Before this update, BIND would attempt to do a full recursive resolution
process for each query received if the requested rrset had its ttl
expired. If the resolution fails for any reason, only then BIND would
check for stale rrset in cache (if 'stale-cache-enable' and
'stale-answer-enable' is on).
The problem with this approach is that if an authoritative server is
unreachable or is failing to respond, it is very unlikely that the
problem will be fixed in the next seconds.
A better approach to improve performance in those cases, is to mark the
moment in which a resolution failed, and if new queries arrive for that
same rrset, try to respond directly from the stale cache, and do that
for a window of time configured via 'stale-refresh-time'.
Only when this interval expires we then try to do a normal refresh of
the rrset.
The logic behind this commit is as following:
- In query.c / query_gotanswer(), if the test of 'result' variable falls
to the default case, an error is assumed to have happened, and a call
to 'query_usestale()' is made to check if serving of stale rrset is
enabled in configuration.
- If serving of stale answers is enabled, a flag will be turned on in
the query context to look for stale records:
query.c:6839
qctx->client->query.dboptions |= DNS_DBFIND_STALEOK;
- A call to query_lookup() will be made again, inside it a call to
'dns_db_findext()' is made, which in turn will invoke rbdb.c /
cache_find().
- In rbtdb.c / cache_find() the important bits of this change is the
call to 'check_stale_header()', which is a function that yields true
if we should skip the stale entry, or false if we should consider it.
- In check_stale_header() we now check if the DNS_DBFIND_STALEOK option
is set, if that is the case we know that this new search for stale
records was made due to a failure in a normal resolution, so we keep
track of the time in which the failured occured in rbtdb.c:4559:
header->last_refresh_fail_ts = search->now;
- In check_stale_header(), if DNS_DBFIND_STALEOK is not set, then we
know this is a normal lookup, if the record is stale and the query
time is between last failure time + stale-refresh-time window, then
we return false so cache_find() knows it can consider this stale
rrset entry to return as a response.
The last additions are two new methods to the database interface:
- setservestale_refresh
- getservestale_refresh
Those were added so rbtdb can be aware of the value set in configuration
option, since in that level we have no access to the view object.
Created isc_refcount_decrement_expect macro to test conditionally
the return value to ensure it is in expected range. Converted
unchecked isc_refcount_decrement to use isc_refcount_decrement_expect.
Converted INSIST(isc_refcount_decrement()...) to isc_refcount_decrement_expect.
There were several problems with rbt hashtable implementation:
1. Our internal hashing function returns uint64_t value, but it was
silently truncated to unsigned int in dns_name_hash() and
dns_name_fullhash() functions. As the SipHash 2-4 higher bits are
more random, we need to use the upper half of the return value.
2. The hashtable implementation in rbt.c was using modulo to pick the
slot number for the hash table. This has several problems because
modulo is: a) slow, b) oblivious to patterns in the input data. This
could lead to very uneven distribution of the hashed data in the
hashtable. Combined with the single-linked lists we use, it could
really hog-down the lookup and removal of the nodes from the rbt
tree[a]. The Fibonacci Hashing is much better fit for the hashtable
function here. For longer description, read "Fibonacci Hashing: The
Optimization that the World Forgot"[b] or just look at the Linux
kernel. Also this will make Diego very happy :).
3. The hashtable would rehash every time the number of nodes in the rbt
tree would exceed 3 * (hashtable size). The overcommit will make the
uneven distribution in the hashtable even worse, but the main problem
lies in the rehashing - every time the database grows beyond the
limit, each subsequent rehashing will be much slower. The mitigation
here is letting the rbt know how big the cache can grown and
pre-allocate the hashtable to be big enough to actually never need to
rehash. This will consume more memory at the start, but since the
size of the hashtable is capped to `1 << 32` (e.g. 4 mio entries), it
will only consume maximum of 32GB of memory for hashtable in the
worst case (and max-cache-size would need to be set to more than
4TB). Calling the dns_db_adjusthashsize() will also cap the maximum
size of the hashtable to the pre-computed number of bits, so it won't
try to consume more gigabytes of memory than available for the
database.
FIXME: What is the average size of the rbt node that gets hashed? I
chose the pagesize (4k) as initial value to precompute the size of
the hashtable, but the value is based on feeling and not any real
data.
For future work, there are more places where we use result of the hash
value modulo some small number and that would benefit from Fibonacci
Hashing to get better distribution.
Notes:
a. A doubly linked list should be used here to speedup the removal of
the entries from the hashtable.
b. https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
Also disable the semantic patch as the code needs tweaks here and there because
some destroy functions might not destroy the object and return early if the
object is still in use.
The isc_buffer_allocate() function now cannot fail with ISC_R_MEMORY.
This commit removes all the checks on the return code using the semantic
patch from previous commit, as isc_buffer_allocate() now returns void.
The indentation for dumping the master zone was driven by two
global variables dns_master_indent and dns_master_indentstr. In
threaded mode, this becomes prone to data access races, so this commit
converts the global variables into a local per-context tuple that
consist of count and string.
This second commit uses second semantic patch to replace the calls to
dns_name_copy() with NULL as third argument where the result was stored in a
isc_result_t variable. As the dns_name_copy(..., NULL) cannot fail gracefully
when the third argument is NULL, it was just a bunch of dead code.
Couple of manual tweaks (removing dead labels and unused variables) were
manually applied on top of the semantic patch.
This commit add RUNTIME_CHECK() around all simple dns_name_copy() calls where
the third argument is NULL using the semantic patch from the previous commit.
Using isc_mem_put(mctx, ...) + isc_mem_detach(mctx) required juggling with the
local variables when mctx was part of the freed object. The isc_mem_putanddetach
function can handle this case internally, but it wasn't used everywhere. This
commit apply the semantic patching plus bit of manual work to replace all such
occurrences with proper usage of isc_mem_putanddetach().
In some cases, setting qctx->result to DNS_R_SERVFAIL causes the value
of a 'result' variable containing a more specific failure reason to be
effectively discarded. This may cause certain query error log messages
to lack specificity despite a more accurate problem cause being
determined during query processing.
In other cases, qctx->result is set to DNS_R_SERVFAIL even though a more
specific error (e.g. ISC_R_NOMEMORY) could be explicitly indicated.
Since the response message's RCODE is derived from qctx->result using
dns_result_torcode(), which handles a number of possible isc_result_t
values and returns SERVFAIL for anything not explicitly listed, it is
fine to set qctx->result to something more specific than DNS_R_SERVFAIL
(in fact, this is already being done in a few cases). Modify most
QUERY_ERROR() calls so that qctx->result is set to a more specific error
code when possible. Adjust query_error() so that statistics are still
calculated properly. Remove the RECURSE_ERROR() macro which was
introduced exactly because qctx->result could be set to DNS_R_SERVFAIL
instead of DNS_R_DUPLICATE or DNS_R_DROP, which need special handling.
Modify dns_sdlz_putrr() so that it returns DNS_R_SERVFAIL when a DLZ
driver returns invalid RDATA, in order to prevent setting RCODE to
FORMERR (which is what dns_result_torcode() translates e.g. DNS_R_SYNTAX
to) while responding authoritatively.
Replace dns_fixedname_init() calls followed by dns_fixedname_name()
calls with calls to dns_fixedname_initname() where it is possible
without affecting current behavior and/or performance.
This patch was mostly prepared using Coccinelle and the following
semantic patch:
@@
expression fixedname, name;
@@
- dns_fixedname_init(&fixedname);
...
- name = dns_fixedname_name(&fixedname);
+ name = dns_fixedname_initname(&fixedname);
The resulting set of changes was then manually reviewed to exclude false
positives and apply minor tweaks.
It is likely that more occurrences of this pattern can be refactored in
an identical way. This commit only takes care of the low-hanging fruit.
allows named to provide stale cached answers when
the authoritative server is under attack.
See max-stale-ttl, stale-answer-enable,
stale-answer-ttl. [RT #44790]
