GEN-9 Arbitrated Bandwidth WM WA's & IPC
Subsystems (2)
v7 cover GEN-9 Arbitrated Bandwidth WM WA's & IPC
This series implements following set of functionality
Implement IPC WA's for Broxton/KBL
Enable IPC in supported platforms
Convert WM calculation to fixed point calculation Calculation of System memory Bandwidth for SKL/KBL/BXT Implementation of Arbitrated memory Bandwidth related WM WA's
Mahesh Kumar (8):
drm/i915/skl: Add variables to check x_tile and y_tile drm/i915/bxt: IPC WA for Broxton
drm/i915/kbl: IPC workaround for kabylake drm/i915/bxt: Enable IPC support
drm/i915/skl+: change WM calc to fixed point 16.16 drm/i915: Add intel_atomic_get_existing_crtc_state function drm/i915: Decode system memory bandwidth drm/i915/gen9: WM memory bandwidth related workaround
drivers/gpu/drm/i915/i915_drv.c | 175 ++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 109 ++++++++++++++ drivers/gpu/drm/i915/i915_reg.h | 38 +++++ drivers/gpu/drm/i915/intel_display.c | 24 ++++ drivers/gpu/drm/i915/intel_drv.h | 15 ++ drivers/gpu/drm/i915/intel_pm.c | 272 +++++++++++++++++++++++++++++------ 6 files changed, 586 insertions(+), 47 deletions(-)
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 1/8 drm/i915/skl: Add variables to check x_tile and y_tile
This patch adds variable to check for X_tiled & y_tiled planes, instead of always checking against framebuffer-modifiers.
Changes:
- Created separate patch as per Paulo's comment
- Added x_tiled variable as well
Changes since V2:
- Incorporate Paulo's comments
- Rebase
--- drivers/gpu/drm/i915/intel_pm.c | 22 ++++++++++++---------- 1 file changed, 12 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/intel_pm.c +12 -10
@@ -3579,13 +3579,18 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,struct intel_atomic_state *state =to_intel_atomic_state (cstate->base.state);bool apply_memory_bw_wa =skl_needs_memory_bw_wa (state); +bool y_tiled, x_tiled;if (latency ==0 || !cstate->base.active || !intel_pstate->base.visible) { *enabled = false;return 0 ; } -if (apply_memory_bw_wa && fb->modifier ==I915_FORMAT_MOD_X_TILED ) + y_tiled = fb->modifier ==I915_FORMAT_MOD_Y_TILED || + fb->modifier == I915_FORMAT_MOD_Yf_TILED; + x_tiled = fb->modifier ==I915_FORMAT_MOD_X_TILED ; + +if (apply_memory_bw_wa && x_tiled) latency +=15 ; width =drm_rect_width (&intel_pstate->base.src) >>16 ;
@@ -3624,16 +3629,15 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, y_min_scanlines *=2 ; plane_bytes_per_line = width * cpp; -if (fb->modifier ==I915_FORMAT_MOD_Y_TILED || - fb->modifier == I915_FORMAT_MOD_Yf_TILED) { +if (y_tiled) { plane_blocks_per_line =DIV_ROUND_UP (plane_bytes_per_line * y_min_scanlines,512 ); plane_blocks_per_line /= y_min_scanlines; - }else if (fb->modifier ==DRM_FORMAT_MOD_NONE ) { + }else if (x_tiled) { + plane_blocks_per_line =DIV_ROUND_UP (plane_bytes_per_line,512 ); + }else { plane_blocks_per_line =DIV_ROUND_UP (plane_bytes_per_line,512 ) +1 ; - }else { - plane_blocks_per_line =DIV_ROUND_UP (plane_bytes_per_line,512 ); } method1 =skl_wm_method1 (plane_pixel_rate, cpp, latency);
@@ -3644,8 +3648,7 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, y_tile_minimum = plane_blocks_per_line * y_min_scanlines; -if (fb->modifier ==I915_FORMAT_MOD_Y_TILED || - fb->modifier == I915_FORMAT_MOD_Yf_TILED) { +if (y_tiled) { selected_result =max (method2, y_tile_minimum); }else { if ((cpp * cstate->base.adjusted_mode.crtc_htotal / 512 < 1) &&
@@ -3661,8 +3664,7 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, res_lines =DIV_ROUND_UP (selected_result, plane_blocks_per_line);if (level >=1 && level <=7 ) { -if (fb->modifier ==I915_FORMAT_MOD_Y_TILED || - fb->modifier == I915_FORMAT_MOD_Yf_TILED) { +if (y_tiled) { res_blocks += y_tile_minimum; res_lines += y_min_scanlines; }else {
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 2/8 drm/i915/bxt: IPC WA for Broxton
In reply to Mahesh KumarDisplay Workarounds #1135
If IPC is enabled in BXT, display underruns are observed.
WA: The Line Time programmed in the WM_LINETIME register should be half of the actual calculated Line Time.
Programmed Line Time = 1/2*Calculated Line Time
Changes since V1:
- Add Workaround number in commit & code
--- drivers/gpu/drm/i915/i915_drv.c | 2 ++ drivers/gpu/drm/i915/i915_drv.h | 2 ++ drivers/gpu/drm/i915/intel_pm.c | 12 ++++++++++-- 3 files changed, 14 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +2 -0
@@ -1251,6 +1251,8 @@ int i915_driver_load(struct pci_dev *pdev, const struct pci_device_id *ent)intel_runtime_pm_enable (dev_priv); + dev_priv->ipc_enabled = false; +/* Everything is in place, we can now relax! */ DRM_INFO("Initialized %s %d.%d.%d %s for %s on minor %d\n", driver.name, driver.major, driver.minor, driver.patchlevel,
drivers/gpu/drm/i915/i915_drv.h +2 -0
@@ -2306,6 +2306,8 @@ struct drm_i915_private {/* perform PHY state sanity checks? */ bool chv_phy_assert[2 ]; +bool ipc_enabled; +/* Used to save the pipe-to-encoder mapping for audio */ struct intel_encoder *av_enc_map[I915_MAX_PIPES ];
drivers/gpu/drm/i915/intel_pm.c +10 -2
@@ -3755,7 +3755,10 @@ skl_compute_wm_level(const struct drm_i915_private *dev_priv,static uint32_t skl_compute_linetime_wm (struct intel_crtc_state *cstate) { +struct drm_atomic_state *state = cstate->base.state; +struct drm_i915_private *dev_priv =to_i915 (state->dev);uint32_t pixel_rate; +uint32_t linetime_wm;if (!cstate->base.active)return 0 ;
@@ -3765,8 +3768,13 @@ skl_compute_linetime_wm(struct intel_crtc_state *cstate)if (WARN_ON (pixel_rate ==0 ))return 0 ; -return DIV_ROUND_UP (8 * cstate->base.adjusted_mode.crtc_htotal *1000 , - pixel_rate); + linetime_wm =DIV_ROUND_UP (8 * cstate->base.adjusted_mode.crtc_htotal * +1000 , pixel_rate); +/* Display WA #1135 for BXT:All */ +if (IS_BROXTON (dev_priv) && dev_priv->ipc_enabled) + linetime_wm =DIV_ROUND_UP (linetime_wm,2 ); + +return linetime_wm; } static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 3/8 drm/i915/kbl: IPC workaround for kabylake
In reply to Mahesh KumarDisplay Workarounds #1141
IPC (Isoch Priority Control) may cause underflows.
KBL WA: When IPC is enabled, watermark latency values must be increased by 4us across all levels. This brings level 0 up to 6us.
Changes since V1:
- Add Workaround number in commit & code
--- drivers/gpu/drm/i915/intel_pm.c | 4 ++++ 1 file changed, 4 insertions(+)
drivers/gpu/drm/i915/intel_pm.c +4 -0
@@ -3590,6 +3590,10 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, fb->modifier == I915_FORMAT_MOD_Yf_TILED; x_tiled = fb->modifier ==I915_FORMAT_MOD_X_TILED ; +/* IPC WA for kabylake Display WA #1141 */ +if (IS_KABYLAKE (dev_priv) && dev_priv->ipc_enabled) + latency +=4 ; +if (apply_memory_bw_wa && x_tiled) latency +=15 ;
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 4/8 drm/i915/bxt: Enable IPC support
In reply to Mahesh KumarThis patch adds IPC support for platforms. This patch enables IPC only for BXT/KBL platform as for SKL recommendation is to keep is disabled. IPC (Isochronous Priority Control) is the hardware feature, which dynamically controles the memory read priority of Display.
When IPC is enabled, plane read requests are sent at high priority until filling above the transition watermark, then the requests are sent at lower priority until dropping below the level 0 watermark. The lower priority requests allow other memory clients to have better memory access. When IPC is disabled, all plane read requests are sent at high priority.
Changes since V1:
- Remove commandline parameter to disable ipc
- Address Paulo's comments
Changes since V2:
- Address review comments
- Set ipc_enabled flag
Changes since V3:
- move ipc_enabled flag assignment inside intel_ipc_enable function
--- drivers/gpu/drm/i915/i915_drv.c | 2 +- drivers/gpu/drm/i915/i915_reg.h | 1 + drivers/gpu/drm/i915/intel_drv.h | 1 + drivers/gpu/drm/i915/intel_pm.c | 16 ++++++++++++++++ 4 files changed, 19 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_drv.c +1 -1
@@ -1251,7 +1251,7 @@ int i915_driver_load(struct pci_dev *pdev, const struct pci_device_id *ent)intel_runtime_pm_enable (dev_priv); - dev_priv->ipc_enabled = false; +intel_enable_ipc (dev_priv);/* Everything is in place, we can now relax! */ DRM_INFO("Initialized %s %d.%d.%d %s for %s on minor %d\n",
drivers/gpu/drm/i915/i915_reg.h +1 -0
@@ -6418,6 +6418,7 @@ enum {#define DISP_FBC_WM_DIS (1<<15)#define DISP_ARB_CTL2 _MMIO(0x45004)#define DISP_DATA_PARTITION_5_6 (1<<6) +#define DISP_IPC_ENABLE (1<<3)#define DBUF_CTL _MMIO(0x45008)#define DBUF_POWER_REQUEST (1<<31)#define DBUF_POWER_STATE (1<<30)
drivers/gpu/drm/i915/intel_drv.h +1 -0
@@ -1755,6 +1755,7 @@ bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry **entries,uint32_t ilk_pipe_pixel_rate (const struct intel_crtc_state *pipe_config);bool ilk_disable_lp_wm (struct drm_device *dev);int sanitize_rc6_option (struct drm_i915_private *dev_priv,int enable_rc6); +void intel_enable_ipc (struct drm_i915_private *dev_priv);static inline int intel_enable_rc6 (void ) {return i915.enable_rc6;
drivers/gpu/drm/i915/intel_pm.c +16 -0
@@ -4655,6 +4655,22 @@ void intel_update_watermarks(struct intel_crtc *crtc) dev_priv->display.update_wm (crtc); } +void intel_enable_ipc (struct drm_i915_private *dev_priv) +{ +u32 val; + + dev_priv->ipc_enabled = false; +if (!(IS_BROXTON (dev_priv) ||IS_KABYLAKE (dev_priv))) +return ; + + val =I915_READ (DISP_ARB_CTL2 ); + + val |=DISP_IPC_ENABLE ; + +I915_WRITE (DISP_ARB_CTL2 , val); + dev_priv->ipc_enabled = true; +} +/* * Lock protecting IPS related data structures */
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 5/8 drm/i915/skl+: change WM calc to fixed point 16.16
In reply to Mahesh KumarThis patch changes Watermak calculation to fixed point calculation. Problem with current calculation is during plane_blocks_per_line calculation we divide intermediate blocks with min_scanlines and takes floor of the result because of integer operation. hence we end-up assigning less blocks than required. Which leads to flickers.
Changes since V1:
- Add fixed point data type as per Paulo's review
Changes since V2:
- use fixed_point instead of fp_16_16
Changes since V3:
- rebase
--- drivers/gpu/drm/i915/i915_drv.h | 84 +++++++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/intel_pm.c | 69 +++++++++++++++++++-------------- 2 files changed, 124 insertions(+), 29 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +84 -0
@@ -119,6 +119,90 @@ bool __i915_inject_load_failure(const char *func, int line);#define i915_inject_load_failure () \ __i915_inject_load_failure(__func__, __LINE__) +typedef struct { +uint32_t val; +}uint_fixed_16_16_t ; + +#define FP_16_16_MAX ({ \ + uint_fixed_16_16_t fp; \ + fp.val = UINT_MAX; \ + fp; \ +}) + +static inline uint_fixed_16_16_t u32_to_fixed_point (uint32_t val) +{ +uint_fixed_16_16_t fp; + +WARN_ON (val >>16 ); + + fp.val = val <<16 ; +return fp; +} + +static inline uint32_t fixed_point_to_u32_round_up (uint_fixed_16_16_t fp) +{ +return DIV_ROUND_UP (fp.val,1 <<16 ); +} + +static inline uint32_t fixed_point_to_u32 (uint_fixed_16_16_t fp) +{ +return (fp.val / (1 <<16 )); +} + +static inline uint_fixed_16_16_t min_fixed_point (uint_fixed_16_16_t min1, +uint_fixed_16_16_t min2) +{ +uint_fixed_16_16_t min; + + min.val =min (min1.val, min2.val); +return min; +} + +static inline uint_fixed_16_16_t max_fixed_point (uint_fixed_16_16_t max1, +uint_fixed_16_16_t max2) +{ +uint_fixed_16_16_t max; + + max.val =max (max1.val, max2.val); +return max; +} + +static inline uint_fixed_16_16_t fixed_point_div_round_up (uint32_t val, +uint32_t d) +{ +uint_fixed_16_16_t fp, res; + + fp =u32_to_fixed_point (val); + res.val =DIV_ROUND_UP (fp.val, d); +return res; +} + +static inline uint_fixed_16_16_t fixed_point_div_round_up_u64 (uint32_t val, +uint32_t d) +{ +uint_fixed_16_16_t res; +uint64_t interm_val; + + interm_val = (uint64_t )val <<16 ; + interm_val =DIV_ROUND_UP_ULL (interm_val, d); +WARN_ON (interm_val >>32 ); + res.val = (uint32_t ) interm_val; + +return res; +} + +static inline uint_fixed_16_16_t mul_fixed_point_u32 (uint32_t val, +uint_fixed_16_16_t mul) +{ +uint64_t intermediate_val; +uint_fixed_16_16_t fp; + + intermediate_val = (uint64_t ) val * mul.val; +WARN_ON (intermediate_val >>32 ); + fp.val = (uint32_t ) intermediate_val; +return fp; +} +static inline const char *yesno (bool v) {return v ?"yes" :"no" ;
drivers/gpu/drm/i915/intel_pm.c +40 -29
126 lines not on this page · raw message
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 6/8 drm/i915: Add intel_atomic_get_existing_crtc_state function
In reply to Mahesh KumarThis patch Adds a function to extract intel_crtc_state from the atomic_state, if not available it returns NULL.
--- drivers/gpu/drm/i915/intel_drv.h | 14 ++++++++++++++ 1 file changed, 14 insertions(+)
drivers/gpu/drm/i915/intel_drv.h +14 -0
@@ -1804,6 +1804,20 @@ intel_atomic_get_crtc_state(struct drm_atomic_state *state,return to_intel_crtc_state (crtc_state); } +static inline struct intel_crtc_state * +intel_atomic_get_existing_crtc_state (struct drm_atomic_state *state, +struct intel_crtc *crtc) +{ +struct drm_crtc_state *crtc_state; + + crtc_state =drm_atomic_get_existing_crtc_state (state, &crtc->base); + +if (crtc_state) +return to_intel_crtc_state (crtc_state); +else +return NULL ; +} +static inline struct intel_plane_state *intel_atomic_get_existing_plane_state (struct drm_atomic_state *state,struct intel_plane *plane)
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 7/8 drm/i915: Decode system memory bandwidth
In reply to Mahesh KumarThis patch adds support to decode system memory bandwidth which will be used for arbitrated display memory percentage calculation in GEN9 based system.
Changes from v1:
- Address comments from Paulo
- implement decode function for SKL/KBL also
Changes from v2:
- Rewrite the code as per HW team inputs
- Addresses review comments
Changes from v3:
- Fix compilation warning
--- drivers/gpu/drm/i915/i915_drv.c | 173 ++++++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 12 +++ drivers/gpu/drm/i915/i915_reg.h | 37 +++++++++ 3 files changed, 222 insertions(+)
drivers/gpu/drm/i915/i915_drv.c +173 -0
@@ -979,6 +979,173 @@ static void intel_sanitize_options(struct drm_i915_private *dev_priv)DRM_DEBUG_DRIVER ("use GPU sempahores? %s\n" ,yesno (i915.semaphores)); } +static inline enum rank skl_memdev_get_channel_rank (uint32_t val) +{ +uint8_t l_rank, s_rank; +uint8_t l_size, s_size; +enum rank ch_rank =DRAM_RANK_SINGLE ; + + l_size = (val >>SKL_DRAM_SIZE_L_SHIFT ) &SKL_DRAM_SIZE_MASK ; + s_size = (val >>SKL_DRAM_SIZE_S_SHIFT ) &SKL_DRAM_SIZE_MASK ; + l_rank = (val >>SKL_DRAM_RANK_L_SHIFT ) &SKL_DRAM_RANK_MASK ; + s_rank = (val >>SKL_DRAM_RANK_S_SHIFT ) &SKL_DRAM_RANK_MASK ; + +/* +* If any of the slot has dual rank memory consider +* dual rank memory channel +*/ +if (l_rank ==SKL_DRAM_RANK_DUAL || s_rank ==SKL_DRAM_RANK_DUAL ) + ch_rank =DRAM_RANK_DUAL ; + +/* +* If both the slot has single rank memory then configuration +* is dual rank memory +*/ +if ((l_size && l_rank ==SKL_DRAM_RANK_SINGLE ) && + (s_size && s_rank ==SKL_DRAM_RANK_SINGLE )) + ch_rank =DRAM_RANK_DUAL ; +return ch_rank; +} + +static int +skl_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t mem_freq_khz; +uint32_t val; +enum rank ch0_rank =DRAM_RANK_INVALID , ch1_rank =DRAM_RANK_INVALID ; + + val =I915_READ (SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU ); + mem_freq_khz = (val &SKL_REQ_DATA_MASK ) * +SKL_MEMORY_FREQ_MULTIPLIER_KHZ ; + + val =I915_READ (SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch0_rank =skl_memdev_get_channel_rank (val); + } + + val =I915_READ (SKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch1_rank =skl_memdev_get_channel_rank (val); + } + +if (memdev_info->num_channels ==0 ) { +DRM_ERROR ("Number of mem channels are zero\n" ); +return -EINVAL ; + } + + memdev_info->bandwidth_kbps = (memdev_info->num_channels * + mem_freq_khz *8 ); + +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true; + +/* +* If any of channel is single rank channel, +* consider single rank memory +*/ +if (ch0_rank ==DRAM_RANK_SINGLE || ch1_rank ==DRAM_RANK_SINGLE ) + memdev_info->rank =DRAM_RANK_SINGLE ; +else + memdev_info->rank =max (ch0_rank, ch1_rank); + +return 0 ; +} + +static int +bxt_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t dram_channels; +uint32_t mem_freq_khz, val; +uint8_t num_active_channels; +int i; + + val =I915_READ (BXT_P_CR_MC_BIOS_REQ_0_0_0 ); + mem_freq_khz = ((val &BXT_REQ_DATA_MASK ) * +BXT_MEMORY_FREQ_MULTIPLIER_KHZ ); + + dram_channels = (val >>BXT_DRAM_CHANNEL_ACTIVE_SHIFT ) & +BXT_DRAM_CHANNEL_ACTIVE_MASK ; + num_active_channels =hweight32 (dram_channels); + + memdev_info->bandwidth_kbps = (mem_freq_khz * num_active_channels *4 ); + +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true; + +/* +* Now read each DUNIT8/9/10/11 to check the rank of each dimms. +*/ +for (i =0 ; i <BXT_D_CR_DRP0_DUNIT_MAX ; i++) { + val =I915_READ (BXT_D_CR_DRP0_DUNIT (i)); +if (val !=0xFFFFFFFF ) { +uint8_t rank; +enum rank ch_rank; + + memdev_info->num_channels++; + rank = val &BXT_DRAM_RANK_MASK ; +if (rank ==BXT_DRAM_RANK_SINGLE ) + ch_rank =DRAM_RANK_SINGLE ; +else if (rank ==BXT_DRAM_RANK_DUAL ) + ch_rank =DRAM_RANK_DUAL ; +else + ch_rank =DRAM_RANK_INVALID ; + +/* +* If any of channel is having single rank memory +* consider memory as single rank +*/ +if (memdev_info->rank ==DRAM_RANK_INVALID ) + memdev_info->rank = ch_rank; +else if (ch_rank ==DRAM_RANK_SINGLE ) + memdev_info->rank =DRAM_RANK_SINGLE ; + } + } +return 0 ; +} + +static void +intel_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +int ret; + + memdev_info->valid = false; + memdev_info->rank =DRAM_RANK_INVALID ; + memdev_info->num_channels =0 ; + +if (!IS_GEN9 (dev_priv)) +return ; + +if (IS_BROXTON (dev_priv)) + ret =bxt_get_memdev_info (dev_priv); +else + ret =skl_get_memdev_info (dev_priv); +if (ret) +return ; + +DRM_DEBUG_DRIVER ("DRAM bandwidth: %u KBps total-channels: %u\n" , + memdev_info->bandwidth_kbps, + memdev_info->num_channels); +if (memdev_info->rank ==DRAM_RANK_INVALID ) +DRM_INFO ("Counld not get memory rank info\n" ); +else { +DRM_DEBUG_DRIVER ("DRAM rank: %s rank\n" , + (memdev_info->rank ==DRAM_RANK_DUAL ) ? +"dual" :"single" ); + } +} + +/** * i915_driver_init_hw - setup state requiring device access * @dev_priv: device private
@@ -1081,6 +1248,12 @@ static int i915_driver_init_hw(struct drm_i915_private *dev_priv)DRM_DEBUG_DRIVER ("can't enable MSI" ); } +/* +* Fill the memdev structure to get the system raw bandwidth +* This will be used by WM algorithm, to implement GEN9 based WA +*/ +intel_get_memdev_info (dev_priv); +return 0 ; out_ggtt:
drivers/gpu/drm/i915/i915_drv.h +12 -0
@@ -2297,6 +2297,18 @@ struct drm_i915_private {bool distrust_bios_wm; } wm; +struct memdev_info { +bool valid; +uint32_t bandwidth_kbps; +uint8_t num_channels; +enum rank { +DRAM_RANK_INVALID =0 , +DRAM_RANK_SINGLE , +DRAM_RANK_DUAL + } rank; + } memdev_info; + +struct i915_runtime_pm pm;struct {
drivers/gpu/drm/i915/i915_reg.h +37 -0
@@ -8015,6 +8015,43 @@ enum {#define DC_STATE_DEBUG_MASK_CORES (1<<0)#define DC_STATE_DEBUG_MASK_MEMORY_UP (1<<1) +#define BXT_P_CR_MC_BIOS_REQ_0_0_0 _MMIO(MCHBAR_MIRROR_BASE_SNB + 0x7114) +#define BXT_REQ_DATA_MASK 0x3F +#define BXT_DRAM_ACTIVE_CHANNEL_SHIFT 12 +#define BXT_DRAM_ACTIVE_CHANNEL_MASK 0xF +/* +* BIOS programs this field of REQ_DATA [5:0] in integer +* multiple of 133333 KHz (133.33MHz) +*/ +#define BXT_MEMORY_FREQ_MULTIPLIER_KHZ 133333 +#define BXT_D_CR_DRP0_DUNIT8 0x1000 +#define BXT_D_CR_DRP0_DUNIT9 0x1200 +#define BXT_D_CR_DRP0_DUNIT_MAX 4 +#define _MMIO_MCHBAR_DUNIT (x, a, b) _MMIO(MCHBAR_MIRROR_BASE_SNB + (a) + (x)*((b)-(a))) +#define BXT_D_CR_DRP0_DUNIT (x) _MMIO_MCHBAR_DUNIT(x, BXT_D_CR_DRP0_DUNIT8, BXT_D_CR_DRP0_DUNIT9) +#define BXT_DRAM_CHANNEL_ACTIVE_SHIFT 12 +#define BXT_DRAM_CHANNEL_ACTIVE_MASK 0xF +#define BXT_DRAM_RANK_MASK 0x3 +#define BXT_DRAM_RANK_SINGLE 0x1 +#define BXT_DRAM_RANK_DUAL 0x3 + +/* +* SKL memory frequeny multiplier is 266667 KHz (266.67 MHz) +*/ +#define SKL_MEMORY_FREQ_MULTIPLIER_KHZ 266667 +#define SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU _MMIO(MCHBAR_MIRROR_BASE_SNB + 0x5E04) +#define SKL_REQ_DATA_MASK (0xF << 0) +#define SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN _MMIO(MCHBAR_MIRROR_BASE_SNB + 0x500C) +#define SKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN _MMIO(MCHBAR_MIRROR_BASE_SNB + 0x5010) +#define SKL_DRAM_SIZE_MASK 0x1F +#define SKL_DRAM_SIZE_L_SHIFT 0 +#define SKL_DRAM_SIZE_S_SHIFT 16 +#define SKL_DRAM_RANK_MASK 0x1 +#define SKL_DRAM_RANK_L_SHIFT 10 +#define SKL_DRAM_RANK_S_SHIFT 26 +#define SKL_DRAM_RANK_SINGLE 0x0 +#define SKL_DRAM_RANK_DUAL 0x1 +/* Please see hsw_read_dcomp() and hsw_write_dcomp() before using this register, * since on HSW we can't write to it using I915_WRITE. */ #define D_COMP_HSW _MMIO(MCHBAR_MIRROR_BASE_SNB + 0x5F0C)
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v7 7/8] drm/i915: Decode system memory bandwidth
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 13 lines
This patch adds support to decode system memory bandwidth which will be used for arbitrated display memory percentage calculation in GEN9 based system.
Changes from v1:
- Address comments from Paulo
- implement decode function for SKL/KBL also
Changes from v2:
- Rewrite the code as per HW team inputs
- Addresses review comments
Changes from v3:
- Fix compilation warning
Quoted 1 line
Signed-off-by Mahesh Kumar <[email protected]>
As a general comment, indentation is weird on all multi-line statements. Also, most comments are missing periods and are not 80- column aligned. And a lot of the comments just say what the code already does instead of explaining why the code does what it does, so perhaps we could just remove them. Comments should explain why the code is written the way it is, not translate from C to English.
Quoted 17 lines
---
drivers/gpu/drm/i915/i915_drv.c | 173
++++++++++++++++++++++++++++++++++++++++
drivers/gpu/drm/i915/i915_drv.h | 12 +++
drivers/gpu/drm/i915/i915_reg.h | 37 +++++++++3 files changed, 222 insertions(+)drivers/gpu/drm/i915/i915_drv.c +0
-0@@ -979,6 +979,173 @@ static void intel_sanitize_options(struct drm_i915_private *dev_priv) DRM_DEBUG_DRIVER("use GPU sempahores? %s\n", yesno(i915.semaphores)); }
Quoted 2 lines
+static inline enum rankskl_memdev_get_channel_rank(uint32_t val)
Why inline? This is already static, leave it up to the compiler to decide if it's better to inline or not.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 986
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+{ +uint8_t l_rank, s_rank; +uint8_t l_size, s_size; +enum rank ch_rank =DRAM_RANK_SINGLE ;
Optional suggestion: get rid of this variable, just return the appropriate values once we discover them.
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +2
-0+ + l_size = (val >> SKL_DRAM_SIZE_L_SHIFT) &
SKL_DRAM_SIZE_MASK;drivers/gpu/drm/i915/i915_drv.c +1
-0+ s_size = (val >> SKL_DRAM_SIZE_S_SHIFT) &
SKL_DRAM_SIZE_MASK;drivers/gpu/drm/i915/i915_drv.c +1
-0+ l_rank = (val >> SKL_DRAM_RANK_L_SHIFT) &
SKL_DRAM_RANK_MASK;drivers/gpu/drm/i915/i915_drv.c +1
-0+ s_rank = (val >> SKL_DRAM_RANK_S_SHIFT) &
SKL_DRAM_RANK_MASK;
Validate our assumptions:
WARN_ON(l_size == 0 && s_size == 0);
Or we could even do the appropriate check and return DRAM_RANK_INVALID, but then we'd have to restructure how our caller calls us, maybe moving some code to this function.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 996
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +5
-0+ +/* +* If any of the slot has dual rank memory consider +* dual rank memory channel +*/
The comment says just what the code already says. What would be interesting to see in the comment is an explanation of why we do it the way we do.
drivers/gpu/drm/i915/i915_drv.c
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +1
-0+ if (l_rank == SKL_DRAM_RANK_DUAL || s_rank ==
SKL_DRAM_RANK_DUAL)drivers/gpu/drm/i915/i915_drv.c +4
-0+ ch_rank =DRAM_RANK_DUAL ; + +/* +* If both the slot has single rank memory then configuration
drivers/gpu/drm/i915/i915_drv.c +2
-0+* is dual rank memory +*/
The comment says just what the code already says. What would be interesting to see in the comment is an explanation of why we do it the way we do.
Quoted 14 lines
drivers/gpu/drm/i915/i915_drv.c +13
-0+if ((l_size && l_rank ==SKL_DRAM_RANK_SINGLE ) && + (s_size && s_rank ==SKL_DRAM_RANK_SINGLE )) + ch_rank =DRAM_RANK_DUAL ; +return ch_rank; +} + +static int +skl_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t mem_freq_khz; +uint32_t val; + enum rank ch0_rank = DRAM_RANK_INVALID, ch1_rank =
DRAM_RANK_INVALID;
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1020
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+ + val =I915_READ (SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU ); + mem_freq_khz = (val &SKL_REQ_DATA_MASK ) * +SKL_MEMORY_FREQ_MULTIPLIER_KHZ ;
Optional suggestion: perhaps extend the memory_freq_multiplier to HZ and then later cut the last 3 digits so the rounding errors get restricted to the units we'll cut? Also, if we do the math using the "4
* num / 3" we can try to reduce the rounding errors even more.
I get annoyed that it says my system bandwidth is 25600032, while the correct number is exactly 25600000.
This applies to the BXT code too.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1026
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+ + val =I915_READ (SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch0_rank =skl_memdev_get_channel_rank (val); + }
Optional suggestion:
Here's how I would have implemented this:
ch0_rank = skl_memdev_get_channel_rank(0);
if (ch0_rank != DRAM_RANK_INVALID)
memdev_info->num_channels++;
This way we'd move all the logic around this register to skl_memdev_get_channel_rank(), and we'd also be able to get rid of the initializations here.
But this is just a suggestion in case you agree with me. Feel free to leave the code the way it is.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1035
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+ + val =I915_READ (SKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch1_rank =skl_memdev_get_channel_rank (val); + } + +if (memdev_info->num_channels ==0 ) { +DRM_ERROR ("Number of mem channels are zero\n" );
I'm not a native English speaker, but I would suppose that s/are/is/ would make the sentence correct (the number is zero). While at it, why not s/mem/memory/ too to make the code sound a little more formal?
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1046
Quoted 12 lines
drivers/gpu/drm/i915/i915_drv.c +5
-0+return -EINVAL ; + } + + memdev_info->bandwidth_kbps = (memdev_info->num_channels * + mem_freq_khz* 8);
drivers/gpu/drm/i915/i915_drv.c +6
-0+ +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true;
My previous comment about ownership of this variable still applies.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1049
Quoted 3 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+ +/* +* If any of channel is single rank channel,
Again, not a native English speaker, but "any of channel" sounds incorrect to me. But anyway, the comment only says what the code already says. Why do we do it like this?
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1055
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+* consider single rank memory +*/ + if (ch0_rank == DRAM_RANK_SINGLE || ch1_rank ==
DRAM_RANK_SINGLE)drivers/gpu/drm/i915/i915_drv.c +3
-0+ memdev_info->rank =DRAM_RANK_SINGLE ; + else + memdev_info->rank = max(ch0_rank, ch1_rank);
I can't think of any situation where this wouldn't end with DRAM_RANK_DUAL, so we might just assign it to the variable.
Quoted 34 lines
drivers/gpu/drm/i915/i915_drv.c +19
-0+ +return 0 ; +} + +static int +bxt_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t dram_channels; +uint32_t mem_freq_khz, val; +uint8_t num_active_channels; +int i; + + val =I915_READ (BXT_P_CR_MC_BIOS_REQ_0_0_0 ); + mem_freq_khz = ((val &BXT_REQ_DATA_MASK ) * +BXT_MEMORY_FREQ_MULTIPLIER_KHZ ); + + dram_channels = (val >>BXT_DRAM_CHANNEL_ACTIVE_SHIFT ) & +BXT_DRAM_CHANNEL_ACTIVE_MASK ;
drivers/gpu/drm/i915/i915_drv.c +3
-0+ num_active_channels =hweight32 (dram_channels); + + memdev_info->bandwidth_kbps = (mem_freq_khz *
num_active_channels* 4);drivers/gpu/drm/i915/i915_drv.c +9
-0+ +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true; + +/* +* Now read each DUNIT8/9/10/11 to check the rank of each dimms.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1095
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+*/ +for (i =0 ; i <BXT_D_CR_DRP0_DUNIT_MAX ; i++) { + val =I915_READ (BXT_D_CR_DRP0_DUNIT (i)); +if (val !=0xFFFFFFFF ) { +uint8_t rank; +enum rank ch_rank; + + memdev_info->num_channels++; + rank = val &BXT_DRAM_RANK_MASK ;
Shouldn't the code here be: "if one of these two bits are enabled then we're RANK_SINGLE, if both bits are enabled then we're RANK_DUAL, otherwise invalid"? If not, why?
In other words: is rank=0x2 really invalid? Isn't it just single rank?
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+if (rank ==BXT_DRAM_RANK_SINGLE ) + ch_rank =DRAM_RANK_SINGLE ; +else if (rank ==BXT_DRAM_RANK_DUAL ) + ch_rank =DRAM_RANK_DUAL ; +else + ch_rank =DRAM_RANK_INVALID ; + +/* +* If any of channel is having single rank memory
Again, "any of channel".
drivers/gpu/drm/i915/i915_drv.c
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+* consider memory as single rank +*/ +if (memdev_info->rank ==DRAM_RANK_INVALID ) + memdev_info->rank = ch_rank; +else if (ch_rank ==DRAM_RANK_SINGLE ) + memdev_info->rank =
DRAM_RANK_SINGLE;drivers/gpu/drm/i915/i915_drv.c +2
-0+ } + }
SKL returns -EINVAL if we end up with DRAM_RANK_INVALID in memdev_info-
Quoted 1 line
rank. I think we could do that in BXT too: either we have all the
information to not apply the workaround, or we give up.
Quoted 25 lines
drivers/gpu/drm/i915/i915_drv.c +24
-0+return 0 ; +} + +static void +intel_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +int ret; + + memdev_info->valid = false; + memdev_info->rank =DRAM_RANK_INVALID ; + memdev_info->num_channels =0 ; + +if (!IS_GEN9 (dev_priv)) +return ; + +if (IS_BROXTON (dev_priv)) + ret =bxt_get_memdev_info (dev_priv); +else + ret =skl_get_memdev_info (dev_priv); +if (ret) +return ; + + DRM_DEBUG_DRIVER("DRAM bandwidth: %u KBps total-channels:%u\n",
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1140
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+ memdev_info->bandwidth_kbps, + memdev_info->num_channels); +if (memdev_info->rank ==DRAM_RANK_INVALID ) +DRM_INFO ("Counld not get memory rank info\n" );
Spelling mistake here. And since you use "Couldn't" instead of "Could not" in all the other messages, I'd keep the same pattern.
But if you accept my suggestion above, we'll never print this message anyway.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1148
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+else { + DRM_DEBUG_DRIVER("DRAM rank: %s rank\n", + (memdev_info->rank ==
DRAM_RANK_DUAL) ?drivers/gpu/drm/i915/i915_drv.c +5
-0+ "dual" : "single"); + } +} + +
Two white spaces here.
Quoted 27 lines
/**
*i915_driver_init_hw- setup state requiring device access
* @dev_priv: device privatedrivers/gpu/drm/i915/i915_drv.c +0
-0@@ -1081,6 +1248,12 @@ static int i915_driver_init_hw(struct drm_i915_private *dev_priv) DRM_DEBUG_DRIVER("can't enable MSI"); }drivers/gpu/drm/i915/i915_drv.c +3
-0+/* +* Fill the memdev structure to get the system raw bandwidth +* This will be used by WM algorithm, to implement GEN9 based WA
drivers/gpu/drm/i915/i915_drv.c +3
-0+*/ +intel_get_memdev_info (dev_priv); +return 0;
out_ggtt:drivers/gpu/drm/i915/i915_drv.h +0
-0@@ -2297,6 +2297,18 @@ struct drm_i915_private { bool distrust_bios_wm; } wm;
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_private { · line 2308
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.h +9
-0+struct memdev_info { +bool valid; +uint32_t bandwidth_kbps; +uint8_t num_channels; +enum rank { +DRAM_RANK_INVALID =0 , +DRAM_RANK_SINGLE , +DRAM_RANK_DUAL + } rank;
This was previously an anonymous enum (enum { ... } rank;), but now it's "enum rank", which is too generic for a .h file. Either use a more specific name (enum memdev_rank?) or try to make it anonymous again.
Quoted 17 lines
drivers/gpu/drm/i915/i915_drv.h +3
-0+ } memdev_info; + +struct
i915_runtime_pmpm;
struct {drivers/gpu/drm/i915/i915_reg.h +0
-0@@ -8015,6 +8015,43 @@ enum { #define DC_STATE_DEBUG_MASK_CORES (1<<0) #define DC_STATE_DEBUG_MASK_MEMORY_UP (1<<1)
+#defineBXT_P_CR_MC_BIOS_REQ_0_0_0_MMIO(MCHBAR_MIRROR_BASE_S
NB +0x7114)
drivers/gpu/drm/i915/i915_reg.h enum { · line 8021
Quoted 3 lines
drivers/gpu/drm/i915/i915_reg.h +3
-0+#define BXT_REQ_DATA_MASK 0x3F +#define BXT_DRAM_ACTIVE_CHANNEL_SHIFT 12 +#define BXT_DRAM_ACTIVE_CHANNEL_MASK 0xF
We're not using these two definitions anywhere.
Quoted 4 lines
+/*
+ * BIOS programs this field ofREQ_DATA[5:0] in integer
+ * multiple of 133333 KHz (133.33MHz)
+ */
Now that we're using decimal on the definition, I don't think this comment is useful.
Quoted 1 line
+#define
BXT_MEMORY_FREQ_MULTIPLIER_KHZ133333
We don't use tabs like that in our definitions.
Quoted 7 lines
drivers/gpu/drm/i915/i915_reg.h +4
-0+#define BXT_D_CR_DRP0_DUNIT8 0x1000 +#define BXT_D_CR_DRP0_DUNIT9 0x1200 +#define BXT_D_CR_DRP0_DUNIT_MAX 4 +#define _MMIO_MCHBAR_DUNIT (x, a, b) _MMIO(MCHBAR_MIRROR_BASE_SNB +(a) + (x)*((b)-(a)))
+#defineBXT_D_CR_DRP0_DUNIT(x)_MMIO_MCHBAR_DUNIT(x,BXT_D_CR_DRP0_DUNIT8,BXT_D_CR_DRP0_DUNIT9)
But then BXT_D_CR_DRP0_DUNIT(1) means DUNIT9 instead of DUNIT1... That's very unintuitive.
drivers/gpu/drm/i915/i915_reg.h enum { · line 8033
Quoted 2 lines
drivers/gpu/drm/i915/i915_reg.h +2
-0+#define BXT_DRAM_CHANNEL_ACTIVE_SHIFT 12 +#define BXT_DRAM_CHANNEL_ACTIVE_MASK 0xF
These two definitions don't belong to the register immediately above.
drivers/gpu/drm/i915/i915_reg.h enum { · line 8039
Quoted 6 lines
drivers/gpu/drm/i915/i915_reg.h +6
-0+#define BXT_DRAM_RANK_MASK 0x3 +#define BXT_DRAM_RANK_SINGLE 0x1 +#define BXT_DRAM_RANK_DUAL 0x3 + +/* +* SKL memory frequeny multiplier is 266667 KHz (266.67 MHz)
s/frequeny/frequency/
Also, this is not how we do one-line comments.
And now that the number is in decimal we probably don't even need the comment here since the definition is a little more obvious.
Quoted 2 lines
+ */
+#defineSKL_MEMORY_FREQ_MULTIPLIER_KHZ266667
We don't use tabs like that in our definitions.
Quoted 3 lines
+#define
SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU_MMIO(MCHBAR_MIRROR_BASE_SNB+0x5E04)
+#defineSKL_REQ_DATA_MASK(0xF<< 0)
Blank lines separating different registers would be good.
Quoted 5 lines
+#define
SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN_MMIO(MCHBAR_MIRROR_BASE_SNB+0x500C)
+#defineSKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN_MMIO(MCHBAR_MIRROR_BASE_SNB+0x5010)
+#defineSKL_DRAM_SIZE_MASK0x1F
0x3F
Quoted 8 lines
drivers/gpu/drm/i915/i915_reg.h +8
-0+#define SKL_DRAM_SIZE_L_SHIFT 0 +#define SKL_DRAM_SIZE_S_SHIFT 16 +#define SKL_DRAM_RANK_MASK 0x1 +#define SKL_DRAM_RANK_L_SHIFT 10 +#define SKL_DRAM_RANK_S_SHIFT 26 +#define SKL_DRAM_RANK_SINGLE 0x0 +#define SKL_DRAM_RANK_DUAL 0x1 +
Quoted 5 lines
/* Please see
hsw_read_dcomp()andhsw_write_dcomp()before using this register,
* since on HSW we can't write to it usingI915_WRITE. */
#defineD_COMP_HSW_MMIO(MCHBAR_MIRROR_BASE_S
NB +0x5F0C)
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v7 7/8] drm/i915: Decode system memory bandwidth
Hi,
Quoted 41 lines · Paulo Zanoni
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 13 lines
This patch adds support to decode system memory bandwidth which will be used for arbitrated display memory percentage calculation in GEN9 based system.
Changes from v1:
- Address comments from Paulo
- implement decode function for SKL/KBL also
Changes from v2:
- Rewrite the code as per HW team inputs
- Addresses review comments
Changes from v3:
- Fix compilation warningQuoted 1 line
Signed-off-by Mahesh Kumar <[email protected]>As a general comment, indentation is weird on all multi-line statements. Also, most comments are missing periods and are not 80- column aligned. And a lot of the comments just say what the code already does instead of explaining why the code does what it does, so perhaps we could just remove them. Comments should explain why the code is written the way it is, not translate from C to English.
Quoted 18 lines
--- drivers/gpu/drm/i915/i915_drv.c | 173 ++++++++++++++++++++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 12 +++ drivers/gpu/drm/i915/i915_reg.h | 37 +++++++++ 3 files changed, 222 insertions(+)drivers/gpu/drm/i915/i915_drv.c +0
-0@@ -979,6 +979,173 @@ static void intel_sanitize_options(struct drm_i915_private *dev_priv) DRM_DEBUG_DRIVER("use GPU sempahores? %s\n", yesno(i915.semaphores)); }
Quoted 3 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 982
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.c +1
-0+static inline enum rank skl_memdev_get_channel_rank (uint32_t val)Why inline? This is already static, leave it up to the compiler to decide if it's better to inline or not.
removed inline in new series
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 986
Quoted 5 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+{ +uint8_t l_rank, s_rank; +uint8_t l_size, s_size; +enum rank ch_rank =DRAM_RANK_SINGLE ;Optional suggestion: get rid of this variable, just return the appropriate values once we discover them.
done
Quoted 12 lines
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+ + l_size = (val >>SKL_DRAM_SIZE_L_SHIFT ) &SKL_DRAM_SIZE_MASK ; + s_size = (val >> SKL_DRAM_SIZE_S_SHIFT) &
SKL_DRAM_SIZE_MASK;drivers/gpu/drm/i915/i915_drv.c +1
-0+ l_rank = (val >> SKL_DRAM_RANK_L_SHIFT) &
SKL_DRAM_RANK_MASK;drivers/gpu/drm/i915/i915_drv.c +1
-0+ s_rank = (val >> SKL_DRAM_RANK_S_SHIFT) &
SKL_DRAM_RANK_MASK;Validate our assumptions:
Quoted 5 lines
WARN_ON(l_size == 0 && s_size == 0);Or we could even do the appropriate check and return
DRAM_RANK_INVALID, but then we'd have to restructure how our caller calls us, maybe moving some code to this function.
made the adjustment
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 996
Quoted 6 lines
drivers/gpu/drm/i915/i915_drv.c +5
-0+ +/* +* If any of the slot has dual rank memory consider +* dual rank memory channel +*/ The comment says just what the code already says. What would be interesting to see in the comment is an explanation of why we do it the way we do.
updated
Quoted 39 lines
drivers/gpu/drm/i915/i915_drv.c
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c +1
-0+ if (l_rank == SKL_DRAM_RANK_DUAL || s_rank ==
SKL_DRAM_RANK_DUAL)drivers/gpu/drm/i915/i915_drv.c +4
-0+ ch_rank =DRAM_RANK_DUAL ; + +/* +* If both the slot has single rank memory then configuration
drivers/gpu/drm/i915/i915_drv.c +2
-0+* is dual rank memory +*/ The comment says just what the code already says. What would be interesting to see in the comment is an explanation of why we do it the way we do.
Quoted 14 lines
drivers/gpu/drm/i915/i915_drv.c +13
-0+if ((l_size && l_rank ==SKL_DRAM_RANK_SINGLE ) && + (s_size && s_rank ==SKL_DRAM_RANK_SINGLE )) + ch_rank =DRAM_RANK_DUAL ; +return ch_rank; +} + +static int +skl_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t mem_freq_khz; +uint32_t val; + enum rank ch0_rank = DRAM_RANK_INVALID, ch1_rank =
DRAM_RANK_INVALID;drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1020
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+ + val =I915_READ (SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU ); + mem_freq_khz = (val &SKL_REQ_DATA_MASK ) * +SKL_MEMORY_FREQ_MULTIPLIER_KHZ ;Optional suggestion: perhaps extend the
memory_freq_multiplierto HZ and then later cut the last 3 digits so the rounding errors get restricted to the units we'll cut? Also, if we do the math using the "4
* num / 3" we can try to reduce the rounding errors even more.
Quoted 4 lines
I get annoyed that it says my system bandwidth is 25600032, while the correct number is exactly 25600000.
This applies to the BXT code too.
Modified to work on Hz & then convert to KHz
Quoted 13 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1026
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+ + val =I915_READ (SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch0_rank =skl_memdev_get_channel_rank (val); + }Optional suggestion:
Here's how I would have implemented this:
ch0_rank=skl_memdev_get_channel_rank(0);
if (ch0_rank!=DRAM_RANK_INVALID)
Quoted 8 lines
memdev_info->num_channels++;This way we'd move all the logic around this register to
skl_memdev_get_channel_rank(), and we'd also be able to get rid of the initializations here.But this is just a suggestion in case you agree with me. Feel free to leave the code the way it is.
suggestion taken :)
Quoted 15 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1035
Quoted 10 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+ + val =I915_READ (SKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN ); +if (val !=0x0 ) { + memdev_info->num_channels++; + ch1_rank =skl_memdev_get_channel_rank (val); + } + +if (memdev_info->num_channels ==0 ) { +DRM_ERROR ("Number of mem channels are zero\n" );I'm not a native English speaker, but I would suppose that s/are/is/ would make the sentence correct (the number is zero). While at it, why not s/mem/memory/ too to make the code sound a little more formal?
Quoted 14 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1046
Quoted 12 lines
drivers/gpu/drm/i915/i915_drv.c +5
-0+return -EINVAL ; + } + + memdev_info->bandwidth_kbps = (memdev_info->num_channels * + mem_freq_khz* 8);
drivers/gpu/drm/i915/i915_drv.c +6
-0+ +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true;My previous comment about ownership of this variable still applies.
Now assigning memdev_info-> valid to true, only if bandwidth & rank both the information are valid
Quoted 8 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1049
Quoted 3 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+ +/* +* If any of channel is single rank channel, Again, not a native English speaker, but "any of channel" sounds incorrect to me. But anyway, the comment only says what the code already says. Why do we do it like this?
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1055
Quoted 7 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+* consider single rank memory +*/ + if (ch0_rank == DRAM_RANK_SINGLE || ch1_rank ==
DRAM_RANK_SINGLE)drivers/gpu/drm/i915/i915_drv.c +3
-0+ memdev_info->rank =DRAM_RANK_SINGLE ; + else + memdev_info->rank = max(ch0_rank, ch1_rank);I can't think of any situation where this wouldn't end with
DRAM_RANK_DUAL, so we might just assign it to the variable.
If there is only one dual-rank DRAM installed out of 4-dimms, then either ch0_rank or ch1_rank will be I915_DRAM_DUAL_RANK & other will be INVALID, in that case we'll have rank as DRAM_RANK_DUAL.
Quoted 48 lines
Quoted 35 lines
drivers/gpu/drm/i915/i915_drv.c +19
-0+ +return 0 ; +} + +static int +bxt_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +uint32_t dram_channels; +uint32_t mem_freq_khz, val; +uint8_t num_active_channels; +int i; + + val =I915_READ (BXT_P_CR_MC_BIOS_REQ_0_0_0 ); + mem_freq_khz = ((val &BXT_REQ_DATA_MASK ) * +BXT_MEMORY_FREQ_MULTIPLIER_KHZ ); + + dram_channels = (val >>BXT_DRAM_CHANNEL_ACTIVE_SHIFT ) & +BXT_DRAM_CHANNEL_ACTIVE_MASK ;
drivers/gpu/drm/i915/i915_drv.c +3
-0+ num_active_channels =hweight32 (dram_channels); + + memdev_info->bandwidth_kbps = (mem_freq_khz *
num_active_channels* 4);drivers/gpu/drm/i915/i915_drv.c +9
-0+ +if (memdev_info->bandwidth_kbps ==0 ) { +DRM_ERROR ("Couldn't get system memory bandwidth\n" ); +return -EINVAL ; + } + memdev_info->valid = true; + +/* +* Now read each DUNIT8/9/10/11 to check the rank of each dimms.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1095
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+*/ +for (i =0 ; i <BXT_D_CR_DRP0_DUNIT_MAX ; i++) { + val =I915_READ (BXT_D_CR_DRP0_DUNIT (i)); +if (val !=0xFFFFFFFF ) { +uint8_t rank; +enum rank ch_rank; + + memdev_info->num_channels++; + rank = val &BXT_DRAM_RANK_MASK ;Shouldn't the code here be: "if one of these two bits are enabled then we're
RANK_SINGLE, if both bits are enabled then we'reRANK_DUAL, otherwise invalid"? If not, why?
Quoted 1 line
In other words: is rank=0x2 really invalid? Isn't it just single rank?
rank=0x2 is an invalid entry. "Enable Rank 0: Must be set to 1 to enable use of this rank. Note: Setting this bit to 0 is not a functional mode" "Enable Rank 0" bit can't be a 0 anyhow.
Quoted 24 lines
Quoted 11 lines
drivers/gpu/drm/i915/i915_drv.c +9
-0+if (rank ==BXT_DRAM_RANK_SINGLE ) + ch_rank =DRAM_RANK_SINGLE ; +else if (rank ==BXT_DRAM_RANK_DUAL ) + ch_rank =DRAM_RANK_DUAL ; +else + ch_rank =DRAM_RANK_INVALID ; + +/* +* If any of channel is having single rank Quoted 1 line
memory
Again, "any of channel".
drivers/gpu/drm/i915/i915_drv.c
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +6
-0+* consider memory as single rank +*/ +if (memdev_info->rank ==DRAM_RANK_INVALID ) + memdev_info->rank = ch_rank; +else if (ch_rank ==DRAM_RANK_SINGLE ) + memdev_info->rank =
DRAM_RANK_SINGLE;drivers/gpu/drm/i915/i915_drv.c +2
-0+ } + }SKL returns
-EINVALif we end up withDRAM_RANK_INVALIDin memdev_info-
Quoted 2 lines
Quoted 1 line
rank. I think we could do that in BXT too: either we have all the
information to not apply the workaround, or we give up.
done, returning invalid if rank information is not available.
Quoted 77 lines
Quoted 27 lines
drivers/gpu/drm/i915/i915_drv.c +24
-0+return 0 ; +} + +static void +intel_get_memdev_info (struct drm_i915_private *dev_priv) +{ +struct memdev_info *memdev_info = &dev_priv->memdev_info; +int ret; + + memdev_info->valid = false; + memdev_info->rank =DRAM_RANK_INVALID ; + memdev_info->num_channels =0 ; + +if (!IS_GEN9 (dev_priv)) +return ; + +if (IS_BROXTON (dev_priv)) + ret =bxt_get_memdev_info (dev_priv); +else + ret =skl_get_memdev_info (dev_priv); +if (ret) +return ; + + DRM_DEBUG_DRIVER("DRAM bandwidth: %u KBps total-channels:%u\n",
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1140
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.c +4
-0+ memdev_info->bandwidth_kbps, + memdev_info->num_channels); +if (memdev_info->rank ==DRAM_RANK_INVALID ) +DRM_INFO ("Counld not get memory rank info\n" );Spelling mistake here. And since you use "Couldn't" instead of "Could not" in all the other messages, I'd keep the same pattern.
But if you accept my suggestion above, we'll never print this message anyway.
drivers/gpu/drm/i915/i915_drv.c static void intel_sanitize_options(struct drm_i915_private *dev_priv) · line 1148
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c +3
-0+else { + DRM_DEBUG_DRIVER("DRAM rank: %s rank\n", + (memdev_info->rank ==
DRAM_RANK_DUAL) ?drivers/gpu/drm/i915/i915_drv.c +5
-0+ "dual" : "single"); + } +} + +Two white spaces here.
Quoted 27 lines
/**
*i915_driver_init_hw- setup state requiring device access
* @dev_priv: device privatedrivers/gpu/drm/i915/i915_drv.c +3
-0@@ -1081,6 +1248,12 @@ static int i915_driver_init_hw(struct drm_i915_private *dev_priv) DRM_DEBUG_DRIVER("can't enable MSI"); } +/* +* Fill the memdev structure to get the system raw bandwidth +* This will be used by WM algorithm, to implement GEN9 based WA
drivers/gpu/drm/i915/i915_drv.c +3
-0+*/ +intel_get_memdev_info (dev_priv); +return 0 ; out_ggtt:drivers/gpu/drm/i915/i915_drv.h +0
-0@@ -2297,6 +2297,18 @@ struct drm_i915_private {bool distrust_bios_wm; } wm;
Quoted 12 lines
drivers/gpu/drm/i915/i915_drv.h struct drm_i915_private { · line 2308
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.h +9
-0+struct memdev_info { +bool valid; +uint32_t bandwidth_kbps; +uint8_t num_channels; +enum rank { +DRAM_RANK_INVALID =0 , +DRAM_RANK_SINGLE , +DRAM_RANK_DUAL + } rank;This was previously an anonymous enum (enum { ... } rank;), but now it's "enum rank", which is too generic for a .h file. Either use a more specific name (enum
memdev_rank?) or try to make it anonymous again.
changed to memdev_rank.
Quoted 16 lines
Quoted 16 lines
drivers/gpu/drm/i915/i915_drv.h +3
-0+ } memdev_info; + +struct i915_runtime_pm pm;struct {drivers/gpu/drm/i915/i915_reg.h +0
-0@@ -8015,6 +8015,43 @@ enum {#define DC_STATE_DEBUG_MASK_CORES (1<<0)#define DC_STATE_DEBUG_MASK_MEMORY_UP (1<<1)
Quoted 6 lines
drivers/gpu/drm/i915/i915_reg.h enum { · line 8021
Quoted 5 lines
drivers/gpu/drm/i915/i915_reg.h +4
-0+#define BXT_P_CR_MC_BIOS_REQ_0_0_0 _MMIO(MCHBAR_MIRROR_BASE_S NB + 0x7114) +#define BXT_REQ_DATA_MASK 0x3F +#define BXT_DRAM_ACTIVE_CHANNEL_SHIFT 12 +#define BXT_DRAM_ACTIVE_CHANNEL_MASK 0xFWe're not using these two definitions anywhere.
These are duplicate for definition below, my mistake, removing them.
Quoted 13 lines
drivers/gpu/drm/i915/i915_reg.h enum { · line 8025
Quoted 5 lines
drivers/gpu/drm/i915/i915_reg.h +4
-0+/* +* BIOS programs this field of REQ_DATA [5:0] in integer +* multiple of 133333 KHz (133.33MHz) +*/ Now that we're using decimal on the definition, I don't think this comment is useful.
drivers/gpu/drm/i915/i915_reg.h enum { · line 8026
Quoted 1 line
drivers/gpu/drm/i915/i915_reg.h +1
-0+#define BXT_MEMORY_FREQ_MULTIPLIER_KHZ 133333We don't use tabs like that in our definitions.
Quoted 9 lines
drivers/gpu/drm/i915/i915_reg.h enum { · line 8034
Quoted 7 lines
drivers/gpu/drm/i915/i915_reg.h +5
-0+#define BXT_D_CR_DRP0_DUNIT8 0x1000 +#define BXT_D_CR_DRP0_DUNIT9 0x1200 +#define BXT_D_CR_DRP0_DUNIT_MAX 4 +#define _MMIO_MCHBAR_DUNIT (x, a, b) _MMIO(MCHBAR_MIRROR_BASE_SNB + (a) + (x)*((b)-(a))) +#define BXT_D_CR_DRP0_DUNIT (x) _MMIO_MCHBAR_DUNIT(x, BXT_D_CR_DRP0_DUNIT8, BXT_D_CR_DRP0_DUNIT9)But then
BXT_D_CR_DRP0_DUNIT(1)means DUNIT9 instead of DUNIT1... That's very unintuitive.
now changed it to pass 8-11 to this macro, (from start to end of dunit)
Quoted 31 lines
drivers/gpu/drm/i915/i915_reg.h enum { · line 8033
Quoted 4 lines
drivers/gpu/drm/i915/i915_reg.h +2
-0+#define BXT_DRAM_CHANNEL_ACTIVE_SHIFT 12 +#define BXT_DRAM_CHANNEL_ACTIVE_MASK 0xFThese two definitions don't belong to the register immediately above.
drivers/gpu/drm/i915/i915_reg.h enum { · line 8039
Quoted 6 lines
drivers/gpu/drm/i915/i915_reg.h +6
-0+#define BXT_DRAM_RANK_MASK 0x3 +#define BXT_DRAM_RANK_SINGLE 0x1 +#define BXT_DRAM_RANK_DUAL 0x3 + +/* +* SKL memory frequeny multiplier is 266667 KHz (266.67 MHz) s/frequeny/frequency/
Also, this is not how we do one-line comments.
And now that the number is in decimal we probably don't even need the comment here since the definition is a little more obvious.
drivers/gpu/drm/i915/i915_reg.h enum { · line 8044
Quoted 2 lines
drivers/gpu/drm/i915/i915_reg.h +2
-0+*/ +#define SKL_MEMORY_FREQ_MULTIPLIER_KHZ 266667We don't use tabs like that in our definitions.
Quoted 3 lines
drivers/gpu/drm/i915/i915_reg.h +1
-0+#define SKL_MC_BIOS_DATA_0_0_0_MCHBAR_PCU _MMIO(MCHBAR_MIRROR
_BASE_SNB+0x5E04)
+#defineSKL_REQ_DATA_MASK(0xF<< 0)Blank lines separating different registers would be good.
Quoted 6 lines
Quoted 5 lines
+#define
SKL_MAD_DIMM_CH0_0_0_0_MCHBAR_MCMAIN_MMIO(MCHBAR_MIRROR_BASE_SNB+0x500C)
+#defineSKL_MAD_DIMM_CH1_0_0_0_MCHBAR_MCMAIN_MMIO(MCHBAR_MIRROR_BASE_SNB+0x5010)
+#defineSKL_DRAM_SIZE_MASK0x1F
0x3F
fixed.
patch is pushed as a new series @ https://patchwork.freedesktop.org/series/18842/
Regards,
-Mahesh
Quoted 15 lines
Quoted 11 lines
drivers/gpu/drm/i915/i915_reg.h +8
-0+#define SKL_DRAM_SIZE_L_SHIFT 0 +#define SKL_DRAM_SIZE_S_SHIFT 16 +#define SKL_DRAM_RANK_MASK 0x1 +#define SKL_DRAM_RANK_L_SHIFT 10 +#define SKL_DRAM_RANK_S_SHIFT 26 +#define SKL_DRAM_RANK_SINGLE 0x0 +#define SKL_DRAM_RANK_DUAL 0x1 +/* Please see hsw_read_dcomp() and hsw_write_dcomp() before using Quoted 4 lines
this register,
* since on HSW we can't write to it usingI915_WRITE. */
#defineD_COMP_HSW_MMIO(MCHBAR_MIRROR_BASE_S
NB +0x5F0C)
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v7 8/8 drm/i915/gen9: WM memory bandwidth related workaround
In reply to Mahesh KumarThis patch implemnets Workariunds related to display arbitrated memory bandwidth. These WA are applicabe for all gen-9 based platforms.
Changes since v1:
- Rebase on top of Paulo's patch series
Changes since v2:
- Address review comments
- Rebase/rework as per other patch changes in series
Changes since v3:
- Rework based on Maarten's comments
--- drivers/gpu/drm/i915/i915_drv.h | 11 +++ drivers/gpu/drm/i915/intel_display.c | 24 ++++++ drivers/gpu/drm/i915/intel_pm.c | 155 +++++++++++++++++++++++++++++++++-- 3 files changed, 181 insertions(+), 9 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +11 -0
@@ -1216,6 +1216,13 @@ enum intel_sbi_destination { SBI_MPHY, }; +/* SKL+ Watermark arbitrated display bandwidth Workarounds */ +enum watermark_memory_wa { +WATERMARK_WA_NONE , +WATERMARK_WA_X_TILED , +WATERMARK_WA_Y_TILED , +}; +#define QUIRK_PIPEA_FORCE (1<<0)#define QUIRK_LVDS_SSC_DISABLE (1<<1)#define QUIRK_INVERT_BRIGHTNESS (1<<2)
@@ -1788,6 +1795,10 @@ struct skl_ddb_allocation {struct skl_wm_values {unsigned dirty_pipes; +/* any WaterMark memory workaround Required */ +enum watermark_memory_wa mem_wa; +uint32_t pipe_bw_kbps[I915_MAX_PIPES ]; +bool pipe_ytiled[I915_MAX_PIPES ];struct skl_ddb_allocation ddb; };
drivers/gpu/drm/i915/intel_display.c +24 -0
@@ -14574,6 +14574,28 @@ static void intel_atomic_track_fbs(struct drm_atomic_state *state) to_intel_plane(plane)->frontbuffer_bit); } +static void +intel_update_wm_bw_wa (struct drm_atomic_state *state) +{ +struct intel_atomic_state *intel_state =to_intel_atomic_state (state); +struct drm_i915_private *dev_priv =to_i915 (state->dev); +const struct drm_crtc *crtc; +const struct drm_crtc_state *cstate; +struct skl_wm_values *results = &intel_state->wm_results; +struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw; +int i; + +if (!IS_GEN9 (dev_priv)) +return ; + +for_each_crtc_in_state (state, crtc, cstate, i) { + hw_vals->pipe_bw_kbps[i] = results->pipe_bw_kbps[i]; + hw_vals->pipe_ytiled[i] = results->pipe_ytiled[i]; + } + + hw_vals->mem_wa = results->mem_wa; +} +/** * intel_atomic_commit - commit validated state object * @dev: DRM device
@@ -14614,6 +14636,8 @@ static int intel_atomic_commit(struct drm_device *dev,intel_shared_dpll_commit (state);intel_atomic_track_fbs (state); +intel_update_wm_bw_wa (state); +if (intel_state->modeset) {memcpy (dev_priv->min_pixclk, intel_state->min_pixclk,sizeof (intel_state->min_pixclk));
drivers/gpu/drm/i915/intel_pm.c +146 -9
203 lines not on this page · raw message
Signature · 6 lines
2.10.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v7 8/8] drm/i915/gen9: WM memory bandwidth related workaround
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 3 lines
This patch implemnets Workariunds related to display arbitrated memory
bandwidth. These WA are applicabe for all gen-9 based platforms.
3 typos above.
The WA is already implemented. What the patch does is that it opts-out of the WA in case we conclude it's safe. Please say this in the commit message.
Quoted 40 lines
Changes since v1:
- Rebase on top of Paulo's patch series Changes since v2:
- Address review comments
- Rebase/rework as per other patch changes in series Changes since v3:
- Rework based on Maarten's commentsSigned-off-by Mahesh Kumar <[email protected]>---
drivers/gpu/drm/i915/i915_drv.h | 11 +++drivers/gpu/drm/i915/intel_display.c| 24 ++++++ drivers/gpu/drm/i915/intel_pm.c | 155
+++++++++++++++++++++++++++++++++--3 files changed, 181 insertions(+), 9 deletions(-)drivers/gpu/drm/i915/i915_drv.h +0
-0@@ -1216,6 +1216,13 @@ enum intel_sbi_destination { SBI_MPHY, };drivers/gpu/drm/i915/i915_drv.h +7
-0+/* SKL+ Watermark arbitrated display bandwidth Workarounds */ +enum watermark_memory_wa { +WATERMARK_WA_NONE , +WATERMARK_WA_X_TILED , +WATERMARK_WA_Y_TILED , +}; +#define
QUIRK_PIPEA_FORCE(1<<0)
#defineQUIRK_LVDS_SSC_DISABLE(1<<1)
#defineQUIRK_INVERT_BRIGHTNESS(1<<2)drivers/gpu/drm/i915/i915_drv.h +0
-0@@ -1788,6 +1795,10 @@ struct skl_ddb_allocation {
structskl_wm_values{
unsigneddirty_pipes;
drivers/gpu/drm/i915/i915_drv.h struct skl_ddb_allocation { · line 1798
Quoted 1 line
drivers/gpu/drm/i915/i915_drv.h +1
-0+/* any WaterMark memory workaround Required */
CapitaliZation is weird Here. Besides, no need for the comment.
Quoted 29 lines
drivers/gpu/drm/i915/i915_drv.h +3
-0+enum watermark_memory_wa mem_wa; +uint32_t pipe_bw_kbps[I915_MAX_PIPES ]; +bool pipe_ytiled[I915_MAX_PIPES ];struct
skl_ddb_allocationddb;
};drivers/gpu/drm/i915/intel_display.c +0
-0@@ -14574,6 +14574,28 @@ static void intel_atomic_track_fbs(struct drm_atomic_state *state)to_intel_plane(plane)-
Quoted 1 line
frontbuffer_bit);}
drivers/gpu/drm/i915/intel_display.c +4
-0+static void +intel_update_wm_bw_wa (struct drm_atomic_state *state) +{ + struct intel_atomic_state *intel_state =
to_intel_atomic_state(state);drivers/gpu/drm/i915/intel_display.c +7
-0+struct drm_i915_private *dev_priv =to_i915 (state->dev); +const struct drm_crtc *crtc; +const struct drm_crtc_state *cstate; +struct skl_wm_values *results = &intel_state->wm_results; +struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw; +int i; +
drivers/gpu/drm/i915/intel_display.c static void intel_atomic_track_fbs(struct drm_atomic_state *state) · line 14597
Quoted 10 lines
drivers/gpu/drm/i915/intel_display.c +10
-0+if (!IS_GEN9 (dev_priv)) +return ; + +for_each_crtc_in_state (state, crtc, cstate, i) { + hw_vals->pipe_bw_kbps[i] = results->pipe_bw_kbps[i]; + hw_vals->pipe_ytiled[i] = results->pipe_ytiled[i]; + } + + hw_vals->mem_wa = results->mem_wa; +}
I think we can just patch skl_copy_wm_for_pipe() to also copy the fields we need instead of adding this function. Whouldn't that be much simpler? Anyway, this one looks correct, so no need to change if you don't want.
Quoted 31 lines
+
/**
*intel_atomic_commit- commit validated state object
* @dev: DRM devicedrivers/gpu/drm/i915/intel_display.c +0
-0@@ -14614,6 +14636,8 @@ static int intel_atomic_commit(structdrm_device *dev, intel_shared_dpll_commit(state); intel_atomic_track_fbs(state);drivers/gpu/drm/i915/intel_display.c +2
-0+intel_update_wm_bw_wa (state); +if (
intel_state->modeset) { memcpy(dev_priv->min_pixclk, intel_state-Quoted 1 line
min_pixclk,sizeof(intel_state->min_pixclk));
drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -2878,11 +2878,7 @@ bool ilk_disable_lp_wm(struct drm_device *dev)
#defineSKL_SAGV_BLOCK_TIME30 /* µs */
-/*
- * FIXME: We still don't have the proper code detect if we need to apply the WA,
- * so assume we'll always need it in order to avoid underruns.
- */drivers/gpu/drm/i915/intel_pm.c +1
-1-static bool skl_needs_memory_bw_wa (struct intel_atomic_state *state) +static bool intel_needs_memory_bw_wa(struct intel_atomic_state
Quoted 1 line
*state)
Why are we renaming this function?
Also, in my last review I suggested that we kill this function. I don't think it makes sense to have this function anymore.
Besides, function intel_can_enable_sagv() needs to look at the value we assigned to enum watermark_memory_wa, not get a true/false based on platform version.
Quoted 30 lines
{
structdrm_i915_private*dev_priv= to_i915(state-Quoted 1 line
base.dev);
drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -3056,7 +3052,7 @@ bool intel_can_enable_sagv(struct drm_atomic_state *state)
latency = dev_priv->wm.skl_latency[level];drivers/gpu/drm/i915/intel_pm.c +1
-1- if (skl_needs_memory_bw_wa(intel_state) && + if (intel_needs_memory_bw_wa(intel_state) &&plane->base.state->fb->modifier ==
I915_FORMAT_MOD_X_TILED) latency += 15;drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -3584,7 +3580,7 @@ static int skl_compute_plane_wm(const structdrm_i915_private *dev_priv, uint32_t y_min_scanlines;struct
intel_atomic_state*state =to_intel_atomic_state(cstate->base.state);drivers/gpu/drm/i915/intel_pm.c +1
-1-bool apply_memory_bw_wa =skl_needs_memory_bw_wa (state); +enum watermark_memory_wa mem_wa;bool
y_tiled,x_tiled;
if (latency == 0 || !cstate->base.active || !intel_pstate-Quoted 1 line
base.visible) {
drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -3600,7 +3596,8 @@ static int skl_compute_plane_wm(const structdrm_i915_private *dev_priv, if (IS_KABYLAKE(dev_priv) && dev_priv->ipc_enabled) latency += 4;
drivers/gpu/drm/i915/intel_pm.c static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, · line 3599
Quoted 2 lines
drivers/gpu/drm/i915/intel_pm.c +1
-1-if (apply_memory_bw_wa && x_tiled) + mem_wa = state->wm_results.mem_wa;
This assignment could have been done during declaration.
Quoted 24 lines
drivers/gpu/drm/i915/intel_pm.c +1
-0+if (mem_wa !=WATERMARK_WA_NONE && x_tiled)latency += 15;
width =drm_rect_width(&intel_pstate->base.src)>> 16;drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -3635,7 +3632,7 @@ static int skl_compute_plane_wm(const structdrm_i915_private *dev_priv, y_min_scanlines = 4; }drivers/gpu/drm/i915/intel_pm.c +1
-1-if (apply_memory_bw_wa) +if (mem_wa ==WATERMARK_WA_Y_TILED )
y_min_scanlines*= 2;
plane_bytes_per_line= width * cpp;drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -4063,6 +4060,15 @@ skl_compute_ddb(struct drm_atomic_state *state) }
/*drivers/gpu/drm/i915/intel_pm.c +4
-0+* If Watermark workaround is changed we need to recalculate +* watermark values for all active pipes +*/ + if (intel_state->wm_results.mem_wa != dev_priv-Quoted 1 line
wm.skl_hw.mem_wa) {
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4069
Quoted 3 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ realloc_pipes = ~0 ; + intel_state->wm_results.dirty_pipes = ~0 ; + }
But then skl_ddb_add_affected_planes() only actually adds to the commit the planes that have different DDB partitioning. It seems to me that it may be possible to have a case where the WA changes and the DDB stays the same, so we need to find a way to include every CRTC there. Maybe we could just go and call the function that adds every CRTC here.
Quoted 25 lines
drivers/gpu/drm/i915/intel_pm.c +2
-0+ +/* * We're not recomputing for the pipes not included in the commit, so
* make sure we start with the current state. */drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -4087,6 +4093,133 @@ skl_compute_ddb(struct drm_atomic_state *state) return 0; }drivers/gpu/drm/i915/intel_pm.c +6
-0+static int +skl_compute_memory_bandwidth_wm_wa (struct drm_atomic_state *state) +{ +struct drm_device *dev = state->dev; +struct drm_i915_private *dev_priv =to_i915 (dev); + struct intel_atomic_state *intel_state =
to_intel_atomic_state(state);drivers/gpu/drm/i915/intel_pm.c +7
-0+struct memdev_info *memdev_info = &dev_priv->memdev_info; +struct skl_wm_values *results = &intel_state->wm_results; +struct drm_crtc *crtc; +struct drm_crtc_state *cstate; +int active_pipes =0 ; +uint32_t max_pipe_bw_kbps =0 , total_pipe_bw_kbps; +int display_bw_percentage;
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4118
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c +10
-0+bool y_tile_enabled = false; +int ret, i; + +if (!intel_needs_memory_bw_wa (intel_state)) { + results->mem_wa =WATERMARK_WA_NONE ; +return 0 ; + } + +if (!memdev_info->valid) + goto exit;
There's no reason to use a label if it's only called here. Just move all that code to the if statement.
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4123
Quoted 5 lines
drivers/gpu/drm/i915/intel_pm.c +5
-0+ + ret =drm_modeset_lock (&dev->mode_config.connection_mutex, + state->acquire_ctx); +if (ret) +return ret;
Why are we getting this here? Why this lock specifically? What's it protecting?
Doc says:
- ww mutex protecting connector state and routing - protects connector->encoder and encoder->crtc links
We're not touching any of that here.
Quoted 75 lines
drivers/gpu/drm/i915/intel_pm.c +2
-0+ + memcpy(results->pipe_bw_kbps, dev_priv-Quoted 1 line
wm.skl_hw.pipe_bw_kbps,
drivers/gpu/drm/i915/intel_pm.c +2
-0+ sizeof(results->pipe_bw_kbps)); + memcpy(results->pipe_ytiled, dev_priv-Quoted 1 line
wm.skl_hw.pipe_ytiled,
drivers/gpu/drm/i915/intel_pm.c +16
-0+ sizeof(results->pipe_ytiled)); + +for_each_crtc_in_state (state, crtc, cstate, i) { +struct drm_plane *plane; +const struct drm_plane_state *pstate; +int active_planes =0 ; +uint32_t max_plane_bw_kbps =0 ; + + results->pipe_ytiled[i] = false; + +if (!cstate->active) { + results->pipe_bw_kbps[i] =0 ; +continue ; + } + + drm_atomic_crtc_state_for_each_plane_state(plane,pstate,
drivers/gpu/drm/i915/intel_pm.c +1
-0+ cstate) {
drivers/gpu/drm/i915/intel_pm.c +3
-0+struct drm_framebuffer *fb; +uint32_t plane_bw_kbps; + enum plane_id id = to_intel_plane(plane)-Quoted 1 line
id;
drivers/gpu/drm/i915/intel_pm.c +12
-0+ +if (!pstate->visible) +continue ; + +if (WARN_ON (!pstate->fb)) +continue ; + +if (id ==PLANE_CURSOR ) +continue ; + + fb = pstate->fb; + if ((fb->modifier == I915_FORMAT_MOD_Y_TILED||
drivers/gpu/drm/i915/intel_pm.c +1
-0+ fb->modifier ==
I915_FORMAT_MOD_Yf_TILED))drivers/gpu/drm/i915/intel_pm.c +3
-0+ results->pipe_ytiled[i] = true; + + plane_bw_kbps =skl_adjusted_plane_pixel_rate(
drivers/gpu/drm/i915/intel_pm.c +1
-0+ to_intel_crtc_state(cstate),
drivers/gpu/drm/i915/intel_pm.c +1
-0+ to_intel_plane_state(pstate));
drivers/gpu/drm/i915/intel_pm.c +2
-0+ max_plane_bw_kbps =max (plane_bw_kbps, + max_plane_bw
_kbps);drivers/gpu/drm/i915/intel_pm.c +3
-0+ active_planes++; + } + results->pipe_bw_kbps[i] = max_plane_bw_kbps *
active_planes;drivers/gpu/drm/i915/intel_pm.c +14
-0+ } + +for_each_pipe (dev_priv, i) { +if (results->pipe_bw_kbps[i]) { + max_pipe_bw_kbps =max (max_pipe_bw_kbps, + results->pipe_bw_kbps[i]); + active_pipes++; + } +if (results->pipe_ytiled[i]) + y_tile_enabled = true; + } + + total_pipe_bw_kbps = max_pipe_bw_kbps * active_pipes; + display_bw_percentage = DIV_ROUND_UP_ULL(total_pipe_bw_kbps* 100,
drivers/gpu/drm/i915/intel_pm.c +1
-0+ memdev_info-
Quoted 1 line
Quoted 1 line
bandwidth_kbps);
Shouldn't this be just DIV_ROUND_UP()? I don't see 64 bit variables here, nor the possibility of overflows.
Quoted 12 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ +/* +* If there is any Ytile plane enabled and arbitrated display
drivers/gpu/drm/i915/intel_pm.c +4
-0+ * bandwidth > 20% of raw system memory bandwidth + * Enale Y-tile related WA + * + * If memory is dual channel single rank, Xtile limit = 35%,else Xtile
drivers/gpu/drm/i915/intel_pm.c +3
-0+ * limit = 60% + * If there is no Ytile plane enabled and + * arbitrated display bandwidth > Xtile limit
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4205
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c +10
-0+* Enable X-tile realated WA +*/ +if (y_tile_enabled && (display_bw_percentage >20 )) + results->mem_wa =WATERMARK_WA_Y_TILED ; +else { +int x_tile_percentage =60 ; +enum rank rank =DRAM_RANK_SINGLE ; + +if (memdev_info->rank !=DRAM_RANK_INVALID ) + rank = memdev_info->rank;
I really think that the previous patch should prevent this. If memdev_info->rank is invalid then memdev_info->valid should also be false and we'd have returned at the beginning of the function. With that, this part of the code could be simplified a little bit since then we'd have no need to choose a default rank.
Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ + if ((rank == DRAM_RANK_SINGLE) && + (memdev_info->num_channels== 2))
drivers/gpu/drm/i915/intel_pm.c +4
-0+ x_tile_percentage =35 ; + +if (display_bw_percentage > x_tile_percentage) + results->mem_wa =WATERMARK_WA_X_TILED ;
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4222
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c +10
-0+ } + +return 0 ; + +exit: + results->mem_wa =WATERMARK_WA_Y_TILED ; +return 0 ; +} + +
Two blank lines here.
Quoted 11 lines
static void
skl_copy_wm_for_pipe(structskl_wm_values*dst,
structskl_wm_values*src,drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -4162,6 +4295,10 @@ skl_compute_wm(struct drm_atomic_state *state) /* Clear all dirty flags */ results->dirty_pipes = 0;drivers/gpu/drm/i915/intel_pm.c +4
-0+ ret =skl_compute_memory_bandwidth_wm_wa (state); +if (ret) +return ret; +
Quoted 3 lines
ret =
skl_compute_ddb(state);
if (ret)
return ret;
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v7 8/8] drm/i915/gen9: WM memory bandwidth related workaround
Hi,
Quoted 9 lines · Paulo Zanoni
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 3 lines
This patch implemnets Workariunds related to display arbitrated memory
bandwidth. These WA are applicabe for all gen-9 based platforms.3 typos above.
The WA is already implemented. What the patch does is that it opts-out of the WA in case we conclude it's safe. Please say this in the commit message.
updated the commit msg
Quoted 122 lines
Quoted 36 lines
Changes since v1:
- Rebase on top of Paulo's patch series
Changes since v2:
- Address review comments
- Rebase/rework as per other patch changes in series
Changes since v3:
- Rework based on Maarten's commentsSigned-off-by Mahesh Kumar <[email protected]>--- drivers/gpu/drm/i915/i915_drv.h | 11 +++ drivers/gpu/drm/i915/intel_display.c | 24 ++++++ drivers/gpu/drm/i915/intel_pm.c | 155 +++++++++++++++++++++++++++++++++-- 3 files changed, 181 insertions(+), 9 deletions(-)drivers/gpu/drm/i915/i915_drv.h +7
-0@@ -1216,6 +1216,13 @@ enum intel_sbi_destination { SBI_MPHY, }; +/* SKL+ Watermark arbitrated display bandwidth Workarounds */ +enum watermark_memory_wa { +WATERMARK_WA_NONE , +WATERMARK_WA_X_TILED , +WATERMARK_WA_Y_TILED , +}; +#define QUIRK_PIPEA_FORCE (1<<0)#define QUIRK_LVDS_SSC_DISABLE (1<<1)#define QUIRK_INVERT_BRIGHTNESS (1<<2)drivers/gpu/drm/i915/i915_drv.h struct skl_ddb_allocation { · line 1798
Quoted 4 lines
drivers/gpu/drm/i915/i915_drv.h +1
-0@@ -1788,6 +1795,10 @@ struct skl_ddb_allocation {struct skl_wm_values {unsigned dirty_pipes; +/* any WaterMark memory workaround Required */ CapitaliZation is weird Here. Besides, no need for the comment.
Quoted 29 lines
drivers/gpu/drm/i915/i915_drv.h +3
-0+enum watermark_memory_wa mem_wa; +uint32_t pipe_bw_kbps[I915_MAX_PIPES ]; +bool pipe_ytiled[I915_MAX_PIPES ];struct skl_ddb_allocation ddb; };drivers/gpu/drm/i915/intel_display.c +11
-0@@ -14574,6 +14574,28 @@ static void intel_atomic_track_fbs(struct drm_atomic_state *state) to_intel_plane(plane)- > frontbuffer_bit); } +static void +intel_update_wm_bw_wa(struct drm_atomic_state *state) +{ + struct intel_atomic_state *intel_state = to_intel_atomic_state(state); + struct drm_i915_private *dev_priv = to_i915(state->dev); + const struct drm_crtc *crtc; + const struct drm_crtc_state *cstate; + struct skl_wm_values *results = &intel_state->wm_results; + struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw; + int i; +drivers/gpu/drm/i915/intel_display.c static void intel_atomic_track_fbs(struct · line 14600
Quoted 10 lines
drivers/gpu/drm/i915/intel_display.c +10
-0+if (!IS_GEN9 (dev_priv)) +return ; + +for_each_crtc_in_state (state, crtc, cstate, i) { + hw_vals->pipe_bw_kbps[i] = results->pipe_bw_kbps[i]; + hw_vals->pipe_ytiled[i] = results->pipe_ytiled[i]; + } + + hw_vals->mem_wa = results->mem_wa; +}I think we can just patch
skl_copy_wm_for_pipe()to also copy the fields we need instead of adding this function. Whouldn't that be much simpler? Anyway, this one looks correct, so no need to change if you don't want.Quoted 22 lines
drivers/gpu/drm/i915/intel_display.c +1
-0+/**
*intel_atomic_commit- commit validated state object
* @dev: DRM devicedrivers/gpu/drm/i915/intel_display.c +2
-0@@ -14614,6 +14636,8 @@ static int intel_atomic_commit(structdrm_device *dev,intel_shared_dpll_commit (state );intel_atomic_track_fbs (state); +intel_update_wm_bw_wa (state); +if (intel_state->modeset) { memcpy(dev_priv->min_pixclk, intel_state-Quoted 1 line
min_pixclk,
sizeof(intel_state->min_pixclk));drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -2878,11 +2878,7 @@ bool ilk_disable_lp_wm(struct drm_device *dev)drivers/gpu/drm/i915/intel_pm.c bool ilk_disable_lp_wm(struct drm_device *dev) · line 2883
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c +1
-5#define SKL_SAGV_BLOCK_TIME 30/* µs */ -/* -* FIXME: We still don't have the proper code detect if we need to apply the WA, -* so assume we'll always need it in order to avoid underruns. -*/ -static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state) +static bool intel_needs_memory_bw_wa(struct intel_atomic_state *state)Why are we renaming this function?
Quoted 6 lines
Also, in my last review I suggested that we kill this function. I don't think it makes sense to have this function anymore.
Besides, function
intel_can_enable_sagv()needs to look at the value we assigned to enumwatermark_memory_wa, not get a true/false based on platform version.
removed the function
Quoted 27 lines
Quoted 27 lines
drivers/gpu/drm/i915/intel_pm.c +0
-0{struct
drm_i915_private*dev_priv= to_i915(state-Quoted 1 line
base.dev);
drivers/gpu/drm/i915/intel_pm.c +1
-1@@ -3056,7 +3052,7 @@ bool intel_can_enable_sagv(struct drm_atomic_state *state) latency = dev_priv->wm.skl_latency[level]; -if (skl_needs_memory_bw_wa (intel_state) && +if (intel_needs_memory_bw_wa (intel_state) && plane->base.state->fb->modifier ==I915_FORMAT_MOD_X_TILED )latency += 15;drivers/gpu/drm/i915/intel_pm.c +1
-1@@ -3584,7 +3580,7 @@ static int skl_compute_plane_wm(const structdrm_i915_private *dev_priv, uint32_t y_min_scanlines;struct intel_atomic_state *state =to_intel_atomic_state (cstate->base.state); -bool apply_memory_bw_wa =skl_needs_memory_bw_wa (state); +enum watermark_memory_wa mem_wa;bool y_tiled, x_tiled;if (latency == 0 || !cstate->base.active || !intel_pstate-
Quoted 1 line
base.visible) {
Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c static int skl_compute_plane_wm(const struct · line 3600
Quoted 7 lines
drivers/gpu/drm/i915/intel_pm.c +1
-1@@ -3600,7 +3596,8 @@ static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv, if (IS_KABYLAKE(dev_priv) && dev_priv->ipc_enabled) latency += 4; -if (apply_memory_bw_wa && x_tiled) + mem_wa = state->wm_results.mem_wa;This assignment could have been done during declaration.
done
Quoted 26 lines
Quoted 26 lines
drivers/gpu/drm/i915/intel_pm.c +1
-0+if (mem_wa !=WATERMARK_WA_NONE && x_tiled) latency +=15 ;width = drm_rect_width(&intel_pstate->base.src) >> 16;drivers/gpu/drm/i915/intel_pm.c +1
-1@@ -3635,7 +3632,7 @@ static int skl_compute_plane_wm(const structdrm_i915_private *dev_priv, y_min_scanlines =4 ; } -if (apply_memory_bw_wa) +if (mem_wa ==WATERMARK_WA_Y_TILED ) y_min_scanlines *=2 ;plane_bytes_per_line = width * cpp;drivers/gpu/drm/i915/intel_pm.c +4
-0@@ -4063,6 +4060,15 @@ skl_compute_ddb(struct drm_atomic_state *state) }/* +* If Watermark workaround is changed we need to recalculate +* watermark values for all active pipes +*/ +if (intel_state->wm_results.mem_wa != dev_priv- > wm.skl_hw.mem_wa) {
Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4069
Quoted 3 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ realloc_pipes = ~0 ; + intel_state->wm_results.dirty_pipes = ~0 ; + }But then
skl_ddb_add_affected_planes()only actually adds to the commit the planes that have different DDB partitioning. It seems to me that it may be possible to have a case where the WA changes and the DDB stays the same, so we need to find a way to include every CRTC there. Maybe we could just go and call the function that adds every CRTC here.
modified this part of code to calculate WM for all pipes if WA is changing & add pipe only if WM values are changing. We are holding a ww_mutex, so this operation should be race condition free.
Quoted 30 lines
Quoted 30 lines
drivers/gpu/drm/i915/intel_pm.c +2
-0+ +/* * We're not recomputing for the pipes not included in the commit, so
* make sure we start with the current state. */drivers/gpu/drm/i915/intel_pm.c +17
-0@@ -4087,6 +4093,133 @@ skl_compute_ddb(struct drm_atomic_state *state)return 0 ; } +static int +skl_compute_memory_bandwidth_wm_wa (struct drm_atomic_state *state) +{ +struct drm_device *dev = state->dev; +struct drm_i915_private *dev_priv =to_i915 (dev); +struct intel_atomic_state *intel_state =to_intel_atomic_state (state); +struct memdev_info *memdev_info = &dev_priv->memdev_info; +struct skl_wm_values *results = &intel_state->wm_results; +struct drm_crtc *crtc; +struct drm_crtc_state *cstate; +int active_pipes =0 ; +uint32_t max_pipe_bw_kbps =0 , total_pipe_bw_kbps; +int display_bw_percentage; +bool y_tile_enabled = false; +int ret, i; + +if (!intel_needs_memory_bw_wa (intel_state)) {
Quoted 8 lines
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4118
Quoted 6 lines
drivers/gpu/drm/i915/intel_pm.c +6
-0+ results->mem_wa =WATERMARK_WA_NONE ; +return 0 ; + } + +if (!memdev_info->valid) + goto exit;There's no reason to use a label if it's only called here. Just move all that code to the if statement.
done
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4123
Quoted 7 lines
drivers/gpu/drm/i915/intel_pm.c +5
-0+ + ret =drm_modeset_lock (&dev->mode_config.connection_mutex, + state->acquire_ctx); +if (ret) +return ret;Why are we getting this here? Why this lock specifically? What's it protecting?
Quoted 5 lines
Doc says:
- ww mutex protecting connector state and routing
- protectsconnector->encoderandencoder->crtclinksWe're not touching any of that here.
Created a new ww_mutex lock to prevent wm related calculation, when it may affect other CRTC's also. using that now instead of connection_mutex.
Quoted 77 lines
Quoted 77 lines
drivers/gpu/drm/i915/intel_pm.c +4
-0+ +memcpy (results->pipe_bw_kbps, dev_priv- > wm.skl_hw.pipe_bw_kbps, +sizeof (results->pipe_bw_kbps)); + memcpy(results->pipe_ytiled, dev_priv-Quoted 1 line
wm.skl_hw.pipe_ytiled,
drivers/gpu/drm/i915/intel_pm.c +16
-0+ sizeof(results->pipe_ytiled)); + +for_each_crtc_in_state (state, crtc, cstate, i) { +struct drm_plane *plane; +const struct drm_plane_state *pstate; +int active_planes =0 ; +uint32_t max_plane_bw_kbps =0 ; + + results->pipe_ytiled[i] = false; + +if (!cstate->active) { + results->pipe_bw_kbps[i] =0 ; +continue ; + } + + drm_atomic_crtc_state_for_each_plane_state(plane,pstate,
drivers/gpu/drm/i915/intel_pm.c +1
-0+ cstate) {
drivers/gpu/drm/i915/intel_pm.c +3
-0+struct drm_framebuffer *fb; +uint32_t plane_bw_kbps; + enum plane_id id = to_intel_plane(plane)-Quoted 1 line
id;
drivers/gpu/drm/i915/intel_pm.c +12
-0+ +if (!pstate->visible) +continue ; + +if (WARN_ON (!pstate->fb)) +continue ; + +if (id ==PLANE_CURSOR ) +continue ; + + fb = pstate->fb; + if ((fb->modifier == I915_FORMAT_MOD_Y_TILED||
drivers/gpu/drm/i915/intel_pm.c +1
-0+ fb->modifier ==
I915_FORMAT_MOD_Yf_TILED))drivers/gpu/drm/i915/intel_pm.c +3
-0+ results->pipe_ytiled[i] = true; + + plane_bw_kbps =skl_adjusted_plane_pixel_rate(
drivers/gpu/drm/i915/intel_pm.c +1
-0+ to_intel_crtc_state(cstate),
drivers/gpu/drm/i915/intel_pm.c +1
-0+ to_intel_plane_state(pstate));
drivers/gpu/drm/i915/intel_pm.c +2
-0+ max_plane_bw_kbps =max (plane_bw_kbps, + max_plane_bw
_kbps);drivers/gpu/drm/i915/intel_pm.c +3
-0+ active_planes++; + } + results->pipe_bw_kbps[i] = max_plane_bw_kbps *
active_planes;drivers/gpu/drm/i915/intel_pm.c +14
-0+ } + +for_each_pipe (dev_priv, i) { +if (results->pipe_bw_kbps[i]) { + max_pipe_bw_kbps =max (max_pipe_bw_kbps, + results->pipe_bw_kbps[i]); + active_pipes++; + } +if (results->pipe_ytiled[i]) + y_tile_enabled = true; + } + + total_pipe_bw_kbps = max_pipe_bw_kbps * active_pipes; + display_bw_percentage = DIV_ROUND_UP_ULL(total_pipe_bw_kbps* 100,
drivers/gpu/drm/i915/intel_pm.c +1
-0+ memdev_info-
Quoted 3 lines
Quoted 1 line
Quoted 1 line
bandwidth_kbps);Shouldn't this be just
DIV_ROUND_UP()? I don't see 64 bit variables here, nor the possibility of overflows.
If we have 3 4K pannels & each have 3 planes enabled with scaling then this value theoretically may go beyond 32bits,
type-casted the variable now to 64bit.
536MHz * 3 planes * 3 pipes * pipe_scaler_ratio * plane_scaler_ratio * 100 = 536000 * 3 * 3 * (1.9 *1.9 (hscale & vscale)) * (1.9 * 1.9 (pipe hscale & vscale)) * 100 = 6286685040 = 0x1 76B7 3370
Quoted 21 lines
Quoted 21 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ +/* +* If there is any Ytile plane enabled and arbitrated display
drivers/gpu/drm/i915/intel_pm.c +4
-0+* bandwidth > 20% of raw system memory bandwidth +* Enale Y-tile related WA +* +* If memory is dual channel single rank, Xtile limit = 35%, else Xtile
drivers/gpu/drm/i915/intel_pm.c +10
-0+* limit = 60% +* If there is no Ytile plane enabled and +* arbitrated display bandwidth > Xtile limit +* Enable X-tile realated WA +*/ +if (y_tile_enabled && (display_bw_percentage >20 )) + results->mem_wa =WATERMARK_WA_Y_TILED ; +else { +int x_tile_percentage =60 ; +enum rank rank =DRAM_RANK_SINGLE ;
Quoted 9 lines
drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4205
Quoted 3 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ +if (memdev_info->rank !=DRAM_RANK_INVALID ) + rank = memdev_info->rank;I really think that the previous patch should prevent this. If
memdev_info->rankis invalid thenmemdev_info->validshould also be false and we'd have returned at the beginning of the function. With that, this part of the code could be simplified a little bit since then we'd have no need to choose a default rank.
done
Regards,
-Mahesh
Quoted 29 lines
Quoted 9 lines
drivers/gpu/drm/i915/intel_pm.c +3
-0+ + if ((rank == DRAM_RANK_SINGLE) && + (memdev_info->num_channels== 2))
drivers/gpu/drm/i915/intel_pm.c +4
-0+ x_tile_percentage =35 ; + +if (display_bw_percentage > x_tile_percentage) + results->mem_wa =WATERMARK_WA_X_TILED ;drivers/gpu/drm/i915/intel_pm.c skl_compute_ddb(struct drm_atomic_state *state) · line 4222
Quoted 10 lines
drivers/gpu/drm/i915/intel_pm.c +10
-0+ } + +return 0 ; + +exit: + results->mem_wa =WATERMARK_WA_Y_TILED ; +return 0 ; +} + +Two blank lines here.
Quoted 7 lines
static void
skl_copy_wm_for_pipe(structskl_wm_values*dst,
structskl_wm_values*src,drivers/gpu/drm/i915/intel_pm.c +0
-0@@ -4162,6 +4295,10 @@ skl_compute_wm(struct drm_atomic_state *state)/* Clear all dirty flags */ results->dirty_pipes =0 ;
Quoted 7 lines
drivers/gpu/drm/i915/intel_pm.c skl_compute_wm(struct drm_atomic_state *state) · line 4304
Quoted 7 lines
drivers/gpu/drm/i915/intel_pm.c +4
-0+ ret =skl_compute_memory_bandwidth_wm_wa (state); +if (ret) +return ret; + ret =skl_compute_ddb (state);if (ret)return ret;
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Patchwork
CI✓ Fi.CI.BAT: success for GEN-9 Arbitrated Bandwidth WM WA's & IPC (rev3)
In reply to Mahesh Kumar== Series Details ==
== Series Details ==
Series: GEN-9 Arbitrated Bandwidth WM WA's & IPC (rev3)
URL : https://patchwork.freedesktop.org/series/15562/
State : success
== Summary ==
Series 15562v3 GEN-9 Arbitrated Bandwidth WM WA's & IPC
https://patchwork.freedesktop.org/api/1.0/series/15562/revisions/3/mbox/
fi-bdw-5557u total:245 pass:226 dwarn:0 dfail:0 fail:4 skip:15 fi-bsw-n3050 total:245 pass:203 dwarn:0 dfail:0 fail:2 skip:40 fi-bxt-t5700 total:245 pass:213 dwarn:0 dfail:0 fail:4 skip:28 fi-byt-j1900 total:245 pass:214 dwarn:0 dfail:0 fail:3 skip:28 fi-byt-n2820 total:245 pass:210 dwarn:0 dfail:0 fail:3 skip:32 fi-hsw-4770 total:245 pass:221 dwarn:0 dfail:0 fail:4 skip:20 fi-hsw-4770r total:245 pass:221 dwarn:0 dfail:0 fail:4 skip:20 fi-ilk-650 total:245 pass:189 dwarn:0 dfail:0 fail:3 skip:53 fi-ivb-3520m total:245 pass:219 dwarn:0 dfail:0 fail:4 skip:22 fi-ivb-3770 total:245 pass:219 dwarn:0 dfail:0 fail:4 skip:22 fi-kbl-7500u total:245 pass:219 dwarn:0 dfail:0 fail:4 skip:22 fi-skl-6260u total:245 pass:227 dwarn:0 dfail:0 fail:4 skip:14 fi-skl-6700hq total:245 pass:220 dwarn:0 dfail:0 fail:4 skip:21 fi-skl-6700k total:245 pass:219 dwarn:1 dfail:0 fail:4 skip:21 fi-skl-6770hq total:245 pass:227 dwarn:0 dfail:0 fail:4 skip:14 fi-snb-2520m total:245 pass:210 dwarn:0 dfail:0 fail:3 skip:32 fi-snb-2600 total:245 pass:209 dwarn:0 dfail:0 fail:3 skip:33
15f4397c9b50b0212dcb99fb2b5b8a3867133e81 drm-tip: 2016y-12m-01d-13h-26m-26s UTC integration manifest b0ac9a9 drm/i915/gen9: WM memory bandwidth related workaround
f84d26c drm/i915: Decode system memory bandwidth
338d933 drm/i915: Add intel_atomic_get_existing_crtc_state function2b11985b drm/i915/skl+: change WM calc to fixed point 16.16
8e23bea drm/i915/bxt: Enable IPC support
fcc35ec drm/i915/kbl: IPC workaround for kabylake
31150d7 drm/i915/bxt: IPC WA for Broxton
c11e0c0 drm/i915/skl: Add variables to check x_tile and y_tile
== Logs ==
For more details see: https://intel-gfx-ci.01.org/CI/Patchwork_3163/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 6 lines
This series implements following set of functionality
Implement IPC WA's for Broxton/KBL
Enable IPC in supported platforms
Convert WM calculation to fixed point calculation Calculation of System memory Bandwidth for SKL/KBL/BXT Implementation of Arbitrated memory Bandwidth related WM WA's
Pushed patches 1, 2, 3, 5 and 6, with small bikesheds applied. Thanks for the patches!
(actually I had a little problem with dim, so right now only dinq has the patches, drm-tip doesn't, soon they will appear)
Let's resume the discussion on the remaining patches. For a possible next series, make sure to rebase and send only what's not applied.
Quoted 12 lines
Mahesh Kumar (8):
drm/i915/skl: Add variables to checkx_tileandy_tiledrm/i915/bxt: IPC WA for Broxton
drm/i915/kbl: IPC workaround for kabylake drm/i915/bxt: Enable IPC support
drm/i915/skl+: change WM calc to fixed point 16.16 drm/i915: Addintel_atomic_get_existing_crtc_statefunction drm/i915: Decode system memory bandwidth drm/i915/gen9: WM memory bandwidth related workaround
Quoted 9 lines
drivers/gpu/drm/i915/i915_drv.c | 175 ++++++++++++++++++++++ drivers/gpu/drm/i915/i915_drv.h | 109 ++++++++++++++ drivers/gpu/drm/i915/i915_reg.h | 38 +++++drivers/gpu/drm/i915/intel_display.c | 24 ++++
drivers/gpu/drm/i915/intel_drv.h | 15 ++
drivers/gpu/drm/i915/intel_pm.c | 272
+++++++++++++++++++++++++++++------6 files changed, 586 insertions(+), 47 deletions(-)
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 8 lines · Paulo Zanoni
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 6 lines
This series implements following set of functionality
Implement IPC WA's for Broxton/KBL
Enable IPC in supported platforms
Convert WM calculation to fixed point calculation Calculation of System memory Bandwidth for SKL/KBL/BXT Implementation of Arbitrated memory Bandwidth related WM WA's
Quoted 5 lines
Pushed patches 1, 2, 3, 5 and 6, with small bikesheds applied. Thanks for the patches!
(actually I had a little problem with dim, so right now only dinq has the patches, drm-tip doesn't, soon they will appear)
Hm, what did go boom with dim? If the pushing worked, then you can re-run just the drm-tip rebuilding with
$ dim rebuild-tip
Might also be worth it to upgrade to latest dim, we've dropped a few bugfixes in the branch push/pull logic in a few places since the big conversion, you might hit them.
If all this doesn't help, pls ping me on irc with a pastebin of what's going on. For testing you can always run
$ dim pq
To repush dinq - dim will complain if you'll do a non-ff push and abort, so it's safe.
-Daniel
Signature · 7 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Em Qui, 2016-12-08 às 17:00 +0100, Daniel Vetter escreveu:
Quoted 15 lines · Paulo Zanoni
Quoted 14 lines
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 7 lines
This series implements following set of functionality
Implement IPC WA's for Broxton/KBL
Enable IPC in supported platforms
Convert WM calculation to fixed point calculation Calculation of System memory Bandwidth for SKL/KBL/BXT Implementation of Arbitrated memory Bandwidth related WM WA'sPushed patches 1, 2, 3, 5 and 6, with small bikesheds applied. Thanks
for the patches!
Quoted 5 lines
Quoted 3 lines
(actually I had a little problem with dim, so right now only dinq has
the patches, drm-tip doesn't, soon they will appear)Hm, what did go boom with dim?
This was my first commit/push since the change. I had followed the instructions on your email (including the dim setup stage), and then when I did "dim push-queued" it complained that I didn't have the audio remotes in my tree. After the error message I was not sure what was the correct way to proceed.
I suppose we could change dim setup so that it really adds every single remote that's needed? Also, why doesn't it just add the remotes itself?
Quoted 4 lines
If the pushing worked, then you can re-run
just the drm-tip rebuilding with$ dim rebuild-tip
Good to know, thanks for that.
Quoted 4 lines
Might also be worth it to upgrade to latest dim, we've dropped a few bugfixes in the branch push/pull logic in a few places since the big conversion, you might hit them.
I always upgrade to the latest version before using it.
Quoted 4 lines
If all this doesn't help, pls ping me on irc with a pastebin of what's
going on. For testing you can always run
I sent you an email yesterday with the error message pasted... Maybe you'll find it soon :).
Quoted 7 lines
$ dim pqTo repush dinq - dim will complain if you'll do a non-ff push and abort,
so it's safe.
-Daniel
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 22 lines · Paulo R
Em Qui, 2016-12-08 às 17:00 +0100, Daniel Vetter escreveu:
Quoted 15 lines · Paulo Zanoni
Quoted 14 lines
Em Qui, 2016-12-01 às 21:19 +0530, Mahesh Kumar escreveu:
Quoted 7 lines
This series implements following set of functionality
Implement IPC WA's for Broxton/KBL
Enable IPC in supported platforms
Convert WM calculation to fixed point calculation Calculation of System memory Bandwidth for SKL/KBL/BXT Implementation of Arbitrated memory Bandwidth related WM WA'sPushed patches 1, 2, 3, 5 and 6, with small bikesheds applied. Thanks
for the patches!Quoted 5 lines
Quoted 3 lines
(actually I had a little problem with dim, so right now only dinq has
the patches, drm-tip doesn't, soon they will appear)Hm, what did go boom with dim?
Quoted 8 lines
This was my first commit/push since the change. I had followed the instructions on your email (including the dim setup stage), and then when I did "dim push-queued" it complained that I didn't have the audio remotes in my tree. After the error message I was not sure what was the correct way to proceed.
I suppose we could change dim setup so that it really adds every single remote that's needed? Also, why doesn't it just add the remotes itself?
dim doesn't require that you have a separate kernel repo for it, so for creating somewhat invasive stuff like naming remotes it defers to the user. It should print the exact git cmdline you need to run, only thing to fill in would be the name you want to give that remote.
I guess we could switch over to just auto-creating remotes. If you want to do that, the function you need to change is url_to_remote. Instead of the error output simply create the remote, and then set the local remote variable to the remote name. I'd be happy to ack such a patch for dim, since you're not the only one who got confused by this. Seems better to just create the remote ...
Quoted 15 lines
Quoted 4 lines
If the pushing worked, then you can re-run
just the drm-tip rebuilding with$ dim rebuild-tipGood to know, thanks for that.
Quoted 4 lines
Might also be worth it to upgrade to latest dim, we've dropped a few bugfixes in the branch push/pull logic in a few places since the big conversion, you might hit them.
I always upgrade to the latest version before using it.
Quoted 1 line
Quoted 6 lines
Quoted 3 lines
If all this doesn't help, pls ping me on irc with a pastebin of what's
going on. For testing you can always runI sent you an email yesterday with the error message pasted... Maybe you'll find it soon :).
I random walk through my various inboxes ;-)
-Daniel
Quoted 9 lines
Quoted 9 lines
$ dim pqTo repush dinq - dim will complain if you'll do a non-ff push and abort,
so it's safe.
-Daniel
Signature · 7 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx