[curl][Win] very high CPU load on page with WebSocket

• Component: Service Workers
• Status: RESOLVED FIXED

[Martijn]

Created attachment 432351 [details]
images and documents to reproduce

I've build WebKit on windows and I'm playing around with the MiniBrowser. I've made a local html page with some javascript which uses a websocket to connect to a local server (seperate application, using boost). As soon as the websocket is opened, the cpu load goes skyhigh, cpu goes in full speed, 1 core full time busy. I've traced it to some libcurl functions, see attachments.
I've like to help fixing it, but I do need some pointers given my very limited knowledge about webkit.

attached is some information to reproduce the issue:
- html document which includes javascript file
- javascript file which opens websocket and does nothing
- screenshot of task manager
- screenshot of visual studio attached to WebKitNetworkProcess.exe

Starting of the build:

c:\[...]>perl tools\Scripts\build-webkit --wincairo --release
Updating WebKitRequirementsWin64.zip...
Found existing release: v2021.05.26
Seeking latest release from WebKitForWindows/WebKitRequirements...
Found release to download: v2021.05.26
Already up-to-date!

[Alexey Proskuryakov]

FWIW, I cannot reproduce this with macOS Safari (needed to adjust the test to run, app2.js -> app3.js).

[Martijn]

hmm, thats unfortunate, I've tried to debug it a little bit but as soon as I add this:

...
} while (rc == -1 && errno == EINTR);
std::cout << std::chrono::system_clock::now().time_since_epoch().count() / 1000 << " errno: " << errno << " rc: " << rc <<  std::endl;
for (auto& stream : m_streamList.values())
...

in CurlStreamScheduler.cpp, the issue goes away (and nothing shows in any terminal unfortunately, though I added AllocConsole() earlier)

It seems like ::select() is returning constantly, that could maybe explain the hot path as shown in the attachment earlier.

I'm not sure how to continue.

[Martijn]

rc = ::select(maxfd + 1, &readfds, &writefds, &exceptfds, &timeout);
} while (rc == -1 && errno == EINTR);
for (auto& stream : m_streamList.values())

I've increased the timeout value to 2000000 us, my webpage is loaded, websocket connection is made, no data is transfering (checked with wireshark) and I put a breakpoint at the line starting with 'for (...'

the breakpoint triggers instantly and what I see is that readfds has a count of 1, writefds has a count of 0, exceptfds has a count of 1. they all contain the same data:
[0] 1084
[1] 1475739....
rest is all equal to [1] (uninitialized I assume)

continuing the function 'TryToTransfer' results in curl returning CURLE_AGAIN.

I would not expect the readfds to return when no data is transfered, does exceptfds give a clue?

[Martijn]

just wondering:

at the bottom of appendMonitoringFd():

if (maxfd < *socket)
  maxfd = *socket;

at the end the the call, maxfd is -1, and the 'maxfx = *socket' is not executed when I step through with the debugger.

changing it to:

int a = *socket;
if (maxfd < a)
  maxfd = a;

somehow resolved this :s

[Radar WebKit Bug Importer]

<rdar://problem/80150503>

[youenn fablet]

Given this is windows/curl specific, maybe windows port maintainer could help you further investigating the issue.

[Fujii Hironori]

Created attachment 432906 [details]
test case

Thank you very much for testing WinCairo and reporting the bug.
I can reproduce the CPU 100% WebKitNetwork.exe problem. Will check.

And, I found that WinCairo has the critical regression at the moment.
Bug 227694 – [curl][Win] wss: WebSocket doesn't work since r271170

[Fujii Hironori]

curl_socket_t is SOCKET on Windows. SOCKET is UINT_PTR in WinSock.

https://github.com/WebKit/WebKit/blob/eab9b2b8ee9405ca95bb415f5ddf58766afe0159/Source/WebCore/platform/network/curl/CurlStream.cpp#L10

>    if (maxfd < socket)

This condition can't be true.

[Fujii Hironori]

Created attachment 432908 [details]
WIP patch

[Fujii Hironori]

Created attachment 432910 [details]
WIP patch

[Fujii Hironori]

Created attachment 432911 [details]
Patch

[Fujii Hironori]

Comment on attachment 432911 [details]
Patch

Clearing flags on attachment: 432911

Committed r279576 (239407@main): <https://commits.webkit.org/239407@main>

[Fujii Hironori]

All reviewed patches have been landed.  Closing bug.