CI fails[1] with this series, but I think it's unrelated- its only chg, and test-http-bad-server.t fails for me locally in a different way, but identically with and without this series applied.

--- /mnt/c/Users/Matt/hg/tests/test-http-bad-server.t
+++ /mnt/c/Users/Matt/hg/tests/test-http-bad-server.t.err
@@ -42,7 +42,7 @@
   $ cat hg.pid > $DAEMON_PIDS

   $ hg clone http://localhost:$HGPORT/ clone
-  abort: error: (\$ECONNRESET\$|\$EADDRNOTAVAIL\$) (re)
+  abort: error: bad HTTP status line: ''
   [100]

[1] https://foss.heptapod.net/mercurial/mercurial-devel/-/jobs/552052

pickle.load() requires that the passed file’s read() returns exactly as many bytes as was passed to its argument and readinto() fills the passed buffer completely; otherwise it raises the “pickle data was truncated” error or EOFError if unambiguously encountering EOF.

For buffered streams this is the case if they are non-interactive and EOF is not reached (which we handle specifically) (according to https://docs.python.org/3/library/io.html#io.BufferedIOBase.read).

However, we can’t use buffered streams because the selector checks whether the unbuffered stream is readable while we need to check that the stream passed to pickle.load() is readable (the bug resulting from this was fixed in 12491abf93bd87b057cb6826e36606afa1cee88a).

_blockingreader wraps the unbuffered stream to meet the expectations of pickle.load() while not reading too much data to create the problem described in the previous paragraph.

In D12628#193389, @mjacob wrote:

pickle.load() requires that the passed file’s read() returns exactly as many bytes as was passed to its argument and readinto() fills the passed buffer completely; otherwise it raises the “pickle data was truncated” error or EOFError if unambiguously encountering EOF.
For buffered streams this is the case if they are non-interactive and EOF is not reached (which we handle specifically) (according to https://docs.python.org/3/library/io.html#io.BufferedIOBase.read).
However, we can’t use buffered streams because the selector checks whether the unbuffered stream is readable while we need to check that the stream passed to pickle.load() is readable (the bug resulting from this was fixed in 12491abf93bd87b057cb6826e36606afa1cee88a).
_blockingreader wraps the unbuffered stream to meet the expectations of pickle.load() while not reading too much data to create the problem described in the previous paragraph.

Interresting. Can I convince you to send a patch adding this explanation to the class docstring ? (and maybe fixing the class name)

In D12628#193391, @marmoute wrote:

Interresting. Can I convince you to send a patch adding this explanation to the class docstring ? (and maybe fixing the class name)

Yes, I planned to do so. Should this go to the default branch?

In D12628#193397, @mjacob wrote:

In D12628#193391, @marmoute wrote:

Interresting. Can I convince you to send a patch adding this explanation to the class docstring ? (and maybe fixing the class name)

Yes, I planned to do so. Should this go to the default branch?

Documentation changes are OK for stable, but renames aren't, so putting everything on default is probably easiest.

In D12628#193398, @mharbison72 wrote:

In D12628#193397, @mjacob wrote:

In D12628#193391, @marmoute wrote:

Interresting. Can I convince you to send a patch adding this explanation to the class docstring ? (and maybe fixing the class name)

Yes, I planned to do so. Should this go to the default branch?

Documentation changes are OK for stable, but renames aren't, so putting everything on default is probably easiest.

Yeah overall the philosophy of stable includes "restrict churn to minimum".

Okay, I sent some patches adding documentation and comments (to the mailing list).