tree 71a78c4a90c4f70a4ae4eb136bd63fcb26b22d60
parent 0ba54bb40102787ce875c5d15cb0fb4193f178ec
author Tim Duesterhus <duesterhus@woltlab.com> 1624021768 +0200
committer Willy Tarreau <w@1wt.eu> 1624024100 +0200
encoding latin1

BUG/MINOR: cache: Correctly handle existing-but-empty 'accept-encoding' header

RFC 7231#5.3.4 makes a difference between a completely missing
'accept-encoding' header and an 'accept-encoding' header without any values.

This case was already correctly handled by accident, because an empty accept
encoding does not match any known encoding. However this resulted in the
'other' encoding being added to the bitmap. Usually this also succeeds in
serving cached responses, because the cached response likely has no
'content-encoding', thus matching the identity case instead of not serving the
response, due to the 'other' encoding. But it's technically not 100% correct.

Fix this by special-casing 'accept-encoding' values with a length of zero and
extend the test to check that an empty accept-encoding is correctly handled.
Due to the reasons given above the test also passes without the change in
cache.c.

Vary support was added in HAProxy 2.4. This fix should be backported to 2.4+.