Re: [PATCH v9 13/13] lpfc: vmid: Introducing vmid in io path.

[James Smart <jsmart2021@gmail.com>]

On 4/8/2021 1:46 AM, Hannes Reinecke wrote:
> On 4/7/21 1:06 AM, Muneendra wrote:
>> From: Gaurav Srivastava <gaurav.srivastava@broadcom.com>
...
>> +        /* while the read lock was released, in case the entry was */
>> +        /* added by other context or is in process of being added */
>> +        if (vmp && vmp->flag & LPFC_VMID_REGISTERED) {
>> +            lpfc_vmid_update_entry(vport, cmd, vmp, tag);
>> +            write_unlock(&vport->vmid_lock);
>> +            return 0;
>> +        } else if (vmp && vmp->flag & LPFC_VMID_REQ_REGISTER) {
>> +            write_unlock(&vport->vmid_lock);
>> +            return -EBUSY;
>> +        }
>> +
>> +        /* else search and allocate a free slot in the hash table */
>> +        if (vport->cur_vmid_cnt < vport->max_vmid) {
>> +            for (i = 0; i < vport->max_vmid; ++i) {
>> +                vmp = vport->vmid + i;
>> +                if (vmp->flag == LPFC_VMID_SLOT_FREE) {
>> +                    vmp = vport->vmid + i;

delete this last line and adjust parens - really odd that it replicates 
the assignment 2 lines earlier.

>> +                    break;
>> +                }
>> +            }

I would prefer that if the table is expended and no slots free, that 
-ENOMEM is returned here. Rather than falling down below and qualifying 
by slot free, then by pending (set only if slot free).  I can't believe 
there is a reason the idle timer has to be started if no slots free as 
all the other fail cases don't bother with it.

This also helps indentation levels below.

>> +        } else {
>> +            write_unlock(&vport->vmid_lock);
>> +            return -ENOMEM;
>> +        }
>> +
>> +        if (vmp && (vmp->flag == LPFC_VMID_SLOT_FREE)) {
>> +            /* Add the vmid and register  */
>> +            lpfc_put_vmid_in_hashtable(vport, hash, vmp);
>> +            vmp->vmid_len = len;
>> +            memcpy(vmp->host_vmid, uuid, vmp->vmid_len);
>> +            vmp->io_rd_cnt = 0;
>> +            vmp->io_wr_cnt = 0;
>> +            vmp->flag = LPFC_VMID_SLOT_USED;
>> +
>> +            vmp->delete_inactive =
>> +                vport->vmid_inactivity_timeout ? 1 : 0;
>> +
>> +            /* if type priority tag, get next available vmid */
>> +            if (lpfc_vmid_is_type_priority_tag(vport))
>> +                lpfc_vmid_assign_cs_ctl(vport, vmp);
>> +
>> +            /* allocate the per cpu variable for holding */
>> +            /* the last access time stamp only if vmid is enabled */
>> +            if (!vmp->last_io_time)
>> +                vmp->last_io_time =
>> +                    __alloc_percpu(sizeof(u64),
>> +                           __alignof__(struct
>> +                                   lpfc_vmid));
>> +
>> +            /* registration pending */
>> +            pending = 1;
>> +        } else {
>> +            rc = -ENOMEM;
>> +        }
>> +        write_unlock(&vport->vmid_lock);
>> +
>> +        /* complete transaction with switch */
>> +        if (pending) {
>> +            if (lpfc_vmid_is_type_priority_tag(vport))
>> +                rc = lpfc_vmid_uvem(vport, vmp, true);
>> +            else
>> +                rc = lpfc_vmid_cmd(vport,
>> +                           SLI_CTAS_RAPP_IDENT,
>> +                           vmp);
>> +            if (!rc) {
>> +                write_lock(&vport->vmid_lock);
>> +                vport->cur_vmid_cnt++;
>> +                vmp->flag |= LPFC_VMID_REQ_REGISTER;
>> +                write_unlock(&vport->vmid_lock);
>> +            }
>> +        }
>> +
>> +        /* finally, enable the idle timer once */
>> +        if (!(vport->phba->pport->vmid_flag & LPFC_VMID_TIMER_ENBLD)) {
>> +            mod_timer(&vport->phba->inactive_vmid_poll,
>> +                  jiffies +
>> +                  msecs_to_jiffies(1000 * LPFC_VMID_TIMER));
>> +            vport->phba->pport->vmid_flag |= LPFC_VMID_TIMER_ENBLD;
>> +        }
>> +    }
>> +    return rc;
>> +}
>> +
>> +/*
>> + * lpfc_is_command_vm_io - get the uuid from blk cgroup
>> + * @cmd:Pointer to scsi_cmnd data structure
>> + * Returns uuid if present if not null
>> + */
>> +static char *lpfc_is_command_vm_io(struct scsi_cmnd *cmd)
>> +{
>> +    char *uuid = NULL;
>> +
>> +    if (cmd->request) {
>> +        if (cmd->request->bio)
>> +            uuid = blkcg_get_fc_appid(cmd->request->bio);
>> +    }
>> +    return uuid;
>> +}
>> +
>>   /**
>>    * lpfc_queuecommand - scsi_host_template queuecommand entry point
>>    * @shost: kernel scsi host pointer.
>> @@ -5288,6 +5437,7 @@ lpfc_queuecommand(struct Scsi_Host *shost, 
>> struct scsi_cmnd *cmnd)
>>       int err, idx;
>>   #ifdef CONFIG_SCSI_LPFC_DEBUG_FS
>>       uint64_t start = 0L;
>> +    u8 *uuid = NULL;
>>       if (phba->ktime_on)
>>           start = ktime_get_ns();
>> @@ -5415,6 +5565,25 @@ lpfc_queuecommand(struct Scsi_Host *shost, 
>> struct scsi_cmnd *cmnd)
>>       }
>> +    /* check the necessary and sufficient condition to support VMID */
>> +    if (lpfc_is_vmid_enabled(phba) &&
>> +        (ndlp->vmid_support ||
>> +         phba->pport->vmid_priority_tagging ==
>> +         LPFC_VMID_PRIO_TAG_ALL_TARGETS)) {
>> +        /* is the IO generated by a VM, get the associated virtual */
>> +        /* entity id */
>> +        uuid = lpfc_is_command_vm_io(cmnd);
>> +
>> +        if (uuid) {
>> +            err = lpfc_vmid_get_appid(vport, uuid, cmnd,
>> +                (union lpfc_vmid_io_tag *)
>> +                    &lpfc_cmd->cur_iocbq.vmid_tag);
>> +            if (!err)
>> +                lpfc_cmd->cur_iocbq.iocb_flag |= LPFC_IO_VMID;
>> +        }
>> +    }
>> +
>> +    atomic_inc(&ndlp->cmd_pending);
>>   #ifdef CONFIG_SCSI_LPFC_DEBUG_FS
>>       if (unlikely(phba->hdwqstat_on & LPFC_CHECK_SCSI_IO))
>>           this_cpu_inc(phba->sli4_hba.c_stat->xmt_io);
>>
> And that's the bit which I don't particular like.
> 
> Essentially we'll have to inject additional ELS commands _on each I/O_ 
> to get a valid VMID.
> Where there are _so_ many things which might get wrong, causing an I/O 
> stall.

I don't follow you - yes ELS's are injected when there isn't an entry 
for the VM, but once there is, there isn't further ELS's. That is the 
cost. as we don't know what uuid's I/O will be sent to before hand, so 
it has to be siphoned off during the I/O flow.  I/O's can be sent 
non-tagged while the ELS's are completing (and there aren't multiple 
sets of ELS's as long as it's the same uuid), which is fine.

so I disagree with "_on each I/O_".

> I would have vastly preferred if we could _avoid_ having to do 
> additional ELS commands for VMID registration in the I/O path
> (ie only allow for I/O with a valid VMID), and reject the I/O otherwise 
> until VMID registration is complete.
> 
> IE return 'BUSY' (or even a command retry?) when no valid VMID for this 
> particular I/O is found, register the VMID (preferably in another 
> thread), and restart the queue once the VMID is registered.

Why does it bother you with the I/O path ?  It's actually happening in 
parallel with no real relation between the two.

I seriously disagree with reject if no vmid tag. Why?  what do you gain 
? This actually introduces more disruption than the parallel flow with 
the ELSs.   Also, after link bounce, where all VMID's have to be done, 
it adds a stall window after link up right when things are trying to 
resume after rports rejoin. Why add the i/o rebouncing ? There no real 
benefit. Issuing a few untagged I/O doesn't hurt.

> 
> That way we have a clear separation, and the I/O path will always work 
> with valid VMIDs.
> 
> Hmm?
> 
> Cheers,
> 
> Hannes

-- james


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme