Created 2020-03-20T113857.rxp on branch 8.0.

subject: Pike 8.0: SSL, Search, Postgres, Protocols.HTTP, describe_backtrace, tzdata2019c
from: 9f8e0579
to: 5020c16a
originator: Henrik Grubbstrom
depends: 2019-08-09T164123
restart: true

Multiple fixes:

• Search.Filter.HTML: Return the first <title>.

There may be multiple <title>-tags in an html document where the
later are related to SVG.

Fixes #10020.

• pgsql: Toggle cache_autoprepared_statements default.

• Documentation [Standards.X509]: Fixed some typos.

Fixes #10012.

• pgsql: Simplify code.

• pgsql: Simplify error handling for users.

• pgsql: Disentangle potentially shared empty arrays.

• SSL: Do not accept MD2, MD5 and SHA1 signatures any longer

Web browsers and others have distrusted SHA1 a long while ago, MD2 and
MD5 are at this point simply dangerous to support. However, we now
support configuration of supported signature algorithms via
@[SSL.Context] in case any use case simply needs to support older
signature algorithms.

• Standard.X509: Make signature algorithms configurable

• pgsql: Cast integers to floats, when floats are desired.

• pgsql: Avoid output_to() race on large requests.

• pgsql: Cater for destructed objects.

• pgsql: Do not pass along strings with NUL characters in UTF-8.

• pgsql: 08P01 is a protocol error that is fatal.

• Graphics.Graph: Prevent dumping of create_graph.pike

"Resolves" an issue with incorrect dumping of this module.

• SSL.File: Signal errors during handshake via close callback

We relatively recently stopped claiming that a connection was
peer_closed when in fact we gave up on the SSL connection with a fatal
alert. This caused us to signal some errors via the write callback
instead. This restores the old signalling behaviour by checking for
local_fatal state in some places we only checked for peer_closed before.

• SSL: Fix ticket_enabled logic a little bit

In practice this ought to have little effect though, as to accept a
ticket we also need to be in STATE_wait_for_ticket anyway.

• SSL: Do not purge sessions for crudely closed connections

Not negotiating the close at the TLS level is in spec, and should not
lead to sessions being purged.

• SSL: Only accept wildcards for one level

* in CN / SubjectAlternativeName:dNS is supposed to only match one level
of the hostname, i.e. *.example.com matches foo.example.com but not
www.foo.example.com.

• Let client code detect support for new Search.Utils.Logger constructor, and
simplify whitespace generation.

• Master: Fix issue where describe_string() could lengthen the output.

Fixes edge-case where the string to be formatted (m) is shorter than
the maxlen, but the formatted string (t) is longer. This caused the
truncation code to instead lengthen the string...

Fixes [PIKE-224].

• Search.Utils.Logger: Added support for specifying indentation width when creating a Logger instance. [WS-581]

• Protocols.HTTP.Server.Request: Added http header 'Transfer-Encoding'.

• pgsql: Closed portals should shut down the connection.

• pgsql: Track back deadlock, explicitly release lock, speeds up code.

• Protocols.HTTP: Use Protocols.DNS.async_host_to_ip().

Most Protocols.HTTP.Query objects had their own Protocols.DNS.async_client
(and related Stdio.UDP).

Fixes [PIKE-195].

• SSL.ClientConnection: Fix session ticket support

The server signals reuse of the session by echoing the session_id we
provide (as well as by a shortened handshake/server hello). Presence of
the session ticket extension informs of the server's intention to send a
new session ticket (or replace an existing one), hence tickets_enabled
does not need to be true (and apparently rarely is) when we resume
sessions based on tickets with other SSL implementations on the server
side.

• SSL.clientConnection: Session tickets (RFC 4507 and RFC 5077).

Client side support for session tickets.

Implementation only verified against itself.

Backported from 95ad6e43 by Tobias
Josefowitz.

• SSL.ServerConnection: Session tickets (RFC 4507 and RFC 5077).

Server side support for session tickets.

Note that the default ticket encoding is to use the session_id,
it thus uses server side state. The ticket encoding can be changed
by overriding {en,de}code_ticket() in SSL.Context.

Implementation verified against OpenSSL's s_client.

Backported from 372b2a05 by Tobias
Josefowitz.

• Updated to tzdata2019c.

• Thread.ResourceCountKey: Survive being used with old pike binaries.

Old pike 8.0 binaries do not have Pike.DestructImmediate, and it is
not meaningful to attempt to kludge around it, as the corresponding
program flag is not propagated with them anyway.

• Pike: Move DestructImmediate to the Pike module.

• Thread.ResourceCountKey: Make destruct immediate

• SSL.File: Flush data before regarding the connection closed.

• SSL.Connection: Improved diagnostics when handshaking.

• SSL.File: Return EPIPE frome write() if the peer has closed the connection.

• SSL.File: query_{suite,version}() now both return -1 when unknown.

Fixes some testsuite failures.

• SSL.File: Only call the write_callback if it exists...

• SSL.Session: Initialize cipher_suite to SSL_invalid_suite.

This makes it easy to see whether the handshake has completed
successfully.

• SSL.Connection: Only set peer_closed when the peer actually has closed.

• SSL.Constants: Added SSL_invalid_suite.

• SSL.File: Adjusted some state macros.

SSL_HANDSHAKING is now terminated when the connection has
entered the CONNECTION_local_fatal state.

SSL_INTERNAL_WRITING now requires there to not be any write errors.

• SSL.File: Improved diagnostics in SSL3_DEBUG mode.

• SSL.File: The write callback must not be called during handshaking.

Fixes several testsuite failures.

• SSL.File: Call the write_callback when pending write error.

• Protocols.HTTP: Keep proxy CONNECT and request headers separate.

This makes sure that we do not leak Authorization- and Cookie-
headers to the proxy.

Fixes [PIKE-216].

• Protocols.HTTP: Remove line breaks from base64 basic auth tokens

• Master: Avoid adding end of line padding in describe_backtrace().

describe_bactrace() sometimes used the rough-linebreak mode of sprintf()
to handle long lines, this caused all lines to be padded to the linewith
at end of line (ie often with several 1000s of extra spaces per line).

Fixes [PIKE-211].

• TURBO2-1398: Allow disabling of wrapping of describe_backtrace output.

The wrapping can add lots of spaces to the end of lines to pad
them, resulting in enormous backtrace files in some circumstances.

• Calendar.Language: Fixed spelling for spanish septiembre.

Thanks to David Tapia <davidtapia> for the report.

Fixes [pikelang/pike #36].