Re: Re: Re: [PATCH] scsi: ufs: qcom: avoid re-init quirk when gears match

[Eric Chanudet]

On Fri, Jan 19, 2024 at 02:33:32PM -0600, Andrew Halaney wrote:
> On Fri, Jan 19, 2024 at 02:07:15PM -0600, Andrew Halaney wrote:
> > On Fri, Jan 19, 2024 at 01:55:47PM -0500, Eric Chanudet wrote:
> > > On sa8775p-ride, probing the hba will go through the
> > > UFSHCD_QUIRK_REINIT_AFTER_MAX_GEAR_SWITCH path although the power info
> > > are same during the second init.
> > > 
> > > If the host is at least v4, ufs_qcom_get_hs_gear() picked the highest
> > > supported gear when setting the host_params. After the negotiation, if
> > > the host and device are on the same gear, it is the highest gear
> > > supported between the two. Skip the re-init to save some time.
> > > 
> > > Signed-off-by: Eric Chanudet <echanude@xxxxxxxxxx>
> > > ---
> > > 
> > > "trace_event=ufs:ufshcd_init" reports the time spent where the re-init
> > > quirk is performed. On sa8775p-ride:
> > > Baseline:
> > >   0.355879: ufshcd_init: 1d84000.ufs: took 103377 usecs, dev_state: UFS_ACTIVE_PWR_MODE, link_state: UIC_LINK_ACTIVE_STATE, err 0
> > > With this patch:
> > >   0.297676: ufshcd_init: 1d84000.ufs: took 43553 usecs, dev_state: UFS_ACTIVE_PWR_MODE, link_state: UIC_LINK_ACTIVE_STATE, err 0
> > > 
> > >  drivers/ufs/host/ufs-qcom.c | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/ufs/host/ufs-qcom.c b/drivers/ufs/host/ufs-qcom.c
> > > index 39eef470f8fa..f9f161340e78 100644
> > > --- a/drivers/ufs/host/ufs-qcom.c
> > > +++ b/drivers/ufs/host/ufs-qcom.c
> > > @@ -738,8 +738,12 @@ static int ufs_qcom_pwr_change_notify(struct ufs_hba *hba,
> > >  		 * the second init can program the optimal PHY settings. This allows one to start
> > >  		 * the first init with either the minimum or the maximum support gear.
> > >  		 */
> > > -		if (hba->ufshcd_state == UFSHCD_STATE_RESET)
> > > +		if (hba->ufshcd_state == UFSHCD_STATE_RESET) {
> > > +			if (host->hw_ver.major >= 0x4 &&
> > 
> > Is this check really necessary?

I *think* so.

For example, if hw_ver < 4, ufs_qcom_set_phy_gear() has a comment saying
"power up the PHY using minimum supported gear (UFS_HS_G2). Switching to
max gear will be performed during reinit if supported."

> > 
> > The initial phy_gear state is something like this (my phrasing of
> > ufs_qcom_set_phy_gear()):
> > 
> >     if hw_ver < 4:
> >         # Comments about powering up with minimum gear (with no
> >         # reasoning in the comment afaict), and mentions switching
> >         # to higher gear in reinit quirk. This is opposite of the later
> >         # versions which start at the max and scale down
> >         phy_gear = UFS_HS_G2

IIUC, the device would not be able to negotiate a gear higher than the
minimum set for the phy_gear on initialization.

ufshcd_init_host_params() and ufs_qcom_get_hs_gear() both set the
controller <v4 host_params to G3. So if the device is HS capable, the
re-init would set G3, instead of the G2 selected by
ufs_qcom_set_phy_gear().

Assuming I'm not loosing track somewhere, the sequence of calls would go
something like this:

ufshcd_init
 ufshcd_hba_init
  ufshcd_variant_hba_init
   ufshcd_vops_init
    ufs_qcom_init
     ufs_qcom_set_host_params /* if hw_ver < 4: phy_gear = G2 */
 ufshcd_hba_enable
  ufshcd_hba_execute_hce
   ufshcd_vops_hce_enable_notify(PRE_CHANGE)
    ufs_qcom_hce_enable_notify /* vops.hce_enable_notify */
     ufs_qcom_power_up_sequence
      phy_set_mode_ext(phy, mode, host->phy_gear);
 async_schedule(ufshcd_async_scan, hba)
 ...
 ufshcd_async_scan
  ufshcd_device_init
  ufshcd_probe_hba /* where the re-init quirk happens */
   ufshcd_device_init
    ufshcd_vops_pwr_change_notify(PRE_CHANGE)
  
So that would limit the device to G2? In this circumstances, the re-init
would instead re-initialize to G3.

> I guess what I'm saying is shouldn't the check be something like:
> 
>     if (dev_req_params->gear_tx == host->phy_gear) {
>         // Skip reinit since we started up in the agreed upon gear
>         hba->quirks &= ~UFSHCD_QUIRK_REINIT_AFTER_MAX_GEAR_SWITCH;
>     }

-- 
Eric Chanudet