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.
`libirs` used to be a reference implementation of `getaddrinfo` and
related modern resolver APIs. It was stripped down in BIND 9.18
leaving only the `irs_resconf` module, which parses
`/etc/resolv.conf`. I have kept its include path and namespace prefix,
so it remains a little fragment of libirs now embedded in libdns.
This "quiescent state based reclamation" module provides support for
the qp-trie module in dns/qp. It is a replacement for liburcu, written
without reference to the urcu source code, and in fact it works in a
significantly different way.
A few specifics of BIND make this variant of QSBR somewhat simpler:
* We can require that wait-free access to a qp-trie only happens in
an isc_loop callback. The loop provides a natural quiescent state,
after the callbacks are done, when no qp-trie access occurs.
* We can dispense with any API like rcu_synchronize(). In practice,
it takes far too long to wait for a grace period to elapse for each
write to a data structure.
* We use the idea of "phases" (aka epochs or eras) from EBR to
reduce the amount of bookkeeping needed to track memory that is no
longer needed, knowing that the qp-trie does most of that work
already.
I considered hazard pointers for safe memory reclamation. They have
more read-side overhead (updating the hazard pointers) and it wasn't
clear to me how to nicely schedule the cleanup work. Another
alternative, epoch-based reclamation, is designed for fine-grained
lock-free updates, so it needs some rethinking to work well with the
heavily read-biased design of the qp-trie. QSBR has the fastest read
side of the basic SMR algorithms (with no barriers), and fits well
into a libuv loop. More recent hybrid SMR algorithms do not appear to
have enough benefits to justify the extra complexity.
Add a singly-linked stack that supports lock-free prepend and drain (to
empty the list and clean up its elements). Intended for use with QSBR
to collect objects that need safe memory reclamation, or any other user
that works with adding objects to the stack and then draining them in
one go like various work queues.
In <isc/atomic.h>, add an `atomic_ptr()` macro to make type
declarations a little less abominable, and clean up a duplicate
definition of `atomic_compare_exchange_strong_acq_rel()`
Unfortunately, C still lacks a standard function for pause (x86,
sparc) or yeild (arm) instructions, for use in spin lock or CAS loops.
BIND has its own based on vendor intrinsics or inline asm.
Previously, it was buried in the `isc_rwlock` implementation. This
commit renames `isc_rwlock_pause()` to `isc_pause()` and moves
it into <isc/pause.h>.
This commit also fixes the configure script so that it detects ARM
yield support on systems that identify as `aarch*` instead of `arm*`.
On 64-bit ARM systems we now use the ISB (instruction synchronization
barrier) instruction in preference to yield. The ISB instruction
pauses the CPU for longer, several nanoseconds, which is more like the
x86 pause instruction. There are more details in a Rust pull request,
which also refers to MySQL making the same change:
https://github.com/rust-lang/rust/pull/84725
* 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.)
ISC_MEM_ZERO requires great care to use when the space returned by
the allocator is larger than the requested space, and when memory is
reallocated. You must ensure that _every_ call to allocate or
reallocate a particular block of memory uses ISC_MEM_ZERO, to ensure
that the extra space is zeroed as expected. (When ISC_MEMFLAG_FILL
is set, the extra space will definitely be non-zero.)
When BIND is built without jemalloc, ISC_MEM_ZERO is implemented in
`jemalloc_shim.h`. This had a bug on systems that have malloc_size()
or malloc_usable_size(): memory was only zeroed up to the requested
size, not the allocated size. When an oversized allocation was
returned, and subsequently reallocated larger, memory between the
original requested size and the original allocated size could
contain unexpected nonzero junk. The realloc call does not know the
original requested size and only zeroes from the original allocated
size onwards.
After this change, `jemalloc_shim.h` always zeroes up to the
allocated size, not the requested size.
