Handle limited open Connections due to keepalive connections

[sgothel]

• Resolves: Handle limited open Connections due to keepalive connections (cool#9621) #9833
• Target version: master

Summary

• ProtocolHandlerInterface::checkTimeout(..)
  is now called via StreamSocket::checkRemoval()
  to ensure tests even w/o input data (was: handlePoll())

• ProtocolHandlerInterface::checkTimeout(..)

  • Add bool return value: true -> shutdown connection, caller shall stop processing
  • Implemented for http::Session
    • Timeout (30s) net::Defaults::HTTPTimeout with missing response
  • Implemented for WebSocketHandler
    • Timeout (2s = net::Defaults::WSPingTimeout)
      after missing WS native frame ping/pong (server only)

• StreamSocket -> Socket (properties moved)

  • bytes sent/received
  • closed state

• Socket (added properties)

  • creation- and last-seen -time
  • socket type and port
  • checkRemoval(..)
    • called directly from SocketPoll::poll()
    • only for IPv4/v6 network connections
    • similar to ProtocolHandlerInterface::checkTimeout(..)
    • added further criteria (age, throughput, ..)
      • Timeout (64s = net::Defaults::SocketPollTimeout)
        if (now - lastSeen) > timeout
      • Timeout (net::Defaults::MinBytesPerSec)
        if Throughput < MinBytesPerSec (disabled by default)
      • TODO: Add maximimal IPv4/IPv6 socket-count criteria, drop oldest.
      • age via Socket::creationTime is currently not used as requested

• SocketPoll::poll()

  • Additionally erases if !socket->isOpen() || socket->checkRemoval()

Changes to original PR

• net::Config -> net::Defaults, without any xml config file side-effects
  • only allow net::Defaults to be changed in unit tests via dedicated callback
• removed Socket age criteria via Socket::creationTime and net::defaults::MaxDuration
  • Socket::creationTime still relevant for minimum throughput criteria
  • dropped the net::defaults::MaxDuration
• max-connections
  • removed under- overflow checks
  • using SC atomic ops (not just relaxed)
• enhanced API doc of Util::TimeAverage (clarification)
• unit tests in one place / commit
• more unification/squashing to single semantic commits
• more sensible order of commits
• removed dead code
• dropped noexcept patch

Further:

• cleaned up the commit messages
• removed the COOLWSD.cpp default config reordering
• reordered the 'Unify socket stats', as it must be upfront the actual connection count limitation
  and before moved _lastSeenHTTPHeader
• fixed another intermediate breakage in the commit chain
• bisectable proof (but no unit test runs ofc) .. via
  • passed git rebase -i --exec "make -j 20" for all commits
• Use scoped enums with efficient types (STATE_ENUM)
  • now using desired STATE_ENUM

Unit tests: +7 files, 879 insertions(+), remaining changed lines +800, -222

Checklist

• I have run make prettier-write and formatted the code.
• All commits have Change-Id
• I have run tests with make check
• Passed certain unit tests and cypress desktop annotations.js
• I have issued make run and manually verified that everything looks okay
• Documentation (manuals or wiki) has been updated or is not required

[sgothel]

This change set created a linkage error (android), the 'usual' type_info, vtable thing.
Can't figure a missed 'virtual' or even pure virtual '=0' virtual key in the code, also reviewed the MOBILEAPP macro.
Hence I need to reproduce and build the android locally here .. WIP.

[caolanm]

might need to move this impl to after

"#endif // !MOBILEAPP" (if mobile needs it) and/or put ifdef MOBILEAPP around it in the header if it doesn't or a !MOBILE stub, something like that anyway

[sgothel]

yeah. Great eyes. Thx!

[Ashod]

Thanks, @sgothel. There are some really good stuff here!

I understand this is a draft, but it reads like a mix of random fixes and improvements (some of it is good and useful, like adding const and = default, make_shared, etc.), re-inventing methods for doing things already supported, and debugging/tracing support for your understanding (to be removed).

It's very hard to read, as there is no clear theme or purpose.

I left various comments, but I didn't dive too deep in a number of areas because it was unclear why we are either replacing an existing approach, or inventing a new one when we can improve on the existing one (where there are shortcomings).

At a minimum, I'd separate the backtrace improvements from the socket improvements, and likely the misc. improvements from new code/features/tests.

Finally, I'm troubled by vformatString and the re-working a lot of the socket states and internals with a huge set of new members that change the existing semantics with unclear benefits.

Perhaps there is some context I'm missing. Perhaps this was agreed on and I'm unaware.

[sgothel]

Thank you @Ashod for the review.

I understand this is a draft, but it reads like a mix of random fixes and improvements (some of it is good and useful, like adding const and = default, make_shared, etc.), re-inventing methods for doing things already supported, and debugging/tracing support for your understanding (to be removed).

The draft status was motivated by my intention to discuss the actual semantics and goals,
i.e. mitigate limited connections - or better, ensure connections are closed as well as
to reduce the risk of running out of sockets/file-handles.

Which brings us to the next point I guess.

It's very hard to read, as there is no clear theme or purpose.

When reading commit by commit, I believe everything is well split apart semantically.
I believed the purpose was well described in #9833 and the commit 3ef99c5
actually starting to address (matching this drafts' description) was meaningful.

Indeed, I added a few unrelated single commits into this review. Understanding that we allow
reasonable small changes w/o issue report, esp of intrinsic nature, I believe it should be OK.
(admin vs workflow compromise as discussed earlier)

I left various comments, but I didn't dive too deep in a number of areas because it was unclear why we are either replacing an existing approach, or inventing a new one when we can improve on the existing one (where there are shortcomings).

I am not aware of really replacing anything. I merely moved and added functionality.

At a minimum, I'd separate the backtrace improvements from the socket improvements, and likely the misc. improvements from new code/features/tests.

As described above, a zero side-effect change in its own commit.

The Backtrace object in particular is useful for all, as we can produce an instance
capturing the current stack, move it around and output where desired.
As discussed, I did not add addr2line or libunwind dependencies, which I understand ofc.

Finally, I'm troubled by vformatString and the re-working a lot of the socket states and internals with a huge set of new members that change the existing semantics with unclear benefits.

I just like printf formatting as it is more performant and flexible (at least to me).
IMHO I use it in a most reasonable save fashion, while being supported by the compiler to
remove the risk of wrong format specifiers and stack values. The latter is still a weakness, which I am aware of.

Perhaps there is some context I'm missing. Perhaps this was agreed on and I'm unaware.

Allow me to continue answering your other remarks below.

[caolanm]

I think it'll be easier to manage this if its split up so that the mostly uncontroversial misc cleanups of existing stuff where no change in logic is intended, the make_unique, make_shared, privator ctor -> = delete, constifying, expansion of auto& to real type, emplace_back, etc goes into a separate pr.

The backtrace debugging aid looks probably useful to have to me, but can also be its own pr I think (and probably not worth fighting the "CodeQL scanning - action / Analyze" step which apparently can't handle the constexpr use which is blocking ci from passing this pr).

And generally trim this pr down to the core change to make it easier to think about it

[Ashod]

Thanks for the responses, @sgothel! We've discussed some offline, so didn't reply here (e.g. printf-style formatting being fast--in fact it's very slow compared to only serialization).

Indeed, I added a few unrelated single commits into this review. Understanding that we allow reasonable small changes w/o issue report, esp of intrinsic nature, I believe it should be OK. (admin vs workflow compromise as discussed earlier)

In general, agreed. But I'll echo @caolanm's comment here that breaking up into PRs, that can be decided to merge independent of other parts, is always helpful.

I should also point out that this PR is well above a 1000 lines. This is a rare size. Considering that some parts change the interface of a large hierarchy (the sockets), which is a fundamental core part of the whole project and very delicate (can easily break things catastrophically), it requires reading and studying far more code (i.e. the context and consequences) to review properly. An isolated and low-risk change of a few 100 lines is fine. A complex few 100 line change (such as in the sockets) is challenging and time consuming. Sometimes we have no choice because we can't break it up. But a 1000 lines that can be easily broken up is hardly necessary.

Finally, even though we can review individual commits, it's preferable to have a single theme to a PR. Helps the reviewer. In some cases a few misc. changes can go together (comment fixes, cosmetics, formatting, logging, etc.). But mixing functional changes with non-functional changes dilutes the process and can reduce the quality of the review. More so with high-risk changes (i.e. socket code) mixed with debugging code (i.e. backtracing).

Regardless, you have a lot of great improvements here, and I don't want to undermine the importance of your contribution!
Thank you.

[sgothel]

Thank you both for your review and kind words.

I have split up single commits into 4 branches, i.e. this pull-request branch and

• Provide snprintf string formatting. #9952
• Add Backtrace class, allowing to conveniently capture a stacktrace #9953
• C++ Cleanup: Default ctor/deleter, drop virtual of virtual-overrides and fix Util::getSystemClockAsString() #9954

For this feature work, I need to complete the unit test incl. the timeout configuration.
Looks OK .. shall provide it in 1-2 days.

[Ashod]

@sgothel, there were multiple suggestions to split this large PR up to reduce its size. That was when it had 18 commits and was "only" a 1000 lines. Now it's past 1500 lines and is 1 commit! I'm sensing this isn't going in the direction of the feedback.

What's the plan, please?

[mmeeks]

As said needs to be split up; lets try to get the less controversial pieces in and merged as separate pull requests - to slim this down =) Can you work on that Sven ?

[sgothel]

As said needs to be split up; lets try to get the less controversial pieces in and merged as separate pull requests - to slim this down =) Can you work on that Sven ?

will do the split later then (tomorrow) + finishing the tests

[sgothel]

@sgothel, there were multiple suggestions to split this large PR up to reduce its size. That was when it had 18 commits and was "only" a 1000 lines. Now it's past 1500 lines and is 1 commit! I'm sensing this isn't going in the direction of the feedback.

What's the plan, please?

As I wrote, I will split out the config and TimeAverage (~400 lines or so) and put the remainder at the bottom -> 3 commits within this pull request. Perhaps more, w/ a fresh look.

But splitting this up in multiple pull request (PR) makes would probably makes no sense, as these pieces are required and each PR must be tested. IMHO.

[sgothel]

Added minor changes addressing @Ashod earlier notes.

[sgothel]

Note: Unit tests: +7 files, 879 insertions(+), remaining changed lines +800, -222

[mmeeks]

Ok - this is great; thanks Sven & Caolan - much easier to review, can go in at the beginning of the next Red merge window; ignore my comments about unit tests - I see them at the end; other minor cleanups can come later / get tacked on.
Nice work - thanks ! =)

[mmeeks]

Actually - I lie the c1,2,3,4 thing really needs fixing or explaining; its not clear to me that it will work =)

[sgothel]

@mmeeks @caolanm addresses Michael's notes

• inlined the changes in rebased commits
• rebased to (Monday's) master

[sgothel]

* rebased to (Monday's) master

I apologize: Should have pushed with a rebase to master 1st, then another push with the changes.
This would have allowed us to see the push delta on the fixed notes only using this github UI diff.
Now it shows the total change :-(

[sgothel]

resolved merge conflict due to changed net/Socket.*