Introduce dsl_dir_diduse_transfer_space(). #12300

amotin · 2021-06-30T04:39:46Z

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time
is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling
them one after another is a waste of time and even more contention.
Doing that twice for each rewritten block within dbuf_write_done()
via dsl_dataset_block_kill() and dsl_dataset_block_born() created one
of the biggest CPU overheads in case of small blocks rewrite.

dsl_dir_diduse_transfer_space() combines functionality of these two
functions for cases where it is needed, but without double overhead,
practically for the cost of dsl_dir_diduse_space() or even cheaper.

While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space()
and dsl_dir_transfer_space(). It seems Clang detects some aliasing
there, repeating dd->dd_dbuf->db_data dereference multiple times,
increasing dd_lock scope and contention.

How Has This Been Tested?

On 80-thread FreeBSD system doing random 4KB ZVOLs writes profiler shows reduction of total active CPU cycles by 13% and reduction of system-wide locking overhead from 18.4% to 7.8% (out of which 2% inside metaslab_alloc_dva() and 3.6% on vq_lock) while keeping the same IOPS (bottlenecked by sync thread ZIO issue rate).

Types of changes

Bug fix (non-breaking change which fixes an issue)
New feature (non-breaking change which adds functionality)
Performance enhancement (non-breaking change which improves efficiency)
Code cleanup (non-breaking change which makes code smaller or more readable)
Breaking change (fix or feature that would cause existing functionality to change)
Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
Documentation (a change to man pages or other documentation)

Checklist:

My code follows the OpenZFS code style requirements.
I have updated the documentation accordingly.
I have read the contributing document.
I have added tests to cover my changes.
I have run the ZFS Test Suite with this change applied.
All commit messages are properly formatted and contain Signed-off-by.

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling them one after another is a waste of time and even more contention. Doing that twice for each rewritten block within dbuf_write_done() via dsl_dataset_block_kill() and dsl_dataset_block_born() created one of the biggest CPU overheads in case of small blocks rewrite. dsl_dir_diduse_transfer_space() combines functionality of these two functions for cases where it is needed, but without double overhead, practically for the cost of dsl_dir_diduse_space() or even cheaper. While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space() and dsl_dir_transfer_space(). It seems Clang detects some aliasing there, repeating dd->dd_dbuf->db_data dereference multiple times, increasing dd_lock scope and contention. Signed-off-by: Alexander Motin <[email protected]> Sponsored-By: iXsystems, Inc.

Commit d1d4769 takes into account the encryption key version to decide if the local_mac could be zeroed out. However, this could lead to failure mounting encrypted datasets created with intermediate versions of ZFS encryption available in master between major releases. In order to prevent this situation revert d1d4769 pending a more comprehensive fix which addresses the mount failure case. Reviewed-by: George Amanakis <[email protected]> Signed-off-by: Brian Behlendorf <[email protected]> Issue openzfs#11294 Issue openzfs#12025 Issue openzfs#12300 Closes openzfs#12033

jumbi77 · 2021-07-03T19:55:10Z

Seems like the 6 year old Illumos ticket 5551 could finally be closed :)

ahrens · 2021-07-09T03:51:49Z

module/zfs/dsl_dir.c

@@ -1517,6 +1517,11 @@ dsl_dir_diduse_space(dsl_dir_t *dd, dd_used_t type,
 {


It would be nice if dsl_dir_diduse_space() and dsl_dir_transfer_space() could be implemented as wrappers around dsl_dir_diduse_transfer_space(), rather than duplicating the code.

For dsl_dir_transfer_space() it would be a functional overkill. For dsl_dir_diduse_space() it would be good, but it would require replication of MUTEX_HELD(&dd->dd_lock) trick, which I found pretty expensive due to permanent cache miss.

Commit d1d4769 takes into account the encryption key version to decide if the local_mac could be zeroed out. However, this could lead to failure mounting encrypted datasets created with intermediate versions of ZFS encryption available in master between major releases. In order to prevent this situation revert d1d4769 pending a more comprehensive fix which addresses the mount failure case. Reviewed-by: George Amanakis <[email protected]> Signed-off-by: Brian Behlendorf <[email protected]> Issue openzfs#11294 Issue openzfs#12025 Issue openzfs#12300 Closes openzfs#12033

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling them one after another is a waste of time and even more contention. Doing that twice for each rewritten block within dbuf_write_done() via dsl_dataset_block_kill() and dsl_dataset_block_born() created one of the biggest CPU overheads in case of small blocks rewrite. dsl_dir_diduse_transfer_space() combines functionality of these two functions for cases where it is needed, but without double overhead, practically for the cost of dsl_dir_diduse_space() or even cheaper. While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space() and dsl_dir_transfer_space(). It seems Clang detects some aliasing there, repeating dd->dd_dbuf->db_data dereference multiple times, increasing dd_lock scope and contention. Reviewed-by: Brian Behlendorf <[email protected]> Reviewed-by: Matthew Ahrens <[email protected]> Author: Ryan Moeller <[email protected]> Signed-off-by: Alexander Motin <[email protected]> Sponsored-By: iXsystems, Inc. Closes openzfs#12300

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling them one after another is a waste of time and even more contention. Doing that twice for each rewritten block within dbuf_write_done() via dsl_dataset_block_kill() and dsl_dataset_block_born() created one of the biggest CPU overheads in case of small blocks rewrite. dsl_dir_diduse_transfer_space() combines functionality of these two functions for cases where it is needed, but without double overhead, practically for the cost of dsl_dir_diduse_space() or even cheaper. While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space() and dsl_dir_transfer_space(). It seems Clang detects some aliasing there, repeating dd->dd_dbuf->db_data dereference multiple times, increasing dd_lock scope and contention. Reviewed-by: Brian Behlendorf <[email protected]> Reviewed-by: Matthew Ahrens <[email protected]> Author: Ryan Moeller <[email protected]> Signed-off-by: Alexander Motin <[email protected]> Closes openzfs#12300

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling them one after another is a waste of time and even more contention. Doing that twice for each rewritten block within dbuf_write_done() via dsl_dataset_block_kill() and dsl_dataset_block_born() created one of the biggest CPU overheads in case of small blocks rewrite. dsl_dir_diduse_transfer_space() combines functionality of these two functions for cases where it is needed, but without double overhead, practically for the cost of dsl_dir_diduse_space() or even cheaper. While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space() and dsl_dir_transfer_space(). It seems Clang detects some aliasing there, repeating dd->dd_dbuf->db_data dereference multiple times, increasing dd_lock scope and contention. Reviewed-by: Brian Behlendorf <[email protected]> Reviewed-by: Matthew Ahrens <[email protected]> Author: Ryan Moeller <[email protected]> Signed-off-by: Alexander Motin <[email protected]> Closes #12300

Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling them one after another is a waste of time and even more contention. Doing that twice for each rewritten block within dbuf_write_done() via dsl_dataset_block_kill() and dsl_dataset_block_born() created one of the biggest CPU overheads in case of small blocks rewrite. dsl_dir_diduse_transfer_space() combines functionality of these two functions for cases where it is needed, but without double overhead, practically for the cost of dsl_dir_diduse_space() or even cheaper. While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space() and dsl_dir_transfer_space(). It seems Clang detects some aliasing there, repeating dd->dd_dbuf->db_data dereference multiple times, increasing dd_lock scope and contention. Reviewed-by: Brian Behlendorf <[email protected]> Reviewed-by: Matthew Ahrens <[email protected]> Author: Ryan Moeller <[email protected]> Signed-off-by: Alexander Motin <[email protected]> Closes openzfs#12300

amotin added Type: Performance Performance improvement or performance problem Status: Code Review Needed Ready for review and testing labels Jun 30, 2021

amotin requested review from behlendorf and ahrens June 30, 2021 04:39

amotin force-pushed the dsl_dir branch from fdbb43e to ad104fc Compare June 30, 2021 04:46

behlendorf approved these changes Jul 6, 2021

View reviewed changes

ghost approved these changes Jul 7, 2021

View reviewed changes

ahrens assigned mmaybee Jul 8, 2021

ahrens approved these changes Jul 9, 2021

View reviewed changes

behlendorf added Status: Accepted Ready to integrate (reviewed, tested) and removed Status: Code Review Needed Ready for review and testing labels Jul 13, 2021

mmaybee merged commit c1b5869 into openzfs:master Jul 16, 2021

amotin deleted the dsl_dir branch August 24, 2021 20:17

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Introduce dsl_dir_diduse_transfer_space(). #12300

Introduce dsl_dir_diduse_transfer_space(). #12300

amotin commented Jun 30, 2021

jumbi77 commented Jul 3, 2021

ahrens Jul 9, 2021

amotin Jul 9, 2021

		@@ -1517,6 +1517,11 @@ dsl_dir_diduse_space(dsl_dir_t *dd, dd_used_t type,
		{

Introduce dsl_dir_diduse_transfer_space(). #12300

Introduce dsl_dir_diduse_transfer_space(). #12300

Conversation

amotin commented Jun 30, 2021

How Has This Been Tested?

Types of changes

Checklist:

jumbi77 commented Jul 3, 2021

ahrens Jul 9, 2021

Choose a reason for hiding this comment

amotin Jul 9, 2021

Choose a reason for hiding this comment