A reused TCP/TLS dispatch with no outstanding responses was left in the
reuse pool with no read pending, so a peer closing the idle connection
went unnoticed: the socket lingered in CLOSE-WAIT and the dead dispatch
was later handed to a new query, which failed and the fetch timed out.
Keep a read pending on an idle connected dispatch, bounded by an idle
timeout, so the close is seen promptly and the connection is dropped
from the pool instead of reused.
The idle read may only be (re)armed while the dispatch is still
connected; arming it on a dispatch that is already shutting down
re-reads a dying handle and double-schedules a netmgr job.
On shutdown, close the connection as soon as the dispatch reaches its
terminal state instead of waiting for the last reference to drop, so an
unexpected read (or a peer-side close) cannot leave the socket in
CLOSE-WAIT while a reference still lingers.
With the advent of the new development model involving security-*
branches and autorebasing, the value added by the pre-release testing
mechanism dropped drastically. The only remaining benefit of
pre-release testing is flagging in-progress security fixes targeting
open source branches that conflict with the corresponding bind-9.x-sub
branches. However, such conflicts are a rare occurrence and can be
handled after merging anyway.
Remove the Danger check related to pre-release testing.
Merge branch 'michal/drop-danger-check-related-to-pre-release-testing' into 'main'
See merge request isc-projects/bind9!12313
With the advent of the new development model involving security-*
branches and autorebasing, the value added by the pre-release testing
mechanism dropped drastically. The only remaining benefit of
pre-release testing is flagging in-progress security fixes targeting
open source branches that conflict with the corresponding bind-9.x-sub
branches. However, such conflicts are a rare occurrence and can be
handled after merging anyway.
Remove the Danger check related to pre-release testing.
RPM build jobs work by pushing a commits to main in a repository. Each
repository is identified by the `SERVICE variable and, since race
conditions may happen, the jobs are run serially via GitLab's
resource_group mechanism.
Merge branch 'andoni/use-resource-group-for-rpms-copr-ci-job' into 'main'
See merge request isc-projects/bind9!12295
RPM build jobs push commits to Git repositories. If multiple such jobs
are triggered simultaneously, some of these pushes may fail due to the
same Git branch getting updated by one job while another one attempts to
do the same thing in parallel. Use GitLab's resource group mechanism to
prevent such races: group jobs by the Git repository they push to, which
is indicated by the $SERVICE variable set for each job.
When an upstream server returned a truncated reply to a query that BIND had
signed with TSIG, the resolver could keep waiting for a follow-up UDP packet
that never arrived, so the query stalled until it hit resolver-query-timeout
and the client received no answer. BIND now treats any reply it cannot
authenticate as an immediate failure and returns SERVFAIL right away as a
defense in depth.
Closes#6028
Merge branch '6028-tsig-truncated-tsig-response' into 'main'
See merge request isc-projects/bind9!12080
A response that failed the signature check with a missing or unexpected
TSIG used to set nextitem, so the resolver kept reading the dispatch for
another response. When the response was truncated with the TSIG cut off
the end of the wire, no further response ever arrived and the fetch
stalled until resolver-query-timeout.
Treat an unauthenticated response like every other signature-check
failure and finish the fetch immediately. A response carrying a missing
or bogus TSIG now yields SERVFAIL instead of being skipped in favour of
a later one; the cookie system test that fed a spoofed TSIG response is
updated to expect that. The unauthenticated data is still never
returned.
render_xsl() served the static XSL stylesheet by casting away the const
qualifier of xslmsg and handing the pointer to isc_buffer_reinit():
p = UNCONST(xslmsg);
isc_buffer_reinit(b, p, strlen(xslmsg));
isc_buffer_reinit() copies any pre-existing buffer content into the new
base with memmove(), so the call would write into xslmsg, which is a
'const char[]' living in read-only memory. This is safe today only
because the supplied bodybuffer is always freshly initialized with
length 0, so the memmove() never runs -- a fragile, action-at-a-distance
invariant that GCC's -fanalyzer flags as a write to a const object
(-Wanalyzer-write-to-const).
Use isc_buffer_constinit(), the primitive intended for pointing a buffer
at constant data: it goes through isc_buffer_init() and never writes to
the base. This drops the UNCONST cast, keeps xslmsg in read-only
memory, and silences the analyzer warning.
Assisted-by: Claude:claude-opus-4-8
Merge branch 'mnowak/render-xsl-const-buffer' into 'main'
See merge request isc-projects/bind9!12168
render_xsl() cast away the const of the static xslmsg stylesheet and
passed it to isc_buffer_reinit(), which memmove()s into the base when
the buffer is non-empty -- a write to read-only memory that -fanalyzer
flags (-Wanalyzer-write-to-const). It only stayed safe because the
body buffer is always empty here.
Use isc_buffer_constinit(), which never writes to the base, and assert
the empty-buffer contract with REQUIRE(isc_buffer_length(b) == 0).
Assisted-by: Claude:claude-opus-4-8
- Remove the "autorebase-merge-request" CI job
- Fix formatting of autorebase failure notifications
- Send Zulip notifications for autorebase failures
Merge branch 'michal/autorebase-ci-tweaks' into 'main'
See merge request isc-projects/bind9!12288
Use the GitLab-to-Zulip username map available in the BIND 9 QA
repository to determine the Zulip username of the developer who happened
to author a breaking base branch change, so that a Zulip notification
can be triggered for that developer.
Using a plain "echo" command does not turn "\n" into newline
characters, breaking the formatting of the Zulip notifications sent upon
autorebase failures. Fix by using "echo -e" instead, which enables
interpreting backslash sequences in the provided input.
The "autorebase-merge-request" CI job is not useful in practice as the
Danger check whose complaints it was supposed to address does not look
for commit hashes in the original merge request's target branch, but
rather in the original merge request itself - and those commit hashes
remain stable over time. Furthermore, after a backport gets merged, any
cherry-pick references its commits might contain will be maintained by
other autorebasing jobs. Given the above, remove the
"autorebase-merge-request" CI job as it serves no useful purpose.
When a cached name held both a CNAME and records of another type — one stale,
the other still fresh — named with serve-stale could return the expired set
instead of the fresh one, in either direction. It now prefers whichever is fresh.
Merge branch 'ondrej/fix-serve_stale-cname-and-type' into 'main'
See merge request isc-projects/bind9!12282
With serve-stale enabled, stale rdataset headers are kept at a node so
they can be served as a last resort. The find loop, however, accepted a
stale CNAME or stale record of the requested type as a final answer and
broke out of the iteration early, returning stale data even when a fresh
header for the same name appeared later in the scan. Treat STALE(found)
like a missing answer so the loop keeps looking and only falls back to
the stale header when no fresh answer is found.
In rare cases the :iscman:`named` process could terminate unexpectedly
when processing authorized DNS UPDATE messages in quick procession
which are updating a zone with inline-signing enabled. This has been
fixed.
Closes#5816
Merge branch '5816-inline-signing-concurrent-update-fix' into 'main'
See merge request isc-projects/bind9!11982
The dns_update_signaturesinc() updates zone signatures in chunks,
keeping its current state in 'zone->rss_state'. When a zone shuts
down, the signature update process is canceled, and all the data
in the state is not freed.
Create a new dns_update_state_clear() function which can be called
from dns_zone_free() to free the memory.
When a DNS UPDATE messages is received, the zone_send_secureserial()
function can schedule a new receive_secure_serial() call with a new
'rss' object before a previous one had a chance to be fully processed.
This can cause an assertion failure in receive_secure_serial() with a
new 'rss' object (when the old one was rescheduled because it got
DNS_R_CONTINUE from dns_update_signaturesinc()). In other words:
1. receive_secure_serial() called with rss, sets zone->rss = rss,
reschedules because of DNS_R_CONTINUE
2. receive_secure_serial() called with rss_new, INSIST fails because
zone->rss != rss_new), i.e. this was called before the old 'rss'
was fully processed
Change the code logic by introducing a new 'rss_next' field and making
sure the old 'rss' is complete before starting processing the new
one.
This new check floods the server with DNS UPDATE messages for an
'inline-signing yes; sig-signing-signatures 1;' zone to see if
it manages to process the updates correctly.
Small cleanup. Found with ninja -C build-dir/ scan-build.
Patch submitted by Tim Rühsen.
Closes#6163
Merge branch '6163-double-initialization-copy-tuple' into 'main'
See merge request isc-projects/bind9!12292
Internal refactoring of the cache database (qpcache) with no functional
change. The slab headers that hold cached rdatasets are now reference
counted and own their memory context and node reference directly, so a
header can outlive the cleaning of its node and be reclaimed
independently of it. Building on that, the per-type slabtop container is
folded into the slab header itself, removing a level of indirection and
one allocation per cached type.
Merge branch 'ondrej/slabheader-reference-counting' into 'main'
See merge request isc-projects/bind9!12285
When the cache is over its size limit, qpcache_miss() runs LRU eviction
while add() is still in progress. Eviction removes a header together
with its RRSIG/covered 'related' partner, so it could free a header the
add still needs -- the new header, the partner it was just paired with,
or the one about to be displaced for max-types-per-name. With 'related'
now a counted reference, that became a use-after-free reachable under
sustained load.
Run the eviction last, after the new header is linked, bound and any
over-limit header removed, and skip the new header and its partner in
the eviction loop (the partner is marked visited so the SIEVE hand still
advances). Per-header removal is factored into header_delete(),
expire_header() and flush_node().
Cache headers are now unlinked from their node as soon as they expire,
so a slabheader is never left in the "ancient, awaiting cleanup" state
that DNS_SLABHEADERATTR_ANCIENT tracked. Drop the attribute, rename
mark_ancient() to header_delete() to reflect that it now removes the
header rather than flagging it (keying idempotency on list membership),
and remove the ancient RRset statistics counter that recorded the state,
which is now always zero; the rdataset statistics array shrinks to
match.
The rdataset-level 'ancient' flag and 'rndc dumpdb -expired' are
unaffected: expiry is derived from the entry's TTL when the rdataset is
bound, not from the slabheader attribute.
qpcache_find, qpcache_findrdataset, and find_headers each repeated the
same logic to pick a header and its RRSIG out of a slabheader and its
'related' link. Pull it into store_headers(), and rename check_header()
to invalid_header() since it returns true when the header should be
skipped.
With headers removed eagerly in mark_ancient there is at most one header
per type at a node, so the separate per-type dns_slabtop container no
longer earns its keep. Fold its fields onto the header -- the link into
the node's list, the RRSIG/covered related pairing, and the SIEVE-LRU
state -- and link headers directly into the node, dropping a level of
indirection and an allocation per cached type.
Now that a bound rdataset keeps references to both the slabheader and
the node, an ancient header (and the slabtop it leaves empty) can be
removed from the cache immediately in mark_ancient, instead of being
parked on a per-node dirty list and reclaimed only once the node becomes
unreferenced. This drops the dirty list and the clean_cache_* machinery
entirely.
Because the cache structure can now change under a node that still has
external references, the all-rdatasets iterator no longer walks the live
slabtop list: allrdatasets() binds every matching rdataset up front and
the iterator works from that snapshot, cloning out each entry.
Now that the slabheader carries its own memory context, free its
noqname/closest proofs from slabheader_destroy rather than reaching
through the owning node's deletedata method. That was the method's last
caller, so remove dns_db_deletedata entirely; the cache bookkeeping it
performed (rrset statistics and the dirty list) becomes a plain helper
called wherever a header leaves the cache.
The bound rdataset now keeps its own node reference instead of reaching
it through the slabheader, and each slabheader carries its own memory
context so it can free itself once its reference count reaches zero.
The noqname/closest proof rdatasets are views into a slabheader's proof
slabs, so they now hold a reference to that slabheader too, keeping the
proof slab (and the cloned owner name): alive for as long as the proof
rdataset is. Keeping the node on the rdataset means node access stays
valid for the life of the rdataset, independent of the slabheader's own
node pointer.
Instead of only reference counting the enclosing qpcnode, add the
reference counting directly to the slabheaders. This will allow us to
simplify the data model, as the slabheaders reference is now incremented
when we bind the rdataset and decremented when we disassociate the
rdataset. This will allow us to remove the slabheader from the slabtop
directly instead of waiting for qpcnode to become completely
unreferenced.
Follow the method split with a data split: give the proof
pseudo-rdataset its own member in the dns_rdataset union, carrying its
own node, instead of overloading the slab member. The proof methods
now read proof.raw/proof.node independently of the real
slab.raw/slab.noqname/slab.closest, so the two share no state and the
proof path is fully self-contained.
The noqname/closest proof pseudo-rdatasets share the slab data layout
but have a different lifecycle: they reference a bare proof slab owned
by the parent header and are torn down via dns_slabheader_freeproof(),
not the rdataset's own teardown. Give them a dedicated
dns_rdataslab_proof_rdatasetmethods table with their own
disassociate/clone so the proof path can diverge from the real
bound-rdataset path -- a prerequisite for reference counting the real
headers without it ever reaching the proof slabs.
With serve-stale enabled, stale rdataset headers are kept at a node so
they can be served as a last resort. The find loop, however, accepted a
stale CNAME or stale record of the requested type as a final answer and
broke out of the iteration early, returning stale data even when a fresh
header for the same name appeared later in the scan. Treat STALE(found)
like a missing answer so the loop keeps looking and only falls back to
the stale header when no fresh answer is found.
Callbacks for libuv stream reads do not signal zero-length reads as a
failure signal but rather as EAGAIN/EWOULDBLOCK. This can trigger an
assertion when a zero-length read is pushed onto a PROXYv2 endpoint that
has not yet processed the headers as it expects a non-NULL region of
positive length.
Closes#6140
Merge branch '6140-proxyv2-assertion-on-zero-byte-read' into 'main'
See merge request isc-projects/bind9!12261
Callbacks for libuv stream reads do not signal zero-length reads as a
failure signal but rather as EAGAIN/EWOULDBLOCK. This can trigger an
assertion when a zero-length read is pushed onto a PROXYv2 endpoint that
has not yet processed the headers as it expects a non-NULL region of
positive length.
The third-party comparison is right at the 0.4 % threshold, making it
fail quite often. The current range of observed values ranges from
0.3-0.5 %. Raise the threshold to 0.5 %.
The recent-named comparison produces values ranging from 0.05 to 0.12 %,
but appears to be more sensitive to time of the day when the test runs.
Raise the threshold to 0.15 %.
Both results are stable within the specified ranges across the last
three releases in both main and bind-9.20 series.
Merge branch 'nicki/respdiff-threshold-bump' into 'main'
See merge request isc-projects/bind9!12286
The third-party comparison is right at the 0.4 % threshold, making it
fail quite often. The current range of observed values ranges from
0.3-0.5 %. Raise the threshold to 0.5 %.
The recent-named comparison produces values ranging from 0.05 to 0.12 %,
but appears to be more sensitive to time of the day when the test runs.
Raise the threshold to 0.15 %.
Both results are stable within the specified ranges across the last
three releases in both main and bind-9.20 series.
When on an ``rndc reconfig`` the DNSSEC policy changes such that it changes the expected ``CDNSKEY`` and/or ``CDS`` records in the zone, the RRset should
be updated accordingly. This did not happen when removing digests from the configuration, or setting `cdnskey no;`. This has been fixed.
Closes#6166
Merge branch '6166-reconfig-delete-cds' into 'main'
See merge request isc-projects/bind9!12265
With dnssec-signzone smart-signing (-S), the CDS and CDSNKEY are
derived from the key timing metadata and the configuration options (-G).
The test has specific test cases that smart signing (with the fix)
interferes with. Therefor, disable smart-signing in the cds system test.
When adding to dnssec-policy:
cdnskey no;
cds-digest-types { };
and then reconfig the server, named must remove existing CDS and CDNSKEY
records. Note this already worked when adding CDS digest, or setting
'cdnskey yes;', but not when digests were removed from the list, or
when setting 'cdnskey no;'.
When on an 'rndc reconfig' the DNSSEC policy changes such that it
changes the expected CDNSKEY/CDS records in the zone, the RRset should
be updated accordingly.
Add a test case where we reconfigure a zone with a policy such that
these records should be removed, and on a second reconfigure add
them back again.
Note the test deliberately adds a different CDS digest on the
second reconfigure.
Some of the tests in `dnssec/tests_validation.py` worked by iterating through the response message looking for failure conditions, such as excessively high TTL values. In some cases, previous changes caused additional data not to be returned. Since there was nothing to iterate, the tests still "passed".
Tests that don't make sense anymore have been removed. Other tests that iterate through responses have been updated with checks to ensure that the responses actually do contain data.
Merge branch 'each-cleanup-validation-test' into 'main'
See merge request isc-projects/bind9!12269
Some of the tests in in dnssec/tests_validation.py worked by iterating
through the response message looking for failure conditions, such as
excessively high TTL values. In some cases, previous changes caused
additional data not to be returned. Since there was nothing to
iterate, the tests still "passed".
Tests that don't make sense anymore have been removed. Other tests that
iterate through responses have been updated with checks to ensure that
the responses actually do contain data.
expand on the isctest.check.empty_answer() function, adding
empty_authority(), empty_additional(), has_answer(), has_authority(),
and has_additional().
A positive wildcard answer, and the NSEC3 proof that the requested
name doesn't exist in the zone, must both be from the same zone.
Otherwise, an NSEC3 from an ancestor zone could be used to interfere
with validation.
We now retrieve the signer name from a wildcard response's signature.
An NSEC3 record cannot be used as a NOQNAME proof for the
wildcard unless it exactly matches the name one level above the NSEC3.
Closes#5971
Merge branch '5971-wildcard-noqname-mismatch' into 'main'
See merge request isc-projects/bind9!12256
A positive wildcard answer, and the NSEC3 proof that the requested
name doesn't exist in the zone, must both be from the same zone.
Otherwise, an NSEC3 from an ancestor zone could be used to interfere
with validation.
We now retrieve the signer name from a wildcard response's signature.
An NSEC3 record cannot be used as a NOQNAME proof for the wildcard
unless it exactly matches the name one level above the NSEC3.
Fixes: isc-projects/bind9#5971
