Bug: fseg_get_pages_info does not release pages during extent iteration
Problem
In fseg_get_pages_info() (storage/innobase/fsp/fsp0fsp.cc), when scanning the FSEG_NOT_FULL and FSEG_FULL extent lists for a segment, the mtr holds latches on every visited page descriptor (xdes) without releasing them until the entire function completes and mtr_commit() is called.
For tables with a large number of extents (large tables or tables with fragmented segments), this can mean holding hundreds or thousands of page latches simultaneously within a single mini-transaction. This causes:
- Excessive buffer pool frame consumption (each latched page pins a buffer pool frame)
- Increased memory pressure and potential OOM
CHECK INDEX on large tables becomes unreliable (may OOM)
Root Cause
The mtr_t memo stack grows unbounded during extent list traversal in fseg_get_pages_info(). Each call to xdes_lst_get_descriptor() and xdes_mtr_get_bit() adds a page latch to the mtr memo, but these latches are only released at mtr_commit() which happens at the very end of the function.
Fix
Add a new function mtr_t::release_all_after_savepoint() that releases all pages held in the mtr memo after a given savepoint. Then in fseg_get_pages_info(), set a savepoint and release pages every 100 iterations, effectively bounding the number of simultaneously held latches.
Changes
-
storage/innobase/include/mtr0mtr.h: Declare release_all_after_savepoint(ulint savepoint) method and mtr_release_all_after_savepoint() macro.
-
storage/innobase/mtr/mtr0mtr.cc: Implement Release_all_after_savepoint functor and mtr_t::release_all_after_savepoint() method.
-
storage/innobase/fsp/fsp0fsp.cc: In fseg_get_pages_info(), add savepoint and release logic every 100 pages during FSEG_NOT_FULL and FSEG_FULL list scans.
Affected Scenarios
CHECK INDEX on large tables — ha_innobase::check_index() → innobase_get_index_status() → fseg_get_pages_info() for both leaf and top segments of each index
Bug: fseg_get_pages_info does not release pages during extent iteration
Problem
In
fseg_get_pages_info()(storage/innobase/fsp/fsp0fsp.cc), when scanning the FSEG_NOT_FULL and FSEG_FULL extent lists for a segment, the mtr holds latches on every visited page descriptor (xdes) without releasing them until the entire function completes andmtr_commit()is called.For tables with a large number of extents (large tables or tables with fragmented segments), this can mean holding hundreds or thousands of page latches simultaneously within a single mini-transaction. This causes:
CHECK INDEXon large tables becomes unreliable (may OOM)Root Cause
The
mtr_tmemo stack grows unbounded during extent list traversal infseg_get_pages_info(). Each call toxdes_lst_get_descriptor()andxdes_mtr_get_bit()adds a page latch to the mtr memo, but these latches are only released atmtr_commit()which happens at the very end of the function.Fix
Add a new function
mtr_t::release_all_after_savepoint()that releases all pages held in the mtr memo after a given savepoint. Then infseg_get_pages_info(), set a savepoint and release pages every 100 iterations, effectively bounding the number of simultaneously held latches.Changes
storage/innobase/include/mtr0mtr.h: Declarerelease_all_after_savepoint(ulint savepoint)method andmtr_release_all_after_savepoint()macro.storage/innobase/mtr/mtr0mtr.cc: ImplementRelease_all_after_savepointfunctor andmtr_t::release_all_after_savepoint()method.storage/innobase/fsp/fsp0fsp.cc: Infseg_get_pages_info(), add savepoint and release logic every 100 pages during FSEG_NOT_FULL and FSEG_FULL list scans.Affected Scenarios
CHECK INDEXon large tables —ha_innobase::check_index()→innobase_get_index_status()→fseg_get_pages_info()for both leaf and top segments of each index