Avoid fork on mount for overlay2 in common case

[dmcgowan]

In the common case where the user is using /var/lib/docker and
an image with less than 60 layers, forking is not needed. Calculate
whether absolute paths can be used and avoid forking to mount in
those cases.

I do not have any statistics of what percentage of images
are less than 60 layers deep, but my guess is a vast majority and
even larger majority when accounting for images which are run most often.

Mount Benchmarks

Overlay2 (pre-change) vs Overlay2 (updated)

benchmark                         old ns/op     new ns/op     delta
BenchmarkGetSingleBaseMount-8     7744409       231444        -97.01%
BenchmarkGet20BaseMount-8         8324308       307961        -96.30%
BenchmarkGet50BaseMount-8         8482378       420499        -95.04%
BenchmarkGet100BaseMount-8        8252170       7976995       -3.33%

Overlay vs Overlay2 (updated)

benchmark                         old ns/op     new ns/op     delta
BenchmarkGetSingleBaseMount-8     339539        231444        -31.84%
BenchmarkGet20BaseMount-8         249545        307961        +23.41%
BenchmarkGet50BaseMount-8         200676        420499        +109.54%
BenchmarkGet100BaseMount-8        252296        7976995       +3061.76%

AUFS vs Overlay2 (updated)

benchmark                         old ns/op     new ns/op     delta
BenchmarkGetSingleBaseMount-8     305157        231444        -24.16%
BenchmarkGet20BaseMount-8         1830297       307961        -83.17%
BenchmarkGet50BaseMount-8         10263958      420499        -95.90%
BenchmarkGet100BaseMount-8        36276899      7976995       -78.01%

AUFS vs Overlay2 (pre-change)

benchmark                         old ns/op     new ns/op     delta
BenchmarkGetSingleBaseMount-8     305157        7744409       +2437.84%
BenchmarkGet20BaseMount-8         1830297       8324308       +354.81%
BenchmarkGet50BaseMount-8         10263958      8482378       -17.36%
BenchmarkGet100BaseMount-8        36276899      8252170       -77.25%

Note overlay will always mount faster than overlay2 for a larger number of lower directories since it can only have one lower directory (which it achieves at the expense of duplicating files). These are the first benchmarks I have done with aufs vs overlay2 so I included the benchmark comparison with the pre-changed overlay2.

Benchmarks done with https://github.com/dmcgowan/dsdbench

[tonistiigi]

Would it make sense to make this calculation more precise? I think even calculating the whole opts even if it is not used would not be noticeably slow(good thing you have benchmarks to prove me wrong). If not then please add a comment that 512 is a buffer for the contants and label, I was initially confused by that.

[dmcgowan]

I had considered just calculating the opts using absolute paths optimistically, I think this might make even more sense considering I found the threshold to be over 60 layers.

[dmcgowan]

Updated to optimistically create the mount data with absolute paths and only use relative paths when the mount data is larger than a page. Tested to make sure that the extra buffer is not needed when the mount data is fully computed (see dmcgowan/dsdbench@071a526#diff-c16a9df1a4a40f7564a79598949a5199R136).

[AkihiroSuda]

I think it will be much more readable if this branch is moved to L430. i.e.,

mountData = label.FormatMountLabel(opts, mountLabel)
if len(mountData) > pageSize {
...
}

[dmcgowan]

Agree, updated

[tonistiigi]

LGTM

[LK4D4]

LGTM