Thanks to visit codestin.com
Credit goes to github.com

Skip to content

fix: Disable compression for websocket dRPC transport #145

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged

Conversation

bryphe-coder
Copy link
Contributor

@bryphe-coder bryphe-coder commented Feb 2, 2022

There is a race condition in the interop between the websocket and dRPC: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the byte[] being sent between them. This can lead to data races, in which both dRPC and the websocket are writing.

This is just tracking some experimentation to fix that race condition

Run results:

status code 412: The provided project history is running. Wait for it to complete importing!`
    workspacehistory_test.go:122: 
        	Error Trace:	workspacehistory_test.go:122
        	Error:      	Condition never satisfied
        	Test:       	TestWorkspaceHistory/CreateHistory
  • Run 5: peer failure
  • Run 6: Pass ✅
  • Run 7: Peer failure

Open Questions:

Is dRPC or websocket at fault for the data race?

It looks like this condition is specifically happening when dRPC decides to [SendError]). This constructs a new byte payload from MarshalError - so dRPC has created this buffer and owns it.

From dRPC's perspective, the callstack looks like this:

// AppendFrame appends a marshaled form of the frame to the provided buffer.
func AppendFrame(buf []byte, fr Frame) []byte {
...
	out := buf
	out = append(out, control).   // <---------

This should be fine, since dPRC create this buffer, and is taking the byte buffer constructed from MarshalError and tacking a bunch of headers on it to create a proper frame.

Once dRPC is done writing, it hangs onto the buffer and resets it here_: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73

However... the websocket implementation, once it gets the buffer, it runs a statelessDeflate here, which compresses the buffer on the fly. This functionality actually mutates the buffer in place, which is where get our race.

In the case where the byte[] aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing dRPC -> websocket, they both are manipulating it (dRPC is reusing the buffer for the next write, and websocket is compressing on the fly).

Why does cloning on Read fail?

Get a bunch of errors like:

2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0
2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF
2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF
2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0

UPDATE:

We decided we could disable websocket compression, which would avoid the race because the in-place deflate operaton would no longer be run. Trying that out now:

@codecov
Copy link

codecov bot commented Feb 2, 2022

Codecov Report

Merging #145 (eef4d3a) into provisionerdaemon (2a76b60) will decrease coverage by 0.21%.
The diff coverage is 100.00%.

Impacted file tree graph

@@                  Coverage Diff                  @@
##           provisionerdaemon     #145      +/-   ##
=====================================================
- Coverage              68.04%   67.83%   -0.22%     
=====================================================
  Files                    101      101              
  Lines                   5095     5098       +3     
  Branches                  68       68              
=====================================================
- Hits                    3467     3458       -9     
- Misses                  1324     1333       +9     
- Partials                 304      307       +3     
Flag Coverage Δ
unittest-go-macos-latest 64.59% <100.00%> (-0.22%) ⬇️
unittest-go-ubuntu-latest 66.89% <100.00%> (-0.07%) ⬇️
unittest-go-windows-latest 64.30% <100.00%> (-0.08%) ⬇️
unittest-js 64.92% <ø> (ø)
Impacted Files Coverage Δ
coderd/provisionerdaemons.go 48.56% <100.00%> (-0.31%) ⬇️
peerbroker/dial.go 76.19% <0.00%> (-4.77%) ⬇️
peer/channel.go 84.75% <0.00%> (-2.44%) ⬇️
peer/conn.go 80.36% <0.00%> (-0.26%) ⬇️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2a76b60...eef4d3a. Read the comment docs.

@bryphe-coder bryphe-coder self-assigned this Feb 2, 2022
Comment on lines 59 to 61
cpy := make([]byte, len(p))
copy(cpy, p)
return c.conn.Write(cpy)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You can do 1 line slice copies as well!

Suggested change
cpy := make([]byte, len(p))
copy(cpy, p)
return c.conn.Write(cpy)
return c.conn.Write(append(([]byte)(nil), p...))

@bryphe-coder bryphe-coder changed the title experiment: Clone bytes on write between provisionerdaemon <-> websocket fix: Disable compression for websocket dRPC transport Feb 2, 2022
@bryphe-coder bryphe-coder marked this pull request as ready for review February 2, 2022 21:31
@bryphe-coder bryphe-coder merged commit 03ed951 into provisionerdaemon Feb 2, 2022
@bryphe-coder bryphe-coder deleted the bryphe/provisionerdaemon/experiment-race branch February 2, 2022 21:32
kylecarbs added a commit that referenced this pull request Feb 3, 2022
* feat: Add history middleware parameters

These will be used for streaming logs, checking status,
and other operations related to workspace and project
history.

* refactor: Move all HTTP routes to top-level struct

Nesting all structs behind their respective structures
is leaky, and promotes naming conflicts between handlers.

Our HTTP routes cannot have conflicts, so neither should
function naming.

* Add provisioner daemon routes

* Add periodic updates

* Skip pubsub if short

* Return jobs with WorkspaceHistory

* Add endpoints for extracting singular history

* The full end-to-end operation works

* fix: Disable compression for websocket dRPC transport (#145)

There is a race condition in the interop between the websocket and `dRPC`: https://github.com/coder/coder/runs/5038545709?check_suite_focus=true#step:7:117 - it seems both the websocket and dRPC feel like they own the `byte[]` being sent between them. This can lead to data races, in which both `dRPC` and the websocket are writing.

This is just tracking some experimentation to fix that race condition

## Run results: ##
- Run 1: peer test failure
- Run 2: peer test failure
- Run 3: `TestWorkspaceHistory/CreateHistory`  - https://github.com/coder/coder/runs/5040858460?check_suite_focus=true#step:8:45
```
status code 412: The provided project history is running. Wait for it to complete importing!`
```
- Run 4: `TestWorkspaceHistory/CreateHistory` - https://github.com/coder/coder/runs/5040957999?check_suite_focus=true#step:7:176
```
    workspacehistory_test.go:122: 
        	Error Trace:	workspacehistory_test.go:122
        	Error:      	Condition never satisfied
        	Test:       	TestWorkspaceHistory/CreateHistory
```
- Run 5: peer failure
- Run 6: Pass ✅ 
- Run 7: Peer failure

## Open Questions: ##

### Is `dRPC` or `websocket` at fault for the data race?

It looks like this condition is specifically happening when `dRPC` decides to [`SendError`]). This constructs a new byte payload from [`MarshalError`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/error.go#L15) - so `dRPC` has created this buffer and owns it.

From `dRPC`'s perspective, the callstack looks like this:
- [`sendPacket`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcstream/stream.go#L253)
  - [`writeFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L65)
    - [`AppendFrame`](https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/packet.go#L128)
      - with finally the data race happening here:
```go
// AppendFrame appends a marshaled form of the frame to the provided buffer.
func AppendFrame(buf []byte, fr Frame) []byte {
...
	out := buf
	out = append(out, control).   // <---------
```

This should be fine, since `dPRC` create this buffer, and is taking the byte buffer constructed from `MarshalError` and tacking a bunch of headers on it to create a proper frame.

Once `dRPC` is done writing, it _hangs onto the buffer and resets it here__: https://github.com/storj/drpc/blob/f6e369438f636b47ee788095d3fc13062ffbd019/drpcwire/writer.go#L73

However... the websocket implementation, once it gets the buffer, it runs a `statelessDeflate` [here](https://github.com/nhooyr/websocket/blob/8dee580a7f74cf1713400307b4eee514b927870f/write.go#L180), which compresses the buffer on the fly. This functionality actually [mutates the buffer in place](https://github.com/klauspost/compress/blob/a1a9cfc821f00faf2f5231beaa96244344d50391/flate/stateless.go#L94), which is where get our race.

In the case where the `byte[]` aren't being manipulated anywhere else, this compress-in-place operation would be safe, and that's probably the case for most over-the-wire usages. In this case, though, where we're plumbing `dRPC` -> websocket, they both are manipulating it (`dRPC` is reusing the buffer for the next `write`, and `websocket` is compressing on the fly).

### Why does cloning on `Read` fail?

Get a bunch of errors like:
```
2022/02/02 19:26:10 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0
2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF
2022/02/02 19:26:25 [ERR] yamux: Failed to read header: unexpected EOF
2022/02/02 19:26:25 [WARN] yamux: frame for missing stream: Vsn:0 Type:0 Flags:0 StreamID:0 Length:0
```

# UPDATE:

We decided we could disable websocket compression, which would avoid the race because the in-place `deflate` operaton would no longer be run. Trying that out now:

- Run 1: ✅ 
- Run 2: https://github.com/coder/coder/runs/5042645522?check_suite_focus=true#step:8:338
- Run 3: ✅ 
- Run 4: https://github.com/coder/coder/runs/5042988758?check_suite_focus=true#step:7:168
- Run 5: ✅

* fix: Remove race condition with acquiredJobDone channel (#148)

Found another data race while running the tests: https://github.com/coder/coder/runs/5044320845?check_suite_focus=true#step:7:83

__Issue:__ There is a race in the p.acquiredJobDone chan - in particular, there can be a case where we're waiting on the channel to finish (in close) with <-p.acquiredJobDone, but in parallel, an acquireJob could've been started, which would create a new channel for p.acquiredJobDone. There is a similar race in `close(..)`ing the channel, which also came up in test runs.

__Fix:__ Instead of recreating the channel everytime, we can use `sync.WaitGroup` to accomplish the same functionality - a semaphore to make close wait for the current job to wrap up.

* fix: Bump up workspace history timeout (#149)

This is an attempted fix for failures like: https://github.com/coder/coder/runs/5043435263?check_suite_focus=true#step:7:32

Looking at the timing of the test:
```
    t.go:56: 2022-02-02 21:33:21.964 [DEBUG]	(terraform-provisioner)	<provision.go:139>	ran apply
    t.go:56: 2022-02-02 21:33:21.991 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    t.go:56: 2022-02-02 21:33:22.050 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    t.go:56: 2022-02-02 21:33:22.090 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    t.go:56: 2022-02-02 21:33:22.140 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    t.go:56: 2022-02-02 21:33:22.195 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    t.go:56: 2022-02-02 21:33:22.240 [DEBUG]	(provisionerd)	<provisionerd.go:162>	skipping acquire; job is already running
    workspacehistory_test.go:122: 
        	Error Trace:	workspacehistory_test.go:122
        	Error:      	Condition never satisfied
        	Test:       	TestWorkspaceHistory/CreateHistory
```

It  appears that the `terraform apply` job had just finished - with less than a second to spare until our `require.Eventually` completes - but there's still work to be done (ie, collecting the state files). So my suspicion is that terraform might, in some cases, exceed our 5s timeout.

Note that in the setup for this test - there is a similar project history wait that waits for 15s, so I borrowed that here.

In the future - we can look at potentially using a simple echo provider to exercise this in the unit test, in a way that is more reliable in terms of timing. I'll log an issue to track that.

Co-authored-by: Bryan <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants