Centre for Advanced Internet Architectures
Implementing AQM in FreeBSD
* Overview <http://caia.swin.edu.au/freebsd/aqm/index.html>
* Articles, Papers and Presentations
<http://caia.swin.edu.au/freebsd/aqm/papers.html>
* Patches and Tools <http://caia.swin.edu.au/freebsd/aqm/downloads.html>
Overview
Recent years have seen a resurgence of interest in better managing
the depth of bottleneck queues in routers, switches and other places
that get congested. Solutions include transport protocol enhancements
at the end-hosts (such as delay-based or hybrid congestion control
schemes) and active queue management (AQM) schemes applied within
bottleneck queues.
The notion of AQM has been around since at least the late 1990s
(e.g. RFC 2309). In recent years the proliferation of oversized
buffers in all sorts of network devices (aka bufferbloat) has
stimulated keen community interest in four new AQM schemes -- CoDel,
FQ-CoDel, PIE and FQ-PIE.
The IETF AQM working group is looking to document these schemes,
and independent implementations are a corner-stone of the IETF's
process for confirming the clarity of publicly available protocol
descriptions. While significant development work on all three schemes
has occured in the Linux kernel, there is very little in FreeBSD.
Project Goals
This project began in late 2015, and aims to design and implement
functionally-correct versions of CoDel, FQ-CoDel, PIE and FQ_PIE
in FreeBSD (with code BSD-licensed as much as practical). We have
chosen to do this as extensions to FreeBSD's ipfw/dummynet firewall
and traffic shaper. Implementation of these AQM schemes in FreeBSD
will:
* Demonstrate whether the publicly available documentation is
sufficient to enable independent, functionally equivalent implementations
* Provide a broader suite of AQM options for sections the networking
community that rely on FreeBSD platforms
Program Members:
* Rasool Al Saadi (developer)
* Grenville Armitage (project lead)
Acknowledgements:
This project has been made possible in part by a gift from the
Comcast Innovation Fund.
Submitted by: Rasool Al-Saadi <ralsaadi@swin.edu.au>
X-No objection: core
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D6388
We were inconsistent about the use of time_second vs. time_uptime.
Always use time_uptime so the value can be meaningfully compared.
Submitted by: "Max" <maximos@als.nnov.ru>
MFC after: 4 days
into dyn_update_proto_state(). This allows eliminate the second state
lookup in the ipfw_install_state().
Also remove MATCH_* macros, they are defined in ip_fw_private.h as enum.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
objects with the same name in different sets.
Add optional manage_sets() callback to objects rewriting framework.
It is intended to implement handler for moving and swapping named
object's sets. Add ipfw_obj_manage_sets() function that implements
generic sets handler. Use new callback to implement sets support for
lookup tables.
External actions objects are global and they don't support sets.
Modify eaction_findbyname() to reflect this.
ipfw(8) now may fail to move rules or sets, because some named objects
in target set may have conflicting names.
Note that ipfw_obj_ntlv type was changed, but since lookup tables
actually didn't support sets, this change is harmless.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
It allows implementing loadable kernel modules with new actions and
without needing to modify kernel headers and ipfw(8). The module
registers its action handler and keyword string, that will be used
as action name. Using generic syntax user can add rules with this
action. Also ipfw(8) can be easily modified to extend basic syntax
for external actions, that become a part base system.
Sample modules will coming soon.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
the same opcode.
o Reduce number of times classifier callback is called. It is
redundant to call it just after find_op_rw(), since the last
does call it already and can have all results.
o Do immediately opcode rewrite in the ref_opcode_object().
This eliminates additional classifier lookup later on bulk update.
For unresolved opcodes the behavior still the same, we save information
from classifier callback in the obj_idx array, then perform automatic
objects creation, then perform rewriting for opcodes using indeces
from created objects.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
When we guess the nature of the outbound packet (output vs. forwarding) we need
to take bridges into account. When bridging the input interface does not match
the output interface, but we're not forwarding. Similarly, it's possible for the
interface to actually be the bridge interface itself (and not a member interface).
PR: 202351
MFC after: 2 weeks
taskqueue_enqueue() was changed to support both fast and non-fast
taskqueues 10 years ago in r154167. It has been a compat shim ever
since. It's time for the compat shim to go.
Submitted by: Howard Su <howard0su@gmail.com>
Reviewed by: sephe
Differential Revision: https://reviews.freebsd.org/D5131
In the DIOCRSETADDRS ioctl() handler we allocate a table for struct pfr_addrs,
which is processed in pfr_set_addrs(). At the users request we also provide
feedback on the deleted addresses, by storing them after the new list
('bcopy(&ad, addr + size + i, sizeof(ad));' in pfr_set_addrs()).
This means we write outside the bounds of the buffer we've just allocated.
We need to look at pfrio_size2 instead (i.e. the size the user reserved for our
feedback). That'd allow a malicious user to specify a smaller pfrio_size2 than
pfrio_size though, in which case we'd still read outside of the allocated
buffer. Instead we allocate the largest of the two values.
Reported By: Paul J Murphy <paul@inetstat.net>
PR: 207463
MFC after: 5 days
Differential Revision: https://reviews.freebsd.org/D5426
is followed by another structure (rr_schk) whose size must be set
in the schk_datalen field of the descriptor.
Not allocating the memory may cause other memory to be overwritten
(though dn_schk is 192 bytes and rr_schk only 12 so we may be lucky
and end up in the padding after the dn_schk).
This is a merge candidate for stable and 10.3
MFC after: 3 days
in computing a shift index. The error was due to the use of mixed
fls() / __fls() functions in another implementation of qfq.
To avoid that the problem occurs again, properly document which
incarnation of the function we need.
Note that the bug only affects QFQ in FreeBSD head from last july, as
the patch was not merged to other versions.
There are number of radix consumers in kernel land (pf,ipfw,nfs,route)
with different requirements. In fact, first 3 don't have _any_ requirements
and first 2 does not use radix locking. On the other hand, routing
structure do have these requirements (rnh_gen, multipath, custom
to-be-added control plane functions, different locking).
Additionally, radix should not known anything about its consumers internals.
So, radix code now uses tiny 'struct radix_head' structure along with
internal 'struct radix_mask_head' instead of 'struct radix_node_head'.
Existing consumers still uses the same 'struct radix_node_head' with
slight modifications: they need to pass pointer to (embedded)
'struct radix_head' to all radix callbacks.
Routing code now uses new 'struct rib_head' with different locking macro:
RADIX_NODE_HEAD prefix was renamed to RIB_ (which stands for routing
information base).
New net/route_var.h header was added to hold routing subsystem internal
data. 'struct rib_head' was placed there. 'struct rtentry' will also
be moved there soon.
if more than 64 distinct values had been used.
Table value code uses internal objhash API which requires unique key
for each object. For value code, pointer to the actual value data
is used. The actual problem arises from the fact that 'actual' e.g.
runtime data is stored in array and that array is auto-growing. There is
special hook (update_tvalue() function) which is used to update the pointers
after the change. For some reason, object 'key' was not updated.
Fix this by adding update code to the update_tvalue().
Sponsored by: Yandex LLC
compiled into the kernel. Ideally lots more code would just not
be called (or compiled in) in that case but that requires a lot
more surgery. For now try to make IP-less kernels compile again.
panics when unloading the dummynet and IPFW modules:
- The callout drain function can sleep and should not be called having
a non-sleepable lock locked. Remove locks around "ipfw_dyn_uninit(0)".
- Add a new "dn_gone" variable to prevent asynchronous restart of
dummynet callouts when unloading the dummynet kernel module.
- Call "dn_reschedule()" locked so that "dn_gone" can be set and
checked atomically with regard to starting a new callout.
Reviewed by: hiren
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D3855
Vast majority of rtalloc(9) users require only basic info from
route table (e.g. "does the rtentry interface match with the interface
I have?". "what is the MTU?", "Give me the IPv4 source address to use",
etc..).
Instead of hand-rolling lookups, checking if rtentry is up, valid,
dealing with IPv6 mtu, finding "address" ifp (almost never done right),
provide easy-to-use API hiding all the complexity and returning the
needed info into small on-stack structure.
This change also helps hiding route subsystem internals (locking, direct
rtentry accesses).
Additionaly, using this API improves lookup performance since rtentry is not
locked.
(This is safe, since all the rtentry changes happens under both radix WLOCK
and rtentry WLOCK).
Sponsored by: Yandex LLC
It is called when last reference to named object is going to be released
and allows to do additional cleanup for implementation of named objects.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
new return codes of -1 were mistakenly being considered "true". Callout_stop
now returns -1 to indicate the callout had either already completed or
was not running and 0 to indicate it could not be stopped. Also update
the manual page to make it more consistent no non-zero in the callout_stop
or callout_reset descriptions.
MFC after: 1 Month with associated callout change.
r289932 accidentally broke the rule skip calculation. The address family
argument to PF_ANEQ() is now important, and because it was set to 0 the macro
always evaluated to false.
This resulted in incorrect skip values, which in turn broke the rule
evaluations.
Actually, object classify callbacks can skip some opcodes, that could
be rewritten. We will deteremine real numbed of rewritten opcodes a bit
later in this function.
Reported by: David H. Wolfskill <david at catwhisker dot org>
check_ipfw_rule_body() function. This function is intended to just
determine that rule has some opcodes that can be rewrited. Then the
ref_rule_objects() function will determine real number of rewritten
opcodes using classify callback.
Reviewed by: melifaro
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
object name correctness. Each type of object can do more strict checking
in own implementation. Do such checks for tables in check_table_name().
Reviewed by: melifaro
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
When using route-to (or reply-to) pf sends the packet directly to the output
interface. If that interface doesn't support checksum offloading the checksum
has to be calculated in software.
That was already done in the IPv4 case, but not for the IPv6 case. As a result
we'd emit packets with pseudo-header checksums (i.e. incorrect checksums).
This issue was exposed by the changes in r289316 when pf stopped performing full
checksum calculations for all packets.
Submitted by: Luoqi Chen
MFC after: 1 week
In certain configurations (mostly but not exclusively as a VM on Xen) pf
produced packets with an invalid TCP checksum.
The problem was that pf could only handle packets with a full checksum. The
FreeBSD IP stack produces TCP packets with a pseudo-header checksum (only
addresses, length and protocol).
Certain network interfaces expect to see the pseudo-header checksum, so they
end up producing packets with invalid checksums.
To fix this stop calculating the full checksum and teach pf to only update TCP
checksums if TSO is disabled or the change affects the pseudo-header checksum.
PR: 154428, 193579, 198868
Reviewed by: sbruno
MFC after: 1 week
Relnotes: yes
Sponsored by: RootBSD
Differential Revision: https://reviews.freebsd.org/D3779
Problem description:
How do we currently perform layer 2 resolution and header imposition:
For IPv4 we have the following chain:
ip_output() -> (ether|atm|whatever)_output() -> arpresolve()
Lookup is done in proper place (link-layer output routine) and it is possible
to provide cached lle data.
For IPv6 situation is more complex:
ip6_output() -> nd6_output() -> nd6_output_ifp() -> (whatever)_output() ->
nd6_storelladdr()
We have ip6_ouput() which calls nd6_output() instead of link output routine.
nd6_output() does the following:
* checks if lle exists, creates it if needed (similar to arpresolve())
* performes lle state transitions (similar to arpresolve())
* calls nd6_output_ifp() which pushes packets to link output routine along
with running SeND/MAC hooks regardless of lle state
(e.g. works as run-hooks placeholder).
After that, iface output routine like ether_output() calls nd6_storelladdr()
which performs lle lookup once again.
As a result, we perform lookup twice for each outgoing packet for most types
of interfaces. We also need to maintain runtime-checked table of 'nd6-free'
interfaces (see nd6_need_cache()).
Fix this behavior by eliminating first ND lookup. To be more specific:
* make all nd6_output() consumers use nd6_output_ifp() instead
* rename nd6_output[_slow]() to nd6_resolve_[slow]()
* convert nd6_resolve() and nd6_resolve_slow() to arpresolve() semantics,
e.g. copy L2 address to buffer instead of pushing packet towards lower
layers
* Make all nd6_storelladdr() users use nd6_resolve()
* eliminate nd6_storelladdr()
The resulting callchain is the following:
ip6_output() -> nd6_output_ifp() -> (whatever)_output() -> nd6_resolve()
Error handling:
Currently sending packet to non-existing la results in ip6_<output|forward>
-> nd6_output() -> nd6_output _lle() which returns 0.
In new scenario packet is propagated to <ether|whatever>_output() ->
nd6_resolve() which will return EWOULDBLOCK, and that result
will be converted to 0.
(And EWOULDBLOCK is actually used by IB/TOE code).
Sponsored by: Yandex LLC
Differential Revision: https://reviews.freebsd.org/D1469
If net.link.bridge.pfil_bridge is set we can end up thinking we're forwarding in
pf_test6() because the rcvif and the ifp (output interface) are different.
In that case we're bridging though, and the rcvif the the bridge member on which
the packet was received and ifp is the bridge itself.
If we'd set dir to PF_FWD we'd end up calling ip6_forward() which is incorrect.
Instead check if the rcvif is a member of the ifp bridge. (In other words, the
if_bridge is the ifp's softc). If that's the case we're not forwarding but
bridging.
PR: 202351
Reviewed by: eri
Differential Revision: https://reviews.freebsd.org/D3534
The crop/drop-ovl fragment scrub modes are not very useful and likely to confuse
users into making poor choices.
It's also a fairly large amount of complex code, so just remove the support
altogether.
Users who have 'scrub fragment crop|drop-ovl' in their pf configuration will be
implicitly converted to 'scrub fragment reassemble'.
Reviewed by: gnn, eri
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D3466
OpenBSD pf 4.5).
Fix argument ordering to memcpy as well as the size of the copy in the
(theoretical) case that pfi_buffer_cnt should be greater than ~_max.
This fix the failure when you hit the self table size and force it to be
resized.
MFC after: 3 days
Sponsored by: Rubicon Communications (Netgate)
CoDel is a parameterless queue discipline that handles variable bandwidth
and RTT.
It can be used as the single queue discipline on an interface or as a sub
discipline of existing queue disciplines such as PRIQ, CBQ, HFSC, FAIRQ.
Differential Revision: https://reviews.freebsd.org/D3272
Reviewd by: rpaulo, gnn (previous version)
Obtained from: pfSense
Sponsored by: Rubicon Communications (Netgate)
The size2 is the maximum userland buffer size (used when the addresses are
copied back to userland).
Obtained from: pfSense
MFC after: 3 days
Sponsored by: Rubicon Communications (Netgate)
Skip checks for IPv6 multicast addresses.
Use in6_localip() for global unicast.
And for IPv6 link-local addresses do search in the IPv6 addresses list.
Since LLA are stored in the kernel internal form, use
IN6_ARE_MASKED_ADDR_EQUAL() macro with lla_mask for addresses comparison.
lla_mask has zero bits in the second word, where we keep sin6_scope_id.
Obtained from: Yandex LLC
Sponsored by: Yandex LLC
When we allocate the struct pf_fragment in pf_fillup_fragment() we forgot to
initialise the fr_flags field. As a result we sometimes mistakenly thought the
fragment to not be a buffered fragment. This resulted in panics because we'd end
up freeing the pf_fragment but not removing it from V_pf_fragqueue (believing it
to be part of V_pf_cachequeue).
The next time we iterated V_pf_fragqueue we'd use a freed object and panic.
While here also fix a pf_fragment use after free in pf_normalize_ip().
pf_reassemble() frees the pf_fragment, so we can't use it any more.
PR: 201879, 201932
MFC after: 5 days
- use 1ULL to avoid shift truncations
- recompute the sum of weight dynamically to provide better fairness
- fix an erroneous constant in the computation of the slot
- preserve timestamp correctness when the old timestamp is stale.
We don't use the direction of the fragments for anything. The frc_direction
field is assigned, but never read.
Just remove it.
Differential Revision: https://reviews.freebsd.org/D2773
Approved by: philip (mentor)
When we try to look up a pf_fragment with pf_find_fragment() we compare (see
pf_frag_compare()) addresses (and family), id but also protocol. We failed to
save the protocol to the pf_fragment in pf_fragcache(), resulting in failing
reassembly.
Differential Revision: https://reviews.freebsd.org/D2772
Fix a panic when handling fragmented ip4 packets with 'drop-ovl' set.
In that scenario we take a different branch in pf_normalize_ip(), taking us to
pf_fragcache() (rather than pf_reassemble()). In pf_fragcache() we create a
pf_fragment, but do not set the address family. This leads to a panic when we
try to insert that into pf_frag_tree because pf_addr_cmp(), which is used to
compare the pf_fragments doesn't know what to do if the address family is not
set.
Simply ensure that the address family is set correctly (always AF_INET in this
path).
PR: 200330
Differential Revision: https://reviews.freebsd.org/D2769
Approved by: philip (mentor), gnn (mentor)
years for head. However, it is continuously misused as the mpsafe argument
for callout_init(9). Deprecate the flag and clean up callout_init() calls
to make them more consistent.
Differential Revision: https://reviews.freebsd.org/D2613
Reviewed by: jhb
MFC after: 2 weeks
even if there was non-zero number of restarts, we would unref/clear
all value references and start ipfw_link_table_values() once again
with (mostly) cleared "tei" buffer.
Additionally, ptei->ptv stores only to-be-added values, not existing ones.
This is a forgotten piece of previous value refconting implementation,
and now it is simply incorrect.
Currently we have tables identified by their names in userland
with internal kernel-assigned indices. This works the following way:
When userland wishes to communicate with kernel to add or change rule(s),
it makes indexed sorted array of table names
(internally ipfw_obj_ntlv entries), and refer to indices in that
array in rule manipulation.
Prior to committing new rule to the ruleset kernel
a) finds all referenced tables, bump their refcounts and change
values inside the opcodes to be real kernel indices
b) auto-creates all referenced but not existing tables and then
do a) for them.
Kernel does almost the same when exporting rules to userland:
prepares array of used tables in all rules in range, and
prepends it before the actual ruleset retaining actual in-kernel
indexes for that.
There is also special translation layer for legacy clients which is
able to provide 'real' indices for table names (basically doing atoi()).
While it is arguable that every subsystem really needs names instead of
numbers, there are several things that should be noted:
1) every non-singleton subsystem needs to store its runtime state
somewhere inside ipfw chain (and be able to get it fast)
2) we can't assume object numbers provided by humans will be dense.
Existing nat implementation (O(n) access and LIST inside chain) is a
good example.
Hence the following:
* Convert table-centric rewrite code to be more generic, callback-based
* Move most of the code from ip_fw_table.c to ip_fw_sockopt.c
* Provide abstract API to permit subsystems convert their objects
between userland string identifier and in-kernel index.
(See struct opcode_obj_rewrite) for more details
* Create another per-chain index (in next commit) shared among all subsystems
* Convert current NAT44 implementation to use new API, O(1) lookups,
shared index and names instead of numbers (in next commit).
Sponsored by: Yandex LLC