[HPDD-discuss] Error Handle in ldlm_cli_cancel_req()

Hi, All. I want you to discuss an issue about an error handle in ldlm_cli_cancel_req(). According to the code of Lustre-2.3.64, ldlm_cli_cancel_req() never re-send ldlm_cancel req when it has got -EAGAIN from ptlrpc_queue_wait(), and ptlrpc_queue_wait() return -EAGAIN when a state of import object which is target of the ldlm_cancel req is in recovery states such as DISCON, CONNECTING etc. And, In my experience, it can happen so often, especially in a large-scale system. Which is why, I suggest that ldlm_cli_cancel_req() resend ldlm_cancel request when getting -EAGAIN. Because, if not, the client which failed to send ldlm_cancel req will be evicted from a server when the server sent blocking callback req to the client. If someone agree my idea, I'd like to make a ticket for the issue on Jira. Best regard. ----------------------------------- Hiroya Nozaki nozaki.hiroya(a)jp.fujitsu.com Next Generation Technical Computing Unit Fujitsu, Ltd Tel: 044-754-8769 Ext: 7103-8594 ----------------------------------- Hiroya Nozaki nozaki.hiroya(a)jp.fujitsu.com Next Generation Technical Computing Unit Fujitsu, Ltd Tel: 044-754-8769 Ext: 7103-8594