Quelle  vlv_suspend.c   Sprache: C

// SPDX-License-Identifier: MIT
/*
 * Copyright © 2020 Intel Corporation
 */

#include <linux/string_helpers.h>
#include <linux/kernel.h>

#include <drm/drm_print.h>

#include "i915_drv.h"
#include "i915_reg.h"
#include "i915_trace.h"
#include "i915_utils.h"
#include "intel_clock_gating.h"
#include "intel_uncore_trace.h"
#include "vlv_suspend.h"

#include "gt/intel_gt_regs.h"

struct vlv_s0ix_state {
 /* GAM */
 u32 wr_watermark;
 u32 gfx_prio_ctrl;
 u32 arb_mode;
 u32 gfx_pend_tlb0;
 u32 gfx_pend_tlb1;
 u32 lra_limits[GEN7_LRA_LIMITS_REG_NUM];
 u32 media_max_req_count;
 u32 gfx_max_req_count;
 u32 render_hwsp;
 u32 ecochk;
 u32 bsd_hwsp;
 u32 blt_hwsp;
 u32 tlb_rd_addr;

 /* MBC */
 u32 g3dctl;
 u32 gsckgctl;
 u32 mbctl;

 /* GCP */
 u32 ucgctl1;
 u32 ucgctl3;
 u32 rcgctl1;
 u32 rcgctl2;
 u32 rstctl;
 u32 misccpctl;

 /* GPM */
 u32 gfxpause;
 u32 rpdeuhwtc;
 u32 rpdeuc;
 u32 ecobus;
 u32 pwrdwnupctl;
 u32 rp_down_timeout;
 u32 rp_deucsw;
 u32 rcubmabdtmr;
 u32 rcedata;
 u32 spare2gh;

 /* Display 1 CZ domain */
 u32 gt_imr;
 u32 gt_ier;
 u32 pm_imr;
 u32 pm_ier;
 u32 gt_scratch[GEN7_GT_SCRATCH_REG_NUM];

 /* GT SA CZ domain */
 u32 tilectl;
 u32 gt_fifoctl;
 u32 gtlc_wake_ctrl;
 u32 gtlc_survive;
 u32 pmwgicz;

 /* Display 2 CZ domain */
 u32 gu_ctl0;
 u32 gu_ctl1;
 u32 pcbr;
 u32 clock_gate_dis2;
};

/*
 * Save all Gunit registers that may be lost after a D3 and a subsequent
 * S0i[R123] transition. The list of registers needing a save/restore is
 * defined in the VLV2_S0IXRegs document. This documents marks all Gunit
 * registers in the following way:
 * - Driver: saved/restored by the driver
 * - Punit : saved/restored by the Punit firmware
 * - No, w/o marking: no need to save/restore, since the register is R/O or
 *                    used internally by the HW in a way that doesn't depend
 *                    keeping the content across a suspend/resume.
 * - Debug : used for debugging
 *
 * We save/restore all registers marked with 'Driver', with the following
 * exceptions:
 * - Registers out of use, including also registers marked with 'Debug'.
 *   These have no effect on the driver's operation, so we don't save/restore
 *   them to reduce the overhead.
 * - Registers that are fully setup by an initialization function called from
 *   the resume path. For example many clock gating and RPS/RC6 registers.
 * - Registers that provide the right functionality with their reset defaults.
 *
 * TODO: Except for registers that based on the above 3 criteria can be safely
 * ignored, we save/restore all others, practically treating the HW context as
 * a black-box for the driver. Further investigation is needed to reduce the
 * saved/restored registers even further, by following the same 3 criteria.
 */
static void vlv_save_gunit_s0ix_state(struct drm_i915_private *i915)
{
 struct vlv_s0ix_state *s = i915->vlv_s0ix_state;
 struct intel_uncore *uncore = &i915->uncore;
 int i;

 if (!s)
  return;

 /* GAM 0x4000-0x4770 */
 s->wr_watermark = intel_uncore_read(uncore, GEN7_WR_WATERMARK);
 s->gfx_prio_ctrl = intel_uncore_read(uncore, GEN7_GFX_PRIO_CTRL);
 s->arb_mode = intel_uncore_read(uncore, ARB_MODE);
 s->gfx_pend_tlb0 = intel_uncore_read(uncore, GEN7_GFX_PEND_TLB0);
 s->gfx_pend_tlb1 = intel_uncore_read(uncore, GEN7_GFX_PEND_TLB1);

 for (i = 0; i < ARRAY_SIZE(s->lra_limits); i++)
  s->lra_limits[i] = intel_uncore_read(uncore, GEN7_LRA_LIMITS(i));

 s->media_max_req_count = intel_uncore_read(uncore, GEN7_MEDIA_MAX_REQ_COUNT);
 s->gfx_max_req_count = intel_uncore_read(uncore, GEN7_GFX_MAX_REQ_COUNT);

 s->render_hwsp = intel_uncore_read(uncore, RENDER_HWS_PGA_GEN7);
 s->ecochk = intel_uncore_read(uncore, GAM_ECOCHK);
 s->bsd_hwsp = intel_uncore_read(uncore, BSD_HWS_PGA_GEN7);
 s->blt_hwsp = intel_uncore_read(uncore, BLT_HWS_PGA_GEN7);

 s->tlb_rd_addr = intel_uncore_read(uncore, GEN7_TLB_RD_ADDR);

 /* MBC 0x9024-0x91D0, 0x8500 */
 s->g3dctl = intel_uncore_read(uncore, VLV_G3DCTL);
 s->gsckgctl = intel_uncore_read(uncore, VLV_GSCKGCTL);
 s->mbctl = intel_uncore_read(uncore, GEN6_MBCTL);

 /* GCP 0x9400-0x9424, 0x8100-0x810C */
 s->ucgctl1 = intel_uncore_read(uncore, GEN6_UCGCTL1);
 s->ucgctl3 = intel_uncore_read(uncore, GEN6_UCGCTL3);
 s->rcgctl1 = intel_uncore_read(uncore, GEN6_RCGCTL1);
 s->rcgctl2 = intel_uncore_read(uncore, GEN6_RCGCTL2);
 s->rstctl = intel_uncore_read(uncore, GEN6_RSTCTL);
 s->misccpctl = intel_uncore_read(uncore, GEN7_MISCCPCTL);

 /* GPM 0xA000-0xAA84, 0x8000-0x80FC */
 s->gfxpause = intel_uncore_read(uncore, GEN6_GFXPAUSE);
 s->rpdeuhwtc = intel_uncore_read(uncore, GEN6_RPDEUHWTC);
 s->rpdeuc = intel_uncore_read(uncore, GEN6_RPDEUC);
 s->ecobus = intel_uncore_read(uncore, ECOBUS);
 s->pwrdwnupctl = intel_uncore_read(uncore, VLV_PWRDWNUPCTL);
 s->rp_down_timeout = intel_uncore_read(uncore, GEN6_RP_DOWN_TIMEOUT);
 s->rp_deucsw = intel_uncore_read(uncore, GEN6_RPDEUCSW);
 s->rcubmabdtmr = intel_uncore_read(uncore, GEN6_RCUBMABDTMR);
 s->rcedata = intel_uncore_read(uncore, VLV_RCEDATA);
 s->spare2gh = intel_uncore_read(uncore, VLV_SPAREG2H);

 /* Display CZ domain, 0x4400C-0x4402C, 0x4F000-0x4F11F */
 s->gt_imr = intel_uncore_read(uncore, GTIMR);
 s->gt_ier = intel_uncore_read(uncore, GTIER);
 s->pm_imr = intel_uncore_read(uncore, GEN6_PMIMR);
 s->pm_ier = intel_uncore_read(uncore, GEN6_PMIER);

 for (i = 0; i < ARRAY_SIZE(s->gt_scratch); i++)
  s->gt_scratch[i] = intel_uncore_read(uncore, GEN7_GT_SCRATCH(i));

 /* GT SA CZ domain, 0x100000-0x138124 */
 s->tilectl = intel_uncore_read(uncore, TILECTL);
 s->gt_fifoctl = intel_uncore_read(uncore, GTFIFOCTL);
 s->gtlc_wake_ctrl = intel_uncore_read(uncore, VLV_GTLC_WAKE_CTRL);
 s->gtlc_survive = intel_uncore_read(uncore, VLV_GTLC_SURVIVABILITY_REG);
 s->pmwgicz = intel_uncore_read(uncore, VLV_PMWGICZ);

 /* Gunit-Display CZ domain, 0x182028-0x1821CF */
 s->gu_ctl0 = intel_uncore_read(uncore, VLV_GU_CTL0);
 s->gu_ctl1 = intel_uncore_read(uncore, VLV_GU_CTL1);
 s->pcbr = intel_uncore_read(uncore, VLV_PCBR);
 s->clock_gate_dis2 = intel_uncore_read(uncore, VLV_GUNIT_CLOCK_GATE2);

 /*
 * Not saving any of:
 * DFT, 0x9800-0x9EC0
 * SARB, 0xB000-0xB1FC
 * GAC, 0x5208-0x524C, 0x14000-0x14C000
 * PCI CFG
 */
}

static void vlv_restore_gunit_s0ix_state(struct drm_i915_private *i915)
{
 struct vlv_s0ix_state *s = i915->vlv_s0ix_state;
 struct intel_uncore *uncore = &i915->uncore;
 int i;

 if (!s)
  return;

 /* GAM 0x4000-0x4770 */
 intel_uncore_write(uncore, GEN7_WR_WATERMARK, s->wr_watermark);
 intel_uncore_write(uncore, GEN7_GFX_PRIO_CTRL, s->gfx_prio_ctrl);
 intel_uncore_write(uncore, ARB_MODE, s->arb_mode | (0xffff << 16));
 intel_uncore_write(uncore, GEN7_GFX_PEND_TLB0, s->gfx_pend_tlb0);
 intel_uncore_write(uncore, GEN7_GFX_PEND_TLB1, s->gfx_pend_tlb1);

 for (i = 0; i < ARRAY_SIZE(s->lra_limits); i++)
  intel_uncore_write(uncore, GEN7_LRA_LIMITS(i), s->lra_limits[i]);

 intel_uncore_write(uncore, GEN7_MEDIA_MAX_REQ_COUNT, s->media_max_req_count);
 intel_uncore_write(uncore, GEN7_GFX_MAX_REQ_COUNT, s->gfx_max_req_count);

 intel_uncore_write(uncore, RENDER_HWS_PGA_GEN7, s->render_hwsp);
 intel_uncore_write(uncore, GAM_ECOCHK, s->ecochk);
 intel_uncore_write(uncore, BSD_HWS_PGA_GEN7, s->bsd_hwsp);
 intel_uncore_write(uncore, BLT_HWS_PGA_GEN7, s->blt_hwsp);

 intel_uncore_write(uncore, GEN7_TLB_RD_ADDR, s->tlb_rd_addr);

 /* MBC 0x9024-0x91D0, 0x8500 */
 intel_uncore_write(uncore, VLV_G3DCTL, s->g3dctl);
 intel_uncore_write(uncore, VLV_GSCKGCTL, s->gsckgctl);
 intel_uncore_write(uncore, GEN6_MBCTL, s->mbctl);

 /* GCP 0x9400-0x9424, 0x8100-0x810C */
 intel_uncore_write(uncore, GEN6_UCGCTL1, s->ucgctl1);
 intel_uncore_write(uncore, GEN6_UCGCTL3, s->ucgctl3);
 intel_uncore_write(uncore, GEN6_RCGCTL1, s->rcgctl1);
 intel_uncore_write(uncore, GEN6_RCGCTL2, s->rcgctl2);
 intel_uncore_write(uncore, GEN6_RSTCTL, s->rstctl);
 intel_uncore_write(uncore, GEN7_MISCCPCTL, s->misccpctl);

 /* GPM 0xA000-0xAA84, 0x8000-0x80FC */
 intel_uncore_write(uncore, GEN6_GFXPAUSE, s->gfxpause);
 intel_uncore_write(uncore, GEN6_RPDEUHWTC, s->rpdeuhwtc);
 intel_uncore_write(uncore, GEN6_RPDEUC, s->rpdeuc);
 intel_uncore_write(uncore, ECOBUS, s->ecobus);
 intel_uncore_write(uncore, VLV_PWRDWNUPCTL, s->pwrdwnupctl);
 intel_uncore_write(uncore, GEN6_RP_DOWN_TIMEOUT, s->rp_down_timeout);
 intel_uncore_write(uncore, GEN6_RPDEUCSW, s->rp_deucsw);
 intel_uncore_write(uncore, GEN6_RCUBMABDTMR, s->rcubmabdtmr);
 intel_uncore_write(uncore, VLV_RCEDATA, s->rcedata);
 intel_uncore_write(uncore, VLV_SPAREG2H, s->spare2gh);

 /* Display CZ domain, 0x4400C-0x4402C, 0x4F000-0x4F11F */
 intel_uncore_write(uncore, GTIMR, s->gt_imr);
 intel_uncore_write(uncore, GTIER, s->gt_ier);
 intel_uncore_write(uncore, GEN6_PMIMR, s->pm_imr);
 intel_uncore_write(uncore, GEN6_PMIER, s->pm_ier);

 for (i = 0; i < ARRAY_SIZE(s->gt_scratch); i++)
  intel_uncore_write(uncore, GEN7_GT_SCRATCH(i), s->gt_scratch[i]);

 /* GT SA CZ domain, 0x100000-0x138124 */
 intel_uncore_write(uncore, TILECTL, s->tilectl);
 intel_uncore_write(uncore, GTFIFOCTL, s->gt_fifoctl);
 /*
 * Preserve the GT allow wake and GFX force clock bit, they are not
 * be restored, as they are used to control the s0ix suspend/resume
 * sequence by the caller.
 */
 intel_uncore_rmw(uncore, VLV_GTLC_WAKE_CTRL, ~VLV_GTLC_ALLOWWAKEREQ,
    s->gtlc_wake_ctrl & ~VLV_GTLC_ALLOWWAKEREQ);

 intel_uncore_rmw(uncore, VLV_GTLC_SURVIVABILITY_REG, ~VLV_GFX_CLK_FORCE_ON_BIT,
    s->gtlc_survive & ~VLV_GFX_CLK_FORCE_ON_BIT);

 intel_uncore_write(uncore, VLV_PMWGICZ, s->pmwgicz);

 /* Gunit-Display CZ domain, 0x182028-0x1821CF */
 intel_uncore_write(uncore, VLV_GU_CTL0, s->gu_ctl0);
 intel_uncore_write(uncore, VLV_GU_CTL1, s->gu_ctl1);
 intel_uncore_write(uncore, VLV_PCBR, s->pcbr);
 intel_uncore_write(uncore, VLV_GUNIT_CLOCK_GATE2, s->clock_gate_dis2);
}

static int vlv_wait_for_pw_status(struct drm_i915_private *i915,
      u32 mask, u32 val)
{
 i915_reg_t reg = VLV_GTLC_PW_STATUS;
 u32 reg_value;
 int ret;

 /* The HW does not like us polling for PW_STATUS frequently, so
 * use the sleeping loop rather than risk the busy spin within
 * intel_wait_for_register().
 *
 * Transitioning between RC6 states should be at most 2ms (see
 * valleyview_enable_rps) so use a 3ms timeout.
 */
 ret = wait_for(((reg_value =
    intel_uncore_read_notrace(&i915->uncore, reg)) & mask)
         == val, 3);

 /* just trace the final value */
 trace_i915_reg_rw(false, reg, reg_value, sizeof(reg_value), true);

 return ret;
}

static int vlv_force_gfx_clock(struct drm_i915_private *i915, bool force_on)
{
 struct intel_uncore *uncore = &i915->uncore;
 int err;

 intel_uncore_rmw(uncore, VLV_GTLC_SURVIVABILITY_REG, VLV_GFX_CLK_FORCE_ON_BIT,
    force_on ? VLV_GFX_CLK_FORCE_ON_BIT : 0);

 if (!force_on)
  return 0;

 err = intel_wait_for_register(uncore,
          VLV_GTLC_SURVIVABILITY_REG,
          VLV_GFX_CLK_STATUS_BIT,
          VLV_GFX_CLK_STATUS_BIT,
          20);
 if (err)
  drm_err(&i915->drm,
   "timeout waiting for GFX clock force-on (%08x)\n",
   intel_uncore_read(uncore, VLV_GTLC_SURVIVABILITY_REG));

 return err;
}

static int vlv_allow_gt_wake(struct drm_i915_private *i915, bool allow)
{
 struct intel_uncore *uncore = &i915->uncore;
 u32 mask;
 u32 val;
 int err;

 intel_uncore_rmw(uncore, VLV_GTLC_WAKE_CTRL, VLV_GTLC_ALLOWWAKEREQ,
    allow ? VLV_GTLC_ALLOWWAKEREQ : 0);
 intel_uncore_posting_read(uncore, VLV_GTLC_WAKE_CTRL);

 mask = VLV_GTLC_ALLOWWAKEACK;
 val = allow ? mask : 0;

 err = vlv_wait_for_pw_status(i915, mask, val);
 if (err)
  drm_err(&i915->drm, "timeout disabling GT waking\n");

 return err;
}

static void vlv_wait_for_gt_wells(struct drm_i915_private *dev_priv,
      bool wait_for_on)
{
 u32 mask;
 u32 val;

 mask = VLV_GTLC_PW_MEDIA_STATUS_MASK | VLV_GTLC_PW_RENDER_STATUS_MASK;
 val = wait_for_on ? mask : 0;

 /*
 * RC6 transitioning can be delayed up to 2 msec (see
 * valleyview_enable_rps), use 3 msec for safety.
 *
 * This can fail to turn off the rc6 if the GPU is stuck after a failed
 * reset and we are trying to force the machine to sleep.
 */
 if (vlv_wait_for_pw_status(dev_priv, mask, val))
  drm_dbg(&dev_priv->drm,
   "timeout waiting for GT wells to go %s\n",
   str_on_off(wait_for_on));
}

static void vlv_check_no_gt_access(struct drm_i915_private *i915)
{
 struct intel_uncore *uncore = &i915->uncore;

 if (!(intel_uncore_read(uncore, VLV_GTLC_PW_STATUS) & VLV_GTLC_ALLOWWAKEERR))
  return;

 drm_dbg(&i915->drm, "GT register access while GT waking disabled\n");
 intel_uncore_write(uncore, VLV_GTLC_PW_STATUS, VLV_GTLC_ALLOWWAKEERR);
}

int vlv_suspend_complete(struct drm_i915_private *dev_priv)
{
 u32 mask;
 int err;

 if (!IS_VALLEYVIEW(dev_priv) && !IS_CHERRYVIEW(dev_priv))
  return 0;

 /*
 * Bspec defines the following GT well on flags as debug only, so
 * don't treat them as hard failures.
 */
 vlv_wait_for_gt_wells(dev_priv, false);

 mask = VLV_GTLC_RENDER_CTX_EXISTS | VLV_GTLC_MEDIA_CTX_EXISTS;
 drm_WARN_ON(&dev_priv->drm,
      (intel_uncore_read(&dev_priv->uncore, VLV_GTLC_WAKE_CTRL) & mask) != mask);

 vlv_check_no_gt_access(dev_priv);

 err = vlv_force_gfx_clock(dev_priv, true);
 if (err)
  goto err1;

 err = vlv_allow_gt_wake(dev_priv, false);
 if (err)
  goto err2;

 vlv_save_gunit_s0ix_state(dev_priv);

 err = vlv_force_gfx_clock(dev_priv, false);
 if (err)
  goto err2;

 return 0;

err2:
 /* For safety always re-enable waking and disable gfx clock forcing */
 vlv_allow_gt_wake(dev_priv, true);
err1:
 vlv_force_gfx_clock(dev_priv, false);

 return err;
}

int vlv_resume_prepare(struct drm_i915_private *dev_priv, bool rpm_resume)
{
 int err;
 int ret;

 if (!IS_VALLEYVIEW(dev_priv) && !IS_CHERRYVIEW(dev_priv))
  return 0;

 /*
 * If any of the steps fail just try to continue, that's the best we
 * can do at this point. Return the first error code (which will also
 * leave RPM permanently disabled).
 */
 ret = vlv_force_gfx_clock(dev_priv, true);

 vlv_restore_gunit_s0ix_state(dev_priv);

 err = vlv_allow_gt_wake(dev_priv, true);
 if (!ret)
  ret = err;

 err = vlv_force_gfx_clock(dev_priv, false);
 if (!ret)
  ret = err;

 vlv_check_no_gt_access(dev_priv);

 if (rpm_resume)
  intel_clock_gating_init(dev_priv);

 return ret;
}

int vlv_suspend_init(struct drm_i915_private *i915)
{
 if (!IS_VALLEYVIEW(i915))
  return 0;

 /* we write all the values in the struct, so no need to zero it out */
 i915->vlv_s0ix_state = kmalloc(sizeof(*i915->vlv_s0ix_state),
           GFP_KERNEL);
 if (!i915->vlv_s0ix_state)
  return -ENOMEM;

 return 0;
}

void vlv_suspend_cleanup(struct drm_i915_private *i915)
{
 if (!i915->vlv_s0ix_state)
  return;

 kfree(i915->vlv_s0ix_state);
 i915->vlv_s0ix_state = NULL;
}