uboot-mvebu: backport two patches for Marvell A38x
authorJosef Schlehofer <pepe.schlehofer@gmail.com>
Fri, 28 Jan 2022 12:39:45 +0000 (13:39 +0100)
committerHauke Mehrtens <hauke@hauke-m.de>
Thu, 3 Feb 2022 20:24:26 +0000 (21:24 +0100)
This solves issue with DDR training on Turris Omnia.

Log:
********   DRAM initialization Failed (res 0x1)   ********
DDR3 Training Sequence - FAILED
ERROR ### Please RESET the board ###

Signed-off-by: Josef Schlehofer <pepe.schlehofer@gmail.com>
package/boot/uboot-mvebu/patches/010-ddr-marvell-a38x-fix-split-out-mix.patch [new file with mode: 0644]
package/boot/uboot-mvebu/patches/011-ddr-marvell-a38x-fix-synchronous-asynchronous-mode.patch [new file with mode: 0644]

diff --git a/package/boot/uboot-mvebu/patches/010-ddr-marvell-a38x-fix-split-out-mix.patch b/package/boot/uboot-mvebu/patches/010-ddr-marvell-a38x-fix-split-out-mix.patch
new file mode 100644 (file)
index 0000000..17ff862
--- /dev/null
@@ -0,0 +1,116 @@
+From 3fc92a215b69ad448c151489228eb340df9a8703 Mon Sep 17 00:00:00 2001
+From: =?UTF-8?q?Marek=20Beh=C3=BAn?= <marek.behun@nic.cz>
+Date: Wed, 12 Jan 2022 17:06:59 +0100
+Subject: [PATCH] ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+This is a cleaned up and fixed version of a patch
+  mv_ddr: a380: fix SPLIT_OUT_MIX state decision
+
+  in each pattern cycle the bus state can be changed
+  in order to avoide it, need to back to the same bus state on each
+  pattern cycle
+by
+  Moti Boskula <motib@marvell.com>
+
+The original patch is not in Marvell's mv-ddr-marvell repository. It was
+gives to us by Marvell to fix an issues with DDR training on some
+boards, but it cannot be applied as is to mv-ddr-marvell, because it is
+a very dirty draft patch that would certainly break other things, mainly
+DDR4 training code in mv-ddr-marvell, since it changes common functions.
+
+I have cleaned up the patch and removed stuff that seemed unnecessary
+(when removed, it still fixed things). Note that I don't understand
+completely what the code does exactly, since I haven't studied the DDR
+training code extensively (and I suspect that no one besides some few
+people in Marvell understand the code completely).
+
+Anyway after the cleanup the patch still fixes isssues with DDR training
+on the failing boards.
+
+There was also a problem with the original patch on some of the Allied
+Telesis' x530 boards, reported by Chris Packham. I have asked Chris to
+send me some logs, and managed to fix it:
+- if you look at the change, you'll notice that it introduces
+  subtraction of cur_start_win[] and cur_end_win[] members, depending on
+  a bit set in the current_byte_status variable
+- the original patch subtracted cur_start_win[] if either
+  BYTE_SPLIT_OUT_MIX or BYTE_HOMOGENEOUS_SPLIT_OUT bits were set, but
+  subtracted cur_end_win[] only if the first one (BYTE_SPLIT_OUT_MIX)
+  was set
+- from Chris Packham logs I discovered that the x530 board where the
+  original patch introduced DDR training failure, only the
+  BYTE_HOMOGENEOUS_SPLIT_OUT bit was set, and on our boards where the
+  patch is needed only the BYTE_SPLIT_OUT_MIX is set in the
+  current_byte_status variable
+- this led me to the hypothesis that both cur_start_win[] and
+  cur_end_win[] should be subtracted only if BYTE_SPLIT_OUT_MIX bit is
+  set, the BYTE_HOMOGENEOUS_SPLIT_OUT bit shouldn't be considered at all
+- this hypothesis also gains credibility when considering the commit
+  title ("fix SPLIT_OUT_MIX state decision")
+
+Hopefully this will fix things without breaking anything else.
+
+Signed-off-by: Marek BehĂșn <marek.behun@nic.cz>
+Reviewed-by: Stefan Roese <sr@denx.de>
+Tested-by: Chris Packham <judge.packham@gmail.com>
+---
+ .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
+ 1 file changed, 26 insertions(+)
+
+--- a/drivers/ddr/marvell/a38x/ddr3_training_centralization.c
++++ b/drivers/ddr/marvell/a38x/ddr3_training_centralization.c
+@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 d
+       enum hws_training_ip_stat training_result[MAX_INTERFACE_NUM];
+       u32 if_id, pattern_id, bit_id;
+       u8 bus_id;
++      u8 current_byte_status;
+       u8 cur_start_win[BUS_WIDTH_IN_BITS];
+       u8 centralization_result[MAX_INTERFACE_NUM][BUS_WIDTH_IN_BITS];
+       u8 cur_end_win[BUS_WIDTH_IN_BITS];
+@@ -166,6 +167,10 @@ static int ddr3_tip_centralization(u32 d
+                                                 result[search_dir_id][7]));
+                               }
++                              current_byte_status =
++                                      mv_ddr_tip_sub_phy_byte_status_get(if_id,
++                                                                         bus_id);
++
+                               for (bit_id = 0; bit_id < BUS_WIDTH_IN_BITS;
+                                    bit_id++) {
+                                       /* check if this code is valid for 2 edge, probably not :( */
+@@ -174,11 +179,32 @@ static int ddr3_tip_centralization(u32 d
+                                                              [HWS_LOW2HIGH]
+                                                              [bit_id],
+                                                              EDGE_1);
++                                      if (current_byte_status &
++                                          BYTE_SPLIT_OUT_MIX) {
++                                              if (cur_start_win[bit_id] >= 64)
++                                                      cur_start_win[bit_id] -= 64;
++                                              else
++                                                      cur_start_win[bit_id] = 0;
++                                              DEBUG_CENTRALIZATION_ENGINE
++                                                      (DEBUG_LEVEL_INFO,
++                                                       ("pattern %d IF %d pup %d bit %d subtract 64 adll from start\n",
++                                                        pattern_id, if_id, bus_id, bit_id));
++                                      }
+                                       cur_end_win[bit_id] =
+                                               GET_TAP_RESULT(result
+                                                              [HWS_HIGH2LOW]
+                                                              [bit_id],
+                                                              EDGE_1);
++                                      if (cur_end_win[bit_id] >= 64 &&
++                                          (current_byte_status &
++                                           BYTE_SPLIT_OUT_MIX)) {
++                                              cur_end_win[bit_id] -= 64;
++                                              DEBUG_CENTRALIZATION_ENGINE
++                                                      (DEBUG_LEVEL_INFO,
++                                                       ("pattern %d IF %d pup %d bit %d subtract 64 adll from end\n",
++                                                        pattern_id, if_id, bus_id, bit_id));
++                                      }
++
+                                       /* window length */
+                                       current_window[bit_id] =
+                                               cur_end_win[bit_id] -
diff --git a/package/boot/uboot-mvebu/patches/011-ddr-marvell-a38x-fix-synchronous-asynchronous-mode.patch b/package/boot/uboot-mvebu/patches/011-ddr-marvell-a38x-fix-synchronous-asynchronous-mode.patch
new file mode 100644 (file)
index 0000000..9da1459
--- /dev/null
@@ -0,0 +1,98 @@
+From eadc4f512fb43bba2fa4e842c982da919da664be Mon Sep 17 00:00:00 2001
+From: =?UTF-8?q?Marek=20Beh=C3=BAn?= <marek.behun@nic.cz>
+Date: Tue, 4 Jan 2022 15:57:49 +0100
+Subject: [PATCH] ddr: marvell: a38x: Fix Synchronous vs Asynchronous mode
+ determination
+MIME-Version: 1.0
+Content-Type: text/plain; charset=UTF-8
+Content-Transfer-Encoding: 8bit
+
+Before commit 4c289425752f ("mv_ddr: a38x: add support for ddr async
+mode"), Asynchornous Mode was only used when the CPU Subsystem Clock
+Options[4:0] field in the SAR1 register was set to value 0x13: CPU at
+2 GHz and DDR at 933 MHz.
+
+Then commit 4c289425752f ("mv_ddr: a38x: add support for ddr async
+mode") added support for Asynchornous Modes with frequencies other than
+933 MHz (but at least 467 MHz), but the code it added to check for
+whether Asynchornous Mode should be used is wrong: it checks whether the
+frequency setting in board DDR topology map is set to value other than
+MV_DDR_FREQ_SAR.
+
+Thus boards which define a specific value, greater than 400 MHz, for DDR
+frequency in their board topology (e.g. Turris Omnia defines
+MV_DDR_FREQ_800), are incorrectly put into Asynchornous Mode after that
+commit.
+
+The A38x Functional Specification, section 10.12 DRAM Clocking, says:
+  In Synchornous mode, the DRAM and CPU clocks are edge aligned and run
+  in 1:2 or 1:3 CPU to DRAM frequency ratios.
+
+Change the check for whether Asynchornous Mode should be used according
+to this explanation in Functional Specification.
+
+Signed-off-by: Marek BehĂșn <marek.behun@nic.cz>
+Tested-by: Chris Packham <judge.packham@gmail.com>
+Reviewed-by: Stefan Roese <sr@denx.de>
+---
+ drivers/ddr/marvell/a38x/mv_ddr_plat.c | 19 ++++++++-----------
+ 1 file changed, 8 insertions(+), 11 deletions(-)
+
+--- a/drivers/ddr/marvell/a38x/mv_ddr_plat.c
++++ b/drivers/ddr/marvell/a38x/mv_ddr_plat.c
+@@ -167,8 +167,6 @@ static u16 a38x_vco_freq_per_sar_ref_clk
+ };
+-static u32 async_mode_at_tf;
+-
+ static u32 dq_bit_map_2_phy_pin[] = {
+       1, 0, 2, 6, 9, 8, 3, 7, /* 0 */
+       8, 9, 1, 7, 2, 6, 3, 0, /* 1 */
+@@ -734,7 +732,8 @@ static int ddr3_tip_a38x_set_divider(u8
+       u32 divider = 0;
+       u32 sar_val, ref_clk_satr;
+       u32 async_val;
+-      u32 freq = mv_ddr_freq_get(frequency);
++      u32 cpu_freq;
++      u32 ddr_freq = mv_ddr_freq_get(frequency);
+       if (if_id != 0) {
+               DEBUG_TRAINING_ACCESS(DEBUG_LEVEL_ERROR,
+@@ -751,11 +750,14 @@ static int ddr3_tip_a38x_set_divider(u8
+       ref_clk_satr = reg_read(DEVICE_SAMPLE_AT_RESET2_REG);
+       if (((ref_clk_satr >> DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_OFFSET) & 0x1) ==
+           DEVICE_SAMPLE_AT_RESET2_REG_REFCLK_25MHZ)
+-              divider = a38x_vco_freq_per_sar_ref_clk_25_mhz[sar_val] / freq;
++              cpu_freq = a38x_vco_freq_per_sar_ref_clk_25_mhz[sar_val];
+       else
+-              divider = a38x_vco_freq_per_sar_ref_clk_40_mhz[sar_val] / freq;
++              cpu_freq = a38x_vco_freq_per_sar_ref_clk_40_mhz[sar_val];
++
++      divider = cpu_freq / ddr_freq;
+-      if ((async_mode_at_tf == 1) && (freq > 400)) {
++      if (((cpu_freq % ddr_freq != 0) || (divider != 2 && divider != 3)) &&
++          (ddr_freq > 400)) {
+               /* Set async mode */
+               dunit_write(0x20220, 0x1000, 0x1000);
+               dunit_write(0xe42f4, 0x200, 0x200);
+@@ -869,8 +871,6 @@ int ddr3_tip_ext_write(u32 dev_num, u32
+ int mv_ddr_early_init(void)
+ {
+-      struct mv_ddr_topology_map *tm = mv_ddr_topology_map_get();
+-
+       /* FIXME: change this configuration per ddr type
+        * configure a380 and a390 to work with receiver odt timing
+        * the odt_config is defined:
+@@ -882,9 +882,6 @@ int mv_ddr_early_init(void)
+       mv_ddr_sw_db_init(0, 0);
+-      if (tm->interface_params[0].memory_freq != MV_DDR_FREQ_SAR)
+-              async_mode_at_tf = 1;
+-
+       return MV_OK;
+ }