CreateDirectory: eliminate some syscalls.

[tmds]

This eliminates three syscalls per Directory.CreateDirectory when the path doesn't exist.

One is the initial check if the directory exist.

The other two are eliminating by using mkdir to check parent directory existence.
Instead of finding the first parent that exist using stat, we now end up creating the first parent that doesn't exist.

When Directory.CreateDirectory is called with a path that does exist, we are now making two syscalls instead of one.
If we want, we can avoid that regression by keeping the initial existence check. That costs us a syscall in the non-exists case.
A user can also avoid it by calling Directory.Exists before calling Directory.CreateDirectory.

Similar changes can be made to the Windows implementation.

@adamsitnik@stephentoub ptal.

[ghost]

Tagging subscribers to this area: @dotnet/area-system-io
See info in area-owners.md if you want to be subscribed.

Issue Details

---

This eliminates three syscalls per Directory.CreateDirectory when the path doesn't exist.

One is the initial check if the directory exist.

The other two are eliminating by using mkdir to check parent directory existence.
Instead of finding the first parent that exist using stat, we now end up creating the first parent that doesn't exist.

When Directory.CreateDirectory is called with a path that does exist, we are now making two syscalls instead of one.
If we want, we can avoid that regression by keeping the initial existence check. That costs us a syscall in the non-exists case.
A user can also avoid it by calling Directory.Exists before calling Directory.CreateDirectory.

Similar changes can be made to the Windows implementation.

@adamsitnik@stephentoub ptal.

Author:	tmds
Assignees:	-
Labels:	area-System.IO
Milestone:	-

[jeffhandley]

Looks good to me, but I'd like a secondary review from @dotnet/area-system-io.

[adamsitnik]

Overall LGTM, but it would be great to refactor the code a little bit and if it's safe replace List<string> with List<int>

@tmds thank you for another great contribution!

[tmds]

@adamsitnik thanks for the review! I've addressed your comments.
I kept the List<string> since I need a string to pass to the MkDir PInvoke.

[adamsitnik]

LGTM, again thank you @tmds !

[carlossanlop]

@tmds@adamsitnik just for completeness, can we get perf numbers before and after this change?

[tmds]

can we get perf numbers before and after this change?

Is there some place we can see the diff on the continuous benchmarking of the dotnet/performance repo?

[adamsitnik]

Is there some place we can see the diff on the continuous benchmarking of the dotnet/performance repo?

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory/TestHistoryIndexIndex.html

But the existing Directory APIs benchmarks are very noisy:

https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04%2fSystem.IO.Tests.Perf_Directory.CreateDirectory.html
https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04%2fSystem.IO.Tests.Perf_Directory.RecursiveCreateDeleteDirectory(depth%3a%2010).html
https://pvscmdupload.blob.core.windows.net/reports/allTestHistory%2frefs%2fheads%2fmain_x64_ubuntu%2018.04%2fSystem.IO.Tests.Perf_Directory.RecursiveCreateDeleteDirectory(depth%3a%20100).html

[tmds]

The change wasn't run by the benchmarks yet. Let's see if we can see it in a couple of days.

@adamsitnik what generates this website? Is it part of BenchmarkDotNet?

[adamsitnik]

The change wasn't run by the benchmarks yet

It's strange, our automation is supposed to run at least few times a day.

what generates this website? Is it part of BenchmarkDotNet?

it's closed source tool maintained by @DrewScoggins

[DrewScoggins]

Long story short the data on the test history index page only updates about once a day, because it takes about a day to generate. I am looking to make some changes to greatly reduce the time it takes to generate them, but for now that is where we are. As a result they can sometimes be a little behind.

I went ahead and generated a report for just the Directory tests, and though one is marked as a regression, it looks like just noise in the test. I don't see any significant regression or improvement as the result of this change. Report

[tmds]

By now, the benchmark ran a few times with the changes.

[danmoseley]

do we need an up for grabs issue to do the same for Windows?

[tmds]

do we need an up for grabs issue to do the same for Windows?

I've created #61954.

[danmoseley]

thanks.