Skip to contentSkip to conversation

ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

PATCH In review Marek Behún 4 participants 7 messages Jan 12, 2022 – Jan 17, 2022

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

From: Marek Behún <[email protected]>

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 <[email protected]>

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.

---
 .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +26 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
 	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 dev_num, u32 mode)
 						  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 dev_num, u32 mode)
 							       [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] -
Signature · 1 line
2.34.1
Reviewed-by on v1 1/1
#1 lore raw

+Cc Chris

Quoted 47 lines · Marek Behún

From: Marek Behún <[email protected]>

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 <[email protected]>

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")

Quoted 3 lines

Hopefully this will fix things without breaking anything else.

I would be good though, if Chris (and others) could also add a comment (RB or TB tag).

Thanks,
Stefan

Quoted 51 lines
---
  .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
  1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +19 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
  	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 dev_num, u32 mode)
  						  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 dev_num, u32 mode)
  							       [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)) {

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode) · line 210

Quoted 10 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +7 -0
+						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] -

Viele Grüße,
Stefan Roese

Signature · 3 lines
DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-51 Fax: (+49)-8142-66989-80 Email: [email protected]
Tested-by
#2 lore raw
Quoted 50 lines · Stefan Roese,

+Cc Chris

Quoted 47 lines · Marek Behún

From: Marek Behún <[email protected]>

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 <[email protected]>

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")

Quoted 9 lines
Quoted 3 lines

Hopefully this will fix things without breaking anything else.

I would be good though, if Chris (and others) could also add a comment (RB or TB tag).

Yes this is the same patch that Marek sent me privately.

Quoted 91 lines

Thanks,
Stefan

Review

Quoted 5 lines
---
  .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
  1 file changed, 26 insertions(+)
(quoted diff) +0 -0

b/drivers/ddr/marvell/a38x/ddr3_training_centralization.c

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32

Quoted 4 lines

index 648b37ef6f..42308b6965 100644

drivers/ddr/marvell/a38x/ddr3_training_centralization.c +0 -0
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32

mode)

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32

Quoted 8 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +1 -0
      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 dev_num, u32

mode)

Quoted 1 line

result[search_dir_id][7]));

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode) · line 173

Quoted 4 lines
                              }
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +2 -0
+                             current_byte_status =
+

mv_ddr_tip_sub_phy_byte_status_get(if_id,

Quoted 1 line

+

bus_id);

Quoted 2 lines

+

                              for (bit_id = 0; bit_id <

BUS_WIDTH_IN_BITS;

Quoted 2 lines
                                   bit_id++) {
                                      /* check if this code is valid for

2 edge, probably not :( */

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num,

Quoted 1 line
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +0 -0
@@ -174,11 +179,32 @@ static int ddr3_tip_centralization(u32 dev_num,

u32 mode)

Quoted 1 line

[HWS_LOW2HIGH]

Quoted 7 lines
                                                             [bit_id],
                                                             EDGE_1);
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +3 -0
+                                     if (current_byte_status &
+                                         BYTE_SPLIT_OUT_MIX) {
+                                             if (cur_start_win[bit_id]

= 64)
+

cur_start_win[bit_id] -= 64;

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode) · line 207

Quoted 2 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +2 -0
+                                             else
+

cur_start_win[bit_id] = 0;

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode)

Quoted 3 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +3 -0
+                                             DEBUG_CENTRALIZATION_ENGINE
+                                                     (DEBUG_LEVEL_INFO,
+                                                      ("pattern %d IF

%d pup %d bit %d subtract 64 adll from start\n",

Quoted 1 line

+ pattern_id,

if_id, bus_id, bit_id));

Quoted 4 lines

+ }

                                      cur_end_win[bit_id] =
                                              GET_TAP_RESULT(result

[HWS_HIGH2LOW]

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode)

Quoted 9 lines
                                                             [bit_id],
                                                             EDGE_1);
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +7 -0
+                                     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",

Quoted 1 line

+ pattern_id,

if_id, bus_id, bit_id));

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode) · line 210

Quoted 6 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +2 -0
+                                     }
+
                                      /* window length */
                                      current_window[bit_id] =
                                              cur_end_win[bit_id] -

Viele Grüße,
Stefan Roese

--
DENX Software Engineering GmbH, Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-51 Fax: (+49)-8142-66989-80 Email: [email protected]

#3 lore raw
In reply to Marek Behún
Quoted 47 lines · Marek Behún

From: Marek Behún <[email protected]>

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 <[email protected]>

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")

Quoted 3 lines

Hopefully this will fix things without breaking anything else.

Applied to u-boot-marvell/master

Thanks,
Stefan

Quoted 51 lines
---
  .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
  1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +19 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
  	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 dev_num, u32 mode)
  						  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 dev_num, u32 mode)
  							       [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)) {

drivers/ddr/marvell/a38x/ddr3_training_centralization.c static int ddr3_tip_centralization(u32 dev_num, u32 mode) · line 210

Quoted 10 lines
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +7 -0
+						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] -

Viele Grüße,
Stefan Roese

Signature · 3 lines
DENX Software Engineering GmbH,      Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-51 Fax: (+49)-8142-66989-80 Email: [email protected]
#4 lore raw
In reply to Marek Behún

Hi Assaf,
I've got this email a few days ago.
Is it related to what you’ve send me?
Thanks

-----Original Message-----
From: Marek Behún <[email protected]>
Sent: Wednesday, January 12, 2022 6:07 PM
To: Stefan Roese <[email protected]>; Mario Six <[email protected]>; Dennis Gilmore <[email protected]>; Kostya Porotchkin <[email protected]>
Cc: Pali Rohár <[email protected]>; [email protected]; Moti Buskila <[email protected]>; Marek Behún <[email protected]>
Subject: [EXT] [PATCH u-boot-marvell] ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

External Email

----------------------------------------------------------------------
From: Marek Behún <[email protected]>

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 <[email protected]>

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.

---
 .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +5 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
 	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 dev_num, u32 mode)
 						  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 dev_num, u32 mode)

[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] -
--
2.34.1

#5 lore raw

Hello Moti,

since you're the author of the original version of this patch, could you please review it and if it is okay, put it into mv-ddr-marvell?

Thanks.

Marek

Quoted 118 lines · On Mon, 17 Jan 2022 06:52:08 +0000 Moti Buskila <[email protected]> wrote:

Hi Assaf,
I've got this email a few days ago.
Is it related to what you’ve send me?
Thanks

-----Original Message-----
From: Marek Behún <[email protected]>
Sent: Wednesday, January 12, 2022 6:07 PM
To: Stefan Roese <[email protected]>; Mario Six <[email protected]>; Dennis Gilmore <[email protected]>; Kostya Porotchkin <[email protected]>
Cc: Pali Rohár <[email protected]>; [email protected]; Moti Buskila <[email protected]>; Marek Behún <[email protected]>
Subject: [EXT] [PATCH u-boot-marvell] ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

External Email

----------------------------------------------------------------------
From: Marek Behún <[email protected]>

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 <[email protected]>

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.

---
 .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +5 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
 	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 dev_num, u32 mode)
 						  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 dev_num, u32 mode)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +21 -1
 							       [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] -
--

2.34.1

#6 lore raw

Hi Marek
This mail was send by accident.
Assaf H, is taking care of that.
Thanks

-----Original Message-----
From: Marek Behún <[email protected]>
Sent: Monday, January 17, 2022 4:23 PM
To: Moti Buskila <[email protected]>
Cc: Stefan Roese <[email protected]>; Mario Six <[email protected]>; Dennis Gilmore <[email protected]>; Kostya Porotchkin <[email protected]>; Pali Rohár <[email protected]>; [email protected]; Marek Behún <[email protected]>
Subject: Re: [EXT] [PATCH u-boot-marvell] ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

Hello Moti,

since you're the author of the original version of this patch, could you please review it and if it is okay, put it into mv-ddr-marvell?

Thanks.

Marek

Quoted 127 lines · On Mon, 17 Jan 2022 06:52:08 +0000 Moti Buskila <[email protected]> wrote:

Hi Assaf,
I've got this email a few days ago.
Is it related to what you’ve send me?
Thanks

-----Original Message-----
From: Marek Behún <[email protected]>
Sent: Wednesday, January 12, 2022 6:07 PM
To: Stefan Roese <[email protected]>; Mario Six <[email protected]>; Dennis Gilmore <[email protected]>; Kostya Porotchkin <[email protected]>
Cc: Pali Rohár <[email protected]>; [email protected]; Moti Buskila <[email protected]>; Marek Behún <[email protected]>
Subject: [EXT] [PATCH u-boot-marvell] ddr: marvell: a38x: fix SPLIT_OUT_MIX state decision

External Email

----------------------------------------------------------------------
From: Marek Behún <[email protected]>

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 <[email protected]>

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.

---
 .../a38x/ddr3_training_centralization.c       | 26 +++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +5 -0
index 648b37ef6f..42308b6965 100644
@@ -55,6 +55,7 @@ static int ddr3_tip_centralization(u32 dev_num, u32 mode)
 	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 dev_num, u32 mode)
 						  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 dev_num, u32 mode)
drivers/ddr/marvell/a38x/ddr3_training_centralization.c +21 -1
 							       [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] -
--

2.34.1