Alex Rousskov [Mon, 27 Jul 2009 01:51:24 +0000 (19:51 -0600)]
Polished r9854 "crash on shutdown" changes by removing explicit String
creation. It should be sufficient to just declare TheNullServices as
a constant C string.
When a C string constant is in use, the code becomes less efficient, but
this will be solved when we can go back to using a String class.
Alex Rousskov [Mon, 27 Jul 2009 01:41:02 +0000 (19:41 -0600)]
Bug 2723 fix: enable PURGE requests if PURGE method ACL is present.
PURGE requests were always denied, probably since 2009-06-28 (r9772) changes.
PURGE was denied because Config2.onoff.enable_purge changes done in
ACLMethodData::parse() are lost when Squid memsets Config2 to zero before
interpreting the changes. Config2 is meant for storing values _derived_ from
the primary configuration phase so it is reset after that phase is over.
This patch solves the above problem by storing ACLMethodData::parse() changes
in an ACLMethodData static member. The member is reset before
[re]configuration.
There is probably another problem with r9772 (or earlier) changes. Special
PURGE method processing is enabled whenever a PURGE ACL is detected in the
Squid configuration file, even if the ACL is unused or used in an http_access
option that does not match. This is specifically what r4363 tried to avoid in
year 2000:
users complain that defining an ACL that is never used in
an access list shouldn't trip this flag.
The patch does not solve this other problem. The right solution may be adding
a dedicated "purge" option that will have an ACL that controls what PURGE
requests, if any, are allowed to purge. That option would be in addition to
any http_access controls.
Just the basics for some needed right now. This needs to be expanded.
Also, required to break some of the mime header parsing calls out into
mime_header.cc in preparation for splitting the icon handling from the
header handling and reduce dependencies on the new tests.
Harden the sanity checks to detect negative status and other syntax issues
before they have a chance to become problems. This applies to replies and
responses both in varying ways.
Also document the sanity check logics. sanityCheck* is supposed to fill
out the error status for what it detects with each fail result.
Henrik Nordstrom [Fri, 24 Jul 2009 15:12:47 +0000 (17:12 +0200)]
Kick any pending *ufs write/close operations alive when the previous write completes
When using uufs only part of the object got written out to the disk,
forgetting to write out the last butes and closing the file.
This should have been seen at least in in diskd as well even if it
for some reason did not seem to show up in simple tests.
* only build if bootstrap works
* display last 20 lines of test result on failure
we want to see the real error plus, not just "make Error 1" messages
* use chmod to improve cleanup removal after (un)install test failures
* Squid Project lisence without Harvest addendum
* Rosetta Project requires Berkley open distribution lisence
* Squid Translators adopting Berkley open distribution lisence
for compatibility with Squid code lisence and Rosetta lisence
Alex Rousskov [Thu, 16 Jul 2009 18:40:00 +0000 (12:40 -0600)]
Bug fix: Server transaction stuck in RESPMOD ACL check if no services matched.
AccessCheck job code assumed it was running asynchronously. However, the
job was started synchronously. If the access check result was available
immediately, without any async ACL checks, then AccessCheck would call
the callback immediately, before returning from AccessCheck::Start. The
server-side code would then get confused because it uses the
AccessCheck::Start return value to change its state, and that value was
stale when AccessCheck::Start returned _after_ calling the callback.
We now start AccessCheck job properly, via AsyncStart. All checks are now
performed in a detached async state.
TODO: There are other jobs that are not started asynchronously. Fix.
Alex Rousskov [Mon, 13 Jul 2009 01:20:26 +0000 (19:20 -0600)]
Support adaptation sets and chains, including dynamic ICAP chains:
- Support adaptation service sets and chains
(adaptation_service_set and adaptation_service_chain)
- Dynamically form chains based on ICAP X-Next-Services header
(icap_service routing=on)
- Support cross-transactional ICAP header exchange
(adaptation_masterx_shared_names)
An adaptation service set contains similar, interchangeable services. No more
than one service is successfully applied. If one service is down or fails,
Squid can use another service. Think "hot standby" or "spare" ICAP servers.
Sets may seem similar to the existing "service bypass" feature, but they allow
the failed adaptation to be retried and succeed if a replacement service is
available. The services in a set may be all optional or all essential,
depending on whether ignoring the entire set is acceptable. The mixture of
optional and essential services in a set is supported, but yields results that
may be difficult for a human to anticipate or interpret. Squid warns when it
detects such a mixture.
When performing adaptations with a set, failures at a service (optional or
essential, does not matter) are retried with a different service if possible.
If there are no more replacement services left to try, the failure is treated
depending on whether the last service tried was optional or essential: Squid
either tries to ignore the failure and proceed or terminates the master
transaction.
An adaptation chain is a list of different services applied one after another,
forming an adaptation pipeline. Services in a chain may be optional or
essential. When performing adaptations, failures at an optional service are
ignored as if the service did not exist in the chain.
Request satisfaction terminates the adaptation chain.
When forming a set or chain for a given transaction, optional down services
are ignored as if they did not exist.
ICAP and eCAP services can be mixed and matched in an adaptation set or chain.
Merged from 3p1-plus branch at r9513.
* Implementation notes
The notes below focus on _changes_. Adaptation terminology and current layers
are now being documented in src/adaptation/notes.dox
Service sets and chains are implemented as ServiceGroup class kids. They are
very similar in most code aspects. The primary external difference is that
ServiceSet can "replace" a service and ServiceChain can find the "next"
service. The internal search code is implemented in ServiceGroup parent and
is parametrized by the kids.
Before the adaptation starts, Squid calculates the adaptation "plan", which is
just an iterator into the ServiceGroup. The client- and server-side adaptation
initiators used to deal with Service pointers. They now deal with ServiceGroup
pointers. The only interesting difference is that a ServiceGroup does not have
a notion of being optional or essential. Thus, if adaptation start fails, we
do not know whether the failure can be bypassed. Fortunately, starting an
adaptation does not require anything that depends on the adaptation services,
so we now simply assert that the start succeeds.
If the entire adaptation fails, the callers are notified as before. They are
told whether they can ignore the failure as before. No changes there.
A new Adaptation::Iterator class has been added to execute the adaptation
plan. That class is responsible for iterating the services in a service group
until the plan is exhausted or cannot progress due to a final failure.
Dynamically form adaptation chains based on the ICAP X-Next-Services header.
If an ICAP service with the routing=1 option in squid.conf returns an ICAP
X-Next-Services response header during a successful REQMOD or RESPMOD
transaction, Squid abandons the original adaptation plan and forms a new
adaptation chain consisting of services identified in the X-Next-Services
header value (using a comma-separated list of adaptation service names from
squid.conf). The dynamically created chain is destroyed once the new plan is
completed or replaced.
This feature is useful when a custom adaptation service knows which other
services are applicable to the message being adapted.
Limit adaptation iterations to adaptation_service_iteration_limit to protect
Squid from infinite adaptation loops caused by ICAP services constantly
including themselves in the dynamic adaptation chain they request. When the
limit is exceeded, the master transaction fails. The default limit of 16
should be large enough to not require an explicit configuration in most
environments yet may be small enough to limit side-effects of loops.
TODO: Add metadata support to eCAP API and honor X-Next-Services there as
well. Currently, only ICAP services can form dynamic chains but the formed
chains may contain eCAP services.
Other improvements:
Polished adaptation service configuration in squid.conf. Old format with an
anonymous bypass option is deprecated but still supported. Quit with a fatal
message if an adaptation service is misconfigured (debugging level-0 messages
do not seem to work at that stage, but that is probably another, general bug).
Polished HttpRequest::adaptHistory() interface so that the code that knows the
history is needed can force history creation without complex
configuration-time preparations and state. Currently, all adaptation history
users but the logging-related ones know runtime whether the history must be
created (e.g., when a certain ICAP header is received).
Fixed "canonical" Request URL maintenance when ICAP clones requests.
TODO: The urlCanonical() must become HttpRequest::canonical(), hiding the
often out-of-sync canonical data member.
Fixed ICAP request parsing (for ICAP logging). We used to parse Request-Line
as if it were the first header. TODO: optimize by parsing only when needed.
Fixed AccessCheck case where a service group disappears during a nb ACL check.
Replaced "done" member with an existing AsyncJob mustStop mechanism. Removed
extra async call as unneeded because ACL callbacks are already async.
Author: Henrik Nordstrom <henrik@henriknordstrom.net>
http_port allow-direct option to allow direct forwarding in accelerator mode
normally direct forwarding is disabled in accelerator mode unless overridden
by always_direct, to avoid unintentional security loops. But there is setups
where it makes sense to not have this restriction as this has effects on
peer selection as well.
Alex Rousskov [Sat, 11 Jul 2009 05:39:44 +0000 (23:39 -0600)]
Limit X-Forwarded-For growth.
X-Forwarded-For growth leads to String size limit assertions and probably
other problems.
We now replace huge XFF values with a string "error", warn the admin the
first 100 times, and hope that something will stop the loop (if it is a
loop). TODO: we should probably deny requests with huge XFF.
To make growth-associated problems visible during forwarding loops, the
loop breaking code must be disabled (no Via) or not applicable (direct
forwarding) and request_header_max_size has to be raised or disabled.
The X-Forwarded-For header value may also grow too large for reasons
unrelated to forwarding loops.
This change also prevents most cases of pointless computation of the
original X-Forwarded-For value list. That computation can be quite
expensive.
Shuffle cachemgr.cgi.8 into tools/ with cachemgr.cgi code
To perform this shuffle the SUBSTITUTE make code pushed into a sub-include
at doc/manuals/Substitute.am for shared use.
Also clean up the manual titles to reflect the binary they apply to
rather than the generic Squid project release name.
The use of generic @SYSCONFDIR@ and specific filenames allows the
translation strings to be more specific, readable for translators and
reduces Makefile complexity at once. squid.conf etc are widely mentioned
and used enough that allowing for alternative names at this point seems
meaningless.
Attempts to translate, and installs the error pages content if present.
Both srcdir and builddir are scanned used since the file location varies
depending on whether a source bundle or VCS is used to install.
Alex Rousskov [Thu, 2 Jul 2009 16:36:36 +0000 (10:36 -0600)]
Truncate too-long HTTP response bodies to match their Content-Length header.
Sometimes a broken server sends more than Content-Length bytes in the
response. For example, a 302 redirect message with "Content-Length: 0" header
may include an HTML body. Squid used to send "everything" it read to the
client, even if it read more than the Content-Length bytes. That may have
helped in some cases, but we should be more conservative when dealing
with broken servers to combat message smuggling attacks and other bad
side-effects for clients.
We now do not forward more than the advertised content length and declare the
connection with a broken server non-persistent.
Chunked responses (that HTTP/1.0 Squid should not receive and that must not
have a Content-Length header) are not truncated because RFC 2616 says we
MUST ignore their Content-Length header.
TODO: Do not cache the truncated entry and purge the cached version, if any.
Alex Rousskov [Thu, 2 Jul 2009 15:24:24 +0000 (09:24 -0600)]
Break forwarding loops for "transparent" or "intercept" http_ports.
Squid detected forwarding loops in most configurations, but broke
them (using a customizable HTTP_FORBIDDEN response) only when working as
an accelerator. Squid now breaks loops when working as a transparent
proxy as well.
A persistent loop is going to be broken anyway, when the Via and
X-Forwarded-For headers exceed header size limit, but that wastes a lot of
resources and may also crash misconfigured Squids.
TODO: Consider breaking all loops, regardless of the http_port options.
TODO: Consider adding a specific and/or configurable error page for this case
instead of using hard-coded ACCESS_DENIED.
projects / thirdparty / squid.git / log
