src: add public API to create isolate and context #20639

helloshuangzi · 2018-05-09T18:45:49Z

This change is used to support addon / embedding scenarios.
An addon/embedding scenario has requirement to create isolate / context with the same setting/tweak that is used in node main thread, instead of duplicating that part of logic by themself, which may lead to conflict in the future.

Please see this commit of napajs as an example.

Checklist

make -j4 test (UNIX), or vcbuild test (Windows) passes
commit message follows commit guidelines

addaleax · 2018-05-09T18:49:23Z

/cc @gireeshpunathil @cmfcmf

jasnell · 2018-05-09T22:39:16Z

Ping @nodejs/n-api ... would it make sense to have n-api support for this?

devsnek · 2018-05-09T22:47:55Z

i'm not quite sure how we would be able to represent contexts in napi but something like napi_create_env would probably work for isolates

addaleax · 2018-05-09T22:57:34Z

@jasnell I think it’s going to make sense to work out and eventually provide an N-API-style embedder API for Node.js, yes. But this makes sense regardless, and we’d probably put more thoughts into a whole new API that we can design from scratch.

gabrielschulhof · 2018-05-10T02:35:30Z

I was thinking something as simple as

napi_env napi_create_env();

... and then you can use napi_run_script() to launch something which, in turn, can require() native modules, for which the Node.js instance will use the special symbol if the modules don't self-register.

Now, if only we could figure out how not to leak the dlopen handles.

gabrielschulhof · 2018-05-10T02:36:22Z

... or

napi_status napi_create_env(napi_env *result);

in keeping with our existing API look and feel.

bnoordhuis

Besides the bug, I also have some misgivings about haphazardly exposing yet another piece of internal plumbing.

It doesn't result in a coherent public API but meanwhile makes maintenance harder.

bnoordhuis · 2018-05-10T06:31:52Z

src/node.cc

@@ -279,6 +279,8 @@ static v8::Isolate* node_isolate;

 DebugOptions debug_options;

+static ArrayBufferAllocator array_buffer_allocator;


This can't be static, it's tied to the isolate. That will be problematic because of the zeroFill field - that field is flipped on and off at run-time and that's observable (read: behaves erratically) between isolates.

@bnoordhuis zeroFill is tied with this condition if (auto zero_fill_field = env->isolate_data()->zero_fill_field()). Currently, the API of CreateIsolateData doesn't support parameter of zero_fill_field for embedding/addon users. It means zeroFill is only tied to the node main thread/environment. This is my understanding at first, and I thought it was not a bad design of node public API.
I can change it to let each isolate have its own ArrayBufferAllocator instance if you think the above understanding of the current node APIs doesn't work or just not right.

@helloshuangzi In that case, it might make sense to expose a way to create a Node.js-style ArrayBufferAllocator given an IsolateData* pointer?

@helloshuangzi here is why it's not safe:

node/lib/buffer.js

Lines 115 to 122 in 46d335c

function createUnsafeArrayBuffer(size) {

zeroFill[0] = 0;

try {

return new ArrayBuffer(size);

} finally {

zeroFill[0] = 1;

}

}

That's the 'flipped on and off at run-time' I mentioned that is observable from another isolate.

@bnoordhuis , Yes, but 'zeroFill' is only available in node main isolate/environment because of if (auto zero_fill_field = env->isolate_data()->zero_fill_field()) based on current node API.
I am thinking that it might be not a bad idea to make all isolates comply with node main isolate/environment about zero_fill_field。

yes, that's what I said all isolates will comply with node main isolate about setting of zero_fill_field, while only node main isolate can change it.

yes, that's what I said all isolates will comply with node main isolate about setting of zero_fill_field, while only node main isolate can change it.

I don’t think that’s a good idea. I agree with Ben, it’s not a good idea for one Isolate to be able to change the allocation mode of other Isolates.

To reinforce the point: what you'd see in other isolates is new arraybuffers randomly being filled with garbage instead of zeroes like they're supposed to.

Okay, @addaleax and @bnoordhuis, thanks for the feedback about this point. Let me change it to give an ArrayBufferAllocator instance for each isolate.
@addaleax, as you commented above, we will need:

an API to create node.js-style ArrayBufferAllocator, and get its zero_fill_filed.

a new API of CreateIsolateData which can take the parameter of zero_fill_field.

fixed, please help review!

gabrielschulhof · 2018-05-10T14:31:08Z

@devsnek the napi_env is stored in a private field on the global object. We retrieve the global object from the context, which we receive during addon initialization from Node.js. Therefore napi_env is per-context.

mhdawson · 2018-05-10T18:21:08Z

I share @bnoordhuis's concern for creating new public API without the complete picture. I also understand it may not be helpful to get in the way when it may take a while to come up with a complete API.

@helloshuangzi I'm wondering how urgent this is from your perspective? Is it significantly blocking your forward progress or something that is a nice to have in the short term?

helloshuangzi · 2018-05-10T19:32:39Z

@mhdawson and @bnoordhuis , thanks for your feedback.
About the coherence of node public API, given APIs like CreateIsolateData and CreateEnvironment are already there, I think NewIsolate and NewContext can be seen as a set of helpers for ease of use to complement those mentioned ones. It should not conflict with the future work to create a complete API.

@mhdawson, thanks for your asking about the urgency. We are working on re-implementing napajs, a node addon for js mutli-threading based on node public APIs (before we do most work by ourselves). Unfortunately, it is a blocker for this work.

addaleax · 2018-05-10T19:49:17Z

Unfortunately, it is a blocker for this work.

@helloshuangzi Can you explain in what way it is a blocker? Generally, one should be able to get by using only the V8 APIs for creating Isolates + Contexts, so that seems like helpful information. (Like you said, the APIs introduced here would be “helpers”, and I am not sure in what way they are necessary themselves, tbh.)

devsnek · 2018-05-10T19:50:46Z

@addaleax isolates that aren't tracked by node's platform tend to screwed over pretty badly through the lifetime. i think there's another pr to expose node's platform because of that.

addaleax · 2018-05-10T19:57:27Z

@devsnek That’s not really related to this PR – CreateIsolateData() is the things which registers an isolate with the platform. The NewIsolate() API here does no such thing.

helloshuangzi · 2018-05-10T20:49:22Z

@addaleax, I say it's a blocker because we are making every effort to maximize the consistency between napa worker and node default isolate/environment (or say it a node standard way to create isolate/context). The consistency is important to us to give napa users a smooth experience. That's why we are trying to implement napa based on node public API. Without the proposed APIs, addon/embedding users will always need to catch up with node changes upon this part.

addaleax · 2018-05-11T07:06:09Z

src/node.cc

@@ -4425,15 +4425,22 @@ int EmitExit(Environment* env) {
 }


-IsolateData* CreateIsolateData(Isolate* isolate, uv_loop_t* loop) {
-  return new IsolateData(isolate, loop, nullptr);
+ArrayBufferAllocatorBase* CreateArrayBufferAllocator() {


I think we could also use a std::unique_ptr, and/or drop FreeArrayBufferAllocator because that would be implicit in using std::unique_ptr.

I just followed the pattern of the existing APIs like FreeIsolateData, FreeEnvironment and FreePlatform. I think they were made in the current pattern to avoid trouble of unique_ptr for ABI.
Do you think I can keep this pattern for ArrayBufferAllocator?

addaleax · 2018-05-11T07:10:52Z

src/node.h

 NODE_EXTERN IsolateData* CreateIsolateData(
    v8::Isolate* isolate,
    struct uv_loop_s* loop,
-    MultiIsolatePlatform* platform);
+    MultiIsolatePlatform* platform = nullptr,
+    uint32_t* zero_fill_field = nullptr);


Just fyi, removal of the first variant and adding the field here would be an ABI breakage and therefore semver-major (adding the label for now). I think you might first want to go the route of adding Yet Another Overload, maybe leaving a TODO() comment for switching to default parameters in a follow-up PR.

thanks for the kind reminder. Just added a new overload to avoid semver-major for this PR, and added TODO().

addaleax · 2018-05-11T07:14:30Z

src/node.h

+
+ protected:
+  uint32_t zero_fill_field_ = 1;  // Boolean but exposed as uint32 to JS land.
+};


Is the intention that we expose zero_fill_field() for public use? I don’t think so. It might be easier/expose less internal API surface to just forward-declare ArrayBufferAllocator and add an ArrayBufferAllocator* parameter to CreateIsolateData()?

You are right, no public use of zero_fill_field. Just updated it as your suggestion.

addaleax · 2018-05-11T07:18:16Z

Oh, and: Please avoid merge commits. :) It would be great if you could rebase them out of this PR; it you’d prefer not to, the person landing this PR would have to do that, which can get a bit messy if there are merge conflicts along the way.

bnoordhuis · 2018-05-11T18:49:37Z

src/node.cc

@@ -4425,10 +4425,21 @@ int EmitExit(Environment* env) {
 }


+ArrayBufferAllocator* CreateArrayBufferAllocator() {
+  return new ArrayBufferAllocator;


new ArrayBufferAllocator() - i.e., include parentheses.

bnoordhuis · 2018-05-11T18:52:02Z

src/node.cc

+inline int Start(uv_loop_t* event_loop,
+                 int argc, const char* const* argv,
+                 int exec_argc, const char* const* exec_argv) {
+  ArrayBufferAllocator* allocator = CreateArrayBufferAllocator();


Use a std::unique_ptr with a custom deleter, or simply use a stack-allocated ArrayBufferAllocator.

I just followed the pattern of the existing APIs like FreeIsolateData, FreeEnvironment and FreePlatform. I think they were made in the current pattern to avoid trouble of unique_ptr for ABI.
Do you think I can keep this pattern for ArrayBufferAllocator?

Updated with unique_ptr, and I don't have to customize deleter for usage in node.cc because it can get the full definition here. As to the public API/ABI, do you think I can keep it with raw pointer by following the existing pattern?

Yes, raw pointers are the way to go, that avoids trouble when Node.js and the add-on use different allocators.

bnoordhuis · 2018-05-11T18:52:34Z

src/node.cc

@@ -4611,15 +4639,16 @@ inline int Start(uv_loop_t* event_loop,
    Locker locker(isolate);
    Isolate::Scope isolate_scope(isolate);
    HandleScope handle_scope(isolate);
-    IsolateData isolate_data(
+    IsolateData* isolate_data = CreateIsolateData(


bnoordhuis

LGTM with a request.

I want it understood that there's no promise we'll carry these APIs forever. If they become a maintenance hassle, we'll simply remove them again in the next major release.

bnoordhuis · 2018-05-12T12:06:28Z

src/node.cc

+inline int Start(uv_loop_t* event_loop,
+                 int argc, const char* const* argv,
+                 int exec_argc, const char* const* exec_argv) {
+  std::unique_ptr<ArrayBufferAllocator> allocator(CreateArrayBufferAllocator());


I prefer simply ArrayBufferAllocator allocator; if you're not using a custom deleter, no point in heap-allocating it. Likewise for IsolateData.

Thanks for the detail description. I just want to reuse the APIs to help keep consistency between node-main-thread and addons when the relevant logic gets updated in the future.
Updated with custom deleter for both ArrayBufferAllocator and IsolateData.

PR-URL: #20639 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: James M Snell <[email protected]> Reviewed-By: Gireesh Punathil <[email protected]> Reviewed-By: Khaidi Chu <[email protected]> Reviewed-By: Ben Noordhuis <[email protected]>

PR-URL: nodejs#20639 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: James M Snell <[email protected]> Reviewed-By: Gireesh Punathil <[email protected]> Reviewed-By: Khaidi Chu <[email protected]> Reviewed-By: Ben Noordhuis <[email protected]>

Notable Changes: * **addons**: - Fixed a memory leak for users of `AsyncResource` and N-API. (Michael Dawson) [#20668](#20668) * **assert**: - The `error` parameter of `assert.throws()` can be an object containing regular expressions now. (Ruben Bridgewater) [#20485](#20485) * **crypto**: - The `authTagLength` option has been made more flexible. (Tobias Nießen) [#20235](#20235), [#20039](#20039) * **http**: - Handling of `close` and `aborted` events has been made more consistent. (Robert Nagy) [#20075](#20075), [#20611](#20611) * Embedder support: - Functions for creating V8 `Isolate` and `Context` objects with Node.js-specific behaviour have been added to the API. (helloshuangzi) [#20639](#20639) - Node.js `Environment`s clean up resources before exiting now. (Anna Henningsen) [#19377](#19377) - Support for multi-threaded embedding has been improved. (Anna Henningsen) [#20542](#20542), [#20539](#20539), [#20541](#20541) * **timers**: - `timeout.refresh()` has been added to the public API. (Jeremiah Senkpiel) [#20298](#20298) PR-URL: #20724

Notable Changes: * **addons**: - Fixed a memory leak for users of `AsyncResource` and N-API. (Michael Dawson) [#20668](#20668) * **assert**: - The `error` parameter of `assert.throws()` can be an object containing regular expressions now. (Ruben Bridgewater) [#20485](#20485) * **crypto**: - The `authTagLength` option has been made more flexible. (Tobias Nießen) [#20235](#20235), [#20039](#20039) * **http**: - Handling of `close` and `aborted` events has been made more consistent. (Robert Nagy) [#20075](#20075), [#20611](#20611) * Embedder support: - Functions for creating V8 `Isolate` and `Context` objects with Node.js-specific behaviour have been added to the API. (helloshuangzi) [#20639](#20639) - Node.js `Environment`s clean up resources before exiting now. (Anna Henningsen) [#19377](#19377) - Support for multi-threaded embedding has been improved. (Anna Henningsen) [#20542](#20542), [#20539](#20539), [#20541](#20541) * **esm**: - Builtin modules (e.g. `fs`) now provide named exports in ES6 modules. (Gus Caplan) [#20403](#20403) * **timers**: - `timeout.refresh()` has been added to the public API. (Jeremiah Senkpiel) [#20298](#20298) PR-URL: #20724

Notable Changes: * **addons**: - Fixed a memory leak for users of `AsyncResource` and N-API. (Michael Dawson) [#20668](#20668) * **assert**: - The `error` parameter of `assert.throws()` can be an object containing regular expressions now. (Ruben Bridgewater) [#20485](#20485) * **crypto**: - The `authTagLength` option has been made more flexible. (Tobias Nießen) [#20235](#20235), [#20039](#20039) * **http**: - Handling of `close` and `aborted` events has been made more consistent. (Robert Nagy) [#20075](#20075), [#20611](#20611) * Embedder support: - Functions for creating V8 `Isolate` and `Context` objects with Node.js-specific behaviour have been added to the API. (Allen Yonghuang Wang) [#20639](#20639) - Node.js `Environment`s clean up resources before exiting now. (Anna Henningsen) [#19377](#19377) - Support for multi-threaded embedding has been improved. (Anna Henningsen) [#20542](#20542), [#20539](#20539), [#20541](#20541) * **esm**: - Builtin modules (e.g. `fs`) now provide named exports in ES6 modules. (Gus Caplan) [#20403](#20403) * **timers**: - `timeout.refresh()` has been added to the public API. (Jeremiah Senkpiel) [#20298](#20298) PR-URL: #20724

* addons: - Fixed a memory leak for users of `AsyncResource` and N-API. (Michael Dawson) #20668 * assert: - The `error` parameter of `assert.throws()` can be an object containing regular expressions now. (Ruben Bridgewater) #20485 * crypto: - The `authTagLength` option has been made more flexible (Tobias Nießen) #20235) #20039 * esm: - Builtin modules (e.g. `fs`) now provide named exports in ES6 modules. (Gus Caplan) #20403 * http: - Handling of `close` and `aborted` events has been made more consistent. (Robert Nagy) #20075 #20611 * module: - add --preserve-symlinks-main (David Goldstein) #19911 * timers: - `timeout.refresh()` has been added to the public API. (Jeremiah Senkpiel) #20298 * Embedder support: - Functions for creating V8 `Isolate` and `Context` objects with Node.js-specific behaviour have been added to the API. (Allen Yonghuang Wang) #20639 - Node.js `Environment`s clean up resources before exiting now. (Anna Henningsen) #19377 - Support for multi-threaded embedding has been improved. (Anna Henningsen) #20542 #20539 #20541 PR-URL: #20724

nodejs-github-bot added the c++ Issues and PRs that require attention from people who are familiar with C++. label May 9, 2018

addaleax approved these changes May 9, 2018

View reviewed changes

addaleax added semver-minor PRs that contain new features and should be released in the next minor version. embedding Issues and PRs related to embedding Node.js in another project. labels May 9, 2018

jasnell approved these changes May 9, 2018

View reviewed changes

gireeshpunathil approved these changes May 10, 2018

View reviewed changes

XadillaX approved these changes May 10, 2018

View reviewed changes

bnoordhuis requested changes May 10, 2018

View reviewed changes

addaleax reviewed May 11, 2018

View reviewed changes

addaleax added semver-major PRs that contain breaking changes and should be released in the next major version. and removed semver-minor PRs that contain new features and should be released in the next minor version. labels May 11, 2018

src: add public API to create isolate and context

4c48b52

helloshuangzi force-pushed the nodeAPI branch from 30c5b52 to 4c48b52 Compare May 11, 2018 18:38

bnoordhuis reviewed May 11, 2018

View reviewed changes

src: use unique_ptr to save Free calls

32f988b

bnoordhuis approved these changes May 12, 2018

View reviewed changes

helloshuangzi deleted the nodeAPI branch May 14, 2018 17:47

addaleax mentioned this pull request May 14, 2018

v10.2.0 proposal #20724

Merged

yous mentioned this pull request Jun 8, 2018

Translate "Node v10.2.0 (Current)" nodejs/nodejs-ko#727

Merged

		@@ -279,6 +279,8 @@ static v8::Isolate* node_isolate;

		DebugOptions debug_options;

		static ArrayBufferAllocator array_buffer_allocator;

	function createUnsafeArrayBuffer(size) {
	zeroFill[0] = 0;
	try {
	return new ArrayBuffer(size);
	} finally {
	zeroFill[0] = 1;
	}
	}

src: add public API to create isolate and context #20639

src: add public API to create isolate and context #20639

Conversation

helloshuangzi commented May 9, 2018

Checklist

addaleax commented May 9, 2018

jasnell commented May 9, 2018

devsnek commented May 9, 2018 • edited Loading

addaleax commented May 9, 2018

gabrielschulhof commented May 10, 2018

gabrielschulhof commented May 10, 2018

bnoordhuis left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gabrielschulhof commented May 10, 2018

mhdawson commented May 10, 2018

helloshuangzi commented May 10, 2018 • edited Loading

addaleax commented May 10, 2018

devsnek commented May 10, 2018 • edited Loading

addaleax commented May 10, 2018

helloshuangzi commented May 10, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

addaleax commented May 11, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bnoordhuis left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

devsnek commented May 9, 2018 •

edited

Loading

helloshuangzi commented May 10, 2018 •

edited

Loading

devsnek commented May 10, 2018 •

edited

Loading