Home: Too many open files when restoring on Mac

[moozzyk]

Looks like the fix for https://github.com/NuGet/Home/issues/2004 was not enough. I am still hitting the same issue on some repos.

I am using dotnet-osx-x64.1.0.0.001496.tar.gz which contains the fix for #2004

To repro:

• clone https://github.com/aspnet/Security
• remove all cached packages
• run ./build.sh

Log from the CI:
https://s3.amazonaws.com/archive.travis-ci.org/jobs/111327930/log.txt

[moozzyk]

You can use opensnoop to track open files or lsof to take a snapshot of open files.

[zhili1208]

@moozzyk just verified with latest bit, looks like it's fixed by @emgarten 's fix, feel free to reopen it if it still repro

[moozzyk]

I hope the new dotnet cli we will pick will have this fix so I will be able to verify this end to end.

[zhili1208]

@moozzyk please try the new dotnet, I think it has this already

[agocke]

I just hit this with 1.0.0.002060

[zhili1208]

@agocke we just found one issue related to this, will fix it soon, Can you provide repro solution? then we can verify it with the fix.

[agocke]

@zhili1208 It's the Roslyn solution -- I'm in the process of making a PR for this. If you get a build, lemme know and I can try it. Otherwise I'll set up a repro for you.

[agocke]

I can confirm this is still happening:

Unhandled Exception: System.IO.IOException: Too many open files
   at Interop.ThrowExceptionForIoErrno(ErrorInfo errorInfo, String path, Boolean isDirectory, Func`2 errorRewriter)
   at Interop.CheckIo[TSafeHandle](TSafeHandle handle, String path, Boolean isDirectory, Func`2 errorRewriter)
   at System.ConsolePal.OpenStandardOutput()
   at System.Console.<>c.<get_Out>b__25_0()
   at System.Console.EnsureInitialized[T](T& field, Func`1 initializer)
   at System.Console.WriteLine(String value)
   at Microsoft.DotNet.Tools.Restore.RestoreCommand.<>c__DisplayClass1_0.<Run>b__2()
   at Microsoft.Dnx.Runtime.Common.CommandLine.CommandLineApplication.Execute(String[] args)
   at Microsoft.DotNet.Cli.Program.ProcessArgs(String[] args, ITelemetry telemetryClient)
   at Microsoft.DotNet.Cli.Program.Main(String[] args)
Restore failed

$ ./dotnet --version
1.0.0-rc2-002378

To repro

1. clone Roslyn
2. remove the ulimit setting in build/scripts/restore.sh
3. run make restore
4. Replace the binaries in Binaries/toolset/roslyn.mac.5/dotnet-cli with the latest mac dotnet build.
5. Delete Binaries/toolset/restore.semaphore and ~/.nuget
6. Re-run make restore

[zhili1208]

@agocke yes, this is still happening, the issue comes from httpclient, dotnet core team will provide us an API to limit the open file number, but not now, so we are still waiting for them. If this blocks you, please use ulimit for now

[jaredpar]

Why is this issue still in the closed state? It's 100% reproducable in Roslyn and our build is completely broken unless we resort to ulimit hackiness to unblock us.

the issue comes from httpclient, dotnet core team will provide us an API to limit the open file number, but not now, so we are still waiting for them

That's fine but still seems like there is work to do on your end. Shouldn't this bug be tracking the work?

[zhili1208]

https://github.com/dotnet/cli/issues/809

[joelverhagen]

Did some digging, related: https://github.com/dotnet/corefx/issues/7932

Still investigating.

[agocke]

Note that we've also started seeing this on Linux: https://github.com/dotnet/roslyn/pull/10768

[joelverhagen]

@agocke, we are going to pull in a fix for dotnet/corefx#7932 today, which addresses this issue. After that point, this problem should go away.

[agocke]

@joelverhagen Thanks! Can you try the Roslyn repro I posted above or let me know when a build is available so I can try it myself?

[joelverhagen]

The latest dotnet CLI has been fixed to address this problem. Basically you should look for a version of System.Net.Http to have a build number of 24022 or greater. This was bumped in CLI with https://github.com/dotnet/cli/commit/e420515c2bde047c568cc86156b340bdf06259e5 and in NuGet with https://github.com/NuGet/NuGet.Client/commit/b3fecd48a02bfbd9c7c4536c9435259b8c9b2220.

I just tried restoring Roslyn latest with dotnet CLI latest on OS X and had no issue. My dotnet CLI version is 1.0.0-rc2-002487.

[starquake]

Just hit on this while doing

dotnet new --install "Microsoft.AspNetCore.SpaTemplates::*"

executing ulimit -n 512 fixes it

The version of dotnet I'm Using is 1.0.0-rc4-004771. Am I correct in assuming this will be fixed in 1.0.0-rtm?

[zhili1208]

@starquake I can't repro this with dotnet 1.0.0-rc4-004823, Could you try latest dotnet cli and also share your nuget.config file from ~/.nuget/NuGet ? Sometimes this is related to the number of package source.

[starquake]

@zhili1208 I will share the nuget config when I'm at work tomorrow. Tried it on another system and it executed without any problems.

If you want to close this ticket because it's not reproducible that would be fine too.

[rrelyea]

Moving out until we have a dependable repro to consider.

[starquake]

Is there a dotnet cli that is newer than 1.0.0-rc4? Where can I get it?

This is in my nuget.config:

<?xml version="1.0" encoding="utf-8"?>
<configuration>
  <packageSources>
    <add key="nuget.org" value="https://api.nuget.org/v3/index.json" protocolVersion="3" />
  </packageSources>
</configuration>%

[mrward]

@starquake - More recent rc4 dotnet cli builds are available from https://github.com/dotnet/cli

[starquake]

Installed 1.0.0-rc4-004834 and it still occurs.

I did some troubleshooting.

ulimit -n tells me 256 is the default
If I set it to 266 the error appears later.
If I set it to 276 the error appears even later.
If I set it 286 there is no error anymore.

So it looks like I'm just over the boundary.

When I look at my open files for the shell with lsof, it's not that much:

➜  ~ lsof -p 4496 | wc -l
      19

[joelverhagen]

Could you run lsof to determine what kind of handles are most prevalent? We have seen issues in the past where HttpClient was not properly closing sockets. We also open a lot of file handles so perhaps there is an issue there.

[starquake]

It's quite a bit. See here:
https://gist.github.com/starquake/b29f18e41d9e0ae6490e998ce14ede57

[am11]

Same situation with SDK v1.0.0 and v1.0.1 .. https://travis-ci.org/am11/MaxMind-DB-Reader-dotnet/jobs/209511300#L974-L996

Based on this comment: https://github.com/dotnet/cli/issues/809#issuecomment-194407913, would it be possible to limit the max connections (or MaxDegreeOfParallelism) to eight?

[Meligy]

I'm having this (as mentioned in https://github.com/dotnet/cli/issues/6014#issuecomment-286060904), even though when I run ulimit, I get unlimited. Is that expected? (as in, is unlimited actually a bad value?)

Thanks.

[am11]

@Meligy, ulimit -n shows the actual value. For TravisCI, we are setting it like this.

[Meligy]

Aha. Makes total sense (for the record mine shows 256 for ulimit -n, but probably everybody here knows that already).

Thanks a lot.

[rrelyea]

@emgarten - can you make sure we get the right "ClosedAs" label set here? won't fix? by design?

[zhili1208]

@rrelyea I remember there was a fix in template, they opened many nupkg files in memory but didn't close them. I think this issue is fixed by their fix.

[rrelyea]

From @mlorbetske:
that issue was fixed as of the 1.0.4 CLI. Either updating to a later CLI build or raising the value of ulimit should fix it.

will mark as external and notrepro - given that. if anybody still hits despite that, please chime in.