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.
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.
Fixed tpool::pread() and tpool::pwrite() to return SSIZE_T on Windows,
so that huge numbers are not converted to negatives.
Also, make sure to never attempt reading/writing more bytes than
DWORD can accomodate (4G)
MDEV-23855 and MDEV-23399 already moved some transient data fields
from buffer pool page descriptors to IORequest, but the write buffer
of PAGE_COMPRESSED or ENCRYPTED tables was missed. Since is only needed
during asynchronous page write requests, it belongs to IORequest.
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.
In commit 5e62b6a5e0 (MDEV-16264)
the logic of os_aio_init() was changed so that it will never fail,
but instead automatically disable innodb_use_native_aio (which is
enabled by default) if the io_setup() system call would fail due
to resource limits being exceeded. This is questionable, especially
because falling back to simulated AIO may lead to significantly
reduced performance.
srv_n_file_io_threads, srv_n_read_io_threads, srv_n_write_io_threads:
Change the data type from ulong to uint.
os_aio_init(): Remove the parameters, and actually return an error code.
thread_pool::configure_aio(): Do not silently fall back to simulated AIO.
Reviewed by: Vladislav Vaintroub
tpool::aio::N_PENDING: Replaces OS_AIO_N_PENDING_IOS_PER_THREAD.
This limits two similar things: the number of outstanding requests
that a thread may io_submit(), and the number of completed requests
collected at a time by io_getevents().
os_aio_userdata_t: Remove. It was basically duplicating IORequest.
buf_page_write_complete(): Take only IORequest as a parameter.
os_aio_func(), pfs_os_aio_func(): Replaced with os_aio() that has
no redundant parameters. There is only one caller, so there is no
point to pass __FILE__, __LINE__ as a parameter.
User-visible changes: The INFORMATION_SCHEMA views INNODB_BUFFER_PAGE
and INNODB_BUFFER_PAGE_LRU will report a dummy value FLUSH_TYPE=0
and will no longer report the PAGE_STATE value READY_FOR_USE.
We will remove some fields from buf_page_t and move much code to
member functions of buf_pool_t and buf_page_t, so that the access
rules of data members can be enforced consistently.
Evicting or adding pages in buf_pool.LRU will remain covered by
buf_pool.mutex.
Evicting or adding pages in buf_pool.page_hash will remain
covered by both buf_pool.mutex and the buf_pool.page_hash X-latch.
After this fix, buf_pool.page_hash lookups can entirely
avoid acquiring buf_pool.mutex, only relying on
buf_pool.hash_lock_get() S-latch.
Similarly, buf_flush_check_neighbors() can will rely solely on
buf_pool.mutex, no buf_pool.page_hash latch at all.
The buf_pool.mutex is rather contended in I/O heavy benchmarks,
especially when the workload does not fit in the buffer pool.
The first attempt to alleviate the contention was the
buf_pool_t::mutex split in
commit 4ed7082eef
which introduced buf_block_t::mutex, which we are now removing.
Later, multiple instances of buf_pool_t were introduced
in commit c18084f71b
and recently removed by us in
commit 1a6f708ec5 (MDEV-15058).
UNIV_BUF_DEBUG: Remove. This option to enable some buffer pool
related debugging in otherwise non-debug builds has not been used
for years. Instead, we have been using UNIV_DEBUG, which is enabled
in CMAKE_BUILD_TYPE=Debug.
buf_block_t::mutex, buf_pool_t::zip_mutex: Remove. We can mainly rely on
std::atomic and the buf_pool.page_hash latches, and in some cases
depend on buf_pool.mutex or buf_pool.flush_list_mutex just like before.
We must always release buf_block_t::lock before invoking
unfix() or io_unfix(), to prevent a glitch where a block that was
added to the buf_pool.free list would apper X-latched. See
commit c5883debd6 how this glitch
was finally caught in a debug environment.
We move some buf_pool_t::page_hash specific code from the
ha and hash modules to buf_pool, for improved readability.
buf_pool_t::close(): Assert that all blocks are clean, except
on aborted startup or crash-like shutdown.
buf_pool_t::validate(): No longer attempt to validate
n_flush[] against the number of BUF_IO_WRITE fixed blocks,
because buf_page_t::flush_type no longer exists.
buf_pool_t::watch_set(): Replaces buf_pool_watch_set().
Reduce mutex contention by separating the buf_pool.watch[]
allocation and the insert into buf_pool.page_hash.
buf_pool_t::page_hash_lock<bool exclusive>(): Acquire a
buf_pool.page_hash latch.
Replaces and extends buf_page_hash_lock_s_confirm()
and buf_page_hash_lock_x_confirm().
buf_pool_t::READ_AHEAD_PAGES: Renamed from BUF_READ_AHEAD_PAGES.
buf_pool_t::curr_size, old_size, read_ahead_area, n_pend_reads:
Use Atomic_counter.
buf_pool_t::running_out(): Replaces buf_LRU_buf_pool_running_out().
buf_pool_t::LRU_remove(): Remove a block from the LRU list
and return its predecessor. Incorporates buf_LRU_adjust_hp(),
which was removed.
buf_page_get_gen(): Remove a redundant call of fsp_is_system_temporary(),
for mode == BUF_GET_IF_IN_POOL_OR_WATCH, which is only used by
BTR_DELETE_OP (purge), which is never invoked on temporary tables.
buf_free_from_unzip_LRU_list_batch(): Avoid redundant assignments.
buf_LRU_free_from_unzip_LRU_list(): Simplify the loop condition.
buf_LRU_free_page(): Clarify the function comment.
buf_flush_check_neighbor(), buf_flush_check_neighbors():
Rewrite the construction of the page hash range. We will hold
the buf_pool.mutex for up to buf_pool.read_ahead_area (at most 64)
consecutive lookups of buf_pool.page_hash.
buf_flush_page_and_try_neighbors(): Remove.
Merge to its only callers, and remove redundant operations in
buf_flush_LRU_list_batch().
buf_read_ahead_random(), buf_read_ahead_linear(): Rewrite.
Do not acquire buf_pool.mutex, and iterate directly with page_id_t.
ut_2_power_up(): Remove. my_round_up_to_next_power() is inlined
and avoids any loops.
fil_page_get_prev(), fil_page_get_next(), fil_addr_is_null(): Remove.
buf_flush_page(): Add a fil_space_t* parameter. Minimize the
buf_pool.mutex hold time. buf_pool.n_flush[] is no longer updated
atomically with the io_fix, and we will protect most buf_block_t
fields with buf_block_t::lock. The function
buf_flush_write_block_low() is removed and merged here.
buf_page_init_for_read(): Use static linkage. Initialize the newly
allocated block and acquire the exclusive buf_block_t::lock while not
holding any mutex.
IORequest::IORequest(): Remove the body. We only need to invoke
set_punch_hole() in buf_flush_page() and nowhere else.
buf_page_t::flush_type: Remove. Replaced by IORequest::flush_type.
This field is only used during a fil_io() call.
That function already takes IORequest as a parameter, so we had
better introduce for the rarely changing field.
buf_block_t::init(): Replaces buf_page_init().
buf_page_t::init(): Replaces buf_page_init_low().
buf_block_t::initialise(): Initialise many fields, but
keep the buf_page_t::state(). Both buf_pool_t::validate() and
buf_page_optimistic_get() requires that buf_page_t::in_file()
be protected atomically with buf_page_t::in_page_hash
and buf_page_t::in_LRU_list.
buf_page_optimistic_get(): Now that buf_block_t::mutex
no longer exists, we must check buf_page_t::io_fix()
after acquiring the buf_pool.page_hash lock, to detect
whether buf_page_init_for_read() has been initiated.
We will also check the io_fix() before acquiring hash_lock
in order to avoid unnecessary computation.
The field buf_block_t::modify_clock (protected by buf_block_t::lock)
allows buf_page_optimistic_get() to validate the block.
buf_page_t::real_size: Remove. It was only used while flushing
pages of page_compressed tables.
buf_page_encrypt(): Add an output parameter that allows us ot eliminate
buf_page_t::real_size. Replace a condition with debug assertion.
buf_page_should_punch_hole(): Remove.
buf_dblwr_t::add_to_batch(): Replaces buf_dblwr_add_to_batch().
Add the parameter size (to replace buf_page_t::real_size).
buf_dblwr_t::write_single_page(): Replaces buf_dblwr_write_single_page().
Add the parameter size (to replace buf_page_t::real_size).
fil_system_t::detach(): Replaces fil_space_detach().
Ensure that fil_validate() will not be violated even if
fil_system.mutex is released and reacquired.
fil_node_t::complete_io(): Renamed from fil_node_complete_io().
fil_node_t::close_to_free(): Replaces fil_node_close_to_free().
Avoid invoking fil_node_t::close() because fil_system.n_open
has already been decremented in fil_space_t::detach().
BUF_BLOCK_READY_FOR_USE: Remove. Directly use BUF_BLOCK_MEMORY.
BUF_BLOCK_ZIP_DIRTY: Remove. Directly use BUF_BLOCK_ZIP_PAGE,
and distinguish dirty pages by buf_page_t::oldest_modification().
BUF_BLOCK_POOL_WATCH: Remove. Use BUF_BLOCK_NOT_USED instead.
This state was only being used for buf_page_t that are in
buf_pool.watch.
buf_pool_t::watch[]: Remove pointer indirection.
buf_page_t::in_flush_list: Remove. It was set if and only if
buf_page_t::oldest_modification() is nonzero.
buf_page_decrypt_after_read(), buf_corrupt_page_release(),
buf_page_check_corrupt(): Change the const fil_space_t* parameter
to const fil_node_t& so that we can report the correct file name.
buf_page_monitor(): Declare as an ATTRIBUTE_COLD global function.
buf_page_io_complete(): Split to buf_page_read_complete() and
buf_page_write_complete().
buf_dblwr_t::in_use: Remove.
buf_dblwr_t::buf_block_array: Add IORequest::flush_t.
buf_dblwr_sync_datafiles(): Remove. It was a useless wrapper of
os_aio_wait_until_no_pending_writes().
buf_flush_write_complete(): Declare static, not global.
Add the parameter IORequest::flush_t.
buf_flush_freed_page(): Simplify the code.
recv_sys_t::flush_lru: Renamed from flush_type and changed to bool.
fil_read(), fil_write(): Replaced with direct use of fil_io().
fil_buffering_disabled(): Remove. Check srv_file_flush_method directly.
fil_mutex_enter_and_prepare_for_io(): Return the resolved
fil_space_t* to avoid a duplicated lookup in the caller.
fil_report_invalid_page_access(): Clean up the parameters.
fil_io(): Return fil_io_t, which comprises fil_node_t and error code.
Always invoke fil_space_t::acquire_for_io() and let either the
sync=true caller or fil_aio_callback() invoke
fil_space_t::release_for_io().
fil_aio_callback(): Rewrite to replace buf_page_io_complete().
fil_check_pending_operations(): Remove a parameter, and remove some
redundant lookups.
fil_node_close_to_free(): Wait for n_pending==0. Because we no longer
do an extra lookup of the tablespace between fil_io() and the
completion of the operation, we must give fil_node_t::complete_io() a
chance to decrement the counter.
fil_close_tablespace(): Remove unused parameter trx, and document
that this is only invoked during the error handling of IMPORT TABLESPACE.
row_import_discard_changes(): Merged with the only caller,
row_import_cleanup(). Do not lock up the data dictionary while
invoking fil_close_tablespace().
logs_empty_and_mark_files_at_shutdown(): Do not invoke
fil_close_all_files(), to avoid a !needs_flush assertion failure
on fil_node_t::close().
innodb_shutdown(): Invoke os_aio_free() before fil_close_all_files().
fil_close_all_files(): Invoke fil_flush_file_spaces()
to ensure proper durability.
thread_pool::unbind(): Fix a crash that would occur on Windows
after srv_thread_pool->disable_aio() and os_file_close().
This fix was submitted by Vladislav Vaintroub.
Thanks to Matthias Leich and Axel Schwenke for extensive testing,
Vladislav Vaintroub for helpful comments, and Eugene Kosov for a review.
Since commit 5e62b6a5e0 (MDEV-16264),
purge_sys_t::stop() no longer waited for all purge activity to stop.
This caused problems on FLUSH TABLES...FOR EXPORT because of
purge running concurrently with the buffer pool flush.
The assertion at the end of buf_flush_dirty_pages() could fail.
The, implemented by Vladislav Vaintroub, aims to eliminate race
conditions when stopping or resuming purge:
waitable_task::disable(): Wait for the task to complete, then replace
the task callback function with noop.
waitable_task::enable(): Restore the original task callback function
after disable().
purge_sys_t::stop(): Invoke purge_coordinator_task.disable().
purge_sys_t::resume(): Invoke purge_coordinator_task.enable().
purge_sys_t::running(): Add const qualifier, and clarify the comment.
The purge coordinator task will remain active as long as any purge
worker task is active.
purge_worker_callback(): Assert purge_sys.running().
srv_purge_wakeup(): Merge with the only caller purge_sys_t::resume().
purge_coordinator_task: Use static linkage.
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.
aio_linux::m_max_io_count: Unused data member; remove.
aiocb::m_ret_len: Declare as the more compatible type size_t.
Unfortunately, ssize_t is not available on Microsoft Visual Studio.
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.
