If there are no more previous leaves, it means the queried name
precedes the entire range of names in the database, so we should just
move the iterator one step back and return, instead of continuing our
search for the predecessor.
This is similar to an earlier bug fixed in an earlier commit:
ea9a8cb392ff59438a911485742b220d40f24d6f
previously, an RCU critical section was held open for the duration
of a snapshot. this should not be necessary, as the snapshot makes
local copies of QP trie metadata, and it causes problems when a
DB iterator is held open between two loop events. we now call
rcu_read_unlock() after setting up the snapshot.
it was possible for fix_iterator() to get stuck in a loop while
trying to find the predecessor of a missing node. this has been
fixed and a regression test has been added.
the fix_iterator() function moves an iterator so that it points
to the predecessor of the searched-for name when that name doesn't
exist in the database. the tests only checked the correctness of
the top of the stack, however, and missed some cases where interior
branches in the stack could be missing or duplicated. in these
cases, the iterator would produce inconsistent results when walked.
the predecessors test case in qp_test has been updated to walk
each iterator to the end and ensure that the expected number of
nodes are found.
Move the code to find the predecessor into one function, as it is shares
quite some similarities: In both cases we first need to find the
immediate predecessor/successor, then we need to find the immediate
predecessor if the iterator is not already pointing at it.
This one is similar to the bug when searching for a key, reaching a
dead-end branch that doesn't match, because the branch offset point
is after the point where the search key differs.
This fixes the case where we are multiple levels deep. In other
words, we had a more-than-one matches *after* the point where the
search key differs.
For example, consider the following qp-trie:
branch: "[e]", "[m]":
- leaf: "a.b.c.d.e"
- branch: "moo[g]", "moo[k]", "moo[n]":
- leaf: "moog"
- branch: "mook[e]", "mook[o]"
- leaf: "mooker"
- leaf: "mooko"
- leaf: "moon"
If searching for a key "monky", we would reach the branch with
twigs "moo[k]" and "moo[n]". The key matches on the 'k' on offset=4,
and reaches the branch with twigs "mook[e]" and "mook[o]". This time
we cannot find a twig that matches our key at offset=5, there is no
twig for 'y'. The closest name we found was "mooker".
Note that on a branch it can't detect it is on a dead branch because the
key is not encapsulated in a branch node.
In the previous code we considered "mooker" to be the successor of
"monky" and so we needed to the predecessor of "mooker" to find the
predecessor for "monky". However, since the search key alread differed
before entering this branch, this is not enough. We would be left with
"moog" as the predecessor of "monky", while in this example "a.b.c.d.e"
is the actual predecessor.
Instead, we need to go up a level, find the predecessor and check
again if we are on the right branch, and repeat the process until we
are.
Unit tests to cover the scenario are now added.
There was yet another edge case in which an iterator could be
positioned at the wrong node after dns_qp_lookup(). When searching for
a key, it's possible to reach a leaf that matches at the given offset,
but because the offset point is *after* the point where the search key
differs from the leaf's contents, we are now at the wrong leaf.
In other words, the bug fixed the previous commit for dead-end branches
must also be applied on matched leaves.
For example, if searching for the key "monpop", we could reach a branch
containing "moop" and "moor". the branch offset point - i.e., the point
after which the branch's leaves differ from each other - is the
fourth character ("p" or "r"). The search key matches the fourth
character "p", and takes that twig to the next node (which can be
a branch for names starting with "moop", or could be a leaf node for
"moop").
The old code failed to detect this condition, and would have
incorrectly left the iterator pointing at some successor, and not
at the predecessor of the "moop".
To find the right predecessor in this case, we need to get to the
previous branch and get the previous from there.
This has been fixed and the unit test now includes several new
scenarios for testing search names that match and unmatch on the
offset but have a different character before the offset.
there was another edge case in which an iterator could be positioned at
the wrong node after dns_qp_lookup(). when searching for a key, it's
possible to reach a dead-end branch that doesn't match, because the
branch offset point is *after* the point where the search key differs
from the branch's contents.
for example, if searching for the key "mop", we could reach a branch
containing "moon" and "moor". the branch offset point - i.e., the
point after which the branch's leaves differ from each other - is the
fourth character ("n" or "r"). however, both leaves differ from the
search key at position *three* ("o" or "p"). the old code failed to
detect this condition, and would have incorrectly left the iterator
pointing at some lower value and not at "moor".
this has been fixed and the unit test now includes this scenario.
in some cases it was possible for the iterator to be positioned in the
wrong place by dns_qp_lookup(). previously, when a leaf node was found
which matched the search key at its parent branch's offset point, but
did not match after that point, the code incorrectly assumed the leaf
it had found was a successor to the searched-for name, and stepped the
iterator back to find a predecessor. however, it was possible for the
non-matching leaf to be the predecessor, in which case stepping the
iterator back was wrong.
(for example: a branch contains "aba" and "abcd", and we are searching
for "abcde". we step down to the twig matching the letter "c" in
position 3. "abcd" is the predecessor of "abcde", so the iterator is
already correctly positioned, but because the twig was an exact match,
we would have moved it back one step to "aba".)
this previously went unnoticed due to a mistake in the qp_test unit
test, which had the wrong expected result for the test case that should
have detected the error. both the code and the test have been fixed.
the 'predecessor' argument to dns_qp_lookup() turns out not to
be sufficient for our needs: the predecessor node in a QP database
could have become "empty" (for the current version) because of an
update or because cache data expired, and in that case the caller
would have to iterate more than one step back to find the predecessor
node that it needs.
it may also be necessary for a caller to iterate forward, in
order to determine whether a node has any children.
for both of these reasons, we now replace the 'predecessor'
argument with an 'iter' argument. if set, this points to memory
with enough space for a dns_qpiter object.
when an exact match is found by the lookup, the iterator will be
pointing to the matching node. if not, it will be pointing to the
lexical predecessor of the nae that was searched for.
a dns_qpiter_current() method has been added for examining
the current value of the iterator without moving it in either
direction.
The `.reader` member of dns_qpmulti_t was accessed without RCU
protection; reader_open() calls rcu_dereference() on it, and this
call needs to be inside an RCU critical section.
A similar problem was identified in the dns_qpmulti_snapshot() - the
RCU critical section was completely missing.
These are relicts of the isc_qsbr - in the QSBR mode the rcu_read_lock()
and rcu_read_unlock() are no-ops and whole event loop is a critical section.
It was possible to reach add_link() without visiting an
intermediate node first, and the check for a duplicate entry
could then cause a crash.
Credit to OSS-Fuzz for discovering this error.
depending on how the QP trie is traversed during a lookup, it is
possible for a search to terminate on a leaf which is a partial
match, without that leaf being added to the chain. to ensure the
chain is correct in this case, when a partial match condition is
detected via qpkey_compare(), we will call add_link() again, just
in case. (add_link() will check for a duplicated node, so it will
be harmless if it was already done.)
dns_qp_findname_ancestor() now takes an optional 'predecessor'
parameter, which if non-NULL is updated to contain the DNSSEC
predecessor of the name searched for. this is done by constructing
an iterator stack while carrying out the search, so it can be used
to step backward if needed.
since dns_qp_findname_ancestor() can now return a chain object, it is no
longer necessary to provide a _NOEXACT search option. if we want to look
up the closest ancestor of a name, we can just do a normal search, and
if successful, retrieve the second-to-last node from the QP chain.
this makes ancestor lookups slightly more complicated for the caller,
but allows us to simplify the code in dns_qp_findname_ancestor(), making
it easier to ensure correctness. this was a fairly rare use case:
outside of unit tests, DNS_QPFIND_NOEXACT was only used in the zone
table, which has now been updated to use the QP chain. the equivalent
RBT feature is only used by the resolver for cache lookups of 'atparent'
types (i.e, DS records).
- make iterators reversible: refactor dns_qpiter_next() and add a new
dns_qpiter_prev() function to support iterating both forwards and
backwards through a QP trie.
- added a 'name' parameter to dns_qpiter_next() (as well as _prev())
to make it easier to retrieve the nodename while iterating, without
having to construct it from pointer value data.
- the helper functions for accessing twigs beneath a branch
(branch_twig_pos(), branch_twig_ptr(), etc) were somewhat confusing
to read, since several of them were implemented by calling other
helper functions. they now all show what they're really doing.
- branch_twigs_vector() has been renamed to simply branch_twigs().
- revised some unrelated comments in qp_p.h for clarity.
dns_qp_findname_ancestor() now takes an optional 'chain' parameter;
if set, the dns_qpchain object it points to will be updated with an
array of pointers to the populated nodes between the tree root and the
requested name. the number of nodes in the chain can then be accessed
using dns_qpchain_length() and the individual nodes using
dns_qpchain_node().
add a 'foundname' parameter to dns_qp_findname_ancestor(),
and use it to set the found name in dns_nametree.
this required adding a dns_qpkey_toname() function; that was
done by moving qp_test_keytoname() from the test library to qp.c.
added some more test cases and fixed bugs with the handling of
relative and empty names.
this function finds the closest matching ancestor, but the function
name could be read to imply that it returns the direct parent node;
this commit suggests a slightly less misleading name.
Make the `pval_r` and `ival_r` out arguments optional.
Add `pval_r` and `ival_r` out arguments to `dns_qp_deletekey()`
and `dns_qp_deletename()`, to return the deleted leaf.
With ThreadSanitizer support added to the Userspace RCU, we no longer
need to wrap the call_rcu and caa_container_of with
__tsan_{acquire,release} hints. Remove the direct calls to
__tsan_{acquire,release} and the isc_urcu_{container,cleanup} macros.
Thread sanitizer warns that parts of the qp-trie are accessed
both with and without the mutex; the unlocked accesses happen during
destruction, so they should be benign, but there's no harm locking
anyway to convince tsan it is clean.
Also, ensure .tsan-suppress and .tsan-suppress-extra are in sync.
Move registration and deregistration of the main thread from
`isc_loopmgr_run()` into `isc__initialize()` / `isc__shutdown()`:
liburcu-qsbr fails an assertion if we try to use it from an
unregistered thread, and we need to be able to use it when the
event loops are not running.
Use `rcu_assign_pointer()` and `rcu_dereference()` in qp-trie
transactions so that they properly mark threads as online. The
RCU-protected pointer is no longer declared atomic because
liburcu does not (yet) use standard C atomics.
Fix the definition of `isc_qsbr_rcu_dereference()` to return
the referenced value, and to call the right function inside
liburcu.
Change the thread sanitizer suppressions to match any variant of
`rcu_*_barrier()`
All the places the qp-trie code was using `call_rcu()` needed
`__tsan_release()` and `__tsan_acquire()` annotations, so
add a couple of wrappers to encapsulate this pattern.
With these wrappers, the tests run almost clean under thread
sanitizer. The remaining problems are due to `rcu_barrier()`
which can be suppressed using `.tsan-suppress`. It does not
suppress the whole of `liburcu`, because we would like thread
sanitizer to detect problems in `call_rcu()` callbacks, which
are called from `liburcu`.
The CI jobs have been updated to use `.tsan-suppress` by
default, except for a special-case job that needs the
additional suppressions in `.tsan-suppress-extra`.
We might be able to get rid of some of this after liburcu gains
support for thread sanitizer.
Note: the `rcu_barrier()` suppression is not entirely effective:
tsan sometimes reports races that originate inside `rcu_barrier()`
but tsan has discarded the stack so it does not have the
information required to suppress the report. These "races" can
be made much easier to reproduce by adding `atexit_sleep_ms=1000`
to `TSAN_OPTIONS`. The problem with tsan's short memory can be
addressed by increasing `history_size`: when it is large enough
(6 or 7) the `rcu_barrier()` stack usually survives long enough
for suppression to work.
A `dns_qmpulti_t` no longer needs to know about its loopmgr. We no
longer keep a linked list of `dns_qpmulti_t` that have reclamation
work, and we no longer mark chunks with the phase in which they are to
be reclaimed. Instead, empty chunks are listed in an array in a
`qp_rcu_t`, which is passed to call_rcu().
It can be fairly long-winded to allocate space for a struct with a
flexible array member: in general we need the size of the struct, the
size of the member, and the number of elements. Wrap them all up in a
STRUCT_FLEX_SIZE() macro, and use the new macro for the flexible
arrays in isc_ht and dns_qp.
Revert refcount debug tracing (commit a8b29f0365), there are better
ways to do it.
Use the dns_qpmethods_t typedef where appropriate.
Some stylistic improvements.
Commit 0858514ae8 enriched dns_qp_compact() to give callers more
control over how thoroughly the trie should be compacted.
In the DNS_QPGC_ALL case, if the trie is small it might be compacted
to a new position in the same memory chunk. In this situation it will
still be holding references to old leaf objects which have been
removed from the trie but will not be completely detached until the
chunk containing the references is freed.
This change resets the qp-trie allocator to a fresh chunk before a
DNS_QPGC_ALL compaction, so all the old memory chunks will be
evacuated and old leaf objects can be detached sooner.
It is sometimes necessary to access a qp-trie outside an isc_loop,
such as in tests or an isc_work callback. The best option was to use
a `dns_qpmulti_write()` transaction, but that has overheads that are
not necessary for read-only access, such as committing a new version
of the trie even when nothing changed.
So this commit adds a `dns_qpmulti_read()` transaction, which is
nearly as lightweight as a query transaction, but it takes the mutex
like a write transaction.
This is the first of the "fancy" searches that know how the DNS
namespace maps on to the structure of a qp-trie. For example, it will
find the closest enclosing zone in the zone tree.
My original idea had been that the core qp-trie code would be mostly
independent of the storage for keys, so I did not make it check at run
time that key lengths are sensible. However, the qp-trie search
routines need to get keys out of leaf objects, for which they provide
storage on the stack, which is particularly dangerous for unchecked
buffer overflows. So this change checks that key lengths are in bounds
at the API boundary between the qp-trie code and the rest of BIND, and
there is no more pretence that keys might be longer.
In general, it's better to do one thorough compaction when a batch of
work is complete, which is the way that `update` transactions work.
Conversely, `write` transactions are designed so that lots of little
transactions are not too inefficient, but they need explicit
compaction. This changes `dns_qp_compact()` so that it is easier to
compact any time that makes sense, if there isn't a better way to
schedule compaction. And `dns_qpmulti_commit()` only recycles garbage
when there is enough to make it worthwhile.
The first working multi-threaded qp-trie was stuck with an unpleasant
trade-off:
* Use `isc_rwlock`, which has acceptable write performance, but
terrible read scalability because the qp-trie made all accesses
through a single lock.
* Use `liburcu`, which has great read scalability, but terrible
write performance, because I was relying on `rcu_synchronize()`
which is rather slow. And `liburcu` is LGPL.
To get the best of both worlds, we need our own scalable read side,
which we now have with `isc_qsbr`. And we need to modify the write
side so that it is not blocked by readers.
Better write performance requires an async cleanup function like
`call_rcu()`, instead of the blocking `rcu_synchronize()`. (There
is no blocking cleanup in `isc_qsbr`, because I have concluded
that it would be an attractive nuisance.)
Until now, all my multithreading qp-trie designs have been based
around two versions, read-only and mutable. This is too few to
work with asynchronous cleanup. The bare minimum (as in epoch
based reclamation) is three, but it makes more sense to support an
arbitrary number. Doing multi-version support "properly" makes
fewer assumptions about how safe memory reclamation works, and it
makes snapshots and rollbacks simpler.
To avoid making the memory management even more complicated, I
have introduced a new kind of "packed reader node" to anchor the
root of a version of the trie. This is simpler because it re-uses
the existing chunk lifetime logic - see the discussion under
"packed reader nodes" in `qp_p.h`.
I have also made the chunk lifetime logic simpler. The idea of a
"generation" is gone; instead, chunks are either mutable or
immutable. And the QSBR phase number is used to indicate when a
chunk can be reclaimed.
Instead of the `shared_base` flag (which was basically a one-bit
reference count, with a two version limit) the base array now has a
refcount, which replaces the confusing ad-hoc lifetime logic with
something more familiar and systematic.
Adjust the dns_qp_memusage() and dns_qp_compact() functions
to be more informative and flexible about handling fragmentation.
Avoid wasting space in runt chunks.
Switch from twigs_mutable() to cells_immutable() because that is the
sense we usually want.
Drop the redundant evacuate() function and rename evacuate_twigs() to
evacuate(). Move some chunk test functions closer to their point of
use.
Clarify compact_recursive(). Some small cleanups to comments.
Use isc_time_monotonic() for qp-trie timing stats.
Use #define constants to control debug logging.
Set up DNS name label offsets in dns_qpkey_fromname() so it is easier
to use in cases where the name is not fully hydrated.
The error occurred when:
* The bump chunk was re-used across multiple write transactions.
In this situation the bump chunk is marked immutable, but the
immutable flag is disregarded for cells after the fender, which
were allocated in the current transaction.
* The bump chunk fills up during an insert operation, so that the
enlarged twigs vector is allocated from a new bump chunk.
* Before this happened, we should have (but didn't) made the twigs
vector mutable. This would have adjusted its refcounts as necessary.
* However, moving to a new bump chunk has a side effect: twigs that
were previously considered mutable because they are after the
fender become immutable.
* Because of this, the old twigs vector was not destroyed as expected.
* So leaves were duplicated without their refcounts being increased.
The effect is that the refcounts were lower than they should have
been, and underflowed. The tests failed to check for refcount
underflow, so this mistake was detected much later than it ideally
could have been.
After the fix, it is now correct not to ensure the twigs are mutable,
because they are about to be copied to a larger vector. Instead, we
need to find out whether `squash_twigs()` destroyed the old twigs, and
adjust the refcounts accordingly.
A qp-trie is a kind of radix tree that is particularly well-suited to
DNS servers. I invented the qp-trie in 2015, based on Dan Bernstein's
crit-bit trees and Phil Bagwell's HAMT. https://dotat.at/prog/qp/
This code incorporates some new ideas that I prototyped using
NLnet Labs NSD in 2020 (optimizations for DNS names as keys)
and 2021 (custom allocator and garbage collector).
https://dotat.at/cgi/git/nsd.git
The BIND version of my qp-trie code has a number of improvements
compared to the prototype developed for NSD.
* The main omission in the prototype was the very sketchy outline of
how locking might work. Now the locking has been implemented,
using a reader/writer lock and a mutex. However, it is designed to
benefit from liburcu if that is available.
* The prototype was designed for two-version concurrency, one
version for readers and one for the writer. The new code supports
multiversion concurrency, to provide a basis for BIND's dbversion
machinery, so that updates are not blocked by long-running zone
transfers.
* There are now two kinds of transaction that modify the trie: an
`update` aims to support many very small zones without wasting
memory; a `write` avoids unnecessary allocation to help the
performance of many small changes to the cache.
* There is also a single-threaded interface for situations where
concurrent access is not necessary.
* The API makes better use of types to make it more clear which
operations are permitted when.
* The lookup table used to convert a DNS name to a qp-trie key is
now initialized by a run-time constructor instead of a programmer
using copy-and-paste. Key conversion is more flexible, so the
qp-trie can be used with keys other than DNS names.
* There has been much refactoring and re-arranging things to improve
the terminology and order of presentation in the code, and the
internal documentation has been moved from a comment into a file
of its own.
Some of the required functionality has been stripped out, to be
brought back later after the basics are known to work.
* Garbage collector performance statistics are missing.
* Fancy searches are missing, such as longest match and
nearest match.
* Iteration is missing.
* Search for update is missing, for cases where the caller needs to
know if the value object is mutable or not.
