On 2017-11-21 20:45, Subhash Jadavani wrote:
> On 2017-11-13 01:14, Michal Potomski wrote:
> > From: Michał Potomski <michalx.potomski@xxxxxxxxx>
> >
> > Recent testing of Runtime PM for UFS has shown it's not working as
> > intended. To acheive fully working Runtime PM, first we have to put
> > back scsi_device autopm reference counter.
> >
> > Existing implementation was prone to races and not working for
> > tranfsers to sleeping devices. This commit aims to fix this to acheive
> > fully working environment. Due to runtime PM being previously disabled
> > by not putting scsi_device autopm counter, the Runtime PM is set to be
> > forbidden as default, to not cause problems with hosts and devices,
> > which do not fully support different Device and Link states.
> >
> > It can be still enabled through sysFS power/control attribute.
> >
> > Changes since v1:
> > - Fix compilation error for certain kernel configs,
> > - Move pm_mark_last_busy(), only to relevant context.
> >
> > Signed-off-by: Michał Potomski <michalx.potomski@xxxxxxxxx>
> > ---
> > drivers/scsi/ufs/ufshcd-pci.c | 4 ++-
> > drivers/scsi/ufs/ufshcd.c | 64
> > +++++++++++++++++++++++++++++++++++--------
> > 2 files changed, 56 insertions(+), 12 deletions(-)
> >
> > diff --git a/drivers/scsi/ufs/ufshcd-pci.c
> > b/drivers/scsi/ufs/ufshcd-pci.c index 925b0ec7ec54..4fffb1123876
> > 100644
> > --- a/drivers/scsi/ufs/ufshcd-pci.c
> > +++ b/drivers/scsi/ufs/ufshcd-pci.c
> > @@ -184,8 +184,10 @@ ufshcd_pci_probe(struct pci_dev *pdev, const
> > struct pci_device_id *id)
> > }
> >
> > pci_set_drvdata(pdev, hba);
> > + pm_runtime_set_autosuspend_delay(&pdev->dev, 3000);
> > + pm_runtime_use_autosuspend(&pdev->dev);
> > pm_runtime_put_noidle(&pdev->dev);
> > - pm_runtime_allow(&pdev->dev);
> > + pm_runtime_forbid(&pdev->dev);
> >
> > return 0;
> > }
> > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> > index 011c3369082c..7ff7f51afe26 100644
> > --- a/drivers/scsi/ufs/ufshcd.c
> > +++ b/drivers/scsi/ufs/ufshcd.c
> > @@ -240,6 +240,37 @@ static inline bool ufshcd_valid_tag(struct
> > ufs_hba *hba, int tag)
> > return tag >= 0 && tag < hba->nutrs; }
> >
> > +#ifdef CONFIG_PM
> > +static void ufshcd_pm_runtime_get(struct device *dev) {
> > + // Don't trigger PM events while in transition states
> > + if (dev->power.runtime_status == RPM_RESUMING ||
> > + dev->power.runtime_status == RPM_SUSPENDING)
>
> Why do we need to check this? Isn't this check race condition prone?
We need to check this, because we might end up in triggering resume, while
suspending or resuming, which will cause deadlock. This is especially important
in ufshcd_queuecommand directive, through which we pass e.g. START/STOP Unit
SCSI command, when resuming or suspending.
> > + pm_runtime_get_noresume(dev);
> > + else
> > + pm_runtime_get_sync(dev);
> > +}
> > +
> > +static void ufshcd_pm_runtime_put(struct device *dev) {
> > + // Don't trigger PM events while in transition states
> > + if (dev->power.runtime_status == RPM_RESUMING ||
> > + dev->power.runtime_status == RPM_SUSPENDING)
> > + pm_runtime_put_noidle(dev);
> > + else {
> > + pm_runtime_mark_last_busy(dev);
> > + pm_runtime_put_autosuspend(dev);
> > + }
> > +}
> > +#else
> > +static void ufshcd_pm_runtime_get(struct device *dev) { } static void
> > +ufshcd_pm_runtime_put(struct device *dev) { } #endif
> > +
> > static inline int ufshcd_enable_irq(struct ufs_hba *hba) {
> > int ret = 0;
> > @@ -1345,7 +1376,7 @@ static ssize_t
> > ufshcd_clkscale_enable_store(struct device *dev,
> > if (value == hba->clk_scaling.is_allowed)
> > goto out;
> >
> > - pm_runtime_get_sync(hba->dev);
> > + ufshcd_pm_runtime_get(hba->dev);
> > ufshcd_hold(hba, false);
> >
> > cancel_work_sync(&hba->clk_scaling.suspend_work);
> > @@ -1364,7 +1395,7 @@ static ssize_t
> > ufshcd_clkscale_enable_store(struct device *dev,
> > }
> >
> > ufshcd_release(hba);
> > - pm_runtime_put_sync(hba->dev);
> > + ufshcd_pm_runtime_put(hba->dev);
> > out:
> > return count;
> > }
> > @@ -1948,6 +1979,7 @@ ufshcd_send_uic_cmd(struct ufs_hba *hba, struct
> > uic_command *uic_cmd)
> > unsigned long flags;
> >
> > ufshcd_hold(hba, false);
> > + ufshcd_pm_runtime_get(hba->dev);
> > mutex_lock(&hba->uic_cmd_mutex);
> > ufshcd_add_delay_before_dme_cmd(hba);
> >
> > @@ -1959,6 +1991,7 @@ ufshcd_send_uic_cmd(struct ufs_hba *hba, struct
> > uic_command *uic_cmd)
> >
> > mutex_unlock(&hba->uic_cmd_mutex);
> >
> > + ufshcd_pm_runtime_put(hba->dev);
> > ufshcd_release(hba);
> > return ret;
> > }
> > @@ -2345,6 +2378,7 @@ static int ufshcd_queuecommand(struct Scsi_Host
> > *host, struct scsi_cmnd *cmd)
> > clear_bit_unlock(tag, &hba->lrb_in_use);
> > goto out;
> > }
> > + ufshcd_pm_runtime_get(hba->dev);
>
>
> why do we need to hold PM reference count here? I thought as ufs host is
> parent of scsi_host/scsi_device, scsi would make sure that UFS host is runtime
> resumed before any access to it.
Theoretically yes, but we end up suspended, before getting completion on the
command, which causes abort. Thanks to this reference count we are sure, that
actual device will be suspended after command is completed.
>
> > WARN_ON(hba->clk_gating.state != CLKS_ON);
> >
> > lrbp = &hba->lrb[tag];
> > @@ -3555,6 +3589,7 @@ static int ufshcd_uic_pwr_ctrl(struct ufs_hba
> > *hba, struct uic_command *cmd)
> > int ret;
> > bool reenable_intr = false;
> >
> > + ufshcd_pm_runtime_get(hba->dev);
> > mutex_lock(&hba->uic_cmd_mutex);
> > init_completion(&uic_async_done);
> > ufshcd_add_delay_before_dme_cmd(hba);
> > @@ -3609,6 +3644,7 @@ static int ufshcd_uic_pwr_ctrl(struct ufs_hba
> > *hba, struct uic_command *cmd)
> > ufshcd_enable_intr(hba, UIC_COMMAND_COMPL);
> > spin_unlock_irqrestore(hba->host->host_lock, flags);
> > mutex_unlock(&hba->uic_cmd_mutex);
> > + ufshcd_pm_runtime_put(hba->dev);
> >
> > return ret;
> > }
> > @@ -4386,6 +4422,7 @@ static int ufshcd_slave_configure(struct
> > scsi_device *sdev)
> >
> > blk_queue_update_dma_pad(q, PRDT_DATA_BYTE_COUNT_PAD - 1);
> > blk_queue_max_segment_size(q, PRDT_DATA_BYTE_COUNT_MAX);
> > + scsi_autopm_put_device(sdev);
>
>
> scsi_sysfs_add_sdev() calls scsi_autopm_put_device() so not sure why this is
> required here?
While I was testing runtime PM, I stumbled upon problem, that device was not
Going to suspend due to active kids. After some research I found a commit
bc4f2401 stating:
"If a high-level driver does
nothing then the device will not be eligible for runtime-suspend
because of the elevated usage count. If a high-level driver wants to
use runtime PM then it can call scsi_autopm_put_device()"
After using this routine only I observed, that runtime PM states has
begun to be triggered, which was not the case earlier (although still
not working properly without the other changes).
>
> >
> > return 0;
> > }
> > @@ -4622,6 +4659,7 @@ static void __ufshcd_transfer_req_compl(struct
> > ufs_hba *hba,
> > /* Do not touch lrbp after scsi done */
> > cmd->scsi_done(cmd);
> > __ufshcd_release(hba);
> > + ufshcd_pm_runtime_put(hba->dev);
> > } else if (lrbp->command_type ==
> UTP_CMD_TYPE_DEV_MANAGE ||
> > lrbp->command_type ==
> UTP_CMD_TYPE_UFS_STORAGE) {
> > if (hba->dev_cmd.complete) {
> > @@ -4951,7 +4989,7 @@ static void
> > ufshcd_exception_event_handler(struct work_struct *work)
> > u32 status = 0;
> > hba = container_of(work, struct ufs_hba, eeh_work);
> >
> > - pm_runtime_get_sync(hba->dev);
> > + ufshcd_pm_runtime_get(hba->dev);
> > err = ufshcd_get_ee_status(hba, &status);
> > if (err) {
> > dev_err(hba->dev, "%s: failed to get exception status %d\n",
> @@
> > -4965,7 +5003,7 @@ static void ufshcd_exception_event_handler(struct
> > work_struct *work)
> > ufshcd_bkops_exception_event_handler(hba);
> >
> > out:
> > - pm_runtime_put_sync(hba->dev);
> > + ufshcd_pm_runtime_put(hba->dev);
> > return;
> > }
> >
> > @@ -5065,7 +5103,7 @@ static void ufshcd_err_handler(struct
> > work_struct
> > *work)
> >
> > hba = container_of(work, struct ufs_hba, eh_work);
> >
> > - pm_runtime_get_sync(hba->dev);
> > + ufshcd_pm_runtime_get(hba->dev);
> > ufshcd_hold(hba, false);
> >
> > spin_lock_irqsave(hba->host->host_lock, flags); @@ -5177,7 +5215,7
> > @@ static void ufshcd_err_handler(struct work_struct
> > *work)
> > spin_unlock_irqrestore(hba->host->host_lock, flags);
> > scsi_unblock_requests(hba->host);
> > ufshcd_release(hba);
> > - pm_runtime_put_sync(hba->dev);
> > + ufshcd_pm_runtime_put(hba->dev);
> > }
> >
> > static void ufshcd_update_uic_reg_hist(struct ufs_uic_err_reg_hist
> > *reg_hist, @@ -6425,7 +6463,7 @@ static int ufshcd_probe_hba(struct
> > ufs_hba *hba)
> > }
> >
> > scsi_scan_host(hba->host);
> > - pm_runtime_put_sync(hba->dev);
> > + ufshcd_pm_runtime_put(hba->dev);
> > }
> >
> > if (!hba->is_init_prefetch)
> > @@ -6437,7 +6475,7 @@ static int ufshcd_probe_hba(struct ufs_hba *hba)
> > * present, turn off the power/clocks etc.
> > */
> > if (ret && !ufshcd_eh_in_progress(hba) && !hba->pm_op_in_progress) {
> > - pm_runtime_put_sync(hba->dev);
> > + ufshcd_pm_runtime_put(hba->dev);
> > ufshcd_hba_exit(hba);
> > }
> >
> > @@ -7747,6 +7785,8 @@ EXPORT_SYMBOL(ufshcd_shutdown); void
> > ufshcd_remove(struct ufs_hba *hba) {
> > ufshcd_remove_sysfs_nodes(hba);
> > + // Resume if suspended for sync
> > + ufshcd_pm_runtime_get(hba->dev);
> > scsi_remove_host(hba->host);
> > /* disable interrupts */
> > ufshcd_disable_intr(hba, hba->intr_mask); @@ -7756,6 +7796,8 @@
> void
> > ufshcd_remove(struct ufs_hba *hba)
> > if (ufshcd_is_clkscaling_supported(hba))
> > device_remove_file(hba->dev, &hba->clk_scaling.enable_attr);
> > ufshcd_hba_exit(hba);
> > + // Final sync finished - put it back
> > + ufshcd_pm_runtime_put(hba->dev);
> > }
> > EXPORT_SYMBOL_GPL(ufshcd_remove);
> >
> > @@ -7978,11 +8020,11 @@ int ufshcd_init(struct ufs_hba *hba, void
> > __iomem *mmio_base, unsigned int irq)
> > UFS_SLEEP_PWR_MODE,
> > UIC_LINK_HIBERN8_STATE);
> > hba->spm_lvl = ufs_get_desired_pm_lvl_for_dev_link_state(
> > - UFS_SLEEP_PWR_MODE,
> > - UIC_LINK_HIBERN8_STATE);
> > +
> UFS_POWERDOWN_PWR_MODE,
> > + UIC_LINK_OFF_STATE);
>
> This may add lot of latencies on resume so not sure we want to change the
> default.
It will possibly add latencies, but since we are going to system suspend, we'd
most likely to save power at system suspend at the highest possible rate, while
we are more interested in moderate latency increase vs. lesser power save in
runtime suspend.
>
> >
> > /* Hold auto suspend until async scan completes */
> > - pm_runtime_get_sync(dev);
> > + ufshcd_pm_runtime_get(dev);
> >
> > /*
> > * We are assuming that device wasn't put in sleep/power-down
>
> --
> The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a
> Linux Foundation Collaborative Project
--------------------------------------------------------------------
Intel Technology Poland sp. z o.o.
ul. Slowackiego 173 | 80-298 Gdansk | Sad Rejonowy Gdansk Polnoc | VII Wydzial Gospodarczy Krajowego Rejestru Sadowego - KRS 101882 | NIP 957-07-52-316 | Kapital zakladowy 200.000 PLN.
Ta wiadomosc wraz z zalacznikami jest przeznaczona dla okreslonego adresata i moze zawierac informacje poufne. W razie przypadkowego otrzymania tej wiadomosci, prosimy o powiadomienie nadawcy oraz trwale jej usuniecie; jakiekolwiek
przegladanie lub rozpowszechnianie jest zabronione.
This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). If you are not the intended recipient, please contact the sender and delete all copies; any review or distribution by
others is strictly prohibited.
