[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-Id: <20200206152408.24165-1-longman@redhat.com>
Date: Thu, 6 Feb 2020 10:24:02 -0500
From: Waiman Long <longman@...hat.com>
To: Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...hat.com>,
Will Deacon <will.deacon@....com>
Cc: linux-kernel@...r.kernel.org, Bart Van Assche <bvanassche@....org>,
Waiman Long <longman@...hat.com>
Subject: [PATCH v6 0/6] locking/lockdep: Reuse zapped chain_hlocks entries
v6:
- Drop patch 7 for now.
- Incorporate further changes as suggested by PeterZ.
- Increase MAX_CHAIN_BUCKETS to 16 to further minimize the possibility
of having more than one chain block in bucket 0.
v5:
- Fix more build errors reported by kbuild test robot.
v4:
- Fix build errors reported by kbuild test robot.
- Adopt the single chain block allocator code suggested by PeterZ which
combine the last 3 patches of v3 series.
- Add another patch to introduce a fast path in the chain block
allocator.
- In patch 1, move the inc_chains() out of CONFIG_TRACE_IRQFLAGS
conditional compilation block.
v3:
- Move the bug fix patches to the beginning of the series.
- Include a number of changes as suggested by PeterZ.
- Increase MAX_CHAIN_BUCKETS from 8 to 10 to reduce the chance of using
the unsized list.
- Add patch 7 to add a lockdep_early_init() call.
- Add patch 8 to allocate chain hlocks by splitting large chain block
as a last resort.
v2:
- Revamp the chain_hlocks reuse patch to store the freed chain_hlocks
information in the chain_hlocks entries themselves avoiding the
need of a separate set of tracking structures. This, however,
requires a minimum allocation size of at least 2. Thanks to PeterZ
for his review and inspiring this change.
- Remove the leakage counter as it is no longer applicable.
- Add patch 6 to make the output of /proc/lockdep_chains more readable.
It was found that when running a workload that kept on adding lock
classes and then zapping them repetitively, the system will eventually
run out of chain_hlocks[] entries even though there were still plenty
of other lockdep data buffers available.
[ 4318.443670] BUG: MAX_LOCKDEP_CHAIN_HLOCKS too low!
[ 4318.444809] turning off the locking correctness validator.
In order to fix this problem, we have to make chain_hlocks[] entries
reusable just like other lockdep arrays. Besides that, the patchset
also adds some zapped class and chain_hlocks counters to be tracked by
/proc/lockdep_stats. It also fixes leakage in the irq context counters
and makes the output of /proc/lockdep_chains more readable.
Waiman Long (6):
locking/lockdep: Decrement irq context counters when removing lock
chain
locking/lockdep: Display irq_context names in /proc/lockdep_chains
locking/lockdep: Track number of zapped classes
locking/lockdep: Throw away all lock chains with zapped class
locking/lockdep: Track number of zapped lock chains
locking/lockdep: Reuse freed chain_hlocks entries
kernel/locking/lockdep.c | 332 ++++++++++++++++++++++++-----
kernel/locking/lockdep_internals.h | 14 +-
kernel/locking/lockdep_proc.c | 31 ++-
3 files changed, 312 insertions(+), 65 deletions(-)
--
2.18.1
Powered by blists - more mailing lists