| Message ID | 1510827497-25188-3-git-send-email-souvik.k.chakravarty@intel.com (mailing list archive) |
|---|---|
| State | Changes Requested, archived |
| Headers | show |
On Thu, Nov 16, 2017 at 12:18 PM, Souvik Kumar Chakravarty <souvik.k.chakravarty@intel.com> wrote: > Suspend stats are not reported consistently due to a limitation in the PMC > firmware. This limitation causes a delay in updating the s0ix counters and > residencies in the telemetry log upon s0ix exit. As a consequence, reading > these counters from the suspend-exit notifier may result in zero read. > > This patch fixes this issue by cross-verifying the s0ix residencies from > the GCR TELEM registers in case the counters are not incremented in the > telemetry log after suspend. > > This fixes https://bugzilla.kernel.org/show_bug.cgi?id=197833 > > Reported-and-tested-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@intel.com> > Signed-off-by: Souvik Kumar Chakravarty <souvik.k.chakravarty@intel.com> > +#define TELEM_LO_HI_TO64(lo, hi) ((u64)(lo) + ((u64)(hi)<<32)) > + > + /* Due to some design limitations in the firmware, sometimes the > + * counters do not get updated by the time we reach here. As a > + * workaround, we try to see if this was a genuine case of sleep > + * failure or not by cross-checking from PMC GCR registers directly. > + */ Comments style! > + if ((suspend_shlw_ctr_exit == suspend_shlw_ctr_temp) && > + (suspend_deep_ctr_exit == suspend_deep_ctr_temp)) { Redundant parens. > + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_SHLW_S0IX_LO_REG, > + &shlw_lo); Okay, from here is now obvious what you did the first patch. I think you need to fold the changes here. > + if (ret < 0) > + goto out; > + > + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_SHLW_S0IX_HI_REG, > + &shlw_hi); > + if (ret < 0) > + goto out; ...and now we have a problem. Each of your call takes a spin lock. What happened in between? If I understand this correctly you need to introduce intel_pmc_gcr_readl() and intel_pmc_gcr_readq(). > + > + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_DEEP_S0IX_LO_REG, > + &deep_lo); > + if (ret < 0) > + goto out; > + > + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_DEEP_S0IX_HI_REG, > + &deep_hi); > + if (ret < 0) > + goto out; Same story.
diff --git a/drivers/platform/x86/intel_telemetry_debugfs.c b/drivers/platform/x86/intel_telemetry_debugfs.c index 4249e826..d0fce8c 100644 --- a/drivers/platform/x86/intel_telemetry_debugfs.c +++ b/drivers/platform/x86/intel_telemetry_debugfs.c @@ -74,6 +74,8 @@ #define TELEM_IOSS_DX_D0IX_EVTS 25 #define TELEM_IOSS_PG_EVTS 30 +#define TELEM_LO_HI_TO64(lo, hi) ((u64)(lo) + ((u64)(hi)<<32)) + #define TELEM_DEBUGFS_CPU(model, data) \ { X86_VENDOR_INTEL, 6, model, X86_FEATURE_ANY, (unsigned long)&data} @@ -858,6 +860,7 @@ static int pm_suspend_exit_cb(void) static u32 suspend_shlw_ctr_exit, suspend_deep_ctr_exit; static u64 suspend_shlw_res_exit, suspend_deep_res_exit; struct telemetry_debugfs_conf *conf = debugfs_conf; + u32 shlw_lo, shlw_hi, deep_lo, deep_hi; int ret, index; if (!suspend_prep_ok) @@ -890,6 +893,42 @@ static int pm_suspend_exit_cb(void) goto out; } + /* Due to some design limitations in the firmware, sometimes the + * counters do not get updated by the time we reach here. As a + * workaround, we try to see if this was a genuine case of sleep + * failure or not by cross-checking from PMC GCR registers directly. + */ + if ((suspend_shlw_ctr_exit == suspend_shlw_ctr_temp) && + (suspend_deep_ctr_exit == suspend_deep_ctr_temp)) { + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_SHLW_S0IX_LO_REG, + &shlw_lo); + if (ret < 0) + goto out; + + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_SHLW_S0IX_HI_REG, + &shlw_hi); + if (ret < 0) + goto out; + + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_DEEP_S0IX_LO_REG, + &deep_lo); + if (ret < 0) + goto out; + + ret = intel_pmc_gcr_read(PMC_GCR_TELEM_DEEP_S0IX_HI_REG, + &deep_hi); + if (ret < 0) + goto out; + + suspend_shlw_res_exit = TELEM_LO_HI_TO64(shlw_lo, shlw_hi); + if (suspend_shlw_res_exit > suspend_shlw_res_temp) + suspend_shlw_ctr_exit++; + + suspend_deep_res_exit = TELEM_LO_HI_TO64(deep_lo, deep_hi); + if (suspend_deep_res_exit > suspend_deep_res_temp) + suspend_deep_ctr_exit++; + } + suspend_shlw_ctr_exit -= suspend_shlw_ctr_temp; suspend_deep_ctr_exit -= suspend_deep_ctr_temp; suspend_shlw_res_exit -= suspend_shlw_res_temp;
Suspend stats are not reported consistently due to a limitation in the PMC firmware. This limitation causes a delay in updating the s0ix counters and residencies in the telemetry log upon s0ix exit. As a consequence, reading these counters from the suspend-exit notifier may result in zero read. This patch fixes this issue by cross-verifying the s0ix residencies from the GCR TELEM registers in case the counters are not incremented in the telemetry log after suspend. This fixes https://bugzilla.kernel.org/show_bug.cgi?id=197833 Reported-and-tested-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@intel.com> Signed-off-by: Souvik Kumar Chakravarty <souvik.k.chakravarty@intel.com> --- drivers/platform/x86/intel_telemetry_debugfs.c | 39 ++++++++++++++++++++++++++ 1 file changed, 39 insertions(+)