This is a server hang and not an issue with backup. While concurrent
DDLs in server gets in hanged state, mariabackup waits for DDLs to
finish trying to acquire MDL_BACKUP_BLOCK_DDL.
The server hang is serious in nature and caused by thread pool state
being incorrectly set to thread creation pending state while no creation
is actually pending. Once a thread pool reaches such state no new thread
gets created in the pool.
While it could possibly affect all thread pools in server, the innodb
thread pool is the victim in current bug where IO job gets blocked when
the pool is stuck with much less number of threads than intended.
Available workers are blocked in purge waiting for page lock to be
released by IO write (SX lock) causing a complete deadlock.
The issue is caused by the state variable m_thread_creation_pending
introduced by MDEV-31095: 9e62ab7aaf. We check and set the variable
early while attempting to create a new thread in pool but fail to reset
it if we exit the flow for other reasons like maximum threads reached
or get into thread creation throttling path.
Fix: The simple fix is to make sure that the state is reset back in case
we don't actually attempt to create the thread.
Previous solution, that would entirely switch timer off, turned out
to be deadlock prone.
This patch fixed previous attempt to switch between long/short interval
periods in MDEV-24295. Now, initial state of the timer is fixed (it is ON).
Also, avoid switching timer to longer periods if there is any activity in
the pool.
Before patch, maintenance timer will tick every 0.4 seconds.
After this patch, timer will tick every 0.4 seconds when necessary(
there are delayed thread creation), switching off completely after 20
seconds of being idle.
Add threadpool functionality to restrict concurrency during "batch"
periods (where tasks are added in rapid succession).
This will throttle thread creation more agressively than usual, while
keeping performance at least on-par.
One of these cases is bufferpool load, where async read IOs are executed
without any throttling. There can be as much as 650K read IOs for
loading 10GB buffer pool.
Another one is recovery, where "fake read" IOs are executed.
Why there are more threads than we expect?
Worker threads are not be recognized as idle, until they return to the
standby list, and to return to that list, they need to acquire
mutex currently held in the submit_task(). In those cases, submit_task()
has no worker to wake, and would create threads until default concurrency
level (2*ncpus) is satisfied. Only after that throttling would happen.
Fixed tpool timer implementation on POSIX.
Prior to this patch, under some specific rare circumstances (concurrency
related), timer callback execution might be skipped.
Table_cache_instance: Define the structure aligned at
the CPU cache line, and remove a pad[] data member.
Krunal Bauskar reported this to improve performance on ARMv8.
aligned_malloc(): Wrapper for the Microsoft _aligned_malloc()
and the ISO/IEC 9899:2011 <stdlib.h> aligned_alloc().
Note: The parameters are in the Microsoft order (size, alignment),
opposite of aligned_alloc(alignment, size).
Note: The standard defines that size must be an integer multiple
of alignment. It is enforced by AddressSanitizer but not by GNU libc
on Linux.
aligned_free(): Wrapper for the Microsoft _aligned_free() and
the standard free().
HAVE_ALIGNED_ALLOC: A new test. Unfortunately, support for
aligned_alloc() may still be missing on some platforms.
We will fall back to posix_memalign() for those cases.
HAVE_MEMALIGN: Remove, along with any use of the nonstandard memalign().
PFS_ALIGNEMENT (sic): Removed; we will use CPU_LEVEL1_DCACHE_LINESIZE.
PFS_ALIGNED: Defined using the C++11 keyword alignas.
buf_pool_t::page_hash_table::create(),
lock_sys_t::hash_table::create():
lock_sys_t::hash_table::resize(): Pad the allocation size to an
integer multiple of the alignment.
Reviewed by: Vladislav Vaintroub
As btrfs showed, a partial read of data in AIO /O_DIRECT circumstances can
really confuse MariaDB.
Filipe Manana (SuSE)[1] showed how database programmers can assume
O_DIRECT is all or nothing.
While a fix was done in the kernel side, we can do better in our code by
requesting that the rest of the block be read/written synchronously if
we do only get a partial read/write.
Per the APIs, a partial read/write can occur before an error, so
reattempting the request will leave the caller with a concrete error to
handle.
[1] https://lore.kernel.org/linux-btrfs/CABVffENfbsC6HjGbskRZGR2NvxbnQi17gAuW65eOM+QRzsr8Bg@mail.gmail.com/T/#mb2738e675e48e0e0778a2e8d1537dec5ec0d3d3a
Also spell synchronously correctly in other files.
liburing is a new optional dependency (WITH_URING=auto|yes|no)
that replaces libaio when it is available.
aio_uring: class which wraps io_uring stuff
aio_uring::bind()/unbind(): optional optimization
aio_uring::submit_io(): mutex prevents data race. liburing calls are
thread-unsafe. But if you look into it's implementation you'll see
atomic operations. They're used for synchronization between kernel and
user-space only. That's why our own synchronization is still needed.
For systemd, we add LimitMEMLOCK=524288 (ulimit -l 524288)
because the io_uring_setup system call that is invoked
by io_uring_queue_init() requests locked memory. The value
was found empirically; with 262144, we would occasionally
fail to enable io_uring when using the maximum values of
innodb_read_io_threads=64 and innodb_write_io_threads=64.
aio_uring::thread_routine(): Tolerate -EINTR return from
io_uring_wait_cqe(), because it may occur on shutdown
on Ubuntu 20.10 (Groovy Gorilla).
This was mostly implemented by Eugene Kosov. Systemd integration
and improved startup/shutdown error handling by Marko Mäkelä.
If maintenance timer does not do much for prolonged time, it will
wake up less frequently, once every 4 seconds instead of once every 0.4
second.
It will wakeup more often if thread creation is throttled, to avoid stalls.
While waiting for mutex, thread_pool_generic::wait_begin(),
current task can be marked long-running. This is done by periodic
mantainence task, that runs in parallel.
Fix to recheck is_long_task() after the mutex acquisition.
maybe_wake_or_create_thread()
A task that is executed,could be counted as waiting (after wait_begin()
before wait_end()) or as long-running (callback runs for a long time).
If task is both marked waiting and long running, then calculation of
current concurrency (# of executing tasks - # of long tasks - #of waiting tasks)
is wrong, as task is counted twice.
Thus current concurrency could go negative, but with unsigned arithmetic
it will become a huge number.
As a result, maybe_wake_or_create_thread() would neither wake or create
a thread, when it should. Which may result in a deadlock.
1. Fix places where data race warnings were relevant.
tls_worker_data::m_state should be modified under mutex protection,
since both maintainence timer and current worker set this flag.
2. Suppress warnings that are legitimate, yet harmless.
Apparently, the dirty reads in waitable_task::get_ref_count() or
write_slots->pending_io_count()
Avoiding race entirely without side-effects here is tricky,
and the effects of race is harmless.
The worst thing that can happen due to race is an extra wait notification,
under rare circumstances.
- wait notification, tpool_wait_begin/tpool_wait_end - to notify the
threadpool that current thread is going to wait
Use it to wait for IOs to complete and also when purge waits for workers.
The library is capable of
- asynchronous execution of tasks (and optionally waiting for them)
- asynchronous file IO
This is implemented using libaio on Linux and completion ports on
Windows. Elsewhere, async io is "simulated", which means worker threads
are performing synchronous IO.
- timers, scheduling work asynchronously in some point of the future.
Also periodic timers are implemented.
