On 19.01.22 16:15, David Hildenbrand wrote: > On 19.01.22 16:08, Boqun Feng wrote: >> Hi, >> >> On Wed, Jan 19, 2022 at 12:37:02PM +0100, David Hildenbrand wrote: >>> Lockdep complains that we do during mmap of the vmcore: >>> down_write(mmap_lock); >>> down_read(vmcore_cb_rwsem); >>> And during read of the vmcore: >>> down_read(vmcore_cb_rwsem); >>> down_read(mmap_lock); >>> >>> We cannot possibly deadlock when only taking vmcore_cb_rwsem in read >>> mode, however, it's hard to teach that to lockdep. >>> >> >> Lockdep warned about the above sequences because rw_semaphore is a fair >> read-write lock, and the following can cause a deadlock: >> >> TASK 1 TASK 2 TASK 3 >> ====== ====== ====== >> down_write(mmap_lock); >> down_read(vmcore_cb_rwsem) >> down_write(vmcore_cb_rwsem); // blocked >> down_read(vmcore_cb_rwsem); // cannot get the lock because of the fairness >> down_read(mmap_lock); // blocked >> >> IOW, a reader can block another read if there is a writer queued by the >> second reader and the lock is fair. >> >> So there is a deadlock possiblity. > > Task 3 will never take the mmap_lock before doing a > down_write(vmcore_cb_rwsem). > > How would this happen? Ah, I get it, nevermind. I'll adjust the patch description. Thanks! -- Thanks, David / dhildenb
