Intel-XE Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: Jani Nikula <jani.nikula@linux.intel.com>
To: "Ville Syrjälä" <ville.syrjala@linux.intel.com>,
	intel-gfx@lists.freedesktop.org
Cc: intel-xe@lists.freedesktop.org
Subject: Re: [PATCH 4/8] drm/i915/de: Make mmio_debug lockless
Date: Wed, 18 Nov 2026 11:46:55 +0200	[thread overview]
Message-ID: <8090655c8cb0670c31293d12045072f57893e395@intel.com> (raw)
In-Reply-To: <avc7C3GbK39VmAzo@intel.com>

On Fri, 13 Nov 2026, Ville Syrjälä <ville.syrjala@linux.intel.com> wrote:
> On Fri, Nov 13, 2026 at 03:43:51PM +0200, Ville Syrjala wrote:
>> From: Ville Syrjälä <ville.syrjala@linux.intel.com>
>> 
>> Currently we protect the mmio_debug state using mmio.lock.
>> We can't just drop the locks or we risk the read-modify-writes
>> to the modparam to get out of whack. But if we get rid of the
>> read-modify-writes and just switch to simple booleans all
>> we need are loads and stores which are naturally atomic.
>> 
>> This is of course racy, but quite harmlessly so. Perhaps the
>> worst thing that can happen is that we mistake which register
>> caused the unclaimed access, but that can already happen since
>> we use the completely lockless intel_de_{read,write}_fw() in
>> many places. And a pair of threads getting a bit confused
>> about the state of the booleans should also be harmless. Perhaps
>> we end up doing a few extra unclaimed mmio debug checks while the
>> first thread is already about to disable the thing. No big deal.
>> 
>> And for good measure, let's annotate the booleans with
>> READ_ONCE()/WRITE_ONCE() to highlight the lockless nature,
>> though I don't think we need those for correctness.
>> 
>> With the booleans we do lose the capability to set mmio_debug
>> to a specific value, and then automagically disable it once
>> that many unclaimed register accesses have been detected.
>> With the boolean it's mmio_debug all the time or none of
>> the time. Not sure if anyone actually used that feature
>> anyway, so doesn't seem like a big loss to me.
>> 
>> I do want to preserve the oneshot detection thing automagically
>> armed by intel_de_arm_unclaimed_mmio_detection(). So we need a
>> second boolean for that.
>> 
>> And the third boolean will be mmio.unclaimed_mmio_check. That
>> was already effectively a boolean, but still being adjusted
>> with rmw.
>> 
>> The FPGA_DBG/CLAIM_ER accesses done from the read/write vfuncs
>> will still be conditionally protected by the mmio.lock, based
>> on mmio_flags. But the accesses from intel_de_unclaimed_mmio_detected()
>> and intel_de_arm_unclaimed_mmio_detection() need another
>> solution. For that I'm just adding another boolean into the
>> mmio funcs that tells us whether the lock should be taken
>> there or not. Calls to these functions are exceedingly rare
>> so no need to come up with any clever solution here.
>> 
>> Note that the modparam will now refuse any non-0/1 value, so
>> if anyone has left this enabled accidentally the module will
>> fail to load. But reminding people to turn this off in
>> production is probably a good thing :P
>
> Argh. Apparently igt uses the modparam for some reason. But it
> only sets it to -1, which is the default w/ the debug .config
> anyway. I guess it might make sense for people running the
> w/o all the debug knobs enabled in .config...

Maybe keep the modparam an int, but treat it as bool?

>
>> 
>> Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
>> ---
>>  drivers/gpu/drm/i915/Kconfig.debug            |  2 +-
>>  drivers/gpu/drm/i915/display/intel_de.c       | 58 +++++++++----------
>>  drivers/gpu/drm/i915/display/intel_de.h       |  1 +
>>  .../gpu/drm/i915/display/intel_display_core.h |  7 +--
>>  .../drm/i915/display/intel_display_params.c   |  4 +-
>>  .../drm/i915/display/intel_display_params.h   |  2 +-
>>  6 files changed, 34 insertions(+), 40 deletions(-)
>> 
>> diff --git a/drivers/gpu/drm/i915/Kconfig.debug b/drivers/gpu/drm/i915/Kconfig.debug
>> index dc43dcfbadb6..6a020adc43ee 100644
>> --- a/drivers/gpu/drm/i915/Kconfig.debug
>> +++ b/drivers/gpu/drm/i915/Kconfig.debug
>> @@ -76,7 +76,7 @@ config DRM_I915_DEBUG_MMIO
>>  	help
>>  	  By default, always enables the extra sanity checks (extra register
>>  	  reads) around every mmio (register) access that will slow the system
>> -	  down. This sets the default value of i915.mmio_debug to -1 and can
>> +	  down. This sets the default value of i915.mmio_debug to Y and can
>>  	  be overridden at module load.
>>  
>>  	  Recommended for driver developers only.
>> diff --git a/drivers/gpu/drm/i915/display/intel_de.c b/drivers/gpu/drm/i915/display/intel_de.c
>> index 06d91570bee2..5a74cf173cb8 100644
>> --- a/drivers/gpu/drm/i915/display/intel_de.c
>> +++ b/drivers/gpu/drm/i915/display/intel_de.c
>> @@ -242,12 +242,16 @@ static bool vlv_check_for_unclaimed_mmio(struct intel_display *display)
>>  	return true;
>>  }
>>  
>> +static bool mmio_debug_is_enabled(struct intel_display *display)
>> +{
>> +	return READ_ONCE(display->params.mmio_debug) ||
>> +		READ_ONCE(display->mmio.mmio_debug_oneshot);
>> +}
>> +
>>  static void check_unclaimed_mmio_pre(struct intel_display *display,
>>  				     intel_reg_t reg, bool write)
>>  {
>> -	lockdep_assert_held(&display->mmio.lock);
>> -
>> -	if (!display->params.mmio_debug)
>> +	if (!mmio_debug_is_enabled(display))
>>  		return;
>>  
>>  	if (!display->mmio.funcs->check_for_unclaimed_mmio(display))
>> @@ -262,9 +266,7 @@ static void check_unclaimed_mmio_pre(struct intel_display *display,
>>  static void check_unclaimed_mmio_post(struct intel_display *display,
>>  				      intel_reg_t reg, bool write)
>>  {
>> -	lockdep_assert_held(&display->mmio.lock);
>> -
>> -	if (!display->params.mmio_debug)
>> +	if (!mmio_debug_is_enabled(display))
>>  		return;
>>  
>>  	if (!display->mmio.funcs->check_for_unclaimed_mmio(display))
>> @@ -275,8 +277,7 @@ static void check_unclaimed_mmio_post(struct intel_display *display,
>>  		 write ? "write to" : "read from",
>>  		 i915_mmio_reg_offset(reg));
>>  
>> -	/* Only report the first N failures */
>> -	display->params.mmio_debug--;
>> +	WRITE_ONCE(display->mmio.mmio_debug_oneshot, false);
>>  }
>>  
>>  bool intel_de_unclaimed_mmio_detected(struct intel_display *display)
>> @@ -286,43 +287,36 @@ bool intel_de_unclaimed_mmio_detected(struct intel_display *display)
>>  	if (!display->mmio.funcs->check_for_unclaimed_mmio)
>>  		return false;
>>  
>> -	spin_lock_irq(&display->mmio.lock);
>> -	ret = display->mmio.funcs->check_for_unclaimed_mmio(display);
>> -	spin_unlock_irq(&display->mmio.lock);
>> +	if (display->mmio.funcs->check_for_unclaimed_mmio_needs_lock) {
>> +		spin_lock_irq(&display->mmio.lock);
>> +		ret = display->mmio.funcs->check_for_unclaimed_mmio(display);
>> +		spin_unlock_irq(&display->mmio.lock);
>> +	} else {
>> +		ret = display->mmio.funcs->check_for_unclaimed_mmio(display);
>> +	}
>>  
>>  	return ret;
>>  }
>>  
>>  bool intel_de_arm_unclaimed_mmio_detection(struct intel_display *display)
>>  {
>> -	bool ret = false;
>> -
>> -	if (!display->mmio.funcs->check_for_unclaimed_mmio)
>> +	if (!READ_ONCE(display->mmio.unclaimed_mmio_check))
>>  		return false;
>>  
>> -	spin_lock_irq(&display->mmio.lock);
>> -
>> -	if (display->mmio.unclaimed_mmio_check <= 0)
>> -		goto out;
>> -
>> -	if (!display->mmio.funcs->check_for_unclaimed_mmio(display))
>> -		goto out;
>> +	if (!intel_de_unclaimed_mmio_detected(display))
>> +		return false;
>>  
>> -	if (!display->params.mmio_debug) {
>> +	if (!READ_ONCE(display->params.mmio_debug)) {
>>  		drm_dbg_kms(display->drm,
>>  			    "Unclaimed register detected, "
>>  			    "enabling oneshot unclaimed register reporting. "
>> -			    "Please use i915.mmio_debug=N for more information.\n");
>> -		display->params.mmio_debug++;
>> +			    "Please use i915.mmio_debug=Y for more information.\n");
>> +		WRITE_ONCE(display->mmio.mmio_debug_oneshot, true);
>>  	}
>>  
>> -	display->mmio.unclaimed_mmio_check--;
>> -	ret = true;
>> +	WRITE_ONCE(display->mmio.unclaimed_mmio_check, false);
>>  
>> -out:
>> -	spin_unlock_irq(&display->mmio.lock);
>> -
>> -	return ret;
>> +	return true;
>>  }
>>  
>>  #define __MMIO_WAKELOCK		BIT(0)
>> @@ -474,12 +468,14 @@ static const struct intel_mmio_funcs hsw_mmio_funcs = {
>>  	.read = hsw_mmio_read,
>>  	.write = hsw_mmio_write,
>>  	.check_for_unclaimed_mmio = hsw_check_for_unclaimed_mmio,
>> +	.check_for_unclaimed_mmio_needs_lock = true,
>>  };
>>  
>>  static const struct intel_mmio_funcs vlv_mmio_funcs = {
>>  	.read = hsw_mmio_read,
>>  	.write = hsw_mmio_write,
>>  	.check_for_unclaimed_mmio = vlv_check_for_unclaimed_mmio,
>> +	.check_for_unclaimed_mmio_needs_lock = true,
>>  };
>>  
>>  #define LNL_MMIO_FLAGS __MMIO_WAKELOCK
>> @@ -539,7 +535,7 @@ void intel_de_init(struct intel_display *display)
>>  	}
>>  
>>  	if (display->mmio.funcs->check_for_unclaimed_mmio)
>> -		display->mmio.unclaimed_mmio_check = 1;
>> +		WRITE_ONCE(display->mmio.unclaimed_mmio_check, true);
>>  
>>  	/* clear out unclaimed reg detection bit someone else left behind */
>>  	if (intel_de_unclaimed_mmio_detected(display))
>> diff --git a/drivers/gpu/drm/i915/display/intel_de.h b/drivers/gpu/drm/i915/display/intel_de.h
>> index 2c469a2d6a20..cbf44975dced 100644
>> --- a/drivers/gpu/drm/i915/display/intel_de.h
>> +++ b/drivers/gpu/drm/i915/display/intel_de.h
>> @@ -13,6 +13,7 @@ struct intel_mmio_funcs {
>>  	u32 (*read)(struct intel_display *display, intel_reg_t reg);
>>  	void (*write)(struct intel_display *display, intel_reg_t reg, u32 val);
>>  	bool (*check_for_unclaimed_mmio)(struct intel_display *display);
>> +	bool check_for_unclaimed_mmio_needs_lock;
>>  };
>>  
>>  void intel_de_init(struct intel_display *display);
>> diff --git a/drivers/gpu/drm/i915/display/intel_display_core.h b/drivers/gpu/drm/i915/display/intel_display_core.h
>> index d68ecc19b8a8..92a3ce4247b1 100644
>> --- a/drivers/gpu/drm/i915/display/intel_display_core.h
>> +++ b/drivers/gpu/drm/i915/display/intel_display_core.h
>> @@ -539,17 +539,14 @@ struct intel_display {
>>  		 * serializes all display register accesses done
>>  		 * using intel_de_{read,write}(), but not those
>>  		 * done using intel_de_{read,write}_fw().
>> -		 *
>> -		 * The lock also protects mmio.unclaimed_mmio_check
>> -		 * and read-modify-write of display->params.mmio_debug.
>>  		 */
>>  		spinlock_t lock;
>>  
>>  		struct intel_uncore *uncore;
>>  		const struct intel_mmio_funcs *funcs;
>>  
>> -		/* protected by mmio.lock */
>> -		int unclaimed_mmio_check;
>> +		bool mmio_debug_oneshot;
>> +		bool unclaimed_mmio_check;
>>  	} mmio;
>>  
>>  	struct {
>> diff --git a/drivers/gpu/drm/i915/display/intel_display_params.c b/drivers/gpu/drm/i915/display/intel_display_params.c
>> index ace3081461de..7ef974861c8e 100644
>> --- a/drivers/gpu/drm/i915/display/intel_display_params.c
>> +++ b/drivers/gpu/drm/i915/display/intel_display_params.c
>> @@ -143,8 +143,8 @@ intel_display_param_named_unsafe(enable_dmc_wl, int, 0400,
>>  	"(-1=use per-chip default, 0=disabled, 1=enabled, 2=match any register, 3=always locked) "
>>  	"Default: -1");
>>  
>> -intel_display_param_named(mmio_debug, int, 0400,
>> -	"Enable the MMIO debug code for the first N failures (default: off). "
>> +intel_display_param_named(mmio_debug, bool, 0400,
>> +	"Enable MMIO debug code (default: false). "
>>  	"This may negatively affect performance.");
>>  
>>  __maybe_unused
>> diff --git a/drivers/gpu/drm/i915/display/intel_display_params.h b/drivers/gpu/drm/i915/display/intel_display_params.h
>> index bcbca95318d6..f6b9c294840c 100644
>> --- a/drivers/gpu/drm/i915/display/intel_display_params.h
>> +++ b/drivers/gpu/drm/i915/display/intel_display_params.h
>> @@ -51,7 +51,7 @@ struct drm_printer;
>>  	param(bool, psr_safest_params, false, 0400) \
>>  	param(bool, enable_psr2_sel_fetch, true, 0400) \
>>  	param(int, enable_dmc_wl, -1, 0400) \
>> -	param(int, mmio_debug, -IS_ENABLED(CONFIG_DRM_I915_DEBUG_MMIO), 0600) \
>> +	param(bool, mmio_debug, IS_ENABLED(CONFIG_DRM_I915_DEBUG_MMIO), 0600) \
>>  	/*
>>  	 * Debugfs-only knob (per-device): no matching module_param is registered
>>  	 * in intel_display_params.c on purpose. Runtime-toggle via debugfs.
>> -- 
>> 2.56.0

-- 
Jani Nikula, Intel

  reply	other threads:[~2026-11-18  9:47 UTC|newest]

Thread overview: 27+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-13 13:43 [PATCH 0/8] drm/{i915, xe}: Make unclaimed mmio detection lockless and enable it for xe Ville Syrjala
2026-11-13 13:43 ` [PATCH 1/8] drm/i915/de: Uninline intel_de_{read,write}_fw() Ville Syrjala
2026-11-13 14:11   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 2/8] drm/i915/de: Pimp the intel_de_{read,write}_fw() docs Ville Syrjala
2026-11-13 14:12   ` [PATCH 2/8] drm/i915/de: Pimp the intel_de_{read, write}_fw() docs Jani Nikula
2026-11-13 14:31     ` Ville Syrjälä
2026-11-18  9:45       ` Jani Nikula
2026-11-13 13:43 ` [PATCH 3/8] drm/i915/de: Flatten the unclaimed mmio code Ville Syrjala
2026-11-13 14:25   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 4/8] drm/i915/de: Make mmio_debug lockless Ville Syrjala
2026-11-13 16:30   ` Ville Syrjälä
2026-11-18  9:46     ` Jani Nikula [this message]
2026-11-18 10:04   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 5/8] drm/i915/de: Optimize the mmio_debug checks in intel_de_{read, write}() Ville Syrjala
2026-11-18 10:10   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 6/8] drm/i915/de: Make display registers unlocked for TGL+ on i915 Ville Syrjala
2026-11-18 10:11   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 7/8] drm/xe: Enable unclaimed register detection on xe Ville Syrjala
2026-11-18 10:14   ` Jani Nikula
2026-11-13 13:43 ` [PATCH 8/8] drm/i915/de: Nuke driver_is_i915() Ville Syrjala
2026-11-18 10:15   ` Jani Nikula
2026-11-13 14:33 ` ✗ CI.checkpatch: warning for drm/{i915, xe}: Make unclaimed mmio detection lockless and enable it for xe Patchwork
2026-11-13 14:35 ` ✓ CI.KUnit: success " Patchwork
2026-11-13 14:52 ` ✗ CI.checksparse: warning " Patchwork
2026-11-13 15:07 ` ✗ Xe.CI.BAT: failure " Patchwork
2026-11-13 16:41   ` Ville Syrjälä
2026-11-14  0:50 ` ✗ Xe.CI.FULL: " Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=8090655c8cb0670c31293d12045072f57893e395@intel.com \
    --to=jani.nikula@linux.intel.com \
    --cc=intel-gfx@lists.freedesktop.org \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=ville.syrjala@linux.intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox