[#970] fstree: Add linux-specific file writer using O_TMPFILE

O_TMPFILE is implemented for all modern FSes and it's much easier and safer to
use. If application crashes in the middle of writing this file would be gone
and won't leave any garbage.

Notice that this implementation makes a different choice wrt EEXIST handling,
generic one always overwrites, while this one keeps the old data.

There is no real performance difference.

SSD (slow&old), XFS, Core i7-8565U:

Sync
```
name                                  old time/op    new time/op    delta
Put/size=1024,thread=1/fstree-8         1.74ms ± 3%    0.06ms ± 7%  -96.31%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8        10.0ms ±41%     1.1ms ±18%  -88.95%  (p=0.000 n=9+10)
Put/size=1024,thread=100/fstree-8       32.3ms ±60%     6.5ms ±14%  -79.97%  (p=0.000 n=10+10)
Put/size=1048576,thread=1/fstree-8      17.8ms ±90%     3.4ms ±70%  -81.08%  (p=0.000 n=10+10)
Put/size=1048576,thread=20/fstree-8     103ms ±174%    112ms ±158%     ~     (p=0.971 n=10+10)
Put/size=1048576,thread=100/fstree-8     949ms ±78%    583ms ±132%     ~     (p=0.089 n=10+10)

name                                  old alloc/op   new alloc/op   delta
Put/size=1024,thread=1/fstree-8         3.17kB ± 1%    1.96kB ± 0%  -38.09%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8        59.6kB ± 1%    39.2kB ± 1%  -34.30%  (p=0.000 n=8+10)
Put/size=1024,thread=100/fstree-8        299kB ± 0%     198kB ± 0%  -33.90%  (p=0.000 n=7+9)
Put/size=1048576,thread=1/fstree-8      3.38kB ± 1%    2.36kB ± 1%  -30.22%  (p=0.000 n=10+10)
Put/size=1048576,thread=20/fstree-8     65.7kB ± 4%    47.7kB ± 6%  -27.27%  (p=0.000 n=10+10)
Put/size=1048576,thread=100/fstree-8     351kB ± 8%     245kB ± 8%  -30.22%  (p=0.000 n=10+10)

name                                  old allocs/op  new allocs/op  delta
Put/size=1024,thread=1/fstree-8           30.3 ± 2%      21.0 ± 0%  -30.69%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8           554 ± 1%       413 ± 0%  -25.35%  (p=0.000 n=8+10)
Put/size=1024,thread=100/fstree-8        2.77k ± 0%     2.07k ± 0%  -25.27%  (p=0.000 n=7+10)
Put/size=1048576,thread=1/fstree-8        32.0 ± 0%      25.0 ± 0%  -21.88%  (p=0.000 n=9+8)
Put/size=1048576,thread=20/fstree-8        609 ± 5%       494 ± 6%  -18.93%  (p=0.000 n=10+10)
Put/size=1048576,thread=100/fstree-8     3.25k ± 9%     2.50k ± 8%  -23.21%  (p=0.000 n=10+10)
```

No sync
```
name                                  old time/op    new time/op    delta
Put/size=1024,thread=1/fstree-8         71.3µs ±10%    59.8µs ±10%  -16.21%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8        1.43ms ± 6%    1.22ms ±13%  -14.53%  (p=0.000 n=10+10)
Put/size=1024,thread=100/fstree-8       8.12ms ± 3%    6.36ms ± 2%  -21.67%  (p=0.000 n=8+9)
Put/size=1048576,thread=1/fstree-8      1.88ms ±70%    1.61ms ±78%     ~     (p=0.393 n=10+10)
Put/size=1048576,thread=20/fstree-8     32.7ms ±28%   34.2ms ±112%     ~     (p=0.968 n=9+10)
Put/size=1048576,thread=100/fstree-8     262ms ±56%     226ms ±34%     ~     (p=0.447 n=10+9)

name                                  old alloc/op   new alloc/op   delta
Put/size=1024,thread=1/fstree-8         2.89kB ± 0%    1.96kB ± 0%  -32.28%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8        58.2kB ± 0%    39.5kB ± 0%  -32.09%  (p=0.000 n=8+8)
Put/size=1024,thread=100/fstree-8        291kB ± 0%     198kB ± 0%  -32.19%  (p=0.000 n=9+9)
Put/size=1048576,thread=1/fstree-8      3.05kB ± 1%    2.13kB ± 1%  -30.16%  (p=0.000 n=10+9)
Put/size=1048576,thread=20/fstree-8     62.6kB ± 0%    44.3kB ± 0%  -29.23%  (p=0.000 n=9+9)
Put/size=1048576,thread=100/fstree-8     302kB ± 0%     210kB ± 1%  -30.39%  (p=0.000 n=9+9)

name                                  old allocs/op  new allocs/op  delta
Put/size=1024,thread=1/fstree-8           27.0 ± 0%      21.0 ± 0%  -22.22%  (p=0.000 n=10+10)
Put/size=1024,thread=20/fstree-8           539 ± 0%       415 ± 0%  -22.98%  (p=0.000 n=10+10)
Put/size=1024,thread=100/fstree-8        2.69k ± 0%     2.07k ± 0%  -23.09%  (p=0.000 n=9+9)
Put/size=1048576,thread=1/fstree-8        28.0 ± 0%      22.3 ± 3%  -20.36%  (p=0.000 n=8+10)
Put/size=1048576,thread=20/fstree-8        577 ± 0%       458 ± 0%  -20.72%  (p=0.000 n=9+9)
Put/size=1048576,thread=100/fstree-8     2.76k ± 0%     2.15k ± 0%  -22.05%  (p=0.000 n=9+8)
```

HDD (LVM), ext4, Ryzen 5 1600:

Sync
```
                                      │ fs.sync-generic │            fs.sync-linux            │
                                      │     sec/op      │    sec/op     vs base               │
Put/size=1024,thread=1/fstree-12           34.70m ± 19%   33.59m ± 16%       ~ (p=0.529 n=10)
Put/size=1024,thread=20/fstree-12          188.8m ±  8%   189.2m ± 16%       ~ (p=0.739 n=10)
Put/size=1024,thread=100/fstree-12         264.8m ± 22%   273.6m ± 28%       ~ (p=0.353 n=10)
Put/size=1048576,thread=1/fstree-12        54.90m ± 14%   47.08m ± 18%       ~ (p=0.063 n=10)
Put/size=1048576,thread=20/fstree-12       244.1m ± 14%   220.4m ± 22%       ~ (p=0.579 n=10)
Put/size=1048576,thread=100/fstree-12      847.2m ±  5%   893.6m ±  3%  +5.48% (p=0.000 n=10)
geomean                                    164.3m         158.9m        -3.29%

                                      │ fs.sync-generic │            fs.sync-linux             │
                                      │      B/op       │     B/op      vs base                │
Put/size=1024,thread=1/fstree-12           3.375Ki ± 1%   2.471Ki ± 1%  -26.80% (p=0.000 n=10)
Put/size=1024,thread=20/fstree-12          66.62Ki ± 6%   49.21Ki ± 6%  -26.15% (p=0.000 n=10)
Put/size=1024,thread=100/fstree-12         319.2Ki ± 1%   230.9Ki ± 2%  -27.64% (p=0.000 n=10)
Put/size=1048576,thread=1/fstree-12        3.457Ki ± 1%   2.559Ki ± 1%  -25.97% (p=0.000 n=10)
Put/size=1048576,thread=20/fstree-12       66.91Ki ± 1%   49.16Ki ± 1%  -26.52% (p=0.000 n=10)
Put/size=1048576,thread=100/fstree-12      338.8Ki ± 2%   252.3Ki ± 3%  -25.54% (p=0.000 n=10)
geomean                                    42.17Ki        31.02Ki       -26.44%

                                      │ fs.sync-generic │            fs.sync-linux            │
                                      │    allocs/op    │  allocs/op   vs base                │
Put/size=1024,thread=1/fstree-12             33.00 ± 0%    27.00 ± 0%  -18.18% (p=0.000 n=10)
Put/size=1024,thread=20/fstree-12            639.5 ± 1%    519.0 ± 2%  -18.84% (p=0.000 n=10)
Put/size=1024,thread=100/fstree-12          3.059k ± 1%   2.478k ± 2%  -18.99% (p=0.000 n=10)
Put/size=1048576,thread=1/fstree-12          33.50 ± 1%    28.00 ± 4%  -16.42% (p=0.000 n=10)
Put/size=1048576,thread=20/fstree-12         638.5 ± 1%    520.0 ± 1%  -18.56% (p=0.000 n=10)
Put/size=1048576,thread=100/fstree-12       3.209k ± 2%   2.655k ± 2%  -17.28% (p=0.000 n=10)
geomean                                      405.3         332.1       -18.05%
```

No sync
```
                                      │ fs.nosync-generic │             fs.nosync-linux              │
                                      │      sec/op       │    sec/op     vs base                    │
Put/size=1024,thread=1/fstree-12           148.2µ ± 20%     136.6µ ± 19%   -7.89% (p=0.029 n=10)
Put/size=1024,thread=20/fstree-12          1.140m ± 26%     1.364m ± 16%        ~ (p=0.143 n=10)
Put/size=1024,thread=100/fstree-12         11.93m ± 68%     26.89m ± 62%        ~ (p=0.123 n=10)
Put/size=1048576,thread=1/fstree-12        1.302m ±  3%     1.287m ±  5%        ~ (p=0.481 n=10)
Put/size=1048576,thread=20/fstree-12       77.52m ±  8%     74.07m ±  7%        ~ (p=0.278 n=10+9)
Put/size=1048576,thread=100/fstree-12      226.1m ±   ∞ ¹
geomean                                    5.986m           3.434m        +18.60%                  ²
¹ need >= 6 samples for confidence interval at level 0.95
² benchmark set differs from baseline; geomeans may not be comparable

                                      │ fs.nosync-generic │             fs.nosync-linux              │
                                      │       B/op        │     B/op      vs base                    │
Put/size=1024,thread=1/fstree-12           2.879Ki ± 0%     1.972Ki ± 0%  -31.51% (p=0.000 n=10)
Put/size=1024,thread=20/fstree-12          55.94Ki ± 1%     37.90Ki ± 1%  -32.25% (p=0.000 n=10)
Put/size=1024,thread=100/fstree-12         272.6Ki ± 0%     182.1Ki ± 9%  -33.21% (p=0.000 n=10)
Put/size=1048576,thread=1/fstree-12        3.158Ki ± 0%     2.259Ki ± 0%  -28.46% (p=0.000 n=10)
Put/size=1048576,thread=20/fstree-12       58.87Ki ± 0%     41.03Ki ± 0%  -30.30% (p=0.000 n=10+9)
Put/size=1048576,thread=100/fstree-12      299.8Ki ±  ∞ ¹
geomean                                    36.71Ki          16.60Ki       -31.17%                  ²
¹ need >= 6 samples for confidence interval at level 0.95
² benchmark set differs from baseline; geomeans may not be comparable

                                      │ fs.nosync-generic │            fs.nosync-linux            │
                                      │     allocs/op     │  allocs/op   vs base                  │
Put/size=1024,thread=1/fstree-12             28.00 ± 0%      22.00 ± 0%  -21.43% (p=0.000 n=10)
Put/size=1024,thread=20/fstree-12            530.0 ± 0%      407.5 ± 1%  -23.11% (p=0.000 n=10)
Put/size=1024,thread=100/fstree-12          2.567k ± 0%     1.956k ± 9%  -23.77% (p=0.000 n=10)
Put/size=1048576,thread=1/fstree-12          30.00 ± 0%      24.00 ± 0%  -20.00% (p=0.000 n=10)
Put/size=1048576,thread=20/fstree-12         553.5 ± 0%      434.0 ± 0%  -21.59% (n=10+9)
Put/size=1048576,thread=100/fstree-12       2.803k ±  ∞ ¹
geomean                                      347.9           178.8       -21.99%                ²
¹ need >= 6 samples for confidence interval at level 0.95
² benchmark set differs from baseline; geomeans may not be comparable
```

Signed-off-by: Roman Khimov <roman@nspcc.ru>
Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
pull/975/head
Roman Khimov 2023-09-09 11:16:06 +03:00 committed by Evgenii Stratonikov
parent ff488b53a1
commit fc31b9c947
3 changed files with 51 additions and 1 deletions

2
go.mod
View File

@ -37,6 +37,7 @@ require (
go.uber.org/zap v1.26.0
golang.org/x/exp v0.0.0-20240119083558-1b970713d09a
golang.org/x/sync v0.6.0
golang.org/x/sys v0.16.0
golang.org/x/term v0.16.0
google.golang.org/grpc v1.61.0
google.golang.org/protobuf v1.32.0
@ -120,7 +121,6 @@ require (
go.uber.org/multierr v1.11.0 // indirect
golang.org/x/crypto v0.18.0 // indirect
golang.org/x/net v0.20.0 // indirect
golang.org/x/sys v0.16.0 // indirect
golang.org/x/text v0.14.0 // indirect
google.golang.org/genproto/googleapis/api v0.0.0-20240123012728-ef4313101c80 // indirect
google.golang.org/genproto/googleapis/rpc v0.0.0-20240123012728-ef4313101c80 // indirect

View File

@ -1,3 +1,5 @@
//go:build !linux
package fstree
import (

View File

@ -0,0 +1,48 @@
//go:build linux
package fstree
import (
"errors"
"strconv"
"git.frostfs.info/TrueCloudLab/frostfs-node/pkg/local_object_storage/blobstor/common"
"golang.org/x/sys/unix"
)
func (t *FSTree) writeData(p string, data []byte) error {
err := t.writeFile(p, data)
if errors.Is(err, unix.ENOSPC) {
return common.ErrNoSpace
}
return err
}
func (t *FSTree) writeFile(p string, data []byte) error {
flags := unix.O_WRONLY | unix.O_TMPFILE | unix.O_CLOEXEC
if !t.noSync {
flags |= unix.O_DSYNC
}
fd, err := unix.Open(t.RootPath, flags, uint32(t.Permissions))
if err != nil {
return err
}
tmpPath := "/proc/self/fd/" + strconv.FormatUint(uint64(fd), 10)
n, err := unix.Write(fd, data)
if err == nil {
if n == len(data) {
err = unix.Linkat(unix.AT_FDCWD, tmpPath, unix.AT_FDCWD, p, unix.AT_SYMLINK_FOLLOW)
if errors.Is(err, unix.EEXIST) {
// https://github.com/nspcc-dev/neofs-node/issues/2563
err = nil
}
} else {
err = errors.New("incomplete write")
}
}
errClose := unix.Close(fd)
if err != nil {
return err // Close() error is ignored, we have a better one.
}
return errClose
}