m5.783s