[1.0.2] non-atomic append fallback to random_access_file for old kernels

[spoonincode]

random_access_file, used for state_history in Spring, makes use of a kernel feature added in 4.16. I didn't really think much of that requirement because 4.16 was released well over 6 years (before EOS genesis!), and is well outside the scope of our officially supported platforms.

Yet we've now seen two reports of users stumbling on this:

• A user running Ubuntu 18.04¹, which for non-desktop installs will use 4.15 by default. It's possible to opt in to the HWE kernel (default on desktop installs) that moves up to 5.4. This was seen on Telegram somewhere I can't find the link at the moment.
• A user running a Ubuntu 22.04 container on Amazon Linux 2. AL2² actually has three supported kernels: 4.14, 5.10, and 5.15, with 5.10 being the default for new installs now. But the user had an old configuration still set to 4.14. Seen in trace-history/chain-state-history option of state_history_plugin breaking with v1.0.1's snapshot-import/node running #776

state_history doesn't actually need this atomic append, the implementation was just done this way for cleanliness, conciseness, and performance. So adding a fallback for old kernels here doesn't impact correctness for state_history. Still not really sure I want to add this additional chatter to the code though.

Footnotes

1. I regret a little not bumping the reproducible build's runtime requirement before 1.0.0 release so that it simply would not work on Ubuntu 18 -- other projects have made this bump recently, for example https://github.com/bitcoin/bitcoin/pull/29987 ↩

2. AL2 is in its 11th hour of support. It was originally scheduled to be EOL July 2023 but they extended it to July 2025. ↩

[heifner]

pwritev2 ignores offset (except -1) when given RWF_APPEND.
pwritev you need size because you want to append. I agree it is cleaner and more concise. Is it actually more performant?

I wonder if pwritev should just be used.

[heifner]

Oh, never mind. It is more performant because you don't have to call size().

[heifner]

When I saw the report about amznlinux2, I wondered if we shouldn't try to support it. This resolves #776 then. My vote would be to put it in 1.0.2.

[ericpassmore]

Note:start
category: Other
component: Internal
summary: Support Linux Kernel 4.15, by falling back to random_access_file for the older kernel.
Note:end