On 12/29/22 10:41, Amish wrote:
On 29/12/22 20:23, Alex Rousskov wrote:
On 12/28/22 23:17, Amish wrote:
But now what?
If your Squid never peeks at origin servers (i.e. it always stares)
and your proxy never serves/secures plain-text "GET https" requests,
then you can run with the createClientContext(true) hack until
somebody volunteers a long-term solution (as discussed below).
What is plain text "GET https" request?
It is an unencrypted HTTP GET request arriving at http_port and using an
https URI scheme in its target URI. These requests are relatively rare,
mostly seen in environments where specialized HTTP clients outsource TLS
work to their forward proxies.
I do use squid as intercept proxy for http (port 80) and https (port 443
and stare bump). I hope above hack will not break that.
I share that hope, but cannot guarantee success.
There are several ways to fix this bug long-term, including these two:
Minimal: Create a TLS context object dedicated to peeking at origin
servers. It will probably have to be admin-configurable to accomodate
various TLS v1.2 (and earlier) corner cases, but we can try to start
without adding support for such configuration. Continue to use the
existing configurable context for staring and other needs but call
createClientContext(true) for that existing context.
I wish I could code it. But I have no idea about TLS and OpenSSL APIs
and also squid code.
I believe as more people switch to OpenSSL 3.0, we will see more people
(using sslbump) complain about squid not connecting to unpatched origin
servers.
I think easiest option to tackle this issue could be to have
tls_outgoing_sslopflags directive which gets applied (ORed) to all SSL
(peek, stare, splice) connections. And then those who wants squid to be
able to connect to unpatched origin server can set it to 0x4. And then
we do not need above hack. Peek and stare both will continue to work as
it is currently.
If it is acceptable solution then I may try my hands on creating PR if
it is not too complicated.
Why would adding such a new directive be better than committing the
createClientContext(true) change that you have tested? AFAICT, such a
commit would enable the old directive to do what the new directive is
meant to do. Just like a new directive, this commit will break either
peek or stare cases (depending on the problematic origin server).
The only advantage I see is that the new directive you propose would
allow to preserve the old configuration behavior, but since that
behavior is wrong/exceptional/surprising, I doubt it is worth preserving
long-term (which is what adding such a new directive would do).
I thought squid (intentionally) supported multiple
tls_outgoing_options. I had asked that question long back.
http://lists.squid-cache.org/pipermail/squid-users/2018-July/018582.html
IMO, Squid support for multiple tls_outgoing_options directives is
* poor (and probably buggy)
* unnecessary (one can always get the same result without it)
* a configuration design mistake (we should have prohibited it)
* significantly complicates fixing/enhancing the underlying code
> Hope there is no plan to remove that feature(!) in future else my
> scripts will break.
If it were my call, it would have been removed (while fixing the problem
you have reported on this thread): It is better to adjust your scripts
(a one-time pain) than suffer from this mistake long-term. It is _not_
my call, and Squid is full of other features that cause ongoing pains
and complicate progress because it is very difficult to reach consensus
that something should be _removed_.
Cheers,
Alex.
Flexible: Replace tls_outgoing_options with tls_outgoing or a similar
directive that supports ACLs. The new directive must be used once for
each set of context configuration parameters (unlike the existing
tls_outgoing_options that could be (mis)used multiple times,
accumulating some parameters and overwriting others). At configuration
time, Squid will create as many TLS context objects as there are
tls_outgoing directives. At runtime, Squid will evaluate ACLs and pick
the right context object for the current outgoing TLS connection
attempt. For example, you would be able to configure Squid to use one
context for peeking at servers and another for staring at them.
Unfortunately, even the minimal solution requires non-trivial
development work. I do not have enough free time to give you a
ready-to-use blueprint for its implementation.
https://wiki.squid-cache.org/SquidFaq/AboutSquid#how-to-add-a-new-squid-feature-enhance-of-fix-something
_______________________________________________
squid-users mailing list
squid-users@xxxxxxxxxxxxxxxxxxxxx
http://lists.squid-cache.org/listinfo/squid-users