Merge tag 'drm-intel-fixes-2024-03-28' of...

Merge tag 'drm-intel-fixes-2024-03-28' of https://anongit.freedesktop.org/git/drm/drm-intel into drm-fixes

Core/GT Fixes:
- Fix for BUG_ON/BUILD_BUG_ON IN I915_memcpy.c (Joonas)
- Update a MTL workaround (Tejas)
- Fix locking inversion in hwmon's sysfs (Janusz)
- Remove a bogus error message around PXP (Jose)
- Fix UAF on VMA (Janusz)
- Reset queue_priority_hint on parking (Chris)

Display Fixes:
- Remove duplicated audio enable/disable on SDVO and DP (Ville)
- Disable AuxCCS for Xe driver (Juha-Pekka)
- Revert init order of MIPI DSI (Ville)
- DRRS debugfs fix with an extra refactor patch (Bhanuprakash)
- VRR related fixes (Ville)
- Fix a JSL eDP corruption (Jonathon)
- Fix the cursor physical dma address (Ville)
- BIOS VBT related fix (Ville)
Signed-off-by: Dave Airlie <airlied@redhat.com>

From: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/ZgYaIVgjIs30mIvS@intel.com

drivers/gpu/drm/i915/display/g4x_dp.c

@@ -717,7 +717,6 @@ static void g4x_enable_dp(struct intel_atomic_state *state,
 {
 	intel_enable_dp(state, encoder, pipe_config, conn_state);
 	intel_edp_backlight_on(pipe_config, conn_state);
-	encoder->audio_enable(encoder, pipe_config, conn_state);
 }
 
 static void vlv_enable_dp(struct intel_atomic_state *state,
@@ -726,7 +725,6 @@ static void vlv_enable_dp(struct intel_atomic_state *state,
 			  const struct drm_connector_state *conn_state)
 {
 	intel_edp_backlight_on(pipe_config, conn_state);
-	encoder->audio_enable(encoder, pipe_config, conn_state);
 }
 
 static void g4x_pre_enable_dp(struct intel_atomic_state *state,

drivers/gpu/drm/i915/display/icl_dsi.c

@@ -1155,7 +1155,6 @@ static void gen11_dsi_powerup_panel(struct intel_encoder *encoder)
 	}
 
 	intel_dsi_vbt_exec_sequence(intel_dsi, MIPI_SEQ_INIT_OTP);
-	intel_dsi_vbt_exec_sequence(intel_dsi, MIPI_SEQ_DISPLAY_ON);
 
 	/* ensure all panel commands dispatched before enabling transcoder */
 	wait_for_cmds_dispatched_to_panel(encoder);
@@ -1256,6 +1255,8 @@ static void gen11_dsi_enable(struct intel_atomic_state *state,
 	/* step6d: enable dsi transcoder */
 	gen11_dsi_enable_transcoder(encoder);
 
+	intel_dsi_vbt_exec_sequence(intel_dsi, MIPI_SEQ_DISPLAY_ON);
+
 	/* step7: enable backlight */
 	intel_backlight_enable(crtc_state, conn_state);
 	intel_dsi_vbt_exec_sequence(intel_dsi, MIPI_SEQ_BACKLIGHT_ON);

drivers/gpu/drm/i915/display/intel_bios.c

@@ -1955,16 +1955,12 @@ static int get_init_otp_deassert_fragment_len(struct drm_i915_private *i915,
  * these devices we split the init OTP sequence into a deassert sequence and
  * the actual init OTP part.
  */
-static void fixup_mipi_sequences(struct drm_i915_private *i915,
-				 struct intel_panel *panel)
+static void vlv_fixup_mipi_sequences(struct drm_i915_private *i915,
+				     struct intel_panel *panel)
 {
 	u8 *init_otp;
 	int len;
 
-	/* Limit this to VLV for now. */
-	if (!IS_VALLEYVIEW(i915))
-		return;
-
 	/* Limit this to v1 vid-mode sequences */
 	if (panel->vbt.dsi.config->is_cmd_mode ||
 	    panel->vbt.dsi.seq_version != 1)
@@ -2000,6 +1996,41 @@ static void fixup_mipi_sequences(struct drm_i915_private *i915,
 	panel->vbt.dsi.sequence[MIPI_SEQ_INIT_OTP] = init_otp + len - 1;
 }
 
+/*
+ * Some machines (eg. Lenovo 82TQ) appear to have broken
+ * VBT sequences:
+ * - INIT_OTP is not present at all
+ * - what should be in INIT_OTP is in DISPLAY_ON
+ * - what should be in DISPLAY_ON is in BACKLIGHT_ON
+ *   (along with the actual backlight stuff)
+ *
+ * To make those work we simply swap DISPLAY_ON and INIT_OTP.
+ *
+ * TODO: Do we need to limit this to specific machines,
+ *       or examine the contents of the sequences to
+ *       avoid false positives?
+ */
+static void icl_fixup_mipi_sequences(struct drm_i915_private *i915,
+				     struct intel_panel *panel)
+{
+	if (!panel->vbt.dsi.sequence[MIPI_SEQ_INIT_OTP] &&
+	    panel->vbt.dsi.sequence[MIPI_SEQ_DISPLAY_ON]) {
+		drm_dbg_kms(&i915->drm, "Broken VBT: Swapping INIT_OTP and DISPLAY_ON sequences\n");
+
+		swap(panel->vbt.dsi.sequence[MIPI_SEQ_INIT_OTP],
+		     panel->vbt.dsi.sequence[MIPI_SEQ_DISPLAY_ON]);
+	}
+}
+
+static void fixup_mipi_sequences(struct drm_i915_private *i915,
+				 struct intel_panel *panel)
+{
+	if (DISPLAY_VER(i915) >= 11)
+		icl_fixup_mipi_sequences(i915, panel);
+	else if (IS_VALLEYVIEW(i915))
+		vlv_fixup_mipi_sequences(i915, panel);
+}
+
 static void
 parse_mipi_sequence(struct drm_i915_private *i915,
 		    struct intel_panel *panel)
@@ -3351,6 +3382,9 @@ bool intel_bios_encoder_supports_dp_dual_mode(const struct intel_bios_encoder_da
 {
 	const struct child_device_config *child = &devdata->child;
 
+	if (!devdata)
+		return false;
+
 	if (!intel_bios_encoder_supports_dp(devdata) ||
 	    !intel_bios_encoder_supports_hdmi(devdata))
 		return false;

drivers/gpu/drm/i915/display/intel_cursor.c

@@ -36,12 +36,10 @@ static u32 intel_cursor_base(const struct intel_plane_state *plane_state)
 {
 	struct drm_i915_private *dev_priv =
 		to_i915(plane_state->uapi.plane->dev);
-	const struct drm_framebuffer *fb = plane_state->hw.fb;
-	struct drm_i915_gem_object *obj = intel_fb_obj(fb);
 	u32 base;
 
 	if (DISPLAY_INFO(dev_priv)->cursor_needs_physical)
-		base = i915_gem_object_get_dma_address(obj, 0);
+		base = plane_state->phys_dma_addr;
 	else
 		base = intel_plane_ggtt_offset(plane_state);
 

drivers/gpu/drm/i915/display/intel_display_types.h

@@ -727,6 +727,7 @@ struct intel_plane_state {
 #define PLANE_HAS_FENCE BIT(0)
 
 	struct intel_fb_view view;
+	u32 phys_dma_addr; /* for cursor_needs_physical */
 
 	/* Plane pxp decryption state */
 	bool decrypt;

drivers/gpu/drm/i915/display/intel_dp.c

@@ -67,6 +67,7 @@
 #include "intel_dp_tunnel.h"
 #include "intel_dpio_phy.h"
 #include "intel_dpll.h"
+#include "intel_drrs.h"
 #include "intel_fifo_underrun.h"
 #include "intel_hdcp.h"
 #include "intel_hdmi.h"
@@ -2683,15 +2684,6 @@ intel_dp_compute_hdr_metadata_infoframe_sdp(struct intel_dp *intel_dp,
 		intel_hdmi_infoframe_enable(HDMI_PACKET_TYPE_GAMUT_METADATA);
 }
 
-static bool cpu_transcoder_has_drrs(struct drm_i915_private *i915,
-				    enum transcoder cpu_transcoder)
-{
-	if (HAS_DOUBLE_BUFFERED_M_N(i915))
-		return true;
-
-	return intel_cpu_transcoder_has_m2_n2(i915, cpu_transcoder);
-}
-
 static bool can_enable_drrs(struct intel_connector *connector,
 			    const struct intel_crtc_state *pipe_config,
 			    const struct drm_display_mode *downclock_mode)
@@ -2714,7 +2706,7 @@ static bool can_enable_drrs(struct intel_connector *connector,
 	if (pipe_config->has_pch_encoder)
 		return false;
 
-	if (!cpu_transcoder_has_drrs(i915, pipe_config->cpu_transcoder))
+	if (!intel_cpu_transcoder_has_drrs(i915, pipe_config->cpu_transcoder))
 		return false;
 
 	return downclock_mode &&

drivers/gpu/drm/i915/display/intel_dpll_mgr.c

@@ -2554,7 +2554,7 @@ static void icl_wrpll_params_populate(struct skl_wrpll_params *params,
 static bool
 ehl_combo_pll_div_frac_wa_needed(struct drm_i915_private *i915)
 {
-	return (((IS_ELKHARTLAKE(i915) || IS_JASPERLAKE(i915)) &&
+	return ((IS_ELKHARTLAKE(i915) &&
 		 IS_DISPLAY_STEP(i915, STEP_B0, STEP_FOREVER)) ||
 		 IS_TIGERLAKE(i915) || IS_ALDERLAKE_S(i915) || IS_ALDERLAKE_P(i915)) &&
 		 i915->display.dpll.ref_clks.nssc == 38400;

drivers/gpu/drm/i915/display/intel_drrs.c

@@ -63,6 +63,15 @@ const char *intel_drrs_type_str(enum drrs_type drrs_type)
 	return str[drrs_type];
 }
 
+bool intel_cpu_transcoder_has_drrs(struct drm_i915_private *i915,
+				   enum transcoder cpu_transcoder)
+{
+	if (HAS_DOUBLE_BUFFERED_M_N(i915))
+		return true;
+
+	return intel_cpu_transcoder_has_m2_n2(i915, cpu_transcoder);
+}
+
 static void
 intel_drrs_set_refresh_rate_pipeconf(struct intel_crtc *crtc,
 				     enum drrs_refresh_rate refresh_rate)
@@ -312,9 +321,8 @@ static int intel_drrs_debugfs_status_show(struct seq_file *m, void *unused)
 	mutex_lock(&crtc->drrs.mutex);
 
 	seq_printf(m, "DRRS capable: %s\n",
-		   str_yes_no(crtc_state->has_drrs ||
-			      HAS_DOUBLE_BUFFERED_M_N(i915) ||
-			      intel_cpu_transcoder_has_m2_n2(i915, crtc_state->cpu_transcoder)));
+		   str_yes_no(intel_cpu_transcoder_has_drrs(i915,
+							    crtc_state->cpu_transcoder)));
 
 	seq_printf(m, "DRRS enabled: %s\n",
 		   str_yes_no(crtc_state->has_drrs));

drivers/gpu/drm/i915/display/intel_drrs.h

@@ -9,12 +9,15 @@
 #include <linux/types.h>
 
 enum drrs_type;
+enum transcoder;
 struct drm_i915_private;
 struct intel_atomic_state;
 struct intel_crtc;
 struct intel_crtc_state;
 struct intel_connector;
 
+bool intel_cpu_transcoder_has_drrs(struct drm_i915_private *i915,
+				   enum transcoder cpu_transcoder);
 const char *intel_drrs_type_str(enum drrs_type drrs_type);
 bool intel_drrs_is_active(struct intel_crtc *crtc);
 void intel_drrs_activate(const struct intel_crtc_state *crtc_state);

drivers/gpu/drm/i915/display/intel_dsb.c

@@ -340,6 +340,17 @@ static int intel_dsb_dewake_scanline(const struct intel_crtc_state *crtc_state)
 	return max(0, vblank_start - intel_usecs_to_scanlines(adjusted_mode, latency));
 }
 
+static u32 dsb_chicken(struct intel_crtc *crtc)
+{
+	if (crtc->mode_flags & I915_MODE_FLAG_VRR)
+		return DSB_CTRL_WAIT_SAFE_WINDOW |
+			DSB_CTRL_NO_WAIT_VBLANK |
+			DSB_INST_WAIT_SAFE_WINDOW |
+			DSB_INST_NO_WAIT_VBLANK;
+	else
+		return 0;
+}
+
 static void _intel_dsb_commit(struct intel_dsb *dsb, u32 ctrl,
 			      int dewake_scanline)
 {
@@ -361,6 +372,9 @@ static void _intel_dsb_commit(struct intel_dsb *dsb, u32 ctrl,
 	intel_de_write_fw(dev_priv, DSB_CTRL(pipe, dsb->id),
 			  ctrl | DSB_ENABLE);
 
+	intel_de_write_fw(dev_priv, DSB_CHICKEN(pipe, dsb->id),
+			  dsb_chicken(crtc));
+
 	intel_de_write_fw(dev_priv, DSB_HEAD(pipe, dsb->id),
 			  intel_dsb_buffer_ggtt_offset(&dsb->dsb_buf));
 

drivers/gpu/drm/i915/display/intel_fb_pin.c

@@ -255,6 +255,16 @@ int intel_plane_pin_fb(struct intel_plane_state *plane_state)
 			return PTR_ERR(vma);
 
 		plane_state->ggtt_vma = vma;
+
+		/*
+		 * Pre-populate the dma address before we enter the vblank
+		 * evade critical section as i915_gem_object_get_dma_address()
+		 * will trigger might_sleep() even if it won't actually sleep,
+		 * which is the case when the fb has already been pinned.
+		 */
+		if (phys_cursor)
+			plane_state->phys_dma_addr =
+				i915_gem_object_get_dma_address(intel_fb_obj(fb), 0);
 	} else {
 		struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 

drivers/gpu/drm/i915/display/intel_sdvo.c

@@ -1842,8 +1842,6 @@ static void intel_disable_sdvo(struct intel_atomic_state *state,
 	struct intel_crtc *crtc = to_intel_crtc(old_crtc_state->uapi.crtc);
 	u32 temp;
 
-	encoder->audio_disable(encoder, old_crtc_state, conn_state);
-
 	intel_sdvo_set_active_outputs(intel_sdvo, 0);
 	if (0)
 		intel_sdvo_set_encoder_power_state(intel_sdvo,
@@ -1935,8 +1933,6 @@ static void intel_enable_sdvo(struct intel_atomic_state *state,
 		intel_sdvo_set_encoder_power_state(intel_sdvo,
 						   DRM_MODE_DPMS_ON);
 	intel_sdvo_set_active_outputs(intel_sdvo, intel_sdvo_connector->output_flag);
-
-	encoder->audio_enable(encoder, pipe_config, conn_state);
 }
 
 static enum drm_mode_status

drivers/gpu/drm/i915/display/intel_vrr.c

@@ -187,10 +187,11 @@ void intel_vrr_set_transcoder_timings(const struct intel_crtc_state *crtc_state)
 	enum transcoder cpu_transcoder = crtc_state->cpu_transcoder;
 
 	/*
-	 * TRANS_SET_CONTEXT_LATENCY with VRR enabled
-	 * requires this chicken bit on ADL/DG2.
+	 * This bit seems to have two meanings depending on the platform:
+	 * TGL: generate VRR "safe window" for DSB vblank waits
+	 * ADL/DG2: make TRANS_SET_CONTEXT_LATENCY effective with VRR
 	 */
-	if (DISPLAY_VER(dev_priv) == 13)
+	if (IS_DISPLAY_VER(dev_priv, 12, 13))
 		intel_de_rmw(dev_priv, CHICKEN_TRANS(cpu_transcoder),
 			     0, PIPE_VBLANK_WITH_DELAY);
 

drivers/gpu/drm/i915/display/skl_universal_plane.c

@@ -2295,6 +2295,9 @@ static u8 skl_get_plane_caps(struct drm_i915_private *i915,
 	if (HAS_4TILE(i915))
 		caps |= INTEL_PLANE_CAP_TILING_4;
 
+	if (!IS_ENABLED(I915) && !HAS_FLAT_CCS(i915))
+		return caps;
+
 	if (skl_plane_has_rc_ccs(i915, pipe, plane_id)) {
 		caps |= INTEL_PLANE_CAP_CCS_RC;
 		if (DISPLAY_VER(i915) >= 12)

drivers/gpu/drm/i915/gt/intel_engine_pm.c

@@ -279,9 +279,6 @@ static int __engine_park(struct intel_wakeref *wf)
 	intel_engine_park_heartbeat(engine);
 	intel_breadcrumbs_park(engine->breadcrumbs);
 
-	/* Must be reset upon idling, or we may miss the busy wakeup. */
-	GEM_BUG_ON(engine->sched_engine->queue_priority_hint != INT_MIN);
-
 	if (engine->park)
 		engine->park(engine);
 

drivers/gpu/drm/i915/gt/intel_execlists_submission.c

@@ -3272,6 +3272,9 @@ static void execlists_park(struct intel_engine_cs *engine)
 {
 	cancel_timer(&engine->execlists.timer);
 	cancel_timer(&engine->execlists.preempt);
+
+	/* Reset upon idling, or we may delay the busy wakeup. */
+	WRITE_ONCE(engine->sched_engine->queue_priority_hint, INT_MIN);
 }
 
 static void add_to_engine(struct i915_request *rq)

drivers/gpu/drm/i915/gt/intel_workarounds.c

@@ -1653,6 +1653,7 @@ static void
 xelpg_gt_workarounds_init(struct intel_gt *gt, struct i915_wa_list *wal)
 {
 	/* Wa_14018575942 / Wa_18018781329 */
+	wa_mcr_write_or(wal, RENDER_MOD_CTRL, FORCE_MISS_FTLB);
 	wa_mcr_write_or(wal, COMP_MOD_CTRL, FORCE_MISS_FTLB);
 
 	/* Wa_22016670082 */

drivers/gpu/drm/i915/i915_driver.c

@@ -800,7 +800,7 @@ int i915_driver_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 		goto out_cleanup_modeset2;
 
 	ret = intel_pxp_init(i915);
-	if (ret != -ENODEV)
+	if (ret && ret != -ENODEV)
 		drm_dbg(&i915->drm, "pxp init failed with %d\n", ret);
 
 	ret = intel_display_driver_probe(i915);

drivers/gpu/drm/i915/i915_hwmon.c

@@ -72,12 +72,13 @@ hwm_locked_with_pm_intel_uncore_rmw(struct hwm_drvdata *ddat,
 	struct intel_uncore *uncore = ddat->uncore;
 	intel_wakeref_t wakeref;
 
-	mutex_lock(&hwmon->hwmon_lock);
+	with_intel_runtime_pm(uncore->rpm, wakeref) {
+		mutex_lock(&hwmon->hwmon_lock);
 
-	with_intel_runtime_pm(uncore->rpm, wakeref)
 		intel_uncore_rmw(uncore, reg, clear, set);
 
-	mutex_unlock(&hwmon->hwmon_lock);
+		mutex_unlock(&hwmon->hwmon_lock);
+	}
 }
 
 /*
@@ -136,20 +137,21 @@ hwm_energy(struct hwm_drvdata *ddat, long *energy)
 	else
 		rgaddr = hwmon->rg.energy_status_all;
 
-	mutex_lock(&hwmon->hwmon_lock);
+	with_intel_runtime_pm(uncore->rpm, wakeref) {
+		mutex_lock(&hwmon->hwmon_lock);
 
-	with_intel_runtime_pm(uncore->rpm, wakeref)
 		reg_val = intel_uncore_read(uncore, rgaddr);
 
-	if (reg_val >= ei->reg_val_prev)
-		ei->accum_energy += reg_val - ei->reg_val_prev;
-	else
-		ei->accum_energy += UINT_MAX - ei->reg_val_prev + reg_val;
-	ei->reg_val_prev = reg_val;
+		if (reg_val >= ei->reg_val_prev)
+			ei->accum_energy += reg_val - ei->reg_val_prev;
+		else
+			ei->accum_energy += UINT_MAX - ei->reg_val_prev + reg_val;
+		ei->reg_val_prev = reg_val;
 
-	*energy = mul_u64_u32_shr(ei->accum_energy, SF_ENERGY,
-				  hwmon->scl_shift_energy);
-	mutex_unlock(&hwmon->hwmon_lock);
+		*energy = mul_u64_u32_shr(ei->accum_energy, SF_ENERGY,
+					  hwmon->scl_shift_energy);
+		mutex_unlock(&hwmon->hwmon_lock);
+	}
 }
 
 static ssize_t
@@ -404,6 +406,7 @@ hwm_power_max_write(struct hwm_drvdata *ddat, long val)
 
 	/* Block waiting for GuC reset to complete when needed */
 	for (;;) {
+		wakeref = intel_runtime_pm_get(ddat->uncore->rpm);
 		mutex_lock(&hwmon->hwmon_lock);
 
 		prepare_to_wait(&ddat->waitq, &wait, TASK_INTERRUPTIBLE);
@@ -417,14 +420,13 @@ hwm_power_max_write(struct hwm_drvdata *ddat, long val)
 		}
 
 		mutex_unlock(&hwmon->hwmon_lock);
+		intel_runtime_pm_put(ddat->uncore->rpm, wakeref);
 
 		schedule();
 	}
 	finish_wait(&ddat->waitq, &wait);
 	if (ret)
-		goto unlock;
-
-	wakeref = intel_runtime_pm_get(ddat->uncore->rpm);
+		goto exit;
 
 	/* Disable PL1 limit and verify, because the limit cannot be disabled on all platforms */
 	if (val == PL1_DISABLE) {
@@ -444,9 +446,8 @@ hwm_power_max_write(struct hwm_drvdata *ddat, long val)
 	intel_uncore_rmw(ddat->uncore, hwmon->rg.pkg_rapl_limit,
 			 PKG_PWR_LIM_1_EN | PKG_PWR_LIM_1, nval);
 exit:
-	intel_runtime_pm_put(ddat->uncore->rpm, wakeref);
-unlock:
 	mutex_unlock(&hwmon->hwmon_lock);
+	intel_runtime_pm_put(ddat->uncore->rpm, wakeref);
 	return ret;
 }
 

drivers/gpu/drm/i915/i915_memcpy.c

@@ -26,6 +26,7 @@
 #include <linux/string.h>
 #include <linux/cpufeature.h>
 #include <linux/bug.h>
+#include <linux/build_bug.h>
 #include <asm/fpu/api.h>
 
 #include "i915_memcpy.h"

drivers/gpu/drm/i915/i915_reg.h

@@ -4599,7 +4599,7 @@
 #define MTL_CHICKEN_TRANS(trans)	_MMIO_TRANS((trans), \
 						    _MTL_CHICKEN_TRANS_A, \
 						    _MTL_CHICKEN_TRANS_B)
-#define   PIPE_VBLANK_WITH_DELAY	REG_BIT(31) /* ADL/DG2 */
+#define   PIPE_VBLANK_WITH_DELAY	REG_BIT(31) /* tgl+ */
 #define   SKL_UNMASK_VBL_TO_PIPE_IN_SRD	REG_BIT(30) /* skl+ */
 #define   HSW_FRAME_START_DELAY_MASK	REG_GENMASK(28, 27)
 #define   HSW_FRAME_START_DELAY(x)	REG_FIELD_PREP(HSW_FRAME_START_DELAY_MASK, x)

drivers/gpu/drm/i915/i915_vma.c

@@ -34,6 +34,7 @@
 #include "gt/intel_engine.h"
 #include "gt/intel_engine_heartbeat.h"
 #include "gt/intel_gt.h"
+#include "gt/intel_gt_pm.h"
 #include "gt/intel_gt_requests.h"
 #include "gt/intel_tlb.h"
 
@@ -103,12 +104,42 @@ static inline struct i915_vma *active_to_vma(struct i915_active *ref)
 
 static int __i915_vma_active(struct i915_active *ref)
 {
-	return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
+	struct i915_vma *vma = active_to_vma(ref);
+
+	if (!i915_vma_tryget(vma))
+		return -ENOENT;
+
+	/*
+	 * Exclude global GTT VMA from holding a GT wakeref
+	 * while active, otherwise GPU never goes idle.
+	 */
+	if (!i915_vma_is_ggtt(vma)) {
+		/*
+		 * Since we and our _retire() counterpart can be
+		 * called asynchronously, storing a wakeref tracking
+		 * handle inside struct i915_vma is not safe, and
+		 * there is no other good place for that.  Hence,
+		 * use untracked variants of intel_gt_pm_get/put().
+		 */
+		intel_gt_pm_get_untracked(vma->vm->gt);
+	}
+
+	return 0;
 }
 
 static void __i915_vma_retire(struct i915_active *ref)
 {
-	i915_vma_put(active_to_vma(ref));
+	struct i915_vma *vma = active_to_vma(ref);
+
+	if (!i915_vma_is_ggtt(vma)) {
+		/*
+		 * Since we can be called from atomic contexts,
+		 * use an async variant of intel_gt_pm_put().
+		 */
+		intel_gt_pm_put_async_untracked(vma->vm->gt);
+	}
+
+	i915_vma_put(vma);
 }
 
 static struct i915_vma *
@@ -1404,7 +1435,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 	struct i915_vma_work *work = NULL;
 	struct dma_fence *moving = NULL;
 	struct i915_vma_resource *vma_res = NULL;
-	intel_wakeref_t wakeref = 0;
+	intel_wakeref_t wakeref;
 	unsigned int bound;
 	int err;
 
@@ -1424,8 +1455,14 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 	if (err)
 		return err;
 
-	if (flags & PIN_GLOBAL)
-		wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
+	/*
+	 * In case of a global GTT, we must hold a runtime-pm wakeref
+	 * while global PTEs are updated.  In other cases, we hold
+	 * the rpm reference while the VMA is active.  Since runtime
+	 * resume may require allocations, which are forbidden inside
+	 * vm->mutex, get the first rpm wakeref outside of the mutex.
+	 */
+	wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
 
 	if (flags & vma->vm->bind_async_flags) {
 		/* lock VM */
@@ -1561,8 +1598,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
 	if (work)
 		dma_fence_work_commit_imm(&work->base);
 err_rpm:
-	if (wakeref)
-		intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
+	intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
 
 	if (moving)
 		dma_fence_put(moving);