On Mon, Apr 25, 2016 at 6:18 PM, Yu Zhao <yuz...@google.com> wrote: > On Mon, Apr 25, 2016 at 05:20:10PM -0400, Dan Streetman wrote: >> Add a work_struct to struct zpool, and change zpool_destroy_pool to >> defer calling the pool implementation destroy. >> >> The zsmalloc pool destroy function, which is one of the zpool >> implementations, may sleep during destruction of the pool. However >> zswap, which uses zpool, may call zpool_destroy_pool from atomic >> context. So we need to defer the call to the zpool implementation >> to destroy the pool. >> >> This is essentially the same as Yu Zhao's proposed patch to zsmalloc, >> but moved to zpool. > > Thanks, Dan. Sergey also mentioned another call path that triggers the > same problem (BUG: scheduling while atomic): > rcu_process_callbacks() > __zswap_pool_release() > zswap_pool_destroy() > zswap_cpu_comp_destroy() > cpu_notifier_register_begin() > mutex_lock(&cpu_add_remove_lock); > So I was thinking zswap_pool_destroy() might be done in workqueue in zswap.c. > This way we fix both call paths.
Yes, you're right, I took so long to get around to this I forgot the details :-) I'll send a new patch to zswap. > > Or you have another patch to fix the second call path?
