Re: [PATCH V3 10/25] smartpqi: add stream detection

[Martin Wilck <mwilck@suse.com>]

On Thu, 2020-12-10 at 14:35 -0600, Don Brace wrote:
> * Enhance performance by adding sequential stream detection.
>   for R5/R6 sequential write requests.
>   * Reduce stripe lock contention with full-stripe write
>     operations.

I suppose that "stripe lock" is used by the firmware? Could you
elaborate a bit more how this technique improves performance?

> 
> Reviewed-by: Scott Benesh <scott.benesh@microchip.com>
> Reviewed-by: Scott Teel <scott.teel@microchip.com>
> Signed-off-by: Don Brace <don.brace@microchip.com>
> ---
>  drivers/scsi/smartpqi/smartpqi.h      |    8 +++
>  drivers/scsi/smartpqi/smartpqi_init.c |   87
> +++++++++++++++++++++++++++++++--
>  2 files changed, 89 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/scsi/smartpqi/smartpqi.h
> b/drivers/scsi/smartpqi/smartpqi.h
> index a5e271dd2742..343f06e44220 100644
> --- a/drivers/scsi/smartpqi/smartpqi.h
> +++ b/drivers/scsi/smartpqi/smartpqi.h
> @@ -1042,6 +1042,12 @@ struct pqi_scsi_dev_raid_map_data {
>  
>  #define RAID_CTLR_LUNID                "\0\0\0\0\0\0\0\0"
>  
> +#define NUM_STREAMS_PER_LUN    8
> +
> +struct pqi_stream_data {
> +       u64     next_lba;
> +       u32     last_accessed;
> +};
>  
>  struct pqi_scsi_dev {
>         int     devtype;                /* as reported by INQUIRY
> commmand */
> @@ -1097,6 +1103,7 @@ struct pqi_scsi_dev {
>         struct list_head add_list_entry;
>         struct list_head delete_list_entry;
>  
> +       struct pqi_stream_data stream_data[NUM_STREAMS_PER_LUN];
>         atomic_t scsi_cmds_outstanding;
>         atomic_t raid_bypass_cnt;
>  };
> @@ -1296,6 +1303,7 @@ struct pqi_ctrl_info {
>         u8              enable_r5_writes : 1;
>         u8              enable_r6_writes : 1;
>         u8              lv_drive_type_mix_valid : 1;
> +       u8              enable_stream_detection : 1;
>  
>         u8              ciss_report_log_flags;
>         u32             max_transfer_encrypted_sas_sata;
> diff --git a/drivers/scsi/smartpqi/smartpqi_init.c
> b/drivers/scsi/smartpqi/smartpqi_init.c
> index fc8fafab480d..96383d047a88 100644
> --- a/drivers/scsi/smartpqi/smartpqi_init.c
> +++ b/drivers/scsi/smartpqi/smartpqi_init.c
> @@ -5721,8 +5721,82 @@ void pqi_prep_for_scsi_done(struct scsi_cmnd
> *scmd)
>         atomic_dec(&device->scsi_cmds_outstanding);
>  }
>  
> -static int pqi_scsi_queue_command(struct Scsi_Host *shost,
> +static bool pqi_is_parity_write_stream(struct pqi_ctrl_info
> *ctrl_info,
>         struct scsi_cmnd *scmd)
> +{
> +       u32 oldest_jiffies;
> +       u8 lru_index;
> +       int i;
> +       int rc;
> +       struct pqi_scsi_dev *device;
> +       struct pqi_stream_data *pqi_stream_data;
> +       struct pqi_scsi_dev_raid_map_data rmd;
> +
> +       if (!ctrl_info->enable_stream_detection)
> +               return false;
> +
> +       rc = pqi_get_aio_lba_and_block_count(scmd, &rmd);
> +       if (rc)
> +               return false;
> +
> +       /* Check writes only. */
> +       if (!rmd.is_write)
> +               return false;
> +
> +       device = scmd->device->hostdata;
> +
> +       /* Check for RAID 5/6 streams. */
> +       if (device->raid_level != SA_RAID_5 && device->raid_level !=
> SA_RAID_6)
> +               return false;
> +
> +       /*
> +        * If controller does not support AIO RAID{5,6} writes, need
> to send
> +        * requests down non-AIO path.
> +        */
> +       if ((device->raid_level == SA_RAID_5 && !ctrl_info-
> >enable_r5_writes) ||
> +               (device->raid_level == SA_RAID_6 && !ctrl_info-
> >enable_r6_writes))
> +               return true;
> +
> +       lru_index = 0;
> +       oldest_jiffies = INT_MAX;
> +       for (i = 0; i < NUM_STREAMS_PER_LUN; i++) {
> +               pqi_stream_data = &device->stream_data[i];
> +               /*
> +                * Check for adjacent request or request is within
> +                * the previous request.
> +                */
> +               if ((pqi_stream_data->next_lba &&
> +                       rmd.first_block >= pqi_stream_data->next_lba)
> &&
> +                       rmd.first_block <= pqi_stream_data->next_lba
> +
> +                               rmd.block_cnt) {

Here you seem to assume that the previous write had the same block_cnt.
What's the justification for that?

> +                       pqi_stream_data->next_lba = rmd.first_block +
> +                               rmd.block_cnt;
> +                       pqi_stream_data->last_accessed = jiffies;
> +                       return true;
> +               }
> +
> +               /* unused entry */
> +               if (pqi_stream_data->last_accessed == 0) {
> +                       lru_index = i;
> +                       break;
> +               }
> +
> +               /* Find entry with oldest last accessed time. */
> +               if (pqi_stream_data->last_accessed <= oldest_jiffies)
> {
> +                       oldest_jiffies = pqi_stream_data-
> >last_accessed;
> +                       lru_index = i;
> +               }
> +       }
> +
> +       /* Set LRU entry. */
> +       pqi_stream_data = &device->stream_data[lru_index];
> +       pqi_stream_data->last_accessed = jiffies;
> +       pqi_stream_data->next_lba = rmd.first_block + rmd.block_cnt;
> +
> +       return false;
> +}
> +
> +static int pqi_scsi_queue_command(struct Scsi_Host *shost, struct
> scsi_cmnd *scmd)
>  {
>         int rc;
>         struct pqi_ctrl_info *ctrl_info;
> @@ -5768,11 +5842,12 @@ static int pqi_scsi_queue_command(struct
> Scsi_Host *shost,
>                 raid_bypassed = false;
>                 if (device->raid_bypass_enabled &&
>                         !blk_rq_is_passthrough(scmd->request)) {
> -                       rc =
> pqi_raid_bypass_submit_scsi_cmd(ctrl_info, device,
> -                               scmd, queue_group);
> -                       if (rc == 0 || rc == SCSI_MLQUEUE_HOST_BUSY)
> {
> -                               raid_bypassed = true;
> -                               atomic_inc(&device->raid_bypass_cnt);
> +                       if (!pqi_is_parity_write_stream(ctrl_info,
> scmd)) {
> +                               rc =
> pqi_raid_bypass_submit_scsi_cmd(ctrl_info, device, scmd,
> queue_group);
> +                               if (rc == 0 || rc ==
> SCSI_MLQUEUE_HOST_BUSY) {
> +                                       raid_bypassed = true;
> +                                       atomic_inc(&device-
> >raid_bypass_cnt);
> +                               }
>                         }
>                 }
>                 if (!raid_bypassed)
>