src: track memory retainer fields #26161

gireeshpunathil · 2019-02-17T12:46:38Z

If retainers are embedded in retainers, direct tracking
those lead to double tracking. Instead, use the base tracker
to track the field objects.

cc @addaleax [ I should admit that I don't have full understanding of this feature, so let me know if this needs rewrite ]

Checklist

make -j4 test (UNIX), or vcbuild test (Windows) passes
tests and/or benchmarks are included
documentation is changed or added
commit message follows commit guidelines

nodejs-github-bot · 2019-02-17T12:46:40Z

@gireeshpunathil build started: https://ci.nodejs.org/blue/organizations/jenkins/node-test-pull-request-lite-pipeline/detail/node-test-pull-request-lite-pipeline/2641/pipeline

addaleax · 2019-02-17T16:34:43Z

src/memory_tracker-inl.h

@@ -81,6 +81,10 @@ void MemoryTracker::TrackFieldWithSize(const char* edge_name,
  if (size > 0) AddNode(GetNodeName(node_name, edge_name), size, edge_name);
 }

+void MemoryTracker::TrackInlineField(const MemoryRetainer& field) {
+  field.MemoryInfo(this);


I feel like if we want to provide this as an utility, it might make sense to perform something similar to Track() and/or re-use it? That would be nice if we do want to see these fields as full “objects” on their own, and want to show them that way in heap dumps.

Then again, I think you’re making a good point about this approach being a good way to address the problem itself; I just didn’t think of a direct approach like this. 😄 Adding 2 calls à la thread_stopper_.MemoryInfo(tracker); in #26099 would also be a good way to address this, as far as that PR is concerned in particular?

gireeshpunathil · 2019-02-18T11:19:10Z

@addaleax - pushed a change; ptal. I tested the two inline fields in question using test/pummel/test-heapdump-worker.js and got this. Please note the size in reduction for the Worker from 344 to 200 through the new logic. does this approch sound reasonable to you?

  { name: 'Node / Worker',
    isRoot: true,
    size: 200,
    edges:
     [ { name: 'wrapped',
         to:
          { name: '<JS Node>',
            isRoot: false,
            size: 0,
            edges: [Array],
            value: [Worker] } },
       { name: 'parent_port',
         to:
          { name: 'Node / MessagePort',
            isRoot: true,
            size: 88,
            edges: [Array] } },
       { name: 'on_thread_finished_',
         to:
          { name: 'Node / AsyncRequest',
            isRoot: false,
            size: 72,
            edges: [Array] } },
       { name: 'thread_stopper_',
         to:
          { name: 'Node / AsyncRequest',
            isRoot: false,
            size: 72,
            edges: [Array] } } ] }

src/memory_tracker-inl.h

addaleax

Thanks, LGTM!

gireeshpunathil · 2019-02-18T13:43:25Z

CI: https://ci.nodejs.org/job/node-test-pull-request/20869/

addaleax · 2019-02-21T21:30:11Z

CI is green … Maybe somebody has time to give this a second review?

joyeecheung · 2019-02-22T00:18:46Z

src/memory_tracker.h

+  // be one, of the container object which the current field is part of.
+  // Reduce the size of memory from the container so that retentions
+  // are accounted properly.
+  inline void TrackInlineField(const MemoryRetainer* retainer,


I am not sure what this is trying to achieve, is there a use case of this API?

Normally the size of the pointer to the retainer should be accounted into the self size of the parent but self size of the retainer should not be accounted, so I can't tell when it's necessary to do a subtraction - maybe there is a misuse of the API?

If this is trying to address the issues that came up from #26099 there are two ways to solve that:

Do not use SET_SELF_SIZE(Worker), which is merely a convenience macro that implements Worker::SelfSize() with sizeof(Worker) to calculate the self size of the worker - that's where we encounter the first tracking of the inline field. Instead, override Worker::SelfSize() to exclude the size of the non-pointer fields, and track them later in Worker::MemoryInfo().

Keep using SET_SELF_SIZE(Worker) for the initial calculation, but rename this method into something like SplitNonPointerField() (SplitStackAllocatedField()?), because this does not necessarily add more memory into the tracker - instead it's splitting off existing memory that have already been tracked, and potentially adding more (if the retainer overrides MemoryInfo() to add additional fields that are not accounted by a sizeof())

@joyeecheung - thanks.

Probably these many scenarios exists (with a component cmp and a container cnt with cnt a retainer):

cmp is a normal primitive field: self size of cnt covers size of cmp

cmp is a composite object (with or without primitive fields): self size of cnt covers size of cmp

cmp is a pointer to a composite object: self size of cnt covers size of cmp (size of a pointer)

cmp is a composite object (with pointer type fields): self size of cnt covers size of cmp, not the dynamic memory [ example: messageport in worker ]

cmp is a composite retainer object (with pointer types): [ current use case ]

cmp is a pointer to a retainer object (with pointer types): [ none at the moment? ]

We have Worker with 200 bytes excluding 2 inline AsyncRequest fields of 72 bytes each. So the question is: should we add those to the worker (because those belongs to worker) or add separately - as they are retainers themselves.

If I follow your approach 1, we need to perform that in all such scenarios? in which case arguably the parent retainer (Worker's referencing object if any) could do the same for worker, and potentially render the Retainer abstraction meaningless?

second proposal looks fair to me, but would love to hear from @addaleax as well. Can SplitStackAllocatedField be somewhat misleading, as there is no memory that is stack-allocated here?

because this does not necessarily add more memory into the tracker - instead it's splitting off existing memory that have already been tracked, and potentially adding more (if the retainer overrides MemoryInfo() to add additional fields that are not accounted by a sizeof())

I’m not sure, but I don’t really see this as a issue/as not “tracking” these fields? Then again, re-naming is not a big deal, so I’m okay with anything.

should we add those to the worker (because those belongs to worker) or add separately - as they are retainers themselves.

That depends on whether we want to display AsyncRequest separately as a Node referenced by the Worker node, and any other fields referenced by those AsyncRequests recursively.

If I follow your approach 1, we need to perform that in all such scenarios? in which case arguably the parent retainer (Worker's referencing object if any) could do the same for worker, and potentially render the Retainer abstraction meaningless?

The abstraction helps adding nodes recursively instead of tracking everything from the root, allowing the MemoryRetainer implementations define how they want their fields to be tracked.

Can SplitStackAllocatedField be somewhat misleading, as there is no memory that is stack-allocated here?

Right, the fields are not really stack allocated if the whole thing is not stack allocated.

I’m not sure, but I don’t really see this as a issue/as not “tracking” these fields?

I am fine with the current naming if the comments are changed to reflect when this method should be used and what issue it is trying to address. Use when a retainer is embedded in another. and Reduce the size of memory from the container so that retentions are accounted properly. is too ambiguous IMO, the core of the issue this method is trying to solve arise from that the parent does not implement its SelfSize() correctly to exclude the memory that is supposed to be tracked as the self size of another node.

@joyeecheung - thanks, I will amend the comments, aligning with above conversation.

@joyeecheung - amended the comment, ptal!

If retainers are embedded in retainers, direct tracking those lead to double tracking. Instead, use a special tracker that adjusts the tracking for the container object. PR-URL: nodejs#26161 Reviewed-By: Anna Henningsen <[email protected]>

gireeshpunathil · 2019-02-28T04:34:32Z

landed as 17b4949

If retainers are embedded in retainers, direct tracking those lead to double tracking. Instead, use a special tracker that adjusts the tracking for the container object. PR-URL: #26161 Reviewed-By: Anna Henningsen <[email protected]>

nodejs-github-bot added the c++ Issues and PRs that require attention from people who are familiar with C++. label Feb 17, 2019

gireeshpunathil requested a review from addaleax February 17, 2019 12:48

addaleax reviewed Feb 17, 2019

View reviewed changes

gireeshpunathil force-pushed the memtrackex branch from 56f5969 to 6055693 Compare February 18, 2019 11:18

addaleax reviewed Feb 18, 2019

View reviewed changes

src/memory_tracker-inl.h Show resolved Hide resolved

addaleax approved these changes Feb 18, 2019

View reviewed changes

BridgeAR added the author ready PRs that have at least one approval, no pending requests for changes, and a CI started. label Feb 20, 2019

joyeecheung reviewed Feb 22, 2019

View reviewed changes

joyeecheung mentioned this pull request Feb 23, 2019

doc: fix the example implementation of MemoryRetainer #26262

Closed

3 tasks

gireeshpunathil removed the author ready PRs that have at least one approval, no pending requests for changes, and a CI started. label Feb 28, 2019

src: track memory retainer fields

17b4949

If retainers are embedded in retainers, direct tracking those lead to double tracking. Instead, use a special tracker that adjusts the tracking for the container object. PR-URL: nodejs#26161 Reviewed-By: Anna Henningsen <[email protected]>

gireeshpunathil force-pushed the memtrackex branch from 726350f to 17b4949 Compare February 28, 2019 04:34

gireeshpunathil merged commit 17b4949 into nodejs:master Feb 28, 2019

gireeshpunathil mentioned this pull request Mar 1, 2019

worker: refactor thread life cycle management #26099

Closed

4 tasks

BridgeAR mentioned this pull request Mar 4, 2019

v11.11.0 proposal #26322

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

src: track memory retainer fields #26161

src: track memory retainer fields #26161

gireeshpunathil commented Feb 17, 2019

nodejs-github-bot commented Feb 17, 2019

addaleax Feb 17, 2019

gireeshpunathil commented Feb 18, 2019

addaleax left a comment

gireeshpunathil commented Feb 18, 2019

addaleax commented Feb 21, 2019

joyeecheung Feb 22, 2019 •

edited

Loading

joyeecheung Feb 22, 2019 •

edited

Loading

gireeshpunathil Feb 22, 2019

addaleax Feb 22, 2019

joyeecheung Feb 22, 2019

gireeshpunathil Feb 22, 2019

gireeshpunathil Feb 23, 2019

gireeshpunathil commented Feb 28, 2019

src: track memory retainer fields #26161

src: track memory retainer fields #26161

Conversation

gireeshpunathil commented Feb 17, 2019

Checklist

nodejs-github-bot commented Feb 17, 2019

addaleax Feb 17, 2019

Choose a reason for hiding this comment

gireeshpunathil commented Feb 18, 2019

addaleax left a comment

Choose a reason for hiding this comment

gireeshpunathil commented Feb 18, 2019

addaleax commented Feb 21, 2019

joyeecheung Feb 22, 2019 • edited Loading

Choose a reason for hiding this comment

joyeecheung Feb 22, 2019 • edited Loading

Choose a reason for hiding this comment

gireeshpunathil Feb 22, 2019

Choose a reason for hiding this comment

addaleax Feb 22, 2019

Choose a reason for hiding this comment

joyeecheung Feb 22, 2019

Choose a reason for hiding this comment

gireeshpunathil Feb 22, 2019

Choose a reason for hiding this comment

gireeshpunathil Feb 23, 2019

Choose a reason for hiding this comment

gireeshpunathil commented Feb 28, 2019

joyeecheung Feb 22, 2019 •

edited

Loading

joyeecheung Feb 22, 2019 •

edited

Loading