Open rainersigwald opened 7 years ago
@nguerrera I think this might be a cleaner way to address @jaredpar's scenario from https://github.com/Microsoft/msbuild/issues/2095.
I'm not sure this would work for Roslyn: we don't use the CLI to build. The CLI won't be able to build Roslyn anytime in the foreseeable future due to the number of desktop specific MSBuild extensions our build contains: WPF, VS SDK, SWIX, etc ...
@rainersigwald This is exactly what I had in mind when I wrote:
(Now, that actually overlaps with another feature that's evolving and we will likely land in a place where you can edit global.json to get this behavior without setting any environment variables, but the mechanism here applies more generally to arbitrary resolvers with arbitrary input.)
There wasn't an issue for this yet (thanks for starting one), just mail threads so far, with a meeting scheduled soon to hash out the details. There's an existing mechanism called "multilevel lookup" in hostfxr that adds a well-known user profile location to the sdk search. This would be an evolution of that.
I did not expect Microsoft/msbuild#2095 to be controversial, but I understand your concerns now. I wasn't intending for this and that to be mutually exclusive. I was imagining that being able to pass arbitrary data down to arbitrary resolvers would be generally useful.
@jaredpar
I'm not sure this would work for Roslyn: we don't use the CLI to build.
The resolver called by desktop msbuild will respect global.json too. That's part of it's main purpose: to pick the same msbuild targets for VS that would be used by the CLI msbuild invoked on the same project.
Do you build with CLI on Mac/Linux? If your process installs .NET Core SDK (official name of entire CLI not just dotnet/sdk) to a local location and sets this feature up via global.json, then you can have one mechanism for pinning down the precise set of things that come from the .NET Core SDK across platforms, and not require it to be globally installed anywhere. On Windows, you'd only use the tasks and targets, but if somebody wanted to use dotnet
on the portable projects on Windows it would work. That could give you a simple way to reproduce issues that only happen on Mac or Linux official builds on a Windows dev box.
This could be equivalent to setting DOTNET_MSBUILD_SDK_RESOLVER_SDKS_DIR
global.json selects the full .NET Core SDK (CLI) used, not just the msbuild SDKs. We don't even have a resolver when we're invoked via dotnet
as it will have taken us to a universe where we have matching msbuild SDKs for the given version.
So global.json should instead specify equivalent of DOTNET_MSBUILD_SDK_RESOLVER_CLI_DIR (matching layout of C:\Program Files\dotnet). More granular control over the resolver than that has to go somewhere other than global.json.
Can I get more information on global.json: in particular where it lives, how it's found, etc ...?
Do you build with CLI on Mac/Linux? I
Yes and depending on the answer above it seems like a problem. The path to the downloaded SDK is going to be very different between the two systems.
{
"sdk": {
"version": "1.0.0",
"path": "tools/downloadedsdk"
}
}
In order for this to solve the problem the SDK essentially has to be within the cone of the global.json. Guessing that mean that it has to be within the cone of the repo. No other tool has that requirement for our build and it's going to take some doing to make that work.
Still overall feels clunky. MSBuild properties can today define everything about my build. Makes it dynamic, very easy to manage. Now we have JSON (again) and I have to write generators to translate between my MSBuild files and this new JSON file.
In order for this to solve the problem the SDK essentially has to be within the cone of the global.json. Guessing that mean that it has to be within the cone of the repo. No other tool has that requirement for our build and it's going to take some doing to make that work.
I think this will be easy to make work in our repos. The only thing that needs to be restored to a directory specified in global.json
is the immediate SDK used by the projects (which is not necessarily .NET SDK). In RepoToolset repos this is RepoToolset SDK (projects use <Project Sdk="RoslynTools.RepoToolset"/>
).
Our current global.json
for symreader repo looks like so:
{
"sdk": {
"version": "2.1.100-preview-007366"
},
"msbuild-sdks": {
"RoslynTools.RepoToolset": "1.0.0-beta2-62705-02"
}
}
Let's say we could add the path like so:
{
"sdk": {
"version": "2.1.100-preview-007366"
},
"msbuild-sdks": {
"RoslynTools.RepoToolset": { "version": "1.0.0-beta2-62705-02", "path": ".sdk" }
}
}
and this would restore the RepoToolset to {repo-root}\.sdk
directory (next to global.json
)
The RepoToolset restores .NET SDK by importing it in its sdk\Sdk.props
file:
<Import Project="Sdk.props" Sdk="Microsoft.NET.Sdk" />
at which point we can use arbitrary msbuild properties to specify the path to restore to.
@tmat that makes sense. I think that's workable.
This would be a very welcome addition to the rather inflexible global.json
config options. Currently, not being able to point to a local path, and the VS installer constantly uninstalling previous SDK versions, makes it pretty hard to do any decent SDK-dependent development (like the F# repo project files, see dotnet/fsharp#10193, which should be fixed on 3.1.302, but the VS installer removes this in favor of 3.1.402).
Since it is trivial to create an msbuild script that downloads a specific SDK version and installs it locally, it would be very, very nice if such behavior could be baked into global.json
as well:
"local": true
and standardize the path to be .dotnet
All in all, that would create a much better experience then the now rather hard-to-understand "cannot load project or solution" message, which baffles newcomers and experienced users alike...
This will have a significant positive impact on the developer experience for .NET repositories, such as https://github.com/dotnet/runtime/, https://github.com/dotnet/winforms, etc. Currently it is impossible to double click on a solution file to open the solutions in VS, because we build against nightly SDKs, which are located in custom locations and are forever changing.
As a workaround, I keep around a .bat file with this example content:
set PATH=C:\msbuild\.dotnet;%PATH%
set DOTNET_INSTALL_DIR=C:\msbuild\.dotnet
set DOTNET_MULTILEVEL_LOOKUP=0
If you go to the repo directory (C:\msbuild in my case) and run this .bat file, then run devenv.exe mysolution.sln, it will open with these environment variables set and the SDK resolution by MSBuild will use the .dotnet that's private to the repo.
That's what other dotnet repos are doing AFAIK in some shape or form. And this is not the best developer experience. It also makes it hard to run ad-hoc tests...
Oh, you don't have to convince me ;) I know it's terrible and I've spilled blood, sweat and tears to even arrive at this workaround by debugging deep into the SDK resolution process.
Lots of pain and friction caused by this over the years.
To include other points form offline chats and summarise:
There a number of use cases that really benefit from this (or comparable) functionality:
If it is difficult to provide the download functionality, we could probably do this in staggered approach, i.e. provide the ability to resolve an SDK from a custom location.
Maybe add another property to specify an error message that can describe how to bootstrap or where to download from. E.g.:
{
"sdk": {
"version": "2.1.100-preview-007366"
},
"msbuild-sdks": {
"RoslynTools.RepoToolset": {
"version": "1.0.0-beta2-62705-02",
"path": ".sdk",
"instructions": "Run .\restore.cmd to download the latest SDK.",
}
}
}
@vitek-karas We explored an option offline of just supporting this within the sdk resolver to find the SDK (essentially just DOTNET_MSBUILD_SDK_RESOLVER_CLI_DIR). Doing only that would unblock build but it would leave run (and potentially testing) still blocked as teams like WindowsDesktop have a test app that they would want to build and launch from VS against a version of the runtime listed in the global.json file.
Thoughts on finding the runtime from a path in global.json as well? I think the scenario for dotnet repos would have the runtime and SDK all in the same directory so potentially we only need one path still for both.
To add to @marcpopMSFT's post.
Starting VS without settings any paths or anything:
Starting VS with env:DOTNET_MSBUILD_SDK_RESOLVER_CLI_DIR
settings:
PS C:\Development\winforms> $env:DOTNET_MSBUILD_SDK_RESOLVER_CLI_DIR='C:\Development\winforms\.dotnet'
PS C:\Development\winforms> start .\Winforms.sln
On a second F6 it all build successfully... ¯\(ツ)\/¯
But I was unable to launch a project from the VS:
From runtime's perspective there's a risk of performance regressions. This would require the host to go looking for global.json
(or some other similar file) performing file existence checks. The deeper the folder hierarchy, the more expensive the check. This would have to be done by EVERY application, EVERY time. Given that this feature is expected to be used almost exclusively by developers that doesn't feel like the right tradeoff.
Doing this for SDK only is in theory possible, but could open issues with compatibility between SDK and the necessary runtime/frameworks for it (since we could not rely on the dotnet.exe
to be the right one anymore).
I absolutely understand the scenario and why it's important, but so far I was not able to figure out a solution which would not have a negative effect when the feature is not is use. (pay-for-play)
/cc @agocke
@RussKie Thoughts on only supporting this for the SDK as originally suggested given Vitek's feedback above? It wouldn't allow for launching test apps from the IDE but would allow for building. It's unclear if tests would work for not.
@agocke thoughts? Seems there's agreement that it would be high value to solve this as it will make us and the community more productive in our repos.
I (think) I like this. :smile: However, pinning the SDK in global.json is already a significant problem within Visual Studio, due to mismatches between the MSBuild engine in Visual Studio and the targets/tasks loaded from the .NET SDK. I have some worry that this might exasperate that problem more.
@KathleenDollard, do you have any thoughts?
Is it possible to start with an ENV solution only? That would mitigate Vitek's concerns. I'd like to reduce the number of places we have to search on disk not invent more.
Having repo-local SDKs make sense and I think is tractable. The team talked about repo-local runtimes and that one seems much more difficult. Not only would it probably be a big perf problem, as more intermediate tools use .NET Core (I hope), it's less and less likely that you want everything to run on local runtime. For instance, would you want new VS components that run on .NET Core to suddenly pick your repo local runtime?
I think SDK is a good place to start, if we can make that work.
If we had a repo root story, then putting a global.json at repo root would be workable and I suspect mitigate Vitek's concerns (depending on how repo-root worked; if it was an ENV, it would work awesome).
@RussKie Thoughts on only supporting this for the SDK as originally suggested given Vitek's feedback above? It wouldn't allow for launching test apps from the IDE but would allow for building. It's unclear if tests would work for not.
This wouldn't work for us. I can build from command line and don't need VS for that. I'd like a seamless devex for VS: double click sln, F5, debug.
RE: perf - we can make a call, and only look for .dotnet folder in the same folder as global.json. We have many things that are convention based, we can probably add yet another thing 😊
@RussKie You're confusing the SDK with the runtime. global.json
is an SDK configuration option. Right now there's no way to configure the muxer dotnet.exe
to look somewhere else for the runtime aside from environment variables. Looking for global.json
would run into the same problem -- anything you add to the muxer is overhead on every execution of a dotnet app. We've already removed a bunch of file system accesses from the host startup because they had unacceptable startup penalty. I don't consider adding any new filesystem checks to be an acceptable approach.
Then, why not do a context separation with global.json
and other env checks?
If the file is found along with other env checks, then, the muxer should be in Dev/SDK mode; else it'll be in Runtime/Execution mode. Do what you are doing today with the exec mode or simplify further if possible since every bit of perf is important. But in Dev/SDK mode, do things that improves the developer productivity such as loading dev-specific options (look for new SDK/Runtime paths, etc…) from both the env variables and the global.json
file. Almost every project, if it needs customizability, will have a global.json
in the root. So, we could reduce the checks to only look for global.json
to determine the context.
To add on further, you could have either --mode <dev|run|…>
or --dev
and --no-dev
to force the muxer into a specific context.
For instance, would you want new VS components that run on .NET Core to suddenly pick your repo local runtime?
This is a good question. I think the answer is "it depends", with likely "yes".
So, we could reduce the checks to only look for global.json to determine the context.
This is the problem - currently the muxer doesn't look for this file when running an app. So doing that would be a perf hit. Note that this is not as cheap as it seems. Basically every non-dev scenario (which is a vast majority of .NET app executions) will not have global.json
, but in order to figure that out the muxer would have to look at every directory from where the application is all the way up to the root of the file system. This means it's actually worse for non-dev - since in those cases there will be no global.json
and so the muxer will always go all the way up to the file system root (unlike dev scenarios, which would find global.json
sooner).
Currently the only reasonable solution I can think of is to require at least 1 environment variable. Something which tells the muxer to "work in the dev mode" or something similar. I don't think we can make the muxer do this work always.
Running VS components on repo-local runtime: I think the answer is "it depends", with likely "yes".
I don't agree. Note that this would apply to all kinds of things. For example today if I type code .
from the repo root I get a VS Code which works. If we were to force it to use the repo local runtime, it's possible that for example the C# language service would not work. Let's say the C# language service is currently running on .NET 5 and has the default roll-forward (Minor). This means that it would not find a compatible runtime in the repo-local runtime set. I don't know if the VS Code C# language service carries its own runtime, maybe it does, but any component which doesn't could run into these problems.
Also in the cases of our .NET repos, they frequently use previews or even nightly builds of the runtime as repo-local as they need a relatively fast update cycle to get new features. Running all my dev tools on top of such runtime can easily lead to breaks (we're not perfect, and especially nightly builds are from time to time a bit breaking).
Fortunately this is a moot discussion - even if we were to look for global.json
the muxer would use the location of the app to run as the starting point for the search. So things like VS, VS Code and so on would search somewhere in Program Files and not in the repo directory and would not find the file.
currently the muxer doesn't look for this file when running an app.
Sorry If I've got this all wrong. Doesn't it use different .NET SDK when we have a global.json
with a particular sdk
version mentioned? If it's not the muxer doing this, then what else?
Running all my dev tools on top of such runtime
Not the case here. As a runtime or a dev tools contributor, I'd want only certain invokes of dotnet
to be run in a developer context. Like testing the repo with a just built components of runtime (e.g.: the SDKs, MSBuild or NuGet targets) or in the case of core, winforms and wpf runtime, the entire runtime.
The muxer (dotnet.exe
) does two main things:
dotnet app.dll
, or dotnet exec app.dll
for example)dotnet build
or dotnet run
for example)When it detects (by looking at the command line basically) that it's to run CLI, it switches to the CLI mode, in which case it will go look for global.json
. But if it's running in the "app" mode, it will not do that.
The discussion is about modifying the "app" mode to go search for global.json
.
Do we require modifying the exec mode to look for global.json
? Isn't that what the runtimeconfig.json
is for?
What I was asking and hoping to achieve here is to extend global.json
to select more than just SDK version and have a list of MSBuild SDK packages, to customize lookup/restore options for the MSBuild SDKs, and of-course the .NET SDK itself.
The discussion is about modifying the "app" mode to go search for
global.json
.
This is not exactly what I had in mind. I look at it from a layman point of view, and I was asking what can be done to make the developer experience contributing to .NET repos to the same level as building and debugging other (normal) .NET solutions, i.e. "double click on a sln file, have VS resolve the required/expected version of SDK/runtime, F5". By "VS" (in layman terms) I mean everything involved to build and debug an .NET app, the editor, the compiler, analyzers, MSBuild, etc.
I understand the high-level scenario, but in my mind it basically boils down to some mechanism through which running a .NET app/command from the repo directory tree can be made to use repo-local SDK/runtime. Ultimately, if I build a test app from within the repo, and run it - it should get the repo-local runtime. Running tests, or using VS to run them from me, is just an extension of that I think.
If we were to scope this to only VS (for example), then there are more options, as we can (in theory) make VS understand repo-local things and run the tests/apps with that knowledge. But I think everything should work directly from command line as well. And the "F5" from command line is basically /myrepo/artifacts/tests/mytest.exe
- and we want to make this use repo-local runtime.
(Note that app.exe
is called apphost
and it's almost the same code as dotnet.exe
(called the "muxer") - so the discussion above about the "muxer" applies to the "apphost" as well).
Do we require modifying the exec mode to look for global.json? Isn't that what the runtimeconfig.json is for?
We could encode something into the runtimeconfig.json
, but that would mean that all of the runtimeconfig.json
files in the repo, would need to have this "special" thing in them. Which is problematic for official builds - those would have to turn it off (shipping apps would not want that), but then again testing the official build could not be done from within the repo... and so on.
Potentially we could encode something like this into the runtimeconfig.dev.json
- but there's still the problem of official build (which would need to know to exclude that file).
I don't think we can scope a solution to only VS and per an earlier comment, we cannot fix this with environment variables either (as we already have that solution in place and this feedback comes even with env variable options being in use). We stopped generating the .dev.json file by default. How would we make it generate it for local builds but not CI builds?
I'm sorry I didn't make my previous comment clear. I don't think it's a good solution to encode something into the .runtimeconfig.json
(or the dev
version of that file). Both solutions have the problem that they require us to modify produced binaries, which is very problematic. It's also very possible such solution would not just work for testing either (since tests generate their own .runtimeconfig.json
) - would need even more "weird" changes.
Just thinking out loud (it could be a terrible idea): New environment variable which would tell the hosts to look for global.json
(or similar file) even for app invocation. We would require developers in affected repos to set this variable (probably globally on their machine) once. It would imply the perf impact of the host on pretty much all apps on the machine, but it would solve the problem of only affecting users which need this functionality. (There are potential improvements on this idea as well obviously).
We would require developers in affected repos to set this variable (probably globally on their machine) once.
I don't think this will work either, e.g. I work on multiple dotnet/* repos, with at least two - actively. Resetting an env variable every time I need to launch a repo in VS would be a nightmare.
Given that the issue has been opened since 2017, and we've determined that we can't make any practical changes to accommodate the seamless VS developer experience for dotnet/* repos, should we close this?
I am coming back to this issue after writing up a breaking change on multi-level-lookup. See: https://github.com/dotnet/core/issues/7131. I'm thinking that it will help this scenario significantly.
For the CLI experience, it's easy to use "C:\PrivateSDK\dotnet build" and end up using a global SDK when you want to use a private one. Super frustrating. Resolving that is the point of the break.
Going forward, I'd offer the following as best practice guidance for the private SDK scenario:
DOTNET_ROOT=C:\PrivateSDK
You could then type the following:
PS C:\Users\rich\fakerepo> C:\PrivateSDK\dotnet build
PS C:\Users\rich\fakerepo> .\bin\Debug\net7.0\fakeapp
App launched on .NET 7 from C:\PrivateSDK
If you wanted to avoid typing "C:\PrivateSDK" all the time, you could add the location to the path, with the following:
set PATH=C:\PrivateSDK;%PATH%
This scenario works super well if you can set these ENVs via a script or manually. It doesn't work at all if you cannot.
The original idea (I think) was to effectively expose the analog of DOTNET_ROOT
in global.json. I think that's an interesting direction IFF we have a new scenario for Visual Studio that it needs to respect that (which includes testing). Alternatively, we could create a convention where Visual Studio looks for a .dotnet
directory at repo root to achieve the same end. I actually prefer that. Again, I think that's a Visual Studio scenario, not a .NET one. Ideally, there would be a Visual Studio way to say "look in this repo-local directory for the dev platform" and it would work equally well for node.js (for example).
Am I on the right page (or even the right book)?
The original issue is mainly around VS (at least that's my understanding). Specifically: clone a repo, maybe run build
from root for the first time, double-click the .sln
file in the root -> everything works.
That said, it would be really nice if we came up with a solution which works from CLI as well. For SDK, using global.json
makes sense. There might be some technical challenges how to implement it, but I don't see any big architectural problems. But for runtime, so far we haven't come up with anything good. The obvious requirement is, that whatever we do here is 100% opt-in.
My take was that this was about getting VS to use a different toolset than the one VS ships (.NET SDK or MSBuild/Roslyn/NuGet that comes with VS). I think that's what you are saying, too. I think this scenario needs to come from the VS team, not the SDK team. It's primarily challenges/constraints in VS that precent this scenario from working. If they want to prioritize the scenario, we can certainly do our part.
It's primarily challenges/constraints in VS that precent this scenario from working.
I don't think that's true? If the .NET SDK resolver could find one based on a path in the global.json
everything would work more or less fine, right? Just like today VS can load different SDKs specified by version, or via env var trickery.
I was referring to the core toolset (MSBuild, Roslyn, Nuget, Common T&Ts). Related to this conversation: https://github.com/dotnet/sdk/issues/8254#issuecomment-301249426.
I've heard time and again that global.json (independent of this topic) is not transparent to VS, that switching to an older or newer (or just different, given the strangeness of SDK bands) can break VS. We need to resolve that topic, right?
Zooming out, can we enable such a model transparently from VS, such that it is none-the-wiser?
Perhaps we're talking past one another. Happy to participate in a call on this to help us focus on the core issues.
The customer feedback I've heard internally is specifically about wanting to clone a repo and double click the solution file in file explorer. That doesn't work today as you have to set environment variables first which we've been told is a barrier to entry for some customers. Hence the ask to put it in global.json so a repo owner can do it and the end developer doesn't have to know though that only works if they actually have the needed sdk installed...
If I have an SDK in hand why do I need to go through the trouble of setting an environment variable to use it? That means that it's a command line tool I can't reliably run unless I muck with some environment variables.
I still strongly think that I should be able to use dotnet build
to use the SDK that came with the exe to build my application. Maybe I need to provide a command line switch like --this-one
but the solution should be in the CLI itself. It shouldn't require any weird pathing.
Every other language framework in existing except .NET provides this solution.
I'm currently OOF and writing from a phone.
My main motivation for this change is outlined in https://github.com/dotnet/sdk/issues/8254#issuecomment-860995649.
Just adding my vote to this, particularly https://github.com/dotnet/sdk/issues/8254#issuecomment-1023551867. If I'm understanding correctly, this would really help in the https://github.com/dotnet/razor-tooling/ repo, which currently requires setting a bunch of envars set. Having to ensure these are set before running VS is a pain, particularly if you're trying to use multiple versions of VS to verify things (switching between developer command prompts etc.), or if you're having to use multiple repos with similar set ups (as @RussKie mentioned too).
@vitek-karas and I are working on a spec that possibly includes what you want. Will share (very) soon. In any case, it will be a good starting point.
The original idea (I think) was to effectively expose the analog of DOTNET_ROOT in global.json. I think that's an interesting direction IFF we have a new scenario for Visual Studio that it needs to respect that (which includes testing). Alternatively, we could create a convention where Visual Studio looks for a .dotnet directory at repo root to achieve the same end. I actually prefer that. Again, I think that's a Visual Studio scenario, not a .NET one.
@richlander IMO it should be in global.json. Irregardless, please make sure to loop in the Visual Studio for Mac team for any scenario.
We backed a way a little from this. We decided NOT to put any new location content into global.json
. It is too easy to trick developers into running malicious code. Instead, we're starting with a new CLI argument. VS* products can add support for this flag if they like. It won't happen automagically due to a .NET SDK feature.
@richlander can you provide some more detail on the malicious scenario here? I don't understand how this is riskier than restoring nuget packages.
Adding this info to global.json
seems like the best way forward for painless IDE support of custom SDK locations on a per-solution basis.
Builders of large systems often want to be able to explicitly specify an SDK version and avoid installing the CLI to a machine-global location.
Currently,
global.json
allows the version lock, but the specified version must be installed in either a machine-global location or a nonstandard location must be specified by the environment variableDOTNET_MSBUILD_SDK_RESOLVER_SDKS_DIR
.That environment means that you must launch Visual Studio from a specific environment to get the downloaded/private SDKs.
There could be an extension to the SDK resolver to respect a path specified in the
global.json
. Something likecould be equivalent to setting
DOTNET_MSBUILD_SDK_RESOLVER_SDKS_DIR=%GlobalJsonPath%\tools\downloadedsdk
before invoking the resolver.This would be visible by any invocation (
dotnet
,msbuild
, VS, or MSBuild API) since it's file-based.