MDEV-34431: Avoid spin loops on page I/O waits #4172
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
|
5 tasks
dr-m
commented
Jul 8, 2025
There was a problem hiding this comment.
I may revise this further to make the I/O wait reporting more consistent. I wanted to publish this as is to enable some performance testing as early as possible.
One thing that is outside the scope of this is waits for page writes, in case a user thread is requesting write access to a buffer page that is currently being written back to the data file. This could be addressed separately later.
dr-m
commented
Jul 10, 2025
c32d5d3 to
23f973d
Compare
5 tasks
2b3806e to
f760918
Compare
The Microsoft Windows implementations of SRWLOCK and WaitOnAddress() include some spin loop logic before entering the operating system kernel. Let us avoid duplicating some of that spin loop logic. Thanks to Vladislav Vaintroub for this fix.
While waiting for for I/O completion, let us skip spin loops. Even on fast storage, reading a page into the buffer pool takes so long that a spin loop would only end up wasting CPU time. block_lock::s_lock_nospin(): A no-spin variant of acquiring a shared buffer page latch. Regular s_lock() always involveis a spin loop. ssux_lock_impl::rd_lock_spin(), ssux_lock_impl::rd_lock_nospin(): Split from rd_wait(). ssux_lock_impl::rd_lock(): Invoke either rd_lock_nospin() or rd_lock_try() and rd_lock_spin(). buf_page_get_low(): After acquiring a page latch on an io-fixed block, try to optimize operations on the page latch.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
While waiting for for I/O completion, let us skip spin loops. Even on fast storage, reading a page into the buffer pool takes so long time that wasting CPU in a spin loop does not make any sense.
block_lock::s_lock_nospin(): A no-spin variant of acquiring a shared buffer page latch. Regulars_lock()always involveis a spin loop.ssux_lock_impl::rd_lock_spin(),ssux_lock_impl::rd_lock_nospin(): Split fromrd_wait().ssux_lock_impl::rd_lock(): Invoke eitherrd_lock_nospin()orrd_lock_try()andrd_lock_spin().The above changes are also part of #4168.
buf_read_page(): Return a pointer to a buffer-fixed, non-read-fixed page, ornullptrin case of an error.buf_inc_get(): Remove the parameter.IORequest::read_complete(): Assert that the page is both read-fixed and buffer-fixed. Samplerecv_sys.recovery_ononly once. Buffer-unfix the page when the asynchronous read completes.buf_page_t::read_complete(): Assert that the page is both read-fixed and buffer-fixed.buf_page_init_for_read(): Return a pointer to a buffer-fixed block descriptor pointer, bitwise-ORed with 1 in case the block already exists in the buffer pool.buf_read_ahead_update(),buf_read_ahead_update_sql(): Common code for updating some statistics counters.buf_read_page_low(): Replace the parametersyncwitherr, which will return an error code to a synchronous caller. Add a parameter for thread-localmariadb_stats. Return the pointer to the block, or the special valuesnullptr(read failure) or-1or-2for asynchronous reads. Increment the statistics when a synchronous read was requested. In a synchronous read, if the page has already been allocated in the buffer pool but it is read-fixed, wait for the read to complete.buf_read_page_background(): Update the statistics if supplied.Release Notes
When a buffer pool lookup needs to wait for page I/O, InnoDB will avoid excessive CPU usage.
How can this PR be tested?
Sysbench or HammerDB on a workload that is larger than the buffer pool. This should be most prominent when there are synchronous waits for pages in the buffer pool.
Measuring the CPU usage while running the test case in MDEV-32067 should show some impact.
Basing the PR against the correct MariaDB version
mainbranch.The second change (refactoring
buf_read_page()) depends on earlier changes f27e9c8 and d6aed21, and 11.4 is the earliest maintained branch that contains them.PR quality check