On Mon, Feb 28, 2022 at 06:56:39PM +0900, Byungchul Park wrote: > I didn't want to bother you so I was planning to send the next spin > after making more progress. However, PATCH v2 reports too many false > positives because Dept tracked the bit_wait_table[] wrong way - I > apologize for that. So I decided to send PATCH v3 first before going > further for those who want to run Dept for now. > > There might still be some false positives but not overwhelming. > Hello Byungchul, I'm running DEPT v3 on my system and I see report below. Looking at the kmemleak code and comment, I think kmemleak tried to avoid lockdep recursive warning but detected by DEPT? =================================================== DEPT: Circular dependency has been detected. 5.17.0-rc1+ #1 Tainted: G W --------------------------------------------------- summary --------------------------------------------------- *** AA DEADLOCK *** context A [S] __raw_spin_lock_irqsave(&object->lock:0) [W] _raw_spin_lock_nested(&object->lock:0) [E] spin_unlock(&object->lock:0) [S]: start of the event context [W]: the wait blocked [E]: the event not reachable --------------------------------------------------- context A's detail --------------------------------------------------- context A [S] __raw_spin_lock_irqsave(&object->lock:0) [W] _raw_spin_lock_nested(&object->lock:0) [E] spin_unlock(&object->lock:0) [S] __raw_spin_lock_irqsave(&object->lock:0): [<ffffffc00810302c>] scan_gray_list+0x84/0x13c stacktrace: dept_ecxt_enter+0x88/0xf4 _raw_spin_lock_irqsave+0xf0/0x1c4 scan_gray_list+0x84/0x13c kmemleak_scan+0x2d8/0x54c kmemleak_scan_thread+0xac/0xd4 kthread+0xd4/0xe4 ret_from_fork+0x10/0x20 [E] spin_unlock(&object->lock:0): [<ffffffc008102ee0>] scan_block+0x60/0x128 --------------------------------------------------- information that might be helpful --------------------------------------------------- CPU: 1 PID: 38 Comm: kmemleak Tainted: G W 5.17.0-rc1+ #1 Hardware name: linux,dummy-virt (DT) Call trace: dump_backtrace.part.0+0x9c/0xc4 show_stack+0x14/0x28 dump_stack_lvl+0x9c/0xcc dump_stack+0x14/0x2c print_circle+0x2d4/0x438 cb_check_dl+0x44/0x70 bfs+0x60/0x168 add_dep+0x88/0x11c add_wait+0x2d0/0x2dc __dept_wait+0x8c/0xa4 dept_wait+0x6c/0x88 _raw_spin_lock_nested+0xa8/0x1b0 scan_block+0xb4/0x128 scan_gray_list+0xc4/0x13c kmemleak_scan+0x2d8/0x54c kmemleak_scan_thread+0xac/0xd4 kthread+0xd4/0xe4 ret_from_fork+0x10/0x20 > --- > > Hi Linus and folks, > > I've been developing a tool for detecting deadlock possibilities by > tracking wait/event rather than lock(?) acquisition order to try to > cover all synchonization machanisms. It's done on v5.17-rc1 tag. > > https://github.com/lgebyungchulpark/linux-dept/commits/dept1.14_on_v5.17-rc1 > [...] > Benifit: > > 0. Works with all lock primitives. > 1. Works with wait_for_completion()/complete(). > 2. Works with 'wait' on PG_locked. > 3. Works with 'wait' on PG_writeback. > 4. Works with swait/wakeup. > 5. Works with waitqueue. > 6. Multiple reports are allowed. > 7. Deduplication control on multiple reports. > 8. Withstand false positives thanks to 6. > 9. Easy to tag any wait/event. > > Future work: > > 0. To make it more stable. > 1. To separates Dept from Lockdep. > 2. To improves performance in terms of time and space. > 3. To use Dept as a dependency engine for Lockdep. > 4. To add any missing tags of wait/event in the kernel. > 5. To deduplicate stack trace. > > How to interpret reports: > > 1. E(event) in each context cannot be triggered because of the > W(wait) that cannot be woken. > 2. The stack trace helping find the problematic code is located > in each conext's detail. > > Thanks, > Byungchul > > --- > > Changes from v2: > > 1. Disable Dept on bit_wait_table[] in sched/wait_bit.c > reporting a lot of false positives, which is my fault. > Wait/event for bit_wait_table[] should've been tagged in a > higher layer for better work, which is a future work. > (feedback from Jan Kara) > 2. Disable Dept on crypto_larval's completion to prevent a false > positive. > > Changes from v1: > > 1. Fix coding style and typo. (feedback from Steven) > 2. Distinguish each work context from another in workqueue. > 3. Skip checking lock acquisition with nest_lock, which is about > correct lock usage that should be checked by Lockdep. > > Changes from RFC: > > 1. Prevent adding a wait tag at prepare_to_wait() but __schedule(). > (feedback from Linus and Matthew) > 2. Use try version at lockdep_acquire_cpus_lock() annotation. > 3. Distinguish each syscall context from another. [ ... ] -- Thank you, You are awesome! Hyeonggon :-)