Remove the complicated mechanism that could be (in theory) used by
external libraries to register new categories and modules with
statically defined lists in <isc/log.h>. This is similar to what we
have done for <isc/result.h> result codes. All the libraries are now
internal to BIND 9, so we don't need to provide a mechanism to register
extra categories and modules.
The log message for commit 24381cc36d8528f5a4046fb2614451aeac4cdfc1
explained:
In some older BIND 9 branches, the extra queuing overhead eliminated by
this change could be remotely exploited to cause excessive memory use.
Due to architectural shift, this branch is not vulnerable to that issue,
but applying the fix to the latter is nevertheless deemed prudent for
consistency and to make the code future-proof.
However, it turned out that having a single queue for the nodes to be
pruned increased lock contention to a level where cleaning up nodes from
the RBTDB took too long, causing the amount of memory used by the cache
to grow indefinitely over time.
This commit reverts the change to the pruning mechanism introduced by
commit 24381cc36d8528f5a4046fb2614451aeac4cdfc1 as BIND branches newer
than 9.16 were not affected by the excessive event queueing overhead
issue mentioned in the log message for the above commit.
- create_node() in rbt.c cannot fail
- the dns_rbt_*name() functions, which are wrappers around
dns_rbt_[add|find|delete]node(), were never used except in tests.
this change isn't really necessary since RBT is likely to go away
eventually anyway. but keeping the API as simple as possible while it
persists is a good thing, and may reduce confusion while QPDB is being
developed from RBTDB code.
these values pertain to whether a node is in the main, nsec, or nsec3
tree of an RBTDB. they need to be moved to a more generic location so
they can also be used by QPDB.
(this is in db.h rather than db_p.h because rbt.c needs access to it.
technically, that's a layer violation, but it's a long-existing one;
refactoring to get rid of it would be a large hassle, and eventually
we expect to remove rbt.c anyway.)
Instead of issuing a separate isc_async_run() call for every RBTDB node
that triggers tree pruning, maintain a list of nodes from which tree
pruning can be started from and only issue an isc_async_run() call if
pruning has not yet been triggered by another RBTDB node.
In some older BIND 9 branches, the extra queuing overhead eliminated by
this change could be remotely exploited to cause excessive memory use.
Due to architectural shift, this branch is not vulnerable to that issue,
but applying the fix to the latter is nevertheless deemed prudent for
consistency and to make the code future-proof.
Clang Static Analyzer is unable to grasp that when dns_rbt_addnode()
returns ISC_R_EXISTS, it always sets the pointer passed to it via its
'nodep' parameter to a non-NULL value. Add an extra safety check in the
conditional expression used in dns_rbt_addname() to silence that
warning.
As we are in overmem state we want to free more memory than we are
adding so we need to add in an allowance for the rbtnodes that may
have been added and the names stored with them. There is the node
for the owner name and a possible ENT node if there was a node split.
1. Replace the "high-performance" macros that were only used if
DNS_NAME_USEINLINE was defined before including <dns/name.h> with
inline header functions with assertion checks and thus use them
everywhere.
2. Replace the old struct initializers with C99 designated initializers
for better understanding what is happening in these macros.
Use the new isc_mem_c*() calloc-like API for allocations that are
zeroed.
In turn, this also fixes couple of incorrect usage of the ISC_MEM_ZERO
for structures that need to be zeroed explicitly.
There are few places where isc_mem_cput() is used on structures with a
flexible member (or similar).
The only place where dns_name_hash() was being used is the old hash
table in the dns_badcache unit. Squash the dns_name_fullhash() and
dns_name_hash() into single dns_name_hash() function that's always
case-insensitive as it doesn't make to do case-sensitive hashing of the
domain names and we were not using this anywhere.
This implements node reference tracing that passes all the internal
layers from dns_db API (and friends) to increment_reference() and
decrement_reference().
It can be enabled by #defining DNS_DB_NODETRACE in <dns/trace.h> header.
The output then looks like this:
incr:node:check_address_records:rootns.c:409:0x7f67f5a55a40->references = 1
decr:node:check_address_records:rootns.c:449:0x7f67f5a55a40->references = 0
incr:nodelock:check_address_records:rootns.c:409:0x7f67f5a55a40:0x7f68304d7040->references = 1
decr:nodelock:check_address_records:rootns.c:449:0x7f67f5a55a40:0x7f68304d7040->references = 0
There's associated python script to find the missing detach located at:
https://gitlab.isc.org/isc-projects/bind9/-/snippets/1038
* rbt node chains were sized to allow for bitstring labels, so they
had 256 levels; but in the absence of bistrings, 128 is enough.
* dns_byaddr_createptrname() had a redundant options argument,
and a very outdated doc comment.
* A number of comments referred to bitstring labels in a way that is
no longer helpful. (A few informative comments remain.)
sizeof(dns_name_t) did not change but the boolean attributes are now
separated as one-bit structure members. This allows debuggers to
pretty-print dns_name_t attributes without any special hacks, plus we
got rid of manual bit manipulation code.
Originally RBT node stored three lowest bits from dns_name_t attributes.
This had a curious side-effect noticed by Tony Finch:
If you create an rbt node from a DYNAMIC name then the flag will be
propagated through dns_rbt_namefromnode() ... if you subsequently call
dns_name_free() it will try to isc_mem_put() a piece of an rbt node ...
but dns_name_free() REQUIRE()s that the name is dynamic so in the usual
case where rbt nodes are created from non-dynamic names, this kind of
code will fail an assertion.
This is a bug it dates back to june 1999 when NAMEATTR_DYNAMIC was
invented.
Apparently it does not happen often :-)
I'm planning to get rid of DNS_NAMEATTR_ definitions and bit operations,
so removal of this "three-bit-subset" assignment is a first step.
We can keep only the ABSOLUTE flag in RBT node and nothing else because
names attached to rbt nodes are always readonly: The internal node_name()
function always sets the NAMEATTR_READONLY when making a dns_name that
refers to the node's name, so the READONLY flag will be set in the name
returned by dns_rbt_namefromnode().
Co-authored-by: Tony Finch <fanf@isc.org>
Add new semantic patch to replace the straightfoward uses of:
ptr = isc_mem_{get,allocate}(..., size);
memset(ptr, 0, size);
with the new API call:
ptr = isc_mem_{get,allocate}x(..., size, ISC_MEM_ZERO);
The rbtnode->rpz flag was left behind when rbt and rpz were disentangled
by CHANGES #4576. Removing it makes the comment above correct again.
This reduces the flags so they fit in a 32 bit word again. On 64
bit systems there is still padding so it doesn't change the size
of an rbtnode. On 32 bit systems it reduces an rbtnode by 4 bytes.
Fibonacci hashing was implemented in four separate places (rbt.c,
rbtdb.c, resolver.c, zone.c). This commit combines them into a single
implementation. The hash_32() function is now replaced with
isc_hash_bits32().
Historically, the inline keyword was a strong suggestion to the compiler
that it should inline the function marked inline. As compilers became
better at optimising, this functionality has receded, and using inline
as a suggestion to inline a function is obsolete. The compiler will
happily ignore it and inline something else entirely if it finds that's
a better optimisation.
Therefore, remove all the occurences of the inline keyword with static
functions inside single compilation unit and leave the decision whether
to inline a function or not entirely on the compiler
NOTE: We keep the usage the inline keyword when the purpose is to change
the linkage behaviour.
Previously, the unreachable code paths would have to be tagged with:
INSIST(0);
ISC_UNREACHABLE();
There was also older parts of the code that used comment annotation:
/* NOTREACHED */
Unify the handling of unreachable code paths to just use:
UNREACHABLE();
The UNREACHABLE() macro now asserts when reached and also uses
__builtin_unreachable(); when such builtin is available in the compiler.
This commit converts the license handling to adhere to the REUSE
specification. It specifically:
1. Adds used licnses to LICENSES/ directory
2. Add "isc" template for adding the copyright boilerplate
3. Changes all source files to include copyright and SPDX license
header, this includes all the C sources, documentation, zone files,
configuration files. There are notes in the doc/dev/copyrights file
on how to add correct headers to the new files.
4. Handle the rest that can't be modified via .reuse/dep5 file. The
binary (or otherwise unmodifiable) files could have license places
next to them in <foo>.license file, but this would lead to cluttered
repository and most of the files handled in the .reuse/dep5 file are
system test files.
The __builtin_expect() can be used to provide the compiler with branch
prediction information. The Gcc manual says[1] on the subject:
In general, you should prefer to use actual profile feedback for
this (-fprofile-arcs), as programmers are notoriously bad at
predicting how their programs actually perform.
Stop using __builtin_expect() and ISC_LIKELY() and ISC_UNLIKELY() macros
to provide the branch prediction information as the performance testing
shows that named performs better when the __builtin_expect() is not
being used.
1. https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html#index-_005f_005fbuiltin_005fexpect
Originally, the hash table used in RBT database would be resized when it
reached certain number of elements (defined by overcommit). This was
causing resolution brownouts for busy resolvers, because the rehashing
could take several seconds to complete. This was mitigated by
pre-allocating the hash table in the RBT database used for caching to be
large-enough as determined by max-cache-size. The downside of this
solution was that the pre-allocated hash table could take a significant
chunk of the memory even when the resolver cache would be otherwise
empty because the default value for max-cache-size is 90% of available
memory.
Implement incremental resizing[1] to perform the rehashing gradually:
1. During the resize, allocate the new hash table, but keep the old
table unchanged.
2. In each lookup or delete operation, check both tables.
3. Perform insertion operations only in the new table.
4. At each insertion also move r elements from the old table to the new
table.
5. When all elements are removed from the old table, deallocate it.
To ensure that the old table is completely copied over before the new
table itself needs to be enlarged, it is necessary to increase the
size of the table by a factor of at least (r + 1)/r during resizing.
In our implementation r is equal to 1.
The downside of this approach is that the old table and the new table
could stay in memory for longer when there are no new insertions into
the hash table for prolonged periods of time as the incremental
rehashing happens only during the insertions.
The upside of this approach is that it's no longer necessary to
pre-allocate large hash table, because the RBT hash table rehashing
doesn't cause resolution brownouts anymore and thus we can use the
memory as needed.
1. https://en.m.wikipedia.org/wiki/Hash_table#Dynamic_resizing
Remove the dynamic registration of result codes. Convert isc_result_t
from unsigned + #defines into 32-bit enum type in grand unified
<isc/result.h> header. Keep the existing values of the result codes
even at the expense of the description and identifier tables being
unnecessary large.
Additionally, add couple of:
switch (result) {
[...]
default:
break;
}
statements where compiler now complains about missing enum values in the
switch statement.
- fixed a size comparison using "signed int" that failed if the file
size was more than 2GB, since that was treated as a negative number.
- incidentally renamed deserialize32() to just deserialize(). we no
longer have separate 32 and 64 bit rbtdb implementations.
The isc/platform.h header was left empty which things either already
moved to config.h or to appropriate headers. This is just the final
cleanup commit.
When "max-cache-size" is changed to "unlimited" (or "0") for a running
named instance (using "rndc reconfig"), the hash table size limit for
each affected cache DB is not reset to the maximum possible value,
preventing those hash tables from being allowed to grow as a result of
new nodes being added.
Extend dns_rbt_adjusthashsize() to interpret "size" set to 0 as a signal
to remove any previously imposed limits on the hash table size. Adjust
API documentation for dns_db_adjusthashsize() accordingly. Move the
call to dns_db_adjusthashsize() from dns_cache_setcachesize() so that it
also happens when "size" is set to 0.
Upon creation, each dns_rbt_t structure has its "maxhashbits" field
initialized to the value of the RBT_HASH_MAX_BITS preprocessor macro,
i.e. 32. When the dns_rbt_adjusthashsize() function is called for the
first time for a given RBT (for cache RBTs, this happens when they are
first created, i.e. upon named startup), it lowers the value of the
"maxhashbits" field to the number of bits required to index the
requested number of hash table slots. When a larger hash table size is
subsequently requested, the value of the "maxhashbits" field should be
increased accordingly, up to RBT_HASH_MAX_BITS. However, the loop in
the rehash_bits() function currently ensures that the number of bits
necessary to index the resized hash table will not be larger than
rbt->maxhashbits instead of RBT_HASH_MAX_BITS, preventing the hash table
from being grown once the "maxhashbits" field of a given dns_rbt_t
structure is set to any value lower than RBT_HASH_MAX_BITS.
Fix by tweaking the loop guard condition in the rehash_bits() function
so that it compares the new number of bits used for indexing the hash
table against RBT_HASH_MAX_BITS rather than rbt->maxhashbits.
When calculating the new hashtable bitsize, there was an off-by-one
error that would allow the new bitsize to be larger than maximum allowed
causing assertion failure in the rehash() function.
The hash table rework MRs (!3865, !3871) increased the default RBT hash
table size from 64 to 65,536 entries (for 64-bit architectures, that is
512 bytes before vs. 524,288 bytes after). This works fine for RBTs
used for cache databases, but since three separate RBT databases are
created for every zone loaded (RRs, NSEC, NSEC3), memory usage would
skyrocket when BIND 9 is used as an authoritative DNS server with many
zones.
The default RBT hash table size before the rework was 64 entries, this
commit reduces it to 16 entries because our educated guess is that most
zones are just couple of entries (SOA, NS, A, AAAA, MX) and rehashing
small hash tables is actually cheap. The rework we did in the previous
MRs tries to avoid growing the hash tables for big-to-huge caches where
growing the hash table comes at a price because the whole cache needs to
be locked.
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/
