Hi James,
> On Nov 7, 2017, at 9:05 AM, James Smart <james.smart@xxxxxxxxxxxx> wrote:
>
> On 11/6/2017 11:55 AM, Himanshu Madhani wrote:
>> From: Anil Gurumurthy <anil.gurumurthy@xxxxxxxxxx>
>>
>> +static struct nvmet_fc_target_template qla_nvmet_fc_transport = {
>> + .targetport_delete = qla_nvmet_targetport_delete,
>> + .xmt_ls_rsp = qla_nvmet_ls_rsp,
>> + .fcp_op = qla_nvmet_fcp_op,
>> + .fcp_abort = qla_nvmet_fcp_abort,
>> + .fcp_req_release = qla_nvmet_fcp_req_release,
>> + .max_hw_queues = 8,
>> + .max_sgl_segments = 128,
>> + .max_dif_sgl_segments = 64,
>> + .dma_boundary = 0xFFFFFFFF,
>> + .target_features = NVMET_FCTGTFEAT_READDATA_RSP |
>> + NVMET_FCTGTFEAT_CMD_IN_ISR |
>> + NVMET_FCTGTFEAT_OPDONE_IN_ISR,
>> + .target_priv_sz = sizeof(struct nvme_private),
>> +};
>> +#endif
>>
>
> Do you really need the xxx_IN_ISR features ? e.g. are you calling the nvmet_fc cmd receive and op done calls in ISR context or something that can't/shouldn't continue into the nvmet layers ?
>
> I was looking to remove those flags and the work_q items behind it as I believed the qla2xxx driver called everything in a deferred callback.
>
Agree we do nvme_fc* callbacks in deferred context, but without the xxx_IN_ISR flag during NVMe Target template registration, we were running into crash due to recursive spin_lock held as part of CTIO response in our driver.
> -- james
>
>
Thanks,
- Himanshu
