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.
the isc_glob module was originally needed to support posix-style glob
processing on Windows, but is now just an unnecessary wrapper around
glob(3). this commit removes it.
Previously, the async job queue would use a locked-list (ISC_LIST).
With introduction of atomic stack (that has to be drained at once), we
could use it to remove some contention between the threads and simplify
the async queue.
Fortunately, the reverse order still works for us - instead of append
and tail/prev operation on the list, we are now using prepend and
head/next operation on the atomic stack.
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()`
removed references in code comments, doc/dev documentation, etc, to
isc_task, isc_timer_reset(), and isc_timertype_inactive. also removed a
coccinelle patch related to isc_timer_reset() that was no longer needed.
as there is no further use of isc_task in BIND, this commit removes
it, along with isc_taskmgr, isc_event, and all other related types.
functions that accepted taskmgr as a parameter have been cleaned up.
as a result of this change, some functions can no longer fail, so
they've been changed to type void, and their callers have been
updated accordingly.
the tasks table has been removed from the statistics channel and
the stats version has been updated. dns_dyndbctx has been changed
to reference the loopmgr instead of taskmgr, and DNS_DYNDB_VERSION
has been udpated as well.
change functions using isc_taskmgr_beginexclusive() to use
isc_loopmgr_pause() instead.
also, removed an unnecessary use of exclusive mode in
named_server_tcptimeouts().
most functions that were implemented as task events because they needed
to be running in a task to use exclusive mode have now been changed
into loop callbacks instead. (the exception is catz, which is being
changed in a separate commit because it's a particularly complex change.)
This changes the internal isc_rwlock implementation to:
Irina Calciu, Dave Dice, Yossi Lev, Victor Luchangco, Virendra
J. Marathe, and Nir Shavit. 2013. NUMA-aware reader-writer locks.
SIGPLAN Not. 48, 8 (August 2013), 157–166.
DOI:https://doi.org/10.1145/2517327.24425
(The full article available from:
http://mcg.cs.tau.ac.il/papers/ppopp2013-rwlocks.pdf)
The implementation is based on the The Writer-Preference Lock (C-RW-WP)
variant (see the 3.4 section of the paper for the rationale).
The implemented algorithm has been modified for simplicity and for usage
patterns in rbtdb.c.
The changes compared to the original algorithm:
* We haven't implemented the cohort locks because that would require a
knowledge of NUMA nodes, instead a simple atomic_bool is used as
synchronization point for writer lock.
* The per-thread reader counters are not being used - this would
require the internal thread id (isc_tid_v) to be always initialized,
even in the utilities; the change has a slight performance penalty,
so we might revisit this change in the future. However, this change
also saves a lot of memory, because cache-line aligned counters were
used, so on 32-core machine, the rwlock would be 4096+ bytes big.
* The readers use a writer_barrier that will raise after a while when
readers lock can't be acquired to prevent readers starvation.
* Separate ingress and egress readers counters queues to reduce both
inter and intra-thread contention.
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
removed some functions that are no longer used and unlikely to
be resurrected, and also some that were only used to support Windows
and can now be replaced with generic versions.
isc_bind9 was a global bool used to indicate whether the library
was being used internally by BIND or by an external caller. external
use is no longer supported, but the variable was retained for use
by dyndb, which needed it only when being built without libtool.
building without libtool is *also* no longer supported, so the variable
can go away.
libuv support for receiving multiple UDP messages in a single system
call (recvmmsg()) has been tweaked several times between libuv versions
1.35.0 and 1.40.0. Mixing and matching libuv versions within that span
may lead to assertion failures and is therefore considered harmful, so
try to limit potential damage be preventing users from mixing libuv
versions with distinct sets of recvmmsg()-related flags.
The implementation of UDP recvmmsg in libuv 1.35 and 1.36 is
incomplete and could cause assertion failure under certain
circumstances.
Modify the configure and runtime checks to report a fatal error when
trying to compile or run with the affected versions.
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.
the rate limter now uses loop callbacks rather than task events.
the API for isc_ratelimiter_enqueue() has been changed; we now pass
in a loop, a callback function and a callback argument, and
receive back a rate limiter event object (isc_rlevent_t). it
is no longer necessary for the caller to allocate the event.
the callback argument needs to include a pointer to the rlevent
object so that it can be freed using isc_rlevent_free(), or by
dequeueing.
This restores the Malloced memory counter and it's now always equal to
InUse counter. This is only for backwards compatibility reason and
there is no separate counter.
The commit also cleanups little things like structure with a single
item (summary.inuse), and shuts up a wrong cppcheck warning (the
notorious NULL check after assignment).
Instead of enforcing stronger synchronization between threads, make all
the atomic operations relaxed. We are not really interested in exact
numbers at all times - the single place where we need the exact number
is when the memory context is being destroyed. Even when there's a
overmem counter, we don't care about exact ordering or exact number.
The Lost memory counter would count the memory "lost" by external
libraries. There's really no such thing as `named` require the memory
contexts to be clean on destroy.
The stats buckets were again more useful for internal allocator, because
we would see the individual "block" caches where the allocations would
fall into. Remove the stats buckets, and if needed, we can pull more
detailed statistics out of the jemalloc.
The total memory counter had again little or no meaning when we removed
the internal memory allocator. It was just a monotonic counter that
would count add the allocation sizes but never subtracted anything, so
it would be just a "big number".
The maxinuse memory counter indicated the highest amount of
memory allocated in the past. Checking and updating this high-
water mark value every time memory was allocated had an impact
on server performance, so it has been removed. Memory size can
be monitored more efficiently via an external tool logging RSS.
The malloced and maxmalloced memory counters were mostly useless since
we removed the internal allocator blocks - it would only differ from
inuse by the memory context size itself.
Use isc_job_run() instead of isc_task_send() for dnssec-signzone
worker threads.
Also fix the issue where the additional assignwork() would be run only
from the main thread effectively serializing all the signing.
The old code didn't handle race conditions and errors on systems
with non load balancing sockets gracefully. Look for an error on
any child socket and if found close all the child sockets and return
an error.
The old code didn't handle race conditions and errors on systems
with non load balancing sockets gracefully. Look for an error on
any child socket and if found close all the child sockets and return
an error.
This commit ensures that BIND and supplementary tools still can be
built on newer versions of DragonFly BSD. It used to be the case, but
somewhere between versions 6.2 and 6.4 the OS developers rearranged
headers and moved some function definitions around.
Before that the fact that it worked was more like a coincidence, this
time we, at least, looked at the related man pages included with the
OS.
No in depth testing has been done on this OS as we do not really
support this platform - so it is more like a goodwill act. We can,
however, use this platform for testing purposes, too. Also, we know
that the OS users do use BIND, as it is included in its ports
directory.
Building with './configure' and './configure --without-jemalloc' have
been fixed and are known to work at the time the commit is made.
Return 'isc_result_t' type value instead of 'bool' to indicate
the actual failure. Rename the function to something not suggesting
a boolean type result. Make changes in the places where the API
function is being used to check for the result code instead of
a boolean value.
Cherry-pick small fixup commit from 9.18/9.16 branches needed for
thread-safety. This fixup commit is not needed for 9.19+ because of
reworked application setup, but it decouples isc_iterated_hash and
isc_md units and keeps all the branches in sync.
As this code is on hot path (NSEC3) this introduces an additional
optimization of the EVP_MD API - instead of calling EVP_MD_CTX_new() on
every call to isc_iterated_hash(), we create two thread_local objects
for each thread - a basectx and mdctx, initialize basectx once and then
use EVP_MD_CTX_copy_ex() to flip the initialized state into mdctx. This
saves us couple more valuable microseconds from the isc_iterated_hash()
call.
If the OpenSSL SHA1_{Init,Update,Final} API is still available, use it.
The API has been deprecated in OpenSSL 3.0, but it is significantly
faster than EVP_MD API, so make an exception here and keep using it
until we can't.
