Squid Release Notes 1.1


Ident (RFC 931) lookups

Squid will make an RFC931/ident request for client connections if 'ident_lookup' is enabled in the config file. Currently, the ident value is only logged with the request in the access.log. It is not currently possible to use the ident return value for access control purposes.

URL Redirector

Squid now has the ability to rewrite requested URLs. Implemented as an external process (similar to a dnsserver), Squid can be configured to pass every incoming URL through a 'redirector' process that returns either a new URL, or a blank line to indicate no change.

The redirector program is NOT a standard part of the Squid package. However there are a couple of user-contributed redirectors in the "contrib/" directory. Since everyone has different needs, it is up to the individual administrators to write their own implementation. For testing, and a place to start, this very simple Perl script can be used:
    print while (<>);
The redirector program must read URLs (one per line) on standard input, and write rewritten URLs or blank lines on standard output. Note that the redirector program can not use buffered I/O. Squid writes additional information after the URL which a redirector can use to make a decision. The input line consists of four fields:
    URL ip-address/fqdn ident method
The ip-address is always given, the fqdn and ident fields will be given if available, or will be "-" otherwise. Note that the ident value will only be available if 'ident_lookup' in enabled in the config file. The requestrequest method is GET, POST, etc.

Note that when used in conjunction with the -V option (on a virtual hosted machine) this provides a mechanism to use a single Squid cache as a front end to numerous servers on different machines. URLs written to the redirector will look like:
The redirector program might be this Perl script:
    while (<>) {
You may receive statistics on the redirector usage by requesting the following 'cache_object' URL:
    % client cache_object://localhost/stats/redirector

Reverse IP Lookups, client hostname ACLs.

Squid now has a address-to-hostname cache ("fqdncache") much like the name-to-address cache ("ipcache"). This means Squid can now write client hostnames in the access log, and that client domain names can be used in ACL expressions.

If you would like to log hostnames instead of addresses, enable 'log_fqdn' in your config file. This causes a reverse-lookup to be started just after the client connection has been accepted. If the reverse lookup has completed by the time the entry gets logged, the fully qualified domain name will be used, otherwise the IP address is still logged. Squid does not wait for the reverse lookup before logging the access.

A new ACL type has been added for matching client hostnames:
    acl Myusers srcdomain foo.org
The use of this ACL type may cause noticeable delay in serving objects through the cache. However, so long as allowed clients are local, the reverse lookup should not take very long and the delay may not be noticed.

Only the FQDN (i.e. the h_name field) is used for the comparison, host aliases are *not* checked.

If a reverse lookup fails, the word "none" will be used for the comparison. If you wanted to deny access to clients which did not map back to valid names, you could use
    acl BadClients srcdomain none
    http_access deny BadClients
NOTE: DNS has a number of known security problems. Squid does not make any effort to guarantee the validity of data returned from gethostbyname() or gethostbyaddr() calls.

Cache directory structure changes

The following improvements to the cache directory structure are due to Mark Treacy (

Squid-1.0 used 100 first-level directories for each 'cache_dir'. For very large caches, this meant between 5,000-10,000 files per directory, which isn't good for performance on any unix system. As well as the directory search times being slow, the amount of disk traffic due to directory operations was quite large (due to directory fragmentation (variable length filenames) each directory was about 100k in size).

To reduce the number of files per directory it was necessary to increase the number of directories used. If this was done using a single level directory structure we would have a single 'cache_dir' with an excessive number of directories in it. Hence we went to a 2 level structure. We wanted to keep each directory smaller than a filesystem block (usually 4-8k), and also wanted to be able to accommodate 1M+ objects. Assuming approximately 256 objects per directory, we settled on 16 first-level (L1) and 256 second-level (L2) directories for a total of 16x256x256 = 1,048,576 objects.

The number of L1 and L2 directories to use is configurable in the squid.conf file (swap_level1_dirs, swap_level2_dirs). To estimate the optimal numbers for your installation, we recommend the following formula:
	DS = amount of 'cache_swap' / number of 'cache_dir's
	OS = avg object size = 20k
	NO = objects per L2 directory = 256

	L1 = number of L1 directories
	L2 = number of L2 directories

such that:
	L1 x L2 = DS / OS / NO

Getting true DNS TTL info into Squid's IP cache

If you have source for BIND, you can modify it as indicated in the diff below. It causes the global variable _dns_ttl_ to be set with the TTL of the most recent lookup. Then, when you compile Squid, the configure script will look for the _dns_ttl_ symbol in libresolv.a. If found, dnsserver will return the TTL value for every lookup.

This hack was contributed by Endre Balint Nagy

diff -ru bind-4.9.4-orig/res/gethnamaddr.c bind-4.9.4/res/gethnamaddr.c
--- bind-4.9.4-orig/res/gethnamaddr.c	Mon Aug  5 02:31:35 1996
+++ bind-4.9.4/res/gethnamaddr.c	Tue Aug 27 15:33:11 1996
@@ -133,6 +133,7 @@
 } align;
 extern int h_errno;
+int _dns_ttl_;
 #ifdef DEBUG
 static void
@@ -223,6 +224,7 @@
 	host.h_addr_list = h_addr_ptrs;
 	haveanswer = 0;
 	had_error = 0;
+	_dns_ttl_ = -1;
 	while (ancount-- > 0 && cp < eom && !had_error) {
 		n = dn_expand(answer->buf, eom, cp, bp, buflen);
 		if ((n < 0) || !(*name_ok)(bp)) {
@@ -232,8 +234,11 @@
 		cp += n;			/* name */
 		type = _getshort(cp);
  		cp += INT16SZ;			/* type */
-		class = _getshort(cp);
- 		cp += INT16SZ + INT32SZ;	/* class, TTL */
+		class = _getshort(cp);  
+		cp += INT16SZ;                  /* class */
+		if (qtype == T_A  && type == T_A)
+			_dns_ttl_ = _getlong(cp);
+		cp += INT32SZ;                  /* TTL */
 		n = _getshort(cp);
 		cp += INT16SZ;			/* len */
 		if (class != C_IN) {

Using a neighbor as both a parent and a sibling

The only difference between a sibling and a parent is that cache misses are NOT fetched from siblings. In some cases it may be desirable to use a neighbor as a parent for some domains and as a sibling for others. This can now be accomplished with the 'neighbor_type_domain' configuration tag. For example:
	cache_host  parent cache.foo.org 3128 3130
	neighbor_type_domain cache.foo.org sibling .com .net
	neighbor_type_domain cache.foo.org sibling .au .de
Note that neighbor_type_domain is totally separate from the cache_host_domain option (which controls whether or not to query the neighbor). In the absence of cache_host_domain restrictions, the neighbor cache.foo.org will be queried for all requests.

If the URL host domain is .com, .net, .au, or .de then cache.foo.org is treated as a sibling (and MISSES will NOT be fetched through cache.foo.org). Otherwise it will be treated as a parent (which is the default from the cache_host line.

Forcing your neighbors to use you as a sibling

In a distributed cache hierarchy, you may need to force your peer caches to use you as a sibling and not a parent. I.e., its okay for them to fetch HITs from you, but not okay to resolve MISSes through your cache (using your resources).

This can be accomplished by using the 'miss_access' config line. The miss_access ACL list is very similar to the 'http_access' list. This functionality is implemented as a separate access list because when we check the http_access list, we don't yet know if the request will be a hit or miss. The sequence of events goes something like this: Note that in order to get to the point where miss_access is checked, the request must have also passed the http_access check.

You probably only want to use 'src' type ACL's with miss_access, although you can use any of the access control types.

If you are restricting your neighbors, be sure to allow miss_access to your local clients (e.g. users at browsers)!

Refresh Rules and If-Modified-Since

Squid 1.1 switched from a Time-To-Live based expiration model to a Refresh-Rate model. Objects are no longer purged from the cache when they expire. Instead of assigning TTL's when the object enters the cache, we now check freshness requirements when objects are requested. If an object is "fresh" it is given directly to the client. If it is "stale" then we make an If-Modified-Since request for it.

When checking the object freshness, we calculate these values:
    AGE is how much the object has aged *since* it was retrieved:

    LM_AGE is how old the object was *when* it was retrieved:


    LM_FACTOR is the ratio of AGE to LM_AGE:


    CLIENT_MAX_AGE is the (optional) maximum object age the client will
    accept as taken from the HTTP/1.1 Cache-Control request header.

    EXPIRES is the (optional) expiry time from the server reply headers.
These values are compared with the parameters of the 'refresh_pattern' rules. The refresh parameters are: The URL regular expressions are checked in the order listed until a match is found. Then this algorithm is applied for determining if an object is fresh or stale:
        if (AGE > CLIENT_MAX_AGE)
            return STALE
    if (AGE <= MIN_AGE)
        return FRESH
    if (EXPIRES) {
        if (EXPIRES <= NOW)
            return STALE
            return FRESH
    if (AGE > MAX_AGE)
        return STALE
        return FRESH
    return STALE
Note that the Max-Age in a client request takes the highest precedence. The 'MIN' value should normally be set to zero since it has higher precedence than the server's Expires: value. But if you wish to override the Expires: headers, you may use the MIN value.

Overriding neighbor refresh rules

The refresh rules also have an effect on the requests your cache makes to its neighbors. Squid uses the MAX_AGE value in the HTTP/1.1 "Cache-Control: Max-age=nnn" request header for outgoing requests. This solves the problem where neighbors with more relaxed refresh policies would send you stale objects (by your configuration).

Object Purge Policy

Squid attempts to keep the size of the disk cache relatively "smooth" or "flat." That is, objects are removed at the same rate they are added. Earlier versions had a "sawtooth" behavior where objects were removed only when disk space reached an upper limit.

Squid uses a Least-Recently-Used (LRU) replacement algorithm. Objects with large LRU age values are removed before objects with small LRU age values. We dynamically calculate the LRU age threshold, above which objects are removed. The threshold is calculated as an exponential function between the high and low water marks. When the store swap size is near the low water mark, the LRU threshold is large. This encourages more objects to be cached. When the store swap size is near the high water mark, the LRU threshold is small, encouraging more objects to be removed. The 'reference_age' configuration parameter specifies the upper limit on the LRU age threshold.

The Squid cache storage is implemented as a hash table with some number of "hash buckets." Squid scans one bucket at a time and sorts all the objects in the bucket by their LRU age. Objects with an LRU age over the threshold are removed. The scan rate is adjusted so that it takes approximately 24 hours to scan the entire cache. The store buckets are randomized so that we don't always scan the same buckets at the same time of the day.

If the store swap size somehow exceeds the high water mark, Squid performs an "emergency" object purge. We sort up to 256 objects in a store bucket and remove the eight (8) least recently used ones. This continues until the disk space is below the low water mark.

X-Forwarded-For request header

Squid used to add a request header called "Forwarded" which appeared in some early HTTP/1.1 draft documents. This header had the format
    Forwarded: by cache-host for client-address
Current HTTP/1.1 draft documents instead use the "Via" header, but it does not provide any standard way of indicating the client address in the request. Since a number of people missed having the originating client address in the request, Squid now adds its own request header called "X-Forwarded-For" which looks like this:
    X-Forwarded-For:, unknown,
Entries are always IP addresses, or the word "unknown" if the address could not be determined or if it has been disabled with the 'forwarded_for' configuration option.

We must note that access controls based on this header are extremely weak and simple to fake. Anyone may hand-enter a request with any IP address whatsoever. This is perhaps the reason why client IP addresses have been omitted from the HTTP/1.1 specification.

Using ICMP to Measure the Network

As of version 1.1.9, Squid is able to utilize ICMP Round-Trip-Time (RTT) measurements to select the optimal location to forward a cache miss. Previously, cache misses would be forwarded to the parent cache which returned the first ICP reply message. These were logged with FIRST_PARENT_MISS in the access.log file. Now we can select the parent which is closest (RTT-wise) to the origin server.

  1. Supporting ICMP in your Squid cache

    It is more important that your parent caches enable the ICMP features. If you are acting as a parent, then you may want to enable ICMP on your cache. Also, if your cache makes RTT measurements, it will fetch objects directly if your cache is closer than any of the parents.

    If you want your Squid cache to measure RTT's to origin servers, Squid must be compiled with the USE_ICMP option. This is easily accomplished by uncommenting "-DUSE_ICMP=1" in src/Makefile and src/Makefile.in.

    An external program called 'pinger' is responsible for sending and receiving ICMP packets. It must run with root privileges. After Squid has been compiled, the pinger program must be installed separately. A special Makefile target will install 'pinger' with appropriate permissions.

    % make install
    % su
    # make install-pinger

    There are three configuration file options for tuning the measurement database on your cache. 'netdb_low' and 'netdb_high' specify high and low water marks for keeping the database to a certain size (e.g. just like with the IP cache). The 'netdb_ttl' option specifies the minimum rate for pinging a site. If 'netdb_ttl' is set to 300 seconds (5 minutes) then an ICMP packet will not be sent to the same site more than once every five minutes. Note that a site is only pinged when an HTTP request for the site is received.

    Another option, 'minimum_direct_hops' can be used to try finding servers which are close to your cache. If the measured hop count to the origin server is less than or equal to minimum_direct_hops, the request will be forwarded directly to the origin server.

  2. Utilizing your parents database

    Your parent caches can be asked to include the RTT measurements in their ICP replies. To do this, you must enable 'query_icmp' in your config file:

    query_icmp on
    This causes a flag to be set in your outgoing ICP queries.

    If your parent caches return ICMP RTT measurements then the eighth column of your access.log will have lines similar to:


    In this case, it means that 'it.cache.nlanr.net' returned the lowest RTT to the origin server. If your cache measured a lower RTT than any of the parents, the request will be logged with


  3. Inspecting the database

    The measurement database can be viewed from the cachemgr by selecting "Network Probe Database." Hostnames are aggregated into /24 networks. All measurements made are averaged over time. Measurements are made to specific hosts, taken from the URLs of HTTP requests. The recv and sent fields are the number of ICMP packets sent and received. At this time they are only informational.

    A typical database entry looks something like this:

    Network          recv/sent     RTT  Hops Hostnames        20/  21    82.3   6.0 www.jisedu.org www.dozo.com 
        bo.cache.nlanr.net        42.0   7.0
        uc.cache.nlanr.net        48.0  10.0
        pb.cache.nlanr.net        55.0  10.0
        it.cache.nlanr.net       185.0  13.0

    This means we have sent 21 pings to both www.jisedu.org and www.dozo.com. The average RTT is 82.3 milliseconds. The next four lines show the measured values from our parent caches. Since 'bo.cache.nlanr.net' has the lowest RTT, it would be selected as the location to forward a request for a www.jisedu.org or www.dozo.com URL.

Planning for Squid's Memory Usage

Squid-1.1 has better memory management, although still not ideal. Squid uses memory in a variety of ways, but the bulk of memory usage falls into two categories: per-object metadata and in-transit objects.

The per-object metadata consists of a StoreEntry data structure, plus the URL for every object your cache knows about. This usually averages out to about 100 bytes per object. If you assume that the objects themselves average 20 KB each, then given your disk size ('cache_swap') you need 1/200th as much for in-memory object metadata.

The other big memory use is due to in-transit objects. The amount of memory required for this will depend on your cache's usage patterns. Obviously a more busy cache will require more memory for in-transit objects.

The 'cache_mem' parameter places a soft upper limit on the amount of memory Squid allocates for holding whole objects in VM. The 'cache_mem' memory is allocated as a pool of 4k blocks. Objects held in memory are stored in blocks from this pool. The 'cache_mem_low' and 'cache_mem_high' values affect the memory reclamation algorithm.

There are two types of in-memory objects: in-transit objects and completed objects. The in-transit objects are "locked" in memory until they are completed. The completed objects may be either normal or "negative-cached" objects.

Whenever new memory is needed for in-transit objects and current usage is above the high water mark, Squid purges some completed objects from memory. The in-memory objects are sorted by time of last use and then removed in order until memory usage is below the low water mark.

Occasionally Squid may need to exceed the maximum number of blocks. This will happen if all of the in-transit objects will not fit within the 'cache_mem' pool size. You will see this warning in your cache.log file:
    WARNING: Exceeded 'cache_mem' size (4122K > 4096K)
If this warning occurs frequently then you need to consider either increasing the 'cache_mem' value or decreasing the 'maximum_object_size' value. If the cache_mem usage is above the low water mark, then Squid will check for objects larger than 'maximum_object_size.' Any such objects are put into "delete behind" mode which means Squid releases the section of the object which has been delivered to all clients reading from it.

As a rule-of-thumb, you should probably set 'cache_mem' to 1/3 of your machine's physical memory amount. You can plan on another 1/3 being used by the per-object metadata. And the final 1/3 will be used by other data structures, unaccounted memory, and malloc() overhead. Note, this assumes that the machine will be dedicated to running Squid. If there are other services on the machine, the memory estimates should be lowered.

Default Parent

Squid has the ability to mark parent caches as the 'default' way to fetch objects. This is probably only useful with the 'no-query' option. For example, the configuration
     cache_host N1 sibling 3128 3130
     cache_host N2 sibling 3128 3130
     cache_host N3 sibling 3128 3130
     cache_host P1 parent 3128 3130 no-query default
will result in ICP queries to sibling caches N1, N2, and N3. If none of the siblings has the requested object then it will be retrieved through parent P1 due to the 'default' designation. Note that 'default' does not conflict with any 'cache_host_domain' restrictions which might be placed on a neighbor.

We do not normally recommend use of the default option. If your parent cache(s) uses ICP then you should also send them ICP queries. If your default parent is unreachable then Squid will return error messages, it will not attempt direct connections to the source.

Cachemgr Passwords

Squid-1.1 allows cachemgr passwords to be specified in the squid.conf file (instead of an /etc/passwd entry). There may be a different password for each cachemgr operation, but only one password per operation. Some sensitive operations require a password, others may be executed if no passwords are given in the squid.conf file. Operations may be disabled by setting the password to "none." See squid.conf for a full list of cachemgr operations.

Round-Robin IP

When a hostname resolves to multiple IP addresses, Squid now cycles to the next address after each connection. If a connection to an address fails, it is removed from the list. If a hostname runs out of addresses, it is removed from the IP cache.

Store Hash Configuration Options

Squid's internal hash table for holding objects has a couple of configuration options (thanks to Mark Treacy). Given the following configuration parameters:
	store_avg_object_size		# default 20K
	store_objects_per_bucket	# default 20
We first estimate the number of objects your cache can hold: Then we estimate the number of hash buckets needed: We want Squid to scan the entire hash table, one bucket at a time, over the course of about a day. We also need NUM_BUCKETS to be a prime number for optimal distribution of the hash table. NUM_BUCKETS is rounded off so that the number of buckets and maintenance rate are taken from this table:
	store_buckets	store_maintain_rate
	     7951		10 sec
	    12149		 7 sec
	    16231		 5 sec
	    33493		 2 sec
	    65357		 1 sec
If you want to increase the maintenance rate then decrease the store_objects_per_bucket parameter.

GNU malloc

Many users have reported significant performance improvements when Squid is linked with the GNU malloc library. A check for 'libgnumalloc.a' has therefore been added to the configure script. If libgnumalloc.a is found, it is automatically linked in.

GNU regex

Squid's configure script attempts to determine whether or not it should compile the GNU regex functions supplied in the source distribution. If your system appears to have its own POSIX compliant regex functions then configure may decide to use those instead of GNU regex.

Using Multicast ICP

As of Squid-1.1.6, ICP queries can be sent via multicast. Use of multicast requires the following config file entries:

  1. A cache which wants to *receive* multicast ICP queries must be configured to join a multicast address. This is done with the 'mcast_groups' directive. For example:


  2. A cache which wants to *send* multicast ICP queries must add a "multicast group" neighbor. For example:

    cache_host multicast 3128 3130 ttl=64

    In this situation, the HTTP port (3128) is ignored, but the ICP port (3130) must be correct. All multicast group members must use the same ICP port number. The 'ttl=' option specifies the IP Multicast TTL value to be used when sending a multicast datagram.

  3. Because Squid does not trust ICP replies received from unknown peers, you must specify all acceptable neighbors which might respond to your multicast query. These appear as normal parents or siblings, but with the special 'multicast-responder' option. For example:

    cache_host foo.sample.com sibling 3128 3130 multicast-responder

Use of multicast creates an interesting dilemma; normally Squid sends N queries and expects N replies. But with multicast Squid doesn't really know how many replies to expect. Somehow Squid must know roughly how many ICP replies to expect, but at the same time it must be careful to not over-estimate and therefore incur many ICP query timeouts.

The current approach is this: Squid periodically (every 15 minutes) sends fake ICP queries to only multicast peers. The replies are counted, up until the 'neighbor_timeout' time. The count is averaged over time with a fast decay so that it adjusts relatively quickly. The number of replies to expect is rounded down to the nearest whole integer to minimize the chance of suffering the neighbor timeout on real ICP queries.

Notes for running Squid under NEXTSTEP

When running Squid under NEXTSTEP 3.x, and when that NEXTSTEP system runs a BIND named process (most NEXTSTEPS handle that through netinfo and netinfo might contact a BIND named on another system) squid can trigger an error in the older BIND named that comes with NEXTSTEP 3.x. It is therefore necessary for systems running NEXTSTEP 3.x, which run their own BIND named, to run a more recent version of BIND. Luckily you don't have to compile BIND yourself, a fat (m68k i486 hppa sparc) Installer package for BIND-4.9.5 is available through ftp://ftp.nluug.nl/pub/comp/next/Internet.

NB: It might be necessary to have BIND running to run Squid under NEXTSTEP releases before NEXTSTEP 3.3+patch. Earlier releases of NEXTSTEP did not have a multithreaded netinfo resolver, which means that Squid's use of multiple dnsserver processes to prevent blocking is thwarted by netinfo blocking on every request.

Gerben Wierda <Gerben_Wierda@RnA.nl>
Ref: $Id: Release-Notes-1.1.txt,v 1997/06/13 17:00:07 wessels Exp $
(Squid 1.1.13)
Jens Elkner <elkner@irb.cs.uni-magdeburg.de>
Last modified: Sat Jul 12 05:37:51 1997